1 // SPDX-License-Identifier: GPL-2.0+
3 * Copyright (C) 2017 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_mount.h"
12 #include "xfs_log_format.h"
13 #include "xfs_trans.h"
14 #include "xfs_inode.h"
15 #include "xfs_quota.h"
17 #include "xfs_errortag.h"
18 #include "xfs_error.h"
19 #include "scrub/scrub.h"
20 #include "scrub/common.h"
21 #include "scrub/trace.h"
22 #include "scrub/repair.h"
23 #include "scrub/health.h"
26 * Online Scrub and Repair
28 * Traditionally, XFS (the kernel driver) did not know how to check or
29 * repair on-disk data structures. That task was left to the xfs_check
30 * and xfs_repair tools, both of which require taking the filesystem
31 * offline for a thorough but time consuming examination. Online
32 * scrub & repair, on the other hand, enables us to check the metadata
33 * for obvious errors while carefully stepping around the filesystem's
34 * ongoing operations, locking rules, etc.
36 * Given that most XFS metadata consist of records stored in a btree,
37 * most of the checking functions iterate the btree blocks themselves
38 * looking for irregularities. When a record block is encountered, each
39 * record can be checked for obviously bad values. Record values can
40 * also be cross-referenced against other btrees to look for potential
41 * misunderstandings between pieces of metadata.
43 * It is expected that the checkers responsible for per-AG metadata
44 * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
45 * metadata structure, and perform any relevant cross-referencing before
46 * unlocking the AG and returning the results to userspace. These
47 * scrubbers must not keep an AG locked for too long to avoid tying up
48 * the block and inode allocators.
50 * Block maps and b-trees rooted in an inode present a special challenge
51 * because they can involve extents from any AG. The general scrubber
52 * structure of lock -> check -> xref -> unlock still holds, but AG
53 * locking order rules /must/ be obeyed to avoid deadlocks. The
54 * ordering rule, of course, is that we must lock in increasing AG
55 * order. Helper functions are provided to track which AG headers we've
56 * already locked. If we detect an imminent locking order violation, we
57 * can signal a potential deadlock, in which case the scrubber can jump
58 * out to the top level, lock all the AGs in order, and retry the scrub.
60 * For file data (directories, extended attributes, symlinks) scrub, we
61 * can simply lock the inode and walk the data. For btree data
62 * (directories and attributes) we follow the same btree-scrubbing
63 * strategy outlined previously to check the records.
65 * We use a bit of trickery with transactions to avoid buffer deadlocks
66 * if there is a cycle in the metadata. The basic problem is that
67 * travelling down a btree involves locking the current buffer at each
68 * tree level. If a pointer should somehow point back to a buffer that
69 * we've already examined, we will deadlock due to the second buffer
70 * locking attempt. Note however that grabbing a buffer in transaction
71 * context links the locked buffer to the transaction. If we try to
72 * re-grab the buffer in the context of the same transaction, we avoid
73 * the second lock attempt and continue. Between the verifier and the
74 * scrubber, something will notice that something is amiss and report
75 * the corruption. Therefore, each scrubber will allocate an empty
76 * transaction, attach buffers to it, and cancel the transaction at the
77 * end of the scrub run. Cancelling a non-dirty transaction simply
78 * unlocks the buffers.
80 * There are four pieces of data that scrub can communicate to
81 * userspace. The first is the error code (errno), which can be used to
82 * communicate operational errors in performing the scrub. There are
83 * also three flags that can be set in the scrub context. If the data
84 * structure itself is corrupt, the CORRUPT flag will be set. If
85 * the metadata is correct but otherwise suboptimal, the PREEN flag
88 * We perform secondary validation of filesystem metadata by
89 * cross-referencing every record with all other available metadata.
90 * For example, for block mapping extents, we verify that there are no
91 * records in the free space and inode btrees corresponding to that
92 * space extent and that there is a corresponding entry in the reverse
93 * mapping btree. Inconsistent metadata is noted by setting the
94 * XCORRUPT flag; btree query function errors are noted by setting the
95 * XFAIL flag and deleting the cursor to prevent further attempts to
96 * cross-reference with a defective btree.
98 * If a piece of metadata proves corrupt or suboptimal, the userspace
99 * program can ask the kernel to apply some tender loving care (TLC) to
100 * the metadata object by setting the REPAIR flag and re-calling the
101 * scrub ioctl. "Corruption" is defined by metadata violating the
102 * on-disk specification; operations cannot continue if the violation is
103 * left untreated. It is possible for XFS to continue if an object is
104 * "suboptimal", however performance may be degraded. Repairs are
105 * usually performed by rebuilding the metadata entirely out of
106 * redundant metadata. Optimizing, on the other hand, can sometimes be
107 * done without rebuilding entire structures.
109 * Generally speaking, the repair code has the following code structure:
110 * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
111 * The first check helps us figure out if we need to rebuild or simply
112 * optimize the structure so that the rebuild knows what to do. The
113 * second check evaluates the completeness of the repair; that is what
114 * is reported to userspace.
116 * A quick note on symbol prefixes:
117 * - "xfs_" are general XFS symbols.
118 * - "xchk_" are symbols related to metadata checking.
119 * - "xrep_" are symbols related to metadata repair.
120 * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
124 * Scrub probe -- userspace uses this to probe if we're willing to scrub
125 * or repair a given mountpoint. This will be used by xfs_scrub to
126 * probe the kernel's abilities to scrub (and repair) the metadata. We
127 * do this by validating the ioctl inputs from userspace, preparing the
128 * filesystem for a scrub (or a repair) operation, and immediately
129 * returning to userspace. Userspace can use the returned errno and
130 * structure state to decide (in broad terms) if scrub/repair are
131 * supported by the running kernel.
135 struct xfs_scrub *sc)
139 if (xchk_should_terminate(sc, &error))
145 /* Scrub setup and teardown */
147 /* Free all the resources and finish the transactions. */
150 struct xfs_scrub *sc,
151 struct xfs_inode *ip_in,
154 xchk_ag_free(sc, &sc->sa);
156 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
157 error = xfs_trans_commit(sc->tp);
159 xfs_trans_cancel(sc->tp);
164 xfs_iunlock(sc->ip, sc->ilock_flags);
165 if (sc->ip != ip_in &&
166 !xfs_internal_inum(sc->mp, sc->ip->i_ino))
170 if (sc->flags & XCHK_REAPING_DISABLED)
171 xchk_start_reaping(sc);
172 if (sc->flags & XCHK_HAS_QUOTAOFFLOCK) {
173 mutex_unlock(&sc->mp->m_quotainfo->qi_quotaofflock);
174 sc->flags &= ~XCHK_HAS_QUOTAOFFLOCK;
183 /* Scrubbing dispatch. */
185 static const struct xchk_meta_ops meta_scrub_ops[] = {
186 [XFS_SCRUB_TYPE_PROBE] = { /* ioctl presence test */
188 .setup = xchk_setup_fs,
190 .repair = xrep_probe,
192 [XFS_SCRUB_TYPE_SB] = { /* superblock */
194 .setup = xchk_setup_fs,
195 .scrub = xchk_superblock,
196 .repair = xrep_superblock,
198 [XFS_SCRUB_TYPE_AGF] = { /* agf */
200 .setup = xchk_setup_fs,
204 [XFS_SCRUB_TYPE_AGFL]= { /* agfl */
206 .setup = xchk_setup_fs,
210 [XFS_SCRUB_TYPE_AGI] = { /* agi */
212 .setup = xchk_setup_fs,
216 [XFS_SCRUB_TYPE_BNOBT] = { /* bnobt */
218 .setup = xchk_setup_ag_allocbt,
220 .repair = xrep_notsupported,
222 [XFS_SCRUB_TYPE_CNTBT] = { /* cntbt */
224 .setup = xchk_setup_ag_allocbt,
226 .repair = xrep_notsupported,
228 [XFS_SCRUB_TYPE_INOBT] = { /* inobt */
230 .setup = xchk_setup_ag_iallocbt,
232 .repair = xrep_notsupported,
234 [XFS_SCRUB_TYPE_FINOBT] = { /* finobt */
236 .setup = xchk_setup_ag_iallocbt,
237 .scrub = xchk_finobt,
238 .has = xfs_sb_version_hasfinobt,
239 .repair = xrep_notsupported,
241 [XFS_SCRUB_TYPE_RMAPBT] = { /* rmapbt */
243 .setup = xchk_setup_ag_rmapbt,
244 .scrub = xchk_rmapbt,
245 .has = xfs_sb_version_hasrmapbt,
246 .repair = xrep_notsupported,
248 [XFS_SCRUB_TYPE_REFCNTBT] = { /* refcountbt */
250 .setup = xchk_setup_ag_refcountbt,
251 .scrub = xchk_refcountbt,
252 .has = xfs_sb_version_hasreflink,
253 .repair = xrep_notsupported,
255 [XFS_SCRUB_TYPE_INODE] = { /* inode record */
257 .setup = xchk_setup_inode,
259 .repair = xrep_notsupported,
261 [XFS_SCRUB_TYPE_BMBTD] = { /* inode data fork */
263 .setup = xchk_setup_inode_bmap,
264 .scrub = xchk_bmap_data,
265 .repair = xrep_notsupported,
267 [XFS_SCRUB_TYPE_BMBTA] = { /* inode attr fork */
269 .setup = xchk_setup_inode_bmap,
270 .scrub = xchk_bmap_attr,
271 .repair = xrep_notsupported,
273 [XFS_SCRUB_TYPE_BMBTC] = { /* inode CoW fork */
275 .setup = xchk_setup_inode_bmap,
276 .scrub = xchk_bmap_cow,
277 .repair = xrep_notsupported,
279 [XFS_SCRUB_TYPE_DIR] = { /* directory */
281 .setup = xchk_setup_directory,
282 .scrub = xchk_directory,
283 .repair = xrep_notsupported,
285 [XFS_SCRUB_TYPE_XATTR] = { /* extended attributes */
287 .setup = xchk_setup_xattr,
289 .repair = xrep_notsupported,
291 [XFS_SCRUB_TYPE_SYMLINK] = { /* symbolic link */
293 .setup = xchk_setup_symlink,
294 .scrub = xchk_symlink,
295 .repair = xrep_notsupported,
297 [XFS_SCRUB_TYPE_PARENT] = { /* parent pointers */
299 .setup = xchk_setup_parent,
300 .scrub = xchk_parent,
301 .repair = xrep_notsupported,
303 [XFS_SCRUB_TYPE_RTBITMAP] = { /* realtime bitmap */
305 .setup = xchk_setup_rt,
306 .scrub = xchk_rtbitmap,
307 .has = xfs_sb_version_hasrealtime,
308 .repair = xrep_notsupported,
310 [XFS_SCRUB_TYPE_RTSUM] = { /* realtime summary */
312 .setup = xchk_setup_rt,
313 .scrub = xchk_rtsummary,
314 .has = xfs_sb_version_hasrealtime,
315 .repair = xrep_notsupported,
317 [XFS_SCRUB_TYPE_UQUOTA] = { /* user quota */
319 .setup = xchk_setup_quota,
321 .repair = xrep_notsupported,
323 [XFS_SCRUB_TYPE_GQUOTA] = { /* group quota */
325 .setup = xchk_setup_quota,
327 .repair = xrep_notsupported,
329 [XFS_SCRUB_TYPE_PQUOTA] = { /* project quota */
331 .setup = xchk_setup_quota,
333 .repair = xrep_notsupported,
335 [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */
337 .setup = xchk_setup_fscounters,
338 .scrub = xchk_fscounters,
339 .repair = xrep_notsupported,
343 /* This isn't a stable feature, warn once per day. */
345 xchk_experimental_warning(
346 struct xfs_mount *mp)
348 static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
349 "xchk_warning", 86400 * HZ, 1);
350 ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
352 if (__ratelimit(&scrub_warning))
354 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
358 xchk_validate_inputs(
359 struct xfs_mount *mp,
360 struct xfs_scrub_metadata *sm)
363 const struct xchk_meta_ops *ops;
366 /* Check our inputs. */
367 sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
368 if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
370 /* sm_reserved[] must be zero */
371 if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
375 /* Do we know about this type of metadata? */
376 if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
378 ops = &meta_scrub_ops[sm->sm_type];
379 if (ops->setup == NULL || ops->scrub == NULL)
381 /* Does this fs even support this type of metadata? */
382 if (ops->has && !ops->has(&mp->m_sb))
386 /* restricting fields must be appropriate for type */
390 if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
394 if (sm->sm_ino || sm->sm_gen ||
395 sm->sm_agno >= mp->m_sb.sb_agcount)
399 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
407 * We only want to repair read-write v5+ filesystems. Defer the check
408 * for ops->repair until after our scrub confirms that we need to
409 * perform repairs so that we avoid failing due to not supporting
410 * repairing an object that doesn't need repairs.
412 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
414 if (!xfs_sb_version_hascrc(&mp->m_sb))
418 if (mp->m_flags & XFS_MOUNT_RDONLY)
427 #ifdef CONFIG_XFS_ONLINE_REPAIR
428 static inline void xchk_postmortem(struct xfs_scrub *sc)
431 * Userspace asked us to repair something, we repaired it, rescanned
432 * it, and the rescan says it's still broken. Scream about this in
435 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
436 (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
437 XFS_SCRUB_OFLAG_XCORRUPT)))
438 xrep_failure(sc->mp);
441 static inline void xchk_postmortem(struct xfs_scrub *sc)
444 * Userspace asked us to scrub something, it's broken, and we have no
445 * way of fixing it. Scream in the logs.
447 if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
448 XFS_SCRUB_OFLAG_XCORRUPT))
449 xfs_alert_ratelimited(sc->mp,
450 "Corruption detected during scrub.");
452 #endif /* CONFIG_XFS_ONLINE_REPAIR */
454 /* Dispatch metadata scrubbing. */
457 struct xfs_inode *ip,
458 struct xfs_scrub_metadata *sm)
460 struct xfs_scrub sc = {
464 .agno = NULLAGNUMBER,
467 struct xfs_mount *mp = ip->i_mount;
470 BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
471 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
473 trace_xchk_start(ip, sm, error);
475 /* Forbidden if we are shut down or mounted norecovery. */
477 if (XFS_FORCED_SHUTDOWN(mp))
479 error = -ENOTRECOVERABLE;
480 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
483 error = xchk_validate_inputs(mp, sm);
487 xchk_experimental_warning(mp);
489 sc.ops = &meta_scrub_ops[sm->sm_type];
490 sc.sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
492 /* Set up for the operation. */
493 error = sc.ops->setup(&sc, ip);
497 /* Scrub for errors. */
498 error = sc.ops->scrub(&sc);
499 if (!(sc.flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
501 * Scrubbers return -EDEADLOCK to mean 'try harder'.
502 * Tear down everything we hold, then set up again with
503 * preparation for worst-case scenarios.
505 error = xchk_teardown(&sc, ip, 0);
508 sc.flags |= XCHK_TRY_HARDER;
513 xchk_update_health(&sc);
515 if ((sc.sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
516 !(sc.flags & XREP_ALREADY_FIXED)) {
519 /* Let debug users force us into the repair routines. */
520 if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR))
521 sc.sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
523 needs_fix = (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
524 XFS_SCRUB_OFLAG_XCORRUPT |
525 XFS_SCRUB_OFLAG_PREEN));
527 * If userspace asked for a repair but it wasn't necessary,
528 * report that back to userspace.
531 sc.sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
536 * If it's broken, userspace wants us to fix it, and we haven't
537 * already tried to fix it, then attempt a repair.
539 error = xrep_attempt(ip, &sc);
540 if (error == -EAGAIN) {
542 * Either the repair function succeeded or it couldn't
543 * get all the resources it needs; either way, we go
544 * back to the beginning and call the scrub function.
546 error = xchk_teardown(&sc, ip, 0);
556 xchk_postmortem(&sc);
558 error = xchk_teardown(&sc, ip, error);
560 trace_xchk_done(ip, sm, error);
561 if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
562 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;