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
28struct kmem_cache *xfs_trans_cache;
29
30#if defined(CONFIG_TRACEPOINTS)
31static void
32xfs_trans_trace_reservations(
33 struct xfs_mount *mp)
34{
35 struct xfs_trans_res *res;
36 struct xfs_trans_res *end_res;
37 int i;
38
39 res = (struct xfs_trans_res *)M_RES(mp);
40 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
41 for (i = 0; res < end_res; i++, res++)
42 trace_xfs_trans_resv_calc(mp, i, res);
43}
44#else
45# define xfs_trans_trace_reservations(mp)
46#endif
47
48/*
49 * Initialize the precomputed transaction reservation values
50 * in the mount structure.
51 */
52void
53xfs_trans_init(
54 struct xfs_mount *mp)
55{
56 xfs_trans_resv_calc(mp, M_RES(mp));
57 xfs_trans_trace_reservations(mp);
58}
59
60/*
61 * Free the transaction structure. If there is more clean up
62 * to do when the structure is freed, add it here.
63 */
64STATIC void
65xfs_trans_free(
66 struct xfs_trans *tp)
67{
68 xfs_extent_busy_sort(&tp->t_busy);
69 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
70
71 trace_xfs_trans_free(tp, _RET_IP_);
72 xfs_trans_clear_context(tp);
73 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
74 sb_end_intwrite(tp->t_mountp->m_super);
75 xfs_trans_free_dqinfo(tp);
76 kmem_cache_free(xfs_trans_cache, tp);
77}
78
79/*
80 * This is called to create a new transaction which will share the
81 * permanent log reservation of the given transaction. The remaining
82 * unused block and rt extent reservations are also inherited. This
83 * implies that the original transaction is no longer allowed to allocate
84 * blocks. Locks and log items, however, are no inherited. They must
85 * be added to the new transaction explicitly.
86 */
87STATIC struct xfs_trans *
88xfs_trans_dup(
89 struct xfs_trans *tp)
90{
91 struct xfs_trans *ntp;
92
93 trace_xfs_trans_dup(tp, _RET_IP_);
94
95 ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
96
97 /*
98 * Initialize the new transaction structure.
99 */
100 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
101 ntp->t_mountp = tp->t_mountp;
102 INIT_LIST_HEAD(&ntp->t_items);
103 INIT_LIST_HEAD(&ntp->t_busy);
104 INIT_LIST_HEAD(&ntp->t_dfops);
105 ntp->t_firstblock = NULLFSBLOCK;
106
107 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
108 ASSERT(tp->t_ticket != NULL);
109
110 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
111 (tp->t_flags & XFS_TRANS_RESERVE) |
112 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
113 (tp->t_flags & XFS_TRANS_RES_FDBLKS);
114 /* We gave our writer reference to the new transaction */
115 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
116 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
117
118 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
119 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
120 tp->t_blk_res = tp->t_blk_res_used;
121
122 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
123 tp->t_rtx_res = tp->t_rtx_res_used;
124
125 xfs_trans_switch_context(tp, ntp);
126
127 /* move deferred ops over to the new tp */
128 xfs_defer_move(ntp, tp);
129
130 xfs_trans_dup_dqinfo(tp, ntp);
131 return ntp;
132}
133
134/*
135 * This is called to reserve free disk blocks and log space for the
136 * given transaction. This must be done before allocating any resources
137 * within the transaction.
138 *
139 * This will return ENOSPC if there are not enough blocks available.
140 * It will sleep waiting for available log space.
141 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
142 * is used by long running transactions. If any one of the reservations
143 * fails then they will all be backed out.
144 *
145 * This does not do quota reservations. That typically is done by the
146 * caller afterwards.
147 */
148static int
149xfs_trans_reserve(
150 struct xfs_trans *tp,
151 struct xfs_trans_res *resp,
152 uint blocks,
153 uint rtextents)
154{
155 struct xfs_mount *mp = tp->t_mountp;
156 int error = 0;
157 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
158
159 /*
160 * Attempt to reserve the needed disk blocks by decrementing
161 * the number needed from the number available. This will
162 * fail if the count would go below zero.
163 */
164 if (blocks > 0) {
165 error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
166 if (error != 0)
167 return -ENOSPC;
168 tp->t_blk_res += blocks;
169 }
170
171 /*
172 * Reserve the log space needed for this transaction.
173 */
174 if (resp->tr_logres > 0) {
175 bool permanent = false;
176
177 ASSERT(tp->t_log_res == 0 ||
178 tp->t_log_res == resp->tr_logres);
179 ASSERT(tp->t_log_count == 0 ||
180 tp->t_log_count == resp->tr_logcount);
181
182 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
183 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
184 permanent = true;
185 } else {
186 ASSERT(tp->t_ticket == NULL);
187 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
188 }
189
190 if (tp->t_ticket != NULL) {
191 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
192 error = xfs_log_regrant(mp, tp->t_ticket);
193 } else {
194 error = xfs_log_reserve(mp, resp->tr_logres,
195 resp->tr_logcount,
196 &tp->t_ticket, permanent);
197 }
198
199 if (error)
200 goto undo_blocks;
201
202 tp->t_log_res = resp->tr_logres;
203 tp->t_log_count = resp->tr_logcount;
204 }
205
206 /*
207 * Attempt to reserve the needed realtime extents by decrementing
208 * the number needed from the number available. This will
209 * fail if the count would go below zero.
210 */
211 if (rtextents > 0) {
212 error = xfs_mod_frextents(mp, -((int64_t)rtextents));
213 if (error) {
214 error = -ENOSPC;
215 goto undo_log;
216 }
217 tp->t_rtx_res += rtextents;
218 }
219
220 return 0;
221
222 /*
223 * Error cases jump to one of these labels to undo any
224 * reservations which have already been performed.
225 */
226undo_log:
227 if (resp->tr_logres > 0) {
228 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
229 tp->t_ticket = NULL;
230 tp->t_log_res = 0;
231 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
232 }
233
234undo_blocks:
235 if (blocks > 0) {
236 xfs_mod_fdblocks(mp, (int64_t)blocks, rsvd);
237 tp->t_blk_res = 0;
238 }
239 return error;
240}
241
242int
243xfs_trans_alloc(
244 struct xfs_mount *mp,
245 struct xfs_trans_res *resp,
246 uint blocks,
247 uint rtextents,
248 uint flags,
249 struct xfs_trans **tpp)
250{
251 struct xfs_trans *tp;
252 bool want_retry = true;
253 int error;
254
255 /*
256 * Allocate the handle before we do our freeze accounting and setting up
257 * GFP_NOFS allocation context so that we avoid lockdep false positives
258 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
259 */
260retry:
261 tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
262 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
263 sb_start_intwrite(mp->m_super);
264 xfs_trans_set_context(tp);
265
266 /*
267 * Zero-reservation ("empty") transactions can't modify anything, so
268 * they're allowed to run while we're frozen.
269 */
270 WARN_ON(resp->tr_logres > 0 &&
271 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
272 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
273 xfs_has_lazysbcount(mp));
274
275 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
276 tp->t_flags = flags;
277 tp->t_mountp = mp;
278 INIT_LIST_HEAD(&tp->t_items);
279 INIT_LIST_HEAD(&tp->t_busy);
280 INIT_LIST_HEAD(&tp->t_dfops);
281 tp->t_firstblock = NULLFSBLOCK;
282
283 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
284 if (error == -ENOSPC && want_retry) {
285 xfs_trans_cancel(tp);
286
287 /*
288 * We weren't able to reserve enough space for the transaction.
289 * Flush the other speculative space allocations to free space.
290 * Do not perform a synchronous scan because callers can hold
291 * other locks.
292 */
293 xfs_blockgc_flush_all(mp);
294 want_retry = false;
295 goto retry;
296 }
297 if (error) {
298 xfs_trans_cancel(tp);
299 return error;
300 }
301
302 trace_xfs_trans_alloc(tp, _RET_IP_);
303
304 *tpp = tp;
305 return 0;
306}
307
308/*
309 * Create an empty transaction with no reservation. This is a defensive
310 * mechanism for routines that query metadata without actually modifying them --
311 * if the metadata being queried is somehow cross-linked (think a btree block
312 * pointer that points higher in the tree), we risk deadlock. However, blocks
313 * grabbed as part of a transaction can be re-grabbed. The verifiers will
314 * notice the corrupt block and the operation will fail back to userspace
315 * without deadlocking.
316 *
317 * Note the zero-length reservation; this transaction MUST be cancelled without
318 * any dirty data.
319 *
320 * Callers should obtain freeze protection to avoid a conflict with fs freezing
321 * where we can be grabbing buffers at the same time that freeze is trying to
322 * drain the buffer LRU list.
323 */
324int
325xfs_trans_alloc_empty(
326 struct xfs_mount *mp,
327 struct xfs_trans **tpp)
328{
329 struct xfs_trans_res resv = {0};
330
331 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
332}
333
334/*
335 * Record the indicated change to the given field for application
336 * to the file system's superblock when the transaction commits.
337 * For now, just store the change in the transaction structure.
338 *
339 * Mark the transaction structure to indicate that the superblock
340 * needs to be updated before committing.
341 *
342 * Because we may not be keeping track of allocated/free inodes and
343 * used filesystem blocks in the superblock, we do not mark the
344 * superblock dirty in this transaction if we modify these fields.
345 * We still need to update the transaction deltas so that they get
346 * applied to the incore superblock, but we don't want them to
347 * cause the superblock to get locked and logged if these are the
348 * only fields in the superblock that the transaction modifies.
349 */
350void
351xfs_trans_mod_sb(
352 xfs_trans_t *tp,
353 uint field,
354 int64_t delta)
355{
356 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
357 xfs_mount_t *mp = tp->t_mountp;
358
359 switch (field) {
360 case XFS_TRANS_SB_ICOUNT:
361 tp->t_icount_delta += delta;
362 if (xfs_has_lazysbcount(mp))
363 flags &= ~XFS_TRANS_SB_DIRTY;
364 break;
365 case XFS_TRANS_SB_IFREE:
366 tp->t_ifree_delta += delta;
367 if (xfs_has_lazysbcount(mp))
368 flags &= ~XFS_TRANS_SB_DIRTY;
369 break;
370 case XFS_TRANS_SB_FDBLOCKS:
371 /*
372 * Track the number of blocks allocated in the transaction.
373 * Make sure it does not exceed the number reserved. If so,
374 * shutdown as this can lead to accounting inconsistency.
375 */
376 if (delta < 0) {
377 tp->t_blk_res_used += (uint)-delta;
378 if (tp->t_blk_res_used > tp->t_blk_res)
379 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
380 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
381 int64_t blkres_delta;
382
383 /*
384 * Return freed blocks directly to the reservation
385 * instead of the global pool, being careful not to
386 * overflow the trans counter. This is used to preserve
387 * reservation across chains of transaction rolls that
388 * repeatedly free and allocate blocks.
389 */
390 blkres_delta = min_t(int64_t, delta,
391 UINT_MAX - tp->t_blk_res);
392 tp->t_blk_res += blkres_delta;
393 delta -= blkres_delta;
394 }
395 tp->t_fdblocks_delta += delta;
396 if (xfs_has_lazysbcount(mp))
397 flags &= ~XFS_TRANS_SB_DIRTY;
398 break;
399 case XFS_TRANS_SB_RES_FDBLOCKS:
400 /*
401 * The allocation has already been applied to the
402 * in-core superblock's counter. This should only
403 * be applied to the on-disk superblock.
404 */
405 tp->t_res_fdblocks_delta += delta;
406 if (xfs_has_lazysbcount(mp))
407 flags &= ~XFS_TRANS_SB_DIRTY;
408 break;
409 case XFS_TRANS_SB_FREXTENTS:
410 /*
411 * Track the number of blocks allocated in the
412 * transaction. Make sure it does not exceed the
413 * number reserved.
414 */
415 if (delta < 0) {
416 tp->t_rtx_res_used += (uint)-delta;
417 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
418 }
419 tp->t_frextents_delta += delta;
420 break;
421 case XFS_TRANS_SB_RES_FREXTENTS:
422 /*
423 * The allocation has already been applied to the
424 * in-core superblock's counter. This should only
425 * be applied to the on-disk superblock.
426 */
427 ASSERT(delta < 0);
428 tp->t_res_frextents_delta += delta;
429 break;
430 case XFS_TRANS_SB_DBLOCKS:
431 tp->t_dblocks_delta += delta;
432 break;
433 case XFS_TRANS_SB_AGCOUNT:
434 ASSERT(delta > 0);
435 tp->t_agcount_delta += delta;
436 break;
437 case XFS_TRANS_SB_IMAXPCT:
438 tp->t_imaxpct_delta += delta;
439 break;
440 case XFS_TRANS_SB_REXTSIZE:
441 tp->t_rextsize_delta += delta;
442 break;
443 case XFS_TRANS_SB_RBMBLOCKS:
444 tp->t_rbmblocks_delta += delta;
445 break;
446 case XFS_TRANS_SB_RBLOCKS:
447 tp->t_rblocks_delta += delta;
448 break;
449 case XFS_TRANS_SB_REXTENTS:
450 tp->t_rextents_delta += delta;
451 break;
452 case XFS_TRANS_SB_REXTSLOG:
453 tp->t_rextslog_delta += delta;
454 break;
455 default:
456 ASSERT(0);
457 return;
458 }
459
460 tp->t_flags |= flags;
461}
462
463/*
464 * xfs_trans_apply_sb_deltas() is called from the commit code
465 * to bring the superblock buffer into the current transaction
466 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
467 *
468 * For now we just look at each field allowed to change and change
469 * it if necessary.
470 */
471STATIC void
472xfs_trans_apply_sb_deltas(
473 xfs_trans_t *tp)
474{
475 struct xfs_dsb *sbp;
476 struct xfs_buf *bp;
477 int whole = 0;
478
479 bp = xfs_trans_getsb(tp);
480 sbp = bp->b_addr;
481
482 /*
483 * Only update the superblock counters if we are logging them
484 */
485 if (!xfs_has_lazysbcount((tp->t_mountp))) {
486 if (tp->t_icount_delta)
487 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
488 if (tp->t_ifree_delta)
489 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
490 if (tp->t_fdblocks_delta)
491 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
492 if (tp->t_res_fdblocks_delta)
493 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
494 }
495
496 /*
497 * Updating frextents requires careful handling because it does not
498 * behave like the lazysb counters because we cannot rely on log
499 * recovery in older kenels to recompute the value from the rtbitmap.
500 * This means that the ondisk frextents must be consistent with the
501 * rtbitmap.
502 *
503 * Therefore, log the frextents change to the ondisk superblock and
504 * update the incore superblock so that future calls to xfs_log_sb
505 * write the correct value ondisk.
506 *
507 * Don't touch m_frextents because it includes incore reservations,
508 * and those are handled by the unreserve function.
509 */
510 if (tp->t_frextents_delta || tp->t_res_frextents_delta) {
511 struct xfs_mount *mp = tp->t_mountp;
512 int64_t rtxdelta;
513
514 rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;
515
516 spin_lock(&mp->m_sb_lock);
517 be64_add_cpu(&sbp->sb_frextents, rtxdelta);
518 mp->m_sb.sb_frextents += rtxdelta;
519 spin_unlock(&mp->m_sb_lock);
520 }
521
522 if (tp->t_dblocks_delta) {
523 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
524 whole = 1;
525 }
526 if (tp->t_agcount_delta) {
527 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
528 whole = 1;
529 }
530 if (tp->t_imaxpct_delta) {
531 sbp->sb_imax_pct += tp->t_imaxpct_delta;
532 whole = 1;
533 }
534 if (tp->t_rextsize_delta) {
535 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
536 whole = 1;
537 }
538 if (tp->t_rbmblocks_delta) {
539 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
540 whole = 1;
541 }
542 if (tp->t_rblocks_delta) {
543 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
544 whole = 1;
545 }
546 if (tp->t_rextents_delta) {
547 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
548 whole = 1;
549 }
550 if (tp->t_rextslog_delta) {
551 sbp->sb_rextslog += tp->t_rextslog_delta;
552 whole = 1;
553 }
554
555 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
556 if (whole)
557 /*
558 * Log the whole thing, the fields are noncontiguous.
559 */
560 xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
561 else
562 /*
563 * Since all the modifiable fields are contiguous, we
564 * can get away with this.
565 */
566 xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
567 offsetof(struct xfs_dsb, sb_frextents) +
568 sizeof(sbp->sb_frextents) - 1);
569}
570
571/*
572 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
573 * apply superblock counter changes to the in-core superblock. The
574 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
575 * applied to the in-core superblock. The idea is that that has already been
576 * done.
577 *
578 * If we are not logging superblock counters, then the inode allocated/free and
579 * used block counts are not updated in the on disk superblock. In this case,
580 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
581 * still need to update the incore superblock with the changes.
582 *
583 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
584 * so we don't need to take the counter lock on every update.
585 */
586#define XFS_ICOUNT_BATCH 128
587
588void
589xfs_trans_unreserve_and_mod_sb(
590 struct xfs_trans *tp)
591{
592 struct xfs_mount *mp = tp->t_mountp;
593 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
594 int64_t blkdelta = 0;
595 int64_t rtxdelta = 0;
596 int64_t idelta = 0;
597 int64_t ifreedelta = 0;
598 int error;
599
600 /* calculate deltas */
601 if (tp->t_blk_res > 0)
602 blkdelta = tp->t_blk_res;
603 if ((tp->t_fdblocks_delta != 0) &&
604 (xfs_has_lazysbcount(mp) ||
605 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
606 blkdelta += tp->t_fdblocks_delta;
607
608 if (tp->t_rtx_res > 0)
609 rtxdelta = tp->t_rtx_res;
610 if ((tp->t_frextents_delta != 0) &&
611 (tp->t_flags & XFS_TRANS_SB_DIRTY))
612 rtxdelta += tp->t_frextents_delta;
613
614 if (xfs_has_lazysbcount(mp) ||
615 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
616 idelta = tp->t_icount_delta;
617 ifreedelta = tp->t_ifree_delta;
618 }
619
620 /* apply the per-cpu counters */
621 if (blkdelta) {
622 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
623 ASSERT(!error);
624 }
625
626 if (idelta)
627 percpu_counter_add_batch(&mp->m_icount, idelta,
628 XFS_ICOUNT_BATCH);
629
630 if (ifreedelta)
631 percpu_counter_add(&mp->m_ifree, ifreedelta);
632
633 if (rtxdelta) {
634 error = xfs_mod_frextents(mp, rtxdelta);
635 ASSERT(!error);
636 }
637
638 if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))
639 return;
640
641 /* apply remaining deltas */
642 spin_lock(&mp->m_sb_lock);
643 mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
644 mp->m_sb.sb_icount += idelta;
645 mp->m_sb.sb_ifree += ifreedelta;
646 /*
647 * Do not touch sb_frextents here because we are dealing with incore
648 * reservation. sb_frextents is not part of the lazy sb counters so it
649 * must be consistent with the ondisk rtbitmap and must never include
650 * incore reservations.
651 */
652 mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
653 mp->m_sb.sb_agcount += tp->t_agcount_delta;
654 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
655 mp->m_sb.sb_rextsize += tp->t_rextsize_delta;
656 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
657 mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
658 mp->m_sb.sb_rextents += tp->t_rextents_delta;
659 mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
660 spin_unlock(&mp->m_sb_lock);
661
662 /*
663 * Debug checks outside of the spinlock so they don't lock up the
664 * machine if they fail.
665 */
666 ASSERT(mp->m_sb.sb_imax_pct >= 0);
667 ASSERT(mp->m_sb.sb_rextslog >= 0);
668 return;
669}
670
671/* Add the given log item to the transaction's list of log items. */
672void
673xfs_trans_add_item(
674 struct xfs_trans *tp,
675 struct xfs_log_item *lip)
676{
677 ASSERT(lip->li_log == tp->t_mountp->m_log);
678 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
679 ASSERT(list_empty(&lip->li_trans));
680 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
681
682 list_add_tail(&lip->li_trans, &tp->t_items);
683 trace_xfs_trans_add_item(tp, _RET_IP_);
684}
685
686/*
687 * Unlink the log item from the transaction. the log item is no longer
688 * considered dirty in this transaction, as the linked transaction has
689 * finished, either by abort or commit completion.
690 */
691void
692xfs_trans_del_item(
693 struct xfs_log_item *lip)
694{
695 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
696 list_del_init(&lip->li_trans);
697}
698
699/* Detach and unlock all of the items in a transaction */
700static void
701xfs_trans_free_items(
702 struct xfs_trans *tp,
703 bool abort)
704{
705 struct xfs_log_item *lip, *next;
706
707 trace_xfs_trans_free_items(tp, _RET_IP_);
708
709 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
710 xfs_trans_del_item(lip);
711 if (abort)
712 set_bit(XFS_LI_ABORTED, &lip->li_flags);
713 if (lip->li_ops->iop_release)
714 lip->li_ops->iop_release(lip);
715 }
716}
717
718static inline void
719xfs_log_item_batch_insert(
720 struct xfs_ail *ailp,
721 struct xfs_ail_cursor *cur,
722 struct xfs_log_item **log_items,
723 int nr_items,
724 xfs_lsn_t commit_lsn)
725{
726 int i;
727
728 spin_lock(&ailp->ail_lock);
729 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
730 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
731
732 for (i = 0; i < nr_items; i++) {
733 struct xfs_log_item *lip = log_items[i];
734
735 if (lip->li_ops->iop_unpin)
736 lip->li_ops->iop_unpin(lip, 0);
737 }
738}
739
740/*
741 * Bulk operation version of xfs_trans_committed that takes a log vector of
742 * items to insert into the AIL. This uses bulk AIL insertion techniques to
743 * minimise lock traffic.
744 *
745 * If we are called with the aborted flag set, it is because a log write during
746 * a CIL checkpoint commit has failed. In this case, all the items in the
747 * checkpoint have already gone through iop_committed and iop_committing, which
748 * means that checkpoint commit abort handling is treated exactly the same
749 * as an iclog write error even though we haven't started any IO yet. Hence in
750 * this case all we need to do is iop_committed processing, followed by an
751 * iop_unpin(aborted) call.
752 *
753 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
754 * at the end of the AIL, the insert cursor avoids the need to walk
755 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
756 * call. This saves a lot of needless list walking and is a net win, even
757 * though it slightly increases that amount of AIL lock traffic to set it up
758 * and tear it down.
759 */
760void
761xfs_trans_committed_bulk(
762 struct xfs_ail *ailp,
763 struct list_head *lv_chain,
764 xfs_lsn_t commit_lsn,
765 bool aborted)
766{
767#define LOG_ITEM_BATCH_SIZE 32
768 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
769 struct xfs_log_vec *lv;
770 struct xfs_ail_cursor cur;
771 int i = 0;
772
773 spin_lock(&ailp->ail_lock);
774 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
775 spin_unlock(&ailp->ail_lock);
776
777 /* unpin all the log items */
778 list_for_each_entry(lv, lv_chain, lv_list) {
779 struct xfs_log_item *lip = lv->lv_item;
780 xfs_lsn_t item_lsn;
781
782 if (aborted)
783 set_bit(XFS_LI_ABORTED, &lip->li_flags);
784
785 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
786 lip->li_ops->iop_release(lip);
787 continue;
788 }
789
790 if (lip->li_ops->iop_committed)
791 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
792 else
793 item_lsn = commit_lsn;
794
795 /* item_lsn of -1 means the item needs no further processing */
796 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
797 continue;
798
799 /*
800 * if we are aborting the operation, no point in inserting the
801 * object into the AIL as we are in a shutdown situation.
802 */
803 if (aborted) {
804 ASSERT(xlog_is_shutdown(ailp->ail_log));
805 if (lip->li_ops->iop_unpin)
806 lip->li_ops->iop_unpin(lip, 1);
807 continue;
808 }
809
810 if (item_lsn != commit_lsn) {
811
812 /*
813 * Not a bulk update option due to unusual item_lsn.
814 * Push into AIL immediately, rechecking the lsn once
815 * we have the ail lock. Then unpin the item. This does
816 * not affect the AIL cursor the bulk insert path is
817 * using.
818 */
819 spin_lock(&ailp->ail_lock);
820 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
821 xfs_trans_ail_update(ailp, lip, item_lsn);
822 else
823 spin_unlock(&ailp->ail_lock);
824 if (lip->li_ops->iop_unpin)
825 lip->li_ops->iop_unpin(lip, 0);
826 continue;
827 }
828
829 /* Item is a candidate for bulk AIL insert. */
830 log_items[i++] = lv->lv_item;
831 if (i >= LOG_ITEM_BATCH_SIZE) {
832 xfs_log_item_batch_insert(ailp, &cur, log_items,
833 LOG_ITEM_BATCH_SIZE, commit_lsn);
834 i = 0;
835 }
836 }
837
838 /* make sure we insert the remainder! */
839 if (i)
840 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
841
842 spin_lock(&ailp->ail_lock);
843 xfs_trans_ail_cursor_done(&cur);
844 spin_unlock(&ailp->ail_lock);
845}
846
847/*
848 * Sort transaction items prior to running precommit operations. This will
849 * attempt to order the items such that they will always be locked in the same
850 * order. Items that have no sort function are moved to the end of the list
851 * and so are locked last.
852 *
853 * This may need refinement as different types of objects add sort functions.
854 *
855 * Function is more complex than it needs to be because we are comparing 64 bit
856 * values and the function only returns 32 bit values.
857 */
858static int
859xfs_trans_precommit_sort(
860 void *unused_arg,
861 const struct list_head *a,
862 const struct list_head *b)
863{
864 struct xfs_log_item *lia = container_of(a,
865 struct xfs_log_item, li_trans);
866 struct xfs_log_item *lib = container_of(b,
867 struct xfs_log_item, li_trans);
868 int64_t diff;
869
870 /*
871 * If both items are non-sortable, leave them alone. If only one is
872 * sortable, move the non-sortable item towards the end of the list.
873 */
874 if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)
875 return 0;
876 if (!lia->li_ops->iop_sort)
877 return 1;
878 if (!lib->li_ops->iop_sort)
879 return -1;
880
881 diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);
882 if (diff < 0)
883 return -1;
884 if (diff > 0)
885 return 1;
886 return 0;
887}
888
889/*
890 * Run transaction precommit functions.
891 *
892 * If there is an error in any of the callouts, then stop immediately and
893 * trigger a shutdown to abort the transaction. There is no recovery possible
894 * from errors at this point as the transaction is dirty....
895 */
896static int
897xfs_trans_run_precommits(
898 struct xfs_trans *tp)
899{
900 struct xfs_mount *mp = tp->t_mountp;
901 struct xfs_log_item *lip, *n;
902 int error = 0;
903
904 /*
905 * Sort the item list to avoid ABBA deadlocks with other transactions
906 * running precommit operations that lock multiple shared items such as
907 * inode cluster buffers.
908 */
909 list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);
910
911 /*
912 * Precommit operations can remove the log item from the transaction
913 * if the log item exists purely to delay modifications until they
914 * can be ordered against other operations. Hence we have to use
915 * list_for_each_entry_safe() here.
916 */
917 list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {
918 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
919 continue;
920 if (lip->li_ops->iop_precommit) {
921 error = lip->li_ops->iop_precommit(tp, lip);
922 if (error)
923 break;
924 }
925 }
926 if (error)
927 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
928 return error;
929}
930
931/*
932 * Commit the given transaction to the log.
933 *
934 * XFS disk error handling mechanism is not based on a typical
935 * transaction abort mechanism. Logically after the filesystem
936 * gets marked 'SHUTDOWN', we can't let any new transactions
937 * be durable - ie. committed to disk - because some metadata might
938 * be inconsistent. In such cases, this returns an error, and the
939 * caller may assume that all locked objects joined to the transaction
940 * have already been unlocked as if the commit had succeeded.
941 * Do not reference the transaction structure after this call.
942 */
943static int
944__xfs_trans_commit(
945 struct xfs_trans *tp,
946 bool regrant)
947{
948 struct xfs_mount *mp = tp->t_mountp;
949 struct xlog *log = mp->m_log;
950 xfs_csn_t commit_seq = 0;
951 int error = 0;
952 int sync = tp->t_flags & XFS_TRANS_SYNC;
953
954 trace_xfs_trans_commit(tp, _RET_IP_);
955
956 error = xfs_trans_run_precommits(tp);
957 if (error) {
958 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
959 xfs_defer_cancel(tp);
960 goto out_unreserve;
961 }
962
963 /*
964 * Finish deferred items on final commit. Only permanent transactions
965 * should ever have deferred ops.
966 */
967 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
968 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
969 if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
970 error = xfs_defer_finish_noroll(&tp);
971 if (error)
972 goto out_unreserve;
973 }
974
975 /*
976 * If there is nothing to be logged by the transaction,
977 * then unlock all of the items associated with the
978 * transaction and free the transaction structure.
979 * Also make sure to return any reserved blocks to
980 * the free pool.
981 */
982 if (!(tp->t_flags & XFS_TRANS_DIRTY))
983 goto out_unreserve;
984
985 /*
986 * We must check against log shutdown here because we cannot abort log
987 * items and leave them dirty, inconsistent and unpinned in memory while
988 * the log is active. This leaves them open to being written back to
989 * disk, and that will lead to on-disk corruption.
990 */
991 if (xlog_is_shutdown(log)) {
992 error = -EIO;
993 goto out_unreserve;
994 }
995
996 ASSERT(tp->t_ticket != NULL);
997
998 /*
999 * If we need to update the superblock, then do it now.
1000 */
1001 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
1002 xfs_trans_apply_sb_deltas(tp);
1003 xfs_trans_apply_dquot_deltas(tp);
1004
1005 xlog_cil_commit(log, tp, &commit_seq, regrant);
1006
1007 xfs_trans_free(tp);
1008
1009 /*
1010 * If the transaction needs to be synchronous, then force the
1011 * log out now and wait for it.
1012 */
1013 if (sync) {
1014 error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
1015 XFS_STATS_INC(mp, xs_trans_sync);
1016 } else {
1017 XFS_STATS_INC(mp, xs_trans_async);
1018 }
1019
1020 return error;
1021
1022out_unreserve:
1023 xfs_trans_unreserve_and_mod_sb(tp);
1024
1025 /*
1026 * It is indeed possible for the transaction to be not dirty but
1027 * the dqinfo portion to be. All that means is that we have some
1028 * (non-persistent) quota reservations that need to be unreserved.
1029 */
1030 xfs_trans_unreserve_and_mod_dquots(tp);
1031 if (tp->t_ticket) {
1032 if (regrant && !xlog_is_shutdown(log))
1033 xfs_log_ticket_regrant(log, tp->t_ticket);
1034 else
1035 xfs_log_ticket_ungrant(log, tp->t_ticket);
1036 tp->t_ticket = NULL;
1037 }
1038 xfs_trans_free_items(tp, !!error);
1039 xfs_trans_free(tp);
1040
1041 XFS_STATS_INC(mp, xs_trans_empty);
1042 return error;
1043}
1044
1045int
1046xfs_trans_commit(
1047 struct xfs_trans *tp)
1048{
1049 return __xfs_trans_commit(tp, false);
1050}
1051
1052/*
1053 * Unlock all of the transaction's items and free the transaction. If the
1054 * transaction is dirty, we must shut down the filesystem because there is no
1055 * way to restore them to their previous state.
1056 *
1057 * If the transaction has made a log reservation, make sure to release it as
1058 * well.
1059 *
1060 * This is a high level function (equivalent to xfs_trans_commit()) and so can
1061 * be called after the transaction has effectively been aborted due to the mount
1062 * being shut down. However, if the mount has not been shut down and the
1063 * transaction is dirty we will shut the mount down and, in doing so, that
1064 * guarantees that the log is shut down, too. Hence we don't need to be as
1065 * careful with shutdown state and dirty items here as we need to be in
1066 * xfs_trans_commit().
1067 */
1068void
1069xfs_trans_cancel(
1070 struct xfs_trans *tp)
1071{
1072 struct xfs_mount *mp = tp->t_mountp;
1073 struct xlog *log = mp->m_log;
1074 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
1075
1076 trace_xfs_trans_cancel(tp, _RET_IP_);
1077
1078 /*
1079 * It's never valid to cancel a transaction with deferred ops attached,
1080 * because the transaction is effectively dirty. Complain about this
1081 * loudly before freeing the in-memory defer items.
1082 */
1083 if (!list_empty(&tp->t_dfops)) {
1084 ASSERT(xfs_is_shutdown(mp) || list_empty(&tp->t_dfops));
1085 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1086 dirty = true;
1087 xfs_defer_cancel(tp);
1088 }
1089
1090 /*
1091 * See if the caller is relying on us to shut down the filesystem. We
1092 * only want an error report if there isn't already a shutdown in
1093 * progress, so we only need to check against the mount shutdown state
1094 * here.
1095 */
1096 if (dirty && !xfs_is_shutdown(mp)) {
1097 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1098 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1099 }
1100#ifdef DEBUG
1101 /* Log items need to be consistent until the log is shut down. */
1102 if (!dirty && !xlog_is_shutdown(log)) {
1103 struct xfs_log_item *lip;
1104
1105 list_for_each_entry(lip, &tp->t_items, li_trans)
1106 ASSERT(!xlog_item_is_intent_done(lip));
1107 }
1108#endif
1109 xfs_trans_unreserve_and_mod_sb(tp);
1110 xfs_trans_unreserve_and_mod_dquots(tp);
1111
1112 if (tp->t_ticket) {
1113 xfs_log_ticket_ungrant(log, tp->t_ticket);
1114 tp->t_ticket = NULL;
1115 }
1116
1117 xfs_trans_free_items(tp, dirty);
1118 xfs_trans_free(tp);
1119}
1120
1121/*
1122 * Roll from one trans in the sequence of PERMANENT transactions to
1123 * the next: permanent transactions are only flushed out when
1124 * committed with xfs_trans_commit(), but we still want as soon
1125 * as possible to let chunks of it go to the log. So we commit the
1126 * chunk we've been working on and get a new transaction to continue.
1127 */
1128int
1129xfs_trans_roll(
1130 struct xfs_trans **tpp)
1131{
1132 struct xfs_trans *trans = *tpp;
1133 struct xfs_trans_res tres;
1134 int error;
1135
1136 trace_xfs_trans_roll(trans, _RET_IP_);
1137
1138 /*
1139 * Copy the critical parameters from one trans to the next.
1140 */
1141 tres.tr_logres = trans->t_log_res;
1142 tres.tr_logcount = trans->t_log_count;
1143
1144 *tpp = xfs_trans_dup(trans);
1145
1146 /*
1147 * Commit the current transaction.
1148 * If this commit failed, then it'd just unlock those items that
1149 * are not marked ihold. That also means that a filesystem shutdown
1150 * is in progress. The caller takes the responsibility to cancel
1151 * the duplicate transaction that gets returned.
1152 */
1153 error = __xfs_trans_commit(trans, true);
1154 if (error)
1155 return error;
1156
1157 /*
1158 * Reserve space in the log for the next transaction.
1159 * This also pushes items in the "AIL", the list of logged items,
1160 * out to disk if they are taking up space at the tail of the log
1161 * that we want to use. This requires that either nothing be locked
1162 * across this call, or that anything that is locked be logged in
1163 * the prior and the next transactions.
1164 */
1165 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1166 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1167}
1168
1169/*
1170 * Allocate an transaction, lock and join the inode to it, and reserve quota.
1171 *
1172 * The caller must ensure that the on-disk dquots attached to this inode have
1173 * already been allocated and initialized. The caller is responsible for
1174 * releasing ILOCK_EXCL if a new transaction is returned.
1175 */
1176int
1177xfs_trans_alloc_inode(
1178 struct xfs_inode *ip,
1179 struct xfs_trans_res *resv,
1180 unsigned int dblocks,
1181 unsigned int rblocks,
1182 bool force,
1183 struct xfs_trans **tpp)
1184{
1185 struct xfs_trans *tp;
1186 struct xfs_mount *mp = ip->i_mount;
1187 bool retried = false;
1188 int error;
1189
1190retry:
1191 error = xfs_trans_alloc(mp, resv, dblocks,
1192 rblocks / mp->m_sb.sb_rextsize,
1193 force ? XFS_TRANS_RESERVE : 0, &tp);
1194 if (error)
1195 return error;
1196
1197 xfs_ilock(ip, XFS_ILOCK_EXCL);
1198 xfs_trans_ijoin(tp, ip, 0);
1199
1200 error = xfs_qm_dqattach_locked(ip, false);
1201 if (error) {
1202 /* Caller should have allocated the dquots! */
1203 ASSERT(error != -ENOENT);
1204 goto out_cancel;
1205 }
1206
1207 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1208 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1209 xfs_trans_cancel(tp);
1210 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1211 xfs_blockgc_free_quota(ip, 0);
1212 retried = true;
1213 goto retry;
1214 }
1215 if (error)
1216 goto out_cancel;
1217
1218 *tpp = tp;
1219 return 0;
1220
1221out_cancel:
1222 xfs_trans_cancel(tp);
1223 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1224 return error;
1225}
1226
1227/*
1228 * Allocate an transaction in preparation for inode creation by reserving quota
1229 * against the given dquots. Callers are not required to hold any inode locks.
1230 */
1231int
1232xfs_trans_alloc_icreate(
1233 struct xfs_mount *mp,
1234 struct xfs_trans_res *resv,
1235 struct xfs_dquot *udqp,
1236 struct xfs_dquot *gdqp,
1237 struct xfs_dquot *pdqp,
1238 unsigned int dblocks,
1239 struct xfs_trans **tpp)
1240{
1241 struct xfs_trans *tp;
1242 bool retried = false;
1243 int error;
1244
1245retry:
1246 error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1247 if (error)
1248 return error;
1249
1250 error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1251 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1252 xfs_trans_cancel(tp);
1253 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1254 retried = true;
1255 goto retry;
1256 }
1257 if (error) {
1258 xfs_trans_cancel(tp);
1259 return error;
1260 }
1261
1262 *tpp = tp;
1263 return 0;
1264}
1265
1266/*
1267 * Allocate an transaction, lock and join the inode to it, and reserve quota
1268 * in preparation for inode attribute changes that include uid, gid, or prid
1269 * changes.
1270 *
1271 * The caller must ensure that the on-disk dquots attached to this inode have
1272 * already been allocated and initialized. The ILOCK will be dropped when the
1273 * transaction is committed or cancelled.
1274 */
1275int
1276xfs_trans_alloc_ichange(
1277 struct xfs_inode *ip,
1278 struct xfs_dquot *new_udqp,
1279 struct xfs_dquot *new_gdqp,
1280 struct xfs_dquot *new_pdqp,
1281 bool force,
1282 struct xfs_trans **tpp)
1283{
1284 struct xfs_trans *tp;
1285 struct xfs_mount *mp = ip->i_mount;
1286 struct xfs_dquot *udqp;
1287 struct xfs_dquot *gdqp;
1288 struct xfs_dquot *pdqp;
1289 bool retried = false;
1290 int error;
1291
1292retry:
1293 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1294 if (error)
1295 return error;
1296
1297 xfs_ilock(ip, XFS_ILOCK_EXCL);
1298 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1299
1300 error = xfs_qm_dqattach_locked(ip, false);
1301 if (error) {
1302 /* Caller should have allocated the dquots! */
1303 ASSERT(error != -ENOENT);
1304 goto out_cancel;
1305 }
1306
1307 /*
1308 * For each quota type, skip quota reservations if the inode's dquots
1309 * now match the ones that came from the caller, or the caller didn't
1310 * pass one in. The inode's dquots can change if we drop the ILOCK to
1311 * perform a blockgc scan, so we must preserve the caller's arguments.
1312 */
1313 udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1314 gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1315 pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1316 if (udqp || gdqp || pdqp) {
1317 unsigned int qflags = XFS_QMOPT_RES_REGBLKS;
1318
1319 if (force)
1320 qflags |= XFS_QMOPT_FORCE_RES;
1321
1322 /*
1323 * Reserve enough quota to handle blocks on disk and reserved
1324 * for a delayed allocation. We'll actually transfer the
1325 * delalloc reservation between dquots at chown time, even
1326 * though that part is only semi-transactional.
1327 */
1328 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1329 pdqp, ip->i_nblocks + ip->i_delayed_blks,
1330 1, qflags);
1331 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1332 xfs_trans_cancel(tp);
1333 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1334 retried = true;
1335 goto retry;
1336 }
1337 if (error)
1338 goto out_cancel;
1339 }
1340
1341 *tpp = tp;
1342 return 0;
1343
1344out_cancel:
1345 xfs_trans_cancel(tp);
1346 return error;
1347}
1348
1349/*
1350 * Allocate an transaction, lock and join the directory and child inodes to it,
1351 * and reserve quota for a directory update. If there isn't sufficient space,
1352 * @dblocks will be set to zero for a reservationless directory update and
1353 * @nospace_error will be set to a negative errno describing the space
1354 * constraint we hit.
1355 *
1356 * The caller must ensure that the on-disk dquots attached to this inode have
1357 * already been allocated and initialized. The ILOCKs will be dropped when the
1358 * transaction is committed or cancelled.
1359 */
1360int
1361xfs_trans_alloc_dir(
1362 struct xfs_inode *dp,
1363 struct xfs_trans_res *resv,
1364 struct xfs_inode *ip,
1365 unsigned int *dblocks,
1366 struct xfs_trans **tpp,
1367 int *nospace_error)
1368{
1369 struct xfs_trans *tp;
1370 struct xfs_mount *mp = ip->i_mount;
1371 unsigned int resblks;
1372 bool retried = false;
1373 int error;
1374
1375retry:
1376 *nospace_error = 0;
1377 resblks = *dblocks;
1378 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1379 if (error == -ENOSPC) {
1380 *nospace_error = error;
1381 resblks = 0;
1382 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1383 }
1384 if (error)
1385 return error;
1386
1387 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
1388
1389 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1390 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1391
1392 error = xfs_qm_dqattach_locked(dp, false);
1393 if (error) {
1394 /* Caller should have allocated the dquots! */
1395 ASSERT(error != -ENOENT);
1396 goto out_cancel;
1397 }
1398
1399 error = xfs_qm_dqattach_locked(ip, false);
1400 if (error) {
1401 /* Caller should have allocated the dquots! */
1402 ASSERT(error != -ENOENT);
1403 goto out_cancel;
1404 }
1405
1406 if (resblks == 0)
1407 goto done;
1408
1409 error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);
1410 if (error == -EDQUOT || error == -ENOSPC) {
1411 if (!retried) {
1412 xfs_trans_cancel(tp);
1413 xfs_blockgc_free_quota(dp, 0);
1414 retried = true;
1415 goto retry;
1416 }
1417
1418 *nospace_error = error;
1419 resblks = 0;
1420 error = 0;
1421 }
1422 if (error)
1423 goto out_cancel;
1424
1425done:
1426 *tpp = tp;
1427 *dblocks = resblks;
1428 return 0;
1429
1430out_cancel:
1431 xfs_trans_cancel(tp);
1432 return error;
1433}