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 handling
3000 * other resources that could cause deadlock.
3002 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)))
3004 ump = VFSTOUFS(vp->v_mount);
3006 if (journal_space(ump, 0)) {
3012 if (waitok == MNT_NOWAIT)
3015 * Attempt to sync this vnode once to flush any journal
3016 * work attached to it.
3018 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0)
3019 ffs_syncvnode(vp, waitok, 0);
3021 process_removes(vp);
3022 process_truncates(vp);
3023 if (journal_space(ump, 0) == 0) {
3024 softdep_speedup(ump);
3025 if (journal_space(ump, 1) == 0)
3026 journal_suspend(ump);
3034 * Before adjusting a link count on a vnode verify that we have sufficient
3035 * journal space. If not, process operations that depend on the currently
3036 * locked pair of vnodes to try to flush space as the syncer, buf daemon,
3037 * and softdep flush threads can not acquire these locks to reclaim space.
3040 softdep_prelink(dvp, vp)
3044 struct ufsmount *ump;
3046 ump = VFSTOUFS(dvp->v_mount);
3049 * Nothing to do if we have sufficient journal space.
3050 * If we currently hold the snapshot lock, we must avoid
3051 * handling other resources that could cause deadlock.
3053 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp))))
3058 ffs_syncvnode(vp, MNT_NOWAIT, 0);
3059 ffs_syncvnode(dvp, MNT_WAIT, 0);
3061 /* Process vp before dvp as it may create .. removes. */
3063 process_removes(vp);
3064 process_truncates(vp);
3066 process_removes(dvp);
3067 process_truncates(dvp);
3068 softdep_speedup(ump);
3069 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT);
3070 if (journal_space(ump, 0) == 0) {
3071 softdep_speedup(ump);
3072 if (journal_space(ump, 1) == 0)
3073 journal_suspend(ump);
3078 jseg_write(ump, jseg, data)
3079 struct ufsmount *ump;
3083 struct jsegrec *rec;
3085 rec = (struct jsegrec *)data;
3086 rec->jsr_seq = jseg->js_seq;
3087 rec->jsr_oldest = jseg->js_oldseq;
3088 rec->jsr_cnt = jseg->js_cnt;
3089 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize;
3091 rec->jsr_time = ump->um_fs->fs_mtime;
3095 inoref_write(inoref, jseg, rec)
3096 struct inoref *inoref;
3098 struct jrefrec *rec;
3101 inoref->if_jsegdep->jd_seg = jseg;
3102 rec->jr_ino = inoref->if_ino;
3103 rec->jr_parent = inoref->if_parent;
3104 rec->jr_nlink = inoref->if_nlink;
3105 rec->jr_mode = inoref->if_mode;
3106 rec->jr_diroff = inoref->if_diroff;
3110 jaddref_write(jaddref, jseg, data)
3111 struct jaddref *jaddref;
3115 struct jrefrec *rec;
3117 rec = (struct jrefrec *)data;
3118 rec->jr_op = JOP_ADDREF;
3119 inoref_write(&jaddref->ja_ref, jseg, rec);
3123 jremref_write(jremref, jseg, data)
3124 struct jremref *jremref;
3128 struct jrefrec *rec;
3130 rec = (struct jrefrec *)data;
3131 rec->jr_op = JOP_REMREF;
3132 inoref_write(&jremref->jr_ref, jseg, rec);
3136 jmvref_write(jmvref, jseg, data)
3137 struct jmvref *jmvref;
3143 rec = (struct jmvrec *)data;
3144 rec->jm_op = JOP_MVREF;
3145 rec->jm_ino = jmvref->jm_ino;
3146 rec->jm_parent = jmvref->jm_parent;
3147 rec->jm_oldoff = jmvref->jm_oldoff;
3148 rec->jm_newoff = jmvref->jm_newoff;
3152 jnewblk_write(jnewblk, jseg, data)
3153 struct jnewblk *jnewblk;
3157 struct jblkrec *rec;
3159 jnewblk->jn_jsegdep->jd_seg = jseg;
3160 rec = (struct jblkrec *)data;
3161 rec->jb_op = JOP_NEWBLK;
3162 rec->jb_ino = jnewblk->jn_ino;
3163 rec->jb_blkno = jnewblk->jn_blkno;
3164 rec->jb_lbn = jnewblk->jn_lbn;
3165 rec->jb_frags = jnewblk->jn_frags;
3166 rec->jb_oldfrags = jnewblk->jn_oldfrags;
3170 jfreeblk_write(jfreeblk, jseg, data)
3171 struct jfreeblk *jfreeblk;
3175 struct jblkrec *rec;
3177 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg;
3178 rec = (struct jblkrec *)data;
3179 rec->jb_op = JOP_FREEBLK;
3180 rec->jb_ino = jfreeblk->jf_ino;
3181 rec->jb_blkno = jfreeblk->jf_blkno;
3182 rec->jb_lbn = jfreeblk->jf_lbn;
3183 rec->jb_frags = jfreeblk->jf_frags;
3184 rec->jb_oldfrags = 0;
3188 jfreefrag_write(jfreefrag, jseg, data)
3189 struct jfreefrag *jfreefrag;
3193 struct jblkrec *rec;
3195 jfreefrag->fr_jsegdep->jd_seg = jseg;
3196 rec = (struct jblkrec *)data;
3197 rec->jb_op = JOP_FREEBLK;
3198 rec->jb_ino = jfreefrag->fr_ino;
3199 rec->jb_blkno = jfreefrag->fr_blkno;
3200 rec->jb_lbn = jfreefrag->fr_lbn;
3201 rec->jb_frags = jfreefrag->fr_frags;
3202 rec->jb_oldfrags = 0;
3206 jtrunc_write(jtrunc, jseg, data)
3207 struct jtrunc *jtrunc;
3211 struct jtrncrec *rec;
3213 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg;
3214 rec = (struct jtrncrec *)data;
3215 rec->jt_op = JOP_TRUNC;
3216 rec->jt_ino = jtrunc->jt_ino;
3217 rec->jt_size = jtrunc->jt_size;
3218 rec->jt_extsize = jtrunc->jt_extsize;
3222 jfsync_write(jfsync, jseg, data)
3223 struct jfsync *jfsync;
3227 struct jtrncrec *rec;
3229 rec = (struct jtrncrec *)data;
3230 rec->jt_op = JOP_SYNC;
3231 rec->jt_ino = jfsync->jfs_ino;
3232 rec->jt_size = jfsync->jfs_size;
3233 rec->jt_extsize = jfsync->jfs_extsize;
3237 softdep_flushjournal(mp)
3240 struct jblocks *jblocks;
3241 struct ufsmount *ump;
3243 if (MOUNTEDSUJ(mp) == 0)
3246 jblocks = ump->softdep_jblocks;
3248 while (ump->softdep_on_journal) {
3249 jblocks->jb_needseg = 1;
3250 softdep_process_journal(mp, NULL, MNT_WAIT);
3255 static void softdep_synchronize_completed(struct bio *);
3256 static void softdep_synchronize(struct bio *, struct ufsmount *, void *);
3259 softdep_synchronize_completed(bp)
3262 struct jseg *oldest;
3264 struct ufsmount *ump;
3267 * caller1 marks the last segment written before we issued the
3268 * synchronize cache.
3270 jseg = bp->bio_caller1;
3275 ump = VFSTOUFS(jseg->js_list.wk_mp);
3279 * Mark all the journal entries waiting on the synchronize cache
3280 * as completed so they may continue on.
3282 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) {
3283 jseg->js_state |= COMPLETE;
3285 jseg = TAILQ_PREV(jseg, jseglst, js_next);
3288 * Restart deferred journal entry processing from the oldest
3292 complete_jsegs(oldest);
3299 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering
3300 * barriers. The journal must be written prior to any blocks that depend
3301 * on it and the journal can not be released until the blocks have be
3302 * written. This code handles both barriers simultaneously.
3305 softdep_synchronize(bp, ump, caller1)
3307 struct ufsmount *ump;
3311 bp->bio_cmd = BIO_FLUSH;
3312 bp->bio_flags |= BIO_ORDERED;
3313 bp->bio_data = NULL;
3314 bp->bio_offset = ump->um_cp->provider->mediasize;
3316 bp->bio_done = softdep_synchronize_completed;
3317 bp->bio_caller1 = caller1;
3319 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private);
3323 * Flush some journal records to disk.
3326 softdep_process_journal(mp, needwk, flags)
3328 struct worklist *needwk;
3331 struct jblocks *jblocks;
3332 struct ufsmount *ump;
3333 struct worklist *wk;
3341 int jrecmin; /* Minimum records per block. */
3342 int jrecmax; /* Maximum records per block. */
3348 if (MOUNTEDSUJ(mp) == 0)
3350 shouldflush = softdep_flushcache;
3356 jblocks = ump->softdep_jblocks;
3357 devbsize = ump->um_devvp->v_bufobj.bo_bsize;
3359 * We write anywhere between a disk block and fs block. The upper
3360 * bound is picked to prevent buffer cache fragmentation and limit
3361 * processing time per I/O.
3363 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */
3364 jrecmax = (fs->fs_bsize / devbsize) * jrecmin;
3367 cnt = ump->softdep_on_journal;
3369 * Criteria for writing a segment:
3370 * 1) We have a full block.
3371 * 2) We're called from jwait() and haven't found the
3373 * 3) Always write if needseg is set.
3374 * 4) If we are called from process_worklist and have
3375 * not yet written anything we write a partial block
3376 * to enforce a 1 second maximum latency on journal
3379 if (cnt < (jrecmax - 1) && needwk == NULL &&
3380 jblocks->jb_needseg == 0 && (segwritten || cnt == 0))
3384 * Verify some free journal space. softdep_prealloc() should
3385 * guarantee that we don't run out so this is indicative of
3386 * a problem with the flow control. Try to recover
3387 * gracefully in any event.
3389 while (jblocks->jb_free == 0) {
3390 if (flags != MNT_WAIT)
3392 printf("softdep: Out of journal space!\n");
3393 softdep_speedup(ump);
3394 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz);
3397 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS);
3398 workitem_alloc(&jseg->js_list, D_JSEG, mp);
3399 LIST_INIT(&jseg->js_entries);
3400 LIST_INIT(&jseg->js_indirs);
3401 jseg->js_state = ATTACHED;
3402 if (shouldflush == 0)
3403 jseg->js_state |= COMPLETE;
3404 else if (bio == NULL)
3405 bio = g_alloc_bio();
3406 jseg->js_jblocks = jblocks;
3407 bp = geteblk(fs->fs_bsize, 0);
3410 * If there was a race while we were allocating the block
3411 * and jseg the entry we care about was likely written.
3412 * We bail out in both the WAIT and NOWAIT case and assume
3413 * the caller will loop if the entry it cares about is
3416 cnt = ump->softdep_on_journal;
3417 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) {
3418 bp->b_flags |= B_INVAL | B_NOCACHE;
3419 WORKITEM_FREE(jseg, D_JSEG);
3426 * Calculate the disk block size required for the available
3427 * records rounded to the min size.
3431 else if (cnt < jrecmax)
3432 size = howmany(cnt, jrecmin) * devbsize;
3434 size = fs->fs_bsize;
3436 * Allocate a disk block for this journal data and account
3437 * for truncation of the requested size if enough contiguous
3438 * space was not available.
3440 bp->b_blkno = jblocks_alloc(jblocks, size, &size);
3441 bp->b_lblkno = bp->b_blkno;
3442 bp->b_offset = bp->b_blkno * DEV_BSIZE;
3443 bp->b_bcount = size;
3444 bp->b_flags &= ~B_INVAL;
3445 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY;
3447 * Initialize our jseg with cnt records. Assign the next
3448 * sequence number to it and link it in-order.
3450 cnt = MIN(cnt, (size / devbsize) * jrecmin);
3453 jseg->js_refs = cnt + 1; /* Self ref. */
3454 jseg->js_size = size;
3455 jseg->js_seq = jblocks->jb_nextseq++;
3456 if (jblocks->jb_oldestseg == NULL)
3457 jblocks->jb_oldestseg = jseg;
3458 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq;
3459 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next);
3460 if (jblocks->jb_writeseg == NULL)
3461 jblocks->jb_writeseg = jseg;
3463 * Start filling in records from the pending list.
3469 * Always put a header on the first block.
3470 * XXX As with below, there might not be a chance to get
3471 * into the loop. Ensure that something valid is written.
3473 jseg_write(ump, jseg, data);
3475 data = bp->b_data + off;
3478 * XXX Something is wrong here. There's no work to do,
3479 * but we need to perform and I/O and allow it to complete
3482 if (LIST_EMPTY(&ump->softdep_journal_pending))
3483 stat_emptyjblocks++;
3485 while ((wk = LIST_FIRST(&ump->softdep_journal_pending))
3489 /* Place a segment header on every device block. */
3490 if ((off % devbsize) == 0) {
3491 jseg_write(ump, jseg, data);
3493 data = bp->b_data + off;
3497 remove_from_journal(wk);
3498 wk->wk_state |= INPROGRESS;
3499 WORKLIST_INSERT(&jseg->js_entries, wk);
3500 switch (wk->wk_type) {
3502 jaddref_write(WK_JADDREF(wk), jseg, data);
3505 jremref_write(WK_JREMREF(wk), jseg, data);
3508 jmvref_write(WK_JMVREF(wk), jseg, data);
3511 jnewblk_write(WK_JNEWBLK(wk), jseg, data);
3514 jfreeblk_write(WK_JFREEBLK(wk), jseg, data);
3517 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data);
3520 jtrunc_write(WK_JTRUNC(wk), jseg, data);
3523 jfsync_write(WK_JFSYNC(wk), jseg, data);
3526 panic("process_journal: Unknown type %s",
3527 TYPENAME(wk->wk_type));
3531 data = bp->b_data + off;
3535 /* Clear any remaining space so we don't leak kernel data */
3537 bzero(data, size - off);
3540 * Write this one buffer and continue.
3543 jblocks->jb_needseg = 0;
3544 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list);
3546 pbgetvp(ump->um_devvp, bp);
3548 * We only do the blocking wait once we find the journal
3549 * entry we're looking for.
3551 if (needwk == NULL && flags == MNT_WAIT)
3558 * If we wrote a segment issue a synchronize cache so the journal
3559 * is reflected on disk before the data is written. Since reclaiming
3560 * journal space also requires writing a journal record this
3561 * process also enforces a barrier before reclamation.
3563 if (segwritten && shouldflush) {
3564 softdep_synchronize(bio, ump,
3565 TAILQ_LAST(&jblocks->jb_segs, jseglst));
3569 * If we've suspended the filesystem because we ran out of journal
3570 * space either try to sync it here to make some progress or
3571 * unsuspend it if we already have.
3573 if (flags == 0 && jblocks->jb_suspended) {
3574 if (journal_unsuspend(ump))
3577 VFS_SYNC(mp, MNT_NOWAIT);
3578 ffs_sbupdate(ump, MNT_WAIT, 0);
3584 * Complete a jseg, allowing all dependencies awaiting journal writes
3585 * to proceed. Each journal dependency also attaches a jsegdep to dependent
3586 * structures so that the journal segment can be freed to reclaim space.
3592 struct worklist *wk;
3593 struct jmvref *jmvref;
3599 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
3600 WORKLIST_REMOVE(wk);
3601 waiting = wk->wk_state & IOWAITING;
3602 wk->wk_state &= ~(INPROGRESS | IOWAITING);
3603 wk->wk_state |= COMPLETE;
3604 KASSERT(i++ < jseg->js_cnt,
3605 ("handle_written_jseg: overflow %d >= %d",
3606 i - 1, jseg->js_cnt));
3607 switch (wk->wk_type) {
3609 handle_written_jaddref(WK_JADDREF(wk));
3612 handle_written_jremref(WK_JREMREF(wk));
3615 rele_jseg(jseg); /* No jsegdep. */
3616 jmvref = WK_JMVREF(wk);
3617 LIST_REMOVE(jmvref, jm_deps);
3618 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0)
3619 free_pagedep(jmvref->jm_pagedep);
3620 WORKITEM_FREE(jmvref, D_JMVREF);
3623 handle_written_jnewblk(WK_JNEWBLK(wk));
3626 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep);
3629 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep);
3632 rele_jseg(jseg); /* No jsegdep. */
3633 WORKITEM_FREE(wk, D_JFSYNC);
3636 handle_written_jfreefrag(WK_JFREEFRAG(wk));
3639 panic("handle_written_jseg: Unknown type %s",
3640 TYPENAME(wk->wk_type));
3646 /* Release the self reference so the structure may be freed. */
3651 * Determine which jsegs are ready for completion processing. Waits for
3652 * synchronize cache to complete as well as forcing in-order completion
3653 * of journal entries.
3656 complete_jsegs(jseg)
3659 struct jblocks *jblocks;
3662 jblocks = jseg->js_jblocks;
3664 * Don't allow out of order completions. If this isn't the first
3665 * block wait for it to write before we're done.
3667 if (jseg != jblocks->jb_writeseg)
3669 /* Iterate through available jsegs processing their entries. */
3670 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) {
3671 jblocks->jb_oldestwrseq = jseg->js_oldseq;
3672 jsegn = TAILQ_NEXT(jseg, js_next);
3673 complete_jseg(jseg);
3676 jblocks->jb_writeseg = jseg;
3678 * Attempt to free jsegs now that oldestwrseq may have advanced.
3680 free_jsegs(jblocks);
3684 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle
3685 * the final completions.
3688 handle_written_jseg(jseg, bp)
3693 if (jseg->js_refs == 0)
3694 panic("handle_written_jseg: No self-reference on %p", jseg);
3695 jseg->js_state |= DEPCOMPLETE;
3697 * We'll never need this buffer again, set flags so it will be
3700 bp->b_flags |= B_INVAL | B_NOCACHE;
3702 complete_jsegs(jseg);
3705 static inline struct jsegdep *
3707 struct inoref *inoref;
3709 struct jsegdep *jsegdep;
3711 jsegdep = inoref->if_jsegdep;
3712 inoref->if_jsegdep = NULL;
3718 * Called once a jremref has made it to stable store. The jremref is marked
3719 * complete and we attempt to free it. Any pagedeps writes sleeping waiting
3720 * for the jremref to complete will be awoken by free_jremref.
3723 handle_written_jremref(jremref)
3724 struct jremref *jremref;
3726 struct inodedep *inodedep;
3727 struct jsegdep *jsegdep;
3728 struct dirrem *dirrem;
3730 /* Grab the jsegdep. */
3731 jsegdep = inoref_jseg(&jremref->jr_ref);
3733 * Remove us from the inoref list.
3735 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
3737 panic("handle_written_jremref: Lost inodedep");
3738 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
3740 * Complete the dirrem.
3742 dirrem = jremref->jr_dirrem;
3743 jremref->jr_dirrem = NULL;
3744 LIST_REMOVE(jremref, jr_deps);
3745 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
3746 jwork_insert(&dirrem->dm_jwork, jsegdep);
3747 if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
3748 (dirrem->dm_state & COMPLETE) != 0)
3749 add_to_worklist(&dirrem->dm_list, 0);
3750 free_jremref(jremref);
3754 * Called once a jaddref has made it to stable store. The dependency is
3755 * marked complete and any dependent structures are added to the inode
3756 * bufwait list to be completed as soon as it is written. If a bitmap write
3757 * depends on this entry we move the inode into the inodedephd of the
3758 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
3761 handle_written_jaddref(jaddref)
3762 struct jaddref *jaddref;
3764 struct jsegdep *jsegdep;
3765 struct inodedep *inodedep;
3766 struct diradd *diradd;
3767 struct mkdir *mkdir;
3769 /* Grab the jsegdep. */
3770 jsegdep = inoref_jseg(&jaddref->ja_ref);
3773 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3775 panic("handle_written_jaddref: Lost inodedep.");
3776 if (jaddref->ja_diradd == NULL)
3777 panic("handle_written_jaddref: No dependency");
3778 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
3779 diradd = jaddref->ja_diradd;
3780 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
3781 } else if (jaddref->ja_state & MKDIR_PARENT) {
3782 mkdir = jaddref->ja_mkdir;
3783 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
3784 } else if (jaddref->ja_state & MKDIR_BODY)
3785 mkdir = jaddref->ja_mkdir;
3787 panic("handle_written_jaddref: Unknown dependency %p",
3788 jaddref->ja_diradd);
3789 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */
3791 * Remove us from the inode list.
3793 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
3795 * The mkdir may be waiting on the jaddref to clear before freeing.
3798 KASSERT(mkdir->md_list.wk_type == D_MKDIR,
3799 ("handle_written_jaddref: Incorrect type for mkdir %s",
3800 TYPENAME(mkdir->md_list.wk_type)));
3801 mkdir->md_jaddref = NULL;
3802 diradd = mkdir->md_diradd;
3803 mkdir->md_state |= DEPCOMPLETE;
3804 complete_mkdir(mkdir);
3806 jwork_insert(&diradd->da_jwork, jsegdep);
3807 if (jaddref->ja_state & NEWBLOCK) {
3808 inodedep->id_state |= ONDEPLIST;
3809 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
3812 free_jaddref(jaddref);
3816 * Called once a jnewblk journal is written. The allocdirect or allocindir
3817 * is placed in the bmsafemap to await notification of a written bitmap. If
3818 * the operation was canceled we add the segdep to the appropriate
3819 * dependency to free the journal space once the canceling operation
3823 handle_written_jnewblk(jnewblk)
3824 struct jnewblk *jnewblk;
3826 struct bmsafemap *bmsafemap;
3827 struct freefrag *freefrag;
3828 struct freework *freework;
3829 struct jsegdep *jsegdep;
3830 struct newblk *newblk;
3832 /* Grab the jsegdep. */
3833 jsegdep = jnewblk->jn_jsegdep;
3834 jnewblk->jn_jsegdep = NULL;
3835 if (jnewblk->jn_dep == NULL)
3836 panic("handle_written_jnewblk: No dependency for the segdep.");
3837 switch (jnewblk->jn_dep->wk_type) {
3842 * Add the written block to the bmsafemap so it can
3843 * be notified when the bitmap is on disk.
3845 newblk = WK_NEWBLK(jnewblk->jn_dep);
3846 newblk->nb_jnewblk = NULL;
3847 if ((newblk->nb_state & GOINGAWAY) == 0) {
3848 bmsafemap = newblk->nb_bmsafemap;
3849 newblk->nb_state |= ONDEPLIST;
3850 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk,
3853 jwork_insert(&newblk->nb_jwork, jsegdep);
3857 * A newblock being removed by a freefrag when replaced by
3860 freefrag = WK_FREEFRAG(jnewblk->jn_dep);
3861 freefrag->ff_jdep = NULL;
3862 jwork_insert(&freefrag->ff_jwork, jsegdep);
3866 * A direct block was removed by truncate.
3868 freework = WK_FREEWORK(jnewblk->jn_dep);
3869 freework->fw_jnewblk = NULL;
3870 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep);
3873 panic("handle_written_jnewblk: Unknown type %d.",
3874 jnewblk->jn_dep->wk_type);
3876 jnewblk->jn_dep = NULL;
3877 free_jnewblk(jnewblk);
3881 * Cancel a jfreefrag that won't be needed, probably due to colliding with
3882 * an in-flight allocation that has not yet been committed. Divorce us
3883 * from the freefrag and mark it DEPCOMPLETE so that it may be added
3887 cancel_jfreefrag(jfreefrag)
3888 struct jfreefrag *jfreefrag;
3890 struct freefrag *freefrag;
3892 if (jfreefrag->fr_jsegdep) {
3893 free_jsegdep(jfreefrag->fr_jsegdep);
3894 jfreefrag->fr_jsegdep = NULL;
3896 freefrag = jfreefrag->fr_freefrag;
3897 jfreefrag->fr_freefrag = NULL;
3898 free_jfreefrag(jfreefrag);
3899 freefrag->ff_state |= DEPCOMPLETE;
3900 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno);
3904 * Free a jfreefrag when the parent freefrag is rendered obsolete.
3907 free_jfreefrag(jfreefrag)
3908 struct jfreefrag *jfreefrag;
3911 if (jfreefrag->fr_state & INPROGRESS)
3912 WORKLIST_REMOVE(&jfreefrag->fr_list);
3913 else if (jfreefrag->fr_state & ONWORKLIST)
3914 remove_from_journal(&jfreefrag->fr_list);
3915 if (jfreefrag->fr_freefrag != NULL)
3916 panic("free_jfreefrag: Still attached to a freefrag.");
3917 WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
3921 * Called when the journal write for a jfreefrag completes. The parent
3922 * freefrag is added to the worklist if this completes its dependencies.
3925 handle_written_jfreefrag(jfreefrag)
3926 struct jfreefrag *jfreefrag;
3928 struct jsegdep *jsegdep;
3929 struct freefrag *freefrag;
3931 /* Grab the jsegdep. */
3932 jsegdep = jfreefrag->fr_jsegdep;
3933 jfreefrag->fr_jsegdep = NULL;
3934 freefrag = jfreefrag->fr_freefrag;
3935 if (freefrag == NULL)
3936 panic("handle_written_jfreefrag: No freefrag.");
3937 freefrag->ff_state |= DEPCOMPLETE;
3938 freefrag->ff_jdep = NULL;
3939 jwork_insert(&freefrag->ff_jwork, jsegdep);
3940 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
3941 add_to_worklist(&freefrag->ff_list, 0);
3942 jfreefrag->fr_freefrag = NULL;
3943 free_jfreefrag(jfreefrag);
3947 * Called when the journal write for a jfreeblk completes. The jfreeblk
3948 * is removed from the freeblks list of pending journal writes and the
3949 * jsegdep is moved to the freeblks jwork to be completed when all blocks
3950 * have been reclaimed.
3953 handle_written_jblkdep(jblkdep)
3954 struct jblkdep *jblkdep;
3956 struct freeblks *freeblks;
3957 struct jsegdep *jsegdep;
3959 /* Grab the jsegdep. */
3960 jsegdep = jblkdep->jb_jsegdep;
3961 jblkdep->jb_jsegdep = NULL;
3962 freeblks = jblkdep->jb_freeblks;
3963 LIST_REMOVE(jblkdep, jb_deps);
3964 jwork_insert(&freeblks->fb_jwork, jsegdep);
3966 * If the freeblks is all journaled, we can add it to the worklist.
3968 if (LIST_EMPTY(&freeblks->fb_jblkdephd) &&
3969 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
3970 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
3972 free_jblkdep(jblkdep);
3975 static struct jsegdep *
3976 newjsegdep(struct worklist *wk)
3978 struct jsegdep *jsegdep;
3980 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
3981 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
3982 jsegdep->jd_seg = NULL;
3987 static struct jmvref *
3988 newjmvref(dp, ino, oldoff, newoff)
3994 struct jmvref *jmvref;
3996 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
3997 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump));
3998 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
3999 jmvref->jm_parent = dp->i_number;
4000 jmvref->jm_ino = ino;
4001 jmvref->jm_oldoff = oldoff;
4002 jmvref->jm_newoff = newoff;
4008 * Allocate a new jremref that tracks the removal of ip from dp with the
4009 * directory entry offset of diroff. Mark the entry as ATTACHED and
4010 * DEPCOMPLETE as we have all the information required for the journal write
4011 * and the directory has already been removed from the buffer. The caller
4012 * is responsible for linking the jremref into the pagedep and adding it
4013 * to the journal to write. The MKDIR_PARENT flag is set if we're doing
4014 * a DOTDOT addition so handle_workitem_remove() can properly assign
4015 * the jsegdep when we're done.
4017 static struct jremref *
4018 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
4019 off_t diroff, nlink_t nlink)
4021 struct jremref *jremref;
4023 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
4024 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump));
4025 jremref->jr_state = ATTACHED;
4026 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
4028 jremref->jr_dirrem = dirrem;
4034 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
4035 nlink_t nlink, uint16_t mode)
4038 inoref->if_jsegdep = newjsegdep(&inoref->if_list);
4039 inoref->if_diroff = diroff;
4040 inoref->if_ino = ino;
4041 inoref->if_parent = parent;
4042 inoref->if_nlink = nlink;
4043 inoref->if_mode = mode;
4047 * Allocate a new jaddref to track the addition of ino to dp at diroff. The
4048 * directory offset may not be known until later. The caller is responsible
4049 * adding the entry to the journal when this information is available. nlink
4050 * should be the link count prior to the addition and mode is only required
4051 * to have the correct FMT.
4053 static struct jaddref *
4054 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
4057 struct jaddref *jaddref;
4059 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
4060 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump));
4061 jaddref->ja_state = ATTACHED;
4062 jaddref->ja_mkdir = NULL;
4063 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
4069 * Create a new free dependency for a freework. The caller is responsible
4070 * for adjusting the reference count when it has the lock held. The freedep
4071 * will track an outstanding bitmap write that will ultimately clear the
4072 * freework to continue.
4074 static struct freedep *
4075 newfreedep(struct freework *freework)
4077 struct freedep *freedep;
4079 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
4080 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
4081 freedep->fd_freework = freework;
4087 * Free a freedep structure once the buffer it is linked to is written. If
4088 * this is the last reference to the freework schedule it for completion.
4091 free_freedep(freedep)
4092 struct freedep *freedep;
4094 struct freework *freework;
4096 freework = freedep->fd_freework;
4097 freework->fw_freeblks->fb_cgwait--;
4098 if (--freework->fw_ref == 0)
4099 freework_enqueue(freework);
4100 WORKITEM_FREE(freedep, D_FREEDEP);
4104 * Allocate a new freework structure that may be a level in an indirect
4105 * when parent is not NULL or a top level block when it is. The top level
4106 * freework structures are allocated without the per-filesystem lock held
4107 * and before the freeblks is visible outside of softdep_setup_freeblocks().
4109 static struct freework *
4110 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal)
4111 struct ufsmount *ump;
4112 struct freeblks *freeblks;
4113 struct freework *parent;
4120 struct freework *freework;
4122 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
4123 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
4124 freework->fw_state = ATTACHED;
4125 freework->fw_jnewblk = NULL;
4126 freework->fw_freeblks = freeblks;
4127 freework->fw_parent = parent;
4128 freework->fw_lbn = lbn;
4129 freework->fw_blkno = nb;
4130 freework->fw_frags = frags;
4131 freework->fw_indir = NULL;
4132 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR)
4133 ? 0 : NINDIR(ump->um_fs) + 1;
4134 freework->fw_start = freework->fw_off = off;
4136 newjfreeblk(freeblks, lbn, nb, frags);
4137 if (parent == NULL) {
4139 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
4148 * Eliminate a jfreeblk for a block that does not need journaling.
4151 cancel_jfreeblk(freeblks, blkno)
4152 struct freeblks *freeblks;
4155 struct jfreeblk *jfreeblk;
4156 struct jblkdep *jblkdep;
4158 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) {
4159 if (jblkdep->jb_list.wk_type != D_JFREEBLK)
4161 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list);
4162 if (jfreeblk->jf_blkno == blkno)
4165 if (jblkdep == NULL)
4167 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno);
4168 free_jsegdep(jblkdep->jb_jsegdep);
4169 LIST_REMOVE(jblkdep, jb_deps);
4170 WORKITEM_FREE(jfreeblk, D_JFREEBLK);
4174 * Allocate a new jfreeblk to journal top level block pointer when truncating
4175 * a file. The caller must add this to the worklist when the per-filesystem
4178 static struct jfreeblk *
4179 newjfreeblk(freeblks, lbn, blkno, frags)
4180 struct freeblks *freeblks;
4185 struct jfreeblk *jfreeblk;
4187 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
4188 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK,
4189 freeblks->fb_list.wk_mp);
4190 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list);
4191 jfreeblk->jf_dep.jb_freeblks = freeblks;
4192 jfreeblk->jf_ino = freeblks->fb_inum;
4193 jfreeblk->jf_lbn = lbn;
4194 jfreeblk->jf_blkno = blkno;
4195 jfreeblk->jf_frags = frags;
4196 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps);
4202 * The journal is only prepared to handle full-size block numbers, so we
4203 * have to adjust the record to reflect the change to a full-size block.
4204 * For example, suppose we have a block made up of fragments 8-15 and
4205 * want to free its last two fragments. We are given a request that says:
4206 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0
4207 * where frags are the number of fragments to free and oldfrags are the
4208 * number of fragments to keep. To block align it, we have to change it to
4209 * have a valid full-size blkno, so it becomes:
4210 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6
4213 adjust_newfreework(freeblks, frag_offset)
4214 struct freeblks *freeblks;
4217 struct jfreeblk *jfreeblk;
4219 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL &&
4220 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK),
4221 ("adjust_newfreework: Missing freeblks dependency"));
4223 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd));
4224 jfreeblk->jf_blkno -= frag_offset;
4225 jfreeblk->jf_frags += frag_offset;
4229 * Allocate a new jtrunc to track a partial truncation.
4231 static struct jtrunc *
4232 newjtrunc(freeblks, size, extsize)
4233 struct freeblks *freeblks;
4237 struct jtrunc *jtrunc;
4239 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
4240 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC,
4241 freeblks->fb_list.wk_mp);
4242 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list);
4243 jtrunc->jt_dep.jb_freeblks = freeblks;
4244 jtrunc->jt_ino = freeblks->fb_inum;
4245 jtrunc->jt_size = size;
4246 jtrunc->jt_extsize = extsize;
4247 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps);
4253 * If we're canceling a new bitmap we have to search for another ref
4254 * to move into the bmsafemap dep. This might be better expressed
4255 * with another structure.
4258 move_newblock_dep(jaddref, inodedep)
4259 struct jaddref *jaddref;
4260 struct inodedep *inodedep;
4262 struct inoref *inoref;
4263 struct jaddref *jaddrefn;
4266 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4267 inoref = TAILQ_NEXT(inoref, if_deps)) {
4268 if ((jaddref->ja_state & NEWBLOCK) &&
4269 inoref->if_list.wk_type == D_JADDREF) {
4270 jaddrefn = (struct jaddref *)inoref;
4274 if (jaddrefn == NULL)
4276 jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
4277 jaddrefn->ja_state |= jaddref->ja_state &
4278 (ATTACHED | UNDONE | NEWBLOCK);
4279 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
4280 jaddref->ja_state |= ATTACHED;
4281 LIST_REMOVE(jaddref, ja_bmdeps);
4282 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
4287 * Cancel a jaddref either before it has been written or while it is being
4288 * written. This happens when a link is removed before the add reaches
4289 * the disk. The jaddref dependency is kept linked into the bmsafemap
4290 * and inode to prevent the link count or bitmap from reaching the disk
4291 * until handle_workitem_remove() re-adjusts the counts and bitmaps as
4294 * Returns 1 if the canceled addref requires journaling of the remove and
4298 cancel_jaddref(jaddref, inodedep, wkhd)
4299 struct jaddref *jaddref;
4300 struct inodedep *inodedep;
4301 struct workhead *wkhd;
4303 struct inoref *inoref;
4304 struct jsegdep *jsegdep;
4307 KASSERT((jaddref->ja_state & COMPLETE) == 0,
4308 ("cancel_jaddref: Canceling complete jaddref"));
4309 if (jaddref->ja_state & (INPROGRESS | COMPLETE))
4313 if (inodedep == NULL)
4314 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
4316 panic("cancel_jaddref: Lost inodedep");
4318 * We must adjust the nlink of any reference operation that follows
4319 * us so that it is consistent with the in-memory reference. This
4320 * ensures that inode nlink rollbacks always have the correct link.
4323 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
4324 inoref = TAILQ_NEXT(inoref, if_deps)) {
4325 if (inoref->if_state & GOINGAWAY)
4330 jsegdep = inoref_jseg(&jaddref->ja_ref);
4331 if (jaddref->ja_state & NEWBLOCK)
4332 move_newblock_dep(jaddref, inodedep);
4333 wake_worklist(&jaddref->ja_list);
4334 jaddref->ja_mkdir = NULL;
4335 if (jaddref->ja_state & INPROGRESS) {
4336 jaddref->ja_state &= ~INPROGRESS;
4337 WORKLIST_REMOVE(&jaddref->ja_list);
4338 jwork_insert(wkhd, jsegdep);
4340 free_jsegdep(jsegdep);
4341 if (jaddref->ja_state & DEPCOMPLETE)
4342 remove_from_journal(&jaddref->ja_list);
4344 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE);
4346 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
4347 * can arrange for them to be freed with the bitmap. Otherwise we
4348 * no longer need this addref attached to the inoreflst and it
4349 * will incorrectly adjust nlink if we leave it.
4351 if ((jaddref->ja_state & NEWBLOCK) == 0) {
4352 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
4354 jaddref->ja_state |= COMPLETE;
4355 free_jaddref(jaddref);
4359 * Leave the head of the list for jsegdeps for fast merging.
4361 if (LIST_FIRST(wkhd) != NULL) {
4362 jaddref->ja_state |= ONWORKLIST;
4363 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
4365 WORKLIST_INSERT(wkhd, &jaddref->ja_list);
4371 * Attempt to free a jaddref structure when some work completes. This
4372 * should only succeed once the entry is written and all dependencies have
4376 free_jaddref(jaddref)
4377 struct jaddref *jaddref;
4380 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
4382 if (jaddref->ja_ref.if_jsegdep)
4383 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
4384 jaddref, jaddref->ja_state);
4385 if (jaddref->ja_state & NEWBLOCK)
4386 LIST_REMOVE(jaddref, ja_bmdeps);
4387 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST))
4388 panic("free_jaddref: Bad state %p(0x%X)",
4389 jaddref, jaddref->ja_state);
4390 if (jaddref->ja_mkdir != NULL)
4391 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
4392 WORKITEM_FREE(jaddref, D_JADDREF);
4396 * Free a jremref structure once it has been written or discarded.
4399 free_jremref(jremref)
4400 struct jremref *jremref;
4403 if (jremref->jr_ref.if_jsegdep)
4404 free_jsegdep(jremref->jr_ref.if_jsegdep);
4405 if (jremref->jr_state & INPROGRESS)
4406 panic("free_jremref: IO still pending");
4407 WORKITEM_FREE(jremref, D_JREMREF);
4411 * Free a jnewblk structure.
4414 free_jnewblk(jnewblk)
4415 struct jnewblk *jnewblk;
4418 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
4420 LIST_REMOVE(jnewblk, jn_deps);
4421 if (jnewblk->jn_dep != NULL)
4422 panic("free_jnewblk: Dependency still attached.");
4423 WORKITEM_FREE(jnewblk, D_JNEWBLK);
4427 * Cancel a jnewblk which has been been made redundant by frag extension.
4430 cancel_jnewblk(jnewblk, wkhd)
4431 struct jnewblk *jnewblk;
4432 struct workhead *wkhd;
4434 struct jsegdep *jsegdep;
4436 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno);
4437 jsegdep = jnewblk->jn_jsegdep;
4438 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL)
4439 panic("cancel_jnewblk: Invalid state");
4440 jnewblk->jn_jsegdep = NULL;
4441 jnewblk->jn_dep = NULL;
4442 jnewblk->jn_state |= GOINGAWAY;
4443 if (jnewblk->jn_state & INPROGRESS) {
4444 jnewblk->jn_state &= ~INPROGRESS;
4445 WORKLIST_REMOVE(&jnewblk->jn_list);
4446 jwork_insert(wkhd, jsegdep);
4448 free_jsegdep(jsegdep);
4449 remove_from_journal(&jnewblk->jn_list);
4451 wake_worklist(&jnewblk->jn_list);
4452 WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
4456 free_jblkdep(jblkdep)
4457 struct jblkdep *jblkdep;
4460 if (jblkdep->jb_list.wk_type == D_JFREEBLK)
4461 WORKITEM_FREE(jblkdep, D_JFREEBLK);
4462 else if (jblkdep->jb_list.wk_type == D_JTRUNC)
4463 WORKITEM_FREE(jblkdep, D_JTRUNC);
4465 panic("free_jblkdep: Unexpected type %s",
4466 TYPENAME(jblkdep->jb_list.wk_type));
4470 * Free a single jseg once it is no longer referenced in memory or on
4471 * disk. Reclaim journal blocks and dependencies waiting for the segment
4475 free_jseg(jseg, jblocks)
4477 struct jblocks *jblocks;
4479 struct freework *freework;
4482 * Free freework structures that were lingering to indicate freed
4483 * indirect blocks that forced journal write ordering on reallocate.
4485 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL)
4486 indirblk_remove(freework);
4487 if (jblocks->jb_oldestseg == jseg)
4488 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next);
4489 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
4490 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
4491 KASSERT(LIST_EMPTY(&jseg->js_entries),
4492 ("free_jseg: Freed jseg has valid entries."));
4493 WORKITEM_FREE(jseg, D_JSEG);
4497 * Free all jsegs that meet the criteria for being reclaimed and update
4502 struct jblocks *jblocks;
4507 * Free only those jsegs which have none allocated before them to
4508 * preserve the journal space ordering.
4510 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
4512 * Only reclaim space when nothing depends on this journal
4513 * set and another set has written that it is no longer
4516 if (jseg->js_refs != 0) {
4517 jblocks->jb_oldestseg = jseg;
4520 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE)
4522 if (jseg->js_seq > jblocks->jb_oldestwrseq)
4525 * We can free jsegs that didn't write entries when
4526 * oldestwrseq == js_seq.
4528 if (jseg->js_seq == jblocks->jb_oldestwrseq &&
4531 free_jseg(jseg, jblocks);
4534 * If we exited the loop above we still must discover the
4535 * oldest valid segment.
4538 for (jseg = jblocks->jb_oldestseg; jseg != NULL;
4539 jseg = TAILQ_NEXT(jseg, js_next))
4540 if (jseg->js_refs != 0)
4542 jblocks->jb_oldestseg = jseg;
4544 * The journal has no valid records but some jsegs may still be
4545 * waiting on oldestwrseq to advance. We force a small record
4546 * out to permit these lingering records to be reclaimed.
4548 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs))
4549 jblocks->jb_needseg = 1;
4553 * Release one reference to a jseg and free it if the count reaches 0. This
4554 * should eventually reclaim journal space as well.
4561 KASSERT(jseg->js_refs > 0,
4562 ("free_jseg: Invalid refcnt %d", jseg->js_refs));
4563 if (--jseg->js_refs != 0)
4565 free_jsegs(jseg->js_jblocks);
4569 * Release a jsegdep and decrement the jseg count.
4572 free_jsegdep(jsegdep)
4573 struct jsegdep *jsegdep;
4576 if (jsegdep->jd_seg)
4577 rele_jseg(jsegdep->jd_seg);
4578 WORKITEM_FREE(jsegdep, D_JSEGDEP);
4582 * Wait for a journal item to make it to disk. Initiate journal processing
4587 struct worklist *wk;
4591 LOCK_OWNED(VFSTOUFS(wk->wk_mp));
4593 * Blocking journal waits cause slow synchronous behavior. Record
4594 * stats on the frequency of these blocking operations.
4596 if (waitfor == MNT_WAIT) {
4597 stat_journal_wait++;
4598 switch (wk->wk_type) {
4601 stat_jwait_filepage++;
4605 stat_jwait_freeblks++;
4608 stat_jwait_newblk++;
4618 * If IO has not started we process the journal. We can't mark the
4619 * worklist item as IOWAITING because we drop the lock while
4620 * processing the journal and the worklist entry may be freed after
4621 * this point. The caller may call back in and re-issue the request.
4623 if ((wk->wk_state & INPROGRESS) == 0) {
4624 softdep_process_journal(wk->wk_mp, wk, waitfor);
4625 if (waitfor != MNT_WAIT)
4629 if (waitfor != MNT_WAIT)
4631 wait_worklist(wk, "jwait");
4636 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
4637 * appropriate. This is a convenience function to reduce duplicate code
4638 * for the setup and revert functions below.
4640 static struct inodedep *
4641 inodedep_lookup_ip(ip)
4644 struct inodedep *inodedep;
4646 KASSERT(ip->i_nlink >= ip->i_effnlink,
4647 ("inodedep_lookup_ip: bad delta"));
4648 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
4650 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
4651 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
4657 * Called prior to creating a new inode and linking it to a directory. The
4658 * jaddref structure must already be allocated by softdep_setup_inomapdep
4659 * and it is discovered here so we can initialize the mode and update
4663 softdep_setup_create(dp, ip)
4667 struct inodedep *inodedep;
4668 struct jaddref *jaddref;
4671 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4672 ("softdep_setup_create called on non-softdep filesystem"));
4673 KASSERT(ip->i_nlink == 1,
4674 ("softdep_setup_create: Invalid link count."));
4676 ACQUIRE_LOCK(dp->i_ump);
4677 inodedep = inodedep_lookup_ip(ip);
4678 if (DOINGSUJ(dvp)) {
4679 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4681 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
4682 ("softdep_setup_create: No addref structure present."));
4684 softdep_prelink(dvp, NULL);
4685 FREE_LOCK(dp->i_ump);
4689 * Create a jaddref structure to track the addition of a DOTDOT link when
4690 * we are reparenting an inode as part of a rename. This jaddref will be
4691 * found by softdep_setup_directory_change. Adjusts nlinkdelta for
4692 * non-journaling softdep.
4695 softdep_setup_dotdot_link(dp, ip)
4699 struct inodedep *inodedep;
4700 struct jaddref *jaddref;
4703 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4704 ("softdep_setup_dotdot_link called on non-softdep filesystem"));
4708 * We don't set MKDIR_PARENT as this is not tied to a mkdir and
4709 * is used as a normal link would be.
4712 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4713 dp->i_effnlink - 1, dp->i_mode);
4714 ACQUIRE_LOCK(dp->i_ump);
4715 inodedep = inodedep_lookup_ip(dp);
4717 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4719 softdep_prelink(dvp, ITOV(ip));
4720 FREE_LOCK(dp->i_ump);
4724 * Create a jaddref structure to track a new link to an inode. The directory
4725 * offset is not known until softdep_setup_directory_add or
4726 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling
4730 softdep_setup_link(dp, ip)
4734 struct inodedep *inodedep;
4735 struct jaddref *jaddref;
4738 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4739 ("softdep_setup_link called on non-softdep filesystem"));
4743 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
4745 ACQUIRE_LOCK(dp->i_ump);
4746 inodedep = inodedep_lookup_ip(ip);
4748 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4750 softdep_prelink(dvp, ITOV(ip));
4751 FREE_LOCK(dp->i_ump);
4755 * Called to create the jaddref structures to track . and .. references as
4756 * well as lookup and further initialize the incomplete jaddref created
4757 * by softdep_setup_inomapdep when the inode was allocated. Adjusts
4758 * nlinkdelta for non-journaling softdep.
4761 softdep_setup_mkdir(dp, ip)
4765 struct inodedep *inodedep;
4766 struct jaddref *dotdotaddref;
4767 struct jaddref *dotaddref;
4768 struct jaddref *jaddref;
4771 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4772 ("softdep_setup_mkdir called on non-softdep filesystem"));
4774 dotaddref = dotdotaddref = NULL;
4775 if (DOINGSUJ(dvp)) {
4776 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
4778 dotaddref->ja_state |= MKDIR_BODY;
4779 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
4780 dp->i_effnlink - 1, dp->i_mode);
4781 dotdotaddref->ja_state |= MKDIR_PARENT;
4783 ACQUIRE_LOCK(dp->i_ump);
4784 inodedep = inodedep_lookup_ip(ip);
4785 if (DOINGSUJ(dvp)) {
4786 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4788 KASSERT(jaddref != NULL,
4789 ("softdep_setup_mkdir: No addref structure present."));
4790 KASSERT(jaddref->ja_parent == dp->i_number,
4791 ("softdep_setup_mkdir: bad parent %ju",
4792 (uintmax_t)jaddref->ja_parent));
4793 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
4796 inodedep = inodedep_lookup_ip(dp);
4798 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
4799 &dotdotaddref->ja_ref, if_deps);
4800 softdep_prelink(ITOV(dp), NULL);
4801 FREE_LOCK(dp->i_ump);
4805 * Called to track nlinkdelta of the inode and parent directories prior to
4806 * unlinking a directory.
4809 softdep_setup_rmdir(dp, ip)
4815 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4816 ("softdep_setup_rmdir called on non-softdep filesystem"));
4818 ACQUIRE_LOCK(dp->i_ump);
4819 (void) inodedep_lookup_ip(ip);
4820 (void) inodedep_lookup_ip(dp);
4821 softdep_prelink(dvp, ITOV(ip));
4822 FREE_LOCK(dp->i_ump);
4826 * Called to track nlinkdelta of the inode and parent directories prior to
4830 softdep_setup_unlink(dp, ip)
4836 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4837 ("softdep_setup_unlink called on non-softdep filesystem"));
4839 ACQUIRE_LOCK(dp->i_ump);
4840 (void) inodedep_lookup_ip(ip);
4841 (void) inodedep_lookup_ip(dp);
4842 softdep_prelink(dvp, ITOV(ip));
4843 FREE_LOCK(dp->i_ump);
4847 * Called to release the journal structures created by a failed non-directory
4848 * creation. Adjusts nlinkdelta for non-journaling softdep.
4851 softdep_revert_create(dp, ip)
4855 struct inodedep *inodedep;
4856 struct jaddref *jaddref;
4859 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4860 ("softdep_revert_create called on non-softdep filesystem"));
4862 ACQUIRE_LOCK(dp->i_ump);
4863 inodedep = inodedep_lookup_ip(ip);
4864 if (DOINGSUJ(dvp)) {
4865 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4867 KASSERT(jaddref->ja_parent == dp->i_number,
4868 ("softdep_revert_create: addref parent mismatch"));
4869 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4871 FREE_LOCK(dp->i_ump);
4875 * Called to release the journal structures created by a failed link
4876 * addition. Adjusts nlinkdelta for non-journaling softdep.
4879 softdep_revert_link(dp, ip)
4883 struct inodedep *inodedep;
4884 struct jaddref *jaddref;
4887 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4888 ("softdep_revert_link called on non-softdep filesystem"));
4890 ACQUIRE_LOCK(dp->i_ump);
4891 inodedep = inodedep_lookup_ip(ip);
4892 if (DOINGSUJ(dvp)) {
4893 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4895 KASSERT(jaddref->ja_parent == dp->i_number,
4896 ("softdep_revert_link: addref parent mismatch"));
4897 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4899 FREE_LOCK(dp->i_ump);
4903 * Called to release the journal structures created by a failed mkdir
4904 * attempt. Adjusts nlinkdelta for non-journaling softdep.
4907 softdep_revert_mkdir(dp, ip)
4911 struct inodedep *inodedep;
4912 struct jaddref *jaddref;
4913 struct jaddref *dotaddref;
4916 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4917 ("softdep_revert_mkdir called on non-softdep filesystem"));
4920 ACQUIRE_LOCK(dp->i_ump);
4921 inodedep = inodedep_lookup_ip(dp);
4922 if (DOINGSUJ(dvp)) {
4923 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4925 KASSERT(jaddref->ja_parent == ip->i_number,
4926 ("softdep_revert_mkdir: dotdot addref parent mismatch"));
4927 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4929 inodedep = inodedep_lookup_ip(ip);
4930 if (DOINGSUJ(dvp)) {
4931 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4933 KASSERT(jaddref->ja_parent == dp->i_number,
4934 ("softdep_revert_mkdir: addref parent mismatch"));
4935 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
4936 inoreflst, if_deps);
4937 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4938 KASSERT(dotaddref->ja_parent == ip->i_number,
4939 ("softdep_revert_mkdir: dot addref parent mismatch"));
4940 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait);
4942 FREE_LOCK(dp->i_ump);
4946 * Called to correct nlinkdelta after a failed rmdir.
4949 softdep_revert_rmdir(dp, ip)
4954 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0,
4955 ("softdep_revert_rmdir called on non-softdep filesystem"));
4956 ACQUIRE_LOCK(dp->i_ump);
4957 (void) inodedep_lookup_ip(ip);
4958 (void) inodedep_lookup_ip(dp);
4959 FREE_LOCK(dp->i_ump);
4963 * Protecting the freemaps (or bitmaps).
4965 * To eliminate the need to execute fsck before mounting a filesystem
4966 * after a power failure, one must (conservatively) guarantee that the
4967 * on-disk copy of the bitmaps never indicate that a live inode or block is
4968 * free. So, when a block or inode is allocated, the bitmap should be
4969 * updated (on disk) before any new pointers. When a block or inode is
4970 * freed, the bitmap should not be updated until all pointers have been
4971 * reset. The latter dependency is handled by the delayed de-allocation
4972 * approach described below for block and inode de-allocation. The former
4973 * dependency is handled by calling the following procedure when a block or
4974 * inode is allocated. When an inode is allocated an "inodedep" is created
4975 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
4976 * Each "inodedep" is also inserted into the hash indexing structure so
4977 * that any additional link additions can be made dependent on the inode
4980 * The ufs filesystem maintains a number of free block counts (e.g., per
4981 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
4982 * in addition to the bitmaps. These counts are used to improve efficiency
4983 * during allocation and therefore must be consistent with the bitmaps.
4984 * There is no convenient way to guarantee post-crash consistency of these
4985 * counts with simple update ordering, for two main reasons: (1) The counts
4986 * and bitmaps for a single cylinder group block are not in the same disk
4987 * sector. If a disk write is interrupted (e.g., by power failure), one may
4988 * be written and the other not. (2) Some of the counts are located in the
4989 * superblock rather than the cylinder group block. So, we focus our soft
4990 * updates implementation on protecting the bitmaps. When mounting a
4991 * filesystem, we recompute the auxiliary counts from the bitmaps.
4995 * Called just after updating the cylinder group block to allocate an inode.
4998 softdep_setup_inomapdep(bp, ip, newinum, mode)
4999 struct buf *bp; /* buffer for cylgroup block with inode map */
5000 struct inode *ip; /* inode related to allocation */
5001 ino_t newinum; /* new inode number being allocated */
5004 struct inodedep *inodedep;
5005 struct bmsafemap *bmsafemap;
5006 struct jaddref *jaddref;
5010 mp = UFSTOVFS(ip->i_ump);
5011 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5012 ("softdep_setup_inomapdep called on non-softdep filesystem"));
5013 fs = ip->i_ump->um_fs;
5017 * Allocate the journal reference add structure so that the bitmap
5018 * can be dependent on it.
5020 if (MOUNTEDSUJ(mp)) {
5021 jaddref = newjaddref(ip, newinum, 0, 0, mode);
5022 jaddref->ja_state |= NEWBLOCK;
5026 * Create a dependency for the newly allocated inode.
5027 * Panic if it already exists as something is seriously wrong.
5028 * Otherwise add it to the dependency list for the buffer holding
5029 * the cylinder group map from which it was allocated.
5031 * We have to preallocate a bmsafemap entry in case it is needed
5032 * in bmsafemap_lookup since once we allocate the inodedep, we
5033 * have to finish initializing it before we can FREE_LOCK().
5034 * By preallocating, we avoid FREE_LOCK() while doing a malloc
5035 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before
5036 * creating the inodedep as it can be freed during the time
5037 * that we FREE_LOCK() while allocating the inodedep. We must
5038 * call workitem_alloc() before entering the locked section as
5039 * it also acquires the lock and we must avoid trying doing so
5042 bmsafemap = malloc(sizeof(struct bmsafemap),
5043 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5044 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5045 ACQUIRE_LOCK(ip->i_ump);
5046 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep)))
5047 panic("softdep_setup_inomapdep: dependency %p for new"
5048 "inode already exists", inodedep);
5049 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap);
5051 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
5052 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
5055 inodedep->id_state |= ONDEPLIST;
5056 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
5058 inodedep->id_bmsafemap = bmsafemap;
5059 inodedep->id_state &= ~DEPCOMPLETE;
5060 FREE_LOCK(ip->i_ump);
5064 * Called just after updating the cylinder group block to
5065 * allocate block or fragment.
5068 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
5069 struct buf *bp; /* buffer for cylgroup block with block map */
5070 struct mount *mp; /* filesystem doing allocation */
5071 ufs2_daddr_t newblkno; /* number of newly allocated block */
5072 int frags; /* Number of fragments. */
5073 int oldfrags; /* Previous number of fragments for extend. */
5075 struct newblk *newblk;
5076 struct bmsafemap *bmsafemap;
5077 struct jnewblk *jnewblk;
5078 struct ufsmount *ump;
5081 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5082 ("softdep_setup_blkmapdep called on non-softdep filesystem"));
5087 * Create a dependency for the newly allocated block.
5088 * Add it to the dependency list for the buffer holding
5089 * the cylinder group map from which it was allocated.
5091 if (MOUNTEDSUJ(mp)) {
5092 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
5093 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
5094 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
5095 jnewblk->jn_state = ATTACHED;
5096 jnewblk->jn_blkno = newblkno;
5097 jnewblk->jn_frags = frags;
5098 jnewblk->jn_oldfrags = oldfrags;
5106 cgp = (struct cg *)bp->b_data;
5107 blksfree = cg_blksfree(cgp);
5108 bno = dtogd(fs, jnewblk->jn_blkno);
5109 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
5111 if (isset(blksfree, bno + i))
5112 panic("softdep_setup_blkmapdep: "
5113 "free fragment %d from %d-%d "
5114 "state 0x%X dep %p", i,
5115 jnewblk->jn_oldfrags,
5125 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d",
5126 newblkno, frags, oldfrags);
5128 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
5129 panic("softdep_setup_blkmapdep: found block");
5130 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
5131 dtog(fs, newblkno), NULL);
5133 jnewblk->jn_dep = (struct worklist *)newblk;
5134 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
5136 newblk->nb_state |= ONDEPLIST;
5137 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
5139 newblk->nb_bmsafemap = bmsafemap;
5140 newblk->nb_jnewblk = jnewblk;
5144 #define BMSAFEMAP_HASH(ump, cg) \
5145 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size])
5148 bmsafemap_find(bmsafemaphd, cg, bmsafemapp)
5149 struct bmsafemap_hashhead *bmsafemaphd;
5151 struct bmsafemap **bmsafemapp;
5153 struct bmsafemap *bmsafemap;
5155 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
5156 if (bmsafemap->sm_cg == cg)
5159 *bmsafemapp = bmsafemap;
5168 * Find the bmsafemap associated with a cylinder group buffer.
5169 * If none exists, create one. The buffer must be locked when
5170 * this routine is called and this routine must be called with
5171 * the softdep lock held. To avoid giving up the lock while
5172 * allocating a new bmsafemap, a preallocated bmsafemap may be
5173 * provided. If it is provided but not needed, it is freed.
5175 static struct bmsafemap *
5176 bmsafemap_lookup(mp, bp, cg, newbmsafemap)
5180 struct bmsafemap *newbmsafemap;
5182 struct bmsafemap_hashhead *bmsafemaphd;
5183 struct bmsafemap *bmsafemap, *collision;
5184 struct worklist *wk;
5185 struct ufsmount *ump;
5189 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer"));
5190 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5191 if (wk->wk_type == D_BMSAFEMAP) {
5193 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5194 return (WK_BMSAFEMAP(wk));
5197 bmsafemaphd = BMSAFEMAP_HASH(ump, cg);
5198 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) {
5200 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
5204 bmsafemap = newbmsafemap;
5207 bmsafemap = malloc(sizeof(struct bmsafemap),
5208 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
5209 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
5212 bmsafemap->sm_buf = bp;
5213 LIST_INIT(&bmsafemap->sm_inodedephd);
5214 LIST_INIT(&bmsafemap->sm_inodedepwr);
5215 LIST_INIT(&bmsafemap->sm_newblkhd);
5216 LIST_INIT(&bmsafemap->sm_newblkwr);
5217 LIST_INIT(&bmsafemap->sm_jaddrefhd);
5218 LIST_INIT(&bmsafemap->sm_jnewblkhd);
5219 LIST_INIT(&bmsafemap->sm_freehd);
5220 LIST_INIT(&bmsafemap->sm_freewr);
5221 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) {
5222 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
5225 bmsafemap->sm_cg = cg;
5226 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
5227 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
5228 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
5233 * Direct block allocation dependencies.
5235 * When a new block is allocated, the corresponding disk locations must be
5236 * initialized (with zeros or new data) before the on-disk inode points to
5237 * them. Also, the freemap from which the block was allocated must be
5238 * updated (on disk) before the inode's pointer. These two dependencies are
5239 * independent of each other and are needed for all file blocks and indirect
5240 * blocks that are pointed to directly by the inode. Just before the
5241 * "in-core" version of the inode is updated with a newly allocated block
5242 * number, a procedure (below) is called to setup allocation dependency
5243 * structures. These structures are removed when the corresponding
5244 * dependencies are satisfied or when the block allocation becomes obsolete
5245 * (i.e., the file is deleted, the block is de-allocated, or the block is a
5246 * fragment that gets upgraded). All of these cases are handled in
5247 * procedures described later.
5249 * When a file extension causes a fragment to be upgraded, either to a larger
5250 * fragment or to a full block, the on-disk location may change (if the
5251 * previous fragment could not simply be extended). In this case, the old
5252 * fragment must be de-allocated, but not until after the inode's pointer has
5253 * been updated. In most cases, this is handled by later procedures, which
5254 * will construct a "freefrag" structure to be added to the workitem queue
5255 * when the inode update is complete (or obsolete). The main exception to
5256 * this is when an allocation occurs while a pending allocation dependency
5257 * (for the same block pointer) remains. This case is handled in the main
5258 * allocation dependency setup procedure by immediately freeing the
5259 * unreferenced fragments.
5262 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5263 struct inode *ip; /* inode to which block is being added */
5264 ufs_lbn_t off; /* block pointer within inode */
5265 ufs2_daddr_t newblkno; /* disk block number being added */
5266 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */
5267 long newsize; /* size of new block */
5268 long oldsize; /* size of new block */
5269 struct buf *bp; /* bp for allocated block */
5271 struct allocdirect *adp, *oldadp;
5272 struct allocdirectlst *adphead;
5273 struct freefrag *freefrag;
5274 struct inodedep *inodedep;
5275 struct pagedep *pagedep;
5276 struct jnewblk *jnewblk;
5277 struct newblk *newblk;
5282 mp = UFSTOVFS(ip->i_ump);
5283 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5284 ("softdep_setup_allocdirect called on non-softdep filesystem"));
5285 if (oldblkno && oldblkno != newblkno)
5286 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5291 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd "
5292 "off %jd newsize %ld oldsize %d",
5293 ip->i_number, newblkno, oldblkno, off, newsize, oldsize);
5294 ACQUIRE_LOCK(ip->i_ump);
5295 if (off >= NDADDR) {
5297 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
5299 /* allocating an indirect block */
5301 panic("softdep_setup_allocdirect: non-zero indir");
5304 panic("softdep_setup_allocdirect: lbn %jd != off %jd",
5307 * Allocating a direct block.
5309 * If we are allocating a directory block, then we must
5310 * allocate an associated pagedep to track additions and
5313 if ((ip->i_mode & IFMT) == IFDIR)
5314 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC,
5317 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5318 panic("softdep_setup_allocdirect: lost block");
5319 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5320 ("softdep_setup_allocdirect: newblk already initialized"));
5322 * Convert the newblk to an allocdirect.
5324 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5325 adp = (struct allocdirect *)newblk;
5326 newblk->nb_freefrag = freefrag;
5327 adp->ad_offset = off;
5328 adp->ad_oldblkno = oldblkno;
5329 adp->ad_newsize = newsize;
5330 adp->ad_oldsize = oldsize;
5333 * Finish initializing the journal.
5335 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5336 jnewblk->jn_ino = ip->i_number;
5337 jnewblk->jn_lbn = lbn;
5338 add_to_journal(&jnewblk->jn_list);
5340 if (freefrag && freefrag->ff_jdep != NULL &&
5341 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5342 add_to_journal(freefrag->ff_jdep);
5343 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5344 adp->ad_inodedep = inodedep;
5346 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5348 * The list of allocdirects must be kept in sorted and ascending
5349 * order so that the rollback routines can quickly determine the
5350 * first uncommitted block (the size of the file stored on disk
5351 * ends at the end of the lowest committed fragment, or if there
5352 * are no fragments, at the end of the highest committed block).
5353 * Since files generally grow, the typical case is that the new
5354 * block is to be added at the end of the list. We speed this
5355 * special case by checking against the last allocdirect in the
5356 * list before laboriously traversing the list looking for the
5359 adphead = &inodedep->id_newinoupdt;
5360 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5361 if (oldadp == NULL || oldadp->ad_offset <= off) {
5362 /* insert at end of list */
5363 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5364 if (oldadp != NULL && oldadp->ad_offset == off)
5365 allocdirect_merge(adphead, adp, oldadp);
5366 FREE_LOCK(ip->i_ump);
5369 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5370 if (oldadp->ad_offset >= off)
5374 panic("softdep_setup_allocdirect: lost entry");
5375 /* insert in middle of list */
5376 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5377 if (oldadp->ad_offset == off)
5378 allocdirect_merge(adphead, adp, oldadp);
5380 FREE_LOCK(ip->i_ump);
5384 * Merge a newer and older journal record to be stored either in a
5385 * newblock or freefrag. This handles aggregating journal records for
5386 * fragment allocation into a second record as well as replacing a
5387 * journal free with an aborted journal allocation. A segment for the
5388 * oldest record will be placed on wkhd if it has been written. If not
5389 * the segment for the newer record will suffice.
5391 static struct worklist *
5392 jnewblk_merge(new, old, wkhd)
5393 struct worklist *new;
5394 struct worklist *old;
5395 struct workhead *wkhd;
5397 struct jnewblk *njnewblk;
5398 struct jnewblk *jnewblk;
5400 /* Handle NULLs to simplify callers. */
5405 /* Replace a jfreefrag with a jnewblk. */
5406 if (new->wk_type == D_JFREEFRAG) {
5407 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno)
5408 panic("jnewblk_merge: blkno mismatch: %p, %p",
5410 cancel_jfreefrag(WK_JFREEFRAG(new));
5413 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK)
5414 panic("jnewblk_merge: Bad type: old %d new %d\n",
5415 old->wk_type, new->wk_type);
5417 * Handle merging of two jnewblk records that describe
5418 * different sets of fragments in the same block.
5420 jnewblk = WK_JNEWBLK(old);
5421 njnewblk = WK_JNEWBLK(new);
5422 if (jnewblk->jn_blkno != njnewblk->jn_blkno)
5423 panic("jnewblk_merge: Merging disparate blocks.");
5425 * The record may be rolled back in the cg.
5427 if (jnewblk->jn_state & UNDONE) {
5428 jnewblk->jn_state &= ~UNDONE;
5429 njnewblk->jn_state |= UNDONE;
5430 njnewblk->jn_state &= ~ATTACHED;
5433 * We modify the newer addref and free the older so that if neither
5434 * has been written the most up-to-date copy will be on disk. If
5435 * both have been written but rolled back we only temporarily need
5436 * one of them to fix the bits when the cg write completes.
5438 jnewblk->jn_state |= ATTACHED | COMPLETE;
5439 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
5440 cancel_jnewblk(jnewblk, wkhd);
5441 WORKLIST_REMOVE(&jnewblk->jn_list);
5442 free_jnewblk(jnewblk);
5447 * Replace an old allocdirect dependency with a newer one.
5448 * This routine must be called with splbio interrupts blocked.
5451 allocdirect_merge(adphead, newadp, oldadp)
5452 struct allocdirectlst *adphead; /* head of list holding allocdirects */
5453 struct allocdirect *newadp; /* allocdirect being added */
5454 struct allocdirect *oldadp; /* existing allocdirect being checked */
5456 struct worklist *wk;
5457 struct freefrag *freefrag;
5460 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp));
5461 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
5462 newadp->ad_oldsize != oldadp->ad_newsize ||
5463 newadp->ad_offset >= NDADDR)
5464 panic("%s %jd != new %jd || old size %ld != new %ld",
5465 "allocdirect_merge: old blkno",
5466 (intmax_t)newadp->ad_oldblkno,
5467 (intmax_t)oldadp->ad_newblkno,
5468 newadp->ad_oldsize, oldadp->ad_newsize);
5469 newadp->ad_oldblkno = oldadp->ad_oldblkno;
5470 newadp->ad_oldsize = oldadp->ad_oldsize;
5472 * If the old dependency had a fragment to free or had never
5473 * previously had a block allocated, then the new dependency
5474 * can immediately post its freefrag and adopt the old freefrag.
5475 * This action is done by swapping the freefrag dependencies.
5476 * The new dependency gains the old one's freefrag, and the
5477 * old one gets the new one and then immediately puts it on
5478 * the worklist when it is freed by free_newblk. It is
5479 * not possible to do this swap when the old dependency had a
5480 * non-zero size but no previous fragment to free. This condition
5481 * arises when the new block is an extension of the old block.
5482 * Here, the first part of the fragment allocated to the new
5483 * dependency is part of the block currently claimed on disk by
5484 * the old dependency, so cannot legitimately be freed until the
5485 * conditions for the new dependency are fulfilled.
5487 freefrag = newadp->ad_freefrag;
5488 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
5489 newadp->ad_freefrag = oldadp->ad_freefrag;
5490 oldadp->ad_freefrag = freefrag;
5493 * If we are tracking a new directory-block allocation,
5494 * move it from the old allocdirect to the new allocdirect.
5496 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
5497 WORKLIST_REMOVE(wk);
5498 if (!LIST_EMPTY(&oldadp->ad_newdirblk))
5499 panic("allocdirect_merge: extra newdirblk");
5500 WORKLIST_INSERT(&newadp->ad_newdirblk, wk);
5502 TAILQ_REMOVE(adphead, oldadp, ad_next);
5504 * We need to move any journal dependencies over to the freefrag
5505 * that releases this block if it exists. Otherwise we are
5506 * extending an existing block and we'll wait until that is
5507 * complete to release the journal space and extend the
5508 * new journal to cover this old space as well.
5510 if (freefrag == NULL) {
5511 if (oldadp->ad_newblkno != newadp->ad_newblkno)
5512 panic("allocdirect_merge: %jd != %jd",
5513 oldadp->ad_newblkno, newadp->ad_newblkno);
5514 newadp->ad_block.nb_jnewblk = (struct jnewblk *)
5515 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list,
5516 &oldadp->ad_block.nb_jnewblk->jn_list,
5517 &newadp->ad_block.nb_jwork);
5518 oldadp->ad_block.nb_jnewblk = NULL;
5519 cancel_newblk(&oldadp->ad_block, NULL,
5520 &newadp->ad_block.nb_jwork);
5522 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block,
5523 &freefrag->ff_list, &freefrag->ff_jwork);
5524 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk,
5525 &freefrag->ff_jwork);
5527 free_newblk(&oldadp->ad_block);
5531 * Allocate a jfreefrag structure to journal a single block free.
5533 static struct jfreefrag *
5534 newjfreefrag(freefrag, ip, blkno, size, lbn)
5535 struct freefrag *freefrag;
5541 struct jfreefrag *jfreefrag;
5545 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
5547 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump));
5548 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
5549 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
5550 jfreefrag->fr_ino = ip->i_number;
5551 jfreefrag->fr_lbn = lbn;
5552 jfreefrag->fr_blkno = blkno;
5553 jfreefrag->fr_frags = numfrags(fs, size);
5554 jfreefrag->fr_freefrag = freefrag;
5560 * Allocate a new freefrag structure.
5562 static struct freefrag *
5563 newfreefrag(ip, blkno, size, lbn)
5569 struct freefrag *freefrag;
5572 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd",
5573 ip->i_number, blkno, size, lbn);
5575 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
5576 panic("newfreefrag: frag size");
5577 freefrag = malloc(sizeof(struct freefrag),
5578 M_FREEFRAG, M_SOFTDEP_FLAGS);
5579 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
5580 freefrag->ff_state = ATTACHED;
5581 LIST_INIT(&freefrag->ff_jwork);
5582 freefrag->ff_inum = ip->i_number;
5583 freefrag->ff_vtype = ITOV(ip)->v_type;
5584 freefrag->ff_blkno = blkno;
5585 freefrag->ff_fragsize = size;
5587 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) {
5588 freefrag->ff_jdep = (struct worklist *)
5589 newjfreefrag(freefrag, ip, blkno, size, lbn);
5591 freefrag->ff_state |= DEPCOMPLETE;
5592 freefrag->ff_jdep = NULL;
5599 * This workitem de-allocates fragments that were replaced during
5600 * file block allocation.
5603 handle_workitem_freefrag(freefrag)
5604 struct freefrag *freefrag;
5606 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
5607 struct workhead wkhd;
5610 "handle_workitem_freefrag: ino %d blkno %jd size %ld",
5611 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize);
5613 * It would be illegal to add new completion items to the
5614 * freefrag after it was schedule to be done so it must be
5615 * safe to modify the list head here.
5619 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
5621 * If the journal has not been written we must cancel it here.
5623 if (freefrag->ff_jdep) {
5624 if (freefrag->ff_jdep->wk_type != D_JNEWBLK)
5625 panic("handle_workitem_freefrag: Unexpected type %d\n",
5626 freefrag->ff_jdep->wk_type);
5627 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd);
5630 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
5631 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd);
5633 WORKITEM_FREE(freefrag, D_FREEFRAG);
5638 * Set up a dependency structure for an external attributes data block.
5639 * This routine follows much of the structure of softdep_setup_allocdirect.
5640 * See the description of softdep_setup_allocdirect above for details.
5643 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
5646 ufs2_daddr_t newblkno;
5647 ufs2_daddr_t oldblkno;
5652 struct allocdirect *adp, *oldadp;
5653 struct allocdirectlst *adphead;
5654 struct freefrag *freefrag;
5655 struct inodedep *inodedep;
5656 struct jnewblk *jnewblk;
5657 struct newblk *newblk;
5661 mp = UFSTOVFS(ip->i_ump);
5662 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5663 ("softdep_setup_allocext called on non-softdep filesystem"));
5664 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR",
5668 if (oldblkno && oldblkno != newblkno)
5669 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
5673 ACQUIRE_LOCK(ip->i_ump);
5674 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
5675 panic("softdep_setup_allocext: lost block");
5676 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5677 ("softdep_setup_allocext: newblk already initialized"));
5679 * Convert the newblk to an allocdirect.
5681 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
5682 adp = (struct allocdirect *)newblk;
5683 newblk->nb_freefrag = freefrag;
5684 adp->ad_offset = off;
5685 adp->ad_oldblkno = oldblkno;
5686 adp->ad_newsize = newsize;
5687 adp->ad_oldsize = oldsize;
5688 adp->ad_state |= EXTDATA;
5691 * Finish initializing the journal.
5693 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5694 jnewblk->jn_ino = ip->i_number;
5695 jnewblk->jn_lbn = lbn;
5696 add_to_journal(&jnewblk->jn_list);
5698 if (freefrag && freefrag->ff_jdep != NULL &&
5699 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5700 add_to_journal(freefrag->ff_jdep);
5701 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5702 adp->ad_inodedep = inodedep;
5704 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
5706 * The list of allocdirects must be kept in sorted and ascending
5707 * order so that the rollback routines can quickly determine the
5708 * first uncommitted block (the size of the file stored on disk
5709 * ends at the end of the lowest committed fragment, or if there
5710 * are no fragments, at the end of the highest committed block).
5711 * Since files generally grow, the typical case is that the new
5712 * block is to be added at the end of the list. We speed this
5713 * special case by checking against the last allocdirect in the
5714 * list before laboriously traversing the list looking for the
5717 adphead = &inodedep->id_newextupdt;
5718 oldadp = TAILQ_LAST(adphead, allocdirectlst);
5719 if (oldadp == NULL || oldadp->ad_offset <= off) {
5720 /* insert at end of list */
5721 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
5722 if (oldadp != NULL && oldadp->ad_offset == off)
5723 allocdirect_merge(adphead, adp, oldadp);
5724 FREE_LOCK(ip->i_ump);
5727 TAILQ_FOREACH(oldadp, adphead, ad_next) {
5728 if (oldadp->ad_offset >= off)
5732 panic("softdep_setup_allocext: lost entry");
5733 /* insert in middle of list */
5734 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
5735 if (oldadp->ad_offset == off)
5736 allocdirect_merge(adphead, adp, oldadp);
5737 FREE_LOCK(ip->i_ump);
5741 * Indirect block allocation dependencies.
5743 * The same dependencies that exist for a direct block also exist when
5744 * a new block is allocated and pointed to by an entry in a block of
5745 * indirect pointers. The undo/redo states described above are also
5746 * used here. Because an indirect block contains many pointers that
5747 * may have dependencies, a second copy of the entire in-memory indirect
5748 * block is kept. The buffer cache copy is always completely up-to-date.
5749 * The second copy, which is used only as a source for disk writes,
5750 * contains only the safe pointers (i.e., those that have no remaining
5751 * update dependencies). The second copy is freed when all pointers
5752 * are safe. The cache is not allowed to replace indirect blocks with
5753 * pending update dependencies. If a buffer containing an indirect
5754 * block with dependencies is written, these routines will mark it
5755 * dirty again. It can only be successfully written once all the
5756 * dependencies are removed. The ffs_fsync routine in conjunction with
5757 * softdep_sync_metadata work together to get all the dependencies
5758 * removed so that a file can be successfully written to disk. Three
5759 * procedures are used when setting up indirect block pointer
5760 * dependencies. The division is necessary because of the organization
5761 * of the "balloc" routine and because of the distinction between file
5762 * pages and file metadata blocks.
5766 * Allocate a new allocindir structure.
5768 static struct allocindir *
5769 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
5770 struct inode *ip; /* inode for file being extended */
5771 int ptrno; /* offset of pointer in indirect block */
5772 ufs2_daddr_t newblkno; /* disk block number being added */
5773 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5776 struct newblk *newblk;
5777 struct allocindir *aip;
5778 struct freefrag *freefrag;
5779 struct jnewblk *jnewblk;
5782 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn);
5785 ACQUIRE_LOCK(ip->i_ump);
5786 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0)
5787 panic("new_allocindir: lost block");
5788 KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
5789 ("newallocindir: newblk already initialized"));
5790 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR);
5791 newblk->nb_freefrag = freefrag;
5792 aip = (struct allocindir *)newblk;
5793 aip->ai_offset = ptrno;
5794 aip->ai_oldblkno = oldblkno;
5796 if ((jnewblk = newblk->nb_jnewblk) != NULL) {
5797 jnewblk->jn_ino = ip->i_number;
5798 jnewblk->jn_lbn = lbn;
5799 add_to_journal(&jnewblk->jn_list);
5801 if (freefrag && freefrag->ff_jdep != NULL &&
5802 freefrag->ff_jdep->wk_type == D_JFREEFRAG)
5803 add_to_journal(freefrag->ff_jdep);
5808 * Called just before setting an indirect block pointer
5809 * to a newly allocated file page.
5812 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
5813 struct inode *ip; /* inode for file being extended */
5814 ufs_lbn_t lbn; /* allocated block number within file */
5815 struct buf *bp; /* buffer with indirect blk referencing page */
5816 int ptrno; /* offset of pointer in indirect block */
5817 ufs2_daddr_t newblkno; /* disk block number being added */
5818 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
5819 struct buf *nbp; /* buffer holding allocated page */
5821 struct inodedep *inodedep;
5822 struct freefrag *freefrag;
5823 struct allocindir *aip;
5824 struct pagedep *pagedep;
5827 mp = UFSTOVFS(ip->i_ump);
5828 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
5829 ("softdep_setup_allocindir_page called on non-softdep filesystem"));
5830 KASSERT(lbn == nbp->b_lblkno,
5831 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
5832 lbn, bp->b_lblkno));
5834 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd "
5835 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn);
5836 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
5837 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
5838 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5840 * If we are allocating a directory page, then we must
5841 * allocate an associated pagedep to track additions and
5844 if ((ip->i_mode & IFMT) == IFDIR)
5845 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep);
5846 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5847 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
5848 FREE_LOCK(ip->i_ump);
5850 handle_workitem_freefrag(freefrag);
5854 * Called just before setting an indirect block pointer to a
5855 * newly allocated indirect block.
5858 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
5859 struct buf *nbp; /* newly allocated indirect block */
5860 struct inode *ip; /* inode for file being extended */
5861 struct buf *bp; /* indirect block referencing allocated block */
5862 int ptrno; /* offset of pointer in indirect block */
5863 ufs2_daddr_t newblkno; /* disk block number being added */
5865 struct inodedep *inodedep;
5866 struct allocindir *aip;
5869 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
5870 ("softdep_setup_allocindir_meta called on non-softdep filesystem"));
5872 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d",
5873 ip->i_number, newblkno, ptrno);
5874 lbn = nbp->b_lblkno;
5875 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
5876 aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
5877 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
5879 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
5880 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
5881 panic("softdep_setup_allocindir_meta: Block already existed");
5882 FREE_LOCK(ip->i_ump);
5886 indirdep_complete(indirdep)
5887 struct indirdep *indirdep;
5889 struct allocindir *aip;
5891 LIST_REMOVE(indirdep, ir_next);
5892 indirdep->ir_state |= DEPCOMPLETE;
5894 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
5895 LIST_REMOVE(aip, ai_next);
5896 free_newblk(&aip->ai_block);
5899 * If this indirdep is not attached to a buf it was simply waiting
5900 * on completion to clear completehd. free_indirdep() asserts
5901 * that nothing is dangling.
5903 if ((indirdep->ir_state & ONWORKLIST) == 0)
5904 free_indirdep(indirdep);
5907 static struct indirdep *
5908 indirdep_lookup(mp, ip, bp)
5913 struct indirdep *indirdep, *newindirdep;
5914 struct newblk *newblk;
5915 struct ufsmount *ump;
5916 struct worklist *wk;
5926 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5927 if (wk->wk_type != D_INDIRDEP)
5929 indirdep = WK_INDIRDEP(wk);
5932 /* Found on the buffer worklist, no new structure to free. */
5933 if (indirdep != NULL && newindirdep == NULL)
5935 if (indirdep != NULL && newindirdep != NULL)
5936 panic("indirdep_lookup: simultaneous create");
5937 /* None found on the buffer and a new structure is ready. */
5938 if (indirdep == NULL && newindirdep != NULL)
5940 /* None found and no new structure available. */
5942 newindirdep = malloc(sizeof(struct indirdep),
5943 M_INDIRDEP, M_SOFTDEP_FLAGS);
5944 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
5945 newindirdep->ir_state = ATTACHED;
5946 if (ip->i_ump->um_fstype == UFS1)
5947 newindirdep->ir_state |= UFS1FMT;
5948 TAILQ_INIT(&newindirdep->ir_trunc);
5949 newindirdep->ir_saveddata = NULL;
5950 LIST_INIT(&newindirdep->ir_deplisthd);
5951 LIST_INIT(&newindirdep->ir_donehd);
5952 LIST_INIT(&newindirdep->ir_writehd);
5953 LIST_INIT(&newindirdep->ir_completehd);
5954 if (bp->b_blkno == bp->b_lblkno) {
5955 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
5957 bp->b_blkno = blkno;
5959 newindirdep->ir_freeblks = NULL;
5960 newindirdep->ir_savebp =
5961 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
5962 newindirdep->ir_bp = bp;
5963 BUF_KERNPROC(newindirdep->ir_savebp);
5964 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
5967 indirdep = newindirdep;
5968 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
5970 * If the block is not yet allocated we don't set DEPCOMPLETE so
5971 * that we don't free dependencies until the pointers are valid.
5972 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather
5973 * than using the hash.
5975 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk))
5976 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next);
5978 indirdep->ir_state |= DEPCOMPLETE;
5983 * Called to finish the allocation of the "aip" allocated
5984 * by one of the two routines above.
5986 static struct freefrag *
5987 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
5988 struct buf *bp; /* in-memory copy of the indirect block */
5989 struct inode *ip; /* inode for file being extended */
5990 struct inodedep *inodedep; /* Inodedep for ip */
5991 struct allocindir *aip; /* allocindir allocated by the above routines */
5992 ufs_lbn_t lbn; /* Logical block number for this block. */
5995 struct indirdep *indirdep;
5996 struct allocindir *oldaip;
5997 struct freefrag *freefrag;
6000 LOCK_OWNED(ip->i_ump);
6001 mp = UFSTOVFS(ip->i_ump);
6003 if (bp->b_lblkno >= 0)
6004 panic("setup_allocindir_phase2: not indir blk");
6005 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs),
6006 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset));
6007 indirdep = indirdep_lookup(mp, ip, bp);
6008 KASSERT(indirdep->ir_savebp != NULL,
6009 ("setup_allocindir_phase2 NULL ir_savebp"));
6010 aip->ai_indirdep = indirdep;
6012 * Check for an unwritten dependency for this indirect offset. If
6013 * there is, merge the old dependency into the new one. This happens
6014 * as a result of reallocblk only.
6017 if (aip->ai_oldblkno != 0) {
6018 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) {
6019 if (oldaip->ai_offset == aip->ai_offset) {
6020 freefrag = allocindir_merge(aip, oldaip);
6024 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) {
6025 if (oldaip->ai_offset == aip->ai_offset) {
6026 freefrag = allocindir_merge(aip, oldaip);
6032 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
6037 * Merge two allocindirs which refer to the same block. Move newblock
6038 * dependencies and setup the freefrags appropriately.
6040 static struct freefrag *
6041 allocindir_merge(aip, oldaip)
6042 struct allocindir *aip;
6043 struct allocindir *oldaip;
6045 struct freefrag *freefrag;
6046 struct worklist *wk;
6048 if (oldaip->ai_newblkno != aip->ai_oldblkno)
6049 panic("allocindir_merge: blkno");
6050 aip->ai_oldblkno = oldaip->ai_oldblkno;
6051 freefrag = aip->ai_freefrag;
6052 aip->ai_freefrag = oldaip->ai_freefrag;
6053 oldaip->ai_freefrag = NULL;
6054 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
6056 * If we are tracking a new directory-block allocation,
6057 * move it from the old allocindir to the new allocindir.
6059 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
6060 WORKLIST_REMOVE(wk);
6061 if (!LIST_EMPTY(&oldaip->ai_newdirblk))
6062 panic("allocindir_merge: extra newdirblk");
6063 WORKLIST_INSERT(&aip->ai_newdirblk, wk);
6066 * We can skip journaling for this freefrag and just complete
6067 * any pending journal work for the allocindir that is being
6068 * removed after the freefrag completes.
6070 if (freefrag->ff_jdep)
6071 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep));
6072 LIST_REMOVE(oldaip, ai_next);
6073 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block,
6074 &freefrag->ff_list, &freefrag->ff_jwork);
6075 free_newblk(&oldaip->ai_block);
6081 setup_freedirect(freeblks, ip, i, needj)
6082 struct freeblks *freeblks;
6090 blkno = DIP(ip, i_db[i]);
6093 DIP_SET(ip, i_db[i], 0);
6094 frags = sblksize(ip->i_fs, ip->i_size, i);
6095 frags = numfrags(ip->i_fs, frags);
6096 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj);
6100 setup_freeext(freeblks, ip, i, needj)
6101 struct freeblks *freeblks;
6109 blkno = ip->i_din2->di_extb[i];
6112 ip->i_din2->di_extb[i] = 0;
6113 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i);
6114 frags = numfrags(ip->i_fs, frags);
6115 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj);
6119 setup_freeindir(freeblks, ip, i, lbn, needj)
6120 struct freeblks *freeblks;
6128 blkno = DIP(ip, i_ib[i]);
6131 DIP_SET(ip, i_ib[i], 0);
6132 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag,
6136 static inline struct freeblks *
6141 struct freeblks *freeblks;
6143 freeblks = malloc(sizeof(struct freeblks),
6144 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
6145 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
6146 LIST_INIT(&freeblks->fb_jblkdephd);
6147 LIST_INIT(&freeblks->fb_jwork);
6148 freeblks->fb_ref = 0;
6149 freeblks->fb_cgwait = 0;
6150 freeblks->fb_state = ATTACHED;
6151 freeblks->fb_uid = ip->i_uid;
6152 freeblks->fb_inum = ip->i_number;
6153 freeblks->fb_vtype = ITOV(ip)->v_type;
6154 freeblks->fb_modrev = DIP(ip, i_modrev);
6155 freeblks->fb_devvp = ip->i_devvp;
6156 freeblks->fb_chkcnt = 0;
6157 freeblks->fb_len = 0;
6163 trunc_indirdep(indirdep, freeblks, bp, off)
6164 struct indirdep *indirdep;
6165 struct freeblks *freeblks;
6169 struct allocindir *aip, *aipn;
6172 * The first set of allocindirs won't be in savedbp.
6174 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn)
6175 if (aip->ai_offset > off)
6176 cancel_allocindir(aip, bp, freeblks, 1);
6177 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn)
6178 if (aip->ai_offset > off)
6179 cancel_allocindir(aip, bp, freeblks, 1);
6181 * These will exist in savedbp.
6183 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn)
6184 if (aip->ai_offset > off)
6185 cancel_allocindir(aip, NULL, freeblks, 0);
6186 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn)
6187 if (aip->ai_offset > off)
6188 cancel_allocindir(aip, NULL, freeblks, 0);
6192 * Follow the chain of indirects down to lastlbn creating a freework
6193 * structure for each. This will be used to start indir_trunc() at
6194 * the right offset and create the journal records for the parrtial
6195 * truncation. A second step will handle the truncated dependencies.
6198 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno)
6199 struct freeblks *freeblks;
6205 struct indirdep *indirdep;
6206 struct indirdep *indirn;
6207 struct freework *freework;
6208 struct newblk *newblk;
6222 mp = freeblks->fb_list.wk_mp;
6223 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
6224 if ((bp->b_flags & B_CACHE) == 0) {
6225 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno);
6226 bp->b_iocmd = BIO_READ;
6227 bp->b_flags &= ~B_INVAL;
6228 bp->b_ioflags &= ~BIO_ERROR;
6229 vfs_busy_pages(bp, 0);
6230 bp->b_iooffset = dbtob(bp->b_blkno);
6232 curthread->td_ru.ru_inblock++;
6233 error = bufwait(bp);
6239 level = lbn_level(lbn);
6240 lbnadd = lbn_offset(ip->i_fs, level);
6242 * Compute the offset of the last block we want to keep. Store
6243 * in the freework the first block we want to completely free.
6245 off = (lastlbn - -(lbn + level)) / lbnadd;
6246 if (off + 1 == NINDIR(ip->i_fs))
6248 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1,
6251 * Link the freework into the indirdep. This will prevent any new
6252 * allocations from proceeding until we are finished with the
6253 * truncate and the block is written.
6255 ACQUIRE_LOCK(ip->i_ump);
6256 indirdep = indirdep_lookup(mp, ip, bp);
6257 if (indirdep->ir_freeblks)
6258 panic("setup_trunc_indir: indirdep already truncated.");
6259 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next);
6260 freework->fw_indir = indirdep;
6262 * Cancel any allocindirs that will not make it to disk.
6263 * We have to do this for all copies of the indirdep that
6264 * live on this newblk.
6266 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
6267 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk);
6268 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next)
6269 trunc_indirdep(indirn, freeblks, bp, off);
6271 trunc_indirdep(indirdep, freeblks, bp, off);
6272 FREE_LOCK(ip->i_ump);
6274 * Creation is protected by the buf lock. The saveddata is only
6275 * needed if a full truncation follows a partial truncation but it
6276 * is difficult to allocate in that case so we fetch it anyway.
6278 if (indirdep->ir_saveddata == NULL)
6279 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
6282 /* Fetch the blkno of the child and the zero start offset. */
6283 if (ip->i_ump->um_fstype == UFS1) {
6284 blkno = ((ufs1_daddr_t *)bp->b_data)[off];
6285 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1];
6287 blkno = ((ufs2_daddr_t *)bp->b_data)[off];
6288 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1];
6291 /* Zero the truncated pointers. */
6292 end = bp->b_data + bp->b_bcount;
6293 bzero(start, end - start);
6299 lbn++; /* adjust level */
6300 lbn -= (off * lbnadd);
6301 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno);
6305 * Complete the partial truncation of an indirect block setup by
6306 * setup_trunc_indir(). This zeros the truncated pointers in the saved
6307 * copy and writes them to disk before the freeblks is allowed to complete.
6310 complete_trunc_indir(freework)
6311 struct freework *freework;
6313 struct freework *fwn;
6314 struct indirdep *indirdep;
6315 struct ufsmount *ump;
6320 ump = VFSTOUFS(freework->fw_list.wk_mp);
6322 indirdep = freework->fw_indir;
6324 bp = indirdep->ir_bp;
6325 /* See if the block was discarded. */
6328 /* Inline part of getdirtybuf(). We dont want bremfree. */
6329 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0)
6331 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
6332 LOCK_PTR(ump)) == 0)
6336 freework->fw_state |= DEPCOMPLETE;
6337 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next);
6339 * Zero the pointers in the saved copy.
6341 if (indirdep->ir_state & UFS1FMT)
6342 start = sizeof(ufs1_daddr_t);
6344 start = sizeof(ufs2_daddr_t);
6345 start *= freework->fw_start;
6346 count = indirdep->ir_savebp->b_bcount - start;
6347 start += (uintptr_t)indirdep->ir_savebp->b_data;
6348 bzero((char *)start, count);
6350 * We need to start the next truncation in the list if it has not
6353 fwn = TAILQ_FIRST(&indirdep->ir_trunc);
6355 if (fwn->fw_freeblks == indirdep->ir_freeblks)
6356 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next);
6357 if ((fwn->fw_state & ONWORKLIST) == 0)
6358 freework_enqueue(fwn);
6361 * If bp is NULL the block was fully truncated, restore
6362 * the saved block list otherwise free it if it is no
6365 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
6367 bcopy(indirdep->ir_saveddata,
6368 indirdep->ir_savebp->b_data,
6369 indirdep->ir_savebp->b_bcount);
6370 free(indirdep->ir_saveddata, M_INDIRDEP);
6371 indirdep->ir_saveddata = NULL;
6374 * When bp is NULL there is a full truncation pending. We
6375 * must wait for this full truncation to be journaled before
6376 * we can release this freework because the disk pointers will
6377 * never be written as zero.
6380 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd))
6381 handle_written_freework(freework);
6383 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd,
6384 &freework->fw_list);
6386 /* Complete when the real copy is written. */
6387 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list);
6393 * Calculate the number of blocks we are going to release where datablocks
6394 * is the current total and length is the new file size.
6397 blkcount(fs, datablocks, length)
6399 ufs2_daddr_t datablocks;
6402 off_t totblks, numblks;
6405 numblks = howmany(length, fs->fs_bsize);
6406 if (numblks <= NDADDR) {
6407 totblks = howmany(length, fs->fs_fsize);
6410 totblks = blkstofrags(fs, numblks);
6413 * Count all single, then double, then triple indirects required.
6414 * Subtracting one indirects worth of blocks for each pass
6415 * acknowledges one of each pointed to by the inode.
6418 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs)));
6419 numblks -= NINDIR(fs);
6422 numblks = howmany(numblks, NINDIR(fs));
6425 totblks = fsbtodb(fs, totblks);
6427 * Handle sparse files. We can't reclaim more blocks than the inode
6428 * references. We will correct it later in handle_complete_freeblks()
6429 * when we know the real count.
6431 if (totblks > datablocks)
6433 return (datablocks - totblks);
6437 * Handle freeblocks for journaled softupdate filesystems.
6439 * Contrary to normal softupdates, we must preserve the block pointers in
6440 * indirects until their subordinates are free. This is to avoid journaling
6441 * every block that is freed which may consume more space than the journal
6442 * itself. The recovery program will see the free block journals at the
6443 * base of the truncated area and traverse them to reclaim space. The
6444 * pointers in the inode may be cleared immediately after the journal
6445 * records are written because each direct and indirect pointer in the
6446 * inode is recorded in a journal. This permits full truncation to proceed
6447 * asynchronously. The write order is journal -> inode -> cgs -> indirects.
6449 * The algorithm is as follows:
6450 * 1) Traverse the in-memory state and create journal entries to release
6451 * the relevant blocks and full indirect trees.
6452 * 2) Traverse the indirect block chain adding partial truncation freework
6453 * records to indirects in the path to lastlbn. The freework will
6454 * prevent new allocation dependencies from being satisfied in this
6455 * indirect until the truncation completes.
6456 * 3) Read and lock the inode block, performing an update with the new size
6457 * and pointers. This prevents truncated data from becoming valid on
6458 * disk through step 4.
6459 * 4) Reap unsatisfied dependencies that are beyond the truncated area,
6460 * eliminate journal work for those records that do not require it.
6461 * 5) Schedule the journal records to be written followed by the inode block.
6462 * 6) Allocate any necessary frags for the end of file.
6463 * 7) Zero any partially truncated blocks.
6465 * From this truncation proceeds asynchronously using the freework and
6466 * indir_trunc machinery. The file will not be extended again into a
6467 * partially truncated indirect block until all work is completed but
6468 * the normal dependency mechanism ensures that it is rolled back/forward
6469 * as appropriate. Further truncation may occur without delay and is
6470 * serialized in indir_trunc().
6473 softdep_journal_freeblocks(ip, cred, length, flags)
6474 struct inode *ip; /* The inode whose length is to be reduced */
6476 off_t length; /* The new length for the file */
6477 int flags; /* IO_EXT and/or IO_NORMAL */
6479 struct freeblks *freeblks, *fbn;
6480 struct worklist *wk, *wkn;
6481 struct inodedep *inodedep;
6482 struct jblkdep *jblkdep;
6483 struct allocdirect *adp, *adpn;
6484 struct ufsmount *ump;
6489 ufs2_daddr_t extblocks, datablocks;
6490 ufs_lbn_t tmpval, lbn, lastlbn;
6491 int frags, lastoff, iboff, allocblock, needj, error, i;
6496 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6497 ("softdep_journal_freeblocks called on non-softdep filesystem"));
6505 freeblks = newfreeblks(mp, ip);
6508 * If we're truncating a removed file that will never be written
6509 * we don't need to journal the block frees. The canceled journals
6510 * for the allocations will suffice.
6512 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6513 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
6516 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d",
6517 ip->i_number, length, needj);
6520 * Calculate the lbn that we are truncating to. This results in -1
6521 * if we're truncating the 0 bytes. So it is the last lbn we want
6522 * to keep, not the first lbn we want to truncate.
6524 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1;
6525 lastoff = blkoff(fs, length);
6527 * Compute frags we are keeping in lastlbn. 0 means all.
6529 if (lastlbn >= 0 && lastlbn < NDADDR) {
6530 frags = fragroundup(fs, lastoff);
6531 /* adp offset of last valid allocdirect. */
6533 } else if (lastlbn > 0)
6535 if (fs->fs_magic == FS_UFS2_MAGIC)
6536 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6538 * Handle normal data blocks and indirects. This section saves
6539 * values used after the inode update to complete frag and indirect
6542 if ((flags & IO_NORMAL) != 0) {
6544 * Handle truncation of whole direct and indirect blocks.
6546 for (i = iboff + 1; i < NDADDR; i++)
6547 setup_freedirect(freeblks, ip, i, needj);
6548 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6549 i++, lbn += tmpval, tmpval *= NINDIR(fs)) {
6550 /* Release a whole indirect tree. */
6551 if (lbn > lastlbn) {
6552 setup_freeindir(freeblks, ip, i, -lbn -i,
6558 * Traverse partially truncated indirect tree.
6560 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn)
6561 setup_trunc_indir(freeblks, ip, -lbn - i,
6562 lastlbn, DIP(ip, i_ib[i]));
6565 * Handle partial truncation to a frag boundary.
6571 oldfrags = blksize(fs, ip, lastlbn);
6572 blkno = DIP(ip, i_db[lastlbn]);
6573 if (blkno && oldfrags != frags) {
6575 oldfrags = numfrags(ip->i_fs, oldfrags);
6576 blkno += numfrags(ip->i_fs, frags);
6577 newfreework(ump, freeblks, NULL, lastlbn,
6578 blkno, oldfrags, 0, needj);
6580 adjust_newfreework(freeblks,
6581 numfrags(ip->i_fs, frags));
6582 } else if (blkno == 0)
6586 * Add a journal record for partial truncate if we are
6587 * handling indirect blocks. Non-indirects need no extra
6590 if (length != 0 && lastlbn >= NDADDR) {
6591 ip->i_flag |= IN_TRUNCATED;
6592 newjtrunc(freeblks, length, 0);
6594 ip->i_size = length;
6595 DIP_SET(ip, i_size, ip->i_size);
6596 datablocks = DIP(ip, i_blocks) - extblocks;
6598 datablocks = blkcount(ip->i_fs, datablocks, length);
6599 freeblks->fb_len = length;
6601 if ((flags & IO_EXT) != 0) {
6602 for (i = 0; i < NXADDR; i++)
6603 setup_freeext(freeblks, ip, i, needj);
6604 ip->i_din2->di_extsize = 0;
6605 datablocks += extblocks;
6608 /* Reference the quotas in case the block count is wrong in the end. */
6609 quotaref(vp, freeblks->fb_quota);
6610 (void) chkdq(ip, -datablocks, NOCRED, 0);
6612 freeblks->fb_chkcnt = -datablocks;
6614 fs->fs_pendingblocks += datablocks;
6616 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6618 * Handle truncation of incomplete alloc direct dependencies. We
6619 * hold the inode block locked to prevent incomplete dependencies
6620 * from reaching the disk while we are eliminating those that
6621 * have been truncated. This is a partially inlined ffs_update().
6624 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
6625 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6626 (int)fs->fs_bsize, cred, &bp);
6629 softdep_error("softdep_journal_freeblocks", error);
6632 if (bp->b_bufsize == fs->fs_bsize)
6633 bp->b_flags |= B_CLUSTEROK;
6634 softdep_update_inodeblock(ip, bp, 0);
6635 if (ump->um_fstype == UFS1)
6636 *((struct ufs1_dinode *)bp->b_data +
6637 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
6639 *((struct ufs2_dinode *)bp->b_data +
6640 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
6642 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6643 if ((inodedep->id_state & IOSTARTED) != 0)
6644 panic("softdep_setup_freeblocks: inode busy");
6646 * Add the freeblks structure to the list of operations that
6647 * must await the zero'ed inode being written to disk. If we
6648 * still have a bitmap dependency (needj), then the inode
6649 * has never been written to disk, so we can process the
6650 * freeblks below once we have deleted the dependencies.
6653 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6655 freeblks->fb_state |= COMPLETE;
6656 if ((flags & IO_NORMAL) != 0) {
6657 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) {
6658 if (adp->ad_offset > iboff)
6659 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6662 * Truncate the allocdirect. We could eliminate
6663 * or modify journal records as well.
6665 else if (adp->ad_offset == iboff && frags)
6666 adp->ad_newsize = frags;
6669 if ((flags & IO_EXT) != 0)
6670 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6671 cancel_allocdirect(&inodedep->id_extupdt, adp,
6674 * Scan the bufwait list for newblock dependencies that will never
6677 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) {
6678 if (wk->wk_type != D_ALLOCDIRECT)
6680 adp = WK_ALLOCDIRECT(wk);
6681 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) ||
6682 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) {
6683 cancel_jfreeblk(freeblks, adp->ad_newblkno);
6684 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork);
6685 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
6691 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps)
6692 add_to_journal(&jblkdep->jb_list);
6696 * Truncate dependency structures beyond length.
6698 trunc_dependencies(ip, freeblks, lastlbn, frags, flags);
6700 * This is only set when we need to allocate a fragment because
6701 * none existed at the end of a frag-sized file. It handles only
6702 * allocating a new, zero filled block.
6705 ip->i_size = length - lastoff;
6706 DIP_SET(ip, i_size, ip->i_size);
6707 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp);
6709 softdep_error("softdep_journal_freeblks", error);
6712 ip->i_size = length;
6713 DIP_SET(ip, i_size, length);
6714 ip->i_flag |= IN_CHANGE | IN_UPDATE;
6715 allocbuf(bp, frags);
6718 } else if (lastoff != 0 && vp->v_type != VDIR) {
6722 * Zero the end of a truncated frag or block.
6724 size = sblksize(fs, length, lastlbn);
6725 error = bread(vp, lastlbn, size, cred, &bp);
6727 softdep_error("softdep_journal_freeblks", error);
6730 bzero((char *)bp->b_data + lastoff, size - lastoff);
6735 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6736 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
6737 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
6739 * We zero earlier truncations so they don't erroneously
6742 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0)
6743 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next)
6745 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE &&
6746 LIST_EMPTY(&freeblks->fb_jblkdephd))
6747 freeblks->fb_state |= INPROGRESS;
6752 handle_workitem_freeblocks(freeblks, 0);
6753 trunc_pages(ip, length, extblocks, flags);
6758 * Flush a JOP_SYNC to the journal.
6761 softdep_journal_fsync(ip)
6764 struct jfsync *jfsync;
6766 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
6767 ("softdep_journal_fsync called on non-softdep filesystem"));
6768 if ((ip->i_flag & IN_TRUNCATED) == 0)
6770 ip->i_flag &= ~IN_TRUNCATED;
6771 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO);
6772 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump));
6773 jfsync->jfs_size = ip->i_size;
6774 jfsync->jfs_ino = ip->i_number;
6775 ACQUIRE_LOCK(ip->i_ump);
6776 add_to_journal(&jfsync->jfs_list);
6777 jwait(&jfsync->jfs_list, MNT_WAIT);
6778 FREE_LOCK(ip->i_ump);
6782 * Block de-allocation dependencies.
6784 * When blocks are de-allocated, the on-disk pointers must be nullified before
6785 * the blocks are made available for use by other files. (The true
6786 * requirement is that old pointers must be nullified before new on-disk
6787 * pointers are set. We chose this slightly more stringent requirement to
6788 * reduce complexity.) Our implementation handles this dependency by updating
6789 * the inode (or indirect block) appropriately but delaying the actual block
6790 * de-allocation (i.e., freemap and free space count manipulation) until
6791 * after the updated versions reach stable storage. After the disk is
6792 * updated, the blocks can be safely de-allocated whenever it is convenient.
6793 * This implementation handles only the common case of reducing a file's
6794 * length to zero. Other cases are handled by the conventional synchronous
6797 * The ffs implementation with which we worked double-checks
6798 * the state of the block pointers and file size as it reduces
6799 * a file's length. Some of this code is replicated here in our
6800 * soft updates implementation. The freeblks->fb_chkcnt field is
6801 * used to transfer a part of this information to the procedure
6802 * that eventually de-allocates the blocks.
6804 * This routine should be called from the routine that shortens
6805 * a file's length, before the inode's size or block pointers
6806 * are modified. It will save the block pointer information for
6807 * later release and zero the inode so that the calling routine
6811 softdep_setup_freeblocks(ip, length, flags)
6812 struct inode *ip; /* The inode whose length is to be reduced */
6813 off_t length; /* The new length for the file */
6814 int flags; /* IO_EXT and/or IO_NORMAL */
6816 struct ufs1_dinode *dp1;
6817 struct ufs2_dinode *dp2;
6818 struct freeblks *freeblks;
6819 struct inodedep *inodedep;
6820 struct allocdirect *adp;
6821 struct ufsmount *ump;
6824 ufs2_daddr_t extblocks, datablocks;
6826 int i, delay, error;
6832 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
6833 ("softdep_setup_freeblocks called on non-softdep filesystem"));
6834 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld",
6835 ip->i_number, length);
6836 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length"));
6838 freeblks = newfreeblks(mp, ip);
6841 if (fs->fs_magic == FS_UFS2_MAGIC)
6842 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
6843 if ((flags & IO_NORMAL) != 0) {
6844 for (i = 0; i < NDADDR; i++)
6845 setup_freedirect(freeblks, ip, i, 0);
6846 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
6847 i++, lbn += tmpval, tmpval *= NINDIR(fs))
6848 setup_freeindir(freeblks, ip, i, -lbn -i, 0);
6850 DIP_SET(ip, i_size, 0);
6851 datablocks = DIP(ip, i_blocks) - extblocks;
6853 if ((flags & IO_EXT) != 0) {
6854 for (i = 0; i < NXADDR; i++)
6855 setup_freeext(freeblks, ip, i, 0);
6856 ip->i_din2->di_extsize = 0;
6857 datablocks += extblocks;
6860 /* Reference the quotas in case the block count is wrong in the end. */
6861 quotaref(ITOV(ip), freeblks->fb_quota);
6862 (void) chkdq(ip, -datablocks, NOCRED, 0);
6864 freeblks->fb_chkcnt = -datablocks;
6866 fs->fs_pendingblocks += datablocks;
6868 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
6870 * Push the zero'ed inode to to its disk buffer so that we are free
6871 * to delete its dependencies below. Once the dependencies are gone
6872 * the buffer can be safely released.
6874 if ((error = bread(ip->i_devvp,
6875 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
6876 (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
6878 softdep_error("softdep_setup_freeblocks", error);
6880 if (ump->um_fstype == UFS1) {
6881 dp1 = ((struct ufs1_dinode *)bp->b_data +
6882 ino_to_fsbo(fs, ip->i_number));
6883 ip->i_din1->di_freelink = dp1->di_freelink;
6886 dp2 = ((struct ufs2_dinode *)bp->b_data +
6887 ino_to_fsbo(fs, ip->i_number));
6888 ip->i_din2->di_freelink = dp2->di_freelink;
6892 * Find and eliminate any inode dependencies.
6895 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
6896 if ((inodedep->id_state & IOSTARTED) != 0)
6897 panic("softdep_setup_freeblocks: inode busy");
6899 * Add the freeblks structure to the list of operations that
6900 * must await the zero'ed inode being written to disk. If we
6901 * still have a bitmap dependency (delay == 0), then the inode
6902 * has never been written to disk, so we can process the
6903 * freeblks below once we have deleted the dependencies.
6905 delay = (inodedep->id_state & DEPCOMPLETE);
6907 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
6909 freeblks->fb_state |= COMPLETE;
6911 * Because the file length has been truncated to zero, any
6912 * pending block allocation dependency structures associated
6913 * with this inode are obsolete and can simply be de-allocated.
6914 * We must first merge the two dependency lists to get rid of
6915 * any duplicate freefrag structures, then purge the merged list.
6916 * If we still have a bitmap dependency, then the inode has never
6917 * been written to disk, so we can free any fragments without delay.
6919 if (flags & IO_NORMAL) {
6920 merge_inode_lists(&inodedep->id_newinoupdt,
6921 &inodedep->id_inoupdt);
6922 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
6923 cancel_allocdirect(&inodedep->id_inoupdt, adp,
6926 if (flags & IO_EXT) {
6927 merge_inode_lists(&inodedep->id_newextupdt,
6928 &inodedep->id_extupdt);
6929 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
6930 cancel_allocdirect(&inodedep->id_extupdt, adp,
6935 trunc_dependencies(ip, freeblks, -1, 0, flags);
6937 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
6938 (void) free_inodedep(inodedep);
6939 freeblks->fb_state |= DEPCOMPLETE;
6941 * If the inode with zeroed block pointers is now on disk
6942 * we can start freeing blocks.
6944 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
6945 freeblks->fb_state |= INPROGRESS;
6950 handle_workitem_freeblocks(freeblks, 0);
6951 trunc_pages(ip, length, extblocks, flags);
6955 * Eliminate pages from the page cache that back parts of this inode and
6956 * adjust the vnode pager's idea of our size. This prevents stale data
6957 * from hanging around in the page cache.
6960 trunc_pages(ip, length, extblocks, flags)
6963 ufs2_daddr_t extblocks;
6973 extend = OFF_TO_IDX(lblktosize(fs, -extblocks));
6974 if ((flags & IO_EXT) != 0)
6975 vn_pages_remove(vp, extend, 0);
6976 if ((flags & IO_NORMAL) == 0)
6978 BO_LOCK(&vp->v_bufobj);
6980 BO_UNLOCK(&vp->v_bufobj);
6982 * The vnode pager eliminates file pages we eliminate indirects
6985 vnode_pager_setsize(vp, length);
6987 * Calculate the end based on the last indirect we want to keep. If
6988 * the block extends into indirects we can just use the negative of
6989 * its lbn. Doubles and triples exist at lower numbers so we must
6990 * be careful not to remove those, if they exist. double and triple
6991 * indirect lbns do not overlap with others so it is not important
6992 * to verify how many levels are required.
6994 lbn = lblkno(fs, length);
6995 if (lbn >= NDADDR) {
6996 /* Calculate the virtual lbn of the triple indirect. */
6997 lbn = -lbn - (NIADDR - 1);
6998 end = OFF_TO_IDX(lblktosize(fs, lbn));
7001 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end);
7005 * See if the buf bp is in the range eliminated by truncation.
7008 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags)
7018 /* Only match ext/normal blocks as appropriate. */
7019 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
7020 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0))
7022 /* ALTDATA is always a full truncation. */
7023 if ((bp->b_xflags & BX_ALTDATA) != 0)
7025 /* -1 is full truncation. */
7029 * If this is a partial truncate we only want those
7030 * blocks and indirect blocks that cover the range
7035 lbn = -(lbn + lbn_level(lbn));
7038 /* Here we only truncate lblkno if it's partial. */
7039 if (lbn == lastlbn) {
7048 * Eliminate any dependencies that exist in memory beyond lblkno:off
7051 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags)
7053 struct freeblks *freeblks;
7064 * We must wait for any I/O in progress to finish so that
7065 * all potential buffers on the dirty list will be visible.
7066 * Once they are all there, walk the list and get rid of
7073 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
7074 bp->b_vflags &= ~BV_SCANNED;
7076 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
7077 if (bp->b_vflags & BV_SCANNED)
7079 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7080 bp->b_vflags |= BV_SCANNED;
7083 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer"));
7084 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL)
7087 if (deallocate_dependencies(bp, freeblks, blkoff))
7095 * Now do the work of vtruncbuf while also matching indirect blocks.
7097 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs)
7098 bp->b_vflags &= ~BV_SCANNED;
7100 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) {
7101 if (bp->b_vflags & BV_SCANNED)
7103 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
7104 bp->b_vflags |= BV_SCANNED;
7108 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
7109 BO_LOCKPTR(bo)) == ENOLCK) {
7113 bp->b_vflags |= BV_SCANNED;
7116 allocbuf(bp, blkoff);
7119 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF;
7130 cancel_pagedep(pagedep, freeblks, blkoff)
7131 struct pagedep *pagedep;
7132 struct freeblks *freeblks;
7135 struct jremref *jremref;
7136 struct jmvref *jmvref;
7137 struct dirrem *dirrem, *tmp;
7141 * Copy any directory remove dependencies to the list
7142 * to be processed after the freeblks proceeds. If
7143 * directory entry never made it to disk they
7144 * can be dumped directly onto the work list.
7146 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) {
7147 /* Skip this directory removal if it is intended to remain. */
7148 if (dirrem->dm_offset < blkoff)
7151 * If there are any dirrems we wait for the journal write
7152 * to complete and then restart the buf scan as the lock
7155 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
7156 jwait(&jremref->jr_list, MNT_WAIT);
7159 LIST_REMOVE(dirrem, dm_next);
7160 dirrem->dm_dirinum = pagedep->pd_ino;
7161 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list);
7163 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
7164 jwait(&jmvref->jm_list, MNT_WAIT);
7168 * When we're partially truncating a pagedep we just want to flush
7169 * journal entries and return. There can not be any adds in the
7170 * truncated portion of the directory and newblk must remain if
7171 * part of the block remains.
7176 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
7177 if (dap->da_offset > blkoff)
7178 panic("cancel_pagedep: diradd %p off %d > %d",
7179 dap, dap->da_offset, blkoff);
7180 for (i = 0; i < DAHASHSZ; i++)
7181 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist)
7182 if (dap->da_offset > blkoff)
7183 panic("cancel_pagedep: diradd %p off %d > %d",
7184 dap, dap->da_offset, blkoff);
7188 * There should be no directory add dependencies present
7189 * as the directory could not be truncated until all
7190 * children were removed.
7192 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
7193 ("deallocate_dependencies: pendinghd != NULL"));
7194 for (i = 0; i < DAHASHSZ; i++)
7195 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
7196 ("deallocate_dependencies: diraddhd != NULL"));
7197 if ((pagedep->pd_state & NEWBLOCK) != 0)
7198 free_newdirblk(pagedep->pd_newdirblk);
7199 if (free_pagedep(pagedep) == 0)
7200 panic("Failed to free pagedep %p", pagedep);
7205 * Reclaim any dependency structures from a buffer that is about to
7206 * be reallocated to a new vnode. The buffer must be locked, thus,
7207 * no I/O completion operations can occur while we are manipulating
7208 * its associated dependencies. The mutex is held so that other I/O's
7209 * associated with related dependencies do not occur.
7212 deallocate_dependencies(bp, freeblks, off)
7214 struct freeblks *freeblks;
7217 struct indirdep *indirdep;
7218 struct pagedep *pagedep;
7219 struct worklist *wk, *wkn;
7220 struct ufsmount *ump;
7222 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
7224 ump = VFSTOUFS(wk->wk_mp);
7226 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) {
7227 switch (wk->wk_type) {
7229 indirdep = WK_INDIRDEP(wk);
7230 if (bp->b_lblkno >= 0 ||
7231 bp->b_blkno != indirdep->ir_savebp->b_lblkno)
7232 panic("deallocate_dependencies: not indir");
7233 cancel_indirdep(indirdep, bp, freeblks);
7237 pagedep = WK_PAGEDEP(wk);
7238 if (cancel_pagedep(pagedep, freeblks, off)) {
7246 * Simply remove the allocindir, we'll find it via
7247 * the indirdep where we can clear pointers if
7250 WORKLIST_REMOVE(wk);
7255 * A truncation is waiting for the zero'd pointers
7256 * to be written. It can be freed when the freeblks
7259 WORKLIST_REMOVE(wk);
7260 wk->wk_state |= ONDEPLIST;
7261 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
7269 panic("deallocate_dependencies: Unexpected type %s",
7270 TYPENAME(wk->wk_type));
7277 * Don't throw away this buf, we were partially truncating and
7278 * some deps may always remain.
7282 bp->b_vflags |= BV_SCANNED;
7285 bp->b_flags |= B_INVAL | B_NOCACHE;
7291 * An allocdirect is being canceled due to a truncate. We must make sure
7292 * the journal entry is released in concert with the blkfree that releases
7293 * the storage. Completed journal entries must not be released until the
7294 * space is no longer pointed to by the inode or in the bitmap.
7297 cancel_allocdirect(adphead, adp, freeblks)
7298 struct allocdirectlst *adphead;
7299 struct allocdirect *adp;
7300 struct freeblks *freeblks;
7302 struct freework *freework;
7303 struct newblk *newblk;
7304 struct worklist *wk;
7306 TAILQ_REMOVE(adphead, adp, ad_next);
7307 newblk = (struct newblk *)adp;
7310 * Find the correct freework structure.
7312 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
7313 if (wk->wk_type != D_FREEWORK)
7315 freework = WK_FREEWORK(wk);
7316 if (freework->fw_blkno == newblk->nb_newblkno)
7319 if (freework == NULL)
7320 panic("cancel_allocdirect: Freework not found");
7322 * If a newblk exists at all we still have the journal entry that
7323 * initiated the allocation so we do not need to journal the free.
7325 cancel_jfreeblk(freeblks, freework->fw_blkno);
7327 * If the journal hasn't been written the jnewblk must be passed
7328 * to the call to ffs_blkfree that reclaims the space. We accomplish
7329 * this by linking the journal dependency into the freework to be
7330 * freed when freework_freeblock() is called. If the journal has
7331 * been written we can simply reclaim the journal space when the
7332 * freeblks work is complete.
7334 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list,
7335 &freeblks->fb_jwork);
7336 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
7341 * Cancel a new block allocation. May be an indirect or direct block. We
7342 * remove it from various lists and return any journal record that needs to
7343 * be resolved by the caller.
7345 * A special consideration is made for indirects which were never pointed
7346 * at on disk and will never be found once this block is released.
7348 static struct jnewblk *
7349 cancel_newblk(newblk, wk, wkhd)
7350 struct newblk *newblk;
7351 struct worklist *wk;
7352 struct workhead *wkhd;
7354 struct jnewblk *jnewblk;
7356 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno);
7358 newblk->nb_state |= GOINGAWAY;
7360 * Previously we traversed the completedhd on each indirdep
7361 * attached to this newblk to cancel them and gather journal
7362 * work. Since we need only the oldest journal segment and
7363 * the lowest point on the tree will always have the oldest
7364 * journal segment we are free to release the segments
7365 * of any subordinates and may leave the indirdep list to
7366 * indirdep_complete() when this newblk is freed.
7368 if (newblk->nb_state & ONDEPLIST) {
7369 newblk->nb_state &= ~ONDEPLIST;
7370 LIST_REMOVE(newblk, nb_deps);
7372 if (newblk->nb_state & ONWORKLIST)
7373 WORKLIST_REMOVE(&newblk->nb_list);
7375 * If the journal entry hasn't been written we save a pointer to
7376 * the dependency that frees it until it is written or the
7377 * superseding operation completes.
7379 jnewblk = newblk->nb_jnewblk;
7380 if (jnewblk != NULL && wk != NULL) {
7381 newblk->nb_jnewblk = NULL;
7382 jnewblk->jn_dep = wk;
7384 if (!LIST_EMPTY(&newblk->nb_jwork))
7385 jwork_move(wkhd, &newblk->nb_jwork);
7387 * When truncating we must free the newdirblk early to remove
7388 * the pagedep from the hash before returning.
7390 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7391 free_newdirblk(WK_NEWDIRBLK(wk));
7392 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7393 panic("cancel_newblk: extra newdirblk");
7399 * Schedule the freefrag associated with a newblk to be released once
7400 * the pointers are written and the previous block is no longer needed.
7403 newblk_freefrag(newblk)
7404 struct newblk *newblk;
7406 struct freefrag *freefrag;
7408 if (newblk->nb_freefrag == NULL)
7410 freefrag = newblk->nb_freefrag;
7411 newblk->nb_freefrag = NULL;
7412 freefrag->ff_state |= COMPLETE;
7413 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
7414 add_to_worklist(&freefrag->ff_list, 0);
7418 * Free a newblk. Generate a new freefrag work request if appropriate.
7419 * This must be called after the inode pointer and any direct block pointers
7420 * are valid or fully removed via truncate or frag extension.
7424 struct newblk *newblk;
7426 struct indirdep *indirdep;
7427 struct worklist *wk;
7429 KASSERT(newblk->nb_jnewblk == NULL,
7430 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk));
7431 KASSERT(newblk->nb_list.wk_type != D_NEWBLK,
7432 ("free_newblk: unclaimed newblk"));
7433 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp));
7434 newblk_freefrag(newblk);
7435 if (newblk->nb_state & ONDEPLIST)
7436 LIST_REMOVE(newblk, nb_deps);
7437 if (newblk->nb_state & ONWORKLIST)
7438 WORKLIST_REMOVE(&newblk->nb_list);
7439 LIST_REMOVE(newblk, nb_hash);
7440 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
7441 free_newdirblk(WK_NEWDIRBLK(wk));
7442 if (!LIST_EMPTY(&newblk->nb_newdirblk))
7443 panic("free_newblk: extra newdirblk");
7444 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL)
7445 indirdep_complete(indirdep);
7446 handle_jwork(&newblk->nb_jwork);
7447 WORKITEM_FREE(newblk, D_NEWBLK);
7451 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
7452 * This routine must be called with splbio interrupts blocked.
7455 free_newdirblk(newdirblk)
7456 struct newdirblk *newdirblk;
7458 struct pagedep *pagedep;
7460 struct worklist *wk;
7462 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp));
7463 WORKLIST_REMOVE(&newdirblk->db_list);
7465 * If the pagedep is still linked onto the directory buffer
7466 * dependency chain, then some of the entries on the
7467 * pd_pendinghd list may not be committed to disk yet. In
7468 * this case, we will simply clear the NEWBLOCK flag and
7469 * let the pd_pendinghd list be processed when the pagedep
7470 * is next written. If the pagedep is no longer on the buffer
7471 * dependency chain, then all the entries on the pd_pending
7472 * list are committed to disk and we can free them here.
7474 pagedep = newdirblk->db_pagedep;
7475 pagedep->pd_state &= ~NEWBLOCK;
7476 if ((pagedep->pd_state & ONWORKLIST) == 0) {
7477 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
7478 free_diradd(dap, NULL);
7480 * If no dependencies remain, the pagedep will be freed.
7482 free_pagedep(pagedep);
7484 /* Should only ever be one item in the list. */
7485 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
7486 WORKLIST_REMOVE(wk);
7487 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
7489 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
7493 * Prepare an inode to be freed. The actual free operation is not
7494 * done until the zero'ed inode has been written to disk.
7497 softdep_freefile(pvp, ino, mode)
7502 struct inode *ip = VTOI(pvp);
7503 struct inodedep *inodedep;
7504 struct freefile *freefile;
7505 struct freeblks *freeblks;
7506 struct ufsmount *ump;
7509 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
7510 ("softdep_freefile called on non-softdep filesystem"));
7512 * This sets up the inode de-allocation dependency.
7514 freefile = malloc(sizeof(struct freefile),
7515 M_FREEFILE, M_SOFTDEP_FLAGS);
7516 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
7517 freefile->fx_mode = mode;
7518 freefile->fx_oldinum = ino;
7519 freefile->fx_devvp = ip->i_devvp;
7520 LIST_INIT(&freefile->fx_jwork);
7522 ip->i_fs->fs_pendinginodes += 1;
7526 * If the inodedep does not exist, then the zero'ed inode has
7527 * been written to disk. If the allocated inode has never been
7528 * written to disk, then the on-disk inode is zero'ed. In either
7529 * case we can free the file immediately. If the journal was
7530 * canceled before being written the inode will never make it to
7531 * disk and we must send the canceled journal entrys to
7532 * ffs_freefile() to be cleared in conjunction with the bitmap.
7533 * Any blocks waiting on the inode to write can be safely freed
7534 * here as it will never been written.
7537 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7540 * Clear out freeblks that no longer need to reference
7544 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) {
7545 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks,
7547 freeblks->fb_state &= ~ONDEPLIST;
7550 * Remove this inode from the unlinked list.
7552 if (inodedep->id_state & UNLINKED) {
7554 * Save the journal work to be freed with the bitmap
7555 * before we clear UNLINKED. Otherwise it can be lost
7556 * if the inode block is written.
7558 handle_bufwait(inodedep, &freefile->fx_jwork);
7559 clear_unlinked_inodedep(inodedep);
7561 * Re-acquire inodedep as we've dropped the
7562 * per-filesystem lock in clear_unlinked_inodedep().
7564 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
7567 if (inodedep == NULL || check_inode_unwritten(inodedep)) {
7569 handle_workitem_freefile(freefile);
7572 if ((inodedep->id_state & DEPCOMPLETE) == 0)
7573 inodedep->id_state |= GOINGAWAY;
7574 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
7576 if (ip->i_number == ino)
7577 ip->i_flag |= IN_MODIFIED;
7581 * Check to see if an inode has never been written to disk. If
7582 * so free the inodedep and return success, otherwise return failure.
7583 * This routine must be called with splbio interrupts blocked.
7585 * If we still have a bitmap dependency, then the inode has never
7586 * been written to disk. Drop the dependency as it is no longer
7587 * necessary since the inode is being deallocated. We set the
7588 * ALLCOMPLETE flags since the bitmap now properly shows that the
7589 * inode is not allocated. Even if the inode is actively being
7590 * written, it has been rolled back to its zero'ed state, so we
7591 * are ensured that a zero inode is what is on the disk. For short
7592 * lived files, this change will usually result in removing all the
7593 * dependencies from the inode so that it can be freed immediately.
7596 check_inode_unwritten(inodedep)
7597 struct inodedep *inodedep;
7600 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7602 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
7603 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7604 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7605 !LIST_EMPTY(&inodedep->id_bufwait) ||
7606 !LIST_EMPTY(&inodedep->id_inowait) ||
7607 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7608 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7609 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7610 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7611 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7612 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7613 inodedep->id_mkdiradd != NULL ||
7614 inodedep->id_nlinkdelta != 0)
7617 * Another process might be in initiate_write_inodeblock_ufs[12]
7618 * trying to allocate memory without holding "Softdep Lock".
7620 if ((inodedep->id_state & IOSTARTED) != 0 &&
7621 inodedep->id_savedino1 == NULL)
7624 if (inodedep->id_state & ONDEPLIST)
7625 LIST_REMOVE(inodedep, id_deps);
7626 inodedep->id_state &= ~ONDEPLIST;
7627 inodedep->id_state |= ALLCOMPLETE;
7628 inodedep->id_bmsafemap = NULL;
7629 if (inodedep->id_state & ONWORKLIST)
7630 WORKLIST_REMOVE(&inodedep->id_list);
7631 if (inodedep->id_savedino1 != NULL) {
7632 free(inodedep->id_savedino1, M_SAVEDINO);
7633 inodedep->id_savedino1 = NULL;
7635 if (free_inodedep(inodedep) == 0)
7636 panic("check_inode_unwritten: busy inode");
7641 check_inodedep_free(inodedep)
7642 struct inodedep *inodedep;
7645 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7646 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
7647 !LIST_EMPTY(&inodedep->id_dirremhd) ||
7648 !LIST_EMPTY(&inodedep->id_pendinghd) ||
7649 !LIST_EMPTY(&inodedep->id_bufwait) ||
7650 !LIST_EMPTY(&inodedep->id_inowait) ||
7651 !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
7652 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
7653 !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
7654 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
7655 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
7656 !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
7657 inodedep->id_mkdiradd != NULL ||
7658 inodedep->id_nlinkdelta != 0 ||
7659 inodedep->id_savedino1 != NULL)
7665 * Try to free an inodedep structure. Return 1 if it could be freed.
7668 free_inodedep(inodedep)
7669 struct inodedep *inodedep;
7672 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp));
7673 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
7674 !check_inodedep_free(inodedep))
7676 if (inodedep->id_state & ONDEPLIST)
7677 LIST_REMOVE(inodedep, id_deps);
7678 LIST_REMOVE(inodedep, id_hash);
7679 WORKITEM_FREE(inodedep, D_INODEDEP);
7684 * Free the block referenced by a freework structure. The parent freeblks
7685 * structure is released and completed when the final cg bitmap reaches
7686 * the disk. This routine may be freeing a jnewblk which never made it to
7687 * disk in which case we do not have to wait as the operation is undone
7688 * in memory immediately.
7691 freework_freeblock(freework)
7692 struct freework *freework;
7694 struct freeblks *freeblks;
7695 struct jnewblk *jnewblk;
7696 struct ufsmount *ump;
7697 struct workhead wkhd;
7702 ump = VFSTOUFS(freework->fw_list.wk_mp);
7705 * Handle partial truncate separately.
7707 if (freework->fw_indir) {
7708 complete_trunc_indir(freework);
7711 freeblks = freework->fw_freeblks;
7713 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0;
7714 bsize = lfragtosize(fs, freework->fw_frags);
7717 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives
7718 * on the indirblk hashtable and prevents premature freeing.
7720 freework->fw_state |= DEPCOMPLETE;
7722 * SUJ needs to wait for the segment referencing freed indirect
7723 * blocks to expire so that we know the checker will not confuse
7724 * a re-allocated indirect block with its old contents.
7726 if (needj && freework->fw_lbn <= -NDADDR)
7727 indirblk_insert(freework);
7729 * If we are canceling an existing jnewblk pass it to the free
7730 * routine, otherwise pass the freeblk which will ultimately
7731 * release the freeblks. If we're not journaling, we can just
7732 * free the freeblks immediately.
7734 jnewblk = freework->fw_jnewblk;
7735 if (jnewblk != NULL) {
7736 cancel_jnewblk(jnewblk, &wkhd);
7739 freework->fw_state |= DELAYEDFREE;
7740 freeblks->fb_cgwait++;
7741 WORKLIST_INSERT(&wkhd, &freework->fw_list);
7744 freeblks_free(ump, freeblks, btodb(bsize));
7746 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld",
7747 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize);
7748 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize,
7749 freeblks->fb_inum, freeblks->fb_vtype, &wkhd);
7752 * The jnewblk will be discarded and the bits in the map never
7753 * made it to disk. We can immediately free the freeblk.
7756 handle_written_freework(freework);
7760 * We enqueue freework items that need processing back on the freeblks and
7761 * add the freeblks to the worklist. This makes it easier to find all work
7762 * required to flush a truncation in process_truncates().
7765 freework_enqueue(freework)
7766 struct freework *freework;
7768 struct freeblks *freeblks;
7770 freeblks = freework->fw_freeblks;
7771 if ((freework->fw_state & INPROGRESS) == 0)
7772 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
7773 if ((freeblks->fb_state &
7774 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE &&
7775 LIST_EMPTY(&freeblks->fb_jblkdephd))
7776 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7780 * Start, continue, or finish the process of freeing an indirect block tree.
7781 * The free operation may be paused at any point with fw_off containing the
7782 * offset to restart from. This enables us to implement some flow control
7783 * for large truncates which may fan out and generate a huge number of
7787 handle_workitem_indirblk(freework)
7788 struct freework *freework;
7790 struct freeblks *freeblks;
7791 struct ufsmount *ump;
7794 freeblks = freework->fw_freeblks;
7795 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7797 if (freework->fw_state & DEPCOMPLETE) {
7798 handle_written_freework(freework);
7801 if (freework->fw_off == NINDIR(fs)) {
7802 freework_freeblock(freework);
7805 freework->fw_state |= INPROGRESS;
7807 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
7813 * Called when a freework structure attached to a cg buf is written. The
7814 * ref on either the parent or the freeblks structure is released and
7815 * the freeblks is added back to the worklist if there is more work to do.
7818 handle_written_freework(freework)
7819 struct freework *freework;
7821 struct freeblks *freeblks;
7822 struct freework *parent;
7824 freeblks = freework->fw_freeblks;
7825 parent = freework->fw_parent;
7826 if (freework->fw_state & DELAYEDFREE)
7827 freeblks->fb_cgwait--;
7828 freework->fw_state |= COMPLETE;
7829 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
7830 WORKITEM_FREE(freework, D_FREEWORK);
7832 if (--parent->fw_ref == 0)
7833 freework_enqueue(parent);
7836 if (--freeblks->fb_ref != 0)
7838 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) ==
7839 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd))
7840 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
7844 * This workitem routine performs the block de-allocation.
7845 * The workitem is added to the pending list after the updated
7846 * inode block has been written to disk. As mentioned above,
7847 * checks regarding the number of blocks de-allocated (compared
7848 * to the number of blocks allocated for the file) are also
7849 * performed in this function.
7852 handle_workitem_freeblocks(freeblks, flags)
7853 struct freeblks *freeblks;
7856 struct freework *freework;
7857 struct newblk *newblk;
7858 struct allocindir *aip;
7859 struct ufsmount *ump;
7860 struct worklist *wk;
7862 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd),
7863 ("handle_workitem_freeblocks: Journal entries not written."));
7864 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7866 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
7867 WORKLIST_REMOVE(wk);
7868 switch (wk->wk_type) {
7870 wk->wk_state |= COMPLETE;
7871 add_to_worklist(wk, 0);
7875 free_newblk(WK_NEWBLK(wk));
7879 aip = WK_ALLOCINDIR(wk);
7881 if (aip->ai_state & DELAYEDFREE) {
7883 freework = newfreework(ump, freeblks, NULL,
7884 aip->ai_lbn, aip->ai_newblkno,
7885 ump->um_fs->fs_frag, 0, 0);
7888 newblk = WK_NEWBLK(wk);
7889 if (newblk->nb_jnewblk) {
7890 freework->fw_jnewblk = newblk->nb_jnewblk;
7891 newblk->nb_jnewblk->jn_dep = &freework->fw_list;
7892 newblk->nb_jnewblk = NULL;
7894 free_newblk(newblk);
7898 freework = WK_FREEWORK(wk);
7899 if (freework->fw_lbn <= -NDADDR)
7900 handle_workitem_indirblk(freework);
7902 freework_freeblock(freework);
7905 panic("handle_workitem_freeblocks: Unknown type %s",
7906 TYPENAME(wk->wk_type));
7909 if (freeblks->fb_ref != 0) {
7910 freeblks->fb_state &= ~INPROGRESS;
7911 wake_worklist(&freeblks->fb_list);
7916 return handle_complete_freeblocks(freeblks, flags);
7921 * Handle completion of block free via truncate. This allows fs_pending
7922 * to track the actual free block count more closely than if we only updated
7923 * it at the end. We must be careful to handle cases where the block count
7924 * on free was incorrect.
7927 freeblks_free(ump, freeblks, blocks)
7928 struct ufsmount *ump;
7929 struct freeblks *freeblks;
7933 ufs2_daddr_t remain;
7936 remain = -freeblks->fb_chkcnt;
7937 freeblks->fb_chkcnt += blocks;
7939 if (remain < blocks)
7942 fs->fs_pendingblocks -= blocks;
7948 * Once all of the freework workitems are complete we can retire the
7949 * freeblocks dependency and any journal work awaiting completion. This
7950 * can not be called until all other dependencies are stable on disk.
7953 handle_complete_freeblocks(freeblks, flags)
7954 struct freeblks *freeblks;
7957 struct inodedep *inodedep;
7961 struct ufsmount *ump;
7964 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
7966 flags = LK_EXCLUSIVE | flags;
7967 spare = freeblks->fb_chkcnt;
7970 * If we did not release the expected number of blocks we may have
7971 * to adjust the inode block count here. Only do so if it wasn't
7972 * a truncation to zero and the modrev still matches.
7974 if (spare && freeblks->fb_len != 0) {
7975 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum,
7976 flags, &vp, FFSV_FORCEINSMQ) != 0)
7979 if (DIP(ip, i_modrev) == freeblks->fb_modrev) {
7980 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare);
7981 ip->i_flag |= IN_CHANGE;
7983 * We must wait so this happens before the
7984 * journal is reclaimed.
7992 fs->fs_pendingblocks += spare;
7998 quotaadj(freeblks->fb_quota, ump, -spare);
7999 quotarele(freeblks->fb_quota);
8002 if (freeblks->fb_state & ONDEPLIST) {
8003 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum,
8005 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next);
8006 freeblks->fb_state &= ~ONDEPLIST;
8007 if (TAILQ_EMPTY(&inodedep->id_freeblklst))
8008 free_inodedep(inodedep);
8011 * All of the freeblock deps must be complete prior to this call
8012 * so it's now safe to complete earlier outstanding journal entries.
8014 handle_jwork(&freeblks->fb_jwork);
8015 WORKITEM_FREE(freeblks, D_FREEBLKS);
8021 * Release blocks associated with the freeblks and stored in the indirect
8022 * block dbn. If level is greater than SINGLE, the block is an indirect block
8023 * and recursive calls to indirtrunc must be used to cleanse other indirect
8026 * This handles partial and complete truncation of blocks. Partial is noted
8027 * with goingaway == 0. In this case the freework is completed after the
8028 * zero'd indirects are written to disk. For full truncation the freework
8029 * is completed after the block is freed.
8032 indir_trunc(freework, dbn, lbn)
8033 struct freework *freework;
8037 struct freework *nfreework;
8038 struct workhead wkhd;
8039 struct freeblks *freeblks;
8042 struct indirdep *indirdep;
8043 struct ufsmount *ump;
8045 ufs2_daddr_t nb, nnb, *bap2;
8046 ufs_lbn_t lbnadd, nlbn;
8047 int i, nblocks, ufs1fmt;
8055 freeblks = freework->fw_freeblks;
8056 ump = VFSTOUFS(freeblks->fb_list.wk_mp);
8059 * Get buffer of block pointers to be freed. There are three cases:
8061 * 1) Partial truncate caches the indirdep pointer in the freework
8062 * which provides us a back copy to the save bp which holds the
8063 * pointers we want to clear. When this completes the zero
8064 * pointers are written to the real copy.
8065 * 2) The indirect is being completely truncated, cancel_indirdep()
8066 * eliminated the real copy and placed the indirdep on the saved
8067 * copy. The indirdep and buf are discarded when this completes.
8068 * 3) The indirect was not in memory, we read a copy off of the disk
8069 * using the devvp and drop and invalidate the buffer when we're
8074 if (freework->fw_indir != NULL) {
8076 indirdep = freework->fw_indir;
8077 bp = indirdep->ir_savebp;
8078 if (bp == NULL || bp->b_blkno != dbn)
8079 panic("indir_trunc: Bad saved buf %p blkno %jd",
8081 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) {
8083 * The lock prevents the buf dep list from changing and
8084 * indirects on devvp should only ever have one dependency.
8086 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep));
8087 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0)
8088 panic("indir_trunc: Bad indirdep %p from buf %p",
8090 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
8091 NOCRED, &bp) != 0) {
8096 /* Protects against a race with complete_trunc_indir(). */
8097 freework->fw_state &= ~INPROGRESS;
8099 * If we have an indirdep we need to enforce the truncation order
8100 * and discard it when it is complete.
8103 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) &&
8104 !TAILQ_EMPTY(&indirdep->ir_trunc)) {
8106 * Add the complete truncate to the list on the
8107 * indirdep to enforce in-order processing.
8109 if (freework->fw_indir == NULL)
8110 TAILQ_INSERT_TAIL(&indirdep->ir_trunc,
8116 * If we're goingaway, free the indirdep. Otherwise it will
8117 * linger until the write completes.
8120 free_indirdep(indirdep);
8123 /* Initialize pointers depending on block size. */
8124 if (ump->um_fstype == UFS1) {
8125 bap1 = (ufs1_daddr_t *)bp->b_data;
8126 nb = bap1[freework->fw_off];
8130 bap2 = (ufs2_daddr_t *)bp->b_data;
8131 nb = bap2[freework->fw_off];
8135 level = lbn_level(lbn);
8136 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0;
8137 lbnadd = lbn_offset(fs, level);
8138 nblocks = btodb(fs->fs_bsize);
8139 nfreework = freework;
8143 * Reclaim blocks. Traverses into nested indirect levels and
8144 * arranges for the current level to be freed when subordinates
8145 * are free when journaling.
8147 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
8148 if (i != NINDIR(fs) - 1) {
8159 nlbn = (lbn + 1) - (i * lbnadd);
8161 nfreework = newfreework(ump, freeblks, freework,
8162 nlbn, nb, fs->fs_frag, 0, 0);
8165 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
8167 struct freedep *freedep;
8170 * Attempt to aggregate freedep dependencies for
8171 * all blocks being released to the same CG.
8175 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
8176 freedep = newfreedep(freework);
8177 WORKLIST_INSERT_UNLOCKED(&wkhd,
8182 "indir_trunc: ino %d blkno %jd size %ld",
8183 freeblks->fb_inum, nb, fs->fs_bsize);
8184 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
8185 fs->fs_bsize, freeblks->fb_inum,
8186 freeblks->fb_vtype, &wkhd);
8190 bp->b_flags |= B_INVAL | B_NOCACHE;
8195 freedblocks = (nblocks * cnt);
8197 freedblocks += nblocks;
8198 freeblks_free(ump, freeblks, freedblocks);
8200 * If we are journaling set up the ref counts and offset so this
8201 * indirect can be completed when its children are free.
8205 freework->fw_off = i;
8206 freework->fw_ref += freedeps;
8207 freework->fw_ref -= NINDIR(fs) + 1;
8209 freeblks->fb_cgwait += freedeps;
8210 if (freework->fw_ref == 0)
8211 freework_freeblock(freework);
8216 * If we're not journaling we can free the indirect now.
8218 dbn = dbtofsb(fs, dbn);
8220 "indir_trunc 2: ino %d blkno %jd size %ld",
8221 freeblks->fb_inum, dbn, fs->fs_bsize);
8222 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
8223 freeblks->fb_inum, freeblks->fb_vtype, NULL);
8224 /* Non SUJ softdep does single-threaded truncations. */
8225 if (freework->fw_blkno == dbn) {
8226 freework->fw_state |= ALLCOMPLETE;
8228 handle_written_freework(freework);
8235 * Cancel an allocindir when it is removed via truncation. When bp is not
8236 * NULL the indirect never appeared on disk and is scheduled to be freed
8237 * independently of the indir so we can more easily track journal work.
8240 cancel_allocindir(aip, bp, freeblks, trunc)
8241 struct allocindir *aip;
8243 struct freeblks *freeblks;
8246 struct indirdep *indirdep;
8247 struct freefrag *freefrag;
8248 struct newblk *newblk;
8250 newblk = (struct newblk *)aip;
8251 LIST_REMOVE(aip, ai_next);
8253 * We must eliminate the pointer in bp if it must be freed on its
8254 * own due to partial truncate or pending journal work.
8256 if (bp && (trunc || newblk->nb_jnewblk)) {
8258 * Clear the pointer and mark the aip to be freed
8259 * directly if it never existed on disk.
8261 aip->ai_state |= DELAYEDFREE;
8262 indirdep = aip->ai_indirdep;
8263 if (indirdep->ir_state & UFS1FMT)
8264 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8266 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
8269 * When truncating the previous pointer will be freed via
8270 * savedbp. Eliminate the freefrag which would dup free.
8272 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) {
8273 newblk->nb_freefrag = NULL;
8274 if (freefrag->ff_jdep)
8276 WK_JFREEFRAG(freefrag->ff_jdep));
8277 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork);
8278 WORKITEM_FREE(freefrag, D_FREEFRAG);
8281 * If the journal hasn't been written the jnewblk must be passed
8282 * to the call to ffs_blkfree that reclaims the space. We accomplish
8283 * this by leaving the journal dependency on the newblk to be freed
8284 * when a freework is created in handle_workitem_freeblocks().
8286 cancel_newblk(newblk, NULL, &freeblks->fb_jwork);
8287 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
8291 * Create the mkdir dependencies for . and .. in a new directory. Link them
8292 * in to a newdirblk so any subsequent additions are tracked properly. The
8293 * caller is responsible for adding the mkdir1 dependency to the journal
8294 * and updating id_mkdiradd. This function returns with the per-filesystem
8297 static struct mkdir *
8298 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
8302 struct buf *newdirbp;
8303 struct mkdir **mkdirp;
8305 struct newblk *newblk;
8306 struct pagedep *pagedep;
8307 struct inodedep *inodedep;
8308 struct newdirblk *newdirblk;
8309 struct mkdir *mkdir1, *mkdir2;
8310 struct worklist *wk;
8311 struct jaddref *jaddref;
8312 struct ufsmount *ump;
8315 mp = dap->da_list.wk_mp;
8317 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
8319 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8320 LIST_INIT(&newdirblk->db_mkdir);
8321 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8322 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
8323 mkdir1->md_state = ATTACHED | MKDIR_BODY;
8324 mkdir1->md_diradd = dap;
8325 mkdir1->md_jaddref = NULL;
8326 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
8327 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
8328 mkdir2->md_state = ATTACHED | MKDIR_PARENT;
8329 mkdir2->md_diradd = dap;
8330 mkdir2->md_jaddref = NULL;
8331 if (MOUNTEDSUJ(mp) == 0) {
8332 mkdir1->md_state |= DEPCOMPLETE;
8333 mkdir2->md_state |= DEPCOMPLETE;
8336 * Dependency on "." and ".." being written to disk.
8338 mkdir1->md_buf = newdirbp;
8339 ACQUIRE_LOCK(VFSTOUFS(mp));
8340 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs);
8342 * We must link the pagedep, allocdirect, and newdirblk for
8343 * the initial file page so the pointer to the new directory
8344 * is not written until the directory contents are live and
8345 * any subsequent additions are not marked live until the
8346 * block is reachable via the inode.
8348 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0)
8349 panic("setup_newdir: lost pagedep");
8350 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
8351 if (wk->wk_type == D_ALLOCDIRECT)
8354 panic("setup_newdir: lost allocdirect");
8355 if (pagedep->pd_state & NEWBLOCK)
8356 panic("setup_newdir: NEWBLOCK already set");
8357 newblk = WK_NEWBLK(wk);
8358 pagedep->pd_state |= NEWBLOCK;
8359 pagedep->pd_newdirblk = newdirblk;
8360 newdirblk->db_pagedep = pagedep;
8361 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8362 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
8364 * Look up the inodedep for the parent directory so that we
8365 * can link mkdir2 into the pending dotdot jaddref or
8366 * the inode write if there is none. If the inode is
8367 * ALLCOMPLETE and no jaddref is present all dependencies have
8368 * been satisfied and mkdir2 can be freed.
8370 inodedep_lookup(mp, dinum, 0, &inodedep);
8371 if (MOUNTEDSUJ(mp)) {
8372 if (inodedep == NULL)
8373 panic("setup_newdir: Lost parent.");
8374 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8376 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
8377 (jaddref->ja_state & MKDIR_PARENT),
8378 ("setup_newdir: bad dotdot jaddref %p", jaddref));
8379 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8380 mkdir2->md_jaddref = jaddref;
8381 jaddref->ja_mkdir = mkdir2;
8382 } else if (inodedep == NULL ||
8383 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
8384 dap->da_state &= ~MKDIR_PARENT;
8385 WORKITEM_FREE(mkdir2, D_MKDIR);
8388 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs);
8389 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list);
8397 * Directory entry addition dependencies.
8399 * When adding a new directory entry, the inode (with its incremented link
8400 * count) must be written to disk before the directory entry's pointer to it.
8401 * Also, if the inode is newly allocated, the corresponding freemap must be
8402 * updated (on disk) before the directory entry's pointer. These requirements
8403 * are met via undo/redo on the directory entry's pointer, which consists
8404 * simply of the inode number.
8406 * As directory entries are added and deleted, the free space within a
8407 * directory block can become fragmented. The ufs filesystem will compact
8408 * a fragmented directory block to make space for a new entry. When this
8409 * occurs, the offsets of previously added entries change. Any "diradd"
8410 * dependency structures corresponding to these entries must be updated with
8415 * This routine is called after the in-memory inode's link
8416 * count has been incremented, but before the directory entry's
8417 * pointer to the inode has been set.
8420 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
8421 struct buf *bp; /* buffer containing directory block */
8422 struct inode *dp; /* inode for directory */
8423 off_t diroffset; /* offset of new entry in directory */
8424 ino_t newinum; /* inode referenced by new directory entry */
8425 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
8426 int isnewblk; /* entry is in a newly allocated block */
8428 int offset; /* offset of new entry within directory block */
8429 ufs_lbn_t lbn; /* block in directory containing new entry */
8432 struct newblk *newblk;
8433 struct pagedep *pagedep;
8434 struct inodedep *inodedep;
8435 struct newdirblk *newdirblk;
8436 struct mkdir *mkdir1, *mkdir2;
8437 struct jaddref *jaddref;
8438 struct ufsmount *ump;
8444 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8445 ("softdep_setup_directory_add called on non-softdep filesystem"));
8447 * Whiteouts have no dependencies.
8449 if (newinum == WINO) {
8450 if (newdirbp != NULL)
8455 mkdir1 = mkdir2 = NULL;
8457 lbn = lblkno(fs, diroffset);
8458 offset = blkoff(fs, diroffset);
8459 dap = malloc(sizeof(struct diradd), M_DIRADD,
8460 M_SOFTDEP_FLAGS|M_ZERO);
8461 workitem_alloc(&dap->da_list, D_DIRADD, mp);
8462 dap->da_offset = offset;
8463 dap->da_newinum = newinum;
8464 dap->da_state = ATTACHED;
8465 LIST_INIT(&dap->da_jwork);
8466 isindir = bp->b_lblkno >= NDADDR;
8469 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
8470 newdirblk = malloc(sizeof(struct newdirblk),
8471 M_NEWDIRBLK, M_SOFTDEP_FLAGS);
8472 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
8473 LIST_INIT(&newdirblk->db_mkdir);
8476 * If we're creating a new directory setup the dependencies and set
8477 * the dap state to wait for them. Otherwise it's COMPLETE and
8480 if (newdirbp == NULL) {
8481 dap->da_state |= DEPCOMPLETE;
8484 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
8485 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
8489 * Link into parent directory pagedep to await its being written.
8491 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep);
8493 if (diradd_lookup(pagedep, offset) != NULL)
8494 panic("softdep_setup_directory_add: %p already at off %d\n",
8495 diradd_lookup(pagedep, offset), offset);
8497 dap->da_pagedep = pagedep;
8498 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
8500 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
8502 * If we're journaling, link the diradd into the jaddref so it
8503 * may be completed after the journal entry is written. Otherwise,
8504 * link the diradd into its inodedep. If the inode is not yet
8505 * written place it on the bufwait list, otherwise do the post-inode
8506 * write processing to put it on the id_pendinghd list.
8508 if (MOUNTEDSUJ(mp)) {
8509 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
8511 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
8512 ("softdep_setup_directory_add: bad jaddref %p", jaddref));
8513 jaddref->ja_diroff = diroffset;
8514 jaddref->ja_diradd = dap;
8515 add_to_journal(&jaddref->ja_list);
8516 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
8517 diradd_inode_written(dap, inodedep);
8519 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
8521 * Add the journal entries for . and .. links now that the primary
8524 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) {
8525 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
8526 inoreflst, if_deps);
8527 KASSERT(jaddref != NULL &&
8528 jaddref->ja_ino == jaddref->ja_parent &&
8529 (jaddref->ja_state & MKDIR_BODY),
8530 ("softdep_setup_directory_add: bad dot jaddref %p",
8532 mkdir1->md_jaddref = jaddref;
8533 jaddref->ja_mkdir = mkdir1;
8535 * It is important that the dotdot journal entry
8536 * is added prior to the dot entry since dot writes
8537 * both the dot and dotdot links. These both must
8538 * be added after the primary link for the journal
8539 * to remain consistent.
8541 add_to_journal(&mkdir2->md_jaddref->ja_list);
8542 add_to_journal(&jaddref->ja_list);
8545 * If we are adding a new directory remember this diradd so that if
8546 * we rename it we can keep the dot and dotdot dependencies. If
8547 * we are adding a new name for an inode that has a mkdiradd we
8548 * must be in rename and we have to move the dot and dotdot
8549 * dependencies to this new name. The old name is being orphaned
8552 if (mkdir1 != NULL) {
8553 if (inodedep->id_mkdiradd != NULL)
8554 panic("softdep_setup_directory_add: Existing mkdir");
8555 inodedep->id_mkdiradd = dap;
8556 } else if (inodedep->id_mkdiradd)
8557 merge_diradd(inodedep, dap);
8558 if (newdirblk != NULL) {
8560 * There is nothing to do if we are already tracking
8563 if ((pagedep->pd_state & NEWBLOCK) != 0) {
8564 WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
8568 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
8570 panic("softdep_setup_directory_add: lost entry");
8571 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
8572 pagedep->pd_state |= NEWBLOCK;
8573 pagedep->pd_newdirblk = newdirblk;
8574 newdirblk->db_pagedep = pagedep;
8577 * If we extended into an indirect signal direnter to sync.
8588 * This procedure is called to change the offset of a directory
8589 * entry when compacting a directory block which must be owned
8590 * exclusively by the caller. Note that the actual entry movement
8591 * must be done in this procedure to ensure that no I/O completions
8592 * occur while the move is in progress.
8595 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
8596 struct buf *bp; /* Buffer holding directory block. */
8597 struct inode *dp; /* inode for directory */
8598 caddr_t base; /* address of dp->i_offset */
8599 caddr_t oldloc; /* address of old directory location */
8600 caddr_t newloc; /* address of new directory location */
8601 int entrysize; /* size of directory entry */
8603 int offset, oldoffset, newoffset;
8604 struct pagedep *pagedep;
8605 struct jmvref *jmvref;
8612 mp = UFSTOVFS(dp->i_ump);
8613 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
8614 ("softdep_change_directoryentry_offset called on "
8615 "non-softdep filesystem"));
8616 de = (struct direct *)oldloc;
8620 * Moves are always journaled as it would be too complex to
8621 * determine if any affected adds or removes are present in the
8624 if (MOUNTEDSUJ(mp)) {
8626 jmvref = newjmvref(dp, de->d_ino,
8627 dp->i_offset + (oldloc - base),
8628 dp->i_offset + (newloc - base));
8630 lbn = lblkno(dp->i_fs, dp->i_offset);
8631 offset = blkoff(dp->i_fs, dp->i_offset);
8632 oldoffset = offset + (oldloc - base);
8633 newoffset = offset + (newloc - base);
8634 ACQUIRE_LOCK(dp->i_ump);
8635 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0)
8637 dap = diradd_lookup(pagedep, oldoffset);
8639 dap->da_offset = newoffset;
8640 newoffset = DIRADDHASH(newoffset);
8641 oldoffset = DIRADDHASH(oldoffset);
8642 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
8643 newoffset != oldoffset) {
8644 LIST_REMOVE(dap, da_pdlist);
8645 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
8651 jmvref->jm_pagedep = pagedep;
8652 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
8653 add_to_journal(&jmvref->jm_list);
8655 bcopy(oldloc, newloc, entrysize);
8656 FREE_LOCK(dp->i_ump);
8660 * Move the mkdir dependencies and journal work from one diradd to another
8661 * when renaming a directory. The new name must depend on the mkdir deps
8662 * completing as the old name did. Directories can only have one valid link
8663 * at a time so one must be canonical.
8666 merge_diradd(inodedep, newdap)
8667 struct inodedep *inodedep;
8668 struct diradd *newdap;
8670 struct diradd *olddap;
8671 struct mkdir *mkdir, *nextmd;
8672 struct ufsmount *ump;
8675 olddap = inodedep->id_mkdiradd;
8676 inodedep->id_mkdiradd = newdap;
8677 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8678 newdap->da_state &= ~DEPCOMPLETE;
8679 ump = VFSTOUFS(inodedep->id_list.wk_mp);
8680 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8682 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8683 if (mkdir->md_diradd != olddap)
8685 mkdir->md_diradd = newdap;
8686 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
8687 newdap->da_state |= state;
8688 olddap->da_state &= ~state;
8689 if ((olddap->da_state &
8690 (MKDIR_PARENT | MKDIR_BODY)) == 0)
8693 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8694 panic("merge_diradd: unfound ref");
8697 * Any mkdir related journal items are not safe to be freed until
8698 * the new name is stable.
8700 jwork_move(&newdap->da_jwork, &olddap->da_jwork);
8701 olddap->da_state |= DEPCOMPLETE;
8702 complete_diradd(olddap);
8706 * Move the diradd to the pending list when all diradd dependencies are
8710 complete_diradd(dap)
8713 struct pagedep *pagedep;
8715 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
8716 if (dap->da_state & DIRCHG)
8717 pagedep = dap->da_previous->dm_pagedep;
8719 pagedep = dap->da_pagedep;
8720 LIST_REMOVE(dap, da_pdlist);
8721 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
8726 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal
8727 * add entries and conditonally journal the remove.
8730 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
8732 struct dirrem *dirrem;
8733 struct jremref *jremref;
8734 struct jremref *dotremref;
8735 struct jremref *dotdotremref;
8737 struct inodedep *inodedep;
8738 struct jaddref *jaddref;
8739 struct inoref *inoref;
8740 struct ufsmount *ump;
8741 struct mkdir *mkdir;
8744 * If no remove references were allocated we're on a non-journaled
8745 * filesystem and can skip the cancel step.
8747 if (jremref == NULL) {
8748 free_diradd(dap, NULL);
8752 * Cancel the primary name an free it if it does not require
8755 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
8756 0, &inodedep) != 0) {
8757 /* Abort the addref that reference this diradd. */
8758 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
8759 if (inoref->if_list.wk_type != D_JADDREF)
8761 jaddref = (struct jaddref *)inoref;
8762 if (jaddref->ja_diradd != dap)
8764 if (cancel_jaddref(jaddref, inodedep,
8765 &dirrem->dm_jwork) == 0) {
8766 free_jremref(jremref);
8773 * Cancel subordinate names and free them if they do not require
8776 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8777 ump = VFSTOUFS(dap->da_list.wk_mp);
8778 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) {
8779 if (mkdir->md_diradd != dap)
8781 if ((jaddref = mkdir->md_jaddref) == NULL)
8783 mkdir->md_jaddref = NULL;
8784 if (mkdir->md_state & MKDIR_PARENT) {
8785 if (cancel_jaddref(jaddref, NULL,
8786 &dirrem->dm_jwork) == 0) {
8787 free_jremref(dotdotremref);
8788 dotdotremref = NULL;
8791 if (cancel_jaddref(jaddref, inodedep,
8792 &dirrem->dm_jwork) == 0) {
8793 free_jremref(dotremref);
8801 journal_jremref(dirrem, jremref, inodedep);
8803 journal_jremref(dirrem, dotremref, inodedep);
8805 journal_jremref(dirrem, dotdotremref, NULL);
8806 jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
8807 free_diradd(dap, &dirrem->dm_jwork);
8811 * Free a diradd dependency structure. This routine must be called
8812 * with splbio interrupts blocked.
8815 free_diradd(dap, wkhd)
8817 struct workhead *wkhd;
8819 struct dirrem *dirrem;
8820 struct pagedep *pagedep;
8821 struct inodedep *inodedep;
8822 struct mkdir *mkdir, *nextmd;
8823 struct ufsmount *ump;
8825 ump = VFSTOUFS(dap->da_list.wk_mp);
8827 LIST_REMOVE(dap, da_pdlist);
8828 if (dap->da_state & ONWORKLIST)
8829 WORKLIST_REMOVE(&dap->da_list);
8830 if ((dap->da_state & DIRCHG) == 0) {
8831 pagedep = dap->da_pagedep;
8833 dirrem = dap->da_previous;
8834 pagedep = dirrem->dm_pagedep;
8835 dirrem->dm_dirinum = pagedep->pd_ino;
8836 dirrem->dm_state |= COMPLETE;
8837 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
8838 add_to_worklist(&dirrem->dm_list, 0);
8840 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
8842 if (inodedep->id_mkdiradd == dap)
8843 inodedep->id_mkdiradd = NULL;
8844 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
8845 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
8847 nextmd = LIST_NEXT(mkdir, md_mkdirs);
8848 if (mkdir->md_diradd != dap)
8851 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
8852 LIST_REMOVE(mkdir, md_mkdirs);
8853 if (mkdir->md_state & ONWORKLIST)
8854 WORKLIST_REMOVE(&mkdir->md_list);
8855 if (mkdir->md_jaddref != NULL)
8856 panic("free_diradd: Unexpected jaddref");
8857 WORKITEM_FREE(mkdir, D_MKDIR);
8858 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
8861 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
8862 panic("free_diradd: unfound ref");
8865 free_inodedep(inodedep);
8867 * Free any journal segments waiting for the directory write.
8869 handle_jwork(&dap->da_jwork);
8870 WORKITEM_FREE(dap, D_DIRADD);
8874 * Directory entry removal dependencies.
8876 * When removing a directory entry, the entry's inode pointer must be
8877 * zero'ed on disk before the corresponding inode's link count is decremented
8878 * (possibly freeing the inode for re-use). This dependency is handled by
8879 * updating the directory entry but delaying the inode count reduction until
8880 * after the directory block has been written to disk. After this point, the
8881 * inode count can be decremented whenever it is convenient.
8885 * This routine should be called immediately after removing
8886 * a directory entry. The inode's link count should not be
8887 * decremented by the calling procedure -- the soft updates
8888 * code will do this task when it is safe.
8891 softdep_setup_remove(bp, dp, ip, isrmdir)
8892 struct buf *bp; /* buffer containing directory block */
8893 struct inode *dp; /* inode for the directory being modified */
8894 struct inode *ip; /* inode for directory entry being removed */
8895 int isrmdir; /* indicates if doing RMDIR */
8897 struct dirrem *dirrem, *prevdirrem;
8898 struct inodedep *inodedep;
8901 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
8902 ("softdep_setup_remove called on non-softdep filesystem"));
8904 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want
8905 * newdirrem() to setup the full directory remove which requires
8908 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
8910 * Add the dirrem to the inodedep's pending remove list for quick
8913 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
8915 panic("softdep_setup_remove: Lost inodedep.");
8916 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
8917 dirrem->dm_state |= ONDEPLIST;
8918 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
8921 * If the COMPLETE flag is clear, then there were no active
8922 * entries and we want to roll back to a zeroed entry until
8923 * the new inode is committed to disk. If the COMPLETE flag is
8924 * set then we have deleted an entry that never made it to
8925 * disk. If the entry we deleted resulted from a name change,
8926 * then the old name still resides on disk. We cannot delete
8927 * its inode (returned to us in prevdirrem) until the zeroed
8928 * directory entry gets to disk. The new inode has never been
8929 * referenced on the disk, so can be deleted immediately.
8931 if ((dirrem->dm_state & COMPLETE) == 0) {
8932 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
8934 FREE_LOCK(ip->i_ump);
8936 if (prevdirrem != NULL)
8937 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
8938 prevdirrem, dm_next);
8939 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
8940 direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
8941 FREE_LOCK(ip->i_ump);
8943 handle_workitem_remove(dirrem, 0);
8948 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
8949 * pd_pendinghd list of a pagedep.
8951 static struct diradd *
8952 diradd_lookup(pagedep, offset)
8953 struct pagedep *pagedep;
8958 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
8959 if (dap->da_offset == offset)
8961 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
8962 if (dap->da_offset == offset)
8968 * Search for a .. diradd dependency in a directory that is being removed.
8969 * If the directory was renamed to a new parent we have a diradd rather
8970 * than a mkdir for the .. entry. We need to cancel it now before
8971 * it is found in truncate().
8973 static struct jremref *
8974 cancel_diradd_dotdot(ip, dirrem, jremref)
8976 struct dirrem *dirrem;
8977 struct jremref *jremref;
8979 struct pagedep *pagedep;
8981 struct worklist *wk;
8983 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0,
8986 dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
8989 cancel_diradd(dap, dirrem, jremref, NULL, NULL);
8991 * Mark any journal work as belonging to the parent so it is freed
8992 * with the .. reference.
8994 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
8995 wk->wk_state |= MKDIR_PARENT;
9000 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
9001 * replace it with a dirrem/diradd pair as a result of re-parenting a
9002 * directory. This ensures that we don't simultaneously have a mkdir and
9003 * a diradd for the same .. entry.
9005 static struct jremref *
9006 cancel_mkdir_dotdot(ip, dirrem, jremref)
9008 struct dirrem *dirrem;
9009 struct jremref *jremref;
9011 struct inodedep *inodedep;
9012 struct jaddref *jaddref;
9013 struct ufsmount *ump;
9014 struct mkdir *mkdir;
9017 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
9020 dap = inodedep->id_mkdiradd;
9021 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
9023 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9024 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir;
9025 mkdir = LIST_NEXT(mkdir, md_mkdirs))
9026 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
9029 panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
9030 if ((jaddref = mkdir->md_jaddref) != NULL) {
9031 mkdir->md_jaddref = NULL;
9032 jaddref->ja_state &= ~MKDIR_PARENT;
9033 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0,
9035 panic("cancel_mkdir_dotdot: Lost parent inodedep");
9036 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
9037 journal_jremref(dirrem, jremref, inodedep);
9041 if (mkdir->md_state & ONWORKLIST)
9042 WORKLIST_REMOVE(&mkdir->md_list);
9043 mkdir->md_state |= ALLCOMPLETE;
9044 complete_mkdir(mkdir);
9049 journal_jremref(dirrem, jremref, inodedep)
9050 struct dirrem *dirrem;
9051 struct jremref *jremref;
9052 struct inodedep *inodedep;
9055 if (inodedep == NULL)
9056 if (inodedep_lookup(jremref->jr_list.wk_mp,
9057 jremref->jr_ref.if_ino, 0, &inodedep) == 0)
9058 panic("journal_jremref: Lost inodedep");
9059 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
9060 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
9061 add_to_journal(&jremref->jr_list);
9065 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
9066 struct dirrem *dirrem;
9067 struct jremref *jremref;
9068 struct jremref *dotremref;
9069 struct jremref *dotdotremref;
9071 struct inodedep *inodedep;
9074 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
9076 panic("dirrem_journal: Lost inodedep");
9077 journal_jremref(dirrem, jremref, inodedep);
9079 journal_jremref(dirrem, dotremref, inodedep);
9081 journal_jremref(dirrem, dotdotremref, NULL);
9085 * Allocate a new dirrem if appropriate and return it along with
9086 * its associated pagedep. Called without a lock, returns with lock.
9088 static struct dirrem *
9089 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
9090 struct buf *bp; /* buffer containing directory block */
9091 struct inode *dp; /* inode for the directory being modified */
9092 struct inode *ip; /* inode for directory entry being removed */
9093 int isrmdir; /* indicates if doing RMDIR */
9094 struct dirrem **prevdirremp; /* previously referenced inode, if any */
9099 struct dirrem *dirrem;
9100 struct pagedep *pagedep;
9101 struct jremref *jremref;
9102 struct jremref *dotremref;
9103 struct jremref *dotdotremref;
9107 * Whiteouts have no deletion dependencies.
9110 panic("newdirrem: whiteout");
9113 * If the system is over its limit and our filesystem is
9114 * responsible for more than our share of that usage and
9115 * we are not a snapshot, request some inodedep cleanup.
9116 * Limiting the number of dirrem structures will also limit
9117 * the number of freefile and freeblks structures.
9119 ACQUIRE_LOCK(ip->i_ump);
9120 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ip->i_ump, D_DIRREM))
9121 schedule_cleanup(ITOV(dp)->v_mount);
9123 FREE_LOCK(ip->i_ump);
9124 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS |
9126 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
9127 LIST_INIT(&dirrem->dm_jremrefhd);
9128 LIST_INIT(&dirrem->dm_jwork);
9129 dirrem->dm_state = isrmdir ? RMDIR : 0;
9130 dirrem->dm_oldinum = ip->i_number;
9131 *prevdirremp = NULL;
9133 * Allocate remove reference structures to track journal write
9134 * dependencies. We will always have one for the link and
9135 * when doing directories we will always have one more for dot.
9136 * When renaming a directory we skip the dotdot link change so
9137 * this is not needed.
9139 jremref = dotremref = dotdotremref = NULL;
9140 if (DOINGSUJ(dvp)) {
9142 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9143 ip->i_effnlink + 2);
9144 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
9145 ip->i_effnlink + 1);
9146 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
9147 dp->i_effnlink + 1);
9148 dotdotremref->jr_state |= MKDIR_PARENT;
9150 jremref = newjremref(dirrem, dp, ip, dp->i_offset,
9151 ip->i_effnlink + 1);
9153 ACQUIRE_LOCK(ip->i_ump);
9154 lbn = lblkno(dp->i_fs, dp->i_offset);
9155 offset = blkoff(dp->i_fs, dp->i_offset);
9156 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC,
9158 dirrem->dm_pagedep = pagedep;
9159 dirrem->dm_offset = offset;
9161 * If we're renaming a .. link to a new directory, cancel any
9162 * existing MKDIR_PARENT mkdir. If it has already been canceled
9163 * the jremref is preserved for any potential diradd in this
9164 * location. This can not coincide with a rmdir.
9166 if (dp->i_offset == DOTDOT_OFFSET) {
9168 panic("newdirrem: .. directory change during remove?");
9169 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
9172 * If we're removing a directory search for the .. dependency now and
9173 * cancel it. Any pending journal work will be added to the dirrem
9174 * to be completed when the workitem remove completes.
9177 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
9179 * Check for a diradd dependency for the same directory entry.
9180 * If present, then both dependencies become obsolete and can
9183 dap = diradd_lookup(pagedep, offset);
9186 * Link the jremref structures into the dirrem so they are
9187 * written prior to the pagedep.
9190 dirrem_journal(dirrem, jremref, dotremref,
9195 * Must be ATTACHED at this point.
9197 if ((dap->da_state & ATTACHED) == 0)
9198 panic("newdirrem: not ATTACHED");
9199 if (dap->da_newinum != ip->i_number)
9200 panic("newdirrem: inum %ju should be %ju",
9201 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum);
9203 * If we are deleting a changed name that never made it to disk,
9204 * then return the dirrem describing the previous inode (which
9205 * represents the inode currently referenced from this entry on disk).
9207 if ((dap->da_state & DIRCHG) != 0) {
9208 *prevdirremp = dap->da_previous;
9209 dap->da_state &= ~DIRCHG;
9210 dap->da_pagedep = pagedep;
9213 * We are deleting an entry that never made it to disk.
9214 * Mark it COMPLETE so we can delete its inode immediately.
9216 dirrem->dm_state |= COMPLETE;
9217 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
9220 struct worklist *wk;
9222 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
9223 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
9224 panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
9232 * Directory entry change dependencies.
9234 * Changing an existing directory entry requires that an add operation
9235 * be completed first followed by a deletion. The semantics for the addition
9236 * are identical to the description of adding a new entry above except
9237 * that the rollback is to the old inode number rather than zero. Once
9238 * the addition dependency is completed, the removal is done as described
9239 * in the removal routine above.
9243 * This routine should be called immediately after changing
9244 * a directory entry. The inode's link count should not be
9245 * decremented by the calling procedure -- the soft updates
9246 * code will perform this task when it is safe.
9249 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
9250 struct buf *bp; /* buffer containing directory block */
9251 struct inode *dp; /* inode for the directory being modified */
9252 struct inode *ip; /* inode for directory entry being removed */
9253 ino_t newinum; /* new inode number for changed entry */
9254 int isrmdir; /* indicates if doing RMDIR */
9257 struct diradd *dap = NULL;
9258 struct dirrem *dirrem, *prevdirrem;
9259 struct pagedep *pagedep;
9260 struct inodedep *inodedep;
9261 struct jaddref *jaddref;
9264 offset = blkoff(dp->i_fs, dp->i_offset);
9265 mp = UFSTOVFS(dp->i_ump);
9266 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
9267 ("softdep_setup_directory_change called on non-softdep filesystem"));
9270 * Whiteouts do not need diradd dependencies.
9272 if (newinum != WINO) {
9273 dap = malloc(sizeof(struct diradd),
9274 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
9275 workitem_alloc(&dap->da_list, D_DIRADD, mp);
9276 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
9277 dap->da_offset = offset;
9278 dap->da_newinum = newinum;
9279 LIST_INIT(&dap->da_jwork);
9283 * Allocate a new dirrem and ACQUIRE_LOCK.
9285 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
9286 pagedep = dirrem->dm_pagedep;
9288 * The possible values for isrmdir:
9289 * 0 - non-directory file rename
9290 * 1 - directory rename within same directory
9291 * inum - directory rename to new directory of given inode number
9292 * When renaming to a new directory, we are both deleting and
9293 * creating a new directory entry, so the link count on the new
9294 * directory should not change. Thus we do not need the followup
9295 * dirrem which is usually done in handle_workitem_remove. We set
9296 * the DIRCHG flag to tell handle_workitem_remove to skip the
9300 dirrem->dm_state |= DIRCHG;
9303 * Whiteouts have no additional dependencies,
9304 * so just put the dirrem on the correct list.
9306 if (newinum == WINO) {
9307 if ((dirrem->dm_state & COMPLETE) == 0) {
9308 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
9311 dirrem->dm_dirinum = pagedep->pd_ino;
9312 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9313 add_to_worklist(&dirrem->dm_list, 0);
9315 FREE_LOCK(dp->i_ump);
9319 * Add the dirrem to the inodedep's pending remove list for quick
9320 * discovery later. A valid nlinkdelta ensures that this lookup
9323 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
9324 panic("softdep_setup_directory_change: Lost inodedep.");
9325 dirrem->dm_state |= ONDEPLIST;
9326 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9329 * If the COMPLETE flag is clear, then there were no active
9330 * entries and we want to roll back to the previous inode until
9331 * the new inode is committed to disk. If the COMPLETE flag is
9332 * set, then we have deleted an entry that never made it to disk.
9333 * If the entry we deleted resulted from a name change, then the old
9334 * inode reference still resides on disk. Any rollback that we do
9335 * needs to be to that old inode (returned to us in prevdirrem). If
9336 * the entry we deleted resulted from a create, then there is
9337 * no entry on the disk, so we want to roll back to zero rather
9338 * than the uncommitted inode. In either of the COMPLETE cases we
9339 * want to immediately free the unwritten and unreferenced inode.
9341 if ((dirrem->dm_state & COMPLETE) == 0) {
9342 dap->da_previous = dirrem;
9344 if (prevdirrem != NULL) {
9345 dap->da_previous = prevdirrem;
9347 dap->da_state &= ~DIRCHG;
9348 dap->da_pagedep = pagedep;
9350 dirrem->dm_dirinum = pagedep->pd_ino;
9351 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
9352 add_to_worklist(&dirrem->dm_list, 0);
9355 * Lookup the jaddref for this journal entry. We must finish
9356 * initializing it and make the diradd write dependent on it.
9357 * If we're not journaling, put it on the id_bufwait list if the
9358 * inode is not yet written. If it is written, do the post-inode
9359 * write processing to put it on the id_pendinghd list.
9361 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
9362 if (MOUNTEDSUJ(mp)) {
9363 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
9365 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
9366 ("softdep_setup_directory_change: bad jaddref %p",
9368 jaddref->ja_diroff = dp->i_offset;
9369 jaddref->ja_diradd = dap;
9370 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9372 add_to_journal(&jaddref->ja_list);
9373 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
9374 dap->da_state |= COMPLETE;
9375 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
9376 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
9378 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
9380 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
9383 * If we're making a new name for a directory that has not been
9384 * committed when need to move the dot and dotdot references to
9387 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
9388 merge_diradd(inodedep, dap);
9389 FREE_LOCK(dp->i_ump);
9393 * Called whenever the link count on an inode is changed.
9394 * It creates an inode dependency so that the new reference(s)
9395 * to the inode cannot be committed to disk until the updated
9396 * inode has been written.
9399 softdep_change_linkcnt(ip)
9400 struct inode *ip; /* the inode with the increased link count */
9402 struct inodedep *inodedep;
9404 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
9405 ("softdep_change_linkcnt called on non-softdep filesystem"));
9406 ACQUIRE_LOCK(ip->i_ump);
9407 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC,
9409 if (ip->i_nlink < ip->i_effnlink)
9410 panic("softdep_change_linkcnt: bad delta");
9411 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9412 FREE_LOCK(ip->i_ump);
9416 * Attach a sbdep dependency to the superblock buf so that we can keep
9417 * track of the head of the linked list of referenced but unlinked inodes.
9420 softdep_setup_sbupdate(ump, fs, bp)
9421 struct ufsmount *ump;
9425 struct sbdep *sbdep;
9426 struct worklist *wk;
9428 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
9429 ("softdep_setup_sbupdate called on non-softdep filesystem"));
9430 LIST_FOREACH(wk, &bp->b_dep, wk_list)
9431 if (wk->wk_type == D_SBDEP)
9435 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
9436 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
9438 sbdep->sb_ump = ump;
9440 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
9445 * Return the first unlinked inodedep which is ready to be the head of the
9446 * list. The inodedep and all those after it must have valid next pointers.
9448 static struct inodedep *
9449 first_unlinked_inodedep(ump)
9450 struct ufsmount *ump;
9452 struct inodedep *inodedep;
9453 struct inodedep *idp;
9456 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
9457 inodedep; inodedep = idp) {
9458 if ((inodedep->id_state & UNLINKNEXT) == 0)
9460 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9461 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
9463 if ((inodedep->id_state & UNLINKPREV) == 0)
9470 * Set the sujfree unlinked head pointer prior to writing a superblock.
9473 initiate_write_sbdep(sbdep)
9474 struct sbdep *sbdep;
9476 struct inodedep *inodedep;
9480 bpfs = sbdep->sb_fs;
9481 fs = sbdep->sb_ump->um_fs;
9482 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9484 fs->fs_sujfree = inodedep->id_ino;
9485 inodedep->id_state |= UNLINKPREV;
9488 bpfs->fs_sujfree = fs->fs_sujfree;
9492 * After a superblock is written determine whether it must be written again
9493 * due to a changing unlinked list head.
9496 handle_written_sbdep(sbdep, bp)
9497 struct sbdep *sbdep;
9500 struct inodedep *inodedep;
9503 LOCK_OWNED(sbdep->sb_ump);
9506 * If the superblock doesn't match the in-memory list start over.
9508 inodedep = first_unlinked_inodedep(sbdep->sb_ump);
9509 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
9510 (inodedep == NULL && fs->fs_sujfree != 0)) {
9514 WORKITEM_FREE(sbdep, D_SBDEP);
9515 if (fs->fs_sujfree == 0)
9518 * Now that we have a record of this inode in stable store allow it
9519 * to be written to free up pending work. Inodes may see a lot of
9520 * write activity after they are unlinked which we must not hold up.
9522 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
9523 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
9524 panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
9525 inodedep, inodedep->id_state);
9526 if (inodedep->id_state & UNLINKONLIST)
9528 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
9535 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
9538 unlinked_inodedep(mp, inodedep)
9540 struct inodedep *inodedep;
9542 struct ufsmount *ump;
9546 if (MOUNTEDSUJ(mp) == 0)
9548 ump->um_fs->fs_fmod = 1;
9549 if (inodedep->id_state & UNLINKED)
9550 panic("unlinked_inodedep: %p already unlinked\n", inodedep);
9551 inodedep->id_state |= UNLINKED;
9552 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
9556 * Remove an inodedep from the unlinked inodedep list. This may require
9557 * disk writes if the inode has made it that far.
9560 clear_unlinked_inodedep(inodedep)
9561 struct inodedep *inodedep;
9563 struct ufsmount *ump;
9564 struct inodedep *idp;
9565 struct inodedep *idn;
9573 ump = VFSTOUFS(inodedep->id_list.wk_mp);
9575 ino = inodedep->id_ino;
9579 KASSERT((inodedep->id_state & UNLINKED) != 0,
9580 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9583 * If nothing has yet been written simply remove us from
9584 * the in memory list and return. This is the most common
9585 * case where handle_workitem_remove() loses the final
9588 if ((inodedep->id_state & UNLINKLINKS) == 0)
9591 * If we have a NEXT pointer and no PREV pointer we can simply
9592 * clear NEXT's PREV and remove ourselves from the list. Be
9593 * careful not to clear PREV if the superblock points at
9596 idn = TAILQ_NEXT(inodedep, id_unlinked);
9597 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
9598 if (idn && fs->fs_sujfree != idn->id_ino)
9599 idn->id_state &= ~UNLINKPREV;
9603 * Here we have an inodedep which is actually linked into
9604 * the list. We must remove it by forcing a write to the
9605 * link before us, whether it be the superblock or an inode.
9606 * Unfortunately the list may change while we're waiting
9607 * on the buf lock for either resource so we must loop until
9608 * we lock the right one. If both the superblock and an
9609 * inode point to this inode we must clear the inode first
9610 * followed by the superblock.
9612 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9614 if (idp && (idp->id_state & UNLINKNEXT))
9618 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9619 (int)fs->fs_sbsize, 0, 0, 0);
9621 error = bread(ump->um_devvp,
9622 fsbtodb(fs, ino_to_fsba(fs, pino)),
9623 (int)fs->fs_bsize, NOCRED, &bp);
9630 /* If the list has changed restart the loop. */
9631 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
9633 if (idp && (idp->id_state & UNLINKNEXT))
9636 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
9643 idn = TAILQ_NEXT(inodedep, id_unlinked);
9647 * Remove us from the in memory list. After this we cannot
9648 * access the inodedep.
9650 KASSERT((inodedep->id_state & UNLINKED) != 0,
9651 ("clear_unlinked_inodedep: inodedep %p not unlinked",
9653 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9654 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9657 * The predecessor's next pointer is manually updated here
9658 * so that the NEXT flag is never cleared for an element
9659 * that is in the list.
9662 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9663 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9664 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9666 } else if (fs->fs_magic == FS_UFS1_MAGIC)
9667 ((struct ufs1_dinode *)bp->b_data +
9668 ino_to_fsbo(fs, pino))->di_freelink = nino;
9670 ((struct ufs2_dinode *)bp->b_data +
9671 ino_to_fsbo(fs, pino))->di_freelink = nino;
9673 * If the bwrite fails we have no recourse to recover. The
9674 * filesystem is corrupted already.
9679 * If the superblock pointer still needs to be cleared force
9682 if (fs->fs_sujfree == ino) {
9684 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
9685 (int)fs->fs_sbsize, 0, 0, 0);
9686 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
9687 ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
9688 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
9694 if (fs->fs_sujfree != ino)
9696 panic("clear_unlinked_inodedep: Failed to clear free head");
9698 if (inodedep->id_ino == fs->fs_sujfree)
9699 panic("clear_unlinked_inodedep: Freeing head of free list");
9700 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
9701 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
9706 * This workitem decrements the inode's link count.
9707 * If the link count reaches zero, the file is removed.
9710 handle_workitem_remove(dirrem, flags)
9711 struct dirrem *dirrem;
9714 struct inodedep *inodedep;
9715 struct workhead dotdotwk;
9716 struct worklist *wk;
9717 struct ufsmount *ump;
9723 if (dirrem->dm_state & ONWORKLIST)
9724 panic("handle_workitem_remove: dirrem %p still on worklist",
9726 oldinum = dirrem->dm_oldinum;
9727 mp = dirrem->dm_list.wk_mp;
9729 flags |= LK_EXCLUSIVE;
9730 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0)
9734 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
9735 panic("handle_workitem_remove: lost inodedep");
9736 if (dirrem->dm_state & ONDEPLIST)
9737 LIST_REMOVE(dirrem, dm_inonext);
9738 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
9739 ("handle_workitem_remove: Journal entries not written."));
9742 * Move all dependencies waiting on the remove to complete
9743 * from the dirrem to the inode inowait list to be completed
9744 * after the inode has been updated and written to disk. Any
9745 * marked MKDIR_PARENT are saved to be completed when the .. ref
9748 LIST_INIT(&dotdotwk);
9749 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
9750 WORKLIST_REMOVE(wk);
9751 if (wk->wk_state & MKDIR_PARENT) {
9752 wk->wk_state &= ~MKDIR_PARENT;
9753 WORKLIST_INSERT(&dotdotwk, wk);
9756 WORKLIST_INSERT(&inodedep->id_inowait, wk);
9758 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
9760 * Normal file deletion.
9762 if ((dirrem->dm_state & RMDIR) == 0) {
9764 DIP_SET(ip, i_nlink, ip->i_nlink);
9765 ip->i_flag |= IN_CHANGE;
9766 if (ip->i_nlink < ip->i_effnlink)
9767 panic("handle_workitem_remove: bad file delta");
9768 if (ip->i_nlink == 0)
9769 unlinked_inodedep(mp, inodedep);
9770 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9771 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9772 ("handle_workitem_remove: worklist not empty. %s",
9773 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
9774 WORKITEM_FREE(dirrem, D_DIRREM);
9779 * Directory deletion. Decrement reference count for both the
9780 * just deleted parent directory entry and the reference for ".".
9781 * Arrange to have the reference count on the parent decremented
9782 * to account for the loss of "..".
9785 DIP_SET(ip, i_nlink, ip->i_nlink);
9786 ip->i_flag |= IN_CHANGE;
9787 if (ip->i_nlink < ip->i_effnlink)
9788 panic("handle_workitem_remove: bad dir delta");
9789 if (ip->i_nlink == 0)
9790 unlinked_inodedep(mp, inodedep);
9791 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
9793 * Rename a directory to a new parent. Since, we are both deleting
9794 * and creating a new directory entry, the link count on the new
9795 * directory should not change. Thus we skip the followup dirrem.
9797 if (dirrem->dm_state & DIRCHG) {
9798 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
9799 ("handle_workitem_remove: DIRCHG and worklist not empty."));
9800 WORKITEM_FREE(dirrem, D_DIRREM);
9804 dirrem->dm_state = ONDEPLIST;
9805 dirrem->dm_oldinum = dirrem->dm_dirinum;
9807 * Place the dirrem on the parent's diremhd list.
9809 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
9810 panic("handle_workitem_remove: lost dir inodedep");
9811 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
9813 * If the allocated inode has never been written to disk, then
9814 * the on-disk inode is zero'ed and we can remove the file
9815 * immediately. When journaling if the inode has been marked
9816 * unlinked and not DEPCOMPLETE we know it can never be written.
9818 inodedep_lookup(mp, oldinum, 0, &inodedep);
9819 if (inodedep == NULL ||
9820 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
9821 check_inode_unwritten(inodedep)) {
9824 return handle_workitem_remove(dirrem, flags);
9826 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
9828 ip->i_flag |= IN_CHANGE;
9836 * Inode de-allocation dependencies.
9838 * When an inode's link count is reduced to zero, it can be de-allocated. We
9839 * found it convenient to postpone de-allocation until after the inode is
9840 * written to disk with its new link count (zero). At this point, all of the
9841 * on-disk inode's block pointers are nullified and, with careful dependency
9842 * list ordering, all dependencies related to the inode will be satisfied and
9843 * the corresponding dependency structures de-allocated. So, if/when the
9844 * inode is reused, there will be no mixing of old dependencies with new
9845 * ones. This artificial dependency is set up by the block de-allocation
9846 * procedure above (softdep_setup_freeblocks) and completed by the
9847 * following procedure.
9850 handle_workitem_freefile(freefile)
9851 struct freefile *freefile;
9853 struct workhead wkhd;
9855 struct inodedep *idp;
9856 struct ufsmount *ump;
9859 ump = VFSTOUFS(freefile->fx_list.wk_mp);
9863 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
9866 panic("handle_workitem_freefile: inodedep %p survived", idp);
9869 fs->fs_pendinginodes -= 1;
9872 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
9873 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
9874 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
9875 softdep_error("handle_workitem_freefile", error);
9877 WORKITEM_FREE(freefile, D_FREEFILE);
9883 * Helper function which unlinks marker element from work list and returns
9884 * the next element on the list.
9886 static __inline struct worklist *
9887 markernext(struct worklist *marker)
9889 struct worklist *next;
9891 next = LIST_NEXT(marker, wk_list);
9892 LIST_REMOVE(marker, wk_list);
9899 * The dependency structures constructed above are most actively used when file
9900 * system blocks are written to disk. No constraints are placed on when a
9901 * block can be written, but unsatisfied update dependencies are made safe by
9902 * modifying (or replacing) the source memory for the duration of the disk
9903 * write. When the disk write completes, the memory block is again brought
9906 * In-core inode structure reclamation.
9908 * Because there are a finite number of "in-core" inode structures, they are
9909 * reused regularly. By transferring all inode-related dependencies to the
9910 * in-memory inode block and indexing them separately (via "inodedep"s), we
9911 * can allow "in-core" inode structures to be reused at any time and avoid
9912 * any increase in contention.
9914 * Called just before entering the device driver to initiate a new disk I/O.
9915 * The buffer must be locked, thus, no I/O completion operations can occur
9916 * while we are manipulating its associated dependencies.
9919 softdep_disk_io_initiation(bp)
9920 struct buf *bp; /* structure describing disk write to occur */
9922 struct worklist *wk;
9923 struct worklist marker;
9924 struct inodedep *inodedep;
9925 struct freeblks *freeblks;
9926 struct jblkdep *jblkdep;
9927 struct newblk *newblk;
9928 struct ufsmount *ump;
9931 * We only care about write operations. There should never
9932 * be dependencies for reads.
9934 if (bp->b_iocmd != BIO_WRITE)
9935 panic("softdep_disk_io_initiation: not write");
9937 if (bp->b_vflags & BV_BKGRDINPROG)
9938 panic("softdep_disk_io_initiation: Writing buffer with "
9939 "background write in progress: %p", bp);
9941 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
9943 ump = VFSTOUFS(wk->wk_mp);
9945 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
9946 PHOLD(curproc); /* Don't swap out kernel stack */
9949 * Do any necessary pre-I/O processing.
9951 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
9952 wk = markernext(&marker)) {
9953 LIST_INSERT_AFTER(wk, &marker, wk_list);
9954 switch (wk->wk_type) {
9957 initiate_write_filepage(WK_PAGEDEP(wk), bp);
9961 inodedep = WK_INODEDEP(wk);
9962 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
9963 initiate_write_inodeblock_ufs1(inodedep, bp);
9965 initiate_write_inodeblock_ufs2(inodedep, bp);
9969 initiate_write_indirdep(WK_INDIRDEP(wk), bp);
9973 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
9977 WK_JSEG(wk)->js_buf = NULL;
9981 freeblks = WK_FREEBLKS(wk);
9982 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd);
9984 * We have to wait for the freeblks to be journaled
9985 * before we can write an inodeblock with updated
9986 * pointers. Be careful to arrange the marker so
9987 * we revisit the freeblks if it's not removed by
9988 * the first jwait().
9990 if (jblkdep != NULL) {
9991 LIST_REMOVE(&marker, wk_list);
9992 LIST_INSERT_BEFORE(wk, &marker, wk_list);
9993 jwait(&jblkdep->jb_list, MNT_WAIT);
9999 * We have to wait for the jnewblk to be journaled
10000 * before we can write to a block if the contents
10001 * may be confused with an earlier file's indirect
10002 * at recovery time. Handle the marker as described
10005 newblk = WK_NEWBLK(wk);
10006 if (newblk->nb_jnewblk != NULL &&
10007 indirblk_lookup(newblk->nb_list.wk_mp,
10008 newblk->nb_newblkno)) {
10009 LIST_REMOVE(&marker, wk_list);
10010 LIST_INSERT_BEFORE(wk, &marker, wk_list);
10011 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
10016 initiate_write_sbdep(WK_SBDEP(wk));
10026 panic("handle_disk_io_initiation: Unexpected type %s",
10027 TYPENAME(wk->wk_type));
10032 PRELE(curproc); /* Allow swapout of kernel stack */
10036 * Called from within the procedure above to deal with unsatisfied
10037 * allocation dependencies in a directory. The buffer must be locked,
10038 * thus, no I/O completion operations can occur while we are
10039 * manipulating its associated dependencies.
10042 initiate_write_filepage(pagedep, bp)
10043 struct pagedep *pagedep;
10046 struct jremref *jremref;
10047 struct jmvref *jmvref;
10048 struct dirrem *dirrem;
10049 struct diradd *dap;
10053 if (pagedep->pd_state & IOSTARTED) {
10055 * This can only happen if there is a driver that does not
10056 * understand chaining. Here biodone will reissue the call
10057 * to strategy for the incomplete buffers.
10059 printf("initiate_write_filepage: already started\n");
10062 pagedep->pd_state |= IOSTARTED;
10064 * Wait for all journal remove dependencies to hit the disk.
10065 * We can not allow any potentially conflicting directory adds
10066 * to be visible before removes and rollback is too difficult.
10067 * The per-filesystem lock may be dropped and re-acquired, however
10068 * we hold the buf locked so the dependency can not go away.
10070 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
10071 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL)
10072 jwait(&jremref->jr_list, MNT_WAIT);
10073 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL)
10074 jwait(&jmvref->jm_list, MNT_WAIT);
10075 for (i = 0; i < DAHASHSZ; i++) {
10076 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
10077 ep = (struct direct *)
10078 ((char *)bp->b_data + dap->da_offset);
10079 if (ep->d_ino != dap->da_newinum)
10080 panic("%s: dir inum %ju != new %ju",
10081 "initiate_write_filepage",
10082 (uintmax_t)ep->d_ino,
10083 (uintmax_t)dap->da_newinum);
10084 if (dap->da_state & DIRCHG)
10085 ep->d_ino = dap->da_previous->dm_oldinum;
10088 dap->da_state &= ~ATTACHED;
10089 dap->da_state |= UNDONE;
10095 * Version of initiate_write_inodeblock that handles UFS1 dinodes.
10096 * Note that any bug fixes made to this routine must be done in the
10097 * version found below.
10099 * Called from within the procedure above to deal with unsatisfied
10100 * allocation dependencies in an inodeblock. The buffer must be
10101 * locked, thus, no I/O completion operations can occur while we
10102 * are manipulating its associated dependencies.
10105 initiate_write_inodeblock_ufs1(inodedep, bp)
10106 struct inodedep *inodedep;
10107 struct buf *bp; /* The inode block */
10109 struct allocdirect *adp, *lastadp;
10110 struct ufs1_dinode *dp;
10111 struct ufs1_dinode *sip;
10112 struct inoref *inoref;
10113 struct ufsmount *ump;
10117 ufs_lbn_t prevlbn = 0;
10121 if (inodedep->id_state & IOSTARTED)
10122 panic("initiate_write_inodeblock_ufs1: already started");
10123 inodedep->id_state |= IOSTARTED;
10124 fs = inodedep->id_fs;
10125 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10127 dp = (struct ufs1_dinode *)bp->b_data +
10128 ino_to_fsbo(fs, inodedep->id_ino);
10131 * If we're on the unlinked list but have not yet written our
10132 * next pointer initialize it here.
10134 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10135 struct inodedep *inon;
10137 inon = TAILQ_NEXT(inodedep, id_unlinked);
10138 dp->di_freelink = inon ? inon->id_ino : 0;
10141 * If the bitmap is not yet written, then the allocated
10142 * inode cannot be written to disk.
10144 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10145 if (inodedep->id_savedino1 != NULL)
10146 panic("initiate_write_inodeblock_ufs1: I/O underway");
10148 sip = malloc(sizeof(struct ufs1_dinode),
10149 M_SAVEDINO, M_SOFTDEP_FLAGS);
10151 inodedep->id_savedino1 = sip;
10152 *inodedep->id_savedino1 = *dp;
10153 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
10154 dp->di_gen = inodedep->id_savedino1->di_gen;
10155 dp->di_freelink = inodedep->id_savedino1->di_freelink;
10159 * If no dependencies, then there is nothing to roll back.
10161 inodedep->id_savedsize = dp->di_size;
10162 inodedep->id_savedextsize = 0;
10163 inodedep->id_savednlink = dp->di_nlink;
10164 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10165 TAILQ_EMPTY(&inodedep->id_inoreflst))
10168 * Revert the link count to that of the first unwritten journal entry.
10170 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10172 dp->di_nlink = inoref->if_nlink;
10174 * Set the dependencies to busy.
10176 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10177 adp = TAILQ_NEXT(adp, ad_next)) {
10179 if (deplist != 0 && prevlbn >= adp->ad_offset)
10180 panic("softdep_write_inodeblock: lbn order");
10181 prevlbn = adp->ad_offset;
10182 if (adp->ad_offset < NDADDR &&
10183 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10184 panic("%s: direct pointer #%jd mismatch %d != %jd",
10185 "softdep_write_inodeblock",
10186 (intmax_t)adp->ad_offset,
10187 dp->di_db[adp->ad_offset],
10188 (intmax_t)adp->ad_newblkno);
10189 if (adp->ad_offset >= NDADDR &&
10190 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10191 panic("%s: indirect pointer #%jd mismatch %d != %jd",
10192 "softdep_write_inodeblock",
10193 (intmax_t)adp->ad_offset - NDADDR,
10194 dp->di_ib[adp->ad_offset - NDADDR],
10195 (intmax_t)adp->ad_newblkno);
10196 deplist |= 1 << adp->ad_offset;
10197 if ((adp->ad_state & ATTACHED) == 0)
10198 panic("softdep_write_inodeblock: Unknown state 0x%x",
10200 #endif /* INVARIANTS */
10201 adp->ad_state &= ~ATTACHED;
10202 adp->ad_state |= UNDONE;
10205 * The on-disk inode cannot claim to be any larger than the last
10206 * fragment that has been written. Otherwise, the on-disk inode
10207 * might have fragments that were not the last block in the file
10208 * which would corrupt the filesystem.
10210 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10211 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10212 if (adp->ad_offset >= NDADDR)
10214 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10215 /* keep going until hitting a rollback to a frag */
10216 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10218 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10219 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10221 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10222 panic("softdep_write_inodeblock: lost dep1");
10223 #endif /* INVARIANTS */
10226 for (i = 0; i < NIADDR; i++) {
10228 if (dp->di_ib[i] != 0 &&
10229 (deplist & ((1 << NDADDR) << i)) == 0)
10230 panic("softdep_write_inodeblock: lost dep2");
10231 #endif /* INVARIANTS */
10237 * If we have zero'ed out the last allocated block of the file,
10238 * roll back the size to the last currently allocated block.
10239 * We know that this last allocated block is a full-sized as
10240 * we already checked for fragments in the loop above.
10242 if (lastadp != NULL &&
10243 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10244 for (i = lastadp->ad_offset; i >= 0; i--)
10245 if (dp->di_db[i] != 0)
10247 dp->di_size = (i + 1) * fs->fs_bsize;
10250 * The only dependencies are for indirect blocks.
10252 * The file size for indirect block additions is not guaranteed.
10253 * Such a guarantee would be non-trivial to achieve. The conventional
10254 * synchronous write implementation also does not make this guarantee.
10255 * Fsck should catch and fix discrepancies. Arguably, the file size
10256 * can be over-estimated without destroying integrity when the file
10257 * moves into the indirect blocks (i.e., is large). If we want to
10258 * postpone fsck, we are stuck with this argument.
10260 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10261 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10265 * Version of initiate_write_inodeblock that handles UFS2 dinodes.
10266 * Note that any bug fixes made to this routine must be done in the
10267 * version found above.
10269 * Called from within the procedure above to deal with unsatisfied
10270 * allocation dependencies in an inodeblock. The buffer must be
10271 * locked, thus, no I/O completion operations can occur while we
10272 * are manipulating its associated dependencies.
10275 initiate_write_inodeblock_ufs2(inodedep, bp)
10276 struct inodedep *inodedep;
10277 struct buf *bp; /* The inode block */
10279 struct allocdirect *adp, *lastadp;
10280 struct ufs2_dinode *dp;
10281 struct ufs2_dinode *sip;
10282 struct inoref *inoref;
10283 struct ufsmount *ump;
10287 ufs_lbn_t prevlbn = 0;
10291 if (inodedep->id_state & IOSTARTED)
10292 panic("initiate_write_inodeblock_ufs2: already started");
10293 inodedep->id_state |= IOSTARTED;
10294 fs = inodedep->id_fs;
10295 ump = VFSTOUFS(inodedep->id_list.wk_mp);
10297 dp = (struct ufs2_dinode *)bp->b_data +
10298 ino_to_fsbo(fs, inodedep->id_ino);
10301 * If we're on the unlinked list but have not yet written our
10302 * next pointer initialize it here.
10304 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10305 struct inodedep *inon;
10307 inon = TAILQ_NEXT(inodedep, id_unlinked);
10308 dp->di_freelink = inon ? inon->id_ino : 0;
10311 * If the bitmap is not yet written, then the allocated
10312 * inode cannot be written to disk.
10314 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10315 if (inodedep->id_savedino2 != NULL)
10316 panic("initiate_write_inodeblock_ufs2: I/O underway");
10318 sip = malloc(sizeof(struct ufs2_dinode),
10319 M_SAVEDINO, M_SOFTDEP_FLAGS);
10321 inodedep->id_savedino2 = sip;
10322 *inodedep->id_savedino2 = *dp;
10323 bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
10324 dp->di_gen = inodedep->id_savedino2->di_gen;
10325 dp->di_freelink = inodedep->id_savedino2->di_freelink;
10329 * If no dependencies, then there is nothing to roll back.
10331 inodedep->id_savedsize = dp->di_size;
10332 inodedep->id_savedextsize = dp->di_extsize;
10333 inodedep->id_savednlink = dp->di_nlink;
10334 if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10335 TAILQ_EMPTY(&inodedep->id_extupdt) &&
10336 TAILQ_EMPTY(&inodedep->id_inoreflst))
10339 * Revert the link count to that of the first unwritten journal entry.
10341 inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10343 dp->di_nlink = inoref->if_nlink;
10346 * Set the ext data dependencies to busy.
10348 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10349 adp = TAILQ_NEXT(adp, ad_next)) {
10351 if (deplist != 0 && prevlbn >= adp->ad_offset)
10352 panic("softdep_write_inodeblock: lbn order");
10353 prevlbn = adp->ad_offset;
10354 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
10355 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10356 "softdep_write_inodeblock",
10357 (intmax_t)adp->ad_offset,
10358 (intmax_t)dp->di_extb[adp->ad_offset],
10359 (intmax_t)adp->ad_newblkno);
10360 deplist |= 1 << adp->ad_offset;
10361 if ((adp->ad_state & ATTACHED) == 0)
10362 panic("softdep_write_inodeblock: Unknown state 0x%x",
10364 #endif /* INVARIANTS */
10365 adp->ad_state &= ~ATTACHED;
10366 adp->ad_state |= UNDONE;
10369 * The on-disk inode cannot claim to be any larger than the last
10370 * fragment that has been written. Otherwise, the on-disk inode
10371 * might have fragments that were not the last block in the ext
10372 * data which would corrupt the filesystem.
10374 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10375 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10376 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
10377 /* keep going until hitting a rollback to a frag */
10378 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10380 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10381 for (i = adp->ad_offset + 1; i < NXADDR; i++) {
10383 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
10384 panic("softdep_write_inodeblock: lost dep1");
10385 #endif /* INVARIANTS */
10386 dp->di_extb[i] = 0;
10392 * If we have zero'ed out the last allocated block of the ext
10393 * data, roll back the size to the last currently allocated block.
10394 * We know that this last allocated block is a full-sized as
10395 * we already checked for fragments in the loop above.
10397 if (lastadp != NULL &&
10398 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10399 for (i = lastadp->ad_offset; i >= 0; i--)
10400 if (dp->di_extb[i] != 0)
10402 dp->di_extsize = (i + 1) * fs->fs_bsize;
10405 * Set the file data dependencies to busy.
10407 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10408 adp = TAILQ_NEXT(adp, ad_next)) {
10410 if (deplist != 0 && prevlbn >= adp->ad_offset)
10411 panic("softdep_write_inodeblock: lbn order");
10412 if ((adp->ad_state & ATTACHED) == 0)
10413 panic("inodedep %p and adp %p not attached", inodedep, adp);
10414 prevlbn = adp->ad_offset;
10415 if (adp->ad_offset < NDADDR &&
10416 dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10417 panic("%s: direct pointer #%jd mismatch %jd != %jd",
10418 "softdep_write_inodeblock",
10419 (intmax_t)adp->ad_offset,
10420 (intmax_t)dp->di_db[adp->ad_offset],
10421 (intmax_t)adp->ad_newblkno);
10422 if (adp->ad_offset >= NDADDR &&
10423 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10424 panic("%s indirect pointer #%jd mismatch %jd != %jd",
10425 "softdep_write_inodeblock:",
10426 (intmax_t)adp->ad_offset - NDADDR,
10427 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR],
10428 (intmax_t)adp->ad_newblkno);
10429 deplist |= 1 << adp->ad_offset;
10430 if ((adp->ad_state & ATTACHED) == 0)
10431 panic("softdep_write_inodeblock: Unknown state 0x%x",
10433 #endif /* INVARIANTS */
10434 adp->ad_state &= ~ATTACHED;
10435 adp->ad_state |= UNDONE;
10438 * The on-disk inode cannot claim to be any larger than the last
10439 * fragment that has been written. Otherwise, the on-disk inode
10440 * might have fragments that were not the last block in the file
10441 * which would corrupt the filesystem.
10443 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10444 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10445 if (adp->ad_offset >= NDADDR)
10447 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10448 /* keep going until hitting a rollback to a frag */
10449 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10451 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10452 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10454 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10455 panic("softdep_write_inodeblock: lost dep2");
10456 #endif /* INVARIANTS */
10459 for (i = 0; i < NIADDR; i++) {
10461 if (dp->di_ib[i] != 0 &&
10462 (deplist & ((1 << NDADDR) << i)) == 0)
10463 panic("softdep_write_inodeblock: lost dep3");
10464 #endif /* INVARIANTS */
10470 * If we have zero'ed out the last allocated block of the file,
10471 * roll back the size to the last currently allocated block.
10472 * We know that this last allocated block is a full-sized as
10473 * we already checked for fragments in the loop above.
10475 if (lastadp != NULL &&
10476 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10477 for (i = lastadp->ad_offset; i >= 0; i--)
10478 if (dp->di_db[i] != 0)
10480 dp->di_size = (i + 1) * fs->fs_bsize;
10483 * The only dependencies are for indirect blocks.
10485 * The file size for indirect block additions is not guaranteed.
10486 * Such a guarantee would be non-trivial to achieve. The conventional
10487 * synchronous write implementation also does not make this guarantee.
10488 * Fsck should catch and fix discrepancies. Arguably, the file size
10489 * can be over-estimated without destroying integrity when the file
10490 * moves into the indirect blocks (i.e., is large). If we want to
10491 * postpone fsck, we are stuck with this argument.
10493 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10494 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10498 * Cancel an indirdep as a result of truncation. Release all of the
10499 * children allocindirs and place their journal work on the appropriate
10503 cancel_indirdep(indirdep, bp, freeblks)
10504 struct indirdep *indirdep;
10506 struct freeblks *freeblks;
10508 struct allocindir *aip;
10511 * None of the indirect pointers will ever be visible,
10512 * so they can simply be tossed. GOINGAWAY ensures
10513 * that allocated pointers will be saved in the buffer
10514 * cache until they are freed. Note that they will
10515 * only be able to be found by their physical address
10516 * since the inode mapping the logical address will
10517 * be gone. The save buffer used for the safe copy
10518 * was allocated in setup_allocindir_phase2 using
10519 * the physical address so it could be used for this
10520 * purpose. Hence we swap the safe copy with the real
10521 * copy, allowing the safe copy to be freed and holding
10522 * on to the real copy for later use in indir_trunc.
10524 if (indirdep->ir_state & GOINGAWAY)
10525 panic("cancel_indirdep: already gone");
10526 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
10527 indirdep->ir_state |= DEPCOMPLETE;
10528 LIST_REMOVE(indirdep, ir_next);
10530 indirdep->ir_state |= GOINGAWAY;
10532 * Pass in bp for blocks still have journal writes
10533 * pending so we can cancel them on their own.
10535 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL)
10536 cancel_allocindir(aip, bp, freeblks, 0);
10537 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL)
10538 cancel_allocindir(aip, NULL, freeblks, 0);
10539 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL)
10540 cancel_allocindir(aip, NULL, freeblks, 0);
10541 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL)
10542 cancel_allocindir(aip, NULL, freeblks, 0);
10544 * If there are pending partial truncations we need to keep the
10545 * old block copy around until they complete. This is because
10546 * the current b_data is not a perfect superset of the available
10549 if (TAILQ_EMPTY(&indirdep->ir_trunc))
10550 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
10552 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10553 WORKLIST_REMOVE(&indirdep->ir_list);
10554 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
10555 indirdep->ir_bp = NULL;
10556 indirdep->ir_freeblks = freeblks;
10560 * Free an indirdep once it no longer has new pointers to track.
10563 free_indirdep(indirdep)
10564 struct indirdep *indirdep;
10567 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc),
10568 ("free_indirdep: Indir trunc list not empty."));
10569 KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
10570 ("free_indirdep: Complete head not empty."));
10571 KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
10572 ("free_indirdep: write head not empty."));
10573 KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
10574 ("free_indirdep: done head not empty."));
10575 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
10576 ("free_indirdep: deplist head not empty."));
10577 KASSERT((indirdep->ir_state & DEPCOMPLETE),
10578 ("free_indirdep: %p still on newblk list.", indirdep));
10579 KASSERT(indirdep->ir_saveddata == NULL,
10580 ("free_indirdep: %p still has saved data.", indirdep));
10581 if (indirdep->ir_state & ONWORKLIST)
10582 WORKLIST_REMOVE(&indirdep->ir_list);
10583 WORKITEM_FREE(indirdep, D_INDIRDEP);
10587 * Called before a write to an indirdep. This routine is responsible for
10588 * rolling back pointers to a safe state which includes only those
10589 * allocindirs which have been completed.
10592 initiate_write_indirdep(indirdep, bp)
10593 struct indirdep *indirdep;
10596 struct ufsmount *ump;
10598 indirdep->ir_state |= IOSTARTED;
10599 if (indirdep->ir_state & GOINGAWAY)
10600 panic("disk_io_initiation: indirdep gone");
10602 * If there are no remaining dependencies, this will be writing
10603 * the real pointers.
10605 if (LIST_EMPTY(&indirdep->ir_deplisthd) &&
10606 TAILQ_EMPTY(&indirdep->ir_trunc))
10609 * Replace up-to-date version with safe version.
10611 if (indirdep->ir_saveddata == NULL) {
10612 ump = VFSTOUFS(indirdep->ir_list.wk_mp);
10615 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
10619 indirdep->ir_state &= ~ATTACHED;
10620 indirdep->ir_state |= UNDONE;
10621 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10622 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
10627 * Called when an inode has been cleared in a cg bitmap. This finally
10628 * eliminates any canceled jaddrefs
10631 softdep_setup_inofree(mp, bp, ino, wkhd)
10635 struct workhead *wkhd;
10637 struct worklist *wk, *wkn;
10638 struct inodedep *inodedep;
10639 struct ufsmount *ump;
10644 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
10645 ("softdep_setup_inofree called on non-softdep filesystem"));
10646 ump = VFSTOUFS(mp);
10649 cgp = (struct cg *)bp->b_data;
10650 inosused = cg_inosused(cgp);
10651 if (isset(inosused, ino % fs->fs_ipg))
10652 panic("softdep_setup_inofree: inode %ju not freed.",
10654 if (inodedep_lookup(mp, ino, 0, &inodedep))
10655 panic("softdep_setup_inofree: ino %ju has existing inodedep %p",
10656 (uintmax_t)ino, inodedep);
10658 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
10659 if (wk->wk_type != D_JADDREF)
10661 WORKLIST_REMOVE(wk);
10663 * We can free immediately even if the jaddref
10664 * isn't attached in a background write as now
10665 * the bitmaps are reconciled.
10667 wk->wk_state |= COMPLETE | ATTACHED;
10668 free_jaddref(WK_JADDREF(wk));
10670 jwork_move(&bp->b_dep, wkhd);
10677 * Called via ffs_blkfree() after a set of frags has been cleared from a cg
10678 * map. Any dependencies waiting for the write to clear are added to the
10679 * buf's list and any jnewblks that are being canceled are discarded
10683 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
10686 ufs2_daddr_t blkno;
10688 struct workhead *wkhd;
10690 struct bmsafemap *bmsafemap;
10691 struct jnewblk *jnewblk;
10692 struct ufsmount *ump;
10693 struct worklist *wk;
10698 ufs2_daddr_t jstart;
10706 "softdep_setup_blkfree: blkno %jd frags %d wk head %p",
10707 blkno, frags, wkhd);
10709 ump = VFSTOUFS(mp);
10710 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0,
10711 ("softdep_setup_blkfree called on non-softdep filesystem"));
10713 /* Lookup the bmsafemap so we track when it is dirty. */
10715 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10717 * Detach any jnewblks which have been canceled. They must linger
10718 * until the bitmap is cleared again by ffs_blkfree() to prevent
10719 * an unjournaled allocation from hitting the disk.
10722 while ((wk = LIST_FIRST(wkhd)) != NULL) {
10724 "softdep_setup_blkfree: blkno %jd wk type %d",
10725 blkno, wk->wk_type);
10726 WORKLIST_REMOVE(wk);
10727 if (wk->wk_type != D_JNEWBLK) {
10728 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk);
10731 jnewblk = WK_JNEWBLK(wk);
10732 KASSERT(jnewblk->jn_state & GOINGAWAY,
10733 ("softdep_setup_blkfree: jnewblk not canceled."));
10736 * Assert that this block is free in the bitmap
10737 * before we discard the jnewblk.
10739 cgp = (struct cg *)bp->b_data;
10740 blksfree = cg_blksfree(cgp);
10741 bno = dtogd(fs, jnewblk->jn_blkno);
10742 for (i = jnewblk->jn_oldfrags;
10743 i < jnewblk->jn_frags; i++) {
10744 if (isset(blksfree, bno + i))
10746 panic("softdep_setup_blkfree: not free");
10750 * Even if it's not attached we can free immediately
10751 * as the new bitmap is correct.
10753 wk->wk_state |= COMPLETE | ATTACHED;
10754 free_jnewblk(jnewblk);
10760 * Assert that we are not freeing a block which has an outstanding
10761 * allocation dependency.
10763 fs = VFSTOUFS(mp)->um_fs;
10764 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10765 end = blkno + frags;
10766 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10768 * Don't match against blocks that will be freed when the
10769 * background write is done.
10771 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
10772 (COMPLETE | DEPCOMPLETE))
10774 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
10775 jend = jnewblk->jn_blkno + jnewblk->jn_frags;
10776 if ((blkno >= jstart && blkno < jend) ||
10777 (end > jstart && end <= jend)) {
10778 printf("state 0x%X %jd - %d %d dep %p\n",
10779 jnewblk->jn_state, jnewblk->jn_blkno,
10780 jnewblk->jn_oldfrags, jnewblk->jn_frags,
10782 panic("softdep_setup_blkfree: "
10783 "%jd-%jd(%d) overlaps with %jd-%jd",
10784 blkno, end, frags, jstart, jend);
10792 * Revert a block allocation when the journal record that describes it
10793 * is not yet written.
10796 jnewblk_rollback(jnewblk, fs, cgp, blksfree)
10797 struct jnewblk *jnewblk;
10802 ufs1_daddr_t fragno;
10808 cgbno = dtogd(fs, jnewblk->jn_blkno);
10810 * We have to test which frags need to be rolled back. We may
10811 * be operating on a stale copy when doing background writes.
10813 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++)
10814 if (isclr(blksfree, cgbno + i))
10819 * This is mostly ffs_blkfree() sans some validation and
10820 * superblock updates.
10822 if (frags == fs->fs_frag) {
10823 fragno = fragstoblks(fs, cgbno);
10824 ffs_setblock(fs, blksfree, fragno);
10825 ffs_clusteracct(fs, cgp, fragno, 1);
10826 cgp->cg_cs.cs_nbfree++;
10828 cgbno += jnewblk->jn_oldfrags;
10829 bbase = cgbno - fragnum(fs, cgbno);
10830 /* Decrement the old frags. */
10831 blk = blkmap(fs, blksfree, bbase);
10832 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
10833 /* Deallocate the fragment */
10834 for (i = 0; i < frags; i++)
10835 setbit(blksfree, cgbno + i);
10836 cgp->cg_cs.cs_nffree += frags;
10837 /* Add back in counts associated with the new frags */
10838 blk = blkmap(fs, blksfree, bbase);
10839 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
10840 /* If a complete block has been reassembled, account for it. */
10841 fragno = fragstoblks(fs, bbase);
10842 if (ffs_isblock(fs, blksfree, fragno)) {
10843 cgp->cg_cs.cs_nffree -= fs->fs_frag;
10844 ffs_clusteracct(fs, cgp, fragno, 1);
10845 cgp->cg_cs.cs_nbfree++;
10849 jnewblk->jn_state &= ~ATTACHED;
10850 jnewblk->jn_state |= UNDONE;
10856 initiate_write_bmsafemap(bmsafemap, bp)
10857 struct bmsafemap *bmsafemap;
10858 struct buf *bp; /* The cg block. */
10860 struct jaddref *jaddref;
10861 struct jnewblk *jnewblk;
10868 if (bmsafemap->sm_state & IOSTARTED)
10870 bmsafemap->sm_state |= IOSTARTED;
10872 * Clear any inode allocations which are pending journal writes.
10874 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) {
10875 cgp = (struct cg *)bp->b_data;
10876 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10877 inosused = cg_inosused(cgp);
10878 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) {
10879 ino = jaddref->ja_ino % fs->fs_ipg;
10880 if (isset(inosused, ino)) {
10881 if ((jaddref->ja_mode & IFMT) == IFDIR)
10882 cgp->cg_cs.cs_ndir--;
10883 cgp->cg_cs.cs_nifree++;
10884 clrbit(inosused, ino);
10885 jaddref->ja_state &= ~ATTACHED;
10886 jaddref->ja_state |= UNDONE;
10889 panic("initiate_write_bmsafemap: inode %ju "
10890 "marked free", (uintmax_t)jaddref->ja_ino);
10894 * Clear any block allocations which are pending journal writes.
10896 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
10897 cgp = (struct cg *)bp->b_data;
10898 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10899 blksfree = cg_blksfree(cgp);
10900 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10901 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree))
10903 panic("initiate_write_bmsafemap: block %jd "
10904 "marked free", jnewblk->jn_blkno);
10908 * Move allocation lists to the written lists so they can be
10909 * cleared once the block write is complete.
10911 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr,
10912 inodedep, id_deps);
10913 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
10915 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist,
10920 * This routine is called during the completion interrupt
10921 * service routine for a disk write (from the procedure called
10922 * by the device driver to inform the filesystem caches of
10923 * a request completion). It should be called early in this
10924 * procedure, before the block is made available to other
10925 * processes or other routines are called.
10929 softdep_disk_write_complete(bp)
10930 struct buf *bp; /* describes the completed disk write */
10932 struct worklist *wk;
10933 struct worklist *owk;
10934 struct ufsmount *ump;
10935 struct workhead reattach;
10936 struct freeblks *freeblks;
10940 * If an error occurred while doing the write, then the data
10941 * has not hit the disk and the dependencies cannot be unrolled.
10943 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
10945 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
10947 ump = VFSTOUFS(wk->wk_mp);
10948 LIST_INIT(&reattach);
10950 * This lock must not be released anywhere in this code segment.
10955 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
10956 WORKLIST_REMOVE(wk);
10957 atomic_add_long(&dep_write[wk->wk_type], 1);
10959 panic("duplicate worklist: %p\n", wk);
10961 switch (wk->wk_type) {
10964 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
10965 WORKLIST_INSERT(&reattach, wk);
10969 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
10970 WORKLIST_INSERT(&reattach, wk);
10974 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp))
10975 WORKLIST_INSERT(&reattach, wk);
10979 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
10982 case D_ALLOCDIRECT:
10983 wk->wk_state |= COMPLETE;
10984 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
10988 wk->wk_state |= COMPLETE;
10989 handle_allocindir_partdone(WK_ALLOCINDIR(wk));
10993 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp))
10994 WORKLIST_INSERT(&reattach, wk);
10998 wk->wk_state |= COMPLETE;
10999 freeblks = WK_FREEBLKS(wk);
11000 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE &&
11001 LIST_EMPTY(&freeblks->fb_jblkdephd))
11002 add_to_worklist(wk, WK_NODELAY);
11006 handle_written_freework(WK_FREEWORK(wk));
11010 free_jsegdep(WK_JSEGDEP(wk));
11014 handle_written_jseg(WK_JSEG(wk), bp);
11018 if (handle_written_sbdep(WK_SBDEP(wk), bp))
11019 WORKLIST_INSERT(&reattach, wk);
11023 free_freedep(WK_FREEDEP(wk));
11027 panic("handle_disk_write_complete: Unknown type %s",
11028 TYPENAME(wk->wk_type));
11033 * Reattach any requests that must be redone.
11035 while ((wk = LIST_FIRST(&reattach)) != NULL) {
11036 WORKLIST_REMOVE(wk);
11037 WORKLIST_INSERT(&bp->b_dep, wk);
11045 * Called from within softdep_disk_write_complete above. Note that
11046 * this routine is always called from interrupt level with further
11047 * splbio interrupts blocked.
11050 handle_allocdirect_partdone(adp, wkhd)
11051 struct allocdirect *adp; /* the completed allocdirect */
11052 struct workhead *wkhd; /* Work to do when inode is writtne. */
11054 struct allocdirectlst *listhead;
11055 struct allocdirect *listadp;
11056 struct inodedep *inodedep;
11059 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11062 * The on-disk inode cannot claim to be any larger than the last
11063 * fragment that has been written. Otherwise, the on-disk inode
11064 * might have fragments that were not the last block in the file
11065 * which would corrupt the filesystem. Thus, we cannot free any
11066 * allocdirects after one whose ad_oldblkno claims a fragment as
11067 * these blocks must be rolled back to zero before writing the inode.
11068 * We check the currently active set of allocdirects in id_inoupdt
11069 * or id_extupdt as appropriate.
11071 inodedep = adp->ad_inodedep;
11072 bsize = inodedep->id_fs->fs_bsize;
11073 if (adp->ad_state & EXTDATA)
11074 listhead = &inodedep->id_extupdt;
11076 listhead = &inodedep->id_inoupdt;
11077 TAILQ_FOREACH(listadp, listhead, ad_next) {
11078 /* found our block */
11079 if (listadp == adp)
11081 /* continue if ad_oldlbn is not a fragment */
11082 if (listadp->ad_oldsize == 0 ||
11083 listadp->ad_oldsize == bsize)
11085 /* hit a fragment */
11089 * If we have reached the end of the current list without
11090 * finding the just finished dependency, then it must be
11091 * on the future dependency list. Future dependencies cannot
11092 * be freed until they are moved to the current list.
11094 if (listadp == NULL) {
11096 if (adp->ad_state & EXTDATA)
11097 listhead = &inodedep->id_newextupdt;
11099 listhead = &inodedep->id_newinoupdt;
11100 TAILQ_FOREACH(listadp, listhead, ad_next)
11101 /* found our block */
11102 if (listadp == adp)
11104 if (listadp == NULL)
11105 panic("handle_allocdirect_partdone: lost dep");
11110 * If we have found the just finished dependency, then queue
11111 * it along with anything that follows it that is complete.
11112 * Since the pointer has not yet been written in the inode
11113 * as the dependency prevents it, place the allocdirect on the
11114 * bufwait list where it will be freed once the pointer is
11118 wkhd = &inodedep->id_bufwait;
11119 for (; adp; adp = listadp) {
11120 listadp = TAILQ_NEXT(adp, ad_next);
11121 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
11123 TAILQ_REMOVE(listhead, adp, ad_next);
11124 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
11129 * Called from within softdep_disk_write_complete above. This routine
11130 * completes successfully written allocindirs.
11133 handle_allocindir_partdone(aip)
11134 struct allocindir *aip; /* the completed allocindir */
11136 struct indirdep *indirdep;
11138 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
11140 indirdep = aip->ai_indirdep;
11141 LIST_REMOVE(aip, ai_next);
11143 * Don't set a pointer while the buffer is undergoing IO or while
11144 * we have active truncations.
11146 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) {
11147 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
11150 if (indirdep->ir_state & UFS1FMT)
11151 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11154 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
11157 * Await the pointer write before freeing the allocindir.
11159 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
11163 * Release segments held on a jwork list.
11167 struct workhead *wkhd;
11169 struct worklist *wk;
11171 while ((wk = LIST_FIRST(wkhd)) != NULL) {
11172 WORKLIST_REMOVE(wk);
11173 switch (wk->wk_type) {
11175 free_jsegdep(WK_JSEGDEP(wk));
11178 free_freedep(WK_FREEDEP(wk));
11181 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep));
11182 WORKITEM_FREE(wk, D_FREEFRAG);
11185 handle_written_freework(WK_FREEWORK(wk));
11188 panic("handle_jwork: Unknown type %s\n",
11189 TYPENAME(wk->wk_type));
11195 * Handle the bufwait list on an inode when it is safe to release items
11196 * held there. This normally happens after an inode block is written but
11197 * may be delayed and handled later if there are pending journal items that
11198 * are not yet safe to be released.
11200 static struct freefile *
11201 handle_bufwait(inodedep, refhd)
11202 struct inodedep *inodedep;
11203 struct workhead *refhd;
11205 struct jaddref *jaddref;
11206 struct freefile *freefile;
11207 struct worklist *wk;
11210 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
11211 WORKLIST_REMOVE(wk);
11212 switch (wk->wk_type) {
11215 * We defer adding freefile to the worklist
11216 * until all other additions have been made to
11217 * ensure that it will be done after all the
11218 * old blocks have been freed.
11220 if (freefile != NULL)
11221 panic("handle_bufwait: freefile");
11222 freefile = WK_FREEFILE(wk);
11226 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
11230 diradd_inode_written(WK_DIRADD(wk), inodedep);
11234 wk->wk_state |= COMPLETE;
11235 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
11236 add_to_worklist(wk, 0);
11240 wk->wk_state |= COMPLETE;
11241 add_to_worklist(wk, 0);
11244 case D_ALLOCDIRECT:
11246 free_newblk(WK_NEWBLK(wk));
11250 wk->wk_state |= COMPLETE;
11251 free_jnewblk(WK_JNEWBLK(wk));
11255 * Save freed journal segments and add references on
11256 * the supplied list which will delay their release
11257 * until the cg bitmap is cleared on disk.
11261 free_jsegdep(WK_JSEGDEP(wk));
11263 WORKLIST_INSERT(refhd, wk);
11267 jaddref = WK_JADDREF(wk);
11268 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
11271 * Transfer any jaddrefs to the list to be freed with
11272 * the bitmap if we're handling a removed file.
11274 if (refhd == NULL) {
11275 wk->wk_state |= COMPLETE;
11276 free_jaddref(jaddref);
11278 WORKLIST_INSERT(refhd, wk);
11282 panic("handle_bufwait: Unknown type %p(%s)",
11283 wk, TYPENAME(wk->wk_type));
11290 * Called from within softdep_disk_write_complete above to restore
11291 * in-memory inode block contents to their most up-to-date state. Note
11292 * that this routine is always called from interrupt level with further
11293 * splbio interrupts blocked.
11296 handle_written_inodeblock(inodedep, bp)
11297 struct inodedep *inodedep;
11298 struct buf *bp; /* buffer containing the inode block */
11300 struct freefile *freefile;
11301 struct allocdirect *adp, *nextadp;
11302 struct ufs1_dinode *dp1 = NULL;
11303 struct ufs2_dinode *dp2 = NULL;
11304 struct workhead wkhd;
11305 int hadchanges, fstype;
11311 if ((inodedep->id_state & IOSTARTED) == 0)
11312 panic("handle_written_inodeblock: not started");
11313 inodedep->id_state &= ~IOSTARTED;
11314 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
11316 dp1 = (struct ufs1_dinode *)bp->b_data +
11317 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11318 freelink = dp1->di_freelink;
11321 dp2 = (struct ufs2_dinode *)bp->b_data +
11322 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11323 freelink = dp2->di_freelink;
11326 * Leave this inodeblock dirty until it's in the list.
11328 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) {
11329 struct inodedep *inon;
11331 inon = TAILQ_NEXT(inodedep, id_unlinked);
11332 if ((inon == NULL && freelink == 0) ||
11333 (inon && inon->id_ino == freelink)) {
11335 inon->id_state |= UNLINKPREV;
11336 inodedep->id_state |= UNLINKNEXT;
11341 * If we had to rollback the inode allocation because of
11342 * bitmaps being incomplete, then simply restore it.
11343 * Keep the block dirty so that it will not be reclaimed until
11344 * all associated dependencies have been cleared and the
11345 * corresponding updates written to disk.
11347 if (inodedep->id_savedino1 != NULL) {
11349 if (fstype == UFS1)
11350 *dp1 = *inodedep->id_savedino1;
11352 *dp2 = *inodedep->id_savedino2;
11353 free(inodedep->id_savedino1, M_SAVEDINO);
11354 inodedep->id_savedino1 = NULL;
11355 if ((bp->b_flags & B_DELWRI) == 0)
11356 stat_inode_bitmap++;
11359 * If the inode is clear here and GOINGAWAY it will never
11360 * be written. Process the bufwait and clear any pending
11361 * work which may include the freefile.
11363 if (inodedep->id_state & GOINGAWAY)
11367 inodedep->id_state |= COMPLETE;
11369 * Roll forward anything that had to be rolled back before
11370 * the inode could be updated.
11372 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
11373 nextadp = TAILQ_NEXT(adp, ad_next);
11374 if (adp->ad_state & ATTACHED)
11375 panic("handle_written_inodeblock: new entry");
11376 if (fstype == UFS1) {
11377 if (adp->ad_offset < NDADDR) {
11378 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11379 panic("%s %s #%jd mismatch %d != %jd",
11380 "handle_written_inodeblock:",
11382 (intmax_t)adp->ad_offset,
11383 dp1->di_db[adp->ad_offset],
11384 (intmax_t)adp->ad_oldblkno);
11385 dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
11387 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0)
11388 panic("%s: %s #%jd allocated as %d",
11389 "handle_written_inodeblock",
11390 "indirect pointer",
11391 (intmax_t)adp->ad_offset - NDADDR,
11392 dp1->di_ib[adp->ad_offset - NDADDR]);
11393 dp1->di_ib[adp->ad_offset - NDADDR] =
11397 if (adp->ad_offset < NDADDR) {
11398 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11399 panic("%s: %s #%jd %s %jd != %jd",
11400 "handle_written_inodeblock",
11402 (intmax_t)adp->ad_offset, "mismatch",
11403 (intmax_t)dp2->di_db[adp->ad_offset],
11404 (intmax_t)adp->ad_oldblkno);
11405 dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
11407 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0)
11408 panic("%s: %s #%jd allocated as %jd",
11409 "handle_written_inodeblock",
11410 "indirect pointer",
11411 (intmax_t)adp->ad_offset - NDADDR,
11413 dp2->di_ib[adp->ad_offset - NDADDR]);
11414 dp2->di_ib[adp->ad_offset - NDADDR] =
11418 adp->ad_state &= ~UNDONE;
11419 adp->ad_state |= ATTACHED;
11422 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
11423 nextadp = TAILQ_NEXT(adp, ad_next);
11424 if (adp->ad_state & ATTACHED)
11425 panic("handle_written_inodeblock: new entry");
11426 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
11427 panic("%s: direct pointers #%jd %s %jd != %jd",
11428 "handle_written_inodeblock",
11429 (intmax_t)adp->ad_offset, "mismatch",
11430 (intmax_t)dp2->di_extb[adp->ad_offset],
11431 (intmax_t)adp->ad_oldblkno);
11432 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
11433 adp->ad_state &= ~UNDONE;
11434 adp->ad_state |= ATTACHED;
11437 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
11438 stat_direct_blk_ptrs++;
11440 * Reset the file size to its most up-to-date value.
11442 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
11443 panic("handle_written_inodeblock: bad size");
11444 if (inodedep->id_savednlink > LINK_MAX)
11445 panic("handle_written_inodeblock: Invalid link count "
11446 "%d for inodedep %p", inodedep->id_savednlink, inodedep);
11447 if (fstype == UFS1) {
11448 if (dp1->di_nlink != inodedep->id_savednlink) {
11449 dp1->di_nlink = inodedep->id_savednlink;
11452 if (dp1->di_size != inodedep->id_savedsize) {
11453 dp1->di_size = inodedep->id_savedsize;
11457 if (dp2->di_nlink != inodedep->id_savednlink) {
11458 dp2->di_nlink = inodedep->id_savednlink;
11461 if (dp2->di_size != inodedep->id_savedsize) {
11462 dp2->di_size = inodedep->id_savedsize;
11465 if (dp2->di_extsize != inodedep->id_savedextsize) {
11466 dp2->di_extsize = inodedep->id_savedextsize;
11470 inodedep->id_savedsize = -1;
11471 inodedep->id_savedextsize = -1;
11472 inodedep->id_savednlink = -1;
11474 * If there were any rollbacks in the inode block, then it must be
11475 * marked dirty so that its will eventually get written back in
11476 * its correct form.
11482 * Process any allocdirects that completed during the update.
11484 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
11485 handle_allocdirect_partdone(adp, &wkhd);
11486 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
11487 handle_allocdirect_partdone(adp, &wkhd);
11489 * Process deallocations that were held pending until the
11490 * inode had been written to disk. Freeing of the inode
11491 * is delayed until after all blocks have been freed to
11492 * avoid creation of new <vfsid, inum, lbn> triples
11493 * before the old ones have been deleted. Completely
11494 * unlinked inodes are not processed until the unlinked
11495 * inode list is written or the last reference is removed.
11497 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
11498 freefile = handle_bufwait(inodedep, NULL);
11499 if (freefile && !LIST_EMPTY(&wkhd)) {
11500 WORKLIST_INSERT(&wkhd, &freefile->fx_list);
11505 * Move rolled forward dependency completions to the bufwait list
11506 * now that those that were already written have been processed.
11508 if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
11509 panic("handle_written_inodeblock: bufwait but no changes");
11510 jwork_move(&inodedep->id_bufwait, &wkhd);
11512 if (freefile != NULL) {
11514 * If the inode is goingaway it was never written. Fake up
11515 * the state here so free_inodedep() can succeed.
11517 if (inodedep->id_state & GOINGAWAY)
11518 inodedep->id_state |= COMPLETE | DEPCOMPLETE;
11519 if (free_inodedep(inodedep) == 0)
11520 panic("handle_written_inodeblock: live inodedep %p",
11522 add_to_worklist(&freefile->fx_list, 0);
11527 * If no outstanding dependencies, free it.
11529 if (free_inodedep(inodedep) ||
11530 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
11531 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
11532 TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
11533 LIST_FIRST(&inodedep->id_bufwait) == 0))
11535 return (hadchanges);
11539 handle_written_indirdep(indirdep, bp, bpp)
11540 struct indirdep *indirdep;
11544 struct allocindir *aip;
11548 if (indirdep->ir_state & GOINGAWAY)
11549 panic("handle_written_indirdep: indirdep gone");
11550 if ((indirdep->ir_state & IOSTARTED) == 0)
11551 panic("handle_written_indirdep: IO not started");
11554 * If there were rollbacks revert them here.
11556 if (indirdep->ir_saveddata) {
11557 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
11558 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11559 free(indirdep->ir_saveddata, M_INDIRDEP);
11560 indirdep->ir_saveddata = NULL;
11564 indirdep->ir_state &= ~(UNDONE | IOSTARTED);
11565 indirdep->ir_state |= ATTACHED;
11567 * Move allocindirs with written pointers to the completehd if
11568 * the indirdep's pointer is not yet written. Otherwise
11571 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) {
11572 LIST_REMOVE(aip, ai_next);
11573 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
11574 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
11576 newblk_freefrag(&aip->ai_block);
11579 free_newblk(&aip->ai_block);
11582 * Move allocindirs that have finished dependency processing from
11583 * the done list to the write list after updating the pointers.
11585 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11586 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) {
11587 handle_allocindir_partdone(aip);
11588 if (aip == LIST_FIRST(&indirdep->ir_donehd))
11589 panic("disk_write_complete: not gone");
11594 * Preserve the indirdep if there were any changes or if it is not
11595 * yet valid on disk.
11598 stat_indir_blk_ptrs++;
11603 * If there were no changes we can discard the savedbp and detach
11604 * ourselves from the buf. We are only carrying completed pointers
11607 sbp = indirdep->ir_savebp;
11608 sbp->b_flags |= B_INVAL | B_NOCACHE;
11609 indirdep->ir_savebp = NULL;
11610 indirdep->ir_bp = NULL;
11612 panic("handle_written_indirdep: bp already exists.");
11615 * The indirdep may not be freed until its parent points at it.
11617 if (indirdep->ir_state & DEPCOMPLETE)
11618 free_indirdep(indirdep);
11624 * Process a diradd entry after its dependent inode has been written.
11625 * This routine must be called with splbio interrupts blocked.
11628 diradd_inode_written(dap, inodedep)
11629 struct diradd *dap;
11630 struct inodedep *inodedep;
11633 dap->da_state |= COMPLETE;
11634 complete_diradd(dap);
11635 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
11639 * Returns true if the bmsafemap will have rollbacks when written. Must only
11640 * be called with the per-filesystem lock and the buf lock on the cg held.
11643 bmsafemap_backgroundwrite(bmsafemap, bp)
11644 struct bmsafemap *bmsafemap;
11649 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp));
11650 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) |
11651 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd);
11653 * If we're initiating a background write we need to process the
11654 * rollbacks as they exist now, not as they exist when IO starts.
11655 * No other consumers will look at the contents of the shadowed
11656 * buf so this is safe to do here.
11658 if (bp->b_xflags & BX_BKGRDMARKER)
11659 initiate_write_bmsafemap(bmsafemap, bp);
11665 * Re-apply an allocation when a cg write is complete.
11668 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)
11669 struct jnewblk *jnewblk;
11674 ufs1_daddr_t fragno;
11675 ufs2_daddr_t blkno;
11681 cgbno = dtogd(fs, jnewblk->jn_blkno);
11682 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) {
11683 if (isclr(blksfree, cgbno + i))
11684 panic("jnewblk_rollforward: re-allocated fragment");
11687 if (frags == fs->fs_frag) {
11688 blkno = fragstoblks(fs, cgbno);
11689 ffs_clrblock(fs, blksfree, (long)blkno);
11690 ffs_clusteracct(fs, cgp, blkno, -1);
11691 cgp->cg_cs.cs_nbfree--;
11693 bbase = cgbno - fragnum(fs, cgbno);
11694 cgbno += jnewblk->jn_oldfrags;
11695 /* If a complete block had been reassembled, account for it. */
11696 fragno = fragstoblks(fs, bbase);
11697 if (ffs_isblock(fs, blksfree, fragno)) {
11698 cgp->cg_cs.cs_nffree += fs->fs_frag;
11699 ffs_clusteracct(fs, cgp, fragno, -1);
11700 cgp->cg_cs.cs_nbfree--;
11702 /* Decrement the old frags. */
11703 blk = blkmap(fs, blksfree, bbase);
11704 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
11705 /* Allocate the fragment */
11706 for (i = 0; i < frags; i++)
11707 clrbit(blksfree, cgbno + i);
11708 cgp->cg_cs.cs_nffree -= frags;
11709 /* Add back in counts associated with the new frags */
11710 blk = blkmap(fs, blksfree, bbase);
11711 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
11717 * Complete a write to a bmsafemap structure. Roll forward any bitmap
11718 * changes if it's not a background write. Set all written dependencies
11719 * to DEPCOMPLETE and free the structure if possible.
11722 handle_written_bmsafemap(bmsafemap, bp)
11723 struct bmsafemap *bmsafemap;
11726 struct newblk *newblk;
11727 struct inodedep *inodedep;
11728 struct jaddref *jaddref, *jatmp;
11729 struct jnewblk *jnewblk, *jntmp;
11730 struct ufsmount *ump;
11739 if ((bmsafemap->sm_state & IOSTARTED) == 0)
11740 panic("initiate_write_bmsafemap: Not started\n");
11741 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp);
11743 bmsafemap->sm_state &= ~IOSTARTED;
11744 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0;
11746 * Release journal work that was waiting on the write.
11748 handle_jwork(&bmsafemap->sm_freewr);
11751 * Restore unwritten inode allocation pending jaddref writes.
11753 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
11754 cgp = (struct cg *)bp->b_data;
11755 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11756 inosused = cg_inosused(cgp);
11757 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
11758 ja_bmdeps, jatmp) {
11759 if ((jaddref->ja_state & UNDONE) == 0)
11761 ino = jaddref->ja_ino % fs->fs_ipg;
11762 if (isset(inosused, ino))
11763 panic("handle_written_bmsafemap: "
11764 "re-allocated inode");
11765 /* Do the roll-forward only if it's a real copy. */
11767 if ((jaddref->ja_mode & IFMT) == IFDIR)
11768 cgp->cg_cs.cs_ndir++;
11769 cgp->cg_cs.cs_nifree--;
11770 setbit(inosused, ino);
11773 jaddref->ja_state &= ~UNDONE;
11774 jaddref->ja_state |= ATTACHED;
11775 free_jaddref(jaddref);
11779 * Restore any block allocations which are pending journal writes.
11781 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
11782 cgp = (struct cg *)bp->b_data;
11783 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11784 blksfree = cg_blksfree(cgp);
11785 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
11787 if ((jnewblk->jn_state & UNDONE) == 0)
11789 /* Do the roll-forward only if it's a real copy. */
11791 jnewblk_rollforward(jnewblk, fs, cgp, blksfree))
11793 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
11794 jnewblk->jn_state |= ATTACHED;
11795 free_jnewblk(jnewblk);
11798 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
11799 newblk->nb_state |= DEPCOMPLETE;
11800 newblk->nb_state &= ~ONDEPLIST;
11801 newblk->nb_bmsafemap = NULL;
11802 LIST_REMOVE(newblk, nb_deps);
11803 if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
11804 handle_allocdirect_partdone(
11805 WK_ALLOCDIRECT(&newblk->nb_list), NULL);
11806 else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
11807 handle_allocindir_partdone(
11808 WK_ALLOCINDIR(&newblk->nb_list));
11809 else if (newblk->nb_list.wk_type != D_NEWBLK)
11810 panic("handle_written_bmsafemap: Unexpected type: %s",
11811 TYPENAME(newblk->nb_list.wk_type));
11813 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
11814 inodedep->id_state |= DEPCOMPLETE;
11815 inodedep->id_state &= ~ONDEPLIST;
11816 LIST_REMOVE(inodedep, id_deps);
11817 inodedep->id_bmsafemap = NULL;
11819 LIST_REMOVE(bmsafemap, sm_next);
11820 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
11821 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
11822 LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
11823 LIST_EMPTY(&bmsafemap->sm_inodedephd) &&
11824 LIST_EMPTY(&bmsafemap->sm_freehd)) {
11825 LIST_REMOVE(bmsafemap, sm_hash);
11826 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
11829 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
11836 * Try to free a mkdir dependency.
11839 complete_mkdir(mkdir)
11840 struct mkdir *mkdir;
11842 struct diradd *dap;
11844 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
11846 LIST_REMOVE(mkdir, md_mkdirs);
11847 dap = mkdir->md_diradd;
11848 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
11849 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
11850 dap->da_state |= DEPCOMPLETE;
11851 complete_diradd(dap);
11853 WORKITEM_FREE(mkdir, D_MKDIR);
11857 * Handle the completion of a mkdir dependency.
11860 handle_written_mkdir(mkdir, type)
11861 struct mkdir *mkdir;
11865 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
11866 panic("handle_written_mkdir: bad type");
11867 mkdir->md_state |= COMPLETE;
11868 complete_mkdir(mkdir);
11872 free_pagedep(pagedep)
11873 struct pagedep *pagedep;
11877 if (pagedep->pd_state & NEWBLOCK)
11879 if (!LIST_EMPTY(&pagedep->pd_dirremhd))
11881 for (i = 0; i < DAHASHSZ; i++)
11882 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
11884 if (!LIST_EMPTY(&pagedep->pd_pendinghd))
11886 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
11888 if (pagedep->pd_state & ONWORKLIST)
11889 WORKLIST_REMOVE(&pagedep->pd_list);
11890 LIST_REMOVE(pagedep, pd_hash);
11891 WORKITEM_FREE(pagedep, D_PAGEDEP);
11897 * Called from within softdep_disk_write_complete above.
11898 * A write operation was just completed. Removed inodes can
11899 * now be freed and associated block pointers may be committed.
11900 * Note that this routine is always called from interrupt level
11901 * with further splbio interrupts blocked.
11904 handle_written_filepage(pagedep, bp)
11905 struct pagedep *pagedep;
11906 struct buf *bp; /* buffer containing the written page */
11908 struct dirrem *dirrem;
11909 struct diradd *dap, *nextdap;
11913 if ((pagedep->pd_state & IOSTARTED) == 0)
11914 panic("handle_written_filepage: not started");
11915 pagedep->pd_state &= ~IOSTARTED;
11917 * Process any directory removals that have been committed.
11919 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
11920 LIST_REMOVE(dirrem, dm_next);
11921 dirrem->dm_state |= COMPLETE;
11922 dirrem->dm_dirinum = pagedep->pd_ino;
11923 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
11924 ("handle_written_filepage: Journal entries not written."));
11925 add_to_worklist(&dirrem->dm_list, 0);
11928 * Free any directory additions that have been committed.
11929 * If it is a newly allocated block, we have to wait until
11930 * the on-disk directory inode claims the new block.
11932 if ((pagedep->pd_state & NEWBLOCK) == 0)
11933 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
11934 free_diradd(dap, NULL);
11936 * Uncommitted directory entries must be restored.
11938 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
11939 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
11941 nextdap = LIST_NEXT(dap, da_pdlist);
11942 if (dap->da_state & ATTACHED)
11943 panic("handle_written_filepage: attached");
11944 ep = (struct direct *)
11945 ((char *)bp->b_data + dap->da_offset);
11946 ep->d_ino = dap->da_newinum;
11947 dap->da_state &= ~UNDONE;
11948 dap->da_state |= ATTACHED;
11951 * If the inode referenced by the directory has
11952 * been written out, then the dependency can be
11953 * moved to the pending list.
11955 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
11956 LIST_REMOVE(dap, da_pdlist);
11957 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
11963 * If there were any rollbacks in the directory, then it must be
11964 * marked dirty so that its will eventually get written back in
11965 * its correct form.
11968 if ((bp->b_flags & B_DELWRI) == 0)
11974 * If we are not waiting for a new directory block to be
11975 * claimed by its inode, then the pagedep will be freed.
11976 * Otherwise it will remain to track any new entries on
11977 * the page in case they are fsync'ed.
11979 free_pagedep(pagedep);
11984 * Writing back in-core inode structures.
11986 * The filesystem only accesses an inode's contents when it occupies an
11987 * "in-core" inode structure. These "in-core" structures are separate from
11988 * the page frames used to cache inode blocks. Only the latter are
11989 * transferred to/from the disk. So, when the updated contents of the
11990 * "in-core" inode structure are copied to the corresponding in-memory inode
11991 * block, the dependencies are also transferred. The following procedure is
11992 * called when copying a dirty "in-core" inode to a cached inode block.
11996 * Called when an inode is loaded from disk. If the effective link count
11997 * differed from the actual link count when it was last flushed, then we
11998 * need to ensure that the correct effective link count is put back.
12001 softdep_load_inodeblock(ip)
12002 struct inode *ip; /* the "in_core" copy of the inode */
12004 struct inodedep *inodedep;
12006 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12007 ("softdep_load_inodeblock called on non-softdep filesystem"));
12009 * Check for alternate nlink count.
12011 ip->i_effnlink = ip->i_nlink;
12012 ACQUIRE_LOCK(ip->i_ump);
12013 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
12015 FREE_LOCK(ip->i_ump);
12018 ip->i_effnlink -= inodedep->id_nlinkdelta;
12019 FREE_LOCK(ip->i_ump);
12023 * This routine is called just before the "in-core" inode
12024 * information is to be copied to the in-memory inode block.
12025 * Recall that an inode block contains several inodes. If
12026 * the force flag is set, then the dependencies will be
12027 * cleared so that the update can always be made. Note that
12028 * the buffer is locked when this routine is called, so we
12029 * will never be in the middle of writing the inode block
12033 softdep_update_inodeblock(ip, bp, waitfor)
12034 struct inode *ip; /* the "in_core" copy of the inode */
12035 struct buf *bp; /* the buffer containing the inode block */
12036 int waitfor; /* nonzero => update must be allowed */
12038 struct inodedep *inodedep;
12039 struct inoref *inoref;
12040 struct ufsmount *ump;
12041 struct worklist *wk;
12048 mp = UFSTOVFS(ump);
12049 KASSERT(MOUNTEDSOFTDEP(mp) != 0,
12050 ("softdep_update_inodeblock called on non-softdep filesystem"));
12053 * Preserve the freelink that is on disk. clear_unlinked_inodedep()
12054 * does not have access to the in-core ip so must write directly into
12055 * the inode block buffer when setting freelink.
12057 if (fs->fs_magic == FS_UFS1_MAGIC)
12058 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
12059 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12061 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
12062 ino_to_fsbo(fs, ip->i_number))->di_freelink);
12064 * If the effective link count is not equal to the actual link
12065 * count, then we must track the difference in an inodedep while
12066 * the inode is (potentially) tossed out of the cache. Otherwise,
12067 * if there is no existing inodedep, then there are no dependencies
12072 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12074 if (ip->i_effnlink != ip->i_nlink)
12075 panic("softdep_update_inodeblock: bad link count");
12078 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
12079 panic("softdep_update_inodeblock: bad delta");
12081 * If we're flushing all dependencies we must also move any waiting
12082 * for journal writes onto the bufwait list prior to I/O.
12085 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12086 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12088 jwait(&inoref->if_list, MNT_WAIT);
12094 * Changes have been initiated. Anything depending on these
12095 * changes cannot occur until this inode has been written.
12097 inodedep->id_state &= ~COMPLETE;
12098 if ((inodedep->id_state & ONWORKLIST) == 0)
12099 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
12101 * Any new dependencies associated with the incore inode must
12102 * now be moved to the list associated with the buffer holding
12103 * the in-memory copy of the inode. Once merged process any
12104 * allocdirects that are completed by the merger.
12106 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
12107 if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
12108 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
12110 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
12111 if (!TAILQ_EMPTY(&inodedep->id_extupdt))
12112 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
12115 * Now that the inode has been pushed into the buffer, the
12116 * operations dependent on the inode being written to disk
12117 * can be moved to the id_bufwait so that they will be
12118 * processed when the buffer I/O completes.
12120 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
12121 WORKLIST_REMOVE(wk);
12122 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
12125 * Newly allocated inodes cannot be written until the bitmap
12126 * that allocates them have been written (indicated by
12127 * DEPCOMPLETE being set in id_state). If we are doing a
12128 * forced sync (e.g., an fsync on a file), we force the bitmap
12129 * to be written so that the update can be done.
12131 if (waitfor == 0) {
12136 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
12140 ibp = inodedep->id_bmsafemap->sm_buf;
12141 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT);
12144 * If ibp came back as NULL, the dependency could have been
12145 * freed while we slept. Look it up again, and check to see
12146 * that it has completed.
12148 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
12154 if ((error = bwrite(ibp)) != 0)
12155 softdep_error("softdep_update_inodeblock: bwrite", error);
12159 * Merge the a new inode dependency list (such as id_newinoupdt) into an
12160 * old inode dependency list (such as id_inoupdt). This routine must be
12161 * called with splbio interrupts blocked.
12164 merge_inode_lists(newlisthead, oldlisthead)
12165 struct allocdirectlst *newlisthead;
12166 struct allocdirectlst *oldlisthead;
12168 struct allocdirect *listadp, *newadp;
12170 newadp = TAILQ_FIRST(newlisthead);
12171 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
12172 if (listadp->ad_offset < newadp->ad_offset) {
12173 listadp = TAILQ_NEXT(listadp, ad_next);
12176 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12177 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
12178 if (listadp->ad_offset == newadp->ad_offset) {
12179 allocdirect_merge(oldlisthead, newadp,
12183 newadp = TAILQ_FIRST(newlisthead);
12185 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
12186 TAILQ_REMOVE(newlisthead, newadp, ad_next);
12187 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
12192 * If we are doing an fsync, then we must ensure that any directory
12193 * entries for the inode have been written after the inode gets to disk.
12197 struct vnode *vp; /* the "in_core" copy of the inode */
12199 struct inodedep *inodedep;
12200 struct pagedep *pagedep;
12201 struct inoref *inoref;
12202 struct ufsmount *ump;
12203 struct worklist *wk;
12204 struct diradd *dap;
12210 struct thread *td = curthread;
12211 int error, flushparent, pagedep_new_block;
12219 if (MOUNTEDSOFTDEP(mp) == 0)
12223 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
12227 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12228 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12230 jwait(&inoref->if_list, MNT_WAIT);
12234 if (!LIST_EMPTY(&inodedep->id_inowait) ||
12235 !TAILQ_EMPTY(&inodedep->id_extupdt) ||
12236 !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
12237 !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
12238 !TAILQ_EMPTY(&inodedep->id_newinoupdt))
12239 panic("softdep_fsync: pending ops %p", inodedep);
12240 for (error = 0, flushparent = 0; ; ) {
12241 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
12243 if (wk->wk_type != D_DIRADD)
12244 panic("softdep_fsync: Unexpected type %s",
12245 TYPENAME(wk->wk_type));
12246 dap = WK_DIRADD(wk);
12248 * Flush our parent if this directory entry has a MKDIR_PARENT
12249 * dependency or is contained in a newly allocated block.
12251 if (dap->da_state & DIRCHG)
12252 pagedep = dap->da_previous->dm_pagedep;
12254 pagedep = dap->da_pagedep;
12255 parentino = pagedep->pd_ino;
12256 lbn = pagedep->pd_lbn;
12257 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
12258 panic("softdep_fsync: dirty");
12259 if ((dap->da_state & MKDIR_PARENT) ||
12260 (pagedep->pd_state & NEWBLOCK))
12265 * If we are being fsync'ed as part of vgone'ing this vnode,
12266 * then we will not be able to release and recover the
12267 * vnode below, so we just have to give up on writing its
12268 * directory entry out. It will eventually be written, just
12269 * not now, but then the user was not asking to have it
12270 * written, so we are not breaking any promises.
12272 if (vp->v_iflag & VI_DOOMED)
12275 * We prevent deadlock by always fetching inodes from the
12276 * root, moving down the directory tree. Thus, when fetching
12277 * our parent directory, we first try to get the lock. If
12278 * that fails, we must unlock ourselves before requesting
12279 * the lock on our parent. See the comment in ufs_lookup
12280 * for details on possible races.
12283 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
12284 FFSV_FORCEINSMQ)) {
12285 error = vfs_busy(mp, MBF_NOWAIT);
12289 error = vfs_busy(mp, 0);
12290 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12294 if (vp->v_iflag & VI_DOOMED) {
12300 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
12301 &pvp, FFSV_FORCEINSMQ);
12303 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12304 if (vp->v_iflag & VI_DOOMED) {
12313 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
12314 * that are contained in direct blocks will be resolved by
12315 * doing a ffs_update. Pagedeps contained in indirect blocks
12316 * may require a complete sync'ing of the directory. So, we
12317 * try the cheap and fast ffs_update first, and if that fails,
12318 * then we do the slower ffs_syncvnode of the directory.
12323 if ((error = ffs_update(pvp, 1)) != 0) {
12329 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
12330 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
12331 if (wk->wk_type != D_DIRADD)
12332 panic("softdep_fsync: Unexpected type %s",
12333 TYPENAME(wk->wk_type));
12334 dap = WK_DIRADD(wk);
12335 if (dap->da_state & DIRCHG)
12336 pagedep = dap->da_previous->dm_pagedep;
12338 pagedep = dap->da_pagedep;
12339 pagedep_new_block = pagedep->pd_state & NEWBLOCK;
12342 if (pagedep_new_block && (error =
12343 ffs_syncvnode(pvp, MNT_WAIT, 0))) {
12353 * Flush directory page containing the inode's name.
12355 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
12358 error = bwrite(bp);
12365 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
12373 * Flush all the dirty bitmaps associated with the block device
12374 * before flushing the rest of the dirty blocks so as to reduce
12375 * the number of dependencies that will have to be rolled back.
12380 softdep_fsync_mountdev(vp)
12383 struct buf *bp, *nbp;
12384 struct worklist *wk;
12387 if (!vn_isdisk(vp, NULL))
12388 panic("softdep_fsync_mountdev: vnode not a disk");
12389 bo = &vp->v_bufobj;
12392 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
12394 * If it is already scheduled, skip to the next buffer.
12396 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
12399 if ((bp->b_flags & B_DELWRI) == 0)
12400 panic("softdep_fsync_mountdev: not dirty");
12402 * We are only interested in bitmaps with outstanding
12405 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
12406 wk->wk_type != D_BMSAFEMAP ||
12407 (bp->b_vflags & BV_BKGRDINPROG)) {
12413 (void) bawrite(bp);
12421 * Sync all cylinder groups that were dirty at the time this function is
12422 * called. Newly dirtied cgs will be inserted before the sentinel. This
12423 * is used to flush freedep activity that may be holding up writes to a
12427 sync_cgs(mp, waitfor)
12431 struct bmsafemap *bmsafemap;
12432 struct bmsafemap *sentinel;
12433 struct ufsmount *ump;
12437 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK);
12438 sentinel->sm_cg = -1;
12439 ump = VFSTOUFS(mp);
12442 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next);
12443 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL;
12444 bmsafemap = LIST_NEXT(sentinel, sm_next)) {
12445 /* Skip sentinels and cgs with no work to release. */
12446 if (bmsafemap->sm_cg == -1 ||
12447 (LIST_EMPTY(&bmsafemap->sm_freehd) &&
12448 LIST_EMPTY(&bmsafemap->sm_freewr))) {
12449 LIST_REMOVE(sentinel, sm_next);
12450 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12454 * If we don't get the lock and we're waiting try again, if
12455 * not move on to the next buf and try to sync it.
12457 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor);
12458 if (bp == NULL && waitfor == MNT_WAIT)
12460 LIST_REMOVE(sentinel, sm_next);
12461 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12465 if (waitfor == MNT_NOWAIT)
12468 error = bwrite(bp);
12473 LIST_REMOVE(sentinel, sm_next);
12475 free(sentinel, M_BMSAFEMAP);
12480 * This routine is called when we are trying to synchronously flush a
12481 * file. This routine must eliminate any filesystem metadata dependencies
12482 * so that the syncing routine can succeed.
12485 softdep_sync_metadata(struct vnode *vp)
12491 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
12492 ("softdep_sync_metadata called on non-softdep filesystem"));
12494 * Ensure that any direct block dependencies have been cleared,
12495 * truncations are started, and inode references are journaled.
12497 ACQUIRE_LOCK(ip->i_ump);
12499 * Write all journal records to prevent rollbacks on devvp.
12501 if (vp->v_type == VCHR)
12502 softdep_flushjournal(vp->v_mount);
12503 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number);
12505 * Ensure that all truncates are written so we won't find deps on
12508 process_truncates(vp);
12509 FREE_LOCK(ip->i_ump);
12515 * This routine is called when we are attempting to sync a buf with
12516 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any
12517 * other IO it can but returns EBUSY if the buffer is not yet able to
12518 * be written. Dependencies which will not cause rollbacks will always
12522 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
12524 struct indirdep *indirdep;
12525 struct pagedep *pagedep;
12526 struct allocindir *aip;
12527 struct newblk *newblk;
12528 struct ufsmount *ump;
12530 struct worklist *wk;
12533 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
12534 ("softdep_sync_buf called on non-softdep filesystem"));
12536 * For VCHR we just don't want to force flush any dependencies that
12537 * will cause rollbacks.
12539 if (vp->v_type == VCHR) {
12540 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0))
12544 ump = VTOI(vp)->i_ump;
12547 * As we hold the buffer locked, none of its dependencies
12552 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
12553 switch (wk->wk_type) {
12555 case D_ALLOCDIRECT:
12557 newblk = WK_NEWBLK(wk);
12558 if (newblk->nb_jnewblk != NULL) {
12559 if (waitfor == MNT_NOWAIT) {
12563 jwait(&newblk->nb_jnewblk->jn_list, waitfor);
12566 if (newblk->nb_state & DEPCOMPLETE ||
12567 waitfor == MNT_NOWAIT)
12569 nbp = newblk->nb_bmsafemap->sm_buf;
12570 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12574 if ((error = bwrite(nbp)) != 0)
12580 indirdep = WK_INDIRDEP(wk);
12581 if (waitfor == MNT_NOWAIT) {
12582 if (!TAILQ_EMPTY(&indirdep->ir_trunc) ||
12583 !LIST_EMPTY(&indirdep->ir_deplisthd)) {
12588 if (!TAILQ_EMPTY(&indirdep->ir_trunc))
12589 panic("softdep_sync_buf: truncation pending.");
12591 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
12592 newblk = (struct newblk *)aip;
12593 if (newblk->nb_jnewblk != NULL) {
12594 jwait(&newblk->nb_jnewblk->jn_list,
12598 if (newblk->nb_state & DEPCOMPLETE)
12600 nbp = newblk->nb_bmsafemap->sm_buf;
12601 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor);
12605 if ((error = bwrite(nbp)) != 0)
12614 * Only flush directory entries in synchronous passes.
12616 if (waitfor != MNT_WAIT) {
12621 * While syncing snapshots, we must allow recursive
12626 * We are trying to sync a directory that may
12627 * have dependencies on both its own metadata
12628 * and/or dependencies on the inodes of any
12629 * recently allocated files. We walk its diradd
12630 * lists pushing out the associated inode.
12632 pagedep = WK_PAGEDEP(wk);
12633 for (i = 0; i < DAHASHSZ; i++) {
12634 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
12636 if ((error = flush_pagedep_deps(vp, wk->wk_mp,
12637 &pagedep->pd_diraddhd[i]))) {
12652 panic("softdep_sync_buf: Unknown type %s",
12653 TYPENAME(wk->wk_type));
12664 * Flush the dependencies associated with an inodedep.
12665 * Called with splbio blocked.
12668 flush_inodedep_deps(vp, mp, ino)
12673 struct inodedep *inodedep;
12674 struct inoref *inoref;
12675 struct ufsmount *ump;
12676 int error, waitfor;
12679 * This work is done in two passes. The first pass grabs most
12680 * of the buffers and begins asynchronously writing them. The
12681 * only way to wait for these asynchronous writes is to sleep
12682 * on the filesystem vnode which may stay busy for a long time
12683 * if the filesystem is active. So, instead, we make a second
12684 * pass over the dependencies blocking on each write. In the
12685 * usual case we will be blocking against a write that we
12686 * initiated, so when it is done the dependency will have been
12687 * resolved. Thus the second pass is expected to end quickly.
12688 * We give a brief window at the top of the loop to allow
12689 * any pending I/O to complete.
12691 ump = VFSTOUFS(mp);
12693 for (error = 0, waitfor = MNT_NOWAIT; ; ) {
12699 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
12701 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12702 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12704 jwait(&inoref->if_list, MNT_WAIT);
12708 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
12709 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
12710 flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
12711 flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
12714 * If pass2, we are done, otherwise do pass 2.
12716 if (waitfor == MNT_WAIT)
12718 waitfor = MNT_WAIT;
12721 * Try freeing inodedep in case all dependencies have been removed.
12723 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
12724 (void) free_inodedep(inodedep);
12729 * Flush an inode dependency list.
12730 * Called with splbio blocked.
12733 flush_deplist(listhead, waitfor, errorp)
12734 struct allocdirectlst *listhead;
12738 struct allocdirect *adp;
12739 struct newblk *newblk;
12740 struct ufsmount *ump;
12743 if ((adp = TAILQ_FIRST(listhead)) == NULL)
12745 ump = VFSTOUFS(adp->ad_list.wk_mp);
12747 TAILQ_FOREACH(adp, listhead, ad_next) {
12748 newblk = (struct newblk *)adp;
12749 if (newblk->nb_jnewblk != NULL) {
12750 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12753 if (newblk->nb_state & DEPCOMPLETE)
12755 bp = newblk->nb_bmsafemap->sm_buf;
12756 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor);
12758 if (waitfor == MNT_NOWAIT)
12763 if (waitfor == MNT_NOWAIT)
12766 *errorp = bwrite(bp);
12774 * Flush dependencies associated with an allocdirect block.
12777 flush_newblk_dep(vp, mp, lbn)
12782 struct newblk *newblk;
12783 struct ufsmount *ump;
12787 ufs2_daddr_t blkno;
12791 bo = &vp->v_bufobj;
12793 blkno = DIP(ip, i_db[lbn]);
12795 panic("flush_newblk_dep: Missing block");
12796 ump = VFSTOUFS(mp);
12799 * Loop until all dependencies related to this block are satisfied.
12800 * We must be careful to restart after each sleep in case a write
12801 * completes some part of this process for us.
12804 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
12808 if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
12809 panic("flush_newblk_deps: Bad newblk %p", newblk);
12811 * Flush the journal.
12813 if (newblk->nb_jnewblk != NULL) {
12814 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12818 * Write the bitmap dependency.
12820 if ((newblk->nb_state & DEPCOMPLETE) == 0) {
12821 bp = newblk->nb_bmsafemap->sm_buf;
12822 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
12826 error = bwrite(bp);
12833 * Write the buffer.
12837 bp = gbincore(bo, lbn);
12839 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
12840 LK_INTERLOCK, BO_LOCKPTR(bo));
12841 if (error == ENOLCK) {
12843 continue; /* Slept, retry */
12846 break; /* Failed */
12847 if (bp->b_flags & B_DELWRI) {
12849 error = bwrite(bp);
12857 * We have to wait for the direct pointers to
12858 * point at the newdirblk before the dependency
12861 error = ffs_update(vp, 1);
12870 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
12871 * Called with splbio blocked.
12874 flush_pagedep_deps(pvp, mp, diraddhdp)
12877 struct diraddhd *diraddhdp;
12879 struct inodedep *inodedep;
12880 struct inoref *inoref;
12881 struct ufsmount *ump;
12882 struct diradd *dap;
12887 struct diraddhd unfinished;
12889 LIST_INIT(&unfinished);
12890 ump = VFSTOUFS(mp);
12893 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
12895 * Flush ourselves if this directory entry
12896 * has a MKDIR_PARENT dependency.
12898 if (dap->da_state & MKDIR_PARENT) {
12900 if ((error = ffs_update(pvp, 1)) != 0)
12904 * If that cleared dependencies, go on to next.
12906 if (dap != LIST_FIRST(diraddhdp))
12909 * All MKDIR_PARENT dependencies and all the
12910 * NEWBLOCK pagedeps that are contained in direct
12911 * blocks were resolved by doing above ffs_update.
12912 * Pagedeps contained in indirect blocks may
12913 * require a complete sync'ing of the directory.
12914 * We are in the midst of doing a complete sync,
12915 * so if they are not resolved in this pass we
12916 * defer them for now as they will be sync'ed by
12917 * our caller shortly.
12919 LIST_REMOVE(dap, da_pdlist);
12920 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist);
12924 * A newly allocated directory must have its "." and
12925 * ".." entries written out before its name can be
12926 * committed in its parent.
12928 inum = dap->da_newinum;
12929 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12930 panic("flush_pagedep_deps: lost inode1");
12932 * Wait for any pending journal adds to complete so we don't
12933 * cause rollbacks while syncing.
12935 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12936 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12938 jwait(&inoref->if_list, MNT_WAIT);
12942 if (dap->da_state & MKDIR_BODY) {
12944 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
12947 error = flush_newblk_dep(vp, mp, 0);
12949 * If we still have the dependency we might need to
12950 * update the vnode to sync the new link count to
12953 if (error == 0 && dap == LIST_FIRST(diraddhdp))
12954 error = ffs_update(vp, 1);
12960 * If that cleared dependencies, go on to next.
12962 if (dap != LIST_FIRST(diraddhdp))
12964 if (dap->da_state & MKDIR_BODY) {
12965 inodedep_lookup(UFSTOVFS(ump), inum, 0,
12967 panic("flush_pagedep_deps: MKDIR_BODY "
12968 "inodedep %p dap %p vp %p",
12969 inodedep, dap, vp);
12973 * Flush the inode on which the directory entry depends.
12974 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
12975 * the only remaining dependency is that the updated inode
12976 * count must get pushed to disk. The inode has already
12977 * been pushed into its inode buffer (via VOP_UPDATE) at
12978 * the time of the reference count change. So we need only
12979 * locate that buffer, ensure that there will be no rollback
12980 * caused by a bitmap dependency, then write the inode buffer.
12983 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12984 panic("flush_pagedep_deps: lost inode");
12986 * If the inode still has bitmap dependencies,
12987 * push them to disk.
12989 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
12990 bp = inodedep->id_bmsafemap->sm_buf;
12991 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT);
12995 if ((error = bwrite(bp)) != 0)
12998 if (dap != LIST_FIRST(diraddhdp))
13002 * If the inode is still sitting in a buffer waiting
13003 * to be written or waiting for the link count to be
13004 * adjusted update it here to flush it to disk.
13006 if (dap == LIST_FIRST(diraddhdp)) {
13008 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
13011 error = ffs_update(vp, 1);
13018 * If we have failed to get rid of all the dependencies
13019 * then something is seriously wrong.
13021 if (dap == LIST_FIRST(diraddhdp)) {
13022 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
13023 panic("flush_pagedep_deps: failed to flush "
13024 "inodedep %p ino %ju dap %p",
13025 inodedep, (uintmax_t)inum, dap);
13030 while ((dap = LIST_FIRST(&unfinished)) != NULL) {
13031 LIST_REMOVE(dap, da_pdlist);
13032 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist);
13038 * A large burst of file addition or deletion activity can drive the
13039 * memory load excessively high. First attempt to slow things down
13040 * using the techniques below. If that fails, this routine requests
13041 * the offending operations to fall back to running synchronously
13042 * until the memory load returns to a reasonable level.
13045 softdep_slowdown(vp)
13048 struct ufsmount *ump;
13050 int max_softdeps_hard;
13052 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0,
13053 ("softdep_slowdown called on non-softdep filesystem"));
13054 ump = VFSTOUFS(vp->v_mount);
13058 * Check for journal space if needed.
13060 if (DOINGSUJ(vp)) {
13061 if (journal_space(ump, 0) == 0)
13065 * If the system is under its limits and our filesystem is
13066 * not responsible for more than our share of the usage and
13067 * we are not low on journal space, then no need to slow down.
13069 max_softdeps_hard = max_softdeps * 11 / 10;
13070 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 &&
13071 dep_current[D_INODEDEP] < max_softdeps_hard &&
13072 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 &&
13073 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 &&
13074 ump->softdep_curdeps[D_DIRREM] <
13075 (max_softdeps_hard / 2) / stat_flush_threads &&
13076 ump->softdep_curdeps[D_INODEDEP] <
13077 max_softdeps_hard / stat_flush_threads &&
13078 ump->softdep_curdeps[D_INDIRDEP] <
13079 (max_softdeps_hard / 1000) / stat_flush_threads &&
13080 ump->softdep_curdeps[D_FREEBLKS] <
13081 max_softdeps_hard / stat_flush_threads) {
13086 * If the journal is low or our filesystem is over its limit
13087 * then speedup the cleanup.
13089 if (ump->softdep_curdeps[D_INDIRDEP] <
13090 (max_softdeps_hard / 1000) / stat_flush_threads || jlow)
13091 softdep_speedup(ump);
13092 stat_sync_limit_hit += 1;
13095 * We only slow down the rate at which new dependencies are
13096 * generated if we are not using journaling. With journaling,
13097 * the cleanup should always be sufficient to keep things
13106 * Called by the allocation routines when they are about to fail
13107 * in the hope that we can free up the requested resource (inodes
13110 * First check to see if the work list has anything on it. If it has,
13111 * clean up entries until we successfully free the requested resource.
13112 * Because this process holds inodes locked, we cannot handle any remove
13113 * requests that might block on a locked inode as that could lead to
13114 * deadlock. If the worklist yields none of the requested resource,
13115 * start syncing out vnodes to free up the needed space.
13118 softdep_request_cleanup(fs, vp, cred, resource)
13121 struct ucred *cred;
13124 struct ufsmount *ump;
13126 struct vnode *lvp, *mvp;
13128 ufs2_daddr_t needed;
13132 * If we are being called because of a process doing a
13133 * copy-on-write, then it is not safe to process any
13134 * worklist items as we will recurse into the copyonwrite
13135 * routine. This will result in an incoherent snapshot.
13136 * If the vnode that we hold is a snapshot, we must avoid
13137 * handling other resources that could cause deadlock.
13139 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
13142 if (resource == FLUSH_BLOCKS_WAIT)
13143 stat_cleanup_blkrequests += 1;
13145 stat_cleanup_inorequests += 1;
13148 ump = VFSTOUFS(mp);
13149 mtx_assert(UFS_MTX(ump), MA_OWNED);
13151 error = ffs_update(vp, 1);
13152 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) {
13157 * If we are in need of resources, start by cleaning up
13158 * any block removals associated with our inode.
13161 process_removes(vp);
13162 process_truncates(vp);
13165 * Now clean up at least as many resources as we will need.
13167 * When requested to clean up inodes, the number that are needed
13168 * is set by the number of simultaneous writers (mnt_writeopcount)
13169 * plus a bit of slop (2) in case some more writers show up while
13172 * When requested to free up space, the amount of space that
13173 * we need is enough blocks to allocate a full-sized segment
13174 * (fs_contigsumsize). The number of such segments that will
13175 * be needed is set by the number of simultaneous writers
13176 * (mnt_writeopcount) plus a bit of slop (2) in case some more
13177 * writers show up while we are cleaning.
13179 * Additionally, if we are unpriviledged and allocating space,
13180 * we need to ensure that we clean up enough blocks to get the
13181 * needed number of blocks over the threshhold of the minimum
13182 * number of blocks required to be kept free by the filesystem
13185 if (resource == FLUSH_INODES_WAIT) {
13186 needed = vp->v_mount->mnt_writeopcount + 2;
13187 } else if (resource == FLUSH_BLOCKS_WAIT) {
13188 needed = (vp->v_mount->mnt_writeopcount + 2) *
13189 fs->fs_contigsumsize;
13190 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0))
13191 needed += fragstoblks(fs,
13192 roundup((fs->fs_dsize * fs->fs_minfree / 100) -
13193 fs->fs_cstotal.cs_nffree, fs->fs_frag));
13196 printf("softdep_request_cleanup: Unknown resource type %d\n",
13200 starttime = time_second;
13202 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 &&
13203 fs->fs_cstotal.cs_nbfree <= needed) ||
13204 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13205 fs->fs_cstotal.cs_nifree <= needed)) {
13207 if (ump->softdep_on_worklist > 0 &&
13208 process_worklist_item(UFSTOVFS(ump),
13209 ump->softdep_on_worklist, LK_NOWAIT) != 0)
13210 stat_worklist_push += 1;
13214 * If we still need resources and there are no more worklist
13215 * entries to process to obtain them, we have to start flushing
13216 * the dirty vnodes to force the release of additional requests
13217 * to the worklist that we can then process to reap addition
13218 * resources. We walk the vnodes associated with the mount point
13219 * until we get the needed worklist requests that we can reap.
13221 if ((resource == FLUSH_BLOCKS_WAIT &&
13222 fs->fs_cstotal.cs_nbfree <= needed) ||
13223 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
13224 fs->fs_cstotal.cs_nifree <= needed)) {
13225 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) {
13226 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) {
13230 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
13233 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */
13237 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0);
13240 lvp = ump->um_devvp;
13241 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
13242 VOP_FSYNC(lvp, MNT_NOWAIT, curthread);
13243 VOP_UNLOCK(lvp, 0);
13245 if (ump->softdep_on_worklist > 0) {
13246 stat_cleanup_retries += 1;
13249 stat_cleanup_failures += 1;
13251 if (time_second - starttime > stat_cleanup_high_delay)
13252 stat_cleanup_high_delay = time_second - starttime;
13258 softdep_excess_items(struct ufsmount *ump, int item)
13261 KASSERT(item >= 0 && item < D_LAST, ("item %d", item));
13262 return (dep_current[item] > max_softdeps &&
13263 ump->softdep_curdeps[item] > max_softdeps /
13264 stat_flush_threads);
13268 schedule_cleanup(struct mount *mp)
13270 struct ufsmount *ump;
13273 ump = VFSTOUFS(mp);
13277 if ((td->td_pflags & TDP_KTHREAD) != 0 &&
13278 (td->td_proc->p_flag2 & P2_AST_SU) == 0) {
13280 * No ast is delivered to kernel threads, so nobody
13281 * would deref the mp. Some kernel threads
13282 * explicitely check for AST, e.g. NFS daemon does
13283 * this in the serving loop.
13287 if (td->td_su != NULL)
13288 vfs_rel(td->td_su);
13292 td->td_flags |= TDF_ASTPENDING;
13297 softdep_ast_cleanup_proc(void)
13301 struct ufsmount *ump;
13306 while ((mp = td->td_su) != NULL) {
13308 error = vfs_busy(mp, MBF_NOWAIT);
13312 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) {
13313 ump = VFSTOUFS(mp);
13317 if (softdep_excess_items(ump, D_INODEDEP)) {
13319 request_cleanup(mp, FLUSH_INODES);
13321 if (softdep_excess_items(ump, D_DIRREM)) {
13323 request_cleanup(mp, FLUSH_BLOCKS);
13326 if (softdep_excess_items(ump, D_NEWBLK) ||
13327 softdep_excess_items(ump, D_ALLOCDIRECT) ||
13328 softdep_excess_items(ump, D_ALLOCINDIR)) {
13329 error = vn_start_write(NULL, &mp,
13333 VFS_SYNC(mp, MNT_WAIT);
13334 vn_finished_write(mp);
13337 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req)
13346 * If memory utilization has gotten too high, deliberately slow things
13347 * down and speed up the I/O processing.
13350 request_cleanup(mp, resource)
13354 struct thread *td = curthread;
13355 struct ufsmount *ump;
13357 ump = VFSTOUFS(mp);
13360 * We never hold up the filesystem syncer or buf daemon.
13362 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
13365 * First check to see if the work list has gotten backlogged.
13366 * If it has, co-opt this process to help clean up two entries.
13367 * Because this process may hold inodes locked, we cannot
13368 * handle any remove requests that might block on a locked
13369 * inode as that could lead to deadlock. We set TDP_SOFTDEP
13370 * to avoid recursively processing the worklist.
13372 if (ump->softdep_on_worklist > max_softdeps / 10) {
13373 td->td_pflags |= TDP_SOFTDEP;
13374 process_worklist_item(mp, 2, LK_NOWAIT);
13375 td->td_pflags &= ~TDP_SOFTDEP;
13376 stat_worklist_push += 2;
13380 * Next, we attempt to speed up the syncer process. If that
13381 * is successful, then we allow the process to continue.
13383 if (softdep_speedup(ump) &&
13384 resource != FLUSH_BLOCKS_WAIT &&
13385 resource != FLUSH_INODES_WAIT)
13388 * If we are resource constrained on inode dependencies, try
13389 * flushing some dirty inodes. Otherwise, we are constrained
13390 * by file deletions, so try accelerating flushes of directories
13391 * with removal dependencies. We would like to do the cleanup
13392 * here, but we probably hold an inode locked at this point and
13393 * that might deadlock against one that we try to clean. So,
13394 * the best that we can do is request the syncer daemon to do
13395 * the cleanup for us.
13397 switch (resource) {
13400 case FLUSH_INODES_WAIT:
13401 ACQUIRE_GBLLOCK(&lk);
13402 stat_ino_limit_push += 1;
13403 req_clear_inodedeps += 1;
13405 stat_countp = &stat_ino_limit_hit;
13409 case FLUSH_BLOCKS_WAIT:
13410 ACQUIRE_GBLLOCK(&lk);
13411 stat_blk_limit_push += 1;
13412 req_clear_remove += 1;
13414 stat_countp = &stat_blk_limit_hit;
13418 panic("request_cleanup: unknown type");
13421 * Hopefully the syncer daemon will catch up and awaken us.
13422 * We wait at most tickdelay before proceeding in any case.
13424 ACQUIRE_GBLLOCK(&lk);
13427 if (callout_pending(&softdep_callout) == FALSE)
13428 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13431 if ((td->td_pflags & TDP_KTHREAD) == 0)
13432 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
13440 * Awaken processes pausing in request_cleanup and clear proc_waiting
13441 * to indicate that there is no longer a timer running. Pause_timer
13442 * will be called with the global softdep mutex (&lk) locked.
13449 GBLLOCK_OWNED(&lk);
13451 * The callout_ API has acquired mtx and will hold it around this
13454 *stat_countp += proc_waiting;
13455 wakeup(&proc_waiting);
13459 * If requested, try removing inode or removal dependencies.
13462 check_clear_deps(mp)
13467 * If we are suspended, it may be because of our using
13468 * too many inodedeps, so help clear them out.
13470 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended)
13471 clear_inodedeps(mp);
13473 * General requests for cleanup of backed up dependencies
13475 ACQUIRE_GBLLOCK(&lk);
13476 if (req_clear_inodedeps) {
13477 req_clear_inodedeps -= 1;
13479 clear_inodedeps(mp);
13480 ACQUIRE_GBLLOCK(&lk);
13481 wakeup(&proc_waiting);
13483 if (req_clear_remove) {
13484 req_clear_remove -= 1;
13487 ACQUIRE_GBLLOCK(&lk);
13488 wakeup(&proc_waiting);
13494 * Flush out a directory with at least one removal dependency in an effort to
13495 * reduce the number of dirrem, freefile, and freeblks dependency structures.
13501 struct pagedep_hashhead *pagedephd;
13502 struct pagedep *pagedep;
13503 struct ufsmount *ump;
13509 ump = VFSTOUFS(mp);
13512 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) {
13513 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++];
13514 if (ump->pagedep_nextclean > ump->pagedep_hash_size)
13515 ump->pagedep_nextclean = 0;
13516 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
13517 if (LIST_EMPTY(&pagedep->pd_dirremhd))
13519 ino = pagedep->pd_ino;
13520 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13525 * Let unmount clear deps
13527 error = vfs_busy(mp, MBF_NOWAIT);
13530 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13534 softdep_error("clear_remove: vget", error);
13537 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13538 softdep_error("clear_remove: fsync", error);
13539 bo = &vp->v_bufobj;
13545 vn_finished_write(mp);
13553 * Clear out a block of dirty inodes in an effort to reduce
13554 * the number of inodedep dependency structures.
13557 clear_inodedeps(mp)
13560 struct inodedep_hashhead *inodedephd;
13561 struct inodedep *inodedep;
13562 struct ufsmount *ump;
13566 ino_t firstino, lastino, ino;
13568 ump = VFSTOUFS(mp);
13572 * Pick a random inode dependency to be cleared.
13573 * We will then gather up all the inodes in its block
13574 * that have dependencies and flush them out.
13576 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) {
13577 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++];
13578 if (ump->inodedep_nextclean > ump->inodedep_hash_size)
13579 ump->inodedep_nextclean = 0;
13580 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
13583 if (inodedep == NULL)
13586 * Find the last inode in the block with dependencies.
13588 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
13589 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
13590 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
13593 * Asynchronously push all but the last inode with dependencies.
13594 * Synchronously push the last inode with dependencies to ensure
13595 * that the inode block gets written to free up the inodedeps.
13597 for (ino = firstino; ino <= lastino; ino++) {
13598 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
13600 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13603 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
13605 vn_finished_write(mp);
13609 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13610 FFSV_FORCEINSMQ)) != 0) {
13611 softdep_error("clear_inodedeps: vget", error);
13613 vn_finished_write(mp);
13618 if (ino == lastino) {
13619 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)))
13620 softdep_error("clear_inodedeps: fsync1", error);
13622 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13623 softdep_error("clear_inodedeps: fsync2", error);
13624 BO_LOCK(&vp->v_bufobj);
13626 BO_UNLOCK(&vp->v_bufobj);
13629 vn_finished_write(mp);
13635 softdep_buf_append(bp, wkhd)
13637 struct workhead *wkhd;
13639 struct worklist *wk;
13640 struct ufsmount *ump;
13642 if ((wk = LIST_FIRST(wkhd)) == NULL)
13644 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13645 ("softdep_buf_append called on non-softdep filesystem"));
13646 ump = VFSTOUFS(wk->wk_mp);
13648 while ((wk = LIST_FIRST(wkhd)) != NULL) {
13649 WORKLIST_REMOVE(wk);
13650 WORKLIST_INSERT(&bp->b_dep, wk);
13657 softdep_inode_append(ip, cred, wkhd)
13659 struct ucred *cred;
13660 struct workhead *wkhd;
13666 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0,
13667 ("softdep_inode_append called on non-softdep filesystem"));
13669 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
13670 (int)fs->fs_bsize, cred, &bp);
13673 softdep_freework(wkhd);
13676 softdep_buf_append(bp, wkhd);
13681 softdep_freework(wkhd)
13682 struct workhead *wkhd;
13684 struct worklist *wk;
13685 struct ufsmount *ump;
13687 if ((wk = LIST_FIRST(wkhd)) == NULL)
13689 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0,
13690 ("softdep_freework called on non-softdep filesystem"));
13691 ump = VFSTOUFS(wk->wk_mp);
13693 handle_jwork(wkhd);
13698 * Function to determine if the buffer has outstanding dependencies
13699 * that will cause a roll-back if the buffer is written. If wantcount
13700 * is set, return number of dependencies, otherwise just yes or no.
13703 softdep_count_dependencies(bp, wantcount)
13707 struct worklist *wk;
13708 struct ufsmount *ump;
13709 struct bmsafemap *bmsafemap;
13710 struct freework *freework;
13711 struct inodedep *inodedep;
13712 struct indirdep *indirdep;
13713 struct freeblks *freeblks;
13714 struct allocindir *aip;
13715 struct pagedep *pagedep;
13716 struct dirrem *dirrem;
13717 struct newblk *newblk;
13718 struct mkdir *mkdir;
13719 struct diradd *dap;
13723 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL)
13725 ump = VFSTOUFS(wk->wk_mp);
13727 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
13728 switch (wk->wk_type) {
13731 inodedep = WK_INODEDEP(wk);
13732 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
13733 /* bitmap allocation dependency */
13738 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
13739 /* direct block pointer dependency */
13744 if (TAILQ_FIRST(&inodedep->id_extupdt)) {
13745 /* direct block pointer dependency */
13750 if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
13751 /* Add reference dependency. */
13759 indirdep = WK_INDIRDEP(wk);
13761 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) {
13762 /* indirect truncation dependency */
13768 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
13769 /* indirect block pointer dependency */
13777 pagedep = WK_PAGEDEP(wk);
13778 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
13779 if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
13780 /* Journal remove ref dependency. */
13786 for (i = 0; i < DAHASHSZ; i++) {
13788 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
13789 /* directory entry dependency */
13798 bmsafemap = WK_BMSAFEMAP(wk);
13799 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
13800 /* Add reference dependency. */
13805 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
13806 /* Allocate block dependency. */
13814 freeblks = WK_FREEBLKS(wk);
13815 if (LIST_FIRST(&freeblks->fb_jblkdephd)) {
13816 /* Freeblk journal dependency. */
13823 case D_ALLOCDIRECT:
13825 newblk = WK_NEWBLK(wk);
13826 if (newblk->nb_jnewblk) {
13827 /* Journal allocate dependency. */
13835 mkdir = WK_MKDIR(wk);
13836 if (mkdir->md_jaddref) {
13837 /* Journal reference dependency. */
13849 /* never a dependency on these blocks */
13853 panic("softdep_count_dependencies: Unexpected type %s",
13854 TYPENAME(wk->wk_type));
13864 * Acquire exclusive access to a buffer.
13865 * Must be called with a locked mtx parameter.
13866 * Return acquired buffer or NULL on failure.
13868 static struct buf *
13869 getdirtybuf(bp, lock, waitfor)
13871 struct rwlock *lock;
13876 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
13877 if (waitfor != MNT_WAIT)
13879 error = BUF_LOCK(bp,
13880 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock);
13882 * Even if we sucessfully acquire bp here, we have dropped
13883 * lock, which may violates our guarantee.
13887 else if (error != ENOLCK)
13888 panic("getdirtybuf: inconsistent lock: %d", error);
13892 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13893 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) {
13895 BO_LOCK(bp->b_bufobj);
13897 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13898 bp->b_vflags |= BV_BKGRDWAIT;
13899 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj),
13900 PRIBIO | PDROP, "getbuf", 0);
13902 BO_UNLOCK(bp->b_bufobj);
13907 if (waitfor != MNT_WAIT)
13910 * The lock argument must be bp->b_vp's mutex in
13913 #ifdef DEBUG_VFS_LOCKS
13914 if (bp->b_vp->v_type != VCHR)
13915 ASSERT_BO_WLOCKED(bp->b_bufobj);
13917 bp->b_vflags |= BV_BKGRDWAIT;
13918 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0);
13921 if ((bp->b_flags & B_DELWRI) == 0) {
13931 * Check if it is safe to suspend the file system now. On entry,
13932 * the vnode interlock for devvp should be held. Return 0 with
13933 * the mount interlock held if the file system can be suspended now,
13934 * otherwise return EAGAIN with the mount interlock held.
13937 softdep_check_suspend(struct mount *mp,
13938 struct vnode *devvp,
13939 int softdep_depcnt,
13940 int softdep_accdepcnt,
13941 int secondary_writes,
13942 int secondary_accwrites)
13945 struct ufsmount *ump;
13946 struct inodedep *inodedep;
13947 int error, unlinked;
13949 bo = &devvp->v_bufobj;
13950 ASSERT_BO_WLOCKED(bo);
13953 * If we are not running with soft updates, then we need only
13954 * deal with secondary writes as we try to suspend.
13956 if (MOUNTEDSOFTDEP(mp) == 0) {
13958 while (mp->mnt_secondary_writes != 0) {
13960 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
13961 (PUSER - 1) | PDROP, "secwr", 0);
13967 * Reasons for needing more work before suspend:
13968 * - Dirty buffers on devvp.
13969 * - Secondary writes occurred after start of vnode sync loop
13972 if (bo->bo_numoutput > 0 ||
13973 bo->bo_dirty.bv_cnt > 0 ||
13974 secondary_writes != 0 ||
13975 mp->mnt_secondary_writes != 0 ||
13976 secondary_accwrites != mp->mnt_secondary_accwrites)
13983 * If we are running with soft updates, then we need to coordinate
13984 * with them as we try to suspend.
13986 ump = VFSTOUFS(mp);
13988 if (!TRY_ACQUIRE_LOCK(ump)) {
13996 if (mp->mnt_secondary_writes != 0) {
13999 msleep(&mp->mnt_secondary_writes,
14001 (PUSER - 1) | PDROP, "secwr", 0);
14009 if (MOUNTEDSUJ(mp)) {
14010 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked);
14012 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
14013 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS |
14014 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS |
14016 !check_inodedep_free(inodedep))
14023 * Reasons for needing more work before suspend:
14024 * - Dirty buffers on devvp.
14025 * - Softdep activity occurred after start of vnode sync loop
14026 * - Secondary writes occurred after start of vnode sync loop
14029 if (bo->bo_numoutput > 0 ||
14030 bo->bo_dirty.bv_cnt > 0 ||
14031 softdep_depcnt != unlinked ||
14032 ump->softdep_deps != unlinked ||
14033 softdep_accdepcnt != ump->softdep_accdeps ||
14034 secondary_writes != 0 ||
14035 mp->mnt_secondary_writes != 0 ||
14036 secondary_accwrites != mp->mnt_secondary_accwrites)
14045 * Get the number of dependency structures for the file system, both
14046 * the current number and the total number allocated. These will
14047 * later be used to detect that softdep processing has occurred.
14050 softdep_get_depcounts(struct mount *mp,
14051 int *softdep_depsp,
14052 int *softdep_accdepsp)
14054 struct ufsmount *ump;
14056 if (MOUNTEDSOFTDEP(mp) == 0) {
14057 *softdep_depsp = 0;
14058 *softdep_accdepsp = 0;
14061 ump = VFSTOUFS(mp);
14063 *softdep_depsp = ump->softdep_deps;
14064 *softdep_accdepsp = ump->softdep_accdeps;
14069 * Wait for pending output on a vnode to complete.
14070 * Must be called with vnode lock and interlock locked.
14072 * XXX: Should just be a call to bufobj_wwait().
14080 bo = &vp->v_bufobj;
14081 ASSERT_VOP_LOCKED(vp, "drain_output");
14082 ASSERT_BO_WLOCKED(bo);
14084 while (bo->bo_numoutput) {
14085 bo->bo_flag |= BO_WWAIT;
14086 msleep((caddr_t)&bo->bo_numoutput,
14087 BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0);
14092 * Called whenever a buffer that is being invalidated or reallocated
14093 * contains dependencies. This should only happen if an I/O error has
14094 * occurred. The routine is called with the buffer locked.
14097 softdep_deallocate_dependencies(bp)
14101 if ((bp->b_ioflags & BIO_ERROR) == 0)
14102 panic("softdep_deallocate_dependencies: dangling deps");
14103 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL)
14104 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
14106 printf("softdep_deallocate_dependencies: "
14107 "got error %d while accessing filesystem\n", bp->b_error);
14108 if (bp->b_error != ENXIO)
14109 panic("softdep_deallocate_dependencies: unrecovered I/O error");
14113 * Function to handle asynchronous write errors in the filesystem.
14116 softdep_error(func, error)
14121 /* XXX should do something better! */
14122 printf("%s: got error %d while accessing filesystem\n", func, error);
14128 inodedep_print(struct inodedep *inodedep, int verbose)
14130 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d"
14132 inodedep, inodedep->id_fs, inodedep->id_state,
14133 (intmax_t)inodedep->id_ino,
14134 (intmax_t)fsbtodb(inodedep->id_fs,
14135 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
14136 inodedep->id_nlinkdelta, inodedep->id_savednlink,
14137 inodedep->id_savedino1);
14142 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
14144 LIST_FIRST(&inodedep->id_pendinghd),
14145 LIST_FIRST(&inodedep->id_bufwait),
14146 LIST_FIRST(&inodedep->id_inowait),
14147 TAILQ_FIRST(&inodedep->id_inoreflst),
14148 inodedep->id_mkdiradd);
14149 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
14150 TAILQ_FIRST(&inodedep->id_inoupdt),
14151 TAILQ_FIRST(&inodedep->id_newinoupdt),
14152 TAILQ_FIRST(&inodedep->id_extupdt),
14153 TAILQ_FIRST(&inodedep->id_newextupdt));
14156 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
14159 if (have_addr == 0) {
14160 db_printf("Address required\n");
14163 inodedep_print((struct inodedep*)addr, 1);
14166 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
14168 struct inodedep_hashhead *inodedephd;
14169 struct inodedep *inodedep;
14170 struct ufsmount *ump;
14173 if (have_addr == 0) {
14174 db_printf("Address required\n");
14177 ump = (struct ufsmount *)addr;
14178 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) {
14179 inodedephd = &ump->inodedep_hashtbl[cnt];
14180 LIST_FOREACH(inodedep, inodedephd, id_hash) {
14181 inodedep_print(inodedep, 0);
14186 DB_SHOW_COMMAND(worklist, db_show_worklist)
14188 struct worklist *wk;
14190 if (have_addr == 0) {
14191 db_printf("Address required\n");
14194 wk = (struct worklist *)addr;
14195 printf("worklist: %p type %s state 0x%X\n",
14196 wk, TYPENAME(wk->wk_type), wk->wk_state);
14199 DB_SHOW_COMMAND(workhead, db_show_workhead)
14201 struct workhead *wkhd;
14202 struct worklist *wk;
14205 if (have_addr == 0) {
14206 db_printf("Address required\n");
14209 wkhd = (struct workhead *)addr;
14210 wk = LIST_FIRST(wkhd);
14211 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
14212 db_printf("worklist: %p type %s state 0x%X",
14213 wk, TYPENAME(wk->wk_type), wk->wk_state);
14215 db_printf("workhead overflow");
14220 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
14222 struct mkdirlist *mkdirlisthd;
14223 struct jaddref *jaddref;
14224 struct diradd *diradd;
14225 struct mkdir *mkdir;
14227 if (have_addr == 0) {
14228 db_printf("Address required\n");
14231 mkdirlisthd = (struct mkdirlist *)addr;
14232 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) {
14233 diradd = mkdir->md_diradd;
14234 db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
14235 mkdir, mkdir->md_state, diradd, diradd->da_state);
14236 if ((jaddref = mkdir->md_jaddref) != NULL)
14237 db_printf(" jaddref %p jaddref state 0x%X",
14238 jaddref, jaddref->ja_state);
14243 /* exported to ffs_vfsops.c */
14244 extern void db_print_ffs(struct ufsmount *ump);
14246 db_print_ffs(struct ufsmount *ump)
14248 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n",
14249 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname,
14250 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist,
14251 ump->softdep_deps, ump->softdep_req);
14256 #endif /* SOFTUPDATES */