Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2003 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_format.h"
9#include "xfs_log_format.h"
10#include "xfs_shared.h"
11#include "xfs_trans_resv.h"
12#include "xfs_bit.h"
13#include "xfs_mount.h"
14#include "xfs_defer.h"
15#include "xfs_inode.h"
16#include "xfs_bmap.h"
17#include "xfs_quota.h"
18#include "xfs_trans.h"
19#include "xfs_buf_item.h"
20#include "xfs_trans_space.h"
21#include "xfs_trans_priv.h"
22#include "xfs_qm.h"
23#include "xfs_trace.h"
24#include "xfs_log.h"
25#include "xfs_bmap_btree.h"
26#include "xfs_error.h"
27
28/*
29 * Lock order:
30 *
31 * ip->i_lock
32 * qi->qi_tree_lock
33 * dquot->q_qlock (xfs_dqlock() and friends)
34 * dquot->q_flush (xfs_dqflock() and friends)
35 * qi->qi_lru_lock
36 *
37 * If two dquots need to be locked the order is user before group/project,
38 * otherwise by the lowest id first, see xfs_dqlock2.
39 */
40
41struct kmem_cache *xfs_dqtrx_cache;
42static struct kmem_cache *xfs_dquot_cache;
43
44static struct lock_class_key xfs_dquot_group_class;
45static struct lock_class_key xfs_dquot_project_class;
46
47/*
48 * This is called to free all the memory associated with a dquot
49 */
50void
51xfs_qm_dqdestroy(
52 struct xfs_dquot *dqp)
53{
54 ASSERT(list_empty(&dqp->q_lru));
55
56 kmem_free(dqp->q_logitem.qli_item.li_lv_shadow);
57 mutex_destroy(&dqp->q_qlock);
58
59 XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
60 kmem_cache_free(xfs_dquot_cache, dqp);
61}
62
63/*
64 * If default limits are in force, push them into the dquot now.
65 * We overwrite the dquot limits only if they are zero and this
66 * is not the root dquot.
67 */
68void
69xfs_qm_adjust_dqlimits(
70 struct xfs_dquot *dq)
71{
72 struct xfs_mount *mp = dq->q_mount;
73 struct xfs_quotainfo *q = mp->m_quotainfo;
74 struct xfs_def_quota *defq;
75 int prealloc = 0;
76
77 ASSERT(dq->q_id);
78 defq = xfs_get_defquota(q, xfs_dquot_type(dq));
79
80 if (!dq->q_blk.softlimit) {
81 dq->q_blk.softlimit = defq->blk.soft;
82 prealloc = 1;
83 }
84 if (!dq->q_blk.hardlimit) {
85 dq->q_blk.hardlimit = defq->blk.hard;
86 prealloc = 1;
87 }
88 if (!dq->q_ino.softlimit)
89 dq->q_ino.softlimit = defq->ino.soft;
90 if (!dq->q_ino.hardlimit)
91 dq->q_ino.hardlimit = defq->ino.hard;
92 if (!dq->q_rtb.softlimit)
93 dq->q_rtb.softlimit = defq->rtb.soft;
94 if (!dq->q_rtb.hardlimit)
95 dq->q_rtb.hardlimit = defq->rtb.hard;
96
97 if (prealloc)
98 xfs_dquot_set_prealloc_limits(dq);
99}
100
101/* Set the expiration time of a quota's grace period. */
102time64_t
103xfs_dquot_set_timeout(
104 struct xfs_mount *mp,
105 time64_t timeout)
106{
107 struct xfs_quotainfo *qi = mp->m_quotainfo;
108
109 return clamp_t(time64_t, timeout, qi->qi_expiry_min,
110 qi->qi_expiry_max);
111}
112
113/* Set the length of the default grace period. */
114time64_t
115xfs_dquot_set_grace_period(
116 time64_t grace)
117{
118 return clamp_t(time64_t, grace, XFS_DQ_GRACE_MIN, XFS_DQ_GRACE_MAX);
119}
120
121/*
122 * Determine if this quota counter is over either limit and set the quota
123 * timers as appropriate.
124 */
125static inline void
126xfs_qm_adjust_res_timer(
127 struct xfs_mount *mp,
128 struct xfs_dquot_res *res,
129 struct xfs_quota_limits *qlim)
130{
131 ASSERT(res->hardlimit == 0 || res->softlimit <= res->hardlimit);
132
133 if ((res->softlimit && res->count > res->softlimit) ||
134 (res->hardlimit && res->count > res->hardlimit)) {
135 if (res->timer == 0)
136 res->timer = xfs_dquot_set_timeout(mp,
137 ktime_get_real_seconds() + qlim->time);
138 } else {
139 res->timer = 0;
140 }
141}
142
143/*
144 * Check the limits and timers of a dquot and start or reset timers
145 * if necessary.
146 * This gets called even when quota enforcement is OFF, which makes our
147 * life a little less complicated. (We just don't reject any quota
148 * reservations in that case, when enforcement is off).
149 * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
150 * enforcement's off.
151 * In contrast, warnings are a little different in that they don't
152 * 'automatically' get started when limits get exceeded. They do
153 * get reset to zero, however, when we find the count to be under
154 * the soft limit (they are only ever set non-zero via userspace).
155 */
156void
157xfs_qm_adjust_dqtimers(
158 struct xfs_dquot *dq)
159{
160 struct xfs_mount *mp = dq->q_mount;
161 struct xfs_quotainfo *qi = mp->m_quotainfo;
162 struct xfs_def_quota *defq;
163
164 ASSERT(dq->q_id);
165 defq = xfs_get_defquota(qi, xfs_dquot_type(dq));
166
167 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_blk, &defq->blk);
168 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_ino, &defq->ino);
169 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_rtb, &defq->rtb);
170}
171
172/*
173 * initialize a buffer full of dquots and log the whole thing
174 */
175STATIC void
176xfs_qm_init_dquot_blk(
177 struct xfs_trans *tp,
178 struct xfs_mount *mp,
179 xfs_dqid_t id,
180 xfs_dqtype_t type,
181 struct xfs_buf *bp)
182{
183 struct xfs_quotainfo *q = mp->m_quotainfo;
184 struct xfs_dqblk *d;
185 xfs_dqid_t curid;
186 unsigned int qflag;
187 unsigned int blftype;
188 int i;
189
190 ASSERT(tp);
191 ASSERT(xfs_buf_islocked(bp));
192
193 switch (type) {
194 case XFS_DQTYPE_USER:
195 qflag = XFS_UQUOTA_CHKD;
196 blftype = XFS_BLF_UDQUOT_BUF;
197 break;
198 case XFS_DQTYPE_PROJ:
199 qflag = XFS_PQUOTA_CHKD;
200 blftype = XFS_BLF_PDQUOT_BUF;
201 break;
202 case XFS_DQTYPE_GROUP:
203 qflag = XFS_GQUOTA_CHKD;
204 blftype = XFS_BLF_GDQUOT_BUF;
205 break;
206 default:
207 ASSERT(0);
208 return;
209 }
210
211 d = bp->b_addr;
212
213 /*
214 * ID of the first dquot in the block - id's are zero based.
215 */
216 curid = id - (id % q->qi_dqperchunk);
217 memset(d, 0, BBTOB(q->qi_dqchunklen));
218 for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
219 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
220 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
221 d->dd_diskdq.d_id = cpu_to_be32(curid);
222 d->dd_diskdq.d_type = type;
223 if (curid > 0 && xfs_has_bigtime(mp))
224 d->dd_diskdq.d_type |= XFS_DQTYPE_BIGTIME;
225 if (xfs_has_crc(mp)) {
226 uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
227 xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
228 XFS_DQUOT_CRC_OFF);
229 }
230 }
231
232 xfs_trans_dquot_buf(tp, bp, blftype);
233
234 /*
235 * quotacheck uses delayed writes to update all the dquots on disk in an
236 * efficient manner instead of logging the individual dquot changes as
237 * they are made. However if we log the buffer allocated here and crash
238 * after quotacheck while the logged initialisation is still in the
239 * active region of the log, log recovery can replay the dquot buffer
240 * initialisation over the top of the checked dquots and corrupt quota
241 * accounting.
242 *
243 * To avoid this problem, quotacheck cannot log the initialised buffer.
244 * We must still dirty the buffer and write it back before the
245 * allocation transaction clears the log. Therefore, mark the buffer as
246 * ordered instead of logging it directly. This is safe for quotacheck
247 * because it detects and repairs allocated but initialized dquot blocks
248 * in the quota inodes.
249 */
250 if (!(mp->m_qflags & qflag))
251 xfs_trans_ordered_buf(tp, bp);
252 else
253 xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
254}
255
256/*
257 * Initialize the dynamic speculative preallocation thresholds. The lo/hi
258 * watermarks correspond to the soft and hard limits by default. If a soft limit
259 * is not specified, we use 95% of the hard limit.
260 */
261void
262xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
263{
264 uint64_t space;
265
266 dqp->q_prealloc_hi_wmark = dqp->q_blk.hardlimit;
267 dqp->q_prealloc_lo_wmark = dqp->q_blk.softlimit;
268 if (!dqp->q_prealloc_lo_wmark) {
269 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
270 do_div(dqp->q_prealloc_lo_wmark, 100);
271 dqp->q_prealloc_lo_wmark *= 95;
272 }
273
274 space = dqp->q_prealloc_hi_wmark;
275
276 do_div(space, 100);
277 dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
278 dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
279 dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
280}
281
282/*
283 * Ensure that the given in-core dquot has a buffer on disk backing it, and
284 * return the buffer locked and held. This is called when the bmapi finds a
285 * hole.
286 */
287STATIC int
288xfs_dquot_disk_alloc(
289 struct xfs_dquot *dqp,
290 struct xfs_buf **bpp)
291{
292 struct xfs_bmbt_irec map;
293 struct xfs_trans *tp;
294 struct xfs_mount *mp = dqp->q_mount;
295 struct xfs_buf *bp;
296 xfs_dqtype_t qtype = xfs_dquot_type(dqp);
297 struct xfs_inode *quotip = xfs_quota_inode(mp, qtype);
298 int nmaps = 1;
299 int error;
300
301 trace_xfs_dqalloc(dqp);
302
303 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
304 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
305 if (error)
306 return error;
307
308 xfs_ilock(quotip, XFS_ILOCK_EXCL);
309 xfs_trans_ijoin(tp, quotip, 0);
310
311 if (!xfs_this_quota_on(dqp->q_mount, qtype)) {
312 /*
313 * Return if this type of quotas is turned off while we didn't
314 * have an inode lock
315 */
316 error = -ESRCH;
317 goto err_cancel;
318 }
319
320 error = xfs_iext_count_may_overflow(quotip, XFS_DATA_FORK,
321 XFS_IEXT_ADD_NOSPLIT_CNT);
322 if (error == -EFBIG)
323 error = xfs_iext_count_upgrade(tp, quotip,
324 XFS_IEXT_ADD_NOSPLIT_CNT);
325 if (error)
326 goto err_cancel;
327
328 /* Create the block mapping. */
329 error = xfs_bmapi_write(tp, quotip, dqp->q_fileoffset,
330 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA, 0, &map,
331 &nmaps);
332 if (error)
333 goto err_cancel;
334
335 ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
336 ASSERT(nmaps == 1);
337 ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
338 (map.br_startblock != HOLESTARTBLOCK));
339
340 /*
341 * Keep track of the blkno to save a lookup later
342 */
343 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
344
345 /* now we can just get the buffer (there's nothing to read yet) */
346 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, dqp->q_blkno,
347 mp->m_quotainfo->qi_dqchunklen, 0, &bp);
348 if (error)
349 goto err_cancel;
350 bp->b_ops = &xfs_dquot_buf_ops;
351
352 /*
353 * Make a chunk of dquots out of this buffer and log
354 * the entire thing.
355 */
356 xfs_qm_init_dquot_blk(tp, mp, dqp->q_id, qtype, bp);
357 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
358
359 /*
360 * Hold the buffer and join it to the dfops so that we'll still own
361 * the buffer when we return to the caller. The buffer disposal on
362 * error must be paid attention to very carefully, as it has been
363 * broken since commit efa092f3d4c6 "[XFS] Fixes a bug in the quota
364 * code when allocating a new dquot record" in 2005, and the later
365 * conversion to xfs_defer_ops in commit 310a75a3c6c747 failed to keep
366 * the buffer locked across the _defer_finish call. We can now do
367 * this correctly with xfs_defer_bjoin.
368 *
369 * Above, we allocated a disk block for the dquot information and used
370 * get_buf to initialize the dquot. If the _defer_finish fails, the old
371 * transaction is gone but the new buffer is not joined or held to any
372 * transaction, so we must _buf_relse it.
373 *
374 * If everything succeeds, the caller of this function is returned a
375 * buffer that is locked and held to the transaction. The caller
376 * is responsible for unlocking any buffer passed back, either
377 * manually or by committing the transaction. On error, the buffer is
378 * released and not passed back.
379 *
380 * Keep the quota inode ILOCKed until after the transaction commit to
381 * maintain the atomicity of bmap/rmap updates.
382 */
383 xfs_trans_bhold(tp, bp);
384 error = xfs_trans_commit(tp);
385 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
386 if (error) {
387 xfs_buf_relse(bp);
388 return error;
389 }
390
391 *bpp = bp;
392 return 0;
393
394err_cancel:
395 xfs_trans_cancel(tp);
396 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
397 return error;
398}
399
400/*
401 * Read in the in-core dquot's on-disk metadata and return the buffer.
402 * Returns ENOENT to signal a hole.
403 */
404STATIC int
405xfs_dquot_disk_read(
406 struct xfs_mount *mp,
407 struct xfs_dquot *dqp,
408 struct xfs_buf **bpp)
409{
410 struct xfs_bmbt_irec map;
411 struct xfs_buf *bp;
412 xfs_dqtype_t qtype = xfs_dquot_type(dqp);
413 struct xfs_inode *quotip = xfs_quota_inode(mp, qtype);
414 uint lock_mode;
415 int nmaps = 1;
416 int error;
417
418 lock_mode = xfs_ilock_data_map_shared(quotip);
419 if (!xfs_this_quota_on(mp, qtype)) {
420 /*
421 * Return if this type of quotas is turned off while we
422 * didn't have the quota inode lock.
423 */
424 xfs_iunlock(quotip, lock_mode);
425 return -ESRCH;
426 }
427
428 /*
429 * Find the block map; no allocations yet
430 */
431 error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
432 XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
433 xfs_iunlock(quotip, lock_mode);
434 if (error)
435 return error;
436
437 ASSERT(nmaps == 1);
438 ASSERT(map.br_blockcount >= 1);
439 ASSERT(map.br_startblock != DELAYSTARTBLOCK);
440 if (map.br_startblock == HOLESTARTBLOCK)
441 return -ENOENT;
442
443 trace_xfs_dqtobp_read(dqp);
444
445 /*
446 * store the blkno etc so that we don't have to do the
447 * mapping all the time
448 */
449 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
450
451 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
452 mp->m_quotainfo->qi_dqchunklen, 0, &bp,
453 &xfs_dquot_buf_ops);
454 if (error) {
455 ASSERT(bp == NULL);
456 return error;
457 }
458
459 ASSERT(xfs_buf_islocked(bp));
460 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
461 *bpp = bp;
462
463 return 0;
464}
465
466/* Allocate and initialize everything we need for an incore dquot. */
467STATIC struct xfs_dquot *
468xfs_dquot_alloc(
469 struct xfs_mount *mp,
470 xfs_dqid_t id,
471 xfs_dqtype_t type)
472{
473 struct xfs_dquot *dqp;
474
475 dqp = kmem_cache_zalloc(xfs_dquot_cache, GFP_KERNEL | __GFP_NOFAIL);
476
477 dqp->q_type = type;
478 dqp->q_id = id;
479 dqp->q_mount = mp;
480 INIT_LIST_HEAD(&dqp->q_lru);
481 mutex_init(&dqp->q_qlock);
482 init_waitqueue_head(&dqp->q_pinwait);
483 dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
484 /*
485 * Offset of dquot in the (fixed sized) dquot chunk.
486 */
487 dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
488 sizeof(struct xfs_dqblk);
489
490 /*
491 * Because we want to use a counting completion, complete
492 * the flush completion once to allow a single access to
493 * the flush completion without blocking.
494 */
495 init_completion(&dqp->q_flush);
496 complete(&dqp->q_flush);
497
498 /*
499 * Make sure group quotas have a different lock class than user
500 * quotas.
501 */
502 switch (type) {
503 case XFS_DQTYPE_USER:
504 /* uses the default lock class */
505 break;
506 case XFS_DQTYPE_GROUP:
507 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
508 break;
509 case XFS_DQTYPE_PROJ:
510 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
511 break;
512 default:
513 ASSERT(0);
514 break;
515 }
516
517 xfs_qm_dquot_logitem_init(dqp);
518
519 XFS_STATS_INC(mp, xs_qm_dquot);
520 return dqp;
521}
522
523/* Check the ondisk dquot's id and type match what the incore dquot expects. */
524static bool
525xfs_dquot_check_type(
526 struct xfs_dquot *dqp,
527 struct xfs_disk_dquot *ddqp)
528{
529 uint8_t ddqp_type;
530 uint8_t dqp_type;
531
532 ddqp_type = ddqp->d_type & XFS_DQTYPE_REC_MASK;
533 dqp_type = xfs_dquot_type(dqp);
534
535 if (be32_to_cpu(ddqp->d_id) != dqp->q_id)
536 return false;
537
538 /*
539 * V5 filesystems always expect an exact type match. V4 filesystems
540 * expect an exact match for user dquots and for non-root group and
541 * project dquots.
542 */
543 if (xfs_has_crc(dqp->q_mount) ||
544 dqp_type == XFS_DQTYPE_USER || dqp->q_id != 0)
545 return ddqp_type == dqp_type;
546
547 /*
548 * V4 filesystems support either group or project quotas, but not both
549 * at the same time. The non-user quota file can be switched between
550 * group and project quota uses depending on the mount options, which
551 * means that we can encounter the other type when we try to load quota
552 * defaults. Quotacheck will soon reset the entire quota file
553 * (including the root dquot) anyway, but don't log scary corruption
554 * reports to dmesg.
555 */
556 return ddqp_type == XFS_DQTYPE_GROUP || ddqp_type == XFS_DQTYPE_PROJ;
557}
558
559/* Copy the in-core quota fields in from the on-disk buffer. */
560STATIC int
561xfs_dquot_from_disk(
562 struct xfs_dquot *dqp,
563 struct xfs_buf *bp)
564{
565 struct xfs_disk_dquot *ddqp = bp->b_addr + dqp->q_bufoffset;
566
567 /*
568 * Ensure that we got the type and ID we were looking for.
569 * Everything else was checked by the dquot buffer verifier.
570 */
571 if (!xfs_dquot_check_type(dqp, ddqp)) {
572 xfs_alert_tag(bp->b_mount, XFS_PTAG_VERIFIER_ERROR,
573 "Metadata corruption detected at %pS, quota %u",
574 __this_address, dqp->q_id);
575 xfs_alert(bp->b_mount, "Unmount and run xfs_repair");
576 return -EFSCORRUPTED;
577 }
578
579 /* copy everything from disk dquot to the incore dquot */
580 dqp->q_type = ddqp->d_type;
581 dqp->q_blk.hardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
582 dqp->q_blk.softlimit = be64_to_cpu(ddqp->d_blk_softlimit);
583 dqp->q_ino.hardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
584 dqp->q_ino.softlimit = be64_to_cpu(ddqp->d_ino_softlimit);
585 dqp->q_rtb.hardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
586 dqp->q_rtb.softlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
587
588 dqp->q_blk.count = be64_to_cpu(ddqp->d_bcount);
589 dqp->q_ino.count = be64_to_cpu(ddqp->d_icount);
590 dqp->q_rtb.count = be64_to_cpu(ddqp->d_rtbcount);
591
592 dqp->q_blk.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_btimer);
593 dqp->q_ino.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_itimer);
594 dqp->q_rtb.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_rtbtimer);
595
596 /*
597 * Reservation counters are defined as reservation plus current usage
598 * to avoid having to add every time.
599 */
600 dqp->q_blk.reserved = dqp->q_blk.count;
601 dqp->q_ino.reserved = dqp->q_ino.count;
602 dqp->q_rtb.reserved = dqp->q_rtb.count;
603
604 /* initialize the dquot speculative prealloc thresholds */
605 xfs_dquot_set_prealloc_limits(dqp);
606 return 0;
607}
608
609/* Copy the in-core quota fields into the on-disk buffer. */
610void
611xfs_dquot_to_disk(
612 struct xfs_disk_dquot *ddqp,
613 struct xfs_dquot *dqp)
614{
615 ddqp->d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
616 ddqp->d_version = XFS_DQUOT_VERSION;
617 ddqp->d_type = dqp->q_type;
618 ddqp->d_id = cpu_to_be32(dqp->q_id);
619 ddqp->d_pad0 = 0;
620 ddqp->d_pad = 0;
621
622 ddqp->d_blk_hardlimit = cpu_to_be64(dqp->q_blk.hardlimit);
623 ddqp->d_blk_softlimit = cpu_to_be64(dqp->q_blk.softlimit);
624 ddqp->d_ino_hardlimit = cpu_to_be64(dqp->q_ino.hardlimit);
625 ddqp->d_ino_softlimit = cpu_to_be64(dqp->q_ino.softlimit);
626 ddqp->d_rtb_hardlimit = cpu_to_be64(dqp->q_rtb.hardlimit);
627 ddqp->d_rtb_softlimit = cpu_to_be64(dqp->q_rtb.softlimit);
628
629 ddqp->d_bcount = cpu_to_be64(dqp->q_blk.count);
630 ddqp->d_icount = cpu_to_be64(dqp->q_ino.count);
631 ddqp->d_rtbcount = cpu_to_be64(dqp->q_rtb.count);
632
633 ddqp->d_bwarns = 0;
634 ddqp->d_iwarns = 0;
635 ddqp->d_rtbwarns = 0;
636
637 ddqp->d_btimer = xfs_dquot_to_disk_ts(dqp, dqp->q_blk.timer);
638 ddqp->d_itimer = xfs_dquot_to_disk_ts(dqp, dqp->q_ino.timer);
639 ddqp->d_rtbtimer = xfs_dquot_to_disk_ts(dqp, dqp->q_rtb.timer);
640}
641
642/*
643 * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
644 * and release the buffer immediately. If @can_alloc is true, fill any
645 * holes in the on-disk metadata.
646 */
647static int
648xfs_qm_dqread(
649 struct xfs_mount *mp,
650 xfs_dqid_t id,
651 xfs_dqtype_t type,
652 bool can_alloc,
653 struct xfs_dquot **dqpp)
654{
655 struct xfs_dquot *dqp;
656 struct xfs_buf *bp;
657 int error;
658
659 dqp = xfs_dquot_alloc(mp, id, type);
660 trace_xfs_dqread(dqp);
661
662 /* Try to read the buffer, allocating if necessary. */
663 error = xfs_dquot_disk_read(mp, dqp, &bp);
664 if (error == -ENOENT && can_alloc)
665 error = xfs_dquot_disk_alloc(dqp, &bp);
666 if (error)
667 goto err;
668
669 /*
670 * At this point we should have a clean locked buffer. Copy the data
671 * to the incore dquot and release the buffer since the incore dquot
672 * has its own locking protocol so we needn't tie up the buffer any
673 * further.
674 */
675 ASSERT(xfs_buf_islocked(bp));
676 error = xfs_dquot_from_disk(dqp, bp);
677 xfs_buf_relse(bp);
678 if (error)
679 goto err;
680
681 *dqpp = dqp;
682 return error;
683
684err:
685 trace_xfs_dqread_fail(dqp);
686 xfs_qm_dqdestroy(dqp);
687 *dqpp = NULL;
688 return error;
689}
690
691/*
692 * Advance to the next id in the current chunk, or if at the
693 * end of the chunk, skip ahead to first id in next allocated chunk
694 * using the SEEK_DATA interface.
695 */
696static int
697xfs_dq_get_next_id(
698 struct xfs_mount *mp,
699 xfs_dqtype_t type,
700 xfs_dqid_t *id)
701{
702 struct xfs_inode *quotip = xfs_quota_inode(mp, type);
703 xfs_dqid_t next_id = *id + 1; /* simple advance */
704 uint lock_flags;
705 struct xfs_bmbt_irec got;
706 struct xfs_iext_cursor cur;
707 xfs_fsblock_t start;
708 int error = 0;
709
710 /* If we'd wrap past the max ID, stop */
711 if (next_id < *id)
712 return -ENOENT;
713
714 /* If new ID is within the current chunk, advancing it sufficed */
715 if (next_id % mp->m_quotainfo->qi_dqperchunk) {
716 *id = next_id;
717 return 0;
718 }
719
720 /* Nope, next_id is now past the current chunk, so find the next one */
721 start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
722
723 lock_flags = xfs_ilock_data_map_shared(quotip);
724 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
725 if (error)
726 return error;
727
728 if (xfs_iext_lookup_extent(quotip, "ip->i_df, start, &cur, &got)) {
729 /* contiguous chunk, bump startoff for the id calculation */
730 if (got.br_startoff < start)
731 got.br_startoff = start;
732 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
733 } else {
734 error = -ENOENT;
735 }
736
737 xfs_iunlock(quotip, lock_flags);
738
739 return error;
740}
741
742/*
743 * Look up the dquot in the in-core cache. If found, the dquot is returned
744 * locked and ready to go.
745 */
746static struct xfs_dquot *
747xfs_qm_dqget_cache_lookup(
748 struct xfs_mount *mp,
749 struct xfs_quotainfo *qi,
750 struct radix_tree_root *tree,
751 xfs_dqid_t id)
752{
753 struct xfs_dquot *dqp;
754
755restart:
756 mutex_lock(&qi->qi_tree_lock);
757 dqp = radix_tree_lookup(tree, id);
758 if (!dqp) {
759 mutex_unlock(&qi->qi_tree_lock);
760 XFS_STATS_INC(mp, xs_qm_dqcachemisses);
761 return NULL;
762 }
763
764 xfs_dqlock(dqp);
765 if (dqp->q_flags & XFS_DQFLAG_FREEING) {
766 xfs_dqunlock(dqp);
767 mutex_unlock(&qi->qi_tree_lock);
768 trace_xfs_dqget_freeing(dqp);
769 delay(1);
770 goto restart;
771 }
772
773 dqp->q_nrefs++;
774 mutex_unlock(&qi->qi_tree_lock);
775
776 trace_xfs_dqget_hit(dqp);
777 XFS_STATS_INC(mp, xs_qm_dqcachehits);
778 return dqp;
779}
780
781/*
782 * Try to insert a new dquot into the in-core cache. If an error occurs the
783 * caller should throw away the dquot and start over. Otherwise, the dquot
784 * is returned locked (and held by the cache) as if there had been a cache
785 * hit.
786 */
787static int
788xfs_qm_dqget_cache_insert(
789 struct xfs_mount *mp,
790 struct xfs_quotainfo *qi,
791 struct radix_tree_root *tree,
792 xfs_dqid_t id,
793 struct xfs_dquot *dqp)
794{
795 int error;
796
797 mutex_lock(&qi->qi_tree_lock);
798 error = radix_tree_insert(tree, id, dqp);
799 if (unlikely(error)) {
800 /* Duplicate found! Caller must try again. */
801 WARN_ON(error != -EEXIST);
802 mutex_unlock(&qi->qi_tree_lock);
803 trace_xfs_dqget_dup(dqp);
804 return error;
805 }
806
807 /* Return a locked dquot to the caller, with a reference taken. */
808 xfs_dqlock(dqp);
809 dqp->q_nrefs = 1;
810
811 qi->qi_dquots++;
812 mutex_unlock(&qi->qi_tree_lock);
813
814 return 0;
815}
816
817/* Check our input parameters. */
818static int
819xfs_qm_dqget_checks(
820 struct xfs_mount *mp,
821 xfs_dqtype_t type)
822{
823 switch (type) {
824 case XFS_DQTYPE_USER:
825 if (!XFS_IS_UQUOTA_ON(mp))
826 return -ESRCH;
827 return 0;
828 case XFS_DQTYPE_GROUP:
829 if (!XFS_IS_GQUOTA_ON(mp))
830 return -ESRCH;
831 return 0;
832 case XFS_DQTYPE_PROJ:
833 if (!XFS_IS_PQUOTA_ON(mp))
834 return -ESRCH;
835 return 0;
836 default:
837 WARN_ON_ONCE(0);
838 return -EINVAL;
839 }
840}
841
842/*
843 * Given the file system, id, and type (UDQUOT/GDQUOT/PDQUOT), return a
844 * locked dquot, doing an allocation (if requested) as needed.
845 */
846int
847xfs_qm_dqget(
848 struct xfs_mount *mp,
849 xfs_dqid_t id,
850 xfs_dqtype_t type,
851 bool can_alloc,
852 struct xfs_dquot **O_dqpp)
853{
854 struct xfs_quotainfo *qi = mp->m_quotainfo;
855 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
856 struct xfs_dquot *dqp;
857 int error;
858
859 error = xfs_qm_dqget_checks(mp, type);
860 if (error)
861 return error;
862
863restart:
864 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
865 if (dqp) {
866 *O_dqpp = dqp;
867 return 0;
868 }
869
870 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
871 if (error)
872 return error;
873
874 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
875 if (error) {
876 /*
877 * Duplicate found. Just throw away the new dquot and start
878 * over.
879 */
880 xfs_qm_dqdestroy(dqp);
881 XFS_STATS_INC(mp, xs_qm_dquot_dups);
882 goto restart;
883 }
884
885 trace_xfs_dqget_miss(dqp);
886 *O_dqpp = dqp;
887 return 0;
888}
889
890/*
891 * Given a dquot id and type, read and initialize a dquot from the on-disk
892 * metadata. This function is only for use during quota initialization so
893 * it ignores the dquot cache assuming that the dquot shrinker isn't set up.
894 * The caller is responsible for _qm_dqdestroy'ing the returned dquot.
895 */
896int
897xfs_qm_dqget_uncached(
898 struct xfs_mount *mp,
899 xfs_dqid_t id,
900 xfs_dqtype_t type,
901 struct xfs_dquot **dqpp)
902{
903 int error;
904
905 error = xfs_qm_dqget_checks(mp, type);
906 if (error)
907 return error;
908
909 return xfs_qm_dqread(mp, id, type, 0, dqpp);
910}
911
912/* Return the quota id for a given inode and type. */
913xfs_dqid_t
914xfs_qm_id_for_quotatype(
915 struct xfs_inode *ip,
916 xfs_dqtype_t type)
917{
918 switch (type) {
919 case XFS_DQTYPE_USER:
920 return i_uid_read(VFS_I(ip));
921 case XFS_DQTYPE_GROUP:
922 return i_gid_read(VFS_I(ip));
923 case XFS_DQTYPE_PROJ:
924 return ip->i_projid;
925 }
926 ASSERT(0);
927 return 0;
928}
929
930/*
931 * Return the dquot for a given inode and type. If @can_alloc is true, then
932 * allocate blocks if needed. The inode's ILOCK must be held and it must not
933 * have already had an inode attached.
934 */
935int
936xfs_qm_dqget_inode(
937 struct xfs_inode *ip,
938 xfs_dqtype_t type,
939 bool can_alloc,
940 struct xfs_dquot **O_dqpp)
941{
942 struct xfs_mount *mp = ip->i_mount;
943 struct xfs_quotainfo *qi = mp->m_quotainfo;
944 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
945 struct xfs_dquot *dqp;
946 xfs_dqid_t id;
947 int error;
948
949 error = xfs_qm_dqget_checks(mp, type);
950 if (error)
951 return error;
952
953 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
954 ASSERT(xfs_inode_dquot(ip, type) == NULL);
955
956 id = xfs_qm_id_for_quotatype(ip, type);
957
958restart:
959 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
960 if (dqp) {
961 *O_dqpp = dqp;
962 return 0;
963 }
964
965 /*
966 * Dquot cache miss. We don't want to keep the inode lock across
967 * a (potential) disk read. Also we don't want to deal with the lock
968 * ordering between quotainode and this inode. OTOH, dropping the inode
969 * lock here means dealing with a chown that can happen before
970 * we re-acquire the lock.
971 */
972 xfs_iunlock(ip, XFS_ILOCK_EXCL);
973 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
974 xfs_ilock(ip, XFS_ILOCK_EXCL);
975 if (error)
976 return error;
977
978 /*
979 * A dquot could be attached to this inode by now, since we had
980 * dropped the ilock.
981 */
982 if (xfs_this_quota_on(mp, type)) {
983 struct xfs_dquot *dqp1;
984
985 dqp1 = xfs_inode_dquot(ip, type);
986 if (dqp1) {
987 xfs_qm_dqdestroy(dqp);
988 dqp = dqp1;
989 xfs_dqlock(dqp);
990 goto dqret;
991 }
992 } else {
993 /* inode stays locked on return */
994 xfs_qm_dqdestroy(dqp);
995 return -ESRCH;
996 }
997
998 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
999 if (error) {
1000 /*
1001 * Duplicate found. Just throw away the new dquot and start
1002 * over.
1003 */
1004 xfs_qm_dqdestroy(dqp);
1005 XFS_STATS_INC(mp, xs_qm_dquot_dups);
1006 goto restart;
1007 }
1008
1009dqret:
1010 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1011 trace_xfs_dqget_miss(dqp);
1012 *O_dqpp = dqp;
1013 return 0;
1014}
1015
1016/*
1017 * Starting at @id and progressing upwards, look for an initialized incore
1018 * dquot, lock it, and return it.
1019 */
1020int
1021xfs_qm_dqget_next(
1022 struct xfs_mount *mp,
1023 xfs_dqid_t id,
1024 xfs_dqtype_t type,
1025 struct xfs_dquot **dqpp)
1026{
1027 struct xfs_dquot *dqp;
1028 int error = 0;
1029
1030 *dqpp = NULL;
1031 for (; !error; error = xfs_dq_get_next_id(mp, type, &id)) {
1032 error = xfs_qm_dqget(mp, id, type, false, &dqp);
1033 if (error == -ENOENT)
1034 continue;
1035 else if (error != 0)
1036 break;
1037
1038 if (!XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
1039 *dqpp = dqp;
1040 return 0;
1041 }
1042
1043 xfs_qm_dqput(dqp);
1044 }
1045
1046 return error;
1047}
1048
1049/*
1050 * Release a reference to the dquot (decrement ref-count) and unlock it.
1051 *
1052 * If there is a group quota attached to this dquot, carefully release that
1053 * too without tripping over deadlocks'n'stuff.
1054 */
1055void
1056xfs_qm_dqput(
1057 struct xfs_dquot *dqp)
1058{
1059 ASSERT(dqp->q_nrefs > 0);
1060 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1061
1062 trace_xfs_dqput(dqp);
1063
1064 if (--dqp->q_nrefs == 0) {
1065 struct xfs_quotainfo *qi = dqp->q_mount->m_quotainfo;
1066 trace_xfs_dqput_free(dqp);
1067
1068 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
1069 XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
1070 }
1071 xfs_dqunlock(dqp);
1072}
1073
1074/*
1075 * Release a dquot. Flush it if dirty, then dqput() it.
1076 * dquot must not be locked.
1077 */
1078void
1079xfs_qm_dqrele(
1080 struct xfs_dquot *dqp)
1081{
1082 if (!dqp)
1083 return;
1084
1085 trace_xfs_dqrele(dqp);
1086
1087 xfs_dqlock(dqp);
1088 /*
1089 * We don't care to flush it if the dquot is dirty here.
1090 * That will create stutters that we want to avoid.
1091 * Instead we do a delayed write when we try to reclaim
1092 * a dirty dquot. Also xfs_sync will take part of the burden...
1093 */
1094 xfs_qm_dqput(dqp);
1095}
1096
1097/*
1098 * This is the dquot flushing I/O completion routine. It is called
1099 * from interrupt level when the buffer containing the dquot is
1100 * flushed to disk. It is responsible for removing the dquot logitem
1101 * from the AIL if it has not been re-logged, and unlocking the dquot's
1102 * flush lock. This behavior is very similar to that of inodes..
1103 */
1104static void
1105xfs_qm_dqflush_done(
1106 struct xfs_log_item *lip)
1107{
1108 struct xfs_dq_logitem *qip = (struct xfs_dq_logitem *)lip;
1109 struct xfs_dquot *dqp = qip->qli_dquot;
1110 struct xfs_ail *ailp = lip->li_ailp;
1111 xfs_lsn_t tail_lsn;
1112
1113 /*
1114 * We only want to pull the item from the AIL if its
1115 * location in the log has not changed since we started the flush.
1116 * Thus, we only bother if the dquot's lsn has
1117 * not changed. First we check the lsn outside the lock
1118 * since it's cheaper, and then we recheck while
1119 * holding the lock before removing the dquot from the AIL.
1120 */
1121 if (test_bit(XFS_LI_IN_AIL, &lip->li_flags) &&
1122 ((lip->li_lsn == qip->qli_flush_lsn) ||
1123 test_bit(XFS_LI_FAILED, &lip->li_flags))) {
1124
1125 spin_lock(&ailp->ail_lock);
1126 xfs_clear_li_failed(lip);
1127 if (lip->li_lsn == qip->qli_flush_lsn) {
1128 /* xfs_ail_update_finish() drops the AIL lock */
1129 tail_lsn = xfs_ail_delete_one(ailp, lip);
1130 xfs_ail_update_finish(ailp, tail_lsn);
1131 } else {
1132 spin_unlock(&ailp->ail_lock);
1133 }
1134 }
1135
1136 /*
1137 * Release the dq's flush lock since we're done with it.
1138 */
1139 xfs_dqfunlock(dqp);
1140}
1141
1142void
1143xfs_buf_dquot_iodone(
1144 struct xfs_buf *bp)
1145{
1146 struct xfs_log_item *lip, *n;
1147
1148 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
1149 list_del_init(&lip->li_bio_list);
1150 xfs_qm_dqflush_done(lip);
1151 }
1152}
1153
1154void
1155xfs_buf_dquot_io_fail(
1156 struct xfs_buf *bp)
1157{
1158 struct xfs_log_item *lip;
1159
1160 spin_lock(&bp->b_mount->m_ail->ail_lock);
1161 list_for_each_entry(lip, &bp->b_li_list, li_bio_list)
1162 xfs_set_li_failed(lip, bp);
1163 spin_unlock(&bp->b_mount->m_ail->ail_lock);
1164}
1165
1166/* Check incore dquot for errors before we flush. */
1167static xfs_failaddr_t
1168xfs_qm_dqflush_check(
1169 struct xfs_dquot *dqp)
1170{
1171 xfs_dqtype_t type = xfs_dquot_type(dqp);
1172
1173 if (type != XFS_DQTYPE_USER &&
1174 type != XFS_DQTYPE_GROUP &&
1175 type != XFS_DQTYPE_PROJ)
1176 return __this_address;
1177
1178 if (dqp->q_id == 0)
1179 return NULL;
1180
1181 if (dqp->q_blk.softlimit && dqp->q_blk.count > dqp->q_blk.softlimit &&
1182 !dqp->q_blk.timer)
1183 return __this_address;
1184
1185 if (dqp->q_ino.softlimit && dqp->q_ino.count > dqp->q_ino.softlimit &&
1186 !dqp->q_ino.timer)
1187 return __this_address;
1188
1189 if (dqp->q_rtb.softlimit && dqp->q_rtb.count > dqp->q_rtb.softlimit &&
1190 !dqp->q_rtb.timer)
1191 return __this_address;
1192
1193 /* bigtime flag should never be set on root dquots */
1194 if (dqp->q_type & XFS_DQTYPE_BIGTIME) {
1195 if (!xfs_has_bigtime(dqp->q_mount))
1196 return __this_address;
1197 if (dqp->q_id == 0)
1198 return __this_address;
1199 }
1200
1201 return NULL;
1202}
1203
1204/*
1205 * Write a modified dquot to disk.
1206 * The dquot must be locked and the flush lock too taken by caller.
1207 * The flush lock will not be unlocked until the dquot reaches the disk,
1208 * but the dquot is free to be unlocked and modified by the caller
1209 * in the interim. Dquot is still locked on return. This behavior is
1210 * identical to that of inodes.
1211 */
1212int
1213xfs_qm_dqflush(
1214 struct xfs_dquot *dqp,
1215 struct xfs_buf **bpp)
1216{
1217 struct xfs_mount *mp = dqp->q_mount;
1218 struct xfs_log_item *lip = &dqp->q_logitem.qli_item;
1219 struct xfs_buf *bp;
1220 struct xfs_dqblk *dqblk;
1221 xfs_failaddr_t fa;
1222 int error;
1223
1224 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1225 ASSERT(!completion_done(&dqp->q_flush));
1226
1227 trace_xfs_dqflush(dqp);
1228
1229 *bpp = NULL;
1230
1231 xfs_qm_dqunpin_wait(dqp);
1232
1233 /*
1234 * Get the buffer containing the on-disk dquot
1235 */
1236 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1237 mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK,
1238 &bp, &xfs_dquot_buf_ops);
1239 if (error == -EAGAIN)
1240 goto out_unlock;
1241 if (error)
1242 goto out_abort;
1243
1244 fa = xfs_qm_dqflush_check(dqp);
1245 if (fa) {
1246 xfs_alert(mp, "corrupt dquot ID 0x%x in memory at %pS",
1247 dqp->q_id, fa);
1248 xfs_buf_relse(bp);
1249 error = -EFSCORRUPTED;
1250 goto out_abort;
1251 }
1252
1253 /* Flush the incore dquot to the ondisk buffer. */
1254 dqblk = bp->b_addr + dqp->q_bufoffset;
1255 xfs_dquot_to_disk(&dqblk->dd_diskdq, dqp);
1256
1257 /*
1258 * Clear the dirty field and remember the flush lsn for later use.
1259 */
1260 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1261
1262 xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1263 &dqp->q_logitem.qli_item.li_lsn);
1264
1265 /*
1266 * copy the lsn into the on-disk dquot now while we have the in memory
1267 * dquot here. This can't be done later in the write verifier as we
1268 * can't get access to the log item at that point in time.
1269 *
1270 * We also calculate the CRC here so that the on-disk dquot in the
1271 * buffer always has a valid CRC. This ensures there is no possibility
1272 * of a dquot without an up-to-date CRC getting to disk.
1273 */
1274 if (xfs_has_crc(mp)) {
1275 dqblk->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1276 xfs_update_cksum((char *)dqblk, sizeof(struct xfs_dqblk),
1277 XFS_DQUOT_CRC_OFF);
1278 }
1279
1280 /*
1281 * Attach the dquot to the buffer so that we can remove this dquot from
1282 * the AIL and release the flush lock once the dquot is synced to disk.
1283 */
1284 bp->b_flags |= _XBF_DQUOTS;
1285 list_add_tail(&dqp->q_logitem.qli_item.li_bio_list, &bp->b_li_list);
1286
1287 /*
1288 * If the buffer is pinned then push on the log so we won't
1289 * get stuck waiting in the write for too long.
1290 */
1291 if (xfs_buf_ispinned(bp)) {
1292 trace_xfs_dqflush_force(dqp);
1293 xfs_log_force(mp, 0);
1294 }
1295
1296 trace_xfs_dqflush_done(dqp);
1297 *bpp = bp;
1298 return 0;
1299
1300out_abort:
1301 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1302 xfs_trans_ail_delete(lip, 0);
1303 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1304out_unlock:
1305 xfs_dqfunlock(dqp);
1306 return error;
1307}
1308
1309/*
1310 * Lock two xfs_dquot structures.
1311 *
1312 * To avoid deadlocks we always lock the quota structure with
1313 * the lowerd id first.
1314 */
1315void
1316xfs_dqlock2(
1317 struct xfs_dquot *d1,
1318 struct xfs_dquot *d2)
1319{
1320 if (d1 && d2) {
1321 ASSERT(d1 != d2);
1322 if (d1->q_id > d2->q_id) {
1323 mutex_lock(&d2->q_qlock);
1324 mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1325 } else {
1326 mutex_lock(&d1->q_qlock);
1327 mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1328 }
1329 } else if (d1) {
1330 mutex_lock(&d1->q_qlock);
1331 } else if (d2) {
1332 mutex_lock(&d2->q_qlock);
1333 }
1334}
1335
1336int __init
1337xfs_qm_init(void)
1338{
1339 xfs_dquot_cache = kmem_cache_create("xfs_dquot",
1340 sizeof(struct xfs_dquot),
1341 0, 0, NULL);
1342 if (!xfs_dquot_cache)
1343 goto out;
1344
1345 xfs_dqtrx_cache = kmem_cache_create("xfs_dqtrx",
1346 sizeof(struct xfs_dquot_acct),
1347 0, 0, NULL);
1348 if (!xfs_dqtrx_cache)
1349 goto out_free_dquot_cache;
1350
1351 return 0;
1352
1353out_free_dquot_cache:
1354 kmem_cache_destroy(xfs_dquot_cache);
1355out:
1356 return -ENOMEM;
1357}
1358
1359void
1360xfs_qm_exit(void)
1361{
1362 kmem_cache_destroy(xfs_dqtrx_cache);
1363 kmem_cache_destroy(xfs_dquot_cache);
1364}
1365
1366/*
1367 * Iterate every dquot of a particular type. The caller must ensure that the
1368 * particular quota type is active. iter_fn can return negative error codes,
1369 * or -ECANCELED to indicate that it wants to stop iterating.
1370 */
1371int
1372xfs_qm_dqiterate(
1373 struct xfs_mount *mp,
1374 xfs_dqtype_t type,
1375 xfs_qm_dqiterate_fn iter_fn,
1376 void *priv)
1377{
1378 struct xfs_dquot *dq;
1379 xfs_dqid_t id = 0;
1380 int error;
1381
1382 do {
1383 error = xfs_qm_dqget_next(mp, id, type, &dq);
1384 if (error == -ENOENT)
1385 return 0;
1386 if (error)
1387 return error;
1388
1389 error = iter_fn(dq, type, priv);
1390 id = dq->q_id;
1391 xfs_qm_dqput(dq);
1392 } while (error == 0 && id != 0);
1393
1394 return error;
1395}
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_format.h"
21#include "xfs_log_format.h"
22#include "xfs_shared.h"
23#include "xfs_trans_resv.h"
24#include "xfs_bit.h"
25#include "xfs_mount.h"
26#include "xfs_inode.h"
27#include "xfs_bmap.h"
28#include "xfs_bmap_util.h"
29#include "xfs_alloc.h"
30#include "xfs_quota.h"
31#include "xfs_error.h"
32#include "xfs_trans.h"
33#include "xfs_buf_item.h"
34#include "xfs_trans_space.h"
35#include "xfs_trans_priv.h"
36#include "xfs_qm.h"
37#include "xfs_cksum.h"
38#include "xfs_trace.h"
39#include "xfs_log.h"
40#include "xfs_bmap_btree.h"
41
42/*
43 * Lock order:
44 *
45 * ip->i_lock
46 * qi->qi_tree_lock
47 * dquot->q_qlock (xfs_dqlock() and friends)
48 * dquot->q_flush (xfs_dqflock() and friends)
49 * qi->qi_lru_lock
50 *
51 * If two dquots need to be locked the order is user before group/project,
52 * otherwise by the lowest id first, see xfs_dqlock2.
53 */
54
55#ifdef DEBUG
56xfs_buftarg_t *xfs_dqerror_target;
57int xfs_do_dqerror;
58int xfs_dqreq_num;
59int xfs_dqerror_mod = 33;
60#endif
61
62struct kmem_zone *xfs_qm_dqtrxzone;
63static struct kmem_zone *xfs_qm_dqzone;
64
65static struct lock_class_key xfs_dquot_group_class;
66static struct lock_class_key xfs_dquot_project_class;
67
68/*
69 * This is called to free all the memory associated with a dquot
70 */
71void
72xfs_qm_dqdestroy(
73 xfs_dquot_t *dqp)
74{
75 ASSERT(list_empty(&dqp->q_lru));
76
77 mutex_destroy(&dqp->q_qlock);
78
79 XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
80 kmem_zone_free(xfs_qm_dqzone, dqp);
81}
82
83/*
84 * If default limits are in force, push them into the dquot now.
85 * We overwrite the dquot limits only if they are zero and this
86 * is not the root dquot.
87 */
88void
89xfs_qm_adjust_dqlimits(
90 struct xfs_mount *mp,
91 struct xfs_dquot *dq)
92{
93 struct xfs_quotainfo *q = mp->m_quotainfo;
94 struct xfs_disk_dquot *d = &dq->q_core;
95 struct xfs_def_quota *defq;
96 int prealloc = 0;
97
98 ASSERT(d->d_id);
99 defq = xfs_get_defquota(dq, q);
100
101 if (defq->bsoftlimit && !d->d_blk_softlimit) {
102 d->d_blk_softlimit = cpu_to_be64(defq->bsoftlimit);
103 prealloc = 1;
104 }
105 if (defq->bhardlimit && !d->d_blk_hardlimit) {
106 d->d_blk_hardlimit = cpu_to_be64(defq->bhardlimit);
107 prealloc = 1;
108 }
109 if (defq->isoftlimit && !d->d_ino_softlimit)
110 d->d_ino_softlimit = cpu_to_be64(defq->isoftlimit);
111 if (defq->ihardlimit && !d->d_ino_hardlimit)
112 d->d_ino_hardlimit = cpu_to_be64(defq->ihardlimit);
113 if (defq->rtbsoftlimit && !d->d_rtb_softlimit)
114 d->d_rtb_softlimit = cpu_to_be64(defq->rtbsoftlimit);
115 if (defq->rtbhardlimit && !d->d_rtb_hardlimit)
116 d->d_rtb_hardlimit = cpu_to_be64(defq->rtbhardlimit);
117
118 if (prealloc)
119 xfs_dquot_set_prealloc_limits(dq);
120}
121
122/*
123 * Check the limits and timers of a dquot and start or reset timers
124 * if necessary.
125 * This gets called even when quota enforcement is OFF, which makes our
126 * life a little less complicated. (We just don't reject any quota
127 * reservations in that case, when enforcement is off).
128 * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
129 * enforcement's off.
130 * In contrast, warnings are a little different in that they don't
131 * 'automatically' get started when limits get exceeded. They do
132 * get reset to zero, however, when we find the count to be under
133 * the soft limit (they are only ever set non-zero via userspace).
134 */
135void
136xfs_qm_adjust_dqtimers(
137 xfs_mount_t *mp,
138 xfs_disk_dquot_t *d)
139{
140 ASSERT(d->d_id);
141
142#ifdef DEBUG
143 if (d->d_blk_hardlimit)
144 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
145 be64_to_cpu(d->d_blk_hardlimit));
146 if (d->d_ino_hardlimit)
147 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
148 be64_to_cpu(d->d_ino_hardlimit));
149 if (d->d_rtb_hardlimit)
150 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
151 be64_to_cpu(d->d_rtb_hardlimit));
152#endif
153
154 if (!d->d_btimer) {
155 if ((d->d_blk_softlimit &&
156 (be64_to_cpu(d->d_bcount) >
157 be64_to_cpu(d->d_blk_softlimit))) ||
158 (d->d_blk_hardlimit &&
159 (be64_to_cpu(d->d_bcount) >
160 be64_to_cpu(d->d_blk_hardlimit)))) {
161 d->d_btimer = cpu_to_be32(get_seconds() +
162 mp->m_quotainfo->qi_btimelimit);
163 } else {
164 d->d_bwarns = 0;
165 }
166 } else {
167 if ((!d->d_blk_softlimit ||
168 (be64_to_cpu(d->d_bcount) <=
169 be64_to_cpu(d->d_blk_softlimit))) &&
170 (!d->d_blk_hardlimit ||
171 (be64_to_cpu(d->d_bcount) <=
172 be64_to_cpu(d->d_blk_hardlimit)))) {
173 d->d_btimer = 0;
174 }
175 }
176
177 if (!d->d_itimer) {
178 if ((d->d_ino_softlimit &&
179 (be64_to_cpu(d->d_icount) >
180 be64_to_cpu(d->d_ino_softlimit))) ||
181 (d->d_ino_hardlimit &&
182 (be64_to_cpu(d->d_icount) >
183 be64_to_cpu(d->d_ino_hardlimit)))) {
184 d->d_itimer = cpu_to_be32(get_seconds() +
185 mp->m_quotainfo->qi_itimelimit);
186 } else {
187 d->d_iwarns = 0;
188 }
189 } else {
190 if ((!d->d_ino_softlimit ||
191 (be64_to_cpu(d->d_icount) <=
192 be64_to_cpu(d->d_ino_softlimit))) &&
193 (!d->d_ino_hardlimit ||
194 (be64_to_cpu(d->d_icount) <=
195 be64_to_cpu(d->d_ino_hardlimit)))) {
196 d->d_itimer = 0;
197 }
198 }
199
200 if (!d->d_rtbtimer) {
201 if ((d->d_rtb_softlimit &&
202 (be64_to_cpu(d->d_rtbcount) >
203 be64_to_cpu(d->d_rtb_softlimit))) ||
204 (d->d_rtb_hardlimit &&
205 (be64_to_cpu(d->d_rtbcount) >
206 be64_to_cpu(d->d_rtb_hardlimit)))) {
207 d->d_rtbtimer = cpu_to_be32(get_seconds() +
208 mp->m_quotainfo->qi_rtbtimelimit);
209 } else {
210 d->d_rtbwarns = 0;
211 }
212 } else {
213 if ((!d->d_rtb_softlimit ||
214 (be64_to_cpu(d->d_rtbcount) <=
215 be64_to_cpu(d->d_rtb_softlimit))) &&
216 (!d->d_rtb_hardlimit ||
217 (be64_to_cpu(d->d_rtbcount) <=
218 be64_to_cpu(d->d_rtb_hardlimit)))) {
219 d->d_rtbtimer = 0;
220 }
221 }
222}
223
224/*
225 * initialize a buffer full of dquots and log the whole thing
226 */
227STATIC void
228xfs_qm_init_dquot_blk(
229 xfs_trans_t *tp,
230 xfs_mount_t *mp,
231 xfs_dqid_t id,
232 uint type,
233 xfs_buf_t *bp)
234{
235 struct xfs_quotainfo *q = mp->m_quotainfo;
236 xfs_dqblk_t *d;
237 xfs_dqid_t curid;
238 int i;
239
240 ASSERT(tp);
241 ASSERT(xfs_buf_islocked(bp));
242
243 d = bp->b_addr;
244
245 /*
246 * ID of the first dquot in the block - id's are zero based.
247 */
248 curid = id - (id % q->qi_dqperchunk);
249 memset(d, 0, BBTOB(q->qi_dqchunklen));
250 for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
251 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
252 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
253 d->dd_diskdq.d_id = cpu_to_be32(curid);
254 d->dd_diskdq.d_flags = type;
255 if (xfs_sb_version_hascrc(&mp->m_sb)) {
256 uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
257 xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
258 XFS_DQUOT_CRC_OFF);
259 }
260 }
261
262 xfs_trans_dquot_buf(tp, bp,
263 (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
264 ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
265 XFS_BLF_GDQUOT_BUF)));
266 xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
267}
268
269/*
270 * Initialize the dynamic speculative preallocation thresholds. The lo/hi
271 * watermarks correspond to the soft and hard limits by default. If a soft limit
272 * is not specified, we use 95% of the hard limit.
273 */
274void
275xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
276{
277 __uint64_t space;
278
279 dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
280 dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
281 if (!dqp->q_prealloc_lo_wmark) {
282 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
283 do_div(dqp->q_prealloc_lo_wmark, 100);
284 dqp->q_prealloc_lo_wmark *= 95;
285 }
286
287 space = dqp->q_prealloc_hi_wmark;
288
289 do_div(space, 100);
290 dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
291 dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
292 dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
293}
294
295/*
296 * Allocate a block and fill it with dquots.
297 * This is called when the bmapi finds a hole.
298 */
299STATIC int
300xfs_qm_dqalloc(
301 xfs_trans_t **tpp,
302 xfs_mount_t *mp,
303 xfs_dquot_t *dqp,
304 xfs_inode_t *quotip,
305 xfs_fileoff_t offset_fsb,
306 xfs_buf_t **O_bpp)
307{
308 xfs_fsblock_t firstblock;
309 xfs_bmap_free_t flist;
310 xfs_bmbt_irec_t map;
311 int nmaps, error;
312 xfs_buf_t *bp;
313 xfs_trans_t *tp = *tpp;
314
315 ASSERT(tp != NULL);
316
317 trace_xfs_dqalloc(dqp);
318
319 /*
320 * Initialize the bmap freelist prior to calling bmapi code.
321 */
322 xfs_bmap_init(&flist, &firstblock);
323 xfs_ilock(quotip, XFS_ILOCK_EXCL);
324 /*
325 * Return if this type of quotas is turned off while we didn't
326 * have an inode lock
327 */
328 if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
329 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
330 return -ESRCH;
331 }
332
333 xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
334 nmaps = 1;
335 error = xfs_bmapi_write(tp, quotip, offset_fsb,
336 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
337 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
338 &map, &nmaps, &flist);
339 if (error)
340 goto error0;
341 ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
342 ASSERT(nmaps == 1);
343 ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
344 (map.br_startblock != HOLESTARTBLOCK));
345
346 /*
347 * Keep track of the blkno to save a lookup later
348 */
349 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
350
351 /* now we can just get the buffer (there's nothing to read yet) */
352 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
353 dqp->q_blkno,
354 mp->m_quotainfo->qi_dqchunklen,
355 0);
356 if (!bp) {
357 error = -ENOMEM;
358 goto error1;
359 }
360 bp->b_ops = &xfs_dquot_buf_ops;
361
362 /*
363 * Make a chunk of dquots out of this buffer and log
364 * the entire thing.
365 */
366 xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
367 dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
368
369 /*
370 * xfs_bmap_finish() may commit the current transaction and
371 * start a second transaction if the freelist is not empty.
372 *
373 * Since we still want to modify this buffer, we need to
374 * ensure that the buffer is not released on commit of
375 * the first transaction and ensure the buffer is added to the
376 * second transaction.
377 *
378 * If there is only one transaction then don't stop the buffer
379 * from being released when it commits later on.
380 */
381
382 xfs_trans_bhold(tp, bp);
383
384 error = xfs_bmap_finish(tpp, &flist, NULL);
385 if (error)
386 goto error1;
387
388 /* Transaction was committed? */
389 if (*tpp != tp) {
390 tp = *tpp;
391 xfs_trans_bjoin(tp, bp);
392 } else {
393 xfs_trans_bhold_release(tp, bp);
394 }
395
396 *O_bpp = bp;
397 return 0;
398
399error1:
400 xfs_bmap_cancel(&flist);
401error0:
402 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
403
404 return error;
405}
406
407STATIC int
408xfs_qm_dqrepair(
409 struct xfs_mount *mp,
410 struct xfs_trans *tp,
411 struct xfs_dquot *dqp,
412 xfs_dqid_t firstid,
413 struct xfs_buf **bpp)
414{
415 int error;
416 struct xfs_disk_dquot *ddq;
417 struct xfs_dqblk *d;
418 int i;
419
420 /*
421 * Read the buffer without verification so we get the corrupted
422 * buffer returned to us. make sure we verify it on write, though.
423 */
424 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
425 mp->m_quotainfo->qi_dqchunklen,
426 0, bpp, NULL);
427
428 if (error) {
429 ASSERT(*bpp == NULL);
430 return error;
431 }
432 (*bpp)->b_ops = &xfs_dquot_buf_ops;
433
434 ASSERT(xfs_buf_islocked(*bpp));
435 d = (struct xfs_dqblk *)(*bpp)->b_addr;
436
437 /* Do the actual repair of dquots in this buffer */
438 for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
439 ddq = &d[i].dd_diskdq;
440 error = xfs_dqcheck(mp, ddq, firstid + i,
441 dqp->dq_flags & XFS_DQ_ALLTYPES,
442 XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
443 if (error) {
444 /* repair failed, we're screwed */
445 xfs_trans_brelse(tp, *bpp);
446 return -EIO;
447 }
448 }
449
450 return 0;
451}
452
453/*
454 * Maps a dquot to the buffer containing its on-disk version.
455 * This returns a ptr to the buffer containing the on-disk dquot
456 * in the bpp param, and a ptr to the on-disk dquot within that buffer
457 */
458STATIC int
459xfs_qm_dqtobp(
460 xfs_trans_t **tpp,
461 xfs_dquot_t *dqp,
462 xfs_disk_dquot_t **O_ddpp,
463 xfs_buf_t **O_bpp,
464 uint flags)
465{
466 struct xfs_bmbt_irec map;
467 int nmaps = 1, error;
468 struct xfs_buf *bp;
469 struct xfs_inode *quotip;
470 struct xfs_mount *mp = dqp->q_mount;
471 xfs_dqid_t id = be32_to_cpu(dqp->q_core.d_id);
472 struct xfs_trans *tp = (tpp ? *tpp : NULL);
473 uint lock_mode;
474
475 quotip = xfs_quota_inode(dqp->q_mount, dqp->dq_flags);
476 dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
477
478 lock_mode = xfs_ilock_data_map_shared(quotip);
479 if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
480 /*
481 * Return if this type of quotas is turned off while we
482 * didn't have the quota inode lock.
483 */
484 xfs_iunlock(quotip, lock_mode);
485 return -ESRCH;
486 }
487
488 /*
489 * Find the block map; no allocations yet
490 */
491 error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
492 XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
493
494 xfs_iunlock(quotip, lock_mode);
495 if (error)
496 return error;
497
498 ASSERT(nmaps == 1);
499 ASSERT(map.br_blockcount == 1);
500
501 /*
502 * Offset of dquot in the (fixed sized) dquot chunk.
503 */
504 dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
505 sizeof(xfs_dqblk_t);
506
507 ASSERT(map.br_startblock != DELAYSTARTBLOCK);
508 if (map.br_startblock == HOLESTARTBLOCK) {
509 /*
510 * We don't allocate unless we're asked to
511 */
512 if (!(flags & XFS_QMOPT_DQALLOC))
513 return -ENOENT;
514
515 ASSERT(tp);
516 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
517 dqp->q_fileoffset, &bp);
518 if (error)
519 return error;
520 tp = *tpp;
521 } else {
522 trace_xfs_dqtobp_read(dqp);
523
524 /*
525 * store the blkno etc so that we don't have to do the
526 * mapping all the time
527 */
528 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
529
530 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
531 dqp->q_blkno,
532 mp->m_quotainfo->qi_dqchunklen,
533 0, &bp, &xfs_dquot_buf_ops);
534
535 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
536 xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
537 mp->m_quotainfo->qi_dqperchunk;
538 ASSERT(bp == NULL);
539 error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
540 }
541
542 if (error) {
543 ASSERT(bp == NULL);
544 return error;
545 }
546 }
547
548 ASSERT(xfs_buf_islocked(bp));
549 *O_bpp = bp;
550 *O_ddpp = bp->b_addr + dqp->q_bufoffset;
551
552 return 0;
553}
554
555
556/*
557 * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
558 * and release the buffer immediately.
559 *
560 * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
561 */
562int
563xfs_qm_dqread(
564 struct xfs_mount *mp,
565 xfs_dqid_t id,
566 uint type,
567 uint flags,
568 struct xfs_dquot **O_dqpp)
569{
570 struct xfs_dquot *dqp;
571 struct xfs_disk_dquot *ddqp;
572 struct xfs_buf *bp;
573 struct xfs_trans *tp = NULL;
574 int error;
575
576 dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
577
578 dqp->dq_flags = type;
579 dqp->q_core.d_id = cpu_to_be32(id);
580 dqp->q_mount = mp;
581 INIT_LIST_HEAD(&dqp->q_lru);
582 mutex_init(&dqp->q_qlock);
583 init_waitqueue_head(&dqp->q_pinwait);
584
585 /*
586 * Because we want to use a counting completion, complete
587 * the flush completion once to allow a single access to
588 * the flush completion without blocking.
589 */
590 init_completion(&dqp->q_flush);
591 complete(&dqp->q_flush);
592
593 /*
594 * Make sure group quotas have a different lock class than user
595 * quotas.
596 */
597 switch (type) {
598 case XFS_DQ_USER:
599 /* uses the default lock class */
600 break;
601 case XFS_DQ_GROUP:
602 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
603 break;
604 case XFS_DQ_PROJ:
605 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
606 break;
607 default:
608 ASSERT(0);
609 break;
610 }
611
612 XFS_STATS_INC(mp, xs_qm_dquot);
613
614 trace_xfs_dqread(dqp);
615
616 if (flags & XFS_QMOPT_DQALLOC) {
617 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
618 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_qm_dqalloc,
619 XFS_QM_DQALLOC_SPACE_RES(mp), 0);
620 if (error)
621 goto error1;
622 }
623
624 /*
625 * get a pointer to the on-disk dquot and the buffer containing it
626 * dqp already knows its own type (GROUP/USER).
627 */
628 error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
629 if (error) {
630 /*
631 * This can happen if quotas got turned off (ESRCH),
632 * or if the dquot didn't exist on disk and we ask to
633 * allocate (ENOENT).
634 */
635 trace_xfs_dqread_fail(dqp);
636 goto error1;
637 }
638
639 /* copy everything from disk dquot to the incore dquot */
640 memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
641 xfs_qm_dquot_logitem_init(dqp);
642
643 /*
644 * Reservation counters are defined as reservation plus current usage
645 * to avoid having to add every time.
646 */
647 dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
648 dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
649 dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
650
651 /* initialize the dquot speculative prealloc thresholds */
652 xfs_dquot_set_prealloc_limits(dqp);
653
654 /* Mark the buf so that this will stay incore a little longer */
655 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
656
657 /*
658 * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
659 * So we need to release with xfs_trans_brelse().
660 * The strategy here is identical to that of inodes; we lock
661 * the dquot in xfs_qm_dqget() before making it accessible to
662 * others. This is because dquots, like inodes, need a good level of
663 * concurrency, and we don't want to take locks on the entire buffers
664 * for dquot accesses.
665 * Note also that the dquot buffer may even be dirty at this point, if
666 * this particular dquot was repaired. We still aren't afraid to
667 * brelse it because we have the changes incore.
668 */
669 ASSERT(xfs_buf_islocked(bp));
670 xfs_trans_brelse(tp, bp);
671
672 if (tp) {
673 error = xfs_trans_commit(tp);
674 if (error)
675 goto error0;
676 }
677
678 *O_dqpp = dqp;
679 return error;
680
681error1:
682 if (tp)
683 xfs_trans_cancel(tp);
684error0:
685 xfs_qm_dqdestroy(dqp);
686 *O_dqpp = NULL;
687 return error;
688}
689
690/*
691 * Advance to the next id in the current chunk, or if at the
692 * end of the chunk, skip ahead to first id in next allocated chunk
693 * using the SEEK_DATA interface.
694 */
695int
696xfs_dq_get_next_id(
697 xfs_mount_t *mp,
698 uint type,
699 xfs_dqid_t *id,
700 loff_t eof)
701{
702 struct xfs_inode *quotip;
703 xfs_fsblock_t start;
704 loff_t offset;
705 uint lock;
706 xfs_dqid_t next_id;
707 int error = 0;
708
709 /* Simple advance */
710 next_id = *id + 1;
711
712 /* If new ID is within the current chunk, advancing it sufficed */
713 if (next_id % mp->m_quotainfo->qi_dqperchunk) {
714 *id = next_id;
715 return 0;
716 }
717
718 /* Nope, next_id is now past the current chunk, so find the next one */
719 start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
720
721 quotip = xfs_quota_inode(mp, type);
722 lock = xfs_ilock_data_map_shared(quotip);
723
724 offset = __xfs_seek_hole_data(VFS_I(quotip), XFS_FSB_TO_B(mp, start),
725 eof, SEEK_DATA);
726 if (offset < 0)
727 error = offset;
728
729 xfs_iunlock(quotip, lock);
730
731 /* -ENXIO is essentially "no more data" */
732 if (error)
733 return (error == -ENXIO ? -ENOENT: error);
734
735 /* Convert next data offset back to a quota id */
736 *id = XFS_B_TO_FSB(mp, offset) * mp->m_quotainfo->qi_dqperchunk;
737 return 0;
738}
739
740/*
741 * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
742 * a locked dquot, doing an allocation (if requested) as needed.
743 * When both an inode and an id are given, the inode's id takes precedence.
744 * That is, if the id changes while we don't hold the ilock inside this
745 * function, the new dquot is returned, not necessarily the one requested
746 * in the id argument.
747 */
748int
749xfs_qm_dqget(
750 xfs_mount_t *mp,
751 xfs_inode_t *ip, /* locked inode (optional) */
752 xfs_dqid_t id, /* uid/projid/gid depending on type */
753 uint type, /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
754 uint flags, /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
755 xfs_dquot_t **O_dqpp) /* OUT : locked incore dquot */
756{
757 struct xfs_quotainfo *qi = mp->m_quotainfo;
758 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
759 struct xfs_dquot *dqp;
760 loff_t eof = 0;
761 int error;
762
763 ASSERT(XFS_IS_QUOTA_RUNNING(mp));
764 if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
765 (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
766 (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
767 return -ESRCH;
768 }
769
770#ifdef DEBUG
771 if (xfs_do_dqerror) {
772 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
773 (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
774 xfs_debug(mp, "Returning error in dqget");
775 return -EIO;
776 }
777 }
778
779 ASSERT(type == XFS_DQ_USER ||
780 type == XFS_DQ_PROJ ||
781 type == XFS_DQ_GROUP);
782 if (ip) {
783 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
784 ASSERT(xfs_inode_dquot(ip, type) == NULL);
785 }
786#endif
787
788 /* Get the end of the quota file if we need it */
789 if (flags & XFS_QMOPT_DQNEXT) {
790 struct xfs_inode *quotip;
791 xfs_fileoff_t last;
792 uint lock_mode;
793
794 quotip = xfs_quota_inode(mp, type);
795 lock_mode = xfs_ilock_data_map_shared(quotip);
796 error = xfs_bmap_last_offset(quotip, &last, XFS_DATA_FORK);
797 xfs_iunlock(quotip, lock_mode);
798 if (error)
799 return error;
800 eof = XFS_FSB_TO_B(mp, last);
801 }
802
803restart:
804 mutex_lock(&qi->qi_tree_lock);
805 dqp = radix_tree_lookup(tree, id);
806 if (dqp) {
807 xfs_dqlock(dqp);
808 if (dqp->dq_flags & XFS_DQ_FREEING) {
809 xfs_dqunlock(dqp);
810 mutex_unlock(&qi->qi_tree_lock);
811 trace_xfs_dqget_freeing(dqp);
812 delay(1);
813 goto restart;
814 }
815
816 /* uninit / unused quota found in radix tree, keep looking */
817 if (flags & XFS_QMOPT_DQNEXT) {
818 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
819 xfs_dqunlock(dqp);
820 mutex_unlock(&qi->qi_tree_lock);
821 error = xfs_dq_get_next_id(mp, type, &id, eof);
822 if (error)
823 return error;
824 goto restart;
825 }
826 }
827
828 dqp->q_nrefs++;
829 mutex_unlock(&qi->qi_tree_lock);
830
831 trace_xfs_dqget_hit(dqp);
832 XFS_STATS_INC(mp, xs_qm_dqcachehits);
833 *O_dqpp = dqp;
834 return 0;
835 }
836 mutex_unlock(&qi->qi_tree_lock);
837 XFS_STATS_INC(mp, xs_qm_dqcachemisses);
838
839 /*
840 * Dquot cache miss. We don't want to keep the inode lock across
841 * a (potential) disk read. Also we don't want to deal with the lock
842 * ordering between quotainode and this inode. OTOH, dropping the inode
843 * lock here means dealing with a chown that can happen before
844 * we re-acquire the lock.
845 */
846 if (ip)
847 xfs_iunlock(ip, XFS_ILOCK_EXCL);
848
849 error = xfs_qm_dqread(mp, id, type, flags, &dqp);
850
851 if (ip)
852 xfs_ilock(ip, XFS_ILOCK_EXCL);
853
854 /* If we are asked to find next active id, keep looking */
855 if (error == -ENOENT && (flags & XFS_QMOPT_DQNEXT)) {
856 error = xfs_dq_get_next_id(mp, type, &id, eof);
857 if (!error)
858 goto restart;
859 }
860
861 if (error)
862 return error;
863
864 if (ip) {
865 /*
866 * A dquot could be attached to this inode by now, since
867 * we had dropped the ilock.
868 */
869 if (xfs_this_quota_on(mp, type)) {
870 struct xfs_dquot *dqp1;
871
872 dqp1 = xfs_inode_dquot(ip, type);
873 if (dqp1) {
874 xfs_qm_dqdestroy(dqp);
875 dqp = dqp1;
876 xfs_dqlock(dqp);
877 goto dqret;
878 }
879 } else {
880 /* inode stays locked on return */
881 xfs_qm_dqdestroy(dqp);
882 return -ESRCH;
883 }
884 }
885
886 mutex_lock(&qi->qi_tree_lock);
887 error = radix_tree_insert(tree, id, dqp);
888 if (unlikely(error)) {
889 WARN_ON(error != -EEXIST);
890
891 /*
892 * Duplicate found. Just throw away the new dquot and start
893 * over.
894 */
895 mutex_unlock(&qi->qi_tree_lock);
896 trace_xfs_dqget_dup(dqp);
897 xfs_qm_dqdestroy(dqp);
898 XFS_STATS_INC(mp, xs_qm_dquot_dups);
899 goto restart;
900 }
901
902 /*
903 * We return a locked dquot to the caller, with a reference taken
904 */
905 xfs_dqlock(dqp);
906 dqp->q_nrefs = 1;
907
908 qi->qi_dquots++;
909 mutex_unlock(&qi->qi_tree_lock);
910
911 /* If we are asked to find next active id, keep looking */
912 if (flags & XFS_QMOPT_DQNEXT) {
913 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
914 xfs_qm_dqput(dqp);
915 error = xfs_dq_get_next_id(mp, type, &id, eof);
916 if (error)
917 return error;
918 goto restart;
919 }
920 }
921
922 dqret:
923 ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
924 trace_xfs_dqget_miss(dqp);
925 *O_dqpp = dqp;
926 return 0;
927}
928
929/*
930 * Release a reference to the dquot (decrement ref-count) and unlock it.
931 *
932 * If there is a group quota attached to this dquot, carefully release that
933 * too without tripping over deadlocks'n'stuff.
934 */
935void
936xfs_qm_dqput(
937 struct xfs_dquot *dqp)
938{
939 ASSERT(dqp->q_nrefs > 0);
940 ASSERT(XFS_DQ_IS_LOCKED(dqp));
941
942 trace_xfs_dqput(dqp);
943
944 if (--dqp->q_nrefs == 0) {
945 struct xfs_quotainfo *qi = dqp->q_mount->m_quotainfo;
946 trace_xfs_dqput_free(dqp);
947
948 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
949 XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
950 }
951 xfs_dqunlock(dqp);
952}
953
954/*
955 * Release a dquot. Flush it if dirty, then dqput() it.
956 * dquot must not be locked.
957 */
958void
959xfs_qm_dqrele(
960 xfs_dquot_t *dqp)
961{
962 if (!dqp)
963 return;
964
965 trace_xfs_dqrele(dqp);
966
967 xfs_dqlock(dqp);
968 /*
969 * We don't care to flush it if the dquot is dirty here.
970 * That will create stutters that we want to avoid.
971 * Instead we do a delayed write when we try to reclaim
972 * a dirty dquot. Also xfs_sync will take part of the burden...
973 */
974 xfs_qm_dqput(dqp);
975}
976
977/*
978 * This is the dquot flushing I/O completion routine. It is called
979 * from interrupt level when the buffer containing the dquot is
980 * flushed to disk. It is responsible for removing the dquot logitem
981 * from the AIL if it has not been re-logged, and unlocking the dquot's
982 * flush lock. This behavior is very similar to that of inodes..
983 */
984STATIC void
985xfs_qm_dqflush_done(
986 struct xfs_buf *bp,
987 struct xfs_log_item *lip)
988{
989 xfs_dq_logitem_t *qip = (struct xfs_dq_logitem *)lip;
990 xfs_dquot_t *dqp = qip->qli_dquot;
991 struct xfs_ail *ailp = lip->li_ailp;
992
993 /*
994 * We only want to pull the item from the AIL if its
995 * location in the log has not changed since we started the flush.
996 * Thus, we only bother if the dquot's lsn has
997 * not changed. First we check the lsn outside the lock
998 * since it's cheaper, and then we recheck while
999 * holding the lock before removing the dquot from the AIL.
1000 */
1001 if ((lip->li_flags & XFS_LI_IN_AIL) &&
1002 lip->li_lsn == qip->qli_flush_lsn) {
1003
1004 /* xfs_trans_ail_delete() drops the AIL lock. */
1005 spin_lock(&ailp->xa_lock);
1006 if (lip->li_lsn == qip->qli_flush_lsn)
1007 xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1008 else
1009 spin_unlock(&ailp->xa_lock);
1010 }
1011
1012 /*
1013 * Release the dq's flush lock since we're done with it.
1014 */
1015 xfs_dqfunlock(dqp);
1016}
1017
1018/*
1019 * Write a modified dquot to disk.
1020 * The dquot must be locked and the flush lock too taken by caller.
1021 * The flush lock will not be unlocked until the dquot reaches the disk,
1022 * but the dquot is free to be unlocked and modified by the caller
1023 * in the interim. Dquot is still locked on return. This behavior is
1024 * identical to that of inodes.
1025 */
1026int
1027xfs_qm_dqflush(
1028 struct xfs_dquot *dqp,
1029 struct xfs_buf **bpp)
1030{
1031 struct xfs_mount *mp = dqp->q_mount;
1032 struct xfs_buf *bp;
1033 struct xfs_disk_dquot *ddqp;
1034 int error;
1035
1036 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1037 ASSERT(!completion_done(&dqp->q_flush));
1038
1039 trace_xfs_dqflush(dqp);
1040
1041 *bpp = NULL;
1042
1043 xfs_qm_dqunpin_wait(dqp);
1044
1045 /*
1046 * This may have been unpinned because the filesystem is shutting
1047 * down forcibly. If that's the case we must not write this dquot
1048 * to disk, because the log record didn't make it to disk.
1049 *
1050 * We also have to remove the log item from the AIL in this case,
1051 * as we wait for an emptry AIL as part of the unmount process.
1052 */
1053 if (XFS_FORCED_SHUTDOWN(mp)) {
1054 struct xfs_log_item *lip = &dqp->q_logitem.qli_item;
1055 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1056
1057 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
1058
1059 error = -EIO;
1060 goto out_unlock;
1061 }
1062
1063 /*
1064 * Get the buffer containing the on-disk dquot
1065 */
1066 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1067 mp->m_quotainfo->qi_dqchunklen, 0, &bp,
1068 &xfs_dquot_buf_ops);
1069 if (error)
1070 goto out_unlock;
1071
1072 /*
1073 * Calculate the location of the dquot inside the buffer.
1074 */
1075 ddqp = bp->b_addr + dqp->q_bufoffset;
1076
1077 /*
1078 * A simple sanity check in case we got a corrupted dquot..
1079 */
1080 error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1081 XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1082 if (error) {
1083 xfs_buf_relse(bp);
1084 xfs_dqfunlock(dqp);
1085 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1086 return -EIO;
1087 }
1088
1089 /* This is the only portion of data that needs to persist */
1090 memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1091
1092 /*
1093 * Clear the dirty field and remember the flush lsn for later use.
1094 */
1095 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1096
1097 xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1098 &dqp->q_logitem.qli_item.li_lsn);
1099
1100 /*
1101 * copy the lsn into the on-disk dquot now while we have the in memory
1102 * dquot here. This can't be done later in the write verifier as we
1103 * can't get access to the log item at that point in time.
1104 *
1105 * We also calculate the CRC here so that the on-disk dquot in the
1106 * buffer always has a valid CRC. This ensures there is no possibility
1107 * of a dquot without an up-to-date CRC getting to disk.
1108 */
1109 if (xfs_sb_version_hascrc(&mp->m_sb)) {
1110 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1111
1112 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1113 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1114 XFS_DQUOT_CRC_OFF);
1115 }
1116
1117 /*
1118 * Attach an iodone routine so that we can remove this dquot from the
1119 * AIL and release the flush lock once the dquot is synced to disk.
1120 */
1121 xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1122 &dqp->q_logitem.qli_item);
1123
1124 /*
1125 * If the buffer is pinned then push on the log so we won't
1126 * get stuck waiting in the write for too long.
1127 */
1128 if (xfs_buf_ispinned(bp)) {
1129 trace_xfs_dqflush_force(dqp);
1130 xfs_log_force(mp, 0);
1131 }
1132
1133 trace_xfs_dqflush_done(dqp);
1134 *bpp = bp;
1135 return 0;
1136
1137out_unlock:
1138 xfs_dqfunlock(dqp);
1139 return -EIO;
1140}
1141
1142/*
1143 * Lock two xfs_dquot structures.
1144 *
1145 * To avoid deadlocks we always lock the quota structure with
1146 * the lowerd id first.
1147 */
1148void
1149xfs_dqlock2(
1150 xfs_dquot_t *d1,
1151 xfs_dquot_t *d2)
1152{
1153 if (d1 && d2) {
1154 ASSERT(d1 != d2);
1155 if (be32_to_cpu(d1->q_core.d_id) >
1156 be32_to_cpu(d2->q_core.d_id)) {
1157 mutex_lock(&d2->q_qlock);
1158 mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1159 } else {
1160 mutex_lock(&d1->q_qlock);
1161 mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1162 }
1163 } else if (d1) {
1164 mutex_lock(&d1->q_qlock);
1165 } else if (d2) {
1166 mutex_lock(&d2->q_qlock);
1167 }
1168}
1169
1170int __init
1171xfs_qm_init(void)
1172{
1173 xfs_qm_dqzone =
1174 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1175 if (!xfs_qm_dqzone)
1176 goto out;
1177
1178 xfs_qm_dqtrxzone =
1179 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1180 if (!xfs_qm_dqtrxzone)
1181 goto out_free_dqzone;
1182
1183 return 0;
1184
1185out_free_dqzone:
1186 kmem_zone_destroy(xfs_qm_dqzone);
1187out:
1188 return -ENOMEM;
1189}
1190
1191void
1192xfs_qm_exit(void)
1193{
1194 kmem_zone_destroy(xfs_qm_dqtrxzone);
1195 kmem_zone_destroy(xfs_qm_dqzone);
1196}