xref: /openbmc/linux/fs/xfs/xfs_trans.c (revision 4800cd83)
1 /*
2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3  * Copyright (C) 2010 Red Hat, Inc.
4  * All Rights Reserved.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License as
8  * published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it would be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write the Free Software Foundation,
17  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
18  */
19 #include "xfs.h"
20 #include "xfs_fs.h"
21 #include "xfs_types.h"
22 #include "xfs_bit.h"
23 #include "xfs_log.h"
24 #include "xfs_inum.h"
25 #include "xfs_trans.h"
26 #include "xfs_sb.h"
27 #include "xfs_ag.h"
28 #include "xfs_mount.h"
29 #include "xfs_error.h"
30 #include "xfs_da_btree.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dinode.h"
35 #include "xfs_inode.h"
36 #include "xfs_btree.h"
37 #include "xfs_ialloc.h"
38 #include "xfs_alloc.h"
39 #include "xfs_bmap.h"
40 #include "xfs_quota.h"
41 #include "xfs_trans_priv.h"
42 #include "xfs_trans_space.h"
43 #include "xfs_inode_item.h"
44 #include "xfs_trace.h"
45 
46 kmem_zone_t	*xfs_trans_zone;
47 kmem_zone_t	*xfs_log_item_desc_zone;
48 
49 
50 /*
51  * Various log reservation values.
52  *
53  * These are based on the size of the file system block because that is what
54  * most transactions manipulate.  Each adds in an additional 128 bytes per
55  * item logged to try to account for the overhead of the transaction mechanism.
56  *
57  * Note:  Most of the reservations underestimate the number of allocation
58  * groups into which they could free extents in the xfs_bmap_finish() call.
59  * This is because the number in the worst case is quite high and quite
60  * unusual.  In order to fix this we need to change xfs_bmap_finish() to free
61  * extents in only a single AG at a time.  This will require changes to the
62  * EFI code as well, however, so that the EFI for the extents not freed is
63  * logged again in each transaction.  See SGI PV #261917.
64  *
65  * Reservation functions here avoid a huge stack in xfs_trans_init due to
66  * register overflow from temporaries in the calculations.
67  */
68 
69 
70 /*
71  * In a write transaction we can allocate a maximum of 2
72  * extents.  This gives:
73  *    the inode getting the new extents: inode size
74  *    the inode's bmap btree: max depth * block size
75  *    the agfs of the ags from which the extents are allocated: 2 * sector
76  *    the superblock free block counter: sector size
77  *    the allocation btrees: 2 exts * 2 trees * (2 * max depth - 1) * block size
78  * And the bmap_finish transaction can free bmap blocks in a join:
79  *    the agfs of the ags containing the blocks: 2 * sector size
80  *    the agfls of the ags containing the blocks: 2 * sector size
81  *    the super block free block counter: sector size
82  *    the allocation btrees: 2 exts * 2 trees * (2 * max depth - 1) * block size
83  */
84 STATIC uint
85 xfs_calc_write_reservation(
86 	struct xfs_mount	*mp)
87 {
88 	return XFS_DQUOT_LOGRES(mp) +
89 		MAX((mp->m_sb.sb_inodesize +
90 		     XFS_FSB_TO_B(mp, XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK)) +
91 		     2 * mp->m_sb.sb_sectsize +
92 		     mp->m_sb.sb_sectsize +
93 		     XFS_ALLOCFREE_LOG_RES(mp, 2) +
94 		     128 * (4 + XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK) +
95 			    XFS_ALLOCFREE_LOG_COUNT(mp, 2))),
96 		    (2 * mp->m_sb.sb_sectsize +
97 		     2 * mp->m_sb.sb_sectsize +
98 		     mp->m_sb.sb_sectsize +
99 		     XFS_ALLOCFREE_LOG_RES(mp, 2) +
100 		     128 * (5 + XFS_ALLOCFREE_LOG_COUNT(mp, 2))));
101 }
102 
103 /*
104  * In truncating a file we free up to two extents at once.  We can modify:
105  *    the inode being truncated: inode size
106  *    the inode's bmap btree: (max depth + 1) * block size
107  * And the bmap_finish transaction can free the blocks and bmap blocks:
108  *    the agf for each of the ags: 4 * sector size
109  *    the agfl for each of the ags: 4 * sector size
110  *    the super block to reflect the freed blocks: sector size
111  *    worst case split in allocation btrees per extent assuming 4 extents:
112  *		4 exts * 2 trees * (2 * max depth - 1) * block size
113  *    the inode btree: max depth * blocksize
114  *    the allocation btrees: 2 trees * (max depth - 1) * block size
115  */
116 STATIC uint
117 xfs_calc_itruncate_reservation(
118 	struct xfs_mount	*mp)
119 {
120 	return XFS_DQUOT_LOGRES(mp) +
121 		MAX((mp->m_sb.sb_inodesize +
122 		     XFS_FSB_TO_B(mp, XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK) + 1) +
123 		     128 * (2 + XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK))),
124 		    (4 * mp->m_sb.sb_sectsize +
125 		     4 * mp->m_sb.sb_sectsize +
126 		     mp->m_sb.sb_sectsize +
127 		     XFS_ALLOCFREE_LOG_RES(mp, 4) +
128 		     128 * (9 + XFS_ALLOCFREE_LOG_COUNT(mp, 4)) +
129 		     128 * 5 +
130 		     XFS_ALLOCFREE_LOG_RES(mp, 1) +
131 		     128 * (2 + XFS_IALLOC_BLOCKS(mp) + mp->m_in_maxlevels +
132 			    XFS_ALLOCFREE_LOG_COUNT(mp, 1))));
133 }
134 
135 /*
136  * In renaming a files we can modify:
137  *    the four inodes involved: 4 * inode size
138  *    the two directory btrees: 2 * (max depth + v2) * dir block size
139  *    the two directory bmap btrees: 2 * max depth * block size
140  * And the bmap_finish transaction can free dir and bmap blocks (two sets
141  *	of bmap blocks) giving:
142  *    the agf for the ags in which the blocks live: 3 * sector size
143  *    the agfl for the ags in which the blocks live: 3 * sector size
144  *    the superblock for the free block count: sector size
145  *    the allocation btrees: 3 exts * 2 trees * (2 * max depth - 1) * block size
146  */
147 STATIC uint
148 xfs_calc_rename_reservation(
149 	struct xfs_mount	*mp)
150 {
151 	return XFS_DQUOT_LOGRES(mp) +
152 		MAX((4 * mp->m_sb.sb_inodesize +
153 		     2 * XFS_DIROP_LOG_RES(mp) +
154 		     128 * (4 + 2 * XFS_DIROP_LOG_COUNT(mp))),
155 		    (3 * mp->m_sb.sb_sectsize +
156 		     3 * mp->m_sb.sb_sectsize +
157 		     mp->m_sb.sb_sectsize +
158 		     XFS_ALLOCFREE_LOG_RES(mp, 3) +
159 		     128 * (7 + XFS_ALLOCFREE_LOG_COUNT(mp, 3))));
160 }
161 
162 /*
163  * For creating a link to an inode:
164  *    the parent directory inode: inode size
165  *    the linked inode: inode size
166  *    the directory btree could split: (max depth + v2) * dir block size
167  *    the directory bmap btree could join or split: (max depth + v2) * blocksize
168  * And the bmap_finish transaction can free some bmap blocks giving:
169  *    the agf for the ag in which the blocks live: sector size
170  *    the agfl for the ag in which the blocks live: sector size
171  *    the superblock for the free block count: sector size
172  *    the allocation btrees: 2 trees * (2 * max depth - 1) * block size
173  */
174 STATIC uint
175 xfs_calc_link_reservation(
176 	struct xfs_mount	*mp)
177 {
178 	return XFS_DQUOT_LOGRES(mp) +
179 		MAX((mp->m_sb.sb_inodesize +
180 		     mp->m_sb.sb_inodesize +
181 		     XFS_DIROP_LOG_RES(mp) +
182 		     128 * (2 + XFS_DIROP_LOG_COUNT(mp))),
183 		    (mp->m_sb.sb_sectsize +
184 		     mp->m_sb.sb_sectsize +
185 		     mp->m_sb.sb_sectsize +
186 		     XFS_ALLOCFREE_LOG_RES(mp, 1) +
187 		     128 * (3 + XFS_ALLOCFREE_LOG_COUNT(mp, 1))));
188 }
189 
190 /*
191  * For removing a directory entry we can modify:
192  *    the parent directory inode: inode size
193  *    the removed inode: inode size
194  *    the directory btree could join: (max depth + v2) * dir block size
195  *    the directory bmap btree could join or split: (max depth + v2) * blocksize
196  * And the bmap_finish transaction can free the dir and bmap blocks giving:
197  *    the agf for the ag in which the blocks live: 2 * sector size
198  *    the agfl for the ag in which the blocks live: 2 * sector size
199  *    the superblock for the free block count: sector size
200  *    the allocation btrees: 2 exts * 2 trees * (2 * max depth - 1) * block size
201  */
202 STATIC uint
203 xfs_calc_remove_reservation(
204 	struct xfs_mount	*mp)
205 {
206 	return XFS_DQUOT_LOGRES(mp) +
207 		MAX((mp->m_sb.sb_inodesize +
208 		     mp->m_sb.sb_inodesize +
209 		     XFS_DIROP_LOG_RES(mp) +
210 		     128 * (2 + XFS_DIROP_LOG_COUNT(mp))),
211 		    (2 * mp->m_sb.sb_sectsize +
212 		     2 * mp->m_sb.sb_sectsize +
213 		     mp->m_sb.sb_sectsize +
214 		     XFS_ALLOCFREE_LOG_RES(mp, 2) +
215 		     128 * (5 + XFS_ALLOCFREE_LOG_COUNT(mp, 2))));
216 }
217 
218 /*
219  * For symlink we can modify:
220  *    the parent directory inode: inode size
221  *    the new inode: inode size
222  *    the inode btree entry: 1 block
223  *    the directory btree: (max depth + v2) * dir block size
224  *    the directory inode's bmap btree: (max depth + v2) * block size
225  *    the blocks for the symlink: 1 kB
226  * Or in the first xact we allocate some inodes giving:
227  *    the agi and agf of the ag getting the new inodes: 2 * sectorsize
228  *    the inode blocks allocated: XFS_IALLOC_BLOCKS * blocksize
229  *    the inode btree: max depth * blocksize
230  *    the allocation btrees: 2 trees * (2 * max depth - 1) * block size
231  */
232 STATIC uint
233 xfs_calc_symlink_reservation(
234 	struct xfs_mount	*mp)
235 {
236 	return XFS_DQUOT_LOGRES(mp) +
237 		MAX((mp->m_sb.sb_inodesize +
238 		     mp->m_sb.sb_inodesize +
239 		     XFS_FSB_TO_B(mp, 1) +
240 		     XFS_DIROP_LOG_RES(mp) +
241 		     1024 +
242 		     128 * (4 + XFS_DIROP_LOG_COUNT(mp))),
243 		    (2 * mp->m_sb.sb_sectsize +
244 		     XFS_FSB_TO_B(mp, XFS_IALLOC_BLOCKS(mp)) +
245 		     XFS_FSB_TO_B(mp, mp->m_in_maxlevels) +
246 		     XFS_ALLOCFREE_LOG_RES(mp, 1) +
247 		     128 * (2 + XFS_IALLOC_BLOCKS(mp) + mp->m_in_maxlevels +
248 			    XFS_ALLOCFREE_LOG_COUNT(mp, 1))));
249 }
250 
251 /*
252  * For create we can modify:
253  *    the parent directory inode: inode size
254  *    the new inode: inode size
255  *    the inode btree entry: block size
256  *    the superblock for the nlink flag: sector size
257  *    the directory btree: (max depth + v2) * dir block size
258  *    the directory inode's bmap btree: (max depth + v2) * block size
259  * Or in the first xact we allocate some inodes giving:
260  *    the agi and agf of the ag getting the new inodes: 2 * sectorsize
261  *    the superblock for the nlink flag: sector size
262  *    the inode blocks allocated: XFS_IALLOC_BLOCKS * blocksize
263  *    the inode btree: max depth * blocksize
264  *    the allocation btrees: 2 trees * (max depth - 1) * block size
265  */
266 STATIC uint
267 xfs_calc_create_reservation(
268 	struct xfs_mount	*mp)
269 {
270 	return XFS_DQUOT_LOGRES(mp) +
271 		MAX((mp->m_sb.sb_inodesize +
272 		     mp->m_sb.sb_inodesize +
273 		     mp->m_sb.sb_sectsize +
274 		     XFS_FSB_TO_B(mp, 1) +
275 		     XFS_DIROP_LOG_RES(mp) +
276 		     128 * (3 + XFS_DIROP_LOG_COUNT(mp))),
277 		    (3 * mp->m_sb.sb_sectsize +
278 		     XFS_FSB_TO_B(mp, XFS_IALLOC_BLOCKS(mp)) +
279 		     XFS_FSB_TO_B(mp, mp->m_in_maxlevels) +
280 		     XFS_ALLOCFREE_LOG_RES(mp, 1) +
281 		     128 * (2 + XFS_IALLOC_BLOCKS(mp) + mp->m_in_maxlevels +
282 			    XFS_ALLOCFREE_LOG_COUNT(mp, 1))));
283 }
284 
285 /*
286  * Making a new directory is the same as creating a new file.
287  */
288 STATIC uint
289 xfs_calc_mkdir_reservation(
290 	struct xfs_mount	*mp)
291 {
292 	return xfs_calc_create_reservation(mp);
293 }
294 
295 /*
296  * In freeing an inode we can modify:
297  *    the inode being freed: inode size
298  *    the super block free inode counter: sector size
299  *    the agi hash list and counters: sector size
300  *    the inode btree entry: block size
301  *    the on disk inode before ours in the agi hash list: inode cluster size
302  *    the inode btree: max depth * blocksize
303  *    the allocation btrees: 2 trees * (max depth - 1) * block size
304  */
305 STATIC uint
306 xfs_calc_ifree_reservation(
307 	struct xfs_mount	*mp)
308 {
309 	return XFS_DQUOT_LOGRES(mp) +
310 		mp->m_sb.sb_inodesize +
311 		mp->m_sb.sb_sectsize +
312 		mp->m_sb.sb_sectsize +
313 		XFS_FSB_TO_B(mp, 1) +
314 		MAX((__uint16_t)XFS_FSB_TO_B(mp, 1),
315 		    XFS_INODE_CLUSTER_SIZE(mp)) +
316 		128 * 5 +
317 		XFS_ALLOCFREE_LOG_RES(mp, 1) +
318 		128 * (2 + XFS_IALLOC_BLOCKS(mp) + mp->m_in_maxlevels +
319 		       XFS_ALLOCFREE_LOG_COUNT(mp, 1));
320 }
321 
322 /*
323  * When only changing the inode we log the inode and possibly the superblock
324  * We also add a bit of slop for the transaction stuff.
325  */
326 STATIC uint
327 xfs_calc_ichange_reservation(
328 	struct xfs_mount	*mp)
329 {
330 	return XFS_DQUOT_LOGRES(mp) +
331 		mp->m_sb.sb_inodesize +
332 		mp->m_sb.sb_sectsize +
333 		512;
334 
335 }
336 
337 /*
338  * Growing the data section of the filesystem.
339  *	superblock
340  *	agi and agf
341  *	allocation btrees
342  */
343 STATIC uint
344 xfs_calc_growdata_reservation(
345 	struct xfs_mount	*mp)
346 {
347 	return mp->m_sb.sb_sectsize * 3 +
348 		XFS_ALLOCFREE_LOG_RES(mp, 1) +
349 		128 * (3 + XFS_ALLOCFREE_LOG_COUNT(mp, 1));
350 }
351 
352 /*
353  * Growing the rt section of the filesystem.
354  * In the first set of transactions (ALLOC) we allocate space to the
355  * bitmap or summary files.
356  *	superblock: sector size
357  *	agf of the ag from which the extent is allocated: sector size
358  *	bmap btree for bitmap/summary inode: max depth * blocksize
359  *	bitmap/summary inode: inode size
360  *	allocation btrees for 1 block alloc: 2 * (2 * maxdepth - 1) * blocksize
361  */
362 STATIC uint
363 xfs_calc_growrtalloc_reservation(
364 	struct xfs_mount	*mp)
365 {
366 	return 2 * mp->m_sb.sb_sectsize +
367 		XFS_FSB_TO_B(mp, XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK)) +
368 		mp->m_sb.sb_inodesize +
369 		XFS_ALLOCFREE_LOG_RES(mp, 1) +
370 		128 * (3 + XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK) +
371 		       XFS_ALLOCFREE_LOG_COUNT(mp, 1));
372 }
373 
374 /*
375  * Growing the rt section of the filesystem.
376  * In the second set of transactions (ZERO) we zero the new metadata blocks.
377  *	one bitmap/summary block: blocksize
378  */
379 STATIC uint
380 xfs_calc_growrtzero_reservation(
381 	struct xfs_mount	*mp)
382 {
383 	return mp->m_sb.sb_blocksize + 128;
384 }
385 
386 /*
387  * Growing the rt section of the filesystem.
388  * In the third set of transactions (FREE) we update metadata without
389  * allocating any new blocks.
390  *	superblock: sector size
391  *	bitmap inode: inode size
392  *	summary inode: inode size
393  *	one bitmap block: blocksize
394  *	summary blocks: new summary size
395  */
396 STATIC uint
397 xfs_calc_growrtfree_reservation(
398 	struct xfs_mount	*mp)
399 {
400 	return mp->m_sb.sb_sectsize +
401 		2 * mp->m_sb.sb_inodesize +
402 		mp->m_sb.sb_blocksize +
403 		mp->m_rsumsize +
404 		128 * 5;
405 }
406 
407 /*
408  * Logging the inode modification timestamp on a synchronous write.
409  *	inode
410  */
411 STATIC uint
412 xfs_calc_swrite_reservation(
413 	struct xfs_mount	*mp)
414 {
415 	return mp->m_sb.sb_inodesize + 128;
416 }
417 
418 /*
419  * Logging the inode mode bits when writing a setuid/setgid file
420  *	inode
421  */
422 STATIC uint
423 xfs_calc_writeid_reservation(xfs_mount_t *mp)
424 {
425 	return mp->m_sb.sb_inodesize + 128;
426 }
427 
428 /*
429  * Converting the inode from non-attributed to attributed.
430  *	the inode being converted: inode size
431  *	agf block and superblock (for block allocation)
432  *	the new block (directory sized)
433  *	bmap blocks for the new directory block
434  *	allocation btrees
435  */
436 STATIC uint
437 xfs_calc_addafork_reservation(
438 	struct xfs_mount	*mp)
439 {
440 	return XFS_DQUOT_LOGRES(mp) +
441 		mp->m_sb.sb_inodesize +
442 		mp->m_sb.sb_sectsize * 2 +
443 		mp->m_dirblksize +
444 		XFS_FSB_TO_B(mp, XFS_DAENTER_BMAP1B(mp, XFS_DATA_FORK) + 1) +
445 		XFS_ALLOCFREE_LOG_RES(mp, 1) +
446 		128 * (4 + XFS_DAENTER_BMAP1B(mp, XFS_DATA_FORK) + 1 +
447 		       XFS_ALLOCFREE_LOG_COUNT(mp, 1));
448 }
449 
450 /*
451  * Removing the attribute fork of a file
452  *    the inode being truncated: inode size
453  *    the inode's bmap btree: max depth * block size
454  * And the bmap_finish transaction can free the blocks and bmap blocks:
455  *    the agf for each of the ags: 4 * sector size
456  *    the agfl for each of the ags: 4 * sector size
457  *    the super block to reflect the freed blocks: sector size
458  *    worst case split in allocation btrees per extent assuming 4 extents:
459  *		4 exts * 2 trees * (2 * max depth - 1) * block size
460  */
461 STATIC uint
462 xfs_calc_attrinval_reservation(
463 	struct xfs_mount	*mp)
464 {
465 	return MAX((mp->m_sb.sb_inodesize +
466 		    XFS_FSB_TO_B(mp, XFS_BM_MAXLEVELS(mp, XFS_ATTR_FORK)) +
467 		    128 * (1 + XFS_BM_MAXLEVELS(mp, XFS_ATTR_FORK))),
468 		   (4 * mp->m_sb.sb_sectsize +
469 		    4 * mp->m_sb.sb_sectsize +
470 		    mp->m_sb.sb_sectsize +
471 		    XFS_ALLOCFREE_LOG_RES(mp, 4) +
472 		    128 * (9 + XFS_ALLOCFREE_LOG_COUNT(mp, 4))));
473 }
474 
475 /*
476  * Setting an attribute.
477  *	the inode getting the attribute
478  *	the superblock for allocations
479  *	the agfs extents are allocated from
480  *	the attribute btree * max depth
481  *	the inode allocation btree
482  * Since attribute transaction space is dependent on the size of the attribute,
483  * the calculation is done partially at mount time and partially at runtime.
484  */
485 STATIC uint
486 xfs_calc_attrset_reservation(
487 	struct xfs_mount	*mp)
488 {
489 	return XFS_DQUOT_LOGRES(mp) +
490 		mp->m_sb.sb_inodesize +
491 		mp->m_sb.sb_sectsize +
492 		XFS_FSB_TO_B(mp, XFS_DA_NODE_MAXDEPTH) +
493 		128 * (2 + XFS_DA_NODE_MAXDEPTH);
494 }
495 
496 /*
497  * Removing an attribute.
498  *    the inode: inode size
499  *    the attribute btree could join: max depth * block size
500  *    the inode bmap btree could join or split: max depth * block size
501  * And the bmap_finish transaction can free the attr blocks freed giving:
502  *    the agf for the ag in which the blocks live: 2 * sector size
503  *    the agfl for the ag in which the blocks live: 2 * sector size
504  *    the superblock for the free block count: sector size
505  *    the allocation btrees: 2 exts * 2 trees * (2 * max depth - 1) * block size
506  */
507 STATIC uint
508 xfs_calc_attrrm_reservation(
509 	struct xfs_mount	*mp)
510 {
511 	return XFS_DQUOT_LOGRES(mp) +
512 		MAX((mp->m_sb.sb_inodesize +
513 		     XFS_FSB_TO_B(mp, XFS_DA_NODE_MAXDEPTH) +
514 		     XFS_FSB_TO_B(mp, XFS_BM_MAXLEVELS(mp, XFS_ATTR_FORK)) +
515 		     128 * (1 + XFS_DA_NODE_MAXDEPTH +
516 			    XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK))),
517 		    (2 * mp->m_sb.sb_sectsize +
518 		     2 * mp->m_sb.sb_sectsize +
519 		     mp->m_sb.sb_sectsize +
520 		     XFS_ALLOCFREE_LOG_RES(mp, 2) +
521 		     128 * (5 + XFS_ALLOCFREE_LOG_COUNT(mp, 2))));
522 }
523 
524 /*
525  * Clearing a bad agino number in an agi hash bucket.
526  */
527 STATIC uint
528 xfs_calc_clear_agi_bucket_reservation(
529 	struct xfs_mount	*mp)
530 {
531 	return mp->m_sb.sb_sectsize + 128;
532 }
533 
534 /*
535  * Initialize the precomputed transaction reservation values
536  * in the mount structure.
537  */
538 void
539 xfs_trans_init(
540 	struct xfs_mount	*mp)
541 {
542 	struct xfs_trans_reservations *resp = &mp->m_reservations;
543 
544 	resp->tr_write = xfs_calc_write_reservation(mp);
545 	resp->tr_itruncate = xfs_calc_itruncate_reservation(mp);
546 	resp->tr_rename = xfs_calc_rename_reservation(mp);
547 	resp->tr_link = xfs_calc_link_reservation(mp);
548 	resp->tr_remove = xfs_calc_remove_reservation(mp);
549 	resp->tr_symlink = xfs_calc_symlink_reservation(mp);
550 	resp->tr_create = xfs_calc_create_reservation(mp);
551 	resp->tr_mkdir = xfs_calc_mkdir_reservation(mp);
552 	resp->tr_ifree = xfs_calc_ifree_reservation(mp);
553 	resp->tr_ichange = xfs_calc_ichange_reservation(mp);
554 	resp->tr_growdata = xfs_calc_growdata_reservation(mp);
555 	resp->tr_swrite = xfs_calc_swrite_reservation(mp);
556 	resp->tr_writeid = xfs_calc_writeid_reservation(mp);
557 	resp->tr_addafork = xfs_calc_addafork_reservation(mp);
558 	resp->tr_attrinval = xfs_calc_attrinval_reservation(mp);
559 	resp->tr_attrset = xfs_calc_attrset_reservation(mp);
560 	resp->tr_attrrm = xfs_calc_attrrm_reservation(mp);
561 	resp->tr_clearagi = xfs_calc_clear_agi_bucket_reservation(mp);
562 	resp->tr_growrtalloc = xfs_calc_growrtalloc_reservation(mp);
563 	resp->tr_growrtzero = xfs_calc_growrtzero_reservation(mp);
564 	resp->tr_growrtfree = xfs_calc_growrtfree_reservation(mp);
565 }
566 
567 /*
568  * This routine is called to allocate a transaction structure.
569  * The type parameter indicates the type of the transaction.  These
570  * are enumerated in xfs_trans.h.
571  *
572  * Dynamically allocate the transaction structure from the transaction
573  * zone, initialize it, and return it to the caller.
574  */
575 xfs_trans_t *
576 xfs_trans_alloc(
577 	xfs_mount_t	*mp,
578 	uint		type)
579 {
580 	xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
581 	return _xfs_trans_alloc(mp, type, KM_SLEEP);
582 }
583 
584 xfs_trans_t *
585 _xfs_trans_alloc(
586 	xfs_mount_t	*mp,
587 	uint		type,
588 	uint		memflags)
589 {
590 	xfs_trans_t	*tp;
591 
592 	atomic_inc(&mp->m_active_trans);
593 
594 	tp = kmem_zone_zalloc(xfs_trans_zone, memflags);
595 	tp->t_magic = XFS_TRANS_MAGIC;
596 	tp->t_type = type;
597 	tp->t_mountp = mp;
598 	INIT_LIST_HEAD(&tp->t_items);
599 	INIT_LIST_HEAD(&tp->t_busy);
600 	return tp;
601 }
602 
603 /*
604  * Free the transaction structure.  If there is more clean up
605  * to do when the structure is freed, add it here.
606  */
607 STATIC void
608 xfs_trans_free(
609 	struct xfs_trans	*tp)
610 {
611 	struct xfs_busy_extent	*busyp, *n;
612 
613 	list_for_each_entry_safe(busyp, n, &tp->t_busy, list)
614 		xfs_alloc_busy_clear(tp->t_mountp, busyp);
615 
616 	atomic_dec(&tp->t_mountp->m_active_trans);
617 	xfs_trans_free_dqinfo(tp);
618 	kmem_zone_free(xfs_trans_zone, tp);
619 }
620 
621 /*
622  * This is called to create a new transaction which will share the
623  * permanent log reservation of the given transaction.  The remaining
624  * unused block and rt extent reservations are also inherited.  This
625  * implies that the original transaction is no longer allowed to allocate
626  * blocks.  Locks and log items, however, are no inherited.  They must
627  * be added to the new transaction explicitly.
628  */
629 xfs_trans_t *
630 xfs_trans_dup(
631 	xfs_trans_t	*tp)
632 {
633 	xfs_trans_t	*ntp;
634 
635 	ntp = kmem_zone_zalloc(xfs_trans_zone, KM_SLEEP);
636 
637 	/*
638 	 * Initialize the new transaction structure.
639 	 */
640 	ntp->t_magic = XFS_TRANS_MAGIC;
641 	ntp->t_type = tp->t_type;
642 	ntp->t_mountp = tp->t_mountp;
643 	INIT_LIST_HEAD(&ntp->t_items);
644 	INIT_LIST_HEAD(&ntp->t_busy);
645 
646 	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
647 	ASSERT(tp->t_ticket != NULL);
648 
649 	ntp->t_flags = XFS_TRANS_PERM_LOG_RES | (tp->t_flags & XFS_TRANS_RESERVE);
650 	ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
651 	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
652 	tp->t_blk_res = tp->t_blk_res_used;
653 	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
654 	tp->t_rtx_res = tp->t_rtx_res_used;
655 	ntp->t_pflags = tp->t_pflags;
656 
657 	xfs_trans_dup_dqinfo(tp, ntp);
658 
659 	atomic_inc(&tp->t_mountp->m_active_trans);
660 	return ntp;
661 }
662 
663 /*
664  * This is called to reserve free disk blocks and log space for the
665  * given transaction.  This must be done before allocating any resources
666  * within the transaction.
667  *
668  * This will return ENOSPC if there are not enough blocks available.
669  * It will sleep waiting for available log space.
670  * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which
671  * is used by long running transactions.  If any one of the reservations
672  * fails then they will all be backed out.
673  *
674  * This does not do quota reservations. That typically is done by the
675  * caller afterwards.
676  */
677 int
678 xfs_trans_reserve(
679 	xfs_trans_t	*tp,
680 	uint		blocks,
681 	uint		logspace,
682 	uint		rtextents,
683 	uint		flags,
684 	uint		logcount)
685 {
686 	int		log_flags;
687 	int		error = 0;
688 	int		rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
689 
690 	/* Mark this thread as being in a transaction */
691 	current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
692 
693 	/*
694 	 * Attempt to reserve the needed disk blocks by decrementing
695 	 * the number needed from the number available.  This will
696 	 * fail if the count would go below zero.
697 	 */
698 	if (blocks > 0) {
699 		error = xfs_icsb_modify_counters(tp->t_mountp, XFS_SBS_FDBLOCKS,
700 					  -((int64_t)blocks), rsvd);
701 		if (error != 0) {
702 			current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
703 			return (XFS_ERROR(ENOSPC));
704 		}
705 		tp->t_blk_res += blocks;
706 	}
707 
708 	/*
709 	 * Reserve the log space needed for this transaction.
710 	 */
711 	if (logspace > 0) {
712 		ASSERT((tp->t_log_res == 0) || (tp->t_log_res == logspace));
713 		ASSERT((tp->t_log_count == 0) ||
714 			(tp->t_log_count == logcount));
715 		if (flags & XFS_TRANS_PERM_LOG_RES) {
716 			log_flags = XFS_LOG_PERM_RESERV;
717 			tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
718 		} else {
719 			ASSERT(tp->t_ticket == NULL);
720 			ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
721 			log_flags = 0;
722 		}
723 
724 		error = xfs_log_reserve(tp->t_mountp, logspace, logcount,
725 					&tp->t_ticket,
726 					XFS_TRANSACTION, log_flags, tp->t_type);
727 		if (error) {
728 			goto undo_blocks;
729 		}
730 		tp->t_log_res = logspace;
731 		tp->t_log_count = logcount;
732 	}
733 
734 	/*
735 	 * Attempt to reserve the needed realtime extents by decrementing
736 	 * the number needed from the number available.  This will
737 	 * fail if the count would go below zero.
738 	 */
739 	if (rtextents > 0) {
740 		error = xfs_mod_incore_sb(tp->t_mountp, XFS_SBS_FREXTENTS,
741 					  -((int64_t)rtextents), rsvd);
742 		if (error) {
743 			error = XFS_ERROR(ENOSPC);
744 			goto undo_log;
745 		}
746 		tp->t_rtx_res += rtextents;
747 	}
748 
749 	return 0;
750 
751 	/*
752 	 * Error cases jump to one of these labels to undo any
753 	 * reservations which have already been performed.
754 	 */
755 undo_log:
756 	if (logspace > 0) {
757 		if (flags & XFS_TRANS_PERM_LOG_RES) {
758 			log_flags = XFS_LOG_REL_PERM_RESERV;
759 		} else {
760 			log_flags = 0;
761 		}
762 		xfs_log_done(tp->t_mountp, tp->t_ticket, NULL, log_flags);
763 		tp->t_ticket = NULL;
764 		tp->t_log_res = 0;
765 		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
766 	}
767 
768 undo_blocks:
769 	if (blocks > 0) {
770 		xfs_icsb_modify_counters(tp->t_mountp, XFS_SBS_FDBLOCKS,
771 					 (int64_t)blocks, rsvd);
772 		tp->t_blk_res = 0;
773 	}
774 
775 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
776 
777 	return error;
778 }
779 
780 /*
781  * Record the indicated change to the given field for application
782  * to the file system's superblock when the transaction commits.
783  * For now, just store the change in the transaction structure.
784  *
785  * Mark the transaction structure to indicate that the superblock
786  * needs to be updated before committing.
787  *
788  * Because we may not be keeping track of allocated/free inodes and
789  * used filesystem blocks in the superblock, we do not mark the
790  * superblock dirty in this transaction if we modify these fields.
791  * We still need to update the transaction deltas so that they get
792  * applied to the incore superblock, but we don't want them to
793  * cause the superblock to get locked and logged if these are the
794  * only fields in the superblock that the transaction modifies.
795  */
796 void
797 xfs_trans_mod_sb(
798 	xfs_trans_t	*tp,
799 	uint		field,
800 	int64_t		delta)
801 {
802 	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
803 	xfs_mount_t	*mp = tp->t_mountp;
804 
805 	switch (field) {
806 	case XFS_TRANS_SB_ICOUNT:
807 		tp->t_icount_delta += delta;
808 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
809 			flags &= ~XFS_TRANS_SB_DIRTY;
810 		break;
811 	case XFS_TRANS_SB_IFREE:
812 		tp->t_ifree_delta += delta;
813 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
814 			flags &= ~XFS_TRANS_SB_DIRTY;
815 		break;
816 	case XFS_TRANS_SB_FDBLOCKS:
817 		/*
818 		 * Track the number of blocks allocated in the
819 		 * transaction.  Make sure it does not exceed the
820 		 * number reserved.
821 		 */
822 		if (delta < 0) {
823 			tp->t_blk_res_used += (uint)-delta;
824 			ASSERT(tp->t_blk_res_used <= tp->t_blk_res);
825 		}
826 		tp->t_fdblocks_delta += delta;
827 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
828 			flags &= ~XFS_TRANS_SB_DIRTY;
829 		break;
830 	case XFS_TRANS_SB_RES_FDBLOCKS:
831 		/*
832 		 * The allocation has already been applied to the
833 		 * in-core superblock's counter.  This should only
834 		 * be applied to the on-disk superblock.
835 		 */
836 		ASSERT(delta < 0);
837 		tp->t_res_fdblocks_delta += delta;
838 		if (xfs_sb_version_haslazysbcount(&mp->m_sb))
839 			flags &= ~XFS_TRANS_SB_DIRTY;
840 		break;
841 	case XFS_TRANS_SB_FREXTENTS:
842 		/*
843 		 * Track the number of blocks allocated in the
844 		 * transaction.  Make sure it does not exceed the
845 		 * number reserved.
846 		 */
847 		if (delta < 0) {
848 			tp->t_rtx_res_used += (uint)-delta;
849 			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
850 		}
851 		tp->t_frextents_delta += delta;
852 		break;
853 	case XFS_TRANS_SB_RES_FREXTENTS:
854 		/*
855 		 * The allocation has already been applied to the
856 		 * in-core superblock's counter.  This should only
857 		 * be applied to the on-disk superblock.
858 		 */
859 		ASSERT(delta < 0);
860 		tp->t_res_frextents_delta += delta;
861 		break;
862 	case XFS_TRANS_SB_DBLOCKS:
863 		ASSERT(delta > 0);
864 		tp->t_dblocks_delta += delta;
865 		break;
866 	case XFS_TRANS_SB_AGCOUNT:
867 		ASSERT(delta > 0);
868 		tp->t_agcount_delta += delta;
869 		break;
870 	case XFS_TRANS_SB_IMAXPCT:
871 		tp->t_imaxpct_delta += delta;
872 		break;
873 	case XFS_TRANS_SB_REXTSIZE:
874 		tp->t_rextsize_delta += delta;
875 		break;
876 	case XFS_TRANS_SB_RBMBLOCKS:
877 		tp->t_rbmblocks_delta += delta;
878 		break;
879 	case XFS_TRANS_SB_RBLOCKS:
880 		tp->t_rblocks_delta += delta;
881 		break;
882 	case XFS_TRANS_SB_REXTENTS:
883 		tp->t_rextents_delta += delta;
884 		break;
885 	case XFS_TRANS_SB_REXTSLOG:
886 		tp->t_rextslog_delta += delta;
887 		break;
888 	default:
889 		ASSERT(0);
890 		return;
891 	}
892 
893 	tp->t_flags |= flags;
894 }
895 
896 /*
897  * xfs_trans_apply_sb_deltas() is called from the commit code
898  * to bring the superblock buffer into the current transaction
899  * and modify it as requested by earlier calls to xfs_trans_mod_sb().
900  *
901  * For now we just look at each field allowed to change and change
902  * it if necessary.
903  */
904 STATIC void
905 xfs_trans_apply_sb_deltas(
906 	xfs_trans_t	*tp)
907 {
908 	xfs_dsb_t	*sbp;
909 	xfs_buf_t	*bp;
910 	int		whole = 0;
911 
912 	bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
913 	sbp = XFS_BUF_TO_SBP(bp);
914 
915 	/*
916 	 * Check that superblock mods match the mods made to AGF counters.
917 	 */
918 	ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
919 	       (tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
920 		tp->t_ag_btree_delta));
921 
922 	/*
923 	 * Only update the superblock counters if we are logging them
924 	 */
925 	if (!xfs_sb_version_haslazysbcount(&(tp->t_mountp->m_sb))) {
926 		if (tp->t_icount_delta)
927 			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
928 		if (tp->t_ifree_delta)
929 			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
930 		if (tp->t_fdblocks_delta)
931 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
932 		if (tp->t_res_fdblocks_delta)
933 			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
934 	}
935 
936 	if (tp->t_frextents_delta)
937 		be64_add_cpu(&sbp->sb_frextents, tp->t_frextents_delta);
938 	if (tp->t_res_frextents_delta)
939 		be64_add_cpu(&sbp->sb_frextents, tp->t_res_frextents_delta);
940 
941 	if (tp->t_dblocks_delta) {
942 		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
943 		whole = 1;
944 	}
945 	if (tp->t_agcount_delta) {
946 		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
947 		whole = 1;
948 	}
949 	if (tp->t_imaxpct_delta) {
950 		sbp->sb_imax_pct += tp->t_imaxpct_delta;
951 		whole = 1;
952 	}
953 	if (tp->t_rextsize_delta) {
954 		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
955 		whole = 1;
956 	}
957 	if (tp->t_rbmblocks_delta) {
958 		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
959 		whole = 1;
960 	}
961 	if (tp->t_rblocks_delta) {
962 		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
963 		whole = 1;
964 	}
965 	if (tp->t_rextents_delta) {
966 		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
967 		whole = 1;
968 	}
969 	if (tp->t_rextslog_delta) {
970 		sbp->sb_rextslog += tp->t_rextslog_delta;
971 		whole = 1;
972 	}
973 
974 	if (whole)
975 		/*
976 		 * Log the whole thing, the fields are noncontiguous.
977 		 */
978 		xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_dsb_t) - 1);
979 	else
980 		/*
981 		 * Since all the modifiable fields are contiguous, we
982 		 * can get away with this.
983 		 */
984 		xfs_trans_log_buf(tp, bp, offsetof(xfs_dsb_t, sb_icount),
985 				  offsetof(xfs_dsb_t, sb_frextents) +
986 				  sizeof(sbp->sb_frextents) - 1);
987 }
988 
989 /*
990  * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations
991  * and apply superblock counter changes to the in-core superblock.  The
992  * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
993  * applied to the in-core superblock.  The idea is that that has already been
994  * done.
995  *
996  * This is done efficiently with a single call to xfs_mod_incore_sb_batch().
997  * However, we have to ensure that we only modify each superblock field only
998  * once because the application of the delta values may not be atomic. That can
999  * lead to ENOSPC races occurring if we have two separate modifcations of the
1000  * free space counter to put back the entire reservation and then take away
1001  * what we used.
1002  *
1003  * If we are not logging superblock counters, then the inode allocated/free and
1004  * used block counts are not updated in the on disk superblock. In this case,
1005  * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
1006  * still need to update the incore superblock with the changes.
1007  */
1008 void
1009 xfs_trans_unreserve_and_mod_sb(
1010 	xfs_trans_t	*tp)
1011 {
1012 	xfs_mod_sb_t	msb[9];	/* If you add cases, add entries */
1013 	xfs_mod_sb_t	*msbp;
1014 	xfs_mount_t	*mp = tp->t_mountp;
1015 	/* REFERENCED */
1016 	int		error;
1017 	int		rsvd;
1018 	int64_t		blkdelta = 0;
1019 	int64_t		rtxdelta = 0;
1020 	int64_t		idelta = 0;
1021 	int64_t		ifreedelta = 0;
1022 
1023 	msbp = msb;
1024 	rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
1025 
1026 	/* calculate deltas */
1027 	if (tp->t_blk_res > 0)
1028 		blkdelta = tp->t_blk_res;
1029 	if ((tp->t_fdblocks_delta != 0) &&
1030 	    (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
1031 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)))
1032 	        blkdelta += tp->t_fdblocks_delta;
1033 
1034 	if (tp->t_rtx_res > 0)
1035 		rtxdelta = tp->t_rtx_res;
1036 	if ((tp->t_frextents_delta != 0) &&
1037 	    (tp->t_flags & XFS_TRANS_SB_DIRTY))
1038 		rtxdelta += tp->t_frextents_delta;
1039 
1040 	if (xfs_sb_version_haslazysbcount(&mp->m_sb) ||
1041 	     (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
1042 		idelta = tp->t_icount_delta;
1043 		ifreedelta = tp->t_ifree_delta;
1044 	}
1045 
1046 	/* apply the per-cpu counters */
1047 	if (blkdelta) {
1048 		error = xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
1049 						 blkdelta, rsvd);
1050 		if (error)
1051 			goto out;
1052 	}
1053 
1054 	if (idelta) {
1055 		error = xfs_icsb_modify_counters(mp, XFS_SBS_ICOUNT,
1056 						 idelta, rsvd);
1057 		if (error)
1058 			goto out_undo_fdblocks;
1059 	}
1060 
1061 	if (ifreedelta) {
1062 		error = xfs_icsb_modify_counters(mp, XFS_SBS_IFREE,
1063 						 ifreedelta, rsvd);
1064 		if (error)
1065 			goto out_undo_icount;
1066 	}
1067 
1068 	/* apply remaining deltas */
1069 	if (rtxdelta != 0) {
1070 		msbp->msb_field = XFS_SBS_FREXTENTS;
1071 		msbp->msb_delta = rtxdelta;
1072 		msbp++;
1073 	}
1074 
1075 	if (tp->t_flags & XFS_TRANS_SB_DIRTY) {
1076 		if (tp->t_dblocks_delta != 0) {
1077 			msbp->msb_field = XFS_SBS_DBLOCKS;
1078 			msbp->msb_delta = tp->t_dblocks_delta;
1079 			msbp++;
1080 		}
1081 		if (tp->t_agcount_delta != 0) {
1082 			msbp->msb_field = XFS_SBS_AGCOUNT;
1083 			msbp->msb_delta = tp->t_agcount_delta;
1084 			msbp++;
1085 		}
1086 		if (tp->t_imaxpct_delta != 0) {
1087 			msbp->msb_field = XFS_SBS_IMAX_PCT;
1088 			msbp->msb_delta = tp->t_imaxpct_delta;
1089 			msbp++;
1090 		}
1091 		if (tp->t_rextsize_delta != 0) {
1092 			msbp->msb_field = XFS_SBS_REXTSIZE;
1093 			msbp->msb_delta = tp->t_rextsize_delta;
1094 			msbp++;
1095 		}
1096 		if (tp->t_rbmblocks_delta != 0) {
1097 			msbp->msb_field = XFS_SBS_RBMBLOCKS;
1098 			msbp->msb_delta = tp->t_rbmblocks_delta;
1099 			msbp++;
1100 		}
1101 		if (tp->t_rblocks_delta != 0) {
1102 			msbp->msb_field = XFS_SBS_RBLOCKS;
1103 			msbp->msb_delta = tp->t_rblocks_delta;
1104 			msbp++;
1105 		}
1106 		if (tp->t_rextents_delta != 0) {
1107 			msbp->msb_field = XFS_SBS_REXTENTS;
1108 			msbp->msb_delta = tp->t_rextents_delta;
1109 			msbp++;
1110 		}
1111 		if (tp->t_rextslog_delta != 0) {
1112 			msbp->msb_field = XFS_SBS_REXTSLOG;
1113 			msbp->msb_delta = tp->t_rextslog_delta;
1114 			msbp++;
1115 		}
1116 	}
1117 
1118 	/*
1119 	 * If we need to change anything, do it.
1120 	 */
1121 	if (msbp > msb) {
1122 		error = xfs_mod_incore_sb_batch(tp->t_mountp, msb,
1123 			(uint)(msbp - msb), rsvd);
1124 		if (error)
1125 			goto out_undo_ifreecount;
1126 	}
1127 
1128 	return;
1129 
1130 out_undo_ifreecount:
1131 	if (ifreedelta)
1132 		xfs_icsb_modify_counters(mp, XFS_SBS_IFREE, -ifreedelta, rsvd);
1133 out_undo_icount:
1134 	if (idelta)
1135 		xfs_icsb_modify_counters(mp, XFS_SBS_ICOUNT, -idelta, rsvd);
1136 out_undo_fdblocks:
1137 	if (blkdelta)
1138 		xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS, -blkdelta, rsvd);
1139 out:
1140 	ASSERT(error == 0);
1141 	return;
1142 }
1143 
1144 /*
1145  * Add the given log item to the transaction's list of log items.
1146  *
1147  * The log item will now point to its new descriptor with its li_desc field.
1148  */
1149 void
1150 xfs_trans_add_item(
1151 	struct xfs_trans	*tp,
1152 	struct xfs_log_item	*lip)
1153 {
1154 	struct xfs_log_item_desc *lidp;
1155 
1156 	ASSERT(lip->li_mountp = tp->t_mountp);
1157 	ASSERT(lip->li_ailp = tp->t_mountp->m_ail);
1158 
1159 	lidp = kmem_zone_zalloc(xfs_log_item_desc_zone, KM_SLEEP | KM_NOFS);
1160 
1161 	lidp->lid_item = lip;
1162 	lidp->lid_flags = 0;
1163 	lidp->lid_size = 0;
1164 	list_add_tail(&lidp->lid_trans, &tp->t_items);
1165 
1166 	lip->li_desc = lidp;
1167 }
1168 
1169 STATIC void
1170 xfs_trans_free_item_desc(
1171 	struct xfs_log_item_desc *lidp)
1172 {
1173 	list_del_init(&lidp->lid_trans);
1174 	kmem_zone_free(xfs_log_item_desc_zone, lidp);
1175 }
1176 
1177 /*
1178  * Unlink and free the given descriptor.
1179  */
1180 void
1181 xfs_trans_del_item(
1182 	struct xfs_log_item	*lip)
1183 {
1184 	xfs_trans_free_item_desc(lip->li_desc);
1185 	lip->li_desc = NULL;
1186 }
1187 
1188 /*
1189  * Unlock all of the items of a transaction and free all the descriptors
1190  * of that transaction.
1191  */
1192 void
1193 xfs_trans_free_items(
1194 	struct xfs_trans	*tp,
1195 	xfs_lsn_t		commit_lsn,
1196 	int			flags)
1197 {
1198 	struct xfs_log_item_desc *lidp, *next;
1199 
1200 	list_for_each_entry_safe(lidp, next, &tp->t_items, lid_trans) {
1201 		struct xfs_log_item	*lip = lidp->lid_item;
1202 
1203 		lip->li_desc = NULL;
1204 
1205 		if (commit_lsn != NULLCOMMITLSN)
1206 			IOP_COMMITTING(lip, commit_lsn);
1207 		if (flags & XFS_TRANS_ABORT)
1208 			lip->li_flags |= XFS_LI_ABORTED;
1209 		IOP_UNLOCK(lip);
1210 
1211 		xfs_trans_free_item_desc(lidp);
1212 	}
1213 }
1214 
1215 /*
1216  * Unlock the items associated with a transaction.
1217  *
1218  * Items which were not logged should be freed.  Those which were logged must
1219  * still be tracked so they can be unpinned when the transaction commits.
1220  */
1221 STATIC void
1222 xfs_trans_unlock_items(
1223 	struct xfs_trans	*tp,
1224 	xfs_lsn_t		commit_lsn)
1225 {
1226 	struct xfs_log_item_desc *lidp, *next;
1227 
1228 	list_for_each_entry_safe(lidp, next, &tp->t_items, lid_trans) {
1229 		struct xfs_log_item	*lip = lidp->lid_item;
1230 
1231 		lip->li_desc = NULL;
1232 
1233 		if (commit_lsn != NULLCOMMITLSN)
1234 			IOP_COMMITTING(lip, commit_lsn);
1235 		IOP_UNLOCK(lip);
1236 
1237 		/*
1238 		 * Free the descriptor if the item is not dirty
1239 		 * within this transaction.
1240 		 */
1241 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
1242 			xfs_trans_free_item_desc(lidp);
1243 	}
1244 }
1245 
1246 /*
1247  * Total up the number of log iovecs needed to commit this
1248  * transaction.  The transaction itself needs one for the
1249  * transaction header.  Ask each dirty item in turn how many
1250  * it needs to get the total.
1251  */
1252 static uint
1253 xfs_trans_count_vecs(
1254 	struct xfs_trans	*tp)
1255 {
1256 	int			nvecs;
1257 	struct xfs_log_item_desc *lidp;
1258 
1259 	nvecs = 1;
1260 
1261 	/* In the non-debug case we need to start bailing out if we
1262 	 * didn't find a log_item here, return zero and let trans_commit
1263 	 * deal with it.
1264 	 */
1265 	if (list_empty(&tp->t_items)) {
1266 		ASSERT(0);
1267 		return 0;
1268 	}
1269 
1270 	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
1271 		/*
1272 		 * Skip items which aren't dirty in this transaction.
1273 		 */
1274 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
1275 			continue;
1276 		lidp->lid_size = IOP_SIZE(lidp->lid_item);
1277 		nvecs += lidp->lid_size;
1278 	}
1279 
1280 	return nvecs;
1281 }
1282 
1283 /*
1284  * Fill in the vector with pointers to data to be logged
1285  * by this transaction.  The transaction header takes
1286  * the first vector, and then each dirty item takes the
1287  * number of vectors it indicated it needed in xfs_trans_count_vecs().
1288  *
1289  * As each item fills in the entries it needs, also pin the item
1290  * so that it cannot be flushed out until the log write completes.
1291  */
1292 static void
1293 xfs_trans_fill_vecs(
1294 	struct xfs_trans	*tp,
1295 	struct xfs_log_iovec	*log_vector)
1296 {
1297 	struct xfs_log_item_desc *lidp;
1298 	struct xfs_log_iovec	*vecp;
1299 	uint			nitems;
1300 
1301 	/*
1302 	 * Skip over the entry for the transaction header, we'll
1303 	 * fill that in at the end.
1304 	 */
1305 	vecp = log_vector + 1;
1306 
1307 	nitems = 0;
1308 	ASSERT(!list_empty(&tp->t_items));
1309 	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
1310 		/* Skip items which aren't dirty in this transaction. */
1311 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
1312 			continue;
1313 
1314 		/*
1315 		 * The item may be marked dirty but not log anything.  This can
1316 		 * be used to get called when a transaction is committed.
1317 		 */
1318 		if (lidp->lid_size)
1319 			nitems++;
1320 		IOP_FORMAT(lidp->lid_item, vecp);
1321 		vecp += lidp->lid_size;
1322 		IOP_PIN(lidp->lid_item);
1323 	}
1324 
1325 	/*
1326 	 * Now that we've counted the number of items in this transaction, fill
1327 	 * in the transaction header. Note that the transaction header does not
1328 	 * have a log item.
1329 	 */
1330 	tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
1331 	tp->t_header.th_type = tp->t_type;
1332 	tp->t_header.th_num_items = nitems;
1333 	log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
1334 	log_vector->i_len = sizeof(xfs_trans_header_t);
1335 	log_vector->i_type = XLOG_REG_TYPE_TRANSHDR;
1336 }
1337 
1338 /*
1339  * The committed item processing consists of calling the committed routine of
1340  * each logged item, updating the item's position in the AIL if necessary, and
1341  * unpinning each item.  If the committed routine returns -1, then do nothing
1342  * further with the item because it may have been freed.
1343  *
1344  * Since items are unlocked when they are copied to the incore log, it is
1345  * possible for two transactions to be completing and manipulating the same
1346  * item simultaneously.  The AIL lock will protect the lsn field of each item.
1347  * The value of this field can never go backwards.
1348  *
1349  * We unpin the items after repositioning them in the AIL, because otherwise
1350  * they could be immediately flushed and we'd have to race with the flusher
1351  * trying to pull the item from the AIL as we add it.
1352  */
1353 static void
1354 xfs_trans_item_committed(
1355 	struct xfs_log_item	*lip,
1356 	xfs_lsn_t		commit_lsn,
1357 	int			aborted)
1358 {
1359 	xfs_lsn_t		item_lsn;
1360 	struct xfs_ail		*ailp;
1361 
1362 	if (aborted)
1363 		lip->li_flags |= XFS_LI_ABORTED;
1364 	item_lsn = IOP_COMMITTED(lip, commit_lsn);
1365 
1366 	/* If the committed routine returns -1, item has been freed. */
1367 	if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
1368 		return;
1369 
1370 	/*
1371 	 * If the returned lsn is greater than what it contained before, update
1372 	 * the location of the item in the AIL.  If it is not, then do nothing.
1373 	 * Items can never move backwards in the AIL.
1374 	 *
1375 	 * While the new lsn should usually be greater, it is possible that a
1376 	 * later transaction completing simultaneously with an earlier one
1377 	 * using the same item could complete first with a higher lsn.  This
1378 	 * would cause the earlier transaction to fail the test below.
1379 	 */
1380 	ailp = lip->li_ailp;
1381 	spin_lock(&ailp->xa_lock);
1382 	if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0) {
1383 		/*
1384 		 * This will set the item's lsn to item_lsn and update the
1385 		 * position of the item in the AIL.
1386 		 *
1387 		 * xfs_trans_ail_update() drops the AIL lock.
1388 		 */
1389 		xfs_trans_ail_update(ailp, lip, item_lsn);
1390 	} else {
1391 		spin_unlock(&ailp->xa_lock);
1392 	}
1393 
1394 	/*
1395 	 * Now that we've repositioned the item in the AIL, unpin it so it can
1396 	 * be flushed. Pass information about buffer stale state down from the
1397 	 * log item flags, if anyone else stales the buffer we do not want to
1398 	 * pay any attention to it.
1399 	 */
1400 	IOP_UNPIN(lip, 0);
1401 }
1402 
1403 /*
1404  * This is typically called by the LM when a transaction has been fully
1405  * committed to disk.  It needs to unpin the items which have
1406  * been logged by the transaction and update their positions
1407  * in the AIL if necessary.
1408  *
1409  * This also gets called when the transactions didn't get written out
1410  * because of an I/O error. Abortflag & XFS_LI_ABORTED is set then.
1411  */
1412 STATIC void
1413 xfs_trans_committed(
1414 	void			*arg,
1415 	int			abortflag)
1416 {
1417 	struct xfs_trans	*tp = arg;
1418 	struct xfs_log_item_desc *lidp, *next;
1419 
1420 	list_for_each_entry_safe(lidp, next, &tp->t_items, lid_trans) {
1421 		xfs_trans_item_committed(lidp->lid_item, tp->t_lsn, abortflag);
1422 		xfs_trans_free_item_desc(lidp);
1423 	}
1424 
1425 	xfs_trans_free(tp);
1426 }
1427 
1428 static inline void
1429 xfs_log_item_batch_insert(
1430 	struct xfs_ail		*ailp,
1431 	struct xfs_log_item	**log_items,
1432 	int			nr_items,
1433 	xfs_lsn_t		commit_lsn)
1434 {
1435 	int	i;
1436 
1437 	spin_lock(&ailp->xa_lock);
1438 	/* xfs_trans_ail_update_bulk drops ailp->xa_lock */
1439 	xfs_trans_ail_update_bulk(ailp, log_items, nr_items, commit_lsn);
1440 
1441 	for (i = 0; i < nr_items; i++)
1442 		IOP_UNPIN(log_items[i], 0);
1443 }
1444 
1445 /*
1446  * Bulk operation version of xfs_trans_committed that takes a log vector of
1447  * items to insert into the AIL. This uses bulk AIL insertion techniques to
1448  * minimise lock traffic.
1449  *
1450  * If we are called with the aborted flag set, it is because a log write during
1451  * a CIL checkpoint commit has failed. In this case, all the items in the
1452  * checkpoint have already gone through IOP_COMMITED and IOP_UNLOCK, which
1453  * means that checkpoint commit abort handling is treated exactly the same
1454  * as an iclog write error even though we haven't started any IO yet. Hence in
1455  * this case all we need to do is IOP_COMMITTED processing, followed by an
1456  * IOP_UNPIN(aborted) call.
1457  */
1458 void
1459 xfs_trans_committed_bulk(
1460 	struct xfs_ail		*ailp,
1461 	struct xfs_log_vec	*log_vector,
1462 	xfs_lsn_t		commit_lsn,
1463 	int			aborted)
1464 {
1465 #define LOG_ITEM_BATCH_SIZE	32
1466 	struct xfs_log_item	*log_items[LOG_ITEM_BATCH_SIZE];
1467 	struct xfs_log_vec	*lv;
1468 	int			i = 0;
1469 
1470 	/* unpin all the log items */
1471 	for (lv = log_vector; lv; lv = lv->lv_next ) {
1472 		struct xfs_log_item	*lip = lv->lv_item;
1473 		xfs_lsn_t		item_lsn;
1474 
1475 		if (aborted)
1476 			lip->li_flags |= XFS_LI_ABORTED;
1477 		item_lsn = IOP_COMMITTED(lip, commit_lsn);
1478 
1479 		/* item_lsn of -1 means the item was freed */
1480 		if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)
1481 			continue;
1482 
1483 		/*
1484 		 * if we are aborting the operation, no point in inserting the
1485 		 * object into the AIL as we are in a shutdown situation.
1486 		 */
1487 		if (aborted) {
1488 			ASSERT(XFS_FORCED_SHUTDOWN(ailp->xa_mount));
1489 			IOP_UNPIN(lip, 1);
1490 			continue;
1491 		}
1492 
1493 		if (item_lsn != commit_lsn) {
1494 
1495 			/*
1496 			 * Not a bulk update option due to unusual item_lsn.
1497 			 * Push into AIL immediately, rechecking the lsn once
1498 			 * we have the ail lock. Then unpin the item.
1499 			 */
1500 			spin_lock(&ailp->xa_lock);
1501 			if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)
1502 				xfs_trans_ail_update(ailp, lip, item_lsn);
1503 			else
1504 				spin_unlock(&ailp->xa_lock);
1505 			IOP_UNPIN(lip, 0);
1506 			continue;
1507 		}
1508 
1509 		/* Item is a candidate for bulk AIL insert.  */
1510 		log_items[i++] = lv->lv_item;
1511 		if (i >= LOG_ITEM_BATCH_SIZE) {
1512 			xfs_log_item_batch_insert(ailp, log_items,
1513 					LOG_ITEM_BATCH_SIZE, commit_lsn);
1514 			i = 0;
1515 		}
1516 	}
1517 
1518 	/* make sure we insert the remainder! */
1519 	if (i)
1520 		xfs_log_item_batch_insert(ailp, log_items, i, commit_lsn);
1521 }
1522 
1523 /*
1524  * Called from the trans_commit code when we notice that the filesystem is in
1525  * the middle of a forced shutdown.
1526  *
1527  * When we are called here, we have already pinned all the items in the
1528  * transaction. However, neither IOP_COMMITTING or IOP_UNLOCK has been called
1529  * so we can simply walk the items in the transaction, unpin them with an abort
1530  * flag and then free the items. Note that unpinning the items can result in
1531  * them being freed immediately, so we need to use a safe list traversal method
1532  * here.
1533  */
1534 STATIC void
1535 xfs_trans_uncommit(
1536 	struct xfs_trans	*tp,
1537 	uint			flags)
1538 {
1539 	struct xfs_log_item_desc *lidp, *n;
1540 
1541 	list_for_each_entry_safe(lidp, n, &tp->t_items, lid_trans) {
1542 		if (lidp->lid_flags & XFS_LID_DIRTY)
1543 			IOP_UNPIN(lidp->lid_item, 1);
1544 	}
1545 
1546 	xfs_trans_unreserve_and_mod_sb(tp);
1547 	xfs_trans_unreserve_and_mod_dquots(tp);
1548 
1549 	xfs_trans_free_items(tp, NULLCOMMITLSN, flags);
1550 	xfs_trans_free(tp);
1551 }
1552 
1553 /*
1554  * Format the transaction direct to the iclog. This isolates the physical
1555  * transaction commit operation from the logical operation and hence allows
1556  * other methods to be introduced without affecting the existing commit path.
1557  */
1558 static int
1559 xfs_trans_commit_iclog(
1560 	struct xfs_mount	*mp,
1561 	struct xfs_trans	*tp,
1562 	xfs_lsn_t		*commit_lsn,
1563 	int			flags)
1564 {
1565 	int			shutdown;
1566 	int			error;
1567 	int			log_flags = 0;
1568 	struct xlog_in_core	*commit_iclog;
1569 #define XFS_TRANS_LOGVEC_COUNT  16
1570 	struct xfs_log_iovec	log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
1571 	struct xfs_log_iovec	*log_vector;
1572 	uint			nvec;
1573 
1574 
1575 	/*
1576 	 * Ask each log item how many log_vector entries it will
1577 	 * need so we can figure out how many to allocate.
1578 	 * Try to avoid the kmem_alloc() call in the common case
1579 	 * by using a vector from the stack when it fits.
1580 	 */
1581 	nvec = xfs_trans_count_vecs(tp);
1582 	if (nvec == 0) {
1583 		return ENOMEM;	/* triggers a shutdown! */
1584 	} else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
1585 		log_vector = log_vector_fast;
1586 	} else {
1587 		log_vector = (xfs_log_iovec_t *)kmem_alloc(nvec *
1588 						   sizeof(xfs_log_iovec_t),
1589 						   KM_SLEEP);
1590 	}
1591 
1592 	/*
1593 	 * Fill in the log_vector and pin the logged items, and
1594 	 * then write the transaction to the log.
1595 	 */
1596 	xfs_trans_fill_vecs(tp, log_vector);
1597 
1598 	if (flags & XFS_TRANS_RELEASE_LOG_RES)
1599 		log_flags = XFS_LOG_REL_PERM_RESERV;
1600 
1601 	error = xfs_log_write(mp, log_vector, nvec, tp->t_ticket, &(tp->t_lsn));
1602 
1603 	/*
1604 	 * The transaction is committed incore here, and can go out to disk
1605 	 * at any time after this call.  However, all the items associated
1606 	 * with the transaction are still locked and pinned in memory.
1607 	 */
1608 	*commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
1609 
1610 	tp->t_commit_lsn = *commit_lsn;
1611 	trace_xfs_trans_commit_lsn(tp);
1612 
1613 	if (nvec > XFS_TRANS_LOGVEC_COUNT)
1614 		kmem_free(log_vector);
1615 
1616 	/*
1617 	 * If we got a log write error. Unpin the logitems that we
1618 	 * had pinned, clean up, free trans structure, and return error.
1619 	 */
1620 	if (error || *commit_lsn == -1) {
1621 		current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1622 		xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
1623 		return XFS_ERROR(EIO);
1624 	}
1625 
1626 	/*
1627 	 * Once the transaction has committed, unused
1628 	 * reservations need to be released and changes to
1629 	 * the superblock need to be reflected in the in-core
1630 	 * version.  Do that now.
1631 	 */
1632 	xfs_trans_unreserve_and_mod_sb(tp);
1633 
1634 	/*
1635 	 * Tell the LM to call the transaction completion routine
1636 	 * when the log write with LSN commit_lsn completes (e.g.
1637 	 * when the transaction commit really hits the on-disk log).
1638 	 * After this call we cannot reference tp, because the call
1639 	 * can happen at any time and the call will free the transaction
1640 	 * structure pointed to by tp.  The only case where we call
1641 	 * the completion routine (xfs_trans_committed) directly is
1642 	 * if the log is turned off on a debug kernel or we're
1643 	 * running in simulation mode (the log is explicitly turned
1644 	 * off).
1645 	 */
1646 	tp->t_logcb.cb_func = xfs_trans_committed;
1647 	tp->t_logcb.cb_arg = tp;
1648 
1649 	/*
1650 	 * We need to pass the iclog buffer which was used for the
1651 	 * transaction commit record into this function, and attach
1652 	 * the callback to it. The callback must be attached before
1653 	 * the items are unlocked to avoid racing with other threads
1654 	 * waiting for an item to unlock.
1655 	 */
1656 	shutdown = xfs_log_notify(mp, commit_iclog, &(tp->t_logcb));
1657 
1658 	/*
1659 	 * Mark this thread as no longer being in a transaction
1660 	 */
1661 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1662 
1663 	/*
1664 	 * Once all the items of the transaction have been copied
1665 	 * to the in core log and the callback is attached, the
1666 	 * items can be unlocked.
1667 	 *
1668 	 * This will free descriptors pointing to items which were
1669 	 * not logged since there is nothing more to do with them.
1670 	 * For items which were logged, we will keep pointers to them
1671 	 * so they can be unpinned after the transaction commits to disk.
1672 	 * This will also stamp each modified meta-data item with
1673 	 * the commit lsn of this transaction for dependency tracking
1674 	 * purposes.
1675 	 */
1676 	xfs_trans_unlock_items(tp, *commit_lsn);
1677 
1678 	/*
1679 	 * If we detected a log error earlier, finish committing
1680 	 * the transaction now (unpin log items, etc).
1681 	 *
1682 	 * Order is critical here, to avoid using the transaction
1683 	 * pointer after its been freed (by xfs_trans_committed
1684 	 * either here now, or as a callback).  We cannot do this
1685 	 * step inside xfs_log_notify as was done earlier because
1686 	 * of this issue.
1687 	 */
1688 	if (shutdown)
1689 		xfs_trans_committed(tp, XFS_LI_ABORTED);
1690 
1691 	/*
1692 	 * Now that the xfs_trans_committed callback has been attached,
1693 	 * and the items are released we can finally allow the iclog to
1694 	 * go to disk.
1695 	 */
1696 	return xfs_log_release_iclog(mp, commit_iclog);
1697 }
1698 
1699 /*
1700  * Walk the log items and allocate log vector structures for
1701  * each item large enough to fit all the vectors they require.
1702  * Note that this format differs from the old log vector format in
1703  * that there is no transaction header in these log vectors.
1704  */
1705 STATIC struct xfs_log_vec *
1706 xfs_trans_alloc_log_vecs(
1707 	xfs_trans_t	*tp)
1708 {
1709 	struct xfs_log_item_desc *lidp;
1710 	struct xfs_log_vec	*lv = NULL;
1711 	struct xfs_log_vec	*ret_lv = NULL;
1712 
1713 
1714 	/* Bail out if we didn't find a log item.  */
1715 	if (list_empty(&tp->t_items)) {
1716 		ASSERT(0);
1717 		return NULL;
1718 	}
1719 
1720 	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
1721 		struct xfs_log_vec *new_lv;
1722 
1723 		/* Skip items which aren't dirty in this transaction. */
1724 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
1725 			continue;
1726 
1727 		/* Skip items that do not have any vectors for writing */
1728 		lidp->lid_size = IOP_SIZE(lidp->lid_item);
1729 		if (!lidp->lid_size)
1730 			continue;
1731 
1732 		new_lv = kmem_zalloc(sizeof(*new_lv) +
1733 				lidp->lid_size * sizeof(struct xfs_log_iovec),
1734 				KM_SLEEP);
1735 
1736 		/* The allocated iovec region lies beyond the log vector. */
1737 		new_lv->lv_iovecp = (struct xfs_log_iovec *)&new_lv[1];
1738 		new_lv->lv_niovecs = lidp->lid_size;
1739 		new_lv->lv_item = lidp->lid_item;
1740 		if (!ret_lv)
1741 			ret_lv = new_lv;
1742 		else
1743 			lv->lv_next = new_lv;
1744 		lv = new_lv;
1745 	}
1746 
1747 	return ret_lv;
1748 }
1749 
1750 static int
1751 xfs_trans_commit_cil(
1752 	struct xfs_mount	*mp,
1753 	struct xfs_trans	*tp,
1754 	xfs_lsn_t		*commit_lsn,
1755 	int			flags)
1756 {
1757 	struct xfs_log_vec	*log_vector;
1758 
1759 	/*
1760 	 * Get each log item to allocate a vector structure for
1761 	 * the log item to to pass to the log write code. The
1762 	 * CIL commit code will format the vector and save it away.
1763 	 */
1764 	log_vector = xfs_trans_alloc_log_vecs(tp);
1765 	if (!log_vector)
1766 		return ENOMEM;
1767 
1768 	xfs_log_commit_cil(mp, tp, log_vector, commit_lsn, flags);
1769 
1770 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1771 	xfs_trans_free(tp);
1772 	return 0;
1773 }
1774 
1775 /*
1776  * xfs_trans_commit
1777  *
1778  * Commit the given transaction to the log a/synchronously.
1779  *
1780  * XFS disk error handling mechanism is not based on a typical
1781  * transaction abort mechanism. Logically after the filesystem
1782  * gets marked 'SHUTDOWN', we can't let any new transactions
1783  * be durable - ie. committed to disk - because some metadata might
1784  * be inconsistent. In such cases, this returns an error, and the
1785  * caller may assume that all locked objects joined to the transaction
1786  * have already been unlocked as if the commit had succeeded.
1787  * Do not reference the transaction structure after this call.
1788  */
1789 int
1790 _xfs_trans_commit(
1791 	struct xfs_trans	*tp,
1792 	uint			flags,
1793 	int			*log_flushed)
1794 {
1795 	struct xfs_mount	*mp = tp->t_mountp;
1796 	xfs_lsn_t		commit_lsn = -1;
1797 	int			error = 0;
1798 	int			log_flags = 0;
1799 	int			sync = tp->t_flags & XFS_TRANS_SYNC;
1800 
1801 	/*
1802 	 * Determine whether this commit is releasing a permanent
1803 	 * log reservation or not.
1804 	 */
1805 	if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1806 		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1807 		log_flags = XFS_LOG_REL_PERM_RESERV;
1808 	}
1809 
1810 	/*
1811 	 * If there is nothing to be logged by the transaction,
1812 	 * then unlock all of the items associated with the
1813 	 * transaction and free the transaction structure.
1814 	 * Also make sure to return any reserved blocks to
1815 	 * the free pool.
1816 	 */
1817 	if (!(tp->t_flags & XFS_TRANS_DIRTY))
1818 		goto out_unreserve;
1819 
1820 	if (XFS_FORCED_SHUTDOWN(mp)) {
1821 		error = XFS_ERROR(EIO);
1822 		goto out_unreserve;
1823 	}
1824 
1825 	ASSERT(tp->t_ticket != NULL);
1826 
1827 	/*
1828 	 * If we need to update the superblock, then do it now.
1829 	 */
1830 	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
1831 		xfs_trans_apply_sb_deltas(tp);
1832 	xfs_trans_apply_dquot_deltas(tp);
1833 
1834 	if (mp->m_flags & XFS_MOUNT_DELAYLOG)
1835 		error = xfs_trans_commit_cil(mp, tp, &commit_lsn, flags);
1836 	else
1837 		error = xfs_trans_commit_iclog(mp, tp, &commit_lsn, flags);
1838 
1839 	if (error == ENOMEM) {
1840 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1841 		error = XFS_ERROR(EIO);
1842 		goto out_unreserve;
1843 	}
1844 
1845 	/*
1846 	 * If the transaction needs to be synchronous, then force the
1847 	 * log out now and wait for it.
1848 	 */
1849 	if (sync) {
1850 		if (!error) {
1851 			error = _xfs_log_force_lsn(mp, commit_lsn,
1852 				      XFS_LOG_SYNC, log_flushed);
1853 		}
1854 		XFS_STATS_INC(xs_trans_sync);
1855 	} else {
1856 		XFS_STATS_INC(xs_trans_async);
1857 	}
1858 
1859 	return error;
1860 
1861 out_unreserve:
1862 	xfs_trans_unreserve_and_mod_sb(tp);
1863 
1864 	/*
1865 	 * It is indeed possible for the transaction to be not dirty but
1866 	 * the dqinfo portion to be.  All that means is that we have some
1867 	 * (non-persistent) quota reservations that need to be unreserved.
1868 	 */
1869 	xfs_trans_unreserve_and_mod_dquots(tp);
1870 	if (tp->t_ticket) {
1871 		commit_lsn = xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1872 		if (commit_lsn == -1 && !error)
1873 			error = XFS_ERROR(EIO);
1874 	}
1875 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1876 	xfs_trans_free_items(tp, NULLCOMMITLSN, error ? XFS_TRANS_ABORT : 0);
1877 	xfs_trans_free(tp);
1878 
1879 	XFS_STATS_INC(xs_trans_empty);
1880 	return error;
1881 }
1882 
1883 /*
1884  * Unlock all of the transaction's items and free the transaction.
1885  * The transaction must not have modified any of its items, because
1886  * there is no way to restore them to their previous state.
1887  *
1888  * If the transaction has made a log reservation, make sure to release
1889  * it as well.
1890  */
1891 void
1892 xfs_trans_cancel(
1893 	xfs_trans_t		*tp,
1894 	int			flags)
1895 {
1896 	int			log_flags;
1897 	xfs_mount_t		*mp = tp->t_mountp;
1898 
1899 	/*
1900 	 * See if the caller is being too lazy to figure out if
1901 	 * the transaction really needs an abort.
1902 	 */
1903 	if ((flags & XFS_TRANS_ABORT) && !(tp->t_flags & XFS_TRANS_DIRTY))
1904 		flags &= ~XFS_TRANS_ABORT;
1905 	/*
1906 	 * See if the caller is relying on us to shut down the
1907 	 * filesystem.  This happens in paths where we detect
1908 	 * corruption and decide to give up.
1909 	 */
1910 	if ((tp->t_flags & XFS_TRANS_DIRTY) && !XFS_FORCED_SHUTDOWN(mp)) {
1911 		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
1912 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1913 	}
1914 #ifdef DEBUG
1915 	if (!(flags & XFS_TRANS_ABORT) && !XFS_FORCED_SHUTDOWN(mp)) {
1916 		struct xfs_log_item_desc *lidp;
1917 
1918 		list_for_each_entry(lidp, &tp->t_items, lid_trans)
1919 			ASSERT(!(lidp->lid_item->li_type == XFS_LI_EFD));
1920 	}
1921 #endif
1922 	xfs_trans_unreserve_and_mod_sb(tp);
1923 	xfs_trans_unreserve_and_mod_dquots(tp);
1924 
1925 	if (tp->t_ticket) {
1926 		if (flags & XFS_TRANS_RELEASE_LOG_RES) {
1927 			ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1928 			log_flags = XFS_LOG_REL_PERM_RESERV;
1929 		} else {
1930 			log_flags = 0;
1931 		}
1932 		xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
1933 	}
1934 
1935 	/* mark this thread as no longer being in a transaction */
1936 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
1937 
1938 	xfs_trans_free_items(tp, NULLCOMMITLSN, flags);
1939 	xfs_trans_free(tp);
1940 }
1941 
1942 /*
1943  * Roll from one trans in the sequence of PERMANENT transactions to
1944  * the next: permanent transactions are only flushed out when
1945  * committed with XFS_TRANS_RELEASE_LOG_RES, but we still want as soon
1946  * as possible to let chunks of it go to the log. So we commit the
1947  * chunk we've been working on and get a new transaction to continue.
1948  */
1949 int
1950 xfs_trans_roll(
1951 	struct xfs_trans	**tpp,
1952 	struct xfs_inode	*dp)
1953 {
1954 	struct xfs_trans	*trans;
1955 	unsigned int		logres, count;
1956 	int			error;
1957 
1958 	/*
1959 	 * Ensure that the inode is always logged.
1960 	 */
1961 	trans = *tpp;
1962 	xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
1963 
1964 	/*
1965 	 * Copy the critical parameters from one trans to the next.
1966 	 */
1967 	logres = trans->t_log_res;
1968 	count = trans->t_log_count;
1969 	*tpp = xfs_trans_dup(trans);
1970 
1971 	/*
1972 	 * Commit the current transaction.
1973 	 * If this commit failed, then it'd just unlock those items that
1974 	 * are not marked ihold. That also means that a filesystem shutdown
1975 	 * is in progress. The caller takes the responsibility to cancel
1976 	 * the duplicate transaction that gets returned.
1977 	 */
1978 	error = xfs_trans_commit(trans, 0);
1979 	if (error)
1980 		return (error);
1981 
1982 	trans = *tpp;
1983 
1984 	/*
1985 	 * transaction commit worked ok so we can drop the extra ticket
1986 	 * reference that we gained in xfs_trans_dup()
1987 	 */
1988 	xfs_log_ticket_put(trans->t_ticket);
1989 
1990 
1991 	/*
1992 	 * Reserve space in the log for th next transaction.
1993 	 * This also pushes items in the "AIL", the list of logged items,
1994 	 * out to disk if they are taking up space at the tail of the log
1995 	 * that we want to use.  This requires that either nothing be locked
1996 	 * across this call, or that anything that is locked be logged in
1997 	 * the prior and the next transactions.
1998 	 */
1999 	error = xfs_trans_reserve(trans, 0, logres, 0,
2000 				  XFS_TRANS_PERM_LOG_RES, count);
2001 	/*
2002 	 *  Ensure that the inode is in the new transaction and locked.
2003 	 */
2004 	if (error)
2005 		return error;
2006 
2007 	xfs_trans_ijoin(trans, dp);
2008 	return 0;
2009 }
2010