Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_bit.h"
13#include "xfs_sb.h"
14#include "xfs_mount.h"
15#include "xfs_inode.h"
16#include "xfs_dir2.h"
17#include "xfs_ialloc.h"
18#include "xfs_alloc.h"
19#include "xfs_rtalloc.h"
20#include "xfs_bmap.h"
21#include "xfs_trans.h"
22#include "xfs_trans_priv.h"
23#include "xfs_log.h"
24#include "xfs_error.h"
25#include "xfs_quota.h"
26#include "xfs_fsops.h"
27#include "xfs_icache.h"
28#include "xfs_sysfs.h"
29#include "xfs_rmap_btree.h"
30#include "xfs_refcount_btree.h"
31#include "xfs_reflink.h"
32#include "xfs_extent_busy.h"
33#include "xfs_health.h"
34#include "xfs_trace.h"
35
36static DEFINE_MUTEX(xfs_uuid_table_mutex);
37static int xfs_uuid_table_size;
38static uuid_t *xfs_uuid_table;
39
40void
41xfs_uuid_table_free(void)
42{
43 if (xfs_uuid_table_size == 0)
44 return;
45 kmem_free(xfs_uuid_table);
46 xfs_uuid_table = NULL;
47 xfs_uuid_table_size = 0;
48}
49
50/*
51 * See if the UUID is unique among mounted XFS filesystems.
52 * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
53 */
54STATIC int
55xfs_uuid_mount(
56 struct xfs_mount *mp)
57{
58 uuid_t *uuid = &mp->m_sb.sb_uuid;
59 int hole, i;
60
61 /* Publish UUID in struct super_block */
62 uuid_copy(&mp->m_super->s_uuid, uuid);
63
64 if (mp->m_flags & XFS_MOUNT_NOUUID)
65 return 0;
66
67 if (uuid_is_null(uuid)) {
68 xfs_warn(mp, "Filesystem has null UUID - can't mount");
69 return -EINVAL;
70 }
71
72 mutex_lock(&xfs_uuid_table_mutex);
73 for (i = 0, hole = -1; i < xfs_uuid_table_size; i++) {
74 if (uuid_is_null(&xfs_uuid_table[i])) {
75 hole = i;
76 continue;
77 }
78 if (uuid_equal(uuid, &xfs_uuid_table[i]))
79 goto out_duplicate;
80 }
81
82 if (hole < 0) {
83 xfs_uuid_table = kmem_realloc(xfs_uuid_table,
84 (xfs_uuid_table_size + 1) * sizeof(*xfs_uuid_table),
85 0);
86 hole = xfs_uuid_table_size++;
87 }
88 xfs_uuid_table[hole] = *uuid;
89 mutex_unlock(&xfs_uuid_table_mutex);
90
91 return 0;
92
93 out_duplicate:
94 mutex_unlock(&xfs_uuid_table_mutex);
95 xfs_warn(mp, "Filesystem has duplicate UUID %pU - can't mount", uuid);
96 return -EINVAL;
97}
98
99STATIC void
100xfs_uuid_unmount(
101 struct xfs_mount *mp)
102{
103 uuid_t *uuid = &mp->m_sb.sb_uuid;
104 int i;
105
106 if (mp->m_flags & XFS_MOUNT_NOUUID)
107 return;
108
109 mutex_lock(&xfs_uuid_table_mutex);
110 for (i = 0; i < xfs_uuid_table_size; i++) {
111 if (uuid_is_null(&xfs_uuid_table[i]))
112 continue;
113 if (!uuid_equal(uuid, &xfs_uuid_table[i]))
114 continue;
115 memset(&xfs_uuid_table[i], 0, sizeof(uuid_t));
116 break;
117 }
118 ASSERT(i < xfs_uuid_table_size);
119 mutex_unlock(&xfs_uuid_table_mutex);
120}
121
122
123STATIC void
124__xfs_free_perag(
125 struct rcu_head *head)
126{
127 struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
128
129 ASSERT(atomic_read(&pag->pag_ref) == 0);
130 kmem_free(pag);
131}
132
133/*
134 * Free up the per-ag resources associated with the mount structure.
135 */
136STATIC void
137xfs_free_perag(
138 xfs_mount_t *mp)
139{
140 xfs_agnumber_t agno;
141 struct xfs_perag *pag;
142
143 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
144 spin_lock(&mp->m_perag_lock);
145 pag = radix_tree_delete(&mp->m_perag_tree, agno);
146 spin_unlock(&mp->m_perag_lock);
147 ASSERT(pag);
148 ASSERT(atomic_read(&pag->pag_ref) == 0);
149 xfs_iunlink_destroy(pag);
150 xfs_buf_hash_destroy(pag);
151 call_rcu(&pag->rcu_head, __xfs_free_perag);
152 }
153}
154
155/*
156 * Check size of device based on the (data/realtime) block count.
157 * Note: this check is used by the growfs code as well as mount.
158 */
159int
160xfs_sb_validate_fsb_count(
161 xfs_sb_t *sbp,
162 uint64_t nblocks)
163{
164 ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
165 ASSERT(sbp->sb_blocklog >= BBSHIFT);
166
167 /* Limited by ULONG_MAX of page cache index */
168 if (nblocks >> (PAGE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
169 return -EFBIG;
170 return 0;
171}
172
173int
174xfs_initialize_perag(
175 xfs_mount_t *mp,
176 xfs_agnumber_t agcount,
177 xfs_agnumber_t *maxagi)
178{
179 xfs_agnumber_t index;
180 xfs_agnumber_t first_initialised = NULLAGNUMBER;
181 xfs_perag_t *pag;
182 int error = -ENOMEM;
183
184 /*
185 * Walk the current per-ag tree so we don't try to initialise AGs
186 * that already exist (growfs case). Allocate and insert all the
187 * AGs we don't find ready for initialisation.
188 */
189 for (index = 0; index < agcount; index++) {
190 pag = xfs_perag_get(mp, index);
191 if (pag) {
192 xfs_perag_put(pag);
193 continue;
194 }
195
196 pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
197 if (!pag)
198 goto out_unwind_new_pags;
199 pag->pag_agno = index;
200 pag->pag_mount = mp;
201 spin_lock_init(&pag->pag_ici_lock);
202 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
203 if (xfs_buf_hash_init(pag))
204 goto out_free_pag;
205 init_waitqueue_head(&pag->pagb_wait);
206 spin_lock_init(&pag->pagb_lock);
207 pag->pagb_count = 0;
208 pag->pagb_tree = RB_ROOT;
209
210 if (radix_tree_preload(GFP_NOFS))
211 goto out_hash_destroy;
212
213 spin_lock(&mp->m_perag_lock);
214 if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
215 WARN_ON_ONCE(1);
216 spin_unlock(&mp->m_perag_lock);
217 radix_tree_preload_end();
218 error = -EEXIST;
219 goto out_hash_destroy;
220 }
221 spin_unlock(&mp->m_perag_lock);
222 radix_tree_preload_end();
223 /* first new pag is fully initialized */
224 if (first_initialised == NULLAGNUMBER)
225 first_initialised = index;
226 error = xfs_iunlink_init(pag);
227 if (error)
228 goto out_hash_destroy;
229 spin_lock_init(&pag->pag_state_lock);
230 }
231
232 index = xfs_set_inode_alloc(mp, agcount);
233
234 if (maxagi)
235 *maxagi = index;
236
237 mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
238 return 0;
239
240out_hash_destroy:
241 xfs_buf_hash_destroy(pag);
242out_free_pag:
243 kmem_free(pag);
244out_unwind_new_pags:
245 /* unwind any prior newly initialized pags */
246 for (index = first_initialised; index < agcount; index++) {
247 pag = radix_tree_delete(&mp->m_perag_tree, index);
248 if (!pag)
249 break;
250 xfs_buf_hash_destroy(pag);
251 xfs_iunlink_destroy(pag);
252 kmem_free(pag);
253 }
254 return error;
255}
256
257/*
258 * xfs_readsb
259 *
260 * Does the initial read of the superblock.
261 */
262int
263xfs_readsb(
264 struct xfs_mount *mp,
265 int flags)
266{
267 unsigned int sector_size;
268 struct xfs_buf *bp;
269 struct xfs_sb *sbp = &mp->m_sb;
270 int error;
271 int loud = !(flags & XFS_MFSI_QUIET);
272 const struct xfs_buf_ops *buf_ops;
273
274 ASSERT(mp->m_sb_bp == NULL);
275 ASSERT(mp->m_ddev_targp != NULL);
276
277 /*
278 * For the initial read, we must guess at the sector
279 * size based on the block device. It's enough to
280 * get the sb_sectsize out of the superblock and
281 * then reread with the proper length.
282 * We don't verify it yet, because it may not be complete.
283 */
284 sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
285 buf_ops = NULL;
286
287 /*
288 * Allocate a (locked) buffer to hold the superblock. This will be kept
289 * around at all times to optimize access to the superblock. Therefore,
290 * set XBF_NO_IOACCT to make sure it doesn't hold the buftarg count
291 * elevated.
292 */
293reread:
294 error = xfs_buf_read_uncached(mp->m_ddev_targp, XFS_SB_DADDR,
295 BTOBB(sector_size), XBF_NO_IOACCT, &bp,
296 buf_ops);
297 if (error) {
298 if (loud)
299 xfs_warn(mp, "SB validate failed with error %d.", error);
300 /* bad CRC means corrupted metadata */
301 if (error == -EFSBADCRC)
302 error = -EFSCORRUPTED;
303 return error;
304 }
305
306 /*
307 * Initialize the mount structure from the superblock.
308 */
309 xfs_sb_from_disk(sbp, bp->b_addr);
310
311 /*
312 * If we haven't validated the superblock, do so now before we try
313 * to check the sector size and reread the superblock appropriately.
314 */
315 if (sbp->sb_magicnum != XFS_SB_MAGIC) {
316 if (loud)
317 xfs_warn(mp, "Invalid superblock magic number");
318 error = -EINVAL;
319 goto release_buf;
320 }
321
322 /*
323 * We must be able to do sector-sized and sector-aligned IO.
324 */
325 if (sector_size > sbp->sb_sectsize) {
326 if (loud)
327 xfs_warn(mp, "device supports %u byte sectors (not %u)",
328 sector_size, sbp->sb_sectsize);
329 error = -ENOSYS;
330 goto release_buf;
331 }
332
333 if (buf_ops == NULL) {
334 /*
335 * Re-read the superblock so the buffer is correctly sized,
336 * and properly verified.
337 */
338 xfs_buf_relse(bp);
339 sector_size = sbp->sb_sectsize;
340 buf_ops = loud ? &xfs_sb_buf_ops : &xfs_sb_quiet_buf_ops;
341 goto reread;
342 }
343
344 xfs_reinit_percpu_counters(mp);
345
346 /* no need to be quiet anymore, so reset the buf ops */
347 bp->b_ops = &xfs_sb_buf_ops;
348
349 mp->m_sb_bp = bp;
350 xfs_buf_unlock(bp);
351 return 0;
352
353release_buf:
354 xfs_buf_relse(bp);
355 return error;
356}
357
358/*
359 * If the sunit/swidth change would move the precomputed root inode value, we
360 * must reject the ondisk change because repair will stumble over that.
361 * However, we allow the mount to proceed because we never rejected this
362 * combination before. Returns true to update the sb, false otherwise.
363 */
364static inline int
365xfs_check_new_dalign(
366 struct xfs_mount *mp,
367 int new_dalign,
368 bool *update_sb)
369{
370 struct xfs_sb *sbp = &mp->m_sb;
371 xfs_ino_t calc_ino;
372
373 calc_ino = xfs_ialloc_calc_rootino(mp, new_dalign);
374 trace_xfs_check_new_dalign(mp, new_dalign, calc_ino);
375
376 if (sbp->sb_rootino == calc_ino) {
377 *update_sb = true;
378 return 0;
379 }
380
381 xfs_warn(mp,
382"Cannot change stripe alignment; would require moving root inode.");
383
384 /*
385 * XXX: Next time we add a new incompat feature, this should start
386 * returning -EINVAL to fail the mount. Until then, spit out a warning
387 * that we're ignoring the administrator's instructions.
388 */
389 xfs_warn(mp, "Skipping superblock stripe alignment update.");
390 *update_sb = false;
391 return 0;
392}
393
394/*
395 * If we were provided with new sunit/swidth values as mount options, make sure
396 * that they pass basic alignment and superblock feature checks, and convert
397 * them into the same units (FSB) that everything else expects. This step
398 * /must/ be done before computing the inode geometry.
399 */
400STATIC int
401xfs_validate_new_dalign(
402 struct xfs_mount *mp)
403{
404 if (mp->m_dalign == 0)
405 return 0;
406
407 /*
408 * If stripe unit and stripe width are not multiples
409 * of the fs blocksize turn off alignment.
410 */
411 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
412 (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
413 xfs_warn(mp,
414 "alignment check failed: sunit/swidth vs. blocksize(%d)",
415 mp->m_sb.sb_blocksize);
416 return -EINVAL;
417 } else {
418 /*
419 * Convert the stripe unit and width to FSBs.
420 */
421 mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
422 if (mp->m_dalign && (mp->m_sb.sb_agblocks % mp->m_dalign)) {
423 xfs_warn(mp,
424 "alignment check failed: sunit/swidth vs. agsize(%d)",
425 mp->m_sb.sb_agblocks);
426 return -EINVAL;
427 } else if (mp->m_dalign) {
428 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
429 } else {
430 xfs_warn(mp,
431 "alignment check failed: sunit(%d) less than bsize(%d)",
432 mp->m_dalign, mp->m_sb.sb_blocksize);
433 return -EINVAL;
434 }
435 }
436
437 if (!xfs_sb_version_hasdalign(&mp->m_sb)) {
438 xfs_warn(mp,
439"cannot change alignment: superblock does not support data alignment");
440 return -EINVAL;
441 }
442
443 return 0;
444}
445
446/* Update alignment values based on mount options and sb values. */
447STATIC int
448xfs_update_alignment(
449 struct xfs_mount *mp)
450{
451 struct xfs_sb *sbp = &mp->m_sb;
452
453 if (mp->m_dalign) {
454 bool update_sb;
455 int error;
456
457 if (sbp->sb_unit == mp->m_dalign &&
458 sbp->sb_width == mp->m_swidth)
459 return 0;
460
461 error = xfs_check_new_dalign(mp, mp->m_dalign, &update_sb);
462 if (error || !update_sb)
463 return error;
464
465 sbp->sb_unit = mp->m_dalign;
466 sbp->sb_width = mp->m_swidth;
467 mp->m_update_sb = true;
468 } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
469 xfs_sb_version_hasdalign(&mp->m_sb)) {
470 mp->m_dalign = sbp->sb_unit;
471 mp->m_swidth = sbp->sb_width;
472 }
473
474 return 0;
475}
476
477/*
478 * precalculate the low space thresholds for dynamic speculative preallocation.
479 */
480void
481xfs_set_low_space_thresholds(
482 struct xfs_mount *mp)
483{
484 int i;
485
486 for (i = 0; i < XFS_LOWSP_MAX; i++) {
487 uint64_t space = mp->m_sb.sb_dblocks;
488
489 do_div(space, 100);
490 mp->m_low_space[i] = space * (i + 1);
491 }
492}
493
494/*
495 * Check that the data (and log if separate) is an ok size.
496 */
497STATIC int
498xfs_check_sizes(
499 struct xfs_mount *mp)
500{
501 struct xfs_buf *bp;
502 xfs_daddr_t d;
503 int error;
504
505 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
506 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
507 xfs_warn(mp, "filesystem size mismatch detected");
508 return -EFBIG;
509 }
510 error = xfs_buf_read_uncached(mp->m_ddev_targp,
511 d - XFS_FSS_TO_BB(mp, 1),
512 XFS_FSS_TO_BB(mp, 1), 0, &bp, NULL);
513 if (error) {
514 xfs_warn(mp, "last sector read failed");
515 return error;
516 }
517 xfs_buf_relse(bp);
518
519 if (mp->m_logdev_targp == mp->m_ddev_targp)
520 return 0;
521
522 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
523 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
524 xfs_warn(mp, "log size mismatch detected");
525 return -EFBIG;
526 }
527 error = xfs_buf_read_uncached(mp->m_logdev_targp,
528 d - XFS_FSB_TO_BB(mp, 1),
529 XFS_FSB_TO_BB(mp, 1), 0, &bp, NULL);
530 if (error) {
531 xfs_warn(mp, "log device read failed");
532 return error;
533 }
534 xfs_buf_relse(bp);
535 return 0;
536}
537
538/*
539 * Clear the quotaflags in memory and in the superblock.
540 */
541int
542xfs_mount_reset_sbqflags(
543 struct xfs_mount *mp)
544{
545 mp->m_qflags = 0;
546
547 /* It is OK to look at sb_qflags in the mount path without m_sb_lock. */
548 if (mp->m_sb.sb_qflags == 0)
549 return 0;
550 spin_lock(&mp->m_sb_lock);
551 mp->m_sb.sb_qflags = 0;
552 spin_unlock(&mp->m_sb_lock);
553
554 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
555 return 0;
556
557 return xfs_sync_sb(mp, false);
558}
559
560uint64_t
561xfs_default_resblks(xfs_mount_t *mp)
562{
563 uint64_t resblks;
564
565 /*
566 * We default to 5% or 8192 fsbs of space reserved, whichever is
567 * smaller. This is intended to cover concurrent allocation
568 * transactions when we initially hit enospc. These each require a 4
569 * block reservation. Hence by default we cover roughly 2000 concurrent
570 * allocation reservations.
571 */
572 resblks = mp->m_sb.sb_dblocks;
573 do_div(resblks, 20);
574 resblks = min_t(uint64_t, resblks, 8192);
575 return resblks;
576}
577
578/* Ensure the summary counts are correct. */
579STATIC int
580xfs_check_summary_counts(
581 struct xfs_mount *mp)
582{
583 /*
584 * The AG0 superblock verifier rejects in-progress filesystems,
585 * so we should never see the flag set this far into mounting.
586 */
587 if (mp->m_sb.sb_inprogress) {
588 xfs_err(mp, "sb_inprogress set after log recovery??");
589 WARN_ON(1);
590 return -EFSCORRUPTED;
591 }
592
593 /*
594 * Now the log is mounted, we know if it was an unclean shutdown or
595 * not. If it was, with the first phase of recovery has completed, we
596 * have consistent AG blocks on disk. We have not recovered EFIs yet,
597 * but they are recovered transactionally in the second recovery phase
598 * later.
599 *
600 * If the log was clean when we mounted, we can check the summary
601 * counters. If any of them are obviously incorrect, we can recompute
602 * them from the AGF headers in the next step.
603 */
604 if (XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
605 (mp->m_sb.sb_fdblocks > mp->m_sb.sb_dblocks ||
606 !xfs_verify_icount(mp, mp->m_sb.sb_icount) ||
607 mp->m_sb.sb_ifree > mp->m_sb.sb_icount))
608 xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
609
610 /*
611 * We can safely re-initialise incore superblock counters from the
612 * per-ag data. These may not be correct if the filesystem was not
613 * cleanly unmounted, so we waited for recovery to finish before doing
614 * this.
615 *
616 * If the filesystem was cleanly unmounted or the previous check did
617 * not flag anything weird, then we can trust the values in the
618 * superblock to be correct and we don't need to do anything here.
619 * Otherwise, recalculate the summary counters.
620 */
621 if ((!xfs_sb_version_haslazysbcount(&mp->m_sb) ||
622 XFS_LAST_UNMOUNT_WAS_CLEAN(mp)) &&
623 !xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS))
624 return 0;
625
626 return xfs_initialize_perag_data(mp, mp->m_sb.sb_agcount);
627}
628
629/*
630 * This function does the following on an initial mount of a file system:
631 * - reads the superblock from disk and init the mount struct
632 * - if we're a 32-bit kernel, do a size check on the superblock
633 * so we don't mount terabyte filesystems
634 * - init mount struct realtime fields
635 * - allocate inode hash table for fs
636 * - init directory manager
637 * - perform recovery and init the log manager
638 */
639int
640xfs_mountfs(
641 struct xfs_mount *mp)
642{
643 struct xfs_sb *sbp = &(mp->m_sb);
644 struct xfs_inode *rip;
645 struct xfs_ino_geometry *igeo = M_IGEO(mp);
646 uint64_t resblks;
647 uint quotamount = 0;
648 uint quotaflags = 0;
649 int error = 0;
650
651 xfs_sb_mount_common(mp, sbp);
652
653 /*
654 * Check for a mismatched features2 values. Older kernels read & wrote
655 * into the wrong sb offset for sb_features2 on some platforms due to
656 * xfs_sb_t not being 64bit size aligned when sb_features2 was added,
657 * which made older superblock reading/writing routines swap it as a
658 * 64-bit value.
659 *
660 * For backwards compatibility, we make both slots equal.
661 *
662 * If we detect a mismatched field, we OR the set bits into the existing
663 * features2 field in case it has already been modified; we don't want
664 * to lose any features. We then update the bad location with the ORed
665 * value so that older kernels will see any features2 flags. The
666 * superblock writeback code ensures the new sb_features2 is copied to
667 * sb_bad_features2 before it is logged or written to disk.
668 */
669 if (xfs_sb_has_mismatched_features2(sbp)) {
670 xfs_warn(mp, "correcting sb_features alignment problem");
671 sbp->sb_features2 |= sbp->sb_bad_features2;
672 mp->m_update_sb = true;
673
674 /*
675 * Re-check for ATTR2 in case it was found in bad_features2
676 * slot.
677 */
678 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
679 !(mp->m_flags & XFS_MOUNT_NOATTR2))
680 mp->m_flags |= XFS_MOUNT_ATTR2;
681 }
682
683 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
684 (mp->m_flags & XFS_MOUNT_NOATTR2)) {
685 xfs_sb_version_removeattr2(&mp->m_sb);
686 mp->m_update_sb = true;
687
688 /* update sb_versionnum for the clearing of the morebits */
689 if (!sbp->sb_features2)
690 mp->m_update_sb = true;
691 }
692
693 /* always use v2 inodes by default now */
694 if (!(mp->m_sb.sb_versionnum & XFS_SB_VERSION_NLINKBIT)) {
695 mp->m_sb.sb_versionnum |= XFS_SB_VERSION_NLINKBIT;
696 mp->m_update_sb = true;
697 }
698
699 /*
700 * If we were given new sunit/swidth options, do some basic validation
701 * checks and convert the incore dalign and swidth values to the
702 * same units (FSB) that everything else uses. This /must/ happen
703 * before computing the inode geometry.
704 */
705 error = xfs_validate_new_dalign(mp);
706 if (error)
707 goto out;
708
709 xfs_alloc_compute_maxlevels(mp);
710 xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
711 xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
712 xfs_ialloc_setup_geometry(mp);
713 xfs_rmapbt_compute_maxlevels(mp);
714 xfs_refcountbt_compute_maxlevels(mp);
715
716 /*
717 * Check if sb_agblocks is aligned at stripe boundary. If sb_agblocks
718 * is NOT aligned turn off m_dalign since allocator alignment is within
719 * an ag, therefore ag has to be aligned at stripe boundary. Note that
720 * we must compute the free space and rmap btree geometry before doing
721 * this.
722 */
723 error = xfs_update_alignment(mp);
724 if (error)
725 goto out;
726
727 /* enable fail_at_unmount as default */
728 mp->m_fail_unmount = true;
729
730 error = xfs_sysfs_init(&mp->m_kobj, &xfs_mp_ktype,
731 NULL, mp->m_super->s_id);
732 if (error)
733 goto out;
734
735 error = xfs_sysfs_init(&mp->m_stats.xs_kobj, &xfs_stats_ktype,
736 &mp->m_kobj, "stats");
737 if (error)
738 goto out_remove_sysfs;
739
740 error = xfs_error_sysfs_init(mp);
741 if (error)
742 goto out_del_stats;
743
744 error = xfs_errortag_init(mp);
745 if (error)
746 goto out_remove_error_sysfs;
747
748 error = xfs_uuid_mount(mp);
749 if (error)
750 goto out_remove_errortag;
751
752 /*
753 * Update the preferred write size based on the information from the
754 * on-disk superblock.
755 */
756 mp->m_allocsize_log =
757 max_t(uint32_t, sbp->sb_blocklog, mp->m_allocsize_log);
758 mp->m_allocsize_blocks = 1U << (mp->m_allocsize_log - sbp->sb_blocklog);
759
760 /* set the low space thresholds for dynamic preallocation */
761 xfs_set_low_space_thresholds(mp);
762
763 /*
764 * If enabled, sparse inode chunk alignment is expected to match the
765 * cluster size. Full inode chunk alignment must match the chunk size,
766 * but that is checked on sb read verification...
767 */
768 if (xfs_sb_version_hassparseinodes(&mp->m_sb) &&
769 mp->m_sb.sb_spino_align !=
770 XFS_B_TO_FSBT(mp, igeo->inode_cluster_size_raw)) {
771 xfs_warn(mp,
772 "Sparse inode block alignment (%u) must match cluster size (%llu).",
773 mp->m_sb.sb_spino_align,
774 XFS_B_TO_FSBT(mp, igeo->inode_cluster_size_raw));
775 error = -EINVAL;
776 goto out_remove_uuid;
777 }
778
779 /*
780 * Check that the data (and log if separate) is an ok size.
781 */
782 error = xfs_check_sizes(mp);
783 if (error)
784 goto out_remove_uuid;
785
786 /*
787 * Initialize realtime fields in the mount structure
788 */
789 error = xfs_rtmount_init(mp);
790 if (error) {
791 xfs_warn(mp, "RT mount failed");
792 goto out_remove_uuid;
793 }
794
795 /*
796 * Copies the low order bits of the timestamp and the randomly
797 * set "sequence" number out of a UUID.
798 */
799 mp->m_fixedfsid[0] =
800 (get_unaligned_be16(&sbp->sb_uuid.b[8]) << 16) |
801 get_unaligned_be16(&sbp->sb_uuid.b[4]);
802 mp->m_fixedfsid[1] = get_unaligned_be32(&sbp->sb_uuid.b[0]);
803
804 error = xfs_da_mount(mp);
805 if (error) {
806 xfs_warn(mp, "Failed dir/attr init: %d", error);
807 goto out_remove_uuid;
808 }
809
810 /*
811 * Initialize the precomputed transaction reservations values.
812 */
813 xfs_trans_init(mp);
814
815 /*
816 * Allocate and initialize the per-ag data.
817 */
818 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
819 if (error) {
820 xfs_warn(mp, "Failed per-ag init: %d", error);
821 goto out_free_dir;
822 }
823
824 if (XFS_IS_CORRUPT(mp, !sbp->sb_logblocks)) {
825 xfs_warn(mp, "no log defined");
826 error = -EFSCORRUPTED;
827 goto out_free_perag;
828 }
829
830 /*
831 * Log's mount-time initialization. The first part of recovery can place
832 * some items on the AIL, to be handled when recovery is finished or
833 * cancelled.
834 */
835 error = xfs_log_mount(mp, mp->m_logdev_targp,
836 XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
837 XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
838 if (error) {
839 xfs_warn(mp, "log mount failed");
840 goto out_fail_wait;
841 }
842
843 /* Make sure the summary counts are ok. */
844 error = xfs_check_summary_counts(mp);
845 if (error)
846 goto out_log_dealloc;
847
848 /*
849 * Get and sanity-check the root inode.
850 * Save the pointer to it in the mount structure.
851 */
852 error = xfs_iget(mp, NULL, sbp->sb_rootino, XFS_IGET_UNTRUSTED,
853 XFS_ILOCK_EXCL, &rip);
854 if (error) {
855 xfs_warn(mp,
856 "Failed to read root inode 0x%llx, error %d",
857 sbp->sb_rootino, -error);
858 goto out_log_dealloc;
859 }
860
861 ASSERT(rip != NULL);
862
863 if (XFS_IS_CORRUPT(mp, !S_ISDIR(VFS_I(rip)->i_mode))) {
864 xfs_warn(mp, "corrupted root inode %llu: not a directory",
865 (unsigned long long)rip->i_ino);
866 xfs_iunlock(rip, XFS_ILOCK_EXCL);
867 error = -EFSCORRUPTED;
868 goto out_rele_rip;
869 }
870 mp->m_rootip = rip; /* save it */
871
872 xfs_iunlock(rip, XFS_ILOCK_EXCL);
873
874 /*
875 * Initialize realtime inode pointers in the mount structure
876 */
877 error = xfs_rtmount_inodes(mp);
878 if (error) {
879 /*
880 * Free up the root inode.
881 */
882 xfs_warn(mp, "failed to read RT inodes");
883 goto out_rele_rip;
884 }
885
886 /*
887 * If this is a read-only mount defer the superblock updates until
888 * the next remount into writeable mode. Otherwise we would never
889 * perform the update e.g. for the root filesystem.
890 */
891 if (mp->m_update_sb && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
892 error = xfs_sync_sb(mp, false);
893 if (error) {
894 xfs_warn(mp, "failed to write sb changes");
895 goto out_rtunmount;
896 }
897 }
898
899 /*
900 * Initialise the XFS quota management subsystem for this mount
901 */
902 if (XFS_IS_QUOTA_RUNNING(mp)) {
903 error = xfs_qm_newmount(mp, "amount, "aflags);
904 if (error)
905 goto out_rtunmount;
906 } else {
907 ASSERT(!XFS_IS_QUOTA_ON(mp));
908
909 /*
910 * If a file system had quotas running earlier, but decided to
911 * mount without -o uquota/pquota/gquota options, revoke the
912 * quotachecked license.
913 */
914 if (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_ACCT) {
915 xfs_notice(mp, "resetting quota flags");
916 error = xfs_mount_reset_sbqflags(mp);
917 if (error)
918 goto out_rtunmount;
919 }
920 }
921
922 /*
923 * Finish recovering the file system. This part needed to be delayed
924 * until after the root and real-time bitmap inodes were consistently
925 * read in.
926 */
927 error = xfs_log_mount_finish(mp);
928 if (error) {
929 xfs_warn(mp, "log mount finish failed");
930 goto out_rtunmount;
931 }
932
933 /*
934 * Now the log is fully replayed, we can transition to full read-only
935 * mode for read-only mounts. This will sync all the metadata and clean
936 * the log so that the recovery we just performed does not have to be
937 * replayed again on the next mount.
938 *
939 * We use the same quiesce mechanism as the rw->ro remount, as they are
940 * semantically identical operations.
941 */
942 if ((mp->m_flags & (XFS_MOUNT_RDONLY|XFS_MOUNT_NORECOVERY)) ==
943 XFS_MOUNT_RDONLY) {
944 xfs_quiesce_attr(mp);
945 }
946
947 /*
948 * Complete the quota initialisation, post-log-replay component.
949 */
950 if (quotamount) {
951 ASSERT(mp->m_qflags == 0);
952 mp->m_qflags = quotaflags;
953
954 xfs_qm_mount_quotas(mp);
955 }
956
957 /*
958 * Now we are mounted, reserve a small amount of unused space for
959 * privileged transactions. This is needed so that transaction
960 * space required for critical operations can dip into this pool
961 * when at ENOSPC. This is needed for operations like create with
962 * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
963 * are not allowed to use this reserved space.
964 *
965 * This may drive us straight to ENOSPC on mount, but that implies
966 * we were already there on the last unmount. Warn if this occurs.
967 */
968 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
969 resblks = xfs_default_resblks(mp);
970 error = xfs_reserve_blocks(mp, &resblks, NULL);
971 if (error)
972 xfs_warn(mp,
973 "Unable to allocate reserve blocks. Continuing without reserve pool.");
974
975 /* Recover any CoW blocks that never got remapped. */
976 error = xfs_reflink_recover_cow(mp);
977 if (error) {
978 xfs_err(mp,
979 "Error %d recovering leftover CoW allocations.", error);
980 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
981 goto out_quota;
982 }
983
984 /* Reserve AG blocks for future btree expansion. */
985 error = xfs_fs_reserve_ag_blocks(mp);
986 if (error && error != -ENOSPC)
987 goto out_agresv;
988 }
989
990 return 0;
991
992 out_agresv:
993 xfs_fs_unreserve_ag_blocks(mp);
994 out_quota:
995 xfs_qm_unmount_quotas(mp);
996 out_rtunmount:
997 xfs_rtunmount_inodes(mp);
998 out_rele_rip:
999 xfs_irele(rip);
1000 /* Clean out dquots that might be in memory after quotacheck. */
1001 xfs_qm_unmount(mp);
1002 /*
1003 * Cancel all delayed reclaim work and reclaim the inodes directly.
1004 * We have to do this /after/ rtunmount and qm_unmount because those
1005 * two will have scheduled delayed reclaim for the rt/quota inodes.
1006 *
1007 * This is slightly different from the unmountfs call sequence
1008 * because we could be tearing down a partially set up mount. In
1009 * particular, if log_mount_finish fails we bail out without calling
1010 * qm_unmount_quotas and therefore rely on qm_unmount to release the
1011 * quota inodes.
1012 */
1013 cancel_delayed_work_sync(&mp->m_reclaim_work);
1014 xfs_reclaim_inodes(mp);
1015 xfs_health_unmount(mp);
1016 out_log_dealloc:
1017 mp->m_flags |= XFS_MOUNT_UNMOUNTING;
1018 xfs_log_mount_cancel(mp);
1019 out_fail_wait:
1020 if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1021 xfs_wait_buftarg(mp->m_logdev_targp);
1022 xfs_wait_buftarg(mp->m_ddev_targp);
1023 out_free_perag:
1024 xfs_free_perag(mp);
1025 out_free_dir:
1026 xfs_da_unmount(mp);
1027 out_remove_uuid:
1028 xfs_uuid_unmount(mp);
1029 out_remove_errortag:
1030 xfs_errortag_del(mp);
1031 out_remove_error_sysfs:
1032 xfs_error_sysfs_del(mp);
1033 out_del_stats:
1034 xfs_sysfs_del(&mp->m_stats.xs_kobj);
1035 out_remove_sysfs:
1036 xfs_sysfs_del(&mp->m_kobj);
1037 out:
1038 return error;
1039}
1040
1041/*
1042 * This flushes out the inodes,dquots and the superblock, unmounts the
1043 * log and makes sure that incore structures are freed.
1044 */
1045void
1046xfs_unmountfs(
1047 struct xfs_mount *mp)
1048{
1049 uint64_t resblks;
1050 int error;
1051
1052 xfs_stop_block_reaping(mp);
1053 xfs_fs_unreserve_ag_blocks(mp);
1054 xfs_qm_unmount_quotas(mp);
1055 xfs_rtunmount_inodes(mp);
1056 xfs_irele(mp->m_rootip);
1057
1058 /*
1059 * We can potentially deadlock here if we have an inode cluster
1060 * that has been freed has its buffer still pinned in memory because
1061 * the transaction is still sitting in a iclog. The stale inodes
1062 * on that buffer will have their flush locks held until the
1063 * transaction hits the disk and the callbacks run. the inode
1064 * flush takes the flush lock unconditionally and with nothing to
1065 * push out the iclog we will never get that unlocked. hence we
1066 * need to force the log first.
1067 */
1068 xfs_log_force(mp, XFS_LOG_SYNC);
1069
1070 /*
1071 * Wait for all busy extents to be freed, including completion of
1072 * any discard operation.
1073 */
1074 xfs_extent_busy_wait_all(mp);
1075 flush_workqueue(xfs_discard_wq);
1076
1077 /*
1078 * We now need to tell the world we are unmounting. This will allow
1079 * us to detect that the filesystem is going away and we should error
1080 * out anything that we have been retrying in the background. This will
1081 * prevent neverending retries in AIL pushing from hanging the unmount.
1082 */
1083 mp->m_flags |= XFS_MOUNT_UNMOUNTING;
1084
1085 /*
1086 * Flush all pending changes from the AIL.
1087 */
1088 xfs_ail_push_all_sync(mp->m_ail);
1089
1090 /*
1091 * Reclaim all inodes. At this point there should be no dirty inodes and
1092 * none should be pinned or locked. Stop background inode reclaim here
1093 * if it is still running.
1094 */
1095 cancel_delayed_work_sync(&mp->m_reclaim_work);
1096 xfs_reclaim_inodes(mp);
1097 xfs_health_unmount(mp);
1098
1099 xfs_qm_unmount(mp);
1100
1101 /*
1102 * Unreserve any blocks we have so that when we unmount we don't account
1103 * the reserved free space as used. This is really only necessary for
1104 * lazy superblock counting because it trusts the incore superblock
1105 * counters to be absolutely correct on clean unmount.
1106 *
1107 * We don't bother correcting this elsewhere for lazy superblock
1108 * counting because on mount of an unclean filesystem we reconstruct the
1109 * correct counter value and this is irrelevant.
1110 *
1111 * For non-lazy counter filesystems, this doesn't matter at all because
1112 * we only every apply deltas to the superblock and hence the incore
1113 * value does not matter....
1114 */
1115 resblks = 0;
1116 error = xfs_reserve_blocks(mp, &resblks, NULL);
1117 if (error)
1118 xfs_warn(mp, "Unable to free reserved block pool. "
1119 "Freespace may not be correct on next mount.");
1120
1121 error = xfs_log_sbcount(mp);
1122 if (error)
1123 xfs_warn(mp, "Unable to update superblock counters. "
1124 "Freespace may not be correct on next mount.");
1125
1126
1127 xfs_log_unmount(mp);
1128 xfs_da_unmount(mp);
1129 xfs_uuid_unmount(mp);
1130
1131#if defined(DEBUG)
1132 xfs_errortag_clearall(mp);
1133#endif
1134 xfs_free_perag(mp);
1135
1136 xfs_errortag_del(mp);
1137 xfs_error_sysfs_del(mp);
1138 xfs_sysfs_del(&mp->m_stats.xs_kobj);
1139 xfs_sysfs_del(&mp->m_kobj);
1140}
1141
1142/*
1143 * Determine whether modifications can proceed. The caller specifies the minimum
1144 * freeze level for which modifications should not be allowed. This allows
1145 * certain operations to proceed while the freeze sequence is in progress, if
1146 * necessary.
1147 */
1148bool
1149xfs_fs_writable(
1150 struct xfs_mount *mp,
1151 int level)
1152{
1153 ASSERT(level > SB_UNFROZEN);
1154 if ((mp->m_super->s_writers.frozen >= level) ||
1155 XFS_FORCED_SHUTDOWN(mp) || (mp->m_flags & XFS_MOUNT_RDONLY))
1156 return false;
1157
1158 return true;
1159}
1160
1161/*
1162 * xfs_log_sbcount
1163 *
1164 * Sync the superblock counters to disk.
1165 *
1166 * Note this code can be called during the process of freezing, so we use the
1167 * transaction allocator that does not block when the transaction subsystem is
1168 * in its frozen state.
1169 */
1170int
1171xfs_log_sbcount(xfs_mount_t *mp)
1172{
1173 /* allow this to proceed during the freeze sequence... */
1174 if (!xfs_fs_writable(mp, SB_FREEZE_COMPLETE))
1175 return 0;
1176
1177 /*
1178 * we don't need to do this if we are updating the superblock
1179 * counters on every modification.
1180 */
1181 if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1182 return 0;
1183
1184 return xfs_sync_sb(mp, true);
1185}
1186
1187/*
1188 * Deltas for the block count can vary from 1 to very large, but lock contention
1189 * only occurs on frequent small block count updates such as in the delayed
1190 * allocation path for buffered writes (page a time updates). Hence we set
1191 * a large batch count (1024) to minimise global counter updates except when
1192 * we get near to ENOSPC and we have to be very accurate with our updates.
1193 */
1194#define XFS_FDBLOCKS_BATCH 1024
1195int
1196xfs_mod_fdblocks(
1197 struct xfs_mount *mp,
1198 int64_t delta,
1199 bool rsvd)
1200{
1201 int64_t lcounter;
1202 long long res_used;
1203 s32 batch;
1204
1205 if (delta > 0) {
1206 /*
1207 * If the reserve pool is depleted, put blocks back into it
1208 * first. Most of the time the pool is full.
1209 */
1210 if (likely(mp->m_resblks == mp->m_resblks_avail)) {
1211 percpu_counter_add(&mp->m_fdblocks, delta);
1212 return 0;
1213 }
1214
1215 spin_lock(&mp->m_sb_lock);
1216 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1217
1218 if (res_used > delta) {
1219 mp->m_resblks_avail += delta;
1220 } else {
1221 delta -= res_used;
1222 mp->m_resblks_avail = mp->m_resblks;
1223 percpu_counter_add(&mp->m_fdblocks, delta);
1224 }
1225 spin_unlock(&mp->m_sb_lock);
1226 return 0;
1227 }
1228
1229 /*
1230 * Taking blocks away, need to be more accurate the closer we
1231 * are to zero.
1232 *
1233 * If the counter has a value of less than 2 * max batch size,
1234 * then make everything serialise as we are real close to
1235 * ENOSPC.
1236 */
1237 if (__percpu_counter_compare(&mp->m_fdblocks, 2 * XFS_FDBLOCKS_BATCH,
1238 XFS_FDBLOCKS_BATCH) < 0)
1239 batch = 1;
1240 else
1241 batch = XFS_FDBLOCKS_BATCH;
1242
1243 percpu_counter_add_batch(&mp->m_fdblocks, delta, batch);
1244 if (__percpu_counter_compare(&mp->m_fdblocks, mp->m_alloc_set_aside,
1245 XFS_FDBLOCKS_BATCH) >= 0) {
1246 /* we had space! */
1247 return 0;
1248 }
1249
1250 /*
1251 * lock up the sb for dipping into reserves before releasing the space
1252 * that took us to ENOSPC.
1253 */
1254 spin_lock(&mp->m_sb_lock);
1255 percpu_counter_add(&mp->m_fdblocks, -delta);
1256 if (!rsvd)
1257 goto fdblocks_enospc;
1258
1259 lcounter = (long long)mp->m_resblks_avail + delta;
1260 if (lcounter >= 0) {
1261 mp->m_resblks_avail = lcounter;
1262 spin_unlock(&mp->m_sb_lock);
1263 return 0;
1264 }
1265 xfs_warn_once(mp,
1266"Reserve blocks depleted! Consider increasing reserve pool size.");
1267
1268fdblocks_enospc:
1269 spin_unlock(&mp->m_sb_lock);
1270 return -ENOSPC;
1271}
1272
1273int
1274xfs_mod_frextents(
1275 struct xfs_mount *mp,
1276 int64_t delta)
1277{
1278 int64_t lcounter;
1279 int ret = 0;
1280
1281 spin_lock(&mp->m_sb_lock);
1282 lcounter = mp->m_sb.sb_frextents + delta;
1283 if (lcounter < 0)
1284 ret = -ENOSPC;
1285 else
1286 mp->m_sb.sb_frextents = lcounter;
1287 spin_unlock(&mp->m_sb_lock);
1288 return ret;
1289}
1290
1291/*
1292 * xfs_getsb() is called to obtain the buffer for the superblock.
1293 * The buffer is returned locked and read in from disk.
1294 * The buffer should be released with a call to xfs_brelse().
1295 */
1296struct xfs_buf *
1297xfs_getsb(
1298 struct xfs_mount *mp)
1299{
1300 struct xfs_buf *bp = mp->m_sb_bp;
1301
1302 xfs_buf_lock(bp);
1303 xfs_buf_hold(bp);
1304 ASSERT(bp->b_flags & XBF_DONE);
1305 return bp;
1306}
1307
1308/*
1309 * Used to free the superblock along various error paths.
1310 */
1311void
1312xfs_freesb(
1313 struct xfs_mount *mp)
1314{
1315 struct xfs_buf *bp = mp->m_sb_bp;
1316
1317 xfs_buf_lock(bp);
1318 mp->m_sb_bp = NULL;
1319 xfs_buf_relse(bp);
1320}
1321
1322/*
1323 * If the underlying (data/log/rt) device is readonly, there are some
1324 * operations that cannot proceed.
1325 */
1326int
1327xfs_dev_is_read_only(
1328 struct xfs_mount *mp,
1329 char *message)
1330{
1331 if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1332 xfs_readonly_buftarg(mp->m_logdev_targp) ||
1333 (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1334 xfs_notice(mp, "%s required on read-only device.", message);
1335 xfs_notice(mp, "write access unavailable, cannot proceed.");
1336 return -EROFS;
1337 }
1338 return 0;
1339}
1340
1341/* Force the summary counters to be recalculated at next mount. */
1342void
1343xfs_force_summary_recalc(
1344 struct xfs_mount *mp)
1345{
1346 if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1347 return;
1348
1349 xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
1350}
1351
1352/*
1353 * Update the in-core delayed block counter.
1354 *
1355 * We prefer to update the counter without having to take a spinlock for every
1356 * counter update (i.e. batching). Each change to delayed allocation
1357 * reservations can change can easily exceed the default percpu counter
1358 * batching, so we use a larger batch factor here.
1359 *
1360 * Note that we don't currently have any callers requiring fast summation
1361 * (e.g. percpu_counter_read) so we can use a big batch value here.
1362 */
1363#define XFS_DELALLOC_BATCH (4096)
1364void
1365xfs_mod_delalloc(
1366 struct xfs_mount *mp,
1367 int64_t delta)
1368{
1369 percpu_counter_add_batch(&mp->m_delalloc_blks, delta,
1370 XFS_DELALLOC_BATCH);
1371}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_bit.h"
13#include "xfs_sb.h"
14#include "xfs_mount.h"
15#include "xfs_inode.h"
16#include "xfs_dir2.h"
17#include "xfs_ialloc.h"
18#include "xfs_alloc.h"
19#include "xfs_rtalloc.h"
20#include "xfs_bmap.h"
21#include "xfs_trans.h"
22#include "xfs_trans_priv.h"
23#include "xfs_log.h"
24#include "xfs_error.h"
25#include "xfs_quota.h"
26#include "xfs_fsops.h"
27#include "xfs_icache.h"
28#include "xfs_sysfs.h"
29#include "xfs_rmap_btree.h"
30#include "xfs_refcount_btree.h"
31#include "xfs_reflink.h"
32#include "xfs_extent_busy.h"
33#include "xfs_health.h"
34#include "xfs_trace.h"
35#include "xfs_ag.h"
36
37static DEFINE_MUTEX(xfs_uuid_table_mutex);
38static int xfs_uuid_table_size;
39static uuid_t *xfs_uuid_table;
40
41void
42xfs_uuid_table_free(void)
43{
44 if (xfs_uuid_table_size == 0)
45 return;
46 kmem_free(xfs_uuid_table);
47 xfs_uuid_table = NULL;
48 xfs_uuid_table_size = 0;
49}
50
51/*
52 * See if the UUID is unique among mounted XFS filesystems.
53 * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
54 */
55STATIC int
56xfs_uuid_mount(
57 struct xfs_mount *mp)
58{
59 uuid_t *uuid = &mp->m_sb.sb_uuid;
60 int hole, i;
61
62 /* Publish UUID in struct super_block */
63 uuid_copy(&mp->m_super->s_uuid, uuid);
64
65 if (mp->m_flags & XFS_MOUNT_NOUUID)
66 return 0;
67
68 if (uuid_is_null(uuid)) {
69 xfs_warn(mp, "Filesystem has null UUID - can't mount");
70 return -EINVAL;
71 }
72
73 mutex_lock(&xfs_uuid_table_mutex);
74 for (i = 0, hole = -1; i < xfs_uuid_table_size; i++) {
75 if (uuid_is_null(&xfs_uuid_table[i])) {
76 hole = i;
77 continue;
78 }
79 if (uuid_equal(uuid, &xfs_uuid_table[i]))
80 goto out_duplicate;
81 }
82
83 if (hole < 0) {
84 xfs_uuid_table = krealloc(xfs_uuid_table,
85 (xfs_uuid_table_size + 1) * sizeof(*xfs_uuid_table),
86 GFP_KERNEL | __GFP_NOFAIL);
87 hole = xfs_uuid_table_size++;
88 }
89 xfs_uuid_table[hole] = *uuid;
90 mutex_unlock(&xfs_uuid_table_mutex);
91
92 return 0;
93
94 out_duplicate:
95 mutex_unlock(&xfs_uuid_table_mutex);
96 xfs_warn(mp, "Filesystem has duplicate UUID %pU - can't mount", uuid);
97 return -EINVAL;
98}
99
100STATIC void
101xfs_uuid_unmount(
102 struct xfs_mount *mp)
103{
104 uuid_t *uuid = &mp->m_sb.sb_uuid;
105 int i;
106
107 if (mp->m_flags & XFS_MOUNT_NOUUID)
108 return;
109
110 mutex_lock(&xfs_uuid_table_mutex);
111 for (i = 0; i < xfs_uuid_table_size; i++) {
112 if (uuid_is_null(&xfs_uuid_table[i]))
113 continue;
114 if (!uuid_equal(uuid, &xfs_uuid_table[i]))
115 continue;
116 memset(&xfs_uuid_table[i], 0, sizeof(uuid_t));
117 break;
118 }
119 ASSERT(i < xfs_uuid_table_size);
120 mutex_unlock(&xfs_uuid_table_mutex);
121}
122
123/*
124 * Check size of device based on the (data/realtime) block count.
125 * Note: this check is used by the growfs code as well as mount.
126 */
127int
128xfs_sb_validate_fsb_count(
129 xfs_sb_t *sbp,
130 uint64_t nblocks)
131{
132 ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
133 ASSERT(sbp->sb_blocklog >= BBSHIFT);
134
135 /* Limited by ULONG_MAX of page cache index */
136 if (nblocks >> (PAGE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
137 return -EFBIG;
138 return 0;
139}
140
141/*
142 * xfs_readsb
143 *
144 * Does the initial read of the superblock.
145 */
146int
147xfs_readsb(
148 struct xfs_mount *mp,
149 int flags)
150{
151 unsigned int sector_size;
152 struct xfs_buf *bp;
153 struct xfs_sb *sbp = &mp->m_sb;
154 int error;
155 int loud = !(flags & XFS_MFSI_QUIET);
156 const struct xfs_buf_ops *buf_ops;
157
158 ASSERT(mp->m_sb_bp == NULL);
159 ASSERT(mp->m_ddev_targp != NULL);
160
161 /*
162 * For the initial read, we must guess at the sector
163 * size based on the block device. It's enough to
164 * get the sb_sectsize out of the superblock and
165 * then reread with the proper length.
166 * We don't verify it yet, because it may not be complete.
167 */
168 sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
169 buf_ops = NULL;
170
171 /*
172 * Allocate a (locked) buffer to hold the superblock. This will be kept
173 * around at all times to optimize access to the superblock. Therefore,
174 * set XBF_NO_IOACCT to make sure it doesn't hold the buftarg count
175 * elevated.
176 */
177reread:
178 error = xfs_buf_read_uncached(mp->m_ddev_targp, XFS_SB_DADDR,
179 BTOBB(sector_size), XBF_NO_IOACCT, &bp,
180 buf_ops);
181 if (error) {
182 if (loud)
183 xfs_warn(mp, "SB validate failed with error %d.", error);
184 /* bad CRC means corrupted metadata */
185 if (error == -EFSBADCRC)
186 error = -EFSCORRUPTED;
187 return error;
188 }
189
190 /*
191 * Initialize the mount structure from the superblock.
192 */
193 xfs_sb_from_disk(sbp, bp->b_addr);
194
195 /*
196 * If we haven't validated the superblock, do so now before we try
197 * to check the sector size and reread the superblock appropriately.
198 */
199 if (sbp->sb_magicnum != XFS_SB_MAGIC) {
200 if (loud)
201 xfs_warn(mp, "Invalid superblock magic number");
202 error = -EINVAL;
203 goto release_buf;
204 }
205
206 /*
207 * We must be able to do sector-sized and sector-aligned IO.
208 */
209 if (sector_size > sbp->sb_sectsize) {
210 if (loud)
211 xfs_warn(mp, "device supports %u byte sectors (not %u)",
212 sector_size, sbp->sb_sectsize);
213 error = -ENOSYS;
214 goto release_buf;
215 }
216
217 if (buf_ops == NULL) {
218 /*
219 * Re-read the superblock so the buffer is correctly sized,
220 * and properly verified.
221 */
222 xfs_buf_relse(bp);
223 sector_size = sbp->sb_sectsize;
224 buf_ops = loud ? &xfs_sb_buf_ops : &xfs_sb_quiet_buf_ops;
225 goto reread;
226 }
227
228 xfs_reinit_percpu_counters(mp);
229
230 /* no need to be quiet anymore, so reset the buf ops */
231 bp->b_ops = &xfs_sb_buf_ops;
232
233 mp->m_sb_bp = bp;
234 xfs_buf_unlock(bp);
235 return 0;
236
237release_buf:
238 xfs_buf_relse(bp);
239 return error;
240}
241
242/*
243 * If the sunit/swidth change would move the precomputed root inode value, we
244 * must reject the ondisk change because repair will stumble over that.
245 * However, we allow the mount to proceed because we never rejected this
246 * combination before. Returns true to update the sb, false otherwise.
247 */
248static inline int
249xfs_check_new_dalign(
250 struct xfs_mount *mp,
251 int new_dalign,
252 bool *update_sb)
253{
254 struct xfs_sb *sbp = &mp->m_sb;
255 xfs_ino_t calc_ino;
256
257 calc_ino = xfs_ialloc_calc_rootino(mp, new_dalign);
258 trace_xfs_check_new_dalign(mp, new_dalign, calc_ino);
259
260 if (sbp->sb_rootino == calc_ino) {
261 *update_sb = true;
262 return 0;
263 }
264
265 xfs_warn(mp,
266"Cannot change stripe alignment; would require moving root inode.");
267
268 /*
269 * XXX: Next time we add a new incompat feature, this should start
270 * returning -EINVAL to fail the mount. Until then, spit out a warning
271 * that we're ignoring the administrator's instructions.
272 */
273 xfs_warn(mp, "Skipping superblock stripe alignment update.");
274 *update_sb = false;
275 return 0;
276}
277
278/*
279 * If we were provided with new sunit/swidth values as mount options, make sure
280 * that they pass basic alignment and superblock feature checks, and convert
281 * them into the same units (FSB) that everything else expects. This step
282 * /must/ be done before computing the inode geometry.
283 */
284STATIC int
285xfs_validate_new_dalign(
286 struct xfs_mount *mp)
287{
288 if (mp->m_dalign == 0)
289 return 0;
290
291 /*
292 * If stripe unit and stripe width are not multiples
293 * of the fs blocksize turn off alignment.
294 */
295 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
296 (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
297 xfs_warn(mp,
298 "alignment check failed: sunit/swidth vs. blocksize(%d)",
299 mp->m_sb.sb_blocksize);
300 return -EINVAL;
301 } else {
302 /*
303 * Convert the stripe unit and width to FSBs.
304 */
305 mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
306 if (mp->m_dalign && (mp->m_sb.sb_agblocks % mp->m_dalign)) {
307 xfs_warn(mp,
308 "alignment check failed: sunit/swidth vs. agsize(%d)",
309 mp->m_sb.sb_agblocks);
310 return -EINVAL;
311 } else if (mp->m_dalign) {
312 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
313 } else {
314 xfs_warn(mp,
315 "alignment check failed: sunit(%d) less than bsize(%d)",
316 mp->m_dalign, mp->m_sb.sb_blocksize);
317 return -EINVAL;
318 }
319 }
320
321 if (!xfs_sb_version_hasdalign(&mp->m_sb)) {
322 xfs_warn(mp,
323"cannot change alignment: superblock does not support data alignment");
324 return -EINVAL;
325 }
326
327 return 0;
328}
329
330/* Update alignment values based on mount options and sb values. */
331STATIC int
332xfs_update_alignment(
333 struct xfs_mount *mp)
334{
335 struct xfs_sb *sbp = &mp->m_sb;
336
337 if (mp->m_dalign) {
338 bool update_sb;
339 int error;
340
341 if (sbp->sb_unit == mp->m_dalign &&
342 sbp->sb_width == mp->m_swidth)
343 return 0;
344
345 error = xfs_check_new_dalign(mp, mp->m_dalign, &update_sb);
346 if (error || !update_sb)
347 return error;
348
349 sbp->sb_unit = mp->m_dalign;
350 sbp->sb_width = mp->m_swidth;
351 mp->m_update_sb = true;
352 } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
353 xfs_sb_version_hasdalign(&mp->m_sb)) {
354 mp->m_dalign = sbp->sb_unit;
355 mp->m_swidth = sbp->sb_width;
356 }
357
358 return 0;
359}
360
361/*
362 * precalculate the low space thresholds for dynamic speculative preallocation.
363 */
364void
365xfs_set_low_space_thresholds(
366 struct xfs_mount *mp)
367{
368 int i;
369
370 for (i = 0; i < XFS_LOWSP_MAX; i++) {
371 uint64_t space = mp->m_sb.sb_dblocks;
372
373 do_div(space, 100);
374 mp->m_low_space[i] = space * (i + 1);
375 }
376}
377
378/*
379 * Check that the data (and log if separate) is an ok size.
380 */
381STATIC int
382xfs_check_sizes(
383 struct xfs_mount *mp)
384{
385 struct xfs_buf *bp;
386 xfs_daddr_t d;
387 int error;
388
389 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
390 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
391 xfs_warn(mp, "filesystem size mismatch detected");
392 return -EFBIG;
393 }
394 error = xfs_buf_read_uncached(mp->m_ddev_targp,
395 d - XFS_FSS_TO_BB(mp, 1),
396 XFS_FSS_TO_BB(mp, 1), 0, &bp, NULL);
397 if (error) {
398 xfs_warn(mp, "last sector read failed");
399 return error;
400 }
401 xfs_buf_relse(bp);
402
403 if (mp->m_logdev_targp == mp->m_ddev_targp)
404 return 0;
405
406 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
407 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
408 xfs_warn(mp, "log size mismatch detected");
409 return -EFBIG;
410 }
411 error = xfs_buf_read_uncached(mp->m_logdev_targp,
412 d - XFS_FSB_TO_BB(mp, 1),
413 XFS_FSB_TO_BB(mp, 1), 0, &bp, NULL);
414 if (error) {
415 xfs_warn(mp, "log device read failed");
416 return error;
417 }
418 xfs_buf_relse(bp);
419 return 0;
420}
421
422/*
423 * Clear the quotaflags in memory and in the superblock.
424 */
425int
426xfs_mount_reset_sbqflags(
427 struct xfs_mount *mp)
428{
429 mp->m_qflags = 0;
430
431 /* It is OK to look at sb_qflags in the mount path without m_sb_lock. */
432 if (mp->m_sb.sb_qflags == 0)
433 return 0;
434 spin_lock(&mp->m_sb_lock);
435 mp->m_sb.sb_qflags = 0;
436 spin_unlock(&mp->m_sb_lock);
437
438 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
439 return 0;
440
441 return xfs_sync_sb(mp, false);
442}
443
444uint64_t
445xfs_default_resblks(xfs_mount_t *mp)
446{
447 uint64_t resblks;
448
449 /*
450 * We default to 5% or 8192 fsbs of space reserved, whichever is
451 * smaller. This is intended to cover concurrent allocation
452 * transactions when we initially hit enospc. These each require a 4
453 * block reservation. Hence by default we cover roughly 2000 concurrent
454 * allocation reservations.
455 */
456 resblks = mp->m_sb.sb_dblocks;
457 do_div(resblks, 20);
458 resblks = min_t(uint64_t, resblks, 8192);
459 return resblks;
460}
461
462/* Ensure the summary counts are correct. */
463STATIC int
464xfs_check_summary_counts(
465 struct xfs_mount *mp)
466{
467 /*
468 * The AG0 superblock verifier rejects in-progress filesystems,
469 * so we should never see the flag set this far into mounting.
470 */
471 if (mp->m_sb.sb_inprogress) {
472 xfs_err(mp, "sb_inprogress set after log recovery??");
473 WARN_ON(1);
474 return -EFSCORRUPTED;
475 }
476
477 /*
478 * Now the log is mounted, we know if it was an unclean shutdown or
479 * not. If it was, with the first phase of recovery has completed, we
480 * have consistent AG blocks on disk. We have not recovered EFIs yet,
481 * but they are recovered transactionally in the second recovery phase
482 * later.
483 *
484 * If the log was clean when we mounted, we can check the summary
485 * counters. If any of them are obviously incorrect, we can recompute
486 * them from the AGF headers in the next step.
487 */
488 if (XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
489 (mp->m_sb.sb_fdblocks > mp->m_sb.sb_dblocks ||
490 !xfs_verify_icount(mp, mp->m_sb.sb_icount) ||
491 mp->m_sb.sb_ifree > mp->m_sb.sb_icount))
492 xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
493
494 /*
495 * We can safely re-initialise incore superblock counters from the
496 * per-ag data. These may not be correct if the filesystem was not
497 * cleanly unmounted, so we waited for recovery to finish before doing
498 * this.
499 *
500 * If the filesystem was cleanly unmounted or the previous check did
501 * not flag anything weird, then we can trust the values in the
502 * superblock to be correct and we don't need to do anything here.
503 * Otherwise, recalculate the summary counters.
504 */
505 if ((!xfs_sb_version_haslazysbcount(&mp->m_sb) ||
506 XFS_LAST_UNMOUNT_WAS_CLEAN(mp)) &&
507 !xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS))
508 return 0;
509
510 return xfs_initialize_perag_data(mp, mp->m_sb.sb_agcount);
511}
512
513/*
514 * Flush and reclaim dirty inodes in preparation for unmount. Inodes and
515 * internal inode structures can be sitting in the CIL and AIL at this point,
516 * so we need to unpin them, write them back and/or reclaim them before unmount
517 * can proceed.
518 *
519 * An inode cluster that has been freed can have its buffer still pinned in
520 * memory because the transaction is still sitting in a iclog. The stale inodes
521 * on that buffer will be pinned to the buffer until the transaction hits the
522 * disk and the callbacks run. Pushing the AIL will skip the stale inodes and
523 * may never see the pinned buffer, so nothing will push out the iclog and
524 * unpin the buffer.
525 *
526 * Hence we need to force the log to unpin everything first. However, log
527 * forces don't wait for the discards they issue to complete, so we have to
528 * explicitly wait for them to complete here as well.
529 *
530 * Then we can tell the world we are unmounting so that error handling knows
531 * that the filesystem is going away and we should error out anything that we
532 * have been retrying in the background. This will prevent never-ending
533 * retries in AIL pushing from hanging the unmount.
534 *
535 * Finally, we can push the AIL to clean all the remaining dirty objects, then
536 * reclaim the remaining inodes that are still in memory at this point in time.
537 */
538static void
539xfs_unmount_flush_inodes(
540 struct xfs_mount *mp)
541{
542 xfs_log_force(mp, XFS_LOG_SYNC);
543 xfs_extent_busy_wait_all(mp);
544 flush_workqueue(xfs_discard_wq);
545
546 mp->m_flags |= XFS_MOUNT_UNMOUNTING;
547
548 xfs_ail_push_all_sync(mp->m_ail);
549 cancel_delayed_work_sync(&mp->m_reclaim_work);
550 xfs_reclaim_inodes(mp);
551 xfs_health_unmount(mp);
552}
553
554static void
555xfs_mount_setup_inode_geom(
556 struct xfs_mount *mp)
557{
558 struct xfs_ino_geometry *igeo = M_IGEO(mp);
559
560 igeo->attr_fork_offset = xfs_bmap_compute_attr_offset(mp);
561 ASSERT(igeo->attr_fork_offset < XFS_LITINO(mp));
562
563 xfs_ialloc_setup_geometry(mp);
564}
565
566/*
567 * This function does the following on an initial mount of a file system:
568 * - reads the superblock from disk and init the mount struct
569 * - if we're a 32-bit kernel, do a size check on the superblock
570 * so we don't mount terabyte filesystems
571 * - init mount struct realtime fields
572 * - allocate inode hash table for fs
573 * - init directory manager
574 * - perform recovery and init the log manager
575 */
576int
577xfs_mountfs(
578 struct xfs_mount *mp)
579{
580 struct xfs_sb *sbp = &(mp->m_sb);
581 struct xfs_inode *rip;
582 struct xfs_ino_geometry *igeo = M_IGEO(mp);
583 uint64_t resblks;
584 uint quotamount = 0;
585 uint quotaflags = 0;
586 int error = 0;
587
588 xfs_sb_mount_common(mp, sbp);
589
590 /*
591 * Check for a mismatched features2 values. Older kernels read & wrote
592 * into the wrong sb offset for sb_features2 on some platforms due to
593 * xfs_sb_t not being 64bit size aligned when sb_features2 was added,
594 * which made older superblock reading/writing routines swap it as a
595 * 64-bit value.
596 *
597 * For backwards compatibility, we make both slots equal.
598 *
599 * If we detect a mismatched field, we OR the set bits into the existing
600 * features2 field in case it has already been modified; we don't want
601 * to lose any features. We then update the bad location with the ORed
602 * value so that older kernels will see any features2 flags. The
603 * superblock writeback code ensures the new sb_features2 is copied to
604 * sb_bad_features2 before it is logged or written to disk.
605 */
606 if (xfs_sb_has_mismatched_features2(sbp)) {
607 xfs_warn(mp, "correcting sb_features alignment problem");
608 sbp->sb_features2 |= sbp->sb_bad_features2;
609 mp->m_update_sb = true;
610
611 /*
612 * Re-check for ATTR2 in case it was found in bad_features2
613 * slot.
614 */
615 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
616 !(mp->m_flags & XFS_MOUNT_NOATTR2))
617 mp->m_flags |= XFS_MOUNT_ATTR2;
618 }
619
620 if (xfs_sb_version_hasattr2(&mp->m_sb) &&
621 (mp->m_flags & XFS_MOUNT_NOATTR2)) {
622 xfs_sb_version_removeattr2(&mp->m_sb);
623 mp->m_update_sb = true;
624
625 /* update sb_versionnum for the clearing of the morebits */
626 if (!sbp->sb_features2)
627 mp->m_update_sb = true;
628 }
629
630 /* always use v2 inodes by default now */
631 if (!(mp->m_sb.sb_versionnum & XFS_SB_VERSION_NLINKBIT)) {
632 mp->m_sb.sb_versionnum |= XFS_SB_VERSION_NLINKBIT;
633 mp->m_update_sb = true;
634 }
635
636 /*
637 * If we were given new sunit/swidth options, do some basic validation
638 * checks and convert the incore dalign and swidth values to the
639 * same units (FSB) that everything else uses. This /must/ happen
640 * before computing the inode geometry.
641 */
642 error = xfs_validate_new_dalign(mp);
643 if (error)
644 goto out;
645
646 xfs_alloc_compute_maxlevels(mp);
647 xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
648 xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
649 xfs_mount_setup_inode_geom(mp);
650 xfs_rmapbt_compute_maxlevels(mp);
651 xfs_refcountbt_compute_maxlevels(mp);
652
653 /*
654 * Check if sb_agblocks is aligned at stripe boundary. If sb_agblocks
655 * is NOT aligned turn off m_dalign since allocator alignment is within
656 * an ag, therefore ag has to be aligned at stripe boundary. Note that
657 * we must compute the free space and rmap btree geometry before doing
658 * this.
659 */
660 error = xfs_update_alignment(mp);
661 if (error)
662 goto out;
663
664 /* enable fail_at_unmount as default */
665 mp->m_fail_unmount = true;
666
667 error = xfs_sysfs_init(&mp->m_kobj, &xfs_mp_ktype,
668 NULL, mp->m_super->s_id);
669 if (error)
670 goto out;
671
672 error = xfs_sysfs_init(&mp->m_stats.xs_kobj, &xfs_stats_ktype,
673 &mp->m_kobj, "stats");
674 if (error)
675 goto out_remove_sysfs;
676
677 error = xfs_error_sysfs_init(mp);
678 if (error)
679 goto out_del_stats;
680
681 error = xfs_errortag_init(mp);
682 if (error)
683 goto out_remove_error_sysfs;
684
685 error = xfs_uuid_mount(mp);
686 if (error)
687 goto out_remove_errortag;
688
689 /*
690 * Update the preferred write size based on the information from the
691 * on-disk superblock.
692 */
693 mp->m_allocsize_log =
694 max_t(uint32_t, sbp->sb_blocklog, mp->m_allocsize_log);
695 mp->m_allocsize_blocks = 1U << (mp->m_allocsize_log - sbp->sb_blocklog);
696
697 /* set the low space thresholds for dynamic preallocation */
698 xfs_set_low_space_thresholds(mp);
699
700 /*
701 * If enabled, sparse inode chunk alignment is expected to match the
702 * cluster size. Full inode chunk alignment must match the chunk size,
703 * but that is checked on sb read verification...
704 */
705 if (xfs_sb_version_hassparseinodes(&mp->m_sb) &&
706 mp->m_sb.sb_spino_align !=
707 XFS_B_TO_FSBT(mp, igeo->inode_cluster_size_raw)) {
708 xfs_warn(mp,
709 "Sparse inode block alignment (%u) must match cluster size (%llu).",
710 mp->m_sb.sb_spino_align,
711 XFS_B_TO_FSBT(mp, igeo->inode_cluster_size_raw));
712 error = -EINVAL;
713 goto out_remove_uuid;
714 }
715
716 /*
717 * Check that the data (and log if separate) is an ok size.
718 */
719 error = xfs_check_sizes(mp);
720 if (error)
721 goto out_remove_uuid;
722
723 /*
724 * Initialize realtime fields in the mount structure
725 */
726 error = xfs_rtmount_init(mp);
727 if (error) {
728 xfs_warn(mp, "RT mount failed");
729 goto out_remove_uuid;
730 }
731
732 /*
733 * Copies the low order bits of the timestamp and the randomly
734 * set "sequence" number out of a UUID.
735 */
736 mp->m_fixedfsid[0] =
737 (get_unaligned_be16(&sbp->sb_uuid.b[8]) << 16) |
738 get_unaligned_be16(&sbp->sb_uuid.b[4]);
739 mp->m_fixedfsid[1] = get_unaligned_be32(&sbp->sb_uuid.b[0]);
740
741 error = xfs_da_mount(mp);
742 if (error) {
743 xfs_warn(mp, "Failed dir/attr init: %d", error);
744 goto out_remove_uuid;
745 }
746
747 /*
748 * Initialize the precomputed transaction reservations values.
749 */
750 xfs_trans_init(mp);
751
752 /*
753 * Allocate and initialize the per-ag data.
754 */
755 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
756 if (error) {
757 xfs_warn(mp, "Failed per-ag init: %d", error);
758 goto out_free_dir;
759 }
760
761 if (XFS_IS_CORRUPT(mp, !sbp->sb_logblocks)) {
762 xfs_warn(mp, "no log defined");
763 error = -EFSCORRUPTED;
764 goto out_free_perag;
765 }
766
767 /*
768 * Log's mount-time initialization. The first part of recovery can place
769 * some items on the AIL, to be handled when recovery is finished or
770 * cancelled.
771 */
772 error = xfs_log_mount(mp, mp->m_logdev_targp,
773 XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
774 XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
775 if (error) {
776 xfs_warn(mp, "log mount failed");
777 goto out_fail_wait;
778 }
779
780 /* Make sure the summary counts are ok. */
781 error = xfs_check_summary_counts(mp);
782 if (error)
783 goto out_log_dealloc;
784
785 /*
786 * Get and sanity-check the root inode.
787 * Save the pointer to it in the mount structure.
788 */
789 error = xfs_iget(mp, NULL, sbp->sb_rootino, XFS_IGET_UNTRUSTED,
790 XFS_ILOCK_EXCL, &rip);
791 if (error) {
792 xfs_warn(mp,
793 "Failed to read root inode 0x%llx, error %d",
794 sbp->sb_rootino, -error);
795 goto out_log_dealloc;
796 }
797
798 ASSERT(rip != NULL);
799
800 if (XFS_IS_CORRUPT(mp, !S_ISDIR(VFS_I(rip)->i_mode))) {
801 xfs_warn(mp, "corrupted root inode %llu: not a directory",
802 (unsigned long long)rip->i_ino);
803 xfs_iunlock(rip, XFS_ILOCK_EXCL);
804 error = -EFSCORRUPTED;
805 goto out_rele_rip;
806 }
807 mp->m_rootip = rip; /* save it */
808
809 xfs_iunlock(rip, XFS_ILOCK_EXCL);
810
811 /*
812 * Initialize realtime inode pointers in the mount structure
813 */
814 error = xfs_rtmount_inodes(mp);
815 if (error) {
816 /*
817 * Free up the root inode.
818 */
819 xfs_warn(mp, "failed to read RT inodes");
820 goto out_rele_rip;
821 }
822
823 /*
824 * If this is a read-only mount defer the superblock updates until
825 * the next remount into writeable mode. Otherwise we would never
826 * perform the update e.g. for the root filesystem.
827 */
828 if (mp->m_update_sb && !(mp->m_flags & XFS_MOUNT_RDONLY)) {
829 error = xfs_sync_sb(mp, false);
830 if (error) {
831 xfs_warn(mp, "failed to write sb changes");
832 goto out_rtunmount;
833 }
834 }
835
836 /*
837 * Initialise the XFS quota management subsystem for this mount
838 */
839 if (XFS_IS_QUOTA_RUNNING(mp)) {
840 error = xfs_qm_newmount(mp, "amount, "aflags);
841 if (error)
842 goto out_rtunmount;
843 } else {
844 ASSERT(!XFS_IS_QUOTA_ON(mp));
845
846 /*
847 * If a file system had quotas running earlier, but decided to
848 * mount without -o uquota/pquota/gquota options, revoke the
849 * quotachecked license.
850 */
851 if (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_ACCT) {
852 xfs_notice(mp, "resetting quota flags");
853 error = xfs_mount_reset_sbqflags(mp);
854 if (error)
855 goto out_rtunmount;
856 }
857 }
858
859 /*
860 * Finish recovering the file system. This part needed to be delayed
861 * until after the root and real-time bitmap inodes were consistently
862 * read in. Temporarily create per-AG space reservations for metadata
863 * btree shape changes because space freeing transactions (for inode
864 * inactivation) require the per-AG reservation in lieu of reserving
865 * blocks.
866 */
867 error = xfs_fs_reserve_ag_blocks(mp);
868 if (error && error == -ENOSPC)
869 xfs_warn(mp,
870 "ENOSPC reserving per-AG metadata pool, log recovery may fail.");
871 error = xfs_log_mount_finish(mp);
872 xfs_fs_unreserve_ag_blocks(mp);
873 if (error) {
874 xfs_warn(mp, "log mount finish failed");
875 goto out_rtunmount;
876 }
877
878 /*
879 * Now the log is fully replayed, we can transition to full read-only
880 * mode for read-only mounts. This will sync all the metadata and clean
881 * the log so that the recovery we just performed does not have to be
882 * replayed again on the next mount.
883 *
884 * We use the same quiesce mechanism as the rw->ro remount, as they are
885 * semantically identical operations.
886 */
887 if ((mp->m_flags & (XFS_MOUNT_RDONLY|XFS_MOUNT_NORECOVERY)) ==
888 XFS_MOUNT_RDONLY) {
889 xfs_log_clean(mp);
890 }
891
892 /*
893 * Complete the quota initialisation, post-log-replay component.
894 */
895 if (quotamount) {
896 ASSERT(mp->m_qflags == 0);
897 mp->m_qflags = quotaflags;
898
899 xfs_qm_mount_quotas(mp);
900 }
901
902 /*
903 * Now we are mounted, reserve a small amount of unused space for
904 * privileged transactions. This is needed so that transaction
905 * space required for critical operations can dip into this pool
906 * when at ENOSPC. This is needed for operations like create with
907 * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
908 * are not allowed to use this reserved space.
909 *
910 * This may drive us straight to ENOSPC on mount, but that implies
911 * we were already there on the last unmount. Warn if this occurs.
912 */
913 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
914 resblks = xfs_default_resblks(mp);
915 error = xfs_reserve_blocks(mp, &resblks, NULL);
916 if (error)
917 xfs_warn(mp,
918 "Unable to allocate reserve blocks. Continuing without reserve pool.");
919
920 /* Recover any CoW blocks that never got remapped. */
921 error = xfs_reflink_recover_cow(mp);
922 if (error) {
923 xfs_err(mp,
924 "Error %d recovering leftover CoW allocations.", error);
925 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
926 goto out_quota;
927 }
928
929 /* Reserve AG blocks for future btree expansion. */
930 error = xfs_fs_reserve_ag_blocks(mp);
931 if (error && error != -ENOSPC)
932 goto out_agresv;
933 }
934
935 return 0;
936
937 out_agresv:
938 xfs_fs_unreserve_ag_blocks(mp);
939 out_quota:
940 xfs_qm_unmount_quotas(mp);
941 out_rtunmount:
942 xfs_rtunmount_inodes(mp);
943 out_rele_rip:
944 xfs_irele(rip);
945 /* Clean out dquots that might be in memory after quotacheck. */
946 xfs_qm_unmount(mp);
947 /*
948 * Flush all inode reclamation work and flush the log.
949 * We have to do this /after/ rtunmount and qm_unmount because those
950 * two will have scheduled delayed reclaim for the rt/quota inodes.
951 *
952 * This is slightly different from the unmountfs call sequence
953 * because we could be tearing down a partially set up mount. In
954 * particular, if log_mount_finish fails we bail out without calling
955 * qm_unmount_quotas and therefore rely on qm_unmount to release the
956 * quota inodes.
957 */
958 xfs_unmount_flush_inodes(mp);
959 out_log_dealloc:
960 xfs_log_mount_cancel(mp);
961 out_fail_wait:
962 if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
963 xfs_buftarg_drain(mp->m_logdev_targp);
964 xfs_buftarg_drain(mp->m_ddev_targp);
965 out_free_perag:
966 xfs_free_perag(mp);
967 out_free_dir:
968 xfs_da_unmount(mp);
969 out_remove_uuid:
970 xfs_uuid_unmount(mp);
971 out_remove_errortag:
972 xfs_errortag_del(mp);
973 out_remove_error_sysfs:
974 xfs_error_sysfs_del(mp);
975 out_del_stats:
976 xfs_sysfs_del(&mp->m_stats.xs_kobj);
977 out_remove_sysfs:
978 xfs_sysfs_del(&mp->m_kobj);
979 out:
980 return error;
981}
982
983/*
984 * This flushes out the inodes,dquots and the superblock, unmounts the
985 * log and makes sure that incore structures are freed.
986 */
987void
988xfs_unmountfs(
989 struct xfs_mount *mp)
990{
991 uint64_t resblks;
992 int error;
993
994 xfs_blockgc_stop(mp);
995 xfs_fs_unreserve_ag_blocks(mp);
996 xfs_qm_unmount_quotas(mp);
997 xfs_rtunmount_inodes(mp);
998 xfs_irele(mp->m_rootip);
999
1000 xfs_unmount_flush_inodes(mp);
1001
1002 xfs_qm_unmount(mp);
1003
1004 /*
1005 * Unreserve any blocks we have so that when we unmount we don't account
1006 * the reserved free space as used. This is really only necessary for
1007 * lazy superblock counting because it trusts the incore superblock
1008 * counters to be absolutely correct on clean unmount.
1009 *
1010 * We don't bother correcting this elsewhere for lazy superblock
1011 * counting because on mount of an unclean filesystem we reconstruct the
1012 * correct counter value and this is irrelevant.
1013 *
1014 * For non-lazy counter filesystems, this doesn't matter at all because
1015 * we only every apply deltas to the superblock and hence the incore
1016 * value does not matter....
1017 */
1018 resblks = 0;
1019 error = xfs_reserve_blocks(mp, &resblks, NULL);
1020 if (error)
1021 xfs_warn(mp, "Unable to free reserved block pool. "
1022 "Freespace may not be correct on next mount.");
1023
1024 xfs_log_unmount(mp);
1025 xfs_da_unmount(mp);
1026 xfs_uuid_unmount(mp);
1027
1028#if defined(DEBUG)
1029 xfs_errortag_clearall(mp);
1030#endif
1031 xfs_free_perag(mp);
1032
1033 xfs_errortag_del(mp);
1034 xfs_error_sysfs_del(mp);
1035 xfs_sysfs_del(&mp->m_stats.xs_kobj);
1036 xfs_sysfs_del(&mp->m_kobj);
1037}
1038
1039/*
1040 * Determine whether modifications can proceed. The caller specifies the minimum
1041 * freeze level for which modifications should not be allowed. This allows
1042 * certain operations to proceed while the freeze sequence is in progress, if
1043 * necessary.
1044 */
1045bool
1046xfs_fs_writable(
1047 struct xfs_mount *mp,
1048 int level)
1049{
1050 ASSERT(level > SB_UNFROZEN);
1051 if ((mp->m_super->s_writers.frozen >= level) ||
1052 XFS_FORCED_SHUTDOWN(mp) || (mp->m_flags & XFS_MOUNT_RDONLY))
1053 return false;
1054
1055 return true;
1056}
1057
1058/*
1059 * Deltas for the block count can vary from 1 to very large, but lock contention
1060 * only occurs on frequent small block count updates such as in the delayed
1061 * allocation path for buffered writes (page a time updates). Hence we set
1062 * a large batch count (1024) to minimise global counter updates except when
1063 * we get near to ENOSPC and we have to be very accurate with our updates.
1064 */
1065#define XFS_FDBLOCKS_BATCH 1024
1066int
1067xfs_mod_fdblocks(
1068 struct xfs_mount *mp,
1069 int64_t delta,
1070 bool rsvd)
1071{
1072 int64_t lcounter;
1073 long long res_used;
1074 s32 batch;
1075 uint64_t set_aside;
1076
1077 if (delta > 0) {
1078 /*
1079 * If the reserve pool is depleted, put blocks back into it
1080 * first. Most of the time the pool is full.
1081 */
1082 if (likely(mp->m_resblks == mp->m_resblks_avail)) {
1083 percpu_counter_add(&mp->m_fdblocks, delta);
1084 return 0;
1085 }
1086
1087 spin_lock(&mp->m_sb_lock);
1088 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1089
1090 if (res_used > delta) {
1091 mp->m_resblks_avail += delta;
1092 } else {
1093 delta -= res_used;
1094 mp->m_resblks_avail = mp->m_resblks;
1095 percpu_counter_add(&mp->m_fdblocks, delta);
1096 }
1097 spin_unlock(&mp->m_sb_lock);
1098 return 0;
1099 }
1100
1101 /*
1102 * Taking blocks away, need to be more accurate the closer we
1103 * are to zero.
1104 *
1105 * If the counter has a value of less than 2 * max batch size,
1106 * then make everything serialise as we are real close to
1107 * ENOSPC.
1108 */
1109 if (__percpu_counter_compare(&mp->m_fdblocks, 2 * XFS_FDBLOCKS_BATCH,
1110 XFS_FDBLOCKS_BATCH) < 0)
1111 batch = 1;
1112 else
1113 batch = XFS_FDBLOCKS_BATCH;
1114
1115 /*
1116 * Set aside allocbt blocks because these blocks are tracked as free
1117 * space but not available for allocation. Technically this means that a
1118 * single reservation cannot consume all remaining free space, but the
1119 * ratio of allocbt blocks to usable free blocks should be rather small.
1120 * The tradeoff without this is that filesystems that maintain high
1121 * perag block reservations can over reserve physical block availability
1122 * and fail physical allocation, which leads to much more serious
1123 * problems (i.e. transaction abort, pagecache discards, etc.) than
1124 * slightly premature -ENOSPC.
1125 */
1126 set_aside = mp->m_alloc_set_aside + atomic64_read(&mp->m_allocbt_blks);
1127 percpu_counter_add_batch(&mp->m_fdblocks, delta, batch);
1128 if (__percpu_counter_compare(&mp->m_fdblocks, set_aside,
1129 XFS_FDBLOCKS_BATCH) >= 0) {
1130 /* we had space! */
1131 return 0;
1132 }
1133
1134 /*
1135 * lock up the sb for dipping into reserves before releasing the space
1136 * that took us to ENOSPC.
1137 */
1138 spin_lock(&mp->m_sb_lock);
1139 percpu_counter_add(&mp->m_fdblocks, -delta);
1140 if (!rsvd)
1141 goto fdblocks_enospc;
1142
1143 lcounter = (long long)mp->m_resblks_avail + delta;
1144 if (lcounter >= 0) {
1145 mp->m_resblks_avail = lcounter;
1146 spin_unlock(&mp->m_sb_lock);
1147 return 0;
1148 }
1149 xfs_warn_once(mp,
1150"Reserve blocks depleted! Consider increasing reserve pool size.");
1151
1152fdblocks_enospc:
1153 spin_unlock(&mp->m_sb_lock);
1154 return -ENOSPC;
1155}
1156
1157int
1158xfs_mod_frextents(
1159 struct xfs_mount *mp,
1160 int64_t delta)
1161{
1162 int64_t lcounter;
1163 int ret = 0;
1164
1165 spin_lock(&mp->m_sb_lock);
1166 lcounter = mp->m_sb.sb_frextents + delta;
1167 if (lcounter < 0)
1168 ret = -ENOSPC;
1169 else
1170 mp->m_sb.sb_frextents = lcounter;
1171 spin_unlock(&mp->m_sb_lock);
1172 return ret;
1173}
1174
1175/*
1176 * Used to free the superblock along various error paths.
1177 */
1178void
1179xfs_freesb(
1180 struct xfs_mount *mp)
1181{
1182 struct xfs_buf *bp = mp->m_sb_bp;
1183
1184 xfs_buf_lock(bp);
1185 mp->m_sb_bp = NULL;
1186 xfs_buf_relse(bp);
1187}
1188
1189/*
1190 * If the underlying (data/log/rt) device is readonly, there are some
1191 * operations that cannot proceed.
1192 */
1193int
1194xfs_dev_is_read_only(
1195 struct xfs_mount *mp,
1196 char *message)
1197{
1198 if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1199 xfs_readonly_buftarg(mp->m_logdev_targp) ||
1200 (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1201 xfs_notice(mp, "%s required on read-only device.", message);
1202 xfs_notice(mp, "write access unavailable, cannot proceed.");
1203 return -EROFS;
1204 }
1205 return 0;
1206}
1207
1208/* Force the summary counters to be recalculated at next mount. */
1209void
1210xfs_force_summary_recalc(
1211 struct xfs_mount *mp)
1212{
1213 if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1214 return;
1215
1216 xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
1217}
1218
1219/*
1220 * Update the in-core delayed block counter.
1221 *
1222 * We prefer to update the counter without having to take a spinlock for every
1223 * counter update (i.e. batching). Each change to delayed allocation
1224 * reservations can change can easily exceed the default percpu counter
1225 * batching, so we use a larger batch factor here.
1226 *
1227 * Note that we don't currently have any callers requiring fast summation
1228 * (e.g. percpu_counter_read) so we can use a big batch value here.
1229 */
1230#define XFS_DELALLOC_BATCH (4096)
1231void
1232xfs_mod_delalloc(
1233 struct xfs_mount *mp,
1234 int64_t delta)
1235{
1236 percpu_counter_add_batch(&mp->m_delalloc_blks, delta,
1237 XFS_DELALLOC_BATCH);
1238}