Loading...
1/*
2 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3 * Copyright (C) 2010 Red Hat, Inc.
4 * All Rights Reserved.
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License as
8 * published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it would be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 */
19#include "xfs.h"
20#include "xfs_fs.h"
21#include "xfs_shared.h"
22#include "xfs_format.h"
23#include "xfs_log_format.h"
24#include "xfs_trans_resv.h"
25#include "xfs_mount.h"
26#include "xfs_inode.h"
27#include "xfs_extent_busy.h"
28#include "xfs_quota.h"
29#include "xfs_trans.h"
30#include "xfs_trans_priv.h"
31#include "xfs_log.h"
32#include "xfs_trace.h"
33#include "xfs_error.h"
34
35kmem_zone_t *xfs_trans_zone;
36kmem_zone_t *xfs_log_item_desc_zone;
37
38#if defined(CONFIG_TRACEPOINTS)
39static void
40xfs_trans_trace_reservations(
41 struct xfs_mount *mp)
42{
43 struct xfs_trans_res resv;
44 struct xfs_trans_res *res;
45 struct xfs_trans_res *end_res;
46 int i;
47
48 res = (struct xfs_trans_res *)M_RES(mp);
49 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
50 for (i = 0; res < end_res; i++, res++)
51 trace_xfs_trans_resv_calc(mp, i, res);
52 xfs_log_get_max_trans_res(mp, &resv);
53 trace_xfs_trans_resv_calc(mp, -1, &resv);
54}
55#else
56# define xfs_trans_trace_reservations(mp)
57#endif
58
59/*
60 * Initialize the precomputed transaction reservation values
61 * in the mount structure.
62 */
63void
64xfs_trans_init(
65 struct xfs_mount *mp)
66{
67 xfs_trans_resv_calc(mp, M_RES(mp));
68 xfs_trans_trace_reservations(mp);
69}
70
71/*
72 * Free the transaction structure. If there is more clean up
73 * to do when the structure is freed, add it here.
74 */
75STATIC void
76xfs_trans_free(
77 struct xfs_trans *tp)
78{
79 xfs_extent_busy_sort(&tp->t_busy);
80 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
81
82 atomic_dec(&tp->t_mountp->m_active_trans);
83 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
84 sb_end_intwrite(tp->t_mountp->m_super);
85 xfs_trans_free_dqinfo(tp);
86 kmem_zone_free(xfs_trans_zone, tp);
87}
88
89/*
90 * This is called to create a new transaction which will share the
91 * permanent log reservation of the given transaction. The remaining
92 * unused block and rt extent reservations are also inherited. This
93 * implies that the original transaction is no longer allowed to allocate
94 * blocks. Locks and log items, however, are no inherited. They must
95 * be added to the new transaction explicitly.
96 */
97STATIC xfs_trans_t *
98xfs_trans_dup(
99 xfs_trans_t *tp)
100{
101 xfs_trans_t *ntp;
102
103 ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
104
105 /*
106 * Initialize the new transaction structure.
107 */
108 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
109 ntp->t_mountp = tp->t_mountp;
110 INIT_LIST_HEAD(&ntp->t_items);
111 INIT_LIST_HEAD(&ntp->t_busy);
112
113 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
114 ASSERT(tp->t_ticket != NULL);
115
116 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
117 (tp->t_flags & XFS_TRANS_RESERVE) |
118 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT);
119 /* We gave our writer reference to the new transaction */
120 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
121 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
122
123 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
124 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
125 tp->t_blk_res = tp->t_blk_res_used;
126
127 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
128 tp->t_rtx_res = tp->t_rtx_res_used;
129 ntp->t_pflags = tp->t_pflags;
130
131 xfs_trans_dup_dqinfo(tp, ntp);
132
133 atomic_inc(&tp->t_mountp->m_active_trans);
134 return ntp;
135}
136
137/*
138 * This is called to reserve free disk blocks and log space for the
139 * given transaction. This must be done before allocating any resources
140 * within the transaction.
141 *
142 * This will return ENOSPC if there are not enough blocks available.
143 * It will sleep waiting for available log space.
144 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
145 * is used by long running transactions. If any one of the reservations
146 * fails then they will all be backed out.
147 *
148 * This does not do quota reservations. That typically is done by the
149 * caller afterwards.
150 */
151static int
152xfs_trans_reserve(
153 struct xfs_trans *tp,
154 struct xfs_trans_res *resp,
155 uint blocks,
156 uint rtextents)
157{
158 int error = 0;
159 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
160
161 /* Mark this thread as being in a transaction */
162 current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
163
164 /*
165 * Attempt to reserve the needed disk blocks by decrementing
166 * the number needed from the number available. This will
167 * fail if the count would go below zero.
168 */
169 if (blocks > 0) {
170 error = xfs_mod_fdblocks(tp->t_mountp, -((int64_t)blocks), rsvd);
171 if (error != 0) {
172 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
173 return -ENOSPC;
174 }
175 tp->t_blk_res += blocks;
176 }
177
178 /*
179 * Reserve the log space needed for this transaction.
180 */
181 if (resp->tr_logres > 0) {
182 bool permanent = false;
183
184 ASSERT(tp->t_log_res == 0 ||
185 tp->t_log_res == resp->tr_logres);
186 ASSERT(tp->t_log_count == 0 ||
187 tp->t_log_count == resp->tr_logcount);
188
189 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
190 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
191 permanent = true;
192 } else {
193 ASSERT(tp->t_ticket == NULL);
194 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
195 }
196
197 if (tp->t_ticket != NULL) {
198 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
199 error = xfs_log_regrant(tp->t_mountp, tp->t_ticket);
200 } else {
201 error = xfs_log_reserve(tp->t_mountp,
202 resp->tr_logres,
203 resp->tr_logcount,
204 &tp->t_ticket, XFS_TRANSACTION,
205 permanent);
206 }
207
208 if (error)
209 goto undo_blocks;
210
211 tp->t_log_res = resp->tr_logres;
212 tp->t_log_count = resp->tr_logcount;
213 }
214
215 /*
216 * Attempt to reserve the needed realtime extents by decrementing
217 * the number needed from the number available. This will
218 * fail if the count would go below zero.
219 */
220 if (rtextents > 0) {
221 error = xfs_mod_frextents(tp->t_mountp, -((int64_t)rtextents));
222 if (error) {
223 error = -ENOSPC;
224 goto undo_log;
225 }
226 tp->t_rtx_res += rtextents;
227 }
228
229 return 0;
230
231 /*
232 * Error cases jump to one of these labels to undo any
233 * reservations which have already been performed.
234 */
235undo_log:
236 if (resp->tr_logres > 0) {
237 xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, false);
238 tp->t_ticket = NULL;
239 tp->t_log_res = 0;
240 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
241 }
242
243undo_blocks:
244 if (blocks > 0) {
245 xfs_mod_fdblocks(tp->t_mountp, (int64_t)blocks, rsvd);
246 tp->t_blk_res = 0;
247 }
248
249 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
250
251 return error;
252}
253
254int
255xfs_trans_alloc(
256 struct xfs_mount *mp,
257 struct xfs_trans_res *resp,
258 uint blocks,
259 uint rtextents,
260 uint flags,
261 struct xfs_trans **tpp)
262{
263 struct xfs_trans *tp;
264 int error;
265
266 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
267 sb_start_intwrite(mp->m_super);
268
269 WARN_ON(mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
270 atomic_inc(&mp->m_active_trans);
271
272 tp = kmem_zone_zalloc(xfs_trans_zone,
273 (flags & XFS_TRANS_NOFS) ? KM_NOFS : KM_SLEEP);
274 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
275 tp->t_flags = flags;
276 tp->t_mountp = mp;
277 INIT_LIST_HEAD(&tp->t_items);
278 INIT_LIST_HEAD(&tp->t_busy);
279
280 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
281 if (error) {
282 xfs_trans_cancel(tp);
283 return error;
284 }
285
286 *tpp = tp;
287 return 0;
288}
289
290/*
291 * Create an empty transaction with no reservation. This is a defensive
292 * mechanism for routines that query metadata without actually modifying
293 * them -- if the metadata being queried is somehow cross-linked (think a
294 * btree block pointer that points higher in the tree), we risk deadlock.
295 * However, blocks grabbed as part of a transaction can be re-grabbed.
296 * The verifiers will notice the corrupt block and the operation will fail
297 * back to userspace without deadlocking.
298 *
299 * Note the zero-length reservation; this transaction MUST be cancelled
300 * without any dirty data.
301 */
302int
303xfs_trans_alloc_empty(
304 struct xfs_mount *mp,
305 struct xfs_trans **tpp)
306{
307 struct xfs_trans_res resv = {0};
308
309 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
310}
311
312/*
313 * Record the indicated change to the given field for application
314 * to the file system's superblock when the transaction commits.
315 * For now, just store the change in the transaction structure.
316 *
317 * Mark the transaction structure to indicate that the superblock
318 * needs to be updated before committing.
319 *
320 * Because we may not be keeping track of allocated/free inodes and
321 * used filesystem blocks in the superblock, we do not mark the
322 * superblock dirty in this transaction if we modify these fields.
323 * We still need to update the transaction deltas so that they get
324 * applied to the incore superblock, but we don't want them to
325 * cause the superblock to get locked and logged if these are the
326 * only fields in the superblock that the transaction modifies.
327 */
328void
329xfs_trans_mod_sb(
330 xfs_trans_t *tp,
331 uint field,
332 int64_t delta)
333{
334 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
335 xfs_mount_t *mp = tp->t_mountp;
336
337 switch (field) {
338 case XFS_TRANS_SB_ICOUNT:
339 tp->t_icount_delta += delta;
340 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
341 flags &= ~XFS_TRANS_SB_DIRTY;
342 break;
343 case XFS_TRANS_SB_IFREE:
344 tp->t_ifree_delta += delta;
345 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
346 flags &= ~XFS_TRANS_SB_DIRTY;
347 break;
348 case XFS_TRANS_SB_FDBLOCKS:
349 /*
350 * Track the number of blocks allocated in the transaction.
351 * Make sure it does not exceed the number reserved. If so,
352 * shutdown as this can lead to accounting inconsistency.
353 */
354 if (delta < 0) {
355 tp->t_blk_res_used += (uint)-delta;
356 if (tp->t_blk_res_used > tp->t_blk_res)
357 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
358 }
359 tp->t_fdblocks_delta += delta;
360 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
361 flags &= ~XFS_TRANS_SB_DIRTY;
362 break;
363 case XFS_TRANS_SB_RES_FDBLOCKS:
364 /*
365 * The allocation has already been applied to the
366 * in-core superblock's counter. This should only
367 * be applied to the on-disk superblock.
368 */
369 tp->t_res_fdblocks_delta += delta;
370 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
371 flags &= ~XFS_TRANS_SB_DIRTY;
372 break;
373 case XFS_TRANS_SB_FREXTENTS:
374 /*
375 * Track the number of blocks allocated in the
376 * transaction. Make sure it does not exceed the
377 * number reserved.
378 */
379 if (delta < 0) {
380 tp->t_rtx_res_used += (uint)-delta;
381 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
382 }
383 tp->t_frextents_delta += delta;
384 break;
385 case XFS_TRANS_SB_RES_FREXTENTS:
386 /*
387 * The allocation has already been applied to the
388 * in-core superblock's counter. This should only
389 * be applied to the on-disk superblock.
390 */
391 ASSERT(delta < 0);
392 tp->t_res_frextents_delta += delta;
393 break;
394 case XFS_TRANS_SB_DBLOCKS:
395 ASSERT(delta > 0);
396 tp->t_dblocks_delta += delta;
397 break;
398 case XFS_TRANS_SB_AGCOUNT:
399 ASSERT(delta > 0);
400 tp->t_agcount_delta += delta;
401 break;
402 case XFS_TRANS_SB_IMAXPCT:
403 tp->t_imaxpct_delta += delta;
404 break;
405 case XFS_TRANS_SB_REXTSIZE:
406 tp->t_rextsize_delta += delta;
407 break;
408 case XFS_TRANS_SB_RBMBLOCKS:
409 tp->t_rbmblocks_delta += delta;
410 break;
411 case XFS_TRANS_SB_RBLOCKS:
412 tp->t_rblocks_delta += delta;
413 break;
414 case XFS_TRANS_SB_REXTENTS:
415 tp->t_rextents_delta += delta;
416 break;
417 case XFS_TRANS_SB_REXTSLOG:
418 tp->t_rextslog_delta += delta;
419 break;
420 default:
421 ASSERT(0);
422 return;
423 }
424
425 tp->t_flags |= flags;
426}
427
428/*
429 * xfs_trans_apply_sb_deltas() is called from the commit code
430 * to bring the superblock buffer into the current transaction
431 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
432 *
433 * For now we just look at each field allowed to change and change
434 * it if necessary.
435 */
436STATIC void
437xfs_trans_apply_sb_deltas(
438 xfs_trans_t *tp)
439{
440 xfs_dsb_t *sbp;
441 xfs_buf_t *bp;
442 int whole = 0;
443
444 bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
445 sbp = XFS_BUF_TO_SBP(bp);
446
447 /*
448 * Check that superblock mods match the mods made to AGF counters.
449 */
450 ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
451 (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
452 tp->t_ag_btree_delta));
453
454 /*
455 * Only update the superblock counters if we are logging them
456 */
457 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
458 if (tp->t_icount_delta)
459 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
460 if (tp->t_ifree_delta)
461 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
462 if (tp->t_fdblocks_delta)
463 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
464 if (tp->t_res_fdblocks_delta)
465 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
466 }
467
468 if (tp->t_frextents_delta)
469 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
470 if (tp->t_res_frextents_delta)
471 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
472
473 if (tp->t_dblocks_delta) {
474 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
475 whole = 1;
476 }
477 if (tp->t_agcount_delta) {
478 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
479 whole = 1;
480 }
481 if (tp->t_imaxpct_delta) {
482 sbp->sb_imax_pct += tp->t_imaxpct_delta;
483 whole = 1;
484 }
485 if (tp->t_rextsize_delta) {
486 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
487 whole = 1;
488 }
489 if (tp->t_rbmblocks_delta) {
490 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
491 whole = 1;
492 }
493 if (tp->t_rblocks_delta) {
494 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
495 whole = 1;
496 }
497 if (tp->t_rextents_delta) {
498 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
499 whole = 1;
500 }
501 if (tp->t_rextslog_delta) {
502 sbp->sb_rextslog += tp->t_rextslog_delta;
503 whole = 1;
504 }
505
506 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
507 if (whole)
508 /*
509 * Log the whole thing, the fields are noncontiguous.
510 */
511 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
512 else
513 /*
514 * Since all the modifiable fields are contiguous, we
515 * can get away with this.
516 */
517 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
518 offsetof(xfs_dsb_t, sb_frextents) +
519 sizeof(sbp->sb_frextents) - 1);
520}
521
522STATIC int
523xfs_sb_mod8(
524 uint8_t *field,
525 int8_t delta)
526{
527 int8_t counter = *field;
528
529 counter += delta;
530 if (counter < 0) {
531 ASSERT(0);
532 return -EINVAL;
533 }
534 *field = counter;
535 return 0;
536}
537
538STATIC int
539xfs_sb_mod32(
540 uint32_t *field,
541 int32_t delta)
542{
543 int32_t counter = *field;
544
545 counter += delta;
546 if (counter < 0) {
547 ASSERT(0);
548 return -EINVAL;
549 }
550 *field = counter;
551 return 0;
552}
553
554STATIC int
555xfs_sb_mod64(
556 uint64_t *field,
557 int64_t delta)
558{
559 int64_t counter = *field;
560
561 counter += delta;
562 if (counter < 0) {
563 ASSERT(0);
564 return -EINVAL;
565 }
566 *field = counter;
567 return 0;
568}
569
570/*
571 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
572 * and apply superblock counter changes to the in-core superblock. The
573 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
574 * applied to the in-core superblock. The idea is that that has already been
575 * done.
576 *
577 * If we are not logging superblock counters, then the inode allocated/free and
578 * used block counts are not updated in the on disk superblock. In this case,
579 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
580 * still need to update the incore superblock with the changes.
581 */
582void
583xfs_trans_unreserve_and_mod_sb(
584 struct xfs_trans *tp)
585{
586 struct xfs_mount *mp = tp->t_mountp;
587 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
588 int64_t blkdelta = 0;
589 int64_t rtxdelta = 0;
590 int64_t idelta = 0;
591 int64_t ifreedelta = 0;
592 int error;
593
594 /* calculate deltas */
595 if (tp->t_blk_res > 0)
596 blkdelta = tp->t_blk_res;
597 if ((tp->t_fdblocks_delta != 0) &&
598 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
599 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
600 blkdelta += tp->t_fdblocks_delta;
601
602 if (tp->t_rtx_res > 0)
603 rtxdelta = tp->t_rtx_res;
604 if ((tp->t_frextents_delta != 0) &&
605 (tp->t_flags & XFS_TRANS_SB_DIRTY))
606 rtxdelta += tp->t_frextents_delta;
607
608 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
609 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
610 idelta = tp->t_icount_delta;
611 ifreedelta = tp->t_ifree_delta;
612 }
613
614 /* apply the per-cpu counters */
615 if (blkdelta) {
616 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
617 if (error)
618 goto out;
619 }
620
621 if (idelta) {
622 error = xfs_mod_icount(mp, idelta);
623 if (error)
624 goto out_undo_fdblocks;
625 }
626
627 if (ifreedelta) {
628 error = xfs_mod_ifree(mp, ifreedelta);
629 if (error)
630 goto out_undo_icount;
631 }
632
633 if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
634 return;
635
636 /* apply remaining deltas */
637 spin_lock(&mp->m_sb_lock);
638 if (rtxdelta) {
639 error = xfs_sb_mod64(&mp->m_sb.sb_frextents, rtxdelta);
640 if (error)
641 goto out_undo_ifree;
642 }
643
644 if (tp->t_dblocks_delta != 0) {
645 error = xfs_sb_mod64(&mp->m_sb.sb_dblocks, tp->t_dblocks_delta);
646 if (error)
647 goto out_undo_frextents;
648 }
649 if (tp->t_agcount_delta != 0) {
650 error = xfs_sb_mod32(&mp->m_sb.sb_agcount, tp->t_agcount_delta);
651 if (error)
652 goto out_undo_dblocks;
653 }
654 if (tp->t_imaxpct_delta != 0) {
655 error = xfs_sb_mod8(&mp->m_sb.sb_imax_pct, tp->t_imaxpct_delta);
656 if (error)
657 goto out_undo_agcount;
658 }
659 if (tp->t_rextsize_delta != 0) {
660 error = xfs_sb_mod32(&mp->m_sb.sb_rextsize,
661 tp->t_rextsize_delta);
662 if (error)
663 goto out_undo_imaxpct;
664 }
665 if (tp->t_rbmblocks_delta != 0) {
666 error = xfs_sb_mod32(&mp->m_sb.sb_rbmblocks,
667 tp->t_rbmblocks_delta);
668 if (error)
669 goto out_undo_rextsize;
670 }
671 if (tp->t_rblocks_delta != 0) {
672 error = xfs_sb_mod64(&mp->m_sb.sb_rblocks, tp->t_rblocks_delta);
673 if (error)
674 goto out_undo_rbmblocks;
675 }
676 if (tp->t_rextents_delta != 0) {
677 error = xfs_sb_mod64(&mp->m_sb.sb_rextents,
678 tp->t_rextents_delta);
679 if (error)
680 goto out_undo_rblocks;
681 }
682 if (tp->t_rextslog_delta != 0) {
683 error = xfs_sb_mod8(&mp->m_sb.sb_rextslog,
684 tp->t_rextslog_delta);
685 if (error)
686 goto out_undo_rextents;
687 }
688 spin_unlock(&mp->m_sb_lock);
689 return;
690
691out_undo_rextents:
692 if (tp->t_rextents_delta)
693 xfs_sb_mod64(&mp->m_sb.sb_rextents, -tp->t_rextents_delta);
694out_undo_rblocks:
695 if (tp->t_rblocks_delta)
696 xfs_sb_mod64(&mp->m_sb.sb_rblocks, -tp->t_rblocks_delta);
697out_undo_rbmblocks:
698 if (tp->t_rbmblocks_delta)
699 xfs_sb_mod32(&mp->m_sb.sb_rbmblocks, -tp->t_rbmblocks_delta);
700out_undo_rextsize:
701 if (tp->t_rextsize_delta)
702 xfs_sb_mod32(&mp->m_sb.sb_rextsize, -tp->t_rextsize_delta);
703out_undo_imaxpct:
704 if (tp->t_rextsize_delta)
705 xfs_sb_mod8(&mp->m_sb.sb_imax_pct, -tp->t_imaxpct_delta);
706out_undo_agcount:
707 if (tp->t_agcount_delta)
708 xfs_sb_mod32(&mp->m_sb.sb_agcount, -tp->t_agcount_delta);
709out_undo_dblocks:
710 if (tp->t_dblocks_delta)
711 xfs_sb_mod64(&mp->m_sb.sb_dblocks, -tp->t_dblocks_delta);
712out_undo_frextents:
713 if (rtxdelta)
714 xfs_sb_mod64(&mp->m_sb.sb_frextents, -rtxdelta);
715out_undo_ifree:
716 spin_unlock(&mp->m_sb_lock);
717 if (ifreedelta)
718 xfs_mod_ifree(mp, -ifreedelta);
719out_undo_icount:
720 if (idelta)
721 xfs_mod_icount(mp, -idelta);
722out_undo_fdblocks:
723 if (blkdelta)
724 xfs_mod_fdblocks(mp, -blkdelta, rsvd);
725out:
726 ASSERT(error == 0);
727 return;
728}
729
730/*
731 * Add the given log item to the transaction's list of log items.
732 *
733 * The log item will now point to its new descriptor with its li_desc field.
734 */
735void
736xfs_trans_add_item(
737 struct xfs_trans *tp,
738 struct xfs_log_item *lip)
739{
740 struct xfs_log_item_desc *lidp;
741
742 ASSERT(lip->li_mountp == tp->t_mountp);
743 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
744
745 lidp = kmem_zone_zalloc(xfs_log_item_desc_zone, KM_SLEEP | KM_NOFS);
746
747 lidp->lid_item = lip;
748 lidp->lid_flags = 0;
749 list_add_tail(&lidp->lid_trans, &tp->t_items);
750
751 lip->li_desc = lidp;
752}
753
754STATIC void
755xfs_trans_free_item_desc(
756 struct xfs_log_item_desc *lidp)
757{
758 list_del_init(&lidp->lid_trans);
759 kmem_zone_free(xfs_log_item_desc_zone, lidp);
760}
761
762/*
763 * Unlink and free the given descriptor.
764 */
765void
766xfs_trans_del_item(
767 struct xfs_log_item *lip)
768{
769 xfs_trans_free_item_desc(lip->li_desc);
770 lip->li_desc = NULL;
771}
772
773/*
774 * Unlock all of the items of a transaction and free all the descriptors
775 * of that transaction.
776 */
777void
778xfs_trans_free_items(
779 struct xfs_trans *tp,
780 xfs_lsn_t commit_lsn,
781 bool abort)
782{
783 struct xfs_log_item_desc *lidp, *next;
784
785 list_for_each_entry_safe(lidp, next, &tp->t_items, lid_trans) {
786 struct xfs_log_item *lip = lidp->lid_item;
787
788 lip->li_desc = NULL;
789
790 if (commit_lsn != NULLCOMMITLSN)
791 lip->li_ops->iop_committing(lip, commit_lsn);
792 if (abort)
793 lip->li_flags |= XFS_LI_ABORTED;
794 lip->li_ops->iop_unlock(lip);
795
796 xfs_trans_free_item_desc(lidp);
797 }
798}
799
800static inline void
801xfs_log_item_batch_insert(
802 struct xfs_ail *ailp,
803 struct xfs_ail_cursor *cur,
804 struct xfs_log_item **log_items,
805 int nr_items,
806 xfs_lsn_t commit_lsn)
807{
808 int i;
809
810 spin_lock(&ailp->ail_lock);
811 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
812 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
813
814 for (i = 0; i < nr_items; i++) {
815 struct xfs_log_item *lip = log_items[i];
816
817 lip->li_ops->iop_unpin(lip, 0);
818 }
819}
820
821/*
822 * Bulk operation version of xfs_trans_committed that takes a log vector of
823 * items to insert into the AIL. This uses bulk AIL insertion techniques to
824 * minimise lock traffic.
825 *
826 * If we are called with the aborted flag set, it is because a log write during
827 * a CIL checkpoint commit has failed. In this case, all the items in the
828 * checkpoint have already gone through iop_commited and iop_unlock, which
829 * means that checkpoint commit abort handling is treated exactly the same
830 * as an iclog write error even though we haven't started any IO yet. Hence in
831 * this case all we need to do is iop_committed processing, followed by an
832 * iop_unpin(aborted) call.
833 *
834 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
835 * at the end of the AIL, the insert cursor avoids the need to walk
836 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
837 * call. This saves a lot of needless list walking and is a net win, even
838 * though it slightly increases that amount of AIL lock traffic to set it up
839 * and tear it down.
840 */
841void
842xfs_trans_committed_bulk(
843 struct xfs_ail *ailp,
844 struct xfs_log_vec *log_vector,
845 xfs_lsn_t commit_lsn,
846 int aborted)
847{
848#define LOG_ITEM_BATCH_SIZE 32
849 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
850 struct xfs_log_vec *lv;
851 struct xfs_ail_cursor cur;
852 int i = 0;
853
854 spin_lock(&ailp->ail_lock);
855 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
856 spin_unlock(&ailp->ail_lock);
857
858 /* unpin all the log items */
859 for (lv = log_vector; lv; lv = lv->lv_next ) {
860 struct xfs_log_item *lip = lv->lv_item;
861 xfs_lsn_t item_lsn;
862
863 if (aborted)
864 lip->li_flags |= XFS_LI_ABORTED;
865 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
866
867 /* item_lsn of -1 means the item needs no further processing */
868 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
869 continue;
870
871 /*
872 * if we are aborting the operation, no point in inserting the
873 * object into the AIL as we are in a shutdown situation.
874 */
875 if (aborted) {
876 ASSERT(XFS_FORCED_SHUTDOWN(ailp->ail_mount));
877 lip->li_ops->iop_unpin(lip, 1);
878 continue;
879 }
880
881 if (item_lsn != commit_lsn) {
882
883 /*
884 * Not a bulk update option due to unusual item_lsn.
885 * Push into AIL immediately, rechecking the lsn once
886 * we have the ail lock. Then unpin the item. This does
887 * not affect the AIL cursor the bulk insert path is
888 * using.
889 */
890 spin_lock(&ailp->ail_lock);
891 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
892 xfs_trans_ail_update(ailp, lip, item_lsn);
893 else
894 spin_unlock(&ailp->ail_lock);
895 lip->li_ops->iop_unpin(lip, 0);
896 continue;
897 }
898
899 /* Item is a candidate for bulk AIL insert. */
900 log_items[i++] = lv->lv_item;
901 if (i >= LOG_ITEM_BATCH_SIZE) {
902 xfs_log_item_batch_insert(ailp, &cur, log_items,
903 LOG_ITEM_BATCH_SIZE, commit_lsn);
904 i = 0;
905 }
906 }
907
908 /* make sure we insert the remainder! */
909 if (i)
910 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
911
912 spin_lock(&ailp->ail_lock);
913 xfs_trans_ail_cursor_done(&cur);
914 spin_unlock(&ailp->ail_lock);
915}
916
917/*
918 * Commit the given transaction to the log.
919 *
920 * XFS disk error handling mechanism is not based on a typical
921 * transaction abort mechanism. Logically after the filesystem
922 * gets marked 'SHUTDOWN', we can't let any new transactions
923 * be durable - ie. committed to disk - because some metadata might
924 * be inconsistent. In such cases, this returns an error, and the
925 * caller may assume that all locked objects joined to the transaction
926 * have already been unlocked as if the commit had succeeded.
927 * Do not reference the transaction structure after this call.
928 */
929static int
930__xfs_trans_commit(
931 struct xfs_trans *tp,
932 bool regrant)
933{
934 struct xfs_mount *mp = tp->t_mountp;
935 xfs_lsn_t commit_lsn = -1;
936 int error = 0;
937 int sync = tp->t_flags & XFS_TRANS_SYNC;
938
939 /*
940 * If there is nothing to be logged by the transaction,
941 * then unlock all of the items associated with the
942 * transaction and free the transaction structure.
943 * Also make sure to return any reserved blocks to
944 * the free pool.
945 */
946 if (!(tp->t_flags & XFS_TRANS_DIRTY))
947 goto out_unreserve;
948
949 if (XFS_FORCED_SHUTDOWN(mp)) {
950 error = -EIO;
951 goto out_unreserve;
952 }
953
954 ASSERT(tp->t_ticket != NULL);
955
956 /*
957 * If we need to update the superblock, then do it now.
958 */
959 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
960 xfs_trans_apply_sb_deltas(tp);
961 xfs_trans_apply_dquot_deltas(tp);
962
963 xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
964
965 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
966 xfs_trans_free(tp);
967
968 /*
969 * If the transaction needs to be synchronous, then force the
970 * log out now and wait for it.
971 */
972 if (sync) {
973 error = xfs_log_force_lsn(mp, commit_lsn, XFS_LOG_SYNC, NULL);
974 XFS_STATS_INC(mp, xs_trans_sync);
975 } else {
976 XFS_STATS_INC(mp, xs_trans_async);
977 }
978
979 return error;
980
981out_unreserve:
982 xfs_trans_unreserve_and_mod_sb(tp);
983
984 /*
985 * It is indeed possible for the transaction to be not dirty but
986 * the dqinfo portion to be. All that means is that we have some
987 * (non-persistent) quota reservations that need to be unreserved.
988 */
989 xfs_trans_unreserve_and_mod_dquots(tp);
990 if (tp->t_ticket) {
991 commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, regrant);
992 if (commit_lsn == -1 && !error)
993 error = -EIO;
994 }
995 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
996 xfs_trans_free_items(tp, NULLCOMMITLSN, !!error);
997 xfs_trans_free(tp);
998
999 XFS_STATS_INC(mp, xs_trans_empty);
1000 return error;
1001}
1002
1003int
1004xfs_trans_commit(
1005 struct xfs_trans *tp)
1006{
1007 return __xfs_trans_commit(tp, false);
1008}
1009
1010/*
1011 * Unlock all of the transaction's items and free the transaction.
1012 * The transaction must not have modified any of its items, because
1013 * there is no way to restore them to their previous state.
1014 *
1015 * If the transaction has made a log reservation, make sure to release
1016 * it as well.
1017 */
1018void
1019xfs_trans_cancel(
1020 struct xfs_trans *tp)
1021{
1022 struct xfs_mount *mp = tp->t_mountp;
1023 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
1024
1025 /*
1026 * See if the caller is relying on us to shut down the
1027 * filesystem. This happens in paths where we detect
1028 * corruption and decide to give up.
1029 */
1030 if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1031 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1032 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1033 }
1034#ifdef DEBUG
1035 if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1036 struct xfs_log_item_desc *lidp;
1037
1038 list_for_each_entry(lidp, &tp->t_items, lid_trans)
1039 ASSERT(!(lidp->lid_item->li_type == XFS_LI_EFD));
1040 }
1041#endif
1042 xfs_trans_unreserve_and_mod_sb(tp);
1043 xfs_trans_unreserve_and_mod_dquots(tp);
1044
1045 if (tp->t_ticket)
1046 xfs_log_done(mp, tp->t_ticket, NULL, false);
1047
1048 /* mark this thread as no longer being in a transaction */
1049 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
1050
1051 xfs_trans_free_items(tp, NULLCOMMITLSN, dirty);
1052 xfs_trans_free(tp);
1053}
1054
1055/*
1056 * Roll from one trans in the sequence of PERMANENT transactions to
1057 * the next: permanent transactions are only flushed out when
1058 * committed with xfs_trans_commit(), but we still want as soon
1059 * as possible to let chunks of it go to the log. So we commit the
1060 * chunk we've been working on and get a new transaction to continue.
1061 */
1062int
1063xfs_trans_roll(
1064 struct xfs_trans **tpp)
1065{
1066 struct xfs_trans *trans = *tpp;
1067 struct xfs_trans_res tres;
1068 int error;
1069
1070 /*
1071 * Copy the critical parameters from one trans to the next.
1072 */
1073 tres.tr_logres = trans->t_log_res;
1074 tres.tr_logcount = trans->t_log_count;
1075
1076 *tpp = xfs_trans_dup(trans);
1077
1078 /*
1079 * Commit the current transaction.
1080 * If this commit failed, then it'd just unlock those items that
1081 * are not marked ihold. That also means that a filesystem shutdown
1082 * is in progress. The caller takes the responsibility to cancel
1083 * the duplicate transaction that gets returned.
1084 */
1085 error = __xfs_trans_commit(trans, true);
1086 if (error)
1087 return error;
1088
1089 /*
1090 * Reserve space in the log for the next transaction.
1091 * This also pushes items in the "AIL", the list of logged items,
1092 * out to disk if they are taking up space at the tail of the log
1093 * that we want to use. This requires that either nothing be locked
1094 * across this call, or that anything that is locked be logged in
1095 * the prior and the next transactions.
1096 */
1097 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1098 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1099}
1/*
2 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3 * Copyright (C) 2010 Red Hat, Inc.
4 * All Rights Reserved.
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License as
8 * published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it would be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 */
19#include "xfs.h"
20#include "xfs_fs.h"
21#include "xfs_shared.h"
22#include "xfs_format.h"
23#include "xfs_log_format.h"
24#include "xfs_trans_resv.h"
25#include "xfs_mount.h"
26#include "xfs_inode.h"
27#include "xfs_extent_busy.h"
28#include "xfs_quota.h"
29#include "xfs_trans.h"
30#include "xfs_trans_priv.h"
31#include "xfs_log.h"
32#include "xfs_trace.h"
33#include "xfs_error.h"
34
35kmem_zone_t *xfs_trans_zone;
36kmem_zone_t *xfs_log_item_desc_zone;
37
38/*
39 * Initialize the precomputed transaction reservation values
40 * in the mount structure.
41 */
42void
43xfs_trans_init(
44 struct xfs_mount *mp)
45{
46 xfs_trans_resv_calc(mp, M_RES(mp));
47}
48
49/*
50 * Free the transaction structure. If there is more clean up
51 * to do when the structure is freed, add it here.
52 */
53STATIC void
54xfs_trans_free(
55 struct xfs_trans *tp)
56{
57 xfs_extent_busy_sort(&tp->t_busy);
58 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
59
60 atomic_dec(&tp->t_mountp->m_active_trans);
61 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
62 sb_end_intwrite(tp->t_mountp->m_super);
63 xfs_trans_free_dqinfo(tp);
64 kmem_zone_free(xfs_trans_zone, tp);
65}
66
67/*
68 * This is called to create a new transaction which will share the
69 * permanent log reservation of the given transaction. The remaining
70 * unused block and rt extent reservations are also inherited. This
71 * implies that the original transaction is no longer allowed to allocate
72 * blocks. Locks and log items, however, are no inherited. They must
73 * be added to the new transaction explicitly.
74 */
75STATIC xfs_trans_t *
76xfs_trans_dup(
77 xfs_trans_t *tp)
78{
79 xfs_trans_t *ntp;
80
81 ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
82
83 /*
84 * Initialize the new transaction structure.
85 */
86 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
87 ntp->t_mountp = tp->t_mountp;
88 INIT_LIST_HEAD(&ntp->t_items);
89 INIT_LIST_HEAD(&ntp->t_busy);
90
91 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
92 ASSERT(tp->t_ticket != NULL);
93
94 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
95 (tp->t_flags & XFS_TRANS_RESERVE) |
96 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT);
97 /* We gave our writer reference to the new transaction */
98 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
99 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
100 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
101 tp->t_blk_res = tp->t_blk_res_used;
102 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
103 tp->t_rtx_res = tp->t_rtx_res_used;
104 ntp->t_pflags = tp->t_pflags;
105
106 xfs_trans_dup_dqinfo(tp, ntp);
107
108 atomic_inc(&tp->t_mountp->m_active_trans);
109 return ntp;
110}
111
112/*
113 * This is called to reserve free disk blocks and log space for the
114 * given transaction. This must be done before allocating any resources
115 * within the transaction.
116 *
117 * This will return ENOSPC if there are not enough blocks available.
118 * It will sleep waiting for available log space.
119 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
120 * is used by long running transactions. If any one of the reservations
121 * fails then they will all be backed out.
122 *
123 * This does not do quota reservations. That typically is done by the
124 * caller afterwards.
125 */
126static int
127xfs_trans_reserve(
128 struct xfs_trans *tp,
129 struct xfs_trans_res *resp,
130 uint blocks,
131 uint rtextents)
132{
133 int error = 0;
134 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
135
136 /* Mark this thread as being in a transaction */
137 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
138
139 /*
140 * Attempt to reserve the needed disk blocks by decrementing
141 * the number needed from the number available. This will
142 * fail if the count would go below zero.
143 */
144 if (blocks > 0) {
145 error = xfs_mod_fdblocks(tp->t_mountp, -((int64_t)blocks), rsvd);
146 if (error != 0) {
147 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
148 return -ENOSPC;
149 }
150 tp->t_blk_res += blocks;
151 }
152
153 /*
154 * Reserve the log space needed for this transaction.
155 */
156 if (resp->tr_logres > 0) {
157 bool permanent = false;
158
159 ASSERT(tp->t_log_res == 0 ||
160 tp->t_log_res == resp->tr_logres);
161 ASSERT(tp->t_log_count == 0 ||
162 tp->t_log_count == resp->tr_logcount);
163
164 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
165 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
166 permanent = true;
167 } else {
168 ASSERT(tp->t_ticket == NULL);
169 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
170 }
171
172 if (tp->t_ticket != NULL) {
173 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
174 error = xfs_log_regrant(tp->t_mountp, tp->t_ticket);
175 } else {
176 error = xfs_log_reserve(tp->t_mountp,
177 resp->tr_logres,
178 resp->tr_logcount,
179 &tp->t_ticket, XFS_TRANSACTION,
180 permanent);
181 }
182
183 if (error)
184 goto undo_blocks;
185
186 tp->t_log_res = resp->tr_logres;
187 tp->t_log_count = resp->tr_logcount;
188 }
189
190 /*
191 * Attempt to reserve the needed realtime extents by decrementing
192 * the number needed from the number available. This will
193 * fail if the count would go below zero.
194 */
195 if (rtextents > 0) {
196 error = xfs_mod_frextents(tp->t_mountp, -((int64_t)rtextents));
197 if (error) {
198 error = -ENOSPC;
199 goto undo_log;
200 }
201 tp->t_rtx_res += rtextents;
202 }
203
204 return 0;
205
206 /*
207 * Error cases jump to one of these labels to undo any
208 * reservations which have already been performed.
209 */
210undo_log:
211 if (resp->tr_logres > 0) {
212 xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, false);
213 tp->t_ticket = NULL;
214 tp->t_log_res = 0;
215 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
216 }
217
218undo_blocks:
219 if (blocks > 0) {
220 xfs_mod_fdblocks(tp->t_mountp, (int64_t)blocks, rsvd);
221 tp->t_blk_res = 0;
222 }
223
224 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
225
226 return error;
227}
228
229int
230xfs_trans_alloc(
231 struct xfs_mount *mp,
232 struct xfs_trans_res *resp,
233 uint blocks,
234 uint rtextents,
235 uint flags,
236 struct xfs_trans **tpp)
237{
238 struct xfs_trans *tp;
239 int error;
240
241 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
242 sb_start_intwrite(mp->m_super);
243
244 WARN_ON(mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
245 atomic_inc(&mp->m_active_trans);
246
247 tp = kmem_zone_zalloc(xfs_trans_zone,
248 (flags & XFS_TRANS_NOFS) ? KM_NOFS : KM_SLEEP);
249 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
250 tp->t_flags = flags;
251 tp->t_mountp = mp;
252 INIT_LIST_HEAD(&tp->t_items);
253 INIT_LIST_HEAD(&tp->t_busy);
254
255 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
256 if (error) {
257 xfs_trans_cancel(tp);
258 return error;
259 }
260
261 *tpp = tp;
262 return 0;
263}
264
265/*
266 * Record the indicated change to the given field for application
267 * to the file system's superblock when the transaction commits.
268 * For now, just store the change in the transaction structure.
269 *
270 * Mark the transaction structure to indicate that the superblock
271 * needs to be updated before committing.
272 *
273 * Because we may not be keeping track of allocated/free inodes and
274 * used filesystem blocks in the superblock, we do not mark the
275 * superblock dirty in this transaction if we modify these fields.
276 * We still need to update the transaction deltas so that they get
277 * applied to the incore superblock, but we don't want them to
278 * cause the superblock to get locked and logged if these are the
279 * only fields in the superblock that the transaction modifies.
280 */
281void
282xfs_trans_mod_sb(
283 xfs_trans_t *tp,
284 uint field,
285 int64_t delta)
286{
287 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
288 xfs_mount_t *mp = tp->t_mountp;
289
290 switch (field) {
291 case XFS_TRANS_SB_ICOUNT:
292 tp->t_icount_delta += delta;
293 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
294 flags &= ~XFS_TRANS_SB_DIRTY;
295 break;
296 case XFS_TRANS_SB_IFREE:
297 tp->t_ifree_delta += delta;
298 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
299 flags &= ~XFS_TRANS_SB_DIRTY;
300 break;
301 case XFS_TRANS_SB_FDBLOCKS:
302 /*
303 * Track the number of blocks allocated in the
304 * transaction. Make sure it does not exceed the
305 * number reserved.
306 */
307 if (delta < 0) {
308 tp->t_blk_res_used += (uint)-delta;
309 ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
310 }
311 tp->t_fdblocks_delta += delta;
312 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
313 flags &= ~XFS_TRANS_SB_DIRTY;
314 break;
315 case XFS_TRANS_SB_RES_FDBLOCKS:
316 /*
317 * The allocation has already been applied to the
318 * in-core superblock's counter. This should only
319 * be applied to the on-disk superblock.
320 */
321 tp->t_res_fdblocks_delta += delta;
322 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
323 flags &= ~XFS_TRANS_SB_DIRTY;
324 break;
325 case XFS_TRANS_SB_FREXTENTS:
326 /*
327 * Track the number of blocks allocated in the
328 * transaction. Make sure it does not exceed the
329 * number reserved.
330 */
331 if (delta < 0) {
332 tp->t_rtx_res_used += (uint)-delta;
333 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
334 }
335 tp->t_frextents_delta += delta;
336 break;
337 case XFS_TRANS_SB_RES_FREXTENTS:
338 /*
339 * The allocation has already been applied to the
340 * in-core superblock's counter. This should only
341 * be applied to the on-disk superblock.
342 */
343 ASSERT(delta < 0);
344 tp->t_res_frextents_delta += delta;
345 break;
346 case XFS_TRANS_SB_DBLOCKS:
347 ASSERT(delta > 0);
348 tp->t_dblocks_delta += delta;
349 break;
350 case XFS_TRANS_SB_AGCOUNT:
351 ASSERT(delta > 0);
352 tp->t_agcount_delta += delta;
353 break;
354 case XFS_TRANS_SB_IMAXPCT:
355 tp->t_imaxpct_delta += delta;
356 break;
357 case XFS_TRANS_SB_REXTSIZE:
358 tp->t_rextsize_delta += delta;
359 break;
360 case XFS_TRANS_SB_RBMBLOCKS:
361 tp->t_rbmblocks_delta += delta;
362 break;
363 case XFS_TRANS_SB_RBLOCKS:
364 tp->t_rblocks_delta += delta;
365 break;
366 case XFS_TRANS_SB_REXTENTS:
367 tp->t_rextents_delta += delta;
368 break;
369 case XFS_TRANS_SB_REXTSLOG:
370 tp->t_rextslog_delta += delta;
371 break;
372 default:
373 ASSERT(0);
374 return;
375 }
376
377 tp->t_flags |= flags;
378}
379
380/*
381 * xfs_trans_apply_sb_deltas() is called from the commit code
382 * to bring the superblock buffer into the current transaction
383 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
384 *
385 * For now we just look at each field allowed to change and change
386 * it if necessary.
387 */
388STATIC void
389xfs_trans_apply_sb_deltas(
390 xfs_trans_t *tp)
391{
392 xfs_dsb_t *sbp;
393 xfs_buf_t *bp;
394 int whole = 0;
395
396 bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
397 sbp = XFS_BUF_TO_SBP(bp);
398
399 /*
400 * Check that superblock mods match the mods made to AGF counters.
401 */
402 ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
403 (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
404 tp->t_ag_btree_delta));
405
406 /*
407 * Only update the superblock counters if we are logging them
408 */
409 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
410 if (tp->t_icount_delta)
411 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
412 if (tp->t_ifree_delta)
413 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
414 if (tp->t_fdblocks_delta)
415 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
416 if (tp->t_res_fdblocks_delta)
417 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
418 }
419
420 if (tp->t_frextents_delta)
421 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
422 if (tp->t_res_frextents_delta)
423 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
424
425 if (tp->t_dblocks_delta) {
426 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
427 whole = 1;
428 }
429 if (tp->t_agcount_delta) {
430 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
431 whole = 1;
432 }
433 if (tp->t_imaxpct_delta) {
434 sbp->sb_imax_pct += tp->t_imaxpct_delta;
435 whole = 1;
436 }
437 if (tp->t_rextsize_delta) {
438 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
439 whole = 1;
440 }
441 if (tp->t_rbmblocks_delta) {
442 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
443 whole = 1;
444 }
445 if (tp->t_rblocks_delta) {
446 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
447 whole = 1;
448 }
449 if (tp->t_rextents_delta) {
450 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
451 whole = 1;
452 }
453 if (tp->t_rextslog_delta) {
454 sbp->sb_rextslog += tp->t_rextslog_delta;
455 whole = 1;
456 }
457
458 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
459 if (whole)
460 /*
461 * Log the whole thing, the fields are noncontiguous.
462 */
463 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
464 else
465 /*
466 * Since all the modifiable fields are contiguous, we
467 * can get away with this.
468 */
469 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
470 offsetof(xfs_dsb_t, sb_frextents) +
471 sizeof(sbp->sb_frextents) - 1);
472}
473
474STATIC int
475xfs_sb_mod8(
476 uint8_t *field,
477 int8_t delta)
478{
479 int8_t counter = *field;
480
481 counter += delta;
482 if (counter < 0) {
483 ASSERT(0);
484 return -EINVAL;
485 }
486 *field = counter;
487 return 0;
488}
489
490STATIC int
491xfs_sb_mod32(
492 uint32_t *field,
493 int32_t delta)
494{
495 int32_t counter = *field;
496
497 counter += delta;
498 if (counter < 0) {
499 ASSERT(0);
500 return -EINVAL;
501 }
502 *field = counter;
503 return 0;
504}
505
506STATIC int
507xfs_sb_mod64(
508 uint64_t *field,
509 int64_t delta)
510{
511 int64_t counter = *field;
512
513 counter += delta;
514 if (counter < 0) {
515 ASSERT(0);
516 return -EINVAL;
517 }
518 *field = counter;
519 return 0;
520}
521
522/*
523 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
524 * and apply superblock counter changes to the in-core superblock. The
525 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
526 * applied to the in-core superblock. The idea is that that has already been
527 * done.
528 *
529 * If we are not logging superblock counters, then the inode allocated/free and
530 * used block counts are not updated in the on disk superblock. In this case,
531 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
532 * still need to update the incore superblock with the changes.
533 */
534void
535xfs_trans_unreserve_and_mod_sb(
536 struct xfs_trans *tp)
537{
538 struct xfs_mount *mp = tp->t_mountp;
539 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
540 int64_t blkdelta = 0;
541 int64_t rtxdelta = 0;
542 int64_t idelta = 0;
543 int64_t ifreedelta = 0;
544 int error;
545
546 /* calculate deltas */
547 if (tp->t_blk_res > 0)
548 blkdelta = tp->t_blk_res;
549 if ((tp->t_fdblocks_delta != 0) &&
550 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
551 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
552 blkdelta += tp->t_fdblocks_delta;
553
554 if (tp->t_rtx_res > 0)
555 rtxdelta = tp->t_rtx_res;
556 if ((tp->t_frextents_delta != 0) &&
557 (tp->t_flags & XFS_TRANS_SB_DIRTY))
558 rtxdelta += tp->t_frextents_delta;
559
560 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
561 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
562 idelta = tp->t_icount_delta;
563 ifreedelta = tp->t_ifree_delta;
564 }
565
566 /* apply the per-cpu counters */
567 if (blkdelta) {
568 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
569 if (error)
570 goto out;
571 }
572
573 if (idelta) {
574 error = xfs_mod_icount(mp, idelta);
575 if (error)
576 goto out_undo_fdblocks;
577 }
578
579 if (ifreedelta) {
580 error = xfs_mod_ifree(mp, ifreedelta);
581 if (error)
582 goto out_undo_icount;
583 }
584
585 if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
586 return;
587
588 /* apply remaining deltas */
589 spin_lock(&mp->m_sb_lock);
590 if (rtxdelta) {
591 error = xfs_sb_mod64(&mp->m_sb.sb_frextents, rtxdelta);
592 if (error)
593 goto out_undo_ifree;
594 }
595
596 if (tp->t_dblocks_delta != 0) {
597 error = xfs_sb_mod64(&mp->m_sb.sb_dblocks, tp->t_dblocks_delta);
598 if (error)
599 goto out_undo_frextents;
600 }
601 if (tp->t_agcount_delta != 0) {
602 error = xfs_sb_mod32(&mp->m_sb.sb_agcount, tp->t_agcount_delta);
603 if (error)
604 goto out_undo_dblocks;
605 }
606 if (tp->t_imaxpct_delta != 0) {
607 error = xfs_sb_mod8(&mp->m_sb.sb_imax_pct, tp->t_imaxpct_delta);
608 if (error)
609 goto out_undo_agcount;
610 }
611 if (tp->t_rextsize_delta != 0) {
612 error = xfs_sb_mod32(&mp->m_sb.sb_rextsize,
613 tp->t_rextsize_delta);
614 if (error)
615 goto out_undo_imaxpct;
616 }
617 if (tp->t_rbmblocks_delta != 0) {
618 error = xfs_sb_mod32(&mp->m_sb.sb_rbmblocks,
619 tp->t_rbmblocks_delta);
620 if (error)
621 goto out_undo_rextsize;
622 }
623 if (tp->t_rblocks_delta != 0) {
624 error = xfs_sb_mod64(&mp->m_sb.sb_rblocks, tp->t_rblocks_delta);
625 if (error)
626 goto out_undo_rbmblocks;
627 }
628 if (tp->t_rextents_delta != 0) {
629 error = xfs_sb_mod64(&mp->m_sb.sb_rextents,
630 tp->t_rextents_delta);
631 if (error)
632 goto out_undo_rblocks;
633 }
634 if (tp->t_rextslog_delta != 0) {
635 error = xfs_sb_mod8(&mp->m_sb.sb_rextslog,
636 tp->t_rextslog_delta);
637 if (error)
638 goto out_undo_rextents;
639 }
640 spin_unlock(&mp->m_sb_lock);
641 return;
642
643out_undo_rextents:
644 if (tp->t_rextents_delta)
645 xfs_sb_mod64(&mp->m_sb.sb_rextents, -tp->t_rextents_delta);
646out_undo_rblocks:
647 if (tp->t_rblocks_delta)
648 xfs_sb_mod64(&mp->m_sb.sb_rblocks, -tp->t_rblocks_delta);
649out_undo_rbmblocks:
650 if (tp->t_rbmblocks_delta)
651 xfs_sb_mod32(&mp->m_sb.sb_rbmblocks, -tp->t_rbmblocks_delta);
652out_undo_rextsize:
653 if (tp->t_rextsize_delta)
654 xfs_sb_mod32(&mp->m_sb.sb_rextsize, -tp->t_rextsize_delta);
655out_undo_imaxpct:
656 if (tp->t_rextsize_delta)
657 xfs_sb_mod8(&mp->m_sb.sb_imax_pct, -tp->t_imaxpct_delta);
658out_undo_agcount:
659 if (tp->t_agcount_delta)
660 xfs_sb_mod32(&mp->m_sb.sb_agcount, -tp->t_agcount_delta);
661out_undo_dblocks:
662 if (tp->t_dblocks_delta)
663 xfs_sb_mod64(&mp->m_sb.sb_dblocks, -tp->t_dblocks_delta);
664out_undo_frextents:
665 if (rtxdelta)
666 xfs_sb_mod64(&mp->m_sb.sb_frextents, -rtxdelta);
667out_undo_ifree:
668 spin_unlock(&mp->m_sb_lock);
669 if (ifreedelta)
670 xfs_mod_ifree(mp, -ifreedelta);
671out_undo_icount:
672 if (idelta)
673 xfs_mod_icount(mp, -idelta);
674out_undo_fdblocks:
675 if (blkdelta)
676 xfs_mod_fdblocks(mp, -blkdelta, rsvd);
677out:
678 ASSERT(error == 0);
679 return;
680}
681
682/*
683 * Add the given log item to the transaction's list of log items.
684 *
685 * The log item will now point to its new descriptor with its li_desc field.
686 */
687void
688xfs_trans_add_item(
689 struct xfs_trans *tp,
690 struct xfs_log_item *lip)
691{
692 struct xfs_log_item_desc *lidp;
693
694 ASSERT(lip->li_mountp == tp->t_mountp);
695 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
696
697 lidp = kmem_zone_zalloc(xfs_log_item_desc_zone, KM_SLEEP | KM_NOFS);
698
699 lidp->lid_item = lip;
700 lidp->lid_flags = 0;
701 list_add_tail(&lidp->lid_trans, &tp->t_items);
702
703 lip->li_desc = lidp;
704}
705
706STATIC void
707xfs_trans_free_item_desc(
708 struct xfs_log_item_desc *lidp)
709{
710 list_del_init(&lidp->lid_trans);
711 kmem_zone_free(xfs_log_item_desc_zone, lidp);
712}
713
714/*
715 * Unlink and free the given descriptor.
716 */
717void
718xfs_trans_del_item(
719 struct xfs_log_item *lip)
720{
721 xfs_trans_free_item_desc(lip->li_desc);
722 lip->li_desc = NULL;
723}
724
725/*
726 * Unlock all of the items of a transaction and free all the descriptors
727 * of that transaction.
728 */
729void
730xfs_trans_free_items(
731 struct xfs_trans *tp,
732 xfs_lsn_t commit_lsn,
733 bool abort)
734{
735 struct xfs_log_item_desc *lidp, *next;
736
737 list_for_each_entry_safe(lidp, next, &tp->t_items, lid_trans) {
738 struct xfs_log_item *lip = lidp->lid_item;
739
740 lip->li_desc = NULL;
741
742 if (commit_lsn != NULLCOMMITLSN)
743 lip->li_ops->iop_committing(lip, commit_lsn);
744 if (abort)
745 lip->li_flags |= XFS_LI_ABORTED;
746 lip->li_ops->iop_unlock(lip);
747
748 xfs_trans_free_item_desc(lidp);
749 }
750}
751
752static inline void
753xfs_log_item_batch_insert(
754 struct xfs_ail *ailp,
755 struct xfs_ail_cursor *cur,
756 struct xfs_log_item **log_items,
757 int nr_items,
758 xfs_lsn_t commit_lsn)
759{
760 int i;
761
762 spin_lock(&ailp->xa_lock);
763 /* xfs_trans_ail_update_bulk drops ailp->xa_lock */
764 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
765
766 for (i = 0; i < nr_items; i++) {
767 struct xfs_log_item *lip = log_items[i];
768
769 lip->li_ops->iop_unpin(lip, 0);
770 }
771}
772
773/*
774 * Bulk operation version of xfs_trans_committed that takes a log vector of
775 * items to insert into the AIL. This uses bulk AIL insertion techniques to
776 * minimise lock traffic.
777 *
778 * If we are called with the aborted flag set, it is because a log write during
779 * a CIL checkpoint commit has failed. In this case, all the items in the
780 * checkpoint have already gone through iop_commited and iop_unlock, which
781 * means that checkpoint commit abort handling is treated exactly the same
782 * as an iclog write error even though we haven't started any IO yet. Hence in
783 * this case all we need to do is iop_committed processing, followed by an
784 * iop_unpin(aborted) call.
785 *
786 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
787 * at the end of the AIL, the insert cursor avoids the need to walk
788 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
789 * call. This saves a lot of needless list walking and is a net win, even
790 * though it slightly increases that amount of AIL lock traffic to set it up
791 * and tear it down.
792 */
793void
794xfs_trans_committed_bulk(
795 struct xfs_ail *ailp,
796 struct xfs_log_vec *log_vector,
797 xfs_lsn_t commit_lsn,
798 int aborted)
799{
800#define LOG_ITEM_BATCH_SIZE 32
801 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
802 struct xfs_log_vec *lv;
803 struct xfs_ail_cursor cur;
804 int i = 0;
805
806 spin_lock(&ailp->xa_lock);
807 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
808 spin_unlock(&ailp->xa_lock);
809
810 /* unpin all the log items */
811 for (lv = log_vector; lv; lv = lv->lv_next ) {
812 struct xfs_log_item *lip = lv->lv_item;
813 xfs_lsn_t item_lsn;
814
815 if (aborted)
816 lip->li_flags |= XFS_LI_ABORTED;
817 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
818
819 /* item_lsn of -1 means the item needs no further processing */
820 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
821 continue;
822
823 /*
824 * if we are aborting the operation, no point in inserting the
825 * object into the AIL as we are in a shutdown situation.
826 */
827 if (aborted) {
828 ASSERT(XFS_FORCED_SHUTDOWN(ailp->xa_mount));
829 lip->li_ops->iop_unpin(lip, 1);
830 continue;
831 }
832
833 if (item_lsn != commit_lsn) {
834
835 /*
836 * Not a bulk update option due to unusual item_lsn.
837 * Push into AIL immediately, rechecking the lsn once
838 * we have the ail lock. Then unpin the item. This does
839 * not affect the AIL cursor the bulk insert path is
840 * using.
841 */
842 spin_lock(&ailp->xa_lock);
843 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
844 xfs_trans_ail_update(ailp, lip, item_lsn);
845 else
846 spin_unlock(&ailp->xa_lock);
847 lip->li_ops->iop_unpin(lip, 0);
848 continue;
849 }
850
851 /* Item is a candidate for bulk AIL insert. */
852 log_items[i++] = lv->lv_item;
853 if (i >= LOG_ITEM_BATCH_SIZE) {
854 xfs_log_item_batch_insert(ailp, &cur, log_items,
855 LOG_ITEM_BATCH_SIZE, commit_lsn);
856 i = 0;
857 }
858 }
859
860 /* make sure we insert the remainder! */
861 if (i)
862 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
863
864 spin_lock(&ailp->xa_lock);
865 xfs_trans_ail_cursor_done(&cur);
866 spin_unlock(&ailp->xa_lock);
867}
868
869/*
870 * Commit the given transaction to the log.
871 *
872 * XFS disk error handling mechanism is not based on a typical
873 * transaction abort mechanism. Logically after the filesystem
874 * gets marked 'SHUTDOWN', we can't let any new transactions
875 * be durable - ie. committed to disk - because some metadata might
876 * be inconsistent. In such cases, this returns an error, and the
877 * caller may assume that all locked objects joined to the transaction
878 * have already been unlocked as if the commit had succeeded.
879 * Do not reference the transaction structure after this call.
880 */
881static int
882__xfs_trans_commit(
883 struct xfs_trans *tp,
884 bool regrant)
885{
886 struct xfs_mount *mp = tp->t_mountp;
887 xfs_lsn_t commit_lsn = -1;
888 int error = 0;
889 int sync = tp->t_flags & XFS_TRANS_SYNC;
890
891 /*
892 * If there is nothing to be logged by the transaction,
893 * then unlock all of the items associated with the
894 * transaction and free the transaction structure.
895 * Also make sure to return any reserved blocks to
896 * the free pool.
897 */
898 if (!(tp->t_flags & XFS_TRANS_DIRTY))
899 goto out_unreserve;
900
901 if (XFS_FORCED_SHUTDOWN(mp)) {
902 error = -EIO;
903 goto out_unreserve;
904 }
905
906 ASSERT(tp->t_ticket != NULL);
907
908 /*
909 * If we need to update the superblock, then do it now.
910 */
911 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
912 xfs_trans_apply_sb_deltas(tp);
913 xfs_trans_apply_dquot_deltas(tp);
914
915 xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
916
917 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
918 xfs_trans_free(tp);
919
920 /*
921 * If the transaction needs to be synchronous, then force the
922 * log out now and wait for it.
923 */
924 if (sync) {
925 error = _xfs_log_force_lsn(mp, commit_lsn, XFS_LOG_SYNC, NULL);
926 XFS_STATS_INC(mp, xs_trans_sync);
927 } else {
928 XFS_STATS_INC(mp, xs_trans_async);
929 }
930
931 return error;
932
933out_unreserve:
934 xfs_trans_unreserve_and_mod_sb(tp);
935
936 /*
937 * It is indeed possible for the transaction to be not dirty but
938 * the dqinfo portion to be. All that means is that we have some
939 * (non-persistent) quota reservations that need to be unreserved.
940 */
941 xfs_trans_unreserve_and_mod_dquots(tp);
942 if (tp->t_ticket) {
943 commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, regrant);
944 if (commit_lsn == -1 && !error)
945 error = -EIO;
946 }
947 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
948 xfs_trans_free_items(tp, NULLCOMMITLSN, !!error);
949 xfs_trans_free(tp);
950
951 XFS_STATS_INC(mp, xs_trans_empty);
952 return error;
953}
954
955int
956xfs_trans_commit(
957 struct xfs_trans *tp)
958{
959 return __xfs_trans_commit(tp, false);
960}
961
962/*
963 * Unlock all of the transaction's items and free the transaction.
964 * The transaction must not have modified any of its items, because
965 * there is no way to restore them to their previous state.
966 *
967 * If the transaction has made a log reservation, make sure to release
968 * it as well.
969 */
970void
971xfs_trans_cancel(
972 struct xfs_trans *tp)
973{
974 struct xfs_mount *mp = tp->t_mountp;
975 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
976
977 /*
978 * See if the caller is relying on us to shut down the
979 * filesystem. This happens in paths where we detect
980 * corruption and decide to give up.
981 */
982 if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
983 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
984 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
985 }
986#ifdef DEBUG
987 if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
988 struct xfs_log_item_desc *lidp;
989
990 list_for_each_entry(lidp, &tp->t_items, lid_trans)
991 ASSERT(!(lidp->lid_item->li_type == XFS_LI_EFD));
992 }
993#endif
994 xfs_trans_unreserve_and_mod_sb(tp);
995 xfs_trans_unreserve_and_mod_dquots(tp);
996
997 if (tp->t_ticket)
998 xfs_log_done(mp, tp->t_ticket, NULL, false);
999
1000 /* mark this thread as no longer being in a transaction */
1001 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1002
1003 xfs_trans_free_items(tp, NULLCOMMITLSN, dirty);
1004 xfs_trans_free(tp);
1005}
1006
1007/*
1008 * Roll from one trans in the sequence of PERMANENT transactions to
1009 * the next: permanent transactions are only flushed out when
1010 * committed with xfs_trans_commit(), but we still want as soon
1011 * as possible to let chunks of it go to the log. So we commit the
1012 * chunk we've been working on and get a new transaction to continue.
1013 */
1014int
1015__xfs_trans_roll(
1016 struct xfs_trans **tpp,
1017 struct xfs_inode *dp,
1018 int *committed)
1019{
1020 struct xfs_trans *trans;
1021 struct xfs_trans_res tres;
1022 int error;
1023
1024 *committed = 0;
1025
1026 /*
1027 * Ensure that the inode is always logged.
1028 */
1029 trans = *tpp;
1030 if (dp)
1031 xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
1032
1033 /*
1034 * Copy the critical parameters from one trans to the next.
1035 */
1036 tres.tr_logres = trans->t_log_res;
1037 tres.tr_logcount = trans->t_log_count;
1038 *tpp = xfs_trans_dup(trans);
1039
1040 /*
1041 * Commit the current transaction.
1042 * If this commit failed, then it'd just unlock those items that
1043 * are not marked ihold. That also means that a filesystem shutdown
1044 * is in progress. The caller takes the responsibility to cancel
1045 * the duplicate transaction that gets returned.
1046 */
1047 error = __xfs_trans_commit(trans, true);
1048 if (error)
1049 return error;
1050
1051 *committed = 1;
1052 trans = *tpp;
1053
1054 /*
1055 * Reserve space in the log for th next transaction.
1056 * This also pushes items in the "AIL", the list of logged items,
1057 * out to disk if they are taking up space at the tail of the log
1058 * that we want to use. This requires that either nothing be locked
1059 * across this call, or that anything that is locked be logged in
1060 * the prior and the next transactions.
1061 */
1062 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1063 error = xfs_trans_reserve(trans, &tres, 0, 0);
1064 /*
1065 * Ensure that the inode is in the new transaction and locked.
1066 */
1067 if (error)
1068 return error;
1069
1070 if (dp)
1071 xfs_trans_ijoin(trans, dp, 0);
1072 return 0;
1073}
1074
1075int
1076xfs_trans_roll(
1077 struct xfs_trans **tpp,
1078 struct xfs_inode *dp)
1079{
1080 int committed;
1081 return __xfs_trans_roll(tpp, dp, &committed);
1082}