xref: /linux/fs/xfs/scrub/scrub.c (revision 18f90d372cf35b387663f1567de701e5393f6eb5)
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_defer.h"
13 #include "xfs_btree.h"
14 #include "xfs_bit.h"
15 #include "xfs_log_format.h"
16 #include "xfs_trans.h"
17 #include "xfs_sb.h"
18 #include "xfs_inode.h"
19 #include "xfs_icache.h"
20 #include "xfs_itable.h"
21 #include "xfs_alloc.h"
22 #include "xfs_alloc_btree.h"
23 #include "xfs_bmap.h"
24 #include "xfs_bmap_btree.h"
25 #include "xfs_ialloc.h"
26 #include "xfs_ialloc_btree.h"
27 #include "xfs_refcount.h"
28 #include "xfs_refcount_btree.h"
29 #include "xfs_rmap.h"
30 #include "xfs_rmap_btree.h"
31 #include "xfs_quota.h"
32 #include "xfs_qm.h"
33 #include "xfs_errortag.h"
34 #include "xfs_error.h"
35 #include "xfs_log.h"
36 #include "xfs_trans_priv.h"
37 #include "scrub/xfs_scrub.h"
38 #include "scrub/scrub.h"
39 #include "scrub/common.h"
40 #include "scrub/trace.h"
41 #include "scrub/btree.h"
42 #include "scrub/repair.h"
43 
44 /*
45  * Online Scrub and Repair
46  *
47  * Traditionally, XFS (the kernel driver) did not know how to check or
48  * repair on-disk data structures.  That task was left to the xfs_check
49  * and xfs_repair tools, both of which require taking the filesystem
50  * offline for a thorough but time consuming examination.  Online
51  * scrub & repair, on the other hand, enables us to check the metadata
52  * for obvious errors while carefully stepping around the filesystem's
53  * ongoing operations, locking rules, etc.
54  *
55  * Given that most XFS metadata consist of records stored in a btree,
56  * most of the checking functions iterate the btree blocks themselves
57  * looking for irregularities.  When a record block is encountered, each
58  * record can be checked for obviously bad values.  Record values can
59  * also be cross-referenced against other btrees to look for potential
60  * misunderstandings between pieces of metadata.
61  *
62  * It is expected that the checkers responsible for per-AG metadata
63  * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
64  * metadata structure, and perform any relevant cross-referencing before
65  * unlocking the AG and returning the results to userspace.  These
66  * scrubbers must not keep an AG locked for too long to avoid tying up
67  * the block and inode allocators.
68  *
69  * Block maps and b-trees rooted in an inode present a special challenge
70  * because they can involve extents from any AG.  The general scrubber
71  * structure of lock -> check -> xref -> unlock still holds, but AG
72  * locking order rules /must/ be obeyed to avoid deadlocks.  The
73  * ordering rule, of course, is that we must lock in increasing AG
74  * order.  Helper functions are provided to track which AG headers we've
75  * already locked.  If we detect an imminent locking order violation, we
76  * can signal a potential deadlock, in which case the scrubber can jump
77  * out to the top level, lock all the AGs in order, and retry the scrub.
78  *
79  * For file data (directories, extended attributes, symlinks) scrub, we
80  * can simply lock the inode and walk the data.  For btree data
81  * (directories and attributes) we follow the same btree-scrubbing
82  * strategy outlined previously to check the records.
83  *
84  * We use a bit of trickery with transactions to avoid buffer deadlocks
85  * if there is a cycle in the metadata.  The basic problem is that
86  * travelling down a btree involves locking the current buffer at each
87  * tree level.  If a pointer should somehow point back to a buffer that
88  * we've already examined, we will deadlock due to the second buffer
89  * locking attempt.  Note however that grabbing a buffer in transaction
90  * context links the locked buffer to the transaction.  If we try to
91  * re-grab the buffer in the context of the same transaction, we avoid
92  * the second lock attempt and continue.  Between the verifier and the
93  * scrubber, something will notice that something is amiss and report
94  * the corruption.  Therefore, each scrubber will allocate an empty
95  * transaction, attach buffers to it, and cancel the transaction at the
96  * end of the scrub run.  Cancelling a non-dirty transaction simply
97  * unlocks the buffers.
98  *
99  * There are four pieces of data that scrub can communicate to
100  * userspace.  The first is the error code (errno), which can be used to
101  * communicate operational errors in performing the scrub.  There are
102  * also three flags that can be set in the scrub context.  If the data
103  * structure itself is corrupt, the CORRUPT flag will be set.  If
104  * the metadata is correct but otherwise suboptimal, the PREEN flag
105  * will be set.
106  *
107  * We perform secondary validation of filesystem metadata by
108  * cross-referencing every record with all other available metadata.
109  * For example, for block mapping extents, we verify that there are no
110  * records in the free space and inode btrees corresponding to that
111  * space extent and that there is a corresponding entry in the reverse
112  * mapping btree.  Inconsistent metadata is noted by setting the
113  * XCORRUPT flag; btree query function errors are noted by setting the
114  * XFAIL flag and deleting the cursor to prevent further attempts to
115  * cross-reference with a defective btree.
116  *
117  * If a piece of metadata proves corrupt or suboptimal, the userspace
118  * program can ask the kernel to apply some tender loving care (TLC) to
119  * the metadata object by setting the REPAIR flag and re-calling the
120  * scrub ioctl.  "Corruption" is defined by metadata violating the
121  * on-disk specification; operations cannot continue if the violation is
122  * left untreated.  It is possible for XFS to continue if an object is
123  * "suboptimal", however performance may be degraded.  Repairs are
124  * usually performed by rebuilding the metadata entirely out of
125  * redundant metadata.  Optimizing, on the other hand, can sometimes be
126  * done without rebuilding entire structures.
127  *
128  * Generally speaking, the repair code has the following code structure:
129  * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
130  * The first check helps us figure out if we need to rebuild or simply
131  * optimize the structure so that the rebuild knows what to do.  The
132  * second check evaluates the completeness of the repair; that is what
133  * is reported to userspace.
134  *
135  * A quick note on symbol prefixes:
136  * - "xfs_" are general XFS symbols.
137  * - "xchk_" are symbols related to metadata checking.
138  * - "xrep_" are symbols related to metadata repair.
139  * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
140  */
141 
142 /*
143  * Scrub probe -- userspace uses this to probe if we're willing to scrub
144  * or repair a given mountpoint.  This will be used by xfs_scrub to
145  * probe the kernel's abilities to scrub (and repair) the metadata.  We
146  * do this by validating the ioctl inputs from userspace, preparing the
147  * filesystem for a scrub (or a repair) operation, and immediately
148  * returning to userspace.  Userspace can use the returned errno and
149  * structure state to decide (in broad terms) if scrub/repair are
150  * supported by the running kernel.
151  */
152 static int
153 xchk_probe(
154 	struct xfs_scrub	*sc)
155 {
156 	int			error = 0;
157 
158 	if (xchk_should_terminate(sc, &error))
159 		return error;
160 
161 	return 0;
162 }
163 
164 /* Scrub setup and teardown */
165 
166 /* Free all the resources and finish the transactions. */
167 STATIC int
168 xchk_teardown(
169 	struct xfs_scrub	*sc,
170 	struct xfs_inode	*ip_in,
171 	int			error)
172 {
173 	xchk_ag_free(sc, &sc->sa);
174 	if (sc->tp) {
175 		if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
176 			error = xfs_trans_commit(sc->tp);
177 		else
178 			xfs_trans_cancel(sc->tp);
179 		sc->tp = NULL;
180 	}
181 	if (sc->ip) {
182 		if (sc->ilock_flags)
183 			xfs_iunlock(sc->ip, sc->ilock_flags);
184 		if (sc->ip != ip_in &&
185 		    !xfs_internal_inum(sc->mp, sc->ip->i_ino))
186 			xfs_irele(sc->ip);
187 		sc->ip = NULL;
188 	}
189 	if (sc->has_quotaofflock)
190 		mutex_unlock(&sc->mp->m_quotainfo->qi_quotaofflock);
191 	if (sc->buf) {
192 		kmem_free(sc->buf);
193 		sc->buf = NULL;
194 	}
195 	return error;
196 }
197 
198 /* Scrubbing dispatch. */
199 
200 static const struct xchk_meta_ops meta_scrub_ops[] = {
201 	[XFS_SCRUB_TYPE_PROBE] = {	/* ioctl presence test */
202 		.type	= ST_NONE,
203 		.setup	= xchk_setup_fs,
204 		.scrub	= xchk_probe,
205 		.repair = xrep_probe,
206 	},
207 	[XFS_SCRUB_TYPE_SB] = {		/* superblock */
208 		.type	= ST_PERAG,
209 		.setup	= xchk_setup_fs,
210 		.scrub	= xchk_superblock,
211 		.repair	= xrep_superblock,
212 	},
213 	[XFS_SCRUB_TYPE_AGF] = {	/* agf */
214 		.type	= ST_PERAG,
215 		.setup	= xchk_setup_fs,
216 		.scrub	= xchk_agf,
217 		.repair	= xrep_agf,
218 	},
219 	[XFS_SCRUB_TYPE_AGFL]= {	/* agfl */
220 		.type	= ST_PERAG,
221 		.setup	= xchk_setup_fs,
222 		.scrub	= xchk_agfl,
223 		.repair	= xrep_agfl,
224 	},
225 	[XFS_SCRUB_TYPE_AGI] = {	/* agi */
226 		.type	= ST_PERAG,
227 		.setup	= xchk_setup_fs,
228 		.scrub	= xchk_agi,
229 		.repair	= xrep_agi,
230 	},
231 	[XFS_SCRUB_TYPE_BNOBT] = {	/* bnobt */
232 		.type	= ST_PERAG,
233 		.setup	= xchk_setup_ag_allocbt,
234 		.scrub	= xchk_bnobt,
235 		.repair	= xrep_notsupported,
236 	},
237 	[XFS_SCRUB_TYPE_CNTBT] = {	/* cntbt */
238 		.type	= ST_PERAG,
239 		.setup	= xchk_setup_ag_allocbt,
240 		.scrub	= xchk_cntbt,
241 		.repair	= xrep_notsupported,
242 	},
243 	[XFS_SCRUB_TYPE_INOBT] = {	/* inobt */
244 		.type	= ST_PERAG,
245 		.setup	= xchk_setup_ag_iallocbt,
246 		.scrub	= xchk_inobt,
247 		.repair	= xrep_notsupported,
248 	},
249 	[XFS_SCRUB_TYPE_FINOBT] = {	/* finobt */
250 		.type	= ST_PERAG,
251 		.setup	= xchk_setup_ag_iallocbt,
252 		.scrub	= xchk_finobt,
253 		.has	= xfs_sb_version_hasfinobt,
254 		.repair	= xrep_notsupported,
255 	},
256 	[XFS_SCRUB_TYPE_RMAPBT] = {	/* rmapbt */
257 		.type	= ST_PERAG,
258 		.setup	= xchk_setup_ag_rmapbt,
259 		.scrub	= xchk_rmapbt,
260 		.has	= xfs_sb_version_hasrmapbt,
261 		.repair	= xrep_notsupported,
262 	},
263 	[XFS_SCRUB_TYPE_REFCNTBT] = {	/* refcountbt */
264 		.type	= ST_PERAG,
265 		.setup	= xchk_setup_ag_refcountbt,
266 		.scrub	= xchk_refcountbt,
267 		.has	= xfs_sb_version_hasreflink,
268 		.repair	= xrep_notsupported,
269 	},
270 	[XFS_SCRUB_TYPE_INODE] = {	/* inode record */
271 		.type	= ST_INODE,
272 		.setup	= xchk_setup_inode,
273 		.scrub	= xchk_inode,
274 		.repair	= xrep_notsupported,
275 	},
276 	[XFS_SCRUB_TYPE_BMBTD] = {	/* inode data fork */
277 		.type	= ST_INODE,
278 		.setup	= xchk_setup_inode_bmap,
279 		.scrub	= xchk_bmap_data,
280 		.repair	= xrep_notsupported,
281 	},
282 	[XFS_SCRUB_TYPE_BMBTA] = {	/* inode attr fork */
283 		.type	= ST_INODE,
284 		.setup	= xchk_setup_inode_bmap,
285 		.scrub	= xchk_bmap_attr,
286 		.repair	= xrep_notsupported,
287 	},
288 	[XFS_SCRUB_TYPE_BMBTC] = {	/* inode CoW fork */
289 		.type	= ST_INODE,
290 		.setup	= xchk_setup_inode_bmap,
291 		.scrub	= xchk_bmap_cow,
292 		.repair	= xrep_notsupported,
293 	},
294 	[XFS_SCRUB_TYPE_DIR] = {	/* directory */
295 		.type	= ST_INODE,
296 		.setup	= xchk_setup_directory,
297 		.scrub	= xchk_directory,
298 		.repair	= xrep_notsupported,
299 	},
300 	[XFS_SCRUB_TYPE_XATTR] = {	/* extended attributes */
301 		.type	= ST_INODE,
302 		.setup	= xchk_setup_xattr,
303 		.scrub	= xchk_xattr,
304 		.repair	= xrep_notsupported,
305 	},
306 	[XFS_SCRUB_TYPE_SYMLINK] = {	/* symbolic link */
307 		.type	= ST_INODE,
308 		.setup	= xchk_setup_symlink,
309 		.scrub	= xchk_symlink,
310 		.repair	= xrep_notsupported,
311 	},
312 	[XFS_SCRUB_TYPE_PARENT] = {	/* parent pointers */
313 		.type	= ST_INODE,
314 		.setup	= xchk_setup_parent,
315 		.scrub	= xchk_parent,
316 		.repair	= xrep_notsupported,
317 	},
318 	[XFS_SCRUB_TYPE_RTBITMAP] = {	/* realtime bitmap */
319 		.type	= ST_FS,
320 		.setup	= xchk_setup_rt,
321 		.scrub	= xchk_rtbitmap,
322 		.has	= xfs_sb_version_hasrealtime,
323 		.repair	= xrep_notsupported,
324 	},
325 	[XFS_SCRUB_TYPE_RTSUM] = {	/* realtime summary */
326 		.type	= ST_FS,
327 		.setup	= xchk_setup_rt,
328 		.scrub	= xchk_rtsummary,
329 		.has	= xfs_sb_version_hasrealtime,
330 		.repair	= xrep_notsupported,
331 	},
332 	[XFS_SCRUB_TYPE_UQUOTA] = {	/* user quota */
333 		.type	= ST_FS,
334 		.setup	= xchk_setup_quota,
335 		.scrub	= xchk_quota,
336 		.repair	= xrep_notsupported,
337 	},
338 	[XFS_SCRUB_TYPE_GQUOTA] = {	/* group quota */
339 		.type	= ST_FS,
340 		.setup	= xchk_setup_quota,
341 		.scrub	= xchk_quota,
342 		.repair	= xrep_notsupported,
343 	},
344 	[XFS_SCRUB_TYPE_PQUOTA] = {	/* project quota */
345 		.type	= ST_FS,
346 		.setup	= xchk_setup_quota,
347 		.scrub	= xchk_quota,
348 		.repair	= xrep_notsupported,
349 	},
350 };
351 
352 /* This isn't a stable feature, warn once per day. */
353 static inline void
354 xchk_experimental_warning(
355 	struct xfs_mount	*mp)
356 {
357 	static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
358 			"xchk_warning", 86400 * HZ, 1);
359 	ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
360 
361 	if (__ratelimit(&scrub_warning))
362 		xfs_alert(mp,
363 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
364 }
365 
366 static int
367 xchk_validate_inputs(
368 	struct xfs_mount		*mp,
369 	struct xfs_scrub_metadata	*sm)
370 {
371 	int				error;
372 	const struct xchk_meta_ops	*ops;
373 
374 	error = -EINVAL;
375 	/* Check our inputs. */
376 	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
377 	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
378 		goto out;
379 	/* sm_reserved[] must be zero */
380 	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
381 		goto out;
382 
383 	error = -ENOENT;
384 	/* Do we know about this type of metadata? */
385 	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
386 		goto out;
387 	ops = &meta_scrub_ops[sm->sm_type];
388 	if (ops->setup == NULL || ops->scrub == NULL)
389 		goto out;
390 	/* Does this fs even support this type of metadata? */
391 	if (ops->has && !ops->has(&mp->m_sb))
392 		goto out;
393 
394 	error = -EINVAL;
395 	/* restricting fields must be appropriate for type */
396 	switch (ops->type) {
397 	case ST_NONE:
398 	case ST_FS:
399 		if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
400 			goto out;
401 		break;
402 	case ST_PERAG:
403 		if (sm->sm_ino || sm->sm_gen ||
404 		    sm->sm_agno >= mp->m_sb.sb_agcount)
405 			goto out;
406 		break;
407 	case ST_INODE:
408 		if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
409 			goto out;
410 		break;
411 	default:
412 		goto out;
413 	}
414 
415 	/*
416 	 * We only want to repair read-write v5+ filesystems.  Defer the check
417 	 * for ops->repair until after our scrub confirms that we need to
418 	 * perform repairs so that we avoid failing due to not supporting
419 	 * repairing an object that doesn't need repairs.
420 	 */
421 	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
422 		error = -EOPNOTSUPP;
423 		if (!xfs_sb_version_hascrc(&mp->m_sb))
424 			goto out;
425 
426 		error = -EROFS;
427 		if (mp->m_flags & XFS_MOUNT_RDONLY)
428 			goto out;
429 	}
430 
431 	error = 0;
432 out:
433 	return error;
434 }
435 
436 #ifdef CONFIG_XFS_ONLINE_REPAIR
437 static inline void xchk_postmortem(struct xfs_scrub *sc)
438 {
439 	/*
440 	 * Userspace asked us to repair something, we repaired it, rescanned
441 	 * it, and the rescan says it's still broken.  Scream about this in
442 	 * the system logs.
443 	 */
444 	if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
445 	    (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
446 				 XFS_SCRUB_OFLAG_XCORRUPT)))
447 		xrep_failure(sc->mp);
448 }
449 #else
450 static inline void xchk_postmortem(struct xfs_scrub *sc)
451 {
452 	/*
453 	 * Userspace asked us to scrub something, it's broken, and we have no
454 	 * way of fixing it.  Scream in the logs.
455 	 */
456 	if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
457 				XFS_SCRUB_OFLAG_XCORRUPT))
458 		xfs_alert_ratelimited(sc->mp,
459 				"Corruption detected during scrub.");
460 }
461 #endif /* CONFIG_XFS_ONLINE_REPAIR */
462 
463 /* Dispatch metadata scrubbing. */
464 int
465 xfs_scrub_metadata(
466 	struct xfs_inode		*ip,
467 	struct xfs_scrub_metadata	*sm)
468 {
469 	struct xfs_scrub		sc;
470 	struct xfs_mount		*mp = ip->i_mount;
471 	bool				try_harder = false;
472 	bool				already_fixed = false;
473 	int				error = 0;
474 
475 	BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
476 		(sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
477 
478 	trace_xchk_start(ip, sm, error);
479 
480 	/* Forbidden if we are shut down or mounted norecovery. */
481 	error = -ESHUTDOWN;
482 	if (XFS_FORCED_SHUTDOWN(mp))
483 		goto out;
484 	error = -ENOTRECOVERABLE;
485 	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
486 		goto out;
487 
488 	error = xchk_validate_inputs(mp, sm);
489 	if (error)
490 		goto out;
491 
492 	xchk_experimental_warning(mp);
493 
494 retry_op:
495 	/* Set up for the operation. */
496 	memset(&sc, 0, sizeof(sc));
497 	sc.mp = ip->i_mount;
498 	sc.sm = sm;
499 	sc.ops = &meta_scrub_ops[sm->sm_type];
500 	sc.try_harder = try_harder;
501 	sc.sa.agno = NULLAGNUMBER;
502 	error = sc.ops->setup(&sc, ip);
503 	if (error)
504 		goto out_teardown;
505 
506 	/* Scrub for errors. */
507 	error = sc.ops->scrub(&sc);
508 	if (!try_harder && error == -EDEADLOCK) {
509 		/*
510 		 * Scrubbers return -EDEADLOCK to mean 'try harder'.
511 		 * Tear down everything we hold, then set up again with
512 		 * preparation for worst-case scenarios.
513 		 */
514 		error = xchk_teardown(&sc, ip, 0);
515 		if (error)
516 			goto out;
517 		try_harder = true;
518 		goto retry_op;
519 	} else if (error)
520 		goto out_teardown;
521 
522 	if ((sc.sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) && !already_fixed) {
523 		bool needs_fix;
524 
525 		/* Let debug users force us into the repair routines. */
526 		if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR))
527 			sc.sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
528 
529 		needs_fix = (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
530 						XFS_SCRUB_OFLAG_XCORRUPT |
531 						XFS_SCRUB_OFLAG_PREEN));
532 		/*
533 		 * If userspace asked for a repair but it wasn't necessary,
534 		 * report that back to userspace.
535 		 */
536 		if (!needs_fix) {
537 			sc.sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
538 			goto out_nofix;
539 		}
540 
541 		/*
542 		 * If it's broken, userspace wants us to fix it, and we haven't
543 		 * already tried to fix it, then attempt a repair.
544 		 */
545 		error = xrep_attempt(ip, &sc, &already_fixed);
546 		if (error == -EAGAIN) {
547 			if (sc.try_harder)
548 				try_harder = true;
549 			error = xchk_teardown(&sc, ip, 0);
550 			if (error) {
551 				xrep_failure(mp);
552 				goto out;
553 			}
554 			goto retry_op;
555 		}
556 	}
557 
558 out_nofix:
559 	xchk_postmortem(&sc);
560 out_teardown:
561 	error = xchk_teardown(&sc, ip, error);
562 out:
563 	trace_xchk_done(ip, sm, error);
564 	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
565 		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
566 		error = 0;
567 	}
568 	return error;
569 }
570