1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Copyright (C) 2017 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
5 */
6 #include "xfs.h"
7 #include "xfs_fs.h"
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"
16 #include "xfs_qm.h"
17 #include "xfs_errortag.h"
18 #include "xfs_error.h"
19 #include "xfs_scrub.h"
20 #include "scrub/scrub.h"
21 #include "scrub/common.h"
22 #include "scrub/trace.h"
23 #include "scrub/repair.h"
24 #include "scrub/health.h"
25
26 /*
27 * Online Scrub and Repair
28 *
29 * Traditionally, XFS (the kernel driver) did not know how to check or
30 * repair on-disk data structures. That task was left to the xfs_check
31 * and xfs_repair tools, both of which require taking the filesystem
32 * offline for a thorough but time consuming examination. Online
33 * scrub & repair, on the other hand, enables us to check the metadata
34 * for obvious errors while carefully stepping around the filesystem's
35 * ongoing operations, locking rules, etc.
36 *
37 * Given that most XFS metadata consist of records stored in a btree,
38 * most of the checking functions iterate the btree blocks themselves
39 * looking for irregularities. When a record block is encountered, each
40 * record can be checked for obviously bad values. Record values can
41 * also be cross-referenced against other btrees to look for potential
42 * misunderstandings between pieces of metadata.
43 *
44 * It is expected that the checkers responsible for per-AG metadata
45 * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
46 * metadata structure, and perform any relevant cross-referencing before
47 * unlocking the AG and returning the results to userspace. These
48 * scrubbers must not keep an AG locked for too long to avoid tying up
49 * the block and inode allocators.
50 *
51 * Block maps and b-trees rooted in an inode present a special challenge
52 * because they can involve extents from any AG. The general scrubber
53 * structure of lock -> check -> xref -> unlock still holds, but AG
54 * locking order rules /must/ be obeyed to avoid deadlocks. The
55 * ordering rule, of course, is that we must lock in increasing AG
56 * order. Helper functions are provided to track which AG headers we've
57 * already locked. If we detect an imminent locking order violation, we
58 * can signal a potential deadlock, in which case the scrubber can jump
59 * out to the top level, lock all the AGs in order, and retry the scrub.
60 *
61 * For file data (directories, extended attributes, symlinks) scrub, we
62 * can simply lock the inode and walk the data. For btree data
63 * (directories and attributes) we follow the same btree-scrubbing
64 * strategy outlined previously to check the records.
65 *
66 * We use a bit of trickery with transactions to avoid buffer deadlocks
67 * if there is a cycle in the metadata. The basic problem is that
68 * travelling down a btree involves locking the current buffer at each
69 * tree level. If a pointer should somehow point back to a buffer that
70 * we've already examined, we will deadlock due to the second buffer
71 * locking attempt. Note however that grabbing a buffer in transaction
72 * context links the locked buffer to the transaction. If we try to
73 * re-grab the buffer in the context of the same transaction, we avoid
74 * the second lock attempt and continue. Between the verifier and the
75 * scrubber, something will notice that something is amiss and report
76 * the corruption. Therefore, each scrubber will allocate an empty
77 * transaction, attach buffers to it, and cancel the transaction at the
78 * end of the scrub run. Cancelling a non-dirty transaction simply
79 * unlocks the buffers.
80 *
81 * There are four pieces of data that scrub can communicate to
82 * userspace. The first is the error code (errno), which can be used to
83 * communicate operational errors in performing the scrub. There are
84 * also three flags that can be set in the scrub context. If the data
85 * structure itself is corrupt, the CORRUPT flag will be set. If
86 * the metadata is correct but otherwise suboptimal, the PREEN flag
87 * will be set.
88 *
89 * We perform secondary validation of filesystem metadata by
90 * cross-referencing every record with all other available metadata.
91 * For example, for block mapping extents, we verify that there are no
92 * records in the free space and inode btrees corresponding to that
93 * space extent and that there is a corresponding entry in the reverse
94 * mapping btree. Inconsistent metadata is noted by setting the
95 * XCORRUPT flag; btree query function errors are noted by setting the
96 * XFAIL flag and deleting the cursor to prevent further attempts to
97 * cross-reference with a defective btree.
98 *
99 * If a piece of metadata proves corrupt or suboptimal, the userspace
100 * program can ask the kernel to apply some tender loving care (TLC) to
101 * the metadata object by setting the REPAIR flag and re-calling the
102 * scrub ioctl. "Corruption" is defined by metadata violating the
103 * on-disk specification; operations cannot continue if the violation is
104 * left untreated. It is possible for XFS to continue if an object is
105 * "suboptimal", however performance may be degraded. Repairs are
106 * usually performed by rebuilding the metadata entirely out of
107 * redundant metadata. Optimizing, on the other hand, can sometimes be
108 * done without rebuilding entire structures.
109 *
110 * Generally speaking, the repair code has the following code structure:
111 * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
112 * The first check helps us figure out if we need to rebuild or simply
113 * optimize the structure so that the rebuild knows what to do. The
114 * second check evaluates the completeness of the repair; that is what
115 * is reported to userspace.
116 *
117 * A quick note on symbol prefixes:
118 * - "xfs_" are general XFS symbols.
119 * - "xchk_" are symbols related to metadata checking.
120 * - "xrep_" are symbols related to metadata repair.
121 * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
122 */
123
124 /*
125 * Scrub probe -- userspace uses this to probe if we're willing to scrub
126 * or repair a given mountpoint. This will be used by xfs_scrub to
127 * probe the kernel's abilities to scrub (and repair) the metadata. We
128 * do this by validating the ioctl inputs from userspace, preparing the
129 * filesystem for a scrub (or a repair) operation, and immediately
130 * returning to userspace. Userspace can use the returned errno and
131 * structure state to decide (in broad terms) if scrub/repair are
132 * supported by the running kernel.
133 */
134 static int
xchk_probe(struct xfs_scrub * sc)135 xchk_probe(
136 struct xfs_scrub *sc)
137 {
138 int error = 0;
139
140 if (xchk_should_terminate(sc, &error))
141 return error;
142
143 return 0;
144 }
145
146 /* Scrub setup and teardown */
147
148 /* Free all the resources and finish the transactions. */
149 STATIC int
xchk_teardown(struct xfs_scrub * sc,struct xfs_inode * ip_in,int error)150 xchk_teardown(
151 struct xfs_scrub *sc,
152 struct xfs_inode *ip_in,
153 int error)
154 {
155 xchk_ag_free(sc, &sc->sa);
156 if (sc->tp) {
157 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
158 error = xfs_trans_commit(sc->tp);
159 else
160 xfs_trans_cancel(sc->tp);
161 sc->tp = NULL;
162 }
163 if (sc->ip) {
164 if (sc->ilock_flags)
165 xfs_iunlock(sc->ip, sc->ilock_flags);
166 if (sc->ip != ip_in &&
167 !xfs_internal_inum(sc->mp, sc->ip->i_ino))
168 xfs_irele(sc->ip);
169 sc->ip = NULL;
170 }
171 sb_end_write(sc->mp->m_super);
172 if (sc->flags & XCHK_REAPING_DISABLED)
173 xchk_start_reaping(sc);
174 if (sc->flags & XCHK_HAS_QUOTAOFFLOCK) {
175 mutex_unlock(&sc->mp->m_quotainfo->qi_quotaofflock);
176 sc->flags &= ~XCHK_HAS_QUOTAOFFLOCK;
177 }
178 if (sc->buf) {
179 kmem_free(sc->buf);
180 sc->buf = NULL;
181 }
182 return error;
183 }
184
185 /* Scrubbing dispatch. */
186
187 static const struct xchk_meta_ops meta_scrub_ops[] = {
188 [XFS_SCRUB_TYPE_PROBE] = { /* ioctl presence test */
189 .type = ST_NONE,
190 .setup = xchk_setup_fs,
191 .scrub = xchk_probe,
192 .repair = xrep_probe,
193 },
194 [XFS_SCRUB_TYPE_SB] = { /* superblock */
195 .type = ST_PERAG,
196 .setup = xchk_setup_fs,
197 .scrub = xchk_superblock,
198 .repair = xrep_superblock,
199 },
200 [XFS_SCRUB_TYPE_AGF] = { /* agf */
201 .type = ST_PERAG,
202 .setup = xchk_setup_fs,
203 .scrub = xchk_agf,
204 .repair = xrep_agf,
205 },
206 [XFS_SCRUB_TYPE_AGFL]= { /* agfl */
207 .type = ST_PERAG,
208 .setup = xchk_setup_fs,
209 .scrub = xchk_agfl,
210 .repair = xrep_agfl,
211 },
212 [XFS_SCRUB_TYPE_AGI] = { /* agi */
213 .type = ST_PERAG,
214 .setup = xchk_setup_fs,
215 .scrub = xchk_agi,
216 .repair = xrep_agi,
217 },
218 [XFS_SCRUB_TYPE_BNOBT] = { /* bnobt */
219 .type = ST_PERAG,
220 .setup = xchk_setup_ag_allocbt,
221 .scrub = xchk_bnobt,
222 .repair = xrep_notsupported,
223 },
224 [XFS_SCRUB_TYPE_CNTBT] = { /* cntbt */
225 .type = ST_PERAG,
226 .setup = xchk_setup_ag_allocbt,
227 .scrub = xchk_cntbt,
228 .repair = xrep_notsupported,
229 },
230 [XFS_SCRUB_TYPE_INOBT] = { /* inobt */
231 .type = ST_PERAG,
232 .setup = xchk_setup_ag_iallocbt,
233 .scrub = xchk_inobt,
234 .repair = xrep_notsupported,
235 },
236 [XFS_SCRUB_TYPE_FINOBT] = { /* finobt */
237 .type = ST_PERAG,
238 .setup = xchk_setup_ag_iallocbt,
239 .scrub = xchk_finobt,
240 .has = xfs_sb_version_hasfinobt,
241 .repair = xrep_notsupported,
242 },
243 [XFS_SCRUB_TYPE_RMAPBT] = { /* rmapbt */
244 .type = ST_PERAG,
245 .setup = xchk_setup_ag_rmapbt,
246 .scrub = xchk_rmapbt,
247 .has = xfs_sb_version_hasrmapbt,
248 .repair = xrep_notsupported,
249 },
250 [XFS_SCRUB_TYPE_REFCNTBT] = { /* refcountbt */
251 .type = ST_PERAG,
252 .setup = xchk_setup_ag_refcountbt,
253 .scrub = xchk_refcountbt,
254 .has = xfs_sb_version_hasreflink,
255 .repair = xrep_notsupported,
256 },
257 [XFS_SCRUB_TYPE_INODE] = { /* inode record */
258 .type = ST_INODE,
259 .setup = xchk_setup_inode,
260 .scrub = xchk_inode,
261 .repair = xrep_notsupported,
262 },
263 [XFS_SCRUB_TYPE_BMBTD] = { /* inode data fork */
264 .type = ST_INODE,
265 .setup = xchk_setup_inode_bmap,
266 .scrub = xchk_bmap_data,
267 .repair = xrep_notsupported,
268 },
269 [XFS_SCRUB_TYPE_BMBTA] = { /* inode attr fork */
270 .type = ST_INODE,
271 .setup = xchk_setup_inode_bmap,
272 .scrub = xchk_bmap_attr,
273 .repair = xrep_notsupported,
274 },
275 [XFS_SCRUB_TYPE_BMBTC] = { /* inode CoW fork */
276 .type = ST_INODE,
277 .setup = xchk_setup_inode_bmap,
278 .scrub = xchk_bmap_cow,
279 .repair = xrep_notsupported,
280 },
281 [XFS_SCRUB_TYPE_DIR] = { /* directory */
282 .type = ST_INODE,
283 .setup = xchk_setup_directory,
284 .scrub = xchk_directory,
285 .repair = xrep_notsupported,
286 },
287 [XFS_SCRUB_TYPE_XATTR] = { /* extended attributes */
288 .type = ST_INODE,
289 .setup = xchk_setup_xattr,
290 .scrub = xchk_xattr,
291 .repair = xrep_notsupported,
292 },
293 [XFS_SCRUB_TYPE_SYMLINK] = { /* symbolic link */
294 .type = ST_INODE,
295 .setup = xchk_setup_symlink,
296 .scrub = xchk_symlink,
297 .repair = xrep_notsupported,
298 },
299 [XFS_SCRUB_TYPE_PARENT] = { /* parent pointers */
300 .type = ST_INODE,
301 .setup = xchk_setup_parent,
302 .scrub = xchk_parent,
303 .repair = xrep_notsupported,
304 },
305 [XFS_SCRUB_TYPE_RTBITMAP] = { /* realtime bitmap */
306 .type = ST_FS,
307 .setup = xchk_setup_rt,
308 .scrub = xchk_rtbitmap,
309 .has = xfs_sb_version_hasrealtime,
310 .repair = xrep_notsupported,
311 },
312 [XFS_SCRUB_TYPE_RTSUM] = { /* realtime summary */
313 .type = ST_FS,
314 .setup = xchk_setup_rt,
315 .scrub = xchk_rtsummary,
316 .has = xfs_sb_version_hasrealtime,
317 .repair = xrep_notsupported,
318 },
319 [XFS_SCRUB_TYPE_UQUOTA] = { /* user quota */
320 .type = ST_FS,
321 .setup = xchk_setup_quota,
322 .scrub = xchk_quota,
323 .repair = xrep_notsupported,
324 },
325 [XFS_SCRUB_TYPE_GQUOTA] = { /* group quota */
326 .type = ST_FS,
327 .setup = xchk_setup_quota,
328 .scrub = xchk_quota,
329 .repair = xrep_notsupported,
330 },
331 [XFS_SCRUB_TYPE_PQUOTA] = { /* project quota */
332 .type = ST_FS,
333 .setup = xchk_setup_quota,
334 .scrub = xchk_quota,
335 .repair = xrep_notsupported,
336 },
337 [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */
338 .type = ST_FS,
339 .setup = xchk_setup_fscounters,
340 .scrub = xchk_fscounters,
341 .repair = xrep_notsupported,
342 },
343 };
344
345 /* This isn't a stable feature, warn once per day. */
346 static inline void
xchk_experimental_warning(struct xfs_mount * mp)347 xchk_experimental_warning(
348 struct xfs_mount *mp)
349 {
350 static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
351 "xchk_warning", 86400 * HZ, 1);
352 ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
353
354 if (__ratelimit(&scrub_warning))
355 xfs_alert(mp,
356 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
357 }
358
359 static int
xchk_validate_inputs(struct xfs_mount * mp,struct xfs_scrub_metadata * sm)360 xchk_validate_inputs(
361 struct xfs_mount *mp,
362 struct xfs_scrub_metadata *sm)
363 {
364 int error;
365 const struct xchk_meta_ops *ops;
366
367 error = -EINVAL;
368 /* Check our inputs. */
369 sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
370 if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
371 goto out;
372 /* sm_reserved[] must be zero */
373 if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
374 goto out;
375
376 error = -ENOENT;
377 /* Do we know about this type of metadata? */
378 if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
379 goto out;
380 ops = &meta_scrub_ops[sm->sm_type];
381 if (ops->setup == NULL || ops->scrub == NULL)
382 goto out;
383 /* Does this fs even support this type of metadata? */
384 if (ops->has && !ops->has(&mp->m_sb))
385 goto out;
386
387 error = -EINVAL;
388 /* restricting fields must be appropriate for type */
389 switch (ops->type) {
390 case ST_NONE:
391 case ST_FS:
392 if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
393 goto out;
394 break;
395 case ST_PERAG:
396 if (sm->sm_ino || sm->sm_gen ||
397 sm->sm_agno >= mp->m_sb.sb_agcount)
398 goto out;
399 break;
400 case ST_INODE:
401 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
402 goto out;
403 break;
404 default:
405 goto out;
406 }
407
408 /*
409 * We only want to repair read-write v5+ filesystems. Defer the check
410 * for ops->repair until after our scrub confirms that we need to
411 * perform repairs so that we avoid failing due to not supporting
412 * repairing an object that doesn't need repairs.
413 */
414 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
415 error = -EOPNOTSUPP;
416 if (!xfs_sb_version_hascrc(&mp->m_sb))
417 goto out;
418
419 error = -EROFS;
420 if (mp->m_flags & XFS_MOUNT_RDONLY)
421 goto out;
422 }
423
424 error = 0;
425 out:
426 return error;
427 }
428
429 #ifdef CONFIG_XFS_ONLINE_REPAIR
xchk_postmortem(struct xfs_scrub * sc)430 static inline void xchk_postmortem(struct xfs_scrub *sc)
431 {
432 /*
433 * Userspace asked us to repair something, we repaired it, rescanned
434 * it, and the rescan says it's still broken. Scream about this in
435 * the system logs.
436 */
437 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
438 (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
439 XFS_SCRUB_OFLAG_XCORRUPT)))
440 xrep_failure(sc->mp);
441 }
442 #else
xchk_postmortem(struct xfs_scrub * sc)443 static inline void xchk_postmortem(struct xfs_scrub *sc)
444 {
445 /*
446 * Userspace asked us to scrub something, it's broken, and we have no
447 * way of fixing it. Scream in the logs.
448 */
449 if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
450 XFS_SCRUB_OFLAG_XCORRUPT))
451 xfs_alert_ratelimited(sc->mp,
452 "Corruption detected during scrub.");
453 }
454 #endif /* CONFIG_XFS_ONLINE_REPAIR */
455
456 /* Dispatch metadata scrubbing. */
457 int
xfs_scrub_metadata(struct xfs_inode * ip,struct xfs_scrub_metadata * sm)458 xfs_scrub_metadata(
459 struct xfs_inode *ip,
460 struct xfs_scrub_metadata *sm)
461 {
462 struct xfs_scrub sc = {
463 .mp = ip->i_mount,
464 .sm = sm,
465 .sa = {
466 .agno = NULLAGNUMBER,
467 },
468 };
469 struct xfs_mount *mp = ip->i_mount;
470 int error = 0;
471
472 BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
473 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
474
475 trace_xchk_start(ip, sm, error);
476
477 /* Forbidden if we are shut down or mounted norecovery. */
478 error = -ESHUTDOWN;
479 if (XFS_FORCED_SHUTDOWN(mp))
480 goto out;
481 error = -ENOTRECOVERABLE;
482 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
483 goto out;
484
485 error = xchk_validate_inputs(mp, sm);
486 if (error)
487 goto out;
488
489 xchk_experimental_warning(mp);
490
491 sc.ops = &meta_scrub_ops[sm->sm_type];
492 sc.sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
493 retry_op:
494 /*
495 * If freeze runs concurrently with a scrub, the freeze can be delayed
496 * indefinitely as we walk the filesystem and iterate over metadata
497 * buffers. Freeze quiesces the log (which waits for the buffer LRU to
498 * be emptied) and that won't happen while checking is running.
499 */
500 sb_start_write(mp->m_super);
501
502 /* Set up for the operation. */
503 error = sc.ops->setup(&sc, ip);
504 if (error)
505 goto out_teardown;
506
507 /* Scrub for errors. */
508 error = sc.ops->scrub(&sc);
509 if (!(sc.flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
510 /*
511 * Scrubbers return -EDEADLOCK to mean 'try harder'.
512 * Tear down everything we hold, then set up again with
513 * preparation for worst-case scenarios.
514 */
515 error = xchk_teardown(&sc, ip, 0);
516 if (error)
517 goto out;
518 sc.flags |= XCHK_TRY_HARDER;
519 goto retry_op;
520 } else if (error)
521 goto out_teardown;
522
523 xchk_update_health(&sc);
524
525 if ((sc.sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
526 !(sc.flags & XREP_ALREADY_FIXED)) {
527 bool needs_fix;
528
529 /* Let debug users force us into the repair routines. */
530 if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR))
531 sc.sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
532
533 needs_fix = (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
534 XFS_SCRUB_OFLAG_XCORRUPT |
535 XFS_SCRUB_OFLAG_PREEN));
536 /*
537 * If userspace asked for a repair but it wasn't necessary,
538 * report that back to userspace.
539 */
540 if (!needs_fix) {
541 sc.sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
542 goto out_nofix;
543 }
544
545 /*
546 * If it's broken, userspace wants us to fix it, and we haven't
547 * already tried to fix it, then attempt a repair.
548 */
549 error = xrep_attempt(ip, &sc);
550 if (error == -EAGAIN) {
551 /*
552 * Either the repair function succeeded or it couldn't
553 * get all the resources it needs; either way, we go
554 * back to the beginning and call the scrub function.
555 */
556 error = xchk_teardown(&sc, ip, 0);
557 if (error) {
558 xrep_failure(mp);
559 goto out;
560 }
561 goto retry_op;
562 }
563 }
564
565 out_nofix:
566 xchk_postmortem(&sc);
567 out_teardown:
568 error = xchk_teardown(&sc, ip, error);
569 out:
570 trace_xchk_done(ip, sm, error);
571 if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
572 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
573 error = 0;
574 }
575 return error;
576 }
577