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_da_format.h"
15 #include "xfs_da_btree.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_inode_item.h"
19 #include "xfs_bmap.h"
20 #include "xfs_bmap_util.h"
21 #include "xfs_error.h"
22 #include "xfs_dir2.h"
23 #include "xfs_dir2_priv.h"
24 #include "xfs_ioctl.h"
25 #include "xfs_trace.h"
26 #include "xfs_log.h"
27 #include "xfs_icache.h"
28 #include "xfs_pnfs.h"
29 #include "xfs_btree.h"
30 #include "xfs_refcount_btree.h"
31 #include "xfs_refcount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_trans_space.h"
34 #include "xfs_bit.h"
35 #include "xfs_alloc.h"
36 #include "xfs_quota_defs.h"
37 #include "xfs_quota.h"
38 #include "xfs_reflink.h"
39 #include "xfs_iomap.h"
40 #include "xfs_rmap_btree.h"
41 #include "xfs_sb.h"
42 #include "xfs_ag_resv.h"
44 /*
45 * Copy on Write of Shared Blocks
46 *
47 * XFS must preserve "the usual" file semantics even when two files share
48 * the same physical blocks. This means that a write to one file must not
49 * alter the blocks in a different file; the way that we'll do that is
50 * through the use of a copy-on-write mechanism. At a high level, that
51 * means that when we want to write to a shared block, we allocate a new
52 * block, write the data to the new block, and if that succeeds we map the
53 * new block into the file.
54 *
55 * XFS provides a "delayed allocation" mechanism that defers the allocation
56 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
57 * possible. This reduces fragmentation by enabling the filesystem to ask
58 * for bigger chunks less often, which is exactly what we want for CoW.
59 *
60 * The delalloc mechanism begins when the kernel wants to make a block
61 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
62 * create a delalloc mapping, which is a regular in-core extent, but without
63 * a real startblock. (For delalloc mappings, the startblock encodes both
64 * a flag that this is a delalloc mapping, and a worst-case estimate of how
65 * many blocks might be required to put the mapping into the BMBT.) delalloc
66 * mappings are a reservation against the free space in the filesystem;
67 * adjacent mappings can also be combined into fewer larger mappings.
68 *
69 * As an optimization, the CoW extent size hint (cowextsz) creates
70 * outsized aligned delalloc reservations in the hope of landing out of
71 * order nearby CoW writes in a single extent on disk, thereby reducing
72 * fragmentation and improving future performance.
73 *
74 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
75 * C: ------DDDDDDD--------- (CoW fork)
76 *
77 * When dirty pages are being written out (typically in writepage), the
78 * delalloc reservations are converted into unwritten mappings by
79 * allocating blocks and replacing the delalloc mapping with real ones.
80 * A delalloc mapping can be replaced by several unwritten ones if the
81 * free space is fragmented.
82 *
83 * D: --RRRRRRSSSRRRRRRRR---
84 * C: ------UUUUUUU---------
85 *
86 * We want to adapt the delalloc mechanism for copy-on-write, since the
87 * write paths are similar. The first two steps (creating the reservation
88 * and allocating the blocks) are exactly the same as delalloc except that
89 * the mappings must be stored in a separate CoW fork because we do not want
90 * to disturb the mapping in the data fork until we're sure that the write
91 * succeeded. IO completion in this case is the process of removing the old
92 * mapping from the data fork and moving the new mapping from the CoW fork to
93 * the data fork. This will be discussed shortly.
94 *
95 * For now, unaligned directio writes will be bounced back to the page cache.
96 * Block-aligned directio writes will use the same mechanism as buffered
97 * writes.
98 *
99 * Just prior to submitting the actual disk write requests, we convert
100 * the extents representing the range of the file actually being written
101 * (as opposed to extra pieces created for the cowextsize hint) to real
102 * extents. This will become important in the next step:
103 *
104 * D: --RRRRRRSSSRRRRRRRR---
105 * C: ------UUrrUUU---------
106 *
107 * CoW remapping must be done after the data block write completes,
108 * because we don't want to destroy the old data fork map until we're sure
109 * the new block has been written. Since the new mappings are kept in a
110 * separate fork, we can simply iterate these mappings to find the ones
111 * that cover the file blocks that we just CoW'd. For each extent, simply
112 * unmap the corresponding range in the data fork, map the new range into
113 * the data fork, and remove the extent from the CoW fork. Because of
114 * the presence of the cowextsize hint, however, we must be careful
115 * only to remap the blocks that we've actually written out -- we must
116 * never remap delalloc reservations nor CoW staging blocks that have
117 * yet to be written. This corresponds exactly to the real extents in
118 * the CoW fork:
119 *
120 * D: --RRRRRRrrSRRRRRRRR---
121 * C: ------UU--UUU---------
122 *
123 * Since the remapping operation can be applied to an arbitrary file
124 * range, we record the need for the remap step as a flag in the ioend
125 * instead of declaring a new IO type. This is required for direct io
126 * because we only have ioend for the whole dio, and we have to be able to
127 * remember the presence of unwritten blocks and CoW blocks with a single
128 * ioend structure. Better yet, the more ground we can cover with one
129 * ioend, the better.
130 */
132 /*
133 * Given an AG extent, find the lowest-numbered run of shared blocks
134 * within that range and return the range in fbno/flen. If
135 * find_end_of_shared is true, return the longest contiguous extent of
136 * shared blocks. If there are no shared extents, fbno and flen will
137 * be set to NULLAGBLOCK and 0, respectively.
138 */
139 int
140 xfs_reflink_find_shared(
141 struct xfs_mount *mp,
142 struct xfs_trans *tp,
143 xfs_agnumber_t agno,
144 xfs_agblock_t agbno,
145 xfs_extlen_t aglen,
146 xfs_agblock_t *fbno,
147 xfs_extlen_t *flen,
148 bool find_end_of_shared)
149 {
150 struct xfs_buf *agbp;
151 struct xfs_btree_cur *cur;
152 int error;
154 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
155 if (error)
156 return error;
157 if (!agbp)
158 return -ENOMEM;
160 cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
162 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
163 find_end_of_shared);
165 xfs_btree_del_cursor(cur, error);
167 xfs_trans_brelse(tp, agbp);
168 return error;
169 }
171 /*
172 * Trim the mapping to the next block where there's a change in the
173 * shared/unshared status. More specifically, this means that we
174 * find the lowest-numbered extent of shared blocks that coincides with
175 * the given block mapping. If the shared extent overlaps the start of
176 * the mapping, trim the mapping to the end of the shared extent. If
177 * the shared region intersects the mapping, trim the mapping to the
178 * start of the shared extent. If there are no shared regions that
179 * overlap, just return the original extent.
180 */
181 int
182 xfs_reflink_trim_around_shared(
183 struct xfs_inode *ip,
184 struct xfs_bmbt_irec *irec,
185 bool *shared,
186 bool *trimmed)
187 {
188 xfs_agnumber_t agno;
189 xfs_agblock_t agbno;
190 xfs_extlen_t aglen;
191 xfs_agblock_t fbno;
192 xfs_extlen_t flen;
193 int error = 0;
195 /* Holes, unwritten, and delalloc extents cannot be shared */
196 if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
197 *shared = false;
198 return 0;
199 }
201 trace_xfs_reflink_trim_around_shared(ip, irec);
203 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
204 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
205 aglen = irec->br_blockcount;
207 error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
208 aglen, &fbno, &flen, true);
209 if (error)
210 return error;
212 *shared = *trimmed = false;
213 if (fbno == NULLAGBLOCK) {
214 /* No shared blocks at all. */
215 return 0;
216 } else if (fbno == agbno) {
217 /*
218 * The start of this extent is shared. Truncate the
219 * mapping at the end of the shared region so that a
220 * subsequent iteration starts at the start of the
221 * unshared region.
222 */
223 irec->br_blockcount = flen;
224 *shared = true;
225 if (flen != aglen)
226 *trimmed = true;
227 return 0;
228 } else {
229 /*
230 * There's a shared extent midway through this extent.
231 * Truncate the mapping at the start of the shared
232 * extent so that a subsequent iteration starts at the
233 * start of the shared region.
234 */
235 irec->br_blockcount = fbno - agbno;
236 *trimmed = true;
237 return 0;
238 }
239 }
241 /*
242 * Trim the passed in imap to the next shared/unshared extent boundary, and
243 * if imap->br_startoff points to a shared extent reserve space for it in the
244 * COW fork. In this case *shared is set to true, else to false.
245 *
246 * Note that imap will always contain the block numbers for the existing blocks
247 * in the data fork, as the upper layers need them for read-modify-write
248 * operations.
249 */
250 int
251 xfs_reflink_reserve_cow(
252 struct xfs_inode *ip,
253 struct xfs_bmbt_irec *imap,
254 bool *shared)
255 {
256 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
257 struct xfs_bmbt_irec got;
258 int error = 0;
259 bool eof = false, trimmed;
260 struct xfs_iext_cursor icur;
262 /*
263 * Search the COW fork extent list first. This serves two purposes:
264 * first this implement the speculative preallocation using cowextisze,
265 * so that we also unshared block adjacent to shared blocks instead
266 * of just the shared blocks themselves. Second the lookup in the
267 * extent list is generally faster than going out to the shared extent
268 * tree.
269 */
271 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
272 eof = true;
273 if (!eof && got.br_startoff <= imap->br_startoff) {
274 trace_xfs_reflink_cow_found(ip, imap);
275 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
277 *shared = true;
278 return 0;
279 }
281 /* Trim the mapping to the nearest shared extent boundary. */
282 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
283 if (error)
284 return error;
286 /* Not shared? Just report the (potentially capped) extent. */
287 if (!*shared)
288 return 0;
290 /*
291 * Fork all the shared blocks from our write offset until the end of
292 * the extent.
293 */
294 error = xfs_qm_dqattach_locked(ip, false);
295 if (error)
296 return error;
298 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
299 imap->br_blockcount, 0, &got, &icur, eof);
300 if (error == -ENOSPC || error == -EDQUOT)
301 trace_xfs_reflink_cow_enospc(ip, imap);
302 if (error)
303 return error;
305 trace_xfs_reflink_cow_alloc(ip, &got);
306 return 0;
307 }
309 /* Convert part of an unwritten CoW extent to a real one. */
310 STATIC int
311 xfs_reflink_convert_cow_extent(
312 struct xfs_inode *ip,
313 struct xfs_bmbt_irec *imap,
314 xfs_fileoff_t offset_fsb,
315 xfs_filblks_t count_fsb)
316 {
317 int nimaps = 1;
319 if (imap->br_state == XFS_EXT_NORM)
320 return 0;
322 xfs_trim_extent(imap, offset_fsb, count_fsb);
323 trace_xfs_reflink_convert_cow(ip, imap);
324 if (imap->br_blockcount == 0)
325 return 0;
326 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
327 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, 0, imap,
328 &nimaps);
329 }
331 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
332 int
333 xfs_reflink_convert_cow(
334 struct xfs_inode *ip,
335 xfs_off_t offset,
336 xfs_off_t count)
337 {
338 struct xfs_mount *mp = ip->i_mount;
339 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
340 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
341 xfs_filblks_t count_fsb = end_fsb - offset_fsb;
342 struct xfs_bmbt_irec imap;
343 int nimaps = 1, error = 0;
345 ASSERT(count != 0);
347 xfs_ilock(ip, XFS_ILOCK_EXCL);
348 error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
349 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
350 XFS_BMAPI_CONVERT_ONLY, 0, &imap, &nimaps);
351 xfs_iunlock(ip, XFS_ILOCK_EXCL);
352 return error;
353 }
355 /*
356 * Find the extent that maps the given range in the COW fork. Even if the extent
357 * is not shared we might have a preallocation for it in the COW fork. If so we
358 * use it that rather than trigger a new allocation.
359 */
360 static int
361 xfs_find_trim_cow_extent(
362 struct xfs_inode *ip,
363 struct xfs_bmbt_irec *imap,
364 bool *shared,
365 bool *found)
366 {
367 xfs_fileoff_t offset_fsb = imap->br_startoff;
368 xfs_filblks_t count_fsb = imap->br_blockcount;
369 struct xfs_iext_cursor icur;
370 struct xfs_bmbt_irec got;
371 bool trimmed;
373 *found = false;
375 /*
376 * If we don't find an overlapping extent, trim the range we need to
377 * allocate to fit the hole we found.
378 */
379 if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) ||
380 got.br_startoff > offset_fsb)
381 return xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
383 *shared = true;
384 if (isnullstartblock(got.br_startblock)) {
385 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
386 return 0;
387 }
389 /* real extent found - no need to allocate */
390 xfs_trim_extent(&got, offset_fsb, count_fsb);
391 *imap = got;
392 *found = true;
393 return 0;
394 }
396 /* Allocate all CoW reservations covering a range of blocks in a file. */
397 int
398 xfs_reflink_allocate_cow(
399 struct xfs_inode *ip,
400 struct xfs_bmbt_irec *imap,
401 bool *shared,
402 uint *lockmode)
403 {
404 struct xfs_mount *mp = ip->i_mount;
405 xfs_fileoff_t offset_fsb = imap->br_startoff;
406 xfs_filblks_t count_fsb = imap->br_blockcount;
407 struct xfs_trans *tp;
408 int nimaps, error = 0;
409 bool found;
410 xfs_filblks_t resaligned;
411 xfs_extlen_t resblks = 0;
413 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
414 ASSERT(xfs_is_reflink_inode(ip));
416 error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
417 if (error || !*shared)
418 return error;
419 if (found)
420 goto convert;
422 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
423 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
424 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
426 xfs_iunlock(ip, *lockmode);
427 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
428 *lockmode = XFS_ILOCK_EXCL;
429 xfs_ilock(ip, *lockmode);
431 if (error)
432 return error;
434 error = xfs_qm_dqattach_locked(ip, false);
435 if (error)
436 goto out_trans_cancel;
438 /*
439 * Check for an overlapping extent again now that we dropped the ilock.
440 */
441 error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
442 if (error || !*shared)
443 goto out_trans_cancel;
444 if (found) {
445 xfs_trans_cancel(tp);
446 goto convert;
447 }
449 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
450 XFS_QMOPT_RES_REGBLKS);
451 if (error)
452 goto out_trans_cancel;
454 xfs_trans_ijoin(tp, ip, 0);
456 /* Allocate the entire reservation as unwritten blocks. */
457 nimaps = 1;
458 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
459 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC,
460 resblks, imap, &nimaps);
461 if (error)
462 goto out_unreserve;
464 xfs_inode_set_cowblocks_tag(ip);
465 error = xfs_trans_commit(tp);
466 if (error)
467 return error;
469 /*
470 * Allocation succeeded but the requested range was not even partially
471 * satisfied? Bail out!
472 */
473 if (nimaps == 0)
474 return -ENOSPC;
475 convert:
476 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb);
478 out_unreserve:
479 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
480 XFS_QMOPT_RES_REGBLKS);
481 out_trans_cancel:
482 xfs_trans_cancel(tp);
483 return error;
484 }
486 /*
487 * Cancel CoW reservations for some block range of an inode.
488 *
489 * If cancel_real is true this function cancels all COW fork extents for the
490 * inode; if cancel_real is false, real extents are not cleared.
491 *
492 * Caller must have already joined the inode to the current transaction. The
493 * inode will be joined to the transaction returned to the caller.
494 */
495 int
496 xfs_reflink_cancel_cow_blocks(
497 struct xfs_inode *ip,
498 struct xfs_trans **tpp,
499 xfs_fileoff_t offset_fsb,
500 xfs_fileoff_t end_fsb,
501 bool cancel_real)
502 {
503 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
504 struct xfs_bmbt_irec got, del;
505 struct xfs_iext_cursor icur;
506 int error = 0;
508 if (!xfs_inode_has_cow_data(ip))
509 return 0;
510 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
511 return 0;
513 /* Walk backwards until we're out of the I/O range... */
514 while (got.br_startoff + got.br_blockcount > offset_fsb) {
515 del = got;
516 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
518 /* Extent delete may have bumped ext forward */
519 if (!del.br_blockcount) {
520 xfs_iext_prev(ifp, &icur);
521 goto next_extent;
522 }
524 trace_xfs_reflink_cancel_cow(ip, &del);
526 if (isnullstartblock(del.br_startblock)) {
527 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
528 &icur, &got, &del);
529 if (error)
530 break;
531 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
532 ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
534 /* Free the CoW orphan record. */
535 error = xfs_refcount_free_cow_extent(*tpp,
536 del.br_startblock, del.br_blockcount);
537 if (error)
538 break;
540 xfs_bmap_add_free(*tpp, del.br_startblock,
541 del.br_blockcount, NULL);
543 /* Roll the transaction */
544 error = xfs_defer_finish(tpp);
545 if (error)
546 break;
548 /* Remove the mapping from the CoW fork. */
549 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
551 /* Remove the quota reservation */
552 error = xfs_trans_reserve_quota_nblks(NULL, ip,
553 -(long)del.br_blockcount, 0,
554 XFS_QMOPT_RES_REGBLKS);
555 if (error)
556 break;
557 } else {
558 /* Didn't do anything, push cursor back. */
559 xfs_iext_prev(ifp, &icur);
560 }
561 next_extent:
562 if (!xfs_iext_get_extent(ifp, &icur, &got))
563 break;
564 }
566 /* clear tag if cow fork is emptied */
567 if (!ifp->if_bytes)
568 xfs_inode_clear_cowblocks_tag(ip);
569 return error;
570 }
572 /*
573 * Cancel CoW reservations for some byte range of an inode.
574 *
575 * If cancel_real is true this function cancels all COW fork extents for the
576 * inode; if cancel_real is false, real extents are not cleared.
577 */
578 int
579 xfs_reflink_cancel_cow_range(
580 struct xfs_inode *ip,
581 xfs_off_t offset,
582 xfs_off_t count,
583 bool cancel_real)
584 {
585 struct xfs_trans *tp;
586 xfs_fileoff_t offset_fsb;
587 xfs_fileoff_t end_fsb;
588 int error;
590 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
591 ASSERT(xfs_is_reflink_inode(ip));
593 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
594 if (count == NULLFILEOFF)
595 end_fsb = NULLFILEOFF;
596 else
597 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
599 /* Start a rolling transaction to remove the mappings */
600 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
601 0, 0, XFS_TRANS_NOFS, &tp);
602 if (error)
603 goto out;
605 xfs_ilock(ip, XFS_ILOCK_EXCL);
606 xfs_trans_ijoin(tp, ip, 0);
608 /* Scrape out the old CoW reservations */
609 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
610 cancel_real);
611 if (error)
612 goto out_cancel;
614 error = xfs_trans_commit(tp);
616 xfs_iunlock(ip, XFS_ILOCK_EXCL);
617 return error;
619 out_cancel:
620 xfs_trans_cancel(tp);
621 xfs_iunlock(ip, XFS_ILOCK_EXCL);
622 out:
623 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
624 return error;
625 }
627 /*
628 * Remap parts of a file's data fork after a successful CoW.
629 */
630 int
631 xfs_reflink_end_cow(
632 struct xfs_inode *ip,
633 xfs_off_t offset,
634 xfs_off_t count)
635 {
636 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
637 struct xfs_bmbt_irec got, del;
638 struct xfs_trans *tp;
639 xfs_fileoff_t offset_fsb;
640 xfs_fileoff_t end_fsb;
641 int error;
642 unsigned int resblks;
643 xfs_filblks_t rlen;
644 struct xfs_iext_cursor icur;
646 trace_xfs_reflink_end_cow(ip, offset, count);
648 /* No COW extents? That's easy! */
649 if (ifp->if_bytes == 0)
650 return 0;
652 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
653 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
655 /*
656 * Start a rolling transaction to switch the mappings. We're
657 * unlikely ever to have to remap 16T worth of single-block
658 * extents, so just cap the worst case extent count to 2^32-1.
659 * Stick a warning in just in case, and avoid 64-bit division.
660 */
661 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
662 if (end_fsb - offset_fsb > UINT_MAX) {
663 error = -EFSCORRUPTED;
664 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
665 ASSERT(0);
666 goto out;
667 }
668 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
669 (unsigned int)(end_fsb - offset_fsb),
670 XFS_DATA_FORK);
671 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
672 resblks, 0, XFS_TRANS_RESERVE | XFS_TRANS_NOFS, &tp);
673 if (error)
674 goto out;
676 xfs_ilock(ip, XFS_ILOCK_EXCL);
677 xfs_trans_ijoin(tp, ip, 0);
679 /*
680 * In case of racing, overlapping AIO writes no COW extents might be
681 * left by the time I/O completes for the loser of the race. In that
682 * case we are done.
683 */
684 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
685 goto out_cancel;
687 /* Walk backwards until we're out of the I/O range... */
688 while (got.br_startoff + got.br_blockcount > offset_fsb) {
689 del = got;
690 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
692 /* Extent delete may have bumped ext forward */
693 if (!del.br_blockcount)
694 goto prev_extent;
696 /*
697 * Only remap real extent that contain data. With AIO
698 * speculatively preallocations can leak into the range we
699 * are called upon, and we need to skip them.
700 */
701 if (!xfs_bmap_is_real_extent(&got))
702 goto prev_extent;
704 /* Unmap the old blocks in the data fork. */
705 ASSERT(tp->t_firstblock == NULLFSBLOCK);
706 rlen = del.br_blockcount;
707 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
708 if (error)
709 goto out_cancel;
711 /* Trim the extent to whatever got unmapped. */
712 if (rlen) {
713 xfs_trim_extent(&del, del.br_startoff + rlen,
714 del.br_blockcount - rlen);
715 }
716 trace_xfs_reflink_cow_remap(ip, &del);
718 /* Free the CoW orphan record. */
719 error = xfs_refcount_free_cow_extent(tp, del.br_startblock,
720 del.br_blockcount);
721 if (error)
722 goto out_cancel;
724 /* Map the new blocks into the data fork. */
725 error = xfs_bmap_map_extent(tp, ip, &del);
726 if (error)
727 goto out_cancel;
729 /* Charge this new data fork mapping to the on-disk quota. */
730 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
731 (long)del.br_blockcount);
733 /* Remove the mapping from the CoW fork. */
734 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
736 error = xfs_defer_finish(&tp);
737 if (error)
738 goto out_cancel;
739 if (!xfs_iext_get_extent(ifp, &icur, &got))
740 break;
741 continue;
742 prev_extent:
743 if (!xfs_iext_prev_extent(ifp, &icur, &got))
744 break;
745 }
747 error = xfs_trans_commit(tp);
748 xfs_iunlock(ip, XFS_ILOCK_EXCL);
749 if (error)
750 goto out;
751 return 0;
753 out_cancel:
754 xfs_trans_cancel(tp);
755 xfs_iunlock(ip, XFS_ILOCK_EXCL);
756 out:
757 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
758 return error;
759 }
761 /*
762 * Free leftover CoW reservations that didn't get cleaned out.
763 */
764 int
765 xfs_reflink_recover_cow(
766 struct xfs_mount *mp)
767 {
768 xfs_agnumber_t agno;
769 int error = 0;
771 if (!xfs_sb_version_hasreflink(&mp->m_sb))
772 return 0;
774 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
775 error = xfs_refcount_recover_cow_leftovers(mp, agno);
776 if (error)
777 break;
778 }
780 return error;
781 }
783 /*
784 * Reflinking (Block) Ranges of Two Files Together
785 *
786 * First, ensure that the reflink flag is set on both inodes. The flag is an
787 * optimization to avoid unnecessary refcount btree lookups in the write path.
788 *
789 * Now we can iteratively remap the range of extents (and holes) in src to the
790 * corresponding ranges in dest. Let drange and srange denote the ranges of
791 * logical blocks in dest and src touched by the reflink operation.
792 *
793 * While the length of drange is greater than zero,
794 * - Read src's bmbt at the start of srange ("imap")
795 * - If imap doesn't exist, make imap appear to start at the end of srange
796 * with zero length.
797 * - If imap starts before srange, advance imap to start at srange.
798 * - If imap goes beyond srange, truncate imap to end at the end of srange.
799 * - Punch (imap start - srange start + imap len) blocks from dest at
800 * offset (drange start).
801 * - If imap points to a real range of pblks,
802 * > Increase the refcount of the imap's pblks
803 * > Map imap's pblks into dest at the offset
804 * (drange start + imap start - srange start)
805 * - Advance drange and srange by (imap start - srange start + imap len)
806 *
807 * Finally, if the reflink made dest longer, update both the in-core and
808 * on-disk file sizes.
809 *
810 * ASCII Art Demonstration:
811 *
812 * Let's say we want to reflink this source file:
813 *
814 * ----SSSSSSS-SSSSS----SSSSSS (src file)
815 * <-------------------->
816 *
817 * into this destination file:
818 *
819 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
820 * <-------------------->
821 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
822 * Observe that the range has different logical offsets in either file.
823 *
824 * Consider that the first extent in the source file doesn't line up with our
825 * reflink range. Unmapping and remapping are separate operations, so we can
826 * unmap more blocks from the destination file than we remap.
827 *
828 * ----SSSSSSS-SSSSS----SSSSSS
829 * <------->
830 * --DDDDD---------DDDDD--DDD
831 * <------->
832 *
833 * Now remap the source extent into the destination file:
834 *
835 * ----SSSSSSS-SSSSS----SSSSSS
836 * <------->
837 * --DDDDD--SSSSSSSDDDDD--DDD
838 * <------->
839 *
840 * Do likewise with the second hole and extent in our range. Holes in the
841 * unmap range don't affect our operation.
842 *
843 * ----SSSSSSS-SSSSS----SSSSSS
844 * <---->
845 * --DDDDD--SSSSSSS-SSSSS-DDD
846 * <---->
847 *
848 * Finally, unmap and remap part of the third extent. This will increase the
849 * size of the destination file.
850 *
851 * ----SSSSSSS-SSSSS----SSSSSS
852 * <----->
853 * --DDDDD--SSSSSSS-SSSSS----SSS
854 * <----->
855 *
856 * Once we update the destination file's i_size, we're done.
857 */
859 /*
860 * Ensure the reflink bit is set in both inodes.
861 */
862 STATIC int
863 xfs_reflink_set_inode_flag(
864 struct xfs_inode *src,
865 struct xfs_inode *dest)
866 {
867 struct xfs_mount *mp = src->i_mount;
868 int error;
869 struct xfs_trans *tp;
871 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
872 return 0;
874 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
875 if (error)
876 goto out_error;
878 /* Lock both files against IO */
879 if (src->i_ino == dest->i_ino)
880 xfs_ilock(src, XFS_ILOCK_EXCL);
881 else
882 xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
884 if (!xfs_is_reflink_inode(src)) {
885 trace_xfs_reflink_set_inode_flag(src);
886 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
887 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
888 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
889 xfs_ifork_init_cow(src);
890 } else
891 xfs_iunlock(src, XFS_ILOCK_EXCL);
893 if (src->i_ino == dest->i_ino)
894 goto commit_flags;
896 if (!xfs_is_reflink_inode(dest)) {
897 trace_xfs_reflink_set_inode_flag(dest);
898 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
899 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
900 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
901 xfs_ifork_init_cow(dest);
902 } else
903 xfs_iunlock(dest, XFS_ILOCK_EXCL);
905 commit_flags:
906 error = xfs_trans_commit(tp);
907 if (error)
908 goto out_error;
909 return error;
911 out_error:
912 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
913 return error;
914 }
916 /*
917 * Update destination inode size & cowextsize hint, if necessary.
918 */
919 STATIC int
920 xfs_reflink_update_dest(
921 struct xfs_inode *dest,
922 xfs_off_t newlen,
923 xfs_extlen_t cowextsize,
924 bool is_dedupe)
925 {
926 struct xfs_mount *mp = dest->i_mount;
927 struct xfs_trans *tp;
928 int error;
930 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
931 return 0;
933 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
934 if (error)
935 goto out_error;
937 xfs_ilock(dest, XFS_ILOCK_EXCL);
938 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
940 if (newlen > i_size_read(VFS_I(dest))) {
941 trace_xfs_reflink_update_inode_size(dest, newlen);
942 i_size_write(VFS_I(dest), newlen);
943 dest->i_d.di_size = newlen;
944 }
946 if (cowextsize) {
947 dest->i_d.di_cowextsize = cowextsize;
948 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
949 }
951 if (!is_dedupe) {
952 xfs_trans_ichgtime(tp, dest,
953 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
954 }
955 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
957 error = xfs_trans_commit(tp);
958 if (error)
959 goto out_error;
960 return error;
962 out_error:
963 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
964 return error;
965 }
967 /*
968 * Do we have enough reserve in this AG to handle a reflink? The refcount
969 * btree already reserved all the space it needs, but the rmap btree can grow
970 * infinitely, so we won't allow more reflinks when the AG is down to the
971 * btree reserves.
972 */
973 static int
974 xfs_reflink_ag_has_free_space(
975 struct xfs_mount *mp,
976 xfs_agnumber_t agno)
977 {
978 struct xfs_perag *pag;
979 int error = 0;
981 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
982 return 0;
984 pag = xfs_perag_get(mp, agno);
985 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
986 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
987 error = -ENOSPC;
988 xfs_perag_put(pag);
989 return error;
990 }
992 /*
993 * Unmap a range of blocks from a file, then map other blocks into the hole.
994 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
995 * The extent irec is mapped into dest at irec->br_startoff.
996 */
997 STATIC int
998 xfs_reflink_remap_extent(
999 struct xfs_inode *ip,
1000 struct xfs_bmbt_irec *irec,
1001 xfs_fileoff_t destoff,
1002 xfs_off_t new_isize)
1003 {
1004 struct xfs_mount *mp = ip->i_mount;
1005 bool real_extent = xfs_bmap_is_real_extent(irec);
1006 struct xfs_trans *tp;
1007 unsigned int resblks;
1008 struct xfs_bmbt_irec uirec;
1009 xfs_filblks_t rlen;
1010 xfs_filblks_t unmap_len;
1011 xfs_off_t newlen;
1012 int error;
1014 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1015 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1017 /* No reflinking if we're low on space */
1018 if (real_extent) {
1019 error = xfs_reflink_ag_has_free_space(mp,
1020 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1021 if (error)
1022 goto out;
1023 }
1025 /* Start a rolling transaction to switch the mappings */
1026 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1027 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1028 if (error)
1029 goto out;
1031 xfs_ilock(ip, XFS_ILOCK_EXCL);
1032 xfs_trans_ijoin(tp, ip, 0);
1034 /* If we're not just clearing space, then do we have enough quota? */
1035 if (real_extent) {
1036 error = xfs_trans_reserve_quota_nblks(tp, ip,
1037 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1038 if (error)
1039 goto out_cancel;
1040 }
1042 trace_xfs_reflink_remap(ip, irec->br_startoff,
1043 irec->br_blockcount, irec->br_startblock);
1045 /* Unmap the old blocks in the data fork. */
1046 rlen = unmap_len;
1047 while (rlen) {
1048 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1049 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1);
1050 if (error)
1051 goto out_cancel;
1053 /*
1054 * Trim the extent to whatever got unmapped.
1055 * Remember, bunmapi works backwards.
1056 */
1057 uirec.br_startblock = irec->br_startblock + rlen;
1058 uirec.br_startoff = irec->br_startoff + rlen;
1059 uirec.br_blockcount = unmap_len - rlen;
1060 unmap_len = rlen;
1062 /* If this isn't a real mapping, we're done. */
1063 if (!real_extent || uirec.br_blockcount == 0)
1064 goto next_extent;
1066 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1067 uirec.br_blockcount, uirec.br_startblock);
1069 /* Update the refcount tree */
1070 error = xfs_refcount_increase_extent(tp, &uirec);
1071 if (error)
1072 goto out_cancel;
1074 /* Map the new blocks into the data fork. */
1075 error = xfs_bmap_map_extent(tp, ip, &uirec);
1076 if (error)
1077 goto out_cancel;
1079 /* Update quota accounting. */
1080 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1081 uirec.br_blockcount);
1083 /* Update dest isize if needed. */
1084 newlen = XFS_FSB_TO_B(mp,
1085 uirec.br_startoff + uirec.br_blockcount);
1086 newlen = min_t(xfs_off_t, newlen, new_isize);
1087 if (newlen > i_size_read(VFS_I(ip))) {
1088 trace_xfs_reflink_update_inode_size(ip, newlen);
1089 i_size_write(VFS_I(ip), newlen);
1090 ip->i_d.di_size = newlen;
1091 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1092 }
1094 next_extent:
1095 /* Process all the deferred stuff. */
1096 error = xfs_defer_finish(&tp);
1097 if (error)
1098 goto out_cancel;
1099 }
1101 error = xfs_trans_commit(tp);
1102 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1103 if (error)
1104 goto out;
1105 return 0;
1107 out_cancel:
1108 xfs_trans_cancel(tp);
1109 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1110 out:
1111 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1112 return error;
1113 }
1115 /*
1116 * Iteratively remap one file's extents (and holes) to another's.
1117 */
1118 STATIC int
1119 xfs_reflink_remap_blocks(
1120 struct xfs_inode *src,
1121 xfs_fileoff_t srcoff,
1122 struct xfs_inode *dest,
1123 xfs_fileoff_t destoff,
1124 xfs_filblks_t len,
1125 xfs_off_t new_isize)
1126 {
1127 struct xfs_bmbt_irec imap;
1128 int nimaps;
1129 int error = 0;
1130 xfs_filblks_t range_len;
1132 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1133 while (len) {
1134 uint lock_mode;
1136 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1137 dest, destoff);
1139 /* Read extent from the source file */
1140 nimaps = 1;
1141 lock_mode = xfs_ilock_data_map_shared(src);
1142 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1143 xfs_iunlock(src, lock_mode);
1144 if (error)
1145 goto err;
1146 ASSERT(nimaps == 1);
1148 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1149 &imap);
1151 /* Translate imap into the destination file. */
1152 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1153 imap.br_startoff += destoff - srcoff;
1155 /* Clear dest from destoff to the end of imap and map it in. */
1156 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1157 new_isize);
1158 if (error)
1159 goto err;
1161 if (fatal_signal_pending(current)) {
1162 error = -EINTR;
1163 goto err;
1164 }
1166 /* Advance drange/srange */
1167 srcoff += range_len;
1168 destoff += range_len;
1169 len -= range_len;
1170 }
1172 return 0;
1174 err:
1175 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1176 return error;
1177 }
1179 /*
1180 * Grab the exclusive iolock for a data copy from src to dest, making
1181 * sure to abide vfs locking order (lowest pointer value goes first) and
1182 * breaking the pnfs layout leases on dest before proceeding. The loop
1183 * is needed because we cannot call the blocking break_layout() with the
1184 * src iolock held, and therefore have to back out both locks.
1185 */
1186 static int
1187 xfs_iolock_two_inodes_and_break_layout(
1188 struct inode *src,
1189 struct inode *dest)
1190 {
1191 int error;
1193 retry:
1194 if (src < dest) {
1195 inode_lock_shared(src);
1196 inode_lock_nested(dest, I_MUTEX_NONDIR2);
1197 } else {
1198 /* src >= dest */
1199 inode_lock(dest);
1200 }
1202 error = break_layout(dest, false);
1203 if (error == -EWOULDBLOCK) {
1204 inode_unlock(dest);
1205 if (src < dest)
1206 inode_unlock_shared(src);
1207 error = break_layout(dest, true);
1208 if (error)
1209 return error;
1210 goto retry;
1211 }
1212 if (error) {
1213 inode_unlock(dest);
1214 if (src < dest)
1215 inode_unlock_shared(src);
1216 return error;
1217 }
1218 if (src > dest)
1219 inode_lock_shared_nested(src, I_MUTEX_NONDIR2);
1220 return 0;
1221 }
1223 /* Unlock both inodes after they've been prepped for a range clone. */
1224 STATIC void
1225 xfs_reflink_remap_unlock(
1226 struct file *file_in,
1227 struct file *file_out)
1228 {
1229 struct inode *inode_in = file_inode(file_in);
1230 struct xfs_inode *src = XFS_I(inode_in);
1231 struct inode *inode_out = file_inode(file_out);
1232 struct xfs_inode *dest = XFS_I(inode_out);
1233 bool same_inode = (inode_in == inode_out);
1235 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1236 if (!same_inode)
1237 xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
1238 inode_unlock(inode_out);
1239 if (!same_inode)
1240 inode_unlock_shared(inode_in);
1241 }
1243 /*
1244 * If we're reflinking to a point past the destination file's EOF, we must
1245 * zero any speculative post-EOF preallocations that sit between the old EOF
1246 * and the destination file offset.
1247 */
1248 static int
1249 xfs_reflink_zero_posteof(
1250 struct xfs_inode *ip,
1251 loff_t pos)
1252 {
1253 loff_t isize = i_size_read(VFS_I(ip));
1255 if (pos <= isize)
1256 return 0;
1258 trace_xfs_zero_eof(ip, isize, pos - isize);
1259 return iomap_zero_range(VFS_I(ip), isize, pos - isize, NULL,
1260 &xfs_iomap_ops);
1261 }
1263 /*
1264 * Prepare two files for range cloning. Upon a successful return both inodes
1265 * will have the iolock and mmaplock held, the page cache of the out file will
1266 * be truncated, and any leases on the out file will have been broken. This
1267 * function borrows heavily from xfs_file_aio_write_checks.
1268 *
1269 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1270 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1271 * EOF block in the source dedupe range because it's not a complete block match,
1272 * hence can introduce a corruption into the file that has it's block replaced.
1273 *
1274 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1275 * "block aligned" for the purposes of cloning entire files. However, if the
1276 * source file range includes the EOF block and it lands within the existing EOF
1277 * of the destination file, then we can expose stale data from beyond the source
1278 * file EOF in the destination file.
1279 *
1280 * XFS doesn't support partial block sharing, so in both cases we have check
1281 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1282 * down to the previous whole block and ignore the partial EOF block. While this
1283 * means we can't dedupe the last block of a file, this is an acceptible
1284 * tradeoff for simplicity on implementation.
1285 *
1286 * For cloning, we want to share the partial EOF block if it is also the new EOF
1287 * block of the destination file. If the partial EOF block lies inside the
1288 * existing destination EOF, then we have to abort the clone to avoid exposing
1289 * stale data in the destination file. Hence we reject these clone attempts with
1290 * -EINVAL in this case.
1291 */
1292 STATIC int
1293 xfs_reflink_remap_prep(
1294 struct file *file_in,
1295 loff_t pos_in,
1296 struct file *file_out,
1297 loff_t pos_out,
1298 u64 *len,
1299 bool is_dedupe)
1300 {
1301 struct inode *inode_in = file_inode(file_in);
1302 struct xfs_inode *src = XFS_I(inode_in);
1303 struct inode *inode_out = file_inode(file_out);
1304 struct xfs_inode *dest = XFS_I(inode_out);
1305 bool same_inode = (inode_in == inode_out);
1306 u64 blkmask = i_blocksize(inode_in) - 1;
1307 ssize_t ret;
1309 /* Lock both files against IO */
1310 ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1311 if (ret)
1312 return ret;
1313 if (same_inode)
1314 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1315 else
1316 xfs_lock_two_inodes(src, XFS_MMAPLOCK_SHARED, dest,
1317 XFS_MMAPLOCK_EXCL);
1319 /* Check file eligibility and prepare for block sharing. */
1320 ret = -EINVAL;
1321 /* Don't reflink realtime inodes */
1322 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1323 goto out_unlock;
1325 /* Don't share DAX file data for now. */
1326 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1327 goto out_unlock;
1329 ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1330 len, is_dedupe);
1331 if (ret <= 0)
1332 goto out_unlock;
1334 /*
1335 * If the dedupe data matches, chop off the partial EOF block
1336 * from the source file so we don't try to dedupe the partial
1337 * EOF block.
1338 */
1339 if (is_dedupe) {
1340 *len &= ~blkmask;
1341 } else if (*len & blkmask) {
1342 /*
1343 * The user is attempting to share a partial EOF block,
1344 * if it's inside the destination EOF then reject it.
1345 */
1346 if (pos_out + *len < i_size_read(inode_out)) {
1347 ret = -EINVAL;
1348 goto out_unlock;
1349 }
1350 }
1352 /* Attach dquots to dest inode before changing block map */
1353 ret = xfs_qm_dqattach(dest);
1354 if (ret)
1355 goto out_unlock;
1357 /*
1358 * Zero existing post-eof speculative preallocations in the destination
1359 * file.
1360 */
1361 ret = xfs_reflink_zero_posteof(dest, pos_out);
1362 if (ret)
1363 goto out_unlock;
1365 /* Set flags and remap blocks. */
1366 ret = xfs_reflink_set_inode_flag(src, dest);
1367 if (ret)
1368 goto out_unlock;
1370 /* Zap any page cache for the destination file's range. */
1371 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1372 PAGE_ALIGN(pos_out + *len) - 1);
1374 /* If we're altering the file contents... */
1375 if (!is_dedupe) {
1376 /*
1377 * ...update the timestamps (which will grab the ilock again
1378 * from xfs_fs_dirty_inode, so we have to call it before we
1379 * take the ilock).
1380 */
1381 if (!(file_out->f_mode & FMODE_NOCMTIME)) {
1382 ret = file_update_time(file_out);
1383 if (ret)
1384 goto out_unlock;
1385 }
1387 /*
1388 * ...clear the security bits if the process is not being run
1389 * by root. This keeps people from modifying setuid and setgid
1390 * binaries.
1391 */
1392 ret = file_remove_privs(file_out);
1393 if (ret)
1394 goto out_unlock;
1395 }
1397 return 1;
1398 out_unlock:
1399 xfs_reflink_remap_unlock(file_in, file_out);
1400 return ret;
1401 }
1403 /*
1404 * Link a range of blocks from one file to another.
1405 */
1406 int
1407 xfs_reflink_remap_range(
1408 struct file *file_in,
1409 loff_t pos_in,
1410 struct file *file_out,
1411 loff_t pos_out,
1412 u64 len,
1413 bool is_dedupe)
1414 {
1415 struct inode *inode_in = file_inode(file_in);
1416 struct xfs_inode *src = XFS_I(inode_in);
1417 struct inode *inode_out = file_inode(file_out);
1418 struct xfs_inode *dest = XFS_I(inode_out);
1419 struct xfs_mount *mp = src->i_mount;
1420 xfs_fileoff_t sfsbno, dfsbno;
1421 xfs_filblks_t fsblen;
1422 xfs_extlen_t cowextsize;
1423 ssize_t ret;
1425 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1426 return -EOPNOTSUPP;
1428 if (XFS_FORCED_SHUTDOWN(mp))
1429 return -EIO;
1431 /* Prepare and then clone file data. */
1432 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1433 &len, is_dedupe);
1434 if (ret <= 0)
1435 return ret;
1437 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1439 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1440 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1441 fsblen = XFS_B_TO_FSB(mp, len);
1442 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1443 pos_out + len);
1444 if (ret)
1445 goto out_unlock;
1447 /*
1448 * Carry the cowextsize hint from src to dest if we're sharing the
1449 * entire source file to the entire destination file, the source file
1450 * has a cowextsize hint, and the destination file does not.
1451 */
1452 cowextsize = 0;
1453 if (pos_in == 0 && len == i_size_read(inode_in) &&
1454 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1455 pos_out == 0 && len >= i_size_read(inode_out) &&
1456 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1457 cowextsize = src->i_d.di_cowextsize;
1459 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1460 is_dedupe);
1462 out_unlock:
1463 xfs_reflink_remap_unlock(file_in, file_out);
1464 if (ret)
1465 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1466 return ret;
1467 }
1469 /*
1470 * The user wants to preemptively CoW all shared blocks in this file,
1471 * which enables us to turn off the reflink flag. Iterate all
1472 * extents which are not prealloc/delalloc to see which ranges are
1473 * mentioned in the refcount tree, then read those blocks into the
1474 * pagecache, dirty them, fsync them back out, and then we can update
1475 * the inode flag. What happens if we run out of memory? :)
1476 */
1477 STATIC int
1478 xfs_reflink_dirty_extents(
1479 struct xfs_inode *ip,
1480 xfs_fileoff_t fbno,
1481 xfs_filblks_t end,
1482 xfs_off_t isize)
1483 {
1484 struct xfs_mount *mp = ip->i_mount;
1485 xfs_agnumber_t agno;
1486 xfs_agblock_t agbno;
1487 xfs_extlen_t aglen;
1488 xfs_agblock_t rbno;
1489 xfs_extlen_t rlen;
1490 xfs_off_t fpos;
1491 xfs_off_t flen;
1492 struct xfs_bmbt_irec map[2];
1493 int nmaps;
1494 int error = 0;
1496 while (end - fbno > 0) {
1497 nmaps = 1;
1498 /*
1499 * Look for extents in the file. Skip holes, delalloc, or
1500 * unwritten extents; they can't be reflinked.
1501 */
1502 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1503 if (error)
1504 goto out;
1505 if (nmaps == 0)
1506 break;
1507 if (!xfs_bmap_is_real_extent(&map[0]))
1508 goto next;
1510 map[1] = map[0];
1511 while (map[1].br_blockcount) {
1512 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1513 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1514 aglen = map[1].br_blockcount;
1516 error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1517 aglen, &rbno, &rlen, true);
1518 if (error)
1519 goto out;
1520 if (rbno == NULLAGBLOCK)
1521 break;
1523 /* Dirty the pages */
1524 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1525 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1526 (rbno - agbno));
1527 flen = XFS_FSB_TO_B(mp, rlen);
1528 if (fpos + flen > isize)
1529 flen = isize - fpos;
1530 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1531 &xfs_iomap_ops);
1532 xfs_ilock(ip, XFS_ILOCK_EXCL);
1533 if (error)
1534 goto out;
1536 map[1].br_blockcount -= (rbno - agbno + rlen);
1537 map[1].br_startoff += (rbno - agbno + rlen);
1538 map[1].br_startblock += (rbno - agbno + rlen);
1539 }
1541 next:
1542 fbno = map[0].br_startoff + map[0].br_blockcount;
1543 }
1544 out:
1545 return error;
1546 }
1548 /* Does this inode need the reflink flag? */
1549 int
1550 xfs_reflink_inode_has_shared_extents(
1551 struct xfs_trans *tp,
1552 struct xfs_inode *ip,
1553 bool *has_shared)
1554 {
1555 struct xfs_bmbt_irec got;
1556 struct xfs_mount *mp = ip->i_mount;
1557 struct xfs_ifork *ifp;
1558 xfs_agnumber_t agno;
1559 xfs_agblock_t agbno;
1560 xfs_extlen_t aglen;
1561 xfs_agblock_t rbno;
1562 xfs_extlen_t rlen;
1563 struct xfs_iext_cursor icur;
1564 bool found;
1565 int error;
1567 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1568 if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1569 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1570 if (error)
1571 return error;
1572 }
1574 *has_shared = false;
1575 found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1576 while (found) {
1577 if (isnullstartblock(got.br_startblock) ||
1578 got.br_state != XFS_EXT_NORM)
1579 goto next;
1580 agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1581 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1582 aglen = got.br_blockcount;
1584 error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1585 &rbno, &rlen, false);
1586 if (error)
1587 return error;
1588 /* Is there still a shared block here? */
1589 if (rbno != NULLAGBLOCK) {
1590 *has_shared = true;
1591 return 0;
1592 }
1593 next:
1594 found = xfs_iext_next_extent(ifp, &icur, &got);
1595 }
1597 return 0;
1598 }
1600 /*
1601 * Clear the inode reflink flag if there are no shared extents.
1602 *
1603 * The caller is responsible for joining the inode to the transaction passed in.
1604 * The inode will be joined to the transaction that is returned to the caller.
1605 */
1606 int
1607 xfs_reflink_clear_inode_flag(
1608 struct xfs_inode *ip,
1609 struct xfs_trans **tpp)
1610 {
1611 bool needs_flag;
1612 int error = 0;
1614 ASSERT(xfs_is_reflink_inode(ip));
1616 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1617 if (error || needs_flag)
1618 return error;
1620 /*
1621 * We didn't find any shared blocks so turn off the reflink flag.
1622 * First, get rid of any leftover CoW mappings.
1623 */
1624 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1625 if (error)
1626 return error;
1628 /* Clear the inode flag. */
1629 trace_xfs_reflink_unset_inode_flag(ip);
1630 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1631 xfs_inode_clear_cowblocks_tag(ip);
1632 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1634 return error;
1635 }
1637 /*
1638 * Clear the inode reflink flag if there are no shared extents and the size
1639 * hasn't changed.
1640 */
1641 STATIC int
1642 xfs_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;
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;
1654 xfs_ilock(ip, XFS_ILOCK_EXCL);
1655 xfs_trans_ijoin(tp, ip, 0);
1657 error = xfs_reflink_clear_inode_flag(ip, &tp);
1658 if (error)
1659 goto cancel;
1661 error = xfs_trans_commit(tp);
1662 if (error)
1663 goto out;
1665 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1666 return 0;
1667 cancel:
1668 xfs_trans_cancel(tp);
1669 out:
1670 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1671 return error;
1672 }
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 */
1678 int
1679 xfs_reflink_unshare(
1680 struct xfs_inode *ip,
1681 xfs_off_t offset,
1682 xfs_off_t len)
1683 {
1684 struct xfs_mount *mp = ip->i_mount;
1685 xfs_fileoff_t fbno;
1686 xfs_filblks_t end;
1687 xfs_off_t isize;
1688 int error;
1690 if (!xfs_is_reflink_inode(ip))
1691 return 0;
1693 trace_xfs_reflink_unshare(ip, offset, len);
1695 inode_dio_wait(VFS_I(ip));
1697 /* Try to CoW the selected ranges */
1698 xfs_ilock(ip, XFS_ILOCK_EXCL);
1699 fbno = XFS_B_TO_FSBT(mp, offset);
1700 isize = i_size_read(VFS_I(ip));
1701 end = XFS_B_TO_FSB(mp, offset + len);
1702 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1703 if (error)
1704 goto out_unlock;
1705 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1707 /* Wait for the IO to finish */
1708 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1709 if (error)
1710 goto out;
1712 /* Turn off the reflink flag if possible. */
1713 error = xfs_reflink_try_clear_inode_flag(ip);
1714 if (error)
1715 goto out;
1717 return 0;
1719 out_unlock:
1720 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1721 out:
1722 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1723 return error;
1724 }