2 * Copyright 1998, 2000 Marshall Kirk McKusick.
3 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org>
6 * The soft updates code is derived from the appendix of a University
7 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
8 * "Soft Updates: A Solution to the Metadata Update Problem in File
9 * Systems", CSE-TR-254-95, August 1995).
11 * Further information about soft updates can be obtained from:
13 * Marshall Kirk McKusick http://www.mckusick.com/softdep/
14 * 1614 Oxford Street mckusick@mckusick.com
15 * Berkeley, CA 94709-1608 +1-510-843-9542
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions
22 * 1. Redistributions of source code must retain the above copyright
23 * notice, this list of conditions and the following disclaimer.
24 * 2. Redistributions in binary form must reproduce the above copyright
25 * notice, this list of conditions and the following disclaimer in the
26 * documentation and/or other materials provided with the distribution.
28 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
29 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
30 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
31 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT,
32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
34 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
35 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
36 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
37 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00
42 #include <sys/cdefs.h>
43 __FBSDID("$FreeBSD$");
46 #include "opt_quota.h"
50 * For now we want the safety net that the DEBUG flag provides.
56 #include <sys/param.h>
57 #include <sys/kernel.h>
58 #include <sys/systm.h>
62 #include <sys/kthread.h>
64 #include <sys/limits.h>
66 #include <sys/malloc.h>
67 #include <sys/mount.h>
68 #include <sys/mutex.h>
69 #include <sys/namei.h>
72 #include <sys/rwlock.h>
74 #include <sys/sysctl.h>
75 #include <sys/syslog.h>
76 #include <sys/vnode.h>
79 #include <ufs/ufs/dir.h>
80 #include <ufs/ufs/extattr.h>
81 #include <ufs/ufs/quota.h>
82 #include <ufs/ufs/inode.h>
83 #include <ufs/ufs/ufsmount.h>
84 #include <ufs/ffs/fs.h>
85 #include <ufs/ffs/softdep.h>
86 #include <ufs/ffs/ffs_extern.h>
87 #include <ufs/ufs/ufs_extern.h>
90 #include <vm/vm_extern.h>
91 #include <vm/vm_object.h>
93 #include <geom/geom.h>
97 #define KTR_SUJ 0 /* Define to KTR_SPARE. */
102 softdep_flushfiles(oldmnt, flags, td)
103 struct mount *oldmnt;
108 panic("softdep_flushfiles called");
112 softdep_mount(devvp, mp, fs, cred)
130 softdep_uninitialize()
141 panic("softdep_unmount called");
145 softdep_setup_sbupdate(ump, fs, bp)
146 struct ufsmount *ump;
151 panic("softdep_setup_sbupdate called");
155 softdep_setup_inomapdep(bp, ip, newinum, mode)
162 panic("softdep_setup_inomapdep called");
166 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
169 ufs2_daddr_t newblkno;
174 panic("softdep_setup_blkmapdep called");
178 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
181 ufs2_daddr_t newblkno;
182 ufs2_daddr_t oldblkno;
188 panic("softdep_setup_allocdirect called");
192 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
195 ufs2_daddr_t newblkno;
196 ufs2_daddr_t oldblkno;
202 panic("softdep_setup_allocext called");
206 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
211 ufs2_daddr_t newblkno;
212 ufs2_daddr_t oldblkno;
216 panic("softdep_setup_allocindir_page called");
220 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
225 ufs2_daddr_t newblkno;
228 panic("softdep_setup_allocindir_meta called");
232 softdep_journal_freeblocks(ip, cred, length, flags)
239 panic("softdep_journal_freeblocks called");
243 softdep_journal_fsync(ip)
247 panic("softdep_journal_fsync called");
251 softdep_setup_freeblocks(ip, length, flags)
257 panic("softdep_setup_freeblocks called");
261 softdep_freefile(pvp, ino, mode)
267 panic("softdep_freefile called");
271 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
276 struct buf *newdirbp;
280 panic("softdep_setup_directory_add called");
284 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
293 panic("softdep_change_directoryentry_offset called");
297 softdep_setup_remove(bp, dp, ip, isrmdir)
304 panic("softdep_setup_remove called");
308 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
316 panic("softdep_setup_directory_change called");
320 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
325 struct workhead *wkhd;
328 panic("%s called", __FUNCTION__);
332 softdep_setup_inofree(mp, bp, ino, wkhd)
336 struct workhead *wkhd;
339 panic("%s called", __FUNCTION__);
343 softdep_setup_unlink(dp, ip)
348 panic("%s called", __FUNCTION__);
352 softdep_setup_link(dp, ip)
357 panic("%s called", __FUNCTION__);
361 softdep_revert_link(dp, ip)
366 panic("%s called", __FUNCTION__);
370 softdep_setup_rmdir(dp, ip)
375 panic("%s called", __FUNCTION__);
379 softdep_revert_rmdir(dp, ip)
384 panic("%s called", __FUNCTION__);
388 softdep_setup_create(dp, ip)
393 panic("%s called", __FUNCTION__);
397 softdep_revert_create(dp, ip)
402 panic("%s called", __FUNCTION__);
406 softdep_setup_mkdir(dp, ip)
411 panic("%s called", __FUNCTION__);
415 softdep_revert_mkdir(dp, ip)
420 panic("%s called", __FUNCTION__);
424 softdep_setup_dotdot_link(dp, ip)
429 panic("%s called", __FUNCTION__);
433 softdep_prealloc(vp, waitok)
438 panic("%s called", __FUNCTION__);
442 softdep_journal_lookup(mp, vpp)
451 softdep_change_linkcnt(ip)
455 panic("softdep_change_linkcnt called");
459 softdep_load_inodeblock(ip)
463 panic("softdep_load_inodeblock called");
467 softdep_update_inodeblock(ip, bp, waitfor)
473 panic("softdep_update_inodeblock called");
478 struct vnode *vp; /* the "in_core" copy of the inode */
485 softdep_fsync_mountdev(vp)
493 softdep_flushworklist(oldmnt, countp, td)
494 struct mount *oldmnt;
504 softdep_sync_metadata(struct vnode *vp)
507 panic("softdep_sync_metadata called");
511 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
514 panic("softdep_sync_buf called");
522 panic("softdep_slowdown called");
526 softdep_request_cleanup(fs, vp, cred, resource)
537 softdep_check_suspend(struct mount *mp,
540 int softdep_accdepcnt,
541 int secondary_writes,
542 int secondary_accwrites)
547 (void) softdep_depcnt,
548 (void) softdep_accdepcnt;
550 bo = &devvp->v_bufobj;
551 ASSERT_BO_WLOCKED(bo);
554 while (mp->mnt_secondary_writes != 0) {
556 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
557 (PUSER - 1) | PDROP, "secwr", 0);
563 * Reasons for needing more work before suspend:
564 * - Dirty buffers on devvp.
565 * - Secondary writes occurred after start of vnode sync loop
568 if (bo->bo_numoutput > 0 ||
569 bo->bo_dirty.bv_cnt > 0 ||
570 secondary_writes != 0 ||
571 mp->mnt_secondary_writes != 0 ||
572 secondary_accwrites != mp->mnt_secondary_accwrites)
579 softdep_get_depcounts(struct mount *mp,
581 int *softdepactiveaccp)
585 *softdepactiveaccp = 0;
589 softdep_buf_append(bp, wkhd)
591 struct workhead *wkhd;
594 panic("softdep_buf_appendwork called");
598 softdep_inode_append(ip, cred, wkhd)
601 struct workhead *wkhd;
604 panic("softdep_inode_appendwork called");
608 softdep_freework(wkhd)
609 struct workhead *wkhd;
612 panic("softdep_freework called");
617 FEATURE(softupdates, "FFS soft-updates support");
619 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0,
620 "soft updates stats");
621 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0,
622 "total dependencies allocated");
623 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0,
624 "high use dependencies allocated");
625 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0,
626 "current dependencies allocated");
627 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0,
628 "current dependencies written");
630 unsigned long dep_current[D_LAST + 1];
631 unsigned long dep_highuse[D_LAST + 1];
632 unsigned long dep_total[D_LAST + 1];
633 unsigned long dep_write[D_LAST + 1];
635 #define SOFTDEP_TYPE(type, str, long) \
636 static MALLOC_DEFINE(M_ ## type, #str, long); \
637 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \
638 &dep_total[D_ ## type], 0, ""); \
639 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \
640 &dep_current[D_ ## type], 0, ""); \
641 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \
642 &dep_highuse[D_ ## type], 0, ""); \
643 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \
644 &dep_write[D_ ## type], 0, "");
646 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies");
647 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies");
648 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap,
649 "Block or frag allocated from cyl group map");
650 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency");
651 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode");
652 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies");
653 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block");
654 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode");
655 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode");
656 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated");
657 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry");
658 SOFTDEP_TYPE(MKDIR, mkdir, "New directory");
659 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted");
660 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block");
661 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block");
662 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free");
663 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add");
664 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove");
665 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move");
666 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block");
667 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block");
668 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag");
669 SOFTDEP_TYPE(JSEG, jseg, "Journal segment");
670 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete");
671 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency");
672 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation");
673 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete");
675 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel");
677 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes");
678 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations");
679 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data");
681 #define M_SOFTDEP_FLAGS (M_WAITOK)
684 * translate from workitem type to memory type
685 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
687 static struct malloc_type *memtype[] = {
718 #define DtoM(type) (memtype[type])
721 * Names of malloc types.
723 #define TYPENAME(type) \
724 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???")
726 * End system adaptation definitions.
729 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino)
730 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino)
733 * Internal function prototypes.
735 static void check_clear_deps(struct mount *);
736 static void softdep_error(char *, int);
737 static int softdep_process_worklist(struct mount *, int);
738 static int softdep_waitidle(struct mount *, int);
739 static void drain_output(struct vnode *);
740 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int);
741 static int check_inodedep_free(struct inodedep *);
742 static void clear_remove(struct mount *);
743 static void clear_inodedeps(struct mount *);
744 static void unlinked_inodedep(struct mount *, struct inodedep *);
745 static void clear_unlinked_inodedep(struct inodedep *);
746 static struct inodedep *first_unlinked_inodedep(struct ufsmount *);
747 static int flush_pagedep_deps(struct vnode *, struct mount *,
749 static int free_pagedep(struct pagedep *);
750 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t);
751 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t);
752 static int flush_deplist(struct allocdirectlst *, int, int *);
753 static int sync_cgs(struct mount *, int);
754 static int handle_written_filepage(struct pagedep *, struct buf *, int);
755 static int handle_written_sbdep(struct sbdep *, struct buf *);
756 static void initiate_write_sbdep(struct sbdep *);
757 static void diradd_inode_written(struct diradd *, struct inodedep *);
758 static int handle_written_indirdep(struct indirdep *, struct buf *,
760 static int handle_written_inodeblock(struct inodedep *, struct buf *, int);
761 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *,
763 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int);
764 static void handle_written_jaddref(struct jaddref *);
765 static void handle_written_jremref(struct jremref *);
766 static void handle_written_jseg(struct jseg *, struct buf *);
767 static void handle_written_jnewblk(struct jnewblk *);
768 static void handle_written_jblkdep(struct jblkdep *);
769 static void handle_written_jfreefrag(struct jfreefrag *);
770 static void complete_jseg(struct jseg *);
771 static void complete_jsegs(struct jseg *);
772 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *);
773 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *);
774 static void jremref_write(struct jremref *, struct jseg *, uint8_t *);
775 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *);
776 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *);
777 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data);
778 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *);
779 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *);
780 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *);
781 static inline void inoref_write(struct inoref *, struct jseg *,
783 static void handle_allocdirect_partdone(struct allocdirect *,
785 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *,
787 static void indirdep_complete(struct indirdep *);
788 static int indirblk_lookup(struct mount *, ufs2_daddr_t);
789 static void indirblk_insert(struct freework *);
790 static void indirblk_remove(struct freework *);
791 static void handle_allocindir_partdone(struct allocindir *);
792 static void initiate_write_filepage(struct pagedep *, struct buf *);
793 static void initiate_write_indirdep(struct indirdep*, struct buf *);
794 static void handle_written_mkdir(struct mkdir *, int);
795 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *,
797 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *);
798 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
799 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
800 static void handle_workitem_freefile(struct freefile *);
801 static int handle_workitem_remove(struct dirrem *, int);
802 static struct dirrem *newdirrem(struct buf *, struct inode *,
803 struct inode *, int, struct dirrem **);
804 static struct indirdep *indirdep_lookup(struct mount *, struct inode *,
806 static void cancel_indirdep(struct indirdep *, struct buf *,
808 static void free_indirdep(struct indirdep *);
809 static void free_diradd(struct diradd *, struct workhead *);
810 static void merge_diradd(struct inodedep *, struct diradd *);
811 static void complete_diradd(struct diradd *);
812 static struct diradd *diradd_lookup(struct pagedep *, int);
813 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *,
815 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *,
817 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *,
818 struct jremref *, struct jremref *);
819 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *,
821 static void cancel_allocindir(struct allocindir *, struct buf *bp,
822 struct freeblks *, int);
823 static int setup_trunc_indir(struct freeblks *, struct inode *,
824 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t);
825 static void complete_trunc_indir(struct freework *);
826 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *,
828 static void complete_mkdir(struct mkdir *);
829 static void free_newdirblk(struct newdirblk *);
830 static void free_jremref(struct jremref *);
831 static void free_jaddref(struct jaddref *);
832 static void free_jsegdep(struct jsegdep *);
833 static void free_jsegs(struct jblocks *);
834 static void rele_jseg(struct jseg *);
835 static void free_jseg(struct jseg *, struct jblocks *);
836 static void free_jnewblk(struct jnewblk *);
837 static void free_jblkdep(struct jblkdep *);
838 static void free_jfreefrag(struct jfreefrag *);
839 static void free_freedep(struct freedep *);
840 static void journal_jremref(struct dirrem *, struct jremref *,
842 static void cancel_jnewblk(struct jnewblk *, struct workhead *);
843 static int cancel_jaddref(struct jaddref *, struct inodedep *,
845 static void cancel_jfreefrag(struct jfreefrag *);
846 static inline void setup_freedirect(struct freeblks *, struct inode *,
848 static inline void setup_freeext(struct freeblks *, struct inode *, int, int);
849 static inline void setup_freeindir(struct freeblks *, struct inode *, int,
851 static inline struct freeblks *newfreeblks(struct mount *, struct inode *);
852 static void freeblks_free(struct ufsmount *, struct freeblks *, int);
853 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t);
854 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t);
855 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int);
856 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t,
858 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int);
859 static int cancel_pagedep(struct pagedep *, struct freeblks *, int);
860 static int deallocate_dependencies(struct buf *, struct freeblks *, int);
861 static void newblk_freefrag(struct newblk*);
862 static void free_newblk(struct newblk *);
863 static void cancel_allocdirect(struct allocdirectlst *,
864 struct allocdirect *, struct freeblks *);
865 static int check_inode_unwritten(struct inodedep *);
866 static int free_inodedep(struct inodedep *);
867 static void freework_freeblock(struct freework *);
868 static void freework_enqueue(struct freework *);
869 static int handle_workitem_freeblocks(struct freeblks *, int);
870 static int handle_complete_freeblocks(struct freeblks *, int);
871 static void handle_workitem_indirblk(struct freework *);
872 static void handle_written_freework(struct freework *);
873 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
874 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *,
876 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *,
877 struct inodedep *, struct allocindir *, ufs_lbn_t);
878 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
879 ufs2_daddr_t, ufs_lbn_t);
880 static void handle_workitem_freefrag(struct freefrag *);
881 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long,
883 static void allocdirect_merge(struct allocdirectlst *,
884 struct allocdirect *, struct allocdirect *);
885 static struct freefrag *allocindir_merge(struct allocindir *,
886 struct allocindir *);
887 static int bmsafemap_find(struct bmsafemap_hashhead *, int,
888 struct bmsafemap **);
889 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *,
890 int cg, struct bmsafemap *);
891 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int,
893 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **);
894 static int inodedep_find(struct inodedep_hashhead *, ino_t,
896 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
897 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t,
898 int, struct pagedep **);
899 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
901 static void pause_timer(void *);
902 static int request_cleanup(struct mount *, int);
903 static void schedule_cleanup(struct mount *);
904 static void softdep_ast_cleanup_proc(void);
905 static int process_worklist_item(struct mount *, int, int);
906 static void process_removes(struct vnode *);
907 static void process_truncates(struct vnode *);
908 static void jwork_move(struct workhead *, struct workhead *);
909 static void jwork_insert(struct workhead *, struct jsegdep *);
910 static void add_to_worklist(struct worklist *, int);
911 static void wake_worklist(struct worklist *);
912 static void wait_worklist(struct worklist *, char *);
913 static void remove_from_worklist(struct worklist *);
914 static void softdep_flush(void *);
915 static void softdep_flushjournal(struct mount *);
916 static int softdep_speedup(struct ufsmount *);
917 static void worklist_speedup(struct mount *);
918 static int journal_mount(struct mount *, struct fs *, struct ucred *);
919 static void journal_unmount(struct ufsmount *);
920 static int journal_space(struct ufsmount *, int);
921 static void journal_suspend(struct ufsmount *);
922 static int journal_unsuspend(struct ufsmount *ump);
923 static void softdep_prelink(struct vnode *, struct vnode *);
924 static void add_to_journal(struct worklist *);
925 static void remove_from_journal(struct worklist *);
926 static bool softdep_excess_items(struct ufsmount *, int);
927 static void softdep_process_journal(struct mount *, struct worklist *, int);
928 static struct jremref *newjremref(struct dirrem *, struct inode *,
929 struct inode *ip, off_t, nlink_t);
930 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t,
932 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t,
934 static inline struct jsegdep *inoref_jseg(struct inoref *);
935 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t);
936 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t,
938 static void adjust_newfreework(struct freeblks *, int);
939 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int);
940 static void move_newblock_dep(struct jaddref *, struct inodedep *);
941 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t);
942 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *,
943 ufs2_daddr_t, long, ufs_lbn_t);
944 static struct freework *newfreework(struct ufsmount *, struct freeblks *,
945 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int);
946 static int jwait(struct worklist *, int);
947 static struct inodedep *inodedep_lookup_ip(struct inode *);
948 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *);
949 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *);
950 static void handle_jwork(struct workhead *);
951 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *,
953 static struct jblocks *jblocks_create(void);
954 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *);
955 static void jblocks_free(struct jblocks *, struct mount *, int);
956 static void jblocks_destroy(struct jblocks *);
957 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int);
960 * Exported softdep operations.
962 static void softdep_disk_io_initiation(struct buf *);
963 static void softdep_disk_write_complete(struct buf *);
964 static void softdep_deallocate_dependencies(struct buf *);
965 static int softdep_count_dependencies(struct buf *bp, int);
968 * Global lock over all of soft updates.
970 static struct mtx lk;
971 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF);
973 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk)
974 #define FREE_GBLLOCK(lk) mtx_unlock(lk)
975 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED)
978 * Per-filesystem soft-updates locking.
980 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock)
981 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock)
982 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock)
983 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock)
984 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \
987 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock)
988 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock)
991 * Worklist queue management.
992 * These routines require that the lock be held.
994 #ifndef /* NOT */ DEBUG
995 #define WORKLIST_INSERT(head, item) do { \
996 (item)->wk_state |= ONWORKLIST; \
997 LIST_INSERT_HEAD(head, item, wk_list); \
999 #define WORKLIST_REMOVE(item) do { \
1000 (item)->wk_state &= ~ONWORKLIST; \
1001 LIST_REMOVE(item, wk_list); \
1003 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT
1004 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE
1007 static void worklist_insert(struct workhead *, struct worklist *, int);
1008 static void worklist_remove(struct worklist *, int);
1010 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1)
1011 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0)
1012 #define WORKLIST_REMOVE(item) worklist_remove(item, 1)
1013 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0)
1016 worklist_insert(head, item, locked)
1017 struct workhead *head;
1018 struct worklist *item;
1023 LOCK_OWNED(VFSTOUFS(item->wk_mp));
1024 if (item->wk_state & ONWORKLIST)
1025 panic("worklist_insert: %p %s(0x%X) already on list",
1026 item, TYPENAME(item->wk_type), item->wk_state);
1027 item->wk_state |= ONWORKLIST;
1028 LIST_INSERT_HEAD(head, item, wk_list);
1032 worklist_remove(item, locked)
1033 struct worklist *item;
1038 LOCK_OWNED(VFSTOUFS(item->wk_mp));
1039 if ((item->wk_state & ONWORKLIST) == 0)
1040 panic("worklist_remove: %p %s(0x%X) not on list",
1041 item, TYPENAME(item->wk_type), item->wk_state);
1042 item->wk_state &= ~ONWORKLIST;
1043 LIST_REMOVE(item, wk_list);
1048 * Merge two jsegdeps keeping only the oldest one as newer references
1049 * can't be discarded until after older references.
1051 static inline struct jsegdep *
1052 jsegdep_merge(struct jsegdep *one, struct jsegdep *two)
1054 struct jsegdep *swp;
1059 if (one->jd_seg->js_seq > two->jd_seg->js_seq) {
1064 WORKLIST_REMOVE(&two->jd_list);
1071 * If two freedeps are compatible free one to reduce list size.
1073 static inline struct freedep *
1074 freedep_merge(struct freedep *one, struct freedep *two)
1079 if (one->fd_freework == two->fd_freework) {
1080 WORKLIST_REMOVE(&two->fd_list);
1087 * Move journal work from one list to another. Duplicate freedeps and
1088 * jsegdeps are coalesced to keep the lists as small as possible.
1091 jwork_move(dst, src)
1092 struct workhead *dst;
1093 struct workhead *src;
1095 struct freedep *freedep;
1096 struct jsegdep *jsegdep;
1097 struct worklist *wkn;
1098 struct worklist *wk;
1101 ("jwork_move: dst == src"));
1104 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) {
1105 if (wk->wk_type == D_JSEGDEP)
1106 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1107 if (wk->wk_type == D_FREEDEP)
1108 freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1111 while ((wk = LIST_FIRST(src)) != NULL) {
1112 WORKLIST_REMOVE(wk);
1113 WORKLIST_INSERT(dst, wk);
1114 if (wk->wk_type == D_JSEGDEP) {
1115 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1118 if (wk->wk_type == D_FREEDEP)
1119 freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1124 jwork_insert(dst, jsegdep)
1125 struct workhead *dst;
1126 struct jsegdep *jsegdep;
1128 struct jsegdep *jsegdepn;
1129 struct worklist *wk;
1131 LIST_FOREACH(wk, dst, wk_list)
1132 if (wk->wk_type == D_JSEGDEP)
1135 WORKLIST_INSERT(dst, &jsegdep->jd_list);
1138 jsegdepn = WK_JSEGDEP(wk);
1139 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) {
1140 WORKLIST_REMOVE(wk);
1141 free_jsegdep(jsegdepn);
1142 WORKLIST_INSERT(dst, &jsegdep->jd_list);
1144 free_jsegdep(jsegdep);
1148 * Routines for tracking and managing workitems.
1150 static void workitem_free(struct worklist *, int);
1151 static void workitem_alloc(struct worklist *, int, struct mount *);
1152 static void workitem_reassign(struct worklist *, int);
1154 #define WORKITEM_FREE(item, type) \
1155 workitem_free((struct worklist *)(item), (type))
1156 #define WORKITEM_REASSIGN(item, type) \
1157 workitem_reassign((struct worklist *)(item), (type))
1160 workitem_free(item, type)
1161 struct worklist *item;
1164 struct ufsmount *ump;
1167 if (item->wk_state & ONWORKLIST)
1168 panic("workitem_free: %s(0x%X) still on list",
1169 TYPENAME(item->wk_type), item->wk_state);
1170 if (item->wk_type != type && type != D_NEWBLK)
1171 panic("workitem_free: type mismatch %s != %s",
1172 TYPENAME(item->wk_type), TYPENAME(type));
1174 if (item->wk_state & IOWAITING)
1176 ump = VFSTOUFS(item->wk_mp);
1178 KASSERT(ump->softdep_deps > 0,
1179 ("workitem_free: %s: softdep_deps going negative",
1180 ump->um_fs->fs_fsmnt));
1181 if (--ump->softdep_deps == 0 && ump->softdep_req)
1182 wakeup(&ump->softdep_deps);
1183 KASSERT(dep_current[item->wk_type] > 0,
1184 ("workitem_free: %s: dep_current[%s] going negative",
1185 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1186 KASSERT(ump->softdep_curdeps[item->wk_type] > 0,
1187 ("workitem_free: %s: softdep_curdeps[%s] going negative",
1188 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1189 atomic_subtract_long(&dep_current[item->wk_type], 1);
1190 ump->softdep_curdeps[item->wk_type] -= 1;
1191 free(item, DtoM(type));
1195 workitem_alloc(item, type, mp)
1196 struct worklist *item;
1200 struct ufsmount *ump;
1202 item->wk_type = type;
1207 ACQUIRE_GBLLOCK(&lk);
1208 dep_current[type]++;
1209 if (dep_current[type] > dep_highuse[type])
1210 dep_highuse[type] = dep_current[type];
1214 ump->softdep_curdeps[type] += 1;
1215 ump->softdep_deps++;
1216 ump->softdep_accdeps++;
1221 workitem_reassign(item, newtype)
1222 struct worklist *item;
1225 struct ufsmount *ump;
1227 ump = VFSTOUFS(item->wk_mp);
1229 KASSERT(ump->softdep_curdeps[item->wk_type] > 0,
1230 ("workitem_reassign: %s: softdep_curdeps[%s] going negative",
1231 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1232 ump->softdep_curdeps[item->wk_type] -= 1;
1233 ump->softdep_curdeps[newtype] += 1;
1234 KASSERT(dep_current[item->wk_type] > 0,
1235 ("workitem_reassign: %s: dep_current[%s] going negative",
1236 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
1237 ACQUIRE_GBLLOCK(&lk);
1238 dep_current[newtype]++;
1239 dep_current[item->wk_type]--;
1240 if (dep_current[newtype] > dep_highuse[newtype])
1241 dep_highuse[newtype] = dep_current[newtype];
1242 dep_total[newtype]++;
1244 item->wk_type = newtype;
1248 * Workitem queue management
1250 static int max_softdeps; /* maximum number of structs before slowdown */
1251 static int tickdelay = 2; /* number of ticks to pause during slowdown */
1252 static int proc_waiting; /* tracks whether we have a timeout posted */
1253 static int *stat_countp; /* statistic to count in proc_waiting timeout */
1254 static struct callout softdep_callout;
1255 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
1256 static int req_clear_remove; /* syncer process flush some freeblks */
1257 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */
1260 * runtime statistics
1262 static int stat_flush_threads; /* number of softdep flushing threads */
1263 static int stat_worklist_push; /* number of worklist cleanups */
1264 static int stat_blk_limit_push; /* number of times block limit neared */
1265 static int stat_ino_limit_push; /* number of times inode limit neared */
1266 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
1267 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
1268 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
1269 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
1270 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
1271 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
1272 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
1273 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */
1274 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */
1275 static int stat_journal_min; /* Times hit journal min threshold */
1276 static int stat_journal_low; /* Times hit journal low threshold */
1277 static int stat_journal_wait; /* Times blocked in jwait(). */
1278 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */
1279 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */
1280 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */
1281 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */
1282 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */
1283 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */
1284 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */
1285 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */
1286 static int stat_cleanup_failures; /* Number of cleanup requests that failed */
1287 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */
1289 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW,
1290 &max_softdeps, 0, "");
1291 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW,
1293 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD,
1294 &stat_flush_threads, 0, "");
1295 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW,
1296 &stat_worklist_push, 0,"");
1297 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW,
1298 &stat_blk_limit_push, 0,"");
1299 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW,
1300 &stat_ino_limit_push, 0,"");
1301 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW,
1302 &stat_blk_limit_hit, 0, "");
1303 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW,
1304 &stat_ino_limit_hit, 0, "");
1305 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW,
1306 &stat_sync_limit_hit, 0, "");
1307 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW,
1308 &stat_indir_blk_ptrs, 0, "");
1309 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW,
1310 &stat_inode_bitmap, 0, "");
1311 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW,
1312 &stat_direct_blk_ptrs, 0, "");
1313 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW,
1314 &stat_dir_entry, 0, "");
1315 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW,
1316 &stat_jaddref, 0, "");
1317 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW,
1318 &stat_jnewblk, 0, "");
1319 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW,
1320 &stat_journal_low, 0, "");
1321 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW,
1322 &stat_journal_min, 0, "");
1323 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW,
1324 &stat_journal_wait, 0, "");
1325 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW,
1326 &stat_jwait_filepage, 0, "");
1327 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW,
1328 &stat_jwait_freeblks, 0, "");
1329 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW,
1330 &stat_jwait_inode, 0, "");
1331 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW,
1332 &stat_jwait_newblk, 0, "");
1333 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW,
1334 &stat_cleanup_blkrequests, 0, "");
1335 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW,
1336 &stat_cleanup_inorequests, 0, "");
1337 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW,
1338 &stat_cleanup_high_delay, 0, "");
1339 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW,
1340 &stat_cleanup_retries, 0, "");
1341 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW,
1342 &stat_cleanup_failures, 0, "");
1343 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW,
1344 &softdep_flushcache, 0, "");
1345 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD,
1346 &stat_emptyjblocks, 0, "");
1348 SYSCTL_DECL(_vfs_ffs);
1350 /* Whether to recompute the summary at mount time */
1351 static int compute_summary_at_mount = 0;
1352 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
1353 &compute_summary_at_mount, 0, "Recompute summary at mount");
1354 static int print_threads = 0;
1355 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW,
1356 &print_threads, 0, "Notify flusher thread start/stop");
1358 /* List of all filesystems mounted with soft updates */
1359 static TAILQ_HEAD(, mount_softdeps) softdepmounts;
1362 * This function cleans the worklist for a filesystem.
1363 * Each filesystem running with soft dependencies gets its own
1364 * thread to run in this function. The thread is started up in
1365 * softdep_mount and shutdown in softdep_unmount. They show up
1366 * as part of the kernel "bufdaemon" process whose process
1367 * entry is available in bufdaemonproc.
1369 static int searchfailed;
1370 extern struct proc *bufdaemonproc;
1377 struct ufsmount *ump;
1380 td->td_pflags |= TDP_NORUNNINGBUF;
1381 mp = (struct mount *)addr;
1383 atomic_add_int(&stat_flush_threads, 1);
1385 ump->softdep_flags &= ~FLUSH_STARTING;
1386 wakeup(&ump->softdep_flushtd);
1388 if (print_threads) {
1389 if (stat_flush_threads == 1)
1390 printf("Running %s at pid %d\n", bufdaemonproc->p_comm,
1391 bufdaemonproc->p_pid);
1392 printf("Start thread %s\n", td->td_name);
1395 while (softdep_process_worklist(mp, 0) > 0 ||
1397 VFSTOUFS(mp)->softdep_jblocks->jb_suspended))
1398 kthread_suspend_check();
1400 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1401 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM,
1403 ump->softdep_flags &= ~FLUSH_CLEANUP;
1405 * Check to see if we are done and need to exit.
1407 if ((ump->softdep_flags & FLUSH_EXIT) == 0) {
1411 ump->softdep_flags &= ~FLUSH_EXIT;
1413 wakeup(&ump->softdep_flags);
1415 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups);
1416 atomic_subtract_int(&stat_flush_threads, 1);
1418 panic("kthread_exit failed\n");
1423 worklist_speedup(mp)
1426 struct ufsmount *ump;
1430 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1431 ump->softdep_flags |= FLUSH_CLEANUP;
1432 wakeup(&ump->softdep_flushtd);
1436 softdep_speedup(ump)
1437 struct ufsmount *ump;
1439 struct ufsmount *altump;
1440 struct mount_softdeps *sdp;
1443 worklist_speedup(ump->um_mountp);
1446 * If we have global shortages, then we need other
1447 * filesystems to help with the cleanup. Here we wakeup a
1448 * flusher thread for a filesystem that is over its fair
1449 * share of resources.
1451 if (req_clear_inodedeps || req_clear_remove) {
1452 ACQUIRE_GBLLOCK(&lk);
1453 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) {
1454 if ((altump = sdp->sd_ump) == ump)
1456 if (((req_clear_inodedeps &&
1457 altump->softdep_curdeps[D_INODEDEP] >
1458 max_softdeps / stat_flush_threads) ||
1459 (req_clear_remove &&
1460 altump->softdep_curdeps[D_DIRREM] >
1461 (max_softdeps / 2) / stat_flush_threads)) &&
1462 TRY_ACQUIRE_LOCK(altump))
1470 * Move to the end of the list so we pick a
1471 * different one on out next try.
1473 TAILQ_REMOVE(&softdepmounts, sdp, sd_next);
1474 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next);
1476 if ((altump->softdep_flags &
1477 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0)
1478 altump->softdep_flags |= FLUSH_CLEANUP;
1479 altump->um_softdep->sd_cleanups++;
1480 wakeup(&altump->softdep_flushtd);
1484 return (speedup_syncer());
1488 * Add an item to the end of the work queue.
1489 * This routine requires that the lock be held.
1490 * This is the only routine that adds items to the list.
1491 * The following routine is the only one that removes items
1492 * and does so in order from first to last.
1495 #define WK_HEAD 0x0001 /* Add to HEAD. */
1496 #define WK_NODELAY 0x0002 /* Process immediately. */
1499 add_to_worklist(wk, flags)
1500 struct worklist *wk;
1503 struct ufsmount *ump;
1505 ump = VFSTOUFS(wk->wk_mp);
1507 if (wk->wk_state & ONWORKLIST)
1508 panic("add_to_worklist: %s(0x%X) already on list",
1509 TYPENAME(wk->wk_type), wk->wk_state);
1510 wk->wk_state |= ONWORKLIST;
1511 if (ump->softdep_on_worklist == 0) {
1512 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
1513 ump->softdep_worklist_tail = wk;
1514 } else if (flags & WK_HEAD) {
1515 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
1517 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
1518 ump->softdep_worklist_tail = wk;
1520 ump->softdep_on_worklist += 1;
1521 if (flags & WK_NODELAY)
1522 worklist_speedup(wk->wk_mp);
1526 * Remove the item to be processed. If we are removing the last
1527 * item on the list, we need to recalculate the tail pointer.
1530 remove_from_worklist(wk)
1531 struct worklist *wk;
1533 struct ufsmount *ump;
1535 ump = VFSTOUFS(wk->wk_mp);
1536 WORKLIST_REMOVE(wk);
1537 if (ump->softdep_worklist_tail == wk)
1538 ump->softdep_worklist_tail =
1539 (struct worklist *)wk->wk_list.le_prev;
1540 ump->softdep_on_worklist -= 1;
1545 struct worklist *wk;
1547 if (wk->wk_state & IOWAITING) {
1548 wk->wk_state &= ~IOWAITING;
1554 wait_worklist(wk, wmesg)
1555 struct worklist *wk;
1558 struct ufsmount *ump;
1560 ump = VFSTOUFS(wk->wk_mp);
1561 wk->wk_state |= IOWAITING;
1562 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0);
1566 * Process that runs once per second to handle items in the background queue.
1568 * Note that we ensure that everything is done in the order in which they
1569 * appear in the queue. The code below depends on this property to ensure
1570 * that blocks of a file are freed before the inode itself is freed. This
1571 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
1572 * until all the old ones have been purged from the dependency lists.
1575 softdep_process_worklist(mp, full)
1580 struct ufsmount *ump;
1583 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
1584 if (MOUNTEDSOFTDEP(mp) == 0)
1589 starttime = time_second;
1590 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0);
1591 check_clear_deps(mp);
1592 while (ump->softdep_on_worklist > 0) {
1593 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0)
1597 check_clear_deps(mp);
1599 * We do not generally want to stop for buffer space, but if
1600 * we are really being a buffer hog, we will stop and wait.
1602 if (should_yield()) {
1604 kern_yield(PRI_USER);
1609 * Never allow processing to run for more than one
1610 * second. This gives the syncer thread the opportunity
1611 * to pause if appropriate.
1613 if (!full && starttime != time_second)
1617 journal_unsuspend(ump);
1623 * Process all removes associated with a vnode if we are running out of
1624 * journal space. Any other process which attempts to flush these will
1625 * be unable as we have the vnodes locked.
1631 struct inodedep *inodedep;
1632 struct dirrem *dirrem;
1633 struct ufsmount *ump;
1640 inum = VTOI(vp)->i_number;
1643 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
1645 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) {
1647 * If another thread is trying to lock this vnode
1648 * it will fail but we must wait for it to do so
1649 * before we can proceed.
1651 if (dirrem->dm_state & INPROGRESS) {
1652 wait_worklist(&dirrem->dm_list, "pwrwait");
1655 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) ==
1656 (COMPLETE | ONWORKLIST))
1661 remove_from_worklist(&dirrem->dm_list);
1663 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1664 panic("process_removes: suspended filesystem");
1665 handle_workitem_remove(dirrem, 0);
1666 vn_finished_secondary_write(mp);
1672 * Process all truncations associated with a vnode if we are running out
1673 * of journal space. This is called when the vnode lock is already held
1674 * and no other process can clear the truncation. This function returns
1675 * a value greater than zero if it did any work.
1678 process_truncates(vp)
1681 struct inodedep *inodedep;
1682 struct freeblks *freeblks;
1683 struct ufsmount *ump;
1691 inum = VTOI(vp)->i_number;
1693 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
1696 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) {
1697 /* Journal entries not yet written. */
1698 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) {
1700 &freeblks->fb_jblkdephd)->jb_list,
1704 /* Another thread is executing this item. */
1705 if (freeblks->fb_state & INPROGRESS) {
1706 wait_worklist(&freeblks->fb_list, "ptrwait");
1709 /* Freeblks is waiting on a inode write. */
1710 if ((freeblks->fb_state & COMPLETE) == 0) {
1716 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) ==
1717 (ALLCOMPLETE | ONWORKLIST)) {
1718 remove_from_worklist(&freeblks->fb_list);
1719 freeblks->fb_state |= INPROGRESS;
1721 if (vn_start_secondary_write(NULL, &mp,
1723 panic("process_truncates: "
1724 "suspended filesystem");
1725 handle_workitem_freeblocks(freeblks, 0);
1726 vn_finished_secondary_write(mp);
1730 if (freeblks->fb_cgwait)
1735 sync_cgs(mp, MNT_WAIT);
1736 ffs_sync_snap(mp, MNT_WAIT);
1740 if (freeblks == NULL)
1747 * Process one item on the worklist.
1750 process_worklist_item(mp, target, flags)
1755 struct worklist sentinel;
1756 struct worklist *wk;
1757 struct ufsmount *ump;
1761 KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
1763 * If we are being called because of a process doing a
1764 * copy-on-write, then it is not safe to write as we may
1765 * recurse into the copy-on-write routine.
1767 if (curthread->td_pflags & TDP_COWINPROGRESS)
1769 PHOLD(curproc); /* Don't let the stack go away. */
1773 sentinel.wk_mp = NULL;
1774 sentinel.wk_type = D_SENTINEL;
1775 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list);
1776 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL;
1777 wk = LIST_NEXT(&sentinel, wk_list)) {
1778 if (wk->wk_type == D_SENTINEL) {
1779 LIST_REMOVE(&sentinel, wk_list);
1780 LIST_INSERT_AFTER(wk, &sentinel, wk_list);
1783 if (wk->wk_state & INPROGRESS)
1784 panic("process_worklist_item: %p already in progress.",
1786 wk->wk_state |= INPROGRESS;
1787 remove_from_worklist(wk);
1789 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1790 panic("process_worklist_item: suspended filesystem");
1791 switch (wk->wk_type) {
1793 /* removal of a directory entry */
1794 error = handle_workitem_remove(WK_DIRREM(wk), flags);
1798 /* releasing blocks and/or fragments from a file */
1799 error = handle_workitem_freeblocks(WK_FREEBLKS(wk),
1804 /* releasing a fragment when replaced as a file grows */
1805 handle_workitem_freefrag(WK_FREEFRAG(wk));
1810 /* releasing an inode when its link count drops to 0 */
1811 handle_workitem_freefile(WK_FREEFILE(wk));
1816 panic("%s_process_worklist: Unknown type %s",
1817 "softdep", TYPENAME(wk->wk_type));
1820 vn_finished_secondary_write(mp);
1823 if (++matchcnt == target)
1828 * We have to retry the worklist item later. Wake up any
1829 * waiters who may be able to complete it immediately and
1830 * add the item back to the head so we don't try to execute
1833 wk->wk_state &= ~INPROGRESS;
1835 add_to_worklist(wk, WK_HEAD);
1837 LIST_REMOVE(&sentinel, wk_list);
1838 /* Sentinal could've become the tail from remove_from_worklist. */
1839 if (ump->softdep_worklist_tail == &sentinel)
1840 ump->softdep_worklist_tail =
1841 (struct worklist *)sentinel.wk_list.le_prev;
1847 * Move dependencies from one buffer to another.
1850 softdep_move_dependencies(oldbp, newbp)
1854 struct worklist *wk, *wktail;
1855 struct ufsmount *ump;
1858 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL)
1860 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
1861 ("softdep_move_dependencies called on non-softdep filesystem"));
1864 ump = VFSTOUFS(wk->wk_mp);
1866 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
1867 LIST_REMOVE(wk, wk_list);
1868 if (wk->wk_type == D_BMSAFEMAP &&
1869 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp))
1872 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1874 LIST_INSERT_AFTER(wktail, wk, wk_list);
1883 * Purge the work list of all items associated with a particular mount point.
1886 softdep_flushworklist(oldmnt, countp, td)
1887 struct mount *oldmnt;
1891 struct vnode *devvp;
1892 struct ufsmount *ump;
1896 * Alternately flush the block device associated with the mount
1897 * point and process any dependencies that the flushing
1898 * creates. We continue until no more worklist dependencies
1903 ump = VFSTOUFS(oldmnt);
1904 devvp = ump->um_devvp;
1905 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1907 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1908 error = VOP_FSYNC(devvp, MNT_WAIT, td);
1909 VOP_UNLOCK(devvp, 0);
1916 #define SU_WAITIDLE_RETRIES 20
1918 softdep_waitidle(struct mount *mp, int flags __unused)
1920 struct ufsmount *ump;
1921 struct vnode *devvp;
1926 devvp = ump->um_devvp;
1930 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) {
1931 ump->softdep_req = 1;
1932 KASSERT((flags & FORCECLOSE) == 0 ||
1933 ump->softdep_on_worklist == 0,
1934 ("softdep_waitidle: work added after flush"));
1935 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP,
1936 "softdeps", 10 * hz);
1937 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1938 error = VOP_FSYNC(devvp, MNT_WAIT, td);
1939 VOP_UNLOCK(devvp, 0);
1944 ump->softdep_req = 0;
1945 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) {
1947 printf("softdep_waitidle: Failed to flush worklist for %p\n",
1955 * Flush all vnodes and worklist items associated with a specified mount point.
1958 softdep_flushfiles(oldmnt, flags, td)
1959 struct mount *oldmnt;
1964 struct ufsmount *ump;
1967 int error, early, depcount, loopcnt, retry_flush_count, retry;
1970 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0,
1971 ("softdep_flushfiles called on non-softdep filesystem"));
1973 retry_flush_count = 3;
1978 * Alternately flush the vnodes associated with the mount
1979 * point and process any dependencies that the flushing
1980 * creates. In theory, this loop can happen at most twice,
1981 * but we give it a few extra just to be sure.
1983 for (; loopcnt > 0; loopcnt--) {
1985 * Do another flush in case any vnodes were brought in
1986 * as part of the cleanup operations.
1988 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag &
1989 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH;
1990 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0)
1992 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
1997 * If we are unmounting then it is an error to fail. If we
1998 * are simply trying to downgrade to read-only, then filesystem
1999 * activity can keep us busy forever, so we just fail with EBUSY.
2002 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
2003 panic("softdep_flushfiles: looping");
2007 error = softdep_waitidle(oldmnt, flags);
2009 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
2012 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
2013 ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
2014 morework = oldmnt->mnt_nvnodelistsize > 0;
2016 ump = VFSTOUFS(oldmnt);
2018 for (i = 0; i < MAXQUOTAS; i++) {
2019 if (ump->um_quotas[i] != NULLVP)
2025 if (--retry_flush_count > 0) {
2031 MNT_IUNLOCK(oldmnt);
2040 * Structure hashing.
2042 * There are four types of structures that can be looked up:
2043 * 1) pagedep structures identified by mount point, inode number,
2044 * and logical block.
2045 * 2) inodedep structures identified by mount point and inode number.
2046 * 3) newblk structures identified by mount point and
2047 * physical block number.
2048 * 4) bmsafemap structures identified by mount point and
2049 * cylinder group number.
2051 * The "pagedep" and "inodedep" dependency structures are hashed
2052 * separately from the file blocks and inodes to which they correspond.
2053 * This separation helps when the in-memory copy of an inode or
2054 * file block must be replaced. It also obviates the need to access
2055 * an inode or file page when simply updating (or de-allocating)
2056 * dependency structures. Lookup of newblk structures is needed to
2057 * find newly allocated blocks when trying to associate them with
2058 * their allocdirect or allocindir structure.
2060 * The lookup routines optionally create and hash a new instance when
2061 * an existing entry is not found. The bmsafemap lookup routine always
2062 * allocates a new structure if an existing one is not found.
2064 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
2067 * Structures and routines associated with pagedep caching.
2069 #define PAGEDEP_HASH(ump, inum, lbn) \
2070 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size])
2073 pagedep_find(pagedephd, ino, lbn, pagedeppp)
2074 struct pagedep_hashhead *pagedephd;
2077 struct pagedep **pagedeppp;
2079 struct pagedep *pagedep;
2081 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
2082 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) {
2083 *pagedeppp = pagedep;
2091 * Look up a pagedep. Return 1 if found, 0 otherwise.
2092 * If not found, allocate if DEPALLOC flag is passed.
2093 * Found or allocated entry is returned in pagedeppp.
2094 * This routine must be called with splbio interrupts blocked.
2097 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp)
2103 struct pagedep **pagedeppp;
2105 struct pagedep *pagedep;
2106 struct pagedep_hashhead *pagedephd;
2107 struct worklist *wk;
2108 struct ufsmount *ump;
2115 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2116 if (wk->wk_type == D_PAGEDEP) {
2117 *pagedeppp = WK_PAGEDEP(wk);
2122 pagedephd = PAGEDEP_HASH(ump, ino, lbn);
2123 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp);
2125 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp)
2126 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list);
2129 if ((flags & DEPALLOC) == 0)
2132 pagedep = malloc(sizeof(struct pagedep),
2133 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
2134 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
2136 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp);
2139 * This should never happen since we only create pagedeps
2140 * with the vnode lock held. Could be an assert.
2142 WORKITEM_FREE(pagedep, D_PAGEDEP);
2145 pagedep->pd_ino = ino;
2146 pagedep->pd_lbn = lbn;
2147 LIST_INIT(&pagedep->pd_dirremhd);
2148 LIST_INIT(&pagedep->pd_pendinghd);
2149 for (i = 0; i < DAHASHSZ; i++)
2150 LIST_INIT(&pagedep->pd_diraddhd[i]);
2151 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
2152 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2153 *pagedeppp = pagedep;
2158 * Structures and routines associated with inodedep caching.
2160 #define INODEDEP_HASH(ump, inum) \
2161 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size])
2164 inodedep_find(inodedephd, inum, inodedeppp)
2165 struct inodedep_hashhead *inodedephd;
2167 struct inodedep **inodedeppp;
2169 struct inodedep *inodedep;
2171 LIST_FOREACH(inodedep, inodedephd, id_hash)
2172 if (inum == inodedep->id_ino)
2175 *inodedeppp = inodedep;
2183 * Look up an inodedep. Return 1 if found, 0 if not found.
2184 * If not found, allocate if DEPALLOC flag is passed.
2185 * Found or allocated entry is returned in inodedeppp.
2186 * This routine must be called with splbio interrupts blocked.
2189 inodedep_lookup(mp, inum, flags, inodedeppp)
2193 struct inodedep **inodedeppp;
2195 struct inodedep *inodedep;
2196 struct inodedep_hashhead *inodedephd;
2197 struct ufsmount *ump;
2203 inodedephd = INODEDEP_HASH(ump, inum);
2205 if (inodedep_find(inodedephd, inum, inodedeppp))
2207 if ((flags & DEPALLOC) == 0)
2210 * If the system is over its limit and our filesystem is
2211 * responsible for more than our share of that usage and
2212 * we are not in a rush, request some inodedep cleanup.
2214 if (softdep_excess_items(ump, D_INODEDEP))
2215 schedule_cleanup(mp);
2218 inodedep = malloc(sizeof(struct inodedep),
2219 M_INODEDEP, M_SOFTDEP_FLAGS);
2220 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
2222 if (inodedep_find(inodedephd, inum, inodedeppp)) {
2223 WORKITEM_FREE(inodedep, D_INODEDEP);
2226 inodedep->id_fs = fs;
2227 inodedep->id_ino = inum;
2228 inodedep->id_state = ALLCOMPLETE;
2229 inodedep->id_nlinkdelta = 0;
2230 inodedep->id_savedino1 = NULL;
2231 inodedep->id_savedsize = -1;
2232 inodedep->id_savedextsize = -1;
2233 inodedep->id_savednlink = -1;
2234 inodedep->id_bmsafemap = NULL;
2235 inodedep->id_mkdiradd = NULL;
2236 LIST_INIT(&inodedep->id_dirremhd);
2237 LIST_INIT(&inodedep->id_pendinghd);
2238 LIST_INIT(&inodedep->id_inowait);
2239 LIST_INIT(&inodedep->id_bufwait);
2240 TAILQ_INIT(&inodedep->id_inoreflst);
2241 TAILQ_INIT(&inodedep->id_inoupdt);
2242 TAILQ_INIT(&inodedep->id_newinoupdt);
2243 TAILQ_INIT(&inodedep->id_extupdt);
2244 TAILQ_INIT(&inodedep->id_newextupdt);
2245 TAILQ_INIT(&inodedep->id_freeblklst);
2246 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
2247 *inodedeppp = inodedep;
2252 * Structures and routines associated with newblk caching.
2254 #define NEWBLK_HASH(ump, inum) \
2255 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size])
2258 newblk_find(newblkhd, newblkno, flags, newblkpp)
2259 struct newblk_hashhead *newblkhd;
2260 ufs2_daddr_t newblkno;
2262 struct newblk **newblkpp;
2264 struct newblk *newblk;
2266 LIST_FOREACH(newblk, newblkhd, nb_hash) {
2267 if (newblkno != newblk->nb_newblkno)
2270 * If we're creating a new dependency don't match those that
2271 * have already been converted to allocdirects. This is for
2274 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK)
2287 * Look up a newblk. Return 1 if found, 0 if not found.
2288 * If not found, allocate if DEPALLOC flag is passed.
2289 * Found or allocated entry is returned in newblkpp.
2292 newblk_lookup(mp, newblkno, flags, newblkpp)
2294 ufs2_daddr_t newblkno;
2296 struct newblk **newblkpp;
2298 struct newblk *newblk;
2299 struct newblk_hashhead *newblkhd;
2300 struct ufsmount *ump;
2304 newblkhd = NEWBLK_HASH(ump, newblkno);
2305 if (newblk_find(newblkhd, newblkno, flags, newblkpp))
2307 if ((flags & DEPALLOC) == 0)
2309 if (softdep_excess_items(ump, D_NEWBLK) ||
2310 softdep_excess_items(ump, D_ALLOCDIRECT) ||
2311 softdep_excess_items(ump, D_ALLOCINDIR))
2312 schedule_cleanup(mp);
2315 newblk = malloc(sizeof(union allblk), M_NEWBLK,
2316 M_SOFTDEP_FLAGS | M_ZERO);
2317 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp);
2319 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) {
2320 WORKITEM_FREE(newblk, D_NEWBLK);
2323 newblk->nb_freefrag = NULL;
2324 LIST_INIT(&newblk->nb_indirdeps);
2325 LIST_INIT(&newblk->nb_newdirblk);
2326 LIST_INIT(&newblk->nb_jwork);
2327 newblk->nb_state = ATTACHED;
2328 newblk->nb_newblkno = newblkno;
2329 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
2335 * Structures and routines associated with freed indirect block caching.
2337 #define INDIR_HASH(ump, blkno) \
2338 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size])
2341 * Lookup an indirect block in the indir hash table. The freework is
2342 * removed and potentially freed. The caller must do a blocking journal
2343 * write before writing to the blkno.
2346 indirblk_lookup(mp, blkno)
2350 struct freework *freework;
2351 struct indir_hashhead *wkhd;
2352 struct ufsmount *ump;
2355 wkhd = INDIR_HASH(ump, blkno);
2356 TAILQ_FOREACH(freework, wkhd, fw_next) {
2357 if (freework->fw_blkno != blkno)
2359 indirblk_remove(freework);
2366 * Insert an indirect block represented by freework into the indirblk
2367 * hash table so that it may prevent the block from being re-used prior
2368 * to the journal being written.
2371 indirblk_insert(freework)
2372 struct freework *freework;
2374 struct jblocks *jblocks;
2376 struct ufsmount *ump;
2378 ump = VFSTOUFS(freework->fw_list.wk_mp);
2379 jblocks = ump->softdep_jblocks;
2380 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst);
2384 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs);
2385 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework,
2387 freework->fw_state &= ~DEPCOMPLETE;
2391 indirblk_remove(freework)
2392 struct freework *freework;
2394 struct ufsmount *ump;
2396 ump = VFSTOUFS(freework->fw_list.wk_mp);
2397 LIST_REMOVE(freework, fw_segs);
2398 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next);
2399 freework->fw_state |= DEPCOMPLETE;
2400 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
2401 WORKITEM_FREE(freework, D_FREEWORK);
2405 * Executed during filesystem system initialization before
2406 * mounting any filesystems.
2409 softdep_initialize()
2412 TAILQ_INIT(&softdepmounts);
2414 max_softdeps = desiredvnodes * 4;
2416 max_softdeps = desiredvnodes * 2;
2419 /* initialise bioops hack */
2420 bioops.io_start = softdep_disk_io_initiation;
2421 bioops.io_complete = softdep_disk_write_complete;
2422 bioops.io_deallocate = softdep_deallocate_dependencies;
2423 bioops.io_countdeps = softdep_count_dependencies;
2424 softdep_ast_cleanup = softdep_ast_cleanup_proc;
2426 /* Initialize the callout with an mtx. */
2427 callout_init_mtx(&softdep_callout, &lk, 0);
2431 * Executed after all filesystems have been unmounted during
2432 * filesystem module unload.
2435 softdep_uninitialize()
2438 /* clear bioops hack */
2439 bioops.io_start = NULL;
2440 bioops.io_complete = NULL;
2441 bioops.io_deallocate = NULL;
2442 bioops.io_countdeps = NULL;
2443 softdep_ast_cleanup = NULL;
2445 callout_drain(&softdep_callout);
2449 * Called at mount time to notify the dependency code that a
2450 * filesystem wishes to use it.
2453 softdep_mount(devvp, mp, fs, cred)
2454 struct vnode *devvp;
2459 struct csum_total cstotal;
2460 struct mount_softdeps *sdp;
2461 struct ufsmount *ump;
2466 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA,
2469 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
2470 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
2471 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
2472 MNTK_SOFTDEP | MNTK_NOASYNC;
2475 ump->um_softdep = sdp;
2477 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock");
2479 LIST_INIT(&ump->softdep_workitem_pending);
2480 LIST_INIT(&ump->softdep_journal_pending);
2481 TAILQ_INIT(&ump->softdep_unlinked);
2482 LIST_INIT(&ump->softdep_dirtycg);
2483 ump->softdep_worklist_tail = NULL;
2484 ump->softdep_on_worklist = 0;
2485 ump->softdep_deps = 0;
2486 LIST_INIT(&ump->softdep_mkdirlisthd);
2487 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
2488 &ump->pagedep_hash_size);
2489 ump->pagedep_nextclean = 0;
2490 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP,
2491 &ump->inodedep_hash_size);
2492 ump->inodedep_nextclean = 0;
2493 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK,
2494 &ump->newblk_hash_size);
2495 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP,
2496 &ump->bmsafemap_hash_size);
2497 i = 1 << (ffs(desiredvnodes / 10) - 1);
2498 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead),
2499 M_FREEWORK, M_WAITOK);
2500 ump->indir_hash_size = i - 1;
2501 for (i = 0; i <= ump->indir_hash_size; i++)
2502 TAILQ_INIT(&ump->indir_hashtbl[i]);
2503 ACQUIRE_GBLLOCK(&lk);
2504 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next);
2506 if ((fs->fs_flags & FS_SUJ) &&
2507 (error = journal_mount(mp, fs, cred)) != 0) {
2508 printf("Failed to start journal: %d\n", error);
2509 softdep_unmount(mp);
2513 * Start our flushing thread in the bufdaemon process.
2516 ump->softdep_flags |= FLUSH_STARTING;
2518 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc,
2519 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker",
2520 mp->mnt_stat.f_mntonname);
2522 while ((ump->softdep_flags & FLUSH_STARTING) != 0) {
2523 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart",
2528 * When doing soft updates, the counters in the
2529 * superblock may have gotten out of sync. Recomputation
2530 * can take a long time and can be deferred for background
2531 * fsck. However, the old behavior of scanning the cylinder
2532 * groups and recalculating them at mount time is available
2533 * by setting vfs.ffs.compute_summary_at_mount to one.
2535 if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
2537 bzero(&cstotal, sizeof cstotal);
2538 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
2539 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
2540 fs->fs_cgsize, cred, &bp)) != 0) {
2542 softdep_unmount(mp);
2545 cgp = (struct cg *)bp->b_data;
2546 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
2547 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
2548 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
2549 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
2550 fs->fs_cs(fs, cyl) = cgp->cg_cs;
2554 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
2555 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
2557 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
2565 struct ufsmount *ump;
2570 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
2571 ("softdep_unmount called on non-softdep filesystem"));
2574 mp->mnt_flag &= ~MNT_SOFTDEP;
2575 if (MOUNTEDSUJ(mp) == 0) {
2578 mp->mnt_flag &= ~MNT_SUJ;
2580 journal_unmount(ump);
2583 * Shut down our flushing thread. Check for NULL is if
2584 * softdep_mount errors out before the thread has been created.
2586 if (ump->softdep_flushtd != NULL) {
2588 ump->softdep_flags |= FLUSH_EXIT;
2589 wakeup(&ump->softdep_flushtd);
2590 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP,
2592 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0,
2593 ("Thread shutdown failed"));
2596 * Free up our resources.
2598 ACQUIRE_GBLLOCK(&lk);
2599 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next);
2601 rw_destroy(LOCK_PTR(ump));
2602 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size);
2603 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size);
2604 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size);
2605 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP,
2606 ump->bmsafemap_hash_size);
2607 free(ump->indir_hashtbl, M_FREEWORK);
2609 for (i = 0; i <= D_LAST; i++)
2610 KASSERT(ump->softdep_curdeps[i] == 0,
2611 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt,
2612 TYPENAME(i), ump->softdep_curdeps[i]));
2614 free(ump->um_softdep, M_MOUNTDATA);
2617 static struct jblocks *
2618 jblocks_create(void)
2620 struct jblocks *jblocks;
2622 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO);
2623 TAILQ_INIT(&jblocks->jb_segs);
2624 jblocks->jb_avail = 10;
2625 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2626 M_JBLOCKS, M_WAITOK | M_ZERO);
2632 jblocks_alloc(jblocks, bytes, actual)
2633 struct jblocks *jblocks;
2638 struct jextent *jext;
2642 blocks = bytes / DEV_BSIZE;
2643 jext = &jblocks->jb_extent[jblocks->jb_head];
2644 freecnt = jext->je_blocks - jblocks->jb_off;
2646 jblocks->jb_off = 0;
2647 if (++jblocks->jb_head > jblocks->jb_used)
2648 jblocks->jb_head = 0;
2649 jext = &jblocks->jb_extent[jblocks->jb_head];
2650 freecnt = jext->je_blocks;
2652 if (freecnt > blocks)
2654 *actual = freecnt * DEV_BSIZE;
2655 daddr = jext->je_daddr + jblocks->jb_off;
2656 jblocks->jb_off += freecnt;
2657 jblocks->jb_free -= freecnt;
2663 jblocks_free(jblocks, mp, bytes)
2664 struct jblocks *jblocks;
2669 LOCK_OWNED(VFSTOUFS(mp));
2670 jblocks->jb_free += bytes / DEV_BSIZE;
2671 if (jblocks->jb_suspended)
2672 worklist_speedup(mp);
2677 jblocks_destroy(jblocks)
2678 struct jblocks *jblocks;
2681 if (jblocks->jb_extent)
2682 free(jblocks->jb_extent, M_JBLOCKS);
2683 free(jblocks, M_JBLOCKS);
2687 jblocks_add(jblocks, daddr, blocks)
2688 struct jblocks *jblocks;
2692 struct jextent *jext;
2694 jblocks->jb_blocks += blocks;
2695 jblocks->jb_free += blocks;
2696 jext = &jblocks->jb_extent[jblocks->jb_used];
2697 /* Adding the first block. */
2698 if (jext->je_daddr == 0) {
2699 jext->je_daddr = daddr;
2700 jext->je_blocks = blocks;
2703 /* Extending the last extent. */
2704 if (jext->je_daddr + jext->je_blocks == daddr) {
2705 jext->je_blocks += blocks;
2708 /* Adding a new extent. */
2709 if (++jblocks->jb_used == jblocks->jb_avail) {
2710 jblocks->jb_avail *= 2;
2711 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2712 M_JBLOCKS, M_WAITOK | M_ZERO);
2713 memcpy(jext, jblocks->jb_extent,
2714 sizeof(struct jextent) * jblocks->jb_used);
2715 free(jblocks->jb_extent, M_JBLOCKS);
2716 jblocks->jb_extent = jext;
2718 jext = &jblocks->jb_extent[jblocks->jb_used];
2719 jext->je_daddr = daddr;
2720 jext->je_blocks = blocks;
2725 softdep_journal_lookup(mp, vpp)
2729 struct componentname cnp;
2734 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp);
2737 bzero(&cnp, sizeof(cnp));
2738 cnp.cn_nameiop = LOOKUP;
2739 cnp.cn_flags = ISLASTCN;
2740 cnp.cn_thread = curthread;
2741 cnp.cn_cred = curthread->td_ucred;
2742 cnp.cn_pnbuf = SUJ_FILE;
2743 cnp.cn_nameptr = SUJ_FILE;
2744 cnp.cn_namelen = strlen(SUJ_FILE);
2745 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal);
2749 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp);
2754 * Open and verify the journal file.
2757 journal_mount(mp, fs, cred)
2762 struct jblocks *jblocks;
2763 struct ufsmount *ump;
2772 ump->softdep_journal_tail = NULL;
2773 ump->softdep_on_journal = 0;
2774 ump->softdep_accdeps = 0;
2775 ump->softdep_req = 0;
2776 ump->softdep_jblocks = NULL;
2777 error = softdep_journal_lookup(mp, &vp);
2779 printf("Failed to find journal. Use tunefs to create one\n");
2783 if (ip->i_size < SUJ_MIN) {
2787 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */
2788 jblocks = jblocks_create();
2789 for (i = 0; i < bcount; i++) {
2790 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL);
2793 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag));
2796 jblocks_destroy(jblocks);
2799 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */
2800 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */
2801 ump->softdep_jblocks = jblocks;
2805 mp->mnt_flag |= MNT_SUJ;
2806 mp->mnt_flag &= ~MNT_SOFTDEP;
2809 * Only validate the journal contents if the
2810 * filesystem is clean, otherwise we write the logs
2811 * but they'll never be used. If the filesystem was
2812 * still dirty when we mounted it the journal is
2813 * invalid and a new journal can only be valid if it
2814 * starts from a clean mount.
2817 DIP_SET(ip, i_modrev, fs->fs_mtime);
2818 ip->i_flags |= IN_MODIFIED;
2827 journal_unmount(ump)
2828 struct ufsmount *ump;
2831 if (ump->softdep_jblocks)
2832 jblocks_destroy(ump->softdep_jblocks);
2833 ump->softdep_jblocks = NULL;
2837 * Called when a journal record is ready to be written. Space is allocated
2838 * and the journal entry is created when the journal is flushed to stable
2843 struct worklist *wk;
2845 struct ufsmount *ump;
2847 ump = VFSTOUFS(wk->wk_mp);
2849 if (wk->wk_state & ONWORKLIST)
2850 panic("add_to_journal: %s(0x%X) already on list",
2851 TYPENAME(wk->wk_type), wk->wk_state);
2852 wk->wk_state |= ONWORKLIST | DEPCOMPLETE;
2853 if (LIST_EMPTY(&ump->softdep_journal_pending)) {
2854 ump->softdep_jblocks->jb_age = ticks;
2855 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list);
2857 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list);
2858 ump->softdep_journal_tail = wk;
2859 ump->softdep_on_journal += 1;
2863 * Remove an arbitrary item for the journal worklist maintain the tail
2864 * pointer. This happens when a new operation obviates the need to
2865 * journal an old operation.
2868 remove_from_journal(wk)
2869 struct worklist *wk;
2871 struct ufsmount *ump;
2873 ump = VFSTOUFS(wk->wk_mp);
2877 struct worklist *wkn;
2879 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list)
2883 panic("remove_from_journal: %p is not in journal", wk);
2887 * We emulate a TAILQ to save space in most structures which do not
2888 * require TAILQ semantics. Here we must update the tail position
2889 * when removing the tail which is not the final entry. This works
2890 * only if the worklist linkage are at the beginning of the structure.
2892 if (ump->softdep_journal_tail == wk)
2893 ump->softdep_journal_tail =
2894 (struct worklist *)wk->wk_list.le_prev;
2896 WORKLIST_REMOVE(wk);
2897 ump->softdep_on_journal -= 1;
2901 * Check for journal space as well as dependency limits so the prelink
2902 * code can throttle both journaled and non-journaled filesystems.
2903 * Threshold is 0 for low and 1 for min.
2906 journal_space(ump, thresh)
2907 struct ufsmount *ump;
2910 struct jblocks *jblocks;
2913 jblocks = ump->softdep_jblocks;
2914 if (jblocks == NULL)
2917 * We use a tighter restriction here to prevent request_cleanup()
2918 * running in threads from running into locks we currently hold.
2919 * We have to be over the limit and our filesystem has to be
2920 * responsible for more than our share of that usage.
2922 limit = (max_softdeps / 10) * 9;
2923 if (dep_current[D_INODEDEP] > limit &&
2924 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads)
2927 thresh = jblocks->jb_min;
2929 thresh = jblocks->jb_low;
2930 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE;
2931 avail = jblocks->jb_free - avail;
2933 return (avail > thresh);
2937 journal_suspend(ump)
2938 struct ufsmount *ump;
2940 struct jblocks *jblocks;
2944 jblocks = ump->softdep_jblocks;
2946 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
2948 mp->mnt_kern_flag |= MNTK_SUSPEND;
2949 mp->mnt_susp_owner = ump->softdep_flushtd;
2951 jblocks->jb_suspended = 1;
2956 journal_unsuspend(struct ufsmount *ump)
2958 struct jblocks *jblocks;
2962 jblocks = ump->softdep_jblocks;
2964 if (jblocks != NULL && jblocks->jb_suspended &&
2965 journal_space(ump, jblocks->jb_min)) {
2966 jblocks->jb_suspended = 0;
2968 mp->mnt_susp_owner = curthread;
2969 vfs_write_resume(mp, 0);
2977 * Called before any allocation function to be certain that there is
2978 * sufficient space in the journal prior to creating any new records.
2979 * Since in the case of block allocation we may have multiple locked
2980 * buffers at the time of the actual allocation we can not block
2981 * when the journal records are created. Doing so would create a deadlock
2982 * if any of these buffers needed to be flushed to reclaim space. Instead
2983 * we require a sufficiently large amount of available space such that
2984 * each thread in the system could have passed this allocation check and
2985 * still have sufficient free space. With 20% of a minimum journal size
2986 * of 1MB we have 6553 records available.
2989 softdep_prealloc(vp, waitok)
2993 struct ufsmount *ump;
2995 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
2996 ("softdep_prealloc called on non-softdep filesystem"));
2998 * Nothing to do if we are not running journaled soft updates.
2999 * If we currently hold the snapshot lock, we must avoid
3000 * handling other resources that could cause deadlock. Do not
3001 * touch quotas vnode since it is typically recursed with
3002 * other vnode locks held.
3004 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) ||
3005 (vp->v_vflag & VV_SYSTEM) != 0)
3007 ump = VFSTOUFS(vp->v_mount);
3009 if (journal_space(ump, 0)) {
3015 if (waitok == MNT_NOWAIT)
3018 * Attempt to sync this vnode once to flush any journal
3019 * work attached to it.
3021 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0)
3022 ffs_syncvnode(vp, waitok, 0);
3024 process_removes(vp);
3025 process_truncates(vp);
3026 if (journal_space(ump, 0) == 0) {
3027 softdep_speedup(ump);
3028 if (journal_space(ump, 1) == 0)
3029 journal_suspend(ump);
3037 * Before adjusting a link count on a vnode verify that we have sufficient
3038 * journal space. If not, process operations that depend on the currently
3039 * locked pair of vnodes to try to flush space as the syncer, buf daemon,
3040 * and softdep flush threads can not acquire these locks to reclaim space.
3043 softdep_prelink(dvp, vp)
3047 struct ufsmount *ump;
3049 ump = VFSTOUFS(dvp->v_mount);
3052 * Nothing to do if we have sufficient journal space.
3053 * If we currently hold the snapshot lock, we must avoid
3054 * handling other resources that could cause deadlock.
3056 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp))))
3061 ffs_syncvnode(vp, MNT_NOWAIT, 0);
3062 ffs_syncvnode(dvp, MNT_WAIT, 0);
3064 /* Process vp before dvp as it may create .. removes. */
3066 process_removes(vp);
3067 process_truncates(vp);
3069 process_removes(dvp);
3070 process_truncates(dvp);
3071 softdep_speedup(ump);
3072 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT);
3073 if (journal_space(ump, 0) == 0) {
3074 softdep_speedup(ump);
3075 if (journal_space(ump, 1) == 0)
3076 journal_suspend(ump);
3081 jseg_write(ump, jseg, data)
3082 struct ufsmount *ump;
3086 struct jsegrec *rec;
3088 rec = (struct jsegrec *)data;
3089 rec->jsr_seq = jseg->js_seq;
3090 rec->jsr_oldest = jseg->js_oldseq;
3091 rec->jsr_cnt = jseg->js_cnt;
3092 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize;
3094 rec->jsr_time = ump->um_fs->fs_mtime;
3098 inoref_write(inoref, jseg, rec)
3099 struct inoref *inoref;
3101 struct jrefrec *rec;
3104 inoref->if_jsegdep->jd_seg = jseg;
3105 rec->jr_ino = inoref->if_ino;
3106 rec->jr_parent = inoref->if_parent;
3107 rec->jr_nlink = inoref->if_nlink;
3108 rec->jr_mode = inoref->if_mode;
3109 rec->jr_diroff = inoref->if_diroff;
3113 jaddref_write(jaddref, jseg, data)
3114 struct jaddref *jaddref;
3118 struct jrefrec *rec;
3120 rec = (struct jrefrec *)data;
3121 rec->jr_op = JOP_ADDREF;
3122 inoref_write(&jaddref->ja_ref, jseg, rec);
3126 jremref_write(jremref, jseg, data)
3127 struct jremref *jremref;
3131 struct jrefrec *rec;
3133 rec = (struct jrefrec *)data;
3134 rec->jr_op = JOP_REMREF;
3135 inoref_write(&jremref->jr_ref, jseg, rec);
3139 jmvref_write(jmvref, jseg, data)
3140 struct jmvref *jmvref;
3146 rec = (struct jmvrec *)data;
3147 rec->jm_op = JOP_MVREF;
3148 rec->jm_ino = jmvref->jm_ino;
3149 rec->jm_parent = jmvref->jm_parent;
3150 rec->jm_oldoff = jmvref->jm_oldoff;
3151 rec->jm_newoff = jmvref->jm_newoff;
3155 jnewblk_write(jnewblk, jseg, data)
3156 struct jnewblk *jnewblk;
3160 struct jblkrec *rec;
3162 jnewblk->jn_jsegdep->jd_seg = jseg;
3163 rec = (struct jblkrec *)data;
3164 rec->jb_op = JOP_NEWBLK;
3165 rec->jb_ino = jnewblk->jn_ino;
3166 rec->jb_blkno = jnewblk->jn_blkno;
3167 rec->jb_lbn = jnewblk->jn_lbn;
3168 rec->jb_frags = jnewblk->jn_frags;
3169 rec->jb_oldfrags = jnewblk->jn_oldfrags;
3173 jfreeblk_write(jfreeblk, jseg, data)
3174 struct jfreeblk *jfreeblk;
3178 struct jblkrec *rec;
3180 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg;
3181 rec = (struct jblkrec *)data;
3182 rec->jb_op = JOP_FREEBLK;
3183 rec->jb_ino = jfreeblk->jf_ino;
3184 rec->jb_blkno = jfreeblk->jf_blkno;
3185 rec->jb_lbn = jfreeblk->jf_lbn;
3186 rec->jb_frags = jfreeblk->jf_frags;
3187 rec->jb_oldfrags = 0;
3191 jfreefrag_write(jfreefrag, jseg, data)
3192 struct jfreefrag *jfreefrag;
3196 struct jblkrec *rec;
3198 jfreefrag->fr_jsegdep->jd_seg = jseg;
3199 rec = (struct jblkrec *)data;
3200 rec->jb_op = JOP_FREEBLK;
3201 rec->jb_ino = jfreefrag->fr_ino;
3202 rec->jb_blkno = jfreefrag->fr_blkno;
3203 rec->jb_lbn = jfreefrag->fr_lbn;
3204 rec->jb_frags = jfreefrag->fr_frags;
3205 rec->jb_oldfrags = 0;
3209 jtrunc_write(jtrunc, jseg, data)
3210 struct jtrunc *jtrunc;
3214 struct jtrncrec *rec;
3216 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg;
3217 rec = (struct jtrncrec *)data;
3218 rec->jt_op = JOP_TRUNC;
3219 rec->jt_ino = jtrunc->jt_ino;
3220 rec->jt_size = jtrunc->jt_size;
3221 rec->jt_extsize = jtrunc->jt_extsize;
3225 jfsync_write(jfsync, jseg, data)
3226 struct jfsync *jfsync;
3230 struct jtrncrec *rec;
3232 rec = (struct jtrncrec *)data;
3233 rec->jt_op = JOP_SYNC;
3234 rec->jt_ino = jfsync->jfs_ino;
3235 rec->jt_size = jfsync->jfs_size;
3236 rec->jt_extsize = jfsync->jfs_extsize;
3240 softdep_flushjournal(mp)
3243 struct jblocks *jblocks;
3244 struct ufsmount *ump;
3246 if (MOUNTEDSUJ(mp) == 0)
3249 jblocks = ump->softdep_jblocks;
3251 while (ump->softdep_on_journal) {
3252 jblocks->jb_needseg = 1;
3253 softdep_process_journal(mp, NULL, MNT_WAIT);
3258 static void softdep_synchronize_completed(struct bio *);
3259 static void softdep_synchronize(struct bio *, struct ufsmount *, void *);
3262 softdep_synchronize_completed(bp)
3265 struct jseg *oldest;
3267 struct ufsmount *ump;
3270 * caller1 marks the last segment written before we issued the
3271 * synchronize cache.
3273 jseg = bp->bio_caller1;
3278 ump = VFSTOUFS(jseg->js_list.wk_mp);
3282 * Mark all the journal entries waiting on the synchronize cache
3283 * as completed so they may continue on.
3285 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) {
3286 jseg->js_state |= COMPLETE;
3288 jseg = TAILQ_PREV(jseg, jseglst, js_next);
3291 * Restart deferred journal entry processing from the oldest
3295 complete_jsegs(oldest);
3302 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering
3303 * barriers. The journal must be written prior to any blocks that depend
3304 * on it and the journal can not be released until the blocks have be
3305 * written. This code handles both barriers simultaneously.
3308 softdep_synchronize(bp, ump, caller1)
3310 struct ufsmount *ump;
3314 bp->bio_cmd = BIO_FLUSH;
3315 bp->bio_flags |= BIO_ORDERED;
3316 bp->bio_data = NULL;
3317 bp->bio_offset = ump->um_cp->provider->mediasize;
3319 bp->bio_done = softdep_synchronize_completed;
3320 bp->bio_caller1 = caller1;
3322 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private);
3326 * Flush some journal records to disk.
3329 softdep_process_journal(mp, needwk, flags)
3331 struct worklist *needwk;
3334 struct jblocks *jblocks;
3335 struct ufsmount *ump;
3336 struct worklist *wk;
3344 int jrecmin; /* Minimum records per block. */
3345 int jrecmax; /* Maximum records per block. */
3351 if (MOUNTEDSUJ(mp) == 0)
3353 shouldflush = softdep_flushcache;
3359 jblocks = ump->softdep_jblocks;
3360 devbsize = ump->um_devvp->v_bufobj.bo_bsize;
3362 * We write anywhere between a disk block and fs block. The upper
3363 * bound is picked to prevent buffer cache fragmentation and limit
3364 * processing time per I/O.
3366 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */
3367 jrecmax = (fs->fs_bsize / devbsize) * jrecmin;
3370 cnt = ump->softdep_on_journal;
3372 * Criteria for writing a segment:
3373 * 1) We have a full block.
3374 * 2) We're called from jwait() and haven't found the
3376 * 3) Always write if needseg is set.
3377 * 4) If we are called from process_worklist and have
3378 * not yet written anything we write a partial block
3379 * to enforce a 1 second maximum latency on journal
3382 if (cnt < (jrecmax - 1) && needwk == NULL &&
3383 jblocks->jb_needseg == 0 && (segwritten || cnt == 0))
3387 * Verify some free journal space. softdep_prealloc() should
3388 * guarantee that we don't run out so this is indicative of
3389 * a problem with the flow control. Try to recover
3390 * gracefully in any event.
3392 while (jblocks->jb_free == 0) {
3393 if (flags != MNT_WAIT)
3395 printf("softdep: Out of journal space!\n");
3396 softdep_speedup(ump);
3397 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz);
3400 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS);
3401 workitem_alloc(&jseg->js_list, D_JSEG, mp);
3402 LIST_INIT(&jseg->js_entries);
3403 LIST_INIT(&jseg->js_indirs);
3404 jseg->js_state = ATTACHED;
3405 if (shouldflush == 0)
3406 jseg->js_state |= COMPLETE;
3407 else if (bio == NULL)
3408 bio = g_alloc_bio();
3409 jseg->js_jblocks = jblocks;
3410 bp = geteblk(fs->fs_bsize, 0);
3413 * If there was a race while we were allocating the block
3414 * and jseg the entry we care about was likely written.
3415 * We bail out in both the WAIT and NOWAIT case and assume
3416 * the caller will loop if the entry it cares about is
3419 cnt = ump->softdep_on_journal;
3420 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) {
3421 bp->b_flags |= B_INVAL | B_NOCACHE;
3422 WORKITEM_FREE(jseg, D_JSEG);
3429 * Calculate the disk block size required for the available
3430 * records rounded to the min size.
3434 else if (cnt < jrecmax)
3435 size = howmany(cnt, jrecmin) * devbsize;
3437 size = fs->fs_bsize;
3439 * Allocate a disk block for this journal data and account
3440 * for truncation of the requested size if enough contiguous
3441 * space was not available.
3443 bp->b_blkno = jblocks_alloc(jblocks, size, &size);
3444 bp->b_lblkno = bp->b_blkno;
3445 bp->b_offset = bp->b_blkno * DEV_BSIZE;
3446 bp->b_bcount = size;
3447 bp->b_flags &= ~B_INVAL;
3448 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY;
3450 * Initialize our jseg with cnt records. Assign the next
3451 * sequence number to it and link it in-order.
3453 cnt = MIN(cnt, (size / devbsize) * jrecmin);
3456 jseg->js_refs = cnt + 1; /* Self ref. */
3457 jseg->js_size = size;
3458 jseg->js_seq = jblocks->jb_nextseq++;
3459 if (jblocks->jb_oldestseg == NULL)
3460 jblocks->jb_oldestseg = jseg;
3461 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq;
3462 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next);
3463 if (jblocks->jb_writeseg == NULL)
3464 jblocks->jb_writeseg = jseg;
3466 * Start filling in records from the pending list.
3472 * Always put a header on the first block.
3473 * XXX As with below, there might not be a chance to get
3474 * into the loop. Ensure that something valid is written.
3476 jseg_write(ump, jseg, data);
3478 data = bp->b_data + off;
3481 * XXX Something is wrong here. There's no work to do,
3482 * but we need to perform and I/O and allow it to complete
3485 if (LIST_EMPTY(&ump->softdep_journal_pending))
3486 stat_emptyjblocks++;
3488 while ((wk = LIST_FIRST(&ump->softdep_journal_pending))
3492 /* Place a segment header on every device block. */
3493 if ((off % devbsize) == 0) {
3494 jseg_write(ump, jseg, data);
3496 data = bp->b_data + off;
3500 remove_from_journal(wk);
3501 wk->wk_state |= INPROGRESS;
3502 WORKLIST_INSERT(&jseg->js_entries, wk);
3503 switch (wk->wk_type) {
3505 jaddref_write(WK_JADDREF(wk), jseg, data);
3508 jremref_write(WK_JREMREF(wk), jseg, data);
3511 jmvref_write(WK_JMVREF(wk), jseg, data);
3514 jnewblk_write(WK_JNEWBLK(wk), jseg, data);
3517 jfreeblk_write(WK_JFREEBLK(wk), jseg, data);
3520 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data);
3523 jtrunc_write(WK_JTRUNC(wk), jseg, data);
3526 jfsync_write(WK_JFSYNC(wk), jseg, data);
3529 panic("process_journal: Unknown type %s",
3530 TYPENAME(wk->wk_type));
3534 data = bp->b_data + off;
3538 /* Clear any remaining space so we don't leak kernel data */
3540 bzero(data, size - off);
3543 * Write this one buffer and continue.
3546 jblocks->jb_needseg = 0;
3547 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list);
3549 pbgetvp(ump->um_devvp, bp);
3551 * We only do the blocking wait once we find the journal
3552 * entry we're looking for.
3554 if (needwk == NULL && flags == MNT_WAIT)
3561 * If we wrote a segment issue a synchronize cache so the journal
3562 * is reflected on disk before the data is written. Since reclaiming
3563 * journal space also requires writing a journal record this
3564 * process also enforces a barrier before reclamation.
3566 if (segwritten && shouldflush) {
3567 softdep_synchronize(bio, ump,
3568 TAILQ_LAST(&jblocks->jb_segs, jseglst));
3572 * If we've suspended the filesystem because we ran out of journal
3573 * space either try to sync it here to make some progress or
3574 * unsuspend it if we already have.
3576 if (flags == 0 && jblocks->jb_suspended) {
3577 if (journal_unsuspend(ump))
3580 VFS_SYNC(mp, MNT_NOWAIT);
3581 ffs_sbupdate(ump, MNT_WAIT, 0);
3587 * Complete a jseg, allowing all dependencies awaiting journal writes
3588 * to proceed. Each journal dependency also attaches a jsegdep to dependent
3589 * structures so that the journal segment can be freed to reclaim space.
3595 struct worklist *wk;
3596 struct jmvref *jmvref;
3601 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
3602 WORKLIST_REMOVE(wk);
3603 wk->wk_state &= ~INPROGRESS;
3604 wk->wk_state |= COMPLETE;
3605 KASSERT(i++ < jseg->js_cnt,
3606 ("handle_written_jseg: overflow %d >= %d",
3607 i - 1, jseg->js_cnt));
3608 switch (wk->wk_type) {
3610 handle_written_jaddref(WK_JADDREF(wk));
3613 handle_written_jremref(WK_JREMREF(wk));
3616 rele_jseg(jseg); /* No jsegdep. */
3617 jmvref = WK_JMVREF(wk);
3618 LIST_REMOVE(jmvref, jm_deps);
3619 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0)
3620 free_pagedep(jmvref->jm_pagedep);
3621 WORKITEM_FREE(jmvref, D_JMVREF);
3624 handle_written_jnewblk(WK_JNEWBLK(wk));
3627 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep);
3630 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep);
3633 rele_jseg(jseg); /* No jsegdep. */
3634 WORKITEM_FREE(wk, D_JFSYNC);
3637 handle_written_jfreefrag(WK_JFREEFRAG(wk));
3640 panic("handle_written_jseg: Unknown type %s",
3641 TYPENAME(wk->wk_type));
3645 /* Release the self reference so the structure may be freed. */
3650 * Determine which jsegs are ready for completion processing. Waits for
3651 * synchronize cache to complete as well as forcing in-order completion
3652 * of journal entries.
3655 complete_jsegs(jseg)
3658 struct jblocks *jblocks;
3661 jblocks = jseg->js_jblocks;
3663 * Don't allow out of order completions. If this isn't the first
3664 * block wait for it to write before we're done.
3666 if (jseg != jblocks->jb_writeseg)
3668 /* Iterate through available jsegs processing their entries. */
3669 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) {
3670 jblocks->jb_oldestwrseq = jseg->js_oldseq;
3671 jsegn = TAILQ_NEXT(jseg, js_next);
3672 complete_jseg(jseg);
3675 jblocks->jb_writeseg = jseg;
3677 * Attempt to free jsegs now that oldestwrseq may have advanced.
3679 free_jsegs(jblocks);
3683 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle
3684 * the final completions.
3687 handle_written_jseg(jseg, bp)
3692 if (jseg->js_refs == 0)
3693 panic("handle_written_jseg: No self-reference on %p", jseg);
3694 jseg->js_state |= DEPCOMPLETE;
3696 * We'll never need this buffer again, set flags so it will be
3699 bp->b_flags |= B_INVAL | B_NOCACHE;
3701 complete_jsegs(jseg);
3704 static inline struct jsegdep *
3706 struct inoref *inoref;
3708 struct jsegdep *jsegdep;
3710 jsegdep = inoref->if_jsegdep;
3711 inoref->if_jsegdep = NULL;
3717 * Called once a jremref has made it to stable store. The jremref is marked
3718 * complete and we attempt to free it. Any pagedeps writes sleeping waiting
3719 * for the jremref to complete will be awoken by free_jremref.
3722 handle_written_jremref(jremref)
3723 struct jremref *jremref;
3725 struct inodedep *inodedep;
3726 struct jsegdep *jsegdep;
3727 struct dirrem *dirrem;
3729 /* Grab the jsegdep. */
3730 jsegdep = inoref_jseg(&jremref->jr_ref);
3732 * Remove us from the inoref list.
3734 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
3736 panic("handle_written_jremref: Lost inodedep");
3737 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
3739 * Complete the dirrem.
3741 dirrem = jremref->jr_dirrem;
3742 jremref->jr_dirrem = NULL;
3743 LIST_REMOVE(jremref, jr_deps);
3744 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
3745 jwork_insert(&dirrem->dm_jwork, jsegdep);
3746 if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
3747 (dirrem->dm_state & COMPLETE) != 0)
3748 add_to_worklist(&dirrem->dm_list, 0);
3749 free_jremref(jremref);
3753 * Called once a jaddref has made it to stable store. The dependency is
3754 * marked complete and any dependent structures are added to the inode
3755 * bufwait list to be completed as soon as it is written. If a bitmap write
3756 * depends on this entry we move the inode into the inodedephd of the
3757 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
3760 handle_written_jaddref(jaddref)
3761 struct jaddref *jaddref;
3763 struct jsegdep *jsegdep;
3764 struct inodedep *inodedep;
3765 struct diradd *diradd;
3766 struct mkdir *mkdir;
3768 /* Grab the jsegdep. */
3769 jsegdep = inoref_jseg(&jaddref->ja_ref);
3772 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3774 panic("handle_written_jaddref: Lost inodedep.");
3775 if (jaddref->ja_diradd == NULL)
3776 panic("handle_written_jaddref: No dependency");
3777 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
3778 diradd = jaddref->ja_diradd;
3779 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
3780 } else if (jaddref->ja_state & MKDIR_PARENT) {
3781 mkdir = jaddref->ja_mkdir;
3782 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
3783 } else if (jaddref->ja_state & MKDIR_BODY)
3784 mkdir = jaddref->ja_mkdir;
3786 panic("handle_written_jaddref: Unknown dependency %p",
3787 jaddref->ja_diradd);
3788 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */
3790 * Remove us from the inode list.
3792 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
3794 * The mkdir may be waiting on the jaddref to clear before freeing.
3797 KASSERT(mkdir->md_list.wk_type == D_MKDIR,
3798 ("handle_written_jaddref: Incorrect type for mkdir %s",
3799 TYPENAME(mkdir->md_list.wk_type)));
3800 mkdir->md_jaddref = NULL;
3801 diradd = mkdir->md_diradd;
3802 mkdir->md_state |= DEPCOMPLETE;
3803 complete_mkdir(mkdir);
3805 jwork_insert(&diradd->da_jwork, jsegdep);
3806 if (jaddref->ja_state & NEWBLOCK) {
3807 inodedep->id_state |= ONDEPLIST;
3808 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
3811 free_jaddref(jaddref);
3815 * Called once a jnewblk journal is written. The allocdirect or allocindir
3816 * is placed in the bmsafemap to await notification of a written bitmap. If
3817 * the operation was canceled we add the segdep to the appropriate
3818 * dependency to free the journal space once the canceling operation
3822 handle_written_jnewblk(jnewblk)
3823 struct jnewblk *jnewblk;
3825 struct bmsafemap *bmsafemap;
3826 struct freefrag *freefrag;
3827 struct freework *freework;
3828 struct jsegdep *jsegdep;
3829 struct newblk *newblk;
3831 /* Grab the jsegdep. */
3832 jsegdep = jnewblk->jn_jsegdep;
3833 jnewblk->jn_jsegdep = NULL;
3834 if (jnewblk->jn_dep == NULL)
3835 panic("handle_written_jnewblk: No dependency for the segdep.");
3836 switch (jnewblk->jn_dep->wk_type) {
3841 * Add the written block to the bmsafemap so it can
3842 * be notified when the bitmap is on disk.
3844 newblk = WK_NEWBLK(jnewblk->jn_dep);
3845 newblk->nb_jnewblk = NULL;
3846 if ((newblk->nb_state & GOINGAWAY) == 0) {
3847 bmsafemap = newblk->nb_bmsafemap;
3848 newblk->nb_state |= ONDEPLIST;
3849 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk,
3852 jwork_insert(&newblk->nb_jwork, jsegdep);
3856 * A newblock being removed by a freefrag when replaced by
3859 freefrag = WK_FREEFRAG(jnewblk->jn_dep);
3860 freefrag->ff_jdep = NULL;
3861 jwork_insert(&freefrag->ff_jwork, jsegdep);
3865 * A direct block was removed by truncate.
3867 freework = WK_FREEWORK(jnewblk->jn_dep);
3868 freework->fw_jnewblk = NULL;
3869 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep);
3872 panic("handle_written_jnewblk: Unknown type %d.",
3873 jnewblk->jn_dep->wk_type);
3875 jnewblk->jn_dep = NULL;
3876 free_jnewblk(jnewblk);
3880 * Cancel a jfreefrag that won't be needed, probably due to colliding with
3881 * an in-flight allocation that has not yet been committed. Divorce us
3882 * from the freefrag and mark it DEPCOMPLETE so that it may be added
3886 cancel_jfreefrag(jfreefrag)
3887 struct jfreefrag *jfreefrag;
3889 struct freefrag *freefrag;
3891 if (jfreefrag->fr_jsegdep) {
3892 free_jsegdep(jfreefrag->fr_jsegdep);
3893 jfreefrag->fr_jsegdep = NULL;
3895 freefrag = jfreefrag->fr_freefrag;
3896 jfreefrag->fr_freefrag = NULL;
3897 free_jfreefrag(jfreefrag);
3898 freefrag->ff_state |= DEPCOMPLETE;
3899 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno);
3903 * Free a jfreefrag when the parent freefrag is rendered obsolete.
3906 free_jfreefrag(jfreefrag)
3907 struct jfreefrag *jfreefrag;
3910 if (jfreefrag->fr_state & INPROGRESS)
3911 WORKLIST_REMOVE(&jfreefrag->fr_list);
3912 else if (jfreefrag->fr_state & ONWORKLIST)
3913 remove_from_journal(&jfreefrag->fr_list);
3914 if (jfreefrag->fr_freefrag != NULL)
3915 panic("free_jfreefrag: Still attached to a freefrag.");
3916 WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
3920 * Called when the journal write for a jfreefrag completes. The parent
3921 * freefrag is added to the worklist if this completes its dependencies.
3924 handle_written_jfreefrag(jfreefrag)
3925 struct jfreefrag *jfreefrag;
3927 struct jsegdep *jsegdep;
3928 struct freefrag *freefrag;
3930 /* Grab the jsegdep. */
3931 jsegdep = jfreefrag->fr_jsegdep;
3932 jfreefrag->fr_jsegdep = NULL;
3933 freefrag = jfreefrag->fr_freefrag;
3934 if (freefrag == NULL)
3935 panic("handle_written_jfreefrag: No freefrag.");
3936 freefrag->ff_state |= DEPCOMPLETE;
3937 freefrag->ff_jdep = NULL;
3938 jwork_insert(&freefrag->ff_jwork, jsegdep);
3939 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
3940 add_to_worklist(&freefrag->ff_list, 0);
3941 jfreefrag->fr_freefrag = NULL;
3942 free_jfreefrag(jfreefrag);
3946 * Called when the journal write for a jfreeblk completes. The jfreeblk
3947 * is removed from the freeblks list of pending journal writes and the
3948 * jsegdep is moved to the freeblks jwork to be completed when all blocks
3949 * have been reclaimed.
3952 handle_written_jblkdep(jblkdep)
3953 struct jblkdep *jblkdep;
3955 struct freeblks *freeblks;
3956 struct jsegdep *jsegdep;
3958 /* Grab the jsegdep. */
3959 jsegdep = jblkdep->jb_jsegdep;
3960 jblkdep->jb_jsegdep = NULL;
3961 freeblks = jblkdep->jb_freeblks;
3962 LIST_REMOVE(jblkdep, jb_deps);
3963 jwork_insert(&freeblks->fb_jwork, jsegdep);
3965 * If the freeblks is all journaled, we can add it to the worklist.
3967 if (LIST_EMPTY(&freeblks->fb_jblkdephd) &&
3968 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
3969 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
3971 free_jblkdep(jblkdep);
3974 static struct jsegdep *
3975 newjsegdep(struct worklist *wk)
3977 struct jsegdep *jsegdep;
3979 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
3980 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
3981 jsegdep->jd_seg = NULL;
3986 static struct jmvref *
3987 newjmvref(dp, ino, oldoff, newoff)
3993 struct jmvref *jmvref;
3995 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
3996 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump));
3997 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
3998 jmvref->jm_parent = dp->i_number;
3999 jmvref->jm_ino = ino;
4000 jmvref->jm_oldoff = oldoff;
4001 jmvref->jm_newoff = newoff;
4007 * Allocate a new jremref that tracks the removal of ip from dp with the
4008 * directory entry offset of diroff. Mark the entry as ATTACHED and
4009 * DEPCOMPLETE as we have all the information required for the journal write
4010 * and the directory has already been removed from the buffer. The caller
4011 * is responsible for linking the jremref into the pagedep and adding it
4012 * to the journal to write. The MKDIR_PARENT flag is set if we're doing
4013 * a DOTDOT addition so handle_workitem_remove() can properly assign
4014 * the jsegdep when we're done.
4016 static struct jremref *
4017 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
4018 off_t diroff, nlink_t nlink)
4020 struct jremref *jremref;
4022 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
4023 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump));
4024 jremref->jr_state = ATTACHED;
4025 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
4027 jremref->jr_dirrem = dirrem;
4033 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
4034 nlink_t nlink, uint16_t mode)
4037 inoref->if_jsegdep = newjsegdep(&inoref->if_list);
4038 inoref->if_diroff = diroff;
4039 inoref->if_ino = ino;
4040 inoref->if_parent = parent;
4041 inoref->if_nlink = nlink;
4042 inoref->if_mode = mode;
4046 * Allocate a new jaddref to track the addition of ino to dp at diroff. The
4047 * directory offset may not be known until later. The caller is responsible
4048 * adding the entry to the journal when this information is available. nlink
4049 * should be the link count prior to the addition and mode is only required
4050 * to have the correct FMT.
4052 static struct jaddref *
4053 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
4056 struct jaddref *jaddref;
4058 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
4059 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump));
4060 jaddref->ja_state = ATTACHED;
4061 jaddref->ja_mkdir = NULL;
4062 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
4068 * Create a new free dependency for a freework. The caller is responsible
4069 * for adjusting the reference count when it has the lock held. The freedep
4070 * will track an outstanding bitmap write that will ultimately clear the
4071 * freework to continue.
4073 static struct freedep *
4074 newfreedep(struct freework *freework)
4076 struct freedep *freedep;
4078 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
4079 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
4080 freedep->fd_freework = freework;
4086 * Free a freedep structure once the buffer it is linked to is written. If
4087 * this is the last reference to the freework schedule it for completion.
4090 free_freedep(freedep)
4091 struct freedep *freedep;
4093 struct freework *freework;
4095 freework = freedep->fd_freework;
4096 freework->fw_freeblks->fb_cgwait--;
4097 if (--freework->fw_ref == 0)
4098 freework_enqueue(freework);
4099 WORKITEM_FREE(freedep, D_FREEDEP);
4103 * Allocate a new freework structure that may be a level in an indirect
4104 * when parent is not NULL or a top level block when it is. The top level
4105 * freework structures are allocated without the per-filesystem lock held
4106 * and before the freeblks is visible outside of softdep_setup_freeblocks().
4108 static struct freework *
4109 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal)
4110 struct ufsmount *ump;
4111 struct freeblks *freeblks;
4112 struct freework *parent;
4119 struct freework *freework;
4121 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
4122 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
4123 freework->fw_state = ATTACHED;
4124 freework->fw_jnewblk = NULL;
4125 freework->fw_freeblks = freeblks;
4126 freework->fw_parent = parent;
4127 freework->fw_lbn = lbn;
4128 freework->fw_blkno = nb;
4129 freework->fw_frags = frags;
4130 freework->fw_indir = NULL;
4131 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR)
4132 ? 0 : NINDIR(ump->um_fs) + 1;
4133 freework->fw_start = freework->fw_off = off;
4135 newjfreeblk(freeblks, lbn, nb, frags);
4136 if (parent == NULL) {
4138 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
4147 * Eliminate a jfreeblk for a block that does not need journaling.
4150 cancel_jfreeblk(freeblks, blkno)
4151 struct freeblks *freeblks;
4154 struct jfreeblk *jfreeblk;
4155 struct jblkdep *jblkdep;
4157 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) {
4158 if (jblkdep->jb_list.wk_type != D_JFREEBLK)
4160 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list);
4161 if (jfreeblk->jf_blkno == blkno)
4164 if (jblkdep == NULL)
4166 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno);
4167 free_jsegdep(jblkdep->jb_jsegdep);
4168 LIST_REMOVE(jblkdep, jb_deps);
4169 WORKITEM_FREE(jfreeblk, D_JFREEBLK);
4173 * Allocate a new jfreeblk to journal top level block pointer when truncating
4174 * a file. The caller must add this to the worklist when the per-filesystem
4177 static struct jfreeblk *
4178 newjfreeblk(freeblks, lbn, blkno, frags)
4179 struct freeblks *freeblks;
4184 struct jfreeblk *jfreeblk;
4186 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
4187 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK,
4188 freeblks->fb_list.wk_mp);
4189 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list);
4190 jfreeblk->jf_dep.jb_freeblks = freeblks;
4191 jfreeblk->jf_ino = freeblks->fb_inum;
4192 jfreeblk->jf_lbn = lbn;
4193 jfreeblk->jf_blkno = blkno;
4194 jfreeblk->jf_frags = frags;
4195 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps);
4201 * The journal is only prepared to handle full-size block numbers, so we
4202 * have to adjust the record to reflect the change to a full-size block.
4203 * For example, suppose we have a block made up of fragments 8-15 and
4204 * want to free its last two fragments. We are given a request that says:
4205 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0
4206 * where frags are the number of fragments to free and oldfrags are the
4207 * number of fragments to keep. To block align it, we have to change it to
4208 * have a valid full-size blkno, so it becomes:
4209 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6
4212 adjust_newfreework(freeblks, frag_offset)
4213 struct freeblks *freeblks;
4216 struct jfreeblk *jfreeblk;
4218 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL &&
4219 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK),
4220 ("adjust_newfreework: Missing freeblks dependency"));
4222 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd));
4223 jfreeblk->jf_blkno -= frag_offset;
4224 jfreeblk->jf_frags += frag_offset;
4228 * Allocate a new jtrunc to track a partial truncation.
4230 static struct jtrunc *
4231 newjtrunc(freeblks, size, extsize)
4232 struct freeblks *freeblks;
4236 struct jtrunc *jtrunc;
4238 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
4239 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC,
4240 freeblks->fb_list.wk_mp);
4241 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list);
4242 jtrunc->jt_dep.jb_freeblks = freeblks;
4243 jtrunc->jt_ino = freeblks->fb_inum;
4244 jtrunc->jt_size = size;
4245 jtrunc->jt_extsize = extsize;
4246 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps);
4252 * If we're canceling a new bitmap we have to search for another ref
4253 * to move into the bmsafemap dep. This might be better expressed
4254 * with another structure.
4257 move_newblock_dep(jaddref, inodedep)
4258 struct jaddref *jaddref;
4259 struct inodedep *inodedep;
4261 struct inoref *inoref;
4262 struct jaddref *jaddrefn;
4265 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4266 inoref = TAILQ_NEXT(inoref, if_deps)) {
4267 if ((jaddref->ja_state & NEWBLOCK) &&
4268 inoref->if_list.wk_type == D_JADDREF) {
4269 jaddrefn = (struct jaddref *)inoref;
4273 if (jaddrefn == NULL)
4275 jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
4276 jaddrefn->ja_state |= jaddref->ja_state &
4277 (ATTACHED | UNDONE | NEWBLOCK);
4278 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
4279 jaddref->ja_state |= ATTACHED;
4280 LIST_REMOVE(jaddref, ja_bmdeps);
4281 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
4286 * Cancel a jaddref either before it has been written or while it is being
4287 * written. This happens when a link is removed before the add reaches
4288 * the disk. The jaddref dependency is kept linked into the bmsafemap
4289 * and inode to prevent the link count or bitmap from reaching the disk
4290 * until handle_workitem_remove() re-adjusts the counts and bitmaps as
4293 * Returns 1 if the canceled addref requires journaling of the remove and
4297 cancel_jaddref(jaddref, inodedep, wkhd)
4298 struct jaddref *jaddref;
4299 struct inodedep *inodedep;
4300 struct workhead *wkhd;
4302 struct inoref *inoref;
4303 struct jsegdep *jsegdep;
4306 KASSERT((jaddref->ja_state & COMPLETE) == 0,
4307 ("cancel_jaddref: Canceling complete jaddref"));
4308 if (jaddref->ja_state & (INPROGRESS | COMPLETE))
4312 if (inodedep == NULL)
4313 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
4315 panic("cancel_jaddref: Lost inodedep");
4317 * We must adjust the nlink of any reference operation that follows
4318 * us so that it is consistent with the in-memory reference. This
4319 * ensures that inode nlink rollbacks always have the correct link.
4322 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4323 inoref = TAILQ_NEXT(inoref, if_deps)) {
4324 if (inoref->if_state & GOINGAWAY)
4329 jsegdep = inoref_jseg(&jaddref->ja_ref);
4330 if (jaddref->ja_state & NEWBLOCK)
4331 move_newblock_dep(jaddref, inodedep);
4332 wake_worklist(&jaddref->ja_list);
4333 jaddref->ja_mkdir = NULL;
4334 if (jaddref->ja_state & INPROGRESS) {
4335 jaddref->ja_state &= ~INPROGRESS;
4336 WORKLIST_REMOVE(&jaddref->ja_list);
4337 jwork_insert(wkhd, jsegdep);
4339 free_jsegdep(jsegdep);
4340 if (jaddref->ja_state & DEPCOMPLETE)
4341 remove_from_journal(&jaddref->ja_list);
4343 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE);
4345 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
4346 * can arrange for them to be freed with the bitmap. Otherwise we
4347 * no longer need this addref attached to the inoreflst and it
4348 * will incorrectly adjust nlink if we leave it.
4350 if ((jaddref->ja_state & NEWBLOCK) == 0) {
4351 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
4353 jaddref->ja_state |= COMPLETE;
4354 free_jaddref(jaddref);
4358 * Leave the head of the list for jsegdeps for fast merging.
4360 if (LIST_FIRST(wkhd) != NULL) {
4361 jaddref->ja_state |= ONWORKLIST;
4362 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
4364 WORKLIST_INSERT(wkhd, &jaddref->ja_list);
4370 * Attempt to free a jaddref structure when some work completes. This
4371 * should only succeed once the entry is written and all dependencies have
4375 free_jaddref(jaddref)
4376 struct jaddref *jaddref;
4379 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
4381 if (jaddref->ja_ref.if_jsegdep)
4382 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
4383 jaddref, jaddref->ja_state);
4384 if (jaddref->ja_state & NEWBLOCK)
4385 LIST_REMOVE(jaddref, ja_bmdeps);
4386 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST))
4387 panic("free_jaddref: Bad state %p(0x%X)",
4388 jaddref, jaddref->ja_state);
4389 if (jaddref->ja_mkdir != NULL)
4390 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
4391 WORKITEM_FREE(jaddref, D_JADDREF);
4395 * Free a jremref structure once it has been written or discarded.
4398 free_jremref(jremref)
4399 struct jremref *jremref;
4402 if (jremref->jr_ref.if_jsegdep)
4403 free_jsegdep(jremref->jr_ref.if_jsegdep);
4404 if (jremref->jr_state & INPROGRESS)
4405 panic("free_jremref: IO still pending");
4406 WORKITEM_FREE(jremref, D_JREMREF);
4410 * Free a jnewblk structure.
4413 free_jnewblk(jnewblk)
4414 struct jnewblk *jnewblk;
4417 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
4419 LIST_REMOVE(jnewblk, jn_deps);
4420 if (jnewblk->jn_dep != NULL)
4421 panic("free_jnewblk: Dependency still attached.");
4422 WORKITEM_FREE(jnewblk, D_JNEWBLK);
4426 * Cancel a jnewblk which has been been made redundant by frag extension.
4429 cancel_jnewblk(jnewblk, wkhd)
4430 struct jnewblk *jnewblk;
4431 struct workhead *wkhd;
4433 struct jsegdep *jsegdep;
4435 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno);
4436 jsegdep = jnewblk->jn_jsegdep;
4437 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL)
4438 panic("cancel_jnewblk: Invalid state");
4439 jnewblk->jn_jsegdep = NULL;
4440 jnewblk->jn_dep = NULL;
4441 jnewblk->jn_state |= GOINGAWAY;
4442 if (jnewblk->jn_state & INPROGRESS) {
4443 jnewblk->jn_state &= ~INPROGRESS;
4444 WORKLIST_REMOVE(&jnewblk->jn_list);
4445 jwork_insert(wkhd, jsegdep);
4447 free_jsegdep(jsegdep);
4448 remove_from_journal(&jnewblk->jn_list);
4450 wake_worklist(&jnewblk->jn_list);
4451 WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
4455 free_jblkdep(jblkdep)
4456 struct jblkdep *jblkdep;
4459 if (jblkdep->jb_list.wk_type == D_JFREEBLK)
4460 WORKITEM_FREE(jblkdep, D_JFREEBLK);
4461 else if (jblkdep->jb_list.wk_type == D_JTRUNC)
4462 WORKITEM_FREE(jblkdep, D_JTRUNC);
4464 panic("free_jblkdep: Unexpected type %s",
4465 TYPENAME(jblkdep->jb_list.wk_type));
4469 * Free a single jseg once it is no longer referenced in memory or on
4470 * disk. Reclaim journal blocks and dependencies waiting for the segment
4474 free_jseg(jseg, jblocks)
4476 struct jblocks *jblocks;
4478 struct freework *freework;
4481 * Free freework structures that were lingering to indicate freed
4482 * indirect blocks that forced journal write ordering on reallocate.
4484 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL)
4485 indirblk_remove(freework);
4486 if (jblocks->jb_oldestseg == jseg)
4487 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next);
4488 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
4489 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
4490 KASSERT(LIST_EMPTY(&jseg->js_entries),
4491 ("free_jseg: Freed jseg has valid entries."));
4492 WORKITEM_FREE(jseg, D_JSEG);
4496 * Free all jsegs that meet the criteria for being reclaimed and update
4501 struct jblocks *jblocks;
4506 * Free only those jsegs which have none allocated before them to
4507 * preserve the journal space ordering.
4509 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
4511 * Only reclaim space when nothing depends on this journal
4512 * set and another set has written that it is no longer
4515 if (jseg->js_refs != 0) {
4516 jblocks->jb_oldestseg = jseg;
4519 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE)
4521 if (jseg->js_seq > jblocks->jb_oldestwrseq)
4524 * We can free jsegs that didn't write entries when
4525 * oldestwrseq == js_seq.
4527 if (jseg->js_seq == jblocks->jb_oldestwrseq &&
4530 free_jseg(jseg, jblocks);
4533 * If we exited the loop above we still must discover the
4534 * oldest valid segment.
4537 for (jseg = jblocks->jb_oldestseg; jseg != NULL;
4538 jseg = TAILQ_NEXT(jseg, js_next))
4539 if (jseg->js_refs != 0)
4541 jblocks->jb_oldestseg = jseg;
4543 * The journal has no valid records but some jsegs may still be
4544 * waiting on oldestwrseq to advance. We force a small record
4545 * out to permit these lingering records to be reclaimed.
4547 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs))
4548 jblocks->jb_needseg = 1;
4552 * Release one reference to a jseg and free it if the count reaches 0. This
4553 * should eventually reclaim journal space as well.
4560 KASSERT(jseg->js_refs > 0,
4561 ("free_jseg: Invalid refcnt %d", jseg->js_refs));
4562 if (--jseg->js_refs != 0)
4564 free_jsegs(jseg->js_jblocks);
4568 * Release a jsegdep and decrement the jseg count.
4571 free_jsegdep(jsegdep)
4572 struct jsegdep *jsegdep;
4575 if (jsegdep->jd_seg)
4576 rele_jseg(jsegdep->jd_seg);
4577 WORKITEM_FREE(jsegdep, D_JSEGDEP);
4581 * Wait for a journal item to make it to disk. Initiate journal processing
4586 struct worklist *wk;
4590 LOCK_OWNED(VFSTOUFS(wk->wk_mp));
4592 * Blocking journal waits cause slow synchronous behavior. Record
4593 * stats on the frequency of these blocking operations.
4595 if (waitfor == MNT_WAIT) {
4596 stat_journal_wait++;
4597 switch (wk->wk_type) {
4600 stat_jwait_filepage++;
4604 stat_jwait_freeblks++;
4607 stat_jwait_newblk++;
4617 * If IO has not started we process the journal. We can't mark the
4618 * worklist item as IOWAITING because we drop the lock while
4619 * processing the journal and the worklist entry may be freed after
4620 * this point. The caller may call back in and re-issue the request.
4622 if ((wk->wk_state & INPROGRESS) == 0) {
4623 softdep_process_journal(wk->wk_mp, wk, waitfor);
4624 if (waitfor != MNT_WAIT)
4628 if (waitfor != MNT_WAIT)
4630 wait_worklist(wk, "jwait");
4635 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
4636 * appropriate. This is a convenience function to reduce duplicate code
4637 * for the setup and revert functions below.
4639 static struct inodedep *
4640 inodedep_lookup_ip(ip)
4643 struct inodedep *inodedep;
4645 KASSERT(ip->i_nlink >= ip->i_effnlink,
4646 ("inodedep_lookup_ip: bad delta"));
4647 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
4649 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
4650 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
4656 * Called prior to creating a new inode and linking it to a directory. The
4657 * jaddref structure must already be allocated by softdep_setup_inomapdep
4658 * and it is discovered here so we can initialize the mode and update
4662 softdep_setup_create(dp, ip)
4666 struct inodedep *inodedep;
4667 struct jaddref *jaddref;
4670 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4671 ("softdep_setup_create called on non-softdep filesystem"));
4672 KASSERT(ip->i_nlink == 1,
4673 ("softdep_setup_create: Invalid link count."));
4675 ACQUIRE_LOCK(dp->i_ump);
4676 inodedep = inodedep_lookup_ip(ip);
4677 if (DOINGSUJ(dvp)) {
4678 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4680 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
4681 ("softdep_setup_create: No addref structure present."));
4683 softdep_prelink(dvp, NULL);
4684 FREE_LOCK(dp->i_ump);
4688 * Create a jaddref structure to track the addition of a DOTDOT link when
4689 * we are reparenting an inode as part of a rename. This jaddref will be
4690 * found by softdep_setup_directory_change. Adjusts nlinkdelta for
4691 * non-journaling softdep.
4694 softdep_setup_dotdot_link(dp, ip)
4698 struct inodedep *inodedep;
4699 struct jaddref *jaddref;
4702 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4703 ("softdep_setup_dotdot_link called on non-softdep filesystem"));
4707 * We don't set MKDIR_PARENT as this is not tied to a mkdir and
4708 * is used as a normal link would be.
4711 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4712 dp->i_effnlink - 1, dp->i_mode);
4713 ACQUIRE_LOCK(dp->i_ump);
4714 inodedep = inodedep_lookup_ip(dp);
4716 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4718 softdep_prelink(dvp, ITOV(ip));
4719 FREE_LOCK(dp->i_ump);
4723 * Create a jaddref structure to track a new link to an inode. The directory
4724 * offset is not known until softdep_setup_directory_add or
4725 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling
4729 softdep_setup_link(dp, ip)
4733 struct inodedep *inodedep;
4734 struct jaddref *jaddref;
4737 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4738 ("softdep_setup_link called on non-softdep filesystem"));
4742 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
4744 ACQUIRE_LOCK(dp->i_ump);
4745 inodedep = inodedep_lookup_ip(ip);
4747 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4749 softdep_prelink(dvp, ITOV(ip));
4750 FREE_LOCK(dp->i_ump);
4754 * Called to create the jaddref structures to track . and .. references as
4755 * well as lookup and further initialize the incomplete jaddref created
4756 * by softdep_setup_inomapdep when the inode was allocated. Adjusts
4757 * nlinkdelta for non-journaling softdep.
4760 softdep_setup_mkdir(dp, ip)
4764 struct inodedep *inodedep;
4765 struct jaddref *dotdotaddref;
4766 struct jaddref *dotaddref;
4767 struct jaddref *jaddref;
4770 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4771 ("softdep_setup_mkdir called on non-softdep filesystem"));
4773 dotaddref = dotdotaddref = NULL;
4774 if (DOINGSUJ(dvp)) {
4775 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
4777 dotaddref->ja_state |= MKDIR_BODY;
4778 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4779 dp->i_effnlink - 1, dp->i_mode);
4780 dotdotaddref->ja_state |= MKDIR_PARENT;
4782 ACQUIRE_LOCK(dp->i_ump);
4783 inodedep = inodedep_lookup_ip(ip);
4784 if (DOINGSUJ(dvp)) {
4785 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4787 KASSERT(jaddref != NULL,
4788 ("softdep_setup_mkdir: No addref structure present."));
4789 KASSERT(jaddref->ja_parent == dp->i_number,
4790 ("softdep_setup_mkdir: bad parent %ju",
4791 (uintmax_t)jaddref->ja_parent));
4792 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
4795 inodedep = inodedep_lookup_ip(dp);
4797 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
4798 &dotdotaddref->ja_ref, if_deps);
4799 softdep_prelink(ITOV(dp), NULL);
4800 FREE_LOCK(dp->i_ump);
4804 * Called to track nlinkdelta of the inode and parent directories prior to
4805 * unlinking a directory.
4808 softdep_setup_rmdir(dp, ip)
4814 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4815 ("softdep_setup_rmdir called on non-softdep filesystem"));
4817 ACQUIRE_LOCK(dp->i_ump);
4818 (void) inodedep_lookup_ip(ip);
4819 (void) inodedep_lookup_ip(dp);
4820 softdep_prelink(dvp, ITOV(ip));
4821 FREE_LOCK(dp->i_ump);
4825 * Called to track nlinkdelta of the inode and parent directories prior to
4829 softdep_setup_unlink(dp, ip)
4835 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4836 ("softdep_setup_unlink called on non-softdep filesystem"));
4838 ACQUIRE_LOCK(dp->i_ump);
4839 (void) inodedep_lookup_ip(ip);
4840 (void) inodedep_lookup_ip(dp);
4841 softdep_prelink(dvp, ITOV(ip));
4842 FREE_LOCK(dp->i_ump);
4846 * Called to release the journal structures created by a failed non-directory
4847 * creation. Adjusts nlinkdelta for non-journaling softdep.
4850 softdep_revert_create(dp, ip)
4854 struct inodedep *inodedep;
4855 struct jaddref *jaddref;
4858 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4859 ("softdep_revert_create called on non-softdep filesystem"));
4861 ACQUIRE_LOCK(dp->i_ump);
4862 inodedep = inodedep_lookup_ip(ip);
4863 if (DOINGSUJ(dvp)) {
4864 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4866 KASSERT(jaddref->ja_parent == dp->i_number,
4867 ("softdep_revert_create: addref parent mismatch"));
4868 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4870 FREE_LOCK(dp->i_ump);
4874 * Called to release the journal structures created by a failed link
4875 * addition. Adjusts nlinkdelta for non-journaling softdep.
4878 softdep_revert_link(dp, ip)
4882 struct inodedep *inodedep;
4883 struct jaddref *jaddref;
4886 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4887 ("softdep_revert_link called on non-softdep filesystem"));
4889 ACQUIRE_LOCK(dp->i_ump);
4890 inodedep = inodedep_lookup_ip(ip);
4891 if (DOINGSUJ(dvp)) {
4892 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4894 KASSERT(jaddref->ja_parent == dp->i_number,
4895 ("softdep_revert_link: addref parent mismatch"));
4896 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4898 FREE_LOCK(dp->i_ump);
4902 * Called to release the journal structures created by a failed mkdir
4903 * attempt. Adjusts nlinkdelta for non-journaling softdep.
4906 softdep_revert_mkdir(dp, ip)
4910 struct inodedep *inodedep;
4911 struct jaddref *jaddref;
4912 struct jaddref *dotaddref;
4915 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4916 ("softdep_revert_mkdir called on non-softdep filesystem"));
4919 ACQUIRE_LOCK(dp->i_ump);
4920 inodedep = inodedep_lookup_ip(dp);
4921 if (DOINGSUJ(dvp)) {
4922 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4924 KASSERT(jaddref->ja_parent == ip->i_number,
4925 ("softdep_revert_mkdir: dotdot addref parent mismatch"));
4926 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4928 inodedep = inodedep_lookup_ip(ip);
4929 if (DOINGSUJ(dvp)) {
4930 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4932 KASSERT(jaddref->ja_parent == dp->i_number,
4933 ("softdep_revert_mkdir: addref parent mismatch"));
4934 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
4935 inoreflst, if_deps);
4936 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4937 KASSERT(dotaddref->ja_parent == ip->i_number,
4938 ("softdep_revert_mkdir: dot addref parent mismatch"));
4939 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait);
4941 FREE_LOCK(dp->i_ump);
4945 * Called to correct nlinkdelta after a failed rmdir.
4948 softdep_revert_rmdir(dp, ip)
4953 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4954 ("softdep_revert_rmdir called on non-softdep filesystem"));
4955 ACQUIRE_LOCK(dp->i_ump);
4956 (void) inodedep_lookup_ip(ip);
4957 (void) inodedep_lookup_ip(dp);
4958 FREE_LOCK(dp->i_ump);
4962 * Protecting the freemaps (or bitmaps).
4964 * To eliminate the need to execute fsck before mounting a filesystem
4965 * after a power failure, one must (conservatively) guarantee that the
4966 * on-disk copy of the bitmaps never indicate that a live inode or block is
4967 * free. So, when a block or inode is allocated, the bitmap should be
4968 * updated (on disk) before any new pointers. When a block or inode is
4969 * freed, the bitmap should not be updated until all pointers have been
4970 * reset. The latter dependency is handled by the delayed de-allocation
4971 * approach described below for block and inode de-allocation. The former
4972 * dependency is handled by calling the following procedure when a block or
4973 * inode is allocated. When an inode is allocated an "inodedep" is created
4974 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
4975 * Each "inodedep" is also inserted into the hash indexing structure so
4976 * that any additional link additions can be made dependent on the inode
4979 * The ufs filesystem maintains a number of free block counts (e.g., per
4980 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
4981 * in addition to the bitmaps. These counts are used to improve efficiency
4982 * during allocation and therefore must be consistent with the bitmaps.
4983 * There is no convenient way to guarantee post-crash consistency of these
4984 * counts with simple update ordering, for two main reasons: (1) The counts
4985 * and bitmaps for a single cylinder group block are not in the same disk
4986 * sector. If a disk write is interrupted (e.g., by power failure), one may
4987 * be written and the other not. (2) Some of the counts are located in the
4988 * superblock rather than the cylinder group block. So, we focus our soft
4989 * updates implementation on protecting the bitmaps. When mounting a
4990 * filesystem, we recompute the auxiliary counts from the bitmaps.
4994 * Called just after updating the cylinder group block to allocate an inode.
4997 softdep_setup_inomapdep(bp, ip, newinum, mode)
4998 struct buf *bp; /* buffer for cylgroup block with inode map */
4999 struct inode *ip; /* inode related to allocation */
5000 ino_t newinum; /* new inode number being allocated */
5003 struct inodedep *inodedep;
5004 struct bmsafemap *bmsafemap;
5005 struct jaddref *jaddref;
5009 mp = UFSTOVFS(ip->i_ump);
5010 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5011 ("softdep_setup_inomapdep called on non-softdep filesystem"));
5012 fs = ip->i_ump->um_fs;
5016 * Allocate the journal reference add structure so that the bitmap
5017 * can be dependent on it.
5019 if (MOUNTEDSUJ(mp)) {
5020 jaddref = newjaddref(ip, newinum, 0, 0, mode);
5021 jaddref->ja_state |= NEWBLOCK;
5025 * Create a dependency for the newly allocated inode.
5026 * Panic if it already exists as something is seriously wrong.
5027 * Otherwise add it to the dependency list for the buffer holding
5028 * the cylinder group map from which it was allocated.
5030 * We have to preallocate a bmsafemap entry in case it is needed
5031 * in bmsafemap_lookup since once we allocate the inodedep, we
5032 * have to finish initializing it before we can FREE_LOCK().
5033 * By preallocating, we avoid FREE_LOCK() while doing a malloc
5034 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before
5035 * creating the inodedep as it can be freed during the time
5036 * that we FREE_LOCK() while allocating the inodedep. We must
5037 * call workitem_alloc() before entering the locked section as
5038 * it also acquires the lock and we must avoid trying doing so
5041 bmsafemap = malloc(sizeof(struct bmsafemap),
5042 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5043 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5044 ACQUIRE_LOCK(ip->i_ump);
5045 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep)))
5046 panic("softdep_setup_inomapdep: dependency %p for new"
5047 "inode already exists", inodedep);
5048 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap);
5050 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
5051 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
5054 inodedep->id_state |= ONDEPLIST;
5055 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
5057 inodedep->id_bmsafemap = bmsafemap;
5058 inodedep->id_state &= ~DEPCOMPLETE;
5059 FREE_LOCK(ip->i_ump);
5063 * Called just after updating the cylinder group block to
5064 * allocate block or fragment.
5067 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
5068 struct buf *bp; /* buffer for cylgroup block with block map */
5069 struct mount *mp; /* filesystem doing allocation */
5070 ufs2_daddr_t newblkno; /* number of newly allocated block */
5071 int frags; /* Number of fragments. */
5072 int oldfrags; /* Previous number of fragments for extend. */
5074 struct newblk *newblk;
5075 struct bmsafemap *bmsafemap;
5076 struct jnewblk *jnewblk;
5077 struct ufsmount *ump;
5080 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5081 ("softdep_setup_blkmapdep called on non-softdep filesystem"));
5086 * Create a dependency for the newly allocated block.
5087 * Add it to the dependency list for the buffer holding
5088 * the cylinder group map from which it was allocated.
5090 if (MOUNTEDSUJ(mp)) {
5091 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
5092 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
5093 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
5094 jnewblk->jn_state = ATTACHED;
5095 jnewblk->jn_blkno = newblkno;
5096 jnewblk->jn_frags = frags;
5097 jnewblk->jn_oldfrags = oldfrags;
5105 cgp = (struct cg *)bp->b_data;
5106 blksfree = cg_blksfree(cgp);
5107 bno = dtogd(fs, jnewblk->jn_blkno);
5108 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
5110 if (isset(blksfree, bno + i))
5111 panic("softdep_setup_blkmapdep: "
5112 "free fragment %d from %d-%d "
5113 "state 0x%X dep %p", i,
5114 jnewblk->jn_oldfrags,
5124 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d",
5125 newblkno, frags, oldfrags);
5127 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
5128 panic("softdep_setup_blkmapdep: found block");
5129 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
5130 dtog(fs, newblkno), NULL);
5132 jnewblk->jn_dep = (struct worklist *)newblk;
5133 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
5135 newblk->nb_state |= ONDEPLIST;
5136 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
5138 newblk->nb_bmsafemap = bmsafemap;
5139 newblk->nb_jnewblk = jnewblk;
5143 #define BMSAFEMAP_HASH(ump, cg) \
5144 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size])
5147 bmsafemap_find(bmsafemaphd, cg, bmsafemapp)
5148 struct bmsafemap_hashhead *bmsafemaphd;
5150 struct bmsafemap **bmsafemapp;
5152 struct bmsafemap *bmsafemap;
5154 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
5155 if (bmsafemap->sm_cg == cg)
5158 *bmsafemapp = bmsafemap;
5167 * Find the bmsafemap associated with a cylinder group buffer.
5168 * If none exists, create one. The buffer must be locked when
5169 * this routine is called and this routine must be called with
5170 * the softdep lock held. To avoid giving up the lock while
5171 * allocating a new bmsafemap, a preallocated bmsafemap may be
5172 * provided. If it is provided but not needed, it is freed.
5174 static struct bmsafemap *
5175 bmsafemap_lookup(mp, bp, cg, newbmsafemap)
5179 struct bmsafemap *newbmsafemap;
5181 struct bmsafemap_hashhead *bmsafemaphd;
5182 struct bmsafemap *bmsafemap, *collision;
5183 struct worklist *wk;
5184 struct ufsmount *ump;
5188 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer"));
5189 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5190 if (wk->wk_type == D_BMSAFEMAP) {
5192 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5193 return (WK_BMSAFEMAP(wk));
5196 bmsafemaphd = BMSAFEMAP_HASH(ump, cg);
5197 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) {
5199 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5203 bmsafemap = newbmsafemap;
5206 bmsafemap = malloc(sizeof(struct bmsafemap),
5207 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5208 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5211 bmsafemap->sm_buf = bp;
5212 LIST_INIT(&bmsafemap->sm_inodedephd);
5213 LIST_INIT(&bmsafemap->sm_inodedepwr);
5214 LIST_INIT(&bmsafemap->sm_newblkhd);
5215 LIST_INIT(&bmsafemap->sm_newblkwr);
5216 LIST_INIT(&bmsafemap->sm_jaddrefhd);
5217 LIST_INIT(&bmsafemap->sm_jnewblkhd);
5218 LIST_INIT(&bmsafemap->sm_freehd);
5219 LIST_INIT(&bmsafemap->sm_freewr);
5220 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) {
5221 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
5224 bmsafemap->sm_cg = cg;
5225 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
5226 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
5227 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
5232 * Direct block allocation dependencies.
5234 * When a new block is allocated, the corresponding disk locations must be
5235 * initialized (with zeros or new data) before the on-disk inode points to
5236 * them. Also, the freemap from which the block was allocated must be
5237 * updated (on disk) before the inode's pointer. These two dependencies are
5238 * independent of each other and are needed for all file blocks and indirect
5239 * blocks that are pointed to directly by the inode. Just before the
5240 * "in-core" version of the inode is updated with a newly allocated block
5241 * number, a procedure (below) is called to setup allocation dependency
5242 * structures. These structures are removed when the corresponding
5243 * dependencies are satisfied or when the block allocation becomes obsolete
5244 * (i.e., the file is deleted, the block is de-allocated, or the block is a
5245 * fragment that gets upgraded). All of these cases are handled in
5246 * procedures described later.
5248 * When a file extension causes a fragment to be upgraded, either to a larger
5249 * fragment or to a full block, the on-disk location may change (if the
5250 * previous fragment could not simply be extended). In this case, the old
5251 * fragment must be de-allocated, but not until after the inode's pointer has
5252 * been updated. In most cases, this is handled by later procedures, which
5253 * will construct a "freefrag" structure to be added to the workitem queue
5254 * when the inode update is complete (or obsolete). The main exception to
5255 * this is when an allocation occurs while a pending allocation dependency
5256 * (for the same block pointer) remains. This case is handled in the main
5257 * allocation dependency setup procedure by immediately freeing the
5258 * unreferenced fragments.
5261 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5262 struct inode *ip; /* inode to which block is being added */
5263 ufs_lbn_t off; /* block pointer within inode */
5264 ufs2_daddr_t newblkno; /* disk block number being added */
5265 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */
5266 long newsize; /* size of new block */
5267 long oldsize; /* size of new block */
5268 struct buf *bp; /* bp for allocated block */
5270 struct allocdirect *adp, *oldadp;
5271 struct allocdirectlst *adphead;
5272 struct freefrag *freefrag;
5273 struct inodedep *inodedep;
5274 struct pagedep *pagedep;
5275 struct jnewblk *jnewblk;
5276 struct newblk *newblk;
5281 mp = UFSTOVFS(ip->i_ump);
5282 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5283 ("softdep_setup_allocdirect called on non-softdep filesystem"));
5284 if (oldblkno && oldblkno != newblkno)
5285 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5290 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd "
5291 "off %jd newsize %ld oldsize %d",
5292 ip->i_number, newblkno, oldblkno, off, newsize, oldsize);
5293 ACQUIRE_LOCK(ip->i_ump);
5294 if (off >= NDADDR) {
5296 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
5298 /* allocating an indirect block */
5300 panic("softdep_setup_allocdirect: non-zero indir");
5303 panic("softdep_setup_allocdirect: lbn %jd != off %jd",
5306 * Allocating a direct block.
5308 * If we are allocating a directory block, then we must
5309 * allocate an associated pagedep to track additions and
5312 if ((ip->i_mode & IFMT) == IFDIR)
5313 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC,
5316 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5317 panic("softdep_setup_allocdirect: lost block");
5318 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5319 ("softdep_setup_allocdirect: newblk already initialized"));
5321 * Convert the newblk to an allocdirect.
5323 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5324 adp = (struct allocdirect *)newblk;
5325 newblk->nb_freefrag = freefrag;
5326 adp->ad_offset = off;
5327 adp->ad_oldblkno = oldblkno;
5328 adp->ad_newsize = newsize;
5329 adp->ad_oldsize = oldsize;
5332 * Finish initializing the journal.
5334 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5335 jnewblk->jn_ino = ip->i_number;
5336 jnewblk->jn_lbn = lbn;
5337 add_to_journal(&jnewblk->jn_list);
5339 if (freefrag && freefrag->ff_jdep != NULL &&
5340 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5341 add_to_journal(freefrag->ff_jdep);
5342 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5343 adp->ad_inodedep = inodedep;
5345 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5347 * The list of allocdirects must be kept in sorted and ascending
5348 * order so that the rollback routines can quickly determine the
5349 * first uncommitted block (the size of the file stored on disk
5350 * ends at the end of the lowest committed fragment, or if there
5351 * are no fragments, at the end of the highest committed block).
5352 * Since files generally grow, the typical case is that the new
5353 * block is to be added at the end of the list. We speed this
5354 * special case by checking against the last allocdirect in the
5355 * list before laboriously traversing the list looking for the
5358 adphead = &inodedep->id_newinoupdt;
5359 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5360 if (oldadp == NULL || oldadp->ad_offset <= off) {
5361 /* insert at end of list */
5362 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5363 if (oldadp != NULL && oldadp->ad_offset == off)
5364 allocdirect_merge(adphead, adp, oldadp);
5365 FREE_LOCK(ip->i_ump);
5368 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5369 if (oldadp->ad_offset >= off)
5373 panic("softdep_setup_allocdirect: lost entry");
5374 /* insert in middle of list */
5375 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5376 if (oldadp->ad_offset == off)
5377 allocdirect_merge(adphead, adp, oldadp);
5379 FREE_LOCK(ip->i_ump);
5383 * Merge a newer and older journal record to be stored either in a
5384 * newblock or freefrag. This handles aggregating journal records for
5385 * fragment allocation into a second record as well as replacing a
5386 * journal free with an aborted journal allocation. A segment for the
5387 * oldest record will be placed on wkhd if it has been written. If not
5388 * the segment for the newer record will suffice.
5390 static struct worklist *
5391 jnewblk_merge(new, old, wkhd)
5392 struct worklist *new;
5393 struct worklist *old;
5394 struct workhead *wkhd;
5396 struct jnewblk *njnewblk;
5397 struct jnewblk *jnewblk;
5399 /* Handle NULLs to simplify callers. */
5404 /* Replace a jfreefrag with a jnewblk. */
5405 if (new->wk_type == D_JFREEFRAG) {
5406 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno)
5407 panic("jnewblk_merge: blkno mismatch: %p, %p",
5409 cancel_jfreefrag(WK_JFREEFRAG(new));
5412 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK)
5413 panic("jnewblk_merge: Bad type: old %d new %d\n",
5414 old->wk_type, new->wk_type);
5416 * Handle merging of two jnewblk records that describe
5417 * different sets of fragments in the same block.
5419 jnewblk = WK_JNEWBLK(old);
5420 njnewblk = WK_JNEWBLK(new);
5421 if (jnewblk->jn_blkno != njnewblk->jn_blkno)
5422 panic("jnewblk_merge: Merging disparate blocks.");
5424 * The record may be rolled back in the cg.
5426 if (jnewblk->jn_state & UNDONE) {
5427 jnewblk->jn_state &= ~UNDONE;
5428 njnewblk->jn_state |= UNDONE;
5429 njnewblk->jn_state &= ~ATTACHED;
5432 * We modify the newer addref and free the older so that if neither
5433 * has been written the most up-to-date copy will be on disk. If
5434 * both have been written but rolled back we only temporarily need
5435 * one of them to fix the bits when the cg write completes.
5437 jnewblk->jn_state |= ATTACHED | COMPLETE;
5438 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
5439 cancel_jnewblk(jnewblk, wkhd);
5440 WORKLIST_REMOVE(&jnewblk->jn_list);
5441 free_jnewblk(jnewblk);
5446 * Replace an old allocdirect dependency with a newer one.
5447 * This routine must be called with splbio interrupts blocked.
5450 allocdirect_merge(adphead, newadp, oldadp)
5451 struct allocdirectlst *adphead; /* head of list holding allocdirects */
5452 struct allocdirect *newadp; /* allocdirect being added */
5453 struct allocdirect *oldadp; /* existing allocdirect being checked */
5455 struct worklist *wk;
5456 struct freefrag *freefrag;
5459 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp));
5460 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
5461 newadp->ad_oldsize != oldadp->ad_newsize ||
5462 newadp->ad_offset >= NDADDR)
5463 panic("%s %jd != new %jd || old size %ld != new %ld",
5464 "allocdirect_merge: old blkno",
5465 (intmax_t)newadp->ad_oldblkno,
5466 (intmax_t)oldadp->ad_newblkno,
5467 newadp->ad_oldsize, oldadp->ad_newsize);
5468 newadp->ad_oldblkno = oldadp->ad_oldblkno;
5469 newadp->ad_oldsize = oldadp->ad_oldsize;
5471 * If the old dependency had a fragment to free or had never
5472 * previously had a block allocated, then the new dependency
5473 * can immediately post its freefrag and adopt the old freefrag.
5474 * This action is done by swapping the freefrag dependencies.
5475 * The new dependency gains the old one's freefrag, and the
5476 * old one gets the new one and then immediately puts it on
5477 * the worklist when it is freed by free_newblk. It is
5478 * not possible to do this swap when the old dependency had a
5479 * non-zero size but no previous fragment to free. This condition
5480 * arises when the new block is an extension of the old block.
5481 * Here, the first part of the fragment allocated to the new
5482 * dependency is part of the block currently claimed on disk by
5483 * the old dependency, so cannot legitimately be freed until the
5484 * conditions for the new dependency are fulfilled.
5486 freefrag = newadp->ad_freefrag;
5487 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
5488 newadp->ad_freefrag = oldadp->ad_freefrag;
5489 oldadp->ad_freefrag = freefrag;
5492 * If we are tracking a new directory-block allocation,
5493 * move it from the old allocdirect to the new allocdirect.
5495 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
5496 WORKLIST_REMOVE(wk);
5497 if (!LIST_EMPTY(&oldadp->ad_newdirblk))
5498 panic("allocdirect_merge: extra newdirblk");
5499 WORKLIST_INSERT(&newadp->ad_newdirblk, wk);
5501 TAILQ_REMOVE(adphead, oldadp, ad_next);
5503 * We need to move any journal dependencies over to the freefrag
5504 * that releases this block if it exists. Otherwise we are
5505 * extending an existing block and we'll wait until that is
5506 * complete to release the journal space and extend the
5507 * new journal to cover this old space as well.
5509 if (freefrag == NULL) {
5510 if (oldadp->ad_newblkno != newadp->ad_newblkno)
5511 panic("allocdirect_merge: %jd != %jd",
5512 oldadp->ad_newblkno, newadp->ad_newblkno);
5513 newadp->ad_block.nb_jnewblk = (struct jnewblk *)
5514 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list,
5515 &oldadp->ad_block.nb_jnewblk->jn_list,
5516 &newadp->ad_block.nb_jwork);
5517 oldadp->ad_block.nb_jnewblk = NULL;
5518 cancel_newblk(&oldadp->ad_block, NULL,
5519 &newadp->ad_block.nb_jwork);
5521 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block,
5522 &freefrag->ff_list, &freefrag->ff_jwork);
5523 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk,
5524 &freefrag->ff_jwork);
5526 free_newblk(&oldadp->ad_block);
5530 * Allocate a jfreefrag structure to journal a single block free.
5532 static struct jfreefrag *
5533 newjfreefrag(freefrag, ip, blkno, size, lbn)
5534 struct freefrag *freefrag;
5540 struct jfreefrag *jfreefrag;
5544 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
5546 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump));
5547 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
5548 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
5549 jfreefrag->fr_ino = ip->i_number;
5550 jfreefrag->fr_lbn = lbn;
5551 jfreefrag->fr_blkno = blkno;
5552 jfreefrag->fr_frags = numfrags(fs, size);
5553 jfreefrag->fr_freefrag = freefrag;
5559 * Allocate a new freefrag structure.
5561 static struct freefrag *
5562 newfreefrag(ip, blkno, size, lbn)
5568 struct freefrag *freefrag;
5571 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd",
5572 ip->i_number, blkno, size, lbn);
5574 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
5575 panic("newfreefrag: frag size");
5576 freefrag = malloc(sizeof(struct freefrag),
5577 M_FREEFRAG, M_SOFTDEP_FLAGS);
5578 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
5579 freefrag->ff_state = ATTACHED;
5580 LIST_INIT(&freefrag->ff_jwork);
5581 freefrag->ff_inum = ip->i_number;
5582 freefrag->ff_vtype = ITOV(ip)->v_type;
5583 freefrag->ff_blkno = blkno;
5584 freefrag->ff_fragsize = size;
5586 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) {
5587 freefrag->ff_jdep = (struct worklist *)
5588 newjfreefrag(freefrag, ip, blkno, size, lbn);
5590 freefrag->ff_state |= DEPCOMPLETE;
5591 freefrag->ff_jdep = NULL;
5598 * This workitem de-allocates fragments that were replaced during
5599 * file block allocation.
5602 handle_workitem_freefrag(freefrag)
5603 struct freefrag *freefrag;
5605 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
5606 struct workhead wkhd;
5609 "handle_workitem_freefrag: ino %d blkno %jd size %ld",
5610 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize);
5612 * It would be illegal to add new completion items to the
5613 * freefrag after it was schedule to be done so it must be
5614 * safe to modify the list head here.
5618 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
5620 * If the journal has not been written we must cancel it here.
5622 if (freefrag->ff_jdep) {
5623 if (freefrag->ff_jdep->wk_type != D_JNEWBLK)
5624 panic("handle_workitem_freefrag: Unexpected type %d\n",
5625 freefrag->ff_jdep->wk_type);
5626 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd);
5629 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
5630 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd);
5632 WORKITEM_FREE(freefrag, D_FREEFRAG);
5637 * Set up a dependency structure for an external attributes data block.
5638 * This routine follows much of the structure of softdep_setup_allocdirect.
5639 * See the description of softdep_setup_allocdirect above for details.
5642 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5645 ufs2_daddr_t newblkno;
5646 ufs2_daddr_t oldblkno;
5651 struct allocdirect *adp, *oldadp;
5652 struct allocdirectlst *adphead;
5653 struct freefrag *freefrag;
5654 struct inodedep *inodedep;
5655 struct jnewblk *jnewblk;
5656 struct newblk *newblk;
5660 mp = UFSTOVFS(ip->i_ump);
5661 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5662 ("softdep_setup_allocext called on non-softdep filesystem"));
5663 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR",
5667 if (oldblkno && oldblkno != newblkno)
5668 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5672 ACQUIRE_LOCK(ip->i_ump);
5673 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5674 panic("softdep_setup_allocext: lost block");
5675 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5676 ("softdep_setup_allocext: newblk already initialized"));
5678 * Convert the newblk to an allocdirect.
5680 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5681 adp = (struct allocdirect *)newblk;
5682 newblk->nb_freefrag = freefrag;
5683 adp->ad_offset = off;
5684 adp->ad_oldblkno = oldblkno;
5685 adp->ad_newsize = newsize;
5686 adp->ad_oldsize = oldsize;
5687 adp->ad_state |= EXTDATA;
5690 * Finish initializing the journal.
5692 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5693 jnewblk->jn_ino = ip->i_number;
5694 jnewblk->jn_lbn = lbn;
5695 add_to_journal(&jnewblk->jn_list);
5697 if (freefrag && freefrag->ff_jdep != NULL &&
5698 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5699 add_to_journal(freefrag->ff_jdep);
5700 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5701 adp->ad_inodedep = inodedep;
5703 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5705 * The list of allocdirects must be kept in sorted and ascending
5706 * order so that the rollback routines can quickly determine the
5707 * first uncommitted block (the size of the file stored on disk
5708 * ends at the end of the lowest committed fragment, or if there
5709 * are no fragments, at the end of the highest committed block).
5710 * Since files generally grow, the typical case is that the new
5711 * block is to be added at the end of the list. We speed this
5712 * special case by checking against the last allocdirect in the
5713 * list before laboriously traversing the list looking for the
5716 adphead = &inodedep->id_newextupdt;
5717 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5718 if (oldadp == NULL || oldadp->ad_offset <= off) {
5719 /* insert at end of list */
5720 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5721 if (oldadp != NULL && oldadp->ad_offset == off)
5722 allocdirect_merge(adphead, adp, oldadp);
5723 FREE_LOCK(ip->i_ump);
5726 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5727 if (oldadp->ad_offset >= off)
5731 panic("softdep_setup_allocext: lost entry");
5732 /* insert in middle of list */
5733 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5734 if (oldadp->ad_offset == off)
5735 allocdirect_merge(adphead, adp, oldadp);
5736 FREE_LOCK(ip->i_ump);
5740 * Indirect block allocation dependencies.
5742 * The same dependencies that exist for a direct block also exist when
5743 * a new block is allocated and pointed to by an entry in a block of
5744 * indirect pointers. The undo/redo states described above are also
5745 * used here. Because an indirect block contains many pointers that
5746 * may have dependencies, a second copy of the entire in-memory indirect
5747 * block is kept. The buffer cache copy is always completely up-to-date.
5748 * The second copy, which is used only as a source for disk writes,
5749 * contains only the safe pointers (i.e., those that have no remaining
5750 * update dependencies). The second copy is freed when all pointers
5751 * are safe. The cache is not allowed to replace indirect blocks with
5752 * pending update dependencies. If a buffer containing an indirect
5753 * block with dependencies is written, these routines will mark it
5754 * dirty again. It can only be successfully written once all the
5755 * dependencies are removed. The ffs_fsync routine in conjunction with
5756 * softdep_sync_metadata work together to get all the dependencies
5757 * removed so that a file can be successfully written to disk. Three
5758 * procedures are used when setting up indirect block pointer
5759 * dependencies. The division is necessary because of the organization
5760 * of the "balloc" routine and because of the distinction between file
5761 * pages and file metadata blocks.
5765 * Allocate a new allocindir structure.
5767 static struct allocindir *
5768 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
5769 struct inode *ip; /* inode for file being extended */
5770 int ptrno; /* offset of pointer in indirect block */
5771 ufs2_daddr_t newblkno; /* disk block number being added */
5772 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5775 struct newblk *newblk;
5776 struct allocindir *aip;
5777 struct freefrag *freefrag;
5778 struct jnewblk *jnewblk;
5781 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn);
5784 ACQUIRE_LOCK(ip->i_ump);
5785 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0)
5786 panic("new_allocindir: lost block");
5787 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5788 ("newallocindir: newblk already initialized"));
5789 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR);
5790 newblk->nb_freefrag = freefrag;
5791 aip = (struct allocindir *)newblk;
5792 aip->ai_offset = ptrno;
5793 aip->ai_oldblkno = oldblkno;
5795 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5796 jnewblk->jn_ino = ip->i_number;
5797 jnewblk->jn_lbn = lbn;
5798 add_to_journal(&jnewblk->jn_list);
5800 if (freefrag && freefrag->ff_jdep != NULL &&
5801 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5802 add_to_journal(freefrag->ff_jdep);
5807 * Called just before setting an indirect block pointer
5808 * to a newly allocated file page.
5811 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
5812 struct inode *ip; /* inode for file being extended */
5813 ufs_lbn_t lbn; /* allocated block number within file */
5814 struct buf *bp; /* buffer with indirect blk referencing page */
5815 int ptrno; /* offset of pointer in indirect block */
5816 ufs2_daddr_t newblkno; /* disk block number being added */
5817 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5818 struct buf *nbp; /* buffer holding allocated page */
5820 struct inodedep *inodedep;
5821 struct freefrag *freefrag;
5822 struct allocindir *aip;
5823 struct pagedep *pagedep;
5826 mp = UFSTOVFS(ip->i_ump);
5827 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5828 ("softdep_setup_allocindir_page called on non-softdep filesystem"));
5829 KASSERT(lbn == nbp->b_lblkno,
5830 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
5831 lbn, bp->b_lblkno));
5833 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd "
5834 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn);
5835 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
5836 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
5837 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5839 * If we are allocating a directory page, then we must
5840 * allocate an associated pagedep to track additions and
5843 if ((ip->i_mode & IFMT) == IFDIR)
5844 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep);
5845 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5846 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
5847 FREE_LOCK(ip->i_ump);
5849 handle_workitem_freefrag(freefrag);
5853 * Called just before setting an indirect block pointer to a
5854 * newly allocated indirect block.
5857 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
5858 struct buf *nbp; /* newly allocated indirect block */
5859 struct inode *ip; /* inode for file being extended */
5860 struct buf *bp; /* indirect block referencing allocated block */
5861 int ptrno; /* offset of pointer in indirect block */
5862 ufs2_daddr_t newblkno; /* disk block number being added */
5864 struct inodedep *inodedep;
5865 struct allocindir *aip;
5868 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
5869 ("softdep_setup_allocindir_meta called on non-softdep filesystem"));
5871 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d",
5872 ip->i_number, newblkno, ptrno);
5873 lbn = nbp->b_lblkno;
5874 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
5875 aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
5876 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
5878 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5879 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
5880 panic("softdep_setup_allocindir_meta: Block already existed");
5881 FREE_LOCK(ip->i_ump);
5885 indirdep_complete(indirdep)
5886 struct indirdep *indirdep;
5888 struct allocindir *aip;
5890 LIST_REMOVE(indirdep, ir_next);
5891 indirdep->ir_state |= DEPCOMPLETE;
5893 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
5894 LIST_REMOVE(aip, ai_next);
5895 free_newblk(&aip->ai_block);
5898 * If this indirdep is not attached to a buf it was simply waiting
5899 * on completion to clear completehd. free_indirdep() asserts
5900 * that nothing is dangling.
5902 if ((indirdep->ir_state & ONWORKLIST) == 0)
5903 free_indirdep(indirdep);
5906 static struct indirdep *
5907 indirdep_lookup(mp, ip, bp)
5912 struct indirdep *indirdep, *newindirdep;
5913 struct newblk *newblk;
5914 struct ufsmount *ump;
5915 struct worklist *wk;
5925 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5926 if (wk->wk_type != D_INDIRDEP)
5928 indirdep = WK_INDIRDEP(wk);
5931 /* Found on the buffer worklist, no new structure to free. */
5932 if (indirdep != NULL && newindirdep == NULL)
5934 if (indirdep != NULL && newindirdep != NULL)
5935 panic("indirdep_lookup: simultaneous create");
5936 /* None found on the buffer and a new structure is ready. */
5937 if (indirdep == NULL && newindirdep != NULL)
5939 /* None found and no new structure available. */
5941 newindirdep = malloc(sizeof(struct indirdep),
5942 M_INDIRDEP, M_SOFTDEP_FLAGS);
5943 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
5944 newindirdep->ir_state = ATTACHED;
5945 if (ip->i_ump->um_fstype == UFS1)
5946 newindirdep->ir_state |= UFS1FMT;
5947 TAILQ_INIT(&newindirdep->ir_trunc);
5948 newindirdep->ir_saveddata = NULL;
5949 LIST_INIT(&newindirdep->ir_deplisthd);
5950 LIST_INIT(&newindirdep->ir_donehd);
5951 LIST_INIT(&newindirdep->ir_writehd);
5952 LIST_INIT(&newindirdep->ir_completehd);
5953 if (bp->b_blkno == bp->b_lblkno) {
5954 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
5956 bp->b_blkno = blkno;
5958 newindirdep->ir_freeblks = NULL;
5959 newindirdep->ir_savebp =
5960 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
5961 newindirdep->ir_bp = bp;
5962 BUF_KERNPROC(newindirdep->ir_savebp);
5963 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
5966 indirdep = newindirdep;
5967 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
5969 * If the block is not yet allocated we don't set DEPCOMPLETE so
5970 * that we don't free dependencies until the pointers are valid.
5971 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather
5972 * than using the hash.
5974 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk))
5975 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next);
5977 indirdep->ir_state |= DEPCOMPLETE;
5982 * Called to finish the allocation of the "aip" allocated
5983 * by one of the two routines above.
5985 static struct freefrag *
5986 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
5987 struct buf *bp; /* in-memory copy of the indirect block */
5988 struct inode *ip; /* inode for file being extended */
5989 struct inodedep *inodedep; /* Inodedep for ip */
5990 struct allocindir *aip; /* allocindir allocated by the above routines */
5991 ufs_lbn_t lbn; /* Logical block number for this block. */
5994 struct indirdep *indirdep;
5995 struct allocindir *oldaip;
5996 struct freefrag *freefrag;
5999 LOCK_OWNED(ip->i_ump);
6000 mp = UFSTOVFS(ip->i_ump);
6002 if (bp->b_lblkno >= 0)
6003 panic("setup_allocindir_phase2: not indir blk");
6004 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs),
6005 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset));
6006 indirdep = indirdep_lookup(mp, ip, bp);
6007 KASSERT(indirdep->ir_savebp != NULL,
6008 ("setup_allocindir_phase2 NULL ir_savebp"));
6009 aip->ai_indirdep = indirdep;
6011 * Check for an unwritten dependency for this indirect offset. If
6012 * there is, merge the old dependency into the new one. This happens
6013 * as a result of reallocblk only.
6016 if (aip->ai_oldblkno != 0) {
6017 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) {
6018 if (oldaip->ai_offset == aip->ai_offset) {
6019 freefrag = allocindir_merge(aip, oldaip);
6023 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) {
6024 if (oldaip->ai_offset == aip->ai_offset) {
6025 freefrag = allocindir_merge(aip, oldaip);
6031 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
6036 * Merge two allocindirs which refer to the same block. Move newblock
6037 * dependencies and setup the freefrags appropriately.
6039 static struct freefrag *
6040 allocindir_merge(aip, oldaip)
6041 struct allocindir *aip;
6042 struct allocindir *oldaip;
6044 struct freefrag *freefrag;
6045 struct worklist *wk;
6047 if (oldaip->ai_newblkno != aip->ai_oldblkno)
6048 panic("allocindir_merge: blkno");
6049 aip->ai_oldblkno = oldaip->ai_oldblkno;
6050 freefrag = aip->ai_freefrag;
6051 aip->ai_freefrag = oldaip->ai_freefrag;
6052 oldaip->ai_freefrag = NULL;
6053 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
6055 * If we are tracking a new directory-block allocation,
6056 * move it from the old allocindir to the new allocindir.
6058 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
6059 WORKLIST_REMOVE(wk);
6060 if (!LIST_EMPTY(&oldaip->ai_newdirblk))
6061 panic("allocindir_merge: extra newdirblk");
6062 WORKLIST_INSERT(&aip->ai_newdirblk, wk);
6065 * We can skip journaling for this freefrag and just complete
6066 * any pending journal work for the allocindir that is being
6067 * removed after the freefrag completes.
6069 if (freefrag->ff_jdep)
6070 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep));
6071 LIST_REMOVE(oldaip, ai_next);
6072 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block,
6073 &freefrag->ff_list, &freefrag->ff_jwork);
6074 free_newblk(&oldaip->ai_block);
6080 setup_freedirect(freeblks, ip, i, needj)
6081 struct freeblks *freeblks;
6089 blkno = DIP(ip, i_db[i]);
6092 DIP_SET(ip, i_db[i], 0);
6093 frags = sblksize(ip->i_fs, ip->i_size, i);
6094 frags = numfrags(ip->i_fs, frags);
6095 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj);
6099 setup_freeext(freeblks, ip, i, needj)
6100 struct freeblks *freeblks;
6108 blkno = ip->i_din2->di_extb[i];
6111 ip->i_din2->di_extb[i] = 0;
6112 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i);
6113 frags = numfrags(ip->i_fs, frags);
6114 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj);
6118 setup_freeindir(freeblks, ip, i, lbn, needj)
6119 struct freeblks *freeblks;
6127 blkno = DIP(ip, i_ib[i]);
6130 DIP_SET(ip, i_ib[i], 0);
6131 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag,
6135 static inline struct freeblks *
6140 struct freeblks *freeblks;
6142 freeblks = malloc(sizeof(struct freeblks),
6143 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
6144 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
6145 LIST_INIT(&freeblks->fb_jblkdephd);
6146 LIST_INIT(&freeblks->fb_jwork);
6147 freeblks->fb_ref = 0;
6148 freeblks->fb_cgwait = 0;
6149 freeblks->fb_state = ATTACHED;
6150 freeblks->fb_uid = ip->i_uid;
6151 freeblks->fb_inum = ip->i_number;
6152 freeblks->fb_vtype = ITOV(ip)->v_type;
6153 freeblks->fb_modrev = DIP(ip, i_modrev);
6154 freeblks->fb_devvp = ip->i_devvp;
6155 freeblks->fb_chkcnt = 0;
6156 freeblks->fb_len = 0;
6162 trunc_indirdep(indirdep, freeblks, bp, off)
6163 struct indirdep *indirdep;
6164 struct freeblks *freeblks;
6168 struct allocindir *aip, *aipn;
6171 * The first set of allocindirs won't be in savedbp.
6173 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn)
6174 if (aip->ai_offset > off)
6175 cancel_allocindir(aip, bp, freeblks, 1);
6176 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn)
6177 if (aip->ai_offset > off)
6178 cancel_allocindir(aip, bp, freeblks, 1);
6180 * These will exist in savedbp.
6182 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn)
6183 if (aip->ai_offset > off)
6184 cancel_allocindir(aip, NULL, freeblks, 0);
6185 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn)
6186 if (aip->ai_offset > off)
6187 cancel_allocindir(aip, NULL, freeblks, 0);
6191 * Follow the chain of indirects down to lastlbn creating a freework
6192 * structure for each. This will be used to start indir_trunc() at
6193 * the right offset and create the journal records for the parrtial
6194 * truncation. A second step will handle the truncated dependencies.
6197 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno)
6198 struct freeblks *freeblks;
6204 struct indirdep *indirdep;
6205 struct indirdep *indirn;
6206 struct freework *freework;
6207 struct newblk *newblk;
6221 mp = freeblks->fb_list.wk_mp;
6222 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
6223 if ((bp->b_flags & B_CACHE) == 0) {
6224 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno);
6225 bp->b_iocmd = BIO_READ;
6226 bp->b_flags &= ~B_INVAL;
6227 bp->b_ioflags &= ~BIO_ERROR;
6228 vfs_busy_pages(bp, 0);
6229 bp->b_iooffset = dbtob(bp->b_blkno);
6231 curthread->td_ru.ru_inblock++;
6232 error = bufwait(bp);
6238 level = lbn_level(lbn);
6239 lbnadd = lbn_offset(ip->i_fs, level);
6241 * Compute the offset of the last block we want to keep. Store
6242 * in the freework the first block we want to completely free.
6244 off = (lastlbn - -(lbn + level)) / lbnadd;
6245 if (off + 1 == NINDIR(ip->i_fs))
6247 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1,
6250 * Link the freework into the indirdep. This will prevent any new
6251 * allocations from proceeding until we are finished with the
6252 * truncate and the block is written.
6254 ACQUIRE_LOCK(ip->i_ump);
6255 indirdep = indirdep_lookup(mp, ip, bp);
6256 if (indirdep->ir_freeblks)
6257 panic("setup_trunc_indir: indirdep already truncated.");
6258 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next);
6259 freework->fw_indir = indirdep;
6261 * Cancel any allocindirs that will not make it to disk.
6262 * We have to do this for all copies of the indirdep that
6263 * live on this newblk.
6265 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
6266 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk);
6267 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next)
6268 trunc_indirdep(indirn, freeblks, bp, off);
6270 trunc_indirdep(indirdep, freeblks, bp, off);
6271 FREE_LOCK(ip->i_ump);
6273 * Creation is protected by the buf lock. The saveddata is only
6274 * needed if a full truncation follows a partial truncation but it
6275 * is difficult to allocate in that case so we fetch it anyway.
6277 if (indirdep->ir_saveddata == NULL)
6278 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
6281 /* Fetch the blkno of the child and the zero start offset. */
6282 if (ip->i_ump->um_fstype == UFS1) {
6283 blkno = ((ufs1_daddr_t *)bp->b_data)[off];
6284 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1];
6286 blkno = ((ufs2_daddr_t *)bp->b_data)[off];
6287 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1];
6290 /* Zero the truncated pointers. */
6291 end = bp->b_data + bp->b_bcount;
6292 bzero(start, end - start);
6298 lbn++; /* adjust level */
6299 lbn -= (off * lbnadd);
6300 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno);
6304 * Complete the partial truncation of an indirect block setup by
6305 * setup_trunc_indir(). This zeros the truncated pointers in the saved
6306 * copy and writes them to disk before the freeblks is allowed to complete.
6309 complete_trunc_indir(freework)
6310 struct freework *freework;
6312 struct freework *fwn;
6313 struct indirdep *indirdep;
6314 struct ufsmount *ump;
6319 ump = VFSTOUFS(freework->fw_list.wk_mp);
6321 indirdep = freework->fw_indir;
6323 bp = indirdep->ir_bp;
6324 /* See if the block was discarded. */
6327 /* Inline part of getdirtybuf(). We dont want bremfree. */
6328 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0)
6330 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
6331 LOCK_PTR(ump)) == 0)
6335 freework->fw_state |= DEPCOMPLETE;
6336 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next);
6338 * Zero the pointers in the saved copy.
6340 if (indirdep->ir_state & UFS1FMT)
6341 start = sizeof(ufs1_daddr_t);
6343 start = sizeof(ufs2_daddr_t);
6344 start *= freework->fw_start;
6345 count = indirdep->ir_savebp->b_bcount - start;
6346 start += (uintptr_t)indirdep->ir_savebp->b_data;
6347 bzero((char *)start, count);
6349 * We need to start the next truncation in the list if it has not
6352 fwn = TAILQ_FIRST(&indirdep->ir_trunc);
6354 if (fwn->fw_freeblks == indirdep->ir_freeblks)
6355 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next);
6356 if ((fwn->fw_state & ONWORKLIST) == 0)
6357 freework_enqueue(fwn);
6360 * If bp is NULL the block was fully truncated, restore
6361 * the saved block list otherwise free it if it is no
6364 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
6366 bcopy(indirdep->ir_saveddata,
6367 indirdep->ir_savebp->b_data,
6368 indirdep->ir_savebp->b_bcount);
6369 free(indirdep->ir_saveddata, M_INDIRDEP);
6370 indirdep->ir_saveddata = NULL;
6373 * When bp is NULL there is a full truncation pending. We
6374 * must wait for this full truncation to be journaled before
6375 * we can release this freework because the disk pointers will
6376 * never be written as zero.
6379 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd))
6380 handle_written_freework(freework);
6382 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd,
6383 &freework->fw_list);
6385 /* Complete when the real copy is written. */
6386 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list);
6392 * Calculate the number of blocks we are going to release where datablocks
6393 * is the current total and length is the new file size.
6396 blkcount(fs, datablocks, length)
6398 ufs2_daddr_t datablocks;
6401 off_t totblks, numblks;
6404 numblks = howmany(length, fs->fs_bsize);
6405 if (numblks <= NDADDR) {
6406 totblks = howmany(length, fs->fs_fsize);
6409 totblks = blkstofrags(fs, numblks);
6412 * Count all single, then double, then triple indirects required.
6413 * Subtracting one indirects worth of blocks for each pass
6414 * acknowledges one of each pointed to by the inode.
6417 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs)));
6418 numblks -= NINDIR(fs);
6421 numblks = howmany(numblks, NINDIR(fs));
6424 totblks = fsbtodb(fs, totblks);
6426 * Handle sparse files. We can't reclaim more blocks than the inode
6427 * references. We will correct it later in handle_complete_freeblks()
6428 * when we know the real count.
6430 if (totblks > datablocks)
6432 return (datablocks - totblks);
6436 * Handle freeblocks for journaled softupdate filesystems.
6438 * Contrary to normal softupdates, we must preserve the block pointers in
6439 * indirects until their subordinates are free. This is to avoid journaling
6440 * every block that is freed which may consume more space than the journal
6441 * itself. The recovery program will see the free block journals at the
6442 * base of the truncated area and traverse them to reclaim space. The
6443 * pointers in the inode may be cleared immediately after the journal
6444 * records are written because each direct and indirect pointer in the
6445 * inode is recorded in a journal. This permits full truncation to proceed
6446 * asynchronously. The write order is journal -> inode -> cgs -> indirects.
6448 * The algorithm is as follows:
6449 * 1) Traverse the in-memory state and create journal entries to release
6450 * the relevant blocks and full indirect trees.
6451 * 2) Traverse the indirect block chain adding partial truncation freework
6452 * records to indirects in the path to lastlbn. The freework will
6453 * prevent new allocation dependencies from being satisfied in this
6454 * indirect until the truncation completes.
6455 * 3) Read and lock the inode block, performing an update with the new size
6456 * and pointers. This prevents truncated data from becoming valid on
6457 * disk through step 4.
6458 * 4) Reap unsatisfied dependencies that are beyond the truncated area,
6459 * eliminate journal work for those records that do not require it.
6460 * 5) Schedule the journal records to be written followed by the inode block.
6461 * 6) Allocate any necessary frags for the end of file.
6462 * 7) Zero any partially truncated blocks.
6464 * From this truncation proceeds asynchronously using the freework and
6465 * indir_trunc machinery. The file will not be extended again into a
6466 * partially truncated indirect block until all work is completed but
6467 * the normal dependency mechanism ensures that it is rolled back/forward
6468 * as appropriate. Further truncation may occur without delay and is
6469 * serialized in indir_trunc().
6472 softdep_journal_freeblocks(ip, cred, length, flags)
6473 struct inode *ip; /* The inode whose length is to be reduced */
6475 off_t length; /* The new length for the file */
6476 int flags; /* IO_EXT and/or IO_NORMAL */
6478 struct freeblks *freeblks, *fbn;
6479 struct worklist *wk, *wkn;
6480 struct inodedep *inodedep;
6481 struct jblkdep *jblkdep;
6482 struct allocdirect *adp, *adpn;
6483 struct ufsmount *ump;
6488 ufs2_daddr_t extblocks, datablocks;
6489 ufs_lbn_t tmpval, lbn, lastlbn;
6490 int frags, lastoff, iboff, allocblock, needj, error, i;
6495 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6496 ("softdep_journal_freeblocks called on non-softdep filesystem"));
6504 freeblks = newfreeblks(mp, ip);
6507 * If we're truncating a removed file that will never be written
6508 * we don't need to journal the block frees. The canceled journals
6509 * for the allocations will suffice.
6511 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6512 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
6515 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d",
6516 ip->i_number, length, needj);
6519 * Calculate the lbn that we are truncating to. This results in -1
6520 * if we're truncating the 0 bytes. So it is the last lbn we want
6521 * to keep, not the first lbn we want to truncate.
6523 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1;
6524 lastoff = blkoff(fs, length);
6526 * Compute frags we are keeping in lastlbn. 0 means all.
6528 if (lastlbn >= 0 && lastlbn < NDADDR) {
6529 frags = fragroundup(fs, lastoff);
6530 /* adp offset of last valid allocdirect. */
6532 } else if (lastlbn > 0)
6534 if (fs->fs_magic == FS_UFS2_MAGIC)
6535 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6537 * Handle normal data blocks and indirects. This section saves
6538 * values used after the inode update to complete frag and indirect
6541 if ((flags & IO_NORMAL) != 0) {
6543 * Handle truncation of whole direct and indirect blocks.
6545 for (i = iboff + 1; i < NDADDR; i++)
6546 setup_freedirect(freeblks, ip, i, needj);
6547 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6548 i++, lbn += tmpval, tmpval *= NINDIR(fs)) {
6549 /* Release a whole indirect tree. */
6550 if (lbn > lastlbn) {
6551 setup_freeindir(freeblks, ip, i, -lbn -i,
6557 * Traverse partially truncated indirect tree.
6559 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn)
6560 setup_trunc_indir(freeblks, ip, -lbn - i,
6561 lastlbn, DIP(ip, i_ib[i]));
6564 * Handle partial truncation to a frag boundary.
6570 oldfrags = blksize(fs, ip, lastlbn);
6571 blkno = DIP(ip, i_db[lastlbn]);
6572 if (blkno && oldfrags != frags) {
6574 oldfrags = numfrags(ip->i_fs, oldfrags);
6575 blkno += numfrags(ip->i_fs, frags);
6576 newfreework(ump, freeblks, NULL, lastlbn,
6577 blkno, oldfrags, 0, needj);
6579 adjust_newfreework(freeblks,
6580 numfrags(ip->i_fs, frags));
6581 } else if (blkno == 0)
6585 * Add a journal record for partial truncate if we are
6586 * handling indirect blocks. Non-indirects need no extra
6589 if (length != 0 && lastlbn >= NDADDR) {
6590 ip->i_flag |= IN_TRUNCATED;
6591 newjtrunc(freeblks, length, 0);
6593 ip->i_size = length;
6594 DIP_SET(ip, i_size, ip->i_size);
6595 datablocks = DIP(ip, i_blocks) - extblocks;
6597 datablocks = blkcount(ip->i_fs, datablocks, length);
6598 freeblks->fb_len = length;
6600 if ((flags & IO_EXT) != 0) {
6601 for (i = 0; i < NXADDR; i++)
6602 setup_freeext(freeblks, ip, i, needj);
6603 ip->i_din2->di_extsize = 0;
6604 datablocks += extblocks;
6607 /* Reference the quotas in case the block count is wrong in the end. */
6608 quotaref(vp, freeblks->fb_quota);
6609 (void) chkdq(ip, -datablocks, NOCRED, 0);
6611 freeblks->fb_chkcnt = -datablocks;
6613 fs->fs_pendingblocks += datablocks;
6615 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6617 * Handle truncation of incomplete alloc direct dependencies. We
6618 * hold the inode block locked to prevent incomplete dependencies
6619 * from reaching the disk while we are eliminating those that
6620 * have been truncated. This is a partially inlined ffs_update().
6623 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
6624 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6625 (int)fs->fs_bsize, cred, &bp);
6628 softdep_error("softdep_journal_freeblocks", error);
6631 if (bp->b_bufsize == fs->fs_bsize)
6632 bp->b_flags |= B_CLUSTEROK;
6633 softdep_update_inodeblock(ip, bp, 0);
6634 if (ump->um_fstype == UFS1)
6635 *((struct ufs1_dinode *)bp->b_data +
6636 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
6638 *((struct ufs2_dinode *)bp->b_data +
6639 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
6641 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6642 if ((inodedep->id_state & IOSTARTED) != 0)
6643 panic("softdep_setup_freeblocks: inode busy");
6645 * Add the freeblks structure to the list of operations that
6646 * must await the zero'ed inode being written to disk. If we
6647 * still have a bitmap dependency (needj), then the inode
6648 * has never been written to disk, so we can process the
6649 * freeblks below once we have deleted the dependencies.
6652 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6654 freeblks->fb_state |= COMPLETE;
6655 if ((flags & IO_NORMAL) != 0) {
6656 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) {
6657 if (adp->ad_offset > iboff)
6658 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6661 * Truncate the allocdirect. We could eliminate
6662 * or modify journal records as well.
6664 else if (adp->ad_offset == iboff && frags)
6665 adp->ad_newsize = frags;
6668 if ((flags & IO_EXT) != 0)
6669 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6670 cancel_allocdirect(&inodedep->id_extupdt, adp,
6673 * Scan the bufwait list for newblock dependencies that will never
6676 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) {
6677 if (wk->wk_type != D_ALLOCDIRECT)
6679 adp = WK_ALLOCDIRECT(wk);
6680 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) ||
6681 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) {
6682 cancel_jfreeblk(freeblks, adp->ad_newblkno);
6683 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork);
6684 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
6690 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps)
6691 add_to_journal(&jblkdep->jb_list);
6695 * Truncate dependency structures beyond length.
6697 trunc_dependencies(ip, freeblks, lastlbn, frags, flags);
6699 * This is only set when we need to allocate a fragment because
6700 * none existed at the end of a frag-sized file. It handles only
6701 * allocating a new, zero filled block.
6704 ip->i_size = length - lastoff;
6705 DIP_SET(ip, i_size, ip->i_size);
6706 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp);
6708 softdep_error("softdep_journal_freeblks", error);
6711 ip->i_size = length;
6712 DIP_SET(ip, i_size, length);
6713 ip->i_flag |= IN_CHANGE | IN_UPDATE;
6714 allocbuf(bp, frags);
6717 } else if (lastoff != 0 && vp->v_type != VDIR) {
6721 * Zero the end of a truncated frag or block.
6723 size = sblksize(fs, length, lastlbn);
6724 error = bread(vp, lastlbn, size, cred, &bp);
6726 softdep_error("softdep_journal_freeblks", error);
6729 bzero((char *)bp->b_data + lastoff, size - lastoff);
6734 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6735 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
6736 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
6738 * We zero earlier truncations so they don't erroneously
6741 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0)
6742 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next)
6744 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE &&
6745 LIST_EMPTY(&freeblks->fb_jblkdephd))
6746 freeblks->fb_state |= INPROGRESS;
6751 handle_workitem_freeblocks(freeblks, 0);
6752 trunc_pages(ip, length, extblocks, flags);
6757 * Flush a JOP_SYNC to the journal.
6760 softdep_journal_fsync(ip)
6763 struct jfsync *jfsync;
6765 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
6766 ("softdep_journal_fsync called on non-softdep filesystem"));
6767 if ((ip->i_flag & IN_TRUNCATED) == 0)
6769 ip->i_flag &= ~IN_TRUNCATED;
6770 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO);
6771 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump));
6772 jfsync->jfs_size = ip->i_size;
6773 jfsync->jfs_ino = ip->i_number;
6774 ACQUIRE_LOCK(ip->i_ump);
6775 add_to_journal(&jfsync->jfs_list);
6776 jwait(&jfsync->jfs_list, MNT_WAIT);
6777 FREE_LOCK(ip->i_ump);
6781 * Block de-allocation dependencies.
6783 * When blocks are de-allocated, the on-disk pointers must be nullified before
6784 * the blocks are made available for use by other files. (The true
6785 * requirement is that old pointers must be nullified before new on-disk
6786 * pointers are set. We chose this slightly more stringent requirement to
6787 * reduce complexity.) Our implementation handles this dependency by updating
6788 * the inode (or indirect block) appropriately but delaying the actual block
6789 * de-allocation (i.e., freemap and free space count manipulation) until
6790 * after the updated versions reach stable storage. After the disk is
6791 * updated, the blocks can be safely de-allocated whenever it is convenient.
6792 * This implementation handles only the common case of reducing a file's
6793 * length to zero. Other cases are handled by the conventional synchronous
6796 * The ffs implementation with which we worked double-checks
6797 * the state of the block pointers and file size as it reduces
6798 * a file's length. Some of this code is replicated here in our
6799 * soft updates implementation. The freeblks->fb_chkcnt field is
6800 * used to transfer a part of this information to the procedure
6801 * that eventually de-allocates the blocks.
6803 * This routine should be called from the routine that shortens
6804 * a file's length, before the inode's size or block pointers
6805 * are modified. It will save the block pointer information for
6806 * later release and zero the inode so that the calling routine
6810 softdep_setup_freeblocks(ip, length, flags)
6811 struct inode *ip; /* The inode whose length is to be reduced */
6812 off_t length; /* The new length for the file */
6813 int flags; /* IO_EXT and/or IO_NORMAL */
6815 struct ufs1_dinode *dp1;
6816 struct ufs2_dinode *dp2;
6817 struct freeblks *freeblks;
6818 struct inodedep *inodedep;
6819 struct allocdirect *adp;
6820 struct ufsmount *ump;
6823 ufs2_daddr_t extblocks, datablocks;
6825 int i, delay, error;
6831 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6832 ("softdep_setup_freeblocks called on non-softdep filesystem"));
6833 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld",
6834 ip->i_number, length);
6835 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length"));
6837 freeblks = newfreeblks(mp, ip);
6840 if (fs->fs_magic == FS_UFS2_MAGIC)
6841 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6842 if ((flags & IO_NORMAL) != 0) {
6843 for (i = 0; i < NDADDR; i++)
6844 setup_freedirect(freeblks, ip, i, 0);
6845 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6846 i++, lbn += tmpval, tmpval *= NINDIR(fs))
6847 setup_freeindir(freeblks, ip, i, -lbn -i, 0);
6849 DIP_SET(ip, i_size, 0);
6850 datablocks = DIP(ip, i_blocks) - extblocks;
6852 if ((flags & IO_EXT) != 0) {
6853 for (i = 0; i < NXADDR; i++)
6854 setup_freeext(freeblks, ip, i, 0);
6855 ip->i_din2->di_extsize = 0;
6856 datablocks += extblocks;
6859 /* Reference the quotas in case the block count is wrong in the end. */
6860 quotaref(ITOV(ip), freeblks->fb_quota);
6861 (void) chkdq(ip, -datablocks, NOCRED, 0);
6863 freeblks->fb_chkcnt = -datablocks;
6865 fs->fs_pendingblocks += datablocks;
6867 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6869 * Push the zero'ed inode to to its disk buffer so that we are free
6870 * to delete its dependencies below. Once the dependencies are gone
6871 * the buffer can be safely released.
6873 if ((error = bread(ip->i_devvp,
6874 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6875 (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
6877 softdep_error("softdep_setup_freeblocks", error);
6879 if (ump->um_fstype == UFS1) {
6880 dp1 = ((struct ufs1_dinode *)bp->b_data +
6881 ino_to_fsbo(fs, ip->i_number));
6882 ip->i_din1->di_freelink = dp1->di_freelink;
6885 dp2 = ((struct ufs2_dinode *)bp->b_data +
6886 ino_to_fsbo(fs, ip->i_number));
6887 ip->i_din2->di_freelink = dp2->di_freelink;
6891 * Find and eliminate any inode dependencies.
6894 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6895 if ((inodedep->id_state & IOSTARTED) != 0)
6896 panic("softdep_setup_freeblocks: inode busy");
6898 * Add the freeblks structure to the list of operations that
6899 * must await the zero'ed inode being written to disk. If we
6900 * still have a bitmap dependency (delay == 0), then the inode
6901 * has never been written to disk, so we can process the
6902 * freeblks below once we have deleted the dependencies.
6904 delay = (inodedep->id_state & DEPCOMPLETE);
6906 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6908 freeblks->fb_state |= COMPLETE;
6910 * Because the file length has been truncated to zero, any
6911 * pending block allocation dependency structures associated
6912 * with this inode are obsolete and can simply be de-allocated.
6913 * We must first merge the two dependency lists to get rid of
6914 * any duplicate freefrag structures, then purge the merged list.
6915 * If we still have a bitmap dependency, then the inode has never
6916 * been written to disk, so we can free any fragments without delay.
6918 if (flags & IO_NORMAL) {
6919 merge_inode_lists(&inodedep->id_newinoupdt,
6920 &inodedep->id_inoupdt);
6921 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
6922 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6925 if (flags & IO_EXT) {
6926 merge_inode_lists(&inodedep->id_newextupdt,
6927 &inodedep->id_extupdt);
6928 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6929 cancel_allocdirect(&inodedep->id_extupdt, adp,
6934 trunc_dependencies(ip, freeblks, -1, 0, flags);
6936 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
6937 (void) free_inodedep(inodedep);
6938 freeblks->fb_state |= DEPCOMPLETE;
6940 * If the inode with zeroed block pointers is now on disk
6941 * we can start freeing blocks.
6943 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
6944 freeblks->fb_state |= INPROGRESS;
6949 handle_workitem_freeblocks(freeblks, 0);
6950 trunc_pages(ip, length, extblocks, flags);
6954 * Eliminate pages from the page cache that back parts of this inode and
6955 * adjust the vnode pager's idea of our size. This prevents stale data
6956 * from hanging around in the page cache.
6959 trunc_pages(ip, length, extblocks, flags)
6962 ufs2_daddr_t extblocks;
6972 extend = OFF_TO_IDX(lblktosize(fs, -extblocks));
6973 if ((flags & IO_EXT) != 0)
6974 vn_pages_remove(vp, extend, 0);
6975 if ((flags & IO_NORMAL) == 0)
6977 BO_LOCK(&vp->v_bufobj);
6979 BO_UNLOCK(&vp->v_bufobj);
6981 * The vnode pager eliminates file pages we eliminate indirects
6984 vnode_pager_setsize(vp, length);
6986 * Calculate the end based on the last indirect we want to keep. If
6987 * the block extends into indirects we can just use the negative of
6988 * its lbn. Doubles and triples exist at lower numbers so we must
6989 * be careful not to remove those, if they exist. double and triple
6990 * indirect lbns do not overlap with others so it is not important
6991 * to verify how many levels are required.
6993 lbn = lblkno(fs, length);
6994 if (lbn >= NDADDR) {
6995 /* Calculate the virtual lbn of the triple indirect. */
6996 lbn = -lbn - (NIADDR - 1);
6997 end = OFF_TO_IDX(lblktosize(fs, lbn));
7000 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end);
7004 * See if the buf bp is in the range eliminated by truncation.
7007 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags)
7017 /* Only match ext/normal blocks as appropriate. */
7018 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
7019 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0))
7021 /* ALTDATA is always a full truncation. */
7022 if ((bp->b_xflags & BX_ALTDATA) != 0)
7024 /* -1 is full truncation. */
7028 * If this is a partial truncate we only want those
7029 * blocks and indirect blocks that cover the range
7034 lbn = -(lbn + lbn_level(lbn));
7037 /* Here we only truncate lblkno if it's partial. */
7038 if (lbn == lastlbn) {
7047 * Eliminate any dependencies that exist in memory beyond lblkno:off
7050 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags)
7052 struct freeblks *freeblks;
7063 * We must wait for any I/O in progress to finish so that
7064 * all potential buffers on the dirty list will be visible.
7065 * Once they are all there, walk the list and get rid of
7072 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
7073 bp->b_vflags &= ~BV_SCANNED;
7075 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
7076 if (bp->b_vflags & BV_SCANNED)
7078 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7079 bp->b_vflags |= BV_SCANNED;
7082 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer"));
7083 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL)
7086 if (deallocate_dependencies(bp, freeblks, blkoff))
7094 * Now do the work of vtruncbuf while also matching indirect blocks.
7096 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs)
7097 bp->b_vflags &= ~BV_SCANNED;
7099 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) {
7100 if (bp->b_vflags & BV_SCANNED)
7102 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7103 bp->b_vflags |= BV_SCANNED;
7107 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
7108 BO_LOCKPTR(bo)) == ENOLCK) {
7112 bp->b_vflags |= BV_SCANNED;
7115 allocbuf(bp, blkoff);
7118 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF;
7129 cancel_pagedep(pagedep, freeblks, blkoff)
7130 struct pagedep *pagedep;
7131 struct freeblks *freeblks;
7134 struct jremref *jremref;
7135 struct jmvref *jmvref;
7136 struct dirrem *dirrem, *tmp;
7140 * Copy any directory remove dependencies to the list
7141 * to be processed after the freeblks proceeds. If
7142 * directory entry never made it to disk they
7143 * can be dumped directly onto the work list.
7145 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) {
7146 /* Skip this directory removal if it is intended to remain. */
7147 if (dirrem->dm_offset < blkoff)
7150 * If there are any dirrems we wait for the journal write
7151 * to complete and then restart the buf scan as the lock
7154 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
7155 jwait(&jremref->jr_list, MNT_WAIT);
7158 LIST_REMOVE(dirrem, dm_next);
7159 dirrem->dm_dirinum = pagedep->pd_ino;
7160 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list);
7162 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
7163 jwait(&jmvref->jm_list, MNT_WAIT);
7167 * When we're partially truncating a pagedep we just want to flush
7168 * journal entries and return. There can not be any adds in the
7169 * truncated portion of the directory and newblk must remain if
7170 * part of the block remains.
7175 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
7176 if (dap->da_offset > blkoff)
7177 panic("cancel_pagedep: diradd %p off %d > %d",
7178 dap, dap->da_offset, blkoff);
7179 for (i = 0; i < DAHASHSZ; i++)
7180 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist)
7181 if (dap->da_offset > blkoff)
7182 panic("cancel_pagedep: diradd %p off %d > %d",
7183 dap, dap->da_offset, blkoff);
7187 * There should be no directory add dependencies present
7188 * as the directory could not be truncated until all
7189 * children were removed.
7191 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
7192 ("deallocate_dependencies: pendinghd != NULL"));
7193 for (i = 0; i < DAHASHSZ; i++)
7194 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
7195 ("deallocate_dependencies: diraddhd != NULL"));
7196 if ((pagedep->pd_state & NEWBLOCK) != 0)
7197 free_newdirblk(pagedep->pd_newdirblk);
7198 if (free_pagedep(pagedep) == 0)
7199 panic("Failed to free pagedep %p", pagedep);
7204 * Reclaim any dependency structures from a buffer that is about to
7205 * be reallocated to a new vnode. The buffer must be locked, thus,
7206 * no I/O completion operations can occur while we are manipulating
7207 * its associated dependencies. The mutex is held so that other I/O's
7208 * associated with related dependencies do not occur.
7211 deallocate_dependencies(bp, freeblks, off)
7213 struct freeblks *freeblks;
7216 struct indirdep *indirdep;
7217 struct pagedep *pagedep;
7218 struct worklist *wk, *wkn;
7219 struct ufsmount *ump;
7221 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
7223 ump = VFSTOUFS(wk->wk_mp);
7225 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) {
7226 switch (wk->wk_type) {
7228 indirdep = WK_INDIRDEP(wk);
7229 if (bp->b_lblkno >= 0 ||
7230 bp->b_blkno != indirdep->ir_savebp->b_lblkno)
7231 panic("deallocate_dependencies: not indir");
7232 cancel_indirdep(indirdep, bp, freeblks);
7236 pagedep = WK_PAGEDEP(wk);
7237 if (cancel_pagedep(pagedep, freeblks, off)) {
7245 * Simply remove the allocindir, we'll find it via
7246 * the indirdep where we can clear pointers if
7249 WORKLIST_REMOVE(wk);
7254 * A truncation is waiting for the zero'd pointers
7255 * to be written. It can be freed when the freeblks
7258 WORKLIST_REMOVE(wk);
7259 wk->wk_state |= ONDEPLIST;
7260 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
7268 panic("deallocate_dependencies: Unexpected type %s",
7269 TYPENAME(wk->wk_type));
7276 * Don't throw away this buf, we were partially truncating and
7277 * some deps may always remain.
7281 bp->b_vflags |= BV_SCANNED;
7284 bp->b_flags |= B_INVAL | B_NOCACHE;
7290 * An allocdirect is being canceled due to a truncate. We must make sure
7291 * the journal entry is released in concert with the blkfree that releases
7292 * the storage. Completed journal entries must not be released until the
7293 * space is no longer pointed to by the inode or in the bitmap.
7296 cancel_allocdirect(adphead, adp, freeblks)
7297 struct allocdirectlst *adphead;
7298 struct allocdirect *adp;
7299 struct freeblks *freeblks;
7301 struct freework *freework;
7302 struct newblk *newblk;
7303 struct worklist *wk;
7305 TAILQ_REMOVE(adphead, adp, ad_next);
7306 newblk = (struct newblk *)adp;
7309 * Find the correct freework structure.
7311 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
7312 if (wk->wk_type != D_FREEWORK)
7314 freework = WK_FREEWORK(wk);
7315 if (freework->fw_blkno == newblk->nb_newblkno)
7318 if (freework == NULL)
7319 panic("cancel_allocdirect: Freework not found");
7321 * If a newblk exists at all we still have the journal entry that
7322 * initiated the allocation so we do not need to journal the free.
7324 cancel_jfreeblk(freeblks, freework->fw_blkno);
7326 * If the journal hasn't been written the jnewblk must be passed
7327 * to the call to ffs_blkfree that reclaims the space. We accomplish
7328 * this by linking the journal dependency into the freework to be
7329 * freed when freework_freeblock() is called. If the journal has
7330 * been written we can simply reclaim the journal space when the
7331 * freeblks work is complete.
7333 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list,
7334 &freeblks->fb_jwork);
7335 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
7340 * Cancel a new block allocation. May be an indirect or direct block. We
7341 * remove it from various lists and return any journal record that needs to
7342 * be resolved by the caller.
7344 * A special consideration is made for indirects which were never pointed
7345 * at on disk and will never be found once this block is released.
7347 static struct jnewblk *
7348 cancel_newblk(newblk, wk, wkhd)
7349 struct newblk *newblk;
7350 struct worklist *wk;
7351 struct workhead *wkhd;
7353 struct jnewblk *jnewblk;
7355 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno);
7357 newblk->nb_state |= GOINGAWAY;
7359 * Previously we traversed the completedhd on each indirdep
7360 * attached to this newblk to cancel them and gather journal
7361 * work. Since we need only the oldest journal segment and
7362 * the lowest point on the tree will always have the oldest
7363 * journal segment we are free to release the segments
7364 * of any subordinates and may leave the indirdep list to
7365 * indirdep_complete() when this newblk is freed.
7367 if (newblk->nb_state & ONDEPLIST) {
7368 newblk->nb_state &= ~ONDEPLIST;
7369 LIST_REMOVE(newblk, nb_deps);
7371 if (newblk->nb_state & ONWORKLIST)
7372 WORKLIST_REMOVE(&newblk->nb_list);
7374 * If the journal entry hasn't been written we save a pointer to
7375 * the dependency that frees it until it is written or the
7376 * superseding operation completes.
7378 jnewblk = newblk->nb_jnewblk;
7379 if (jnewblk != NULL && wk != NULL) {
7380 newblk->nb_jnewblk = NULL;
7381 jnewblk->jn_dep = wk;
7383 if (!LIST_EMPTY(&newblk->nb_jwork))
7384 jwork_move(wkhd, &newblk->nb_jwork);
7386 * When truncating we must free the newdirblk early to remove
7387 * the pagedep from the hash before returning.
7389 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7390 free_newdirblk(WK_NEWDIRBLK(wk));
7391 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7392 panic("cancel_newblk: extra newdirblk");
7398 * Schedule the freefrag associated with a newblk to be released once
7399 * the pointers are written and the previous block is no longer needed.
7402 newblk_freefrag(newblk)
7403 struct newblk *newblk;
7405 struct freefrag *freefrag;
7407 if (newblk->nb_freefrag == NULL)
7409 freefrag = newblk->nb_freefrag;
7410 newblk->nb_freefrag = NULL;
7411 freefrag->ff_state |= COMPLETE;
7412 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
7413 add_to_worklist(&freefrag->ff_list, 0);
7417 * Free a newblk. Generate a new freefrag work request if appropriate.
7418 * This must be called after the inode pointer and any direct block pointers
7419 * are valid or fully removed via truncate or frag extension.
7423 struct newblk *newblk;
7425 struct indirdep *indirdep;
7426 struct worklist *wk;
7428 KASSERT(newblk->nb_jnewblk == NULL,
7429 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk));
7430 KASSERT(newblk->nb_list.wk_type != D_NEWBLK,
7431 ("free_newblk: unclaimed newblk"));
7432 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp));
7433 newblk_freefrag(newblk);
7434 if (newblk->nb_state & ONDEPLIST)
7435 LIST_REMOVE(newblk, nb_deps);
7436 if (newblk->nb_state & ONWORKLIST)
7437 WORKLIST_REMOVE(&newblk->nb_list);
7438 LIST_REMOVE(newblk, nb_hash);
7439 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7440 free_newdirblk(WK_NEWDIRBLK(wk));
7441 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7442 panic("free_newblk: extra newdirblk");
7443 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL)
7444 indirdep_complete(indirdep);
7445 handle_jwork(&newblk->nb_jwork);
7446 WORKITEM_FREE(newblk, D_NEWBLK);
7450 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
7451 * This routine must be called with splbio interrupts blocked.
7454 free_newdirblk(newdirblk)
7455 struct newdirblk *newdirblk;
7457 struct pagedep *pagedep;
7459 struct worklist *wk;
7461 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp));
7462 WORKLIST_REMOVE(&newdirblk->db_list);
7464 * If the pagedep is still linked onto the directory buffer
7465 * dependency chain, then some of the entries on the
7466 * pd_pendinghd list may not be committed to disk yet. In
7467 * this case, we will simply clear the NEWBLOCK flag and
7468 * let the pd_pendinghd list be processed when the pagedep
7469 * is next written. If the pagedep is no longer on the buffer
7470 * dependency chain, then all the entries on the pd_pending
7471 * list are committed to disk and we can free them here.
7473 pagedep = newdirblk->db_pagedep;
7474 pagedep->pd_state &= ~NEWBLOCK;
7475 if ((pagedep->pd_state & ONWORKLIST) == 0) {
7476 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
7477 free_diradd(dap, NULL);
7479 * If no dependencies remain, the pagedep will be freed.
7481 free_pagedep(pagedep);
7483 /* Should only ever be one item in the list. */
7484 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
7485 WORKLIST_REMOVE(wk);
7486 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
7488 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
7492 * Prepare an inode to be freed. The actual free operation is not
7493 * done until the zero'ed inode has been written to disk.
7496 softdep_freefile(pvp, ino, mode)
7501 struct inode *ip = VTOI(pvp);
7502 struct inodedep *inodedep;
7503 struct freefile *freefile;
7504 struct freeblks *freeblks;
7505 struct ufsmount *ump;
7508 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
7509 ("softdep_freefile called on non-softdep filesystem"));
7511 * This sets up the inode de-allocation dependency.
7513 freefile = malloc(sizeof(struct freefile),
7514 M_FREEFILE, M_SOFTDEP_FLAGS);
7515 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
7516 freefile->fx_mode = mode;
7517 freefile->fx_oldinum = ino;
7518 freefile->fx_devvp = ip->i_devvp;
7519 LIST_INIT(&freefile->fx_jwork);
7521 ip->i_fs->fs_pendinginodes += 1;
7525 * If the inodedep does not exist, then the zero'ed inode has
7526 * been written to disk. If the allocated inode has never been
7527 * written to disk, then the on-disk inode is zero'ed. In either
7528 * case we can free the file immediately. If the journal was
7529 * canceled before being written the inode will never make it to
7530 * disk and we must send the canceled journal entrys to
7531 * ffs_freefile() to be cleared in conjunction with the bitmap.
7532 * Any blocks waiting on the inode to write can be safely freed
7533 * here as it will never been written.
7536 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7539 * Clear out freeblks that no longer need to reference
7543 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) {
7544 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks,
7546 freeblks->fb_state &= ~ONDEPLIST;
7549 * Remove this inode from the unlinked list.
7551 if (inodedep->id_state & UNLINKED) {
7553 * Save the journal work to be freed with the bitmap
7554 * before we clear UNLINKED. Otherwise it can be lost
7555 * if the inode block is written.
7557 handle_bufwait(inodedep, &freefile->fx_jwork);
7558 clear_unlinked_inodedep(inodedep);
7560 * Re-acquire inodedep as we've dropped the
7561 * per-filesystem lock in clear_unlinked_inodedep().
7563 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7566 if (inodedep == NULL || check_inode_unwritten(inodedep)) {
7568 handle_workitem_freefile(freefile);
7571 if ((inodedep->id_state & DEPCOMPLETE) == 0)
7572 inodedep->id_state |= GOINGAWAY;
7573 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
7575 if (ip->i_number == ino)
7576 ip->i_flag |= IN_MODIFIED;
7580 * Check to see if an inode has never been written to disk. If
7581 * so free the inodedep and return success, otherwise return failure.
7582 * This routine must be called with splbio interrupts blocked.
7584 * If we still have a bitmap dependency, then the inode has never
7585 * been written to disk. Drop the dependency as it is no longer
7586 * necessary since the inode is being deallocated. We set the
7587 * ALLCOMPLETE flags since the bitmap now properly shows that the
7588 * inode is not allocated. Even if the inode is actively being
7589 * written, it has been rolled back to its zero'ed state, so we
7590 * are ensured that a zero inode is what is on the disk. For short
7591 * lived files, this change will usually result in removing all the
7592 * dependencies from the inode so that it can be freed immediately.
7595 check_inode_unwritten(inodedep)
7596 struct inodedep *inodedep;
7599 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7601 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
7602 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7603 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7604 !LIST_EMPTY(&inodedep->id_bufwait) ||
7605 !LIST_EMPTY(&inodedep->id_inowait) ||
7606 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7607 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7608 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7609 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7610 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7611 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7612 inodedep->id_mkdiradd != NULL ||
7613 inodedep->id_nlinkdelta != 0)
7616 * Another process might be in initiate_write_inodeblock_ufs[12]
7617 * trying to allocate memory without holding "Softdep Lock".
7619 if ((inodedep->id_state & IOSTARTED) != 0 &&
7620 inodedep->id_savedino1 == NULL)
7623 if (inodedep->id_state & ONDEPLIST)
7624 LIST_REMOVE(inodedep, id_deps);
7625 inodedep->id_state &= ~ONDEPLIST;
7626 inodedep->id_state |= ALLCOMPLETE;
7627 inodedep->id_bmsafemap = NULL;
7628 if (inodedep->id_state & ONWORKLIST)
7629 WORKLIST_REMOVE(&inodedep->id_list);
7630 if (inodedep->id_savedino1 != NULL) {
7631 free(inodedep->id_savedino1, M_SAVEDINO);
7632 inodedep->id_savedino1 = NULL;
7634 if (free_inodedep(inodedep) == 0)
7635 panic("check_inode_unwritten: busy inode");
7640 check_inodedep_free(inodedep)
7641 struct inodedep *inodedep;
7644 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7645 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
7646 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7647 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7648 !LIST_EMPTY(&inodedep->id_bufwait) ||
7649 !LIST_EMPTY(&inodedep->id_inowait) ||
7650 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7651 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7652 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7653 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7654 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7655 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7656 inodedep->id_mkdiradd != NULL ||
7657 inodedep->id_nlinkdelta != 0 ||
7658 inodedep->id_savedino1 != NULL)
7664 * Try to free an inodedep structure. Return 1 if it could be freed.
7667 free_inodedep(inodedep)
7668 struct inodedep *inodedep;
7671 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7672 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
7673 !check_inodedep_free(inodedep))
7675 if (inodedep->id_state & ONDEPLIST)
7676 LIST_REMOVE(inodedep, id_deps);
7677 LIST_REMOVE(inodedep, id_hash);
7678 WORKITEM_FREE(inodedep, D_INODEDEP);
7683 * Free the block referenced by a freework structure. The parent freeblks
7684 * structure is released and completed when the final cg bitmap reaches
7685 * the disk. This routine may be freeing a jnewblk which never made it to
7686 * disk in which case we do not have to wait as the operation is undone
7687 * in memory immediately.
7690 freework_freeblock(freework)
7691 struct freework *freework;
7693 struct freeblks *freeblks;
7694 struct jnewblk *jnewblk;
7695 struct ufsmount *ump;
7696 struct workhead wkhd;
7701 ump = VFSTOUFS(freework->fw_list.wk_mp);
7704 * Handle partial truncate separately.
7706 if (freework->fw_indir) {
7707 complete_trunc_indir(freework);
7710 freeblks = freework->fw_freeblks;
7712 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0;
7713 bsize = lfragtosize(fs, freework->fw_frags);
7716 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives
7717 * on the indirblk hashtable and prevents premature freeing.
7719 freework->fw_state |= DEPCOMPLETE;
7721 * SUJ needs to wait for the segment referencing freed indirect
7722 * blocks to expire so that we know the checker will not confuse
7723 * a re-allocated indirect block with its old contents.
7725 if (needj && freework->fw_lbn <= -NDADDR)
7726 indirblk_insert(freework);
7728 * If we are canceling an existing jnewblk pass it to the free
7729 * routine, otherwise pass the freeblk which will ultimately
7730 * release the freeblks. If we're not journaling, we can just
7731 * free the freeblks immediately.
7733 jnewblk = freework->fw_jnewblk;
7734 if (jnewblk != NULL) {
7735 cancel_jnewblk(jnewblk, &wkhd);
7738 freework->fw_state |= DELAYEDFREE;
7739 freeblks->fb_cgwait++;
7740 WORKLIST_INSERT(&wkhd, &freework->fw_list);
7743 freeblks_free(ump, freeblks, btodb(bsize));
7745 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld",
7746 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize);
7747 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize,
7748 freeblks->fb_inum, freeblks->fb_vtype, &wkhd);
7751 * The jnewblk will be discarded and the bits in the map never
7752 * made it to disk. We can immediately free the freeblk.
7755 handle_written_freework(freework);
7759 * We enqueue freework items that need processing back on the freeblks and
7760 * add the freeblks to the worklist. This makes it easier to find all work
7761 * required to flush a truncation in process_truncates().
7764 freework_enqueue(freework)
7765 struct freework *freework;
7767 struct freeblks *freeblks;
7769 freeblks = freework->fw_freeblks;
7770 if ((freework->fw_state & INPROGRESS) == 0)
7771 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
7772 if ((freeblks->fb_state &
7773 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE &&
7774 LIST_EMPTY(&freeblks->fb_jblkdephd))
7775 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7779 * Start, continue, or finish the process of freeing an indirect block tree.
7780 * The free operation may be paused at any point with fw_off containing the
7781 * offset to restart from. This enables us to implement some flow control
7782 * for large truncates which may fan out and generate a huge number of
7786 handle_workitem_indirblk(freework)
7787 struct freework *freework;
7789 struct freeblks *freeblks;
7790 struct ufsmount *ump;
7793 freeblks = freework->fw_freeblks;
7794 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7796 if (freework->fw_state & DEPCOMPLETE) {
7797 handle_written_freework(freework);
7800 if (freework->fw_off == NINDIR(fs)) {
7801 freework_freeblock(freework);
7804 freework->fw_state |= INPROGRESS;
7806 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
7812 * Called when a freework structure attached to a cg buf is written. The
7813 * ref on either the parent or the freeblks structure is released and
7814 * the freeblks is added back to the worklist if there is more work to do.
7817 handle_written_freework(freework)
7818 struct freework *freework;
7820 struct freeblks *freeblks;
7821 struct freework *parent;
7823 freeblks = freework->fw_freeblks;
7824 parent = freework->fw_parent;
7825 if (freework->fw_state & DELAYEDFREE)
7826 freeblks->fb_cgwait--;
7827 freework->fw_state |= COMPLETE;
7828 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
7829 WORKITEM_FREE(freework, D_FREEWORK);
7831 if (--parent->fw_ref == 0)
7832 freework_enqueue(parent);
7835 if (--freeblks->fb_ref != 0)
7837 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) ==
7838 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd))
7839 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7843 * This workitem routine performs the block de-allocation.
7844 * The workitem is added to the pending list after the updated
7845 * inode block has been written to disk. As mentioned above,
7846 * checks regarding the number of blocks de-allocated (compared
7847 * to the number of blocks allocated for the file) are also
7848 * performed in this function.
7851 handle_workitem_freeblocks(freeblks, flags)
7852 struct freeblks *freeblks;
7855 struct freework *freework;
7856 struct newblk *newblk;
7857 struct allocindir *aip;
7858 struct ufsmount *ump;
7859 struct worklist *wk;
7861 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd),
7862 ("handle_workitem_freeblocks: Journal entries not written."));
7863 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7865 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
7866 WORKLIST_REMOVE(wk);
7867 switch (wk->wk_type) {
7869 wk->wk_state |= COMPLETE;
7870 add_to_worklist(wk, 0);
7874 free_newblk(WK_NEWBLK(wk));
7878 aip = WK_ALLOCINDIR(wk);
7880 if (aip->ai_state & DELAYEDFREE) {
7882 freework = newfreework(ump, freeblks, NULL,
7883 aip->ai_lbn, aip->ai_newblkno,
7884 ump->um_fs->fs_frag, 0, 0);
7887 newblk = WK_NEWBLK(wk);
7888 if (newblk->nb_jnewblk) {
7889 freework->fw_jnewblk = newblk->nb_jnewblk;
7890 newblk->nb_jnewblk->jn_dep = &freework->fw_list;
7891 newblk->nb_jnewblk = NULL;
7893 free_newblk(newblk);
7897 freework = WK_FREEWORK(wk);
7898 if (freework->fw_lbn <= -NDADDR)
7899 handle_workitem_indirblk(freework);
7901 freework_freeblock(freework);
7904 panic("handle_workitem_freeblocks: Unknown type %s",
7905 TYPENAME(wk->wk_type));
7908 if (freeblks->fb_ref != 0) {
7909 freeblks->fb_state &= ~INPROGRESS;
7910 wake_worklist(&freeblks->fb_list);
7915 return handle_complete_freeblocks(freeblks, flags);
7920 * Handle completion of block free via truncate. This allows fs_pending
7921 * to track the actual free block count more closely than if we only updated
7922 * it at the end. We must be careful to handle cases where the block count
7923 * on free was incorrect.
7926 freeblks_free(ump, freeblks, blocks)
7927 struct ufsmount *ump;
7928 struct freeblks *freeblks;
7932 ufs2_daddr_t remain;
7935 remain = -freeblks->fb_chkcnt;
7936 freeblks->fb_chkcnt += blocks;
7938 if (remain < blocks)
7941 fs->fs_pendingblocks -= blocks;
7947 * Once all of the freework workitems are complete we can retire the
7948 * freeblocks dependency and any journal work awaiting completion. This
7949 * can not be called until all other dependencies are stable on disk.
7952 handle_complete_freeblocks(freeblks, flags)
7953 struct freeblks *freeblks;
7956 struct inodedep *inodedep;
7960 struct ufsmount *ump;
7963 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7965 flags = LK_EXCLUSIVE | flags;
7966 spare = freeblks->fb_chkcnt;
7969 * If we did not release the expected number of blocks we may have
7970 * to adjust the inode block count here. Only do so if it wasn't
7971 * a truncation to zero and the modrev still matches.
7973 if (spare && freeblks->fb_len != 0) {
7974 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum,
7975 flags, &vp, FFSV_FORCEINSMQ) != 0)
7978 if (DIP(ip, i_modrev) == freeblks->fb_modrev) {
7979 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare);
7980 ip->i_flag |= IN_CHANGE;
7982 * We must wait so this happens before the
7983 * journal is reclaimed.
7991 fs->fs_pendingblocks += spare;
7997 quotaadj(freeblks->fb_quota, ump, -spare);
7998 quotarele(freeblks->fb_quota);
8001 if (freeblks->fb_state & ONDEPLIST) {
8002 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum,
8004 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next);
8005 freeblks->fb_state &= ~ONDEPLIST;
8006 if (TAILQ_EMPTY(&inodedep->id_freeblklst))
8007 free_inodedep(inodedep);
8010 * All of the freeblock deps must be complete prior to this call
8011 * so it's now safe to complete earlier outstanding journal entries.
8013 handle_jwork(&freeblks->fb_jwork);
8014 WORKITEM_FREE(freeblks, D_FREEBLKS);
8020 * Release blocks associated with the freeblks and stored in the indirect
8021 * block dbn. If level is greater than SINGLE, the block is an indirect block
8022 * and recursive calls to indirtrunc must be used to cleanse other indirect
8025 * This handles partial and complete truncation of blocks. Partial is noted
8026 * with goingaway == 0. In this case the freework is completed after the
8027 * zero'd indirects are written to disk. For full truncation the freework
8028 * is completed after the block is freed.
8031 indir_trunc(freework, dbn, lbn)
8032 struct freework *freework;
8036 struct freework *nfreework;
8037 struct workhead wkhd;
8038 struct freeblks *freeblks;
8041 struct indirdep *indirdep;
8042 struct ufsmount *ump;
8044 ufs2_daddr_t nb, nnb, *bap2;
8045 ufs_lbn_t lbnadd, nlbn;
8046 int i, nblocks, ufs1fmt;
8054 freeblks = freework->fw_freeblks;
8055 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
8058 * Get buffer of block pointers to be freed. There are three cases:
8060 * 1) Partial truncate caches the indirdep pointer in the freework
8061 * which provides us a back copy to the save bp which holds the
8062 * pointers we want to clear. When this completes the zero
8063 * pointers are written to the real copy.
8064 * 2) The indirect is being completely truncated, cancel_indirdep()
8065 * eliminated the real copy and placed the indirdep on the saved
8066 * copy. The indirdep and buf are discarded when this completes.
8067 * 3) The indirect was not in memory, we read a copy off of the disk
8068 * using the devvp and drop and invalidate the buffer when we're
8073 if (freework->fw_indir != NULL) {
8075 indirdep = freework->fw_indir;
8076 bp = indirdep->ir_savebp;
8077 if (bp == NULL || bp->b_blkno != dbn)
8078 panic("indir_trunc: Bad saved buf %p blkno %jd",
8080 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) {
8082 * The lock prevents the buf dep list from changing and
8083 * indirects on devvp should only ever have one dependency.
8085 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep));
8086 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0)
8087 panic("indir_trunc: Bad indirdep %p from buf %p",
8089 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
8090 NOCRED, &bp) != 0) {
8095 /* Protects against a race with complete_trunc_indir(). */
8096 freework->fw_state &= ~INPROGRESS;
8098 * If we have an indirdep we need to enforce the truncation order
8099 * and discard it when it is complete.
8102 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) &&
8103 !TAILQ_EMPTY(&indirdep->ir_trunc)) {
8105 * Add the complete truncate to the list on the
8106 * indirdep to enforce in-order processing.
8108 if (freework->fw_indir == NULL)
8109 TAILQ_INSERT_TAIL(&indirdep->ir_trunc,
8115 * If we're goingaway, free the indirdep. Otherwise it will
8116 * linger until the write completes.
8119 free_indirdep(indirdep);
8122 /* Initialize pointers depending on block size. */
8123 if (ump->um_fstype == UFS1) {
8124 bap1 = (ufs1_daddr_t *)bp->b_data;
8125 nb = bap1[freework->fw_off];
8129 bap2 = (ufs2_daddr_t *)bp->b_data;
8130 nb = bap2[freework->fw_off];
8134 level = lbn_level(lbn);
8135 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0;
8136 lbnadd = lbn_offset(fs, level);
8137 nblocks = btodb(fs->fs_bsize);
8138 nfreework = freework;
8142 * Reclaim blocks. Traverses into nested indirect levels and
8143 * arranges for the current level to be freed when subordinates
8144 * are free when journaling.
8146 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
8147 if (i != NINDIR(fs) - 1) {
8158 nlbn = (lbn + 1) - (i * lbnadd);
8160 nfreework = newfreework(ump, freeblks, freework,
8161 nlbn, nb, fs->fs_frag, 0, 0);
8164 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
8166 struct freedep *freedep;
8169 * Attempt to aggregate freedep dependencies for
8170 * all blocks being released to the same CG.
8174 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
8175 freedep = newfreedep(freework);
8176 WORKLIST_INSERT_UNLOCKED(&wkhd,
8181 "indir_trunc: ino %d blkno %jd size %ld",
8182 freeblks->fb_inum, nb, fs->fs_bsize);
8183 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
8184 fs->fs_bsize, freeblks->fb_inum,
8185 freeblks->fb_vtype, &wkhd);
8189 bp->b_flags |= B_INVAL | B_NOCACHE;
8194 freedblocks = (nblocks * cnt);
8196 freedblocks += nblocks;
8197 freeblks_free(ump, freeblks, freedblocks);
8199 * If we are journaling set up the ref counts and offset so this
8200 * indirect can be completed when its children are free.
8204 freework->fw_off = i;
8205 freework->fw_ref += freedeps;
8206 freework->fw_ref -= NINDIR(fs) + 1;
8208 freeblks->fb_cgwait += freedeps;
8209 if (freework->fw_ref == 0)
8210 freework_freeblock(freework);
8215 * If we're not journaling we can free the indirect now.
8217 dbn = dbtofsb(fs, dbn);
8219 "indir_trunc 2: ino %d blkno %jd size %ld",
8220 freeblks->fb_inum, dbn, fs->fs_bsize);
8221 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
8222 freeblks->fb_inum, freeblks->fb_vtype, NULL);
8223 /* Non SUJ softdep does single-threaded truncations. */
8224 if (freework->fw_blkno == dbn) {
8225 freework->fw_state |= ALLCOMPLETE;
8227 handle_written_freework(freework);
8234 * Cancel an allocindir when it is removed via truncation. When bp is not
8235 * NULL the indirect never appeared on disk and is scheduled to be freed
8236 * independently of the indir so we can more easily track journal work.
8239 cancel_allocindir(aip, bp, freeblks, trunc)
8240 struct allocindir *aip;
8242 struct freeblks *freeblks;
8245 struct indirdep *indirdep;
8246 struct freefrag *freefrag;
8247 struct newblk *newblk;
8249 newblk = (struct newblk *)aip;
8250 LIST_REMOVE(aip, ai_next);
8252 * We must eliminate the pointer in bp if it must be freed on its
8253 * own due to partial truncate or pending journal work.
8255 if (bp && (trunc || newblk->nb_jnewblk)) {
8257 * Clear the pointer and mark the aip to be freed
8258 * directly if it never existed on disk.
8260 aip->ai_state |= DELAYEDFREE;
8261 indirdep = aip->ai_indirdep;
8262 if (indirdep->ir_state & UFS1FMT)
8263 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8265 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8268 * When truncating the previous pointer will be freed via
8269 * savedbp. Eliminate the freefrag which would dup free.
8271 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) {
8272 newblk->nb_freefrag = NULL;
8273 if (freefrag->ff_jdep)
8275 WK_JFREEFRAG(freefrag->ff_jdep));
8276 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork);
8277 WORKITEM_FREE(freefrag, D_FREEFRAG);
8280 * If the journal hasn't been written the jnewblk must be passed
8281 * to the call to ffs_blkfree that reclaims the space. We accomplish
8282 * this by leaving the journal dependency on the newblk to be freed
8283 * when a freework is created in handle_workitem_freeblocks().
8285 cancel_newblk(newblk, NULL, &freeblks->fb_jwork);
8286 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
8290 * Create the mkdir dependencies for . and .. in a new directory. Link them
8291 * in to a newdirblk so any subsequent additions are tracked properly. The
8292 * caller is responsible for adding the mkdir1 dependency to the journal
8293 * and updating id_mkdiradd. This function returns with the per-filesystem
8296 static struct mkdir *
8297 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
8301 struct buf *newdirbp;
8302 struct mkdir **mkdirp;
8304 struct newblk *newblk;
8305 struct pagedep *pagedep;
8306 struct inodedep *inodedep;
8307 struct newdirblk *newdirblk;
8308 struct mkdir *mkdir1, *mkdir2;
8309 struct worklist *wk;
8310 struct jaddref *jaddref;
8311 struct ufsmount *ump;
8314 mp = dap->da_list.wk_mp;
8316 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
8318 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8319 LIST_INIT(&newdirblk->db_mkdir);
8320 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8321 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
8322 mkdir1->md_state = ATTACHED | MKDIR_BODY;
8323 mkdir1->md_diradd = dap;
8324 mkdir1->md_jaddref = NULL;
8325 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8326 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
8327 mkdir2->md_state = ATTACHED | MKDIR_PARENT;
8328 mkdir2->md_diradd = dap;
8329 mkdir2->md_jaddref = NULL;
8330 if (MOUNTEDSUJ(mp) == 0) {
8331 mkdir1->md_state |= DEPCOMPLETE;
8332 mkdir2->md_state |= DEPCOMPLETE;
8335 * Dependency on "." and ".." being written to disk.
8337 mkdir1->md_buf = newdirbp;
8338 ACQUIRE_LOCK(VFSTOUFS(mp));
8339 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs);
8341 * We must link the pagedep, allocdirect, and newdirblk for
8342 * the initial file page so the pointer to the new directory
8343 * is not written until the directory contents are live and
8344 * any subsequent additions are not marked live until the
8345 * block is reachable via the inode.
8347 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0)
8348 panic("setup_newdir: lost pagedep");
8349 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
8350 if (wk->wk_type == D_ALLOCDIRECT)
8353 panic("setup_newdir: lost allocdirect");
8354 if (pagedep->pd_state & NEWBLOCK)
8355 panic("setup_newdir: NEWBLOCK already set");
8356 newblk = WK_NEWBLK(wk);
8357 pagedep->pd_state |= NEWBLOCK;
8358 pagedep->pd_newdirblk = newdirblk;
8359 newdirblk->db_pagedep = pagedep;
8360 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8361 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
8363 * Look up the inodedep for the parent directory so that we
8364 * can link mkdir2 into the pending dotdot jaddref or
8365 * the inode write if there is none. If the inode is
8366 * ALLCOMPLETE and no jaddref is present all dependencies have
8367 * been satisfied and mkdir2 can be freed.
8369 inodedep_lookup(mp, dinum, 0, &inodedep);
8370 if (MOUNTEDSUJ(mp)) {
8371 if (inodedep == NULL)
8372 panic("setup_newdir: Lost parent.");
8373 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8375 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
8376 (jaddref->ja_state & MKDIR_PARENT),
8377 ("setup_newdir: bad dotdot jaddref %p", jaddref));
8378 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8379 mkdir2->md_jaddref = jaddref;
8380 jaddref->ja_mkdir = mkdir2;
8381 } else if (inodedep == NULL ||
8382 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
8383 dap->da_state &= ~MKDIR_PARENT;
8384 WORKITEM_FREE(mkdir2, D_MKDIR);
8387 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8388 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list);
8396 * Directory entry addition dependencies.
8398 * When adding a new directory entry, the inode (with its incremented link
8399 * count) must be written to disk before the directory entry's pointer to it.
8400 * Also, if the inode is newly allocated, the corresponding freemap must be
8401 * updated (on disk) before the directory entry's pointer. These requirements
8402 * are met via undo/redo on the directory entry's pointer, which consists
8403 * simply of the inode number.
8405 * As directory entries are added and deleted, the free space within a
8406 * directory block can become fragmented. The ufs filesystem will compact
8407 * a fragmented directory block to make space for a new entry. When this
8408 * occurs, the offsets of previously added entries change. Any "diradd"
8409 * dependency structures corresponding to these entries must be updated with
8414 * This routine is called after the in-memory inode's link
8415 * count has been incremented, but before the directory entry's
8416 * pointer to the inode has been set.
8419 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
8420 struct buf *bp; /* buffer containing directory block */
8421 struct inode *dp; /* inode for directory */
8422 off_t diroffset; /* offset of new entry in directory */
8423 ino_t newinum; /* inode referenced by new directory entry */
8424 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
8425 int isnewblk; /* entry is in a newly allocated block */
8427 int offset; /* offset of new entry within directory block */
8428 ufs_lbn_t lbn; /* block in directory containing new entry */
8431 struct newblk *newblk;
8432 struct pagedep *pagedep;
8433 struct inodedep *inodedep;
8434 struct newdirblk *newdirblk;
8435 struct mkdir *mkdir1, *mkdir2;
8436 struct jaddref *jaddref;
8437 struct ufsmount *ump;
8443 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8444 ("softdep_setup_directory_add called on non-softdep filesystem"));
8446 * Whiteouts have no dependencies.
8448 if (newinum == WINO) {
8449 if (newdirbp != NULL)
8454 mkdir1 = mkdir2 = NULL;
8456 lbn = lblkno(fs, diroffset);
8457 offset = blkoff(fs, diroffset);
8458 dap = malloc(sizeof(struct diradd), M_DIRADD,
8459 M_SOFTDEP_FLAGS|M_ZERO);
8460 workitem_alloc(&dap->da_list, D_DIRADD, mp);
8461 dap->da_offset = offset;
8462 dap->da_newinum = newinum;
8463 dap->da_state = ATTACHED;
8464 LIST_INIT(&dap->da_jwork);
8465 isindir = bp->b_lblkno >= NDADDR;
8468 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
8469 newdirblk = malloc(sizeof(struct newdirblk),
8470 M_NEWDIRBLK, M_SOFTDEP_FLAGS);
8471 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8472 LIST_INIT(&newdirblk->db_mkdir);
8475 * If we're creating a new directory setup the dependencies and set
8476 * the dap state to wait for them. Otherwise it's COMPLETE and
8479 if (newdirbp == NULL) {
8480 dap->da_state |= DEPCOMPLETE;
8483 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
8484 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
8488 * Link into parent directory pagedep to await its being written.
8490 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep);
8492 if (diradd_lookup(pagedep, offset) != NULL)
8493 panic("softdep_setup_directory_add: %p already at off %d\n",
8494 diradd_lookup(pagedep, offset), offset);
8496 dap->da_pagedep = pagedep;
8497 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
8499 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
8501 * If we're journaling, link the diradd into the jaddref so it
8502 * may be completed after the journal entry is written. Otherwise,
8503 * link the diradd into its inodedep. If the inode is not yet
8504 * written place it on the bufwait list, otherwise do the post-inode
8505 * write processing to put it on the id_pendinghd list.
8507 if (MOUNTEDSUJ(mp)) {
8508 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8510 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
8511 ("softdep_setup_directory_add: bad jaddref %p", jaddref));
8512 jaddref->ja_diroff = diroffset;
8513 jaddref->ja_diradd = dap;
8514 add_to_journal(&jaddref->ja_list);
8515 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
8516 diradd_inode_written(dap, inodedep);
8518 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
8520 * Add the journal entries for . and .. links now that the primary
8523 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) {
8524 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
8525 inoreflst, if_deps);
8526 KASSERT(jaddref != NULL &&
8527 jaddref->ja_ino == jaddref->ja_parent &&
8528 (jaddref->ja_state & MKDIR_BODY),
8529 ("softdep_setup_directory_add: bad dot jaddref %p",
8531 mkdir1->md_jaddref = jaddref;
8532 jaddref->ja_mkdir = mkdir1;
8534 * It is important that the dotdot journal entry
8535 * is added prior to the dot entry since dot writes
8536 * both the dot and dotdot links. These both must
8537 * be added after the primary link for the journal
8538 * to remain consistent.
8540 add_to_journal(&mkdir2->md_jaddref->ja_list);
8541 add_to_journal(&jaddref->ja_list);
8544 * If we are adding a new directory remember this diradd so that if
8545 * we rename it we can keep the dot and dotdot dependencies. If
8546 * we are adding a new name for an inode that has a mkdiradd we
8547 * must be in rename and we have to move the dot and dotdot
8548 * dependencies to this new name. The old name is being orphaned
8551 if (mkdir1 != NULL) {
8552 if (inodedep->id_mkdiradd != NULL)
8553 panic("softdep_setup_directory_add: Existing mkdir");
8554 inodedep->id_mkdiradd = dap;
8555 } else if (inodedep->id_mkdiradd)
8556 merge_diradd(inodedep, dap);
8557 if (newdirblk != NULL) {
8559 * There is nothing to do if we are already tracking
8562 if ((pagedep->pd_state & NEWBLOCK) != 0) {
8563 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
8567 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
8569 panic("softdep_setup_directory_add: lost entry");
8570 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8571 pagedep->pd_state |= NEWBLOCK;
8572 pagedep->pd_newdirblk = newdirblk;
8573 newdirblk->db_pagedep = pagedep;
8576 * If we extended into an indirect signal direnter to sync.
8587 * This procedure is called to change the offset of a directory
8588 * entry when compacting a directory block which must be owned
8589 * exclusively by the caller. Note that the actual entry movement
8590 * must be done in this procedure to ensure that no I/O completions
8591 * occur while the move is in progress.
8594 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
8595 struct buf *bp; /* Buffer holding directory block. */
8596 struct inode *dp; /* inode for directory */
8597 caddr_t base; /* address of dp->i_offset */
8598 caddr_t oldloc; /* address of old directory location */
8599 caddr_t newloc; /* address of new directory location */
8600 int entrysize; /* size of directory entry */
8602 int offset, oldoffset, newoffset;
8603 struct pagedep *pagedep;
8604 struct jmvref *jmvref;
8611 mp = UFSTOVFS(dp->i_ump);
8612 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8613 ("softdep_change_directoryentry_offset called on "
8614 "non-softdep filesystem"));
8615 de = (struct direct *)oldloc;
8619 * Moves are always journaled as it would be too complex to
8620 * determine if any affected adds or removes are present in the
8623 if (MOUNTEDSUJ(mp)) {
8625 jmvref = newjmvref(dp, de->d_ino,
8626 dp->i_offset + (oldloc - base),
8627 dp->i_offset + (newloc - base));
8629 lbn = lblkno(dp->i_fs, dp->i_offset);
8630 offset = blkoff(dp->i_fs, dp->i_offset);
8631 oldoffset = offset + (oldloc - base);
8632 newoffset = offset + (newloc - base);
8633 ACQUIRE_LOCK(dp->i_ump);
8634 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0)
8636 dap = diradd_lookup(pagedep, oldoffset);
8638 dap->da_offset = newoffset;
8639 newoffset = DIRADDHASH(newoffset);
8640 oldoffset = DIRADDHASH(oldoffset);
8641 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
8642 newoffset != oldoffset) {
8643 LIST_REMOVE(dap, da_pdlist);
8644 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
8650 jmvref->jm_pagedep = pagedep;
8651 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
8652 add_to_journal(&jmvref->jm_list);
8654 bcopy(oldloc, newloc, entrysize);
8655 FREE_LOCK(dp->i_ump);
8659 * Move the mkdir dependencies and journal work from one diradd to another
8660 * when renaming a directory. The new name must depend on the mkdir deps
8661 * completing as the old name did. Directories can only have one valid link
8662 * at a time so one must be canonical.
8665 merge_diradd(inodedep, newdap)
8666 struct inodedep *inodedep;
8667 struct diradd *newdap;
8669 struct diradd *olddap;
8670 struct mkdir *mkdir, *nextmd;
8671 struct ufsmount *ump;
8674 olddap = inodedep->id_mkdiradd;
8675 inodedep->id_mkdiradd = newdap;
8676 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8677 newdap->da_state &= ~DEPCOMPLETE;
8678 ump = VFSTOUFS(inodedep->id_list.wk_mp);
8679 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8681 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8682 if (mkdir->md_diradd != olddap)
8684 mkdir->md_diradd = newdap;
8685 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
8686 newdap->da_state |= state;
8687 olddap->da_state &= ~state;
8688 if ((olddap->da_state &
8689 (MKDIR_PARENT | MKDIR_BODY)) == 0)
8692 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8693 panic("merge_diradd: unfound ref");
8696 * Any mkdir related journal items are not safe to be freed until
8697 * the new name is stable.
8699 jwork_move(&newdap->da_jwork, &olddap->da_jwork);
8700 olddap->da_state |= DEPCOMPLETE;
8701 complete_diradd(olddap);
8705 * Move the diradd to the pending list when all diradd dependencies are
8709 complete_diradd(dap)
8712 struct pagedep *pagedep;
8714 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
8715 if (dap->da_state & DIRCHG)
8716 pagedep = dap->da_previous->dm_pagedep;
8718 pagedep = dap->da_pagedep;
8719 LIST_REMOVE(dap, da_pdlist);
8720 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
8725 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal
8726 * add entries and conditonally journal the remove.
8729 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
8731 struct dirrem *dirrem;
8732 struct jremref *jremref;
8733 struct jremref *dotremref;
8734 struct jremref *dotdotremref;
8736 struct inodedep *inodedep;
8737 struct jaddref *jaddref;
8738 struct inoref *inoref;
8739 struct ufsmount *ump;
8740 struct mkdir *mkdir;
8743 * If no remove references were allocated we're on a non-journaled
8744 * filesystem and can skip the cancel step.
8746 if (jremref == NULL) {
8747 free_diradd(dap, NULL);
8751 * Cancel the primary name an free it if it does not require
8754 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
8755 0, &inodedep) != 0) {
8756 /* Abort the addref that reference this diradd. */
8757 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
8758 if (inoref->if_list.wk_type != D_JADDREF)
8760 jaddref = (struct jaddref *)inoref;
8761 if (jaddref->ja_diradd != dap)
8763 if (cancel_jaddref(jaddref, inodedep,
8764 &dirrem->dm_jwork) == 0) {
8765 free_jremref(jremref);
8772 * Cancel subordinate names and free them if they do not require
8775 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8776 ump = VFSTOUFS(dap->da_list.wk_mp);
8777 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) {
8778 if (mkdir->md_diradd != dap)
8780 if ((jaddref = mkdir->md_jaddref) == NULL)
8782 mkdir->md_jaddref = NULL;
8783 if (mkdir->md_state & MKDIR_PARENT) {
8784 if (cancel_jaddref(jaddref, NULL,
8785 &dirrem->dm_jwork) == 0) {
8786 free_jremref(dotdotremref);
8787 dotdotremref = NULL;
8790 if (cancel_jaddref(jaddref, inodedep,
8791 &dirrem->dm_jwork) == 0) {
8792 free_jremref(dotremref);
8800 journal_jremref(dirrem, jremref, inodedep);
8802 journal_jremref(dirrem, dotremref, inodedep);
8804 journal_jremref(dirrem, dotdotremref, NULL);
8805 jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
8806 free_diradd(dap, &dirrem->dm_jwork);
8810 * Free a diradd dependency structure. This routine must be called
8811 * with splbio interrupts blocked.
8814 free_diradd(dap, wkhd)
8816 struct workhead *wkhd;
8818 struct dirrem *dirrem;
8819 struct pagedep *pagedep;
8820 struct inodedep *inodedep;
8821 struct mkdir *mkdir, *nextmd;
8822 struct ufsmount *ump;
8824 ump = VFSTOUFS(dap->da_list.wk_mp);
8826 LIST_REMOVE(dap, da_pdlist);
8827 if (dap->da_state & ONWORKLIST)
8828 WORKLIST_REMOVE(&dap->da_list);
8829 if ((dap->da_state & DIRCHG) == 0) {
8830 pagedep = dap->da_pagedep;
8832 dirrem = dap->da_previous;
8833 pagedep = dirrem->dm_pagedep;
8834 dirrem->dm_dirinum = pagedep->pd_ino;
8835 dirrem->dm_state |= COMPLETE;
8836 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
8837 add_to_worklist(&dirrem->dm_list, 0);
8839 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
8841 if (inodedep->id_mkdiradd == dap)
8842 inodedep->id_mkdiradd = NULL;
8843 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8844 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8846 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8847 if (mkdir->md_diradd != dap)
8850 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
8851 LIST_REMOVE(mkdir, md_mkdirs);
8852 if (mkdir->md_state & ONWORKLIST)
8853 WORKLIST_REMOVE(&mkdir->md_list);
8854 if (mkdir->md_jaddref != NULL)
8855 panic("free_diradd: Unexpected jaddref");
8856 WORKITEM_FREE(mkdir, D_MKDIR);
8857 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
8860 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8861 panic("free_diradd: unfound ref");
8864 free_inodedep(inodedep);
8866 * Free any journal segments waiting for the directory write.
8868 handle_jwork(&dap->da_jwork);
8869 WORKITEM_FREE(dap, D_DIRADD);
8873 * Directory entry removal dependencies.
8875 * When removing a directory entry, the entry's inode pointer must be
8876 * zero'ed on disk before the corresponding inode's link count is decremented
8877 * (possibly freeing the inode for re-use). This dependency is handled by
8878 * updating the directory entry but delaying the inode count reduction until
8879 * after the directory block has been written to disk. After this point, the
8880 * inode count can be decremented whenever it is convenient.
8884 * This routine should be called immediately after removing
8885 * a directory entry. The inode's link count should not be
8886 * decremented by the calling procedure -- the soft updates
8887 * code will do this task when it is safe.
8890 softdep_setup_remove(bp, dp, ip, isrmdir)
8891 struct buf *bp; /* buffer containing directory block */
8892 struct inode *dp; /* inode for the directory being modified */
8893 struct inode *ip; /* inode for directory entry being removed */
8894 int isrmdir; /* indicates if doing RMDIR */
8896 struct dirrem *dirrem, *prevdirrem;
8897 struct inodedep *inodedep;
8900 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
8901 ("softdep_setup_remove called on non-softdep filesystem"));
8903 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want
8904 * newdirrem() to setup the full directory remove which requires
8907 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
8909 * Add the dirrem to the inodedep's pending remove list for quick
8912 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
8914 panic("softdep_setup_remove: Lost inodedep.");
8915 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
8916 dirrem->dm_state |= ONDEPLIST;
8917 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
8920 * If the COMPLETE flag is clear, then there were no active
8921 * entries and we want to roll back to a zeroed entry until
8922 * the new inode is committed to disk. If the COMPLETE flag is
8923 * set then we have deleted an entry that never made it to
8924 * disk. If the entry we deleted resulted from a name change,
8925 * then the old name still resides on disk. We cannot delete
8926 * its inode (returned to us in prevdirrem) until the zeroed
8927 * directory entry gets to disk. The new inode has never been
8928 * referenced on the disk, so can be deleted immediately.
8930 if ((dirrem->dm_state & COMPLETE) == 0) {
8931 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
8933 FREE_LOCK(ip->i_ump);
8935 if (prevdirrem != NULL)
8936 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
8937 prevdirrem, dm_next);
8938 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
8939 direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
8940 FREE_LOCK(ip->i_ump);
8942 handle_workitem_remove(dirrem, 0);
8947 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
8948 * pd_pendinghd list of a pagedep.
8950 static struct diradd *
8951 diradd_lookup(pagedep, offset)
8952 struct pagedep *pagedep;
8957 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
8958 if (dap->da_offset == offset)
8960 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
8961 if (dap->da_offset == offset)
8967 * Search for a .. diradd dependency in a directory that is being removed.
8968 * If the directory was renamed to a new parent we have a diradd rather
8969 * than a mkdir for the .. entry. We need to cancel it now before
8970 * it is found in truncate().
8972 static struct jremref *
8973 cancel_diradd_dotdot(ip, dirrem, jremref)
8975 struct dirrem *dirrem;
8976 struct jremref *jremref;
8978 struct pagedep *pagedep;
8980 struct worklist *wk;
8982 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0,
8985 dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
8988 cancel_diradd(dap, dirrem, jremref, NULL, NULL);
8990 * Mark any journal work as belonging to the parent so it is freed
8991 * with the .. reference.
8993 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
8994 wk->wk_state |= MKDIR_PARENT;
8999 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
9000 * replace it with a dirrem/diradd pair as a result of re-parenting a
9001 * directory. This ensures that we don't simultaneously have a mkdir and
9002 * a diradd for the same .. entry.
9004 static struct jremref *
9005 cancel_mkdir_dotdot(ip, dirrem, jremref)
9007 struct dirrem *dirrem;
9008 struct jremref *jremref;
9010 struct inodedep *inodedep;
9011 struct jaddref *jaddref;
9012 struct ufsmount *ump;
9013 struct mkdir *mkdir;
9016 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
9019 dap = inodedep->id_mkdiradd;
9020 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
9022 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9023 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
9024 mkdir = LIST_NEXT(mkdir, md_mkdirs))
9025 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
9028 panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
9029 if ((jaddref = mkdir->md_jaddref) != NULL) {
9030 mkdir->md_jaddref = NULL;
9031 jaddref->ja_state &= ~MKDIR_PARENT;
9032 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0,
9034 panic("cancel_mkdir_dotdot: Lost parent inodedep");
9035 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
9036 journal_jremref(dirrem, jremref, inodedep);
9040 if (mkdir->md_state & ONWORKLIST)
9041 WORKLIST_REMOVE(&mkdir->md_list);
9042 mkdir->md_state |= ALLCOMPLETE;
9043 complete_mkdir(mkdir);
9048 journal_jremref(dirrem, jremref, inodedep)
9049 struct dirrem *dirrem;
9050 struct jremref *jremref;
9051 struct inodedep *inodedep;
9054 if (inodedep == NULL)
9055 if (inodedep_lookup(jremref->jr_list.wk_mp,
9056 jremref->jr_ref.if_ino, 0, &inodedep) == 0)
9057 panic("journal_jremref: Lost inodedep");
9058 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
9059 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
9060 add_to_journal(&jremref->jr_list);
9064 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
9065 struct dirrem *dirrem;
9066 struct jremref *jremref;
9067 struct jremref *dotremref;
9068 struct jremref *dotdotremref;
9070 struct inodedep *inodedep;
9073 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
9075 panic("dirrem_journal: Lost inodedep");
9076 journal_jremref(dirrem, jremref, inodedep);
9078 journal_jremref(dirrem, dotremref, inodedep);
9080 journal_jremref(dirrem, dotdotremref, NULL);
9084 * Allocate a new dirrem if appropriate and return it along with
9085 * its associated pagedep. Called without a lock, returns with lock.
9087 static struct dirrem *
9088 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
9089 struct buf *bp; /* buffer containing directory block */
9090 struct inode *dp; /* inode for the directory being modified */
9091 struct inode *ip; /* inode for directory entry being removed */
9092 int isrmdir; /* indicates if doing RMDIR */
9093 struct dirrem **prevdirremp; /* previously referenced inode, if any */
9098 struct dirrem *dirrem;
9099 struct pagedep *pagedep;
9100 struct jremref *jremref;
9101 struct jremref *dotremref;
9102 struct jremref *dotdotremref;
9106 * Whiteouts have no deletion dependencies.
9109 panic("newdirrem: whiteout");
9112 * If the system is over its limit and our filesystem is
9113 * responsible for more than our share of that usage and
9114 * we are not a snapshot, request some inodedep cleanup.
9115 * Limiting the number of dirrem structures will also limit
9116 * the number of freefile and freeblks structures.
9118 ACQUIRE_LOCK(ip->i_ump);
9119 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ip->i_ump, D_DIRREM))
9120 schedule_cleanup(ITOV(dp)->v_mount);
9122 FREE_LOCK(ip->i_ump);
9123 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS |
9125 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
9126 LIST_INIT(&dirrem->dm_jremrefhd);
9127 LIST_INIT(&dirrem->dm_jwork);
9128 dirrem->dm_state = isrmdir ? RMDIR : 0;
9129 dirrem->dm_oldinum = ip->i_number;
9130 *prevdirremp = NULL;
9132 * Allocate remove reference structures to track journal write
9133 * dependencies. We will always have one for the link and
9134 * when doing directories we will always have one more for dot.
9135 * When renaming a directory we skip the dotdot link change so
9136 * this is not needed.
9138 jremref = dotremref = dotdotremref = NULL;
9139 if (DOINGSUJ(dvp)) {
9141 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9142 ip->i_effnlink + 2);
9143 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
9144 ip->i_effnlink + 1);
9145 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
9146 dp->i_effnlink + 1);
9147 dotdotremref->jr_state |= MKDIR_PARENT;
9149 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9150 ip->i_effnlink + 1);
9152 ACQUIRE_LOCK(ip->i_ump);
9153 lbn = lblkno(dp->i_fs, dp->i_offset);
9154 offset = blkoff(dp->i_fs, dp->i_offset);
9155 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC,
9157 dirrem->dm_pagedep = pagedep;
9158 dirrem->dm_offset = offset;
9160 * If we're renaming a .. link to a new directory, cancel any
9161 * existing MKDIR_PARENT mkdir. If it has already been canceled
9162 * the jremref is preserved for any potential diradd in this
9163 * location. This can not coincide with a rmdir.
9165 if (dp->i_offset == DOTDOT_OFFSET) {
9167 panic("newdirrem: .. directory change during remove?");
9168 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
9171 * If we're removing a directory search for the .. dependency now and
9172 * cancel it. Any pending journal work will be added to the dirrem
9173 * to be completed when the workitem remove completes.
9176 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
9178 * Check for a diradd dependency for the same directory entry.
9179 * If present, then both dependencies become obsolete and can
9182 dap = diradd_lookup(pagedep, offset);
9185 * Link the jremref structures into the dirrem so they are
9186 * written prior to the pagedep.
9189 dirrem_journal(dirrem, jremref, dotremref,
9194 * Must be ATTACHED at this point.
9196 if ((dap->da_state & ATTACHED) == 0)
9197 panic("newdirrem: not ATTACHED");
9198 if (dap->da_newinum != ip->i_number)
9199 panic("newdirrem: inum %ju should be %ju",
9200 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum);
9202 * If we are deleting a changed name that never made it to disk,
9203 * then return the dirrem describing the previous inode (which
9204 * represents the inode currently referenced from this entry on disk).
9206 if ((dap->da_state & DIRCHG) != 0) {
9207 *prevdirremp = dap->da_previous;
9208 dap->da_state &= ~DIRCHG;
9209 dap->da_pagedep = pagedep;
9212 * We are deleting an entry that never made it to disk.
9213 * Mark it COMPLETE so we can delete its inode immediately.
9215 dirrem->dm_state |= COMPLETE;
9216 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
9219 struct worklist *wk;
9221 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
9222 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
9223 panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
9231 * Directory entry change dependencies.
9233 * Changing an existing directory entry requires that an add operation
9234 * be completed first followed by a deletion. The semantics for the addition
9235 * are identical to the description of adding a new entry above except
9236 * that the rollback is to the old inode number rather than zero. Once
9237 * the addition dependency is completed, the removal is done as described
9238 * in the removal routine above.
9242 * This routine should be called immediately after changing
9243 * a directory entry. The inode's link count should not be
9244 * decremented by the calling procedure -- the soft updates
9245 * code will perform this task when it is safe.
9248 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
9249 struct buf *bp; /* buffer containing directory block */
9250 struct inode *dp; /* inode for the directory being modified */
9251 struct inode *ip; /* inode for directory entry being removed */
9252 ino_t newinum; /* new inode number for changed entry */
9253 int isrmdir; /* indicates if doing RMDIR */
9256 struct diradd *dap = NULL;
9257 struct dirrem *dirrem, *prevdirrem;
9258 struct pagedep *pagedep;
9259 struct inodedep *inodedep;
9260 struct jaddref *jaddref;
9263 offset = blkoff(dp->i_fs, dp->i_offset);
9264 mp = UFSTOVFS(dp->i_ump);
9265 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
9266 ("softdep_setup_directory_change called on non-softdep filesystem"));
9269 * Whiteouts do not need diradd dependencies.
9271 if (newinum != WINO) {
9272 dap = malloc(sizeof(struct diradd),
9273 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
9274 workitem_alloc(&dap->da_list, D_DIRADD, mp);
9275 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
9276 dap->da_offset = offset;
9277 dap->da_newinum = newinum;
9278 LIST_INIT(&dap->da_jwork);
9282 * Allocate a new dirrem and ACQUIRE_LOCK.
9284 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
9285 pagedep = dirrem->dm_pagedep;
9287 * The possible values for isrmdir:
9288 * 0 - non-directory file rename
9289 * 1 - directory rename within same directory
9290 * inum - directory rename to new directory of given inode number
9291 * When renaming to a new directory, we are both deleting and
9292 * creating a new directory entry, so the link count on the new
9293 * directory should not change. Thus we do not need the followup
9294 * dirrem which is usually done in handle_workitem_remove. We set
9295 * the DIRCHG flag to tell handle_workitem_remove to skip the
9299 dirrem->dm_state |= DIRCHG;
9302 * Whiteouts have no additional dependencies,
9303 * so just put the dirrem on the correct list.
9305 if (newinum == WINO) {
9306 if ((dirrem->dm_state & COMPLETE) == 0) {
9307 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
9310 dirrem->dm_dirinum = pagedep->pd_ino;
9311 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9312 add_to_worklist(&dirrem->dm_list, 0);
9314 FREE_LOCK(dp->i_ump);
9318 * Add the dirrem to the inodedep's pending remove list for quick
9319 * discovery later. A valid nlinkdelta ensures that this lookup
9322 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
9323 panic("softdep_setup_directory_change: Lost inodedep.");
9324 dirrem->dm_state |= ONDEPLIST;
9325 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9328 * If the COMPLETE flag is clear, then there were no active
9329 * entries and we want to roll back to the previous inode until
9330 * the new inode is committed to disk. If the COMPLETE flag is
9331 * set, then we have deleted an entry that never made it to disk.
9332 * If the entry we deleted resulted from a name change, then the old
9333 * inode reference still resides on disk. Any rollback that we do
9334 * needs to be to that old inode (returned to us in prevdirrem). If
9335 * the entry we deleted resulted from a create, then there is
9336 * no entry on the disk, so we want to roll back to zero rather
9337 * than the uncommitted inode. In either of the COMPLETE cases we
9338 * want to immediately free the unwritten and unreferenced inode.
9340 if ((dirrem->dm_state & COMPLETE) == 0) {
9341 dap->da_previous = dirrem;
9343 if (prevdirrem != NULL) {
9344 dap->da_previous = prevdirrem;
9346 dap->da_state &= ~DIRCHG;
9347 dap->da_pagedep = pagedep;
9349 dirrem->dm_dirinum = pagedep->pd_ino;
9350 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9351 add_to_worklist(&dirrem->dm_list, 0);
9354 * Lookup the jaddref for this journal entry. We must finish
9355 * initializing it and make the diradd write dependent on it.
9356 * If we're not journaling, put it on the id_bufwait list if the
9357 * inode is not yet written. If it is written, do the post-inode
9358 * write processing to put it on the id_pendinghd list.
9360 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
9361 if (MOUNTEDSUJ(mp)) {
9362 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
9364 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
9365 ("softdep_setup_directory_change: bad jaddref %p",
9367 jaddref->ja_diroff = dp->i_offset;
9368 jaddref->ja_diradd = dap;
9369 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9371 add_to_journal(&jaddref->ja_list);
9372 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
9373 dap->da_state |= COMPLETE;
9374 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
9375 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
9377 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9379 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
9382 * If we're making a new name for a directory that has not been
9383 * committed when need to move the dot and dotdot references to
9386 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
9387 merge_diradd(inodedep, dap);
9388 FREE_LOCK(dp->i_ump);
9392 * Called whenever the link count on an inode is changed.
9393 * It creates an inode dependency so that the new reference(s)
9394 * to the inode cannot be committed to disk until the updated
9395 * inode has been written.
9398 softdep_change_linkcnt(ip)
9399 struct inode *ip; /* the inode with the increased link count */
9401 struct inodedep *inodedep;
9403 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
9404 ("softdep_change_linkcnt called on non-softdep filesystem"));
9405 ACQUIRE_LOCK(ip->i_ump);
9406 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
9408 if (ip->i_nlink < ip->i_effnlink)
9409 panic("softdep_change_linkcnt: bad delta");
9410 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9411 FREE_LOCK(ip->i_ump);
9415 * Attach a sbdep dependency to the superblock buf so that we can keep
9416 * track of the head of the linked list of referenced but unlinked inodes.
9419 softdep_setup_sbupdate(ump, fs, bp)
9420 struct ufsmount *ump;
9424 struct sbdep *sbdep;
9425 struct worklist *wk;
9427 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
9428 ("softdep_setup_sbupdate called on non-softdep filesystem"));
9429 LIST_FOREACH(wk, &bp->b_dep, wk_list)
9430 if (wk->wk_type == D_SBDEP)
9434 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
9435 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
9437 sbdep->sb_ump = ump;
9439 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
9444 * Return the first unlinked inodedep which is ready to be the head of the
9445 * list. The inodedep and all those after it must have valid next pointers.
9447 static struct inodedep *
9448 first_unlinked_inodedep(ump)
9449 struct ufsmount *ump;
9451 struct inodedep *inodedep;
9452 struct inodedep *idp;
9455 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
9456 inodedep; inodedep = idp) {
9457 if ((inodedep->id_state & UNLINKNEXT) == 0)
9459 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9460 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
9462 if ((inodedep->id_state & UNLINKPREV) == 0)
9469 * Set the sujfree unlinked head pointer prior to writing a superblock.
9472 initiate_write_sbdep(sbdep)
9473 struct sbdep *sbdep;
9475 struct inodedep *inodedep;
9479 bpfs = sbdep->sb_fs;
9480 fs = sbdep->sb_ump->um_fs;
9481 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9483 fs->fs_sujfree = inodedep->id_ino;
9484 inodedep->id_state |= UNLINKPREV;
9487 bpfs->fs_sujfree = fs->fs_sujfree;
9491 * After a superblock is written determine whether it must be written again
9492 * due to a changing unlinked list head.
9495 handle_written_sbdep(sbdep, bp)
9496 struct sbdep *sbdep;
9499 struct inodedep *inodedep;
9502 LOCK_OWNED(sbdep->sb_ump);
9505 * If the superblock doesn't match the in-memory list start over.
9507 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9508 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
9509 (inodedep == NULL && fs->fs_sujfree != 0)) {
9513 WORKITEM_FREE(sbdep, D_SBDEP);
9514 if (fs->fs_sujfree == 0)
9517 * Now that we have a record of this inode in stable store allow it
9518 * to be written to free up pending work. Inodes may see a lot of
9519 * write activity after they are unlinked which we must not hold up.
9521 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
9522 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
9523 panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
9524 inodedep, inodedep->id_state);
9525 if (inodedep->id_state & UNLINKONLIST)
9527 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
9534 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
9537 unlinked_inodedep(mp, inodedep)
9539 struct inodedep *inodedep;
9541 struct ufsmount *ump;
9545 if (MOUNTEDSUJ(mp) == 0)
9547 ump->um_fs->fs_fmod = 1;
9548 if (inodedep->id_state & UNLINKED)
9549 panic("unlinked_inodedep: %p already unlinked\n", inodedep);
9550 inodedep->id_state |= UNLINKED;
9551 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
9555 * Remove an inodedep from the unlinked inodedep list. This may require
9556 * disk writes if the inode has made it that far.
9559 clear_unlinked_inodedep(inodedep)
9560 struct inodedep *inodedep;
9562 struct ufsmount *ump;
9563 struct inodedep *idp;
9564 struct inodedep *idn;
9572 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9574 ino = inodedep->id_ino;
9578 KASSERT((inodedep->id_state & UNLINKED) != 0,
9579 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9582 * If nothing has yet been written simply remove us from
9583 * the in memory list and return. This is the most common
9584 * case where handle_workitem_remove() loses the final
9587 if ((inodedep->id_state & UNLINKLINKS) == 0)
9590 * If we have a NEXT pointer and no PREV pointer we can simply
9591 * clear NEXT's PREV and remove ourselves from the list. Be
9592 * careful not to clear PREV if the superblock points at
9595 idn = TAILQ_NEXT(inodedep, id_unlinked);
9596 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
9597 if (idn && fs->fs_sujfree != idn->id_ino)
9598 idn->id_state &= ~UNLINKPREV;
9602 * Here we have an inodedep which is actually linked into
9603 * the list. We must remove it by forcing a write to the
9604 * link before us, whether it be the superblock or an inode.
9605 * Unfortunately the list may change while we're waiting
9606 * on the buf lock for either resource so we must loop until
9607 * we lock the right one. If both the superblock and an
9608 * inode point to this inode we must clear the inode first
9609 * followed by the superblock.
9611 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9613 if (idp && (idp->id_state & UNLINKNEXT))
9617 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9618 (int)fs->fs_sbsize, 0, 0, 0);
9620 error = bread(ump->um_devvp,
9621 fsbtodb(fs, ino_to_fsba(fs, pino)),
9622 (int)fs->fs_bsize, NOCRED, &bp);
9629 /* If the list has changed restart the loop. */
9630 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9632 if (idp && (idp->id_state & UNLINKNEXT))
9635 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
9642 idn = TAILQ_NEXT(inodedep, id_unlinked);
9646 * Remove us from the in memory list. After this we cannot
9647 * access the inodedep.
9649 KASSERT((inodedep->id_state & UNLINKED) != 0,
9650 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9652 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9653 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9656 * The predecessor's next pointer is manually updated here
9657 * so that the NEXT flag is never cleared for an element
9658 * that is in the list.
9661 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9662 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9663 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9665 } else if (fs->fs_magic == FS_UFS1_MAGIC)
9666 ((struct ufs1_dinode *)bp->b_data +
9667 ino_to_fsbo(fs, pino))->di_freelink = nino;
9669 ((struct ufs2_dinode *)bp->b_data +
9670 ino_to_fsbo(fs, pino))->di_freelink = nino;
9672 * If the bwrite fails we have no recourse to recover. The
9673 * filesystem is corrupted already.
9678 * If the superblock pointer still needs to be cleared force
9681 if (fs->fs_sujfree == ino) {
9683 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9684 (int)fs->fs_sbsize, 0, 0, 0);
9685 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9686 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9687 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9693 if (fs->fs_sujfree != ino)
9695 panic("clear_unlinked_inodedep: Failed to clear free head");
9697 if (inodedep->id_ino == fs->fs_sujfree)
9698 panic("clear_unlinked_inodedep: Freeing head of free list");
9699 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9700 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9705 * This workitem decrements the inode's link count.
9706 * If the link count reaches zero, the file is removed.
9709 handle_workitem_remove(dirrem, flags)
9710 struct dirrem *dirrem;
9713 struct inodedep *inodedep;
9714 struct workhead dotdotwk;
9715 struct worklist *wk;
9716 struct ufsmount *ump;
9722 if (dirrem->dm_state & ONWORKLIST)
9723 panic("handle_workitem_remove: dirrem %p still on worklist",
9725 oldinum = dirrem->dm_oldinum;
9726 mp = dirrem->dm_list.wk_mp;
9728 flags |= LK_EXCLUSIVE;
9729 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0)
9733 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
9734 panic("handle_workitem_remove: lost inodedep");
9735 if (dirrem->dm_state & ONDEPLIST)
9736 LIST_REMOVE(dirrem, dm_inonext);
9737 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
9738 ("handle_workitem_remove: Journal entries not written."));
9741 * Move all dependencies waiting on the remove to complete
9742 * from the dirrem to the inode inowait list to be completed
9743 * after the inode has been updated and written to disk. Any
9744 * marked MKDIR_PARENT are saved to be completed when the .. ref
9747 LIST_INIT(&dotdotwk);
9748 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
9749 WORKLIST_REMOVE(wk);
9750 if (wk->wk_state & MKDIR_PARENT) {
9751 wk->wk_state &= ~MKDIR_PARENT;
9752 WORKLIST_INSERT(&dotdotwk, wk);
9755 WORKLIST_INSERT(&inodedep->id_inowait, wk);
9757 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
9759 * Normal file deletion.
9761 if ((dirrem->dm_state & RMDIR) == 0) {
9763 DIP_SET(ip, i_nlink, ip->i_nlink);
9764 ip->i_flag |= IN_CHANGE;
9765 if (ip->i_nlink < ip->i_effnlink)
9766 panic("handle_workitem_remove: bad file delta");
9767 if (ip->i_nlink == 0)
9768 unlinked_inodedep(mp, inodedep);
9769 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9770 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9771 ("handle_workitem_remove: worklist not empty. %s",
9772 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
9773 WORKITEM_FREE(dirrem, D_DIRREM);
9778 * Directory deletion. Decrement reference count for both the
9779 * just deleted parent directory entry and the reference for ".".
9780 * Arrange to have the reference count on the parent decremented
9781 * to account for the loss of "..".
9784 DIP_SET(ip, i_nlink, ip->i_nlink);
9785 ip->i_flag |= IN_CHANGE;
9786 if (ip->i_nlink < ip->i_effnlink)
9787 panic("handle_workitem_remove: bad dir delta");
9788 if (ip->i_nlink == 0)
9789 unlinked_inodedep(mp, inodedep);
9790 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9792 * Rename a directory to a new parent. Since, we are both deleting
9793 * and creating a new directory entry, the link count on the new
9794 * directory should not change. Thus we skip the followup dirrem.
9796 if (dirrem->dm_state & DIRCHG) {
9797 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9798 ("handle_workitem_remove: DIRCHG and worklist not empty."));
9799 WORKITEM_FREE(dirrem, D_DIRREM);
9803 dirrem->dm_state = ONDEPLIST;
9804 dirrem->dm_oldinum = dirrem->dm_dirinum;
9806 * Place the dirrem on the parent's diremhd list.
9808 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
9809 panic("handle_workitem_remove: lost dir inodedep");
9810 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9812 * If the allocated inode has never been written to disk, then
9813 * the on-disk inode is zero'ed and we can remove the file
9814 * immediately. When journaling if the inode has been marked
9815 * unlinked and not DEPCOMPLETE we know it can never be written.
9817 inodedep_lookup(mp, oldinum, 0, &inodedep);
9818 if (inodedep == NULL ||
9819 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
9820 check_inode_unwritten(inodedep)) {
9823 return handle_workitem_remove(dirrem, flags);
9825 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
9827 ip->i_flag |= IN_CHANGE;
9835 * Inode de-allocation dependencies.
9837 * When an inode's link count is reduced to zero, it can be de-allocated. We
9838 * found it convenient to postpone de-allocation until after the inode is
9839 * written to disk with its new link count (zero). At this point, all of the
9840 * on-disk inode's block pointers are nullified and, with careful dependency
9841 * list ordering, all dependencies related to the inode will be satisfied and
9842 * the corresponding dependency structures de-allocated. So, if/when the
9843 * inode is reused, there will be no mixing of old dependencies with new
9844 * ones. This artificial dependency is set up by the block de-allocation
9845 * procedure above (softdep_setup_freeblocks) and completed by the
9846 * following procedure.
9849 handle_workitem_freefile(freefile)
9850 struct freefile *freefile;
9852 struct workhead wkhd;
9854 struct inodedep *idp;
9855 struct ufsmount *ump;
9858 ump = VFSTOUFS(freefile->fx_list.wk_mp);
9862 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
9865 panic("handle_workitem_freefile: inodedep %p survived", idp);
9868 fs->fs_pendinginodes -= 1;
9871 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
9872 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
9873 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
9874 softdep_error("handle_workitem_freefile", error);
9876 WORKITEM_FREE(freefile, D_FREEFILE);
9882 * Helper function which unlinks marker element from work list and returns
9883 * the next element on the list.
9885 static __inline struct worklist *
9886 markernext(struct worklist *marker)
9888 struct worklist *next;
9890 next = LIST_NEXT(marker, wk_list);
9891 LIST_REMOVE(marker, wk_list);
9898 * The dependency structures constructed above are most actively used when file
9899 * system blocks are written to disk. No constraints are placed on when a
9900 * block can be written, but unsatisfied update dependencies are made safe by
9901 * modifying (or replacing) the source memory for the duration of the disk
9902 * write. When the disk write completes, the memory block is again brought
9905 * In-core inode structure reclamation.
9907 * Because there are a finite number of "in-core" inode structures, they are
9908 * reused regularly. By transferring all inode-related dependencies to the
9909 * in-memory inode block and indexing them separately (via "inodedep"s), we
9910 * can allow "in-core" inode structures to be reused at any time and avoid
9911 * any increase in contention.
9913 * Called just before entering the device driver to initiate a new disk I/O.
9914 * The buffer must be locked, thus, no I/O completion operations can occur
9915 * while we are manipulating its associated dependencies.
9918 softdep_disk_io_initiation(bp)
9919 struct buf *bp; /* structure describing disk write to occur */
9921 struct worklist *wk;
9922 struct worklist marker;
9923 struct inodedep *inodedep;
9924 struct freeblks *freeblks;
9925 struct jblkdep *jblkdep;
9926 struct newblk *newblk;
9927 struct ufsmount *ump;
9930 * We only care about write operations. There should never
9931 * be dependencies for reads.
9933 if (bp->b_iocmd != BIO_WRITE)
9934 panic("softdep_disk_io_initiation: not write");
9936 if (bp->b_vflags & BV_BKGRDINPROG)
9937 panic("softdep_disk_io_initiation: Writing buffer with "
9938 "background write in progress: %p", bp);
9940 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
9942 ump = VFSTOUFS(wk->wk_mp);
9944 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
9945 PHOLD(curproc); /* Don't swap out kernel stack */
9948 * Do any necessary pre-I/O processing.
9950 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
9951 wk = markernext(&marker)) {
9952 LIST_INSERT_AFTER(wk, &marker, wk_list);
9953 switch (wk->wk_type) {
9956 initiate_write_filepage(WK_PAGEDEP(wk), bp);
9960 inodedep = WK_INODEDEP(wk);
9961 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
9962 initiate_write_inodeblock_ufs1(inodedep, bp);
9964 initiate_write_inodeblock_ufs2(inodedep, bp);
9968 initiate_write_indirdep(WK_INDIRDEP(wk), bp);
9972 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
9976 WK_JSEG(wk)->js_buf = NULL;
9980 freeblks = WK_FREEBLKS(wk);
9981 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd);
9983 * We have to wait for the freeblks to be journaled
9984 * before we can write an inodeblock with updated
9985 * pointers. Be careful to arrange the marker so
9986 * we revisit the freeblks if it's not removed by
9987 * the first jwait().
9989 if (jblkdep != NULL) {
9990 LIST_REMOVE(&marker, wk_list);
9991 LIST_INSERT_BEFORE(wk, &marker, wk_list);
9992 jwait(&jblkdep->jb_list, MNT_WAIT);
9998 * We have to wait for the jnewblk to be journaled
9999 * before we can write to a block if the contents
10000 * may be confused with an earlier file's indirect
10001 * at recovery time. Handle the marker as described
10004 newblk = WK_NEWBLK(wk);
10005 if (newblk->nb_jnewblk != NULL &&
10006 indirblk_lookup(newblk->nb_list.wk_mp,
10007 newblk->nb_newblkno)) {
10008 LIST_REMOVE(&marker, wk_list);
10009 LIST_INSERT_BEFORE(wk, &marker, wk_list);
10010 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
10015 initiate_write_sbdep(WK_SBDEP(wk));
10025 panic("handle_disk_io_initiation: Unexpected type %s",
10026 TYPENAME(wk->wk_type));
10031 PRELE(curproc); /* Allow swapout of kernel stack */
10035 * Called from within the procedure above to deal with unsatisfied
10036 * allocation dependencies in a directory. The buffer must be locked,
10037 * thus, no I/O completion operations can occur while we are
10038 * manipulating its associated dependencies.
10041 initiate_write_filepage(pagedep, bp)
10042 struct pagedep *pagedep;
10045 struct jremref *jremref;
10046 struct jmvref *jmvref;
10047 struct dirrem *dirrem;
10048 struct diradd *dap;
10052 if (pagedep->pd_state & IOSTARTED) {
10054 * This can only happen if there is a driver that does not
10055 * understand chaining. Here biodone will reissue the call
10056 * to strategy for the incomplete buffers.
10058 printf("initiate_write_filepage: already started\n");
10061 pagedep->pd_state |= IOSTARTED;
10063 * Wait for all journal remove dependencies to hit the disk.
10064 * We can not allow any potentially conflicting directory adds
10065 * to be visible before removes and rollback is too difficult.
10066 * The per-filesystem lock may be dropped and re-acquired, however
10067 * we hold the buf locked so the dependency can not go away.
10069 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
10070 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL)
10071 jwait(&jremref->jr_list, MNT_WAIT);
10072 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL)
10073 jwait(&jmvref->jm_list, MNT_WAIT);
10074 for (i = 0; i < DAHASHSZ; i++) {
10075 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
10076 ep = (struct direct *)
10077 ((char *)bp->b_data + dap->da_offset);
10078 if (ep->d_ino != dap->da_newinum)
10079 panic("%s: dir inum %ju != new %ju",
10080 "initiate_write_filepage",
10081 (uintmax_t)ep->d_ino,
10082 (uintmax_t)dap->da_newinum);
10083 if (dap->da_state & DIRCHG)
10084 ep->d_ino = dap->da_previous->dm_oldinum;
10087 dap->da_state &= ~ATTACHED;
10088 dap->da_state |= UNDONE;
10094 * Version of initiate_write_inodeblock that handles UFS1 dinodes.
10095 * Note that any bug fixes made to this routine must be done in the
10096 * version found below.
10098 * Called from within the procedure above to deal with unsatisfied
10099 * allocation dependencies in an inodeblock. The buffer must be
10100 * locked, thus, no I/O completion operations can occur while we
10101 * are manipulating its associated dependencies.
10104 initiate_write_inodeblock_ufs1(inodedep, bp)
10105 struct inodedep *inodedep;
10106 struct buf *bp; /* The inode block */
10108 struct allocdirect *adp, *lastadp;
10109 struct ufs1_dinode *dp;
10110 struct ufs1_dinode *sip;
10111 struct inoref *inoref;
10112 struct ufsmount *ump;
10116 ufs_lbn_t prevlbn = 0;
10120 if (inodedep->id_state & IOSTARTED)
10121 panic("initiate_write_inodeblock_ufs1: already started");
10122 inodedep->id_state |= IOSTARTED;
10123 fs = inodedep->id_fs;
10124 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10126 dp = (struct ufs1_dinode *)bp->b_data +
10127 ino_to_fsbo(fs, inodedep->id_ino);
10130 * If we're on the unlinked list but have not yet written our
10131 * next pointer initialize it here.
10133 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10134 struct inodedep *inon;
10136 inon = TAILQ_NEXT(inodedep, id_unlinked);
10137 dp->di_freelink = inon ? inon->id_ino : 0;
10140 * If the bitmap is not yet written, then the allocated
10141 * inode cannot be written to disk.
10143 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10144 if (inodedep->id_savedino1 != NULL)
10145 panic("initiate_write_inodeblock_ufs1: I/O underway");
10147 sip = malloc(sizeof(struct ufs1_dinode),
10148 M_SAVEDINO, M_SOFTDEP_FLAGS);
10150 inodedep->id_savedino1 = sip;
10151 *inodedep->id_savedino1 = *dp;
10152 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
10153 dp->di_gen = inodedep->id_savedino1->di_gen;
10154 dp->di_freelink = inodedep->id_savedino1->di_freelink;
10158 * If no dependencies, then there is nothing to roll back.
10160 inodedep->id_savedsize = dp->di_size;
10161 inodedep->id_savedextsize = 0;
10162 inodedep->id_savednlink = dp->di_nlink;
10163 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10164 TAILQ_EMPTY(&inodedep->id_inoreflst))
10167 * Revert the link count to that of the first unwritten journal entry.
10169 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10171 dp->di_nlink = inoref->if_nlink;
10173 * Set the dependencies to busy.
10175 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10176 adp = TAILQ_NEXT(adp, ad_next)) {
10178 if (deplist != 0 && prevlbn >= adp->ad_offset)
10179 panic("softdep_write_inodeblock: lbn order");
10180 prevlbn = adp->ad_offset;
10181 if (adp->ad_offset < NDADDR &&
10182 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10183 panic("%s: direct pointer #%jd mismatch %d != %jd",
10184 "softdep_write_inodeblock",
10185 (intmax_t)adp->ad_offset,
10186 dp->di_db[adp->ad_offset],
10187 (intmax_t)adp->ad_newblkno);
10188 if (adp->ad_offset >= NDADDR &&
10189 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10190 panic("%s: indirect pointer #%jd mismatch %d != %jd",
10191 "softdep_write_inodeblock",
10192 (intmax_t)adp->ad_offset - NDADDR,
10193 dp->di_ib[adp->ad_offset - NDADDR],
10194 (intmax_t)adp->ad_newblkno);
10195 deplist |= 1 << adp->ad_offset;
10196 if ((adp->ad_state & ATTACHED) == 0)
10197 panic("softdep_write_inodeblock: Unknown state 0x%x",
10199 #endif /* INVARIANTS */
10200 adp->ad_state &= ~ATTACHED;
10201 adp->ad_state |= UNDONE;
10204 * The on-disk inode cannot claim to be any larger than the last
10205 * fragment that has been written. Otherwise, the on-disk inode
10206 * might have fragments that were not the last block in the file
10207 * which would corrupt the filesystem.
10209 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10210 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10211 if (adp->ad_offset >= NDADDR)
10213 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10214 /* keep going until hitting a rollback to a frag */
10215 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10217 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10218 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10220 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10221 panic("softdep_write_inodeblock: lost dep1");
10222 #endif /* INVARIANTS */
10225 for (i = 0; i < NIADDR; i++) {
10227 if (dp->di_ib[i] != 0 &&
10228 (deplist & ((1 << NDADDR) << i)) == 0)
10229 panic("softdep_write_inodeblock: lost dep2");
10230 #endif /* INVARIANTS */
10236 * If we have zero'ed out the last allocated block of the file,
10237 * roll back the size to the last currently allocated block.
10238 * We know that this last allocated block is a full-sized as
10239 * we already checked for fragments in the loop above.
10241 if (lastadp != NULL &&
10242 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10243 for (i = lastadp->ad_offset; i >= 0; i--)
10244 if (dp->di_db[i] != 0)
10246 dp->di_size = (i + 1) * fs->fs_bsize;
10249 * The only dependencies are for indirect blocks.
10251 * The file size for indirect block additions is not guaranteed.
10252 * Such a guarantee would be non-trivial to achieve. The conventional
10253 * synchronous write implementation also does not make this guarantee.
10254 * Fsck should catch and fix discrepancies. Arguably, the file size
10255 * can be over-estimated without destroying integrity when the file
10256 * moves into the indirect blocks (i.e., is large). If we want to
10257 * postpone fsck, we are stuck with this argument.
10259 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10260 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10264 * Version of initiate_write_inodeblock that handles UFS2 dinodes.
10265 * Note that any bug fixes made to this routine must be done in the
10266 * version found above.
10268 * Called from within the procedure above to deal with unsatisfied
10269 * allocation dependencies in an inodeblock. The buffer must be
10270 * locked, thus, no I/O completion operations can occur while we
10271 * are manipulating its associated dependencies.
10274 initiate_write_inodeblock_ufs2(inodedep, bp)
10275 struct inodedep *inodedep;
10276 struct buf *bp; /* The inode block */
10278 struct allocdirect *adp, *lastadp;
10279 struct ufs2_dinode *dp;
10280 struct ufs2_dinode *sip;
10281 struct inoref *inoref;
10282 struct ufsmount *ump;
10286 ufs_lbn_t prevlbn = 0;
10290 if (inodedep->id_state & IOSTARTED)
10291 panic("initiate_write_inodeblock_ufs2: already started");
10292 inodedep->id_state |= IOSTARTED;
10293 fs = inodedep->id_fs;
10294 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10296 dp = (struct ufs2_dinode *)bp->b_data +
10297 ino_to_fsbo(fs, inodedep->id_ino);
10300 * If we're on the unlinked list but have not yet written our
10301 * next pointer initialize it here.
10303 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10304 struct inodedep *inon;
10306 inon = TAILQ_NEXT(inodedep, id_unlinked);
10307 dp->di_freelink = inon ? inon->id_ino : 0;
10310 * If the bitmap is not yet written, then the allocated
10311 * inode cannot be written to disk.
10313 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10314 if (inodedep->id_savedino2 != NULL)
10315 panic("initiate_write_inodeblock_ufs2: I/O underway");
10317 sip = malloc(sizeof(struct ufs2_dinode),
10318 M_SAVEDINO, M_SOFTDEP_FLAGS);
10320 inodedep->id_savedino2 = sip;
10321 *inodedep->id_savedino2 = *dp;
10322 bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
10323 dp->di_gen = inodedep->id_savedino2->di_gen;
10324 dp->di_freelink = inodedep->id_savedino2->di_freelink;
10328 * If no dependencies, then there is nothing to roll back.
10330 inodedep->id_savedsize = dp->di_size;
10331 inodedep->id_savedextsize = dp->di_extsize;
10332 inodedep->id_savednlink = dp->di_nlink;
10333 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10334 TAILQ_EMPTY(&inodedep->id_extupdt) &&
10335 TAILQ_EMPTY(&inodedep->id_inoreflst))
10338 * Revert the link count to that of the first unwritten journal entry.
10340 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10342 dp->di_nlink = inoref->if_nlink;
10345 * Set the ext data dependencies to busy.
10347 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10348 adp = TAILQ_NEXT(adp, ad_next)) {
10350 if (deplist != 0 && prevlbn >= adp->ad_offset)
10351 panic("softdep_write_inodeblock: lbn order");
10352 prevlbn = adp->ad_offset;
10353 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
10354 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10355 "softdep_write_inodeblock",
10356 (intmax_t)adp->ad_offset,
10357 (intmax_t)dp->di_extb[adp->ad_offset],
10358 (intmax_t)adp->ad_newblkno);
10359 deplist |= 1 << adp->ad_offset;
10360 if ((adp->ad_state & ATTACHED) == 0)
10361 panic("softdep_write_inodeblock: Unknown state 0x%x",
10363 #endif /* INVARIANTS */
10364 adp->ad_state &= ~ATTACHED;
10365 adp->ad_state |= UNDONE;
10368 * The on-disk inode cannot claim to be any larger than the last
10369 * fragment that has been written. Otherwise, the on-disk inode
10370 * might have fragments that were not the last block in the ext
10371 * data which would corrupt the filesystem.
10373 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10374 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10375 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
10376 /* keep going until hitting a rollback to a frag */
10377 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10379 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10380 for (i = adp->ad_offset + 1; i < NXADDR; i++) {
10382 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
10383 panic("softdep_write_inodeblock: lost dep1");
10384 #endif /* INVARIANTS */
10385 dp->di_extb[i] = 0;
10391 * If we have zero'ed out the last allocated block of the ext
10392 * data, roll back the size to the last currently allocated block.
10393 * We know that this last allocated block is a full-sized as
10394 * we already checked for fragments in the loop above.
10396 if (lastadp != NULL &&
10397 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10398 for (i = lastadp->ad_offset; i >= 0; i--)
10399 if (dp->di_extb[i] != 0)
10401 dp->di_extsize = (i + 1) * fs->fs_bsize;
10404 * Set the file data dependencies to busy.
10406 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10407 adp = TAILQ_NEXT(adp, ad_next)) {
10409 if (deplist != 0 && prevlbn >= adp->ad_offset)
10410 panic("softdep_write_inodeblock: lbn order");
10411 if ((adp->ad_state & ATTACHED) == 0)
10412 panic("inodedep %p and adp %p not attached", inodedep, adp);
10413 prevlbn = adp->ad_offset;
10414 if (adp->ad_offset < NDADDR &&
10415 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10416 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10417 "softdep_write_inodeblock",
10418 (intmax_t)adp->ad_offset,
10419 (intmax_t)dp->di_db[adp->ad_offset],
10420 (intmax_t)adp->ad_newblkno);
10421 if (adp->ad_offset >= NDADDR &&
10422 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10423 panic("%s indirect pointer #%jd mismatch %jd != %jd",
10424 "softdep_write_inodeblock:",
10425 (intmax_t)adp->ad_offset - NDADDR,
10426 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR],
10427 (intmax_t)adp->ad_newblkno);
10428 deplist |= 1 << adp->ad_offset;
10429 if ((adp->ad_state & ATTACHED) == 0)
10430 panic("softdep_write_inodeblock: Unknown state 0x%x",
10432 #endif /* INVARIANTS */
10433 adp->ad_state &= ~ATTACHED;
10434 adp->ad_state |= UNDONE;
10437 * The on-disk inode cannot claim to be any larger than the last
10438 * fragment that has been written. Otherwise, the on-disk inode
10439 * might have fragments that were not the last block in the file
10440 * which would corrupt the filesystem.
10442 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10443 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10444 if (adp->ad_offset >= NDADDR)
10446 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10447 /* keep going until hitting a rollback to a frag */
10448 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10450 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10451 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10453 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10454 panic("softdep_write_inodeblock: lost dep2");
10455 #endif /* INVARIANTS */
10458 for (i = 0; i < NIADDR; i++) {
10460 if (dp->di_ib[i] != 0 &&
10461 (deplist & ((1 << NDADDR) << i)) == 0)
10462 panic("softdep_write_inodeblock: lost dep3");
10463 #endif /* INVARIANTS */
10469 * If we have zero'ed out the last allocated block of the file,
10470 * roll back the size to the last currently allocated block.
10471 * We know that this last allocated block is a full-sized as
10472 * we already checked for fragments in the loop above.
10474 if (lastadp != NULL &&
10475 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10476 for (i = lastadp->ad_offset; i >= 0; i--)
10477 if (dp->di_db[i] != 0)
10479 dp->di_size = (i + 1) * fs->fs_bsize;
10482 * The only dependencies are for indirect blocks.
10484 * The file size for indirect block additions is not guaranteed.
10485 * Such a guarantee would be non-trivial to achieve. The conventional
10486 * synchronous write implementation also does not make this guarantee.
10487 * Fsck should catch and fix discrepancies. Arguably, the file size
10488 * can be over-estimated without destroying integrity when the file
10489 * moves into the indirect blocks (i.e., is large). If we want to
10490 * postpone fsck, we are stuck with this argument.
10492 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10493 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10497 * Cancel an indirdep as a result of truncation. Release all of the
10498 * children allocindirs and place their journal work on the appropriate
10502 cancel_indirdep(indirdep, bp, freeblks)
10503 struct indirdep *indirdep;
10505 struct freeblks *freeblks;
10507 struct allocindir *aip;
10510 * None of the indirect pointers will ever be visible,
10511 * so they can simply be tossed. GOINGAWAY ensures
10512 * that allocated pointers will be saved in the buffer
10513 * cache until they are freed. Note that they will
10514 * only be able to be found by their physical address
10515 * since the inode mapping the logical address will
10516 * be gone. The save buffer used for the safe copy
10517 * was allocated in setup_allocindir_phase2 using
10518 * the physical address so it could be used for this
10519 * purpose. Hence we swap the safe copy with the real
10520 * copy, allowing the safe copy to be freed and holding
10521 * on to the real copy for later use in indir_trunc.
10523 if (indirdep->ir_state & GOINGAWAY)
10524 panic("cancel_indirdep: already gone");
10525 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
10526 indirdep->ir_state |= DEPCOMPLETE;
10527 LIST_REMOVE(indirdep, ir_next);
10529 indirdep->ir_state |= GOINGAWAY;
10531 * Pass in bp for blocks still have journal writes
10532 * pending so we can cancel them on their own.
10534 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL)
10535 cancel_allocindir(aip, bp, freeblks, 0);
10536 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL)
10537 cancel_allocindir(aip, NULL, freeblks, 0);
10538 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL)
10539 cancel_allocindir(aip, NULL, freeblks, 0);
10540 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL)
10541 cancel_allocindir(aip, NULL, freeblks, 0);
10543 * If there are pending partial truncations we need to keep the
10544 * old block copy around until they complete. This is because
10545 * the current b_data is not a perfect superset of the available
10548 if (TAILQ_EMPTY(&indirdep->ir_trunc))
10549 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
10551 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10552 WORKLIST_REMOVE(&indirdep->ir_list);
10553 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
10554 indirdep->ir_bp = NULL;
10555 indirdep->ir_freeblks = freeblks;
10559 * Free an indirdep once it no longer has new pointers to track.
10562 free_indirdep(indirdep)
10563 struct indirdep *indirdep;
10566 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc),
10567 ("free_indirdep: Indir trunc list not empty."));
10568 KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
10569 ("free_indirdep: Complete head not empty."));
10570 KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
10571 ("free_indirdep: write head not empty."));
10572 KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
10573 ("free_indirdep: done head not empty."));
10574 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
10575 ("free_indirdep: deplist head not empty."));
10576 KASSERT((indirdep->ir_state & DEPCOMPLETE),
10577 ("free_indirdep: %p still on newblk list.", indirdep));
10578 KASSERT(indirdep->ir_saveddata == NULL,
10579 ("free_indirdep: %p still has saved data.", indirdep));
10580 if (indirdep->ir_state & ONWORKLIST)
10581 WORKLIST_REMOVE(&indirdep->ir_list);
10582 WORKITEM_FREE(indirdep, D_INDIRDEP);
10586 * Called before a write to an indirdep. This routine is responsible for
10587 * rolling back pointers to a safe state which includes only those
10588 * allocindirs which have been completed.
10591 initiate_write_indirdep(indirdep, bp)
10592 struct indirdep *indirdep;
10595 struct ufsmount *ump;
10597 indirdep->ir_state |= IOSTARTED;
10598 if (indirdep->ir_state & GOINGAWAY)
10599 panic("disk_io_initiation: indirdep gone");
10601 * If there are no remaining dependencies, this will be writing
10602 * the real pointers.
10604 if (LIST_EMPTY(&indirdep->ir_deplisthd) &&
10605 TAILQ_EMPTY(&indirdep->ir_trunc))
10608 * Replace up-to-date version with safe version.
10610 if (indirdep->ir_saveddata == NULL) {
10611 ump = VFSTOUFS(indirdep->ir_list.wk_mp);
10614 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
10618 indirdep->ir_state &= ~ATTACHED;
10619 indirdep->ir_state |= UNDONE;
10620 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10621 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
10626 * Called when an inode has been cleared in a cg bitmap. This finally
10627 * eliminates any canceled jaddrefs
10630 softdep_setup_inofree(mp, bp, ino, wkhd)
10634 struct workhead *wkhd;
10636 struct worklist *wk, *wkn;
10637 struct inodedep *inodedep;
10638 struct ufsmount *ump;
10643 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
10644 ("softdep_setup_inofree called on non-softdep filesystem"));
10645 ump = VFSTOUFS(mp);
10648 cgp = (struct cg *)bp->b_data;
10649 inosused = cg_inosused(cgp);
10650 if (isset(inosused, ino % fs->fs_ipg))
10651 panic("softdep_setup_inofree: inode %ju not freed.",
10653 if (inodedep_lookup(mp, ino, 0, &inodedep))
10654 panic("softdep_setup_inofree: ino %ju has existing inodedep %p",
10655 (uintmax_t)ino, inodedep);
10657 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
10658 if (wk->wk_type != D_JADDREF)
10660 WORKLIST_REMOVE(wk);
10662 * We can free immediately even if the jaddref
10663 * isn't attached in a background write as now
10664 * the bitmaps are reconciled.
10666 wk->wk_state |= COMPLETE | ATTACHED;
10667 free_jaddref(WK_JADDREF(wk));
10669 jwork_move(&bp->b_dep, wkhd);
10676 * Called via ffs_blkfree() after a set of frags has been cleared from a cg
10677 * map. Any dependencies waiting for the write to clear are added to the
10678 * buf's list and any jnewblks that are being canceled are discarded
10682 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
10685 ufs2_daddr_t blkno;
10687 struct workhead *wkhd;
10689 struct bmsafemap *bmsafemap;
10690 struct jnewblk *jnewblk;
10691 struct ufsmount *ump;
10692 struct worklist *wk;
10697 ufs2_daddr_t jstart;
10705 "softdep_setup_blkfree: blkno %jd frags %d wk head %p",
10706 blkno, frags, wkhd);
10708 ump = VFSTOUFS(mp);
10709 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
10710 ("softdep_setup_blkfree called on non-softdep filesystem"));
10712 /* Lookup the bmsafemap so we track when it is dirty. */
10714 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10716 * Detach any jnewblks which have been canceled. They must linger
10717 * until the bitmap is cleared again by ffs_blkfree() to prevent
10718 * an unjournaled allocation from hitting the disk.
10721 while ((wk = LIST_FIRST(wkhd)) != NULL) {
10723 "softdep_setup_blkfree: blkno %jd wk type %d",
10724 blkno, wk->wk_type);
10725 WORKLIST_REMOVE(wk);
10726 if (wk->wk_type != D_JNEWBLK) {
10727 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk);
10730 jnewblk = WK_JNEWBLK(wk);
10731 KASSERT(jnewblk->jn_state & GOINGAWAY,
10732 ("softdep_setup_blkfree: jnewblk not canceled."));
10735 * Assert that this block is free in the bitmap
10736 * before we discard the jnewblk.
10738 cgp = (struct cg *)bp->b_data;
10739 blksfree = cg_blksfree(cgp);
10740 bno = dtogd(fs, jnewblk->jn_blkno);
10741 for (i = jnewblk->jn_oldfrags;
10742 i < jnewblk->jn_frags; i++) {
10743 if (isset(blksfree, bno + i))
10745 panic("softdep_setup_blkfree: not free");
10749 * Even if it's not attached we can free immediately
10750 * as the new bitmap is correct.
10752 wk->wk_state |= COMPLETE | ATTACHED;
10753 free_jnewblk(jnewblk);
10759 * Assert that we are not freeing a block which has an outstanding
10760 * allocation dependency.
10762 fs = VFSTOUFS(mp)->um_fs;
10763 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10764 end = blkno + frags;
10765 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10767 * Don't match against blocks that will be freed when the
10768 * background write is done.
10770 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
10771 (COMPLETE | DEPCOMPLETE))
10773 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
10774 jend = jnewblk->jn_blkno + jnewblk->jn_frags;
10775 if ((blkno >= jstart && blkno < jend) ||
10776 (end > jstart && end <= jend)) {
10777 printf("state 0x%X %jd - %d %d dep %p\n",
10778 jnewblk->jn_state, jnewblk->jn_blkno,
10779 jnewblk->jn_oldfrags, jnewblk->jn_frags,
10781 panic("softdep_setup_blkfree: "
10782 "%jd-%jd(%d) overlaps with %jd-%jd",
10783 blkno, end, frags, jstart, jend);
10791 * Revert a block allocation when the journal record that describes it
10792 * is not yet written.
10795 jnewblk_rollback(jnewblk, fs, cgp, blksfree)
10796 struct jnewblk *jnewblk;
10801 ufs1_daddr_t fragno;
10807 cgbno = dtogd(fs, jnewblk->jn_blkno);
10809 * We have to test which frags need to be rolled back. We may
10810 * be operating on a stale copy when doing background writes.
10812 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++)
10813 if (isclr(blksfree, cgbno + i))
10818 * This is mostly ffs_blkfree() sans some validation and
10819 * superblock updates.
10821 if (frags == fs->fs_frag) {
10822 fragno = fragstoblks(fs, cgbno);
10823 ffs_setblock(fs, blksfree, fragno);
10824 ffs_clusteracct(fs, cgp, fragno, 1);
10825 cgp->cg_cs.cs_nbfree++;
10827 cgbno += jnewblk->jn_oldfrags;
10828 bbase = cgbno - fragnum(fs, cgbno);
10829 /* Decrement the old frags. */
10830 blk = blkmap(fs, blksfree, bbase);
10831 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
10832 /* Deallocate the fragment */
10833 for (i = 0; i < frags; i++)
10834 setbit(blksfree, cgbno + i);
10835 cgp->cg_cs.cs_nffree += frags;
10836 /* Add back in counts associated with the new frags */
10837 blk = blkmap(fs, blksfree, bbase);
10838 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
10839 /* If a complete block has been reassembled, account for it. */
10840 fragno = fragstoblks(fs, bbase);
10841 if (ffs_isblock(fs, blksfree, fragno)) {
10842 cgp->cg_cs.cs_nffree -= fs->fs_frag;
10843 ffs_clusteracct(fs, cgp, fragno, 1);
10844 cgp->cg_cs.cs_nbfree++;
10848 jnewblk->jn_state &= ~ATTACHED;
10849 jnewblk->jn_state |= UNDONE;
10855 initiate_write_bmsafemap(bmsafemap, bp)
10856 struct bmsafemap *bmsafemap;
10857 struct buf *bp; /* The cg block. */
10859 struct jaddref *jaddref;
10860 struct jnewblk *jnewblk;
10868 * If this is a background write, we did this at the time that
10869 * the copy was made, so do not need to do it again.
10871 if (bmsafemap->sm_state & IOSTARTED)
10873 bmsafemap->sm_state |= IOSTARTED;
10875 * Clear any inode allocations which are pending journal writes.
10877 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) {
10878 cgp = (struct cg *)bp->b_data;
10879 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10880 inosused = cg_inosused(cgp);
10881 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) {
10882 ino = jaddref->ja_ino % fs->fs_ipg;
10883 if (isset(inosused, ino)) {
10884 if ((jaddref->ja_mode & IFMT) == IFDIR)
10885 cgp->cg_cs.cs_ndir--;
10886 cgp->cg_cs.cs_nifree++;
10887 clrbit(inosused, ino);
10888 jaddref->ja_state &= ~ATTACHED;
10889 jaddref->ja_state |= UNDONE;
10892 panic("initiate_write_bmsafemap: inode %ju "
10893 "marked free", (uintmax_t)jaddref->ja_ino);
10897 * Clear any block allocations which are pending journal writes.
10899 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
10900 cgp = (struct cg *)bp->b_data;
10901 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10902 blksfree = cg_blksfree(cgp);
10903 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10904 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree))
10906 panic("initiate_write_bmsafemap: block %jd "
10907 "marked free", jnewblk->jn_blkno);
10911 * Move allocation lists to the written lists so they can be
10912 * cleared once the block write is complete.
10914 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr,
10915 inodedep, id_deps);
10916 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
10918 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist,
10923 * This routine is called during the completion interrupt
10924 * service routine for a disk write (from the procedure called
10925 * by the device driver to inform the filesystem caches of
10926 * a request completion). It should be called early in this
10927 * procedure, before the block is made available to other
10928 * processes or other routines are called.
10932 softdep_disk_write_complete(bp)
10933 struct buf *bp; /* describes the completed disk write */
10935 struct worklist *wk;
10936 struct worklist *owk;
10937 struct ufsmount *ump;
10938 struct workhead reattach;
10939 struct freeblks *freeblks;
10943 * If an error occurred while doing the write, then the data
10944 * has not hit the disk and the dependencies cannot be processed.
10945 * But we do have to go through and roll forward any dependencies
10946 * that were rolled back before the disk write.
10948 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) {
10949 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
10950 switch (wk->wk_type) {
10953 handle_written_filepage(WK_PAGEDEP(wk), bp, 0);
10957 handle_written_inodeblock(WK_INODEDEP(wk),
10962 handle_written_bmsafemap(WK_BMSAFEMAP(wk),
10967 handle_written_indirdep(WK_INDIRDEP(wk),
10971 /* nothing to roll forward */
10977 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
10979 ump = VFSTOUFS(wk->wk_mp);
10980 LIST_INIT(&reattach);
10982 * This lock must not be released anywhere in this code segment.
10987 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
10988 WORKLIST_REMOVE(wk);
10989 atomic_add_long(&dep_write[wk->wk_type], 1);
10991 panic("duplicate worklist: %p\n", wk);
10993 switch (wk->wk_type) {
10996 if (handle_written_filepage(WK_PAGEDEP(wk), bp,
10998 WORKLIST_INSERT(&reattach, wk);
11002 if (handle_written_inodeblock(WK_INODEDEP(wk), bp,
11004 WORKLIST_INSERT(&reattach, wk);
11008 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp,
11010 WORKLIST_INSERT(&reattach, wk);
11014 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
11017 case D_ALLOCDIRECT:
11018 wk->wk_state |= COMPLETE;
11019 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
11023 wk->wk_state |= COMPLETE;
11024 handle_allocindir_partdone(WK_ALLOCINDIR(wk));
11028 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp,
11030 WORKLIST_INSERT(&reattach, wk);
11034 wk->wk_state |= COMPLETE;
11035 freeblks = WK_FREEBLKS(wk);
11036 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE &&
11037 LIST_EMPTY(&freeblks->fb_jblkdephd))
11038 add_to_worklist(wk, WK_NODELAY);
11042 handle_written_freework(WK_FREEWORK(wk));
11046 free_jsegdep(WK_JSEGDEP(wk));
11050 handle_written_jseg(WK_JSEG(wk), bp);
11054 if (handle_written_sbdep(WK_SBDEP(wk), bp))
11055 WORKLIST_INSERT(&reattach, wk);
11059 free_freedep(WK_FREEDEP(wk));
11063 panic("handle_disk_write_complete: Unknown type %s",
11064 TYPENAME(wk->wk_type));
11069 * Reattach any requests that must be redone.
11071 while ((wk = LIST_FIRST(&reattach)) != NULL) {
11072 WORKLIST_REMOVE(wk);
11073 WORKLIST_INSERT(&bp->b_dep, wk);
11081 * Called from within softdep_disk_write_complete above. Note that
11082 * this routine is always called from interrupt level with further
11083 * splbio interrupts blocked.
11086 handle_allocdirect_partdone(adp, wkhd)
11087 struct allocdirect *adp; /* the completed allocdirect */
11088 struct workhead *wkhd; /* Work to do when inode is writtne. */
11090 struct allocdirectlst *listhead;
11091 struct allocdirect *listadp;
11092 struct inodedep *inodedep;
11095 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11098 * The on-disk inode cannot claim to be any larger than the last
11099 * fragment that has been written. Otherwise, the on-disk inode
11100 * might have fragments that were not the last block in the file
11101 * which would corrupt the filesystem. Thus, we cannot free any
11102 * allocdirects after one whose ad_oldblkno claims a fragment as
11103 * these blocks must be rolled back to zero before writing the inode.
11104 * We check the currently active set of allocdirects in id_inoupdt
11105 * or id_extupdt as appropriate.
11107 inodedep = adp->ad_inodedep;
11108 bsize = inodedep->id_fs->fs_bsize;
11109 if (adp->ad_state & EXTDATA)
11110 listhead = &inodedep->id_extupdt;
11112 listhead = &inodedep->id_inoupdt;
11113 TAILQ_FOREACH(listadp, listhead, ad_next) {
11114 /* found our block */
11115 if (listadp == adp)
11117 /* continue if ad_oldlbn is not a fragment */
11118 if (listadp->ad_oldsize == 0 ||
11119 listadp->ad_oldsize == bsize)
11121 /* hit a fragment */
11125 * If we have reached the end of the current list without
11126 * finding the just finished dependency, then it must be
11127 * on the future dependency list. Future dependencies cannot
11128 * be freed until they are moved to the current list.
11130 if (listadp == NULL) {
11132 if (adp->ad_state & EXTDATA)
11133 listhead = &inodedep->id_newextupdt;
11135 listhead = &inodedep->id_newinoupdt;
11136 TAILQ_FOREACH(listadp, listhead, ad_next)
11137 /* found our block */
11138 if (listadp == adp)
11140 if (listadp == NULL)
11141 panic("handle_allocdirect_partdone: lost dep");
11146 * If we have found the just finished dependency, then queue
11147 * it along with anything that follows it that is complete.
11148 * Since the pointer has not yet been written in the inode
11149 * as the dependency prevents it, place the allocdirect on the
11150 * bufwait list where it will be freed once the pointer is
11154 wkhd = &inodedep->id_bufwait;
11155 for (; adp; adp = listadp) {
11156 listadp = TAILQ_NEXT(adp, ad_next);
11157 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11159 TAILQ_REMOVE(listhead, adp, ad_next);
11160 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
11165 * Called from within softdep_disk_write_complete above. This routine
11166 * completes successfully written allocindirs.
11169 handle_allocindir_partdone(aip)
11170 struct allocindir *aip; /* the completed allocindir */
11172 struct indirdep *indirdep;
11174 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
11176 indirdep = aip->ai_indirdep;
11177 LIST_REMOVE(aip, ai_next);
11179 * Don't set a pointer while the buffer is undergoing IO or while
11180 * we have active truncations.
11182 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) {
11183 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
11186 if (indirdep->ir_state & UFS1FMT)
11187 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11190 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11193 * Await the pointer write before freeing the allocindir.
11195 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
11199 * Release segments held on a jwork list.
11203 struct workhead *wkhd;
11205 struct worklist *wk;
11207 while ((wk = LIST_FIRST(wkhd)) != NULL) {
11208 WORKLIST_REMOVE(wk);
11209 switch (wk->wk_type) {
11211 free_jsegdep(WK_JSEGDEP(wk));
11214 free_freedep(WK_FREEDEP(wk));
11217 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep));
11218 WORKITEM_FREE(wk, D_FREEFRAG);
11221 handle_written_freework(WK_FREEWORK(wk));
11224 panic("handle_jwork: Unknown type %s\n",
11225 TYPENAME(wk->wk_type));
11231 * Handle the bufwait list on an inode when it is safe to release items
11232 * held there. This normally happens after an inode block is written but
11233 * may be delayed and handled later if there are pending journal items that
11234 * are not yet safe to be released.
11236 static struct freefile *
11237 handle_bufwait(inodedep, refhd)
11238 struct inodedep *inodedep;
11239 struct workhead *refhd;
11241 struct jaddref *jaddref;
11242 struct freefile *freefile;
11243 struct worklist *wk;
11246 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
11247 WORKLIST_REMOVE(wk);
11248 switch (wk->wk_type) {
11251 * We defer adding freefile to the worklist
11252 * until all other additions have been made to
11253 * ensure that it will be done after all the
11254 * old blocks have been freed.
11256 if (freefile != NULL)
11257 panic("handle_bufwait: freefile");
11258 freefile = WK_FREEFILE(wk);
11262 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
11266 diradd_inode_written(WK_DIRADD(wk), inodedep);
11270 wk->wk_state |= COMPLETE;
11271 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
11272 add_to_worklist(wk, 0);
11276 wk->wk_state |= COMPLETE;
11277 add_to_worklist(wk, 0);
11280 case D_ALLOCDIRECT:
11282 free_newblk(WK_NEWBLK(wk));
11286 wk->wk_state |= COMPLETE;
11287 free_jnewblk(WK_JNEWBLK(wk));
11291 * Save freed journal segments and add references on
11292 * the supplied list which will delay their release
11293 * until the cg bitmap is cleared on disk.
11297 free_jsegdep(WK_JSEGDEP(wk));
11299 WORKLIST_INSERT(refhd, wk);
11303 jaddref = WK_JADDREF(wk);
11304 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
11307 * Transfer any jaddrefs to the list to be freed with
11308 * the bitmap if we're handling a removed file.
11310 if (refhd == NULL) {
11311 wk->wk_state |= COMPLETE;
11312 free_jaddref(jaddref);
11314 WORKLIST_INSERT(refhd, wk);
11318 panic("handle_bufwait: Unknown type %p(%s)",
11319 wk, TYPENAME(wk->wk_type));
11326 * Called from within softdep_disk_write_complete above to restore
11327 * in-memory inode block contents to their most up-to-date state. Note
11328 * that this routine is always called from interrupt level with further
11329 * interrupts from this device blocked.
11331 * If the write did not succeed, we will do all the roll-forward
11332 * operations, but we will not take the actions that will allow its
11333 * dependencies to be processed.
11336 handle_written_inodeblock(inodedep, bp, flags)
11337 struct inodedep *inodedep;
11338 struct buf *bp; /* buffer containing the inode block */
11341 struct freefile *freefile;
11342 struct allocdirect *adp, *nextadp;
11343 struct ufs1_dinode *dp1 = NULL;
11344 struct ufs2_dinode *dp2 = NULL;
11345 struct workhead wkhd;
11346 int hadchanges, fstype;
11352 if ((inodedep->id_state & IOSTARTED) == 0)
11353 panic("handle_written_inodeblock: not started");
11354 inodedep->id_state &= ~IOSTARTED;
11355 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
11357 dp1 = (struct ufs1_dinode *)bp->b_data +
11358 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11359 freelink = dp1->di_freelink;
11362 dp2 = (struct ufs2_dinode *)bp->b_data +
11363 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11364 freelink = dp2->di_freelink;
11367 * Leave this inodeblock dirty until it's in the list.
11369 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED &&
11370 (flags & WRITESUCCEEDED)) {
11371 struct inodedep *inon;
11373 inon = TAILQ_NEXT(inodedep, id_unlinked);
11374 if ((inon == NULL && freelink == 0) ||
11375 (inon && inon->id_ino == freelink)) {
11377 inon->id_state |= UNLINKPREV;
11378 inodedep->id_state |= UNLINKNEXT;
11383 * If we had to rollback the inode allocation because of
11384 * bitmaps being incomplete, then simply restore it.
11385 * Keep the block dirty so that it will not be reclaimed until
11386 * all associated dependencies have been cleared and the
11387 * corresponding updates written to disk.
11389 if (inodedep->id_savedino1 != NULL) {
11391 if (fstype == UFS1)
11392 *dp1 = *inodedep->id_savedino1;
11394 *dp2 = *inodedep->id_savedino2;
11395 free(inodedep->id_savedino1, M_SAVEDINO);
11396 inodedep->id_savedino1 = NULL;
11397 if ((bp->b_flags & B_DELWRI) == 0)
11398 stat_inode_bitmap++;
11401 * If the inode is clear here and GOINGAWAY it will never
11402 * be written. Process the bufwait and clear any pending
11403 * work which may include the freefile.
11405 if (inodedep->id_state & GOINGAWAY)
11409 if (flags & WRITESUCCEEDED)
11410 inodedep->id_state |= COMPLETE;
11412 * Roll forward anything that had to be rolled back before
11413 * the inode could be updated.
11415 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
11416 nextadp = TAILQ_NEXT(adp, ad_next);
11417 if (adp->ad_state & ATTACHED)
11418 panic("handle_written_inodeblock: new entry");
11419 if (fstype == UFS1) {
11420 if (adp->ad_offset < NDADDR) {
11421 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11422 panic("%s %s #%jd mismatch %d != %jd",
11423 "handle_written_inodeblock:",
11425 (intmax_t)adp->ad_offset,
11426 dp1->di_db[adp->ad_offset],
11427 (intmax_t)adp->ad_oldblkno);
11428 dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
11430 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0)
11431 panic("%s: %s #%jd allocated as %d",
11432 "handle_written_inodeblock",
11433 "indirect pointer",
11434 (intmax_t)adp->ad_offset - NDADDR,
11435 dp1->di_ib[adp->ad_offset - NDADDR]);
11436 dp1->di_ib[adp->ad_offset - NDADDR] =
11440 if (adp->ad_offset < NDADDR) {
11441 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11442 panic("%s: %s #%jd %s %jd != %jd",
11443 "handle_written_inodeblock",
11445 (intmax_t)adp->ad_offset, "mismatch",
11446 (intmax_t)dp2->di_db[adp->ad_offset],
11447 (intmax_t)adp->ad_oldblkno);
11448 dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
11450 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0)
11451 panic("%s: %s #%jd allocated as %jd",
11452 "handle_written_inodeblock",
11453 "indirect pointer",
11454 (intmax_t)adp->ad_offset - NDADDR,
11456 dp2->di_ib[adp->ad_offset - NDADDR]);
11457 dp2->di_ib[adp->ad_offset - NDADDR] =
11461 adp->ad_state &= ~UNDONE;
11462 adp->ad_state |= ATTACHED;
11465 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
11466 nextadp = TAILQ_NEXT(adp, ad_next);
11467 if (adp->ad_state & ATTACHED)
11468 panic("handle_written_inodeblock: new entry");
11469 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
11470 panic("%s: direct pointers #%jd %s %jd != %jd",
11471 "handle_written_inodeblock",
11472 (intmax_t)adp->ad_offset, "mismatch",
11473 (intmax_t)dp2->di_extb[adp->ad_offset],
11474 (intmax_t)adp->ad_oldblkno);
11475 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
11476 adp->ad_state &= ~UNDONE;
11477 adp->ad_state |= ATTACHED;
11480 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
11481 stat_direct_blk_ptrs++;
11483 * Reset the file size to its most up-to-date value.
11485 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
11486 panic("handle_written_inodeblock: bad size");
11487 if (inodedep->id_savednlink > LINK_MAX)
11488 panic("handle_written_inodeblock: Invalid link count "
11489 "%d for inodedep %p", inodedep->id_savednlink, inodedep);
11490 if (fstype == UFS1) {
11491 if (dp1->di_nlink != inodedep->id_savednlink) {
11492 dp1->di_nlink = inodedep->id_savednlink;
11495 if (dp1->di_size != inodedep->id_savedsize) {
11496 dp1->di_size = inodedep->id_savedsize;
11500 if (dp2->di_nlink != inodedep->id_savednlink) {
11501 dp2->di_nlink = inodedep->id_savednlink;
11504 if (dp2->di_size != inodedep->id_savedsize) {
11505 dp2->di_size = inodedep->id_savedsize;
11508 if (dp2->di_extsize != inodedep->id_savedextsize) {
11509 dp2->di_extsize = inodedep->id_savedextsize;
11513 inodedep->id_savedsize = -1;
11514 inodedep->id_savedextsize = -1;
11515 inodedep->id_savednlink = -1;
11517 * If there were any rollbacks in the inode block, then it must be
11518 * marked dirty so that its will eventually get written back in
11519 * its correct form.
11525 * If the write did not succeed, we have done all the roll-forward
11526 * operations, but we cannot take the actions that will allow its
11527 * dependencies to be processed.
11529 if ((flags & WRITESUCCEEDED) == 0)
11530 return (hadchanges);
11532 * Process any allocdirects that completed during the update.
11534 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
11535 handle_allocdirect_partdone(adp, &wkhd);
11536 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
11537 handle_allocdirect_partdone(adp, &wkhd);
11539 * Process deallocations that were held pending until the
11540 * inode had been written to disk. Freeing of the inode
11541 * is delayed until after all blocks have been freed to
11542 * avoid creation of new <vfsid, inum, lbn> triples
11543 * before the old ones have been deleted. Completely
11544 * unlinked inodes are not processed until the unlinked
11545 * inode list is written or the last reference is removed.
11547 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
11548 freefile = handle_bufwait(inodedep, NULL);
11549 if (freefile && !LIST_EMPTY(&wkhd)) {
11550 WORKLIST_INSERT(&wkhd, &freefile->fx_list);
11555 * Move rolled forward dependency completions to the bufwait list
11556 * now that those that were already written have been processed.
11558 if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
11559 panic("handle_written_inodeblock: bufwait but no changes");
11560 jwork_move(&inodedep->id_bufwait, &wkhd);
11562 if (freefile != NULL) {
11564 * If the inode is goingaway it was never written. Fake up
11565 * the state here so free_inodedep() can succeed.
11567 if (inodedep->id_state & GOINGAWAY)
11568 inodedep->id_state |= COMPLETE | DEPCOMPLETE;
11569 if (free_inodedep(inodedep) == 0)
11570 panic("handle_written_inodeblock: live inodedep %p",
11572 add_to_worklist(&freefile->fx_list, 0);
11577 * If no outstanding dependencies, free it.
11579 if (free_inodedep(inodedep) ||
11580 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
11581 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
11582 TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
11583 LIST_FIRST(&inodedep->id_bufwait) == 0))
11585 return (hadchanges);
11589 * Perform needed roll-forwards and kick off any dependencies that
11590 * can now be processed.
11592 * If the write did not succeed, we will do all the roll-forward
11593 * operations, but we will not take the actions that will allow its
11594 * dependencies to be processed.
11597 handle_written_indirdep(indirdep, bp, bpp, flags)
11598 struct indirdep *indirdep;
11603 struct allocindir *aip;
11607 if (indirdep->ir_state & GOINGAWAY)
11608 panic("handle_written_indirdep: indirdep gone");
11609 if ((indirdep->ir_state & IOSTARTED) == 0)
11610 panic("handle_written_indirdep: IO not started");
11613 * If there were rollbacks revert them here.
11615 if (indirdep->ir_saveddata) {
11616 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
11617 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11618 free(indirdep->ir_saveddata, M_INDIRDEP);
11619 indirdep->ir_saveddata = NULL;
11623 indirdep->ir_state &= ~(UNDONE | IOSTARTED);
11624 indirdep->ir_state |= ATTACHED;
11626 * If the write did not succeed, we have done all the roll-forward
11627 * operations, but we cannot take the actions that will allow its
11628 * dependencies to be processed.
11630 if ((flags & WRITESUCCEEDED) == 0) {
11631 stat_indir_blk_ptrs++;
11636 * Move allocindirs with written pointers to the completehd if
11637 * the indirdep's pointer is not yet written. Otherwise
11640 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) {
11641 LIST_REMOVE(aip, ai_next);
11642 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
11643 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
11645 newblk_freefrag(&aip->ai_block);
11648 free_newblk(&aip->ai_block);
11651 * Move allocindirs that have finished dependency processing from
11652 * the done list to the write list after updating the pointers.
11654 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11655 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) {
11656 handle_allocindir_partdone(aip);
11657 if (aip == LIST_FIRST(&indirdep->ir_donehd))
11658 panic("disk_write_complete: not gone");
11663 * Preserve the indirdep if there were any changes or if it is not
11664 * yet valid on disk.
11667 stat_indir_blk_ptrs++;
11672 * If there were no changes we can discard the savedbp and detach
11673 * ourselves from the buf. We are only carrying completed pointers
11676 sbp = indirdep->ir_savebp;
11677 sbp->b_flags |= B_INVAL | B_NOCACHE;
11678 indirdep->ir_savebp = NULL;
11679 indirdep->ir_bp = NULL;
11681 panic("handle_written_indirdep: bp already exists.");
11684 * The indirdep may not be freed until its parent points at it.
11686 if (indirdep->ir_state & DEPCOMPLETE)
11687 free_indirdep(indirdep);
11693 * Process a diradd entry after its dependent inode has been written.
11694 * This routine must be called with splbio interrupts blocked.
11697 diradd_inode_written(dap, inodedep)
11698 struct diradd *dap;
11699 struct inodedep *inodedep;
11702 dap->da_state |= COMPLETE;
11703 complete_diradd(dap);
11704 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
11708 * Returns true if the bmsafemap will have rollbacks when written. Must only
11709 * be called with the per-filesystem lock and the buf lock on the cg held.
11712 bmsafemap_backgroundwrite(bmsafemap, bp)
11713 struct bmsafemap *bmsafemap;
11718 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp));
11719 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) |
11720 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd);
11722 * If we're initiating a background write we need to process the
11723 * rollbacks as they exist now, not as they exist when IO starts.
11724 * No other consumers will look at the contents of the shadowed
11725 * buf so this is safe to do here.
11727 if (bp->b_xflags & BX_BKGRDMARKER)
11728 initiate_write_bmsafemap(bmsafemap, bp);
11734 * Re-apply an allocation when a cg write is complete.
11737 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)
11738 struct jnewblk *jnewblk;
11743 ufs1_daddr_t fragno;
11744 ufs2_daddr_t blkno;
11750 cgbno = dtogd(fs, jnewblk->jn_blkno);
11751 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) {
11752 if (isclr(blksfree, cgbno + i))
11753 panic("jnewblk_rollforward: re-allocated fragment");
11756 if (frags == fs->fs_frag) {
11757 blkno = fragstoblks(fs, cgbno);
11758 ffs_clrblock(fs, blksfree, (long)blkno);
11759 ffs_clusteracct(fs, cgp, blkno, -1);
11760 cgp->cg_cs.cs_nbfree--;
11762 bbase = cgbno - fragnum(fs, cgbno);
11763 cgbno += jnewblk->jn_oldfrags;
11764 /* If a complete block had been reassembled, account for it. */
11765 fragno = fragstoblks(fs, bbase);
11766 if (ffs_isblock(fs, blksfree, fragno)) {
11767 cgp->cg_cs.cs_nffree += fs->fs_frag;
11768 ffs_clusteracct(fs, cgp, fragno, -1);
11769 cgp->cg_cs.cs_nbfree--;
11771 /* Decrement the old frags. */
11772 blk = blkmap(fs, blksfree, bbase);
11773 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
11774 /* Allocate the fragment */
11775 for (i = 0; i < frags; i++)
11776 clrbit(blksfree, cgbno + i);
11777 cgp->cg_cs.cs_nffree -= frags;
11778 /* Add back in counts associated with the new frags */
11779 blk = blkmap(fs, blksfree, bbase);
11780 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
11786 * Complete a write to a bmsafemap structure. Roll forward any bitmap
11787 * changes if it's not a background write. Set all written dependencies
11788 * to DEPCOMPLETE and free the structure if possible.
11790 * If the write did not succeed, we will do all the roll-forward
11791 * operations, but we will not take the actions that will allow its
11792 * dependencies to be processed.
11795 handle_written_bmsafemap(bmsafemap, bp, flags)
11796 struct bmsafemap *bmsafemap;
11800 struct newblk *newblk;
11801 struct inodedep *inodedep;
11802 struct jaddref *jaddref, *jatmp;
11803 struct jnewblk *jnewblk, *jntmp;
11804 struct ufsmount *ump;
11813 if ((bmsafemap->sm_state & IOSTARTED) == 0)
11814 panic("handle_written_bmsafemap: Not started\n");
11815 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp);
11817 bmsafemap->sm_state &= ~IOSTARTED;
11818 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0;
11820 * If write was successful, release journal work that was waiting
11821 * on the write. Otherwise move the work back.
11823 if (flags & WRITESUCCEEDED)
11824 handle_jwork(&bmsafemap->sm_freewr);
11826 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr,
11827 worklist, wk_list);
11830 * Restore unwritten inode allocation pending jaddref writes.
11832 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
11833 cgp = (struct cg *)bp->b_data;
11834 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11835 inosused = cg_inosused(cgp);
11836 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
11837 ja_bmdeps, jatmp) {
11838 if ((jaddref->ja_state & UNDONE) == 0)
11840 ino = jaddref->ja_ino % fs->fs_ipg;
11841 if (isset(inosused, ino))
11842 panic("handle_written_bmsafemap: "
11843 "re-allocated inode");
11844 /* Do the roll-forward only if it's a real copy. */
11846 if ((jaddref->ja_mode & IFMT) == IFDIR)
11847 cgp->cg_cs.cs_ndir++;
11848 cgp->cg_cs.cs_nifree--;
11849 setbit(inosused, ino);
11852 jaddref->ja_state &= ~UNDONE;
11853 jaddref->ja_state |= ATTACHED;
11854 free_jaddref(jaddref);
11858 * Restore any block allocations which are pending journal writes.
11860 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
11861 cgp = (struct cg *)bp->b_data;
11862 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11863 blksfree = cg_blksfree(cgp);
11864 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
11866 if ((jnewblk->jn_state & UNDONE) == 0)
11868 /* Do the roll-forward only if it's a real copy. */
11870 jnewblk_rollforward(jnewblk, fs, cgp, blksfree))
11872 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
11873 jnewblk->jn_state |= ATTACHED;
11874 free_jnewblk(jnewblk);
11878 * If the write did not succeed, we have done all the roll-forward
11879 * operations, but we cannot take the actions that will allow its
11880 * dependencies to be processed.
11882 if ((flags & WRITESUCCEEDED) == 0) {
11883 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
11885 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr,
11886 worklist, wk_list);
11891 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
11892 newblk->nb_state |= DEPCOMPLETE;
11893 newblk->nb_state &= ~ONDEPLIST;
11894 newblk->nb_bmsafemap = NULL;
11895 LIST_REMOVE(newblk, nb_deps);
11896 if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
11897 handle_allocdirect_partdone(
11898 WK_ALLOCDIRECT(&newblk->nb_list), NULL);
11899 else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
11900 handle_allocindir_partdone(
11901 WK_ALLOCINDIR(&newblk->nb_list));
11902 else if (newblk->nb_list.wk_type != D_NEWBLK)
11903 panic("handle_written_bmsafemap: Unexpected type: %s",
11904 TYPENAME(newblk->nb_list.wk_type));
11906 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
11907 inodedep->id_state |= DEPCOMPLETE;
11908 inodedep->id_state &= ~ONDEPLIST;
11909 LIST_REMOVE(inodedep, id_deps);
11910 inodedep->id_bmsafemap = NULL;
11912 LIST_REMOVE(bmsafemap, sm_next);
11913 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
11914 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
11915 LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
11916 LIST_EMPTY(&bmsafemap->sm_inodedephd) &&
11917 LIST_EMPTY(&bmsafemap->sm_freehd)) {
11918 LIST_REMOVE(bmsafemap, sm_hash);
11919 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
11922 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
11929 * Try to free a mkdir dependency.
11932 complete_mkdir(mkdir)
11933 struct mkdir *mkdir;
11935 struct diradd *dap;
11937 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
11939 LIST_REMOVE(mkdir, md_mkdirs);
11940 dap = mkdir->md_diradd;
11941 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
11942 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
11943 dap->da_state |= DEPCOMPLETE;
11944 complete_diradd(dap);
11946 WORKITEM_FREE(mkdir, D_MKDIR);
11950 * Handle the completion of a mkdir dependency.
11953 handle_written_mkdir(mkdir, type)
11954 struct mkdir *mkdir;
11958 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
11959 panic("handle_written_mkdir: bad type");
11960 mkdir->md_state |= COMPLETE;
11961 complete_mkdir(mkdir);
11965 free_pagedep(pagedep)
11966 struct pagedep *pagedep;
11970 if (pagedep->pd_state & NEWBLOCK)
11972 if (!LIST_EMPTY(&pagedep->pd_dirremhd))
11974 for (i = 0; i < DAHASHSZ; i++)
11975 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
11977 if (!LIST_EMPTY(&pagedep->pd_pendinghd))
11979 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
11981 if (pagedep->pd_state & ONWORKLIST)
11982 WORKLIST_REMOVE(&pagedep->pd_list);
11983 LIST_REMOVE(pagedep, pd_hash);
11984 WORKITEM_FREE(pagedep, D_PAGEDEP);
11990 * Called from within softdep_disk_write_complete above.
11991 * A write operation was just completed. Removed inodes can
11992 * now be freed and associated block pointers may be committed.
11993 * Note that this routine is always called from interrupt level
11994 * with further interrupts from this device blocked.
11996 * If the write did not succeed, we will do all the roll-forward
11997 * operations, but we will not take the actions that will allow its
11998 * dependencies to be processed.
12001 handle_written_filepage(pagedep, bp, flags)
12002 struct pagedep *pagedep;
12003 struct buf *bp; /* buffer containing the written page */
12006 struct dirrem *dirrem;
12007 struct diradd *dap, *nextdap;
12011 if ((pagedep->pd_state & IOSTARTED) == 0)
12012 panic("handle_written_filepage: not started");
12013 pagedep->pd_state &= ~IOSTARTED;
12014 if ((flags & WRITESUCCEEDED) == 0)
12017 * Process any directory removals that have been committed.
12019 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
12020 LIST_REMOVE(dirrem, dm_next);
12021 dirrem->dm_state |= COMPLETE;
12022 dirrem->dm_dirinum = pagedep->pd_ino;
12023 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
12024 ("handle_written_filepage: Journal entries not written."));
12025 add_to_worklist(&dirrem->dm_list, 0);
12028 * Free any directory additions that have been committed.
12029 * If it is a newly allocated block, we have to wait until
12030 * the on-disk directory inode claims the new block.
12032 if ((pagedep->pd_state & NEWBLOCK) == 0)
12033 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
12034 free_diradd(dap, NULL);
12037 * Uncommitted directory entries must be restored.
12039 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
12040 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
12042 nextdap = LIST_NEXT(dap, da_pdlist);
12043 if (dap->da_state & ATTACHED)
12044 panic("handle_written_filepage: attached");
12045 ep = (struct direct *)
12046 ((char *)bp->b_data + dap->da_offset);
12047 ep->d_ino = dap->da_newinum;
12048 dap->da_state &= ~UNDONE;
12049 dap->da_state |= ATTACHED;
12052 * If the inode referenced by the directory has
12053 * been written out, then the dependency can be
12054 * moved to the pending list.
12056 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
12057 LIST_REMOVE(dap, da_pdlist);
12058 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
12064 * If there were any rollbacks in the directory, then it must be
12065 * marked dirty so that its will eventually get written back in
12066 * its correct form.
12068 if (chgs || (flags & WRITESUCCEEDED) == 0) {
12069 if ((bp->b_flags & B_DELWRI) == 0)
12075 * If we are not waiting for a new directory block to be
12076 * claimed by its inode, then the pagedep will be freed.
12077 * Otherwise it will remain to track any new entries on
12078 * the page in case they are fsync'ed.
12080 free_pagedep(pagedep);
12085 * Writing back in-core inode structures.
12087 * The filesystem only accesses an inode's contents when it occupies an
12088 * "in-core" inode structure. These "in-core" structures are separate from
12089 * the page frames used to cache inode blocks. Only the latter are
12090 * transferred to/from the disk. So, when the updated contents of the
12091 * "in-core" inode structure are copied to the corresponding in-memory inode
12092 * block, the dependencies are also transferred. The following procedure is
12093 * called when copying a dirty "in-core" inode to a cached inode block.
12097 * Called when an inode is loaded from disk. If the effective link count
12098 * differed from the actual link count when it was last flushed, then we
12099 * need to ensure that the correct effective link count is put back.
12102 softdep_load_inodeblock(ip)
12103 struct inode *ip; /* the "in_core" copy of the inode */
12105 struct inodedep *inodedep;
12107 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12108 ("softdep_load_inodeblock called on non-softdep filesystem"));
12110 * Check for alternate nlink count.
12112 ip->i_effnlink = ip->i_nlink;
12113 ACQUIRE_LOCK(ip->i_ump);
12114 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
12116 FREE_LOCK(ip->i_ump);
12119 ip->i_effnlink -= inodedep->id_nlinkdelta;
12120 FREE_LOCK(ip->i_ump);
12124 * This routine is called just before the "in-core" inode
12125 * information is to be copied to the in-memory inode block.
12126 * Recall that an inode block contains several inodes. If
12127 * the force flag is set, then the dependencies will be
12128 * cleared so that the update can always be made. Note that
12129 * the buffer is locked when this routine is called, so we
12130 * will never be in the middle of writing the inode block
12134 softdep_update_inodeblock(ip, bp, waitfor)
12135 struct inode *ip; /* the "in_core" copy of the inode */
12136 struct buf *bp; /* the buffer containing the inode block */
12137 int waitfor; /* nonzero => update must be allowed */
12139 struct inodedep *inodedep;
12140 struct inoref *inoref;
12141 struct ufsmount *ump;
12142 struct worklist *wk;
12149 mp = UFSTOVFS(ump);
12150 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
12151 ("softdep_update_inodeblock called on non-softdep filesystem"));
12154 * Preserve the freelink that is on disk. clear_unlinked_inodedep()
12155 * does not have access to the in-core ip so must write directly into
12156 * the inode block buffer when setting freelink.
12158 if (fs->fs_magic == FS_UFS1_MAGIC)
12159 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
12160 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12162 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
12163 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12165 * If the effective link count is not equal to the actual link
12166 * count, then we must track the difference in an inodedep while
12167 * the inode is (potentially) tossed out of the cache. Otherwise,
12168 * if there is no existing inodedep, then there are no dependencies
12173 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12175 if (ip->i_effnlink != ip->i_nlink)
12176 panic("softdep_update_inodeblock: bad link count");
12179 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
12180 panic("softdep_update_inodeblock: bad delta");
12182 * If we're flushing all dependencies we must also move any waiting
12183 * for journal writes onto the bufwait list prior to I/O.
12186 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12187 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12189 jwait(&inoref->if_list, MNT_WAIT);
12195 * Changes have been initiated. Anything depending on these
12196 * changes cannot occur until this inode has been written.
12198 inodedep->id_state &= ~COMPLETE;
12199 if ((inodedep->id_state & ONWORKLIST) == 0)
12200 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
12202 * Any new dependencies associated with the incore inode must
12203 * now be moved to the list associated with the buffer holding
12204 * the in-memory copy of the inode. Once merged process any
12205 * allocdirects that are completed by the merger.
12207 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
12208 if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
12209 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
12211 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
12212 if (!TAILQ_EMPTY(&inodedep->id_extupdt))
12213 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
12216 * Now that the inode has been pushed into the buffer, the
12217 * operations dependent on the inode being written to disk
12218 * can be moved to the id_bufwait so that they will be
12219 * processed when the buffer I/O completes.
12221 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
12222 WORKLIST_REMOVE(wk);
12223 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
12226 * Newly allocated inodes cannot be written until the bitmap
12227 * that allocates them have been written (indicated by
12228 * DEPCOMPLETE being set in id_state). If we are doing a
12229 * forced sync (e.g., an fsync on a file), we force the bitmap
12230 * to be written so that the update can be done.
12232 if (waitfor == 0) {
12237 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
12241 ibp = inodedep->id_bmsafemap->sm_buf;
12242 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT);
12245 * If ibp came back as NULL, the dependency could have been
12246 * freed while we slept. Look it up again, and check to see
12247 * that it has completed.
12249 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
12255 if ((error = bwrite(ibp)) != 0)
12256 softdep_error("softdep_update_inodeblock: bwrite", error);
12260 * Merge the a new inode dependency list (such as id_newinoupdt) into an
12261 * old inode dependency list (such as id_inoupdt). This routine must be
12262 * called with splbio interrupts blocked.
12265 merge_inode_lists(newlisthead, oldlisthead)
12266 struct allocdirectlst *newlisthead;
12267 struct allocdirectlst *oldlisthead;
12269 struct allocdirect *listadp, *newadp;
12271 newadp = TAILQ_FIRST(newlisthead);
12272 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
12273 if (listadp->ad_offset < newadp->ad_offset) {
12274 listadp = TAILQ_NEXT(listadp, ad_next);
12277 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12278 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
12279 if (listadp->ad_offset == newadp->ad_offset) {
12280 allocdirect_merge(oldlisthead, newadp,
12284 newadp = TAILQ_FIRST(newlisthead);
12286 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
12287 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12288 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
12293 * If we are doing an fsync, then we must ensure that any directory
12294 * entries for the inode have been written after the inode gets to disk.
12298 struct vnode *vp; /* the "in_core" copy of the inode */
12300 struct inodedep *inodedep;
12301 struct pagedep *pagedep;
12302 struct inoref *inoref;
12303 struct ufsmount *ump;
12304 struct worklist *wk;
12305 struct diradd *dap;
12311 struct thread *td = curthread;
12312 int error, flushparent, pagedep_new_block;
12320 if (MOUNTEDSOFTDEP(mp) == 0)
12324 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12328 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12329 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12331 jwait(&inoref->if_list, MNT_WAIT);
12335 if (!LIST_EMPTY(&inodedep->id_inowait) ||
12336 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
12337 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
12338 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
12339 !TAILQ_EMPTY(&inodedep->id_newinoupdt))
12340 panic("softdep_fsync: pending ops %p", inodedep);
12341 for (error = 0, flushparent = 0; ; ) {
12342 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
12344 if (wk->wk_type != D_DIRADD)
12345 panic("softdep_fsync: Unexpected type %s",
12346 TYPENAME(wk->wk_type));
12347 dap = WK_DIRADD(wk);
12349 * Flush our parent if this directory entry has a MKDIR_PARENT
12350 * dependency or is contained in a newly allocated block.
12352 if (dap->da_state & DIRCHG)
12353 pagedep = dap->da_previous->dm_pagedep;
12355 pagedep = dap->da_pagedep;
12356 parentino = pagedep->pd_ino;
12357 lbn = pagedep->pd_lbn;
12358 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
12359 panic("softdep_fsync: dirty");
12360 if ((dap->da_state & MKDIR_PARENT) ||
12361 (pagedep->pd_state & NEWBLOCK))
12366 * If we are being fsync'ed as part of vgone'ing this vnode,
12367 * then we will not be able to release and recover the
12368 * vnode below, so we just have to give up on writing its
12369 * directory entry out. It will eventually be written, just
12370 * not now, but then the user was not asking to have it
12371 * written, so we are not breaking any promises.
12373 if (vp->v_iflag & VI_DOOMED)
12376 * We prevent deadlock by always fetching inodes from the
12377 * root, moving down the directory tree. Thus, when fetching
12378 * our parent directory, we first try to get the lock. If
12379 * that fails, we must unlock ourselves before requesting
12380 * the lock on our parent. See the comment in ufs_lookup
12381 * for details on possible races.
12384 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
12385 FFSV_FORCEINSMQ)) {
12386 error = vfs_busy(mp, MBF_NOWAIT);
12390 error = vfs_busy(mp, 0);
12391 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12395 if (vp->v_iflag & VI_DOOMED) {
12401 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
12402 &pvp, FFSV_FORCEINSMQ);
12404 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12405 if (vp->v_iflag & VI_DOOMED) {
12414 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
12415 * that are contained in direct blocks will be resolved by
12416 * doing a ffs_update. Pagedeps contained in indirect blocks
12417 * may require a complete sync'ing of the directory. So, we
12418 * try the cheap and fast ffs_update first, and if that fails,
12419 * then we do the slower ffs_syncvnode of the directory.
12424 if ((error = ffs_update(pvp, 1)) != 0) {
12430 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
12431 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
12432 if (wk->wk_type != D_DIRADD)
12433 panic("softdep_fsync: Unexpected type %s",
12434 TYPENAME(wk->wk_type));
12435 dap = WK_DIRADD(wk);
12436 if (dap->da_state & DIRCHG)
12437 pagedep = dap->da_previous->dm_pagedep;
12439 pagedep = dap->da_pagedep;
12440 pagedep_new_block = pagedep->pd_state & NEWBLOCK;
12443 if (pagedep_new_block && (error =
12444 ffs_syncvnode(pvp, MNT_WAIT, 0))) {
12454 * Flush directory page containing the inode's name.
12456 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
12459 error = bwrite(bp);
12466 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
12474 * Flush all the dirty bitmaps associated with the block device
12475 * before flushing the rest of the dirty blocks so as to reduce
12476 * the number of dependencies that will have to be rolled back.
12481 softdep_fsync_mountdev(vp)
12484 struct buf *bp, *nbp;
12485 struct worklist *wk;
12488 if (!vn_isdisk(vp, NULL))
12489 panic("softdep_fsync_mountdev: vnode not a disk");
12490 bo = &vp->v_bufobj;
12493 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
12495 * If it is already scheduled, skip to the next buffer.
12497 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
12500 if ((bp->b_flags & B_DELWRI) == 0)
12501 panic("softdep_fsync_mountdev: not dirty");
12503 * We are only interested in bitmaps with outstanding
12506 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
12507 wk->wk_type != D_BMSAFEMAP ||
12508 (bp->b_vflags & BV_BKGRDINPROG)) {
12514 (void) bawrite(bp);
12522 * Sync all cylinder groups that were dirty at the time this function is
12523 * called. Newly dirtied cgs will be inserted before the sentinel. This
12524 * is used to flush freedep activity that may be holding up writes to a
12528 sync_cgs(mp, waitfor)
12532 struct bmsafemap *bmsafemap;
12533 struct bmsafemap *sentinel;
12534 struct ufsmount *ump;
12538 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK);
12539 sentinel->sm_cg = -1;
12540 ump = VFSTOUFS(mp);
12543 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next);
12544 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL;
12545 bmsafemap = LIST_NEXT(sentinel, sm_next)) {
12546 /* Skip sentinels and cgs with no work to release. */
12547 if (bmsafemap->sm_cg == -1 ||
12548 (LIST_EMPTY(&bmsafemap->sm_freehd) &&
12549 LIST_EMPTY(&bmsafemap->sm_freewr))) {
12550 LIST_REMOVE(sentinel, sm_next);
12551 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12555 * If we don't get the lock and we're waiting try again, if
12556 * not move on to the next buf and try to sync it.
12558 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor);
12559 if (bp == NULL && waitfor == MNT_WAIT)
12561 LIST_REMOVE(sentinel, sm_next);
12562 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12566 if (waitfor == MNT_NOWAIT)
12569 error = bwrite(bp);
12574 LIST_REMOVE(sentinel, sm_next);
12576 free(sentinel, M_BMSAFEMAP);
12581 * This routine is called when we are trying to synchronously flush a
12582 * file. This routine must eliminate any filesystem metadata dependencies
12583 * so that the syncing routine can succeed.
12586 softdep_sync_metadata(struct vnode *vp)
12592 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12593 ("softdep_sync_metadata called on non-softdep filesystem"));
12595 * Ensure that any direct block dependencies have been cleared,
12596 * truncations are started, and inode references are journaled.
12598 ACQUIRE_LOCK(ip->i_ump);
12600 * Write all journal records to prevent rollbacks on devvp.
12602 if (vp->v_type == VCHR)
12603 softdep_flushjournal(vp->v_mount);
12604 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number);
12606 * Ensure that all truncates are written so we won't find deps on
12609 process_truncates(vp);
12610 FREE_LOCK(ip->i_ump);
12616 * This routine is called when we are attempting to sync a buf with
12617 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any
12618 * other IO it can but returns EBUSY if the buffer is not yet able to
12619 * be written. Dependencies which will not cause rollbacks will always
12623 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
12625 struct indirdep *indirdep;
12626 struct pagedep *pagedep;
12627 struct allocindir *aip;
12628 struct newblk *newblk;
12629 struct ufsmount *ump;
12631 struct worklist *wk;
12634 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
12635 ("softdep_sync_buf called on non-softdep filesystem"));
12637 * For VCHR we just don't want to force flush any dependencies that
12638 * will cause rollbacks.
12640 if (vp->v_type == VCHR) {
12641 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0))
12645 ump = VTOI(vp)->i_ump;
12648 * As we hold the buffer locked, none of its dependencies
12653 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
12654 switch (wk->wk_type) {
12656 case D_ALLOCDIRECT:
12658 newblk = WK_NEWBLK(wk);
12659 if (newblk->nb_jnewblk != NULL) {
12660 if (waitfor == MNT_NOWAIT) {
12664 jwait(&newblk->nb_jnewblk->jn_list, waitfor);
12667 if (newblk->nb_state & DEPCOMPLETE ||
12668 waitfor == MNT_NOWAIT)
12670 nbp = newblk->nb_bmsafemap->sm_buf;
12671 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12675 if ((error = bwrite(nbp)) != 0)
12681 indirdep = WK_INDIRDEP(wk);
12682 if (waitfor == MNT_NOWAIT) {
12683 if (!TAILQ_EMPTY(&indirdep->ir_trunc) ||
12684 !LIST_EMPTY(&indirdep->ir_deplisthd)) {
12689 if (!TAILQ_EMPTY(&indirdep->ir_trunc))
12690 panic("softdep_sync_buf: truncation pending.");
12692 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
12693 newblk = (struct newblk *)aip;
12694 if (newblk->nb_jnewblk != NULL) {
12695 jwait(&newblk->nb_jnewblk->jn_list,
12699 if (newblk->nb_state & DEPCOMPLETE)
12701 nbp = newblk->nb_bmsafemap->sm_buf;
12702 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12706 if ((error = bwrite(nbp)) != 0)
12715 * Only flush directory entries in synchronous passes.
12717 if (waitfor != MNT_WAIT) {
12722 * While syncing snapshots, we must allow recursive
12727 * We are trying to sync a directory that may
12728 * have dependencies on both its own metadata
12729 * and/or dependencies on the inodes of any
12730 * recently allocated files. We walk its diradd
12731 * lists pushing out the associated inode.
12733 pagedep = WK_PAGEDEP(wk);
12734 for (i = 0; i < DAHASHSZ; i++) {
12735 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
12737 if ((error = flush_pagedep_deps(vp, wk->wk_mp,
12738 &pagedep->pd_diraddhd[i]))) {
12753 panic("softdep_sync_buf: Unknown type %s",
12754 TYPENAME(wk->wk_type));
12765 * Flush the dependencies associated with an inodedep.
12766 * Called with splbio blocked.
12769 flush_inodedep_deps(vp, mp, ino)
12774 struct inodedep *inodedep;
12775 struct inoref *inoref;
12776 struct ufsmount *ump;
12777 int error, waitfor;
12780 * This work is done in two passes. The first pass grabs most
12781 * of the buffers and begins asynchronously writing them. The
12782 * only way to wait for these asynchronous writes is to sleep
12783 * on the filesystem vnode which may stay busy for a long time
12784 * if the filesystem is active. So, instead, we make a second
12785 * pass over the dependencies blocking on each write. In the
12786 * usual case we will be blocking against a write that we
12787 * initiated, so when it is done the dependency will have been
12788 * resolved. Thus the second pass is expected to end quickly.
12789 * We give a brief window at the top of the loop to allow
12790 * any pending I/O to complete.
12792 ump = VFSTOUFS(mp);
12794 for (error = 0, waitfor = MNT_NOWAIT; ; ) {
12800 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
12802 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12803 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12805 jwait(&inoref->if_list, MNT_WAIT);
12809 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
12810 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
12811 flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
12812 flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
12815 * If pass2, we are done, otherwise do pass 2.
12817 if (waitfor == MNT_WAIT)
12819 waitfor = MNT_WAIT;
12822 * Try freeing inodedep in case all dependencies have been removed.
12824 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
12825 (void) free_inodedep(inodedep);
12830 * Flush an inode dependency list.
12831 * Called with splbio blocked.
12834 flush_deplist(listhead, waitfor, errorp)
12835 struct allocdirectlst *listhead;
12839 struct allocdirect *adp;
12840 struct newblk *newblk;
12841 struct ufsmount *ump;
12844 if ((adp = TAILQ_FIRST(listhead)) == NULL)
12846 ump = VFSTOUFS(adp->ad_list.wk_mp);
12848 TAILQ_FOREACH(adp, listhead, ad_next) {
12849 newblk = (struct newblk *)adp;
12850 if (newblk->nb_jnewblk != NULL) {
12851 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12854 if (newblk->nb_state & DEPCOMPLETE)
12856 bp = newblk->nb_bmsafemap->sm_buf;
12857 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor);
12859 if (waitfor == MNT_NOWAIT)
12864 if (waitfor == MNT_NOWAIT)
12867 *errorp = bwrite(bp);
12875 * Flush dependencies associated with an allocdirect block.
12878 flush_newblk_dep(vp, mp, lbn)
12883 struct newblk *newblk;
12884 struct ufsmount *ump;
12888 ufs2_daddr_t blkno;
12892 bo = &vp->v_bufobj;
12894 blkno = DIP(ip, i_db[lbn]);
12896 panic("flush_newblk_dep: Missing block");
12897 ump = VFSTOUFS(mp);
12900 * Loop until all dependencies related to this block are satisfied.
12901 * We must be careful to restart after each sleep in case a write
12902 * completes some part of this process for us.
12905 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
12909 if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
12910 panic("flush_newblk_deps: Bad newblk %p", newblk);
12912 * Flush the journal.
12914 if (newblk->nb_jnewblk != NULL) {
12915 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12919 * Write the bitmap dependency.
12921 if ((newblk->nb_state & DEPCOMPLETE) == 0) {
12922 bp = newblk->nb_bmsafemap->sm_buf;
12923 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
12927 error = bwrite(bp);
12934 * Write the buffer.
12938 bp = gbincore(bo, lbn);
12940 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
12941 LK_INTERLOCK, BO_LOCKPTR(bo));
12942 if (error == ENOLCK) {
12945 continue; /* Slept, retry */
12948 break; /* Failed */
12949 if (bp->b_flags & B_DELWRI) {
12951 error = bwrite(bp);
12959 * We have to wait for the direct pointers to
12960 * point at the newdirblk before the dependency
12963 error = ffs_update(vp, 1);
12972 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
12973 * Called with splbio blocked.
12976 flush_pagedep_deps(pvp, mp, diraddhdp)
12979 struct diraddhd *diraddhdp;
12981 struct inodedep *inodedep;
12982 struct inoref *inoref;
12983 struct ufsmount *ump;
12984 struct diradd *dap;
12989 struct diraddhd unfinished;
12991 LIST_INIT(&unfinished);
12992 ump = VFSTOUFS(mp);
12995 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
12997 * Flush ourselves if this directory entry
12998 * has a MKDIR_PARENT dependency.
13000 if (dap->da_state & MKDIR_PARENT) {
13002 if ((error = ffs_update(pvp, 1)) != 0)
13006 * If that cleared dependencies, go on to next.
13008 if (dap != LIST_FIRST(diraddhdp))
13011 * All MKDIR_PARENT dependencies and all the
13012 * NEWBLOCK pagedeps that are contained in direct
13013 * blocks were resolved by doing above ffs_update.
13014 * Pagedeps contained in indirect blocks may
13015 * require a complete sync'ing of the directory.
13016 * We are in the midst of doing a complete sync,
13017 * so if they are not resolved in this pass we
13018 * defer them for now as they will be sync'ed by
13019 * our caller shortly.
13021 LIST_REMOVE(dap, da_pdlist);
13022 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist);
13026 * A newly allocated directory must have its "." and
13027 * ".." entries written out before its name can be
13028 * committed in its parent.
13030 inum = dap->da_newinum;
13031 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
13032 panic("flush_pagedep_deps: lost inode1");
13034 * Wait for any pending journal adds to complete so we don't
13035 * cause rollbacks while syncing.
13037 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
13038 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
13040 jwait(&inoref->if_list, MNT_WAIT);
13044 if (dap->da_state & MKDIR_BODY) {
13046 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13049 error = flush_newblk_dep(vp, mp, 0);
13051 * If we still have the dependency we might need to
13052 * update the vnode to sync the new link count to
13055 if (error == 0 && dap == LIST_FIRST(diraddhdp))
13056 error = ffs_update(vp, 1);
13062 * If that cleared dependencies, go on to next.
13064 if (dap != LIST_FIRST(diraddhdp))
13066 if (dap->da_state & MKDIR_BODY) {
13067 inodedep_lookup(UFSTOVFS(ump), inum, 0,
13069 panic("flush_pagedep_deps: MKDIR_BODY "
13070 "inodedep %p dap %p vp %p",
13071 inodedep, dap, vp);
13075 * Flush the inode on which the directory entry depends.
13076 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
13077 * the only remaining dependency is that the updated inode
13078 * count must get pushed to disk. The inode has already
13079 * been pushed into its inode buffer (via VOP_UPDATE) at
13080 * the time of the reference count change. So we need only
13081 * locate that buffer, ensure that there will be no rollback
13082 * caused by a bitmap dependency, then write the inode buffer.
13085 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
13086 panic("flush_pagedep_deps: lost inode");
13088 * If the inode still has bitmap dependencies,
13089 * push them to disk.
13091 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
13092 bp = inodedep->id_bmsafemap->sm_buf;
13093 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
13097 if ((error = bwrite(bp)) != 0)
13100 if (dap != LIST_FIRST(diraddhdp))
13104 * If the inode is still sitting in a buffer waiting
13105 * to be written or waiting for the link count to be
13106 * adjusted update it here to flush it to disk.
13108 if (dap == LIST_FIRST(diraddhdp)) {
13110 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13113 error = ffs_update(vp, 1);
13120 * If we have failed to get rid of all the dependencies
13121 * then something is seriously wrong.
13123 if (dap == LIST_FIRST(diraddhdp)) {
13124 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
13125 panic("flush_pagedep_deps: failed to flush "
13126 "inodedep %p ino %ju dap %p",
13127 inodedep, (uintmax_t)inum, dap);
13132 while ((dap = LIST_FIRST(&unfinished)) != NULL) {
13133 LIST_REMOVE(dap, da_pdlist);
13134 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist);
13140 * A large burst of file addition or deletion activity can drive the
13141 * memory load excessively high. First attempt to slow things down
13142 * using the techniques below. If that fails, this routine requests
13143 * the offending operations to fall back to running synchronously
13144 * until the memory load returns to a reasonable level.
13147 softdep_slowdown(vp)
13150 struct ufsmount *ump;
13152 int max_softdeps_hard;
13154 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
13155 ("softdep_slowdown called on non-softdep filesystem"));
13156 ump = VFSTOUFS(vp->v_mount);
13160 * Check for journal space if needed.
13162 if (DOINGSUJ(vp)) {
13163 if (journal_space(ump, 0) == 0)
13167 * If the system is under its limits and our filesystem is
13168 * not responsible for more than our share of the usage and
13169 * we are not low on journal space, then no need to slow down.
13171 max_softdeps_hard = max_softdeps * 11 / 10;
13172 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 &&
13173 dep_current[D_INODEDEP] < max_softdeps_hard &&
13174 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 &&
13175 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 &&
13176 ump->softdep_curdeps[D_DIRREM] <
13177 (max_softdeps_hard / 2) / stat_flush_threads &&
13178 ump->softdep_curdeps[D_INODEDEP] <
13179 max_softdeps_hard / stat_flush_threads &&
13180 ump->softdep_curdeps[D_INDIRDEP] <
13181 (max_softdeps_hard / 1000) / stat_flush_threads &&
13182 ump->softdep_curdeps[D_FREEBLKS] <
13183 max_softdeps_hard / stat_flush_threads) {
13188 * If the journal is low or our filesystem is over its limit
13189 * then speedup the cleanup.
13191 if (ump->softdep_curdeps[D_INDIRDEP] <
13192 (max_softdeps_hard / 1000) / stat_flush_threads || jlow)
13193 softdep_speedup(ump);
13194 stat_sync_limit_hit += 1;
13197 * We only slow down the rate at which new dependencies are
13198 * generated if we are not using journaling. With journaling,
13199 * the cleanup should always be sufficient to keep things
13208 * Called by the allocation routines when they are about to fail
13209 * in the hope that we can free up the requested resource (inodes
13212 * First check to see if the work list has anything on it. If it has,
13213 * clean up entries until we successfully free the requested resource.
13214 * Because this process holds inodes locked, we cannot handle any remove
13215 * requests that might block on a locked inode as that could lead to
13216 * deadlock. If the worklist yields none of the requested resource,
13217 * start syncing out vnodes to free up the needed space.
13220 softdep_request_cleanup(fs, vp, cred, resource)
13223 struct ucred *cred;
13226 struct ufsmount *ump;
13228 struct vnode *lvp, *mvp;
13230 ufs2_daddr_t needed;
13234 * If we are being called because of a process doing a
13235 * copy-on-write, then it is not safe to process any
13236 * worklist items as we will recurse into the copyonwrite
13237 * routine. This will result in an incoherent snapshot.
13238 * If the vnode that we hold is a snapshot, we must avoid
13239 * handling other resources that could cause deadlock.
13241 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
13244 if (resource == FLUSH_BLOCKS_WAIT)
13245 stat_cleanup_blkrequests += 1;
13247 stat_cleanup_inorequests += 1;
13250 ump = VFSTOUFS(mp);
13251 mtx_assert(UFS_MTX(ump), MA_OWNED);
13253 error = ffs_update(vp, 1);
13254 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) {
13259 * If we are in need of resources, start by cleaning up
13260 * any block removals associated with our inode.
13263 process_removes(vp);
13264 process_truncates(vp);
13267 * Now clean up at least as many resources as we will need.
13269 * When requested to clean up inodes, the number that are needed
13270 * is set by the number of simultaneous writers (mnt_writeopcount)
13271 * plus a bit of slop (2) in case some more writers show up while
13274 * When requested to free up space, the amount of space that
13275 * we need is enough blocks to allocate a full-sized segment
13276 * (fs_contigsumsize). The number of such segments that will
13277 * be needed is set by the number of simultaneous writers
13278 * (mnt_writeopcount) plus a bit of slop (2) in case some more
13279 * writers show up while we are cleaning.
13281 * Additionally, if we are unpriviledged and allocating space,
13282 * we need to ensure that we clean up enough blocks to get the
13283 * needed number of blocks over the threshhold of the minimum
13284 * number of blocks required to be kept free by the filesystem
13287 if (resource == FLUSH_INODES_WAIT) {
13288 needed = vp->v_mount->mnt_writeopcount + 2;
13289 } else if (resource == FLUSH_BLOCKS_WAIT) {
13290 needed = (vp->v_mount->mnt_writeopcount + 2) *
13291 fs->fs_contigsumsize;
13292 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0))
13293 needed += fragstoblks(fs,
13294 roundup((fs->fs_dsize * fs->fs_minfree / 100) -
13295 fs->fs_cstotal.cs_nffree, fs->fs_frag));
13298 printf("softdep_request_cleanup: Unknown resource type %d\n",
13302 starttime = time_second;
13304 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 &&
13305 fs->fs_cstotal.cs_nbfree <= needed) ||
13306 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13307 fs->fs_cstotal.cs_nifree <= needed)) {
13309 if (ump->softdep_on_worklist > 0 &&
13310 process_worklist_item(UFSTOVFS(ump),
13311 ump->softdep_on_worklist, LK_NOWAIT) != 0)
13312 stat_worklist_push += 1;
13316 * If we still need resources and there are no more worklist
13317 * entries to process to obtain them, we have to start flushing
13318 * the dirty vnodes to force the release of additional requests
13319 * to the worklist that we can then process to reap addition
13320 * resources. We walk the vnodes associated with the mount point
13321 * until we get the needed worklist requests that we can reap.
13323 if ((resource == FLUSH_BLOCKS_WAIT &&
13324 fs->fs_cstotal.cs_nbfree <= needed) ||
13325 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13326 fs->fs_cstotal.cs_nifree <= needed)) {
13327 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) {
13328 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) {
13332 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
13335 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */
13339 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0);
13342 lvp = ump->um_devvp;
13343 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
13344 VOP_FSYNC(lvp, MNT_NOWAIT, curthread);
13345 VOP_UNLOCK(lvp, 0);
13347 if (ump->softdep_on_worklist > 0) {
13348 stat_cleanup_retries += 1;
13351 stat_cleanup_failures += 1;
13353 if (time_second - starttime > stat_cleanup_high_delay)
13354 stat_cleanup_high_delay = time_second - starttime;
13360 softdep_excess_items(struct ufsmount *ump, int item)
13363 KASSERT(item >= 0 && item < D_LAST, ("item %d", item));
13364 return (dep_current[item] > max_softdeps &&
13365 ump->softdep_curdeps[item] > max_softdeps /
13366 stat_flush_threads);
13370 schedule_cleanup(struct mount *mp)
13372 struct ufsmount *ump;
13375 ump = VFSTOUFS(mp);
13379 if ((td->td_pflags & TDP_KTHREAD) != 0 &&
13380 (td->td_proc->p_flag2 & P2_AST_SU) == 0) {
13382 * No ast is delivered to kernel threads, so nobody
13383 * would deref the mp. Some kernel threads
13384 * explicitely check for AST, e.g. NFS daemon does
13385 * this in the serving loop.
13389 if (td->td_su != NULL)
13390 vfs_rel(td->td_su);
13394 td->td_flags |= TDF_ASTPENDING;
13399 softdep_ast_cleanup_proc(void)
13403 struct ufsmount *ump;
13408 while ((mp = td->td_su) != NULL) {
13410 error = vfs_busy(mp, MBF_NOWAIT);
13414 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) {
13415 ump = VFSTOUFS(mp);
13419 if (softdep_excess_items(ump, D_INODEDEP)) {
13421 request_cleanup(mp, FLUSH_INODES);
13423 if (softdep_excess_items(ump, D_DIRREM)) {
13425 request_cleanup(mp, FLUSH_BLOCKS);
13428 if (softdep_excess_items(ump, D_NEWBLK) ||
13429 softdep_excess_items(ump, D_ALLOCDIRECT) ||
13430 softdep_excess_items(ump, D_ALLOCINDIR)) {
13431 error = vn_start_write(NULL, &mp,
13435 VFS_SYNC(mp, MNT_WAIT);
13436 vn_finished_write(mp);
13439 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req)
13448 * If memory utilization has gotten too high, deliberately slow things
13449 * down and speed up the I/O processing.
13452 request_cleanup(mp, resource)
13456 struct thread *td = curthread;
13457 struct ufsmount *ump;
13459 ump = VFSTOUFS(mp);
13462 * We never hold up the filesystem syncer or buf daemon.
13464 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
13467 * First check to see if the work list has gotten backlogged.
13468 * If it has, co-opt this process to help clean up two entries.
13469 * Because this process may hold inodes locked, we cannot
13470 * handle any remove requests that might block on a locked
13471 * inode as that could lead to deadlock. We set TDP_SOFTDEP
13472 * to avoid recursively processing the worklist.
13474 if (ump->softdep_on_worklist > max_softdeps / 10) {
13475 td->td_pflags |= TDP_SOFTDEP;
13476 process_worklist_item(mp, 2, LK_NOWAIT);
13477 td->td_pflags &= ~TDP_SOFTDEP;
13478 stat_worklist_push += 2;
13482 * Next, we attempt to speed up the syncer process. If that
13483 * is successful, then we allow the process to continue.
13485 if (softdep_speedup(ump) &&
13486 resource != FLUSH_BLOCKS_WAIT &&
13487 resource != FLUSH_INODES_WAIT)
13490 * If we are resource constrained on inode dependencies, try
13491 * flushing some dirty inodes. Otherwise, we are constrained
13492 * by file deletions, so try accelerating flushes of directories
13493 * with removal dependencies. We would like to do the cleanup
13494 * here, but we probably hold an inode locked at this point and
13495 * that might deadlock against one that we try to clean. So,
13496 * the best that we can do is request the syncer daemon to do
13497 * the cleanup for us.
13499 switch (resource) {
13502 case FLUSH_INODES_WAIT:
13503 ACQUIRE_GBLLOCK(&lk);
13504 stat_ino_limit_push += 1;
13505 req_clear_inodedeps += 1;
13507 stat_countp = &stat_ino_limit_hit;
13511 case FLUSH_BLOCKS_WAIT:
13512 ACQUIRE_GBLLOCK(&lk);
13513 stat_blk_limit_push += 1;
13514 req_clear_remove += 1;
13516 stat_countp = &stat_blk_limit_hit;
13520 panic("request_cleanup: unknown type");
13523 * Hopefully the syncer daemon will catch up and awaken us.
13524 * We wait at most tickdelay before proceeding in any case.
13526 ACQUIRE_GBLLOCK(&lk);
13529 if (callout_pending(&softdep_callout) == FALSE)
13530 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13533 if ((td->td_pflags & TDP_KTHREAD) == 0)
13534 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
13542 * Awaken processes pausing in request_cleanup and clear proc_waiting
13543 * to indicate that there is no longer a timer running. Pause_timer
13544 * will be called with the global softdep mutex (&lk) locked.
13551 GBLLOCK_OWNED(&lk);
13553 * The callout_ API has acquired mtx and will hold it around this
13556 *stat_countp += proc_waiting;
13557 wakeup(&proc_waiting);
13561 * If requested, try removing inode or removal dependencies.
13564 check_clear_deps(mp)
13569 * If we are suspended, it may be because of our using
13570 * too many inodedeps, so help clear them out.
13572 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended)
13573 clear_inodedeps(mp);
13575 * General requests for cleanup of backed up dependencies
13577 ACQUIRE_GBLLOCK(&lk);
13578 if (req_clear_inodedeps) {
13579 req_clear_inodedeps -= 1;
13581 clear_inodedeps(mp);
13582 ACQUIRE_GBLLOCK(&lk);
13583 wakeup(&proc_waiting);
13585 if (req_clear_remove) {
13586 req_clear_remove -= 1;
13589 ACQUIRE_GBLLOCK(&lk);
13590 wakeup(&proc_waiting);
13596 * Flush out a directory with at least one removal dependency in an effort to
13597 * reduce the number of dirrem, freefile, and freeblks dependency structures.
13603 struct pagedep_hashhead *pagedephd;
13604 struct pagedep *pagedep;
13605 struct ufsmount *ump;
13611 ump = VFSTOUFS(mp);
13614 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) {
13615 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++];
13616 if (ump->pagedep_nextclean > ump->pagedep_hash_size)
13617 ump->pagedep_nextclean = 0;
13618 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
13619 if (LIST_EMPTY(&pagedep->pd_dirremhd))
13621 ino = pagedep->pd_ino;
13622 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13627 * Let unmount clear deps
13629 error = vfs_busy(mp, MBF_NOWAIT);
13632 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13636 softdep_error("clear_remove: vget", error);
13639 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13640 softdep_error("clear_remove: fsync", error);
13641 bo = &vp->v_bufobj;
13647 vn_finished_write(mp);
13655 * Clear out a block of dirty inodes in an effort to reduce
13656 * the number of inodedep dependency structures.
13659 clear_inodedeps(mp)
13662 struct inodedep_hashhead *inodedephd;
13663 struct inodedep *inodedep;
13664 struct ufsmount *ump;
13668 ino_t firstino, lastino, ino;
13670 ump = VFSTOUFS(mp);
13674 * Pick a random inode dependency to be cleared.
13675 * We will then gather up all the inodes in its block
13676 * that have dependencies and flush them out.
13678 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) {
13679 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++];
13680 if (ump->inodedep_nextclean > ump->inodedep_hash_size)
13681 ump->inodedep_nextclean = 0;
13682 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
13685 if (inodedep == NULL)
13688 * Find the last inode in the block with dependencies.
13690 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
13691 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
13692 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
13695 * Asynchronously push all but the last inode with dependencies.
13696 * Synchronously push the last inode with dependencies to ensure
13697 * that the inode block gets written to free up the inodedeps.
13699 for (ino = firstino; ino <= lastino; ino++) {
13700 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
13702 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13705 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
13707 vn_finished_write(mp);
13711 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13712 FFSV_FORCEINSMQ)) != 0) {
13713 softdep_error("clear_inodedeps: vget", error);
13715 vn_finished_write(mp);
13720 if (ino == lastino) {
13721 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)))
13722 softdep_error("clear_inodedeps: fsync1", error);
13724 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13725 softdep_error("clear_inodedeps: fsync2", error);
13726 BO_LOCK(&vp->v_bufobj);
13728 BO_UNLOCK(&vp->v_bufobj);
13731 vn_finished_write(mp);
13737 softdep_buf_append(bp, wkhd)
13739 struct workhead *wkhd;
13741 struct worklist *wk;
13742 struct ufsmount *ump;
13744 if ((wk = LIST_FIRST(wkhd)) == NULL)
13746 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13747 ("softdep_buf_append called on non-softdep filesystem"));
13748 ump = VFSTOUFS(wk->wk_mp);
13750 while ((wk = LIST_FIRST(wkhd)) != NULL) {
13751 WORKLIST_REMOVE(wk);
13752 WORKLIST_INSERT(&bp->b_dep, wk);
13759 softdep_inode_append(ip, cred, wkhd)
13761 struct ucred *cred;
13762 struct workhead *wkhd;
13768 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
13769 ("softdep_inode_append called on non-softdep filesystem"));
13771 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
13772 (int)fs->fs_bsize, cred, &bp);
13775 softdep_freework(wkhd);
13778 softdep_buf_append(bp, wkhd);
13783 softdep_freework(wkhd)
13784 struct workhead *wkhd;
13786 struct worklist *wk;
13787 struct ufsmount *ump;
13789 if ((wk = LIST_FIRST(wkhd)) == NULL)
13791 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13792 ("softdep_freework called on non-softdep filesystem"));
13793 ump = VFSTOUFS(wk->wk_mp);
13795 handle_jwork(wkhd);
13800 * Function to determine if the buffer has outstanding dependencies
13801 * that will cause a roll-back if the buffer is written. If wantcount
13802 * is set, return number of dependencies, otherwise just yes or no.
13805 softdep_count_dependencies(bp, wantcount)
13809 struct worklist *wk;
13810 struct ufsmount *ump;
13811 struct bmsafemap *bmsafemap;
13812 struct freework *freework;
13813 struct inodedep *inodedep;
13814 struct indirdep *indirdep;
13815 struct freeblks *freeblks;
13816 struct allocindir *aip;
13817 struct pagedep *pagedep;
13818 struct dirrem *dirrem;
13819 struct newblk *newblk;
13820 struct mkdir *mkdir;
13821 struct diradd *dap;
13825 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
13827 ump = VFSTOUFS(wk->wk_mp);
13829 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
13830 switch (wk->wk_type) {
13833 inodedep = WK_INODEDEP(wk);
13834 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
13835 /* bitmap allocation dependency */
13840 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
13841 /* direct block pointer dependency */
13846 if (TAILQ_FIRST(&inodedep->id_extupdt)) {
13847 /* direct block pointer dependency */
13852 if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
13853 /* Add reference dependency. */
13861 indirdep = WK_INDIRDEP(wk);
13863 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) {
13864 /* indirect truncation dependency */
13870 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
13871 /* indirect block pointer dependency */
13879 pagedep = WK_PAGEDEP(wk);
13880 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
13881 if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
13882 /* Journal remove ref dependency. */
13888 for (i = 0; i < DAHASHSZ; i++) {
13890 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
13891 /* directory entry dependency */
13900 bmsafemap = WK_BMSAFEMAP(wk);
13901 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
13902 /* Add reference dependency. */
13907 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
13908 /* Allocate block dependency. */
13916 freeblks = WK_FREEBLKS(wk);
13917 if (LIST_FIRST(&freeblks->fb_jblkdephd)) {
13918 /* Freeblk journal dependency. */
13925 case D_ALLOCDIRECT:
13927 newblk = WK_NEWBLK(wk);
13928 if (newblk->nb_jnewblk) {
13929 /* Journal allocate dependency. */
13937 mkdir = WK_MKDIR(wk);
13938 if (mkdir->md_jaddref) {
13939 /* Journal reference dependency. */
13951 /* never a dependency on these blocks */
13955 panic("softdep_count_dependencies: Unexpected type %s",
13956 TYPENAME(wk->wk_type));
13966 * Acquire exclusive access to a buffer.
13967 * Must be called with a locked mtx parameter.
13968 * Return acquired buffer or NULL on failure.
13970 static struct buf *
13971 getdirtybuf(bp, lock, waitfor)
13973 struct rwlock *lock;
13978 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
13979 if (waitfor != MNT_WAIT)
13981 error = BUF_LOCK(bp,
13982 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock);
13984 * Even if we sucessfully acquire bp here, we have dropped
13985 * lock, which may violates our guarantee.
13989 else if (error != ENOLCK)
13990 panic("getdirtybuf: inconsistent lock: %d", error);
13994 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13995 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) {
13997 BO_LOCK(bp->b_bufobj);
13999 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
14000 bp->b_vflags |= BV_BKGRDWAIT;
14001 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj),
14002 PRIBIO | PDROP, "getbuf", 0);
14004 BO_UNLOCK(bp->b_bufobj);
14009 if (waitfor != MNT_WAIT)
14012 * The lock argument must be bp->b_vp's mutex in
14015 #ifdef DEBUG_VFS_LOCKS
14016 if (bp->b_vp->v_type != VCHR)
14017 ASSERT_BO_WLOCKED(bp->b_bufobj);
14019 bp->b_vflags |= BV_BKGRDWAIT;
14020 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0);
14023 if ((bp->b_flags & B_DELWRI) == 0) {
14033 * Check if it is safe to suspend the file system now. On entry,
14034 * the vnode interlock for devvp should be held. Return 0 with
14035 * the mount interlock held if the file system can be suspended now,
14036 * otherwise return EAGAIN with the mount interlock held.
14039 softdep_check_suspend(struct mount *mp,
14040 struct vnode *devvp,
14041 int softdep_depcnt,
14042 int softdep_accdepcnt,
14043 int secondary_writes,
14044 int secondary_accwrites)
14047 struct ufsmount *ump;
14048 struct inodedep *inodedep;
14049 int error, unlinked;
14051 bo = &devvp->v_bufobj;
14052 ASSERT_BO_WLOCKED(bo);
14055 * If we are not running with soft updates, then we need only
14056 * deal with secondary writes as we try to suspend.
14058 if (MOUNTEDSOFTDEP(mp) == 0) {
14060 while (mp->mnt_secondary_writes != 0) {
14062 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
14063 (PUSER - 1) | PDROP, "secwr", 0);
14069 * Reasons for needing more work before suspend:
14070 * - Dirty buffers on devvp.
14071 * - Secondary writes occurred after start of vnode sync loop
14074 if (bo->bo_numoutput > 0 ||
14075 bo->bo_dirty.bv_cnt > 0 ||
14076 secondary_writes != 0 ||
14077 mp->mnt_secondary_writes != 0 ||
14078 secondary_accwrites != mp->mnt_secondary_accwrites)
14085 * If we are running with soft updates, then we need to coordinate
14086 * with them as we try to suspend.
14088 ump = VFSTOUFS(mp);
14090 if (!TRY_ACQUIRE_LOCK(ump)) {
14098 if (mp->mnt_secondary_writes != 0) {
14101 msleep(&mp->mnt_secondary_writes,
14103 (PUSER - 1) | PDROP, "secwr", 0);
14111 if (MOUNTEDSUJ(mp)) {
14112 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked);
14114 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
14115 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS |
14116 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS |
14118 !check_inodedep_free(inodedep))
14125 * Reasons for needing more work before suspend:
14126 * - Dirty buffers on devvp.
14127 * - Softdep activity occurred after start of vnode sync loop
14128 * - Secondary writes occurred after start of vnode sync loop
14131 if (bo->bo_numoutput > 0 ||
14132 bo->bo_dirty.bv_cnt > 0 ||
14133 softdep_depcnt != unlinked ||
14134 ump->softdep_deps != unlinked ||
14135 softdep_accdepcnt != ump->softdep_accdeps ||
14136 secondary_writes != 0 ||
14137 mp->mnt_secondary_writes != 0 ||
14138 secondary_accwrites != mp->mnt_secondary_accwrites)
14147 * Get the number of dependency structures for the file system, both
14148 * the current number and the total number allocated. These will
14149 * later be used to detect that softdep processing has occurred.
14152 softdep_get_depcounts(struct mount *mp,
14153 int *softdep_depsp,
14154 int *softdep_accdepsp)
14156 struct ufsmount *ump;
14158 if (MOUNTEDSOFTDEP(mp) == 0) {
14159 *softdep_depsp = 0;
14160 *softdep_accdepsp = 0;
14163 ump = VFSTOUFS(mp);
14165 *softdep_depsp = ump->softdep_deps;
14166 *softdep_accdepsp = ump->softdep_accdeps;
14171 * Wait for pending output on a vnode to complete.
14172 * Must be called with vnode lock and interlock locked.
14174 * XXX: Should just be a call to bufobj_wwait().
14182 bo = &vp->v_bufobj;
14183 ASSERT_VOP_LOCKED(vp, "drain_output");
14184 ASSERT_BO_WLOCKED(bo);
14186 while (bo->bo_numoutput) {
14187 bo->bo_flag |= BO_WWAIT;
14188 msleep((caddr_t)&bo->bo_numoutput,
14189 BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0);
14194 * Called whenever a buffer that is being invalidated or reallocated
14195 * contains dependencies. This should only happen if an I/O error has
14196 * occurred. The routine is called with the buffer locked.
14199 softdep_deallocate_dependencies(bp)
14203 if ((bp->b_ioflags & BIO_ERROR) == 0)
14204 panic("softdep_deallocate_dependencies: dangling deps");
14205 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL)
14206 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
14208 printf("softdep_deallocate_dependencies: "
14209 "got error %d while accessing filesystem\n", bp->b_error);
14210 if (bp->b_error != ENXIO)
14211 panic("softdep_deallocate_dependencies: unrecovered I/O error");
14215 * Function to handle asynchronous write errors in the filesystem.
14218 softdep_error(func, error)
14223 /* XXX should do something better! */
14224 printf("%s: got error %d while accessing filesystem\n", func, error);
14230 inodedep_print(struct inodedep *inodedep, int verbose)
14232 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d"
14234 inodedep, inodedep->id_fs, inodedep->id_state,
14235 (intmax_t)inodedep->id_ino,
14236 (intmax_t)fsbtodb(inodedep->id_fs,
14237 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
14238 inodedep->id_nlinkdelta, inodedep->id_savednlink,
14239 inodedep->id_savedino1);
14244 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
14246 LIST_FIRST(&inodedep->id_pendinghd),
14247 LIST_FIRST(&inodedep->id_bufwait),
14248 LIST_FIRST(&inodedep->id_inowait),
14249 TAILQ_FIRST(&inodedep->id_inoreflst),
14250 inodedep->id_mkdiradd);
14251 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
14252 TAILQ_FIRST(&inodedep->id_inoupdt),
14253 TAILQ_FIRST(&inodedep->id_newinoupdt),
14254 TAILQ_FIRST(&inodedep->id_extupdt),
14255 TAILQ_FIRST(&inodedep->id_newextupdt));
14258 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
14261 if (have_addr == 0) {
14262 db_printf("Address required\n");
14265 inodedep_print((struct inodedep*)addr, 1);
14268 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
14270 struct inodedep_hashhead *inodedephd;
14271 struct inodedep *inodedep;
14272 struct ufsmount *ump;
14275 if (have_addr == 0) {
14276 db_printf("Address required\n");
14279 ump = (struct ufsmount *)addr;
14280 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) {
14281 inodedephd = &ump->inodedep_hashtbl[cnt];
14282 LIST_FOREACH(inodedep, inodedephd, id_hash) {
14283 inodedep_print(inodedep, 0);
14288 DB_SHOW_COMMAND(worklist, db_show_worklist)
14290 struct worklist *wk;
14292 if (have_addr == 0) {
14293 db_printf("Address required\n");
14296 wk = (struct worklist *)addr;
14297 printf("worklist: %p type %s state 0x%X\n",
14298 wk, TYPENAME(wk->wk_type), wk->wk_state);
14301 DB_SHOW_COMMAND(workhead, db_show_workhead)
14303 struct workhead *wkhd;
14304 struct worklist *wk;
14307 if (have_addr == 0) {
14308 db_printf("Address required\n");
14311 wkhd = (struct workhead *)addr;
14312 wk = LIST_FIRST(wkhd);
14313 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
14314 db_printf("worklist: %p type %s state 0x%X",
14315 wk, TYPENAME(wk->wk_type), wk->wk_state);
14317 db_printf("workhead overflow");
14322 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
14324 struct mkdirlist *mkdirlisthd;
14325 struct jaddref *jaddref;
14326 struct diradd *diradd;
14327 struct mkdir *mkdir;
14329 if (have_addr == 0) {
14330 db_printf("Address required\n");
14333 mkdirlisthd = (struct mkdirlist *)addr;
14334 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) {
14335 diradd = mkdir->md_diradd;
14336 db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
14337 mkdir, mkdir->md_state, diradd, diradd->da_state);
14338 if ((jaddref = mkdir->md_jaddref) != NULL)
14339 db_printf(" jaddref %p jaddref state 0x%X",
14340 jaddref, jaddref->ja_state);
14345 /* exported to ffs_vfsops.c */
14346 extern void db_print_ffs(struct ufsmount *ump);
14348 db_print_ffs(struct ufsmount *ump)
14350 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n",
14351 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname,
14352 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist,
14353 ump->softdep_deps, ump->softdep_req);
14358 #endif /* SOFTUPDATES */