Loading...
Note: File does not exist in v3.1.
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_highest_agno == NULLAGNUMBER);
614
615 /* Free the CoW orphan record. */
616 xfs_refcount_free_cow_extent(*tpp, del.br_startblock,
617 del.br_blockcount);
618
619 error = xfs_free_extent_later(*tpp, del.br_startblock,
620 del.br_blockcount, NULL,
621 XFS_AG_RESV_NONE, false);
622 if (error)
623 break;
624
625 /* Roll the transaction */
626 error = xfs_defer_finish(tpp);
627 if (error)
628 break;
629
630 /* Remove the mapping from the CoW fork. */
631 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
632
633 /* Remove the quota reservation */
634 error = xfs_quota_unreserve_blkres(ip,
635 del.br_blockcount);
636 if (error)
637 break;
638 } else {
639 /* Didn't do anything, push cursor back. */
640 xfs_iext_prev(ifp, &icur);
641 }
642next_extent:
643 if (!xfs_iext_get_extent(ifp, &icur, &got))
644 break;
645 }
646
647 /* clear tag if cow fork is emptied */
648 if (!ifp->if_bytes)
649 xfs_inode_clear_cowblocks_tag(ip);
650 return error;
651}
652
653/*
654 * Cancel CoW reservations for some byte range of an inode.
655 *
656 * If cancel_real is true this function cancels all COW fork extents for the
657 * inode; if cancel_real is false, real extents are not cleared.
658 */
659int
660xfs_reflink_cancel_cow_range(
661 struct xfs_inode *ip,
662 xfs_off_t offset,
663 xfs_off_t count,
664 bool cancel_real)
665{
666 struct xfs_trans *tp;
667 xfs_fileoff_t offset_fsb;
668 xfs_fileoff_t end_fsb;
669 int error;
670
671 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
672 ASSERT(ip->i_cowfp);
673
674 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
675 if (count == NULLFILEOFF)
676 end_fsb = NULLFILEOFF;
677 else
678 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
679
680 /* Start a rolling transaction to remove the mappings */
681 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
682 0, 0, 0, &tp);
683 if (error)
684 goto out;
685
686 xfs_ilock(ip, XFS_ILOCK_EXCL);
687 xfs_trans_ijoin(tp, ip, 0);
688
689 /* Scrape out the old CoW reservations */
690 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
691 cancel_real);
692 if (error)
693 goto out_cancel;
694
695 error = xfs_trans_commit(tp);
696
697 xfs_iunlock(ip, XFS_ILOCK_EXCL);
698 return error;
699
700out_cancel:
701 xfs_trans_cancel(tp);
702 xfs_iunlock(ip, XFS_ILOCK_EXCL);
703out:
704 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
705 return error;
706}
707
708/*
709 * Remap part of the CoW fork into the data fork.
710 *
711 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
712 * into the data fork; this function will remap what it can (at the end of the
713 * range) and update @end_fsb appropriately. Each remap gets its own
714 * transaction because we can end up merging and splitting bmbt blocks for
715 * every remap operation and we'd like to keep the block reservation
716 * requirements as low as possible.
717 */
718STATIC int
719xfs_reflink_end_cow_extent(
720 struct xfs_inode *ip,
721 xfs_fileoff_t *offset_fsb,
722 xfs_fileoff_t end_fsb)
723{
724 struct xfs_iext_cursor icur;
725 struct xfs_bmbt_irec got, del, data;
726 struct xfs_mount *mp = ip->i_mount;
727 struct xfs_trans *tp;
728 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
729 unsigned int resblks;
730 int nmaps;
731 int error;
732
733 /* No COW extents? That's easy! */
734 if (ifp->if_bytes == 0) {
735 *offset_fsb = end_fsb;
736 return 0;
737 }
738
739 resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
740 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
741 XFS_TRANS_RESERVE, &tp);
742 if (error)
743 return error;
744
745 /*
746 * Lock the inode. We have to ijoin without automatic unlock because
747 * the lead transaction is the refcountbt record deletion; the data
748 * fork update follows as a deferred log item.
749 */
750 xfs_ilock(ip, XFS_ILOCK_EXCL);
751 xfs_trans_ijoin(tp, ip, 0);
752
753 error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
754 XFS_IEXT_REFLINK_END_COW_CNT);
755 if (error == -EFBIG)
756 error = xfs_iext_count_upgrade(tp, ip,
757 XFS_IEXT_REFLINK_END_COW_CNT);
758 if (error)
759 goto out_cancel;
760
761 /*
762 * In case of racing, overlapping AIO writes no COW extents might be
763 * left by the time I/O completes for the loser of the race. In that
764 * case we are done.
765 */
766 if (!xfs_iext_lookup_extent(ip, ifp, *offset_fsb, &icur, &got) ||
767 got.br_startoff >= end_fsb) {
768 *offset_fsb = end_fsb;
769 goto out_cancel;
770 }
771
772 /*
773 * Only remap real extents that contain data. With AIO, speculative
774 * preallocations can leak into the range we are called upon, and we
775 * need to skip them. Preserve @got for the eventual CoW fork
776 * deletion; from now on @del represents the mapping that we're
777 * actually remapping.
778 */
779 while (!xfs_bmap_is_written_extent(&got)) {
780 if (!xfs_iext_next_extent(ifp, &icur, &got) ||
781 got.br_startoff >= end_fsb) {
782 *offset_fsb = end_fsb;
783 goto out_cancel;
784 }
785 }
786 del = got;
787 xfs_trim_extent(&del, *offset_fsb, end_fsb - *offset_fsb);
788
789 /* Grab the corresponding mapping in the data fork. */
790 nmaps = 1;
791 error = xfs_bmapi_read(ip, del.br_startoff, del.br_blockcount, &data,
792 &nmaps, 0);
793 if (error)
794 goto out_cancel;
795
796 /* We can only remap the smaller of the two extent sizes. */
797 data.br_blockcount = min(data.br_blockcount, del.br_blockcount);
798 del.br_blockcount = data.br_blockcount;
799
800 trace_xfs_reflink_cow_remap_from(ip, &del);
801 trace_xfs_reflink_cow_remap_to(ip, &data);
802
803 if (xfs_bmap_is_real_extent(&data)) {
804 /*
805 * If the extent we're remapping is backed by storage (written
806 * or not), unmap the extent and drop its refcount.
807 */
808 xfs_bmap_unmap_extent(tp, ip, &data);
809 xfs_refcount_decrease_extent(tp, &data);
810 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
811 -data.br_blockcount);
812 } else if (data.br_startblock == DELAYSTARTBLOCK) {
813 int done;
814
815 /*
816 * If the extent we're remapping is a delalloc reservation,
817 * we can use the regular bunmapi function to release the
818 * incore state. Dropping the delalloc reservation takes care
819 * of the quota reservation for us.
820 */
821 error = xfs_bunmapi(NULL, ip, data.br_startoff,
822 data.br_blockcount, 0, 1, &done);
823 if (error)
824 goto out_cancel;
825 ASSERT(done);
826 }
827
828 /* Free the CoW orphan record. */
829 xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);
830
831 /* Map the new blocks into the data fork. */
832 xfs_bmap_map_extent(tp, ip, &del);
833
834 /* Charge this new data fork mapping to the on-disk quota. */
835 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
836 (long)del.br_blockcount);
837
838 /* Remove the mapping from the CoW fork. */
839 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
840
841 error = xfs_trans_commit(tp);
842 xfs_iunlock(ip, XFS_ILOCK_EXCL);
843 if (error)
844 return error;
845
846 /* Update the caller about how much progress we made. */
847 *offset_fsb = del.br_startoff + del.br_blockcount;
848 return 0;
849
850out_cancel:
851 xfs_trans_cancel(tp);
852 xfs_iunlock(ip, XFS_ILOCK_EXCL);
853 return error;
854}
855
856/*
857 * Remap parts of a file's data fork after a successful CoW.
858 */
859int
860xfs_reflink_end_cow(
861 struct xfs_inode *ip,
862 xfs_off_t offset,
863 xfs_off_t count)
864{
865 xfs_fileoff_t offset_fsb;
866 xfs_fileoff_t end_fsb;
867 int error = 0;
868
869 trace_xfs_reflink_end_cow(ip, offset, count);
870
871 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
872 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
873
874 /*
875 * Walk forwards until we've remapped the I/O range. The loop function
876 * repeatedly cycles the ILOCK to allocate one transaction per remapped
877 * extent.
878 *
879 * If we're being called by writeback then the pages will still
880 * have PageWriteback set, which prevents races with reflink remapping
881 * and truncate. Reflink remapping prevents races with writeback by
882 * taking the iolock and mmaplock before flushing the pages and
883 * remapping, which means there won't be any further writeback or page
884 * cache dirtying until the reflink completes.
885 *
886 * We should never have two threads issuing writeback for the same file
887 * region. There are also have post-eof checks in the writeback
888 * preparation code so that we don't bother writing out pages that are
889 * about to be truncated.
890 *
891 * If we're being called as part of directio write completion, the dio
892 * count is still elevated, which reflink and truncate will wait for.
893 * Reflink remapping takes the iolock and mmaplock and waits for
894 * pending dio to finish, which should prevent any directio until the
895 * remap completes. Multiple concurrent directio writes to the same
896 * region are handled by end_cow processing only occurring for the
897 * threads which succeed; the outcome of multiple overlapping direct
898 * writes is not well defined anyway.
899 *
900 * It's possible that a buffered write and a direct write could collide
901 * here (the buffered write stumbles in after the dio flushes and
902 * invalidates the page cache and immediately queues writeback), but we
903 * have never supported this 100%. If either disk write succeeds the
904 * blocks will be remapped.
905 */
906 while (end_fsb > offset_fsb && !error)
907 error = xfs_reflink_end_cow_extent(ip, &offset_fsb, end_fsb);
908
909 if (error)
910 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
911 return error;
912}
913
914/*
915 * Free all CoW staging blocks that are still referenced by the ondisk refcount
916 * metadata. The ondisk metadata does not track which inode created the
917 * staging extent, so callers must ensure that there are no cached inodes with
918 * live CoW staging extents.
919 */
920int
921xfs_reflink_recover_cow(
922 struct xfs_mount *mp)
923{
924 struct xfs_perag *pag;
925 xfs_agnumber_t agno;
926 int error = 0;
927
928 if (!xfs_has_reflink(mp))
929 return 0;
930
931 for_each_perag(mp, agno, pag) {
932 error = xfs_refcount_recover_cow_leftovers(mp, pag);
933 if (error) {
934 xfs_perag_rele(pag);
935 break;
936 }
937 }
938
939 return error;
940}
941
942/*
943 * Reflinking (Block) Ranges of Two Files Together
944 *
945 * First, ensure that the reflink flag is set on both inodes. The flag is an
946 * optimization to avoid unnecessary refcount btree lookups in the write path.
947 *
948 * Now we can iteratively remap the range of extents (and holes) in src to the
949 * corresponding ranges in dest. Let drange and srange denote the ranges of
950 * logical blocks in dest and src touched by the reflink operation.
951 *
952 * While the length of drange is greater than zero,
953 * - Read src's bmbt at the start of srange ("imap")
954 * - If imap doesn't exist, make imap appear to start at the end of srange
955 * with zero length.
956 * - If imap starts before srange, advance imap to start at srange.
957 * - If imap goes beyond srange, truncate imap to end at the end of srange.
958 * - Punch (imap start - srange start + imap len) blocks from dest at
959 * offset (drange start).
960 * - If imap points to a real range of pblks,
961 * > Increase the refcount of the imap's pblks
962 * > Map imap's pblks into dest at the offset
963 * (drange start + imap start - srange start)
964 * - Advance drange and srange by (imap start - srange start + imap len)
965 *
966 * Finally, if the reflink made dest longer, update both the in-core and
967 * on-disk file sizes.
968 *
969 * ASCII Art Demonstration:
970 *
971 * Let's say we want to reflink this source file:
972 *
973 * ----SSSSSSS-SSSSS----SSSSSS (src file)
974 * <-------------------->
975 *
976 * into this destination file:
977 *
978 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
979 * <-------------------->
980 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
981 * Observe that the range has different logical offsets in either file.
982 *
983 * Consider that the first extent in the source file doesn't line up with our
984 * reflink range. Unmapping and remapping are separate operations, so we can
985 * unmap more blocks from the destination file than we remap.
986 *
987 * ----SSSSSSS-SSSSS----SSSSSS
988 * <------->
989 * --DDDDD---------DDDDD--DDD
990 * <------->
991 *
992 * Now remap the source extent into the destination file:
993 *
994 * ----SSSSSSS-SSSSS----SSSSSS
995 * <------->
996 * --DDDDD--SSSSSSSDDDDD--DDD
997 * <------->
998 *
999 * Do likewise with the second hole and extent in our range. Holes in the
1000 * unmap range don't affect our operation.
1001 *
1002 * ----SSSSSSS-SSSSS----SSSSSS
1003 * <---->
1004 * --DDDDD--SSSSSSS-SSSSS-DDD
1005 * <---->
1006 *
1007 * Finally, unmap and remap part of the third extent. This will increase the
1008 * size of the destination file.
1009 *
1010 * ----SSSSSSS-SSSSS----SSSSSS
1011 * <----->
1012 * --DDDDD--SSSSSSS-SSSSS----SSS
1013 * <----->
1014 *
1015 * Once we update the destination file's i_size, we're done.
1016 */
1017
1018/*
1019 * Ensure the reflink bit is set in both inodes.
1020 */
1021STATIC int
1022xfs_reflink_set_inode_flag(
1023 struct xfs_inode *src,
1024 struct xfs_inode *dest)
1025{
1026 struct xfs_mount *mp = src->i_mount;
1027 int error;
1028 struct xfs_trans *tp;
1029
1030 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
1031 return 0;
1032
1033 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1034 if (error)
1035 goto out_error;
1036
1037 /* Lock both files against IO */
1038 if (src->i_ino == dest->i_ino)
1039 xfs_ilock(src, XFS_ILOCK_EXCL);
1040 else
1041 xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
1042
1043 if (!xfs_is_reflink_inode(src)) {
1044 trace_xfs_reflink_set_inode_flag(src);
1045 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
1046 src->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1047 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
1048 xfs_ifork_init_cow(src);
1049 } else
1050 xfs_iunlock(src, XFS_ILOCK_EXCL);
1051
1052 if (src->i_ino == dest->i_ino)
1053 goto commit_flags;
1054
1055 if (!xfs_is_reflink_inode(dest)) {
1056 trace_xfs_reflink_set_inode_flag(dest);
1057 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1058 dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1059 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1060 xfs_ifork_init_cow(dest);
1061 } else
1062 xfs_iunlock(dest, XFS_ILOCK_EXCL);
1063
1064commit_flags:
1065 error = xfs_trans_commit(tp);
1066 if (error)
1067 goto out_error;
1068 return error;
1069
1070out_error:
1071 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
1072 return error;
1073}
1074
1075/*
1076 * Update destination inode size & cowextsize hint, if necessary.
1077 */
1078int
1079xfs_reflink_update_dest(
1080 struct xfs_inode *dest,
1081 xfs_off_t newlen,
1082 xfs_extlen_t cowextsize,
1083 unsigned int remap_flags)
1084{
1085 struct xfs_mount *mp = dest->i_mount;
1086 struct xfs_trans *tp;
1087 int error;
1088
1089 if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1090 return 0;
1091
1092 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1093 if (error)
1094 goto out_error;
1095
1096 xfs_ilock(dest, XFS_ILOCK_EXCL);
1097 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1098
1099 if (newlen > i_size_read(VFS_I(dest))) {
1100 trace_xfs_reflink_update_inode_size(dest, newlen);
1101 i_size_write(VFS_I(dest), newlen);
1102 dest->i_disk_size = newlen;
1103 }
1104
1105 if (cowextsize) {
1106 dest->i_cowextsize = cowextsize;
1107 dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
1108 }
1109
1110 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1111
1112 error = xfs_trans_commit(tp);
1113 if (error)
1114 goto out_error;
1115 return error;
1116
1117out_error:
1118 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1119 return error;
1120}
1121
1122/*
1123 * Do we have enough reserve in this AG to handle a reflink? The refcount
1124 * btree already reserved all the space it needs, but the rmap btree can grow
1125 * infinitely, so we won't allow more reflinks when the AG is down to the
1126 * btree reserves.
1127 */
1128static int
1129xfs_reflink_ag_has_free_space(
1130 struct xfs_mount *mp,
1131 xfs_agnumber_t agno)
1132{
1133 struct xfs_perag *pag;
1134 int error = 0;
1135
1136 if (!xfs_has_rmapbt(mp))
1137 return 0;
1138
1139 pag = xfs_perag_get(mp, agno);
1140 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
1141 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1142 error = -ENOSPC;
1143 xfs_perag_put(pag);
1144 return error;
1145}
1146
1147/*
1148 * Remap the given extent into the file. The dmap blockcount will be set to
1149 * the number of blocks that were actually remapped.
1150 */
1151STATIC int
1152xfs_reflink_remap_extent(
1153 struct xfs_inode *ip,
1154 struct xfs_bmbt_irec *dmap,
1155 xfs_off_t new_isize)
1156{
1157 struct xfs_bmbt_irec smap;
1158 struct xfs_mount *mp = ip->i_mount;
1159 struct xfs_trans *tp;
1160 xfs_off_t newlen;
1161 int64_t qdelta = 0;
1162 unsigned int resblks;
1163 bool quota_reserved = true;
1164 bool smap_real;
1165 bool dmap_written = xfs_bmap_is_written_extent(dmap);
1166 int iext_delta = 0;
1167 int nimaps;
1168 int error;
1169
1170 /*
1171 * Start a rolling transaction to switch the mappings.
1172 *
1173 * Adding a written extent to the extent map can cause a bmbt split,
1174 * and removing a mapped extent from the extent can cause a bmbt split.
1175 * The two operations cannot both cause a split since they operate on
1176 * the same index in the bmap btree, so we only need a reservation for
1177 * one bmbt split if either thing is happening. However, we haven't
1178 * locked the inode yet, so we reserve assuming this is the case.
1179 *
1180 * The first allocation call tries to reserve enough space to handle
1181 * mapping dmap into a sparse part of the file plus the bmbt split. We
1182 * haven't locked the inode or read the existing mapping yet, so we do
1183 * not know for sure that we need the space. This should succeed most
1184 * of the time.
1185 *
1186 * If the first attempt fails, try again but reserving only enough
1187 * space to handle a bmbt split. This is the hard minimum requirement,
1188 * and we revisit quota reservations later when we know more about what
1189 * we're remapping.
1190 */
1191 resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1192 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1193 resblks + dmap->br_blockcount, 0, false, &tp);
1194 if (error == -EDQUOT || error == -ENOSPC) {
1195 quota_reserved = false;
1196 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1197 resblks, 0, false, &tp);
1198 }
1199 if (error)
1200 goto out;
1201
1202 /*
1203 * Read what's currently mapped in the destination file into smap.
1204 * If smap isn't a hole, we will have to remove it before we can add
1205 * dmap to the destination file.
1206 */
1207 nimaps = 1;
1208 error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,
1209 &smap, &nimaps, 0);
1210 if (error)
1211 goto out_cancel;
1212 ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);
1213 smap_real = xfs_bmap_is_real_extent(&smap);
1214
1215 /*
1216 * We can only remap as many blocks as the smaller of the two extent
1217 * maps, because we can only remap one extent at a time.
1218 */
1219 dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);
1220 ASSERT(dmap->br_blockcount == smap.br_blockcount);
1221
1222 trace_xfs_reflink_remap_extent_dest(ip, &smap);
1223
1224 /*
1225 * Two extents mapped to the same physical block must not have
1226 * different states; that's filesystem corruption. Move on to the next
1227 * extent if they're both holes or both the same physical extent.
1228 */
1229 if (dmap->br_startblock == smap.br_startblock) {
1230 if (dmap->br_state != smap.br_state)
1231 error = -EFSCORRUPTED;
1232 goto out_cancel;
1233 }
1234
1235 /* If both extents are unwritten, leave them alone. */
1236 if (dmap->br_state == XFS_EXT_UNWRITTEN &&
1237 smap.br_state == XFS_EXT_UNWRITTEN)
1238 goto out_cancel;
1239
1240 /* No reflinking if the AG of the dest mapping is low on space. */
1241 if (dmap_written) {
1242 error = xfs_reflink_ag_has_free_space(mp,
1243 XFS_FSB_TO_AGNO(mp, dmap->br_startblock));
1244 if (error)
1245 goto out_cancel;
1246 }
1247
1248 /*
1249 * Increase quota reservation if we think the quota block counter for
1250 * this file could increase.
1251 *
1252 * If we are mapping a written extent into the file, we need to have
1253 * enough quota block count reservation to handle the blocks in that
1254 * extent. We log only the delta to the quota block counts, so if the
1255 * extent we're unmapping also has blocks allocated to it, we don't
1256 * need a quota reservation for the extent itself.
1257 *
1258 * Note that if we're replacing a delalloc reservation with a written
1259 * extent, we have to take the full quota reservation because removing
1260 * the delalloc reservation gives the block count back to the quota
1261 * count. This is suboptimal, but the VFS flushed the dest range
1262 * before we started. That should have removed all the delalloc
1263 * reservations, but we code defensively.
1264 *
1265 * xfs_trans_alloc_inode above already tried to grab an even larger
1266 * quota reservation, and kicked off a blockgc scan if it couldn't.
1267 * If we can't get a potentially smaller quota reservation now, we're
1268 * done.
1269 */
1270 if (!quota_reserved && !smap_real && dmap_written) {
1271 error = xfs_trans_reserve_quota_nblks(tp, ip,
1272 dmap->br_blockcount, 0, false);
1273 if (error)
1274 goto out_cancel;
1275 }
1276
1277 if (smap_real)
1278 ++iext_delta;
1279
1280 if (dmap_written)
1281 ++iext_delta;
1282
1283 error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK, iext_delta);
1284 if (error == -EFBIG)
1285 error = xfs_iext_count_upgrade(tp, ip, iext_delta);
1286 if (error)
1287 goto out_cancel;
1288
1289 if (smap_real) {
1290 /*
1291 * If the extent we're unmapping is backed by storage (written
1292 * or not), unmap the extent and drop its refcount.
1293 */
1294 xfs_bmap_unmap_extent(tp, ip, &smap);
1295 xfs_refcount_decrease_extent(tp, &smap);
1296 qdelta -= smap.br_blockcount;
1297 } else if (smap.br_startblock == DELAYSTARTBLOCK) {
1298 int done;
1299
1300 /*
1301 * If the extent we're unmapping is a delalloc reservation,
1302 * we can use the regular bunmapi function to release the
1303 * incore state. Dropping the delalloc reservation takes care
1304 * of the quota reservation for us.
1305 */
1306 error = xfs_bunmapi(NULL, ip, smap.br_startoff,
1307 smap.br_blockcount, 0, 1, &done);
1308 if (error)
1309 goto out_cancel;
1310 ASSERT(done);
1311 }
1312
1313 /*
1314 * If the extent we're sharing is backed by written storage, increase
1315 * its refcount and map it into the file.
1316 */
1317 if (dmap_written) {
1318 xfs_refcount_increase_extent(tp, dmap);
1319 xfs_bmap_map_extent(tp, ip, dmap);
1320 qdelta += dmap->br_blockcount;
1321 }
1322
1323 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, qdelta);
1324
1325 /* Update dest isize if needed. */
1326 newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);
1327 newlen = min_t(xfs_off_t, newlen, new_isize);
1328 if (newlen > i_size_read(VFS_I(ip))) {
1329 trace_xfs_reflink_update_inode_size(ip, newlen);
1330 i_size_write(VFS_I(ip), newlen);
1331 ip->i_disk_size = newlen;
1332 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1333 }
1334
1335 /* Commit everything and unlock. */
1336 error = xfs_trans_commit(tp);
1337 goto out_unlock;
1338
1339out_cancel:
1340 xfs_trans_cancel(tp);
1341out_unlock:
1342 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1343out:
1344 if (error)
1345 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1346 return error;
1347}
1348
1349/* Remap a range of one file to the other. */
1350int
1351xfs_reflink_remap_blocks(
1352 struct xfs_inode *src,
1353 loff_t pos_in,
1354 struct xfs_inode *dest,
1355 loff_t pos_out,
1356 loff_t remap_len,
1357 loff_t *remapped)
1358{
1359 struct xfs_bmbt_irec imap;
1360 struct xfs_mount *mp = src->i_mount;
1361 xfs_fileoff_t srcoff = XFS_B_TO_FSBT(mp, pos_in);
1362 xfs_fileoff_t destoff = XFS_B_TO_FSBT(mp, pos_out);
1363 xfs_filblks_t len;
1364 xfs_filblks_t remapped_len = 0;
1365 xfs_off_t new_isize = pos_out + remap_len;
1366 int nimaps;
1367 int error = 0;
1368
1369 len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),
1370 XFS_MAX_FILEOFF);
1371
1372 trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);
1373
1374 while (len > 0) {
1375 unsigned int lock_mode;
1376
1377 /* Read extent from the source file */
1378 nimaps = 1;
1379 lock_mode = xfs_ilock_data_map_shared(src);
1380 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1381 xfs_iunlock(src, lock_mode);
1382 if (error)
1383 break;
1384 /*
1385 * The caller supposedly flushed all dirty pages in the source
1386 * file range, which means that writeback should have allocated
1387 * or deleted all delalloc reservations in that range. If we
1388 * find one, that's a good sign that something is seriously
1389 * wrong here.
1390 */
1391 ASSERT(nimaps == 1 && imap.br_startoff == srcoff);
1392 if (imap.br_startblock == DELAYSTARTBLOCK) {
1393 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1394 error = -EFSCORRUPTED;
1395 break;
1396 }
1397
1398 trace_xfs_reflink_remap_extent_src(src, &imap);
1399
1400 /* Remap into the destination file at the given offset. */
1401 imap.br_startoff = destoff;
1402 error = xfs_reflink_remap_extent(dest, &imap, new_isize);
1403 if (error)
1404 break;
1405
1406 if (fatal_signal_pending(current)) {
1407 error = -EINTR;
1408 break;
1409 }
1410
1411 /* Advance drange/srange */
1412 srcoff += imap.br_blockcount;
1413 destoff += imap.br_blockcount;
1414 len -= imap.br_blockcount;
1415 remapped_len += imap.br_blockcount;
1416 }
1417
1418 if (error)
1419 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1420 *remapped = min_t(loff_t, remap_len,
1421 XFS_FSB_TO_B(src->i_mount, remapped_len));
1422 return error;
1423}
1424
1425/*
1426 * If we're reflinking to a point past the destination file's EOF, we must
1427 * zero any speculative post-EOF preallocations that sit between the old EOF
1428 * and the destination file offset.
1429 */
1430static int
1431xfs_reflink_zero_posteof(
1432 struct xfs_inode *ip,
1433 loff_t pos)
1434{
1435 loff_t isize = i_size_read(VFS_I(ip));
1436
1437 if (pos <= isize)
1438 return 0;
1439
1440 trace_xfs_zero_eof(ip, isize, pos - isize);
1441 return xfs_zero_range(ip, isize, pos - isize, NULL);
1442}
1443
1444/*
1445 * Prepare two files for range cloning. Upon a successful return both inodes
1446 * will have the iolock and mmaplock held, the page cache of the out file will
1447 * be truncated, and any leases on the out file will have been broken. This
1448 * function borrows heavily from xfs_file_aio_write_checks.
1449 *
1450 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1451 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1452 * EOF block in the source dedupe range because it's not a complete block match,
1453 * hence can introduce a corruption into the file that has it's block replaced.
1454 *
1455 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1456 * "block aligned" for the purposes of cloning entire files. However, if the
1457 * source file range includes the EOF block and it lands within the existing EOF
1458 * of the destination file, then we can expose stale data from beyond the source
1459 * file EOF in the destination file.
1460 *
1461 * XFS doesn't support partial block sharing, so in both cases we have check
1462 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1463 * down to the previous whole block and ignore the partial EOF block. While this
1464 * means we can't dedupe the last block of a file, this is an acceptible
1465 * tradeoff for simplicity on implementation.
1466 *
1467 * For cloning, we want to share the partial EOF block if it is also the new EOF
1468 * block of the destination file. If the partial EOF block lies inside the
1469 * existing destination EOF, then we have to abort the clone to avoid exposing
1470 * stale data in the destination file. Hence we reject these clone attempts with
1471 * -EINVAL in this case.
1472 */
1473int
1474xfs_reflink_remap_prep(
1475 struct file *file_in,
1476 loff_t pos_in,
1477 struct file *file_out,
1478 loff_t pos_out,
1479 loff_t *len,
1480 unsigned int remap_flags)
1481{
1482 struct inode *inode_in = file_inode(file_in);
1483 struct xfs_inode *src = XFS_I(inode_in);
1484 struct inode *inode_out = file_inode(file_out);
1485 struct xfs_inode *dest = XFS_I(inode_out);
1486 int ret;
1487
1488 /* Lock both files against IO */
1489 ret = xfs_ilock2_io_mmap(src, dest);
1490 if (ret)
1491 return ret;
1492
1493 /* Check file eligibility and prepare for block sharing. */
1494 ret = -EINVAL;
1495 /* Don't reflink realtime inodes */
1496 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1497 goto out_unlock;
1498
1499 /* Don't share DAX file data with non-DAX file. */
1500 if (IS_DAX(inode_in) != IS_DAX(inode_out))
1501 goto out_unlock;
1502
1503 if (!IS_DAX(inode_in))
1504 ret = generic_remap_file_range_prep(file_in, pos_in, file_out,
1505 pos_out, len, remap_flags);
1506 else
1507 ret = dax_remap_file_range_prep(file_in, pos_in, file_out,
1508 pos_out, len, remap_flags, &xfs_read_iomap_ops);
1509 if (ret || *len == 0)
1510 goto out_unlock;
1511
1512 /* Attach dquots to dest inode before changing block map */
1513 ret = xfs_qm_dqattach(dest);
1514 if (ret)
1515 goto out_unlock;
1516
1517 /*
1518 * Zero existing post-eof speculative preallocations in the destination
1519 * file.
1520 */
1521 ret = xfs_reflink_zero_posteof(dest, pos_out);
1522 if (ret)
1523 goto out_unlock;
1524
1525 /* Set flags and remap blocks. */
1526 ret = xfs_reflink_set_inode_flag(src, dest);
1527 if (ret)
1528 goto out_unlock;
1529
1530 /*
1531 * If pos_out > EOF, we may have dirtied blocks between EOF and
1532 * pos_out. In that case, we need to extend the flush and unmap to cover
1533 * from EOF to the end of the copy length.
1534 */
1535 if (pos_out > XFS_ISIZE(dest)) {
1536 loff_t flen = *len + (pos_out - XFS_ISIZE(dest));
1537 ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1538 } else {
1539 ret = xfs_flush_unmap_range(dest, pos_out, *len);
1540 }
1541 if (ret)
1542 goto out_unlock;
1543
1544 xfs_iflags_set(src, XFS_IREMAPPING);
1545 if (inode_in != inode_out)
1546 xfs_ilock_demote(src, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);
1547
1548 return 0;
1549out_unlock:
1550 xfs_iunlock2_io_mmap(src, dest);
1551 return ret;
1552}
1553
1554/* Does this inode need the reflink flag? */
1555int
1556xfs_reflink_inode_has_shared_extents(
1557 struct xfs_trans *tp,
1558 struct xfs_inode *ip,
1559 bool *has_shared)
1560{
1561 struct xfs_bmbt_irec got;
1562 struct xfs_mount *mp = ip->i_mount;
1563 struct xfs_ifork *ifp;
1564 struct xfs_iext_cursor icur;
1565 bool found;
1566 int error;
1567
1568 ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1569 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1570 if (error)
1571 return error;
1572
1573 *has_shared = false;
1574 found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1575 while (found) {
1576 struct xfs_perag *pag;
1577 xfs_agblock_t agbno;
1578 xfs_extlen_t aglen;
1579 xfs_agblock_t rbno;
1580 xfs_extlen_t rlen;
1581
1582 if (isnullstartblock(got.br_startblock) ||
1583 got.br_state != XFS_EXT_NORM)
1584 goto next;
1585
1586 pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, got.br_startblock));
1587 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1588 aglen = got.br_blockcount;
1589 error = xfs_reflink_find_shared(pag, tp, agbno, aglen,
1590 &rbno, &rlen, false);
1591 xfs_perag_put(pag);
1592 if (error)
1593 return error;
1594
1595 /* Is there still a shared block here? */
1596 if (rbno != NULLAGBLOCK) {
1597 *has_shared = true;
1598 return 0;
1599 }
1600next:
1601 found = xfs_iext_next_extent(ifp, &icur, &got);
1602 }
1603
1604 return 0;
1605}
1606
1607/*
1608 * Clear the inode reflink flag if there are no shared extents.
1609 *
1610 * The caller is responsible for joining the inode to the transaction passed in.
1611 * The inode will be joined to the transaction that is returned to the caller.
1612 */
1613int
1614xfs_reflink_clear_inode_flag(
1615 struct xfs_inode *ip,
1616 struct xfs_trans **tpp)
1617{
1618 bool needs_flag;
1619 int error = 0;
1620
1621 ASSERT(xfs_is_reflink_inode(ip));
1622
1623 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1624 if (error || needs_flag)
1625 return error;
1626
1627 /*
1628 * We didn't find any shared blocks so turn off the reflink flag.
1629 * First, get rid of any leftover CoW mappings.
1630 */
1631 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,
1632 true);
1633 if (error)
1634 return error;
1635
1636 /* Clear the inode flag. */
1637 trace_xfs_reflink_unset_inode_flag(ip);
1638 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1639 xfs_inode_clear_cowblocks_tag(ip);
1640 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1641
1642 return error;
1643}
1644
1645/*
1646 * Clear the inode reflink flag if there are no shared extents and the size
1647 * hasn't changed.
1648 */
1649STATIC int
1650xfs_reflink_try_clear_inode_flag(
1651 struct xfs_inode *ip)
1652{
1653 struct xfs_mount *mp = ip->i_mount;
1654 struct xfs_trans *tp;
1655 int error = 0;
1656
1657 /* Start a rolling transaction to remove the mappings */
1658 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1659 if (error)
1660 return error;
1661
1662 xfs_ilock(ip, XFS_ILOCK_EXCL);
1663 xfs_trans_ijoin(tp, ip, 0);
1664
1665 error = xfs_reflink_clear_inode_flag(ip, &tp);
1666 if (error)
1667 goto cancel;
1668
1669 error = xfs_trans_commit(tp);
1670 if (error)
1671 goto out;
1672
1673 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1674 return 0;
1675cancel:
1676 xfs_trans_cancel(tp);
1677out:
1678 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1679 return error;
1680}
1681
1682/*
1683 * Pre-COW all shared blocks within a given byte range of a file and turn off
1684 * the reflink flag if we unshare all of the file's blocks.
1685 */
1686int
1687xfs_reflink_unshare(
1688 struct xfs_inode *ip,
1689 xfs_off_t offset,
1690 xfs_off_t len)
1691{
1692 struct inode *inode = VFS_I(ip);
1693 int error;
1694
1695 if (!xfs_is_reflink_inode(ip))
1696 return 0;
1697
1698 trace_xfs_reflink_unshare(ip, offset, len);
1699
1700 inode_dio_wait(inode);
1701
1702 if (IS_DAX(inode))
1703 error = dax_file_unshare(inode, offset, len,
1704 &xfs_dax_write_iomap_ops);
1705 else
1706 error = iomap_file_unshare(inode, offset, len,
1707 &xfs_buffered_write_iomap_ops);
1708 if (error)
1709 goto out;
1710
1711 error = filemap_write_and_wait_range(inode->i_mapping, offset,
1712 offset + len - 1);
1713 if (error)
1714 goto out;
1715
1716 /* Turn off the reflink flag if possible. */
1717 error = xfs_reflink_try_clear_inode_flag(ip);
1718 if (error)
1719 goto out;
1720 return 0;
1721
1722out:
1723 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1724 return error;
1725}