xref: /linux/fs/xfs/scrub/refcount_repair.c (revision 79790b6818e96c58fe2bffee1b418c16e64e7b80)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2018-2023 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_defer.h"
13 #include "xfs_btree.h"
14 #include "xfs_btree_staging.h"
15 #include "xfs_inode.h"
16 #include "xfs_bit.h"
17 #include "xfs_log_format.h"
18 #include "xfs_trans.h"
19 #include "xfs_sb.h"
20 #include "xfs_alloc.h"
21 #include "xfs_ialloc.h"
22 #include "xfs_rmap.h"
23 #include "xfs_rmap_btree.h"
24 #include "xfs_refcount.h"
25 #include "xfs_refcount_btree.h"
26 #include "xfs_error.h"
27 #include "xfs_ag.h"
28 #include "xfs_health.h"
29 #include "scrub/xfs_scrub.h"
30 #include "scrub/scrub.h"
31 #include "scrub/common.h"
32 #include "scrub/btree.h"
33 #include "scrub/trace.h"
34 #include "scrub/repair.h"
35 #include "scrub/bitmap.h"
36 #include "scrub/agb_bitmap.h"
37 #include "scrub/xfile.h"
38 #include "scrub/xfarray.h"
39 #include "scrub/newbt.h"
40 #include "scrub/reap.h"
41 #include "scrub/rcbag.h"
42 
43 /*
44  * Rebuilding the Reference Count Btree
45  * ====================================
46  *
47  * This algorithm is "borrowed" from xfs_repair.  Imagine the rmap
48  * entries as rectangles representing extents of physical blocks, and
49  * that the rectangles can be laid down to allow them to overlap each
50  * other; then we know that we must emit a refcnt btree entry wherever
51  * the amount of overlap changes, i.e. the emission stimulus is
52  * level-triggered:
53  *
54  *                 -    ---
55  *       --      ----- ----   ---        ------
56  * --   ----     ----------- ----     ---------
57  * -------------------------------- -----------
58  * ^ ^  ^^ ^^    ^ ^^ ^^^  ^^^^  ^ ^^ ^  ^     ^
59  * 2 1  23 21    3 43 234  2123  1 01 2  3     0
60  *
61  * For our purposes, a rmap is a tuple (startblock, len, fileoff, owner).
62  *
63  * Note that in the actual refcnt btree we don't store the refcount < 2
64  * cases because the bnobt tells us which blocks are free; single-use
65  * blocks aren't recorded in the bnobt or the refcntbt.  If the rmapbt
66  * supports storing multiple entries covering a given block we could
67  * theoretically dispense with the refcntbt and simply count rmaps, but
68  * that's inefficient in the (hot) write path, so we'll take the cost of
69  * the extra tree to save time.  Also there's no guarantee that rmap
70  * will be enabled.
71  *
72  * Given an array of rmaps sorted by physical block number, a starting
73  * physical block (sp), a bag to hold rmaps that cover sp, and the next
74  * physical block where the level changes (np), we can reconstruct the
75  * refcount btree as follows:
76  *
77  * While there are still unprocessed rmaps in the array,
78  *  - Set sp to the physical block (pblk) of the next unprocessed rmap.
79  *  - Add to the bag all rmaps in the array where startblock == sp.
80  *  - Set np to the physical block where the bag size will change.  This
81  *    is the minimum of (the pblk of the next unprocessed rmap) and
82  *    (startblock + len of each rmap in the bag).
83  *  - Record the bag size as old_bag_size.
84  *
85  *  - While the bag isn't empty,
86  *     - Remove from the bag all rmaps where startblock + len == np.
87  *     - Add to the bag all rmaps in the array where startblock == np.
88  *     - If the bag size isn't old_bag_size, store the refcount entry
89  *       (sp, np - sp, bag_size) in the refcnt btree.
90  *     - If the bag is empty, break out of the inner loop.
91  *     - Set old_bag_size to the bag size
92  *     - Set sp = np.
93  *     - Set np to the physical block where the bag size will change.
94  *       This is the minimum of (the pblk of the next unprocessed rmap)
95  *       and (startblock + len of each rmap in the bag).
96  *
97  * Like all the other repairers, we make a list of all the refcount
98  * records we need, then reinitialize the refcount btree root and
99  * insert all the records.
100  */
101 
102 struct xrep_refc {
103 	/* refcount extents */
104 	struct xfarray		*refcount_records;
105 
106 	/* new refcountbt information */
107 	struct xrep_newbt	new_btree;
108 
109 	/* old refcountbt blocks */
110 	struct xagb_bitmap	old_refcountbt_blocks;
111 
112 	struct xfs_scrub	*sc;
113 
114 	/* get_records()'s position in the refcount record array. */
115 	xfarray_idx_t		array_cur;
116 
117 	/* # of refcountbt blocks */
118 	xfs_extlen_t		btblocks;
119 };
120 
121 /* Set us up to repair refcount btrees. */
122 int
xrep_setup_ag_refcountbt(struct xfs_scrub * sc)123 xrep_setup_ag_refcountbt(
124 	struct xfs_scrub	*sc)
125 {
126 	char			*descr;
127 	int			error;
128 
129 	descr = xchk_xfile_ag_descr(sc, "rmap record bag");
130 	error = xrep_setup_xfbtree(sc, descr);
131 	kfree(descr);
132 	return error;
133 }
134 
135 /* Check for any obvious conflicts with this shared/CoW staging extent. */
136 STATIC int
xrep_refc_check_ext(struct xfs_scrub * sc,const struct xfs_refcount_irec * rec)137 xrep_refc_check_ext(
138 	struct xfs_scrub		*sc,
139 	const struct xfs_refcount_irec	*rec)
140 {
141 	enum xbtree_recpacking		outcome;
142 	int				error;
143 
144 	if (xfs_refcount_check_irec(sc->sa.pag, rec) != NULL)
145 		return -EFSCORRUPTED;
146 
147 	/* Make sure this isn't free space. */
148 	error = xfs_alloc_has_records(sc->sa.bno_cur, rec->rc_startblock,
149 			rec->rc_blockcount, &outcome);
150 	if (error)
151 		return error;
152 	if (outcome != XBTREE_RECPACKING_EMPTY)
153 		return -EFSCORRUPTED;
154 
155 	/* Must not be an inode chunk. */
156 	error = xfs_ialloc_has_inodes_at_extent(sc->sa.ino_cur,
157 			rec->rc_startblock, rec->rc_blockcount, &outcome);
158 	if (error)
159 		return error;
160 	if (outcome != XBTREE_RECPACKING_EMPTY)
161 		return -EFSCORRUPTED;
162 
163 	return 0;
164 }
165 
166 /* Record a reference count extent. */
167 STATIC int
xrep_refc_stash(struct xrep_refc * rr,enum xfs_refc_domain domain,xfs_agblock_t agbno,xfs_extlen_t len,uint64_t refcount)168 xrep_refc_stash(
169 	struct xrep_refc		*rr,
170 	enum xfs_refc_domain		domain,
171 	xfs_agblock_t			agbno,
172 	xfs_extlen_t			len,
173 	uint64_t			refcount)
174 {
175 	struct xfs_refcount_irec	irec = {
176 		.rc_startblock		= agbno,
177 		.rc_blockcount		= len,
178 		.rc_domain		= domain,
179 	};
180 	struct xfs_scrub		*sc = rr->sc;
181 	int				error = 0;
182 
183 	if (xchk_should_terminate(sc, &error))
184 		return error;
185 
186 	irec.rc_refcount = min_t(uint64_t, MAXREFCOUNT, refcount);
187 
188 	error = xrep_refc_check_ext(rr->sc, &irec);
189 	if (error)
190 		return error;
191 
192 	trace_xrep_refc_found(sc->sa.pag, &irec);
193 
194 	return xfarray_append(rr->refcount_records, &irec);
195 }
196 
197 /* Record a CoW staging extent. */
198 STATIC int
xrep_refc_stash_cow(struct xrep_refc * rr,xfs_agblock_t agbno,xfs_extlen_t len)199 xrep_refc_stash_cow(
200 	struct xrep_refc		*rr,
201 	xfs_agblock_t			agbno,
202 	xfs_extlen_t			len)
203 {
204 	return xrep_refc_stash(rr, XFS_REFC_DOMAIN_COW, agbno, len, 1);
205 }
206 
207 /* Decide if an rmap could describe a shared extent. */
208 static inline bool
xrep_refc_rmap_shareable(struct xfs_mount * mp,const struct xfs_rmap_irec * rmap)209 xrep_refc_rmap_shareable(
210 	struct xfs_mount		*mp,
211 	const struct xfs_rmap_irec	*rmap)
212 {
213 	/* AG metadata are never sharable */
214 	if (XFS_RMAP_NON_INODE_OWNER(rmap->rm_owner))
215 		return false;
216 
217 	/* Metadata in files are never shareable */
218 	if (xfs_internal_inum(mp, rmap->rm_owner))
219 		return false;
220 
221 	/* Metadata and unwritten file blocks are not shareable. */
222 	if (rmap->rm_flags & (XFS_RMAP_ATTR_FORK | XFS_RMAP_BMBT_BLOCK |
223 			      XFS_RMAP_UNWRITTEN))
224 		return false;
225 
226 	return true;
227 }
228 
229 /*
230  * Walk along the reverse mapping records until we find one that could describe
231  * a shared extent.
232  */
233 STATIC int
xrep_refc_walk_rmaps(struct xrep_refc * rr,struct xfs_rmap_irec * rmap,bool * have_rec)234 xrep_refc_walk_rmaps(
235 	struct xrep_refc	*rr,
236 	struct xfs_rmap_irec	*rmap,
237 	bool			*have_rec)
238 {
239 	struct xfs_btree_cur	*cur = rr->sc->sa.rmap_cur;
240 	struct xfs_mount	*mp = cur->bc_mp;
241 	int			have_gt;
242 	int			error = 0;
243 
244 	*have_rec = false;
245 
246 	/*
247 	 * Loop through the remaining rmaps.  Remember CoW staging
248 	 * extents and the refcountbt blocks from the old tree for later
249 	 * disposal.  We can only share written data fork extents, so
250 	 * keep looping until we find an rmap for one.
251 	 */
252 	do {
253 		if (xchk_should_terminate(rr->sc, &error))
254 			return error;
255 
256 		error = xfs_btree_increment(cur, 0, &have_gt);
257 		if (error)
258 			return error;
259 		if (!have_gt)
260 			return 0;
261 
262 		error = xfs_rmap_get_rec(cur, rmap, &have_gt);
263 		if (error)
264 			return error;
265 		if (XFS_IS_CORRUPT(mp, !have_gt)) {
266 			xfs_btree_mark_sick(cur);
267 			return -EFSCORRUPTED;
268 		}
269 
270 		if (rmap->rm_owner == XFS_RMAP_OWN_COW) {
271 			error = xrep_refc_stash_cow(rr, rmap->rm_startblock,
272 					rmap->rm_blockcount);
273 			if (error)
274 				return error;
275 		} else if (rmap->rm_owner == XFS_RMAP_OWN_REFC) {
276 			/* refcountbt block, dump it when we're done. */
277 			rr->btblocks += rmap->rm_blockcount;
278 			error = xagb_bitmap_set(&rr->old_refcountbt_blocks,
279 					rmap->rm_startblock,
280 					rmap->rm_blockcount);
281 			if (error)
282 				return error;
283 		}
284 	} while (!xrep_refc_rmap_shareable(mp, rmap));
285 
286 	*have_rec = true;
287 	return 0;
288 }
289 
290 static inline uint32_t
xrep_refc_encode_startblock(const struct xfs_refcount_irec * irec)291 xrep_refc_encode_startblock(
292 	const struct xfs_refcount_irec	*irec)
293 {
294 	uint32_t			start;
295 
296 	start = irec->rc_startblock & ~XFS_REFC_COWFLAG;
297 	if (irec->rc_domain == XFS_REFC_DOMAIN_COW)
298 		start |= XFS_REFC_COWFLAG;
299 
300 	return start;
301 }
302 
303 /* Sort in the same order as the ondisk records. */
304 static int
xrep_refc_extent_cmp(const void * a,const void * b)305 xrep_refc_extent_cmp(
306 	const void			*a,
307 	const void			*b)
308 {
309 	const struct xfs_refcount_irec	*ap = a;
310 	const struct xfs_refcount_irec	*bp = b;
311 	uint32_t			sa, sb;
312 
313 	sa = xrep_refc_encode_startblock(ap);
314 	sb = xrep_refc_encode_startblock(bp);
315 
316 	if (sa > sb)
317 		return 1;
318 	if (sa < sb)
319 		return -1;
320 	return 0;
321 }
322 
323 /*
324  * Sort the refcount extents by startblock or else the btree records will be in
325  * the wrong order.  Make sure the records do not overlap in physical space.
326  */
327 STATIC int
xrep_refc_sort_records(struct xrep_refc * rr)328 xrep_refc_sort_records(
329 	struct xrep_refc		*rr)
330 {
331 	struct xfs_refcount_irec	irec;
332 	xfarray_idx_t			cur;
333 	enum xfs_refc_domain		dom = XFS_REFC_DOMAIN_SHARED;
334 	xfs_agblock_t			next_agbno = 0;
335 	int				error;
336 
337 	error = xfarray_sort(rr->refcount_records, xrep_refc_extent_cmp,
338 			XFARRAY_SORT_KILLABLE);
339 	if (error)
340 		return error;
341 
342 	foreach_xfarray_idx(rr->refcount_records, cur) {
343 		if (xchk_should_terminate(rr->sc, &error))
344 			return error;
345 
346 		error = xfarray_load(rr->refcount_records, cur, &irec);
347 		if (error)
348 			return error;
349 
350 		if (dom == XFS_REFC_DOMAIN_SHARED &&
351 		    irec.rc_domain == XFS_REFC_DOMAIN_COW) {
352 			dom = irec.rc_domain;
353 			next_agbno = 0;
354 		}
355 
356 		if (dom != irec.rc_domain)
357 			return -EFSCORRUPTED;
358 		if (irec.rc_startblock < next_agbno)
359 			return -EFSCORRUPTED;
360 
361 		next_agbno = irec.rc_startblock + irec.rc_blockcount;
362 	}
363 
364 	return error;
365 }
366 
367 /*
368  * Walk forward through the rmap btree to collect all rmaps starting at
369  * @bno in @rmap_bag.  These represent the file(s) that share ownership of
370  * the current block.  Upon return, the rmap cursor points to the last record
371  * satisfying the startblock constraint.
372  */
373 static int
xrep_refc_push_rmaps_at(struct xrep_refc * rr,struct rcbag * rcstack,xfs_agblock_t bno,struct xfs_rmap_irec * rmap,bool * have)374 xrep_refc_push_rmaps_at(
375 	struct xrep_refc	*rr,
376 	struct rcbag		*rcstack,
377 	xfs_agblock_t		bno,
378 	struct xfs_rmap_irec	*rmap,
379 	bool			*have)
380 {
381 	struct xfs_scrub	*sc = rr->sc;
382 	int			have_gt;
383 	int			error;
384 
385 	while (*have && rmap->rm_startblock == bno) {
386 		error = rcbag_add(rcstack, rr->sc->tp, rmap);
387 		if (error)
388 			return error;
389 
390 		error = xrep_refc_walk_rmaps(rr, rmap, have);
391 		if (error)
392 			return error;
393 	}
394 
395 	error = xfs_btree_decrement(sc->sa.rmap_cur, 0, &have_gt);
396 	if (error)
397 		return error;
398 	if (XFS_IS_CORRUPT(sc->mp, !have_gt)) {
399 		xfs_btree_mark_sick(sc->sa.rmap_cur);
400 		return -EFSCORRUPTED;
401 	}
402 
403 	return 0;
404 }
405 
406 /* Iterate all the rmap records to generate reference count data. */
407 STATIC int
xrep_refc_find_refcounts(struct xrep_refc * rr)408 xrep_refc_find_refcounts(
409 	struct xrep_refc	*rr)
410 {
411 	struct xfs_scrub	*sc = rr->sc;
412 	struct rcbag		*rcstack;
413 	uint64_t		old_stack_height;
414 	xfs_agblock_t		sbno;
415 	xfs_agblock_t		cbno;
416 	xfs_agblock_t		nbno;
417 	bool			have;
418 	int			error;
419 
420 	xrep_ag_btcur_init(sc, &sc->sa);
421 
422 	/*
423 	 * Set up a bag to store all the rmap records that we're tracking to
424 	 * generate a reference count record.  If the size of the bag exceeds
425 	 * MAXREFCOUNT, we clamp rc_refcount.
426 	 */
427 	error = rcbag_init(sc->mp, sc->xmbtp, &rcstack);
428 	if (error)
429 		goto out_cur;
430 
431 	/* Start the rmapbt cursor to the left of all records. */
432 	error = xfs_btree_goto_left_edge(sc->sa.rmap_cur);
433 	if (error)
434 		goto out_bag;
435 
436 	/* Process reverse mappings into refcount data. */
437 	while (xfs_btree_has_more_records(sc->sa.rmap_cur)) {
438 		struct xfs_rmap_irec	rmap;
439 
440 		/* Push all rmaps with pblk == sbno onto the stack */
441 		error = xrep_refc_walk_rmaps(rr, &rmap, &have);
442 		if (error)
443 			goto out_bag;
444 		if (!have)
445 			break;
446 		sbno = cbno = rmap.rm_startblock;
447 		error = xrep_refc_push_rmaps_at(rr, rcstack, sbno, &rmap,
448 				&have);
449 		if (error)
450 			goto out_bag;
451 
452 		/* Set nbno to the bno of the next refcount change */
453 		error = rcbag_next_edge(rcstack, sc->tp, &rmap, have, &nbno);
454 		if (error)
455 			goto out_bag;
456 
457 		ASSERT(nbno > sbno);
458 		old_stack_height = rcbag_count(rcstack);
459 
460 		/* While stack isn't empty... */
461 		while (rcbag_count(rcstack) > 0) {
462 			/* Pop all rmaps that end at nbno */
463 			error = rcbag_remove_ending_at(rcstack, sc->tp, nbno);
464 			if (error)
465 				goto out_bag;
466 
467 			/* Push array items that start at nbno */
468 			error = xrep_refc_walk_rmaps(rr, &rmap, &have);
469 			if (error)
470 				goto out_bag;
471 			if (have) {
472 				error = xrep_refc_push_rmaps_at(rr, rcstack,
473 						nbno, &rmap, &have);
474 				if (error)
475 					goto out_bag;
476 			}
477 
478 			/* Emit refcount if necessary */
479 			ASSERT(nbno > cbno);
480 			if (rcbag_count(rcstack) != old_stack_height) {
481 				if (old_stack_height > 1) {
482 					error = xrep_refc_stash(rr,
483 							XFS_REFC_DOMAIN_SHARED,
484 							cbno, nbno - cbno,
485 							old_stack_height);
486 					if (error)
487 						goto out_bag;
488 				}
489 				cbno = nbno;
490 			}
491 
492 			/* Stack empty, go find the next rmap */
493 			if (rcbag_count(rcstack) == 0)
494 				break;
495 			old_stack_height = rcbag_count(rcstack);
496 			sbno = nbno;
497 
498 			/* Set nbno to the bno of the next refcount change */
499 			error = rcbag_next_edge(rcstack, sc->tp, &rmap, have,
500 					&nbno);
501 			if (error)
502 				goto out_bag;
503 
504 			ASSERT(nbno > sbno);
505 		}
506 	}
507 
508 	ASSERT(rcbag_count(rcstack) == 0);
509 out_bag:
510 	rcbag_free(&rcstack);
511 out_cur:
512 	xchk_ag_btcur_free(&sc->sa);
513 	return error;
514 }
515 
516 /* Retrieve refcountbt data for bulk load. */
517 STATIC int
xrep_refc_get_records(struct xfs_btree_cur * cur,unsigned int idx,struct xfs_btree_block * block,unsigned int nr_wanted,void * priv)518 xrep_refc_get_records(
519 	struct xfs_btree_cur		*cur,
520 	unsigned int			idx,
521 	struct xfs_btree_block		*block,
522 	unsigned int			nr_wanted,
523 	void				*priv)
524 {
525 	struct xfs_refcount_irec	*irec = &cur->bc_rec.rc;
526 	struct xrep_refc		*rr = priv;
527 	union xfs_btree_rec		*block_rec;
528 	unsigned int			loaded;
529 	int				error;
530 
531 	for (loaded = 0; loaded < nr_wanted; loaded++, idx++) {
532 		error = xfarray_load(rr->refcount_records, rr->array_cur++,
533 				irec);
534 		if (error)
535 			return error;
536 
537 		block_rec = xfs_btree_rec_addr(cur, idx, block);
538 		cur->bc_ops->init_rec_from_cur(cur, block_rec);
539 	}
540 
541 	return loaded;
542 }
543 
544 /* Feed one of the new btree blocks to the bulk loader. */
545 STATIC int
xrep_refc_claim_block(struct xfs_btree_cur * cur,union xfs_btree_ptr * ptr,void * priv)546 xrep_refc_claim_block(
547 	struct xfs_btree_cur	*cur,
548 	union xfs_btree_ptr	*ptr,
549 	void			*priv)
550 {
551 	struct xrep_refc        *rr = priv;
552 
553 	return xrep_newbt_claim_block(cur, &rr->new_btree, ptr);
554 }
555 
556 /* Update the AGF counters. */
557 STATIC int
xrep_refc_reset_counters(struct xrep_refc * rr)558 xrep_refc_reset_counters(
559 	struct xrep_refc	*rr)
560 {
561 	struct xfs_scrub	*sc = rr->sc;
562 	struct xfs_perag	*pag = sc->sa.pag;
563 
564 	/*
565 	 * After we commit the new btree to disk, it is possible that the
566 	 * process to reap the old btree blocks will race with the AIL trying
567 	 * to checkpoint the old btree blocks into the filesystem.  If the new
568 	 * tree is shorter than the old one, the refcountbt write verifier will
569 	 * fail and the AIL will shut down the filesystem.
570 	 *
571 	 * To avoid this, save the old incore btree height values as the alt
572 	 * height values before re-initializing the perag info from the updated
573 	 * AGF to capture all the new values.
574 	 */
575 	pag->pagf_repair_refcount_level = pag->pagf_refcount_level;
576 
577 	/* Reinitialize with the values we just logged. */
578 	return xrep_reinit_pagf(sc);
579 }
580 
581 /*
582  * Use the collected refcount information to stage a new refcount btree.  If
583  * this is successful we'll return with the new btree root information logged
584  * to the repair transaction but not yet committed.
585  */
586 STATIC int
xrep_refc_build_new_tree(struct xrep_refc * rr)587 xrep_refc_build_new_tree(
588 	struct xrep_refc	*rr)
589 {
590 	struct xfs_scrub	*sc = rr->sc;
591 	struct xfs_btree_cur	*refc_cur;
592 	struct xfs_perag	*pag = sc->sa.pag;
593 	xfs_fsblock_t		fsbno;
594 	int			error;
595 
596 	error = xrep_refc_sort_records(rr);
597 	if (error)
598 		return error;
599 
600 	/*
601 	 * Prepare to construct the new btree by reserving disk space for the
602 	 * new btree and setting up all the accounting information we'll need
603 	 * to root the new btree while it's under construction and before we
604 	 * attach it to the AG header.
605 	 */
606 	fsbno = XFS_AGB_TO_FSB(sc->mp, pag->pag_agno, xfs_refc_block(sc->mp));
607 	xrep_newbt_init_ag(&rr->new_btree, sc, &XFS_RMAP_OINFO_REFC, fsbno,
608 			XFS_AG_RESV_METADATA);
609 	rr->new_btree.bload.get_records = xrep_refc_get_records;
610 	rr->new_btree.bload.claim_block = xrep_refc_claim_block;
611 
612 	/* Compute how many blocks we'll need. */
613 	refc_cur = xfs_refcountbt_init_cursor(sc->mp, NULL, NULL, pag);
614 	xfs_btree_stage_afakeroot(refc_cur, &rr->new_btree.afake);
615 	error = xfs_btree_bload_compute_geometry(refc_cur,
616 			&rr->new_btree.bload,
617 			xfarray_length(rr->refcount_records));
618 	if (error)
619 		goto err_cur;
620 
621 	/* Last chance to abort before we start committing fixes. */
622 	if (xchk_should_terminate(sc, &error))
623 		goto err_cur;
624 
625 	/* Reserve the space we'll need for the new btree. */
626 	error = xrep_newbt_alloc_blocks(&rr->new_btree,
627 			rr->new_btree.bload.nr_blocks);
628 	if (error)
629 		goto err_cur;
630 
631 	/*
632 	 * Due to btree slack factors, it's possible for a new btree to be one
633 	 * level taller than the old btree.  Update the incore btree height so
634 	 * that we don't trip the verifiers when writing the new btree blocks
635 	 * to disk.
636 	 */
637 	pag->pagf_repair_refcount_level = rr->new_btree.bload.btree_height;
638 
639 	/* Add all observed refcount records. */
640 	rr->array_cur = XFARRAY_CURSOR_INIT;
641 	error = xfs_btree_bload(refc_cur, &rr->new_btree.bload, rr);
642 	if (error)
643 		goto err_level;
644 
645 	/*
646 	 * Install the new btree in the AG header.  After this point the old
647 	 * btree is no longer accessible and the new tree is live.
648 	 */
649 	xfs_refcountbt_commit_staged_btree(refc_cur, sc->tp, sc->sa.agf_bp);
650 	xfs_btree_del_cursor(refc_cur, 0);
651 
652 	/* Reset the AGF counters now that we've changed the btree shape. */
653 	error = xrep_refc_reset_counters(rr);
654 	if (error)
655 		goto err_newbt;
656 
657 	/* Dispose of any unused blocks and the accounting information. */
658 	error = xrep_newbt_commit(&rr->new_btree);
659 	if (error)
660 		return error;
661 
662 	return xrep_roll_ag_trans(sc);
663 
664 err_level:
665 	pag->pagf_repair_refcount_level = 0;
666 err_cur:
667 	xfs_btree_del_cursor(refc_cur, error);
668 err_newbt:
669 	xrep_newbt_cancel(&rr->new_btree);
670 	return error;
671 }
672 
673 /*
674  * Now that we've logged the roots of the new btrees, invalidate all of the
675  * old blocks and free them.
676  */
677 STATIC int
xrep_refc_remove_old_tree(struct xrep_refc * rr)678 xrep_refc_remove_old_tree(
679 	struct xrep_refc	*rr)
680 {
681 	struct xfs_scrub	*sc = rr->sc;
682 	struct xfs_perag	*pag = sc->sa.pag;
683 	int			error;
684 
685 	/* Free the old refcountbt blocks if they're not in use. */
686 	error = xrep_reap_agblocks(sc, &rr->old_refcountbt_blocks,
687 			&XFS_RMAP_OINFO_REFC, XFS_AG_RESV_METADATA);
688 	if (error)
689 		return error;
690 
691 	/*
692 	 * Now that we've zapped all the old refcountbt blocks we can turn off
693 	 * the alternate height mechanism and reset the per-AG space
694 	 * reservations.
695 	 */
696 	pag->pagf_repair_refcount_level = 0;
697 	sc->flags |= XREP_RESET_PERAG_RESV;
698 	return 0;
699 }
700 
701 /* Rebuild the refcount btree. */
702 int
xrep_refcountbt(struct xfs_scrub * sc)703 xrep_refcountbt(
704 	struct xfs_scrub	*sc)
705 {
706 	struct xrep_refc	*rr;
707 	struct xfs_mount	*mp = sc->mp;
708 	char			*descr;
709 	int			error;
710 
711 	/* We require the rmapbt to rebuild anything. */
712 	if (!xfs_has_rmapbt(mp))
713 		return -EOPNOTSUPP;
714 
715 	rr = kzalloc(sizeof(struct xrep_refc), XCHK_GFP_FLAGS);
716 	if (!rr)
717 		return -ENOMEM;
718 	rr->sc = sc;
719 
720 	/* Set up enough storage to handle one refcount record per block. */
721 	descr = xchk_xfile_ag_descr(sc, "reference count records");
722 	error = xfarray_create(descr, mp->m_sb.sb_agblocks,
723 			sizeof(struct xfs_refcount_irec),
724 			&rr->refcount_records);
725 	kfree(descr);
726 	if (error)
727 		goto out_rr;
728 
729 	/* Collect all reference counts. */
730 	xagb_bitmap_init(&rr->old_refcountbt_blocks);
731 	error = xrep_refc_find_refcounts(rr);
732 	if (error)
733 		goto out_bitmap;
734 
735 	/* Rebuild the refcount information. */
736 	error = xrep_refc_build_new_tree(rr);
737 	if (error)
738 		goto out_bitmap;
739 
740 	/* Kill the old tree. */
741 	error = xrep_refc_remove_old_tree(rr);
742 	if (error)
743 		goto out_bitmap;
744 
745 out_bitmap:
746 	xagb_bitmap_destroy(&rr->old_refcountbt_blocks);
747 	xfarray_destroy(rr->refcount_records);
748 out_rr:
749 	kfree(rr);
750 	return error;
751 }
752