2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright 1998, 2000 Marshall Kirk McKusick.
5 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org>
8 * The soft updates code is derived from the appendix of a University
9 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
10 * "Soft Updates: A Solution to the Metadata Update Problem in File
11 * Systems", CSE-TR-254-95, August 1995).
13 * Further information about soft updates can be obtained from:
15 * Marshall Kirk McKusick http://www.mckusick.com/softdep/
16 * 1614 Oxford Street mckusick@mckusick.com
17 * Berkeley, CA 94709-1608 +1-510-843-9542
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * 1. Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * 2. Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in the
28 * documentation and/or other materials provided with the distribution.
30 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
31 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
32 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
33 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT,
34 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
35 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
36 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
37 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
38 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
39 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00
44 #include <sys/cdefs.h>
45 __FBSDID("$FreeBSD$");
48 #include "opt_quota.h"
52 * For now we want the safety net that the DEBUG flag provides.
58 #include <sys/param.h>
59 #include <sys/kernel.h>
60 #include <sys/systm.h>
64 #include <sys/kthread.h>
66 #include <sys/limits.h>
68 #include <sys/malloc.h>
69 #include <sys/mount.h>
70 #include <sys/mutex.h>
71 #include <sys/namei.h>
74 #include <sys/racct.h>
75 #include <sys/rwlock.h>
77 #include <sys/sysctl.h>
78 #include <sys/syslog.h>
79 #include <sys/vnode.h>
82 #include <ufs/ufs/dir.h>
83 #include <ufs/ufs/extattr.h>
84 #include <ufs/ufs/quota.h>
85 #include <ufs/ufs/inode.h>
86 #include <ufs/ufs/ufsmount.h>
87 #include <ufs/ffs/fs.h>
88 #include <ufs/ffs/softdep.h>
89 #include <ufs/ffs/ffs_extern.h>
90 #include <ufs/ufs/ufs_extern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_object.h>
96 #include <geom/geom.h>
100 #define KTR_SUJ 0 /* Define to KTR_SPARE. */
105 softdep_flushfiles(oldmnt, flags, td)
106 struct mount *oldmnt;
111 panic("softdep_flushfiles called");
115 softdep_mount(devvp, mp, fs, cred)
133 softdep_uninitialize()
144 panic("softdep_unmount called");
148 softdep_setup_sbupdate(ump, fs, bp)
149 struct ufsmount *ump;
154 panic("softdep_setup_sbupdate called");
158 softdep_setup_inomapdep(bp, ip, newinum, mode)
165 panic("softdep_setup_inomapdep called");
169 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
172 ufs2_daddr_t newblkno;
177 panic("softdep_setup_blkmapdep called");
181 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
184 ufs2_daddr_t newblkno;
185 ufs2_daddr_t oldblkno;
191 panic("softdep_setup_allocdirect called");
195 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
198 ufs2_daddr_t newblkno;
199 ufs2_daddr_t oldblkno;
205 panic("softdep_setup_allocext called");
209 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
214 ufs2_daddr_t newblkno;
215 ufs2_daddr_t oldblkno;
219 panic("softdep_setup_allocindir_page called");
223 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
228 ufs2_daddr_t newblkno;
231 panic("softdep_setup_allocindir_meta called");
235 softdep_journal_freeblocks(ip, cred, length, flags)
242 panic("softdep_journal_freeblocks called");
246 softdep_journal_fsync(ip)
250 panic("softdep_journal_fsync called");
254 softdep_setup_freeblocks(ip, length, flags)
260 panic("softdep_setup_freeblocks called");
264 softdep_freefile(pvp, ino, mode)
270 panic("softdep_freefile called");
274 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
279 struct buf *newdirbp;
283 panic("softdep_setup_directory_add called");
287 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
296 panic("softdep_change_directoryentry_offset called");
300 softdep_setup_remove(bp, dp, ip, isrmdir)
307 panic("softdep_setup_remove called");
311 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
319 panic("softdep_setup_directory_change called");
323 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
328 struct workhead *wkhd;
331 panic("%s called", __FUNCTION__);
335 softdep_setup_inofree(mp, bp, ino, wkhd)
339 struct workhead *wkhd;
342 panic("%s called", __FUNCTION__);
346 softdep_setup_unlink(dp, ip)
351 panic("%s called", __FUNCTION__);
355 softdep_setup_link(dp, ip)
360 panic("%s called", __FUNCTION__);
364 softdep_revert_link(dp, ip)
369 panic("%s called", __FUNCTION__);
373 softdep_setup_rmdir(dp, ip)
378 panic("%s called", __FUNCTION__);
382 softdep_revert_rmdir(dp, ip)
387 panic("%s called", __FUNCTION__);
391 softdep_setup_create(dp, ip)
396 panic("%s called", __FUNCTION__);
400 softdep_revert_create(dp, ip)
405 panic("%s called", __FUNCTION__);
409 softdep_setup_mkdir(dp, ip)
414 panic("%s called", __FUNCTION__);
418 softdep_revert_mkdir(dp, ip)
423 panic("%s called", __FUNCTION__);
427 softdep_setup_dotdot_link(dp, ip)
432 panic("%s called", __FUNCTION__);
436 softdep_prealloc(vp, waitok)
441 panic("%s called", __FUNCTION__);
445 softdep_journal_lookup(mp, vpp)
454 softdep_change_linkcnt(ip)
458 panic("softdep_change_linkcnt called");
462 softdep_load_inodeblock(ip)
466 panic("softdep_load_inodeblock called");
470 softdep_update_inodeblock(ip, bp, waitfor)
476 panic("softdep_update_inodeblock called");
481 struct vnode *vp; /* the "in_core" copy of the inode */
488 softdep_fsync_mountdev(vp)
496 softdep_flushworklist(oldmnt, countp, td)
497 struct mount *oldmnt;
507 softdep_sync_metadata(struct vnode *vp)
510 panic("softdep_sync_metadata called");
514 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
517 panic("softdep_sync_buf called");
525 panic("softdep_slowdown called");
529 softdep_request_cleanup(fs, vp, cred, resource)
540 softdep_check_suspend(struct mount *mp,
543 int softdep_accdepcnt,
544 int secondary_writes,
545 int secondary_accwrites)
550 (void) softdep_depcnt,
551 (void) softdep_accdepcnt;
553 bo = &devvp->v_bufobj;
554 ASSERT_BO_WLOCKED(bo);
557 while (mp->mnt_secondary_writes != 0) {
559 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
560 (PUSER - 1) | PDROP, "secwr", 0);
566 * Reasons for needing more work before suspend:
567 * - Dirty buffers on devvp.
568 * - Secondary writes occurred after start of vnode sync loop
571 if (bo->bo_numoutput > 0 ||
572 bo->bo_dirty.bv_cnt > 0 ||
573 secondary_writes != 0 ||
574 mp->mnt_secondary_writes != 0 ||
575 secondary_accwrites != mp->mnt_secondary_accwrites)
582 softdep_get_depcounts(struct mount *mp,
584 int *softdepactiveaccp)
588 *softdepactiveaccp = 0;
592 softdep_buf_append(bp, wkhd)
594 struct workhead *wkhd;
597 panic("softdep_buf_appendwork called");
601 softdep_inode_append(ip, cred, wkhd)
604 struct workhead *wkhd;
607 panic("softdep_inode_appendwork called");
611 softdep_freework(wkhd)
612 struct workhead *wkhd;
615 panic("softdep_freework called");
620 FEATURE(softupdates, "FFS soft-updates support");
622 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0,
623 "soft updates stats");
624 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0,
625 "total dependencies allocated");
626 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0,
627 "high use dependencies allocated");
628 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0,
629 "current dependencies allocated");
630 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0,
631 "current dependencies written");
633 unsigned long dep_current[D_LAST + 1];
634 unsigned long dep_highuse[D_LAST + 1];
635 unsigned long dep_total[D_LAST + 1];
636 unsigned long dep_write[D_LAST + 1];
638 #define SOFTDEP_TYPE(type, str, long) \
639 static MALLOC_DEFINE(M_ ## type, #str, long); \
640 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \
641 &dep_total[D_ ## type], 0, ""); \
642 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \
643 &dep_current[D_ ## type], 0, ""); \
644 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \
645 &dep_highuse[D_ ## type], 0, ""); \
646 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \
647 &dep_write[D_ ## type], 0, "");
649 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies");
650 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies");
651 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap,
652 "Block or frag allocated from cyl group map");
653 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency");
654 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode");
655 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies");
656 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block");
657 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode");
658 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode");
659 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated");
660 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry");
661 SOFTDEP_TYPE(MKDIR, mkdir, "New directory");
662 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted");
663 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block");
664 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block");
665 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free");
666 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add");
667 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove");
668 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move");
669 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block");
670 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block");
671 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag");
672 SOFTDEP_TYPE(JSEG, jseg, "Journal segment");
673 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete");
674 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency");
675 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation");
676 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete");
678 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel");
680 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes");
681 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations");
682 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data");
684 #define M_SOFTDEP_FLAGS (M_WAITOK)
687 * translate from workitem type to memory type
688 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
690 static struct malloc_type *memtype[] = {
722 #define DtoM(type) (memtype[type])
725 * Names of malloc types.
727 #define TYPENAME(type) \
728 ((unsigned)(type) <= D_LAST && (unsigned)(type) >= D_FIRST ? \
729 memtype[type]->ks_shortdesc : "???")
731 * End system adaptation definitions.
734 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino)
735 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino)
738 * Internal function prototypes.
740 static void check_clear_deps(struct mount *);
741 static void softdep_error(char *, int);
742 static int softdep_process_worklist(struct mount *, int);
743 static int softdep_waitidle(struct mount *, int);
744 static void drain_output(struct vnode *);
745 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int);
746 static int check_inodedep_free(struct inodedep *);
747 static void clear_remove(struct mount *);
748 static void clear_inodedeps(struct mount *);
749 static void unlinked_inodedep(struct mount *, struct inodedep *);
750 static void clear_unlinked_inodedep(struct inodedep *);
751 static struct inodedep *first_unlinked_inodedep(struct ufsmount *);
752 static int flush_pagedep_deps(struct vnode *, struct mount *,
754 static int free_pagedep(struct pagedep *);
755 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t);
756 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t);
757 static int flush_deplist(struct allocdirectlst *, int, int *);
758 static int sync_cgs(struct mount *, int);
759 static int handle_written_filepage(struct pagedep *, struct buf *, int);
760 static int handle_written_sbdep(struct sbdep *, struct buf *);
761 static void initiate_write_sbdep(struct sbdep *);
762 static void diradd_inode_written(struct diradd *, struct inodedep *);
763 static int handle_written_indirdep(struct indirdep *, struct buf *,
765 static int handle_written_inodeblock(struct inodedep *, struct buf *, int);
766 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *,
768 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int);
769 static void handle_written_jaddref(struct jaddref *);
770 static void handle_written_jremref(struct jremref *);
771 static void handle_written_jseg(struct jseg *, struct buf *);
772 static void handle_written_jnewblk(struct jnewblk *);
773 static void handle_written_jblkdep(struct jblkdep *);
774 static void handle_written_jfreefrag(struct jfreefrag *);
775 static void complete_jseg(struct jseg *);
776 static void complete_jsegs(struct jseg *);
777 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *);
778 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *);
779 static void jremref_write(struct jremref *, struct jseg *, uint8_t *);
780 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *);
781 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *);
782 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data);
783 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *);
784 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *);
785 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *);
786 static inline void inoref_write(struct inoref *, struct jseg *,
788 static void handle_allocdirect_partdone(struct allocdirect *,
790 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *,
792 static void indirdep_complete(struct indirdep *);
793 static int indirblk_lookup(struct mount *, ufs2_daddr_t);
794 static void indirblk_insert(struct freework *);
795 static void indirblk_remove(struct freework *);
796 static void handle_allocindir_partdone(struct allocindir *);
797 static void initiate_write_filepage(struct pagedep *, struct buf *);
798 static void initiate_write_indirdep(struct indirdep*, struct buf *);
799 static void handle_written_mkdir(struct mkdir *, int);
800 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *,
802 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *);
803 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
804 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
805 static void handle_workitem_freefile(struct freefile *);
806 static int handle_workitem_remove(struct dirrem *, int);
807 static struct dirrem *newdirrem(struct buf *, struct inode *,
808 struct inode *, int, struct dirrem **);
809 static struct indirdep *indirdep_lookup(struct mount *, struct inode *,
811 static void cancel_indirdep(struct indirdep *, struct buf *,
813 static void free_indirdep(struct indirdep *);
814 static void free_diradd(struct diradd *, struct workhead *);
815 static void merge_diradd(struct inodedep *, struct diradd *);
816 static void complete_diradd(struct diradd *);
817 static struct diradd *diradd_lookup(struct pagedep *, int);
818 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *,
820 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *,
822 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *,
823 struct jremref *, struct jremref *);
824 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *,
826 static void cancel_allocindir(struct allocindir *, struct buf *bp,
827 struct freeblks *, int);
828 static int setup_trunc_indir(struct freeblks *, struct inode *,
829 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t);
830 static void complete_trunc_indir(struct freework *);
831 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *,
833 static void complete_mkdir(struct mkdir *);
834 static void free_newdirblk(struct newdirblk *);
835 static void free_jremref(struct jremref *);
836 static void free_jaddref(struct jaddref *);
837 static void free_jsegdep(struct jsegdep *);
838 static void free_jsegs(struct jblocks *);
839 static void rele_jseg(struct jseg *);
840 static void free_jseg(struct jseg *, struct jblocks *);
841 static void free_jnewblk(struct jnewblk *);
842 static void free_jblkdep(struct jblkdep *);
843 static void free_jfreefrag(struct jfreefrag *);
844 static void free_freedep(struct freedep *);
845 static void journal_jremref(struct dirrem *, struct jremref *,
847 static void cancel_jnewblk(struct jnewblk *, struct workhead *);
848 static int cancel_jaddref(struct jaddref *, struct inodedep *,
850 static void cancel_jfreefrag(struct jfreefrag *);
851 static inline void setup_freedirect(struct freeblks *, struct inode *,
853 static inline void setup_freeext(struct freeblks *, struct inode *, int, int);
854 static inline void setup_freeindir(struct freeblks *, struct inode *, int,
856 static inline struct freeblks *newfreeblks(struct mount *, struct inode *);
857 static void freeblks_free(struct ufsmount *, struct freeblks *, int);
858 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t);
859 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t);
860 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int);
861 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t,
863 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int);
864 static int cancel_pagedep(struct pagedep *, struct freeblks *, int);
865 static int deallocate_dependencies(struct buf *, struct freeblks *, int);
866 static void newblk_freefrag(struct newblk*);
867 static void free_newblk(struct newblk *);
868 static void cancel_allocdirect(struct allocdirectlst *,
869 struct allocdirect *, struct freeblks *);
870 static int check_inode_unwritten(struct inodedep *);
871 static int free_inodedep(struct inodedep *);
872 static void freework_freeblock(struct freework *, u_long);
873 static void freework_enqueue(struct freework *);
874 static int handle_workitem_freeblocks(struct freeblks *, int);
875 static int handle_complete_freeblocks(struct freeblks *, int);
876 static void handle_workitem_indirblk(struct freework *);
877 static void handle_written_freework(struct freework *);
878 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
879 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *,
881 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *,
882 struct inodedep *, struct allocindir *, ufs_lbn_t);
883 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
884 ufs2_daddr_t, ufs_lbn_t);
885 static void handle_workitem_freefrag(struct freefrag *);
886 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long,
888 static void allocdirect_merge(struct allocdirectlst *,
889 struct allocdirect *, struct allocdirect *);
890 static struct freefrag *allocindir_merge(struct allocindir *,
891 struct allocindir *);
892 static int bmsafemap_find(struct bmsafemap_hashhead *, int,
893 struct bmsafemap **);
894 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *,
895 int cg, struct bmsafemap *);
896 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int,
898 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **);
899 static int inodedep_find(struct inodedep_hashhead *, ino_t,
901 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
902 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t,
903 int, struct pagedep **);
904 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
906 static void pause_timer(void *);
907 static int request_cleanup(struct mount *, int);
908 static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *);
909 static void schedule_cleanup(struct mount *);
910 static void softdep_ast_cleanup_proc(struct thread *);
911 static struct ufsmount *softdep_bp_to_mp(struct buf *bp);
912 static int process_worklist_item(struct mount *, int, int);
913 static void process_removes(struct vnode *);
914 static void process_truncates(struct vnode *);
915 static void jwork_move(struct workhead *, struct workhead *);
916 static void jwork_insert(struct workhead *, struct jsegdep *);
917 static void add_to_worklist(struct worklist *, int);
918 static void wake_worklist(struct worklist *);
919 static void wait_worklist(struct worklist *, char *);
920 static void remove_from_worklist(struct worklist *);
921 static void softdep_flush(void *);
922 static void softdep_flushjournal(struct mount *);
923 static int softdep_speedup(struct ufsmount *);
924 static void worklist_speedup(struct mount *);
925 static int journal_mount(struct mount *, struct fs *, struct ucred *);
926 static void journal_unmount(struct ufsmount *);
927 static int journal_space(struct ufsmount *, int);
928 static void journal_suspend(struct ufsmount *);
929 static int journal_unsuspend(struct ufsmount *ump);
930 static void softdep_prelink(struct vnode *, struct vnode *);
931 static void add_to_journal(struct worklist *);
932 static void remove_from_journal(struct worklist *);
933 static bool softdep_excess_items(struct ufsmount *, int);
934 static void softdep_process_journal(struct mount *, struct worklist *, int);
935 static struct jremref *newjremref(struct dirrem *, struct inode *,
936 struct inode *ip, off_t, nlink_t);
937 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t,
939 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t,
941 static inline struct jsegdep *inoref_jseg(struct inoref *);
942 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t);
943 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t,
945 static void adjust_newfreework(struct freeblks *, int);
946 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int);
947 static void move_newblock_dep(struct jaddref *, struct inodedep *);
948 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t);
949 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *,
950 ufs2_daddr_t, long, ufs_lbn_t);
951 static struct freework *newfreework(struct ufsmount *, struct freeblks *,
952 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int);
953 static int jwait(struct worklist *, int);
954 static struct inodedep *inodedep_lookup_ip(struct inode *);
955 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *);
956 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *);
957 static void handle_jwork(struct workhead *);
958 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *,
960 static struct jblocks *jblocks_create(void);
961 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *);
962 static void jblocks_free(struct jblocks *, struct mount *, int);
963 static void jblocks_destroy(struct jblocks *);
964 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int);
967 * Exported softdep operations.
969 static void softdep_disk_io_initiation(struct buf *);
970 static void softdep_disk_write_complete(struct buf *);
971 static void softdep_deallocate_dependencies(struct buf *);
972 static int softdep_count_dependencies(struct buf *bp, int);
975 * Global lock over all of soft updates.
977 static struct mtx lk;
978 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF);
980 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk)
981 #define FREE_GBLLOCK(lk) mtx_unlock(lk)
982 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED)
985 * Per-filesystem soft-updates locking.
987 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock)
988 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock)
989 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock)
990 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock)
991 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \
994 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock)
995 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock)
998 * Worklist queue management.
999 * These routines require that the lock be held.
1001 #ifndef /* NOT */ DEBUG
1002 #define WORKLIST_INSERT(head, item) do { \
1003 (item)->wk_state |= ONWORKLIST; \
1004 LIST_INSERT_HEAD(head, item, wk_list); \
1006 #define WORKLIST_REMOVE(item) do { \
1007 (item)->wk_state &= ~ONWORKLIST; \
1008 LIST_REMOVE(item, wk_list); \
1010 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT
1011 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE
1014 static void worklist_insert(struct workhead *, struct worklist *, int);
1015 static void worklist_remove(struct worklist *, int);
1017 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1)
1018 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0)
1019 #define WORKLIST_REMOVE(item) worklist_remove(item, 1)
1020 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0)
1023 worklist_insert(head, item, locked)
1024 struct workhead *head;
1025 struct worklist *item;
1030 LOCK_OWNED(VFSTOUFS(item->wk_mp));
1031 if (item->wk_state & ONWORKLIST)
1032 panic("worklist_insert: %p %s(0x%X) already on list",
1033 item, TYPENAME(item->wk_type), item->wk_state);
1034 item->wk_state |= ONWORKLIST;
1035 LIST_INSERT_HEAD(head, item, wk_list);
1039 worklist_remove(item, locked)
1040 struct worklist *item;
1045 LOCK_OWNED(VFSTOUFS(item->wk_mp));
1046 if ((item->wk_state & ONWORKLIST) == 0)
1047 panic("worklist_remove: %p %s(0x%X) not on list",
1048 item, TYPENAME(item->wk_type), item->wk_state);
1049 item->wk_state &= ~ONWORKLIST;
1050 LIST_REMOVE(item, wk_list);
1055 * Merge two jsegdeps keeping only the oldest one as newer references
1056 * can't be discarded until after older references.
1058 static inline struct jsegdep *
1059 jsegdep_merge(struct jsegdep *one, struct jsegdep *two)
1061 struct jsegdep *swp;
1066 if (one->jd_seg->js_seq > two->jd_seg->js_seq) {
1071 WORKLIST_REMOVE(&two->jd_list);
1078 * If two freedeps are compatible free one to reduce list size.
1080 static inline struct freedep *
1081 freedep_merge(struct freedep *one, struct freedep *two)
1086 if (one->fd_freework == two->fd_freework) {
1087 WORKLIST_REMOVE(&two->fd_list);
1094 * Move journal work from one list to another. Duplicate freedeps and
1095 * jsegdeps are coalesced to keep the lists as small as possible.
1098 jwork_move(dst, src)
1099 struct workhead *dst;
1100 struct workhead *src;
1102 struct freedep *freedep;
1103 struct jsegdep *jsegdep;
1104 struct worklist *wkn;
1105 struct worklist *wk;
1108 ("jwork_move: dst == src"));
1111 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) {
1112 if (wk->wk_type == D_JSEGDEP)
1113 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1114 else if (wk->wk_type == D_FREEDEP)
1115 freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1118 while ((wk = LIST_FIRST(src)) != NULL) {
1119 WORKLIST_REMOVE(wk);
1120 WORKLIST_INSERT(dst, wk);
1121 if (wk->wk_type == D_JSEGDEP) {
1122 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1125 if (wk->wk_type == D_FREEDEP)
1126 freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1131 jwork_insert(dst, jsegdep)
1132 struct workhead *dst;
1133 struct jsegdep *jsegdep;
1135 struct jsegdep *jsegdepn;
1136 struct worklist *wk;
1138 LIST_FOREACH(wk, dst, wk_list)
1139 if (wk->wk_type == D_JSEGDEP)
1142 WORKLIST_INSERT(dst, &jsegdep->jd_list);
1145 jsegdepn = WK_JSEGDEP(wk);
1146 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) {
1147 WORKLIST_REMOVE(wk);
1148 free_jsegdep(jsegdepn);
1149 WORKLIST_INSERT(dst, &jsegdep->jd_list);
1151 free_jsegdep(jsegdep);
1155 * Routines for tracking and managing workitems.
1157 static void workitem_free(struct worklist *, int);
1158 static void workitem_alloc(struct worklist *, int, struct mount *);
1159 static void workitem_reassign(struct worklist *, int);
1161 #define WORKITEM_FREE(item, type) \
1162 workitem_free((struct worklist *)(item), (type))
1163 #define WORKITEM_REASSIGN(item, type) \
1164 workitem_reassign((struct worklist *)(item), (type))
1167 workitem_free(item, type)
1168 struct worklist *item;
1171 struct ufsmount *ump;
1174 if (item->wk_state & ONWORKLIST)
1175 panic("workitem_free: %s(0x%X) still on list",
1176 TYPENAME(item->wk_type), item->wk_state);
1177 if (item->wk_type != type && type != D_NEWBLK)
1178 panic("workitem_free: type mismatch %s != %s",
1179 TYPENAME(item->wk_type), TYPENAME(type));
1181 if (item->wk_state & IOWAITING)
1183 ump = VFSTOUFS(item->wk_mp);
1185 KASSERT(ump->softdep_deps > 0,
1186 ("workitem_free: %s: softdep_deps going negative",
1187 ump->um_fs->fs_fsmnt));
1188 if (--ump->softdep_deps == 0 && ump->softdep_req)
1189 wakeup(&ump->softdep_deps);
1190 KASSERT(dep_current[item->wk_type] > 0,
1191 ("workitem_free: %s: dep_current[%s] going negative",
1192 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1193 KASSERT(ump->softdep_curdeps[item->wk_type] > 0,
1194 ("workitem_free: %s: softdep_curdeps[%s] going negative",
1195 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1196 atomic_subtract_long(&dep_current[item->wk_type], 1);
1197 ump->softdep_curdeps[item->wk_type] -= 1;
1198 free(item, DtoM(type));
1202 workitem_alloc(item, type, mp)
1203 struct worklist *item;
1207 struct ufsmount *ump;
1209 item->wk_type = type;
1214 ACQUIRE_GBLLOCK(&lk);
1215 dep_current[type]++;
1216 if (dep_current[type] > dep_highuse[type])
1217 dep_highuse[type] = dep_current[type];
1221 ump->softdep_curdeps[type] += 1;
1222 ump->softdep_deps++;
1223 ump->softdep_accdeps++;
1228 workitem_reassign(item, newtype)
1229 struct worklist *item;
1232 struct ufsmount *ump;
1234 ump = VFSTOUFS(item->wk_mp);
1236 KASSERT(ump->softdep_curdeps[item->wk_type] > 0,
1237 ("workitem_reassign: %s: softdep_curdeps[%s] going negative",
1238 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1239 ump->softdep_curdeps[item->wk_type] -= 1;
1240 ump->softdep_curdeps[newtype] += 1;
1241 KASSERT(dep_current[item->wk_type] > 0,
1242 ("workitem_reassign: %s: dep_current[%s] going negative",
1243 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1244 ACQUIRE_GBLLOCK(&lk);
1245 dep_current[newtype]++;
1246 dep_current[item->wk_type]--;
1247 if (dep_current[newtype] > dep_highuse[newtype])
1248 dep_highuse[newtype] = dep_current[newtype];
1249 dep_total[newtype]++;
1251 item->wk_type = newtype;
1255 * Workitem queue management
1257 static int max_softdeps; /* maximum number of structs before slowdown */
1258 static int tickdelay = 2; /* number of ticks to pause during slowdown */
1259 static int proc_waiting; /* tracks whether we have a timeout posted */
1260 static int *stat_countp; /* statistic to count in proc_waiting timeout */
1261 static struct callout softdep_callout;
1262 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
1263 static int req_clear_remove; /* syncer process flush some freeblks */
1264 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */
1267 * runtime statistics
1269 static int stat_flush_threads; /* number of softdep flushing threads */
1270 static int stat_worklist_push; /* number of worklist cleanups */
1271 static int stat_blk_limit_push; /* number of times block limit neared */
1272 static int stat_ino_limit_push; /* number of times inode limit neared */
1273 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
1274 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
1275 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
1276 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
1277 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
1278 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
1279 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
1280 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */
1281 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */
1282 static int stat_journal_min; /* Times hit journal min threshold */
1283 static int stat_journal_low; /* Times hit journal low threshold */
1284 static int stat_journal_wait; /* Times blocked in jwait(). */
1285 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */
1286 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */
1287 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */
1288 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */
1289 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */
1290 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */
1291 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */
1292 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */
1293 static int stat_cleanup_failures; /* Number of cleanup requests that failed */
1294 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */
1296 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW,
1297 &max_softdeps, 0, "");
1298 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW,
1300 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD,
1301 &stat_flush_threads, 0, "");
1302 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW,
1303 &stat_worklist_push, 0,"");
1304 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW,
1305 &stat_blk_limit_push, 0,"");
1306 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW,
1307 &stat_ino_limit_push, 0,"");
1308 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW,
1309 &stat_blk_limit_hit, 0, "");
1310 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW,
1311 &stat_ino_limit_hit, 0, "");
1312 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW,
1313 &stat_sync_limit_hit, 0, "");
1314 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW,
1315 &stat_indir_blk_ptrs, 0, "");
1316 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW,
1317 &stat_inode_bitmap, 0, "");
1318 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW,
1319 &stat_direct_blk_ptrs, 0, "");
1320 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW,
1321 &stat_dir_entry, 0, "");
1322 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW,
1323 &stat_jaddref, 0, "");
1324 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW,
1325 &stat_jnewblk, 0, "");
1326 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW,
1327 &stat_journal_low, 0, "");
1328 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW,
1329 &stat_journal_min, 0, "");
1330 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW,
1331 &stat_journal_wait, 0, "");
1332 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW,
1333 &stat_jwait_filepage, 0, "");
1334 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW,
1335 &stat_jwait_freeblks, 0, "");
1336 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW,
1337 &stat_jwait_inode, 0, "");
1338 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW,
1339 &stat_jwait_newblk, 0, "");
1340 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW,
1341 &stat_cleanup_blkrequests, 0, "");
1342 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW,
1343 &stat_cleanup_inorequests, 0, "");
1344 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW,
1345 &stat_cleanup_high_delay, 0, "");
1346 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW,
1347 &stat_cleanup_retries, 0, "");
1348 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW,
1349 &stat_cleanup_failures, 0, "");
1350 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW,
1351 &softdep_flushcache, 0, "");
1352 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD,
1353 &stat_emptyjblocks, 0, "");
1355 SYSCTL_DECL(_vfs_ffs);
1357 /* Whether to recompute the summary at mount time */
1358 static int compute_summary_at_mount = 0;
1359 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
1360 &compute_summary_at_mount, 0, "Recompute summary at mount");
1361 static int print_threads = 0;
1362 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW,
1363 &print_threads, 0, "Notify flusher thread start/stop");
1365 /* List of all filesystems mounted with soft updates */
1366 static TAILQ_HEAD(, mount_softdeps) softdepmounts;
1369 * This function cleans the worklist for a filesystem.
1370 * Each filesystem running with soft dependencies gets its own
1371 * thread to run in this function. The thread is started up in
1372 * softdep_mount and shutdown in softdep_unmount. They show up
1373 * as part of the kernel "bufdaemon" process whose process
1374 * entry is available in bufdaemonproc.
1376 static int searchfailed;
1377 extern struct proc *bufdaemonproc;
1384 struct ufsmount *ump;
1387 td->td_pflags |= TDP_NORUNNINGBUF;
1388 mp = (struct mount *)addr;
1390 atomic_add_int(&stat_flush_threads, 1);
1392 ump->softdep_flags &= ~FLUSH_STARTING;
1393 wakeup(&ump->softdep_flushtd);
1395 if (print_threads) {
1396 if (stat_flush_threads == 1)
1397 printf("Running %s at pid %d\n", bufdaemonproc->p_comm,
1398 bufdaemonproc->p_pid);
1399 printf("Start thread %s\n", td->td_name);
1402 while (softdep_process_worklist(mp, 0) > 0 ||
1404 VFSTOUFS(mp)->softdep_jblocks->jb_suspended))
1405 kthread_suspend_check();
1407 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1408 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM,
1410 ump->softdep_flags &= ~FLUSH_CLEANUP;
1412 * Check to see if we are done and need to exit.
1414 if ((ump->softdep_flags & FLUSH_EXIT) == 0) {
1418 ump->softdep_flags &= ~FLUSH_EXIT;
1420 wakeup(&ump->softdep_flags);
1422 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups);
1423 atomic_subtract_int(&stat_flush_threads, 1);
1425 panic("kthread_exit failed\n");
1430 worklist_speedup(mp)
1433 struct ufsmount *ump;
1437 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1438 ump->softdep_flags |= FLUSH_CLEANUP;
1439 wakeup(&ump->softdep_flushtd);
1443 softdep_speedup(ump)
1444 struct ufsmount *ump;
1446 struct ufsmount *altump;
1447 struct mount_softdeps *sdp;
1450 worklist_speedup(ump->um_mountp);
1453 * If we have global shortages, then we need other
1454 * filesystems to help with the cleanup. Here we wakeup a
1455 * flusher thread for a filesystem that is over its fair
1456 * share of resources.
1458 if (req_clear_inodedeps || req_clear_remove) {
1459 ACQUIRE_GBLLOCK(&lk);
1460 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) {
1461 if ((altump = sdp->sd_ump) == ump)
1463 if (((req_clear_inodedeps &&
1464 altump->softdep_curdeps[D_INODEDEP] >
1465 max_softdeps / stat_flush_threads) ||
1466 (req_clear_remove &&
1467 altump->softdep_curdeps[D_DIRREM] >
1468 (max_softdeps / 2) / stat_flush_threads)) &&
1469 TRY_ACQUIRE_LOCK(altump))
1477 * Move to the end of the list so we pick a
1478 * different one on out next try.
1480 TAILQ_REMOVE(&softdepmounts, sdp, sd_next);
1481 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next);
1483 if ((altump->softdep_flags &
1484 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1485 altump->softdep_flags |= FLUSH_CLEANUP;
1486 altump->um_softdep->sd_cleanups++;
1487 wakeup(&altump->softdep_flushtd);
1491 return (speedup_syncer());
1495 * Add an item to the end of the work queue.
1496 * This routine requires that the lock be held.
1497 * This is the only routine that adds items to the list.
1498 * The following routine is the only one that removes items
1499 * and does so in order from first to last.
1502 #define WK_HEAD 0x0001 /* Add to HEAD. */
1503 #define WK_NODELAY 0x0002 /* Process immediately. */
1506 add_to_worklist(wk, flags)
1507 struct worklist *wk;
1510 struct ufsmount *ump;
1512 ump = VFSTOUFS(wk->wk_mp);
1514 if (wk->wk_state & ONWORKLIST)
1515 panic("add_to_worklist: %s(0x%X) already on list",
1516 TYPENAME(wk->wk_type), wk->wk_state);
1517 wk->wk_state |= ONWORKLIST;
1518 if (ump->softdep_on_worklist == 0) {
1519 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
1520 ump->softdep_worklist_tail = wk;
1521 } else if (flags & WK_HEAD) {
1522 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
1524 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
1525 ump->softdep_worklist_tail = wk;
1527 ump->softdep_on_worklist += 1;
1528 if (flags & WK_NODELAY)
1529 worklist_speedup(wk->wk_mp);
1533 * Remove the item to be processed. If we are removing the last
1534 * item on the list, we need to recalculate the tail pointer.
1537 remove_from_worklist(wk)
1538 struct worklist *wk;
1540 struct ufsmount *ump;
1542 ump = VFSTOUFS(wk->wk_mp);
1543 if (ump->softdep_worklist_tail == wk)
1544 ump->softdep_worklist_tail =
1545 (struct worklist *)wk->wk_list.le_prev;
1546 WORKLIST_REMOVE(wk);
1547 ump->softdep_on_worklist -= 1;
1552 struct worklist *wk;
1554 if (wk->wk_state & IOWAITING) {
1555 wk->wk_state &= ~IOWAITING;
1561 wait_worklist(wk, wmesg)
1562 struct worklist *wk;
1565 struct ufsmount *ump;
1567 ump = VFSTOUFS(wk->wk_mp);
1568 wk->wk_state |= IOWAITING;
1569 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0);
1573 * Process that runs once per second to handle items in the background queue.
1575 * Note that we ensure that everything is done in the order in which they
1576 * appear in the queue. The code below depends on this property to ensure
1577 * that blocks of a file are freed before the inode itself is freed. This
1578 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
1579 * until all the old ones have been purged from the dependency lists.
1582 softdep_process_worklist(mp, full)
1587 struct ufsmount *ump;
1590 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
1591 if (MOUNTEDSOFTDEP(mp) == 0)
1596 starttime = time_second;
1597 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0);
1598 check_clear_deps(mp);
1599 while (ump->softdep_on_worklist > 0) {
1600 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0)
1604 check_clear_deps(mp);
1606 * We do not generally want to stop for buffer space, but if
1607 * we are really being a buffer hog, we will stop and wait.
1609 if (should_yield()) {
1611 kern_yield(PRI_USER);
1616 * Never allow processing to run for more than one
1617 * second. This gives the syncer thread the opportunity
1618 * to pause if appropriate.
1620 if (!full && starttime != time_second)
1624 journal_unsuspend(ump);
1630 * Process all removes associated with a vnode if we are running out of
1631 * journal space. Any other process which attempts to flush these will
1632 * be unable as we have the vnodes locked.
1638 struct inodedep *inodedep;
1639 struct dirrem *dirrem;
1640 struct ufsmount *ump;
1647 inum = VTOI(vp)->i_number;
1650 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
1652 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) {
1654 * If another thread is trying to lock this vnode
1655 * it will fail but we must wait for it to do so
1656 * before we can proceed.
1658 if (dirrem->dm_state & INPROGRESS) {
1659 wait_worklist(&dirrem->dm_list, "pwrwait");
1662 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) ==
1663 (COMPLETE | ONWORKLIST))
1668 remove_from_worklist(&dirrem->dm_list);
1670 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1671 panic("process_removes: suspended filesystem");
1672 handle_workitem_remove(dirrem, 0);
1673 vn_finished_secondary_write(mp);
1679 * Process all truncations associated with a vnode if we are running out
1680 * of journal space. This is called when the vnode lock is already held
1681 * and no other process can clear the truncation. This function returns
1682 * a value greater than zero if it did any work.
1685 process_truncates(vp)
1688 struct inodedep *inodedep;
1689 struct freeblks *freeblks;
1690 struct ufsmount *ump;
1698 inum = VTOI(vp)->i_number;
1700 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
1703 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) {
1704 /* Journal entries not yet written. */
1705 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) {
1707 &freeblks->fb_jblkdephd)->jb_list,
1711 /* Another thread is executing this item. */
1712 if (freeblks->fb_state & INPROGRESS) {
1713 wait_worklist(&freeblks->fb_list, "ptrwait");
1716 /* Freeblks is waiting on a inode write. */
1717 if ((freeblks->fb_state & COMPLETE) == 0) {
1723 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) ==
1724 (ALLCOMPLETE | ONWORKLIST)) {
1725 remove_from_worklist(&freeblks->fb_list);
1726 freeblks->fb_state |= INPROGRESS;
1728 if (vn_start_secondary_write(NULL, &mp,
1730 panic("process_truncates: "
1731 "suspended filesystem");
1732 handle_workitem_freeblocks(freeblks, 0);
1733 vn_finished_secondary_write(mp);
1737 if (freeblks->fb_cgwait)
1742 sync_cgs(mp, MNT_WAIT);
1743 ffs_sync_snap(mp, MNT_WAIT);
1747 if (freeblks == NULL)
1754 * Process one item on the worklist.
1757 process_worklist_item(mp, target, flags)
1762 struct worklist sentinel;
1763 struct worklist *wk;
1764 struct ufsmount *ump;
1768 KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
1770 * If we are being called because of a process doing a
1771 * copy-on-write, then it is not safe to write as we may
1772 * recurse into the copy-on-write routine.
1774 if (curthread->td_pflags & TDP_COWINPROGRESS)
1776 PHOLD(curproc); /* Don't let the stack go away. */
1780 sentinel.wk_mp = NULL;
1781 sentinel.wk_type = D_SENTINEL;
1782 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list);
1783 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL;
1784 wk = LIST_NEXT(&sentinel, wk_list)) {
1785 if (wk->wk_type == D_SENTINEL) {
1786 LIST_REMOVE(&sentinel, wk_list);
1787 LIST_INSERT_AFTER(wk, &sentinel, wk_list);
1790 if (wk->wk_state & INPROGRESS)
1791 panic("process_worklist_item: %p already in progress.",
1793 wk->wk_state |= INPROGRESS;
1794 remove_from_worklist(wk);
1796 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1797 panic("process_worklist_item: suspended filesystem");
1798 switch (wk->wk_type) {
1800 /* removal of a directory entry */
1801 error = handle_workitem_remove(WK_DIRREM(wk), flags);
1805 /* releasing blocks and/or fragments from a file */
1806 error = handle_workitem_freeblocks(WK_FREEBLKS(wk),
1811 /* releasing a fragment when replaced as a file grows */
1812 handle_workitem_freefrag(WK_FREEFRAG(wk));
1817 /* releasing an inode when its link count drops to 0 */
1818 handle_workitem_freefile(WK_FREEFILE(wk));
1823 panic("%s_process_worklist: Unknown type %s",
1824 "softdep", TYPENAME(wk->wk_type));
1827 vn_finished_secondary_write(mp);
1830 if (++matchcnt == target)
1835 * We have to retry the worklist item later. Wake up any
1836 * waiters who may be able to complete it immediately and
1837 * add the item back to the head so we don't try to execute
1840 wk->wk_state &= ~INPROGRESS;
1842 add_to_worklist(wk, WK_HEAD);
1844 /* Sentinal could've become the tail from remove_from_worklist. */
1845 if (ump->softdep_worklist_tail == &sentinel)
1846 ump->softdep_worklist_tail =
1847 (struct worklist *)sentinel.wk_list.le_prev;
1848 LIST_REMOVE(&sentinel, wk_list);
1854 * Move dependencies from one buffer to another.
1857 softdep_move_dependencies(oldbp, newbp)
1861 struct worklist *wk, *wktail;
1862 struct ufsmount *ump;
1865 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL)
1867 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
1868 ("softdep_move_dependencies called on non-softdep filesystem"));
1871 ump = VFSTOUFS(wk->wk_mp);
1873 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
1874 LIST_REMOVE(wk, wk_list);
1875 if (wk->wk_type == D_BMSAFEMAP &&
1876 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp))
1879 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1881 LIST_INSERT_AFTER(wktail, wk, wk_list);
1890 * Purge the work list of all items associated with a particular mount point.
1893 softdep_flushworklist(oldmnt, countp, td)
1894 struct mount *oldmnt;
1898 struct vnode *devvp;
1899 struct ufsmount *ump;
1903 * Alternately flush the block device associated with the mount
1904 * point and process any dependencies that the flushing
1905 * creates. We continue until no more worklist dependencies
1910 ump = VFSTOUFS(oldmnt);
1911 devvp = ump->um_devvp;
1912 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1914 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1915 error = VOP_FSYNC(devvp, MNT_WAIT, td);
1916 VOP_UNLOCK(devvp, 0);
1923 #define SU_WAITIDLE_RETRIES 20
1925 softdep_waitidle(struct mount *mp, int flags __unused)
1927 struct ufsmount *ump;
1928 struct vnode *devvp;
1933 devvp = ump->um_devvp;
1937 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) {
1938 ump->softdep_req = 1;
1939 KASSERT((flags & FORCECLOSE) == 0 ||
1940 ump->softdep_on_worklist == 0,
1941 ("softdep_waitidle: work added after flush"));
1942 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP,
1943 "softdeps", 10 * hz);
1944 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1945 error = VOP_FSYNC(devvp, MNT_WAIT, td);
1946 VOP_UNLOCK(devvp, 0);
1951 ump->softdep_req = 0;
1952 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) {
1954 printf("softdep_waitidle: Failed to flush worklist for %p\n",
1962 * Flush all vnodes and worklist items associated with a specified mount point.
1965 softdep_flushfiles(oldmnt, flags, td)
1966 struct mount *oldmnt;
1971 struct ufsmount *ump;
1974 int error, early, depcount, loopcnt, retry_flush_count, retry;
1977 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0,
1978 ("softdep_flushfiles called on non-softdep filesystem"));
1980 retry_flush_count = 3;
1985 * Alternately flush the vnodes associated with the mount
1986 * point and process any dependencies that the flushing
1987 * creates. In theory, this loop can happen at most twice,
1988 * but we give it a few extra just to be sure.
1990 for (; loopcnt > 0; loopcnt--) {
1992 * Do another flush in case any vnodes were brought in
1993 * as part of the cleanup operations.
1995 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag &
1996 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH;
1997 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0)
1999 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
2004 * If we are unmounting then it is an error to fail. If we
2005 * are simply trying to downgrade to read-only, then filesystem
2006 * activity can keep us busy forever, so we just fail with EBUSY.
2009 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
2010 panic("softdep_flushfiles: looping");
2014 error = softdep_waitidle(oldmnt, flags);
2016 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
2019 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
2020 ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
2021 morework = oldmnt->mnt_nvnodelistsize > 0;
2023 ump = VFSTOUFS(oldmnt);
2025 for (i = 0; i < MAXQUOTAS; i++) {
2026 if (ump->um_quotas[i] != NULLVP)
2032 if (--retry_flush_count > 0) {
2038 MNT_IUNLOCK(oldmnt);
2047 * Structure hashing.
2049 * There are four types of structures that can be looked up:
2050 * 1) pagedep structures identified by mount point, inode number,
2051 * and logical block.
2052 * 2) inodedep structures identified by mount point and inode number.
2053 * 3) newblk structures identified by mount point and
2054 * physical block number.
2055 * 4) bmsafemap structures identified by mount point and
2056 * cylinder group number.
2058 * The "pagedep" and "inodedep" dependency structures are hashed
2059 * separately from the file blocks and inodes to which they correspond.
2060 * This separation helps when the in-memory copy of an inode or
2061 * file block must be replaced. It also obviates the need to access
2062 * an inode or file page when simply updating (or de-allocating)
2063 * dependency structures. Lookup of newblk structures is needed to
2064 * find newly allocated blocks when trying to associate them with
2065 * their allocdirect or allocindir structure.
2067 * The lookup routines optionally create and hash a new instance when
2068 * an existing entry is not found. The bmsafemap lookup routine always
2069 * allocates a new structure if an existing one is not found.
2071 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
2074 * Structures and routines associated with pagedep caching.
2076 #define PAGEDEP_HASH(ump, inum, lbn) \
2077 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size])
2080 pagedep_find(pagedephd, ino, lbn, pagedeppp)
2081 struct pagedep_hashhead *pagedephd;
2084 struct pagedep **pagedeppp;
2086 struct pagedep *pagedep;
2088 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
2089 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) {
2090 *pagedeppp = pagedep;
2098 * Look up a pagedep. Return 1 if found, 0 otherwise.
2099 * If not found, allocate if DEPALLOC flag is passed.
2100 * Found or allocated entry is returned in pagedeppp.
2103 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp)
2109 struct pagedep **pagedeppp;
2111 struct pagedep *pagedep;
2112 struct pagedep_hashhead *pagedephd;
2113 struct worklist *wk;
2114 struct ufsmount *ump;
2121 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2122 if (wk->wk_type == D_PAGEDEP) {
2123 *pagedeppp = WK_PAGEDEP(wk);
2128 pagedephd = PAGEDEP_HASH(ump, ino, lbn);
2129 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp);
2131 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp)
2132 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list);
2135 if ((flags & DEPALLOC) == 0)
2138 pagedep = malloc(sizeof(struct pagedep),
2139 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
2140 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
2142 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp);
2145 * This should never happen since we only create pagedeps
2146 * with the vnode lock held. Could be an assert.
2148 WORKITEM_FREE(pagedep, D_PAGEDEP);
2151 pagedep->pd_ino = ino;
2152 pagedep->pd_lbn = lbn;
2153 LIST_INIT(&pagedep->pd_dirremhd);
2154 LIST_INIT(&pagedep->pd_pendinghd);
2155 for (i = 0; i < DAHASHSZ; i++)
2156 LIST_INIT(&pagedep->pd_diraddhd[i]);
2157 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
2158 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2159 *pagedeppp = pagedep;
2164 * Structures and routines associated with inodedep caching.
2166 #define INODEDEP_HASH(ump, inum) \
2167 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size])
2170 inodedep_find(inodedephd, inum, inodedeppp)
2171 struct inodedep_hashhead *inodedephd;
2173 struct inodedep **inodedeppp;
2175 struct inodedep *inodedep;
2177 LIST_FOREACH(inodedep, inodedephd, id_hash)
2178 if (inum == inodedep->id_ino)
2181 *inodedeppp = inodedep;
2189 * Look up an inodedep. Return 1 if found, 0 if not found.
2190 * If not found, allocate if DEPALLOC flag is passed.
2191 * Found or allocated entry is returned in inodedeppp.
2194 inodedep_lookup(mp, inum, flags, inodedeppp)
2198 struct inodedep **inodedeppp;
2200 struct inodedep *inodedep;
2201 struct inodedep_hashhead *inodedephd;
2202 struct ufsmount *ump;
2208 inodedephd = INODEDEP_HASH(ump, inum);
2210 if (inodedep_find(inodedephd, inum, inodedeppp))
2212 if ((flags & DEPALLOC) == 0)
2215 * If the system is over its limit and our filesystem is
2216 * responsible for more than our share of that usage and
2217 * we are not in a rush, request some inodedep cleanup.
2219 if (softdep_excess_items(ump, D_INODEDEP))
2220 schedule_cleanup(mp);
2223 inodedep = malloc(sizeof(struct inodedep),
2224 M_INODEDEP, M_SOFTDEP_FLAGS);
2225 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
2227 if (inodedep_find(inodedephd, inum, inodedeppp)) {
2228 WORKITEM_FREE(inodedep, D_INODEDEP);
2231 inodedep->id_fs = fs;
2232 inodedep->id_ino = inum;
2233 inodedep->id_state = ALLCOMPLETE;
2234 inodedep->id_nlinkdelta = 0;
2235 inodedep->id_savedino1 = NULL;
2236 inodedep->id_savedsize = -1;
2237 inodedep->id_savedextsize = -1;
2238 inodedep->id_savednlink = -1;
2239 inodedep->id_bmsafemap = NULL;
2240 inodedep->id_mkdiradd = NULL;
2241 LIST_INIT(&inodedep->id_dirremhd);
2242 LIST_INIT(&inodedep->id_pendinghd);
2243 LIST_INIT(&inodedep->id_inowait);
2244 LIST_INIT(&inodedep->id_bufwait);
2245 TAILQ_INIT(&inodedep->id_inoreflst);
2246 TAILQ_INIT(&inodedep->id_inoupdt);
2247 TAILQ_INIT(&inodedep->id_newinoupdt);
2248 TAILQ_INIT(&inodedep->id_extupdt);
2249 TAILQ_INIT(&inodedep->id_newextupdt);
2250 TAILQ_INIT(&inodedep->id_freeblklst);
2251 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
2252 *inodedeppp = inodedep;
2257 * Structures and routines associated with newblk caching.
2259 #define NEWBLK_HASH(ump, inum) \
2260 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size])
2263 newblk_find(newblkhd, newblkno, flags, newblkpp)
2264 struct newblk_hashhead *newblkhd;
2265 ufs2_daddr_t newblkno;
2267 struct newblk **newblkpp;
2269 struct newblk *newblk;
2271 LIST_FOREACH(newblk, newblkhd, nb_hash) {
2272 if (newblkno != newblk->nb_newblkno)
2275 * If we're creating a new dependency don't match those that
2276 * have already been converted to allocdirects. This is for
2279 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK)
2292 * Look up a newblk. Return 1 if found, 0 if not found.
2293 * If not found, allocate if DEPALLOC flag is passed.
2294 * Found or allocated entry is returned in newblkpp.
2297 newblk_lookup(mp, newblkno, flags, newblkpp)
2299 ufs2_daddr_t newblkno;
2301 struct newblk **newblkpp;
2303 struct newblk *newblk;
2304 struct newblk_hashhead *newblkhd;
2305 struct ufsmount *ump;
2309 newblkhd = NEWBLK_HASH(ump, newblkno);
2310 if (newblk_find(newblkhd, newblkno, flags, newblkpp))
2312 if ((flags & DEPALLOC) == 0)
2314 if (softdep_excess_items(ump, D_NEWBLK) ||
2315 softdep_excess_items(ump, D_ALLOCDIRECT) ||
2316 softdep_excess_items(ump, D_ALLOCINDIR))
2317 schedule_cleanup(mp);
2320 newblk = malloc(sizeof(union allblk), M_NEWBLK,
2321 M_SOFTDEP_FLAGS | M_ZERO);
2322 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp);
2324 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) {
2325 WORKITEM_FREE(newblk, D_NEWBLK);
2328 newblk->nb_freefrag = NULL;
2329 LIST_INIT(&newblk->nb_indirdeps);
2330 LIST_INIT(&newblk->nb_newdirblk);
2331 LIST_INIT(&newblk->nb_jwork);
2332 newblk->nb_state = ATTACHED;
2333 newblk->nb_newblkno = newblkno;
2334 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
2340 * Structures and routines associated with freed indirect block caching.
2342 #define INDIR_HASH(ump, blkno) \
2343 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size])
2346 * Lookup an indirect block in the indir hash table. The freework is
2347 * removed and potentially freed. The caller must do a blocking journal
2348 * write before writing to the blkno.
2351 indirblk_lookup(mp, blkno)
2355 struct freework *freework;
2356 struct indir_hashhead *wkhd;
2357 struct ufsmount *ump;
2360 wkhd = INDIR_HASH(ump, blkno);
2361 TAILQ_FOREACH(freework, wkhd, fw_next) {
2362 if (freework->fw_blkno != blkno)
2364 indirblk_remove(freework);
2371 * Insert an indirect block represented by freework into the indirblk
2372 * hash table so that it may prevent the block from being re-used prior
2373 * to the journal being written.
2376 indirblk_insert(freework)
2377 struct freework *freework;
2379 struct jblocks *jblocks;
2381 struct ufsmount *ump;
2383 ump = VFSTOUFS(freework->fw_list.wk_mp);
2384 jblocks = ump->softdep_jblocks;
2385 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst);
2389 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs);
2390 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework,
2392 freework->fw_state &= ~DEPCOMPLETE;
2396 indirblk_remove(freework)
2397 struct freework *freework;
2399 struct ufsmount *ump;
2401 ump = VFSTOUFS(freework->fw_list.wk_mp);
2402 LIST_REMOVE(freework, fw_segs);
2403 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next);
2404 freework->fw_state |= DEPCOMPLETE;
2405 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
2406 WORKITEM_FREE(freework, D_FREEWORK);
2410 * Executed during filesystem system initialization before
2411 * mounting any filesystems.
2414 softdep_initialize()
2417 TAILQ_INIT(&softdepmounts);
2419 max_softdeps = desiredvnodes * 4;
2421 max_softdeps = desiredvnodes * 2;
2424 /* initialise bioops hack */
2425 bioops.io_start = softdep_disk_io_initiation;
2426 bioops.io_complete = softdep_disk_write_complete;
2427 bioops.io_deallocate = softdep_deallocate_dependencies;
2428 bioops.io_countdeps = softdep_count_dependencies;
2429 softdep_ast_cleanup = softdep_ast_cleanup_proc;
2431 /* Initialize the callout with an mtx. */
2432 callout_init_mtx(&softdep_callout, &lk, 0);
2436 * Executed after all filesystems have been unmounted during
2437 * filesystem module unload.
2440 softdep_uninitialize()
2443 /* clear bioops hack */
2444 bioops.io_start = NULL;
2445 bioops.io_complete = NULL;
2446 bioops.io_deallocate = NULL;
2447 bioops.io_countdeps = NULL;
2448 softdep_ast_cleanup = NULL;
2450 callout_drain(&softdep_callout);
2454 * Called at mount time to notify the dependency code that a
2455 * filesystem wishes to use it.
2458 softdep_mount(devvp, mp, fs, cred)
2459 struct vnode *devvp;
2464 struct csum_total cstotal;
2465 struct mount_softdeps *sdp;
2466 struct ufsmount *ump;
2472 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA,
2475 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
2476 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
2477 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
2478 MNTK_SOFTDEP | MNTK_NOASYNC;
2481 ump->um_softdep = sdp;
2483 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock");
2485 LIST_INIT(&ump->softdep_workitem_pending);
2486 LIST_INIT(&ump->softdep_journal_pending);
2487 TAILQ_INIT(&ump->softdep_unlinked);
2488 LIST_INIT(&ump->softdep_dirtycg);
2489 ump->softdep_worklist_tail = NULL;
2490 ump->softdep_on_worklist = 0;
2491 ump->softdep_deps = 0;
2492 LIST_INIT(&ump->softdep_mkdirlisthd);
2493 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
2494 &ump->pagedep_hash_size);
2495 ump->pagedep_nextclean = 0;
2496 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP,
2497 &ump->inodedep_hash_size);
2498 ump->inodedep_nextclean = 0;
2499 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK,
2500 &ump->newblk_hash_size);
2501 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP,
2502 &ump->bmsafemap_hash_size);
2503 i = 1 << (ffs(desiredvnodes / 10) - 1);
2504 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead),
2505 M_FREEWORK, M_WAITOK);
2506 ump->indir_hash_size = i - 1;
2507 for (i = 0; i <= ump->indir_hash_size; i++)
2508 TAILQ_INIT(&ump->indir_hashtbl[i]);
2509 ACQUIRE_GBLLOCK(&lk);
2510 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next);
2512 if ((fs->fs_flags & FS_SUJ) &&
2513 (error = journal_mount(mp, fs, cred)) != 0) {
2514 printf("Failed to start journal: %d\n", error);
2515 softdep_unmount(mp);
2519 * Start our flushing thread in the bufdaemon process.
2522 ump->softdep_flags |= FLUSH_STARTING;
2524 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc,
2525 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker",
2526 mp->mnt_stat.f_mntonname);
2528 while ((ump->softdep_flags & FLUSH_STARTING) != 0) {
2529 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart",
2534 * When doing soft updates, the counters in the
2535 * superblock may have gotten out of sync. Recomputation
2536 * can take a long time and can be deferred for background
2537 * fsck. However, the old behavior of scanning the cylinder
2538 * groups and recalculating them at mount time is available
2539 * by setting vfs.ffs.compute_summary_at_mount to one.
2541 if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
2543 bzero(&cstotal, sizeof cstotal);
2544 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
2545 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
2546 fs->fs_cgsize, cred, &bp)) != 0) {
2548 softdep_unmount(mp);
2551 cgp = (struct cg *)bp->b_data;
2552 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
2553 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
2554 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
2555 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
2556 fs->fs_cs(fs, cyl) = cgp->cg_cs;
2560 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
2561 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
2563 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
2571 struct ufsmount *ump;
2576 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
2577 ("softdep_unmount called on non-softdep filesystem"));
2580 mp->mnt_flag &= ~MNT_SOFTDEP;
2581 if (MOUNTEDSUJ(mp) == 0) {
2584 mp->mnt_flag &= ~MNT_SUJ;
2586 journal_unmount(ump);
2589 * Shut down our flushing thread. Check for NULL is if
2590 * softdep_mount errors out before the thread has been created.
2592 if (ump->softdep_flushtd != NULL) {
2594 ump->softdep_flags |= FLUSH_EXIT;
2595 wakeup(&ump->softdep_flushtd);
2596 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP,
2598 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0,
2599 ("Thread shutdown failed"));
2602 * Free up our resources.
2604 ACQUIRE_GBLLOCK(&lk);
2605 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next);
2607 rw_destroy(LOCK_PTR(ump));
2608 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size);
2609 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size);
2610 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size);
2611 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP,
2612 ump->bmsafemap_hash_size);
2613 free(ump->indir_hashtbl, M_FREEWORK);
2615 for (i = 0; i <= D_LAST; i++)
2616 KASSERT(ump->softdep_curdeps[i] == 0,
2617 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt,
2618 TYPENAME(i), ump->softdep_curdeps[i]));
2620 free(ump->um_softdep, M_MOUNTDATA);
2623 static struct jblocks *
2624 jblocks_create(void)
2626 struct jblocks *jblocks;
2628 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO);
2629 TAILQ_INIT(&jblocks->jb_segs);
2630 jblocks->jb_avail = 10;
2631 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2632 M_JBLOCKS, M_WAITOK | M_ZERO);
2638 jblocks_alloc(jblocks, bytes, actual)
2639 struct jblocks *jblocks;
2644 struct jextent *jext;
2648 blocks = bytes / DEV_BSIZE;
2649 jext = &jblocks->jb_extent[jblocks->jb_head];
2650 freecnt = jext->je_blocks - jblocks->jb_off;
2652 jblocks->jb_off = 0;
2653 if (++jblocks->jb_head > jblocks->jb_used)
2654 jblocks->jb_head = 0;
2655 jext = &jblocks->jb_extent[jblocks->jb_head];
2656 freecnt = jext->je_blocks;
2658 if (freecnt > blocks)
2660 *actual = freecnt * DEV_BSIZE;
2661 daddr = jext->je_daddr + jblocks->jb_off;
2662 jblocks->jb_off += freecnt;
2663 jblocks->jb_free -= freecnt;
2669 jblocks_free(jblocks, mp, bytes)
2670 struct jblocks *jblocks;
2675 LOCK_OWNED(VFSTOUFS(mp));
2676 jblocks->jb_free += bytes / DEV_BSIZE;
2677 if (jblocks->jb_suspended)
2678 worklist_speedup(mp);
2683 jblocks_destroy(jblocks)
2684 struct jblocks *jblocks;
2687 if (jblocks->jb_extent)
2688 free(jblocks->jb_extent, M_JBLOCKS);
2689 free(jblocks, M_JBLOCKS);
2693 jblocks_add(jblocks, daddr, blocks)
2694 struct jblocks *jblocks;
2698 struct jextent *jext;
2700 jblocks->jb_blocks += blocks;
2701 jblocks->jb_free += blocks;
2702 jext = &jblocks->jb_extent[jblocks->jb_used];
2703 /* Adding the first block. */
2704 if (jext->je_daddr == 0) {
2705 jext->je_daddr = daddr;
2706 jext->je_blocks = blocks;
2709 /* Extending the last extent. */
2710 if (jext->je_daddr + jext->je_blocks == daddr) {
2711 jext->je_blocks += blocks;
2714 /* Adding a new extent. */
2715 if (++jblocks->jb_used == jblocks->jb_avail) {
2716 jblocks->jb_avail *= 2;
2717 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2718 M_JBLOCKS, M_WAITOK | M_ZERO);
2719 memcpy(jext, jblocks->jb_extent,
2720 sizeof(struct jextent) * jblocks->jb_used);
2721 free(jblocks->jb_extent, M_JBLOCKS);
2722 jblocks->jb_extent = jext;
2724 jext = &jblocks->jb_extent[jblocks->jb_used];
2725 jext->je_daddr = daddr;
2726 jext->je_blocks = blocks;
2731 softdep_journal_lookup(mp, vpp)
2735 struct componentname cnp;
2740 error = VFS_VGET(mp, UFS_ROOTINO, LK_EXCLUSIVE, &dvp);
2743 bzero(&cnp, sizeof(cnp));
2744 cnp.cn_nameiop = LOOKUP;
2745 cnp.cn_flags = ISLASTCN;
2746 cnp.cn_thread = curthread;
2747 cnp.cn_cred = curthread->td_ucred;
2748 cnp.cn_pnbuf = SUJ_FILE;
2749 cnp.cn_nameptr = SUJ_FILE;
2750 cnp.cn_namelen = strlen(SUJ_FILE);
2751 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal);
2755 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp);
2760 * Open and verify the journal file.
2763 journal_mount(mp, fs, cred)
2768 struct jblocks *jblocks;
2769 struct ufsmount *ump;
2778 ump->softdep_journal_tail = NULL;
2779 ump->softdep_on_journal = 0;
2780 ump->softdep_accdeps = 0;
2781 ump->softdep_req = 0;
2782 ump->softdep_jblocks = NULL;
2783 error = softdep_journal_lookup(mp, &vp);
2785 printf("Failed to find journal. Use tunefs to create one\n");
2789 if (ip->i_size < SUJ_MIN) {
2793 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */
2794 jblocks = jblocks_create();
2795 for (i = 0; i < bcount; i++) {
2796 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL);
2799 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag));
2802 jblocks_destroy(jblocks);
2805 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */
2806 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */
2807 ump->softdep_jblocks = jblocks;
2811 mp->mnt_flag |= MNT_SUJ;
2812 mp->mnt_flag &= ~MNT_SOFTDEP;
2815 * Only validate the journal contents if the
2816 * filesystem is clean, otherwise we write the logs
2817 * but they'll never be used. If the filesystem was
2818 * still dirty when we mounted it the journal is
2819 * invalid and a new journal can only be valid if it
2820 * starts from a clean mount.
2823 DIP_SET(ip, i_modrev, fs->fs_mtime);
2824 ip->i_flags |= IN_MODIFIED;
2833 journal_unmount(ump)
2834 struct ufsmount *ump;
2837 if (ump->softdep_jblocks)
2838 jblocks_destroy(ump->softdep_jblocks);
2839 ump->softdep_jblocks = NULL;
2843 * Called when a journal record is ready to be written. Space is allocated
2844 * and the journal entry is created when the journal is flushed to stable
2849 struct worklist *wk;
2851 struct ufsmount *ump;
2853 ump = VFSTOUFS(wk->wk_mp);
2855 if (wk->wk_state & ONWORKLIST)
2856 panic("add_to_journal: %s(0x%X) already on list",
2857 TYPENAME(wk->wk_type), wk->wk_state);
2858 wk->wk_state |= ONWORKLIST | DEPCOMPLETE;
2859 if (LIST_EMPTY(&ump->softdep_journal_pending)) {
2860 ump->softdep_jblocks->jb_age = ticks;
2861 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list);
2863 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list);
2864 ump->softdep_journal_tail = wk;
2865 ump->softdep_on_journal += 1;
2869 * Remove an arbitrary item for the journal worklist maintain the tail
2870 * pointer. This happens when a new operation obviates the need to
2871 * journal an old operation.
2874 remove_from_journal(wk)
2875 struct worklist *wk;
2877 struct ufsmount *ump;
2879 ump = VFSTOUFS(wk->wk_mp);
2883 struct worklist *wkn;
2885 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list)
2889 panic("remove_from_journal: %p is not in journal", wk);
2893 * We emulate a TAILQ to save space in most structures which do not
2894 * require TAILQ semantics. Here we must update the tail position
2895 * when removing the tail which is not the final entry. This works
2896 * only if the worklist linkage are at the beginning of the structure.
2898 if (ump->softdep_journal_tail == wk)
2899 ump->softdep_journal_tail =
2900 (struct worklist *)wk->wk_list.le_prev;
2901 WORKLIST_REMOVE(wk);
2902 ump->softdep_on_journal -= 1;
2906 * Check for journal space as well as dependency limits so the prelink
2907 * code can throttle both journaled and non-journaled filesystems.
2908 * Threshold is 0 for low and 1 for min.
2911 journal_space(ump, thresh)
2912 struct ufsmount *ump;
2915 struct jblocks *jblocks;
2918 jblocks = ump->softdep_jblocks;
2919 if (jblocks == NULL)
2922 * We use a tighter restriction here to prevent request_cleanup()
2923 * running in threads from running into locks we currently hold.
2924 * We have to be over the limit and our filesystem has to be
2925 * responsible for more than our share of that usage.
2927 limit = (max_softdeps / 10) * 9;
2928 if (dep_current[D_INODEDEP] > limit &&
2929 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads)
2932 thresh = jblocks->jb_min;
2934 thresh = jblocks->jb_low;
2935 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE;
2936 avail = jblocks->jb_free - avail;
2938 return (avail > thresh);
2942 journal_suspend(ump)
2943 struct ufsmount *ump;
2945 struct jblocks *jblocks;
2949 jblocks = ump->softdep_jblocks;
2951 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
2953 mp->mnt_kern_flag |= MNTK_SUSPEND;
2954 mp->mnt_susp_owner = ump->softdep_flushtd;
2956 jblocks->jb_suspended = 1;
2961 journal_unsuspend(struct ufsmount *ump)
2963 struct jblocks *jblocks;
2967 jblocks = ump->softdep_jblocks;
2969 if (jblocks != NULL && jblocks->jb_suspended &&
2970 journal_space(ump, jblocks->jb_min)) {
2971 jblocks->jb_suspended = 0;
2973 mp->mnt_susp_owner = curthread;
2974 vfs_write_resume(mp, 0);
2982 * Called before any allocation function to be certain that there is
2983 * sufficient space in the journal prior to creating any new records.
2984 * Since in the case of block allocation we may have multiple locked
2985 * buffers at the time of the actual allocation we can not block
2986 * when the journal records are created. Doing so would create a deadlock
2987 * if any of these buffers needed to be flushed to reclaim space. Instead
2988 * we require a sufficiently large amount of available space such that
2989 * each thread in the system could have passed this allocation check and
2990 * still have sufficient free space. With 20% of a minimum journal size
2991 * of 1MB we have 6553 records available.
2994 softdep_prealloc(vp, waitok)
2998 struct ufsmount *ump;
3000 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
3001 ("softdep_prealloc called on non-softdep filesystem"));
3003 * Nothing to do if we are not running journaled soft updates.
3004 * If we currently hold the snapshot lock, we must avoid
3005 * handling other resources that could cause deadlock. Do not
3006 * touch quotas vnode since it is typically recursed with
3007 * other vnode locks held.
3009 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) ||
3010 (vp->v_vflag & VV_SYSTEM) != 0)
3012 ump = VFSTOUFS(vp->v_mount);
3014 if (journal_space(ump, 0)) {
3020 if (waitok == MNT_NOWAIT)
3023 * Attempt to sync this vnode once to flush any journal
3024 * work attached to it.
3026 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0)
3027 ffs_syncvnode(vp, waitok, 0);
3029 process_removes(vp);
3030 process_truncates(vp);
3031 if (journal_space(ump, 0) == 0) {
3032 softdep_speedup(ump);
3033 if (journal_space(ump, 1) == 0)
3034 journal_suspend(ump);
3042 * Before adjusting a link count on a vnode verify that we have sufficient
3043 * journal space. If not, process operations that depend on the currently
3044 * locked pair of vnodes to try to flush space as the syncer, buf daemon,
3045 * and softdep flush threads can not acquire these locks to reclaim space.
3048 softdep_prelink(dvp, vp)
3052 struct ufsmount *ump;
3054 ump = VFSTOUFS(dvp->v_mount);
3057 * Nothing to do if we have sufficient journal space.
3058 * If we currently hold the snapshot lock, we must avoid
3059 * handling other resources that could cause deadlock.
3061 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp))))
3066 ffs_syncvnode(vp, MNT_NOWAIT, 0);
3067 ffs_syncvnode(dvp, MNT_WAIT, 0);
3069 /* Process vp before dvp as it may create .. removes. */
3071 process_removes(vp);
3072 process_truncates(vp);
3074 process_removes(dvp);
3075 process_truncates(dvp);
3076 softdep_speedup(ump);
3077 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT);
3078 if (journal_space(ump, 0) == 0) {
3079 softdep_speedup(ump);
3080 if (journal_space(ump, 1) == 0)
3081 journal_suspend(ump);
3086 jseg_write(ump, jseg, data)
3087 struct ufsmount *ump;
3091 struct jsegrec *rec;
3093 rec = (struct jsegrec *)data;
3094 rec->jsr_seq = jseg->js_seq;
3095 rec->jsr_oldest = jseg->js_oldseq;
3096 rec->jsr_cnt = jseg->js_cnt;
3097 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize;
3099 rec->jsr_time = ump->um_fs->fs_mtime;
3103 inoref_write(inoref, jseg, rec)
3104 struct inoref *inoref;
3106 struct jrefrec *rec;
3109 inoref->if_jsegdep->jd_seg = jseg;
3110 rec->jr_ino = inoref->if_ino;
3111 rec->jr_parent = inoref->if_parent;
3112 rec->jr_nlink = inoref->if_nlink;
3113 rec->jr_mode = inoref->if_mode;
3114 rec->jr_diroff = inoref->if_diroff;
3118 jaddref_write(jaddref, jseg, data)
3119 struct jaddref *jaddref;
3123 struct jrefrec *rec;
3125 rec = (struct jrefrec *)data;
3126 rec->jr_op = JOP_ADDREF;
3127 inoref_write(&jaddref->ja_ref, jseg, rec);
3131 jremref_write(jremref, jseg, data)
3132 struct jremref *jremref;
3136 struct jrefrec *rec;
3138 rec = (struct jrefrec *)data;
3139 rec->jr_op = JOP_REMREF;
3140 inoref_write(&jremref->jr_ref, jseg, rec);
3144 jmvref_write(jmvref, jseg, data)
3145 struct jmvref *jmvref;
3151 rec = (struct jmvrec *)data;
3152 rec->jm_op = JOP_MVREF;
3153 rec->jm_ino = jmvref->jm_ino;
3154 rec->jm_parent = jmvref->jm_parent;
3155 rec->jm_oldoff = jmvref->jm_oldoff;
3156 rec->jm_newoff = jmvref->jm_newoff;
3160 jnewblk_write(jnewblk, jseg, data)
3161 struct jnewblk *jnewblk;
3165 struct jblkrec *rec;
3167 jnewblk->jn_jsegdep->jd_seg = jseg;
3168 rec = (struct jblkrec *)data;
3169 rec->jb_op = JOP_NEWBLK;
3170 rec->jb_ino = jnewblk->jn_ino;
3171 rec->jb_blkno = jnewblk->jn_blkno;
3172 rec->jb_lbn = jnewblk->jn_lbn;
3173 rec->jb_frags = jnewblk->jn_frags;
3174 rec->jb_oldfrags = jnewblk->jn_oldfrags;
3178 jfreeblk_write(jfreeblk, jseg, data)
3179 struct jfreeblk *jfreeblk;
3183 struct jblkrec *rec;
3185 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg;
3186 rec = (struct jblkrec *)data;
3187 rec->jb_op = JOP_FREEBLK;
3188 rec->jb_ino = jfreeblk->jf_ino;
3189 rec->jb_blkno = jfreeblk->jf_blkno;
3190 rec->jb_lbn = jfreeblk->jf_lbn;
3191 rec->jb_frags = jfreeblk->jf_frags;
3192 rec->jb_oldfrags = 0;
3196 jfreefrag_write(jfreefrag, jseg, data)
3197 struct jfreefrag *jfreefrag;
3201 struct jblkrec *rec;
3203 jfreefrag->fr_jsegdep->jd_seg = jseg;
3204 rec = (struct jblkrec *)data;
3205 rec->jb_op = JOP_FREEBLK;
3206 rec->jb_ino = jfreefrag->fr_ino;
3207 rec->jb_blkno = jfreefrag->fr_blkno;
3208 rec->jb_lbn = jfreefrag->fr_lbn;
3209 rec->jb_frags = jfreefrag->fr_frags;
3210 rec->jb_oldfrags = 0;
3214 jtrunc_write(jtrunc, jseg, data)
3215 struct jtrunc *jtrunc;
3219 struct jtrncrec *rec;
3221 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg;
3222 rec = (struct jtrncrec *)data;
3223 rec->jt_op = JOP_TRUNC;
3224 rec->jt_ino = jtrunc->jt_ino;
3225 rec->jt_size = jtrunc->jt_size;
3226 rec->jt_extsize = jtrunc->jt_extsize;
3230 jfsync_write(jfsync, jseg, data)
3231 struct jfsync *jfsync;
3235 struct jtrncrec *rec;
3237 rec = (struct jtrncrec *)data;
3238 rec->jt_op = JOP_SYNC;
3239 rec->jt_ino = jfsync->jfs_ino;
3240 rec->jt_size = jfsync->jfs_size;
3241 rec->jt_extsize = jfsync->jfs_extsize;
3245 softdep_flushjournal(mp)
3248 struct jblocks *jblocks;
3249 struct ufsmount *ump;
3251 if (MOUNTEDSUJ(mp) == 0)
3254 jblocks = ump->softdep_jblocks;
3256 while (ump->softdep_on_journal) {
3257 jblocks->jb_needseg = 1;
3258 softdep_process_journal(mp, NULL, MNT_WAIT);
3263 static void softdep_synchronize_completed(struct bio *);
3264 static void softdep_synchronize(struct bio *, struct ufsmount *, void *);
3267 softdep_synchronize_completed(bp)
3270 struct jseg *oldest;
3272 struct ufsmount *ump;
3275 * caller1 marks the last segment written before we issued the
3276 * synchronize cache.
3278 jseg = bp->bio_caller1;
3283 ump = VFSTOUFS(jseg->js_list.wk_mp);
3287 * Mark all the journal entries waiting on the synchronize cache
3288 * as completed so they may continue on.
3290 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) {
3291 jseg->js_state |= COMPLETE;
3293 jseg = TAILQ_PREV(jseg, jseglst, js_next);
3296 * Restart deferred journal entry processing from the oldest
3300 complete_jsegs(oldest);
3307 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering
3308 * barriers. The journal must be written prior to any blocks that depend
3309 * on it and the journal can not be released until the blocks have be
3310 * written. This code handles both barriers simultaneously.
3313 softdep_synchronize(bp, ump, caller1)
3315 struct ufsmount *ump;
3319 bp->bio_cmd = BIO_FLUSH;
3320 bp->bio_flags |= BIO_ORDERED;
3321 bp->bio_data = NULL;
3322 bp->bio_offset = ump->um_cp->provider->mediasize;
3324 bp->bio_done = softdep_synchronize_completed;
3325 bp->bio_caller1 = caller1;
3327 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private);
3331 * Flush some journal records to disk.
3334 softdep_process_journal(mp, needwk, flags)
3336 struct worklist *needwk;
3339 struct jblocks *jblocks;
3340 struct ufsmount *ump;
3341 struct worklist *wk;
3349 int jrecmin; /* Minimum records per block. */
3350 int jrecmax; /* Maximum records per block. */
3356 if (MOUNTEDSUJ(mp) == 0)
3358 shouldflush = softdep_flushcache;
3364 jblocks = ump->softdep_jblocks;
3365 devbsize = ump->um_devvp->v_bufobj.bo_bsize;
3367 * We write anywhere between a disk block and fs block. The upper
3368 * bound is picked to prevent buffer cache fragmentation and limit
3369 * processing time per I/O.
3371 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */
3372 jrecmax = (fs->fs_bsize / devbsize) * jrecmin;
3375 cnt = ump->softdep_on_journal;
3377 * Criteria for writing a segment:
3378 * 1) We have a full block.
3379 * 2) We're called from jwait() and haven't found the
3381 * 3) Always write if needseg is set.
3382 * 4) If we are called from process_worklist and have
3383 * not yet written anything we write a partial block
3384 * to enforce a 1 second maximum latency on journal
3387 if (cnt < (jrecmax - 1) && needwk == NULL &&
3388 jblocks->jb_needseg == 0 && (segwritten || cnt == 0))
3392 * Verify some free journal space. softdep_prealloc() should
3393 * guarantee that we don't run out so this is indicative of
3394 * a problem with the flow control. Try to recover
3395 * gracefully in any event.
3397 while (jblocks->jb_free == 0) {
3398 if (flags != MNT_WAIT)
3400 printf("softdep: Out of journal space!\n");
3401 softdep_speedup(ump);
3402 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz);
3405 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS);
3406 workitem_alloc(&jseg->js_list, D_JSEG, mp);
3407 LIST_INIT(&jseg->js_entries);
3408 LIST_INIT(&jseg->js_indirs);
3409 jseg->js_state = ATTACHED;
3410 if (shouldflush == 0)
3411 jseg->js_state |= COMPLETE;
3412 else if (bio == NULL)
3413 bio = g_alloc_bio();
3414 jseg->js_jblocks = jblocks;
3415 bp = geteblk(fs->fs_bsize, 0);
3418 * If there was a race while we were allocating the block
3419 * and jseg the entry we care about was likely written.
3420 * We bail out in both the WAIT and NOWAIT case and assume
3421 * the caller will loop if the entry it cares about is
3424 cnt = ump->softdep_on_journal;
3425 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) {
3426 bp->b_flags |= B_INVAL | B_NOCACHE;
3427 WORKITEM_FREE(jseg, D_JSEG);
3434 * Calculate the disk block size required for the available
3435 * records rounded to the min size.
3439 else if (cnt < jrecmax)
3440 size = howmany(cnt, jrecmin) * devbsize;
3442 size = fs->fs_bsize;
3444 * Allocate a disk block for this journal data and account
3445 * for truncation of the requested size if enough contiguous
3446 * space was not available.
3448 bp->b_blkno = jblocks_alloc(jblocks, size, &size);
3449 bp->b_lblkno = bp->b_blkno;
3450 bp->b_offset = bp->b_blkno * DEV_BSIZE;
3451 bp->b_bcount = size;
3452 bp->b_flags &= ~B_INVAL;
3453 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY;
3455 * Initialize our jseg with cnt records. Assign the next
3456 * sequence number to it and link it in-order.
3458 cnt = MIN(cnt, (size / devbsize) * jrecmin);
3461 jseg->js_refs = cnt + 1; /* Self ref. */
3462 jseg->js_size = size;
3463 jseg->js_seq = jblocks->jb_nextseq++;
3464 if (jblocks->jb_oldestseg == NULL)
3465 jblocks->jb_oldestseg = jseg;
3466 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq;
3467 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next);
3468 if (jblocks->jb_writeseg == NULL)
3469 jblocks->jb_writeseg = jseg;
3471 * Start filling in records from the pending list.
3477 * Always put a header on the first block.
3478 * XXX As with below, there might not be a chance to get
3479 * into the loop. Ensure that something valid is written.
3481 jseg_write(ump, jseg, data);
3483 data = bp->b_data + off;
3486 * XXX Something is wrong here. There's no work to do,
3487 * but we need to perform and I/O and allow it to complete
3490 if (LIST_EMPTY(&ump->softdep_journal_pending))
3491 stat_emptyjblocks++;
3493 while ((wk = LIST_FIRST(&ump->softdep_journal_pending))
3497 /* Place a segment header on every device block. */
3498 if ((off % devbsize) == 0) {
3499 jseg_write(ump, jseg, data);
3501 data = bp->b_data + off;
3505 remove_from_journal(wk);
3506 wk->wk_state |= INPROGRESS;
3507 WORKLIST_INSERT(&jseg->js_entries, wk);
3508 switch (wk->wk_type) {
3510 jaddref_write(WK_JADDREF(wk), jseg, data);
3513 jremref_write(WK_JREMREF(wk), jseg, data);
3516 jmvref_write(WK_JMVREF(wk), jseg, data);
3519 jnewblk_write(WK_JNEWBLK(wk), jseg, data);
3522 jfreeblk_write(WK_JFREEBLK(wk), jseg, data);
3525 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data);
3528 jtrunc_write(WK_JTRUNC(wk), jseg, data);
3531 jfsync_write(WK_JFSYNC(wk), jseg, data);
3534 panic("process_journal: Unknown type %s",
3535 TYPENAME(wk->wk_type));
3539 data = bp->b_data + off;
3543 /* Clear any remaining space so we don't leak kernel data */
3545 bzero(data, size - off);
3548 * Write this one buffer and continue.
3551 jblocks->jb_needseg = 0;
3552 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list);
3554 pbgetvp(ump->um_devvp, bp);
3556 * We only do the blocking wait once we find the journal
3557 * entry we're looking for.
3559 if (needwk == NULL && flags == MNT_WAIT)
3566 * If we wrote a segment issue a synchronize cache so the journal
3567 * is reflected on disk before the data is written. Since reclaiming
3568 * journal space also requires writing a journal record this
3569 * process also enforces a barrier before reclamation.
3571 if (segwritten && shouldflush) {
3572 softdep_synchronize(bio, ump,
3573 TAILQ_LAST(&jblocks->jb_segs, jseglst));
3577 * If we've suspended the filesystem because we ran out of journal
3578 * space either try to sync it here to make some progress or
3579 * unsuspend it if we already have.
3581 if (flags == 0 && jblocks->jb_suspended) {
3582 if (journal_unsuspend(ump))
3585 VFS_SYNC(mp, MNT_NOWAIT);
3586 ffs_sbupdate(ump, MNT_WAIT, 0);
3592 * Complete a jseg, allowing all dependencies awaiting journal writes
3593 * to proceed. Each journal dependency also attaches a jsegdep to dependent
3594 * structures so that the journal segment can be freed to reclaim space.
3600 struct worklist *wk;
3601 struct jmvref *jmvref;
3606 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
3607 WORKLIST_REMOVE(wk);
3608 wk->wk_state &= ~INPROGRESS;
3609 wk->wk_state |= COMPLETE;
3610 KASSERT(i++ < jseg->js_cnt,
3611 ("handle_written_jseg: overflow %d >= %d",
3612 i - 1, jseg->js_cnt));
3613 switch (wk->wk_type) {
3615 handle_written_jaddref(WK_JADDREF(wk));
3618 handle_written_jremref(WK_JREMREF(wk));
3621 rele_jseg(jseg); /* No jsegdep. */
3622 jmvref = WK_JMVREF(wk);
3623 LIST_REMOVE(jmvref, jm_deps);
3624 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0)
3625 free_pagedep(jmvref->jm_pagedep);
3626 WORKITEM_FREE(jmvref, D_JMVREF);
3629 handle_written_jnewblk(WK_JNEWBLK(wk));
3632 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep);
3635 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep);
3638 rele_jseg(jseg); /* No jsegdep. */
3639 WORKITEM_FREE(wk, D_JFSYNC);
3642 handle_written_jfreefrag(WK_JFREEFRAG(wk));
3645 panic("handle_written_jseg: Unknown type %s",
3646 TYPENAME(wk->wk_type));
3650 /* Release the self reference so the structure may be freed. */
3655 * Determine which jsegs are ready for completion processing. Waits for
3656 * synchronize cache to complete as well as forcing in-order completion
3657 * of journal entries.
3660 complete_jsegs(jseg)
3663 struct jblocks *jblocks;
3666 jblocks = jseg->js_jblocks;
3668 * Don't allow out of order completions. If this isn't the first
3669 * block wait for it to write before we're done.
3671 if (jseg != jblocks->jb_writeseg)
3673 /* Iterate through available jsegs processing their entries. */
3674 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) {
3675 jblocks->jb_oldestwrseq = jseg->js_oldseq;
3676 jsegn = TAILQ_NEXT(jseg, js_next);
3677 complete_jseg(jseg);
3680 jblocks->jb_writeseg = jseg;
3682 * Attempt to free jsegs now that oldestwrseq may have advanced.
3684 free_jsegs(jblocks);
3688 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle
3689 * the final completions.
3692 handle_written_jseg(jseg, bp)
3697 if (jseg->js_refs == 0)
3698 panic("handle_written_jseg: No self-reference on %p", jseg);
3699 jseg->js_state |= DEPCOMPLETE;
3701 * We'll never need this buffer again, set flags so it will be
3704 bp->b_flags |= B_INVAL | B_NOCACHE;
3706 complete_jsegs(jseg);
3709 static inline struct jsegdep *
3711 struct inoref *inoref;
3713 struct jsegdep *jsegdep;
3715 jsegdep = inoref->if_jsegdep;
3716 inoref->if_jsegdep = NULL;
3722 * Called once a jremref has made it to stable store. The jremref is marked
3723 * complete and we attempt to free it. Any pagedeps writes sleeping waiting
3724 * for the jremref to complete will be awoken by free_jremref.
3727 handle_written_jremref(jremref)
3728 struct jremref *jremref;
3730 struct inodedep *inodedep;
3731 struct jsegdep *jsegdep;
3732 struct dirrem *dirrem;
3734 /* Grab the jsegdep. */
3735 jsegdep = inoref_jseg(&jremref->jr_ref);
3737 * Remove us from the inoref list.
3739 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
3741 panic("handle_written_jremref: Lost inodedep");
3742 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
3744 * Complete the dirrem.
3746 dirrem = jremref->jr_dirrem;
3747 jremref->jr_dirrem = NULL;
3748 LIST_REMOVE(jremref, jr_deps);
3749 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
3750 jwork_insert(&dirrem->dm_jwork, jsegdep);
3751 if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
3752 (dirrem->dm_state & COMPLETE) != 0)
3753 add_to_worklist(&dirrem->dm_list, 0);
3754 free_jremref(jremref);
3758 * Called once a jaddref has made it to stable store. The dependency is
3759 * marked complete and any dependent structures are added to the inode
3760 * bufwait list to be completed as soon as it is written. If a bitmap write
3761 * depends on this entry we move the inode into the inodedephd of the
3762 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
3765 handle_written_jaddref(jaddref)
3766 struct jaddref *jaddref;
3768 struct jsegdep *jsegdep;
3769 struct inodedep *inodedep;
3770 struct diradd *diradd;
3771 struct mkdir *mkdir;
3773 /* Grab the jsegdep. */
3774 jsegdep = inoref_jseg(&jaddref->ja_ref);
3777 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3779 panic("handle_written_jaddref: Lost inodedep.");
3780 if (jaddref->ja_diradd == NULL)
3781 panic("handle_written_jaddref: No dependency");
3782 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
3783 diradd = jaddref->ja_diradd;
3784 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
3785 } else if (jaddref->ja_state & MKDIR_PARENT) {
3786 mkdir = jaddref->ja_mkdir;
3787 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
3788 } else if (jaddref->ja_state & MKDIR_BODY)
3789 mkdir = jaddref->ja_mkdir;
3791 panic("handle_written_jaddref: Unknown dependency %p",
3792 jaddref->ja_diradd);
3793 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */
3795 * Remove us from the inode list.
3797 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
3799 * The mkdir may be waiting on the jaddref to clear before freeing.
3802 KASSERT(mkdir->md_list.wk_type == D_MKDIR,
3803 ("handle_written_jaddref: Incorrect type for mkdir %s",
3804 TYPENAME(mkdir->md_list.wk_type)));
3805 mkdir->md_jaddref = NULL;
3806 diradd = mkdir->md_diradd;
3807 mkdir->md_state |= DEPCOMPLETE;
3808 complete_mkdir(mkdir);
3810 jwork_insert(&diradd->da_jwork, jsegdep);
3811 if (jaddref->ja_state & NEWBLOCK) {
3812 inodedep->id_state |= ONDEPLIST;
3813 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
3816 free_jaddref(jaddref);
3820 * Called once a jnewblk journal is written. The allocdirect or allocindir
3821 * is placed in the bmsafemap to await notification of a written bitmap. If
3822 * the operation was canceled we add the segdep to the appropriate
3823 * dependency to free the journal space once the canceling operation
3827 handle_written_jnewblk(jnewblk)
3828 struct jnewblk *jnewblk;
3830 struct bmsafemap *bmsafemap;
3831 struct freefrag *freefrag;
3832 struct freework *freework;
3833 struct jsegdep *jsegdep;
3834 struct newblk *newblk;
3836 /* Grab the jsegdep. */
3837 jsegdep = jnewblk->jn_jsegdep;
3838 jnewblk->jn_jsegdep = NULL;
3839 if (jnewblk->jn_dep == NULL)
3840 panic("handle_written_jnewblk: No dependency for the segdep.");
3841 switch (jnewblk->jn_dep->wk_type) {
3846 * Add the written block to the bmsafemap so it can
3847 * be notified when the bitmap is on disk.
3849 newblk = WK_NEWBLK(jnewblk->jn_dep);
3850 newblk->nb_jnewblk = NULL;
3851 if ((newblk->nb_state & GOINGAWAY) == 0) {
3852 bmsafemap = newblk->nb_bmsafemap;
3853 newblk->nb_state |= ONDEPLIST;
3854 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk,
3857 jwork_insert(&newblk->nb_jwork, jsegdep);
3861 * A newblock being removed by a freefrag when replaced by
3864 freefrag = WK_FREEFRAG(jnewblk->jn_dep);
3865 freefrag->ff_jdep = NULL;
3866 jwork_insert(&freefrag->ff_jwork, jsegdep);
3870 * A direct block was removed by truncate.
3872 freework = WK_FREEWORK(jnewblk->jn_dep);
3873 freework->fw_jnewblk = NULL;
3874 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep);
3877 panic("handle_written_jnewblk: Unknown type %d.",
3878 jnewblk->jn_dep->wk_type);
3880 jnewblk->jn_dep = NULL;
3881 free_jnewblk(jnewblk);
3885 * Cancel a jfreefrag that won't be needed, probably due to colliding with
3886 * an in-flight allocation that has not yet been committed. Divorce us
3887 * from the freefrag and mark it DEPCOMPLETE so that it may be added
3891 cancel_jfreefrag(jfreefrag)
3892 struct jfreefrag *jfreefrag;
3894 struct freefrag *freefrag;
3896 if (jfreefrag->fr_jsegdep) {
3897 free_jsegdep(jfreefrag->fr_jsegdep);
3898 jfreefrag->fr_jsegdep = NULL;
3900 freefrag = jfreefrag->fr_freefrag;
3901 jfreefrag->fr_freefrag = NULL;
3902 free_jfreefrag(jfreefrag);
3903 freefrag->ff_state |= DEPCOMPLETE;
3904 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno);
3908 * Free a jfreefrag when the parent freefrag is rendered obsolete.
3911 free_jfreefrag(jfreefrag)
3912 struct jfreefrag *jfreefrag;
3915 if (jfreefrag->fr_state & INPROGRESS)
3916 WORKLIST_REMOVE(&jfreefrag->fr_list);
3917 else if (jfreefrag->fr_state & ONWORKLIST)
3918 remove_from_journal(&jfreefrag->fr_list);
3919 if (jfreefrag->fr_freefrag != NULL)
3920 panic("free_jfreefrag: Still attached to a freefrag.");
3921 WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
3925 * Called when the journal write for a jfreefrag completes. The parent
3926 * freefrag is added to the worklist if this completes its dependencies.
3929 handle_written_jfreefrag(jfreefrag)
3930 struct jfreefrag *jfreefrag;
3932 struct jsegdep *jsegdep;
3933 struct freefrag *freefrag;
3935 /* Grab the jsegdep. */
3936 jsegdep = jfreefrag->fr_jsegdep;
3937 jfreefrag->fr_jsegdep = NULL;
3938 freefrag = jfreefrag->fr_freefrag;
3939 if (freefrag == NULL)
3940 panic("handle_written_jfreefrag: No freefrag.");
3941 freefrag->ff_state |= DEPCOMPLETE;
3942 freefrag->ff_jdep = NULL;
3943 jwork_insert(&freefrag->ff_jwork, jsegdep);
3944 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
3945 add_to_worklist(&freefrag->ff_list, 0);
3946 jfreefrag->fr_freefrag = NULL;
3947 free_jfreefrag(jfreefrag);
3951 * Called when the journal write for a jfreeblk completes. The jfreeblk
3952 * is removed from the freeblks list of pending journal writes and the
3953 * jsegdep is moved to the freeblks jwork to be completed when all blocks
3954 * have been reclaimed.
3957 handle_written_jblkdep(jblkdep)
3958 struct jblkdep *jblkdep;
3960 struct freeblks *freeblks;
3961 struct jsegdep *jsegdep;
3963 /* Grab the jsegdep. */
3964 jsegdep = jblkdep->jb_jsegdep;
3965 jblkdep->jb_jsegdep = NULL;
3966 freeblks = jblkdep->jb_freeblks;
3967 LIST_REMOVE(jblkdep, jb_deps);
3968 jwork_insert(&freeblks->fb_jwork, jsegdep);
3970 * If the freeblks is all journaled, we can add it to the worklist.
3972 if (LIST_EMPTY(&freeblks->fb_jblkdephd) &&
3973 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
3974 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
3976 free_jblkdep(jblkdep);
3979 static struct jsegdep *
3980 newjsegdep(struct worklist *wk)
3982 struct jsegdep *jsegdep;
3984 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
3985 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
3986 jsegdep->jd_seg = NULL;
3991 static struct jmvref *
3992 newjmvref(dp, ino, oldoff, newoff)
3998 struct jmvref *jmvref;
4000 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
4001 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp));
4002 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
4003 jmvref->jm_parent = dp->i_number;
4004 jmvref->jm_ino = ino;
4005 jmvref->jm_oldoff = oldoff;
4006 jmvref->jm_newoff = newoff;
4012 * Allocate a new jremref that tracks the removal of ip from dp with the
4013 * directory entry offset of diroff. Mark the entry as ATTACHED and
4014 * DEPCOMPLETE as we have all the information required for the journal write
4015 * and the directory has already been removed from the buffer. The caller
4016 * is responsible for linking the jremref into the pagedep and adding it
4017 * to the journal to write. The MKDIR_PARENT flag is set if we're doing
4018 * a DOTDOT addition so handle_workitem_remove() can properly assign
4019 * the jsegdep when we're done.
4021 static struct jremref *
4022 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
4023 off_t diroff, nlink_t nlink)
4025 struct jremref *jremref;
4027 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
4028 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp));
4029 jremref->jr_state = ATTACHED;
4030 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
4032 jremref->jr_dirrem = dirrem;
4038 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
4039 nlink_t nlink, uint16_t mode)
4042 inoref->if_jsegdep = newjsegdep(&inoref->if_list);
4043 inoref->if_diroff = diroff;
4044 inoref->if_ino = ino;
4045 inoref->if_parent = parent;
4046 inoref->if_nlink = nlink;
4047 inoref->if_mode = mode;
4051 * Allocate a new jaddref to track the addition of ino to dp at diroff. The
4052 * directory offset may not be known until later. The caller is responsible
4053 * adding the entry to the journal when this information is available. nlink
4054 * should be the link count prior to the addition and mode is only required
4055 * to have the correct FMT.
4057 static struct jaddref *
4058 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
4061 struct jaddref *jaddref;
4063 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
4064 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp));
4065 jaddref->ja_state = ATTACHED;
4066 jaddref->ja_mkdir = NULL;
4067 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
4073 * Create a new free dependency for a freework. The caller is responsible
4074 * for adjusting the reference count when it has the lock held. The freedep
4075 * will track an outstanding bitmap write that will ultimately clear the
4076 * freework to continue.
4078 static struct freedep *
4079 newfreedep(struct freework *freework)
4081 struct freedep *freedep;
4083 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
4084 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
4085 freedep->fd_freework = freework;
4091 * Free a freedep structure once the buffer it is linked to is written. If
4092 * this is the last reference to the freework schedule it for completion.
4095 free_freedep(freedep)
4096 struct freedep *freedep;
4098 struct freework *freework;
4100 freework = freedep->fd_freework;
4101 freework->fw_freeblks->fb_cgwait--;
4102 if (--freework->fw_ref == 0)
4103 freework_enqueue(freework);
4104 WORKITEM_FREE(freedep, D_FREEDEP);
4108 * Allocate a new freework structure that may be a level in an indirect
4109 * when parent is not NULL or a top level block when it is. The top level
4110 * freework structures are allocated without the per-filesystem lock held
4111 * and before the freeblks is visible outside of softdep_setup_freeblocks().
4113 static struct freework *
4114 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal)
4115 struct ufsmount *ump;
4116 struct freeblks *freeblks;
4117 struct freework *parent;
4124 struct freework *freework;
4126 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
4127 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
4128 freework->fw_state = ATTACHED;
4129 freework->fw_jnewblk = NULL;
4130 freework->fw_freeblks = freeblks;
4131 freework->fw_parent = parent;
4132 freework->fw_lbn = lbn;
4133 freework->fw_blkno = nb;
4134 freework->fw_frags = frags;
4135 freework->fw_indir = NULL;
4136 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 ||
4137 lbn >= -UFS_NXADDR) ? 0 : NINDIR(ump->um_fs) + 1;
4138 freework->fw_start = freework->fw_off = off;
4140 newjfreeblk(freeblks, lbn, nb, frags);
4141 if (parent == NULL) {
4143 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
4152 * Eliminate a jfreeblk for a block that does not need journaling.
4155 cancel_jfreeblk(freeblks, blkno)
4156 struct freeblks *freeblks;
4159 struct jfreeblk *jfreeblk;
4160 struct jblkdep *jblkdep;
4162 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) {
4163 if (jblkdep->jb_list.wk_type != D_JFREEBLK)
4165 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list);
4166 if (jfreeblk->jf_blkno == blkno)
4169 if (jblkdep == NULL)
4171 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno);
4172 free_jsegdep(jblkdep->jb_jsegdep);
4173 LIST_REMOVE(jblkdep, jb_deps);
4174 WORKITEM_FREE(jfreeblk, D_JFREEBLK);
4178 * Allocate a new jfreeblk to journal top level block pointer when truncating
4179 * a file. The caller must add this to the worklist when the per-filesystem
4182 static struct jfreeblk *
4183 newjfreeblk(freeblks, lbn, blkno, frags)
4184 struct freeblks *freeblks;
4189 struct jfreeblk *jfreeblk;
4191 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
4192 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK,
4193 freeblks->fb_list.wk_mp);
4194 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list);
4195 jfreeblk->jf_dep.jb_freeblks = freeblks;
4196 jfreeblk->jf_ino = freeblks->fb_inum;
4197 jfreeblk->jf_lbn = lbn;
4198 jfreeblk->jf_blkno = blkno;
4199 jfreeblk->jf_frags = frags;
4200 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps);
4206 * The journal is only prepared to handle full-size block numbers, so we
4207 * have to adjust the record to reflect the change to a full-size block.
4208 * For example, suppose we have a block made up of fragments 8-15 and
4209 * want to free its last two fragments. We are given a request that says:
4210 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0
4211 * where frags are the number of fragments to free and oldfrags are the
4212 * number of fragments to keep. To block align it, we have to change it to
4213 * have a valid full-size blkno, so it becomes:
4214 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6
4217 adjust_newfreework(freeblks, frag_offset)
4218 struct freeblks *freeblks;
4221 struct jfreeblk *jfreeblk;
4223 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL &&
4224 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK),
4225 ("adjust_newfreework: Missing freeblks dependency"));
4227 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd));
4228 jfreeblk->jf_blkno -= frag_offset;
4229 jfreeblk->jf_frags += frag_offset;
4233 * Allocate a new jtrunc to track a partial truncation.
4235 static struct jtrunc *
4236 newjtrunc(freeblks, size, extsize)
4237 struct freeblks *freeblks;
4241 struct jtrunc *jtrunc;
4243 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
4244 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC,
4245 freeblks->fb_list.wk_mp);
4246 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list);
4247 jtrunc->jt_dep.jb_freeblks = freeblks;
4248 jtrunc->jt_ino = freeblks->fb_inum;
4249 jtrunc->jt_size = size;
4250 jtrunc->jt_extsize = extsize;
4251 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps);
4257 * If we're canceling a new bitmap we have to search for another ref
4258 * to move into the bmsafemap dep. This might be better expressed
4259 * with another structure.
4262 move_newblock_dep(jaddref, inodedep)
4263 struct jaddref *jaddref;
4264 struct inodedep *inodedep;
4266 struct inoref *inoref;
4267 struct jaddref *jaddrefn;
4270 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4271 inoref = TAILQ_NEXT(inoref, if_deps)) {
4272 if ((jaddref->ja_state & NEWBLOCK) &&
4273 inoref->if_list.wk_type == D_JADDREF) {
4274 jaddrefn = (struct jaddref *)inoref;
4278 if (jaddrefn == NULL)
4280 jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
4281 jaddrefn->ja_state |= jaddref->ja_state &
4282 (ATTACHED | UNDONE | NEWBLOCK);
4283 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
4284 jaddref->ja_state |= ATTACHED;
4285 LIST_REMOVE(jaddref, ja_bmdeps);
4286 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
4291 * Cancel a jaddref either before it has been written or while it is being
4292 * written. This happens when a link is removed before the add reaches
4293 * the disk. The jaddref dependency is kept linked into the bmsafemap
4294 * and inode to prevent the link count or bitmap from reaching the disk
4295 * until handle_workitem_remove() re-adjusts the counts and bitmaps as
4298 * Returns 1 if the canceled addref requires journaling of the remove and
4302 cancel_jaddref(jaddref, inodedep, wkhd)
4303 struct jaddref *jaddref;
4304 struct inodedep *inodedep;
4305 struct workhead *wkhd;
4307 struct inoref *inoref;
4308 struct jsegdep *jsegdep;
4311 KASSERT((jaddref->ja_state & COMPLETE) == 0,
4312 ("cancel_jaddref: Canceling complete jaddref"));
4313 if (jaddref->ja_state & (INPROGRESS | COMPLETE))
4317 if (inodedep == NULL)
4318 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
4320 panic("cancel_jaddref: Lost inodedep");
4322 * We must adjust the nlink of any reference operation that follows
4323 * us so that it is consistent with the in-memory reference. This
4324 * ensures that inode nlink rollbacks always have the correct link.
4327 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4328 inoref = TAILQ_NEXT(inoref, if_deps)) {
4329 if (inoref->if_state & GOINGAWAY)
4334 jsegdep = inoref_jseg(&jaddref->ja_ref);
4335 if (jaddref->ja_state & NEWBLOCK)
4336 move_newblock_dep(jaddref, inodedep);
4337 wake_worklist(&jaddref->ja_list);
4338 jaddref->ja_mkdir = NULL;
4339 if (jaddref->ja_state & INPROGRESS) {
4340 jaddref->ja_state &= ~INPROGRESS;
4341 WORKLIST_REMOVE(&jaddref->ja_list);
4342 jwork_insert(wkhd, jsegdep);
4344 free_jsegdep(jsegdep);
4345 if (jaddref->ja_state & DEPCOMPLETE)
4346 remove_from_journal(&jaddref->ja_list);
4348 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE);
4350 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
4351 * can arrange for them to be freed with the bitmap. Otherwise we
4352 * no longer need this addref attached to the inoreflst and it
4353 * will incorrectly adjust nlink if we leave it.
4355 if ((jaddref->ja_state & NEWBLOCK) == 0) {
4356 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
4358 jaddref->ja_state |= COMPLETE;
4359 free_jaddref(jaddref);
4363 * Leave the head of the list for jsegdeps for fast merging.
4365 if (LIST_FIRST(wkhd) != NULL) {
4366 jaddref->ja_state |= ONWORKLIST;
4367 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
4369 WORKLIST_INSERT(wkhd, &jaddref->ja_list);
4375 * Attempt to free a jaddref structure when some work completes. This
4376 * should only succeed once the entry is written and all dependencies have
4380 free_jaddref(jaddref)
4381 struct jaddref *jaddref;
4384 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
4386 if (jaddref->ja_ref.if_jsegdep)
4387 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
4388 jaddref, jaddref->ja_state);
4389 if (jaddref->ja_state & NEWBLOCK)
4390 LIST_REMOVE(jaddref, ja_bmdeps);
4391 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST))
4392 panic("free_jaddref: Bad state %p(0x%X)",
4393 jaddref, jaddref->ja_state);
4394 if (jaddref->ja_mkdir != NULL)
4395 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
4396 WORKITEM_FREE(jaddref, D_JADDREF);
4400 * Free a jremref structure once it has been written or discarded.
4403 free_jremref(jremref)
4404 struct jremref *jremref;
4407 if (jremref->jr_ref.if_jsegdep)
4408 free_jsegdep(jremref->jr_ref.if_jsegdep);
4409 if (jremref->jr_state & INPROGRESS)
4410 panic("free_jremref: IO still pending");
4411 WORKITEM_FREE(jremref, D_JREMREF);
4415 * Free a jnewblk structure.
4418 free_jnewblk(jnewblk)
4419 struct jnewblk *jnewblk;
4422 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
4424 LIST_REMOVE(jnewblk, jn_deps);
4425 if (jnewblk->jn_dep != NULL)
4426 panic("free_jnewblk: Dependency still attached.");
4427 WORKITEM_FREE(jnewblk, D_JNEWBLK);
4431 * Cancel a jnewblk which has been been made redundant by frag extension.
4434 cancel_jnewblk(jnewblk, wkhd)
4435 struct jnewblk *jnewblk;
4436 struct workhead *wkhd;
4438 struct jsegdep *jsegdep;
4440 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno);
4441 jsegdep = jnewblk->jn_jsegdep;
4442 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL)
4443 panic("cancel_jnewblk: Invalid state");
4444 jnewblk->jn_jsegdep = NULL;
4445 jnewblk->jn_dep = NULL;
4446 jnewblk->jn_state |= GOINGAWAY;
4447 if (jnewblk->jn_state & INPROGRESS) {
4448 jnewblk->jn_state &= ~INPROGRESS;
4449 WORKLIST_REMOVE(&jnewblk->jn_list);
4450 jwork_insert(wkhd, jsegdep);
4452 free_jsegdep(jsegdep);
4453 remove_from_journal(&jnewblk->jn_list);
4455 wake_worklist(&jnewblk->jn_list);
4456 WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
4460 free_jblkdep(jblkdep)
4461 struct jblkdep *jblkdep;
4464 if (jblkdep->jb_list.wk_type == D_JFREEBLK)
4465 WORKITEM_FREE(jblkdep, D_JFREEBLK);
4466 else if (jblkdep->jb_list.wk_type == D_JTRUNC)
4467 WORKITEM_FREE(jblkdep, D_JTRUNC);
4469 panic("free_jblkdep: Unexpected type %s",
4470 TYPENAME(jblkdep->jb_list.wk_type));
4474 * Free a single jseg once it is no longer referenced in memory or on
4475 * disk. Reclaim journal blocks and dependencies waiting for the segment
4479 free_jseg(jseg, jblocks)
4481 struct jblocks *jblocks;
4483 struct freework *freework;
4486 * Free freework structures that were lingering to indicate freed
4487 * indirect blocks that forced journal write ordering on reallocate.
4489 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL)
4490 indirblk_remove(freework);
4491 if (jblocks->jb_oldestseg == jseg)
4492 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next);
4493 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
4494 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
4495 KASSERT(LIST_EMPTY(&jseg->js_entries),
4496 ("free_jseg: Freed jseg has valid entries."));
4497 WORKITEM_FREE(jseg, D_JSEG);
4501 * Free all jsegs that meet the criteria for being reclaimed and update
4506 struct jblocks *jblocks;
4511 * Free only those jsegs which have none allocated before them to
4512 * preserve the journal space ordering.
4514 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
4516 * Only reclaim space when nothing depends on this journal
4517 * set and another set has written that it is no longer
4520 if (jseg->js_refs != 0) {
4521 jblocks->jb_oldestseg = jseg;
4524 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE)
4526 if (jseg->js_seq > jblocks->jb_oldestwrseq)
4529 * We can free jsegs that didn't write entries when
4530 * oldestwrseq == js_seq.
4532 if (jseg->js_seq == jblocks->jb_oldestwrseq &&
4535 free_jseg(jseg, jblocks);
4538 * If we exited the loop above we still must discover the
4539 * oldest valid segment.
4542 for (jseg = jblocks->jb_oldestseg; jseg != NULL;
4543 jseg = TAILQ_NEXT(jseg, js_next))
4544 if (jseg->js_refs != 0)
4546 jblocks->jb_oldestseg = jseg;
4548 * The journal has no valid records but some jsegs may still be
4549 * waiting on oldestwrseq to advance. We force a small record
4550 * out to permit these lingering records to be reclaimed.
4552 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs))
4553 jblocks->jb_needseg = 1;
4557 * Release one reference to a jseg and free it if the count reaches 0. This
4558 * should eventually reclaim journal space as well.
4565 KASSERT(jseg->js_refs > 0,
4566 ("free_jseg: Invalid refcnt %d", jseg->js_refs));
4567 if (--jseg->js_refs != 0)
4569 free_jsegs(jseg->js_jblocks);
4573 * Release a jsegdep and decrement the jseg count.
4576 free_jsegdep(jsegdep)
4577 struct jsegdep *jsegdep;
4580 if (jsegdep->jd_seg)
4581 rele_jseg(jsegdep->jd_seg);
4582 WORKITEM_FREE(jsegdep, D_JSEGDEP);
4586 * Wait for a journal item to make it to disk. Initiate journal processing
4591 struct worklist *wk;
4595 LOCK_OWNED(VFSTOUFS(wk->wk_mp));
4597 * Blocking journal waits cause slow synchronous behavior. Record
4598 * stats on the frequency of these blocking operations.
4600 if (waitfor == MNT_WAIT) {
4601 stat_journal_wait++;
4602 switch (wk->wk_type) {
4605 stat_jwait_filepage++;
4609 stat_jwait_freeblks++;
4612 stat_jwait_newblk++;
4622 * If IO has not started we process the journal. We can't mark the
4623 * worklist item as IOWAITING because we drop the lock while
4624 * processing the journal and the worklist entry may be freed after
4625 * this point. The caller may call back in and re-issue the request.
4627 if ((wk->wk_state & INPROGRESS) == 0) {
4628 softdep_process_journal(wk->wk_mp, wk, waitfor);
4629 if (waitfor != MNT_WAIT)
4633 if (waitfor != MNT_WAIT)
4635 wait_worklist(wk, "jwait");
4640 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
4641 * appropriate. This is a convenience function to reduce duplicate code
4642 * for the setup and revert functions below.
4644 static struct inodedep *
4645 inodedep_lookup_ip(ip)
4648 struct inodedep *inodedep;
4650 KASSERT(ip->i_nlink >= ip->i_effnlink,
4651 ("inodedep_lookup_ip: bad delta"));
4652 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC,
4654 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
4655 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
4661 * Called prior to creating a new inode and linking it to a directory. The
4662 * jaddref structure must already be allocated by softdep_setup_inomapdep
4663 * and it is discovered here so we can initialize the mode and update
4667 softdep_setup_create(dp, ip)
4671 struct inodedep *inodedep;
4672 struct jaddref *jaddref;
4675 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4676 ("softdep_setup_create called on non-softdep filesystem"));
4677 KASSERT(ip->i_nlink == 1,
4678 ("softdep_setup_create: Invalid link count."));
4680 ACQUIRE_LOCK(ITOUMP(dp));
4681 inodedep = inodedep_lookup_ip(ip);
4682 if (DOINGSUJ(dvp)) {
4683 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4685 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
4686 ("softdep_setup_create: No addref structure present."));
4688 softdep_prelink(dvp, NULL);
4689 FREE_LOCK(ITOUMP(dp));
4693 * Create a jaddref structure to track the addition of a DOTDOT link when
4694 * we are reparenting an inode as part of a rename. This jaddref will be
4695 * found by softdep_setup_directory_change. Adjusts nlinkdelta for
4696 * non-journaling softdep.
4699 softdep_setup_dotdot_link(dp, ip)
4703 struct inodedep *inodedep;
4704 struct jaddref *jaddref;
4707 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4708 ("softdep_setup_dotdot_link called on non-softdep filesystem"));
4712 * We don't set MKDIR_PARENT as this is not tied to a mkdir and
4713 * is used as a normal link would be.
4716 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4717 dp->i_effnlink - 1, dp->i_mode);
4718 ACQUIRE_LOCK(ITOUMP(dp));
4719 inodedep = inodedep_lookup_ip(dp);
4721 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4723 softdep_prelink(dvp, ITOV(ip));
4724 FREE_LOCK(ITOUMP(dp));
4728 * Create a jaddref structure to track a new link to an inode. The directory
4729 * offset is not known until softdep_setup_directory_add or
4730 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling
4734 softdep_setup_link(dp, ip)
4738 struct inodedep *inodedep;
4739 struct jaddref *jaddref;
4742 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4743 ("softdep_setup_link called on non-softdep filesystem"));
4747 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
4749 ACQUIRE_LOCK(ITOUMP(dp));
4750 inodedep = inodedep_lookup_ip(ip);
4752 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4754 softdep_prelink(dvp, ITOV(ip));
4755 FREE_LOCK(ITOUMP(dp));
4759 * Called to create the jaddref structures to track . and .. references as
4760 * well as lookup and further initialize the incomplete jaddref created
4761 * by softdep_setup_inomapdep when the inode was allocated. Adjusts
4762 * nlinkdelta for non-journaling softdep.
4765 softdep_setup_mkdir(dp, ip)
4769 struct inodedep *inodedep;
4770 struct jaddref *dotdotaddref;
4771 struct jaddref *dotaddref;
4772 struct jaddref *jaddref;
4775 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4776 ("softdep_setup_mkdir called on non-softdep filesystem"));
4778 dotaddref = dotdotaddref = NULL;
4779 if (DOINGSUJ(dvp)) {
4780 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
4782 dotaddref->ja_state |= MKDIR_BODY;
4783 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4784 dp->i_effnlink - 1, dp->i_mode);
4785 dotdotaddref->ja_state |= MKDIR_PARENT;
4787 ACQUIRE_LOCK(ITOUMP(dp));
4788 inodedep = inodedep_lookup_ip(ip);
4789 if (DOINGSUJ(dvp)) {
4790 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4792 KASSERT(jaddref != NULL,
4793 ("softdep_setup_mkdir: No addref structure present."));
4794 KASSERT(jaddref->ja_parent == dp->i_number,
4795 ("softdep_setup_mkdir: bad parent %ju",
4796 (uintmax_t)jaddref->ja_parent));
4797 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
4800 inodedep = inodedep_lookup_ip(dp);
4802 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
4803 &dotdotaddref->ja_ref, if_deps);
4804 softdep_prelink(ITOV(dp), NULL);
4805 FREE_LOCK(ITOUMP(dp));
4809 * Called to track nlinkdelta of the inode and parent directories prior to
4810 * unlinking a directory.
4813 softdep_setup_rmdir(dp, ip)
4819 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4820 ("softdep_setup_rmdir called on non-softdep filesystem"));
4822 ACQUIRE_LOCK(ITOUMP(dp));
4823 (void) inodedep_lookup_ip(ip);
4824 (void) inodedep_lookup_ip(dp);
4825 softdep_prelink(dvp, ITOV(ip));
4826 FREE_LOCK(ITOUMP(dp));
4830 * Called to track nlinkdelta of the inode and parent directories prior to
4834 softdep_setup_unlink(dp, ip)
4840 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4841 ("softdep_setup_unlink called on non-softdep filesystem"));
4843 ACQUIRE_LOCK(ITOUMP(dp));
4844 (void) inodedep_lookup_ip(ip);
4845 (void) inodedep_lookup_ip(dp);
4846 softdep_prelink(dvp, ITOV(ip));
4847 FREE_LOCK(ITOUMP(dp));
4851 * Called to release the journal structures created by a failed non-directory
4852 * creation. Adjusts nlinkdelta for non-journaling softdep.
4855 softdep_revert_create(dp, ip)
4859 struct inodedep *inodedep;
4860 struct jaddref *jaddref;
4863 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0,
4864 ("softdep_revert_create called on non-softdep filesystem"));
4866 ACQUIRE_LOCK(ITOUMP(dp));
4867 inodedep = inodedep_lookup_ip(ip);
4868 if (DOINGSUJ(dvp)) {
4869 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4871 KASSERT(jaddref->ja_parent == dp->i_number,
4872 ("softdep_revert_create: addref parent mismatch"));
4873 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4875 FREE_LOCK(ITOUMP(dp));
4879 * Called to release the journal structures created by a failed link
4880 * addition. Adjusts nlinkdelta for non-journaling softdep.
4883 softdep_revert_link(dp, ip)
4887 struct inodedep *inodedep;
4888 struct jaddref *jaddref;
4891 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4892 ("softdep_revert_link called on non-softdep filesystem"));
4894 ACQUIRE_LOCK(ITOUMP(dp));
4895 inodedep = inodedep_lookup_ip(ip);
4896 if (DOINGSUJ(dvp)) {
4897 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4899 KASSERT(jaddref->ja_parent == dp->i_number,
4900 ("softdep_revert_link: addref parent mismatch"));
4901 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4903 FREE_LOCK(ITOUMP(dp));
4907 * Called to release the journal structures created by a failed mkdir
4908 * attempt. Adjusts nlinkdelta for non-journaling softdep.
4911 softdep_revert_mkdir(dp, ip)
4915 struct inodedep *inodedep;
4916 struct jaddref *jaddref;
4917 struct jaddref *dotaddref;
4920 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4921 ("softdep_revert_mkdir called on non-softdep filesystem"));
4924 ACQUIRE_LOCK(ITOUMP(dp));
4925 inodedep = inodedep_lookup_ip(dp);
4926 if (DOINGSUJ(dvp)) {
4927 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4929 KASSERT(jaddref->ja_parent == ip->i_number,
4930 ("softdep_revert_mkdir: dotdot addref parent mismatch"));
4931 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4933 inodedep = inodedep_lookup_ip(ip);
4934 if (DOINGSUJ(dvp)) {
4935 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4937 KASSERT(jaddref->ja_parent == dp->i_number,
4938 ("softdep_revert_mkdir: addref parent mismatch"));
4939 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
4940 inoreflst, if_deps);
4941 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4942 KASSERT(dotaddref->ja_parent == ip->i_number,
4943 ("softdep_revert_mkdir: dot addref parent mismatch"));
4944 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait);
4946 FREE_LOCK(ITOUMP(dp));
4950 * Called to correct nlinkdelta after a failed rmdir.
4953 softdep_revert_rmdir(dp, ip)
4958 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0,
4959 ("softdep_revert_rmdir called on non-softdep filesystem"));
4960 ACQUIRE_LOCK(ITOUMP(dp));
4961 (void) inodedep_lookup_ip(ip);
4962 (void) inodedep_lookup_ip(dp);
4963 FREE_LOCK(ITOUMP(dp));
4967 * Protecting the freemaps (or bitmaps).
4969 * To eliminate the need to execute fsck before mounting a filesystem
4970 * after a power failure, one must (conservatively) guarantee that the
4971 * on-disk copy of the bitmaps never indicate that a live inode or block is
4972 * free. So, when a block or inode is allocated, the bitmap should be
4973 * updated (on disk) before any new pointers. When a block or inode is
4974 * freed, the bitmap should not be updated until all pointers have been
4975 * reset. The latter dependency is handled by the delayed de-allocation
4976 * approach described below for block and inode de-allocation. The former
4977 * dependency is handled by calling the following procedure when a block or
4978 * inode is allocated. When an inode is allocated an "inodedep" is created
4979 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
4980 * Each "inodedep" is also inserted into the hash indexing structure so
4981 * that any additional link additions can be made dependent on the inode
4984 * The ufs filesystem maintains a number of free block counts (e.g., per
4985 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
4986 * in addition to the bitmaps. These counts are used to improve efficiency
4987 * during allocation and therefore must be consistent with the bitmaps.
4988 * There is no convenient way to guarantee post-crash consistency of these
4989 * counts with simple update ordering, for two main reasons: (1) The counts
4990 * and bitmaps for a single cylinder group block are not in the same disk
4991 * sector. If a disk write is interrupted (e.g., by power failure), one may
4992 * be written and the other not. (2) Some of the counts are located in the
4993 * superblock rather than the cylinder group block. So, we focus our soft
4994 * updates implementation on protecting the bitmaps. When mounting a
4995 * filesystem, we recompute the auxiliary counts from the bitmaps.
4999 * Called just after updating the cylinder group block to allocate an inode.
5002 softdep_setup_inomapdep(bp, ip, newinum, mode)
5003 struct buf *bp; /* buffer for cylgroup block with inode map */
5004 struct inode *ip; /* inode related to allocation */
5005 ino_t newinum; /* new inode number being allocated */
5008 struct inodedep *inodedep;
5009 struct bmsafemap *bmsafemap;
5010 struct jaddref *jaddref;
5015 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5016 ("softdep_setup_inomapdep called on non-softdep filesystem"));
5017 fs = VFSTOUFS(mp)->um_fs;
5021 * Allocate the journal reference add structure so that the bitmap
5022 * can be dependent on it.
5024 if (MOUNTEDSUJ(mp)) {
5025 jaddref = newjaddref(ip, newinum, 0, 0, mode);
5026 jaddref->ja_state |= NEWBLOCK;
5030 * Create a dependency for the newly allocated inode.
5031 * Panic if it already exists as something is seriously wrong.
5032 * Otherwise add it to the dependency list for the buffer holding
5033 * the cylinder group map from which it was allocated.
5035 * We have to preallocate a bmsafemap entry in case it is needed
5036 * in bmsafemap_lookup since once we allocate the inodedep, we
5037 * have to finish initializing it before we can FREE_LOCK().
5038 * By preallocating, we avoid FREE_LOCK() while doing a malloc
5039 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before
5040 * creating the inodedep as it can be freed during the time
5041 * that we FREE_LOCK() while allocating the inodedep. We must
5042 * call workitem_alloc() before entering the locked section as
5043 * it also acquires the lock and we must avoid trying doing so
5046 bmsafemap = malloc(sizeof(struct bmsafemap),
5047 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5048 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5049 ACQUIRE_LOCK(ITOUMP(ip));
5050 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep)))
5051 panic("softdep_setup_inomapdep: dependency %p for new"
5052 "inode already exists", inodedep);
5053 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap);
5055 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
5056 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
5059 inodedep->id_state |= ONDEPLIST;
5060 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
5062 inodedep->id_bmsafemap = bmsafemap;
5063 inodedep->id_state &= ~DEPCOMPLETE;
5064 FREE_LOCK(ITOUMP(ip));
5068 * Called just after updating the cylinder group block to
5069 * allocate block or fragment.
5072 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
5073 struct buf *bp; /* buffer for cylgroup block with block map */
5074 struct mount *mp; /* filesystem doing allocation */
5075 ufs2_daddr_t newblkno; /* number of newly allocated block */
5076 int frags; /* Number of fragments. */
5077 int oldfrags; /* Previous number of fragments for extend. */
5079 struct newblk *newblk;
5080 struct bmsafemap *bmsafemap;
5081 struct jnewblk *jnewblk;
5082 struct ufsmount *ump;
5085 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5086 ("softdep_setup_blkmapdep called on non-softdep filesystem"));
5091 * Create a dependency for the newly allocated block.
5092 * Add it to the dependency list for the buffer holding
5093 * the cylinder group map from which it was allocated.
5095 if (MOUNTEDSUJ(mp)) {
5096 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
5097 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
5098 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
5099 jnewblk->jn_state = ATTACHED;
5100 jnewblk->jn_blkno = newblkno;
5101 jnewblk->jn_frags = frags;
5102 jnewblk->jn_oldfrags = oldfrags;
5110 cgp = (struct cg *)bp->b_data;
5111 blksfree = cg_blksfree(cgp);
5112 bno = dtogd(fs, jnewblk->jn_blkno);
5113 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
5115 if (isset(blksfree, bno + i))
5116 panic("softdep_setup_blkmapdep: "
5117 "free fragment %d from %d-%d "
5118 "state 0x%X dep %p", i,
5119 jnewblk->jn_oldfrags,
5129 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d",
5130 newblkno, frags, oldfrags);
5132 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
5133 panic("softdep_setup_blkmapdep: found block");
5134 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
5135 dtog(fs, newblkno), NULL);
5137 jnewblk->jn_dep = (struct worklist *)newblk;
5138 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
5140 newblk->nb_state |= ONDEPLIST;
5141 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
5143 newblk->nb_bmsafemap = bmsafemap;
5144 newblk->nb_jnewblk = jnewblk;
5148 #define BMSAFEMAP_HASH(ump, cg) \
5149 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size])
5152 bmsafemap_find(bmsafemaphd, cg, bmsafemapp)
5153 struct bmsafemap_hashhead *bmsafemaphd;
5155 struct bmsafemap **bmsafemapp;
5157 struct bmsafemap *bmsafemap;
5159 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
5160 if (bmsafemap->sm_cg == cg)
5163 *bmsafemapp = bmsafemap;
5172 * Find the bmsafemap associated with a cylinder group buffer.
5173 * If none exists, create one. The buffer must be locked when
5174 * this routine is called and this routine must be called with
5175 * the softdep lock held. To avoid giving up the lock while
5176 * allocating a new bmsafemap, a preallocated bmsafemap may be
5177 * provided. If it is provided but not needed, it is freed.
5179 static struct bmsafemap *
5180 bmsafemap_lookup(mp, bp, cg, newbmsafemap)
5184 struct bmsafemap *newbmsafemap;
5186 struct bmsafemap_hashhead *bmsafemaphd;
5187 struct bmsafemap *bmsafemap, *collision;
5188 struct worklist *wk;
5189 struct ufsmount *ump;
5193 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer"));
5194 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5195 if (wk->wk_type == D_BMSAFEMAP) {
5197 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5198 return (WK_BMSAFEMAP(wk));
5201 bmsafemaphd = BMSAFEMAP_HASH(ump, cg);
5202 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) {
5204 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5208 bmsafemap = newbmsafemap;
5211 bmsafemap = malloc(sizeof(struct bmsafemap),
5212 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5213 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5216 bmsafemap->sm_buf = bp;
5217 LIST_INIT(&bmsafemap->sm_inodedephd);
5218 LIST_INIT(&bmsafemap->sm_inodedepwr);
5219 LIST_INIT(&bmsafemap->sm_newblkhd);
5220 LIST_INIT(&bmsafemap->sm_newblkwr);
5221 LIST_INIT(&bmsafemap->sm_jaddrefhd);
5222 LIST_INIT(&bmsafemap->sm_jnewblkhd);
5223 LIST_INIT(&bmsafemap->sm_freehd);
5224 LIST_INIT(&bmsafemap->sm_freewr);
5225 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) {
5226 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
5229 bmsafemap->sm_cg = cg;
5230 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
5231 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
5232 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
5237 * Direct block allocation dependencies.
5239 * When a new block is allocated, the corresponding disk locations must be
5240 * initialized (with zeros or new data) before the on-disk inode points to
5241 * them. Also, the freemap from which the block was allocated must be
5242 * updated (on disk) before the inode's pointer. These two dependencies are
5243 * independent of each other and are needed for all file blocks and indirect
5244 * blocks that are pointed to directly by the inode. Just before the
5245 * "in-core" version of the inode is updated with a newly allocated block
5246 * number, a procedure (below) is called to setup allocation dependency
5247 * structures. These structures are removed when the corresponding
5248 * dependencies are satisfied or when the block allocation becomes obsolete
5249 * (i.e., the file is deleted, the block is de-allocated, or the block is a
5250 * fragment that gets upgraded). All of these cases are handled in
5251 * procedures described later.
5253 * When a file extension causes a fragment to be upgraded, either to a larger
5254 * fragment or to a full block, the on-disk location may change (if the
5255 * previous fragment could not simply be extended). In this case, the old
5256 * fragment must be de-allocated, but not until after the inode's pointer has
5257 * been updated. In most cases, this is handled by later procedures, which
5258 * will construct a "freefrag" structure to be added to the workitem queue
5259 * when the inode update is complete (or obsolete). The main exception to
5260 * this is when an allocation occurs while a pending allocation dependency
5261 * (for the same block pointer) remains. This case is handled in the main
5262 * allocation dependency setup procedure by immediately freeing the
5263 * unreferenced fragments.
5266 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5267 struct inode *ip; /* inode to which block is being added */
5268 ufs_lbn_t off; /* block pointer within inode */
5269 ufs2_daddr_t newblkno; /* disk block number being added */
5270 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */
5271 long newsize; /* size of new block */
5272 long oldsize; /* size of new block */
5273 struct buf *bp; /* bp for allocated block */
5275 struct allocdirect *adp, *oldadp;
5276 struct allocdirectlst *adphead;
5277 struct freefrag *freefrag;
5278 struct inodedep *inodedep;
5279 struct pagedep *pagedep;
5280 struct jnewblk *jnewblk;
5281 struct newblk *newblk;
5287 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5288 ("softdep_setup_allocdirect called on non-softdep filesystem"));
5289 if (oldblkno && oldblkno != newblkno)
5291 * The usual case is that a smaller fragment that
5292 * was just allocated has been replaced with a bigger
5293 * fragment or a full-size block. If it is marked as
5294 * B_DELWRI, the current contents have not been written
5295 * to disk. It is possible that the block was written
5296 * earlier, but very uncommon. If the block has never
5297 * been written, there is no need to send a BIO_DELETE
5298 * for it when it is freed. The gain from avoiding the
5299 * TRIMs for the common case of unwritten blocks far
5300 * exceeds the cost of the write amplification for the
5301 * uncommon case of failing to send a TRIM for a block
5302 * that had been written.
5304 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn,
5305 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY);
5310 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd "
5311 "off %jd newsize %ld oldsize %d",
5312 ip->i_number, newblkno, oldblkno, off, newsize, oldsize);
5313 ACQUIRE_LOCK(ITOUMP(ip));
5314 if (off >= UFS_NDADDR) {
5316 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
5318 /* allocating an indirect block */
5320 panic("softdep_setup_allocdirect: non-zero indir");
5323 panic("softdep_setup_allocdirect: lbn %jd != off %jd",
5326 * Allocating a direct block.
5328 * If we are allocating a directory block, then we must
5329 * allocate an associated pagedep to track additions and
5332 if ((ip->i_mode & IFMT) == IFDIR)
5333 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC,
5336 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5337 panic("softdep_setup_allocdirect: lost block");
5338 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5339 ("softdep_setup_allocdirect: newblk already initialized"));
5341 * Convert the newblk to an allocdirect.
5343 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5344 adp = (struct allocdirect *)newblk;
5345 newblk->nb_freefrag = freefrag;
5346 adp->ad_offset = off;
5347 adp->ad_oldblkno = oldblkno;
5348 adp->ad_newsize = newsize;
5349 adp->ad_oldsize = oldsize;
5352 * Finish initializing the journal.
5354 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5355 jnewblk->jn_ino = ip->i_number;
5356 jnewblk->jn_lbn = lbn;
5357 add_to_journal(&jnewblk->jn_list);
5359 if (freefrag && freefrag->ff_jdep != NULL &&
5360 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5361 add_to_journal(freefrag->ff_jdep);
5362 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5363 adp->ad_inodedep = inodedep;
5365 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5367 * The list of allocdirects must be kept in sorted and ascending
5368 * order so that the rollback routines can quickly determine the
5369 * first uncommitted block (the size of the file stored on disk
5370 * ends at the end of the lowest committed fragment, or if there
5371 * are no fragments, at the end of the highest committed block).
5372 * Since files generally grow, the typical case is that the new
5373 * block is to be added at the end of the list. We speed this
5374 * special case by checking against the last allocdirect in the
5375 * list before laboriously traversing the list looking for the
5378 adphead = &inodedep->id_newinoupdt;
5379 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5380 if (oldadp == NULL || oldadp->ad_offset <= off) {
5381 /* insert at end of list */
5382 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5383 if (oldadp != NULL && oldadp->ad_offset == off)
5384 allocdirect_merge(adphead, adp, oldadp);
5385 FREE_LOCK(ITOUMP(ip));
5388 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5389 if (oldadp->ad_offset >= off)
5393 panic("softdep_setup_allocdirect: lost entry");
5394 /* insert in middle of list */
5395 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5396 if (oldadp->ad_offset == off)
5397 allocdirect_merge(adphead, adp, oldadp);
5399 FREE_LOCK(ITOUMP(ip));
5403 * Merge a newer and older journal record to be stored either in a
5404 * newblock or freefrag. This handles aggregating journal records for
5405 * fragment allocation into a second record as well as replacing a
5406 * journal free with an aborted journal allocation. A segment for the
5407 * oldest record will be placed on wkhd if it has been written. If not
5408 * the segment for the newer record will suffice.
5410 static struct worklist *
5411 jnewblk_merge(new, old, wkhd)
5412 struct worklist *new;
5413 struct worklist *old;
5414 struct workhead *wkhd;
5416 struct jnewblk *njnewblk;
5417 struct jnewblk *jnewblk;
5419 /* Handle NULLs to simplify callers. */
5424 /* Replace a jfreefrag with a jnewblk. */
5425 if (new->wk_type == D_JFREEFRAG) {
5426 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno)
5427 panic("jnewblk_merge: blkno mismatch: %p, %p",
5429 cancel_jfreefrag(WK_JFREEFRAG(new));
5432 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK)
5433 panic("jnewblk_merge: Bad type: old %d new %d\n",
5434 old->wk_type, new->wk_type);
5436 * Handle merging of two jnewblk records that describe
5437 * different sets of fragments in the same block.
5439 jnewblk = WK_JNEWBLK(old);
5440 njnewblk = WK_JNEWBLK(new);
5441 if (jnewblk->jn_blkno != njnewblk->jn_blkno)
5442 panic("jnewblk_merge: Merging disparate blocks.");
5444 * The record may be rolled back in the cg.
5446 if (jnewblk->jn_state & UNDONE) {
5447 jnewblk->jn_state &= ~UNDONE;
5448 njnewblk->jn_state |= UNDONE;
5449 njnewblk->jn_state &= ~ATTACHED;
5452 * We modify the newer addref and free the older so that if neither
5453 * has been written the most up-to-date copy will be on disk. If
5454 * both have been written but rolled back we only temporarily need
5455 * one of them to fix the bits when the cg write completes.
5457 jnewblk->jn_state |= ATTACHED | COMPLETE;
5458 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
5459 cancel_jnewblk(jnewblk, wkhd);
5460 WORKLIST_REMOVE(&jnewblk->jn_list);
5461 free_jnewblk(jnewblk);
5466 * Replace an old allocdirect dependency with a newer one.
5469 allocdirect_merge(adphead, newadp, oldadp)
5470 struct allocdirectlst *adphead; /* head of list holding allocdirects */
5471 struct allocdirect *newadp; /* allocdirect being added */
5472 struct allocdirect *oldadp; /* existing allocdirect being checked */
5474 struct worklist *wk;
5475 struct freefrag *freefrag;
5478 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp));
5479 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
5480 newadp->ad_oldsize != oldadp->ad_newsize ||
5481 newadp->ad_offset >= UFS_NDADDR)
5482 panic("%s %jd != new %jd || old size %ld != new %ld",
5483 "allocdirect_merge: old blkno",
5484 (intmax_t)newadp->ad_oldblkno,
5485 (intmax_t)oldadp->ad_newblkno,
5486 newadp->ad_oldsize, oldadp->ad_newsize);
5487 newadp->ad_oldblkno = oldadp->ad_oldblkno;
5488 newadp->ad_oldsize = oldadp->ad_oldsize;
5490 * If the old dependency had a fragment to free or had never
5491 * previously had a block allocated, then the new dependency
5492 * can immediately post its freefrag and adopt the old freefrag.
5493 * This action is done by swapping the freefrag dependencies.
5494 * The new dependency gains the old one's freefrag, and the
5495 * old one gets the new one and then immediately puts it on
5496 * the worklist when it is freed by free_newblk. It is
5497 * not possible to do this swap when the old dependency had a
5498 * non-zero size but no previous fragment to free. This condition
5499 * arises when the new block is an extension of the old block.
5500 * Here, the first part of the fragment allocated to the new
5501 * dependency is part of the block currently claimed on disk by
5502 * the old dependency, so cannot legitimately be freed until the
5503 * conditions for the new dependency are fulfilled.
5505 freefrag = newadp->ad_freefrag;
5506 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
5507 newadp->ad_freefrag = oldadp->ad_freefrag;
5508 oldadp->ad_freefrag = freefrag;
5511 * If we are tracking a new directory-block allocation,
5512 * move it from the old allocdirect to the new allocdirect.
5514 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
5515 WORKLIST_REMOVE(wk);
5516 if (!LIST_EMPTY(&oldadp->ad_newdirblk))
5517 panic("allocdirect_merge: extra newdirblk");
5518 WORKLIST_INSERT(&newadp->ad_newdirblk, wk);
5520 TAILQ_REMOVE(adphead, oldadp, ad_next);
5522 * We need to move any journal dependencies over to the freefrag
5523 * that releases this block if it exists. Otherwise we are
5524 * extending an existing block and we'll wait until that is
5525 * complete to release the journal space and extend the
5526 * new journal to cover this old space as well.
5528 if (freefrag == NULL) {
5529 if (oldadp->ad_newblkno != newadp->ad_newblkno)
5530 panic("allocdirect_merge: %jd != %jd",
5531 oldadp->ad_newblkno, newadp->ad_newblkno);
5532 newadp->ad_block.nb_jnewblk = (struct jnewblk *)
5533 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list,
5534 &oldadp->ad_block.nb_jnewblk->jn_list,
5535 &newadp->ad_block.nb_jwork);
5536 oldadp->ad_block.nb_jnewblk = NULL;
5537 cancel_newblk(&oldadp->ad_block, NULL,
5538 &newadp->ad_block.nb_jwork);
5540 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block,
5541 &freefrag->ff_list, &freefrag->ff_jwork);
5542 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk,
5543 &freefrag->ff_jwork);
5545 free_newblk(&oldadp->ad_block);
5549 * Allocate a jfreefrag structure to journal a single block free.
5551 static struct jfreefrag *
5552 newjfreefrag(freefrag, ip, blkno, size, lbn)
5553 struct freefrag *freefrag;
5559 struct jfreefrag *jfreefrag;
5563 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
5565 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip));
5566 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
5567 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
5568 jfreefrag->fr_ino = ip->i_number;
5569 jfreefrag->fr_lbn = lbn;
5570 jfreefrag->fr_blkno = blkno;
5571 jfreefrag->fr_frags = numfrags(fs, size);
5572 jfreefrag->fr_freefrag = freefrag;
5578 * Allocate a new freefrag structure.
5580 static struct freefrag *
5581 newfreefrag(ip, blkno, size, lbn, key)
5588 struct freefrag *freefrag;
5589 struct ufsmount *ump;
5592 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd",
5593 ip->i_number, blkno, size, lbn);
5596 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
5597 panic("newfreefrag: frag size");
5598 freefrag = malloc(sizeof(struct freefrag),
5599 M_FREEFRAG, M_SOFTDEP_FLAGS);
5600 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump));
5601 freefrag->ff_state = ATTACHED;
5602 LIST_INIT(&freefrag->ff_jwork);
5603 freefrag->ff_inum = ip->i_number;
5604 freefrag->ff_vtype = ITOV(ip)->v_type;
5605 freefrag->ff_blkno = blkno;
5606 freefrag->ff_fragsize = size;
5607 freefrag->ff_key = key;
5609 if (MOUNTEDSUJ(UFSTOVFS(ump))) {
5610 freefrag->ff_jdep = (struct worklist *)
5611 newjfreefrag(freefrag, ip, blkno, size, lbn);
5613 freefrag->ff_state |= DEPCOMPLETE;
5614 freefrag->ff_jdep = NULL;
5621 * This workitem de-allocates fragments that were replaced during
5622 * file block allocation.
5625 handle_workitem_freefrag(freefrag)
5626 struct freefrag *freefrag;
5628 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
5629 struct workhead wkhd;
5632 "handle_workitem_freefrag: ino %d blkno %jd size %ld",
5633 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize);
5635 * It would be illegal to add new completion items to the
5636 * freefrag after it was schedule to be done so it must be
5637 * safe to modify the list head here.
5641 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
5643 * If the journal has not been written we must cancel it here.
5645 if (freefrag->ff_jdep) {
5646 if (freefrag->ff_jdep->wk_type != D_JNEWBLK)
5647 panic("handle_workitem_freefrag: Unexpected type %d\n",
5648 freefrag->ff_jdep->wk_type);
5649 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd);
5652 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
5653 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype,
5654 &wkhd, freefrag->ff_key);
5656 WORKITEM_FREE(freefrag, D_FREEFRAG);
5661 * Set up a dependency structure for an external attributes data block.
5662 * This routine follows much of the structure of softdep_setup_allocdirect.
5663 * See the description of softdep_setup_allocdirect above for details.
5666 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5669 ufs2_daddr_t newblkno;
5670 ufs2_daddr_t oldblkno;
5675 struct allocdirect *adp, *oldadp;
5676 struct allocdirectlst *adphead;
5677 struct freefrag *freefrag;
5678 struct inodedep *inodedep;
5679 struct jnewblk *jnewblk;
5680 struct newblk *newblk;
5682 struct ufsmount *ump;
5687 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5688 ("softdep_setup_allocext called on non-softdep filesystem"));
5689 KASSERT(off < UFS_NXADDR,
5690 ("softdep_setup_allocext: lbn %lld > UFS_NXADDR", (long long)off));
5693 if (oldblkno && oldblkno != newblkno)
5695 * The usual case is that a smaller fragment that
5696 * was just allocated has been replaced with a bigger
5697 * fragment or a full-size block. If it is marked as
5698 * B_DELWRI, the current contents have not been written
5699 * to disk. It is possible that the block was written
5700 * earlier, but very uncommon. If the block has never
5701 * been written, there is no need to send a BIO_DELETE
5702 * for it when it is freed. The gain from avoiding the
5703 * TRIMs for the common case of unwritten blocks far
5704 * exceeds the cost of the write amplification for the
5705 * uncommon case of failing to send a TRIM for a block
5706 * that had been written.
5708 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn,
5709 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY);
5714 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5715 panic("softdep_setup_allocext: lost block");
5716 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5717 ("softdep_setup_allocext: newblk already initialized"));
5719 * Convert the newblk to an allocdirect.
5721 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5722 adp = (struct allocdirect *)newblk;
5723 newblk->nb_freefrag = freefrag;
5724 adp->ad_offset = off;
5725 adp->ad_oldblkno = oldblkno;
5726 adp->ad_newsize = newsize;
5727 adp->ad_oldsize = oldsize;
5728 adp->ad_state |= EXTDATA;
5731 * Finish initializing the journal.
5733 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5734 jnewblk->jn_ino = ip->i_number;
5735 jnewblk->jn_lbn = lbn;
5736 add_to_journal(&jnewblk->jn_list);
5738 if (freefrag && freefrag->ff_jdep != NULL &&
5739 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5740 add_to_journal(freefrag->ff_jdep);
5741 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5742 adp->ad_inodedep = inodedep;
5744 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5746 * The list of allocdirects must be kept in sorted and ascending
5747 * order so that the rollback routines can quickly determine the
5748 * first uncommitted block (the size of the file stored on disk
5749 * ends at the end of the lowest committed fragment, or if there
5750 * are no fragments, at the end of the highest committed block).
5751 * Since files generally grow, the typical case is that the new
5752 * block is to be added at the end of the list. We speed this
5753 * special case by checking against the last allocdirect in the
5754 * list before laboriously traversing the list looking for the
5757 adphead = &inodedep->id_newextupdt;
5758 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5759 if (oldadp == NULL || oldadp->ad_offset <= off) {
5760 /* insert at end of list */
5761 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5762 if (oldadp != NULL && oldadp->ad_offset == off)
5763 allocdirect_merge(adphead, adp, oldadp);
5767 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5768 if (oldadp->ad_offset >= off)
5772 panic("softdep_setup_allocext: lost entry");
5773 /* insert in middle of list */
5774 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5775 if (oldadp->ad_offset == off)
5776 allocdirect_merge(adphead, adp, oldadp);
5781 * Indirect block allocation dependencies.
5783 * The same dependencies that exist for a direct block also exist when
5784 * a new block is allocated and pointed to by an entry in a block of
5785 * indirect pointers. The undo/redo states described above are also
5786 * used here. Because an indirect block contains many pointers that
5787 * may have dependencies, a second copy of the entire in-memory indirect
5788 * block is kept. The buffer cache copy is always completely up-to-date.
5789 * The second copy, which is used only as a source for disk writes,
5790 * contains only the safe pointers (i.e., those that have no remaining
5791 * update dependencies). The second copy is freed when all pointers
5792 * are safe. The cache is not allowed to replace indirect blocks with
5793 * pending update dependencies. If a buffer containing an indirect
5794 * block with dependencies is written, these routines will mark it
5795 * dirty again. It can only be successfully written once all the
5796 * dependencies are removed. The ffs_fsync routine in conjunction with
5797 * softdep_sync_metadata work together to get all the dependencies
5798 * removed so that a file can be successfully written to disk. Three
5799 * procedures are used when setting up indirect block pointer
5800 * dependencies. The division is necessary because of the organization
5801 * of the "balloc" routine and because of the distinction between file
5802 * pages and file metadata blocks.
5806 * Allocate a new allocindir structure.
5808 static struct allocindir *
5809 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
5810 struct inode *ip; /* inode for file being extended */
5811 int ptrno; /* offset of pointer in indirect block */
5812 ufs2_daddr_t newblkno; /* disk block number being added */
5813 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5816 struct newblk *newblk;
5817 struct allocindir *aip;
5818 struct freefrag *freefrag;
5819 struct jnewblk *jnewblk;
5822 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn,
5826 ACQUIRE_LOCK(ITOUMP(ip));
5827 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0)
5828 panic("new_allocindir: lost block");
5829 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5830 ("newallocindir: newblk already initialized"));
5831 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR);
5832 newblk->nb_freefrag = freefrag;
5833 aip = (struct allocindir *)newblk;
5834 aip->ai_offset = ptrno;
5835 aip->ai_oldblkno = oldblkno;
5837 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5838 jnewblk->jn_ino = ip->i_number;
5839 jnewblk->jn_lbn = lbn;
5840 add_to_journal(&jnewblk->jn_list);
5842 if (freefrag && freefrag->ff_jdep != NULL &&
5843 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5844 add_to_journal(freefrag->ff_jdep);
5849 * Called just before setting an indirect block pointer
5850 * to a newly allocated file page.
5853 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
5854 struct inode *ip; /* inode for file being extended */
5855 ufs_lbn_t lbn; /* allocated block number within file */
5856 struct buf *bp; /* buffer with indirect blk referencing page */
5857 int ptrno; /* offset of pointer in indirect block */
5858 ufs2_daddr_t newblkno; /* disk block number being added */
5859 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5860 struct buf *nbp; /* buffer holding allocated page */
5862 struct inodedep *inodedep;
5863 struct freefrag *freefrag;
5864 struct allocindir *aip;
5865 struct pagedep *pagedep;
5867 struct ufsmount *ump;
5871 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5872 ("softdep_setup_allocindir_page called on non-softdep filesystem"));
5873 KASSERT(lbn == nbp->b_lblkno,
5874 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
5875 lbn, bp->b_lblkno));
5877 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd "
5878 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn);
5879 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
5880 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
5881 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5883 * If we are allocating a directory page, then we must
5884 * allocate an associated pagedep to track additions and
5887 if ((ip->i_mode & IFMT) == IFDIR)
5888 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep);
5889 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5890 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
5893 handle_workitem_freefrag(freefrag);
5897 * Called just before setting an indirect block pointer to a
5898 * newly allocated indirect block.
5901 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
5902 struct buf *nbp; /* newly allocated indirect block */
5903 struct inode *ip; /* inode for file being extended */
5904 struct buf *bp; /* indirect block referencing allocated block */
5905 int ptrno; /* offset of pointer in indirect block */
5906 ufs2_daddr_t newblkno; /* disk block number being added */
5908 struct inodedep *inodedep;
5909 struct allocindir *aip;
5910 struct ufsmount *ump;
5914 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
5915 ("softdep_setup_allocindir_meta called on non-softdep filesystem"));
5917 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d",
5918 ip->i_number, newblkno, ptrno);
5919 lbn = nbp->b_lblkno;
5920 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
5921 aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
5922 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep);
5923 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5924 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
5925 panic("softdep_setup_allocindir_meta: Block already existed");
5930 indirdep_complete(indirdep)
5931 struct indirdep *indirdep;
5933 struct allocindir *aip;
5935 LIST_REMOVE(indirdep, ir_next);
5936 indirdep->ir_state |= DEPCOMPLETE;
5938 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
5939 LIST_REMOVE(aip, ai_next);
5940 free_newblk(&aip->ai_block);
5943 * If this indirdep is not attached to a buf it was simply waiting
5944 * on completion to clear completehd. free_indirdep() asserts
5945 * that nothing is dangling.
5947 if ((indirdep->ir_state & ONWORKLIST) == 0)
5948 free_indirdep(indirdep);
5951 static struct indirdep *
5952 indirdep_lookup(mp, ip, bp)
5957 struct indirdep *indirdep, *newindirdep;
5958 struct newblk *newblk;
5959 struct ufsmount *ump;
5960 struct worklist *wk;
5970 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5971 if (wk->wk_type != D_INDIRDEP)
5973 indirdep = WK_INDIRDEP(wk);
5976 /* Found on the buffer worklist, no new structure to free. */
5977 if (indirdep != NULL && newindirdep == NULL)
5979 if (indirdep != NULL && newindirdep != NULL)
5980 panic("indirdep_lookup: simultaneous create");
5981 /* None found on the buffer and a new structure is ready. */
5982 if (indirdep == NULL && newindirdep != NULL)
5984 /* None found and no new structure available. */
5986 newindirdep = malloc(sizeof(struct indirdep),
5987 M_INDIRDEP, M_SOFTDEP_FLAGS);
5988 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
5989 newindirdep->ir_state = ATTACHED;
5991 newindirdep->ir_state |= UFS1FMT;
5992 TAILQ_INIT(&newindirdep->ir_trunc);
5993 newindirdep->ir_saveddata = NULL;
5994 LIST_INIT(&newindirdep->ir_deplisthd);
5995 LIST_INIT(&newindirdep->ir_donehd);
5996 LIST_INIT(&newindirdep->ir_writehd);
5997 LIST_INIT(&newindirdep->ir_completehd);
5998 if (bp->b_blkno == bp->b_lblkno) {
5999 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
6001 bp->b_blkno = blkno;
6003 newindirdep->ir_freeblks = NULL;
6004 newindirdep->ir_savebp =
6005 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
6006 newindirdep->ir_bp = bp;
6007 BUF_KERNPROC(newindirdep->ir_savebp);
6008 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
6011 indirdep = newindirdep;
6012 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
6014 * If the block is not yet allocated we don't set DEPCOMPLETE so
6015 * that we don't free dependencies until the pointers are valid.
6016 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather
6017 * than using the hash.
6019 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk))
6020 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next);
6022 indirdep->ir_state |= DEPCOMPLETE;
6027 * Called to finish the allocation of the "aip" allocated
6028 * by one of the two routines above.
6030 static struct freefrag *
6031 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
6032 struct buf *bp; /* in-memory copy of the indirect block */
6033 struct inode *ip; /* inode for file being extended */
6034 struct inodedep *inodedep; /* Inodedep for ip */
6035 struct allocindir *aip; /* allocindir allocated by the above routines */
6036 ufs_lbn_t lbn; /* Logical block number for this block. */
6039 struct indirdep *indirdep;
6040 struct allocindir *oldaip;
6041 struct freefrag *freefrag;
6043 struct ufsmount *ump;
6049 if (bp->b_lblkno >= 0)
6050 panic("setup_allocindir_phase2: not indir blk");
6051 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs),
6052 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset));
6053 indirdep = indirdep_lookup(mp, ip, bp);
6054 KASSERT(indirdep->ir_savebp != NULL,
6055 ("setup_allocindir_phase2 NULL ir_savebp"));
6056 aip->ai_indirdep = indirdep;
6058 * Check for an unwritten dependency for this indirect offset. If
6059 * there is, merge the old dependency into the new one. This happens
6060 * as a result of reallocblk only.
6063 if (aip->ai_oldblkno != 0) {
6064 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) {
6065 if (oldaip->ai_offset == aip->ai_offset) {
6066 freefrag = allocindir_merge(aip, oldaip);
6070 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) {
6071 if (oldaip->ai_offset == aip->ai_offset) {
6072 freefrag = allocindir_merge(aip, oldaip);
6078 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
6083 * Merge two allocindirs which refer to the same block. Move newblock
6084 * dependencies and setup the freefrags appropriately.
6086 static struct freefrag *
6087 allocindir_merge(aip, oldaip)
6088 struct allocindir *aip;
6089 struct allocindir *oldaip;
6091 struct freefrag *freefrag;
6092 struct worklist *wk;
6094 if (oldaip->ai_newblkno != aip->ai_oldblkno)
6095 panic("allocindir_merge: blkno");
6096 aip->ai_oldblkno = oldaip->ai_oldblkno;
6097 freefrag = aip->ai_freefrag;
6098 aip->ai_freefrag = oldaip->ai_freefrag;
6099 oldaip->ai_freefrag = NULL;
6100 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
6102 * If we are tracking a new directory-block allocation,
6103 * move it from the old allocindir to the new allocindir.
6105 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
6106 WORKLIST_REMOVE(wk);
6107 if (!LIST_EMPTY(&oldaip->ai_newdirblk))
6108 panic("allocindir_merge: extra newdirblk");
6109 WORKLIST_INSERT(&aip->ai_newdirblk, wk);
6112 * We can skip journaling for this freefrag and just complete
6113 * any pending journal work for the allocindir that is being
6114 * removed after the freefrag completes.
6116 if (freefrag->ff_jdep)
6117 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep));
6118 LIST_REMOVE(oldaip, ai_next);
6119 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block,
6120 &freefrag->ff_list, &freefrag->ff_jwork);
6121 free_newblk(&oldaip->ai_block);
6127 setup_freedirect(freeblks, ip, i, needj)
6128 struct freeblks *freeblks;
6133 struct ufsmount *ump;
6137 blkno = DIP(ip, i_db[i]);
6140 DIP_SET(ip, i_db[i], 0);
6142 frags = sblksize(ump->um_fs, ip->i_size, i);
6143 frags = numfrags(ump->um_fs, frags);
6144 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj);
6148 setup_freeext(freeblks, ip, i, needj)
6149 struct freeblks *freeblks;
6154 struct ufsmount *ump;
6158 blkno = ip->i_din2->di_extb[i];
6161 ip->i_din2->di_extb[i] = 0;
6163 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i);
6164 frags = numfrags(ump->um_fs, frags);
6165 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj);
6169 setup_freeindir(freeblks, ip, i, lbn, needj)
6170 struct freeblks *freeblks;
6176 struct ufsmount *ump;
6179 blkno = DIP(ip, i_ib[i]);
6182 DIP_SET(ip, i_ib[i], 0);
6184 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag,
6188 static inline struct freeblks *
6193 struct freeblks *freeblks;
6195 freeblks = malloc(sizeof(struct freeblks),
6196 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
6197 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
6198 LIST_INIT(&freeblks->fb_jblkdephd);
6199 LIST_INIT(&freeblks->fb_jwork);
6200 freeblks->fb_ref = 0;
6201 freeblks->fb_cgwait = 0;
6202 freeblks->fb_state = ATTACHED;
6203 freeblks->fb_uid = ip->i_uid;
6204 freeblks->fb_inum = ip->i_number;
6205 freeblks->fb_vtype = ITOV(ip)->v_type;
6206 freeblks->fb_modrev = DIP(ip, i_modrev);
6207 freeblks->fb_devvp = ITODEVVP(ip);
6208 freeblks->fb_chkcnt = 0;
6209 freeblks->fb_len = 0;
6215 trunc_indirdep(indirdep, freeblks, bp, off)
6216 struct indirdep *indirdep;
6217 struct freeblks *freeblks;
6221 struct allocindir *aip, *aipn;
6224 * The first set of allocindirs won't be in savedbp.
6226 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn)
6227 if (aip->ai_offset > off)
6228 cancel_allocindir(aip, bp, freeblks, 1);
6229 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn)
6230 if (aip->ai_offset > off)
6231 cancel_allocindir(aip, bp, freeblks, 1);
6233 * These will exist in savedbp.
6235 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn)
6236 if (aip->ai_offset > off)
6237 cancel_allocindir(aip, NULL, freeblks, 0);
6238 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn)
6239 if (aip->ai_offset > off)
6240 cancel_allocindir(aip, NULL, freeblks, 0);
6244 * Follow the chain of indirects down to lastlbn creating a freework
6245 * structure for each. This will be used to start indir_trunc() at
6246 * the right offset and create the journal records for the parrtial
6247 * truncation. A second step will handle the truncated dependencies.
6250 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno)
6251 struct freeblks *freeblks;
6257 struct indirdep *indirdep;
6258 struct indirdep *indirn;
6259 struct freework *freework;
6260 struct newblk *newblk;
6262 struct ufsmount *ump;
6275 mp = freeblks->fb_list.wk_mp;
6277 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
6278 if ((bp->b_flags & B_CACHE) == 0) {
6279 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno);
6280 bp->b_iocmd = BIO_READ;
6281 bp->b_flags &= ~B_INVAL;
6282 bp->b_ioflags &= ~BIO_ERROR;
6283 vfs_busy_pages(bp, 0);
6284 bp->b_iooffset = dbtob(bp->b_blkno);
6289 racct_add_buf(curproc, bp, 0);
6290 PROC_UNLOCK(curproc);
6293 curthread->td_ru.ru_inblock++;
6294 error = bufwait(bp);
6300 level = lbn_level(lbn);
6301 lbnadd = lbn_offset(ump->um_fs, level);
6303 * Compute the offset of the last block we want to keep. Store
6304 * in the freework the first block we want to completely free.
6306 off = (lastlbn - -(lbn + level)) / lbnadd;
6307 if (off + 1 == NINDIR(ump->um_fs))
6309 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0);
6311 * Link the freework into the indirdep. This will prevent any new
6312 * allocations from proceeding until we are finished with the
6313 * truncate and the block is written.
6316 indirdep = indirdep_lookup(mp, ip, bp);
6317 if (indirdep->ir_freeblks)
6318 panic("setup_trunc_indir: indirdep already truncated.");
6319 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next);
6320 freework->fw_indir = indirdep;
6322 * Cancel any allocindirs that will not make it to disk.
6323 * We have to do this for all copies of the indirdep that
6324 * live on this newblk.
6326 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
6327 if (newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0,
6329 panic("setup_trunc_indir: lost block");
6330 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next)
6331 trunc_indirdep(indirn, freeblks, bp, off);
6333 trunc_indirdep(indirdep, freeblks, bp, off);
6336 * Creation is protected by the buf lock. The saveddata is only
6337 * needed if a full truncation follows a partial truncation but it
6338 * is difficult to allocate in that case so we fetch it anyway.
6340 if (indirdep->ir_saveddata == NULL)
6341 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
6344 /* Fetch the blkno of the child and the zero start offset. */
6345 if (I_IS_UFS1(ip)) {
6346 blkno = ((ufs1_daddr_t *)bp->b_data)[off];
6347 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1];
6349 blkno = ((ufs2_daddr_t *)bp->b_data)[off];
6350 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1];
6353 /* Zero the truncated pointers. */
6354 end = bp->b_data + bp->b_bcount;
6355 bzero(start, end - start);
6361 lbn++; /* adjust level */
6362 lbn -= (off * lbnadd);
6363 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno);
6367 * Complete the partial truncation of an indirect block setup by
6368 * setup_trunc_indir(). This zeros the truncated pointers in the saved
6369 * copy and writes them to disk before the freeblks is allowed to complete.
6372 complete_trunc_indir(freework)
6373 struct freework *freework;
6375 struct freework *fwn;
6376 struct indirdep *indirdep;
6377 struct ufsmount *ump;
6382 ump = VFSTOUFS(freework->fw_list.wk_mp);
6384 indirdep = freework->fw_indir;
6386 bp = indirdep->ir_bp;
6387 /* See if the block was discarded. */
6390 /* Inline part of getdirtybuf(). We dont want bremfree. */
6391 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0)
6393 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
6394 LOCK_PTR(ump)) == 0)
6398 freework->fw_state |= DEPCOMPLETE;
6399 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next);
6401 * Zero the pointers in the saved copy.
6403 if (indirdep->ir_state & UFS1FMT)
6404 start = sizeof(ufs1_daddr_t);
6406 start = sizeof(ufs2_daddr_t);
6407 start *= freework->fw_start;
6408 count = indirdep->ir_savebp->b_bcount - start;
6409 start += (uintptr_t)indirdep->ir_savebp->b_data;
6410 bzero((char *)start, count);
6412 * We need to start the next truncation in the list if it has not
6415 fwn = TAILQ_FIRST(&indirdep->ir_trunc);
6417 if (fwn->fw_freeblks == indirdep->ir_freeblks)
6418 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next);
6419 if ((fwn->fw_state & ONWORKLIST) == 0)
6420 freework_enqueue(fwn);
6423 * If bp is NULL the block was fully truncated, restore
6424 * the saved block list otherwise free it if it is no
6427 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
6429 bcopy(indirdep->ir_saveddata,
6430 indirdep->ir_savebp->b_data,
6431 indirdep->ir_savebp->b_bcount);
6432 free(indirdep->ir_saveddata, M_INDIRDEP);
6433 indirdep->ir_saveddata = NULL;
6436 * When bp is NULL there is a full truncation pending. We
6437 * must wait for this full truncation to be journaled before
6438 * we can release this freework because the disk pointers will
6439 * never be written as zero.
6442 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd))
6443 handle_written_freework(freework);
6445 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd,
6446 &freework->fw_list);
6448 /* Complete when the real copy is written. */
6449 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list);
6455 * Calculate the number of blocks we are going to release where datablocks
6456 * is the current total and length is the new file size.
6459 blkcount(fs, datablocks, length)
6461 ufs2_daddr_t datablocks;
6464 off_t totblks, numblks;
6467 numblks = howmany(length, fs->fs_bsize);
6468 if (numblks <= UFS_NDADDR) {
6469 totblks = howmany(length, fs->fs_fsize);
6472 totblks = blkstofrags(fs, numblks);
6473 numblks -= UFS_NDADDR;
6475 * Count all single, then double, then triple indirects required.
6476 * Subtracting one indirects worth of blocks for each pass
6477 * acknowledges one of each pointed to by the inode.
6480 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs)));
6481 numblks -= NINDIR(fs);
6484 numblks = howmany(numblks, NINDIR(fs));
6487 totblks = fsbtodb(fs, totblks);
6489 * Handle sparse files. We can't reclaim more blocks than the inode
6490 * references. We will correct it later in handle_complete_freeblks()
6491 * when we know the real count.
6493 if (totblks > datablocks)
6495 return (datablocks - totblks);
6499 * Handle freeblocks for journaled softupdate filesystems.
6501 * Contrary to normal softupdates, we must preserve the block pointers in
6502 * indirects until their subordinates are free. This is to avoid journaling
6503 * every block that is freed which may consume more space than the journal
6504 * itself. The recovery program will see the free block journals at the
6505 * base of the truncated area and traverse them to reclaim space. The
6506 * pointers in the inode may be cleared immediately after the journal
6507 * records are written because each direct and indirect pointer in the
6508 * inode is recorded in a journal. This permits full truncation to proceed
6509 * asynchronously. The write order is journal -> inode -> cgs -> indirects.
6511 * The algorithm is as follows:
6512 * 1) Traverse the in-memory state and create journal entries to release
6513 * the relevant blocks and full indirect trees.
6514 * 2) Traverse the indirect block chain adding partial truncation freework
6515 * records to indirects in the path to lastlbn. The freework will
6516 * prevent new allocation dependencies from being satisfied in this
6517 * indirect until the truncation completes.
6518 * 3) Read and lock the inode block, performing an update with the new size
6519 * and pointers. This prevents truncated data from becoming valid on
6520 * disk through step 4.
6521 * 4) Reap unsatisfied dependencies that are beyond the truncated area,
6522 * eliminate journal work for those records that do not require it.
6523 * 5) Schedule the journal records to be written followed by the inode block.
6524 * 6) Allocate any necessary frags for the end of file.
6525 * 7) Zero any partially truncated blocks.
6527 * From this truncation proceeds asynchronously using the freework and
6528 * indir_trunc machinery. The file will not be extended again into a
6529 * partially truncated indirect block until all work is completed but
6530 * the normal dependency mechanism ensures that it is rolled back/forward
6531 * as appropriate. Further truncation may occur without delay and is
6532 * serialized in indir_trunc().
6535 softdep_journal_freeblocks(ip, cred, length, flags)
6536 struct inode *ip; /* The inode whose length is to be reduced */
6538 off_t length; /* The new length for the file */
6539 int flags; /* IO_EXT and/or IO_NORMAL */
6541 struct freeblks *freeblks, *fbn;
6542 struct worklist *wk, *wkn;
6543 struct inodedep *inodedep;
6544 struct jblkdep *jblkdep;
6545 struct allocdirect *adp, *adpn;
6546 struct ufsmount *ump;
6551 ufs2_daddr_t extblocks, datablocks;
6552 ufs_lbn_t tmpval, lbn, lastlbn;
6553 int frags, lastoff, iboff, allocblock, needj, error, i;
6558 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6559 ("softdep_journal_freeblocks called on non-softdep filesystem"));
6567 freeblks = newfreeblks(mp, ip);
6570 * If we're truncating a removed file that will never be written
6571 * we don't need to journal the block frees. The canceled journals
6572 * for the allocations will suffice.
6574 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6575 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
6578 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d",
6579 ip->i_number, length, needj);
6582 * Calculate the lbn that we are truncating to. This results in -1
6583 * if we're truncating the 0 bytes. So it is the last lbn we want
6584 * to keep, not the first lbn we want to truncate.
6586 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1;
6587 lastoff = blkoff(fs, length);
6589 * Compute frags we are keeping in lastlbn. 0 means all.
6591 if (lastlbn >= 0 && lastlbn < UFS_NDADDR) {
6592 frags = fragroundup(fs, lastoff);
6593 /* adp offset of last valid allocdirect. */
6595 } else if (lastlbn > 0)
6597 if (fs->fs_magic == FS_UFS2_MAGIC)
6598 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6600 * Handle normal data blocks and indirects. This section saves
6601 * values used after the inode update to complete frag and indirect
6604 if ((flags & IO_NORMAL) != 0) {
6606 * Handle truncation of whole direct and indirect blocks.
6608 for (i = iboff + 1; i < UFS_NDADDR; i++)
6609 setup_freedirect(freeblks, ip, i, needj);
6610 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR;
6612 i++, lbn += tmpval, tmpval *= NINDIR(fs)) {
6613 /* Release a whole indirect tree. */
6614 if (lbn > lastlbn) {
6615 setup_freeindir(freeblks, ip, i, -lbn -i,
6619 iboff = i + UFS_NDADDR;
6621 * Traverse partially truncated indirect tree.
6623 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn)
6624 setup_trunc_indir(freeblks, ip, -lbn - i,
6625 lastlbn, DIP(ip, i_ib[i]));
6628 * Handle partial truncation to a frag boundary.
6634 oldfrags = blksize(fs, ip, lastlbn);
6635 blkno = DIP(ip, i_db[lastlbn]);
6636 if (blkno && oldfrags != frags) {
6638 oldfrags = numfrags(fs, oldfrags);
6639 blkno += numfrags(fs, frags);
6640 newfreework(ump, freeblks, NULL, lastlbn,
6641 blkno, oldfrags, 0, needj);
6643 adjust_newfreework(freeblks,
6644 numfrags(fs, frags));
6645 } else if (blkno == 0)
6649 * Add a journal record for partial truncate if we are
6650 * handling indirect blocks. Non-indirects need no extra
6653 if (length != 0 && lastlbn >= UFS_NDADDR) {
6654 ip->i_flag |= IN_TRUNCATED;
6655 newjtrunc(freeblks, length, 0);
6657 ip->i_size = length;
6658 DIP_SET(ip, i_size, ip->i_size);
6659 datablocks = DIP(ip, i_blocks) - extblocks;
6661 datablocks = blkcount(fs, datablocks, length);
6662 freeblks->fb_len = length;
6664 if ((flags & IO_EXT) != 0) {
6665 for (i = 0; i < UFS_NXADDR; i++)
6666 setup_freeext(freeblks, ip, i, needj);
6667 ip->i_din2->di_extsize = 0;
6668 datablocks += extblocks;
6671 /* Reference the quotas in case the block count is wrong in the end. */
6672 quotaref(vp, freeblks->fb_quota);
6673 (void) chkdq(ip, -datablocks, NOCRED, FORCE);
6675 freeblks->fb_chkcnt = -datablocks;
6677 fs->fs_pendingblocks += datablocks;
6679 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6681 * Handle truncation of incomplete alloc direct dependencies. We
6682 * hold the inode block locked to prevent incomplete dependencies
6683 * from reaching the disk while we are eliminating those that
6684 * have been truncated. This is a partially inlined ffs_update().
6687 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
6688 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6689 (int)fs->fs_bsize, cred, &bp);
6692 softdep_error("softdep_journal_freeblocks", error);
6695 if (bp->b_bufsize == fs->fs_bsize)
6696 bp->b_flags |= B_CLUSTEROK;
6697 softdep_update_inodeblock(ip, bp, 0);
6698 if (ump->um_fstype == UFS1)
6699 *((struct ufs1_dinode *)bp->b_data +
6700 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
6702 *((struct ufs2_dinode *)bp->b_data +
6703 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
6705 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6706 if ((inodedep->id_state & IOSTARTED) != 0)
6707 panic("softdep_setup_freeblocks: inode busy");
6709 * Add the freeblks structure to the list of operations that
6710 * must await the zero'ed inode being written to disk. If we
6711 * still have a bitmap dependency (needj), then the inode
6712 * has never been written to disk, so we can process the
6713 * freeblks below once we have deleted the dependencies.
6716 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6718 freeblks->fb_state |= COMPLETE;
6719 if ((flags & IO_NORMAL) != 0) {
6720 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) {
6721 if (adp->ad_offset > iboff)
6722 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6725 * Truncate the allocdirect. We could eliminate
6726 * or modify journal records as well.
6728 else if (adp->ad_offset == iboff && frags)
6729 adp->ad_newsize = frags;
6732 if ((flags & IO_EXT) != 0)
6733 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6734 cancel_allocdirect(&inodedep->id_extupdt, adp,
6737 * Scan the bufwait list for newblock dependencies that will never
6740 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) {
6741 if (wk->wk_type != D_ALLOCDIRECT)
6743 adp = WK_ALLOCDIRECT(wk);
6744 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) ||
6745 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) {
6746 cancel_jfreeblk(freeblks, adp->ad_newblkno);
6747 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork);
6748 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
6754 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps)
6755 add_to_journal(&jblkdep->jb_list);
6759 * Truncate dependency structures beyond length.
6761 trunc_dependencies(ip, freeblks, lastlbn, frags, flags);
6763 * This is only set when we need to allocate a fragment because
6764 * none existed at the end of a frag-sized file. It handles only
6765 * allocating a new, zero filled block.
6768 ip->i_size = length - lastoff;
6769 DIP_SET(ip, i_size, ip->i_size);
6770 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp);
6772 softdep_error("softdep_journal_freeblks", error);
6775 ip->i_size = length;
6776 DIP_SET(ip, i_size, length);
6777 ip->i_flag |= IN_CHANGE | IN_UPDATE;
6778 allocbuf(bp, frags);
6781 } else if (lastoff != 0 && vp->v_type != VDIR) {
6785 * Zero the end of a truncated frag or block.
6787 size = sblksize(fs, length, lastlbn);
6788 error = bread(vp, lastlbn, size, cred, &bp);
6790 softdep_error("softdep_journal_freeblks", error);
6793 bzero((char *)bp->b_data + lastoff, size - lastoff);
6798 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6799 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
6800 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
6802 * We zero earlier truncations so they don't erroneously
6805 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0)
6806 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next)
6808 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE &&
6809 LIST_EMPTY(&freeblks->fb_jblkdephd))
6810 freeblks->fb_state |= INPROGRESS;
6815 handle_workitem_freeblocks(freeblks, 0);
6816 trunc_pages(ip, length, extblocks, flags);
6821 * Flush a JOP_SYNC to the journal.
6824 softdep_journal_fsync(ip)
6827 struct jfsync *jfsync;
6828 struct ufsmount *ump;
6831 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
6832 ("softdep_journal_fsync called on non-softdep filesystem"));
6833 if ((ip->i_flag & IN_TRUNCATED) == 0)
6835 ip->i_flag &= ~IN_TRUNCATED;
6836 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO);
6837 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump));
6838 jfsync->jfs_size = ip->i_size;
6839 jfsync->jfs_ino = ip->i_number;
6841 add_to_journal(&jfsync->jfs_list);
6842 jwait(&jfsync->jfs_list, MNT_WAIT);
6847 * Block de-allocation dependencies.
6849 * When blocks are de-allocated, the on-disk pointers must be nullified before
6850 * the blocks are made available for use by other files. (The true
6851 * requirement is that old pointers must be nullified before new on-disk
6852 * pointers are set. We chose this slightly more stringent requirement to
6853 * reduce complexity.) Our implementation handles this dependency by updating
6854 * the inode (or indirect block) appropriately but delaying the actual block
6855 * de-allocation (i.e., freemap and free space count manipulation) until
6856 * after the updated versions reach stable storage. After the disk is
6857 * updated, the blocks can be safely de-allocated whenever it is convenient.
6858 * This implementation handles only the common case of reducing a file's
6859 * length to zero. Other cases are handled by the conventional synchronous
6862 * The ffs implementation with which we worked double-checks
6863 * the state of the block pointers and file size as it reduces
6864 * a file's length. Some of this code is replicated here in our
6865 * soft updates implementation. The freeblks->fb_chkcnt field is
6866 * used to transfer a part of this information to the procedure
6867 * that eventually de-allocates the blocks.
6869 * This routine should be called from the routine that shortens
6870 * a file's length, before the inode's size or block pointers
6871 * are modified. It will save the block pointer information for
6872 * later release and zero the inode so that the calling routine
6876 softdep_setup_freeblocks(ip, length, flags)
6877 struct inode *ip; /* The inode whose length is to be reduced */
6878 off_t length; /* The new length for the file */
6879 int flags; /* IO_EXT and/or IO_NORMAL */
6881 struct ufs1_dinode *dp1;
6882 struct ufs2_dinode *dp2;
6883 struct freeblks *freeblks;
6884 struct inodedep *inodedep;
6885 struct allocdirect *adp;
6886 struct ufsmount *ump;
6889 ufs2_daddr_t extblocks, datablocks;
6891 int i, delay, error;
6897 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6898 ("softdep_setup_freeblocks called on non-softdep filesystem"));
6899 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld",
6900 ip->i_number, length);
6901 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length"));
6903 if ((error = bread(ump->um_devvp,
6904 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6905 (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
6907 softdep_error("softdep_setup_freeblocks", error);
6910 freeblks = newfreeblks(mp, ip);
6913 if (fs->fs_magic == FS_UFS2_MAGIC)
6914 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6915 if ((flags & IO_NORMAL) != 0) {
6916 for (i = 0; i < UFS_NDADDR; i++)
6917 setup_freedirect(freeblks, ip, i, 0);
6918 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR;
6920 i++, lbn += tmpval, tmpval *= NINDIR(fs))
6921 setup_freeindir(freeblks, ip, i, -lbn -i, 0);
6923 DIP_SET(ip, i_size, 0);
6924 datablocks = DIP(ip, i_blocks) - extblocks;
6926 if ((flags & IO_EXT) != 0) {
6927 for (i = 0; i < UFS_NXADDR; i++)
6928 setup_freeext(freeblks, ip, i, 0);
6929 ip->i_din2->di_extsize = 0;
6930 datablocks += extblocks;
6933 /* Reference the quotas in case the block count is wrong in the end. */
6934 quotaref(ITOV(ip), freeblks->fb_quota);
6935 (void) chkdq(ip, -datablocks, NOCRED, FORCE);
6937 freeblks->fb_chkcnt = -datablocks;
6939 fs->fs_pendingblocks += datablocks;
6941 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6943 * Push the zero'ed inode to its disk buffer so that we are free
6944 * to delete its dependencies below. Once the dependencies are gone
6945 * the buffer can be safely released.
6947 if (ump->um_fstype == UFS1) {
6948 dp1 = ((struct ufs1_dinode *)bp->b_data +
6949 ino_to_fsbo(fs, ip->i_number));
6950 ip->i_din1->di_freelink = dp1->di_freelink;
6953 dp2 = ((struct ufs2_dinode *)bp->b_data +
6954 ino_to_fsbo(fs, ip->i_number));
6955 ip->i_din2->di_freelink = dp2->di_freelink;
6959 * Find and eliminate any inode dependencies.
6962 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6963 if ((inodedep->id_state & IOSTARTED) != 0)
6964 panic("softdep_setup_freeblocks: inode busy");
6966 * Add the freeblks structure to the list of operations that
6967 * must await the zero'ed inode being written to disk. If we
6968 * still have a bitmap dependency (delay == 0), then the inode
6969 * has never been written to disk, so we can process the
6970 * freeblks below once we have deleted the dependencies.
6972 delay = (inodedep->id_state & DEPCOMPLETE);
6974 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6976 freeblks->fb_state |= COMPLETE;
6978 * Because the file length has been truncated to zero, any
6979 * pending block allocation dependency structures associated
6980 * with this inode are obsolete and can simply be de-allocated.
6981 * We must first merge the two dependency lists to get rid of
6982 * any duplicate freefrag structures, then purge the merged list.
6983 * If we still have a bitmap dependency, then the inode has never
6984 * been written to disk, so we can free any fragments without delay.
6986 if (flags & IO_NORMAL) {
6987 merge_inode_lists(&inodedep->id_newinoupdt,
6988 &inodedep->id_inoupdt);
6989 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
6990 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6993 if (flags & IO_EXT) {
6994 merge_inode_lists(&inodedep->id_newextupdt,
6995 &inodedep->id_extupdt);
6996 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6997 cancel_allocdirect(&inodedep->id_extupdt, adp,
7002 trunc_dependencies(ip, freeblks, -1, 0, flags);
7004 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
7005 (void) free_inodedep(inodedep);
7006 freeblks->fb_state |= DEPCOMPLETE;
7008 * If the inode with zeroed block pointers is now on disk
7009 * we can start freeing blocks.
7011 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
7012 freeblks->fb_state |= INPROGRESS;
7017 handle_workitem_freeblocks(freeblks, 0);
7018 trunc_pages(ip, length, extblocks, flags);
7022 * Eliminate pages from the page cache that back parts of this inode and
7023 * adjust the vnode pager's idea of our size. This prevents stale data
7024 * from hanging around in the page cache.
7027 trunc_pages(ip, length, extblocks, flags)
7030 ufs2_daddr_t extblocks;
7040 extend = OFF_TO_IDX(lblktosize(fs, -extblocks));
7041 if ((flags & IO_EXT) != 0)
7042 vn_pages_remove(vp, extend, 0);
7043 if ((flags & IO_NORMAL) == 0)
7045 BO_LOCK(&vp->v_bufobj);
7047 BO_UNLOCK(&vp->v_bufobj);
7049 * The vnode pager eliminates file pages we eliminate indirects
7052 vnode_pager_setsize(vp, length);
7054 * Calculate the end based on the last indirect we want to keep. If
7055 * the block extends into indirects we can just use the negative of
7056 * its lbn. Doubles and triples exist at lower numbers so we must
7057 * be careful not to remove those, if they exist. double and triple
7058 * indirect lbns do not overlap with others so it is not important
7059 * to verify how many levels are required.
7061 lbn = lblkno(fs, length);
7062 if (lbn >= UFS_NDADDR) {
7063 /* Calculate the virtual lbn of the triple indirect. */
7064 lbn = -lbn - (UFS_NIADDR - 1);
7065 end = OFF_TO_IDX(lblktosize(fs, lbn));
7068 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end);
7072 * See if the buf bp is in the range eliminated by truncation.
7075 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags)
7085 /* Only match ext/normal blocks as appropriate. */
7086 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
7087 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0))
7089 /* ALTDATA is always a full truncation. */
7090 if ((bp->b_xflags & BX_ALTDATA) != 0)
7092 /* -1 is full truncation. */
7096 * If this is a partial truncate we only want those
7097 * blocks and indirect blocks that cover the range
7102 lbn = -(lbn + lbn_level(lbn));
7105 /* Here we only truncate lblkno if it's partial. */
7106 if (lbn == lastlbn) {
7115 * Eliminate any dependencies that exist in memory beyond lblkno:off
7118 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags)
7120 struct freeblks *freeblks;
7131 * We must wait for any I/O in progress to finish so that
7132 * all potential buffers on the dirty list will be visible.
7133 * Once they are all there, walk the list and get rid of
7140 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
7141 bp->b_vflags &= ~BV_SCANNED;
7143 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
7144 if (bp->b_vflags & BV_SCANNED)
7146 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7147 bp->b_vflags |= BV_SCANNED;
7150 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer"));
7151 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL)
7154 if (deallocate_dependencies(bp, freeblks, blkoff))
7162 * Now do the work of vtruncbuf while also matching indirect blocks.
7164 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs)
7165 bp->b_vflags &= ~BV_SCANNED;
7167 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) {
7168 if (bp->b_vflags & BV_SCANNED)
7170 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7171 bp->b_vflags |= BV_SCANNED;
7175 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
7176 BO_LOCKPTR(bo)) == ENOLCK) {
7180 bp->b_vflags |= BV_SCANNED;
7183 allocbuf(bp, blkoff);
7186 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF;
7197 cancel_pagedep(pagedep, freeblks, blkoff)
7198 struct pagedep *pagedep;
7199 struct freeblks *freeblks;
7202 struct jremref *jremref;
7203 struct jmvref *jmvref;
7204 struct dirrem *dirrem, *tmp;
7208 * Copy any directory remove dependencies to the list
7209 * to be processed after the freeblks proceeds. If
7210 * directory entry never made it to disk they
7211 * can be dumped directly onto the work list.
7213 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) {
7214 /* Skip this directory removal if it is intended to remain. */
7215 if (dirrem->dm_offset < blkoff)
7218 * If there are any dirrems we wait for the journal write
7219 * to complete and then restart the buf scan as the lock
7222 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
7223 jwait(&jremref->jr_list, MNT_WAIT);
7226 LIST_REMOVE(dirrem, dm_next);
7227 dirrem->dm_dirinum = pagedep->pd_ino;
7228 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list);
7230 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
7231 jwait(&jmvref->jm_list, MNT_WAIT);
7235 * When we're partially truncating a pagedep we just want to flush
7236 * journal entries and return. There can not be any adds in the
7237 * truncated portion of the directory and newblk must remain if
7238 * part of the block remains.
7243 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
7244 if (dap->da_offset > blkoff)
7245 panic("cancel_pagedep: diradd %p off %d > %d",
7246 dap, dap->da_offset, blkoff);
7247 for (i = 0; i < DAHASHSZ; i++)
7248 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist)
7249 if (dap->da_offset > blkoff)
7250 panic("cancel_pagedep: diradd %p off %d > %d",
7251 dap, dap->da_offset, blkoff);
7255 * There should be no directory add dependencies present
7256 * as the directory could not be truncated until all
7257 * children were removed.
7259 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
7260 ("deallocate_dependencies: pendinghd != NULL"));
7261 for (i = 0; i < DAHASHSZ; i++)
7262 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
7263 ("deallocate_dependencies: diraddhd != NULL"));
7264 if ((pagedep->pd_state & NEWBLOCK) != 0)
7265 free_newdirblk(pagedep->pd_newdirblk);
7266 if (free_pagedep(pagedep) == 0)
7267 panic("Failed to free pagedep %p", pagedep);
7272 * Reclaim any dependency structures from a buffer that is about to
7273 * be reallocated to a new vnode. The buffer must be locked, thus,
7274 * no I/O completion operations can occur while we are manipulating
7275 * its associated dependencies. The mutex is held so that other I/O's
7276 * associated with related dependencies do not occur.
7279 deallocate_dependencies(bp, freeblks, off)
7281 struct freeblks *freeblks;
7284 struct indirdep *indirdep;
7285 struct pagedep *pagedep;
7286 struct worklist *wk, *wkn;
7287 struct ufsmount *ump;
7289 ump = softdep_bp_to_mp(bp);
7293 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) {
7294 switch (wk->wk_type) {
7296 indirdep = WK_INDIRDEP(wk);
7297 if (bp->b_lblkno >= 0 ||
7298 bp->b_blkno != indirdep->ir_savebp->b_lblkno)
7299 panic("deallocate_dependencies: not indir");
7300 cancel_indirdep(indirdep, bp, freeblks);
7304 pagedep = WK_PAGEDEP(wk);
7305 if (cancel_pagedep(pagedep, freeblks, off)) {
7313 * Simply remove the allocindir, we'll find it via
7314 * the indirdep where we can clear pointers if
7317 WORKLIST_REMOVE(wk);
7322 * A truncation is waiting for the zero'd pointers
7323 * to be written. It can be freed when the freeblks
7326 WORKLIST_REMOVE(wk);
7327 wk->wk_state |= ONDEPLIST;
7328 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
7336 panic("deallocate_dependencies: Unexpected type %s",
7337 TYPENAME(wk->wk_type));
7344 * Don't throw away this buf, we were partially truncating and
7345 * some deps may always remain.
7349 bp->b_vflags |= BV_SCANNED;
7352 bp->b_flags |= B_INVAL | B_NOCACHE;
7358 * An allocdirect is being canceled due to a truncate. We must make sure
7359 * the journal entry is released in concert with the blkfree that releases
7360 * the storage. Completed journal entries must not be released until the
7361 * space is no longer pointed to by the inode or in the bitmap.
7364 cancel_allocdirect(adphead, adp, freeblks)
7365 struct allocdirectlst *adphead;
7366 struct allocdirect *adp;
7367 struct freeblks *freeblks;
7369 struct freework *freework;
7370 struct newblk *newblk;
7371 struct worklist *wk;
7373 TAILQ_REMOVE(adphead, adp, ad_next);
7374 newblk = (struct newblk *)adp;
7377 * Find the correct freework structure.
7379 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
7380 if (wk->wk_type != D_FREEWORK)
7382 freework = WK_FREEWORK(wk);
7383 if (freework->fw_blkno == newblk->nb_newblkno)
7386 if (freework == NULL)
7387 panic("cancel_allocdirect: Freework not found");
7389 * If a newblk exists at all we still have the journal entry that
7390 * initiated the allocation so we do not need to journal the free.
7392 cancel_jfreeblk(freeblks, freework->fw_blkno);
7394 * If the journal hasn't been written the jnewblk must be passed
7395 * to the call to ffs_blkfree that reclaims the space. We accomplish
7396 * this by linking the journal dependency into the freework to be
7397 * freed when freework_freeblock() is called. If the journal has
7398 * been written we can simply reclaim the journal space when the
7399 * freeblks work is complete.
7401 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list,
7402 &freeblks->fb_jwork);
7403 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
7408 * Cancel a new block allocation. May be an indirect or direct block. We
7409 * remove it from various lists and return any journal record that needs to
7410 * be resolved by the caller.
7412 * A special consideration is made for indirects which were never pointed
7413 * at on disk and will never be found once this block is released.
7415 static struct jnewblk *
7416 cancel_newblk(newblk, wk, wkhd)
7417 struct newblk *newblk;
7418 struct worklist *wk;
7419 struct workhead *wkhd;
7421 struct jnewblk *jnewblk;
7423 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno);
7425 newblk->nb_state |= GOINGAWAY;
7427 * Previously we traversed the completedhd on each indirdep
7428 * attached to this newblk to cancel them and gather journal
7429 * work. Since we need only the oldest journal segment and
7430 * the lowest point on the tree will always have the oldest
7431 * journal segment we are free to release the segments
7432 * of any subordinates and may leave the indirdep list to
7433 * indirdep_complete() when this newblk is freed.
7435 if (newblk->nb_state & ONDEPLIST) {
7436 newblk->nb_state &= ~ONDEPLIST;
7437 LIST_REMOVE(newblk, nb_deps);
7439 if (newblk->nb_state & ONWORKLIST)
7440 WORKLIST_REMOVE(&newblk->nb_list);
7442 * If the journal entry hasn't been written we save a pointer to
7443 * the dependency that frees it until it is written or the
7444 * superseding operation completes.
7446 jnewblk = newblk->nb_jnewblk;
7447 if (jnewblk != NULL && wk != NULL) {
7448 newblk->nb_jnewblk = NULL;
7449 jnewblk->jn_dep = wk;
7451 if (!LIST_EMPTY(&newblk->nb_jwork))
7452 jwork_move(wkhd, &newblk->nb_jwork);
7454 * When truncating we must free the newdirblk early to remove
7455 * the pagedep from the hash before returning.
7457 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7458 free_newdirblk(WK_NEWDIRBLK(wk));
7459 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7460 panic("cancel_newblk: extra newdirblk");
7466 * Schedule the freefrag associated with a newblk to be released once
7467 * the pointers are written and the previous block is no longer needed.
7470 newblk_freefrag(newblk)
7471 struct newblk *newblk;
7473 struct freefrag *freefrag;
7475 if (newblk->nb_freefrag == NULL)
7477 freefrag = newblk->nb_freefrag;
7478 newblk->nb_freefrag = NULL;
7479 freefrag->ff_state |= COMPLETE;
7480 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
7481 add_to_worklist(&freefrag->ff_list, 0);
7485 * Free a newblk. Generate a new freefrag work request if appropriate.
7486 * This must be called after the inode pointer and any direct block pointers
7487 * are valid or fully removed via truncate or frag extension.
7491 struct newblk *newblk;
7493 struct indirdep *indirdep;
7494 struct worklist *wk;
7496 KASSERT(newblk->nb_jnewblk == NULL,
7497 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk));
7498 KASSERT(newblk->nb_list.wk_type != D_NEWBLK,
7499 ("free_newblk: unclaimed newblk"));
7500 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp));
7501 newblk_freefrag(newblk);
7502 if (newblk->nb_state & ONDEPLIST)
7503 LIST_REMOVE(newblk, nb_deps);
7504 if (newblk->nb_state & ONWORKLIST)
7505 WORKLIST_REMOVE(&newblk->nb_list);
7506 LIST_REMOVE(newblk, nb_hash);
7507 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7508 free_newdirblk(WK_NEWDIRBLK(wk));
7509 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7510 panic("free_newblk: extra newdirblk");
7511 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL)
7512 indirdep_complete(indirdep);
7513 handle_jwork(&newblk->nb_jwork);
7514 WORKITEM_FREE(newblk, D_NEWBLK);
7518 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
7521 free_newdirblk(newdirblk)
7522 struct newdirblk *newdirblk;
7524 struct pagedep *pagedep;
7526 struct worklist *wk;
7528 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp));
7529 WORKLIST_REMOVE(&newdirblk->db_list);
7531 * If the pagedep is still linked onto the directory buffer
7532 * dependency chain, then some of the entries on the
7533 * pd_pendinghd list may not be committed to disk yet. In
7534 * this case, we will simply clear the NEWBLOCK flag and
7535 * let the pd_pendinghd list be processed when the pagedep
7536 * is next written. If the pagedep is no longer on the buffer
7537 * dependency chain, then all the entries on the pd_pending
7538 * list are committed to disk and we can free them here.
7540 pagedep = newdirblk->db_pagedep;
7541 pagedep->pd_state &= ~NEWBLOCK;
7542 if ((pagedep->pd_state & ONWORKLIST) == 0) {
7543 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
7544 free_diradd(dap, NULL);
7546 * If no dependencies remain, the pagedep will be freed.
7548 free_pagedep(pagedep);
7550 /* Should only ever be one item in the list. */
7551 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
7552 WORKLIST_REMOVE(wk);
7553 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
7555 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
7559 * Prepare an inode to be freed. The actual free operation is not
7560 * done until the zero'ed inode has been written to disk.
7563 softdep_freefile(pvp, ino, mode)
7568 struct inode *ip = VTOI(pvp);
7569 struct inodedep *inodedep;
7570 struct freefile *freefile;
7571 struct freeblks *freeblks;
7572 struct ufsmount *ump;
7575 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
7576 ("softdep_freefile called on non-softdep filesystem"));
7578 * This sets up the inode de-allocation dependency.
7580 freefile = malloc(sizeof(struct freefile),
7581 M_FREEFILE, M_SOFTDEP_FLAGS);
7582 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
7583 freefile->fx_mode = mode;
7584 freefile->fx_oldinum = ino;
7585 freefile->fx_devvp = ump->um_devvp;
7586 LIST_INIT(&freefile->fx_jwork);
7588 ump->um_fs->fs_pendinginodes += 1;
7592 * If the inodedep does not exist, then the zero'ed inode has
7593 * been written to disk. If the allocated inode has never been
7594 * written to disk, then the on-disk inode is zero'ed. In either
7595 * case we can free the file immediately. If the journal was
7596 * canceled before being written the inode will never make it to
7597 * disk and we must send the canceled journal entrys to
7598 * ffs_freefile() to be cleared in conjunction with the bitmap.
7599 * Any blocks waiting on the inode to write can be safely freed
7600 * here as it will never been written.
7603 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7606 * Clear out freeblks that no longer need to reference
7610 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) {
7611 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks,
7613 freeblks->fb_state &= ~ONDEPLIST;
7616 * Remove this inode from the unlinked list.
7618 if (inodedep->id_state & UNLINKED) {
7620 * Save the journal work to be freed with the bitmap
7621 * before we clear UNLINKED. Otherwise it can be lost
7622 * if the inode block is written.
7624 handle_bufwait(inodedep, &freefile->fx_jwork);
7625 clear_unlinked_inodedep(inodedep);
7627 * Re-acquire inodedep as we've dropped the
7628 * per-filesystem lock in clear_unlinked_inodedep().
7630 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7633 if (inodedep == NULL || check_inode_unwritten(inodedep)) {
7635 handle_workitem_freefile(freefile);
7638 if ((inodedep->id_state & DEPCOMPLETE) == 0)
7639 inodedep->id_state |= GOINGAWAY;
7640 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
7642 if (ip->i_number == ino)
7643 ip->i_flag |= IN_MODIFIED;
7647 * Check to see if an inode has never been written to disk. If
7648 * so free the inodedep and return success, otherwise return failure.
7650 * If we still have a bitmap dependency, then the inode has never
7651 * been written to disk. Drop the dependency as it is no longer
7652 * necessary since the inode is being deallocated. We set the
7653 * ALLCOMPLETE flags since the bitmap now properly shows that the
7654 * inode is not allocated. Even if the inode is actively being
7655 * written, it has been rolled back to its zero'ed state, so we
7656 * are ensured that a zero inode is what is on the disk. For short
7657 * lived files, this change will usually result in removing all the
7658 * dependencies from the inode so that it can be freed immediately.
7661 check_inode_unwritten(inodedep)
7662 struct inodedep *inodedep;
7665 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7667 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
7668 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7669 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7670 !LIST_EMPTY(&inodedep->id_bufwait) ||
7671 !LIST_EMPTY(&inodedep->id_inowait) ||
7672 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7673 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7674 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7675 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7676 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7677 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7678 inodedep->id_mkdiradd != NULL ||
7679 inodedep->id_nlinkdelta != 0)
7682 * Another process might be in initiate_write_inodeblock_ufs[12]
7683 * trying to allocate memory without holding "Softdep Lock".
7685 if ((inodedep->id_state & IOSTARTED) != 0 &&
7686 inodedep->id_savedino1 == NULL)
7689 if (inodedep->id_state & ONDEPLIST)
7690 LIST_REMOVE(inodedep, id_deps);
7691 inodedep->id_state &= ~ONDEPLIST;
7692 inodedep->id_state |= ALLCOMPLETE;
7693 inodedep->id_bmsafemap = NULL;
7694 if (inodedep->id_state & ONWORKLIST)
7695 WORKLIST_REMOVE(&inodedep->id_list);
7696 if (inodedep->id_savedino1 != NULL) {
7697 free(inodedep->id_savedino1, M_SAVEDINO);
7698 inodedep->id_savedino1 = NULL;
7700 if (free_inodedep(inodedep) == 0)
7701 panic("check_inode_unwritten: busy inode");
7706 check_inodedep_free(inodedep)
7707 struct inodedep *inodedep;
7710 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7711 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
7712 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7713 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7714 !LIST_EMPTY(&inodedep->id_bufwait) ||
7715 !LIST_EMPTY(&inodedep->id_inowait) ||
7716 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7717 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7718 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7719 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7720 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7721 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7722 inodedep->id_mkdiradd != NULL ||
7723 inodedep->id_nlinkdelta != 0 ||
7724 inodedep->id_savedino1 != NULL)
7730 * Try to free an inodedep structure. Return 1 if it could be freed.
7733 free_inodedep(inodedep)
7734 struct inodedep *inodedep;
7737 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7738 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
7739 !check_inodedep_free(inodedep))
7741 if (inodedep->id_state & ONDEPLIST)
7742 LIST_REMOVE(inodedep, id_deps);
7743 LIST_REMOVE(inodedep, id_hash);
7744 WORKITEM_FREE(inodedep, D_INODEDEP);
7749 * Free the block referenced by a freework structure. The parent freeblks
7750 * structure is released and completed when the final cg bitmap reaches
7751 * the disk. This routine may be freeing a jnewblk which never made it to
7752 * disk in which case we do not have to wait as the operation is undone
7753 * in memory immediately.
7756 freework_freeblock(freework, key)
7757 struct freework *freework;
7760 struct freeblks *freeblks;
7761 struct jnewblk *jnewblk;
7762 struct ufsmount *ump;
7763 struct workhead wkhd;
7768 ump = VFSTOUFS(freework->fw_list.wk_mp);
7771 * Handle partial truncate separately.
7773 if (freework->fw_indir) {
7774 complete_trunc_indir(freework);
7777 freeblks = freework->fw_freeblks;
7779 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0;
7780 bsize = lfragtosize(fs, freework->fw_frags);
7783 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives
7784 * on the indirblk hashtable and prevents premature freeing.
7786 freework->fw_state |= DEPCOMPLETE;
7788 * SUJ needs to wait for the segment referencing freed indirect
7789 * blocks to expire so that we know the checker will not confuse
7790 * a re-allocated indirect block with its old contents.
7792 if (needj && freework->fw_lbn <= -UFS_NDADDR)
7793 indirblk_insert(freework);
7795 * If we are canceling an existing jnewblk pass it to the free
7796 * routine, otherwise pass the freeblk which will ultimately
7797 * release the freeblks. If we're not journaling, we can just
7798 * free the freeblks immediately.
7800 jnewblk = freework->fw_jnewblk;
7801 if (jnewblk != NULL) {
7802 cancel_jnewblk(jnewblk, &wkhd);
7805 freework->fw_state |= DELAYEDFREE;
7806 freeblks->fb_cgwait++;
7807 WORKLIST_INSERT(&wkhd, &freework->fw_list);
7810 freeblks_free(ump, freeblks, btodb(bsize));
7812 "freework_freeblock: ino %jd blkno %jd lbn %jd size %d",
7813 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize);
7814 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize,
7815 freeblks->fb_inum, freeblks->fb_vtype, &wkhd, key);
7818 * The jnewblk will be discarded and the bits in the map never
7819 * made it to disk. We can immediately free the freeblk.
7822 handle_written_freework(freework);
7826 * We enqueue freework items that need processing back on the freeblks and
7827 * add the freeblks to the worklist. This makes it easier to find all work
7828 * required to flush a truncation in process_truncates().
7831 freework_enqueue(freework)
7832 struct freework *freework;
7834 struct freeblks *freeblks;
7836 freeblks = freework->fw_freeblks;
7837 if ((freework->fw_state & INPROGRESS) == 0)
7838 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
7839 if ((freeblks->fb_state &
7840 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE &&
7841 LIST_EMPTY(&freeblks->fb_jblkdephd))
7842 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7846 * Start, continue, or finish the process of freeing an indirect block tree.
7847 * The free operation may be paused at any point with fw_off containing the
7848 * offset to restart from. This enables us to implement some flow control
7849 * for large truncates which may fan out and generate a huge number of
7853 handle_workitem_indirblk(freework)
7854 struct freework *freework;
7856 struct freeblks *freeblks;
7857 struct ufsmount *ump;
7860 freeblks = freework->fw_freeblks;
7861 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7863 if (freework->fw_state & DEPCOMPLETE) {
7864 handle_written_freework(freework);
7867 if (freework->fw_off == NINDIR(fs)) {
7868 freework_freeblock(freework, SINGLETON_KEY);
7871 freework->fw_state |= INPROGRESS;
7873 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
7879 * Called when a freework structure attached to a cg buf is written. The
7880 * ref on either the parent or the freeblks structure is released and
7881 * the freeblks is added back to the worklist if there is more work to do.
7884 handle_written_freework(freework)
7885 struct freework *freework;
7887 struct freeblks *freeblks;
7888 struct freework *parent;
7890 freeblks = freework->fw_freeblks;
7891 parent = freework->fw_parent;
7892 if (freework->fw_state & DELAYEDFREE)
7893 freeblks->fb_cgwait--;
7894 freework->fw_state |= COMPLETE;
7895 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
7896 WORKITEM_FREE(freework, D_FREEWORK);
7898 if (--parent->fw_ref == 0)
7899 freework_enqueue(parent);
7902 if (--freeblks->fb_ref != 0)
7904 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) ==
7905 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd))
7906 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7910 * This workitem routine performs the block de-allocation.
7911 * The workitem is added to the pending list after the updated
7912 * inode block has been written to disk. As mentioned above,
7913 * checks regarding the number of blocks de-allocated (compared
7914 * to the number of blocks allocated for the file) are also
7915 * performed in this function.
7918 handle_workitem_freeblocks(freeblks, flags)
7919 struct freeblks *freeblks;
7922 struct freework *freework;
7923 struct newblk *newblk;
7924 struct allocindir *aip;
7925 struct ufsmount *ump;
7926 struct worklist *wk;
7929 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd),
7930 ("handle_workitem_freeblocks: Journal entries not written."));
7931 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7932 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum);
7934 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
7935 WORKLIST_REMOVE(wk);
7936 switch (wk->wk_type) {
7938 wk->wk_state |= COMPLETE;
7939 add_to_worklist(wk, 0);
7943 free_newblk(WK_NEWBLK(wk));
7947 aip = WK_ALLOCINDIR(wk);
7949 if (aip->ai_state & DELAYEDFREE) {
7951 freework = newfreework(ump, freeblks, NULL,
7952 aip->ai_lbn, aip->ai_newblkno,
7953 ump->um_fs->fs_frag, 0, 0);
7956 newblk = WK_NEWBLK(wk);
7957 if (newblk->nb_jnewblk) {
7958 freework->fw_jnewblk = newblk->nb_jnewblk;
7959 newblk->nb_jnewblk->jn_dep = &freework->fw_list;
7960 newblk->nb_jnewblk = NULL;
7962 free_newblk(newblk);
7966 freework = WK_FREEWORK(wk);
7967 if (freework->fw_lbn <= -UFS_NDADDR)
7968 handle_workitem_indirblk(freework);
7970 freework_freeblock(freework, key);
7973 panic("handle_workitem_freeblocks: Unknown type %s",
7974 TYPENAME(wk->wk_type));
7977 if (freeblks->fb_ref != 0) {
7978 freeblks->fb_state &= ~INPROGRESS;
7979 wake_worklist(&freeblks->fb_list);
7983 ffs_blkrelease_finish(ump, key);
7985 return handle_complete_freeblocks(freeblks, flags);
7990 * Handle completion of block free via truncate. This allows fs_pending
7991 * to track the actual free block count more closely than if we only updated
7992 * it at the end. We must be careful to handle cases where the block count
7993 * on free was incorrect.
7996 freeblks_free(ump, freeblks, blocks)
7997 struct ufsmount *ump;
7998 struct freeblks *freeblks;
8002 ufs2_daddr_t remain;
8005 remain = -freeblks->fb_chkcnt;
8006 freeblks->fb_chkcnt += blocks;
8008 if (remain < blocks)
8011 fs->fs_pendingblocks -= blocks;
8017 * Once all of the freework workitems are complete we can retire the
8018 * freeblocks dependency and any journal work awaiting completion. This
8019 * can not be called until all other dependencies are stable on disk.
8022 handle_complete_freeblocks(freeblks, flags)
8023 struct freeblks *freeblks;
8026 struct inodedep *inodedep;
8030 struct ufsmount *ump;
8033 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
8035 flags = LK_EXCLUSIVE | flags;
8036 spare = freeblks->fb_chkcnt;
8039 * If we did not release the expected number of blocks we may have
8040 * to adjust the inode block count here. Only do so if it wasn't
8041 * a truncation to zero and the modrev still matches.
8043 if (spare && freeblks->fb_len != 0) {
8044 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum,
8045 flags, &vp, FFSV_FORCEINSMQ) != 0)
8048 if (DIP(ip, i_modrev) == freeblks->fb_modrev) {
8049 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare);
8050 ip->i_flag |= IN_CHANGE;
8052 * We must wait so this happens before the
8053 * journal is reclaimed.
8061 fs->fs_pendingblocks += spare;
8067 quotaadj(freeblks->fb_quota, ump, -spare);
8068 quotarele(freeblks->fb_quota);
8071 if (freeblks->fb_state & ONDEPLIST) {
8072 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum,
8074 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next);
8075 freeblks->fb_state &= ~ONDEPLIST;
8076 if (TAILQ_EMPTY(&inodedep->id_freeblklst))
8077 free_inodedep(inodedep);
8080 * All of the freeblock deps must be complete prior to this call
8081 * so it's now safe to complete earlier outstanding journal entries.
8083 handle_jwork(&freeblks->fb_jwork);
8084 WORKITEM_FREE(freeblks, D_FREEBLKS);
8090 * Release blocks associated with the freeblks and stored in the indirect
8091 * block dbn. If level is greater than SINGLE, the block is an indirect block
8092 * and recursive calls to indirtrunc must be used to cleanse other indirect
8095 * This handles partial and complete truncation of blocks. Partial is noted
8096 * with goingaway == 0. In this case the freework is completed after the
8097 * zero'd indirects are written to disk. For full truncation the freework
8098 * is completed after the block is freed.
8101 indir_trunc(freework, dbn, lbn)
8102 struct freework *freework;
8106 struct freework *nfreework;
8107 struct workhead wkhd;
8108 struct freeblks *freeblks;
8111 struct indirdep *indirdep;
8113 struct ufsmount *ump;
8115 ufs2_daddr_t nb, nnb, *bap2;
8116 ufs_lbn_t lbnadd, nlbn;
8118 int nblocks, ufs1fmt, freedblocks;
8119 int goingaway, freedeps, needj, level, cnt, i;
8121 freeblks = freework->fw_freeblks;
8122 mp = freeblks->fb_list.wk_mp;
8126 * Get buffer of block pointers to be freed. There are three cases:
8128 * 1) Partial truncate caches the indirdep pointer in the freework
8129 * which provides us a back copy to the save bp which holds the
8130 * pointers we want to clear. When this completes the zero
8131 * pointers are written to the real copy.
8132 * 2) The indirect is being completely truncated, cancel_indirdep()
8133 * eliminated the real copy and placed the indirdep on the saved
8134 * copy. The indirdep and buf are discarded when this completes.
8135 * 3) The indirect was not in memory, we read a copy off of the disk
8136 * using the devvp and drop and invalidate the buffer when we're
8141 if (freework->fw_indir != NULL) {
8143 indirdep = freework->fw_indir;
8144 bp = indirdep->ir_savebp;
8145 if (bp == NULL || bp->b_blkno != dbn)
8146 panic("indir_trunc: Bad saved buf %p blkno %jd",
8148 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) {
8150 * The lock prevents the buf dep list from changing and
8151 * indirects on devvp should only ever have one dependency.
8153 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep));
8154 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0)
8155 panic("indir_trunc: Bad indirdep %p from buf %p",
8157 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
8158 NOCRED, &bp) != 0) {
8163 /* Protects against a race with complete_trunc_indir(). */
8164 freework->fw_state &= ~INPROGRESS;
8166 * If we have an indirdep we need to enforce the truncation order
8167 * and discard it when it is complete.
8170 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) &&
8171 !TAILQ_EMPTY(&indirdep->ir_trunc)) {
8173 * Add the complete truncate to the list on the
8174 * indirdep to enforce in-order processing.
8176 if (freework->fw_indir == NULL)
8177 TAILQ_INSERT_TAIL(&indirdep->ir_trunc,
8183 * If we're goingaway, free the indirdep. Otherwise it will
8184 * linger until the write completes.
8187 free_indirdep(indirdep);
8190 /* Initialize pointers depending on block size. */
8191 if (ump->um_fstype == UFS1) {
8192 bap1 = (ufs1_daddr_t *)bp->b_data;
8193 nb = bap1[freework->fw_off];
8197 bap2 = (ufs2_daddr_t *)bp->b_data;
8198 nb = bap2[freework->fw_off];
8202 level = lbn_level(lbn);
8203 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0;
8204 lbnadd = lbn_offset(fs, level);
8205 nblocks = btodb(fs->fs_bsize);
8206 nfreework = freework;
8210 * Reclaim blocks. Traverses into nested indirect levels and
8211 * arranges for the current level to be freed when subordinates
8212 * are free when journaling.
8214 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum);
8215 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
8216 if (UFS_CHECK_BLKNO(mp, freeblks->fb_inum, nb,
8219 if (i != NINDIR(fs) - 1) {
8230 nlbn = (lbn + 1) - (i * lbnadd);
8232 nfreework = newfreework(ump, freeblks, freework,
8233 nlbn, nb, fs->fs_frag, 0, 0);
8236 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
8238 struct freedep *freedep;
8241 * Attempt to aggregate freedep dependencies for
8242 * all blocks being released to the same CG.
8246 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
8247 freedep = newfreedep(freework);
8248 WORKLIST_INSERT_UNLOCKED(&wkhd,
8253 "indir_trunc: ino %jd blkno %jd size %d",
8254 freeblks->fb_inum, nb, fs->fs_bsize);
8255 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
8256 fs->fs_bsize, freeblks->fb_inum,
8257 freeblks->fb_vtype, &wkhd, key);
8260 ffs_blkrelease_finish(ump, key);
8262 bp->b_flags |= B_INVAL | B_NOCACHE;
8267 freedblocks = (nblocks * cnt);
8269 freedblocks += nblocks;
8270 freeblks_free(ump, freeblks, freedblocks);
8272 * If we are journaling set up the ref counts and offset so this
8273 * indirect can be completed when its children are free.
8277 freework->fw_off = i;
8278 freework->fw_ref += freedeps;
8279 freework->fw_ref -= NINDIR(fs) + 1;
8281 freeblks->fb_cgwait += freedeps;
8282 if (freework->fw_ref == 0)
8283 freework_freeblock(freework, SINGLETON_KEY);
8288 * If we're not journaling we can free the indirect now.
8290 dbn = dbtofsb(fs, dbn);
8292 "indir_trunc 2: ino %jd blkno %jd size %d",
8293 freeblks->fb_inum, dbn, fs->fs_bsize);
8294 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
8295 freeblks->fb_inum, freeblks->fb_vtype, NULL, SINGLETON_KEY);
8296 /* Non SUJ softdep does single-threaded truncations. */
8297 if (freework->fw_blkno == dbn) {
8298 freework->fw_state |= ALLCOMPLETE;
8300 handle_written_freework(freework);
8307 * Cancel an allocindir when it is removed via truncation. When bp is not
8308 * NULL the indirect never appeared on disk and is scheduled to be freed
8309 * independently of the indir so we can more easily track journal work.
8312 cancel_allocindir(aip, bp, freeblks, trunc)
8313 struct allocindir *aip;
8315 struct freeblks *freeblks;
8318 struct indirdep *indirdep;
8319 struct freefrag *freefrag;
8320 struct newblk *newblk;
8322 newblk = (struct newblk *)aip;
8323 LIST_REMOVE(aip, ai_next);
8325 * We must eliminate the pointer in bp if it must be freed on its
8326 * own due to partial truncate or pending journal work.
8328 if (bp && (trunc || newblk->nb_jnewblk)) {
8330 * Clear the pointer and mark the aip to be freed
8331 * directly if it never existed on disk.
8333 aip->ai_state |= DELAYEDFREE;
8334 indirdep = aip->ai_indirdep;
8335 if (indirdep->ir_state & UFS1FMT)
8336 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8338 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8341 * When truncating the previous pointer will be freed via
8342 * savedbp. Eliminate the freefrag which would dup free.
8344 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) {
8345 newblk->nb_freefrag = NULL;
8346 if (freefrag->ff_jdep)
8348 WK_JFREEFRAG(freefrag->ff_jdep));
8349 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork);
8350 WORKITEM_FREE(freefrag, D_FREEFRAG);
8353 * If the journal hasn't been written the jnewblk must be passed
8354 * to the call to ffs_blkfree that reclaims the space. We accomplish
8355 * this by leaving the journal dependency on the newblk to be freed
8356 * when a freework is created in handle_workitem_freeblocks().
8358 cancel_newblk(newblk, NULL, &freeblks->fb_jwork);
8359 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
8363 * Create the mkdir dependencies for . and .. in a new directory. Link them
8364 * in to a newdirblk so any subsequent additions are tracked properly. The
8365 * caller is responsible for adding the mkdir1 dependency to the journal
8366 * and updating id_mkdiradd. This function returns with the per-filesystem
8369 static struct mkdir *
8370 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
8374 struct buf *newdirbp;
8375 struct mkdir **mkdirp;
8377 struct newblk *newblk;
8378 struct pagedep *pagedep;
8379 struct inodedep *inodedep;
8380 struct newdirblk *newdirblk;
8381 struct mkdir *mkdir1, *mkdir2;
8382 struct worklist *wk;
8383 struct jaddref *jaddref;
8384 struct ufsmount *ump;
8387 mp = dap->da_list.wk_mp;
8389 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
8391 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8392 LIST_INIT(&newdirblk->db_mkdir);
8393 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8394 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
8395 mkdir1->md_state = ATTACHED | MKDIR_BODY;
8396 mkdir1->md_diradd = dap;
8397 mkdir1->md_jaddref = NULL;
8398 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8399 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
8400 mkdir2->md_state = ATTACHED | MKDIR_PARENT;
8401 mkdir2->md_diradd = dap;
8402 mkdir2->md_jaddref = NULL;
8403 if (MOUNTEDSUJ(mp) == 0) {
8404 mkdir1->md_state |= DEPCOMPLETE;
8405 mkdir2->md_state |= DEPCOMPLETE;
8408 * Dependency on "." and ".." being written to disk.
8410 mkdir1->md_buf = newdirbp;
8411 ACQUIRE_LOCK(VFSTOUFS(mp));
8412 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs);
8414 * We must link the pagedep, allocdirect, and newdirblk for
8415 * the initial file page so the pointer to the new directory
8416 * is not written until the directory contents are live and
8417 * any subsequent additions are not marked live until the
8418 * block is reachable via the inode.
8420 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0)
8421 panic("setup_newdir: lost pagedep");
8422 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
8423 if (wk->wk_type == D_ALLOCDIRECT)
8426 panic("setup_newdir: lost allocdirect");
8427 if (pagedep->pd_state & NEWBLOCK)
8428 panic("setup_newdir: NEWBLOCK already set");
8429 newblk = WK_NEWBLK(wk);
8430 pagedep->pd_state |= NEWBLOCK;
8431 pagedep->pd_newdirblk = newdirblk;
8432 newdirblk->db_pagedep = pagedep;
8433 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8434 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
8436 * Look up the inodedep for the parent directory so that we
8437 * can link mkdir2 into the pending dotdot jaddref or
8438 * the inode write if there is none. If the inode is
8439 * ALLCOMPLETE and no jaddref is present all dependencies have
8440 * been satisfied and mkdir2 can be freed.
8442 inodedep_lookup(mp, dinum, 0, &inodedep);
8443 if (MOUNTEDSUJ(mp)) {
8444 if (inodedep == NULL)
8445 panic("setup_newdir: Lost parent.");
8446 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8448 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
8449 (jaddref->ja_state & MKDIR_PARENT),
8450 ("setup_newdir: bad dotdot jaddref %p", jaddref));
8451 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8452 mkdir2->md_jaddref = jaddref;
8453 jaddref->ja_mkdir = mkdir2;
8454 } else if (inodedep == NULL ||
8455 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
8456 dap->da_state &= ~MKDIR_PARENT;
8457 WORKITEM_FREE(mkdir2, D_MKDIR);
8460 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8461 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list);
8469 * Directory entry addition dependencies.
8471 * When adding a new directory entry, the inode (with its incremented link
8472 * count) must be written to disk before the directory entry's pointer to it.
8473 * Also, if the inode is newly allocated, the corresponding freemap must be
8474 * updated (on disk) before the directory entry's pointer. These requirements
8475 * are met via undo/redo on the directory entry's pointer, which consists
8476 * simply of the inode number.
8478 * As directory entries are added and deleted, the free space within a
8479 * directory block can become fragmented. The ufs filesystem will compact
8480 * a fragmented directory block to make space for a new entry. When this
8481 * occurs, the offsets of previously added entries change. Any "diradd"
8482 * dependency structures corresponding to these entries must be updated with
8487 * This routine is called after the in-memory inode's link
8488 * count has been incremented, but before the directory entry's
8489 * pointer to the inode has been set.
8492 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
8493 struct buf *bp; /* buffer containing directory block */
8494 struct inode *dp; /* inode for directory */
8495 off_t diroffset; /* offset of new entry in directory */
8496 ino_t newinum; /* inode referenced by new directory entry */
8497 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
8498 int isnewblk; /* entry is in a newly allocated block */
8500 int offset; /* offset of new entry within directory block */
8501 ufs_lbn_t lbn; /* block in directory containing new entry */
8504 struct newblk *newblk;
8505 struct pagedep *pagedep;
8506 struct inodedep *inodedep;
8507 struct newdirblk *newdirblk;
8508 struct mkdir *mkdir1, *mkdir2;
8509 struct jaddref *jaddref;
8510 struct ufsmount *ump;
8516 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8517 ("softdep_setup_directory_add called on non-softdep filesystem"));
8519 * Whiteouts have no dependencies.
8521 if (newinum == UFS_WINO) {
8522 if (newdirbp != NULL)
8527 mkdir1 = mkdir2 = NULL;
8529 lbn = lblkno(fs, diroffset);
8530 offset = blkoff(fs, diroffset);
8531 dap = malloc(sizeof(struct diradd), M_DIRADD,
8532 M_SOFTDEP_FLAGS|M_ZERO);
8533 workitem_alloc(&dap->da_list, D_DIRADD, mp);
8534 dap->da_offset = offset;
8535 dap->da_newinum = newinum;
8536 dap->da_state = ATTACHED;
8537 LIST_INIT(&dap->da_jwork);
8538 isindir = bp->b_lblkno >= UFS_NDADDR;
8541 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
8542 newdirblk = malloc(sizeof(struct newdirblk),
8543 M_NEWDIRBLK, M_SOFTDEP_FLAGS);
8544 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8545 LIST_INIT(&newdirblk->db_mkdir);
8548 * If we're creating a new directory setup the dependencies and set
8549 * the dap state to wait for them. Otherwise it's COMPLETE and
8552 if (newdirbp == NULL) {
8553 dap->da_state |= DEPCOMPLETE;
8556 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
8557 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
8561 * Link into parent directory pagedep to await its being written.
8563 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep);
8565 if (diradd_lookup(pagedep, offset) != NULL)
8566 panic("softdep_setup_directory_add: %p already at off %d\n",
8567 diradd_lookup(pagedep, offset), offset);
8569 dap->da_pagedep = pagedep;
8570 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
8572 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
8574 * If we're journaling, link the diradd into the jaddref so it
8575 * may be completed after the journal entry is written. Otherwise,
8576 * link the diradd into its inodedep. If the inode is not yet
8577 * written place it on the bufwait list, otherwise do the post-inode
8578 * write processing to put it on the id_pendinghd list.
8580 if (MOUNTEDSUJ(mp)) {
8581 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8583 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
8584 ("softdep_setup_directory_add: bad jaddref %p", jaddref));
8585 jaddref->ja_diroff = diroffset;
8586 jaddref->ja_diradd = dap;
8587 add_to_journal(&jaddref->ja_list);
8588 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
8589 diradd_inode_written(dap, inodedep);
8591 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
8593 * Add the journal entries for . and .. links now that the primary
8596 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) {
8597 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
8598 inoreflst, if_deps);
8599 KASSERT(jaddref != NULL &&
8600 jaddref->ja_ino == jaddref->ja_parent &&
8601 (jaddref->ja_state & MKDIR_BODY),
8602 ("softdep_setup_directory_add: bad dot jaddref %p",
8604 mkdir1->md_jaddref = jaddref;
8605 jaddref->ja_mkdir = mkdir1;
8607 * It is important that the dotdot journal entry
8608 * is added prior to the dot entry since dot writes
8609 * both the dot and dotdot links. These both must
8610 * be added after the primary link for the journal
8611 * to remain consistent.
8613 add_to_journal(&mkdir2->md_jaddref->ja_list);
8614 add_to_journal(&jaddref->ja_list);
8617 * If we are adding a new directory remember this diradd so that if
8618 * we rename it we can keep the dot and dotdot dependencies. If
8619 * we are adding a new name for an inode that has a mkdiradd we
8620 * must be in rename and we have to move the dot and dotdot
8621 * dependencies to this new name. The old name is being orphaned
8624 if (mkdir1 != NULL) {
8625 if (inodedep->id_mkdiradd != NULL)
8626 panic("softdep_setup_directory_add: Existing mkdir");
8627 inodedep->id_mkdiradd = dap;
8628 } else if (inodedep->id_mkdiradd)
8629 merge_diradd(inodedep, dap);
8630 if (newdirblk != NULL) {
8632 * There is nothing to do if we are already tracking
8635 if ((pagedep->pd_state & NEWBLOCK) != 0) {
8636 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
8640 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
8642 panic("softdep_setup_directory_add: lost entry");
8643 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8644 pagedep->pd_state |= NEWBLOCK;
8645 pagedep->pd_newdirblk = newdirblk;
8646 newdirblk->db_pagedep = pagedep;
8649 * If we extended into an indirect signal direnter to sync.
8660 * This procedure is called to change the offset of a directory
8661 * entry when compacting a directory block which must be owned
8662 * exclusively by the caller. Note that the actual entry movement
8663 * must be done in this procedure to ensure that no I/O completions
8664 * occur while the move is in progress.
8667 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
8668 struct buf *bp; /* Buffer holding directory block. */
8669 struct inode *dp; /* inode for directory */
8670 caddr_t base; /* address of dp->i_offset */
8671 caddr_t oldloc; /* address of old directory location */
8672 caddr_t newloc; /* address of new directory location */
8673 int entrysize; /* size of directory entry */
8675 int offset, oldoffset, newoffset;
8676 struct pagedep *pagedep;
8677 struct jmvref *jmvref;
8681 struct ufsmount *ump;
8687 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8688 ("softdep_change_directoryentry_offset called on "
8689 "non-softdep filesystem"));
8690 de = (struct direct *)oldloc;
8694 * Moves are always journaled as it would be too complex to
8695 * determine if any affected adds or removes are present in the
8698 if (MOUNTEDSUJ(mp)) {
8700 jmvref = newjmvref(dp, de->d_ino,
8701 dp->i_offset + (oldloc - base),
8702 dp->i_offset + (newloc - base));
8704 lbn = lblkno(ump->um_fs, dp->i_offset);
8705 offset = blkoff(ump->um_fs, dp->i_offset);
8706 oldoffset = offset + (oldloc - base);
8707 newoffset = offset + (newloc - base);
8709 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0)
8711 dap = diradd_lookup(pagedep, oldoffset);
8713 dap->da_offset = newoffset;
8714 newoffset = DIRADDHASH(newoffset);
8715 oldoffset = DIRADDHASH(oldoffset);
8716 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
8717 newoffset != oldoffset) {
8718 LIST_REMOVE(dap, da_pdlist);
8719 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
8725 jmvref->jm_pagedep = pagedep;
8726 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
8727 add_to_journal(&jmvref->jm_list);
8729 bcopy(oldloc, newloc, entrysize);
8734 * Move the mkdir dependencies and journal work from one diradd to another
8735 * when renaming a directory. The new name must depend on the mkdir deps
8736 * completing as the old name did. Directories can only have one valid link
8737 * at a time so one must be canonical.
8740 merge_diradd(inodedep, newdap)
8741 struct inodedep *inodedep;
8742 struct diradd *newdap;
8744 struct diradd *olddap;
8745 struct mkdir *mkdir, *nextmd;
8746 struct ufsmount *ump;
8749 olddap = inodedep->id_mkdiradd;
8750 inodedep->id_mkdiradd = newdap;
8751 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8752 newdap->da_state &= ~DEPCOMPLETE;
8753 ump = VFSTOUFS(inodedep->id_list.wk_mp);
8754 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8756 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8757 if (mkdir->md_diradd != olddap)
8759 mkdir->md_diradd = newdap;
8760 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
8761 newdap->da_state |= state;
8762 olddap->da_state &= ~state;
8763 if ((olddap->da_state &
8764 (MKDIR_PARENT | MKDIR_BODY)) == 0)
8767 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8768 panic("merge_diradd: unfound ref");
8771 * Any mkdir related journal items are not safe to be freed until
8772 * the new name is stable.
8774 jwork_move(&newdap->da_jwork, &olddap->da_jwork);
8775 olddap->da_state |= DEPCOMPLETE;
8776 complete_diradd(olddap);
8780 * Move the diradd to the pending list when all diradd dependencies are
8784 complete_diradd(dap)
8787 struct pagedep *pagedep;
8789 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
8790 if (dap->da_state & DIRCHG)
8791 pagedep = dap->da_previous->dm_pagedep;
8793 pagedep = dap->da_pagedep;
8794 LIST_REMOVE(dap, da_pdlist);
8795 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
8800 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal
8801 * add entries and conditonally journal the remove.
8804 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
8806 struct dirrem *dirrem;
8807 struct jremref *jremref;
8808 struct jremref *dotremref;
8809 struct jremref *dotdotremref;
8811 struct inodedep *inodedep;
8812 struct jaddref *jaddref;
8813 struct inoref *inoref;
8814 struct ufsmount *ump;
8815 struct mkdir *mkdir;
8818 * If no remove references were allocated we're on a non-journaled
8819 * filesystem and can skip the cancel step.
8821 if (jremref == NULL) {
8822 free_diradd(dap, NULL);
8826 * Cancel the primary name an free it if it does not require
8829 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
8830 0, &inodedep) != 0) {
8831 /* Abort the addref that reference this diradd. */
8832 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
8833 if (inoref->if_list.wk_type != D_JADDREF)
8835 jaddref = (struct jaddref *)inoref;
8836 if (jaddref->ja_diradd != dap)
8838 if (cancel_jaddref(jaddref, inodedep,
8839 &dirrem->dm_jwork) == 0) {
8840 free_jremref(jremref);
8847 * Cancel subordinate names and free them if they do not require
8850 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8851 ump = VFSTOUFS(dap->da_list.wk_mp);
8852 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) {
8853 if (mkdir->md_diradd != dap)
8855 if ((jaddref = mkdir->md_jaddref) == NULL)
8857 mkdir->md_jaddref = NULL;
8858 if (mkdir->md_state & MKDIR_PARENT) {
8859 if (cancel_jaddref(jaddref, NULL,
8860 &dirrem->dm_jwork) == 0) {
8861 free_jremref(dotdotremref);
8862 dotdotremref = NULL;
8865 if (cancel_jaddref(jaddref, inodedep,
8866 &dirrem->dm_jwork) == 0) {
8867 free_jremref(dotremref);
8875 journal_jremref(dirrem, jremref, inodedep);
8877 journal_jremref(dirrem, dotremref, inodedep);
8879 journal_jremref(dirrem, dotdotremref, NULL);
8880 jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
8881 free_diradd(dap, &dirrem->dm_jwork);
8885 * Free a diradd dependency structure.
8888 free_diradd(dap, wkhd)
8890 struct workhead *wkhd;
8892 struct dirrem *dirrem;
8893 struct pagedep *pagedep;
8894 struct inodedep *inodedep;
8895 struct mkdir *mkdir, *nextmd;
8896 struct ufsmount *ump;
8898 ump = VFSTOUFS(dap->da_list.wk_mp);
8900 LIST_REMOVE(dap, da_pdlist);
8901 if (dap->da_state & ONWORKLIST)
8902 WORKLIST_REMOVE(&dap->da_list);
8903 if ((dap->da_state & DIRCHG) == 0) {
8904 pagedep = dap->da_pagedep;
8906 dirrem = dap->da_previous;
8907 pagedep = dirrem->dm_pagedep;
8908 dirrem->dm_dirinum = pagedep->pd_ino;
8909 dirrem->dm_state |= COMPLETE;
8910 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
8911 add_to_worklist(&dirrem->dm_list, 0);
8913 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
8915 if (inodedep->id_mkdiradd == dap)
8916 inodedep->id_mkdiradd = NULL;
8917 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8918 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8920 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8921 if (mkdir->md_diradd != dap)
8924 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
8925 LIST_REMOVE(mkdir, md_mkdirs);
8926 if (mkdir->md_state & ONWORKLIST)
8927 WORKLIST_REMOVE(&mkdir->md_list);
8928 if (mkdir->md_jaddref != NULL)
8929 panic("free_diradd: Unexpected jaddref");
8930 WORKITEM_FREE(mkdir, D_MKDIR);
8931 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
8934 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8935 panic("free_diradd: unfound ref");
8938 free_inodedep(inodedep);
8940 * Free any journal segments waiting for the directory write.
8942 handle_jwork(&dap->da_jwork);
8943 WORKITEM_FREE(dap, D_DIRADD);
8947 * Directory entry removal dependencies.
8949 * When removing a directory entry, the entry's inode pointer must be
8950 * zero'ed on disk before the corresponding inode's link count is decremented
8951 * (possibly freeing the inode for re-use). This dependency is handled by
8952 * updating the directory entry but delaying the inode count reduction until
8953 * after the directory block has been written to disk. After this point, the
8954 * inode count can be decremented whenever it is convenient.
8958 * This routine should be called immediately after removing
8959 * a directory entry. The inode's link count should not be
8960 * decremented by the calling procedure -- the soft updates
8961 * code will do this task when it is safe.
8964 softdep_setup_remove(bp, dp, ip, isrmdir)
8965 struct buf *bp; /* buffer containing directory block */
8966 struct inode *dp; /* inode for the directory being modified */
8967 struct inode *ip; /* inode for directory entry being removed */
8968 int isrmdir; /* indicates if doing RMDIR */
8970 struct dirrem *dirrem, *prevdirrem;
8971 struct inodedep *inodedep;
8972 struct ufsmount *ump;
8976 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
8977 ("softdep_setup_remove called on non-softdep filesystem"));
8979 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want
8980 * newdirrem() to setup the full directory remove which requires
8983 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
8985 * Add the dirrem to the inodedep's pending remove list for quick
8988 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0)
8989 panic("softdep_setup_remove: Lost inodedep.");
8990 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
8991 dirrem->dm_state |= ONDEPLIST;
8992 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
8995 * If the COMPLETE flag is clear, then there were no active
8996 * entries and we want to roll back to a zeroed entry until
8997 * the new inode is committed to disk. If the COMPLETE flag is
8998 * set then we have deleted an entry that never made it to
8999 * disk. If the entry we deleted resulted from a name change,
9000 * then the old name still resides on disk. We cannot delete
9001 * its inode (returned to us in prevdirrem) until the zeroed
9002 * directory entry gets to disk. The new inode has never been
9003 * referenced on the disk, so can be deleted immediately.
9005 if ((dirrem->dm_state & COMPLETE) == 0) {
9006 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
9010 if (prevdirrem != NULL)
9011 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
9012 prevdirrem, dm_next);
9013 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
9014 direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
9017 handle_workitem_remove(dirrem, 0);
9022 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
9023 * pd_pendinghd list of a pagedep.
9025 static struct diradd *
9026 diradd_lookup(pagedep, offset)
9027 struct pagedep *pagedep;
9032 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
9033 if (dap->da_offset == offset)
9035 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
9036 if (dap->da_offset == offset)
9042 * Search for a .. diradd dependency in a directory that is being removed.
9043 * If the directory was renamed to a new parent we have a diradd rather
9044 * than a mkdir for the .. entry. We need to cancel it now before
9045 * it is found in truncate().
9047 static struct jremref *
9048 cancel_diradd_dotdot(ip, dirrem, jremref)
9050 struct dirrem *dirrem;
9051 struct jremref *jremref;
9053 struct pagedep *pagedep;
9055 struct worklist *wk;
9057 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0)
9059 dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
9062 cancel_diradd(dap, dirrem, jremref, NULL, NULL);
9064 * Mark any journal work as belonging to the parent so it is freed
9065 * with the .. reference.
9067 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
9068 wk->wk_state |= MKDIR_PARENT;
9073 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
9074 * replace it with a dirrem/diradd pair as a result of re-parenting a
9075 * directory. This ensures that we don't simultaneously have a mkdir and
9076 * a diradd for the same .. entry.
9078 static struct jremref *
9079 cancel_mkdir_dotdot(ip, dirrem, jremref)
9081 struct dirrem *dirrem;
9082 struct jremref *jremref;
9084 struct inodedep *inodedep;
9085 struct jaddref *jaddref;
9086 struct ufsmount *ump;
9087 struct mkdir *mkdir;
9092 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
9094 dap = inodedep->id_mkdiradd;
9095 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
9097 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9098 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
9099 mkdir = LIST_NEXT(mkdir, md_mkdirs))
9100 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
9103 panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
9104 if ((jaddref = mkdir->md_jaddref) != NULL) {
9105 mkdir->md_jaddref = NULL;
9106 jaddref->ja_state &= ~MKDIR_PARENT;
9107 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0)
9108 panic("cancel_mkdir_dotdot: Lost parent inodedep");
9109 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
9110 journal_jremref(dirrem, jremref, inodedep);
9114 if (mkdir->md_state & ONWORKLIST)
9115 WORKLIST_REMOVE(&mkdir->md_list);
9116 mkdir->md_state |= ALLCOMPLETE;
9117 complete_mkdir(mkdir);
9122 journal_jremref(dirrem, jremref, inodedep)
9123 struct dirrem *dirrem;
9124 struct jremref *jremref;
9125 struct inodedep *inodedep;
9128 if (inodedep == NULL)
9129 if (inodedep_lookup(jremref->jr_list.wk_mp,
9130 jremref->jr_ref.if_ino, 0, &inodedep) == 0)
9131 panic("journal_jremref: Lost inodedep");
9132 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
9133 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
9134 add_to_journal(&jremref->jr_list);
9138 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
9139 struct dirrem *dirrem;
9140 struct jremref *jremref;
9141 struct jremref *dotremref;
9142 struct jremref *dotdotremref;
9144 struct inodedep *inodedep;
9147 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
9149 panic("dirrem_journal: Lost inodedep");
9150 journal_jremref(dirrem, jremref, inodedep);
9152 journal_jremref(dirrem, dotremref, inodedep);
9154 journal_jremref(dirrem, dotdotremref, NULL);
9158 * Allocate a new dirrem if appropriate and return it along with
9159 * its associated pagedep. Called without a lock, returns with lock.
9161 static struct dirrem *
9162 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
9163 struct buf *bp; /* buffer containing directory block */
9164 struct inode *dp; /* inode for the directory being modified */
9165 struct inode *ip; /* inode for directory entry being removed */
9166 int isrmdir; /* indicates if doing RMDIR */
9167 struct dirrem **prevdirremp; /* previously referenced inode, if any */
9172 struct dirrem *dirrem;
9173 struct pagedep *pagedep;
9174 struct jremref *jremref;
9175 struct jremref *dotremref;
9176 struct jremref *dotdotremref;
9178 struct ufsmount *ump;
9181 * Whiteouts have no deletion dependencies.
9184 panic("newdirrem: whiteout");
9189 * If the system is over its limit and our filesystem is
9190 * responsible for more than our share of that usage and
9191 * we are not a snapshot, request some inodedep cleanup.
9192 * Limiting the number of dirrem structures will also limit
9193 * the number of freefile and freeblks structures.
9196 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM))
9197 schedule_cleanup(UFSTOVFS(ump));
9200 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS |
9202 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
9203 LIST_INIT(&dirrem->dm_jremrefhd);
9204 LIST_INIT(&dirrem->dm_jwork);
9205 dirrem->dm_state = isrmdir ? RMDIR : 0;
9206 dirrem->dm_oldinum = ip->i_number;
9207 *prevdirremp = NULL;
9209 * Allocate remove reference structures to track journal write
9210 * dependencies. We will always have one for the link and
9211 * when doing directories we will always have one more for dot.
9212 * When renaming a directory we skip the dotdot link change so
9213 * this is not needed.
9215 jremref = dotremref = dotdotremref = NULL;
9216 if (DOINGSUJ(dvp)) {
9218 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9219 ip->i_effnlink + 2);
9220 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
9221 ip->i_effnlink + 1);
9222 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
9223 dp->i_effnlink + 1);
9224 dotdotremref->jr_state |= MKDIR_PARENT;
9226 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9227 ip->i_effnlink + 1);
9230 lbn = lblkno(ump->um_fs, dp->i_offset);
9231 offset = blkoff(ump->um_fs, dp->i_offset);
9232 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC,
9234 dirrem->dm_pagedep = pagedep;
9235 dirrem->dm_offset = offset;
9237 * If we're renaming a .. link to a new directory, cancel any
9238 * existing MKDIR_PARENT mkdir. If it has already been canceled
9239 * the jremref is preserved for any potential diradd in this
9240 * location. This can not coincide with a rmdir.
9242 if (dp->i_offset == DOTDOT_OFFSET) {
9244 panic("newdirrem: .. directory change during remove?");
9245 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
9248 * If we're removing a directory search for the .. dependency now and
9249 * cancel it. Any pending journal work will be added to the dirrem
9250 * to be completed when the workitem remove completes.
9253 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
9255 * Check for a diradd dependency for the same directory entry.
9256 * If present, then both dependencies become obsolete and can
9259 dap = diradd_lookup(pagedep, offset);
9262 * Link the jremref structures into the dirrem so they are
9263 * written prior to the pagedep.
9266 dirrem_journal(dirrem, jremref, dotremref,
9271 * Must be ATTACHED at this point.
9273 if ((dap->da_state & ATTACHED) == 0)
9274 panic("newdirrem: not ATTACHED");
9275 if (dap->da_newinum != ip->i_number)
9276 panic("newdirrem: inum %ju should be %ju",
9277 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum);
9279 * If we are deleting a changed name that never made it to disk,
9280 * then return the dirrem describing the previous inode (which
9281 * represents the inode currently referenced from this entry on disk).
9283 if ((dap->da_state & DIRCHG) != 0) {
9284 *prevdirremp = dap->da_previous;
9285 dap->da_state &= ~DIRCHG;
9286 dap->da_pagedep = pagedep;
9289 * We are deleting an entry that never made it to disk.
9290 * Mark it COMPLETE so we can delete its inode immediately.
9292 dirrem->dm_state |= COMPLETE;
9293 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
9296 struct worklist *wk;
9298 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
9299 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
9300 panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
9308 * Directory entry change dependencies.
9310 * Changing an existing directory entry requires that an add operation
9311 * be completed first followed by a deletion. The semantics for the addition
9312 * are identical to the description of adding a new entry above except
9313 * that the rollback is to the old inode number rather than zero. Once
9314 * the addition dependency is completed, the removal is done as described
9315 * in the removal routine above.
9319 * This routine should be called immediately after changing
9320 * a directory entry. The inode's link count should not be
9321 * decremented by the calling procedure -- the soft updates
9322 * code will perform this task when it is safe.
9325 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
9326 struct buf *bp; /* buffer containing directory block */
9327 struct inode *dp; /* inode for the directory being modified */
9328 struct inode *ip; /* inode for directory entry being removed */
9329 ino_t newinum; /* new inode number for changed entry */
9330 int isrmdir; /* indicates if doing RMDIR */
9333 struct diradd *dap = NULL;
9334 struct dirrem *dirrem, *prevdirrem;
9335 struct pagedep *pagedep;
9336 struct inodedep *inodedep;
9337 struct jaddref *jaddref;
9339 struct ufsmount *ump;
9343 offset = blkoff(ump->um_fs, dp->i_offset);
9344 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
9345 ("softdep_setup_directory_change called on non-softdep filesystem"));
9348 * Whiteouts do not need diradd dependencies.
9350 if (newinum != UFS_WINO) {
9351 dap = malloc(sizeof(struct diradd),
9352 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
9353 workitem_alloc(&dap->da_list, D_DIRADD, mp);
9354 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
9355 dap->da_offset = offset;
9356 dap->da_newinum = newinum;
9357 LIST_INIT(&dap->da_jwork);
9361 * Allocate a new dirrem and ACQUIRE_LOCK.
9363 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
9364 pagedep = dirrem->dm_pagedep;
9366 * The possible values for isrmdir:
9367 * 0 - non-directory file rename
9368 * 1 - directory rename within same directory
9369 * inum - directory rename to new directory of given inode number
9370 * When renaming to a new directory, we are both deleting and
9371 * creating a new directory entry, so the link count on the new
9372 * directory should not change. Thus we do not need the followup
9373 * dirrem which is usually done in handle_workitem_remove. We set
9374 * the DIRCHG flag to tell handle_workitem_remove to skip the
9378 dirrem->dm_state |= DIRCHG;
9381 * Whiteouts have no additional dependencies,
9382 * so just put the dirrem on the correct list.
9384 if (newinum == UFS_WINO) {
9385 if ((dirrem->dm_state & COMPLETE) == 0) {
9386 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
9389 dirrem->dm_dirinum = pagedep->pd_ino;
9390 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9391 add_to_worklist(&dirrem->dm_list, 0);
9397 * Add the dirrem to the inodedep's pending remove list for quick
9398 * discovery later. A valid nlinkdelta ensures that this lookup
9401 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
9402 panic("softdep_setup_directory_change: Lost inodedep.");
9403 dirrem->dm_state |= ONDEPLIST;
9404 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9407 * If the COMPLETE flag is clear, then there were no active
9408 * entries and we want to roll back to the previous inode until
9409 * the new inode is committed to disk. If the COMPLETE flag is
9410 * set, then we have deleted an entry that never made it to disk.
9411 * If the entry we deleted resulted from a name change, then the old
9412 * inode reference still resides on disk. Any rollback that we do
9413 * needs to be to that old inode (returned to us in prevdirrem). If
9414 * the entry we deleted resulted from a create, then there is
9415 * no entry on the disk, so we want to roll back to zero rather
9416 * than the uncommitted inode. In either of the COMPLETE cases we
9417 * want to immediately free the unwritten and unreferenced inode.
9419 if ((dirrem->dm_state & COMPLETE) == 0) {
9420 dap->da_previous = dirrem;
9422 if (prevdirrem != NULL) {
9423 dap->da_previous = prevdirrem;
9425 dap->da_state &= ~DIRCHG;
9426 dap->da_pagedep = pagedep;
9428 dirrem->dm_dirinum = pagedep->pd_ino;
9429 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9430 add_to_worklist(&dirrem->dm_list, 0);
9433 * Lookup the jaddref for this journal entry. We must finish
9434 * initializing it and make the diradd write dependent on it.
9435 * If we're not journaling, put it on the id_bufwait list if the
9436 * inode is not yet written. If it is written, do the post-inode
9437 * write processing to put it on the id_pendinghd list.
9439 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
9440 if (MOUNTEDSUJ(mp)) {
9441 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
9443 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
9444 ("softdep_setup_directory_change: bad jaddref %p",
9446 jaddref->ja_diroff = dp->i_offset;
9447 jaddref->ja_diradd = dap;
9448 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9450 add_to_journal(&jaddref->ja_list);
9451 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
9452 dap->da_state |= COMPLETE;
9453 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
9454 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
9456 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9458 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
9461 * If we're making a new name for a directory that has not been
9462 * committed when need to move the dot and dotdot references to
9465 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
9466 merge_diradd(inodedep, dap);
9471 * Called whenever the link count on an inode is changed.
9472 * It creates an inode dependency so that the new reference(s)
9473 * to the inode cannot be committed to disk until the updated
9474 * inode has been written.
9477 softdep_change_linkcnt(ip)
9478 struct inode *ip; /* the inode with the increased link count */
9480 struct inodedep *inodedep;
9481 struct ufsmount *ump;
9484 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
9485 ("softdep_change_linkcnt called on non-softdep filesystem"));
9487 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep);
9488 if (ip->i_nlink < ip->i_effnlink)
9489 panic("softdep_change_linkcnt: bad delta");
9490 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9495 * Attach a sbdep dependency to the superblock buf so that we can keep
9496 * track of the head of the linked list of referenced but unlinked inodes.
9499 softdep_setup_sbupdate(ump, fs, bp)
9500 struct ufsmount *ump;
9504 struct sbdep *sbdep;
9505 struct worklist *wk;
9507 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
9508 ("softdep_setup_sbupdate called on non-softdep filesystem"));
9509 LIST_FOREACH(wk, &bp->b_dep, wk_list)
9510 if (wk->wk_type == D_SBDEP)
9514 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
9515 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
9517 sbdep->sb_ump = ump;
9519 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
9524 * Return the first unlinked inodedep which is ready to be the head of the
9525 * list. The inodedep and all those after it must have valid next pointers.
9527 static struct inodedep *
9528 first_unlinked_inodedep(ump)
9529 struct ufsmount *ump;
9531 struct inodedep *inodedep;
9532 struct inodedep *idp;
9535 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
9536 inodedep; inodedep = idp) {
9537 if ((inodedep->id_state & UNLINKNEXT) == 0)
9539 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9540 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
9542 if ((inodedep->id_state & UNLINKPREV) == 0)
9549 * Set the sujfree unlinked head pointer prior to writing a superblock.
9552 initiate_write_sbdep(sbdep)
9553 struct sbdep *sbdep;
9555 struct inodedep *inodedep;
9559 bpfs = sbdep->sb_fs;
9560 fs = sbdep->sb_ump->um_fs;
9561 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9563 fs->fs_sujfree = inodedep->id_ino;
9564 inodedep->id_state |= UNLINKPREV;
9567 bpfs->fs_sujfree = fs->fs_sujfree;
9571 * After a superblock is written determine whether it must be written again
9572 * due to a changing unlinked list head.
9575 handle_written_sbdep(sbdep, bp)
9576 struct sbdep *sbdep;
9579 struct inodedep *inodedep;
9582 LOCK_OWNED(sbdep->sb_ump);
9585 * If the superblock doesn't match the in-memory list start over.
9587 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9588 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
9589 (inodedep == NULL && fs->fs_sujfree != 0)) {
9593 WORKITEM_FREE(sbdep, D_SBDEP);
9594 if (fs->fs_sujfree == 0)
9597 * Now that we have a record of this inode in stable store allow it
9598 * to be written to free up pending work. Inodes may see a lot of
9599 * write activity after they are unlinked which we must not hold up.
9601 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
9602 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
9603 panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
9604 inodedep, inodedep->id_state);
9605 if (inodedep->id_state & UNLINKONLIST)
9607 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
9614 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
9617 unlinked_inodedep(mp, inodedep)
9619 struct inodedep *inodedep;
9621 struct ufsmount *ump;
9625 if (MOUNTEDSUJ(mp) == 0)
9627 ump->um_fs->fs_fmod = 1;
9628 if (inodedep->id_state & UNLINKED)
9629 panic("unlinked_inodedep: %p already unlinked\n", inodedep);
9630 inodedep->id_state |= UNLINKED;
9631 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
9635 * Remove an inodedep from the unlinked inodedep list. This may require
9636 * disk writes if the inode has made it that far.
9639 clear_unlinked_inodedep(inodedep)
9640 struct inodedep *inodedep;
9642 struct ufsmount *ump;
9643 struct inodedep *idp;
9644 struct inodedep *idn;
9652 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9654 ino = inodedep->id_ino;
9658 KASSERT((inodedep->id_state & UNLINKED) != 0,
9659 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9662 * If nothing has yet been written simply remove us from
9663 * the in memory list and return. This is the most common
9664 * case where handle_workitem_remove() loses the final
9667 if ((inodedep->id_state & UNLINKLINKS) == 0)
9670 * If we have a NEXT pointer and no PREV pointer we can simply
9671 * clear NEXT's PREV and remove ourselves from the list. Be
9672 * careful not to clear PREV if the superblock points at
9675 idn = TAILQ_NEXT(inodedep, id_unlinked);
9676 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
9677 if (idn && fs->fs_sujfree != idn->id_ino)
9678 idn->id_state &= ~UNLINKPREV;
9682 * Here we have an inodedep which is actually linked into
9683 * the list. We must remove it by forcing a write to the
9684 * link before us, whether it be the superblock or an inode.
9685 * Unfortunately the list may change while we're waiting
9686 * on the buf lock for either resource so we must loop until
9687 * we lock the right one. If both the superblock and an
9688 * inode point to this inode we must clear the inode first
9689 * followed by the superblock.
9691 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9693 if (idp && (idp->id_state & UNLINKNEXT))
9697 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9698 (int)fs->fs_sbsize, 0, 0, 0);
9700 error = bread(ump->um_devvp,
9701 fsbtodb(fs, ino_to_fsba(fs, pino)),
9702 (int)fs->fs_bsize, NOCRED, &bp);
9709 /* If the list has changed restart the loop. */
9710 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9712 if (idp && (idp->id_state & UNLINKNEXT))
9715 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
9722 idn = TAILQ_NEXT(inodedep, id_unlinked);
9726 * Remove us from the in memory list. After this we cannot
9727 * access the inodedep.
9729 KASSERT((inodedep->id_state & UNLINKED) != 0,
9730 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9732 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9733 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9736 * The predecessor's next pointer is manually updated here
9737 * so that the NEXT flag is never cleared for an element
9738 * that is in the list.
9741 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9742 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9743 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9745 } else if (fs->fs_magic == FS_UFS1_MAGIC)
9746 ((struct ufs1_dinode *)bp->b_data +
9747 ino_to_fsbo(fs, pino))->di_freelink = nino;
9749 ((struct ufs2_dinode *)bp->b_data +
9750 ino_to_fsbo(fs, pino))->di_freelink = nino;
9752 * If the bwrite fails we have no recourse to recover. The
9753 * filesystem is corrupted already.
9758 * If the superblock pointer still needs to be cleared force
9761 if (fs->fs_sujfree == ino) {
9763 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9764 (int)fs->fs_sbsize, 0, 0, 0);
9765 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9766 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9767 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9773 if (fs->fs_sujfree != ino)
9775 panic("clear_unlinked_inodedep: Failed to clear free head");
9777 if (inodedep->id_ino == fs->fs_sujfree)
9778 panic("clear_unlinked_inodedep: Freeing head of free list");
9779 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9780 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9785 * This workitem decrements the inode's link count.
9786 * If the link count reaches zero, the file is removed.
9789 handle_workitem_remove(dirrem, flags)
9790 struct dirrem *dirrem;
9793 struct inodedep *inodedep;
9794 struct workhead dotdotwk;
9795 struct worklist *wk;
9796 struct ufsmount *ump;
9802 if (dirrem->dm_state & ONWORKLIST)
9803 panic("handle_workitem_remove: dirrem %p still on worklist",
9805 oldinum = dirrem->dm_oldinum;
9806 mp = dirrem->dm_list.wk_mp;
9808 flags |= LK_EXCLUSIVE;
9809 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0)
9813 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
9814 panic("handle_workitem_remove: lost inodedep");
9815 if (dirrem->dm_state & ONDEPLIST)
9816 LIST_REMOVE(dirrem, dm_inonext);
9817 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
9818 ("handle_workitem_remove: Journal entries not written."));
9821 * Move all dependencies waiting on the remove to complete
9822 * from the dirrem to the inode inowait list to be completed
9823 * after the inode has been updated and written to disk. Any
9824 * marked MKDIR_PARENT are saved to be completed when the .. ref
9827 LIST_INIT(&dotdotwk);
9828 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
9829 WORKLIST_REMOVE(wk);
9830 if (wk->wk_state & MKDIR_PARENT) {
9831 wk->wk_state &= ~MKDIR_PARENT;
9832 WORKLIST_INSERT(&dotdotwk, wk);
9835 WORKLIST_INSERT(&inodedep->id_inowait, wk);
9837 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
9839 * Normal file deletion.
9841 if ((dirrem->dm_state & RMDIR) == 0) {
9843 DIP_SET(ip, i_nlink, ip->i_nlink);
9844 ip->i_flag |= IN_CHANGE;
9845 if (ip->i_nlink < ip->i_effnlink)
9846 panic("handle_workitem_remove: bad file delta");
9847 if (ip->i_nlink == 0)
9848 unlinked_inodedep(mp, inodedep);
9849 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9850 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9851 ("handle_workitem_remove: worklist not empty. %s",
9852 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
9853 WORKITEM_FREE(dirrem, D_DIRREM);
9858 * Directory deletion. Decrement reference count for both the
9859 * just deleted parent directory entry and the reference for ".".
9860 * Arrange to have the reference count on the parent decremented
9861 * to account for the loss of "..".
9864 DIP_SET(ip, i_nlink, ip->i_nlink);
9865 ip->i_flag |= IN_CHANGE;
9866 if (ip->i_nlink < ip->i_effnlink)
9867 panic("handle_workitem_remove: bad dir delta");
9868 if (ip->i_nlink == 0)
9869 unlinked_inodedep(mp, inodedep);
9870 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9872 * Rename a directory to a new parent. Since, we are both deleting
9873 * and creating a new directory entry, the link count on the new
9874 * directory should not change. Thus we skip the followup dirrem.
9876 if (dirrem->dm_state & DIRCHG) {
9877 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9878 ("handle_workitem_remove: DIRCHG and worklist not empty."));
9879 WORKITEM_FREE(dirrem, D_DIRREM);
9883 dirrem->dm_state = ONDEPLIST;
9884 dirrem->dm_oldinum = dirrem->dm_dirinum;
9886 * Place the dirrem on the parent's diremhd list.
9888 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
9889 panic("handle_workitem_remove: lost dir inodedep");
9890 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9892 * If the allocated inode has never been written to disk, then
9893 * the on-disk inode is zero'ed and we can remove the file
9894 * immediately. When journaling if the inode has been marked
9895 * unlinked and not DEPCOMPLETE we know it can never be written.
9897 inodedep_lookup(mp, oldinum, 0, &inodedep);
9898 if (inodedep == NULL ||
9899 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
9900 check_inode_unwritten(inodedep)) {
9903 return handle_workitem_remove(dirrem, flags);
9905 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
9907 ip->i_flag |= IN_CHANGE;
9915 * Inode de-allocation dependencies.
9917 * When an inode's link count is reduced to zero, it can be de-allocated. We
9918 * found it convenient to postpone de-allocation until after the inode is
9919 * written to disk with its new link count (zero). At this point, all of the
9920 * on-disk inode's block pointers are nullified and, with careful dependency
9921 * list ordering, all dependencies related to the inode will be satisfied and
9922 * the corresponding dependency structures de-allocated. So, if/when the
9923 * inode is reused, there will be no mixing of old dependencies with new
9924 * ones. This artificial dependency is set up by the block de-allocation
9925 * procedure above (softdep_setup_freeblocks) and completed by the
9926 * following procedure.
9929 handle_workitem_freefile(freefile)
9930 struct freefile *freefile;
9932 struct workhead wkhd;
9934 struct inodedep *idp;
9935 struct ufsmount *ump;
9938 ump = VFSTOUFS(freefile->fx_list.wk_mp);
9942 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
9945 panic("handle_workitem_freefile: inodedep %p survived", idp);
9948 fs->fs_pendinginodes -= 1;
9951 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
9952 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
9953 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
9954 softdep_error("handle_workitem_freefile", error);
9956 WORKITEM_FREE(freefile, D_FREEFILE);
9962 * Helper function which unlinks marker element from work list and returns
9963 * the next element on the list.
9965 static __inline struct worklist *
9966 markernext(struct worklist *marker)
9968 struct worklist *next;
9970 next = LIST_NEXT(marker, wk_list);
9971 LIST_REMOVE(marker, wk_list);
9978 * The dependency structures constructed above are most actively used when file
9979 * system blocks are written to disk. No constraints are placed on when a
9980 * block can be written, but unsatisfied update dependencies are made safe by
9981 * modifying (or replacing) the source memory for the duration of the disk
9982 * write. When the disk write completes, the memory block is again brought
9985 * In-core inode structure reclamation.
9987 * Because there are a finite number of "in-core" inode structures, they are
9988 * reused regularly. By transferring all inode-related dependencies to the
9989 * in-memory inode block and indexing them separately (via "inodedep"s), we
9990 * can allow "in-core" inode structures to be reused at any time and avoid
9991 * any increase in contention.
9993 * Called just before entering the device driver to initiate a new disk I/O.
9994 * The buffer must be locked, thus, no I/O completion operations can occur
9995 * while we are manipulating its associated dependencies.
9998 softdep_disk_io_initiation(bp)
9999 struct buf *bp; /* structure describing disk write to occur */
10001 struct worklist *wk;
10002 struct worklist marker;
10003 struct inodedep *inodedep;
10004 struct freeblks *freeblks;
10005 struct jblkdep *jblkdep;
10006 struct newblk *newblk;
10007 struct ufsmount *ump;
10010 * We only care about write operations. There should never
10011 * be dependencies for reads.
10013 if (bp->b_iocmd != BIO_WRITE)
10014 panic("softdep_disk_io_initiation: not write");
10016 if (bp->b_vflags & BV_BKGRDINPROG)
10017 panic("softdep_disk_io_initiation: Writing buffer with "
10018 "background write in progress: %p", bp);
10020 ump = softdep_bp_to_mp(bp);
10024 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
10025 PHOLD(curproc); /* Don't swap out kernel stack */
10028 * Do any necessary pre-I/O processing.
10030 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
10031 wk = markernext(&marker)) {
10032 LIST_INSERT_AFTER(wk, &marker, wk_list);
10033 switch (wk->wk_type) {
10036 initiate_write_filepage(WK_PAGEDEP(wk), bp);
10040 inodedep = WK_INODEDEP(wk);
10041 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
10042 initiate_write_inodeblock_ufs1(inodedep, bp);
10044 initiate_write_inodeblock_ufs2(inodedep, bp);
10048 initiate_write_indirdep(WK_INDIRDEP(wk), bp);
10052 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
10056 WK_JSEG(wk)->js_buf = NULL;
10060 freeblks = WK_FREEBLKS(wk);
10061 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd);
10063 * We have to wait for the freeblks to be journaled
10064 * before we can write an inodeblock with updated
10065 * pointers. Be careful to arrange the marker so
10066 * we revisit the freeblks if it's not removed by
10067 * the first jwait().
10069 if (jblkdep != NULL) {
10070 LIST_REMOVE(&marker, wk_list);
10071 LIST_INSERT_BEFORE(wk, &marker, wk_list);
10072 jwait(&jblkdep->jb_list, MNT_WAIT);
10075 case D_ALLOCDIRECT:
10078 * We have to wait for the jnewblk to be journaled
10079 * before we can write to a block if the contents
10080 * may be confused with an earlier file's indirect
10081 * at recovery time. Handle the marker as described
10084 newblk = WK_NEWBLK(wk);
10085 if (newblk->nb_jnewblk != NULL &&
10086 indirblk_lookup(newblk->nb_list.wk_mp,
10087 newblk->nb_newblkno)) {
10088 LIST_REMOVE(&marker, wk_list);
10089 LIST_INSERT_BEFORE(wk, &marker, wk_list);
10090 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
10095 initiate_write_sbdep(WK_SBDEP(wk));
10105 panic("handle_disk_io_initiation: Unexpected type %s",
10106 TYPENAME(wk->wk_type));
10111 PRELE(curproc); /* Allow swapout of kernel stack */
10115 * Called from within the procedure above to deal with unsatisfied
10116 * allocation dependencies in a directory. The buffer must be locked,
10117 * thus, no I/O completion operations can occur while we are
10118 * manipulating its associated dependencies.
10121 initiate_write_filepage(pagedep, bp)
10122 struct pagedep *pagedep;
10125 struct jremref *jremref;
10126 struct jmvref *jmvref;
10127 struct dirrem *dirrem;
10128 struct diradd *dap;
10132 if (pagedep->pd_state & IOSTARTED) {
10134 * This can only happen if there is a driver that does not
10135 * understand chaining. Here biodone will reissue the call
10136 * to strategy for the incomplete buffers.
10138 printf("initiate_write_filepage: already started\n");
10141 pagedep->pd_state |= IOSTARTED;
10143 * Wait for all journal remove dependencies to hit the disk.
10144 * We can not allow any potentially conflicting directory adds
10145 * to be visible before removes and rollback is too difficult.
10146 * The per-filesystem lock may be dropped and re-acquired, however
10147 * we hold the buf locked so the dependency can not go away.
10149 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
10150 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL)
10151 jwait(&jremref->jr_list, MNT_WAIT);
10152 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL)
10153 jwait(&jmvref->jm_list, MNT_WAIT);
10154 for (i = 0; i < DAHASHSZ; i++) {
10155 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
10156 ep = (struct direct *)
10157 ((char *)bp->b_data + dap->da_offset);
10158 if (ep->d_ino != dap->da_newinum)
10159 panic("%s: dir inum %ju != new %ju",
10160 "initiate_write_filepage",
10161 (uintmax_t)ep->d_ino,
10162 (uintmax_t)dap->da_newinum);
10163 if (dap->da_state & DIRCHG)
10164 ep->d_ino = dap->da_previous->dm_oldinum;
10167 dap->da_state &= ~ATTACHED;
10168 dap->da_state |= UNDONE;
10174 * Version of initiate_write_inodeblock that handles UFS1 dinodes.
10175 * Note that any bug fixes made to this routine must be done in the
10176 * version found below.
10178 * Called from within the procedure above to deal with unsatisfied
10179 * allocation dependencies in an inodeblock. The buffer must be
10180 * locked, thus, no I/O completion operations can occur while we
10181 * are manipulating its associated dependencies.
10184 initiate_write_inodeblock_ufs1(inodedep, bp)
10185 struct inodedep *inodedep;
10186 struct buf *bp; /* The inode block */
10188 struct allocdirect *adp, *lastadp;
10189 struct ufs1_dinode *dp;
10190 struct ufs1_dinode *sip;
10191 struct inoref *inoref;
10192 struct ufsmount *ump;
10196 ufs_lbn_t prevlbn = 0;
10200 if (inodedep->id_state & IOSTARTED)
10201 panic("initiate_write_inodeblock_ufs1: already started");
10202 inodedep->id_state |= IOSTARTED;
10203 fs = inodedep->id_fs;
10204 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10206 dp = (struct ufs1_dinode *)bp->b_data +
10207 ino_to_fsbo(fs, inodedep->id_ino);
10210 * If we're on the unlinked list but have not yet written our
10211 * next pointer initialize it here.
10213 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10214 struct inodedep *inon;
10216 inon = TAILQ_NEXT(inodedep, id_unlinked);
10217 dp->di_freelink = inon ? inon->id_ino : 0;
10220 * If the bitmap is not yet written, then the allocated
10221 * inode cannot be written to disk.
10223 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10224 if (inodedep->id_savedino1 != NULL)
10225 panic("initiate_write_inodeblock_ufs1: I/O underway");
10227 sip = malloc(sizeof(struct ufs1_dinode),
10228 M_SAVEDINO, M_SOFTDEP_FLAGS);
10230 inodedep->id_savedino1 = sip;
10231 *inodedep->id_savedino1 = *dp;
10232 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
10233 dp->di_gen = inodedep->id_savedino1->di_gen;
10234 dp->di_freelink = inodedep->id_savedino1->di_freelink;
10238 * If no dependencies, then there is nothing to roll back.
10240 inodedep->id_savedsize = dp->di_size;
10241 inodedep->id_savedextsize = 0;
10242 inodedep->id_savednlink = dp->di_nlink;
10243 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10244 TAILQ_EMPTY(&inodedep->id_inoreflst))
10247 * Revert the link count to that of the first unwritten journal entry.
10249 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10251 dp->di_nlink = inoref->if_nlink;
10253 * Set the dependencies to busy.
10255 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10256 adp = TAILQ_NEXT(adp, ad_next)) {
10258 if (deplist != 0 && prevlbn >= adp->ad_offset)
10259 panic("softdep_write_inodeblock: lbn order");
10260 prevlbn = adp->ad_offset;
10261 if (adp->ad_offset < UFS_NDADDR &&
10262 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10263 panic("initiate_write_inodeblock_ufs1: "
10264 "direct pointer #%jd mismatch %d != %jd",
10265 (intmax_t)adp->ad_offset,
10266 dp->di_db[adp->ad_offset],
10267 (intmax_t)adp->ad_newblkno);
10268 if (adp->ad_offset >= UFS_NDADDR &&
10269 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno)
10270 panic("initiate_write_inodeblock_ufs1: "
10271 "indirect pointer #%jd mismatch %d != %jd",
10272 (intmax_t)adp->ad_offset - UFS_NDADDR,
10273 dp->di_ib[adp->ad_offset - UFS_NDADDR],
10274 (intmax_t)adp->ad_newblkno);
10275 deplist |= 1 << adp->ad_offset;
10276 if ((adp->ad_state & ATTACHED) == 0)
10277 panic("initiate_write_inodeblock_ufs1: "
10278 "Unknown state 0x%x", adp->ad_state);
10279 #endif /* INVARIANTS */
10280 adp->ad_state &= ~ATTACHED;
10281 adp->ad_state |= UNDONE;
10284 * The on-disk inode cannot claim to be any larger than the last
10285 * fragment that has been written. Otherwise, the on-disk inode
10286 * might have fragments that were not the last block in the file
10287 * which would corrupt the filesystem.
10289 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10290 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10291 if (adp->ad_offset >= UFS_NDADDR)
10293 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10294 /* keep going until hitting a rollback to a frag */
10295 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10297 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10298 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) {
10300 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10301 panic("initiate_write_inodeblock_ufs1: "
10303 #endif /* INVARIANTS */
10306 for (i = 0; i < UFS_NIADDR; i++) {
10308 if (dp->di_ib[i] != 0 &&
10309 (deplist & ((1 << UFS_NDADDR) << i)) == 0)
10310 panic("initiate_write_inodeblock_ufs1: "
10312 #endif /* INVARIANTS */
10318 * If we have zero'ed out the last allocated block of the file,
10319 * roll back the size to the last currently allocated block.
10320 * We know that this last allocated block is a full-sized as
10321 * we already checked for fragments in the loop above.
10323 if (lastadp != NULL &&
10324 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10325 for (i = lastadp->ad_offset; i >= 0; i--)
10326 if (dp->di_db[i] != 0)
10328 dp->di_size = (i + 1) * fs->fs_bsize;
10331 * The only dependencies are for indirect blocks.
10333 * The file size for indirect block additions is not guaranteed.
10334 * Such a guarantee would be non-trivial to achieve. The conventional
10335 * synchronous write implementation also does not make this guarantee.
10336 * Fsck should catch and fix discrepancies. Arguably, the file size
10337 * can be over-estimated without destroying integrity when the file
10338 * moves into the indirect blocks (i.e., is large). If we want to
10339 * postpone fsck, we are stuck with this argument.
10341 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10342 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0;
10346 * Version of initiate_write_inodeblock that handles UFS2 dinodes.
10347 * Note that any bug fixes made to this routine must be done in the
10348 * version found above.
10350 * Called from within the procedure above to deal with unsatisfied
10351 * allocation dependencies in an inodeblock. The buffer must be
10352 * locked, thus, no I/O completion operations can occur while we
10353 * are manipulating its associated dependencies.
10356 initiate_write_inodeblock_ufs2(inodedep, bp)
10357 struct inodedep *inodedep;
10358 struct buf *bp; /* The inode block */
10360 struct allocdirect *adp, *lastadp;
10361 struct ufs2_dinode *dp;
10362 struct ufs2_dinode *sip;
10363 struct inoref *inoref;
10364 struct ufsmount *ump;
10368 ufs_lbn_t prevlbn = 0;
10372 if (inodedep->id_state & IOSTARTED)
10373 panic("initiate_write_inodeblock_ufs2: already started");
10374 inodedep->id_state |= IOSTARTED;
10375 fs = inodedep->id_fs;
10376 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10378 dp = (struct ufs2_dinode *)bp->b_data +
10379 ino_to_fsbo(fs, inodedep->id_ino);
10382 * If we're on the unlinked list but have not yet written our
10383 * next pointer initialize it here.
10385 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10386 struct inodedep *inon;
10388 inon = TAILQ_NEXT(inodedep, id_unlinked);
10389 dp->di_freelink = inon ? inon->id_ino : 0;
10392 * If the bitmap is not yet written, then the allocated
10393 * inode cannot be written to disk.
10395 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10396 if (inodedep->id_savedino2 != NULL)
10397 panic("initiate_write_inodeblock_ufs2: I/O underway");
10399 sip = malloc(sizeof(struct ufs2_dinode),
10400 M_SAVEDINO, M_SOFTDEP_FLAGS);
10402 inodedep->id_savedino2 = sip;
10403 *inodedep->id_savedino2 = *dp;
10404 bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
10405 dp->di_gen = inodedep->id_savedino2->di_gen;
10406 dp->di_freelink = inodedep->id_savedino2->di_freelink;
10410 * If no dependencies, then there is nothing to roll back.
10412 inodedep->id_savedsize = dp->di_size;
10413 inodedep->id_savedextsize = dp->di_extsize;
10414 inodedep->id_savednlink = dp->di_nlink;
10415 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10416 TAILQ_EMPTY(&inodedep->id_extupdt) &&
10417 TAILQ_EMPTY(&inodedep->id_inoreflst))
10420 * Revert the link count to that of the first unwritten journal entry.
10422 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10424 dp->di_nlink = inoref->if_nlink;
10427 * Set the ext data dependencies to busy.
10429 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10430 adp = TAILQ_NEXT(adp, ad_next)) {
10432 if (deplist != 0 && prevlbn >= adp->ad_offset)
10433 panic("initiate_write_inodeblock_ufs2: lbn order");
10434 prevlbn = adp->ad_offset;
10435 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
10436 panic("initiate_write_inodeblock_ufs2: "
10437 "ext pointer #%jd mismatch %jd != %jd",
10438 (intmax_t)adp->ad_offset,
10439 (intmax_t)dp->di_extb[adp->ad_offset],
10440 (intmax_t)adp->ad_newblkno);
10441 deplist |= 1 << adp->ad_offset;
10442 if ((adp->ad_state & ATTACHED) == 0)
10443 panic("initiate_write_inodeblock_ufs2: Unknown "
10444 "state 0x%x", adp->ad_state);
10445 #endif /* INVARIANTS */
10446 adp->ad_state &= ~ATTACHED;
10447 adp->ad_state |= UNDONE;
10450 * The on-disk inode cannot claim to be any larger than the last
10451 * fragment that has been written. Otherwise, the on-disk inode
10452 * might have fragments that were not the last block in the ext
10453 * data which would corrupt the filesystem.
10455 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10456 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10457 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
10458 /* keep going until hitting a rollback to a frag */
10459 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10461 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10462 for (i = adp->ad_offset + 1; i < UFS_NXADDR; i++) {
10464 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
10465 panic("initiate_write_inodeblock_ufs2: "
10467 #endif /* INVARIANTS */
10468 dp->di_extb[i] = 0;
10474 * If we have zero'ed out the last allocated block of the ext
10475 * data, roll back the size to the last currently allocated block.
10476 * We know that this last allocated block is a full-sized as
10477 * we already checked for fragments in the loop above.
10479 if (lastadp != NULL &&
10480 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10481 for (i = lastadp->ad_offset; i >= 0; i--)
10482 if (dp->di_extb[i] != 0)
10484 dp->di_extsize = (i + 1) * fs->fs_bsize;
10487 * Set the file data dependencies to busy.
10489 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10490 adp = TAILQ_NEXT(adp, ad_next)) {
10492 if (deplist != 0 && prevlbn >= adp->ad_offset)
10493 panic("softdep_write_inodeblock: lbn order");
10494 if ((adp->ad_state & ATTACHED) == 0)
10495 panic("inodedep %p and adp %p not attached", inodedep, adp);
10496 prevlbn = adp->ad_offset;
10497 if (adp->ad_offset < UFS_NDADDR &&
10498 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10499 panic("initiate_write_inodeblock_ufs2: "
10500 "direct pointer #%jd mismatch %jd != %jd",
10501 (intmax_t)adp->ad_offset,
10502 (intmax_t)dp->di_db[adp->ad_offset],
10503 (intmax_t)adp->ad_newblkno);
10504 if (adp->ad_offset >= UFS_NDADDR &&
10505 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno)
10506 panic("initiate_write_inodeblock_ufs2: "
10507 "indirect pointer #%jd mismatch %jd != %jd",
10508 (intmax_t)adp->ad_offset - UFS_NDADDR,
10509 (intmax_t)dp->di_ib[adp->ad_offset - UFS_NDADDR],
10510 (intmax_t)adp->ad_newblkno);
10511 deplist |= 1 << adp->ad_offset;
10512 if ((adp->ad_state & ATTACHED) == 0)
10513 panic("initiate_write_inodeblock_ufs2: Unknown "
10514 "state 0x%x", adp->ad_state);
10515 #endif /* INVARIANTS */
10516 adp->ad_state &= ~ATTACHED;
10517 adp->ad_state |= UNDONE;
10520 * The on-disk inode cannot claim to be any larger than the last
10521 * fragment that has been written. Otherwise, the on-disk inode
10522 * might have fragments that were not the last block in the file
10523 * which would corrupt the filesystem.
10525 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10526 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10527 if (adp->ad_offset >= UFS_NDADDR)
10529 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10530 /* keep going until hitting a rollback to a frag */
10531 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10533 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10534 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) {
10536 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10537 panic("initiate_write_inodeblock_ufs2: "
10539 #endif /* INVARIANTS */
10542 for (i = 0; i < UFS_NIADDR; i++) {
10544 if (dp->di_ib[i] != 0 &&
10545 (deplist & ((1 << UFS_NDADDR) << i)) == 0)
10546 panic("initiate_write_inodeblock_ufs2: "
10548 #endif /* INVARIANTS */
10554 * If we have zero'ed out the last allocated block of the file,
10555 * roll back the size to the last currently allocated block.
10556 * We know that this last allocated block is a full-sized as
10557 * we already checked for fragments in the loop above.
10559 if (lastadp != NULL &&
10560 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10561 for (i = lastadp->ad_offset; i >= 0; i--)
10562 if (dp->di_db[i] != 0)
10564 dp->di_size = (i + 1) * fs->fs_bsize;
10567 * The only dependencies are for indirect blocks.
10569 * The file size for indirect block additions is not guaranteed.
10570 * Such a guarantee would be non-trivial to achieve. The conventional
10571 * synchronous write implementation also does not make this guarantee.
10572 * Fsck should catch and fix discrepancies. Arguably, the file size
10573 * can be over-estimated without destroying integrity when the file
10574 * moves into the indirect blocks (i.e., is large). If we want to
10575 * postpone fsck, we are stuck with this argument.
10577 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10578 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0;
10582 * Cancel an indirdep as a result of truncation. Release all of the
10583 * children allocindirs and place their journal work on the appropriate
10587 cancel_indirdep(indirdep, bp, freeblks)
10588 struct indirdep *indirdep;
10590 struct freeblks *freeblks;
10592 struct allocindir *aip;
10595 * None of the indirect pointers will ever be visible,
10596 * so they can simply be tossed. GOINGAWAY ensures
10597 * that allocated pointers will be saved in the buffer
10598 * cache until they are freed. Note that they will
10599 * only be able to be found by their physical address
10600 * since the inode mapping the logical address will
10601 * be gone. The save buffer used for the safe copy
10602 * was allocated in setup_allocindir_phase2 using
10603 * the physical address so it could be used for this
10604 * purpose. Hence we swap the safe copy with the real
10605 * copy, allowing the safe copy to be freed and holding
10606 * on to the real copy for later use in indir_trunc.
10608 if (indirdep->ir_state & GOINGAWAY)
10609 panic("cancel_indirdep: already gone");
10610 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
10611 indirdep->ir_state |= DEPCOMPLETE;
10612 LIST_REMOVE(indirdep, ir_next);
10614 indirdep->ir_state |= GOINGAWAY;
10616 * Pass in bp for blocks still have journal writes
10617 * pending so we can cancel them on their own.
10619 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL)
10620 cancel_allocindir(aip, bp, freeblks, 0);
10621 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL)
10622 cancel_allocindir(aip, NULL, freeblks, 0);
10623 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL)
10624 cancel_allocindir(aip, NULL, freeblks, 0);
10625 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL)
10626 cancel_allocindir(aip, NULL, freeblks, 0);
10628 * If there are pending partial truncations we need to keep the
10629 * old block copy around until they complete. This is because
10630 * the current b_data is not a perfect superset of the available
10633 if (TAILQ_EMPTY(&indirdep->ir_trunc))
10634 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
10636 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10637 WORKLIST_REMOVE(&indirdep->ir_list);
10638 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
10639 indirdep->ir_bp = NULL;
10640 indirdep->ir_freeblks = freeblks;
10644 * Free an indirdep once it no longer has new pointers to track.
10647 free_indirdep(indirdep)
10648 struct indirdep *indirdep;
10651 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc),
10652 ("free_indirdep: Indir trunc list not empty."));
10653 KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
10654 ("free_indirdep: Complete head not empty."));
10655 KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
10656 ("free_indirdep: write head not empty."));
10657 KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
10658 ("free_indirdep: done head not empty."));
10659 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
10660 ("free_indirdep: deplist head not empty."));
10661 KASSERT((indirdep->ir_state & DEPCOMPLETE),
10662 ("free_indirdep: %p still on newblk list.", indirdep));
10663 KASSERT(indirdep->ir_saveddata == NULL,
10664 ("free_indirdep: %p still has saved data.", indirdep));
10665 if (indirdep->ir_state & ONWORKLIST)
10666 WORKLIST_REMOVE(&indirdep->ir_list);
10667 WORKITEM_FREE(indirdep, D_INDIRDEP);
10671 * Called before a write to an indirdep. This routine is responsible for
10672 * rolling back pointers to a safe state which includes only those
10673 * allocindirs which have been completed.
10676 initiate_write_indirdep(indirdep, bp)
10677 struct indirdep *indirdep;
10680 struct ufsmount *ump;
10682 indirdep->ir_state |= IOSTARTED;
10683 if (indirdep->ir_state & GOINGAWAY)
10684 panic("disk_io_initiation: indirdep gone");
10686 * If there are no remaining dependencies, this will be writing
10687 * the real pointers.
10689 if (LIST_EMPTY(&indirdep->ir_deplisthd) &&
10690 TAILQ_EMPTY(&indirdep->ir_trunc))
10693 * Replace up-to-date version with safe version.
10695 if (indirdep->ir_saveddata == NULL) {
10696 ump = VFSTOUFS(indirdep->ir_list.wk_mp);
10699 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
10703 indirdep->ir_state &= ~ATTACHED;
10704 indirdep->ir_state |= UNDONE;
10705 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10706 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
10711 * Called when an inode has been cleared in a cg bitmap. This finally
10712 * eliminates any canceled jaddrefs
10715 softdep_setup_inofree(mp, bp, ino, wkhd)
10719 struct workhead *wkhd;
10721 struct worklist *wk, *wkn;
10722 struct inodedep *inodedep;
10723 struct ufsmount *ump;
10728 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
10729 ("softdep_setup_inofree called on non-softdep filesystem"));
10730 ump = VFSTOUFS(mp);
10733 cgp = (struct cg *)bp->b_data;
10734 inosused = cg_inosused(cgp);
10735 if (isset(inosused, ino % fs->fs_ipg))
10736 panic("softdep_setup_inofree: inode %ju not freed.",
10738 if (inodedep_lookup(mp, ino, 0, &inodedep))
10739 panic("softdep_setup_inofree: ino %ju has existing inodedep %p",
10740 (uintmax_t)ino, inodedep);
10742 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
10743 if (wk->wk_type != D_JADDREF)
10745 WORKLIST_REMOVE(wk);
10747 * We can free immediately even if the jaddref
10748 * isn't attached in a background write as now
10749 * the bitmaps are reconciled.
10751 wk->wk_state |= COMPLETE | ATTACHED;
10752 free_jaddref(WK_JADDREF(wk));
10754 jwork_move(&bp->b_dep, wkhd);
10761 * Called via ffs_blkfree() after a set of frags has been cleared from a cg
10762 * map. Any dependencies waiting for the write to clear are added to the
10763 * buf's list and any jnewblks that are being canceled are discarded
10767 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
10770 ufs2_daddr_t blkno;
10772 struct workhead *wkhd;
10774 struct bmsafemap *bmsafemap;
10775 struct jnewblk *jnewblk;
10776 struct ufsmount *ump;
10777 struct worklist *wk;
10782 ufs2_daddr_t jstart;
10790 "softdep_setup_blkfree: blkno %jd frags %d wk head %p",
10791 blkno, frags, wkhd);
10793 ump = VFSTOUFS(mp);
10794 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
10795 ("softdep_setup_blkfree called on non-softdep filesystem"));
10797 /* Lookup the bmsafemap so we track when it is dirty. */
10799 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10801 * Detach any jnewblks which have been canceled. They must linger
10802 * until the bitmap is cleared again by ffs_blkfree() to prevent
10803 * an unjournaled allocation from hitting the disk.
10806 while ((wk = LIST_FIRST(wkhd)) != NULL) {
10808 "softdep_setup_blkfree: blkno %jd wk type %d",
10809 blkno, wk->wk_type);
10810 WORKLIST_REMOVE(wk);
10811 if (wk->wk_type != D_JNEWBLK) {
10812 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk);
10815 jnewblk = WK_JNEWBLK(wk);
10816 KASSERT(jnewblk->jn_state & GOINGAWAY,
10817 ("softdep_setup_blkfree: jnewblk not canceled."));
10820 * Assert that this block is free in the bitmap
10821 * before we discard the jnewblk.
10823 cgp = (struct cg *)bp->b_data;
10824 blksfree = cg_blksfree(cgp);
10825 bno = dtogd(fs, jnewblk->jn_blkno);
10826 for (i = jnewblk->jn_oldfrags;
10827 i < jnewblk->jn_frags; i++) {
10828 if (isset(blksfree, bno + i))
10830 panic("softdep_setup_blkfree: not free");
10834 * Even if it's not attached we can free immediately
10835 * as the new bitmap is correct.
10837 wk->wk_state |= COMPLETE | ATTACHED;
10838 free_jnewblk(jnewblk);
10844 * Assert that we are not freeing a block which has an outstanding
10845 * allocation dependency.
10847 fs = VFSTOUFS(mp)->um_fs;
10848 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10849 end = blkno + frags;
10850 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10852 * Don't match against blocks that will be freed when the
10853 * background write is done.
10855 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
10856 (COMPLETE | DEPCOMPLETE))
10858 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
10859 jend = jnewblk->jn_blkno + jnewblk->jn_frags;
10860 if ((blkno >= jstart && blkno < jend) ||
10861 (end > jstart && end <= jend)) {
10862 printf("state 0x%X %jd - %d %d dep %p\n",
10863 jnewblk->jn_state, jnewblk->jn_blkno,
10864 jnewblk->jn_oldfrags, jnewblk->jn_frags,
10866 panic("softdep_setup_blkfree: "
10867 "%jd-%jd(%d) overlaps with %jd-%jd",
10868 blkno, end, frags, jstart, jend);
10876 * Revert a block allocation when the journal record that describes it
10877 * is not yet written.
10880 jnewblk_rollback(jnewblk, fs, cgp, blksfree)
10881 struct jnewblk *jnewblk;
10886 ufs1_daddr_t fragno;
10892 cgbno = dtogd(fs, jnewblk->jn_blkno);
10894 * We have to test which frags need to be rolled back. We may
10895 * be operating on a stale copy when doing background writes.
10897 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++)
10898 if (isclr(blksfree, cgbno + i))
10903 * This is mostly ffs_blkfree() sans some validation and
10904 * superblock updates.
10906 if (frags == fs->fs_frag) {
10907 fragno = fragstoblks(fs, cgbno);
10908 ffs_setblock(fs, blksfree, fragno);
10909 ffs_clusteracct(fs, cgp, fragno, 1);
10910 cgp->cg_cs.cs_nbfree++;
10912 cgbno += jnewblk->jn_oldfrags;
10913 bbase = cgbno - fragnum(fs, cgbno);
10914 /* Decrement the old frags. */
10915 blk = blkmap(fs, blksfree, bbase);
10916 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
10917 /* Deallocate the fragment */
10918 for (i = 0; i < frags; i++)
10919 setbit(blksfree, cgbno + i);
10920 cgp->cg_cs.cs_nffree += frags;
10921 /* Add back in counts associated with the new frags */
10922 blk = blkmap(fs, blksfree, bbase);
10923 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
10924 /* If a complete block has been reassembled, account for it. */
10925 fragno = fragstoblks(fs, bbase);
10926 if (ffs_isblock(fs, blksfree, fragno)) {
10927 cgp->cg_cs.cs_nffree -= fs->fs_frag;
10928 ffs_clusteracct(fs, cgp, fragno, 1);
10929 cgp->cg_cs.cs_nbfree++;
10933 jnewblk->jn_state &= ~ATTACHED;
10934 jnewblk->jn_state |= UNDONE;
10940 initiate_write_bmsafemap(bmsafemap, bp)
10941 struct bmsafemap *bmsafemap;
10942 struct buf *bp; /* The cg block. */
10944 struct jaddref *jaddref;
10945 struct jnewblk *jnewblk;
10953 * If this is a background write, we did this at the time that
10954 * the copy was made, so do not need to do it again.
10956 if (bmsafemap->sm_state & IOSTARTED)
10958 bmsafemap->sm_state |= IOSTARTED;
10960 * Clear any inode allocations which are pending journal writes.
10962 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) {
10963 cgp = (struct cg *)bp->b_data;
10964 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10965 inosused = cg_inosused(cgp);
10966 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) {
10967 ino = jaddref->ja_ino % fs->fs_ipg;
10968 if (isset(inosused, ino)) {
10969 if ((jaddref->ja_mode & IFMT) == IFDIR)
10970 cgp->cg_cs.cs_ndir--;
10971 cgp->cg_cs.cs_nifree++;
10972 clrbit(inosused, ino);
10973 jaddref->ja_state &= ~ATTACHED;
10974 jaddref->ja_state |= UNDONE;
10977 panic("initiate_write_bmsafemap: inode %ju "
10978 "marked free", (uintmax_t)jaddref->ja_ino);
10982 * Clear any block allocations which are pending journal writes.
10984 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
10985 cgp = (struct cg *)bp->b_data;
10986 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10987 blksfree = cg_blksfree(cgp);
10988 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10989 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree))
10991 panic("initiate_write_bmsafemap: block %jd "
10992 "marked free", jnewblk->jn_blkno);
10996 * Move allocation lists to the written lists so they can be
10997 * cleared once the block write is complete.
10999 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr,
11000 inodedep, id_deps);
11001 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
11003 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist,
11008 * This routine is called during the completion interrupt
11009 * service routine for a disk write (from the procedure called
11010 * by the device driver to inform the filesystem caches of
11011 * a request completion). It should be called early in this
11012 * procedure, before the block is made available to other
11013 * processes or other routines are called.
11017 softdep_disk_write_complete(bp)
11018 struct buf *bp; /* describes the completed disk write */
11020 struct worklist *wk;
11021 struct worklist *owk;
11022 struct ufsmount *ump;
11023 struct workhead reattach;
11024 struct freeblks *freeblks;
11027 ump = softdep_bp_to_mp(bp);
11034 * If an error occurred while doing the write, then the data
11035 * has not hit the disk and the dependencies cannot be processed.
11036 * But we do have to go through and roll forward any dependencies
11037 * that were rolled back before the disk write.
11040 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) {
11041 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
11042 switch (wk->wk_type) {
11045 handle_written_filepage(WK_PAGEDEP(wk), bp, 0);
11049 handle_written_inodeblock(WK_INODEDEP(wk),
11054 handle_written_bmsafemap(WK_BMSAFEMAP(wk),
11059 handle_written_indirdep(WK_INDIRDEP(wk),
11063 /* nothing to roll forward */
11070 LIST_INIT(&reattach);
11073 * Ump SU lock must not be released anywhere in this code segment.
11076 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
11077 WORKLIST_REMOVE(wk);
11078 atomic_add_long(&dep_write[wk->wk_type], 1);
11080 panic("duplicate worklist: %p\n", wk);
11082 switch (wk->wk_type) {
11085 if (handle_written_filepage(WK_PAGEDEP(wk), bp,
11087 WORKLIST_INSERT(&reattach, wk);
11091 if (handle_written_inodeblock(WK_INODEDEP(wk), bp,
11093 WORKLIST_INSERT(&reattach, wk);
11097 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp,
11099 WORKLIST_INSERT(&reattach, wk);
11103 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
11106 case D_ALLOCDIRECT:
11107 wk->wk_state |= COMPLETE;
11108 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
11112 wk->wk_state |= COMPLETE;
11113 handle_allocindir_partdone(WK_ALLOCINDIR(wk));
11117 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp,
11119 WORKLIST_INSERT(&reattach, wk);
11123 wk->wk_state |= COMPLETE;
11124 freeblks = WK_FREEBLKS(wk);
11125 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE &&
11126 LIST_EMPTY(&freeblks->fb_jblkdephd))
11127 add_to_worklist(wk, WK_NODELAY);
11131 handle_written_freework(WK_FREEWORK(wk));
11135 free_jsegdep(WK_JSEGDEP(wk));
11139 handle_written_jseg(WK_JSEG(wk), bp);
11143 if (handle_written_sbdep(WK_SBDEP(wk), bp))
11144 WORKLIST_INSERT(&reattach, wk);
11148 free_freedep(WK_FREEDEP(wk));
11152 panic("handle_disk_write_complete: Unknown type %s",
11153 TYPENAME(wk->wk_type));
11158 * Reattach any requests that must be redone.
11160 while ((wk = LIST_FIRST(&reattach)) != NULL) {
11161 WORKLIST_REMOVE(wk);
11162 WORKLIST_INSERT(&bp->b_dep, wk);
11170 * Called from within softdep_disk_write_complete above.
11173 handle_allocdirect_partdone(adp, wkhd)
11174 struct allocdirect *adp; /* the completed allocdirect */
11175 struct workhead *wkhd; /* Work to do when inode is writtne. */
11177 struct allocdirectlst *listhead;
11178 struct allocdirect *listadp;
11179 struct inodedep *inodedep;
11182 LOCK_OWNED(VFSTOUFS(adp->ad_block.nb_list.wk_mp));
11183 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11186 * The on-disk inode cannot claim to be any larger than the last
11187 * fragment that has been written. Otherwise, the on-disk inode
11188 * might have fragments that were not the last block in the file
11189 * which would corrupt the filesystem. Thus, we cannot free any
11190 * allocdirects after one whose ad_oldblkno claims a fragment as
11191 * these blocks must be rolled back to zero before writing the inode.
11192 * We check the currently active set of allocdirects in id_inoupdt
11193 * or id_extupdt as appropriate.
11195 inodedep = adp->ad_inodedep;
11196 bsize = inodedep->id_fs->fs_bsize;
11197 if (adp->ad_state & EXTDATA)
11198 listhead = &inodedep->id_extupdt;
11200 listhead = &inodedep->id_inoupdt;
11201 TAILQ_FOREACH(listadp, listhead, ad_next) {
11202 /* found our block */
11203 if (listadp == adp)
11205 /* continue if ad_oldlbn is not a fragment */
11206 if (listadp->ad_oldsize == 0 ||
11207 listadp->ad_oldsize == bsize)
11209 /* hit a fragment */
11213 * If we have reached the end of the current list without
11214 * finding the just finished dependency, then it must be
11215 * on the future dependency list. Future dependencies cannot
11216 * be freed until they are moved to the current list.
11218 if (listadp == NULL) {
11220 if (adp->ad_state & EXTDATA)
11221 listhead = &inodedep->id_newextupdt;
11223 listhead = &inodedep->id_newinoupdt;
11224 TAILQ_FOREACH(listadp, listhead, ad_next)
11225 /* found our block */
11226 if (listadp == adp)
11228 if (listadp == NULL)
11229 panic("handle_allocdirect_partdone: lost dep");
11234 * If we have found the just finished dependency, then queue
11235 * it along with anything that follows it that is complete.
11236 * Since the pointer has not yet been written in the inode
11237 * as the dependency prevents it, place the allocdirect on the
11238 * bufwait list where it will be freed once the pointer is
11242 wkhd = &inodedep->id_bufwait;
11243 for (; adp; adp = listadp) {
11244 listadp = TAILQ_NEXT(adp, ad_next);
11245 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11247 TAILQ_REMOVE(listhead, adp, ad_next);
11248 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
11253 * Called from within softdep_disk_write_complete above. This routine
11254 * completes successfully written allocindirs.
11257 handle_allocindir_partdone(aip)
11258 struct allocindir *aip; /* the completed allocindir */
11260 struct indirdep *indirdep;
11262 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
11264 indirdep = aip->ai_indirdep;
11265 LIST_REMOVE(aip, ai_next);
11267 * Don't set a pointer while the buffer is undergoing IO or while
11268 * we have active truncations.
11270 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) {
11271 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
11274 if (indirdep->ir_state & UFS1FMT)
11275 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11278 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11281 * Await the pointer write before freeing the allocindir.
11283 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
11287 * Release segments held on a jwork list.
11291 struct workhead *wkhd;
11293 struct worklist *wk;
11295 while ((wk = LIST_FIRST(wkhd)) != NULL) {
11296 WORKLIST_REMOVE(wk);
11297 switch (wk->wk_type) {
11299 free_jsegdep(WK_JSEGDEP(wk));
11302 free_freedep(WK_FREEDEP(wk));
11305 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep));
11306 WORKITEM_FREE(wk, D_FREEFRAG);
11309 handle_written_freework(WK_FREEWORK(wk));
11312 panic("handle_jwork: Unknown type %s\n",
11313 TYPENAME(wk->wk_type));
11319 * Handle the bufwait list on an inode when it is safe to release items
11320 * held there. This normally happens after an inode block is written but
11321 * may be delayed and handled later if there are pending journal items that
11322 * are not yet safe to be released.
11324 static struct freefile *
11325 handle_bufwait(inodedep, refhd)
11326 struct inodedep *inodedep;
11327 struct workhead *refhd;
11329 struct jaddref *jaddref;
11330 struct freefile *freefile;
11331 struct worklist *wk;
11334 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
11335 WORKLIST_REMOVE(wk);
11336 switch (wk->wk_type) {
11339 * We defer adding freefile to the worklist
11340 * until all other additions have been made to
11341 * ensure that it will be done after all the
11342 * old blocks have been freed.
11344 if (freefile != NULL)
11345 panic("handle_bufwait: freefile");
11346 freefile = WK_FREEFILE(wk);
11350 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
11354 diradd_inode_written(WK_DIRADD(wk), inodedep);
11358 wk->wk_state |= COMPLETE;
11359 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
11360 add_to_worklist(wk, 0);
11364 wk->wk_state |= COMPLETE;
11365 add_to_worklist(wk, 0);
11368 case D_ALLOCDIRECT:
11370 free_newblk(WK_NEWBLK(wk));
11374 wk->wk_state |= COMPLETE;
11375 free_jnewblk(WK_JNEWBLK(wk));
11379 * Save freed journal segments and add references on
11380 * the supplied list which will delay their release
11381 * until the cg bitmap is cleared on disk.
11385 free_jsegdep(WK_JSEGDEP(wk));
11387 WORKLIST_INSERT(refhd, wk);
11391 jaddref = WK_JADDREF(wk);
11392 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
11395 * Transfer any jaddrefs to the list to be freed with
11396 * the bitmap if we're handling a removed file.
11398 if (refhd == NULL) {
11399 wk->wk_state |= COMPLETE;
11400 free_jaddref(jaddref);
11402 WORKLIST_INSERT(refhd, wk);
11406 panic("handle_bufwait: Unknown type %p(%s)",
11407 wk, TYPENAME(wk->wk_type));
11414 * Called from within softdep_disk_write_complete above to restore
11415 * in-memory inode block contents to their most up-to-date state. Note
11416 * that this routine is always called from interrupt level with further
11417 * interrupts from this device blocked.
11419 * If the write did not succeed, we will do all the roll-forward
11420 * operations, but we will not take the actions that will allow its
11421 * dependencies to be processed.
11424 handle_written_inodeblock(inodedep, bp, flags)
11425 struct inodedep *inodedep;
11426 struct buf *bp; /* buffer containing the inode block */
11429 struct freefile *freefile;
11430 struct allocdirect *adp, *nextadp;
11431 struct ufs1_dinode *dp1 = NULL;
11432 struct ufs2_dinode *dp2 = NULL;
11433 struct workhead wkhd;
11434 int hadchanges, fstype;
11440 if ((inodedep->id_state & IOSTARTED) == 0)
11441 panic("handle_written_inodeblock: not started");
11442 inodedep->id_state &= ~IOSTARTED;
11443 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
11445 dp1 = (struct ufs1_dinode *)bp->b_data +
11446 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11447 freelink = dp1->di_freelink;
11450 dp2 = (struct ufs2_dinode *)bp->b_data +
11451 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11452 freelink = dp2->di_freelink;
11455 * Leave this inodeblock dirty until it's in the list.
11457 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED &&
11458 (flags & WRITESUCCEEDED)) {
11459 struct inodedep *inon;
11461 inon = TAILQ_NEXT(inodedep, id_unlinked);
11462 if ((inon == NULL && freelink == 0) ||
11463 (inon && inon->id_ino == freelink)) {
11465 inon->id_state |= UNLINKPREV;
11466 inodedep->id_state |= UNLINKNEXT;
11471 * If we had to rollback the inode allocation because of
11472 * bitmaps being incomplete, then simply restore it.
11473 * Keep the block dirty so that it will not be reclaimed until
11474 * all associated dependencies have been cleared and the
11475 * corresponding updates written to disk.
11477 if (inodedep->id_savedino1 != NULL) {
11479 if (fstype == UFS1)
11480 *dp1 = *inodedep->id_savedino1;
11482 *dp2 = *inodedep->id_savedino2;
11483 free(inodedep->id_savedino1, M_SAVEDINO);
11484 inodedep->id_savedino1 = NULL;
11485 if ((bp->b_flags & B_DELWRI) == 0)
11486 stat_inode_bitmap++;
11489 * If the inode is clear here and GOINGAWAY it will never
11490 * be written. Process the bufwait and clear any pending
11491 * work which may include the freefile.
11493 if (inodedep->id_state & GOINGAWAY)
11497 if (flags & WRITESUCCEEDED)
11498 inodedep->id_state |= COMPLETE;
11500 * Roll forward anything that had to be rolled back before
11501 * the inode could be updated.
11503 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
11504 nextadp = TAILQ_NEXT(adp, ad_next);
11505 if (adp->ad_state & ATTACHED)
11506 panic("handle_written_inodeblock: new entry");
11507 if (fstype == UFS1) {
11508 if (adp->ad_offset < UFS_NDADDR) {
11509 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11510 panic("%s %s #%jd mismatch %d != %jd",
11511 "handle_written_inodeblock:",
11513 (intmax_t)adp->ad_offset,
11514 dp1->di_db[adp->ad_offset],
11515 (intmax_t)adp->ad_oldblkno);
11516 dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
11518 if (dp1->di_ib[adp->ad_offset - UFS_NDADDR] !=
11520 panic("%s: %s #%jd allocated as %d",
11521 "handle_written_inodeblock",
11522 "indirect pointer",
11523 (intmax_t)adp->ad_offset -
11525 dp1->di_ib[adp->ad_offset -
11527 dp1->di_ib[adp->ad_offset - UFS_NDADDR] =
11531 if (adp->ad_offset < UFS_NDADDR) {
11532 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11533 panic("%s: %s #%jd %s %jd != %jd",
11534 "handle_written_inodeblock",
11536 (intmax_t)adp->ad_offset, "mismatch",
11537 (intmax_t)dp2->di_db[adp->ad_offset],
11538 (intmax_t)adp->ad_oldblkno);
11539 dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
11541 if (dp2->di_ib[adp->ad_offset - UFS_NDADDR] !=
11543 panic("%s: %s #%jd allocated as %jd",
11544 "handle_written_inodeblock",
11545 "indirect pointer",
11546 (intmax_t)adp->ad_offset -
11549 dp2->di_ib[adp->ad_offset -
11551 dp2->di_ib[adp->ad_offset - UFS_NDADDR] =
11555 adp->ad_state &= ~UNDONE;
11556 adp->ad_state |= ATTACHED;
11559 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
11560 nextadp = TAILQ_NEXT(adp, ad_next);
11561 if (adp->ad_state & ATTACHED)
11562 panic("handle_written_inodeblock: new entry");
11563 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
11564 panic("%s: direct pointers #%jd %s %jd != %jd",
11565 "handle_written_inodeblock",
11566 (intmax_t)adp->ad_offset, "mismatch",
11567 (intmax_t)dp2->di_extb[adp->ad_offset],
11568 (intmax_t)adp->ad_oldblkno);
11569 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
11570 adp->ad_state &= ~UNDONE;
11571 adp->ad_state |= ATTACHED;
11574 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
11575 stat_direct_blk_ptrs++;
11577 * Reset the file size to its most up-to-date value.
11579 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
11580 panic("handle_written_inodeblock: bad size");
11581 if (inodedep->id_savednlink > UFS_LINK_MAX)
11582 panic("handle_written_inodeblock: Invalid link count "
11583 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink,
11585 if (fstype == UFS1) {
11586 if (dp1->di_nlink != inodedep->id_savednlink) {
11587 dp1->di_nlink = inodedep->id_savednlink;
11590 if (dp1->di_size != inodedep->id_savedsize) {
11591 dp1->di_size = inodedep->id_savedsize;
11595 if (dp2->di_nlink != inodedep->id_savednlink) {
11596 dp2->di_nlink = inodedep->id_savednlink;
11599 if (dp2->di_size != inodedep->id_savedsize) {
11600 dp2->di_size = inodedep->id_savedsize;
11603 if (dp2->di_extsize != inodedep->id_savedextsize) {
11604 dp2->di_extsize = inodedep->id_savedextsize;
11608 inodedep->id_savedsize = -1;
11609 inodedep->id_savedextsize = -1;
11610 inodedep->id_savednlink = -1;
11612 * If there were any rollbacks in the inode block, then it must be
11613 * marked dirty so that its will eventually get written back in
11614 * its correct form.
11620 * If the write did not succeed, we have done all the roll-forward
11621 * operations, but we cannot take the actions that will allow its
11622 * dependencies to be processed.
11624 if ((flags & WRITESUCCEEDED) == 0)
11625 return (hadchanges);
11627 * Process any allocdirects that completed during the update.
11629 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
11630 handle_allocdirect_partdone(adp, &wkhd);
11631 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
11632 handle_allocdirect_partdone(adp, &wkhd);
11634 * Process deallocations that were held pending until the
11635 * inode had been written to disk. Freeing of the inode
11636 * is delayed until after all blocks have been freed to
11637 * avoid creation of new <vfsid, inum, lbn> triples
11638 * before the old ones have been deleted. Completely
11639 * unlinked inodes are not processed until the unlinked
11640 * inode list is written or the last reference is removed.
11642 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
11643 freefile = handle_bufwait(inodedep, NULL);
11644 if (freefile && !LIST_EMPTY(&wkhd)) {
11645 WORKLIST_INSERT(&wkhd, &freefile->fx_list);
11650 * Move rolled forward dependency completions to the bufwait list
11651 * now that those that were already written have been processed.
11653 if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
11654 panic("handle_written_inodeblock: bufwait but no changes");
11655 jwork_move(&inodedep->id_bufwait, &wkhd);
11657 if (freefile != NULL) {
11659 * If the inode is goingaway it was never written. Fake up
11660 * the state here so free_inodedep() can succeed.
11662 if (inodedep->id_state & GOINGAWAY)
11663 inodedep->id_state |= COMPLETE | DEPCOMPLETE;
11664 if (free_inodedep(inodedep) == 0)
11665 panic("handle_written_inodeblock: live inodedep %p",
11667 add_to_worklist(&freefile->fx_list, 0);
11672 * If no outstanding dependencies, free it.
11674 if (free_inodedep(inodedep) ||
11675 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
11676 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
11677 TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
11678 LIST_FIRST(&inodedep->id_bufwait) == 0))
11680 return (hadchanges);
11684 * Perform needed roll-forwards and kick off any dependencies that
11685 * can now be processed.
11687 * If the write did not succeed, we will do all the roll-forward
11688 * operations, but we will not take the actions that will allow its
11689 * dependencies to be processed.
11692 handle_written_indirdep(indirdep, bp, bpp, flags)
11693 struct indirdep *indirdep;
11698 struct allocindir *aip;
11702 if (indirdep->ir_state & GOINGAWAY)
11703 panic("handle_written_indirdep: indirdep gone");
11704 if ((indirdep->ir_state & IOSTARTED) == 0)
11705 panic("handle_written_indirdep: IO not started");
11708 * If there were rollbacks revert them here.
11710 if (indirdep->ir_saveddata) {
11711 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
11712 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11713 free(indirdep->ir_saveddata, M_INDIRDEP);
11714 indirdep->ir_saveddata = NULL;
11718 indirdep->ir_state &= ~(UNDONE | IOSTARTED);
11719 indirdep->ir_state |= ATTACHED;
11721 * If the write did not succeed, we have done all the roll-forward
11722 * operations, but we cannot take the actions that will allow its
11723 * dependencies to be processed.
11725 if ((flags & WRITESUCCEEDED) == 0) {
11726 stat_indir_blk_ptrs++;
11731 * Move allocindirs with written pointers to the completehd if
11732 * the indirdep's pointer is not yet written. Otherwise
11735 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) {
11736 LIST_REMOVE(aip, ai_next);
11737 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
11738 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
11740 newblk_freefrag(&aip->ai_block);
11743 free_newblk(&aip->ai_block);
11746 * Move allocindirs that have finished dependency processing from
11747 * the done list to the write list after updating the pointers.
11749 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11750 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) {
11751 handle_allocindir_partdone(aip);
11752 if (aip == LIST_FIRST(&indirdep->ir_donehd))
11753 panic("disk_write_complete: not gone");
11758 * Preserve the indirdep if there were any changes or if it is not
11759 * yet valid on disk.
11762 stat_indir_blk_ptrs++;
11767 * If there were no changes we can discard the savedbp and detach
11768 * ourselves from the buf. We are only carrying completed pointers
11771 sbp = indirdep->ir_savebp;
11772 sbp->b_flags |= B_INVAL | B_NOCACHE;
11773 indirdep->ir_savebp = NULL;
11774 indirdep->ir_bp = NULL;
11776 panic("handle_written_indirdep: bp already exists.");
11779 * The indirdep may not be freed until its parent points at it.
11781 if (indirdep->ir_state & DEPCOMPLETE)
11782 free_indirdep(indirdep);
11788 * Process a diradd entry after its dependent inode has been written.
11791 diradd_inode_written(dap, inodedep)
11792 struct diradd *dap;
11793 struct inodedep *inodedep;
11796 LOCK_OWNED(VFSTOUFS(dap->da_list.wk_mp));
11797 dap->da_state |= COMPLETE;
11798 complete_diradd(dap);
11799 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
11803 * Returns true if the bmsafemap will have rollbacks when written. Must only
11804 * be called with the per-filesystem lock and the buf lock on the cg held.
11807 bmsafemap_backgroundwrite(bmsafemap, bp)
11808 struct bmsafemap *bmsafemap;
11813 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp));
11814 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) |
11815 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd);
11817 * If we're initiating a background write we need to process the
11818 * rollbacks as they exist now, not as they exist when IO starts.
11819 * No other consumers will look at the contents of the shadowed
11820 * buf so this is safe to do here.
11822 if (bp->b_xflags & BX_BKGRDMARKER)
11823 initiate_write_bmsafemap(bmsafemap, bp);
11829 * Re-apply an allocation when a cg write is complete.
11832 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)
11833 struct jnewblk *jnewblk;
11838 ufs1_daddr_t fragno;
11839 ufs2_daddr_t blkno;
11845 cgbno = dtogd(fs, jnewblk->jn_blkno);
11846 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) {
11847 if (isclr(blksfree, cgbno + i))
11848 panic("jnewblk_rollforward: re-allocated fragment");
11851 if (frags == fs->fs_frag) {
11852 blkno = fragstoblks(fs, cgbno);
11853 ffs_clrblock(fs, blksfree, (long)blkno);
11854 ffs_clusteracct(fs, cgp, blkno, -1);
11855 cgp->cg_cs.cs_nbfree--;
11857 bbase = cgbno - fragnum(fs, cgbno);
11858 cgbno += jnewblk->jn_oldfrags;
11859 /* If a complete block had been reassembled, account for it. */
11860 fragno = fragstoblks(fs, bbase);
11861 if (ffs_isblock(fs, blksfree, fragno)) {
11862 cgp->cg_cs.cs_nffree += fs->fs_frag;
11863 ffs_clusteracct(fs, cgp, fragno, -1);
11864 cgp->cg_cs.cs_nbfree--;
11866 /* Decrement the old frags. */
11867 blk = blkmap(fs, blksfree, bbase);
11868 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
11869 /* Allocate the fragment */
11870 for (i = 0; i < frags; i++)
11871 clrbit(blksfree, cgbno + i);
11872 cgp->cg_cs.cs_nffree -= frags;
11873 /* Add back in counts associated with the new frags */
11874 blk = blkmap(fs, blksfree, bbase);
11875 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
11881 * Complete a write to a bmsafemap structure. Roll forward any bitmap
11882 * changes if it's not a background write. Set all written dependencies
11883 * to DEPCOMPLETE and free the structure if possible.
11885 * If the write did not succeed, we will do all the roll-forward
11886 * operations, but we will not take the actions that will allow its
11887 * dependencies to be processed.
11890 handle_written_bmsafemap(bmsafemap, bp, flags)
11891 struct bmsafemap *bmsafemap;
11895 struct newblk *newblk;
11896 struct inodedep *inodedep;
11897 struct jaddref *jaddref, *jatmp;
11898 struct jnewblk *jnewblk, *jntmp;
11899 struct ufsmount *ump;
11908 if ((bmsafemap->sm_state & IOSTARTED) == 0)
11909 panic("handle_written_bmsafemap: Not started\n");
11910 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp);
11912 bmsafemap->sm_state &= ~IOSTARTED;
11913 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0;
11915 * If write was successful, release journal work that was waiting
11916 * on the write. Otherwise move the work back.
11918 if (flags & WRITESUCCEEDED)
11919 handle_jwork(&bmsafemap->sm_freewr);
11921 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr,
11922 worklist, wk_list);
11925 * Restore unwritten inode allocation pending jaddref writes.
11927 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
11928 cgp = (struct cg *)bp->b_data;
11929 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11930 inosused = cg_inosused(cgp);
11931 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
11932 ja_bmdeps, jatmp) {
11933 if ((jaddref->ja_state & UNDONE) == 0)
11935 ino = jaddref->ja_ino % fs->fs_ipg;
11936 if (isset(inosused, ino))
11937 panic("handle_written_bmsafemap: "
11938 "re-allocated inode");
11939 /* Do the roll-forward only if it's a real copy. */
11941 if ((jaddref->ja_mode & IFMT) == IFDIR)
11942 cgp->cg_cs.cs_ndir++;
11943 cgp->cg_cs.cs_nifree--;
11944 setbit(inosused, ino);
11947 jaddref->ja_state &= ~UNDONE;
11948 jaddref->ja_state |= ATTACHED;
11949 free_jaddref(jaddref);
11953 * Restore any block allocations which are pending journal writes.
11955 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
11956 cgp = (struct cg *)bp->b_data;
11957 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11958 blksfree = cg_blksfree(cgp);
11959 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
11961 if ((jnewblk->jn_state & UNDONE) == 0)
11963 /* Do the roll-forward only if it's a real copy. */
11965 jnewblk_rollforward(jnewblk, fs, cgp, blksfree))
11967 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
11968 jnewblk->jn_state |= ATTACHED;
11969 free_jnewblk(jnewblk);
11973 * If the write did not succeed, we have done all the roll-forward
11974 * operations, but we cannot take the actions that will allow its
11975 * dependencies to be processed.
11977 if ((flags & WRITESUCCEEDED) == 0) {
11978 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
11980 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr,
11981 worklist, wk_list);
11986 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
11987 newblk->nb_state |= DEPCOMPLETE;
11988 newblk->nb_state &= ~ONDEPLIST;
11989 newblk->nb_bmsafemap = NULL;
11990 LIST_REMOVE(newblk, nb_deps);
11991 if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
11992 handle_allocdirect_partdone(
11993 WK_ALLOCDIRECT(&newblk->nb_list), NULL);
11994 else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
11995 handle_allocindir_partdone(
11996 WK_ALLOCINDIR(&newblk->nb_list));
11997 else if (newblk->nb_list.wk_type != D_NEWBLK)
11998 panic("handle_written_bmsafemap: Unexpected type: %s",
11999 TYPENAME(newblk->nb_list.wk_type));
12001 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
12002 inodedep->id_state |= DEPCOMPLETE;
12003 inodedep->id_state &= ~ONDEPLIST;
12004 LIST_REMOVE(inodedep, id_deps);
12005 inodedep->id_bmsafemap = NULL;
12007 LIST_REMOVE(bmsafemap, sm_next);
12008 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
12009 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
12010 LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
12011 LIST_EMPTY(&bmsafemap->sm_inodedephd) &&
12012 LIST_EMPTY(&bmsafemap->sm_freehd)) {
12013 LIST_REMOVE(bmsafemap, sm_hash);
12014 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
12017 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
12024 * Try to free a mkdir dependency.
12027 complete_mkdir(mkdir)
12028 struct mkdir *mkdir;
12030 struct diradd *dap;
12032 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
12034 LIST_REMOVE(mkdir, md_mkdirs);
12035 dap = mkdir->md_diradd;
12036 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
12037 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
12038 dap->da_state |= DEPCOMPLETE;
12039 complete_diradd(dap);
12041 WORKITEM_FREE(mkdir, D_MKDIR);
12045 * Handle the completion of a mkdir dependency.
12048 handle_written_mkdir(mkdir, type)
12049 struct mkdir *mkdir;
12053 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
12054 panic("handle_written_mkdir: bad type");
12055 mkdir->md_state |= COMPLETE;
12056 complete_mkdir(mkdir);
12060 free_pagedep(pagedep)
12061 struct pagedep *pagedep;
12065 if (pagedep->pd_state & NEWBLOCK)
12067 if (!LIST_EMPTY(&pagedep->pd_dirremhd))
12069 for (i = 0; i < DAHASHSZ; i++)
12070 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
12072 if (!LIST_EMPTY(&pagedep->pd_pendinghd))
12074 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
12076 if (pagedep->pd_state & ONWORKLIST)
12077 WORKLIST_REMOVE(&pagedep->pd_list);
12078 LIST_REMOVE(pagedep, pd_hash);
12079 WORKITEM_FREE(pagedep, D_PAGEDEP);
12085 * Called from within softdep_disk_write_complete above.
12086 * A write operation was just completed. Removed inodes can
12087 * now be freed and associated block pointers may be committed.
12088 * Note that this routine is always called from interrupt level
12089 * with further interrupts from this device blocked.
12091 * If the write did not succeed, we will do all the roll-forward
12092 * operations, but we will not take the actions that will allow its
12093 * dependencies to be processed.
12096 handle_written_filepage(pagedep, bp, flags)
12097 struct pagedep *pagedep;
12098 struct buf *bp; /* buffer containing the written page */
12101 struct dirrem *dirrem;
12102 struct diradd *dap, *nextdap;
12106 if ((pagedep->pd_state & IOSTARTED) == 0)
12107 panic("handle_written_filepage: not started");
12108 pagedep->pd_state &= ~IOSTARTED;
12109 if ((flags & WRITESUCCEEDED) == 0)
12112 * Process any directory removals that have been committed.
12114 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
12115 LIST_REMOVE(dirrem, dm_next);
12116 dirrem->dm_state |= COMPLETE;
12117 dirrem->dm_dirinum = pagedep->pd_ino;
12118 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
12119 ("handle_written_filepage: Journal entries not written."));
12120 add_to_worklist(&dirrem->dm_list, 0);
12123 * Free any directory additions that have been committed.
12124 * If it is a newly allocated block, we have to wait until
12125 * the on-disk directory inode claims the new block.
12127 if ((pagedep->pd_state & NEWBLOCK) == 0)
12128 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
12129 free_diradd(dap, NULL);
12132 * Uncommitted directory entries must be restored.
12134 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
12135 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
12137 nextdap = LIST_NEXT(dap, da_pdlist);
12138 if (dap->da_state & ATTACHED)
12139 panic("handle_written_filepage: attached");
12140 ep = (struct direct *)
12141 ((char *)bp->b_data + dap->da_offset);
12142 ep->d_ino = dap->da_newinum;
12143 dap->da_state &= ~UNDONE;
12144 dap->da_state |= ATTACHED;
12147 * If the inode referenced by the directory has
12148 * been written out, then the dependency can be
12149 * moved to the pending list.
12151 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
12152 LIST_REMOVE(dap, da_pdlist);
12153 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
12159 * If there were any rollbacks in the directory, then it must be
12160 * marked dirty so that its will eventually get written back in
12161 * its correct form.
12163 if (chgs || (flags & WRITESUCCEEDED) == 0) {
12164 if ((bp->b_flags & B_DELWRI) == 0)
12170 * If we are not waiting for a new directory block to be
12171 * claimed by its inode, then the pagedep will be freed.
12172 * Otherwise it will remain to track any new entries on
12173 * the page in case they are fsync'ed.
12175 free_pagedep(pagedep);
12180 * Writing back in-core inode structures.
12182 * The filesystem only accesses an inode's contents when it occupies an
12183 * "in-core" inode structure. These "in-core" structures are separate from
12184 * the page frames used to cache inode blocks. Only the latter are
12185 * transferred to/from the disk. So, when the updated contents of the
12186 * "in-core" inode structure are copied to the corresponding in-memory inode
12187 * block, the dependencies are also transferred. The following procedure is
12188 * called when copying a dirty "in-core" inode to a cached inode block.
12192 * Called when an inode is loaded from disk. If the effective link count
12193 * differed from the actual link count when it was last flushed, then we
12194 * need to ensure that the correct effective link count is put back.
12197 softdep_load_inodeblock(ip)
12198 struct inode *ip; /* the "in_core" copy of the inode */
12200 struct inodedep *inodedep;
12201 struct ufsmount *ump;
12204 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
12205 ("softdep_load_inodeblock called on non-softdep filesystem"));
12207 * Check for alternate nlink count.
12209 ip->i_effnlink = ip->i_nlink;
12211 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) {
12215 ip->i_effnlink -= inodedep->id_nlinkdelta;
12220 * This routine is called just before the "in-core" inode
12221 * information is to be copied to the in-memory inode block.
12222 * Recall that an inode block contains several inodes. If
12223 * the force flag is set, then the dependencies will be
12224 * cleared so that the update can always be made. Note that
12225 * the buffer is locked when this routine is called, so we
12226 * will never be in the middle of writing the inode block
12230 softdep_update_inodeblock(ip, bp, waitfor)
12231 struct inode *ip; /* the "in_core" copy of the inode */
12232 struct buf *bp; /* the buffer containing the inode block */
12233 int waitfor; /* nonzero => update must be allowed */
12235 struct inodedep *inodedep;
12236 struct inoref *inoref;
12237 struct ufsmount *ump;
12238 struct worklist *wk;
12245 mp = UFSTOVFS(ump);
12246 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
12247 ("softdep_update_inodeblock called on non-softdep filesystem"));
12250 * Preserve the freelink that is on disk. clear_unlinked_inodedep()
12251 * does not have access to the in-core ip so must write directly into
12252 * the inode block buffer when setting freelink.
12254 if (fs->fs_magic == FS_UFS1_MAGIC)
12255 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
12256 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12258 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
12259 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12261 * If the effective link count is not equal to the actual link
12262 * count, then we must track the difference in an inodedep while
12263 * the inode is (potentially) tossed out of the cache. Otherwise,
12264 * if there is no existing inodedep, then there are no dependencies
12269 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12271 if (ip->i_effnlink != ip->i_nlink)
12272 panic("softdep_update_inodeblock: bad link count");
12275 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
12276 panic("softdep_update_inodeblock: bad delta");
12278 * If we're flushing all dependencies we must also move any waiting
12279 * for journal writes onto the bufwait list prior to I/O.
12282 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12283 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12285 jwait(&inoref->if_list, MNT_WAIT);
12291 * Changes have been initiated. Anything depending on these
12292 * changes cannot occur until this inode has been written.
12294 inodedep->id_state &= ~COMPLETE;
12295 if ((inodedep->id_state & ONWORKLIST) == 0)
12296 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
12298 * Any new dependencies associated with the incore inode must
12299 * now be moved to the list associated with the buffer holding
12300 * the in-memory copy of the inode. Once merged process any
12301 * allocdirects that are completed by the merger.
12303 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
12304 if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
12305 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
12307 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
12308 if (!TAILQ_EMPTY(&inodedep->id_extupdt))
12309 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
12312 * Now that the inode has been pushed into the buffer, the
12313 * operations dependent on the inode being written to disk
12314 * can be moved to the id_bufwait so that they will be
12315 * processed when the buffer I/O completes.
12317 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
12318 WORKLIST_REMOVE(wk);
12319 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
12322 * Newly allocated inodes cannot be written until the bitmap
12323 * that allocates them have been written (indicated by
12324 * DEPCOMPLETE being set in id_state). If we are doing a
12325 * forced sync (e.g., an fsync on a file), we force the bitmap
12326 * to be written so that the update can be done.
12328 if (waitfor == 0) {
12333 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
12337 ibp = inodedep->id_bmsafemap->sm_buf;
12338 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT);
12341 * If ibp came back as NULL, the dependency could have been
12342 * freed while we slept. Look it up again, and check to see
12343 * that it has completed.
12345 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
12351 if ((error = bwrite(ibp)) != 0)
12352 softdep_error("softdep_update_inodeblock: bwrite", error);
12356 * Merge the a new inode dependency list (such as id_newinoupdt) into an
12357 * old inode dependency list (such as id_inoupdt).
12360 merge_inode_lists(newlisthead, oldlisthead)
12361 struct allocdirectlst *newlisthead;
12362 struct allocdirectlst *oldlisthead;
12364 struct allocdirect *listadp, *newadp;
12366 newadp = TAILQ_FIRST(newlisthead);
12367 if (newadp != NULL)
12368 LOCK_OWNED(VFSTOUFS(newadp->ad_block.nb_list.wk_mp));
12369 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
12370 if (listadp->ad_offset < newadp->ad_offset) {
12371 listadp = TAILQ_NEXT(listadp, ad_next);
12374 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12375 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
12376 if (listadp->ad_offset == newadp->ad_offset) {
12377 allocdirect_merge(oldlisthead, newadp,
12381 newadp = TAILQ_FIRST(newlisthead);
12383 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
12384 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12385 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
12390 * If we are doing an fsync, then we must ensure that any directory
12391 * entries for the inode have been written after the inode gets to disk.
12395 struct vnode *vp; /* the "in_core" copy of the inode */
12397 struct inodedep *inodedep;
12398 struct pagedep *pagedep;
12399 struct inoref *inoref;
12400 struct ufsmount *ump;
12401 struct worklist *wk;
12402 struct diradd *dap;
12408 struct thread *td = curthread;
12409 int error, flushparent, pagedep_new_block;
12415 ump = VFSTOUFS(mp);
12417 if (MOUNTEDSOFTDEP(mp) == 0)
12421 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12425 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12426 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12428 jwait(&inoref->if_list, MNT_WAIT);
12432 if (!LIST_EMPTY(&inodedep->id_inowait) ||
12433 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
12434 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
12435 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
12436 !TAILQ_EMPTY(&inodedep->id_newinoupdt))
12437 panic("softdep_fsync: pending ops %p", inodedep);
12438 for (error = 0, flushparent = 0; ; ) {
12439 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
12441 if (wk->wk_type != D_DIRADD)
12442 panic("softdep_fsync: Unexpected type %s",
12443 TYPENAME(wk->wk_type));
12444 dap = WK_DIRADD(wk);
12446 * Flush our parent if this directory entry has a MKDIR_PARENT
12447 * dependency or is contained in a newly allocated block.
12449 if (dap->da_state & DIRCHG)
12450 pagedep = dap->da_previous->dm_pagedep;
12452 pagedep = dap->da_pagedep;
12453 parentino = pagedep->pd_ino;
12454 lbn = pagedep->pd_lbn;
12455 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
12456 panic("softdep_fsync: dirty");
12457 if ((dap->da_state & MKDIR_PARENT) ||
12458 (pagedep->pd_state & NEWBLOCK))
12463 * If we are being fsync'ed as part of vgone'ing this vnode,
12464 * then we will not be able to release and recover the
12465 * vnode below, so we just have to give up on writing its
12466 * directory entry out. It will eventually be written, just
12467 * not now, but then the user was not asking to have it
12468 * written, so we are not breaking any promises.
12470 if (vp->v_iflag & VI_DOOMED)
12473 * We prevent deadlock by always fetching inodes from the
12474 * root, moving down the directory tree. Thus, when fetching
12475 * our parent directory, we first try to get the lock. If
12476 * that fails, we must unlock ourselves before requesting
12477 * the lock on our parent. See the comment in ufs_lookup
12478 * for details on possible races.
12481 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
12482 FFSV_FORCEINSMQ)) {
12483 error = vfs_busy(mp, MBF_NOWAIT);
12487 error = vfs_busy(mp, 0);
12488 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12492 if (vp->v_iflag & VI_DOOMED) {
12498 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
12499 &pvp, FFSV_FORCEINSMQ);
12501 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12502 if (vp->v_iflag & VI_DOOMED) {
12511 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
12512 * that are contained in direct blocks will be resolved by
12513 * doing a ffs_update. Pagedeps contained in indirect blocks
12514 * may require a complete sync'ing of the directory. So, we
12515 * try the cheap and fast ffs_update first, and if that fails,
12516 * then we do the slower ffs_syncvnode of the directory.
12521 if ((error = ffs_update(pvp, 1)) != 0) {
12527 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
12528 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
12529 if (wk->wk_type != D_DIRADD)
12530 panic("softdep_fsync: Unexpected type %s",
12531 TYPENAME(wk->wk_type));
12532 dap = WK_DIRADD(wk);
12533 if (dap->da_state & DIRCHG)
12534 pagedep = dap->da_previous->dm_pagedep;
12536 pagedep = dap->da_pagedep;
12537 pagedep_new_block = pagedep->pd_state & NEWBLOCK;
12540 if (pagedep_new_block && (error =
12541 ffs_syncvnode(pvp, MNT_WAIT, 0))) {
12551 * Flush directory page containing the inode's name.
12553 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
12556 error = bwrite(bp);
12563 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
12571 * Flush all the dirty bitmaps associated with the block device
12572 * before flushing the rest of the dirty blocks so as to reduce
12573 * the number of dependencies that will have to be rolled back.
12578 softdep_fsync_mountdev(vp)
12581 struct buf *bp, *nbp;
12582 struct worklist *wk;
12585 if (!vn_isdisk(vp, NULL))
12586 panic("softdep_fsync_mountdev: vnode not a disk");
12587 bo = &vp->v_bufobj;
12590 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
12592 * If it is already scheduled, skip to the next buffer.
12594 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
12597 if ((bp->b_flags & B_DELWRI) == 0)
12598 panic("softdep_fsync_mountdev: not dirty");
12600 * We are only interested in bitmaps with outstanding
12603 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
12604 wk->wk_type != D_BMSAFEMAP ||
12605 (bp->b_vflags & BV_BKGRDINPROG)) {
12611 (void) bawrite(bp);
12619 * Sync all cylinder groups that were dirty at the time this function is
12620 * called. Newly dirtied cgs will be inserted before the sentinel. This
12621 * is used to flush freedep activity that may be holding up writes to a
12625 sync_cgs(mp, waitfor)
12629 struct bmsafemap *bmsafemap;
12630 struct bmsafemap *sentinel;
12631 struct ufsmount *ump;
12635 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK);
12636 sentinel->sm_cg = -1;
12637 ump = VFSTOUFS(mp);
12640 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next);
12641 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL;
12642 bmsafemap = LIST_NEXT(sentinel, sm_next)) {
12643 /* Skip sentinels and cgs with no work to release. */
12644 if (bmsafemap->sm_cg == -1 ||
12645 (LIST_EMPTY(&bmsafemap->sm_freehd) &&
12646 LIST_EMPTY(&bmsafemap->sm_freewr))) {
12647 LIST_REMOVE(sentinel, sm_next);
12648 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12652 * If we don't get the lock and we're waiting try again, if
12653 * not move on to the next buf and try to sync it.
12655 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor);
12656 if (bp == NULL && waitfor == MNT_WAIT)
12658 LIST_REMOVE(sentinel, sm_next);
12659 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12663 if (waitfor == MNT_NOWAIT)
12666 error = bwrite(bp);
12671 LIST_REMOVE(sentinel, sm_next);
12673 free(sentinel, M_BMSAFEMAP);
12678 * This routine is called when we are trying to synchronously flush a
12679 * file. This routine must eliminate any filesystem metadata dependencies
12680 * so that the syncing routine can succeed.
12683 softdep_sync_metadata(struct vnode *vp)
12689 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
12690 ("softdep_sync_metadata called on non-softdep filesystem"));
12692 * Ensure that any direct block dependencies have been cleared,
12693 * truncations are started, and inode references are journaled.
12695 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount));
12697 * Write all journal records to prevent rollbacks on devvp.
12699 if (vp->v_type == VCHR)
12700 softdep_flushjournal(vp->v_mount);
12701 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number);
12703 * Ensure that all truncates are written so we won't find deps on
12706 process_truncates(vp);
12707 FREE_LOCK(VFSTOUFS(vp->v_mount));
12713 * This routine is called when we are attempting to sync a buf with
12714 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any
12715 * other IO it can but returns EBUSY if the buffer is not yet able to
12716 * be written. Dependencies which will not cause rollbacks will always
12720 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
12722 struct indirdep *indirdep;
12723 struct pagedep *pagedep;
12724 struct allocindir *aip;
12725 struct newblk *newblk;
12726 struct ufsmount *ump;
12728 struct worklist *wk;
12731 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
12732 ("softdep_sync_buf called on non-softdep filesystem"));
12734 * For VCHR we just don't want to force flush any dependencies that
12735 * will cause rollbacks.
12737 if (vp->v_type == VCHR) {
12738 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0))
12742 ump = VFSTOUFS(vp->v_mount);
12745 * As we hold the buffer locked, none of its dependencies
12750 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
12751 switch (wk->wk_type) {
12753 case D_ALLOCDIRECT:
12755 newblk = WK_NEWBLK(wk);
12756 if (newblk->nb_jnewblk != NULL) {
12757 if (waitfor == MNT_NOWAIT) {
12761 jwait(&newblk->nb_jnewblk->jn_list, waitfor);
12764 if (newblk->nb_state & DEPCOMPLETE ||
12765 waitfor == MNT_NOWAIT)
12767 nbp = newblk->nb_bmsafemap->sm_buf;
12768 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12772 if ((error = bwrite(nbp)) != 0)
12778 indirdep = WK_INDIRDEP(wk);
12779 if (waitfor == MNT_NOWAIT) {
12780 if (!TAILQ_EMPTY(&indirdep->ir_trunc) ||
12781 !LIST_EMPTY(&indirdep->ir_deplisthd)) {
12786 if (!TAILQ_EMPTY(&indirdep->ir_trunc))
12787 panic("softdep_sync_buf: truncation pending.");
12789 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
12790 newblk = (struct newblk *)aip;
12791 if (newblk->nb_jnewblk != NULL) {
12792 jwait(&newblk->nb_jnewblk->jn_list,
12796 if (newblk->nb_state & DEPCOMPLETE)
12798 nbp = newblk->nb_bmsafemap->sm_buf;
12799 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12803 if ((error = bwrite(nbp)) != 0)
12812 * Only flush directory entries in synchronous passes.
12814 if (waitfor != MNT_WAIT) {
12819 * While syncing snapshots, we must allow recursive
12824 * We are trying to sync a directory that may
12825 * have dependencies on both its own metadata
12826 * and/or dependencies on the inodes of any
12827 * recently allocated files. We walk its diradd
12828 * lists pushing out the associated inode.
12830 pagedep = WK_PAGEDEP(wk);
12831 for (i = 0; i < DAHASHSZ; i++) {
12832 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
12834 if ((error = flush_pagedep_deps(vp, wk->wk_mp,
12835 &pagedep->pd_diraddhd[i]))) {
12850 panic("softdep_sync_buf: Unknown type %s",
12851 TYPENAME(wk->wk_type));
12862 * Flush the dependencies associated with an inodedep.
12865 flush_inodedep_deps(vp, mp, ino)
12870 struct inodedep *inodedep;
12871 struct inoref *inoref;
12872 struct ufsmount *ump;
12873 int error, waitfor;
12876 * This work is done in two passes. The first pass grabs most
12877 * of the buffers and begins asynchronously writing them. The
12878 * only way to wait for these asynchronous writes is to sleep
12879 * on the filesystem vnode which may stay busy for a long time
12880 * if the filesystem is active. So, instead, we make a second
12881 * pass over the dependencies blocking on each write. In the
12882 * usual case we will be blocking against a write that we
12883 * initiated, so when it is done the dependency will have been
12884 * resolved. Thus the second pass is expected to end quickly.
12885 * We give a brief window at the top of the loop to allow
12886 * any pending I/O to complete.
12888 ump = VFSTOUFS(mp);
12890 for (error = 0, waitfor = MNT_NOWAIT; ; ) {
12896 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
12898 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12899 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12901 jwait(&inoref->if_list, MNT_WAIT);
12905 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
12906 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
12907 flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
12908 flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
12911 * If pass2, we are done, otherwise do pass 2.
12913 if (waitfor == MNT_WAIT)
12915 waitfor = MNT_WAIT;
12918 * Try freeing inodedep in case all dependencies have been removed.
12920 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
12921 (void) free_inodedep(inodedep);
12926 * Flush an inode dependency list.
12929 flush_deplist(listhead, waitfor, errorp)
12930 struct allocdirectlst *listhead;
12934 struct allocdirect *adp;
12935 struct newblk *newblk;
12936 struct ufsmount *ump;
12939 if ((adp = TAILQ_FIRST(listhead)) == NULL)
12941 ump = VFSTOUFS(adp->ad_list.wk_mp);
12943 TAILQ_FOREACH(adp, listhead, ad_next) {
12944 newblk = (struct newblk *)adp;
12945 if (newblk->nb_jnewblk != NULL) {
12946 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12949 if (newblk->nb_state & DEPCOMPLETE)
12951 bp = newblk->nb_bmsafemap->sm_buf;
12952 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor);
12954 if (waitfor == MNT_NOWAIT)
12959 if (waitfor == MNT_NOWAIT)
12962 *errorp = bwrite(bp);
12970 * Flush dependencies associated with an allocdirect block.
12973 flush_newblk_dep(vp, mp, lbn)
12978 struct newblk *newblk;
12979 struct ufsmount *ump;
12983 ufs2_daddr_t blkno;
12987 bo = &vp->v_bufobj;
12989 blkno = DIP(ip, i_db[lbn]);
12991 panic("flush_newblk_dep: Missing block");
12992 ump = VFSTOUFS(mp);
12995 * Loop until all dependencies related to this block are satisfied.
12996 * We must be careful to restart after each sleep in case a write
12997 * completes some part of this process for us.
13000 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
13004 if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
13005 panic("flush_newblk_dep: Bad newblk %p", newblk);
13007 * Flush the journal.
13009 if (newblk->nb_jnewblk != NULL) {
13010 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
13014 * Write the bitmap dependency.
13016 if ((newblk->nb_state & DEPCOMPLETE) == 0) {
13017 bp = newblk->nb_bmsafemap->sm_buf;
13018 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
13022 error = bwrite(bp);
13029 * Write the buffer.
13033 bp = gbincore(bo, lbn);
13035 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
13036 LK_INTERLOCK, BO_LOCKPTR(bo));
13037 if (error == ENOLCK) {
13040 continue; /* Slept, retry */
13043 break; /* Failed */
13044 if (bp->b_flags & B_DELWRI) {
13046 error = bwrite(bp);
13054 * We have to wait for the direct pointers to
13055 * point at the newdirblk before the dependency
13058 error = ffs_update(vp, 1);
13067 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
13070 flush_pagedep_deps(pvp, mp, diraddhdp)
13073 struct diraddhd *diraddhdp;
13075 struct inodedep *inodedep;
13076 struct inoref *inoref;
13077 struct ufsmount *ump;
13078 struct diradd *dap;
13083 struct diraddhd unfinished;
13085 LIST_INIT(&unfinished);
13086 ump = VFSTOUFS(mp);
13089 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
13091 * Flush ourselves if this directory entry
13092 * has a MKDIR_PARENT dependency.
13094 if (dap->da_state & MKDIR_PARENT) {
13096 if ((error = ffs_update(pvp, 1)) != 0)
13100 * If that cleared dependencies, go on to next.
13102 if (dap != LIST_FIRST(diraddhdp))
13105 * All MKDIR_PARENT dependencies and all the
13106 * NEWBLOCK pagedeps that are contained in direct
13107 * blocks were resolved by doing above ffs_update.
13108 * Pagedeps contained in indirect blocks may
13109 * require a complete sync'ing of the directory.
13110 * We are in the midst of doing a complete sync,
13111 * so if they are not resolved in this pass we
13112 * defer them for now as they will be sync'ed by
13113 * our caller shortly.
13115 LIST_REMOVE(dap, da_pdlist);
13116 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist);
13120 * A newly allocated directory must have its "." and
13121 * ".." entries written out before its name can be
13122 * committed in its parent.
13124 inum = dap->da_newinum;
13125 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
13126 panic("flush_pagedep_deps: lost inode1");
13128 * Wait for any pending journal adds to complete so we don't
13129 * cause rollbacks while syncing.
13131 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
13132 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
13134 jwait(&inoref->if_list, MNT_WAIT);
13138 if (dap->da_state & MKDIR_BODY) {
13140 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13143 error = flush_newblk_dep(vp, mp, 0);
13145 * If we still have the dependency we might need to
13146 * update the vnode to sync the new link count to
13149 if (error == 0 && dap == LIST_FIRST(diraddhdp))
13150 error = ffs_update(vp, 1);
13156 * If that cleared dependencies, go on to next.
13158 if (dap != LIST_FIRST(diraddhdp))
13160 if (dap->da_state & MKDIR_BODY) {
13161 inodedep_lookup(UFSTOVFS(ump), inum, 0,
13163 panic("flush_pagedep_deps: MKDIR_BODY "
13164 "inodedep %p dap %p vp %p",
13165 inodedep, dap, vp);
13169 * Flush the inode on which the directory entry depends.
13170 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
13171 * the only remaining dependency is that the updated inode
13172 * count must get pushed to disk. The inode has already
13173 * been pushed into its inode buffer (via VOP_UPDATE) at
13174 * the time of the reference count change. So we need only
13175 * locate that buffer, ensure that there will be no rollback
13176 * caused by a bitmap dependency, then write the inode buffer.
13179 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
13180 panic("flush_pagedep_deps: lost inode");
13182 * If the inode still has bitmap dependencies,
13183 * push them to disk.
13185 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
13186 bp = inodedep->id_bmsafemap->sm_buf;
13187 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
13191 if ((error = bwrite(bp)) != 0)
13194 if (dap != LIST_FIRST(diraddhdp))
13198 * If the inode is still sitting in a buffer waiting
13199 * to be written or waiting for the link count to be
13200 * adjusted update it here to flush it to disk.
13202 if (dap == LIST_FIRST(diraddhdp)) {
13204 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13207 error = ffs_update(vp, 1);
13214 * If we have failed to get rid of all the dependencies
13215 * then something is seriously wrong.
13217 if (dap == LIST_FIRST(diraddhdp)) {
13218 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
13219 panic("flush_pagedep_deps: failed to flush "
13220 "inodedep %p ino %ju dap %p",
13221 inodedep, (uintmax_t)inum, dap);
13226 while ((dap = LIST_FIRST(&unfinished)) != NULL) {
13227 LIST_REMOVE(dap, da_pdlist);
13228 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist);
13234 * A large burst of file addition or deletion activity can drive the
13235 * memory load excessively high. First attempt to slow things down
13236 * using the techniques below. If that fails, this routine requests
13237 * the offending operations to fall back to running synchronously
13238 * until the memory load returns to a reasonable level.
13241 softdep_slowdown(vp)
13244 struct ufsmount *ump;
13246 int max_softdeps_hard;
13248 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
13249 ("softdep_slowdown called on non-softdep filesystem"));
13250 ump = VFSTOUFS(vp->v_mount);
13254 * Check for journal space if needed.
13256 if (DOINGSUJ(vp)) {
13257 if (journal_space(ump, 0) == 0)
13261 * If the system is under its limits and our filesystem is
13262 * not responsible for more than our share of the usage and
13263 * we are not low on journal space, then no need to slow down.
13265 max_softdeps_hard = max_softdeps * 11 / 10;
13266 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 &&
13267 dep_current[D_INODEDEP] < max_softdeps_hard &&
13268 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 &&
13269 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 &&
13270 ump->softdep_curdeps[D_DIRREM] <
13271 (max_softdeps_hard / 2) / stat_flush_threads &&
13272 ump->softdep_curdeps[D_INODEDEP] <
13273 max_softdeps_hard / stat_flush_threads &&
13274 ump->softdep_curdeps[D_INDIRDEP] <
13275 (max_softdeps_hard / 1000) / stat_flush_threads &&
13276 ump->softdep_curdeps[D_FREEBLKS] <
13277 max_softdeps_hard / stat_flush_threads) {
13282 * If the journal is low or our filesystem is over its limit
13283 * then speedup the cleanup.
13285 if (ump->softdep_curdeps[D_INDIRDEP] <
13286 (max_softdeps_hard / 1000) / stat_flush_threads || jlow)
13287 softdep_speedup(ump);
13288 stat_sync_limit_hit += 1;
13291 * We only slow down the rate at which new dependencies are
13292 * generated if we are not using journaling. With journaling,
13293 * the cleanup should always be sufficient to keep things
13302 * Called by the allocation routines when they are about to fail
13303 * in the hope that we can free up the requested resource (inodes
13306 * First check to see if the work list has anything on it. If it has,
13307 * clean up entries until we successfully free the requested resource.
13308 * Because this process holds inodes locked, we cannot handle any remove
13309 * requests that might block on a locked inode as that could lead to
13310 * deadlock. If the worklist yields none of the requested resource,
13311 * start syncing out vnodes to free up the needed space.
13314 softdep_request_cleanup(fs, vp, cred, resource)
13317 struct ucred *cred;
13320 struct ufsmount *ump;
13323 ufs2_daddr_t needed;
13324 int error, failed_vnode;
13327 * If we are being called because of a process doing a
13328 * copy-on-write, then it is not safe to process any
13329 * worklist items as we will recurse into the copyonwrite
13330 * routine. This will result in an incoherent snapshot.
13331 * If the vnode that we hold is a snapshot, we must avoid
13332 * handling other resources that could cause deadlock.
13334 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
13337 if (resource == FLUSH_BLOCKS_WAIT)
13338 stat_cleanup_blkrequests += 1;
13340 stat_cleanup_inorequests += 1;
13343 ump = VFSTOUFS(mp);
13344 mtx_assert(UFS_MTX(ump), MA_OWNED);
13346 error = ffs_update(vp, 1);
13347 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) {
13352 * If we are in need of resources, start by cleaning up
13353 * any block removals associated with our inode.
13356 process_removes(vp);
13357 process_truncates(vp);
13360 * Now clean up at least as many resources as we will need.
13362 * When requested to clean up inodes, the number that are needed
13363 * is set by the number of simultaneous writers (mnt_writeopcount)
13364 * plus a bit of slop (2) in case some more writers show up while
13367 * When requested to free up space, the amount of space that
13368 * we need is enough blocks to allocate a full-sized segment
13369 * (fs_contigsumsize). The number of such segments that will
13370 * be needed is set by the number of simultaneous writers
13371 * (mnt_writeopcount) plus a bit of slop (2) in case some more
13372 * writers show up while we are cleaning.
13374 * Additionally, if we are unpriviledged and allocating space,
13375 * we need to ensure that we clean up enough blocks to get the
13376 * needed number of blocks over the threshold of the minimum
13377 * number of blocks required to be kept free by the filesystem
13380 if (resource == FLUSH_INODES_WAIT) {
13381 needed = vp->v_mount->mnt_writeopcount + 2;
13382 } else if (resource == FLUSH_BLOCKS_WAIT) {
13383 needed = (vp->v_mount->mnt_writeopcount + 2) *
13384 fs->fs_contigsumsize;
13385 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0))
13386 needed += fragstoblks(fs,
13387 roundup((fs->fs_dsize * fs->fs_minfree / 100) -
13388 fs->fs_cstotal.cs_nffree, fs->fs_frag));
13391 printf("softdep_request_cleanup: Unknown resource type %d\n",
13395 starttime = time_second;
13397 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 &&
13398 fs->fs_cstotal.cs_nbfree <= needed) ||
13399 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13400 fs->fs_cstotal.cs_nifree <= needed)) {
13402 if (ump->softdep_on_worklist > 0 &&
13403 process_worklist_item(UFSTOVFS(ump),
13404 ump->softdep_on_worklist, LK_NOWAIT) != 0)
13405 stat_worklist_push += 1;
13409 * If we still need resources and there are no more worklist
13410 * entries to process to obtain them, we have to start flushing
13411 * the dirty vnodes to force the release of additional requests
13412 * to the worklist that we can then process to reap addition
13413 * resources. We walk the vnodes associated with the mount point
13414 * until we get the needed worklist requests that we can reap.
13416 * If there are several threads all needing to clean the same
13417 * mount point, only one is allowed to walk the mount list.
13418 * When several threads all try to walk the same mount list,
13419 * they end up competing with each other and often end up in
13420 * livelock. This approach ensures that forward progress is
13421 * made at the cost of occational ENOSPC errors being returned
13422 * that might otherwise have been avoided.
13425 if ((resource == FLUSH_BLOCKS_WAIT &&
13426 fs->fs_cstotal.cs_nbfree <= needed) ||
13427 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13428 fs->fs_cstotal.cs_nifree <= needed)) {
13430 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) {
13431 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE;
13433 failed_vnode = softdep_request_cleanup_flush(mp, ump);
13435 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE;
13437 if (ump->softdep_on_worklist > 0) {
13438 stat_cleanup_retries += 1;
13446 stat_cleanup_failures += 1;
13448 if (time_second - starttime > stat_cleanup_high_delay)
13449 stat_cleanup_high_delay = time_second - starttime;
13455 * Scan the vnodes for the specified mount point flushing out any
13456 * vnodes that can be locked without waiting. Finally, try to flush
13457 * the device associated with the mount point if it can be locked
13460 * We return 0 if we were able to lock every vnode in our scan.
13461 * If we had to skip one or more vnodes, we return 1.
13464 softdep_request_cleanup_flush(mp, ump)
13466 struct ufsmount *ump;
13469 struct vnode *lvp, *mvp;
13474 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) {
13475 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) {
13479 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
13484 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */
13488 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0);
13491 lvp = ump->um_devvp;
13492 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
13493 VOP_FSYNC(lvp, MNT_NOWAIT, td);
13494 VOP_UNLOCK(lvp, 0);
13496 return (failed_vnode);
13500 softdep_excess_items(struct ufsmount *ump, int item)
13503 KASSERT(item >= 0 && item < D_LAST, ("item %d", item));
13504 return (dep_current[item] > max_softdeps &&
13505 ump->softdep_curdeps[item] > max_softdeps /
13506 stat_flush_threads);
13510 schedule_cleanup(struct mount *mp)
13512 struct ufsmount *ump;
13515 ump = VFSTOUFS(mp);
13519 if ((td->td_pflags & TDP_KTHREAD) != 0 &&
13520 (td->td_proc->p_flag2 & P2_AST_SU) == 0) {
13522 * No ast is delivered to kernel threads, so nobody
13523 * would deref the mp. Some kernel threads
13524 * explicitely check for AST, e.g. NFS daemon does
13525 * this in the serving loop.
13529 if (td->td_su != NULL)
13530 vfs_rel(td->td_su);
13534 td->td_flags |= TDF_ASTPENDING;
13539 softdep_ast_cleanup_proc(struct thread *td)
13542 struct ufsmount *ump;
13546 while ((mp = td->td_su) != NULL) {
13548 error = vfs_busy(mp, MBF_NOWAIT);
13552 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) {
13553 ump = VFSTOUFS(mp);
13557 if (softdep_excess_items(ump, D_INODEDEP)) {
13559 request_cleanup(mp, FLUSH_INODES);
13561 if (softdep_excess_items(ump, D_DIRREM)) {
13563 request_cleanup(mp, FLUSH_BLOCKS);
13566 if (softdep_excess_items(ump, D_NEWBLK) ||
13567 softdep_excess_items(ump, D_ALLOCDIRECT) ||
13568 softdep_excess_items(ump, D_ALLOCINDIR)) {
13569 error = vn_start_write(NULL, &mp,
13573 VFS_SYNC(mp, MNT_WAIT);
13574 vn_finished_write(mp);
13577 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req)
13583 if ((mp = td->td_su) != NULL) {
13590 * If memory utilization has gotten too high, deliberately slow things
13591 * down and speed up the I/O processing.
13594 request_cleanup(mp, resource)
13598 struct thread *td = curthread;
13599 struct ufsmount *ump;
13601 ump = VFSTOUFS(mp);
13604 * We never hold up the filesystem syncer or buf daemon.
13606 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
13609 * First check to see if the work list has gotten backlogged.
13610 * If it has, co-opt this process to help clean up two entries.
13611 * Because this process may hold inodes locked, we cannot
13612 * handle any remove requests that might block on a locked
13613 * inode as that could lead to deadlock. We set TDP_SOFTDEP
13614 * to avoid recursively processing the worklist.
13616 if (ump->softdep_on_worklist > max_softdeps / 10) {
13617 td->td_pflags |= TDP_SOFTDEP;
13618 process_worklist_item(mp, 2, LK_NOWAIT);
13619 td->td_pflags &= ~TDP_SOFTDEP;
13620 stat_worklist_push += 2;
13624 * Next, we attempt to speed up the syncer process. If that
13625 * is successful, then we allow the process to continue.
13627 if (softdep_speedup(ump) &&
13628 resource != FLUSH_BLOCKS_WAIT &&
13629 resource != FLUSH_INODES_WAIT)
13632 * If we are resource constrained on inode dependencies, try
13633 * flushing some dirty inodes. Otherwise, we are constrained
13634 * by file deletions, so try accelerating flushes of directories
13635 * with removal dependencies. We would like to do the cleanup
13636 * here, but we probably hold an inode locked at this point and
13637 * that might deadlock against one that we try to clean. So,
13638 * the best that we can do is request the syncer daemon to do
13639 * the cleanup for us.
13641 switch (resource) {
13644 case FLUSH_INODES_WAIT:
13645 ACQUIRE_GBLLOCK(&lk);
13646 stat_ino_limit_push += 1;
13647 req_clear_inodedeps += 1;
13649 stat_countp = &stat_ino_limit_hit;
13653 case FLUSH_BLOCKS_WAIT:
13654 ACQUIRE_GBLLOCK(&lk);
13655 stat_blk_limit_push += 1;
13656 req_clear_remove += 1;
13658 stat_countp = &stat_blk_limit_hit;
13662 panic("request_cleanup: unknown type");
13665 * Hopefully the syncer daemon will catch up and awaken us.
13666 * We wait at most tickdelay before proceeding in any case.
13668 ACQUIRE_GBLLOCK(&lk);
13671 if (callout_pending(&softdep_callout) == FALSE)
13672 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13675 if ((td->td_pflags & TDP_KTHREAD) == 0)
13676 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
13684 * Awaken processes pausing in request_cleanup and clear proc_waiting
13685 * to indicate that there is no longer a timer running. Pause_timer
13686 * will be called with the global softdep mutex (&lk) locked.
13693 GBLLOCK_OWNED(&lk);
13695 * The callout_ API has acquired mtx and will hold it around this
13698 *stat_countp += proc_waiting;
13699 wakeup(&proc_waiting);
13703 * If requested, try removing inode or removal dependencies.
13706 check_clear_deps(mp)
13711 * If we are suspended, it may be because of our using
13712 * too many inodedeps, so help clear them out.
13714 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended)
13715 clear_inodedeps(mp);
13717 * General requests for cleanup of backed up dependencies
13719 ACQUIRE_GBLLOCK(&lk);
13720 if (req_clear_inodedeps) {
13721 req_clear_inodedeps -= 1;
13723 clear_inodedeps(mp);
13724 ACQUIRE_GBLLOCK(&lk);
13725 wakeup(&proc_waiting);
13727 if (req_clear_remove) {
13728 req_clear_remove -= 1;
13731 ACQUIRE_GBLLOCK(&lk);
13732 wakeup(&proc_waiting);
13738 * Flush out a directory with at least one removal dependency in an effort to
13739 * reduce the number of dirrem, freefile, and freeblks dependency structures.
13745 struct pagedep_hashhead *pagedephd;
13746 struct pagedep *pagedep;
13747 struct ufsmount *ump;
13753 ump = VFSTOUFS(mp);
13756 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) {
13757 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++];
13758 if (ump->pagedep_nextclean > ump->pagedep_hash_size)
13759 ump->pagedep_nextclean = 0;
13760 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
13761 if (LIST_EMPTY(&pagedep->pd_dirremhd))
13763 ino = pagedep->pd_ino;
13764 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13769 * Let unmount clear deps
13771 error = vfs_busy(mp, MBF_NOWAIT);
13774 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13778 softdep_error("clear_remove: vget", error);
13781 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13782 softdep_error("clear_remove: fsync", error);
13783 bo = &vp->v_bufobj;
13789 vn_finished_write(mp);
13797 * Clear out a block of dirty inodes in an effort to reduce
13798 * the number of inodedep dependency structures.
13801 clear_inodedeps(mp)
13804 struct inodedep_hashhead *inodedephd;
13805 struct inodedep *inodedep;
13806 struct ufsmount *ump;
13810 ino_t firstino, lastino, ino;
13812 ump = VFSTOUFS(mp);
13816 * Pick a random inode dependency to be cleared.
13817 * We will then gather up all the inodes in its block
13818 * that have dependencies and flush them out.
13820 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) {
13821 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++];
13822 if (ump->inodedep_nextclean > ump->inodedep_hash_size)
13823 ump->inodedep_nextclean = 0;
13824 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
13827 if (inodedep == NULL)
13830 * Find the last inode in the block with dependencies.
13832 firstino = rounddown2(inodedep->id_ino, INOPB(fs));
13833 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
13834 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
13837 * Asynchronously push all but the last inode with dependencies.
13838 * Synchronously push the last inode with dependencies to ensure
13839 * that the inode block gets written to free up the inodedeps.
13841 for (ino = firstino; ino <= lastino; ino++) {
13842 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
13844 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13847 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
13849 vn_finished_write(mp);
13853 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13854 FFSV_FORCEINSMQ)) != 0) {
13855 softdep_error("clear_inodedeps: vget", error);
13857 vn_finished_write(mp);
13862 if (ino == lastino) {
13863 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)))
13864 softdep_error("clear_inodedeps: fsync1", error);
13866 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13867 softdep_error("clear_inodedeps: fsync2", error);
13868 BO_LOCK(&vp->v_bufobj);
13870 BO_UNLOCK(&vp->v_bufobj);
13873 vn_finished_write(mp);
13879 softdep_buf_append(bp, wkhd)
13881 struct workhead *wkhd;
13883 struct worklist *wk;
13884 struct ufsmount *ump;
13886 if ((wk = LIST_FIRST(wkhd)) == NULL)
13888 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13889 ("softdep_buf_append called on non-softdep filesystem"));
13890 ump = VFSTOUFS(wk->wk_mp);
13892 while ((wk = LIST_FIRST(wkhd)) != NULL) {
13893 WORKLIST_REMOVE(wk);
13894 WORKLIST_INSERT(&bp->b_dep, wk);
13901 softdep_inode_append(ip, cred, wkhd)
13903 struct ucred *cred;
13904 struct workhead *wkhd;
13908 struct ufsmount *ump;
13912 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
13913 ("softdep_inode_append called on non-softdep filesystem"));
13915 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
13916 (int)fs->fs_bsize, cred, &bp);
13919 softdep_freework(wkhd);
13922 softdep_buf_append(bp, wkhd);
13927 softdep_freework(wkhd)
13928 struct workhead *wkhd;
13930 struct worklist *wk;
13931 struct ufsmount *ump;
13933 if ((wk = LIST_FIRST(wkhd)) == NULL)
13935 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13936 ("softdep_freework called on non-softdep filesystem"));
13937 ump = VFSTOUFS(wk->wk_mp);
13939 handle_jwork(wkhd);
13943 static struct ufsmount *
13944 softdep_bp_to_mp(bp)
13950 if (LIST_EMPTY(&bp->b_dep))
13953 KASSERT(vp != NULL,
13954 ("%s, buffer with dependencies lacks vnode", __func__));
13957 * The ump mount point is stable after we get a correct
13958 * pointer, since bp is locked and this prevents unmount from
13959 * proceeding. But to get to it, we cannot dereference bp->b_dep
13960 * head wk_mp, because we do not yet own SU ump lock and
13961 * workitem might be freed while dereferenced.
13964 switch (vp->v_type) {
13967 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL;
13980 vn_printf(vp, "softdep_bp_to_mp: unexpected block device\n");
13988 vn_printf(vp, "unknown vnode type");
13992 return (VFSTOUFS(mp));
13996 * Function to determine if the buffer has outstanding dependencies
13997 * that will cause a roll-back if the buffer is written. If wantcount
13998 * is set, return number of dependencies, otherwise just yes or no.
14001 softdep_count_dependencies(bp, wantcount)
14005 struct worklist *wk;
14006 struct ufsmount *ump;
14007 struct bmsafemap *bmsafemap;
14008 struct freework *freework;
14009 struct inodedep *inodedep;
14010 struct indirdep *indirdep;
14011 struct freeblks *freeblks;
14012 struct allocindir *aip;
14013 struct pagedep *pagedep;
14014 struct dirrem *dirrem;
14015 struct newblk *newblk;
14016 struct mkdir *mkdir;
14017 struct diradd *dap;
14020 ump = softdep_bp_to_mp(bp);
14025 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
14026 switch (wk->wk_type) {
14029 inodedep = WK_INODEDEP(wk);
14030 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
14031 /* bitmap allocation dependency */
14036 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
14037 /* direct block pointer dependency */
14042 if (TAILQ_FIRST(&inodedep->id_extupdt)) {
14043 /* direct block pointer dependency */
14048 if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
14049 /* Add reference dependency. */
14057 indirdep = WK_INDIRDEP(wk);
14059 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) {
14060 /* indirect truncation dependency */
14066 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
14067 /* indirect block pointer dependency */
14075 pagedep = WK_PAGEDEP(wk);
14076 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
14077 if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
14078 /* Journal remove ref dependency. */
14084 for (i = 0; i < DAHASHSZ; i++) {
14086 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
14087 /* directory entry dependency */
14096 bmsafemap = WK_BMSAFEMAP(wk);
14097 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
14098 /* Add reference dependency. */
14103 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
14104 /* Allocate block dependency. */
14112 freeblks = WK_FREEBLKS(wk);
14113 if (LIST_FIRST(&freeblks->fb_jblkdephd)) {
14114 /* Freeblk journal dependency. */
14121 case D_ALLOCDIRECT:
14123 newblk = WK_NEWBLK(wk);
14124 if (newblk->nb_jnewblk) {
14125 /* Journal allocate dependency. */
14133 mkdir = WK_MKDIR(wk);
14134 if (mkdir->md_jaddref) {
14135 /* Journal reference dependency. */
14147 /* never a dependency on these blocks */
14151 panic("softdep_count_dependencies: Unexpected type %s",
14152 TYPENAME(wk->wk_type));
14162 * Acquire exclusive access to a buffer.
14163 * Must be called with a locked mtx parameter.
14164 * Return acquired buffer or NULL on failure.
14166 static struct buf *
14167 getdirtybuf(bp, lock, waitfor)
14169 struct rwlock *lock;
14174 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
14175 if (waitfor != MNT_WAIT)
14177 error = BUF_LOCK(bp,
14178 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock);
14180 * Even if we successfully acquire bp here, we have dropped
14181 * lock, which may violates our guarantee.
14185 else if (error != ENOLCK)
14186 panic("getdirtybuf: inconsistent lock: %d", error);
14190 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
14191 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) {
14193 BO_LOCK(bp->b_bufobj);
14195 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
14196 bp->b_vflags |= BV_BKGRDWAIT;
14197 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj),
14198 PRIBIO | PDROP, "getbuf", 0);
14200 BO_UNLOCK(bp->b_bufobj);
14205 if (waitfor != MNT_WAIT)
14207 #ifdef DEBUG_VFS_LOCKS
14208 if (bp->b_vp->v_type != VCHR)
14209 ASSERT_BO_WLOCKED(bp->b_bufobj);
14211 bp->b_vflags |= BV_BKGRDWAIT;
14212 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0);
14215 if ((bp->b_flags & B_DELWRI) == 0) {
14225 * Check if it is safe to suspend the file system now. On entry,
14226 * the vnode interlock for devvp should be held. Return 0 with
14227 * the mount interlock held if the file system can be suspended now,
14228 * otherwise return EAGAIN with the mount interlock held.
14231 softdep_check_suspend(struct mount *mp,
14232 struct vnode *devvp,
14233 int softdep_depcnt,
14234 int softdep_accdepcnt,
14235 int secondary_writes,
14236 int secondary_accwrites)
14239 struct ufsmount *ump;
14240 struct inodedep *inodedep;
14241 int error, unlinked;
14243 bo = &devvp->v_bufobj;
14244 ASSERT_BO_WLOCKED(bo);
14247 * If we are not running with soft updates, then we need only
14248 * deal with secondary writes as we try to suspend.
14250 if (MOUNTEDSOFTDEP(mp) == 0) {
14252 while (mp->mnt_secondary_writes != 0) {
14254 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
14255 (PUSER - 1) | PDROP, "secwr", 0);
14261 * Reasons for needing more work before suspend:
14262 * - Dirty buffers on devvp.
14263 * - Secondary writes occurred after start of vnode sync loop
14266 if (bo->bo_numoutput > 0 ||
14267 bo->bo_dirty.bv_cnt > 0 ||
14268 secondary_writes != 0 ||
14269 mp->mnt_secondary_writes != 0 ||
14270 secondary_accwrites != mp->mnt_secondary_accwrites)
14277 * If we are running with soft updates, then we need to coordinate
14278 * with them as we try to suspend.
14280 ump = VFSTOUFS(mp);
14282 if (!TRY_ACQUIRE_LOCK(ump)) {
14290 if (mp->mnt_secondary_writes != 0) {
14293 msleep(&mp->mnt_secondary_writes,
14295 (PUSER - 1) | PDROP, "secwr", 0);
14303 if (MOUNTEDSUJ(mp)) {
14304 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked);
14306 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
14307 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS |
14308 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS |
14310 !check_inodedep_free(inodedep))
14317 * Reasons for needing more work before suspend:
14318 * - Dirty buffers on devvp.
14319 * - Softdep activity occurred after start of vnode sync loop
14320 * - Secondary writes occurred after start of vnode sync loop
14323 if (bo->bo_numoutput > 0 ||
14324 bo->bo_dirty.bv_cnt > 0 ||
14325 softdep_depcnt != unlinked ||
14326 ump->softdep_deps != unlinked ||
14327 softdep_accdepcnt != ump->softdep_accdeps ||
14328 secondary_writes != 0 ||
14329 mp->mnt_secondary_writes != 0 ||
14330 secondary_accwrites != mp->mnt_secondary_accwrites)
14339 * Get the number of dependency structures for the file system, both
14340 * the current number and the total number allocated. These will
14341 * later be used to detect that softdep processing has occurred.
14344 softdep_get_depcounts(struct mount *mp,
14345 int *softdep_depsp,
14346 int *softdep_accdepsp)
14348 struct ufsmount *ump;
14350 if (MOUNTEDSOFTDEP(mp) == 0) {
14351 *softdep_depsp = 0;
14352 *softdep_accdepsp = 0;
14355 ump = VFSTOUFS(mp);
14357 *softdep_depsp = ump->softdep_deps;
14358 *softdep_accdepsp = ump->softdep_accdeps;
14363 * Wait for pending output on a vnode to complete.
14370 ASSERT_VOP_LOCKED(vp, "drain_output");
14371 (void)bufobj_wwait(&vp->v_bufobj, 0, 0);
14375 * Called whenever a buffer that is being invalidated or reallocated
14376 * contains dependencies. This should only happen if an I/O error has
14377 * occurred. The routine is called with the buffer locked.
14380 softdep_deallocate_dependencies(bp)
14384 if ((bp->b_ioflags & BIO_ERROR) == 0)
14385 panic("softdep_deallocate_dependencies: dangling deps");
14386 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL)
14387 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
14389 printf("softdep_deallocate_dependencies: "
14390 "got error %d while accessing filesystem\n", bp->b_error);
14391 if (bp->b_error != ENXIO)
14392 panic("softdep_deallocate_dependencies: unrecovered I/O error");
14396 * Function to handle asynchronous write errors in the filesystem.
14399 softdep_error(func, error)
14404 /* XXX should do something better! */
14405 printf("%s: got error %d while accessing filesystem\n", func, error);
14411 inodedep_print(struct inodedep *inodedep, int verbose)
14413 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %jd nlink %jd"
14415 inodedep, inodedep->id_fs, inodedep->id_state,
14416 (intmax_t)inodedep->id_ino,
14417 (intmax_t)fsbtodb(inodedep->id_fs,
14418 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
14419 (intmax_t)inodedep->id_nlinkdelta,
14420 (intmax_t)inodedep->id_savednlink,
14421 inodedep->id_savedino1);
14426 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
14428 LIST_FIRST(&inodedep->id_pendinghd),
14429 LIST_FIRST(&inodedep->id_bufwait),
14430 LIST_FIRST(&inodedep->id_inowait),
14431 TAILQ_FIRST(&inodedep->id_inoreflst),
14432 inodedep->id_mkdiradd);
14433 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
14434 TAILQ_FIRST(&inodedep->id_inoupdt),
14435 TAILQ_FIRST(&inodedep->id_newinoupdt),
14436 TAILQ_FIRST(&inodedep->id_extupdt),
14437 TAILQ_FIRST(&inodedep->id_newextupdt));
14440 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
14443 if (have_addr == 0) {
14444 db_printf("Address required\n");
14447 inodedep_print((struct inodedep*)addr, 1);
14450 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
14452 struct inodedep_hashhead *inodedephd;
14453 struct inodedep *inodedep;
14454 struct ufsmount *ump;
14457 if (have_addr == 0) {
14458 db_printf("Address required\n");
14461 ump = (struct ufsmount *)addr;
14462 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) {
14463 inodedephd = &ump->inodedep_hashtbl[cnt];
14464 LIST_FOREACH(inodedep, inodedephd, id_hash) {
14465 inodedep_print(inodedep, 0);
14470 DB_SHOW_COMMAND(worklist, db_show_worklist)
14472 struct worklist *wk;
14474 if (have_addr == 0) {
14475 db_printf("Address required\n");
14478 wk = (struct worklist *)addr;
14479 printf("worklist: %p type %s state 0x%X\n",
14480 wk, TYPENAME(wk->wk_type), wk->wk_state);
14483 DB_SHOW_COMMAND(workhead, db_show_workhead)
14485 struct workhead *wkhd;
14486 struct worklist *wk;
14489 if (have_addr == 0) {
14490 db_printf("Address required\n");
14493 wkhd = (struct workhead *)addr;
14494 wk = LIST_FIRST(wkhd);
14495 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
14496 db_printf("worklist: %p type %s state 0x%X",
14497 wk, TYPENAME(wk->wk_type), wk->wk_state);
14499 db_printf("workhead overflow");
14504 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
14506 struct mkdirlist *mkdirlisthd;
14507 struct jaddref *jaddref;
14508 struct diradd *diradd;
14509 struct mkdir *mkdir;
14511 if (have_addr == 0) {
14512 db_printf("Address required\n");
14515 mkdirlisthd = (struct mkdirlist *)addr;
14516 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) {
14517 diradd = mkdir->md_diradd;
14518 db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
14519 mkdir, mkdir->md_state, diradd, diradd->da_state);
14520 if ((jaddref = mkdir->md_jaddref) != NULL)
14521 db_printf(" jaddref %p jaddref state 0x%X",
14522 jaddref, jaddref->ja_state);
14527 /* exported to ffs_vfsops.c */
14528 extern void db_print_ffs(struct ufsmount *ump);
14530 db_print_ffs(struct ufsmount *ump)
14532 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n",
14533 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname,
14534 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist,
14535 ump->softdep_deps, ump->softdep_req);
14540 #endif /* SOFTUPDATES */