Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2017-2023 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <djwong@kernel.org>
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_scrub.h"
18#include "xfs_buf_mem.h"
19#include "xfs_rmap.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#include "scrub/stats.h"
26#include "scrub/xfile.h"
27
28/*
29 * Online Scrub and Repair
30 *
31 * Traditionally, XFS (the kernel driver) did not know how to check or
32 * repair on-disk data structures. That task was left to the xfs_check
33 * and xfs_repair tools, both of which require taking the filesystem
34 * offline for a thorough but time consuming examination. Online
35 * scrub & repair, on the other hand, enables us to check the metadata
36 * for obvious errors while carefully stepping around the filesystem's
37 * ongoing operations, locking rules, etc.
38 *
39 * Given that most XFS metadata consist of records stored in a btree,
40 * most of the checking functions iterate the btree blocks themselves
41 * looking for irregularities. When a record block is encountered, each
42 * record can be checked for obviously bad values. Record values can
43 * also be cross-referenced against other btrees to look for potential
44 * misunderstandings between pieces of metadata.
45 *
46 * It is expected that the checkers responsible for per-AG metadata
47 * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
48 * metadata structure, and perform any relevant cross-referencing before
49 * unlocking the AG and returning the results to userspace. These
50 * scrubbers must not keep an AG locked for too long to avoid tying up
51 * the block and inode allocators.
52 *
53 * Block maps and b-trees rooted in an inode present a special challenge
54 * because they can involve extents from any AG. The general scrubber
55 * structure of lock -> check -> xref -> unlock still holds, but AG
56 * locking order rules /must/ be obeyed to avoid deadlocks. The
57 * ordering rule, of course, is that we must lock in increasing AG
58 * order. Helper functions are provided to track which AG headers we've
59 * already locked. If we detect an imminent locking order violation, we
60 * can signal a potential deadlock, in which case the scrubber can jump
61 * out to the top level, lock all the AGs in order, and retry the scrub.
62 *
63 * For file data (directories, extended attributes, symlinks) scrub, we
64 * can simply lock the inode and walk the data. For btree data
65 * (directories and attributes) we follow the same btree-scrubbing
66 * strategy outlined previously to check the records.
67 *
68 * We use a bit of trickery with transactions to avoid buffer deadlocks
69 * if there is a cycle in the metadata. The basic problem is that
70 * travelling down a btree involves locking the current buffer at each
71 * tree level. If a pointer should somehow point back to a buffer that
72 * we've already examined, we will deadlock due to the second buffer
73 * locking attempt. Note however that grabbing a buffer in transaction
74 * context links the locked buffer to the transaction. If we try to
75 * re-grab the buffer in the context of the same transaction, we avoid
76 * the second lock attempt and continue. Between the verifier and the
77 * scrubber, something will notice that something is amiss and report
78 * the corruption. Therefore, each scrubber will allocate an empty
79 * transaction, attach buffers to it, and cancel the transaction at the
80 * end of the scrub run. Cancelling a non-dirty transaction simply
81 * unlocks the buffers.
82 *
83 * There are four pieces of data that scrub can communicate to
84 * userspace. The first is the error code (errno), which can be used to
85 * communicate operational errors in performing the scrub. There are
86 * also three flags that can be set in the scrub context. If the data
87 * structure itself is corrupt, the CORRUPT flag will be set. If
88 * the metadata is correct but otherwise suboptimal, the PREEN flag
89 * will be set.
90 *
91 * We perform secondary validation of filesystem metadata by
92 * cross-referencing every record with all other available metadata.
93 * For example, for block mapping extents, we verify that there are no
94 * records in the free space and inode btrees corresponding to that
95 * space extent and that there is a corresponding entry in the reverse
96 * mapping btree. Inconsistent metadata is noted by setting the
97 * XCORRUPT flag; btree query function errors are noted by setting the
98 * XFAIL flag and deleting the cursor to prevent further attempts to
99 * cross-reference with a defective btree.
100 *
101 * If a piece of metadata proves corrupt or suboptimal, the userspace
102 * program can ask the kernel to apply some tender loving care (TLC) to
103 * the metadata object by setting the REPAIR flag and re-calling the
104 * scrub ioctl. "Corruption" is defined by metadata violating the
105 * on-disk specification; operations cannot continue if the violation is
106 * left untreated. It is possible for XFS to continue if an object is
107 * "suboptimal", however performance may be degraded. Repairs are
108 * usually performed by rebuilding the metadata entirely out of
109 * redundant metadata. Optimizing, on the other hand, can sometimes be
110 * done without rebuilding entire structures.
111 *
112 * Generally speaking, the repair code has the following code structure:
113 * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
114 * The first check helps us figure out if we need to rebuild or simply
115 * optimize the structure so that the rebuild knows what to do. The
116 * second check evaluates the completeness of the repair; that is what
117 * is reported to userspace.
118 *
119 * A quick note on symbol prefixes:
120 * - "xfs_" are general XFS symbols.
121 * - "xchk_" are symbols related to metadata checking.
122 * - "xrep_" are symbols related to metadata repair.
123 * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
124 */
125
126/*
127 * Scrub probe -- userspace uses this to probe if we're willing to scrub
128 * or repair a given mountpoint. This will be used by xfs_scrub to
129 * probe the kernel's abilities to scrub (and repair) the metadata. We
130 * do this by validating the ioctl inputs from userspace, preparing the
131 * filesystem for a scrub (or a repair) operation, and immediately
132 * returning to userspace. Userspace can use the returned errno and
133 * structure state to decide (in broad terms) if scrub/repair are
134 * supported by the running kernel.
135 */
136static int
137xchk_probe(
138 struct xfs_scrub *sc)
139{
140 int error = 0;
141
142 if (xchk_should_terminate(sc, &error))
143 return error;
144
145 return 0;
146}
147
148/* Scrub setup and teardown */
149
150static inline void
151xchk_fsgates_disable(
152 struct xfs_scrub *sc)
153{
154 if (!(sc->flags & XCHK_FSGATES_ALL))
155 return;
156
157 trace_xchk_fsgates_disable(sc, sc->flags & XCHK_FSGATES_ALL);
158
159 if (sc->flags & XCHK_FSGATES_DRAIN)
160 xfs_drain_wait_disable();
161
162 if (sc->flags & XCHK_FSGATES_QUOTA)
163 xfs_dqtrx_hook_disable();
164
165 if (sc->flags & XCHK_FSGATES_DIRENTS)
166 xfs_dir_hook_disable();
167
168 if (sc->flags & XCHK_FSGATES_RMAP)
169 xfs_rmap_hook_disable();
170
171 sc->flags &= ~XCHK_FSGATES_ALL;
172}
173
174/* Free all the resources and finish the transactions. */
175STATIC int
176xchk_teardown(
177 struct xfs_scrub *sc,
178 int error)
179{
180 xchk_ag_free(sc, &sc->sa);
181 if (sc->tp) {
182 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
183 error = xfs_trans_commit(sc->tp);
184 else
185 xfs_trans_cancel(sc->tp);
186 sc->tp = NULL;
187 }
188 if (sc->ip) {
189 if (sc->ilock_flags)
190 xchk_iunlock(sc, sc->ilock_flags);
191 xchk_irele(sc, sc->ip);
192 sc->ip = NULL;
193 }
194 if (sc->flags & XCHK_HAVE_FREEZE_PROT) {
195 sc->flags &= ~XCHK_HAVE_FREEZE_PROT;
196 mnt_drop_write_file(sc->file);
197 }
198 if (sc->xmbtp) {
199 xmbuf_free(sc->xmbtp);
200 sc->xmbtp = NULL;
201 }
202 if (sc->xfile) {
203 xfile_destroy(sc->xfile);
204 sc->xfile = NULL;
205 }
206 if (sc->buf) {
207 if (sc->buf_cleanup)
208 sc->buf_cleanup(sc->buf);
209 kvfree(sc->buf);
210 sc->buf_cleanup = NULL;
211 sc->buf = NULL;
212 }
213
214 xchk_fsgates_disable(sc);
215 return error;
216}
217
218/* Scrubbing dispatch. */
219
220static const struct xchk_meta_ops meta_scrub_ops[] = {
221 [XFS_SCRUB_TYPE_PROBE] = { /* ioctl presence test */
222 .type = ST_NONE,
223 .setup = xchk_setup_fs,
224 .scrub = xchk_probe,
225 .repair = xrep_probe,
226 },
227 [XFS_SCRUB_TYPE_SB] = { /* superblock */
228 .type = ST_PERAG,
229 .setup = xchk_setup_agheader,
230 .scrub = xchk_superblock,
231 .repair = xrep_superblock,
232 },
233 [XFS_SCRUB_TYPE_AGF] = { /* agf */
234 .type = ST_PERAG,
235 .setup = xchk_setup_agheader,
236 .scrub = xchk_agf,
237 .repair = xrep_agf,
238 },
239 [XFS_SCRUB_TYPE_AGFL]= { /* agfl */
240 .type = ST_PERAG,
241 .setup = xchk_setup_agheader,
242 .scrub = xchk_agfl,
243 .repair = xrep_agfl,
244 },
245 [XFS_SCRUB_TYPE_AGI] = { /* agi */
246 .type = ST_PERAG,
247 .setup = xchk_setup_agheader,
248 .scrub = xchk_agi,
249 .repair = xrep_agi,
250 },
251 [XFS_SCRUB_TYPE_BNOBT] = { /* bnobt */
252 .type = ST_PERAG,
253 .setup = xchk_setup_ag_allocbt,
254 .scrub = xchk_allocbt,
255 .repair = xrep_allocbt,
256 .repair_eval = xrep_revalidate_allocbt,
257 },
258 [XFS_SCRUB_TYPE_CNTBT] = { /* cntbt */
259 .type = ST_PERAG,
260 .setup = xchk_setup_ag_allocbt,
261 .scrub = xchk_allocbt,
262 .repair = xrep_allocbt,
263 .repair_eval = xrep_revalidate_allocbt,
264 },
265 [XFS_SCRUB_TYPE_INOBT] = { /* inobt */
266 .type = ST_PERAG,
267 .setup = xchk_setup_ag_iallocbt,
268 .scrub = xchk_iallocbt,
269 .repair = xrep_iallocbt,
270 .repair_eval = xrep_revalidate_iallocbt,
271 },
272 [XFS_SCRUB_TYPE_FINOBT] = { /* finobt */
273 .type = ST_PERAG,
274 .setup = xchk_setup_ag_iallocbt,
275 .scrub = xchk_iallocbt,
276 .has = xfs_has_finobt,
277 .repair = xrep_iallocbt,
278 .repair_eval = xrep_revalidate_iallocbt,
279 },
280 [XFS_SCRUB_TYPE_RMAPBT] = { /* rmapbt */
281 .type = ST_PERAG,
282 .setup = xchk_setup_ag_rmapbt,
283 .scrub = xchk_rmapbt,
284 .has = xfs_has_rmapbt,
285 .repair = xrep_rmapbt,
286 },
287 [XFS_SCRUB_TYPE_REFCNTBT] = { /* refcountbt */
288 .type = ST_PERAG,
289 .setup = xchk_setup_ag_refcountbt,
290 .scrub = xchk_refcountbt,
291 .has = xfs_has_reflink,
292 .repair = xrep_refcountbt,
293 },
294 [XFS_SCRUB_TYPE_INODE] = { /* inode record */
295 .type = ST_INODE,
296 .setup = xchk_setup_inode,
297 .scrub = xchk_inode,
298 .repair = xrep_inode,
299 },
300 [XFS_SCRUB_TYPE_BMBTD] = { /* inode data fork */
301 .type = ST_INODE,
302 .setup = xchk_setup_inode_bmap,
303 .scrub = xchk_bmap_data,
304 .repair = xrep_bmap_data,
305 },
306 [XFS_SCRUB_TYPE_BMBTA] = { /* inode attr fork */
307 .type = ST_INODE,
308 .setup = xchk_setup_inode_bmap,
309 .scrub = xchk_bmap_attr,
310 .repair = xrep_bmap_attr,
311 },
312 [XFS_SCRUB_TYPE_BMBTC] = { /* inode CoW fork */
313 .type = ST_INODE,
314 .setup = xchk_setup_inode_bmap,
315 .scrub = xchk_bmap_cow,
316 .repair = xrep_bmap_cow,
317 },
318 [XFS_SCRUB_TYPE_DIR] = { /* directory */
319 .type = ST_INODE,
320 .setup = xchk_setup_directory,
321 .scrub = xchk_directory,
322 .repair = xrep_notsupported,
323 },
324 [XFS_SCRUB_TYPE_XATTR] = { /* extended attributes */
325 .type = ST_INODE,
326 .setup = xchk_setup_xattr,
327 .scrub = xchk_xattr,
328 .repair = xrep_notsupported,
329 },
330 [XFS_SCRUB_TYPE_SYMLINK] = { /* symbolic link */
331 .type = ST_INODE,
332 .setup = xchk_setup_symlink,
333 .scrub = xchk_symlink,
334 .repair = xrep_notsupported,
335 },
336 [XFS_SCRUB_TYPE_PARENT] = { /* parent pointers */
337 .type = ST_INODE,
338 .setup = xchk_setup_parent,
339 .scrub = xchk_parent,
340 .repair = xrep_notsupported,
341 },
342 [XFS_SCRUB_TYPE_RTBITMAP] = { /* realtime bitmap */
343 .type = ST_FS,
344 .setup = xchk_setup_rtbitmap,
345 .scrub = xchk_rtbitmap,
346 .repair = xrep_rtbitmap,
347 },
348 [XFS_SCRUB_TYPE_RTSUM] = { /* realtime summary */
349 .type = ST_FS,
350 .setup = xchk_setup_rtsummary,
351 .scrub = xchk_rtsummary,
352 .repair = xrep_notsupported,
353 },
354 [XFS_SCRUB_TYPE_UQUOTA] = { /* user quota */
355 .type = ST_FS,
356 .setup = xchk_setup_quota,
357 .scrub = xchk_quota,
358 .repair = xrep_quota,
359 },
360 [XFS_SCRUB_TYPE_GQUOTA] = { /* group quota */
361 .type = ST_FS,
362 .setup = xchk_setup_quota,
363 .scrub = xchk_quota,
364 .repair = xrep_quota,
365 },
366 [XFS_SCRUB_TYPE_PQUOTA] = { /* project quota */
367 .type = ST_FS,
368 .setup = xchk_setup_quota,
369 .scrub = xchk_quota,
370 .repair = xrep_quota,
371 },
372 [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */
373 .type = ST_FS,
374 .setup = xchk_setup_fscounters,
375 .scrub = xchk_fscounters,
376 .repair = xrep_fscounters,
377 },
378 [XFS_SCRUB_TYPE_QUOTACHECK] = { /* quota counters */
379 .type = ST_FS,
380 .setup = xchk_setup_quotacheck,
381 .scrub = xchk_quotacheck,
382 .repair = xrep_quotacheck,
383 },
384 [XFS_SCRUB_TYPE_NLINKS] = { /* inode link counts */
385 .type = ST_FS,
386 .setup = xchk_setup_nlinks,
387 .scrub = xchk_nlinks,
388 .repair = xrep_nlinks,
389 },
390 [XFS_SCRUB_TYPE_HEALTHY] = { /* fs healthy; clean all reminders */
391 .type = ST_FS,
392 .setup = xchk_setup_fs,
393 .scrub = xchk_health_record,
394 .repair = xrep_notsupported,
395 },
396};
397
398static int
399xchk_validate_inputs(
400 struct xfs_mount *mp,
401 struct xfs_scrub_metadata *sm)
402{
403 int error;
404 const struct xchk_meta_ops *ops;
405
406 error = -EINVAL;
407 /* Check our inputs. */
408 sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
409 if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
410 goto out;
411 /* sm_reserved[] must be zero */
412 if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
413 goto out;
414
415 error = -ENOENT;
416 /* Do we know about this type of metadata? */
417 if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
418 goto out;
419 ops = &meta_scrub_ops[sm->sm_type];
420 if (ops->setup == NULL || ops->scrub == NULL)
421 goto out;
422 /* Does this fs even support this type of metadata? */
423 if (ops->has && !ops->has(mp))
424 goto out;
425
426 error = -EINVAL;
427 /* restricting fields must be appropriate for type */
428 switch (ops->type) {
429 case ST_NONE:
430 case ST_FS:
431 if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
432 goto out;
433 break;
434 case ST_PERAG:
435 if (sm->sm_ino || sm->sm_gen ||
436 sm->sm_agno >= mp->m_sb.sb_agcount)
437 goto out;
438 break;
439 case ST_INODE:
440 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
441 goto out;
442 break;
443 default:
444 goto out;
445 }
446
447 /* No rebuild without repair. */
448 if ((sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) &&
449 !(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
450 return -EINVAL;
451
452 /*
453 * We only want to repair read-write v5+ filesystems. Defer the check
454 * for ops->repair until after our scrub confirms that we need to
455 * perform repairs so that we avoid failing due to not supporting
456 * repairing an object that doesn't need repairs.
457 */
458 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
459 error = -EOPNOTSUPP;
460 if (!xfs_has_crc(mp))
461 goto out;
462
463 error = -EROFS;
464 if (xfs_is_readonly(mp))
465 goto out;
466 }
467
468 error = 0;
469out:
470 return error;
471}
472
473#ifdef CONFIG_XFS_ONLINE_REPAIR
474static inline void xchk_postmortem(struct xfs_scrub *sc)
475{
476 /*
477 * Userspace asked us to repair something, we repaired it, rescanned
478 * it, and the rescan says it's still broken. Scream about this in
479 * the system logs.
480 */
481 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
482 (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
483 XFS_SCRUB_OFLAG_XCORRUPT)))
484 xrep_failure(sc->mp);
485}
486#else
487static inline void xchk_postmortem(struct xfs_scrub *sc)
488{
489 /*
490 * Userspace asked us to scrub something, it's broken, and we have no
491 * way of fixing it. Scream in the logs.
492 */
493 if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
494 XFS_SCRUB_OFLAG_XCORRUPT))
495 xfs_alert_ratelimited(sc->mp,
496 "Corruption detected during scrub.");
497}
498#endif /* CONFIG_XFS_ONLINE_REPAIR */
499
500/* Dispatch metadata scrubbing. */
501int
502xfs_scrub_metadata(
503 struct file *file,
504 struct xfs_scrub_metadata *sm)
505{
506 struct xchk_stats_run run = { };
507 struct xfs_scrub *sc;
508 struct xfs_mount *mp = XFS_I(file_inode(file))->i_mount;
509 u64 check_start;
510 int error = 0;
511
512 BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
513 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
514
515 trace_xchk_start(XFS_I(file_inode(file)), sm, error);
516
517 /* Forbidden if we are shut down or mounted norecovery. */
518 error = -ESHUTDOWN;
519 if (xfs_is_shutdown(mp))
520 goto out;
521 error = -ENOTRECOVERABLE;
522 if (xfs_has_norecovery(mp))
523 goto out;
524
525 error = xchk_validate_inputs(mp, sm);
526 if (error)
527 goto out;
528
529 xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SCRUB,
530 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
531
532 sc = kzalloc(sizeof(struct xfs_scrub), XCHK_GFP_FLAGS);
533 if (!sc) {
534 error = -ENOMEM;
535 goto out;
536 }
537
538 sc->mp = mp;
539 sc->file = file;
540 sc->sm = sm;
541 sc->ops = &meta_scrub_ops[sm->sm_type];
542 sc->sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
543retry_op:
544 /*
545 * When repairs are allowed, prevent freezing or readonly remount while
546 * scrub is running with a real transaction.
547 */
548 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
549 error = mnt_want_write_file(sc->file);
550 if (error)
551 goto out_sc;
552
553 sc->flags |= XCHK_HAVE_FREEZE_PROT;
554 }
555
556 /* Set up for the operation. */
557 error = sc->ops->setup(sc);
558 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
559 goto try_harder;
560 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
561 goto need_drain;
562 if (error)
563 goto out_teardown;
564
565 /* Scrub for errors. */
566 check_start = xchk_stats_now();
567 if ((sc->flags & XREP_ALREADY_FIXED) && sc->ops->repair_eval != NULL)
568 error = sc->ops->repair_eval(sc);
569 else
570 error = sc->ops->scrub(sc);
571 run.scrub_ns += xchk_stats_elapsed_ns(check_start);
572 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
573 goto try_harder;
574 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
575 goto need_drain;
576 if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE))
577 goto out_teardown;
578
579 xchk_update_health(sc);
580
581 if (xchk_could_repair(sc)) {
582 /*
583 * If userspace asked for a repair but it wasn't necessary,
584 * report that back to userspace.
585 */
586 if (!xrep_will_attempt(sc)) {
587 sc->sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
588 goto out_nofix;
589 }
590
591 /*
592 * If it's broken, userspace wants us to fix it, and we haven't
593 * already tried to fix it, then attempt a repair.
594 */
595 error = xrep_attempt(sc, &run);
596 if (error == -EAGAIN) {
597 /*
598 * Either the repair function succeeded or it couldn't
599 * get all the resources it needs; either way, we go
600 * back to the beginning and call the scrub function.
601 */
602 error = xchk_teardown(sc, 0);
603 if (error) {
604 xrep_failure(mp);
605 goto out_sc;
606 }
607 goto retry_op;
608 }
609 }
610
611out_nofix:
612 xchk_postmortem(sc);
613out_teardown:
614 error = xchk_teardown(sc, error);
615out_sc:
616 if (error != -ENOENT)
617 xchk_stats_merge(mp, sm, &run);
618 kfree(sc);
619out:
620 trace_xchk_done(XFS_I(file_inode(file)), sm, error);
621 if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
622 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
623 error = 0;
624 }
625 return error;
626need_drain:
627 error = xchk_teardown(sc, 0);
628 if (error)
629 goto out_sc;
630 sc->flags |= XCHK_NEED_DRAIN;
631 run.retries++;
632 goto retry_op;
633try_harder:
634 /*
635 * Scrubbers return -EDEADLOCK to mean 'try harder'. Tear down
636 * everything we hold, then set up again with preparation for
637 * worst-case scenarios.
638 */
639 error = xchk_teardown(sc, 0);
640 if (error)
641 goto out_sc;
642 sc->flags |= XCHK_TRY_HARDER;
643 run.retries++;
644 goto retry_op;
645}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2017-2023 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <djwong@kernel.org>
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_scrub.h"
18#include "scrub/scrub.h"
19#include "scrub/common.h"
20#include "scrub/trace.h"
21#include "scrub/repair.h"
22#include "scrub/health.h"
23#include "scrub/stats.h"
24#include "scrub/xfile.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 */
134static int
135xchk_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
148static inline void
149xchk_fsgates_disable(
150 struct xfs_scrub *sc)
151{
152 if (!(sc->flags & XCHK_FSGATES_ALL))
153 return;
154
155 trace_xchk_fsgates_disable(sc, sc->flags & XCHK_FSGATES_ALL);
156
157 if (sc->flags & XCHK_FSGATES_DRAIN)
158 xfs_drain_wait_disable();
159
160 sc->flags &= ~XCHK_FSGATES_ALL;
161}
162
163/* Free all the resources and finish the transactions. */
164STATIC int
165xchk_teardown(
166 struct xfs_scrub *sc,
167 int error)
168{
169 xchk_ag_free(sc, &sc->sa);
170 if (sc->tp) {
171 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
172 error = xfs_trans_commit(sc->tp);
173 else
174 xfs_trans_cancel(sc->tp);
175 sc->tp = NULL;
176 }
177 if (sc->ip) {
178 if (sc->ilock_flags)
179 xchk_iunlock(sc, sc->ilock_flags);
180 xchk_irele(sc, sc->ip);
181 sc->ip = NULL;
182 }
183 if (sc->flags & XCHK_HAVE_FREEZE_PROT) {
184 sc->flags &= ~XCHK_HAVE_FREEZE_PROT;
185 mnt_drop_write_file(sc->file);
186 }
187 if (sc->xfile) {
188 xfile_destroy(sc->xfile);
189 sc->xfile = NULL;
190 }
191 if (sc->buf) {
192 if (sc->buf_cleanup)
193 sc->buf_cleanup(sc->buf);
194 kvfree(sc->buf);
195 sc->buf_cleanup = NULL;
196 sc->buf = NULL;
197 }
198
199 xchk_fsgates_disable(sc);
200 return error;
201}
202
203/* Scrubbing dispatch. */
204
205static const struct xchk_meta_ops meta_scrub_ops[] = {
206 [XFS_SCRUB_TYPE_PROBE] = { /* ioctl presence test */
207 .type = ST_NONE,
208 .setup = xchk_setup_fs,
209 .scrub = xchk_probe,
210 .repair = xrep_probe,
211 },
212 [XFS_SCRUB_TYPE_SB] = { /* superblock */
213 .type = ST_PERAG,
214 .setup = xchk_setup_agheader,
215 .scrub = xchk_superblock,
216 .repair = xrep_superblock,
217 },
218 [XFS_SCRUB_TYPE_AGF] = { /* agf */
219 .type = ST_PERAG,
220 .setup = xchk_setup_agheader,
221 .scrub = xchk_agf,
222 .repair = xrep_agf,
223 },
224 [XFS_SCRUB_TYPE_AGFL]= { /* agfl */
225 .type = ST_PERAG,
226 .setup = xchk_setup_agheader,
227 .scrub = xchk_agfl,
228 .repair = xrep_agfl,
229 },
230 [XFS_SCRUB_TYPE_AGI] = { /* agi */
231 .type = ST_PERAG,
232 .setup = xchk_setup_agheader,
233 .scrub = xchk_agi,
234 .repair = xrep_agi,
235 },
236 [XFS_SCRUB_TYPE_BNOBT] = { /* bnobt */
237 .type = ST_PERAG,
238 .setup = xchk_setup_ag_allocbt,
239 .scrub = xchk_allocbt,
240 .repair = xrep_allocbt,
241 .repair_eval = xrep_revalidate_allocbt,
242 },
243 [XFS_SCRUB_TYPE_CNTBT] = { /* cntbt */
244 .type = ST_PERAG,
245 .setup = xchk_setup_ag_allocbt,
246 .scrub = xchk_allocbt,
247 .repair = xrep_allocbt,
248 .repair_eval = xrep_revalidate_allocbt,
249 },
250 [XFS_SCRUB_TYPE_INOBT] = { /* inobt */
251 .type = ST_PERAG,
252 .setup = xchk_setup_ag_iallocbt,
253 .scrub = xchk_iallocbt,
254 .repair = xrep_iallocbt,
255 .repair_eval = xrep_revalidate_iallocbt,
256 },
257 [XFS_SCRUB_TYPE_FINOBT] = { /* finobt */
258 .type = ST_PERAG,
259 .setup = xchk_setup_ag_iallocbt,
260 .scrub = xchk_iallocbt,
261 .has = xfs_has_finobt,
262 .repair = xrep_iallocbt,
263 .repair_eval = xrep_revalidate_iallocbt,
264 },
265 [XFS_SCRUB_TYPE_RMAPBT] = { /* rmapbt */
266 .type = ST_PERAG,
267 .setup = xchk_setup_ag_rmapbt,
268 .scrub = xchk_rmapbt,
269 .has = xfs_has_rmapbt,
270 .repair = xrep_notsupported,
271 },
272 [XFS_SCRUB_TYPE_REFCNTBT] = { /* refcountbt */
273 .type = ST_PERAG,
274 .setup = xchk_setup_ag_refcountbt,
275 .scrub = xchk_refcountbt,
276 .has = xfs_has_reflink,
277 .repair = xrep_refcountbt,
278 },
279 [XFS_SCRUB_TYPE_INODE] = { /* inode record */
280 .type = ST_INODE,
281 .setup = xchk_setup_inode,
282 .scrub = xchk_inode,
283 .repair = xrep_inode,
284 },
285 [XFS_SCRUB_TYPE_BMBTD] = { /* inode data fork */
286 .type = ST_INODE,
287 .setup = xchk_setup_inode_bmap,
288 .scrub = xchk_bmap_data,
289 .repair = xrep_bmap_data,
290 },
291 [XFS_SCRUB_TYPE_BMBTA] = { /* inode attr fork */
292 .type = ST_INODE,
293 .setup = xchk_setup_inode_bmap,
294 .scrub = xchk_bmap_attr,
295 .repair = xrep_bmap_attr,
296 },
297 [XFS_SCRUB_TYPE_BMBTC] = { /* inode CoW fork */
298 .type = ST_INODE,
299 .setup = xchk_setup_inode_bmap,
300 .scrub = xchk_bmap_cow,
301 .repair = xrep_bmap_cow,
302 },
303 [XFS_SCRUB_TYPE_DIR] = { /* directory */
304 .type = ST_INODE,
305 .setup = xchk_setup_directory,
306 .scrub = xchk_directory,
307 .repair = xrep_notsupported,
308 },
309 [XFS_SCRUB_TYPE_XATTR] = { /* extended attributes */
310 .type = ST_INODE,
311 .setup = xchk_setup_xattr,
312 .scrub = xchk_xattr,
313 .repair = xrep_notsupported,
314 },
315 [XFS_SCRUB_TYPE_SYMLINK] = { /* symbolic link */
316 .type = ST_INODE,
317 .setup = xchk_setup_symlink,
318 .scrub = xchk_symlink,
319 .repair = xrep_notsupported,
320 },
321 [XFS_SCRUB_TYPE_PARENT] = { /* parent pointers */
322 .type = ST_INODE,
323 .setup = xchk_setup_parent,
324 .scrub = xchk_parent,
325 .repair = xrep_notsupported,
326 },
327 [XFS_SCRUB_TYPE_RTBITMAP] = { /* realtime bitmap */
328 .type = ST_FS,
329 .setup = xchk_setup_rtbitmap,
330 .scrub = xchk_rtbitmap,
331 .repair = xrep_rtbitmap,
332 },
333 [XFS_SCRUB_TYPE_RTSUM] = { /* realtime summary */
334 .type = ST_FS,
335 .setup = xchk_setup_rtsummary,
336 .scrub = xchk_rtsummary,
337 .repair = xrep_notsupported,
338 },
339 [XFS_SCRUB_TYPE_UQUOTA] = { /* user quota */
340 .type = ST_FS,
341 .setup = xchk_setup_quota,
342 .scrub = xchk_quota,
343 .repair = xrep_quota,
344 },
345 [XFS_SCRUB_TYPE_GQUOTA] = { /* group quota */
346 .type = ST_FS,
347 .setup = xchk_setup_quota,
348 .scrub = xchk_quota,
349 .repair = xrep_quota,
350 },
351 [XFS_SCRUB_TYPE_PQUOTA] = { /* project quota */
352 .type = ST_FS,
353 .setup = xchk_setup_quota,
354 .scrub = xchk_quota,
355 .repair = xrep_quota,
356 },
357 [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */
358 .type = ST_FS,
359 .setup = xchk_setup_fscounters,
360 .scrub = xchk_fscounters,
361 .repair = xrep_notsupported,
362 },
363};
364
365static int
366xchk_validate_inputs(
367 struct xfs_mount *mp,
368 struct xfs_scrub_metadata *sm)
369{
370 int error;
371 const struct xchk_meta_ops *ops;
372
373 error = -EINVAL;
374 /* Check our inputs. */
375 sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
376 if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
377 goto out;
378 /* sm_reserved[] must be zero */
379 if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
380 goto out;
381
382 error = -ENOENT;
383 /* Do we know about this type of metadata? */
384 if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
385 goto out;
386 ops = &meta_scrub_ops[sm->sm_type];
387 if (ops->setup == NULL || ops->scrub == NULL)
388 goto out;
389 /* Does this fs even support this type of metadata? */
390 if (ops->has && !ops->has(mp))
391 goto out;
392
393 error = -EINVAL;
394 /* restricting fields must be appropriate for type */
395 switch (ops->type) {
396 case ST_NONE:
397 case ST_FS:
398 if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
399 goto out;
400 break;
401 case ST_PERAG:
402 if (sm->sm_ino || sm->sm_gen ||
403 sm->sm_agno >= mp->m_sb.sb_agcount)
404 goto out;
405 break;
406 case ST_INODE:
407 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
408 goto out;
409 break;
410 default:
411 goto out;
412 }
413
414 /* No rebuild without repair. */
415 if ((sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) &&
416 !(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
417 return -EINVAL;
418
419 /*
420 * We only want to repair read-write v5+ filesystems. Defer the check
421 * for ops->repair until after our scrub confirms that we need to
422 * perform repairs so that we avoid failing due to not supporting
423 * repairing an object that doesn't need repairs.
424 */
425 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
426 error = -EOPNOTSUPP;
427 if (!xfs_has_crc(mp))
428 goto out;
429
430 error = -EROFS;
431 if (xfs_is_readonly(mp))
432 goto out;
433 }
434
435 error = 0;
436out:
437 return error;
438}
439
440#ifdef CONFIG_XFS_ONLINE_REPAIR
441static inline void xchk_postmortem(struct xfs_scrub *sc)
442{
443 /*
444 * Userspace asked us to repair something, we repaired it, rescanned
445 * it, and the rescan says it's still broken. Scream about this in
446 * the system logs.
447 */
448 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
449 (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
450 XFS_SCRUB_OFLAG_XCORRUPT)))
451 xrep_failure(sc->mp);
452}
453#else
454static inline void xchk_postmortem(struct xfs_scrub *sc)
455{
456 /*
457 * Userspace asked us to scrub something, it's broken, and we have no
458 * way of fixing it. Scream in the logs.
459 */
460 if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
461 XFS_SCRUB_OFLAG_XCORRUPT))
462 xfs_alert_ratelimited(sc->mp,
463 "Corruption detected during scrub.");
464}
465#endif /* CONFIG_XFS_ONLINE_REPAIR */
466
467/* Dispatch metadata scrubbing. */
468int
469xfs_scrub_metadata(
470 struct file *file,
471 struct xfs_scrub_metadata *sm)
472{
473 struct xchk_stats_run run = { };
474 struct xfs_scrub *sc;
475 struct xfs_mount *mp = XFS_I(file_inode(file))->i_mount;
476 u64 check_start;
477 int error = 0;
478
479 BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
480 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
481
482 trace_xchk_start(XFS_I(file_inode(file)), sm, error);
483
484 /* Forbidden if we are shut down or mounted norecovery. */
485 error = -ESHUTDOWN;
486 if (xfs_is_shutdown(mp))
487 goto out;
488 error = -ENOTRECOVERABLE;
489 if (xfs_has_norecovery(mp))
490 goto out;
491
492 error = xchk_validate_inputs(mp, sm);
493 if (error)
494 goto out;
495
496 xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SCRUB,
497 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
498
499 sc = kzalloc(sizeof(struct xfs_scrub), XCHK_GFP_FLAGS);
500 if (!sc) {
501 error = -ENOMEM;
502 goto out;
503 }
504
505 sc->mp = mp;
506 sc->file = file;
507 sc->sm = sm;
508 sc->ops = &meta_scrub_ops[sm->sm_type];
509 sc->sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
510retry_op:
511 /*
512 * When repairs are allowed, prevent freezing or readonly remount while
513 * scrub is running with a real transaction.
514 */
515 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
516 error = mnt_want_write_file(sc->file);
517 if (error)
518 goto out_sc;
519
520 sc->flags |= XCHK_HAVE_FREEZE_PROT;
521 }
522
523 /* Set up for the operation. */
524 error = sc->ops->setup(sc);
525 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
526 goto try_harder;
527 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
528 goto need_drain;
529 if (error)
530 goto out_teardown;
531
532 /* Scrub for errors. */
533 check_start = xchk_stats_now();
534 if ((sc->flags & XREP_ALREADY_FIXED) && sc->ops->repair_eval != NULL)
535 error = sc->ops->repair_eval(sc);
536 else
537 error = sc->ops->scrub(sc);
538 run.scrub_ns += xchk_stats_elapsed_ns(check_start);
539 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
540 goto try_harder;
541 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
542 goto need_drain;
543 if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE))
544 goto out_teardown;
545
546 xchk_update_health(sc);
547
548 if (xchk_could_repair(sc)) {
549 /*
550 * If userspace asked for a repair but it wasn't necessary,
551 * report that back to userspace.
552 */
553 if (!xrep_will_attempt(sc)) {
554 sc->sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
555 goto out_nofix;
556 }
557
558 /*
559 * If it's broken, userspace wants us to fix it, and we haven't
560 * already tried to fix it, then attempt a repair.
561 */
562 error = xrep_attempt(sc, &run);
563 if (error == -EAGAIN) {
564 /*
565 * Either the repair function succeeded or it couldn't
566 * get all the resources it needs; either way, we go
567 * back to the beginning and call the scrub function.
568 */
569 error = xchk_teardown(sc, 0);
570 if (error) {
571 xrep_failure(mp);
572 goto out_sc;
573 }
574 goto retry_op;
575 }
576 }
577
578out_nofix:
579 xchk_postmortem(sc);
580out_teardown:
581 error = xchk_teardown(sc, error);
582out_sc:
583 if (error != -ENOENT)
584 xchk_stats_merge(mp, sm, &run);
585 kfree(sc);
586out:
587 trace_xchk_done(XFS_I(file_inode(file)), sm, error);
588 if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
589 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
590 error = 0;
591 }
592 return error;
593need_drain:
594 error = xchk_teardown(sc, 0);
595 if (error)
596 goto out_sc;
597 sc->flags |= XCHK_NEED_DRAIN;
598 run.retries++;
599 goto retry_op;
600try_harder:
601 /*
602 * Scrubbers return -EDEADLOCK to mean 'try harder'. Tear down
603 * everything we hold, then set up again with preparation for
604 * worst-case scenarios.
605 */
606 error = xchk_teardown(sc, 0);
607 if (error)
608 goto out_sc;
609 sc->flags |= XCHK_TRY_HARDER;
610 run.retries++;
611 goto retry_op;
612}