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 *
xfs_qm_dqinit(xfs_mount_t * mp,xfs_dqid_t id,uint type)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
xfs_qm_dqdestroy(xfs_dquot_t * dqp)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
xfs_qm_dqinit_core(xfs_dqid_t id,uint type,xfs_dqblk_t * d)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
xfs_qm_adjust_dqlimits(xfs_mount_t * mp,xfs_disk_dquot_t * d)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
xfs_qm_adjust_dqtimers(xfs_mount_t * mp,xfs_disk_dquot_t * d)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 QUOTADEBUG
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 	if (!d->d_btimer) {
235 		if ((d->d_blk_softlimit &&
236 		     (be64_to_cpu(d->d_bcount) >=
237 		      be64_to_cpu(d->d_blk_softlimit))) ||
238 		    (d->d_blk_hardlimit &&
239 		     (be64_to_cpu(d->d_bcount) >=
240 		      be64_to_cpu(d->d_blk_hardlimit)))) {
241 			d->d_btimer = cpu_to_be32(get_seconds() +
242 					mp->m_quotainfo->qi_btimelimit);
243 		} else {
244 			d->d_bwarns = 0;
245 		}
246 	} else {
247 		if ((!d->d_blk_softlimit ||
248 		     (be64_to_cpu(d->d_bcount) <
249 		      be64_to_cpu(d->d_blk_softlimit))) &&
250 		    (!d->d_blk_hardlimit ||
251 		    (be64_to_cpu(d->d_bcount) <
252 		     be64_to_cpu(d->d_blk_hardlimit)))) {
253 			d->d_btimer = 0;
254 		}
255 	}
256 
257 	if (!d->d_itimer) {
258 		if ((d->d_ino_softlimit &&
259 		     (be64_to_cpu(d->d_icount) >=
260 		      be64_to_cpu(d->d_ino_softlimit))) ||
261 		    (d->d_ino_hardlimit &&
262 		     (be64_to_cpu(d->d_icount) >=
263 		      be64_to_cpu(d->d_ino_hardlimit)))) {
264 			d->d_itimer = cpu_to_be32(get_seconds() +
265 					mp->m_quotainfo->qi_itimelimit);
266 		} else {
267 			d->d_iwarns = 0;
268 		}
269 	} else {
270 		if ((!d->d_ino_softlimit ||
271 		     (be64_to_cpu(d->d_icount) <
272 		      be64_to_cpu(d->d_ino_softlimit)))  &&
273 		    (!d->d_ino_hardlimit ||
274 		     (be64_to_cpu(d->d_icount) <
275 		      be64_to_cpu(d->d_ino_hardlimit)))) {
276 			d->d_itimer = 0;
277 		}
278 	}
279 
280 	if (!d->d_rtbtimer) {
281 		if ((d->d_rtb_softlimit &&
282 		     (be64_to_cpu(d->d_rtbcount) >=
283 		      be64_to_cpu(d->d_rtb_softlimit))) ||
284 		    (d->d_rtb_hardlimit &&
285 		     (be64_to_cpu(d->d_rtbcount) >=
286 		      be64_to_cpu(d->d_rtb_hardlimit)))) {
287 			d->d_rtbtimer = cpu_to_be32(get_seconds() +
288 					mp->m_quotainfo->qi_rtbtimelimit);
289 		} else {
290 			d->d_rtbwarns = 0;
291 		}
292 	} else {
293 		if ((!d->d_rtb_softlimit ||
294 		     (be64_to_cpu(d->d_rtbcount) <
295 		      be64_to_cpu(d->d_rtb_softlimit))) &&
296 		    (!d->d_rtb_hardlimit ||
297 		     (be64_to_cpu(d->d_rtbcount) <
298 		      be64_to_cpu(d->d_rtb_hardlimit)))) {
299 			d->d_rtbtimer = 0;
300 		}
301 	}
302 }
303 
304 /*
305  * initialize a buffer full of dquots and log the whole thing
306  */
307 STATIC void
xfs_qm_init_dquot_blk(xfs_trans_t * tp,xfs_mount_t * mp,xfs_dqid_t id,uint type,xfs_buf_t * bp)308 xfs_qm_init_dquot_blk(
309 	xfs_trans_t	*tp,
310 	xfs_mount_t	*mp,
311 	xfs_dqid_t	id,
312 	uint		type,
313 	xfs_buf_t	*bp)
314 {
315 	struct xfs_quotainfo	*q = mp->m_quotainfo;
316 	xfs_dqblk_t	*d;
317 	int		curid, i;
318 
319 	ASSERT(tp);
320 	ASSERT(XFS_BUF_ISBUSY(bp));
321 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
322 
323 	d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
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
xfs_qm_dqalloc(xfs_trans_t ** tpp,xfs_mount_t * mp,xfs_dquot_t * dqp,xfs_inode_t * quotip,xfs_fileoff_t offset_fsb,xfs_buf_t ** O_bpp)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
xfs_qm_dqtobp(xfs_trans_t ** tpp,xfs_dquot_t * dqp,xfs_disk_dquot_t ** O_ddpp,xfs_buf_t ** O_bpp,uint flags)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_ISBUSY(bp));
537 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
538 
539 	/*
540 	 * calculate the location of the dquot inside the buffer.
541 	 */
542 	ddq = (struct xfs_disk_dquot *)(XFS_BUF_PTR(bp) + dqp->q_bufoffset);
543 
544 	/*
545 	 * A simple sanity check in case we got a corrupted dquot...
546 	 */
547 	error = xfs_qm_dqcheck(mp, ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
548 			   flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
549 			   "dqtobp");
550 	if (error) {
551 		if (!(flags & XFS_QMOPT_DQREPAIR)) {
552 			xfs_trans_brelse(tp, bp);
553 			return XFS_ERROR(EIO);
554 		}
555 		XFS_BUF_BUSY(bp); /* We dirtied this */
556 	}
557 
558 	*O_bpp = bp;
559 	*O_ddpp = ddq;
560 
561 	return (0);
562 }
563 
564 
565 /*
566  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
567  * and release the buffer immediately.
568  *
569  */
570 /* ARGSUSED */
571 STATIC int
xfs_qm_dqread(xfs_trans_t ** tpp,xfs_dqid_t id,xfs_dquot_t * dqp,uint flags)572 xfs_qm_dqread(
573 	xfs_trans_t	**tpp,
574 	xfs_dqid_t	id,
575 	xfs_dquot_t	*dqp,	/* dquot to get filled in */
576 	uint		flags)
577 {
578 	xfs_disk_dquot_t *ddqp;
579 	xfs_buf_t	 *bp;
580 	int		 error;
581 	xfs_trans_t	 *tp;
582 
583 	ASSERT(tpp);
584 
585 	trace_xfs_dqread(dqp);
586 
587 	/*
588 	 * get a pointer to the on-disk dquot and the buffer containing it
589 	 * dqp already knows its own type (GROUP/USER).
590 	 */
591 	if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
592 		return (error);
593 	}
594 	tp = *tpp;
595 
596 	/* copy everything from disk dquot to the incore dquot */
597 	memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
598 	ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
599 	xfs_qm_dquot_logitem_init(dqp);
600 
601 	/*
602 	 * Reservation counters are defined as reservation plus current usage
603 	 * to avoid having to add every time.
604 	 */
605 	dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
606 	dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
607 	dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
608 
609 	/* Mark the buf so that this will stay incore a little longer */
610 	XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
611 
612 	/*
613 	 * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
614 	 * So we need to release with xfs_trans_brelse().
615 	 * The strategy here is identical to that of inodes; we lock
616 	 * the dquot in xfs_qm_dqget() before making it accessible to
617 	 * others. This is because dquots, like inodes, need a good level of
618 	 * concurrency, and we don't want to take locks on the entire buffers
619 	 * for dquot accesses.
620 	 * Note also that the dquot buffer may even be dirty at this point, if
621 	 * this particular dquot was repaired. We still aren't afraid to
622 	 * brelse it because we have the changes incore.
623 	 */
624 	ASSERT(XFS_BUF_ISBUSY(bp));
625 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
626 	xfs_trans_brelse(tp, bp);
627 
628 	return (error);
629 }
630 
631 
632 /*
633  * allocate an incore dquot from the kernel heap,
634  * and fill its core with quota information kept on disk.
635  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
636  * if it wasn't already allocated.
637  */
638 STATIC int
xfs_qm_idtodq(xfs_mount_t * mp,xfs_dqid_t id,uint type,uint flags,xfs_dquot_t ** O_dqpp)639 xfs_qm_idtodq(
640 	xfs_mount_t	*mp,
641 	xfs_dqid_t	id,	 /* gid or uid, depending on type */
642 	uint		type,	 /* UDQUOT or GDQUOT */
643 	uint		flags,	 /* DQALLOC, DQREPAIR */
644 	xfs_dquot_t	**O_dqpp)/* OUT : incore dquot, not locked */
645 {
646 	xfs_dquot_t	*dqp;
647 	int		error;
648 	xfs_trans_t	*tp;
649 	int		cancelflags=0;
650 
651 	dqp = xfs_qm_dqinit(mp, id, type);
652 	tp = NULL;
653 	if (flags & XFS_QMOPT_DQALLOC) {
654 		tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
655 		error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
656 				XFS_WRITE_LOG_RES(mp) +
657 				BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 +
658 				128,
659 				0,
660 				XFS_TRANS_PERM_LOG_RES,
661 				XFS_WRITE_LOG_COUNT);
662 		if (error) {
663 			cancelflags = 0;
664 			goto error0;
665 		}
666 		cancelflags = XFS_TRANS_RELEASE_LOG_RES;
667 	}
668 
669 	/*
670 	 * Read it from disk; xfs_dqread() takes care of
671 	 * all the necessary initialization of dquot's fields (locks, etc)
672 	 */
673 	if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
674 		/*
675 		 * This can happen if quotas got turned off (ESRCH),
676 		 * or if the dquot didn't exist on disk and we ask to
677 		 * allocate (ENOENT).
678 		 */
679 		trace_xfs_dqread_fail(dqp);
680 		cancelflags |= XFS_TRANS_ABORT;
681 		goto error0;
682 	}
683 	if (tp) {
684 		if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
685 			goto error1;
686 	}
687 
688 	*O_dqpp = dqp;
689 	return (0);
690 
691  error0:
692 	ASSERT(error);
693 	if (tp)
694 		xfs_trans_cancel(tp, cancelflags);
695  error1:
696 	xfs_qm_dqdestroy(dqp);
697 	*O_dqpp = NULL;
698 	return (error);
699 }
700 
701 /*
702  * Lookup a dquot in the incore dquot hashtable. We keep two separate
703  * hashtables for user and group dquots; and, these are global tables
704  * inside the XQM, not per-filesystem tables.
705  * The hash chain must be locked by caller, and it is left locked
706  * on return. Returning dquot is locked.
707  */
708 STATIC int
xfs_qm_dqlookup(xfs_mount_t * mp,xfs_dqid_t id,xfs_dqhash_t * qh,xfs_dquot_t ** O_dqpp)709 xfs_qm_dqlookup(
710 	xfs_mount_t		*mp,
711 	xfs_dqid_t		id,
712 	xfs_dqhash_t		*qh,
713 	xfs_dquot_t		**O_dqpp)
714 {
715 	xfs_dquot_t		*dqp;
716 	uint			flist_locked;
717 
718 	ASSERT(mutex_is_locked(&qh->qh_lock));
719 
720 	flist_locked = B_FALSE;
721 
722 	/*
723 	 * Traverse the hashchain looking for a match
724 	 */
725 	list_for_each_entry(dqp, &qh->qh_list, q_hashlist) {
726 		/*
727 		 * We already have the hashlock. We don't need the
728 		 * dqlock to look at the id field of the dquot, since the
729 		 * id can't be modified without the hashlock anyway.
730 		 */
731 		if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
732 			trace_xfs_dqlookup_found(dqp);
733 
734 			/*
735 			 * All in core dquots must be on the dqlist of mp
736 			 */
737 			ASSERT(!list_empty(&dqp->q_mplist));
738 
739 			xfs_dqlock(dqp);
740 			if (dqp->q_nrefs == 0) {
741 				ASSERT(!list_empty(&dqp->q_freelist));
742 				if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
743 					trace_xfs_dqlookup_want(dqp);
744 
745 					/*
746 					 * We may have raced with dqreclaim_one()
747 					 * (and lost). So, flag that we don't
748 					 * want the dquot to be reclaimed.
749 					 */
750 					dqp->dq_flags |= XFS_DQ_WANT;
751 					xfs_dqunlock(dqp);
752 					mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
753 					xfs_dqlock(dqp);
754 					dqp->dq_flags &= ~(XFS_DQ_WANT);
755 				}
756 				flist_locked = B_TRUE;
757 			}
758 
759 			/*
760 			 * id couldn't have changed; we had the hashlock all
761 			 * along
762 			 */
763 			ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
764 
765 			if (flist_locked) {
766 				if (dqp->q_nrefs != 0) {
767 					mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
768 					flist_locked = B_FALSE;
769 				} else {
770 					/* take it off the freelist */
771 					trace_xfs_dqlookup_freelist(dqp);
772 					list_del_init(&dqp->q_freelist);
773 					xfs_Gqm->qm_dqfrlist_cnt--;
774 				}
775 			}
776 
777 			XFS_DQHOLD(dqp);
778 
779 			if (flist_locked)
780 				mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
781 			/*
782 			 * move the dquot to the front of the hashchain
783 			 */
784 			ASSERT(mutex_is_locked(&qh->qh_lock));
785 			list_move(&dqp->q_hashlist, &qh->qh_list);
786 			trace_xfs_dqlookup_done(dqp);
787 			*O_dqpp = dqp;
788 			return 0;
789 		}
790 	}
791 
792 	*O_dqpp = NULL;
793 	ASSERT(mutex_is_locked(&qh->qh_lock));
794 	return (1);
795 }
796 
797 /*
798  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
799  * a locked dquot, doing an allocation (if requested) as needed.
800  * When both an inode and an id are given, the inode's id takes precedence.
801  * That is, if the id changes while we don't hold the ilock inside this
802  * function, the new dquot is returned, not necessarily the one requested
803  * in the id argument.
804  */
805 int
xfs_qm_dqget(xfs_mount_t * mp,xfs_inode_t * ip,xfs_dqid_t id,uint type,uint flags,xfs_dquot_t ** O_dqpp)806 xfs_qm_dqget(
807 	xfs_mount_t	*mp,
808 	xfs_inode_t	*ip,	  /* locked inode (optional) */
809 	xfs_dqid_t	id,	  /* uid/projid/gid depending on type */
810 	uint		type,	  /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
811 	uint		flags,	  /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
812 	xfs_dquot_t	**O_dqpp) /* OUT : locked incore dquot */
813 {
814 	xfs_dquot_t	*dqp;
815 	xfs_dqhash_t	*h;
816 	uint		version;
817 	int		error;
818 
819 	ASSERT(XFS_IS_QUOTA_RUNNING(mp));
820 	if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
821 	    (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
822 	    (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
823 		return (ESRCH);
824 	}
825 	h = XFS_DQ_HASH(mp, id, type);
826 
827 #ifdef DEBUG
828 	if (xfs_do_dqerror) {
829 		if ((xfs_dqerror_target == mp->m_ddev_targp) &&
830 		    (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
831 			xfs_debug(mp, "Returning error in dqget");
832 			return (EIO);
833 		}
834 	}
835 #endif
836 
837  again:
838 
839 #ifdef DEBUG
840 	ASSERT(type == XFS_DQ_USER ||
841 	       type == XFS_DQ_PROJ ||
842 	       type == XFS_DQ_GROUP);
843 	if (ip) {
844 		ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
845 		if (type == XFS_DQ_USER)
846 			ASSERT(ip->i_udquot == NULL);
847 		else
848 			ASSERT(ip->i_gdquot == NULL);
849 	}
850 #endif
851 	mutex_lock(&h->qh_lock);
852 
853 	/*
854 	 * Look in the cache (hashtable).
855 	 * The chain is kept locked during lookup.
856 	 */
857 	if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
858 		XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
859 		/*
860 		 * The dquot was found, moved to the front of the chain,
861 		 * taken off the freelist if it was on it, and locked
862 		 * at this point. Just unlock the hashchain and return.
863 		 */
864 		ASSERT(*O_dqpp);
865 		ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
866 		mutex_unlock(&h->qh_lock);
867 		trace_xfs_dqget_hit(*O_dqpp);
868 		return (0);	/* success */
869 	}
870 	XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
871 
872 	/*
873 	 * Dquot cache miss. We don't want to keep the inode lock across
874 	 * a (potential) disk read. Also we don't want to deal with the lock
875 	 * ordering between quotainode and this inode. OTOH, dropping the inode
876 	 * lock here means dealing with a chown that can happen before
877 	 * we re-acquire the lock.
878 	 */
879 	if (ip)
880 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
881 	/*
882 	 * Save the hashchain version stamp, and unlock the chain, so that
883 	 * we don't keep the lock across a disk read
884 	 */
885 	version = h->qh_version;
886 	mutex_unlock(&h->qh_lock);
887 
888 	/*
889 	 * Allocate the dquot on the kernel heap, and read the ondisk
890 	 * portion off the disk. Also, do all the necessary initialization
891 	 * This can return ENOENT if dquot didn't exist on disk and we didn't
892 	 * ask it to allocate; ESRCH if quotas got turned off suddenly.
893 	 */
894 	if ((error = xfs_qm_idtodq(mp, id, type,
895 				  flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
896 					   XFS_QMOPT_DOWARN),
897 				  &dqp))) {
898 		if (ip)
899 			xfs_ilock(ip, XFS_ILOCK_EXCL);
900 		return (error);
901 	}
902 
903 	/*
904 	 * See if this is mount code calling to look at the overall quota limits
905 	 * which are stored in the id == 0 user or group's dquot.
906 	 * Since we may not have done a quotacheck by this point, just return
907 	 * the dquot without attaching it to any hashtables, lists, etc, or even
908 	 * taking a reference.
909 	 * The caller must dqdestroy this once done.
910 	 */
911 	if (flags & XFS_QMOPT_DQSUSER) {
912 		ASSERT(id == 0);
913 		ASSERT(! ip);
914 		goto dqret;
915 	}
916 
917 	/*
918 	 * Dquot lock comes after hashlock in the lock ordering
919 	 */
920 	if (ip) {
921 		xfs_ilock(ip, XFS_ILOCK_EXCL);
922 
923 		/*
924 		 * A dquot could be attached to this inode by now, since
925 		 * we had dropped the ilock.
926 		 */
927 		if (type == XFS_DQ_USER) {
928 			if (!XFS_IS_UQUOTA_ON(mp)) {
929 				/* inode stays locked on return */
930 				xfs_qm_dqdestroy(dqp);
931 				return XFS_ERROR(ESRCH);
932 			}
933 			if (ip->i_udquot) {
934 				xfs_qm_dqdestroy(dqp);
935 				dqp = ip->i_udquot;
936 				xfs_dqlock(dqp);
937 				goto dqret;
938 			}
939 		} else {
940 			if (!XFS_IS_OQUOTA_ON(mp)) {
941 				/* inode stays locked on return */
942 				xfs_qm_dqdestroy(dqp);
943 				return XFS_ERROR(ESRCH);
944 			}
945 			if (ip->i_gdquot) {
946 				xfs_qm_dqdestroy(dqp);
947 				dqp = ip->i_gdquot;
948 				xfs_dqlock(dqp);
949 				goto dqret;
950 			}
951 		}
952 	}
953 
954 	/*
955 	 * Hashlock comes after ilock in lock order
956 	 */
957 	mutex_lock(&h->qh_lock);
958 	if (version != h->qh_version) {
959 		xfs_dquot_t *tmpdqp;
960 		/*
961 		 * Now, see if somebody else put the dquot in the
962 		 * hashtable before us. This can happen because we didn't
963 		 * keep the hashchain lock. We don't have to worry about
964 		 * lock order between the two dquots here since dqp isn't
965 		 * on any findable lists yet.
966 		 */
967 		if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
968 			/*
969 			 * Duplicate found. Just throw away the new dquot
970 			 * and start over.
971 			 */
972 			xfs_qm_dqput(tmpdqp);
973 			mutex_unlock(&h->qh_lock);
974 			xfs_qm_dqdestroy(dqp);
975 			XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
976 			goto again;
977 		}
978 	}
979 
980 	/*
981 	 * Put the dquot at the beginning of the hash-chain and mp's list
982 	 * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
983 	 */
984 	ASSERT(mutex_is_locked(&h->qh_lock));
985 	dqp->q_hash = h;
986 	list_add(&dqp->q_hashlist, &h->qh_list);
987 	h->qh_version++;
988 
989 	/*
990 	 * Attach this dquot to this filesystem's list of all dquots,
991 	 * kept inside the mount structure in m_quotainfo field
992 	 */
993 	mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
994 
995 	/*
996 	 * We return a locked dquot to the caller, with a reference taken
997 	 */
998 	xfs_dqlock(dqp);
999 	dqp->q_nrefs = 1;
1000 
1001 	list_add(&dqp->q_mplist, &mp->m_quotainfo->qi_dqlist);
1002 	mp->m_quotainfo->qi_dquots++;
1003 	mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1004 	mutex_unlock(&h->qh_lock);
1005  dqret:
1006 	ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1007 	trace_xfs_dqget_miss(dqp);
1008 	*O_dqpp = dqp;
1009 	return (0);
1010 }
1011 
1012 
1013 /*
1014  * Release a reference to the dquot (decrement ref-count)
1015  * and unlock it. If there is a group quota attached to this
1016  * dquot, carefully release that too without tripping over
1017  * deadlocks'n'stuff.
1018  */
1019 void
xfs_qm_dqput(xfs_dquot_t * dqp)1020 xfs_qm_dqput(
1021 	xfs_dquot_t	*dqp)
1022 {
1023 	xfs_dquot_t	*gdqp;
1024 
1025 	ASSERT(dqp->q_nrefs > 0);
1026 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
1027 
1028 	trace_xfs_dqput(dqp);
1029 
1030 	if (dqp->q_nrefs != 1) {
1031 		dqp->q_nrefs--;
1032 		xfs_dqunlock(dqp);
1033 		return;
1034 	}
1035 
1036 	/*
1037 	 * drop the dqlock and acquire the freelist and dqlock
1038 	 * in the right order; but try to get it out-of-order first
1039 	 */
1040 	if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
1041 		trace_xfs_dqput_wait(dqp);
1042 		xfs_dqunlock(dqp);
1043 		mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1044 		xfs_dqlock(dqp);
1045 	}
1046 
1047 	while (1) {
1048 		gdqp = NULL;
1049 
1050 		/* We can't depend on nrefs being == 1 here */
1051 		if (--dqp->q_nrefs == 0) {
1052 			trace_xfs_dqput_free(dqp);
1053 
1054 			list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
1055 			xfs_Gqm->qm_dqfrlist_cnt++;
1056 
1057 			/*
1058 			 * If we just added a udquot to the freelist, then
1059 			 * we want to release the gdquot reference that
1060 			 * it (probably) has. Otherwise it'll keep the
1061 			 * gdquot from getting reclaimed.
1062 			 */
1063 			if ((gdqp = dqp->q_gdquot)) {
1064 				/*
1065 				 * Avoid a recursive dqput call
1066 				 */
1067 				xfs_dqlock(gdqp);
1068 				dqp->q_gdquot = NULL;
1069 			}
1070 		}
1071 		xfs_dqunlock(dqp);
1072 
1073 		/*
1074 		 * If we had a group quota inside the user quota as a hint,
1075 		 * release it now.
1076 		 */
1077 		if (! gdqp)
1078 			break;
1079 		dqp = gdqp;
1080 	}
1081 	mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1082 }
1083 
1084 /*
1085  * Release a dquot. Flush it if dirty, then dqput() it.
1086  * dquot must not be locked.
1087  */
1088 void
xfs_qm_dqrele(xfs_dquot_t * dqp)1089 xfs_qm_dqrele(
1090 	xfs_dquot_t	*dqp)
1091 {
1092 	if (!dqp)
1093 		return;
1094 
1095 	trace_xfs_dqrele(dqp);
1096 
1097 	xfs_dqlock(dqp);
1098 	/*
1099 	 * We don't care to flush it if the dquot is dirty here.
1100 	 * That will create stutters that we want to avoid.
1101 	 * Instead we do a delayed write when we try to reclaim
1102 	 * a dirty dquot. Also xfs_sync will take part of the burden...
1103 	 */
1104 	xfs_qm_dqput(dqp);
1105 }
1106 
1107 /*
1108  * This is the dquot flushing I/O completion routine.  It is called
1109  * from interrupt level when the buffer containing the dquot is
1110  * flushed to disk.  It is responsible for removing the dquot logitem
1111  * from the AIL if it has not been re-logged, and unlocking the dquot's
1112  * flush lock. This behavior is very similar to that of inodes..
1113  */
1114 STATIC void
xfs_qm_dqflush_done(struct xfs_buf * bp,struct xfs_log_item * lip)1115 xfs_qm_dqflush_done(
1116 	struct xfs_buf		*bp,
1117 	struct xfs_log_item	*lip)
1118 {
1119 	xfs_dq_logitem_t	*qip = (struct xfs_dq_logitem *)lip;
1120 	xfs_dquot_t		*dqp = qip->qli_dquot;
1121 	struct xfs_ail		*ailp = lip->li_ailp;
1122 
1123 	/*
1124 	 * We only want to pull the item from the AIL if its
1125 	 * location in the log has not changed since we started the flush.
1126 	 * Thus, we only bother if the dquot's lsn has
1127 	 * not changed. First we check the lsn outside the lock
1128 	 * since it's cheaper, and then we recheck while
1129 	 * holding the lock before removing the dquot from the AIL.
1130 	 */
1131 	if ((lip->li_flags & XFS_LI_IN_AIL) &&
1132 	    lip->li_lsn == qip->qli_flush_lsn) {
1133 
1134 		/* xfs_trans_ail_delete() drops the AIL lock. */
1135 		spin_lock(&ailp->xa_lock);
1136 		if (lip->li_lsn == qip->qli_flush_lsn)
1137 			xfs_trans_ail_delete(ailp, lip);
1138 		else
1139 			spin_unlock(&ailp->xa_lock);
1140 	}
1141 
1142 	/*
1143 	 * Release the dq's flush lock since we're done with it.
1144 	 */
1145 	xfs_dqfunlock(dqp);
1146 }
1147 
1148 /*
1149  * Write a modified dquot to disk.
1150  * The dquot must be locked and the flush lock too taken by caller.
1151  * The flush lock will not be unlocked until the dquot reaches the disk,
1152  * but the dquot is free to be unlocked and modified by the caller
1153  * in the interim. Dquot is still locked on return. This behavior is
1154  * identical to that of inodes.
1155  */
1156 int
xfs_qm_dqflush(xfs_dquot_t * dqp,uint flags)1157 xfs_qm_dqflush(
1158 	xfs_dquot_t		*dqp,
1159 	uint			flags)
1160 {
1161 	struct xfs_mount	*mp = dqp->q_mount;
1162 	struct xfs_buf		*bp;
1163 	struct xfs_disk_dquot	*ddqp;
1164 	int			error;
1165 
1166 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
1167 	ASSERT(!completion_done(&dqp->q_flush));
1168 
1169 	trace_xfs_dqflush(dqp);
1170 
1171 	/*
1172 	 * If not dirty, or it's pinned and we are not supposed to block, nada.
1173 	 */
1174 	if (!XFS_DQ_IS_DIRTY(dqp) ||
1175 	    (!(flags & SYNC_WAIT) && atomic_read(&dqp->q_pincount) > 0)) {
1176 		xfs_dqfunlock(dqp);
1177 		return 0;
1178 	}
1179 	xfs_qm_dqunpin_wait(dqp);
1180 
1181 	/*
1182 	 * This may have been unpinned because the filesystem is shutting
1183 	 * down forcibly. If that's the case we must not write this dquot
1184 	 * to disk, because the log record didn't make it to disk!
1185 	 */
1186 	if (XFS_FORCED_SHUTDOWN(mp)) {
1187 		dqp->dq_flags &= ~XFS_DQ_DIRTY;
1188 		xfs_dqfunlock(dqp);
1189 		return XFS_ERROR(EIO);
1190 	}
1191 
1192 	/*
1193 	 * Get the buffer containing the on-disk dquot
1194 	 */
1195 	error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1196 				   mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1197 	if (error) {
1198 		ASSERT(error != ENOENT);
1199 		xfs_dqfunlock(dqp);
1200 		return error;
1201 	}
1202 
1203 	/*
1204 	 * Calculate the location of the dquot inside the buffer.
1205 	 */
1206 	ddqp = (struct xfs_disk_dquot *)(XFS_BUF_PTR(bp) + dqp->q_bufoffset);
1207 
1208 	/*
1209 	 * A simple sanity check in case we got a corrupted dquot..
1210 	 */
1211 	error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1212 			   XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1213 	if (error) {
1214 		xfs_buf_relse(bp);
1215 		xfs_dqfunlock(dqp);
1216 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1217 		return XFS_ERROR(EIO);
1218 	}
1219 
1220 	/* This is the only portion of data that needs to persist */
1221 	memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1222 
1223 	/*
1224 	 * Clear the dirty field and remember the flush lsn for later use.
1225 	 */
1226 	dqp->dq_flags &= ~XFS_DQ_DIRTY;
1227 
1228 	xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1229 					&dqp->q_logitem.qli_item.li_lsn);
1230 
1231 	/*
1232 	 * Attach an iodone routine so that we can remove this dquot from the
1233 	 * AIL and release the flush lock once the dquot is synced to disk.
1234 	 */
1235 	xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1236 				  &dqp->q_logitem.qli_item);
1237 
1238 	/*
1239 	 * If the buffer is pinned then push on the log so we won't
1240 	 * get stuck waiting in the write for too long.
1241 	 */
1242 	if (XFS_BUF_ISPINNED(bp)) {
1243 		trace_xfs_dqflush_force(dqp);
1244 		xfs_log_force(mp, 0);
1245 	}
1246 
1247 	if (flags & SYNC_WAIT)
1248 		error = xfs_bwrite(mp, bp);
1249 	else
1250 		xfs_bdwrite(mp, bp);
1251 
1252 	trace_xfs_dqflush_done(dqp);
1253 
1254 	/*
1255 	 * dqp is still locked, but caller is free to unlock it now.
1256 	 */
1257 	return error;
1258 
1259 }
1260 
1261 int
xfs_qm_dqlock_nowait(xfs_dquot_t * dqp)1262 xfs_qm_dqlock_nowait(
1263 	xfs_dquot_t *dqp)
1264 {
1265 	return mutex_trylock(&dqp->q_qlock);
1266 }
1267 
1268 void
xfs_dqlock(xfs_dquot_t * dqp)1269 xfs_dqlock(
1270 	xfs_dquot_t *dqp)
1271 {
1272 	mutex_lock(&dqp->q_qlock);
1273 }
1274 
1275 void
xfs_dqunlock(xfs_dquot_t * dqp)1276 xfs_dqunlock(
1277 	xfs_dquot_t *dqp)
1278 {
1279 	mutex_unlock(&(dqp->q_qlock));
1280 	if (dqp->q_logitem.qli_dquot == dqp) {
1281 		/* Once was dqp->q_mount, but might just have been cleared */
1282 		xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1283 					(xfs_log_item_t*)&(dqp->q_logitem));
1284 	}
1285 }
1286 
1287 
1288 void
xfs_dqunlock_nonotify(xfs_dquot_t * dqp)1289 xfs_dqunlock_nonotify(
1290 	xfs_dquot_t *dqp)
1291 {
1292 	mutex_unlock(&(dqp->q_qlock));
1293 }
1294 
1295 /*
1296  * Lock two xfs_dquot structures.
1297  *
1298  * To avoid deadlocks we always lock the quota structure with
1299  * the lowerd id first.
1300  */
1301 void
xfs_dqlock2(xfs_dquot_t * d1,xfs_dquot_t * d2)1302 xfs_dqlock2(
1303 	xfs_dquot_t	*d1,
1304 	xfs_dquot_t	*d2)
1305 {
1306 	if (d1 && d2) {
1307 		ASSERT(d1 != d2);
1308 		if (be32_to_cpu(d1->q_core.d_id) >
1309 		    be32_to_cpu(d2->q_core.d_id)) {
1310 			mutex_lock(&d2->q_qlock);
1311 			mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1312 		} else {
1313 			mutex_lock(&d1->q_qlock);
1314 			mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1315 		}
1316 	} else if (d1) {
1317 		mutex_lock(&d1->q_qlock);
1318 	} else if (d2) {
1319 		mutex_lock(&d2->q_qlock);
1320 	}
1321 }
1322 
1323 
1324 /*
1325  * Take a dquot out of the mount's dqlist as well as the hashlist.
1326  * This is called via unmount as well as quotaoff, and the purge
1327  * will always succeed unless there are soft (temp) references
1328  * outstanding.
1329  *
1330  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1331  * that we're returning! XXXsup - not cool.
1332  */
1333 /* ARGSUSED */
1334 int
xfs_qm_dqpurge(xfs_dquot_t * dqp)1335 xfs_qm_dqpurge(
1336 	xfs_dquot_t	*dqp)
1337 {
1338 	xfs_dqhash_t	*qh = dqp->q_hash;
1339 	xfs_mount_t	*mp = dqp->q_mount;
1340 
1341 	ASSERT(mutex_is_locked(&mp->m_quotainfo->qi_dqlist_lock));
1342 	ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1343 
1344 	xfs_dqlock(dqp);
1345 	/*
1346 	 * We really can't afford to purge a dquot that is
1347 	 * referenced, because these are hard refs.
1348 	 * It shouldn't happen in general because we went thru _all_ inodes in
1349 	 * dqrele_all_inodes before calling this and didn't let the mountlock go.
1350 	 * However it is possible that we have dquots with temporary
1351 	 * references that are not attached to an inode. e.g. see xfs_setattr().
1352 	 */
1353 	if (dqp->q_nrefs != 0) {
1354 		xfs_dqunlock(dqp);
1355 		mutex_unlock(&dqp->q_hash->qh_lock);
1356 		return (1);
1357 	}
1358 
1359 	ASSERT(!list_empty(&dqp->q_freelist));
1360 
1361 	/*
1362 	 * If we're turning off quotas, we have to make sure that, for
1363 	 * example, we don't delete quota disk blocks while dquots are
1364 	 * in the process of getting written to those disk blocks.
1365 	 * This dquot might well be on AIL, and we can't leave it there
1366 	 * if we're turning off quotas. Basically, we need this flush
1367 	 * lock, and are willing to block on it.
1368 	 */
1369 	if (!xfs_dqflock_nowait(dqp)) {
1370 		/*
1371 		 * Block on the flush lock after nudging dquot buffer,
1372 		 * if it is incore.
1373 		 */
1374 		xfs_qm_dqflock_pushbuf_wait(dqp);
1375 	}
1376 
1377 	/*
1378 	 * XXXIf we're turning this type of quotas off, we don't care
1379 	 * about the dirty metadata sitting in this dquot. OTOH, if
1380 	 * we're unmounting, we do care, so we flush it and wait.
1381 	 */
1382 	if (XFS_DQ_IS_DIRTY(dqp)) {
1383 		int	error;
1384 
1385 		/* dqflush unlocks dqflock */
1386 		/*
1387 		 * Given that dqpurge is a very rare occurrence, it is OK
1388 		 * that we're holding the hashlist and mplist locks
1389 		 * across the disk write. But, ... XXXsup
1390 		 *
1391 		 * We don't care about getting disk errors here. We need
1392 		 * to purge this dquot anyway, so we go ahead regardless.
1393 		 */
1394 		error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1395 		if (error)
1396 			xfs_warn(mp, "%s: dquot %p flush failed",
1397 				__func__, dqp);
1398 		xfs_dqflock(dqp);
1399 	}
1400 	ASSERT(atomic_read(&dqp->q_pincount) == 0);
1401 	ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1402 	       !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1403 
1404 	list_del_init(&dqp->q_hashlist);
1405 	qh->qh_version++;
1406 	list_del_init(&dqp->q_mplist);
1407 	mp->m_quotainfo->qi_dqreclaims++;
1408 	mp->m_quotainfo->qi_dquots--;
1409 	/*
1410 	 * XXX Move this to the front of the freelist, if we can get the
1411 	 * freelist lock.
1412 	 */
1413 	ASSERT(!list_empty(&dqp->q_freelist));
1414 
1415 	dqp->q_mount = NULL;
1416 	dqp->q_hash = NULL;
1417 	dqp->dq_flags = XFS_DQ_INACTIVE;
1418 	memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1419 	xfs_dqfunlock(dqp);
1420 	xfs_dqunlock(dqp);
1421 	mutex_unlock(&qh->qh_lock);
1422 	return (0);
1423 }
1424 
1425 
1426 #ifdef QUOTADEBUG
1427 void
xfs_qm_dqprint(xfs_dquot_t * dqp)1428 xfs_qm_dqprint(xfs_dquot_t *dqp)
1429 {
1430 	struct xfs_mount	*mp = dqp->q_mount;
1431 
1432 	xfs_debug(mp, "-----------KERNEL DQUOT----------------");
1433 	xfs_debug(mp, "---- dquotID =  %d",
1434 		(int)be32_to_cpu(dqp->q_core.d_id));
1435 	xfs_debug(mp, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1436 	xfs_debug(mp, "---- fs      =  0x%p", dqp->q_mount);
1437 	xfs_debug(mp, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1438 	xfs_debug(mp, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1439 	xfs_debug(mp, "---- blkhlimit =  %Lu (0x%x)",
1440 		be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1441 		(int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1442 	xfs_debug(mp, "---- blkslimit =  %Lu (0x%x)",
1443 		be64_to_cpu(dqp->q_core.d_blk_softlimit),
1444 		(int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1445 	xfs_debug(mp, "---- inohlimit =  %Lu (0x%x)",
1446 		be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1447 		(int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1448 	xfs_debug(mp, "---- inoslimit =  %Lu (0x%x)",
1449 		be64_to_cpu(dqp->q_core.d_ino_softlimit),
1450 		(int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1451 	xfs_debug(mp, "---- bcount  =  %Lu (0x%x)",
1452 		be64_to_cpu(dqp->q_core.d_bcount),
1453 		(int)be64_to_cpu(dqp->q_core.d_bcount));
1454 	xfs_debug(mp, "---- icount  =  %Lu (0x%x)",
1455 		be64_to_cpu(dqp->q_core.d_icount),
1456 		(int)be64_to_cpu(dqp->q_core.d_icount));
1457 	xfs_debug(mp, "---- btimer  =  %d",
1458 		(int)be32_to_cpu(dqp->q_core.d_btimer));
1459 	xfs_debug(mp, "---- itimer  =  %d",
1460 		(int)be32_to_cpu(dqp->q_core.d_itimer));
1461 	xfs_debug(mp, "---------------------------");
1462 }
1463 #endif
1464 
1465 /*
1466  * Give the buffer a little push if it is incore and
1467  * wait on the flush lock.
1468  */
1469 void
xfs_qm_dqflock_pushbuf_wait(xfs_dquot_t * dqp)1470 xfs_qm_dqflock_pushbuf_wait(
1471 	xfs_dquot_t	*dqp)
1472 {
1473 	xfs_mount_t	*mp = dqp->q_mount;
1474 	xfs_buf_t	*bp;
1475 
1476 	/*
1477 	 * Check to see if the dquot has been flushed delayed
1478 	 * write.  If so, grab its buffer and send it
1479 	 * out immediately.  We'll be able to acquire
1480 	 * the flush lock when the I/O completes.
1481 	 */
1482 	bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1483 			mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1484 	if (!bp)
1485 		goto out_lock;
1486 
1487 	if (XFS_BUF_ISDELAYWRITE(bp)) {
1488 		if (XFS_BUF_ISPINNED(bp))
1489 			xfs_log_force(mp, 0);
1490 		xfs_buf_delwri_promote(bp);
1491 		wake_up_process(bp->b_target->bt_task);
1492 	}
1493 	xfs_buf_relse(bp);
1494 out_lock:
1495 	xfs_dqflock(dqp);
1496 }
1497