xref: /openbmc/linux/fs/xfs/xfs_trans.c (revision 98ddec80)
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_inode.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 
24 kmem_zone_t	*xfs_trans_zone;
25 
26 #if defined(CONFIG_TRACEPOINTS)
27 static void
28 xfs_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  */
51 void
52 xfs_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  */
63 STATIC void
64 xfs_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 	atomic_dec(&tp->t_mountp->m_active_trans);
72 	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
73 		sb_end_intwrite(tp->t_mountp->m_super);
74 	xfs_trans_free_dqinfo(tp);
75 	kmem_zone_free(xfs_trans_zone, tp);
76 }
77 
78 /*
79  * This is called to create a new transaction which will share the
80  * permanent log reservation of the given transaction.  The remaining
81  * unused block and rt extent reservations are also inherited.  This
82  * implies that the original transaction is no longer allowed to allocate
83  * blocks.  Locks and log items, however, are no inherited.  They must
84  * be added to the new transaction explicitly.
85  */
86 STATIC struct xfs_trans *
87 xfs_trans_dup(
88 	struct xfs_trans	*tp)
89 {
90 	struct xfs_trans	*ntp;
91 
92 	trace_xfs_trans_dup(tp, _RET_IP_);
93 
94 	ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
95 
96 	/*
97 	 * Initialize the new transaction structure.
98 	 */
99 	ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
100 	ntp->t_mountp = tp->t_mountp;
101 	INIT_LIST_HEAD(&ntp->t_items);
102 	INIT_LIST_HEAD(&ntp->t_busy);
103 
104 	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
105 	ASSERT(tp->t_ticket != NULL);
106 
107 	ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
108 		       (tp->t_flags & XFS_TRANS_RESERVE) |
109 		       (tp->t_flags & XFS_TRANS_NO_WRITECOUNT);
110 	/* We gave our writer reference to the new transaction */
111 	tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
112 	ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
113 
114 	ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
115 	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
116 	tp->t_blk_res = tp->t_blk_res_used;
117 
118 	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
119 	tp->t_rtx_res = tp->t_rtx_res_used;
120 	ntp->t_pflags = tp->t_pflags;
121 	ntp->t_agfl_dfops = tp->t_agfl_dfops;
122 
123 	xfs_trans_dup_dqinfo(tp, ntp);
124 
125 	atomic_inc(&tp->t_mountp->m_active_trans);
126 	return ntp;
127 }
128 
129 /*
130  * This is called to reserve free disk blocks and log space for the
131  * given transaction.  This must be done before allocating any resources
132  * within the transaction.
133  *
134  * This will return ENOSPC if there are not enough blocks available.
135  * It will sleep waiting for available log space.
136  * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
137  * is used by long running transactions.  If any one of the reservations
138  * fails then they will all be backed out.
139  *
140  * This does not do quota reservations. That typically is done by the
141  * caller afterwards.
142  */
143 static int
144 xfs_trans_reserve(
145 	struct xfs_trans	*tp,
146 	struct xfs_trans_res	*resp,
147 	uint			blocks,
148 	uint			rtextents)
149 {
150 	int		error = 0;
151 	bool		rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
152 
153 	/* Mark this thread as being in a transaction */
154 	current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
155 
156 	/*
157 	 * Attempt to reserve the needed disk blocks by decrementing
158 	 * the number needed from the number available.  This will
159 	 * fail if the count would go below zero.
160 	 */
161 	if (blocks > 0) {
162 		error = xfs_mod_fdblocks(tp->t_mountp, -((int64_t)blocks), rsvd);
163 		if (error != 0) {
164 			current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
165 			return -ENOSPC;
166 		}
167 		tp->t_blk_res += blocks;
168 	}
169 
170 	/*
171 	 * Reserve the log space needed for this transaction.
172 	 */
173 	if (resp->tr_logres > 0) {
174 		bool	permanent = false;
175 
176 		ASSERT(tp->t_log_res == 0 ||
177 		       tp->t_log_res == resp->tr_logres);
178 		ASSERT(tp->t_log_count == 0 ||
179 		       tp->t_log_count == resp->tr_logcount);
180 
181 		if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {
182 			tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
183 			permanent = true;
184 		} else {
185 			ASSERT(tp->t_ticket == NULL);
186 			ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
187 		}
188 
189 		if (tp->t_ticket != NULL) {
190 			ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);
191 			error = xfs_log_regrant(tp->t_mountp, tp->t_ticket);
192 		} else {
193 			error = xfs_log_reserve(tp->t_mountp,
194 						resp->tr_logres,
195 						resp->tr_logcount,
196 						&tp->t_ticket, XFS_TRANSACTION,
197 						permanent);
198 		}
199 
200 		if (error)
201 			goto undo_blocks;
202 
203 		tp->t_log_res = resp->tr_logres;
204 		tp->t_log_count = resp->tr_logcount;
205 	}
206 
207 	/*
208 	 * Attempt to reserve the needed realtime extents by decrementing
209 	 * the number needed from the number available.  This will
210 	 * fail if the count would go below zero.
211 	 */
212 	if (rtextents > 0) {
213 		error = xfs_mod_frextents(tp->t_mountp, -((int64_t)rtextents));
214 		if (error) {
215 			error = -ENOSPC;
216 			goto undo_log;
217 		}
218 		tp->t_rtx_res += rtextents;
219 	}
220 
221 	return 0;
222 
223 	/*
224 	 * Error cases jump to one of these labels to undo any
225 	 * reservations which have already been performed.
226 	 */
227 undo_log:
228 	if (resp->tr_logres > 0) {
229 		xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, false);
230 		tp->t_ticket = NULL;
231 		tp->t_log_res = 0;
232 		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
233 	}
234 
235 undo_blocks:
236 	if (blocks > 0) {
237 		xfs_mod_fdblocks(tp->t_mountp, (int64_t)blocks, rsvd);
238 		tp->t_blk_res = 0;
239 	}
240 
241 	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
242 
243 	return error;
244 }
245 
246 int
247 xfs_trans_alloc(
248 	struct xfs_mount	*mp,
249 	struct xfs_trans_res	*resp,
250 	uint			blocks,
251 	uint			rtextents,
252 	uint			flags,
253 	struct xfs_trans	**tpp)
254 {
255 	struct xfs_trans	*tp;
256 	int			error;
257 
258 	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
259 		sb_start_intwrite(mp->m_super);
260 
261 	WARN_ON(mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
262 	atomic_inc(&mp->m_active_trans);
263 
264 	tp = kmem_zone_zalloc(xfs_trans_zone,
265 		(flags & XFS_TRANS_NOFS) ? KM_NOFS : KM_SLEEP);
266 	tp->t_magic = XFS_TRANS_HEADER_MAGIC;
267 	tp->t_flags = flags;
268 	tp->t_mountp = mp;
269 	INIT_LIST_HEAD(&tp->t_items);
270 	INIT_LIST_HEAD(&tp->t_busy);
271 
272 	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
273 	if (error) {
274 		xfs_trans_cancel(tp);
275 		return error;
276 	}
277 
278 	trace_xfs_trans_alloc(tp, _RET_IP_);
279 
280 	*tpp = tp;
281 	return 0;
282 }
283 
284 /*
285  * Create an empty transaction with no reservation.  This is a defensive
286  * mechanism for routines that query metadata without actually modifying
287  * them -- if the metadata being queried is somehow cross-linked (think a
288  * btree block pointer that points higher in the tree), we risk deadlock.
289  * However, blocks grabbed as part of a transaction can be re-grabbed.
290  * The verifiers will notice the corrupt block and the operation will fail
291  * back to userspace without deadlocking.
292  *
293  * Note the zero-length reservation; this transaction MUST be cancelled
294  * without any dirty data.
295  */
296 int
297 xfs_trans_alloc_empty(
298 	struct xfs_mount		*mp,
299 	struct xfs_trans		**tpp)
300 {
301 	struct xfs_trans_res		resv = {0};
302 
303 	return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
304 }
305 
306 /*
307  * Record the indicated change to the given field for application
308  * to the file system's superblock when the transaction commits.
309  * For now, just store the change in the transaction structure.
310  *
311  * Mark the transaction structure to indicate that the superblock
312  * needs to be updated before committing.
313  *
314  * Because we may not be keeping track of allocated/free inodes and
315  * used filesystem blocks in the superblock, we do not mark the
316  * superblock dirty in this transaction if we modify these fields.
317  * We still need to update the transaction deltas so that they get
318  * applied to the incore superblock, but we don't want them to
319  * cause the superblock to get locked and logged if these are the
320  * only fields in the superblock that the transaction modifies.
321  */
322 void
323 xfs_trans_mod_sb(
324 	xfs_trans_t	*tp,
325 	uint		field,
326 	int64_t		delta)
327 {
328 	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
329 	xfs_mount_t	*mp = tp->t_mountp;
330 
331 	switch (field) {
332 	case XFS_TRANS_SB_ICOUNT:
333 		tp->t_icount_delta += delta;
334 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
335 			flags &= ~XFS_TRANS_SB_DIRTY;
336 		break;
337 	case XFS_TRANS_SB_IFREE:
338 		tp->t_ifree_delta += delta;
339 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
340 			flags &= ~XFS_TRANS_SB_DIRTY;
341 		break;
342 	case XFS_TRANS_SB_FDBLOCKS:
343 		/*
344 		 * Track the number of blocks allocated in the transaction.
345 		 * Make sure it does not exceed the number reserved. If so,
346 		 * shutdown as this can lead to accounting inconsistency.
347 		 */
348 		if (delta < 0) {
349 			tp->t_blk_res_used += (uint)-delta;
350 			if (tp->t_blk_res_used > tp->t_blk_res)
351 				xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
352 		}
353 		tp->t_fdblocks_delta += delta;
354 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
355 			flags &= ~XFS_TRANS_SB_DIRTY;
356 		break;
357 	case XFS_TRANS_SB_RES_FDBLOCKS:
358 		/*
359 		 * The allocation has already been applied to the
360 		 * in-core superblock's counter.  This should only
361 		 * be applied to the on-disk superblock.
362 		 */
363 		tp->t_res_fdblocks_delta += delta;
364 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
365 			flags &= ~XFS_TRANS_SB_DIRTY;
366 		break;
367 	case XFS_TRANS_SB_FREXTENTS:
368 		/*
369 		 * Track the number of blocks allocated in the
370 		 * transaction.  Make sure it does not exceed the
371 		 * number reserved.
372 		 */
373 		if (delta < 0) {
374 			tp->t_rtx_res_used += (uint)-delta;
375 			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
376 		}
377 		tp->t_frextents_delta += delta;
378 		break;
379 	case XFS_TRANS_SB_RES_FREXTENTS:
380 		/*
381 		 * The allocation has already been applied to the
382 		 * in-core superblock's counter.  This should only
383 		 * be applied to the on-disk superblock.
384 		 */
385 		ASSERT(delta < 0);
386 		tp->t_res_frextents_delta += delta;
387 		break;
388 	case XFS_TRANS_SB_DBLOCKS:
389 		ASSERT(delta > 0);
390 		tp->t_dblocks_delta += delta;
391 		break;
392 	case XFS_TRANS_SB_AGCOUNT:
393 		ASSERT(delta > 0);
394 		tp->t_agcount_delta += delta;
395 		break;
396 	case XFS_TRANS_SB_IMAXPCT:
397 		tp->t_imaxpct_delta += delta;
398 		break;
399 	case XFS_TRANS_SB_REXTSIZE:
400 		tp->t_rextsize_delta += delta;
401 		break;
402 	case XFS_TRANS_SB_RBMBLOCKS:
403 		tp->t_rbmblocks_delta += delta;
404 		break;
405 	case XFS_TRANS_SB_RBLOCKS:
406 		tp->t_rblocks_delta += delta;
407 		break;
408 	case XFS_TRANS_SB_REXTENTS:
409 		tp->t_rextents_delta += delta;
410 		break;
411 	case XFS_TRANS_SB_REXTSLOG:
412 		tp->t_rextslog_delta += delta;
413 		break;
414 	default:
415 		ASSERT(0);
416 		return;
417 	}
418 
419 	tp->t_flags |= flags;
420 }
421 
422 /*
423  * xfs_trans_apply_sb_deltas() is called from the commit code
424  * to bring the superblock buffer into the current transaction
425  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
426  *
427  * For now we just look at each field allowed to change and change
428  * it if necessary.
429  */
430 STATIC void
431 xfs_trans_apply_sb_deltas(
432 	xfs_trans_t	*tp)
433 {
434 	xfs_dsb_t	*sbp;
435 	xfs_buf_t	*bp;
436 	int		whole = 0;
437 
438 	bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
439 	sbp = XFS_BUF_TO_SBP(bp);
440 
441 	/*
442 	 * Check that superblock mods match the mods made to AGF counters.
443 	 */
444 	ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
445 	       (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
446 		tp->t_ag_btree_delta));
447 
448 	/*
449 	 * Only update the superblock counters if we are logging them
450 	 */
451 	if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
452 		if (tp->t_icount_delta)
453 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
454 		if (tp->t_ifree_delta)
455 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
456 		if (tp->t_fdblocks_delta)
457 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
458 		if (tp->t_res_fdblocks_delta)
459 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
460 	}
461 
462 	if (tp->t_frextents_delta)
463 		be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
464 	if (tp->t_res_frextents_delta)
465 		be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
466 
467 	if (tp->t_dblocks_delta) {
468 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
469 		whole = 1;
470 	}
471 	if (tp->t_agcount_delta) {
472 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
473 		whole = 1;
474 	}
475 	if (tp->t_imaxpct_delta) {
476 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
477 		whole = 1;
478 	}
479 	if (tp->t_rextsize_delta) {
480 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
481 		whole = 1;
482 	}
483 	if (tp->t_rbmblocks_delta) {
484 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
485 		whole = 1;
486 	}
487 	if (tp->t_rblocks_delta) {
488 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
489 		whole = 1;
490 	}
491 	if (tp->t_rextents_delta) {
492 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
493 		whole = 1;
494 	}
495 	if (tp->t_rextslog_delta) {
496 		sbp->sb_rextslog += tp->t_rextslog_delta;
497 		whole = 1;
498 	}
499 
500 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
501 	if (whole)
502 		/*
503 		 * Log the whole thing, the fields are noncontiguous.
504 		 */
505 		xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
506 	else
507 		/*
508 		 * Since all the modifiable fields are contiguous, we
509 		 * can get away with this.
510 		 */
511 		xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
512 				  offsetof(xfs_dsb_t, sb_frextents) +
513 				  sizeof(sbp->sb_frextents) - 1);
514 }
515 
516 STATIC int
517 xfs_sb_mod8(
518 	uint8_t			*field,
519 	int8_t			delta)
520 {
521 	int8_t			counter = *field;
522 
523 	counter += delta;
524 	if (counter < 0) {
525 		ASSERT(0);
526 		return -EINVAL;
527 	}
528 	*field = counter;
529 	return 0;
530 }
531 
532 STATIC int
533 xfs_sb_mod32(
534 	uint32_t		*field,
535 	int32_t			delta)
536 {
537 	int32_t			counter = *field;
538 
539 	counter += delta;
540 	if (counter < 0) {
541 		ASSERT(0);
542 		return -EINVAL;
543 	}
544 	*field = counter;
545 	return 0;
546 }
547 
548 STATIC int
549 xfs_sb_mod64(
550 	uint64_t		*field,
551 	int64_t			delta)
552 {
553 	int64_t			counter = *field;
554 
555 	counter += delta;
556 	if (counter < 0) {
557 		ASSERT(0);
558 		return -EINVAL;
559 	}
560 	*field = counter;
561 	return 0;
562 }
563 
564 /*
565  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
566  * and apply superblock counter changes to the in-core superblock.  The
567  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
568  * applied to the in-core superblock.  The idea is that that has already been
569  * done.
570  *
571  * If we are not logging superblock counters, then the inode allocated/free and
572  * used block counts are not updated in the on disk superblock. In this case,
573  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
574  * still need to update the incore superblock with the changes.
575  */
576 void
577 xfs_trans_unreserve_and_mod_sb(
578 	struct xfs_trans	*tp)
579 {
580 	struct xfs_mount	*mp = tp->t_mountp;
581 	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
582 	int64_t			blkdelta = 0;
583 	int64_t			rtxdelta = 0;
584 	int64_t			idelta = 0;
585 	int64_t			ifreedelta = 0;
586 	int			error;
587 
588 	/* calculate deltas */
589 	if (tp->t_blk_res > 0)
590 		blkdelta = tp->t_blk_res;
591 	if ((tp->t_fdblocks_delta != 0) &&
592 	    (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
593 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)))
594 	        blkdelta += tp->t_fdblocks_delta;
595 
596 	if (tp->t_rtx_res > 0)
597 		rtxdelta = tp->t_rtx_res;
598 	if ((tp->t_frextents_delta != 0) &&
599 	    (tp->t_flags & XFS_TRANS_SB_DIRTY))
600 		rtxdelta += tp->t_frextents_delta;
601 
602 	if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
603 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
604 		idelta = tp->t_icount_delta;
605 		ifreedelta = tp->t_ifree_delta;
606 	}
607 
608 	/* apply the per-cpu counters */
609 	if (blkdelta) {
610 		error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
611 		if (error)
612 			goto out;
613 	}
614 
615 	if (idelta) {
616 		error = xfs_mod_icount(mp, idelta);
617 		if (error)
618 			goto out_undo_fdblocks;
619 	}
620 
621 	if (ifreedelta) {
622 		error = xfs_mod_ifree(mp, ifreedelta);
623 		if (error)
624 			goto out_undo_icount;
625 	}
626 
627 	if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
628 		return;
629 
630 	/* apply remaining deltas */
631 	spin_lock(&mp->m_sb_lock);
632 	if (rtxdelta) {
633 		error = xfs_sb_mod64(&mp->m_sb.sb_frextents, rtxdelta);
634 		if (error)
635 			goto out_undo_ifree;
636 	}
637 
638 	if (tp->t_dblocks_delta != 0) {
639 		error = xfs_sb_mod64(&mp->m_sb.sb_dblocks, tp->t_dblocks_delta);
640 		if (error)
641 			goto out_undo_frextents;
642 	}
643 	if (tp->t_agcount_delta != 0) {
644 		error = xfs_sb_mod32(&mp->m_sb.sb_agcount, tp->t_agcount_delta);
645 		if (error)
646 			goto out_undo_dblocks;
647 	}
648 	if (tp->t_imaxpct_delta != 0) {
649 		error = xfs_sb_mod8(&mp->m_sb.sb_imax_pct, tp->t_imaxpct_delta);
650 		if (error)
651 			goto out_undo_agcount;
652 	}
653 	if (tp->t_rextsize_delta != 0) {
654 		error = xfs_sb_mod32(&mp->m_sb.sb_rextsize,
655 				     tp->t_rextsize_delta);
656 		if (error)
657 			goto out_undo_imaxpct;
658 	}
659 	if (tp->t_rbmblocks_delta != 0) {
660 		error = xfs_sb_mod32(&mp->m_sb.sb_rbmblocks,
661 				     tp->t_rbmblocks_delta);
662 		if (error)
663 			goto out_undo_rextsize;
664 	}
665 	if (tp->t_rblocks_delta != 0) {
666 		error = xfs_sb_mod64(&mp->m_sb.sb_rblocks, tp->t_rblocks_delta);
667 		if (error)
668 			goto out_undo_rbmblocks;
669 	}
670 	if (tp->t_rextents_delta != 0) {
671 		error = xfs_sb_mod64(&mp->m_sb.sb_rextents,
672 				     tp->t_rextents_delta);
673 		if (error)
674 			goto out_undo_rblocks;
675 	}
676 	if (tp->t_rextslog_delta != 0) {
677 		error = xfs_sb_mod8(&mp->m_sb.sb_rextslog,
678 				     tp->t_rextslog_delta);
679 		if (error)
680 			goto out_undo_rextents;
681 	}
682 	spin_unlock(&mp->m_sb_lock);
683 	return;
684 
685 out_undo_rextents:
686 	if (tp->t_rextents_delta)
687 		xfs_sb_mod64(&mp->m_sb.sb_rextents, -tp->t_rextents_delta);
688 out_undo_rblocks:
689 	if (tp->t_rblocks_delta)
690 		xfs_sb_mod64(&mp->m_sb.sb_rblocks, -tp->t_rblocks_delta);
691 out_undo_rbmblocks:
692 	if (tp->t_rbmblocks_delta)
693 		xfs_sb_mod32(&mp->m_sb.sb_rbmblocks, -tp->t_rbmblocks_delta);
694 out_undo_rextsize:
695 	if (tp->t_rextsize_delta)
696 		xfs_sb_mod32(&mp->m_sb.sb_rextsize, -tp->t_rextsize_delta);
697 out_undo_imaxpct:
698 	if (tp->t_rextsize_delta)
699 		xfs_sb_mod8(&mp->m_sb.sb_imax_pct, -tp->t_imaxpct_delta);
700 out_undo_agcount:
701 	if (tp->t_agcount_delta)
702 		xfs_sb_mod32(&mp->m_sb.sb_agcount, -tp->t_agcount_delta);
703 out_undo_dblocks:
704 	if (tp->t_dblocks_delta)
705 		xfs_sb_mod64(&mp->m_sb.sb_dblocks, -tp->t_dblocks_delta);
706 out_undo_frextents:
707 	if (rtxdelta)
708 		xfs_sb_mod64(&mp->m_sb.sb_frextents, -rtxdelta);
709 out_undo_ifree:
710 	spin_unlock(&mp->m_sb_lock);
711 	if (ifreedelta)
712 		xfs_mod_ifree(mp, -ifreedelta);
713 out_undo_icount:
714 	if (idelta)
715 		xfs_mod_icount(mp, -idelta);
716 out_undo_fdblocks:
717 	if (blkdelta)
718 		xfs_mod_fdblocks(mp, -blkdelta, rsvd);
719 out:
720 	ASSERT(error == 0);
721 	return;
722 }
723 
724 /* Add the given log item to the transaction's list of log items. */
725 void
726 xfs_trans_add_item(
727 	struct xfs_trans	*tp,
728 	struct xfs_log_item	*lip)
729 {
730 	ASSERT(lip->li_mountp == tp->t_mountp);
731 	ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
732 	ASSERT(list_empty(&lip->li_trans));
733 	ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
734 
735 	list_add_tail(&lip->li_trans, &tp->t_items);
736 	trace_xfs_trans_add_item(tp, _RET_IP_);
737 }
738 
739 /*
740  * Unlink the log item from the transaction. the log item is no longer
741  * considered dirty in this transaction, as the linked transaction has
742  * finished, either by abort or commit completion.
743  */
744 void
745 xfs_trans_del_item(
746 	struct xfs_log_item	*lip)
747 {
748 	clear_bit(XFS_LI_DIRTY, &lip->li_flags);
749 	list_del_init(&lip->li_trans);
750 }
751 
752 /* Detach and unlock all of the items in a transaction */
753 void
754 xfs_trans_free_items(
755 	struct xfs_trans	*tp,
756 	xfs_lsn_t		commit_lsn,
757 	bool			abort)
758 {
759 	struct xfs_log_item	*lip, *next;
760 
761 	trace_xfs_trans_free_items(tp, _RET_IP_);
762 
763 	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
764 		xfs_trans_del_item(lip);
765 		if (commit_lsn != NULLCOMMITLSN)
766 			lip->li_ops->iop_committing(lip, commit_lsn);
767 		if (abort)
768 			set_bit(XFS_LI_ABORTED, &lip->li_flags);
769 		lip->li_ops->iop_unlock(lip);
770 	}
771 }
772 
773 static inline void
774 xfs_log_item_batch_insert(
775 	struct xfs_ail		*ailp,
776 	struct xfs_ail_cursor	*cur,
777 	struct xfs_log_item	**log_items,
778 	int			nr_items,
779 	xfs_lsn_t		commit_lsn)
780 {
781 	int	i;
782 
783 	spin_lock(&ailp->ail_lock);
784 	/* xfs_trans_ail_update_bulk drops ailp->ail_lock */
785 	xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
786 
787 	for (i = 0; i < nr_items; i++) {
788 		struct xfs_log_item *lip = log_items[i];
789 
790 		lip->li_ops->iop_unpin(lip, 0);
791 	}
792 }
793 
794 /*
795  * Bulk operation version of xfs_trans_committed that takes a log vector of
796  * items to insert into the AIL. This uses bulk AIL insertion techniques to
797  * minimise lock traffic.
798  *
799  * If we are called with the aborted flag set, it is because a log write during
800  * a CIL checkpoint commit has failed. In this case, all the items in the
801  * checkpoint have already gone through iop_commited and iop_unlock, which
802  * means that checkpoint commit abort handling is treated exactly the same
803  * as an iclog write error even though we haven't started any IO yet. Hence in
804  * this case all we need to do is iop_committed processing, followed by an
805  * iop_unpin(aborted) call.
806  *
807  * The AIL cursor is used to optimise the insert process. If commit_lsn is not
808  * at the end of the AIL, the insert cursor avoids the need to walk
809  * the AIL to find the insertion point on every xfs_log_item_batch_insert()
810  * call. This saves a lot of needless list walking and is a net win, even
811  * though it slightly increases that amount of AIL lock traffic to set it up
812  * and tear it down.
813  */
814 void
815 xfs_trans_committed_bulk(
816 	struct xfs_ail		*ailp,
817 	struct xfs_log_vec	*log_vector,
818 	xfs_lsn_t		commit_lsn,
819 	int			aborted)
820 {
821 #define LOG_ITEM_BATCH_SIZE	32
822 	struct xfs_log_item	*log_items[LOG_ITEM_BATCH_SIZE];
823 	struct xfs_log_vec	*lv;
824 	struct xfs_ail_cursor	cur;
825 	int			i = 0;
826 
827 	spin_lock(&ailp->ail_lock);
828 	xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
829 	spin_unlock(&ailp->ail_lock);
830 
831 	/* unpin all the log items */
832 	for (lv = log_vector; lv; lv = lv->lv_next ) {
833 		struct xfs_log_item	*lip = lv->lv_item;
834 		xfs_lsn_t		item_lsn;
835 
836 		if (aborted)
837 			set_bit(XFS_LI_ABORTED, &lip->li_flags);
838 		item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
839 
840 		/* item_lsn of -1 means the item needs no further processing */
841 		if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
842 			continue;
843 
844 		/*
845 		 * if we are aborting the operation, no point in inserting the
846 		 * object into the AIL as we are in a shutdown situation.
847 		 */
848 		if (aborted) {
849 			ASSERT(XFS_FORCED_SHUTDOWN(ailp->ail_mount));
850 			lip->li_ops->iop_unpin(lip, 1);
851 			continue;
852 		}
853 
854 		if (item_lsn != commit_lsn) {
855 
856 			/*
857 			 * Not a bulk update option due to unusual item_lsn.
858 			 * Push into AIL immediately, rechecking the lsn once
859 			 * we have the ail lock. Then unpin the item. This does
860 			 * not affect the AIL cursor the bulk insert path is
861 			 * using.
862 			 */
863 			spin_lock(&ailp->ail_lock);
864 			if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
865 				xfs_trans_ail_update(ailp, lip, item_lsn);
866 			else
867 				spin_unlock(&ailp->ail_lock);
868 			lip->li_ops->iop_unpin(lip, 0);
869 			continue;
870 		}
871 
872 		/* Item is a candidate for bulk AIL insert.  */
873 		log_items[i++] = lv->lv_item;
874 		if (i >= LOG_ITEM_BATCH_SIZE) {
875 			xfs_log_item_batch_insert(ailp, &cur, log_items,
876 					LOG_ITEM_BATCH_SIZE, commit_lsn);
877 			i = 0;
878 		}
879 	}
880 
881 	/* make sure we insert the remainder! */
882 	if (i)
883 		xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
884 
885 	spin_lock(&ailp->ail_lock);
886 	xfs_trans_ail_cursor_done(&cur);
887 	spin_unlock(&ailp->ail_lock);
888 }
889 
890 /*
891  * Commit the given transaction to the log.
892  *
893  * XFS disk error handling mechanism is not based on a typical
894  * transaction abort mechanism. Logically after the filesystem
895  * gets marked 'SHUTDOWN', we can't let any new transactions
896  * be durable - ie. committed to disk - because some metadata might
897  * be inconsistent. In such cases, this returns an error, and the
898  * caller may assume that all locked objects joined to the transaction
899  * have already been unlocked as if the commit had succeeded.
900  * Do not reference the transaction structure after this call.
901  */
902 static int
903 __xfs_trans_commit(
904 	struct xfs_trans	*tp,
905 	bool			regrant)
906 {
907 	struct xfs_mount	*mp = tp->t_mountp;
908 	xfs_lsn_t		commit_lsn = -1;
909 	int			error = 0;
910 	int			sync = tp->t_flags & XFS_TRANS_SYNC;
911 
912 	ASSERT(!tp->t_agfl_dfops ||
913 	       !xfs_defer_has_unfinished_work(tp->t_agfl_dfops) || regrant);
914 
915 	trace_xfs_trans_commit(tp, _RET_IP_);
916 
917 	/*
918 	 * If there is nothing to be logged by the transaction,
919 	 * then unlock all of the items associated with the
920 	 * transaction and free the transaction structure.
921 	 * Also make sure to return any reserved blocks to
922 	 * the free pool.
923 	 */
924 	if (!(tp->t_flags & XFS_TRANS_DIRTY))
925 		goto out_unreserve;
926 
927 	if (XFS_FORCED_SHUTDOWN(mp)) {
928 		error = -EIO;
929 		goto out_unreserve;
930 	}
931 
932 	ASSERT(tp->t_ticket != NULL);
933 
934 	/*
935 	 * If we need to update the superblock, then do it now.
936 	 */
937 	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
938 		xfs_trans_apply_sb_deltas(tp);
939 	xfs_trans_apply_dquot_deltas(tp);
940 
941 	xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
942 
943 	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
944 	xfs_trans_free(tp);
945 
946 	/*
947 	 * If the transaction needs to be synchronous, then force the
948 	 * log out now and wait for it.
949 	 */
950 	if (sync) {
951 		error = xfs_log_force_lsn(mp, commit_lsn, XFS_LOG_SYNC, NULL);
952 		XFS_STATS_INC(mp, xs_trans_sync);
953 	} else {
954 		XFS_STATS_INC(mp, xs_trans_async);
955 	}
956 
957 	return error;
958 
959 out_unreserve:
960 	xfs_trans_unreserve_and_mod_sb(tp);
961 
962 	/*
963 	 * It is indeed possible for the transaction to be not dirty but
964 	 * the dqinfo portion to be.  All that means is that we have some
965 	 * (non-persistent) quota reservations that need to be unreserved.
966 	 */
967 	xfs_trans_unreserve_and_mod_dquots(tp);
968 	if (tp->t_ticket) {
969 		commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, regrant);
970 		if (commit_lsn == -1 && !error)
971 			error = -EIO;
972 		tp->t_ticket = NULL;
973 	}
974 	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
975 	xfs_trans_free_items(tp, NULLCOMMITLSN, !!error);
976 	xfs_trans_free(tp);
977 
978 	XFS_STATS_INC(mp, xs_trans_empty);
979 	return error;
980 }
981 
982 int
983 xfs_trans_commit(
984 	struct xfs_trans	*tp)
985 {
986 	return __xfs_trans_commit(tp, false);
987 }
988 
989 /*
990  * Unlock all of the transaction's items and free the transaction.
991  * The transaction must not have modified any of its items, because
992  * there is no way to restore them to their previous state.
993  *
994  * If the transaction has made a log reservation, make sure to release
995  * it as well.
996  */
997 void
998 xfs_trans_cancel(
999 	struct xfs_trans	*tp)
1000 {
1001 	struct xfs_mount	*mp = tp->t_mountp;
1002 	bool			dirty = (tp->t_flags & XFS_TRANS_DIRTY);
1003 
1004 	trace_xfs_trans_cancel(tp, _RET_IP_);
1005 
1006 	/*
1007 	 * See if the caller is relying on us to shut down the
1008 	 * filesystem.  This happens in paths where we detect
1009 	 * corruption and decide to give up.
1010 	 */
1011 	if (dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1012 		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1013 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1014 	}
1015 #ifdef DEBUG
1016 	if (!dirty && !XFS_FORCED_SHUTDOWN(mp)) {
1017 		struct xfs_log_item *lip;
1018 
1019 		list_for_each_entry(lip, &tp->t_items, li_trans)
1020 			ASSERT(!(lip->li_type == XFS_LI_EFD));
1021 	}
1022 #endif
1023 	xfs_trans_unreserve_and_mod_sb(tp);
1024 	xfs_trans_unreserve_and_mod_dquots(tp);
1025 
1026 	if (tp->t_ticket) {
1027 		xfs_log_done(mp, tp->t_ticket, NULL, false);
1028 		tp->t_ticket = NULL;
1029 	}
1030 
1031 	/* mark this thread as no longer being in a transaction */
1032 	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
1033 
1034 	xfs_trans_free_items(tp, NULLCOMMITLSN, dirty);
1035 	xfs_trans_free(tp);
1036 }
1037 
1038 /*
1039  * Roll from one trans in the sequence of PERMANENT transactions to
1040  * the next: permanent transactions are only flushed out when
1041  * committed with xfs_trans_commit(), but we still want as soon
1042  * as possible to let chunks of it go to the log. So we commit the
1043  * chunk we've been working on and get a new transaction to continue.
1044  */
1045 int
1046 xfs_trans_roll(
1047 	struct xfs_trans	**tpp)
1048 {
1049 	struct xfs_trans	*trans = *tpp;
1050 	struct xfs_trans_res	tres;
1051 	int			error;
1052 
1053 	trace_xfs_trans_roll(trans, _RET_IP_);
1054 
1055 	/*
1056 	 * Copy the critical parameters from one trans to the next.
1057 	 */
1058 	tres.tr_logres = trans->t_log_res;
1059 	tres.tr_logcount = trans->t_log_count;
1060 
1061 	*tpp = xfs_trans_dup(trans);
1062 
1063 	/*
1064 	 * Commit the current transaction.
1065 	 * If this commit failed, then it'd just unlock those items that
1066 	 * are not marked ihold. That also means that a filesystem shutdown
1067 	 * is in progress. The caller takes the responsibility to cancel
1068 	 * the duplicate transaction that gets returned.
1069 	 */
1070 	error = __xfs_trans_commit(trans, true);
1071 	if (error)
1072 		return error;
1073 
1074 	/*
1075 	 * Reserve space in the log for the next transaction.
1076 	 * This also pushes items in the "AIL", the list of logged items,
1077 	 * out to disk if they are taking up space at the tail of the log
1078 	 * that we want to use.  This requires that either nothing be locked
1079 	 * across this call, or that anything that is locked be logged in
1080 	 * the prior and the next transactions.
1081 	 */
1082 	tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1083 	return xfs_trans_reserve(*tpp, &tres, 0, 0);
1084 }
1085