1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2003 Silicon Graphics, Inc.
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_shared.h"
11 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_defer.h"
15 #include "xfs_inode.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"
23 #include "xfs_trace.h"
25 #include "xfs_bmap_btree.h"
26 #include "xfs_error.h"
27 #include "xfs_health.h"
34 * dquot->q_qlock (xfs_dqlock() and friends)
35 * dquot->q_flush (xfs_dqflock() and friends)
38 * If two dquots need to be locked the order is user before group/project,
39 * otherwise by the lowest id first, see xfs_dqlock2.
42 struct kmem_cache *xfs_dqtrx_cache;
43 static struct kmem_cache *xfs_dquot_cache;
45 static struct lock_class_key xfs_dquot_group_class;
46 static struct lock_class_key xfs_dquot_project_class;
48 /* Record observations of quota corruption with the health tracking system. */
51 struct xfs_dquot *dqp)
53 struct xfs_mount *mp = dqp->q_mount;
55 switch (dqp->q_type) {
57 xfs_fs_mark_sick(mp, XFS_SICK_FS_UQUOTA);
59 case XFS_DQTYPE_GROUP:
60 xfs_fs_mark_sick(mp, XFS_SICK_FS_GQUOTA);
63 xfs_fs_mark_sick(mp, XFS_SICK_FS_PQUOTA);
72 * This is called to free all the memory associated with a dquot
76 struct xfs_dquot *dqp)
78 ASSERT(list_empty(&dqp->q_lru));
80 kvfree(dqp->q_logitem.qli_item.li_lv_shadow);
81 mutex_destroy(&dqp->q_qlock);
83 XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
84 kmem_cache_free(xfs_dquot_cache, dqp);
88 * If default limits are in force, push them into the dquot now.
89 * We overwrite the dquot limits only if they are zero and this
90 * is not the root dquot.
93 xfs_qm_adjust_dqlimits(
96 struct xfs_mount *mp = dq->q_mount;
97 struct xfs_quotainfo *q = mp->m_quotainfo;
98 struct xfs_def_quota *defq;
102 defq = xfs_get_defquota(q, xfs_dquot_type(dq));
104 if (!dq->q_blk.softlimit) {
105 dq->q_blk.softlimit = defq->blk.soft;
108 if (!dq->q_blk.hardlimit) {
109 dq->q_blk.hardlimit = defq->blk.hard;
112 if (!dq->q_ino.softlimit)
113 dq->q_ino.softlimit = defq->ino.soft;
114 if (!dq->q_ino.hardlimit)
115 dq->q_ino.hardlimit = defq->ino.hard;
116 if (!dq->q_rtb.softlimit)
117 dq->q_rtb.softlimit = defq->rtb.soft;
118 if (!dq->q_rtb.hardlimit)
119 dq->q_rtb.hardlimit = defq->rtb.hard;
122 xfs_dquot_set_prealloc_limits(dq);
125 /* Set the expiration time of a quota's grace period. */
127 xfs_dquot_set_timeout(
128 struct xfs_mount *mp,
131 struct xfs_quotainfo *qi = mp->m_quotainfo;
133 return clamp_t(time64_t, timeout, qi->qi_expiry_min,
137 /* Set the length of the default grace period. */
139 xfs_dquot_set_grace_period(
142 return clamp_t(time64_t, grace, XFS_DQ_GRACE_MIN, XFS_DQ_GRACE_MAX);
146 * Determine if this quota counter is over either limit and set the quota
147 * timers as appropriate.
150 xfs_qm_adjust_res_timer(
151 struct xfs_mount *mp,
152 struct xfs_dquot_res *res,
153 struct xfs_quota_limits *qlim)
155 ASSERT(res->hardlimit == 0 || res->softlimit <= res->hardlimit);
157 if ((res->softlimit && res->count > res->softlimit) ||
158 (res->hardlimit && res->count > res->hardlimit)) {
160 res->timer = xfs_dquot_set_timeout(mp,
161 ktime_get_real_seconds() + qlim->time);
168 * Check the limits and timers of a dquot and start or reset timers
170 * This gets called even when quota enforcement is OFF, which makes our
171 * life a little less complicated. (We just don't reject any quota
172 * reservations in that case, when enforcement is off).
173 * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
175 * In contrast, warnings are a little different in that they don't
176 * 'automatically' get started when limits get exceeded. They do
177 * get reset to zero, however, when we find the count to be under
178 * the soft limit (they are only ever set non-zero via userspace).
181 xfs_qm_adjust_dqtimers(
182 struct xfs_dquot *dq)
184 struct xfs_mount *mp = dq->q_mount;
185 struct xfs_quotainfo *qi = mp->m_quotainfo;
186 struct xfs_def_quota *defq;
189 defq = xfs_get_defquota(qi, xfs_dquot_type(dq));
191 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_blk, &defq->blk);
192 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_ino, &defq->ino);
193 xfs_qm_adjust_res_timer(dq->q_mount, &dq->q_rtb, &defq->rtb);
197 * initialize a buffer full of dquots and log the whole thing
200 xfs_qm_init_dquot_blk(
201 struct xfs_trans *tp,
206 struct xfs_mount *mp = tp->t_mountp;
207 struct xfs_quotainfo *q = mp->m_quotainfo;
211 unsigned int blftype;
215 ASSERT(xfs_buf_islocked(bp));
218 case XFS_DQTYPE_USER:
219 qflag = XFS_UQUOTA_CHKD;
220 blftype = XFS_BLF_UDQUOT_BUF;
222 case XFS_DQTYPE_PROJ:
223 qflag = XFS_PQUOTA_CHKD;
224 blftype = XFS_BLF_PDQUOT_BUF;
226 case XFS_DQTYPE_GROUP:
227 qflag = XFS_GQUOTA_CHKD;
228 blftype = XFS_BLF_GDQUOT_BUF;
238 * ID of the first dquot in the block - id's are zero based.
240 curid = id - (id % q->qi_dqperchunk);
241 memset(d, 0, BBTOB(q->qi_dqchunklen));
242 for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
243 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
244 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
245 d->dd_diskdq.d_id = cpu_to_be32(curid);
246 d->dd_diskdq.d_type = type;
247 if (curid > 0 && xfs_has_bigtime(mp))
248 d->dd_diskdq.d_type |= XFS_DQTYPE_BIGTIME;
249 if (xfs_has_crc(mp)) {
250 uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
251 xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
256 xfs_trans_dquot_buf(tp, bp, blftype);
259 * quotacheck uses delayed writes to update all the dquots on disk in an
260 * efficient manner instead of logging the individual dquot changes as
261 * they are made. However if we log the buffer allocated here and crash
262 * after quotacheck while the logged initialisation is still in the
263 * active region of the log, log recovery can replay the dquot buffer
264 * initialisation over the top of the checked dquots and corrupt quota
267 * To avoid this problem, quotacheck cannot log the initialised buffer.
268 * We must still dirty the buffer and write it back before the
269 * allocation transaction clears the log. Therefore, mark the buffer as
270 * ordered instead of logging it directly. This is safe for quotacheck
271 * because it detects and repairs allocated but initialized dquot blocks
272 * in the quota inodes.
274 if (!(mp->m_qflags & qflag))
275 xfs_trans_ordered_buf(tp, bp);
277 xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
281 * Initialize the dynamic speculative preallocation thresholds. The lo/hi
282 * watermarks correspond to the soft and hard limits by default. If a soft limit
283 * is not specified, we use 95% of the hard limit.
286 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
290 dqp->q_prealloc_hi_wmark = dqp->q_blk.hardlimit;
291 dqp->q_prealloc_lo_wmark = dqp->q_blk.softlimit;
292 if (!dqp->q_prealloc_lo_wmark) {
293 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
294 do_div(dqp->q_prealloc_lo_wmark, 100);
295 dqp->q_prealloc_lo_wmark *= 95;
298 space = dqp->q_prealloc_hi_wmark;
301 dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
302 dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
303 dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
307 * Ensure that the given in-core dquot has a buffer on disk backing it, and
308 * return the buffer locked and held. This is called when the bmapi finds a
312 xfs_dquot_disk_alloc(
313 struct xfs_dquot *dqp,
314 struct xfs_buf **bpp)
316 struct xfs_bmbt_irec map;
317 struct xfs_trans *tp;
318 struct xfs_mount *mp = dqp->q_mount;
320 xfs_dqtype_t qtype = xfs_dquot_type(dqp);
321 struct xfs_inode *quotip = xfs_quota_inode(mp, qtype);
325 trace_xfs_dqalloc(dqp);
327 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
328 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
332 xfs_ilock(quotip, XFS_ILOCK_EXCL);
333 xfs_trans_ijoin(tp, quotip, 0);
335 if (!xfs_this_quota_on(dqp->q_mount, qtype)) {
337 * Return if this type of quotas is turned off while we didn't
344 error = xfs_iext_count_may_overflow(quotip, XFS_DATA_FORK,
345 XFS_IEXT_ADD_NOSPLIT_CNT);
347 error = xfs_iext_count_upgrade(tp, quotip,
348 XFS_IEXT_ADD_NOSPLIT_CNT);
352 /* Create the block mapping. */
353 error = xfs_bmapi_write(tp, quotip, dqp->q_fileoffset,
354 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA, 0, &map,
359 ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
361 ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
362 (map.br_startblock != HOLESTARTBLOCK));
365 * Keep track of the blkno to save a lookup later
367 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
369 /* now we can just get the buffer (there's nothing to read yet) */
370 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, dqp->q_blkno,
371 mp->m_quotainfo->qi_dqchunklen, 0, &bp);
374 bp->b_ops = &xfs_dquot_buf_ops;
377 * Make a chunk of dquots out of this buffer and log
380 xfs_qm_init_dquot_blk(tp, dqp->q_id, qtype, bp);
381 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
384 * Hold the buffer and join it to the dfops so that we'll still own
385 * the buffer when we return to the caller. The buffer disposal on
386 * error must be paid attention to very carefully, as it has been
387 * broken since commit efa092f3d4c6 "[XFS] Fixes a bug in the quota
388 * code when allocating a new dquot record" in 2005, and the later
389 * conversion to xfs_defer_ops in commit 310a75a3c6c747 failed to keep
390 * the buffer locked across the _defer_finish call. We can now do
391 * this correctly with xfs_defer_bjoin.
393 * Above, we allocated a disk block for the dquot information and used
394 * get_buf to initialize the dquot. If the _defer_finish fails, the old
395 * transaction is gone but the new buffer is not joined or held to any
396 * transaction, so we must _buf_relse it.
398 * If everything succeeds, the caller of this function is returned a
399 * buffer that is locked and held to the transaction. The caller
400 * is responsible for unlocking any buffer passed back, either
401 * manually or by committing the transaction. On error, the buffer is
402 * released and not passed back.
404 * Keep the quota inode ILOCKed until after the transaction commit to
405 * maintain the atomicity of bmap/rmap updates.
407 xfs_trans_bhold(tp, bp);
408 error = xfs_trans_commit(tp);
409 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
419 xfs_trans_cancel(tp);
420 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
425 * Read in the in-core dquot's on-disk metadata and return the buffer.
426 * Returns ENOENT to signal a hole.
430 struct xfs_mount *mp,
431 struct xfs_dquot *dqp,
432 struct xfs_buf **bpp)
434 struct xfs_bmbt_irec map;
436 xfs_dqtype_t qtype = xfs_dquot_type(dqp);
437 struct xfs_inode *quotip = xfs_quota_inode(mp, qtype);
442 lock_mode = xfs_ilock_data_map_shared(quotip);
443 if (!xfs_this_quota_on(mp, qtype)) {
445 * Return if this type of quotas is turned off while we
446 * didn't have the quota inode lock.
448 xfs_iunlock(quotip, lock_mode);
453 * Find the block map; no allocations yet
455 error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
456 XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
457 xfs_iunlock(quotip, lock_mode);
462 ASSERT(map.br_blockcount >= 1);
463 ASSERT(map.br_startblock != DELAYSTARTBLOCK);
464 if (map.br_startblock == HOLESTARTBLOCK)
467 trace_xfs_dqtobp_read(dqp);
470 * store the blkno etc so that we don't have to do the
471 * mapping all the time
473 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
475 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
476 mp->m_quotainfo->qi_dqchunklen, 0, &bp,
478 if (xfs_metadata_is_sick(error))
479 xfs_dquot_mark_sick(dqp);
485 ASSERT(xfs_buf_islocked(bp));
486 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
492 /* Allocate and initialize everything we need for an incore dquot. */
493 STATIC struct xfs_dquot *
495 struct xfs_mount *mp,
499 struct xfs_dquot *dqp;
501 dqp = kmem_cache_zalloc(xfs_dquot_cache, GFP_KERNEL | __GFP_NOFAIL);
506 INIT_LIST_HEAD(&dqp->q_lru);
507 mutex_init(&dqp->q_qlock);
508 init_waitqueue_head(&dqp->q_pinwait);
509 dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
511 * Offset of dquot in the (fixed sized) dquot chunk.
513 dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
514 sizeof(struct xfs_dqblk);
517 * Because we want to use a counting completion, complete
518 * the flush completion once to allow a single access to
519 * the flush completion without blocking.
521 init_completion(&dqp->q_flush);
522 complete(&dqp->q_flush);
525 * Make sure group quotas have a different lock class than user
529 case XFS_DQTYPE_USER:
530 /* uses the default lock class */
532 case XFS_DQTYPE_GROUP:
533 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
535 case XFS_DQTYPE_PROJ:
536 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
543 xfs_qm_dquot_logitem_init(dqp);
545 XFS_STATS_INC(mp, xs_qm_dquot);
549 /* Check the ondisk dquot's id and type match what the incore dquot expects. */
551 xfs_dquot_check_type(
552 struct xfs_dquot *dqp,
553 struct xfs_disk_dquot *ddqp)
558 ddqp_type = ddqp->d_type & XFS_DQTYPE_REC_MASK;
559 dqp_type = xfs_dquot_type(dqp);
561 if (be32_to_cpu(ddqp->d_id) != dqp->q_id)
565 * V5 filesystems always expect an exact type match. V4 filesystems
566 * expect an exact match for user dquots and for non-root group and
569 if (xfs_has_crc(dqp->q_mount) ||
570 dqp_type == XFS_DQTYPE_USER || dqp->q_id != 0)
571 return ddqp_type == dqp_type;
574 * V4 filesystems support either group or project quotas, but not both
575 * at the same time. The non-user quota file can be switched between
576 * group and project quota uses depending on the mount options, which
577 * means that we can encounter the other type when we try to load quota
578 * defaults. Quotacheck will soon reset the entire quota file
579 * (including the root dquot) anyway, but don't log scary corruption
582 return ddqp_type == XFS_DQTYPE_GROUP || ddqp_type == XFS_DQTYPE_PROJ;
585 /* Copy the in-core quota fields in from the on-disk buffer. */
588 struct xfs_dquot *dqp,
591 struct xfs_dqblk *dqb = xfs_buf_offset(bp, dqp->q_bufoffset);
592 struct xfs_disk_dquot *ddqp = &dqb->dd_diskdq;
595 * Ensure that we got the type and ID we were looking for.
596 * Everything else was checked by the dquot buffer verifier.
598 if (!xfs_dquot_check_type(dqp, ddqp)) {
599 xfs_alert_tag(bp->b_mount, XFS_PTAG_VERIFIER_ERROR,
600 "Metadata corruption detected at %pS, quota %u",
601 __this_address, dqp->q_id);
602 xfs_alert(bp->b_mount, "Unmount and run xfs_repair");
603 xfs_dquot_mark_sick(dqp);
604 return -EFSCORRUPTED;
607 /* copy everything from disk dquot to the incore dquot */
608 dqp->q_type = ddqp->d_type;
609 dqp->q_blk.hardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
610 dqp->q_blk.softlimit = be64_to_cpu(ddqp->d_blk_softlimit);
611 dqp->q_ino.hardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
612 dqp->q_ino.softlimit = be64_to_cpu(ddqp->d_ino_softlimit);
613 dqp->q_rtb.hardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
614 dqp->q_rtb.softlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
616 dqp->q_blk.count = be64_to_cpu(ddqp->d_bcount);
617 dqp->q_ino.count = be64_to_cpu(ddqp->d_icount);
618 dqp->q_rtb.count = be64_to_cpu(ddqp->d_rtbcount);
620 dqp->q_blk.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_btimer);
621 dqp->q_ino.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_itimer);
622 dqp->q_rtb.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_rtbtimer);
625 * Reservation counters are defined as reservation plus current usage
626 * to avoid having to add every time.
628 dqp->q_blk.reserved = dqp->q_blk.count;
629 dqp->q_ino.reserved = dqp->q_ino.count;
630 dqp->q_rtb.reserved = dqp->q_rtb.count;
632 /* initialize the dquot speculative prealloc thresholds */
633 xfs_dquot_set_prealloc_limits(dqp);
637 /* Copy the in-core quota fields into the on-disk buffer. */
640 struct xfs_disk_dquot *ddqp,
641 struct xfs_dquot *dqp)
643 ddqp->d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
644 ddqp->d_version = XFS_DQUOT_VERSION;
645 ddqp->d_type = dqp->q_type;
646 ddqp->d_id = cpu_to_be32(dqp->q_id);
650 ddqp->d_blk_hardlimit = cpu_to_be64(dqp->q_blk.hardlimit);
651 ddqp->d_blk_softlimit = cpu_to_be64(dqp->q_blk.softlimit);
652 ddqp->d_ino_hardlimit = cpu_to_be64(dqp->q_ino.hardlimit);
653 ddqp->d_ino_softlimit = cpu_to_be64(dqp->q_ino.softlimit);
654 ddqp->d_rtb_hardlimit = cpu_to_be64(dqp->q_rtb.hardlimit);
655 ddqp->d_rtb_softlimit = cpu_to_be64(dqp->q_rtb.softlimit);
657 ddqp->d_bcount = cpu_to_be64(dqp->q_blk.count);
658 ddqp->d_icount = cpu_to_be64(dqp->q_ino.count);
659 ddqp->d_rtbcount = cpu_to_be64(dqp->q_rtb.count);
663 ddqp->d_rtbwarns = 0;
665 ddqp->d_btimer = xfs_dquot_to_disk_ts(dqp, dqp->q_blk.timer);
666 ddqp->d_itimer = xfs_dquot_to_disk_ts(dqp, dqp->q_ino.timer);
667 ddqp->d_rtbtimer = xfs_dquot_to_disk_ts(dqp, dqp->q_rtb.timer);
671 * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
672 * and release the buffer immediately. If @can_alloc is true, fill any
673 * holes in the on-disk metadata.
677 struct xfs_mount *mp,
681 struct xfs_dquot **dqpp)
683 struct xfs_dquot *dqp;
687 dqp = xfs_dquot_alloc(mp, id, type);
688 trace_xfs_dqread(dqp);
690 /* Try to read the buffer, allocating if necessary. */
691 error = xfs_dquot_disk_read(mp, dqp, &bp);
692 if (error == -ENOENT && can_alloc)
693 error = xfs_dquot_disk_alloc(dqp, &bp);
698 * At this point we should have a clean locked buffer. Copy the data
699 * to the incore dquot and release the buffer since the incore dquot
700 * has its own locking protocol so we needn't tie up the buffer any
703 ASSERT(xfs_buf_islocked(bp));
704 error = xfs_dquot_from_disk(dqp, bp);
713 trace_xfs_dqread_fail(dqp);
714 xfs_qm_dqdestroy(dqp);
720 * Advance to the next id in the current chunk, or if at the
721 * end of the chunk, skip ahead to first id in next allocated chunk
722 * using the SEEK_DATA interface.
726 struct xfs_mount *mp,
730 struct xfs_inode *quotip = xfs_quota_inode(mp, type);
731 xfs_dqid_t next_id = *id + 1; /* simple advance */
733 struct xfs_bmbt_irec got;
734 struct xfs_iext_cursor cur;
738 /* If we'd wrap past the max ID, stop */
742 /* If new ID is within the current chunk, advancing it sufficed */
743 if (next_id % mp->m_quotainfo->qi_dqperchunk) {
748 /* Nope, next_id is now past the current chunk, so find the next one */
749 start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
751 lock_flags = xfs_ilock_data_map_shared(quotip);
752 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
756 if (xfs_iext_lookup_extent(quotip, "ip->i_df, start, &cur, &got)) {
757 /* contiguous chunk, bump startoff for the id calculation */
758 if (got.br_startoff < start)
759 got.br_startoff = start;
760 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
765 xfs_iunlock(quotip, lock_flags);
771 * Look up the dquot in the in-core cache. If found, the dquot is returned
772 * locked and ready to go.
774 static struct xfs_dquot *
775 xfs_qm_dqget_cache_lookup(
776 struct xfs_mount *mp,
777 struct xfs_quotainfo *qi,
778 struct radix_tree_root *tree,
781 struct xfs_dquot *dqp;
784 mutex_lock(&qi->qi_tree_lock);
785 dqp = radix_tree_lookup(tree, id);
787 mutex_unlock(&qi->qi_tree_lock);
788 XFS_STATS_INC(mp, xs_qm_dqcachemisses);
793 if (dqp->q_flags & XFS_DQFLAG_FREEING) {
795 mutex_unlock(&qi->qi_tree_lock);
796 trace_xfs_dqget_freeing(dqp);
802 mutex_unlock(&qi->qi_tree_lock);
804 trace_xfs_dqget_hit(dqp);
805 XFS_STATS_INC(mp, xs_qm_dqcachehits);
810 * Try to insert a new dquot into the in-core cache. If an error occurs the
811 * caller should throw away the dquot and start over. Otherwise, the dquot
812 * is returned locked (and held by the cache) as if there had been a cache
815 * The insert needs to be done under memalloc_nofs context because the radix
816 * tree can do memory allocation during insert. The qi->qi_tree_lock is taken in
817 * memory reclaim when freeing unused dquots, so we cannot have the radix tree
818 * node allocation recursing into filesystem reclaim whilst we hold the
822 xfs_qm_dqget_cache_insert(
823 struct xfs_mount *mp,
824 struct xfs_quotainfo *qi,
825 struct radix_tree_root *tree,
827 struct xfs_dquot *dqp)
829 unsigned int nofs_flags;
832 nofs_flags = memalloc_nofs_save();
833 mutex_lock(&qi->qi_tree_lock);
834 error = radix_tree_insert(tree, id, dqp);
835 if (unlikely(error)) {
836 /* Duplicate found! Caller must try again. */
837 trace_xfs_dqget_dup(dqp);
841 /* Return a locked dquot to the caller, with a reference taken. */
847 mutex_unlock(&qi->qi_tree_lock);
848 memalloc_nofs_restore(nofs_flags);
852 /* Check our input parameters. */
855 struct xfs_mount *mp,
859 case XFS_DQTYPE_USER:
860 if (!XFS_IS_UQUOTA_ON(mp))
863 case XFS_DQTYPE_GROUP:
864 if (!XFS_IS_GQUOTA_ON(mp))
867 case XFS_DQTYPE_PROJ:
868 if (!XFS_IS_PQUOTA_ON(mp))
878 * Given the file system, id, and type (UDQUOT/GDQUOT/PDQUOT), return a
879 * locked dquot, doing an allocation (if requested) as needed.
883 struct xfs_mount *mp,
887 struct xfs_dquot **O_dqpp)
889 struct xfs_quotainfo *qi = mp->m_quotainfo;
890 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
891 struct xfs_dquot *dqp;
894 error = xfs_qm_dqget_checks(mp, type);
899 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
905 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
909 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
912 * Duplicate found. Just throw away the new dquot and start
915 xfs_qm_dqdestroy(dqp);
916 XFS_STATS_INC(mp, xs_qm_dquot_dups);
920 trace_xfs_dqget_miss(dqp);
926 * Given a dquot id and type, read and initialize a dquot from the on-disk
927 * metadata. This function is only for use during quota initialization so
928 * it ignores the dquot cache assuming that the dquot shrinker isn't set up.
929 * The caller is responsible for _qm_dqdestroy'ing the returned dquot.
932 xfs_qm_dqget_uncached(
933 struct xfs_mount *mp,
936 struct xfs_dquot **dqpp)
940 error = xfs_qm_dqget_checks(mp, type);
944 return xfs_qm_dqread(mp, id, type, 0, dqpp);
947 /* Return the quota id for a given inode and type. */
949 xfs_qm_id_for_quotatype(
950 struct xfs_inode *ip,
954 case XFS_DQTYPE_USER:
955 return i_uid_read(VFS_I(ip));
956 case XFS_DQTYPE_GROUP:
957 return i_gid_read(VFS_I(ip));
958 case XFS_DQTYPE_PROJ:
966 * Return the dquot for a given inode and type. If @can_alloc is true, then
967 * allocate blocks if needed. The inode's ILOCK must be held and it must not
968 * have already had an inode attached.
972 struct xfs_inode *ip,
975 struct xfs_dquot **O_dqpp)
977 struct xfs_mount *mp = ip->i_mount;
978 struct xfs_quotainfo *qi = mp->m_quotainfo;
979 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
980 struct xfs_dquot *dqp;
984 error = xfs_qm_dqget_checks(mp, type);
988 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
989 ASSERT(xfs_inode_dquot(ip, type) == NULL);
991 id = xfs_qm_id_for_quotatype(ip, type);
994 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
1001 * Dquot cache miss. We don't want to keep the inode lock across
1002 * a (potential) disk read. Also we don't want to deal with the lock
1003 * ordering between quotainode and this inode. OTOH, dropping the inode
1004 * lock here means dealing with a chown that can happen before
1005 * we re-acquire the lock.
1007 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1008 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
1009 xfs_ilock(ip, XFS_ILOCK_EXCL);
1014 * A dquot could be attached to this inode by now, since we had
1015 * dropped the ilock.
1017 if (xfs_this_quota_on(mp, type)) {
1018 struct xfs_dquot *dqp1;
1020 dqp1 = xfs_inode_dquot(ip, type);
1022 xfs_qm_dqdestroy(dqp);
1028 /* inode stays locked on return */
1029 xfs_qm_dqdestroy(dqp);
1033 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
1036 * Duplicate found. Just throw away the new dquot and start
1039 xfs_qm_dqdestroy(dqp);
1040 XFS_STATS_INC(mp, xs_qm_dquot_dups);
1045 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1046 trace_xfs_dqget_miss(dqp);
1052 * Starting at @id and progressing upwards, look for an initialized incore
1053 * dquot, lock it, and return it.
1057 struct xfs_mount *mp,
1060 struct xfs_dquot **dqpp)
1062 struct xfs_dquot *dqp;
1066 for (; !error; error = xfs_dq_get_next_id(mp, type, &id)) {
1067 error = xfs_qm_dqget(mp, id, type, false, &dqp);
1068 if (error == -ENOENT)
1070 else if (error != 0)
1073 if (!XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
1085 * Release a reference to the dquot (decrement ref-count) and unlock it.
1087 * If there is a group quota attached to this dquot, carefully release that
1088 * too without tripping over deadlocks'n'stuff.
1092 struct xfs_dquot *dqp)
1094 ASSERT(dqp->q_nrefs > 0);
1095 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1097 trace_xfs_dqput(dqp);
1099 if (--dqp->q_nrefs == 0) {
1100 struct xfs_quotainfo *qi = dqp->q_mount->m_quotainfo;
1101 trace_xfs_dqput_free(dqp);
1103 if (list_lru_add_obj(&qi->qi_lru, &dqp->q_lru))
1104 XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
1110 * Release a dquot. Flush it if dirty, then dqput() it.
1111 * dquot must not be locked.
1115 struct xfs_dquot *dqp)
1120 trace_xfs_dqrele(dqp);
1124 * We don't care to flush it if the dquot is dirty here.
1125 * That will create stutters that we want to avoid.
1126 * Instead we do a delayed write when we try to reclaim
1127 * a dirty dquot. Also xfs_sync will take part of the burden...
1133 * This is the dquot flushing I/O completion routine. It is called
1134 * from interrupt level when the buffer containing the dquot is
1135 * flushed to disk. It is responsible for removing the dquot logitem
1136 * from the AIL if it has not been re-logged, and unlocking the dquot's
1137 * flush lock. This behavior is very similar to that of inodes..
1140 xfs_qm_dqflush_done(
1141 struct xfs_log_item *lip)
1143 struct xfs_dq_logitem *qip = (struct xfs_dq_logitem *)lip;
1144 struct xfs_dquot *dqp = qip->qli_dquot;
1145 struct xfs_ail *ailp = lip->li_ailp;
1149 * We only want to pull the item from the AIL if its
1150 * location in the log has not changed since we started the flush.
1151 * Thus, we only bother if the dquot's lsn has
1152 * not changed. First we check the lsn outside the lock
1153 * since it's cheaper, and then we recheck while
1154 * holding the lock before removing the dquot from the AIL.
1156 if (test_bit(XFS_LI_IN_AIL, &lip->li_flags) &&
1157 ((lip->li_lsn == qip->qli_flush_lsn) ||
1158 test_bit(XFS_LI_FAILED, &lip->li_flags))) {
1160 spin_lock(&ailp->ail_lock);
1161 xfs_clear_li_failed(lip);
1162 if (lip->li_lsn == qip->qli_flush_lsn) {
1163 /* xfs_ail_update_finish() drops the AIL lock */
1164 tail_lsn = xfs_ail_delete_one(ailp, lip);
1165 xfs_ail_update_finish(ailp, tail_lsn);
1167 spin_unlock(&ailp->ail_lock);
1172 * Release the dq's flush lock since we're done with it.
1178 xfs_buf_dquot_iodone(
1181 struct xfs_log_item *lip, *n;
1183 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
1184 list_del_init(&lip->li_bio_list);
1185 xfs_qm_dqflush_done(lip);
1190 xfs_buf_dquot_io_fail(
1193 struct xfs_log_item *lip;
1195 spin_lock(&bp->b_mount->m_ail->ail_lock);
1196 list_for_each_entry(lip, &bp->b_li_list, li_bio_list)
1197 xfs_set_li_failed(lip, bp);
1198 spin_unlock(&bp->b_mount->m_ail->ail_lock);
1201 /* Check incore dquot for errors before we flush. */
1202 static xfs_failaddr_t
1203 xfs_qm_dqflush_check(
1204 struct xfs_dquot *dqp)
1206 xfs_dqtype_t type = xfs_dquot_type(dqp);
1208 if (type != XFS_DQTYPE_USER &&
1209 type != XFS_DQTYPE_GROUP &&
1210 type != XFS_DQTYPE_PROJ)
1211 return __this_address;
1216 if (dqp->q_blk.softlimit && dqp->q_blk.count > dqp->q_blk.softlimit &&
1218 return __this_address;
1220 if (dqp->q_ino.softlimit && dqp->q_ino.count > dqp->q_ino.softlimit &&
1222 return __this_address;
1224 if (dqp->q_rtb.softlimit && dqp->q_rtb.count > dqp->q_rtb.softlimit &&
1226 return __this_address;
1228 /* bigtime flag should never be set on root dquots */
1229 if (dqp->q_type & XFS_DQTYPE_BIGTIME) {
1230 if (!xfs_has_bigtime(dqp->q_mount))
1231 return __this_address;
1233 return __this_address;
1240 * Write a modified dquot to disk.
1241 * The dquot must be locked and the flush lock too taken by caller.
1242 * The flush lock will not be unlocked until the dquot reaches the disk,
1243 * but the dquot is free to be unlocked and modified by the caller
1244 * in the interim. Dquot is still locked on return. This behavior is
1245 * identical to that of inodes.
1249 struct xfs_dquot *dqp,
1250 struct xfs_buf **bpp)
1252 struct xfs_mount *mp = dqp->q_mount;
1253 struct xfs_log_item *lip = &dqp->q_logitem.qli_item;
1255 struct xfs_dqblk *dqblk;
1259 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1260 ASSERT(!completion_done(&dqp->q_flush));
1262 trace_xfs_dqflush(dqp);
1266 xfs_qm_dqunpin_wait(dqp);
1269 * Get the buffer containing the on-disk dquot
1271 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1272 mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK,
1273 &bp, &xfs_dquot_buf_ops);
1274 if (error == -EAGAIN)
1276 if (xfs_metadata_is_sick(error))
1277 xfs_dquot_mark_sick(dqp);
1281 fa = xfs_qm_dqflush_check(dqp);
1283 xfs_alert(mp, "corrupt dquot ID 0x%x in memory at %pS",
1286 xfs_dquot_mark_sick(dqp);
1287 error = -EFSCORRUPTED;
1291 /* Flush the incore dquot to the ondisk buffer. */
1292 dqblk = xfs_buf_offset(bp, dqp->q_bufoffset);
1293 xfs_dquot_to_disk(&dqblk->dd_diskdq, dqp);
1296 * Clear the dirty field and remember the flush lsn for later use.
1298 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1300 xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1301 &dqp->q_logitem.qli_item.li_lsn);
1304 * copy the lsn into the on-disk dquot now while we have the in memory
1305 * dquot here. This can't be done later in the write verifier as we
1306 * can't get access to the log item at that point in time.
1308 * We also calculate the CRC here so that the on-disk dquot in the
1309 * buffer always has a valid CRC. This ensures there is no possibility
1310 * of a dquot without an up-to-date CRC getting to disk.
1312 if (xfs_has_crc(mp)) {
1313 dqblk->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1314 xfs_update_cksum((char *)dqblk, sizeof(struct xfs_dqblk),
1319 * Attach the dquot to the buffer so that we can remove this dquot from
1320 * the AIL and release the flush lock once the dquot is synced to disk.
1322 bp->b_flags |= _XBF_DQUOTS;
1323 list_add_tail(&dqp->q_logitem.qli_item.li_bio_list, &bp->b_li_list);
1326 * If the buffer is pinned then push on the log so we won't
1327 * get stuck waiting in the write for too long.
1329 if (xfs_buf_ispinned(bp)) {
1330 trace_xfs_dqflush_force(dqp);
1331 xfs_log_force(mp, 0);
1334 trace_xfs_dqflush_done(dqp);
1339 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1340 xfs_trans_ail_delete(lip, 0);
1341 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1348 * Lock two xfs_dquot structures.
1350 * To avoid deadlocks we always lock the quota structure with
1351 * the lowerd id first.
1355 struct xfs_dquot *d1,
1356 struct xfs_dquot *d2)
1360 if (d1->q_id > d2->q_id) {
1361 mutex_lock(&d2->q_qlock);
1362 mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1364 mutex_lock(&d1->q_qlock);
1365 mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1368 mutex_lock(&d1->q_qlock);
1370 mutex_lock(&d2->q_qlock);
1377 xfs_dquot_cache = kmem_cache_create("xfs_dquot",
1378 sizeof(struct xfs_dquot),
1380 if (!xfs_dquot_cache)
1383 xfs_dqtrx_cache = kmem_cache_create("xfs_dqtrx",
1384 sizeof(struct xfs_dquot_acct),
1386 if (!xfs_dqtrx_cache)
1387 goto out_free_dquot_cache;
1391 out_free_dquot_cache:
1392 kmem_cache_destroy(xfs_dquot_cache);
1400 kmem_cache_destroy(xfs_dqtrx_cache);
1401 kmem_cache_destroy(xfs_dquot_cache);