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
29struct kmem_cache *xfs_trans_cache;
30
31#if defined(CONFIG_TRACEPOINTS)
32static void
33xfs_trans_trace_reservations(
34 struct xfs_mount *mp)
35{
36 struct xfs_trans_res *res;
37 struct xfs_trans_res *end_res;
38 int i;
39
40 res = (struct xfs_trans_res *)M_RES(mp);
41 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
42 for (i = 0; res < end_res; i++, res++)
43 trace_xfs_trans_resv_calc(mp, i, res);
44}
45#else
46# define xfs_trans_trace_reservations(mp)
47#endif
48
49/*
50 * Initialize the precomputed transaction reservation values
51 * in the mount structure.
52 */
53void
54xfs_trans_init(
55 struct xfs_mount *mp)
56{
57 xfs_trans_resv_calc(mp, M_RES(mp));
58 xfs_trans_trace_reservations(mp);
59}
60
61/*
62 * Free the transaction structure. If there is more clean up
63 * to do when the structure is freed, add it here.
64 */
65STATIC void
66xfs_trans_free(
67 struct xfs_trans *tp)
68{
69 xfs_extent_busy_sort(&tp->t_busy);
70 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
71
72 trace_xfs_trans_free(tp, _RET_IP_);
73 xfs_trans_clear_context(tp);
74 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
75 sb_end_intwrite(tp->t_mountp->m_super);
76 xfs_trans_free_dqinfo(tp);
77 kmem_cache_free(xfs_trans_cache, tp);
78}
79
80/*
81 * This is called to create a new transaction which will share the
82 * permanent log reservation of the given transaction. The remaining
83 * unused block and rt extent reservations are also inherited. This
84 * implies that the original transaction is no longer allowed to allocate
85 * blocks. Locks and log items, however, are no inherited. They must
86 * be added to the new transaction explicitly.
87 */
88STATIC struct xfs_trans *
89xfs_trans_dup(
90 struct xfs_trans *tp)
91{
92 struct xfs_trans *ntp;
93
94 trace_xfs_trans_dup(tp, _RET_IP_);
95
96 ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
97
98 /*
99 * Initialize the new transaction structure.
100 */
101 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
102 ntp->t_mountp = tp->t_mountp;
103 INIT_LIST_HEAD(&ntp->t_items);
104 INIT_LIST_HEAD(&ntp->t_busy);
105 INIT_LIST_HEAD(&ntp->t_dfops);
106 ntp->t_highest_agno = NULLAGNUMBER;
107
108 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
109 ASSERT(tp->t_ticket != NULL);
110
111 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
112 (tp->t_flags & XFS_TRANS_RESERVE) |
113 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
114 (tp->t_flags & XFS_TRANS_RES_FDBLKS);
115 /* We gave our writer reference to the new transaction */
116 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
117 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
118
119 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
120 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
121 tp->t_blk_res = tp->t_blk_res_used;
122
123 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
124 tp->t_rtx_res = tp->t_rtx_res_used;
125
126 xfs_trans_switch_context(tp, ntp);
127
128 /* move deferred ops over to the new tp */
129 xfs_defer_move(ntp, tp);
130
131 xfs_trans_dup_dqinfo(tp, ntp);
132 return ntp;
133}
134
135/*
136 * This is called to reserve free disk blocks and log space for the
137 * given transaction. This must be done before allocating any resources
138 * within the transaction.
139 *
140 * This will return ENOSPC if there are not enough blocks available.
141 * It will sleep waiting for available log space.
142 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
143 * is used by long running transactions. If any one of the reservations
144 * fails then they will all be backed out.
145 *
146 * This does not do quota reservations. That typically is done by the
147 * caller afterwards.
148 */
149static int
150xfs_trans_reserve(
151 struct xfs_trans *tp,
152 struct xfs_trans_res *resp,
153 uint blocks,
154 uint rtextents)
155{
156 struct xfs_mount *mp = tp->t_mountp;
157 int error = 0;
158 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
159
160 /*
161 * Attempt to reserve the needed disk blocks by decrementing
162 * the number needed from the number available. This will
163 * fail if the count would go below zero.
164 */
165 if (blocks > 0) {
166 error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
167 if (error != 0)
168 return -ENOSPC;
169 tp->t_blk_res += blocks;
170 }
171
172 /*
173 * Reserve the log space needed for this transaction.
174 */
175 if (resp->tr_logres > 0) {
176 bool permanent = false;
177
178 ASSERT(tp->t_log_res == 0 ||
179 tp->t_log_res == resp->tr_logres);
180 ASSERT(tp->t_log_count == 0 ||
181 tp->t_log_count == resp->tr_logcount);
182
183 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
184 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
185 permanent = true;
186 } else {
187 ASSERT(tp->t_ticket == NULL);
188 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
189 }
190
191 if (tp->t_ticket != NULL) {
192 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
193 error = xfs_log_regrant(mp, tp->t_ticket);
194 } else {
195 error = xfs_log_reserve(mp, resp->tr_logres,
196 resp->tr_logcount,
197 &tp->t_ticket, permanent);
198 }
199
200 if (error)
201 goto undo_blocks;
202
203 tp->t_log_res = resp->tr_logres;
204 tp->t_log_count = resp->tr_logcount;
205 }
206
207 /*
208 * Attempt to reserve the needed realtime extents by decrementing
209 * the number needed from the number available. This will
210 * fail if the count would go below zero.
211 */
212 if (rtextents > 0) {
213 error = xfs_mod_frextents(mp, -((int64_t)rtextents));
214 if (error) {
215 error = -ENOSPC;
216 goto undo_log;
217 }
218 tp->t_rtx_res += rtextents;
219 }
220
221 return 0;
222
223 /*
224 * Error cases jump to one of these labels to undo any
225 * reservations which have already been performed.
226 */
227undo_log:
228 if (resp->tr_logres > 0) {
229 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
230 tp->t_ticket = NULL;
231 tp->t_log_res = 0;
232 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
233 }
234
235undo_blocks:
236 if (blocks > 0) {
237 xfs_mod_fdblocks(mp, (int64_t)blocks, rsvd);
238 tp->t_blk_res = 0;
239 }
240 return error;
241}
242
243int
244xfs_trans_alloc(
245 struct xfs_mount *mp,
246 struct xfs_trans_res *resp,
247 uint blocks,
248 uint rtextents,
249 uint flags,
250 struct xfs_trans **tpp)
251{
252 struct xfs_trans *tp;
253 bool want_retry = true;
254 int error;
255
256 /*
257 * Allocate the handle before we do our freeze accounting and setting up
258 * GFP_NOFS allocation context so that we avoid lockdep false positives
259 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
260 */
261retry:
262 tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
263 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
264 sb_start_intwrite(mp->m_super);
265 xfs_trans_set_context(tp);
266
267 /*
268 * Zero-reservation ("empty") transactions can't modify anything, so
269 * they're allowed to run while we're frozen.
270 */
271 WARN_ON(resp->tr_logres > 0 &&
272 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
273 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
274 xfs_has_lazysbcount(mp));
275
276 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
277 tp->t_flags = flags;
278 tp->t_mountp = mp;
279 INIT_LIST_HEAD(&tp->t_items);
280 INIT_LIST_HEAD(&tp->t_busy);
281 INIT_LIST_HEAD(&tp->t_dfops);
282 tp->t_highest_agno = NULLAGNUMBER;
283
284 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
285 if (error == -ENOSPC && want_retry) {
286 xfs_trans_cancel(tp);
287
288 /*
289 * We weren't able to reserve enough space for the transaction.
290 * Flush the other speculative space allocations to free space.
291 * Do not perform a synchronous scan because callers can hold
292 * other locks.
293 */
294 error = xfs_blockgc_flush_all(mp);
295 if (error)
296 return error;
297 want_retry = false;
298 goto retry;
299 }
300 if (error) {
301 xfs_trans_cancel(tp);
302 return error;
303 }
304
305 trace_xfs_trans_alloc(tp, _RET_IP_);
306
307 *tpp = tp;
308 return 0;
309}
310
311/*
312 * Create an empty transaction with no reservation. This is a defensive
313 * mechanism for routines that query metadata without actually modifying them --
314 * if the metadata being queried is somehow cross-linked (think a btree block
315 * pointer that points higher in the tree), we risk deadlock. However, blocks
316 * grabbed as part of a transaction can be re-grabbed. The verifiers will
317 * notice the corrupt block and the operation will fail back to userspace
318 * without deadlocking.
319 *
320 * Note the zero-length reservation; this transaction MUST be cancelled without
321 * any dirty data.
322 *
323 * Callers should obtain freeze protection to avoid a conflict with fs freezing
324 * where we can be grabbing buffers at the same time that freeze is trying to
325 * drain the buffer LRU list.
326 */
327int
328xfs_trans_alloc_empty(
329 struct xfs_mount *mp,
330 struct xfs_trans **tpp)
331{
332 struct xfs_trans_res resv = {0};
333
334 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
335}
336
337/*
338 * Record the indicated change to the given field for application
339 * to the file system's superblock when the transaction commits.
340 * For now, just store the change in the transaction structure.
341 *
342 * Mark the transaction structure to indicate that the superblock
343 * needs to be updated before committing.
344 *
345 * Because we may not be keeping track of allocated/free inodes and
346 * used filesystem blocks in the superblock, we do not mark the
347 * superblock dirty in this transaction if we modify these fields.
348 * We still need to update the transaction deltas so that they get
349 * applied to the incore superblock, but we don't want them to
350 * cause the superblock to get locked and logged if these are the
351 * only fields in the superblock that the transaction modifies.
352 */
353void
354xfs_trans_mod_sb(
355 xfs_trans_t *tp,
356 uint field,
357 int64_t delta)
358{
359 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
360 xfs_mount_t *mp = tp->t_mountp;
361
362 switch (field) {
363 case XFS_TRANS_SB_ICOUNT:
364 tp->t_icount_delta += delta;
365 if (xfs_has_lazysbcount(mp))
366 flags &= ~XFS_TRANS_SB_DIRTY;
367 break;
368 case XFS_TRANS_SB_IFREE:
369 tp->t_ifree_delta += delta;
370 if (xfs_has_lazysbcount(mp))
371 flags &= ~XFS_TRANS_SB_DIRTY;
372 break;
373 case XFS_TRANS_SB_FDBLOCKS:
374 /*
375 * Track the number of blocks allocated in the transaction.
376 * Make sure it does not exceed the number reserved. If so,
377 * shutdown as this can lead to accounting inconsistency.
378 */
379 if (delta < 0) {
380 tp->t_blk_res_used += (uint)-delta;
381 if (tp->t_blk_res_used > tp->t_blk_res)
382 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
383 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
384 int64_t blkres_delta;
385
386 /*
387 * Return freed blocks directly to the reservation
388 * instead of the global pool, being careful not to
389 * overflow the trans counter. This is used to preserve
390 * reservation across chains of transaction rolls that
391 * repeatedly free and allocate blocks.
392 */
393 blkres_delta = min_t(int64_t, delta,
394 UINT_MAX - tp->t_blk_res);
395 tp->t_blk_res += blkres_delta;
396 delta -= blkres_delta;
397 }
398 tp->t_fdblocks_delta += delta;
399 if (xfs_has_lazysbcount(mp))
400 flags &= ~XFS_TRANS_SB_DIRTY;
401 break;
402 case XFS_TRANS_SB_RES_FDBLOCKS:
403 /*
404 * The allocation has already been applied to the
405 * in-core superblock's counter. This should only
406 * be applied to the on-disk superblock.
407 */
408 tp->t_res_fdblocks_delta += delta;
409 if (xfs_has_lazysbcount(mp))
410 flags &= ~XFS_TRANS_SB_DIRTY;
411 break;
412 case XFS_TRANS_SB_FREXTENTS:
413 /*
414 * Track the number of blocks allocated in the
415 * transaction. Make sure it does not exceed the
416 * number reserved.
417 */
418 if (delta < 0) {
419 tp->t_rtx_res_used += (uint)-delta;
420 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
421 }
422 tp->t_frextents_delta += delta;
423 break;
424 case XFS_TRANS_SB_RES_FREXTENTS:
425 /*
426 * The allocation has already been applied to the
427 * in-core superblock's counter. This should only
428 * be applied to the on-disk superblock.
429 */
430 ASSERT(delta < 0);
431 tp->t_res_frextents_delta += delta;
432 break;
433 case XFS_TRANS_SB_DBLOCKS:
434 tp->t_dblocks_delta += delta;
435 break;
436 case XFS_TRANS_SB_AGCOUNT:
437 ASSERT(delta > 0);
438 tp->t_agcount_delta += delta;
439 break;
440 case XFS_TRANS_SB_IMAXPCT:
441 tp->t_imaxpct_delta += delta;
442 break;
443 case XFS_TRANS_SB_REXTSIZE:
444 tp->t_rextsize_delta += delta;
445 break;
446 case XFS_TRANS_SB_RBMBLOCKS:
447 tp->t_rbmblocks_delta += delta;
448 break;
449 case XFS_TRANS_SB_RBLOCKS:
450 tp->t_rblocks_delta += delta;
451 break;
452 case XFS_TRANS_SB_REXTENTS:
453 tp->t_rextents_delta += delta;
454 break;
455 case XFS_TRANS_SB_REXTSLOG:
456 tp->t_rextslog_delta += delta;
457 break;
458 default:
459 ASSERT(0);
460 return;
461 }
462
463 tp->t_flags |= flags;
464}
465
466/*
467 * xfs_trans_apply_sb_deltas() is called from the commit code
468 * to bring the superblock buffer into the current transaction
469 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
470 *
471 * For now we just look at each field allowed to change and change
472 * it if necessary.
473 */
474STATIC void
475xfs_trans_apply_sb_deltas(
476 xfs_trans_t *tp)
477{
478 struct xfs_dsb *sbp;
479 struct xfs_buf *bp;
480 int whole = 0;
481
482 bp = xfs_trans_getsb(tp);
483 sbp = bp->b_addr;
484
485 /*
486 * Only update the superblock counters if we are logging them
487 */
488 if (!xfs_has_lazysbcount((tp->t_mountp))) {
489 if (tp->t_icount_delta)
490 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
491 if (tp->t_ifree_delta)
492 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
493 if (tp->t_fdblocks_delta)
494 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
495 if (tp->t_res_fdblocks_delta)
496 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
497 }
498
499 /*
500 * Updating frextents requires careful handling because it does not
501 * behave like the lazysb counters because we cannot rely on log
502 * recovery in older kenels to recompute the value from the rtbitmap.
503 * This means that the ondisk frextents must be consistent with the
504 * rtbitmap.
505 *
506 * Therefore, log the frextents change to the ondisk superblock and
507 * update the incore superblock so that future calls to xfs_log_sb
508 * write the correct value ondisk.
509 *
510 * Don't touch m_frextents because it includes incore reservations,
511 * and those are handled by the unreserve function.
512 */
513 if (tp->t_frextents_delta || tp->t_res_frextents_delta) {
514 struct xfs_mount *mp = tp->t_mountp;
515 int64_t rtxdelta;
516
517 rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;
518
519 spin_lock(&mp->m_sb_lock);
520 be64_add_cpu(&sbp->sb_frextents, rtxdelta);
521 mp->m_sb.sb_frextents += rtxdelta;
522 spin_unlock(&mp->m_sb_lock);
523 }
524
525 if (tp->t_dblocks_delta) {
526 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
527 whole = 1;
528 }
529 if (tp->t_agcount_delta) {
530 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
531 whole = 1;
532 }
533 if (tp->t_imaxpct_delta) {
534 sbp->sb_imax_pct += tp->t_imaxpct_delta;
535 whole = 1;
536 }
537 if (tp->t_rextsize_delta) {
538 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
539 whole = 1;
540 }
541 if (tp->t_rbmblocks_delta) {
542 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
543 whole = 1;
544 }
545 if (tp->t_rblocks_delta) {
546 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
547 whole = 1;
548 }
549 if (tp->t_rextents_delta) {
550 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
551 whole = 1;
552 }
553 if (tp->t_rextslog_delta) {
554 sbp->sb_rextslog += tp->t_rextslog_delta;
555 whole = 1;
556 }
557
558 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
559 if (whole)
560 /*
561 * Log the whole thing, the fields are noncontiguous.
562 */
563 xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
564 else
565 /*
566 * Since all the modifiable fields are contiguous, we
567 * can get away with this.
568 */
569 xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
570 offsetof(struct xfs_dsb, sb_frextents) +
571 sizeof(sbp->sb_frextents) - 1);
572}
573
574/*
575 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
576 * apply superblock counter changes to the in-core superblock. The
577 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
578 * applied to the in-core superblock. The idea is that that has already been
579 * done.
580 *
581 * If we are not logging superblock counters, then the inode allocated/free and
582 * used block counts are not updated in the on disk superblock. In this case,
583 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
584 * still need to update the incore superblock with the changes.
585 *
586 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
587 * so we don't need to take the counter lock on every update.
588 */
589#define XFS_ICOUNT_BATCH 128
590
591void
592xfs_trans_unreserve_and_mod_sb(
593 struct xfs_trans *tp)
594{
595 struct xfs_mount *mp = tp->t_mountp;
596 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
597 int64_t blkdelta = 0;
598 int64_t rtxdelta = 0;
599 int64_t idelta = 0;
600 int64_t ifreedelta = 0;
601 int error;
602
603 /* calculate deltas */
604 if (tp->t_blk_res > 0)
605 blkdelta = tp->t_blk_res;
606 if ((tp->t_fdblocks_delta != 0) &&
607 (xfs_has_lazysbcount(mp) ||
608 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
609 blkdelta += tp->t_fdblocks_delta;
610
611 if (tp->t_rtx_res > 0)
612 rtxdelta = tp->t_rtx_res;
613 if ((tp->t_frextents_delta != 0) &&
614 (tp->t_flags & XFS_TRANS_SB_DIRTY))
615 rtxdelta += tp->t_frextents_delta;
616
617 if (xfs_has_lazysbcount(mp) ||
618 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
619 idelta = tp->t_icount_delta;
620 ifreedelta = tp->t_ifree_delta;
621 }
622
623 /* apply the per-cpu counters */
624 if (blkdelta) {
625 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
626 ASSERT(!error);
627 }
628
629 if (idelta)
630 percpu_counter_add_batch(&mp->m_icount, idelta,
631 XFS_ICOUNT_BATCH);
632
633 if (ifreedelta)
634 percpu_counter_add(&mp->m_ifree, ifreedelta);
635
636 if (rtxdelta) {
637 error = xfs_mod_frextents(mp, rtxdelta);
638 ASSERT(!error);
639 }
640
641 if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))
642 return;
643
644 /* apply remaining deltas */
645 spin_lock(&mp->m_sb_lock);
646 mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
647 mp->m_sb.sb_icount += idelta;
648 mp->m_sb.sb_ifree += ifreedelta;
649 /*
650 * Do not touch sb_frextents here because we are dealing with incore
651 * reservation. sb_frextents is not part of the lazy sb counters so it
652 * must be consistent with the ondisk rtbitmap and must never include
653 * incore reservations.
654 */
655 mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
656 mp->m_sb.sb_agcount += tp->t_agcount_delta;
657 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
658 mp->m_sb.sb_rextsize += tp->t_rextsize_delta;
659 if (tp->t_rextsize_delta) {
660 mp->m_rtxblklog = log2_if_power2(mp->m_sb.sb_rextsize);
661 mp->m_rtxblkmask = mask64_if_power2(mp->m_sb.sb_rextsize);
662 }
663 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
664 mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
665 mp->m_sb.sb_rextents += tp->t_rextents_delta;
666 mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
667 spin_unlock(&mp->m_sb_lock);
668
669 /*
670 * Debug checks outside of the spinlock so they don't lock up the
671 * machine if they fail.
672 */
673 ASSERT(mp->m_sb.sb_imax_pct >= 0);
674 ASSERT(mp->m_sb.sb_rextslog >= 0);
675 return;
676}
677
678/* Add the given log item to the transaction's list of log items. */
679void
680xfs_trans_add_item(
681 struct xfs_trans *tp,
682 struct xfs_log_item *lip)
683{
684 ASSERT(lip->li_log == tp->t_mountp->m_log);
685 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
686 ASSERT(list_empty(&lip->li_trans));
687 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
688
689 list_add_tail(&lip->li_trans, &tp->t_items);
690 trace_xfs_trans_add_item(tp, _RET_IP_);
691}
692
693/*
694 * Unlink the log item from the transaction. the log item is no longer
695 * considered dirty in this transaction, as the linked transaction has
696 * finished, either by abort or commit completion.
697 */
698void
699xfs_trans_del_item(
700 struct xfs_log_item *lip)
701{
702 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
703 list_del_init(&lip->li_trans);
704}
705
706/* Detach and unlock all of the items in a transaction */
707static void
708xfs_trans_free_items(
709 struct xfs_trans *tp,
710 bool abort)
711{
712 struct xfs_log_item *lip, *next;
713
714 trace_xfs_trans_free_items(tp, _RET_IP_);
715
716 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
717 xfs_trans_del_item(lip);
718 if (abort)
719 set_bit(XFS_LI_ABORTED, &lip->li_flags);
720 if (lip->li_ops->iop_release)
721 lip->li_ops->iop_release(lip);
722 }
723}
724
725static inline void
726xfs_log_item_batch_insert(
727 struct xfs_ail *ailp,
728 struct xfs_ail_cursor *cur,
729 struct xfs_log_item **log_items,
730 int nr_items,
731 xfs_lsn_t commit_lsn)
732{
733 int i;
734
735 spin_lock(&ailp->ail_lock);
736 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
737 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
738
739 for (i = 0; i < nr_items; i++) {
740 struct xfs_log_item *lip = log_items[i];
741
742 if (lip->li_ops->iop_unpin)
743 lip->li_ops->iop_unpin(lip, 0);
744 }
745}
746
747/*
748 * Bulk operation version of xfs_trans_committed that takes a log vector of
749 * items to insert into the AIL. This uses bulk AIL insertion techniques to
750 * minimise lock traffic.
751 *
752 * If we are called with the aborted flag set, it is because a log write during
753 * a CIL checkpoint commit has failed. In this case, all the items in the
754 * checkpoint have already gone through iop_committed and iop_committing, which
755 * means that checkpoint commit abort handling is treated exactly the same
756 * as an iclog write error even though we haven't started any IO yet. Hence in
757 * this case all we need to do is iop_committed processing, followed by an
758 * iop_unpin(aborted) call.
759 *
760 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
761 * at the end of the AIL, the insert cursor avoids the need to walk
762 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
763 * call. This saves a lot of needless list walking and is a net win, even
764 * though it slightly increases that amount of AIL lock traffic to set it up
765 * and tear it down.
766 */
767void
768xfs_trans_committed_bulk(
769 struct xfs_ail *ailp,
770 struct list_head *lv_chain,
771 xfs_lsn_t commit_lsn,
772 bool aborted)
773{
774#define LOG_ITEM_BATCH_SIZE 32
775 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
776 struct xfs_log_vec *lv;
777 struct xfs_ail_cursor cur;
778 int i = 0;
779
780 spin_lock(&ailp->ail_lock);
781 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
782 spin_unlock(&ailp->ail_lock);
783
784 /* unpin all the log items */
785 list_for_each_entry(lv, lv_chain, lv_list) {
786 struct xfs_log_item *lip = lv->lv_item;
787 xfs_lsn_t item_lsn;
788
789 if (aborted)
790 set_bit(XFS_LI_ABORTED, &lip->li_flags);
791
792 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
793 lip->li_ops->iop_release(lip);
794 continue;
795 }
796
797 if (lip->li_ops->iop_committed)
798 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
799 else
800 item_lsn = commit_lsn;
801
802 /* item_lsn of -1 means the item needs no further processing */
803 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
804 continue;
805
806 /*
807 * if we are aborting the operation, no point in inserting the
808 * object into the AIL as we are in a shutdown situation.
809 */
810 if (aborted) {
811 ASSERT(xlog_is_shutdown(ailp->ail_log));
812 if (lip->li_ops->iop_unpin)
813 lip->li_ops->iop_unpin(lip, 1);
814 continue;
815 }
816
817 if (item_lsn != commit_lsn) {
818
819 /*
820 * Not a bulk update option due to unusual item_lsn.
821 * Push into AIL immediately, rechecking the lsn once
822 * we have the ail lock. Then unpin the item. This does
823 * not affect the AIL cursor the bulk insert path is
824 * using.
825 */
826 spin_lock(&ailp->ail_lock);
827 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
828 xfs_trans_ail_update(ailp, lip, item_lsn);
829 else
830 spin_unlock(&ailp->ail_lock);
831 if (lip->li_ops->iop_unpin)
832 lip->li_ops->iop_unpin(lip, 0);
833 continue;
834 }
835
836 /* Item is a candidate for bulk AIL insert. */
837 log_items[i++] = lv->lv_item;
838 if (i >= LOG_ITEM_BATCH_SIZE) {
839 xfs_log_item_batch_insert(ailp, &cur, log_items,
840 LOG_ITEM_BATCH_SIZE, commit_lsn);
841 i = 0;
842 }
843 }
844
845 /* make sure we insert the remainder! */
846 if (i)
847 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
848
849 spin_lock(&ailp->ail_lock);
850 xfs_trans_ail_cursor_done(&cur);
851 spin_unlock(&ailp->ail_lock);
852}
853
854/*
855 * Sort transaction items prior to running precommit operations. This will
856 * attempt to order the items such that they will always be locked in the same
857 * order. Items that have no sort function are moved to the end of the list
858 * and so are locked last.
859 *
860 * This may need refinement as different types of objects add sort functions.
861 *
862 * Function is more complex than it needs to be because we are comparing 64 bit
863 * values and the function only returns 32 bit values.
864 */
865static int
866xfs_trans_precommit_sort(
867 void *unused_arg,
868 const struct list_head *a,
869 const struct list_head *b)
870{
871 struct xfs_log_item *lia = container_of(a,
872 struct xfs_log_item, li_trans);
873 struct xfs_log_item *lib = container_of(b,
874 struct xfs_log_item, li_trans);
875 int64_t diff;
876
877 /*
878 * If both items are non-sortable, leave them alone. If only one is
879 * sortable, move the non-sortable item towards the end of the list.
880 */
881 if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)
882 return 0;
883 if (!lia->li_ops->iop_sort)
884 return 1;
885 if (!lib->li_ops->iop_sort)
886 return -1;
887
888 diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);
889 if (diff < 0)
890 return -1;
891 if (diff > 0)
892 return 1;
893 return 0;
894}
895
896/*
897 * Run transaction precommit functions.
898 *
899 * If there is an error in any of the callouts, then stop immediately and
900 * trigger a shutdown to abort the transaction. There is no recovery possible
901 * from errors at this point as the transaction is dirty....
902 */
903static int
904xfs_trans_run_precommits(
905 struct xfs_trans *tp)
906{
907 struct xfs_mount *mp = tp->t_mountp;
908 struct xfs_log_item *lip, *n;
909 int error = 0;
910
911 /*
912 * Sort the item list to avoid ABBA deadlocks with other transactions
913 * running precommit operations that lock multiple shared items such as
914 * inode cluster buffers.
915 */
916 list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);
917
918 /*
919 * Precommit operations can remove the log item from the transaction
920 * if the log item exists purely to delay modifications until they
921 * can be ordered against other operations. Hence we have to use
922 * list_for_each_entry_safe() here.
923 */
924 list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {
925 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
926 continue;
927 if (lip->li_ops->iop_precommit) {
928 error = lip->li_ops->iop_precommit(tp, lip);
929 if (error)
930 break;
931 }
932 }
933 if (error)
934 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
935 return error;
936}
937
938/*
939 * Commit the given transaction to the log.
940 *
941 * XFS disk error handling mechanism is not based on a typical
942 * transaction abort mechanism. Logically after the filesystem
943 * gets marked 'SHUTDOWN', we can't let any new transactions
944 * be durable - ie. committed to disk - because some metadata might
945 * be inconsistent. In such cases, this returns an error, and the
946 * caller may assume that all locked objects joined to the transaction
947 * have already been unlocked as if the commit had succeeded.
948 * Do not reference the transaction structure after this call.
949 */
950static int
951__xfs_trans_commit(
952 struct xfs_trans *tp,
953 bool regrant)
954{
955 struct xfs_mount *mp = tp->t_mountp;
956 struct xlog *log = mp->m_log;
957 xfs_csn_t commit_seq = 0;
958 int error = 0;
959 int sync = tp->t_flags & XFS_TRANS_SYNC;
960
961 trace_xfs_trans_commit(tp, _RET_IP_);
962
963 error = xfs_trans_run_precommits(tp);
964 if (error) {
965 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
966 xfs_defer_cancel(tp);
967 goto out_unreserve;
968 }
969
970 /*
971 * Finish deferred items on final commit. Only permanent transactions
972 * should ever have deferred ops.
973 */
974 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
975 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
976 if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
977 error = xfs_defer_finish_noroll(&tp);
978 if (error)
979 goto out_unreserve;
980
981 /* Run precommits from final tx in defer chain. */
982 error = xfs_trans_run_precommits(tp);
983 if (error)
984 goto out_unreserve;
985 }
986
987 /*
988 * If there is nothing to be logged by the transaction,
989 * then unlock all of the items associated with the
990 * transaction and free the transaction structure.
991 * Also make sure to return any reserved blocks to
992 * the free pool.
993 */
994 if (!(tp->t_flags & XFS_TRANS_DIRTY))
995 goto out_unreserve;
996
997 /*
998 * We must check against log shutdown here because we cannot abort log
999 * items and leave them dirty, inconsistent and unpinned in memory while
1000 * the log is active. This leaves them open to being written back to
1001 * disk, and that will lead to on-disk corruption.
1002 */
1003 if (xlog_is_shutdown(log)) {
1004 error = -EIO;
1005 goto out_unreserve;
1006 }
1007
1008 ASSERT(tp->t_ticket != NULL);
1009
1010 /*
1011 * If we need to update the superblock, then do it now.
1012 */
1013 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
1014 xfs_trans_apply_sb_deltas(tp);
1015 xfs_trans_apply_dquot_deltas(tp);
1016
1017 xlog_cil_commit(log, tp, &commit_seq, regrant);
1018
1019 xfs_trans_free(tp);
1020
1021 /*
1022 * If the transaction needs to be synchronous, then force the
1023 * log out now and wait for it.
1024 */
1025 if (sync) {
1026 error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
1027 XFS_STATS_INC(mp, xs_trans_sync);
1028 } else {
1029 XFS_STATS_INC(mp, xs_trans_async);
1030 }
1031
1032 return error;
1033
1034out_unreserve:
1035 xfs_trans_unreserve_and_mod_sb(tp);
1036
1037 /*
1038 * It is indeed possible for the transaction to be not dirty but
1039 * the dqinfo portion to be. All that means is that we have some
1040 * (non-persistent) quota reservations that need to be unreserved.
1041 */
1042 xfs_trans_unreserve_and_mod_dquots(tp);
1043 if (tp->t_ticket) {
1044 if (regrant && !xlog_is_shutdown(log))
1045 xfs_log_ticket_regrant(log, tp->t_ticket);
1046 else
1047 xfs_log_ticket_ungrant(log, tp->t_ticket);
1048 tp->t_ticket = NULL;
1049 }
1050 xfs_trans_free_items(tp, !!error);
1051 xfs_trans_free(tp);
1052
1053 XFS_STATS_INC(mp, xs_trans_empty);
1054 return error;
1055}
1056
1057int
1058xfs_trans_commit(
1059 struct xfs_trans *tp)
1060{
1061 return __xfs_trans_commit(tp, false);
1062}
1063
1064/*
1065 * Unlock all of the transaction's items and free the transaction. If the
1066 * transaction is dirty, we must shut down the filesystem because there is no
1067 * way to restore them to their previous state.
1068 *
1069 * If the transaction has made a log reservation, make sure to release it as
1070 * well.
1071 *
1072 * This is a high level function (equivalent to xfs_trans_commit()) and so can
1073 * be called after the transaction has effectively been aborted due to the mount
1074 * being shut down. However, if the mount has not been shut down and the
1075 * transaction is dirty we will shut the mount down and, in doing so, that
1076 * guarantees that the log is shut down, too. Hence we don't need to be as
1077 * careful with shutdown state and dirty items here as we need to be in
1078 * xfs_trans_commit().
1079 */
1080void
1081xfs_trans_cancel(
1082 struct xfs_trans *tp)
1083{
1084 struct xfs_mount *mp = tp->t_mountp;
1085 struct xlog *log = mp->m_log;
1086 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
1087
1088 trace_xfs_trans_cancel(tp, _RET_IP_);
1089
1090 /*
1091 * It's never valid to cancel a transaction with deferred ops attached,
1092 * because the transaction is effectively dirty. Complain about this
1093 * loudly before freeing the in-memory defer items and shutting down the
1094 * filesystem.
1095 */
1096 if (!list_empty(&tp->t_dfops)) {
1097 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1098 dirty = true;
1099 xfs_defer_cancel(tp);
1100 }
1101
1102 /*
1103 * See if the caller is relying on us to shut down the filesystem. We
1104 * only want an error report if there isn't already a shutdown in
1105 * progress, so we only need to check against the mount shutdown state
1106 * here.
1107 */
1108 if (dirty && !xfs_is_shutdown(mp)) {
1109 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1110 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1111 }
1112#ifdef DEBUG
1113 /* Log items need to be consistent until the log is shut down. */
1114 if (!dirty && !xlog_is_shutdown(log)) {
1115 struct xfs_log_item *lip;
1116
1117 list_for_each_entry(lip, &tp->t_items, li_trans)
1118 ASSERT(!xlog_item_is_intent_done(lip));
1119 }
1120#endif
1121 xfs_trans_unreserve_and_mod_sb(tp);
1122 xfs_trans_unreserve_and_mod_dquots(tp);
1123
1124 if (tp->t_ticket) {
1125 xfs_log_ticket_ungrant(log, tp->t_ticket);
1126 tp->t_ticket = NULL;
1127 }
1128
1129 xfs_trans_free_items(tp, dirty);
1130 xfs_trans_free(tp);
1131}
1132
1133/*
1134 * Roll from one trans in the sequence of PERMANENT transactions to
1135 * the next: permanent transactions are only flushed out when
1136 * committed with xfs_trans_commit(), but we still want as soon
1137 * as possible to let chunks of it go to the log. So we commit the
1138 * chunk we've been working on and get a new transaction to continue.
1139 */
1140int
1141xfs_trans_roll(
1142 struct xfs_trans **tpp)
1143{
1144 struct xfs_trans *trans = *tpp;
1145 struct xfs_trans_res tres;
1146 int error;
1147
1148 trace_xfs_trans_roll(trans, _RET_IP_);
1149
1150 /*
1151 * Copy the critical parameters from one trans to the next.
1152 */
1153 tres.tr_logres = trans->t_log_res;
1154 tres.tr_logcount = trans->t_log_count;
1155
1156 *tpp = xfs_trans_dup(trans);
1157
1158 /*
1159 * Commit the current transaction.
1160 * If this commit failed, then it'd just unlock those items that
1161 * are not marked ihold. That also means that a filesystem shutdown
1162 * is in progress. The caller takes the responsibility to cancel
1163 * the duplicate transaction that gets returned.
1164 */
1165 error = __xfs_trans_commit(trans, true);
1166 if (error)
1167 return error;
1168
1169 /*
1170 * Reserve space in the log for the next transaction.
1171 * This also pushes items in the "AIL", the list of logged items,
1172 * out to disk if they are taking up space at the tail of the log
1173 * that we want to use. This requires that either nothing be locked
1174 * across this call, or that anything that is locked be logged in
1175 * the prior and the next transactions.
1176 */
1177 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1178 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1179}
1180
1181/*
1182 * Allocate an transaction, lock and join the inode to it, and reserve quota.
1183 *
1184 * The caller must ensure that the on-disk dquots attached to this inode have
1185 * already been allocated and initialized. The caller is responsible for
1186 * releasing ILOCK_EXCL if a new transaction is returned.
1187 */
1188int
1189xfs_trans_alloc_inode(
1190 struct xfs_inode *ip,
1191 struct xfs_trans_res *resv,
1192 unsigned int dblocks,
1193 unsigned int rblocks,
1194 bool force,
1195 struct xfs_trans **tpp)
1196{
1197 struct xfs_trans *tp;
1198 struct xfs_mount *mp = ip->i_mount;
1199 bool retried = false;
1200 int error;
1201
1202retry:
1203 error = xfs_trans_alloc(mp, resv, dblocks,
1204 xfs_extlen_to_rtxlen(mp, rblocks),
1205 force ? XFS_TRANS_RESERVE : 0, &tp);
1206 if (error)
1207 return error;
1208
1209 xfs_ilock(ip, XFS_ILOCK_EXCL);
1210 xfs_trans_ijoin(tp, ip, 0);
1211
1212 error = xfs_qm_dqattach_locked(ip, false);
1213 if (error) {
1214 /* Caller should have allocated the dquots! */
1215 ASSERT(error != -ENOENT);
1216 goto out_cancel;
1217 }
1218
1219 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1220 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1221 xfs_trans_cancel(tp);
1222 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1223 xfs_blockgc_free_quota(ip, 0);
1224 retried = true;
1225 goto retry;
1226 }
1227 if (error)
1228 goto out_cancel;
1229
1230 *tpp = tp;
1231 return 0;
1232
1233out_cancel:
1234 xfs_trans_cancel(tp);
1235 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1236 return error;
1237}
1238
1239/*
1240 * Try to reserve more blocks for a transaction.
1241 *
1242 * This is for callers that need to attach resources to a transaction, scan
1243 * those resources to determine the space reservation requirements, and then
1244 * modify the attached resources. In other words, online repair. This can
1245 * fail due to ENOSPC, so the caller must be able to cancel the transaction
1246 * without shutting down the fs.
1247 */
1248int
1249xfs_trans_reserve_more(
1250 struct xfs_trans *tp,
1251 unsigned int blocks,
1252 unsigned int rtextents)
1253{
1254 struct xfs_trans_res resv = { };
1255
1256 return xfs_trans_reserve(tp, &resv, blocks, rtextents);
1257}
1258
1259/*
1260 * Try to reserve more blocks and file quota for a transaction. Same
1261 * conditions of usage as xfs_trans_reserve_more.
1262 */
1263int
1264xfs_trans_reserve_more_inode(
1265 struct xfs_trans *tp,
1266 struct xfs_inode *ip,
1267 unsigned int dblocks,
1268 unsigned int rblocks,
1269 bool force_quota)
1270{
1271 struct xfs_trans_res resv = { };
1272 struct xfs_mount *mp = ip->i_mount;
1273 unsigned int rtx = xfs_extlen_to_rtxlen(mp, rblocks);
1274 int error;
1275
1276 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1277
1278 error = xfs_trans_reserve(tp, &resv, dblocks, rtx);
1279 if (error)
1280 return error;
1281
1282 if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino))
1283 return 0;
1284
1285 if (tp->t_flags & XFS_TRANS_RESERVE)
1286 force_quota = true;
1287
1288 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,
1289 force_quota);
1290 if (!error)
1291 return 0;
1292
1293 /* Quota failed, give back the new reservation. */
1294 xfs_mod_fdblocks(mp, dblocks, tp->t_flags & XFS_TRANS_RESERVE);
1295 tp->t_blk_res -= dblocks;
1296 xfs_mod_frextents(mp, rtx);
1297 tp->t_rtx_res -= rtx;
1298 return error;
1299}
1300
1301/*
1302 * Allocate an transaction in preparation for inode creation by reserving quota
1303 * against the given dquots. Callers are not required to hold any inode locks.
1304 */
1305int
1306xfs_trans_alloc_icreate(
1307 struct xfs_mount *mp,
1308 struct xfs_trans_res *resv,
1309 struct xfs_dquot *udqp,
1310 struct xfs_dquot *gdqp,
1311 struct xfs_dquot *pdqp,
1312 unsigned int dblocks,
1313 struct xfs_trans **tpp)
1314{
1315 struct xfs_trans *tp;
1316 bool retried = false;
1317 int error;
1318
1319retry:
1320 error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1321 if (error)
1322 return error;
1323
1324 error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1325 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1326 xfs_trans_cancel(tp);
1327 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1328 retried = true;
1329 goto retry;
1330 }
1331 if (error) {
1332 xfs_trans_cancel(tp);
1333 return error;
1334 }
1335
1336 *tpp = tp;
1337 return 0;
1338}
1339
1340/*
1341 * Allocate an transaction, lock and join the inode to it, and reserve quota
1342 * in preparation for inode attribute changes that include uid, gid, or prid
1343 * changes.
1344 *
1345 * The caller must ensure that the on-disk dquots attached to this inode have
1346 * already been allocated and initialized. The ILOCK will be dropped when the
1347 * transaction is committed or cancelled.
1348 */
1349int
1350xfs_trans_alloc_ichange(
1351 struct xfs_inode *ip,
1352 struct xfs_dquot *new_udqp,
1353 struct xfs_dquot *new_gdqp,
1354 struct xfs_dquot *new_pdqp,
1355 bool force,
1356 struct xfs_trans **tpp)
1357{
1358 struct xfs_trans *tp;
1359 struct xfs_mount *mp = ip->i_mount;
1360 struct xfs_dquot *udqp;
1361 struct xfs_dquot *gdqp;
1362 struct xfs_dquot *pdqp;
1363 bool retried = false;
1364 int error;
1365
1366retry:
1367 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1368 if (error)
1369 return error;
1370
1371 xfs_ilock(ip, XFS_ILOCK_EXCL);
1372 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1373
1374 error = xfs_qm_dqattach_locked(ip, false);
1375 if (error) {
1376 /* Caller should have allocated the dquots! */
1377 ASSERT(error != -ENOENT);
1378 goto out_cancel;
1379 }
1380
1381 /*
1382 * For each quota type, skip quota reservations if the inode's dquots
1383 * now match the ones that came from the caller, or the caller didn't
1384 * pass one in. The inode's dquots can change if we drop the ILOCK to
1385 * perform a blockgc scan, so we must preserve the caller's arguments.
1386 */
1387 udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1388 gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1389 pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1390 if (udqp || gdqp || pdqp) {
1391 unsigned int qflags = XFS_QMOPT_RES_REGBLKS;
1392
1393 if (force)
1394 qflags |= XFS_QMOPT_FORCE_RES;
1395
1396 /*
1397 * Reserve enough quota to handle blocks on disk and reserved
1398 * for a delayed allocation. We'll actually transfer the
1399 * delalloc reservation between dquots at chown time, even
1400 * though that part is only semi-transactional.
1401 */
1402 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1403 pdqp, ip->i_nblocks + ip->i_delayed_blks,
1404 1, qflags);
1405 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1406 xfs_trans_cancel(tp);
1407 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1408 retried = true;
1409 goto retry;
1410 }
1411 if (error)
1412 goto out_cancel;
1413 }
1414
1415 *tpp = tp;
1416 return 0;
1417
1418out_cancel:
1419 xfs_trans_cancel(tp);
1420 return error;
1421}
1422
1423/*
1424 * Allocate an transaction, lock and join the directory and child inodes to it,
1425 * and reserve quota for a directory update. If there isn't sufficient space,
1426 * @dblocks will be set to zero for a reservationless directory update and
1427 * @nospace_error will be set to a negative errno describing the space
1428 * constraint we hit.
1429 *
1430 * The caller must ensure that the on-disk dquots attached to this inode have
1431 * already been allocated and initialized. The ILOCKs will be dropped when the
1432 * transaction is committed or cancelled.
1433 */
1434int
1435xfs_trans_alloc_dir(
1436 struct xfs_inode *dp,
1437 struct xfs_trans_res *resv,
1438 struct xfs_inode *ip,
1439 unsigned int *dblocks,
1440 struct xfs_trans **tpp,
1441 int *nospace_error)
1442{
1443 struct xfs_trans *tp;
1444 struct xfs_mount *mp = ip->i_mount;
1445 unsigned int resblks;
1446 bool retried = false;
1447 int error;
1448
1449retry:
1450 *nospace_error = 0;
1451 resblks = *dblocks;
1452 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1453 if (error == -ENOSPC) {
1454 *nospace_error = error;
1455 resblks = 0;
1456 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1457 }
1458 if (error)
1459 return error;
1460
1461 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
1462
1463 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1464 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1465
1466 error = xfs_qm_dqattach_locked(dp, false);
1467 if (error) {
1468 /* Caller should have allocated the dquots! */
1469 ASSERT(error != -ENOENT);
1470 goto out_cancel;
1471 }
1472
1473 error = xfs_qm_dqattach_locked(ip, false);
1474 if (error) {
1475 /* Caller should have allocated the dquots! */
1476 ASSERT(error != -ENOENT);
1477 goto out_cancel;
1478 }
1479
1480 if (resblks == 0)
1481 goto done;
1482
1483 error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);
1484 if (error == -EDQUOT || error == -ENOSPC) {
1485 if (!retried) {
1486 xfs_trans_cancel(tp);
1487 xfs_blockgc_free_quota(dp, 0);
1488 retried = true;
1489 goto retry;
1490 }
1491
1492 *nospace_error = error;
1493 resblks = 0;
1494 error = 0;
1495 }
1496 if (error)
1497 goto out_cancel;
1498
1499done:
1500 *tpp = tp;
1501 *dblocks = resblks;
1502 return 0;
1503
1504out_cancel:
1505 xfs_trans_cancel(tp);
1506 return error;
1507}