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// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
4 * Copyright (C) 2010 Red Hat, Inc.
5 * All Rights Reserved.
6 */
7#include "xfs.h"
8#include "xfs_fs.h"
9#include "xfs_shared.h"
10#include "xfs_format.h"
11#include "xfs_log_format.h"
12#include "xfs_trans_resv.h"
13#include "xfs_mount.h"
14#include "xfs_extent_busy.h"
15#include "xfs_quota.h"
16#include "xfs_trans.h"
17#include "xfs_trans_priv.h"
18#include "xfs_log.h"
19#include "xfs_trace.h"
20#include "xfs_error.h"
21#include "xfs_defer.h"
22
23kmem_zone_t *xfs_trans_zone;
24
25#if defined(CONFIG_TRACEPOINTS)
26static void
27xfs_trans_trace_reservations(
28 struct xfs_mount *mp)
29{
30 struct xfs_trans_res resv;
31 struct xfs_trans_res *res;
32 struct xfs_trans_res *end_res;
33 int i;
34
35 res = (struct xfs_trans_res *)M_RES(mp);
36 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
37 for (i = 0; res < end_res; i++, res++)
38 trace_xfs_trans_resv_calc(mp, i, res);
39 xfs_log_get_max_trans_res(mp, &resv);
40 trace_xfs_trans_resv_calc(mp, -1, &resv);
41}
42#else
43# define xfs_trans_trace_reservations(mp)
44#endif
45
46/*
47 * Initialize the precomputed transaction reservation values
48 * in the mount structure.
49 */
50void
51xfs_trans_init(
52 struct xfs_mount *mp)
53{
54 xfs_trans_resv_calc(mp, M_RES(mp));
55 xfs_trans_trace_reservations(mp);
56}
57
58/*
59 * Free the transaction structure. If there is more clean up
60 * to do when the structure is freed, add it here.
61 */
62STATIC void
63xfs_trans_free(
64 struct xfs_trans *tp)
65{
66 xfs_extent_busy_sort(&tp->t_busy);
67 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
68
69 trace_xfs_trans_free(tp, _RET_IP_);
70 atomic_dec(&tp->t_mountp->m_active_trans);
71 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
72 sb_end_intwrite(tp->t_mountp->m_super);
73 xfs_trans_free_dqinfo(tp);
74 kmem_zone_free(xfs_trans_zone, tp);
75}
76
77/*
78 * This is called to create a new transaction which will share the
79 * permanent log reservation of the given transaction. The remaining
80 * unused block and rt extent reservations are also inherited. This
81 * implies that the original transaction is no longer allowed to allocate
82 * blocks. Locks and log items, however, are no inherited. They must
83 * be added to the new transaction explicitly.
84 */
85STATIC struct xfs_trans *
86xfs_trans_dup(
87 struct xfs_trans *tp)
88{
89 struct xfs_trans *ntp;
90
91 trace_xfs_trans_dup(tp, _RET_IP_);
92
93 ntp = kmem_zone_zalloc(xfs_trans_zone, 0);
94
95 /*
96 * Initialize the new transaction structure.
97 */
98 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
99 ntp->t_mountp = tp->t_mountp;
100 INIT_LIST_HEAD(&ntp->t_items);
101 INIT_LIST_HEAD(&ntp->t_busy);
102 INIT_LIST_HEAD(&ntp->t_dfops);
103 ntp->t_firstblock = NULLFSBLOCK;
104
105 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
106 ASSERT(tp->t_ticket != NULL);
107
108 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
109 (tp->t_flags & XFS_TRANS_RESERVE) |
110 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT);
111 /* We gave our writer reference to the new transaction */
112 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
113 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
114
115 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
116 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
117 tp->t_blk_res = tp->t_blk_res_used;
118
119 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
120 tp->t_rtx_res = tp->t_rtx_res_used;
121 ntp->t_pflags = tp->t_pflags;
122
123 /* move deferred ops over to the new tp */
124 xfs_defer_move(ntp, tp);
125
126 xfs_trans_dup_dqinfo(tp, ntp);
127
128 atomic_inc(&tp->t_mountp->m_active_trans);
129 return ntp;
130}
131
132/*
133 * This is called to reserve free disk blocks and log space for the
134 * given transaction. This must be done before allocating any resources
135 * within the transaction.
136 *
137 * This will return ENOSPC if there are not enough blocks available.
138 * It will sleep waiting for available log space.
139 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
140 * is used by long running transactions. If any one of the reservations
141 * fails then they will all be backed out.
142 *
143 * This does not do quota reservations. That typically is done by the
144 * caller afterwards.
145 */
146static int
147xfs_trans_reserve(
148 struct xfs_trans *tp,
149 struct xfs_trans_res *resp,
150 uint blocks,
151 uint rtextents)
152{
153 int error = 0;
154 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
155
156 /* Mark this thread as being in a transaction */
157 current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
158
159 /*
160 * Attempt to reserve the needed disk blocks by decrementing
161 * the number needed from the number available. This will
162 * fail if the count would go below zero.
163 */
164 if (blocks > 0) {
165 error = xfs_mod_fdblocks(tp->t_mountp, -((int64_t)blocks), rsvd);
166 if (error != 0) {
167 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
168 return -ENOSPC;
169 }
170 tp->t_blk_res += blocks;
171 }
172
173 /*
174 * Reserve the log space needed for this transaction.
175 */
176 if (resp->tr_logres > 0) {
177 bool permanent = false;
178
179 ASSERT(tp->t_log_res == 0 ||
180 tp->t_log_res == resp->tr_logres);
181 ASSERT(tp->t_log_count == 0 ||
182 tp->t_log_count == resp->tr_logcount);
183
184 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
185 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
186 permanent = true;
187 } else {
188 ASSERT(tp->t_ticket == NULL);
189 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
190 }
191
192 if (tp->t_ticket != NULL) {
193 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
194 error = xfs_log_regrant(tp->t_mountp, tp->t_ticket);
195 } else {
196 error = xfs_log_reserve(tp->t_mountp,
197 resp->tr_logres,
198 resp->tr_logcount,
199 &tp->t_ticket, XFS_TRANSACTION,
200 permanent);
201 }
202
203 if (error)
204 goto undo_blocks;
205
206 tp->t_log_res = resp->tr_logres;
207 tp->t_log_count = resp->tr_logcount;
208 }
209
210 /*
211 * Attempt to reserve the needed realtime extents by decrementing
212 * the number needed from the number available. This will
213 * fail if the count would go below zero.
214 */
215 if (rtextents > 0) {
216 error = xfs_mod_frextents(tp->t_mountp, -((int64_t)rtextents));
217 if (error) {
218 error = -ENOSPC;
219 goto undo_log;
220 }
221 tp->t_rtx_res += rtextents;
222 }
223
224 return 0;
225
226 /*
227 * Error cases jump to one of these labels to undo any
228 * reservations which have already been performed.
229 */
230undo_log:
231 if (resp->tr_logres > 0) {
232 xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, false);
233 tp->t_ticket = NULL;
234 tp->t_log_res = 0;
235 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
236 }
237
238undo_blocks:
239 if (blocks > 0) {
240 xfs_mod_fdblocks(tp->t_mountp, (int64_t)blocks, rsvd);
241 tp->t_blk_res = 0;
242 }
243
244 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
245
246 return error;
247}
248
249int
250xfs_trans_alloc(
251 struct xfs_mount *mp,
252 struct xfs_trans_res *resp,
253 uint blocks,
254 uint rtextents,
255 uint flags,
256 struct xfs_trans **tpp)
257{
258 struct xfs_trans *tp;
259 int error;
260
261 /*
262 * Allocate the handle before we do our freeze accounting and setting up
263 * GFP_NOFS allocation context so that we avoid lockdep false positives
264 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
265 */
266 tp = kmem_zone_zalloc(xfs_trans_zone, 0);
267 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
268 sb_start_intwrite(mp->m_super);
269
270 /*
271 * Zero-reservation ("empty") transactions can't modify anything, so
272 * they're allowed to run while we're frozen.
273 */
274 WARN_ON(resp->tr_logres > 0 &&
275 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
276 atomic_inc(&mp->m_active_trans);
277
278 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
279 tp->t_flags = flags;
280 tp->t_mountp = mp;
281 INIT_LIST_HEAD(&tp->t_items);
282 INIT_LIST_HEAD(&tp->t_busy);
283 INIT_LIST_HEAD(&tp->t_dfops);
284 tp->t_firstblock = NULLFSBLOCK;
285
286 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
287 if (error) {
288 xfs_trans_cancel(tp);
289 return error;
290 }
291
292 trace_xfs_trans_alloc(tp, _RET_IP_);
293
294 *tpp = tp;
295 return 0;
296}
297
298/*
299 * Create an empty transaction with no reservation. This is a defensive
300 * mechanism for routines that query metadata without actually modifying
301 * them -- if the metadata being queried is somehow cross-linked (think a
302 * btree block pointer that points higher in the tree), we risk deadlock.
303 * However, blocks grabbed as part of a transaction can be re-grabbed.
304 * The verifiers will notice the corrupt block and the operation will fail
305 * back to userspace without deadlocking.
306 *
307 * Note the zero-length reservation; this transaction MUST be cancelled
308 * without any dirty data.
309 */
310int
311xfs_trans_alloc_empty(
312 struct xfs_mount *mp,
313 struct xfs_trans **tpp)
314{
315 struct xfs_trans_res resv = {0};
316
317 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
318}
319
320/*
321 * Record the indicated change to the given field for application
322 * to the file system's superblock when the transaction commits.
323 * For now, just store the change in the transaction structure.
324 *
325 * Mark the transaction structure to indicate that the superblock
326 * needs to be updated before committing.
327 *
328 * Because we may not be keeping track of allocated/free inodes and
329 * used filesystem blocks in the superblock, we do not mark the
330 * superblock dirty in this transaction if we modify these fields.
331 * We still need to update the transaction deltas so that they get
332 * applied to the incore superblock, but we don't want them to
333 * cause the superblock to get locked and logged if these are the
334 * only fields in the superblock that the transaction modifies.
335 */
336void
337xfs_trans_mod_sb(
338 xfs_trans_t *tp,
339 uint field,
340 int64_t delta)
341{
342 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
343 xfs_mount_t *mp = tp->t_mountp;
344
345 switch (field) {
346 case XFS_TRANS_SB_ICOUNT:
347 tp->t_icount_delta += delta;
348 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
349 flags &= ~XFS_TRANS_SB_DIRTY;
350 break;
351 case XFS_TRANS_SB_IFREE:
352 tp->t_ifree_delta += delta;
353 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
354 flags &= ~XFS_TRANS_SB_DIRTY;
355 break;
356 case XFS_TRANS_SB_FDBLOCKS:
357 /*
358 * Track the number of blocks allocated in the transaction.
359 * Make sure it does not exceed the number reserved. If so,
360 * shutdown as this can lead to accounting inconsistency.
361 */
362 if (delta < 0) {
363 tp->t_blk_res_used += (uint)-delta;
364 if (tp->t_blk_res_used > tp->t_blk_res)
365 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
366 }
367 tp->t_fdblocks_delta += delta;
368 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
369 flags &= ~XFS_TRANS_SB_DIRTY;
370 break;
371 case XFS_TRANS_SB_RES_FDBLOCKS:
372 /*
373 * The allocation has already been applied to the
374 * in-core superblock's counter. This should only
375 * be applied to the on-disk superblock.
376 */
377 tp->t_res_fdblocks_delta += delta;
378 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
379 flags &= ~XFS_TRANS_SB_DIRTY;
380 break;
381 case XFS_TRANS_SB_FREXTENTS:
382 /*
383 * Track the number of blocks allocated in the
384 * transaction. Make sure it does not exceed the
385 * number reserved.
386 */
387 if (delta < 0) {
388 tp->t_rtx_res_used += (uint)-delta;
389 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
390 }
391 tp->t_frextents_delta += delta;
392 break;
393 case XFS_TRANS_SB_RES_FREXTENTS:
394 /*
395 * The allocation has already been applied to the
396 * in-core superblock's counter. This should only
397 * be applied to the on-disk superblock.
398 */
399 ASSERT(delta < 0);
400 tp->t_res_frextents_delta += delta;
401 break;
402 case XFS_TRANS_SB_DBLOCKS:
403 ASSERT(delta > 0);
404 tp->t_dblocks_delta += delta;
405 break;
406 case XFS_TRANS_SB_AGCOUNT:
407 ASSERT(delta > 0);
408 tp->t_agcount_delta += delta;
409 break;
410 case XFS_TRANS_SB_IMAXPCT:
411 tp->t_imaxpct_delta += delta;
412 break;
413 case XFS_TRANS_SB_REXTSIZE:
414 tp->t_rextsize_delta += delta;
415 break;
416 case XFS_TRANS_SB_RBMBLOCKS:
417 tp->t_rbmblocks_delta += delta;
418 break;
419 case XFS_TRANS_SB_RBLOCKS:
420 tp->t_rblocks_delta += delta;
421 break;
422 case XFS_TRANS_SB_REXTENTS:
423 tp->t_rextents_delta += delta;
424 break;
425 case XFS_TRANS_SB_REXTSLOG:
426 tp->t_rextslog_delta += delta;
427 break;
428 default:
429 ASSERT(0);
430 return;
431 }
432
433 tp->t_flags |= flags;
434}
435
436/*
437 * xfs_trans_apply_sb_deltas() is called from the commit code
438 * to bring the superblock buffer into the current transaction
439 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
440 *
441 * For now we just look at each field allowed to change and change
442 * it if necessary.
443 */
444STATIC void
445xfs_trans_apply_sb_deltas(
446 xfs_trans_t *tp)
447{
448 xfs_dsb_t *sbp;
449 xfs_buf_t *bp;
450 int whole = 0;
451
452 bp = xfs_trans_getsb(tp, tp->t_mountp);
453 sbp = XFS_BUF_TO_SBP(bp);
454
455 /*
456 * Check that superblock mods match the mods made to AGF counters.
457 */
458 ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
459 (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
460 tp->t_ag_btree_delta));
461
462 /*
463 * Only update the superblock counters if we are logging them
464 */
465 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
466 if (tp->t_icount_delta)
467 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
468 if (tp->t_ifree_delta)
469 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
470 if (tp->t_fdblocks_delta)
471 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
472 if (tp->t_res_fdblocks_delta)
473 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
474 }
475
476 if (tp->t_frextents_delta)
477 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
478 if (tp->t_res_frextents_delta)
479 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
480
481 if (tp->t_dblocks_delta) {
482 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
483 whole = 1;
484 }
485 if (tp->t_agcount_delta) {
486 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
487 whole = 1;
488 }
489 if (tp->t_imaxpct_delta) {
490 sbp->sb_imax_pct += tp->t_imaxpct_delta;
491 whole = 1;
492 }
493 if (tp->t_rextsize_delta) {
494 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
495 whole = 1;
496 }
497 if (tp->t_rbmblocks_delta) {
498 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
499 whole = 1;
500 }
501 if (tp->t_rblocks_delta) {
502 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
503 whole = 1;
504 }
505 if (tp->t_rextents_delta) {
506 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
507 whole = 1;
508 }
509 if (tp->t_rextslog_delta) {
510 sbp->sb_rextslog += tp->t_rextslog_delta;
511 whole = 1;
512 }
513
514 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
515 if (whole)
516 /*
517 * Log the whole thing, the fields are noncontiguous.
518 */
519 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
520 else
521 /*
522 * Since all the modifiable fields are contiguous, we
523 * can get away with this.
524 */
525 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
526 offsetof(xfs_dsb_t, sb_frextents) +
527 sizeof(sbp->sb_frextents) - 1);
528}
529
530STATIC int
531xfs_sb_mod8(
532 uint8_t *field,
533 int8_t delta)
534{
535 int8_t counter = *field;
536
537 counter += delta;
538 if (counter < 0) {
539 ASSERT(0);
540 return -EINVAL;
541 }
542 *field = counter;
543 return 0;
544}
545
546STATIC int
547xfs_sb_mod32(
548 uint32_t *field,
549 int32_t delta)
550{
551 int32_t counter = *field;
552
553 counter += delta;
554 if (counter < 0) {
555 ASSERT(0);
556 return -EINVAL;
557 }
558 *field = counter;
559 return 0;
560}
561
562STATIC int
563xfs_sb_mod64(
564 uint64_t *field,
565 int64_t delta)
566{
567 int64_t counter = *field;
568
569 counter += delta;
570 if (counter < 0) {
571 ASSERT(0);
572 return -EINVAL;
573 }
574 *field = counter;
575 return 0;
576}
577
578/*
579 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
580 * and apply superblock counter changes to the in-core superblock. The
581 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
582 * applied to the in-core superblock. The idea is that that has already been
583 * done.
584 *
585 * If we are not logging superblock counters, then the inode allocated/free and
586 * used block counts are not updated in the on disk superblock. In this case,
587 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
588 * still need to update the incore superblock with the changes.
589 */
590void
591xfs_trans_unreserve_and_mod_sb(
592 struct xfs_trans *tp)
593{
594 struct xfs_mount *mp = tp->t_mountp;
595 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
596 int64_t blkdelta = 0;
597 int64_t rtxdelta = 0;
598 int64_t idelta = 0;
599 int64_t ifreedelta = 0;
600 int error;
601
602 /* calculate deltas */
603 if (tp->t_blk_res > 0)
604 blkdelta = tp->t_blk_res;
605 if ((tp->t_fdblocks_delta != 0) &&
606 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
607 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
608 blkdelta += tp->t_fdblocks_delta;
609
610 if (tp->t_rtx_res > 0)
611 rtxdelta = tp->t_rtx_res;
612 if ((tp->t_frextents_delta != 0) &&
613 (tp->t_flags & XFS_TRANS_SB_DIRTY))
614 rtxdelta += tp->t_frextents_delta;
615
616 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
617 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
618 idelta = tp->t_icount_delta;
619 ifreedelta = tp->t_ifree_delta;
620 }
621
622 /* apply the per-cpu counters */
623 if (blkdelta) {
624 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
625 if (error)
626 goto out;
627 }
628
629 if (idelta) {
630 error = xfs_mod_icount(mp, idelta);
631 if (error)
632 goto out_undo_fdblocks;
633 }
634
635 if (ifreedelta) {
636 error = xfs_mod_ifree(mp, ifreedelta);
637 if (error)
638 goto out_undo_icount;
639 }
640
641 if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
642 return;
643
644 /* apply remaining deltas */
645 spin_lock(&mp->m_sb_lock);
646 if (rtxdelta) {
647 error = xfs_sb_mod64(&mp->m_sb.sb_frextents, rtxdelta);
648 if (error)
649 goto out_undo_ifree;
650 }
651
652 if (tp->t_dblocks_delta != 0) {
653 error = xfs_sb_mod64(&mp->m_sb.sb_dblocks, tp->t_dblocks_delta);
654 if (error)
655 goto out_undo_frextents;
656 }
657 if (tp->t_agcount_delta != 0) {
658 error = xfs_sb_mod32(&mp->m_sb.sb_agcount, tp->t_agcount_delta);
659 if (error)
660 goto out_undo_dblocks;
661 }
662 if (tp->t_imaxpct_delta != 0) {
663 error = xfs_sb_mod8(&mp->m_sb.sb_imax_pct, tp->t_imaxpct_delta);
664 if (error)
665 goto out_undo_agcount;
666 }
667 if (tp->t_rextsize_delta != 0) {
668 error = xfs_sb_mod32(&mp->m_sb.sb_rextsize,
669 tp->t_rextsize_delta);
670 if (error)
671 goto out_undo_imaxpct;
672 }
673 if (tp->t_rbmblocks_delta != 0) {
674 error = xfs_sb_mod32(&mp->m_sb.sb_rbmblocks,
675 tp->t_rbmblocks_delta);
676 if (error)
677 goto out_undo_rextsize;
678 }
679 if (tp->t_rblocks_delta != 0) {
680 error = xfs_sb_mod64(&mp->m_sb.sb_rblocks, tp->t_rblocks_delta);
681 if (error)
682 goto out_undo_rbmblocks;
683 }
684 if (tp->t_rextents_delta != 0) {
685 error = xfs_sb_mod64(&mp->m_sb.sb_rextents,
686 tp->t_rextents_delta);
687 if (error)
688 goto out_undo_rblocks;
689 }
690 if (tp->t_rextslog_delta != 0) {
691 error = xfs_sb_mod8(&mp->m_sb.sb_rextslog,
692 tp->t_rextslog_delta);
693 if (error)
694 goto out_undo_rextents;
695 }
696 spin_unlock(&mp->m_sb_lock);
697 return;
698
699out_undo_rextents:
700 if (tp->t_rextents_delta)
701 xfs_sb_mod64(&mp->m_sb.sb_rextents, -tp->t_rextents_delta);
702out_undo_rblocks:
703 if (tp->t_rblocks_delta)
704 xfs_sb_mod64(&mp->m_sb.sb_rblocks, -tp->t_rblocks_delta);
705out_undo_rbmblocks:
706 if (tp->t_rbmblocks_delta)
707 xfs_sb_mod32(&mp->m_sb.sb_rbmblocks, -tp->t_rbmblocks_delta);
708out_undo_rextsize:
709 if (tp->t_rextsize_delta)
710 xfs_sb_mod32(&mp->m_sb.sb_rextsize, -tp->t_rextsize_delta);
711out_undo_imaxpct:
712 if (tp->t_rextsize_delta)
713 xfs_sb_mod8(&mp->m_sb.sb_imax_pct, -tp->t_imaxpct_delta);
714out_undo_agcount:
715 if (tp->t_agcount_delta)
716 xfs_sb_mod32(&mp->m_sb.sb_agcount, -tp->t_agcount_delta);
717out_undo_dblocks:
718 if (tp->t_dblocks_delta)
719 xfs_sb_mod64(&mp->m_sb.sb_dblocks, -tp->t_dblocks_delta);
720out_undo_frextents:
721 if (rtxdelta)
722 xfs_sb_mod64(&mp->m_sb.sb_frextents, -rtxdelta);
723out_undo_ifree:
724 spin_unlock(&mp->m_sb_lock);
725 if (ifreedelta)
726 xfs_mod_ifree(mp, -ifreedelta);
727out_undo_icount:
728 if (idelta)
729 xfs_mod_icount(mp, -idelta);
730out_undo_fdblocks:
731 if (blkdelta)
732 xfs_mod_fdblocks(mp, -blkdelta, rsvd);
733out:
734 ASSERT(error == 0);
735 return;
736}
737
738/* Add the given log item to the transaction's list of log items. */
739void
740xfs_trans_add_item(
741 struct xfs_trans *tp,
742 struct xfs_log_item *lip)
743{
744 ASSERT(lip->li_mountp == tp->t_mountp);
745 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
746 ASSERT(list_empty(&lip->li_trans));
747 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
748
749 list_add_tail(&lip->li_trans, &tp->t_items);
750 trace_xfs_trans_add_item(tp, _RET_IP_);
751}
752
753/*
754 * Unlink the log item from the transaction. the log item is no longer
755 * considered dirty in this transaction, as the linked transaction has
756 * finished, either by abort or commit completion.
757 */
758void
759xfs_trans_del_item(
760 struct xfs_log_item *lip)
761{
762 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
763 list_del_init(&lip->li_trans);
764}
765
766/* Detach and unlock all of the items in a transaction */
767static void
768xfs_trans_free_items(
769 struct xfs_trans *tp,
770 bool abort)
771{
772 struct xfs_log_item *lip, *next;
773
774 trace_xfs_trans_free_items(tp, _RET_IP_);
775
776 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
777 xfs_trans_del_item(lip);
778 if (abort)
779 set_bit(XFS_LI_ABORTED, &lip->li_flags);
780 if (lip->li_ops->iop_release)
781 lip->li_ops->iop_release(lip);
782 }
783}
784
785static inline void
786xfs_log_item_batch_insert(
787 struct xfs_ail *ailp,
788 struct xfs_ail_cursor *cur,
789 struct xfs_log_item **log_items,
790 int nr_items,
791 xfs_lsn_t commit_lsn)
792{
793 int i;
794
795 spin_lock(&ailp->ail_lock);
796 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
797 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
798
799 for (i = 0; i < nr_items; i++) {
800 struct xfs_log_item *lip = log_items[i];
801
802 if (lip->li_ops->iop_unpin)
803 lip->li_ops->iop_unpin(lip, 0);
804 }
805}
806
807/*
808 * Bulk operation version of xfs_trans_committed that takes a log vector of
809 * items to insert into the AIL. This uses bulk AIL insertion techniques to
810 * minimise lock traffic.
811 *
812 * If we are called with the aborted flag set, it is because a log write during
813 * a CIL checkpoint commit has failed. In this case, all the items in the
814 * checkpoint have already gone through iop_committed and iop_committing, which
815 * means that checkpoint commit abort handling is treated exactly the same
816 * as an iclog write error even though we haven't started any IO yet. Hence in
817 * this case all we need to do is iop_committed processing, followed by an
818 * iop_unpin(aborted) call.
819 *
820 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
821 * at the end of the AIL, the insert cursor avoids the need to walk
822 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
823 * call. This saves a lot of needless list walking and is a net win, even
824 * though it slightly increases that amount of AIL lock traffic to set it up
825 * and tear it down.
826 */
827void
828xfs_trans_committed_bulk(
829 struct xfs_ail *ailp,
830 struct xfs_log_vec *log_vector,
831 xfs_lsn_t commit_lsn,
832 bool aborted)
833{
834#define LOG_ITEM_BATCH_SIZE 32
835 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
836 struct xfs_log_vec *lv;
837 struct xfs_ail_cursor cur;
838 int i = 0;
839
840 spin_lock(&ailp->ail_lock);
841 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
842 spin_unlock(&ailp->ail_lock);
843
844 /* unpin all the log items */
845 for (lv = log_vector; lv; lv = lv->lv_next ) {
846 struct xfs_log_item *lip = lv->lv_item;
847 xfs_lsn_t item_lsn;
848
849 if (aborted)
850 set_bit(XFS_LI_ABORTED, &lip->li_flags);
851
852 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
853 lip->li_ops->iop_release(lip);
854 continue;
855 }
856
857 if (lip->li_ops->iop_committed)
858 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
859 else
860 item_lsn = commit_lsn;
861
862 /* item_lsn of -1 means the item needs no further processing */
863 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
864 continue;
865
866 /*
867 * if we are aborting the operation, no point in inserting the
868 * object into the AIL as we are in a shutdown situation.
869 */
870 if (aborted) {
871 ASSERT(XFS_FORCED_SHUTDOWN(ailp->ail_mount));
872 if (lip->li_ops->iop_unpin)
873 lip->li_ops->iop_unpin(lip, 1);
874 continue;
875 }
876
877 if (item_lsn != commit_lsn) {
878
879 /*
880 * Not a bulk update option due to unusual item_lsn.
881 * Push into AIL immediately, rechecking the lsn once
882 * we have the ail lock. Then unpin the item. This does
883 * not affect the AIL cursor the bulk insert path is
884 * using.
885 */
886 spin_lock(&ailp->ail_lock);
887 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
888 xfs_trans_ail_update(ailp, lip, item_lsn);
889 else
890 spin_unlock(&ailp->ail_lock);
891 if (lip->li_ops->iop_unpin)
892 lip->li_ops->iop_unpin(lip, 0);
893 continue;
894 }
895
896 /* Item is a candidate for bulk AIL insert. */
897 log_items[i++] = lv->lv_item;
898 if (i >= LOG_ITEM_BATCH_SIZE) {
899 xfs_log_item_batch_insert(ailp, &cur, log_items,
900 LOG_ITEM_BATCH_SIZE, commit_lsn);
901 i = 0;
902 }
903 }
904
905 /* make sure we insert the remainder! */
906 if (i)
907 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
908
909 spin_lock(&ailp->ail_lock);
910 xfs_trans_ail_cursor_done(&cur);
911 spin_unlock(&ailp->ail_lock);
912}
913
914/*
915 * Commit the given transaction to the log.
916 *
917 * XFS disk error handling mechanism is not based on a typical
918 * transaction abort mechanism. Logically after the filesystem
919 * gets marked 'SHUTDOWN', we can't let any new transactions
920 * be durable - ie. committed to disk - because some metadata might
921 * be inconsistent. In such cases, this returns an error, and the
922 * caller may assume that all locked objects joined to the transaction
923 * have already been unlocked as if the commit had succeeded.
924 * Do not reference the transaction structure after this call.
925 */
926static int
927__xfs_trans_commit(
928 struct xfs_trans *tp,
929 bool regrant)
930{
931 struct xfs_mount *mp = tp->t_mountp;
932 xfs_lsn_t commit_lsn = -1;
933 int error = 0;
934 int sync = tp->t_flags & XFS_TRANS_SYNC;
935
936 trace_xfs_trans_commit(tp, _RET_IP_);
937
938 /*
939 * Finish deferred items on final commit. Only permanent transactions
940 * should ever have deferred ops.
941 */
942 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
943 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
944 if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
945 error = xfs_defer_finish_noroll(&tp);
946 if (error)
947 goto out_unreserve;
948 }
949
950 /*
951 * If there is nothing to be logged by the transaction,
952 * then unlock all of the items associated with the
953 * transaction and free the transaction structure.
954 * Also make sure to return any reserved blocks to
955 * the free pool.
956 */
957 if (!(tp->t_flags & XFS_TRANS_DIRTY))
958 goto out_unreserve;
959
960 if (XFS_FORCED_SHUTDOWN(mp)) {
961 error = -EIO;
962 goto out_unreserve;
963 }
964
965 ASSERT(tp->t_ticket != NULL);
966
967 /*
968 * If we need to update the superblock, then do it now.
969 */
970 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
971 xfs_trans_apply_sb_deltas(tp);
972 xfs_trans_apply_dquot_deltas(tp);
973
974 xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
975
976 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
977 xfs_trans_free(tp);
978
979 /*
980 * If the transaction needs to be synchronous, then force the
981 * log out now and wait for it.
982 */
983 if (sync) {
984 error = xfs_log_force_lsn(mp, commit_lsn, XFS_LOG_SYNC, NULL);
985 XFS_STATS_INC(mp, xs_trans_sync);
986 } else {
987 XFS_STATS_INC(mp, xs_trans_async);
988 }
989
990 return error;
991
992out_unreserve:
993 xfs_trans_unreserve_and_mod_sb(tp);
994
995 /*
996 * It is indeed possible for the transaction to be not dirty but
997 * the dqinfo portion to be. All that means is that we have some
998 * (non-persistent) quota reservations that need to be unreserved.
999 */
1000 xfs_trans_unreserve_and_mod_dquots(tp);
1001 if (tp->t_ticket) {
1002 commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, regrant);
1003 if (commit_lsn == -1 && !error)
1004 error = -EIO;
1005 tp->t_ticket = NULL;
1006 }
1007 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
1008 xfs_trans_free_items(tp, !!error);
1009 xfs_trans_free(tp);
1010
1011 XFS_STATS_INC(mp, xs_trans_empty);
1012 return error;
1013}
1014
1015int
1016xfs_trans_commit(
1017 struct xfs_trans *tp)
1018{
1019 return __xfs_trans_commit(tp, false);
1020}
1021
1022/*
1023 * Unlock all of the transaction's items and free the transaction.
1024 * The transaction must not have modified any of its items, because
1025 * there is no way to restore them to their previous state.
1026 *
1027 * If the transaction has made a log reservation, make sure to release
1028 * it as well.
1029 */
1030void
1031xfs_trans_cancel(
1032 struct xfs_trans *tp)
1033{
1034 struct xfs_mount *mp = tp->t_mountp;
1035 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
1036
1037 trace_xfs_trans_cancel(tp, _RET_IP_);
1038
1039 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
1040 xfs_defer_cancel(tp);
1041
1042 /*
1043 * See if the caller is relying on us to shut down the
1044 * filesystem. This happens in paths where we detect
1045 * corruption and decide to give up.
1046 */
1047 if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1048 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1049 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1050 }
1051#ifdef DEBUG
1052 if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1053 struct xfs_log_item *lip;
1054
1055 list_for_each_entry(lip, &tp->t_items, li_trans)
1056 ASSERT(!(lip->li_type == XFS_LI_EFD));
1057 }
1058#endif
1059 xfs_trans_unreserve_and_mod_sb(tp);
1060 xfs_trans_unreserve_and_mod_dquots(tp);
1061
1062 if (tp->t_ticket) {
1063 xfs_log_done(mp, tp->t_ticket, NULL, false);
1064 tp->t_ticket = NULL;
1065 }
1066
1067 /* mark this thread as no longer being in a transaction */
1068 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
1069
1070 xfs_trans_free_items(tp, dirty);
1071 xfs_trans_free(tp);
1072}
1073
1074/*
1075 * Roll from one trans in the sequence of PERMANENT transactions to
1076 * the next: permanent transactions are only flushed out when
1077 * committed with xfs_trans_commit(), but we still want as soon
1078 * as possible to let chunks of it go to the log. So we commit the
1079 * chunk we've been working on and get a new transaction to continue.
1080 */
1081int
1082xfs_trans_roll(
1083 struct xfs_trans **tpp)
1084{
1085 struct xfs_trans *trans = *tpp;
1086 struct xfs_trans_res tres;
1087 int error;
1088
1089 trace_xfs_trans_roll(trans, _RET_IP_);
1090
1091 /*
1092 * Copy the critical parameters from one trans to the next.
1093 */
1094 tres.tr_logres = trans->t_log_res;
1095 tres.tr_logcount = trans->t_log_count;
1096
1097 *tpp = xfs_trans_dup(trans);
1098
1099 /*
1100 * Commit the current transaction.
1101 * If this commit failed, then it'd just unlock those items that
1102 * are not marked ihold. That also means that a filesystem shutdown
1103 * is in progress. The caller takes the responsibility to cancel
1104 * the duplicate transaction that gets returned.
1105 */
1106 error = __xfs_trans_commit(trans, true);
1107 if (error)
1108 return error;
1109
1110 /*
1111 * Reserve space in the log for the next transaction.
1112 * This also pushes items in the "AIL", the list of logged items,
1113 * out to disk if they are taking up space at the tail of the log
1114 * that we want to use. This requires that either nothing be locked
1115 * across this call, or that anything that is locked be logged in
1116 * the prior and the next transactions.
1117 */
1118 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1119 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1120}