xref: /openbmc/linux/fs/xfs/xfs_dquot.c (revision 9c1f8594)
1 /*
2  * Copyright (c) 2000-2003 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_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_trans.h"
24 #include "xfs_sb.h"
25 #include "xfs_ag.h"
26 #include "xfs_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_mount.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_bmap.h"
32 #include "xfs_rtalloc.h"
33 #include "xfs_error.h"
34 #include "xfs_itable.h"
35 #include "xfs_attr.h"
36 #include "xfs_buf_item.h"
37 #include "xfs_trans_space.h"
38 #include "xfs_trans_priv.h"
39 #include "xfs_qm.h"
40 #include "xfs_trace.h"
41 
42 
43 /*
44    LOCK ORDER
45 
46    inode lock		    (ilock)
47    dquot hash-chain lock    (hashlock)
48    xqm dquot freelist lock  (freelistlock
49    mount's dquot list lock  (mplistlock)
50    user dquot lock - lock ordering among dquots is based on the uid or gid
51    group dquot lock - similar to udquots. Between the two dquots, the udquot
52 		      has to be locked first.
53    pin lock - the dquot lock must be held to take this lock.
54    flush lock - ditto.
55 */
56 
57 #ifdef DEBUG
58 xfs_buftarg_t *xfs_dqerror_target;
59 int xfs_do_dqerror;
60 int xfs_dqreq_num;
61 int xfs_dqerror_mod = 33;
62 #endif
63 
64 static struct lock_class_key xfs_dquot_other_class;
65 
66 /*
67  * Allocate and initialize a dquot. We don't always allocate fresh memory;
68  * we try to reclaim a free dquot if the number of incore dquots are above
69  * a threshold.
70  * The only field inside the core that gets initialized at this point
71  * is the d_id field. The idea is to fill in the entire q_core
72  * when we read in the on disk dquot.
73  */
74 STATIC xfs_dquot_t *
75 xfs_qm_dqinit(
76 	xfs_mount_t  *mp,
77 	xfs_dqid_t   id,
78 	uint	     type)
79 {
80 	xfs_dquot_t	*dqp;
81 	boolean_t	brandnewdquot;
82 
83 	brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
84 	dqp->dq_flags = type;
85 	dqp->q_core.d_id = cpu_to_be32(id);
86 	dqp->q_mount = mp;
87 
88 	/*
89 	 * No need to re-initialize these if this is a reclaimed dquot.
90 	 */
91 	if (brandnewdquot) {
92 		INIT_LIST_HEAD(&dqp->q_freelist);
93 		mutex_init(&dqp->q_qlock);
94 		init_waitqueue_head(&dqp->q_pinwait);
95 
96 		/*
97 		 * Because we want to use a counting completion, complete
98 		 * the flush completion once to allow a single access to
99 		 * the flush completion without blocking.
100 		 */
101 		init_completion(&dqp->q_flush);
102 		complete(&dqp->q_flush);
103 
104 		trace_xfs_dqinit(dqp);
105 	} else {
106 		/*
107 		 * Only the q_core portion was zeroed in dqreclaim_one().
108 		 * So, we need to reset others.
109 		 */
110 		dqp->q_nrefs = 0;
111 		dqp->q_blkno = 0;
112 		INIT_LIST_HEAD(&dqp->q_mplist);
113 		INIT_LIST_HEAD(&dqp->q_hashlist);
114 		dqp->q_bufoffset = 0;
115 		dqp->q_fileoffset = 0;
116 		dqp->q_transp = NULL;
117 		dqp->q_gdquot = NULL;
118 		dqp->q_res_bcount = 0;
119 		dqp->q_res_icount = 0;
120 		dqp->q_res_rtbcount = 0;
121 		atomic_set(&dqp->q_pincount, 0);
122 		dqp->q_hash = NULL;
123 		ASSERT(list_empty(&dqp->q_freelist));
124 
125 		trace_xfs_dqreuse(dqp);
126 	}
127 
128 	/*
129 	 * In either case we need to make sure group quotas have a different
130 	 * lock class than user quotas, to make sure lockdep knows we can
131 	 * locks of one of each at the same time.
132 	 */
133 	if (!(type & XFS_DQ_USER))
134 		lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
135 
136 	/*
137 	 * log item gets initialized later
138 	 */
139 	return (dqp);
140 }
141 
142 /*
143  * This is called to free all the memory associated with a dquot
144  */
145 void
146 xfs_qm_dqdestroy(
147 	xfs_dquot_t	*dqp)
148 {
149 	ASSERT(list_empty(&dqp->q_freelist));
150 
151 	mutex_destroy(&dqp->q_qlock);
152 	kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
153 
154 	atomic_dec(&xfs_Gqm->qm_totaldquots);
155 }
156 
157 /*
158  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
159  */
160 STATIC void
161 xfs_qm_dqinit_core(
162 	xfs_dqid_t	id,
163 	uint		type,
164 	xfs_dqblk_t	*d)
165 {
166 	/*
167 	 * Caller has zero'd the entire dquot 'chunk' already.
168 	 */
169 	d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
170 	d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
171 	d->dd_diskdq.d_id = cpu_to_be32(id);
172 	d->dd_diskdq.d_flags = type;
173 }
174 
175 /*
176  * If default limits are in force, push them into the dquot now.
177  * We overwrite the dquot limits only if they are zero and this
178  * is not the root dquot.
179  */
180 void
181 xfs_qm_adjust_dqlimits(
182 	xfs_mount_t		*mp,
183 	xfs_disk_dquot_t	*d)
184 {
185 	xfs_quotainfo_t		*q = mp->m_quotainfo;
186 
187 	ASSERT(d->d_id);
188 
189 	if (q->qi_bsoftlimit && !d->d_blk_softlimit)
190 		d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
191 	if (q->qi_bhardlimit && !d->d_blk_hardlimit)
192 		d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
193 	if (q->qi_isoftlimit && !d->d_ino_softlimit)
194 		d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
195 	if (q->qi_ihardlimit && !d->d_ino_hardlimit)
196 		d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
197 	if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
198 		d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
199 	if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
200 		d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
201 }
202 
203 /*
204  * Check the limits and timers of a dquot and start or reset timers
205  * if necessary.
206  * This gets called even when quota enforcement is OFF, which makes our
207  * life a little less complicated. (We just don't reject any quota
208  * reservations in that case, when enforcement is off).
209  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
210  * enforcement's off.
211  * In contrast, warnings are a little different in that they don't
212  * 'automatically' get started when limits get exceeded.  They do
213  * get reset to zero, however, when we find the count to be under
214  * the soft limit (they are only ever set non-zero via userspace).
215  */
216 void
217 xfs_qm_adjust_dqtimers(
218 	xfs_mount_t		*mp,
219 	xfs_disk_dquot_t	*d)
220 {
221 	ASSERT(d->d_id);
222 
223 #ifdef DEBUG
224 	if (d->d_blk_hardlimit)
225 		ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
226 		       be64_to_cpu(d->d_blk_hardlimit));
227 	if (d->d_ino_hardlimit)
228 		ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
229 		       be64_to_cpu(d->d_ino_hardlimit));
230 	if (d->d_rtb_hardlimit)
231 		ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
232 		       be64_to_cpu(d->d_rtb_hardlimit));
233 #endif
234 
235 	if (!d->d_btimer) {
236 		if ((d->d_blk_softlimit &&
237 		     (be64_to_cpu(d->d_bcount) >=
238 		      be64_to_cpu(d->d_blk_softlimit))) ||
239 		    (d->d_blk_hardlimit &&
240 		     (be64_to_cpu(d->d_bcount) >=
241 		      be64_to_cpu(d->d_blk_hardlimit)))) {
242 			d->d_btimer = cpu_to_be32(get_seconds() +
243 					mp->m_quotainfo->qi_btimelimit);
244 		} else {
245 			d->d_bwarns = 0;
246 		}
247 	} else {
248 		if ((!d->d_blk_softlimit ||
249 		     (be64_to_cpu(d->d_bcount) <
250 		      be64_to_cpu(d->d_blk_softlimit))) &&
251 		    (!d->d_blk_hardlimit ||
252 		    (be64_to_cpu(d->d_bcount) <
253 		     be64_to_cpu(d->d_blk_hardlimit)))) {
254 			d->d_btimer = 0;
255 		}
256 	}
257 
258 	if (!d->d_itimer) {
259 		if ((d->d_ino_softlimit &&
260 		     (be64_to_cpu(d->d_icount) >=
261 		      be64_to_cpu(d->d_ino_softlimit))) ||
262 		    (d->d_ino_hardlimit &&
263 		     (be64_to_cpu(d->d_icount) >=
264 		      be64_to_cpu(d->d_ino_hardlimit)))) {
265 			d->d_itimer = cpu_to_be32(get_seconds() +
266 					mp->m_quotainfo->qi_itimelimit);
267 		} else {
268 			d->d_iwarns = 0;
269 		}
270 	} else {
271 		if ((!d->d_ino_softlimit ||
272 		     (be64_to_cpu(d->d_icount) <
273 		      be64_to_cpu(d->d_ino_softlimit)))  &&
274 		    (!d->d_ino_hardlimit ||
275 		     (be64_to_cpu(d->d_icount) <
276 		      be64_to_cpu(d->d_ino_hardlimit)))) {
277 			d->d_itimer = 0;
278 		}
279 	}
280 
281 	if (!d->d_rtbtimer) {
282 		if ((d->d_rtb_softlimit &&
283 		     (be64_to_cpu(d->d_rtbcount) >=
284 		      be64_to_cpu(d->d_rtb_softlimit))) ||
285 		    (d->d_rtb_hardlimit &&
286 		     (be64_to_cpu(d->d_rtbcount) >=
287 		      be64_to_cpu(d->d_rtb_hardlimit)))) {
288 			d->d_rtbtimer = cpu_to_be32(get_seconds() +
289 					mp->m_quotainfo->qi_rtbtimelimit);
290 		} else {
291 			d->d_rtbwarns = 0;
292 		}
293 	} else {
294 		if ((!d->d_rtb_softlimit ||
295 		     (be64_to_cpu(d->d_rtbcount) <
296 		      be64_to_cpu(d->d_rtb_softlimit))) &&
297 		    (!d->d_rtb_hardlimit ||
298 		     (be64_to_cpu(d->d_rtbcount) <
299 		      be64_to_cpu(d->d_rtb_hardlimit)))) {
300 			d->d_rtbtimer = 0;
301 		}
302 	}
303 }
304 
305 /*
306  * initialize a buffer full of dquots and log the whole thing
307  */
308 STATIC void
309 xfs_qm_init_dquot_blk(
310 	xfs_trans_t	*tp,
311 	xfs_mount_t	*mp,
312 	xfs_dqid_t	id,
313 	uint		type,
314 	xfs_buf_t	*bp)
315 {
316 	struct xfs_quotainfo	*q = mp->m_quotainfo;
317 	xfs_dqblk_t	*d;
318 	int		curid, i;
319 
320 	ASSERT(tp);
321 	ASSERT(xfs_buf_islocked(bp));
322 
323 	d = bp->b_addr;
324 
325 	/*
326 	 * ID of the first dquot in the block - id's are zero based.
327 	 */
328 	curid = id - (id % q->qi_dqperchunk);
329 	ASSERT(curid >= 0);
330 	memset(d, 0, BBTOB(q->qi_dqchunklen));
331 	for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++)
332 		xfs_qm_dqinit_core(curid, type, d);
333 	xfs_trans_dquot_buf(tp, bp,
334 			    (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
335 			    ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
336 			     XFS_BLF_GDQUOT_BUF)));
337 	xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
338 }
339 
340 
341 
342 /*
343  * Allocate a block and fill it with dquots.
344  * This is called when the bmapi finds a hole.
345  */
346 STATIC int
347 xfs_qm_dqalloc(
348 	xfs_trans_t	**tpp,
349 	xfs_mount_t	*mp,
350 	xfs_dquot_t	*dqp,
351 	xfs_inode_t	*quotip,
352 	xfs_fileoff_t	offset_fsb,
353 	xfs_buf_t	**O_bpp)
354 {
355 	xfs_fsblock_t	firstblock;
356 	xfs_bmap_free_t flist;
357 	xfs_bmbt_irec_t map;
358 	int		nmaps, error, committed;
359 	xfs_buf_t	*bp;
360 	xfs_trans_t	*tp = *tpp;
361 
362 	ASSERT(tp != NULL);
363 
364 	trace_xfs_dqalloc(dqp);
365 
366 	/*
367 	 * Initialize the bmap freelist prior to calling bmapi code.
368 	 */
369 	xfs_bmap_init(&flist, &firstblock);
370 	xfs_ilock(quotip, XFS_ILOCK_EXCL);
371 	/*
372 	 * Return if this type of quotas is turned off while we didn't
373 	 * have an inode lock
374 	 */
375 	if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
376 		xfs_iunlock(quotip, XFS_ILOCK_EXCL);
377 		return (ESRCH);
378 	}
379 
380 	xfs_trans_ijoin_ref(tp, quotip, XFS_ILOCK_EXCL);
381 	nmaps = 1;
382 	if ((error = xfs_bmapi(tp, quotip,
383 			      offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
384 			      XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
385 			      &firstblock,
386 			      XFS_QM_DQALLOC_SPACE_RES(mp),
387 			      &map, &nmaps, &flist))) {
388 		goto error0;
389 	}
390 	ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
391 	ASSERT(nmaps == 1);
392 	ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
393 	       (map.br_startblock != HOLESTARTBLOCK));
394 
395 	/*
396 	 * Keep track of the blkno to save a lookup later
397 	 */
398 	dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
399 
400 	/* now we can just get the buffer (there's nothing to read yet) */
401 	bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
402 			       dqp->q_blkno,
403 			       mp->m_quotainfo->qi_dqchunklen,
404 			       0);
405 	if (!bp || (error = xfs_buf_geterror(bp)))
406 		goto error1;
407 	/*
408 	 * Make a chunk of dquots out of this buffer and log
409 	 * the entire thing.
410 	 */
411 	xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
412 			      dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
413 
414 	/*
415 	 * xfs_bmap_finish() may commit the current transaction and
416 	 * start a second transaction if the freelist is not empty.
417 	 *
418 	 * Since we still want to modify this buffer, we need to
419 	 * ensure that the buffer is not released on commit of
420 	 * the first transaction and ensure the buffer is added to the
421 	 * second transaction.
422 	 *
423 	 * If there is only one transaction then don't stop the buffer
424 	 * from being released when it commits later on.
425 	 */
426 
427 	xfs_trans_bhold(tp, bp);
428 
429 	if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
430 		goto error1;
431 	}
432 
433 	if (committed) {
434 		tp = *tpp;
435 		xfs_trans_bjoin(tp, bp);
436 	} else {
437 		xfs_trans_bhold_release(tp, bp);
438 	}
439 
440 	*O_bpp = bp;
441 	return 0;
442 
443       error1:
444 	xfs_bmap_cancel(&flist);
445       error0:
446 	xfs_iunlock(quotip, XFS_ILOCK_EXCL);
447 
448 	return (error);
449 }
450 
451 /*
452  * Maps a dquot to the buffer containing its on-disk version.
453  * This returns a ptr to the buffer containing the on-disk dquot
454  * in the bpp param, and a ptr to the on-disk dquot within that buffer
455  */
456 STATIC int
457 xfs_qm_dqtobp(
458 	xfs_trans_t		**tpp,
459 	xfs_dquot_t		*dqp,
460 	xfs_disk_dquot_t	**O_ddpp,
461 	xfs_buf_t		**O_bpp,
462 	uint			flags)
463 {
464 	xfs_bmbt_irec_t map;
465 	int		nmaps = 1, error;
466 	xfs_buf_t	*bp;
467 	xfs_inode_t	*quotip = XFS_DQ_TO_QIP(dqp);
468 	xfs_mount_t	*mp = dqp->q_mount;
469 	xfs_disk_dquot_t *ddq;
470 	xfs_dqid_t	id = be32_to_cpu(dqp->q_core.d_id);
471 	xfs_trans_t	*tp = (tpp ? *tpp : NULL);
472 
473 	dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
474 
475 	xfs_ilock(quotip, XFS_ILOCK_SHARED);
476 	if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
477 		/*
478 		 * Return if this type of quotas is turned off while we
479 		 * didn't have the quota inode lock.
480 		 */
481 		xfs_iunlock(quotip, XFS_ILOCK_SHARED);
482 		return ESRCH;
483 	}
484 
485 	/*
486 	 * Find the block map; no allocations yet
487 	 */
488 	error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
489 			  XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
490 			  NULL, 0, &map, &nmaps, NULL);
491 
492 	xfs_iunlock(quotip, XFS_ILOCK_SHARED);
493 	if (error)
494 		return error;
495 
496 	ASSERT(nmaps == 1);
497 	ASSERT(map.br_blockcount == 1);
498 
499 	/*
500 	 * Offset of dquot in the (fixed sized) dquot chunk.
501 	 */
502 	dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
503 		sizeof(xfs_dqblk_t);
504 
505 	ASSERT(map.br_startblock != DELAYSTARTBLOCK);
506 	if (map.br_startblock == HOLESTARTBLOCK) {
507 		/*
508 		 * We don't allocate unless we're asked to
509 		 */
510 		if (!(flags & XFS_QMOPT_DQALLOC))
511 			return ENOENT;
512 
513 		ASSERT(tp);
514 		error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
515 					dqp->q_fileoffset, &bp);
516 		if (error)
517 			return error;
518 		tp = *tpp;
519 	} else {
520 		trace_xfs_dqtobp_read(dqp);
521 
522 		/*
523 		 * store the blkno etc so that we don't have to do the
524 		 * mapping all the time
525 		 */
526 		dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
527 
528 		error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
529 					   dqp->q_blkno,
530 					   mp->m_quotainfo->qi_dqchunklen,
531 					   0, &bp);
532 		if (error || !bp)
533 			return XFS_ERROR(error);
534 	}
535 
536 	ASSERT(xfs_buf_islocked(bp));
537 
538 	/*
539 	 * calculate the location of the dquot inside the buffer.
540 	 */
541 	ddq = bp->b_addr + dqp->q_bufoffset;
542 
543 	/*
544 	 * A simple sanity check in case we got a corrupted dquot...
545 	 */
546 	error = xfs_qm_dqcheck(mp, ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
547 			   flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
548 			   "dqtobp");
549 	if (error) {
550 		if (!(flags & XFS_QMOPT_DQREPAIR)) {
551 			xfs_trans_brelse(tp, bp);
552 			return XFS_ERROR(EIO);
553 		}
554 	}
555 
556 	*O_bpp = bp;
557 	*O_ddpp = ddq;
558 
559 	return (0);
560 }
561 
562 
563 /*
564  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
565  * and release the buffer immediately.
566  *
567  */
568 /* ARGSUSED */
569 STATIC int
570 xfs_qm_dqread(
571 	xfs_trans_t	**tpp,
572 	xfs_dqid_t	id,
573 	xfs_dquot_t	*dqp,	/* dquot to get filled in */
574 	uint		flags)
575 {
576 	xfs_disk_dquot_t *ddqp;
577 	xfs_buf_t	 *bp;
578 	int		 error;
579 	xfs_trans_t	 *tp;
580 
581 	ASSERT(tpp);
582 
583 	trace_xfs_dqread(dqp);
584 
585 	/*
586 	 * get a pointer to the on-disk dquot and the buffer containing it
587 	 * dqp already knows its own type (GROUP/USER).
588 	 */
589 	if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
590 		return (error);
591 	}
592 	tp = *tpp;
593 
594 	/* copy everything from disk dquot to the incore dquot */
595 	memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
596 	ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
597 	xfs_qm_dquot_logitem_init(dqp);
598 
599 	/*
600 	 * Reservation counters are defined as reservation plus current usage
601 	 * to avoid having to add every time.
602 	 */
603 	dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
604 	dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
605 	dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
606 
607 	/* Mark the buf so that this will stay incore a little longer */
608 	XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
609 
610 	/*
611 	 * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
612 	 * So we need to release with xfs_trans_brelse().
613 	 * The strategy here is identical to that of inodes; we lock
614 	 * the dquot in xfs_qm_dqget() before making it accessible to
615 	 * others. This is because dquots, like inodes, need a good level of
616 	 * concurrency, and we don't want to take locks on the entire buffers
617 	 * for dquot accesses.
618 	 * Note also that the dquot buffer may even be dirty at this point, if
619 	 * this particular dquot was repaired. We still aren't afraid to
620 	 * brelse it because we have the changes incore.
621 	 */
622 	ASSERT(xfs_buf_islocked(bp));
623 	xfs_trans_brelse(tp, bp);
624 
625 	return (error);
626 }
627 
628 
629 /*
630  * allocate an incore dquot from the kernel heap,
631  * and fill its core with quota information kept on disk.
632  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
633  * if it wasn't already allocated.
634  */
635 STATIC int
636 xfs_qm_idtodq(
637 	xfs_mount_t	*mp,
638 	xfs_dqid_t	id,	 /* gid or uid, depending on type */
639 	uint		type,	 /* UDQUOT or GDQUOT */
640 	uint		flags,	 /* DQALLOC, DQREPAIR */
641 	xfs_dquot_t	**O_dqpp)/* OUT : incore dquot, not locked */
642 {
643 	xfs_dquot_t	*dqp;
644 	int		error;
645 	xfs_trans_t	*tp;
646 	int		cancelflags=0;
647 
648 	dqp = xfs_qm_dqinit(mp, id, type);
649 	tp = NULL;
650 	if (flags & XFS_QMOPT_DQALLOC) {
651 		tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
652 		error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
653 				XFS_WRITE_LOG_RES(mp) +
654 				BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 +
655 				128,
656 				0,
657 				XFS_TRANS_PERM_LOG_RES,
658 				XFS_WRITE_LOG_COUNT);
659 		if (error) {
660 			cancelflags = 0;
661 			goto error0;
662 		}
663 		cancelflags = XFS_TRANS_RELEASE_LOG_RES;
664 	}
665 
666 	/*
667 	 * Read it from disk; xfs_dqread() takes care of
668 	 * all the necessary initialization of dquot's fields (locks, etc)
669 	 */
670 	if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
671 		/*
672 		 * This can happen if quotas got turned off (ESRCH),
673 		 * or if the dquot didn't exist on disk and we ask to
674 		 * allocate (ENOENT).
675 		 */
676 		trace_xfs_dqread_fail(dqp);
677 		cancelflags |= XFS_TRANS_ABORT;
678 		goto error0;
679 	}
680 	if (tp) {
681 		if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
682 			goto error1;
683 	}
684 
685 	*O_dqpp = dqp;
686 	return (0);
687 
688  error0:
689 	ASSERT(error);
690 	if (tp)
691 		xfs_trans_cancel(tp, cancelflags);
692  error1:
693 	xfs_qm_dqdestroy(dqp);
694 	*O_dqpp = NULL;
695 	return (error);
696 }
697 
698 /*
699  * Lookup a dquot in the incore dquot hashtable. We keep two separate
700  * hashtables for user and group dquots; and, these are global tables
701  * inside the XQM, not per-filesystem tables.
702  * The hash chain must be locked by caller, and it is left locked
703  * on return. Returning dquot is locked.
704  */
705 STATIC int
706 xfs_qm_dqlookup(
707 	xfs_mount_t		*mp,
708 	xfs_dqid_t		id,
709 	xfs_dqhash_t		*qh,
710 	xfs_dquot_t		**O_dqpp)
711 {
712 	xfs_dquot_t		*dqp;
713 	uint			flist_locked;
714 
715 	ASSERT(mutex_is_locked(&qh->qh_lock));
716 
717 	flist_locked = B_FALSE;
718 
719 	/*
720 	 * Traverse the hashchain looking for a match
721 	 */
722 	list_for_each_entry(dqp, &qh->qh_list, q_hashlist) {
723 		/*
724 		 * We already have the hashlock. We don't need the
725 		 * dqlock to look at the id field of the dquot, since the
726 		 * id can't be modified without the hashlock anyway.
727 		 */
728 		if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
729 			trace_xfs_dqlookup_found(dqp);
730 
731 			/*
732 			 * All in core dquots must be on the dqlist of mp
733 			 */
734 			ASSERT(!list_empty(&dqp->q_mplist));
735 
736 			xfs_dqlock(dqp);
737 			if (dqp->q_nrefs == 0) {
738 				ASSERT(!list_empty(&dqp->q_freelist));
739 				if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
740 					trace_xfs_dqlookup_want(dqp);
741 
742 					/*
743 					 * We may have raced with dqreclaim_one()
744 					 * (and lost). So, flag that we don't
745 					 * want the dquot to be reclaimed.
746 					 */
747 					dqp->dq_flags |= XFS_DQ_WANT;
748 					xfs_dqunlock(dqp);
749 					mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
750 					xfs_dqlock(dqp);
751 					dqp->dq_flags &= ~(XFS_DQ_WANT);
752 				}
753 				flist_locked = B_TRUE;
754 			}
755 
756 			/*
757 			 * id couldn't have changed; we had the hashlock all
758 			 * along
759 			 */
760 			ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
761 
762 			if (flist_locked) {
763 				if (dqp->q_nrefs != 0) {
764 					mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
765 					flist_locked = B_FALSE;
766 				} else {
767 					/* take it off the freelist */
768 					trace_xfs_dqlookup_freelist(dqp);
769 					list_del_init(&dqp->q_freelist);
770 					xfs_Gqm->qm_dqfrlist_cnt--;
771 				}
772 			}
773 
774 			XFS_DQHOLD(dqp);
775 
776 			if (flist_locked)
777 				mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
778 			/*
779 			 * move the dquot to the front of the hashchain
780 			 */
781 			ASSERT(mutex_is_locked(&qh->qh_lock));
782 			list_move(&dqp->q_hashlist, &qh->qh_list);
783 			trace_xfs_dqlookup_done(dqp);
784 			*O_dqpp = dqp;
785 			return 0;
786 		}
787 	}
788 
789 	*O_dqpp = NULL;
790 	ASSERT(mutex_is_locked(&qh->qh_lock));
791 	return (1);
792 }
793 
794 /*
795  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
796  * a locked dquot, doing an allocation (if requested) as needed.
797  * When both an inode and an id are given, the inode's id takes precedence.
798  * That is, if the id changes while we don't hold the ilock inside this
799  * function, the new dquot is returned, not necessarily the one requested
800  * in the id argument.
801  */
802 int
803 xfs_qm_dqget(
804 	xfs_mount_t	*mp,
805 	xfs_inode_t	*ip,	  /* locked inode (optional) */
806 	xfs_dqid_t	id,	  /* uid/projid/gid depending on type */
807 	uint		type,	  /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
808 	uint		flags,	  /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
809 	xfs_dquot_t	**O_dqpp) /* OUT : locked incore dquot */
810 {
811 	xfs_dquot_t	*dqp;
812 	xfs_dqhash_t	*h;
813 	uint		version;
814 	int		error;
815 
816 	ASSERT(XFS_IS_QUOTA_RUNNING(mp));
817 	if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
818 	    (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
819 	    (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
820 		return (ESRCH);
821 	}
822 	h = XFS_DQ_HASH(mp, id, type);
823 
824 #ifdef DEBUG
825 	if (xfs_do_dqerror) {
826 		if ((xfs_dqerror_target == mp->m_ddev_targp) &&
827 		    (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
828 			xfs_debug(mp, "Returning error in dqget");
829 			return (EIO);
830 		}
831 	}
832 #endif
833 
834  again:
835 
836 #ifdef DEBUG
837 	ASSERT(type == XFS_DQ_USER ||
838 	       type == XFS_DQ_PROJ ||
839 	       type == XFS_DQ_GROUP);
840 	if (ip) {
841 		ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
842 		if (type == XFS_DQ_USER)
843 			ASSERT(ip->i_udquot == NULL);
844 		else
845 			ASSERT(ip->i_gdquot == NULL);
846 	}
847 #endif
848 	mutex_lock(&h->qh_lock);
849 
850 	/*
851 	 * Look in the cache (hashtable).
852 	 * The chain is kept locked during lookup.
853 	 */
854 	if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
855 		XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
856 		/*
857 		 * The dquot was found, moved to the front of the chain,
858 		 * taken off the freelist if it was on it, and locked
859 		 * at this point. Just unlock the hashchain and return.
860 		 */
861 		ASSERT(*O_dqpp);
862 		ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
863 		mutex_unlock(&h->qh_lock);
864 		trace_xfs_dqget_hit(*O_dqpp);
865 		return (0);	/* success */
866 	}
867 	XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
868 
869 	/*
870 	 * Dquot cache miss. We don't want to keep the inode lock across
871 	 * a (potential) disk read. Also we don't want to deal with the lock
872 	 * ordering between quotainode and this inode. OTOH, dropping the inode
873 	 * lock here means dealing with a chown that can happen before
874 	 * we re-acquire the lock.
875 	 */
876 	if (ip)
877 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
878 	/*
879 	 * Save the hashchain version stamp, and unlock the chain, so that
880 	 * we don't keep the lock across a disk read
881 	 */
882 	version = h->qh_version;
883 	mutex_unlock(&h->qh_lock);
884 
885 	/*
886 	 * Allocate the dquot on the kernel heap, and read the ondisk
887 	 * portion off the disk. Also, do all the necessary initialization
888 	 * This can return ENOENT if dquot didn't exist on disk and we didn't
889 	 * ask it to allocate; ESRCH if quotas got turned off suddenly.
890 	 */
891 	if ((error = xfs_qm_idtodq(mp, id, type,
892 				  flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
893 					   XFS_QMOPT_DOWARN),
894 				  &dqp))) {
895 		if (ip)
896 			xfs_ilock(ip, XFS_ILOCK_EXCL);
897 		return (error);
898 	}
899 
900 	/*
901 	 * See if this is mount code calling to look at the overall quota limits
902 	 * which are stored in the id == 0 user or group's dquot.
903 	 * Since we may not have done a quotacheck by this point, just return
904 	 * the dquot without attaching it to any hashtables, lists, etc, or even
905 	 * taking a reference.
906 	 * The caller must dqdestroy this once done.
907 	 */
908 	if (flags & XFS_QMOPT_DQSUSER) {
909 		ASSERT(id == 0);
910 		ASSERT(! ip);
911 		goto dqret;
912 	}
913 
914 	/*
915 	 * Dquot lock comes after hashlock in the lock ordering
916 	 */
917 	if (ip) {
918 		xfs_ilock(ip, XFS_ILOCK_EXCL);
919 
920 		/*
921 		 * A dquot could be attached to this inode by now, since
922 		 * we had dropped the ilock.
923 		 */
924 		if (type == XFS_DQ_USER) {
925 			if (!XFS_IS_UQUOTA_ON(mp)) {
926 				/* inode stays locked on return */
927 				xfs_qm_dqdestroy(dqp);
928 				return XFS_ERROR(ESRCH);
929 			}
930 			if (ip->i_udquot) {
931 				xfs_qm_dqdestroy(dqp);
932 				dqp = ip->i_udquot;
933 				xfs_dqlock(dqp);
934 				goto dqret;
935 			}
936 		} else {
937 			if (!XFS_IS_OQUOTA_ON(mp)) {
938 				/* inode stays locked on return */
939 				xfs_qm_dqdestroy(dqp);
940 				return XFS_ERROR(ESRCH);
941 			}
942 			if (ip->i_gdquot) {
943 				xfs_qm_dqdestroy(dqp);
944 				dqp = ip->i_gdquot;
945 				xfs_dqlock(dqp);
946 				goto dqret;
947 			}
948 		}
949 	}
950 
951 	/*
952 	 * Hashlock comes after ilock in lock order
953 	 */
954 	mutex_lock(&h->qh_lock);
955 	if (version != h->qh_version) {
956 		xfs_dquot_t *tmpdqp;
957 		/*
958 		 * Now, see if somebody else put the dquot in the
959 		 * hashtable before us. This can happen because we didn't
960 		 * keep the hashchain lock. We don't have to worry about
961 		 * lock order between the two dquots here since dqp isn't
962 		 * on any findable lists yet.
963 		 */
964 		if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
965 			/*
966 			 * Duplicate found. Just throw away the new dquot
967 			 * and start over.
968 			 */
969 			xfs_qm_dqput(tmpdqp);
970 			mutex_unlock(&h->qh_lock);
971 			xfs_qm_dqdestroy(dqp);
972 			XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
973 			goto again;
974 		}
975 	}
976 
977 	/*
978 	 * Put the dquot at the beginning of the hash-chain and mp's list
979 	 * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
980 	 */
981 	ASSERT(mutex_is_locked(&h->qh_lock));
982 	dqp->q_hash = h;
983 	list_add(&dqp->q_hashlist, &h->qh_list);
984 	h->qh_version++;
985 
986 	/*
987 	 * Attach this dquot to this filesystem's list of all dquots,
988 	 * kept inside the mount structure in m_quotainfo field
989 	 */
990 	mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
991 
992 	/*
993 	 * We return a locked dquot to the caller, with a reference taken
994 	 */
995 	xfs_dqlock(dqp);
996 	dqp->q_nrefs = 1;
997 
998 	list_add(&dqp->q_mplist, &mp->m_quotainfo->qi_dqlist);
999 	mp->m_quotainfo->qi_dquots++;
1000 	mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1001 	mutex_unlock(&h->qh_lock);
1002  dqret:
1003 	ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1004 	trace_xfs_dqget_miss(dqp);
1005 	*O_dqpp = dqp;
1006 	return (0);
1007 }
1008 
1009 
1010 /*
1011  * Release a reference to the dquot (decrement ref-count)
1012  * and unlock it. If there is a group quota attached to this
1013  * dquot, carefully release that too without tripping over
1014  * deadlocks'n'stuff.
1015  */
1016 void
1017 xfs_qm_dqput(
1018 	xfs_dquot_t	*dqp)
1019 {
1020 	xfs_dquot_t	*gdqp;
1021 
1022 	ASSERT(dqp->q_nrefs > 0);
1023 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
1024 
1025 	trace_xfs_dqput(dqp);
1026 
1027 	if (dqp->q_nrefs != 1) {
1028 		dqp->q_nrefs--;
1029 		xfs_dqunlock(dqp);
1030 		return;
1031 	}
1032 
1033 	/*
1034 	 * drop the dqlock and acquire the freelist and dqlock
1035 	 * in the right order; but try to get it out-of-order first
1036 	 */
1037 	if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
1038 		trace_xfs_dqput_wait(dqp);
1039 		xfs_dqunlock(dqp);
1040 		mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1041 		xfs_dqlock(dqp);
1042 	}
1043 
1044 	while (1) {
1045 		gdqp = NULL;
1046 
1047 		/* We can't depend on nrefs being == 1 here */
1048 		if (--dqp->q_nrefs == 0) {
1049 			trace_xfs_dqput_free(dqp);
1050 
1051 			list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
1052 			xfs_Gqm->qm_dqfrlist_cnt++;
1053 
1054 			/*
1055 			 * If we just added a udquot to the freelist, then
1056 			 * we want to release the gdquot reference that
1057 			 * it (probably) has. Otherwise it'll keep the
1058 			 * gdquot from getting reclaimed.
1059 			 */
1060 			if ((gdqp = dqp->q_gdquot)) {
1061 				/*
1062 				 * Avoid a recursive dqput call
1063 				 */
1064 				xfs_dqlock(gdqp);
1065 				dqp->q_gdquot = NULL;
1066 			}
1067 		}
1068 		xfs_dqunlock(dqp);
1069 
1070 		/*
1071 		 * If we had a group quota inside the user quota as a hint,
1072 		 * release it now.
1073 		 */
1074 		if (! gdqp)
1075 			break;
1076 		dqp = gdqp;
1077 	}
1078 	mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1079 }
1080 
1081 /*
1082  * Release a dquot. Flush it if dirty, then dqput() it.
1083  * dquot must not be locked.
1084  */
1085 void
1086 xfs_qm_dqrele(
1087 	xfs_dquot_t	*dqp)
1088 {
1089 	if (!dqp)
1090 		return;
1091 
1092 	trace_xfs_dqrele(dqp);
1093 
1094 	xfs_dqlock(dqp);
1095 	/*
1096 	 * We don't care to flush it if the dquot is dirty here.
1097 	 * That will create stutters that we want to avoid.
1098 	 * Instead we do a delayed write when we try to reclaim
1099 	 * a dirty dquot. Also xfs_sync will take part of the burden...
1100 	 */
1101 	xfs_qm_dqput(dqp);
1102 }
1103 
1104 /*
1105  * This is the dquot flushing I/O completion routine.  It is called
1106  * from interrupt level when the buffer containing the dquot is
1107  * flushed to disk.  It is responsible for removing the dquot logitem
1108  * from the AIL if it has not been re-logged, and unlocking the dquot's
1109  * flush lock. This behavior is very similar to that of inodes..
1110  */
1111 STATIC void
1112 xfs_qm_dqflush_done(
1113 	struct xfs_buf		*bp,
1114 	struct xfs_log_item	*lip)
1115 {
1116 	xfs_dq_logitem_t	*qip = (struct xfs_dq_logitem *)lip;
1117 	xfs_dquot_t		*dqp = qip->qli_dquot;
1118 	struct xfs_ail		*ailp = lip->li_ailp;
1119 
1120 	/*
1121 	 * We only want to pull the item from the AIL if its
1122 	 * location in the log has not changed since we started the flush.
1123 	 * Thus, we only bother if the dquot's lsn has
1124 	 * not changed. First we check the lsn outside the lock
1125 	 * since it's cheaper, and then we recheck while
1126 	 * holding the lock before removing the dquot from the AIL.
1127 	 */
1128 	if ((lip->li_flags & XFS_LI_IN_AIL) &&
1129 	    lip->li_lsn == qip->qli_flush_lsn) {
1130 
1131 		/* xfs_trans_ail_delete() drops the AIL lock. */
1132 		spin_lock(&ailp->xa_lock);
1133 		if (lip->li_lsn == qip->qli_flush_lsn)
1134 			xfs_trans_ail_delete(ailp, lip);
1135 		else
1136 			spin_unlock(&ailp->xa_lock);
1137 	}
1138 
1139 	/*
1140 	 * Release the dq's flush lock since we're done with it.
1141 	 */
1142 	xfs_dqfunlock(dqp);
1143 }
1144 
1145 /*
1146  * Write a modified dquot to disk.
1147  * The dquot must be locked and the flush lock too taken by caller.
1148  * The flush lock will not be unlocked until the dquot reaches the disk,
1149  * but the dquot is free to be unlocked and modified by the caller
1150  * in the interim. Dquot is still locked on return. This behavior is
1151  * identical to that of inodes.
1152  */
1153 int
1154 xfs_qm_dqflush(
1155 	xfs_dquot_t		*dqp,
1156 	uint			flags)
1157 {
1158 	struct xfs_mount	*mp = dqp->q_mount;
1159 	struct xfs_buf		*bp;
1160 	struct xfs_disk_dquot	*ddqp;
1161 	int			error;
1162 
1163 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
1164 	ASSERT(!completion_done(&dqp->q_flush));
1165 
1166 	trace_xfs_dqflush(dqp);
1167 
1168 	/*
1169 	 * If not dirty, or it's pinned and we are not supposed to block, nada.
1170 	 */
1171 	if (!XFS_DQ_IS_DIRTY(dqp) ||
1172 	    (!(flags & SYNC_WAIT) && atomic_read(&dqp->q_pincount) > 0)) {
1173 		xfs_dqfunlock(dqp);
1174 		return 0;
1175 	}
1176 	xfs_qm_dqunpin_wait(dqp);
1177 
1178 	/*
1179 	 * This may have been unpinned because the filesystem is shutting
1180 	 * down forcibly. If that's the case we must not write this dquot
1181 	 * to disk, because the log record didn't make it to disk!
1182 	 */
1183 	if (XFS_FORCED_SHUTDOWN(mp)) {
1184 		dqp->dq_flags &= ~XFS_DQ_DIRTY;
1185 		xfs_dqfunlock(dqp);
1186 		return XFS_ERROR(EIO);
1187 	}
1188 
1189 	/*
1190 	 * Get the buffer containing the on-disk dquot
1191 	 */
1192 	error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1193 				   mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1194 	if (error) {
1195 		ASSERT(error != ENOENT);
1196 		xfs_dqfunlock(dqp);
1197 		return error;
1198 	}
1199 
1200 	/*
1201 	 * Calculate the location of the dquot inside the buffer.
1202 	 */
1203 	ddqp = bp->b_addr + dqp->q_bufoffset;
1204 
1205 	/*
1206 	 * A simple sanity check in case we got a corrupted dquot..
1207 	 */
1208 	error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1209 			   XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1210 	if (error) {
1211 		xfs_buf_relse(bp);
1212 		xfs_dqfunlock(dqp);
1213 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1214 		return XFS_ERROR(EIO);
1215 	}
1216 
1217 	/* This is the only portion of data that needs to persist */
1218 	memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1219 
1220 	/*
1221 	 * Clear the dirty field and remember the flush lsn for later use.
1222 	 */
1223 	dqp->dq_flags &= ~XFS_DQ_DIRTY;
1224 
1225 	xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1226 					&dqp->q_logitem.qli_item.li_lsn);
1227 
1228 	/*
1229 	 * Attach an iodone routine so that we can remove this dquot from the
1230 	 * AIL and release the flush lock once the dquot is synced to disk.
1231 	 */
1232 	xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1233 				  &dqp->q_logitem.qli_item);
1234 
1235 	/*
1236 	 * If the buffer is pinned then push on the log so we won't
1237 	 * get stuck waiting in the write for too long.
1238 	 */
1239 	if (xfs_buf_ispinned(bp)) {
1240 		trace_xfs_dqflush_force(dqp);
1241 		xfs_log_force(mp, 0);
1242 	}
1243 
1244 	if (flags & SYNC_WAIT)
1245 		error = xfs_bwrite(mp, bp);
1246 	else
1247 		xfs_bdwrite(mp, bp);
1248 
1249 	trace_xfs_dqflush_done(dqp);
1250 
1251 	/*
1252 	 * dqp is still locked, but caller is free to unlock it now.
1253 	 */
1254 	return error;
1255 
1256 }
1257 
1258 int
1259 xfs_qm_dqlock_nowait(
1260 	xfs_dquot_t *dqp)
1261 {
1262 	return mutex_trylock(&dqp->q_qlock);
1263 }
1264 
1265 void
1266 xfs_dqlock(
1267 	xfs_dquot_t *dqp)
1268 {
1269 	mutex_lock(&dqp->q_qlock);
1270 }
1271 
1272 void
1273 xfs_dqunlock(
1274 	xfs_dquot_t *dqp)
1275 {
1276 	mutex_unlock(&(dqp->q_qlock));
1277 	if (dqp->q_logitem.qli_dquot == dqp) {
1278 		/* Once was dqp->q_mount, but might just have been cleared */
1279 		xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1280 					(xfs_log_item_t*)&(dqp->q_logitem));
1281 	}
1282 }
1283 
1284 
1285 void
1286 xfs_dqunlock_nonotify(
1287 	xfs_dquot_t *dqp)
1288 {
1289 	mutex_unlock(&(dqp->q_qlock));
1290 }
1291 
1292 /*
1293  * Lock two xfs_dquot structures.
1294  *
1295  * To avoid deadlocks we always lock the quota structure with
1296  * the lowerd id first.
1297  */
1298 void
1299 xfs_dqlock2(
1300 	xfs_dquot_t	*d1,
1301 	xfs_dquot_t	*d2)
1302 {
1303 	if (d1 && d2) {
1304 		ASSERT(d1 != d2);
1305 		if (be32_to_cpu(d1->q_core.d_id) >
1306 		    be32_to_cpu(d2->q_core.d_id)) {
1307 			mutex_lock(&d2->q_qlock);
1308 			mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1309 		} else {
1310 			mutex_lock(&d1->q_qlock);
1311 			mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1312 		}
1313 	} else if (d1) {
1314 		mutex_lock(&d1->q_qlock);
1315 	} else if (d2) {
1316 		mutex_lock(&d2->q_qlock);
1317 	}
1318 }
1319 
1320 
1321 /*
1322  * Take a dquot out of the mount's dqlist as well as the hashlist.
1323  * This is called via unmount as well as quotaoff, and the purge
1324  * will always succeed unless there are soft (temp) references
1325  * outstanding.
1326  *
1327  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1328  * that we're returning! XXXsup - not cool.
1329  */
1330 /* ARGSUSED */
1331 int
1332 xfs_qm_dqpurge(
1333 	xfs_dquot_t	*dqp)
1334 {
1335 	xfs_dqhash_t	*qh = dqp->q_hash;
1336 	xfs_mount_t	*mp = dqp->q_mount;
1337 
1338 	ASSERT(mutex_is_locked(&mp->m_quotainfo->qi_dqlist_lock));
1339 	ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1340 
1341 	xfs_dqlock(dqp);
1342 	/*
1343 	 * We really can't afford to purge a dquot that is
1344 	 * referenced, because these are hard refs.
1345 	 * It shouldn't happen in general because we went thru _all_ inodes in
1346 	 * dqrele_all_inodes before calling this and didn't let the mountlock go.
1347 	 * However it is possible that we have dquots with temporary
1348 	 * references that are not attached to an inode. e.g. see xfs_setattr().
1349 	 */
1350 	if (dqp->q_nrefs != 0) {
1351 		xfs_dqunlock(dqp);
1352 		mutex_unlock(&dqp->q_hash->qh_lock);
1353 		return (1);
1354 	}
1355 
1356 	ASSERT(!list_empty(&dqp->q_freelist));
1357 
1358 	/*
1359 	 * If we're turning off quotas, we have to make sure that, for
1360 	 * example, we don't delete quota disk blocks while dquots are
1361 	 * in the process of getting written to those disk blocks.
1362 	 * This dquot might well be on AIL, and we can't leave it there
1363 	 * if we're turning off quotas. Basically, we need this flush
1364 	 * lock, and are willing to block on it.
1365 	 */
1366 	if (!xfs_dqflock_nowait(dqp)) {
1367 		/*
1368 		 * Block on the flush lock after nudging dquot buffer,
1369 		 * if it is incore.
1370 		 */
1371 		xfs_qm_dqflock_pushbuf_wait(dqp);
1372 	}
1373 
1374 	/*
1375 	 * XXXIf we're turning this type of quotas off, we don't care
1376 	 * about the dirty metadata sitting in this dquot. OTOH, if
1377 	 * we're unmounting, we do care, so we flush it and wait.
1378 	 */
1379 	if (XFS_DQ_IS_DIRTY(dqp)) {
1380 		int	error;
1381 
1382 		/* dqflush unlocks dqflock */
1383 		/*
1384 		 * Given that dqpurge is a very rare occurrence, it is OK
1385 		 * that we're holding the hashlist and mplist locks
1386 		 * across the disk write. But, ... XXXsup
1387 		 *
1388 		 * We don't care about getting disk errors here. We need
1389 		 * to purge this dquot anyway, so we go ahead regardless.
1390 		 */
1391 		error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1392 		if (error)
1393 			xfs_warn(mp, "%s: dquot %p flush failed",
1394 				__func__, dqp);
1395 		xfs_dqflock(dqp);
1396 	}
1397 	ASSERT(atomic_read(&dqp->q_pincount) == 0);
1398 	ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1399 	       !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1400 
1401 	list_del_init(&dqp->q_hashlist);
1402 	qh->qh_version++;
1403 	list_del_init(&dqp->q_mplist);
1404 	mp->m_quotainfo->qi_dqreclaims++;
1405 	mp->m_quotainfo->qi_dquots--;
1406 	/*
1407 	 * XXX Move this to the front of the freelist, if we can get the
1408 	 * freelist lock.
1409 	 */
1410 	ASSERT(!list_empty(&dqp->q_freelist));
1411 
1412 	dqp->q_mount = NULL;
1413 	dqp->q_hash = NULL;
1414 	dqp->dq_flags = XFS_DQ_INACTIVE;
1415 	memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1416 	xfs_dqfunlock(dqp);
1417 	xfs_dqunlock(dqp);
1418 	mutex_unlock(&qh->qh_lock);
1419 	return (0);
1420 }
1421 
1422 
1423 /*
1424  * Give the buffer a little push if it is incore and
1425  * wait on the flush lock.
1426  */
1427 void
1428 xfs_qm_dqflock_pushbuf_wait(
1429 	xfs_dquot_t	*dqp)
1430 {
1431 	xfs_mount_t	*mp = dqp->q_mount;
1432 	xfs_buf_t	*bp;
1433 
1434 	/*
1435 	 * Check to see if the dquot has been flushed delayed
1436 	 * write.  If so, grab its buffer and send it
1437 	 * out immediately.  We'll be able to acquire
1438 	 * the flush lock when the I/O completes.
1439 	 */
1440 	bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1441 			mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1442 	if (!bp)
1443 		goto out_lock;
1444 
1445 	if (XFS_BUF_ISDELAYWRITE(bp)) {
1446 		if (xfs_buf_ispinned(bp))
1447 			xfs_log_force(mp, 0);
1448 		xfs_buf_delwri_promote(bp);
1449 		wake_up_process(bp->b_target->bt_task);
1450 	}
1451 	xfs_buf_relse(bp);
1452 out_lock:
1453 	xfs_dqflock(dqp);
1454 }
1455