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->buf) {
177		kmem_free(sc->buf);
178		sc->buf = NULL;
179	}
180	return error;
181}
182
183/* Scrubbing dispatch. */
184
185static const struct xchk_meta_ops meta_scrub_ops[] = {
186	[XFS_SCRUB_TYPE_PROBE] = {	/* ioctl presence test */
187		.type	= ST_NONE,
188		.setup	= xchk_setup_fs,
189		.scrub	= xchk_probe,
190		.repair = xrep_probe,
191	},
192	[XFS_SCRUB_TYPE_SB] = {		/* superblock */
193		.type	= ST_PERAG,
194		.setup	= xchk_setup_fs,
195		.scrub	= xchk_superblock,
196		.repair	= xrep_superblock,
197	},
198	[XFS_SCRUB_TYPE_AGF] = {	/* agf */
199		.type	= ST_PERAG,
200		.setup	= xchk_setup_fs,
201		.scrub	= xchk_agf,
202		.repair	= xrep_agf,
203	},
204	[XFS_SCRUB_TYPE_AGFL]= {	/* agfl */
205		.type	= ST_PERAG,
206		.setup	= xchk_setup_fs,
207		.scrub	= xchk_agfl,
208		.repair	= xrep_agfl,
209	},
210	[XFS_SCRUB_TYPE_AGI] = {	/* agi */
211		.type	= ST_PERAG,
212		.setup	= xchk_setup_fs,
213		.scrub	= xchk_agi,
214		.repair	= xrep_agi,
215	},
216	[XFS_SCRUB_TYPE_BNOBT] = {	/* bnobt */
217		.type	= ST_PERAG,
218		.setup	= xchk_setup_ag_allocbt,
219		.scrub	= xchk_bnobt,
220		.repair	= xrep_notsupported,
221	},
222	[XFS_SCRUB_TYPE_CNTBT] = {	/* cntbt */
223		.type	= ST_PERAG,
224		.setup	= xchk_setup_ag_allocbt,
225		.scrub	= xchk_cntbt,
226		.repair	= xrep_notsupported,
227	},
228	[XFS_SCRUB_TYPE_INOBT] = {	/* inobt */
229		.type	= ST_PERAG,
230		.setup	= xchk_setup_ag_iallocbt,
231		.scrub	= xchk_inobt,
232		.repair	= xrep_notsupported,
233	},
234	[XFS_SCRUB_TYPE_FINOBT] = {	/* finobt */
235		.type	= ST_PERAG,
236		.setup	= xchk_setup_ag_iallocbt,
237		.scrub	= xchk_finobt,
238		.has	= xfs_has_finobt,
239		.repair	= xrep_notsupported,
240	},
241	[XFS_SCRUB_TYPE_RMAPBT] = {	/* rmapbt */
242		.type	= ST_PERAG,
243		.setup	= xchk_setup_ag_rmapbt,
244		.scrub	= xchk_rmapbt,
245		.has	= xfs_has_rmapbt,
246		.repair	= xrep_notsupported,
247	},
248	[XFS_SCRUB_TYPE_REFCNTBT] = {	/* refcountbt */
249		.type	= ST_PERAG,
250		.setup	= xchk_setup_ag_refcountbt,
251		.scrub	= xchk_refcountbt,
252		.has	= xfs_has_reflink,
253		.repair	= xrep_notsupported,
254	},
255	[XFS_SCRUB_TYPE_INODE] = {	/* inode record */
256		.type	= ST_INODE,
257		.setup	= xchk_setup_inode,
258		.scrub	= xchk_inode,
259		.repair	= xrep_notsupported,
260	},
261	[XFS_SCRUB_TYPE_BMBTD] = {	/* inode data fork */
262		.type	= ST_INODE,
263		.setup	= xchk_setup_inode_bmap,
264		.scrub	= xchk_bmap_data,
265		.repair	= xrep_notsupported,
266	},
267	[XFS_SCRUB_TYPE_BMBTA] = {	/* inode attr fork */
268		.type	= ST_INODE,
269		.setup	= xchk_setup_inode_bmap,
270		.scrub	= xchk_bmap_attr,
271		.repair	= xrep_notsupported,
272	},
273	[XFS_SCRUB_TYPE_BMBTC] = {	/* inode CoW fork */
274		.type	= ST_INODE,
275		.setup	= xchk_setup_inode_bmap,
276		.scrub	= xchk_bmap_cow,
277		.repair	= xrep_notsupported,
278	},
279	[XFS_SCRUB_TYPE_DIR] = {	/* directory */
280		.type	= ST_INODE,
281		.setup	= xchk_setup_directory,
282		.scrub	= xchk_directory,
283		.repair	= xrep_notsupported,
284	},
285	[XFS_SCRUB_TYPE_XATTR] = {	/* extended attributes */
286		.type	= ST_INODE,
287		.setup	= xchk_setup_xattr,
288		.scrub	= xchk_xattr,
289		.repair	= xrep_notsupported,
290	},
291	[XFS_SCRUB_TYPE_SYMLINK] = {	/* symbolic link */
292		.type	= ST_INODE,
293		.setup	= xchk_setup_symlink,
294		.scrub	= xchk_symlink,
295		.repair	= xrep_notsupported,
296	},
297	[XFS_SCRUB_TYPE_PARENT] = {	/* parent pointers */
298		.type	= ST_INODE,
299		.setup	= xchk_setup_parent,
300		.scrub	= xchk_parent,
301		.repair	= xrep_notsupported,
302	},
303	[XFS_SCRUB_TYPE_RTBITMAP] = {	/* realtime bitmap */
304		.type	= ST_FS,
305		.setup	= xchk_setup_rt,
306		.scrub	= xchk_rtbitmap,
307		.has	= xfs_has_realtime,
308		.repair	= xrep_notsupported,
309	},
310	[XFS_SCRUB_TYPE_RTSUM] = {	/* realtime summary */
311		.type	= ST_FS,
312		.setup	= xchk_setup_rt,
313		.scrub	= xchk_rtsummary,
314		.has	= xfs_has_realtime,
315		.repair	= xrep_notsupported,
316	},
317	[XFS_SCRUB_TYPE_UQUOTA] = {	/* user quota */
318		.type	= ST_FS,
319		.setup	= xchk_setup_quota,
320		.scrub	= xchk_quota,
321		.repair	= xrep_notsupported,
322	},
323	[XFS_SCRUB_TYPE_GQUOTA] = {	/* group quota */
324		.type	= ST_FS,
325		.setup	= xchk_setup_quota,
326		.scrub	= xchk_quota,
327		.repair	= xrep_notsupported,
328	},
329	[XFS_SCRUB_TYPE_PQUOTA] = {	/* project quota */
330		.type	= ST_FS,
331		.setup	= xchk_setup_quota,
332		.scrub	= xchk_quota,
333		.repair	= xrep_notsupported,
334	},
335	[XFS_SCRUB_TYPE_FSCOUNTERS] = {	/* fs summary counters */
336		.type	= ST_FS,
337		.setup	= xchk_setup_fscounters,
338		.scrub	= xchk_fscounters,
339		.repair	= xrep_notsupported,
340	},
341};
342
343/* This isn't a stable feature, warn once per day. */
344static inline void
345xchk_experimental_warning(
346	struct xfs_mount	*mp)
347{
348	static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
349			"xchk_warning", 86400 * HZ, 1);
350	ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
351
352	if (__ratelimit(&scrub_warning))
353		xfs_alert(mp,
354"EXPERIMENTAL online scrub feature in use. Use at your own risk!");
355}
356
357static int
358xchk_validate_inputs(
359	struct xfs_mount		*mp,
360	struct xfs_scrub_metadata	*sm)
361{
362	int				error;
363	const struct xchk_meta_ops	*ops;
364
365	error = -EINVAL;
366	/* Check our inputs. */
367	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
368	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
369		goto out;
370	/* sm_reserved[] must be zero */
371	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
372		goto out;
373
374	error = -ENOENT;
375	/* Do we know about this type of metadata? */
376	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
377		goto out;
378	ops = &meta_scrub_ops[sm->sm_type];
379	if (ops->setup == NULL || ops->scrub == NULL)
380		goto out;
381	/* Does this fs even support this type of metadata? */
382	if (ops->has && !ops->has(mp))
383		goto out;
384
385	error = -EINVAL;
386	/* restricting fields must be appropriate for type */
387	switch (ops->type) {
388	case ST_NONE:
389	case ST_FS:
390		if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
391			goto out;
392		break;
393	case ST_PERAG:
394		if (sm->sm_ino || sm->sm_gen ||
395		    sm->sm_agno >= mp->m_sb.sb_agcount)
396			goto out;
397		break;
398	case ST_INODE:
399		if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
400			goto out;
401		break;
402	default:
403		goto out;
404	}
405
406	/*
407	 * We only want to repair read-write v5+ filesystems.  Defer the check
408	 * for ops->repair until after our scrub confirms that we need to
409	 * perform repairs so that we avoid failing due to not supporting
410	 * repairing an object that doesn't need repairs.
411	 */
412	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
413		error = -EOPNOTSUPP;
414		if (!xfs_has_crc(mp))
415			goto out;
416
417		error = -EROFS;
418		if (xfs_is_readonly(mp))
419			goto out;
420	}
421
422	error = 0;
423out:
424	return error;
425}
426
427#ifdef CONFIG_XFS_ONLINE_REPAIR
428static inline void xchk_postmortem(struct xfs_scrub *sc)
429{
430	/*
431	 * Userspace asked us to repair something, we repaired it, rescanned
432	 * it, and the rescan says it's still broken.  Scream about this in
433	 * the system logs.
434	 */
435	if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
436	    (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
437				 XFS_SCRUB_OFLAG_XCORRUPT)))
438		xrep_failure(sc->mp);
439}
440#else
441static inline void xchk_postmortem(struct xfs_scrub *sc)
442{
443	/*
444	 * Userspace asked us to scrub something, it's broken, and we have no
445	 * way of fixing it.  Scream in the logs.
446	 */
447	if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
448				XFS_SCRUB_OFLAG_XCORRUPT))
449		xfs_alert_ratelimited(sc->mp,
450				"Corruption detected during scrub.");
451}
452#endif /* CONFIG_XFS_ONLINE_REPAIR */
453
454/* Dispatch metadata scrubbing. */
455int
456xfs_scrub_metadata(
457	struct file			*file,
458	struct xfs_scrub_metadata	*sm)
459{
460	struct xfs_scrub		*sc;
461	struct xfs_mount		*mp = XFS_I(file_inode(file))->i_mount;
462	int				error = 0;
463
464	BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
465		(sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
466
467	trace_xchk_start(XFS_I(file_inode(file)), sm, error);
468
469	/* Forbidden if we are shut down or mounted norecovery. */
470	error = -ESHUTDOWN;
471	if (xfs_is_shutdown(mp))
472		goto out;
473	error = -ENOTRECOVERABLE;
474	if (xfs_has_norecovery(mp))
475		goto out;
476
477	error = xchk_validate_inputs(mp, sm);
478	if (error)
479		goto out;
480
481	xchk_experimental_warning(mp);
482
483	sc = kmem_zalloc(sizeof(struct xfs_scrub), KM_NOFS | KM_MAYFAIL);
484	if (!sc) {
485		error = -ENOMEM;
486		goto out;
487	}
488
489	sc->mp = mp;
490	sc->file = file;
491	sc->sm = sm;
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_sc;
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_sc;
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_sc;
563			}
564			goto retry_op;
565		}
566	}
567
568out_nofix:
569	xchk_postmortem(sc);
570out_teardown:
571	error = xchk_teardown(sc, error);
572out_sc:
573	kmem_free(sc);
574out:
575	trace_xchk_done(XFS_I(file_inode(file)), sm, error);
576	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
577		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
578		error = 0;
579	}
580	return error;
581}
582