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 *);
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 *);
761 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *,
763 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *);
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;
3602 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
3603 WORKLIST_REMOVE(wk);
3604 waiting = wk->wk_state & IOWAITING;
3605 wk->wk_state &= ~(INPROGRESS | IOWAITING);
3606 wk->wk_state |= COMPLETE;
3607 KASSERT(i++ < jseg->js_cnt,
3608 ("handle_written_jseg: overflow %d >= %d",
3609 i - 1, jseg->js_cnt));
3610 switch (wk->wk_type) {
3612 handle_written_jaddref(WK_JADDREF(wk));
3615 handle_written_jremref(WK_JREMREF(wk));
3618 rele_jseg(jseg); /* No jsegdep. */
3619 jmvref = WK_JMVREF(wk);
3620 LIST_REMOVE(jmvref, jm_deps);
3621 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0)
3622 free_pagedep(jmvref->jm_pagedep);
3623 WORKITEM_FREE(jmvref, D_JMVREF);
3626 handle_written_jnewblk(WK_JNEWBLK(wk));
3629 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep);
3632 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep);
3635 rele_jseg(jseg); /* No jsegdep. */
3636 WORKITEM_FREE(wk, D_JFSYNC);
3639 handle_written_jfreefrag(WK_JFREEFRAG(wk));
3642 panic("handle_written_jseg: Unknown type %s",
3643 TYPENAME(wk->wk_type));
3649 /* Release the self reference so the structure may be freed. */
3654 * Determine which jsegs are ready for completion processing. Waits for
3655 * synchronize cache to complete as well as forcing in-order completion
3656 * of journal entries.
3659 complete_jsegs(jseg)
3662 struct jblocks *jblocks;
3665 jblocks = jseg->js_jblocks;
3667 * Don't allow out of order completions. If this isn't the first
3668 * block wait for it to write before we're done.
3670 if (jseg != jblocks->jb_writeseg)
3672 /* Iterate through available jsegs processing their entries. */
3673 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) {
3674 jblocks->jb_oldestwrseq = jseg->js_oldseq;
3675 jsegn = TAILQ_NEXT(jseg, js_next);
3676 complete_jseg(jseg);
3679 jblocks->jb_writeseg = jseg;
3681 * Attempt to free jsegs now that oldestwrseq may have advanced.
3683 free_jsegs(jblocks);
3687 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle
3688 * the final completions.
3691 handle_written_jseg(jseg, bp)
3696 if (jseg->js_refs == 0)
3697 panic("handle_written_jseg: No self-reference on %p", jseg);
3698 jseg->js_state |= DEPCOMPLETE;
3700 * We'll never need this buffer again, set flags so it will be
3703 bp->b_flags |= B_INVAL | B_NOCACHE;
3705 complete_jsegs(jseg);
3708 static inline struct jsegdep *
3710 struct inoref *inoref;
3712 struct jsegdep *jsegdep;
3714 jsegdep = inoref->if_jsegdep;
3715 inoref->if_jsegdep = NULL;
3721 * Called once a jremref has made it to stable store. The jremref is marked
3722 * complete and we attempt to free it. Any pagedeps writes sleeping waiting
3723 * for the jremref to complete will be awoken by free_jremref.
3726 handle_written_jremref(jremref)
3727 struct jremref *jremref;
3729 struct inodedep *inodedep;
3730 struct jsegdep *jsegdep;
3731 struct dirrem *dirrem;
3733 /* Grab the jsegdep. */
3734 jsegdep = inoref_jseg(&jremref->jr_ref);
3736 * Remove us from the inoref list.
3738 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
3740 panic("handle_written_jremref: Lost inodedep");
3741 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
3743 * Complete the dirrem.
3745 dirrem = jremref->jr_dirrem;
3746 jremref->jr_dirrem = NULL;
3747 LIST_REMOVE(jremref, jr_deps);
3748 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
3749 jwork_insert(&dirrem->dm_jwork, jsegdep);
3750 if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
3751 (dirrem->dm_state & COMPLETE) != 0)
3752 add_to_worklist(&dirrem->dm_list, 0);
3753 free_jremref(jremref);
3757 * Called once a jaddref has made it to stable store. The dependency is
3758 * marked complete and any dependent structures are added to the inode
3759 * bufwait list to be completed as soon as it is written. If a bitmap write
3760 * depends on this entry we move the inode into the inodedephd of the
3761 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
3764 handle_written_jaddref(jaddref)
3765 struct jaddref *jaddref;
3767 struct jsegdep *jsegdep;
3768 struct inodedep *inodedep;
3769 struct diradd *diradd;
3770 struct mkdir *mkdir;
3772 /* Grab the jsegdep. */
3773 jsegdep = inoref_jseg(&jaddref->ja_ref);
3776 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3778 panic("handle_written_jaddref: Lost inodedep.");
3779 if (jaddref->ja_diradd == NULL)
3780 panic("handle_written_jaddref: No dependency");
3781 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
3782 diradd = jaddref->ja_diradd;
3783 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
3784 } else if (jaddref->ja_state & MKDIR_PARENT) {
3785 mkdir = jaddref->ja_mkdir;
3786 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
3787 } else if (jaddref->ja_state & MKDIR_BODY)
3788 mkdir = jaddref->ja_mkdir;
3790 panic("handle_written_jaddref: Unknown dependency %p",
3791 jaddref->ja_diradd);
3792 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */
3794 * Remove us from the inode list.
3796 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
3798 * The mkdir may be waiting on the jaddref to clear before freeing.
3801 KASSERT(mkdir->md_list.wk_type == D_MKDIR,
3802 ("handle_written_jaddref: Incorrect type for mkdir %s",
3803 TYPENAME(mkdir->md_list.wk_type)));
3804 mkdir->md_jaddref = NULL;
3805 diradd = mkdir->md_diradd;
3806 mkdir->md_state |= DEPCOMPLETE;
3807 complete_mkdir(mkdir);
3809 jwork_insert(&diradd->da_jwork, jsegdep);
3810 if (jaddref->ja_state & NEWBLOCK) {
3811 inodedep->id_state |= ONDEPLIST;
3812 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
3815 free_jaddref(jaddref);
3819 * Called once a jnewblk journal is written. The allocdirect or allocindir
3820 * is placed in the bmsafemap to await notification of a written bitmap. If
3821 * the operation was canceled we add the segdep to the appropriate
3822 * dependency to free the journal space once the canceling operation
3826 handle_written_jnewblk(jnewblk)
3827 struct jnewblk *jnewblk;
3829 struct bmsafemap *bmsafemap;
3830 struct freefrag *freefrag;
3831 struct freework *freework;
3832 struct jsegdep *jsegdep;
3833 struct newblk *newblk;
3835 /* Grab the jsegdep. */
3836 jsegdep = jnewblk->jn_jsegdep;
3837 jnewblk->jn_jsegdep = NULL;
3838 if (jnewblk->jn_dep == NULL)
3839 panic("handle_written_jnewblk: No dependency for the segdep.");
3840 switch (jnewblk->jn_dep->wk_type) {
3845 * Add the written block to the bmsafemap so it can
3846 * be notified when the bitmap is on disk.
3848 newblk = WK_NEWBLK(jnewblk->jn_dep);
3849 newblk->nb_jnewblk = NULL;
3850 if ((newblk->nb_state & GOINGAWAY) == 0) {
3851 bmsafemap = newblk->nb_bmsafemap;
3852 newblk->nb_state |= ONDEPLIST;
3853 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk,
3856 jwork_insert(&newblk->nb_jwork, jsegdep);
3860 * A newblock being removed by a freefrag when replaced by
3863 freefrag = WK_FREEFRAG(jnewblk->jn_dep);
3864 freefrag->ff_jdep = NULL;
3865 jwork_insert(&freefrag->ff_jwork, jsegdep);
3869 * A direct block was removed by truncate.
3871 freework = WK_FREEWORK(jnewblk->jn_dep);
3872 freework->fw_jnewblk = NULL;
3873 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep);
3876 panic("handle_written_jnewblk: Unknown type %d.",
3877 jnewblk->jn_dep->wk_type);
3879 jnewblk->jn_dep = NULL;
3880 free_jnewblk(jnewblk);
3884 * Cancel a jfreefrag that won't be needed, probably due to colliding with
3885 * an in-flight allocation that has not yet been committed. Divorce us
3886 * from the freefrag and mark it DEPCOMPLETE so that it may be added
3890 cancel_jfreefrag(jfreefrag)
3891 struct jfreefrag *jfreefrag;
3893 struct freefrag *freefrag;
3895 if (jfreefrag->fr_jsegdep) {
3896 free_jsegdep(jfreefrag->fr_jsegdep);
3897 jfreefrag->fr_jsegdep = NULL;
3899 freefrag = jfreefrag->fr_freefrag;
3900 jfreefrag->fr_freefrag = NULL;
3901 free_jfreefrag(jfreefrag);
3902 freefrag->ff_state |= DEPCOMPLETE;
3903 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno);
3907 * Free a jfreefrag when the parent freefrag is rendered obsolete.
3910 free_jfreefrag(jfreefrag)
3911 struct jfreefrag *jfreefrag;
3914 if (jfreefrag->fr_state & INPROGRESS)
3915 WORKLIST_REMOVE(&jfreefrag->fr_list);
3916 else if (jfreefrag->fr_state & ONWORKLIST)
3917 remove_from_journal(&jfreefrag->fr_list);
3918 if (jfreefrag->fr_freefrag != NULL)
3919 panic("free_jfreefrag: Still attached to a freefrag.");
3920 WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
3924 * Called when the journal write for a jfreefrag completes. The parent
3925 * freefrag is added to the worklist if this completes its dependencies.
3928 handle_written_jfreefrag(jfreefrag)
3929 struct jfreefrag *jfreefrag;
3931 struct jsegdep *jsegdep;
3932 struct freefrag *freefrag;
3934 /* Grab the jsegdep. */
3935 jsegdep = jfreefrag->fr_jsegdep;
3936 jfreefrag->fr_jsegdep = NULL;
3937 freefrag = jfreefrag->fr_freefrag;
3938 if (freefrag == NULL)
3939 panic("handle_written_jfreefrag: No freefrag.");
3940 freefrag->ff_state |= DEPCOMPLETE;
3941 freefrag->ff_jdep = NULL;
3942 jwork_insert(&freefrag->ff_jwork, jsegdep);
3943 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
3944 add_to_worklist(&freefrag->ff_list, 0);
3945 jfreefrag->fr_freefrag = NULL;
3946 free_jfreefrag(jfreefrag);
3950 * Called when the journal write for a jfreeblk completes. The jfreeblk
3951 * is removed from the freeblks list of pending journal writes and the
3952 * jsegdep is moved to the freeblks jwork to be completed when all blocks
3953 * have been reclaimed.
3956 handle_written_jblkdep(jblkdep)
3957 struct jblkdep *jblkdep;
3959 struct freeblks *freeblks;
3960 struct jsegdep *jsegdep;
3962 /* Grab the jsegdep. */
3963 jsegdep = jblkdep->jb_jsegdep;
3964 jblkdep->jb_jsegdep = NULL;
3965 freeblks = jblkdep->jb_freeblks;
3966 LIST_REMOVE(jblkdep, jb_deps);
3967 jwork_insert(&freeblks->fb_jwork, jsegdep);
3969 * If the freeblks is all journaled, we can add it to the worklist.
3971 if (LIST_EMPTY(&freeblks->fb_jblkdephd) &&
3972 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
3973 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
3975 free_jblkdep(jblkdep);
3978 static struct jsegdep *
3979 newjsegdep(struct worklist *wk)
3981 struct jsegdep *jsegdep;
3983 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
3984 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
3985 jsegdep->jd_seg = NULL;
3990 static struct jmvref *
3991 newjmvref(dp, ino, oldoff, newoff)
3997 struct jmvref *jmvref;
3999 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
4000 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump));
4001 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
4002 jmvref->jm_parent = dp->i_number;
4003 jmvref->jm_ino = ino;
4004 jmvref->jm_oldoff = oldoff;
4005 jmvref->jm_newoff = newoff;
4011 * Allocate a new jremref that tracks the removal of ip from dp with the
4012 * directory entry offset of diroff. Mark the entry as ATTACHED and
4013 * DEPCOMPLETE as we have all the information required for the journal write
4014 * and the directory has already been removed from the buffer. The caller
4015 * is responsible for linking the jremref into the pagedep and adding it
4016 * to the journal to write. The MKDIR_PARENT flag is set if we're doing
4017 * a DOTDOT addition so handle_workitem_remove() can properly assign
4018 * the jsegdep when we're done.
4020 static struct jremref *
4021 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
4022 off_t diroff, nlink_t nlink)
4024 struct jremref *jremref;
4026 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
4027 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump));
4028 jremref->jr_state = ATTACHED;
4029 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
4031 jremref->jr_dirrem = dirrem;
4037 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
4038 nlink_t nlink, uint16_t mode)
4041 inoref->if_jsegdep = newjsegdep(&inoref->if_list);
4042 inoref->if_diroff = diroff;
4043 inoref->if_ino = ino;
4044 inoref->if_parent = parent;
4045 inoref->if_nlink = nlink;
4046 inoref->if_mode = mode;
4050 * Allocate a new jaddref to track the addition of ino to dp at diroff. The
4051 * directory offset may not be known until later. The caller is responsible
4052 * adding the entry to the journal when this information is available. nlink
4053 * should be the link count prior to the addition and mode is only required
4054 * to have the correct FMT.
4056 static struct jaddref *
4057 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
4060 struct jaddref *jaddref;
4062 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
4063 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump));
4064 jaddref->ja_state = ATTACHED;
4065 jaddref->ja_mkdir = NULL;
4066 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
4072 * Create a new free dependency for a freework. The caller is responsible
4073 * for adjusting the reference count when it has the lock held. The freedep
4074 * will track an outstanding bitmap write that will ultimately clear the
4075 * freework to continue.
4077 static struct freedep *
4078 newfreedep(struct freework *freework)
4080 struct freedep *freedep;
4082 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
4083 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
4084 freedep->fd_freework = freework;
4090 * Free a freedep structure once the buffer it is linked to is written. If
4091 * this is the last reference to the freework schedule it for completion.
4094 free_freedep(freedep)
4095 struct freedep *freedep;
4097 struct freework *freework;
4099 freework = freedep->fd_freework;
4100 freework->fw_freeblks->fb_cgwait--;
4101 if (--freework->fw_ref == 0)
4102 freework_enqueue(freework);
4103 WORKITEM_FREE(freedep, D_FREEDEP);
4107 * Allocate a new freework structure that may be a level in an indirect
4108 * when parent is not NULL or a top level block when it is. The top level
4109 * freework structures are allocated without the per-filesystem lock held
4110 * and before the freeblks is visible outside of softdep_setup_freeblocks().
4112 static struct freework *
4113 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal)
4114 struct ufsmount *ump;
4115 struct freeblks *freeblks;
4116 struct freework *parent;
4123 struct freework *freework;
4125 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
4126 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
4127 freework->fw_state = ATTACHED;
4128 freework->fw_jnewblk = NULL;
4129 freework->fw_freeblks = freeblks;
4130 freework->fw_parent = parent;
4131 freework->fw_lbn = lbn;
4132 freework->fw_blkno = nb;
4133 freework->fw_frags = frags;
4134 freework->fw_indir = NULL;
4135 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR)
4136 ? 0 : NINDIR(ump->um_fs) + 1;
4137 freework->fw_start = freework->fw_off = off;
4139 newjfreeblk(freeblks, lbn, nb, frags);
4140 if (parent == NULL) {
4142 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
4151 * Eliminate a jfreeblk for a block that does not need journaling.
4154 cancel_jfreeblk(freeblks, blkno)
4155 struct freeblks *freeblks;
4158 struct jfreeblk *jfreeblk;
4159 struct jblkdep *jblkdep;
4161 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) {
4162 if (jblkdep->jb_list.wk_type != D_JFREEBLK)
4164 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list);
4165 if (jfreeblk->jf_blkno == blkno)
4168 if (jblkdep == NULL)
4170 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno);
4171 free_jsegdep(jblkdep->jb_jsegdep);
4172 LIST_REMOVE(jblkdep, jb_deps);
4173 WORKITEM_FREE(jfreeblk, D_JFREEBLK);
4177 * Allocate a new jfreeblk to journal top level block pointer when truncating
4178 * a file. The caller must add this to the worklist when the per-filesystem
4181 static struct jfreeblk *
4182 newjfreeblk(freeblks, lbn, blkno, frags)
4183 struct freeblks *freeblks;
4188 struct jfreeblk *jfreeblk;
4190 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
4191 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK,
4192 freeblks->fb_list.wk_mp);
4193 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list);
4194 jfreeblk->jf_dep.jb_freeblks = freeblks;
4195 jfreeblk->jf_ino = freeblks->fb_inum;
4196 jfreeblk->jf_lbn = lbn;
4197 jfreeblk->jf_blkno = blkno;
4198 jfreeblk->jf_frags = frags;
4199 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps);
4205 * The journal is only prepared to handle full-size block numbers, so we
4206 * have to adjust the record to reflect the change to a full-size block.
4207 * For example, suppose we have a block made up of fragments 8-15 and
4208 * want to free its last two fragments. We are given a request that says:
4209 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0
4210 * where frags are the number of fragments to free and oldfrags are the
4211 * number of fragments to keep. To block align it, we have to change it to
4212 * have a valid full-size blkno, so it becomes:
4213 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6
4216 adjust_newfreework(freeblks, frag_offset)
4217 struct freeblks *freeblks;
4220 struct jfreeblk *jfreeblk;
4222 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL &&
4223 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK),
4224 ("adjust_newfreework: Missing freeblks dependency"));
4226 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd));
4227 jfreeblk->jf_blkno -= frag_offset;
4228 jfreeblk->jf_frags += frag_offset;
4232 * Allocate a new jtrunc to track a partial truncation.
4234 static struct jtrunc *
4235 newjtrunc(freeblks, size, extsize)
4236 struct freeblks *freeblks;
4240 struct jtrunc *jtrunc;
4242 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
4243 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC,
4244 freeblks->fb_list.wk_mp);
4245 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list);
4246 jtrunc->jt_dep.jb_freeblks = freeblks;
4247 jtrunc->jt_ino = freeblks->fb_inum;
4248 jtrunc->jt_size = size;
4249 jtrunc->jt_extsize = extsize;
4250 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps);
4256 * If we're canceling a new bitmap we have to search for another ref
4257 * to move into the bmsafemap dep. This might be better expressed
4258 * with another structure.
4261 move_newblock_dep(jaddref, inodedep)
4262 struct jaddref *jaddref;
4263 struct inodedep *inodedep;
4265 struct inoref *inoref;
4266 struct jaddref *jaddrefn;
4269 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4270 inoref = TAILQ_NEXT(inoref, if_deps)) {
4271 if ((jaddref->ja_state & NEWBLOCK) &&
4272 inoref->if_list.wk_type == D_JADDREF) {
4273 jaddrefn = (struct jaddref *)inoref;
4277 if (jaddrefn == NULL)
4279 jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
4280 jaddrefn->ja_state |= jaddref->ja_state &
4281 (ATTACHED | UNDONE | NEWBLOCK);
4282 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
4283 jaddref->ja_state |= ATTACHED;
4284 LIST_REMOVE(jaddref, ja_bmdeps);
4285 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
4290 * Cancel a jaddref either before it has been written or while it is being
4291 * written. This happens when a link is removed before the add reaches
4292 * the disk. The jaddref dependency is kept linked into the bmsafemap
4293 * and inode to prevent the link count or bitmap from reaching the disk
4294 * until handle_workitem_remove() re-adjusts the counts and bitmaps as
4297 * Returns 1 if the canceled addref requires journaling of the remove and
4301 cancel_jaddref(jaddref, inodedep, wkhd)
4302 struct jaddref *jaddref;
4303 struct inodedep *inodedep;
4304 struct workhead *wkhd;
4306 struct inoref *inoref;
4307 struct jsegdep *jsegdep;
4310 KASSERT((jaddref->ja_state & COMPLETE) == 0,
4311 ("cancel_jaddref: Canceling complete jaddref"));
4312 if (jaddref->ja_state & (INPROGRESS | COMPLETE))
4316 if (inodedep == NULL)
4317 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
4319 panic("cancel_jaddref: Lost inodedep");
4321 * We must adjust the nlink of any reference operation that follows
4322 * us so that it is consistent with the in-memory reference. This
4323 * ensures that inode nlink rollbacks always have the correct link.
4326 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4327 inoref = TAILQ_NEXT(inoref, if_deps)) {
4328 if (inoref->if_state & GOINGAWAY)
4333 jsegdep = inoref_jseg(&jaddref->ja_ref);
4334 if (jaddref->ja_state & NEWBLOCK)
4335 move_newblock_dep(jaddref, inodedep);
4336 wake_worklist(&jaddref->ja_list);
4337 jaddref->ja_mkdir = NULL;
4338 if (jaddref->ja_state & INPROGRESS) {
4339 jaddref->ja_state &= ~INPROGRESS;
4340 WORKLIST_REMOVE(&jaddref->ja_list);
4341 jwork_insert(wkhd, jsegdep);
4343 free_jsegdep(jsegdep);
4344 if (jaddref->ja_state & DEPCOMPLETE)
4345 remove_from_journal(&jaddref->ja_list);
4347 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE);
4349 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
4350 * can arrange for them to be freed with the bitmap. Otherwise we
4351 * no longer need this addref attached to the inoreflst and it
4352 * will incorrectly adjust nlink if we leave it.
4354 if ((jaddref->ja_state & NEWBLOCK) == 0) {
4355 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
4357 jaddref->ja_state |= COMPLETE;
4358 free_jaddref(jaddref);
4362 * Leave the head of the list for jsegdeps for fast merging.
4364 if (LIST_FIRST(wkhd) != NULL) {
4365 jaddref->ja_state |= ONWORKLIST;
4366 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
4368 WORKLIST_INSERT(wkhd, &jaddref->ja_list);
4374 * Attempt to free a jaddref structure when some work completes. This
4375 * should only succeed once the entry is written and all dependencies have
4379 free_jaddref(jaddref)
4380 struct jaddref *jaddref;
4383 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
4385 if (jaddref->ja_ref.if_jsegdep)
4386 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
4387 jaddref, jaddref->ja_state);
4388 if (jaddref->ja_state & NEWBLOCK)
4389 LIST_REMOVE(jaddref, ja_bmdeps);
4390 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST))
4391 panic("free_jaddref: Bad state %p(0x%X)",
4392 jaddref, jaddref->ja_state);
4393 if (jaddref->ja_mkdir != NULL)
4394 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
4395 WORKITEM_FREE(jaddref, D_JADDREF);
4399 * Free a jremref structure once it has been written or discarded.
4402 free_jremref(jremref)
4403 struct jremref *jremref;
4406 if (jremref->jr_ref.if_jsegdep)
4407 free_jsegdep(jremref->jr_ref.if_jsegdep);
4408 if (jremref->jr_state & INPROGRESS)
4409 panic("free_jremref: IO still pending");
4410 WORKITEM_FREE(jremref, D_JREMREF);
4414 * Free a jnewblk structure.
4417 free_jnewblk(jnewblk)
4418 struct jnewblk *jnewblk;
4421 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
4423 LIST_REMOVE(jnewblk, jn_deps);
4424 if (jnewblk->jn_dep != NULL)
4425 panic("free_jnewblk: Dependency still attached.");
4426 WORKITEM_FREE(jnewblk, D_JNEWBLK);
4430 * Cancel a jnewblk which has been been made redundant by frag extension.
4433 cancel_jnewblk(jnewblk, wkhd)
4434 struct jnewblk *jnewblk;
4435 struct workhead *wkhd;
4437 struct jsegdep *jsegdep;
4439 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno);
4440 jsegdep = jnewblk->jn_jsegdep;
4441 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL)
4442 panic("cancel_jnewblk: Invalid state");
4443 jnewblk->jn_jsegdep = NULL;
4444 jnewblk->jn_dep = NULL;
4445 jnewblk->jn_state |= GOINGAWAY;
4446 if (jnewblk->jn_state & INPROGRESS) {
4447 jnewblk->jn_state &= ~INPROGRESS;
4448 WORKLIST_REMOVE(&jnewblk->jn_list);
4449 jwork_insert(wkhd, jsegdep);
4451 free_jsegdep(jsegdep);
4452 remove_from_journal(&jnewblk->jn_list);
4454 wake_worklist(&jnewblk->jn_list);
4455 WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
4459 free_jblkdep(jblkdep)
4460 struct jblkdep *jblkdep;
4463 if (jblkdep->jb_list.wk_type == D_JFREEBLK)
4464 WORKITEM_FREE(jblkdep, D_JFREEBLK);
4465 else if (jblkdep->jb_list.wk_type == D_JTRUNC)
4466 WORKITEM_FREE(jblkdep, D_JTRUNC);
4468 panic("free_jblkdep: Unexpected type %s",
4469 TYPENAME(jblkdep->jb_list.wk_type));
4473 * Free a single jseg once it is no longer referenced in memory or on
4474 * disk. Reclaim journal blocks and dependencies waiting for the segment
4478 free_jseg(jseg, jblocks)
4480 struct jblocks *jblocks;
4482 struct freework *freework;
4485 * Free freework structures that were lingering to indicate freed
4486 * indirect blocks that forced journal write ordering on reallocate.
4488 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL)
4489 indirblk_remove(freework);
4490 if (jblocks->jb_oldestseg == jseg)
4491 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next);
4492 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
4493 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
4494 KASSERT(LIST_EMPTY(&jseg->js_entries),
4495 ("free_jseg: Freed jseg has valid entries."));
4496 WORKITEM_FREE(jseg, D_JSEG);
4500 * Free all jsegs that meet the criteria for being reclaimed and update
4505 struct jblocks *jblocks;
4510 * Free only those jsegs which have none allocated before them to
4511 * preserve the journal space ordering.
4513 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
4515 * Only reclaim space when nothing depends on this journal
4516 * set and another set has written that it is no longer
4519 if (jseg->js_refs != 0) {
4520 jblocks->jb_oldestseg = jseg;
4523 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE)
4525 if (jseg->js_seq > jblocks->jb_oldestwrseq)
4528 * We can free jsegs that didn't write entries when
4529 * oldestwrseq == js_seq.
4531 if (jseg->js_seq == jblocks->jb_oldestwrseq &&
4534 free_jseg(jseg, jblocks);
4537 * If we exited the loop above we still must discover the
4538 * oldest valid segment.
4541 for (jseg = jblocks->jb_oldestseg; jseg != NULL;
4542 jseg = TAILQ_NEXT(jseg, js_next))
4543 if (jseg->js_refs != 0)
4545 jblocks->jb_oldestseg = jseg;
4547 * The journal has no valid records but some jsegs may still be
4548 * waiting on oldestwrseq to advance. We force a small record
4549 * out to permit these lingering records to be reclaimed.
4551 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs))
4552 jblocks->jb_needseg = 1;
4556 * Release one reference to a jseg and free it if the count reaches 0. This
4557 * should eventually reclaim journal space as well.
4564 KASSERT(jseg->js_refs > 0,
4565 ("free_jseg: Invalid refcnt %d", jseg->js_refs));
4566 if (--jseg->js_refs != 0)
4568 free_jsegs(jseg->js_jblocks);
4572 * Release a jsegdep and decrement the jseg count.
4575 free_jsegdep(jsegdep)
4576 struct jsegdep *jsegdep;
4579 if (jsegdep->jd_seg)
4580 rele_jseg(jsegdep->jd_seg);
4581 WORKITEM_FREE(jsegdep, D_JSEGDEP);
4585 * Wait for a journal item to make it to disk. Initiate journal processing
4590 struct worklist *wk;
4594 LOCK_OWNED(VFSTOUFS(wk->wk_mp));
4596 * Blocking journal waits cause slow synchronous behavior. Record
4597 * stats on the frequency of these blocking operations.
4599 if (waitfor == MNT_WAIT) {
4600 stat_journal_wait++;
4601 switch (wk->wk_type) {
4604 stat_jwait_filepage++;
4608 stat_jwait_freeblks++;
4611 stat_jwait_newblk++;
4621 * If IO has not started we process the journal. We can't mark the
4622 * worklist item as IOWAITING because we drop the lock while
4623 * processing the journal and the worklist entry may be freed after
4624 * this point. The caller may call back in and re-issue the request.
4626 if ((wk->wk_state & INPROGRESS) == 0) {
4627 softdep_process_journal(wk->wk_mp, wk, waitfor);
4628 if (waitfor != MNT_WAIT)
4632 if (waitfor != MNT_WAIT)
4634 wait_worklist(wk, "jwait");
4639 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
4640 * appropriate. This is a convenience function to reduce duplicate code
4641 * for the setup and revert functions below.
4643 static struct inodedep *
4644 inodedep_lookup_ip(ip)
4647 struct inodedep *inodedep;
4649 KASSERT(ip->i_nlink >= ip->i_effnlink,
4650 ("inodedep_lookup_ip: bad delta"));
4651 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
4653 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
4654 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
4660 * Called prior to creating a new inode and linking it to a directory. The
4661 * jaddref structure must already be allocated by softdep_setup_inomapdep
4662 * and it is discovered here so we can initialize the mode and update
4666 softdep_setup_create(dp, ip)
4670 struct inodedep *inodedep;
4671 struct jaddref *jaddref;
4674 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4675 ("softdep_setup_create called on non-softdep filesystem"));
4676 KASSERT(ip->i_nlink == 1,
4677 ("softdep_setup_create: Invalid link count."));
4679 ACQUIRE_LOCK(dp->i_ump);
4680 inodedep = inodedep_lookup_ip(ip);
4681 if (DOINGSUJ(dvp)) {
4682 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4684 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
4685 ("softdep_setup_create: No addref structure present."));
4687 softdep_prelink(dvp, NULL);
4688 FREE_LOCK(dp->i_ump);
4692 * Create a jaddref structure to track the addition of a DOTDOT link when
4693 * we are reparenting an inode as part of a rename. This jaddref will be
4694 * found by softdep_setup_directory_change. Adjusts nlinkdelta for
4695 * non-journaling softdep.
4698 softdep_setup_dotdot_link(dp, ip)
4702 struct inodedep *inodedep;
4703 struct jaddref *jaddref;
4706 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4707 ("softdep_setup_dotdot_link called on non-softdep filesystem"));
4711 * We don't set MKDIR_PARENT as this is not tied to a mkdir and
4712 * is used as a normal link would be.
4715 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4716 dp->i_effnlink - 1, dp->i_mode);
4717 ACQUIRE_LOCK(dp->i_ump);
4718 inodedep = inodedep_lookup_ip(dp);
4720 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4722 softdep_prelink(dvp, ITOV(ip));
4723 FREE_LOCK(dp->i_ump);
4727 * Create a jaddref structure to track a new link to an inode. The directory
4728 * offset is not known until softdep_setup_directory_add or
4729 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling
4733 softdep_setup_link(dp, ip)
4737 struct inodedep *inodedep;
4738 struct jaddref *jaddref;
4741 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4742 ("softdep_setup_link called on non-softdep filesystem"));
4746 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
4748 ACQUIRE_LOCK(dp->i_ump);
4749 inodedep = inodedep_lookup_ip(ip);
4751 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4753 softdep_prelink(dvp, ITOV(ip));
4754 FREE_LOCK(dp->i_ump);
4758 * Called to create the jaddref structures to track . and .. references as
4759 * well as lookup and further initialize the incomplete jaddref created
4760 * by softdep_setup_inomapdep when the inode was allocated. Adjusts
4761 * nlinkdelta for non-journaling softdep.
4764 softdep_setup_mkdir(dp, ip)
4768 struct inodedep *inodedep;
4769 struct jaddref *dotdotaddref;
4770 struct jaddref *dotaddref;
4771 struct jaddref *jaddref;
4774 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4775 ("softdep_setup_mkdir called on non-softdep filesystem"));
4777 dotaddref = dotdotaddref = NULL;
4778 if (DOINGSUJ(dvp)) {
4779 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
4781 dotaddref->ja_state |= MKDIR_BODY;
4782 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4783 dp->i_effnlink - 1, dp->i_mode);
4784 dotdotaddref->ja_state |= MKDIR_PARENT;
4786 ACQUIRE_LOCK(dp->i_ump);
4787 inodedep = inodedep_lookup_ip(ip);
4788 if (DOINGSUJ(dvp)) {
4789 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4791 KASSERT(jaddref != NULL,
4792 ("softdep_setup_mkdir: No addref structure present."));
4793 KASSERT(jaddref->ja_parent == dp->i_number,
4794 ("softdep_setup_mkdir: bad parent %ju",
4795 (uintmax_t)jaddref->ja_parent));
4796 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
4799 inodedep = inodedep_lookup_ip(dp);
4801 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
4802 &dotdotaddref->ja_ref, if_deps);
4803 softdep_prelink(ITOV(dp), NULL);
4804 FREE_LOCK(dp->i_ump);
4808 * Called to track nlinkdelta of the inode and parent directories prior to
4809 * unlinking a directory.
4812 softdep_setup_rmdir(dp, ip)
4818 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4819 ("softdep_setup_rmdir called on non-softdep filesystem"));
4821 ACQUIRE_LOCK(dp->i_ump);
4822 (void) inodedep_lookup_ip(ip);
4823 (void) inodedep_lookup_ip(dp);
4824 softdep_prelink(dvp, ITOV(ip));
4825 FREE_LOCK(dp->i_ump);
4829 * Called to track nlinkdelta of the inode and parent directories prior to
4833 softdep_setup_unlink(dp, ip)
4839 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4840 ("softdep_setup_unlink called on non-softdep filesystem"));
4842 ACQUIRE_LOCK(dp->i_ump);
4843 (void) inodedep_lookup_ip(ip);
4844 (void) inodedep_lookup_ip(dp);
4845 softdep_prelink(dvp, ITOV(ip));
4846 FREE_LOCK(dp->i_ump);
4850 * Called to release the journal structures created by a failed non-directory
4851 * creation. Adjusts nlinkdelta for non-journaling softdep.
4854 softdep_revert_create(dp, ip)
4858 struct inodedep *inodedep;
4859 struct jaddref *jaddref;
4862 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4863 ("softdep_revert_create called on non-softdep filesystem"));
4865 ACQUIRE_LOCK(dp->i_ump);
4866 inodedep = inodedep_lookup_ip(ip);
4867 if (DOINGSUJ(dvp)) {
4868 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4870 KASSERT(jaddref->ja_parent == dp->i_number,
4871 ("softdep_revert_create: addref parent mismatch"));
4872 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4874 FREE_LOCK(dp->i_ump);
4878 * Called to release the journal structures created by a failed link
4879 * addition. Adjusts nlinkdelta for non-journaling softdep.
4882 softdep_revert_link(dp, ip)
4886 struct inodedep *inodedep;
4887 struct jaddref *jaddref;
4890 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4891 ("softdep_revert_link called on non-softdep filesystem"));
4893 ACQUIRE_LOCK(dp->i_ump);
4894 inodedep = inodedep_lookup_ip(ip);
4895 if (DOINGSUJ(dvp)) {
4896 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4898 KASSERT(jaddref->ja_parent == dp->i_number,
4899 ("softdep_revert_link: addref parent mismatch"));
4900 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4902 FREE_LOCK(dp->i_ump);
4906 * Called to release the journal structures created by a failed mkdir
4907 * attempt. Adjusts nlinkdelta for non-journaling softdep.
4910 softdep_revert_mkdir(dp, ip)
4914 struct inodedep *inodedep;
4915 struct jaddref *jaddref;
4916 struct jaddref *dotaddref;
4919 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4920 ("softdep_revert_mkdir called on non-softdep filesystem"));
4923 ACQUIRE_LOCK(dp->i_ump);
4924 inodedep = inodedep_lookup_ip(dp);
4925 if (DOINGSUJ(dvp)) {
4926 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4928 KASSERT(jaddref->ja_parent == ip->i_number,
4929 ("softdep_revert_mkdir: dotdot addref parent mismatch"));
4930 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4932 inodedep = inodedep_lookup_ip(ip);
4933 if (DOINGSUJ(dvp)) {
4934 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4936 KASSERT(jaddref->ja_parent == dp->i_number,
4937 ("softdep_revert_mkdir: addref parent mismatch"));
4938 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
4939 inoreflst, if_deps);
4940 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4941 KASSERT(dotaddref->ja_parent == ip->i_number,
4942 ("softdep_revert_mkdir: dot addref parent mismatch"));
4943 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait);
4945 FREE_LOCK(dp->i_ump);
4949 * Called to correct nlinkdelta after a failed rmdir.
4952 softdep_revert_rmdir(dp, ip)
4957 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4958 ("softdep_revert_rmdir called on non-softdep filesystem"));
4959 ACQUIRE_LOCK(dp->i_ump);
4960 (void) inodedep_lookup_ip(ip);
4961 (void) inodedep_lookup_ip(dp);
4962 FREE_LOCK(dp->i_ump);
4966 * Protecting the freemaps (or bitmaps).
4968 * To eliminate the need to execute fsck before mounting a filesystem
4969 * after a power failure, one must (conservatively) guarantee that the
4970 * on-disk copy of the bitmaps never indicate that a live inode or block is
4971 * free. So, when a block or inode is allocated, the bitmap should be
4972 * updated (on disk) before any new pointers. When a block or inode is
4973 * freed, the bitmap should not be updated until all pointers have been
4974 * reset. The latter dependency is handled by the delayed de-allocation
4975 * approach described below for block and inode de-allocation. The former
4976 * dependency is handled by calling the following procedure when a block or
4977 * inode is allocated. When an inode is allocated an "inodedep" is created
4978 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
4979 * Each "inodedep" is also inserted into the hash indexing structure so
4980 * that any additional link additions can be made dependent on the inode
4983 * The ufs filesystem maintains a number of free block counts (e.g., per
4984 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
4985 * in addition to the bitmaps. These counts are used to improve efficiency
4986 * during allocation and therefore must be consistent with the bitmaps.
4987 * There is no convenient way to guarantee post-crash consistency of these
4988 * counts with simple update ordering, for two main reasons: (1) The counts
4989 * and bitmaps for a single cylinder group block are not in the same disk
4990 * sector. If a disk write is interrupted (e.g., by power failure), one may
4991 * be written and the other not. (2) Some of the counts are located in the
4992 * superblock rather than the cylinder group block. So, we focus our soft
4993 * updates implementation on protecting the bitmaps. When mounting a
4994 * filesystem, we recompute the auxiliary counts from the bitmaps.
4998 * Called just after updating the cylinder group block to allocate an inode.
5001 softdep_setup_inomapdep(bp, ip, newinum, mode)
5002 struct buf *bp; /* buffer for cylgroup block with inode map */
5003 struct inode *ip; /* inode related to allocation */
5004 ino_t newinum; /* new inode number being allocated */
5007 struct inodedep *inodedep;
5008 struct bmsafemap *bmsafemap;
5009 struct jaddref *jaddref;
5013 mp = UFSTOVFS(ip->i_ump);
5014 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5015 ("softdep_setup_inomapdep called on non-softdep filesystem"));
5016 fs = ip->i_ump->um_fs;
5020 * Allocate the journal reference add structure so that the bitmap
5021 * can be dependent on it.
5023 if (MOUNTEDSUJ(mp)) {
5024 jaddref = newjaddref(ip, newinum, 0, 0, mode);
5025 jaddref->ja_state |= NEWBLOCK;
5029 * Create a dependency for the newly allocated inode.
5030 * Panic if it already exists as something is seriously wrong.
5031 * Otherwise add it to the dependency list for the buffer holding
5032 * the cylinder group map from which it was allocated.
5034 * We have to preallocate a bmsafemap entry in case it is needed
5035 * in bmsafemap_lookup since once we allocate the inodedep, we
5036 * have to finish initializing it before we can FREE_LOCK().
5037 * By preallocating, we avoid FREE_LOCK() while doing a malloc
5038 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before
5039 * creating the inodedep as it can be freed during the time
5040 * that we FREE_LOCK() while allocating the inodedep. We must
5041 * call workitem_alloc() before entering the locked section as
5042 * it also acquires the lock and we must avoid trying doing so
5045 bmsafemap = malloc(sizeof(struct bmsafemap),
5046 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5047 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5048 ACQUIRE_LOCK(ip->i_ump);
5049 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep)))
5050 panic("softdep_setup_inomapdep: dependency %p for new"
5051 "inode already exists", inodedep);
5052 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap);
5054 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
5055 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
5058 inodedep->id_state |= ONDEPLIST;
5059 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
5061 inodedep->id_bmsafemap = bmsafemap;
5062 inodedep->id_state &= ~DEPCOMPLETE;
5063 FREE_LOCK(ip->i_ump);
5067 * Called just after updating the cylinder group block to
5068 * allocate block or fragment.
5071 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
5072 struct buf *bp; /* buffer for cylgroup block with block map */
5073 struct mount *mp; /* filesystem doing allocation */
5074 ufs2_daddr_t newblkno; /* number of newly allocated block */
5075 int frags; /* Number of fragments. */
5076 int oldfrags; /* Previous number of fragments for extend. */
5078 struct newblk *newblk;
5079 struct bmsafemap *bmsafemap;
5080 struct jnewblk *jnewblk;
5081 struct ufsmount *ump;
5084 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5085 ("softdep_setup_blkmapdep called on non-softdep filesystem"));
5090 * Create a dependency for the newly allocated block.
5091 * Add it to the dependency list for the buffer holding
5092 * the cylinder group map from which it was allocated.
5094 if (MOUNTEDSUJ(mp)) {
5095 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
5096 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
5097 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
5098 jnewblk->jn_state = ATTACHED;
5099 jnewblk->jn_blkno = newblkno;
5100 jnewblk->jn_frags = frags;
5101 jnewblk->jn_oldfrags = oldfrags;
5109 cgp = (struct cg *)bp->b_data;
5110 blksfree = cg_blksfree(cgp);
5111 bno = dtogd(fs, jnewblk->jn_blkno);
5112 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
5114 if (isset(blksfree, bno + i))
5115 panic("softdep_setup_blkmapdep: "
5116 "free fragment %d from %d-%d "
5117 "state 0x%X dep %p", i,
5118 jnewblk->jn_oldfrags,
5128 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d",
5129 newblkno, frags, oldfrags);
5131 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
5132 panic("softdep_setup_blkmapdep: found block");
5133 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
5134 dtog(fs, newblkno), NULL);
5136 jnewblk->jn_dep = (struct worklist *)newblk;
5137 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
5139 newblk->nb_state |= ONDEPLIST;
5140 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
5142 newblk->nb_bmsafemap = bmsafemap;
5143 newblk->nb_jnewblk = jnewblk;
5147 #define BMSAFEMAP_HASH(ump, cg) \
5148 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size])
5151 bmsafemap_find(bmsafemaphd, cg, bmsafemapp)
5152 struct bmsafemap_hashhead *bmsafemaphd;
5154 struct bmsafemap **bmsafemapp;
5156 struct bmsafemap *bmsafemap;
5158 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
5159 if (bmsafemap->sm_cg == cg)
5162 *bmsafemapp = bmsafemap;
5171 * Find the bmsafemap associated with a cylinder group buffer.
5172 * If none exists, create one. The buffer must be locked when
5173 * this routine is called and this routine must be called with
5174 * the softdep lock held. To avoid giving up the lock while
5175 * allocating a new bmsafemap, a preallocated bmsafemap may be
5176 * provided. If it is provided but not needed, it is freed.
5178 static struct bmsafemap *
5179 bmsafemap_lookup(mp, bp, cg, newbmsafemap)
5183 struct bmsafemap *newbmsafemap;
5185 struct bmsafemap_hashhead *bmsafemaphd;
5186 struct bmsafemap *bmsafemap, *collision;
5187 struct worklist *wk;
5188 struct ufsmount *ump;
5192 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer"));
5193 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5194 if (wk->wk_type == D_BMSAFEMAP) {
5196 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5197 return (WK_BMSAFEMAP(wk));
5200 bmsafemaphd = BMSAFEMAP_HASH(ump, cg);
5201 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) {
5203 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5207 bmsafemap = newbmsafemap;
5210 bmsafemap = malloc(sizeof(struct bmsafemap),
5211 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5212 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5215 bmsafemap->sm_buf = bp;
5216 LIST_INIT(&bmsafemap->sm_inodedephd);
5217 LIST_INIT(&bmsafemap->sm_inodedepwr);
5218 LIST_INIT(&bmsafemap->sm_newblkhd);
5219 LIST_INIT(&bmsafemap->sm_newblkwr);
5220 LIST_INIT(&bmsafemap->sm_jaddrefhd);
5221 LIST_INIT(&bmsafemap->sm_jnewblkhd);
5222 LIST_INIT(&bmsafemap->sm_freehd);
5223 LIST_INIT(&bmsafemap->sm_freewr);
5224 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) {
5225 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
5228 bmsafemap->sm_cg = cg;
5229 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
5230 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
5231 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
5236 * Direct block allocation dependencies.
5238 * When a new block is allocated, the corresponding disk locations must be
5239 * initialized (with zeros or new data) before the on-disk inode points to
5240 * them. Also, the freemap from which the block was allocated must be
5241 * updated (on disk) before the inode's pointer. These two dependencies are
5242 * independent of each other and are needed for all file blocks and indirect
5243 * blocks that are pointed to directly by the inode. Just before the
5244 * "in-core" version of the inode is updated with a newly allocated block
5245 * number, a procedure (below) is called to setup allocation dependency
5246 * structures. These structures are removed when the corresponding
5247 * dependencies are satisfied or when the block allocation becomes obsolete
5248 * (i.e., the file is deleted, the block is de-allocated, or the block is a
5249 * fragment that gets upgraded). All of these cases are handled in
5250 * procedures described later.
5252 * When a file extension causes a fragment to be upgraded, either to a larger
5253 * fragment or to a full block, the on-disk location may change (if the
5254 * previous fragment could not simply be extended). In this case, the old
5255 * fragment must be de-allocated, but not until after the inode's pointer has
5256 * been updated. In most cases, this is handled by later procedures, which
5257 * will construct a "freefrag" structure to be added to the workitem queue
5258 * when the inode update is complete (or obsolete). The main exception to
5259 * this is when an allocation occurs while a pending allocation dependency
5260 * (for the same block pointer) remains. This case is handled in the main
5261 * allocation dependency setup procedure by immediately freeing the
5262 * unreferenced fragments.
5265 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5266 struct inode *ip; /* inode to which block is being added */
5267 ufs_lbn_t off; /* block pointer within inode */
5268 ufs2_daddr_t newblkno; /* disk block number being added */
5269 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */
5270 long newsize; /* size of new block */
5271 long oldsize; /* size of new block */
5272 struct buf *bp; /* bp for allocated block */
5274 struct allocdirect *adp, *oldadp;
5275 struct allocdirectlst *adphead;
5276 struct freefrag *freefrag;
5277 struct inodedep *inodedep;
5278 struct pagedep *pagedep;
5279 struct jnewblk *jnewblk;
5280 struct newblk *newblk;
5285 mp = UFSTOVFS(ip->i_ump);
5286 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5287 ("softdep_setup_allocdirect called on non-softdep filesystem"));
5288 if (oldblkno && oldblkno != newblkno)
5289 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5294 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd "
5295 "off %jd newsize %ld oldsize %d",
5296 ip->i_number, newblkno, oldblkno, off, newsize, oldsize);
5297 ACQUIRE_LOCK(ip->i_ump);
5298 if (off >= NDADDR) {
5300 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
5302 /* allocating an indirect block */
5304 panic("softdep_setup_allocdirect: non-zero indir");
5307 panic("softdep_setup_allocdirect: lbn %jd != off %jd",
5310 * Allocating a direct block.
5312 * If we are allocating a directory block, then we must
5313 * allocate an associated pagedep to track additions and
5316 if ((ip->i_mode & IFMT) == IFDIR)
5317 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC,
5320 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5321 panic("softdep_setup_allocdirect: lost block");
5322 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5323 ("softdep_setup_allocdirect: newblk already initialized"));
5325 * Convert the newblk to an allocdirect.
5327 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5328 adp = (struct allocdirect *)newblk;
5329 newblk->nb_freefrag = freefrag;
5330 adp->ad_offset = off;
5331 adp->ad_oldblkno = oldblkno;
5332 adp->ad_newsize = newsize;
5333 adp->ad_oldsize = oldsize;
5336 * Finish initializing the journal.
5338 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5339 jnewblk->jn_ino = ip->i_number;
5340 jnewblk->jn_lbn = lbn;
5341 add_to_journal(&jnewblk->jn_list);
5343 if (freefrag && freefrag->ff_jdep != NULL &&
5344 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5345 add_to_journal(freefrag->ff_jdep);
5346 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5347 adp->ad_inodedep = inodedep;
5349 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5351 * The list of allocdirects must be kept in sorted and ascending
5352 * order so that the rollback routines can quickly determine the
5353 * first uncommitted block (the size of the file stored on disk
5354 * ends at the end of the lowest committed fragment, or if there
5355 * are no fragments, at the end of the highest committed block).
5356 * Since files generally grow, the typical case is that the new
5357 * block is to be added at the end of the list. We speed this
5358 * special case by checking against the last allocdirect in the
5359 * list before laboriously traversing the list looking for the
5362 adphead = &inodedep->id_newinoupdt;
5363 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5364 if (oldadp == NULL || oldadp->ad_offset <= off) {
5365 /* insert at end of list */
5366 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5367 if (oldadp != NULL && oldadp->ad_offset == off)
5368 allocdirect_merge(adphead, adp, oldadp);
5369 FREE_LOCK(ip->i_ump);
5372 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5373 if (oldadp->ad_offset >= off)
5377 panic("softdep_setup_allocdirect: lost entry");
5378 /* insert in middle of list */
5379 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5380 if (oldadp->ad_offset == off)
5381 allocdirect_merge(adphead, adp, oldadp);
5383 FREE_LOCK(ip->i_ump);
5387 * Merge a newer and older journal record to be stored either in a
5388 * newblock or freefrag. This handles aggregating journal records for
5389 * fragment allocation into a second record as well as replacing a
5390 * journal free with an aborted journal allocation. A segment for the
5391 * oldest record will be placed on wkhd if it has been written. If not
5392 * the segment for the newer record will suffice.
5394 static struct worklist *
5395 jnewblk_merge(new, old, wkhd)
5396 struct worklist *new;
5397 struct worklist *old;
5398 struct workhead *wkhd;
5400 struct jnewblk *njnewblk;
5401 struct jnewblk *jnewblk;
5403 /* Handle NULLs to simplify callers. */
5408 /* Replace a jfreefrag with a jnewblk. */
5409 if (new->wk_type == D_JFREEFRAG) {
5410 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno)
5411 panic("jnewblk_merge: blkno mismatch: %p, %p",
5413 cancel_jfreefrag(WK_JFREEFRAG(new));
5416 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK)
5417 panic("jnewblk_merge: Bad type: old %d new %d\n",
5418 old->wk_type, new->wk_type);
5420 * Handle merging of two jnewblk records that describe
5421 * different sets of fragments in the same block.
5423 jnewblk = WK_JNEWBLK(old);
5424 njnewblk = WK_JNEWBLK(new);
5425 if (jnewblk->jn_blkno != njnewblk->jn_blkno)
5426 panic("jnewblk_merge: Merging disparate blocks.");
5428 * The record may be rolled back in the cg.
5430 if (jnewblk->jn_state & UNDONE) {
5431 jnewblk->jn_state &= ~UNDONE;
5432 njnewblk->jn_state |= UNDONE;
5433 njnewblk->jn_state &= ~ATTACHED;
5436 * We modify the newer addref and free the older so that if neither
5437 * has been written the most up-to-date copy will be on disk. If
5438 * both have been written but rolled back we only temporarily need
5439 * one of them to fix the bits when the cg write completes.
5441 jnewblk->jn_state |= ATTACHED | COMPLETE;
5442 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
5443 cancel_jnewblk(jnewblk, wkhd);
5444 WORKLIST_REMOVE(&jnewblk->jn_list);
5445 free_jnewblk(jnewblk);
5450 * Replace an old allocdirect dependency with a newer one.
5451 * This routine must be called with splbio interrupts blocked.
5454 allocdirect_merge(adphead, newadp, oldadp)
5455 struct allocdirectlst *adphead; /* head of list holding allocdirects */
5456 struct allocdirect *newadp; /* allocdirect being added */
5457 struct allocdirect *oldadp; /* existing allocdirect being checked */
5459 struct worklist *wk;
5460 struct freefrag *freefrag;
5463 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp));
5464 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
5465 newadp->ad_oldsize != oldadp->ad_newsize ||
5466 newadp->ad_offset >= NDADDR)
5467 panic("%s %jd != new %jd || old size %ld != new %ld",
5468 "allocdirect_merge: old blkno",
5469 (intmax_t)newadp->ad_oldblkno,
5470 (intmax_t)oldadp->ad_newblkno,
5471 newadp->ad_oldsize, oldadp->ad_newsize);
5472 newadp->ad_oldblkno = oldadp->ad_oldblkno;
5473 newadp->ad_oldsize = oldadp->ad_oldsize;
5475 * If the old dependency had a fragment to free or had never
5476 * previously had a block allocated, then the new dependency
5477 * can immediately post its freefrag and adopt the old freefrag.
5478 * This action is done by swapping the freefrag dependencies.
5479 * The new dependency gains the old one's freefrag, and the
5480 * old one gets the new one and then immediately puts it on
5481 * the worklist when it is freed by free_newblk. It is
5482 * not possible to do this swap when the old dependency had a
5483 * non-zero size but no previous fragment to free. This condition
5484 * arises when the new block is an extension of the old block.
5485 * Here, the first part of the fragment allocated to the new
5486 * dependency is part of the block currently claimed on disk by
5487 * the old dependency, so cannot legitimately be freed until the
5488 * conditions for the new dependency are fulfilled.
5490 freefrag = newadp->ad_freefrag;
5491 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
5492 newadp->ad_freefrag = oldadp->ad_freefrag;
5493 oldadp->ad_freefrag = freefrag;
5496 * If we are tracking a new directory-block allocation,
5497 * move it from the old allocdirect to the new allocdirect.
5499 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
5500 WORKLIST_REMOVE(wk);
5501 if (!LIST_EMPTY(&oldadp->ad_newdirblk))
5502 panic("allocdirect_merge: extra newdirblk");
5503 WORKLIST_INSERT(&newadp->ad_newdirblk, wk);
5505 TAILQ_REMOVE(adphead, oldadp, ad_next);
5507 * We need to move any journal dependencies over to the freefrag
5508 * that releases this block if it exists. Otherwise we are
5509 * extending an existing block and we'll wait until that is
5510 * complete to release the journal space and extend the
5511 * new journal to cover this old space as well.
5513 if (freefrag == NULL) {
5514 if (oldadp->ad_newblkno != newadp->ad_newblkno)
5515 panic("allocdirect_merge: %jd != %jd",
5516 oldadp->ad_newblkno, newadp->ad_newblkno);
5517 newadp->ad_block.nb_jnewblk = (struct jnewblk *)
5518 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list,
5519 &oldadp->ad_block.nb_jnewblk->jn_list,
5520 &newadp->ad_block.nb_jwork);
5521 oldadp->ad_block.nb_jnewblk = NULL;
5522 cancel_newblk(&oldadp->ad_block, NULL,
5523 &newadp->ad_block.nb_jwork);
5525 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block,
5526 &freefrag->ff_list, &freefrag->ff_jwork);
5527 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk,
5528 &freefrag->ff_jwork);
5530 free_newblk(&oldadp->ad_block);
5534 * Allocate a jfreefrag structure to journal a single block free.
5536 static struct jfreefrag *
5537 newjfreefrag(freefrag, ip, blkno, size, lbn)
5538 struct freefrag *freefrag;
5544 struct jfreefrag *jfreefrag;
5548 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
5550 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump));
5551 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
5552 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
5553 jfreefrag->fr_ino = ip->i_number;
5554 jfreefrag->fr_lbn = lbn;
5555 jfreefrag->fr_blkno = blkno;
5556 jfreefrag->fr_frags = numfrags(fs, size);
5557 jfreefrag->fr_freefrag = freefrag;
5563 * Allocate a new freefrag structure.
5565 static struct freefrag *
5566 newfreefrag(ip, blkno, size, lbn)
5572 struct freefrag *freefrag;
5575 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd",
5576 ip->i_number, blkno, size, lbn);
5578 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
5579 panic("newfreefrag: frag size");
5580 freefrag = malloc(sizeof(struct freefrag),
5581 M_FREEFRAG, M_SOFTDEP_FLAGS);
5582 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
5583 freefrag->ff_state = ATTACHED;
5584 LIST_INIT(&freefrag->ff_jwork);
5585 freefrag->ff_inum = ip->i_number;
5586 freefrag->ff_vtype = ITOV(ip)->v_type;
5587 freefrag->ff_blkno = blkno;
5588 freefrag->ff_fragsize = size;
5590 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) {
5591 freefrag->ff_jdep = (struct worklist *)
5592 newjfreefrag(freefrag, ip, blkno, size, lbn);
5594 freefrag->ff_state |= DEPCOMPLETE;
5595 freefrag->ff_jdep = NULL;
5602 * This workitem de-allocates fragments that were replaced during
5603 * file block allocation.
5606 handle_workitem_freefrag(freefrag)
5607 struct freefrag *freefrag;
5609 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
5610 struct workhead wkhd;
5613 "handle_workitem_freefrag: ino %d blkno %jd size %ld",
5614 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize);
5616 * It would be illegal to add new completion items to the
5617 * freefrag after it was schedule to be done so it must be
5618 * safe to modify the list head here.
5622 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
5624 * If the journal has not been written we must cancel it here.
5626 if (freefrag->ff_jdep) {
5627 if (freefrag->ff_jdep->wk_type != D_JNEWBLK)
5628 panic("handle_workitem_freefrag: Unexpected type %d\n",
5629 freefrag->ff_jdep->wk_type);
5630 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd);
5633 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
5634 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd);
5636 WORKITEM_FREE(freefrag, D_FREEFRAG);
5641 * Set up a dependency structure for an external attributes data block.
5642 * This routine follows much of the structure of softdep_setup_allocdirect.
5643 * See the description of softdep_setup_allocdirect above for details.
5646 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5649 ufs2_daddr_t newblkno;
5650 ufs2_daddr_t oldblkno;
5655 struct allocdirect *adp, *oldadp;
5656 struct allocdirectlst *adphead;
5657 struct freefrag *freefrag;
5658 struct inodedep *inodedep;
5659 struct jnewblk *jnewblk;
5660 struct newblk *newblk;
5664 mp = UFSTOVFS(ip->i_ump);
5665 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5666 ("softdep_setup_allocext called on non-softdep filesystem"));
5667 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR",
5671 if (oldblkno && oldblkno != newblkno)
5672 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5676 ACQUIRE_LOCK(ip->i_ump);
5677 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5678 panic("softdep_setup_allocext: lost block");
5679 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5680 ("softdep_setup_allocext: newblk already initialized"));
5682 * Convert the newblk to an allocdirect.
5684 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5685 adp = (struct allocdirect *)newblk;
5686 newblk->nb_freefrag = freefrag;
5687 adp->ad_offset = off;
5688 adp->ad_oldblkno = oldblkno;
5689 adp->ad_newsize = newsize;
5690 adp->ad_oldsize = oldsize;
5691 adp->ad_state |= EXTDATA;
5694 * Finish initializing the journal.
5696 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5697 jnewblk->jn_ino = ip->i_number;
5698 jnewblk->jn_lbn = lbn;
5699 add_to_journal(&jnewblk->jn_list);
5701 if (freefrag && freefrag->ff_jdep != NULL &&
5702 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5703 add_to_journal(freefrag->ff_jdep);
5704 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5705 adp->ad_inodedep = inodedep;
5707 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5709 * The list of allocdirects must be kept in sorted and ascending
5710 * order so that the rollback routines can quickly determine the
5711 * first uncommitted block (the size of the file stored on disk
5712 * ends at the end of the lowest committed fragment, or if there
5713 * are no fragments, at the end of the highest committed block).
5714 * Since files generally grow, the typical case is that the new
5715 * block is to be added at the end of the list. We speed this
5716 * special case by checking against the last allocdirect in the
5717 * list before laboriously traversing the list looking for the
5720 adphead = &inodedep->id_newextupdt;
5721 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5722 if (oldadp == NULL || oldadp->ad_offset <= off) {
5723 /* insert at end of list */
5724 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5725 if (oldadp != NULL && oldadp->ad_offset == off)
5726 allocdirect_merge(adphead, adp, oldadp);
5727 FREE_LOCK(ip->i_ump);
5730 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5731 if (oldadp->ad_offset >= off)
5735 panic("softdep_setup_allocext: lost entry");
5736 /* insert in middle of list */
5737 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5738 if (oldadp->ad_offset == off)
5739 allocdirect_merge(adphead, adp, oldadp);
5740 FREE_LOCK(ip->i_ump);
5744 * Indirect block allocation dependencies.
5746 * The same dependencies that exist for a direct block also exist when
5747 * a new block is allocated and pointed to by an entry in a block of
5748 * indirect pointers. The undo/redo states described above are also
5749 * used here. Because an indirect block contains many pointers that
5750 * may have dependencies, a second copy of the entire in-memory indirect
5751 * block is kept. The buffer cache copy is always completely up-to-date.
5752 * The second copy, which is used only as a source for disk writes,
5753 * contains only the safe pointers (i.e., those that have no remaining
5754 * update dependencies). The second copy is freed when all pointers
5755 * are safe. The cache is not allowed to replace indirect blocks with
5756 * pending update dependencies. If a buffer containing an indirect
5757 * block with dependencies is written, these routines will mark it
5758 * dirty again. It can only be successfully written once all the
5759 * dependencies are removed. The ffs_fsync routine in conjunction with
5760 * softdep_sync_metadata work together to get all the dependencies
5761 * removed so that a file can be successfully written to disk. Three
5762 * procedures are used when setting up indirect block pointer
5763 * dependencies. The division is necessary because of the organization
5764 * of the "balloc" routine and because of the distinction between file
5765 * pages and file metadata blocks.
5769 * Allocate a new allocindir structure.
5771 static struct allocindir *
5772 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
5773 struct inode *ip; /* inode for file being extended */
5774 int ptrno; /* offset of pointer in indirect block */
5775 ufs2_daddr_t newblkno; /* disk block number being added */
5776 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5779 struct newblk *newblk;
5780 struct allocindir *aip;
5781 struct freefrag *freefrag;
5782 struct jnewblk *jnewblk;
5785 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn);
5788 ACQUIRE_LOCK(ip->i_ump);
5789 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0)
5790 panic("new_allocindir: lost block");
5791 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5792 ("newallocindir: newblk already initialized"));
5793 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR);
5794 newblk->nb_freefrag = freefrag;
5795 aip = (struct allocindir *)newblk;
5796 aip->ai_offset = ptrno;
5797 aip->ai_oldblkno = oldblkno;
5799 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5800 jnewblk->jn_ino = ip->i_number;
5801 jnewblk->jn_lbn = lbn;
5802 add_to_journal(&jnewblk->jn_list);
5804 if (freefrag && freefrag->ff_jdep != NULL &&
5805 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5806 add_to_journal(freefrag->ff_jdep);
5811 * Called just before setting an indirect block pointer
5812 * to a newly allocated file page.
5815 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
5816 struct inode *ip; /* inode for file being extended */
5817 ufs_lbn_t lbn; /* allocated block number within file */
5818 struct buf *bp; /* buffer with indirect blk referencing page */
5819 int ptrno; /* offset of pointer in indirect block */
5820 ufs2_daddr_t newblkno; /* disk block number being added */
5821 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5822 struct buf *nbp; /* buffer holding allocated page */
5824 struct inodedep *inodedep;
5825 struct freefrag *freefrag;
5826 struct allocindir *aip;
5827 struct pagedep *pagedep;
5830 mp = UFSTOVFS(ip->i_ump);
5831 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5832 ("softdep_setup_allocindir_page called on non-softdep filesystem"));
5833 KASSERT(lbn == nbp->b_lblkno,
5834 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
5835 lbn, bp->b_lblkno));
5837 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd "
5838 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn);
5839 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
5840 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
5841 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5843 * If we are allocating a directory page, then we must
5844 * allocate an associated pagedep to track additions and
5847 if ((ip->i_mode & IFMT) == IFDIR)
5848 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep);
5849 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5850 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
5851 FREE_LOCK(ip->i_ump);
5853 handle_workitem_freefrag(freefrag);
5857 * Called just before setting an indirect block pointer to a
5858 * newly allocated indirect block.
5861 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
5862 struct buf *nbp; /* newly allocated indirect block */
5863 struct inode *ip; /* inode for file being extended */
5864 struct buf *bp; /* indirect block referencing allocated block */
5865 int ptrno; /* offset of pointer in indirect block */
5866 ufs2_daddr_t newblkno; /* disk block number being added */
5868 struct inodedep *inodedep;
5869 struct allocindir *aip;
5872 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
5873 ("softdep_setup_allocindir_meta called on non-softdep filesystem"));
5875 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d",
5876 ip->i_number, newblkno, ptrno);
5877 lbn = nbp->b_lblkno;
5878 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
5879 aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
5880 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
5882 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5883 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
5884 panic("softdep_setup_allocindir_meta: Block already existed");
5885 FREE_LOCK(ip->i_ump);
5889 indirdep_complete(indirdep)
5890 struct indirdep *indirdep;
5892 struct allocindir *aip;
5894 LIST_REMOVE(indirdep, ir_next);
5895 indirdep->ir_state |= DEPCOMPLETE;
5897 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
5898 LIST_REMOVE(aip, ai_next);
5899 free_newblk(&aip->ai_block);
5902 * If this indirdep is not attached to a buf it was simply waiting
5903 * on completion to clear completehd. free_indirdep() asserts
5904 * that nothing is dangling.
5906 if ((indirdep->ir_state & ONWORKLIST) == 0)
5907 free_indirdep(indirdep);
5910 static struct indirdep *
5911 indirdep_lookup(mp, ip, bp)
5916 struct indirdep *indirdep, *newindirdep;
5917 struct newblk *newblk;
5918 struct ufsmount *ump;
5919 struct worklist *wk;
5929 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5930 if (wk->wk_type != D_INDIRDEP)
5932 indirdep = WK_INDIRDEP(wk);
5935 /* Found on the buffer worklist, no new structure to free. */
5936 if (indirdep != NULL && newindirdep == NULL)
5938 if (indirdep != NULL && newindirdep != NULL)
5939 panic("indirdep_lookup: simultaneous create");
5940 /* None found on the buffer and a new structure is ready. */
5941 if (indirdep == NULL && newindirdep != NULL)
5943 /* None found and no new structure available. */
5945 newindirdep = malloc(sizeof(struct indirdep),
5946 M_INDIRDEP, M_SOFTDEP_FLAGS);
5947 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
5948 newindirdep->ir_state = ATTACHED;
5949 if (ip->i_ump->um_fstype == UFS1)
5950 newindirdep->ir_state |= UFS1FMT;
5951 TAILQ_INIT(&newindirdep->ir_trunc);
5952 newindirdep->ir_saveddata = NULL;
5953 LIST_INIT(&newindirdep->ir_deplisthd);
5954 LIST_INIT(&newindirdep->ir_donehd);
5955 LIST_INIT(&newindirdep->ir_writehd);
5956 LIST_INIT(&newindirdep->ir_completehd);
5957 if (bp->b_blkno == bp->b_lblkno) {
5958 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
5960 bp->b_blkno = blkno;
5962 newindirdep->ir_freeblks = NULL;
5963 newindirdep->ir_savebp =
5964 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
5965 newindirdep->ir_bp = bp;
5966 BUF_KERNPROC(newindirdep->ir_savebp);
5967 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
5970 indirdep = newindirdep;
5971 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
5973 * If the block is not yet allocated we don't set DEPCOMPLETE so
5974 * that we don't free dependencies until the pointers are valid.
5975 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather
5976 * than using the hash.
5978 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk))
5979 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next);
5981 indirdep->ir_state |= DEPCOMPLETE;
5986 * Called to finish the allocation of the "aip" allocated
5987 * by one of the two routines above.
5989 static struct freefrag *
5990 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
5991 struct buf *bp; /* in-memory copy of the indirect block */
5992 struct inode *ip; /* inode for file being extended */
5993 struct inodedep *inodedep; /* Inodedep for ip */
5994 struct allocindir *aip; /* allocindir allocated by the above routines */
5995 ufs_lbn_t lbn; /* Logical block number for this block. */
5998 struct indirdep *indirdep;
5999 struct allocindir *oldaip;
6000 struct freefrag *freefrag;
6003 LOCK_OWNED(ip->i_ump);
6004 mp = UFSTOVFS(ip->i_ump);
6006 if (bp->b_lblkno >= 0)
6007 panic("setup_allocindir_phase2: not indir blk");
6008 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs),
6009 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset));
6010 indirdep = indirdep_lookup(mp, ip, bp);
6011 KASSERT(indirdep->ir_savebp != NULL,
6012 ("setup_allocindir_phase2 NULL ir_savebp"));
6013 aip->ai_indirdep = indirdep;
6015 * Check for an unwritten dependency for this indirect offset. If
6016 * there is, merge the old dependency into the new one. This happens
6017 * as a result of reallocblk only.
6020 if (aip->ai_oldblkno != 0) {
6021 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) {
6022 if (oldaip->ai_offset == aip->ai_offset) {
6023 freefrag = allocindir_merge(aip, oldaip);
6027 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) {
6028 if (oldaip->ai_offset == aip->ai_offset) {
6029 freefrag = allocindir_merge(aip, oldaip);
6035 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
6040 * Merge two allocindirs which refer to the same block. Move newblock
6041 * dependencies and setup the freefrags appropriately.
6043 static struct freefrag *
6044 allocindir_merge(aip, oldaip)
6045 struct allocindir *aip;
6046 struct allocindir *oldaip;
6048 struct freefrag *freefrag;
6049 struct worklist *wk;
6051 if (oldaip->ai_newblkno != aip->ai_oldblkno)
6052 panic("allocindir_merge: blkno");
6053 aip->ai_oldblkno = oldaip->ai_oldblkno;
6054 freefrag = aip->ai_freefrag;
6055 aip->ai_freefrag = oldaip->ai_freefrag;
6056 oldaip->ai_freefrag = NULL;
6057 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
6059 * If we are tracking a new directory-block allocation,
6060 * move it from the old allocindir to the new allocindir.
6062 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
6063 WORKLIST_REMOVE(wk);
6064 if (!LIST_EMPTY(&oldaip->ai_newdirblk))
6065 panic("allocindir_merge: extra newdirblk");
6066 WORKLIST_INSERT(&aip->ai_newdirblk, wk);
6069 * We can skip journaling for this freefrag and just complete
6070 * any pending journal work for the allocindir that is being
6071 * removed after the freefrag completes.
6073 if (freefrag->ff_jdep)
6074 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep));
6075 LIST_REMOVE(oldaip, ai_next);
6076 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block,
6077 &freefrag->ff_list, &freefrag->ff_jwork);
6078 free_newblk(&oldaip->ai_block);
6084 setup_freedirect(freeblks, ip, i, needj)
6085 struct freeblks *freeblks;
6093 blkno = DIP(ip, i_db[i]);
6096 DIP_SET(ip, i_db[i], 0);
6097 frags = sblksize(ip->i_fs, ip->i_size, i);
6098 frags = numfrags(ip->i_fs, frags);
6099 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj);
6103 setup_freeext(freeblks, ip, i, needj)
6104 struct freeblks *freeblks;
6112 blkno = ip->i_din2->di_extb[i];
6115 ip->i_din2->di_extb[i] = 0;
6116 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i);
6117 frags = numfrags(ip->i_fs, frags);
6118 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj);
6122 setup_freeindir(freeblks, ip, i, lbn, needj)
6123 struct freeblks *freeblks;
6131 blkno = DIP(ip, i_ib[i]);
6134 DIP_SET(ip, i_ib[i], 0);
6135 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag,
6139 static inline struct freeblks *
6144 struct freeblks *freeblks;
6146 freeblks = malloc(sizeof(struct freeblks),
6147 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
6148 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
6149 LIST_INIT(&freeblks->fb_jblkdephd);
6150 LIST_INIT(&freeblks->fb_jwork);
6151 freeblks->fb_ref = 0;
6152 freeblks->fb_cgwait = 0;
6153 freeblks->fb_state = ATTACHED;
6154 freeblks->fb_uid = ip->i_uid;
6155 freeblks->fb_inum = ip->i_number;
6156 freeblks->fb_vtype = ITOV(ip)->v_type;
6157 freeblks->fb_modrev = DIP(ip, i_modrev);
6158 freeblks->fb_devvp = ip->i_devvp;
6159 freeblks->fb_chkcnt = 0;
6160 freeblks->fb_len = 0;
6166 trunc_indirdep(indirdep, freeblks, bp, off)
6167 struct indirdep *indirdep;
6168 struct freeblks *freeblks;
6172 struct allocindir *aip, *aipn;
6175 * The first set of allocindirs won't be in savedbp.
6177 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn)
6178 if (aip->ai_offset > off)
6179 cancel_allocindir(aip, bp, freeblks, 1);
6180 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn)
6181 if (aip->ai_offset > off)
6182 cancel_allocindir(aip, bp, freeblks, 1);
6184 * These will exist in savedbp.
6186 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn)
6187 if (aip->ai_offset > off)
6188 cancel_allocindir(aip, NULL, freeblks, 0);
6189 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn)
6190 if (aip->ai_offset > off)
6191 cancel_allocindir(aip, NULL, freeblks, 0);
6195 * Follow the chain of indirects down to lastlbn creating a freework
6196 * structure for each. This will be used to start indir_trunc() at
6197 * the right offset and create the journal records for the parrtial
6198 * truncation. A second step will handle the truncated dependencies.
6201 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno)
6202 struct freeblks *freeblks;
6208 struct indirdep *indirdep;
6209 struct indirdep *indirn;
6210 struct freework *freework;
6211 struct newblk *newblk;
6225 mp = freeblks->fb_list.wk_mp;
6226 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
6227 if ((bp->b_flags & B_CACHE) == 0) {
6228 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno);
6229 bp->b_iocmd = BIO_READ;
6230 bp->b_flags &= ~B_INVAL;
6231 bp->b_ioflags &= ~BIO_ERROR;
6232 vfs_busy_pages(bp, 0);
6233 bp->b_iooffset = dbtob(bp->b_blkno);
6235 curthread->td_ru.ru_inblock++;
6236 error = bufwait(bp);
6242 level = lbn_level(lbn);
6243 lbnadd = lbn_offset(ip->i_fs, level);
6245 * Compute the offset of the last block we want to keep. Store
6246 * in the freework the first block we want to completely free.
6248 off = (lastlbn - -(lbn + level)) / lbnadd;
6249 if (off + 1 == NINDIR(ip->i_fs))
6251 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1,
6254 * Link the freework into the indirdep. This will prevent any new
6255 * allocations from proceeding until we are finished with the
6256 * truncate and the block is written.
6258 ACQUIRE_LOCK(ip->i_ump);
6259 indirdep = indirdep_lookup(mp, ip, bp);
6260 if (indirdep->ir_freeblks)
6261 panic("setup_trunc_indir: indirdep already truncated.");
6262 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next);
6263 freework->fw_indir = indirdep;
6265 * Cancel any allocindirs that will not make it to disk.
6266 * We have to do this for all copies of the indirdep that
6267 * live on this newblk.
6269 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
6270 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk);
6271 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next)
6272 trunc_indirdep(indirn, freeblks, bp, off);
6274 trunc_indirdep(indirdep, freeblks, bp, off);
6275 FREE_LOCK(ip->i_ump);
6277 * Creation is protected by the buf lock. The saveddata is only
6278 * needed if a full truncation follows a partial truncation but it
6279 * is difficult to allocate in that case so we fetch it anyway.
6281 if (indirdep->ir_saveddata == NULL)
6282 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
6285 /* Fetch the blkno of the child and the zero start offset. */
6286 if (ip->i_ump->um_fstype == UFS1) {
6287 blkno = ((ufs1_daddr_t *)bp->b_data)[off];
6288 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1];
6290 blkno = ((ufs2_daddr_t *)bp->b_data)[off];
6291 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1];
6294 /* Zero the truncated pointers. */
6295 end = bp->b_data + bp->b_bcount;
6296 bzero(start, end - start);
6302 lbn++; /* adjust level */
6303 lbn -= (off * lbnadd);
6304 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno);
6308 * Complete the partial truncation of an indirect block setup by
6309 * setup_trunc_indir(). This zeros the truncated pointers in the saved
6310 * copy and writes them to disk before the freeblks is allowed to complete.
6313 complete_trunc_indir(freework)
6314 struct freework *freework;
6316 struct freework *fwn;
6317 struct indirdep *indirdep;
6318 struct ufsmount *ump;
6323 ump = VFSTOUFS(freework->fw_list.wk_mp);
6325 indirdep = freework->fw_indir;
6327 bp = indirdep->ir_bp;
6328 /* See if the block was discarded. */
6331 /* Inline part of getdirtybuf(). We dont want bremfree. */
6332 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0)
6334 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
6335 LOCK_PTR(ump)) == 0)
6339 freework->fw_state |= DEPCOMPLETE;
6340 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next);
6342 * Zero the pointers in the saved copy.
6344 if (indirdep->ir_state & UFS1FMT)
6345 start = sizeof(ufs1_daddr_t);
6347 start = sizeof(ufs2_daddr_t);
6348 start *= freework->fw_start;
6349 count = indirdep->ir_savebp->b_bcount - start;
6350 start += (uintptr_t)indirdep->ir_savebp->b_data;
6351 bzero((char *)start, count);
6353 * We need to start the next truncation in the list if it has not
6356 fwn = TAILQ_FIRST(&indirdep->ir_trunc);
6358 if (fwn->fw_freeblks == indirdep->ir_freeblks)
6359 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next);
6360 if ((fwn->fw_state & ONWORKLIST) == 0)
6361 freework_enqueue(fwn);
6364 * If bp is NULL the block was fully truncated, restore
6365 * the saved block list otherwise free it if it is no
6368 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
6370 bcopy(indirdep->ir_saveddata,
6371 indirdep->ir_savebp->b_data,
6372 indirdep->ir_savebp->b_bcount);
6373 free(indirdep->ir_saveddata, M_INDIRDEP);
6374 indirdep->ir_saveddata = NULL;
6377 * When bp is NULL there is a full truncation pending. We
6378 * must wait for this full truncation to be journaled before
6379 * we can release this freework because the disk pointers will
6380 * never be written as zero.
6383 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd))
6384 handle_written_freework(freework);
6386 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd,
6387 &freework->fw_list);
6389 /* Complete when the real copy is written. */
6390 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list);
6396 * Calculate the number of blocks we are going to release where datablocks
6397 * is the current total and length is the new file size.
6400 blkcount(fs, datablocks, length)
6402 ufs2_daddr_t datablocks;
6405 off_t totblks, numblks;
6408 numblks = howmany(length, fs->fs_bsize);
6409 if (numblks <= NDADDR) {
6410 totblks = howmany(length, fs->fs_fsize);
6413 totblks = blkstofrags(fs, numblks);
6416 * Count all single, then double, then triple indirects required.
6417 * Subtracting one indirects worth of blocks for each pass
6418 * acknowledges one of each pointed to by the inode.
6421 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs)));
6422 numblks -= NINDIR(fs);
6425 numblks = howmany(numblks, NINDIR(fs));
6428 totblks = fsbtodb(fs, totblks);
6430 * Handle sparse files. We can't reclaim more blocks than the inode
6431 * references. We will correct it later in handle_complete_freeblks()
6432 * when we know the real count.
6434 if (totblks > datablocks)
6436 return (datablocks - totblks);
6440 * Handle freeblocks for journaled softupdate filesystems.
6442 * Contrary to normal softupdates, we must preserve the block pointers in
6443 * indirects until their subordinates are free. This is to avoid journaling
6444 * every block that is freed which may consume more space than the journal
6445 * itself. The recovery program will see the free block journals at the
6446 * base of the truncated area and traverse them to reclaim space. The
6447 * pointers in the inode may be cleared immediately after the journal
6448 * records are written because each direct and indirect pointer in the
6449 * inode is recorded in a journal. This permits full truncation to proceed
6450 * asynchronously. The write order is journal -> inode -> cgs -> indirects.
6452 * The algorithm is as follows:
6453 * 1) Traverse the in-memory state and create journal entries to release
6454 * the relevant blocks and full indirect trees.
6455 * 2) Traverse the indirect block chain adding partial truncation freework
6456 * records to indirects in the path to lastlbn. The freework will
6457 * prevent new allocation dependencies from being satisfied in this
6458 * indirect until the truncation completes.
6459 * 3) Read and lock the inode block, performing an update with the new size
6460 * and pointers. This prevents truncated data from becoming valid on
6461 * disk through step 4.
6462 * 4) Reap unsatisfied dependencies that are beyond the truncated area,
6463 * eliminate journal work for those records that do not require it.
6464 * 5) Schedule the journal records to be written followed by the inode block.
6465 * 6) Allocate any necessary frags for the end of file.
6466 * 7) Zero any partially truncated blocks.
6468 * From this truncation proceeds asynchronously using the freework and
6469 * indir_trunc machinery. The file will not be extended again into a
6470 * partially truncated indirect block until all work is completed but
6471 * the normal dependency mechanism ensures that it is rolled back/forward
6472 * as appropriate. Further truncation may occur without delay and is
6473 * serialized in indir_trunc().
6476 softdep_journal_freeblocks(ip, cred, length, flags)
6477 struct inode *ip; /* The inode whose length is to be reduced */
6479 off_t length; /* The new length for the file */
6480 int flags; /* IO_EXT and/or IO_NORMAL */
6482 struct freeblks *freeblks, *fbn;
6483 struct worklist *wk, *wkn;
6484 struct inodedep *inodedep;
6485 struct jblkdep *jblkdep;
6486 struct allocdirect *adp, *adpn;
6487 struct ufsmount *ump;
6492 ufs2_daddr_t extblocks, datablocks;
6493 ufs_lbn_t tmpval, lbn, lastlbn;
6494 int frags, lastoff, iboff, allocblock, needj, error, i;
6499 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6500 ("softdep_journal_freeblocks called on non-softdep filesystem"));
6508 freeblks = newfreeblks(mp, ip);
6511 * If we're truncating a removed file that will never be written
6512 * we don't need to journal the block frees. The canceled journals
6513 * for the allocations will suffice.
6515 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6516 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
6519 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d",
6520 ip->i_number, length, needj);
6523 * Calculate the lbn that we are truncating to. This results in -1
6524 * if we're truncating the 0 bytes. So it is the last lbn we want
6525 * to keep, not the first lbn we want to truncate.
6527 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1;
6528 lastoff = blkoff(fs, length);
6530 * Compute frags we are keeping in lastlbn. 0 means all.
6532 if (lastlbn >= 0 && lastlbn < NDADDR) {
6533 frags = fragroundup(fs, lastoff);
6534 /* adp offset of last valid allocdirect. */
6536 } else if (lastlbn > 0)
6538 if (fs->fs_magic == FS_UFS2_MAGIC)
6539 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6541 * Handle normal data blocks and indirects. This section saves
6542 * values used after the inode update to complete frag and indirect
6545 if ((flags & IO_NORMAL) != 0) {
6547 * Handle truncation of whole direct and indirect blocks.
6549 for (i = iboff + 1; i < NDADDR; i++)
6550 setup_freedirect(freeblks, ip, i, needj);
6551 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6552 i++, lbn += tmpval, tmpval *= NINDIR(fs)) {
6553 /* Release a whole indirect tree. */
6554 if (lbn > lastlbn) {
6555 setup_freeindir(freeblks, ip, i, -lbn -i,
6561 * Traverse partially truncated indirect tree.
6563 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn)
6564 setup_trunc_indir(freeblks, ip, -lbn - i,
6565 lastlbn, DIP(ip, i_ib[i]));
6568 * Handle partial truncation to a frag boundary.
6574 oldfrags = blksize(fs, ip, lastlbn);
6575 blkno = DIP(ip, i_db[lastlbn]);
6576 if (blkno && oldfrags != frags) {
6578 oldfrags = numfrags(ip->i_fs, oldfrags);
6579 blkno += numfrags(ip->i_fs, frags);
6580 newfreework(ump, freeblks, NULL, lastlbn,
6581 blkno, oldfrags, 0, needj);
6583 adjust_newfreework(freeblks,
6584 numfrags(ip->i_fs, frags));
6585 } else if (blkno == 0)
6589 * Add a journal record for partial truncate if we are
6590 * handling indirect blocks. Non-indirects need no extra
6593 if (length != 0 && lastlbn >= NDADDR) {
6594 ip->i_flag |= IN_TRUNCATED;
6595 newjtrunc(freeblks, length, 0);
6597 ip->i_size = length;
6598 DIP_SET(ip, i_size, ip->i_size);
6599 datablocks = DIP(ip, i_blocks) - extblocks;
6601 datablocks = blkcount(ip->i_fs, datablocks, length);
6602 freeblks->fb_len = length;
6604 if ((flags & IO_EXT) != 0) {
6605 for (i = 0; i < NXADDR; i++)
6606 setup_freeext(freeblks, ip, i, needj);
6607 ip->i_din2->di_extsize = 0;
6608 datablocks += extblocks;
6611 /* Reference the quotas in case the block count is wrong in the end. */
6612 quotaref(vp, freeblks->fb_quota);
6613 (void) chkdq(ip, -datablocks, NOCRED, 0);
6615 freeblks->fb_chkcnt = -datablocks;
6617 fs->fs_pendingblocks += datablocks;
6619 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6621 * Handle truncation of incomplete alloc direct dependencies. We
6622 * hold the inode block locked to prevent incomplete dependencies
6623 * from reaching the disk while we are eliminating those that
6624 * have been truncated. This is a partially inlined ffs_update().
6627 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
6628 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6629 (int)fs->fs_bsize, cred, &bp);
6632 softdep_error("softdep_journal_freeblocks", error);
6635 if (bp->b_bufsize == fs->fs_bsize)
6636 bp->b_flags |= B_CLUSTEROK;
6637 softdep_update_inodeblock(ip, bp, 0);
6638 if (ump->um_fstype == UFS1)
6639 *((struct ufs1_dinode *)bp->b_data +
6640 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
6642 *((struct ufs2_dinode *)bp->b_data +
6643 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
6645 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6646 if ((inodedep->id_state & IOSTARTED) != 0)
6647 panic("softdep_setup_freeblocks: inode busy");
6649 * Add the freeblks structure to the list of operations that
6650 * must await the zero'ed inode being written to disk. If we
6651 * still have a bitmap dependency (needj), then the inode
6652 * has never been written to disk, so we can process the
6653 * freeblks below once we have deleted the dependencies.
6656 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6658 freeblks->fb_state |= COMPLETE;
6659 if ((flags & IO_NORMAL) != 0) {
6660 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) {
6661 if (adp->ad_offset > iboff)
6662 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6665 * Truncate the allocdirect. We could eliminate
6666 * or modify journal records as well.
6668 else if (adp->ad_offset == iboff && frags)
6669 adp->ad_newsize = frags;
6672 if ((flags & IO_EXT) != 0)
6673 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6674 cancel_allocdirect(&inodedep->id_extupdt, adp,
6677 * Scan the bufwait list for newblock dependencies that will never
6680 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) {
6681 if (wk->wk_type != D_ALLOCDIRECT)
6683 adp = WK_ALLOCDIRECT(wk);
6684 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) ||
6685 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) {
6686 cancel_jfreeblk(freeblks, adp->ad_newblkno);
6687 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork);
6688 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
6694 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps)
6695 add_to_journal(&jblkdep->jb_list);
6699 * Truncate dependency structures beyond length.
6701 trunc_dependencies(ip, freeblks, lastlbn, frags, flags);
6703 * This is only set when we need to allocate a fragment because
6704 * none existed at the end of a frag-sized file. It handles only
6705 * allocating a new, zero filled block.
6708 ip->i_size = length - lastoff;
6709 DIP_SET(ip, i_size, ip->i_size);
6710 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp);
6712 softdep_error("softdep_journal_freeblks", error);
6715 ip->i_size = length;
6716 DIP_SET(ip, i_size, length);
6717 ip->i_flag |= IN_CHANGE | IN_UPDATE;
6718 allocbuf(bp, frags);
6721 } else if (lastoff != 0 && vp->v_type != VDIR) {
6725 * Zero the end of a truncated frag or block.
6727 size = sblksize(fs, length, lastlbn);
6728 error = bread(vp, lastlbn, size, cred, &bp);
6730 softdep_error("softdep_journal_freeblks", error);
6733 bzero((char *)bp->b_data + lastoff, size - lastoff);
6738 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6739 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
6740 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
6742 * We zero earlier truncations so they don't erroneously
6745 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0)
6746 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next)
6748 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE &&
6749 LIST_EMPTY(&freeblks->fb_jblkdephd))
6750 freeblks->fb_state |= INPROGRESS;
6755 handle_workitem_freeblocks(freeblks, 0);
6756 trunc_pages(ip, length, extblocks, flags);
6761 * Flush a JOP_SYNC to the journal.
6764 softdep_journal_fsync(ip)
6767 struct jfsync *jfsync;
6769 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
6770 ("softdep_journal_fsync called on non-softdep filesystem"));
6771 if ((ip->i_flag & IN_TRUNCATED) == 0)
6773 ip->i_flag &= ~IN_TRUNCATED;
6774 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO);
6775 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump));
6776 jfsync->jfs_size = ip->i_size;
6777 jfsync->jfs_ino = ip->i_number;
6778 ACQUIRE_LOCK(ip->i_ump);
6779 add_to_journal(&jfsync->jfs_list);
6780 jwait(&jfsync->jfs_list, MNT_WAIT);
6781 FREE_LOCK(ip->i_ump);
6785 * Block de-allocation dependencies.
6787 * When blocks are de-allocated, the on-disk pointers must be nullified before
6788 * the blocks are made available for use by other files. (The true
6789 * requirement is that old pointers must be nullified before new on-disk
6790 * pointers are set. We chose this slightly more stringent requirement to
6791 * reduce complexity.) Our implementation handles this dependency by updating
6792 * the inode (or indirect block) appropriately but delaying the actual block
6793 * de-allocation (i.e., freemap and free space count manipulation) until
6794 * after the updated versions reach stable storage. After the disk is
6795 * updated, the blocks can be safely de-allocated whenever it is convenient.
6796 * This implementation handles only the common case of reducing a file's
6797 * length to zero. Other cases are handled by the conventional synchronous
6800 * The ffs implementation with which we worked double-checks
6801 * the state of the block pointers and file size as it reduces
6802 * a file's length. Some of this code is replicated here in our
6803 * soft updates implementation. The freeblks->fb_chkcnt field is
6804 * used to transfer a part of this information to the procedure
6805 * that eventually de-allocates the blocks.
6807 * This routine should be called from the routine that shortens
6808 * a file's length, before the inode's size or block pointers
6809 * are modified. It will save the block pointer information for
6810 * later release and zero the inode so that the calling routine
6814 softdep_setup_freeblocks(ip, length, flags)
6815 struct inode *ip; /* The inode whose length is to be reduced */
6816 off_t length; /* The new length for the file */
6817 int flags; /* IO_EXT and/or IO_NORMAL */
6819 struct ufs1_dinode *dp1;
6820 struct ufs2_dinode *dp2;
6821 struct freeblks *freeblks;
6822 struct inodedep *inodedep;
6823 struct allocdirect *adp;
6824 struct ufsmount *ump;
6827 ufs2_daddr_t extblocks, datablocks;
6829 int i, delay, error;
6835 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6836 ("softdep_setup_freeblocks called on non-softdep filesystem"));
6837 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld",
6838 ip->i_number, length);
6839 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length"));
6841 freeblks = newfreeblks(mp, ip);
6844 if (fs->fs_magic == FS_UFS2_MAGIC)
6845 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6846 if ((flags & IO_NORMAL) != 0) {
6847 for (i = 0; i < NDADDR; i++)
6848 setup_freedirect(freeblks, ip, i, 0);
6849 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6850 i++, lbn += tmpval, tmpval *= NINDIR(fs))
6851 setup_freeindir(freeblks, ip, i, -lbn -i, 0);
6853 DIP_SET(ip, i_size, 0);
6854 datablocks = DIP(ip, i_blocks) - extblocks;
6856 if ((flags & IO_EXT) != 0) {
6857 for (i = 0; i < NXADDR; i++)
6858 setup_freeext(freeblks, ip, i, 0);
6859 ip->i_din2->di_extsize = 0;
6860 datablocks += extblocks;
6863 /* Reference the quotas in case the block count is wrong in the end. */
6864 quotaref(ITOV(ip), freeblks->fb_quota);
6865 (void) chkdq(ip, -datablocks, NOCRED, 0);
6867 freeblks->fb_chkcnt = -datablocks;
6869 fs->fs_pendingblocks += datablocks;
6871 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6873 * Push the zero'ed inode to to its disk buffer so that we are free
6874 * to delete its dependencies below. Once the dependencies are gone
6875 * the buffer can be safely released.
6877 if ((error = bread(ip->i_devvp,
6878 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6879 (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
6881 softdep_error("softdep_setup_freeblocks", error);
6883 if (ump->um_fstype == UFS1) {
6884 dp1 = ((struct ufs1_dinode *)bp->b_data +
6885 ino_to_fsbo(fs, ip->i_number));
6886 ip->i_din1->di_freelink = dp1->di_freelink;
6889 dp2 = ((struct ufs2_dinode *)bp->b_data +
6890 ino_to_fsbo(fs, ip->i_number));
6891 ip->i_din2->di_freelink = dp2->di_freelink;
6895 * Find and eliminate any inode dependencies.
6898 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6899 if ((inodedep->id_state & IOSTARTED) != 0)
6900 panic("softdep_setup_freeblocks: inode busy");
6902 * Add the freeblks structure to the list of operations that
6903 * must await the zero'ed inode being written to disk. If we
6904 * still have a bitmap dependency (delay == 0), then the inode
6905 * has never been written to disk, so we can process the
6906 * freeblks below once we have deleted the dependencies.
6908 delay = (inodedep->id_state & DEPCOMPLETE);
6910 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6912 freeblks->fb_state |= COMPLETE;
6914 * Because the file length has been truncated to zero, any
6915 * pending block allocation dependency structures associated
6916 * with this inode are obsolete and can simply be de-allocated.
6917 * We must first merge the two dependency lists to get rid of
6918 * any duplicate freefrag structures, then purge the merged list.
6919 * If we still have a bitmap dependency, then the inode has never
6920 * been written to disk, so we can free any fragments without delay.
6922 if (flags & IO_NORMAL) {
6923 merge_inode_lists(&inodedep->id_newinoupdt,
6924 &inodedep->id_inoupdt);
6925 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
6926 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6929 if (flags & IO_EXT) {
6930 merge_inode_lists(&inodedep->id_newextupdt,
6931 &inodedep->id_extupdt);
6932 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6933 cancel_allocdirect(&inodedep->id_extupdt, adp,
6938 trunc_dependencies(ip, freeblks, -1, 0, flags);
6940 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
6941 (void) free_inodedep(inodedep);
6942 freeblks->fb_state |= DEPCOMPLETE;
6944 * If the inode with zeroed block pointers is now on disk
6945 * we can start freeing blocks.
6947 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
6948 freeblks->fb_state |= INPROGRESS;
6953 handle_workitem_freeblocks(freeblks, 0);
6954 trunc_pages(ip, length, extblocks, flags);
6958 * Eliminate pages from the page cache that back parts of this inode and
6959 * adjust the vnode pager's idea of our size. This prevents stale data
6960 * from hanging around in the page cache.
6963 trunc_pages(ip, length, extblocks, flags)
6966 ufs2_daddr_t extblocks;
6976 extend = OFF_TO_IDX(lblktosize(fs, -extblocks));
6977 if ((flags & IO_EXT) != 0)
6978 vn_pages_remove(vp, extend, 0);
6979 if ((flags & IO_NORMAL) == 0)
6981 BO_LOCK(&vp->v_bufobj);
6983 BO_UNLOCK(&vp->v_bufobj);
6985 * The vnode pager eliminates file pages we eliminate indirects
6988 vnode_pager_setsize(vp, length);
6990 * Calculate the end based on the last indirect we want to keep. If
6991 * the block extends into indirects we can just use the negative of
6992 * its lbn. Doubles and triples exist at lower numbers so we must
6993 * be careful not to remove those, if they exist. double and triple
6994 * indirect lbns do not overlap with others so it is not important
6995 * to verify how many levels are required.
6997 lbn = lblkno(fs, length);
6998 if (lbn >= NDADDR) {
6999 /* Calculate the virtual lbn of the triple indirect. */
7000 lbn = -lbn - (NIADDR - 1);
7001 end = OFF_TO_IDX(lblktosize(fs, lbn));
7004 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end);
7008 * See if the buf bp is in the range eliminated by truncation.
7011 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags)
7021 /* Only match ext/normal blocks as appropriate. */
7022 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
7023 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0))
7025 /* ALTDATA is always a full truncation. */
7026 if ((bp->b_xflags & BX_ALTDATA) != 0)
7028 /* -1 is full truncation. */
7032 * If this is a partial truncate we only want those
7033 * blocks and indirect blocks that cover the range
7038 lbn = -(lbn + lbn_level(lbn));
7041 /* Here we only truncate lblkno if it's partial. */
7042 if (lbn == lastlbn) {
7051 * Eliminate any dependencies that exist in memory beyond lblkno:off
7054 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags)
7056 struct freeblks *freeblks;
7067 * We must wait for any I/O in progress to finish so that
7068 * all potential buffers on the dirty list will be visible.
7069 * Once they are all there, walk the list and get rid of
7076 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
7077 bp->b_vflags &= ~BV_SCANNED;
7079 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
7080 if (bp->b_vflags & BV_SCANNED)
7082 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7083 bp->b_vflags |= BV_SCANNED;
7086 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer"));
7087 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL)
7090 if (deallocate_dependencies(bp, freeblks, blkoff))
7098 * Now do the work of vtruncbuf while also matching indirect blocks.
7100 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs)
7101 bp->b_vflags &= ~BV_SCANNED;
7103 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) {
7104 if (bp->b_vflags & BV_SCANNED)
7106 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7107 bp->b_vflags |= BV_SCANNED;
7111 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
7112 BO_LOCKPTR(bo)) == ENOLCK) {
7116 bp->b_vflags |= BV_SCANNED;
7119 allocbuf(bp, blkoff);
7122 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF;
7133 cancel_pagedep(pagedep, freeblks, blkoff)
7134 struct pagedep *pagedep;
7135 struct freeblks *freeblks;
7138 struct jremref *jremref;
7139 struct jmvref *jmvref;
7140 struct dirrem *dirrem, *tmp;
7144 * Copy any directory remove dependencies to the list
7145 * to be processed after the freeblks proceeds. If
7146 * directory entry never made it to disk they
7147 * can be dumped directly onto the work list.
7149 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) {
7150 /* Skip this directory removal if it is intended to remain. */
7151 if (dirrem->dm_offset < blkoff)
7154 * If there are any dirrems we wait for the journal write
7155 * to complete and then restart the buf scan as the lock
7158 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
7159 jwait(&jremref->jr_list, MNT_WAIT);
7162 LIST_REMOVE(dirrem, dm_next);
7163 dirrem->dm_dirinum = pagedep->pd_ino;
7164 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list);
7166 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
7167 jwait(&jmvref->jm_list, MNT_WAIT);
7171 * When we're partially truncating a pagedep we just want to flush
7172 * journal entries and return. There can not be any adds in the
7173 * truncated portion of the directory and newblk must remain if
7174 * part of the block remains.
7179 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
7180 if (dap->da_offset > blkoff)
7181 panic("cancel_pagedep: diradd %p off %d > %d",
7182 dap, dap->da_offset, blkoff);
7183 for (i = 0; i < DAHASHSZ; i++)
7184 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist)
7185 if (dap->da_offset > blkoff)
7186 panic("cancel_pagedep: diradd %p off %d > %d",
7187 dap, dap->da_offset, blkoff);
7191 * There should be no directory add dependencies present
7192 * as the directory could not be truncated until all
7193 * children were removed.
7195 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
7196 ("deallocate_dependencies: pendinghd != NULL"));
7197 for (i = 0; i < DAHASHSZ; i++)
7198 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
7199 ("deallocate_dependencies: diraddhd != NULL"));
7200 if ((pagedep->pd_state & NEWBLOCK) != 0)
7201 free_newdirblk(pagedep->pd_newdirblk);
7202 if (free_pagedep(pagedep) == 0)
7203 panic("Failed to free pagedep %p", pagedep);
7208 * Reclaim any dependency structures from a buffer that is about to
7209 * be reallocated to a new vnode. The buffer must be locked, thus,
7210 * no I/O completion operations can occur while we are manipulating
7211 * its associated dependencies. The mutex is held so that other I/O's
7212 * associated with related dependencies do not occur.
7215 deallocate_dependencies(bp, freeblks, off)
7217 struct freeblks *freeblks;
7220 struct indirdep *indirdep;
7221 struct pagedep *pagedep;
7222 struct worklist *wk, *wkn;
7223 struct ufsmount *ump;
7225 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
7227 ump = VFSTOUFS(wk->wk_mp);
7229 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) {
7230 switch (wk->wk_type) {
7232 indirdep = WK_INDIRDEP(wk);
7233 if (bp->b_lblkno >= 0 ||
7234 bp->b_blkno != indirdep->ir_savebp->b_lblkno)
7235 panic("deallocate_dependencies: not indir");
7236 cancel_indirdep(indirdep, bp, freeblks);
7240 pagedep = WK_PAGEDEP(wk);
7241 if (cancel_pagedep(pagedep, freeblks, off)) {
7249 * Simply remove the allocindir, we'll find it via
7250 * the indirdep where we can clear pointers if
7253 WORKLIST_REMOVE(wk);
7258 * A truncation is waiting for the zero'd pointers
7259 * to be written. It can be freed when the freeblks
7262 WORKLIST_REMOVE(wk);
7263 wk->wk_state |= ONDEPLIST;
7264 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
7272 panic("deallocate_dependencies: Unexpected type %s",
7273 TYPENAME(wk->wk_type));
7280 * Don't throw away this buf, we were partially truncating and
7281 * some deps may always remain.
7285 bp->b_vflags |= BV_SCANNED;
7288 bp->b_flags |= B_INVAL | B_NOCACHE;
7294 * An allocdirect is being canceled due to a truncate. We must make sure
7295 * the journal entry is released in concert with the blkfree that releases
7296 * the storage. Completed journal entries must not be released until the
7297 * space is no longer pointed to by the inode or in the bitmap.
7300 cancel_allocdirect(adphead, adp, freeblks)
7301 struct allocdirectlst *adphead;
7302 struct allocdirect *adp;
7303 struct freeblks *freeblks;
7305 struct freework *freework;
7306 struct newblk *newblk;
7307 struct worklist *wk;
7309 TAILQ_REMOVE(adphead, adp, ad_next);
7310 newblk = (struct newblk *)adp;
7313 * Find the correct freework structure.
7315 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
7316 if (wk->wk_type != D_FREEWORK)
7318 freework = WK_FREEWORK(wk);
7319 if (freework->fw_blkno == newblk->nb_newblkno)
7322 if (freework == NULL)
7323 panic("cancel_allocdirect: Freework not found");
7325 * If a newblk exists at all we still have the journal entry that
7326 * initiated the allocation so we do not need to journal the free.
7328 cancel_jfreeblk(freeblks, freework->fw_blkno);
7330 * If the journal hasn't been written the jnewblk must be passed
7331 * to the call to ffs_blkfree that reclaims the space. We accomplish
7332 * this by linking the journal dependency into the freework to be
7333 * freed when freework_freeblock() is called. If the journal has
7334 * been written we can simply reclaim the journal space when the
7335 * freeblks work is complete.
7337 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list,
7338 &freeblks->fb_jwork);
7339 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
7344 * Cancel a new block allocation. May be an indirect or direct block. We
7345 * remove it from various lists and return any journal record that needs to
7346 * be resolved by the caller.
7348 * A special consideration is made for indirects which were never pointed
7349 * at on disk and will never be found once this block is released.
7351 static struct jnewblk *
7352 cancel_newblk(newblk, wk, wkhd)
7353 struct newblk *newblk;
7354 struct worklist *wk;
7355 struct workhead *wkhd;
7357 struct jnewblk *jnewblk;
7359 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno);
7361 newblk->nb_state |= GOINGAWAY;
7363 * Previously we traversed the completedhd on each indirdep
7364 * attached to this newblk to cancel them and gather journal
7365 * work. Since we need only the oldest journal segment and
7366 * the lowest point on the tree will always have the oldest
7367 * journal segment we are free to release the segments
7368 * of any subordinates and may leave the indirdep list to
7369 * indirdep_complete() when this newblk is freed.
7371 if (newblk->nb_state & ONDEPLIST) {
7372 newblk->nb_state &= ~ONDEPLIST;
7373 LIST_REMOVE(newblk, nb_deps);
7375 if (newblk->nb_state & ONWORKLIST)
7376 WORKLIST_REMOVE(&newblk->nb_list);
7378 * If the journal entry hasn't been written we save a pointer to
7379 * the dependency that frees it until it is written or the
7380 * superseding operation completes.
7382 jnewblk = newblk->nb_jnewblk;
7383 if (jnewblk != NULL && wk != NULL) {
7384 newblk->nb_jnewblk = NULL;
7385 jnewblk->jn_dep = wk;
7387 if (!LIST_EMPTY(&newblk->nb_jwork))
7388 jwork_move(wkhd, &newblk->nb_jwork);
7390 * When truncating we must free the newdirblk early to remove
7391 * the pagedep from the hash before returning.
7393 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7394 free_newdirblk(WK_NEWDIRBLK(wk));
7395 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7396 panic("cancel_newblk: extra newdirblk");
7402 * Schedule the freefrag associated with a newblk to be released once
7403 * the pointers are written and the previous block is no longer needed.
7406 newblk_freefrag(newblk)
7407 struct newblk *newblk;
7409 struct freefrag *freefrag;
7411 if (newblk->nb_freefrag == NULL)
7413 freefrag = newblk->nb_freefrag;
7414 newblk->nb_freefrag = NULL;
7415 freefrag->ff_state |= COMPLETE;
7416 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
7417 add_to_worklist(&freefrag->ff_list, 0);
7421 * Free a newblk. Generate a new freefrag work request if appropriate.
7422 * This must be called after the inode pointer and any direct block pointers
7423 * are valid or fully removed via truncate or frag extension.
7427 struct newblk *newblk;
7429 struct indirdep *indirdep;
7430 struct worklist *wk;
7432 KASSERT(newblk->nb_jnewblk == NULL,
7433 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk));
7434 KASSERT(newblk->nb_list.wk_type != D_NEWBLK,
7435 ("free_newblk: unclaimed newblk"));
7436 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp));
7437 newblk_freefrag(newblk);
7438 if (newblk->nb_state & ONDEPLIST)
7439 LIST_REMOVE(newblk, nb_deps);
7440 if (newblk->nb_state & ONWORKLIST)
7441 WORKLIST_REMOVE(&newblk->nb_list);
7442 LIST_REMOVE(newblk, nb_hash);
7443 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7444 free_newdirblk(WK_NEWDIRBLK(wk));
7445 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7446 panic("free_newblk: extra newdirblk");
7447 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL)
7448 indirdep_complete(indirdep);
7449 handle_jwork(&newblk->nb_jwork);
7450 WORKITEM_FREE(newblk, D_NEWBLK);
7454 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
7455 * This routine must be called with splbio interrupts blocked.
7458 free_newdirblk(newdirblk)
7459 struct newdirblk *newdirblk;
7461 struct pagedep *pagedep;
7463 struct worklist *wk;
7465 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp));
7466 WORKLIST_REMOVE(&newdirblk->db_list);
7468 * If the pagedep is still linked onto the directory buffer
7469 * dependency chain, then some of the entries on the
7470 * pd_pendinghd list may not be committed to disk yet. In
7471 * this case, we will simply clear the NEWBLOCK flag and
7472 * let the pd_pendinghd list be processed when the pagedep
7473 * is next written. If the pagedep is no longer on the buffer
7474 * dependency chain, then all the entries on the pd_pending
7475 * list are committed to disk and we can free them here.
7477 pagedep = newdirblk->db_pagedep;
7478 pagedep->pd_state &= ~NEWBLOCK;
7479 if ((pagedep->pd_state & ONWORKLIST) == 0) {
7480 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
7481 free_diradd(dap, NULL);
7483 * If no dependencies remain, the pagedep will be freed.
7485 free_pagedep(pagedep);
7487 /* Should only ever be one item in the list. */
7488 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
7489 WORKLIST_REMOVE(wk);
7490 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
7492 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
7496 * Prepare an inode to be freed. The actual free operation is not
7497 * done until the zero'ed inode has been written to disk.
7500 softdep_freefile(pvp, ino, mode)
7505 struct inode *ip = VTOI(pvp);
7506 struct inodedep *inodedep;
7507 struct freefile *freefile;
7508 struct freeblks *freeblks;
7509 struct ufsmount *ump;
7512 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
7513 ("softdep_freefile called on non-softdep filesystem"));
7515 * This sets up the inode de-allocation dependency.
7517 freefile = malloc(sizeof(struct freefile),
7518 M_FREEFILE, M_SOFTDEP_FLAGS);
7519 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
7520 freefile->fx_mode = mode;
7521 freefile->fx_oldinum = ino;
7522 freefile->fx_devvp = ip->i_devvp;
7523 LIST_INIT(&freefile->fx_jwork);
7525 ip->i_fs->fs_pendinginodes += 1;
7529 * If the inodedep does not exist, then the zero'ed inode has
7530 * been written to disk. If the allocated inode has never been
7531 * written to disk, then the on-disk inode is zero'ed. In either
7532 * case we can free the file immediately. If the journal was
7533 * canceled before being written the inode will never make it to
7534 * disk and we must send the canceled journal entrys to
7535 * ffs_freefile() to be cleared in conjunction with the bitmap.
7536 * Any blocks waiting on the inode to write can be safely freed
7537 * here as it will never been written.
7540 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7543 * Clear out freeblks that no longer need to reference
7547 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) {
7548 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks,
7550 freeblks->fb_state &= ~ONDEPLIST;
7553 * Remove this inode from the unlinked list.
7555 if (inodedep->id_state & UNLINKED) {
7557 * Save the journal work to be freed with the bitmap
7558 * before we clear UNLINKED. Otherwise it can be lost
7559 * if the inode block is written.
7561 handle_bufwait(inodedep, &freefile->fx_jwork);
7562 clear_unlinked_inodedep(inodedep);
7564 * Re-acquire inodedep as we've dropped the
7565 * per-filesystem lock in clear_unlinked_inodedep().
7567 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7570 if (inodedep == NULL || check_inode_unwritten(inodedep)) {
7572 handle_workitem_freefile(freefile);
7575 if ((inodedep->id_state & DEPCOMPLETE) == 0)
7576 inodedep->id_state |= GOINGAWAY;
7577 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
7579 if (ip->i_number == ino)
7580 ip->i_flag |= IN_MODIFIED;
7584 * Check to see if an inode has never been written to disk. If
7585 * so free the inodedep and return success, otherwise return failure.
7586 * This routine must be called with splbio interrupts blocked.
7588 * If we still have a bitmap dependency, then the inode has never
7589 * been written to disk. Drop the dependency as it is no longer
7590 * necessary since the inode is being deallocated. We set the
7591 * ALLCOMPLETE flags since the bitmap now properly shows that the
7592 * inode is not allocated. Even if the inode is actively being
7593 * written, it has been rolled back to its zero'ed state, so we
7594 * are ensured that a zero inode is what is on the disk. For short
7595 * lived files, this change will usually result in removing all the
7596 * dependencies from the inode so that it can be freed immediately.
7599 check_inode_unwritten(inodedep)
7600 struct inodedep *inodedep;
7603 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7605 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
7606 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7607 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7608 !LIST_EMPTY(&inodedep->id_bufwait) ||
7609 !LIST_EMPTY(&inodedep->id_inowait) ||
7610 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7611 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7612 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7613 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7614 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7615 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7616 inodedep->id_mkdiradd != NULL ||
7617 inodedep->id_nlinkdelta != 0)
7620 * Another process might be in initiate_write_inodeblock_ufs[12]
7621 * trying to allocate memory without holding "Softdep Lock".
7623 if ((inodedep->id_state & IOSTARTED) != 0 &&
7624 inodedep->id_savedino1 == NULL)
7627 if (inodedep->id_state & ONDEPLIST)
7628 LIST_REMOVE(inodedep, id_deps);
7629 inodedep->id_state &= ~ONDEPLIST;
7630 inodedep->id_state |= ALLCOMPLETE;
7631 inodedep->id_bmsafemap = NULL;
7632 if (inodedep->id_state & ONWORKLIST)
7633 WORKLIST_REMOVE(&inodedep->id_list);
7634 if (inodedep->id_savedino1 != NULL) {
7635 free(inodedep->id_savedino1, M_SAVEDINO);
7636 inodedep->id_savedino1 = NULL;
7638 if (free_inodedep(inodedep) == 0)
7639 panic("check_inode_unwritten: busy inode");
7644 check_inodedep_free(inodedep)
7645 struct inodedep *inodedep;
7648 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7649 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
7650 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7651 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7652 !LIST_EMPTY(&inodedep->id_bufwait) ||
7653 !LIST_EMPTY(&inodedep->id_inowait) ||
7654 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7655 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7656 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7657 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7658 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7659 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7660 inodedep->id_mkdiradd != NULL ||
7661 inodedep->id_nlinkdelta != 0 ||
7662 inodedep->id_savedino1 != NULL)
7668 * Try to free an inodedep structure. Return 1 if it could be freed.
7671 free_inodedep(inodedep)
7672 struct inodedep *inodedep;
7675 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7676 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
7677 !check_inodedep_free(inodedep))
7679 if (inodedep->id_state & ONDEPLIST)
7680 LIST_REMOVE(inodedep, id_deps);
7681 LIST_REMOVE(inodedep, id_hash);
7682 WORKITEM_FREE(inodedep, D_INODEDEP);
7687 * Free the block referenced by a freework structure. The parent freeblks
7688 * structure is released and completed when the final cg bitmap reaches
7689 * the disk. This routine may be freeing a jnewblk which never made it to
7690 * disk in which case we do not have to wait as the operation is undone
7691 * in memory immediately.
7694 freework_freeblock(freework)
7695 struct freework *freework;
7697 struct freeblks *freeblks;
7698 struct jnewblk *jnewblk;
7699 struct ufsmount *ump;
7700 struct workhead wkhd;
7705 ump = VFSTOUFS(freework->fw_list.wk_mp);
7708 * Handle partial truncate separately.
7710 if (freework->fw_indir) {
7711 complete_trunc_indir(freework);
7714 freeblks = freework->fw_freeblks;
7716 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0;
7717 bsize = lfragtosize(fs, freework->fw_frags);
7720 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives
7721 * on the indirblk hashtable and prevents premature freeing.
7723 freework->fw_state |= DEPCOMPLETE;
7725 * SUJ needs to wait for the segment referencing freed indirect
7726 * blocks to expire so that we know the checker will not confuse
7727 * a re-allocated indirect block with its old contents.
7729 if (needj && freework->fw_lbn <= -NDADDR)
7730 indirblk_insert(freework);
7732 * If we are canceling an existing jnewblk pass it to the free
7733 * routine, otherwise pass the freeblk which will ultimately
7734 * release the freeblks. If we're not journaling, we can just
7735 * free the freeblks immediately.
7737 jnewblk = freework->fw_jnewblk;
7738 if (jnewblk != NULL) {
7739 cancel_jnewblk(jnewblk, &wkhd);
7742 freework->fw_state |= DELAYEDFREE;
7743 freeblks->fb_cgwait++;
7744 WORKLIST_INSERT(&wkhd, &freework->fw_list);
7747 freeblks_free(ump, freeblks, btodb(bsize));
7749 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld",
7750 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize);
7751 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize,
7752 freeblks->fb_inum, freeblks->fb_vtype, &wkhd);
7755 * The jnewblk will be discarded and the bits in the map never
7756 * made it to disk. We can immediately free the freeblk.
7759 handle_written_freework(freework);
7763 * We enqueue freework items that need processing back on the freeblks and
7764 * add the freeblks to the worklist. This makes it easier to find all work
7765 * required to flush a truncation in process_truncates().
7768 freework_enqueue(freework)
7769 struct freework *freework;
7771 struct freeblks *freeblks;
7773 freeblks = freework->fw_freeblks;
7774 if ((freework->fw_state & INPROGRESS) == 0)
7775 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
7776 if ((freeblks->fb_state &
7777 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE &&
7778 LIST_EMPTY(&freeblks->fb_jblkdephd))
7779 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7783 * Start, continue, or finish the process of freeing an indirect block tree.
7784 * The free operation may be paused at any point with fw_off containing the
7785 * offset to restart from. This enables us to implement some flow control
7786 * for large truncates which may fan out and generate a huge number of
7790 handle_workitem_indirblk(freework)
7791 struct freework *freework;
7793 struct freeblks *freeblks;
7794 struct ufsmount *ump;
7797 freeblks = freework->fw_freeblks;
7798 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7800 if (freework->fw_state & DEPCOMPLETE) {
7801 handle_written_freework(freework);
7804 if (freework->fw_off == NINDIR(fs)) {
7805 freework_freeblock(freework);
7808 freework->fw_state |= INPROGRESS;
7810 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
7816 * Called when a freework structure attached to a cg buf is written. The
7817 * ref on either the parent or the freeblks structure is released and
7818 * the freeblks is added back to the worklist if there is more work to do.
7821 handle_written_freework(freework)
7822 struct freework *freework;
7824 struct freeblks *freeblks;
7825 struct freework *parent;
7827 freeblks = freework->fw_freeblks;
7828 parent = freework->fw_parent;
7829 if (freework->fw_state & DELAYEDFREE)
7830 freeblks->fb_cgwait--;
7831 freework->fw_state |= COMPLETE;
7832 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
7833 WORKITEM_FREE(freework, D_FREEWORK);
7835 if (--parent->fw_ref == 0)
7836 freework_enqueue(parent);
7839 if (--freeblks->fb_ref != 0)
7841 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) ==
7842 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd))
7843 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7847 * This workitem routine performs the block de-allocation.
7848 * The workitem is added to the pending list after the updated
7849 * inode block has been written to disk. As mentioned above,
7850 * checks regarding the number of blocks de-allocated (compared
7851 * to the number of blocks allocated for the file) are also
7852 * performed in this function.
7855 handle_workitem_freeblocks(freeblks, flags)
7856 struct freeblks *freeblks;
7859 struct freework *freework;
7860 struct newblk *newblk;
7861 struct allocindir *aip;
7862 struct ufsmount *ump;
7863 struct worklist *wk;
7865 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd),
7866 ("handle_workitem_freeblocks: Journal entries not written."));
7867 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7869 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
7870 WORKLIST_REMOVE(wk);
7871 switch (wk->wk_type) {
7873 wk->wk_state |= COMPLETE;
7874 add_to_worklist(wk, 0);
7878 free_newblk(WK_NEWBLK(wk));
7882 aip = WK_ALLOCINDIR(wk);
7884 if (aip->ai_state & DELAYEDFREE) {
7886 freework = newfreework(ump, freeblks, NULL,
7887 aip->ai_lbn, aip->ai_newblkno,
7888 ump->um_fs->fs_frag, 0, 0);
7891 newblk = WK_NEWBLK(wk);
7892 if (newblk->nb_jnewblk) {
7893 freework->fw_jnewblk = newblk->nb_jnewblk;
7894 newblk->nb_jnewblk->jn_dep = &freework->fw_list;
7895 newblk->nb_jnewblk = NULL;
7897 free_newblk(newblk);
7901 freework = WK_FREEWORK(wk);
7902 if (freework->fw_lbn <= -NDADDR)
7903 handle_workitem_indirblk(freework);
7905 freework_freeblock(freework);
7908 panic("handle_workitem_freeblocks: Unknown type %s",
7909 TYPENAME(wk->wk_type));
7912 if (freeblks->fb_ref != 0) {
7913 freeblks->fb_state &= ~INPROGRESS;
7914 wake_worklist(&freeblks->fb_list);
7919 return handle_complete_freeblocks(freeblks, flags);
7924 * Handle completion of block free via truncate. This allows fs_pending
7925 * to track the actual free block count more closely than if we only updated
7926 * it at the end. We must be careful to handle cases where the block count
7927 * on free was incorrect.
7930 freeblks_free(ump, freeblks, blocks)
7931 struct ufsmount *ump;
7932 struct freeblks *freeblks;
7936 ufs2_daddr_t remain;
7939 remain = -freeblks->fb_chkcnt;
7940 freeblks->fb_chkcnt += blocks;
7942 if (remain < blocks)
7945 fs->fs_pendingblocks -= blocks;
7951 * Once all of the freework workitems are complete we can retire the
7952 * freeblocks dependency and any journal work awaiting completion. This
7953 * can not be called until all other dependencies are stable on disk.
7956 handle_complete_freeblocks(freeblks, flags)
7957 struct freeblks *freeblks;
7960 struct inodedep *inodedep;
7964 struct ufsmount *ump;
7967 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7969 flags = LK_EXCLUSIVE | flags;
7970 spare = freeblks->fb_chkcnt;
7973 * If we did not release the expected number of blocks we may have
7974 * to adjust the inode block count here. Only do so if it wasn't
7975 * a truncation to zero and the modrev still matches.
7977 if (spare && freeblks->fb_len != 0) {
7978 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum,
7979 flags, &vp, FFSV_FORCEINSMQ) != 0)
7982 if (DIP(ip, i_modrev) == freeblks->fb_modrev) {
7983 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare);
7984 ip->i_flag |= IN_CHANGE;
7986 * We must wait so this happens before the
7987 * journal is reclaimed.
7995 fs->fs_pendingblocks += spare;
8001 quotaadj(freeblks->fb_quota, ump, -spare);
8002 quotarele(freeblks->fb_quota);
8005 if (freeblks->fb_state & ONDEPLIST) {
8006 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum,
8008 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next);
8009 freeblks->fb_state &= ~ONDEPLIST;
8010 if (TAILQ_EMPTY(&inodedep->id_freeblklst))
8011 free_inodedep(inodedep);
8014 * All of the freeblock deps must be complete prior to this call
8015 * so it's now safe to complete earlier outstanding journal entries.
8017 handle_jwork(&freeblks->fb_jwork);
8018 WORKITEM_FREE(freeblks, D_FREEBLKS);
8024 * Release blocks associated with the freeblks and stored in the indirect
8025 * block dbn. If level is greater than SINGLE, the block is an indirect block
8026 * and recursive calls to indirtrunc must be used to cleanse other indirect
8029 * This handles partial and complete truncation of blocks. Partial is noted
8030 * with goingaway == 0. In this case the freework is completed after the
8031 * zero'd indirects are written to disk. For full truncation the freework
8032 * is completed after the block is freed.
8035 indir_trunc(freework, dbn, lbn)
8036 struct freework *freework;
8040 struct freework *nfreework;
8041 struct workhead wkhd;
8042 struct freeblks *freeblks;
8045 struct indirdep *indirdep;
8046 struct ufsmount *ump;
8048 ufs2_daddr_t nb, nnb, *bap2;
8049 ufs_lbn_t lbnadd, nlbn;
8050 int i, nblocks, ufs1fmt;
8058 freeblks = freework->fw_freeblks;
8059 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
8062 * Get buffer of block pointers to be freed. There are three cases:
8064 * 1) Partial truncate caches the indirdep pointer in the freework
8065 * which provides us a back copy to the save bp which holds the
8066 * pointers we want to clear. When this completes the zero
8067 * pointers are written to the real copy.
8068 * 2) The indirect is being completely truncated, cancel_indirdep()
8069 * eliminated the real copy and placed the indirdep on the saved
8070 * copy. The indirdep and buf are discarded when this completes.
8071 * 3) The indirect was not in memory, we read a copy off of the disk
8072 * using the devvp and drop and invalidate the buffer when we're
8077 if (freework->fw_indir != NULL) {
8079 indirdep = freework->fw_indir;
8080 bp = indirdep->ir_savebp;
8081 if (bp == NULL || bp->b_blkno != dbn)
8082 panic("indir_trunc: Bad saved buf %p blkno %jd",
8084 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) {
8086 * The lock prevents the buf dep list from changing and
8087 * indirects on devvp should only ever have one dependency.
8089 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep));
8090 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0)
8091 panic("indir_trunc: Bad indirdep %p from buf %p",
8093 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
8094 NOCRED, &bp) != 0) {
8099 /* Protects against a race with complete_trunc_indir(). */
8100 freework->fw_state &= ~INPROGRESS;
8102 * If we have an indirdep we need to enforce the truncation order
8103 * and discard it when it is complete.
8106 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) &&
8107 !TAILQ_EMPTY(&indirdep->ir_trunc)) {
8109 * Add the complete truncate to the list on the
8110 * indirdep to enforce in-order processing.
8112 if (freework->fw_indir == NULL)
8113 TAILQ_INSERT_TAIL(&indirdep->ir_trunc,
8119 * If we're goingaway, free the indirdep. Otherwise it will
8120 * linger until the write completes.
8123 free_indirdep(indirdep);
8126 /* Initialize pointers depending on block size. */
8127 if (ump->um_fstype == UFS1) {
8128 bap1 = (ufs1_daddr_t *)bp->b_data;
8129 nb = bap1[freework->fw_off];
8133 bap2 = (ufs2_daddr_t *)bp->b_data;
8134 nb = bap2[freework->fw_off];
8138 level = lbn_level(lbn);
8139 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0;
8140 lbnadd = lbn_offset(fs, level);
8141 nblocks = btodb(fs->fs_bsize);
8142 nfreework = freework;
8146 * Reclaim blocks. Traverses into nested indirect levels and
8147 * arranges for the current level to be freed when subordinates
8148 * are free when journaling.
8150 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
8151 if (i != NINDIR(fs) - 1) {
8162 nlbn = (lbn + 1) - (i * lbnadd);
8164 nfreework = newfreework(ump, freeblks, freework,
8165 nlbn, nb, fs->fs_frag, 0, 0);
8168 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
8170 struct freedep *freedep;
8173 * Attempt to aggregate freedep dependencies for
8174 * all blocks being released to the same CG.
8178 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
8179 freedep = newfreedep(freework);
8180 WORKLIST_INSERT_UNLOCKED(&wkhd,
8185 "indir_trunc: ino %d blkno %jd size %ld",
8186 freeblks->fb_inum, nb, fs->fs_bsize);
8187 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
8188 fs->fs_bsize, freeblks->fb_inum,
8189 freeblks->fb_vtype, &wkhd);
8193 bp->b_flags |= B_INVAL | B_NOCACHE;
8198 freedblocks = (nblocks * cnt);
8200 freedblocks += nblocks;
8201 freeblks_free(ump, freeblks, freedblocks);
8203 * If we are journaling set up the ref counts and offset so this
8204 * indirect can be completed when its children are free.
8208 freework->fw_off = i;
8209 freework->fw_ref += freedeps;
8210 freework->fw_ref -= NINDIR(fs) + 1;
8212 freeblks->fb_cgwait += freedeps;
8213 if (freework->fw_ref == 0)
8214 freework_freeblock(freework);
8219 * If we're not journaling we can free the indirect now.
8221 dbn = dbtofsb(fs, dbn);
8223 "indir_trunc 2: ino %d blkno %jd size %ld",
8224 freeblks->fb_inum, dbn, fs->fs_bsize);
8225 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
8226 freeblks->fb_inum, freeblks->fb_vtype, NULL);
8227 /* Non SUJ softdep does single-threaded truncations. */
8228 if (freework->fw_blkno == dbn) {
8229 freework->fw_state |= ALLCOMPLETE;
8231 handle_written_freework(freework);
8238 * Cancel an allocindir when it is removed via truncation. When bp is not
8239 * NULL the indirect never appeared on disk and is scheduled to be freed
8240 * independently of the indir so we can more easily track journal work.
8243 cancel_allocindir(aip, bp, freeblks, trunc)
8244 struct allocindir *aip;
8246 struct freeblks *freeblks;
8249 struct indirdep *indirdep;
8250 struct freefrag *freefrag;
8251 struct newblk *newblk;
8253 newblk = (struct newblk *)aip;
8254 LIST_REMOVE(aip, ai_next);
8256 * We must eliminate the pointer in bp if it must be freed on its
8257 * own due to partial truncate or pending journal work.
8259 if (bp && (trunc || newblk->nb_jnewblk)) {
8261 * Clear the pointer and mark the aip to be freed
8262 * directly if it never existed on disk.
8264 aip->ai_state |= DELAYEDFREE;
8265 indirdep = aip->ai_indirdep;
8266 if (indirdep->ir_state & UFS1FMT)
8267 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8269 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8272 * When truncating the previous pointer will be freed via
8273 * savedbp. Eliminate the freefrag which would dup free.
8275 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) {
8276 newblk->nb_freefrag = NULL;
8277 if (freefrag->ff_jdep)
8279 WK_JFREEFRAG(freefrag->ff_jdep));
8280 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork);
8281 WORKITEM_FREE(freefrag, D_FREEFRAG);
8284 * If the journal hasn't been written the jnewblk must be passed
8285 * to the call to ffs_blkfree that reclaims the space. We accomplish
8286 * this by leaving the journal dependency on the newblk to be freed
8287 * when a freework is created in handle_workitem_freeblocks().
8289 cancel_newblk(newblk, NULL, &freeblks->fb_jwork);
8290 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
8294 * Create the mkdir dependencies for . and .. in a new directory. Link them
8295 * in to a newdirblk so any subsequent additions are tracked properly. The
8296 * caller is responsible for adding the mkdir1 dependency to the journal
8297 * and updating id_mkdiradd. This function returns with the per-filesystem
8300 static struct mkdir *
8301 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
8305 struct buf *newdirbp;
8306 struct mkdir **mkdirp;
8308 struct newblk *newblk;
8309 struct pagedep *pagedep;
8310 struct inodedep *inodedep;
8311 struct newdirblk *newdirblk;
8312 struct mkdir *mkdir1, *mkdir2;
8313 struct worklist *wk;
8314 struct jaddref *jaddref;
8315 struct ufsmount *ump;
8318 mp = dap->da_list.wk_mp;
8320 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
8322 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8323 LIST_INIT(&newdirblk->db_mkdir);
8324 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8325 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
8326 mkdir1->md_state = ATTACHED | MKDIR_BODY;
8327 mkdir1->md_diradd = dap;
8328 mkdir1->md_jaddref = NULL;
8329 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8330 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
8331 mkdir2->md_state = ATTACHED | MKDIR_PARENT;
8332 mkdir2->md_diradd = dap;
8333 mkdir2->md_jaddref = NULL;
8334 if (MOUNTEDSUJ(mp) == 0) {
8335 mkdir1->md_state |= DEPCOMPLETE;
8336 mkdir2->md_state |= DEPCOMPLETE;
8339 * Dependency on "." and ".." being written to disk.
8341 mkdir1->md_buf = newdirbp;
8342 ACQUIRE_LOCK(VFSTOUFS(mp));
8343 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs);
8345 * We must link the pagedep, allocdirect, and newdirblk for
8346 * the initial file page so the pointer to the new directory
8347 * is not written until the directory contents are live and
8348 * any subsequent additions are not marked live until the
8349 * block is reachable via the inode.
8351 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0)
8352 panic("setup_newdir: lost pagedep");
8353 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
8354 if (wk->wk_type == D_ALLOCDIRECT)
8357 panic("setup_newdir: lost allocdirect");
8358 if (pagedep->pd_state & NEWBLOCK)
8359 panic("setup_newdir: NEWBLOCK already set");
8360 newblk = WK_NEWBLK(wk);
8361 pagedep->pd_state |= NEWBLOCK;
8362 pagedep->pd_newdirblk = newdirblk;
8363 newdirblk->db_pagedep = pagedep;
8364 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8365 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
8367 * Look up the inodedep for the parent directory so that we
8368 * can link mkdir2 into the pending dotdot jaddref or
8369 * the inode write if there is none. If the inode is
8370 * ALLCOMPLETE and no jaddref is present all dependencies have
8371 * been satisfied and mkdir2 can be freed.
8373 inodedep_lookup(mp, dinum, 0, &inodedep);
8374 if (MOUNTEDSUJ(mp)) {
8375 if (inodedep == NULL)
8376 panic("setup_newdir: Lost parent.");
8377 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8379 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
8380 (jaddref->ja_state & MKDIR_PARENT),
8381 ("setup_newdir: bad dotdot jaddref %p", jaddref));
8382 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8383 mkdir2->md_jaddref = jaddref;
8384 jaddref->ja_mkdir = mkdir2;
8385 } else if (inodedep == NULL ||
8386 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
8387 dap->da_state &= ~MKDIR_PARENT;
8388 WORKITEM_FREE(mkdir2, D_MKDIR);
8391 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8392 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list);
8400 * Directory entry addition dependencies.
8402 * When adding a new directory entry, the inode (with its incremented link
8403 * count) must be written to disk before the directory entry's pointer to it.
8404 * Also, if the inode is newly allocated, the corresponding freemap must be
8405 * updated (on disk) before the directory entry's pointer. These requirements
8406 * are met via undo/redo on the directory entry's pointer, which consists
8407 * simply of the inode number.
8409 * As directory entries are added and deleted, the free space within a
8410 * directory block can become fragmented. The ufs filesystem will compact
8411 * a fragmented directory block to make space for a new entry. When this
8412 * occurs, the offsets of previously added entries change. Any "diradd"
8413 * dependency structures corresponding to these entries must be updated with
8418 * This routine is called after the in-memory inode's link
8419 * count has been incremented, but before the directory entry's
8420 * pointer to the inode has been set.
8423 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
8424 struct buf *bp; /* buffer containing directory block */
8425 struct inode *dp; /* inode for directory */
8426 off_t diroffset; /* offset of new entry in directory */
8427 ino_t newinum; /* inode referenced by new directory entry */
8428 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
8429 int isnewblk; /* entry is in a newly allocated block */
8431 int offset; /* offset of new entry within directory block */
8432 ufs_lbn_t lbn; /* block in directory containing new entry */
8435 struct newblk *newblk;
8436 struct pagedep *pagedep;
8437 struct inodedep *inodedep;
8438 struct newdirblk *newdirblk;
8439 struct mkdir *mkdir1, *mkdir2;
8440 struct jaddref *jaddref;
8441 struct ufsmount *ump;
8447 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8448 ("softdep_setup_directory_add called on non-softdep filesystem"));
8450 * Whiteouts have no dependencies.
8452 if (newinum == WINO) {
8453 if (newdirbp != NULL)
8458 mkdir1 = mkdir2 = NULL;
8460 lbn = lblkno(fs, diroffset);
8461 offset = blkoff(fs, diroffset);
8462 dap = malloc(sizeof(struct diradd), M_DIRADD,
8463 M_SOFTDEP_FLAGS|M_ZERO);
8464 workitem_alloc(&dap->da_list, D_DIRADD, mp);
8465 dap->da_offset = offset;
8466 dap->da_newinum = newinum;
8467 dap->da_state = ATTACHED;
8468 LIST_INIT(&dap->da_jwork);
8469 isindir = bp->b_lblkno >= NDADDR;
8472 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
8473 newdirblk = malloc(sizeof(struct newdirblk),
8474 M_NEWDIRBLK, M_SOFTDEP_FLAGS);
8475 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8476 LIST_INIT(&newdirblk->db_mkdir);
8479 * If we're creating a new directory setup the dependencies and set
8480 * the dap state to wait for them. Otherwise it's COMPLETE and
8483 if (newdirbp == NULL) {
8484 dap->da_state |= DEPCOMPLETE;
8487 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
8488 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
8492 * Link into parent directory pagedep to await its being written.
8494 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep);
8496 if (diradd_lookup(pagedep, offset) != NULL)
8497 panic("softdep_setup_directory_add: %p already at off %d\n",
8498 diradd_lookup(pagedep, offset), offset);
8500 dap->da_pagedep = pagedep;
8501 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
8503 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
8505 * If we're journaling, link the diradd into the jaddref so it
8506 * may be completed after the journal entry is written. Otherwise,
8507 * link the diradd into its inodedep. If the inode is not yet
8508 * written place it on the bufwait list, otherwise do the post-inode
8509 * write processing to put it on the id_pendinghd list.
8511 if (MOUNTEDSUJ(mp)) {
8512 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8514 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
8515 ("softdep_setup_directory_add: bad jaddref %p", jaddref));
8516 jaddref->ja_diroff = diroffset;
8517 jaddref->ja_diradd = dap;
8518 add_to_journal(&jaddref->ja_list);
8519 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
8520 diradd_inode_written(dap, inodedep);
8522 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
8524 * Add the journal entries for . and .. links now that the primary
8527 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) {
8528 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
8529 inoreflst, if_deps);
8530 KASSERT(jaddref != NULL &&
8531 jaddref->ja_ino == jaddref->ja_parent &&
8532 (jaddref->ja_state & MKDIR_BODY),
8533 ("softdep_setup_directory_add: bad dot jaddref %p",
8535 mkdir1->md_jaddref = jaddref;
8536 jaddref->ja_mkdir = mkdir1;
8538 * It is important that the dotdot journal entry
8539 * is added prior to the dot entry since dot writes
8540 * both the dot and dotdot links. These both must
8541 * be added after the primary link for the journal
8542 * to remain consistent.
8544 add_to_journal(&mkdir2->md_jaddref->ja_list);
8545 add_to_journal(&jaddref->ja_list);
8548 * If we are adding a new directory remember this diradd so that if
8549 * we rename it we can keep the dot and dotdot dependencies. If
8550 * we are adding a new name for an inode that has a mkdiradd we
8551 * must be in rename and we have to move the dot and dotdot
8552 * dependencies to this new name. The old name is being orphaned
8555 if (mkdir1 != NULL) {
8556 if (inodedep->id_mkdiradd != NULL)
8557 panic("softdep_setup_directory_add: Existing mkdir");
8558 inodedep->id_mkdiradd = dap;
8559 } else if (inodedep->id_mkdiradd)
8560 merge_diradd(inodedep, dap);
8561 if (newdirblk != NULL) {
8563 * There is nothing to do if we are already tracking
8566 if ((pagedep->pd_state & NEWBLOCK) != 0) {
8567 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
8571 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
8573 panic("softdep_setup_directory_add: lost entry");
8574 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8575 pagedep->pd_state |= NEWBLOCK;
8576 pagedep->pd_newdirblk = newdirblk;
8577 newdirblk->db_pagedep = pagedep;
8580 * If we extended into an indirect signal direnter to sync.
8591 * This procedure is called to change the offset of a directory
8592 * entry when compacting a directory block which must be owned
8593 * exclusively by the caller. Note that the actual entry movement
8594 * must be done in this procedure to ensure that no I/O completions
8595 * occur while the move is in progress.
8598 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
8599 struct buf *bp; /* Buffer holding directory block. */
8600 struct inode *dp; /* inode for directory */
8601 caddr_t base; /* address of dp->i_offset */
8602 caddr_t oldloc; /* address of old directory location */
8603 caddr_t newloc; /* address of new directory location */
8604 int entrysize; /* size of directory entry */
8606 int offset, oldoffset, newoffset;
8607 struct pagedep *pagedep;
8608 struct jmvref *jmvref;
8615 mp = UFSTOVFS(dp->i_ump);
8616 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8617 ("softdep_change_directoryentry_offset called on "
8618 "non-softdep filesystem"));
8619 de = (struct direct *)oldloc;
8623 * Moves are always journaled as it would be too complex to
8624 * determine if any affected adds or removes are present in the
8627 if (MOUNTEDSUJ(mp)) {
8629 jmvref = newjmvref(dp, de->d_ino,
8630 dp->i_offset + (oldloc - base),
8631 dp->i_offset + (newloc - base));
8633 lbn = lblkno(dp->i_fs, dp->i_offset);
8634 offset = blkoff(dp->i_fs, dp->i_offset);
8635 oldoffset = offset + (oldloc - base);
8636 newoffset = offset + (newloc - base);
8637 ACQUIRE_LOCK(dp->i_ump);
8638 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0)
8640 dap = diradd_lookup(pagedep, oldoffset);
8642 dap->da_offset = newoffset;
8643 newoffset = DIRADDHASH(newoffset);
8644 oldoffset = DIRADDHASH(oldoffset);
8645 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
8646 newoffset != oldoffset) {
8647 LIST_REMOVE(dap, da_pdlist);
8648 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
8654 jmvref->jm_pagedep = pagedep;
8655 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
8656 add_to_journal(&jmvref->jm_list);
8658 bcopy(oldloc, newloc, entrysize);
8659 FREE_LOCK(dp->i_ump);
8663 * Move the mkdir dependencies and journal work from one diradd to another
8664 * when renaming a directory. The new name must depend on the mkdir deps
8665 * completing as the old name did. Directories can only have one valid link
8666 * at a time so one must be canonical.
8669 merge_diradd(inodedep, newdap)
8670 struct inodedep *inodedep;
8671 struct diradd *newdap;
8673 struct diradd *olddap;
8674 struct mkdir *mkdir, *nextmd;
8675 struct ufsmount *ump;
8678 olddap = inodedep->id_mkdiradd;
8679 inodedep->id_mkdiradd = newdap;
8680 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8681 newdap->da_state &= ~DEPCOMPLETE;
8682 ump = VFSTOUFS(inodedep->id_list.wk_mp);
8683 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8685 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8686 if (mkdir->md_diradd != olddap)
8688 mkdir->md_diradd = newdap;
8689 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
8690 newdap->da_state |= state;
8691 olddap->da_state &= ~state;
8692 if ((olddap->da_state &
8693 (MKDIR_PARENT | MKDIR_BODY)) == 0)
8696 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8697 panic("merge_diradd: unfound ref");
8700 * Any mkdir related journal items are not safe to be freed until
8701 * the new name is stable.
8703 jwork_move(&newdap->da_jwork, &olddap->da_jwork);
8704 olddap->da_state |= DEPCOMPLETE;
8705 complete_diradd(olddap);
8709 * Move the diradd to the pending list when all diradd dependencies are
8713 complete_diradd(dap)
8716 struct pagedep *pagedep;
8718 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
8719 if (dap->da_state & DIRCHG)
8720 pagedep = dap->da_previous->dm_pagedep;
8722 pagedep = dap->da_pagedep;
8723 LIST_REMOVE(dap, da_pdlist);
8724 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
8729 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal
8730 * add entries and conditonally journal the remove.
8733 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
8735 struct dirrem *dirrem;
8736 struct jremref *jremref;
8737 struct jremref *dotremref;
8738 struct jremref *dotdotremref;
8740 struct inodedep *inodedep;
8741 struct jaddref *jaddref;
8742 struct inoref *inoref;
8743 struct ufsmount *ump;
8744 struct mkdir *mkdir;
8747 * If no remove references were allocated we're on a non-journaled
8748 * filesystem and can skip the cancel step.
8750 if (jremref == NULL) {
8751 free_diradd(dap, NULL);
8755 * Cancel the primary name an free it if it does not require
8758 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
8759 0, &inodedep) != 0) {
8760 /* Abort the addref that reference this diradd. */
8761 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
8762 if (inoref->if_list.wk_type != D_JADDREF)
8764 jaddref = (struct jaddref *)inoref;
8765 if (jaddref->ja_diradd != dap)
8767 if (cancel_jaddref(jaddref, inodedep,
8768 &dirrem->dm_jwork) == 0) {
8769 free_jremref(jremref);
8776 * Cancel subordinate names and free them if they do not require
8779 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8780 ump = VFSTOUFS(dap->da_list.wk_mp);
8781 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) {
8782 if (mkdir->md_diradd != dap)
8784 if ((jaddref = mkdir->md_jaddref) == NULL)
8786 mkdir->md_jaddref = NULL;
8787 if (mkdir->md_state & MKDIR_PARENT) {
8788 if (cancel_jaddref(jaddref, NULL,
8789 &dirrem->dm_jwork) == 0) {
8790 free_jremref(dotdotremref);
8791 dotdotremref = NULL;
8794 if (cancel_jaddref(jaddref, inodedep,
8795 &dirrem->dm_jwork) == 0) {
8796 free_jremref(dotremref);
8804 journal_jremref(dirrem, jremref, inodedep);
8806 journal_jremref(dirrem, dotremref, inodedep);
8808 journal_jremref(dirrem, dotdotremref, NULL);
8809 jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
8810 free_diradd(dap, &dirrem->dm_jwork);
8814 * Free a diradd dependency structure. This routine must be called
8815 * with splbio interrupts blocked.
8818 free_diradd(dap, wkhd)
8820 struct workhead *wkhd;
8822 struct dirrem *dirrem;
8823 struct pagedep *pagedep;
8824 struct inodedep *inodedep;
8825 struct mkdir *mkdir, *nextmd;
8826 struct ufsmount *ump;
8828 ump = VFSTOUFS(dap->da_list.wk_mp);
8830 LIST_REMOVE(dap, da_pdlist);
8831 if (dap->da_state & ONWORKLIST)
8832 WORKLIST_REMOVE(&dap->da_list);
8833 if ((dap->da_state & DIRCHG) == 0) {
8834 pagedep = dap->da_pagedep;
8836 dirrem = dap->da_previous;
8837 pagedep = dirrem->dm_pagedep;
8838 dirrem->dm_dirinum = pagedep->pd_ino;
8839 dirrem->dm_state |= COMPLETE;
8840 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
8841 add_to_worklist(&dirrem->dm_list, 0);
8843 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
8845 if (inodedep->id_mkdiradd == dap)
8846 inodedep->id_mkdiradd = NULL;
8847 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8848 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8850 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8851 if (mkdir->md_diradd != dap)
8854 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
8855 LIST_REMOVE(mkdir, md_mkdirs);
8856 if (mkdir->md_state & ONWORKLIST)
8857 WORKLIST_REMOVE(&mkdir->md_list);
8858 if (mkdir->md_jaddref != NULL)
8859 panic("free_diradd: Unexpected jaddref");
8860 WORKITEM_FREE(mkdir, D_MKDIR);
8861 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
8864 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8865 panic("free_diradd: unfound ref");
8868 free_inodedep(inodedep);
8870 * Free any journal segments waiting for the directory write.
8872 handle_jwork(&dap->da_jwork);
8873 WORKITEM_FREE(dap, D_DIRADD);
8877 * Directory entry removal dependencies.
8879 * When removing a directory entry, the entry's inode pointer must be
8880 * zero'ed on disk before the corresponding inode's link count is decremented
8881 * (possibly freeing the inode for re-use). This dependency is handled by
8882 * updating the directory entry but delaying the inode count reduction until
8883 * after the directory block has been written to disk. After this point, the
8884 * inode count can be decremented whenever it is convenient.
8888 * This routine should be called immediately after removing
8889 * a directory entry. The inode's link count should not be
8890 * decremented by the calling procedure -- the soft updates
8891 * code will do this task when it is safe.
8894 softdep_setup_remove(bp, dp, ip, isrmdir)
8895 struct buf *bp; /* buffer containing directory block */
8896 struct inode *dp; /* inode for the directory being modified */
8897 struct inode *ip; /* inode for directory entry being removed */
8898 int isrmdir; /* indicates if doing RMDIR */
8900 struct dirrem *dirrem, *prevdirrem;
8901 struct inodedep *inodedep;
8904 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
8905 ("softdep_setup_remove called on non-softdep filesystem"));
8907 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want
8908 * newdirrem() to setup the full directory remove which requires
8911 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
8913 * Add the dirrem to the inodedep's pending remove list for quick
8916 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
8918 panic("softdep_setup_remove: Lost inodedep.");
8919 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
8920 dirrem->dm_state |= ONDEPLIST;
8921 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
8924 * If the COMPLETE flag is clear, then there were no active
8925 * entries and we want to roll back to a zeroed entry until
8926 * the new inode is committed to disk. If the COMPLETE flag is
8927 * set then we have deleted an entry that never made it to
8928 * disk. If the entry we deleted resulted from a name change,
8929 * then the old name still resides on disk. We cannot delete
8930 * its inode (returned to us in prevdirrem) until the zeroed
8931 * directory entry gets to disk. The new inode has never been
8932 * referenced on the disk, so can be deleted immediately.
8934 if ((dirrem->dm_state & COMPLETE) == 0) {
8935 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
8937 FREE_LOCK(ip->i_ump);
8939 if (prevdirrem != NULL)
8940 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
8941 prevdirrem, dm_next);
8942 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
8943 direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
8944 FREE_LOCK(ip->i_ump);
8946 handle_workitem_remove(dirrem, 0);
8951 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
8952 * pd_pendinghd list of a pagedep.
8954 static struct diradd *
8955 diradd_lookup(pagedep, offset)
8956 struct pagedep *pagedep;
8961 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
8962 if (dap->da_offset == offset)
8964 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
8965 if (dap->da_offset == offset)
8971 * Search for a .. diradd dependency in a directory that is being removed.
8972 * If the directory was renamed to a new parent we have a diradd rather
8973 * than a mkdir for the .. entry. We need to cancel it now before
8974 * it is found in truncate().
8976 static struct jremref *
8977 cancel_diradd_dotdot(ip, dirrem, jremref)
8979 struct dirrem *dirrem;
8980 struct jremref *jremref;
8982 struct pagedep *pagedep;
8984 struct worklist *wk;
8986 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0,
8989 dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
8992 cancel_diradd(dap, dirrem, jremref, NULL, NULL);
8994 * Mark any journal work as belonging to the parent so it is freed
8995 * with the .. reference.
8997 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
8998 wk->wk_state |= MKDIR_PARENT;
9003 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
9004 * replace it with a dirrem/diradd pair as a result of re-parenting a
9005 * directory. This ensures that we don't simultaneously have a mkdir and
9006 * a diradd for the same .. entry.
9008 static struct jremref *
9009 cancel_mkdir_dotdot(ip, dirrem, jremref)
9011 struct dirrem *dirrem;
9012 struct jremref *jremref;
9014 struct inodedep *inodedep;
9015 struct jaddref *jaddref;
9016 struct ufsmount *ump;
9017 struct mkdir *mkdir;
9020 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
9023 dap = inodedep->id_mkdiradd;
9024 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
9026 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9027 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
9028 mkdir = LIST_NEXT(mkdir, md_mkdirs))
9029 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
9032 panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
9033 if ((jaddref = mkdir->md_jaddref) != NULL) {
9034 mkdir->md_jaddref = NULL;
9035 jaddref->ja_state &= ~MKDIR_PARENT;
9036 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0,
9038 panic("cancel_mkdir_dotdot: Lost parent inodedep");
9039 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
9040 journal_jremref(dirrem, jremref, inodedep);
9044 if (mkdir->md_state & ONWORKLIST)
9045 WORKLIST_REMOVE(&mkdir->md_list);
9046 mkdir->md_state |= ALLCOMPLETE;
9047 complete_mkdir(mkdir);
9052 journal_jremref(dirrem, jremref, inodedep)
9053 struct dirrem *dirrem;
9054 struct jremref *jremref;
9055 struct inodedep *inodedep;
9058 if (inodedep == NULL)
9059 if (inodedep_lookup(jremref->jr_list.wk_mp,
9060 jremref->jr_ref.if_ino, 0, &inodedep) == 0)
9061 panic("journal_jremref: Lost inodedep");
9062 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
9063 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
9064 add_to_journal(&jremref->jr_list);
9068 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
9069 struct dirrem *dirrem;
9070 struct jremref *jremref;
9071 struct jremref *dotremref;
9072 struct jremref *dotdotremref;
9074 struct inodedep *inodedep;
9077 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
9079 panic("dirrem_journal: Lost inodedep");
9080 journal_jremref(dirrem, jremref, inodedep);
9082 journal_jremref(dirrem, dotremref, inodedep);
9084 journal_jremref(dirrem, dotdotremref, NULL);
9088 * Allocate a new dirrem if appropriate and return it along with
9089 * its associated pagedep. Called without a lock, returns with lock.
9091 static struct dirrem *
9092 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
9093 struct buf *bp; /* buffer containing directory block */
9094 struct inode *dp; /* inode for the directory being modified */
9095 struct inode *ip; /* inode for directory entry being removed */
9096 int isrmdir; /* indicates if doing RMDIR */
9097 struct dirrem **prevdirremp; /* previously referenced inode, if any */
9102 struct dirrem *dirrem;
9103 struct pagedep *pagedep;
9104 struct jremref *jremref;
9105 struct jremref *dotremref;
9106 struct jremref *dotdotremref;
9110 * Whiteouts have no deletion dependencies.
9113 panic("newdirrem: whiteout");
9116 * If the system is over its limit and our filesystem is
9117 * responsible for more than our share of that usage and
9118 * we are not a snapshot, request some inodedep cleanup.
9119 * Limiting the number of dirrem structures will also limit
9120 * the number of freefile and freeblks structures.
9122 ACQUIRE_LOCK(ip->i_ump);
9123 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ip->i_ump, D_DIRREM))
9124 schedule_cleanup(ITOV(dp)->v_mount);
9126 FREE_LOCK(ip->i_ump);
9127 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS |
9129 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
9130 LIST_INIT(&dirrem->dm_jremrefhd);
9131 LIST_INIT(&dirrem->dm_jwork);
9132 dirrem->dm_state = isrmdir ? RMDIR : 0;
9133 dirrem->dm_oldinum = ip->i_number;
9134 *prevdirremp = NULL;
9136 * Allocate remove reference structures to track journal write
9137 * dependencies. We will always have one for the link and
9138 * when doing directories we will always have one more for dot.
9139 * When renaming a directory we skip the dotdot link change so
9140 * this is not needed.
9142 jremref = dotremref = dotdotremref = NULL;
9143 if (DOINGSUJ(dvp)) {
9145 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9146 ip->i_effnlink + 2);
9147 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
9148 ip->i_effnlink + 1);
9149 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
9150 dp->i_effnlink + 1);
9151 dotdotremref->jr_state |= MKDIR_PARENT;
9153 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9154 ip->i_effnlink + 1);
9156 ACQUIRE_LOCK(ip->i_ump);
9157 lbn = lblkno(dp->i_fs, dp->i_offset);
9158 offset = blkoff(dp->i_fs, dp->i_offset);
9159 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC,
9161 dirrem->dm_pagedep = pagedep;
9162 dirrem->dm_offset = offset;
9164 * If we're renaming a .. link to a new directory, cancel any
9165 * existing MKDIR_PARENT mkdir. If it has already been canceled
9166 * the jremref is preserved for any potential diradd in this
9167 * location. This can not coincide with a rmdir.
9169 if (dp->i_offset == DOTDOT_OFFSET) {
9171 panic("newdirrem: .. directory change during remove?");
9172 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
9175 * If we're removing a directory search for the .. dependency now and
9176 * cancel it. Any pending journal work will be added to the dirrem
9177 * to be completed when the workitem remove completes.
9180 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
9182 * Check for a diradd dependency for the same directory entry.
9183 * If present, then both dependencies become obsolete and can
9186 dap = diradd_lookup(pagedep, offset);
9189 * Link the jremref structures into the dirrem so they are
9190 * written prior to the pagedep.
9193 dirrem_journal(dirrem, jremref, dotremref,
9198 * Must be ATTACHED at this point.
9200 if ((dap->da_state & ATTACHED) == 0)
9201 panic("newdirrem: not ATTACHED");
9202 if (dap->da_newinum != ip->i_number)
9203 panic("newdirrem: inum %ju should be %ju",
9204 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum);
9206 * If we are deleting a changed name that never made it to disk,
9207 * then return the dirrem describing the previous inode (which
9208 * represents the inode currently referenced from this entry on disk).
9210 if ((dap->da_state & DIRCHG) != 0) {
9211 *prevdirremp = dap->da_previous;
9212 dap->da_state &= ~DIRCHG;
9213 dap->da_pagedep = pagedep;
9216 * We are deleting an entry that never made it to disk.
9217 * Mark it COMPLETE so we can delete its inode immediately.
9219 dirrem->dm_state |= COMPLETE;
9220 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
9223 struct worklist *wk;
9225 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
9226 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
9227 panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
9235 * Directory entry change dependencies.
9237 * Changing an existing directory entry requires that an add operation
9238 * be completed first followed by a deletion. The semantics for the addition
9239 * are identical to the description of adding a new entry above except
9240 * that the rollback is to the old inode number rather than zero. Once
9241 * the addition dependency is completed, the removal is done as described
9242 * in the removal routine above.
9246 * This routine should be called immediately after changing
9247 * a directory entry. The inode's link count should not be
9248 * decremented by the calling procedure -- the soft updates
9249 * code will perform this task when it is safe.
9252 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
9253 struct buf *bp; /* buffer containing directory block */
9254 struct inode *dp; /* inode for the directory being modified */
9255 struct inode *ip; /* inode for directory entry being removed */
9256 ino_t newinum; /* new inode number for changed entry */
9257 int isrmdir; /* indicates if doing RMDIR */
9260 struct diradd *dap = NULL;
9261 struct dirrem *dirrem, *prevdirrem;
9262 struct pagedep *pagedep;
9263 struct inodedep *inodedep;
9264 struct jaddref *jaddref;
9267 offset = blkoff(dp->i_fs, dp->i_offset);
9268 mp = UFSTOVFS(dp->i_ump);
9269 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
9270 ("softdep_setup_directory_change called on non-softdep filesystem"));
9273 * Whiteouts do not need diradd dependencies.
9275 if (newinum != WINO) {
9276 dap = malloc(sizeof(struct diradd),
9277 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
9278 workitem_alloc(&dap->da_list, D_DIRADD, mp);
9279 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
9280 dap->da_offset = offset;
9281 dap->da_newinum = newinum;
9282 LIST_INIT(&dap->da_jwork);
9286 * Allocate a new dirrem and ACQUIRE_LOCK.
9288 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
9289 pagedep = dirrem->dm_pagedep;
9291 * The possible values for isrmdir:
9292 * 0 - non-directory file rename
9293 * 1 - directory rename within same directory
9294 * inum - directory rename to new directory of given inode number
9295 * When renaming to a new directory, we are both deleting and
9296 * creating a new directory entry, so the link count on the new
9297 * directory should not change. Thus we do not need the followup
9298 * dirrem which is usually done in handle_workitem_remove. We set
9299 * the DIRCHG flag to tell handle_workitem_remove to skip the
9303 dirrem->dm_state |= DIRCHG;
9306 * Whiteouts have no additional dependencies,
9307 * so just put the dirrem on the correct list.
9309 if (newinum == WINO) {
9310 if ((dirrem->dm_state & COMPLETE) == 0) {
9311 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
9314 dirrem->dm_dirinum = pagedep->pd_ino;
9315 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9316 add_to_worklist(&dirrem->dm_list, 0);
9318 FREE_LOCK(dp->i_ump);
9322 * Add the dirrem to the inodedep's pending remove list for quick
9323 * discovery later. A valid nlinkdelta ensures that this lookup
9326 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
9327 panic("softdep_setup_directory_change: Lost inodedep.");
9328 dirrem->dm_state |= ONDEPLIST;
9329 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9332 * If the COMPLETE flag is clear, then there were no active
9333 * entries and we want to roll back to the previous inode until
9334 * the new inode is committed to disk. If the COMPLETE flag is
9335 * set, then we have deleted an entry that never made it to disk.
9336 * If the entry we deleted resulted from a name change, then the old
9337 * inode reference still resides on disk. Any rollback that we do
9338 * needs to be to that old inode (returned to us in prevdirrem). If
9339 * the entry we deleted resulted from a create, then there is
9340 * no entry on the disk, so we want to roll back to zero rather
9341 * than the uncommitted inode. In either of the COMPLETE cases we
9342 * want to immediately free the unwritten and unreferenced inode.
9344 if ((dirrem->dm_state & COMPLETE) == 0) {
9345 dap->da_previous = dirrem;
9347 if (prevdirrem != NULL) {
9348 dap->da_previous = prevdirrem;
9350 dap->da_state &= ~DIRCHG;
9351 dap->da_pagedep = pagedep;
9353 dirrem->dm_dirinum = pagedep->pd_ino;
9354 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9355 add_to_worklist(&dirrem->dm_list, 0);
9358 * Lookup the jaddref for this journal entry. We must finish
9359 * initializing it and make the diradd write dependent on it.
9360 * If we're not journaling, put it on the id_bufwait list if the
9361 * inode is not yet written. If it is written, do the post-inode
9362 * write processing to put it on the id_pendinghd list.
9364 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
9365 if (MOUNTEDSUJ(mp)) {
9366 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
9368 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
9369 ("softdep_setup_directory_change: bad jaddref %p",
9371 jaddref->ja_diroff = dp->i_offset;
9372 jaddref->ja_diradd = dap;
9373 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9375 add_to_journal(&jaddref->ja_list);
9376 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
9377 dap->da_state |= COMPLETE;
9378 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
9379 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
9381 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9383 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
9386 * If we're making a new name for a directory that has not been
9387 * committed when need to move the dot and dotdot references to
9390 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
9391 merge_diradd(inodedep, dap);
9392 FREE_LOCK(dp->i_ump);
9396 * Called whenever the link count on an inode is changed.
9397 * It creates an inode dependency so that the new reference(s)
9398 * to the inode cannot be committed to disk until the updated
9399 * inode has been written.
9402 softdep_change_linkcnt(ip)
9403 struct inode *ip; /* the inode with the increased link count */
9405 struct inodedep *inodedep;
9407 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
9408 ("softdep_change_linkcnt called on non-softdep filesystem"));
9409 ACQUIRE_LOCK(ip->i_ump);
9410 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
9412 if (ip->i_nlink < ip->i_effnlink)
9413 panic("softdep_change_linkcnt: bad delta");
9414 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9415 FREE_LOCK(ip->i_ump);
9419 * Attach a sbdep dependency to the superblock buf so that we can keep
9420 * track of the head of the linked list of referenced but unlinked inodes.
9423 softdep_setup_sbupdate(ump, fs, bp)
9424 struct ufsmount *ump;
9428 struct sbdep *sbdep;
9429 struct worklist *wk;
9431 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
9432 ("softdep_setup_sbupdate called on non-softdep filesystem"));
9433 LIST_FOREACH(wk, &bp->b_dep, wk_list)
9434 if (wk->wk_type == D_SBDEP)
9438 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
9439 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
9441 sbdep->sb_ump = ump;
9443 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
9448 * Return the first unlinked inodedep which is ready to be the head of the
9449 * list. The inodedep and all those after it must have valid next pointers.
9451 static struct inodedep *
9452 first_unlinked_inodedep(ump)
9453 struct ufsmount *ump;
9455 struct inodedep *inodedep;
9456 struct inodedep *idp;
9459 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
9460 inodedep; inodedep = idp) {
9461 if ((inodedep->id_state & UNLINKNEXT) == 0)
9463 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9464 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
9466 if ((inodedep->id_state & UNLINKPREV) == 0)
9473 * Set the sujfree unlinked head pointer prior to writing a superblock.
9476 initiate_write_sbdep(sbdep)
9477 struct sbdep *sbdep;
9479 struct inodedep *inodedep;
9483 bpfs = sbdep->sb_fs;
9484 fs = sbdep->sb_ump->um_fs;
9485 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9487 fs->fs_sujfree = inodedep->id_ino;
9488 inodedep->id_state |= UNLINKPREV;
9491 bpfs->fs_sujfree = fs->fs_sujfree;
9495 * After a superblock is written determine whether it must be written again
9496 * due to a changing unlinked list head.
9499 handle_written_sbdep(sbdep, bp)
9500 struct sbdep *sbdep;
9503 struct inodedep *inodedep;
9506 LOCK_OWNED(sbdep->sb_ump);
9509 * If the superblock doesn't match the in-memory list start over.
9511 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9512 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
9513 (inodedep == NULL && fs->fs_sujfree != 0)) {
9517 WORKITEM_FREE(sbdep, D_SBDEP);
9518 if (fs->fs_sujfree == 0)
9521 * Now that we have a record of this inode in stable store allow it
9522 * to be written to free up pending work. Inodes may see a lot of
9523 * write activity after they are unlinked which we must not hold up.
9525 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
9526 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
9527 panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
9528 inodedep, inodedep->id_state);
9529 if (inodedep->id_state & UNLINKONLIST)
9531 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
9538 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
9541 unlinked_inodedep(mp, inodedep)
9543 struct inodedep *inodedep;
9545 struct ufsmount *ump;
9549 if (MOUNTEDSUJ(mp) == 0)
9551 ump->um_fs->fs_fmod = 1;
9552 if (inodedep->id_state & UNLINKED)
9553 panic("unlinked_inodedep: %p already unlinked\n", inodedep);
9554 inodedep->id_state |= UNLINKED;
9555 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
9559 * Remove an inodedep from the unlinked inodedep list. This may require
9560 * disk writes if the inode has made it that far.
9563 clear_unlinked_inodedep(inodedep)
9564 struct inodedep *inodedep;
9566 struct ufsmount *ump;
9567 struct inodedep *idp;
9568 struct inodedep *idn;
9576 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9578 ino = inodedep->id_ino;
9582 KASSERT((inodedep->id_state & UNLINKED) != 0,
9583 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9586 * If nothing has yet been written simply remove us from
9587 * the in memory list and return. This is the most common
9588 * case where handle_workitem_remove() loses the final
9591 if ((inodedep->id_state & UNLINKLINKS) == 0)
9594 * If we have a NEXT pointer and no PREV pointer we can simply
9595 * clear NEXT's PREV and remove ourselves from the list. Be
9596 * careful not to clear PREV if the superblock points at
9599 idn = TAILQ_NEXT(inodedep, id_unlinked);
9600 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
9601 if (idn && fs->fs_sujfree != idn->id_ino)
9602 idn->id_state &= ~UNLINKPREV;
9606 * Here we have an inodedep which is actually linked into
9607 * the list. We must remove it by forcing a write to the
9608 * link before us, whether it be the superblock or an inode.
9609 * Unfortunately the list may change while we're waiting
9610 * on the buf lock for either resource so we must loop until
9611 * we lock the right one. If both the superblock and an
9612 * inode point to this inode we must clear the inode first
9613 * followed by the superblock.
9615 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9617 if (idp && (idp->id_state & UNLINKNEXT))
9621 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9622 (int)fs->fs_sbsize, 0, 0, 0);
9624 error = bread(ump->um_devvp,
9625 fsbtodb(fs, ino_to_fsba(fs, pino)),
9626 (int)fs->fs_bsize, NOCRED, &bp);
9633 /* If the list has changed restart the loop. */
9634 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9636 if (idp && (idp->id_state & UNLINKNEXT))
9639 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
9646 idn = TAILQ_NEXT(inodedep, id_unlinked);
9650 * Remove us from the in memory list. After this we cannot
9651 * access the inodedep.
9653 KASSERT((inodedep->id_state & UNLINKED) != 0,
9654 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9656 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9657 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9660 * The predecessor's next pointer is manually updated here
9661 * so that the NEXT flag is never cleared for an element
9662 * that is in the list.
9665 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9666 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9667 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9669 } else if (fs->fs_magic == FS_UFS1_MAGIC)
9670 ((struct ufs1_dinode *)bp->b_data +
9671 ino_to_fsbo(fs, pino))->di_freelink = nino;
9673 ((struct ufs2_dinode *)bp->b_data +
9674 ino_to_fsbo(fs, pino))->di_freelink = nino;
9676 * If the bwrite fails we have no recourse to recover. The
9677 * filesystem is corrupted already.
9682 * If the superblock pointer still needs to be cleared force
9685 if (fs->fs_sujfree == ino) {
9687 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9688 (int)fs->fs_sbsize, 0, 0, 0);
9689 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9690 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9691 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9697 if (fs->fs_sujfree != ino)
9699 panic("clear_unlinked_inodedep: Failed to clear free head");
9701 if (inodedep->id_ino == fs->fs_sujfree)
9702 panic("clear_unlinked_inodedep: Freeing head of free list");
9703 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9704 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9709 * This workitem decrements the inode's link count.
9710 * If the link count reaches zero, the file is removed.
9713 handle_workitem_remove(dirrem, flags)
9714 struct dirrem *dirrem;
9717 struct inodedep *inodedep;
9718 struct workhead dotdotwk;
9719 struct worklist *wk;
9720 struct ufsmount *ump;
9726 if (dirrem->dm_state & ONWORKLIST)
9727 panic("handle_workitem_remove: dirrem %p still on worklist",
9729 oldinum = dirrem->dm_oldinum;
9730 mp = dirrem->dm_list.wk_mp;
9732 flags |= LK_EXCLUSIVE;
9733 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0)
9737 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
9738 panic("handle_workitem_remove: lost inodedep");
9739 if (dirrem->dm_state & ONDEPLIST)
9740 LIST_REMOVE(dirrem, dm_inonext);
9741 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
9742 ("handle_workitem_remove: Journal entries not written."));
9745 * Move all dependencies waiting on the remove to complete
9746 * from the dirrem to the inode inowait list to be completed
9747 * after the inode has been updated and written to disk. Any
9748 * marked MKDIR_PARENT are saved to be completed when the .. ref
9751 LIST_INIT(&dotdotwk);
9752 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
9753 WORKLIST_REMOVE(wk);
9754 if (wk->wk_state & MKDIR_PARENT) {
9755 wk->wk_state &= ~MKDIR_PARENT;
9756 WORKLIST_INSERT(&dotdotwk, wk);
9759 WORKLIST_INSERT(&inodedep->id_inowait, wk);
9761 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
9763 * Normal file deletion.
9765 if ((dirrem->dm_state & RMDIR) == 0) {
9767 DIP_SET(ip, i_nlink, ip->i_nlink);
9768 ip->i_flag |= IN_CHANGE;
9769 if (ip->i_nlink < ip->i_effnlink)
9770 panic("handle_workitem_remove: bad file delta");
9771 if (ip->i_nlink == 0)
9772 unlinked_inodedep(mp, inodedep);
9773 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9774 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9775 ("handle_workitem_remove: worklist not empty. %s",
9776 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
9777 WORKITEM_FREE(dirrem, D_DIRREM);
9782 * Directory deletion. Decrement reference count for both the
9783 * just deleted parent directory entry and the reference for ".".
9784 * Arrange to have the reference count on the parent decremented
9785 * to account for the loss of "..".
9788 DIP_SET(ip, i_nlink, ip->i_nlink);
9789 ip->i_flag |= IN_CHANGE;
9790 if (ip->i_nlink < ip->i_effnlink)
9791 panic("handle_workitem_remove: bad dir delta");
9792 if (ip->i_nlink == 0)
9793 unlinked_inodedep(mp, inodedep);
9794 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9796 * Rename a directory to a new parent. Since, we are both deleting
9797 * and creating a new directory entry, the link count on the new
9798 * directory should not change. Thus we skip the followup dirrem.
9800 if (dirrem->dm_state & DIRCHG) {
9801 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9802 ("handle_workitem_remove: DIRCHG and worklist not empty."));
9803 WORKITEM_FREE(dirrem, D_DIRREM);
9807 dirrem->dm_state = ONDEPLIST;
9808 dirrem->dm_oldinum = dirrem->dm_dirinum;
9810 * Place the dirrem on the parent's diremhd list.
9812 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
9813 panic("handle_workitem_remove: lost dir inodedep");
9814 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9816 * If the allocated inode has never been written to disk, then
9817 * the on-disk inode is zero'ed and we can remove the file
9818 * immediately. When journaling if the inode has been marked
9819 * unlinked and not DEPCOMPLETE we know it can never be written.
9821 inodedep_lookup(mp, oldinum, 0, &inodedep);
9822 if (inodedep == NULL ||
9823 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
9824 check_inode_unwritten(inodedep)) {
9827 return handle_workitem_remove(dirrem, flags);
9829 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
9831 ip->i_flag |= IN_CHANGE;
9839 * Inode de-allocation dependencies.
9841 * When an inode's link count is reduced to zero, it can be de-allocated. We
9842 * found it convenient to postpone de-allocation until after the inode is
9843 * written to disk with its new link count (zero). At this point, all of the
9844 * on-disk inode's block pointers are nullified and, with careful dependency
9845 * list ordering, all dependencies related to the inode will be satisfied and
9846 * the corresponding dependency structures de-allocated. So, if/when the
9847 * inode is reused, there will be no mixing of old dependencies with new
9848 * ones. This artificial dependency is set up by the block de-allocation
9849 * procedure above (softdep_setup_freeblocks) and completed by the
9850 * following procedure.
9853 handle_workitem_freefile(freefile)
9854 struct freefile *freefile;
9856 struct workhead wkhd;
9858 struct inodedep *idp;
9859 struct ufsmount *ump;
9862 ump = VFSTOUFS(freefile->fx_list.wk_mp);
9866 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
9869 panic("handle_workitem_freefile: inodedep %p survived", idp);
9872 fs->fs_pendinginodes -= 1;
9875 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
9876 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
9877 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
9878 softdep_error("handle_workitem_freefile", error);
9880 WORKITEM_FREE(freefile, D_FREEFILE);
9886 * Helper function which unlinks marker element from work list and returns
9887 * the next element on the list.
9889 static __inline struct worklist *
9890 markernext(struct worklist *marker)
9892 struct worklist *next;
9894 next = LIST_NEXT(marker, wk_list);
9895 LIST_REMOVE(marker, wk_list);
9902 * The dependency structures constructed above are most actively used when file
9903 * system blocks are written to disk. No constraints are placed on when a
9904 * block can be written, but unsatisfied update dependencies are made safe by
9905 * modifying (or replacing) the source memory for the duration of the disk
9906 * write. When the disk write completes, the memory block is again brought
9909 * In-core inode structure reclamation.
9911 * Because there are a finite number of "in-core" inode structures, they are
9912 * reused regularly. By transferring all inode-related dependencies to the
9913 * in-memory inode block and indexing them separately (via "inodedep"s), we
9914 * can allow "in-core" inode structures to be reused at any time and avoid
9915 * any increase in contention.
9917 * Called just before entering the device driver to initiate a new disk I/O.
9918 * The buffer must be locked, thus, no I/O completion operations can occur
9919 * while we are manipulating its associated dependencies.
9922 softdep_disk_io_initiation(bp)
9923 struct buf *bp; /* structure describing disk write to occur */
9925 struct worklist *wk;
9926 struct worklist marker;
9927 struct inodedep *inodedep;
9928 struct freeblks *freeblks;
9929 struct jblkdep *jblkdep;
9930 struct newblk *newblk;
9931 struct ufsmount *ump;
9934 * We only care about write operations. There should never
9935 * be dependencies for reads.
9937 if (bp->b_iocmd != BIO_WRITE)
9938 panic("softdep_disk_io_initiation: not write");
9940 if (bp->b_vflags & BV_BKGRDINPROG)
9941 panic("softdep_disk_io_initiation: Writing buffer with "
9942 "background write in progress: %p", bp);
9944 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
9946 ump = VFSTOUFS(wk->wk_mp);
9948 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
9949 PHOLD(curproc); /* Don't swap out kernel stack */
9952 * Do any necessary pre-I/O processing.
9954 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
9955 wk = markernext(&marker)) {
9956 LIST_INSERT_AFTER(wk, &marker, wk_list);
9957 switch (wk->wk_type) {
9960 initiate_write_filepage(WK_PAGEDEP(wk), bp);
9964 inodedep = WK_INODEDEP(wk);
9965 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
9966 initiate_write_inodeblock_ufs1(inodedep, bp);
9968 initiate_write_inodeblock_ufs2(inodedep, bp);
9972 initiate_write_indirdep(WK_INDIRDEP(wk), bp);
9976 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
9980 WK_JSEG(wk)->js_buf = NULL;
9984 freeblks = WK_FREEBLKS(wk);
9985 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd);
9987 * We have to wait for the freeblks to be journaled
9988 * before we can write an inodeblock with updated
9989 * pointers. Be careful to arrange the marker so
9990 * we revisit the freeblks if it's not removed by
9991 * the first jwait().
9993 if (jblkdep != NULL) {
9994 LIST_REMOVE(&marker, wk_list);
9995 LIST_INSERT_BEFORE(wk, &marker, wk_list);
9996 jwait(&jblkdep->jb_list, MNT_WAIT);
10002 * We have to wait for the jnewblk to be journaled
10003 * before we can write to a block if the contents
10004 * may be confused with an earlier file's indirect
10005 * at recovery time. Handle the marker as described
10008 newblk = WK_NEWBLK(wk);
10009 if (newblk->nb_jnewblk != NULL &&
10010 indirblk_lookup(newblk->nb_list.wk_mp,
10011 newblk->nb_newblkno)) {
10012 LIST_REMOVE(&marker, wk_list);
10013 LIST_INSERT_BEFORE(wk, &marker, wk_list);
10014 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
10019 initiate_write_sbdep(WK_SBDEP(wk));
10029 panic("handle_disk_io_initiation: Unexpected type %s",
10030 TYPENAME(wk->wk_type));
10035 PRELE(curproc); /* Allow swapout of kernel stack */
10039 * Called from within the procedure above to deal with unsatisfied
10040 * allocation dependencies in a directory. The buffer must be locked,
10041 * thus, no I/O completion operations can occur while we are
10042 * manipulating its associated dependencies.
10045 initiate_write_filepage(pagedep, bp)
10046 struct pagedep *pagedep;
10049 struct jremref *jremref;
10050 struct jmvref *jmvref;
10051 struct dirrem *dirrem;
10052 struct diradd *dap;
10056 if (pagedep->pd_state & IOSTARTED) {
10058 * This can only happen if there is a driver that does not
10059 * understand chaining. Here biodone will reissue the call
10060 * to strategy for the incomplete buffers.
10062 printf("initiate_write_filepage: already started\n");
10065 pagedep->pd_state |= IOSTARTED;
10067 * Wait for all journal remove dependencies to hit the disk.
10068 * We can not allow any potentially conflicting directory adds
10069 * to be visible before removes and rollback is too difficult.
10070 * The per-filesystem lock may be dropped and re-acquired, however
10071 * we hold the buf locked so the dependency can not go away.
10073 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
10074 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL)
10075 jwait(&jremref->jr_list, MNT_WAIT);
10076 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL)
10077 jwait(&jmvref->jm_list, MNT_WAIT);
10078 for (i = 0; i < DAHASHSZ; i++) {
10079 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
10080 ep = (struct direct *)
10081 ((char *)bp->b_data + dap->da_offset);
10082 if (ep->d_ino != dap->da_newinum)
10083 panic("%s: dir inum %ju != new %ju",
10084 "initiate_write_filepage",
10085 (uintmax_t)ep->d_ino,
10086 (uintmax_t)dap->da_newinum);
10087 if (dap->da_state & DIRCHG)
10088 ep->d_ino = dap->da_previous->dm_oldinum;
10091 dap->da_state &= ~ATTACHED;
10092 dap->da_state |= UNDONE;
10098 * Version of initiate_write_inodeblock that handles UFS1 dinodes.
10099 * Note that any bug fixes made to this routine must be done in the
10100 * version found below.
10102 * Called from within the procedure above to deal with unsatisfied
10103 * allocation dependencies in an inodeblock. The buffer must be
10104 * locked, thus, no I/O completion operations can occur while we
10105 * are manipulating its associated dependencies.
10108 initiate_write_inodeblock_ufs1(inodedep, bp)
10109 struct inodedep *inodedep;
10110 struct buf *bp; /* The inode block */
10112 struct allocdirect *adp, *lastadp;
10113 struct ufs1_dinode *dp;
10114 struct ufs1_dinode *sip;
10115 struct inoref *inoref;
10116 struct ufsmount *ump;
10120 ufs_lbn_t prevlbn = 0;
10124 if (inodedep->id_state & IOSTARTED)
10125 panic("initiate_write_inodeblock_ufs1: already started");
10126 inodedep->id_state |= IOSTARTED;
10127 fs = inodedep->id_fs;
10128 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10130 dp = (struct ufs1_dinode *)bp->b_data +
10131 ino_to_fsbo(fs, inodedep->id_ino);
10134 * If we're on the unlinked list but have not yet written our
10135 * next pointer initialize it here.
10137 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10138 struct inodedep *inon;
10140 inon = TAILQ_NEXT(inodedep, id_unlinked);
10141 dp->di_freelink = inon ? inon->id_ino : 0;
10144 * If the bitmap is not yet written, then the allocated
10145 * inode cannot be written to disk.
10147 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10148 if (inodedep->id_savedino1 != NULL)
10149 panic("initiate_write_inodeblock_ufs1: I/O underway");
10151 sip = malloc(sizeof(struct ufs1_dinode),
10152 M_SAVEDINO, M_SOFTDEP_FLAGS);
10154 inodedep->id_savedino1 = sip;
10155 *inodedep->id_savedino1 = *dp;
10156 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
10157 dp->di_gen = inodedep->id_savedino1->di_gen;
10158 dp->di_freelink = inodedep->id_savedino1->di_freelink;
10162 * If no dependencies, then there is nothing to roll back.
10164 inodedep->id_savedsize = dp->di_size;
10165 inodedep->id_savedextsize = 0;
10166 inodedep->id_savednlink = dp->di_nlink;
10167 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10168 TAILQ_EMPTY(&inodedep->id_inoreflst))
10171 * Revert the link count to that of the first unwritten journal entry.
10173 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10175 dp->di_nlink = inoref->if_nlink;
10177 * Set the dependencies to busy.
10179 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10180 adp = TAILQ_NEXT(adp, ad_next)) {
10182 if (deplist != 0 && prevlbn >= adp->ad_offset)
10183 panic("softdep_write_inodeblock: lbn order");
10184 prevlbn = adp->ad_offset;
10185 if (adp->ad_offset < NDADDR &&
10186 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10187 panic("%s: direct pointer #%jd mismatch %d != %jd",
10188 "softdep_write_inodeblock",
10189 (intmax_t)adp->ad_offset,
10190 dp->di_db[adp->ad_offset],
10191 (intmax_t)adp->ad_newblkno);
10192 if (adp->ad_offset >= NDADDR &&
10193 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10194 panic("%s: indirect pointer #%jd mismatch %d != %jd",
10195 "softdep_write_inodeblock",
10196 (intmax_t)adp->ad_offset - NDADDR,
10197 dp->di_ib[adp->ad_offset - NDADDR],
10198 (intmax_t)adp->ad_newblkno);
10199 deplist |= 1 << adp->ad_offset;
10200 if ((adp->ad_state & ATTACHED) == 0)
10201 panic("softdep_write_inodeblock: Unknown state 0x%x",
10203 #endif /* INVARIANTS */
10204 adp->ad_state &= ~ATTACHED;
10205 adp->ad_state |= UNDONE;
10208 * The on-disk inode cannot claim to be any larger than the last
10209 * fragment that has been written. Otherwise, the on-disk inode
10210 * might have fragments that were not the last block in the file
10211 * which would corrupt the filesystem.
10213 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10214 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10215 if (adp->ad_offset >= NDADDR)
10217 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10218 /* keep going until hitting a rollback to a frag */
10219 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10221 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10222 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10224 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10225 panic("softdep_write_inodeblock: lost dep1");
10226 #endif /* INVARIANTS */
10229 for (i = 0; i < NIADDR; i++) {
10231 if (dp->di_ib[i] != 0 &&
10232 (deplist & ((1 << NDADDR) << i)) == 0)
10233 panic("softdep_write_inodeblock: lost dep2");
10234 #endif /* INVARIANTS */
10240 * If we have zero'ed out the last allocated block of the file,
10241 * roll back the size to the last currently allocated block.
10242 * We know that this last allocated block is a full-sized as
10243 * we already checked for fragments in the loop above.
10245 if (lastadp != NULL &&
10246 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10247 for (i = lastadp->ad_offset; i >= 0; i--)
10248 if (dp->di_db[i] != 0)
10250 dp->di_size = (i + 1) * fs->fs_bsize;
10253 * The only dependencies are for indirect blocks.
10255 * The file size for indirect block additions is not guaranteed.
10256 * Such a guarantee would be non-trivial to achieve. The conventional
10257 * synchronous write implementation also does not make this guarantee.
10258 * Fsck should catch and fix discrepancies. Arguably, the file size
10259 * can be over-estimated without destroying integrity when the file
10260 * moves into the indirect blocks (i.e., is large). If we want to
10261 * postpone fsck, we are stuck with this argument.
10263 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10264 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10268 * Version of initiate_write_inodeblock that handles UFS2 dinodes.
10269 * Note that any bug fixes made to this routine must be done in the
10270 * version found above.
10272 * Called from within the procedure above to deal with unsatisfied
10273 * allocation dependencies in an inodeblock. The buffer must be
10274 * locked, thus, no I/O completion operations can occur while we
10275 * are manipulating its associated dependencies.
10278 initiate_write_inodeblock_ufs2(inodedep, bp)
10279 struct inodedep *inodedep;
10280 struct buf *bp; /* The inode block */
10282 struct allocdirect *adp, *lastadp;
10283 struct ufs2_dinode *dp;
10284 struct ufs2_dinode *sip;
10285 struct inoref *inoref;
10286 struct ufsmount *ump;
10290 ufs_lbn_t prevlbn = 0;
10294 if (inodedep->id_state & IOSTARTED)
10295 panic("initiate_write_inodeblock_ufs2: already started");
10296 inodedep->id_state |= IOSTARTED;
10297 fs = inodedep->id_fs;
10298 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10300 dp = (struct ufs2_dinode *)bp->b_data +
10301 ino_to_fsbo(fs, inodedep->id_ino);
10304 * If we're on the unlinked list but have not yet written our
10305 * next pointer initialize it here.
10307 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10308 struct inodedep *inon;
10310 inon = TAILQ_NEXT(inodedep, id_unlinked);
10311 dp->di_freelink = inon ? inon->id_ino : 0;
10314 * If the bitmap is not yet written, then the allocated
10315 * inode cannot be written to disk.
10317 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10318 if (inodedep->id_savedino2 != NULL)
10319 panic("initiate_write_inodeblock_ufs2: I/O underway");
10321 sip = malloc(sizeof(struct ufs2_dinode),
10322 M_SAVEDINO, M_SOFTDEP_FLAGS);
10324 inodedep->id_savedino2 = sip;
10325 *inodedep->id_savedino2 = *dp;
10326 bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
10327 dp->di_gen = inodedep->id_savedino2->di_gen;
10328 dp->di_freelink = inodedep->id_savedino2->di_freelink;
10332 * If no dependencies, then there is nothing to roll back.
10334 inodedep->id_savedsize = dp->di_size;
10335 inodedep->id_savedextsize = dp->di_extsize;
10336 inodedep->id_savednlink = dp->di_nlink;
10337 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10338 TAILQ_EMPTY(&inodedep->id_extupdt) &&
10339 TAILQ_EMPTY(&inodedep->id_inoreflst))
10342 * Revert the link count to that of the first unwritten journal entry.
10344 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10346 dp->di_nlink = inoref->if_nlink;
10349 * Set the ext data dependencies to busy.
10351 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10352 adp = TAILQ_NEXT(adp, ad_next)) {
10354 if (deplist != 0 && prevlbn >= adp->ad_offset)
10355 panic("softdep_write_inodeblock: lbn order");
10356 prevlbn = adp->ad_offset;
10357 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
10358 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10359 "softdep_write_inodeblock",
10360 (intmax_t)adp->ad_offset,
10361 (intmax_t)dp->di_extb[adp->ad_offset],
10362 (intmax_t)adp->ad_newblkno);
10363 deplist |= 1 << adp->ad_offset;
10364 if ((adp->ad_state & ATTACHED) == 0)
10365 panic("softdep_write_inodeblock: Unknown state 0x%x",
10367 #endif /* INVARIANTS */
10368 adp->ad_state &= ~ATTACHED;
10369 adp->ad_state |= UNDONE;
10372 * The on-disk inode cannot claim to be any larger than the last
10373 * fragment that has been written. Otherwise, the on-disk inode
10374 * might have fragments that were not the last block in the ext
10375 * data which would corrupt the filesystem.
10377 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10378 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10379 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
10380 /* keep going until hitting a rollback to a frag */
10381 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10383 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10384 for (i = adp->ad_offset + 1; i < NXADDR; i++) {
10386 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
10387 panic("softdep_write_inodeblock: lost dep1");
10388 #endif /* INVARIANTS */
10389 dp->di_extb[i] = 0;
10395 * If we have zero'ed out the last allocated block of the ext
10396 * data, roll back the size to the last currently allocated block.
10397 * We know that this last allocated block is a full-sized as
10398 * we already checked for fragments in the loop above.
10400 if (lastadp != NULL &&
10401 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10402 for (i = lastadp->ad_offset; i >= 0; i--)
10403 if (dp->di_extb[i] != 0)
10405 dp->di_extsize = (i + 1) * fs->fs_bsize;
10408 * Set the file data dependencies to busy.
10410 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10411 adp = TAILQ_NEXT(adp, ad_next)) {
10413 if (deplist != 0 && prevlbn >= adp->ad_offset)
10414 panic("softdep_write_inodeblock: lbn order");
10415 if ((adp->ad_state & ATTACHED) == 0)
10416 panic("inodedep %p and adp %p not attached", inodedep, adp);
10417 prevlbn = adp->ad_offset;
10418 if (adp->ad_offset < NDADDR &&
10419 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10420 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10421 "softdep_write_inodeblock",
10422 (intmax_t)adp->ad_offset,
10423 (intmax_t)dp->di_db[adp->ad_offset],
10424 (intmax_t)adp->ad_newblkno);
10425 if (adp->ad_offset >= NDADDR &&
10426 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10427 panic("%s indirect pointer #%jd mismatch %jd != %jd",
10428 "softdep_write_inodeblock:",
10429 (intmax_t)adp->ad_offset - NDADDR,
10430 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR],
10431 (intmax_t)adp->ad_newblkno);
10432 deplist |= 1 << adp->ad_offset;
10433 if ((adp->ad_state & ATTACHED) == 0)
10434 panic("softdep_write_inodeblock: Unknown state 0x%x",
10436 #endif /* INVARIANTS */
10437 adp->ad_state &= ~ATTACHED;
10438 adp->ad_state |= UNDONE;
10441 * The on-disk inode cannot claim to be any larger than the last
10442 * fragment that has been written. Otherwise, the on-disk inode
10443 * might have fragments that were not the last block in the file
10444 * which would corrupt the filesystem.
10446 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10447 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10448 if (adp->ad_offset >= NDADDR)
10450 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10451 /* keep going until hitting a rollback to a frag */
10452 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10454 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10455 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10457 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10458 panic("softdep_write_inodeblock: lost dep2");
10459 #endif /* INVARIANTS */
10462 for (i = 0; i < NIADDR; i++) {
10464 if (dp->di_ib[i] != 0 &&
10465 (deplist & ((1 << NDADDR) << i)) == 0)
10466 panic("softdep_write_inodeblock: lost dep3");
10467 #endif /* INVARIANTS */
10473 * If we have zero'ed out the last allocated block of the file,
10474 * roll back the size to the last currently allocated block.
10475 * We know that this last allocated block is a full-sized as
10476 * we already checked for fragments in the loop above.
10478 if (lastadp != NULL &&
10479 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10480 for (i = lastadp->ad_offset; i >= 0; i--)
10481 if (dp->di_db[i] != 0)
10483 dp->di_size = (i + 1) * fs->fs_bsize;
10486 * The only dependencies are for indirect blocks.
10488 * The file size for indirect block additions is not guaranteed.
10489 * Such a guarantee would be non-trivial to achieve. The conventional
10490 * synchronous write implementation also does not make this guarantee.
10491 * Fsck should catch and fix discrepancies. Arguably, the file size
10492 * can be over-estimated without destroying integrity when the file
10493 * moves into the indirect blocks (i.e., is large). If we want to
10494 * postpone fsck, we are stuck with this argument.
10496 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10497 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10501 * Cancel an indirdep as a result of truncation. Release all of the
10502 * children allocindirs and place their journal work on the appropriate
10506 cancel_indirdep(indirdep, bp, freeblks)
10507 struct indirdep *indirdep;
10509 struct freeblks *freeblks;
10511 struct allocindir *aip;
10514 * None of the indirect pointers will ever be visible,
10515 * so they can simply be tossed. GOINGAWAY ensures
10516 * that allocated pointers will be saved in the buffer
10517 * cache until they are freed. Note that they will
10518 * only be able to be found by their physical address
10519 * since the inode mapping the logical address will
10520 * be gone. The save buffer used for the safe copy
10521 * was allocated in setup_allocindir_phase2 using
10522 * the physical address so it could be used for this
10523 * purpose. Hence we swap the safe copy with the real
10524 * copy, allowing the safe copy to be freed and holding
10525 * on to the real copy for later use in indir_trunc.
10527 if (indirdep->ir_state & GOINGAWAY)
10528 panic("cancel_indirdep: already gone");
10529 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
10530 indirdep->ir_state |= DEPCOMPLETE;
10531 LIST_REMOVE(indirdep, ir_next);
10533 indirdep->ir_state |= GOINGAWAY;
10535 * Pass in bp for blocks still have journal writes
10536 * pending so we can cancel them on their own.
10538 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL)
10539 cancel_allocindir(aip, bp, freeblks, 0);
10540 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL)
10541 cancel_allocindir(aip, NULL, freeblks, 0);
10542 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL)
10543 cancel_allocindir(aip, NULL, freeblks, 0);
10544 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL)
10545 cancel_allocindir(aip, NULL, freeblks, 0);
10547 * If there are pending partial truncations we need to keep the
10548 * old block copy around until they complete. This is because
10549 * the current b_data is not a perfect superset of the available
10552 if (TAILQ_EMPTY(&indirdep->ir_trunc))
10553 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
10555 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10556 WORKLIST_REMOVE(&indirdep->ir_list);
10557 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
10558 indirdep->ir_bp = NULL;
10559 indirdep->ir_freeblks = freeblks;
10563 * Free an indirdep once it no longer has new pointers to track.
10566 free_indirdep(indirdep)
10567 struct indirdep *indirdep;
10570 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc),
10571 ("free_indirdep: Indir trunc list not empty."));
10572 KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
10573 ("free_indirdep: Complete head not empty."));
10574 KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
10575 ("free_indirdep: write head not empty."));
10576 KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
10577 ("free_indirdep: done head not empty."));
10578 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
10579 ("free_indirdep: deplist head not empty."));
10580 KASSERT((indirdep->ir_state & DEPCOMPLETE),
10581 ("free_indirdep: %p still on newblk list.", indirdep));
10582 KASSERT(indirdep->ir_saveddata == NULL,
10583 ("free_indirdep: %p still has saved data.", indirdep));
10584 if (indirdep->ir_state & ONWORKLIST)
10585 WORKLIST_REMOVE(&indirdep->ir_list);
10586 WORKITEM_FREE(indirdep, D_INDIRDEP);
10590 * Called before a write to an indirdep. This routine is responsible for
10591 * rolling back pointers to a safe state which includes only those
10592 * allocindirs which have been completed.
10595 initiate_write_indirdep(indirdep, bp)
10596 struct indirdep *indirdep;
10599 struct ufsmount *ump;
10601 indirdep->ir_state |= IOSTARTED;
10602 if (indirdep->ir_state & GOINGAWAY)
10603 panic("disk_io_initiation: indirdep gone");
10605 * If there are no remaining dependencies, this will be writing
10606 * the real pointers.
10608 if (LIST_EMPTY(&indirdep->ir_deplisthd) &&
10609 TAILQ_EMPTY(&indirdep->ir_trunc))
10612 * Replace up-to-date version with safe version.
10614 if (indirdep->ir_saveddata == NULL) {
10615 ump = VFSTOUFS(indirdep->ir_list.wk_mp);
10618 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
10622 indirdep->ir_state &= ~ATTACHED;
10623 indirdep->ir_state |= UNDONE;
10624 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10625 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
10630 * Called when an inode has been cleared in a cg bitmap. This finally
10631 * eliminates any canceled jaddrefs
10634 softdep_setup_inofree(mp, bp, ino, wkhd)
10638 struct workhead *wkhd;
10640 struct worklist *wk, *wkn;
10641 struct inodedep *inodedep;
10642 struct ufsmount *ump;
10647 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
10648 ("softdep_setup_inofree called on non-softdep filesystem"));
10649 ump = VFSTOUFS(mp);
10652 cgp = (struct cg *)bp->b_data;
10653 inosused = cg_inosused(cgp);
10654 if (isset(inosused, ino % fs->fs_ipg))
10655 panic("softdep_setup_inofree: inode %ju not freed.",
10657 if (inodedep_lookup(mp, ino, 0, &inodedep))
10658 panic("softdep_setup_inofree: ino %ju has existing inodedep %p",
10659 (uintmax_t)ino, inodedep);
10661 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
10662 if (wk->wk_type != D_JADDREF)
10664 WORKLIST_REMOVE(wk);
10666 * We can free immediately even if the jaddref
10667 * isn't attached in a background write as now
10668 * the bitmaps are reconciled.
10670 wk->wk_state |= COMPLETE | ATTACHED;
10671 free_jaddref(WK_JADDREF(wk));
10673 jwork_move(&bp->b_dep, wkhd);
10680 * Called via ffs_blkfree() after a set of frags has been cleared from a cg
10681 * map. Any dependencies waiting for the write to clear are added to the
10682 * buf's list and any jnewblks that are being canceled are discarded
10686 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
10689 ufs2_daddr_t blkno;
10691 struct workhead *wkhd;
10693 struct bmsafemap *bmsafemap;
10694 struct jnewblk *jnewblk;
10695 struct ufsmount *ump;
10696 struct worklist *wk;
10701 ufs2_daddr_t jstart;
10709 "softdep_setup_blkfree: blkno %jd frags %d wk head %p",
10710 blkno, frags, wkhd);
10712 ump = VFSTOUFS(mp);
10713 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
10714 ("softdep_setup_blkfree called on non-softdep filesystem"));
10716 /* Lookup the bmsafemap so we track when it is dirty. */
10718 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10720 * Detach any jnewblks which have been canceled. They must linger
10721 * until the bitmap is cleared again by ffs_blkfree() to prevent
10722 * an unjournaled allocation from hitting the disk.
10725 while ((wk = LIST_FIRST(wkhd)) != NULL) {
10727 "softdep_setup_blkfree: blkno %jd wk type %d",
10728 blkno, wk->wk_type);
10729 WORKLIST_REMOVE(wk);
10730 if (wk->wk_type != D_JNEWBLK) {
10731 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk);
10734 jnewblk = WK_JNEWBLK(wk);
10735 KASSERT(jnewblk->jn_state & GOINGAWAY,
10736 ("softdep_setup_blkfree: jnewblk not canceled."));
10739 * Assert that this block is free in the bitmap
10740 * before we discard the jnewblk.
10742 cgp = (struct cg *)bp->b_data;
10743 blksfree = cg_blksfree(cgp);
10744 bno = dtogd(fs, jnewblk->jn_blkno);
10745 for (i = jnewblk->jn_oldfrags;
10746 i < jnewblk->jn_frags; i++) {
10747 if (isset(blksfree, bno + i))
10749 panic("softdep_setup_blkfree: not free");
10753 * Even if it's not attached we can free immediately
10754 * as the new bitmap is correct.
10756 wk->wk_state |= COMPLETE | ATTACHED;
10757 free_jnewblk(jnewblk);
10763 * Assert that we are not freeing a block which has an outstanding
10764 * allocation dependency.
10766 fs = VFSTOUFS(mp)->um_fs;
10767 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10768 end = blkno + frags;
10769 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10771 * Don't match against blocks that will be freed when the
10772 * background write is done.
10774 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
10775 (COMPLETE | DEPCOMPLETE))
10777 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
10778 jend = jnewblk->jn_blkno + jnewblk->jn_frags;
10779 if ((blkno >= jstart && blkno < jend) ||
10780 (end > jstart && end <= jend)) {
10781 printf("state 0x%X %jd - %d %d dep %p\n",
10782 jnewblk->jn_state, jnewblk->jn_blkno,
10783 jnewblk->jn_oldfrags, jnewblk->jn_frags,
10785 panic("softdep_setup_blkfree: "
10786 "%jd-%jd(%d) overlaps with %jd-%jd",
10787 blkno, end, frags, jstart, jend);
10795 * Revert a block allocation when the journal record that describes it
10796 * is not yet written.
10799 jnewblk_rollback(jnewblk, fs, cgp, blksfree)
10800 struct jnewblk *jnewblk;
10805 ufs1_daddr_t fragno;
10811 cgbno = dtogd(fs, jnewblk->jn_blkno);
10813 * We have to test which frags need to be rolled back. We may
10814 * be operating on a stale copy when doing background writes.
10816 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++)
10817 if (isclr(blksfree, cgbno + i))
10822 * This is mostly ffs_blkfree() sans some validation and
10823 * superblock updates.
10825 if (frags == fs->fs_frag) {
10826 fragno = fragstoblks(fs, cgbno);
10827 ffs_setblock(fs, blksfree, fragno);
10828 ffs_clusteracct(fs, cgp, fragno, 1);
10829 cgp->cg_cs.cs_nbfree++;
10831 cgbno += jnewblk->jn_oldfrags;
10832 bbase = cgbno - fragnum(fs, cgbno);
10833 /* Decrement the old frags. */
10834 blk = blkmap(fs, blksfree, bbase);
10835 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
10836 /* Deallocate the fragment */
10837 for (i = 0; i < frags; i++)
10838 setbit(blksfree, cgbno + i);
10839 cgp->cg_cs.cs_nffree += frags;
10840 /* Add back in counts associated with the new frags */
10841 blk = blkmap(fs, blksfree, bbase);
10842 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
10843 /* If a complete block has been reassembled, account for it. */
10844 fragno = fragstoblks(fs, bbase);
10845 if (ffs_isblock(fs, blksfree, fragno)) {
10846 cgp->cg_cs.cs_nffree -= fs->fs_frag;
10847 ffs_clusteracct(fs, cgp, fragno, 1);
10848 cgp->cg_cs.cs_nbfree++;
10852 jnewblk->jn_state &= ~ATTACHED;
10853 jnewblk->jn_state |= UNDONE;
10859 initiate_write_bmsafemap(bmsafemap, bp)
10860 struct bmsafemap *bmsafemap;
10861 struct buf *bp; /* The cg block. */
10863 struct jaddref *jaddref;
10864 struct jnewblk *jnewblk;
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 unrolled.
10946 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
10948 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
10950 ump = VFSTOUFS(wk->wk_mp);
10951 LIST_INIT(&reattach);
10953 * This lock must not be released anywhere in this code segment.
10958 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
10959 WORKLIST_REMOVE(wk);
10960 atomic_add_long(&dep_write[wk->wk_type], 1);
10962 panic("duplicate worklist: %p\n", wk);
10964 switch (wk->wk_type) {
10967 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
10968 WORKLIST_INSERT(&reattach, wk);
10972 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
10973 WORKLIST_INSERT(&reattach, wk);
10977 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp))
10978 WORKLIST_INSERT(&reattach, wk);
10982 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
10985 case D_ALLOCDIRECT:
10986 wk->wk_state |= COMPLETE;
10987 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
10991 wk->wk_state |= COMPLETE;
10992 handle_allocindir_partdone(WK_ALLOCINDIR(wk));
10996 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp))
10997 WORKLIST_INSERT(&reattach, wk);
11001 wk->wk_state |= COMPLETE;
11002 freeblks = WK_FREEBLKS(wk);
11003 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE &&
11004 LIST_EMPTY(&freeblks->fb_jblkdephd))
11005 add_to_worklist(wk, WK_NODELAY);
11009 handle_written_freework(WK_FREEWORK(wk));
11013 free_jsegdep(WK_JSEGDEP(wk));
11017 handle_written_jseg(WK_JSEG(wk), bp);
11021 if (handle_written_sbdep(WK_SBDEP(wk), bp))
11022 WORKLIST_INSERT(&reattach, wk);
11026 free_freedep(WK_FREEDEP(wk));
11030 panic("handle_disk_write_complete: Unknown type %s",
11031 TYPENAME(wk->wk_type));
11036 * Reattach any requests that must be redone.
11038 while ((wk = LIST_FIRST(&reattach)) != NULL) {
11039 WORKLIST_REMOVE(wk);
11040 WORKLIST_INSERT(&bp->b_dep, wk);
11048 * Called from within softdep_disk_write_complete above. Note that
11049 * this routine is always called from interrupt level with further
11050 * splbio interrupts blocked.
11053 handle_allocdirect_partdone(adp, wkhd)
11054 struct allocdirect *adp; /* the completed allocdirect */
11055 struct workhead *wkhd; /* Work to do when inode is writtne. */
11057 struct allocdirectlst *listhead;
11058 struct allocdirect *listadp;
11059 struct inodedep *inodedep;
11062 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11065 * The on-disk inode cannot claim to be any larger than the last
11066 * fragment that has been written. Otherwise, the on-disk inode
11067 * might have fragments that were not the last block in the file
11068 * which would corrupt the filesystem. Thus, we cannot free any
11069 * allocdirects after one whose ad_oldblkno claims a fragment as
11070 * these blocks must be rolled back to zero before writing the inode.
11071 * We check the currently active set of allocdirects in id_inoupdt
11072 * or id_extupdt as appropriate.
11074 inodedep = adp->ad_inodedep;
11075 bsize = inodedep->id_fs->fs_bsize;
11076 if (adp->ad_state & EXTDATA)
11077 listhead = &inodedep->id_extupdt;
11079 listhead = &inodedep->id_inoupdt;
11080 TAILQ_FOREACH(listadp, listhead, ad_next) {
11081 /* found our block */
11082 if (listadp == adp)
11084 /* continue if ad_oldlbn is not a fragment */
11085 if (listadp->ad_oldsize == 0 ||
11086 listadp->ad_oldsize == bsize)
11088 /* hit a fragment */
11092 * If we have reached the end of the current list without
11093 * finding the just finished dependency, then it must be
11094 * on the future dependency list. Future dependencies cannot
11095 * be freed until they are moved to the current list.
11097 if (listadp == NULL) {
11099 if (adp->ad_state & EXTDATA)
11100 listhead = &inodedep->id_newextupdt;
11102 listhead = &inodedep->id_newinoupdt;
11103 TAILQ_FOREACH(listadp, listhead, ad_next)
11104 /* found our block */
11105 if (listadp == adp)
11107 if (listadp == NULL)
11108 panic("handle_allocdirect_partdone: lost dep");
11113 * If we have found the just finished dependency, then queue
11114 * it along with anything that follows it that is complete.
11115 * Since the pointer has not yet been written in the inode
11116 * as the dependency prevents it, place the allocdirect on the
11117 * bufwait list where it will be freed once the pointer is
11121 wkhd = &inodedep->id_bufwait;
11122 for (; adp; adp = listadp) {
11123 listadp = TAILQ_NEXT(adp, ad_next);
11124 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11126 TAILQ_REMOVE(listhead, adp, ad_next);
11127 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
11132 * Called from within softdep_disk_write_complete above. This routine
11133 * completes successfully written allocindirs.
11136 handle_allocindir_partdone(aip)
11137 struct allocindir *aip; /* the completed allocindir */
11139 struct indirdep *indirdep;
11141 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
11143 indirdep = aip->ai_indirdep;
11144 LIST_REMOVE(aip, ai_next);
11146 * Don't set a pointer while the buffer is undergoing IO or while
11147 * we have active truncations.
11149 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) {
11150 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
11153 if (indirdep->ir_state & UFS1FMT)
11154 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11157 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11160 * Await the pointer write before freeing the allocindir.
11162 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
11166 * Release segments held on a jwork list.
11170 struct workhead *wkhd;
11172 struct worklist *wk;
11174 while ((wk = LIST_FIRST(wkhd)) != NULL) {
11175 WORKLIST_REMOVE(wk);
11176 switch (wk->wk_type) {
11178 free_jsegdep(WK_JSEGDEP(wk));
11181 free_freedep(WK_FREEDEP(wk));
11184 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep));
11185 WORKITEM_FREE(wk, D_FREEFRAG);
11188 handle_written_freework(WK_FREEWORK(wk));
11191 panic("handle_jwork: Unknown type %s\n",
11192 TYPENAME(wk->wk_type));
11198 * Handle the bufwait list on an inode when it is safe to release items
11199 * held there. This normally happens after an inode block is written but
11200 * may be delayed and handled later if there are pending journal items that
11201 * are not yet safe to be released.
11203 static struct freefile *
11204 handle_bufwait(inodedep, refhd)
11205 struct inodedep *inodedep;
11206 struct workhead *refhd;
11208 struct jaddref *jaddref;
11209 struct freefile *freefile;
11210 struct worklist *wk;
11213 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
11214 WORKLIST_REMOVE(wk);
11215 switch (wk->wk_type) {
11218 * We defer adding freefile to the worklist
11219 * until all other additions have been made to
11220 * ensure that it will be done after all the
11221 * old blocks have been freed.
11223 if (freefile != NULL)
11224 panic("handle_bufwait: freefile");
11225 freefile = WK_FREEFILE(wk);
11229 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
11233 diradd_inode_written(WK_DIRADD(wk), inodedep);
11237 wk->wk_state |= COMPLETE;
11238 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
11239 add_to_worklist(wk, 0);
11243 wk->wk_state |= COMPLETE;
11244 add_to_worklist(wk, 0);
11247 case D_ALLOCDIRECT:
11249 free_newblk(WK_NEWBLK(wk));
11253 wk->wk_state |= COMPLETE;
11254 free_jnewblk(WK_JNEWBLK(wk));
11258 * Save freed journal segments and add references on
11259 * the supplied list which will delay their release
11260 * until the cg bitmap is cleared on disk.
11264 free_jsegdep(WK_JSEGDEP(wk));
11266 WORKLIST_INSERT(refhd, wk);
11270 jaddref = WK_JADDREF(wk);
11271 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
11274 * Transfer any jaddrefs to the list to be freed with
11275 * the bitmap if we're handling a removed file.
11277 if (refhd == NULL) {
11278 wk->wk_state |= COMPLETE;
11279 free_jaddref(jaddref);
11281 WORKLIST_INSERT(refhd, wk);
11285 panic("handle_bufwait: Unknown type %p(%s)",
11286 wk, TYPENAME(wk->wk_type));
11293 * Called from within softdep_disk_write_complete above to restore
11294 * in-memory inode block contents to their most up-to-date state. Note
11295 * that this routine is always called from interrupt level with further
11296 * splbio interrupts blocked.
11299 handle_written_inodeblock(inodedep, bp)
11300 struct inodedep *inodedep;
11301 struct buf *bp; /* buffer containing the inode block */
11303 struct freefile *freefile;
11304 struct allocdirect *adp, *nextadp;
11305 struct ufs1_dinode *dp1 = NULL;
11306 struct ufs2_dinode *dp2 = NULL;
11307 struct workhead wkhd;
11308 int hadchanges, fstype;
11314 if ((inodedep->id_state & IOSTARTED) == 0)
11315 panic("handle_written_inodeblock: not started");
11316 inodedep->id_state &= ~IOSTARTED;
11317 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
11319 dp1 = (struct ufs1_dinode *)bp->b_data +
11320 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11321 freelink = dp1->di_freelink;
11324 dp2 = (struct ufs2_dinode *)bp->b_data +
11325 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11326 freelink = dp2->di_freelink;
11329 * Leave this inodeblock dirty until it's in the list.
11331 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) {
11332 struct inodedep *inon;
11334 inon = TAILQ_NEXT(inodedep, id_unlinked);
11335 if ((inon == NULL && freelink == 0) ||
11336 (inon && inon->id_ino == freelink)) {
11338 inon->id_state |= UNLINKPREV;
11339 inodedep->id_state |= UNLINKNEXT;
11344 * If we had to rollback the inode allocation because of
11345 * bitmaps being incomplete, then simply restore it.
11346 * Keep the block dirty so that it will not be reclaimed until
11347 * all associated dependencies have been cleared and the
11348 * corresponding updates written to disk.
11350 if (inodedep->id_savedino1 != NULL) {
11352 if (fstype == UFS1)
11353 *dp1 = *inodedep->id_savedino1;
11355 *dp2 = *inodedep->id_savedino2;
11356 free(inodedep->id_savedino1, M_SAVEDINO);
11357 inodedep->id_savedino1 = NULL;
11358 if ((bp->b_flags & B_DELWRI) == 0)
11359 stat_inode_bitmap++;
11362 * If the inode is clear here and GOINGAWAY it will never
11363 * be written. Process the bufwait and clear any pending
11364 * work which may include the freefile.
11366 if (inodedep->id_state & GOINGAWAY)
11370 inodedep->id_state |= COMPLETE;
11372 * Roll forward anything that had to be rolled back before
11373 * the inode could be updated.
11375 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
11376 nextadp = TAILQ_NEXT(adp, ad_next);
11377 if (adp->ad_state & ATTACHED)
11378 panic("handle_written_inodeblock: new entry");
11379 if (fstype == UFS1) {
11380 if (adp->ad_offset < NDADDR) {
11381 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11382 panic("%s %s #%jd mismatch %d != %jd",
11383 "handle_written_inodeblock:",
11385 (intmax_t)adp->ad_offset,
11386 dp1->di_db[adp->ad_offset],
11387 (intmax_t)adp->ad_oldblkno);
11388 dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
11390 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0)
11391 panic("%s: %s #%jd allocated as %d",
11392 "handle_written_inodeblock",
11393 "indirect pointer",
11394 (intmax_t)adp->ad_offset - NDADDR,
11395 dp1->di_ib[adp->ad_offset - NDADDR]);
11396 dp1->di_ib[adp->ad_offset - NDADDR] =
11400 if (adp->ad_offset < NDADDR) {
11401 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11402 panic("%s: %s #%jd %s %jd != %jd",
11403 "handle_written_inodeblock",
11405 (intmax_t)adp->ad_offset, "mismatch",
11406 (intmax_t)dp2->di_db[adp->ad_offset],
11407 (intmax_t)adp->ad_oldblkno);
11408 dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
11410 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0)
11411 panic("%s: %s #%jd allocated as %jd",
11412 "handle_written_inodeblock",
11413 "indirect pointer",
11414 (intmax_t)adp->ad_offset - NDADDR,
11416 dp2->di_ib[adp->ad_offset - NDADDR]);
11417 dp2->di_ib[adp->ad_offset - NDADDR] =
11421 adp->ad_state &= ~UNDONE;
11422 adp->ad_state |= ATTACHED;
11425 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
11426 nextadp = TAILQ_NEXT(adp, ad_next);
11427 if (adp->ad_state & ATTACHED)
11428 panic("handle_written_inodeblock: new entry");
11429 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
11430 panic("%s: direct pointers #%jd %s %jd != %jd",
11431 "handle_written_inodeblock",
11432 (intmax_t)adp->ad_offset, "mismatch",
11433 (intmax_t)dp2->di_extb[adp->ad_offset],
11434 (intmax_t)adp->ad_oldblkno);
11435 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
11436 adp->ad_state &= ~UNDONE;
11437 adp->ad_state |= ATTACHED;
11440 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
11441 stat_direct_blk_ptrs++;
11443 * Reset the file size to its most up-to-date value.
11445 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
11446 panic("handle_written_inodeblock: bad size");
11447 if (inodedep->id_savednlink > LINK_MAX)
11448 panic("handle_written_inodeblock: Invalid link count "
11449 "%d for inodedep %p", inodedep->id_savednlink, inodedep);
11450 if (fstype == UFS1) {
11451 if (dp1->di_nlink != inodedep->id_savednlink) {
11452 dp1->di_nlink = inodedep->id_savednlink;
11455 if (dp1->di_size != inodedep->id_savedsize) {
11456 dp1->di_size = inodedep->id_savedsize;
11460 if (dp2->di_nlink != inodedep->id_savednlink) {
11461 dp2->di_nlink = inodedep->id_savednlink;
11464 if (dp2->di_size != inodedep->id_savedsize) {
11465 dp2->di_size = inodedep->id_savedsize;
11468 if (dp2->di_extsize != inodedep->id_savedextsize) {
11469 dp2->di_extsize = inodedep->id_savedextsize;
11473 inodedep->id_savedsize = -1;
11474 inodedep->id_savedextsize = -1;
11475 inodedep->id_savednlink = -1;
11477 * If there were any rollbacks in the inode block, then it must be
11478 * marked dirty so that its will eventually get written back in
11479 * its correct form.
11485 * Process any allocdirects that completed during the update.
11487 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
11488 handle_allocdirect_partdone(adp, &wkhd);
11489 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
11490 handle_allocdirect_partdone(adp, &wkhd);
11492 * Process deallocations that were held pending until the
11493 * inode had been written to disk. Freeing of the inode
11494 * is delayed until after all blocks have been freed to
11495 * avoid creation of new <vfsid, inum, lbn> triples
11496 * before the old ones have been deleted. Completely
11497 * unlinked inodes are not processed until the unlinked
11498 * inode list is written or the last reference is removed.
11500 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
11501 freefile = handle_bufwait(inodedep, NULL);
11502 if (freefile && !LIST_EMPTY(&wkhd)) {
11503 WORKLIST_INSERT(&wkhd, &freefile->fx_list);
11508 * Move rolled forward dependency completions to the bufwait list
11509 * now that those that were already written have been processed.
11511 if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
11512 panic("handle_written_inodeblock: bufwait but no changes");
11513 jwork_move(&inodedep->id_bufwait, &wkhd);
11515 if (freefile != NULL) {
11517 * If the inode is goingaway it was never written. Fake up
11518 * the state here so free_inodedep() can succeed.
11520 if (inodedep->id_state & GOINGAWAY)
11521 inodedep->id_state |= COMPLETE | DEPCOMPLETE;
11522 if (free_inodedep(inodedep) == 0)
11523 panic("handle_written_inodeblock: live inodedep %p",
11525 add_to_worklist(&freefile->fx_list, 0);
11530 * If no outstanding dependencies, free it.
11532 if (free_inodedep(inodedep) ||
11533 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
11534 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
11535 TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
11536 LIST_FIRST(&inodedep->id_bufwait) == 0))
11538 return (hadchanges);
11542 handle_written_indirdep(indirdep, bp, bpp)
11543 struct indirdep *indirdep;
11547 struct allocindir *aip;
11551 if (indirdep->ir_state & GOINGAWAY)
11552 panic("handle_written_indirdep: indirdep gone");
11553 if ((indirdep->ir_state & IOSTARTED) == 0)
11554 panic("handle_written_indirdep: IO not started");
11557 * If there were rollbacks revert them here.
11559 if (indirdep->ir_saveddata) {
11560 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
11561 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11562 free(indirdep->ir_saveddata, M_INDIRDEP);
11563 indirdep->ir_saveddata = NULL;
11567 indirdep->ir_state &= ~(UNDONE | IOSTARTED);
11568 indirdep->ir_state |= ATTACHED;
11570 * Move allocindirs with written pointers to the completehd if
11571 * the indirdep's pointer is not yet written. Otherwise
11574 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) {
11575 LIST_REMOVE(aip, ai_next);
11576 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
11577 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
11579 newblk_freefrag(&aip->ai_block);
11582 free_newblk(&aip->ai_block);
11585 * Move allocindirs that have finished dependency processing from
11586 * the done list to the write list after updating the pointers.
11588 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11589 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) {
11590 handle_allocindir_partdone(aip);
11591 if (aip == LIST_FIRST(&indirdep->ir_donehd))
11592 panic("disk_write_complete: not gone");
11597 * Preserve the indirdep if there were any changes or if it is not
11598 * yet valid on disk.
11601 stat_indir_blk_ptrs++;
11606 * If there were no changes we can discard the savedbp and detach
11607 * ourselves from the buf. We are only carrying completed pointers
11610 sbp = indirdep->ir_savebp;
11611 sbp->b_flags |= B_INVAL | B_NOCACHE;
11612 indirdep->ir_savebp = NULL;
11613 indirdep->ir_bp = NULL;
11615 panic("handle_written_indirdep: bp already exists.");
11618 * The indirdep may not be freed until its parent points at it.
11620 if (indirdep->ir_state & DEPCOMPLETE)
11621 free_indirdep(indirdep);
11627 * Process a diradd entry after its dependent inode has been written.
11628 * This routine must be called with splbio interrupts blocked.
11631 diradd_inode_written(dap, inodedep)
11632 struct diradd *dap;
11633 struct inodedep *inodedep;
11636 dap->da_state |= COMPLETE;
11637 complete_diradd(dap);
11638 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
11642 * Returns true if the bmsafemap will have rollbacks when written. Must only
11643 * be called with the per-filesystem lock and the buf lock on the cg held.
11646 bmsafemap_backgroundwrite(bmsafemap, bp)
11647 struct bmsafemap *bmsafemap;
11652 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp));
11653 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) |
11654 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd);
11656 * If we're initiating a background write we need to process the
11657 * rollbacks as they exist now, not as they exist when IO starts.
11658 * No other consumers will look at the contents of the shadowed
11659 * buf so this is safe to do here.
11661 if (bp->b_xflags & BX_BKGRDMARKER)
11662 initiate_write_bmsafemap(bmsafemap, bp);
11668 * Re-apply an allocation when a cg write is complete.
11671 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)
11672 struct jnewblk *jnewblk;
11677 ufs1_daddr_t fragno;
11678 ufs2_daddr_t blkno;
11684 cgbno = dtogd(fs, jnewblk->jn_blkno);
11685 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) {
11686 if (isclr(blksfree, cgbno + i))
11687 panic("jnewblk_rollforward: re-allocated fragment");
11690 if (frags == fs->fs_frag) {
11691 blkno = fragstoblks(fs, cgbno);
11692 ffs_clrblock(fs, blksfree, (long)blkno);
11693 ffs_clusteracct(fs, cgp, blkno, -1);
11694 cgp->cg_cs.cs_nbfree--;
11696 bbase = cgbno - fragnum(fs, cgbno);
11697 cgbno += jnewblk->jn_oldfrags;
11698 /* If a complete block had been reassembled, account for it. */
11699 fragno = fragstoblks(fs, bbase);
11700 if (ffs_isblock(fs, blksfree, fragno)) {
11701 cgp->cg_cs.cs_nffree += fs->fs_frag;
11702 ffs_clusteracct(fs, cgp, fragno, -1);
11703 cgp->cg_cs.cs_nbfree--;
11705 /* Decrement the old frags. */
11706 blk = blkmap(fs, blksfree, bbase);
11707 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
11708 /* Allocate the fragment */
11709 for (i = 0; i < frags; i++)
11710 clrbit(blksfree, cgbno + i);
11711 cgp->cg_cs.cs_nffree -= frags;
11712 /* Add back in counts associated with the new frags */
11713 blk = blkmap(fs, blksfree, bbase);
11714 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
11720 * Complete a write to a bmsafemap structure. Roll forward any bitmap
11721 * changes if it's not a background write. Set all written dependencies
11722 * to DEPCOMPLETE and free the structure if possible.
11725 handle_written_bmsafemap(bmsafemap, bp)
11726 struct bmsafemap *bmsafemap;
11729 struct newblk *newblk;
11730 struct inodedep *inodedep;
11731 struct jaddref *jaddref, *jatmp;
11732 struct jnewblk *jnewblk, *jntmp;
11733 struct ufsmount *ump;
11742 if ((bmsafemap->sm_state & IOSTARTED) == 0)
11743 panic("initiate_write_bmsafemap: Not started\n");
11744 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp);
11746 bmsafemap->sm_state &= ~IOSTARTED;
11747 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0;
11749 * Release journal work that was waiting on the write.
11751 handle_jwork(&bmsafemap->sm_freewr);
11754 * Restore unwritten inode allocation pending jaddref writes.
11756 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
11757 cgp = (struct cg *)bp->b_data;
11758 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11759 inosused = cg_inosused(cgp);
11760 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
11761 ja_bmdeps, jatmp) {
11762 if ((jaddref->ja_state & UNDONE) == 0)
11764 ino = jaddref->ja_ino % fs->fs_ipg;
11765 if (isset(inosused, ino))
11766 panic("handle_written_bmsafemap: "
11767 "re-allocated inode");
11768 /* Do the roll-forward only if it's a real copy. */
11770 if ((jaddref->ja_mode & IFMT) == IFDIR)
11771 cgp->cg_cs.cs_ndir++;
11772 cgp->cg_cs.cs_nifree--;
11773 setbit(inosused, ino);
11776 jaddref->ja_state &= ~UNDONE;
11777 jaddref->ja_state |= ATTACHED;
11778 free_jaddref(jaddref);
11782 * Restore any block allocations which are pending journal writes.
11784 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
11785 cgp = (struct cg *)bp->b_data;
11786 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11787 blksfree = cg_blksfree(cgp);
11788 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
11790 if ((jnewblk->jn_state & UNDONE) == 0)
11792 /* Do the roll-forward only if it's a real copy. */
11794 jnewblk_rollforward(jnewblk, fs, cgp, blksfree))
11796 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
11797 jnewblk->jn_state |= ATTACHED;
11798 free_jnewblk(jnewblk);
11801 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
11802 newblk->nb_state |= DEPCOMPLETE;
11803 newblk->nb_state &= ~ONDEPLIST;
11804 newblk->nb_bmsafemap = NULL;
11805 LIST_REMOVE(newblk, nb_deps);
11806 if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
11807 handle_allocdirect_partdone(
11808 WK_ALLOCDIRECT(&newblk->nb_list), NULL);
11809 else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
11810 handle_allocindir_partdone(
11811 WK_ALLOCINDIR(&newblk->nb_list));
11812 else if (newblk->nb_list.wk_type != D_NEWBLK)
11813 panic("handle_written_bmsafemap: Unexpected type: %s",
11814 TYPENAME(newblk->nb_list.wk_type));
11816 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
11817 inodedep->id_state |= DEPCOMPLETE;
11818 inodedep->id_state &= ~ONDEPLIST;
11819 LIST_REMOVE(inodedep, id_deps);
11820 inodedep->id_bmsafemap = NULL;
11822 LIST_REMOVE(bmsafemap, sm_next);
11823 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
11824 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
11825 LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
11826 LIST_EMPTY(&bmsafemap->sm_inodedephd) &&
11827 LIST_EMPTY(&bmsafemap->sm_freehd)) {
11828 LIST_REMOVE(bmsafemap, sm_hash);
11829 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
11832 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
11839 * Try to free a mkdir dependency.
11842 complete_mkdir(mkdir)
11843 struct mkdir *mkdir;
11845 struct diradd *dap;
11847 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
11849 LIST_REMOVE(mkdir, md_mkdirs);
11850 dap = mkdir->md_diradd;
11851 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
11852 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
11853 dap->da_state |= DEPCOMPLETE;
11854 complete_diradd(dap);
11856 WORKITEM_FREE(mkdir, D_MKDIR);
11860 * Handle the completion of a mkdir dependency.
11863 handle_written_mkdir(mkdir, type)
11864 struct mkdir *mkdir;
11868 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
11869 panic("handle_written_mkdir: bad type");
11870 mkdir->md_state |= COMPLETE;
11871 complete_mkdir(mkdir);
11875 free_pagedep(pagedep)
11876 struct pagedep *pagedep;
11880 if (pagedep->pd_state & NEWBLOCK)
11882 if (!LIST_EMPTY(&pagedep->pd_dirremhd))
11884 for (i = 0; i < DAHASHSZ; i++)
11885 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
11887 if (!LIST_EMPTY(&pagedep->pd_pendinghd))
11889 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
11891 if (pagedep->pd_state & ONWORKLIST)
11892 WORKLIST_REMOVE(&pagedep->pd_list);
11893 LIST_REMOVE(pagedep, pd_hash);
11894 WORKITEM_FREE(pagedep, D_PAGEDEP);
11900 * Called from within softdep_disk_write_complete above.
11901 * A write operation was just completed. Removed inodes can
11902 * now be freed and associated block pointers may be committed.
11903 * Note that this routine is always called from interrupt level
11904 * with further splbio interrupts blocked.
11907 handle_written_filepage(pagedep, bp)
11908 struct pagedep *pagedep;
11909 struct buf *bp; /* buffer containing the written page */
11911 struct dirrem *dirrem;
11912 struct diradd *dap, *nextdap;
11916 if ((pagedep->pd_state & IOSTARTED) == 0)
11917 panic("handle_written_filepage: not started");
11918 pagedep->pd_state &= ~IOSTARTED;
11920 * Process any directory removals that have been committed.
11922 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
11923 LIST_REMOVE(dirrem, dm_next);
11924 dirrem->dm_state |= COMPLETE;
11925 dirrem->dm_dirinum = pagedep->pd_ino;
11926 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
11927 ("handle_written_filepage: Journal entries not written."));
11928 add_to_worklist(&dirrem->dm_list, 0);
11931 * Free any directory additions that have been committed.
11932 * If it is a newly allocated block, we have to wait until
11933 * the on-disk directory inode claims the new block.
11935 if ((pagedep->pd_state & NEWBLOCK) == 0)
11936 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
11937 free_diradd(dap, NULL);
11939 * Uncommitted directory entries must be restored.
11941 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
11942 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
11944 nextdap = LIST_NEXT(dap, da_pdlist);
11945 if (dap->da_state & ATTACHED)
11946 panic("handle_written_filepage: attached");
11947 ep = (struct direct *)
11948 ((char *)bp->b_data + dap->da_offset);
11949 ep->d_ino = dap->da_newinum;
11950 dap->da_state &= ~UNDONE;
11951 dap->da_state |= ATTACHED;
11954 * If the inode referenced by the directory has
11955 * been written out, then the dependency can be
11956 * moved to the pending list.
11958 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
11959 LIST_REMOVE(dap, da_pdlist);
11960 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
11966 * If there were any rollbacks in the directory, then it must be
11967 * marked dirty so that its will eventually get written back in
11968 * its correct form.
11971 if ((bp->b_flags & B_DELWRI) == 0)
11977 * If we are not waiting for a new directory block to be
11978 * claimed by its inode, then the pagedep will be freed.
11979 * Otherwise it will remain to track any new entries on
11980 * the page in case they are fsync'ed.
11982 free_pagedep(pagedep);
11987 * Writing back in-core inode structures.
11989 * The filesystem only accesses an inode's contents when it occupies an
11990 * "in-core" inode structure. These "in-core" structures are separate from
11991 * the page frames used to cache inode blocks. Only the latter are
11992 * transferred to/from the disk. So, when the updated contents of the
11993 * "in-core" inode structure are copied to the corresponding in-memory inode
11994 * block, the dependencies are also transferred. The following procedure is
11995 * called when copying a dirty "in-core" inode to a cached inode block.
11999 * Called when an inode is loaded from disk. If the effective link count
12000 * differed from the actual link count when it was last flushed, then we
12001 * need to ensure that the correct effective link count is put back.
12004 softdep_load_inodeblock(ip)
12005 struct inode *ip; /* the "in_core" copy of the inode */
12007 struct inodedep *inodedep;
12009 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12010 ("softdep_load_inodeblock called on non-softdep filesystem"));
12012 * Check for alternate nlink count.
12014 ip->i_effnlink = ip->i_nlink;
12015 ACQUIRE_LOCK(ip->i_ump);
12016 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
12018 FREE_LOCK(ip->i_ump);
12021 ip->i_effnlink -= inodedep->id_nlinkdelta;
12022 FREE_LOCK(ip->i_ump);
12026 * This routine is called just before the "in-core" inode
12027 * information is to be copied to the in-memory inode block.
12028 * Recall that an inode block contains several inodes. If
12029 * the force flag is set, then the dependencies will be
12030 * cleared so that the update can always be made. Note that
12031 * the buffer is locked when this routine is called, so we
12032 * will never be in the middle of writing the inode block
12036 softdep_update_inodeblock(ip, bp, waitfor)
12037 struct inode *ip; /* the "in_core" copy of the inode */
12038 struct buf *bp; /* the buffer containing the inode block */
12039 int waitfor; /* nonzero => update must be allowed */
12041 struct inodedep *inodedep;
12042 struct inoref *inoref;
12043 struct ufsmount *ump;
12044 struct worklist *wk;
12051 mp = UFSTOVFS(ump);
12052 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
12053 ("softdep_update_inodeblock called on non-softdep filesystem"));
12056 * Preserve the freelink that is on disk. clear_unlinked_inodedep()
12057 * does not have access to the in-core ip so must write directly into
12058 * the inode block buffer when setting freelink.
12060 if (fs->fs_magic == FS_UFS1_MAGIC)
12061 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
12062 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12064 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
12065 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12067 * If the effective link count is not equal to the actual link
12068 * count, then we must track the difference in an inodedep while
12069 * the inode is (potentially) tossed out of the cache. Otherwise,
12070 * if there is no existing inodedep, then there are no dependencies
12075 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12077 if (ip->i_effnlink != ip->i_nlink)
12078 panic("softdep_update_inodeblock: bad link count");
12081 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
12082 panic("softdep_update_inodeblock: bad delta");
12084 * If we're flushing all dependencies we must also move any waiting
12085 * for journal writes onto the bufwait list prior to I/O.
12088 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12089 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12091 jwait(&inoref->if_list, MNT_WAIT);
12097 * Changes have been initiated. Anything depending on these
12098 * changes cannot occur until this inode has been written.
12100 inodedep->id_state &= ~COMPLETE;
12101 if ((inodedep->id_state & ONWORKLIST) == 0)
12102 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
12104 * Any new dependencies associated with the incore inode must
12105 * now be moved to the list associated with the buffer holding
12106 * the in-memory copy of the inode. Once merged process any
12107 * allocdirects that are completed by the merger.
12109 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
12110 if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
12111 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
12113 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
12114 if (!TAILQ_EMPTY(&inodedep->id_extupdt))
12115 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
12118 * Now that the inode has been pushed into the buffer, the
12119 * operations dependent on the inode being written to disk
12120 * can be moved to the id_bufwait so that they will be
12121 * processed when the buffer I/O completes.
12123 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
12124 WORKLIST_REMOVE(wk);
12125 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
12128 * Newly allocated inodes cannot be written until the bitmap
12129 * that allocates them have been written (indicated by
12130 * DEPCOMPLETE being set in id_state). If we are doing a
12131 * forced sync (e.g., an fsync on a file), we force the bitmap
12132 * to be written so that the update can be done.
12134 if (waitfor == 0) {
12139 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
12143 ibp = inodedep->id_bmsafemap->sm_buf;
12144 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT);
12147 * If ibp came back as NULL, the dependency could have been
12148 * freed while we slept. Look it up again, and check to see
12149 * that it has completed.
12151 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
12157 if ((error = bwrite(ibp)) != 0)
12158 softdep_error("softdep_update_inodeblock: bwrite", error);
12162 * Merge the a new inode dependency list (such as id_newinoupdt) into an
12163 * old inode dependency list (such as id_inoupdt). This routine must be
12164 * called with splbio interrupts blocked.
12167 merge_inode_lists(newlisthead, oldlisthead)
12168 struct allocdirectlst *newlisthead;
12169 struct allocdirectlst *oldlisthead;
12171 struct allocdirect *listadp, *newadp;
12173 newadp = TAILQ_FIRST(newlisthead);
12174 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
12175 if (listadp->ad_offset < newadp->ad_offset) {
12176 listadp = TAILQ_NEXT(listadp, ad_next);
12179 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12180 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
12181 if (listadp->ad_offset == newadp->ad_offset) {
12182 allocdirect_merge(oldlisthead, newadp,
12186 newadp = TAILQ_FIRST(newlisthead);
12188 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
12189 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12190 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
12195 * If we are doing an fsync, then we must ensure that any directory
12196 * entries for the inode have been written after the inode gets to disk.
12200 struct vnode *vp; /* the "in_core" copy of the inode */
12202 struct inodedep *inodedep;
12203 struct pagedep *pagedep;
12204 struct inoref *inoref;
12205 struct ufsmount *ump;
12206 struct worklist *wk;
12207 struct diradd *dap;
12213 struct thread *td = curthread;
12214 int error, flushparent, pagedep_new_block;
12222 if (MOUNTEDSOFTDEP(mp) == 0)
12226 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12230 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12231 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12233 jwait(&inoref->if_list, MNT_WAIT);
12237 if (!LIST_EMPTY(&inodedep->id_inowait) ||
12238 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
12239 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
12240 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
12241 !TAILQ_EMPTY(&inodedep->id_newinoupdt))
12242 panic("softdep_fsync: pending ops %p", inodedep);
12243 for (error = 0, flushparent = 0; ; ) {
12244 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
12246 if (wk->wk_type != D_DIRADD)
12247 panic("softdep_fsync: Unexpected type %s",
12248 TYPENAME(wk->wk_type));
12249 dap = WK_DIRADD(wk);
12251 * Flush our parent if this directory entry has a MKDIR_PARENT
12252 * dependency or is contained in a newly allocated block.
12254 if (dap->da_state & DIRCHG)
12255 pagedep = dap->da_previous->dm_pagedep;
12257 pagedep = dap->da_pagedep;
12258 parentino = pagedep->pd_ino;
12259 lbn = pagedep->pd_lbn;
12260 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
12261 panic("softdep_fsync: dirty");
12262 if ((dap->da_state & MKDIR_PARENT) ||
12263 (pagedep->pd_state & NEWBLOCK))
12268 * If we are being fsync'ed as part of vgone'ing this vnode,
12269 * then we will not be able to release and recover the
12270 * vnode below, so we just have to give up on writing its
12271 * directory entry out. It will eventually be written, just
12272 * not now, but then the user was not asking to have it
12273 * written, so we are not breaking any promises.
12275 if (vp->v_iflag & VI_DOOMED)
12278 * We prevent deadlock by always fetching inodes from the
12279 * root, moving down the directory tree. Thus, when fetching
12280 * our parent directory, we first try to get the lock. If
12281 * that fails, we must unlock ourselves before requesting
12282 * the lock on our parent. See the comment in ufs_lookup
12283 * for details on possible races.
12286 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
12287 FFSV_FORCEINSMQ)) {
12288 error = vfs_busy(mp, MBF_NOWAIT);
12292 error = vfs_busy(mp, 0);
12293 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12297 if (vp->v_iflag & VI_DOOMED) {
12303 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
12304 &pvp, FFSV_FORCEINSMQ);
12306 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12307 if (vp->v_iflag & VI_DOOMED) {
12316 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
12317 * that are contained in direct blocks will be resolved by
12318 * doing a ffs_update. Pagedeps contained in indirect blocks
12319 * may require a complete sync'ing of the directory. So, we
12320 * try the cheap and fast ffs_update first, and if that fails,
12321 * then we do the slower ffs_syncvnode of the directory.
12326 if ((error = ffs_update(pvp, 1)) != 0) {
12332 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
12333 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
12334 if (wk->wk_type != D_DIRADD)
12335 panic("softdep_fsync: Unexpected type %s",
12336 TYPENAME(wk->wk_type));
12337 dap = WK_DIRADD(wk);
12338 if (dap->da_state & DIRCHG)
12339 pagedep = dap->da_previous->dm_pagedep;
12341 pagedep = dap->da_pagedep;
12342 pagedep_new_block = pagedep->pd_state & NEWBLOCK;
12345 if (pagedep_new_block && (error =
12346 ffs_syncvnode(pvp, MNT_WAIT, 0))) {
12356 * Flush directory page containing the inode's name.
12358 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
12361 error = bwrite(bp);
12368 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
12376 * Flush all the dirty bitmaps associated with the block device
12377 * before flushing the rest of the dirty blocks so as to reduce
12378 * the number of dependencies that will have to be rolled back.
12383 softdep_fsync_mountdev(vp)
12386 struct buf *bp, *nbp;
12387 struct worklist *wk;
12390 if (!vn_isdisk(vp, NULL))
12391 panic("softdep_fsync_mountdev: vnode not a disk");
12392 bo = &vp->v_bufobj;
12395 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
12397 * If it is already scheduled, skip to the next buffer.
12399 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
12402 if ((bp->b_flags & B_DELWRI) == 0)
12403 panic("softdep_fsync_mountdev: not dirty");
12405 * We are only interested in bitmaps with outstanding
12408 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
12409 wk->wk_type != D_BMSAFEMAP ||
12410 (bp->b_vflags & BV_BKGRDINPROG)) {
12416 (void) bawrite(bp);
12424 * Sync all cylinder groups that were dirty at the time this function is
12425 * called. Newly dirtied cgs will be inserted before the sentinel. This
12426 * is used to flush freedep activity that may be holding up writes to a
12430 sync_cgs(mp, waitfor)
12434 struct bmsafemap *bmsafemap;
12435 struct bmsafemap *sentinel;
12436 struct ufsmount *ump;
12440 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK);
12441 sentinel->sm_cg = -1;
12442 ump = VFSTOUFS(mp);
12445 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next);
12446 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL;
12447 bmsafemap = LIST_NEXT(sentinel, sm_next)) {
12448 /* Skip sentinels and cgs with no work to release. */
12449 if (bmsafemap->sm_cg == -1 ||
12450 (LIST_EMPTY(&bmsafemap->sm_freehd) &&
12451 LIST_EMPTY(&bmsafemap->sm_freewr))) {
12452 LIST_REMOVE(sentinel, sm_next);
12453 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12457 * If we don't get the lock and we're waiting try again, if
12458 * not move on to the next buf and try to sync it.
12460 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor);
12461 if (bp == NULL && waitfor == MNT_WAIT)
12463 LIST_REMOVE(sentinel, sm_next);
12464 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12468 if (waitfor == MNT_NOWAIT)
12471 error = bwrite(bp);
12476 LIST_REMOVE(sentinel, sm_next);
12478 free(sentinel, M_BMSAFEMAP);
12483 * This routine is called when we are trying to synchronously flush a
12484 * file. This routine must eliminate any filesystem metadata dependencies
12485 * so that the syncing routine can succeed.
12488 softdep_sync_metadata(struct vnode *vp)
12494 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12495 ("softdep_sync_metadata called on non-softdep filesystem"));
12497 * Ensure that any direct block dependencies have been cleared,
12498 * truncations are started, and inode references are journaled.
12500 ACQUIRE_LOCK(ip->i_ump);
12502 * Write all journal records to prevent rollbacks on devvp.
12504 if (vp->v_type == VCHR)
12505 softdep_flushjournal(vp->v_mount);
12506 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number);
12508 * Ensure that all truncates are written so we won't find deps on
12511 process_truncates(vp);
12512 FREE_LOCK(ip->i_ump);
12518 * This routine is called when we are attempting to sync a buf with
12519 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any
12520 * other IO it can but returns EBUSY if the buffer is not yet able to
12521 * be written. Dependencies which will not cause rollbacks will always
12525 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
12527 struct indirdep *indirdep;
12528 struct pagedep *pagedep;
12529 struct allocindir *aip;
12530 struct newblk *newblk;
12531 struct ufsmount *ump;
12533 struct worklist *wk;
12536 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
12537 ("softdep_sync_buf called on non-softdep filesystem"));
12539 * For VCHR we just don't want to force flush any dependencies that
12540 * will cause rollbacks.
12542 if (vp->v_type == VCHR) {
12543 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0))
12547 ump = VTOI(vp)->i_ump;
12550 * As we hold the buffer locked, none of its dependencies
12555 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
12556 switch (wk->wk_type) {
12558 case D_ALLOCDIRECT:
12560 newblk = WK_NEWBLK(wk);
12561 if (newblk->nb_jnewblk != NULL) {
12562 if (waitfor == MNT_NOWAIT) {
12566 jwait(&newblk->nb_jnewblk->jn_list, waitfor);
12569 if (newblk->nb_state & DEPCOMPLETE ||
12570 waitfor == MNT_NOWAIT)
12572 nbp = newblk->nb_bmsafemap->sm_buf;
12573 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12577 if ((error = bwrite(nbp)) != 0)
12583 indirdep = WK_INDIRDEP(wk);
12584 if (waitfor == MNT_NOWAIT) {
12585 if (!TAILQ_EMPTY(&indirdep->ir_trunc) ||
12586 !LIST_EMPTY(&indirdep->ir_deplisthd)) {
12591 if (!TAILQ_EMPTY(&indirdep->ir_trunc))
12592 panic("softdep_sync_buf: truncation pending.");
12594 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
12595 newblk = (struct newblk *)aip;
12596 if (newblk->nb_jnewblk != NULL) {
12597 jwait(&newblk->nb_jnewblk->jn_list,
12601 if (newblk->nb_state & DEPCOMPLETE)
12603 nbp = newblk->nb_bmsafemap->sm_buf;
12604 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12608 if ((error = bwrite(nbp)) != 0)
12617 * Only flush directory entries in synchronous passes.
12619 if (waitfor != MNT_WAIT) {
12624 * While syncing snapshots, we must allow recursive
12629 * We are trying to sync a directory that may
12630 * have dependencies on both its own metadata
12631 * and/or dependencies on the inodes of any
12632 * recently allocated files. We walk its diradd
12633 * lists pushing out the associated inode.
12635 pagedep = WK_PAGEDEP(wk);
12636 for (i = 0; i < DAHASHSZ; i++) {
12637 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
12639 if ((error = flush_pagedep_deps(vp, wk->wk_mp,
12640 &pagedep->pd_diraddhd[i]))) {
12655 panic("softdep_sync_buf: Unknown type %s",
12656 TYPENAME(wk->wk_type));
12667 * Flush the dependencies associated with an inodedep.
12668 * Called with splbio blocked.
12671 flush_inodedep_deps(vp, mp, ino)
12676 struct inodedep *inodedep;
12677 struct inoref *inoref;
12678 struct ufsmount *ump;
12679 int error, waitfor;
12682 * This work is done in two passes. The first pass grabs most
12683 * of the buffers and begins asynchronously writing them. The
12684 * only way to wait for these asynchronous writes is to sleep
12685 * on the filesystem vnode which may stay busy for a long time
12686 * if the filesystem is active. So, instead, we make a second
12687 * pass over the dependencies blocking on each write. In the
12688 * usual case we will be blocking against a write that we
12689 * initiated, so when it is done the dependency will have been
12690 * resolved. Thus the second pass is expected to end quickly.
12691 * We give a brief window at the top of the loop to allow
12692 * any pending I/O to complete.
12694 ump = VFSTOUFS(mp);
12696 for (error = 0, waitfor = MNT_NOWAIT; ; ) {
12702 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
12704 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12705 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12707 jwait(&inoref->if_list, MNT_WAIT);
12711 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
12712 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
12713 flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
12714 flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
12717 * If pass2, we are done, otherwise do pass 2.
12719 if (waitfor == MNT_WAIT)
12721 waitfor = MNT_WAIT;
12724 * Try freeing inodedep in case all dependencies have been removed.
12726 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
12727 (void) free_inodedep(inodedep);
12732 * Flush an inode dependency list.
12733 * Called with splbio blocked.
12736 flush_deplist(listhead, waitfor, errorp)
12737 struct allocdirectlst *listhead;
12741 struct allocdirect *adp;
12742 struct newblk *newblk;
12743 struct ufsmount *ump;
12746 if ((adp = TAILQ_FIRST(listhead)) == NULL)
12748 ump = VFSTOUFS(adp->ad_list.wk_mp);
12750 TAILQ_FOREACH(adp, listhead, ad_next) {
12751 newblk = (struct newblk *)adp;
12752 if (newblk->nb_jnewblk != NULL) {
12753 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12756 if (newblk->nb_state & DEPCOMPLETE)
12758 bp = newblk->nb_bmsafemap->sm_buf;
12759 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor);
12761 if (waitfor == MNT_NOWAIT)
12766 if (waitfor == MNT_NOWAIT)
12769 *errorp = bwrite(bp);
12777 * Flush dependencies associated with an allocdirect block.
12780 flush_newblk_dep(vp, mp, lbn)
12785 struct newblk *newblk;
12786 struct ufsmount *ump;
12790 ufs2_daddr_t blkno;
12794 bo = &vp->v_bufobj;
12796 blkno = DIP(ip, i_db[lbn]);
12798 panic("flush_newblk_dep: Missing block");
12799 ump = VFSTOUFS(mp);
12802 * Loop until all dependencies related to this block are satisfied.
12803 * We must be careful to restart after each sleep in case a write
12804 * completes some part of this process for us.
12807 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
12811 if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
12812 panic("flush_newblk_deps: Bad newblk %p", newblk);
12814 * Flush the journal.
12816 if (newblk->nb_jnewblk != NULL) {
12817 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12821 * Write the bitmap dependency.
12823 if ((newblk->nb_state & DEPCOMPLETE) == 0) {
12824 bp = newblk->nb_bmsafemap->sm_buf;
12825 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
12829 error = bwrite(bp);
12836 * Write the buffer.
12840 bp = gbincore(bo, lbn);
12842 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
12843 LK_INTERLOCK, BO_LOCKPTR(bo));
12844 if (error == ENOLCK) {
12846 continue; /* Slept, retry */
12849 break; /* Failed */
12850 if (bp->b_flags & B_DELWRI) {
12852 error = bwrite(bp);
12860 * We have to wait for the direct pointers to
12861 * point at the newdirblk before the dependency
12864 error = ffs_update(vp, 1);
12873 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
12874 * Called with splbio blocked.
12877 flush_pagedep_deps(pvp, mp, diraddhdp)
12880 struct diraddhd *diraddhdp;
12882 struct inodedep *inodedep;
12883 struct inoref *inoref;
12884 struct ufsmount *ump;
12885 struct diradd *dap;
12890 struct diraddhd unfinished;
12892 LIST_INIT(&unfinished);
12893 ump = VFSTOUFS(mp);
12896 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
12898 * Flush ourselves if this directory entry
12899 * has a MKDIR_PARENT dependency.
12901 if (dap->da_state & MKDIR_PARENT) {
12903 if ((error = ffs_update(pvp, 1)) != 0)
12907 * If that cleared dependencies, go on to next.
12909 if (dap != LIST_FIRST(diraddhdp))
12912 * All MKDIR_PARENT dependencies and all the
12913 * NEWBLOCK pagedeps that are contained in direct
12914 * blocks were resolved by doing above ffs_update.
12915 * Pagedeps contained in indirect blocks may
12916 * require a complete sync'ing of the directory.
12917 * We are in the midst of doing a complete sync,
12918 * so if they are not resolved in this pass we
12919 * defer them for now as they will be sync'ed by
12920 * our caller shortly.
12922 LIST_REMOVE(dap, da_pdlist);
12923 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist);
12927 * A newly allocated directory must have its "." and
12928 * ".." entries written out before its name can be
12929 * committed in its parent.
12931 inum = dap->da_newinum;
12932 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12933 panic("flush_pagedep_deps: lost inode1");
12935 * Wait for any pending journal adds to complete so we don't
12936 * cause rollbacks while syncing.
12938 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12939 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12941 jwait(&inoref->if_list, MNT_WAIT);
12945 if (dap->da_state & MKDIR_BODY) {
12947 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
12950 error = flush_newblk_dep(vp, mp, 0);
12952 * If we still have the dependency we might need to
12953 * update the vnode to sync the new link count to
12956 if (error == 0 && dap == LIST_FIRST(diraddhdp))
12957 error = ffs_update(vp, 1);
12963 * If that cleared dependencies, go on to next.
12965 if (dap != LIST_FIRST(diraddhdp))
12967 if (dap->da_state & MKDIR_BODY) {
12968 inodedep_lookup(UFSTOVFS(ump), inum, 0,
12970 panic("flush_pagedep_deps: MKDIR_BODY "
12971 "inodedep %p dap %p vp %p",
12972 inodedep, dap, vp);
12976 * Flush the inode on which the directory entry depends.
12977 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
12978 * the only remaining dependency is that the updated inode
12979 * count must get pushed to disk. The inode has already
12980 * been pushed into its inode buffer (via VOP_UPDATE) at
12981 * the time of the reference count change. So we need only
12982 * locate that buffer, ensure that there will be no rollback
12983 * caused by a bitmap dependency, then write the inode buffer.
12986 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12987 panic("flush_pagedep_deps: lost inode");
12989 * If the inode still has bitmap dependencies,
12990 * push them to disk.
12992 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
12993 bp = inodedep->id_bmsafemap->sm_buf;
12994 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
12998 if ((error = bwrite(bp)) != 0)
13001 if (dap != LIST_FIRST(diraddhdp))
13005 * If the inode is still sitting in a buffer waiting
13006 * to be written or waiting for the link count to be
13007 * adjusted update it here to flush it to disk.
13009 if (dap == LIST_FIRST(diraddhdp)) {
13011 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13014 error = ffs_update(vp, 1);
13021 * If we have failed to get rid of all the dependencies
13022 * then something is seriously wrong.
13024 if (dap == LIST_FIRST(diraddhdp)) {
13025 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
13026 panic("flush_pagedep_deps: failed to flush "
13027 "inodedep %p ino %ju dap %p",
13028 inodedep, (uintmax_t)inum, dap);
13033 while ((dap = LIST_FIRST(&unfinished)) != NULL) {
13034 LIST_REMOVE(dap, da_pdlist);
13035 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist);
13041 * A large burst of file addition or deletion activity can drive the
13042 * memory load excessively high. First attempt to slow things down
13043 * using the techniques below. If that fails, this routine requests
13044 * the offending operations to fall back to running synchronously
13045 * until the memory load returns to a reasonable level.
13048 softdep_slowdown(vp)
13051 struct ufsmount *ump;
13053 int max_softdeps_hard;
13055 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
13056 ("softdep_slowdown called on non-softdep filesystem"));
13057 ump = VFSTOUFS(vp->v_mount);
13061 * Check for journal space if needed.
13063 if (DOINGSUJ(vp)) {
13064 if (journal_space(ump, 0) == 0)
13068 * If the system is under its limits and our filesystem is
13069 * not responsible for more than our share of the usage and
13070 * we are not low on journal space, then no need to slow down.
13072 max_softdeps_hard = max_softdeps * 11 / 10;
13073 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 &&
13074 dep_current[D_INODEDEP] < max_softdeps_hard &&
13075 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 &&
13076 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 &&
13077 ump->softdep_curdeps[D_DIRREM] <
13078 (max_softdeps_hard / 2) / stat_flush_threads &&
13079 ump->softdep_curdeps[D_INODEDEP] <
13080 max_softdeps_hard / stat_flush_threads &&
13081 ump->softdep_curdeps[D_INDIRDEP] <
13082 (max_softdeps_hard / 1000) / stat_flush_threads &&
13083 ump->softdep_curdeps[D_FREEBLKS] <
13084 max_softdeps_hard / stat_flush_threads) {
13089 * If the journal is low or our filesystem is over its limit
13090 * then speedup the cleanup.
13092 if (ump->softdep_curdeps[D_INDIRDEP] <
13093 (max_softdeps_hard / 1000) / stat_flush_threads || jlow)
13094 softdep_speedup(ump);
13095 stat_sync_limit_hit += 1;
13098 * We only slow down the rate at which new dependencies are
13099 * generated if we are not using journaling. With journaling,
13100 * the cleanup should always be sufficient to keep things
13109 * Called by the allocation routines when they are about to fail
13110 * in the hope that we can free up the requested resource (inodes
13113 * First check to see if the work list has anything on it. If it has,
13114 * clean up entries until we successfully free the requested resource.
13115 * Because this process holds inodes locked, we cannot handle any remove
13116 * requests that might block on a locked inode as that could lead to
13117 * deadlock. If the worklist yields none of the requested resource,
13118 * start syncing out vnodes to free up the needed space.
13121 softdep_request_cleanup(fs, vp, cred, resource)
13124 struct ucred *cred;
13127 struct ufsmount *ump;
13129 struct vnode *lvp, *mvp;
13131 ufs2_daddr_t needed;
13135 * If we are being called because of a process doing a
13136 * copy-on-write, then it is not safe to process any
13137 * worklist items as we will recurse into the copyonwrite
13138 * routine. This will result in an incoherent snapshot.
13139 * If the vnode that we hold is a snapshot, we must avoid
13140 * handling other resources that could cause deadlock.
13142 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
13145 if (resource == FLUSH_BLOCKS_WAIT)
13146 stat_cleanup_blkrequests += 1;
13148 stat_cleanup_inorequests += 1;
13151 ump = VFSTOUFS(mp);
13152 mtx_assert(UFS_MTX(ump), MA_OWNED);
13154 error = ffs_update(vp, 1);
13155 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) {
13160 * If we are in need of resources, start by cleaning up
13161 * any block removals associated with our inode.
13164 process_removes(vp);
13165 process_truncates(vp);
13168 * Now clean up at least as many resources as we will need.
13170 * When requested to clean up inodes, the number that are needed
13171 * is set by the number of simultaneous writers (mnt_writeopcount)
13172 * plus a bit of slop (2) in case some more writers show up while
13175 * When requested to free up space, the amount of space that
13176 * we need is enough blocks to allocate a full-sized segment
13177 * (fs_contigsumsize). The number of such segments that will
13178 * be needed is set by the number of simultaneous writers
13179 * (mnt_writeopcount) plus a bit of slop (2) in case some more
13180 * writers show up while we are cleaning.
13182 * Additionally, if we are unpriviledged and allocating space,
13183 * we need to ensure that we clean up enough blocks to get the
13184 * needed number of blocks over the threshhold of the minimum
13185 * number of blocks required to be kept free by the filesystem
13188 if (resource == FLUSH_INODES_WAIT) {
13189 needed = vp->v_mount->mnt_writeopcount + 2;
13190 } else if (resource == FLUSH_BLOCKS_WAIT) {
13191 needed = (vp->v_mount->mnt_writeopcount + 2) *
13192 fs->fs_contigsumsize;
13193 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0))
13194 needed += fragstoblks(fs,
13195 roundup((fs->fs_dsize * fs->fs_minfree / 100) -
13196 fs->fs_cstotal.cs_nffree, fs->fs_frag));
13199 printf("softdep_request_cleanup: Unknown resource type %d\n",
13203 starttime = time_second;
13205 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 &&
13206 fs->fs_cstotal.cs_nbfree <= needed) ||
13207 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13208 fs->fs_cstotal.cs_nifree <= needed)) {
13210 if (ump->softdep_on_worklist > 0 &&
13211 process_worklist_item(UFSTOVFS(ump),
13212 ump->softdep_on_worklist, LK_NOWAIT) != 0)
13213 stat_worklist_push += 1;
13217 * If we still need resources and there are no more worklist
13218 * entries to process to obtain them, we have to start flushing
13219 * the dirty vnodes to force the release of additional requests
13220 * to the worklist that we can then process to reap addition
13221 * resources. We walk the vnodes associated with the mount point
13222 * until we get the needed worklist requests that we can reap.
13224 if ((resource == FLUSH_BLOCKS_WAIT &&
13225 fs->fs_cstotal.cs_nbfree <= needed) ||
13226 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13227 fs->fs_cstotal.cs_nifree <= needed)) {
13228 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) {
13229 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) {
13233 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
13236 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */
13240 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0);
13243 lvp = ump->um_devvp;
13244 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
13245 VOP_FSYNC(lvp, MNT_NOWAIT, curthread);
13246 VOP_UNLOCK(lvp, 0);
13248 if (ump->softdep_on_worklist > 0) {
13249 stat_cleanup_retries += 1;
13252 stat_cleanup_failures += 1;
13254 if (time_second - starttime > stat_cleanup_high_delay)
13255 stat_cleanup_high_delay = time_second - starttime;
13261 softdep_excess_items(struct ufsmount *ump, int item)
13264 KASSERT(item >= 0 && item < D_LAST, ("item %d", item));
13265 return (dep_current[item] > max_softdeps &&
13266 ump->softdep_curdeps[item] > max_softdeps /
13267 stat_flush_threads);
13271 schedule_cleanup(struct mount *mp)
13273 struct ufsmount *ump;
13276 ump = VFSTOUFS(mp);
13280 if ((td->td_pflags & TDP_KTHREAD) != 0 &&
13281 (td->td_proc->p_flag2 & P2_AST_SU) == 0) {
13283 * No ast is delivered to kernel threads, so nobody
13284 * would deref the mp. Some kernel threads
13285 * explicitely check for AST, e.g. NFS daemon does
13286 * this in the serving loop.
13290 if (td->td_su != NULL)
13291 vfs_rel(td->td_su);
13295 td->td_flags |= TDF_ASTPENDING;
13300 softdep_ast_cleanup_proc(void)
13304 struct ufsmount *ump;
13309 while ((mp = td->td_su) != NULL) {
13311 error = vfs_busy(mp, MBF_NOWAIT);
13315 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) {
13316 ump = VFSTOUFS(mp);
13320 if (softdep_excess_items(ump, D_INODEDEP)) {
13322 request_cleanup(mp, FLUSH_INODES);
13324 if (softdep_excess_items(ump, D_DIRREM)) {
13326 request_cleanup(mp, FLUSH_BLOCKS);
13329 if (softdep_excess_items(ump, D_NEWBLK) ||
13330 softdep_excess_items(ump, D_ALLOCDIRECT) ||
13331 softdep_excess_items(ump, D_ALLOCINDIR)) {
13332 error = vn_start_write(NULL, &mp,
13336 VFS_SYNC(mp, MNT_WAIT);
13337 vn_finished_write(mp);
13340 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req)
13349 * If memory utilization has gotten too high, deliberately slow things
13350 * down and speed up the I/O processing.
13353 request_cleanup(mp, resource)
13357 struct thread *td = curthread;
13358 struct ufsmount *ump;
13360 ump = VFSTOUFS(mp);
13363 * We never hold up the filesystem syncer or buf daemon.
13365 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
13368 * First check to see if the work list has gotten backlogged.
13369 * If it has, co-opt this process to help clean up two entries.
13370 * Because this process may hold inodes locked, we cannot
13371 * handle any remove requests that might block on a locked
13372 * inode as that could lead to deadlock. We set TDP_SOFTDEP
13373 * to avoid recursively processing the worklist.
13375 if (ump->softdep_on_worklist > max_softdeps / 10) {
13376 td->td_pflags |= TDP_SOFTDEP;
13377 process_worklist_item(mp, 2, LK_NOWAIT);
13378 td->td_pflags &= ~TDP_SOFTDEP;
13379 stat_worklist_push += 2;
13383 * Next, we attempt to speed up the syncer process. If that
13384 * is successful, then we allow the process to continue.
13386 if (softdep_speedup(ump) &&
13387 resource != FLUSH_BLOCKS_WAIT &&
13388 resource != FLUSH_INODES_WAIT)
13391 * If we are resource constrained on inode dependencies, try
13392 * flushing some dirty inodes. Otherwise, we are constrained
13393 * by file deletions, so try accelerating flushes of directories
13394 * with removal dependencies. We would like to do the cleanup
13395 * here, but we probably hold an inode locked at this point and
13396 * that might deadlock against one that we try to clean. So,
13397 * the best that we can do is request the syncer daemon to do
13398 * the cleanup for us.
13400 switch (resource) {
13403 case FLUSH_INODES_WAIT:
13404 ACQUIRE_GBLLOCK(&lk);
13405 stat_ino_limit_push += 1;
13406 req_clear_inodedeps += 1;
13408 stat_countp = &stat_ino_limit_hit;
13412 case FLUSH_BLOCKS_WAIT:
13413 ACQUIRE_GBLLOCK(&lk);
13414 stat_blk_limit_push += 1;
13415 req_clear_remove += 1;
13417 stat_countp = &stat_blk_limit_hit;
13421 panic("request_cleanup: unknown type");
13424 * Hopefully the syncer daemon will catch up and awaken us.
13425 * We wait at most tickdelay before proceeding in any case.
13427 ACQUIRE_GBLLOCK(&lk);
13430 if (callout_pending(&softdep_callout) == FALSE)
13431 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13434 if ((td->td_pflags & TDP_KTHREAD) == 0)
13435 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
13443 * Awaken processes pausing in request_cleanup and clear proc_waiting
13444 * to indicate that there is no longer a timer running. Pause_timer
13445 * will be called with the global softdep mutex (&lk) locked.
13452 GBLLOCK_OWNED(&lk);
13454 * The callout_ API has acquired mtx and will hold it around this
13457 *stat_countp += proc_waiting;
13458 wakeup(&proc_waiting);
13462 * If requested, try removing inode or removal dependencies.
13465 check_clear_deps(mp)
13470 * If we are suspended, it may be because of our using
13471 * too many inodedeps, so help clear them out.
13473 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended)
13474 clear_inodedeps(mp);
13476 * General requests for cleanup of backed up dependencies
13478 ACQUIRE_GBLLOCK(&lk);
13479 if (req_clear_inodedeps) {
13480 req_clear_inodedeps -= 1;
13482 clear_inodedeps(mp);
13483 ACQUIRE_GBLLOCK(&lk);
13484 wakeup(&proc_waiting);
13486 if (req_clear_remove) {
13487 req_clear_remove -= 1;
13490 ACQUIRE_GBLLOCK(&lk);
13491 wakeup(&proc_waiting);
13497 * Flush out a directory with at least one removal dependency in an effort to
13498 * reduce the number of dirrem, freefile, and freeblks dependency structures.
13504 struct pagedep_hashhead *pagedephd;
13505 struct pagedep *pagedep;
13506 struct ufsmount *ump;
13512 ump = VFSTOUFS(mp);
13515 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) {
13516 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++];
13517 if (ump->pagedep_nextclean > ump->pagedep_hash_size)
13518 ump->pagedep_nextclean = 0;
13519 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
13520 if (LIST_EMPTY(&pagedep->pd_dirremhd))
13522 ino = pagedep->pd_ino;
13523 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13528 * Let unmount clear deps
13530 error = vfs_busy(mp, MBF_NOWAIT);
13533 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13537 softdep_error("clear_remove: vget", error);
13540 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13541 softdep_error("clear_remove: fsync", error);
13542 bo = &vp->v_bufobj;
13548 vn_finished_write(mp);
13556 * Clear out a block of dirty inodes in an effort to reduce
13557 * the number of inodedep dependency structures.
13560 clear_inodedeps(mp)
13563 struct inodedep_hashhead *inodedephd;
13564 struct inodedep *inodedep;
13565 struct ufsmount *ump;
13569 ino_t firstino, lastino, ino;
13571 ump = VFSTOUFS(mp);
13575 * Pick a random inode dependency to be cleared.
13576 * We will then gather up all the inodes in its block
13577 * that have dependencies and flush them out.
13579 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) {
13580 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++];
13581 if (ump->inodedep_nextclean > ump->inodedep_hash_size)
13582 ump->inodedep_nextclean = 0;
13583 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
13586 if (inodedep == NULL)
13589 * Find the last inode in the block with dependencies.
13591 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
13592 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
13593 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
13596 * Asynchronously push all but the last inode with dependencies.
13597 * Synchronously push the last inode with dependencies to ensure
13598 * that the inode block gets written to free up the inodedeps.
13600 for (ino = firstino; ino <= lastino; ino++) {
13601 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
13603 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13606 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
13608 vn_finished_write(mp);
13612 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13613 FFSV_FORCEINSMQ)) != 0) {
13614 softdep_error("clear_inodedeps: vget", error);
13616 vn_finished_write(mp);
13621 if (ino == lastino) {
13622 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)))
13623 softdep_error("clear_inodedeps: fsync1", error);
13625 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13626 softdep_error("clear_inodedeps: fsync2", error);
13627 BO_LOCK(&vp->v_bufobj);
13629 BO_UNLOCK(&vp->v_bufobj);
13632 vn_finished_write(mp);
13638 softdep_buf_append(bp, wkhd)
13640 struct workhead *wkhd;
13642 struct worklist *wk;
13643 struct ufsmount *ump;
13645 if ((wk = LIST_FIRST(wkhd)) == NULL)
13647 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13648 ("softdep_buf_append called on non-softdep filesystem"));
13649 ump = VFSTOUFS(wk->wk_mp);
13651 while ((wk = LIST_FIRST(wkhd)) != NULL) {
13652 WORKLIST_REMOVE(wk);
13653 WORKLIST_INSERT(&bp->b_dep, wk);
13660 softdep_inode_append(ip, cred, wkhd)
13662 struct ucred *cred;
13663 struct workhead *wkhd;
13669 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
13670 ("softdep_inode_append called on non-softdep filesystem"));
13672 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
13673 (int)fs->fs_bsize, cred, &bp);
13676 softdep_freework(wkhd);
13679 softdep_buf_append(bp, wkhd);
13684 softdep_freework(wkhd)
13685 struct workhead *wkhd;
13687 struct worklist *wk;
13688 struct ufsmount *ump;
13690 if ((wk = LIST_FIRST(wkhd)) == NULL)
13692 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13693 ("softdep_freework called on non-softdep filesystem"));
13694 ump = VFSTOUFS(wk->wk_mp);
13696 handle_jwork(wkhd);
13701 * Function to determine if the buffer has outstanding dependencies
13702 * that will cause a roll-back if the buffer is written. If wantcount
13703 * is set, return number of dependencies, otherwise just yes or no.
13706 softdep_count_dependencies(bp, wantcount)
13710 struct worklist *wk;
13711 struct ufsmount *ump;
13712 struct bmsafemap *bmsafemap;
13713 struct freework *freework;
13714 struct inodedep *inodedep;
13715 struct indirdep *indirdep;
13716 struct freeblks *freeblks;
13717 struct allocindir *aip;
13718 struct pagedep *pagedep;
13719 struct dirrem *dirrem;
13720 struct newblk *newblk;
13721 struct mkdir *mkdir;
13722 struct diradd *dap;
13726 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
13728 ump = VFSTOUFS(wk->wk_mp);
13730 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
13731 switch (wk->wk_type) {
13734 inodedep = WK_INODEDEP(wk);
13735 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
13736 /* bitmap allocation dependency */
13741 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
13742 /* direct block pointer dependency */
13747 if (TAILQ_FIRST(&inodedep->id_extupdt)) {
13748 /* direct block pointer dependency */
13753 if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
13754 /* Add reference dependency. */
13762 indirdep = WK_INDIRDEP(wk);
13764 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) {
13765 /* indirect truncation dependency */
13771 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
13772 /* indirect block pointer dependency */
13780 pagedep = WK_PAGEDEP(wk);
13781 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
13782 if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
13783 /* Journal remove ref dependency. */
13789 for (i = 0; i < DAHASHSZ; i++) {
13791 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
13792 /* directory entry dependency */
13801 bmsafemap = WK_BMSAFEMAP(wk);
13802 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
13803 /* Add reference dependency. */
13808 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
13809 /* Allocate block dependency. */
13817 freeblks = WK_FREEBLKS(wk);
13818 if (LIST_FIRST(&freeblks->fb_jblkdephd)) {
13819 /* Freeblk journal dependency. */
13826 case D_ALLOCDIRECT:
13828 newblk = WK_NEWBLK(wk);
13829 if (newblk->nb_jnewblk) {
13830 /* Journal allocate dependency. */
13838 mkdir = WK_MKDIR(wk);
13839 if (mkdir->md_jaddref) {
13840 /* Journal reference dependency. */
13852 /* never a dependency on these blocks */
13856 panic("softdep_count_dependencies: Unexpected type %s",
13857 TYPENAME(wk->wk_type));
13867 * Acquire exclusive access to a buffer.
13868 * Must be called with a locked mtx parameter.
13869 * Return acquired buffer or NULL on failure.
13871 static struct buf *
13872 getdirtybuf(bp, lock, waitfor)
13874 struct rwlock *lock;
13879 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
13880 if (waitfor != MNT_WAIT)
13882 error = BUF_LOCK(bp,
13883 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock);
13885 * Even if we sucessfully acquire bp here, we have dropped
13886 * lock, which may violates our guarantee.
13890 else if (error != ENOLCK)
13891 panic("getdirtybuf: inconsistent lock: %d", error);
13895 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13896 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) {
13898 BO_LOCK(bp->b_bufobj);
13900 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13901 bp->b_vflags |= BV_BKGRDWAIT;
13902 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj),
13903 PRIBIO | PDROP, "getbuf", 0);
13905 BO_UNLOCK(bp->b_bufobj);
13910 if (waitfor != MNT_WAIT)
13913 * The lock argument must be bp->b_vp's mutex in
13916 #ifdef DEBUG_VFS_LOCKS
13917 if (bp->b_vp->v_type != VCHR)
13918 ASSERT_BO_WLOCKED(bp->b_bufobj);
13920 bp->b_vflags |= BV_BKGRDWAIT;
13921 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0);
13924 if ((bp->b_flags & B_DELWRI) == 0) {
13934 * Check if it is safe to suspend the file system now. On entry,
13935 * the vnode interlock for devvp should be held. Return 0 with
13936 * the mount interlock held if the file system can be suspended now,
13937 * otherwise return EAGAIN with the mount interlock held.
13940 softdep_check_suspend(struct mount *mp,
13941 struct vnode *devvp,
13942 int softdep_depcnt,
13943 int softdep_accdepcnt,
13944 int secondary_writes,
13945 int secondary_accwrites)
13948 struct ufsmount *ump;
13949 struct inodedep *inodedep;
13950 int error, unlinked;
13952 bo = &devvp->v_bufobj;
13953 ASSERT_BO_WLOCKED(bo);
13956 * If we are not running with soft updates, then we need only
13957 * deal with secondary writes as we try to suspend.
13959 if (MOUNTEDSOFTDEP(mp) == 0) {
13961 while (mp->mnt_secondary_writes != 0) {
13963 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
13964 (PUSER - 1) | PDROP, "secwr", 0);
13970 * Reasons for needing more work before suspend:
13971 * - Dirty buffers on devvp.
13972 * - Secondary writes occurred after start of vnode sync loop
13975 if (bo->bo_numoutput > 0 ||
13976 bo->bo_dirty.bv_cnt > 0 ||
13977 secondary_writes != 0 ||
13978 mp->mnt_secondary_writes != 0 ||
13979 secondary_accwrites != mp->mnt_secondary_accwrites)
13986 * If we are running with soft updates, then we need to coordinate
13987 * with them as we try to suspend.
13989 ump = VFSTOUFS(mp);
13991 if (!TRY_ACQUIRE_LOCK(ump)) {
13999 if (mp->mnt_secondary_writes != 0) {
14002 msleep(&mp->mnt_secondary_writes,
14004 (PUSER - 1) | PDROP, "secwr", 0);
14012 if (MOUNTEDSUJ(mp)) {
14013 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked);
14015 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
14016 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS |
14017 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS |
14019 !check_inodedep_free(inodedep))
14026 * Reasons for needing more work before suspend:
14027 * - Dirty buffers on devvp.
14028 * - Softdep activity occurred after start of vnode sync loop
14029 * - Secondary writes occurred after start of vnode sync loop
14032 if (bo->bo_numoutput > 0 ||
14033 bo->bo_dirty.bv_cnt > 0 ||
14034 softdep_depcnt != unlinked ||
14035 ump->softdep_deps != unlinked ||
14036 softdep_accdepcnt != ump->softdep_accdeps ||
14037 secondary_writes != 0 ||
14038 mp->mnt_secondary_writes != 0 ||
14039 secondary_accwrites != mp->mnt_secondary_accwrites)
14048 * Get the number of dependency structures for the file system, both
14049 * the current number and the total number allocated. These will
14050 * later be used to detect that softdep processing has occurred.
14053 softdep_get_depcounts(struct mount *mp,
14054 int *softdep_depsp,
14055 int *softdep_accdepsp)
14057 struct ufsmount *ump;
14059 if (MOUNTEDSOFTDEP(mp) == 0) {
14060 *softdep_depsp = 0;
14061 *softdep_accdepsp = 0;
14064 ump = VFSTOUFS(mp);
14066 *softdep_depsp = ump->softdep_deps;
14067 *softdep_accdepsp = ump->softdep_accdeps;
14072 * Wait for pending output on a vnode to complete.
14073 * Must be called with vnode lock and interlock locked.
14075 * XXX: Should just be a call to bufobj_wwait().
14083 bo = &vp->v_bufobj;
14084 ASSERT_VOP_LOCKED(vp, "drain_output");
14085 ASSERT_BO_WLOCKED(bo);
14087 while (bo->bo_numoutput) {
14088 bo->bo_flag |= BO_WWAIT;
14089 msleep((caddr_t)&bo->bo_numoutput,
14090 BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0);
14095 * Called whenever a buffer that is being invalidated or reallocated
14096 * contains dependencies. This should only happen if an I/O error has
14097 * occurred. The routine is called with the buffer locked.
14100 softdep_deallocate_dependencies(bp)
14104 if ((bp->b_ioflags & BIO_ERROR) == 0)
14105 panic("softdep_deallocate_dependencies: dangling deps");
14106 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL)
14107 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
14109 printf("softdep_deallocate_dependencies: "
14110 "got error %d while accessing filesystem\n", bp->b_error);
14111 if (bp->b_error != ENXIO)
14112 panic("softdep_deallocate_dependencies: unrecovered I/O error");
14116 * Function to handle asynchronous write errors in the filesystem.
14119 softdep_error(func, error)
14124 /* XXX should do something better! */
14125 printf("%s: got error %d while accessing filesystem\n", func, error);
14131 inodedep_print(struct inodedep *inodedep, int verbose)
14133 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d"
14135 inodedep, inodedep->id_fs, inodedep->id_state,
14136 (intmax_t)inodedep->id_ino,
14137 (intmax_t)fsbtodb(inodedep->id_fs,
14138 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
14139 inodedep->id_nlinkdelta, inodedep->id_savednlink,
14140 inodedep->id_savedino1);
14145 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
14147 LIST_FIRST(&inodedep->id_pendinghd),
14148 LIST_FIRST(&inodedep->id_bufwait),
14149 LIST_FIRST(&inodedep->id_inowait),
14150 TAILQ_FIRST(&inodedep->id_inoreflst),
14151 inodedep->id_mkdiradd);
14152 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
14153 TAILQ_FIRST(&inodedep->id_inoupdt),
14154 TAILQ_FIRST(&inodedep->id_newinoupdt),
14155 TAILQ_FIRST(&inodedep->id_extupdt),
14156 TAILQ_FIRST(&inodedep->id_newextupdt));
14159 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
14162 if (have_addr == 0) {
14163 db_printf("Address required\n");
14166 inodedep_print((struct inodedep*)addr, 1);
14169 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
14171 struct inodedep_hashhead *inodedephd;
14172 struct inodedep *inodedep;
14173 struct ufsmount *ump;
14176 if (have_addr == 0) {
14177 db_printf("Address required\n");
14180 ump = (struct ufsmount *)addr;
14181 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) {
14182 inodedephd = &ump->inodedep_hashtbl[cnt];
14183 LIST_FOREACH(inodedep, inodedephd, id_hash) {
14184 inodedep_print(inodedep, 0);
14189 DB_SHOW_COMMAND(worklist, db_show_worklist)
14191 struct worklist *wk;
14193 if (have_addr == 0) {
14194 db_printf("Address required\n");
14197 wk = (struct worklist *)addr;
14198 printf("worklist: %p type %s state 0x%X\n",
14199 wk, TYPENAME(wk->wk_type), wk->wk_state);
14202 DB_SHOW_COMMAND(workhead, db_show_workhead)
14204 struct workhead *wkhd;
14205 struct worklist *wk;
14208 if (have_addr == 0) {
14209 db_printf("Address required\n");
14212 wkhd = (struct workhead *)addr;
14213 wk = LIST_FIRST(wkhd);
14214 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
14215 db_printf("worklist: %p type %s state 0x%X",
14216 wk, TYPENAME(wk->wk_type), wk->wk_state);
14218 db_printf("workhead overflow");
14223 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
14225 struct mkdirlist *mkdirlisthd;
14226 struct jaddref *jaddref;
14227 struct diradd *diradd;
14228 struct mkdir *mkdir;
14230 if (have_addr == 0) {
14231 db_printf("Address required\n");
14234 mkdirlisthd = (struct mkdirlist *)addr;
14235 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) {
14236 diradd = mkdir->md_diradd;
14237 db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
14238 mkdir, mkdir->md_state, diradd, diradd->da_state);
14239 if ((jaddref = mkdir->md_jaddref) != NULL)
14240 db_printf(" jaddref %p jaddref state 0x%X",
14241 jaddref, jaddref->ja_state);
14246 /* exported to ffs_vfsops.c */
14247 extern void db_print_ffs(struct ufsmount *ump);
14249 db_print_ffs(struct ufsmount *ump)
14251 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n",
14252 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname,
14253 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist,
14254 ump->softdep_deps, ump->softdep_req);
14259 #endif /* SOFTUPDATES */