1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * Copyright (c) 2016-2018 Christoph Hellwig.
5 * All Rights Reserved.
6 */
7#include "xfs.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_inode.h"
14#include "xfs_trans.h"
15#include "xfs_iomap.h"
16#include "xfs_trace.h"
17#include "xfs_bmap.h"
18#include "xfs_bmap_util.h"
19#include "xfs_reflink.h"
20#include "xfs_errortag.h"
21#include "xfs_error.h"
22
23struct xfs_writepage_ctx {
24 struct iomap_writepage_ctx ctx;
25 unsigned int data_seq;
26 unsigned int cow_seq;
27};
28
29static inline struct xfs_writepage_ctx *
30XFS_WPC(struct iomap_writepage_ctx *ctx)
31{
32 return container_of(ctx, struct xfs_writepage_ctx, ctx);
33}
34
35/*
36 * Fast and loose check if this write could update the on-disk inode size.
37 */
38static inline bool xfs_ioend_is_append(struct iomap_ioend *ioend)
39{
40 return ioend->io_offset + ioend->io_size >
41 XFS_I(inode: ioend->io_inode)->i_disk_size;
42}
43
44/*
45 * Update on-disk file size now that data has been written to disk.
46 */
47int
48xfs_setfilesize(
49 struct xfs_inode *ip,
50 xfs_off_t offset,
51 size_t size)
52{
53 struct xfs_mount *mp = ip->i_mount;
54 struct xfs_trans *tp;
55 xfs_fsize_t isize;
56 int error;
57
58 error = xfs_trans_alloc(mp, resp: &M_RES(mp)->tr_fsyncts, blocks: 0, rtextents: 0, flags: 0, tpp: &tp);
59 if (error)
60 return error;
61
62 xfs_ilock(ip, XFS_ILOCK_EXCL);
63 isize = xfs_new_eof(ip, offset + size);
64 if (!isize) {
65 xfs_iunlock(ip, XFS_ILOCK_EXCL);
66 xfs_trans_cancel(tp);
67 return 0;
68 }
69
70 trace_xfs_setfilesize(ip, offset, count: size);
71
72 ip->i_disk_size = isize;
73 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
74 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
75
76 return xfs_trans_commit(tp);
77}
78
79/*
80 * IO write completion.
81 */
82STATIC void
83xfs_end_ioend(
84 struct iomap_ioend *ioend)
85{
86 struct xfs_inode *ip = XFS_I(inode: ioend->io_inode);
87 struct xfs_mount *mp = ip->i_mount;
88 xfs_off_t offset = ioend->io_offset;
89 size_t size = ioend->io_size;
90 unsigned int nofs_flag;
91 int error;
92
93 /*
94 * We can allocate memory here while doing writeback on behalf of
95 * memory reclaim. To avoid memory allocation deadlocks set the
96 * task-wide nofs context for the following operations.
97 */
98 nofs_flag = memalloc_nofs_save();
99
100 /*
101 * Just clean up the in-memory structures if the fs has been shut down.
102 */
103 if (xfs_is_shutdown(mp)) {
104 error = -EIO;
105 goto done;
106 }
107
108 /*
109 * Clean up all COW blocks and underlying data fork delalloc blocks on
110 * I/O error. The delalloc punch is required because this ioend was
111 * mapped to blocks in the COW fork and the associated pages are no
112 * longer dirty. If we don't remove delalloc blocks here, they become
113 * stale and can corrupt free space accounting on unmount.
114 */
115 error = blk_status_to_errno(status: ioend->io_bio.bi_status);
116 if (unlikely(error)) {
117 if (ioend->io_flags & IOMAP_F_SHARED) {
118 xfs_reflink_cancel_cow_range(ip, offset, count: size, cancel_real: true);
119 xfs_bmap_punch_delalloc_range(ip, start_byte: offset,
120 end_byte: offset + size);
121 }
122 goto done;
123 }
124
125 /*
126 * Success: commit the COW or unwritten blocks if needed.
127 */
128 if (ioend->io_flags & IOMAP_F_SHARED)
129 error = xfs_reflink_end_cow(ip, offset, count: size);
130 else if (ioend->io_type == IOMAP_UNWRITTEN)
131 error = xfs_iomap_write_unwritten(ip, offset, size, false);
132
133 if (!error && xfs_ioend_is_append(ioend))
134 error = xfs_setfilesize(ip, offset: ioend->io_offset, size: ioend->io_size);
135done:
136 iomap_finish_ioends(ioend, error);
137 memalloc_nofs_restore(flags: nofs_flag);
138}
139
140/*
141 * Finish all pending IO completions that require transactional modifications.
142 *
143 * We try to merge physical and logically contiguous ioends before completion to
144 * minimise the number of transactions we need to perform during IO completion.
145 * Both unwritten extent conversion and COW remapping need to iterate and modify
146 * one physical extent at a time, so we gain nothing by merging physically
147 * discontiguous extents here.
148 *
149 * The ioend chain length that we can be processing here is largely unbound in
150 * length and we may have to perform significant amounts of work on each ioend
151 * to complete it. Hence we have to be careful about holding the CPU for too
152 * long in this loop.
153 */
154void
155xfs_end_io(
156 struct work_struct *work)
157{
158 struct xfs_inode *ip =
159 container_of(work, struct xfs_inode, i_ioend_work);
160 struct iomap_ioend *ioend;
161 struct list_head tmp;
162 unsigned long flags;
163
164 spin_lock_irqsave(&ip->i_ioend_lock, flags);
165 list_replace_init(old: &ip->i_ioend_list, new: &tmp);
166 spin_unlock_irqrestore(lock: &ip->i_ioend_lock, flags);
167
168 iomap_sort_ioends(ioend_list: &tmp);
169 while ((ioend = list_first_entry_or_null(&tmp, struct iomap_ioend,
170 io_list))) {
171 list_del_init(entry: &ioend->io_list);
172 iomap_ioend_try_merge(ioend, more_ioends: &tmp);
173 xfs_end_ioend(ioend);
174 cond_resched();
175 }
176}
177
178STATIC void
179xfs_end_bio(
180 struct bio *bio)
181{
182 struct iomap_ioend *ioend = iomap_ioend_from_bio(bio);
183 struct xfs_inode *ip = XFS_I(inode: ioend->io_inode);
184 unsigned long flags;
185
186 spin_lock_irqsave(&ip->i_ioend_lock, flags);
187 if (list_empty(head: &ip->i_ioend_list))
188 WARN_ON_ONCE(!queue_work(ip->i_mount->m_unwritten_workqueue,
189 &ip->i_ioend_work));
190 list_add_tail(new: &ioend->io_list, head: &ip->i_ioend_list);
191 spin_unlock_irqrestore(lock: &ip->i_ioend_lock, flags);
192}
193
194/*
195 * Fast revalidation of the cached writeback mapping. Return true if the current
196 * mapping is valid, false otherwise.
197 */
198static bool
199xfs_imap_valid(
200 struct iomap_writepage_ctx *wpc,
201 struct xfs_inode *ip,
202 loff_t offset)
203{
204 if (offset < wpc->iomap.offset ||
205 offset >= wpc->iomap.offset + wpc->iomap.length)
206 return false;
207 /*
208 * If this is a COW mapping, it is sufficient to check that the mapping
209 * covers the offset. Be careful to check this first because the caller
210 * can revalidate a COW mapping without updating the data seqno.
211 */
212 if (wpc->iomap.flags & IOMAP_F_SHARED)
213 return true;
214
215 /*
216 * This is not a COW mapping. Check the sequence number of the data fork
217 * because concurrent changes could have invalidated the extent. Check
218 * the COW fork because concurrent changes since the last time we
219 * checked (and found nothing at this offset) could have added
220 * overlapping blocks.
221 */
222 if (XFS_WPC(ctx: wpc)->data_seq != READ_ONCE(ip->i_df.if_seq)) {
223 trace_xfs_wb_data_iomap_invalid(ip, &wpc->iomap,
224 XFS_WPC(wpc)->data_seq, XFS_DATA_FORK);
225 return false;
226 }
227 if (xfs_inode_has_cow_data(ip) &&
228 XFS_WPC(ctx: wpc)->cow_seq != READ_ONCE(ip->i_cowfp->if_seq)) {
229 trace_xfs_wb_cow_iomap_invalid(ip, &wpc->iomap,
230 XFS_WPC(wpc)->cow_seq, XFS_COW_FORK);
231 return false;
232 }
233 return true;
234}
235
236/*
237 * Pass in a dellalloc extent and convert it to real extents, return the real
238 * extent that maps offset_fsb in wpc->iomap.
239 *
240 * The current page is held locked so nothing could have removed the block
241 * backing offset_fsb, although it could have moved from the COW to the data
242 * fork by another thread.
243 */
244static int
245xfs_convert_blocks(
246 struct iomap_writepage_ctx *wpc,
247 struct xfs_inode *ip,
248 int whichfork,
249 loff_t offset)
250{
251 int error;
252 unsigned *seq;
253
254 if (whichfork == XFS_COW_FORK)
255 seq = &XFS_WPC(ctx: wpc)->cow_seq;
256 else
257 seq = &XFS_WPC(ctx: wpc)->data_seq;
258
259 /*
260 * Attempt to allocate whatever delalloc extent currently backs offset
261 * and put the result into wpc->iomap. Allocate in a loop because it
262 * may take several attempts to allocate real blocks for a contiguous
263 * delalloc extent if free space is sufficiently fragmented.
264 */
265 do {
266 error = xfs_bmapi_convert_delalloc(ip, whichfork, offset,
267 &wpc->iomap, seq);
268 if (error)
269 return error;
270 } while (wpc->iomap.offset + wpc->iomap.length <= offset);
271
272 return 0;
273}
274
275static int
276xfs_map_blocks(
277 struct iomap_writepage_ctx *wpc,
278 struct inode *inode,
279 loff_t offset,
280 unsigned int len)
281{
282 struct xfs_inode *ip = XFS_I(inode);
283 struct xfs_mount *mp = ip->i_mount;
284 ssize_t count = i_blocksize(node: inode);
285 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
286 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
287 xfs_fileoff_t cow_fsb;
288 int whichfork;
289 struct xfs_bmbt_irec imap;
290 struct xfs_iext_cursor icur;
291 int retries = 0;
292 int error = 0;
293
294 if (xfs_is_shutdown(mp))
295 return -EIO;
296
297 XFS_ERRORTAG_DELAY(mp, XFS_ERRTAG_WB_DELAY_MS);
298
299 /*
300 * COW fork blocks can overlap data fork blocks even if the blocks
301 * aren't shared. COW I/O always takes precedent, so we must always
302 * check for overlap on reflink inodes unless the mapping is already a
303 * COW one, or the COW fork hasn't changed from the last time we looked
304 * at it.
305 *
306 * It's safe to check the COW fork if_seq here without the ILOCK because
307 * we've indirectly protected against concurrent updates: writeback has
308 * the page locked, which prevents concurrent invalidations by reflink
309 * and directio and prevents concurrent buffered writes to the same
310 * page. Changes to if_seq always happen under i_lock, which protects
311 * against concurrent updates and provides a memory barrier on the way
312 * out that ensures that we always see the current value.
313 */
314 if (xfs_imap_valid(wpc, ip, offset))
315 return 0;
316
317 /*
318 * If we don't have a valid map, now it's time to get a new one for this
319 * offset. This will convert delayed allocations (including COW ones)
320 * into real extents. If we return without a valid map, it means we
321 * landed in a hole and we skip the block.
322 */
323retry:
324 cow_fsb = NULLFILEOFF;
325 whichfork = XFS_DATA_FORK;
326 xfs_ilock(ip, XFS_ILOCK_SHARED);
327 ASSERT(!xfs_need_iread_extents(&ip->i_df));
328
329 /*
330 * Check if this is offset is covered by a COW extents, and if yes use
331 * it directly instead of looking up anything in the data fork.
332 */
333 if (xfs_inode_has_cow_data(ip) &&
334 xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &imap))
335 cow_fsb = imap.br_startoff;
336 if (cow_fsb != NULLFILEOFF && cow_fsb <= offset_fsb) {
337 XFS_WPC(ctx: wpc)->cow_seq = READ_ONCE(ip->i_cowfp->if_seq);
338 xfs_iunlock(ip, XFS_ILOCK_SHARED);
339
340 whichfork = XFS_COW_FORK;
341 goto allocate_blocks;
342 }
343
344 /*
345 * No COW extent overlap. Revalidate now that we may have updated
346 * ->cow_seq. If the data mapping is still valid, we're done.
347 */
348 if (xfs_imap_valid(wpc, ip, offset)) {
349 xfs_iunlock(ip, XFS_ILOCK_SHARED);
350 return 0;
351 }
352
353 /*
354 * If we don't have a valid map, now it's time to get a new one for this
355 * offset. This will convert delayed allocations (including COW ones)
356 * into real extents.
357 */
358 if (!xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap))
359 imap.br_startoff = end_fsb; /* fake a hole past EOF */
360 XFS_WPC(ctx: wpc)->data_seq = READ_ONCE(ip->i_df.if_seq);
361 xfs_iunlock(ip, XFS_ILOCK_SHARED);
362
363 /* landed in a hole or beyond EOF? */
364 if (imap.br_startoff > offset_fsb) {
365 imap.br_blockcount = imap.br_startoff - offset_fsb;
366 imap.br_startoff = offset_fsb;
367 imap.br_startblock = HOLESTARTBLOCK;
368 imap.br_state = XFS_EXT_NORM;
369 }
370
371 /*
372 * Truncate to the next COW extent if there is one. This is the only
373 * opportunity to do this because we can skip COW fork lookups for the
374 * subsequent blocks in the mapping; however, the requirement to treat
375 * the COW range separately remains.
376 */
377 if (cow_fsb != NULLFILEOFF &&
378 cow_fsb < imap.br_startoff + imap.br_blockcount)
379 imap.br_blockcount = cow_fsb - imap.br_startoff;
380
381 /* got a delalloc extent? */
382 if (imap.br_startblock != HOLESTARTBLOCK &&
383 isnullstartblock(imap.br_startblock))
384 goto allocate_blocks;
385
386 xfs_bmbt_to_iomap(ip, iomap: &wpc->iomap, imap: &imap, mapping_flags: 0, iomap_flags: 0, sequence_cookie: XFS_WPC(ctx: wpc)->data_seq);
387 trace_xfs_map_blocks_found(ip, offset, count, whichfork, irec: &imap);
388 return 0;
389allocate_blocks:
390 error = xfs_convert_blocks(wpc, ip, whichfork, offset);
391 if (error) {
392 /*
393 * If we failed to find the extent in the COW fork we might have
394 * raced with a COW to data fork conversion or truncate.
395 * Restart the lookup to catch the extent in the data fork for
396 * the former case, but prevent additional retries to avoid
397 * looping forever for the latter case.
398 */
399 if (error == -EAGAIN && whichfork == XFS_COW_FORK && !retries++)
400 goto retry;
401 ASSERT(error != -EAGAIN);
402 return error;
403 }
404
405 /*
406 * Due to merging the return real extent might be larger than the
407 * original delalloc one. Trim the return extent to the next COW
408 * boundary again to force a re-lookup.
409 */
410 if (whichfork != XFS_COW_FORK && cow_fsb != NULLFILEOFF) {
411 loff_t cow_offset = XFS_FSB_TO_B(mp, cow_fsb);
412
413 if (cow_offset < wpc->iomap.offset + wpc->iomap.length)
414 wpc->iomap.length = cow_offset - wpc->iomap.offset;
415 }
416
417 ASSERT(wpc->iomap.offset <= offset);
418 ASSERT(wpc->iomap.offset + wpc->iomap.length > offset);
419 trace_xfs_map_blocks_alloc(ip, offset, count, whichfork, irec: &imap);
420 return 0;
421}
422
423static int
424xfs_prepare_ioend(
425 struct iomap_ioend *ioend,
426 int status)
427{
428 unsigned int nofs_flag;
429
430 /*
431 * We can allocate memory here while doing writeback on behalf of
432 * memory reclaim. To avoid memory allocation deadlocks set the
433 * task-wide nofs context for the following operations.
434 */
435 nofs_flag = memalloc_nofs_save();
436
437 /* Convert CoW extents to regular */
438 if (!status && (ioend->io_flags & IOMAP_F_SHARED)) {
439 status = xfs_reflink_convert_cow(ip: XFS_I(inode: ioend->io_inode),
440 offset: ioend->io_offset, count: ioend->io_size);
441 }
442
443 memalloc_nofs_restore(flags: nofs_flag);
444
445 /* send ioends that might require a transaction to the completion wq */
446 if (xfs_ioend_is_append(ioend) || ioend->io_type == IOMAP_UNWRITTEN ||
447 (ioend->io_flags & IOMAP_F_SHARED))
448 ioend->io_bio.bi_end_io = xfs_end_bio;
449 return status;
450}
451
452/*
453 * If the folio has delalloc blocks on it, the caller is asking us to punch them
454 * out. If we don't, we can leave a stale delalloc mapping covered by a clean
455 * page that needs to be dirtied again before the delalloc mapping can be
456 * converted. This stale delalloc mapping can trip up a later direct I/O read
457 * operation on the same region.
458 *
459 * We prevent this by truncating away the delalloc regions on the folio. Because
460 * they are delalloc, we can do this without needing a transaction. Indeed - if
461 * we get ENOSPC errors, we have to be able to do this truncation without a
462 * transaction as there is no space left for block reservation (typically why
463 * we see a ENOSPC in writeback).
464 */
465static void
466xfs_discard_folio(
467 struct folio *folio,
468 loff_t pos)
469{
470 struct xfs_inode *ip = XFS_I(inode: folio->mapping->host);
471 struct xfs_mount *mp = ip->i_mount;
472 int error;
473
474 if (xfs_is_shutdown(mp))
475 return;
476
477 xfs_alert_ratelimited(mp,
478 "page discard on page "PTR_FMT", inode 0x%llx, pos %llu.",
479 folio, ip->i_ino, pos);
480
481 /*
482 * The end of the punch range is always the offset of the first
483 * byte of the next folio. Hence the end offset is only dependent on the
484 * folio itself and not the start offset that is passed in.
485 */
486 error = xfs_bmap_punch_delalloc_range(ip, start_byte: pos,
487 end_byte: folio_pos(folio) + folio_size(folio));
488
489 if (error && !xfs_is_shutdown(mp))
490 xfs_alert(mp, "page discard unable to remove delalloc mapping.");
491}
492
493static const struct iomap_writeback_ops xfs_writeback_ops = {
494 .map_blocks = xfs_map_blocks,
495 .prepare_ioend = xfs_prepare_ioend,
496 .discard_folio = xfs_discard_folio,
497};
498
499STATIC int
500xfs_vm_writepages(
501 struct address_space *mapping,
502 struct writeback_control *wbc)
503{
504 struct xfs_writepage_ctx wpc = { };
505
506 xfs_iflags_clear(ip: XFS_I(inode: mapping->host), XFS_ITRUNCATED);
507 return iomap_writepages(mapping, wbc, wpc: &wpc.ctx, ops: &xfs_writeback_ops);
508}
509
510STATIC int
511xfs_dax_writepages(
512 struct address_space *mapping,
513 struct writeback_control *wbc)
514{
515 struct xfs_inode *ip = XFS_I(inode: mapping->host);
516
517 xfs_iflags_clear(ip, XFS_ITRUNCATED);
518 return dax_writeback_mapping_range(mapping,
519 xfs_inode_buftarg(ip)->bt_daxdev, wbc);
520}
521
522STATIC sector_t
523xfs_vm_bmap(
524 struct address_space *mapping,
525 sector_t block)
526{
527 struct xfs_inode *ip = XFS_I(inode: mapping->host);
528
529 trace_xfs_vm_bmap(ip);
530
531 /*
532 * The swap code (ab-)uses ->bmap to get a block mapping and then
533 * bypasses the file system for actual I/O. We really can't allow
534 * that on reflinks inodes, so we have to skip out here. And yes,
535 * 0 is the magic code for a bmap error.
536 *
537 * Since we don't pass back blockdev info, we can't return bmap
538 * information for rt files either.
539 */
540 if (xfs_is_cow_inode(ip) || XFS_IS_REALTIME_INODE(ip))
541 return 0;
542 return iomap_bmap(mapping, bno: block, ops: &xfs_read_iomap_ops);
543}
544
545STATIC int
546xfs_vm_read_folio(
547 struct file *unused,
548 struct folio *folio)
549{
550 return iomap_read_folio(folio, ops: &xfs_read_iomap_ops);
551}
552
553STATIC void
554xfs_vm_readahead(
555 struct readahead_control *rac)
556{
557 iomap_readahead(rac, ops: &xfs_read_iomap_ops);
558}
559
560static int
561xfs_iomap_swapfile_activate(
562 struct swap_info_struct *sis,
563 struct file *swap_file,
564 sector_t *span)
565{
566 sis->bdev = xfs_inode_buftarg(XFS_I(file_inode(swap_file)))->bt_bdev;
567 return iomap_swapfile_activate(sis, swap_file, pagespan: span,
568 ops: &xfs_read_iomap_ops);
569}
570
571const struct address_space_operations xfs_address_space_operations = {
572 .read_folio = xfs_vm_read_folio,
573 .readahead = xfs_vm_readahead,
574 .writepages = xfs_vm_writepages,
575 .dirty_folio = iomap_dirty_folio,
576 .release_folio = iomap_release_folio,
577 .invalidate_folio = iomap_invalidate_folio,
578 .bmap = xfs_vm_bmap,
579 .migrate_folio = filemap_migrate_folio,
580 .is_partially_uptodate = iomap_is_partially_uptodate,
581 .error_remove_folio = generic_error_remove_folio,
582 .swap_activate = xfs_iomap_swapfile_activate,
583};
584
585const struct address_space_operations xfs_dax_aops = {
586 .writepages = xfs_dax_writepages,
587 .dirty_folio = noop_dirty_folio,
588 .swap_activate = xfs_iomap_swapfile_activate,
589};
590

source code of linux/fs/xfs/xfs_aops.c