xref: /openbmc/linux/fs/xfs/libxfs/xfs_rmap_btree.c (revision 2025cf9e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2014 Red Hat, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_bit.h"
13 #include "xfs_sb.h"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_alloc.h"
19 #include "xfs_btree.h"
20 #include "xfs_rmap.h"
21 #include "xfs_rmap_btree.h"
22 #include "xfs_trace.h"
23 #include "xfs_cksum.h"
24 #include "xfs_error.h"
25 #include "xfs_extent_busy.h"
26 #include "xfs_ag_resv.h"
27 
28 /*
29  * Reverse map btree.
30  *
31  * This is a per-ag tree used to track the owner(s) of a given extent. With
32  * reflink it is possible for there to be multiple owners, which is a departure
33  * from classic XFS. Owner records for data extents are inserted when the
34  * extent is mapped and removed when an extent is unmapped.  Owner records for
35  * all other block types (i.e. metadata) are inserted when an extent is
36  * allocated and removed when an extent is freed. There can only be one owner
37  * of a metadata extent, usually an inode or some other metadata structure like
38  * an AG btree.
39  *
40  * The rmap btree is part of the free space management, so blocks for the tree
41  * are sourced from the agfl. Hence we need transaction reservation support for
42  * this tree so that the freelist is always large enough. This also impacts on
43  * the minimum space we need to leave free in the AG.
44  *
45  * The tree is ordered by [ag block, owner, offset]. This is a large key size,
46  * but it is the only way to enforce unique keys when a block can be owned by
47  * multiple files at any offset. There's no need to order/search by extent
48  * size for online updating/management of the tree. It is intended that most
49  * reverse lookups will be to find the owner(s) of a particular block, or to
50  * try to recover tree and file data from corrupt primary metadata.
51  */
52 
53 static struct xfs_btree_cur *
54 xfs_rmapbt_dup_cursor(
55 	struct xfs_btree_cur	*cur)
56 {
57 	return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
58 			cur->bc_private.a.agbp, cur->bc_private.a.agno);
59 }
60 
61 STATIC void
62 xfs_rmapbt_set_root(
63 	struct xfs_btree_cur	*cur,
64 	union xfs_btree_ptr	*ptr,
65 	int			inc)
66 {
67 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
68 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
69 	xfs_agnumber_t		seqno = be32_to_cpu(agf->agf_seqno);
70 	int			btnum = cur->bc_btnum;
71 	struct xfs_perag	*pag = xfs_perag_get(cur->bc_mp, seqno);
72 
73 	ASSERT(ptr->s != 0);
74 
75 	agf->agf_roots[btnum] = ptr->s;
76 	be32_add_cpu(&agf->agf_levels[btnum], inc);
77 	pag->pagf_levels[btnum] += inc;
78 	xfs_perag_put(pag);
79 
80 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
81 }
82 
83 STATIC int
84 xfs_rmapbt_alloc_block(
85 	struct xfs_btree_cur	*cur,
86 	union xfs_btree_ptr	*start,
87 	union xfs_btree_ptr	*new,
88 	int			*stat)
89 {
90 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
91 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
92 	int			error;
93 	xfs_agblock_t		bno;
94 
95 	/* Allocate the new block from the freelist. If we can't, give up.  */
96 	error = xfs_alloc_get_freelist(cur->bc_tp, cur->bc_private.a.agbp,
97 				       &bno, 1);
98 	if (error)
99 		return error;
100 
101 	trace_xfs_rmapbt_alloc_block(cur->bc_mp, cur->bc_private.a.agno,
102 			bno, 1);
103 	if (bno == NULLAGBLOCK) {
104 		*stat = 0;
105 		return 0;
106 	}
107 
108 	xfs_extent_busy_reuse(cur->bc_mp, cur->bc_private.a.agno, bno, 1,
109 			false);
110 
111 	xfs_trans_agbtree_delta(cur->bc_tp, 1);
112 	new->s = cpu_to_be32(bno);
113 	be32_add_cpu(&agf->agf_rmap_blocks, 1);
114 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
115 
116 	xfs_ag_resv_rmapbt_alloc(cur->bc_mp, cur->bc_private.a.agno);
117 
118 	*stat = 1;
119 	return 0;
120 }
121 
122 STATIC int
123 xfs_rmapbt_free_block(
124 	struct xfs_btree_cur	*cur,
125 	struct xfs_buf		*bp)
126 {
127 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
128 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
129 	xfs_agblock_t		bno;
130 	int			error;
131 
132 	bno = xfs_daddr_to_agbno(cur->bc_mp, XFS_BUF_ADDR(bp));
133 	trace_xfs_rmapbt_free_block(cur->bc_mp, cur->bc_private.a.agno,
134 			bno, 1);
135 	be32_add_cpu(&agf->agf_rmap_blocks, -1);
136 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
137 	error = xfs_alloc_put_freelist(cur->bc_tp, agbp, NULL, bno, 1);
138 	if (error)
139 		return error;
140 
141 	xfs_extent_busy_insert(cur->bc_tp, be32_to_cpu(agf->agf_seqno), bno, 1,
142 			      XFS_EXTENT_BUSY_SKIP_DISCARD);
143 	xfs_trans_agbtree_delta(cur->bc_tp, -1);
144 
145 	xfs_ag_resv_rmapbt_free(cur->bc_mp, cur->bc_private.a.agno);
146 
147 	return 0;
148 }
149 
150 STATIC int
151 xfs_rmapbt_get_minrecs(
152 	struct xfs_btree_cur	*cur,
153 	int			level)
154 {
155 	return cur->bc_mp->m_rmap_mnr[level != 0];
156 }
157 
158 STATIC int
159 xfs_rmapbt_get_maxrecs(
160 	struct xfs_btree_cur	*cur,
161 	int			level)
162 {
163 	return cur->bc_mp->m_rmap_mxr[level != 0];
164 }
165 
166 STATIC void
167 xfs_rmapbt_init_key_from_rec(
168 	union xfs_btree_key	*key,
169 	union xfs_btree_rec	*rec)
170 {
171 	key->rmap.rm_startblock = rec->rmap.rm_startblock;
172 	key->rmap.rm_owner = rec->rmap.rm_owner;
173 	key->rmap.rm_offset = rec->rmap.rm_offset;
174 }
175 
176 /*
177  * The high key for a reverse mapping record can be computed by shifting
178  * the startblock and offset to the highest value that would still map
179  * to that record.  In practice this means that we add blockcount-1 to
180  * the startblock for all records, and if the record is for a data/attr
181  * fork mapping, we add blockcount-1 to the offset too.
182  */
183 STATIC void
184 xfs_rmapbt_init_high_key_from_rec(
185 	union xfs_btree_key	*key,
186 	union xfs_btree_rec	*rec)
187 {
188 	uint64_t		off;
189 	int			adj;
190 
191 	adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
192 
193 	key->rmap.rm_startblock = rec->rmap.rm_startblock;
194 	be32_add_cpu(&key->rmap.rm_startblock, adj);
195 	key->rmap.rm_owner = rec->rmap.rm_owner;
196 	key->rmap.rm_offset = rec->rmap.rm_offset;
197 	if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
198 	    XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
199 		return;
200 	off = be64_to_cpu(key->rmap.rm_offset);
201 	off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
202 	key->rmap.rm_offset = cpu_to_be64(off);
203 }
204 
205 STATIC void
206 xfs_rmapbt_init_rec_from_cur(
207 	struct xfs_btree_cur	*cur,
208 	union xfs_btree_rec	*rec)
209 {
210 	rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
211 	rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
212 	rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
213 	rec->rmap.rm_offset = cpu_to_be64(
214 			xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
215 }
216 
217 STATIC void
218 xfs_rmapbt_init_ptr_from_cur(
219 	struct xfs_btree_cur	*cur,
220 	union xfs_btree_ptr	*ptr)
221 {
222 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(cur->bc_private.a.agbp);
223 
224 	ASSERT(cur->bc_private.a.agno == be32_to_cpu(agf->agf_seqno));
225 
226 	ptr->s = agf->agf_roots[cur->bc_btnum];
227 }
228 
229 STATIC int64_t
230 xfs_rmapbt_key_diff(
231 	struct xfs_btree_cur	*cur,
232 	union xfs_btree_key	*key)
233 {
234 	struct xfs_rmap_irec	*rec = &cur->bc_rec.r;
235 	struct xfs_rmap_key	*kp = &key->rmap;
236 	__u64			x, y;
237 	int64_t			d;
238 
239 	d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
240 	if (d)
241 		return d;
242 
243 	x = be64_to_cpu(kp->rm_owner);
244 	y = rec->rm_owner;
245 	if (x > y)
246 		return 1;
247 	else if (y > x)
248 		return -1;
249 
250 	x = XFS_RMAP_OFF(be64_to_cpu(kp->rm_offset));
251 	y = rec->rm_offset;
252 	if (x > y)
253 		return 1;
254 	else if (y > x)
255 		return -1;
256 	return 0;
257 }
258 
259 STATIC int64_t
260 xfs_rmapbt_diff_two_keys(
261 	struct xfs_btree_cur	*cur,
262 	union xfs_btree_key	*k1,
263 	union xfs_btree_key	*k2)
264 {
265 	struct xfs_rmap_key	*kp1 = &k1->rmap;
266 	struct xfs_rmap_key	*kp2 = &k2->rmap;
267 	int64_t			d;
268 	__u64			x, y;
269 
270 	d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
271 		       be32_to_cpu(kp2->rm_startblock);
272 	if (d)
273 		return d;
274 
275 	x = be64_to_cpu(kp1->rm_owner);
276 	y = be64_to_cpu(kp2->rm_owner);
277 	if (x > y)
278 		return 1;
279 	else if (y > x)
280 		return -1;
281 
282 	x = XFS_RMAP_OFF(be64_to_cpu(kp1->rm_offset));
283 	y = XFS_RMAP_OFF(be64_to_cpu(kp2->rm_offset));
284 	if (x > y)
285 		return 1;
286 	else if (y > x)
287 		return -1;
288 	return 0;
289 }
290 
291 static xfs_failaddr_t
292 xfs_rmapbt_verify(
293 	struct xfs_buf		*bp)
294 {
295 	struct xfs_mount	*mp = bp->b_target->bt_mount;
296 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
297 	struct xfs_perag	*pag = bp->b_pag;
298 	xfs_failaddr_t		fa;
299 	unsigned int		level;
300 
301 	/*
302 	 * magic number and level verification
303 	 *
304 	 * During growfs operations, we can't verify the exact level or owner as
305 	 * the perag is not fully initialised and hence not attached to the
306 	 * buffer.  In this case, check against the maximum tree depth.
307 	 *
308 	 * Similarly, during log recovery we will have a perag structure
309 	 * attached, but the agf information will not yet have been initialised
310 	 * from the on disk AGF. Again, we can only check against maximum limits
311 	 * in this case.
312 	 */
313 	if (!xfs_verify_magic(bp, block->bb_magic))
314 		return __this_address;
315 
316 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
317 		return __this_address;
318 	fa = xfs_btree_sblock_v5hdr_verify(bp);
319 	if (fa)
320 		return fa;
321 
322 	level = be16_to_cpu(block->bb_level);
323 	if (pag && pag->pagf_init) {
324 		if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
325 			return __this_address;
326 	} else if (level >= mp->m_rmap_maxlevels)
327 		return __this_address;
328 
329 	return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
330 }
331 
332 static void
333 xfs_rmapbt_read_verify(
334 	struct xfs_buf	*bp)
335 {
336 	xfs_failaddr_t	fa;
337 
338 	if (!xfs_btree_sblock_verify_crc(bp))
339 		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
340 	else {
341 		fa = xfs_rmapbt_verify(bp);
342 		if (fa)
343 			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
344 	}
345 
346 	if (bp->b_error)
347 		trace_xfs_btree_corrupt(bp, _RET_IP_);
348 }
349 
350 static void
351 xfs_rmapbt_write_verify(
352 	struct xfs_buf	*bp)
353 {
354 	xfs_failaddr_t	fa;
355 
356 	fa = xfs_rmapbt_verify(bp);
357 	if (fa) {
358 		trace_xfs_btree_corrupt(bp, _RET_IP_);
359 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
360 		return;
361 	}
362 	xfs_btree_sblock_calc_crc(bp);
363 
364 }
365 
366 const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
367 	.name			= "xfs_rmapbt",
368 	.magic			= { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
369 	.verify_read		= xfs_rmapbt_read_verify,
370 	.verify_write		= xfs_rmapbt_write_verify,
371 	.verify_struct		= xfs_rmapbt_verify,
372 };
373 
374 STATIC int
375 xfs_rmapbt_keys_inorder(
376 	struct xfs_btree_cur	*cur,
377 	union xfs_btree_key	*k1,
378 	union xfs_btree_key	*k2)
379 {
380 	uint32_t		x;
381 	uint32_t		y;
382 	uint64_t		a;
383 	uint64_t		b;
384 
385 	x = be32_to_cpu(k1->rmap.rm_startblock);
386 	y = be32_to_cpu(k2->rmap.rm_startblock);
387 	if (x < y)
388 		return 1;
389 	else if (x > y)
390 		return 0;
391 	a = be64_to_cpu(k1->rmap.rm_owner);
392 	b = be64_to_cpu(k2->rmap.rm_owner);
393 	if (a < b)
394 		return 1;
395 	else if (a > b)
396 		return 0;
397 	a = XFS_RMAP_OFF(be64_to_cpu(k1->rmap.rm_offset));
398 	b = XFS_RMAP_OFF(be64_to_cpu(k2->rmap.rm_offset));
399 	if (a <= b)
400 		return 1;
401 	return 0;
402 }
403 
404 STATIC int
405 xfs_rmapbt_recs_inorder(
406 	struct xfs_btree_cur	*cur,
407 	union xfs_btree_rec	*r1,
408 	union xfs_btree_rec	*r2)
409 {
410 	uint32_t		x;
411 	uint32_t		y;
412 	uint64_t		a;
413 	uint64_t		b;
414 
415 	x = be32_to_cpu(r1->rmap.rm_startblock);
416 	y = be32_to_cpu(r2->rmap.rm_startblock);
417 	if (x < y)
418 		return 1;
419 	else if (x > y)
420 		return 0;
421 	a = be64_to_cpu(r1->rmap.rm_owner);
422 	b = be64_to_cpu(r2->rmap.rm_owner);
423 	if (a < b)
424 		return 1;
425 	else if (a > b)
426 		return 0;
427 	a = XFS_RMAP_OFF(be64_to_cpu(r1->rmap.rm_offset));
428 	b = XFS_RMAP_OFF(be64_to_cpu(r2->rmap.rm_offset));
429 	if (a <= b)
430 		return 1;
431 	return 0;
432 }
433 
434 static const struct xfs_btree_ops xfs_rmapbt_ops = {
435 	.rec_len		= sizeof(struct xfs_rmap_rec),
436 	.key_len		= 2 * sizeof(struct xfs_rmap_key),
437 
438 	.dup_cursor		= xfs_rmapbt_dup_cursor,
439 	.set_root		= xfs_rmapbt_set_root,
440 	.alloc_block		= xfs_rmapbt_alloc_block,
441 	.free_block		= xfs_rmapbt_free_block,
442 	.get_minrecs		= xfs_rmapbt_get_minrecs,
443 	.get_maxrecs		= xfs_rmapbt_get_maxrecs,
444 	.init_key_from_rec	= xfs_rmapbt_init_key_from_rec,
445 	.init_high_key_from_rec	= xfs_rmapbt_init_high_key_from_rec,
446 	.init_rec_from_cur	= xfs_rmapbt_init_rec_from_cur,
447 	.init_ptr_from_cur	= xfs_rmapbt_init_ptr_from_cur,
448 	.key_diff		= xfs_rmapbt_key_diff,
449 	.buf_ops		= &xfs_rmapbt_buf_ops,
450 	.diff_two_keys		= xfs_rmapbt_diff_two_keys,
451 	.keys_inorder		= xfs_rmapbt_keys_inorder,
452 	.recs_inorder		= xfs_rmapbt_recs_inorder,
453 };
454 
455 /*
456  * Allocate a new allocation btree cursor.
457  */
458 struct xfs_btree_cur *
459 xfs_rmapbt_init_cursor(
460 	struct xfs_mount	*mp,
461 	struct xfs_trans	*tp,
462 	struct xfs_buf		*agbp,
463 	xfs_agnumber_t		agno)
464 {
465 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
466 	struct xfs_btree_cur	*cur;
467 
468 	cur = kmem_zone_zalloc(xfs_btree_cur_zone, KM_NOFS);
469 	cur->bc_tp = tp;
470 	cur->bc_mp = mp;
471 	/* Overlapping btree; 2 keys per pointer. */
472 	cur->bc_btnum = XFS_BTNUM_RMAP;
473 	cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
474 	cur->bc_blocklog = mp->m_sb.sb_blocklog;
475 	cur->bc_ops = &xfs_rmapbt_ops;
476 	cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
477 	cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
478 
479 	cur->bc_private.a.agbp = agbp;
480 	cur->bc_private.a.agno = agno;
481 
482 	return cur;
483 }
484 
485 /*
486  * Calculate number of records in an rmap btree block.
487  */
488 int
489 xfs_rmapbt_maxrecs(
490 	int			blocklen,
491 	int			leaf)
492 {
493 	blocklen -= XFS_RMAP_BLOCK_LEN;
494 
495 	if (leaf)
496 		return blocklen / sizeof(struct xfs_rmap_rec);
497 	return blocklen /
498 		(2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
499 }
500 
501 /* Compute the maximum height of an rmap btree. */
502 void
503 xfs_rmapbt_compute_maxlevels(
504 	struct xfs_mount		*mp)
505 {
506 	/*
507 	 * On a non-reflink filesystem, the maximum number of rmap
508 	 * records is the number of blocks in the AG, hence the max
509 	 * rmapbt height is log_$maxrecs($agblocks).  However, with
510 	 * reflink each AG block can have up to 2^32 (per the refcount
511 	 * record format) owners, which means that theoretically we
512 	 * could face up to 2^64 rmap records.
513 	 *
514 	 * That effectively means that the max rmapbt height must be
515 	 * XFS_BTREE_MAXLEVELS.  "Fortunately" we'll run out of AG
516 	 * blocks to feed the rmapbt long before the rmapbt reaches
517 	 * maximum height.  The reflink code uses ag_resv_critical to
518 	 * disallow reflinking when less than 10% of the per-AG metadata
519 	 * block reservation since the fallback is a regular file copy.
520 	 */
521 	if (xfs_sb_version_hasreflink(&mp->m_sb))
522 		mp->m_rmap_maxlevels = XFS_BTREE_MAXLEVELS;
523 	else
524 		mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
525 				mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
526 }
527 
528 /* Calculate the refcount btree size for some records. */
529 xfs_extlen_t
530 xfs_rmapbt_calc_size(
531 	struct xfs_mount	*mp,
532 	unsigned long long	len)
533 {
534 	return xfs_btree_calc_size(mp->m_rmap_mnr, len);
535 }
536 
537 /*
538  * Calculate the maximum refcount btree size.
539  */
540 xfs_extlen_t
541 xfs_rmapbt_max_size(
542 	struct xfs_mount	*mp,
543 	xfs_agblock_t		agblocks)
544 {
545 	/* Bail out if we're uninitialized, which can happen in mkfs. */
546 	if (mp->m_rmap_mxr[0] == 0)
547 		return 0;
548 
549 	return xfs_rmapbt_calc_size(mp, agblocks);
550 }
551 
552 /*
553  * Figure out how many blocks to reserve and how many are used by this btree.
554  */
555 int
556 xfs_rmapbt_calc_reserves(
557 	struct xfs_mount	*mp,
558 	struct xfs_trans	*tp,
559 	xfs_agnumber_t		agno,
560 	xfs_extlen_t		*ask,
561 	xfs_extlen_t		*used)
562 {
563 	struct xfs_buf		*agbp;
564 	struct xfs_agf		*agf;
565 	xfs_agblock_t		agblocks;
566 	xfs_extlen_t		tree_len;
567 	int			error;
568 
569 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
570 		return 0;
571 
572 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
573 	if (error)
574 		return error;
575 
576 	agf = XFS_BUF_TO_AGF(agbp);
577 	agblocks = be32_to_cpu(agf->agf_length);
578 	tree_len = be32_to_cpu(agf->agf_rmap_blocks);
579 	xfs_trans_brelse(tp, agbp);
580 
581 	/*
582 	 * The log is permanently allocated, so the space it occupies will
583 	 * never be available for the kinds of things that would require btree
584 	 * expansion.  We therefore can pretend the space isn't there.
585 	 */
586 	if (mp->m_sb.sb_logstart &&
587 	    XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == agno)
588 		agblocks -= mp->m_sb.sb_logblocks;
589 
590 	/* Reserve 1% of the AG or enough for 1 block per record. */
591 	*ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));
592 	*used += tree_len;
593 
594 	return error;
595 }
596