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_log_priv.h"
20#include "xfs_trace.h"
21#include "xfs_error.h"
22#include "xfs_defer.h"
23#include "xfs_inode.h"
24#include "xfs_dquot_item.h"
25#include "xfs_dquot.h"
26#include "xfs_icache.h"
27#include "xfs_rtbitmap.h"
28#include "xfs_rtgroup.h"
29#include "xfs_sb.h"
30
31struct kmem_cache *xfs_trans_cache;
32
33#if defined(CONFIG_TRACEPOINTS)
34static void
35xfs_trans_trace_reservations(
36 struct xfs_mount *mp)
37{
38 struct xfs_trans_res *res;
39 struct xfs_trans_res *end_res;
40 int i;
41
42 res = (struct xfs_trans_res *)M_RES(mp);
43 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
44 for (i = 0; res < end_res; i++, res++)
45 trace_xfs_trans_resv_calc(mp, i, res);
46}
47#else
48# define xfs_trans_trace_reservations(mp)
49#endif
50
51/*
52 * Initialize the precomputed transaction reservation values
53 * in the mount structure.
54 */
55void
56xfs_trans_init(
57 struct xfs_mount *mp)
58{
59 xfs_trans_resv_calc(mp, M_RES(mp));
60 xfs_trans_trace_reservations(mp);
61}
62
63/*
64 * Free the transaction structure. If there is more clean up
65 * to do when the structure is freed, add it here.
66 */
67STATIC void
68xfs_trans_free(
69 struct xfs_trans *tp)
70{
71 xfs_extent_busy_sort(&tp->t_busy);
72 xfs_extent_busy_clear(&tp->t_busy, false);
73
74 trace_xfs_trans_free(tp, _RET_IP_);
75 xfs_trans_clear_context(tp);
76 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
77 sb_end_intwrite(tp->t_mountp->m_super);
78 xfs_trans_free_dqinfo(tp);
79 kmem_cache_free(xfs_trans_cache, tp);
80}
81
82/*
83 * This is called to create a new transaction which will share the
84 * permanent log reservation of the given transaction. The remaining
85 * unused block and rt extent reservations are also inherited. This
86 * implies that the original transaction is no longer allowed to allocate
87 * blocks. Locks and log items, however, are no inherited. They must
88 * be added to the new transaction explicitly.
89 */
90STATIC struct xfs_trans *
91xfs_trans_dup(
92 struct xfs_trans *tp)
93{
94 struct xfs_trans *ntp;
95
96 trace_xfs_trans_dup(tp, _RET_IP_);
97
98 ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
99
100 /*
101 * Initialize the new transaction structure.
102 */
103 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
104 ntp->t_mountp = tp->t_mountp;
105 INIT_LIST_HEAD(&ntp->t_items);
106 INIT_LIST_HEAD(&ntp->t_busy);
107 INIT_LIST_HEAD(&ntp->t_dfops);
108 ntp->t_highest_agno = NULLAGNUMBER;
109
110 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
111 ASSERT(tp->t_ticket != NULL);
112
113 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
114 (tp->t_flags & XFS_TRANS_RESERVE) |
115 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
116 (tp->t_flags & XFS_TRANS_RES_FDBLKS);
117 /* We gave our writer reference to the new transaction */
118 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
119 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
120
121 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
122 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
123 tp->t_blk_res = tp->t_blk_res_used;
124
125 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
126 tp->t_rtx_res = tp->t_rtx_res_used;
127
128 xfs_trans_switch_context(tp, ntp);
129
130 /* move deferred ops over to the new tp */
131 xfs_defer_move(ntp, tp);
132
133 xfs_trans_dup_dqinfo(tp, ntp);
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 struct xfs_mount *mp = tp->t_mountp;
159 int error = 0;
160 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
161
162 /*
163 * Attempt to reserve the needed disk blocks by decrementing
164 * the number needed from the number available. This will
165 * fail if the count would go below zero.
166 */
167 if (blocks > 0) {
168 error = xfs_dec_fdblocks(mp, blocks, rsvd);
169 if (error != 0)
170 return -ENOSPC;
171 tp->t_blk_res += blocks;
172 }
173
174 /*
175 * Reserve the log space needed for this transaction.
176 */
177 if (resp->tr_logres > 0) {
178 bool permanent = false;
179
180 ASSERT(tp->t_log_res == 0 ||
181 tp->t_log_res == resp->tr_logres);
182 ASSERT(tp->t_log_count == 0 ||
183 tp->t_log_count == resp->tr_logcount);
184
185 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
186 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
187 permanent = true;
188 } else {
189 ASSERT(tp->t_ticket == NULL);
190 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
191 }
192
193 if (tp->t_ticket != NULL) {
194 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
195 error = xfs_log_regrant(mp, tp->t_ticket);
196 } else {
197 error = xfs_log_reserve(mp, resp->tr_logres,
198 resp->tr_logcount,
199 &tp->t_ticket, permanent);
200 }
201
202 if (error)
203 goto undo_blocks;
204
205 tp->t_log_res = resp->tr_logres;
206 tp->t_log_count = resp->tr_logcount;
207 }
208
209 /*
210 * Attempt to reserve the needed realtime extents by decrementing
211 * the number needed from the number available. This will
212 * fail if the count would go below zero.
213 */
214 if (rtextents > 0) {
215 error = xfs_dec_frextents(mp, rtextents);
216 if (error) {
217 error = -ENOSPC;
218 goto undo_log;
219 }
220 tp->t_rtx_res += rtextents;
221 }
222
223 return 0;
224
225 /*
226 * Error cases jump to one of these labels to undo any
227 * reservations which have already been performed.
228 */
229undo_log:
230 if (resp->tr_logres > 0) {
231 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
232 tp->t_ticket = NULL;
233 tp->t_log_res = 0;
234 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
235 }
236
237undo_blocks:
238 if (blocks > 0) {
239 xfs_add_fdblocks(mp, blocks);
240 tp->t_blk_res = 0;
241 }
242 return error;
243}
244
245int
246xfs_trans_alloc(
247 struct xfs_mount *mp,
248 struct xfs_trans_res *resp,
249 uint blocks,
250 uint rtextents,
251 uint flags,
252 struct xfs_trans **tpp)
253{
254 struct xfs_trans *tp;
255 bool want_retry = true;
256 int error;
257
258 /*
259 * Allocate the handle before we do our freeze accounting and setting up
260 * GFP_NOFS allocation context so that we avoid lockdep false positives
261 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
262 */
263retry:
264 tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
265 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
266 sb_start_intwrite(mp->m_super);
267 xfs_trans_set_context(tp);
268
269 /*
270 * Zero-reservation ("empty") transactions can't modify anything, so
271 * they're allowed to run while we're frozen.
272 */
273 WARN_ON(resp->tr_logres > 0 &&
274 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
275 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
276 xfs_has_lazysbcount(mp));
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_highest_agno = NULLAGNUMBER;
285
286 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
287 if (error == -ENOSPC && want_retry) {
288 xfs_trans_cancel(tp);
289
290 /*
291 * We weren't able to reserve enough space for the transaction.
292 * Flush the other speculative space allocations to free space.
293 * Do not perform a synchronous scan because callers can hold
294 * other locks.
295 */
296 error = xfs_blockgc_flush_all(mp);
297 if (error)
298 return error;
299 want_retry = false;
300 goto retry;
301 }
302 if (error) {
303 xfs_trans_cancel(tp);
304 return error;
305 }
306
307 trace_xfs_trans_alloc(tp, _RET_IP_);
308
309 *tpp = tp;
310 return 0;
311}
312
313/*
314 * Create an empty transaction with no reservation. This is a defensive
315 * mechanism for routines that query metadata without actually modifying them --
316 * if the metadata being queried is somehow cross-linked (think a btree block
317 * pointer that points higher in the tree), we risk deadlock. However, blocks
318 * grabbed as part of a transaction can be re-grabbed. The verifiers will
319 * notice the corrupt block and the operation will fail back to userspace
320 * without deadlocking.
321 *
322 * Note the zero-length reservation; this transaction MUST be cancelled without
323 * any dirty data.
324 *
325 * Callers should obtain freeze protection to avoid a conflict with fs freezing
326 * where we can be grabbing buffers at the same time that freeze is trying to
327 * drain the buffer LRU list.
328 */
329int
330xfs_trans_alloc_empty(
331 struct xfs_mount *mp,
332 struct xfs_trans **tpp)
333{
334 struct xfs_trans_res resv = {0};
335
336 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
337}
338
339/*
340 * Record the indicated change to the given field for application
341 * to the file system's superblock when the transaction commits.
342 * For now, just store the change in the transaction structure.
343 *
344 * Mark the transaction structure to indicate that the superblock
345 * needs to be updated before committing.
346 *
347 * Because we may not be keeping track of allocated/free inodes and
348 * used filesystem blocks in the superblock, we do not mark the
349 * superblock dirty in this transaction if we modify these fields.
350 * We still need to update the transaction deltas so that they get
351 * applied to the incore superblock, but we don't want them to
352 * cause the superblock to get locked and logged if these are the
353 * only fields in the superblock that the transaction modifies.
354 */
355void
356xfs_trans_mod_sb(
357 xfs_trans_t *tp,
358 uint field,
359 int64_t delta)
360{
361 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
362 xfs_mount_t *mp = tp->t_mountp;
363
364 switch (field) {
365 case XFS_TRANS_SB_ICOUNT:
366 tp->t_icount_delta += delta;
367 if (xfs_has_lazysbcount(mp))
368 flags &= ~XFS_TRANS_SB_DIRTY;
369 break;
370 case XFS_TRANS_SB_IFREE:
371 tp->t_ifree_delta += delta;
372 if (xfs_has_lazysbcount(mp))
373 flags &= ~XFS_TRANS_SB_DIRTY;
374 break;
375 case XFS_TRANS_SB_FDBLOCKS:
376 /*
377 * Track the number of blocks allocated in the transaction.
378 * Make sure it does not exceed the number reserved. If so,
379 * shutdown as this can lead to accounting inconsistency.
380 */
381 if (delta < 0) {
382 tp->t_blk_res_used += (uint)-delta;
383 if (tp->t_blk_res_used > tp->t_blk_res)
384 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
385 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
386 int64_t blkres_delta;
387
388 /*
389 * Return freed blocks directly to the reservation
390 * instead of the global pool, being careful not to
391 * overflow the trans counter. This is used to preserve
392 * reservation across chains of transaction rolls that
393 * repeatedly free and allocate blocks.
394 */
395 blkres_delta = min_t(int64_t, delta,
396 UINT_MAX - tp->t_blk_res);
397 tp->t_blk_res += blkres_delta;
398 delta -= blkres_delta;
399 }
400 tp->t_fdblocks_delta += delta;
401 if (xfs_has_lazysbcount(mp))
402 flags &= ~XFS_TRANS_SB_DIRTY;
403 break;
404 case XFS_TRANS_SB_RES_FDBLOCKS:
405 /*
406 * The allocation has already been applied to the
407 * in-core superblock's counter. This should only
408 * be applied to the on-disk superblock.
409 */
410 tp->t_res_fdblocks_delta += delta;
411 if (xfs_has_lazysbcount(mp))
412 flags &= ~XFS_TRANS_SB_DIRTY;
413 break;
414 case XFS_TRANS_SB_FREXTENTS:
415 /*
416 * Track the number of blocks allocated in the
417 * transaction. Make sure it does not exceed the
418 * number reserved.
419 */
420 if (delta < 0) {
421 tp->t_rtx_res_used += (uint)-delta;
422 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
423 }
424 tp->t_frextents_delta += delta;
425 if (xfs_has_rtgroups(mp))
426 flags &= ~XFS_TRANS_SB_DIRTY;
427 break;
428 case XFS_TRANS_SB_RES_FREXTENTS:
429 /*
430 * The allocation has already been applied to the
431 * in-core superblock's counter. This should only
432 * be applied to the on-disk superblock.
433 */
434 ASSERT(delta < 0);
435 tp->t_res_frextents_delta += delta;
436 if (xfs_has_rtgroups(mp))
437 flags &= ~XFS_TRANS_SB_DIRTY;
438 break;
439 case XFS_TRANS_SB_DBLOCKS:
440 tp->t_dblocks_delta += delta;
441 break;
442 case XFS_TRANS_SB_AGCOUNT:
443 ASSERT(delta > 0);
444 tp->t_agcount_delta += delta;
445 break;
446 case XFS_TRANS_SB_IMAXPCT:
447 tp->t_imaxpct_delta += delta;
448 break;
449 case XFS_TRANS_SB_REXTSIZE:
450 tp->t_rextsize_delta += delta;
451 break;
452 case XFS_TRANS_SB_RBMBLOCKS:
453 tp->t_rbmblocks_delta += delta;
454 break;
455 case XFS_TRANS_SB_RBLOCKS:
456 tp->t_rblocks_delta += delta;
457 break;
458 case XFS_TRANS_SB_REXTENTS:
459 tp->t_rextents_delta += delta;
460 break;
461 case XFS_TRANS_SB_REXTSLOG:
462 tp->t_rextslog_delta += delta;
463 break;
464 case XFS_TRANS_SB_RGCOUNT:
465 ASSERT(delta > 0);
466 tp->t_rgcount_delta += delta;
467 break;
468 default:
469 ASSERT(0);
470 return;
471 }
472
473 tp->t_flags |= flags;
474}
475
476/*
477 * xfs_trans_apply_sb_deltas() is called from the commit code
478 * to bring the superblock buffer into the current transaction
479 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
480 *
481 * For now we just look at each field allowed to change and change
482 * it if necessary.
483 */
484STATIC void
485xfs_trans_apply_sb_deltas(
486 xfs_trans_t *tp)
487{
488 struct xfs_dsb *sbp;
489 struct xfs_buf *bp;
490 int whole = 0;
491
492 bp = xfs_trans_getsb(tp);
493 sbp = bp->b_addr;
494
495 /*
496 * Only update the superblock counters if we are logging them
497 */
498 if (!xfs_has_lazysbcount((tp->t_mountp))) {
499 if (tp->t_icount_delta)
500 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
501 if (tp->t_ifree_delta)
502 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
503 if (tp->t_fdblocks_delta)
504 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
505 if (tp->t_res_fdblocks_delta)
506 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
507 }
508
509 /*
510 * sb_frextents was added to the lazy sb counters when the rt groups
511 * feature was introduced. This is possible because we know that all
512 * kernels supporting rtgroups will also recompute frextents from the
513 * realtime bitmap.
514 *
515 * For older file systems, updating frextents requires careful handling
516 * because we cannot rely on log recovery in older kernels to recompute
517 * the value from the rtbitmap. This means that the ondisk frextents
518 * must be consistent with the rtbitmap.
519 *
520 * Therefore, log the frextents change to the ondisk superblock and
521 * update the incore superblock so that future calls to xfs_log_sb
522 * write the correct value ondisk.
523 */
524 if ((tp->t_frextents_delta || tp->t_res_frextents_delta) &&
525 !xfs_has_rtgroups(tp->t_mountp)) {
526 struct xfs_mount *mp = tp->t_mountp;
527 int64_t rtxdelta;
528
529 rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;
530
531 spin_lock(&mp->m_sb_lock);
532 be64_add_cpu(&sbp->sb_frextents, rtxdelta);
533 mp->m_sb.sb_frextents += rtxdelta;
534 spin_unlock(&mp->m_sb_lock);
535 }
536
537 if (tp->t_dblocks_delta) {
538 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
539 whole = 1;
540 }
541 if (tp->t_agcount_delta) {
542 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
543 whole = 1;
544 }
545 if (tp->t_imaxpct_delta) {
546 sbp->sb_imax_pct += tp->t_imaxpct_delta;
547 whole = 1;
548 }
549 if (tp->t_rextsize_delta) {
550 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
551
552 /*
553 * Because the ondisk sb records rtgroup size in units of rt
554 * extents, any time we update the rt extent size we have to
555 * recompute the ondisk rtgroup block log. The incore values
556 * will be recomputed in xfs_trans_unreserve_and_mod_sb.
557 */
558 if (xfs_has_rtgroups(tp->t_mountp)) {
559 sbp->sb_rgblklog = xfs_compute_rgblklog(
560 be32_to_cpu(sbp->sb_rgextents),
561 be32_to_cpu(sbp->sb_rextsize));
562 }
563 whole = 1;
564 }
565 if (tp->t_rbmblocks_delta) {
566 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
567 whole = 1;
568 }
569 if (tp->t_rblocks_delta) {
570 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
571 whole = 1;
572 }
573 if (tp->t_rextents_delta) {
574 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
575 whole = 1;
576 }
577 if (tp->t_rextslog_delta) {
578 sbp->sb_rextslog += tp->t_rextslog_delta;
579 whole = 1;
580 }
581 if (tp->t_rgcount_delta) {
582 be32_add_cpu(&sbp->sb_rgcount, tp->t_rgcount_delta);
583 whole = 1;
584 }
585
586 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
587 if (whole)
588 /*
589 * Log the whole thing, the fields are noncontiguous.
590 */
591 xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
592 else
593 /*
594 * Since all the modifiable fields are contiguous, we
595 * can get away with this.
596 */
597 xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
598 offsetof(struct xfs_dsb, sb_frextents) +
599 sizeof(sbp->sb_frextents) - 1);
600}
601
602/*
603 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
604 * apply superblock counter changes to the in-core superblock. The
605 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
606 * applied to the in-core superblock. The idea is that that has already been
607 * done.
608 *
609 * If we are not logging superblock counters, then the inode allocated/free and
610 * used block counts are not updated in the on disk superblock. In this case,
611 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
612 * still need to update the incore superblock with the changes.
613 *
614 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
615 * so we don't need to take the counter lock on every update.
616 */
617#define XFS_ICOUNT_BATCH 128
618
619void
620xfs_trans_unreserve_and_mod_sb(
621 struct xfs_trans *tp)
622{
623 struct xfs_mount *mp = tp->t_mountp;
624 int64_t blkdelta = tp->t_blk_res;
625 int64_t rtxdelta = tp->t_rtx_res;
626 int64_t idelta = 0;
627 int64_t ifreedelta = 0;
628
629 /*
630 * Calculate the deltas.
631 *
632 * t_fdblocks_delta and t_frextents_delta can be positive or negative:
633 *
634 * - positive values indicate blocks freed in the transaction.
635 * - negative values indicate blocks allocated in the transaction
636 *
637 * Negative values can only happen if the transaction has a block
638 * reservation that covers the allocated block. The end result is
639 * that the calculated delta values must always be positive and we
640 * can only put back previous allocated or reserved blocks here.
641 */
642 ASSERT(tp->t_blk_res || tp->t_fdblocks_delta >= 0);
643 if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
644 blkdelta += tp->t_fdblocks_delta;
645 ASSERT(blkdelta >= 0);
646 }
647
648 ASSERT(tp->t_rtx_res || tp->t_frextents_delta >= 0);
649 if (xfs_has_rtgroups(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
650 rtxdelta += tp->t_frextents_delta;
651 ASSERT(rtxdelta >= 0);
652 }
653
654 if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
655 idelta = tp->t_icount_delta;
656 ifreedelta = tp->t_ifree_delta;
657 }
658
659 /* apply the per-cpu counters */
660 if (blkdelta)
661 xfs_add_fdblocks(mp, blkdelta);
662
663 if (idelta)
664 percpu_counter_add_batch(&mp->m_icount, idelta,
665 XFS_ICOUNT_BATCH);
666
667 if (ifreedelta)
668 percpu_counter_add(&mp->m_ifree, ifreedelta);
669
670 if (rtxdelta)
671 xfs_add_frextents(mp, rtxdelta);
672
673 if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))
674 return;
675
676 /* apply remaining deltas */
677 spin_lock(&mp->m_sb_lock);
678 mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
679 mp->m_sb.sb_icount += idelta;
680 mp->m_sb.sb_ifree += ifreedelta;
681 /*
682 * Do not touch sb_frextents here because it is handled in
683 * xfs_trans_apply_sb_deltas for file systems where it isn't a lazy
684 * counter anyway.
685 */
686 mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
687 mp->m_sb.sb_agcount += tp->t_agcount_delta;
688 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
689 if (tp->t_rextsize_delta)
690 xfs_mount_sb_set_rextsize(mp, &mp->m_sb,
691 mp->m_sb.sb_rextsize + tp->t_rextsize_delta);
692 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
693 mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
694 mp->m_sb.sb_rextents += tp->t_rextents_delta;
695 mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
696 mp->m_sb.sb_rgcount += tp->t_rgcount_delta;
697 spin_unlock(&mp->m_sb_lock);
698
699 /*
700 * Debug checks outside of the spinlock so they don't lock up the
701 * machine if they fail.
702 */
703 ASSERT(mp->m_sb.sb_imax_pct >= 0);
704 ASSERT(mp->m_sb.sb_rextslog >= 0);
705}
706
707/* Add the given log item to the transaction's list of log items. */
708void
709xfs_trans_add_item(
710 struct xfs_trans *tp,
711 struct xfs_log_item *lip)
712{
713 ASSERT(lip->li_log == tp->t_mountp->m_log);
714 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
715 ASSERT(list_empty(&lip->li_trans));
716 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
717
718 list_add_tail(&lip->li_trans, &tp->t_items);
719 trace_xfs_trans_add_item(tp, _RET_IP_);
720}
721
722/*
723 * Unlink the log item from the transaction. the log item is no longer
724 * considered dirty in this transaction, as the linked transaction has
725 * finished, either by abort or commit completion.
726 */
727void
728xfs_trans_del_item(
729 struct xfs_log_item *lip)
730{
731 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
732 list_del_init(&lip->li_trans);
733}
734
735/* Detach and unlock all of the items in a transaction */
736static void
737xfs_trans_free_items(
738 struct xfs_trans *tp,
739 bool abort)
740{
741 struct xfs_log_item *lip, *next;
742
743 trace_xfs_trans_free_items(tp, _RET_IP_);
744
745 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
746 xfs_trans_del_item(lip);
747 if (abort)
748 set_bit(XFS_LI_ABORTED, &lip->li_flags);
749 if (lip->li_ops->iop_release)
750 lip->li_ops->iop_release(lip);
751 }
752}
753
754/*
755 * Sort transaction items prior to running precommit operations. This will
756 * attempt to order the items such that they will always be locked in the same
757 * order. Items that have no sort function are moved to the end of the list
758 * and so are locked last.
759 *
760 * This may need refinement as different types of objects add sort functions.
761 *
762 * Function is more complex than it needs to be because we are comparing 64 bit
763 * values and the function only returns 32 bit values.
764 */
765static int
766xfs_trans_precommit_sort(
767 void *unused_arg,
768 const struct list_head *a,
769 const struct list_head *b)
770{
771 struct xfs_log_item *lia = container_of(a,
772 struct xfs_log_item, li_trans);
773 struct xfs_log_item *lib = container_of(b,
774 struct xfs_log_item, li_trans);
775 int64_t diff;
776
777 /*
778 * If both items are non-sortable, leave them alone. If only one is
779 * sortable, move the non-sortable item towards the end of the list.
780 */
781 if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)
782 return 0;
783 if (!lia->li_ops->iop_sort)
784 return 1;
785 if (!lib->li_ops->iop_sort)
786 return -1;
787
788 diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);
789 if (diff < 0)
790 return -1;
791 if (diff > 0)
792 return 1;
793 return 0;
794}
795
796/*
797 * Run transaction precommit functions.
798 *
799 * If there is an error in any of the callouts, then stop immediately and
800 * trigger a shutdown to abort the transaction. There is no recovery possible
801 * from errors at this point as the transaction is dirty....
802 */
803static int
804xfs_trans_run_precommits(
805 struct xfs_trans *tp)
806{
807 struct xfs_mount *mp = tp->t_mountp;
808 struct xfs_log_item *lip, *n;
809 int error = 0;
810
811 /*
812 * Sort the item list to avoid ABBA deadlocks with other transactions
813 * running precommit operations that lock multiple shared items such as
814 * inode cluster buffers.
815 */
816 list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);
817
818 /*
819 * Precommit operations can remove the log item from the transaction
820 * if the log item exists purely to delay modifications until they
821 * can be ordered against other operations. Hence we have to use
822 * list_for_each_entry_safe() here.
823 */
824 list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {
825 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
826 continue;
827 if (lip->li_ops->iop_precommit) {
828 error = lip->li_ops->iop_precommit(tp, lip);
829 if (error)
830 break;
831 }
832 }
833 if (error)
834 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
835 return error;
836}
837
838/*
839 * Commit the given transaction to the log.
840 *
841 * XFS disk error handling mechanism is not based on a typical
842 * transaction abort mechanism. Logically after the filesystem
843 * gets marked 'SHUTDOWN', we can't let any new transactions
844 * be durable - ie. committed to disk - because some metadata might
845 * be inconsistent. In such cases, this returns an error, and the
846 * caller may assume that all locked objects joined to the transaction
847 * have already been unlocked as if the commit had succeeded.
848 * Do not reference the transaction structure after this call.
849 */
850static int
851__xfs_trans_commit(
852 struct xfs_trans *tp,
853 bool regrant)
854{
855 struct xfs_mount *mp = tp->t_mountp;
856 struct xlog *log = mp->m_log;
857 xfs_csn_t commit_seq = 0;
858 int error = 0;
859 int sync = tp->t_flags & XFS_TRANS_SYNC;
860
861 trace_xfs_trans_commit(tp, _RET_IP_);
862
863 /*
864 * Commit per-transaction changes that are not already tracked through
865 * log items. This can add dirty log items to the transaction.
866 */
867 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
868 xfs_trans_apply_sb_deltas(tp);
869 xfs_trans_apply_dquot_deltas(tp);
870
871 error = xfs_trans_run_precommits(tp);
872 if (error)
873 goto out_unreserve;
874
875 /*
876 * If there is nothing to be logged by the transaction,
877 * then unlock all of the items associated with the
878 * transaction and free the transaction structure.
879 * Also make sure to return any reserved blocks to
880 * the free pool.
881 */
882 if (!(tp->t_flags & XFS_TRANS_DIRTY))
883 goto out_unreserve;
884
885 /*
886 * We must check against log shutdown here because we cannot abort log
887 * items and leave them dirty, inconsistent and unpinned in memory while
888 * the log is active. This leaves them open to being written back to
889 * disk, and that will lead to on-disk corruption.
890 */
891 if (xlog_is_shutdown(log)) {
892 error = -EIO;
893 goto out_unreserve;
894 }
895
896 ASSERT(tp->t_ticket != NULL);
897
898 xlog_cil_commit(log, tp, &commit_seq, regrant);
899
900 xfs_trans_free(tp);
901
902 /*
903 * If the transaction needs to be synchronous, then force the
904 * log out now and wait for it.
905 */
906 if (sync) {
907 error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
908 XFS_STATS_INC(mp, xs_trans_sync);
909 } else {
910 XFS_STATS_INC(mp, xs_trans_async);
911 }
912
913 return error;
914
915out_unreserve:
916 xfs_trans_unreserve_and_mod_sb(tp);
917
918 /*
919 * It is indeed possible for the transaction to be not dirty but
920 * the dqinfo portion to be. All that means is that we have some
921 * (non-persistent) quota reservations that need to be unreserved.
922 */
923 xfs_trans_unreserve_and_mod_dquots(tp, true);
924 if (tp->t_ticket) {
925 if (regrant && !xlog_is_shutdown(log))
926 xfs_log_ticket_regrant(log, tp->t_ticket);
927 else
928 xfs_log_ticket_ungrant(log, tp->t_ticket);
929 tp->t_ticket = NULL;
930 }
931 xfs_trans_free_items(tp, !!error);
932 xfs_trans_free(tp);
933
934 XFS_STATS_INC(mp, xs_trans_empty);
935 return error;
936}
937
938int
939xfs_trans_commit(
940 struct xfs_trans *tp)
941{
942 /*
943 * Finish deferred items on final commit. Only permanent transactions
944 * should ever have deferred ops.
945 */
946 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
947 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
948 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES) {
949 int error = xfs_defer_finish_noroll(&tp);
950 if (error) {
951 xfs_trans_cancel(tp);
952 return error;
953 }
954 }
955
956 return __xfs_trans_commit(tp, false);
957}
958
959/*
960 * Unlock all of the transaction's items and free the transaction. If the
961 * transaction is dirty, we must shut down the filesystem because there is no
962 * way to restore them to their previous state.
963 *
964 * If the transaction has made a log reservation, make sure to release it as
965 * well.
966 *
967 * This is a high level function (equivalent to xfs_trans_commit()) and so can
968 * be called after the transaction has effectively been aborted due to the mount
969 * being shut down. However, if the mount has not been shut down and the
970 * transaction is dirty we will shut the mount down and, in doing so, that
971 * guarantees that the log is shut down, too. Hence we don't need to be as
972 * careful with shutdown state and dirty items here as we need to be in
973 * xfs_trans_commit().
974 */
975void
976xfs_trans_cancel(
977 struct xfs_trans *tp)
978{
979 struct xfs_mount *mp = tp->t_mountp;
980 struct xlog *log = mp->m_log;
981 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
982
983 trace_xfs_trans_cancel(tp, _RET_IP_);
984
985 /*
986 * It's never valid to cancel a transaction with deferred ops attached,
987 * because the transaction is effectively dirty. Complain about this
988 * loudly before freeing the in-memory defer items and shutting down the
989 * filesystem.
990 */
991 if (!list_empty(&tp->t_dfops)) {
992 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
993 dirty = true;
994 xfs_defer_cancel(tp);
995 }
996
997 /*
998 * See if the caller is relying on us to shut down the filesystem. We
999 * only want an error report if there isn't already a shutdown in
1000 * progress, so we only need to check against the mount shutdown state
1001 * here.
1002 */
1003 if (dirty && !xfs_is_shutdown(mp)) {
1004 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1005 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1006 }
1007#ifdef DEBUG
1008 /* Log items need to be consistent until the log is shut down. */
1009 if (!dirty && !xlog_is_shutdown(log)) {
1010 struct xfs_log_item *lip;
1011
1012 list_for_each_entry(lip, &tp->t_items, li_trans)
1013 ASSERT(!xlog_item_is_intent_done(lip));
1014 }
1015#endif
1016 xfs_trans_unreserve_and_mod_sb(tp);
1017 xfs_trans_unreserve_and_mod_dquots(tp, false);
1018
1019 if (tp->t_ticket) {
1020 xfs_log_ticket_ungrant(log, tp->t_ticket);
1021 tp->t_ticket = NULL;
1022 }
1023
1024 xfs_trans_free_items(tp, dirty);
1025 xfs_trans_free(tp);
1026}
1027
1028/*
1029 * Roll from one trans in the sequence of PERMANENT transactions to
1030 * the next: permanent transactions are only flushed out when
1031 * committed with xfs_trans_commit(), but we still want as soon
1032 * as possible to let chunks of it go to the log. So we commit the
1033 * chunk we've been working on and get a new transaction to continue.
1034 */
1035int
1036xfs_trans_roll(
1037 struct xfs_trans **tpp)
1038{
1039 struct xfs_trans *trans = *tpp;
1040 struct xfs_trans_res tres;
1041 int error;
1042
1043 trace_xfs_trans_roll(trans, _RET_IP_);
1044
1045 /*
1046 * Copy the critical parameters from one trans to the next.
1047 */
1048 tres.tr_logres = trans->t_log_res;
1049 tres.tr_logcount = trans->t_log_count;
1050
1051 *tpp = xfs_trans_dup(trans);
1052
1053 /*
1054 * Commit the current transaction.
1055 * If this commit failed, then it'd just unlock those items that
1056 * are not marked ihold. That also means that a filesystem shutdown
1057 * is in progress. The caller takes the responsibility to cancel
1058 * the duplicate transaction that gets returned.
1059 */
1060 error = __xfs_trans_commit(trans, true);
1061 if (error)
1062 return error;
1063
1064 /*
1065 * Reserve space in the log for the next transaction.
1066 * This also pushes items in the "AIL", the list of logged items,
1067 * out to disk if they are taking up space at the tail of the log
1068 * that we want to use. This requires that either nothing be locked
1069 * across this call, or that anything that is locked be logged in
1070 * the prior and the next transactions.
1071 */
1072 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1073 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1074}
1075
1076/*
1077 * Allocate an transaction, lock and join the inode to it, and reserve quota.
1078 *
1079 * The caller must ensure that the on-disk dquots attached to this inode have
1080 * already been allocated and initialized. The caller is responsible for
1081 * releasing ILOCK_EXCL if a new transaction is returned.
1082 */
1083int
1084xfs_trans_alloc_inode(
1085 struct xfs_inode *ip,
1086 struct xfs_trans_res *resv,
1087 unsigned int dblocks,
1088 unsigned int rblocks,
1089 bool force,
1090 struct xfs_trans **tpp)
1091{
1092 struct xfs_trans *tp;
1093 struct xfs_mount *mp = ip->i_mount;
1094 bool retried = false;
1095 int error;
1096
1097retry:
1098 error = xfs_trans_alloc(mp, resv, dblocks,
1099 xfs_extlen_to_rtxlen(mp, rblocks),
1100 force ? XFS_TRANS_RESERVE : 0, &tp);
1101 if (error)
1102 return error;
1103
1104 xfs_ilock(ip, XFS_ILOCK_EXCL);
1105 xfs_trans_ijoin(tp, ip, 0);
1106
1107 error = xfs_qm_dqattach_locked(ip, false);
1108 if (error) {
1109 /* Caller should have allocated the dquots! */
1110 ASSERT(error != -ENOENT);
1111 goto out_cancel;
1112 }
1113
1114 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1115 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1116 xfs_trans_cancel(tp);
1117 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1118 xfs_blockgc_free_quota(ip, 0);
1119 retried = true;
1120 goto retry;
1121 }
1122 if (error)
1123 goto out_cancel;
1124
1125 *tpp = tp;
1126 return 0;
1127
1128out_cancel:
1129 xfs_trans_cancel(tp);
1130 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1131 return error;
1132}
1133
1134/*
1135 * Try to reserve more blocks for a transaction.
1136 *
1137 * This is for callers that need to attach resources to a transaction, scan
1138 * those resources to determine the space reservation requirements, and then
1139 * modify the attached resources. In other words, online repair. This can
1140 * fail due to ENOSPC, so the caller must be able to cancel the transaction
1141 * without shutting down the fs.
1142 */
1143int
1144xfs_trans_reserve_more(
1145 struct xfs_trans *tp,
1146 unsigned int blocks,
1147 unsigned int rtextents)
1148{
1149 struct xfs_trans_res resv = { };
1150
1151 return xfs_trans_reserve(tp, &resv, blocks, rtextents);
1152}
1153
1154/*
1155 * Try to reserve more blocks and file quota for a transaction. Same
1156 * conditions of usage as xfs_trans_reserve_more.
1157 */
1158int
1159xfs_trans_reserve_more_inode(
1160 struct xfs_trans *tp,
1161 struct xfs_inode *ip,
1162 unsigned int dblocks,
1163 unsigned int rblocks,
1164 bool force_quota)
1165{
1166 struct xfs_trans_res resv = { };
1167 struct xfs_mount *mp = ip->i_mount;
1168 unsigned int rtx = xfs_extlen_to_rtxlen(mp, rblocks);
1169 int error;
1170
1171 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1172
1173 error = xfs_trans_reserve(tp, &resv, dblocks, rtx);
1174 if (error)
1175 return error;
1176
1177 if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino))
1178 return 0;
1179
1180 if (tp->t_flags & XFS_TRANS_RESERVE)
1181 force_quota = true;
1182
1183 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,
1184 force_quota);
1185 if (!error)
1186 return 0;
1187
1188 /* Quota failed, give back the new reservation. */
1189 xfs_add_fdblocks(mp, dblocks);
1190 tp->t_blk_res -= dblocks;
1191 xfs_add_frextents(mp, rtx);
1192 tp->t_rtx_res -= rtx;
1193 return error;
1194}
1195
1196/*
1197 * Allocate an transaction in preparation for inode creation by reserving quota
1198 * against the given dquots. Callers are not required to hold any inode locks.
1199 */
1200int
1201xfs_trans_alloc_icreate(
1202 struct xfs_mount *mp,
1203 struct xfs_trans_res *resv,
1204 struct xfs_dquot *udqp,
1205 struct xfs_dquot *gdqp,
1206 struct xfs_dquot *pdqp,
1207 unsigned int dblocks,
1208 struct xfs_trans **tpp)
1209{
1210 struct xfs_trans *tp;
1211 bool retried = false;
1212 int error;
1213
1214retry:
1215 error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1216 if (error)
1217 return error;
1218
1219 error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1220 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1221 xfs_trans_cancel(tp);
1222 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1223 retried = true;
1224 goto retry;
1225 }
1226 if (error) {
1227 xfs_trans_cancel(tp);
1228 return error;
1229 }
1230
1231 *tpp = tp;
1232 return 0;
1233}
1234
1235/*
1236 * Allocate an transaction, lock and join the inode to it, and reserve quota
1237 * in preparation for inode attribute changes that include uid, gid, or prid
1238 * changes.
1239 *
1240 * The caller must ensure that the on-disk dquots attached to this inode have
1241 * already been allocated and initialized. The ILOCK will be dropped when the
1242 * transaction is committed or cancelled.
1243 */
1244int
1245xfs_trans_alloc_ichange(
1246 struct xfs_inode *ip,
1247 struct xfs_dquot *new_udqp,
1248 struct xfs_dquot *new_gdqp,
1249 struct xfs_dquot *new_pdqp,
1250 bool force,
1251 struct xfs_trans **tpp)
1252{
1253 struct xfs_trans *tp;
1254 struct xfs_mount *mp = ip->i_mount;
1255 struct xfs_dquot *udqp;
1256 struct xfs_dquot *gdqp;
1257 struct xfs_dquot *pdqp;
1258 bool retried = false;
1259 int error;
1260
1261retry:
1262 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1263 if (error)
1264 return error;
1265
1266 xfs_ilock(ip, XFS_ILOCK_EXCL);
1267 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1268
1269 error = xfs_qm_dqattach_locked(ip, false);
1270 if (error) {
1271 /* Caller should have allocated the dquots! */
1272 ASSERT(error != -ENOENT);
1273 goto out_cancel;
1274 }
1275
1276 /*
1277 * For each quota type, skip quota reservations if the inode's dquots
1278 * now match the ones that came from the caller, or the caller didn't
1279 * pass one in. The inode's dquots can change if we drop the ILOCK to
1280 * perform a blockgc scan, so we must preserve the caller's arguments.
1281 */
1282 udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1283 gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1284 pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1285 if (udqp || gdqp || pdqp) {
1286 xfs_filblks_t dblocks, rblocks;
1287 unsigned int qflags = XFS_QMOPT_RES_REGBLKS;
1288 bool isrt = XFS_IS_REALTIME_INODE(ip);
1289
1290 if (force)
1291 qflags |= XFS_QMOPT_FORCE_RES;
1292
1293 if (isrt) {
1294 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1295 if (error)
1296 goto out_cancel;
1297 }
1298
1299 xfs_inode_count_blocks(tp, ip, &dblocks, &rblocks);
1300
1301 if (isrt)
1302 rblocks += ip->i_delayed_blks;
1303 else
1304 dblocks += ip->i_delayed_blks;
1305
1306 /*
1307 * Reserve enough quota to handle blocks on disk and reserved
1308 * for a delayed allocation. We'll actually transfer the
1309 * delalloc reservation between dquots at chown time, even
1310 * though that part is only semi-transactional.
1311 */
1312 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1313 pdqp, dblocks, 1, qflags);
1314 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1315 xfs_trans_cancel(tp);
1316 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1317 retried = true;
1318 goto retry;
1319 }
1320 if (error)
1321 goto out_cancel;
1322
1323 /* Do the same for realtime. */
1324 qflags = XFS_QMOPT_RES_RTBLKS | (qflags & XFS_QMOPT_FORCE_RES);
1325 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1326 pdqp, rblocks, 0, qflags);
1327 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1328 xfs_trans_cancel(tp);
1329 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1330 retried = true;
1331 goto retry;
1332 }
1333 if (error)
1334 goto out_cancel;
1335 }
1336
1337 *tpp = tp;
1338 return 0;
1339
1340out_cancel:
1341 xfs_trans_cancel(tp);
1342 return error;
1343}
1344
1345/*
1346 * Allocate an transaction, lock and join the directory and child inodes to it,
1347 * and reserve quota for a directory update. If there isn't sufficient space,
1348 * @dblocks will be set to zero for a reservationless directory update and
1349 * @nospace_error will be set to a negative errno describing the space
1350 * constraint we hit.
1351 *
1352 * The caller must ensure that the on-disk dquots attached to this inode have
1353 * already been allocated and initialized. The ILOCKs will be dropped when the
1354 * transaction is committed or cancelled.
1355 *
1356 * Caller is responsible for unlocking the inodes manually upon return
1357 */
1358int
1359xfs_trans_alloc_dir(
1360 struct xfs_inode *dp,
1361 struct xfs_trans_res *resv,
1362 struct xfs_inode *ip,
1363 unsigned int *dblocks,
1364 struct xfs_trans **tpp,
1365 int *nospace_error)
1366{
1367 struct xfs_trans *tp;
1368 struct xfs_mount *mp = ip->i_mount;
1369 unsigned int resblks;
1370 bool retried = false;
1371 int error;
1372
1373retry:
1374 *nospace_error = 0;
1375 resblks = *dblocks;
1376 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1377 if (error == -ENOSPC) {
1378 *nospace_error = error;
1379 resblks = 0;
1380 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1381 }
1382 if (error)
1383 return error;
1384
1385 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
1386
1387 xfs_trans_ijoin(tp, dp, 0);
1388 xfs_trans_ijoin(tp, ip, 0);
1389
1390 error = xfs_qm_dqattach_locked(dp, false);
1391 if (error) {
1392 /* Caller should have allocated the dquots! */
1393 ASSERT(error != -ENOENT);
1394 goto out_cancel;
1395 }
1396
1397 error = xfs_qm_dqattach_locked(ip, false);
1398 if (error) {
1399 /* Caller should have allocated the dquots! */
1400 ASSERT(error != -ENOENT);
1401 goto out_cancel;
1402 }
1403
1404 if (resblks == 0)
1405 goto done;
1406
1407 error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);
1408 if (error == -EDQUOT || error == -ENOSPC) {
1409 if (!retried) {
1410 xfs_trans_cancel(tp);
1411 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1412 if (dp != ip)
1413 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1414 xfs_blockgc_free_quota(dp, 0);
1415 retried = true;
1416 goto retry;
1417 }
1418
1419 *nospace_error = error;
1420 resblks = 0;
1421 error = 0;
1422 }
1423 if (error)
1424 goto out_cancel;
1425
1426done:
1427 *tpp = tp;
1428 *dblocks = resblks;
1429 return 0;
1430
1431out_cancel:
1432 xfs_trans_cancel(tp);
1433 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1434 if (dp != ip)
1435 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1436 return error;
1437}