xref: /linux/fs/xfs/scrub/repair.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2018 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_inode.h"
19 #include "xfs_icache.h"
20 #include "xfs_alloc.h"
21 #include "xfs_alloc_btree.h"
22 #include "xfs_ialloc.h"
23 #include "xfs_ialloc_btree.h"
24 #include "xfs_rmap.h"
25 #include "xfs_rmap_btree.h"
26 #include "xfs_refcount.h"
27 #include "xfs_refcount_btree.h"
28 #include "xfs_extent_busy.h"
29 #include "xfs_ag_resv.h"
30 #include "xfs_trans_space.h"
31 #include "xfs_quota.h"
32 #include "xfs_attr.h"
33 #include "xfs_reflink.h"
34 #include "scrub/xfs_scrub.h"
35 #include "scrub/scrub.h"
36 #include "scrub/common.h"
37 #include "scrub/trace.h"
38 #include "scrub/repair.h"
39 #include "scrub/bitmap.h"
40 
41 /*
42  * Attempt to repair some metadata, if the metadata is corrupt and userspace
43  * told us to fix it.  This function returns -EAGAIN to mean "re-run scrub",
44  * and will set *fixed to true if it thinks it repaired anything.
45  */
46 int
47 xrep_attempt(
48 	struct xfs_inode	*ip,
49 	struct xfs_scrub	*sc)
50 {
51 	int			error = 0;
52 
53 	trace_xrep_attempt(ip, sc->sm, error);
54 
55 	xchk_ag_btcur_free(&sc->sa);
56 
57 	/* Repair whatever's broken. */
58 	ASSERT(sc->ops->repair);
59 	error = sc->ops->repair(sc);
60 	trace_xrep_done(ip, sc->sm, error);
61 	switch (error) {
62 	case 0:
63 		/*
64 		 * Repair succeeded.  Commit the fixes and perform a second
65 		 * scrub so that we can tell userspace if we fixed the problem.
66 		 */
67 		sc->sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT;
68 		sc->flags |= XREP_ALREADY_FIXED;
69 		return -EAGAIN;
70 	case -EDEADLOCK:
71 	case -EAGAIN:
72 		/* Tell the caller to try again having grabbed all the locks. */
73 		if (!(sc->flags & XCHK_TRY_HARDER)) {
74 			sc->flags |= XCHK_TRY_HARDER;
75 			return -EAGAIN;
76 		}
77 		/*
78 		 * We tried harder but still couldn't grab all the resources
79 		 * we needed to fix it.  The corruption has not been fixed,
80 		 * so report back to userspace.
81 		 */
82 		return -EFSCORRUPTED;
83 	default:
84 		return error;
85 	}
86 }
87 
88 /*
89  * Complain about unfixable problems in the filesystem.  We don't log
90  * corruptions when IFLAG_REPAIR wasn't set on the assumption that the driver
91  * program is xfs_scrub, which will call back with IFLAG_REPAIR set if the
92  * administrator isn't running xfs_scrub in no-repairs mode.
93  *
94  * Use this helper function because _ratelimited silently declares a static
95  * structure to track rate limiting information.
96  */
97 void
98 xrep_failure(
99 	struct xfs_mount	*mp)
100 {
101 	xfs_alert_ratelimited(mp,
102 "Corruption not fixed during online repair.  Unmount and run xfs_repair.");
103 }
104 
105 /*
106  * Repair probe -- userspace uses this to probe if we're willing to repair a
107  * given mountpoint.
108  */
109 int
110 xrep_probe(
111 	struct xfs_scrub	*sc)
112 {
113 	int			error = 0;
114 
115 	if (xchk_should_terminate(sc, &error))
116 		return error;
117 
118 	return 0;
119 }
120 
121 /*
122  * Roll a transaction, keeping the AG headers locked and reinitializing
123  * the btree cursors.
124  */
125 int
126 xrep_roll_ag_trans(
127 	struct xfs_scrub	*sc)
128 {
129 	int			error;
130 
131 	/* Keep the AG header buffers locked so we can keep going. */
132 	if (sc->sa.agi_bp)
133 		xfs_trans_bhold(sc->tp, sc->sa.agi_bp);
134 	if (sc->sa.agf_bp)
135 		xfs_trans_bhold(sc->tp, sc->sa.agf_bp);
136 	if (sc->sa.agfl_bp)
137 		xfs_trans_bhold(sc->tp, sc->sa.agfl_bp);
138 
139 	/*
140 	 * Roll the transaction.  We still own the buffer and the buffer lock
141 	 * regardless of whether or not the roll succeeds.  If the roll fails,
142 	 * the buffers will be released during teardown on our way out of the
143 	 * kernel.  If it succeeds, we join them to the new transaction and
144 	 * move on.
145 	 */
146 	error = xfs_trans_roll(&sc->tp);
147 	if (error)
148 		return error;
149 
150 	/* Join AG headers to the new transaction. */
151 	if (sc->sa.agi_bp)
152 		xfs_trans_bjoin(sc->tp, sc->sa.agi_bp);
153 	if (sc->sa.agf_bp)
154 		xfs_trans_bjoin(sc->tp, sc->sa.agf_bp);
155 	if (sc->sa.agfl_bp)
156 		xfs_trans_bjoin(sc->tp, sc->sa.agfl_bp);
157 
158 	return 0;
159 }
160 
161 /*
162  * Does the given AG have enough space to rebuild a btree?  Neither AG
163  * reservation can be critical, and we must have enough space (factoring
164  * in AG reservations) to construct a whole btree.
165  */
166 bool
167 xrep_ag_has_space(
168 	struct xfs_perag	*pag,
169 	xfs_extlen_t		nr_blocks,
170 	enum xfs_ag_resv_type	type)
171 {
172 	return  !xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) &&
173 		!xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA) &&
174 		pag->pagf_freeblks > xfs_ag_resv_needed(pag, type) + nr_blocks;
175 }
176 
177 /*
178  * Figure out how many blocks to reserve for an AG repair.  We calculate the
179  * worst case estimate for the number of blocks we'd need to rebuild one of
180  * any type of per-AG btree.
181  */
182 xfs_extlen_t
183 xrep_calc_ag_resblks(
184 	struct xfs_scrub		*sc)
185 {
186 	struct xfs_mount		*mp = sc->mp;
187 	struct xfs_scrub_metadata	*sm = sc->sm;
188 	struct xfs_perag		*pag;
189 	struct xfs_buf			*bp;
190 	xfs_agino_t			icount = NULLAGINO;
191 	xfs_extlen_t			aglen = NULLAGBLOCK;
192 	xfs_extlen_t			usedlen;
193 	xfs_extlen_t			freelen;
194 	xfs_extlen_t			bnobt_sz;
195 	xfs_extlen_t			inobt_sz;
196 	xfs_extlen_t			rmapbt_sz;
197 	xfs_extlen_t			refcbt_sz;
198 	int				error;
199 
200 	if (!(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR))
201 		return 0;
202 
203 	pag = xfs_perag_get(mp, sm->sm_agno);
204 	if (pag->pagi_init) {
205 		/* Use in-core icount if possible. */
206 		icount = pag->pagi_count;
207 	} else {
208 		/* Try to get the actual counters from disk. */
209 		error = xfs_ialloc_read_agi(mp, NULL, sm->sm_agno, &bp);
210 		if (!error) {
211 			icount = pag->pagi_count;
212 			xfs_buf_relse(bp);
213 		}
214 	}
215 
216 	/* Now grab the block counters from the AGF. */
217 	error = xfs_alloc_read_agf(mp, NULL, sm->sm_agno, 0, &bp);
218 	if (!error) {
219 		aglen = be32_to_cpu(XFS_BUF_TO_AGF(bp)->agf_length);
220 		freelen = be32_to_cpu(XFS_BUF_TO_AGF(bp)->agf_freeblks);
221 		usedlen = aglen - freelen;
222 		xfs_buf_relse(bp);
223 	}
224 	xfs_perag_put(pag);
225 
226 	/* If the icount is impossible, make some worst-case assumptions. */
227 	if (icount == NULLAGINO ||
228 	    !xfs_verify_agino(mp, sm->sm_agno, icount)) {
229 		xfs_agino_t	first, last;
230 
231 		xfs_agino_range(mp, sm->sm_agno, &first, &last);
232 		icount = last - first + 1;
233 	}
234 
235 	/* If the block counts are impossible, make worst-case assumptions. */
236 	if (aglen == NULLAGBLOCK ||
237 	    aglen != xfs_ag_block_count(mp, sm->sm_agno) ||
238 	    freelen >= aglen) {
239 		aglen = xfs_ag_block_count(mp, sm->sm_agno);
240 		freelen = aglen;
241 		usedlen = aglen;
242 	}
243 
244 	trace_xrep_calc_ag_resblks(mp, sm->sm_agno, icount, aglen,
245 			freelen, usedlen);
246 
247 	/*
248 	 * Figure out how many blocks we'd need worst case to rebuild
249 	 * each type of btree.  Note that we can only rebuild the
250 	 * bnobt/cntbt or inobt/finobt as pairs.
251 	 */
252 	bnobt_sz = 2 * xfs_allocbt_calc_size(mp, freelen);
253 	if (xfs_sb_version_hassparseinodes(&mp->m_sb))
254 		inobt_sz = xfs_iallocbt_calc_size(mp, icount /
255 				XFS_INODES_PER_HOLEMASK_BIT);
256 	else
257 		inobt_sz = xfs_iallocbt_calc_size(mp, icount /
258 				XFS_INODES_PER_CHUNK);
259 	if (xfs_sb_version_hasfinobt(&mp->m_sb))
260 		inobt_sz *= 2;
261 	if (xfs_sb_version_hasreflink(&mp->m_sb))
262 		refcbt_sz = xfs_refcountbt_calc_size(mp, usedlen);
263 	else
264 		refcbt_sz = 0;
265 	if (xfs_sb_version_hasrmapbt(&mp->m_sb)) {
266 		/*
267 		 * Guess how many blocks we need to rebuild the rmapbt.
268 		 * For non-reflink filesystems we can't have more records than
269 		 * used blocks.  However, with reflink it's possible to have
270 		 * more than one rmap record per AG block.  We don't know how
271 		 * many rmaps there could be in the AG, so we start off with
272 		 * what we hope is an generous over-estimation.
273 		 */
274 		if (xfs_sb_version_hasreflink(&mp->m_sb))
275 			rmapbt_sz = xfs_rmapbt_calc_size(mp,
276 					(unsigned long long)aglen * 2);
277 		else
278 			rmapbt_sz = xfs_rmapbt_calc_size(mp, usedlen);
279 	} else {
280 		rmapbt_sz = 0;
281 	}
282 
283 	trace_xrep_calc_ag_resblks_btsize(mp, sm->sm_agno, bnobt_sz,
284 			inobt_sz, rmapbt_sz, refcbt_sz);
285 
286 	return max(max(bnobt_sz, inobt_sz), max(rmapbt_sz, refcbt_sz));
287 }
288 
289 /* Allocate a block in an AG. */
290 int
291 xrep_alloc_ag_block(
292 	struct xfs_scrub		*sc,
293 	const struct xfs_owner_info	*oinfo,
294 	xfs_fsblock_t			*fsbno,
295 	enum xfs_ag_resv_type		resv)
296 {
297 	struct xfs_alloc_arg		args = {0};
298 	xfs_agblock_t			bno;
299 	int				error;
300 
301 	switch (resv) {
302 	case XFS_AG_RESV_AGFL:
303 	case XFS_AG_RESV_RMAPBT:
304 		error = xfs_alloc_get_freelist(sc->tp, sc->sa.agf_bp, &bno, 1);
305 		if (error)
306 			return error;
307 		if (bno == NULLAGBLOCK)
308 			return -ENOSPC;
309 		xfs_extent_busy_reuse(sc->mp, sc->sa.agno, bno,
310 				1, false);
311 		*fsbno = XFS_AGB_TO_FSB(sc->mp, sc->sa.agno, bno);
312 		if (resv == XFS_AG_RESV_RMAPBT)
313 			xfs_ag_resv_rmapbt_alloc(sc->mp, sc->sa.agno);
314 		return 0;
315 	default:
316 		break;
317 	}
318 
319 	args.tp = sc->tp;
320 	args.mp = sc->mp;
321 	args.oinfo = *oinfo;
322 	args.fsbno = XFS_AGB_TO_FSB(args.mp, sc->sa.agno, 0);
323 	args.minlen = 1;
324 	args.maxlen = 1;
325 	args.prod = 1;
326 	args.type = XFS_ALLOCTYPE_THIS_AG;
327 	args.resv = resv;
328 
329 	error = xfs_alloc_vextent(&args);
330 	if (error)
331 		return error;
332 	if (args.fsbno == NULLFSBLOCK)
333 		return -ENOSPC;
334 	ASSERT(args.len == 1);
335 	*fsbno = args.fsbno;
336 
337 	return 0;
338 }
339 
340 /* Initialize a new AG btree root block with zero entries. */
341 int
342 xrep_init_btblock(
343 	struct xfs_scrub		*sc,
344 	xfs_fsblock_t			fsb,
345 	struct xfs_buf			**bpp,
346 	xfs_btnum_t			btnum,
347 	const struct xfs_buf_ops	*ops)
348 {
349 	struct xfs_trans		*tp = sc->tp;
350 	struct xfs_mount		*mp = sc->mp;
351 	struct xfs_buf			*bp;
352 
353 	trace_xrep_init_btblock(mp, XFS_FSB_TO_AGNO(mp, fsb),
354 			XFS_FSB_TO_AGBNO(mp, fsb), btnum);
355 
356 	ASSERT(XFS_FSB_TO_AGNO(mp, fsb) == sc->sa.agno);
357 	bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, XFS_FSB_TO_DADDR(mp, fsb),
358 			XFS_FSB_TO_BB(mp, 1), 0);
359 	xfs_buf_zero(bp, 0, BBTOB(bp->b_length));
360 	xfs_btree_init_block(mp, bp, btnum, 0, 0, sc->sa.agno, 0);
361 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_BTREE_BUF);
362 	xfs_trans_log_buf(tp, bp, 0, bp->b_length);
363 	bp->b_ops = ops;
364 	*bpp = bp;
365 
366 	return 0;
367 }
368 
369 /*
370  * Reconstructing per-AG Btrees
371  *
372  * When a space btree is corrupt, we don't bother trying to fix it.  Instead,
373  * we scan secondary space metadata to derive the records that should be in
374  * the damaged btree, initialize a fresh btree root, and insert the records.
375  * Note that for rebuilding the rmapbt we scan all the primary data to
376  * generate the new records.
377  *
378  * However, that leaves the matter of removing all the metadata describing the
379  * old broken structure.  For primary metadata we use the rmap data to collect
380  * every extent with a matching rmap owner (bitmap); we then iterate all other
381  * metadata structures with the same rmap owner to collect the extents that
382  * cannot be removed (sublist).  We then subtract sublist from bitmap to
383  * derive the blocks that were used by the old btree.  These blocks can be
384  * reaped.
385  *
386  * For rmapbt reconstructions we must use different tactics for extent
387  * collection.  First we iterate all primary metadata (this excludes the old
388  * rmapbt, obviously) to generate new rmap records.  The gaps in the rmap
389  * records are collected as bitmap.  The bnobt records are collected as
390  * sublist.  As with the other btrees we subtract sublist from bitmap, and the
391  * result (since the rmapbt lives in the free space) are the blocks from the
392  * old rmapbt.
393  *
394  * Disposal of Blocks from Old per-AG Btrees
395  *
396  * Now that we've constructed a new btree to replace the damaged one, we want
397  * to dispose of the blocks that (we think) the old btree was using.
398  * Previously, we used the rmapbt to collect the extents (bitmap) with the
399  * rmap owner corresponding to the tree we rebuilt, collected extents for any
400  * blocks with the same rmap owner that are owned by another data structure
401  * (sublist), and subtracted sublist from bitmap.  In theory the extents
402  * remaining in bitmap are the old btree's blocks.
403  *
404  * Unfortunately, it's possible that the btree was crosslinked with other
405  * blocks on disk.  The rmap data can tell us if there are multiple owners, so
406  * if the rmapbt says there is an owner of this block other than @oinfo, then
407  * the block is crosslinked.  Remove the reverse mapping and continue.
408  *
409  * If there is one rmap record, we can free the block, which removes the
410  * reverse mapping but doesn't add the block to the free space.  Our repair
411  * strategy is to hope the other metadata objects crosslinked on this block
412  * will be rebuilt (atop different blocks), thereby removing all the cross
413  * links.
414  *
415  * If there are no rmap records at all, we also free the block.  If the btree
416  * being rebuilt lives in the free space (bnobt/cntbt/rmapbt) then there isn't
417  * supposed to be a rmap record and everything is ok.  For other btrees there
418  * had to have been an rmap entry for the block to have ended up on @bitmap,
419  * so if it's gone now there's something wrong and the fs will shut down.
420  *
421  * Note: If there are multiple rmap records with only the same rmap owner as
422  * the btree we're trying to rebuild and the block is indeed owned by another
423  * data structure with the same rmap owner, then the block will be in sublist
424  * and therefore doesn't need disposal.  If there are multiple rmap records
425  * with only the same rmap owner but the block is not owned by something with
426  * the same rmap owner, the block will be freed.
427  *
428  * The caller is responsible for locking the AG headers for the entire rebuild
429  * operation so that nothing else can sneak in and change the AG state while
430  * we're not looking.  We also assume that the caller already invalidated any
431  * buffers associated with @bitmap.
432  */
433 
434 /*
435  * Invalidate buffers for per-AG btree blocks we're dumping.  This function
436  * is not intended for use with file data repairs; we have bunmapi for that.
437  */
438 int
439 xrep_invalidate_blocks(
440 	struct xfs_scrub	*sc,
441 	struct xfs_bitmap	*bitmap)
442 {
443 	struct xfs_bitmap_range	*bmr;
444 	struct xfs_bitmap_range	*n;
445 	struct xfs_buf		*bp;
446 	xfs_fsblock_t		fsbno;
447 
448 	/*
449 	 * For each block in each extent, see if there's an incore buffer for
450 	 * exactly that block; if so, invalidate it.  The buffer cache only
451 	 * lets us look for one buffer at a time, so we have to look one block
452 	 * at a time.  Avoid invalidating AG headers and post-EOFS blocks
453 	 * because we never own those; and if we can't TRYLOCK the buffer we
454 	 * assume it's owned by someone else.
455 	 */
456 	for_each_xfs_bitmap_block(fsbno, bmr, n, bitmap) {
457 		/* Skip AG headers and post-EOFS blocks */
458 		if (!xfs_verify_fsbno(sc->mp, fsbno))
459 			continue;
460 		bp = xfs_buf_incore(sc->mp->m_ddev_targp,
461 				XFS_FSB_TO_DADDR(sc->mp, fsbno),
462 				XFS_FSB_TO_BB(sc->mp, 1), XBF_TRYLOCK);
463 		if (bp) {
464 			xfs_trans_bjoin(sc->tp, bp);
465 			xfs_trans_binval(sc->tp, bp);
466 		}
467 	}
468 
469 	return 0;
470 }
471 
472 /* Ensure the freelist is the correct size. */
473 int
474 xrep_fix_freelist(
475 	struct xfs_scrub	*sc,
476 	bool			can_shrink)
477 {
478 	struct xfs_alloc_arg	args = {0};
479 
480 	args.mp = sc->mp;
481 	args.tp = sc->tp;
482 	args.agno = sc->sa.agno;
483 	args.alignment = 1;
484 	args.pag = sc->sa.pag;
485 
486 	return xfs_alloc_fix_freelist(&args,
487 			can_shrink ? 0 : XFS_ALLOC_FLAG_NOSHRINK);
488 }
489 
490 /*
491  * Put a block back on the AGFL.
492  */
493 STATIC int
494 xrep_put_freelist(
495 	struct xfs_scrub	*sc,
496 	xfs_agblock_t		agbno)
497 {
498 	int			error;
499 
500 	/* Make sure there's space on the freelist. */
501 	error = xrep_fix_freelist(sc, true);
502 	if (error)
503 		return error;
504 
505 	/*
506 	 * Since we're "freeing" a lost block onto the AGFL, we have to
507 	 * create an rmap for the block prior to merging it or else other
508 	 * parts will break.
509 	 */
510 	error = xfs_rmap_alloc(sc->tp, sc->sa.agf_bp, sc->sa.agno, agbno, 1,
511 			&XFS_RMAP_OINFO_AG);
512 	if (error)
513 		return error;
514 
515 	/* Put the block on the AGFL. */
516 	error = xfs_alloc_put_freelist(sc->tp, sc->sa.agf_bp, sc->sa.agfl_bp,
517 			agbno, 0);
518 	if (error)
519 		return error;
520 	xfs_extent_busy_insert(sc->tp, sc->sa.agno, agbno, 1,
521 			XFS_EXTENT_BUSY_SKIP_DISCARD);
522 
523 	return 0;
524 }
525 
526 /* Dispose of a single block. */
527 STATIC int
528 xrep_reap_block(
529 	struct xfs_scrub		*sc,
530 	xfs_fsblock_t			fsbno,
531 	const struct xfs_owner_info	*oinfo,
532 	enum xfs_ag_resv_type		resv)
533 {
534 	struct xfs_btree_cur		*cur;
535 	struct xfs_buf			*agf_bp = NULL;
536 	xfs_agnumber_t			agno;
537 	xfs_agblock_t			agbno;
538 	bool				has_other_rmap;
539 	int				error;
540 
541 	agno = XFS_FSB_TO_AGNO(sc->mp, fsbno);
542 	agbno = XFS_FSB_TO_AGBNO(sc->mp, fsbno);
543 
544 	/*
545 	 * If we are repairing per-inode metadata, we need to read in the AGF
546 	 * buffer.  Otherwise, we're repairing a per-AG structure, so reuse
547 	 * the AGF buffer that the setup functions already grabbed.
548 	 */
549 	if (sc->ip) {
550 		error = xfs_alloc_read_agf(sc->mp, sc->tp, agno, 0, &agf_bp);
551 		if (error)
552 			return error;
553 		if (!agf_bp)
554 			return -ENOMEM;
555 	} else {
556 		agf_bp = sc->sa.agf_bp;
557 	}
558 	cur = xfs_rmapbt_init_cursor(sc->mp, sc->tp, agf_bp, agno);
559 
560 	/* Can we find any other rmappings? */
561 	error = xfs_rmap_has_other_keys(cur, agbno, 1, oinfo, &has_other_rmap);
562 	xfs_btree_del_cursor(cur, error);
563 	if (error)
564 		goto out_free;
565 
566 	/*
567 	 * If there are other rmappings, this block is cross linked and must
568 	 * not be freed.  Remove the reverse mapping and move on.  Otherwise,
569 	 * we were the only owner of the block, so free the extent, which will
570 	 * also remove the rmap.
571 	 *
572 	 * XXX: XFS doesn't support detecting the case where a single block
573 	 * metadata structure is crosslinked with a multi-block structure
574 	 * because the buffer cache doesn't detect aliasing problems, so we
575 	 * can't fix 100% of crosslinking problems (yet).  The verifiers will
576 	 * blow on writeout, the filesystem will shut down, and the admin gets
577 	 * to run xfs_repair.
578 	 */
579 	if (has_other_rmap)
580 		error = xfs_rmap_free(sc->tp, agf_bp, agno, agbno, 1, oinfo);
581 	else if (resv == XFS_AG_RESV_AGFL)
582 		error = xrep_put_freelist(sc, agbno);
583 	else
584 		error = xfs_free_extent(sc->tp, fsbno, 1, oinfo, resv);
585 	if (agf_bp != sc->sa.agf_bp)
586 		xfs_trans_brelse(sc->tp, agf_bp);
587 	if (error)
588 		return error;
589 
590 	if (sc->ip)
591 		return xfs_trans_roll_inode(&sc->tp, sc->ip);
592 	return xrep_roll_ag_trans(sc);
593 
594 out_free:
595 	if (agf_bp != sc->sa.agf_bp)
596 		xfs_trans_brelse(sc->tp, agf_bp);
597 	return error;
598 }
599 
600 /* Dispose of every block of every extent in the bitmap. */
601 int
602 xrep_reap_extents(
603 	struct xfs_scrub		*sc,
604 	struct xfs_bitmap		*bitmap,
605 	const struct xfs_owner_info	*oinfo,
606 	enum xfs_ag_resv_type		type)
607 {
608 	struct xfs_bitmap_range		*bmr;
609 	struct xfs_bitmap_range		*n;
610 	xfs_fsblock_t			fsbno;
611 	int				error = 0;
612 
613 	ASSERT(xfs_sb_version_hasrmapbt(&sc->mp->m_sb));
614 
615 	for_each_xfs_bitmap_block(fsbno, bmr, n, bitmap) {
616 		ASSERT(sc->ip != NULL ||
617 		       XFS_FSB_TO_AGNO(sc->mp, fsbno) == sc->sa.agno);
618 		trace_xrep_dispose_btree_extent(sc->mp,
619 				XFS_FSB_TO_AGNO(sc->mp, fsbno),
620 				XFS_FSB_TO_AGBNO(sc->mp, fsbno), 1);
621 
622 		error = xrep_reap_block(sc, fsbno, oinfo, type);
623 		if (error)
624 			goto out;
625 	}
626 
627 out:
628 	xfs_bitmap_destroy(bitmap);
629 	return error;
630 }
631 
632 /*
633  * Finding per-AG Btree Roots for AGF/AGI Reconstruction
634  *
635  * If the AGF or AGI become slightly corrupted, it may be necessary to rebuild
636  * the AG headers by using the rmap data to rummage through the AG looking for
637  * btree roots.  This is not guaranteed to work if the AG is heavily damaged
638  * or the rmap data are corrupt.
639  *
640  * Callers of xrep_find_ag_btree_roots must lock the AGF and AGFL
641  * buffers if the AGF is being rebuilt; or the AGF and AGI buffers if the
642  * AGI is being rebuilt.  It must maintain these locks until it's safe for
643  * other threads to change the btrees' shapes.  The caller provides
644  * information about the btrees to look for by passing in an array of
645  * xrep_find_ag_btree with the (rmap owner, buf_ops, magic) fields set.
646  * The (root, height) fields will be set on return if anything is found.  The
647  * last element of the array should have a NULL buf_ops to mark the end of the
648  * array.
649  *
650  * For every rmapbt record matching any of the rmap owners in btree_info,
651  * read each block referenced by the rmap record.  If the block is a btree
652  * block from this filesystem matching any of the magic numbers and has a
653  * level higher than what we've already seen, remember the block and the
654  * height of the tree required to have such a block.  When the call completes,
655  * we return the highest block we've found for each btree description; those
656  * should be the roots.
657  */
658 
659 struct xrep_findroot {
660 	struct xfs_scrub		*sc;
661 	struct xfs_buf			*agfl_bp;
662 	struct xfs_agf			*agf;
663 	struct xrep_find_ag_btree	*btree_info;
664 };
665 
666 /* See if our block is in the AGFL. */
667 STATIC int
668 xrep_findroot_agfl_walk(
669 	struct xfs_mount	*mp,
670 	xfs_agblock_t		bno,
671 	void			*priv)
672 {
673 	xfs_agblock_t		*agbno = priv;
674 
675 	return (*agbno == bno) ? XFS_BTREE_QUERY_RANGE_ABORT : 0;
676 }
677 
678 /* Does this block match the btree information passed in? */
679 STATIC int
680 xrep_findroot_block(
681 	struct xrep_findroot		*ri,
682 	struct xrep_find_ag_btree	*fab,
683 	uint64_t			owner,
684 	xfs_agblock_t			agbno,
685 	bool				*done_with_block)
686 {
687 	struct xfs_mount		*mp = ri->sc->mp;
688 	struct xfs_buf			*bp;
689 	struct xfs_btree_block		*btblock;
690 	xfs_daddr_t			daddr;
691 	int				block_level;
692 	int				error = 0;
693 
694 	daddr = XFS_AGB_TO_DADDR(mp, ri->sc->sa.agno, agbno);
695 
696 	/*
697 	 * Blocks in the AGFL have stale contents that might just happen to
698 	 * have a matching magic and uuid.  We don't want to pull these blocks
699 	 * in as part of a tree root, so we have to filter out the AGFL stuff
700 	 * here.  If the AGFL looks insane we'll just refuse to repair.
701 	 */
702 	if (owner == XFS_RMAP_OWN_AG) {
703 		error = xfs_agfl_walk(mp, ri->agf, ri->agfl_bp,
704 				xrep_findroot_agfl_walk, &agbno);
705 		if (error == XFS_BTREE_QUERY_RANGE_ABORT)
706 			return 0;
707 		if (error)
708 			return error;
709 	}
710 
711 	/*
712 	 * Read the buffer into memory so that we can see if it's a match for
713 	 * our btree type.  We have no clue if it is beforehand, and we want to
714 	 * avoid xfs_trans_read_buf's behavior of dumping the DONE state (which
715 	 * will cause needless disk reads in subsequent calls to this function)
716 	 * and logging metadata verifier failures.
717 	 *
718 	 * Therefore, pass in NULL buffer ops.  If the buffer was already in
719 	 * memory from some other caller it will already have b_ops assigned.
720 	 * If it was in memory from a previous unsuccessful findroot_block
721 	 * call, the buffer won't have b_ops but it should be clean and ready
722 	 * for us to try to verify if the read call succeeds.  The same applies
723 	 * if the buffer wasn't in memory at all.
724 	 *
725 	 * Note: If we never match a btree type with this buffer, it will be
726 	 * left in memory with NULL b_ops.  This shouldn't be a problem unless
727 	 * the buffer gets written.
728 	 */
729 	error = xfs_trans_read_buf(mp, ri->sc->tp, mp->m_ddev_targp, daddr,
730 			mp->m_bsize, 0, &bp, NULL);
731 	if (error)
732 		return error;
733 
734 	/* Ensure the block magic matches the btree type we're looking for. */
735 	btblock = XFS_BUF_TO_BLOCK(bp);
736 	ASSERT(fab->buf_ops->magic[1] != 0);
737 	if (btblock->bb_magic != fab->buf_ops->magic[1])
738 		goto out;
739 
740 	/*
741 	 * If the buffer already has ops applied and they're not the ones for
742 	 * this btree type, we know this block doesn't match the btree and we
743 	 * can bail out.
744 	 *
745 	 * If the buffer ops match ours, someone else has already validated
746 	 * the block for us, so we can move on to checking if this is a root
747 	 * block candidate.
748 	 *
749 	 * If the buffer does not have ops, nobody has successfully validated
750 	 * the contents and the buffer cannot be dirty.  If the magic, uuid,
751 	 * and structure match this btree type then we'll move on to checking
752 	 * if it's a root block candidate.  If there is no match, bail out.
753 	 */
754 	if (bp->b_ops) {
755 		if (bp->b_ops != fab->buf_ops)
756 			goto out;
757 	} else {
758 		ASSERT(!xfs_trans_buf_is_dirty(bp));
759 		if (!uuid_equal(&btblock->bb_u.s.bb_uuid,
760 				&mp->m_sb.sb_meta_uuid))
761 			goto out;
762 		/*
763 		 * Read verifiers can reference b_ops, so we set the pointer
764 		 * here.  If the verifier fails we'll reset the buffer state
765 		 * to what it was before we touched the buffer.
766 		 */
767 		bp->b_ops = fab->buf_ops;
768 		fab->buf_ops->verify_read(bp);
769 		if (bp->b_error) {
770 			bp->b_ops = NULL;
771 			bp->b_error = 0;
772 			goto out;
773 		}
774 
775 		/*
776 		 * Some read verifiers will (re)set b_ops, so we must be
777 		 * careful not to change b_ops after running the verifier.
778 		 */
779 	}
780 
781 	/*
782 	 * This block passes the magic/uuid and verifier tests for this btree
783 	 * type.  We don't need the caller to try the other tree types.
784 	 */
785 	*done_with_block = true;
786 
787 	/*
788 	 * Compare this btree block's level to the height of the current
789 	 * candidate root block.
790 	 *
791 	 * If the level matches the root we found previously, throw away both
792 	 * blocks because there can't be two candidate roots.
793 	 *
794 	 * If level is lower in the tree than the root we found previously,
795 	 * ignore this block.
796 	 */
797 	block_level = xfs_btree_get_level(btblock);
798 	if (block_level + 1 == fab->height) {
799 		fab->root = NULLAGBLOCK;
800 		goto out;
801 	} else if (block_level < fab->height) {
802 		goto out;
803 	}
804 
805 	/*
806 	 * This is the highest block in the tree that we've found so far.
807 	 * Update the btree height to reflect what we've learned from this
808 	 * block.
809 	 */
810 	fab->height = block_level + 1;
811 
812 	/*
813 	 * If this block doesn't have sibling pointers, then it's the new root
814 	 * block candidate.  Otherwise, the root will be found farther up the
815 	 * tree.
816 	 */
817 	if (btblock->bb_u.s.bb_leftsib == cpu_to_be32(NULLAGBLOCK) &&
818 	    btblock->bb_u.s.bb_rightsib == cpu_to_be32(NULLAGBLOCK))
819 		fab->root = agbno;
820 	else
821 		fab->root = NULLAGBLOCK;
822 
823 	trace_xrep_findroot_block(mp, ri->sc->sa.agno, agbno,
824 			be32_to_cpu(btblock->bb_magic), fab->height - 1);
825 out:
826 	xfs_trans_brelse(ri->sc->tp, bp);
827 	return error;
828 }
829 
830 /*
831  * Do any of the blocks in this rmap record match one of the btrees we're
832  * looking for?
833  */
834 STATIC int
835 xrep_findroot_rmap(
836 	struct xfs_btree_cur		*cur,
837 	struct xfs_rmap_irec		*rec,
838 	void				*priv)
839 {
840 	struct xrep_findroot		*ri = priv;
841 	struct xrep_find_ag_btree	*fab;
842 	xfs_agblock_t			b;
843 	bool				done;
844 	int				error = 0;
845 
846 	/* Ignore anything that isn't AG metadata. */
847 	if (!XFS_RMAP_NON_INODE_OWNER(rec->rm_owner))
848 		return 0;
849 
850 	/* Otherwise scan each block + btree type. */
851 	for (b = 0; b < rec->rm_blockcount; b++) {
852 		done = false;
853 		for (fab = ri->btree_info; fab->buf_ops; fab++) {
854 			if (rec->rm_owner != fab->rmap_owner)
855 				continue;
856 			error = xrep_findroot_block(ri, fab,
857 					rec->rm_owner, rec->rm_startblock + b,
858 					&done);
859 			if (error)
860 				return error;
861 			if (done)
862 				break;
863 		}
864 	}
865 
866 	return 0;
867 }
868 
869 /* Find the roots of the per-AG btrees described in btree_info. */
870 int
871 xrep_find_ag_btree_roots(
872 	struct xfs_scrub		*sc,
873 	struct xfs_buf			*agf_bp,
874 	struct xrep_find_ag_btree	*btree_info,
875 	struct xfs_buf			*agfl_bp)
876 {
877 	struct xfs_mount		*mp = sc->mp;
878 	struct xrep_findroot		ri;
879 	struct xrep_find_ag_btree	*fab;
880 	struct xfs_btree_cur		*cur;
881 	int				error;
882 
883 	ASSERT(xfs_buf_islocked(agf_bp));
884 	ASSERT(agfl_bp == NULL || xfs_buf_islocked(agfl_bp));
885 
886 	ri.sc = sc;
887 	ri.btree_info = btree_info;
888 	ri.agf = XFS_BUF_TO_AGF(agf_bp);
889 	ri.agfl_bp = agfl_bp;
890 	for (fab = btree_info; fab->buf_ops; fab++) {
891 		ASSERT(agfl_bp || fab->rmap_owner != XFS_RMAP_OWN_AG);
892 		ASSERT(XFS_RMAP_NON_INODE_OWNER(fab->rmap_owner));
893 		fab->root = NULLAGBLOCK;
894 		fab->height = 0;
895 	}
896 
897 	cur = xfs_rmapbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno);
898 	error = xfs_rmap_query_all(cur, xrep_findroot_rmap, &ri);
899 	xfs_btree_del_cursor(cur, error);
900 
901 	return error;
902 }
903 
904 /* Force a quotacheck the next time we mount. */
905 void
906 xrep_force_quotacheck(
907 	struct xfs_scrub	*sc,
908 	uint			dqtype)
909 {
910 	uint			flag;
911 
912 	flag = xfs_quota_chkd_flag(dqtype);
913 	if (!(flag & sc->mp->m_qflags))
914 		return;
915 
916 	sc->mp->m_qflags &= ~flag;
917 	spin_lock(&sc->mp->m_sb_lock);
918 	sc->mp->m_sb.sb_qflags &= ~flag;
919 	spin_unlock(&sc->mp->m_sb_lock);
920 	xfs_log_sb(sc->tp);
921 }
922 
923 /*
924  * Attach dquots to this inode, or schedule quotacheck to fix them.
925  *
926  * This function ensures that the appropriate dquots are attached to an inode.
927  * We cannot allow the dquot code to allocate an on-disk dquot block here
928  * because we're already in transaction context with the inode locked.  The
929  * on-disk dquot should already exist anyway.  If the quota code signals
930  * corruption or missing quota information, schedule quotacheck, which will
931  * repair corruptions in the quota metadata.
932  */
933 int
934 xrep_ino_dqattach(
935 	struct xfs_scrub	*sc)
936 {
937 	int			error;
938 
939 	error = xfs_qm_dqattach_locked(sc->ip, false);
940 	switch (error) {
941 	case -EFSBADCRC:
942 	case -EFSCORRUPTED:
943 	case -ENOENT:
944 		xfs_err_ratelimited(sc->mp,
945 "inode %llu repair encountered quota error %d, quotacheck forced.",
946 				(unsigned long long)sc->ip->i_ino, error);
947 		if (XFS_IS_UQUOTA_ON(sc->mp) && !sc->ip->i_udquot)
948 			xrep_force_quotacheck(sc, XFS_DQ_USER);
949 		if (XFS_IS_GQUOTA_ON(sc->mp) && !sc->ip->i_gdquot)
950 			xrep_force_quotacheck(sc, XFS_DQ_GROUP);
951 		if (XFS_IS_PQUOTA_ON(sc->mp) && !sc->ip->i_pdquot)
952 			xrep_force_quotacheck(sc, XFS_DQ_PROJ);
953 		/* fall through */
954 	case -ESRCH:
955 		error = 0;
956 		break;
957 	default:
958 		break;
959 	}
960 
961 	return error;
962 }
963