xref: /openbmc/linux/fs/xfs/xfs_trans.c (revision a09d2831)
1 /*
2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_da_btree.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_alloc.h"
42 #include "xfs_bmap.h"
43 #include "xfs_quota.h"
44 #include "xfs_trans_priv.h"
45 #include "xfs_trans_space.h"
46 #include "xfs_inode_item.h"
47 
48 
49 STATIC void	xfs_trans_apply_sb_deltas(xfs_trans_t *);
50 STATIC uint	xfs_trans_count_vecs(xfs_trans_t *);
51 STATIC void	xfs_trans_fill_vecs(xfs_trans_t *, xfs_log_iovec_t *);
52 STATIC void	xfs_trans_uncommit(xfs_trans_t *, uint);
53 STATIC void	xfs_trans_committed(xfs_trans_t *, int);
54 STATIC void	xfs_trans_chunk_committed(xfs_log_item_chunk_t *, xfs_lsn_t, int);
55 STATIC void	xfs_trans_free(xfs_trans_t *);
56 
57 kmem_zone_t	*xfs_trans_zone;
58 
59 
60 /*
61  * Reservation functions here avoid a huge stack in xfs_trans_init
62  * due to register overflow from temporaries in the calculations.
63  */
64 
65 STATIC uint
66 xfs_calc_write_reservation(xfs_mount_t *mp)
67 {
68 	return XFS_CALC_WRITE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
69 }
70 
71 STATIC uint
72 xfs_calc_itruncate_reservation(xfs_mount_t *mp)
73 {
74 	return XFS_CALC_ITRUNCATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
75 }
76 
77 STATIC uint
78 xfs_calc_rename_reservation(xfs_mount_t *mp)
79 {
80 	return XFS_CALC_RENAME_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
81 }
82 
83 STATIC uint
84 xfs_calc_link_reservation(xfs_mount_t *mp)
85 {
86 	return XFS_CALC_LINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
87 }
88 
89 STATIC uint
90 xfs_calc_remove_reservation(xfs_mount_t *mp)
91 {
92 	return XFS_CALC_REMOVE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
93 }
94 
95 STATIC uint
96 xfs_calc_symlink_reservation(xfs_mount_t *mp)
97 {
98 	return XFS_CALC_SYMLINK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
99 }
100 
101 STATIC uint
102 xfs_calc_create_reservation(xfs_mount_t *mp)
103 {
104 	return XFS_CALC_CREATE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
105 }
106 
107 STATIC uint
108 xfs_calc_mkdir_reservation(xfs_mount_t *mp)
109 {
110 	return XFS_CALC_MKDIR_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
111 }
112 
113 STATIC uint
114 xfs_calc_ifree_reservation(xfs_mount_t *mp)
115 {
116 	return XFS_CALC_IFREE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
117 }
118 
119 STATIC uint
120 xfs_calc_ichange_reservation(xfs_mount_t *mp)
121 {
122 	return XFS_CALC_ICHANGE_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
123 }
124 
125 STATIC uint
126 xfs_calc_growdata_reservation(xfs_mount_t *mp)
127 {
128 	return XFS_CALC_GROWDATA_LOG_RES(mp);
129 }
130 
131 STATIC uint
132 xfs_calc_growrtalloc_reservation(xfs_mount_t *mp)
133 {
134 	return XFS_CALC_GROWRTALLOC_LOG_RES(mp);
135 }
136 
137 STATIC uint
138 xfs_calc_growrtzero_reservation(xfs_mount_t *mp)
139 {
140 	return XFS_CALC_GROWRTZERO_LOG_RES(mp);
141 }
142 
143 STATIC uint
144 xfs_calc_growrtfree_reservation(xfs_mount_t *mp)
145 {
146 	return XFS_CALC_GROWRTFREE_LOG_RES(mp);
147 }
148 
149 STATIC uint
150 xfs_calc_swrite_reservation(xfs_mount_t *mp)
151 {
152 	return XFS_CALC_SWRITE_LOG_RES(mp);
153 }
154 
155 STATIC uint
156 xfs_calc_writeid_reservation(xfs_mount_t *mp)
157 {
158 	return XFS_CALC_WRITEID_LOG_RES(mp);
159 }
160 
161 STATIC uint
162 xfs_calc_addafork_reservation(xfs_mount_t *mp)
163 {
164 	return XFS_CALC_ADDAFORK_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
165 }
166 
167 STATIC uint
168 xfs_calc_attrinval_reservation(xfs_mount_t *mp)
169 {
170 	return XFS_CALC_ATTRINVAL_LOG_RES(mp);
171 }
172 
173 STATIC uint
174 xfs_calc_attrset_reservation(xfs_mount_t *mp)
175 {
176 	return XFS_CALC_ATTRSET_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
177 }
178 
179 STATIC uint
180 xfs_calc_attrrm_reservation(xfs_mount_t *mp)
181 {
182 	return XFS_CALC_ATTRRM_LOG_RES(mp) + XFS_DQUOT_LOGRES(mp);
183 }
184 
185 STATIC uint
186 xfs_calc_clear_agi_bucket_reservation(xfs_mount_t *mp)
187 {
188 	return XFS_CALC_CLEAR_AGI_BUCKET_LOG_RES(mp);
189 }
190 
191 /*
192  * Initialize the precomputed transaction reservation values
193  * in the mount structure.
194  */
195 void
196 xfs_trans_init(
197 	xfs_mount_t	*mp)
198 {
199 	xfs_trans_reservations_t	*resp;
200 
201 	resp = &(mp->m_reservations);
202 	resp->tr_write = xfs_calc_write_reservation(mp);
203 	resp->tr_itruncate = xfs_calc_itruncate_reservation(mp);
204 	resp->tr_rename = xfs_calc_rename_reservation(mp);
205 	resp->tr_link = xfs_calc_link_reservation(mp);
206 	resp->tr_remove = xfs_calc_remove_reservation(mp);
207 	resp->tr_symlink = xfs_calc_symlink_reservation(mp);
208 	resp->tr_create = xfs_calc_create_reservation(mp);
209 	resp->tr_mkdir = xfs_calc_mkdir_reservation(mp);
210 	resp->tr_ifree = xfs_calc_ifree_reservation(mp);
211 	resp->tr_ichange = xfs_calc_ichange_reservation(mp);
212 	resp->tr_growdata = xfs_calc_growdata_reservation(mp);
213 	resp->tr_swrite = xfs_calc_swrite_reservation(mp);
214 	resp->tr_writeid = xfs_calc_writeid_reservation(mp);
215 	resp->tr_addafork = xfs_calc_addafork_reservation(mp);
216 	resp->tr_attrinval = xfs_calc_attrinval_reservation(mp);
217 	resp->tr_attrset = xfs_calc_attrset_reservation(mp);
218 	resp->tr_attrrm = xfs_calc_attrrm_reservation(mp);
219 	resp->tr_clearagi = xfs_calc_clear_agi_bucket_reservation(mp);
220 	resp->tr_growrtalloc = xfs_calc_growrtalloc_reservation(mp);
221 	resp->tr_growrtzero = xfs_calc_growrtzero_reservation(mp);
222 	resp->tr_growrtfree = xfs_calc_growrtfree_reservation(mp);
223 }
224 
225 /*
226  * This routine is called to allocate a transaction structure.
227  * The type parameter indicates the type of the transaction.  These
228  * are enumerated in xfs_trans.h.
229  *
230  * Dynamically allocate the transaction structure from the transaction
231  * zone, initialize it, and return it to the caller.
232  */
233 xfs_trans_t *
234 xfs_trans_alloc(
235 	xfs_mount_t	*mp,
236 	uint		type)
237 {
238 	xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
239 	return _xfs_trans_alloc(mp, type, KM_SLEEP);
240 }
241 
242 xfs_trans_t *
243 _xfs_trans_alloc(
244 	xfs_mount_t	*mp,
245 	uint		type,
246 	uint		memflags)
247 {
248 	xfs_trans_t	*tp;
249 
250 	atomic_inc(&mp->m_active_trans);
251 
252 	tp = kmem_zone_zalloc(xfs_trans_zone, memflags);
253 	tp->t_magic = XFS_TRANS_MAGIC;
254 	tp->t_type = type;
255 	tp->t_mountp = mp;
256 	tp->t_items_free = XFS_LIC_NUM_SLOTS;
257 	tp->t_busy_free = XFS_LBC_NUM_SLOTS;
258 	xfs_lic_init(&(tp->t_items));
259 	XFS_LBC_INIT(&(tp->t_busy));
260 	return tp;
261 }
262 
263 /*
264  * This is called to create a new transaction which will share the
265  * permanent log reservation of the given transaction.  The remaining
266  * unused block and rt extent reservations are also inherited.  This
267  * implies that the original transaction is no longer allowed to allocate
268  * blocks.  Locks and log items, however, are no inherited.  They must
269  * be added to the new transaction explicitly.
270  */
271 xfs_trans_t *
272 xfs_trans_dup(
273 	xfs_trans_t	*tp)
274 {
275 	xfs_trans_t	*ntp;
276 
277 	ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
278 
279 	/*
280 	 * Initialize the new transaction structure.
281 	 */
282 	ntp->t_magic = XFS_TRANS_MAGIC;
283 	ntp->t_type = tp->t_type;
284 	ntp->t_mountp = tp->t_mountp;
285 	ntp->t_items_free = XFS_LIC_NUM_SLOTS;
286 	ntp->t_busy_free = XFS_LBC_NUM_SLOTS;
287 	xfs_lic_init(&(ntp->t_items));
288 	XFS_LBC_INIT(&(ntp->t_busy));
289 
290 	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
291 	ASSERT(tp->t_ticket != NULL);
292 
293 	ntp->t_flags = XFS_TRANS_PERM_LOG_RES | (tp->t_flags & XFS_TRANS_RESERVE);
294 	ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
295 	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
296 	tp->t_blk_res = tp->t_blk_res_used;
297 	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
298 	tp->t_rtx_res = tp->t_rtx_res_used;
299 	ntp->t_pflags = tp->t_pflags;
300 
301 	xfs_trans_dup_dqinfo(tp, ntp);
302 
303 	atomic_inc(&tp->t_mountp->m_active_trans);
304 	return ntp;
305 }
306 
307 /*
308  * This is called to reserve free disk blocks and log space for the
309  * given transaction.  This must be done before allocating any resources
310  * within the transaction.
311  *
312  * This will return ENOSPC if there are not enough blocks available.
313  * It will sleep waiting for available log space.
314  * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
315  * is used by long running transactions.  If any one of the reservations
316  * fails then they will all be backed out.
317  *
318  * This does not do quota reservations. That typically is done by the
319  * caller afterwards.
320  */
321 int
322 xfs_trans_reserve(
323 	xfs_trans_t	*tp,
324 	uint		blocks,
325 	uint		logspace,
326 	uint		rtextents,
327 	uint		flags,
328 	uint		logcount)
329 {
330 	int		log_flags;
331 	int		error = 0;
332 	int		rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
333 
334 	/* Mark this thread as being in a transaction */
335 	current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
336 
337 	/*
338 	 * Attempt to reserve the needed disk blocks by decrementing
339 	 * the number needed from the number available.  This will
340 	 * fail if the count would go below zero.
341 	 */
342 	if (blocks > 0) {
343 		error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
344 					  -((int64_t)blocks), rsvd);
345 		if (error != 0) {
346 			current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
347 			return (XFS_ERROR(ENOSPC));
348 		}
349 		tp->t_blk_res += blocks;
350 	}
351 
352 	/*
353 	 * Reserve the log space needed for this transaction.
354 	 */
355 	if (logspace > 0) {
356 		ASSERT((tp->t_log_res == 0) || (tp->t_log_res == logspace));
357 		ASSERT((tp->t_log_count == 0) ||
358 			(tp->t_log_count == logcount));
359 		if (flags & XFS_TRANS_PERM_LOG_RES) {
360 			log_flags = XFS_LOG_PERM_RESERV;
361 			tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
362 		} else {
363 			ASSERT(tp->t_ticket == NULL);
364 			ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
365 			log_flags = 0;
366 		}
367 
368 		error = xfs_log_reserve(tp->t_mountp, logspace, logcount,
369 					&tp->t_ticket,
370 					XFS_TRANSACTION, log_flags, tp->t_type);
371 		if (error) {
372 			goto undo_blocks;
373 		}
374 		tp->t_log_res = logspace;
375 		tp->t_log_count = logcount;
376 	}
377 
378 	/*
379 	 * Attempt to reserve the needed realtime extents by decrementing
380 	 * the number needed from the number available.  This will
381 	 * fail if the count would go below zero.
382 	 */
383 	if (rtextents > 0) {
384 		error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FREXTENTS,
385 					  -((int64_t)rtextents), rsvd);
386 		if (error) {
387 			error = XFS_ERROR(ENOSPC);
388 			goto undo_log;
389 		}
390 		tp->t_rtx_res += rtextents;
391 	}
392 
393 	return 0;
394 
395 	/*
396 	 * Error cases jump to one of these labels to undo any
397 	 * reservations which have already been performed.
398 	 */
399 undo_log:
400 	if (logspace > 0) {
401 		if (flags & XFS_TRANS_PERM_LOG_RES) {
402 			log_flags = XFS_LOG_REL_PERM_RESERV;
403 		} else {
404 			log_flags = 0;
405 		}
406 		xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, log_flags);
407 		tp->t_ticket = NULL;
408 		tp->t_log_res = 0;
409 		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
410 	}
411 
412 undo_blocks:
413 	if (blocks > 0) {
414 		(void) xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
415 					 (int64_t)blocks, rsvd);
416 		tp->t_blk_res = 0;
417 	}
418 
419 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
420 
421 	return error;
422 }
423 
424 
425 /*
426  * Record the indicated change to the given field for application
427  * to the file system's superblock when the transaction commits.
428  * For now, just store the change in the transaction structure.
429  *
430  * Mark the transaction structure to indicate that the superblock
431  * needs to be updated before committing.
432  *
433  * Because we may not be keeping track of allocated/free inodes and
434  * used filesystem blocks in the superblock, we do not mark the
435  * superblock dirty in this transaction if we modify these fields.
436  * We still need to update the transaction deltas so that they get
437  * applied to the incore superblock, but we don't want them to
438  * cause the superblock to get locked and logged if these are the
439  * only fields in the superblock that the transaction modifies.
440  */
441 void
442 xfs_trans_mod_sb(
443 	xfs_trans_t	*tp,
444 	uint		field,
445 	int64_t		delta)
446 {
447 	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
448 	xfs_mount_t	*mp = tp->t_mountp;
449 
450 	switch (field) {
451 	case XFS_TRANS_SB_ICOUNT:
452 		tp->t_icount_delta += delta;
453 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
454 			flags &= ~XFS_TRANS_SB_DIRTY;
455 		break;
456 	case XFS_TRANS_SB_IFREE:
457 		tp->t_ifree_delta += delta;
458 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
459 			flags &= ~XFS_TRANS_SB_DIRTY;
460 		break;
461 	case XFS_TRANS_SB_FDBLOCKS:
462 		/*
463 		 * Track the number of blocks allocated in the
464 		 * transaction.  Make sure it does not exceed the
465 		 * number reserved.
466 		 */
467 		if (delta < 0) {
468 			tp->t_blk_res_used += (uint)-delta;
469 			ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
470 		}
471 		tp->t_fdblocks_delta += delta;
472 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
473 			flags &= ~XFS_TRANS_SB_DIRTY;
474 		break;
475 	case XFS_TRANS_SB_RES_FDBLOCKS:
476 		/*
477 		 * The allocation has already been applied to the
478 		 * in-core superblock's counter.  This should only
479 		 * be applied to the on-disk superblock.
480 		 */
481 		ASSERT(delta < 0);
482 		tp->t_res_fdblocks_delta += delta;
483 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
484 			flags &= ~XFS_TRANS_SB_DIRTY;
485 		break;
486 	case XFS_TRANS_SB_FREXTENTS:
487 		/*
488 		 * Track the number of blocks allocated in the
489 		 * transaction.  Make sure it does not exceed the
490 		 * number reserved.
491 		 */
492 		if (delta < 0) {
493 			tp->t_rtx_res_used += (uint)-delta;
494 			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
495 		}
496 		tp->t_frextents_delta += delta;
497 		break;
498 	case XFS_TRANS_SB_RES_FREXTENTS:
499 		/*
500 		 * The allocation has already been applied to the
501 		 * in-core superblock's counter.  This should only
502 		 * be applied to the on-disk superblock.
503 		 */
504 		ASSERT(delta < 0);
505 		tp->t_res_frextents_delta += delta;
506 		break;
507 	case XFS_TRANS_SB_DBLOCKS:
508 		ASSERT(delta > 0);
509 		tp->t_dblocks_delta += delta;
510 		break;
511 	case XFS_TRANS_SB_AGCOUNT:
512 		ASSERT(delta > 0);
513 		tp->t_agcount_delta += delta;
514 		break;
515 	case XFS_TRANS_SB_IMAXPCT:
516 		tp->t_imaxpct_delta += delta;
517 		break;
518 	case XFS_TRANS_SB_REXTSIZE:
519 		tp->t_rextsize_delta += delta;
520 		break;
521 	case XFS_TRANS_SB_RBMBLOCKS:
522 		tp->t_rbmblocks_delta += delta;
523 		break;
524 	case XFS_TRANS_SB_RBLOCKS:
525 		tp->t_rblocks_delta += delta;
526 		break;
527 	case XFS_TRANS_SB_REXTENTS:
528 		tp->t_rextents_delta += delta;
529 		break;
530 	case XFS_TRANS_SB_REXTSLOG:
531 		tp->t_rextslog_delta += delta;
532 		break;
533 	default:
534 		ASSERT(0);
535 		return;
536 	}
537 
538 	tp->t_flags |= flags;
539 }
540 
541 /*
542  * xfs_trans_apply_sb_deltas() is called from the commit code
543  * to bring the superblock buffer into the current transaction
544  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
545  *
546  * For now we just look at each field allowed to change and change
547  * it if necessary.
548  */
549 STATIC void
550 xfs_trans_apply_sb_deltas(
551 	xfs_trans_t	*tp)
552 {
553 	xfs_dsb_t	*sbp;
554 	xfs_buf_t	*bp;
555 	int		whole = 0;
556 
557 	bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
558 	sbp = XFS_BUF_TO_SBP(bp);
559 
560 	/*
561 	 * Check that superblock mods match the mods made to AGF counters.
562 	 */
563 	ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
564 	       (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
565 		tp->t_ag_btree_delta));
566 
567 	/*
568 	 * Only update the superblock counters if we are logging them
569 	 */
570 	if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
571 		if (tp->t_icount_delta)
572 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
573 		if (tp->t_ifree_delta)
574 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
575 		if (tp->t_fdblocks_delta)
576 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
577 		if (tp->t_res_fdblocks_delta)
578 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
579 	}
580 
581 	if (tp->t_frextents_delta)
582 		be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
583 	if (tp->t_res_frextents_delta)
584 		be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
585 
586 	if (tp->t_dblocks_delta) {
587 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
588 		whole = 1;
589 	}
590 	if (tp->t_agcount_delta) {
591 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
592 		whole = 1;
593 	}
594 	if (tp->t_imaxpct_delta) {
595 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
596 		whole = 1;
597 	}
598 	if (tp->t_rextsize_delta) {
599 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
600 		whole = 1;
601 	}
602 	if (tp->t_rbmblocks_delta) {
603 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
604 		whole = 1;
605 	}
606 	if (tp->t_rblocks_delta) {
607 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
608 		whole = 1;
609 	}
610 	if (tp->t_rextents_delta) {
611 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
612 		whole = 1;
613 	}
614 	if (tp->t_rextslog_delta) {
615 		sbp->sb_rextslog += tp->t_rextslog_delta;
616 		whole = 1;
617 	}
618 
619 	if (whole)
620 		/*
621 		 * Log the whole thing, the fields are noncontiguous.
622 		 */
623 		xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
624 	else
625 		/*
626 		 * Since all the modifiable fields are contiguous, we
627 		 * can get away with this.
628 		 */
629 		xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
630 				  offsetof(xfs_dsb_t, sb_frextents) +
631 				  sizeof(sbp->sb_frextents) - 1);
632 }
633 
634 /*
635  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
636  * and apply superblock counter changes to the in-core superblock.  The
637  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
638  * applied to the in-core superblock.  The idea is that that has already been
639  * done.
640  *
641  * This is done efficiently with a single call to xfs_mod_incore_sb_batch().
642  * However, we have to ensure that we only modify each superblock field only
643  * once because the application of the delta values may not be atomic. That can
644  * lead to ENOSPC races occurring if we have two separate modifcations of the
645  * free space counter to put back the entire reservation and then take away
646  * what we used.
647  *
648  * If we are not logging superblock counters, then the inode allocated/free and
649  * used block counts are not updated in the on disk superblock. In this case,
650  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
651  * still need to update the incore superblock with the changes.
652  */
653 STATIC void
654 xfs_trans_unreserve_and_mod_sb(
655 	xfs_trans_t	*tp)
656 {
657 	xfs_mod_sb_t	msb[14];	/* If you add cases, add entries */
658 	xfs_mod_sb_t	*msbp;
659 	xfs_mount_t	*mp = tp->t_mountp;
660 	/* REFERENCED */
661 	int		error;
662 	int		rsvd;
663 	int64_t		blkdelta = 0;
664 	int64_t		rtxdelta = 0;
665 
666 	msbp = msb;
667 	rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
668 
669 	/* calculate free blocks delta */
670 	if (tp->t_blk_res > 0)
671 		blkdelta = tp->t_blk_res;
672 
673 	if ((tp->t_fdblocks_delta != 0) &&
674 	    (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
675 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)))
676 	        blkdelta += tp->t_fdblocks_delta;
677 
678 	if (blkdelta != 0) {
679 		msbp->msb_field = XFS_SBS_FDBLOCKS;
680 		msbp->msb_delta = blkdelta;
681 		msbp++;
682 	}
683 
684 	/* calculate free realtime extents delta */
685 	if (tp->t_rtx_res > 0)
686 		rtxdelta = tp->t_rtx_res;
687 
688 	if ((tp->t_frextents_delta != 0) &&
689 	    (tp->t_flags & XFS_TRANS_SB_DIRTY))
690 		rtxdelta += tp->t_frextents_delta;
691 
692 	if (rtxdelta != 0) {
693 		msbp->msb_field = XFS_SBS_FREXTENTS;
694 		msbp->msb_delta = rtxdelta;
695 		msbp++;
696 	}
697 
698 	/* apply remaining deltas */
699 
700 	if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
701 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
702 		if (tp->t_icount_delta != 0) {
703 			msbp->msb_field = XFS_SBS_ICOUNT;
704 			msbp->msb_delta = tp->t_icount_delta;
705 			msbp++;
706 		}
707 		if (tp->t_ifree_delta != 0) {
708 			msbp->msb_field = XFS_SBS_IFREE;
709 			msbp->msb_delta = tp->t_ifree_delta;
710 			msbp++;
711 		}
712 	}
713 
714 	if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
715 		if (tp->t_dblocks_delta != 0) {
716 			msbp->msb_field = XFS_SBS_DBLOCKS;
717 			msbp->msb_delta = tp->t_dblocks_delta;
718 			msbp++;
719 		}
720 		if (tp->t_agcount_delta != 0) {
721 			msbp->msb_field = XFS_SBS_AGCOUNT;
722 			msbp->msb_delta = tp->t_agcount_delta;
723 			msbp++;
724 		}
725 		if (tp->t_imaxpct_delta != 0) {
726 			msbp->msb_field = XFS_SBS_IMAX_PCT;
727 			msbp->msb_delta = tp->t_imaxpct_delta;
728 			msbp++;
729 		}
730 		if (tp->t_rextsize_delta != 0) {
731 			msbp->msb_field = XFS_SBS_REXTSIZE;
732 			msbp->msb_delta = tp->t_rextsize_delta;
733 			msbp++;
734 		}
735 		if (tp->t_rbmblocks_delta != 0) {
736 			msbp->msb_field = XFS_SBS_RBMBLOCKS;
737 			msbp->msb_delta = tp->t_rbmblocks_delta;
738 			msbp++;
739 		}
740 		if (tp->t_rblocks_delta != 0) {
741 			msbp->msb_field = XFS_SBS_RBLOCKS;
742 			msbp->msb_delta = tp->t_rblocks_delta;
743 			msbp++;
744 		}
745 		if (tp->t_rextents_delta != 0) {
746 			msbp->msb_field = XFS_SBS_REXTENTS;
747 			msbp->msb_delta = tp->t_rextents_delta;
748 			msbp++;
749 		}
750 		if (tp->t_rextslog_delta != 0) {
751 			msbp->msb_field = XFS_SBS_REXTSLOG;
752 			msbp->msb_delta = tp->t_rextslog_delta;
753 			msbp++;
754 		}
755 	}
756 
757 	/*
758 	 * If we need to change anything, do it.
759 	 */
760 	if (msbp > msb) {
761 		error = xfs_mod_incore_sb_batch(tp->t_mountp, msb,
762 			(uint)(msbp - msb), rsvd);
763 		ASSERT(error == 0);
764 	}
765 }
766 
767 
768 /*
769  * xfs_trans_commit
770  *
771  * Commit the given transaction to the log a/synchronously.
772  *
773  * XFS disk error handling mechanism is not based on a typical
774  * transaction abort mechanism. Logically after the filesystem
775  * gets marked 'SHUTDOWN', we can't let any new transactions
776  * be durable - ie. committed to disk - because some metadata might
777  * be inconsistent. In such cases, this returns an error, and the
778  * caller may assume that all locked objects joined to the transaction
779  * have already been unlocked as if the commit had succeeded.
780  * Do not reference the transaction structure after this call.
781  */
782  /*ARGSUSED*/
783 int
784 _xfs_trans_commit(
785 	xfs_trans_t	*tp,
786 	uint		flags,
787 	int		*log_flushed)
788 {
789 	xfs_log_iovec_t		*log_vector;
790 	int			nvec;
791 	xfs_mount_t		*mp;
792 	xfs_lsn_t		commit_lsn;
793 	/* REFERENCED */
794 	int			error;
795 	int			log_flags;
796 	int			sync;
797 #define	XFS_TRANS_LOGVEC_COUNT	16
798 	xfs_log_iovec_t		log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
799 	void			*commit_iclog;
800 	int			shutdown;
801 
802 	commit_lsn = -1;
803 
804 	/*
805 	 * Determine whether this commit is releasing a permanent
806 	 * log reservation or not.
807 	 */
808 	if (flags & XFS_TRANS_RELEASE_LOG_RES) {
809 		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
810 		log_flags = XFS_LOG_REL_PERM_RESERV;
811 	} else {
812 		log_flags = 0;
813 	}
814 	mp = tp->t_mountp;
815 
816 	/*
817 	 * If there is nothing to be logged by the transaction,
818 	 * then unlock all of the items associated with the
819 	 * transaction and free the transaction structure.
820 	 * Also make sure to return any reserved blocks to
821 	 * the free pool.
822 	 */
823 shut_us_down:
824 	shutdown = XFS_FORCED_SHUTDOWN(mp) ? EIO : 0;
825 	if (!(tp->t_flags & XFS_TRANS_DIRTY) || shutdown) {
826 		xfs_trans_unreserve_and_mod_sb(tp);
827 		/*
828 		 * It is indeed possible for the transaction to be
829 		 * not dirty but the dqinfo portion to be. All that
830 		 * means is that we have some (non-persistent) quota
831 		 * reservations that need to be unreserved.
832 		 */
833 		xfs_trans_unreserve_and_mod_dquots(tp);
834 		if (tp->t_ticket) {
835 			commit_lsn = xfs_log_done(mp, tp->t_ticket,
836 							NULL, log_flags);
837 			if (commit_lsn == -1 && !shutdown)
838 				shutdown = XFS_ERROR(EIO);
839 		}
840 		current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
841 		xfs_trans_free_items(tp, shutdown? XFS_TRANS_ABORT : 0);
842 		xfs_trans_free_busy(tp);
843 		xfs_trans_free(tp);
844 		XFS_STATS_INC(xs_trans_empty);
845 		return (shutdown);
846 	}
847 	ASSERT(tp->t_ticket != NULL);
848 
849 	/*
850 	 * If we need to update the superblock, then do it now.
851 	 */
852 	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
853 		xfs_trans_apply_sb_deltas(tp);
854 	xfs_trans_apply_dquot_deltas(tp);
855 
856 	/*
857 	 * Ask each log item how many log_vector entries it will
858 	 * need so we can figure out how many to allocate.
859 	 * Try to avoid the kmem_alloc() call in the common case
860 	 * by using a vector from the stack when it fits.
861 	 */
862 	nvec = xfs_trans_count_vecs(tp);
863 	if (nvec == 0) {
864 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
865 		goto shut_us_down;
866 	} else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
867 		log_vector = log_vector_fast;
868 	} else {
869 		log_vector = (xfs_log_iovec_t *)kmem_alloc(nvec *
870 						   sizeof(xfs_log_iovec_t),
871 						   KM_SLEEP);
872 	}
873 
874 	/*
875 	 * Fill in the log_vector and pin the logged items, and
876 	 * then write the transaction to the log.
877 	 */
878 	xfs_trans_fill_vecs(tp, log_vector);
879 
880 	error = xfs_log_write(mp, log_vector, nvec, tp->t_ticket, &(tp->t_lsn));
881 
882 	/*
883 	 * The transaction is committed incore here, and can go out to disk
884 	 * at any time after this call.  However, all the items associated
885 	 * with the transaction are still locked and pinned in memory.
886 	 */
887 	commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
888 
889 	tp->t_commit_lsn = commit_lsn;
890 	if (nvec > XFS_TRANS_LOGVEC_COUNT) {
891 		kmem_free(log_vector);
892 	}
893 
894 	/*
895 	 * If we got a log write error. Unpin the logitems that we
896 	 * had pinned, clean up, free trans structure, and return error.
897 	 */
898 	if (error || commit_lsn == -1) {
899 		current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
900 		xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
901 		return XFS_ERROR(EIO);
902 	}
903 
904 	/*
905 	 * Once the transaction has committed, unused
906 	 * reservations need to be released and changes to
907 	 * the superblock need to be reflected in the in-core
908 	 * version.  Do that now.
909 	 */
910 	xfs_trans_unreserve_and_mod_sb(tp);
911 
912 	sync = tp->t_flags & XFS_TRANS_SYNC;
913 
914 	/*
915 	 * Tell the LM to call the transaction completion routine
916 	 * when the log write with LSN commit_lsn completes (e.g.
917 	 * when the transaction commit really hits the on-disk log).
918 	 * After this call we cannot reference tp, because the call
919 	 * can happen at any time and the call will free the transaction
920 	 * structure pointed to by tp.  The only case where we call
921 	 * the completion routine (xfs_trans_committed) directly is
922 	 * if the log is turned off on a debug kernel or we're
923 	 * running in simulation mode (the log is explicitly turned
924 	 * off).
925 	 */
926 	tp->t_logcb.cb_func = (void(*)(void*, int))xfs_trans_committed;
927 	tp->t_logcb.cb_arg = tp;
928 
929 	/*
930 	 * We need to pass the iclog buffer which was used for the
931 	 * transaction commit record into this function, and attach
932 	 * the callback to it. The callback must be attached before
933 	 * the items are unlocked to avoid racing with other threads
934 	 * waiting for an item to unlock.
935 	 */
936 	shutdown = xfs_log_notify(mp, commit_iclog, &(tp->t_logcb));
937 
938 	/*
939 	 * Mark this thread as no longer being in a transaction
940 	 */
941 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
942 
943 	/*
944 	 * Once all the items of the transaction have been copied
945 	 * to the in core log and the callback is attached, the
946 	 * items can be unlocked.
947 	 *
948 	 * This will free descriptors pointing to items which were
949 	 * not logged since there is nothing more to do with them.
950 	 * For items which were logged, we will keep pointers to them
951 	 * so they can be unpinned after the transaction commits to disk.
952 	 * This will also stamp each modified meta-data item with
953 	 * the commit lsn of this transaction for dependency tracking
954 	 * purposes.
955 	 */
956 	xfs_trans_unlock_items(tp, commit_lsn);
957 
958 	/*
959 	 * If we detected a log error earlier, finish committing
960 	 * the transaction now (unpin log items, etc).
961 	 *
962 	 * Order is critical here, to avoid using the transaction
963 	 * pointer after its been freed (by xfs_trans_committed
964 	 * either here now, or as a callback).  We cannot do this
965 	 * step inside xfs_log_notify as was done earlier because
966 	 * of this issue.
967 	 */
968 	if (shutdown)
969 		xfs_trans_committed(tp, XFS_LI_ABORTED);
970 
971 	/*
972 	 * Now that the xfs_trans_committed callback has been attached,
973 	 * and the items are released we can finally allow the iclog to
974 	 * go to disk.
975 	 */
976 	error = xfs_log_release_iclog(mp, commit_iclog);
977 
978 	/*
979 	 * If the transaction needs to be synchronous, then force the
980 	 * log out now and wait for it.
981 	 */
982 	if (sync) {
983 		if (!error) {
984 			error = _xfs_log_force(mp, commit_lsn,
985 				      XFS_LOG_FORCE | XFS_LOG_SYNC,
986 				      log_flushed);
987 		}
988 		XFS_STATS_INC(xs_trans_sync);
989 	} else {
990 		XFS_STATS_INC(xs_trans_async);
991 	}
992 
993 	return (error);
994 }
995 
996 
997 /*
998  * Total up the number of log iovecs needed to commit this
999  * transaction.  The transaction itself needs one for the
1000  * transaction header.  Ask each dirty item in turn how many
1001  * it needs to get the total.
1002  */
1003 STATIC uint
1004 xfs_trans_count_vecs(
1005 	xfs_trans_t	*tp)
1006 {
1007 	int			nvecs;
1008 	xfs_log_item_desc_t	*lidp;
1009 
1010 	nvecs = 1;
1011 	lidp = xfs_trans_first_item(tp);
1012 	ASSERT(lidp != NULL);
1013 
1014 	/* In the non-debug case we need to start bailing out if we
1015 	 * didn't find a log_item here, return zero and let trans_commit
1016 	 * deal with it.
1017 	 */
1018 	if (lidp == NULL)
1019 		return 0;
1020 
1021 	while (lidp != NULL) {
1022 		/*
1023 		 * Skip items which aren't dirty in this transaction.
1024 		 */
1025 		if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
1026 			lidp = xfs_trans_next_item(tp, lidp);
1027 			continue;
1028 		}
1029 		lidp->lid_size = IOP_SIZE(lidp->lid_item);
1030 		nvecs += lidp->lid_size;
1031 		lidp = xfs_trans_next_item(tp, lidp);
1032 	}
1033 
1034 	return nvecs;
1035 }
1036 
1037 /*
1038  * Called from the trans_commit code when we notice that
1039  * the filesystem is in the middle of a forced shutdown.
1040  */
1041 STATIC void
1042 xfs_trans_uncommit(
1043 	xfs_trans_t	*tp,
1044 	uint		flags)
1045 {
1046 	xfs_log_item_desc_t	*lidp;
1047 
1048 	for (lidp = xfs_trans_first_item(tp);
1049 	     lidp != NULL;
1050 	     lidp = xfs_trans_next_item(tp, lidp)) {
1051 		/*
1052 		 * Unpin all but those that aren't dirty.
1053 		 */
1054 		if (lidp->lid_flags & XFS_LID_DIRTY)
1055 			IOP_UNPIN_REMOVE(lidp->lid_item, tp);
1056 	}
1057 
1058 	xfs_trans_unreserve_and_mod_sb(tp);
1059 	xfs_trans_unreserve_and_mod_dquots(tp);
1060 
1061 	xfs_trans_free_items(tp, flags);
1062 	xfs_trans_free_busy(tp);
1063 	xfs_trans_free(tp);
1064 }
1065 
1066 /*
1067  * Fill in the vector with pointers to data to be logged
1068  * by this transaction.  The transaction header takes
1069  * the first vector, and then each dirty item takes the
1070  * number of vectors it indicated it needed in xfs_trans_count_vecs().
1071  *
1072  * As each item fills in the entries it needs, also pin the item
1073  * so that it cannot be flushed out until the log write completes.
1074  */
1075 STATIC void
1076 xfs_trans_fill_vecs(
1077 	xfs_trans_t		*tp,
1078 	xfs_log_iovec_t		*log_vector)
1079 {
1080 	xfs_log_item_desc_t	*lidp;
1081 	xfs_log_iovec_t		*vecp;
1082 	uint			nitems;
1083 
1084 	/*
1085 	 * Skip over the entry for the transaction header, we'll
1086 	 * fill that in at the end.
1087 	 */
1088 	vecp = log_vector + 1;		/* pointer arithmetic */
1089 
1090 	nitems = 0;
1091 	lidp = xfs_trans_first_item(tp);
1092 	ASSERT(lidp != NULL);
1093 	while (lidp != NULL) {
1094 		/*
1095 		 * Skip items which aren't dirty in this transaction.
1096 		 */
1097 		if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
1098 			lidp = xfs_trans_next_item(tp, lidp);
1099 			continue;
1100 		}
1101 		/*
1102 		 * The item may be marked dirty but not log anything.
1103 		 * This can be used to get called when a transaction
1104 		 * is committed.
1105 		 */
1106 		if (lidp->lid_size) {
1107 			nitems++;
1108 		}
1109 		IOP_FORMAT(lidp->lid_item, vecp);
1110 		vecp += lidp->lid_size;		/* pointer arithmetic */
1111 		IOP_PIN(lidp->lid_item);
1112 		lidp = xfs_trans_next_item(tp, lidp);
1113 	}
1114 
1115 	/*
1116 	 * Now that we've counted the number of items in this
1117 	 * transaction, fill in the transaction header.
1118 	 */
1119 	tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
1120 	tp->t_header.th_type = tp->t_type;
1121 	tp->t_header.th_num_items = nitems;
1122 	log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
1123 	log_vector->i_len = sizeof(xfs_trans_header_t);
1124 	XLOG_VEC_SET_TYPE(log_vector, XLOG_REG_TYPE_TRANSHDR);
1125 }
1126 
1127 
1128 /*
1129  * Unlock all of the transaction's items and free the transaction.
1130  * The transaction must not have modified any of its items, because
1131  * there is no way to restore them to their previous state.
1132  *
1133  * If the transaction has made a log reservation, make sure to release
1134  * it as well.
1135  */
1136 void
1137 xfs_trans_cancel(
1138 	xfs_trans_t		*tp,
1139 	int			flags)
1140 {
1141 	int			log_flags;
1142 #ifdef DEBUG
1143 	xfs_log_item_chunk_t	*licp;
1144 	xfs_log_item_desc_t	*lidp;
1145 	xfs_log_item_t		*lip;
1146 	int			i;
1147 #endif
1148 	xfs_mount_t		*mp = tp->t_mountp;
1149 
1150 	/*
1151 	 * See if the caller is being too lazy to figure out if
1152 	 * the transaction really needs an abort.
1153 	 */
1154 	if ((flags & XFS_TRANS_ABORT) && !(tp->t_flags & XFS_TRANS_DIRTY))
1155 		flags &= ~XFS_TRANS_ABORT;
1156 	/*
1157 	 * See if the caller is relying on us to shut down the
1158 	 * filesystem.  This happens in paths where we detect
1159 	 * corruption and decide to give up.
1160 	 */
1161 	if ((tp->t_flags & XFS_TRANS_DIRTY) && !XFS_FORCED_SHUTDOWN(mp)) {
1162 		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1163 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1164 	}
1165 #ifdef DEBUG
1166 	if (!(flags & XFS_TRANS_ABORT)) {
1167 		licp = &(tp->t_items);
1168 		while (licp != NULL) {
1169 			lidp = licp->lic_descs;
1170 			for (i = 0; i < licp->lic_unused; i++, lidp++) {
1171 				if (xfs_lic_isfree(licp, i)) {
1172 					continue;
1173 				}
1174 
1175 				lip = lidp->lid_item;
1176 				if (!XFS_FORCED_SHUTDOWN(mp))
1177 					ASSERT(!(lip->li_type == XFS_LI_EFD));
1178 			}
1179 			licp = licp->lic_next;
1180 		}
1181 	}
1182 #endif
1183 	xfs_trans_unreserve_and_mod_sb(tp);
1184 	xfs_trans_unreserve_and_mod_dquots(tp);
1185 
1186 	if (tp->t_ticket) {
1187 		if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1188 			ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1189 			log_flags = XFS_LOG_REL_PERM_RESERV;
1190 		} else {
1191 			log_flags = 0;
1192 		}
1193 		xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1194 	}
1195 
1196 	/* mark this thread as no longer being in a transaction */
1197 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1198 
1199 	xfs_trans_free_items(tp, flags);
1200 	xfs_trans_free_busy(tp);
1201 	xfs_trans_free(tp);
1202 }
1203 
1204 
1205 /*
1206  * Free the transaction structure.  If there is more clean up
1207  * to do when the structure is freed, add it here.
1208  */
1209 STATIC void
1210 xfs_trans_free(
1211 	xfs_trans_t	*tp)
1212 {
1213 	atomic_dec(&tp->t_mountp->m_active_trans);
1214 	xfs_trans_free_dqinfo(tp);
1215 	kmem_zone_free(xfs_trans_zone, tp);
1216 }
1217 
1218 /*
1219  * Roll from one trans in the sequence of PERMANENT transactions to
1220  * the next: permanent transactions are only flushed out when
1221  * committed with XFS_TRANS_RELEASE_LOG_RES, but we still want as soon
1222  * as possible to let chunks of it go to the log. So we commit the
1223  * chunk we've been working on and get a new transaction to continue.
1224  */
1225 int
1226 xfs_trans_roll(
1227 	struct xfs_trans	**tpp,
1228 	struct xfs_inode	*dp)
1229 {
1230 	struct xfs_trans	*trans;
1231 	unsigned int		logres, count;
1232 	int			error;
1233 
1234 	/*
1235 	 * Ensure that the inode is always logged.
1236 	 */
1237 	trans = *tpp;
1238 	xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
1239 
1240 	/*
1241 	 * Copy the critical parameters from one trans to the next.
1242 	 */
1243 	logres = trans->t_log_res;
1244 	count = trans->t_log_count;
1245 	*tpp = xfs_trans_dup(trans);
1246 
1247 	/*
1248 	 * Commit the current transaction.
1249 	 * If this commit failed, then it'd just unlock those items that
1250 	 * are not marked ihold. That also means that a filesystem shutdown
1251 	 * is in progress. The caller takes the responsibility to cancel
1252 	 * the duplicate transaction that gets returned.
1253 	 */
1254 	error = xfs_trans_commit(trans, 0);
1255 	if (error)
1256 		return (error);
1257 
1258 	trans = *tpp;
1259 
1260 	/*
1261 	 * transaction commit worked ok so we can drop the extra ticket
1262 	 * reference that we gained in xfs_trans_dup()
1263 	 */
1264 	xfs_log_ticket_put(trans->t_ticket);
1265 
1266 
1267 	/*
1268 	 * Reserve space in the log for th next transaction.
1269 	 * This also pushes items in the "AIL", the list of logged items,
1270 	 * out to disk if they are taking up space at the tail of the log
1271 	 * that we want to use.  This requires that either nothing be locked
1272 	 * across this call, or that anything that is locked be logged in
1273 	 * the prior and the next transactions.
1274 	 */
1275 	error = xfs_trans_reserve(trans, 0, logres, 0,
1276 				  XFS_TRANS_PERM_LOG_RES, count);
1277 	/*
1278 	 *  Ensure that the inode is in the new transaction and locked.
1279 	 */
1280 	if (error)
1281 		return error;
1282 
1283 	xfs_trans_ijoin(trans, dp, XFS_ILOCK_EXCL);
1284 	xfs_trans_ihold(trans, dp);
1285 	return 0;
1286 }
1287 
1288 /*
1289  * THIS SHOULD BE REWRITTEN TO USE xfs_trans_next_item().
1290  *
1291  * This is typically called by the LM when a transaction has been fully
1292  * committed to disk.  It needs to unpin the items which have
1293  * been logged by the transaction and update their positions
1294  * in the AIL if necessary.
1295  * This also gets called when the transactions didn't get written out
1296  * because of an I/O error. Abortflag & XFS_LI_ABORTED is set then.
1297  *
1298  * Call xfs_trans_chunk_committed() to process the items in
1299  * each chunk.
1300  */
1301 STATIC void
1302 xfs_trans_committed(
1303 	xfs_trans_t	*tp,
1304 	int		abortflag)
1305 {
1306 	xfs_log_item_chunk_t	*licp;
1307 	xfs_log_item_chunk_t	*next_licp;
1308 	xfs_log_busy_chunk_t	*lbcp;
1309 	xfs_log_busy_slot_t	*lbsp;
1310 	int			i;
1311 
1312 	/*
1313 	 * Call the transaction's completion callback if there
1314 	 * is one.
1315 	 */
1316 	if (tp->t_callback != NULL) {
1317 		tp->t_callback(tp, tp->t_callarg);
1318 	}
1319 
1320 	/*
1321 	 * Special case the chunk embedded in the transaction.
1322 	 */
1323 	licp = &(tp->t_items);
1324 	if (!(xfs_lic_are_all_free(licp))) {
1325 		xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
1326 	}
1327 
1328 	/*
1329 	 * Process the items in each chunk in turn.
1330 	 */
1331 	licp = licp->lic_next;
1332 	while (licp != NULL) {
1333 		ASSERT(!xfs_lic_are_all_free(licp));
1334 		xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
1335 		next_licp = licp->lic_next;
1336 		kmem_free(licp);
1337 		licp = next_licp;
1338 	}
1339 
1340 	/*
1341 	 * Clear all the per-AG busy list items listed in this transaction
1342 	 */
1343 	lbcp = &tp->t_busy;
1344 	while (lbcp != NULL) {
1345 		for (i = 0, lbsp = lbcp->lbc_busy; i < lbcp->lbc_unused; i++, lbsp++) {
1346 			if (!XFS_LBC_ISFREE(lbcp, i)) {
1347 				xfs_alloc_clear_busy(tp, lbsp->lbc_ag,
1348 						     lbsp->lbc_idx);
1349 			}
1350 		}
1351 		lbcp = lbcp->lbc_next;
1352 	}
1353 	xfs_trans_free_busy(tp);
1354 
1355 	/*
1356 	 * That's it for the transaction structure.  Free it.
1357 	 */
1358 	xfs_trans_free(tp);
1359 }
1360 
1361 /*
1362  * This is called to perform the commit processing for each
1363  * item described by the given chunk.
1364  *
1365  * The commit processing consists of unlocking items which were
1366  * held locked with the SYNC_UNLOCK attribute, calling the committed
1367  * routine of each logged item, updating the item's position in the AIL
1368  * if necessary, and unpinning each item.  If the committed routine
1369  * returns -1, then do nothing further with the item because it
1370  * may have been freed.
1371  *
1372  * Since items are unlocked when they are copied to the incore
1373  * log, it is possible for two transactions to be completing
1374  * and manipulating the same item simultaneously.  The AIL lock
1375  * will protect the lsn field of each item.  The value of this
1376  * field can never go backwards.
1377  *
1378  * We unpin the items after repositioning them in the AIL, because
1379  * otherwise they could be immediately flushed and we'd have to race
1380  * with the flusher trying to pull the item from the AIL as we add it.
1381  */
1382 STATIC void
1383 xfs_trans_chunk_committed(
1384 	xfs_log_item_chunk_t	*licp,
1385 	xfs_lsn_t		lsn,
1386 	int			aborted)
1387 {
1388 	xfs_log_item_desc_t	*lidp;
1389 	xfs_log_item_t		*lip;
1390 	xfs_lsn_t		item_lsn;
1391 	int			i;
1392 
1393 	lidp = licp->lic_descs;
1394 	for (i = 0; i < licp->lic_unused; i++, lidp++) {
1395 		struct xfs_ail		*ailp;
1396 
1397 		if (xfs_lic_isfree(licp, i)) {
1398 			continue;
1399 		}
1400 
1401 		lip = lidp->lid_item;
1402 		if (aborted)
1403 			lip->li_flags |= XFS_LI_ABORTED;
1404 
1405 		/*
1406 		 * Send in the ABORTED flag to the COMMITTED routine
1407 		 * so that it knows whether the transaction was aborted
1408 		 * or not.
1409 		 */
1410 		item_lsn = IOP_COMMITTED(lip, lsn);
1411 
1412 		/*
1413 		 * If the committed routine returns -1, make
1414 		 * no more references to the item.
1415 		 */
1416 		if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0) {
1417 			continue;
1418 		}
1419 
1420 		/*
1421 		 * If the returned lsn is greater than what it
1422 		 * contained before, update the location of the
1423 		 * item in the AIL.  If it is not, then do nothing.
1424 		 * Items can never move backwards in the AIL.
1425 		 *
1426 		 * While the new lsn should usually be greater, it
1427 		 * is possible that a later transaction completing
1428 		 * simultaneously with an earlier one using the
1429 		 * same item could complete first with a higher lsn.
1430 		 * This would cause the earlier transaction to fail
1431 		 * the test below.
1432 		 */
1433 		ailp = lip->li_ailp;
1434 		spin_lock(&ailp->xa_lock);
1435 		if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0) {
1436 			/*
1437 			 * This will set the item's lsn to item_lsn
1438 			 * and update the position of the item in
1439 			 * the AIL.
1440 			 *
1441 			 * xfs_trans_ail_update() drops the AIL lock.
1442 			 */
1443 			xfs_trans_ail_update(ailp, lip, item_lsn);
1444 		} else {
1445 			spin_unlock(&ailp->xa_lock);
1446 		}
1447 
1448 		/*
1449 		 * Now that we've repositioned the item in the AIL,
1450 		 * unpin it so it can be flushed. Pass information
1451 		 * about buffer stale state down from the log item
1452 		 * flags, if anyone else stales the buffer we do not
1453 		 * want to pay any attention to it.
1454 		 */
1455 		IOP_UNPIN(lip, lidp->lid_flags & XFS_LID_BUF_STALE);
1456 	}
1457 }
1458