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_extend(tp, quotip, XFS_DATA_FORK,
345 XFS_IEXT_ADD_NOSPLIT_CNT);
349 /* Create the block mapping. */
350 error = xfs_bmapi_write(tp, quotip, dqp->q_fileoffset,
351 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA, 0, &map,
356 ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
357 ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
358 (map.br_startblock != HOLESTARTBLOCK));
361 * Keep track of the blkno to save a lookup later
363 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
365 /* now we can just get the buffer (there's nothing to read yet) */
366 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, dqp->q_blkno,
367 mp->m_quotainfo->qi_dqchunklen, 0, &bp);
370 bp->b_ops = &xfs_dquot_buf_ops;
373 * Make a chunk of dquots out of this buffer and log
376 xfs_qm_init_dquot_blk(tp, dqp->q_id, qtype, bp);
377 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
380 * Hold the buffer and join it to the dfops so that we'll still own
381 * the buffer when we return to the caller. The buffer disposal on
382 * error must be paid attention to very carefully, as it has been
383 * broken since commit efa092f3d4c6 "[XFS] Fixes a bug in the quota
384 * code when allocating a new dquot record" in 2005, and the later
385 * conversion to xfs_defer_ops in commit 310a75a3c6c747 failed to keep
386 * the buffer locked across the _defer_finish call. We can now do
387 * this correctly with xfs_defer_bjoin.
389 * Above, we allocated a disk block for the dquot information and used
390 * get_buf to initialize the dquot. If the _defer_finish fails, the old
391 * transaction is gone but the new buffer is not joined or held to any
392 * transaction, so we must _buf_relse it.
394 * If everything succeeds, the caller of this function is returned a
395 * buffer that is locked and held to the transaction. The caller
396 * is responsible for unlocking any buffer passed back, either
397 * manually or by committing the transaction. On error, the buffer is
398 * released and not passed back.
400 * Keep the quota inode ILOCKed until after the transaction commit to
401 * maintain the atomicity of bmap/rmap updates.
403 xfs_trans_bhold(tp, bp);
404 error = xfs_trans_commit(tp);
405 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
415 xfs_trans_cancel(tp);
416 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
421 * Read in the in-core dquot's on-disk metadata and return the buffer.
422 * Returns ENOENT to signal a hole.
426 struct xfs_mount *mp,
427 struct xfs_dquot *dqp,
428 struct xfs_buf **bpp)
430 struct xfs_bmbt_irec map;
432 xfs_dqtype_t qtype = xfs_dquot_type(dqp);
433 struct xfs_inode *quotip = xfs_quota_inode(mp, qtype);
438 lock_mode = xfs_ilock_data_map_shared(quotip);
439 if (!xfs_this_quota_on(mp, qtype)) {
441 * Return if this type of quotas is turned off while we
442 * didn't have the quota inode lock.
444 xfs_iunlock(quotip, lock_mode);
449 * Find the block map; no allocations yet
451 error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
452 XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
453 xfs_iunlock(quotip, lock_mode);
458 ASSERT(map.br_blockcount >= 1);
459 ASSERT(map.br_startblock != DELAYSTARTBLOCK);
460 if (map.br_startblock == HOLESTARTBLOCK)
463 trace_xfs_dqtobp_read(dqp);
466 * store the blkno etc so that we don't have to do the
467 * mapping all the time
469 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
471 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
472 mp->m_quotainfo->qi_dqchunklen, 0, &bp,
474 if (xfs_metadata_is_sick(error))
475 xfs_dquot_mark_sick(dqp);
481 ASSERT(xfs_buf_islocked(bp));
482 xfs_buf_set_ref(bp, XFS_DQUOT_REF);
488 /* Allocate and initialize everything we need for an incore dquot. */
489 STATIC struct xfs_dquot *
491 struct xfs_mount *mp,
495 struct xfs_dquot *dqp;
497 dqp = kmem_cache_zalloc(xfs_dquot_cache, GFP_KERNEL | __GFP_NOFAIL);
502 INIT_LIST_HEAD(&dqp->q_lru);
503 mutex_init(&dqp->q_qlock);
504 init_waitqueue_head(&dqp->q_pinwait);
505 dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
507 * Offset of dquot in the (fixed sized) dquot chunk.
509 dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
510 sizeof(struct xfs_dqblk);
513 * Because we want to use a counting completion, complete
514 * the flush completion once to allow a single access to
515 * the flush completion without blocking.
517 init_completion(&dqp->q_flush);
518 complete(&dqp->q_flush);
521 * Make sure group quotas have a different lock class than user
525 case XFS_DQTYPE_USER:
526 /* uses the default lock class */
528 case XFS_DQTYPE_GROUP:
529 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
531 case XFS_DQTYPE_PROJ:
532 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
539 xfs_qm_dquot_logitem_init(dqp);
541 XFS_STATS_INC(mp, xs_qm_dquot);
545 /* Check the ondisk dquot's id and type match what the incore dquot expects. */
547 xfs_dquot_check_type(
548 struct xfs_dquot *dqp,
549 struct xfs_disk_dquot *ddqp)
554 ddqp_type = ddqp->d_type & XFS_DQTYPE_REC_MASK;
555 dqp_type = xfs_dquot_type(dqp);
557 if (be32_to_cpu(ddqp->d_id) != dqp->q_id)
561 * V5 filesystems always expect an exact type match. V4 filesystems
562 * expect an exact match for user dquots and for non-root group and
565 if (xfs_has_crc(dqp->q_mount) ||
566 dqp_type == XFS_DQTYPE_USER || dqp->q_id != 0)
567 return ddqp_type == dqp_type;
570 * V4 filesystems support either group or project quotas, but not both
571 * at the same time. The non-user quota file can be switched between
572 * group and project quota uses depending on the mount options, which
573 * means that we can encounter the other type when we try to load quota
574 * defaults. Quotacheck will soon reset the entire quota file
575 * (including the root dquot) anyway, but don't log scary corruption
578 return ddqp_type == XFS_DQTYPE_GROUP || ddqp_type == XFS_DQTYPE_PROJ;
581 /* Copy the in-core quota fields in from the on-disk buffer. */
584 struct xfs_dquot *dqp,
587 struct xfs_dqblk *dqb = xfs_buf_offset(bp, dqp->q_bufoffset);
588 struct xfs_disk_dquot *ddqp = &dqb->dd_diskdq;
591 * Ensure that we got the type and ID we were looking for.
592 * Everything else was checked by the dquot buffer verifier.
594 if (!xfs_dquot_check_type(dqp, ddqp)) {
595 xfs_alert_tag(bp->b_mount, XFS_PTAG_VERIFIER_ERROR,
596 "Metadata corruption detected at %pS, quota %u",
597 __this_address, dqp->q_id);
598 xfs_alert(bp->b_mount, "Unmount and run xfs_repair");
599 xfs_dquot_mark_sick(dqp);
600 return -EFSCORRUPTED;
603 /* copy everything from disk dquot to the incore dquot */
604 dqp->q_type = ddqp->d_type;
605 dqp->q_blk.hardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
606 dqp->q_blk.softlimit = be64_to_cpu(ddqp->d_blk_softlimit);
607 dqp->q_ino.hardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
608 dqp->q_ino.softlimit = be64_to_cpu(ddqp->d_ino_softlimit);
609 dqp->q_rtb.hardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
610 dqp->q_rtb.softlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
612 dqp->q_blk.count = be64_to_cpu(ddqp->d_bcount);
613 dqp->q_ino.count = be64_to_cpu(ddqp->d_icount);
614 dqp->q_rtb.count = be64_to_cpu(ddqp->d_rtbcount);
616 dqp->q_blk.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_btimer);
617 dqp->q_ino.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_itimer);
618 dqp->q_rtb.timer = xfs_dquot_from_disk_ts(ddqp, ddqp->d_rtbtimer);
621 * Reservation counters are defined as reservation plus current usage
622 * to avoid having to add every time.
624 dqp->q_blk.reserved = dqp->q_blk.count;
625 dqp->q_ino.reserved = dqp->q_ino.count;
626 dqp->q_rtb.reserved = dqp->q_rtb.count;
628 /* initialize the dquot speculative prealloc thresholds */
629 xfs_dquot_set_prealloc_limits(dqp);
633 /* Copy the in-core quota fields into the on-disk buffer. */
636 struct xfs_disk_dquot *ddqp,
637 struct xfs_dquot *dqp)
639 ddqp->d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
640 ddqp->d_version = XFS_DQUOT_VERSION;
641 ddqp->d_type = dqp->q_type;
642 ddqp->d_id = cpu_to_be32(dqp->q_id);
646 ddqp->d_blk_hardlimit = cpu_to_be64(dqp->q_blk.hardlimit);
647 ddqp->d_blk_softlimit = cpu_to_be64(dqp->q_blk.softlimit);
648 ddqp->d_ino_hardlimit = cpu_to_be64(dqp->q_ino.hardlimit);
649 ddqp->d_ino_softlimit = cpu_to_be64(dqp->q_ino.softlimit);
650 ddqp->d_rtb_hardlimit = cpu_to_be64(dqp->q_rtb.hardlimit);
651 ddqp->d_rtb_softlimit = cpu_to_be64(dqp->q_rtb.softlimit);
653 ddqp->d_bcount = cpu_to_be64(dqp->q_blk.count);
654 ddqp->d_icount = cpu_to_be64(dqp->q_ino.count);
655 ddqp->d_rtbcount = cpu_to_be64(dqp->q_rtb.count);
659 ddqp->d_rtbwarns = 0;
661 ddqp->d_btimer = xfs_dquot_to_disk_ts(dqp, dqp->q_blk.timer);
662 ddqp->d_itimer = xfs_dquot_to_disk_ts(dqp, dqp->q_ino.timer);
663 ddqp->d_rtbtimer = xfs_dquot_to_disk_ts(dqp, dqp->q_rtb.timer);
667 * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
668 * and release the buffer immediately. If @can_alloc is true, fill any
669 * holes in the on-disk metadata.
673 struct xfs_mount *mp,
677 struct xfs_dquot **dqpp)
679 struct xfs_dquot *dqp;
683 dqp = xfs_dquot_alloc(mp, id, type);
684 trace_xfs_dqread(dqp);
686 /* Try to read the buffer, allocating if necessary. */
687 error = xfs_dquot_disk_read(mp, dqp, &bp);
688 if (error == -ENOENT && can_alloc)
689 error = xfs_dquot_disk_alloc(dqp, &bp);
694 * At this point we should have a clean locked buffer. Copy the data
695 * to the incore dquot and release the buffer since the incore dquot
696 * has its own locking protocol so we needn't tie up the buffer any
699 ASSERT(xfs_buf_islocked(bp));
700 error = xfs_dquot_from_disk(dqp, bp);
709 trace_xfs_dqread_fail(dqp);
710 xfs_qm_dqdestroy(dqp);
716 * Advance to the next id in the current chunk, or if at the
717 * end of the chunk, skip ahead to first id in next allocated chunk
718 * using the SEEK_DATA interface.
722 struct xfs_mount *mp,
726 struct xfs_inode *quotip = xfs_quota_inode(mp, type);
727 xfs_dqid_t next_id = *id + 1; /* simple advance */
729 struct xfs_bmbt_irec got;
730 struct xfs_iext_cursor cur;
734 /* If we'd wrap past the max ID, stop */
738 /* If new ID is within the current chunk, advancing it sufficed */
739 if (next_id % mp->m_quotainfo->qi_dqperchunk) {
744 /* Nope, next_id is now past the current chunk, so find the next one */
745 start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
747 lock_flags = xfs_ilock_data_map_shared(quotip);
748 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
752 if (xfs_iext_lookup_extent(quotip, "ip->i_df, start, &cur, &got)) {
753 /* contiguous chunk, bump startoff for the id calculation */
754 if (got.br_startoff < start)
755 got.br_startoff = start;
756 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
761 xfs_iunlock(quotip, lock_flags);
767 * Look up the dquot in the in-core cache. If found, the dquot is returned
768 * locked and ready to go.
770 static struct xfs_dquot *
771 xfs_qm_dqget_cache_lookup(
772 struct xfs_mount *mp,
773 struct xfs_quotainfo *qi,
774 struct radix_tree_root *tree,
777 struct xfs_dquot *dqp;
780 mutex_lock(&qi->qi_tree_lock);
781 dqp = radix_tree_lookup(tree, id);
783 mutex_unlock(&qi->qi_tree_lock);
784 XFS_STATS_INC(mp, xs_qm_dqcachemisses);
789 if (dqp->q_flags & XFS_DQFLAG_FREEING) {
791 mutex_unlock(&qi->qi_tree_lock);
792 trace_xfs_dqget_freeing(dqp);
798 mutex_unlock(&qi->qi_tree_lock);
800 trace_xfs_dqget_hit(dqp);
801 XFS_STATS_INC(mp, xs_qm_dqcachehits);
806 * Try to insert a new dquot into the in-core cache. If an error occurs the
807 * caller should throw away the dquot and start over. Otherwise, the dquot
808 * is returned locked (and held by the cache) as if there had been a cache
811 * The insert needs to be done under memalloc_nofs context because the radix
812 * tree can do memory allocation during insert. The qi->qi_tree_lock is taken in
813 * memory reclaim when freeing unused dquots, so we cannot have the radix tree
814 * node allocation recursing into filesystem reclaim whilst we hold the
818 xfs_qm_dqget_cache_insert(
819 struct xfs_mount *mp,
820 struct xfs_quotainfo *qi,
821 struct radix_tree_root *tree,
823 struct xfs_dquot *dqp)
825 unsigned int nofs_flags;
828 nofs_flags = memalloc_nofs_save();
829 mutex_lock(&qi->qi_tree_lock);
830 error = radix_tree_insert(tree, id, dqp);
831 if (unlikely(error)) {
832 /* Duplicate found! Caller must try again. */
833 trace_xfs_dqget_dup(dqp);
837 /* Return a locked dquot to the caller, with a reference taken. */
843 mutex_unlock(&qi->qi_tree_lock);
844 memalloc_nofs_restore(nofs_flags);
848 /* Check our input parameters. */
851 struct xfs_mount *mp,
855 case XFS_DQTYPE_USER:
856 if (!XFS_IS_UQUOTA_ON(mp))
859 case XFS_DQTYPE_GROUP:
860 if (!XFS_IS_GQUOTA_ON(mp))
863 case XFS_DQTYPE_PROJ:
864 if (!XFS_IS_PQUOTA_ON(mp))
874 * Given the file system, id, and type (UDQUOT/GDQUOT/PDQUOT), return a
875 * locked dquot, doing an allocation (if requested) as needed.
879 struct xfs_mount *mp,
883 struct xfs_dquot **O_dqpp)
885 struct xfs_quotainfo *qi = mp->m_quotainfo;
886 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
887 struct xfs_dquot *dqp;
890 error = xfs_qm_dqget_checks(mp, type);
895 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
901 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
905 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
908 * Duplicate found. Just throw away the new dquot and start
911 xfs_qm_dqdestroy(dqp);
912 XFS_STATS_INC(mp, xs_qm_dquot_dups);
916 trace_xfs_dqget_miss(dqp);
922 * Given a dquot id and type, read and initialize a dquot from the on-disk
923 * metadata. This function is only for use during quota initialization so
924 * it ignores the dquot cache assuming that the dquot shrinker isn't set up.
925 * The caller is responsible for _qm_dqdestroy'ing the returned dquot.
928 xfs_qm_dqget_uncached(
929 struct xfs_mount *mp,
932 struct xfs_dquot **dqpp)
936 error = xfs_qm_dqget_checks(mp, type);
940 return xfs_qm_dqread(mp, id, type, 0, dqpp);
943 /* Return the quota id for a given inode and type. */
945 xfs_qm_id_for_quotatype(
946 struct xfs_inode *ip,
950 case XFS_DQTYPE_USER:
951 return i_uid_read(VFS_I(ip));
952 case XFS_DQTYPE_GROUP:
953 return i_gid_read(VFS_I(ip));
954 case XFS_DQTYPE_PROJ:
962 * Return the dquot for a given inode and type. If @can_alloc is true, then
963 * allocate blocks if needed. The inode's ILOCK must be held and it must not
964 * have already had an inode attached.
968 struct xfs_inode *ip,
971 struct xfs_dquot **O_dqpp)
973 struct xfs_mount *mp = ip->i_mount;
974 struct xfs_quotainfo *qi = mp->m_quotainfo;
975 struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
976 struct xfs_dquot *dqp;
980 error = xfs_qm_dqget_checks(mp, type);
984 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
985 ASSERT(xfs_inode_dquot(ip, type) == NULL);
987 id = xfs_qm_id_for_quotatype(ip, type);
990 dqp = xfs_qm_dqget_cache_lookup(mp, qi, tree, id);
997 * Dquot cache miss. We don't want to keep the inode lock across
998 * a (potential) disk read. Also we don't want to deal with the lock
999 * ordering between quotainode and this inode. OTOH, dropping the inode
1000 * lock here means dealing with a chown that can happen before
1001 * we re-acquire the lock.
1003 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1004 error = xfs_qm_dqread(mp, id, type, can_alloc, &dqp);
1005 xfs_ilock(ip, XFS_ILOCK_EXCL);
1010 * A dquot could be attached to this inode by now, since we had
1011 * dropped the ilock.
1013 if (xfs_this_quota_on(mp, type)) {
1014 struct xfs_dquot *dqp1;
1016 dqp1 = xfs_inode_dquot(ip, type);
1018 xfs_qm_dqdestroy(dqp);
1024 /* inode stays locked on return */
1025 xfs_qm_dqdestroy(dqp);
1029 error = xfs_qm_dqget_cache_insert(mp, qi, tree, id, dqp);
1032 * Duplicate found. Just throw away the new dquot and start
1035 xfs_qm_dqdestroy(dqp);
1036 XFS_STATS_INC(mp, xs_qm_dquot_dups);
1041 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1042 trace_xfs_dqget_miss(dqp);
1048 * Starting at @id and progressing upwards, look for an initialized incore
1049 * dquot, lock it, and return it.
1053 struct xfs_mount *mp,
1056 struct xfs_dquot **dqpp)
1058 struct xfs_dquot *dqp;
1062 for (; !error; error = xfs_dq_get_next_id(mp, type, &id)) {
1063 error = xfs_qm_dqget(mp, id, type, false, &dqp);
1064 if (error == -ENOENT)
1066 else if (error != 0)
1069 if (!XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
1081 * Release a reference to the dquot (decrement ref-count) and unlock it.
1083 * If there is a group quota attached to this dquot, carefully release that
1084 * too without tripping over deadlocks'n'stuff.
1088 struct xfs_dquot *dqp)
1090 ASSERT(dqp->q_nrefs > 0);
1091 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1093 trace_xfs_dqput(dqp);
1095 if (--dqp->q_nrefs == 0) {
1096 struct xfs_quotainfo *qi = dqp->q_mount->m_quotainfo;
1097 trace_xfs_dqput_free(dqp);
1099 if (list_lru_add_obj(&qi->qi_lru, &dqp->q_lru))
1100 XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
1106 * Release a dquot. Flush it if dirty, then dqput() it.
1107 * dquot must not be locked.
1111 struct xfs_dquot *dqp)
1116 trace_xfs_dqrele(dqp);
1120 * We don't care to flush it if the dquot is dirty here.
1121 * That will create stutters that we want to avoid.
1122 * Instead we do a delayed write when we try to reclaim
1123 * a dirty dquot. Also xfs_sync will take part of the burden...
1129 * This is the dquot flushing I/O completion routine. It is called
1130 * from interrupt level when the buffer containing the dquot is
1131 * flushed to disk. It is responsible for removing the dquot logitem
1132 * from the AIL if it has not been re-logged, and unlocking the dquot's
1133 * flush lock. This behavior is very similar to that of inodes..
1136 xfs_qm_dqflush_done(
1137 struct xfs_log_item *lip)
1139 struct xfs_dq_logitem *qip = (struct xfs_dq_logitem *)lip;
1140 struct xfs_dquot *dqp = qip->qli_dquot;
1141 struct xfs_ail *ailp = lip->li_ailp;
1145 * We only want to pull the item from the AIL if its
1146 * location in the log has not changed since we started the flush.
1147 * Thus, we only bother if the dquot's lsn has
1148 * not changed. First we check the lsn outside the lock
1149 * since it's cheaper, and then we recheck while
1150 * holding the lock before removing the dquot from the AIL.
1152 if (test_bit(XFS_LI_IN_AIL, &lip->li_flags) &&
1153 ((lip->li_lsn == qip->qli_flush_lsn) ||
1154 test_bit(XFS_LI_FAILED, &lip->li_flags))) {
1156 spin_lock(&ailp->ail_lock);
1157 xfs_clear_li_failed(lip);
1158 if (lip->li_lsn == qip->qli_flush_lsn) {
1159 /* xfs_ail_update_finish() drops the AIL lock */
1160 tail_lsn = xfs_ail_delete_one(ailp, lip);
1161 xfs_ail_update_finish(ailp, tail_lsn);
1163 spin_unlock(&ailp->ail_lock);
1168 * Release the dq's flush lock since we're done with it.
1174 xfs_buf_dquot_iodone(
1177 struct xfs_log_item *lip, *n;
1179 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
1180 list_del_init(&lip->li_bio_list);
1181 xfs_qm_dqflush_done(lip);
1186 xfs_buf_dquot_io_fail(
1189 struct xfs_log_item *lip;
1191 spin_lock(&bp->b_mount->m_ail->ail_lock);
1192 list_for_each_entry(lip, &bp->b_li_list, li_bio_list)
1193 xfs_set_li_failed(lip, bp);
1194 spin_unlock(&bp->b_mount->m_ail->ail_lock);
1197 /* Check incore dquot for errors before we flush. */
1198 static xfs_failaddr_t
1199 xfs_qm_dqflush_check(
1200 struct xfs_dquot *dqp)
1202 xfs_dqtype_t type = xfs_dquot_type(dqp);
1204 if (type != XFS_DQTYPE_USER &&
1205 type != XFS_DQTYPE_GROUP &&
1206 type != XFS_DQTYPE_PROJ)
1207 return __this_address;
1212 if (dqp->q_blk.softlimit && dqp->q_blk.count > dqp->q_blk.softlimit &&
1214 return __this_address;
1216 if (dqp->q_ino.softlimit && dqp->q_ino.count > dqp->q_ino.softlimit &&
1218 return __this_address;
1220 if (dqp->q_rtb.softlimit && dqp->q_rtb.count > dqp->q_rtb.softlimit &&
1222 return __this_address;
1224 /* bigtime flag should never be set on root dquots */
1225 if (dqp->q_type & XFS_DQTYPE_BIGTIME) {
1226 if (!xfs_has_bigtime(dqp->q_mount))
1227 return __this_address;
1229 return __this_address;
1236 * Write a modified dquot to disk.
1237 * The dquot must be locked and the flush lock too taken by caller.
1238 * The flush lock will not be unlocked until the dquot reaches the disk,
1239 * but the dquot is free to be unlocked and modified by the caller
1240 * in the interim. Dquot is still locked on return. This behavior is
1241 * identical to that of inodes.
1245 struct xfs_dquot *dqp,
1246 struct xfs_buf **bpp)
1248 struct xfs_mount *mp = dqp->q_mount;
1249 struct xfs_log_item *lip = &dqp->q_logitem.qli_item;
1251 struct xfs_dqblk *dqblk;
1255 ASSERT(XFS_DQ_IS_LOCKED(dqp));
1256 ASSERT(!completion_done(&dqp->q_flush));
1258 trace_xfs_dqflush(dqp);
1262 xfs_qm_dqunpin_wait(dqp);
1265 * Get the buffer containing the on-disk dquot
1267 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1268 mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK,
1269 &bp, &xfs_dquot_buf_ops);
1270 if (error == -EAGAIN)
1272 if (xfs_metadata_is_sick(error))
1273 xfs_dquot_mark_sick(dqp);
1277 fa = xfs_qm_dqflush_check(dqp);
1279 xfs_alert(mp, "corrupt dquot ID 0x%x in memory at %pS",
1282 xfs_dquot_mark_sick(dqp);
1283 error = -EFSCORRUPTED;
1287 /* Flush the incore dquot to the ondisk buffer. */
1288 dqblk = xfs_buf_offset(bp, dqp->q_bufoffset);
1289 xfs_dquot_to_disk(&dqblk->dd_diskdq, dqp);
1292 * Clear the dirty field and remember the flush lsn for later use.
1294 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1296 xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1297 &dqp->q_logitem.qli_item.li_lsn);
1300 * copy the lsn into the on-disk dquot now while we have the in memory
1301 * dquot here. This can't be done later in the write verifier as we
1302 * can't get access to the log item at that point in time.
1304 * We also calculate the CRC here so that the on-disk dquot in the
1305 * buffer always has a valid CRC. This ensures there is no possibility
1306 * of a dquot without an up-to-date CRC getting to disk.
1308 if (xfs_has_crc(mp)) {
1309 dqblk->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1310 xfs_update_cksum((char *)dqblk, sizeof(struct xfs_dqblk),
1315 * Attach the dquot to the buffer so that we can remove this dquot from
1316 * the AIL and release the flush lock once the dquot is synced to disk.
1318 bp->b_flags |= _XBF_DQUOTS;
1319 list_add_tail(&dqp->q_logitem.qli_item.li_bio_list, &bp->b_li_list);
1322 * If the buffer is pinned then push on the log so we won't
1323 * get stuck waiting in the write for too long.
1325 if (xfs_buf_ispinned(bp)) {
1326 trace_xfs_dqflush_force(dqp);
1327 xfs_log_force(mp, 0);
1330 trace_xfs_dqflush_done(dqp);
1335 dqp->q_flags &= ~XFS_DQFLAG_DIRTY;
1336 xfs_trans_ail_delete(lip, 0);
1337 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1344 * Lock two xfs_dquot structures.
1346 * To avoid deadlocks we always lock the quota structure with
1347 * the lowerd id first.
1351 struct xfs_dquot *d1,
1352 struct xfs_dquot *d2)
1356 if (d1->q_id > d2->q_id) {
1357 mutex_lock(&d2->q_qlock);
1358 mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1360 mutex_lock(&d1->q_qlock);
1361 mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1364 mutex_lock(&d1->q_qlock);
1366 mutex_lock(&d2->q_qlock);
1375 const struct xfs_dqtrx *qa = a;
1376 const struct xfs_dqtrx *qb = b;
1378 if (qa->qt_dquot->q_id > qb->qt_dquot->q_id)
1380 if (qa->qt_dquot->q_id < qb->qt_dquot->q_id)
1387 struct xfs_dqtrx *q)
1391 BUILD_BUG_ON(XFS_QM_TRANS_MAXDQS > MAX_LOCKDEP_SUBCLASSES);
1393 /* Sort in order of dquot id, do not allow duplicates */
1394 for (i = 0; i < XFS_QM_TRANS_MAXDQS && q[i].qt_dquot != NULL; i++) {
1397 for (j = 0; j < i; j++)
1398 ASSERT(q[i].qt_dquot != q[j].qt_dquot);
1403 sort(q, i, sizeof(struct xfs_dqtrx), xfs_dqtrx_cmp, NULL);
1405 mutex_lock(&q[0].qt_dquot->q_qlock);
1406 for (i = 1; i < XFS_QM_TRANS_MAXDQS && q[i].qt_dquot != NULL; i++)
1407 mutex_lock_nested(&q[i].qt_dquot->q_qlock,
1408 XFS_QLOCK_NESTED + i - 1);
1414 xfs_dquot_cache = kmem_cache_create("xfs_dquot",
1415 sizeof(struct xfs_dquot),
1417 if (!xfs_dquot_cache)
1420 xfs_dqtrx_cache = kmem_cache_create("xfs_dqtrx",
1421 sizeof(struct xfs_dquot_acct),
1423 if (!xfs_dqtrx_cache)
1424 goto out_free_dquot_cache;
1428 out_free_dquot_cache:
1429 kmem_cache_destroy(xfs_dquot_cache);
1437 kmem_cache_destroy(xfs_dqtrx_cache);
1438 kmem_cache_destroy(xfs_dquot_cache);