xref: /openbmc/linux/fs/xfs/libxfs/xfs_rmap_btree.c (revision b8d312aa)
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_sb.h"
13 #include "xfs_mount.h"
14 #include "xfs_trans.h"
15 #include "xfs_alloc.h"
16 #include "xfs_btree.h"
17 #include "xfs_rmap.h"
18 #include "xfs_rmap_btree.h"
19 #include "xfs_trace.h"
20 #include "xfs_error.h"
21 #include "xfs_extent_busy.h"
22 #include "xfs_ag_resv.h"
23 
24 /*
25  * Reverse map btree.
26  *
27  * This is a per-ag tree used to track the owner(s) of a given extent. With
28  * reflink it is possible for there to be multiple owners, which is a departure
29  * from classic XFS. Owner records for data extents are inserted when the
30  * extent is mapped and removed when an extent is unmapped.  Owner records for
31  * all other block types (i.e. metadata) are inserted when an extent is
32  * allocated and removed when an extent is freed. There can only be one owner
33  * of a metadata extent, usually an inode or some other metadata structure like
34  * an AG btree.
35  *
36  * The rmap btree is part of the free space management, so blocks for the tree
37  * are sourced from the agfl. Hence we need transaction reservation support for
38  * this tree so that the freelist is always large enough. This also impacts on
39  * the minimum space we need to leave free in the AG.
40  *
41  * The tree is ordered by [ag block, owner, offset]. This is a large key size,
42  * but it is the only way to enforce unique keys when a block can be owned by
43  * multiple files at any offset. There's no need to order/search by extent
44  * size for online updating/management of the tree. It is intended that most
45  * reverse lookups will be to find the owner(s) of a particular block, or to
46  * try to recover tree and file data from corrupt primary metadata.
47  */
48 
49 static struct xfs_btree_cur *
50 xfs_rmapbt_dup_cursor(
51 	struct xfs_btree_cur	*cur)
52 {
53 	return xfs_rmapbt_init_cursor(cur->bc_mp, cur->bc_tp,
54 			cur->bc_private.a.agbp, cur->bc_private.a.agno);
55 }
56 
57 STATIC void
58 xfs_rmapbt_set_root(
59 	struct xfs_btree_cur	*cur,
60 	union xfs_btree_ptr	*ptr,
61 	int			inc)
62 {
63 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
64 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
65 	xfs_agnumber_t		seqno = be32_to_cpu(agf->agf_seqno);
66 	int			btnum = cur->bc_btnum;
67 	struct xfs_perag	*pag = xfs_perag_get(cur->bc_mp, seqno);
68 
69 	ASSERT(ptr->s != 0);
70 
71 	agf->agf_roots[btnum] = ptr->s;
72 	be32_add_cpu(&agf->agf_levels[btnum], inc);
73 	pag->pagf_levels[btnum] += inc;
74 	xfs_perag_put(pag);
75 
76 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_ROOTS | XFS_AGF_LEVELS);
77 }
78 
79 STATIC int
80 xfs_rmapbt_alloc_block(
81 	struct xfs_btree_cur	*cur,
82 	union xfs_btree_ptr	*start,
83 	union xfs_btree_ptr	*new,
84 	int			*stat)
85 {
86 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
87 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
88 	int			error;
89 	xfs_agblock_t		bno;
90 
91 	/* Allocate the new block from the freelist. If we can't, give up.  */
92 	error = xfs_alloc_get_freelist(cur->bc_tp, cur->bc_private.a.agbp,
93 				       &bno, 1);
94 	if (error)
95 		return error;
96 
97 	trace_xfs_rmapbt_alloc_block(cur->bc_mp, cur->bc_private.a.agno,
98 			bno, 1);
99 	if (bno == NULLAGBLOCK) {
100 		*stat = 0;
101 		return 0;
102 	}
103 
104 	xfs_extent_busy_reuse(cur->bc_mp, cur->bc_private.a.agno, bno, 1,
105 			false);
106 
107 	xfs_trans_agbtree_delta(cur->bc_tp, 1);
108 	new->s = cpu_to_be32(bno);
109 	be32_add_cpu(&agf->agf_rmap_blocks, 1);
110 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
111 
112 	xfs_ag_resv_rmapbt_alloc(cur->bc_mp, cur->bc_private.a.agno);
113 
114 	*stat = 1;
115 	return 0;
116 }
117 
118 STATIC int
119 xfs_rmapbt_free_block(
120 	struct xfs_btree_cur	*cur,
121 	struct xfs_buf		*bp)
122 {
123 	struct xfs_buf		*agbp = cur->bc_private.a.agbp;
124 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
125 	xfs_agblock_t		bno;
126 	int			error;
127 
128 	bno = xfs_daddr_to_agbno(cur->bc_mp, XFS_BUF_ADDR(bp));
129 	trace_xfs_rmapbt_free_block(cur->bc_mp, cur->bc_private.a.agno,
130 			bno, 1);
131 	be32_add_cpu(&agf->agf_rmap_blocks, -1);
132 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_RMAP_BLOCKS);
133 	error = xfs_alloc_put_freelist(cur->bc_tp, agbp, NULL, bno, 1);
134 	if (error)
135 		return error;
136 
137 	xfs_extent_busy_insert(cur->bc_tp, be32_to_cpu(agf->agf_seqno), bno, 1,
138 			      XFS_EXTENT_BUSY_SKIP_DISCARD);
139 	xfs_trans_agbtree_delta(cur->bc_tp, -1);
140 
141 	xfs_ag_resv_rmapbt_free(cur->bc_mp, cur->bc_private.a.agno);
142 
143 	return 0;
144 }
145 
146 STATIC int
147 xfs_rmapbt_get_minrecs(
148 	struct xfs_btree_cur	*cur,
149 	int			level)
150 {
151 	return cur->bc_mp->m_rmap_mnr[level != 0];
152 }
153 
154 STATIC int
155 xfs_rmapbt_get_maxrecs(
156 	struct xfs_btree_cur	*cur,
157 	int			level)
158 {
159 	return cur->bc_mp->m_rmap_mxr[level != 0];
160 }
161 
162 STATIC void
163 xfs_rmapbt_init_key_from_rec(
164 	union xfs_btree_key	*key,
165 	union xfs_btree_rec	*rec)
166 {
167 	key->rmap.rm_startblock = rec->rmap.rm_startblock;
168 	key->rmap.rm_owner = rec->rmap.rm_owner;
169 	key->rmap.rm_offset = rec->rmap.rm_offset;
170 }
171 
172 /*
173  * The high key for a reverse mapping record can be computed by shifting
174  * the startblock and offset to the highest value that would still map
175  * to that record.  In practice this means that we add blockcount-1 to
176  * the startblock for all records, and if the record is for a data/attr
177  * fork mapping, we add blockcount-1 to the offset too.
178  */
179 STATIC void
180 xfs_rmapbt_init_high_key_from_rec(
181 	union xfs_btree_key	*key,
182 	union xfs_btree_rec	*rec)
183 {
184 	uint64_t		off;
185 	int			adj;
186 
187 	adj = be32_to_cpu(rec->rmap.rm_blockcount) - 1;
188 
189 	key->rmap.rm_startblock = rec->rmap.rm_startblock;
190 	be32_add_cpu(&key->rmap.rm_startblock, adj);
191 	key->rmap.rm_owner = rec->rmap.rm_owner;
192 	key->rmap.rm_offset = rec->rmap.rm_offset;
193 	if (XFS_RMAP_NON_INODE_OWNER(be64_to_cpu(rec->rmap.rm_owner)) ||
194 	    XFS_RMAP_IS_BMBT_BLOCK(be64_to_cpu(rec->rmap.rm_offset)))
195 		return;
196 	off = be64_to_cpu(key->rmap.rm_offset);
197 	off = (XFS_RMAP_OFF(off) + adj) | (off & ~XFS_RMAP_OFF_MASK);
198 	key->rmap.rm_offset = cpu_to_be64(off);
199 }
200 
201 STATIC void
202 xfs_rmapbt_init_rec_from_cur(
203 	struct xfs_btree_cur	*cur,
204 	union xfs_btree_rec	*rec)
205 {
206 	rec->rmap.rm_startblock = cpu_to_be32(cur->bc_rec.r.rm_startblock);
207 	rec->rmap.rm_blockcount = cpu_to_be32(cur->bc_rec.r.rm_blockcount);
208 	rec->rmap.rm_owner = cpu_to_be64(cur->bc_rec.r.rm_owner);
209 	rec->rmap.rm_offset = cpu_to_be64(
210 			xfs_rmap_irec_offset_pack(&cur->bc_rec.r));
211 }
212 
213 STATIC void
214 xfs_rmapbt_init_ptr_from_cur(
215 	struct xfs_btree_cur	*cur,
216 	union xfs_btree_ptr	*ptr)
217 {
218 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(cur->bc_private.a.agbp);
219 
220 	ASSERT(cur->bc_private.a.agno == be32_to_cpu(agf->agf_seqno));
221 
222 	ptr->s = agf->agf_roots[cur->bc_btnum];
223 }
224 
225 STATIC int64_t
226 xfs_rmapbt_key_diff(
227 	struct xfs_btree_cur	*cur,
228 	union xfs_btree_key	*key)
229 {
230 	struct xfs_rmap_irec	*rec = &cur->bc_rec.r;
231 	struct xfs_rmap_key	*kp = &key->rmap;
232 	__u64			x, y;
233 	int64_t			d;
234 
235 	d = (int64_t)be32_to_cpu(kp->rm_startblock) - rec->rm_startblock;
236 	if (d)
237 		return d;
238 
239 	x = be64_to_cpu(kp->rm_owner);
240 	y = rec->rm_owner;
241 	if (x > y)
242 		return 1;
243 	else if (y > x)
244 		return -1;
245 
246 	x = XFS_RMAP_OFF(be64_to_cpu(kp->rm_offset));
247 	y = rec->rm_offset;
248 	if (x > y)
249 		return 1;
250 	else if (y > x)
251 		return -1;
252 	return 0;
253 }
254 
255 STATIC int64_t
256 xfs_rmapbt_diff_two_keys(
257 	struct xfs_btree_cur	*cur,
258 	union xfs_btree_key	*k1,
259 	union xfs_btree_key	*k2)
260 {
261 	struct xfs_rmap_key	*kp1 = &k1->rmap;
262 	struct xfs_rmap_key	*kp2 = &k2->rmap;
263 	int64_t			d;
264 	__u64			x, y;
265 
266 	d = (int64_t)be32_to_cpu(kp1->rm_startblock) -
267 		       be32_to_cpu(kp2->rm_startblock);
268 	if (d)
269 		return d;
270 
271 	x = be64_to_cpu(kp1->rm_owner);
272 	y = be64_to_cpu(kp2->rm_owner);
273 	if (x > y)
274 		return 1;
275 	else if (y > x)
276 		return -1;
277 
278 	x = XFS_RMAP_OFF(be64_to_cpu(kp1->rm_offset));
279 	y = XFS_RMAP_OFF(be64_to_cpu(kp2->rm_offset));
280 	if (x > y)
281 		return 1;
282 	else if (y > x)
283 		return -1;
284 	return 0;
285 }
286 
287 static xfs_failaddr_t
288 xfs_rmapbt_verify(
289 	struct xfs_buf		*bp)
290 {
291 	struct xfs_mount	*mp = bp->b_mount;
292 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
293 	struct xfs_perag	*pag = bp->b_pag;
294 	xfs_failaddr_t		fa;
295 	unsigned int		level;
296 
297 	/*
298 	 * magic number and level verification
299 	 *
300 	 * During growfs operations, we can't verify the exact level or owner as
301 	 * the perag is not fully initialised and hence not attached to the
302 	 * buffer.  In this case, check against the maximum tree depth.
303 	 *
304 	 * Similarly, during log recovery we will have a perag structure
305 	 * attached, but the agf information will not yet have been initialised
306 	 * from the on disk AGF. Again, we can only check against maximum limits
307 	 * in this case.
308 	 */
309 	if (!xfs_verify_magic(bp, block->bb_magic))
310 		return __this_address;
311 
312 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
313 		return __this_address;
314 	fa = xfs_btree_sblock_v5hdr_verify(bp);
315 	if (fa)
316 		return fa;
317 
318 	level = be16_to_cpu(block->bb_level);
319 	if (pag && pag->pagf_init) {
320 		if (level >= pag->pagf_levels[XFS_BTNUM_RMAPi])
321 			return __this_address;
322 	} else if (level >= mp->m_rmap_maxlevels)
323 		return __this_address;
324 
325 	return xfs_btree_sblock_verify(bp, mp->m_rmap_mxr[level != 0]);
326 }
327 
328 static void
329 xfs_rmapbt_read_verify(
330 	struct xfs_buf	*bp)
331 {
332 	xfs_failaddr_t	fa;
333 
334 	if (!xfs_btree_sblock_verify_crc(bp))
335 		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
336 	else {
337 		fa = xfs_rmapbt_verify(bp);
338 		if (fa)
339 			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
340 	}
341 
342 	if (bp->b_error)
343 		trace_xfs_btree_corrupt(bp, _RET_IP_);
344 }
345 
346 static void
347 xfs_rmapbt_write_verify(
348 	struct xfs_buf	*bp)
349 {
350 	xfs_failaddr_t	fa;
351 
352 	fa = xfs_rmapbt_verify(bp);
353 	if (fa) {
354 		trace_xfs_btree_corrupt(bp, _RET_IP_);
355 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
356 		return;
357 	}
358 	xfs_btree_sblock_calc_crc(bp);
359 
360 }
361 
362 const struct xfs_buf_ops xfs_rmapbt_buf_ops = {
363 	.name			= "xfs_rmapbt",
364 	.magic			= { 0, cpu_to_be32(XFS_RMAP_CRC_MAGIC) },
365 	.verify_read		= xfs_rmapbt_read_verify,
366 	.verify_write		= xfs_rmapbt_write_verify,
367 	.verify_struct		= xfs_rmapbt_verify,
368 };
369 
370 STATIC int
371 xfs_rmapbt_keys_inorder(
372 	struct xfs_btree_cur	*cur,
373 	union xfs_btree_key	*k1,
374 	union xfs_btree_key	*k2)
375 {
376 	uint32_t		x;
377 	uint32_t		y;
378 	uint64_t		a;
379 	uint64_t		b;
380 
381 	x = be32_to_cpu(k1->rmap.rm_startblock);
382 	y = be32_to_cpu(k2->rmap.rm_startblock);
383 	if (x < y)
384 		return 1;
385 	else if (x > y)
386 		return 0;
387 	a = be64_to_cpu(k1->rmap.rm_owner);
388 	b = be64_to_cpu(k2->rmap.rm_owner);
389 	if (a < b)
390 		return 1;
391 	else if (a > b)
392 		return 0;
393 	a = XFS_RMAP_OFF(be64_to_cpu(k1->rmap.rm_offset));
394 	b = XFS_RMAP_OFF(be64_to_cpu(k2->rmap.rm_offset));
395 	if (a <= b)
396 		return 1;
397 	return 0;
398 }
399 
400 STATIC int
401 xfs_rmapbt_recs_inorder(
402 	struct xfs_btree_cur	*cur,
403 	union xfs_btree_rec	*r1,
404 	union xfs_btree_rec	*r2)
405 {
406 	uint32_t		x;
407 	uint32_t		y;
408 	uint64_t		a;
409 	uint64_t		b;
410 
411 	x = be32_to_cpu(r1->rmap.rm_startblock);
412 	y = be32_to_cpu(r2->rmap.rm_startblock);
413 	if (x < y)
414 		return 1;
415 	else if (x > y)
416 		return 0;
417 	a = be64_to_cpu(r1->rmap.rm_owner);
418 	b = be64_to_cpu(r2->rmap.rm_owner);
419 	if (a < b)
420 		return 1;
421 	else if (a > b)
422 		return 0;
423 	a = XFS_RMAP_OFF(be64_to_cpu(r1->rmap.rm_offset));
424 	b = XFS_RMAP_OFF(be64_to_cpu(r2->rmap.rm_offset));
425 	if (a <= b)
426 		return 1;
427 	return 0;
428 }
429 
430 static const struct xfs_btree_ops xfs_rmapbt_ops = {
431 	.rec_len		= sizeof(struct xfs_rmap_rec),
432 	.key_len		= 2 * sizeof(struct xfs_rmap_key),
433 
434 	.dup_cursor		= xfs_rmapbt_dup_cursor,
435 	.set_root		= xfs_rmapbt_set_root,
436 	.alloc_block		= xfs_rmapbt_alloc_block,
437 	.free_block		= xfs_rmapbt_free_block,
438 	.get_minrecs		= xfs_rmapbt_get_minrecs,
439 	.get_maxrecs		= xfs_rmapbt_get_maxrecs,
440 	.init_key_from_rec	= xfs_rmapbt_init_key_from_rec,
441 	.init_high_key_from_rec	= xfs_rmapbt_init_high_key_from_rec,
442 	.init_rec_from_cur	= xfs_rmapbt_init_rec_from_cur,
443 	.init_ptr_from_cur	= xfs_rmapbt_init_ptr_from_cur,
444 	.key_diff		= xfs_rmapbt_key_diff,
445 	.buf_ops		= &xfs_rmapbt_buf_ops,
446 	.diff_two_keys		= xfs_rmapbt_diff_two_keys,
447 	.keys_inorder		= xfs_rmapbt_keys_inorder,
448 	.recs_inorder		= xfs_rmapbt_recs_inorder,
449 };
450 
451 /*
452  * Allocate a new allocation btree cursor.
453  */
454 struct xfs_btree_cur *
455 xfs_rmapbt_init_cursor(
456 	struct xfs_mount	*mp,
457 	struct xfs_trans	*tp,
458 	struct xfs_buf		*agbp,
459 	xfs_agnumber_t		agno)
460 {
461 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
462 	struct xfs_btree_cur	*cur;
463 
464 	cur = kmem_zone_zalloc(xfs_btree_cur_zone, KM_NOFS);
465 	cur->bc_tp = tp;
466 	cur->bc_mp = mp;
467 	/* Overlapping btree; 2 keys per pointer. */
468 	cur->bc_btnum = XFS_BTNUM_RMAP;
469 	cur->bc_flags = XFS_BTREE_CRC_BLOCKS | XFS_BTREE_OVERLAPPING;
470 	cur->bc_blocklog = mp->m_sb.sb_blocklog;
471 	cur->bc_ops = &xfs_rmapbt_ops;
472 	cur->bc_nlevels = be32_to_cpu(agf->agf_levels[XFS_BTNUM_RMAP]);
473 	cur->bc_statoff = XFS_STATS_CALC_INDEX(xs_rmap_2);
474 
475 	cur->bc_private.a.agbp = agbp;
476 	cur->bc_private.a.agno = agno;
477 
478 	return cur;
479 }
480 
481 /*
482  * Calculate number of records in an rmap btree block.
483  */
484 int
485 xfs_rmapbt_maxrecs(
486 	int			blocklen,
487 	int			leaf)
488 {
489 	blocklen -= XFS_RMAP_BLOCK_LEN;
490 
491 	if (leaf)
492 		return blocklen / sizeof(struct xfs_rmap_rec);
493 	return blocklen /
494 		(2 * sizeof(struct xfs_rmap_key) + sizeof(xfs_rmap_ptr_t));
495 }
496 
497 /* Compute the maximum height of an rmap btree. */
498 void
499 xfs_rmapbt_compute_maxlevels(
500 	struct xfs_mount		*mp)
501 {
502 	/*
503 	 * On a non-reflink filesystem, the maximum number of rmap
504 	 * records is the number of blocks in the AG, hence the max
505 	 * rmapbt height is log_$maxrecs($agblocks).  However, with
506 	 * reflink each AG block can have up to 2^32 (per the refcount
507 	 * record format) owners, which means that theoretically we
508 	 * could face up to 2^64 rmap records.
509 	 *
510 	 * That effectively means that the max rmapbt height must be
511 	 * XFS_BTREE_MAXLEVELS.  "Fortunately" we'll run out of AG
512 	 * blocks to feed the rmapbt long before the rmapbt reaches
513 	 * maximum height.  The reflink code uses ag_resv_critical to
514 	 * disallow reflinking when less than 10% of the per-AG metadata
515 	 * block reservation since the fallback is a regular file copy.
516 	 */
517 	if (xfs_sb_version_hasreflink(&mp->m_sb))
518 		mp->m_rmap_maxlevels = XFS_BTREE_MAXLEVELS;
519 	else
520 		mp->m_rmap_maxlevels = xfs_btree_compute_maxlevels(
521 				mp->m_rmap_mnr, mp->m_sb.sb_agblocks);
522 }
523 
524 /* Calculate the refcount btree size for some records. */
525 xfs_extlen_t
526 xfs_rmapbt_calc_size(
527 	struct xfs_mount	*mp,
528 	unsigned long long	len)
529 {
530 	return xfs_btree_calc_size(mp->m_rmap_mnr, len);
531 }
532 
533 /*
534  * Calculate the maximum refcount btree size.
535  */
536 xfs_extlen_t
537 xfs_rmapbt_max_size(
538 	struct xfs_mount	*mp,
539 	xfs_agblock_t		agblocks)
540 {
541 	/* Bail out if we're uninitialized, which can happen in mkfs. */
542 	if (mp->m_rmap_mxr[0] == 0)
543 		return 0;
544 
545 	return xfs_rmapbt_calc_size(mp, agblocks);
546 }
547 
548 /*
549  * Figure out how many blocks to reserve and how many are used by this btree.
550  */
551 int
552 xfs_rmapbt_calc_reserves(
553 	struct xfs_mount	*mp,
554 	struct xfs_trans	*tp,
555 	xfs_agnumber_t		agno,
556 	xfs_extlen_t		*ask,
557 	xfs_extlen_t		*used)
558 {
559 	struct xfs_buf		*agbp;
560 	struct xfs_agf		*agf;
561 	xfs_agblock_t		agblocks;
562 	xfs_extlen_t		tree_len;
563 	int			error;
564 
565 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
566 		return 0;
567 
568 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
569 	if (error)
570 		return error;
571 
572 	agf = XFS_BUF_TO_AGF(agbp);
573 	agblocks = be32_to_cpu(agf->agf_length);
574 	tree_len = be32_to_cpu(agf->agf_rmap_blocks);
575 	xfs_trans_brelse(tp, agbp);
576 
577 	/*
578 	 * The log is permanently allocated, so the space it occupies will
579 	 * never be available for the kinds of things that would require btree
580 	 * expansion.  We therefore can pretend the space isn't there.
581 	 */
582 	if (mp->m_sb.sb_logstart &&
583 	    XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == agno)
584 		agblocks -= mp->m_sb.sb_logblocks;
585 
586 	/* Reserve 1% of the AG or enough for 1 block per record. */
587 	*ask += max(agblocks / 100, xfs_rmapbt_max_size(mp, agblocks));
588 	*used += tree_len;
589 
590 	return error;
591 }
592