xref: /linux/fs/xfs/xfs_trans.c (revision 9208c05f9fdfd927ea160b97dfef3c379049fff2)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
4  * Copyright (C) 2010 Red Hat, Inc.
5  * All Rights Reserved.
6  */
7 #include "xfs.h"
8 #include "xfs_fs.h"
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_extent_busy.h"
15 #include "xfs_quota.h"
16 #include "xfs_trans.h"
17 #include "xfs_trans_priv.h"
18 #include "xfs_log.h"
19 #include "xfs_log_priv.h"
20 #include "xfs_trace.h"
21 #include "xfs_error.h"
22 #include "xfs_defer.h"
23 #include "xfs_inode.h"
24 #include "xfs_dquot_item.h"
25 #include "xfs_dquot.h"
26 #include "xfs_icache.h"
27 #include "xfs_rtbitmap.h"
28 #include "xfs_rtgroup.h"
29 #include "xfs_sb.h"
30 
31 struct kmem_cache	*xfs_trans_cache;
32 
33 #if defined(CONFIG_TRACEPOINTS)
34 static void
35 xfs_trans_trace_reservations(
36 	struct xfs_mount	*mp)
37 {
38 	struct xfs_trans_res	*res;
39 	struct xfs_trans_res	*end_res;
40 	int			i;
41 
42 	res = (struct xfs_trans_res *)M_RES(mp);
43 	end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
44 	for (i = 0; res < end_res; i++, res++)
45 		trace_xfs_trans_resv_calc(mp, i, res);
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
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
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_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 *
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_highest_agno = NULLAGNUMBER;
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
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_dec_fdblocks(mp, 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, resp->tr_logres,
198 						resp->tr_logcount,
199 						&tp->t_ticket, permanent);
200 		}
201 
202 		if (error)
203 			goto undo_blocks;
204 
205 		tp->t_log_res = resp->tr_logres;
206 		tp->t_log_count = resp->tr_logcount;
207 	}
208 
209 	/*
210 	 * Attempt to reserve the needed realtime extents by decrementing
211 	 * the number needed from the number available.  This will
212 	 * fail if the count would go below zero.
213 	 */
214 	if (rtextents > 0) {
215 		error = xfs_dec_frextents(mp, rtextents);
216 		if (error) {
217 			error = -ENOSPC;
218 			goto undo_log;
219 		}
220 		tp->t_rtx_res += rtextents;
221 	}
222 
223 	return 0;
224 
225 	/*
226 	 * Error cases jump to one of these labels to undo any
227 	 * reservations which have already been performed.
228 	 */
229 undo_log:
230 	if (resp->tr_logres > 0) {
231 		xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
232 		tp->t_ticket = NULL;
233 		tp->t_log_res = 0;
234 		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
235 	}
236 
237 undo_blocks:
238 	if (blocks > 0) {
239 		xfs_add_fdblocks(mp, blocks);
240 		tp->t_blk_res = 0;
241 	}
242 	return error;
243 }
244 
245 int
246 xfs_trans_alloc(
247 	struct xfs_mount	*mp,
248 	struct xfs_trans_res	*resp,
249 	uint			blocks,
250 	uint			rtextents,
251 	uint			flags,
252 	struct xfs_trans	**tpp)
253 {
254 	struct xfs_trans	*tp;
255 	bool			want_retry = true;
256 	int			error;
257 
258 	/*
259 	 * Allocate the handle before we do our freeze accounting and setting up
260 	 * GFP_NOFS allocation context so that we avoid lockdep false positives
261 	 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
262 	 */
263 retry:
264 	tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
265 	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
266 		sb_start_intwrite(mp->m_super);
267 	xfs_trans_set_context(tp);
268 
269 	/*
270 	 * Zero-reservation ("empty") transactions can't modify anything, so
271 	 * they're allowed to run while we're frozen.
272 	 */
273 	WARN_ON(resp->tr_logres > 0 &&
274 		mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
275 	ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||
276 	       xfs_has_lazysbcount(mp));
277 
278 	tp->t_magic = XFS_TRANS_HEADER_MAGIC;
279 	tp->t_flags = flags;
280 	tp->t_mountp = mp;
281 	INIT_LIST_HEAD(&tp->t_items);
282 	INIT_LIST_HEAD(&tp->t_busy);
283 	INIT_LIST_HEAD(&tp->t_dfops);
284 	tp->t_highest_agno = NULLAGNUMBER;
285 
286 	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
287 	if (error == -ENOSPC && want_retry) {
288 		xfs_trans_cancel(tp);
289 
290 		/*
291 		 * We weren't able to reserve enough space for the transaction.
292 		 * Flush the other speculative space allocations to free space.
293 		 * Do not perform a synchronous scan because callers can hold
294 		 * other locks.
295 		 */
296 		error = xfs_blockgc_flush_all(mp);
297 		if (error)
298 			return error;
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
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
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 		if (xfs_has_rtgroups(mp))
426 			flags &= ~XFS_TRANS_SB_DIRTY;
427 		break;
428 	case XFS_TRANS_SB_RES_FREXTENTS:
429 		/*
430 		 * The allocation has already been applied to the
431 		 * in-core superblock's counter.  This should only
432 		 * be applied to the on-disk superblock.
433 		 */
434 		ASSERT(delta < 0);
435 		tp->t_res_frextents_delta += delta;
436 		if (xfs_has_rtgroups(mp))
437 			flags &= ~XFS_TRANS_SB_DIRTY;
438 		break;
439 	case XFS_TRANS_SB_DBLOCKS:
440 		tp->t_dblocks_delta += delta;
441 		break;
442 	case XFS_TRANS_SB_AGCOUNT:
443 		ASSERT(delta > 0);
444 		tp->t_agcount_delta += delta;
445 		break;
446 	case XFS_TRANS_SB_IMAXPCT:
447 		tp->t_imaxpct_delta += delta;
448 		break;
449 	case XFS_TRANS_SB_REXTSIZE:
450 		tp->t_rextsize_delta += delta;
451 		break;
452 	case XFS_TRANS_SB_RBMBLOCKS:
453 		tp->t_rbmblocks_delta += delta;
454 		break;
455 	case XFS_TRANS_SB_RBLOCKS:
456 		tp->t_rblocks_delta += delta;
457 		break;
458 	case XFS_TRANS_SB_REXTENTS:
459 		tp->t_rextents_delta += delta;
460 		break;
461 	case XFS_TRANS_SB_REXTSLOG:
462 		tp->t_rextslog_delta += delta;
463 		break;
464 	case XFS_TRANS_SB_RGCOUNT:
465 		ASSERT(delta > 0);
466 		tp->t_rgcount_delta += delta;
467 		break;
468 	default:
469 		ASSERT(0);
470 		return;
471 	}
472 
473 	tp->t_flags |= flags;
474 }
475 
476 /*
477  * xfs_trans_apply_sb_deltas() is called from the commit code
478  * to bring the superblock buffer into the current transaction
479  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
480  *
481  * For now we just look at each field allowed to change and change
482  * it if necessary.
483  */
484 STATIC void
485 xfs_trans_apply_sb_deltas(
486 	xfs_trans_t	*tp)
487 {
488 	struct xfs_dsb	*sbp;
489 	struct xfs_buf	*bp;
490 	int		whole = 0;
491 
492 	bp = xfs_trans_getsb(tp);
493 	sbp = bp->b_addr;
494 
495 	/*
496 	 * Only update the superblock counters if we are logging them
497 	 */
498 	if (!xfs_has_lazysbcount((tp->t_mountp))) {
499 		if (tp->t_icount_delta)
500 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
501 		if (tp->t_ifree_delta)
502 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
503 		if (tp->t_fdblocks_delta)
504 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
505 		if (tp->t_res_fdblocks_delta)
506 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
507 	}
508 
509 	/*
510 	 * sb_frextents was added to the lazy sb counters when the rt groups
511 	 * feature was introduced.  This is possible because we know that all
512 	 * kernels supporting rtgroups will also recompute frextents from the
513 	 * realtime bitmap.
514 	 *
515 	 * For older file systems, updating frextents requires careful handling
516 	 * because we cannot rely on log recovery in older kernels to recompute
517 	 * the value from the rtbitmap.  This means that the ondisk frextents
518 	 * must be consistent with the rtbitmap.
519 	 *
520 	 * Therefore, log the frextents change to the ondisk superblock and
521 	 * update the incore superblock so that future calls to xfs_log_sb
522 	 * write the correct value ondisk.
523 	 */
524 	if ((tp->t_frextents_delta || tp->t_res_frextents_delta) &&
525 	    !xfs_has_rtgroups(tp->t_mountp)) {
526 		struct xfs_mount	*mp = tp->t_mountp;
527 		int64_t			rtxdelta;
528 
529 		rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;
530 
531 		spin_lock(&mp->m_sb_lock);
532 		be64_add_cpu(&sbp->sb_frextents, rtxdelta);
533 		mp->m_sb.sb_frextents += rtxdelta;
534 		spin_unlock(&mp->m_sb_lock);
535 	}
536 
537 	if (tp->t_dblocks_delta) {
538 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
539 		whole = 1;
540 	}
541 	if (tp->t_agcount_delta) {
542 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
543 		whole = 1;
544 	}
545 	if (tp->t_imaxpct_delta) {
546 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
547 		whole = 1;
548 	}
549 	if (tp->t_rextsize_delta) {
550 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
551 
552 		/*
553 		 * Because the ondisk sb records rtgroup size in units of rt
554 		 * extents, any time we update the rt extent size we have to
555 		 * recompute the ondisk rtgroup block log.  The incore values
556 		 * will be recomputed in xfs_trans_unreserve_and_mod_sb.
557 		 */
558 		if (xfs_has_rtgroups(tp->t_mountp)) {
559 			sbp->sb_rgblklog = xfs_compute_rgblklog(
560 						be32_to_cpu(sbp->sb_rgextents),
561 						be32_to_cpu(sbp->sb_rextsize));
562 		}
563 		whole = 1;
564 	}
565 	if (tp->t_rbmblocks_delta) {
566 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
567 		whole = 1;
568 	}
569 	if (tp->t_rblocks_delta) {
570 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
571 		whole = 1;
572 	}
573 	if (tp->t_rextents_delta) {
574 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
575 		whole = 1;
576 	}
577 	if (tp->t_rextslog_delta) {
578 		sbp->sb_rextslog += tp->t_rextslog_delta;
579 		whole = 1;
580 	}
581 	if (tp->t_rgcount_delta) {
582 		be32_add_cpu(&sbp->sb_rgcount, tp->t_rgcount_delta);
583 		whole = 1;
584 	}
585 
586 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
587 	if (whole)
588 		/*
589 		 * Log the whole thing, the fields are noncontiguous.
590 		 */
591 		xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
592 	else
593 		/*
594 		 * Since all the modifiable fields are contiguous, we
595 		 * can get away with this.
596 		 */
597 		xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
598 				  offsetof(struct xfs_dsb, sb_frextents) +
599 				  sizeof(sbp->sb_frextents) - 1);
600 }
601 
602 /*
603  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
604  * apply superblock counter changes to the in-core superblock.  The
605  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
606  * applied to the in-core superblock.  The idea is that that has already been
607  * done.
608  *
609  * If we are not logging superblock counters, then the inode allocated/free and
610  * used block counts are not updated in the on disk superblock. In this case,
611  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
612  * still need to update the incore superblock with the changes.
613  *
614  * Deltas for the inode count are +/-64, hence we use a large batch size of 128
615  * so we don't need to take the counter lock on every update.
616  */
617 #define XFS_ICOUNT_BATCH	128
618 
619 void
620 xfs_trans_unreserve_and_mod_sb(
621 	struct xfs_trans	*tp)
622 {
623 	struct xfs_mount	*mp = tp->t_mountp;
624 	int64_t			blkdelta = tp->t_blk_res;
625 	int64_t			rtxdelta = tp->t_rtx_res;
626 	int64_t			idelta = 0;
627 	int64_t			ifreedelta = 0;
628 
629 	/*
630 	 * Calculate the deltas.
631 	 *
632 	 * t_fdblocks_delta and t_frextents_delta can be positive or negative:
633 	 *
634 	 *  - positive values indicate blocks freed in the transaction.
635 	 *  - negative values indicate blocks allocated in the transaction
636 	 *
637 	 * Negative values can only happen if the transaction has a block
638 	 * reservation that covers the allocated block.  The end result is
639 	 * that the calculated delta values must always be positive and we
640 	 * can only put back previous allocated or reserved blocks here.
641 	 */
642 	ASSERT(tp->t_blk_res || tp->t_fdblocks_delta >= 0);
643 	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
644 	        blkdelta += tp->t_fdblocks_delta;
645 		ASSERT(blkdelta >= 0);
646 	}
647 
648 	ASSERT(tp->t_rtx_res || tp->t_frextents_delta >= 0);
649 	if (xfs_has_rtgroups(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
650 		rtxdelta += tp->t_frextents_delta;
651 		ASSERT(rtxdelta >= 0);
652 	}
653 
654 	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
655 		idelta = tp->t_icount_delta;
656 		ifreedelta = tp->t_ifree_delta;
657 	}
658 
659 	/* apply the per-cpu counters */
660 	if (blkdelta)
661 		xfs_add_fdblocks(mp, blkdelta);
662 
663 	if (idelta)
664 		percpu_counter_add_batch(&mp->m_icount, idelta,
665 					 XFS_ICOUNT_BATCH);
666 
667 	if (ifreedelta)
668 		percpu_counter_add(&mp->m_ifree, ifreedelta);
669 
670 	if (rtxdelta)
671 		xfs_add_frextents(mp, rtxdelta);
672 
673 	if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))
674 		return;
675 
676 	/* apply remaining deltas */
677 	spin_lock(&mp->m_sb_lock);
678 	mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
679 	mp->m_sb.sb_icount += idelta;
680 	mp->m_sb.sb_ifree += ifreedelta;
681 	/*
682 	 * Do not touch sb_frextents here because it is handled in
683 	 * xfs_trans_apply_sb_deltas for file systems where it isn't a lazy
684 	 * counter anyway.
685 	 */
686 	mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
687 	mp->m_sb.sb_agcount += tp->t_agcount_delta;
688 	mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
689 	if (tp->t_rextsize_delta)
690 		xfs_mount_sb_set_rextsize(mp, &mp->m_sb,
691 				mp->m_sb.sb_rextsize + tp->t_rextsize_delta);
692 	mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
693 	mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
694 	mp->m_sb.sb_rextents += tp->t_rextents_delta;
695 	mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
696 	mp->m_sb.sb_rgcount += tp->t_rgcount_delta;
697 	spin_unlock(&mp->m_sb_lock);
698 
699 	/*
700 	 * Debug checks outside of the spinlock so they don't lock up the
701 	 * machine if they fail.
702 	 */
703 	ASSERT(mp->m_sb.sb_imax_pct >= 0);
704 	ASSERT(mp->m_sb.sb_rextslog >= 0);
705 }
706 
707 /* Add the given log item to the transaction's list of log items. */
708 void
709 xfs_trans_add_item(
710 	struct xfs_trans	*tp,
711 	struct xfs_log_item	*lip)
712 {
713 	ASSERT(lip->li_log == tp->t_mountp->m_log);
714 	ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
715 	ASSERT(list_empty(&lip->li_trans));
716 	ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
717 
718 	list_add_tail(&lip->li_trans, &tp->t_items);
719 	trace_xfs_trans_add_item(tp, _RET_IP_);
720 }
721 
722 /*
723  * Unlink the log item from the transaction. the log item is no longer
724  * considered dirty in this transaction, as the linked transaction has
725  * finished, either by abort or commit completion.
726  */
727 void
728 xfs_trans_del_item(
729 	struct xfs_log_item	*lip)
730 {
731 	clear_bit(XFS_LI_DIRTY, &lip->li_flags);
732 	list_del_init(&lip->li_trans);
733 }
734 
735 /* Detach and unlock all of the items in a transaction */
736 static void
737 xfs_trans_free_items(
738 	struct xfs_trans	*tp,
739 	bool			abort)
740 {
741 	struct xfs_log_item	*lip, *next;
742 
743 	trace_xfs_trans_free_items(tp, _RET_IP_);
744 
745 	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
746 		xfs_trans_del_item(lip);
747 		if (abort)
748 			set_bit(XFS_LI_ABORTED, &lip->li_flags);
749 		if (lip->li_ops->iop_release)
750 			lip->li_ops->iop_release(lip);
751 	}
752 }
753 
754 /*
755  * Sort transaction items prior to running precommit operations. This will
756  * attempt to order the items such that they will always be locked in the same
757  * order. Items that have no sort function are moved to the end of the list
758  * and so are locked last.
759  *
760  * This may need refinement as different types of objects add sort functions.
761  *
762  * Function is more complex than it needs to be because we are comparing 64 bit
763  * values and the function only returns 32 bit values.
764  */
765 static int
766 xfs_trans_precommit_sort(
767 	void			*unused_arg,
768 	const struct list_head	*a,
769 	const struct list_head	*b)
770 {
771 	struct xfs_log_item	*lia = container_of(a,
772 					struct xfs_log_item, li_trans);
773 	struct xfs_log_item	*lib = container_of(b,
774 					struct xfs_log_item, li_trans);
775 	int64_t			diff;
776 
777 	/*
778 	 * If both items are non-sortable, leave them alone. If only one is
779 	 * sortable, move the non-sortable item towards the end of the list.
780 	 */
781 	if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)
782 		return 0;
783 	if (!lia->li_ops->iop_sort)
784 		return 1;
785 	if (!lib->li_ops->iop_sort)
786 		return -1;
787 
788 	diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);
789 	if (diff < 0)
790 		return -1;
791 	if (diff > 0)
792 		return 1;
793 	return 0;
794 }
795 
796 /*
797  * Run transaction precommit functions.
798  *
799  * If there is an error in any of the callouts, then stop immediately and
800  * trigger a shutdown to abort the transaction. There is no recovery possible
801  * from errors at this point as the transaction is dirty....
802  */
803 static int
804 xfs_trans_run_precommits(
805 	struct xfs_trans	*tp)
806 {
807 	struct xfs_mount	*mp = tp->t_mountp;
808 	struct xfs_log_item	*lip, *n;
809 	int			error = 0;
810 
811 	/*
812 	 * Sort the item list to avoid ABBA deadlocks with other transactions
813 	 * running precommit operations that lock multiple shared items such as
814 	 * inode cluster buffers.
815 	 */
816 	list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);
817 
818 	/*
819 	 * Precommit operations can remove the log item from the transaction
820 	 * if the log item exists purely to delay modifications until they
821 	 * can be ordered against other operations. Hence we have to use
822 	 * list_for_each_entry_safe() here.
823 	 */
824 	list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {
825 		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
826 			continue;
827 		if (lip->li_ops->iop_precommit) {
828 			error = lip->li_ops->iop_precommit(tp, lip);
829 			if (error)
830 				break;
831 		}
832 	}
833 	if (error)
834 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
835 	return error;
836 }
837 
838 /*
839  * Commit the given transaction to the log.
840  *
841  * XFS disk error handling mechanism is not based on a typical
842  * transaction abort mechanism. Logically after the filesystem
843  * gets marked 'SHUTDOWN', we can't let any new transactions
844  * be durable - ie. committed to disk - because some metadata might
845  * be inconsistent. In such cases, this returns an error, and the
846  * caller may assume that all locked objects joined to the transaction
847  * have already been unlocked as if the commit had succeeded.
848  * Do not reference the transaction structure after this call.
849  */
850 static int
851 __xfs_trans_commit(
852 	struct xfs_trans	*tp,
853 	bool			regrant)
854 {
855 	struct xfs_mount	*mp = tp->t_mountp;
856 	struct xlog		*log = mp->m_log;
857 	xfs_csn_t		commit_seq = 0;
858 	int			error = 0;
859 	int			sync = tp->t_flags & XFS_TRANS_SYNC;
860 
861 	trace_xfs_trans_commit(tp, _RET_IP_);
862 
863 	error = xfs_trans_run_precommits(tp);
864 	if (error) {
865 		if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
866 			xfs_defer_cancel(tp);
867 		goto out_unreserve;
868 	}
869 
870 	/*
871 	 * Finish deferred items on final commit. Only permanent transactions
872 	 * should ever have deferred ops.
873 	 */
874 	WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
875 		     !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
876 	if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
877 		error = xfs_defer_finish_noroll(&tp);
878 		if (error)
879 			goto out_unreserve;
880 
881 		/* Run precommits from final tx in defer chain. */
882 		error = xfs_trans_run_precommits(tp);
883 		if (error)
884 			goto out_unreserve;
885 	}
886 
887 	/*
888 	 * If there is nothing to be logged by the transaction,
889 	 * then unlock all of the items associated with the
890 	 * transaction and free the transaction structure.
891 	 * Also make sure to return any reserved blocks to
892 	 * the free pool.
893 	 */
894 	if (!(tp->t_flags & XFS_TRANS_DIRTY))
895 		goto out_unreserve;
896 
897 	/*
898 	 * We must check against log shutdown here because we cannot abort log
899 	 * items and leave them dirty, inconsistent and unpinned in memory while
900 	 * the log is active. This leaves them open to being written back to
901 	 * disk, and that will lead to on-disk corruption.
902 	 */
903 	if (xlog_is_shutdown(log)) {
904 		error = -EIO;
905 		goto out_unreserve;
906 	}
907 
908 	ASSERT(tp->t_ticket != NULL);
909 
910 	/*
911 	 * If we need to update the superblock, then do it now.
912 	 */
913 	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
914 		xfs_trans_apply_sb_deltas(tp);
915 	xfs_trans_apply_dquot_deltas(tp);
916 
917 	xlog_cil_commit(log, tp, &commit_seq, regrant);
918 
919 	xfs_trans_free(tp);
920 
921 	/*
922 	 * If the transaction needs to be synchronous, then force the
923 	 * log out now and wait for it.
924 	 */
925 	if (sync) {
926 		error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
927 		XFS_STATS_INC(mp, xs_trans_sync);
928 	} else {
929 		XFS_STATS_INC(mp, xs_trans_async);
930 	}
931 
932 	return error;
933 
934 out_unreserve:
935 	xfs_trans_unreserve_and_mod_sb(tp);
936 
937 	/*
938 	 * It is indeed possible for the transaction to be not dirty but
939 	 * the dqinfo portion to be.  All that means is that we have some
940 	 * (non-persistent) quota reservations that need to be unreserved.
941 	 */
942 	xfs_trans_unreserve_and_mod_dquots(tp);
943 	if (tp->t_ticket) {
944 		if (regrant && !xlog_is_shutdown(log))
945 			xfs_log_ticket_regrant(log, tp->t_ticket);
946 		else
947 			xfs_log_ticket_ungrant(log, tp->t_ticket);
948 		tp->t_ticket = NULL;
949 	}
950 	xfs_trans_free_items(tp, !!error);
951 	xfs_trans_free(tp);
952 
953 	XFS_STATS_INC(mp, xs_trans_empty);
954 	return error;
955 }
956 
957 int
958 xfs_trans_commit(
959 	struct xfs_trans	*tp)
960 {
961 	return __xfs_trans_commit(tp, false);
962 }
963 
964 /*
965  * Unlock all of the transaction's items and free the transaction.  If the
966  * transaction is dirty, we must shut down the filesystem because there is no
967  * way to restore them to their previous state.
968  *
969  * If the transaction has made a log reservation, make sure to release it as
970  * well.
971  *
972  * This is a high level function (equivalent to xfs_trans_commit()) and so can
973  * be called after the transaction has effectively been aborted due to the mount
974  * being shut down. However, if the mount has not been shut down and the
975  * transaction is dirty we will shut the mount down and, in doing so, that
976  * guarantees that the log is shut down, too. Hence we don't need to be as
977  * careful with shutdown state and dirty items here as we need to be in
978  * xfs_trans_commit().
979  */
980 void
981 xfs_trans_cancel(
982 	struct xfs_trans	*tp)
983 {
984 	struct xfs_mount	*mp = tp->t_mountp;
985 	struct xlog		*log = mp->m_log;
986 	bool			dirty = (tp->t_flags & XFS_TRANS_DIRTY);
987 
988 	trace_xfs_trans_cancel(tp, _RET_IP_);
989 
990 	/*
991 	 * It's never valid to cancel a transaction with deferred ops attached,
992 	 * because the transaction is effectively dirty.  Complain about this
993 	 * loudly before freeing the in-memory defer items and shutting down the
994 	 * filesystem.
995 	 */
996 	if (!list_empty(&tp->t_dfops)) {
997 		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
998 		dirty = true;
999 		xfs_defer_cancel(tp);
1000 	}
1001 
1002 	/*
1003 	 * See if the caller is relying on us to shut down the filesystem. We
1004 	 * only want an error report if there isn't already a shutdown in
1005 	 * progress, so we only need to check against the mount shutdown state
1006 	 * here.
1007 	 */
1008 	if (dirty && !xfs_is_shutdown(mp)) {
1009 		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1010 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1011 	}
1012 #ifdef DEBUG
1013 	/* Log items need to be consistent until the log is shut down. */
1014 	if (!dirty && !xlog_is_shutdown(log)) {
1015 		struct xfs_log_item *lip;
1016 
1017 		list_for_each_entry(lip, &tp->t_items, li_trans)
1018 			ASSERT(!xlog_item_is_intent_done(lip));
1019 	}
1020 #endif
1021 	xfs_trans_unreserve_and_mod_sb(tp);
1022 	xfs_trans_unreserve_and_mod_dquots(tp);
1023 
1024 	if (tp->t_ticket) {
1025 		xfs_log_ticket_ungrant(log, tp->t_ticket);
1026 		tp->t_ticket = NULL;
1027 	}
1028 
1029 	xfs_trans_free_items(tp, dirty);
1030 	xfs_trans_free(tp);
1031 }
1032 
1033 /*
1034  * Roll from one trans in the sequence of PERMANENT transactions to
1035  * the next: permanent transactions are only flushed out when
1036  * committed with xfs_trans_commit(), but we still want as soon
1037  * as possible to let chunks of it go to the log. So we commit the
1038  * chunk we've been working on and get a new transaction to continue.
1039  */
1040 int
1041 xfs_trans_roll(
1042 	struct xfs_trans	**tpp)
1043 {
1044 	struct xfs_trans	*trans = *tpp;
1045 	struct xfs_trans_res	tres;
1046 	int			error;
1047 
1048 	trace_xfs_trans_roll(trans, _RET_IP_);
1049 
1050 	/*
1051 	 * Copy the critical parameters from one trans to the next.
1052 	 */
1053 	tres.tr_logres = trans->t_log_res;
1054 	tres.tr_logcount = trans->t_log_count;
1055 
1056 	*tpp = xfs_trans_dup(trans);
1057 
1058 	/*
1059 	 * Commit the current transaction.
1060 	 * If this commit failed, then it'd just unlock those items that
1061 	 * are not marked ihold. That also means that a filesystem shutdown
1062 	 * is in progress. The caller takes the responsibility to cancel
1063 	 * the duplicate transaction that gets returned.
1064 	 */
1065 	error = __xfs_trans_commit(trans, true);
1066 	if (error)
1067 		return error;
1068 
1069 	/*
1070 	 * Reserve space in the log for the next transaction.
1071 	 * This also pushes items in the "AIL", the list of logged items,
1072 	 * out to disk if they are taking up space at the tail of the log
1073 	 * that we want to use.  This requires that either nothing be locked
1074 	 * across this call, or that anything that is locked be logged in
1075 	 * the prior and the next transactions.
1076 	 */
1077 	tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1078 	return xfs_trans_reserve(*tpp, &tres, 0, 0);
1079 }
1080 
1081 /*
1082  * Allocate an transaction, lock and join the inode to it, and reserve quota.
1083  *
1084  * The caller must ensure that the on-disk dquots attached to this inode have
1085  * already been allocated and initialized.  The caller is responsible for
1086  * releasing ILOCK_EXCL if a new transaction is returned.
1087  */
1088 int
1089 xfs_trans_alloc_inode(
1090 	struct xfs_inode	*ip,
1091 	struct xfs_trans_res	*resv,
1092 	unsigned int		dblocks,
1093 	unsigned int		rblocks,
1094 	bool			force,
1095 	struct xfs_trans	**tpp)
1096 {
1097 	struct xfs_trans	*tp;
1098 	struct xfs_mount	*mp = ip->i_mount;
1099 	bool			retried = false;
1100 	int			error;
1101 
1102 retry:
1103 	error = xfs_trans_alloc(mp, resv, dblocks,
1104 			xfs_extlen_to_rtxlen(mp, rblocks),
1105 			force ? XFS_TRANS_RESERVE : 0, &tp);
1106 	if (error)
1107 		return error;
1108 
1109 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1110 	xfs_trans_ijoin(tp, ip, 0);
1111 
1112 	error = xfs_qm_dqattach_locked(ip, false);
1113 	if (error) {
1114 		/* Caller should have allocated the dquots! */
1115 		ASSERT(error != -ENOENT);
1116 		goto out_cancel;
1117 	}
1118 
1119 	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1120 	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1121 		xfs_trans_cancel(tp);
1122 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1123 		xfs_blockgc_free_quota(ip, 0);
1124 		retried = true;
1125 		goto retry;
1126 	}
1127 	if (error)
1128 		goto out_cancel;
1129 
1130 	*tpp = tp;
1131 	return 0;
1132 
1133 out_cancel:
1134 	xfs_trans_cancel(tp);
1135 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1136 	return error;
1137 }
1138 
1139 /*
1140  * Try to reserve more blocks for a transaction.
1141  *
1142  * This is for callers that need to attach resources to a transaction, scan
1143  * those resources to determine the space reservation requirements, and then
1144  * modify the attached resources.  In other words, online repair.  This can
1145  * fail due to ENOSPC, so the caller must be able to cancel the transaction
1146  * without shutting down the fs.
1147  */
1148 int
1149 xfs_trans_reserve_more(
1150 	struct xfs_trans	*tp,
1151 	unsigned int		blocks,
1152 	unsigned int		rtextents)
1153 {
1154 	struct xfs_trans_res	resv = { };
1155 
1156 	return xfs_trans_reserve(tp, &resv, blocks, rtextents);
1157 }
1158 
1159 /*
1160  * Try to reserve more blocks and file quota for a transaction.  Same
1161  * conditions of usage as xfs_trans_reserve_more.
1162  */
1163 int
1164 xfs_trans_reserve_more_inode(
1165 	struct xfs_trans	*tp,
1166 	struct xfs_inode	*ip,
1167 	unsigned int		dblocks,
1168 	unsigned int		rblocks,
1169 	bool			force_quota)
1170 {
1171 	struct xfs_trans_res	resv = { };
1172 	struct xfs_mount	*mp = ip->i_mount;
1173 	unsigned int		rtx = xfs_extlen_to_rtxlen(mp, rblocks);
1174 	int			error;
1175 
1176 	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1177 
1178 	error = xfs_trans_reserve(tp, &resv, dblocks, rtx);
1179 	if (error)
1180 		return error;
1181 
1182 	if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino))
1183 		return 0;
1184 
1185 	if (tp->t_flags & XFS_TRANS_RESERVE)
1186 		force_quota = true;
1187 
1188 	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,
1189 			force_quota);
1190 	if (!error)
1191 		return 0;
1192 
1193 	/* Quota failed, give back the new reservation. */
1194 	xfs_add_fdblocks(mp, dblocks);
1195 	tp->t_blk_res -= dblocks;
1196 	xfs_add_frextents(mp, rtx);
1197 	tp->t_rtx_res -= rtx;
1198 	return error;
1199 }
1200 
1201 /*
1202  * Allocate an transaction in preparation for inode creation by reserving quota
1203  * against the given dquots.  Callers are not required to hold any inode locks.
1204  */
1205 int
1206 xfs_trans_alloc_icreate(
1207 	struct xfs_mount	*mp,
1208 	struct xfs_trans_res	*resv,
1209 	struct xfs_dquot	*udqp,
1210 	struct xfs_dquot	*gdqp,
1211 	struct xfs_dquot	*pdqp,
1212 	unsigned int		dblocks,
1213 	struct xfs_trans	**tpp)
1214 {
1215 	struct xfs_trans	*tp;
1216 	bool			retried = false;
1217 	int			error;
1218 
1219 retry:
1220 	error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1221 	if (error)
1222 		return error;
1223 
1224 	error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1225 	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1226 		xfs_trans_cancel(tp);
1227 		xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1228 		retried = true;
1229 		goto retry;
1230 	}
1231 	if (error) {
1232 		xfs_trans_cancel(tp);
1233 		return error;
1234 	}
1235 
1236 	*tpp = tp;
1237 	return 0;
1238 }
1239 
1240 /*
1241  * Allocate an transaction, lock and join the inode to it, and reserve quota
1242  * in preparation for inode attribute changes that include uid, gid, or prid
1243  * changes.
1244  *
1245  * The caller must ensure that the on-disk dquots attached to this inode have
1246  * already been allocated and initialized.  The ILOCK will be dropped when the
1247  * transaction is committed or cancelled.
1248  */
1249 int
1250 xfs_trans_alloc_ichange(
1251 	struct xfs_inode	*ip,
1252 	struct xfs_dquot	*new_udqp,
1253 	struct xfs_dquot	*new_gdqp,
1254 	struct xfs_dquot	*new_pdqp,
1255 	bool			force,
1256 	struct xfs_trans	**tpp)
1257 {
1258 	struct xfs_trans	*tp;
1259 	struct xfs_mount	*mp = ip->i_mount;
1260 	struct xfs_dquot	*udqp;
1261 	struct xfs_dquot	*gdqp;
1262 	struct xfs_dquot	*pdqp;
1263 	bool			retried = false;
1264 	int			error;
1265 
1266 retry:
1267 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1268 	if (error)
1269 		return error;
1270 
1271 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1272 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1273 
1274 	error = xfs_qm_dqattach_locked(ip, false);
1275 	if (error) {
1276 		/* Caller should have allocated the dquots! */
1277 		ASSERT(error != -ENOENT);
1278 		goto out_cancel;
1279 	}
1280 
1281 	/*
1282 	 * For each quota type, skip quota reservations if the inode's dquots
1283 	 * now match the ones that came from the caller, or the caller didn't
1284 	 * pass one in.  The inode's dquots can change if we drop the ILOCK to
1285 	 * perform a blockgc scan, so we must preserve the caller's arguments.
1286 	 */
1287 	udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1288 	gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1289 	pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1290 	if (udqp || gdqp || pdqp) {
1291 		xfs_filblks_t	dblocks, rblocks;
1292 		unsigned int	qflags = XFS_QMOPT_RES_REGBLKS;
1293 		bool		isrt = XFS_IS_REALTIME_INODE(ip);
1294 
1295 		if (force)
1296 			qflags |= XFS_QMOPT_FORCE_RES;
1297 
1298 		if (isrt) {
1299 			error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1300 			if (error)
1301 				goto out_cancel;
1302 		}
1303 
1304 		xfs_inode_count_blocks(tp, ip, &dblocks, &rblocks);
1305 
1306 		if (isrt)
1307 			rblocks += ip->i_delayed_blks;
1308 		else
1309 			dblocks += ip->i_delayed_blks;
1310 
1311 		/*
1312 		 * Reserve enough quota to handle blocks on disk and reserved
1313 		 * for a delayed allocation.  We'll actually transfer the
1314 		 * delalloc reservation between dquots at chown time, even
1315 		 * though that part is only semi-transactional.
1316 		 */
1317 		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1318 				pdqp, dblocks, 1, qflags);
1319 		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1320 			xfs_trans_cancel(tp);
1321 			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1322 			retried = true;
1323 			goto retry;
1324 		}
1325 		if (error)
1326 			goto out_cancel;
1327 
1328 		/* Do the same for realtime. */
1329 		qflags = XFS_QMOPT_RES_RTBLKS | (qflags & XFS_QMOPT_FORCE_RES);
1330 		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1331 				pdqp, rblocks, 0, qflags);
1332 		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1333 			xfs_trans_cancel(tp);
1334 			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1335 			retried = true;
1336 			goto retry;
1337 		}
1338 		if (error)
1339 			goto out_cancel;
1340 	}
1341 
1342 	*tpp = tp;
1343 	return 0;
1344 
1345 out_cancel:
1346 	xfs_trans_cancel(tp);
1347 	return error;
1348 }
1349 
1350 /*
1351  * Allocate an transaction, lock and join the directory and child inodes to it,
1352  * and reserve quota for a directory update.  If there isn't sufficient space,
1353  * @dblocks will be set to zero for a reservationless directory update and
1354  * @nospace_error will be set to a negative errno describing the space
1355  * constraint we hit.
1356  *
1357  * The caller must ensure that the on-disk dquots attached to this inode have
1358  * already been allocated and initialized.  The ILOCKs will be dropped when the
1359  * transaction is committed or cancelled.
1360  *
1361  * Caller is responsible for unlocking the inodes manually upon return
1362  */
1363 int
1364 xfs_trans_alloc_dir(
1365 	struct xfs_inode	*dp,
1366 	struct xfs_trans_res	*resv,
1367 	struct xfs_inode	*ip,
1368 	unsigned int		*dblocks,
1369 	struct xfs_trans	**tpp,
1370 	int			*nospace_error)
1371 {
1372 	struct xfs_trans	*tp;
1373 	struct xfs_mount	*mp = ip->i_mount;
1374 	unsigned int		resblks;
1375 	bool			retried = false;
1376 	int			error;
1377 
1378 retry:
1379 	*nospace_error = 0;
1380 	resblks = *dblocks;
1381 	error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1382 	if (error == -ENOSPC) {
1383 		*nospace_error = error;
1384 		resblks = 0;
1385 		error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1386 	}
1387 	if (error)
1388 		return error;
1389 
1390 	xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
1391 
1392 	xfs_trans_ijoin(tp, dp, 0);
1393 	xfs_trans_ijoin(tp, ip, 0);
1394 
1395 	error = xfs_qm_dqattach_locked(dp, false);
1396 	if (error) {
1397 		/* Caller should have allocated the dquots! */
1398 		ASSERT(error != -ENOENT);
1399 		goto out_cancel;
1400 	}
1401 
1402 	error = xfs_qm_dqattach_locked(ip, false);
1403 	if (error) {
1404 		/* Caller should have allocated the dquots! */
1405 		ASSERT(error != -ENOENT);
1406 		goto out_cancel;
1407 	}
1408 
1409 	if (resblks == 0)
1410 		goto done;
1411 
1412 	error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);
1413 	if (error == -EDQUOT || error == -ENOSPC) {
1414 		if (!retried) {
1415 			xfs_trans_cancel(tp);
1416 			xfs_iunlock(dp, XFS_ILOCK_EXCL);
1417 			if (dp != ip)
1418 				xfs_iunlock(ip, XFS_ILOCK_EXCL);
1419 			xfs_blockgc_free_quota(dp, 0);
1420 			retried = true;
1421 			goto retry;
1422 		}
1423 
1424 		*nospace_error = error;
1425 		resblks = 0;
1426 		error = 0;
1427 	}
1428 	if (error)
1429 		goto out_cancel;
1430 
1431 done:
1432 	*tpp = tp;
1433 	*dblocks = resblks;
1434 	return 0;
1435 
1436 out_cancel:
1437 	xfs_trans_cancel(tp);
1438 	return error;
1439 }
1440