2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
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.
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.
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
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_da_format.h"
28 #include "xfs_da_btree.h"
29 #include "xfs_inode.h"
31 #include "xfs_ialloc.h"
32 #include "xfs_alloc.h"
33 #include "xfs_rtalloc.h"
35 #include "xfs_trans.h"
36 #include "xfs_trans_priv.h"
38 #include "xfs_error.h"
39 #include "xfs_quota.h"
40 #include "xfs_fsops.h"
41 #include "xfs_trace.h"
42 #include "xfs_icache.h"
43 #include "xfs_sysfs.h"
47 STATIC void xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
49 STATIC void xfs_icsb_balance_counter_locked(xfs_mount_t *, xfs_sb_field_t,
51 STATIC void xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
54 #define xfs_icsb_balance_counter(mp, a, b) do { } while (0)
55 #define xfs_icsb_balance_counter_locked(mp, a, b) do { } while (0)
58 static DEFINE_MUTEX(xfs_uuid_table_mutex);
59 static int xfs_uuid_table_size;
60 static uuid_t *xfs_uuid_table;
63 * See if the UUID is unique among mounted XFS filesystems.
64 * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
70 uuid_t *uuid = &mp->m_sb.sb_uuid;
73 if (mp->m_flags & XFS_MOUNT_NOUUID)
76 if (uuid_is_nil(uuid)) {
77 xfs_warn(mp, "Filesystem has nil UUID - can't mount");
81 mutex_lock(&xfs_uuid_table_mutex);
82 for (i = 0, hole = -1; i < xfs_uuid_table_size; i++) {
83 if (uuid_is_nil(&xfs_uuid_table[i])) {
87 if (uuid_equal(uuid, &xfs_uuid_table[i]))
92 xfs_uuid_table = kmem_realloc(xfs_uuid_table,
93 (xfs_uuid_table_size + 1) * sizeof(*xfs_uuid_table),
94 xfs_uuid_table_size * sizeof(*xfs_uuid_table),
96 hole = xfs_uuid_table_size++;
98 xfs_uuid_table[hole] = *uuid;
99 mutex_unlock(&xfs_uuid_table_mutex);
104 mutex_unlock(&xfs_uuid_table_mutex);
105 xfs_warn(mp, "Filesystem has duplicate UUID %pU - can't mount", uuid);
111 struct xfs_mount *mp)
113 uuid_t *uuid = &mp->m_sb.sb_uuid;
116 if (mp->m_flags & XFS_MOUNT_NOUUID)
119 mutex_lock(&xfs_uuid_table_mutex);
120 for (i = 0; i < xfs_uuid_table_size; i++) {
121 if (uuid_is_nil(&xfs_uuid_table[i]))
123 if (!uuid_equal(uuid, &xfs_uuid_table[i]))
125 memset(&xfs_uuid_table[i], 0, sizeof(uuid_t));
128 ASSERT(i < xfs_uuid_table_size);
129 mutex_unlock(&xfs_uuid_table_mutex);
135 struct rcu_head *head)
137 struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
139 ASSERT(atomic_read(&pag->pag_ref) == 0);
144 * Free up the per-ag resources associated with the mount structure.
151 struct xfs_perag *pag;
153 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
154 spin_lock(&mp->m_perag_lock);
155 pag = radix_tree_delete(&mp->m_perag_tree, agno);
156 spin_unlock(&mp->m_perag_lock);
158 ASSERT(atomic_read(&pag->pag_ref) == 0);
159 call_rcu(&pag->rcu_head, __xfs_free_perag);
164 * Check size of device based on the (data/realtime) block count.
165 * Note: this check is used by the growfs code as well as mount.
168 xfs_sb_validate_fsb_count(
172 ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
173 ASSERT(sbp->sb_blocklog >= BBSHIFT);
175 /* Limited by ULONG_MAX of page cache index */
176 if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
182 xfs_initialize_perag(
184 xfs_agnumber_t agcount,
185 xfs_agnumber_t *maxagi)
187 xfs_agnumber_t index;
188 xfs_agnumber_t first_initialised = 0;
192 xfs_sb_t *sbp = &mp->m_sb;
196 * Walk the current per-ag tree so we don't try to initialise AGs
197 * that already exist (growfs case). Allocate and insert all the
198 * AGs we don't find ready for initialisation.
200 for (index = 0; index < agcount; index++) {
201 pag = xfs_perag_get(mp, index);
206 if (!first_initialised)
207 first_initialised = index;
209 pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
212 pag->pag_agno = index;
214 spin_lock_init(&pag->pag_ici_lock);
215 mutex_init(&pag->pag_ici_reclaim_lock);
216 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
217 spin_lock_init(&pag->pag_buf_lock);
218 pag->pag_buf_tree = RB_ROOT;
220 if (radix_tree_preload(GFP_NOFS))
223 spin_lock(&mp->m_perag_lock);
224 if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
226 spin_unlock(&mp->m_perag_lock);
227 radix_tree_preload_end();
231 spin_unlock(&mp->m_perag_lock);
232 radix_tree_preload_end();
236 * If we mount with the inode64 option, or no inode overflows
237 * the legacy 32-bit address space clear the inode32 option.
239 agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
240 ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
242 if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > XFS_MAXINUMBER_32)
243 mp->m_flags |= XFS_MOUNT_32BITINODES;
245 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
247 if (mp->m_flags & XFS_MOUNT_32BITINODES)
248 index = xfs_set_inode32(mp, agcount);
250 index = xfs_set_inode64(mp, agcount);
258 for (; index > first_initialised; index--) {
259 pag = radix_tree_delete(&mp->m_perag_tree, index);
268 * Does the initial read of the superblock.
272 struct xfs_mount *mp,
275 unsigned int sector_size;
277 struct xfs_sb *sbp = &mp->m_sb;
279 int loud = !(flags & XFS_MFSI_QUIET);
280 const struct xfs_buf_ops *buf_ops;
282 ASSERT(mp->m_sb_bp == NULL);
283 ASSERT(mp->m_ddev_targp != NULL);
286 * For the initial read, we must guess at the sector
287 * size based on the block device. It's enough to
288 * get the sb_sectsize out of the superblock and
289 * then reread with the proper length.
290 * We don't verify it yet, because it may not be complete.
292 sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
296 * Allocate a (locked) buffer to hold the superblock.
297 * This will be kept around at all times to optimize
298 * access to the superblock.
301 error = xfs_buf_read_uncached(mp->m_ddev_targp, XFS_SB_DADDR,
302 BTOBB(sector_size), 0, &bp, buf_ops);
305 xfs_warn(mp, "SB validate failed with error %d.", error);
306 /* bad CRC means corrupted metadata */
307 if (error == -EFSBADCRC)
308 error = -EFSCORRUPTED;
313 * Initialize the mount structure from the superblock.
315 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
318 * If we haven't validated the superblock, do so now before we try
319 * to check the sector size and reread the superblock appropriately.
321 if (sbp->sb_magicnum != XFS_SB_MAGIC) {
323 xfs_warn(mp, "Invalid superblock magic number");
329 * We must be able to do sector-sized and sector-aligned IO.
331 if (sector_size > sbp->sb_sectsize) {
333 xfs_warn(mp, "device supports %u byte sectors (not %u)",
334 sector_size, sbp->sb_sectsize);
339 if (buf_ops == NULL) {
341 * Re-read the superblock so the buffer is correctly sized,
342 * and properly verified.
345 sector_size = sbp->sb_sectsize;
346 buf_ops = loud ? &xfs_sb_buf_ops : &xfs_sb_quiet_buf_ops;
350 /* Initialize per-cpu counters */
351 xfs_icsb_reinit_counters(mp);
353 /* no need to be quiet anymore, so reset the buf ops */
354 bp->b_ops = &xfs_sb_buf_ops;
366 * Update alignment values based on mount options and sb values
369 xfs_update_alignment(xfs_mount_t *mp)
371 xfs_sb_t *sbp = &(mp->m_sb);
375 * If stripe unit and stripe width are not multiples
376 * of the fs blocksize turn off alignment.
378 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
379 (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
381 "alignment check failed: sunit/swidth vs. blocksize(%d)",
386 * Convert the stripe unit and width to FSBs.
388 mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
389 if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
391 "alignment check failed: sunit/swidth vs. agsize(%d)",
394 } else if (mp->m_dalign) {
395 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
398 "alignment check failed: sunit(%d) less than bsize(%d)",
399 mp->m_dalign, sbp->sb_blocksize);
405 * Update superblock with new values
408 if (xfs_sb_version_hasdalign(sbp)) {
409 if (sbp->sb_unit != mp->m_dalign) {
410 sbp->sb_unit = mp->m_dalign;
411 mp->m_update_flags |= XFS_SB_UNIT;
413 if (sbp->sb_width != mp->m_swidth) {
414 sbp->sb_width = mp->m_swidth;
415 mp->m_update_flags |= XFS_SB_WIDTH;
419 "cannot change alignment: superblock does not support data alignment");
422 } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
423 xfs_sb_version_hasdalign(&mp->m_sb)) {
424 mp->m_dalign = sbp->sb_unit;
425 mp->m_swidth = sbp->sb_width;
432 * Set the maximum inode count for this filesystem
435 xfs_set_maxicount(xfs_mount_t *mp)
437 xfs_sb_t *sbp = &(mp->m_sb);
440 if (sbp->sb_imax_pct) {
442 * Make sure the maximum inode count is a multiple
443 * of the units we allocate inodes in.
445 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
447 do_div(icount, mp->m_ialloc_blks);
448 mp->m_maxicount = (icount * mp->m_ialloc_blks) <<
456 * Set the default minimum read and write sizes unless
457 * already specified in a mount option.
458 * We use smaller I/O sizes when the file system
459 * is being used for NFS service (wsync mount option).
462 xfs_set_rw_sizes(xfs_mount_t *mp)
464 xfs_sb_t *sbp = &(mp->m_sb);
465 int readio_log, writeio_log;
467 if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
468 if (mp->m_flags & XFS_MOUNT_WSYNC) {
469 readio_log = XFS_WSYNC_READIO_LOG;
470 writeio_log = XFS_WSYNC_WRITEIO_LOG;
472 readio_log = XFS_READIO_LOG_LARGE;
473 writeio_log = XFS_WRITEIO_LOG_LARGE;
476 readio_log = mp->m_readio_log;
477 writeio_log = mp->m_writeio_log;
480 if (sbp->sb_blocklog > readio_log) {
481 mp->m_readio_log = sbp->sb_blocklog;
483 mp->m_readio_log = readio_log;
485 mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
486 if (sbp->sb_blocklog > writeio_log) {
487 mp->m_writeio_log = sbp->sb_blocklog;
489 mp->m_writeio_log = writeio_log;
491 mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
495 * precalculate the low space thresholds for dynamic speculative preallocation.
498 xfs_set_low_space_thresholds(
499 struct xfs_mount *mp)
503 for (i = 0; i < XFS_LOWSP_MAX; i++) {
504 __uint64_t space = mp->m_sb.sb_dblocks;
507 mp->m_low_space[i] = space * (i + 1);
513 * Set whether we're using inode alignment.
516 xfs_set_inoalignment(xfs_mount_t *mp)
518 if (xfs_sb_version_hasalign(&mp->m_sb) &&
519 mp->m_sb.sb_inoalignmt >=
520 XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
521 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
523 mp->m_inoalign_mask = 0;
525 * If we are using stripe alignment, check whether
526 * the stripe unit is a multiple of the inode alignment
528 if (mp->m_dalign && mp->m_inoalign_mask &&
529 !(mp->m_dalign & mp->m_inoalign_mask))
530 mp->m_sinoalign = mp->m_dalign;
536 * Check that the data (and log if separate) is an ok size.
540 struct xfs_mount *mp)
546 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
547 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
548 xfs_warn(mp, "filesystem size mismatch detected");
551 error = xfs_buf_read_uncached(mp->m_ddev_targp,
552 d - XFS_FSS_TO_BB(mp, 1),
553 XFS_FSS_TO_BB(mp, 1), 0, &bp, NULL);
555 xfs_warn(mp, "last sector read failed");
560 if (mp->m_logdev_targp == mp->m_ddev_targp)
563 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
564 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
565 xfs_warn(mp, "log size mismatch detected");
568 error = xfs_buf_read_uncached(mp->m_logdev_targp,
569 d - XFS_FSB_TO_BB(mp, 1),
570 XFS_FSB_TO_BB(mp, 1), 0, &bp, NULL);
572 xfs_warn(mp, "log device read failed");
580 * Clear the quotaflags in memory and in the superblock.
583 xfs_mount_reset_sbqflags(
584 struct xfs_mount *mp)
587 struct xfs_trans *tp;
592 * It is OK to look at sb_qflags here in mount path,
595 if (mp->m_sb.sb_qflags == 0)
597 spin_lock(&mp->m_sb_lock);
598 mp->m_sb.sb_qflags = 0;
599 spin_unlock(&mp->m_sb_lock);
602 * If the fs is readonly, let the incore superblock run
603 * with quotas off but don't flush the update out to disk
605 if (mp->m_flags & XFS_MOUNT_RDONLY)
608 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
609 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_qm_sbchange, 0, 0);
611 xfs_trans_cancel(tp, 0);
612 xfs_alert(mp, "%s: Superblock update failed!", __func__);
616 xfs_mod_sb(tp, XFS_SB_QFLAGS);
617 return xfs_trans_commit(tp, 0);
621 xfs_default_resblks(xfs_mount_t *mp)
626 * We default to 5% or 8192 fsbs of space reserved, whichever is
627 * smaller. This is intended to cover concurrent allocation
628 * transactions when we initially hit enospc. These each require a 4
629 * block reservation. Hence by default we cover roughly 2000 concurrent
630 * allocation reservations.
632 resblks = mp->m_sb.sb_dblocks;
634 resblks = min_t(__uint64_t, resblks, 8192);
639 * This function does the following on an initial mount of a file system:
640 * - reads the superblock from disk and init the mount struct
641 * - if we're a 32-bit kernel, do a size check on the superblock
642 * so we don't mount terabyte filesystems
643 * - init mount struct realtime fields
644 * - allocate inode hash table for fs
645 * - init directory manager
646 * - perform recovery and init the log manager
652 xfs_sb_t *sbp = &(mp->m_sb);
659 xfs_sb_mount_common(mp, sbp);
662 * Check for a mismatched features2 values. Older kernels
663 * read & wrote into the wrong sb offset for sb_features2
664 * on some platforms due to xfs_sb_t not being 64bit size aligned
665 * when sb_features2 was added, which made older superblock
666 * reading/writing routines swap it as a 64-bit value.
668 * For backwards compatibility, we make both slots equal.
670 * If we detect a mismatched field, we OR the set bits into the
671 * existing features2 field in case it has already been modified; we
672 * don't want to lose any features. We then update the bad location
673 * with the ORed value so that older kernels will see any features2
674 * flags, and mark the two fields as needing updates once the
675 * transaction subsystem is online.
677 if (xfs_sb_has_mismatched_features2(sbp)) {
678 xfs_warn(mp, "correcting sb_features alignment problem");
679 sbp->sb_features2 |= sbp->sb_bad_features2;
680 sbp->sb_bad_features2 = sbp->sb_features2;
681 mp->m_update_flags |= XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2;
684 * Re-check for ATTR2 in case it was found in bad_features2
687 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
688 !(mp->m_flags & XFS_MOUNT_NOATTR2))
689 mp->m_flags |= XFS_MOUNT_ATTR2;
692 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
693 (mp->m_flags & XFS_MOUNT_NOATTR2)) {
694 xfs_sb_version_removeattr2(&mp->m_sb);
695 mp->m_update_flags |= XFS_SB_FEATURES2;
697 /* update sb_versionnum for the clearing of the morebits */
698 if (!sbp->sb_features2)
699 mp->m_update_flags |= XFS_SB_VERSIONNUM;
702 /* always use v2 inodes by default now */
703 if (!(mp->m_sb.sb_versionnum & XFS_SB_VERSION_NLINKBIT)) {
704 mp->m_sb.sb_versionnum |= XFS_SB_VERSION_NLINKBIT;
705 mp->m_update_flags |= XFS_SB_VERSIONNUM;
709 * Check if sb_agblocks is aligned at stripe boundary
710 * If sb_agblocks is NOT aligned turn off m_dalign since
711 * allocator alignment is within an ag, therefore ag has
712 * to be aligned at stripe boundary.
714 error = xfs_update_alignment(mp);
718 xfs_alloc_compute_maxlevels(mp);
719 xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
720 xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
721 xfs_ialloc_compute_maxlevels(mp);
723 xfs_set_maxicount(mp);
725 error = xfs_sysfs_init(&mp->m_kobj, &xfs_mp_ktype, NULL, mp->m_fsname);
729 error = xfs_uuid_mount(mp);
731 goto out_remove_sysfs;
734 * Set the minimum read and write sizes
736 xfs_set_rw_sizes(mp);
738 /* set the low space thresholds for dynamic preallocation */
739 xfs_set_low_space_thresholds(mp);
742 * Set the inode cluster size.
743 * This may still be overridden by the file system
744 * block size if it is larger than the chosen cluster size.
746 * For v5 filesystems, scale the cluster size with the inode size to
747 * keep a constant ratio of inode per cluster buffer, but only if mkfs
748 * has set the inode alignment value appropriately for larger cluster
751 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
752 if (xfs_sb_version_hascrc(&mp->m_sb)) {
753 int new_size = mp->m_inode_cluster_size;
755 new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
756 if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
757 mp->m_inode_cluster_size = new_size;
761 * Set inode alignment fields
763 xfs_set_inoalignment(mp);
766 * Check that the data (and log if separate) is an ok size.
768 error = xfs_check_sizes(mp);
770 goto out_remove_uuid;
773 * Initialize realtime fields in the mount structure
775 error = xfs_rtmount_init(mp);
777 xfs_warn(mp, "RT mount failed");
778 goto out_remove_uuid;
782 * Copies the low order bits of the timestamp and the randomly
783 * set "sequence" number out of a UUID.
785 uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
787 mp->m_dmevmask = 0; /* not persistent; set after each mount */
789 error = xfs_da_mount(mp);
791 xfs_warn(mp, "Failed dir/attr init: %d", error);
792 goto out_remove_uuid;
796 * Initialize the precomputed transaction reservations values.
801 * Allocate and initialize the per-ag data.
803 spin_lock_init(&mp->m_perag_lock);
804 INIT_RADIX_TREE(&mp->m_perag_tree, GFP_ATOMIC);
805 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
807 xfs_warn(mp, "Failed per-ag init: %d", error);
811 if (!sbp->sb_logblocks) {
812 xfs_warn(mp, "no log defined");
813 XFS_ERROR_REPORT("xfs_mountfs", XFS_ERRLEVEL_LOW, mp);
814 error = -EFSCORRUPTED;
819 * log's mount-time initialization. Perform 1st part recovery if needed
821 error = xfs_log_mount(mp, mp->m_logdev_targp,
822 XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
823 XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
825 xfs_warn(mp, "log mount failed");
830 * Now the log is mounted, we know if it was an unclean shutdown or
831 * not. If it was, with the first phase of recovery has completed, we
832 * have consistent AG blocks on disk. We have not recovered EFIs yet,
833 * but they are recovered transactionally in the second recovery phase
836 * Hence we can safely re-initialise incore superblock counters from
837 * the per-ag data. These may not be correct if the filesystem was not
838 * cleanly unmounted, so we need to wait for recovery to finish before
841 * If the filesystem was cleanly unmounted, then we can trust the
842 * values in the superblock to be correct and we don't need to do
845 * If we are currently making the filesystem, the initialisation will
846 * fail as the perag data is in an undefined state.
848 if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
849 !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
850 !mp->m_sb.sb_inprogress) {
851 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
853 goto out_log_dealloc;
857 * Get and sanity-check the root inode.
858 * Save the pointer to it in the mount structure.
860 error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip);
862 xfs_warn(mp, "failed to read root inode");
863 goto out_log_dealloc;
868 if (unlikely(!S_ISDIR(rip->i_d.di_mode))) {
869 xfs_warn(mp, "corrupted root inode %llu: not a directory",
870 (unsigned long long)rip->i_ino);
871 xfs_iunlock(rip, XFS_ILOCK_EXCL);
872 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
874 error = -EFSCORRUPTED;
877 mp->m_rootip = rip; /* save it */
879 xfs_iunlock(rip, XFS_ILOCK_EXCL);
882 * Initialize realtime inode pointers in the mount structure
884 error = xfs_rtmount_inodes(mp);
887 * Free up the root inode.
889 xfs_warn(mp, "failed to read RT inodes");
894 * If this is a read-only mount defer the superblock updates until
895 * the next remount into writeable mode. Otherwise we would never
896 * perform the update e.g. for the root filesystem.
898 if (mp->m_update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
899 error = xfs_mount_log_sb(mp, mp->m_update_flags);
901 xfs_warn(mp, "failed to write sb changes");
907 * Initialise the XFS quota management subsystem for this mount
909 if (XFS_IS_QUOTA_RUNNING(mp)) {
910 error = xfs_qm_newmount(mp, "amount, "aflags);
914 ASSERT(!XFS_IS_QUOTA_ON(mp));
917 * If a file system had quotas running earlier, but decided to
918 * mount without -o uquota/pquota/gquota options, revoke the
919 * quotachecked license.
921 if (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_ACCT) {
922 xfs_notice(mp, "resetting quota flags");
923 error = xfs_mount_reset_sbqflags(mp);
930 * Finish recovering the file system. This part needed to be
931 * delayed until after the root and real-time bitmap inodes
932 * were consistently read in.
934 error = xfs_log_mount_finish(mp);
936 xfs_warn(mp, "log mount finish failed");
941 * Complete the quota initialisation, post-log-replay component.
944 ASSERT(mp->m_qflags == 0);
945 mp->m_qflags = quotaflags;
947 xfs_qm_mount_quotas(mp);
951 * Now we are mounted, reserve a small amount of unused space for
952 * privileged transactions. This is needed so that transaction
953 * space required for critical operations can dip into this pool
954 * when at ENOSPC. This is needed for operations like create with
955 * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
956 * are not allowed to use this reserved space.
958 * This may drive us straight to ENOSPC on mount, but that implies
959 * we were already there on the last unmount. Warn if this occurs.
961 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
962 resblks = xfs_default_resblks(mp);
963 error = xfs_reserve_blocks(mp, &resblks, NULL);
966 "Unable to allocate reserve blocks. Continuing without reserve pool.");
972 xfs_rtunmount_inodes(mp);
978 if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
979 xfs_wait_buftarg(mp->m_logdev_targp);
980 xfs_wait_buftarg(mp->m_ddev_targp);
986 xfs_uuid_unmount(mp);
988 xfs_sysfs_del(&mp->m_kobj);
994 * This flushes out the inodes,dquots and the superblock, unmounts the
995 * log and makes sure that incore structures are freed.
999 struct xfs_mount *mp)
1004 cancel_delayed_work_sync(&mp->m_eofblocks_work);
1006 xfs_qm_unmount_quotas(mp);
1007 xfs_rtunmount_inodes(mp);
1008 IRELE(mp->m_rootip);
1011 * We can potentially deadlock here if we have an inode cluster
1012 * that has been freed has its buffer still pinned in memory because
1013 * the transaction is still sitting in a iclog. The stale inodes
1014 * on that buffer will have their flush locks held until the
1015 * transaction hits the disk and the callbacks run. the inode
1016 * flush takes the flush lock unconditionally and with nothing to
1017 * push out the iclog we will never get that unlocked. hence we
1018 * need to force the log first.
1020 xfs_log_force(mp, XFS_LOG_SYNC);
1023 * Flush all pending changes from the AIL.
1025 xfs_ail_push_all_sync(mp->m_ail);
1028 * And reclaim all inodes. At this point there should be no dirty
1029 * inodes and none should be pinned or locked, but use synchronous
1030 * reclaim just to be sure. We can stop background inode reclaim
1031 * here as well if it is still running.
1033 cancel_delayed_work_sync(&mp->m_reclaim_work);
1034 xfs_reclaim_inodes(mp, SYNC_WAIT);
1039 * Unreserve any blocks we have so that when we unmount we don't account
1040 * the reserved free space as used. This is really only necessary for
1041 * lazy superblock counting because it trusts the incore superblock
1042 * counters to be absolutely correct on clean unmount.
1044 * We don't bother correcting this elsewhere for lazy superblock
1045 * counting because on mount of an unclean filesystem we reconstruct the
1046 * correct counter value and this is irrelevant.
1048 * For non-lazy counter filesystems, this doesn't matter at all because
1049 * we only every apply deltas to the superblock and hence the incore
1050 * value does not matter....
1053 error = xfs_reserve_blocks(mp, &resblks, NULL);
1055 xfs_warn(mp, "Unable to free reserved block pool. "
1056 "Freespace may not be correct on next mount.");
1058 error = xfs_log_sbcount(mp);
1060 xfs_warn(mp, "Unable to update superblock counters. "
1061 "Freespace may not be correct on next mount.");
1063 xfs_log_unmount(mp);
1065 xfs_uuid_unmount(mp);
1068 xfs_errortag_clearall(mp, 0);
1072 xfs_sysfs_del(&mp->m_kobj);
1076 * Determine whether modifications can proceed. The caller specifies the minimum
1077 * freeze level for which modifications should not be allowed. This allows
1078 * certain operations to proceed while the freeze sequence is in progress, if
1083 struct xfs_mount *mp,
1086 ASSERT(level > SB_UNFROZEN);
1087 if ((mp->m_super->s_writers.frozen >= level) ||
1088 XFS_FORCED_SHUTDOWN(mp) || (mp->m_flags & XFS_MOUNT_RDONLY))
1097 * Sync the superblock counters to disk.
1099 * Note this code can be called during the process of freezing, so we use the
1100 * transaction allocator that does not block when the transaction subsystem is
1101 * in its frozen state.
1104 xfs_log_sbcount(xfs_mount_t *mp)
1109 /* allow this to proceed during the freeze sequence... */
1110 if (!xfs_fs_writable(mp, SB_FREEZE_COMPLETE))
1113 xfs_icsb_sync_counters(mp, 0);
1116 * we don't need to do this if we are updating the superblock
1117 * counters on every modification.
1119 if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1122 tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT, KM_SLEEP);
1123 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_sb, 0, 0);
1125 xfs_trans_cancel(tp, 0);
1129 xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1130 xfs_trans_set_sync(tp);
1131 error = xfs_trans_commit(tp, 0);
1136 * xfs_mod_incore_sb_unlocked() is a utility routine commonly used to apply
1137 * a delta to a specified field in the in-core superblock. Simply
1138 * switch on the field indicated and apply the delta to that field.
1139 * Fields are not allowed to dip below zero, so if the delta would
1140 * do this do not apply it and return EINVAL.
1142 * The m_sb_lock must be held when this routine is called.
1145 xfs_mod_incore_sb_unlocked(
1147 xfs_sb_field_t field,
1151 int scounter; /* short counter for 32 bit fields */
1152 long long lcounter; /* long counter for 64 bit fields */
1153 long long res_used, rem;
1156 * With the in-core superblock spin lock held, switch
1157 * on the indicated field. Apply the delta to the
1158 * proper field. If the fields value would dip below
1159 * 0, then do not apply the delta and return EINVAL.
1162 case XFS_SBS_ICOUNT:
1163 lcounter = (long long)mp->m_sb.sb_icount;
1169 mp->m_sb.sb_icount = lcounter;
1172 lcounter = (long long)mp->m_sb.sb_ifree;
1178 mp->m_sb.sb_ifree = lcounter;
1180 case XFS_SBS_FDBLOCKS:
1181 lcounter = (long long)
1182 mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1183 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1185 if (delta > 0) { /* Putting blocks back */
1186 if (res_used > delta) {
1187 mp->m_resblks_avail += delta;
1189 rem = delta - res_used;
1190 mp->m_resblks_avail = mp->m_resblks;
1193 } else { /* Taking blocks away */
1195 if (lcounter >= 0) {
1196 mp->m_sb.sb_fdblocks = lcounter +
1197 XFS_ALLOC_SET_ASIDE(mp);
1202 * We are out of blocks, use any available reserved
1203 * blocks if were allowed to.
1208 lcounter = (long long)mp->m_resblks_avail + delta;
1209 if (lcounter >= 0) {
1210 mp->m_resblks_avail = lcounter;
1213 printk_once(KERN_WARNING
1214 "Filesystem \"%s\": reserve blocks depleted! "
1215 "Consider increasing reserve pool size.",
1220 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1222 case XFS_SBS_FREXTENTS:
1223 lcounter = (long long)mp->m_sb.sb_frextents;
1228 mp->m_sb.sb_frextents = lcounter;
1230 case XFS_SBS_DBLOCKS:
1231 lcounter = (long long)mp->m_sb.sb_dblocks;
1237 mp->m_sb.sb_dblocks = lcounter;
1239 case XFS_SBS_AGCOUNT:
1240 scounter = mp->m_sb.sb_agcount;
1246 mp->m_sb.sb_agcount = scounter;
1248 case XFS_SBS_IMAX_PCT:
1249 scounter = mp->m_sb.sb_imax_pct;
1255 mp->m_sb.sb_imax_pct = scounter;
1257 case XFS_SBS_REXTSIZE:
1258 scounter = mp->m_sb.sb_rextsize;
1264 mp->m_sb.sb_rextsize = scounter;
1266 case XFS_SBS_RBMBLOCKS:
1267 scounter = mp->m_sb.sb_rbmblocks;
1273 mp->m_sb.sb_rbmblocks = scounter;
1275 case XFS_SBS_RBLOCKS:
1276 lcounter = (long long)mp->m_sb.sb_rblocks;
1282 mp->m_sb.sb_rblocks = lcounter;
1284 case XFS_SBS_REXTENTS:
1285 lcounter = (long long)mp->m_sb.sb_rextents;
1291 mp->m_sb.sb_rextents = lcounter;
1293 case XFS_SBS_REXTSLOG:
1294 scounter = mp->m_sb.sb_rextslog;
1300 mp->m_sb.sb_rextslog = scounter;
1309 * xfs_mod_incore_sb() is used to change a field in the in-core
1310 * superblock structure by the specified delta. This modification
1311 * is protected by the m_sb_lock. Just use the xfs_mod_incore_sb_unlocked()
1312 * routine to do the work.
1316 struct xfs_mount *mp,
1317 xfs_sb_field_t field,
1323 #ifdef HAVE_PERCPU_SB
1324 ASSERT(field < XFS_SBS_ICOUNT || field > XFS_SBS_FDBLOCKS);
1326 spin_lock(&mp->m_sb_lock);
1327 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1328 spin_unlock(&mp->m_sb_lock);
1334 * Change more than one field in the in-core superblock structure at a time.
1336 * The fields and changes to those fields are specified in the array of
1337 * xfs_mod_sb structures passed in. Either all of the specified deltas
1338 * will be applied or none of them will. If any modified field dips below 0,
1339 * then all modifications will be backed out and EINVAL will be returned.
1341 * Note that this function may not be used for the superblock values that
1342 * are tracked with the in-memory per-cpu counters - a direct call to
1343 * xfs_icsb_modify_counters is required for these.
1346 xfs_mod_incore_sb_batch(
1347 struct xfs_mount *mp,
1356 * Loop through the array of mod structures and apply each individually.
1357 * If any fail, then back out all those which have already been applied.
1358 * Do all of this within the scope of the m_sb_lock so that all of the
1359 * changes will be atomic.
1361 spin_lock(&mp->m_sb_lock);
1362 for (msbp = msb; msbp < (msb + nmsb); msbp++) {
1363 ASSERT(msbp->msb_field < XFS_SBS_ICOUNT ||
1364 msbp->msb_field > XFS_SBS_FDBLOCKS);
1366 error = xfs_mod_incore_sb_unlocked(mp, msbp->msb_field,
1367 msbp->msb_delta, rsvd);
1371 spin_unlock(&mp->m_sb_lock);
1375 while (--msbp >= msb) {
1376 error = xfs_mod_incore_sb_unlocked(mp, msbp->msb_field,
1377 -msbp->msb_delta, rsvd);
1380 spin_unlock(&mp->m_sb_lock);
1385 * xfs_getsb() is called to obtain the buffer for the superblock.
1386 * The buffer is returned locked and read in from disk.
1387 * The buffer should be released with a call to xfs_brelse().
1389 * If the flags parameter is BUF_TRYLOCK, then we'll only return
1390 * the superblock buffer if it can be locked without sleeping.
1391 * If it can't then we'll return NULL.
1395 struct xfs_mount *mp,
1398 struct xfs_buf *bp = mp->m_sb_bp;
1400 if (!xfs_buf_trylock(bp)) {
1401 if (flags & XBF_TRYLOCK)
1407 ASSERT(XFS_BUF_ISDONE(bp));
1412 * Used to free the superblock along various error paths.
1416 struct xfs_mount *mp)
1418 struct xfs_buf *bp = mp->m_sb_bp;
1426 * Used to log changes to the superblock unit and width fields which could
1427 * be altered by the mount options, as well as any potential sb_features2
1428 * fixup. Only the first superblock is updated.
1438 ASSERT(fields & (XFS_SB_UNIT | XFS_SB_WIDTH | XFS_SB_UUID |
1439 XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2 |
1440 XFS_SB_VERSIONNUM));
1442 tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1443 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_sb, 0, 0);
1445 xfs_trans_cancel(tp, 0);
1448 xfs_mod_sb(tp, fields);
1449 error = xfs_trans_commit(tp, 0);
1454 * If the underlying (data/log/rt) device is readonly, there are some
1455 * operations that cannot proceed.
1458 xfs_dev_is_read_only(
1459 struct xfs_mount *mp,
1462 if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1463 xfs_readonly_buftarg(mp->m_logdev_targp) ||
1464 (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1465 xfs_notice(mp, "%s required on read-only device.", message);
1466 xfs_notice(mp, "write access unavailable, cannot proceed.");
1472 #ifdef HAVE_PERCPU_SB
1474 * Per-cpu incore superblock counters
1476 * Simple concept, difficult implementation
1478 * Basically, replace the incore superblock counters with a distributed per cpu
1479 * counter for contended fields (e.g. free block count).
1481 * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1482 * hence needs to be accurately read when we are running low on space. Hence
1483 * there is a method to enable and disable the per-cpu counters based on how
1484 * much "stuff" is available in them.
1486 * Basically, a counter is enabled if there is enough free resource to justify
1487 * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1488 * ENOSPC), then we disable the counters to synchronise all callers and
1489 * re-distribute the available resources.
1491 * If, once we redistributed the available resources, we still get a failure,
1492 * we disable the per-cpu counter and go through the slow path.
1494 * The slow path is the current xfs_mod_incore_sb() function. This means that
1495 * when we disable a per-cpu counter, we need to drain its resources back to
1496 * the global superblock. We do this after disabling the counter to prevent
1497 * more threads from queueing up on the counter.
1499 * Essentially, this means that we still need a lock in the fast path to enable
1500 * synchronisation between the global counters and the per-cpu counters. This
1501 * is not a problem because the lock will be local to a CPU almost all the time
1502 * and have little contention except when we get to ENOSPC conditions.
1504 * Basically, this lock becomes a barrier that enables us to lock out the fast
1505 * path while we do things like enabling and disabling counters and
1506 * synchronising the counters.
1510 * 1. m_sb_lock before picking up per-cpu locks
1511 * 2. per-cpu locks always picked up via for_each_online_cpu() order
1512 * 3. accurate counter sync requires m_sb_lock + per cpu locks
1513 * 4. modifying per-cpu counters requires holding per-cpu lock
1514 * 5. modifying global counters requires holding m_sb_lock
1515 * 6. enabling or disabling a counter requires holding the m_sb_lock
1516 * and _none_ of the per-cpu locks.
1518 * Disabled counters are only ever re-enabled by a balance operation
1519 * that results in more free resources per CPU than a given threshold.
1520 * To ensure counters don't remain disabled, they are rebalanced when
1521 * the global resource goes above a higher threshold (i.e. some hysteresis
1522 * is present to prevent thrashing).
1525 #ifdef CONFIG_HOTPLUG_CPU
1527 * hot-plug CPU notifier support.
1529 * We need a notifier per filesystem as we need to be able to identify
1530 * the filesystem to balance the counters out. This is achieved by
1531 * having a notifier block embedded in the xfs_mount_t and doing pointer
1532 * magic to get the mount pointer from the notifier block address.
1535 xfs_icsb_cpu_notify(
1536 struct notifier_block *nfb,
1537 unsigned long action,
1540 xfs_icsb_cnts_t *cntp;
1543 mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1544 cntp = (xfs_icsb_cnts_t *)
1545 per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1547 case CPU_UP_PREPARE:
1548 case CPU_UP_PREPARE_FROZEN:
1549 /* Easy Case - initialize the area and locks, and
1550 * then rebalance when online does everything else for us. */
1551 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1554 case CPU_ONLINE_FROZEN:
1556 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1557 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1558 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1559 xfs_icsb_unlock(mp);
1562 case CPU_DEAD_FROZEN:
1563 /* Disable all the counters, then fold the dead cpu's
1564 * count into the total on the global superblock and
1565 * re-enable the counters. */
1567 spin_lock(&mp->m_sb_lock);
1568 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1569 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1570 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1572 mp->m_sb.sb_icount += cntp->icsb_icount;
1573 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1574 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1576 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1578 xfs_icsb_balance_counter_locked(mp, XFS_SBS_ICOUNT, 0);
1579 xfs_icsb_balance_counter_locked(mp, XFS_SBS_IFREE, 0);
1580 xfs_icsb_balance_counter_locked(mp, XFS_SBS_FDBLOCKS, 0);
1581 spin_unlock(&mp->m_sb_lock);
1582 xfs_icsb_unlock(mp);
1588 #endif /* CONFIG_HOTPLUG_CPU */
1591 xfs_icsb_init_counters(
1594 xfs_icsb_cnts_t *cntp;
1597 mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1598 if (mp->m_sb_cnts == NULL)
1601 for_each_online_cpu(i) {
1602 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1603 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1606 mutex_init(&mp->m_icsb_mutex);
1609 * start with all counters disabled so that the
1610 * initial balance kicks us off correctly
1612 mp->m_icsb_counters = -1;
1614 #ifdef CONFIG_HOTPLUG_CPU
1615 mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1616 mp->m_icsb_notifier.priority = 0;
1617 register_hotcpu_notifier(&mp->m_icsb_notifier);
1618 #endif /* CONFIG_HOTPLUG_CPU */
1624 xfs_icsb_reinit_counters(
1629 * start with all counters disabled so that the
1630 * initial balance kicks us off correctly
1632 mp->m_icsb_counters = -1;
1633 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1634 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1635 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1636 xfs_icsb_unlock(mp);
1640 xfs_icsb_destroy_counters(
1643 if (mp->m_sb_cnts) {
1644 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
1645 free_percpu(mp->m_sb_cnts);
1647 mutex_destroy(&mp->m_icsb_mutex);
1652 xfs_icsb_cnts_t *icsbp)
1654 while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1660 xfs_icsb_unlock_cntr(
1661 xfs_icsb_cnts_t *icsbp)
1663 clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1668 xfs_icsb_lock_all_counters(
1671 xfs_icsb_cnts_t *cntp;
1674 for_each_online_cpu(i) {
1675 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1676 xfs_icsb_lock_cntr(cntp);
1681 xfs_icsb_unlock_all_counters(
1684 xfs_icsb_cnts_t *cntp;
1687 for_each_online_cpu(i) {
1688 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1689 xfs_icsb_unlock_cntr(cntp);
1696 xfs_icsb_cnts_t *cnt,
1699 xfs_icsb_cnts_t *cntp;
1702 memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1704 if (!(flags & XFS_ICSB_LAZY_COUNT))
1705 xfs_icsb_lock_all_counters(mp);
1707 for_each_online_cpu(i) {
1708 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1709 cnt->icsb_icount += cntp->icsb_icount;
1710 cnt->icsb_ifree += cntp->icsb_ifree;
1711 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1714 if (!(flags & XFS_ICSB_LAZY_COUNT))
1715 xfs_icsb_unlock_all_counters(mp);
1719 xfs_icsb_counter_disabled(
1721 xfs_sb_field_t field)
1723 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1724 return test_bit(field, &mp->m_icsb_counters);
1728 xfs_icsb_disable_counter(
1730 xfs_sb_field_t field)
1732 xfs_icsb_cnts_t cnt;
1734 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1737 * If we are already disabled, then there is nothing to do
1738 * here. We check before locking all the counters to avoid
1739 * the expensive lock operation when being called in the
1740 * slow path and the counter is already disabled. This is
1741 * safe because the only time we set or clear this state is under
1744 if (xfs_icsb_counter_disabled(mp, field))
1747 xfs_icsb_lock_all_counters(mp);
1748 if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1749 /* drain back to superblock */
1751 xfs_icsb_count(mp, &cnt, XFS_ICSB_LAZY_COUNT);
1753 case XFS_SBS_ICOUNT:
1754 mp->m_sb.sb_icount = cnt.icsb_icount;
1757 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1759 case XFS_SBS_FDBLOCKS:
1760 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1767 xfs_icsb_unlock_all_counters(mp);
1771 xfs_icsb_enable_counter(
1773 xfs_sb_field_t field,
1777 xfs_icsb_cnts_t *cntp;
1780 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1782 xfs_icsb_lock_all_counters(mp);
1783 for_each_online_cpu(i) {
1784 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1786 case XFS_SBS_ICOUNT:
1787 cntp->icsb_icount = count + resid;
1790 cntp->icsb_ifree = count + resid;
1792 case XFS_SBS_FDBLOCKS:
1793 cntp->icsb_fdblocks = count + resid;
1801 clear_bit(field, &mp->m_icsb_counters);
1802 xfs_icsb_unlock_all_counters(mp);
1806 xfs_icsb_sync_counters_locked(
1810 xfs_icsb_cnts_t cnt;
1812 xfs_icsb_count(mp, &cnt, flags);
1814 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1815 mp->m_sb.sb_icount = cnt.icsb_icount;
1816 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1817 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1818 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1819 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1823 * Accurate update of per-cpu counters to incore superblock
1826 xfs_icsb_sync_counters(
1830 spin_lock(&mp->m_sb_lock);
1831 xfs_icsb_sync_counters_locked(mp, flags);
1832 spin_unlock(&mp->m_sb_lock);
1836 * Balance and enable/disable counters as necessary.
1838 * Thresholds for re-enabling counters are somewhat magic. inode counts are
1839 * chosen to be the same number as single on disk allocation chunk per CPU, and
1840 * free blocks is something far enough zero that we aren't going thrash when we
1841 * get near ENOSPC. We also need to supply a minimum we require per cpu to
1842 * prevent looping endlessly when xfs_alloc_space asks for more than will
1843 * be distributed to a single CPU but each CPU has enough blocks to be
1846 * Note that we can be called when counters are already disabled.
1847 * xfs_icsb_disable_counter() optimises the counter locking in this case to
1848 * prevent locking every per-cpu counter needlessly.
1851 #define XFS_ICSB_INO_CNTR_REENABLE (uint64_t)64
1852 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
1853 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
1855 xfs_icsb_balance_counter_locked(
1857 xfs_sb_field_t field,
1860 uint64_t count, resid;
1861 int weight = num_online_cpus();
1862 uint64_t min = (uint64_t)min_per_cpu;
1864 /* disable counter and sync counter */
1865 xfs_icsb_disable_counter(mp, field);
1867 /* update counters - first CPU gets residual*/
1869 case XFS_SBS_ICOUNT:
1870 count = mp->m_sb.sb_icount;
1871 resid = do_div(count, weight);
1872 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
1876 count = mp->m_sb.sb_ifree;
1877 resid = do_div(count, weight);
1878 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
1881 case XFS_SBS_FDBLOCKS:
1882 count = mp->m_sb.sb_fdblocks;
1883 resid = do_div(count, weight);
1884 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
1889 count = resid = 0; /* quiet, gcc */
1893 xfs_icsb_enable_counter(mp, field, count, resid);
1897 xfs_icsb_balance_counter(
1899 xfs_sb_field_t fields,
1902 spin_lock(&mp->m_sb_lock);
1903 xfs_icsb_balance_counter_locked(mp, fields, min_per_cpu);
1904 spin_unlock(&mp->m_sb_lock);
1908 xfs_icsb_modify_counters(
1910 xfs_sb_field_t field,
1914 xfs_icsb_cnts_t *icsbp;
1915 long long lcounter; /* long counter for 64 bit fields */
1921 icsbp = this_cpu_ptr(mp->m_sb_cnts);
1924 * if the counter is disabled, go to slow path
1926 if (unlikely(xfs_icsb_counter_disabled(mp, field)))
1928 xfs_icsb_lock_cntr(icsbp);
1929 if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
1930 xfs_icsb_unlock_cntr(icsbp);
1935 case XFS_SBS_ICOUNT:
1936 lcounter = icsbp->icsb_icount;
1938 if (unlikely(lcounter < 0))
1939 goto balance_counter;
1940 icsbp->icsb_icount = lcounter;
1944 lcounter = icsbp->icsb_ifree;
1946 if (unlikely(lcounter < 0))
1947 goto balance_counter;
1948 icsbp->icsb_ifree = lcounter;
1951 case XFS_SBS_FDBLOCKS:
1952 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
1954 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1956 if (unlikely(lcounter < 0))
1957 goto balance_counter;
1958 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1964 xfs_icsb_unlock_cntr(icsbp);
1972 * serialise with a mutex so we don't burn lots of cpu on
1973 * the superblock lock. We still need to hold the superblock
1974 * lock, however, when we modify the global structures.
1979 * Now running atomically.
1981 * If the counter is enabled, someone has beaten us to rebalancing.
1982 * Drop the lock and try again in the fast path....
1984 if (!(xfs_icsb_counter_disabled(mp, field))) {
1985 xfs_icsb_unlock(mp);
1990 * The counter is currently disabled. Because we are
1991 * running atomically here, we know a rebalance cannot
1992 * be in progress. Hence we can go straight to operating
1993 * on the global superblock. We do not call xfs_mod_incore_sb()
1994 * here even though we need to get the m_sb_lock. Doing so
1995 * will cause us to re-enter this function and deadlock.
1996 * Hence we get the m_sb_lock ourselves and then call
1997 * xfs_mod_incore_sb_unlocked() as the unlocked path operates
1998 * directly on the global counters.
2000 spin_lock(&mp->m_sb_lock);
2001 ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2002 spin_unlock(&mp->m_sb_lock);
2005 * Now that we've modified the global superblock, we
2006 * may be able to re-enable the distributed counters
2007 * (e.g. lots of space just got freed). After that
2011 xfs_icsb_balance_counter(mp, field, 0);
2012 xfs_icsb_unlock(mp);
2016 xfs_icsb_unlock_cntr(icsbp);
2020 * We may have multiple threads here if multiple per-cpu
2021 * counters run dry at the same time. This will mean we can
2022 * do more balances than strictly necessary but it is not
2023 * the common slowpath case.
2028 * running atomically.
2030 * This will leave the counter in the correct state for future
2031 * accesses. After the rebalance, we simply try again and our retry
2032 * will either succeed through the fast path or slow path without
2033 * another balance operation being required.
2035 xfs_icsb_balance_counter(mp, field, delta);
2036 xfs_icsb_unlock(mp);