xref: /linux/fs/xfs/scrub/scrub.c (revision d85fe250f2eb61e19029e9e0d30095c5f646e2f2)
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 "xfs_exchrange.h"
21 #include "xfs_exchmaps.h"
22 #include "scrub/scrub.h"
23 #include "scrub/common.h"
24 #include "scrub/trace.h"
25 #include "scrub/repair.h"
26 #include "scrub/health.h"
27 #include "scrub/stats.h"
28 #include "scrub/xfile.h"
29 #include "scrub/tempfile.h"
30 #include "scrub/orphanage.h"
31 
32 /*
33  * Online Scrub and Repair
34  *
35  * Traditionally, XFS (the kernel driver) did not know how to check or
36  * repair on-disk data structures.  That task was left to the xfs_check
37  * and xfs_repair tools, both of which require taking the filesystem
38  * offline for a thorough but time consuming examination.  Online
39  * scrub & repair, on the other hand, enables us to check the metadata
40  * for obvious errors while carefully stepping around the filesystem's
41  * ongoing operations, locking rules, etc.
42  *
43  * Given that most XFS metadata consist of records stored in a btree,
44  * most of the checking functions iterate the btree blocks themselves
45  * looking for irregularities.  When a record block is encountered, each
46  * record can be checked for obviously bad values.  Record values can
47  * also be cross-referenced against other btrees to look for potential
48  * misunderstandings between pieces of metadata.
49  *
50  * It is expected that the checkers responsible for per-AG metadata
51  * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
52  * metadata structure, and perform any relevant cross-referencing before
53  * unlocking the AG and returning the results to userspace.  These
54  * scrubbers must not keep an AG locked for too long to avoid tying up
55  * the block and inode allocators.
56  *
57  * Block maps and b-trees rooted in an inode present a special challenge
58  * because they can involve extents from any AG.  The general scrubber
59  * structure of lock -> check -> xref -> unlock still holds, but AG
60  * locking order rules /must/ be obeyed to avoid deadlocks.  The
61  * ordering rule, of course, is that we must lock in increasing AG
62  * order.  Helper functions are provided to track which AG headers we've
63  * already locked.  If we detect an imminent locking order violation, we
64  * can signal a potential deadlock, in which case the scrubber can jump
65  * out to the top level, lock all the AGs in order, and retry the scrub.
66  *
67  * For file data (directories, extended attributes, symlinks) scrub, we
68  * can simply lock the inode and walk the data.  For btree data
69  * (directories and attributes) we follow the same btree-scrubbing
70  * strategy outlined previously to check the records.
71  *
72  * We use a bit of trickery with transactions to avoid buffer deadlocks
73  * if there is a cycle in the metadata.  The basic problem is that
74  * travelling down a btree involves locking the current buffer at each
75  * tree level.  If a pointer should somehow point back to a buffer that
76  * we've already examined, we will deadlock due to the second buffer
77  * locking attempt.  Note however that grabbing a buffer in transaction
78  * context links the locked buffer to the transaction.  If we try to
79  * re-grab the buffer in the context of the same transaction, we avoid
80  * the second lock attempt and continue.  Between the verifier and the
81  * scrubber, something will notice that something is amiss and report
82  * the corruption.  Therefore, each scrubber will allocate an empty
83  * transaction, attach buffers to it, and cancel the transaction at the
84  * end of the scrub run.  Cancelling a non-dirty transaction simply
85  * unlocks the buffers.
86  *
87  * There are four pieces of data that scrub can communicate to
88  * userspace.  The first is the error code (errno), which can be used to
89  * communicate operational errors in performing the scrub.  There are
90  * also three flags that can be set in the scrub context.  If the data
91  * structure itself is corrupt, the CORRUPT flag will be set.  If
92  * the metadata is correct but otherwise suboptimal, the PREEN flag
93  * will be set.
94  *
95  * We perform secondary validation of filesystem metadata by
96  * cross-referencing every record with all other available metadata.
97  * For example, for block mapping extents, we verify that there are no
98  * records in the free space and inode btrees corresponding to that
99  * space extent and that there is a corresponding entry in the reverse
100  * mapping btree.  Inconsistent metadata is noted by setting the
101  * XCORRUPT flag; btree query function errors are noted by setting the
102  * XFAIL flag and deleting the cursor to prevent further attempts to
103  * cross-reference with a defective btree.
104  *
105  * If a piece of metadata proves corrupt or suboptimal, the userspace
106  * program can ask the kernel to apply some tender loving care (TLC) to
107  * the metadata object by setting the REPAIR flag and re-calling the
108  * scrub ioctl.  "Corruption" is defined by metadata violating the
109  * on-disk specification; operations cannot continue if the violation is
110  * left untreated.  It is possible for XFS to continue if an object is
111  * "suboptimal", however performance may be degraded.  Repairs are
112  * usually performed by rebuilding the metadata entirely out of
113  * redundant metadata.  Optimizing, on the other hand, can sometimes be
114  * done without rebuilding entire structures.
115  *
116  * Generally speaking, the repair code has the following code structure:
117  * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
118  * The first check helps us figure out if we need to rebuild or simply
119  * optimize the structure so that the rebuild knows what to do.  The
120  * second check evaluates the completeness of the repair; that is what
121  * is reported to userspace.
122  *
123  * A quick note on symbol prefixes:
124  * - "xfs_" are general XFS symbols.
125  * - "xchk_" are symbols related to metadata checking.
126  * - "xrep_" are symbols related to metadata repair.
127  * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
128  */
129 
130 /*
131  * Scrub probe -- userspace uses this to probe if we're willing to scrub
132  * or repair a given mountpoint.  This will be used by xfs_scrub to
133  * probe the kernel's abilities to scrub (and repair) the metadata.  We
134  * do this by validating the ioctl inputs from userspace, preparing the
135  * filesystem for a scrub (or a repair) operation, and immediately
136  * returning to userspace.  Userspace can use the returned errno and
137  * structure state to decide (in broad terms) if scrub/repair are
138  * supported by the running kernel.
139  */
140 static int
141 xchk_probe(
142 	struct xfs_scrub	*sc)
143 {
144 	int			error = 0;
145 
146 	if (xchk_should_terminate(sc, &error))
147 		return error;
148 
149 	return 0;
150 }
151 
152 /* Scrub setup and teardown */
153 
154 #define FSGATES_MASK	(XCHK_FSGATES_ALL | XREP_FSGATES_ALL)
155 static inline void
156 xchk_fsgates_disable(
157 	struct xfs_scrub	*sc)
158 {
159 	if (!(sc->flags & FSGATES_MASK))
160 		return;
161 
162 	trace_xchk_fsgates_disable(sc, sc->flags & FSGATES_MASK);
163 
164 	if (sc->flags & XCHK_FSGATES_DRAIN)
165 		xfs_drain_wait_disable();
166 
167 	if (sc->flags & XCHK_FSGATES_QUOTA)
168 		xfs_dqtrx_hook_disable();
169 
170 	if (sc->flags & XCHK_FSGATES_DIRENTS)
171 		xfs_dir_hook_disable();
172 
173 	if (sc->flags & XCHK_FSGATES_RMAP)
174 		xfs_rmap_hook_disable();
175 
176 	sc->flags &= ~FSGATES_MASK;
177 }
178 #undef FSGATES_MASK
179 
180 /* Free the resources associated with a scrub subtype. */
181 void
182 xchk_scrub_free_subord(
183 	struct xfs_scrub_subord	*sub)
184 {
185 	struct xfs_scrub	*sc = sub->parent_sc;
186 
187 	ASSERT(sc->ip == sub->sc.ip);
188 	ASSERT(sc->orphanage == sub->sc.orphanage);
189 	ASSERT(sc->tempip == sub->sc.tempip);
190 
191 	sc->sm->sm_type = sub->old_smtype;
192 	sc->sm->sm_flags = sub->old_smflags |
193 				(sc->sm->sm_flags & XFS_SCRUB_FLAGS_OUT);
194 	sc->tp = sub->sc.tp;
195 
196 	if (sub->sc.buf) {
197 		if (sub->sc.buf_cleanup)
198 			sub->sc.buf_cleanup(sub->sc.buf);
199 		kvfree(sub->sc.buf);
200 	}
201 	if (sub->sc.xmbtp)
202 		xmbuf_free(sub->sc.xmbtp);
203 	if (sub->sc.xfile)
204 		xfile_destroy(sub->sc.xfile);
205 
206 	sc->ilock_flags = sub->sc.ilock_flags;
207 	sc->orphanage_ilock_flags = sub->sc.orphanage_ilock_flags;
208 	sc->temp_ilock_flags = sub->sc.temp_ilock_flags;
209 
210 	kfree(sub);
211 }
212 
213 /* Free all the resources and finish the transactions. */
214 STATIC int
215 xchk_teardown(
216 	struct xfs_scrub	*sc,
217 	int			error)
218 {
219 	xchk_ag_free(sc, &sc->sa);
220 	if (sc->tp) {
221 		if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
222 			error = xfs_trans_commit(sc->tp);
223 		else
224 			xfs_trans_cancel(sc->tp);
225 		sc->tp = NULL;
226 	}
227 	if (sc->ip) {
228 		if (sc->ilock_flags)
229 			xchk_iunlock(sc, sc->ilock_flags);
230 		xchk_irele(sc, sc->ip);
231 		sc->ip = NULL;
232 	}
233 	if (sc->flags & XCHK_HAVE_FREEZE_PROT) {
234 		sc->flags &= ~XCHK_HAVE_FREEZE_PROT;
235 		mnt_drop_write_file(sc->file);
236 	}
237 	if (sc->xmbtp) {
238 		xmbuf_free(sc->xmbtp);
239 		sc->xmbtp = NULL;
240 	}
241 	if (sc->xfile) {
242 		xfile_destroy(sc->xfile);
243 		sc->xfile = NULL;
244 	}
245 	if (sc->buf) {
246 		if (sc->buf_cleanup)
247 			sc->buf_cleanup(sc->buf);
248 		kvfree(sc->buf);
249 		sc->buf_cleanup = NULL;
250 		sc->buf = NULL;
251 	}
252 
253 	xrep_tempfile_rele(sc);
254 	xrep_orphanage_rele(sc);
255 	xchk_fsgates_disable(sc);
256 	return error;
257 }
258 
259 /* Scrubbing dispatch. */
260 
261 static const struct xchk_meta_ops meta_scrub_ops[] = {
262 	[XFS_SCRUB_TYPE_PROBE] = {	/* ioctl presence test */
263 		.type	= ST_NONE,
264 		.setup	= xchk_setup_fs,
265 		.scrub	= xchk_probe,
266 		.repair = xrep_probe,
267 	},
268 	[XFS_SCRUB_TYPE_SB] = {		/* superblock */
269 		.type	= ST_PERAG,
270 		.setup	= xchk_setup_agheader,
271 		.scrub	= xchk_superblock,
272 		.repair	= xrep_superblock,
273 	},
274 	[XFS_SCRUB_TYPE_AGF] = {	/* agf */
275 		.type	= ST_PERAG,
276 		.setup	= xchk_setup_agheader,
277 		.scrub	= xchk_agf,
278 		.repair	= xrep_agf,
279 	},
280 	[XFS_SCRUB_TYPE_AGFL]= {	/* agfl */
281 		.type	= ST_PERAG,
282 		.setup	= xchk_setup_agheader,
283 		.scrub	= xchk_agfl,
284 		.repair	= xrep_agfl,
285 	},
286 	[XFS_SCRUB_TYPE_AGI] = {	/* agi */
287 		.type	= ST_PERAG,
288 		.setup	= xchk_setup_agheader,
289 		.scrub	= xchk_agi,
290 		.repair	= xrep_agi,
291 	},
292 	[XFS_SCRUB_TYPE_BNOBT] = {	/* bnobt */
293 		.type	= ST_PERAG,
294 		.setup	= xchk_setup_ag_allocbt,
295 		.scrub	= xchk_allocbt,
296 		.repair	= xrep_allocbt,
297 		.repair_eval = xrep_revalidate_allocbt,
298 	},
299 	[XFS_SCRUB_TYPE_CNTBT] = {	/* cntbt */
300 		.type	= ST_PERAG,
301 		.setup	= xchk_setup_ag_allocbt,
302 		.scrub	= xchk_allocbt,
303 		.repair	= xrep_allocbt,
304 		.repair_eval = xrep_revalidate_allocbt,
305 	},
306 	[XFS_SCRUB_TYPE_INOBT] = {	/* inobt */
307 		.type	= ST_PERAG,
308 		.setup	= xchk_setup_ag_iallocbt,
309 		.scrub	= xchk_iallocbt,
310 		.repair	= xrep_iallocbt,
311 		.repair_eval = xrep_revalidate_iallocbt,
312 	},
313 	[XFS_SCRUB_TYPE_FINOBT] = {	/* finobt */
314 		.type	= ST_PERAG,
315 		.setup	= xchk_setup_ag_iallocbt,
316 		.scrub	= xchk_iallocbt,
317 		.has	= xfs_has_finobt,
318 		.repair	= xrep_iallocbt,
319 		.repair_eval = xrep_revalidate_iallocbt,
320 	},
321 	[XFS_SCRUB_TYPE_RMAPBT] = {	/* rmapbt */
322 		.type	= ST_PERAG,
323 		.setup	= xchk_setup_ag_rmapbt,
324 		.scrub	= xchk_rmapbt,
325 		.has	= xfs_has_rmapbt,
326 		.repair	= xrep_rmapbt,
327 	},
328 	[XFS_SCRUB_TYPE_REFCNTBT] = {	/* refcountbt */
329 		.type	= ST_PERAG,
330 		.setup	= xchk_setup_ag_refcountbt,
331 		.scrub	= xchk_refcountbt,
332 		.has	= xfs_has_reflink,
333 		.repair	= xrep_refcountbt,
334 	},
335 	[XFS_SCRUB_TYPE_INODE] = {	/* inode record */
336 		.type	= ST_INODE,
337 		.setup	= xchk_setup_inode,
338 		.scrub	= xchk_inode,
339 		.repair	= xrep_inode,
340 	},
341 	[XFS_SCRUB_TYPE_BMBTD] = {	/* inode data fork */
342 		.type	= ST_INODE,
343 		.setup	= xchk_setup_inode_bmap,
344 		.scrub	= xchk_bmap_data,
345 		.repair	= xrep_bmap_data,
346 	},
347 	[XFS_SCRUB_TYPE_BMBTA] = {	/* inode attr fork */
348 		.type	= ST_INODE,
349 		.setup	= xchk_setup_inode_bmap,
350 		.scrub	= xchk_bmap_attr,
351 		.repair	= xrep_bmap_attr,
352 	},
353 	[XFS_SCRUB_TYPE_BMBTC] = {	/* inode CoW fork */
354 		.type	= ST_INODE,
355 		.setup	= xchk_setup_inode_bmap,
356 		.scrub	= xchk_bmap_cow,
357 		.repair	= xrep_bmap_cow,
358 	},
359 	[XFS_SCRUB_TYPE_DIR] = {	/* directory */
360 		.type	= ST_INODE,
361 		.setup	= xchk_setup_directory,
362 		.scrub	= xchk_directory,
363 		.repair	= xrep_directory,
364 	},
365 	[XFS_SCRUB_TYPE_XATTR] = {	/* extended attributes */
366 		.type	= ST_INODE,
367 		.setup	= xchk_setup_xattr,
368 		.scrub	= xchk_xattr,
369 		.repair	= xrep_xattr,
370 	},
371 	[XFS_SCRUB_TYPE_SYMLINK] = {	/* symbolic link */
372 		.type	= ST_INODE,
373 		.setup	= xchk_setup_symlink,
374 		.scrub	= xchk_symlink,
375 		.repair	= xrep_symlink,
376 	},
377 	[XFS_SCRUB_TYPE_PARENT] = {	/* parent pointers */
378 		.type	= ST_INODE,
379 		.setup	= xchk_setup_parent,
380 		.scrub	= xchk_parent,
381 		.repair	= xrep_parent,
382 	},
383 	[XFS_SCRUB_TYPE_RTBITMAP] = {	/* realtime bitmap */
384 		.type	= ST_FS,
385 		.setup	= xchk_setup_rtbitmap,
386 		.scrub	= xchk_rtbitmap,
387 		.repair	= xrep_rtbitmap,
388 	},
389 	[XFS_SCRUB_TYPE_RTSUM] = {	/* realtime summary */
390 		.type	= ST_FS,
391 		.setup	= xchk_setup_rtsummary,
392 		.scrub	= xchk_rtsummary,
393 		.repair	= xrep_rtsummary,
394 	},
395 	[XFS_SCRUB_TYPE_UQUOTA] = {	/* user quota */
396 		.type	= ST_FS,
397 		.setup	= xchk_setup_quota,
398 		.scrub	= xchk_quota,
399 		.repair	= xrep_quota,
400 	},
401 	[XFS_SCRUB_TYPE_GQUOTA] = {	/* group quota */
402 		.type	= ST_FS,
403 		.setup	= xchk_setup_quota,
404 		.scrub	= xchk_quota,
405 		.repair	= xrep_quota,
406 	},
407 	[XFS_SCRUB_TYPE_PQUOTA] = {	/* project quota */
408 		.type	= ST_FS,
409 		.setup	= xchk_setup_quota,
410 		.scrub	= xchk_quota,
411 		.repair	= xrep_quota,
412 	},
413 	[XFS_SCRUB_TYPE_FSCOUNTERS] = {	/* fs summary counters */
414 		.type	= ST_FS,
415 		.setup	= xchk_setup_fscounters,
416 		.scrub	= xchk_fscounters,
417 		.repair	= xrep_fscounters,
418 	},
419 	[XFS_SCRUB_TYPE_QUOTACHECK] = {	/* quota counters */
420 		.type	= ST_FS,
421 		.setup	= xchk_setup_quotacheck,
422 		.scrub	= xchk_quotacheck,
423 		.repair	= xrep_quotacheck,
424 	},
425 	[XFS_SCRUB_TYPE_NLINKS] = {	/* inode link counts */
426 		.type	= ST_FS,
427 		.setup	= xchk_setup_nlinks,
428 		.scrub	= xchk_nlinks,
429 		.repair	= xrep_nlinks,
430 	},
431 	[XFS_SCRUB_TYPE_HEALTHY] = {	/* fs healthy; clean all reminders */
432 		.type	= ST_FS,
433 		.setup	= xchk_setup_fs,
434 		.scrub	= xchk_health_record,
435 		.repair = xrep_notsupported,
436 	},
437 };
438 
439 static int
440 xchk_validate_inputs(
441 	struct xfs_mount		*mp,
442 	struct xfs_scrub_metadata	*sm)
443 {
444 	int				error;
445 	const struct xchk_meta_ops	*ops;
446 
447 	error = -EINVAL;
448 	/* Check our inputs. */
449 	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
450 	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
451 		goto out;
452 	/* sm_reserved[] must be zero */
453 	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
454 		goto out;
455 
456 	error = -ENOENT;
457 	/* Do we know about this type of metadata? */
458 	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
459 		goto out;
460 	ops = &meta_scrub_ops[sm->sm_type];
461 	if (ops->setup == NULL || ops->scrub == NULL)
462 		goto out;
463 	/* Does this fs even support this type of metadata? */
464 	if (ops->has && !ops->has(mp))
465 		goto out;
466 
467 	error = -EINVAL;
468 	/* restricting fields must be appropriate for type */
469 	switch (ops->type) {
470 	case ST_NONE:
471 	case ST_FS:
472 		if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
473 			goto out;
474 		break;
475 	case ST_PERAG:
476 		if (sm->sm_ino || sm->sm_gen ||
477 		    sm->sm_agno >= mp->m_sb.sb_agcount)
478 			goto out;
479 		break;
480 	case ST_INODE:
481 		if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
482 			goto out;
483 		break;
484 	default:
485 		goto out;
486 	}
487 
488 	/* No rebuild without repair. */
489 	if ((sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) &&
490 	    !(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
491 		return -EINVAL;
492 
493 	/*
494 	 * We only want to repair read-write v5+ filesystems.  Defer the check
495 	 * for ops->repair until after our scrub confirms that we need to
496 	 * perform repairs so that we avoid failing due to not supporting
497 	 * repairing an object that doesn't need repairs.
498 	 */
499 	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
500 		error = -EOPNOTSUPP;
501 		if (!xfs_has_crc(mp))
502 			goto out;
503 
504 		error = -EROFS;
505 		if (xfs_is_readonly(mp))
506 			goto out;
507 	}
508 
509 	error = 0;
510 out:
511 	return error;
512 }
513 
514 #ifdef CONFIG_XFS_ONLINE_REPAIR
515 static inline void xchk_postmortem(struct xfs_scrub *sc)
516 {
517 	/*
518 	 * Userspace asked us to repair something, we repaired it, rescanned
519 	 * it, and the rescan says it's still broken.  Scream about this in
520 	 * the system logs.
521 	 */
522 	if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
523 	    (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
524 				 XFS_SCRUB_OFLAG_XCORRUPT)))
525 		xrep_failure(sc->mp);
526 }
527 #else
528 static inline void xchk_postmortem(struct xfs_scrub *sc)
529 {
530 	/*
531 	 * Userspace asked us to scrub something, it's broken, and we have no
532 	 * way of fixing it.  Scream in the logs.
533 	 */
534 	if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
535 				XFS_SCRUB_OFLAG_XCORRUPT))
536 		xfs_alert_ratelimited(sc->mp,
537 				"Corruption detected during scrub.");
538 }
539 #endif /* CONFIG_XFS_ONLINE_REPAIR */
540 
541 /*
542  * Create a new scrub context from an existing one, but with a different scrub
543  * type.
544  */
545 struct xfs_scrub_subord *
546 xchk_scrub_create_subord(
547 	struct xfs_scrub	*sc,
548 	unsigned int		subtype)
549 {
550 	struct xfs_scrub_subord	*sub;
551 
552 	sub = kzalloc(sizeof(*sub), XCHK_GFP_FLAGS);
553 	if (!sub)
554 		return ERR_PTR(-ENOMEM);
555 
556 	sub->old_smtype = sc->sm->sm_type;
557 	sub->old_smflags = sc->sm->sm_flags;
558 	sub->parent_sc = sc;
559 	memcpy(&sub->sc, sc, sizeof(struct xfs_scrub));
560 	sub->sc.ops = &meta_scrub_ops[subtype];
561 	sub->sc.sm->sm_type = subtype;
562 	sub->sc.sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
563 	sub->sc.buf = NULL;
564 	sub->sc.buf_cleanup = NULL;
565 	sub->sc.xfile = NULL;
566 	sub->sc.xmbtp = NULL;
567 
568 	return sub;
569 }
570 
571 /* Dispatch metadata scrubbing. */
572 int
573 xfs_scrub_metadata(
574 	struct file			*file,
575 	struct xfs_scrub_metadata	*sm)
576 {
577 	struct xchk_stats_run		run = { };
578 	struct xfs_scrub		*sc;
579 	struct xfs_mount		*mp = XFS_I(file_inode(file))->i_mount;
580 	u64				check_start;
581 	int				error = 0;
582 
583 	BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
584 		(sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
585 
586 	trace_xchk_start(XFS_I(file_inode(file)), sm, error);
587 
588 	/* Forbidden if we are shut down or mounted norecovery. */
589 	error = -ESHUTDOWN;
590 	if (xfs_is_shutdown(mp))
591 		goto out;
592 	error = -ENOTRECOVERABLE;
593 	if (xfs_has_norecovery(mp))
594 		goto out;
595 
596 	error = xchk_validate_inputs(mp, sm);
597 	if (error)
598 		goto out;
599 
600 	xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SCRUB,
601  "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
602 
603 	sc = kzalloc(sizeof(struct xfs_scrub), XCHK_GFP_FLAGS);
604 	if (!sc) {
605 		error = -ENOMEM;
606 		goto out;
607 	}
608 
609 	sc->mp = mp;
610 	sc->file = file;
611 	sc->sm = sm;
612 	sc->ops = &meta_scrub_ops[sm->sm_type];
613 	sc->sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
614 retry_op:
615 	/*
616 	 * When repairs are allowed, prevent freezing or readonly remount while
617 	 * scrub is running with a real transaction.
618 	 */
619 	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
620 		error = mnt_want_write_file(sc->file);
621 		if (error)
622 			goto out_sc;
623 
624 		sc->flags |= XCHK_HAVE_FREEZE_PROT;
625 	}
626 
627 	/* Set up for the operation. */
628 	error = sc->ops->setup(sc);
629 	if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
630 		goto try_harder;
631 	if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
632 		goto need_drain;
633 	if (error)
634 		goto out_teardown;
635 
636 	/* Scrub for errors. */
637 	check_start = xchk_stats_now();
638 	if ((sc->flags & XREP_ALREADY_FIXED) && sc->ops->repair_eval != NULL)
639 		error = sc->ops->repair_eval(sc);
640 	else
641 		error = sc->ops->scrub(sc);
642 	run.scrub_ns += xchk_stats_elapsed_ns(check_start);
643 	if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
644 		goto try_harder;
645 	if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
646 		goto need_drain;
647 	if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE))
648 		goto out_teardown;
649 
650 	xchk_update_health(sc);
651 
652 	if (xchk_could_repair(sc)) {
653 		/*
654 		 * If userspace asked for a repair but it wasn't necessary,
655 		 * report that back to userspace.
656 		 */
657 		if (!xrep_will_attempt(sc)) {
658 			sc->sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
659 			goto out_nofix;
660 		}
661 
662 		/*
663 		 * If it's broken, userspace wants us to fix it, and we haven't
664 		 * already tried to fix it, then attempt a repair.
665 		 */
666 		error = xrep_attempt(sc, &run);
667 		if (error == -EAGAIN) {
668 			/*
669 			 * Either the repair function succeeded or it couldn't
670 			 * get all the resources it needs; either way, we go
671 			 * back to the beginning and call the scrub function.
672 			 */
673 			error = xchk_teardown(sc, 0);
674 			if (error) {
675 				xrep_failure(mp);
676 				goto out_sc;
677 			}
678 			goto retry_op;
679 		}
680 	}
681 
682 out_nofix:
683 	xchk_postmortem(sc);
684 out_teardown:
685 	error = xchk_teardown(sc, error);
686 out_sc:
687 	if (error != -ENOENT)
688 		xchk_stats_merge(mp, sm, &run);
689 	kfree(sc);
690 out:
691 	trace_xchk_done(XFS_I(file_inode(file)), sm, error);
692 	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
693 		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
694 		error = 0;
695 	}
696 	return error;
697 need_drain:
698 	error = xchk_teardown(sc, 0);
699 	if (error)
700 		goto out_sc;
701 	sc->flags |= XCHK_NEED_DRAIN;
702 	run.retries++;
703 	goto retry_op;
704 try_harder:
705 	/*
706 	 * Scrubbers return -EDEADLOCK to mean 'try harder'.  Tear down
707 	 * everything we hold, then set up again with preparation for
708 	 * worst-case scenarios.
709 	 */
710 	error = xchk_teardown(sc, 0);
711 	if (error)
712 		goto out_sc;
713 	sc->flags |= XCHK_TRY_HARDER;
714 	run.retries++;
715 	goto retry_op;
716 }
717