Loading...
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_log_priv.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_extent_busy.h"
16#include "xfs_quota.h"
17#include "xfs_trans.h"
18#include "xfs_trans_priv.h"
19#include "xfs_log.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
28kmem_zone_t *xfs_trans_zone;
29
30#if defined(CONFIG_TRACEPOINTS)
31static void
32xfs_trans_trace_reservations(
33 struct xfs_mount *mp)
34{
35 struct xfs_trans_res resv;
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 xfs_log_get_max_trans_res(mp, &resv);
45 trace_xfs_trans_resv_calc(mp, -1, &resv);
46}
47#else
48# define xfs_trans_trace_reservations(mp)
49#endif
50
51/*
52 * Initialize the precomputed transaction reservation values
53 * in the mount structure.
54 */
55void
56xfs_trans_init(
57 struct xfs_mount *mp)
58{
59 xfs_trans_resv_calc(mp, M_RES(mp));
60 xfs_trans_trace_reservations(mp);
61}
62
63/*
64 * Free the transaction structure. If there is more clean up
65 * to do when the structure is freed, add it here.
66 */
67STATIC void
68xfs_trans_free(
69 struct xfs_trans *tp)
70{
71 xfs_extent_busy_sort(&tp->t_busy);
72 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
73
74 trace_xfs_trans_free(tp, _RET_IP_);
75 xfs_trans_clear_context(tp);
76 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
77 sb_end_intwrite(tp->t_mountp->m_super);
78 xfs_trans_free_dqinfo(tp);
79 kmem_cache_free(xfs_trans_zone, tp);
80}
81
82/*
83 * This is called to create a new transaction which will share the
84 * permanent log reservation of the given transaction. The remaining
85 * unused block and rt extent reservations are also inherited. This
86 * implies that the original transaction is no longer allowed to allocate
87 * blocks. Locks and log items, however, are no inherited. They must
88 * be added to the new transaction explicitly.
89 */
90STATIC struct xfs_trans *
91xfs_trans_dup(
92 struct xfs_trans *tp)
93{
94 struct xfs_trans *ntp;
95
96 trace_xfs_trans_dup(tp, _RET_IP_);
97
98 ntp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
99
100 /*
101 * Initialize the new transaction structure.
102 */
103 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
104 ntp->t_mountp = tp->t_mountp;
105 INIT_LIST_HEAD(&ntp->t_items);
106 INIT_LIST_HEAD(&ntp->t_busy);
107 INIT_LIST_HEAD(&ntp->t_dfops);
108 ntp->t_firstblock = NULLFSBLOCK;
109
110 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
111 ASSERT(tp->t_ticket != NULL);
112
113 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
114 (tp->t_flags & XFS_TRANS_RESERVE) |
115 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
116 (tp->t_flags & XFS_TRANS_RES_FDBLKS);
117 /* We gave our writer reference to the new transaction */
118 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
119 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
120
121 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
122 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
123 tp->t_blk_res = tp->t_blk_res_used;
124
125 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
126 tp->t_rtx_res = tp->t_rtx_res_used;
127
128 xfs_trans_switch_context(tp, ntp);
129
130 /* move deferred ops over to the new tp */
131 xfs_defer_move(ntp, tp);
132
133 xfs_trans_dup_dqinfo(tp, ntp);
134 return ntp;
135}
136
137/*
138 * This is called to reserve free disk blocks and log space for the
139 * given transaction. This must be done before allocating any resources
140 * within the transaction.
141 *
142 * This will return ENOSPC if there are not enough blocks available.
143 * It will sleep waiting for available log space.
144 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
145 * is used by long running transactions. If any one of the reservations
146 * fails then they will all be backed out.
147 *
148 * This does not do quota reservations. That typically is done by the
149 * caller afterwards.
150 */
151static int
152xfs_trans_reserve(
153 struct xfs_trans *tp,
154 struct xfs_trans_res *resp,
155 uint blocks,
156 uint rtextents)
157{
158 struct xfs_mount *mp = tp->t_mountp;
159 int error = 0;
160 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
161
162 /*
163 * Attempt to reserve the needed disk blocks by decrementing
164 * the number needed from the number available. This will
165 * fail if the count would go below zero.
166 */
167 if (blocks > 0) {
168 error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
169 if (error != 0)
170 return -ENOSPC;
171 tp->t_blk_res += blocks;
172 }
173
174 /*
175 * Reserve the log space needed for this transaction.
176 */
177 if (resp->tr_logres > 0) {
178 bool permanent = false;
179
180 ASSERT(tp->t_log_res == 0 ||
181 tp->t_log_res == resp->tr_logres);
182 ASSERT(tp->t_log_count == 0 ||
183 tp->t_log_count == resp->tr_logcount);
184
185 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
186 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
187 permanent = true;
188 } else {
189 ASSERT(tp->t_ticket == NULL);
190 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
191 }
192
193 if (tp->t_ticket != NULL) {
194 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
195 error = xfs_log_regrant(mp, tp->t_ticket);
196 } else {
197 error = xfs_log_reserve(mp,
198 resp->tr_logres,
199 resp->tr_logcount,
200 &tp->t_ticket, XFS_TRANSACTION,
201 permanent);
202 }
203
204 if (error)
205 goto undo_blocks;
206
207 tp->t_log_res = resp->tr_logres;
208 tp->t_log_count = resp->tr_logcount;
209 }
210
211 /*
212 * Attempt to reserve the needed realtime extents by decrementing
213 * the number needed from the number available. This will
214 * fail if the count would go below zero.
215 */
216 if (rtextents > 0) {
217 error = xfs_mod_frextents(mp, -((int64_t)rtextents));
218 if (error) {
219 error = -ENOSPC;
220 goto undo_log;
221 }
222 tp->t_rtx_res += rtextents;
223 }
224
225 return 0;
226
227 /*
228 * Error cases jump to one of these labels to undo any
229 * reservations which have already been performed.
230 */
231undo_log:
232 if (resp->tr_logres > 0) {
233 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
234 tp->t_ticket = NULL;
235 tp->t_log_res = 0;
236 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
237 }
238
239undo_blocks:
240 if (blocks > 0) {
241 xfs_mod_fdblocks(mp, (int64_t)blocks, rsvd);
242 tp->t_blk_res = 0;
243 }
244 return error;
245}
246
247int
248xfs_trans_alloc(
249 struct xfs_mount *mp,
250 struct xfs_trans_res *resp,
251 uint blocks,
252 uint rtextents,
253 uint flags,
254 struct xfs_trans **tpp)
255{
256 struct xfs_trans *tp;
257 bool want_retry = true;
258 int error;
259
260 /*
261 * Allocate the handle before we do our freeze accounting and setting up
262 * GFP_NOFS allocation context so that we avoid lockdep false positives
263 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
264 */
265retry:
266 tp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
267 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
268 sb_start_intwrite(mp->m_super);
269 xfs_trans_set_context(tp);
270
271 /*
272 * Zero-reservation ("empty") transactions can't modify anything, so
273 * they're allowed to run while we're frozen.
274 */
275 WARN_ON(resp->tr_logres > 0 &&
276 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
277 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
278 xfs_sb_version_haslazysbcount(&mp->m_sb));
279
280 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
281 tp->t_flags = flags;
282 tp->t_mountp = mp;
283 INIT_LIST_HEAD(&tp->t_items);
284 INIT_LIST_HEAD(&tp->t_busy);
285 INIT_LIST_HEAD(&tp->t_dfops);
286 tp->t_firstblock = NULLFSBLOCK;
287
288 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
289 if (error == -ENOSPC && want_retry) {
290 xfs_trans_cancel(tp);
291
292 /*
293 * We weren't able to reserve enough space for the transaction.
294 * Flush the other speculative space allocations to free space.
295 * Do not perform a synchronous scan because callers can hold
296 * other locks.
297 */
298 error = xfs_blockgc_free_space(mp, NULL);
299 if (error)
300 return error;
301
302 want_retry = false;
303 goto retry;
304 }
305 if (error) {
306 xfs_trans_cancel(tp);
307 return error;
308 }
309
310 trace_xfs_trans_alloc(tp, _RET_IP_);
311
312 *tpp = tp;
313 return 0;
314}
315
316/*
317 * Create an empty transaction with no reservation. This is a defensive
318 * mechanism for routines that query metadata without actually modifying them --
319 * if the metadata being queried is somehow cross-linked (think a btree block
320 * pointer that points higher in the tree), we risk deadlock. However, blocks
321 * grabbed as part of a transaction can be re-grabbed. The verifiers will
322 * notice the corrupt block and the operation will fail back to userspace
323 * without deadlocking.
324 *
325 * Note the zero-length reservation; this transaction MUST be cancelled without
326 * any dirty data.
327 *
328 * Callers should obtain freeze protection to avoid a conflict with fs freezing
329 * where we can be grabbing buffers at the same time that freeze is trying to
330 * drain the buffer LRU list.
331 */
332int
333xfs_trans_alloc_empty(
334 struct xfs_mount *mp,
335 struct xfs_trans **tpp)
336{
337 struct xfs_trans_res resv = {0};
338
339 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
340}
341
342/*
343 * Record the indicated change to the given field for application
344 * to the file system's superblock when the transaction commits.
345 * For now, just store the change in the transaction structure.
346 *
347 * Mark the transaction structure to indicate that the superblock
348 * needs to be updated before committing.
349 *
350 * Because we may not be keeping track of allocated/free inodes and
351 * used filesystem blocks in the superblock, we do not mark the
352 * superblock dirty in this transaction if we modify these fields.
353 * We still need to update the transaction deltas so that they get
354 * applied to the incore superblock, but we don't want them to
355 * cause the superblock to get locked and logged if these are the
356 * only fields in the superblock that the transaction modifies.
357 */
358void
359xfs_trans_mod_sb(
360 xfs_trans_t *tp,
361 uint field,
362 int64_t delta)
363{
364 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
365 xfs_mount_t *mp = tp->t_mountp;
366
367 switch (field) {
368 case XFS_TRANS_SB_ICOUNT:
369 tp->t_icount_delta += delta;
370 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
371 flags &= ~XFS_TRANS_SB_DIRTY;
372 break;
373 case XFS_TRANS_SB_IFREE:
374 tp->t_ifree_delta += delta;
375 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
376 flags &= ~XFS_TRANS_SB_DIRTY;
377 break;
378 case XFS_TRANS_SB_FDBLOCKS:
379 /*
380 * Track the number of blocks allocated in the transaction.
381 * Make sure it does not exceed the number reserved. If so,
382 * shutdown as this can lead to accounting inconsistency.
383 */
384 if (delta < 0) {
385 tp->t_blk_res_used += (uint)-delta;
386 if (tp->t_blk_res_used > tp->t_blk_res)
387 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
388 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
389 int64_t blkres_delta;
390
391 /*
392 * Return freed blocks directly to the reservation
393 * instead of the global pool, being careful not to
394 * overflow the trans counter. This is used to preserve
395 * reservation across chains of transaction rolls that
396 * repeatedly free and allocate blocks.
397 */
398 blkres_delta = min_t(int64_t, delta,
399 UINT_MAX - tp->t_blk_res);
400 tp->t_blk_res += blkres_delta;
401 delta -= blkres_delta;
402 }
403 tp->t_fdblocks_delta += delta;
404 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
405 flags &= ~XFS_TRANS_SB_DIRTY;
406 break;
407 case XFS_TRANS_SB_RES_FDBLOCKS:
408 /*
409 * The allocation has already been applied to the
410 * in-core superblock's counter. This should only
411 * be applied to the on-disk superblock.
412 */
413 tp->t_res_fdblocks_delta += delta;
414 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
415 flags &= ~XFS_TRANS_SB_DIRTY;
416 break;
417 case XFS_TRANS_SB_FREXTENTS:
418 /*
419 * Track the number of blocks allocated in the
420 * transaction. Make sure it does not exceed the
421 * number reserved.
422 */
423 if (delta < 0) {
424 tp->t_rtx_res_used += (uint)-delta;
425 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
426 }
427 tp->t_frextents_delta += delta;
428 break;
429 case XFS_TRANS_SB_RES_FREXTENTS:
430 /*
431 * The allocation has already been applied to the
432 * in-core superblock's counter. This should only
433 * be applied to the on-disk superblock.
434 */
435 ASSERT(delta < 0);
436 tp->t_res_frextents_delta += delta;
437 break;
438 case XFS_TRANS_SB_DBLOCKS:
439 tp->t_dblocks_delta += delta;
440 break;
441 case XFS_TRANS_SB_AGCOUNT:
442 ASSERT(delta > 0);
443 tp->t_agcount_delta += delta;
444 break;
445 case XFS_TRANS_SB_IMAXPCT:
446 tp->t_imaxpct_delta += delta;
447 break;
448 case XFS_TRANS_SB_REXTSIZE:
449 tp->t_rextsize_delta += delta;
450 break;
451 case XFS_TRANS_SB_RBMBLOCKS:
452 tp->t_rbmblocks_delta += delta;
453 break;
454 case XFS_TRANS_SB_RBLOCKS:
455 tp->t_rblocks_delta += delta;
456 break;
457 case XFS_TRANS_SB_REXTENTS:
458 tp->t_rextents_delta += delta;
459 break;
460 case XFS_TRANS_SB_REXTSLOG:
461 tp->t_rextslog_delta += delta;
462 break;
463 default:
464 ASSERT(0);
465 return;
466 }
467
468 tp->t_flags |= flags;
469}
470
471/*
472 * xfs_trans_apply_sb_deltas() is called from the commit code
473 * to bring the superblock buffer into the current transaction
474 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
475 *
476 * For now we just look at each field allowed to change and change
477 * it if necessary.
478 */
479STATIC void
480xfs_trans_apply_sb_deltas(
481 xfs_trans_t *tp)
482{
483 xfs_dsb_t *sbp;
484 struct xfs_buf *bp;
485 int whole = 0;
486
487 bp = xfs_trans_getsb(tp);
488 sbp = bp->b_addr;
489
490 /*
491 * Only update the superblock counters if we are logging them
492 */
493 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
494 if (tp->t_icount_delta)
495 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
496 if (tp->t_ifree_delta)
497 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
498 if (tp->t_fdblocks_delta)
499 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
500 if (tp->t_res_fdblocks_delta)
501 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
502 }
503
504 if (tp->t_frextents_delta)
505 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
506 if (tp->t_res_frextents_delta)
507 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
508
509 if (tp->t_dblocks_delta) {
510 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
511 whole = 1;
512 }
513 if (tp->t_agcount_delta) {
514 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
515 whole = 1;
516 }
517 if (tp->t_imaxpct_delta) {
518 sbp->sb_imax_pct += tp->t_imaxpct_delta;
519 whole = 1;
520 }
521 if (tp->t_rextsize_delta) {
522 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
523 whole = 1;
524 }
525 if (tp->t_rbmblocks_delta) {
526 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
527 whole = 1;
528 }
529 if (tp->t_rblocks_delta) {
530 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
531 whole = 1;
532 }
533 if (tp->t_rextents_delta) {
534 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
535 whole = 1;
536 }
537 if (tp->t_rextslog_delta) {
538 sbp->sb_rextslog += tp->t_rextslog_delta;
539 whole = 1;
540 }
541
542 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
543 if (whole)
544 /*
545 * Log the whole thing, the fields are noncontiguous.
546 */
547 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
548 else
549 /*
550 * Since all the modifiable fields are contiguous, we
551 * can get away with this.
552 */
553 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
554 offsetof(xfs_dsb_t, sb_frextents) +
555 sizeof(sbp->sb_frextents) - 1);
556}
557
558/*
559 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
560 * apply superblock counter changes to the in-core superblock. The
561 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
562 * applied to the in-core superblock. The idea is that that has already been
563 * done.
564 *
565 * If we are not logging superblock counters, then the inode allocated/free and
566 * used block counts are not updated in the on disk superblock. In this case,
567 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
568 * still need to update the incore superblock with the changes.
569 *
570 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
571 * so we don't need to take the counter lock on every update.
572 */
573#define XFS_ICOUNT_BATCH 128
574
575void
576xfs_trans_unreserve_and_mod_sb(
577 struct xfs_trans *tp)
578{
579 struct xfs_mount *mp = tp->t_mountp;
580 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
581 int64_t blkdelta = 0;
582 int64_t rtxdelta = 0;
583 int64_t idelta = 0;
584 int64_t ifreedelta = 0;
585 int error;
586
587 /* calculate deltas */
588 if (tp->t_blk_res > 0)
589 blkdelta = tp->t_blk_res;
590 if ((tp->t_fdblocks_delta != 0) &&
591 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
592 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
593 blkdelta += tp->t_fdblocks_delta;
594
595 if (tp->t_rtx_res > 0)
596 rtxdelta = tp->t_rtx_res;
597 if ((tp->t_frextents_delta != 0) &&
598 (tp->t_flags & XFS_TRANS_SB_DIRTY))
599 rtxdelta += tp->t_frextents_delta;
600
601 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
602 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
603 idelta = tp->t_icount_delta;
604 ifreedelta = tp->t_ifree_delta;
605 }
606
607 /* apply the per-cpu counters */
608 if (blkdelta) {
609 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
610 ASSERT(!error);
611 }
612
613 if (idelta)
614 percpu_counter_add_batch(&mp->m_icount, idelta,
615 XFS_ICOUNT_BATCH);
616
617 if (ifreedelta)
618 percpu_counter_add(&mp->m_ifree, ifreedelta);
619
620 if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
621 return;
622
623 /* apply remaining deltas */
624 spin_lock(&mp->m_sb_lock);
625 mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
626 mp->m_sb.sb_icount += idelta;
627 mp->m_sb.sb_ifree += ifreedelta;
628 mp->m_sb.sb_frextents += rtxdelta;
629 mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
630 mp->m_sb.sb_agcount += tp->t_agcount_delta;
631 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
632 mp->m_sb.sb_rextsize += tp->t_rextsize_delta;
633 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
634 mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
635 mp->m_sb.sb_rextents += tp->t_rextents_delta;
636 mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
637 spin_unlock(&mp->m_sb_lock);
638
639 /*
640 * Debug checks outside of the spinlock so they don't lock up the
641 * machine if they fail.
642 */
643 ASSERT(mp->m_sb.sb_imax_pct >= 0);
644 ASSERT(mp->m_sb.sb_rextslog >= 0);
645 return;
646}
647
648/* Add the given log item to the transaction's list of log items. */
649void
650xfs_trans_add_item(
651 struct xfs_trans *tp,
652 struct xfs_log_item *lip)
653{
654 ASSERT(lip->li_mountp == tp->t_mountp);
655 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
656 ASSERT(list_empty(&lip->li_trans));
657 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
658
659 list_add_tail(&lip->li_trans, &tp->t_items);
660 trace_xfs_trans_add_item(tp, _RET_IP_);
661}
662
663/*
664 * Unlink the log item from the transaction. the log item is no longer
665 * considered dirty in this transaction, as the linked transaction has
666 * finished, either by abort or commit completion.
667 */
668void
669xfs_trans_del_item(
670 struct xfs_log_item *lip)
671{
672 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
673 list_del_init(&lip->li_trans);
674}
675
676/* Detach and unlock all of the items in a transaction */
677static void
678xfs_trans_free_items(
679 struct xfs_trans *tp,
680 bool abort)
681{
682 struct xfs_log_item *lip, *next;
683
684 trace_xfs_trans_free_items(tp, _RET_IP_);
685
686 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
687 xfs_trans_del_item(lip);
688 if (abort)
689 set_bit(XFS_LI_ABORTED, &lip->li_flags);
690 if (lip->li_ops->iop_release)
691 lip->li_ops->iop_release(lip);
692 }
693}
694
695static inline void
696xfs_log_item_batch_insert(
697 struct xfs_ail *ailp,
698 struct xfs_ail_cursor *cur,
699 struct xfs_log_item **log_items,
700 int nr_items,
701 xfs_lsn_t commit_lsn)
702{
703 int i;
704
705 spin_lock(&ailp->ail_lock);
706 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
707 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
708
709 for (i = 0; i < nr_items; i++) {
710 struct xfs_log_item *lip = log_items[i];
711
712 if (lip->li_ops->iop_unpin)
713 lip->li_ops->iop_unpin(lip, 0);
714 }
715}
716
717/*
718 * Bulk operation version of xfs_trans_committed that takes a log vector of
719 * items to insert into the AIL. This uses bulk AIL insertion techniques to
720 * minimise lock traffic.
721 *
722 * If we are called with the aborted flag set, it is because a log write during
723 * a CIL checkpoint commit has failed. In this case, all the items in the
724 * checkpoint have already gone through iop_committed and iop_committing, which
725 * means that checkpoint commit abort handling is treated exactly the same
726 * as an iclog write error even though we haven't started any IO yet. Hence in
727 * this case all we need to do is iop_committed processing, followed by an
728 * iop_unpin(aborted) call.
729 *
730 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
731 * at the end of the AIL, the insert cursor avoids the need to walk
732 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
733 * call. This saves a lot of needless list walking and is a net win, even
734 * though it slightly increases that amount of AIL lock traffic to set it up
735 * and tear it down.
736 */
737void
738xfs_trans_committed_bulk(
739 struct xfs_ail *ailp,
740 struct xfs_log_vec *log_vector,
741 xfs_lsn_t commit_lsn,
742 bool aborted)
743{
744#define LOG_ITEM_BATCH_SIZE 32
745 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
746 struct xfs_log_vec *lv;
747 struct xfs_ail_cursor cur;
748 int i = 0;
749
750 spin_lock(&ailp->ail_lock);
751 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
752 spin_unlock(&ailp->ail_lock);
753
754 /* unpin all the log items */
755 for (lv = log_vector; lv; lv = lv->lv_next ) {
756 struct xfs_log_item *lip = lv->lv_item;
757 xfs_lsn_t item_lsn;
758
759 if (aborted)
760 set_bit(XFS_LI_ABORTED, &lip->li_flags);
761
762 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
763 lip->li_ops->iop_release(lip);
764 continue;
765 }
766
767 if (lip->li_ops->iop_committed)
768 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
769 else
770 item_lsn = commit_lsn;
771
772 /* item_lsn of -1 means the item needs no further processing */
773 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
774 continue;
775
776 /*
777 * if we are aborting the operation, no point in inserting the
778 * object into the AIL as we are in a shutdown situation.
779 */
780 if (aborted) {
781 ASSERT(XFS_FORCED_SHUTDOWN(ailp->ail_mount));
782 if (lip->li_ops->iop_unpin)
783 lip->li_ops->iop_unpin(lip, 1);
784 continue;
785 }
786
787 if (item_lsn != commit_lsn) {
788
789 /*
790 * Not a bulk update option due to unusual item_lsn.
791 * Push into AIL immediately, rechecking the lsn once
792 * we have the ail lock. Then unpin the item. This does
793 * not affect the AIL cursor the bulk insert path is
794 * using.
795 */
796 spin_lock(&ailp->ail_lock);
797 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
798 xfs_trans_ail_update(ailp, lip, item_lsn);
799 else
800 spin_unlock(&ailp->ail_lock);
801 if (lip->li_ops->iop_unpin)
802 lip->li_ops->iop_unpin(lip, 0);
803 continue;
804 }
805
806 /* Item is a candidate for bulk AIL insert. */
807 log_items[i++] = lv->lv_item;
808 if (i >= LOG_ITEM_BATCH_SIZE) {
809 xfs_log_item_batch_insert(ailp, &cur, log_items,
810 LOG_ITEM_BATCH_SIZE, commit_lsn);
811 i = 0;
812 }
813 }
814
815 /* make sure we insert the remainder! */
816 if (i)
817 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
818
819 spin_lock(&ailp->ail_lock);
820 xfs_trans_ail_cursor_done(&cur);
821 spin_unlock(&ailp->ail_lock);
822}
823
824/*
825 * Commit the given transaction to the log.
826 *
827 * XFS disk error handling mechanism is not based on a typical
828 * transaction abort mechanism. Logically after the filesystem
829 * gets marked 'SHUTDOWN', we can't let any new transactions
830 * be durable - ie. committed to disk - because some metadata might
831 * be inconsistent. In such cases, this returns an error, and the
832 * caller may assume that all locked objects joined to the transaction
833 * have already been unlocked as if the commit had succeeded.
834 * Do not reference the transaction structure after this call.
835 */
836static int
837__xfs_trans_commit(
838 struct xfs_trans *tp,
839 bool regrant)
840{
841 struct xfs_mount *mp = tp->t_mountp;
842 xfs_csn_t commit_seq = 0;
843 int error = 0;
844 int sync = tp->t_flags & XFS_TRANS_SYNC;
845
846 trace_xfs_trans_commit(tp, _RET_IP_);
847
848 /*
849 * Finish deferred items on final commit. Only permanent transactions
850 * should ever have deferred ops.
851 */
852 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
853 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
854 if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
855 error = xfs_defer_finish_noroll(&tp);
856 if (error)
857 goto out_unreserve;
858 }
859
860 /*
861 * If there is nothing to be logged by the transaction,
862 * then unlock all of the items associated with the
863 * transaction and free the transaction structure.
864 * Also make sure to return any reserved blocks to
865 * the free pool.
866 */
867 if (!(tp->t_flags & XFS_TRANS_DIRTY))
868 goto out_unreserve;
869
870 if (XFS_FORCED_SHUTDOWN(mp)) {
871 error = -EIO;
872 goto out_unreserve;
873 }
874
875 ASSERT(tp->t_ticket != NULL);
876
877 /*
878 * If we need to update the superblock, then do it now.
879 */
880 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
881 xfs_trans_apply_sb_deltas(tp);
882 xfs_trans_apply_dquot_deltas(tp);
883
884 xlog_cil_commit(mp->m_log, tp, &commit_seq, regrant);
885
886 xfs_trans_free(tp);
887
888 /*
889 * If the transaction needs to be synchronous, then force the
890 * log out now and wait for it.
891 */
892 if (sync) {
893 error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
894 XFS_STATS_INC(mp, xs_trans_sync);
895 } else {
896 XFS_STATS_INC(mp, xs_trans_async);
897 }
898
899 return error;
900
901out_unreserve:
902 xfs_trans_unreserve_and_mod_sb(tp);
903
904 /*
905 * It is indeed possible for the transaction to be not dirty but
906 * the dqinfo portion to be. All that means is that we have some
907 * (non-persistent) quota reservations that need to be unreserved.
908 */
909 xfs_trans_unreserve_and_mod_dquots(tp);
910 if (tp->t_ticket) {
911 if (regrant && !XLOG_FORCED_SHUTDOWN(mp->m_log))
912 xfs_log_ticket_regrant(mp->m_log, tp->t_ticket);
913 else
914 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
915 tp->t_ticket = NULL;
916 }
917 xfs_trans_free_items(tp, !!error);
918 xfs_trans_free(tp);
919
920 XFS_STATS_INC(mp, xs_trans_empty);
921 return error;
922}
923
924int
925xfs_trans_commit(
926 struct xfs_trans *tp)
927{
928 return __xfs_trans_commit(tp, false);
929}
930
931/*
932 * Unlock all of the transaction's items and free the transaction.
933 * The transaction must not have modified any of its items, because
934 * there is no way to restore them to their previous state.
935 *
936 * If the transaction has made a log reservation, make sure to release
937 * it as well.
938 */
939void
940xfs_trans_cancel(
941 struct xfs_trans *tp)
942{
943 struct xfs_mount *mp = tp->t_mountp;
944 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
945
946 trace_xfs_trans_cancel(tp, _RET_IP_);
947
948 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
949 xfs_defer_cancel(tp);
950
951 /*
952 * See if the caller is relying on us to shut down the
953 * filesystem. This happens in paths where we detect
954 * corruption and decide to give up.
955 */
956 if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
957 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
958 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
959 }
960#ifdef DEBUG
961 if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
962 struct xfs_log_item *lip;
963
964 list_for_each_entry(lip, &tp->t_items, li_trans)
965 ASSERT(!xlog_item_is_intent_done(lip));
966 }
967#endif
968 xfs_trans_unreserve_and_mod_sb(tp);
969 xfs_trans_unreserve_and_mod_dquots(tp);
970
971 if (tp->t_ticket) {
972 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
973 tp->t_ticket = NULL;
974 }
975
976 xfs_trans_free_items(tp, dirty);
977 xfs_trans_free(tp);
978}
979
980/*
981 * Roll from one trans in the sequence of PERMANENT transactions to
982 * the next: permanent transactions are only flushed out when
983 * committed with xfs_trans_commit(), but we still want as soon
984 * as possible to let chunks of it go to the log. So we commit the
985 * chunk we've been working on and get a new transaction to continue.
986 */
987int
988xfs_trans_roll(
989 struct xfs_trans **tpp)
990{
991 struct xfs_trans *trans = *tpp;
992 struct xfs_trans_res tres;
993 int error;
994
995 trace_xfs_trans_roll(trans, _RET_IP_);
996
997 /*
998 * Copy the critical parameters from one trans to the next.
999 */
1000 tres.tr_logres = trans->t_log_res;
1001 tres.tr_logcount = trans->t_log_count;
1002
1003 *tpp = xfs_trans_dup(trans);
1004
1005 /*
1006 * Commit the current transaction.
1007 * If this commit failed, then it'd just unlock those items that
1008 * are not marked ihold. That also means that a filesystem shutdown
1009 * is in progress. The caller takes the responsibility to cancel
1010 * the duplicate transaction that gets returned.
1011 */
1012 error = __xfs_trans_commit(trans, true);
1013 if (error)
1014 return error;
1015
1016 /*
1017 * Reserve space in the log for the next transaction.
1018 * This also pushes items in the "AIL", the list of logged items,
1019 * out to disk if they are taking up space at the tail of the log
1020 * that we want to use. This requires that either nothing be locked
1021 * across this call, or that anything that is locked be logged in
1022 * the prior and the next transactions.
1023 */
1024 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1025 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1026}
1027
1028/*
1029 * Allocate an transaction, lock and join the inode to it, and reserve quota.
1030 *
1031 * The caller must ensure that the on-disk dquots attached to this inode have
1032 * already been allocated and initialized. The caller is responsible for
1033 * releasing ILOCK_EXCL if a new transaction is returned.
1034 */
1035int
1036xfs_trans_alloc_inode(
1037 struct xfs_inode *ip,
1038 struct xfs_trans_res *resv,
1039 unsigned int dblocks,
1040 unsigned int rblocks,
1041 bool force,
1042 struct xfs_trans **tpp)
1043{
1044 struct xfs_trans *tp;
1045 struct xfs_mount *mp = ip->i_mount;
1046 bool retried = false;
1047 int error;
1048
1049retry:
1050 error = xfs_trans_alloc(mp, resv, dblocks,
1051 rblocks / mp->m_sb.sb_rextsize,
1052 force ? XFS_TRANS_RESERVE : 0, &tp);
1053 if (error)
1054 return error;
1055
1056 xfs_ilock(ip, XFS_ILOCK_EXCL);
1057 xfs_trans_ijoin(tp, ip, 0);
1058
1059 error = xfs_qm_dqattach_locked(ip, false);
1060 if (error) {
1061 /* Caller should have allocated the dquots! */
1062 ASSERT(error != -ENOENT);
1063 goto out_cancel;
1064 }
1065
1066 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1067 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1068 xfs_trans_cancel(tp);
1069 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1070 xfs_blockgc_free_quota(ip, 0);
1071 retried = true;
1072 goto retry;
1073 }
1074 if (error)
1075 goto out_cancel;
1076
1077 *tpp = tp;
1078 return 0;
1079
1080out_cancel:
1081 xfs_trans_cancel(tp);
1082 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1083 return error;
1084}
1085
1086/*
1087 * Allocate an transaction in preparation for inode creation by reserving quota
1088 * against the given dquots. Callers are not required to hold any inode locks.
1089 */
1090int
1091xfs_trans_alloc_icreate(
1092 struct xfs_mount *mp,
1093 struct xfs_trans_res *resv,
1094 struct xfs_dquot *udqp,
1095 struct xfs_dquot *gdqp,
1096 struct xfs_dquot *pdqp,
1097 unsigned int dblocks,
1098 struct xfs_trans **tpp)
1099{
1100 struct xfs_trans *tp;
1101 bool retried = false;
1102 int error;
1103
1104retry:
1105 error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1106 if (error)
1107 return error;
1108
1109 error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1110 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1111 xfs_trans_cancel(tp);
1112 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1113 retried = true;
1114 goto retry;
1115 }
1116 if (error) {
1117 xfs_trans_cancel(tp);
1118 return error;
1119 }
1120
1121 *tpp = tp;
1122 return 0;
1123}
1124
1125/*
1126 * Allocate an transaction, lock and join the inode to it, and reserve quota
1127 * in preparation for inode attribute changes that include uid, gid, or prid
1128 * changes.
1129 *
1130 * The caller must ensure that the on-disk dquots attached to this inode have
1131 * already been allocated and initialized. The ILOCK will be dropped when the
1132 * transaction is committed or cancelled.
1133 */
1134int
1135xfs_trans_alloc_ichange(
1136 struct xfs_inode *ip,
1137 struct xfs_dquot *new_udqp,
1138 struct xfs_dquot *new_gdqp,
1139 struct xfs_dquot *new_pdqp,
1140 bool force,
1141 struct xfs_trans **tpp)
1142{
1143 struct xfs_trans *tp;
1144 struct xfs_mount *mp = ip->i_mount;
1145 struct xfs_dquot *udqp;
1146 struct xfs_dquot *gdqp;
1147 struct xfs_dquot *pdqp;
1148 bool retried = false;
1149 int error;
1150
1151retry:
1152 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1153 if (error)
1154 return error;
1155
1156 xfs_ilock(ip, XFS_ILOCK_EXCL);
1157 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1158
1159 error = xfs_qm_dqattach_locked(ip, false);
1160 if (error) {
1161 /* Caller should have allocated the dquots! */
1162 ASSERT(error != -ENOENT);
1163 goto out_cancel;
1164 }
1165
1166 /*
1167 * For each quota type, skip quota reservations if the inode's dquots
1168 * now match the ones that came from the caller, or the caller didn't
1169 * pass one in. The inode's dquots can change if we drop the ILOCK to
1170 * perform a blockgc scan, so we must preserve the caller's arguments.
1171 */
1172 udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1173 gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1174 pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1175 if (udqp || gdqp || pdqp) {
1176 unsigned int qflags = XFS_QMOPT_RES_REGBLKS;
1177
1178 if (force)
1179 qflags |= XFS_QMOPT_FORCE_RES;
1180
1181 /*
1182 * Reserve enough quota to handle blocks on disk and reserved
1183 * for a delayed allocation. We'll actually transfer the
1184 * delalloc reservation between dquots at chown time, even
1185 * though that part is only semi-transactional.
1186 */
1187 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1188 pdqp, ip->i_nblocks + ip->i_delayed_blks,
1189 1, qflags);
1190 if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1191 xfs_trans_cancel(tp);
1192 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1193 retried = true;
1194 goto retry;
1195 }
1196 if (error)
1197 goto out_cancel;
1198 }
1199
1200 *tpp = tp;
1201 return 0;
1202
1203out_cancel:
1204 xfs_trans_cancel(tp);
1205 return error;
1206}
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_log_priv.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_extent_busy.h"
16#include "xfs_quota.h"
17#include "xfs_trans.h"
18#include "xfs_trans_priv.h"
19#include "xfs_log.h"
20#include "xfs_trace.h"
21#include "xfs_error.h"
22#include "xfs_defer.h"
23
24kmem_zone_t *xfs_trans_zone;
25
26#if defined(CONFIG_TRACEPOINTS)
27static void
28xfs_trans_trace_reservations(
29 struct xfs_mount *mp)
30{
31 struct xfs_trans_res resv;
32 struct xfs_trans_res *res;
33 struct xfs_trans_res *end_res;
34 int i;
35
36 res = (struct xfs_trans_res *)M_RES(mp);
37 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
38 for (i = 0; res < end_res; i++, res++)
39 trace_xfs_trans_resv_calc(mp, i, res);
40 xfs_log_get_max_trans_res(mp, &resv);
41 trace_xfs_trans_resv_calc(mp, -1, &resv);
42}
43#else
44# define xfs_trans_trace_reservations(mp)
45#endif
46
47/*
48 * Initialize the precomputed transaction reservation values
49 * in the mount structure.
50 */
51void
52xfs_trans_init(
53 struct xfs_mount *mp)
54{
55 xfs_trans_resv_calc(mp, M_RES(mp));
56 xfs_trans_trace_reservations(mp);
57}
58
59/*
60 * Free the transaction structure. If there is more clean up
61 * to do when the structure is freed, add it here.
62 */
63STATIC void
64xfs_trans_free(
65 struct xfs_trans *tp)
66{
67 xfs_extent_busy_sort(&tp->t_busy);
68 xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
69
70 trace_xfs_trans_free(tp, _RET_IP_);
71 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
72 sb_end_intwrite(tp->t_mountp->m_super);
73 xfs_trans_free_dqinfo(tp);
74 kmem_cache_free(xfs_trans_zone, tp);
75}
76
77/*
78 * This is called to create a new transaction which will share the
79 * permanent log reservation of the given transaction. The remaining
80 * unused block and rt extent reservations are also inherited. This
81 * implies that the original transaction is no longer allowed to allocate
82 * blocks. Locks and log items, however, are no inherited. They must
83 * be added to the new transaction explicitly.
84 */
85STATIC struct xfs_trans *
86xfs_trans_dup(
87 struct xfs_trans *tp)
88{
89 struct xfs_trans *ntp;
90
91 trace_xfs_trans_dup(tp, _RET_IP_);
92
93 ntp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
94
95 /*
96 * Initialize the new transaction structure.
97 */
98 ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
99 ntp->t_mountp = tp->t_mountp;
100 INIT_LIST_HEAD(&ntp->t_items);
101 INIT_LIST_HEAD(&ntp->t_busy);
102 INIT_LIST_HEAD(&ntp->t_dfops);
103 ntp->t_firstblock = NULLFSBLOCK;
104
105 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
106 ASSERT(tp->t_ticket != NULL);
107
108 ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
109 (tp->t_flags & XFS_TRANS_RESERVE) |
110 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
111 (tp->t_flags & XFS_TRANS_RES_FDBLKS);
112 /* We gave our writer reference to the new transaction */
113 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
114 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
115
116 ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
117 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
118 tp->t_blk_res = tp->t_blk_res_used;
119
120 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
121 tp->t_rtx_res = tp->t_rtx_res_used;
122 ntp->t_pflags = tp->t_pflags;
123
124 /* move deferred ops over to the new tp */
125 xfs_defer_move(ntp, tp);
126
127 xfs_trans_dup_dqinfo(tp, ntp);
128 return ntp;
129}
130
131/*
132 * This is called to reserve free disk blocks and log space for the
133 * given transaction. This must be done before allocating any resources
134 * within the transaction.
135 *
136 * This will return ENOSPC if there are not enough blocks available.
137 * It will sleep waiting for available log space.
138 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
139 * is used by long running transactions. If any one of the reservations
140 * fails then they will all be backed out.
141 *
142 * This does not do quota reservations. That typically is done by the
143 * caller afterwards.
144 */
145static int
146xfs_trans_reserve(
147 struct xfs_trans *tp,
148 struct xfs_trans_res *resp,
149 uint blocks,
150 uint rtextents)
151{
152 struct xfs_mount *mp = tp->t_mountp;
153 int error = 0;
154 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
155
156 /* Mark this thread as being in a transaction */
157 current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
158
159 /*
160 * Attempt to reserve the needed disk blocks by decrementing
161 * the number needed from the number available. This will
162 * fail if the count would go below zero.
163 */
164 if (blocks > 0) {
165 error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
166 if (error != 0) {
167 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
168 return -ENOSPC;
169 }
170 tp->t_blk_res += blocks;
171 }
172
173 /*
174 * Reserve the log space needed for this transaction.
175 */
176 if (resp->tr_logres > 0) {
177 bool permanent = false;
178
179 ASSERT(tp->t_log_res == 0 ||
180 tp->t_log_res == resp->tr_logres);
181 ASSERT(tp->t_log_count == 0 ||
182 tp->t_log_count == resp->tr_logcount);
183
184 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
185 tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
186 permanent = true;
187 } else {
188 ASSERT(tp->t_ticket == NULL);
189 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
190 }
191
192 if (tp->t_ticket != NULL) {
193 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
194 error = xfs_log_regrant(mp, tp->t_ticket);
195 } else {
196 error = xfs_log_reserve(mp,
197 resp->tr_logres,
198 resp->tr_logcount,
199 &tp->t_ticket, XFS_TRANSACTION,
200 permanent);
201 }
202
203 if (error)
204 goto undo_blocks;
205
206 tp->t_log_res = resp->tr_logres;
207 tp->t_log_count = resp->tr_logcount;
208 }
209
210 /*
211 * Attempt to reserve the needed realtime extents by decrementing
212 * the number needed from the number available. This will
213 * fail if the count would go below zero.
214 */
215 if (rtextents > 0) {
216 error = xfs_mod_frextents(mp, -((int64_t)rtextents));
217 if (error) {
218 error = -ENOSPC;
219 goto undo_log;
220 }
221 tp->t_rtx_res += rtextents;
222 }
223
224 return 0;
225
226 /*
227 * Error cases jump to one of these labels to undo any
228 * reservations which have already been performed.
229 */
230undo_log:
231 if (resp->tr_logres > 0) {
232 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
233 tp->t_ticket = NULL;
234 tp->t_log_res = 0;
235 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
236 }
237
238undo_blocks:
239 if (blocks > 0) {
240 xfs_mod_fdblocks(mp, (int64_t)blocks, rsvd);
241 tp->t_blk_res = 0;
242 }
243
244 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
245
246 return error;
247}
248
249int
250xfs_trans_alloc(
251 struct xfs_mount *mp,
252 struct xfs_trans_res *resp,
253 uint blocks,
254 uint rtextents,
255 uint flags,
256 struct xfs_trans **tpp)
257{
258 struct xfs_trans *tp;
259 int error;
260
261 /*
262 * Allocate the handle before we do our freeze accounting and setting up
263 * GFP_NOFS allocation context so that we avoid lockdep false positives
264 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
265 */
266 tp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
267 if (!(flags & XFS_TRANS_NO_WRITECOUNT))
268 sb_start_intwrite(mp->m_super);
269
270 /*
271 * Zero-reservation ("empty") transactions can't modify anything, so
272 * they're allowed to run while we're frozen.
273 */
274 WARN_ON(resp->tr_logres > 0 &&
275 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
276 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
277 xfs_sb_version_haslazysbcount(&mp->m_sb));
278
279 tp->t_magic = XFS_TRANS_HEADER_MAGIC;
280 tp->t_flags = flags;
281 tp->t_mountp = mp;
282 INIT_LIST_HEAD(&tp->t_items);
283 INIT_LIST_HEAD(&tp->t_busy);
284 INIT_LIST_HEAD(&tp->t_dfops);
285 tp->t_firstblock = NULLFSBLOCK;
286
287 error = xfs_trans_reserve(tp, resp, blocks, rtextents);
288 if (error) {
289 xfs_trans_cancel(tp);
290 return error;
291 }
292
293 trace_xfs_trans_alloc(tp, _RET_IP_);
294
295 *tpp = tp;
296 return 0;
297}
298
299/*
300 * Create an empty transaction with no reservation. This is a defensive
301 * mechanism for routines that query metadata without actually modifying them --
302 * if the metadata being queried is somehow cross-linked (think a btree block
303 * pointer that points higher in the tree), we risk deadlock. However, blocks
304 * grabbed as part of a transaction can be re-grabbed. The verifiers will
305 * notice the corrupt block and the operation will fail back to userspace
306 * without deadlocking.
307 *
308 * Note the zero-length reservation; this transaction MUST be cancelled without
309 * any dirty data.
310 *
311 * Callers should obtain freeze protection to avoid a conflict with fs freezing
312 * where we can be grabbing buffers at the same time that freeze is trying to
313 * drain the buffer LRU list.
314 */
315int
316xfs_trans_alloc_empty(
317 struct xfs_mount *mp,
318 struct xfs_trans **tpp)
319{
320 struct xfs_trans_res resv = {0};
321
322 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
323}
324
325/*
326 * Record the indicated change to the given field for application
327 * to the file system's superblock when the transaction commits.
328 * For now, just store the change in the transaction structure.
329 *
330 * Mark the transaction structure to indicate that the superblock
331 * needs to be updated before committing.
332 *
333 * Because we may not be keeping track of allocated/free inodes and
334 * used filesystem blocks in the superblock, we do not mark the
335 * superblock dirty in this transaction if we modify these fields.
336 * We still need to update the transaction deltas so that they get
337 * applied to the incore superblock, but we don't want them to
338 * cause the superblock to get locked and logged if these are the
339 * only fields in the superblock that the transaction modifies.
340 */
341void
342xfs_trans_mod_sb(
343 xfs_trans_t *tp,
344 uint field,
345 int64_t delta)
346{
347 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
348 xfs_mount_t *mp = tp->t_mountp;
349
350 switch (field) {
351 case XFS_TRANS_SB_ICOUNT:
352 tp->t_icount_delta += delta;
353 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
354 flags &= ~XFS_TRANS_SB_DIRTY;
355 break;
356 case XFS_TRANS_SB_IFREE:
357 tp->t_ifree_delta += delta;
358 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
359 flags &= ~XFS_TRANS_SB_DIRTY;
360 break;
361 case XFS_TRANS_SB_FDBLOCKS:
362 /*
363 * Track the number of blocks allocated in the transaction.
364 * Make sure it does not exceed the number reserved. If so,
365 * shutdown as this can lead to accounting inconsistency.
366 */
367 if (delta < 0) {
368 tp->t_blk_res_used += (uint)-delta;
369 if (tp->t_blk_res_used > tp->t_blk_res)
370 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
371 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
372 int64_t blkres_delta;
373
374 /*
375 * Return freed blocks directly to the reservation
376 * instead of the global pool, being careful not to
377 * overflow the trans counter. This is used to preserve
378 * reservation across chains of transaction rolls that
379 * repeatedly free and allocate blocks.
380 */
381 blkres_delta = min_t(int64_t, delta,
382 UINT_MAX - tp->t_blk_res);
383 tp->t_blk_res += blkres_delta;
384 delta -= blkres_delta;
385 }
386 tp->t_fdblocks_delta += delta;
387 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
388 flags &= ~XFS_TRANS_SB_DIRTY;
389 break;
390 case XFS_TRANS_SB_RES_FDBLOCKS:
391 /*
392 * The allocation has already been applied to the
393 * in-core superblock's counter. This should only
394 * be applied to the on-disk superblock.
395 */
396 tp->t_res_fdblocks_delta += delta;
397 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
398 flags &= ~XFS_TRANS_SB_DIRTY;
399 break;
400 case XFS_TRANS_SB_FREXTENTS:
401 /*
402 * Track the number of blocks allocated in the
403 * transaction. Make sure it does not exceed the
404 * number reserved.
405 */
406 if (delta < 0) {
407 tp->t_rtx_res_used += (uint)-delta;
408 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
409 }
410 tp->t_frextents_delta += delta;
411 break;
412 case XFS_TRANS_SB_RES_FREXTENTS:
413 /*
414 * The allocation has already been applied to the
415 * in-core superblock's counter. This should only
416 * be applied to the on-disk superblock.
417 */
418 ASSERT(delta < 0);
419 tp->t_res_frextents_delta += delta;
420 break;
421 case XFS_TRANS_SB_DBLOCKS:
422 ASSERT(delta > 0);
423 tp->t_dblocks_delta += delta;
424 break;
425 case XFS_TRANS_SB_AGCOUNT:
426 ASSERT(delta > 0);
427 tp->t_agcount_delta += delta;
428 break;
429 case XFS_TRANS_SB_IMAXPCT:
430 tp->t_imaxpct_delta += delta;
431 break;
432 case XFS_TRANS_SB_REXTSIZE:
433 tp->t_rextsize_delta += delta;
434 break;
435 case XFS_TRANS_SB_RBMBLOCKS:
436 tp->t_rbmblocks_delta += delta;
437 break;
438 case XFS_TRANS_SB_RBLOCKS:
439 tp->t_rblocks_delta += delta;
440 break;
441 case XFS_TRANS_SB_REXTENTS:
442 tp->t_rextents_delta += delta;
443 break;
444 case XFS_TRANS_SB_REXTSLOG:
445 tp->t_rextslog_delta += delta;
446 break;
447 default:
448 ASSERT(0);
449 return;
450 }
451
452 tp->t_flags |= flags;
453}
454
455/*
456 * xfs_trans_apply_sb_deltas() is called from the commit code
457 * to bring the superblock buffer into the current transaction
458 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
459 *
460 * For now we just look at each field allowed to change and change
461 * it if necessary.
462 */
463STATIC void
464xfs_trans_apply_sb_deltas(
465 xfs_trans_t *tp)
466{
467 xfs_dsb_t *sbp;
468 xfs_buf_t *bp;
469 int whole = 0;
470
471 bp = xfs_trans_getsb(tp, tp->t_mountp);
472 sbp = bp->b_addr;
473
474 /*
475 * Check that superblock mods match the mods made to AGF counters.
476 */
477 ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
478 (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
479 tp->t_ag_btree_delta));
480
481 /*
482 * Only update the superblock counters if we are logging them
483 */
484 if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
485 if (tp->t_icount_delta)
486 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
487 if (tp->t_ifree_delta)
488 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
489 if (tp->t_fdblocks_delta)
490 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
491 if (tp->t_res_fdblocks_delta)
492 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
493 }
494
495 if (tp->t_frextents_delta)
496 be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
497 if (tp->t_res_frextents_delta)
498 be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
499
500 if (tp->t_dblocks_delta) {
501 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
502 whole = 1;
503 }
504 if (tp->t_agcount_delta) {
505 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
506 whole = 1;
507 }
508 if (tp->t_imaxpct_delta) {
509 sbp->sb_imax_pct += tp->t_imaxpct_delta;
510 whole = 1;
511 }
512 if (tp->t_rextsize_delta) {
513 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
514 whole = 1;
515 }
516 if (tp->t_rbmblocks_delta) {
517 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
518 whole = 1;
519 }
520 if (tp->t_rblocks_delta) {
521 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
522 whole = 1;
523 }
524 if (tp->t_rextents_delta) {
525 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
526 whole = 1;
527 }
528 if (tp->t_rextslog_delta) {
529 sbp->sb_rextslog += tp->t_rextslog_delta;
530 whole = 1;
531 }
532
533 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
534 if (whole)
535 /*
536 * Log the whole thing, the fields are noncontiguous.
537 */
538 xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
539 else
540 /*
541 * Since all the modifiable fields are contiguous, we
542 * can get away with this.
543 */
544 xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
545 offsetof(xfs_dsb_t, sb_frextents) +
546 sizeof(sbp->sb_frextents) - 1);
547}
548
549/*
550 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
551 * apply superblock counter changes to the in-core superblock. The
552 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
553 * applied to the in-core superblock. The idea is that that has already been
554 * done.
555 *
556 * If we are not logging superblock counters, then the inode allocated/free and
557 * used block counts are not updated in the on disk superblock. In this case,
558 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
559 * still need to update the incore superblock with the changes.
560 *
561 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
562 * so we don't need to take the counter lock on every update.
563 */
564#define XFS_ICOUNT_BATCH 128
565
566void
567xfs_trans_unreserve_and_mod_sb(
568 struct xfs_trans *tp)
569{
570 struct xfs_mount *mp = tp->t_mountp;
571 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
572 int64_t blkdelta = 0;
573 int64_t rtxdelta = 0;
574 int64_t idelta = 0;
575 int64_t ifreedelta = 0;
576 int error;
577
578 /* calculate deltas */
579 if (tp->t_blk_res > 0)
580 blkdelta = tp->t_blk_res;
581 if ((tp->t_fdblocks_delta != 0) &&
582 (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
583 (tp->t_flags & XFS_TRANS_SB_DIRTY)))
584 blkdelta += tp->t_fdblocks_delta;
585
586 if (tp->t_rtx_res > 0)
587 rtxdelta = tp->t_rtx_res;
588 if ((tp->t_frextents_delta != 0) &&
589 (tp->t_flags & XFS_TRANS_SB_DIRTY))
590 rtxdelta += tp->t_frextents_delta;
591
592 if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
593 (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
594 idelta = tp->t_icount_delta;
595 ifreedelta = tp->t_ifree_delta;
596 }
597
598 /* apply the per-cpu counters */
599 if (blkdelta) {
600 error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
601 ASSERT(!error);
602 }
603
604 if (idelta) {
605 percpu_counter_add_batch(&mp->m_icount, idelta,
606 XFS_ICOUNT_BATCH);
607 if (idelta < 0)
608 ASSERT(__percpu_counter_compare(&mp->m_icount, 0,
609 XFS_ICOUNT_BATCH) >= 0);
610 }
611
612 if (ifreedelta) {
613 percpu_counter_add(&mp->m_ifree, ifreedelta);
614 if (ifreedelta < 0)
615 ASSERT(percpu_counter_compare(&mp->m_ifree, 0) >= 0);
616 }
617
618 if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
619 return;
620
621 /* apply remaining deltas */
622 spin_lock(&mp->m_sb_lock);
623 mp->m_sb.sb_frextents += rtxdelta;
624 mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
625 mp->m_sb.sb_agcount += tp->t_agcount_delta;
626 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
627 mp->m_sb.sb_rextsize += tp->t_rextsize_delta;
628 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
629 mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
630 mp->m_sb.sb_rextents += tp->t_rextents_delta;
631 mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
632 spin_unlock(&mp->m_sb_lock);
633
634 /*
635 * Debug checks outside of the spinlock so they don't lock up the
636 * machine if they fail.
637 */
638 ASSERT(mp->m_sb.sb_imax_pct >= 0);
639 ASSERT(mp->m_sb.sb_rextslog >= 0);
640 return;
641}
642
643/* Add the given log item to the transaction's list of log items. */
644void
645xfs_trans_add_item(
646 struct xfs_trans *tp,
647 struct xfs_log_item *lip)
648{
649 ASSERT(lip->li_mountp == tp->t_mountp);
650 ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
651 ASSERT(list_empty(&lip->li_trans));
652 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
653
654 list_add_tail(&lip->li_trans, &tp->t_items);
655 trace_xfs_trans_add_item(tp, _RET_IP_);
656}
657
658/*
659 * Unlink the log item from the transaction. the log item is no longer
660 * considered dirty in this transaction, as the linked transaction has
661 * finished, either by abort or commit completion.
662 */
663void
664xfs_trans_del_item(
665 struct xfs_log_item *lip)
666{
667 clear_bit(XFS_LI_DIRTY, &lip->li_flags);
668 list_del_init(&lip->li_trans);
669}
670
671/* Detach and unlock all of the items in a transaction */
672static void
673xfs_trans_free_items(
674 struct xfs_trans *tp,
675 bool abort)
676{
677 struct xfs_log_item *lip, *next;
678
679 trace_xfs_trans_free_items(tp, _RET_IP_);
680
681 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
682 xfs_trans_del_item(lip);
683 if (abort)
684 set_bit(XFS_LI_ABORTED, &lip->li_flags);
685 if (lip->li_ops->iop_release)
686 lip->li_ops->iop_release(lip);
687 }
688}
689
690static inline void
691xfs_log_item_batch_insert(
692 struct xfs_ail *ailp,
693 struct xfs_ail_cursor *cur,
694 struct xfs_log_item **log_items,
695 int nr_items,
696 xfs_lsn_t commit_lsn)
697{
698 int i;
699
700 spin_lock(&ailp->ail_lock);
701 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */
702 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
703
704 for (i = 0; i < nr_items; i++) {
705 struct xfs_log_item *lip = log_items[i];
706
707 if (lip->li_ops->iop_unpin)
708 lip->li_ops->iop_unpin(lip, 0);
709 }
710}
711
712/*
713 * Bulk operation version of xfs_trans_committed that takes a log vector of
714 * items to insert into the AIL. This uses bulk AIL insertion techniques to
715 * minimise lock traffic.
716 *
717 * If we are called with the aborted flag set, it is because a log write during
718 * a CIL checkpoint commit has failed. In this case, all the items in the
719 * checkpoint have already gone through iop_committed and iop_committing, which
720 * means that checkpoint commit abort handling is treated exactly the same
721 * as an iclog write error even though we haven't started any IO yet. Hence in
722 * this case all we need to do is iop_committed processing, followed by an
723 * iop_unpin(aborted) call.
724 *
725 * The AIL cursor is used to optimise the insert process. If commit_lsn is not
726 * at the end of the AIL, the insert cursor avoids the need to walk
727 * the AIL to find the insertion point on every xfs_log_item_batch_insert()
728 * call. This saves a lot of needless list walking and is a net win, even
729 * though it slightly increases that amount of AIL lock traffic to set it up
730 * and tear it down.
731 */
732void
733xfs_trans_committed_bulk(
734 struct xfs_ail *ailp,
735 struct xfs_log_vec *log_vector,
736 xfs_lsn_t commit_lsn,
737 bool aborted)
738{
739#define LOG_ITEM_BATCH_SIZE 32
740 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];
741 struct xfs_log_vec *lv;
742 struct xfs_ail_cursor cur;
743 int i = 0;
744
745 spin_lock(&ailp->ail_lock);
746 xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
747 spin_unlock(&ailp->ail_lock);
748
749 /* unpin all the log items */
750 for (lv = log_vector; lv; lv = lv->lv_next ) {
751 struct xfs_log_item *lip = lv->lv_item;
752 xfs_lsn_t item_lsn;
753
754 if (aborted)
755 set_bit(XFS_LI_ABORTED, &lip->li_flags);
756
757 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
758 lip->li_ops->iop_release(lip);
759 continue;
760 }
761
762 if (lip->li_ops->iop_committed)
763 item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
764 else
765 item_lsn = commit_lsn;
766
767 /* item_lsn of -1 means the item needs no further processing */
768 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
769 continue;
770
771 /*
772 * if we are aborting the operation, no point in inserting the
773 * object into the AIL as we are in a shutdown situation.
774 */
775 if (aborted) {
776 ASSERT(XFS_FORCED_SHUTDOWN(ailp->ail_mount));
777 if (lip->li_ops->iop_unpin)
778 lip->li_ops->iop_unpin(lip, 1);
779 continue;
780 }
781
782 if (item_lsn != commit_lsn) {
783
784 /*
785 * Not a bulk update option due to unusual item_lsn.
786 * Push into AIL immediately, rechecking the lsn once
787 * we have the ail lock. Then unpin the item. This does
788 * not affect the AIL cursor the bulk insert path is
789 * using.
790 */
791 spin_lock(&ailp->ail_lock);
792 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
793 xfs_trans_ail_update(ailp, lip, item_lsn);
794 else
795 spin_unlock(&ailp->ail_lock);
796 if (lip->li_ops->iop_unpin)
797 lip->li_ops->iop_unpin(lip, 0);
798 continue;
799 }
800
801 /* Item is a candidate for bulk AIL insert. */
802 log_items[i++] = lv->lv_item;
803 if (i >= LOG_ITEM_BATCH_SIZE) {
804 xfs_log_item_batch_insert(ailp, &cur, log_items,
805 LOG_ITEM_BATCH_SIZE, commit_lsn);
806 i = 0;
807 }
808 }
809
810 /* make sure we insert the remainder! */
811 if (i)
812 xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
813
814 spin_lock(&ailp->ail_lock);
815 xfs_trans_ail_cursor_done(&cur);
816 spin_unlock(&ailp->ail_lock);
817}
818
819/*
820 * Commit the given transaction to the log.
821 *
822 * XFS disk error handling mechanism is not based on a typical
823 * transaction abort mechanism. Logically after the filesystem
824 * gets marked 'SHUTDOWN', we can't let any new transactions
825 * be durable - ie. committed to disk - because some metadata might
826 * be inconsistent. In such cases, this returns an error, and the
827 * caller may assume that all locked objects joined to the transaction
828 * have already been unlocked as if the commit had succeeded.
829 * Do not reference the transaction structure after this call.
830 */
831static int
832__xfs_trans_commit(
833 struct xfs_trans *tp,
834 bool regrant)
835{
836 struct xfs_mount *mp = tp->t_mountp;
837 xfs_lsn_t commit_lsn = -1;
838 int error = 0;
839 int sync = tp->t_flags & XFS_TRANS_SYNC;
840
841 trace_xfs_trans_commit(tp, _RET_IP_);
842
843 /*
844 * Finish deferred items on final commit. Only permanent transactions
845 * should ever have deferred ops.
846 */
847 WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
848 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
849 if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
850 error = xfs_defer_finish_noroll(&tp);
851 if (error)
852 goto out_unreserve;
853 }
854
855 /*
856 * If there is nothing to be logged by the transaction,
857 * then unlock all of the items associated with the
858 * transaction and free the transaction structure.
859 * Also make sure to return any reserved blocks to
860 * the free pool.
861 */
862 if (!(tp->t_flags & XFS_TRANS_DIRTY))
863 goto out_unreserve;
864
865 if (XFS_FORCED_SHUTDOWN(mp)) {
866 error = -EIO;
867 goto out_unreserve;
868 }
869
870 ASSERT(tp->t_ticket != NULL);
871
872 /*
873 * If we need to update the superblock, then do it now.
874 */
875 if (tp->t_flags & XFS_TRANS_SB_DIRTY)
876 xfs_trans_apply_sb_deltas(tp);
877 xfs_trans_apply_dquot_deltas(tp);
878
879 xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
880
881 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
882 xfs_trans_free(tp);
883
884 /*
885 * If the transaction needs to be synchronous, then force the
886 * log out now and wait for it.
887 */
888 if (sync) {
889 error = xfs_log_force_lsn(mp, commit_lsn, XFS_LOG_SYNC, NULL);
890 XFS_STATS_INC(mp, xs_trans_sync);
891 } else {
892 XFS_STATS_INC(mp, xs_trans_async);
893 }
894
895 return error;
896
897out_unreserve:
898 xfs_trans_unreserve_and_mod_sb(tp);
899
900 /*
901 * It is indeed possible for the transaction to be not dirty but
902 * the dqinfo portion to be. All that means is that we have some
903 * (non-persistent) quota reservations that need to be unreserved.
904 */
905 xfs_trans_unreserve_and_mod_dquots(tp);
906 if (tp->t_ticket) {
907 if (regrant && !XLOG_FORCED_SHUTDOWN(mp->m_log))
908 xfs_log_ticket_regrant(mp->m_log, tp->t_ticket);
909 else
910 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
911 tp->t_ticket = NULL;
912 }
913 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
914 xfs_trans_free_items(tp, !!error);
915 xfs_trans_free(tp);
916
917 XFS_STATS_INC(mp, xs_trans_empty);
918 return error;
919}
920
921int
922xfs_trans_commit(
923 struct xfs_trans *tp)
924{
925 return __xfs_trans_commit(tp, false);
926}
927
928/*
929 * Unlock all of the transaction's items and free the transaction.
930 * The transaction must not have modified any of its items, because
931 * there is no way to restore them to their previous state.
932 *
933 * If the transaction has made a log reservation, make sure to release
934 * it as well.
935 */
936void
937xfs_trans_cancel(
938 struct xfs_trans *tp)
939{
940 struct xfs_mount *mp = tp->t_mountp;
941 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY);
942
943 trace_xfs_trans_cancel(tp, _RET_IP_);
944
945 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
946 xfs_defer_cancel(tp);
947
948 /*
949 * See if the caller is relying on us to shut down the
950 * filesystem. This happens in paths where we detect
951 * corruption and decide to give up.
952 */
953 if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
954 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
955 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
956 }
957#ifdef DEBUG
958 if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
959 struct xfs_log_item *lip;
960
961 list_for_each_entry(lip, &tp->t_items, li_trans)
962 ASSERT(!(lip->li_type == XFS_LI_EFD));
963 }
964#endif
965 xfs_trans_unreserve_and_mod_sb(tp);
966 xfs_trans_unreserve_and_mod_dquots(tp);
967
968 if (tp->t_ticket) {
969 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
970 tp->t_ticket = NULL;
971 }
972
973 /* mark this thread as no longer being in a transaction */
974 current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
975
976 xfs_trans_free_items(tp, dirty);
977 xfs_trans_free(tp);
978}
979
980/*
981 * Roll from one trans in the sequence of PERMANENT transactions to
982 * the next: permanent transactions are only flushed out when
983 * committed with xfs_trans_commit(), but we still want as soon
984 * as possible to let chunks of it go to the log. So we commit the
985 * chunk we've been working on and get a new transaction to continue.
986 */
987int
988xfs_trans_roll(
989 struct xfs_trans **tpp)
990{
991 struct xfs_trans *trans = *tpp;
992 struct xfs_trans_res tres;
993 int error;
994
995 trace_xfs_trans_roll(trans, _RET_IP_);
996
997 /*
998 * Copy the critical parameters from one trans to the next.
999 */
1000 tres.tr_logres = trans->t_log_res;
1001 tres.tr_logcount = trans->t_log_count;
1002
1003 *tpp = xfs_trans_dup(trans);
1004
1005 /*
1006 * Commit the current transaction.
1007 * If this commit failed, then it'd just unlock those items that
1008 * are not marked ihold. That also means that a filesystem shutdown
1009 * is in progress. The caller takes the responsibility to cancel
1010 * the duplicate transaction that gets returned.
1011 */
1012 error = __xfs_trans_commit(trans, true);
1013 if (error)
1014 return error;
1015
1016 /*
1017 * Reserve space in the log for the next transaction.
1018 * This also pushes items in the "AIL", the list of logged items,
1019 * out to disk if they are taking up space at the tail of the log
1020 * that we want to use. This requires that either nothing be locked
1021 * across this call, or that anything that is locked be logged in
1022 * the prior and the next transactions.
1023 */
1024 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1025 return xfs_trans_reserve(*tpp, &tres, 0, 0);
1026}