Merge tag 'mfd-next-4.19' of git://git.kernel.org/pub/scm/linux/kernel/git/lee/mfd
[linux-2.6-block.git] / fs / xfs / scrub / refcount.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_defer.h"
13 #include "xfs_btree.h"
14 #include "xfs_bit.h"
15 #include "xfs_log_format.h"
16 #include "xfs_trans.h"
17 #include "xfs_sb.h"
18 #include "xfs_alloc.h"
19 #include "xfs_rmap.h"
20 #include "xfs_refcount.h"
21 #include "scrub/xfs_scrub.h"
22 #include "scrub/scrub.h"
23 #include "scrub/common.h"
24 #include "scrub/btree.h"
25 #include "scrub/trace.h"
26
27 /*
28  * Set us up to scrub reference count btrees.
29  */
30 int
31 xchk_setup_ag_refcountbt(
32         struct xfs_scrub        *sc,
33         struct xfs_inode        *ip)
34 {
35         return xchk_setup_ag_btree(sc, ip, false);
36 }
37
38 /* Reference count btree scrubber. */
39
40 /*
41  * Confirming Reference Counts via Reverse Mappings
42  *
43  * We want to count the reverse mappings overlapping a refcount record
44  * (bno, len, refcount), allowing for the possibility that some of the
45  * overlap may come from smaller adjoining reverse mappings, while some
46  * comes from single extents which overlap the range entirely.  The
47  * outer loop is as follows:
48  *
49  * 1. For all reverse mappings overlapping the refcount extent,
50  *    a. If a given rmap completely overlaps, mark it as seen.
51  *    b. Otherwise, record the fragment (in agbno order) for later
52  *       processing.
53  *
54  * Once we've seen all the rmaps, we know that for all blocks in the
55  * refcount record we want to find $refcount owners and we've already
56  * visited $seen extents that overlap all the blocks.  Therefore, we
57  * need to find ($refcount - $seen) owners for every block in the
58  * extent; call that quantity $target_nr.  Proceed as follows:
59  *
60  * 2. Pull the first $target_nr fragments from the list; all of them
61  *    should start at or before the start of the extent.
62  *    Call this subset of fragments the working set.
63  * 3. Until there are no more unprocessed fragments,
64  *    a. Find the shortest fragments in the set and remove them.
65  *    b. Note the block number of the end of these fragments.
66  *    c. Pull the same number of fragments from the list.  All of these
67  *       fragments should start at the block number recorded in the
68  *       previous step.
69  *    d. Put those fragments in the set.
70  * 4. Check that there are $target_nr fragments remaining in the list,
71  *    and that they all end at or beyond the end of the refcount extent.
72  *
73  * If the refcount is correct, all the check conditions in the algorithm
74  * should always hold true.  If not, the refcount is incorrect.
75  */
76 struct xchk_refcnt_frag {
77         struct list_head        list;
78         struct xfs_rmap_irec    rm;
79 };
80
81 struct xchk_refcnt_check {
82         struct xfs_scrub        *sc;
83         struct list_head        fragments;
84
85         /* refcount extent we're examining */
86         xfs_agblock_t           bno;
87         xfs_extlen_t            len;
88         xfs_nlink_t             refcount;
89
90         /* number of owners seen */
91         xfs_nlink_t             seen;
92 };
93
94 /*
95  * Decide if the given rmap is large enough that we can redeem it
96  * towards refcount verification now, or if it's a fragment, in
97  * which case we'll hang onto it in the hopes that we'll later
98  * discover that we've collected exactly the correct number of
99  * fragments as the refcountbt says we should have.
100  */
101 STATIC int
102 xchk_refcountbt_rmap_check(
103         struct xfs_btree_cur            *cur,
104         struct xfs_rmap_irec            *rec,
105         void                            *priv)
106 {
107         struct xchk_refcnt_check        *refchk = priv;
108         struct xchk_refcnt_frag         *frag;
109         xfs_agblock_t                   rm_last;
110         xfs_agblock_t                   rc_last;
111         int                             error = 0;
112
113         if (xchk_should_terminate(refchk->sc, &error))
114                 return error;
115
116         rm_last = rec->rm_startblock + rec->rm_blockcount - 1;
117         rc_last = refchk->bno + refchk->len - 1;
118
119         /* Confirm that a single-owner refc extent is a CoW stage. */
120         if (refchk->refcount == 1 && rec->rm_owner != XFS_RMAP_OWN_COW) {
121                 xchk_btree_xref_set_corrupt(refchk->sc, cur, 0);
122                 return 0;
123         }
124
125         if (rec->rm_startblock <= refchk->bno && rm_last >= rc_last) {
126                 /*
127                  * The rmap overlaps the refcount record, so we can confirm
128                  * one refcount owner seen.
129                  */
130                 refchk->seen++;
131         } else {
132                 /*
133                  * This rmap covers only part of the refcount record, so
134                  * save the fragment for later processing.  If the rmapbt
135                  * is healthy each rmap_irec we see will be in agbno order
136                  * so we don't need insertion sort here.
137                  */
138                 frag = kmem_alloc(sizeof(struct xchk_refcnt_frag),
139                                 KM_MAYFAIL);
140                 if (!frag)
141                         return -ENOMEM;
142                 memcpy(&frag->rm, rec, sizeof(frag->rm));
143                 list_add_tail(&frag->list, &refchk->fragments);
144         }
145
146         return 0;
147 }
148
149 /*
150  * Given a bunch of rmap fragments, iterate through them, keeping
151  * a running tally of the refcount.  If this ever deviates from
152  * what we expect (which is the refcountbt's refcount minus the
153  * number of extents that totally covered the refcountbt extent),
154  * we have a refcountbt error.
155  */
156 STATIC void
157 xchk_refcountbt_process_rmap_fragments(
158         struct xchk_refcnt_check        *refchk)
159 {
160         struct list_head                worklist;
161         struct xchk_refcnt_frag         *frag;
162         struct xchk_refcnt_frag         *n;
163         xfs_agblock_t                   bno;
164         xfs_agblock_t                   rbno;
165         xfs_agblock_t                   next_rbno;
166         xfs_nlink_t                     nr;
167         xfs_nlink_t                     target_nr;
168
169         target_nr = refchk->refcount - refchk->seen;
170         if (target_nr == 0)
171                 return;
172
173         /*
174          * There are (refchk->rc.rc_refcount - refchk->nr refcount)
175          * references we haven't found yet.  Pull that many off the
176          * fragment list and figure out where the smallest rmap ends
177          * (and therefore the next rmap should start).  All the rmaps
178          * we pull off should start at or before the beginning of the
179          * refcount record's range.
180          */
181         INIT_LIST_HEAD(&worklist);
182         rbno = NULLAGBLOCK;
183         nr = 1;
184
185         /* Make sure the fragments actually /are/ in agbno order. */
186         bno = 0;
187         list_for_each_entry(frag, &refchk->fragments, list) {
188                 if (frag->rm.rm_startblock < bno)
189                         goto done;
190                 bno = frag->rm.rm_startblock;
191         }
192
193         /*
194          * Find all the rmaps that start at or before the refc extent,
195          * and put them on the worklist.
196          */
197         list_for_each_entry_safe(frag, n, &refchk->fragments, list) {
198                 if (frag->rm.rm_startblock > refchk->bno)
199                         goto done;
200                 bno = frag->rm.rm_startblock + frag->rm.rm_blockcount;
201                 if (bno < rbno)
202                         rbno = bno;
203                 list_move_tail(&frag->list, &worklist);
204                 if (nr == target_nr)
205                         break;
206                 nr++;
207         }
208
209         /*
210          * We should have found exactly $target_nr rmap fragments starting
211          * at or before the refcount extent.
212          */
213         if (nr != target_nr)
214                 goto done;
215
216         while (!list_empty(&refchk->fragments)) {
217                 /* Discard any fragments ending at rbno from the worklist. */
218                 nr = 0;
219                 next_rbno = NULLAGBLOCK;
220                 list_for_each_entry_safe(frag, n, &worklist, list) {
221                         bno = frag->rm.rm_startblock + frag->rm.rm_blockcount;
222                         if (bno != rbno) {
223                                 if (bno < next_rbno)
224                                         next_rbno = bno;
225                                 continue;
226                         }
227                         list_del(&frag->list);
228                         kmem_free(frag);
229                         nr++;
230                 }
231
232                 /* Try to add nr rmaps starting at rbno to the worklist. */
233                 list_for_each_entry_safe(frag, n, &refchk->fragments, list) {
234                         bno = frag->rm.rm_startblock + frag->rm.rm_blockcount;
235                         if (frag->rm.rm_startblock != rbno)
236                                 goto done;
237                         list_move_tail(&frag->list, &worklist);
238                         if (next_rbno > bno)
239                                 next_rbno = bno;
240                         nr--;
241                         if (nr == 0)
242                                 break;
243                 }
244
245                 /*
246                  * If we get here and nr > 0, this means that we added fewer
247                  * items to the worklist than we discarded because the fragment
248                  * list ran out of items.  Therefore, we cannot maintain the
249                  * required refcount.  Something is wrong, so we're done.
250                  */
251                 if (nr)
252                         goto done;
253
254                 rbno = next_rbno;
255         }
256
257         /*
258          * Make sure the last extent we processed ends at or beyond
259          * the end of the refcount extent.
260          */
261         if (rbno < refchk->bno + refchk->len)
262                 goto done;
263
264         /* Actually record us having seen the remaining refcount. */
265         refchk->seen = refchk->refcount;
266 done:
267         /* Delete fragments and work list. */
268         list_for_each_entry_safe(frag, n, &worklist, list) {
269                 list_del(&frag->list);
270                 kmem_free(frag);
271         }
272         list_for_each_entry_safe(frag, n, &refchk->fragments, list) {
273                 list_del(&frag->list);
274                 kmem_free(frag);
275         }
276 }
277
278 /* Use the rmap entries covering this extent to verify the refcount. */
279 STATIC void
280 xchk_refcountbt_xref_rmap(
281         struct xfs_scrub                *sc,
282         xfs_agblock_t                   bno,
283         xfs_extlen_t                    len,
284         xfs_nlink_t                     refcount)
285 {
286         struct xchk_refcnt_check        refchk = {
287                 .sc = sc,
288                 .bno = bno,
289                 .len = len,
290                 .refcount = refcount,
291                 .seen = 0,
292         };
293         struct xfs_rmap_irec            low;
294         struct xfs_rmap_irec            high;
295         struct xchk_refcnt_frag         *frag;
296         struct xchk_refcnt_frag         *n;
297         int                             error;
298
299         if (!sc->sa.rmap_cur || xchk_skip_xref(sc->sm))
300                 return;
301
302         /* Cross-reference with the rmapbt to confirm the refcount. */
303         memset(&low, 0, sizeof(low));
304         low.rm_startblock = bno;
305         memset(&high, 0xFF, sizeof(high));
306         high.rm_startblock = bno + len - 1;
307
308         INIT_LIST_HEAD(&refchk.fragments);
309         error = xfs_rmap_query_range(sc->sa.rmap_cur, &low, &high,
310                         &xchk_refcountbt_rmap_check, &refchk);
311         if (!xchk_should_check_xref(sc, &error, &sc->sa.rmap_cur))
312                 goto out_free;
313
314         xchk_refcountbt_process_rmap_fragments(&refchk);
315         if (refcount != refchk.seen)
316                 xchk_btree_xref_set_corrupt(sc, sc->sa.rmap_cur, 0);
317
318 out_free:
319         list_for_each_entry_safe(frag, n, &refchk.fragments, list) {
320                 list_del(&frag->list);
321                 kmem_free(frag);
322         }
323 }
324
325 /* Cross-reference with the other btrees. */
326 STATIC void
327 xchk_refcountbt_xref(
328         struct xfs_scrub        *sc,
329         xfs_agblock_t           agbno,
330         xfs_extlen_t            len,
331         xfs_nlink_t             refcount)
332 {
333         if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_CORRUPT)
334                 return;
335
336         xchk_xref_is_used_space(sc, agbno, len);
337         xchk_xref_is_not_inode_chunk(sc, agbno, len);
338         xchk_refcountbt_xref_rmap(sc, agbno, len, refcount);
339 }
340
341 /* Scrub a refcountbt record. */
342 STATIC int
343 xchk_refcountbt_rec(
344         struct xchk_btree       *bs,
345         union xfs_btree_rec     *rec)
346 {
347         struct xfs_mount        *mp = bs->cur->bc_mp;
348         xfs_agblock_t           *cow_blocks = bs->private;
349         xfs_agnumber_t          agno = bs->cur->bc_private.a.agno;
350         xfs_agblock_t           bno;
351         xfs_extlen_t            len;
352         xfs_nlink_t             refcount;
353         bool                    has_cowflag;
354         int                     error = 0;
355
356         bno = be32_to_cpu(rec->refc.rc_startblock);
357         len = be32_to_cpu(rec->refc.rc_blockcount);
358         refcount = be32_to_cpu(rec->refc.rc_refcount);
359
360         /* Only CoW records can have refcount == 1. */
361         has_cowflag = (bno & XFS_REFC_COW_START);
362         if ((refcount == 1 && !has_cowflag) || (refcount != 1 && has_cowflag))
363                 xchk_btree_set_corrupt(bs->sc, bs->cur, 0);
364         if (has_cowflag)
365                 (*cow_blocks) += len;
366
367         /* Check the extent. */
368         bno &= ~XFS_REFC_COW_START;
369         if (bno + len <= bno ||
370             !xfs_verify_agbno(mp, agno, bno) ||
371             !xfs_verify_agbno(mp, agno, bno + len - 1))
372                 xchk_btree_set_corrupt(bs->sc, bs->cur, 0);
373
374         if (refcount == 0)
375                 xchk_btree_set_corrupt(bs->sc, bs->cur, 0);
376
377         xchk_refcountbt_xref(bs->sc, bno, len, refcount);
378
379         return error;
380 }
381
382 /* Make sure we have as many refc blocks as the rmap says. */
383 STATIC void
384 xchk_refcount_xref_rmap(
385         struct xfs_scrub        *sc,
386         struct xfs_owner_info   *oinfo,
387         xfs_filblks_t           cow_blocks)
388 {
389         xfs_extlen_t            refcbt_blocks = 0;
390         xfs_filblks_t           blocks;
391         int                     error;
392
393         if (!sc->sa.rmap_cur || xchk_skip_xref(sc->sm))
394                 return;
395
396         /* Check that we saw as many refcbt blocks as the rmap knows about. */
397         error = xfs_btree_count_blocks(sc->sa.refc_cur, &refcbt_blocks);
398         if (!xchk_btree_process_error(sc, sc->sa.refc_cur, 0, &error))
399                 return;
400         error = xchk_count_rmap_ownedby_ag(sc, sc->sa.rmap_cur, oinfo,
401                         &blocks);
402         if (!xchk_should_check_xref(sc, &error, &sc->sa.rmap_cur))
403                 return;
404         if (blocks != refcbt_blocks)
405                 xchk_btree_xref_set_corrupt(sc, sc->sa.rmap_cur, 0);
406
407         /* Check that we saw as many cow blocks as the rmap knows about. */
408         xfs_rmap_ag_owner(oinfo, XFS_RMAP_OWN_COW);
409         error = xchk_count_rmap_ownedby_ag(sc, sc->sa.rmap_cur, oinfo,
410                         &blocks);
411         if (!xchk_should_check_xref(sc, &error, &sc->sa.rmap_cur))
412                 return;
413         if (blocks != cow_blocks)
414                 xchk_btree_xref_set_corrupt(sc, sc->sa.rmap_cur, 0);
415 }
416
417 /* Scrub the refcount btree for some AG. */
418 int
419 xchk_refcountbt(
420         struct xfs_scrub        *sc)
421 {
422         struct xfs_owner_info   oinfo;
423         xfs_agblock_t           cow_blocks = 0;
424         int                     error;
425
426         xfs_rmap_ag_owner(&oinfo, XFS_RMAP_OWN_REFC);
427         error = xchk_btree(sc, sc->sa.refc_cur, xchk_refcountbt_rec,
428                         &oinfo, &cow_blocks);
429         if (error)
430                 return error;
431
432         xchk_refcount_xref_rmap(sc, &oinfo, cow_blocks);
433
434         return 0;
435 }
436
437 /* xref check that a cow staging extent is marked in the refcountbt. */
438 void
439 xchk_xref_is_cow_staging(
440         struct xfs_scrub                *sc,
441         xfs_agblock_t                   agbno,
442         xfs_extlen_t                    len)
443 {
444         struct xfs_refcount_irec        rc;
445         bool                            has_cowflag;
446         int                             has_refcount;
447         int                             error;
448
449         if (!sc->sa.refc_cur || xchk_skip_xref(sc->sm))
450                 return;
451
452         /* Find the CoW staging extent. */
453         error = xfs_refcount_lookup_le(sc->sa.refc_cur,
454                         agbno + XFS_REFC_COW_START, &has_refcount);
455         if (!xchk_should_check_xref(sc, &error, &sc->sa.refc_cur))
456                 return;
457         if (!has_refcount) {
458                 xchk_btree_xref_set_corrupt(sc, sc->sa.refc_cur, 0);
459                 return;
460         }
461
462         error = xfs_refcount_get_rec(sc->sa.refc_cur, &rc, &has_refcount);
463         if (!xchk_should_check_xref(sc, &error, &sc->sa.refc_cur))
464                 return;
465         if (!has_refcount) {
466                 xchk_btree_xref_set_corrupt(sc, sc->sa.refc_cur, 0);
467                 return;
468         }
469
470         /* CoW flag must be set, refcount must be 1. */
471         has_cowflag = (rc.rc_startblock & XFS_REFC_COW_START);
472         if (!has_cowflag || rc.rc_refcount != 1)
473                 xchk_btree_xref_set_corrupt(sc, sc->sa.refc_cur, 0);
474
475         /* Must be at least as long as what was passed in */
476         if (rc.rc_blockcount < len)
477                 xchk_btree_xref_set_corrupt(sc, sc->sa.refc_cur, 0);
478 }
479
480 /*
481  * xref check that the extent is not shared.  Only file data blocks
482  * can have multiple owners.
483  */
484 void
485 xchk_xref_is_not_shared(
486         struct xfs_scrub        *sc,
487         xfs_agblock_t           agbno,
488         xfs_extlen_t            len)
489 {
490         bool                    shared;
491         int                     error;
492
493         if (!sc->sa.refc_cur || xchk_skip_xref(sc->sm))
494                 return;
495
496         error = xfs_refcount_has_record(sc->sa.refc_cur, agbno, len, &shared);
497         if (!xchk_should_check_xref(sc, &error, &sc->sa.refc_cur))
498                 return;
499         if (shared)
500                 xchk_btree_xref_set_corrupt(sc, sc->sa.refc_cur, 0);
501 }