Merge tag 'for-linus-hmm' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
[linux-2.6-microblaze.git] / fs / xfs / scrub / scrub.c
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  */
134 static int
135 xchk_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. */
149 STATIC int
150 xchk_teardown(
151         struct xfs_scrub        *sc,
152         struct xfs_inode        *ip_in,
153         int                     error)
154 {
155         xchk_ag_free(sc, &sc->sa);
156         if (sc->tp) {
157                 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
158                         error = xfs_trans_commit(sc->tp);
159                 else
160                         xfs_trans_cancel(sc->tp);
161                 sc->tp = NULL;
162         }
163         if (sc->ip) {
164                 if (sc->ilock_flags)
165                         xfs_iunlock(sc->ip, sc->ilock_flags);
166                 if (sc->ip != ip_in &&
167                     !xfs_internal_inum(sc->mp, sc->ip->i_ino))
168                         xfs_irele(sc->ip);
169                 sc->ip = NULL;
170         }
171         if (sc->flags & XCHK_REAPING_DISABLED)
172                 xchk_start_reaping(sc);
173         if (sc->flags & XCHK_HAS_QUOTAOFFLOCK) {
174                 mutex_unlock(&sc->mp->m_quotainfo->qi_quotaofflock);
175                 sc->flags &= ~XCHK_HAS_QUOTAOFFLOCK;
176         }
177         if (sc->buf) {
178                 kmem_free(sc->buf);
179                 sc->buf = NULL;
180         }
181         return error;
182 }
183
184 /* Scrubbing dispatch. */
185
186 static const struct xchk_meta_ops meta_scrub_ops[] = {
187         [XFS_SCRUB_TYPE_PROBE] = {      /* ioctl presence test */
188                 .type   = ST_NONE,
189                 .setup  = xchk_setup_fs,
190                 .scrub  = xchk_probe,
191                 .repair = xrep_probe,
192         },
193         [XFS_SCRUB_TYPE_SB] = {         /* superblock */
194                 .type   = ST_PERAG,
195                 .setup  = xchk_setup_fs,
196                 .scrub  = xchk_superblock,
197                 .repair = xrep_superblock,
198         },
199         [XFS_SCRUB_TYPE_AGF] = {        /* agf */
200                 .type   = ST_PERAG,
201                 .setup  = xchk_setup_fs,
202                 .scrub  = xchk_agf,
203                 .repair = xrep_agf,
204         },
205         [XFS_SCRUB_TYPE_AGFL]= {        /* agfl */
206                 .type   = ST_PERAG,
207                 .setup  = xchk_setup_fs,
208                 .scrub  = xchk_agfl,
209                 .repair = xrep_agfl,
210         },
211         [XFS_SCRUB_TYPE_AGI] = {        /* agi */
212                 .type   = ST_PERAG,
213                 .setup  = xchk_setup_fs,
214                 .scrub  = xchk_agi,
215                 .repair = xrep_agi,
216         },
217         [XFS_SCRUB_TYPE_BNOBT] = {      /* bnobt */
218                 .type   = ST_PERAG,
219                 .setup  = xchk_setup_ag_allocbt,
220                 .scrub  = xchk_bnobt,
221                 .repair = xrep_notsupported,
222         },
223         [XFS_SCRUB_TYPE_CNTBT] = {      /* cntbt */
224                 .type   = ST_PERAG,
225                 .setup  = xchk_setup_ag_allocbt,
226                 .scrub  = xchk_cntbt,
227                 .repair = xrep_notsupported,
228         },
229         [XFS_SCRUB_TYPE_INOBT] = {      /* inobt */
230                 .type   = ST_PERAG,
231                 .setup  = xchk_setup_ag_iallocbt,
232                 .scrub  = xchk_inobt,
233                 .repair = xrep_notsupported,
234         },
235         [XFS_SCRUB_TYPE_FINOBT] = {     /* finobt */
236                 .type   = ST_PERAG,
237                 .setup  = xchk_setup_ag_iallocbt,
238                 .scrub  = xchk_finobt,
239                 .has    = xfs_sb_version_hasfinobt,
240                 .repair = xrep_notsupported,
241         },
242         [XFS_SCRUB_TYPE_RMAPBT] = {     /* rmapbt */
243                 .type   = ST_PERAG,
244                 .setup  = xchk_setup_ag_rmapbt,
245                 .scrub  = xchk_rmapbt,
246                 .has    = xfs_sb_version_hasrmapbt,
247                 .repair = xrep_notsupported,
248         },
249         [XFS_SCRUB_TYPE_REFCNTBT] = {   /* refcountbt */
250                 .type   = ST_PERAG,
251                 .setup  = xchk_setup_ag_refcountbt,
252                 .scrub  = xchk_refcountbt,
253                 .has    = xfs_sb_version_hasreflink,
254                 .repair = xrep_notsupported,
255         },
256         [XFS_SCRUB_TYPE_INODE] = {      /* inode record */
257                 .type   = ST_INODE,
258                 .setup  = xchk_setup_inode,
259                 .scrub  = xchk_inode,
260                 .repair = xrep_notsupported,
261         },
262         [XFS_SCRUB_TYPE_BMBTD] = {      /* inode data fork */
263                 .type   = ST_INODE,
264                 .setup  = xchk_setup_inode_bmap,
265                 .scrub  = xchk_bmap_data,
266                 .repair = xrep_notsupported,
267         },
268         [XFS_SCRUB_TYPE_BMBTA] = {      /* inode attr fork */
269                 .type   = ST_INODE,
270                 .setup  = xchk_setup_inode_bmap,
271                 .scrub  = xchk_bmap_attr,
272                 .repair = xrep_notsupported,
273         },
274         [XFS_SCRUB_TYPE_BMBTC] = {      /* inode CoW fork */
275                 .type   = ST_INODE,
276                 .setup  = xchk_setup_inode_bmap,
277                 .scrub  = xchk_bmap_cow,
278                 .repair = xrep_notsupported,
279         },
280         [XFS_SCRUB_TYPE_DIR] = {        /* directory */
281                 .type   = ST_INODE,
282                 .setup  = xchk_setup_directory,
283                 .scrub  = xchk_directory,
284                 .repair = xrep_notsupported,
285         },
286         [XFS_SCRUB_TYPE_XATTR] = {      /* extended attributes */
287                 .type   = ST_INODE,
288                 .setup  = xchk_setup_xattr,
289                 .scrub  = xchk_xattr,
290                 .repair = xrep_notsupported,
291         },
292         [XFS_SCRUB_TYPE_SYMLINK] = {    /* symbolic link */
293                 .type   = ST_INODE,
294                 .setup  = xchk_setup_symlink,
295                 .scrub  = xchk_symlink,
296                 .repair = xrep_notsupported,
297         },
298         [XFS_SCRUB_TYPE_PARENT] = {     /* parent pointers */
299                 .type   = ST_INODE,
300                 .setup  = xchk_setup_parent,
301                 .scrub  = xchk_parent,
302                 .repair = xrep_notsupported,
303         },
304         [XFS_SCRUB_TYPE_RTBITMAP] = {   /* realtime bitmap */
305                 .type   = ST_FS,
306                 .setup  = xchk_setup_rt,
307                 .scrub  = xchk_rtbitmap,
308                 .has    = xfs_sb_version_hasrealtime,
309                 .repair = xrep_notsupported,
310         },
311         [XFS_SCRUB_TYPE_RTSUM] = {      /* realtime summary */
312                 .type   = ST_FS,
313                 .setup  = xchk_setup_rt,
314                 .scrub  = xchk_rtsummary,
315                 .has    = xfs_sb_version_hasrealtime,
316                 .repair = xrep_notsupported,
317         },
318         [XFS_SCRUB_TYPE_UQUOTA] = {     /* user quota */
319                 .type   = ST_FS,
320                 .setup  = xchk_setup_quota,
321                 .scrub  = xchk_quota,
322                 .repair = xrep_notsupported,
323         },
324         [XFS_SCRUB_TYPE_GQUOTA] = {     /* group quota */
325                 .type   = ST_FS,
326                 .setup  = xchk_setup_quota,
327                 .scrub  = xchk_quota,
328                 .repair = xrep_notsupported,
329         },
330         [XFS_SCRUB_TYPE_PQUOTA] = {     /* project quota */
331                 .type   = ST_FS,
332                 .setup  = xchk_setup_quota,
333                 .scrub  = xchk_quota,
334                 .repair = xrep_notsupported,
335         },
336         [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */
337                 .type   = ST_FS,
338                 .setup  = xchk_setup_fscounters,
339                 .scrub  = xchk_fscounters,
340                 .repair = xrep_notsupported,
341         },
342 };
343
344 /* This isn't a stable feature, warn once per day. */
345 static inline void
346 xchk_experimental_warning(
347         struct xfs_mount        *mp)
348 {
349         static struct ratelimit_state scrub_warning = RATELIMIT_STATE_INIT(
350                         "xchk_warning", 86400 * HZ, 1);
351         ratelimit_set_flags(&scrub_warning, RATELIMIT_MSG_ON_RELEASE);
352
353         if (__ratelimit(&scrub_warning))
354                 xfs_alert(mp,
355 "EXPERIMENTAL online scrub feature in use. Use at your own risk!");
356 }
357
358 static int
359 xchk_validate_inputs(
360         struct xfs_mount                *mp,
361         struct xfs_scrub_metadata       *sm)
362 {
363         int                             error;
364         const struct xchk_meta_ops      *ops;
365
366         error = -EINVAL;
367         /* Check our inputs. */
368         sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
369         if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN)
370                 goto out;
371         /* sm_reserved[] must be zero */
372         if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved)))
373                 goto out;
374
375         error = -ENOENT;
376         /* Do we know about this type of metadata? */
377         if (sm->sm_type >= XFS_SCRUB_TYPE_NR)
378                 goto out;
379         ops = &meta_scrub_ops[sm->sm_type];
380         if (ops->setup == NULL || ops->scrub == NULL)
381                 goto out;
382         /* Does this fs even support this type of metadata? */
383         if (ops->has && !ops->has(&mp->m_sb))
384                 goto out;
385
386         error = -EINVAL;
387         /* restricting fields must be appropriate for type */
388         switch (ops->type) {
389         case ST_NONE:
390         case ST_FS:
391                 if (sm->sm_ino || sm->sm_gen || sm->sm_agno)
392                         goto out;
393                 break;
394         case ST_PERAG:
395                 if (sm->sm_ino || sm->sm_gen ||
396                     sm->sm_agno >= mp->m_sb.sb_agcount)
397                         goto out;
398                 break;
399         case ST_INODE:
400                 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino))
401                         goto out;
402                 break;
403         default:
404                 goto out;
405         }
406
407         /*
408          * We only want to repair read-write v5+ filesystems.  Defer the check
409          * for ops->repair until after our scrub confirms that we need to
410          * perform repairs so that we avoid failing due to not supporting
411          * repairing an object that doesn't need repairs.
412          */
413         if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) {
414                 error = -EOPNOTSUPP;
415                 if (!xfs_sb_version_hascrc(&mp->m_sb))
416                         goto out;
417
418                 error = -EROFS;
419                 if (mp->m_flags & XFS_MOUNT_RDONLY)
420                         goto out;
421         }
422
423         error = 0;
424 out:
425         return error;
426 }
427
428 #ifdef CONFIG_XFS_ONLINE_REPAIR
429 static inline void xchk_postmortem(struct xfs_scrub *sc)
430 {
431         /*
432          * Userspace asked us to repair something, we repaired it, rescanned
433          * it, and the rescan says it's still broken.  Scream about this in
434          * the system logs.
435          */
436         if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
437             (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
438                                  XFS_SCRUB_OFLAG_XCORRUPT)))
439                 xrep_failure(sc->mp);
440 }
441 #else
442 static inline void xchk_postmortem(struct xfs_scrub *sc)
443 {
444         /*
445          * Userspace asked us to scrub something, it's broken, and we have no
446          * way of fixing it.  Scream in the logs.
447          */
448         if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
449                                 XFS_SCRUB_OFLAG_XCORRUPT))
450                 xfs_alert_ratelimited(sc->mp,
451                                 "Corruption detected during scrub.");
452 }
453 #endif /* CONFIG_XFS_ONLINE_REPAIR */
454
455 /* Dispatch metadata scrubbing. */
456 int
457 xfs_scrub_metadata(
458         struct xfs_inode                *ip,
459         struct xfs_scrub_metadata       *sm)
460 {
461         struct xfs_scrub                sc = {
462                 .mp                     = ip->i_mount,
463                 .sm                     = sm,
464                 .sa                     = {
465                         .agno           = NULLAGNUMBER,
466                 },
467         };
468         struct xfs_mount                *mp = ip->i_mount;
469         int                             error = 0;
470
471         BUILD_BUG_ON(sizeof(meta_scrub_ops) !=
472                 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR));
473
474         trace_xchk_start(ip, sm, error);
475
476         /* Forbidden if we are shut down or mounted norecovery. */
477         error = -ESHUTDOWN;
478         if (XFS_FORCED_SHUTDOWN(mp))
479                 goto out;
480         error = -ENOTRECOVERABLE;
481         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
482                 goto out;
483
484         error = xchk_validate_inputs(mp, sm);
485         if (error)
486                 goto out;
487
488         xchk_experimental_warning(mp);
489
490         sc.ops = &meta_scrub_ops[sm->sm_type];
491         sc.sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type);
492 retry_op:
493         /* Set up for the operation. */
494         error = sc.ops->setup(&sc, ip);
495         if (error)
496                 goto out_teardown;
497
498         /* Scrub for errors. */
499         error = sc.ops->scrub(&sc);
500         if (!(sc.flags & XCHK_TRY_HARDER) && error == -EDEADLOCK) {
501                 /*
502                  * Scrubbers return -EDEADLOCK to mean 'try harder'.
503                  * Tear down everything we hold, then set up again with
504                  * preparation for worst-case scenarios.
505                  */
506                 error = xchk_teardown(&sc, ip, 0);
507                 if (error)
508                         goto out;
509                 sc.flags |= XCHK_TRY_HARDER;
510                 goto retry_op;
511         } else if (error)
512                 goto out_teardown;
513
514         xchk_update_health(&sc);
515
516         if ((sc.sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) &&
517             !(sc.flags & XREP_ALREADY_FIXED)) {
518                 bool needs_fix;
519
520                 /* Let debug users force us into the repair routines. */
521                 if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR))
522                         sc.sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
523
524                 needs_fix = (sc.sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT |
525                                                 XFS_SCRUB_OFLAG_XCORRUPT |
526                                                 XFS_SCRUB_OFLAG_PREEN));
527                 /*
528                  * If userspace asked for a repair but it wasn't necessary,
529                  * report that back to userspace.
530                  */
531                 if (!needs_fix) {
532                         sc.sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED;
533                         goto out_nofix;
534                 }
535
536                 /*
537                  * If it's broken, userspace wants us to fix it, and we haven't
538                  * already tried to fix it, then attempt a repair.
539                  */
540                 error = xrep_attempt(ip, &sc);
541                 if (error == -EAGAIN) {
542                         /*
543                          * Either the repair function succeeded or it couldn't
544                          * get all the resources it needs; either way, we go
545                          * back to the beginning and call the scrub function.
546                          */
547                         error = xchk_teardown(&sc, ip, 0);
548                         if (error) {
549                                 xrep_failure(mp);
550                                 goto out;
551                         }
552                         goto retry_op;
553                 }
554         }
555
556 out_nofix:
557         xchk_postmortem(&sc);
558 out_teardown:
559         error = xchk_teardown(&sc, ip, error);
560 out:
561         trace_xchk_done(ip, sm, error);
562         if (error == -EFSCORRUPTED || error == -EFSBADCRC) {
563                 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT;
564                 error = 0;
565         }
566         return error;
567 }