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"
28 #include "xfs_mount.h"
29 #include "xfs_da_format.h"
30 #include "xfs_inode.h"
32 #include "xfs_ialloc.h"
33 #include "xfs_alloc.h"
34 #include "xfs_rtalloc.h"
36 #include "xfs_trans.h"
37 #include "xfs_trans_priv.h"
39 #include "xfs_error.h"
40 #include "xfs_quota.h"
41 #include "xfs_fsops.h"
42 #include "xfs_trace.h"
43 #include "xfs_icache.h"
44 #include "xfs_dinode.h"
48 STATIC void xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
50 STATIC void xfs_icsb_balance_counter_locked(xfs_mount_t *, xfs_sb_field_t,
52 STATIC void xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
55 #define xfs_icsb_balance_counter(mp, a, b) do { } while (0)
56 #define xfs_icsb_balance_counter_locked(mp, a, b) do { } while (0)
59 static DEFINE_MUTEX(xfs_uuid_table_mutex);
60 static int xfs_uuid_table_size;
61 static uuid_t *xfs_uuid_table;
64 * See if the UUID is unique among mounted XFS filesystems.
65 * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
71 uuid_t *uuid = &mp->m_sb.sb_uuid;
74 if (mp->m_flags & XFS_MOUNT_NOUUID)
77 if (uuid_is_nil(uuid)) {
78 xfs_warn(mp, "Filesystem has nil UUID - can't mount");
79 return XFS_ERROR(EINVAL);
82 mutex_lock(&xfs_uuid_table_mutex);
83 for (i = 0, hole = -1; i < xfs_uuid_table_size; i++) {
84 if (uuid_is_nil(&xfs_uuid_table[i])) {
88 if (uuid_equal(uuid, &xfs_uuid_table[i]))
93 xfs_uuid_table = kmem_realloc(xfs_uuid_table,
94 (xfs_uuid_table_size + 1) * sizeof(*xfs_uuid_table),
95 xfs_uuid_table_size * sizeof(*xfs_uuid_table),
97 hole = xfs_uuid_table_size++;
99 xfs_uuid_table[hole] = *uuid;
100 mutex_unlock(&xfs_uuid_table_mutex);
105 mutex_unlock(&xfs_uuid_table_mutex);
106 xfs_warn(mp, "Filesystem has duplicate UUID %pU - can't mount", uuid);
107 return XFS_ERROR(EINVAL);
112 struct xfs_mount *mp)
114 uuid_t *uuid = &mp->m_sb.sb_uuid;
117 if (mp->m_flags & XFS_MOUNT_NOUUID)
120 mutex_lock(&xfs_uuid_table_mutex);
121 for (i = 0; i < xfs_uuid_table_size; i++) {
122 if (uuid_is_nil(&xfs_uuid_table[i]))
124 if (!uuid_equal(uuid, &xfs_uuid_table[i]))
126 memset(&xfs_uuid_table[i], 0, sizeof(uuid_t));
129 ASSERT(i < xfs_uuid_table_size);
130 mutex_unlock(&xfs_uuid_table_mutex);
136 struct rcu_head *head)
138 struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
140 ASSERT(atomic_read(&pag->pag_ref) == 0);
145 * Free up the per-ag resources associated with the mount structure.
152 struct xfs_perag *pag;
154 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
155 spin_lock(&mp->m_perag_lock);
156 pag = radix_tree_delete(&mp->m_perag_tree, agno);
157 spin_unlock(&mp->m_perag_lock);
159 ASSERT(atomic_read(&pag->pag_ref) == 0);
160 call_rcu(&pag->rcu_head, __xfs_free_perag);
165 * Check size of device based on the (data/realtime) block count.
166 * Note: this check is used by the growfs code as well as mount.
169 xfs_sb_validate_fsb_count(
173 ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
174 ASSERT(sbp->sb_blocklog >= BBSHIFT);
176 #if XFS_BIG_BLKNOS /* Limited by ULONG_MAX of page cache index */
177 if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
179 #else /* Limited by UINT_MAX of sectors */
180 if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
187 xfs_initialize_perag(
189 xfs_agnumber_t agcount,
190 xfs_agnumber_t *maxagi)
192 xfs_agnumber_t index;
193 xfs_agnumber_t first_initialised = 0;
197 xfs_sb_t *sbp = &mp->m_sb;
201 * Walk the current per-ag tree so we don't try to initialise AGs
202 * that already exist (growfs case). Allocate and insert all the
203 * AGs we don't find ready for initialisation.
205 for (index = 0; index < agcount; index++) {
206 pag = xfs_perag_get(mp, index);
211 if (!first_initialised)
212 first_initialised = index;
214 pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
217 pag->pag_agno = index;
219 spin_lock_init(&pag->pag_ici_lock);
220 mutex_init(&pag->pag_ici_reclaim_lock);
221 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
222 spin_lock_init(&pag->pag_buf_lock);
223 pag->pag_buf_tree = RB_ROOT;
225 if (radix_tree_preload(GFP_NOFS))
228 spin_lock(&mp->m_perag_lock);
229 if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
231 spin_unlock(&mp->m_perag_lock);
232 radix_tree_preload_end();
236 spin_unlock(&mp->m_perag_lock);
237 radix_tree_preload_end();
241 * If we mount with the inode64 option, or no inode overflows
242 * the legacy 32-bit address space clear the inode32 option.
244 agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
245 ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
247 if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > XFS_MAXINUMBER_32)
248 mp->m_flags |= XFS_MOUNT_32BITINODES;
250 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
252 if (mp->m_flags & XFS_MOUNT_32BITINODES)
253 index = xfs_set_inode32(mp);
255 index = xfs_set_inode64(mp);
263 for (; index > first_initialised; index--) {
264 pag = radix_tree_delete(&mp->m_perag_tree, index);
273 * Does the initial read of the superblock.
277 struct xfs_mount *mp,
280 unsigned int sector_size;
282 struct xfs_sb *sbp = &mp->m_sb;
284 int loud = !(flags & XFS_MFSI_QUIET);
285 const struct xfs_buf_ops *buf_ops;
287 ASSERT(mp->m_sb_bp == NULL);
288 ASSERT(mp->m_ddev_targp != NULL);
291 * For the initial read, we must guess at the sector
292 * size based on the block device. It's enough to
293 * get the sb_sectsize out of the superblock and
294 * then reread with the proper length.
295 * We don't verify it yet, because it may not be complete.
297 sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
301 * Allocate a (locked) buffer to hold the superblock.
302 * This will be kept around at all times to optimize
303 * access to the superblock.
306 bp = xfs_buf_read_uncached(mp->m_ddev_targp, XFS_SB_DADDR,
307 BTOBB(sector_size), 0, buf_ops);
310 xfs_warn(mp, "SB buffer read failed");
316 xfs_warn(mp, "SB validate failed with error %d.", error);
317 /* bad CRC means corrupted metadata */
318 if (error == EFSBADCRC)
319 error = EFSCORRUPTED;
324 * Initialize the mount structure from the superblock.
326 xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
327 xfs_sb_quota_from_disk(&mp->m_sb);
330 * We must be able to do sector-sized and sector-aligned IO.
332 if (sector_size > sbp->sb_sectsize) {
334 xfs_warn(mp, "device supports %u byte sectors (not %u)",
335 sector_size, sbp->sb_sectsize);
341 * Re-read the superblock so the buffer is correctly sized,
342 * and properly verified.
344 if (buf_ops == NULL) {
346 sector_size = sbp->sb_sectsize;
347 buf_ops = loud ? &xfs_sb_buf_ops : &xfs_sb_quiet_buf_ops;
351 /* Initialize per-cpu counters */
352 xfs_icsb_reinit_counters(mp);
354 /* no need to be quiet anymore, so reset the buf ops */
355 bp->b_ops = &xfs_sb_buf_ops;
367 * Update alignment values based on mount options and sb values
370 xfs_update_alignment(xfs_mount_t *mp)
372 xfs_sb_t *sbp = &(mp->m_sb);
376 * If stripe unit and stripe width are not multiples
377 * of the fs blocksize turn off alignment.
379 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
380 (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
382 "alignment check failed: sunit/swidth vs. blocksize(%d)",
384 return XFS_ERROR(EINVAL);
387 * Convert the stripe unit and width to FSBs.
389 mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
390 if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
392 "alignment check failed: sunit/swidth vs. agsize(%d)",
394 return XFS_ERROR(EINVAL);
395 } else if (mp->m_dalign) {
396 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
399 "alignment check failed: sunit(%d) less than bsize(%d)",
400 mp->m_dalign, sbp->sb_blocksize);
401 return XFS_ERROR(EINVAL);
406 * Update superblock with new values
409 if (xfs_sb_version_hasdalign(sbp)) {
410 if (sbp->sb_unit != mp->m_dalign) {
411 sbp->sb_unit = mp->m_dalign;
412 mp->m_update_flags |= XFS_SB_UNIT;
414 if (sbp->sb_width != mp->m_swidth) {
415 sbp->sb_width = mp->m_swidth;
416 mp->m_update_flags |= XFS_SB_WIDTH;
420 "cannot change alignment: superblock does not support data alignment");
421 return XFS_ERROR(EINVAL);
423 } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
424 xfs_sb_version_hasdalign(&mp->m_sb)) {
425 mp->m_dalign = sbp->sb_unit;
426 mp->m_swidth = sbp->sb_width;
433 * Set the maximum inode count for this filesystem
436 xfs_set_maxicount(xfs_mount_t *mp)
438 xfs_sb_t *sbp = &(mp->m_sb);
441 if (sbp->sb_imax_pct) {
443 * Make sure the maximum inode count is a multiple
444 * of the units we allocate inodes in.
446 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
448 do_div(icount, mp->m_ialloc_blks);
449 mp->m_maxicount = (icount * mp->m_ialloc_blks) <<
457 * Set the default minimum read and write sizes unless
458 * already specified in a mount option.
459 * We use smaller I/O sizes when the file system
460 * is being used for NFS service (wsync mount option).
463 xfs_set_rw_sizes(xfs_mount_t *mp)
465 xfs_sb_t *sbp = &(mp->m_sb);
466 int readio_log, writeio_log;
468 if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
469 if (mp->m_flags & XFS_MOUNT_WSYNC) {
470 readio_log = XFS_WSYNC_READIO_LOG;
471 writeio_log = XFS_WSYNC_WRITEIO_LOG;
473 readio_log = XFS_READIO_LOG_LARGE;
474 writeio_log = XFS_WRITEIO_LOG_LARGE;
477 readio_log = mp->m_readio_log;
478 writeio_log = mp->m_writeio_log;
481 if (sbp->sb_blocklog > readio_log) {
482 mp->m_readio_log = sbp->sb_blocklog;
484 mp->m_readio_log = readio_log;
486 mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
487 if (sbp->sb_blocklog > writeio_log) {
488 mp->m_writeio_log = sbp->sb_blocklog;
490 mp->m_writeio_log = writeio_log;
492 mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
496 * precalculate the low space thresholds for dynamic speculative preallocation.
499 xfs_set_low_space_thresholds(
500 struct xfs_mount *mp)
504 for (i = 0; i < XFS_LOWSP_MAX; i++) {
505 __uint64_t space = mp->m_sb.sb_dblocks;
508 mp->m_low_space[i] = space * (i + 1);
514 * Set whether we're using inode alignment.
517 xfs_set_inoalignment(xfs_mount_t *mp)
519 if (xfs_sb_version_hasalign(&mp->m_sb) &&
520 mp->m_sb.sb_inoalignmt >=
521 XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
522 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
524 mp->m_inoalign_mask = 0;
526 * If we are using stripe alignment, check whether
527 * the stripe unit is a multiple of the inode alignment
529 if (mp->m_dalign && mp->m_inoalign_mask &&
530 !(mp->m_dalign & mp->m_inoalign_mask))
531 mp->m_sinoalign = mp->m_dalign;
537 * Check that the data (and log if separate) is an ok size.
540 xfs_check_sizes(xfs_mount_t *mp)
545 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
546 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
547 xfs_warn(mp, "filesystem size mismatch detected");
548 return XFS_ERROR(EFBIG);
550 bp = xfs_buf_read_uncached(mp->m_ddev_targp,
551 d - XFS_FSS_TO_BB(mp, 1),
552 XFS_FSS_TO_BB(mp, 1), 0, NULL);
554 xfs_warn(mp, "last sector read failed");
559 if (mp->m_logdev_targp != mp->m_ddev_targp) {
560 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
561 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
562 xfs_warn(mp, "log size mismatch detected");
563 return XFS_ERROR(EFBIG);
565 bp = xfs_buf_read_uncached(mp->m_logdev_targp,
566 d - XFS_FSB_TO_BB(mp, 1),
567 XFS_FSB_TO_BB(mp, 1), 0, NULL);
569 xfs_warn(mp, "log device read failed");
578 * Clear the quotaflags in memory and in the superblock.
581 xfs_mount_reset_sbqflags(
582 struct xfs_mount *mp)
585 struct xfs_trans *tp;
590 * It is OK to look at sb_qflags here in mount path,
593 if (mp->m_sb.sb_qflags == 0)
595 spin_lock(&mp->m_sb_lock);
596 mp->m_sb.sb_qflags = 0;
597 spin_unlock(&mp->m_sb_lock);
600 * If the fs is readonly, let the incore superblock run
601 * with quotas off but don't flush the update out to disk
603 if (mp->m_flags & XFS_MOUNT_RDONLY)
606 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
607 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_qm_sbchange, 0, 0);
609 xfs_trans_cancel(tp, 0);
610 xfs_alert(mp, "%s: Superblock update failed!", __func__);
614 xfs_mod_sb(tp, XFS_SB_QFLAGS);
615 return xfs_trans_commit(tp, 0);
619 xfs_default_resblks(xfs_mount_t *mp)
624 * We default to 5% or 8192 fsbs of space reserved, whichever is
625 * smaller. This is intended to cover concurrent allocation
626 * transactions when we initially hit enospc. These each require a 4
627 * block reservation. Hence by default we cover roughly 2000 concurrent
628 * allocation reservations.
630 resblks = mp->m_sb.sb_dblocks;
632 resblks = min_t(__uint64_t, resblks, 8192);
637 * This function does the following on an initial mount of a file system:
638 * - reads the superblock from disk and init the mount struct
639 * - if we're a 32-bit kernel, do a size check on the superblock
640 * so we don't mount terabyte filesystems
641 * - init mount struct realtime fields
642 * - allocate inode hash table for fs
643 * - init directory manager
644 * - perform recovery and init the log manager
650 xfs_sb_t *sbp = &(mp->m_sb);
657 xfs_sb_mount_common(mp, sbp);
660 * Check for a mismatched features2 values. Older kernels
661 * read & wrote into the wrong sb offset for sb_features2
662 * on some platforms due to xfs_sb_t not being 64bit size aligned
663 * when sb_features2 was added, which made older superblock
664 * reading/writing routines swap it as a 64-bit value.
666 * For backwards compatibility, we make both slots equal.
668 * If we detect a mismatched field, we OR the set bits into the
669 * existing features2 field in case it has already been modified; we
670 * don't want to lose any features. We then update the bad location
671 * with the ORed value so that older kernels will see any features2
672 * flags, and mark the two fields as needing updates once the
673 * transaction subsystem is online.
675 if (xfs_sb_has_mismatched_features2(sbp)) {
676 xfs_warn(mp, "correcting sb_features alignment problem");
677 sbp->sb_features2 |= sbp->sb_bad_features2;
678 sbp->sb_bad_features2 = sbp->sb_features2;
679 mp->m_update_flags |= XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2;
682 * Re-check for ATTR2 in case it was found in bad_features2
685 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
686 !(mp->m_flags & XFS_MOUNT_NOATTR2))
687 mp->m_flags |= XFS_MOUNT_ATTR2;
690 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
691 (mp->m_flags & XFS_MOUNT_NOATTR2)) {
692 xfs_sb_version_removeattr2(&mp->m_sb);
693 mp->m_update_flags |= XFS_SB_FEATURES2;
695 /* update sb_versionnum for the clearing of the morebits */
696 if (!sbp->sb_features2)
697 mp->m_update_flags |= XFS_SB_VERSIONNUM;
701 * Check if sb_agblocks is aligned at stripe boundary
702 * If sb_agblocks is NOT aligned turn off m_dalign since
703 * allocator alignment is within an ag, therefore ag has
704 * to be aligned at stripe boundary.
706 error = xfs_update_alignment(mp);
710 xfs_alloc_compute_maxlevels(mp);
711 xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
712 xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
713 xfs_ialloc_compute_maxlevels(mp);
715 xfs_set_maxicount(mp);
717 error = xfs_uuid_mount(mp);
722 * Set the minimum read and write sizes
724 xfs_set_rw_sizes(mp);
726 /* set the low space thresholds for dynamic preallocation */
727 xfs_set_low_space_thresholds(mp);
730 * Set the inode cluster size.
731 * This may still be overridden by the file system
732 * block size if it is larger than the chosen cluster size.
734 * For v5 filesystems, scale the cluster size with the inode size to
735 * keep a constant ratio of inode per cluster buffer, but only if mkfs
736 * has set the inode alignment value appropriately for larger cluster
739 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
740 if (xfs_sb_version_hascrc(&mp->m_sb)) {
741 int new_size = mp->m_inode_cluster_size;
743 new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
744 if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
745 mp->m_inode_cluster_size = new_size;
746 xfs_info(mp, "Using inode cluster size of %d bytes",
747 mp->m_inode_cluster_size);
751 * Set inode alignment fields
753 xfs_set_inoalignment(mp);
756 * Check that the data (and log if separate) is an ok size.
758 error = xfs_check_sizes(mp);
760 goto out_remove_uuid;
763 * Initialize realtime fields in the mount structure
765 error = xfs_rtmount_init(mp);
767 xfs_warn(mp, "RT mount failed");
768 goto out_remove_uuid;
772 * Copies the low order bits of the timestamp and the randomly
773 * set "sequence" number out of a UUID.
775 uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
777 mp->m_dmevmask = 0; /* not persistent; set after each mount */
782 * Initialize the attribute manager's entries.
784 mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
787 * Initialize the precomputed transaction reservations values.
792 * Allocate and initialize the per-ag data.
794 spin_lock_init(&mp->m_perag_lock);
795 INIT_RADIX_TREE(&mp->m_perag_tree, GFP_ATOMIC);
796 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
798 xfs_warn(mp, "Failed per-ag init: %d", error);
799 goto out_remove_uuid;
802 if (!sbp->sb_logblocks) {
803 xfs_warn(mp, "no log defined");
804 XFS_ERROR_REPORT("xfs_mountfs", XFS_ERRLEVEL_LOW, mp);
805 error = XFS_ERROR(EFSCORRUPTED);
810 * log's mount-time initialization. Perform 1st part recovery if needed
812 error = xfs_log_mount(mp, mp->m_logdev_targp,
813 XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
814 XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
816 xfs_warn(mp, "log mount failed");
821 * Now the log is mounted, we know if it was an unclean shutdown or
822 * not. If it was, with the first phase of recovery has completed, we
823 * have consistent AG blocks on disk. We have not recovered EFIs yet,
824 * but they are recovered transactionally in the second recovery phase
827 * Hence we can safely re-initialise incore superblock counters from
828 * the per-ag data. These may not be correct if the filesystem was not
829 * cleanly unmounted, so we need to wait for recovery to finish before
832 * If the filesystem was cleanly unmounted, then we can trust the
833 * values in the superblock to be correct and we don't need to do
836 * If we are currently making the filesystem, the initialisation will
837 * fail as the perag data is in an undefined state.
839 if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
840 !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
841 !mp->m_sb.sb_inprogress) {
842 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
848 * Get and sanity-check the root inode.
849 * Save the pointer to it in the mount structure.
851 error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip);
853 xfs_warn(mp, "failed to read root inode");
854 goto out_log_dealloc;
859 if (unlikely(!S_ISDIR(rip->i_d.di_mode))) {
860 xfs_warn(mp, "corrupted root inode %llu: not a directory",
861 (unsigned long long)rip->i_ino);
862 xfs_iunlock(rip, XFS_ILOCK_EXCL);
863 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
865 error = XFS_ERROR(EFSCORRUPTED);
868 mp->m_rootip = rip; /* save it */
870 xfs_iunlock(rip, XFS_ILOCK_EXCL);
873 * Initialize realtime inode pointers in the mount structure
875 error = xfs_rtmount_inodes(mp);
878 * Free up the root inode.
880 xfs_warn(mp, "failed to read RT inodes");
885 * If this is a read-only mount defer the superblock updates until
886 * the next remount into writeable mode. Otherwise we would never
887 * perform the update e.g. for the root filesystem.
889 if (mp->m_update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
890 error = xfs_mount_log_sb(mp, mp->m_update_flags);
892 xfs_warn(mp, "failed to write sb changes");
898 * Initialise the XFS quota management subsystem for this mount
900 if (XFS_IS_QUOTA_RUNNING(mp)) {
901 error = xfs_qm_newmount(mp, "amount, "aflags);
905 ASSERT(!XFS_IS_QUOTA_ON(mp));
908 * If a file system had quotas running earlier, but decided to
909 * mount without -o uquota/pquota/gquota options, revoke the
910 * quotachecked license.
912 if (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_ACCT) {
913 xfs_notice(mp, "resetting quota flags");
914 error = xfs_mount_reset_sbqflags(mp);
921 * Finish recovering the file system. This part needed to be
922 * delayed until after the root and real-time bitmap inodes
923 * were consistently read in.
925 error = xfs_log_mount_finish(mp);
927 xfs_warn(mp, "log mount finish failed");
932 * Complete the quota initialisation, post-log-replay component.
935 ASSERT(mp->m_qflags == 0);
936 mp->m_qflags = quotaflags;
938 xfs_qm_mount_quotas(mp);
942 * Now we are mounted, reserve a small amount of unused space for
943 * privileged transactions. This is needed so that transaction
944 * space required for critical operations can dip into this pool
945 * when at ENOSPC. This is needed for operations like create with
946 * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
947 * are not allowed to use this reserved space.
949 * This may drive us straight to ENOSPC on mount, but that implies
950 * we were already there on the last unmount. Warn if this occurs.
952 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
953 resblks = xfs_default_resblks(mp);
954 error = xfs_reserve_blocks(mp, &resblks, NULL);
957 "Unable to allocate reserve blocks. Continuing without reserve pool.");
963 xfs_rtunmount_inodes(mp);
969 if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
970 xfs_wait_buftarg(mp->m_logdev_targp);
971 xfs_wait_buftarg(mp->m_ddev_targp);
975 xfs_uuid_unmount(mp);
981 * This flushes out the inodes,dquots and the superblock, unmounts the
982 * log and makes sure that incore structures are freed.
986 struct xfs_mount *mp)
991 cancel_delayed_work_sync(&mp->m_eofblocks_work);
993 xfs_qm_unmount_quotas(mp);
994 xfs_rtunmount_inodes(mp);
998 * We can potentially deadlock here if we have an inode cluster
999 * that has been freed has its buffer still pinned in memory because
1000 * the transaction is still sitting in a iclog. The stale inodes
1001 * on that buffer will have their flush locks held until the
1002 * transaction hits the disk and the callbacks run. the inode
1003 * flush takes the flush lock unconditionally and with nothing to
1004 * push out the iclog we will never get that unlocked. hence we
1005 * need to force the log first.
1007 xfs_log_force(mp, XFS_LOG_SYNC);
1010 * Flush all pending changes from the AIL.
1012 xfs_ail_push_all_sync(mp->m_ail);
1015 * And reclaim all inodes. At this point there should be no dirty
1016 * inodes and none should be pinned or locked, but use synchronous
1017 * reclaim just to be sure. We can stop background inode reclaim
1018 * here as well if it is still running.
1020 cancel_delayed_work_sync(&mp->m_reclaim_work);
1021 xfs_reclaim_inodes(mp, SYNC_WAIT);
1026 * Unreserve any blocks we have so that when we unmount we don't account
1027 * the reserved free space as used. This is really only necessary for
1028 * lazy superblock counting because it trusts the incore superblock
1029 * counters to be absolutely correct on clean unmount.
1031 * We don't bother correcting this elsewhere for lazy superblock
1032 * counting because on mount of an unclean filesystem we reconstruct the
1033 * correct counter value and this is irrelevant.
1035 * For non-lazy counter filesystems, this doesn't matter at all because
1036 * we only every apply deltas to the superblock and hence the incore
1037 * value does not matter....
1040 error = xfs_reserve_blocks(mp, &resblks, NULL);
1042 xfs_warn(mp, "Unable to free reserved block pool. "
1043 "Freespace may not be correct on next mount.");
1045 error = xfs_log_sbcount(mp);
1047 xfs_warn(mp, "Unable to update superblock counters. "
1048 "Freespace may not be correct on next mount.");
1050 xfs_log_unmount(mp);
1051 xfs_uuid_unmount(mp);
1054 xfs_errortag_clearall(mp, 0);
1060 xfs_fs_writable(xfs_mount_t *mp)
1062 return !(mp->m_super->s_writers.frozen || XFS_FORCED_SHUTDOWN(mp) ||
1063 (mp->m_flags & XFS_MOUNT_RDONLY));
1069 * Sync the superblock counters to disk.
1071 * Note this code can be called during the process of freezing, so
1072 * we may need to use the transaction allocator which does not
1073 * block when the transaction subsystem is in its frozen state.
1076 xfs_log_sbcount(xfs_mount_t *mp)
1081 if (!xfs_fs_writable(mp))
1084 xfs_icsb_sync_counters(mp, 0);
1087 * we don't need to do this if we are updating the superblock
1088 * counters on every modification.
1090 if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1093 tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT, KM_SLEEP);
1094 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_sb, 0, 0);
1096 xfs_trans_cancel(tp, 0);
1100 xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1101 xfs_trans_set_sync(tp);
1102 error = xfs_trans_commit(tp, 0);
1107 * xfs_mod_incore_sb_unlocked() is a utility routine commonly used to apply
1108 * a delta to a specified field in the in-core superblock. Simply
1109 * switch on the field indicated and apply the delta to that field.
1110 * Fields are not allowed to dip below zero, so if the delta would
1111 * do this do not apply it and return EINVAL.
1113 * The m_sb_lock must be held when this routine is called.
1116 xfs_mod_incore_sb_unlocked(
1118 xfs_sb_field_t field,
1122 int scounter; /* short counter for 32 bit fields */
1123 long long lcounter; /* long counter for 64 bit fields */
1124 long long res_used, rem;
1127 * With the in-core superblock spin lock held, switch
1128 * on the indicated field. Apply the delta to the
1129 * proper field. If the fields value would dip below
1130 * 0, then do not apply the delta and return EINVAL.
1133 case XFS_SBS_ICOUNT:
1134 lcounter = (long long)mp->m_sb.sb_icount;
1138 return XFS_ERROR(EINVAL);
1140 mp->m_sb.sb_icount = lcounter;
1143 lcounter = (long long)mp->m_sb.sb_ifree;
1147 return XFS_ERROR(EINVAL);
1149 mp->m_sb.sb_ifree = lcounter;
1151 case XFS_SBS_FDBLOCKS:
1152 lcounter = (long long)
1153 mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1154 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1156 if (delta > 0) { /* Putting blocks back */
1157 if (res_used > delta) {
1158 mp->m_resblks_avail += delta;
1160 rem = delta - res_used;
1161 mp->m_resblks_avail = mp->m_resblks;
1164 } else { /* Taking blocks away */
1166 if (lcounter >= 0) {
1167 mp->m_sb.sb_fdblocks = lcounter +
1168 XFS_ALLOC_SET_ASIDE(mp);
1173 * We are out of blocks, use any available reserved
1174 * blocks if were allowed to.
1177 return XFS_ERROR(ENOSPC);
1179 lcounter = (long long)mp->m_resblks_avail + delta;
1180 if (lcounter >= 0) {
1181 mp->m_resblks_avail = lcounter;
1184 printk_once(KERN_WARNING
1185 "Filesystem \"%s\": reserve blocks depleted! "
1186 "Consider increasing reserve pool size.",
1188 return XFS_ERROR(ENOSPC);
1191 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1193 case XFS_SBS_FREXTENTS:
1194 lcounter = (long long)mp->m_sb.sb_frextents;
1197 return XFS_ERROR(ENOSPC);
1199 mp->m_sb.sb_frextents = lcounter;
1201 case XFS_SBS_DBLOCKS:
1202 lcounter = (long long)mp->m_sb.sb_dblocks;
1206 return XFS_ERROR(EINVAL);
1208 mp->m_sb.sb_dblocks = lcounter;
1210 case XFS_SBS_AGCOUNT:
1211 scounter = mp->m_sb.sb_agcount;
1215 return XFS_ERROR(EINVAL);
1217 mp->m_sb.sb_agcount = scounter;
1219 case XFS_SBS_IMAX_PCT:
1220 scounter = mp->m_sb.sb_imax_pct;
1224 return XFS_ERROR(EINVAL);
1226 mp->m_sb.sb_imax_pct = scounter;
1228 case XFS_SBS_REXTSIZE:
1229 scounter = mp->m_sb.sb_rextsize;
1233 return XFS_ERROR(EINVAL);
1235 mp->m_sb.sb_rextsize = scounter;
1237 case XFS_SBS_RBMBLOCKS:
1238 scounter = mp->m_sb.sb_rbmblocks;
1242 return XFS_ERROR(EINVAL);
1244 mp->m_sb.sb_rbmblocks = scounter;
1246 case XFS_SBS_RBLOCKS:
1247 lcounter = (long long)mp->m_sb.sb_rblocks;
1251 return XFS_ERROR(EINVAL);
1253 mp->m_sb.sb_rblocks = lcounter;
1255 case XFS_SBS_REXTENTS:
1256 lcounter = (long long)mp->m_sb.sb_rextents;
1260 return XFS_ERROR(EINVAL);
1262 mp->m_sb.sb_rextents = lcounter;
1264 case XFS_SBS_REXTSLOG:
1265 scounter = mp->m_sb.sb_rextslog;
1269 return XFS_ERROR(EINVAL);
1271 mp->m_sb.sb_rextslog = scounter;
1275 return XFS_ERROR(EINVAL);
1280 * xfs_mod_incore_sb() is used to change a field in the in-core
1281 * superblock structure by the specified delta. This modification
1282 * is protected by the m_sb_lock. Just use the xfs_mod_incore_sb_unlocked()
1283 * routine to do the work.
1287 struct xfs_mount *mp,
1288 xfs_sb_field_t field,
1294 #ifdef HAVE_PERCPU_SB
1295 ASSERT(field < XFS_SBS_ICOUNT || field > XFS_SBS_FDBLOCKS);
1297 spin_lock(&mp->m_sb_lock);
1298 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1299 spin_unlock(&mp->m_sb_lock);
1305 * Change more than one field in the in-core superblock structure at a time.
1307 * The fields and changes to those fields are specified in the array of
1308 * xfs_mod_sb structures passed in. Either all of the specified deltas
1309 * will be applied or none of them will. If any modified field dips below 0,
1310 * then all modifications will be backed out and EINVAL will be returned.
1312 * Note that this function may not be used for the superblock values that
1313 * are tracked with the in-memory per-cpu counters - a direct call to
1314 * xfs_icsb_modify_counters is required for these.
1317 xfs_mod_incore_sb_batch(
1318 struct xfs_mount *mp,
1327 * Loop through the array of mod structures and apply each individually.
1328 * If any fail, then back out all those which have already been applied.
1329 * Do all of this within the scope of the m_sb_lock so that all of the
1330 * changes will be atomic.
1332 spin_lock(&mp->m_sb_lock);
1333 for (msbp = msb; msbp < (msb + nmsb); msbp++) {
1334 ASSERT(msbp->msb_field < XFS_SBS_ICOUNT ||
1335 msbp->msb_field > XFS_SBS_FDBLOCKS);
1337 error = xfs_mod_incore_sb_unlocked(mp, msbp->msb_field,
1338 msbp->msb_delta, rsvd);
1342 spin_unlock(&mp->m_sb_lock);
1346 while (--msbp >= msb) {
1347 error = xfs_mod_incore_sb_unlocked(mp, msbp->msb_field,
1348 -msbp->msb_delta, rsvd);
1351 spin_unlock(&mp->m_sb_lock);
1356 * xfs_getsb() is called to obtain the buffer for the superblock.
1357 * The buffer is returned locked and read in from disk.
1358 * The buffer should be released with a call to xfs_brelse().
1360 * If the flags parameter is BUF_TRYLOCK, then we'll only return
1361 * the superblock buffer if it can be locked without sleeping.
1362 * If it can't then we'll return NULL.
1366 struct xfs_mount *mp,
1369 struct xfs_buf *bp = mp->m_sb_bp;
1371 if (!xfs_buf_trylock(bp)) {
1372 if (flags & XBF_TRYLOCK)
1378 ASSERT(XFS_BUF_ISDONE(bp));
1383 * Used to free the superblock along various error paths.
1387 struct xfs_mount *mp)
1389 struct xfs_buf *bp = mp->m_sb_bp;
1397 * Used to log changes to the superblock unit and width fields which could
1398 * be altered by the mount options, as well as any potential sb_features2
1399 * fixup. Only the first superblock is updated.
1409 ASSERT(fields & (XFS_SB_UNIT | XFS_SB_WIDTH | XFS_SB_UUID |
1410 XFS_SB_FEATURES2 | XFS_SB_BAD_FEATURES2 |
1411 XFS_SB_VERSIONNUM));
1413 tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1414 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_sb, 0, 0);
1416 xfs_trans_cancel(tp, 0);
1419 xfs_mod_sb(tp, fields);
1420 error = xfs_trans_commit(tp, 0);
1425 * If the underlying (data/log/rt) device is readonly, there are some
1426 * operations that cannot proceed.
1429 xfs_dev_is_read_only(
1430 struct xfs_mount *mp,
1433 if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1434 xfs_readonly_buftarg(mp->m_logdev_targp) ||
1435 (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1436 xfs_notice(mp, "%s required on read-only device.", message);
1437 xfs_notice(mp, "write access unavailable, cannot proceed.");
1443 #ifdef HAVE_PERCPU_SB
1445 * Per-cpu incore superblock counters
1447 * Simple concept, difficult implementation
1449 * Basically, replace the incore superblock counters with a distributed per cpu
1450 * counter for contended fields (e.g. free block count).
1452 * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1453 * hence needs to be accurately read when we are running low on space. Hence
1454 * there is a method to enable and disable the per-cpu counters based on how
1455 * much "stuff" is available in them.
1457 * Basically, a counter is enabled if there is enough free resource to justify
1458 * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1459 * ENOSPC), then we disable the counters to synchronise all callers and
1460 * re-distribute the available resources.
1462 * If, once we redistributed the available resources, we still get a failure,
1463 * we disable the per-cpu counter and go through the slow path.
1465 * The slow path is the current xfs_mod_incore_sb() function. This means that
1466 * when we disable a per-cpu counter, we need to drain its resources back to
1467 * the global superblock. We do this after disabling the counter to prevent
1468 * more threads from queueing up on the counter.
1470 * Essentially, this means that we still need a lock in the fast path to enable
1471 * synchronisation between the global counters and the per-cpu counters. This
1472 * is not a problem because the lock will be local to a CPU almost all the time
1473 * and have little contention except when we get to ENOSPC conditions.
1475 * Basically, this lock becomes a barrier that enables us to lock out the fast
1476 * path while we do things like enabling and disabling counters and
1477 * synchronising the counters.
1481 * 1. m_sb_lock before picking up per-cpu locks
1482 * 2. per-cpu locks always picked up via for_each_online_cpu() order
1483 * 3. accurate counter sync requires m_sb_lock + per cpu locks
1484 * 4. modifying per-cpu counters requires holding per-cpu lock
1485 * 5. modifying global counters requires holding m_sb_lock
1486 * 6. enabling or disabling a counter requires holding the m_sb_lock
1487 * and _none_ of the per-cpu locks.
1489 * Disabled counters are only ever re-enabled by a balance operation
1490 * that results in more free resources per CPU than a given threshold.
1491 * To ensure counters don't remain disabled, they are rebalanced when
1492 * the global resource goes above a higher threshold (i.e. some hysteresis
1493 * is present to prevent thrashing).
1496 #ifdef CONFIG_HOTPLUG_CPU
1498 * hot-plug CPU notifier support.
1500 * We need a notifier per filesystem as we need to be able to identify
1501 * the filesystem to balance the counters out. This is achieved by
1502 * having a notifier block embedded in the xfs_mount_t and doing pointer
1503 * magic to get the mount pointer from the notifier block address.
1506 xfs_icsb_cpu_notify(
1507 struct notifier_block *nfb,
1508 unsigned long action,
1511 xfs_icsb_cnts_t *cntp;
1514 mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1515 cntp = (xfs_icsb_cnts_t *)
1516 per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1518 case CPU_UP_PREPARE:
1519 case CPU_UP_PREPARE_FROZEN:
1520 /* Easy Case - initialize the area and locks, and
1521 * then rebalance when online does everything else for us. */
1522 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1525 case CPU_ONLINE_FROZEN:
1527 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1528 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1529 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1530 xfs_icsb_unlock(mp);
1533 case CPU_DEAD_FROZEN:
1534 /* Disable all the counters, then fold the dead cpu's
1535 * count into the total on the global superblock and
1536 * re-enable the counters. */
1538 spin_lock(&mp->m_sb_lock);
1539 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1540 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1541 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1543 mp->m_sb.sb_icount += cntp->icsb_icount;
1544 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1545 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1547 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1549 xfs_icsb_balance_counter_locked(mp, XFS_SBS_ICOUNT, 0);
1550 xfs_icsb_balance_counter_locked(mp, XFS_SBS_IFREE, 0);
1551 xfs_icsb_balance_counter_locked(mp, XFS_SBS_FDBLOCKS, 0);
1552 spin_unlock(&mp->m_sb_lock);
1553 xfs_icsb_unlock(mp);
1559 #endif /* CONFIG_HOTPLUG_CPU */
1562 xfs_icsb_init_counters(
1565 xfs_icsb_cnts_t *cntp;
1568 mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1569 if (mp->m_sb_cnts == NULL)
1572 for_each_online_cpu(i) {
1573 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1574 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1577 mutex_init(&mp->m_icsb_mutex);
1580 * start with all counters disabled so that the
1581 * initial balance kicks us off correctly
1583 mp->m_icsb_counters = -1;
1585 #ifdef CONFIG_HOTPLUG_CPU
1586 mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1587 mp->m_icsb_notifier.priority = 0;
1588 register_hotcpu_notifier(&mp->m_icsb_notifier);
1589 #endif /* CONFIG_HOTPLUG_CPU */
1595 xfs_icsb_reinit_counters(
1600 * start with all counters disabled so that the
1601 * initial balance kicks us off correctly
1603 mp->m_icsb_counters = -1;
1604 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1605 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1606 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1607 xfs_icsb_unlock(mp);
1611 xfs_icsb_destroy_counters(
1614 if (mp->m_sb_cnts) {
1615 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
1616 free_percpu(mp->m_sb_cnts);
1618 mutex_destroy(&mp->m_icsb_mutex);
1623 xfs_icsb_cnts_t *icsbp)
1625 while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1631 xfs_icsb_unlock_cntr(
1632 xfs_icsb_cnts_t *icsbp)
1634 clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1639 xfs_icsb_lock_all_counters(
1642 xfs_icsb_cnts_t *cntp;
1645 for_each_online_cpu(i) {
1646 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1647 xfs_icsb_lock_cntr(cntp);
1652 xfs_icsb_unlock_all_counters(
1655 xfs_icsb_cnts_t *cntp;
1658 for_each_online_cpu(i) {
1659 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1660 xfs_icsb_unlock_cntr(cntp);
1667 xfs_icsb_cnts_t *cnt,
1670 xfs_icsb_cnts_t *cntp;
1673 memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1675 if (!(flags & XFS_ICSB_LAZY_COUNT))
1676 xfs_icsb_lock_all_counters(mp);
1678 for_each_online_cpu(i) {
1679 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1680 cnt->icsb_icount += cntp->icsb_icount;
1681 cnt->icsb_ifree += cntp->icsb_ifree;
1682 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1685 if (!(flags & XFS_ICSB_LAZY_COUNT))
1686 xfs_icsb_unlock_all_counters(mp);
1690 xfs_icsb_counter_disabled(
1692 xfs_sb_field_t field)
1694 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1695 return test_bit(field, &mp->m_icsb_counters);
1699 xfs_icsb_disable_counter(
1701 xfs_sb_field_t field)
1703 xfs_icsb_cnts_t cnt;
1705 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1708 * If we are already disabled, then there is nothing to do
1709 * here. We check before locking all the counters to avoid
1710 * the expensive lock operation when being called in the
1711 * slow path and the counter is already disabled. This is
1712 * safe because the only time we set or clear this state is under
1715 if (xfs_icsb_counter_disabled(mp, field))
1718 xfs_icsb_lock_all_counters(mp);
1719 if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1720 /* drain back to superblock */
1722 xfs_icsb_count(mp, &cnt, XFS_ICSB_LAZY_COUNT);
1724 case XFS_SBS_ICOUNT:
1725 mp->m_sb.sb_icount = cnt.icsb_icount;
1728 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1730 case XFS_SBS_FDBLOCKS:
1731 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1738 xfs_icsb_unlock_all_counters(mp);
1742 xfs_icsb_enable_counter(
1744 xfs_sb_field_t field,
1748 xfs_icsb_cnts_t *cntp;
1751 ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1753 xfs_icsb_lock_all_counters(mp);
1754 for_each_online_cpu(i) {
1755 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1757 case XFS_SBS_ICOUNT:
1758 cntp->icsb_icount = count + resid;
1761 cntp->icsb_ifree = count + resid;
1763 case XFS_SBS_FDBLOCKS:
1764 cntp->icsb_fdblocks = count + resid;
1772 clear_bit(field, &mp->m_icsb_counters);
1773 xfs_icsb_unlock_all_counters(mp);
1777 xfs_icsb_sync_counters_locked(
1781 xfs_icsb_cnts_t cnt;
1783 xfs_icsb_count(mp, &cnt, flags);
1785 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1786 mp->m_sb.sb_icount = cnt.icsb_icount;
1787 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1788 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1789 if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1790 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1794 * Accurate update of per-cpu counters to incore superblock
1797 xfs_icsb_sync_counters(
1801 spin_lock(&mp->m_sb_lock);
1802 xfs_icsb_sync_counters_locked(mp, flags);
1803 spin_unlock(&mp->m_sb_lock);
1807 * Balance and enable/disable counters as necessary.
1809 * Thresholds for re-enabling counters are somewhat magic. inode counts are
1810 * chosen to be the same number as single on disk allocation chunk per CPU, and
1811 * free blocks is something far enough zero that we aren't going thrash when we
1812 * get near ENOSPC. We also need to supply a minimum we require per cpu to
1813 * prevent looping endlessly when xfs_alloc_space asks for more than will
1814 * be distributed to a single CPU but each CPU has enough blocks to be
1817 * Note that we can be called when counters are already disabled.
1818 * xfs_icsb_disable_counter() optimises the counter locking in this case to
1819 * prevent locking every per-cpu counter needlessly.
1822 #define XFS_ICSB_INO_CNTR_REENABLE (uint64_t)64
1823 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
1824 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
1826 xfs_icsb_balance_counter_locked(
1828 xfs_sb_field_t field,
1831 uint64_t count, resid;
1832 int weight = num_online_cpus();
1833 uint64_t min = (uint64_t)min_per_cpu;
1835 /* disable counter and sync counter */
1836 xfs_icsb_disable_counter(mp, field);
1838 /* update counters - first CPU gets residual*/
1840 case XFS_SBS_ICOUNT:
1841 count = mp->m_sb.sb_icount;
1842 resid = do_div(count, weight);
1843 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
1847 count = mp->m_sb.sb_ifree;
1848 resid = do_div(count, weight);
1849 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
1852 case XFS_SBS_FDBLOCKS:
1853 count = mp->m_sb.sb_fdblocks;
1854 resid = do_div(count, weight);
1855 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
1860 count = resid = 0; /* quiet, gcc */
1864 xfs_icsb_enable_counter(mp, field, count, resid);
1868 xfs_icsb_balance_counter(
1870 xfs_sb_field_t fields,
1873 spin_lock(&mp->m_sb_lock);
1874 xfs_icsb_balance_counter_locked(mp, fields, min_per_cpu);
1875 spin_unlock(&mp->m_sb_lock);
1879 xfs_icsb_modify_counters(
1881 xfs_sb_field_t field,
1885 xfs_icsb_cnts_t *icsbp;
1886 long long lcounter; /* long counter for 64 bit fields */
1892 icsbp = this_cpu_ptr(mp->m_sb_cnts);
1895 * if the counter is disabled, go to slow path
1897 if (unlikely(xfs_icsb_counter_disabled(mp, field)))
1899 xfs_icsb_lock_cntr(icsbp);
1900 if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
1901 xfs_icsb_unlock_cntr(icsbp);
1906 case XFS_SBS_ICOUNT:
1907 lcounter = icsbp->icsb_icount;
1909 if (unlikely(lcounter < 0))
1910 goto balance_counter;
1911 icsbp->icsb_icount = lcounter;
1915 lcounter = icsbp->icsb_ifree;
1917 if (unlikely(lcounter < 0))
1918 goto balance_counter;
1919 icsbp->icsb_ifree = lcounter;
1922 case XFS_SBS_FDBLOCKS:
1923 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
1925 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1927 if (unlikely(lcounter < 0))
1928 goto balance_counter;
1929 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1935 xfs_icsb_unlock_cntr(icsbp);
1943 * serialise with a mutex so we don't burn lots of cpu on
1944 * the superblock lock. We still need to hold the superblock
1945 * lock, however, when we modify the global structures.
1950 * Now running atomically.
1952 * If the counter is enabled, someone has beaten us to rebalancing.
1953 * Drop the lock and try again in the fast path....
1955 if (!(xfs_icsb_counter_disabled(mp, field))) {
1956 xfs_icsb_unlock(mp);
1961 * The counter is currently disabled. Because we are
1962 * running atomically here, we know a rebalance cannot
1963 * be in progress. Hence we can go straight to operating
1964 * on the global superblock. We do not call xfs_mod_incore_sb()
1965 * here even though we need to get the m_sb_lock. Doing so
1966 * will cause us to re-enter this function and deadlock.
1967 * Hence we get the m_sb_lock ourselves and then call
1968 * xfs_mod_incore_sb_unlocked() as the unlocked path operates
1969 * directly on the global counters.
1971 spin_lock(&mp->m_sb_lock);
1972 ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1973 spin_unlock(&mp->m_sb_lock);
1976 * Now that we've modified the global superblock, we
1977 * may be able to re-enable the distributed counters
1978 * (e.g. lots of space just got freed). After that
1982 xfs_icsb_balance_counter(mp, field, 0);
1983 xfs_icsb_unlock(mp);
1987 xfs_icsb_unlock_cntr(icsbp);
1991 * We may have multiple threads here if multiple per-cpu
1992 * counters run dry at the same time. This will mean we can
1993 * do more balances than strictly necessary but it is not
1994 * the common slowpath case.
1999 * running atomically.
2001 * This will leave the counter in the correct state for future
2002 * accesses. After the rebalance, we simply try again and our retry
2003 * will either succeed through the fast path or slow path without
2004 * another balance operation being required.
2006 xfs_icsb_balance_counter(mp, field, delta);
2007 xfs_icsb_unlock(mp);