2 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_da_btree.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_alloc.h"
43 #include "xfs_quota.h"
44 #include "xfs_trans_priv.h"
45 #include "xfs_trans_space.h"
46 #include "xfs_inode_item.h"
47 #include "xfs_trace.h"
49 kmem_zone_t *xfs_trans_zone;
52 * Reservation functions here avoid a huge stack in xfs_trans_init
53 * due to register overflow from temporaries in the calculations.
56 xfs_calc_write_reservation(xfs_mount_t *mp)
58 return XFS_CALC_WRITE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
62 xfs_calc_itruncate_reservation(xfs_mount_t *mp)
64 return XFS_CALC_ITRUNCATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
68 xfs_calc_rename_reservation(xfs_mount_t *mp)
70 return XFS_CALC_RENAME_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
74 xfs_calc_link_reservation(xfs_mount_t *mp)
76 return XFS_CALC_LINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
80 xfs_calc_remove_reservation(xfs_mount_t *mp)
82 return XFS_CALC_REMOVE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
86 xfs_calc_symlink_reservation(xfs_mount_t *mp)
88 return XFS_CALC_SYMLINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
92 xfs_calc_create_reservation(xfs_mount_t *mp)
94 return XFS_CALC_CREATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
98 xfs_calc_mkdir_reservation(xfs_mount_t *mp)
100 return XFS_CALC_MKDIR_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
104 xfs_calc_ifree_reservation(xfs_mount_t *mp)
106 return XFS_CALC_IFREE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
110 xfs_calc_ichange_reservation(xfs_mount_t *mp)
112 return XFS_CALC_ICHANGE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
116 xfs_calc_growdata_reservation(xfs_mount_t *mp)
118 return XFS_CALC_GROWDATA_LOG_RES(mp);
122 xfs_calc_growrtalloc_reservation(xfs_mount_t *mp)
124 return XFS_CALC_GROWRTALLOC_LOG_RES(mp);
128 xfs_calc_growrtzero_reservation(xfs_mount_t *mp)
130 return XFS_CALC_GROWRTZERO_LOG_RES(mp);
134 xfs_calc_growrtfree_reservation(xfs_mount_t *mp)
136 return XFS_CALC_GROWRTFREE_LOG_RES(mp);
140 xfs_calc_swrite_reservation(xfs_mount_t *mp)
142 return XFS_CALC_SWRITE_LOG_RES(mp);
146 xfs_calc_writeid_reservation(xfs_mount_t *mp)
148 return XFS_CALC_WRITEID_LOG_RES(mp);
152 xfs_calc_addafork_reservation(xfs_mount_t *mp)
154 return XFS_CALC_ADDAFORK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
158 xfs_calc_attrinval_reservation(xfs_mount_t *mp)
160 return XFS_CALC_ATTRINVAL_LOG_RES(mp);
164 xfs_calc_attrset_reservation(xfs_mount_t *mp)
166 return XFS_CALC_ATTRSET_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
170 xfs_calc_attrrm_reservation(xfs_mount_t *mp)
172 return XFS_CALC_ATTRRM_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
176 xfs_calc_clear_agi_bucket_reservation(xfs_mount_t *mp)
178 return XFS_CALC_CLEAR_AGI_BUCKET_LOG_RES(mp);
182 * Initialize the precomputed transaction reservation values
183 * in the mount structure.
189 xfs_trans_reservations_t *resp;
191 resp = &(mp->m_reservations);
192 resp->tr_write = xfs_calc_write_reservation(mp);
193 resp->tr_itruncate = xfs_calc_itruncate_reservation(mp);
194 resp->tr_rename = xfs_calc_rename_reservation(mp);
195 resp->tr_link = xfs_calc_link_reservation(mp);
196 resp->tr_remove = xfs_calc_remove_reservation(mp);
197 resp->tr_symlink = xfs_calc_symlink_reservation(mp);
198 resp->tr_create = xfs_calc_create_reservation(mp);
199 resp->tr_mkdir = xfs_calc_mkdir_reservation(mp);
200 resp->tr_ifree = xfs_calc_ifree_reservation(mp);
201 resp->tr_ichange = xfs_calc_ichange_reservation(mp);
202 resp->tr_growdata = xfs_calc_growdata_reservation(mp);
203 resp->tr_swrite = xfs_calc_swrite_reservation(mp);
204 resp->tr_writeid = xfs_calc_writeid_reservation(mp);
205 resp->tr_addafork = xfs_calc_addafork_reservation(mp);
206 resp->tr_attrinval = xfs_calc_attrinval_reservation(mp);
207 resp->tr_attrset = xfs_calc_attrset_reservation(mp);
208 resp->tr_attrrm = xfs_calc_attrrm_reservation(mp);
209 resp->tr_clearagi = xfs_calc_clear_agi_bucket_reservation(mp);
210 resp->tr_growrtalloc = xfs_calc_growrtalloc_reservation(mp);
211 resp->tr_growrtzero = xfs_calc_growrtzero_reservation(mp);
212 resp->tr_growrtfree = xfs_calc_growrtfree_reservation(mp);
216 * This routine is called to allocate a transaction structure.
217 * The type parameter indicates the type of the transaction. These
218 * are enumerated in xfs_trans.h.
220 * Dynamically allocate the transaction structure from the transaction
221 * zone, initialize it, and return it to the caller.
228 xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
229 return _xfs_trans_alloc(mp, type, KM_SLEEP);
240 atomic_inc(&mp->m_active_trans);
242 tp = kmem_zone_zalloc(xfs_trans_zone, memflags);
243 tp->t_magic = XFS_TRANS_MAGIC;
246 tp->t_items_free = XFS_LIC_NUM_SLOTS;
247 xfs_lic_init(&(tp->t_items));
248 INIT_LIST_HEAD(&tp->t_busy);
253 * Free the transaction structure. If there is more clean up
254 * to do when the structure is freed, add it here.
258 struct xfs_trans *tp)
260 struct xfs_busy_extent *busyp, *n;
262 list_for_each_entry_safe(busyp, n, &tp->t_busy, list)
263 xfs_alloc_busy_clear(tp->t_mountp, busyp);
265 atomic_dec(&tp->t_mountp->m_active_trans);
266 xfs_trans_free_dqinfo(tp);
267 kmem_zone_free(xfs_trans_zone, tp);
271 * This is called to create a new transaction which will share the
272 * permanent log reservation of the given transaction. The remaining
273 * unused block and rt extent reservations are also inherited. This
274 * implies that the original transaction is no longer allowed to allocate
275 * blocks. Locks and log items, however, are no inherited. They must
276 * be added to the new transaction explicitly.
284 ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
287 * Initialize the new transaction structure.
289 ntp->t_magic = XFS_TRANS_MAGIC;
290 ntp->t_type = tp->t_type;
291 ntp->t_mountp = tp->t_mountp;
292 ntp->t_items_free = XFS_LIC_NUM_SLOTS;
293 xfs_lic_init(&(ntp->t_items));
294 INIT_LIST_HEAD(&ntp->t_busy);
296 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
297 ASSERT(tp->t_ticket != NULL);
299 ntp->t_flags = XFS_TRANS_PERM_LOG_RES | (tp->t_flags & XFS_TRANS_RESERVE);
300 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
301 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
302 tp->t_blk_res = tp->t_blk_res_used;
303 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
304 tp->t_rtx_res = tp->t_rtx_res_used;
305 ntp->t_pflags = tp->t_pflags;
307 xfs_trans_dup_dqinfo(tp, ntp);
309 atomic_inc(&tp->t_mountp->m_active_trans);
314 * This is called to reserve free disk blocks and log space for the
315 * given transaction. This must be done before allocating any resources
316 * within the transaction.
318 * This will return ENOSPC if there are not enough blocks available.
319 * It will sleep waiting for available log space.
320 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
321 * is used by long running transactions. If any one of the reservations
322 * fails then they will all be backed out.
324 * This does not do quota reservations. That typically is done by the
338 int rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
340 /* Mark this thread as being in a transaction */
341 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
344 * Attempt to reserve the needed disk blocks by decrementing
345 * the number needed from the number available. This will
346 * fail if the count would go below zero.
349 error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
350 -((int64_t)blocks), rsvd);
352 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
353 return (XFS_ERROR(ENOSPC));
355 tp->t_blk_res += blocks;
359 * Reserve the log space needed for this transaction.
362 ASSERT((tp->t_log_res == 0) || (tp->t_log_res == logspace));
363 ASSERT((tp->t_log_count == 0) ||
364 (tp->t_log_count == logcount));
365 if (flags & XFS_TRANS_PERM_LOG_RES) {
366 log_flags = XFS_LOG_PERM_RESERV;
367 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
369 ASSERT(tp->t_ticket == NULL);
370 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
374 error = xfs_log_reserve(tp->t_mountp, logspace, logcount,
376 XFS_TRANSACTION, log_flags, tp->t_type);
380 tp->t_log_res = logspace;
381 tp->t_log_count = logcount;
385 * Attempt to reserve the needed realtime extents by decrementing
386 * the number needed from the number available. This will
387 * fail if the count would go below zero.
390 error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FREXTENTS,
391 -((int64_t)rtextents), rsvd);
393 error = XFS_ERROR(ENOSPC);
396 tp->t_rtx_res += rtextents;
402 * Error cases jump to one of these labels to undo any
403 * reservations which have already been performed.
407 if (flags & XFS_TRANS_PERM_LOG_RES) {
408 log_flags = XFS_LOG_REL_PERM_RESERV;
412 xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, log_flags);
415 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
420 (void) xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
421 (int64_t)blocks, rsvd);
425 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
431 * Record the indicated change to the given field for application
432 * to the file system's superblock when the transaction commits.
433 * For now, just store the change in the transaction structure.
435 * Mark the transaction structure to indicate that the superblock
436 * needs to be updated before committing.
438 * Because we may not be keeping track of allocated/free inodes and
439 * used filesystem blocks in the superblock, we do not mark the
440 * superblock dirty in this transaction if we modify these fields.
441 * We still need to update the transaction deltas so that they get
442 * applied to the incore superblock, but we don't want them to
443 * cause the superblock to get locked and logged if these are the
444 * only fields in the superblock that the transaction modifies.
452 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
453 xfs_mount_t *mp = tp->t_mountp;
456 case XFS_TRANS_SB_ICOUNT:
457 tp->t_icount_delta += delta;
458 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
459 flags &= ~XFS_TRANS_SB_DIRTY;
461 case XFS_TRANS_SB_IFREE:
462 tp->t_ifree_delta += delta;
463 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
464 flags &= ~XFS_TRANS_SB_DIRTY;
466 case XFS_TRANS_SB_FDBLOCKS:
468 * Track the number of blocks allocated in the
469 * transaction. Make sure it does not exceed the
473 tp->t_blk_res_used += (uint)-delta;
474 ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
476 tp->t_fdblocks_delta += delta;
477 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
478 flags &= ~XFS_TRANS_SB_DIRTY;
480 case XFS_TRANS_SB_RES_FDBLOCKS:
482 * The allocation has already been applied to the
483 * in-core superblock's counter. This should only
484 * be applied to the on-disk superblock.
487 tp->t_res_fdblocks_delta += delta;
488 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
489 flags &= ~XFS_TRANS_SB_DIRTY;
491 case XFS_TRANS_SB_FREXTENTS:
493 * Track the number of blocks allocated in the
494 * transaction. Make sure it does not exceed the
498 tp->t_rtx_res_used += (uint)-delta;
499 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
501 tp->t_frextents_delta += delta;
503 case XFS_TRANS_SB_RES_FREXTENTS:
505 * The allocation has already been applied to the
506 * in-core superblock's counter. This should only
507 * be applied to the on-disk superblock.
510 tp->t_res_frextents_delta += delta;
512 case XFS_TRANS_SB_DBLOCKS:
514 tp->t_dblocks_delta += delta;
516 case XFS_TRANS_SB_AGCOUNT:
518 tp->t_agcount_delta += delta;
520 case XFS_TRANS_SB_IMAXPCT:
521 tp->t_imaxpct_delta += delta;
523 case XFS_TRANS_SB_REXTSIZE:
524 tp->t_rextsize_delta += delta;
526 case XFS_TRANS_SB_RBMBLOCKS:
527 tp->t_rbmblocks_delta += delta;
529 case XFS_TRANS_SB_RBLOCKS:
530 tp->t_rblocks_delta += delta;
532 case XFS_TRANS_SB_REXTENTS:
533 tp->t_rextents_delta += delta;
535 case XFS_TRANS_SB_REXTSLOG:
536 tp->t_rextslog_delta += delta;
543 tp->t_flags |= flags;
547 * xfs_trans_apply_sb_deltas() is called from the commit code
548 * to bring the superblock buffer into the current transaction
549 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
551 * For now we just look at each field allowed to change and change
555 xfs_trans_apply_sb_deltas(
562 bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
563 sbp = XFS_BUF_TO_SBP(bp);
566 * Check that superblock mods match the mods made to AGF counters.
568 ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
569 (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
570 tp->t_ag_btree_delta));
573 * Only update the superblock counters if we are logging them
575 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
576 if (tp->t_icount_delta)
577 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
578 if (tp->t_ifree_delta)
579 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
580 if (tp->t_fdblocks_delta)
581 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
582 if (tp->t_res_fdblocks_delta)
583 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
586 if (tp->t_frextents_delta)
587 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
588 if (tp->t_res_frextents_delta)
589 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
591 if (tp->t_dblocks_delta) {
592 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
595 if (tp->t_agcount_delta) {
596 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
599 if (tp->t_imaxpct_delta) {
600 sbp->sb_imax_pct += tp->t_imaxpct_delta;
603 if (tp->t_rextsize_delta) {
604 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
607 if (tp->t_rbmblocks_delta) {
608 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
611 if (tp->t_rblocks_delta) {
612 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
615 if (tp->t_rextents_delta) {
616 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
619 if (tp->t_rextslog_delta) {
620 sbp->sb_rextslog += tp->t_rextslog_delta;
626 * Log the whole thing, the fields are noncontiguous.
628 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
631 * Since all the modifiable fields are contiguous, we
632 * can get away with this.
634 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
635 offsetof(xfs_dsb_t, sb_frextents) +
636 sizeof(sbp->sb_frextents) - 1);
640 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
641 * and apply superblock counter changes to the in-core superblock. The
642 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
643 * applied to the in-core superblock. The idea is that that has already been
646 * This is done efficiently with a single call to xfs_mod_incore_sb_batch().
647 * However, we have to ensure that we only modify each superblock field only
648 * once because the application of the delta values may not be atomic. That can
649 * lead to ENOSPC races occurring if we have two separate modifcations of the
650 * free space counter to put back the entire reservation and then take away
653 * If we are not logging superblock counters, then the inode allocated/free and
654 * used block counts are not updated in the on disk superblock. In this case,
655 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
656 * still need to update the incore superblock with the changes.
659 xfs_trans_unreserve_and_mod_sb(
662 xfs_mod_sb_t msb[14]; /* If you add cases, add entries */
664 xfs_mount_t *mp = tp->t_mountp;
668 int64_t blkdelta = 0;
669 int64_t rtxdelta = 0;
672 rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
674 /* calculate free blocks delta */
675 if (tp->t_blk_res > 0)
676 blkdelta = tp->t_blk_res;
678 if ((tp->t_fdblocks_delta != 0) &&
679 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
680 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
681 blkdelta += tp->t_fdblocks_delta;
684 msbp->msb_field = XFS_SBS_FDBLOCKS;
685 msbp->msb_delta = blkdelta;
689 /* calculate free realtime extents delta */
690 if (tp->t_rtx_res > 0)
691 rtxdelta = tp->t_rtx_res;
693 if ((tp->t_frextents_delta != 0) &&
694 (tp->t_flags & XFS_TRANS_SB_DIRTY))
695 rtxdelta += tp->t_frextents_delta;
698 msbp->msb_field = XFS_SBS_FREXTENTS;
699 msbp->msb_delta = rtxdelta;
703 /* apply remaining deltas */
705 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
706 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
707 if (tp->t_icount_delta != 0) {
708 msbp->msb_field = XFS_SBS_ICOUNT;
709 msbp->msb_delta = tp->t_icount_delta;
712 if (tp->t_ifree_delta != 0) {
713 msbp->msb_field = XFS_SBS_IFREE;
714 msbp->msb_delta = tp->t_ifree_delta;
719 if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
720 if (tp->t_dblocks_delta != 0) {
721 msbp->msb_field = XFS_SBS_DBLOCKS;
722 msbp->msb_delta = tp->t_dblocks_delta;
725 if (tp->t_agcount_delta != 0) {
726 msbp->msb_field = XFS_SBS_AGCOUNT;
727 msbp->msb_delta = tp->t_agcount_delta;
730 if (tp->t_imaxpct_delta != 0) {
731 msbp->msb_field = XFS_SBS_IMAX_PCT;
732 msbp->msb_delta = tp->t_imaxpct_delta;
735 if (tp->t_rextsize_delta != 0) {
736 msbp->msb_field = XFS_SBS_REXTSIZE;
737 msbp->msb_delta = tp->t_rextsize_delta;
740 if (tp->t_rbmblocks_delta != 0) {
741 msbp->msb_field = XFS_SBS_RBMBLOCKS;
742 msbp->msb_delta = tp->t_rbmblocks_delta;
745 if (tp->t_rblocks_delta != 0) {
746 msbp->msb_field = XFS_SBS_RBLOCKS;
747 msbp->msb_delta = tp->t_rblocks_delta;
750 if (tp->t_rextents_delta != 0) {
751 msbp->msb_field = XFS_SBS_REXTENTS;
752 msbp->msb_delta = tp->t_rextents_delta;
755 if (tp->t_rextslog_delta != 0) {
756 msbp->msb_field = XFS_SBS_REXTSLOG;
757 msbp->msb_delta = tp->t_rextslog_delta;
763 * If we need to change anything, do it.
766 error = xfs_mod_incore_sb_batch(tp->t_mountp, msb,
767 (uint)(msbp - msb), rsvd);
773 * Total up the number of log iovecs needed to commit this
774 * transaction. The transaction itself needs one for the
775 * transaction header. Ask each dirty item in turn how many
776 * it needs to get the total.
779 xfs_trans_count_vecs(
780 struct xfs_trans *tp)
783 xfs_log_item_desc_t *lidp;
786 lidp = xfs_trans_first_item(tp);
787 ASSERT(lidp != NULL);
789 /* In the non-debug case we need to start bailing out if we
790 * didn't find a log_item here, return zero and let trans_commit
796 while (lidp != NULL) {
798 * Skip items which aren't dirty in this transaction.
800 if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
801 lidp = xfs_trans_next_item(tp, lidp);
804 lidp->lid_size = IOP_SIZE(lidp->lid_item);
805 nvecs += lidp->lid_size;
806 lidp = xfs_trans_next_item(tp, lidp);
813 * Fill in the vector with pointers to data to be logged
814 * by this transaction. The transaction header takes
815 * the first vector, and then each dirty item takes the
816 * number of vectors it indicated it needed in xfs_trans_count_vecs().
818 * As each item fills in the entries it needs, also pin the item
819 * so that it cannot be flushed out until the log write completes.
823 struct xfs_trans *tp,
824 struct xfs_log_iovec *log_vector)
826 xfs_log_item_desc_t *lidp;
827 struct xfs_log_iovec *vecp;
831 * Skip over the entry for the transaction header, we'll
832 * fill that in at the end.
834 vecp = log_vector + 1;
837 lidp = xfs_trans_first_item(tp);
840 /* Skip items which aren't dirty in this transaction. */
841 if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
842 lidp = xfs_trans_next_item(tp, lidp);
847 * The item may be marked dirty but not log anything. This can
848 * be used to get called when a transaction is committed.
852 IOP_FORMAT(lidp->lid_item, vecp);
853 vecp += lidp->lid_size;
854 IOP_PIN(lidp->lid_item);
855 lidp = xfs_trans_next_item(tp, lidp);
859 * Now that we've counted the number of items in this transaction, fill
860 * in the transaction header. Note that the transaction header does not
863 tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
864 tp->t_header.th_type = tp->t_type;
865 tp->t_header.th_num_items = nitems;
866 log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
867 log_vector->i_len = sizeof(xfs_trans_header_t);
868 log_vector->i_type = XLOG_REG_TYPE_TRANSHDR;
872 * The committed item processing consists of calling the committed routine of
873 * each logged item, updating the item's position in the AIL if necessary, and
874 * unpinning each item. If the committed routine returns -1, then do nothing
875 * further with the item because it may have been freed.
877 * Since items are unlocked when they are copied to the incore log, it is
878 * possible for two transactions to be completing and manipulating the same
879 * item simultaneously. The AIL lock will protect the lsn field of each item.
880 * The value of this field can never go backwards.
882 * We unpin the items after repositioning them in the AIL, because otherwise
883 * they could be immediately flushed and we'd have to race with the flusher
884 * trying to pull the item from the AIL as we add it.
887 xfs_trans_item_committed(
888 struct xfs_log_item *lip,
889 xfs_lsn_t commit_lsn,
893 struct xfs_ail *ailp;
896 lip->li_flags |= XFS_LI_ABORTED;
897 item_lsn = IOP_COMMITTED(lip, commit_lsn);
899 /* If the committed routine returns -1, item has been freed. */
900 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
904 * If the returned lsn is greater than what it contained before, update
905 * the location of the item in the AIL. If it is not, then do nothing.
906 * Items can never move backwards in the AIL.
908 * While the new lsn should usually be greater, it is possible that a
909 * later transaction completing simultaneously with an earlier one
910 * using the same item could complete first with a higher lsn. This
911 * would cause the earlier transaction to fail the test below.
914 spin_lock(&ailp->xa_lock);
915 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0) {
917 * This will set the item's lsn to item_lsn and update the
918 * position of the item in the AIL.
920 * xfs_trans_ail_update() drops the AIL lock.
922 xfs_trans_ail_update(ailp, lip, item_lsn);
924 spin_unlock(&ailp->xa_lock);
928 * Now that we've repositioned the item in the AIL, unpin it so it can
929 * be flushed. Pass information about buffer stale state down from the
930 * log item flags, if anyone else stales the buffer we do not want to
931 * pay any attention to it.
937 * This is typically called by the LM when a transaction has been fully
938 * committed to disk. It needs to unpin the items which have
939 * been logged by the transaction and update their positions
940 * in the AIL if necessary.
942 * This also gets called when the transactions didn't get written out
943 * because of an I/O error. Abortflag & XFS_LI_ABORTED is set then.
947 struct xfs_trans *tp,
950 xfs_log_item_desc_t *lidp;
951 xfs_log_item_chunk_t *licp;
952 xfs_log_item_chunk_t *next_licp;
954 /* Call the transaction's completion callback if there is one. */
955 if (tp->t_callback != NULL)
956 tp->t_callback(tp, tp->t_callarg);
958 for (lidp = xfs_trans_first_item(tp);
960 lidp = xfs_trans_next_item(tp, lidp)) {
961 xfs_trans_item_committed(lidp->lid_item, tp->t_lsn, abortflag);
964 /* free the item chunks, ignoring the embedded chunk */
965 for (licp = tp->t_items.lic_next; licp != NULL; licp = next_licp) {
966 next_licp = licp->lic_next;
974 * Called from the trans_commit code when we notice that
975 * the filesystem is in the middle of a forced shutdown.
979 struct xfs_trans *tp,
982 xfs_log_item_desc_t *lidp;
984 for (lidp = xfs_trans_first_item(tp);
986 lidp = xfs_trans_next_item(tp, lidp)) {
988 * Unpin all but those that aren't dirty.
990 if (lidp->lid_flags & XFS_LID_DIRTY)
991 IOP_UNPIN_REMOVE(lidp->lid_item, tp);
994 xfs_trans_unreserve_and_mod_sb(tp);
995 xfs_trans_unreserve_and_mod_dquots(tp);
997 xfs_trans_free_items(tp, NULLCOMMITLSN, flags);
1002 * Format the transaction direct to the iclog. This isolates the physical
1003 * transaction commit operation from the logical operation and hence allows
1004 * other methods to be introduced without affecting the existing commit path.
1007 xfs_trans_commit_iclog(
1008 struct xfs_mount *mp,
1009 struct xfs_trans *tp,
1010 xfs_lsn_t *commit_lsn,
1016 struct xlog_in_core *commit_iclog;
1017 #define XFS_TRANS_LOGVEC_COUNT 16
1018 struct xfs_log_iovec log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
1019 struct xfs_log_iovec *log_vector;
1024 * Ask each log item how many log_vector entries it will
1025 * need so we can figure out how many to allocate.
1026 * Try to avoid the kmem_alloc() call in the common case
1027 * by using a vector from the stack when it fits.
1029 nvec = xfs_trans_count_vecs(tp);
1031 return ENOMEM; /* triggers a shutdown! */
1032 } else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
1033 log_vector = log_vector_fast;
1035 log_vector = (xfs_log_iovec_t *)kmem_alloc(nvec *
1036 sizeof(xfs_log_iovec_t),
1041 * Fill in the log_vector and pin the logged items, and
1042 * then write the transaction to the log.
1044 xfs_trans_fill_vecs(tp, log_vector);
1046 if (flags & XFS_TRANS_RELEASE_LOG_RES)
1047 log_flags = XFS_LOG_REL_PERM_RESERV;
1049 error = xfs_log_write(mp, log_vector, nvec, tp->t_ticket, &(tp->t_lsn));
1052 * The transaction is committed incore here, and can go out to disk
1053 * at any time after this call. However, all the items associated
1054 * with the transaction are still locked and pinned in memory.
1056 *commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
1058 tp->t_commit_lsn = *commit_lsn;
1059 trace_xfs_trans_commit_lsn(tp);
1061 if (nvec > XFS_TRANS_LOGVEC_COUNT)
1062 kmem_free(log_vector);
1065 * If we got a log write error. Unpin the logitems that we
1066 * had pinned, clean up, free trans structure, and return error.
1068 if (error || *commit_lsn == -1) {
1069 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1070 xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
1071 return XFS_ERROR(EIO);
1075 * Once the transaction has committed, unused
1076 * reservations need to be released and changes to
1077 * the superblock need to be reflected in the in-core
1078 * version. Do that now.
1080 xfs_trans_unreserve_and_mod_sb(tp);
1083 * Tell the LM to call the transaction completion routine
1084 * when the log write with LSN commit_lsn completes (e.g.
1085 * when the transaction commit really hits the on-disk log).
1086 * After this call we cannot reference tp, because the call
1087 * can happen at any time and the call will free the transaction
1088 * structure pointed to by tp. The only case where we call
1089 * the completion routine (xfs_trans_committed) directly is
1090 * if the log is turned off on a debug kernel or we're
1091 * running in simulation mode (the log is explicitly turned
1094 tp->t_logcb.cb_func = (void(*)(void*, int))xfs_trans_committed;
1095 tp->t_logcb.cb_arg = tp;
1098 * We need to pass the iclog buffer which was used for the
1099 * transaction commit record into this function, and attach
1100 * the callback to it. The callback must be attached before
1101 * the items are unlocked to avoid racing with other threads
1102 * waiting for an item to unlock.
1104 shutdown = xfs_log_notify(mp, commit_iclog, &(tp->t_logcb));
1107 * Mark this thread as no longer being in a transaction
1109 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1112 * Once all the items of the transaction have been copied
1113 * to the in core log and the callback is attached, the
1114 * items can be unlocked.
1116 * This will free descriptors pointing to items which were
1117 * not logged since there is nothing more to do with them.
1118 * For items which were logged, we will keep pointers to them
1119 * so they can be unpinned after the transaction commits to disk.
1120 * This will also stamp each modified meta-data item with
1121 * the commit lsn of this transaction for dependency tracking
1124 xfs_trans_unlock_items(tp, *commit_lsn);
1127 * If we detected a log error earlier, finish committing
1128 * the transaction now (unpin log items, etc).
1130 * Order is critical here, to avoid using the transaction
1131 * pointer after its been freed (by xfs_trans_committed
1132 * either here now, or as a callback). We cannot do this
1133 * step inside xfs_log_notify as was done earlier because
1137 xfs_trans_committed(tp, XFS_LI_ABORTED);
1140 * Now that the xfs_trans_committed callback has been attached,
1141 * and the items are released we can finally allow the iclog to
1144 return xfs_log_release_iclog(mp, commit_iclog);
1148 * Walk the log items and allocate log vector structures for
1149 * each item large enough to fit all the vectors they require.
1150 * Note that this format differs from the old log vector format in
1151 * that there is no transaction header in these log vectors.
1153 STATIC struct xfs_log_vec *
1154 xfs_trans_alloc_log_vecs(
1157 xfs_log_item_desc_t *lidp;
1158 struct xfs_log_vec *lv = NULL;
1159 struct xfs_log_vec *ret_lv = NULL;
1161 lidp = xfs_trans_first_item(tp);
1163 /* Bail out if we didn't find a log item. */
1169 while (lidp != NULL) {
1170 struct xfs_log_vec *new_lv;
1172 /* Skip items which aren't dirty in this transaction. */
1173 if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
1174 lidp = xfs_trans_next_item(tp, lidp);
1178 /* Skip items that do not have any vectors for writing */
1179 lidp->lid_size = IOP_SIZE(lidp->lid_item);
1180 if (!lidp->lid_size) {
1181 lidp = xfs_trans_next_item(tp, lidp);
1185 new_lv = kmem_zalloc(sizeof(*new_lv) +
1186 lidp->lid_size * sizeof(struct xfs_log_iovec),
1189 /* The allocated iovec region lies beyond the log vector. */
1190 new_lv->lv_iovecp = (struct xfs_log_iovec *)&new_lv[1];
1191 new_lv->lv_niovecs = lidp->lid_size;
1192 new_lv->lv_item = lidp->lid_item;
1196 lv->lv_next = new_lv;
1198 lidp = xfs_trans_next_item(tp, lidp);
1205 xfs_trans_commit_cil(
1206 struct xfs_mount *mp,
1207 struct xfs_trans *tp,
1208 xfs_lsn_t *commit_lsn,
1211 struct xfs_log_vec *log_vector;
1215 * Get each log item to allocate a vector structure for
1216 * the log item to to pass to the log write code. The
1217 * CIL commit code will format the vector and save it away.
1219 log_vector = xfs_trans_alloc_log_vecs(tp);
1223 error = xfs_log_commit_cil(mp, tp, log_vector, commit_lsn, flags);
1227 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1229 /* xfs_trans_free_items() unlocks them first */
1230 xfs_trans_free_items(tp, *commit_lsn, 0);
1238 * Commit the given transaction to the log a/synchronously.
1240 * XFS disk error handling mechanism is not based on a typical
1241 * transaction abort mechanism. Logically after the filesystem
1242 * gets marked 'SHUTDOWN', we can't let any new transactions
1243 * be durable - ie. committed to disk - because some metadata might
1244 * be inconsistent. In such cases, this returns an error, and the
1245 * caller may assume that all locked objects joined to the transaction
1246 * have already been unlocked as if the commit had succeeded.
1247 * Do not reference the transaction structure after this call.
1251 struct xfs_trans *tp,
1255 struct xfs_mount *mp = tp->t_mountp;
1256 xfs_lsn_t commit_lsn = -1;
1259 int sync = tp->t_flags & XFS_TRANS_SYNC;
1262 * Determine whether this commit is releasing a permanent
1263 * log reservation or not.
1265 if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1266 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1267 log_flags = XFS_LOG_REL_PERM_RESERV;
1271 * If there is nothing to be logged by the transaction,
1272 * then unlock all of the items associated with the
1273 * transaction and free the transaction structure.
1274 * Also make sure to return any reserved blocks to
1277 if (!(tp->t_flags & XFS_TRANS_DIRTY))
1280 if (XFS_FORCED_SHUTDOWN(mp)) {
1281 error = XFS_ERROR(EIO);
1285 ASSERT(tp->t_ticket != NULL);
1288 * If we need to update the superblock, then do it now.
1290 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
1291 xfs_trans_apply_sb_deltas(tp);
1292 xfs_trans_apply_dquot_deltas(tp);
1294 if (mp->m_flags & XFS_MOUNT_DELAYLOG)
1295 error = xfs_trans_commit_cil(mp, tp, &commit_lsn, flags);
1297 error = xfs_trans_commit_iclog(mp, tp, &commit_lsn, flags);
1299 if (error == ENOMEM) {
1300 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1301 error = XFS_ERROR(EIO);
1306 * If the transaction needs to be synchronous, then force the
1307 * log out now and wait for it.
1311 error = _xfs_log_force_lsn(mp, commit_lsn,
1312 XFS_LOG_SYNC, log_flushed);
1314 XFS_STATS_INC(xs_trans_sync);
1316 XFS_STATS_INC(xs_trans_async);
1322 xfs_trans_unreserve_and_mod_sb(tp);
1325 * It is indeed possible for the transaction to be not dirty but
1326 * the dqinfo portion to be. All that means is that we have some
1327 * (non-persistent) quota reservations that need to be unreserved.
1329 xfs_trans_unreserve_and_mod_dquots(tp);
1331 commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1332 if (commit_lsn == -1 && !error)
1333 error = XFS_ERROR(EIO);
1335 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1336 xfs_trans_free_items(tp, NULLCOMMITLSN, error ? XFS_TRANS_ABORT : 0);
1339 XFS_STATS_INC(xs_trans_empty);
1344 * Unlock all of the transaction's items and free the transaction.
1345 * The transaction must not have modified any of its items, because
1346 * there is no way to restore them to their previous state.
1348 * If the transaction has made a log reservation, make sure to release
1358 xfs_log_item_chunk_t *licp;
1359 xfs_log_item_desc_t *lidp;
1360 xfs_log_item_t *lip;
1363 xfs_mount_t *mp = tp->t_mountp;
1366 * See if the caller is being too lazy to figure out if
1367 * the transaction really needs an abort.
1369 if ((flags & XFS_TRANS_ABORT) && !(tp->t_flags & XFS_TRANS_DIRTY))
1370 flags &= ~XFS_TRANS_ABORT;
1372 * See if the caller is relying on us to shut down the
1373 * filesystem. This happens in paths where we detect
1374 * corruption and decide to give up.
1376 if ((tp->t_flags & XFS_TRANS_DIRTY) && !XFS_FORCED_SHUTDOWN(mp)) {
1377 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1378 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1381 if (!(flags & XFS_TRANS_ABORT)) {
1382 licp = &(tp->t_items);
1383 while (licp != NULL) {
1384 lidp = licp->lic_descs;
1385 for (i = 0; i < licp->lic_unused; i++, lidp++) {
1386 if (xfs_lic_isfree(licp, i)) {
1390 lip = lidp->lid_item;
1391 if (!XFS_FORCED_SHUTDOWN(mp))
1392 ASSERT(!(lip->li_type == XFS_LI_EFD));
1394 licp = licp->lic_next;
1398 xfs_trans_unreserve_and_mod_sb(tp);
1399 xfs_trans_unreserve_and_mod_dquots(tp);
1402 if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1403 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1404 log_flags = XFS_LOG_REL_PERM_RESERV;
1408 xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1411 /* mark this thread as no longer being in a transaction */
1412 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1414 xfs_trans_free_items(tp, NULLCOMMITLSN, flags);
1419 * Roll from one trans in the sequence of PERMANENT transactions to
1420 * the next: permanent transactions are only flushed out when
1421 * committed with XFS_TRANS_RELEASE_LOG_RES, but we still want as soon
1422 * as possible to let chunks of it go to the log. So we commit the
1423 * chunk we've been working on and get a new transaction to continue.
1427 struct xfs_trans **tpp,
1428 struct xfs_inode *dp)
1430 struct xfs_trans *trans;
1431 unsigned int logres, count;
1435 * Ensure that the inode is always logged.
1438 xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
1441 * Copy the critical parameters from one trans to the next.
1443 logres = trans->t_log_res;
1444 count = trans->t_log_count;
1445 *tpp = xfs_trans_dup(trans);
1448 * Commit the current transaction.
1449 * If this commit failed, then it'd just unlock those items that
1450 * are not marked ihold. That also means that a filesystem shutdown
1451 * is in progress. The caller takes the responsibility to cancel
1452 * the duplicate transaction that gets returned.
1454 error = xfs_trans_commit(trans, 0);
1461 * transaction commit worked ok so we can drop the extra ticket
1462 * reference that we gained in xfs_trans_dup()
1464 xfs_log_ticket_put(trans->t_ticket);
1468 * Reserve space in the log for th next transaction.
1469 * This also pushes items in the "AIL", the list of logged items,
1470 * out to disk if they are taking up space at the tail of the log
1471 * that we want to use. This requires that either nothing be locked
1472 * across this call, or that anything that is locked be logged in
1473 * the prior and the next transactions.
1475 error = xfs_trans_reserve(trans, 0, logres, 0,
1476 XFS_TRANS_PERM_LOG_RES, count);
1478 * Ensure that the inode is in the new transaction and locked.
1483 xfs_trans_ijoin(trans, dp, XFS_ILOCK_EXCL);
1484 xfs_trans_ihold(trans, dp);