xref: /linux/fs/xfs/scrub/scrub.c (revision cd80e7ee47d2fd5c97563c003ff31ce8240ca2d8)
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 "xfs_dir2.h"
23 #include "xfs_parent.h"
24 #include "xfs_icache.h"
25 #include "scrub/scrub.h"
26 #include "scrub/common.h"
27 #include "scrub/trace.h"
28 #include "scrub/repair.h"
29 #include "scrub/health.h"
30 #include "scrub/stats.h"
31 #include "scrub/xfile.h"
32 #include "scrub/tempfile.h"
33 #include "scrub/orphanage.h"
34 
35 /*
36  * Online Scrub and Repair
37  *
38  * Traditionally, XFS (the kernel driver) did not know how to check or
39  * repair on-disk data structures.  That task was left to the xfs_check
40  * and xfs_repair tools, both of which require taking the filesystem
41  * offline for a thorough but time consuming examination.  Online
42  * scrub & repair, on the other hand, enables us to check the metadata
43  * for obvious errors while carefully stepping around the filesystem's
44  * ongoing operations, locking rules, etc.
45  *
46  * Given that most XFS metadata consist of records stored in a btree,
47  * most of the checking functions iterate the btree blocks themselves
48  * looking for irregularities.  When a record block is encountered, each
49  * record can be checked for obviously bad values.  Record values can
50  * also be cross-referenced against other btrees to look for potential
51  * misunderstandings between pieces of metadata.
52  *
53  * It is expected that the checkers responsible for per-AG metadata
54  * structures will lock the AG headers (AGI, AGF, AGFL), iterate the
55  * metadata structure, and perform any relevant cross-referencing before
56  * unlocking the AG and returning the results to userspace.  These
57  * scrubbers must not keep an AG locked for too long to avoid tying up
58  * the block and inode allocators.
59  *
60  * Block maps and b-trees rooted in an inode present a special challenge
61  * because they can involve extents from any AG.  The general scrubber
62  * structure of lock -> check -> xref -> unlock still holds, but AG
63  * locking order rules /must/ be obeyed to avoid deadlocks.  The
64  * ordering rule, of course, is that we must lock in increasing AG
65  * order.  Helper functions are provided to track which AG headers we've
66  * already locked.  If we detect an imminent locking order violation, we
67  * can signal a potential deadlock, in which case the scrubber can jump
68  * out to the top level, lock all the AGs in order, and retry the scrub.
69  *
70  * For file data (directories, extended attributes, symlinks) scrub, we
71  * can simply lock the inode and walk the data.  For btree data
72  * (directories and attributes) we follow the same btree-scrubbing
73  * strategy outlined previously to check the records.
74  *
75  * We use a bit of trickery with transactions to avoid buffer deadlocks
76  * if there is a cycle in the metadata.  The basic problem is that
77  * travelling down a btree involves locking the current buffer at each
78  * tree level.  If a pointer should somehow point back to a buffer that
79  * we've already examined, we will deadlock due to the second buffer
80  * locking attempt.  Note however that grabbing a buffer in transaction
81  * context links the locked buffer to the transaction.  If we try to
82  * re-grab the buffer in the context of the same transaction, we avoid
83  * the second lock attempt and continue.  Between the verifier and the
84  * scrubber, something will notice that something is amiss and report
85  * the corruption.  Therefore, each scrubber will allocate an empty
86  * transaction, attach buffers to it, and cancel the transaction at the
87  * end of the scrub run.  Cancelling a non-dirty transaction simply
88  * unlocks the buffers.
89  *
90  * There are four pieces of data that scrub can communicate to
91  * userspace.  The first is the error code (errno), which can be used to
92  * communicate operational errors in performing the scrub.  There are
93  * also three flags that can be set in the scrub context.  If the data
94  * structure itself is corrupt, the CORRUPT flag will be set.  If
95  * the metadata is correct but otherwise suboptimal, the PREEN flag
96  * will be set.
97  *
98  * We perform secondary validation of filesystem metadata by
99  * cross-referencing every record with all other available metadata.
100  * For example, for block mapping extents, we verify that there are no
101  * records in the free space and inode btrees corresponding to that
102  * space extent and that there is a corresponding entry in the reverse
103  * mapping btree.  Inconsistent metadata is noted by setting the
104  * XCORRUPT flag; btree query function errors are noted by setting the
105  * XFAIL flag and deleting the cursor to prevent further attempts to
106  * cross-reference with a defective btree.
107  *
108  * If a piece of metadata proves corrupt or suboptimal, the userspace
109  * program can ask the kernel to apply some tender loving care (TLC) to
110  * the metadata object by setting the REPAIR flag and re-calling the
111  * scrub ioctl.  "Corruption" is defined by metadata violating the
112  * on-disk specification; operations cannot continue if the violation is
113  * left untreated.  It is possible for XFS to continue if an object is
114  * "suboptimal", however performance may be degraded.  Repairs are
115  * usually performed by rebuilding the metadata entirely out of
116  * redundant metadata.  Optimizing, on the other hand, can sometimes be
117  * done without rebuilding entire structures.
118  *
119  * Generally speaking, the repair code has the following code structure:
120  * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock.
121  * The first check helps us figure out if we need to rebuild or simply
122  * optimize the structure so that the rebuild knows what to do.  The
123  * second check evaluates the completeness of the repair; that is what
124  * is reported to userspace.
125  *
126  * A quick note on symbol prefixes:
127  * - "xfs_" are general XFS symbols.
128  * - "xchk_" are symbols related to metadata checking.
129  * - "xrep_" are symbols related to metadata repair.
130  * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS.
131  */
132 
133 /*
134  * Scrub probe -- userspace uses this to probe if we're willing to scrub
135  * or repair a given mountpoint.  This will be used by xfs_scrub to
136  * probe the kernel's abilities to scrub (and repair) the metadata.  We
137  * do this by validating the ioctl inputs from userspace, preparing the
138  * filesystem for a scrub (or a repair) operation, and immediately
139  * returning to userspace.  Userspace can use the returned errno and
140  * structure state to decide (in broad terms) if scrub/repair are
141  * supported by the running kernel.
142  */
143 static int
144 xchk_probe(
145 	struct xfs_scrub	*sc)
146 {
147 	int			error = 0;
148 
149 	if (xchk_should_terminate(sc, &error))
150 		return error;
151 
152 	return 0;
153 }
154 
155 /* Scrub setup and teardown */
156 
157 static inline void
158 xchk_fsgates_disable(
159 	struct xfs_scrub	*sc)
160 {
161 	if (!(sc->flags & XCHK_FSGATES_ALL))
162 		return;
163 
164 	trace_xchk_fsgates_disable(sc, sc->flags & XCHK_FSGATES_ALL);
165 
166 	if (sc->flags & XCHK_FSGATES_DRAIN)
167 		xfs_drain_wait_disable();
168 
169 	if (sc->flags & XCHK_FSGATES_QUOTA)
170 		xfs_dqtrx_hook_disable();
171 
172 	if (sc->flags & XCHK_FSGATES_DIRENTS)
173 		xfs_dir_hook_disable();
174 
175 	if (sc->flags & XCHK_FSGATES_RMAP)
176 		xfs_rmap_hook_disable();
177 
178 	sc->flags &= ~XCHK_FSGATES_ALL;
179 }
180 
181 /* Free the resources associated with a scrub subtype. */
182 void
183 xchk_scrub_free_subord(
184 	struct xfs_scrub_subord	*sub)
185 {
186 	struct xfs_scrub	*sc = sub->parent_sc;
187 
188 	ASSERT(sc->ip == sub->sc.ip);
189 	ASSERT(sc->orphanage == sub->sc.orphanage);
190 	ASSERT(sc->tempip == sub->sc.tempip);
191 
192 	sc->sm->sm_type = sub->old_smtype;
193 	sc->sm->sm_flags = sub->old_smflags |
194 				(sc->sm->sm_flags & XFS_SCRUB_FLAGS_OUT);
195 	sc->tp = sub->sc.tp;
196 
197 	if (sub->sc.buf) {
198 		if (sub->sc.buf_cleanup)
199 			sub->sc.buf_cleanup(sub->sc.buf);
200 		kvfree(sub->sc.buf);
201 	}
202 	if (sub->sc.xmbtp)
203 		xmbuf_free(sub->sc.xmbtp);
204 	if (sub->sc.xfile)
205 		xfile_destroy(sub->sc.xfile);
206 
207 	sc->ilock_flags = sub->sc.ilock_flags;
208 	sc->orphanage_ilock_flags = sub->sc.orphanage_ilock_flags;
209 	sc->temp_ilock_flags = sub->sc.temp_ilock_flags;
210 
211 	kfree(sub);
212 }
213 
214 /* Free all the resources and finish the transactions. */
215 STATIC int
216 xchk_teardown(
217 	struct xfs_scrub	*sc,
218 	int			error)
219 {
220 	xchk_ag_free(sc, &sc->sa);
221 	if (sc->tp) {
222 		if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
223 			error = xfs_trans_commit(sc->tp);
224 		else
225 			xfs_trans_cancel(sc->tp);
226 		sc->tp = NULL;
227 	}
228 	if (sc->ip) {
229 		if (sc->ilock_flags)
230 			xchk_iunlock(sc, sc->ilock_flags);
231 		xchk_irele(sc, sc->ip);
232 		sc->ip = NULL;
233 	}
234 	if (sc->flags & XCHK_HAVE_FREEZE_PROT) {
235 		sc->flags &= ~XCHK_HAVE_FREEZE_PROT;
236 		mnt_drop_write_file(sc->file);
237 	}
238 	if (sc->xmbtp) {
239 		xmbuf_free(sc->xmbtp);
240 		sc->xmbtp = NULL;
241 	}
242 	if (sc->xfile) {
243 		xfile_destroy(sc->xfile);
244 		sc->xfile = NULL;
245 	}
246 	if (sc->buf) {
247 		if (sc->buf_cleanup)
248 			sc->buf_cleanup(sc->buf);
249 		kvfree(sc->buf);
250 		sc->buf_cleanup = NULL;
251 		sc->buf = NULL;
252 	}
253 
254 	xrep_tempfile_rele(sc);
255 	xrep_orphanage_rele(sc);
256 	xchk_fsgates_disable(sc);
257 	return error;
258 }
259 
260 /* Scrubbing dispatch. */
261 
262 static const struct xchk_meta_ops meta_scrub_ops[] = {
263 	[XFS_SCRUB_TYPE_PROBE] = {	/* ioctl presence test */
264 		.type	= ST_NONE,
265 		.setup	= xchk_setup_fs,
266 		.scrub	= xchk_probe,
267 		.repair = xrep_probe,
268 	},
269 	[XFS_SCRUB_TYPE_SB] = {		/* superblock */
270 		.type	= ST_PERAG,
271 		.setup	= xchk_setup_agheader,
272 		.scrub	= xchk_superblock,
273 		.repair	= xrep_superblock,
274 	},
275 	[XFS_SCRUB_TYPE_AGF] = {	/* agf */
276 		.type	= ST_PERAG,
277 		.setup	= xchk_setup_agheader,
278 		.scrub	= xchk_agf,
279 		.repair	= xrep_agf,
280 	},
281 	[XFS_SCRUB_TYPE_AGFL]= {	/* agfl */
282 		.type	= ST_PERAG,
283 		.setup	= xchk_setup_agheader,
284 		.scrub	= xchk_agfl,
285 		.repair	= xrep_agfl,
286 	},
287 	[XFS_SCRUB_TYPE_AGI] = {	/* agi */
288 		.type	= ST_PERAG,
289 		.setup	= xchk_setup_agheader,
290 		.scrub	= xchk_agi,
291 		.repair	= xrep_agi,
292 	},
293 	[XFS_SCRUB_TYPE_BNOBT] = {	/* bnobt */
294 		.type	= ST_PERAG,
295 		.setup	= xchk_setup_ag_allocbt,
296 		.scrub	= xchk_allocbt,
297 		.repair	= xrep_allocbt,
298 		.repair_eval = xrep_revalidate_allocbt,
299 	},
300 	[XFS_SCRUB_TYPE_CNTBT] = {	/* cntbt */
301 		.type	= ST_PERAG,
302 		.setup	= xchk_setup_ag_allocbt,
303 		.scrub	= xchk_allocbt,
304 		.repair	= xrep_allocbt,
305 		.repair_eval = xrep_revalidate_allocbt,
306 	},
307 	[XFS_SCRUB_TYPE_INOBT] = {	/* inobt */
308 		.type	= ST_PERAG,
309 		.setup	= xchk_setup_ag_iallocbt,
310 		.scrub	= xchk_iallocbt,
311 		.repair	= xrep_iallocbt,
312 		.repair_eval = xrep_revalidate_iallocbt,
313 	},
314 	[XFS_SCRUB_TYPE_FINOBT] = {	/* finobt */
315 		.type	= ST_PERAG,
316 		.setup	= xchk_setup_ag_iallocbt,
317 		.scrub	= xchk_iallocbt,
318 		.has	= xfs_has_finobt,
319 		.repair	= xrep_iallocbt,
320 		.repair_eval = xrep_revalidate_iallocbt,
321 	},
322 	[XFS_SCRUB_TYPE_RMAPBT] = {	/* rmapbt */
323 		.type	= ST_PERAG,
324 		.setup	= xchk_setup_ag_rmapbt,
325 		.scrub	= xchk_rmapbt,
326 		.has	= xfs_has_rmapbt,
327 		.repair	= xrep_rmapbt,
328 	},
329 	[XFS_SCRUB_TYPE_REFCNTBT] = {	/* refcountbt */
330 		.type	= ST_PERAG,
331 		.setup	= xchk_setup_ag_refcountbt,
332 		.scrub	= xchk_refcountbt,
333 		.has	= xfs_has_reflink,
334 		.repair	= xrep_refcountbt,
335 	},
336 	[XFS_SCRUB_TYPE_INODE] = {	/* inode record */
337 		.type	= ST_INODE,
338 		.setup	= xchk_setup_inode,
339 		.scrub	= xchk_inode,
340 		.repair	= xrep_inode,
341 	},
342 	[XFS_SCRUB_TYPE_BMBTD] = {	/* inode data fork */
343 		.type	= ST_INODE,
344 		.setup	= xchk_setup_inode_bmap,
345 		.scrub	= xchk_bmap_data,
346 		.repair	= xrep_bmap_data,
347 	},
348 	[XFS_SCRUB_TYPE_BMBTA] = {	/* inode attr fork */
349 		.type	= ST_INODE,
350 		.setup	= xchk_setup_inode_bmap,
351 		.scrub	= xchk_bmap_attr,
352 		.repair	= xrep_bmap_attr,
353 	},
354 	[XFS_SCRUB_TYPE_BMBTC] = {	/* inode CoW fork */
355 		.type	= ST_INODE,
356 		.setup	= xchk_setup_inode_bmap,
357 		.scrub	= xchk_bmap_cow,
358 		.repair	= xrep_bmap_cow,
359 	},
360 	[XFS_SCRUB_TYPE_DIR] = {	/* directory */
361 		.type	= ST_INODE,
362 		.setup	= xchk_setup_directory,
363 		.scrub	= xchk_directory,
364 		.repair	= xrep_directory,
365 	},
366 	[XFS_SCRUB_TYPE_XATTR] = {	/* extended attributes */
367 		.type	= ST_INODE,
368 		.setup	= xchk_setup_xattr,
369 		.scrub	= xchk_xattr,
370 		.repair	= xrep_xattr,
371 	},
372 	[XFS_SCRUB_TYPE_SYMLINK] = {	/* symbolic link */
373 		.type	= ST_INODE,
374 		.setup	= xchk_setup_symlink,
375 		.scrub	= xchk_symlink,
376 		.repair	= xrep_symlink,
377 	},
378 	[XFS_SCRUB_TYPE_PARENT] = {	/* parent pointers */
379 		.type	= ST_INODE,
380 		.setup	= xchk_setup_parent,
381 		.scrub	= xchk_parent,
382 		.repair	= xrep_parent,
383 	},
384 	[XFS_SCRUB_TYPE_RTBITMAP] = {	/* realtime bitmap */
385 		.type	= ST_FS,
386 		.setup	= xchk_setup_rtbitmap,
387 		.scrub	= xchk_rtbitmap,
388 		.repair	= xrep_rtbitmap,
389 	},
390 	[XFS_SCRUB_TYPE_RTSUM] = {	/* realtime summary */
391 		.type	= ST_FS,
392 		.setup	= xchk_setup_rtsummary,
393 		.scrub	= xchk_rtsummary,
394 		.repair	= xrep_rtsummary,
395 	},
396 	[XFS_SCRUB_TYPE_UQUOTA] = {	/* user quota */
397 		.type	= ST_FS,
398 		.setup	= xchk_setup_quota,
399 		.scrub	= xchk_quota,
400 		.repair	= xrep_quota,
401 	},
402 	[XFS_SCRUB_TYPE_GQUOTA] = {	/* group quota */
403 		.type	= ST_FS,
404 		.setup	= xchk_setup_quota,
405 		.scrub	= xchk_quota,
406 		.repair	= xrep_quota,
407 	},
408 	[XFS_SCRUB_TYPE_PQUOTA] = {	/* project quota */
409 		.type	= ST_FS,
410 		.setup	= xchk_setup_quota,
411 		.scrub	= xchk_quota,
412 		.repair	= xrep_quota,
413 	},
414 	[XFS_SCRUB_TYPE_FSCOUNTERS] = {	/* fs summary counters */
415 		.type	= ST_FS,
416 		.setup	= xchk_setup_fscounters,
417 		.scrub	= xchk_fscounters,
418 		.repair	= xrep_fscounters,
419 	},
420 	[XFS_SCRUB_TYPE_QUOTACHECK] = {	/* quota counters */
421 		.type	= ST_FS,
422 		.setup	= xchk_setup_quotacheck,
423 		.scrub	= xchk_quotacheck,
424 		.repair	= xrep_quotacheck,
425 	},
426 	[XFS_SCRUB_TYPE_NLINKS] = {	/* inode link counts */
427 		.type	= ST_FS,
428 		.setup	= xchk_setup_nlinks,
429 		.scrub	= xchk_nlinks,
430 		.repair	= xrep_nlinks,
431 	},
432 	[XFS_SCRUB_TYPE_HEALTHY] = {	/* fs healthy; clean all reminders */
433 		.type	= ST_FS,
434 		.setup	= xchk_setup_fs,
435 		.scrub	= xchk_health_record,
436 		.repair = xrep_notsupported,
437 	},
438 	[XFS_SCRUB_TYPE_DIRTREE] = {	/* directory tree structure */
439 		.type	= ST_INODE,
440 		.setup	= xchk_setup_dirtree,
441 		.scrub	= xchk_dirtree,
442 		.has	= xfs_has_parent,
443 		.repair	= xrep_dirtree,
444 	},
445 };
446 
447 static int
448 xchk_validate_inputs(
449 	struct xfs_mount		*mp,
450 	struct xfs_scrub_metadata	*sm)
451 {
452 	int				error;
453 	const struct xchk_meta_ops	*ops;
454 
455 	error = -EINVAL;
456 	/* Check our inputs. */
457 	sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
458 	if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
459 		goto out;
460 	/* sm_reserved[] must be zero */
461 	if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
462 		goto out;
463 
464 	error = -ENOENT;
465 	/* Do we know about this type of metadata? */
466 	if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
467 		goto out;
468 	ops = &meta_scrub_ops[sm->sm_type];
469 	if (ops->setup == NULL || ops->scrub == NULL)
470 		goto out;
471 	/* Does this fs even support this type of metadata? */
472 	if (ops->has && !ops->has(mp))
473 		goto out;
474 
475 	error = -EINVAL;
476 	/* restricting fields must be appropriate for type */
477 	switch (ops->type) {
478 	case ST_NONE:
479 	case ST_FS:
480 		if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
481 			goto out;
482 		break;
483 	case ST_PERAG:
484 		if (sm->sm_ino || sm->sm_gen ||
485 		    sm->sm_agno >= mp->m_sb.sb_agcount)
486 			goto out;
487 		break;
488 	case ST_INODE:
489 		if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
490 			goto out;
491 		break;
492 	default:
493 		goto out;
494 	}
495 
496 	/* No rebuild without repair. */
497 	if ((sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) &&
498 	    !(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
499 		return -EINVAL;
500 
501 	/*
502 	 * We only want to repair read-write v5+ filesystems.  Defer the check
503 	 * for ops->repair until after our scrub confirms that we need to
504 	 * perform repairs so that we avoid failing due to not supporting
505 	 * repairing an object that doesn't need repairs.
506 	 */
507 	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
508 		error = -EOPNOTSUPP;
509 		if (!xfs_has_crc(mp))
510 			goto out;
511 
512 		error = -EROFS;
513 		if (xfs_is_readonly(mp))
514 			goto out;
515 	}
516 
517 	error = 0;
518 out:
519 	return error;
520 }
521 
522 #ifdef CONFIG_XFS_ONLINE_REPAIR
523 static inline void xchk_postmortem(struct xfs_scrub *sc)
524 {
525 	/*
526 	 * Userspace asked us to repair something, we repaired it, rescanned
527 	 * it, and the rescan says it's still broken.  Scream about this in
528 	 * the system logs.
529 	 */
530 	if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
531 	    (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
532 				 XFS_SCRUB_OFLAG_XCORRUPT)))
533 		xrep_failure(sc->mp);
534 }
535 #else
536 static inline void xchk_postmortem(struct xfs_scrub *sc)
537 {
538 	/*
539 	 * Userspace asked us to scrub something, it's broken, and we have no
540 	 * way of fixing it.  Scream in the logs.
541 	 */
542 	if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
543 				XFS_SCRUB_OFLAG_XCORRUPT))
544 		xfs_alert_ratelimited(sc->mp,
545 				"Corruption detected during scrub.");
546 }
547 #endif /* CONFIG_XFS_ONLINE_REPAIR */
548 
549 /*
550  * Create a new scrub context from an existing one, but with a different scrub
551  * type.
552  */
553 struct xfs_scrub_subord *
554 xchk_scrub_create_subord(
555 	struct xfs_scrub	*sc,
556 	unsigned int		subtype)
557 {
558 	struct xfs_scrub_subord	*sub;
559 
560 	sub = kzalloc(sizeof(*sub), XCHK_GFP_FLAGS);
561 	if (!sub)
562 		return ERR_PTR(-ENOMEM);
563 
564 	sub->old_smtype = sc->sm->sm_type;
565 	sub->old_smflags = sc->sm->sm_flags;
566 	sub->parent_sc = sc;
567 	memcpy(&sub->sc, sc, sizeof(struct xfs_scrub));
568 	sub->sc.ops = &meta_scrub_ops[subtype];
569 	sub->sc.sm->sm_type = subtype;
570 	sub->sc.sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
571 	sub->sc.buf = NULL;
572 	sub->sc.buf_cleanup = NULL;
573 	sub->sc.xfile = NULL;
574 	sub->sc.xmbtp = NULL;
575 
576 	return sub;
577 }
578 
579 /* Dispatch metadata scrubbing. */
580 STATIC int
581 xfs_scrub_metadata(
582 	struct file			*file,
583 	struct xfs_scrub_metadata	*sm)
584 {
585 	struct xchk_stats_run		run = { };
586 	struct xfs_scrub		*sc;
587 	struct xfs_mount		*mp = XFS_I(file_inode(file))->i_mount;
588 	u64				check_start;
589 	int				error = 0;
590 
591 	BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
592 		(sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
593 
594 	trace_xchk_start(XFS_I(file_inode(file)), sm, error);
595 
596 	/* Forbidden if we are shut down or mounted norecovery. */
597 	error = -ESHUTDOWN;
598 	if (xfs_is_shutdown(mp))
599 		goto out;
600 	error = -ENOTRECOVERABLE;
601 	if (xfs_has_norecovery(mp))
602 		goto out;
603 
604 	error = xchk_validate_inputs(mp, sm);
605 	if (error)
606 		goto out;
607 
608 	xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SCRUB,
609  "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
610 
611 	sc = kzalloc(sizeof(struct xfs_scrub), XCHK_GFP_FLAGS);
612 	if (!sc) {
613 		error = -ENOMEM;
614 		goto out;
615 	}
616 
617 	sc->mp = mp;
618 	sc->file = file;
619 	sc->sm = sm;
620 	sc->ops = &meta_scrub_ops[sm->sm_type];
621 	sc->sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
622 	sc->relax = INIT_XCHK_RELAX;
623 retry_op:
624 	/*
625 	 * When repairs are allowed, prevent freezing or readonly remount while
626 	 * scrub is running with a real transaction.
627 	 */
628 	if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
629 		error = mnt_want_write_file(sc->file);
630 		if (error)
631 			goto out_sc;
632 
633 		sc->flags |= XCHK_HAVE_FREEZE_PROT;
634 	}
635 
636 	/* Set up for the operation. */
637 	error = sc->ops->setup(sc);
638 	if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
639 		goto try_harder;
640 	if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
641 		goto need_drain;
642 	if (error)
643 		goto out_teardown;
644 
645 	/* Scrub for errors. */
646 	check_start = xchk_stats_now();
647 	if ((sc->flags & XREP_ALREADY_FIXED) && sc->ops->repair_eval != NULL)
648 		error = sc->ops->repair_eval(sc);
649 	else
650 		error = sc->ops->scrub(sc);
651 	run.scrub_ns += xchk_stats_elapsed_ns(check_start);
652 	if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER))
653 		goto try_harder;
654 	if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN))
655 		goto need_drain;
656 	if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE))
657 		goto out_teardown;
658 
659 	xchk_update_health(sc);
660 
661 	if (xchk_could_repair(sc)) {
662 		/*
663 		 * If userspace asked for a repair but it wasn't necessary,
664 		 * report that back to userspace.
665 		 */
666 		if (!xrep_will_attempt(sc)) {
667 			sc->sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
668 			goto out_nofix;
669 		}
670 
671 		/*
672 		 * If it's broken, userspace wants us to fix it, and we haven't
673 		 * already tried to fix it, then attempt a repair.
674 		 */
675 		error = xrep_attempt(sc, &run);
676 		if (error == -EAGAIN) {
677 			/*
678 			 * Either the repair function succeeded or it couldn't
679 			 * get all the resources it needs; either way, we go
680 			 * back to the beginning and call the scrub function.
681 			 */
682 			error = xchk_teardown(sc, 0);
683 			if (error) {
684 				xrep_failure(mp);
685 				goto out_sc;
686 			}
687 			goto retry_op;
688 		}
689 	}
690 
691 out_nofix:
692 	xchk_postmortem(sc);
693 out_teardown:
694 	error = xchk_teardown(sc, error);
695 out_sc:
696 	if (error != -ENOENT)
697 		xchk_stats_merge(mp, sm, &run);
698 	kfree(sc);
699 out:
700 	trace_xchk_done(XFS_I(file_inode(file)), sm, error);
701 	if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
702 		sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
703 		error = 0;
704 	}
705 	return error;
706 need_drain:
707 	error = xchk_teardown(sc, 0);
708 	if (error)
709 		goto out_sc;
710 	sc->flags |= XCHK_NEED_DRAIN;
711 	run.retries++;
712 	goto retry_op;
713 try_harder:
714 	/*
715 	 * Scrubbers return -EDEADLOCK to mean 'try harder'.  Tear down
716 	 * everything we hold, then set up again with preparation for
717 	 * worst-case scenarios.
718 	 */
719 	error = xchk_teardown(sc, 0);
720 	if (error)
721 		goto out_sc;
722 	sc->flags |= XCHK_TRY_HARDER;
723 	run.retries++;
724 	goto retry_op;
725 }
726 
727 /* Scrub one aspect of one piece of metadata. */
728 int
729 xfs_ioc_scrub_metadata(
730 	struct file			*file,
731 	void				__user *arg)
732 {
733 	struct xfs_scrub_metadata	scrub;
734 	int				error;
735 
736 	if (!capable(CAP_SYS_ADMIN))
737 		return -EPERM;
738 
739 	if (copy_from_user(&scrub, arg, sizeof(scrub)))
740 		return -EFAULT;
741 
742 	error = xfs_scrub_metadata(file, &scrub);
743 	if (error)
744 		return error;
745 
746 	if (copy_to_user(arg, &scrub, sizeof(scrub)))
747 		return -EFAULT;
748 
749 	return 0;
750 }
751 
752 /* Decide if there have been any scrub failures up to this point. */
753 static inline int
754 xfs_scrubv_check_barrier(
755 	struct xfs_mount		*mp,
756 	const struct xfs_scrub_vec	*vectors,
757 	const struct xfs_scrub_vec	*stop_vec)
758 {
759 	const struct xfs_scrub_vec	*v;
760 	__u32				failmask;
761 
762 	failmask = stop_vec->sv_flags & XFS_SCRUB_FLAGS_OUT;
763 
764 	for (v = vectors; v < stop_vec; v++) {
765 		if (v->sv_type == XFS_SCRUB_TYPE_BARRIER)
766 			continue;
767 
768 		/*
769 		 * Runtime errors count as a previous failure, except the ones
770 		 * used to ask userspace to retry.
771 		 */
772 		switch (v->sv_ret) {
773 		case -EBUSY:
774 		case -ENOENT:
775 		case -EUSERS:
776 		case 0:
777 			break;
778 		default:
779 			return -ECANCELED;
780 		}
781 
782 		/*
783 		 * If any of the out-flags on the scrub vector match the mask
784 		 * that was set on the barrier vector, that's a previous fail.
785 		 */
786 		if (v->sv_flags & failmask)
787 			return -ECANCELED;
788 	}
789 
790 	return 0;
791 }
792 
793 /*
794  * If the caller provided us with a nonzero inode number that isn't the ioctl
795  * file, try to grab a reference to it to eliminate all further untrusted inode
796  * lookups.  If we can't get the inode, let each scrub function try again.
797  */
798 STATIC struct xfs_inode *
799 xchk_scrubv_open_by_handle(
800 	struct xfs_mount		*mp,
801 	const struct xfs_scrub_vec_head	*head)
802 {
803 	struct xfs_trans		*tp;
804 	struct xfs_inode		*ip;
805 	int				error;
806 
807 	error = xfs_trans_alloc_empty(mp, &tp);
808 	if (error)
809 		return NULL;
810 
811 	error = xfs_iget(mp, tp, head->svh_ino, XCHK_IGET_FLAGS, 0, &ip);
812 	xfs_trans_cancel(tp);
813 	if (error)
814 		return NULL;
815 
816 	if (VFS_I(ip)->i_generation != head->svh_gen) {
817 		xfs_irele(ip);
818 		return NULL;
819 	}
820 
821 	return ip;
822 }
823 
824 /* Vectored scrub implementation to reduce ioctl calls. */
825 int
826 xfs_ioc_scrubv_metadata(
827 	struct file			*file,
828 	void				__user *arg)
829 {
830 	struct xfs_scrub_vec_head	head;
831 	struct xfs_scrub_vec_head	__user *uhead = arg;
832 	struct xfs_scrub_vec		*vectors;
833 	struct xfs_scrub_vec		__user *uvectors;
834 	struct xfs_inode		*ip_in = XFS_I(file_inode(file));
835 	struct xfs_mount		*mp = ip_in->i_mount;
836 	struct xfs_inode		*handle_ip = NULL;
837 	struct xfs_scrub_vec		*v;
838 	size_t				vec_bytes;
839 	unsigned int			i;
840 	int				error = 0;
841 
842 	if (!capable(CAP_SYS_ADMIN))
843 		return -EPERM;
844 
845 	if (copy_from_user(&head, uhead, sizeof(head)))
846 		return -EFAULT;
847 
848 	if (head.svh_reserved)
849 		return -EINVAL;
850 	if (head.svh_flags & ~XFS_SCRUB_VEC_FLAGS_ALL)
851 		return -EINVAL;
852 	if (head.svh_nr == 0)
853 		return 0;
854 
855 	vec_bytes = array_size(head.svh_nr, sizeof(struct xfs_scrub_vec));
856 	if (vec_bytes > PAGE_SIZE)
857 		return -ENOMEM;
858 
859 	uvectors = u64_to_user_ptr(head.svh_vectors);
860 	vectors = memdup_user(uvectors, vec_bytes);
861 	if (IS_ERR(vectors))
862 		return PTR_ERR(vectors);
863 
864 	trace_xchk_scrubv_start(ip_in, &head);
865 
866 	for (i = 0, v = vectors; i < head.svh_nr; i++, v++) {
867 		if (v->sv_reserved) {
868 			error = -EINVAL;
869 			goto out_free;
870 		}
871 
872 		if (v->sv_type == XFS_SCRUB_TYPE_BARRIER &&
873 		    (v->sv_flags & ~XFS_SCRUB_FLAGS_OUT)) {
874 			error = -EINVAL;
875 			goto out_free;
876 		}
877 
878 		trace_xchk_scrubv_item(mp, &head, i, v);
879 	}
880 
881 	/*
882 	 * If the caller wants us to do a scrub-by-handle and the file used to
883 	 * call the ioctl is not the same file, load the incore inode and pin
884 	 * it across all the scrubv actions to avoid repeated UNTRUSTED
885 	 * lookups.  The reference is not passed to deeper layers of scrub
886 	 * because each scrubber gets to decide its own strategy and return
887 	 * values for getting an inode.
888 	 */
889 	if (head.svh_ino && head.svh_ino != ip_in->i_ino)
890 		handle_ip = xchk_scrubv_open_by_handle(mp, &head);
891 
892 	/* Run all the scrubbers. */
893 	for (i = 0, v = vectors; i < head.svh_nr; i++, v++) {
894 		struct xfs_scrub_metadata	sm = {
895 			.sm_type		= v->sv_type,
896 			.sm_flags		= v->sv_flags,
897 			.sm_ino			= head.svh_ino,
898 			.sm_gen			= head.svh_gen,
899 			.sm_agno		= head.svh_agno,
900 		};
901 
902 		if (v->sv_type == XFS_SCRUB_TYPE_BARRIER) {
903 			v->sv_ret = xfs_scrubv_check_barrier(mp, vectors, v);
904 			if (v->sv_ret) {
905 				trace_xchk_scrubv_barrier_fail(mp, &head, i, v);
906 				break;
907 			}
908 
909 			continue;
910 		}
911 
912 		v->sv_ret = xfs_scrub_metadata(file, &sm);
913 		v->sv_flags = sm.sm_flags;
914 
915 		trace_xchk_scrubv_outcome(mp, &head, i, v);
916 
917 		if (head.svh_rest_us) {
918 			ktime_t		expires;
919 
920 			expires = ktime_add_ns(ktime_get(),
921 					head.svh_rest_us * 1000);
922 			set_current_state(TASK_KILLABLE);
923 			schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
924 		}
925 
926 		if (fatal_signal_pending(current)) {
927 			error = -EINTR;
928 			goto out_free;
929 		}
930 	}
931 
932 	if (copy_to_user(uvectors, vectors, vec_bytes) ||
933 	    copy_to_user(uhead, &head, sizeof(head))) {
934 		error = -EFAULT;
935 		goto out_free;
936 	}
937 
938 out_free:
939 	if (handle_ip)
940 		xfs_irele(handle_ip);
941 	kfree(vectors);
942 	return error;
943 }
944