1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
4  * Copyright (C) 2010 Red Hat, Inc.
5  * All Rights Reserved.
6  */
7 #include "xfs.h"
8 #include "xfs_fs.h"
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_extent_busy.h"
15 #include "xfs_quota.h"
16 #include "xfs_trans.h"
17 #include "xfs_trans_priv.h"
18 #include "xfs_log.h"
19 #include "xfs_log_priv.h"
20 #include "xfs_trace.h"
21 #include "xfs_error.h"
22 #include "xfs_defer.h"
23 #include "xfs_inode.h"
24 #include "xfs_dquot_item.h"
25 #include "xfs_dquot.h"
26 #include "xfs_icache.h"
27 
28 struct kmem_cache	*xfs_trans_cache;
29 
30 #if defined(CONFIG_TRACEPOINTS)
31 static void
xfs_trans_trace_reservations(struct xfs_mount * mp)32 xfs_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  */
55 void
xfs_trans_init(struct xfs_mount * mp)56 xfs_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  */
67 STATIC void
xfs_trans_free(struct xfs_trans * tp)68 xfs_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_cache, tp);
80 }
81 
82 /*
83  * This is called to create a new transaction which will share the
84  * permanent log reservation of the given transaction.  The remaining
85  * unused block and rt extent reservations are also inherited.  This
86  * implies that the original transaction is no longer allowed to allocate
87  * blocks.  Locks and log items, however, are no inherited.  They must
88  * be added to the new transaction explicitly.
89  */
90 STATIC struct xfs_trans *
xfs_trans_dup(struct xfs_trans * tp)91 xfs_trans_dup(
92 	struct xfs_trans	*tp)
93 {
94 	struct xfs_trans	*ntp;
95 
96 	trace_xfs_trans_dup(tp, _RET_IP_);
97 
98 	ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
99 
100 	/*
101 	 * Initialize the new transaction structure.
102 	 */
103 	ntp->t_magic = XFS_TRANS_HEADER_MAGIC;
104 	ntp->t_mountp = tp->t_mountp;
105 	INIT_LIST_HEAD(&ntp->t_items);
106 	INIT_LIST_HEAD(&ntp->t_busy);
107 	INIT_LIST_HEAD(&ntp->t_dfops);
108 	ntp->t_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  */
151 static int
xfs_trans_reserve(struct xfs_trans * tp,struct xfs_trans_res * resp,uint blocks,uint rtextents)152 xfs_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 	 */
231 undo_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 
239 undo_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 
247 int
xfs_trans_alloc(struct xfs_mount * mp,struct xfs_trans_res * resp,uint blocks,uint rtextents,uint flags,struct xfs_trans ** tpp)248 xfs_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 	 */
265 retry:
266 	tp = kmem_cache_zalloc(xfs_trans_cache, 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_has_lazysbcount(mp));
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 		xfs_blockgc_flush_all(mp);
299 		want_retry = false;
300 		goto retry;
301 	}
302 	if (error) {
303 		xfs_trans_cancel(tp);
304 		return error;
305 	}
306 
307 	trace_xfs_trans_alloc(tp, _RET_IP_);
308 
309 	*tpp = tp;
310 	return 0;
311 }
312 
313 /*
314  * Create an empty transaction with no reservation.  This is a defensive
315  * mechanism for routines that query metadata without actually modifying them --
316  * if the metadata being queried is somehow cross-linked (think a btree block
317  * pointer that points higher in the tree), we risk deadlock.  However, blocks
318  * grabbed as part of a transaction can be re-grabbed.  The verifiers will
319  * notice the corrupt block and the operation will fail back to userspace
320  * without deadlocking.
321  *
322  * Note the zero-length reservation; this transaction MUST be cancelled without
323  * any dirty data.
324  *
325  * Callers should obtain freeze protection to avoid a conflict with fs freezing
326  * where we can be grabbing buffers at the same time that freeze is trying to
327  * drain the buffer LRU list.
328  */
329 int
xfs_trans_alloc_empty(struct xfs_mount * mp,struct xfs_trans ** tpp)330 xfs_trans_alloc_empty(
331 	struct xfs_mount		*mp,
332 	struct xfs_trans		**tpp)
333 {
334 	struct xfs_trans_res		resv = {0};
335 
336 	return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);
337 }
338 
339 /*
340  * Record the indicated change to the given field for application
341  * to the file system's superblock when the transaction commits.
342  * For now, just store the change in the transaction structure.
343  *
344  * Mark the transaction structure to indicate that the superblock
345  * needs to be updated before committing.
346  *
347  * Because we may not be keeping track of allocated/free inodes and
348  * used filesystem blocks in the superblock, we do not mark the
349  * superblock dirty in this transaction if we modify these fields.
350  * We still need to update the transaction deltas so that they get
351  * applied to the incore superblock, but we don't want them to
352  * cause the superblock to get locked and logged if these are the
353  * only fields in the superblock that the transaction modifies.
354  */
355 void
xfs_trans_mod_sb(xfs_trans_t * tp,uint field,int64_t delta)356 xfs_trans_mod_sb(
357 	xfs_trans_t	*tp,
358 	uint		field,
359 	int64_t		delta)
360 {
361 	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
362 	xfs_mount_t	*mp = tp->t_mountp;
363 
364 	switch (field) {
365 	case XFS_TRANS_SB_ICOUNT:
366 		tp->t_icount_delta += delta;
367 		if (xfs_has_lazysbcount(mp))
368 			flags &= ~XFS_TRANS_SB_DIRTY;
369 		break;
370 	case XFS_TRANS_SB_IFREE:
371 		tp->t_ifree_delta += delta;
372 		if (xfs_has_lazysbcount(mp))
373 			flags &= ~XFS_TRANS_SB_DIRTY;
374 		break;
375 	case XFS_TRANS_SB_FDBLOCKS:
376 		/*
377 		 * Track the number of blocks allocated in the transaction.
378 		 * Make sure it does not exceed the number reserved. If so,
379 		 * shutdown as this can lead to accounting inconsistency.
380 		 */
381 		if (delta < 0) {
382 			tp->t_blk_res_used += (uint)-delta;
383 			if (tp->t_blk_res_used > tp->t_blk_res)
384 				xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
385 		} else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
386 			int64_t	blkres_delta;
387 
388 			/*
389 			 * Return freed blocks directly to the reservation
390 			 * instead of the global pool, being careful not to
391 			 * overflow the trans counter. This is used to preserve
392 			 * reservation across chains of transaction rolls that
393 			 * repeatedly free and allocate blocks.
394 			 */
395 			blkres_delta = min_t(int64_t, delta,
396 					     UINT_MAX - tp->t_blk_res);
397 			tp->t_blk_res += blkres_delta;
398 			delta -= blkres_delta;
399 		}
400 		tp->t_fdblocks_delta += delta;
401 		if (xfs_has_lazysbcount(mp))
402 			flags &= ~XFS_TRANS_SB_DIRTY;
403 		break;
404 	case XFS_TRANS_SB_RES_FDBLOCKS:
405 		/*
406 		 * The allocation has already been applied to the
407 		 * in-core superblock's counter.  This should only
408 		 * be applied to the on-disk superblock.
409 		 */
410 		tp->t_res_fdblocks_delta += delta;
411 		if (xfs_has_lazysbcount(mp))
412 			flags &= ~XFS_TRANS_SB_DIRTY;
413 		break;
414 	case XFS_TRANS_SB_FREXTENTS:
415 		/*
416 		 * Track the number of blocks allocated in the
417 		 * transaction.  Make sure it does not exceed the
418 		 * number reserved.
419 		 */
420 		if (delta < 0) {
421 			tp->t_rtx_res_used += (uint)-delta;
422 			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
423 		}
424 		tp->t_frextents_delta += delta;
425 		break;
426 	case XFS_TRANS_SB_RES_FREXTENTS:
427 		/*
428 		 * The allocation has already been applied to the
429 		 * in-core superblock's counter.  This should only
430 		 * be applied to the on-disk superblock.
431 		 */
432 		ASSERT(delta < 0);
433 		tp->t_res_frextents_delta += delta;
434 		break;
435 	case XFS_TRANS_SB_DBLOCKS:
436 		tp->t_dblocks_delta += delta;
437 		break;
438 	case XFS_TRANS_SB_AGCOUNT:
439 		ASSERT(delta > 0);
440 		tp->t_agcount_delta += delta;
441 		break;
442 	case XFS_TRANS_SB_IMAXPCT:
443 		tp->t_imaxpct_delta += delta;
444 		break;
445 	case XFS_TRANS_SB_REXTSIZE:
446 		tp->t_rextsize_delta += delta;
447 		break;
448 	case XFS_TRANS_SB_RBMBLOCKS:
449 		tp->t_rbmblocks_delta += delta;
450 		break;
451 	case XFS_TRANS_SB_RBLOCKS:
452 		tp->t_rblocks_delta += delta;
453 		break;
454 	case XFS_TRANS_SB_REXTENTS:
455 		tp->t_rextents_delta += delta;
456 		break;
457 	case XFS_TRANS_SB_REXTSLOG:
458 		tp->t_rextslog_delta += delta;
459 		break;
460 	default:
461 		ASSERT(0);
462 		return;
463 	}
464 
465 	tp->t_flags |= flags;
466 }
467 
468 /*
469  * xfs_trans_apply_sb_deltas() is called from the commit code
470  * to bring the superblock buffer into the current transaction
471  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
472  *
473  * For now we just look at each field allowed to change and change
474  * it if necessary.
475  */
476 STATIC void
xfs_trans_apply_sb_deltas(xfs_trans_t * tp)477 xfs_trans_apply_sb_deltas(
478 	xfs_trans_t	*tp)
479 {
480 	struct xfs_dsb	*sbp;
481 	struct xfs_buf	*bp;
482 	int		whole = 0;
483 
484 	bp = xfs_trans_getsb(tp);
485 	sbp = bp->b_addr;
486 
487 	/*
488 	 * Only update the superblock counters if we are logging them
489 	 */
490 	if (!xfs_has_lazysbcount((tp->t_mountp))) {
491 		if (tp->t_icount_delta)
492 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
493 		if (tp->t_ifree_delta)
494 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
495 		if (tp->t_fdblocks_delta)
496 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
497 		if (tp->t_res_fdblocks_delta)
498 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
499 	}
500 
501 	if (tp->t_frextents_delta)
502 		be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
503 	if (tp->t_res_frextents_delta)
504 		be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
505 
506 	if (tp->t_dblocks_delta) {
507 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
508 		whole = 1;
509 	}
510 	if (tp->t_agcount_delta) {
511 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
512 		whole = 1;
513 	}
514 	if (tp->t_imaxpct_delta) {
515 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
516 		whole = 1;
517 	}
518 	if (tp->t_rextsize_delta) {
519 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
520 		whole = 1;
521 	}
522 	if (tp->t_rbmblocks_delta) {
523 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
524 		whole = 1;
525 	}
526 	if (tp->t_rblocks_delta) {
527 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
528 		whole = 1;
529 	}
530 	if (tp->t_rextents_delta) {
531 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
532 		whole = 1;
533 	}
534 	if (tp->t_rextslog_delta) {
535 		sbp->sb_rextslog += tp->t_rextslog_delta;
536 		whole = 1;
537 	}
538 
539 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
540 	if (whole)
541 		/*
542 		 * Log the whole thing, the fields are noncontiguous.
543 		 */
544 		xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
545 	else
546 		/*
547 		 * Since all the modifiable fields are contiguous, we
548 		 * can get away with this.
549 		 */
550 		xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
551 				  offsetof(struct xfs_dsb, sb_frextents) +
552 				  sizeof(sbp->sb_frextents) - 1);
553 }
554 
555 /*
556  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
557  * apply superblock counter changes to the in-core superblock.  The
558  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
559  * applied to the in-core superblock.  The idea is that that has already been
560  * done.
561  *
562  * If we are not logging superblock counters, then the inode allocated/free and
563  * used block counts are not updated in the on disk superblock. In this case,
564  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
565  * still need to update the incore superblock with the changes.
566  *
567  * Deltas for the inode count are +/-64, hence we use a large batch size of 128
568  * so we don't need to take the counter lock on every update.
569  */
570 #define XFS_ICOUNT_BATCH	128
571 
572 void
xfs_trans_unreserve_and_mod_sb(struct xfs_trans * tp)573 xfs_trans_unreserve_and_mod_sb(
574 	struct xfs_trans	*tp)
575 {
576 	struct xfs_mount	*mp = tp->t_mountp;
577 	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
578 	int64_t			blkdelta = 0;
579 	int64_t			rtxdelta = 0;
580 	int64_t			idelta = 0;
581 	int64_t			ifreedelta = 0;
582 	int			error;
583 
584 	/* calculate deltas */
585 	if (tp->t_blk_res > 0)
586 		blkdelta = tp->t_blk_res;
587 	if ((tp->t_fdblocks_delta != 0) &&
588 	    (xfs_has_lazysbcount(mp) ||
589 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)))
590 	        blkdelta += tp->t_fdblocks_delta;
591 
592 	if (tp->t_rtx_res > 0)
593 		rtxdelta = tp->t_rtx_res;
594 	if ((tp->t_frextents_delta != 0) &&
595 	    (tp->t_flags & XFS_TRANS_SB_DIRTY))
596 		rtxdelta += tp->t_frextents_delta;
597 
598 	if (xfs_has_lazysbcount(mp) ||
599 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
600 		idelta = tp->t_icount_delta;
601 		ifreedelta = tp->t_ifree_delta;
602 	}
603 
604 	/* apply the per-cpu counters */
605 	if (blkdelta) {
606 		error = xfs_mod_fdblocks(mp, blkdelta, rsvd);
607 		ASSERT(!error);
608 	}
609 
610 	if (idelta)
611 		percpu_counter_add_batch(&mp->m_icount, idelta,
612 					 XFS_ICOUNT_BATCH);
613 
614 	if (ifreedelta)
615 		percpu_counter_add(&mp->m_ifree, ifreedelta);
616 
617 	if (rtxdelta == 0 && !(tp->t_flags & XFS_TRANS_SB_DIRTY))
618 		return;
619 
620 	/* apply remaining deltas */
621 	spin_lock(&mp->m_sb_lock);
622 	mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
623 	mp->m_sb.sb_icount += idelta;
624 	mp->m_sb.sb_ifree += ifreedelta;
625 	mp->m_sb.sb_frextents += rtxdelta;
626 	mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
627 	mp->m_sb.sb_agcount += tp->t_agcount_delta;
628 	mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
629 	mp->m_sb.sb_rextsize += tp->t_rextsize_delta;
630 	mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
631 	mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
632 	mp->m_sb.sb_rextents += tp->t_rextents_delta;
633 	mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
634 	spin_unlock(&mp->m_sb_lock);
635 
636 	/*
637 	 * Debug checks outside of the spinlock so they don't lock up the
638 	 * machine if they fail.
639 	 */
640 	ASSERT(mp->m_sb.sb_imax_pct >= 0);
641 	ASSERT(mp->m_sb.sb_rextslog >= 0);
642 	return;
643 }
644 
645 /* Add the given log item to the transaction's list of log items. */
646 void
xfs_trans_add_item(struct xfs_trans * tp,struct xfs_log_item * lip)647 xfs_trans_add_item(
648 	struct xfs_trans	*tp,
649 	struct xfs_log_item	*lip)
650 {
651 	ASSERT(lip->li_mountp == tp->t_mountp);
652 	ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
653 	ASSERT(list_empty(&lip->li_trans));
654 	ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
655 
656 	list_add_tail(&lip->li_trans, &tp->t_items);
657 	trace_xfs_trans_add_item(tp, _RET_IP_);
658 }
659 
660 /*
661  * Unlink the log item from the transaction. the log item is no longer
662  * considered dirty in this transaction, as the linked transaction has
663  * finished, either by abort or commit completion.
664  */
665 void
xfs_trans_del_item(struct xfs_log_item * lip)666 xfs_trans_del_item(
667 	struct xfs_log_item	*lip)
668 {
669 	clear_bit(XFS_LI_DIRTY, &lip->li_flags);
670 	list_del_init(&lip->li_trans);
671 }
672 
673 /* Detach and unlock all of the items in a transaction */
674 static void
xfs_trans_free_items(struct xfs_trans * tp,bool abort)675 xfs_trans_free_items(
676 	struct xfs_trans	*tp,
677 	bool			abort)
678 {
679 	struct xfs_log_item	*lip, *next;
680 
681 	trace_xfs_trans_free_items(tp, _RET_IP_);
682 
683 	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
684 		xfs_trans_del_item(lip);
685 		if (abort)
686 			set_bit(XFS_LI_ABORTED, &lip->li_flags);
687 		if (lip->li_ops->iop_release)
688 			lip->li_ops->iop_release(lip);
689 	}
690 }
691 
692 static inline void
xfs_log_item_batch_insert(struct xfs_ail * ailp,struct xfs_ail_cursor * cur,struct xfs_log_item ** log_items,int nr_items,xfs_lsn_t commit_lsn)693 xfs_log_item_batch_insert(
694 	struct xfs_ail		*ailp,
695 	struct xfs_ail_cursor	*cur,
696 	struct xfs_log_item	**log_items,
697 	int			nr_items,
698 	xfs_lsn_t		commit_lsn)
699 {
700 	int	i;
701 
702 	spin_lock(&ailp->ail_lock);
703 	/* xfs_trans_ail_update_bulk drops ailp->ail_lock */
704 	xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);
705 
706 	for (i = 0; i < nr_items; i++) {
707 		struct xfs_log_item *lip = log_items[i];
708 
709 		if (lip->li_ops->iop_unpin)
710 			lip->li_ops->iop_unpin(lip, 0);
711 	}
712 }
713 
714 /*
715  * Bulk operation version of xfs_trans_committed that takes a log vector of
716  * items to insert into the AIL. This uses bulk AIL insertion techniques to
717  * minimise lock traffic.
718  *
719  * If we are called with the aborted flag set, it is because a log write during
720  * a CIL checkpoint commit has failed. In this case, all the items in the
721  * checkpoint have already gone through iop_committed and iop_committing, which
722  * means that checkpoint commit abort handling is treated exactly the same
723  * as an iclog write error even though we haven't started any IO yet. Hence in
724  * this case all we need to do is iop_committed processing, followed by an
725  * iop_unpin(aborted) call.
726  *
727  * The AIL cursor is used to optimise the insert process. If commit_lsn is not
728  * at the end of the AIL, the insert cursor avoids the need to walk
729  * the AIL to find the insertion point on every xfs_log_item_batch_insert()
730  * call. This saves a lot of needless list walking and is a net win, even
731  * though it slightly increases that amount of AIL lock traffic to set it up
732  * and tear it down.
733  */
734 void
xfs_trans_committed_bulk(struct xfs_ail * ailp,struct xfs_log_vec * log_vector,xfs_lsn_t commit_lsn,bool aborted)735 xfs_trans_committed_bulk(
736 	struct xfs_ail		*ailp,
737 	struct xfs_log_vec	*log_vector,
738 	xfs_lsn_t		commit_lsn,
739 	bool			aborted)
740 {
741 #define LOG_ITEM_BATCH_SIZE	32
742 	struct xfs_log_item	*log_items[LOG_ITEM_BATCH_SIZE];
743 	struct xfs_log_vec	*lv;
744 	struct xfs_ail_cursor	cur;
745 	int			i = 0;
746 
747 	spin_lock(&ailp->ail_lock);
748 	xfs_trans_ail_cursor_last(ailp, &cur, commit_lsn);
749 	spin_unlock(&ailp->ail_lock);
750 
751 	/* unpin all the log items */
752 	for (lv = log_vector; lv; lv = lv->lv_next ) {
753 		struct xfs_log_item	*lip = lv->lv_item;
754 		xfs_lsn_t		item_lsn;
755 
756 		if (aborted)
757 			set_bit(XFS_LI_ABORTED, &lip->li_flags);
758 
759 		if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {
760 			lip->li_ops->iop_release(lip);
761 			continue;
762 		}
763 
764 		if (lip->li_ops->iop_committed)
765 			item_lsn = lip->li_ops->iop_committed(lip, commit_lsn);
766 		else
767 			item_lsn = commit_lsn;
768 
769 		/* item_lsn of -1 means the item needs no further processing */
770 		if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
771 			continue;
772 
773 		/*
774 		 * if we are aborting the operation, no point in inserting the
775 		 * object into the AIL as we are in a shutdown situation.
776 		 */
777 		if (aborted) {
778 			ASSERT(xfs_is_shutdown(ailp->ail_mount));
779 			if (lip->li_ops->iop_unpin)
780 				lip->li_ops->iop_unpin(lip, 1);
781 			continue;
782 		}
783 
784 		if (item_lsn != commit_lsn) {
785 
786 			/*
787 			 * Not a bulk update option due to unusual item_lsn.
788 			 * Push into AIL immediately, rechecking the lsn once
789 			 * we have the ail lock. Then unpin the item. This does
790 			 * not affect the AIL cursor the bulk insert path is
791 			 * using.
792 			 */
793 			spin_lock(&ailp->ail_lock);
794 			if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
795 				xfs_trans_ail_update(ailp, lip, item_lsn);
796 			else
797 				spin_unlock(&ailp->ail_lock);
798 			if (lip->li_ops->iop_unpin)
799 				lip->li_ops->iop_unpin(lip, 0);
800 			continue;
801 		}
802 
803 		/* Item is a candidate for bulk AIL insert.  */
804 		log_items[i++] = lv->lv_item;
805 		if (i >= LOG_ITEM_BATCH_SIZE) {
806 			xfs_log_item_batch_insert(ailp, &cur, log_items,
807 					LOG_ITEM_BATCH_SIZE, commit_lsn);
808 			i = 0;
809 		}
810 	}
811 
812 	/* make sure we insert the remainder! */
813 	if (i)
814 		xfs_log_item_batch_insert(ailp, &cur, log_items, i, commit_lsn);
815 
816 	spin_lock(&ailp->ail_lock);
817 	xfs_trans_ail_cursor_done(&cur);
818 	spin_unlock(&ailp->ail_lock);
819 }
820 
821 /*
822  * Commit the given transaction to the log.
823  *
824  * XFS disk error handling mechanism is not based on a typical
825  * transaction abort mechanism. Logically after the filesystem
826  * gets marked 'SHUTDOWN', we can't let any new transactions
827  * be durable - ie. committed to disk - because some metadata might
828  * be inconsistent. In such cases, this returns an error, and the
829  * caller may assume that all locked objects joined to the transaction
830  * have already been unlocked as if the commit had succeeded.
831  * Do not reference the transaction structure after this call.
832  */
833 static int
__xfs_trans_commit(struct xfs_trans * tp,bool regrant)834 __xfs_trans_commit(
835 	struct xfs_trans	*tp,
836 	bool			regrant)
837 {
838 	struct xfs_mount	*mp = tp->t_mountp;
839 	xfs_csn_t		commit_seq = 0;
840 	int			error = 0;
841 	int			sync = tp->t_flags & XFS_TRANS_SYNC;
842 
843 	trace_xfs_trans_commit(tp, _RET_IP_);
844 
845 	/*
846 	 * Finish deferred items on final commit. Only permanent transactions
847 	 * should ever have deferred ops.
848 	 */
849 	WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
850 		     !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
851 	if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
852 		error = xfs_defer_finish_noroll(&tp);
853 		if (error)
854 			goto out_unreserve;
855 	}
856 
857 	/*
858 	 * If there is nothing to be logged by the transaction,
859 	 * then unlock all of the items associated with the
860 	 * transaction and free the transaction structure.
861 	 * Also make sure to return any reserved blocks to
862 	 * the free pool.
863 	 */
864 	if (!(tp->t_flags & XFS_TRANS_DIRTY))
865 		goto out_unreserve;
866 
867 	if (xfs_is_shutdown(mp)) {
868 		error = -EIO;
869 		goto out_unreserve;
870 	}
871 
872 	ASSERT(tp->t_ticket != NULL);
873 
874 	/*
875 	 * If we need to update the superblock, then do it now.
876 	 */
877 	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
878 		xfs_trans_apply_sb_deltas(tp);
879 	xfs_trans_apply_dquot_deltas(tp);
880 
881 	xlog_cil_commit(mp->m_log, tp, &commit_seq, regrant);
882 
883 	xfs_trans_free(tp);
884 
885 	/*
886 	 * If the transaction needs to be synchronous, then force the
887 	 * log out now and wait for it.
888 	 */
889 	if (sync) {
890 		error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
891 		XFS_STATS_INC(mp, xs_trans_sync);
892 	} else {
893 		XFS_STATS_INC(mp, xs_trans_async);
894 	}
895 
896 	return error;
897 
898 out_unreserve:
899 	xfs_trans_unreserve_and_mod_sb(tp);
900 
901 	/*
902 	 * It is indeed possible for the transaction to be not dirty but
903 	 * the dqinfo portion to be.  All that means is that we have some
904 	 * (non-persistent) quota reservations that need to be unreserved.
905 	 */
906 	xfs_trans_unreserve_and_mod_dquots(tp);
907 	if (tp->t_ticket) {
908 		if (regrant && !xlog_is_shutdown(mp->m_log))
909 			xfs_log_ticket_regrant(mp->m_log, tp->t_ticket);
910 		else
911 			xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
912 		tp->t_ticket = NULL;
913 	}
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 
921 int
xfs_trans_commit(struct xfs_trans * tp)922 xfs_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  */
936 void
xfs_trans_cancel(struct xfs_trans * tp)937 xfs_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_is_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_is_shutdown(mp)) {
959 		struct xfs_log_item *lip;
960 
961 		list_for_each_entry(lip, &tp->t_items, li_trans)
962 			ASSERT(!xlog_item_is_intent_done(lip));
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 	xfs_trans_free_items(tp, dirty);
974 	xfs_trans_free(tp);
975 }
976 
977 /*
978  * Roll from one trans in the sequence of PERMANENT transactions to
979  * the next: permanent transactions are only flushed out when
980  * committed with xfs_trans_commit(), but we still want as soon
981  * as possible to let chunks of it go to the log. So we commit the
982  * chunk we've been working on and get a new transaction to continue.
983  */
984 int
xfs_trans_roll(struct xfs_trans ** tpp)985 xfs_trans_roll(
986 	struct xfs_trans	**tpp)
987 {
988 	struct xfs_trans	*trans = *tpp;
989 	struct xfs_trans_res	tres;
990 	int			error;
991 
992 	trace_xfs_trans_roll(trans, _RET_IP_);
993 
994 	/*
995 	 * Copy the critical parameters from one trans to the next.
996 	 */
997 	tres.tr_logres = trans->t_log_res;
998 	tres.tr_logcount = trans->t_log_count;
999 
1000 	*tpp = xfs_trans_dup(trans);
1001 
1002 	/*
1003 	 * Commit the current transaction.
1004 	 * If this commit failed, then it'd just unlock those items that
1005 	 * are not marked ihold. That also means that a filesystem shutdown
1006 	 * is in progress. The caller takes the responsibility to cancel
1007 	 * the duplicate transaction that gets returned.
1008 	 */
1009 	error = __xfs_trans_commit(trans, true);
1010 	if (error)
1011 		return error;
1012 
1013 	/*
1014 	 * Reserve space in the log for the next transaction.
1015 	 * This also pushes items in the "AIL", the list of logged items,
1016 	 * out to disk if they are taking up space at the tail of the log
1017 	 * that we want to use.  This requires that either nothing be locked
1018 	 * across this call, or that anything that is locked be logged in
1019 	 * the prior and the next transactions.
1020 	 */
1021 	tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1022 	return xfs_trans_reserve(*tpp, &tres, 0, 0);
1023 }
1024 
1025 /*
1026  * Allocate an transaction, lock and join the inode to it, and reserve quota.
1027  *
1028  * The caller must ensure that the on-disk dquots attached to this inode have
1029  * already been allocated and initialized.  The caller is responsible for
1030  * releasing ILOCK_EXCL if a new transaction is returned.
1031  */
1032 int
xfs_trans_alloc_inode(struct xfs_inode * ip,struct xfs_trans_res * resv,unsigned int dblocks,unsigned int rblocks,bool force,struct xfs_trans ** tpp)1033 xfs_trans_alloc_inode(
1034 	struct xfs_inode	*ip,
1035 	struct xfs_trans_res	*resv,
1036 	unsigned int		dblocks,
1037 	unsigned int		rblocks,
1038 	bool			force,
1039 	struct xfs_trans	**tpp)
1040 {
1041 	struct xfs_trans	*tp;
1042 	struct xfs_mount	*mp = ip->i_mount;
1043 	bool			retried = false;
1044 	int			error;
1045 
1046 retry:
1047 	error = xfs_trans_alloc(mp, resv, dblocks,
1048 			rblocks / mp->m_sb.sb_rextsize,
1049 			force ? XFS_TRANS_RESERVE : 0, &tp);
1050 	if (error)
1051 		return error;
1052 
1053 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1054 	xfs_trans_ijoin(tp, ip, 0);
1055 
1056 	error = xfs_qm_dqattach_locked(ip, false);
1057 	if (error) {
1058 		/* Caller should have allocated the dquots! */
1059 		ASSERT(error != -ENOENT);
1060 		goto out_cancel;
1061 	}
1062 
1063 	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1064 	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1065 		xfs_trans_cancel(tp);
1066 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1067 		xfs_blockgc_free_quota(ip, 0);
1068 		retried = true;
1069 		goto retry;
1070 	}
1071 	if (error)
1072 		goto out_cancel;
1073 
1074 	*tpp = tp;
1075 	return 0;
1076 
1077 out_cancel:
1078 	xfs_trans_cancel(tp);
1079 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1080 	return error;
1081 }
1082 
1083 /*
1084  * Allocate an transaction in preparation for inode creation by reserving quota
1085  * against the given dquots.  Callers are not required to hold any inode locks.
1086  */
1087 int
xfs_trans_alloc_icreate(struct xfs_mount * mp,struct xfs_trans_res * resv,struct xfs_dquot * udqp,struct xfs_dquot * gdqp,struct xfs_dquot * pdqp,unsigned int dblocks,struct xfs_trans ** tpp)1088 xfs_trans_alloc_icreate(
1089 	struct xfs_mount	*mp,
1090 	struct xfs_trans_res	*resv,
1091 	struct xfs_dquot	*udqp,
1092 	struct xfs_dquot	*gdqp,
1093 	struct xfs_dquot	*pdqp,
1094 	unsigned int		dblocks,
1095 	struct xfs_trans	**tpp)
1096 {
1097 	struct xfs_trans	*tp;
1098 	bool			retried = false;
1099 	int			error;
1100 
1101 retry:
1102 	error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1103 	if (error)
1104 		return error;
1105 
1106 	error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1107 	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1108 		xfs_trans_cancel(tp);
1109 		xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1110 		retried = true;
1111 		goto retry;
1112 	}
1113 	if (error) {
1114 		xfs_trans_cancel(tp);
1115 		return error;
1116 	}
1117 
1118 	*tpp = tp;
1119 	return 0;
1120 }
1121 
1122 /*
1123  * Allocate an transaction, lock and join the inode to it, and reserve quota
1124  * in preparation for inode attribute changes that include uid, gid, or prid
1125  * changes.
1126  *
1127  * The caller must ensure that the on-disk dquots attached to this inode have
1128  * already been allocated and initialized.  The ILOCK will be dropped when the
1129  * transaction is committed or cancelled.
1130  */
1131 int
xfs_trans_alloc_ichange(struct xfs_inode * ip,struct xfs_dquot * new_udqp,struct xfs_dquot * new_gdqp,struct xfs_dquot * new_pdqp,bool force,struct xfs_trans ** tpp)1132 xfs_trans_alloc_ichange(
1133 	struct xfs_inode	*ip,
1134 	struct xfs_dquot	*new_udqp,
1135 	struct xfs_dquot	*new_gdqp,
1136 	struct xfs_dquot	*new_pdqp,
1137 	bool			force,
1138 	struct xfs_trans	**tpp)
1139 {
1140 	struct xfs_trans	*tp;
1141 	struct xfs_mount	*mp = ip->i_mount;
1142 	struct xfs_dquot	*udqp;
1143 	struct xfs_dquot	*gdqp;
1144 	struct xfs_dquot	*pdqp;
1145 	bool			retried = false;
1146 	int			error;
1147 
1148 retry:
1149 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1150 	if (error)
1151 		return error;
1152 
1153 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1154 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1155 
1156 	error = xfs_qm_dqattach_locked(ip, false);
1157 	if (error) {
1158 		/* Caller should have allocated the dquots! */
1159 		ASSERT(error != -ENOENT);
1160 		goto out_cancel;
1161 	}
1162 
1163 	/*
1164 	 * For each quota type, skip quota reservations if the inode's dquots
1165 	 * now match the ones that came from the caller, or the caller didn't
1166 	 * pass one in.  The inode's dquots can change if we drop the ILOCK to
1167 	 * perform a blockgc scan, so we must preserve the caller's arguments.
1168 	 */
1169 	udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1170 	gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1171 	pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1172 	if (udqp || gdqp || pdqp) {
1173 		unsigned int	qflags = XFS_QMOPT_RES_REGBLKS;
1174 
1175 		if (force)
1176 			qflags |= XFS_QMOPT_FORCE_RES;
1177 
1178 		/*
1179 		 * Reserve enough quota to handle blocks on disk and reserved
1180 		 * for a delayed allocation.  We'll actually transfer the
1181 		 * delalloc reservation between dquots at chown time, even
1182 		 * though that part is only semi-transactional.
1183 		 */
1184 		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1185 				pdqp, ip->i_nblocks + ip->i_delayed_blks,
1186 				1, qflags);
1187 		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1188 			xfs_trans_cancel(tp);
1189 			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1190 			retried = true;
1191 			goto retry;
1192 		}
1193 		if (error)
1194 			goto out_cancel;
1195 	}
1196 
1197 	*tpp = tp;
1198 	return 0;
1199 
1200 out_cancel:
1201 	xfs_trans_cancel(tp);
1202 	return error;
1203 }
1204