xref: /openbmc/linux/fs/xfs/scrub/agheader_repair.c (revision 3e30a927)
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_btree.h"
13 #include "xfs_log_format.h"
14 #include "xfs_trans.h"
15 #include "xfs_sb.h"
16 #include "xfs_alloc.h"
17 #include "xfs_alloc_btree.h"
18 #include "xfs_ialloc.h"
19 #include "xfs_ialloc_btree.h"
20 #include "xfs_rmap.h"
21 #include "xfs_rmap_btree.h"
22 #include "xfs_refcount_btree.h"
23 #include "scrub/scrub.h"
24 #include "scrub/common.h"
25 #include "scrub/trace.h"
26 #include "scrub/repair.h"
27 #include "scrub/bitmap.h"
28 
29 /* Superblock */
30 
31 /* Repair the superblock. */
32 int
33 xrep_superblock(
34 	struct xfs_scrub	*sc)
35 {
36 	struct xfs_mount	*mp = sc->mp;
37 	struct xfs_buf		*bp;
38 	xfs_agnumber_t		agno;
39 	int			error;
40 
41 	/* Don't try to repair AG 0's sb; let xfs_repair deal with it. */
42 	agno = sc->sm->sm_agno;
43 	if (agno == 0)
44 		return -EOPNOTSUPP;
45 
46 	error = xfs_sb_get_secondary(mp, sc->tp, agno, &bp);
47 	if (error)
48 		return error;
49 
50 	/* Copy AG 0's superblock to this one. */
51 	xfs_buf_zero(bp, 0, BBTOB(bp->b_length));
52 	xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb);
53 
54 	/* Write this to disk. */
55 	xfs_trans_buf_set_type(sc->tp, bp, XFS_BLFT_SB_BUF);
56 	xfs_trans_log_buf(sc->tp, bp, 0, BBTOB(bp->b_length) - 1);
57 	return error;
58 }
59 
60 /* AGF */
61 
62 struct xrep_agf_allocbt {
63 	struct xfs_scrub	*sc;
64 	xfs_agblock_t		freeblks;
65 	xfs_agblock_t		longest;
66 };
67 
68 /* Record free space shape information. */
69 STATIC int
70 xrep_agf_walk_allocbt(
71 	struct xfs_btree_cur		*cur,
72 	struct xfs_alloc_rec_incore	*rec,
73 	void				*priv)
74 {
75 	struct xrep_agf_allocbt		*raa = priv;
76 	int				error = 0;
77 
78 	if (xchk_should_terminate(raa->sc, &error))
79 		return error;
80 
81 	raa->freeblks += rec->ar_blockcount;
82 	if (rec->ar_blockcount > raa->longest)
83 		raa->longest = rec->ar_blockcount;
84 	return error;
85 }
86 
87 /* Does this AGFL block look sane? */
88 STATIC int
89 xrep_agf_check_agfl_block(
90 	struct xfs_mount	*mp,
91 	xfs_agblock_t		agbno,
92 	void			*priv)
93 {
94 	struct xfs_scrub	*sc = priv;
95 
96 	if (!xfs_verify_agbno(mp, sc->sa.agno, agbno))
97 		return -EFSCORRUPTED;
98 	return 0;
99 }
100 
101 /*
102  * Offset within the xrep_find_ag_btree array for each btree type.  Avoid the
103  * XFS_BTNUM_ names here to avoid creating a sparse array.
104  */
105 enum {
106 	XREP_AGF_BNOBT = 0,
107 	XREP_AGF_CNTBT,
108 	XREP_AGF_RMAPBT,
109 	XREP_AGF_REFCOUNTBT,
110 	XREP_AGF_END,
111 	XREP_AGF_MAX
112 };
113 
114 /* Check a btree root candidate. */
115 static inline bool
116 xrep_check_btree_root(
117 	struct xfs_scrub		*sc,
118 	struct xrep_find_ag_btree	*fab)
119 {
120 	struct xfs_mount		*mp = sc->mp;
121 	xfs_agnumber_t			agno = sc->sm->sm_agno;
122 
123 	return xfs_verify_agbno(mp, agno, fab->root) &&
124 	       fab->height <= XFS_BTREE_MAXLEVELS;
125 }
126 
127 /*
128  * Given the btree roots described by *fab, find the roots, check them for
129  * sanity, and pass the root data back out via *fab.
130  *
131  * This is /also/ a chicken and egg problem because we have to use the rmapbt
132  * (rooted in the AGF) to find the btrees rooted in the AGF.  We also have no
133  * idea if the btrees make any sense.  If we hit obvious corruptions in those
134  * btrees we'll bail out.
135  */
136 STATIC int
137 xrep_agf_find_btrees(
138 	struct xfs_scrub		*sc,
139 	struct xfs_buf			*agf_bp,
140 	struct xrep_find_ag_btree	*fab,
141 	struct xfs_buf			*agfl_bp)
142 {
143 	struct xfs_agf			*old_agf = XFS_BUF_TO_AGF(agf_bp);
144 	int				error;
145 
146 	/* Go find the root data. */
147 	error = xrep_find_ag_btree_roots(sc, agf_bp, fab, agfl_bp);
148 	if (error)
149 		return error;
150 
151 	/* We must find the bnobt, cntbt, and rmapbt roots. */
152 	if (!xrep_check_btree_root(sc, &fab[XREP_AGF_BNOBT]) ||
153 	    !xrep_check_btree_root(sc, &fab[XREP_AGF_CNTBT]) ||
154 	    !xrep_check_btree_root(sc, &fab[XREP_AGF_RMAPBT]))
155 		return -EFSCORRUPTED;
156 
157 	/*
158 	 * We relied on the rmapbt to reconstruct the AGF.  If we get a
159 	 * different root then something's seriously wrong.
160 	 */
161 	if (fab[XREP_AGF_RMAPBT].root !=
162 	    be32_to_cpu(old_agf->agf_roots[XFS_BTNUM_RMAPi]))
163 		return -EFSCORRUPTED;
164 
165 	/* We must find the refcountbt root if that feature is enabled. */
166 	if (xfs_sb_version_hasreflink(&sc->mp->m_sb) &&
167 	    !xrep_check_btree_root(sc, &fab[XREP_AGF_REFCOUNTBT]))
168 		return -EFSCORRUPTED;
169 
170 	return 0;
171 }
172 
173 /*
174  * Reinitialize the AGF header, making an in-core copy of the old contents so
175  * that we know which in-core state needs to be reinitialized.
176  */
177 STATIC void
178 xrep_agf_init_header(
179 	struct xfs_scrub	*sc,
180 	struct xfs_buf		*agf_bp,
181 	struct xfs_agf		*old_agf)
182 {
183 	struct xfs_mount	*mp = sc->mp;
184 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agf_bp);
185 
186 	memcpy(old_agf, agf, sizeof(*old_agf));
187 	memset(agf, 0, BBTOB(agf_bp->b_length));
188 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
189 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
190 	agf->agf_seqno = cpu_to_be32(sc->sa.agno);
191 	agf->agf_length = cpu_to_be32(xfs_ag_block_count(mp, sc->sa.agno));
192 	agf->agf_flfirst = old_agf->agf_flfirst;
193 	agf->agf_fllast = old_agf->agf_fllast;
194 	agf->agf_flcount = old_agf->agf_flcount;
195 	if (xfs_sb_version_hascrc(&mp->m_sb))
196 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
197 
198 	/* Mark the incore AGF data stale until we're done fixing things. */
199 	ASSERT(sc->sa.pag->pagf_init);
200 	sc->sa.pag->pagf_init = 0;
201 }
202 
203 /* Set btree root information in an AGF. */
204 STATIC void
205 xrep_agf_set_roots(
206 	struct xfs_scrub		*sc,
207 	struct xfs_agf			*agf,
208 	struct xrep_find_ag_btree	*fab)
209 {
210 	agf->agf_roots[XFS_BTNUM_BNOi] =
211 			cpu_to_be32(fab[XREP_AGF_BNOBT].root);
212 	agf->agf_levels[XFS_BTNUM_BNOi] =
213 			cpu_to_be32(fab[XREP_AGF_BNOBT].height);
214 
215 	agf->agf_roots[XFS_BTNUM_CNTi] =
216 			cpu_to_be32(fab[XREP_AGF_CNTBT].root);
217 	agf->agf_levels[XFS_BTNUM_CNTi] =
218 			cpu_to_be32(fab[XREP_AGF_CNTBT].height);
219 
220 	agf->agf_roots[XFS_BTNUM_RMAPi] =
221 			cpu_to_be32(fab[XREP_AGF_RMAPBT].root);
222 	agf->agf_levels[XFS_BTNUM_RMAPi] =
223 			cpu_to_be32(fab[XREP_AGF_RMAPBT].height);
224 
225 	if (xfs_sb_version_hasreflink(&sc->mp->m_sb)) {
226 		agf->agf_refcount_root =
227 				cpu_to_be32(fab[XREP_AGF_REFCOUNTBT].root);
228 		agf->agf_refcount_level =
229 				cpu_to_be32(fab[XREP_AGF_REFCOUNTBT].height);
230 	}
231 }
232 
233 /* Update all AGF fields which derive from btree contents. */
234 STATIC int
235 xrep_agf_calc_from_btrees(
236 	struct xfs_scrub	*sc,
237 	struct xfs_buf		*agf_bp)
238 {
239 	struct xrep_agf_allocbt	raa = { .sc = sc };
240 	struct xfs_btree_cur	*cur = NULL;
241 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agf_bp);
242 	struct xfs_mount	*mp = sc->mp;
243 	xfs_agblock_t		btreeblks;
244 	xfs_agblock_t		blocks;
245 	int			error;
246 
247 	/* Update the AGF counters from the bnobt. */
248 	cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
249 			XFS_BTNUM_BNO);
250 	error = xfs_alloc_query_all(cur, xrep_agf_walk_allocbt, &raa);
251 	if (error)
252 		goto err;
253 	error = xfs_btree_count_blocks(cur, &blocks);
254 	if (error)
255 		goto err;
256 	xfs_btree_del_cursor(cur, error);
257 	btreeblks = blocks - 1;
258 	agf->agf_freeblks = cpu_to_be32(raa.freeblks);
259 	agf->agf_longest = cpu_to_be32(raa.longest);
260 
261 	/* Update the AGF counters from the cntbt. */
262 	cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
263 			XFS_BTNUM_CNT);
264 	error = xfs_btree_count_blocks(cur, &blocks);
265 	if (error)
266 		goto err;
267 	xfs_btree_del_cursor(cur, error);
268 	btreeblks += blocks - 1;
269 
270 	/* Update the AGF counters from the rmapbt. */
271 	cur = xfs_rmapbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno);
272 	error = xfs_btree_count_blocks(cur, &blocks);
273 	if (error)
274 		goto err;
275 	xfs_btree_del_cursor(cur, error);
276 	agf->agf_rmap_blocks = cpu_to_be32(blocks);
277 	btreeblks += blocks - 1;
278 
279 	agf->agf_btreeblks = cpu_to_be32(btreeblks);
280 
281 	/* Update the AGF counters from the refcountbt. */
282 	if (xfs_sb_version_hasreflink(&mp->m_sb)) {
283 		cur = xfs_refcountbt_init_cursor(mp, sc->tp, agf_bp,
284 				sc->sa.agno);
285 		error = xfs_btree_count_blocks(cur, &blocks);
286 		if (error)
287 			goto err;
288 		xfs_btree_del_cursor(cur, error);
289 		agf->agf_refcount_blocks = cpu_to_be32(blocks);
290 	}
291 
292 	return 0;
293 err:
294 	xfs_btree_del_cursor(cur, error);
295 	return error;
296 }
297 
298 /* Commit the new AGF and reinitialize the incore state. */
299 STATIC int
300 xrep_agf_commit_new(
301 	struct xfs_scrub	*sc,
302 	struct xfs_buf		*agf_bp)
303 {
304 	struct xfs_perag	*pag;
305 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agf_bp);
306 
307 	/* Trigger fdblocks recalculation */
308 	xfs_force_summary_recalc(sc->mp);
309 
310 	/* Write this to disk. */
311 	xfs_trans_buf_set_type(sc->tp, agf_bp, XFS_BLFT_AGF_BUF);
312 	xfs_trans_log_buf(sc->tp, agf_bp, 0, BBTOB(agf_bp->b_length) - 1);
313 
314 	/* Now reinitialize the in-core counters we changed. */
315 	pag = sc->sa.pag;
316 	pag->pagf_btreeblks = be32_to_cpu(agf->agf_btreeblks);
317 	pag->pagf_freeblks = be32_to_cpu(agf->agf_freeblks);
318 	pag->pagf_longest = be32_to_cpu(agf->agf_longest);
319 	pag->pagf_levels[XFS_BTNUM_BNOi] =
320 			be32_to_cpu(agf->agf_levels[XFS_BTNUM_BNOi]);
321 	pag->pagf_levels[XFS_BTNUM_CNTi] =
322 			be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]);
323 	pag->pagf_levels[XFS_BTNUM_RMAPi] =
324 			be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAPi]);
325 	pag->pagf_refcount_level = be32_to_cpu(agf->agf_refcount_level);
326 	pag->pagf_init = 1;
327 
328 	return 0;
329 }
330 
331 /* Repair the AGF. v5 filesystems only. */
332 int
333 xrep_agf(
334 	struct xfs_scrub		*sc)
335 {
336 	struct xrep_find_ag_btree	fab[XREP_AGF_MAX] = {
337 		[XREP_AGF_BNOBT] = {
338 			.rmap_owner = XFS_RMAP_OWN_AG,
339 			.buf_ops = &xfs_bnobt_buf_ops,
340 		},
341 		[XREP_AGF_CNTBT] = {
342 			.rmap_owner = XFS_RMAP_OWN_AG,
343 			.buf_ops = &xfs_cntbt_buf_ops,
344 		},
345 		[XREP_AGF_RMAPBT] = {
346 			.rmap_owner = XFS_RMAP_OWN_AG,
347 			.buf_ops = &xfs_rmapbt_buf_ops,
348 		},
349 		[XREP_AGF_REFCOUNTBT] = {
350 			.rmap_owner = XFS_RMAP_OWN_REFC,
351 			.buf_ops = &xfs_refcountbt_buf_ops,
352 		},
353 		[XREP_AGF_END] = {
354 			.buf_ops = NULL,
355 		},
356 	};
357 	struct xfs_agf			old_agf;
358 	struct xfs_mount		*mp = sc->mp;
359 	struct xfs_buf			*agf_bp;
360 	struct xfs_buf			*agfl_bp;
361 	struct xfs_agf			*agf;
362 	int				error;
363 
364 	/* We require the rmapbt to rebuild anything. */
365 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
366 		return -EOPNOTSUPP;
367 
368 	xchk_perag_get(sc->mp, &sc->sa);
369 	/*
370 	 * Make sure we have the AGF buffer, as scrub might have decided it
371 	 * was corrupt after xfs_alloc_read_agf failed with -EFSCORRUPTED.
372 	 */
373 	error = xfs_trans_read_buf(mp, sc->tp, mp->m_ddev_targp,
374 			XFS_AG_DADDR(mp, sc->sa.agno, XFS_AGF_DADDR(mp)),
375 			XFS_FSS_TO_BB(mp, 1), 0, &agf_bp, NULL);
376 	if (error)
377 		return error;
378 	agf_bp->b_ops = &xfs_agf_buf_ops;
379 	agf = XFS_BUF_TO_AGF(agf_bp);
380 
381 	/*
382 	 * Load the AGFL so that we can screen out OWN_AG blocks that are on
383 	 * the AGFL now; these blocks might have once been part of the
384 	 * bno/cnt/rmap btrees but are not now.  This is a chicken and egg
385 	 * problem: the AGF is corrupt, so we have to trust the AGFL contents
386 	 * because we can't do any serious cross-referencing with any of the
387 	 * btrees rooted in the AGF.  If the AGFL contents are obviously bad
388 	 * then we'll bail out.
389 	 */
390 	error = xfs_alloc_read_agfl(mp, sc->tp, sc->sa.agno, &agfl_bp);
391 	if (error)
392 		return error;
393 
394 	/*
395 	 * Spot-check the AGFL blocks; if they're obviously corrupt then
396 	 * there's nothing we can do but bail out.
397 	 */
398 	error = xfs_agfl_walk(sc->mp, XFS_BUF_TO_AGF(agf_bp), agfl_bp,
399 			xrep_agf_check_agfl_block, sc);
400 	if (error)
401 		return error;
402 
403 	/*
404 	 * Find the AGF btree roots.  This is also a chicken-and-egg situation;
405 	 * see the function for more details.
406 	 */
407 	error = xrep_agf_find_btrees(sc, agf_bp, fab, agfl_bp);
408 	if (error)
409 		return error;
410 
411 	/* Start rewriting the header and implant the btrees we found. */
412 	xrep_agf_init_header(sc, agf_bp, &old_agf);
413 	xrep_agf_set_roots(sc, agf, fab);
414 	error = xrep_agf_calc_from_btrees(sc, agf_bp);
415 	if (error)
416 		goto out_revert;
417 
418 	/* Commit the changes and reinitialize incore state. */
419 	return xrep_agf_commit_new(sc, agf_bp);
420 
421 out_revert:
422 	/* Mark the incore AGF state stale and revert the AGF. */
423 	sc->sa.pag->pagf_init = 0;
424 	memcpy(agf, &old_agf, sizeof(old_agf));
425 	return error;
426 }
427 
428 /* AGFL */
429 
430 struct xrep_agfl {
431 	/* Bitmap of other OWN_AG metadata blocks. */
432 	struct xfs_bitmap	agmetablocks;
433 
434 	/* Bitmap of free space. */
435 	struct xfs_bitmap	*freesp;
436 
437 	struct xfs_scrub	*sc;
438 };
439 
440 /* Record all OWN_AG (free space btree) information from the rmap data. */
441 STATIC int
442 xrep_agfl_walk_rmap(
443 	struct xfs_btree_cur	*cur,
444 	struct xfs_rmap_irec	*rec,
445 	void			*priv)
446 {
447 	struct xrep_agfl	*ra = priv;
448 	xfs_fsblock_t		fsb;
449 	int			error = 0;
450 
451 	if (xchk_should_terminate(ra->sc, &error))
452 		return error;
453 
454 	/* Record all the OWN_AG blocks. */
455 	if (rec->rm_owner == XFS_RMAP_OWN_AG) {
456 		fsb = XFS_AGB_TO_FSB(cur->bc_mp, cur->bc_private.a.agno,
457 				rec->rm_startblock);
458 		error = xfs_bitmap_set(ra->freesp, fsb, rec->rm_blockcount);
459 		if (error)
460 			return error;
461 	}
462 
463 	return xfs_bitmap_set_btcur_path(&ra->agmetablocks, cur);
464 }
465 
466 /*
467  * Map out all the non-AGFL OWN_AG space in this AG so that we can deduce
468  * which blocks belong to the AGFL.
469  *
470  * Compute the set of old AGFL blocks by subtracting from the list of OWN_AG
471  * blocks the list of blocks owned by all other OWN_AG metadata (bnobt, cntbt,
472  * rmapbt).  These are the old AGFL blocks, so return that list and the number
473  * of blocks we're actually going to put back on the AGFL.
474  */
475 STATIC int
476 xrep_agfl_collect_blocks(
477 	struct xfs_scrub	*sc,
478 	struct xfs_buf		*agf_bp,
479 	struct xfs_bitmap	*agfl_extents,
480 	xfs_agblock_t		*flcount)
481 {
482 	struct xrep_agfl	ra;
483 	struct xfs_mount	*mp = sc->mp;
484 	struct xfs_btree_cur	*cur;
485 	struct xfs_bitmap_range	*br;
486 	struct xfs_bitmap_range	*n;
487 	int			error;
488 
489 	ra.sc = sc;
490 	ra.freesp = agfl_extents;
491 	xfs_bitmap_init(&ra.agmetablocks);
492 
493 	/* Find all space used by the free space btrees & rmapbt. */
494 	cur = xfs_rmapbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno);
495 	error = xfs_rmap_query_all(cur, xrep_agfl_walk_rmap, &ra);
496 	if (error)
497 		goto err;
498 	xfs_btree_del_cursor(cur, error);
499 
500 	/* Find all blocks currently being used by the bnobt. */
501 	cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
502 			XFS_BTNUM_BNO);
503 	error = xfs_bitmap_set_btblocks(&ra.agmetablocks, cur);
504 	if (error)
505 		goto err;
506 	xfs_btree_del_cursor(cur, error);
507 
508 	/* Find all blocks currently being used by the cntbt. */
509 	cur = xfs_allocbt_init_cursor(mp, sc->tp, agf_bp, sc->sa.agno,
510 			XFS_BTNUM_CNT);
511 	error = xfs_bitmap_set_btblocks(&ra.agmetablocks, cur);
512 	if (error)
513 		goto err;
514 
515 	xfs_btree_del_cursor(cur, error);
516 
517 	/*
518 	 * Drop the freesp meta blocks that are in use by btrees.
519 	 * The remaining blocks /should/ be AGFL blocks.
520 	 */
521 	error = xfs_bitmap_disunion(agfl_extents, &ra.agmetablocks);
522 	xfs_bitmap_destroy(&ra.agmetablocks);
523 	if (error)
524 		return error;
525 
526 	/*
527 	 * Calculate the new AGFL size.  If we found more blocks than fit in
528 	 * the AGFL we'll free them later.
529 	 */
530 	*flcount = 0;
531 	for_each_xfs_bitmap_extent(br, n, agfl_extents) {
532 		*flcount += br->len;
533 		if (*flcount > xfs_agfl_size(mp))
534 			break;
535 	}
536 	if (*flcount > xfs_agfl_size(mp))
537 		*flcount = xfs_agfl_size(mp);
538 	return 0;
539 
540 err:
541 	xfs_bitmap_destroy(&ra.agmetablocks);
542 	xfs_btree_del_cursor(cur, error);
543 	return error;
544 }
545 
546 /* Update the AGF and reset the in-core state. */
547 STATIC void
548 xrep_agfl_update_agf(
549 	struct xfs_scrub	*sc,
550 	struct xfs_buf		*agf_bp,
551 	xfs_agblock_t		flcount)
552 {
553 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agf_bp);
554 
555 	ASSERT(flcount <= xfs_agfl_size(sc->mp));
556 
557 	/* Trigger fdblocks recalculation */
558 	xfs_force_summary_recalc(sc->mp);
559 
560 	/* Update the AGF counters. */
561 	if (sc->sa.pag->pagf_init)
562 		sc->sa.pag->pagf_flcount = flcount;
563 	agf->agf_flfirst = cpu_to_be32(0);
564 	agf->agf_flcount = cpu_to_be32(flcount);
565 	agf->agf_fllast = cpu_to_be32(flcount - 1);
566 
567 	xfs_alloc_log_agf(sc->tp, agf_bp,
568 			XFS_AGF_FLFIRST | XFS_AGF_FLLAST | XFS_AGF_FLCOUNT);
569 }
570 
571 /* Write out a totally new AGFL. */
572 STATIC void
573 xrep_agfl_init_header(
574 	struct xfs_scrub	*sc,
575 	struct xfs_buf		*agfl_bp,
576 	struct xfs_bitmap	*agfl_extents,
577 	xfs_agblock_t		flcount)
578 {
579 	struct xfs_mount	*mp = sc->mp;
580 	__be32			*agfl_bno;
581 	struct xfs_bitmap_range	*br;
582 	struct xfs_bitmap_range	*n;
583 	struct xfs_agfl		*agfl;
584 	xfs_agblock_t		agbno;
585 	unsigned int		fl_off;
586 
587 	ASSERT(flcount <= xfs_agfl_size(mp));
588 
589 	/*
590 	 * Start rewriting the header by setting the bno[] array to
591 	 * NULLAGBLOCK, then setting AGFL header fields.
592 	 */
593 	agfl = XFS_BUF_TO_AGFL(agfl_bp);
594 	memset(agfl, 0xFF, BBTOB(agfl_bp->b_length));
595 	agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
596 	agfl->agfl_seqno = cpu_to_be32(sc->sa.agno);
597 	uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
598 
599 	/*
600 	 * Fill the AGFL with the remaining blocks.  If agfl_extents has more
601 	 * blocks than fit in the AGFL, they will be freed in a subsequent
602 	 * step.
603 	 */
604 	fl_off = 0;
605 	agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agfl_bp);
606 	for_each_xfs_bitmap_extent(br, n, agfl_extents) {
607 		agbno = XFS_FSB_TO_AGBNO(mp, br->start);
608 
609 		trace_xrep_agfl_insert(mp, sc->sa.agno, agbno, br->len);
610 
611 		while (br->len > 0 && fl_off < flcount) {
612 			agfl_bno[fl_off] = cpu_to_be32(agbno);
613 			fl_off++;
614 			agbno++;
615 
616 			/*
617 			 * We've now used br->start by putting it in the AGFL,
618 			 * so bump br so that we don't reap the block later.
619 			 */
620 			br->start++;
621 			br->len--;
622 		}
623 
624 		if (br->len)
625 			break;
626 		list_del(&br->list);
627 		kmem_free(br);
628 	}
629 
630 	/* Write new AGFL to disk. */
631 	xfs_trans_buf_set_type(sc->tp, agfl_bp, XFS_BLFT_AGFL_BUF);
632 	xfs_trans_log_buf(sc->tp, agfl_bp, 0, BBTOB(agfl_bp->b_length) - 1);
633 }
634 
635 /* Repair the AGFL. */
636 int
637 xrep_agfl(
638 	struct xfs_scrub	*sc)
639 {
640 	struct xfs_bitmap	agfl_extents;
641 	struct xfs_mount	*mp = sc->mp;
642 	struct xfs_buf		*agf_bp;
643 	struct xfs_buf		*agfl_bp;
644 	xfs_agblock_t		flcount;
645 	int			error;
646 
647 	/* We require the rmapbt to rebuild anything. */
648 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
649 		return -EOPNOTSUPP;
650 
651 	xchk_perag_get(sc->mp, &sc->sa);
652 	xfs_bitmap_init(&agfl_extents);
653 
654 	/*
655 	 * Read the AGF so that we can query the rmapbt.  We hope that there's
656 	 * nothing wrong with the AGF, but all the AG header repair functions
657 	 * have this chicken-and-egg problem.
658 	 */
659 	error = xfs_alloc_read_agf(mp, sc->tp, sc->sa.agno, 0, &agf_bp);
660 	if (error)
661 		return error;
662 
663 	/*
664 	 * Make sure we have the AGFL buffer, as scrub might have decided it
665 	 * was corrupt after xfs_alloc_read_agfl failed with -EFSCORRUPTED.
666 	 */
667 	error = xfs_trans_read_buf(mp, sc->tp, mp->m_ddev_targp,
668 			XFS_AG_DADDR(mp, sc->sa.agno, XFS_AGFL_DADDR(mp)),
669 			XFS_FSS_TO_BB(mp, 1), 0, &agfl_bp, NULL);
670 	if (error)
671 		return error;
672 	agfl_bp->b_ops = &xfs_agfl_buf_ops;
673 
674 	/* Gather all the extents we're going to put on the new AGFL. */
675 	error = xrep_agfl_collect_blocks(sc, agf_bp, &agfl_extents, &flcount);
676 	if (error)
677 		goto err;
678 
679 	/*
680 	 * Update AGF and AGFL.  We reset the global free block counter when
681 	 * we adjust the AGF flcount (which can fail) so avoid updating any
682 	 * buffers until we know that part works.
683 	 */
684 	xrep_agfl_update_agf(sc, agf_bp, flcount);
685 	xrep_agfl_init_header(sc, agfl_bp, &agfl_extents, flcount);
686 
687 	/*
688 	 * Ok, the AGFL should be ready to go now.  Roll the transaction to
689 	 * make the new AGFL permanent before we start using it to return
690 	 * freespace overflow to the freespace btrees.
691 	 */
692 	sc->sa.agf_bp = agf_bp;
693 	sc->sa.agfl_bp = agfl_bp;
694 	error = xrep_roll_ag_trans(sc);
695 	if (error)
696 		goto err;
697 
698 	/* Dump any AGFL overflow. */
699 	return xrep_reap_extents(sc, &agfl_extents, &XFS_RMAP_OINFO_AG,
700 			XFS_AG_RESV_AGFL);
701 err:
702 	xfs_bitmap_destroy(&agfl_extents);
703 	return error;
704 }
705 
706 /* AGI */
707 
708 /*
709  * Offset within the xrep_find_ag_btree array for each btree type.  Avoid the
710  * XFS_BTNUM_ names here to avoid creating a sparse array.
711  */
712 enum {
713 	XREP_AGI_INOBT = 0,
714 	XREP_AGI_FINOBT,
715 	XREP_AGI_END,
716 	XREP_AGI_MAX
717 };
718 
719 /*
720  * Given the inode btree roots described by *fab, find the roots, check them
721  * for sanity, and pass the root data back out via *fab.
722  */
723 STATIC int
724 xrep_agi_find_btrees(
725 	struct xfs_scrub		*sc,
726 	struct xrep_find_ag_btree	*fab)
727 {
728 	struct xfs_buf			*agf_bp;
729 	struct xfs_mount		*mp = sc->mp;
730 	int				error;
731 
732 	/* Read the AGF. */
733 	error = xfs_alloc_read_agf(mp, sc->tp, sc->sa.agno, 0, &agf_bp);
734 	if (error)
735 		return error;
736 
737 	/* Find the btree roots. */
738 	error = xrep_find_ag_btree_roots(sc, agf_bp, fab, NULL);
739 	if (error)
740 		return error;
741 
742 	/* We must find the inobt root. */
743 	if (!xrep_check_btree_root(sc, &fab[XREP_AGI_INOBT]))
744 		return -EFSCORRUPTED;
745 
746 	/* We must find the finobt root if that feature is enabled. */
747 	if (xfs_sb_version_hasfinobt(&mp->m_sb) &&
748 	    !xrep_check_btree_root(sc, &fab[XREP_AGI_FINOBT]))
749 		return -EFSCORRUPTED;
750 
751 	return 0;
752 }
753 
754 /*
755  * Reinitialize the AGI header, making an in-core copy of the old contents so
756  * that we know which in-core state needs to be reinitialized.
757  */
758 STATIC void
759 xrep_agi_init_header(
760 	struct xfs_scrub	*sc,
761 	struct xfs_buf		*agi_bp,
762 	struct xfs_agi		*old_agi)
763 {
764 	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agi_bp);
765 	struct xfs_mount	*mp = sc->mp;
766 
767 	memcpy(old_agi, agi, sizeof(*old_agi));
768 	memset(agi, 0, BBTOB(agi_bp->b_length));
769 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
770 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
771 	agi->agi_seqno = cpu_to_be32(sc->sa.agno);
772 	agi->agi_length = cpu_to_be32(xfs_ag_block_count(mp, sc->sa.agno));
773 	agi->agi_newino = cpu_to_be32(NULLAGINO);
774 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
775 	if (xfs_sb_version_hascrc(&mp->m_sb))
776 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
777 
778 	/* We don't know how to fix the unlinked list yet. */
779 	memcpy(&agi->agi_unlinked, &old_agi->agi_unlinked,
780 			sizeof(agi->agi_unlinked));
781 
782 	/* Mark the incore AGF data stale until we're done fixing things. */
783 	ASSERT(sc->sa.pag->pagi_init);
784 	sc->sa.pag->pagi_init = 0;
785 }
786 
787 /* Set btree root information in an AGI. */
788 STATIC void
789 xrep_agi_set_roots(
790 	struct xfs_scrub		*sc,
791 	struct xfs_agi			*agi,
792 	struct xrep_find_ag_btree	*fab)
793 {
794 	agi->agi_root = cpu_to_be32(fab[XREP_AGI_INOBT].root);
795 	agi->agi_level = cpu_to_be32(fab[XREP_AGI_INOBT].height);
796 
797 	if (xfs_sb_version_hasfinobt(&sc->mp->m_sb)) {
798 		agi->agi_free_root = cpu_to_be32(fab[XREP_AGI_FINOBT].root);
799 		agi->agi_free_level = cpu_to_be32(fab[XREP_AGI_FINOBT].height);
800 	}
801 }
802 
803 /* Update the AGI counters. */
804 STATIC int
805 xrep_agi_calc_from_btrees(
806 	struct xfs_scrub	*sc,
807 	struct xfs_buf		*agi_bp)
808 {
809 	struct xfs_btree_cur	*cur;
810 	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agi_bp);
811 	struct xfs_mount	*mp = sc->mp;
812 	xfs_agino_t		count;
813 	xfs_agino_t		freecount;
814 	int			error;
815 
816 	cur = xfs_inobt_init_cursor(mp, sc->tp, agi_bp, sc->sa.agno,
817 			XFS_BTNUM_INO);
818 	error = xfs_ialloc_count_inodes(cur, &count, &freecount);
819 	if (error)
820 		goto err;
821 	xfs_btree_del_cursor(cur, error);
822 
823 	agi->agi_count = cpu_to_be32(count);
824 	agi->agi_freecount = cpu_to_be32(freecount);
825 	return 0;
826 err:
827 	xfs_btree_del_cursor(cur, error);
828 	return error;
829 }
830 
831 /* Trigger reinitialization of the in-core data. */
832 STATIC int
833 xrep_agi_commit_new(
834 	struct xfs_scrub	*sc,
835 	struct xfs_buf		*agi_bp)
836 {
837 	struct xfs_perag	*pag;
838 	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agi_bp);
839 
840 	/* Trigger inode count recalculation */
841 	xfs_force_summary_recalc(sc->mp);
842 
843 	/* Write this to disk. */
844 	xfs_trans_buf_set_type(sc->tp, agi_bp, XFS_BLFT_AGI_BUF);
845 	xfs_trans_log_buf(sc->tp, agi_bp, 0, BBTOB(agi_bp->b_length) - 1);
846 
847 	/* Now reinitialize the in-core counters if necessary. */
848 	pag = sc->sa.pag;
849 	pag->pagi_count = be32_to_cpu(agi->agi_count);
850 	pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
851 	pag->pagi_init = 1;
852 
853 	return 0;
854 }
855 
856 /* Repair the AGI. */
857 int
858 xrep_agi(
859 	struct xfs_scrub		*sc)
860 {
861 	struct xrep_find_ag_btree	fab[XREP_AGI_MAX] = {
862 		[XREP_AGI_INOBT] = {
863 			.rmap_owner = XFS_RMAP_OWN_INOBT,
864 			.buf_ops = &xfs_inobt_buf_ops,
865 		},
866 		[XREP_AGI_FINOBT] = {
867 			.rmap_owner = XFS_RMAP_OWN_INOBT,
868 			.buf_ops = &xfs_finobt_buf_ops,
869 		},
870 		[XREP_AGI_END] = {
871 			.buf_ops = NULL
872 		},
873 	};
874 	struct xfs_agi			old_agi;
875 	struct xfs_mount		*mp = sc->mp;
876 	struct xfs_buf			*agi_bp;
877 	struct xfs_agi			*agi;
878 	int				error;
879 
880 	/* We require the rmapbt to rebuild anything. */
881 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
882 		return -EOPNOTSUPP;
883 
884 	xchk_perag_get(sc->mp, &sc->sa);
885 	/*
886 	 * Make sure we have the AGI buffer, as scrub might have decided it
887 	 * was corrupt after xfs_ialloc_read_agi failed with -EFSCORRUPTED.
888 	 */
889 	error = xfs_trans_read_buf(mp, sc->tp, mp->m_ddev_targp,
890 			XFS_AG_DADDR(mp, sc->sa.agno, XFS_AGI_DADDR(mp)),
891 			XFS_FSS_TO_BB(mp, 1), 0, &agi_bp, NULL);
892 	if (error)
893 		return error;
894 	agi_bp->b_ops = &xfs_agi_buf_ops;
895 	agi = XFS_BUF_TO_AGI(agi_bp);
896 
897 	/* Find the AGI btree roots. */
898 	error = xrep_agi_find_btrees(sc, fab);
899 	if (error)
900 		return error;
901 
902 	/* Start rewriting the header and implant the btrees we found. */
903 	xrep_agi_init_header(sc, agi_bp, &old_agi);
904 	xrep_agi_set_roots(sc, agi, fab);
905 	error = xrep_agi_calc_from_btrees(sc, agi_bp);
906 	if (error)
907 		goto out_revert;
908 
909 	/* Reinitialize in-core state. */
910 	return xrep_agi_commit_new(sc, agi_bp);
911 
912 out_revert:
913 	/* Mark the incore AGI state stale and revert the AGI. */
914 	sc->sa.pag->pagi_init = 0;
915 	memcpy(agi, &old_agi, sizeof(old_agi));
916 	return error;
917 }
918