Loading...
Note: File does not exist in v3.15.
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Copyright (C) 2016 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_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_defer.h"
14#include "xfs_btree.h"
15#include "xfs_bmap.h"
16#include "xfs_refcount_btree.h"
17#include "xfs_alloc.h"
18#include "xfs_errortag.h"
19#include "xfs_error.h"
20#include "xfs_trace.h"
21#include "xfs_trans.h"
22#include "xfs_bit.h"
23#include "xfs_refcount.h"
24#include "xfs_rmap.h"
25#include "xfs_ag.h"
26
27/* Allowable refcount adjustment amounts. */
28enum xfs_refc_adjust_op {
29 XFS_REFCOUNT_ADJUST_INCREASE = 1,
30 XFS_REFCOUNT_ADJUST_DECREASE = -1,
31 XFS_REFCOUNT_ADJUST_COW_ALLOC = 0,
32 XFS_REFCOUNT_ADJUST_COW_FREE = -1,
33};
34
35STATIC int __xfs_refcount_cow_alloc(struct xfs_btree_cur *rcur,
36 xfs_agblock_t agbno, xfs_extlen_t aglen);
37STATIC int __xfs_refcount_cow_free(struct xfs_btree_cur *rcur,
38 xfs_agblock_t agbno, xfs_extlen_t aglen);
39
40/*
41 * Look up the first record less than or equal to [bno, len] in the btree
42 * given by cur.
43 */
44int
45xfs_refcount_lookup_le(
46 struct xfs_btree_cur *cur,
47 xfs_agblock_t bno,
48 int *stat)
49{
50 trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_ag.pag->pag_agno, bno,
51 XFS_LOOKUP_LE);
52 cur->bc_rec.rc.rc_startblock = bno;
53 cur->bc_rec.rc.rc_blockcount = 0;
54 return xfs_btree_lookup(cur, XFS_LOOKUP_LE, stat);
55}
56
57/*
58 * Look up the first record greater than or equal to [bno, len] in the btree
59 * given by cur.
60 */
61int
62xfs_refcount_lookup_ge(
63 struct xfs_btree_cur *cur,
64 xfs_agblock_t bno,
65 int *stat)
66{
67 trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_ag.pag->pag_agno, bno,
68 XFS_LOOKUP_GE);
69 cur->bc_rec.rc.rc_startblock = bno;
70 cur->bc_rec.rc.rc_blockcount = 0;
71 return xfs_btree_lookup(cur, XFS_LOOKUP_GE, stat);
72}
73
74/*
75 * Look up the first record equal to [bno, len] in the btree
76 * given by cur.
77 */
78int
79xfs_refcount_lookup_eq(
80 struct xfs_btree_cur *cur,
81 xfs_agblock_t bno,
82 int *stat)
83{
84 trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_ag.pag->pag_agno, bno,
85 XFS_LOOKUP_LE);
86 cur->bc_rec.rc.rc_startblock = bno;
87 cur->bc_rec.rc.rc_blockcount = 0;
88 return xfs_btree_lookup(cur, XFS_LOOKUP_EQ, stat);
89}
90
91/* Convert on-disk record to in-core format. */
92void
93xfs_refcount_btrec_to_irec(
94 union xfs_btree_rec *rec,
95 struct xfs_refcount_irec *irec)
96{
97 irec->rc_startblock = be32_to_cpu(rec->refc.rc_startblock);
98 irec->rc_blockcount = be32_to_cpu(rec->refc.rc_blockcount);
99 irec->rc_refcount = be32_to_cpu(rec->refc.rc_refcount);
100}
101
102/*
103 * Get the data from the pointed-to record.
104 */
105int
106xfs_refcount_get_rec(
107 struct xfs_btree_cur *cur,
108 struct xfs_refcount_irec *irec,
109 int *stat)
110{
111 struct xfs_mount *mp = cur->bc_mp;
112 xfs_agnumber_t agno = cur->bc_ag.pag->pag_agno;
113 union xfs_btree_rec *rec;
114 int error;
115 xfs_agblock_t realstart;
116
117 error = xfs_btree_get_rec(cur, &rec, stat);
118 if (error || !*stat)
119 return error;
120
121 xfs_refcount_btrec_to_irec(rec, irec);
122
123 agno = cur->bc_ag.pag->pag_agno;
124 if (irec->rc_blockcount == 0 || irec->rc_blockcount > MAXREFCEXTLEN)
125 goto out_bad_rec;
126
127 /* handle special COW-staging state */
128 realstart = irec->rc_startblock;
129 if (realstart & XFS_REFC_COW_START) {
130 if (irec->rc_refcount != 1)
131 goto out_bad_rec;
132 realstart &= ~XFS_REFC_COW_START;
133 } else if (irec->rc_refcount < 2) {
134 goto out_bad_rec;
135 }
136
137 /* check for valid extent range, including overflow */
138 if (!xfs_verify_agbno(mp, agno, realstart))
139 goto out_bad_rec;
140 if (realstart > realstart + irec->rc_blockcount)
141 goto out_bad_rec;
142 if (!xfs_verify_agbno(mp, agno, realstart + irec->rc_blockcount - 1))
143 goto out_bad_rec;
144
145 if (irec->rc_refcount == 0 || irec->rc_refcount > MAXREFCOUNT)
146 goto out_bad_rec;
147
148 trace_xfs_refcount_get(cur->bc_mp, cur->bc_ag.pag->pag_agno, irec);
149 return 0;
150
151out_bad_rec:
152 xfs_warn(mp,
153 "Refcount BTree record corruption in AG %d detected!", agno);
154 xfs_warn(mp,
155 "Start block 0x%x, block count 0x%x, references 0x%x",
156 irec->rc_startblock, irec->rc_blockcount, irec->rc_refcount);
157 return -EFSCORRUPTED;
158}
159
160/*
161 * Update the record referred to by cur to the value given
162 * by [bno, len, refcount].
163 * This either works (return 0) or gets an EFSCORRUPTED error.
164 */
165STATIC int
166xfs_refcount_update(
167 struct xfs_btree_cur *cur,
168 struct xfs_refcount_irec *irec)
169{
170 union xfs_btree_rec rec;
171 int error;
172
173 trace_xfs_refcount_update(cur->bc_mp, cur->bc_ag.pag->pag_agno, irec);
174 rec.refc.rc_startblock = cpu_to_be32(irec->rc_startblock);
175 rec.refc.rc_blockcount = cpu_to_be32(irec->rc_blockcount);
176 rec.refc.rc_refcount = cpu_to_be32(irec->rc_refcount);
177 error = xfs_btree_update(cur, &rec);
178 if (error)
179 trace_xfs_refcount_update_error(cur->bc_mp,
180 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
181 return error;
182}
183
184/*
185 * Insert the record referred to by cur to the value given
186 * by [bno, len, refcount].
187 * This either works (return 0) or gets an EFSCORRUPTED error.
188 */
189int
190xfs_refcount_insert(
191 struct xfs_btree_cur *cur,
192 struct xfs_refcount_irec *irec,
193 int *i)
194{
195 int error;
196
197 trace_xfs_refcount_insert(cur->bc_mp, cur->bc_ag.pag->pag_agno, irec);
198 cur->bc_rec.rc.rc_startblock = irec->rc_startblock;
199 cur->bc_rec.rc.rc_blockcount = irec->rc_blockcount;
200 cur->bc_rec.rc.rc_refcount = irec->rc_refcount;
201 error = xfs_btree_insert(cur, i);
202 if (error)
203 goto out_error;
204 if (XFS_IS_CORRUPT(cur->bc_mp, *i != 1)) {
205 error = -EFSCORRUPTED;
206 goto out_error;
207 }
208
209out_error:
210 if (error)
211 trace_xfs_refcount_insert_error(cur->bc_mp,
212 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
213 return error;
214}
215
216/*
217 * Remove the record referred to by cur, then set the pointer to the spot
218 * where the record could be re-inserted, in case we want to increment or
219 * decrement the cursor.
220 * This either works (return 0) or gets an EFSCORRUPTED error.
221 */
222STATIC int
223xfs_refcount_delete(
224 struct xfs_btree_cur *cur,
225 int *i)
226{
227 struct xfs_refcount_irec irec;
228 int found_rec;
229 int error;
230
231 error = xfs_refcount_get_rec(cur, &irec, &found_rec);
232 if (error)
233 goto out_error;
234 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
235 error = -EFSCORRUPTED;
236 goto out_error;
237 }
238 trace_xfs_refcount_delete(cur->bc_mp, cur->bc_ag.pag->pag_agno, &irec);
239 error = xfs_btree_delete(cur, i);
240 if (XFS_IS_CORRUPT(cur->bc_mp, *i != 1)) {
241 error = -EFSCORRUPTED;
242 goto out_error;
243 }
244 if (error)
245 goto out_error;
246 error = xfs_refcount_lookup_ge(cur, irec.rc_startblock, &found_rec);
247out_error:
248 if (error)
249 trace_xfs_refcount_delete_error(cur->bc_mp,
250 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
251 return error;
252}
253
254/*
255 * Adjusting the Reference Count
256 *
257 * As stated elsewhere, the reference count btree (refcbt) stores
258 * >1 reference counts for extents of physical blocks. In this
259 * operation, we're either raising or lowering the reference count of
260 * some subrange stored in the tree:
261 *
262 * <------ adjustment range ------>
263 * ----+ +---+-----+ +--+--------+---------
264 * 2 | | 3 | 4 | |17| 55 | 10
265 * ----+ +---+-----+ +--+--------+---------
266 * X axis is physical blocks number;
267 * reference counts are the numbers inside the rectangles
268 *
269 * The first thing we need to do is to ensure that there are no
270 * refcount extents crossing either boundary of the range to be
271 * adjusted. For any extent that does cross a boundary, split it into
272 * two extents so that we can increment the refcount of one of the
273 * pieces later:
274 *
275 * <------ adjustment range ------>
276 * ----+ +---+-----+ +--+--------+----+----
277 * 2 | | 3 | 2 | |17| 55 | 10 | 10
278 * ----+ +---+-----+ +--+--------+----+----
279 *
280 * For this next step, let's assume that all the physical blocks in
281 * the adjustment range are mapped to a file and are therefore in use
282 * at least once. Therefore, we can infer that any gap in the
283 * refcount tree within the adjustment range represents a physical
284 * extent with refcount == 1:
285 *
286 * <------ adjustment range ------>
287 * ----+---+---+-----+-+--+--------+----+----
288 * 2 |"1"| 3 | 2 |1|17| 55 | 10 | 10
289 * ----+---+---+-----+-+--+--------+----+----
290 * ^
291 *
292 * For each extent that falls within the interval range, figure out
293 * which extent is to the left or the right of that extent. Now we
294 * have a left, current, and right extent. If the new reference count
295 * of the center extent enables us to merge left, center, and right
296 * into one record covering all three, do so. If the center extent is
297 * at the left end of the range, abuts the left extent, and its new
298 * reference count matches the left extent's record, then merge them.
299 * If the center extent is at the right end of the range, abuts the
300 * right extent, and the reference counts match, merge those. In the
301 * example, we can left merge (assuming an increment operation):
302 *
303 * <------ adjustment range ------>
304 * --------+---+-----+-+--+--------+----+----
305 * 2 | 3 | 2 |1|17| 55 | 10 | 10
306 * --------+---+-----+-+--+--------+----+----
307 * ^
308 *
309 * For all other extents within the range, adjust the reference count
310 * or delete it if the refcount falls below 2. If we were
311 * incrementing, the end result looks like this:
312 *
313 * <------ adjustment range ------>
314 * --------+---+-----+-+--+--------+----+----
315 * 2 | 4 | 3 |2|18| 56 | 11 | 10
316 * --------+---+-----+-+--+--------+----+----
317 *
318 * The result of a decrement operation looks as such:
319 *
320 * <------ adjustment range ------>
321 * ----+ +---+ +--+--------+----+----
322 * 2 | | 2 | |16| 54 | 9 | 10
323 * ----+ +---+ +--+--------+----+----
324 * DDDD 111111DD
325 *
326 * The blocks marked "D" are freed; the blocks marked "1" are only
327 * referenced once and therefore the record is removed from the
328 * refcount btree.
329 */
330
331/* Next block after this extent. */
332static inline xfs_agblock_t
333xfs_refc_next(
334 struct xfs_refcount_irec *rc)
335{
336 return rc->rc_startblock + rc->rc_blockcount;
337}
338
339/*
340 * Split a refcount extent that crosses agbno.
341 */
342STATIC int
343xfs_refcount_split_extent(
344 struct xfs_btree_cur *cur,
345 xfs_agblock_t agbno,
346 bool *shape_changed)
347{
348 struct xfs_refcount_irec rcext, tmp;
349 int found_rec;
350 int error;
351
352 *shape_changed = false;
353 error = xfs_refcount_lookup_le(cur, agbno, &found_rec);
354 if (error)
355 goto out_error;
356 if (!found_rec)
357 return 0;
358
359 error = xfs_refcount_get_rec(cur, &rcext, &found_rec);
360 if (error)
361 goto out_error;
362 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
363 error = -EFSCORRUPTED;
364 goto out_error;
365 }
366 if (rcext.rc_startblock == agbno || xfs_refc_next(&rcext) <= agbno)
367 return 0;
368
369 *shape_changed = true;
370 trace_xfs_refcount_split_extent(cur->bc_mp, cur->bc_ag.pag->pag_agno,
371 &rcext, agbno);
372
373 /* Establish the right extent. */
374 tmp = rcext;
375 tmp.rc_startblock = agbno;
376 tmp.rc_blockcount -= (agbno - rcext.rc_startblock);
377 error = xfs_refcount_update(cur, &tmp);
378 if (error)
379 goto out_error;
380
381 /* Insert the left extent. */
382 tmp = rcext;
383 tmp.rc_blockcount = agbno - rcext.rc_startblock;
384 error = xfs_refcount_insert(cur, &tmp, &found_rec);
385 if (error)
386 goto out_error;
387 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
388 error = -EFSCORRUPTED;
389 goto out_error;
390 }
391 return error;
392
393out_error:
394 trace_xfs_refcount_split_extent_error(cur->bc_mp,
395 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
396 return error;
397}
398
399/*
400 * Merge the left, center, and right extents.
401 */
402STATIC int
403xfs_refcount_merge_center_extents(
404 struct xfs_btree_cur *cur,
405 struct xfs_refcount_irec *left,
406 struct xfs_refcount_irec *center,
407 struct xfs_refcount_irec *right,
408 unsigned long long extlen,
409 xfs_extlen_t *aglen)
410{
411 int error;
412 int found_rec;
413
414 trace_xfs_refcount_merge_center_extents(cur->bc_mp,
415 cur->bc_ag.pag->pag_agno, left, center, right);
416
417 /*
418 * Make sure the center and right extents are not in the btree.
419 * If the center extent was synthesized, the first delete call
420 * removes the right extent and we skip the second deletion.
421 * If center and right were in the btree, then the first delete
422 * call removes the center and the second one removes the right
423 * extent.
424 */
425 error = xfs_refcount_lookup_ge(cur, center->rc_startblock,
426 &found_rec);
427 if (error)
428 goto out_error;
429 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
430 error = -EFSCORRUPTED;
431 goto out_error;
432 }
433
434 error = xfs_refcount_delete(cur, &found_rec);
435 if (error)
436 goto out_error;
437 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
438 error = -EFSCORRUPTED;
439 goto out_error;
440 }
441
442 if (center->rc_refcount > 1) {
443 error = xfs_refcount_delete(cur, &found_rec);
444 if (error)
445 goto out_error;
446 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
447 error = -EFSCORRUPTED;
448 goto out_error;
449 }
450 }
451
452 /* Enlarge the left extent. */
453 error = xfs_refcount_lookup_le(cur, left->rc_startblock,
454 &found_rec);
455 if (error)
456 goto out_error;
457 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
458 error = -EFSCORRUPTED;
459 goto out_error;
460 }
461
462 left->rc_blockcount = extlen;
463 error = xfs_refcount_update(cur, left);
464 if (error)
465 goto out_error;
466
467 *aglen = 0;
468 return error;
469
470out_error:
471 trace_xfs_refcount_merge_center_extents_error(cur->bc_mp,
472 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
473 return error;
474}
475
476/*
477 * Merge with the left extent.
478 */
479STATIC int
480xfs_refcount_merge_left_extent(
481 struct xfs_btree_cur *cur,
482 struct xfs_refcount_irec *left,
483 struct xfs_refcount_irec *cleft,
484 xfs_agblock_t *agbno,
485 xfs_extlen_t *aglen)
486{
487 int error;
488 int found_rec;
489
490 trace_xfs_refcount_merge_left_extent(cur->bc_mp,
491 cur->bc_ag.pag->pag_agno, left, cleft);
492
493 /* If the extent at agbno (cleft) wasn't synthesized, remove it. */
494 if (cleft->rc_refcount > 1) {
495 error = xfs_refcount_lookup_le(cur, cleft->rc_startblock,
496 &found_rec);
497 if (error)
498 goto out_error;
499 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
500 error = -EFSCORRUPTED;
501 goto out_error;
502 }
503
504 error = xfs_refcount_delete(cur, &found_rec);
505 if (error)
506 goto out_error;
507 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
508 error = -EFSCORRUPTED;
509 goto out_error;
510 }
511 }
512
513 /* Enlarge the left extent. */
514 error = xfs_refcount_lookup_le(cur, left->rc_startblock,
515 &found_rec);
516 if (error)
517 goto out_error;
518 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
519 error = -EFSCORRUPTED;
520 goto out_error;
521 }
522
523 left->rc_blockcount += cleft->rc_blockcount;
524 error = xfs_refcount_update(cur, left);
525 if (error)
526 goto out_error;
527
528 *agbno += cleft->rc_blockcount;
529 *aglen -= cleft->rc_blockcount;
530 return error;
531
532out_error:
533 trace_xfs_refcount_merge_left_extent_error(cur->bc_mp,
534 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
535 return error;
536}
537
538/*
539 * Merge with the right extent.
540 */
541STATIC int
542xfs_refcount_merge_right_extent(
543 struct xfs_btree_cur *cur,
544 struct xfs_refcount_irec *right,
545 struct xfs_refcount_irec *cright,
546 xfs_extlen_t *aglen)
547{
548 int error;
549 int found_rec;
550
551 trace_xfs_refcount_merge_right_extent(cur->bc_mp,
552 cur->bc_ag.pag->pag_agno, cright, right);
553
554 /*
555 * If the extent ending at agbno+aglen (cright) wasn't synthesized,
556 * remove it.
557 */
558 if (cright->rc_refcount > 1) {
559 error = xfs_refcount_lookup_le(cur, cright->rc_startblock,
560 &found_rec);
561 if (error)
562 goto out_error;
563 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
564 error = -EFSCORRUPTED;
565 goto out_error;
566 }
567
568 error = xfs_refcount_delete(cur, &found_rec);
569 if (error)
570 goto out_error;
571 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
572 error = -EFSCORRUPTED;
573 goto out_error;
574 }
575 }
576
577 /* Enlarge the right extent. */
578 error = xfs_refcount_lookup_le(cur, right->rc_startblock,
579 &found_rec);
580 if (error)
581 goto out_error;
582 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
583 error = -EFSCORRUPTED;
584 goto out_error;
585 }
586
587 right->rc_startblock -= cright->rc_blockcount;
588 right->rc_blockcount += cright->rc_blockcount;
589 error = xfs_refcount_update(cur, right);
590 if (error)
591 goto out_error;
592
593 *aglen -= cright->rc_blockcount;
594 return error;
595
596out_error:
597 trace_xfs_refcount_merge_right_extent_error(cur->bc_mp,
598 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
599 return error;
600}
601
602#define XFS_FIND_RCEXT_SHARED 1
603#define XFS_FIND_RCEXT_COW 2
604/*
605 * Find the left extent and the one after it (cleft). This function assumes
606 * that we've already split any extent crossing agbno.
607 */
608STATIC int
609xfs_refcount_find_left_extents(
610 struct xfs_btree_cur *cur,
611 struct xfs_refcount_irec *left,
612 struct xfs_refcount_irec *cleft,
613 xfs_agblock_t agbno,
614 xfs_extlen_t aglen,
615 int flags)
616{
617 struct xfs_refcount_irec tmp;
618 int error;
619 int found_rec;
620
621 left->rc_startblock = cleft->rc_startblock = NULLAGBLOCK;
622 error = xfs_refcount_lookup_le(cur, agbno - 1, &found_rec);
623 if (error)
624 goto out_error;
625 if (!found_rec)
626 return 0;
627
628 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
629 if (error)
630 goto out_error;
631 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
632 error = -EFSCORRUPTED;
633 goto out_error;
634 }
635
636 if (xfs_refc_next(&tmp) != agbno)
637 return 0;
638 if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
639 return 0;
640 if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
641 return 0;
642 /* We have a left extent; retrieve (or invent) the next right one */
643 *left = tmp;
644
645 error = xfs_btree_increment(cur, 0, &found_rec);
646 if (error)
647 goto out_error;
648 if (found_rec) {
649 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
650 if (error)
651 goto out_error;
652 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
653 error = -EFSCORRUPTED;
654 goto out_error;
655 }
656
657 /* if tmp starts at the end of our range, just use that */
658 if (tmp.rc_startblock == agbno)
659 *cleft = tmp;
660 else {
661 /*
662 * There's a gap in the refcntbt at the start of the
663 * range we're interested in (refcount == 1) so
664 * synthesize the implied extent and pass it back.
665 * We assume here that the agbno/aglen range was
666 * passed in from a data fork extent mapping and
667 * therefore is allocated to exactly one owner.
668 */
669 cleft->rc_startblock = agbno;
670 cleft->rc_blockcount = min(aglen,
671 tmp.rc_startblock - agbno);
672 cleft->rc_refcount = 1;
673 }
674 } else {
675 /*
676 * No extents, so pretend that there's one covering the whole
677 * range.
678 */
679 cleft->rc_startblock = agbno;
680 cleft->rc_blockcount = aglen;
681 cleft->rc_refcount = 1;
682 }
683 trace_xfs_refcount_find_left_extent(cur->bc_mp, cur->bc_ag.pag->pag_agno,
684 left, cleft, agbno);
685 return error;
686
687out_error:
688 trace_xfs_refcount_find_left_extent_error(cur->bc_mp,
689 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
690 return error;
691}
692
693/*
694 * Find the right extent and the one before it (cright). This function
695 * assumes that we've already split any extents crossing agbno + aglen.
696 */
697STATIC int
698xfs_refcount_find_right_extents(
699 struct xfs_btree_cur *cur,
700 struct xfs_refcount_irec *right,
701 struct xfs_refcount_irec *cright,
702 xfs_agblock_t agbno,
703 xfs_extlen_t aglen,
704 int flags)
705{
706 struct xfs_refcount_irec tmp;
707 int error;
708 int found_rec;
709
710 right->rc_startblock = cright->rc_startblock = NULLAGBLOCK;
711 error = xfs_refcount_lookup_ge(cur, agbno + aglen, &found_rec);
712 if (error)
713 goto out_error;
714 if (!found_rec)
715 return 0;
716
717 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
718 if (error)
719 goto out_error;
720 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
721 error = -EFSCORRUPTED;
722 goto out_error;
723 }
724
725 if (tmp.rc_startblock != agbno + aglen)
726 return 0;
727 if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
728 return 0;
729 if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
730 return 0;
731 /* We have a right extent; retrieve (or invent) the next left one */
732 *right = tmp;
733
734 error = xfs_btree_decrement(cur, 0, &found_rec);
735 if (error)
736 goto out_error;
737 if (found_rec) {
738 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
739 if (error)
740 goto out_error;
741 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
742 error = -EFSCORRUPTED;
743 goto out_error;
744 }
745
746 /* if tmp ends at the end of our range, just use that */
747 if (xfs_refc_next(&tmp) == agbno + aglen)
748 *cright = tmp;
749 else {
750 /*
751 * There's a gap in the refcntbt at the end of the
752 * range we're interested in (refcount == 1) so
753 * create the implied extent and pass it back.
754 * We assume here that the agbno/aglen range was
755 * passed in from a data fork extent mapping and
756 * therefore is allocated to exactly one owner.
757 */
758 cright->rc_startblock = max(agbno, xfs_refc_next(&tmp));
759 cright->rc_blockcount = right->rc_startblock -
760 cright->rc_startblock;
761 cright->rc_refcount = 1;
762 }
763 } else {
764 /*
765 * No extents, so pretend that there's one covering the whole
766 * range.
767 */
768 cright->rc_startblock = agbno;
769 cright->rc_blockcount = aglen;
770 cright->rc_refcount = 1;
771 }
772 trace_xfs_refcount_find_right_extent(cur->bc_mp, cur->bc_ag.pag->pag_agno,
773 cright, right, agbno + aglen);
774 return error;
775
776out_error:
777 trace_xfs_refcount_find_right_extent_error(cur->bc_mp,
778 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
779 return error;
780}
781
782/* Is this extent valid? */
783static inline bool
784xfs_refc_valid(
785 struct xfs_refcount_irec *rc)
786{
787 return rc->rc_startblock != NULLAGBLOCK;
788}
789
790/*
791 * Try to merge with any extents on the boundaries of the adjustment range.
792 */
793STATIC int
794xfs_refcount_merge_extents(
795 struct xfs_btree_cur *cur,
796 xfs_agblock_t *agbno,
797 xfs_extlen_t *aglen,
798 enum xfs_refc_adjust_op adjust,
799 int flags,
800 bool *shape_changed)
801{
802 struct xfs_refcount_irec left = {0}, cleft = {0};
803 struct xfs_refcount_irec cright = {0}, right = {0};
804 int error;
805 unsigned long long ulen;
806 bool cequal;
807
808 *shape_changed = false;
809 /*
810 * Find the extent just below agbno [left], just above agbno [cleft],
811 * just below (agbno + aglen) [cright], and just above (agbno + aglen)
812 * [right].
813 */
814 error = xfs_refcount_find_left_extents(cur, &left, &cleft, *agbno,
815 *aglen, flags);
816 if (error)
817 return error;
818 error = xfs_refcount_find_right_extents(cur, &right, &cright, *agbno,
819 *aglen, flags);
820 if (error)
821 return error;
822
823 /* No left or right extent to merge; exit. */
824 if (!xfs_refc_valid(&left) && !xfs_refc_valid(&right))
825 return 0;
826
827 cequal = (cleft.rc_startblock == cright.rc_startblock) &&
828 (cleft.rc_blockcount == cright.rc_blockcount);
829
830 /* Try to merge left, cleft, and right. cleft must == cright. */
831 ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount +
832 right.rc_blockcount;
833 if (xfs_refc_valid(&left) && xfs_refc_valid(&right) &&
834 xfs_refc_valid(&cleft) && xfs_refc_valid(&cright) && cequal &&
835 left.rc_refcount == cleft.rc_refcount + adjust &&
836 right.rc_refcount == cleft.rc_refcount + adjust &&
837 ulen < MAXREFCEXTLEN) {
838 *shape_changed = true;
839 return xfs_refcount_merge_center_extents(cur, &left, &cleft,
840 &right, ulen, aglen);
841 }
842
843 /* Try to merge left and cleft. */
844 ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount;
845 if (xfs_refc_valid(&left) && xfs_refc_valid(&cleft) &&
846 left.rc_refcount == cleft.rc_refcount + adjust &&
847 ulen < MAXREFCEXTLEN) {
848 *shape_changed = true;
849 error = xfs_refcount_merge_left_extent(cur, &left, &cleft,
850 agbno, aglen);
851 if (error)
852 return error;
853
854 /*
855 * If we just merged left + cleft and cleft == cright,
856 * we no longer have a cright to merge with right. We're done.
857 */
858 if (cequal)
859 return 0;
860 }
861
862 /* Try to merge cright and right. */
863 ulen = (unsigned long long)right.rc_blockcount + cright.rc_blockcount;
864 if (xfs_refc_valid(&right) && xfs_refc_valid(&cright) &&
865 right.rc_refcount == cright.rc_refcount + adjust &&
866 ulen < MAXREFCEXTLEN) {
867 *shape_changed = true;
868 return xfs_refcount_merge_right_extent(cur, &right, &cright,
869 aglen);
870 }
871
872 return error;
873}
874
875/*
876 * XXX: This is a pretty hand-wavy estimate. The penalty for guessing
877 * true incorrectly is a shutdown FS; the penalty for guessing false
878 * incorrectly is more transaction rolls than might be necessary.
879 * Be conservative here.
880 */
881static bool
882xfs_refcount_still_have_space(
883 struct xfs_btree_cur *cur)
884{
885 unsigned long overhead;
886
887 overhead = cur->bc_ag.refc.shape_changes *
888 xfs_allocfree_log_count(cur->bc_mp, 1);
889 overhead *= cur->bc_mp->m_sb.sb_blocksize;
890
891 /*
892 * Only allow 2 refcount extent updates per transaction if the
893 * refcount continue update "error" has been injected.
894 */
895 if (cur->bc_ag.refc.nr_ops > 2 &&
896 XFS_TEST_ERROR(false, cur->bc_mp,
897 XFS_ERRTAG_REFCOUNT_CONTINUE_UPDATE))
898 return false;
899
900 if (cur->bc_ag.refc.nr_ops == 0)
901 return true;
902 else if (overhead > cur->bc_tp->t_log_res)
903 return false;
904 return cur->bc_tp->t_log_res - overhead >
905 cur->bc_ag.refc.nr_ops * XFS_REFCOUNT_ITEM_OVERHEAD;
906}
907
908/*
909 * Adjust the refcounts of middle extents. At this point we should have
910 * split extents that crossed the adjustment range; merged with adjacent
911 * extents; and updated agbno/aglen to reflect the merges. Therefore,
912 * all we have to do is update the extents inside [agbno, agbno + aglen].
913 */
914STATIC int
915xfs_refcount_adjust_extents(
916 struct xfs_btree_cur *cur,
917 xfs_agblock_t *agbno,
918 xfs_extlen_t *aglen,
919 enum xfs_refc_adjust_op adj,
920 struct xfs_owner_info *oinfo)
921{
922 struct xfs_refcount_irec ext, tmp;
923 int error;
924 int found_rec, found_tmp;
925 xfs_fsblock_t fsbno;
926
927 /* Merging did all the work already. */
928 if (*aglen == 0)
929 return 0;
930
931 error = xfs_refcount_lookup_ge(cur, *agbno, &found_rec);
932 if (error)
933 goto out_error;
934
935 while (*aglen > 0 && xfs_refcount_still_have_space(cur)) {
936 error = xfs_refcount_get_rec(cur, &ext, &found_rec);
937 if (error)
938 goto out_error;
939 if (!found_rec) {
940 ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks;
941 ext.rc_blockcount = 0;
942 ext.rc_refcount = 0;
943 }
944
945 /*
946 * Deal with a hole in the refcount tree; if a file maps to
947 * these blocks and there's no refcountbt record, pretend that
948 * there is one with refcount == 1.
949 */
950 if (ext.rc_startblock != *agbno) {
951 tmp.rc_startblock = *agbno;
952 tmp.rc_blockcount = min(*aglen,
953 ext.rc_startblock - *agbno);
954 tmp.rc_refcount = 1 + adj;
955 trace_xfs_refcount_modify_extent(cur->bc_mp,
956 cur->bc_ag.pag->pag_agno, &tmp);
957
958 /*
959 * Either cover the hole (increment) or
960 * delete the range (decrement).
961 */
962 if (tmp.rc_refcount) {
963 error = xfs_refcount_insert(cur, &tmp,
964 &found_tmp);
965 if (error)
966 goto out_error;
967 if (XFS_IS_CORRUPT(cur->bc_mp,
968 found_tmp != 1)) {
969 error = -EFSCORRUPTED;
970 goto out_error;
971 }
972 cur->bc_ag.refc.nr_ops++;
973 } else {
974 fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
975 cur->bc_ag.pag->pag_agno,
976 tmp.rc_startblock);
977 xfs_bmap_add_free(cur->bc_tp, fsbno,
978 tmp.rc_blockcount, oinfo);
979 }
980
981 (*agbno) += tmp.rc_blockcount;
982 (*aglen) -= tmp.rc_blockcount;
983
984 error = xfs_refcount_lookup_ge(cur, *agbno,
985 &found_rec);
986 if (error)
987 goto out_error;
988 }
989
990 /* Stop if there's nothing left to modify */
991 if (*aglen == 0 || !xfs_refcount_still_have_space(cur))
992 break;
993
994 /*
995 * Adjust the reference count and either update the tree
996 * (incr) or free the blocks (decr).
997 */
998 if (ext.rc_refcount == MAXREFCOUNT)
999 goto skip;
1000 ext.rc_refcount += adj;
1001 trace_xfs_refcount_modify_extent(cur->bc_mp,
1002 cur->bc_ag.pag->pag_agno, &ext);
1003 if (ext.rc_refcount > 1) {
1004 error = xfs_refcount_update(cur, &ext);
1005 if (error)
1006 goto out_error;
1007 cur->bc_ag.refc.nr_ops++;
1008 } else if (ext.rc_refcount == 1) {
1009 error = xfs_refcount_delete(cur, &found_rec);
1010 if (error)
1011 goto out_error;
1012 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
1013 error = -EFSCORRUPTED;
1014 goto out_error;
1015 }
1016 cur->bc_ag.refc.nr_ops++;
1017 goto advloop;
1018 } else {
1019 fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
1020 cur->bc_ag.pag->pag_agno,
1021 ext.rc_startblock);
1022 xfs_bmap_add_free(cur->bc_tp, fsbno, ext.rc_blockcount,
1023 oinfo);
1024 }
1025
1026skip:
1027 error = xfs_btree_increment(cur, 0, &found_rec);
1028 if (error)
1029 goto out_error;
1030
1031advloop:
1032 (*agbno) += ext.rc_blockcount;
1033 (*aglen) -= ext.rc_blockcount;
1034 }
1035
1036 return error;
1037out_error:
1038 trace_xfs_refcount_modify_extent_error(cur->bc_mp,
1039 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
1040 return error;
1041}
1042
1043/* Adjust the reference count of a range of AG blocks. */
1044STATIC int
1045xfs_refcount_adjust(
1046 struct xfs_btree_cur *cur,
1047 xfs_agblock_t agbno,
1048 xfs_extlen_t aglen,
1049 xfs_agblock_t *new_agbno,
1050 xfs_extlen_t *new_aglen,
1051 enum xfs_refc_adjust_op adj,
1052 struct xfs_owner_info *oinfo)
1053{
1054 bool shape_changed;
1055 int shape_changes = 0;
1056 int error;
1057
1058 *new_agbno = agbno;
1059 *new_aglen = aglen;
1060 if (adj == XFS_REFCOUNT_ADJUST_INCREASE)
1061 trace_xfs_refcount_increase(cur->bc_mp, cur->bc_ag.pag->pag_agno,
1062 agbno, aglen);
1063 else
1064 trace_xfs_refcount_decrease(cur->bc_mp, cur->bc_ag.pag->pag_agno,
1065 agbno, aglen);
1066
1067 /*
1068 * Ensure that no rcextents cross the boundary of the adjustment range.
1069 */
1070 error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1071 if (error)
1072 goto out_error;
1073 if (shape_changed)
1074 shape_changes++;
1075
1076 error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1077 if (error)
1078 goto out_error;
1079 if (shape_changed)
1080 shape_changes++;
1081
1082 /*
1083 * Try to merge with the left or right extents of the range.
1084 */
1085 error = xfs_refcount_merge_extents(cur, new_agbno, new_aglen, adj,
1086 XFS_FIND_RCEXT_SHARED, &shape_changed);
1087 if (error)
1088 goto out_error;
1089 if (shape_changed)
1090 shape_changes++;
1091 if (shape_changes)
1092 cur->bc_ag.refc.shape_changes++;
1093
1094 /* Now that we've taken care of the ends, adjust the middle extents */
1095 error = xfs_refcount_adjust_extents(cur, new_agbno, new_aglen,
1096 adj, oinfo);
1097 if (error)
1098 goto out_error;
1099
1100 return 0;
1101
1102out_error:
1103 trace_xfs_refcount_adjust_error(cur->bc_mp, cur->bc_ag.pag->pag_agno,
1104 error, _RET_IP_);
1105 return error;
1106}
1107
1108/* Clean up after calling xfs_refcount_finish_one. */
1109void
1110xfs_refcount_finish_one_cleanup(
1111 struct xfs_trans *tp,
1112 struct xfs_btree_cur *rcur,
1113 int error)
1114{
1115 struct xfs_buf *agbp;
1116
1117 if (rcur == NULL)
1118 return;
1119 agbp = rcur->bc_ag.agbp;
1120 xfs_btree_del_cursor(rcur, error);
1121 if (error)
1122 xfs_trans_brelse(tp, agbp);
1123}
1124
1125/*
1126 * Process one of the deferred refcount operations. We pass back the
1127 * btree cursor to maintain our lock on the btree between calls.
1128 * This saves time and eliminates a buffer deadlock between the
1129 * superblock and the AGF because we'll always grab them in the same
1130 * order.
1131 */
1132int
1133xfs_refcount_finish_one(
1134 struct xfs_trans *tp,
1135 enum xfs_refcount_intent_type type,
1136 xfs_fsblock_t startblock,
1137 xfs_extlen_t blockcount,
1138 xfs_fsblock_t *new_fsb,
1139 xfs_extlen_t *new_len,
1140 struct xfs_btree_cur **pcur)
1141{
1142 struct xfs_mount *mp = tp->t_mountp;
1143 struct xfs_btree_cur *rcur;
1144 struct xfs_buf *agbp = NULL;
1145 int error = 0;
1146 xfs_agblock_t bno;
1147 xfs_agblock_t new_agbno;
1148 unsigned long nr_ops = 0;
1149 int shape_changes = 0;
1150 struct xfs_perag *pag;
1151
1152 pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, startblock));
1153 bno = XFS_FSB_TO_AGBNO(mp, startblock);
1154
1155 trace_xfs_refcount_deferred(mp, XFS_FSB_TO_AGNO(mp, startblock),
1156 type, XFS_FSB_TO_AGBNO(mp, startblock),
1157 blockcount);
1158
1159 if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_REFCOUNT_FINISH_ONE)) {
1160 error = -EIO;
1161 goto out_drop;
1162 }
1163
1164 /*
1165 * If we haven't gotten a cursor or the cursor AG doesn't match
1166 * the startblock, get one now.
1167 */
1168 rcur = *pcur;
1169 if (rcur != NULL && rcur->bc_ag.pag != pag) {
1170 nr_ops = rcur->bc_ag.refc.nr_ops;
1171 shape_changes = rcur->bc_ag.refc.shape_changes;
1172 xfs_refcount_finish_one_cleanup(tp, rcur, 0);
1173 rcur = NULL;
1174 *pcur = NULL;
1175 }
1176 if (rcur == NULL) {
1177 error = xfs_alloc_read_agf(tp->t_mountp, tp, pag->pag_agno,
1178 XFS_ALLOC_FLAG_FREEING, &agbp);
1179 if (error)
1180 goto out_drop;
1181
1182 rcur = xfs_refcountbt_init_cursor(mp, tp, agbp, pag);
1183 rcur->bc_ag.refc.nr_ops = nr_ops;
1184 rcur->bc_ag.refc.shape_changes = shape_changes;
1185 }
1186 *pcur = rcur;
1187
1188 switch (type) {
1189 case XFS_REFCOUNT_INCREASE:
1190 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1191 new_len, XFS_REFCOUNT_ADJUST_INCREASE, NULL);
1192 *new_fsb = XFS_AGB_TO_FSB(mp, pag->pag_agno, new_agbno);
1193 break;
1194 case XFS_REFCOUNT_DECREASE:
1195 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1196 new_len, XFS_REFCOUNT_ADJUST_DECREASE, NULL);
1197 *new_fsb = XFS_AGB_TO_FSB(mp, pag->pag_agno, new_agbno);
1198 break;
1199 case XFS_REFCOUNT_ALLOC_COW:
1200 *new_fsb = startblock + blockcount;
1201 *new_len = 0;
1202 error = __xfs_refcount_cow_alloc(rcur, bno, blockcount);
1203 break;
1204 case XFS_REFCOUNT_FREE_COW:
1205 *new_fsb = startblock + blockcount;
1206 *new_len = 0;
1207 error = __xfs_refcount_cow_free(rcur, bno, blockcount);
1208 break;
1209 default:
1210 ASSERT(0);
1211 error = -EFSCORRUPTED;
1212 }
1213 if (!error && *new_len > 0)
1214 trace_xfs_refcount_finish_one_leftover(mp, pag->pag_agno, type,
1215 bno, blockcount, new_agbno, *new_len);
1216out_drop:
1217 xfs_perag_put(pag);
1218 return error;
1219}
1220
1221/*
1222 * Record a refcount intent for later processing.
1223 */
1224static void
1225__xfs_refcount_add(
1226 struct xfs_trans *tp,
1227 enum xfs_refcount_intent_type type,
1228 xfs_fsblock_t startblock,
1229 xfs_extlen_t blockcount)
1230{
1231 struct xfs_refcount_intent *ri;
1232
1233 trace_xfs_refcount_defer(tp->t_mountp,
1234 XFS_FSB_TO_AGNO(tp->t_mountp, startblock),
1235 type, XFS_FSB_TO_AGBNO(tp->t_mountp, startblock),
1236 blockcount);
1237
1238 ri = kmem_alloc(sizeof(struct xfs_refcount_intent),
1239 KM_NOFS);
1240 INIT_LIST_HEAD(&ri->ri_list);
1241 ri->ri_type = type;
1242 ri->ri_startblock = startblock;
1243 ri->ri_blockcount = blockcount;
1244
1245 xfs_defer_add(tp, XFS_DEFER_OPS_TYPE_REFCOUNT, &ri->ri_list);
1246}
1247
1248/*
1249 * Increase the reference count of the blocks backing a file's extent.
1250 */
1251void
1252xfs_refcount_increase_extent(
1253 struct xfs_trans *tp,
1254 struct xfs_bmbt_irec *PREV)
1255{
1256 if (!xfs_sb_version_hasreflink(&tp->t_mountp->m_sb))
1257 return;
1258
1259 __xfs_refcount_add(tp, XFS_REFCOUNT_INCREASE, PREV->br_startblock,
1260 PREV->br_blockcount);
1261}
1262
1263/*
1264 * Decrease the reference count of the blocks backing a file's extent.
1265 */
1266void
1267xfs_refcount_decrease_extent(
1268 struct xfs_trans *tp,
1269 struct xfs_bmbt_irec *PREV)
1270{
1271 if (!xfs_sb_version_hasreflink(&tp->t_mountp->m_sb))
1272 return;
1273
1274 __xfs_refcount_add(tp, XFS_REFCOUNT_DECREASE, PREV->br_startblock,
1275 PREV->br_blockcount);
1276}
1277
1278/*
1279 * Given an AG extent, find the lowest-numbered run of shared blocks
1280 * within that range and return the range in fbno/flen. If
1281 * find_end_of_shared is set, return the longest contiguous extent of
1282 * shared blocks; if not, just return the first extent we find. If no
1283 * shared blocks are found, fbno and flen will be set to NULLAGBLOCK
1284 * and 0, respectively.
1285 */
1286int
1287xfs_refcount_find_shared(
1288 struct xfs_btree_cur *cur,
1289 xfs_agblock_t agbno,
1290 xfs_extlen_t aglen,
1291 xfs_agblock_t *fbno,
1292 xfs_extlen_t *flen,
1293 bool find_end_of_shared)
1294{
1295 struct xfs_refcount_irec tmp;
1296 int i;
1297 int have;
1298 int error;
1299
1300 trace_xfs_refcount_find_shared(cur->bc_mp, cur->bc_ag.pag->pag_agno,
1301 agbno, aglen);
1302
1303 /* By default, skip the whole range */
1304 *fbno = NULLAGBLOCK;
1305 *flen = 0;
1306
1307 /* Try to find a refcount extent that crosses the start */
1308 error = xfs_refcount_lookup_le(cur, agbno, &have);
1309 if (error)
1310 goto out_error;
1311 if (!have) {
1312 /* No left extent, look at the next one */
1313 error = xfs_btree_increment(cur, 0, &have);
1314 if (error)
1315 goto out_error;
1316 if (!have)
1317 goto done;
1318 }
1319 error = xfs_refcount_get_rec(cur, &tmp, &i);
1320 if (error)
1321 goto out_error;
1322 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1323 error = -EFSCORRUPTED;
1324 goto out_error;
1325 }
1326
1327 /* If the extent ends before the start, look at the next one */
1328 if (tmp.rc_startblock + tmp.rc_blockcount <= agbno) {
1329 error = xfs_btree_increment(cur, 0, &have);
1330 if (error)
1331 goto out_error;
1332 if (!have)
1333 goto done;
1334 error = xfs_refcount_get_rec(cur, &tmp, &i);
1335 if (error)
1336 goto out_error;
1337 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1338 error = -EFSCORRUPTED;
1339 goto out_error;
1340 }
1341 }
1342
1343 /* If the extent starts after the range we want, bail out */
1344 if (tmp.rc_startblock >= agbno + aglen)
1345 goto done;
1346
1347 /* We found the start of a shared extent! */
1348 if (tmp.rc_startblock < agbno) {
1349 tmp.rc_blockcount -= (agbno - tmp.rc_startblock);
1350 tmp.rc_startblock = agbno;
1351 }
1352
1353 *fbno = tmp.rc_startblock;
1354 *flen = min(tmp.rc_blockcount, agbno + aglen - *fbno);
1355 if (!find_end_of_shared)
1356 goto done;
1357
1358 /* Otherwise, find the end of this shared extent */
1359 while (*fbno + *flen < agbno + aglen) {
1360 error = xfs_btree_increment(cur, 0, &have);
1361 if (error)
1362 goto out_error;
1363 if (!have)
1364 break;
1365 error = xfs_refcount_get_rec(cur, &tmp, &i);
1366 if (error)
1367 goto out_error;
1368 if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1369 error = -EFSCORRUPTED;
1370 goto out_error;
1371 }
1372 if (tmp.rc_startblock >= agbno + aglen ||
1373 tmp.rc_startblock != *fbno + *flen)
1374 break;
1375 *flen = min(*flen + tmp.rc_blockcount, agbno + aglen - *fbno);
1376 }
1377
1378done:
1379 trace_xfs_refcount_find_shared_result(cur->bc_mp,
1380 cur->bc_ag.pag->pag_agno, *fbno, *flen);
1381
1382out_error:
1383 if (error)
1384 trace_xfs_refcount_find_shared_error(cur->bc_mp,
1385 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
1386 return error;
1387}
1388
1389/*
1390 * Recovering CoW Blocks After a Crash
1391 *
1392 * Due to the way that the copy on write mechanism works, there's a window of
1393 * opportunity in which we can lose track of allocated blocks during a crash.
1394 * Because CoW uses delayed allocation in the in-core CoW fork, writeback
1395 * causes blocks to be allocated and stored in the CoW fork. The blocks are
1396 * no longer in the free space btree but are not otherwise recorded anywhere
1397 * until the write completes and the blocks are mapped into the file. A crash
1398 * in between allocation and remapping results in the replacement blocks being
1399 * lost. This situation is exacerbated by the CoW extent size hint because
1400 * allocations can hang around for long time.
1401 *
1402 * However, there is a place where we can record these allocations before they
1403 * become mappings -- the reference count btree. The btree does not record
1404 * extents with refcount == 1, so we can record allocations with a refcount of
1405 * 1. Blocks being used for CoW writeout cannot be shared, so there should be
1406 * no conflict with shared block records. These mappings should be created
1407 * when we allocate blocks to the CoW fork and deleted when they're removed
1408 * from the CoW fork.
1409 *
1410 * Minor nit: records for in-progress CoW allocations and records for shared
1411 * extents must never be merged, to preserve the property that (except for CoW
1412 * allocations) there are no refcount btree entries with refcount == 1. The
1413 * only time this could potentially happen is when unsharing a block that's
1414 * adjacent to CoW allocations, so we must be careful to avoid this.
1415 *
1416 * At mount time we recover lost CoW allocations by searching the refcount
1417 * btree for these refcount == 1 mappings. These represent CoW allocations
1418 * that were in progress at the time the filesystem went down, so we can free
1419 * them to get the space back.
1420 *
1421 * This mechanism is superior to creating EFIs for unmapped CoW extents for
1422 * several reasons -- first, EFIs pin the tail of the log and would have to be
1423 * periodically relogged to avoid filling up the log. Second, CoW completions
1424 * will have to file an EFD and create new EFIs for whatever remains in the
1425 * CoW fork; this partially takes care of (1) but extent-size reservations
1426 * will have to periodically relog even if there's no writeout in progress.
1427 * This can happen if the CoW extent size hint is set, which you really want.
1428 * Third, EFIs cannot currently be automatically relogged into newer
1429 * transactions to advance the log tail. Fourth, stuffing the log full of
1430 * EFIs places an upper bound on the number of CoW allocations that can be
1431 * held filesystem-wide at any given time. Recording them in the refcount
1432 * btree doesn't require us to maintain any state in memory and doesn't pin
1433 * the log.
1434 */
1435/*
1436 * Adjust the refcounts of CoW allocations. These allocations are "magic"
1437 * in that they're not referenced anywhere else in the filesystem, so we
1438 * stash them in the refcount btree with a refcount of 1 until either file
1439 * remapping (or CoW cancellation) happens.
1440 */
1441STATIC int
1442xfs_refcount_adjust_cow_extents(
1443 struct xfs_btree_cur *cur,
1444 xfs_agblock_t agbno,
1445 xfs_extlen_t aglen,
1446 enum xfs_refc_adjust_op adj)
1447{
1448 struct xfs_refcount_irec ext, tmp;
1449 int error;
1450 int found_rec, found_tmp;
1451
1452 if (aglen == 0)
1453 return 0;
1454
1455 /* Find any overlapping refcount records */
1456 error = xfs_refcount_lookup_ge(cur, agbno, &found_rec);
1457 if (error)
1458 goto out_error;
1459 error = xfs_refcount_get_rec(cur, &ext, &found_rec);
1460 if (error)
1461 goto out_error;
1462 if (!found_rec) {
1463 ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks +
1464 XFS_REFC_COW_START;
1465 ext.rc_blockcount = 0;
1466 ext.rc_refcount = 0;
1467 }
1468
1469 switch (adj) {
1470 case XFS_REFCOUNT_ADJUST_COW_ALLOC:
1471 /* Adding a CoW reservation, there should be nothing here. */
1472 if (XFS_IS_CORRUPT(cur->bc_mp,
1473 agbno + aglen > ext.rc_startblock)) {
1474 error = -EFSCORRUPTED;
1475 goto out_error;
1476 }
1477
1478 tmp.rc_startblock = agbno;
1479 tmp.rc_blockcount = aglen;
1480 tmp.rc_refcount = 1;
1481 trace_xfs_refcount_modify_extent(cur->bc_mp,
1482 cur->bc_ag.pag->pag_agno, &tmp);
1483
1484 error = xfs_refcount_insert(cur, &tmp,
1485 &found_tmp);
1486 if (error)
1487 goto out_error;
1488 if (XFS_IS_CORRUPT(cur->bc_mp, found_tmp != 1)) {
1489 error = -EFSCORRUPTED;
1490 goto out_error;
1491 }
1492 break;
1493 case XFS_REFCOUNT_ADJUST_COW_FREE:
1494 /* Removing a CoW reservation, there should be one extent. */
1495 if (XFS_IS_CORRUPT(cur->bc_mp, ext.rc_startblock != agbno)) {
1496 error = -EFSCORRUPTED;
1497 goto out_error;
1498 }
1499 if (XFS_IS_CORRUPT(cur->bc_mp, ext.rc_blockcount != aglen)) {
1500 error = -EFSCORRUPTED;
1501 goto out_error;
1502 }
1503 if (XFS_IS_CORRUPT(cur->bc_mp, ext.rc_refcount != 1)) {
1504 error = -EFSCORRUPTED;
1505 goto out_error;
1506 }
1507
1508 ext.rc_refcount = 0;
1509 trace_xfs_refcount_modify_extent(cur->bc_mp,
1510 cur->bc_ag.pag->pag_agno, &ext);
1511 error = xfs_refcount_delete(cur, &found_rec);
1512 if (error)
1513 goto out_error;
1514 if (XFS_IS_CORRUPT(cur->bc_mp, found_rec != 1)) {
1515 error = -EFSCORRUPTED;
1516 goto out_error;
1517 }
1518 break;
1519 default:
1520 ASSERT(0);
1521 }
1522
1523 return error;
1524out_error:
1525 trace_xfs_refcount_modify_extent_error(cur->bc_mp,
1526 cur->bc_ag.pag->pag_agno, error, _RET_IP_);
1527 return error;
1528}
1529
1530/*
1531 * Add or remove refcount btree entries for CoW reservations.
1532 */
1533STATIC int
1534xfs_refcount_adjust_cow(
1535 struct xfs_btree_cur *cur,
1536 xfs_agblock_t agbno,
1537 xfs_extlen_t aglen,
1538 enum xfs_refc_adjust_op adj)
1539{
1540 bool shape_changed;
1541 int error;
1542
1543 agbno += XFS_REFC_COW_START;
1544
1545 /*
1546 * Ensure that no rcextents cross the boundary of the adjustment range.
1547 */
1548 error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1549 if (error)
1550 goto out_error;
1551
1552 error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1553 if (error)
1554 goto out_error;
1555
1556 /*
1557 * Try to merge with the left or right extents of the range.
1558 */
1559 error = xfs_refcount_merge_extents(cur, &agbno, &aglen, adj,
1560 XFS_FIND_RCEXT_COW, &shape_changed);
1561 if (error)
1562 goto out_error;
1563
1564 /* Now that we've taken care of the ends, adjust the middle extents */
1565 error = xfs_refcount_adjust_cow_extents(cur, agbno, aglen, adj);
1566 if (error)
1567 goto out_error;
1568
1569 return 0;
1570
1571out_error:
1572 trace_xfs_refcount_adjust_cow_error(cur->bc_mp, cur->bc_ag.pag->pag_agno,
1573 error, _RET_IP_);
1574 return error;
1575}
1576
1577/*
1578 * Record a CoW allocation in the refcount btree.
1579 */
1580STATIC int
1581__xfs_refcount_cow_alloc(
1582 struct xfs_btree_cur *rcur,
1583 xfs_agblock_t agbno,
1584 xfs_extlen_t aglen)
1585{
1586 trace_xfs_refcount_cow_increase(rcur->bc_mp, rcur->bc_ag.pag->pag_agno,
1587 agbno, aglen);
1588
1589 /* Add refcount btree reservation */
1590 return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1591 XFS_REFCOUNT_ADJUST_COW_ALLOC);
1592}
1593
1594/*
1595 * Remove a CoW allocation from the refcount btree.
1596 */
1597STATIC int
1598__xfs_refcount_cow_free(
1599 struct xfs_btree_cur *rcur,
1600 xfs_agblock_t agbno,
1601 xfs_extlen_t aglen)
1602{
1603 trace_xfs_refcount_cow_decrease(rcur->bc_mp, rcur->bc_ag.pag->pag_agno,
1604 agbno, aglen);
1605
1606 /* Remove refcount btree reservation */
1607 return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1608 XFS_REFCOUNT_ADJUST_COW_FREE);
1609}
1610
1611/* Record a CoW staging extent in the refcount btree. */
1612void
1613xfs_refcount_alloc_cow_extent(
1614 struct xfs_trans *tp,
1615 xfs_fsblock_t fsb,
1616 xfs_extlen_t len)
1617{
1618 struct xfs_mount *mp = tp->t_mountp;
1619
1620 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1621 return;
1622
1623 __xfs_refcount_add(tp, XFS_REFCOUNT_ALLOC_COW, fsb, len);
1624
1625 /* Add rmap entry */
1626 xfs_rmap_alloc_extent(tp, XFS_FSB_TO_AGNO(mp, fsb),
1627 XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1628}
1629
1630/* Forget a CoW staging event in the refcount btree. */
1631void
1632xfs_refcount_free_cow_extent(
1633 struct xfs_trans *tp,
1634 xfs_fsblock_t fsb,
1635 xfs_extlen_t len)
1636{
1637 struct xfs_mount *mp = tp->t_mountp;
1638
1639 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1640 return;
1641
1642 /* Remove rmap entry */
1643 xfs_rmap_free_extent(tp, XFS_FSB_TO_AGNO(mp, fsb),
1644 XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1645 __xfs_refcount_add(tp, XFS_REFCOUNT_FREE_COW, fsb, len);
1646}
1647
1648struct xfs_refcount_recovery {
1649 struct list_head rr_list;
1650 struct xfs_refcount_irec rr_rrec;
1651};
1652
1653/* Stuff an extent on the recovery list. */
1654STATIC int
1655xfs_refcount_recover_extent(
1656 struct xfs_btree_cur *cur,
1657 union xfs_btree_rec *rec,
1658 void *priv)
1659{
1660 struct list_head *debris = priv;
1661 struct xfs_refcount_recovery *rr;
1662
1663 if (XFS_IS_CORRUPT(cur->bc_mp,
1664 be32_to_cpu(rec->refc.rc_refcount) != 1))
1665 return -EFSCORRUPTED;
1666
1667 rr = kmem_alloc(sizeof(struct xfs_refcount_recovery), 0);
1668 xfs_refcount_btrec_to_irec(rec, &rr->rr_rrec);
1669 list_add_tail(&rr->rr_list, debris);
1670
1671 return 0;
1672}
1673
1674/* Find and remove leftover CoW reservations. */
1675int
1676xfs_refcount_recover_cow_leftovers(
1677 struct xfs_mount *mp,
1678 struct xfs_perag *pag)
1679{
1680 struct xfs_trans *tp;
1681 struct xfs_btree_cur *cur;
1682 struct xfs_buf *agbp;
1683 struct xfs_refcount_recovery *rr, *n;
1684 struct list_head debris;
1685 union xfs_btree_irec low;
1686 union xfs_btree_irec high;
1687 xfs_fsblock_t fsb;
1688 xfs_agblock_t agbno;
1689 int error;
1690
1691 if (mp->m_sb.sb_agblocks >= XFS_REFC_COW_START)
1692 return -EOPNOTSUPP;
1693
1694 INIT_LIST_HEAD(&debris);
1695
1696 /*
1697 * In this first part, we use an empty transaction to gather up
1698 * all the leftover CoW extents so that we can subsequently
1699 * delete them. The empty transaction is used to avoid
1700 * a buffer lock deadlock if there happens to be a loop in the
1701 * refcountbt because we're allowed to re-grab a buffer that is
1702 * already attached to our transaction. When we're done
1703 * recording the CoW debris we cancel the (empty) transaction
1704 * and everything goes away cleanly.
1705 */
1706 error = xfs_trans_alloc_empty(mp, &tp);
1707 if (error)
1708 return error;
1709
1710 error = xfs_alloc_read_agf(mp, tp, pag->pag_agno, 0, &agbp);
1711 if (error)
1712 goto out_trans;
1713 cur = xfs_refcountbt_init_cursor(mp, tp, agbp, pag);
1714
1715 /* Find all the leftover CoW staging extents. */
1716 memset(&low, 0, sizeof(low));
1717 memset(&high, 0, sizeof(high));
1718 low.rc.rc_startblock = XFS_REFC_COW_START;
1719 high.rc.rc_startblock = -1U;
1720 error = xfs_btree_query_range(cur, &low, &high,
1721 xfs_refcount_recover_extent, &debris);
1722 xfs_btree_del_cursor(cur, error);
1723 xfs_trans_brelse(tp, agbp);
1724 xfs_trans_cancel(tp);
1725 if (error)
1726 goto out_free;
1727
1728 /* Now iterate the list to free the leftovers */
1729 list_for_each_entry_safe(rr, n, &debris, rr_list) {
1730 /* Set up transaction. */
1731 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1732 if (error)
1733 goto out_free;
1734
1735 trace_xfs_refcount_recover_extent(mp, pag->pag_agno,
1736 &rr->rr_rrec);
1737
1738 /* Free the orphan record */
1739 agbno = rr->rr_rrec.rc_startblock - XFS_REFC_COW_START;
1740 fsb = XFS_AGB_TO_FSB(mp, pag->pag_agno, agbno);
1741 xfs_refcount_free_cow_extent(tp, fsb,
1742 rr->rr_rrec.rc_blockcount);
1743
1744 /* Free the block. */
1745 xfs_bmap_add_free(tp, fsb, rr->rr_rrec.rc_blockcount, NULL);
1746
1747 error = xfs_trans_commit(tp);
1748 if (error)
1749 goto out_free;
1750
1751 list_del(&rr->rr_list);
1752 kmem_free(rr);
1753 }
1754
1755 return error;
1756out_trans:
1757 xfs_trans_cancel(tp);
1758out_free:
1759 /* Free the leftover list */
1760 list_for_each_entry_safe(rr, n, &debris, rr_list) {
1761 list_del(&rr->rr_list);
1762 kmem_free(rr);
1763 }
1764 return error;
1765}
1766
1767/* Is there a record covering a given extent? */
1768int
1769xfs_refcount_has_record(
1770 struct xfs_btree_cur *cur,
1771 xfs_agblock_t bno,
1772 xfs_extlen_t len,
1773 bool *exists)
1774{
1775 union xfs_btree_irec low;
1776 union xfs_btree_irec high;
1777
1778 memset(&low, 0, sizeof(low));
1779 low.rc.rc_startblock = bno;
1780 memset(&high, 0xFF, sizeof(high));
1781 high.rc.rc_startblock = bno + len - 1;
1782
1783 return xfs_btree_has_record(cur, &low, &high, exists);
1784}