scrub.c revision 75c8c50f
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 */
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
148/* Free all the resources and finish the transactions. */
149STATIC int
150xchk_teardown(
151	struct xfs_scrub	*sc,
152	int			error)
153{
154	struct xfs_inode	*ip_in = XFS_I(file_inode(sc->file));
155
156	xchk_ag_free(sc, &sc->sa);
157	if (sc->tp) {
158		if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
159			error = xfs_trans_commit(sc->tp);
160		else
161			xfs_trans_cancel(sc->tp);
162		sc->tp = NULL;
163	}
164	if (sc->ip) {
165		if (sc->ilock_flags)
166			xfs_iunlock(sc->ip, sc->ilock_flags);
167		if (sc->ip != ip_in &&
168		    !xfs_internal_inum(sc->mp, sc->ip->i_ino))
169			xfs_irele(sc->ip);
170		sc->ip = NULL;
171	}
172	if (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR)
173		mnt_drop_write_file(sc->file);
174	if (sc->flags & XCHK_REAPING_DISABLED)
175		xchk_start_reaping(sc);
176	if (sc->flags & XCHK_HAS_QUOTAOFFLOCK) {
177		mutex_unlock(&sc->mp->m_quotainfo->qi_quotaofflock);
178		sc->flags &= ~XCHK_HAS_QUOTAOFFLOCK;
179	}
180	if (sc->buf) {
181		kmem_free(sc->buf);
182		sc->buf = NULL;
183	}
184	return error;
185}
186
187/* Scrubbing dispatch. */
188
189static const struct xchk_meta_ops meta_scrub_ops[] = {
190	[XFS_SCRUB_TYPE_PROBE] = {	/* ioctl presence test */
191		.type	= ST_NONE,
192		.setup	= xchk_setup_fs,
193		.scrub	= xchk_probe,
194		.repair = xrep_probe,
195	},
196	[XFS_SCRUB_TYPE_SB] = {		/* superblock */
197		.type	= ST_PERAG,
198		.setup	= xchk_setup_fs,
199		.scrub	= xchk_superblock,
200		.repair	= xrep_superblock,
201	},
202	[XFS_SCRUB_TYPE_AGF] = {	/* agf */
203		.type	= ST_PERAG,
204		.setup	= xchk_setup_fs,
205		.scrub	= xchk_agf,
206		.repair	= xrep_agf,
207	},
208	[XFS_SCRUB_TYPE_AGFL]= {	/* agfl */
209		.type	= ST_PERAG,
210		.setup	= xchk_setup_fs,
211		.scrub	= xchk_agfl,
212		.repair	= xrep_agfl,
213	},
214	[XFS_SCRUB_TYPE_AGI] = {	/* agi */
215		.type	= ST_PERAG,
216		.setup	= xchk_setup_fs,
217		.scrub	= xchk_agi,
218		.repair	= xrep_agi,
219	},
220	[XFS_SCRUB_TYPE_BNOBT] = {	/* bnobt */
221		.type	= ST_PERAG,
222		.setup	= xchk_setup_ag_allocbt,
223		.scrub	= xchk_bnobt,
224		.repair	= xrep_notsupported,
225	},
226	[XFS_SCRUB_TYPE_CNTBT] = {	/* cntbt */
227		.type	= ST_PERAG,
228		.setup	= xchk_setup_ag_allocbt,
229		.scrub	= xchk_cntbt,
230		.repair	= xrep_notsupported,
231	},
232	[XFS_SCRUB_TYPE_INOBT] = {	/* inobt */
233		.type	= ST_PERAG,
234		.setup	= xchk_setup_ag_iallocbt,
235		.scrub	= xchk_inobt,
236		.repair	= xrep_notsupported,
237	},
238	[XFS_SCRUB_TYPE_FINOBT] = {	/* finobt */
239		.type	= ST_PERAG,
240		.setup	= xchk_setup_ag_iallocbt,
241		.scrub	= xchk_finobt,
242		.has	= xfs_sb_version_hasfinobt,
243		.repair	= xrep_notsupported,
244	},
245	[XFS_SCRUB_TYPE_RMAPBT] = {	/* rmapbt */
246		.type	= ST_PERAG,
247		.setup	= xchk_setup_ag_rmapbt,
248		.scrub	= xchk_rmapbt,
249		.has	= xfs_sb_version_hasrmapbt,
250		.repair	= xrep_notsupported,
251	},
252	[XFS_SCRUB_TYPE_REFCNTBT] = {	/* refcountbt */
253		.type	= ST_PERAG,
254		.setup	= xchk_setup_ag_refcountbt,
255		.scrub	= xchk_refcountbt,
256		.has	= xfs_sb_version_hasreflink,
257		.repair	= xrep_notsupported,
258	},
259	[XFS_SCRUB_TYPE_INODE] = {	/* inode record */
260		.type	= ST_INODE,
261		.setup	= xchk_setup_inode,
262		.scrub	= xchk_inode,
263		.repair	= xrep_notsupported,
264	},
265	[XFS_SCRUB_TYPE_BMBTD] = {	/* inode data fork */
266		.type	= ST_INODE,
267		.setup	= xchk_setup_inode_bmap,
268		.scrub	= xchk_bmap_data,
269		.repair	= xrep_notsupported,
270	},
271	[XFS_SCRUB_TYPE_BMBTA] = {	/* inode attr fork */
272		.type	= ST_INODE,
273		.setup	= xchk_setup_inode_bmap,
274		.scrub	= xchk_bmap_attr,
275		.repair	= xrep_notsupported,
276	},
277	[XFS_SCRUB_TYPE_BMBTC] = {	/* inode CoW fork */
278		.type	= ST_INODE,
279		.setup	= xchk_setup_inode_bmap,
280		.scrub	= xchk_bmap_cow,
281		.repair	= xrep_notsupported,
282	},
283	[XFS_SCRUB_TYPE_DIR] = {	/* directory */
284		.type	= ST_INODE,
285		.setup	= xchk_setup_directory,
286		.scrub	= xchk_directory,
287		.repair	= xrep_notsupported,
288	},
289	[XFS_SCRUB_TYPE_XATTR] = {	/* extended attributes */
290		.type	= ST_INODE,
291		.setup	= xchk_setup_xattr,
292		.scrub	= xchk_xattr,
293		.repair	= xrep_notsupported,
294	},
295	[XFS_SCRUB_TYPE_SYMLINK] = {	/* symbolic link */
296		.type	= ST_INODE,
297		.setup	= xchk_setup_symlink,
298		.scrub	= xchk_symlink,
299		.repair	= xrep_notsupported,
300	},
301	[XFS_SCRUB_TYPE_PARENT] = {	/* parent pointers */
302		.type	= ST_INODE,
303		.setup	= xchk_setup_parent,
304		.scrub	= xchk_parent,
305		.repair	= xrep_notsupported,
306	},
307	[XFS_SCRUB_TYPE_RTBITMAP] = {	/* realtime bitmap */
308		.type	= ST_FS,
309		.setup	= xchk_setup_rt,
310		.scrub	= xchk_rtbitmap,
311		.has	= xfs_sb_version_hasrealtime,
312		.repair	= xrep_notsupported,
313	},
314	[XFS_SCRUB_TYPE_RTSUM] = {	/* realtime summary */
315		.type	= ST_FS,
316		.setup	= xchk_setup_rt,
317		.scrub	= xchk_rtsummary,
318		.has	= xfs_sb_version_hasrealtime,
319		.repair	= xrep_notsupported,
320	},
321	[XFS_SCRUB_TYPE_UQUOTA] = {	/* user quota */
322		.type	= ST_FS,
323		.setup	= xchk_setup_quota,
324		.scrub	= xchk_quota,
325		.repair	= xrep_notsupported,
326	},
327	[XFS_SCRUB_TYPE_GQUOTA] = {	/* group quota */
328		.type	= ST_FS,
329		.setup	= xchk_setup_quota,
330		.scrub	= xchk_quota,
331		.repair	= xrep_notsupported,
332	},
333	[XFS_SCRUB_TYPE_PQUOTA] = {	/* project quota */
334		.type	= ST_FS,
335		.setup	= xchk_setup_quota,
336		.scrub	= xchk_quota,
337		.repair	= xrep_notsupported,
338	},
339	[XFS_SCRUB_TYPE_FSCOUNTERS] = {	/* fs summary counters */
340		.type	= ST_FS,
341		.setup	= xchk_setup_fscounters,
342		.scrub	= xchk_fscounters,
343		.repair	= xrep_notsupported,
344	},
345};
346
347/* This isn't a stable feature, warn once per day. */
348static inline void
349xchk_experimental_warning(
350	struct xfs_mount	*mp)
351{
352	static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
353			"xchk_warning", 86400 * HZ, 1);
354	ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
355
356	if (__ratelimit(&scrub_warning))
357		xfs_alert(mp,
358"EXPERIMENTAL online scrub feature in use. Use at your own risk!");
359}
360
361static int
362xchk_validate_inputs(
363	struct xfs_mount		*mp,
364	struct xfs_scrub_metadata	*sm)
365{
366	int				error;
367	const struct xchk_meta_ops	*ops;
368
369	error = -EINVAL;
370	/* Check our inputs. */
371	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
372	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
373		goto out;
374	/* sm_reserved[] must be zero */
375	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
376		goto out;
377
378	error = -ENOENT;
379	/* Do we know about this type of metadata? */
380	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
381		goto out;
382	ops = &meta_scrub_ops[sm->sm_type];
383	if (ops->setup == NULL || ops->scrub == NULL)
384		goto out;
385	/* Does this fs even support this type of metadata? */
386	if (ops->has && !ops->has(&mp->m_sb))
387		goto out;
388
389	error = -EINVAL;
390	/* restricting fields must be appropriate for type */
391	switch (ops->type) {
392	case ST_NONE:
393	case ST_FS:
394		if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
395			goto out;
396		break;
397	case ST_PERAG:
398		if (sm->sm_ino || sm->sm_gen ||
399		    sm->sm_agno >= mp->m_sb.sb_agcount)
400			goto out;
401		break;
402	case ST_INODE:
403		if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
404			goto out;
405		break;
406	default:
407		goto out;
408	}
409
410	/*
411	 * We only want to repair read-write v5+ filesystems.  Defer the check
412	 * for ops->repair until after our scrub confirms that we need to
413	 * perform repairs so that we avoid failing due to not supporting
414	 * repairing an object that doesn't need repairs.
415	 */
416	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
417		error = -EOPNOTSUPP;
418		if (!xfs_has_crc(mp))
419			goto out;
420
421		error = -EROFS;
422		if (xfs_is_readonly(mp))
423			goto out;
424	}
425
426	error = 0;
427out:
428	return error;
429}
430
431#ifdef CONFIG_XFS_ONLINE_REPAIR
432static inline void xchk_postmortem(struct xfs_scrub *sc)
433{
434	/*
435	 * Userspace asked us to repair something, we repaired it, rescanned
436	 * it, and the rescan says it's still broken.  Scream about this in
437	 * the system logs.
438	 */
439	if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
440	    (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
441				 XFS_SCRUB_OFLAG_XCORRUPT)))
442		xrep_failure(sc->mp);
443}
444#else
445static inline void xchk_postmortem(struct xfs_scrub *sc)
446{
447	/*
448	 * Userspace asked us to scrub something, it's broken, and we have no
449	 * way of fixing it.  Scream in the logs.
450	 */
451	if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
452				XFS_SCRUB_OFLAG_XCORRUPT))
453		xfs_alert_ratelimited(sc->mp,
454				"Corruption detected during scrub.");
455}
456#endif /* CONFIG_XFS_ONLINE_REPAIR */
457
458/* Dispatch metadata scrubbing. */
459int
460xfs_scrub_metadata(
461	struct file			*file,
462	struct xfs_scrub_metadata	*sm)
463{
464	struct xfs_scrub		sc = {
465		.file			= file,
466		.sm			= sm,
467	};
468	struct xfs_mount		*mp = XFS_I(file_inode(file))->i_mount;
469	int				error = 0;
470
471	sc.mp = mp;
472
473	BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
474		(sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
475
476	trace_xchk_start(XFS_I(file_inode(file)), sm, error);
477
478	/* Forbidden if we are shut down or mounted norecovery. */
479	error = -ESHUTDOWN;
480	if (xfs_is_shutdown(mp))
481		goto out;
482	error = -ENOTRECOVERABLE;
483	if (xfs_has_norecovery(mp))
484		goto out;
485
486	error = xchk_validate_inputs(mp, sm);
487	if (error)
488		goto out;
489
490	xchk_experimental_warning(mp);
491
492	sc.ops = &meta_scrub_ops[sm->sm_type];
493	sc.sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
494retry_op:
495	/*
496	 * When repairs are allowed, prevent freezing or readonly remount while
497	 * scrub is running with a real transaction.
498	 */
499	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
500		error = mnt_want_write_file(sc.file);
501		if (error)
502			goto out;
503	}
504
505	/* Set up for the operation. */
506	error = sc.ops->setup(&sc);
507	if (error)
508		goto out_teardown;
509
510	/* Scrub for errors. */
511	error = sc.ops->scrub(&sc);
512	if (!(sc.flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
513		/*
514		 * Scrubbers return -EDEADLOCK to mean 'try harder'.
515		 * Tear down everything we hold, then set up again with
516		 * preparation for worst-case scenarios.
517		 */
518		error = xchk_teardown(&sc, 0);
519		if (error)
520			goto out;
521		sc.flags |= XCHK_TRY_HARDER;
522		goto retry_op;
523	} else if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE))
524		goto out_teardown;
525
526	xchk_update_health(&sc);
527
528	if ((sc.sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
529	    !(sc.flags & XREP_ALREADY_FIXED)) {
530		bool needs_fix;
531
532		/* Let debug users force us into the repair routines. */
533		if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR))
534			sc.sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
535
536		needs_fix = (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
537						XFS_SCRUB_OFLAG_XCORRUPT |
538						XFS_SCRUB_OFLAG_PREEN));
539		/*
540		 * If userspace asked for a repair but it wasn't necessary,
541		 * report that back to userspace.
542		 */
543		if (!needs_fix) {
544			sc.sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
545			goto out_nofix;
546		}
547
548		/*
549		 * If it's broken, userspace wants us to fix it, and we haven't
550		 * already tried to fix it, then attempt a repair.
551		 */
552		error = xrep_attempt(&sc);
553		if (error == -EAGAIN) {
554			/*
555			 * Either the repair function succeeded or it couldn't
556			 * get all the resources it needs; either way, we go
557			 * back to the beginning and call the scrub function.
558			 */
559			error = xchk_teardown(&sc, 0);
560			if (error) {
561				xrep_failure(mp);
562				goto out;
563			}
564			goto retry_op;
565		}
566	}
567
568out_nofix:
569	xchk_postmortem(&sc);
570out_teardown:
571	error = xchk_teardown(&sc, error);
572out:
573	trace_xchk_done(XFS_I(file_inode(file)), sm, error);
574	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
575		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
576		error = 0;
577	}
578	return error;
579}
580