xref: /openbmc/linux/fs/xfs/xfs_trans.c (revision 9ac8d3fb)
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);
240 }
241 
242 xfs_trans_t *
243 _xfs_trans_alloc(
244 	xfs_mount_t	*mp,
245 	uint		type)
246 {
247 	xfs_trans_t	*tp;
248 
249 	atomic_inc(&mp->m_active_trans);
250 
251 	tp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
252 	tp->t_magic = XFS_TRANS_MAGIC;
253 	tp->t_type = type;
254 	tp->t_mountp = mp;
255 	tp->t_items_free = XFS_LIC_NUM_SLOTS;
256 	tp->t_busy_free = XFS_LBC_NUM_SLOTS;
257 	xfs_lic_init(&(tp->t_items));
258 	XFS_LBC_INIT(&(tp->t_busy));
259 	return tp;
260 }
261 
262 /*
263  * This is called to create a new transaction which will share the
264  * permanent log reservation of the given transaction.  The remaining
265  * unused block and rt extent reservations are also inherited.  This
266  * implies that the original transaction is no longer allowed to allocate
267  * blocks.  Locks and log items, however, are no inherited.  They must
268  * be added to the new transaction explicitly.
269  */
270 xfs_trans_t *
271 xfs_trans_dup(
272 	xfs_trans_t	*tp)
273 {
274 	xfs_trans_t	*ntp;
275 
276 	ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
277 
278 	/*
279 	 * Initialize the new transaction structure.
280 	 */
281 	ntp->t_magic = XFS_TRANS_MAGIC;
282 	ntp->t_type = tp->t_type;
283 	ntp->t_mountp = tp->t_mountp;
284 	ntp->t_items_free = XFS_LIC_NUM_SLOTS;
285 	ntp->t_busy_free = XFS_LBC_NUM_SLOTS;
286 	xfs_lic_init(&(ntp->t_items));
287 	XFS_LBC_INIT(&(ntp->t_busy));
288 
289 	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
290 	ASSERT(tp->t_ticket != NULL);
291 
292 	ntp->t_flags = XFS_TRANS_PERM_LOG_RES | (tp->t_flags & XFS_TRANS_RESERVE);
293 	ntp->t_ticket = tp->t_ticket;
294 	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
295 	tp->t_blk_res = tp->t_blk_res_used;
296 	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
297 	tp->t_rtx_res = tp->t_rtx_res_used;
298 	ntp->t_pflags = tp->t_pflags;
299 
300 	XFS_TRANS_DUP_DQINFO(tp->t_mountp, tp, ntp);
301 
302 	atomic_inc(&tp->t_mountp->m_active_trans);
303 	return ntp;
304 }
305 
306 /*
307  * This is called to reserve free disk blocks and log space for the
308  * given transaction.  This must be done before allocating any resources
309  * within the transaction.
310  *
311  * This will return ENOSPC if there are not enough blocks available.
312  * It will sleep waiting for available log space.
313  * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
314  * is used by long running transactions.  If any one of the reservations
315  * fails then they will all be backed out.
316  *
317  * This does not do quota reservations. That typically is done by the
318  * caller afterwards.
319  */
320 int
321 xfs_trans_reserve(
322 	xfs_trans_t	*tp,
323 	uint		blocks,
324 	uint		logspace,
325 	uint		rtextents,
326 	uint		flags,
327 	uint		logcount)
328 {
329 	int		log_flags;
330 	int		error = 0;
331 	int		rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
332 
333 	/* Mark this thread as being in a transaction */
334 	current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
335 
336 	/*
337 	 * Attempt to reserve the needed disk blocks by decrementing
338 	 * the number needed from the number available.  This will
339 	 * fail if the count would go below zero.
340 	 */
341 	if (blocks > 0) {
342 		error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
343 					  -((int64_t)blocks), rsvd);
344 		if (error != 0) {
345 			current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
346 			return (XFS_ERROR(ENOSPC));
347 		}
348 		tp->t_blk_res += blocks;
349 	}
350 
351 	/*
352 	 * Reserve the log space needed for this transaction.
353 	 */
354 	if (logspace > 0) {
355 		ASSERT((tp->t_log_res == 0) || (tp->t_log_res == logspace));
356 		ASSERT((tp->t_log_count == 0) ||
357 			(tp->t_log_count == logcount));
358 		if (flags & XFS_TRANS_PERM_LOG_RES) {
359 			log_flags = XFS_LOG_PERM_RESERV;
360 			tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
361 		} else {
362 			ASSERT(tp->t_ticket == NULL);
363 			ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
364 			log_flags = 0;
365 		}
366 
367 		error = xfs_log_reserve(tp->t_mountp, logspace, logcount,
368 					&tp->t_ticket,
369 					XFS_TRANSACTION, log_flags, tp->t_type);
370 		if (error) {
371 			goto undo_blocks;
372 		}
373 		tp->t_log_res = logspace;
374 		tp->t_log_count = logcount;
375 	}
376 
377 	/*
378 	 * Attempt to reserve the needed realtime extents by decrementing
379 	 * the number needed from the number available.  This will
380 	 * fail if the count would go below zero.
381 	 */
382 	if (rtextents > 0) {
383 		error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FREXTENTS,
384 					  -((int64_t)rtextents), rsvd);
385 		if (error) {
386 			error = XFS_ERROR(ENOSPC);
387 			goto undo_log;
388 		}
389 		tp->t_rtx_res += rtextents;
390 	}
391 
392 	return 0;
393 
394 	/*
395 	 * Error cases jump to one of these labels to undo any
396 	 * reservations which have already been performed.
397 	 */
398 undo_log:
399 	if (logspace > 0) {
400 		if (flags & XFS_TRANS_PERM_LOG_RES) {
401 			log_flags = XFS_LOG_REL_PERM_RESERV;
402 		} else {
403 			log_flags = 0;
404 		}
405 		xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, log_flags);
406 		tp->t_ticket = NULL;
407 		tp->t_log_res = 0;
408 		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
409 	}
410 
411 undo_blocks:
412 	if (blocks > 0) {
413 		(void) xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FDBLOCKS,
414 					 (int64_t)blocks, rsvd);
415 		tp->t_blk_res = 0;
416 	}
417 
418 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
419 
420 	return error;
421 }
422 
423 
424 /*
425  * Record the indicated change to the given field for application
426  * to the file system's superblock when the transaction commits.
427  * For now, just store the change in the transaction structure.
428  *
429  * Mark the transaction structure to indicate that the superblock
430  * needs to be updated before committing.
431  *
432  * Because we may not be keeping track of allocated/free inodes and
433  * used filesystem blocks in the superblock, we do not mark the
434  * superblock dirty in this transaction if we modify these fields.
435  * We still need to update the transaction deltas so that they get
436  * applied to the incore superblock, but we don't want them to
437  * cause the superblock to get locked and logged if these are the
438  * only fields in the superblock that the transaction modifies.
439  */
440 void
441 xfs_trans_mod_sb(
442 	xfs_trans_t	*tp,
443 	uint		field,
444 	int64_t		delta)
445 {
446 	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
447 	xfs_mount_t	*mp = tp->t_mountp;
448 
449 	switch (field) {
450 	case XFS_TRANS_SB_ICOUNT:
451 		tp->t_icount_delta += delta;
452 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
453 			flags &= ~XFS_TRANS_SB_DIRTY;
454 		break;
455 	case XFS_TRANS_SB_IFREE:
456 		tp->t_ifree_delta += delta;
457 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
458 			flags &= ~XFS_TRANS_SB_DIRTY;
459 		break;
460 	case XFS_TRANS_SB_FDBLOCKS:
461 		/*
462 		 * Track the number of blocks allocated in the
463 		 * transaction.  Make sure it does not exceed the
464 		 * number reserved.
465 		 */
466 		if (delta < 0) {
467 			tp->t_blk_res_used += (uint)-delta;
468 			ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
469 		}
470 		tp->t_fdblocks_delta += delta;
471 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
472 			flags &= ~XFS_TRANS_SB_DIRTY;
473 		break;
474 	case XFS_TRANS_SB_RES_FDBLOCKS:
475 		/*
476 		 * The allocation has already been applied to the
477 		 * in-core superblock's counter.  This should only
478 		 * be applied to the on-disk superblock.
479 		 */
480 		ASSERT(delta < 0);
481 		tp->t_res_fdblocks_delta += delta;
482 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
483 			flags &= ~XFS_TRANS_SB_DIRTY;
484 		break;
485 	case XFS_TRANS_SB_FREXTENTS:
486 		/*
487 		 * Track the number of blocks allocated in the
488 		 * transaction.  Make sure it does not exceed the
489 		 * number reserved.
490 		 */
491 		if (delta < 0) {
492 			tp->t_rtx_res_used += (uint)-delta;
493 			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
494 		}
495 		tp->t_frextents_delta += delta;
496 		break;
497 	case XFS_TRANS_SB_RES_FREXTENTS:
498 		/*
499 		 * The allocation has already been applied to the
500 		 * in-core superblock's counter.  This should only
501 		 * be applied to the on-disk superblock.
502 		 */
503 		ASSERT(delta < 0);
504 		tp->t_res_frextents_delta += delta;
505 		break;
506 	case XFS_TRANS_SB_DBLOCKS:
507 		ASSERT(delta > 0);
508 		tp->t_dblocks_delta += delta;
509 		break;
510 	case XFS_TRANS_SB_AGCOUNT:
511 		ASSERT(delta > 0);
512 		tp->t_agcount_delta += delta;
513 		break;
514 	case XFS_TRANS_SB_IMAXPCT:
515 		tp->t_imaxpct_delta += delta;
516 		break;
517 	case XFS_TRANS_SB_REXTSIZE:
518 		tp->t_rextsize_delta += delta;
519 		break;
520 	case XFS_TRANS_SB_RBMBLOCKS:
521 		tp->t_rbmblocks_delta += delta;
522 		break;
523 	case XFS_TRANS_SB_RBLOCKS:
524 		tp->t_rblocks_delta += delta;
525 		break;
526 	case XFS_TRANS_SB_REXTENTS:
527 		tp->t_rextents_delta += delta;
528 		break;
529 	case XFS_TRANS_SB_REXTSLOG:
530 		tp->t_rextslog_delta += delta;
531 		break;
532 	default:
533 		ASSERT(0);
534 		return;
535 	}
536 
537 	tp->t_flags |= flags;
538 }
539 
540 /*
541  * xfs_trans_apply_sb_deltas() is called from the commit code
542  * to bring the superblock buffer into the current transaction
543  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
544  *
545  * For now we just look at each field allowed to change and change
546  * it if necessary.
547  */
548 STATIC void
549 xfs_trans_apply_sb_deltas(
550 	xfs_trans_t	*tp)
551 {
552 	xfs_dsb_t	*sbp;
553 	xfs_buf_t	*bp;
554 	int		whole = 0;
555 
556 	bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
557 	sbp = XFS_BUF_TO_SBP(bp);
558 
559 	/*
560 	 * Check that superblock mods match the mods made to AGF counters.
561 	 */
562 	ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
563 	       (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
564 		tp->t_ag_btree_delta));
565 
566 	/*
567 	 * Only update the superblock counters if we are logging them
568 	 */
569 	if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
570 		if (tp->t_icount_delta)
571 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
572 		if (tp->t_ifree_delta)
573 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
574 		if (tp->t_fdblocks_delta)
575 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
576 		if (tp->t_res_fdblocks_delta)
577 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
578 	}
579 
580 	if (tp->t_frextents_delta)
581 		be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
582 	if (tp->t_res_frextents_delta)
583 		be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
584 
585 	if (tp->t_dblocks_delta) {
586 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
587 		whole = 1;
588 	}
589 	if (tp->t_agcount_delta) {
590 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
591 		whole = 1;
592 	}
593 	if (tp->t_imaxpct_delta) {
594 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
595 		whole = 1;
596 	}
597 	if (tp->t_rextsize_delta) {
598 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
599 		whole = 1;
600 	}
601 	if (tp->t_rbmblocks_delta) {
602 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
603 		whole = 1;
604 	}
605 	if (tp->t_rblocks_delta) {
606 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
607 		whole = 1;
608 	}
609 	if (tp->t_rextents_delta) {
610 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
611 		whole = 1;
612 	}
613 	if (tp->t_rextslog_delta) {
614 		sbp->sb_rextslog += tp->t_rextslog_delta;
615 		whole = 1;
616 	}
617 
618 	if (whole)
619 		/*
620 		 * Log the whole thing, the fields are noncontiguous.
621 		 */
622 		xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
623 	else
624 		/*
625 		 * Since all the modifiable fields are contiguous, we
626 		 * can get away with this.
627 		 */
628 		xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
629 				  offsetof(xfs_dsb_t, sb_frextents) +
630 				  sizeof(sbp->sb_frextents) - 1);
631 
632 	tp->t_mountp->m_super->s_dirt = 1;
633 }
634 
635 /*
636  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
637  * and apply superblock counter changes to the in-core superblock.  The
638  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
639  * applied to the in-core superblock.  The idea is that that has already been
640  * done.
641  *
642  * This is done efficiently with a single call to xfs_mod_incore_sb_batch().
643  * However, we have to ensure that we only modify each superblock field only
644  * once because the application of the delta values may not be atomic. That can
645  * lead to ENOSPC races occurring if we have two separate modifcations of the
646  * free space counter to put back the entire reservation and then take away
647  * what we used.
648  *
649  * If we are not logging superblock counters, then the inode allocated/free and
650  * used block counts are not updated in the on disk superblock. In this case,
651  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
652  * still need to update the incore superblock with the changes.
653  */
654 STATIC void
655 xfs_trans_unreserve_and_mod_sb(
656 	xfs_trans_t	*tp)
657 {
658 	xfs_mod_sb_t	msb[14];	/* If you add cases, add entries */
659 	xfs_mod_sb_t	*msbp;
660 	xfs_mount_t	*mp = tp->t_mountp;
661 	/* REFERENCED */
662 	int		error;
663 	int		rsvd;
664 	int64_t		blkdelta = 0;
665 	int64_t		rtxdelta = 0;
666 
667 	msbp = msb;
668 	rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
669 
670 	/* calculate free blocks delta */
671 	if (tp->t_blk_res > 0)
672 		blkdelta = tp->t_blk_res;
673 
674 	if ((tp->t_fdblocks_delta != 0) &&
675 	    (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
676 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)))
677 	        blkdelta += tp->t_fdblocks_delta;
678 
679 	if (blkdelta != 0) {
680 		msbp->msb_field = XFS_SBS_FDBLOCKS;
681 		msbp->msb_delta = blkdelta;
682 		msbp++;
683 	}
684 
685 	/* calculate free realtime extents delta */
686 	if (tp->t_rtx_res > 0)
687 		rtxdelta = tp->t_rtx_res;
688 
689 	if ((tp->t_frextents_delta != 0) &&
690 	    (tp->t_flags & XFS_TRANS_SB_DIRTY))
691 		rtxdelta += tp->t_frextents_delta;
692 
693 	if (rtxdelta != 0) {
694 		msbp->msb_field = XFS_SBS_FREXTENTS;
695 		msbp->msb_delta = rtxdelta;
696 		msbp++;
697 	}
698 
699 	/* apply remaining deltas */
700 
701 	if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
702 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
703 		if (tp->t_icount_delta != 0) {
704 			msbp->msb_field = XFS_SBS_ICOUNT;
705 			msbp->msb_delta = tp->t_icount_delta;
706 			msbp++;
707 		}
708 		if (tp->t_ifree_delta != 0) {
709 			msbp->msb_field = XFS_SBS_IFREE;
710 			msbp->msb_delta = tp->t_ifree_delta;
711 			msbp++;
712 		}
713 	}
714 
715 	if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
716 		if (tp->t_dblocks_delta != 0) {
717 			msbp->msb_field = XFS_SBS_DBLOCKS;
718 			msbp->msb_delta = tp->t_dblocks_delta;
719 			msbp++;
720 		}
721 		if (tp->t_agcount_delta != 0) {
722 			msbp->msb_field = XFS_SBS_AGCOUNT;
723 			msbp->msb_delta = tp->t_agcount_delta;
724 			msbp++;
725 		}
726 		if (tp->t_imaxpct_delta != 0) {
727 			msbp->msb_field = XFS_SBS_IMAX_PCT;
728 			msbp->msb_delta = tp->t_imaxpct_delta;
729 			msbp++;
730 		}
731 		if (tp->t_rextsize_delta != 0) {
732 			msbp->msb_field = XFS_SBS_REXTSIZE;
733 			msbp->msb_delta = tp->t_rextsize_delta;
734 			msbp++;
735 		}
736 		if (tp->t_rbmblocks_delta != 0) {
737 			msbp->msb_field = XFS_SBS_RBMBLOCKS;
738 			msbp->msb_delta = tp->t_rbmblocks_delta;
739 			msbp++;
740 		}
741 		if (tp->t_rblocks_delta != 0) {
742 			msbp->msb_field = XFS_SBS_RBLOCKS;
743 			msbp->msb_delta = tp->t_rblocks_delta;
744 			msbp++;
745 		}
746 		if (tp->t_rextents_delta != 0) {
747 			msbp->msb_field = XFS_SBS_REXTENTS;
748 			msbp->msb_delta = tp->t_rextents_delta;
749 			msbp++;
750 		}
751 		if (tp->t_rextslog_delta != 0) {
752 			msbp->msb_field = XFS_SBS_REXTSLOG;
753 			msbp->msb_delta = tp->t_rextslog_delta;
754 			msbp++;
755 		}
756 	}
757 
758 	/*
759 	 * If we need to change anything, do it.
760 	 */
761 	if (msbp > msb) {
762 		error = xfs_mod_incore_sb_batch(tp->t_mountp, msb,
763 			(uint)(msbp - msb), rsvd);
764 		ASSERT(error == 0);
765 	}
766 }
767 
768 
769 /*
770  * xfs_trans_commit
771  *
772  * Commit the given transaction to the log a/synchronously.
773  *
774  * XFS disk error handling mechanism is not based on a typical
775  * transaction abort mechanism. Logically after the filesystem
776  * gets marked 'SHUTDOWN', we can't let any new transactions
777  * be durable - ie. committed to disk - because some metadata might
778  * be inconsistent. In such cases, this returns an error, and the
779  * caller may assume that all locked objects joined to the transaction
780  * have already been unlocked as if the commit had succeeded.
781  * Do not reference the transaction structure after this call.
782  */
783  /*ARGSUSED*/
784 int
785 _xfs_trans_commit(
786 	xfs_trans_t	*tp,
787 	uint		flags,
788 	int		*log_flushed)
789 {
790 	xfs_log_iovec_t		*log_vector;
791 	int			nvec;
792 	xfs_mount_t		*mp;
793 	xfs_lsn_t		commit_lsn;
794 	/* REFERENCED */
795 	int			error;
796 	int			log_flags;
797 	int			sync;
798 #define	XFS_TRANS_LOGVEC_COUNT	16
799 	xfs_log_iovec_t		log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
800 	void			*commit_iclog;
801 	int			shutdown;
802 
803 	commit_lsn = -1;
804 
805 	/*
806 	 * Determine whether this commit is releasing a permanent
807 	 * log reservation or not.
808 	 */
809 	if (flags & XFS_TRANS_RELEASE_LOG_RES) {
810 		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
811 		log_flags = XFS_LOG_REL_PERM_RESERV;
812 	} else {
813 		log_flags = 0;
814 	}
815 	mp = tp->t_mountp;
816 
817 	/*
818 	 * If there is nothing to be logged by the transaction,
819 	 * then unlock all of the items associated with the
820 	 * transaction and free the transaction structure.
821 	 * Also make sure to return any reserved blocks to
822 	 * the free pool.
823 	 */
824 shut_us_down:
825 	shutdown = XFS_FORCED_SHUTDOWN(mp) ? EIO : 0;
826 	if (!(tp->t_flags & XFS_TRANS_DIRTY) || shutdown) {
827 		xfs_trans_unreserve_and_mod_sb(tp);
828 		/*
829 		 * It is indeed possible for the transaction to be
830 		 * not dirty but the dqinfo portion to be. All that
831 		 * means is that we have some (non-persistent) quota
832 		 * reservations that need to be unreserved.
833 		 */
834 		XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(mp, tp);
835 		if (tp->t_ticket) {
836 			commit_lsn = xfs_log_done(mp, tp->t_ticket,
837 							NULL, log_flags);
838 			if (commit_lsn == -1 && !shutdown)
839 				shutdown = XFS_ERROR(EIO);
840 		}
841 		current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
842 		xfs_trans_free_items(tp, shutdown? XFS_TRANS_ABORT : 0);
843 		xfs_trans_free_busy(tp);
844 		xfs_trans_free(tp);
845 		XFS_STATS_INC(xs_trans_empty);
846 		return (shutdown);
847 	}
848 	ASSERT(tp->t_ticket != NULL);
849 
850 	/*
851 	 * If we need to update the superblock, then do it now.
852 	 */
853 	if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
854 		xfs_trans_apply_sb_deltas(tp);
855 	}
856 	XFS_TRANS_APPLY_DQUOT_DELTAS(mp, tp);
857 
858 	/*
859 	 * Ask each log item how many log_vector entries it will
860 	 * need so we can figure out how many to allocate.
861 	 * Try to avoid the kmem_alloc() call in the common case
862 	 * by using a vector from the stack when it fits.
863 	 */
864 	nvec = xfs_trans_count_vecs(tp);
865 	if (nvec == 0) {
866 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
867 		goto shut_us_down;
868 	} else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
869 		log_vector = log_vector_fast;
870 	} else {
871 		log_vector = (xfs_log_iovec_t *)kmem_alloc(nvec *
872 						   sizeof(xfs_log_iovec_t),
873 						   KM_SLEEP);
874 	}
875 
876 	/*
877 	 * Fill in the log_vector and pin the logged items, and
878 	 * then write the transaction to the log.
879 	 */
880 	xfs_trans_fill_vecs(tp, log_vector);
881 
882 	error = xfs_log_write(mp, log_vector, nvec, tp->t_ticket, &(tp->t_lsn));
883 
884 	/*
885 	 * The transaction is committed incore here, and can go out to disk
886 	 * at any time after this call.  However, all the items associated
887 	 * with the transaction are still locked and pinned in memory.
888 	 */
889 	commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
890 
891 	tp->t_commit_lsn = commit_lsn;
892 	if (nvec > XFS_TRANS_LOGVEC_COUNT) {
893 		kmem_free(log_vector);
894 	}
895 
896 	/*
897 	 * If we got a log write error. Unpin the logitems that we
898 	 * had pinned, clean up, free trans structure, and return error.
899 	 */
900 	if (error || commit_lsn == -1) {
901 		current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
902 		xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
903 		return XFS_ERROR(EIO);
904 	}
905 
906 	/*
907 	 * Once the transaction has committed, unused
908 	 * reservations need to be released and changes to
909 	 * the superblock need to be reflected in the in-core
910 	 * version.  Do that now.
911 	 */
912 	xfs_trans_unreserve_and_mod_sb(tp);
913 
914 	sync = tp->t_flags & XFS_TRANS_SYNC;
915 
916 	/*
917 	 * Tell the LM to call the transaction completion routine
918 	 * when the log write with LSN commit_lsn completes (e.g.
919 	 * when the transaction commit really hits the on-disk log).
920 	 * After this call we cannot reference tp, because the call
921 	 * can happen at any time and the call will free the transaction
922 	 * structure pointed to by tp.  The only case where we call
923 	 * the completion routine (xfs_trans_committed) directly is
924 	 * if the log is turned off on a debug kernel or we're
925 	 * running in simulation mode (the log is explicitly turned
926 	 * off).
927 	 */
928 	tp->t_logcb.cb_func = (void(*)(void*, int))xfs_trans_committed;
929 	tp->t_logcb.cb_arg = tp;
930 
931 	/*
932 	 * We need to pass the iclog buffer which was used for the
933 	 * transaction commit record into this function, and attach
934 	 * the callback to it. The callback must be attached before
935 	 * the items are unlocked to avoid racing with other threads
936 	 * waiting for an item to unlock.
937 	 */
938 	shutdown = xfs_log_notify(mp, commit_iclog, &(tp->t_logcb));
939 
940 	/*
941 	 * Mark this thread as no longer being in a transaction
942 	 */
943 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
944 
945 	/*
946 	 * Once all the items of the transaction have been copied
947 	 * to the in core log and the callback is attached, the
948 	 * items can be unlocked.
949 	 *
950 	 * This will free descriptors pointing to items which were
951 	 * not logged since there is nothing more to do with them.
952 	 * For items which were logged, we will keep pointers to them
953 	 * so they can be unpinned after the transaction commits to disk.
954 	 * This will also stamp each modified meta-data item with
955 	 * the commit lsn of this transaction for dependency tracking
956 	 * purposes.
957 	 */
958 	xfs_trans_unlock_items(tp, commit_lsn);
959 
960 	/*
961 	 * If we detected a log error earlier, finish committing
962 	 * the transaction now (unpin log items, etc).
963 	 *
964 	 * Order is critical here, to avoid using the transaction
965 	 * pointer after its been freed (by xfs_trans_committed
966 	 * either here now, or as a callback).  We cannot do this
967 	 * step inside xfs_log_notify as was done earlier because
968 	 * of this issue.
969 	 */
970 	if (shutdown)
971 		xfs_trans_committed(tp, XFS_LI_ABORTED);
972 
973 	/*
974 	 * Now that the xfs_trans_committed callback has been attached,
975 	 * and the items are released we can finally allow the iclog to
976 	 * go to disk.
977 	 */
978 	error = xfs_log_release_iclog(mp, commit_iclog);
979 
980 	/*
981 	 * If the transaction needs to be synchronous, then force the
982 	 * log out now and wait for it.
983 	 */
984 	if (sync) {
985 		if (!error) {
986 			error = _xfs_log_force(mp, commit_lsn,
987 				      XFS_LOG_FORCE | XFS_LOG_SYNC,
988 				      log_flushed);
989 		}
990 		XFS_STATS_INC(xs_trans_sync);
991 	} else {
992 		XFS_STATS_INC(xs_trans_async);
993 	}
994 
995 	return (error);
996 }
997 
998 
999 /*
1000  * Total up the number of log iovecs needed to commit this
1001  * transaction.  The transaction itself needs one for the
1002  * transaction header.  Ask each dirty item in turn how many
1003  * it needs to get the total.
1004  */
1005 STATIC uint
1006 xfs_trans_count_vecs(
1007 	xfs_trans_t	*tp)
1008 {
1009 	int			nvecs;
1010 	xfs_log_item_desc_t	*lidp;
1011 
1012 	nvecs = 1;
1013 	lidp = xfs_trans_first_item(tp);
1014 	ASSERT(lidp != NULL);
1015 
1016 	/* In the non-debug case we need to start bailing out if we
1017 	 * didn't find a log_item here, return zero and let trans_commit
1018 	 * deal with it.
1019 	 */
1020 	if (lidp == NULL)
1021 		return 0;
1022 
1023 	while (lidp != NULL) {
1024 		/*
1025 		 * Skip items which aren't dirty in this transaction.
1026 		 */
1027 		if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
1028 			lidp = xfs_trans_next_item(tp, lidp);
1029 			continue;
1030 		}
1031 		lidp->lid_size = IOP_SIZE(lidp->lid_item);
1032 		nvecs += lidp->lid_size;
1033 		lidp = xfs_trans_next_item(tp, lidp);
1034 	}
1035 
1036 	return nvecs;
1037 }
1038 
1039 /*
1040  * Called from the trans_commit code when we notice that
1041  * the filesystem is in the middle of a forced shutdown.
1042  */
1043 STATIC void
1044 xfs_trans_uncommit(
1045 	xfs_trans_t	*tp,
1046 	uint		flags)
1047 {
1048 	xfs_log_item_desc_t	*lidp;
1049 
1050 	for (lidp = xfs_trans_first_item(tp);
1051 	     lidp != NULL;
1052 	     lidp = xfs_trans_next_item(tp, lidp)) {
1053 		/*
1054 		 * Unpin all but those that aren't dirty.
1055 		 */
1056 		if (lidp->lid_flags & XFS_LID_DIRTY)
1057 			IOP_UNPIN_REMOVE(lidp->lid_item, tp);
1058 	}
1059 
1060 	xfs_trans_unreserve_and_mod_sb(tp);
1061 	XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(tp->t_mountp, tp);
1062 
1063 	xfs_trans_free_items(tp, flags);
1064 	xfs_trans_free_busy(tp);
1065 	xfs_trans_free(tp);
1066 }
1067 
1068 /*
1069  * Fill in the vector with pointers to data to be logged
1070  * by this transaction.  The transaction header takes
1071  * the first vector, and then each dirty item takes the
1072  * number of vectors it indicated it needed in xfs_trans_count_vecs().
1073  *
1074  * As each item fills in the entries it needs, also pin the item
1075  * so that it cannot be flushed out until the log write completes.
1076  */
1077 STATIC void
1078 xfs_trans_fill_vecs(
1079 	xfs_trans_t		*tp,
1080 	xfs_log_iovec_t		*log_vector)
1081 {
1082 	xfs_log_item_desc_t	*lidp;
1083 	xfs_log_iovec_t		*vecp;
1084 	uint			nitems;
1085 
1086 	/*
1087 	 * Skip over the entry for the transaction header, we'll
1088 	 * fill that in at the end.
1089 	 */
1090 	vecp = log_vector + 1;		/* pointer arithmetic */
1091 
1092 	nitems = 0;
1093 	lidp = xfs_trans_first_item(tp);
1094 	ASSERT(lidp != NULL);
1095 	while (lidp != NULL) {
1096 		/*
1097 		 * Skip items which aren't dirty in this transaction.
1098 		 */
1099 		if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
1100 			lidp = xfs_trans_next_item(tp, lidp);
1101 			continue;
1102 		}
1103 		/*
1104 		 * The item may be marked dirty but not log anything.
1105 		 * This can be used to get called when a transaction
1106 		 * is committed.
1107 		 */
1108 		if (lidp->lid_size) {
1109 			nitems++;
1110 		}
1111 		IOP_FORMAT(lidp->lid_item, vecp);
1112 		vecp += lidp->lid_size;		/* pointer arithmetic */
1113 		IOP_PIN(lidp->lid_item);
1114 		lidp = xfs_trans_next_item(tp, lidp);
1115 	}
1116 
1117 	/*
1118 	 * Now that we've counted the number of items in this
1119 	 * transaction, fill in the transaction header.
1120 	 */
1121 	tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
1122 	tp->t_header.th_type = tp->t_type;
1123 	tp->t_header.th_num_items = nitems;
1124 	log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
1125 	log_vector->i_len = sizeof(xfs_trans_header_t);
1126 	XLOG_VEC_SET_TYPE(log_vector, XLOG_REG_TYPE_TRANSHDR);
1127 }
1128 
1129 
1130 /*
1131  * Unlock all of the transaction's items and free the transaction.
1132  * The transaction must not have modified any of its items, because
1133  * there is no way to restore them to their previous state.
1134  *
1135  * If the transaction has made a log reservation, make sure to release
1136  * it as well.
1137  */
1138 void
1139 xfs_trans_cancel(
1140 	xfs_trans_t		*tp,
1141 	int			flags)
1142 {
1143 	int			log_flags;
1144 #ifdef DEBUG
1145 	xfs_log_item_chunk_t	*licp;
1146 	xfs_log_item_desc_t	*lidp;
1147 	xfs_log_item_t		*lip;
1148 	int			i;
1149 #endif
1150 	xfs_mount_t		*mp = tp->t_mountp;
1151 
1152 	/*
1153 	 * See if the caller is being too lazy to figure out if
1154 	 * the transaction really needs an abort.
1155 	 */
1156 	if ((flags & XFS_TRANS_ABORT) && !(tp->t_flags & XFS_TRANS_DIRTY))
1157 		flags &= ~XFS_TRANS_ABORT;
1158 	/*
1159 	 * See if the caller is relying on us to shut down the
1160 	 * filesystem.  This happens in paths where we detect
1161 	 * corruption and decide to give up.
1162 	 */
1163 	if ((tp->t_flags & XFS_TRANS_DIRTY) && !XFS_FORCED_SHUTDOWN(mp)) {
1164 		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1165 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1166 	}
1167 #ifdef DEBUG
1168 	if (!(flags & XFS_TRANS_ABORT)) {
1169 		licp = &(tp->t_items);
1170 		while (licp != NULL) {
1171 			lidp = licp->lic_descs;
1172 			for (i = 0; i < licp->lic_unused; i++, lidp++) {
1173 				if (xfs_lic_isfree(licp, i)) {
1174 					continue;
1175 				}
1176 
1177 				lip = lidp->lid_item;
1178 				if (!XFS_FORCED_SHUTDOWN(mp))
1179 					ASSERT(!(lip->li_type == XFS_LI_EFD));
1180 			}
1181 			licp = licp->lic_next;
1182 		}
1183 	}
1184 #endif
1185 	xfs_trans_unreserve_and_mod_sb(tp);
1186 	XFS_TRANS_UNRESERVE_AND_MOD_DQUOTS(mp, tp);
1187 
1188 	if (tp->t_ticket) {
1189 		if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1190 			ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1191 			log_flags = XFS_LOG_REL_PERM_RESERV;
1192 		} else {
1193 			log_flags = 0;
1194 		}
1195 		xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1196 	}
1197 
1198 	/* mark this thread as no longer being in a transaction */
1199 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1200 
1201 	xfs_trans_free_items(tp, flags);
1202 	xfs_trans_free_busy(tp);
1203 	xfs_trans_free(tp);
1204 }
1205 
1206 
1207 /*
1208  * Free the transaction structure.  If there is more clean up
1209  * to do when the structure is freed, add it here.
1210  */
1211 STATIC void
1212 xfs_trans_free(
1213 	xfs_trans_t	*tp)
1214 {
1215 	atomic_dec(&tp->t_mountp->m_active_trans);
1216 	XFS_TRANS_FREE_DQINFO(tp->t_mountp, tp);
1217 	kmem_zone_free(xfs_trans_zone, tp);
1218 }
1219 
1220 /*
1221  * Roll from one trans in the sequence of PERMANENT transactions to
1222  * the next: permanent transactions are only flushed out when
1223  * committed with XFS_TRANS_RELEASE_LOG_RES, but we still want as soon
1224  * as possible to let chunks of it go to the log. So we commit the
1225  * chunk we've been working on and get a new transaction to continue.
1226  */
1227 int
1228 xfs_trans_roll(
1229 	struct xfs_trans	**tpp,
1230 	struct xfs_inode	*dp)
1231 {
1232 	struct xfs_trans	*trans;
1233 	unsigned int		logres, count;
1234 	int			error;
1235 
1236 	/*
1237 	 * Ensure that the inode is always logged.
1238 	 */
1239 	trans = *tpp;
1240 	xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
1241 
1242 	/*
1243 	 * Copy the critical parameters from one trans to the next.
1244 	 */
1245 	logres = trans->t_log_res;
1246 	count = trans->t_log_count;
1247 	*tpp = xfs_trans_dup(trans);
1248 
1249 	/*
1250 	 * Commit the current transaction.
1251 	 * If this commit failed, then it'd just unlock those items that
1252 	 * are not marked ihold. That also means that a filesystem shutdown
1253 	 * is in progress. The caller takes the responsibility to cancel
1254 	 * the duplicate transaction that gets returned.
1255 	 */
1256 	error = xfs_trans_commit(trans, 0);
1257 	if (error)
1258 		return (error);
1259 
1260 	trans = *tpp;
1261 
1262 	/*
1263 	 * Reserve space in the log for th next transaction.
1264 	 * This also pushes items in the "AIL", the list of logged items,
1265 	 * out to disk if they are taking up space at the tail of the log
1266 	 * that we want to use.  This requires that either nothing be locked
1267 	 * across this call, or that anything that is locked be logged in
1268 	 * the prior and the next transactions.
1269 	 */
1270 	error = xfs_trans_reserve(trans, 0, logres, 0,
1271 				  XFS_TRANS_PERM_LOG_RES, count);
1272 	/*
1273 	 *  Ensure that the inode is in the new transaction and locked.
1274 	 */
1275 	if (error)
1276 		return error;
1277 
1278 	xfs_trans_ijoin(trans, dp, XFS_ILOCK_EXCL);
1279 	xfs_trans_ihold(trans, dp);
1280 	return 0;
1281 }
1282 
1283 /*
1284  * THIS SHOULD BE REWRITTEN TO USE xfs_trans_next_item().
1285  *
1286  * This is typically called by the LM when a transaction has been fully
1287  * committed to disk.  It needs to unpin the items which have
1288  * been logged by the transaction and update their positions
1289  * in the AIL if necessary.
1290  * This also gets called when the transactions didn't get written out
1291  * because of an I/O error. Abortflag & XFS_LI_ABORTED is set then.
1292  *
1293  * Call xfs_trans_chunk_committed() to process the items in
1294  * each chunk.
1295  */
1296 STATIC void
1297 xfs_trans_committed(
1298 	xfs_trans_t	*tp,
1299 	int		abortflag)
1300 {
1301 	xfs_log_item_chunk_t	*licp;
1302 	xfs_log_item_chunk_t	*next_licp;
1303 	xfs_log_busy_chunk_t	*lbcp;
1304 	xfs_log_busy_slot_t	*lbsp;
1305 	int			i;
1306 
1307 	/*
1308 	 * Call the transaction's completion callback if there
1309 	 * is one.
1310 	 */
1311 	if (tp->t_callback != NULL) {
1312 		tp->t_callback(tp, tp->t_callarg);
1313 	}
1314 
1315 	/*
1316 	 * Special case the chunk embedded in the transaction.
1317 	 */
1318 	licp = &(tp->t_items);
1319 	if (!(xfs_lic_are_all_free(licp))) {
1320 		xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
1321 	}
1322 
1323 	/*
1324 	 * Process the items in each chunk in turn.
1325 	 */
1326 	licp = licp->lic_next;
1327 	while (licp != NULL) {
1328 		ASSERT(!xfs_lic_are_all_free(licp));
1329 		xfs_trans_chunk_committed(licp, tp->t_lsn, abortflag);
1330 		next_licp = licp->lic_next;
1331 		kmem_free(licp);
1332 		licp = next_licp;
1333 	}
1334 
1335 	/*
1336 	 * Clear all the per-AG busy list items listed in this transaction
1337 	 */
1338 	lbcp = &tp->t_busy;
1339 	while (lbcp != NULL) {
1340 		for (i = 0, lbsp = lbcp->lbc_busy; i < lbcp->lbc_unused; i++, lbsp++) {
1341 			if (!XFS_LBC_ISFREE(lbcp, i)) {
1342 				xfs_alloc_clear_busy(tp, lbsp->lbc_ag,
1343 						     lbsp->lbc_idx);
1344 			}
1345 		}
1346 		lbcp = lbcp->lbc_next;
1347 	}
1348 	xfs_trans_free_busy(tp);
1349 
1350 	/*
1351 	 * That's it for the transaction structure.  Free it.
1352 	 */
1353 	xfs_trans_free(tp);
1354 }
1355 
1356 /*
1357  * This is called to perform the commit processing for each
1358  * item described by the given chunk.
1359  *
1360  * The commit processing consists of unlocking items which were
1361  * held locked with the SYNC_UNLOCK attribute, calling the committed
1362  * routine of each logged item, updating the item's position in the AIL
1363  * if necessary, and unpinning each item.  If the committed routine
1364  * returns -1, then do nothing further with the item because it
1365  * may have been freed.
1366  *
1367  * Since items are unlocked when they are copied to the incore
1368  * log, it is possible for two transactions to be completing
1369  * and manipulating the same item simultaneously.  The AIL lock
1370  * will protect the lsn field of each item.  The value of this
1371  * field can never go backwards.
1372  *
1373  * We unpin the items after repositioning them in the AIL, because
1374  * otherwise they could be immediately flushed and we'd have to race
1375  * with the flusher trying to pull the item from the AIL as we add it.
1376  */
1377 STATIC void
1378 xfs_trans_chunk_committed(
1379 	xfs_log_item_chunk_t	*licp,
1380 	xfs_lsn_t		lsn,
1381 	int			aborted)
1382 {
1383 	xfs_log_item_desc_t	*lidp;
1384 	xfs_log_item_t		*lip;
1385 	xfs_lsn_t		item_lsn;
1386 	struct xfs_mount	*mp;
1387 	int			i;
1388 
1389 	lidp = licp->lic_descs;
1390 	for (i = 0; i < licp->lic_unused; i++, lidp++) {
1391 		if (xfs_lic_isfree(licp, i)) {
1392 			continue;
1393 		}
1394 
1395 		lip = lidp->lid_item;
1396 		if (aborted)
1397 			lip->li_flags |= XFS_LI_ABORTED;
1398 
1399 		/*
1400 		 * Send in the ABORTED flag to the COMMITTED routine
1401 		 * so that it knows whether the transaction was aborted
1402 		 * or not.
1403 		 */
1404 		item_lsn = IOP_COMMITTED(lip, lsn);
1405 
1406 		/*
1407 		 * If the committed routine returns -1, make
1408 		 * no more references to the item.
1409 		 */
1410 		if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0) {
1411 			continue;
1412 		}
1413 
1414 		/*
1415 		 * If the returned lsn is greater than what it
1416 		 * contained before, update the location of the
1417 		 * item in the AIL.  If it is not, then do nothing.
1418 		 * Items can never move backwards in the AIL.
1419 		 *
1420 		 * While the new lsn should usually be greater, it
1421 		 * is possible that a later transaction completing
1422 		 * simultaneously with an earlier one using the
1423 		 * same item could complete first with a higher lsn.
1424 		 * This would cause the earlier transaction to fail
1425 		 * the test below.
1426 		 */
1427 		mp = lip->li_mountp;
1428 		spin_lock(&mp->m_ail_lock);
1429 		if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0) {
1430 			/*
1431 			 * This will set the item's lsn to item_lsn
1432 			 * and update the position of the item in
1433 			 * the AIL.
1434 			 *
1435 			 * xfs_trans_update_ail() drops the AIL lock.
1436 			 */
1437 			xfs_trans_update_ail(mp, lip, item_lsn);
1438 		} else {
1439 			spin_unlock(&mp->m_ail_lock);
1440 		}
1441 
1442 		/*
1443 		 * Now that we've repositioned the item in the AIL,
1444 		 * unpin it so it can be flushed. Pass information
1445 		 * about buffer stale state down from the log item
1446 		 * flags, if anyone else stales the buffer we do not
1447 		 * want to pay any attention to it.
1448 		 */
1449 		IOP_UNPIN(lip, lidp->lid_flags & XFS_LID_BUF_STALE);
1450 	}
1451 }
1452