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_inode.h"
15#include "xfs_trans.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_trace.h"
19#include "xfs_icache.h"
20#include "xfs_btree.h"
21#include "xfs_refcount_btree.h"
22#include "xfs_refcount.h"
23#include "xfs_bmap_btree.h"
24#include "xfs_trans_space.h"
25#include "xfs_bit.h"
26#include "xfs_alloc.h"
27#include "xfs_quota.h"
28#include "xfs_reflink.h"
29#include "xfs_iomap.h"
30#include "xfs_ag.h"
31#include "xfs_ag_resv.h"
32
33/*
34 * Copy on Write of Shared Blocks
35 *
36 * XFS must preserve "the usual" file semantics even when two files share
37 * the same physical blocks. This means that a write to one file must not
38 * alter the blocks in a different file; the way that we'll do that is
39 * through the use of a copy-on-write mechanism. At a high level, that
40 * means that when we want to write to a shared block, we allocate a new
41 * block, write the data to the new block, and if that succeeds we map the
42 * new block into the file.
43 *
44 * XFS provides a "delayed allocation" mechanism that defers the allocation
45 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
46 * possible. This reduces fragmentation by enabling the filesystem to ask
47 * for bigger chunks less often, which is exactly what we want for CoW.
48 *
49 * The delalloc mechanism begins when the kernel wants to make a block
50 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
51 * create a delalloc mapping, which is a regular in-core extent, but without
52 * a real startblock. (For delalloc mappings, the startblock encodes both
53 * a flag that this is a delalloc mapping, and a worst-case estimate of how
54 * many blocks might be required to put the mapping into the BMBT.) delalloc
55 * mappings are a reservation against the free space in the filesystem;
56 * adjacent mappings can also be combined into fewer larger mappings.
57 *
58 * As an optimization, the CoW extent size hint (cowextsz) creates
59 * outsized aligned delalloc reservations in the hope of landing out of
60 * order nearby CoW writes in a single extent on disk, thereby reducing
61 * fragmentation and improving future performance.
62 *
63 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
64 * C: ------DDDDDDD--------- (CoW fork)
65 *
66 * When dirty pages are being written out (typically in writepage), the
67 * delalloc reservations are converted into unwritten mappings by
68 * allocating blocks and replacing the delalloc mapping with real ones.
69 * A delalloc mapping can be replaced by several unwritten ones if the
70 * free space is fragmented.
71 *
72 * D: --RRRRRRSSSRRRRRRRR---
73 * C: ------UUUUUUU---------
74 *
75 * We want to adapt the delalloc mechanism for copy-on-write, since the
76 * write paths are similar. The first two steps (creating the reservation
77 * and allocating the blocks) are exactly the same as delalloc except that
78 * the mappings must be stored in a separate CoW fork because we do not want
79 * to disturb the mapping in the data fork until we're sure that the write
80 * succeeded. IO completion in this case is the process of removing the old
81 * mapping from the data fork and moving the new mapping from the CoW fork to
82 * the data fork. This will be discussed shortly.
83 *
84 * For now, unaligned directio writes will be bounced back to the page cache.
85 * Block-aligned directio writes will use the same mechanism as buffered
86 * writes.
87 *
88 * Just prior to submitting the actual disk write requests, we convert
89 * the extents representing the range of the file actually being written
90 * (as opposed to extra pieces created for the cowextsize hint) to real
91 * extents. This will become important in the next step:
92 *
93 * D: --RRRRRRSSSRRRRRRRR---
94 * C: ------UUrrUUU---------
95 *
96 * CoW remapping must be done after the data block write completes,
97 * because we don't want to destroy the old data fork map until we're sure
98 * the new block has been written. Since the new mappings are kept in a
99 * separate fork, we can simply iterate these mappings to find the ones
100 * that cover the file blocks that we just CoW'd. For each extent, simply
101 * unmap the corresponding range in the data fork, map the new range into
102 * the data fork, and remove the extent from the CoW fork. Because of
103 * the presence of the cowextsize hint, however, we must be careful
104 * only to remap the blocks that we've actually written out -- we must
105 * never remap delalloc reservations nor CoW staging blocks that have
106 * yet to be written. This corresponds exactly to the real extents in
107 * the CoW fork:
108 *
109 * D: --RRRRRRrrSRRRRRRRR---
110 * C: ------UU--UUU---------
111 *
112 * Since the remapping operation can be applied to an arbitrary file
113 * range, we record the need for the remap step as a flag in the ioend
114 * instead of declaring a new IO type. This is required for direct io
115 * because we only have ioend for the whole dio, and we have to be able to
116 * remember the presence of unwritten blocks and CoW blocks with a single
117 * ioend structure. Better yet, the more ground we can cover with one
118 * ioend, the better.
119 */
120
121/*
122 * Given an AG extent, find the lowest-numbered run of shared blocks
123 * within that range and return the range in fbno/flen. If
124 * find_end_of_shared is true, return the longest contiguous extent of
125 * shared blocks. If there are no shared extents, fbno and flen will
126 * be set to NULLAGBLOCK and 0, respectively.
127 */
128static int
129xfs_reflink_find_shared(
130 struct xfs_perag *pag,
131 struct xfs_trans *tp,
132 xfs_agblock_t agbno,
133 xfs_extlen_t aglen,
134 xfs_agblock_t *fbno,
135 xfs_extlen_t *flen,
136 bool find_end_of_shared)
137{
138 struct xfs_buf *agbp;
139 struct xfs_btree_cur *cur;
140 int error;
141
142 error = xfs_alloc_read_agf(pag, tp, 0, &agbp);
143 if (error)
144 return error;
145
146 cur = xfs_refcountbt_init_cursor(pag->pag_mount, tp, agbp, pag);
147
148 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
149 find_end_of_shared);
150
151 xfs_btree_del_cursor(cur, error);
152
153 xfs_trans_brelse(tp, agbp);
154 return error;
155}
156
157/*
158 * Trim the mapping to the next block where there's a change in the
159 * shared/unshared status. More specifically, this means that we
160 * find the lowest-numbered extent of shared blocks that coincides with
161 * the given block mapping. If the shared extent overlaps the start of
162 * the mapping, trim the mapping to the end of the shared extent. If
163 * the shared region intersects the mapping, trim the mapping to the
164 * start of the shared extent. If there are no shared regions that
165 * overlap, just return the original extent.
166 */
167int
168xfs_reflink_trim_around_shared(
169 struct xfs_inode *ip,
170 struct xfs_bmbt_irec *irec,
171 bool *shared)
172{
173 struct xfs_mount *mp = ip->i_mount;
174 struct xfs_perag *pag;
175 xfs_agblock_t agbno;
176 xfs_extlen_t aglen;
177 xfs_agblock_t fbno;
178 xfs_extlen_t flen;
179 int error = 0;
180
181 /* Holes, unwritten, and delalloc extents cannot be shared */
182 if (!xfs_is_cow_inode(ip) || !xfs_bmap_is_written_extent(irec)) {
183 *shared = false;
184 return 0;
185 }
186
187 trace_xfs_reflink_trim_around_shared(ip, irec);
188
189 pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, irec->br_startblock));
190 agbno = XFS_FSB_TO_AGBNO(mp, irec->br_startblock);
191 aglen = irec->br_blockcount;
192
193 error = xfs_reflink_find_shared(pag, NULL, agbno, aglen, &fbno, &flen,
194 true);
195 xfs_perag_put(pag);
196 if (error)
197 return error;
198
199 *shared = false;
200 if (fbno == NULLAGBLOCK) {
201 /* No shared blocks at all. */
202 return 0;
203 }
204
205 if (fbno == agbno) {
206 /*
207 * The start of this extent is shared. Truncate the
208 * mapping at the end of the shared region so that a
209 * subsequent iteration starts at the start of the
210 * unshared region.
211 */
212 irec->br_blockcount = flen;
213 *shared = true;
214 return 0;
215 }
216
217 /*
218 * There's a shared extent midway through this extent.
219 * Truncate the mapping at the start of the shared
220 * extent so that a subsequent iteration starts at the
221 * start of the shared region.
222 */
223 irec->br_blockcount = fbno - agbno;
224 return 0;
225}
226
227int
228xfs_bmap_trim_cow(
229 struct xfs_inode *ip,
230 struct xfs_bmbt_irec *imap,
231 bool *shared)
232{
233 /* We can't update any real extents in always COW mode. */
234 if (xfs_is_always_cow_inode(ip) &&
235 !isnullstartblock(imap->br_startblock)) {
236 *shared = true;
237 return 0;
238 }
239
240 /* Trim the mapping to the nearest shared extent boundary. */
241 return xfs_reflink_trim_around_shared(ip, imap, shared);
242}
243
244static int
245xfs_reflink_convert_cow_locked(
246 struct xfs_inode *ip,
247 xfs_fileoff_t offset_fsb,
248 xfs_filblks_t count_fsb)
249{
250 struct xfs_iext_cursor icur;
251 struct xfs_bmbt_irec got;
252 struct xfs_btree_cur *dummy_cur = NULL;
253 int dummy_logflags;
254 int error = 0;
255
256 if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
257 return 0;
258
259 do {
260 if (got.br_startoff >= offset_fsb + count_fsb)
261 break;
262 if (got.br_state == XFS_EXT_NORM)
263 continue;
264 if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
265 return -EIO;
266
267 xfs_trim_extent(&got, offset_fsb, count_fsb);
268 if (!got.br_blockcount)
269 continue;
270
271 got.br_state = XFS_EXT_NORM;
272 error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
273 XFS_COW_FORK, &icur, &dummy_cur, &got,
274 &dummy_logflags);
275 if (error)
276 return error;
277 } while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
278
279 return error;
280}
281
282/* Convert all of the unwritten CoW extents in a file's range to real ones. */
283int
284xfs_reflink_convert_cow(
285 struct xfs_inode *ip,
286 xfs_off_t offset,
287 xfs_off_t count)
288{
289 struct xfs_mount *mp = ip->i_mount;
290 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
291 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
292 xfs_filblks_t count_fsb = end_fsb - offset_fsb;
293 int error;
294
295 ASSERT(count != 0);
296
297 xfs_ilock(ip, XFS_ILOCK_EXCL);
298 error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
299 xfs_iunlock(ip, XFS_ILOCK_EXCL);
300 return error;
301}
302
303/*
304 * Find the extent that maps the given range in the COW fork. Even if the extent
305 * is not shared we might have a preallocation for it in the COW fork. If so we
306 * use it that rather than trigger a new allocation.
307 */
308static int
309xfs_find_trim_cow_extent(
310 struct xfs_inode *ip,
311 struct xfs_bmbt_irec *imap,
312 struct xfs_bmbt_irec *cmap,
313 bool *shared,
314 bool *found)
315{
316 xfs_fileoff_t offset_fsb = imap->br_startoff;
317 xfs_filblks_t count_fsb = imap->br_blockcount;
318 struct xfs_iext_cursor icur;
319
320 *found = false;
321
322 /*
323 * If we don't find an overlapping extent, trim the range we need to
324 * allocate to fit the hole we found.
325 */
326 if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, cmap))
327 cmap->br_startoff = offset_fsb + count_fsb;
328 if (cmap->br_startoff > offset_fsb) {
329 xfs_trim_extent(imap, imap->br_startoff,
330 cmap->br_startoff - imap->br_startoff);
331 return xfs_bmap_trim_cow(ip, imap, shared);
332 }
333
334 *shared = true;
335 if (isnullstartblock(cmap->br_startblock)) {
336 xfs_trim_extent(imap, cmap->br_startoff, cmap->br_blockcount);
337 return 0;
338 }
339
340 /* real extent found - no need to allocate */
341 xfs_trim_extent(cmap, offset_fsb, count_fsb);
342 *found = true;
343 return 0;
344}
345
346static int
347xfs_reflink_convert_unwritten(
348 struct xfs_inode *ip,
349 struct xfs_bmbt_irec *imap,
350 struct xfs_bmbt_irec *cmap,
351 bool convert_now)
352{
353 xfs_fileoff_t offset_fsb = imap->br_startoff;
354 xfs_filblks_t count_fsb = imap->br_blockcount;
355 int error;
356
357 /*
358 * cmap might larger than imap due to cowextsize hint.
359 */
360 xfs_trim_extent(cmap, offset_fsb, count_fsb);
361
362 /*
363 * COW fork extents are supposed to remain unwritten until we're ready
364 * to initiate a disk write. For direct I/O we are going to write the
365 * data and need the conversion, but for buffered writes we're done.
366 */
367 if (!convert_now || cmap->br_state == XFS_EXT_NORM)
368 return 0;
369
370 trace_xfs_reflink_convert_cow(ip, cmap);
371
372 error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
373 if (!error)
374 cmap->br_state = XFS_EXT_NORM;
375
376 return error;
377}
378
379static int
380xfs_reflink_fill_cow_hole(
381 struct xfs_inode *ip,
382 struct xfs_bmbt_irec *imap,
383 struct xfs_bmbt_irec *cmap,
384 bool *shared,
385 uint *lockmode,
386 bool convert_now)
387{
388 struct xfs_mount *mp = ip->i_mount;
389 struct xfs_trans *tp;
390 xfs_filblks_t resaligned;
391 xfs_extlen_t resblks;
392 int nimaps;
393 int error;
394 bool found;
395
396 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
397 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
398 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
399
400 xfs_iunlock(ip, *lockmode);
401 *lockmode = 0;
402
403 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks, 0,
404 false, &tp);
405 if (error)
406 return error;
407
408 *lockmode = XFS_ILOCK_EXCL;
409
410 error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
411 if (error || !*shared)
412 goto out_trans_cancel;
413
414 if (found) {
415 xfs_trans_cancel(tp);
416 goto convert;
417 }
418
419 /* Allocate the entire reservation as unwritten blocks. */
420 nimaps = 1;
421 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
422 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0, cmap,
423 &nimaps);
424 if (error)
425 goto out_trans_cancel;
426
427 xfs_inode_set_cowblocks_tag(ip);
428 error = xfs_trans_commit(tp);
429 if (error)
430 return error;
431
432 /*
433 * Allocation succeeded but the requested range was not even partially
434 * satisfied? Bail out!
435 */
436 if (nimaps == 0)
437 return -ENOSPC;
438
439convert:
440 return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
441
442out_trans_cancel:
443 xfs_trans_cancel(tp);
444 return error;
445}
446
447static int
448xfs_reflink_fill_delalloc(
449 struct xfs_inode *ip,
450 struct xfs_bmbt_irec *imap,
451 struct xfs_bmbt_irec *cmap,
452 bool *shared,
453 uint *lockmode,
454 bool convert_now)
455{
456 struct xfs_mount *mp = ip->i_mount;
457 struct xfs_trans *tp;
458 int nimaps;
459 int error;
460 bool found;
461
462 do {
463 xfs_iunlock(ip, *lockmode);
464 *lockmode = 0;
465
466 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, 0, 0,
467 false, &tp);
468 if (error)
469 return error;
470
471 *lockmode = XFS_ILOCK_EXCL;
472
473 error = xfs_find_trim_cow_extent(ip, imap, cmap, shared,
474 &found);
475 if (error || !*shared)
476 goto out_trans_cancel;
477
478 if (found) {
479 xfs_trans_cancel(tp);
480 break;
481 }
482
483 ASSERT(isnullstartblock(cmap->br_startblock) ||
484 cmap->br_startblock == DELAYSTARTBLOCK);
485
486 /*
487 * Replace delalloc reservation with an unwritten extent.
488 */
489 nimaps = 1;
490 error = xfs_bmapi_write(tp, ip, cmap->br_startoff,
491 cmap->br_blockcount,
492 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0,
493 cmap, &nimaps);
494 if (error)
495 goto out_trans_cancel;
496
497 xfs_inode_set_cowblocks_tag(ip);
498 error = xfs_trans_commit(tp);
499 if (error)
500 return error;
501
502 /*
503 * Allocation succeeded but the requested range was not even
504 * partially satisfied? Bail out!
505 */
506 if (nimaps == 0)
507 return -ENOSPC;
508 } while (cmap->br_startoff + cmap->br_blockcount <= imap->br_startoff);
509
510 return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
511
512out_trans_cancel:
513 xfs_trans_cancel(tp);
514 return error;
515}
516
517/* Allocate all CoW reservations covering a range of blocks in a file. */
518int
519xfs_reflink_allocate_cow(
520 struct xfs_inode *ip,
521 struct xfs_bmbt_irec *imap,
522 struct xfs_bmbt_irec *cmap,
523 bool *shared,
524 uint *lockmode,
525 bool convert_now)
526{
527 int error;
528 bool found;
529
530 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
531 if (!ip->i_cowfp) {
532 ASSERT(!xfs_is_reflink_inode(ip));
533 xfs_ifork_init_cow(ip);
534 }
535
536 error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
537 if (error || !*shared)
538 return error;
539
540 /* CoW fork has a real extent */
541 if (found)
542 return xfs_reflink_convert_unwritten(ip, imap, cmap,
543 convert_now);
544
545 /*
546 * CoW fork does not have an extent and data extent is shared.
547 * Allocate a real extent in the CoW fork.
548 */
549 if (cmap->br_startoff > imap->br_startoff)
550 return xfs_reflink_fill_cow_hole(ip, imap, cmap, shared,
551 lockmode, convert_now);
552
553 /*
554 * CoW fork has a delalloc reservation. Replace it with a real extent.
555 * There may or may not be a data fork mapping.
556 */
557 if (isnullstartblock(cmap->br_startblock) ||
558 cmap->br_startblock == DELAYSTARTBLOCK)
559 return xfs_reflink_fill_delalloc(ip, imap, cmap, shared,
560 lockmode, convert_now);
561
562 /* Shouldn't get here. */
563 ASSERT(0);
564 return -EFSCORRUPTED;
565}
566
567/*
568 * Cancel CoW reservations for some block range of an inode.
569 *
570 * If cancel_real is true this function cancels all COW fork extents for the
571 * inode; if cancel_real is false, real extents are not cleared.
572 *
573 * Caller must have already joined the inode to the current transaction. The
574 * inode will be joined to the transaction returned to the caller.
575 */
576int
577xfs_reflink_cancel_cow_blocks(
578 struct xfs_inode *ip,
579 struct xfs_trans **tpp,
580 xfs_fileoff_t offset_fsb,
581 xfs_fileoff_t end_fsb,
582 bool cancel_real)
583{
584 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
585 struct xfs_bmbt_irec got, del;
586 struct xfs_iext_cursor icur;
587 int error = 0;
588
589 if (!xfs_inode_has_cow_data(ip))
590 return 0;
591 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
592 return 0;
593
594 /* Walk backwards until we're out of the I/O range... */
595 while (got.br_startoff + got.br_blockcount > offset_fsb) {
596 del = got;
597 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
598
599 /* Extent delete may have bumped ext forward */
600 if (!del.br_blockcount) {
601 xfs_iext_prev(ifp, &icur);
602 goto next_extent;
603 }
604
605 trace_xfs_reflink_cancel_cow(ip, &del);
606
607 if (isnullstartblock(del.br_startblock)) {
608 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
609 &icur, &got, &del);
610 if (error)
611 break;
612 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
613 ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
614
615 /* Free the CoW orphan record. */
616 xfs_refcount_free_cow_extent(*tpp, del.br_startblock,
617 del.br_blockcount);
618
619 xfs_free_extent_later(*tpp, del.br_startblock,
620 del.br_blockcount, NULL);
621
622 /* Roll the transaction */
623 error = xfs_defer_finish(tpp);
624 if (error)
625 break;
626
627 /* Remove the mapping from the CoW fork. */
628 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
629
630 /* Remove the quota reservation */
631 error = xfs_quota_unreserve_blkres(ip,
632 del.br_blockcount);
633 if (error)
634 break;
635 } else {
636 /* Didn't do anything, push cursor back. */
637 xfs_iext_prev(ifp, &icur);
638 }
639next_extent:
640 if (!xfs_iext_get_extent(ifp, &icur, &got))
641 break;
642 }
643
644 /* clear tag if cow fork is emptied */
645 if (!ifp->if_bytes)
646 xfs_inode_clear_cowblocks_tag(ip);
647 return error;
648}
649
650/*
651 * Cancel CoW reservations for some byte range of an inode.
652 *
653 * If cancel_real is true this function cancels all COW fork extents for the
654 * inode; if cancel_real is false, real extents are not cleared.
655 */
656int
657xfs_reflink_cancel_cow_range(
658 struct xfs_inode *ip,
659 xfs_off_t offset,
660 xfs_off_t count,
661 bool cancel_real)
662{
663 struct xfs_trans *tp;
664 xfs_fileoff_t offset_fsb;
665 xfs_fileoff_t end_fsb;
666 int error;
667
668 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
669 ASSERT(ip->i_cowfp);
670
671 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
672 if (count == NULLFILEOFF)
673 end_fsb = NULLFILEOFF;
674 else
675 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
676
677 /* Start a rolling transaction to remove the mappings */
678 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
679 0, 0, 0, &tp);
680 if (error)
681 goto out;
682
683 xfs_ilock(ip, XFS_ILOCK_EXCL);
684 xfs_trans_ijoin(tp, ip, 0);
685
686 /* Scrape out the old CoW reservations */
687 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
688 cancel_real);
689 if (error)
690 goto out_cancel;
691
692 error = xfs_trans_commit(tp);
693
694 xfs_iunlock(ip, XFS_ILOCK_EXCL);
695 return error;
696
697out_cancel:
698 xfs_trans_cancel(tp);
699 xfs_iunlock(ip, XFS_ILOCK_EXCL);
700out:
701 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
702 return error;
703}
704
705/*
706 * Remap part of the CoW fork into the data fork.
707 *
708 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
709 * into the data fork; this function will remap what it can (at the end of the
710 * range) and update @end_fsb appropriately. Each remap gets its own
711 * transaction because we can end up merging and splitting bmbt blocks for
712 * every remap operation and we'd like to keep the block reservation
713 * requirements as low as possible.
714 */
715STATIC int
716xfs_reflink_end_cow_extent(
717 struct xfs_inode *ip,
718 xfs_fileoff_t *offset_fsb,
719 xfs_fileoff_t end_fsb)
720{
721 struct xfs_iext_cursor icur;
722 struct xfs_bmbt_irec got, del, data;
723 struct xfs_mount *mp = ip->i_mount;
724 struct xfs_trans *tp;
725 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
726 unsigned int resblks;
727 int nmaps;
728 int error;
729
730 /* No COW extents? That's easy! */
731 if (ifp->if_bytes == 0) {
732 *offset_fsb = end_fsb;
733 return 0;
734 }
735
736 resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
737 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
738 XFS_TRANS_RESERVE, &tp);
739 if (error)
740 return error;
741
742 /*
743 * Lock the inode. We have to ijoin without automatic unlock because
744 * the lead transaction is the refcountbt record deletion; the data
745 * fork update follows as a deferred log item.
746 */
747 xfs_ilock(ip, XFS_ILOCK_EXCL);
748 xfs_trans_ijoin(tp, ip, 0);
749
750 error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
751 XFS_IEXT_REFLINK_END_COW_CNT);
752 if (error == -EFBIG)
753 error = xfs_iext_count_upgrade(tp, ip,
754 XFS_IEXT_REFLINK_END_COW_CNT);
755 if (error)
756 goto out_cancel;
757
758 /*
759 * In case of racing, overlapping AIO writes no COW extents might be
760 * left by the time I/O completes for the loser of the race. In that
761 * case we are done.
762 */
763 if (!xfs_iext_lookup_extent(ip, ifp, *offset_fsb, &icur, &got) ||
764 got.br_startoff >= end_fsb) {
765 *offset_fsb = end_fsb;
766 goto out_cancel;
767 }
768
769 /*
770 * Only remap real extents that contain data. With AIO, speculative
771 * preallocations can leak into the range we are called upon, and we
772 * need to skip them. Preserve @got for the eventual CoW fork
773 * deletion; from now on @del represents the mapping that we're
774 * actually remapping.
775 */
776 while (!xfs_bmap_is_written_extent(&got)) {
777 if (!xfs_iext_next_extent(ifp, &icur, &got) ||
778 got.br_startoff >= end_fsb) {
779 *offset_fsb = end_fsb;
780 goto out_cancel;
781 }
782 }
783 del = got;
784
785 /* Grab the corresponding mapping in the data fork. */
786 nmaps = 1;
787 error = xfs_bmapi_read(ip, del.br_startoff, del.br_blockcount, &data,
788 &nmaps, 0);
789 if (error)
790 goto out_cancel;
791
792 /* We can only remap the smaller of the two extent sizes. */
793 data.br_blockcount = min(data.br_blockcount, del.br_blockcount);
794 del.br_blockcount = data.br_blockcount;
795
796 trace_xfs_reflink_cow_remap_from(ip, &del);
797 trace_xfs_reflink_cow_remap_to(ip, &data);
798
799 if (xfs_bmap_is_real_extent(&data)) {
800 /*
801 * If the extent we're remapping is backed by storage (written
802 * or not), unmap the extent and drop its refcount.
803 */
804 xfs_bmap_unmap_extent(tp, ip, &data);
805 xfs_refcount_decrease_extent(tp, &data);
806 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
807 -data.br_blockcount);
808 } else if (data.br_startblock == DELAYSTARTBLOCK) {
809 int done;
810
811 /*
812 * If the extent we're remapping is a delalloc reservation,
813 * we can use the regular bunmapi function to release the
814 * incore state. Dropping the delalloc reservation takes care
815 * of the quota reservation for us.
816 */
817 error = xfs_bunmapi(NULL, ip, data.br_startoff,
818 data.br_blockcount, 0, 1, &done);
819 if (error)
820 goto out_cancel;
821 ASSERT(done);
822 }
823
824 /* Free the CoW orphan record. */
825 xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);
826
827 /* Map the new blocks into the data fork. */
828 xfs_bmap_map_extent(tp, ip, &del);
829
830 /* Charge this new data fork mapping to the on-disk quota. */
831 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
832 (long)del.br_blockcount);
833
834 /* Remove the mapping from the CoW fork. */
835 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
836
837 error = xfs_trans_commit(tp);
838 xfs_iunlock(ip, XFS_ILOCK_EXCL);
839 if (error)
840 return error;
841
842 /* Update the caller about how much progress we made. */
843 *offset_fsb = del.br_startoff + del.br_blockcount;
844 return 0;
845
846out_cancel:
847 xfs_trans_cancel(tp);
848 xfs_iunlock(ip, XFS_ILOCK_EXCL);
849 return error;
850}
851
852/*
853 * Remap parts of a file's data fork after a successful CoW.
854 */
855int
856xfs_reflink_end_cow(
857 struct xfs_inode *ip,
858 xfs_off_t offset,
859 xfs_off_t count)
860{
861 xfs_fileoff_t offset_fsb;
862 xfs_fileoff_t end_fsb;
863 int error = 0;
864
865 trace_xfs_reflink_end_cow(ip, offset, count);
866
867 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
868 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
869
870 /*
871 * Walk forwards until we've remapped the I/O range. The loop function
872 * repeatedly cycles the ILOCK to allocate one transaction per remapped
873 * extent.
874 *
875 * If we're being called by writeback then the pages will still
876 * have PageWriteback set, which prevents races with reflink remapping
877 * and truncate. Reflink remapping prevents races with writeback by
878 * taking the iolock and mmaplock before flushing the pages and
879 * remapping, which means there won't be any further writeback or page
880 * cache dirtying until the reflink completes.
881 *
882 * We should never have two threads issuing writeback for the same file
883 * region. There are also have post-eof checks in the writeback
884 * preparation code so that we don't bother writing out pages that are
885 * about to be truncated.
886 *
887 * If we're being called as part of directio write completion, the dio
888 * count is still elevated, which reflink and truncate will wait for.
889 * Reflink remapping takes the iolock and mmaplock and waits for
890 * pending dio to finish, which should prevent any directio until the
891 * remap completes. Multiple concurrent directio writes to the same
892 * region are handled by end_cow processing only occurring for the
893 * threads which succeed; the outcome of multiple overlapping direct
894 * writes is not well defined anyway.
895 *
896 * It's possible that a buffered write and a direct write could collide
897 * here (the buffered write stumbles in after the dio flushes and
898 * invalidates the page cache and immediately queues writeback), but we
899 * have never supported this 100%. If either disk write succeeds the
900 * blocks will be remapped.
901 */
902 while (end_fsb > offset_fsb && !error)
903 error = xfs_reflink_end_cow_extent(ip, &offset_fsb, end_fsb);
904
905 if (error)
906 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
907 return error;
908}
909
910/*
911 * Free all CoW staging blocks that are still referenced by the ondisk refcount
912 * metadata. The ondisk metadata does not track which inode created the
913 * staging extent, so callers must ensure that there are no cached inodes with
914 * live CoW staging extents.
915 */
916int
917xfs_reflink_recover_cow(
918 struct xfs_mount *mp)
919{
920 struct xfs_perag *pag;
921 xfs_agnumber_t agno;
922 int error = 0;
923
924 if (!xfs_has_reflink(mp))
925 return 0;
926
927 for_each_perag(mp, agno, pag) {
928 error = xfs_refcount_recover_cow_leftovers(mp, pag);
929 if (error) {
930 xfs_perag_put(pag);
931 break;
932 }
933 }
934
935 return error;
936}
937
938/*
939 * Reflinking (Block) Ranges of Two Files Together
940 *
941 * First, ensure that the reflink flag is set on both inodes. The flag is an
942 * optimization to avoid unnecessary refcount btree lookups in the write path.
943 *
944 * Now we can iteratively remap the range of extents (and holes) in src to the
945 * corresponding ranges in dest. Let drange and srange denote the ranges of
946 * logical blocks in dest and src touched by the reflink operation.
947 *
948 * While the length of drange is greater than zero,
949 * - Read src's bmbt at the start of srange ("imap")
950 * - If imap doesn't exist, make imap appear to start at the end of srange
951 * with zero length.
952 * - If imap starts before srange, advance imap to start at srange.
953 * - If imap goes beyond srange, truncate imap to end at the end of srange.
954 * - Punch (imap start - srange start + imap len) blocks from dest at
955 * offset (drange start).
956 * - If imap points to a real range of pblks,
957 * > Increase the refcount of the imap's pblks
958 * > Map imap's pblks into dest at the offset
959 * (drange start + imap start - srange start)
960 * - Advance drange and srange by (imap start - srange start + imap len)
961 *
962 * Finally, if the reflink made dest longer, update both the in-core and
963 * on-disk file sizes.
964 *
965 * ASCII Art Demonstration:
966 *
967 * Let's say we want to reflink this source file:
968 *
969 * ----SSSSSSS-SSSSS----SSSSSS (src file)
970 * <-------------------->
971 *
972 * into this destination file:
973 *
974 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
975 * <-------------------->
976 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
977 * Observe that the range has different logical offsets in either file.
978 *
979 * Consider that the first extent in the source file doesn't line up with our
980 * reflink range. Unmapping and remapping are separate operations, so we can
981 * unmap more blocks from the destination file than we remap.
982 *
983 * ----SSSSSSS-SSSSS----SSSSSS
984 * <------->
985 * --DDDDD---------DDDDD--DDD
986 * <------->
987 *
988 * Now remap the source extent into the destination file:
989 *
990 * ----SSSSSSS-SSSSS----SSSSSS
991 * <------->
992 * --DDDDD--SSSSSSSDDDDD--DDD
993 * <------->
994 *
995 * Do likewise with the second hole and extent in our range. Holes in the
996 * unmap range don't affect our operation.
997 *
998 * ----SSSSSSS-SSSSS----SSSSSS
999 * <---->
1000 * --DDDDD--SSSSSSS-SSSSS-DDD
1001 * <---->
1002 *
1003 * Finally, unmap and remap part of the third extent. This will increase the
1004 * size of the destination file.
1005 *
1006 * ----SSSSSSS-SSSSS----SSSSSS
1007 * <----->
1008 * --DDDDD--SSSSSSS-SSSSS----SSS
1009 * <----->
1010 *
1011 * Once we update the destination file's i_size, we're done.
1012 */
1013
1014/*
1015 * Ensure the reflink bit is set in both inodes.
1016 */
1017STATIC int
1018xfs_reflink_set_inode_flag(
1019 struct xfs_inode *src,
1020 struct xfs_inode *dest)
1021{
1022 struct xfs_mount *mp = src->i_mount;
1023 int error;
1024 struct xfs_trans *tp;
1025
1026 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
1027 return 0;
1028
1029 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1030 if (error)
1031 goto out_error;
1032
1033 /* Lock both files against IO */
1034 if (src->i_ino == dest->i_ino)
1035 xfs_ilock(src, XFS_ILOCK_EXCL);
1036 else
1037 xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
1038
1039 if (!xfs_is_reflink_inode(src)) {
1040 trace_xfs_reflink_set_inode_flag(src);
1041 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
1042 src->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1043 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
1044 xfs_ifork_init_cow(src);
1045 } else
1046 xfs_iunlock(src, XFS_ILOCK_EXCL);
1047
1048 if (src->i_ino == dest->i_ino)
1049 goto commit_flags;
1050
1051 if (!xfs_is_reflink_inode(dest)) {
1052 trace_xfs_reflink_set_inode_flag(dest);
1053 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1054 dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1055 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1056 xfs_ifork_init_cow(dest);
1057 } else
1058 xfs_iunlock(dest, XFS_ILOCK_EXCL);
1059
1060commit_flags:
1061 error = xfs_trans_commit(tp);
1062 if (error)
1063 goto out_error;
1064 return error;
1065
1066out_error:
1067 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
1068 return error;
1069}
1070
1071/*
1072 * Update destination inode size & cowextsize hint, if necessary.
1073 */
1074int
1075xfs_reflink_update_dest(
1076 struct xfs_inode *dest,
1077 xfs_off_t newlen,
1078 xfs_extlen_t cowextsize,
1079 unsigned int remap_flags)
1080{
1081 struct xfs_mount *mp = dest->i_mount;
1082 struct xfs_trans *tp;
1083 int error;
1084
1085 if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1086 return 0;
1087
1088 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1089 if (error)
1090 goto out_error;
1091
1092 xfs_ilock(dest, XFS_ILOCK_EXCL);
1093 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1094
1095 if (newlen > i_size_read(VFS_I(dest))) {
1096 trace_xfs_reflink_update_inode_size(dest, newlen);
1097 i_size_write(VFS_I(dest), newlen);
1098 dest->i_disk_size = newlen;
1099 }
1100
1101 if (cowextsize) {
1102 dest->i_cowextsize = cowextsize;
1103 dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
1104 }
1105
1106 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1107
1108 error = xfs_trans_commit(tp);
1109 if (error)
1110 goto out_error;
1111 return error;
1112
1113out_error:
1114 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1115 return error;
1116}
1117
1118/*
1119 * Do we have enough reserve in this AG to handle a reflink? The refcount
1120 * btree already reserved all the space it needs, but the rmap btree can grow
1121 * infinitely, so we won't allow more reflinks when the AG is down to the
1122 * btree reserves.
1123 */
1124static int
1125xfs_reflink_ag_has_free_space(
1126 struct xfs_mount *mp,
1127 xfs_agnumber_t agno)
1128{
1129 struct xfs_perag *pag;
1130 int error = 0;
1131
1132 if (!xfs_has_rmapbt(mp))
1133 return 0;
1134
1135 pag = xfs_perag_get(mp, agno);
1136 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
1137 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1138 error = -ENOSPC;
1139 xfs_perag_put(pag);
1140 return error;
1141}
1142
1143/*
1144 * Remap the given extent into the file. The dmap blockcount will be set to
1145 * the number of blocks that were actually remapped.
1146 */
1147STATIC int
1148xfs_reflink_remap_extent(
1149 struct xfs_inode *ip,
1150 struct xfs_bmbt_irec *dmap,
1151 xfs_off_t new_isize)
1152{
1153 struct xfs_bmbt_irec smap;
1154 struct xfs_mount *mp = ip->i_mount;
1155 struct xfs_trans *tp;
1156 xfs_off_t newlen;
1157 int64_t qdelta = 0;
1158 unsigned int resblks;
1159 bool quota_reserved = true;
1160 bool smap_real;
1161 bool dmap_written = xfs_bmap_is_written_extent(dmap);
1162 int iext_delta = 0;
1163 int nimaps;
1164 int error;
1165
1166 /*
1167 * Start a rolling transaction to switch the mappings.
1168 *
1169 * Adding a written extent to the extent map can cause a bmbt split,
1170 * and removing a mapped extent from the extent can cause a bmbt split.
1171 * The two operations cannot both cause a split since they operate on
1172 * the same index in the bmap btree, so we only need a reservation for
1173 * one bmbt split if either thing is happening. However, we haven't
1174 * locked the inode yet, so we reserve assuming this is the case.
1175 *
1176 * The first allocation call tries to reserve enough space to handle
1177 * mapping dmap into a sparse part of the file plus the bmbt split. We
1178 * haven't locked the inode or read the existing mapping yet, so we do
1179 * not know for sure that we need the space. This should succeed most
1180 * of the time.
1181 *
1182 * If the first attempt fails, try again but reserving only enough
1183 * space to handle a bmbt split. This is the hard minimum requirement,
1184 * and we revisit quota reservations later when we know more about what
1185 * we're remapping.
1186 */
1187 resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1188 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1189 resblks + dmap->br_blockcount, 0, false, &tp);
1190 if (error == -EDQUOT || error == -ENOSPC) {
1191 quota_reserved = false;
1192 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1193 resblks, 0, false, &tp);
1194 }
1195 if (error)
1196 goto out;
1197
1198 /*
1199 * Read what's currently mapped in the destination file into smap.
1200 * If smap isn't a hole, we will have to remove it before we can add
1201 * dmap to the destination file.
1202 */
1203 nimaps = 1;
1204 error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,
1205 &smap, &nimaps, 0);
1206 if (error)
1207 goto out_cancel;
1208 ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);
1209 smap_real = xfs_bmap_is_real_extent(&smap);
1210
1211 /*
1212 * We can only remap as many blocks as the smaller of the two extent
1213 * maps, because we can only remap one extent at a time.
1214 */
1215 dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);
1216 ASSERT(dmap->br_blockcount == smap.br_blockcount);
1217
1218 trace_xfs_reflink_remap_extent_dest(ip, &smap);
1219
1220 /*
1221 * Two extents mapped to the same physical block must not have
1222 * different states; that's filesystem corruption. Move on to the next
1223 * extent if they're both holes or both the same physical extent.
1224 */
1225 if (dmap->br_startblock == smap.br_startblock) {
1226 if (dmap->br_state != smap.br_state)
1227 error = -EFSCORRUPTED;
1228 goto out_cancel;
1229 }
1230
1231 /* If both extents are unwritten, leave them alone. */
1232 if (dmap->br_state == XFS_EXT_UNWRITTEN &&
1233 smap.br_state == XFS_EXT_UNWRITTEN)
1234 goto out_cancel;
1235
1236 /* No reflinking if the AG of the dest mapping is low on space. */
1237 if (dmap_written) {
1238 error = xfs_reflink_ag_has_free_space(mp,
1239 XFS_FSB_TO_AGNO(mp, dmap->br_startblock));
1240 if (error)
1241 goto out_cancel;
1242 }
1243
1244 /*
1245 * Increase quota reservation if we think the quota block counter for
1246 * this file could increase.
1247 *
1248 * If we are mapping a written extent into the file, we need to have
1249 * enough quota block count reservation to handle the blocks in that
1250 * extent. We log only the delta to the quota block counts, so if the
1251 * extent we're unmapping also has blocks allocated to it, we don't
1252 * need a quota reservation for the extent itself.
1253 *
1254 * Note that if we're replacing a delalloc reservation with a written
1255 * extent, we have to take the full quota reservation because removing
1256 * the delalloc reservation gives the block count back to the quota
1257 * count. This is suboptimal, but the VFS flushed the dest range
1258 * before we started. That should have removed all the delalloc
1259 * reservations, but we code defensively.
1260 *
1261 * xfs_trans_alloc_inode above already tried to grab an even larger
1262 * quota reservation, and kicked off a blockgc scan if it couldn't.
1263 * If we can't get a potentially smaller quota reservation now, we're
1264 * done.
1265 */
1266 if (!quota_reserved && !smap_real && dmap_written) {
1267 error = xfs_trans_reserve_quota_nblks(tp, ip,
1268 dmap->br_blockcount, 0, false);
1269 if (error)
1270 goto out_cancel;
1271 }
1272
1273 if (smap_real)
1274 ++iext_delta;
1275
1276 if (dmap_written)
1277 ++iext_delta;
1278
1279 error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK, iext_delta);
1280 if (error == -EFBIG)
1281 error = xfs_iext_count_upgrade(tp, ip, iext_delta);
1282 if (error)
1283 goto out_cancel;
1284
1285 if (smap_real) {
1286 /*
1287 * If the extent we're unmapping is backed by storage (written
1288 * or not), unmap the extent and drop its refcount.
1289 */
1290 xfs_bmap_unmap_extent(tp, ip, &smap);
1291 xfs_refcount_decrease_extent(tp, &smap);
1292 qdelta -= smap.br_blockcount;
1293 } else if (smap.br_startblock == DELAYSTARTBLOCK) {
1294 int done;
1295
1296 /*
1297 * If the extent we're unmapping is a delalloc reservation,
1298 * we can use the regular bunmapi function to release the
1299 * incore state. Dropping the delalloc reservation takes care
1300 * of the quota reservation for us.
1301 */
1302 error = xfs_bunmapi(NULL, ip, smap.br_startoff,
1303 smap.br_blockcount, 0, 1, &done);
1304 if (error)
1305 goto out_cancel;
1306 ASSERT(done);
1307 }
1308
1309 /*
1310 * If the extent we're sharing is backed by written storage, increase
1311 * its refcount and map it into the file.
1312 */
1313 if (dmap_written) {
1314 xfs_refcount_increase_extent(tp, dmap);
1315 xfs_bmap_map_extent(tp, ip, dmap);
1316 qdelta += dmap->br_blockcount;
1317 }
1318
1319 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, qdelta);
1320
1321 /* Update dest isize if needed. */
1322 newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);
1323 newlen = min_t(xfs_off_t, newlen, new_isize);
1324 if (newlen > i_size_read(VFS_I(ip))) {
1325 trace_xfs_reflink_update_inode_size(ip, newlen);
1326 i_size_write(VFS_I(ip), newlen);
1327 ip->i_disk_size = newlen;
1328 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1329 }
1330
1331 /* Commit everything and unlock. */
1332 error = xfs_trans_commit(tp);
1333 goto out_unlock;
1334
1335out_cancel:
1336 xfs_trans_cancel(tp);
1337out_unlock:
1338 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1339out:
1340 if (error)
1341 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1342 return error;
1343}
1344
1345/* Remap a range of one file to the other. */
1346int
1347xfs_reflink_remap_blocks(
1348 struct xfs_inode *src,
1349 loff_t pos_in,
1350 struct xfs_inode *dest,
1351 loff_t pos_out,
1352 loff_t remap_len,
1353 loff_t *remapped)
1354{
1355 struct xfs_bmbt_irec imap;
1356 struct xfs_mount *mp = src->i_mount;
1357 xfs_fileoff_t srcoff = XFS_B_TO_FSBT(mp, pos_in);
1358 xfs_fileoff_t destoff = XFS_B_TO_FSBT(mp, pos_out);
1359 xfs_filblks_t len;
1360 xfs_filblks_t remapped_len = 0;
1361 xfs_off_t new_isize = pos_out + remap_len;
1362 int nimaps;
1363 int error = 0;
1364
1365 len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),
1366 XFS_MAX_FILEOFF);
1367
1368 trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);
1369
1370 while (len > 0) {
1371 unsigned int lock_mode;
1372
1373 /* Read extent from the source file */
1374 nimaps = 1;
1375 lock_mode = xfs_ilock_data_map_shared(src);
1376 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1377 xfs_iunlock(src, lock_mode);
1378 if (error)
1379 break;
1380 /*
1381 * The caller supposedly flushed all dirty pages in the source
1382 * file range, which means that writeback should have allocated
1383 * or deleted all delalloc reservations in that range. If we
1384 * find one, that's a good sign that something is seriously
1385 * wrong here.
1386 */
1387 ASSERT(nimaps == 1 && imap.br_startoff == srcoff);
1388 if (imap.br_startblock == DELAYSTARTBLOCK) {
1389 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1390 error = -EFSCORRUPTED;
1391 break;
1392 }
1393
1394 trace_xfs_reflink_remap_extent_src(src, &imap);
1395
1396 /* Remap into the destination file at the given offset. */
1397 imap.br_startoff = destoff;
1398 error = xfs_reflink_remap_extent(dest, &imap, new_isize);
1399 if (error)
1400 break;
1401
1402 if (fatal_signal_pending(current)) {
1403 error = -EINTR;
1404 break;
1405 }
1406
1407 /* Advance drange/srange */
1408 srcoff += imap.br_blockcount;
1409 destoff += imap.br_blockcount;
1410 len -= imap.br_blockcount;
1411 remapped_len += imap.br_blockcount;
1412 }
1413
1414 if (error)
1415 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1416 *remapped = min_t(loff_t, remap_len,
1417 XFS_FSB_TO_B(src->i_mount, remapped_len));
1418 return error;
1419}
1420
1421/*
1422 * If we're reflinking to a point past the destination file's EOF, we must
1423 * zero any speculative post-EOF preallocations that sit between the old EOF
1424 * and the destination file offset.
1425 */
1426static int
1427xfs_reflink_zero_posteof(
1428 struct xfs_inode *ip,
1429 loff_t pos)
1430{
1431 loff_t isize = i_size_read(VFS_I(ip));
1432
1433 if (pos <= isize)
1434 return 0;
1435
1436 trace_xfs_zero_eof(ip, isize, pos - isize);
1437 return xfs_zero_range(ip, isize, pos - isize, NULL);
1438}
1439
1440/*
1441 * Prepare two files for range cloning. Upon a successful return both inodes
1442 * will have the iolock and mmaplock held, the page cache of the out file will
1443 * be truncated, and any leases on the out file will have been broken. This
1444 * function borrows heavily from xfs_file_aio_write_checks.
1445 *
1446 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1447 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1448 * EOF block in the source dedupe range because it's not a complete block match,
1449 * hence can introduce a corruption into the file that has it's block replaced.
1450 *
1451 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1452 * "block aligned" for the purposes of cloning entire files. However, if the
1453 * source file range includes the EOF block and it lands within the existing EOF
1454 * of the destination file, then we can expose stale data from beyond the source
1455 * file EOF in the destination file.
1456 *
1457 * XFS doesn't support partial block sharing, so in both cases we have check
1458 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1459 * down to the previous whole block and ignore the partial EOF block. While this
1460 * means we can't dedupe the last block of a file, this is an acceptible
1461 * tradeoff for simplicity on implementation.
1462 *
1463 * For cloning, we want to share the partial EOF block if it is also the new EOF
1464 * block of the destination file. If the partial EOF block lies inside the
1465 * existing destination EOF, then we have to abort the clone to avoid exposing
1466 * stale data in the destination file. Hence we reject these clone attempts with
1467 * -EINVAL in this case.
1468 */
1469int
1470xfs_reflink_remap_prep(
1471 struct file *file_in,
1472 loff_t pos_in,
1473 struct file *file_out,
1474 loff_t pos_out,
1475 loff_t *len,
1476 unsigned int remap_flags)
1477{
1478 struct inode *inode_in = file_inode(file_in);
1479 struct xfs_inode *src = XFS_I(inode_in);
1480 struct inode *inode_out = file_inode(file_out);
1481 struct xfs_inode *dest = XFS_I(inode_out);
1482 int ret;
1483
1484 /* Lock both files against IO */
1485 ret = xfs_ilock2_io_mmap(src, dest);
1486 if (ret)
1487 return ret;
1488
1489 /* Check file eligibility and prepare for block sharing. */
1490 ret = -EINVAL;
1491 /* Don't reflink realtime inodes */
1492 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1493 goto out_unlock;
1494
1495 /* Don't share DAX file data with non-DAX file. */
1496 if (IS_DAX(inode_in) != IS_DAX(inode_out))
1497 goto out_unlock;
1498
1499 if (!IS_DAX(inode_in))
1500 ret = generic_remap_file_range_prep(file_in, pos_in, file_out,
1501 pos_out, len, remap_flags);
1502 else
1503 ret = dax_remap_file_range_prep(file_in, pos_in, file_out,
1504 pos_out, len, remap_flags, &xfs_read_iomap_ops);
1505 if (ret || *len == 0)
1506 goto out_unlock;
1507
1508 /* Attach dquots to dest inode before changing block map */
1509 ret = xfs_qm_dqattach(dest);
1510 if (ret)
1511 goto out_unlock;
1512
1513 /*
1514 * Zero existing post-eof speculative preallocations in the destination
1515 * file.
1516 */
1517 ret = xfs_reflink_zero_posteof(dest, pos_out);
1518 if (ret)
1519 goto out_unlock;
1520
1521 /* Set flags and remap blocks. */
1522 ret = xfs_reflink_set_inode_flag(src, dest);
1523 if (ret)
1524 goto out_unlock;
1525
1526 /*
1527 * If pos_out > EOF, we may have dirtied blocks between EOF and
1528 * pos_out. In that case, we need to extend the flush and unmap to cover
1529 * from EOF to the end of the copy length.
1530 */
1531 if (pos_out > XFS_ISIZE(dest)) {
1532 loff_t flen = *len + (pos_out - XFS_ISIZE(dest));
1533 ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1534 } else {
1535 ret = xfs_flush_unmap_range(dest, pos_out, *len);
1536 }
1537 if (ret)
1538 goto out_unlock;
1539
1540 return 0;
1541out_unlock:
1542 xfs_iunlock2_io_mmap(src, dest);
1543 return ret;
1544}
1545
1546/* Does this inode need the reflink flag? */
1547int
1548xfs_reflink_inode_has_shared_extents(
1549 struct xfs_trans *tp,
1550 struct xfs_inode *ip,
1551 bool *has_shared)
1552{
1553 struct xfs_bmbt_irec got;
1554 struct xfs_mount *mp = ip->i_mount;
1555 struct xfs_ifork *ifp;
1556 struct xfs_iext_cursor icur;
1557 bool found;
1558 int error;
1559
1560 ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1561 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1562 if (error)
1563 return error;
1564
1565 *has_shared = false;
1566 found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1567 while (found) {
1568 struct xfs_perag *pag;
1569 xfs_agblock_t agbno;
1570 xfs_extlen_t aglen;
1571 xfs_agblock_t rbno;
1572 xfs_extlen_t rlen;
1573
1574 if (isnullstartblock(got.br_startblock) ||
1575 got.br_state != XFS_EXT_NORM)
1576 goto next;
1577
1578 pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, got.br_startblock));
1579 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1580 aglen = got.br_blockcount;
1581 error = xfs_reflink_find_shared(pag, tp, agbno, aglen,
1582 &rbno, &rlen, false);
1583 xfs_perag_put(pag);
1584 if (error)
1585 return error;
1586
1587 /* Is there still a shared block here? */
1588 if (rbno != NULLAGBLOCK) {
1589 *has_shared = true;
1590 return 0;
1591 }
1592next:
1593 found = xfs_iext_next_extent(ifp, &icur, &got);
1594 }
1595
1596 return 0;
1597}
1598
1599/*
1600 * Clear the inode reflink flag if there are no shared extents.
1601 *
1602 * The caller is responsible for joining the inode to the transaction passed in.
1603 * The inode will be joined to the transaction that is returned to the caller.
1604 */
1605int
1606xfs_reflink_clear_inode_flag(
1607 struct xfs_inode *ip,
1608 struct xfs_trans **tpp)
1609{
1610 bool needs_flag;
1611 int error = 0;
1612
1613 ASSERT(xfs_is_reflink_inode(ip));
1614
1615 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1616 if (error || needs_flag)
1617 return error;
1618
1619 /*
1620 * We didn't find any shared blocks so turn off the reflink flag.
1621 * First, get rid of any leftover CoW mappings.
1622 */
1623 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,
1624 true);
1625 if (error)
1626 return error;
1627
1628 /* Clear the inode flag. */
1629 trace_xfs_reflink_unset_inode_flag(ip);
1630 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1631 xfs_inode_clear_cowblocks_tag(ip);
1632 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1633
1634 return error;
1635}
1636
1637/*
1638 * Clear the inode reflink flag if there are no shared extents and the size
1639 * hasn't changed.
1640 */
1641STATIC int
1642xfs_reflink_try_clear_inode_flag(
1643 struct xfs_inode *ip)
1644{
1645 struct xfs_mount *mp = ip->i_mount;
1646 struct xfs_trans *tp;
1647 int error = 0;
1648
1649 /* Start a rolling transaction to remove the mappings */
1650 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1651 if (error)
1652 return error;
1653
1654 xfs_ilock(ip, XFS_ILOCK_EXCL);
1655 xfs_trans_ijoin(tp, ip, 0);
1656
1657 error = xfs_reflink_clear_inode_flag(ip, &tp);
1658 if (error)
1659 goto cancel;
1660
1661 error = xfs_trans_commit(tp);
1662 if (error)
1663 goto out;
1664
1665 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1666 return 0;
1667cancel:
1668 xfs_trans_cancel(tp);
1669out:
1670 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1671 return error;
1672}
1673
1674/*
1675 * Pre-COW all shared blocks within a given byte range of a file and turn off
1676 * the reflink flag if we unshare all of the file's blocks.
1677 */
1678int
1679xfs_reflink_unshare(
1680 struct xfs_inode *ip,
1681 xfs_off_t offset,
1682 xfs_off_t len)
1683{
1684 struct inode *inode = VFS_I(ip);
1685 int error;
1686
1687 if (!xfs_is_reflink_inode(ip))
1688 return 0;
1689
1690 trace_xfs_reflink_unshare(ip, offset, len);
1691
1692 inode_dio_wait(inode);
1693
1694 if (IS_DAX(inode))
1695 error = dax_file_unshare(inode, offset, len,
1696 &xfs_dax_write_iomap_ops);
1697 else
1698 error = iomap_file_unshare(inode, offset, len,
1699 &xfs_buffered_write_iomap_ops);
1700 if (error)
1701 goto out;
1702
1703 error = filemap_write_and_wait_range(inode->i_mapping, offset,
1704 offset + len - 1);
1705 if (error)
1706 goto out;
1707
1708 /* Turn off the reflink flag if possible. */
1709 error = xfs_reflink_try_clear_inode_flag(ip);
1710 if (error)
1711 goto out;
1712 return 0;
1713
1714out:
1715 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1716 return error;
1717}