1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
4 */
5
6 #include <linux/fs.h>
7 #include <linux/slab.h>
8 #include <linux/highmem.h>
9 #include <linux/pagemap.h>
10 #include <asm/byteorder.h>
11 #include <linux/swap.h>
12 #include <linux/mpage.h>
13 #include <linux/quotaops.h>
14 #include <linux/blkdev.h>
15 #include <linux/uio.h>
16 #include <linux/mm.h>
17
18 #include <cluster/masklog.h>
19
20 #include "ocfs2.h"
21
22 #include "alloc.h"
23 #include "aops.h"
24 #include "dlmglue.h"
25 #include "extent_map.h"
26 #include "file.h"
27 #include "inode.h"
28 #include "journal.h"
29 #include "suballoc.h"
30 #include "super.h"
31 #include "symlink.h"
32 #include "refcounttree.h"
33 #include "ocfs2_trace.h"
34
35 #include "buffer_head_io.h"
36 #include "dir.h"
37 #include "namei.h"
38 #include "sysfile.h"
39
40 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
41 struct buffer_head *bh_result, int create)
42 {
43 int err = -EIO;
44 int status;
45 struct ocfs2_dinode *fe = NULL;
46 struct buffer_head *bh = NULL;
47 struct buffer_head *buffer_cache_bh = NULL;
48 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
49 void *kaddr;
50
51 trace_ocfs2_symlink_get_block(
52 (unsigned long long)OCFS2_I(inode)->ip_blkno,
53 (unsigned long long)iblock, bh_result, create);
54
55 BUG_ON(ocfs2_inode_is_fast_symlink(inode));
56
57 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
58 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
59 (unsigned long long)iblock);
60 goto bail;
61 }
62
63 status = ocfs2_read_inode_block(inode, &bh);
64 if (status < 0) {
65 mlog_errno(status);
66 goto bail;
67 }
68 fe = (struct ocfs2_dinode *) bh->b_data;
69
70 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
71 le32_to_cpu(fe->i_clusters))) {
72 err = -ENOMEM;
73 mlog(ML_ERROR, "block offset is outside the allocated size: "
74 "%llu\n", (unsigned long long)iblock);
75 goto bail;
76 }
77
78 /* We don't use the page cache to create symlink data, so if
79 * need be, copy it over from the buffer cache. */
80 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
81 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
82 iblock;
83 buffer_cache_bh = sb_getblk(osb->sb, blkno);
84 if (!buffer_cache_bh) {
85 err = -ENOMEM;
86 mlog(ML_ERROR, "couldn't getblock for symlink!\n");
87 goto bail;
88 }
89
90 /* we haven't locked out transactions, so a commit
91 * could've happened. Since we've got a reference on
92 * the bh, even if it commits while we're doing the
93 * copy, the data is still good. */
94 if (buffer_jbd(buffer_cache_bh)
95 && ocfs2_inode_is_new(inode)) {
96 kaddr = kmap_atomic(bh_result->b_page);
97 if (!kaddr) {
98 mlog(ML_ERROR, "couldn't kmap!\n");
99 goto bail;
100 }
101 memcpy(kaddr + (bh_result->b_size * iblock),
102 buffer_cache_bh->b_data,
103 bh_result->b_size);
104 kunmap_atomic(kaddr);
105 set_buffer_uptodate(bh_result);
106 }
107 brelse(buffer_cache_bh);
108 }
109
110 map_bh(bh_result, inode->i_sb,
111 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
112
113 err = 0;
114
115 bail:
116 brelse(bh);
117
118 return err;
119 }
120
121 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
122 struct buffer_head *bh_result, int create)
123 {
124 int ret = 0;
125 struct ocfs2_inode_info *oi = OCFS2_I(inode);
126
127 down_read(&oi->ip_alloc_sem);
128 ret = ocfs2_get_block(inode, iblock, bh_result, create);
129 up_read(&oi->ip_alloc_sem);
130
131 return ret;
132 }
133
134 int ocfs2_get_block(struct inode *inode, sector_t iblock,
135 struct buffer_head *bh_result, int create)
136 {
137 int err = 0;
138 unsigned int ext_flags;
139 u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
140 u64 p_blkno, count, past_eof;
141 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
142
143 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
144 (unsigned long long)iblock, bh_result, create);
145
146 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
147 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
148 inode, inode->i_ino);
149
150 if (S_ISLNK(inode->i_mode)) {
151 /* this always does I/O for some reason. */
152 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
153 goto bail;
154 }
155
156 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
157 &ext_flags);
158 if (err) {
159 mlog(ML_ERROR, "get_blocks() failed, inode: 0x%p, "
160 "block: %llu\n", inode, (unsigned long long)iblock);
161 goto bail;
162 }
163
164 if (max_blocks < count)
165 count = max_blocks;
166
167 /*
168 * ocfs2 never allocates in this function - the only time we
169 * need to use BH_New is when we're extending i_size on a file
170 * system which doesn't support holes, in which case BH_New
171 * allows __block_write_begin() to zero.
172 *
173 * If we see this on a sparse file system, then a truncate has
174 * raced us and removed the cluster. In this case, we clear
175 * the buffers dirty and uptodate bits and let the buffer code
176 * ignore it as a hole.
177 */
178 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
179 clear_buffer_dirty(bh_result);
180 clear_buffer_uptodate(bh_result);
181 goto bail;
182 }
183
184 /* Treat the unwritten extent as a hole for zeroing purposes. */
185 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
186 map_bh(bh_result, inode->i_sb, p_blkno);
187
188 bh_result->b_size = count << inode->i_blkbits;
189
190 if (!ocfs2_sparse_alloc(osb)) {
191 if (p_blkno == 0) {
192 err = -EIO;
193 mlog(ML_ERROR,
194 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
195 (unsigned long long)iblock,
196 (unsigned long long)p_blkno,
197 (unsigned long long)OCFS2_I(inode)->ip_blkno);
198 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
199 dump_stack();
200 goto bail;
201 }
202 }
203
204 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
205
206 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
207 (unsigned long long)past_eof);
208 if (create && (iblock >= past_eof))
209 set_buffer_new(bh_result);
210
211 bail:
212 if (err < 0)
213 err = -EIO;
214
215 return err;
216 }
217
218 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
219 struct buffer_head *di_bh)
220 {
221 void *kaddr;
222 loff_t size;
223 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
224
225 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
226 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
227 (unsigned long long)OCFS2_I(inode)->ip_blkno);
228 return -EROFS;
229 }
230
231 size = i_size_read(inode);
232
233 if (size > PAGE_SIZE ||
234 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
235 ocfs2_error(inode->i_sb,
236 "Inode %llu has with inline data has bad size: %Lu\n",
237 (unsigned long long)OCFS2_I(inode)->ip_blkno,
238 (unsigned long long)size);
239 return -EROFS;
240 }
241
242 kaddr = kmap_atomic(page);
243 if (size)
244 memcpy(kaddr, di->id2.i_data.id_data, size);
245 /* Clear the remaining part of the page */
246 memset(kaddr + size, 0, PAGE_SIZE - size);
247 flush_dcache_page(page);
248 kunmap_atomic(kaddr);
249
250 SetPageUptodate(page);
251
252 return 0;
253 }
254
255 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
256 {
257 int ret;
258 struct buffer_head *di_bh = NULL;
259
260 BUG_ON(!PageLocked(page));
261 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
262
263 ret = ocfs2_read_inode_block(inode, &di_bh);
264 if (ret) {
265 mlog_errno(ret);
266 goto out;
267 }
268
269 ret = ocfs2_read_inline_data(inode, page, di_bh);
270 out:
271 unlock_page(page);
272
273 brelse(di_bh);
274 return ret;
275 }
276
277 static int ocfs2_read_folio(struct file *file, struct folio *folio)
278 {
279 struct inode *inode = folio->mapping->host;
280 struct ocfs2_inode_info *oi = OCFS2_I(inode);
281 loff_t start = folio_pos(folio);
282 int ret, unlock = 1;
283
284 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, folio->index);
285
286 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, &folio->page);
287 if (ret != 0) {
288 if (ret == AOP_TRUNCATED_PAGE)
289 unlock = 0;
290 mlog_errno(ret);
291 goto out;
292 }
293
294 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
295 /*
296 * Unlock the folio and cycle ip_alloc_sem so that we don't
297 * busyloop waiting for ip_alloc_sem to unlock
298 */
299 ret = AOP_TRUNCATED_PAGE;
300 folio_unlock(folio);
301 unlock = 0;
302 down_read(&oi->ip_alloc_sem);
303 up_read(&oi->ip_alloc_sem);
304 goto out_inode_unlock;
305 }
306
307 /*
308 * i_size might have just been updated as we grabed the meta lock. We
309 * might now be discovering a truncate that hit on another node.
310 * block_read_full_folio->get_block freaks out if it is asked to read
311 * beyond the end of a file, so we check here. Callers
312 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
313 * and notice that the folio they just read isn't needed.
314 *
315 * XXX sys_readahead() seems to get that wrong?
316 */
317 if (start >= i_size_read(inode)) {
318 folio_zero_segment(folio, 0, folio_size(folio));
319 folio_mark_uptodate(folio);
320 ret = 0;
321 goto out_alloc;
322 }
323
324 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
325 ret = ocfs2_readpage_inline(inode, &folio->page);
326 else
327 ret = block_read_full_folio(folio, ocfs2_get_block);
328 unlock = 0;
329
330 out_alloc:
331 up_read(&oi->ip_alloc_sem);
332 out_inode_unlock:
333 ocfs2_inode_unlock(inode, 0);
334 out:
335 if (unlock)
336 folio_unlock(folio);
337 return ret;
338 }
339
340 /*
341 * This is used only for read-ahead. Failures or difficult to handle
342 * situations are safe to ignore.
343 *
344 * Right now, we don't bother with BH_Boundary - in-inode extent lists
345 * are quite large (243 extents on 4k blocks), so most inodes don't
346 * grow out to a tree. If need be, detecting boundary extents could
347 * trivially be added in a future version of ocfs2_get_block().
348 */
349 static void ocfs2_readahead(struct readahead_control *rac)
350 {
351 int ret;
352 struct inode *inode = rac->mapping->host;
353 struct ocfs2_inode_info *oi = OCFS2_I(inode);
354
355 /*
356 * Use the nonblocking flag for the dlm code to avoid page
357 * lock inversion, but don't bother with retrying.
358 */
359 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
360 if (ret)
361 return;
362
363 if (down_read_trylock(&oi->ip_alloc_sem) == 0)
364 goto out_unlock;
365
366 /*
367 * Don't bother with inline-data. There isn't anything
368 * to read-ahead in that case anyway...
369 */
370 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
371 goto out_up;
372
373 /*
374 * Check whether a remote node truncated this file - we just
375 * drop out in that case as it's not worth handling here.
376 */
377 if (readahead_pos(rac) >= i_size_read(inode))
378 goto out_up;
379
380 mpage_readahead(rac, ocfs2_get_block);
381
382 out_up:
383 up_read(&oi->ip_alloc_sem);
384 out_unlock:
385 ocfs2_inode_unlock(inode, 0);
386 }
387
388 /* Note: Because we don't support holes, our allocation has
389 * already happened (allocation writes zeros to the file data)
390 * so we don't have to worry about ordered writes in
391 * ocfs2_writepages.
392 *
393 * ->writepages is called during the process of invalidating the page cache
394 * during blocked lock processing. It can't block on any cluster locks
395 * to during block mapping. It's relying on the fact that the block
396 * mapping can't have disappeared under the dirty pages that it is
397 * being asked to write back.
398 */
399 static int ocfs2_writepages(struct address_space *mapping,
400 struct writeback_control *wbc)
401 {
402 return mpage_writepages(mapping, wbc, ocfs2_get_block);
403 }
404
405 /* Taken from ext3. We don't necessarily need the full blown
406 * functionality yet, but IMHO it's better to cut and paste the whole
407 * thing so we can avoid introducing our own bugs (and easily pick up
408 * their fixes when they happen) --Mark */
409 int walk_page_buffers( handle_t *handle,
410 struct buffer_head *head,
411 unsigned from,
412 unsigned to,
413 int *partial,
414 int (*fn)( handle_t *handle,
415 struct buffer_head *bh))
416 {
417 struct buffer_head *bh;
418 unsigned block_start, block_end;
419 unsigned blocksize = head->b_size;
420 int err, ret = 0;
421 struct buffer_head *next;
422
423 for ( bh = head, block_start = 0;
424 ret == 0 && (bh != head || !block_start);
425 block_start = block_end, bh = next)
426 {
427 next = bh->b_this_page;
428 block_end = block_start + blocksize;
429 if (block_end <= from || block_start >= to) {
430 if (partial && !buffer_uptodate(bh))
431 *partial = 1;
432 continue;
433 }
434 err = (*fn)(handle, bh);
435 if (!ret)
436 ret = err;
437 }
438 return ret;
439 }
440
441 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
442 {
443 sector_t status;
444 u64 p_blkno = 0;
445 int err = 0;
446 struct inode *inode = mapping->host;
447
448 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
449 (unsigned long long)block);
450
451 /*
452 * The swap code (ab-)uses ->bmap to get a block mapping and then
453 * bypasseѕ the file system for actual I/O. We really can't allow
454 * that on refcounted inodes, so we have to skip out here. And yes,
455 * 0 is the magic code for a bmap error..
456 */
457 if (ocfs2_is_refcount_inode(inode))
458 return 0;
459
460 /* We don't need to lock journal system files, since they aren't
461 * accessed concurrently from multiple nodes.
462 */
463 if (!INODE_JOURNAL(inode)) {
464 err = ocfs2_inode_lock(inode, NULL, 0);
465 if (err) {
466 if (err != -ENOENT)
467 mlog_errno(err);
468 goto bail;
469 }
470 down_read(&OCFS2_I(inode)->ip_alloc_sem);
471 }
472
473 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
474 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
475 NULL);
476
477 if (!INODE_JOURNAL(inode)) {
478 up_read(&OCFS2_I(inode)->ip_alloc_sem);
479 ocfs2_inode_unlock(inode, 0);
480 }
481
482 if (err) {
483 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
484 (unsigned long long)block);
485 mlog_errno(err);
486 goto bail;
487 }
488
489 bail:
490 status = err ? 0 : p_blkno;
491
492 return status;
493 }
494
495 static bool ocfs2_release_folio(struct folio *folio, gfp_t wait)
496 {
497 if (!folio_buffers(folio))
498 return false;
499 return try_to_free_buffers(folio);
500 }
501
502 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
503 u32 cpos,
504 unsigned int *start,
505 unsigned int *end)
506 {
507 unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
508
509 if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
510 unsigned int cpp;
511
512 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
513
514 cluster_start = cpos % cpp;
515 cluster_start = cluster_start << osb->s_clustersize_bits;
516
517 cluster_end = cluster_start + osb->s_clustersize;
518 }
519
520 BUG_ON(cluster_start > PAGE_SIZE);
521 BUG_ON(cluster_end > PAGE_SIZE);
522
523 if (start)
524 *start = cluster_start;
525 if (end)
526 *end = cluster_end;
527 }
528
529 /*
530 * 'from' and 'to' are the region in the page to avoid zeroing.
531 *
532 * If pagesize > clustersize, this function will avoid zeroing outside
533 * of the cluster boundary.
534 *
535 * from == to == 0 is code for "zero the entire cluster region"
536 */
537 static void ocfs2_clear_page_regions(struct page *page,
538 struct ocfs2_super *osb, u32 cpos,
539 unsigned from, unsigned to)
540 {
541 void *kaddr;
542 unsigned int cluster_start, cluster_end;
543
544 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
545
546 kaddr = kmap_atomic(page);
547
548 if (from || to) {
549 if (from > cluster_start)
550 memset(kaddr + cluster_start, 0, from - cluster_start);
551 if (to < cluster_end)
552 memset(kaddr + to, 0, cluster_end - to);
553 } else {
554 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
555 }
556
557 kunmap_atomic(kaddr);
558 }
559
560 /*
561 * Nonsparse file systems fully allocate before we get to the write
562 * code. This prevents ocfs2_write() from tagging the write as an
563 * allocating one, which means ocfs2_map_page_blocks() might try to
564 * read-in the blocks at the tail of our file. Avoid reading them by
565 * testing i_size against each block offset.
566 */
567 static int ocfs2_should_read_blk(struct inode *inode, struct folio *folio,
568 unsigned int block_start)
569 {
570 u64 offset = folio_pos(folio) + block_start;
571
572 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
573 return 1;
574
575 if (i_size_read(inode) > offset)
576 return 1;
577
578 return 0;
579 }
580
581 /*
582 * Some of this taken from __block_write_begin(). We already have our
583 * mapping by now though, and the entire write will be allocating or
584 * it won't, so not much need to use BH_New.
585 *
586 * This will also skip zeroing, which is handled externally.
587 */
588 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
589 struct inode *inode, unsigned int from,
590 unsigned int to, int new)
591 {
592 struct folio *folio = page_folio(page);
593 int ret = 0;
594 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
595 unsigned int block_end, block_start;
596 unsigned int bsize = i_blocksize(inode);
597
598 head = folio_buffers(folio);
599 if (!head)
600 head = create_empty_buffers(folio, bsize, 0);
601
602 for (bh = head, block_start = 0; bh != head || !block_start;
603 bh = bh->b_this_page, block_start += bsize) {
604 block_end = block_start + bsize;
605
606 clear_buffer_new(bh);
607
608 /*
609 * Ignore blocks outside of our i/o range -
610 * they may belong to unallocated clusters.
611 */
612 if (block_start >= to || block_end <= from) {
613 if (folio_test_uptodate(folio))
614 set_buffer_uptodate(bh);
615 continue;
616 }
617
618 /*
619 * For an allocating write with cluster size >= page
620 * size, we always write the entire page.
621 */
622 if (new)
623 set_buffer_new(bh);
624
625 if (!buffer_mapped(bh)) {
626 map_bh(bh, inode->i_sb, *p_blkno);
627 clean_bdev_bh_alias(bh);
628 }
629
630 if (folio_test_uptodate(folio)) {
631 set_buffer_uptodate(bh);
632 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
633 !buffer_new(bh) &&
634 ocfs2_should_read_blk(inode, folio, block_start) &&
635 (block_start < from || block_end > to)) {
636 bh_read_nowait(bh, 0);
637 *wait_bh++=bh;
638 }
639
640 *p_blkno = *p_blkno + 1;
641 }
642
643 /*
644 * If we issued read requests - let them complete.
645 */
646 while(wait_bh > wait) {
647 wait_on_buffer(*--wait_bh);
648 if (!buffer_uptodate(*wait_bh))
649 ret = -EIO;
650 }
651
652 if (ret == 0 || !new)
653 return ret;
654
655 /*
656 * If we get -EIO above, zero out any newly allocated blocks
657 * to avoid exposing stale data.
658 */
659 bh = head;
660 block_start = 0;
661 do {
662 block_end = block_start + bsize;
663 if (block_end <= from)
664 goto next_bh;
665 if (block_start >= to)
666 break;
667
668 folio_zero_range(folio, block_start, bh->b_size);
669 set_buffer_uptodate(bh);
670 mark_buffer_dirty(bh);
671
672 next_bh:
673 block_start = block_end;
674 bh = bh->b_this_page;
675 } while (bh != head);
676
677 return ret;
678 }
679
680 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
681 #define OCFS2_MAX_CTXT_PAGES 1
682 #else
683 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
684 #endif
685
686 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
687
688 struct ocfs2_unwritten_extent {
689 struct list_head ue_node;
690 struct list_head ue_ip_node;
691 u32 ue_cpos;
692 u32 ue_phys;
693 };
694
695 /*
696 * Describe the state of a single cluster to be written to.
697 */
698 struct ocfs2_write_cluster_desc {
699 u32 c_cpos;
700 u32 c_phys;
701 /*
702 * Give this a unique field because c_phys eventually gets
703 * filled.
704 */
705 unsigned c_new;
706 unsigned c_clear_unwritten;
707 unsigned c_needs_zero;
708 };
709
710 struct ocfs2_write_ctxt {
711 /* Logical cluster position / len of write */
712 u32 w_cpos;
713 u32 w_clen;
714
715 /* First cluster allocated in a nonsparse extend */
716 u32 w_first_new_cpos;
717
718 /* Type of caller. Must be one of buffer, mmap, direct. */
719 ocfs2_write_type_t w_type;
720
721 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
722
723 /*
724 * This is true if page_size > cluster_size.
725 *
726 * It triggers a set of special cases during write which might
727 * have to deal with allocating writes to partial pages.
728 */
729 unsigned int w_large_pages;
730
731 /*
732 * Pages involved in this write.
733 *
734 * w_target_page is the page being written to by the user.
735 *
736 * w_pages is an array of pages which always contains
737 * w_target_page, and in the case of an allocating write with
738 * page_size < cluster size, it will contain zero'd and mapped
739 * pages adjacent to w_target_page which need to be written
740 * out in so that future reads from that region will get
741 * zero's.
742 */
743 unsigned int w_num_pages;
744 struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
745 struct page *w_target_page;
746
747 /*
748 * w_target_locked is used for page_mkwrite path indicating no unlocking
749 * against w_target_page in ocfs2_write_end_nolock.
750 */
751 unsigned int w_target_locked:1;
752
753 /*
754 * ocfs2_write_end() uses this to know what the real range to
755 * write in the target should be.
756 */
757 unsigned int w_target_from;
758 unsigned int w_target_to;
759
760 /*
761 * We could use journal_current_handle() but this is cleaner,
762 * IMHO -Mark
763 */
764 handle_t *w_handle;
765
766 struct buffer_head *w_di_bh;
767
768 struct ocfs2_cached_dealloc_ctxt w_dealloc;
769
770 struct list_head w_unwritten_list;
771 unsigned int w_unwritten_count;
772 };
773
774 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
775 {
776 int i;
777
778 for(i = 0; i < num_pages; i++) {
779 if (pages[i]) {
780 unlock_page(pages[i]);
781 mark_page_accessed(pages[i]);
782 put_page(pages[i]);
783 }
784 }
785 }
786
787 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
788 {
789 int i;
790
791 /*
792 * w_target_locked is only set to true in the page_mkwrite() case.
793 * The intent is to allow us to lock the target page from write_begin()
794 * to write_end(). The caller must hold a ref on w_target_page.
795 */
796 if (wc->w_target_locked) {
797 BUG_ON(!wc->w_target_page);
798 for (i = 0; i < wc->w_num_pages; i++) {
799 if (wc->w_target_page == wc->w_pages[i]) {
800 wc->w_pages[i] = NULL;
801 break;
802 }
803 }
804 mark_page_accessed(wc->w_target_page);
805 put_page(wc->w_target_page);
806 }
807 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
808 }
809
810 static void ocfs2_free_unwritten_list(struct inode *inode,
811 struct list_head *head)
812 {
813 struct ocfs2_inode_info *oi = OCFS2_I(inode);
814 struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
815
816 list_for_each_entry_safe(ue, tmp, head, ue_node) {
817 list_del(&ue->ue_node);
818 spin_lock(&oi->ip_lock);
819 list_del(&ue->ue_ip_node);
820 spin_unlock(&oi->ip_lock);
821 kfree(ue);
822 }
823 }
824
825 static void ocfs2_free_write_ctxt(struct inode *inode,
826 struct ocfs2_write_ctxt *wc)
827 {
828 ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
829 ocfs2_unlock_pages(wc);
830 brelse(wc->w_di_bh);
831 kfree(wc);
832 }
833
834 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
835 struct ocfs2_super *osb, loff_t pos,
836 unsigned len, ocfs2_write_type_t type,
837 struct buffer_head *di_bh)
838 {
839 u32 cend;
840 struct ocfs2_write_ctxt *wc;
841
842 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
843 if (!wc)
844 return -ENOMEM;
845
846 wc->w_cpos = pos >> osb->s_clustersize_bits;
847 wc->w_first_new_cpos = UINT_MAX;
848 cend = (pos + len - 1) >> osb->s_clustersize_bits;
849 wc->w_clen = cend - wc->w_cpos + 1;
850 get_bh(di_bh);
851 wc->w_di_bh = di_bh;
852 wc->w_type = type;
853
854 if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
855 wc->w_large_pages = 1;
856 else
857 wc->w_large_pages = 0;
858
859 ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
860 INIT_LIST_HEAD(&wc->w_unwritten_list);
861
862 *wcp = wc;
863
864 return 0;
865 }
866
867 /*
868 * If a page has any new buffers, zero them out here, and mark them uptodate
869 * and dirty so they'll be written out (in order to prevent uninitialised
870 * block data from leaking). And clear the new bit.
871 */
872 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
873 {
874 unsigned int block_start, block_end;
875 struct buffer_head *head, *bh;
876
877 BUG_ON(!PageLocked(page));
878 if (!page_has_buffers(page))
879 return;
880
881 bh = head = page_buffers(page);
882 block_start = 0;
883 do {
884 block_end = block_start + bh->b_size;
885
886 if (buffer_new(bh)) {
887 if (block_end > from && block_start < to) {
888 if (!PageUptodate(page)) {
889 unsigned start, end;
890
891 start = max(from, block_start);
892 end = min(to, block_end);
893
894 zero_user_segment(page, start, end);
895 set_buffer_uptodate(bh);
896 }
897
898 clear_buffer_new(bh);
899 mark_buffer_dirty(bh);
900 }
901 }
902
903 block_start = block_end;
904 bh = bh->b_this_page;
905 } while (bh != head);
906 }
907
908 /*
909 * Only called when we have a failure during allocating write to write
910 * zero's to the newly allocated region.
911 */
912 static void ocfs2_write_failure(struct inode *inode,
913 struct ocfs2_write_ctxt *wc,
914 loff_t user_pos, unsigned user_len)
915 {
916 int i;
917 unsigned from = user_pos & (PAGE_SIZE - 1),
918 to = user_pos + user_len;
919 struct page *tmppage;
920
921 if (wc->w_target_page)
922 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
923
924 for(i = 0; i < wc->w_num_pages; i++) {
925 tmppage = wc->w_pages[i];
926
927 if (tmppage && page_has_buffers(tmppage)) {
928 if (ocfs2_should_order_data(inode))
929 ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
930 user_pos, user_len);
931
932 block_commit_write(tmppage, from, to);
933 }
934 }
935 }
936
937 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
938 struct ocfs2_write_ctxt *wc,
939 struct page *page, u32 cpos,
940 loff_t user_pos, unsigned user_len,
941 int new)
942 {
943 int ret;
944 unsigned int map_from = 0, map_to = 0;
945 unsigned int cluster_start, cluster_end;
946 unsigned int user_data_from = 0, user_data_to = 0;
947
948 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
949 &cluster_start, &cluster_end);
950
951 /* treat the write as new if the a hole/lseek spanned across
952 * the page boundary.
953 */
954 new = new | ((i_size_read(inode) <= page_offset(page)) &&
955 (page_offset(page) <= user_pos));
956
957 if (page == wc->w_target_page) {
958 map_from = user_pos & (PAGE_SIZE - 1);
959 map_to = map_from + user_len;
960
961 if (new)
962 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
963 cluster_start, cluster_end,
964 new);
965 else
966 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
967 map_from, map_to, new);
968 if (ret) {
969 mlog_errno(ret);
970 goto out;
971 }
972
973 user_data_from = map_from;
974 user_data_to = map_to;
975 if (new) {
976 map_from = cluster_start;
977 map_to = cluster_end;
978 }
979 } else {
980 /*
981 * If we haven't allocated the new page yet, we
982 * shouldn't be writing it out without copying user
983 * data. This is likely a math error from the caller.
984 */
985 BUG_ON(!new);
986
987 map_from = cluster_start;
988 map_to = cluster_end;
989
990 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
991 cluster_start, cluster_end, new);
992 if (ret) {
993 mlog_errno(ret);
994 goto out;
995 }
996 }
997
998 /*
999 * Parts of newly allocated pages need to be zero'd.
1000 *
1001 * Above, we have also rewritten 'to' and 'from' - as far as
1002 * the rest of the function is concerned, the entire cluster
1003 * range inside of a page needs to be written.
1004 *
1005 * We can skip this if the page is up to date - it's already
1006 * been zero'd from being read in as a hole.
1007 */
1008 if (new && !PageUptodate(page))
1009 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1010 cpos, user_data_from, user_data_to);
1011
1012 flush_dcache_page(page);
1013
1014 out:
1015 return ret;
1016 }
1017
1018 /*
1019 * This function will only grab one clusters worth of pages.
1020 */
1021 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1022 struct ocfs2_write_ctxt *wc,
1023 u32 cpos, loff_t user_pos,
1024 unsigned user_len, int new,
1025 struct page *mmap_page)
1026 {
1027 int ret = 0, i;
1028 unsigned long start, target_index, end_index, index;
1029 struct inode *inode = mapping->host;
1030 loff_t last_byte;
1031
1032 target_index = user_pos >> PAGE_SHIFT;
1033
1034 /*
1035 * Figure out how many pages we'll be manipulating here. For
1036 * non allocating write, we just change the one
1037 * page. Otherwise, we'll need a whole clusters worth. If we're
1038 * writing past i_size, we only need enough pages to cover the
1039 * last page of the write.
1040 */
1041 if (new) {
1042 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1043 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1044 /*
1045 * We need the index *past* the last page we could possibly
1046 * touch. This is the page past the end of the write or
1047 * i_size, whichever is greater.
1048 */
1049 last_byte = max(user_pos + user_len, i_size_read(inode));
1050 BUG_ON(last_byte < 1);
1051 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1052 if ((start + wc->w_num_pages) > end_index)
1053 wc->w_num_pages = end_index - start;
1054 } else {
1055 wc->w_num_pages = 1;
1056 start = target_index;
1057 }
1058 end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1059
1060 for(i = 0; i < wc->w_num_pages; i++) {
1061 index = start + i;
1062
1063 if (index >= target_index && index <= end_index &&
1064 wc->w_type == OCFS2_WRITE_MMAP) {
1065 /*
1066 * ocfs2_pagemkwrite() is a little different
1067 * and wants us to directly use the page
1068 * passed in.
1069 */
1070 lock_page(mmap_page);
1071
1072 /* Exit and let the caller retry */
1073 if (mmap_page->mapping != mapping) {
1074 WARN_ON(mmap_page->mapping);
1075 unlock_page(mmap_page);
1076 ret = -EAGAIN;
1077 goto out;
1078 }
1079
1080 get_page(mmap_page);
1081 wc->w_pages[i] = mmap_page;
1082 wc->w_target_locked = true;
1083 } else if (index >= target_index && index <= end_index &&
1084 wc->w_type == OCFS2_WRITE_DIRECT) {
1085 /* Direct write has no mapping page. */
1086 wc->w_pages[i] = NULL;
1087 continue;
1088 } else {
1089 wc->w_pages[i] = find_or_create_page(mapping, index,
1090 GFP_NOFS);
1091 if (!wc->w_pages[i]) {
1092 ret = -ENOMEM;
1093 mlog_errno(ret);
1094 goto out;
1095 }
1096 }
1097 wait_for_stable_page(wc->w_pages[i]);
1098
1099 if (index == target_index)
1100 wc->w_target_page = wc->w_pages[i];
1101 }
1102 out:
1103 if (ret)
1104 wc->w_target_locked = false;
1105 return ret;
1106 }
1107
1108 /*
1109 * Prepare a single cluster for write one cluster into the file.
1110 */
1111 static int ocfs2_write_cluster(struct address_space *mapping,
1112 u32 *phys, unsigned int new,
1113 unsigned int clear_unwritten,
1114 unsigned int should_zero,
1115 struct ocfs2_alloc_context *data_ac,
1116 struct ocfs2_alloc_context *meta_ac,
1117 struct ocfs2_write_ctxt *wc, u32 cpos,
1118 loff_t user_pos, unsigned user_len)
1119 {
1120 int ret, i;
1121 u64 p_blkno;
1122 struct inode *inode = mapping->host;
1123 struct ocfs2_extent_tree et;
1124 int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1125
1126 if (new) {
1127 u32 tmp_pos;
1128
1129 /*
1130 * This is safe to call with the page locks - it won't take
1131 * any additional semaphores or cluster locks.
1132 */
1133 tmp_pos = cpos;
1134 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1135 &tmp_pos, 1, !clear_unwritten,
1136 wc->w_di_bh, wc->w_handle,
1137 data_ac, meta_ac, NULL);
1138 /*
1139 * This shouldn't happen because we must have already
1140 * calculated the correct meta data allocation required. The
1141 * internal tree allocation code should know how to increase
1142 * transaction credits itself.
1143 *
1144 * If need be, we could handle -EAGAIN for a
1145 * RESTART_TRANS here.
1146 */
1147 mlog_bug_on_msg(ret == -EAGAIN,
1148 "Inode %llu: EAGAIN return during allocation.\n",
1149 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1150 if (ret < 0) {
1151 mlog_errno(ret);
1152 goto out;
1153 }
1154 } else if (clear_unwritten) {
1155 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1156 wc->w_di_bh);
1157 ret = ocfs2_mark_extent_written(inode, &et,
1158 wc->w_handle, cpos, 1, *phys,
1159 meta_ac, &wc->w_dealloc);
1160 if (ret < 0) {
1161 mlog_errno(ret);
1162 goto out;
1163 }
1164 }
1165
1166 /*
1167 * The only reason this should fail is due to an inability to
1168 * find the extent added.
1169 */
1170 ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1171 if (ret < 0) {
1172 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1173 "at logical cluster %u",
1174 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1175 goto out;
1176 }
1177
1178 BUG_ON(*phys == 0);
1179
1180 p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1181 if (!should_zero)
1182 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1183
1184 for(i = 0; i < wc->w_num_pages; i++) {
1185 int tmpret;
1186
1187 /* This is the direct io target page. */
1188 if (wc->w_pages[i] == NULL) {
1189 p_blkno++;
1190 continue;
1191 }
1192
1193 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1194 wc->w_pages[i], cpos,
1195 user_pos, user_len,
1196 should_zero);
1197 if (tmpret) {
1198 mlog_errno(tmpret);
1199 if (ret == 0)
1200 ret = tmpret;
1201 }
1202 }
1203
1204 /*
1205 * We only have cleanup to do in case of allocating write.
1206 */
1207 if (ret && new)
1208 ocfs2_write_failure(inode, wc, user_pos, user_len);
1209
1210 out:
1211
1212 return ret;
1213 }
1214
1215 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1216 struct ocfs2_alloc_context *data_ac,
1217 struct ocfs2_alloc_context *meta_ac,
1218 struct ocfs2_write_ctxt *wc,
1219 loff_t pos, unsigned len)
1220 {
1221 int ret, i;
1222 loff_t cluster_off;
1223 unsigned int local_len = len;
1224 struct ocfs2_write_cluster_desc *desc;
1225 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1226
1227 for (i = 0; i < wc->w_clen; i++) {
1228 desc = &wc->w_desc[i];
1229
1230 /*
1231 * We have to make sure that the total write passed in
1232 * doesn't extend past a single cluster.
1233 */
1234 local_len = len;
1235 cluster_off = pos & (osb->s_clustersize - 1);
1236 if ((cluster_off + local_len) > osb->s_clustersize)
1237 local_len = osb->s_clustersize - cluster_off;
1238
1239 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1240 desc->c_new,
1241 desc->c_clear_unwritten,
1242 desc->c_needs_zero,
1243 data_ac, meta_ac,
1244 wc, desc->c_cpos, pos, local_len);
1245 if (ret) {
1246 mlog_errno(ret);
1247 goto out;
1248 }
1249
1250 len -= local_len;
1251 pos += local_len;
1252 }
1253
1254 ret = 0;
1255 out:
1256 return ret;
1257 }
1258
1259 /*
1260 * ocfs2_write_end() wants to know which parts of the target page it
1261 * should complete the write on. It's easiest to compute them ahead of
1262 * time when a more complete view of the write is available.
1263 */
1264 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1265 struct ocfs2_write_ctxt *wc,
1266 loff_t pos, unsigned len, int alloc)
1267 {
1268 struct ocfs2_write_cluster_desc *desc;
1269
1270 wc->w_target_from = pos & (PAGE_SIZE - 1);
1271 wc->w_target_to = wc->w_target_from + len;
1272
1273 if (alloc == 0)
1274 return;
1275
1276 /*
1277 * Allocating write - we may have different boundaries based
1278 * on page size and cluster size.
1279 *
1280 * NOTE: We can no longer compute one value from the other as
1281 * the actual write length and user provided length may be
1282 * different.
1283 */
1284
1285 if (wc->w_large_pages) {
1286 /*
1287 * We only care about the 1st and last cluster within
1288 * our range and whether they should be zero'd or not. Either
1289 * value may be extended out to the start/end of a
1290 * newly allocated cluster.
1291 */
1292 desc = &wc->w_desc[0];
1293 if (desc->c_needs_zero)
1294 ocfs2_figure_cluster_boundaries(osb,
1295 desc->c_cpos,
1296 &wc->w_target_from,
1297 NULL);
1298
1299 desc = &wc->w_desc[wc->w_clen - 1];
1300 if (desc->c_needs_zero)
1301 ocfs2_figure_cluster_boundaries(osb,
1302 desc->c_cpos,
1303 NULL,
1304 &wc->w_target_to);
1305 } else {
1306 wc->w_target_from = 0;
1307 wc->w_target_to = PAGE_SIZE;
1308 }
1309 }
1310
1311 /*
1312 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1313 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1314 * by the direct io procedure.
1315 * If this is a new extent that allocated by direct io, we should mark it in
1316 * the ip_unwritten_list.
1317 */
1318 static int ocfs2_unwritten_check(struct inode *inode,
1319 struct ocfs2_write_ctxt *wc,
1320 struct ocfs2_write_cluster_desc *desc)
1321 {
1322 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1323 struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1324 int ret = 0;
1325
1326 if (!desc->c_needs_zero)
1327 return 0;
1328
1329 retry:
1330 spin_lock(&oi->ip_lock);
1331 /* Needs not to zero no metter buffer or direct. The one who is zero
1332 * the cluster is doing zero. And he will clear unwritten after all
1333 * cluster io finished. */
1334 list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1335 if (desc->c_cpos == ue->ue_cpos) {
1336 BUG_ON(desc->c_new);
1337 desc->c_needs_zero = 0;
1338 desc->c_clear_unwritten = 0;
1339 goto unlock;
1340 }
1341 }
1342
1343 if (wc->w_type != OCFS2_WRITE_DIRECT)
1344 goto unlock;
1345
1346 if (new == NULL) {
1347 spin_unlock(&oi->ip_lock);
1348 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1349 GFP_NOFS);
1350 if (new == NULL) {
1351 ret = -ENOMEM;
1352 goto out;
1353 }
1354 goto retry;
1355 }
1356 /* This direct write will doing zero. */
1357 new->ue_cpos = desc->c_cpos;
1358 new->ue_phys = desc->c_phys;
1359 desc->c_clear_unwritten = 0;
1360 list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1361 list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1362 wc->w_unwritten_count++;
1363 new = NULL;
1364 unlock:
1365 spin_unlock(&oi->ip_lock);
1366 out:
1367 kfree(new);
1368 return ret;
1369 }
1370
1371 /*
1372 * Populate each single-cluster write descriptor in the write context
1373 * with information about the i/o to be done.
1374 *
1375 * Returns the number of clusters that will have to be allocated, as
1376 * well as a worst case estimate of the number of extent records that
1377 * would have to be created during a write to an unwritten region.
1378 */
1379 static int ocfs2_populate_write_desc(struct inode *inode,
1380 struct ocfs2_write_ctxt *wc,
1381 unsigned int *clusters_to_alloc,
1382 unsigned int *extents_to_split)
1383 {
1384 int ret;
1385 struct ocfs2_write_cluster_desc *desc;
1386 unsigned int num_clusters = 0;
1387 unsigned int ext_flags = 0;
1388 u32 phys = 0;
1389 int i;
1390
1391 *clusters_to_alloc = 0;
1392 *extents_to_split = 0;
1393
1394 for (i = 0; i < wc->w_clen; i++) {
1395 desc = &wc->w_desc[i];
1396 desc->c_cpos = wc->w_cpos + i;
1397
1398 if (num_clusters == 0) {
1399 /*
1400 * Need to look up the next extent record.
1401 */
1402 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1403 &num_clusters, &ext_flags);
1404 if (ret) {
1405 mlog_errno(ret);
1406 goto out;
1407 }
1408
1409 /* We should already CoW the refcountd extent. */
1410 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1411
1412 /*
1413 * Assume worst case - that we're writing in
1414 * the middle of the extent.
1415 *
1416 * We can assume that the write proceeds from
1417 * left to right, in which case the extent
1418 * insert code is smart enough to coalesce the
1419 * next splits into the previous records created.
1420 */
1421 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1422 *extents_to_split = *extents_to_split + 2;
1423 } else if (phys) {
1424 /*
1425 * Only increment phys if it doesn't describe
1426 * a hole.
1427 */
1428 phys++;
1429 }
1430
1431 /*
1432 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1433 * file that got extended. w_first_new_cpos tells us
1434 * where the newly allocated clusters are so we can
1435 * zero them.
1436 */
1437 if (desc->c_cpos >= wc->w_first_new_cpos) {
1438 BUG_ON(phys == 0);
1439 desc->c_needs_zero = 1;
1440 }
1441
1442 desc->c_phys = phys;
1443 if (phys == 0) {
1444 desc->c_new = 1;
1445 desc->c_needs_zero = 1;
1446 desc->c_clear_unwritten = 1;
1447 *clusters_to_alloc = *clusters_to_alloc + 1;
1448 }
1449
1450 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1451 desc->c_clear_unwritten = 1;
1452 desc->c_needs_zero = 1;
1453 }
1454
1455 ret = ocfs2_unwritten_check(inode, wc, desc);
1456 if (ret) {
1457 mlog_errno(ret);
1458 goto out;
1459 }
1460
1461 num_clusters--;
1462 }
1463
1464 ret = 0;
1465 out:
1466 return ret;
1467 }
1468
1469 static int ocfs2_write_begin_inline(struct address_space *mapping,
1470 struct inode *inode,
1471 struct ocfs2_write_ctxt *wc)
1472 {
1473 int ret;
1474 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1475 struct page *page;
1476 handle_t *handle;
1477 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1478
1479 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1480 if (IS_ERR(handle)) {
1481 ret = PTR_ERR(handle);
1482 mlog_errno(ret);
1483 goto out;
1484 }
1485
1486 page = find_or_create_page(mapping, 0, GFP_NOFS);
1487 if (!page) {
1488 ocfs2_commit_trans(osb, handle);
1489 ret = -ENOMEM;
1490 mlog_errno(ret);
1491 goto out;
1492 }
1493 /*
1494 * If we don't set w_num_pages then this page won't get unlocked
1495 * and freed on cleanup of the write context.
1496 */
1497 wc->w_pages[0] = wc->w_target_page = page;
1498 wc->w_num_pages = 1;
1499
1500 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1501 OCFS2_JOURNAL_ACCESS_WRITE);
1502 if (ret) {
1503 ocfs2_commit_trans(osb, handle);
1504
1505 mlog_errno(ret);
1506 goto out;
1507 }
1508
1509 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1510 ocfs2_set_inode_data_inline(inode, di);
1511
1512 if (!PageUptodate(page)) {
1513 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1514 if (ret) {
1515 ocfs2_commit_trans(osb, handle);
1516
1517 goto out;
1518 }
1519 }
1520
1521 wc->w_handle = handle;
1522 out:
1523 return ret;
1524 }
1525
1526 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1527 {
1528 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1529
1530 if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1531 return 1;
1532 return 0;
1533 }
1534
1535 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1536 struct inode *inode, loff_t pos,
1537 unsigned len, struct page *mmap_page,
1538 struct ocfs2_write_ctxt *wc)
1539 {
1540 int ret, written = 0;
1541 loff_t end = pos + len;
1542 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1543 struct ocfs2_dinode *di = NULL;
1544
1545 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1546 len, (unsigned long long)pos,
1547 oi->ip_dyn_features);
1548
1549 /*
1550 * Handle inodes which already have inline data 1st.
1551 */
1552 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1553 if (mmap_page == NULL &&
1554 ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1555 goto do_inline_write;
1556
1557 /*
1558 * The write won't fit - we have to give this inode an
1559 * inline extent list now.
1560 */
1561 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1562 if (ret)
1563 mlog_errno(ret);
1564 goto out;
1565 }
1566
1567 /*
1568 * Check whether the inode can accept inline data.
1569 */
1570 if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1571 return 0;
1572
1573 /*
1574 * Check whether the write can fit.
1575 */
1576 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1577 if (mmap_page ||
1578 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1579 return 0;
1580
1581 do_inline_write:
1582 ret = ocfs2_write_begin_inline(mapping, inode, wc);
1583 if (ret) {
1584 mlog_errno(ret);
1585 goto out;
1586 }
1587
1588 /*
1589 * This signals to the caller that the data can be written
1590 * inline.
1591 */
1592 written = 1;
1593 out:
1594 return written ? written : ret;
1595 }
1596
1597 /*
1598 * This function only does anything for file systems which can't
1599 * handle sparse files.
1600 *
1601 * What we want to do here is fill in any hole between the current end
1602 * of allocation and the end of our write. That way the rest of the
1603 * write path can treat it as an non-allocating write, which has no
1604 * special case code for sparse/nonsparse files.
1605 */
1606 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1607 struct buffer_head *di_bh,
1608 loff_t pos, unsigned len,
1609 struct ocfs2_write_ctxt *wc)
1610 {
1611 int ret;
1612 loff_t newsize = pos + len;
1613
1614 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1615
1616 if (newsize <= i_size_read(inode))
1617 return 0;
1618
1619 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1620 if (ret)
1621 mlog_errno(ret);
1622
1623 /* There is no wc if this is call from direct. */
1624 if (wc)
1625 wc->w_first_new_cpos =
1626 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1627
1628 return ret;
1629 }
1630
1631 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1632 loff_t pos)
1633 {
1634 int ret = 0;
1635
1636 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1637 if (pos > i_size_read(inode))
1638 ret = ocfs2_zero_extend(inode, di_bh, pos);
1639
1640 return ret;
1641 }
1642
1643 int ocfs2_write_begin_nolock(struct address_space *mapping,
1644 loff_t pos, unsigned len, ocfs2_write_type_t type,
1645 struct page **pagep, void **fsdata,
1646 struct buffer_head *di_bh, struct page *mmap_page)
1647 {
1648 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1649 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1650 struct ocfs2_write_ctxt *wc;
1651 struct inode *inode = mapping->host;
1652 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1653 struct ocfs2_dinode *di;
1654 struct ocfs2_alloc_context *data_ac = NULL;
1655 struct ocfs2_alloc_context *meta_ac = NULL;
1656 handle_t *handle;
1657 struct ocfs2_extent_tree et;
1658 int try_free = 1, ret1;
1659
1660 try_again:
1661 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1662 if (ret) {
1663 mlog_errno(ret);
1664 return ret;
1665 }
1666
1667 if (ocfs2_supports_inline_data(osb)) {
1668 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1669 mmap_page, wc);
1670 if (ret == 1) {
1671 ret = 0;
1672 goto success;
1673 }
1674 if (ret < 0) {
1675 mlog_errno(ret);
1676 goto out;
1677 }
1678 }
1679
1680 /* Direct io change i_size late, should not zero tail here. */
1681 if (type != OCFS2_WRITE_DIRECT) {
1682 if (ocfs2_sparse_alloc(osb))
1683 ret = ocfs2_zero_tail(inode, di_bh, pos);
1684 else
1685 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1686 len, wc);
1687 if (ret) {
1688 mlog_errno(ret);
1689 goto out;
1690 }
1691 }
1692
1693 ret = ocfs2_check_range_for_refcount(inode, pos, len);
1694 if (ret < 0) {
1695 mlog_errno(ret);
1696 goto out;
1697 } else if (ret == 1) {
1698 clusters_need = wc->w_clen;
1699 ret = ocfs2_refcount_cow(inode, di_bh,
1700 wc->w_cpos, wc->w_clen, UINT_MAX);
1701 if (ret) {
1702 mlog_errno(ret);
1703 goto out;
1704 }
1705 }
1706
1707 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1708 &extents_to_split);
1709 if (ret) {
1710 mlog_errno(ret);
1711 goto out;
1712 }
1713 clusters_need += clusters_to_alloc;
1714
1715 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1716
1717 trace_ocfs2_write_begin_nolock(
1718 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1719 (long long)i_size_read(inode),
1720 le32_to_cpu(di->i_clusters),
1721 pos, len, type, mmap_page,
1722 clusters_to_alloc, extents_to_split);
1723
1724 /*
1725 * We set w_target_from, w_target_to here so that
1726 * ocfs2_write_end() knows which range in the target page to
1727 * write out. An allocation requires that we write the entire
1728 * cluster range.
1729 */
1730 if (clusters_to_alloc || extents_to_split) {
1731 /*
1732 * XXX: We are stretching the limits of
1733 * ocfs2_lock_allocators(). It greatly over-estimates
1734 * the work to be done.
1735 */
1736 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1737 wc->w_di_bh);
1738 ret = ocfs2_lock_allocators(inode, &et,
1739 clusters_to_alloc, extents_to_split,
1740 &data_ac, &meta_ac);
1741 if (ret) {
1742 mlog_errno(ret);
1743 goto out;
1744 }
1745
1746 if (data_ac)
1747 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1748
1749 credits = ocfs2_calc_extend_credits(inode->i_sb,
1750 &di->id2.i_list);
1751 } else if (type == OCFS2_WRITE_DIRECT)
1752 /* direct write needs not to start trans if no extents alloc. */
1753 goto success;
1754
1755 /*
1756 * We have to zero sparse allocated clusters, unwritten extent clusters,
1757 * and non-sparse clusters we just extended. For non-sparse writes,
1758 * we know zeros will only be needed in the first and/or last cluster.
1759 */
1760 if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1761 wc->w_desc[wc->w_clen - 1].c_needs_zero))
1762 cluster_of_pages = 1;
1763 else
1764 cluster_of_pages = 0;
1765
1766 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1767
1768 handle = ocfs2_start_trans(osb, credits);
1769 if (IS_ERR(handle)) {
1770 ret = PTR_ERR(handle);
1771 mlog_errno(ret);
1772 goto out;
1773 }
1774
1775 wc->w_handle = handle;
1776
1777 if (clusters_to_alloc) {
1778 ret = dquot_alloc_space_nodirty(inode,
1779 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1780 if (ret)
1781 goto out_commit;
1782 }
1783
1784 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1785 OCFS2_JOURNAL_ACCESS_WRITE);
1786 if (ret) {
1787 mlog_errno(ret);
1788 goto out_quota;
1789 }
1790
1791 /*
1792 * Fill our page array first. That way we've grabbed enough so
1793 * that we can zero and flush if we error after adding the
1794 * extent.
1795 */
1796 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1797 cluster_of_pages, mmap_page);
1798 if (ret) {
1799 /*
1800 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1801 * the target page. In this case, we exit with no error and no target
1802 * page. This will trigger the caller, page_mkwrite(), to re-try
1803 * the operation.
1804 */
1805 if (type == OCFS2_WRITE_MMAP && ret == -EAGAIN) {
1806 BUG_ON(wc->w_target_page);
1807 ret = 0;
1808 goto out_quota;
1809 }
1810
1811 mlog_errno(ret);
1812 goto out_quota;
1813 }
1814
1815 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1816 len);
1817 if (ret) {
1818 mlog_errno(ret);
1819 goto out_quota;
1820 }
1821
1822 if (data_ac)
1823 ocfs2_free_alloc_context(data_ac);
1824 if (meta_ac)
1825 ocfs2_free_alloc_context(meta_ac);
1826
1827 success:
1828 if (pagep)
1829 *pagep = wc->w_target_page;
1830 *fsdata = wc;
1831 return 0;
1832 out_quota:
1833 if (clusters_to_alloc)
1834 dquot_free_space(inode,
1835 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1836 out_commit:
1837 ocfs2_commit_trans(osb, handle);
1838
1839 out:
1840 /*
1841 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1842 * even in case of error here like ENOSPC and ENOMEM. So, we need
1843 * to unlock the target page manually to prevent deadlocks when
1844 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1845 * to VM code.
1846 */
1847 if (wc->w_target_locked)
1848 unlock_page(mmap_page);
1849
1850 ocfs2_free_write_ctxt(inode, wc);
1851
1852 if (data_ac) {
1853 ocfs2_free_alloc_context(data_ac);
1854 data_ac = NULL;
1855 }
1856 if (meta_ac) {
1857 ocfs2_free_alloc_context(meta_ac);
1858 meta_ac = NULL;
1859 }
1860
1861 if (ret == -ENOSPC && try_free) {
1862 /*
1863 * Try to free some truncate log so that we can have enough
1864 * clusters to allocate.
1865 */
1866 try_free = 0;
1867
1868 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1869 if (ret1 == 1)
1870 goto try_again;
1871
1872 if (ret1 < 0)
1873 mlog_errno(ret1);
1874 }
1875
1876 return ret;
1877 }
1878
1879 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1880 loff_t pos, unsigned len,
1881 struct page **pagep, void **fsdata)
1882 {
1883 int ret;
1884 struct buffer_head *di_bh = NULL;
1885 struct inode *inode = mapping->host;
1886
1887 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1888 if (ret) {
1889 mlog_errno(ret);
1890 return ret;
1891 }
1892
1893 /*
1894 * Take alloc sem here to prevent concurrent lookups. That way
1895 * the mapping, zeroing and tree manipulation within
1896 * ocfs2_write() will be safe against ->read_folio(). This
1897 * should also serve to lock out allocation from a shared
1898 * writeable region.
1899 */
1900 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1901
1902 ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1903 pagep, fsdata, di_bh, NULL);
1904 if (ret) {
1905 mlog_errno(ret);
1906 goto out_fail;
1907 }
1908
1909 brelse(di_bh);
1910
1911 return 0;
1912
1913 out_fail:
1914 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1915
1916 brelse(di_bh);
1917 ocfs2_inode_unlock(inode, 1);
1918
1919 return ret;
1920 }
1921
1922 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1923 unsigned len, unsigned *copied,
1924 struct ocfs2_dinode *di,
1925 struct ocfs2_write_ctxt *wc)
1926 {
1927 void *kaddr;
1928
1929 if (unlikely(*copied < len)) {
1930 if (!PageUptodate(wc->w_target_page)) {
1931 *copied = 0;
1932 return;
1933 }
1934 }
1935
1936 kaddr = kmap_atomic(wc->w_target_page);
1937 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1938 kunmap_atomic(kaddr);
1939
1940 trace_ocfs2_write_end_inline(
1941 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1942 (unsigned long long)pos, *copied,
1943 le16_to_cpu(di->id2.i_data.id_count),
1944 le16_to_cpu(di->i_dyn_features));
1945 }
1946
1947 int ocfs2_write_end_nolock(struct address_space *mapping,
1948 loff_t pos, unsigned len, unsigned copied, void *fsdata)
1949 {
1950 int i, ret;
1951 unsigned from, to, start = pos & (PAGE_SIZE - 1);
1952 struct inode *inode = mapping->host;
1953 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1954 struct ocfs2_write_ctxt *wc = fsdata;
1955 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1956 handle_t *handle = wc->w_handle;
1957 struct page *tmppage;
1958
1959 BUG_ON(!list_empty(&wc->w_unwritten_list));
1960
1961 if (handle) {
1962 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1963 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1964 if (ret) {
1965 copied = ret;
1966 mlog_errno(ret);
1967 goto out;
1968 }
1969 }
1970
1971 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1972 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1973 goto out_write_size;
1974 }
1975
1976 if (unlikely(copied < len) && wc->w_target_page) {
1977 loff_t new_isize;
1978
1979 if (!PageUptodate(wc->w_target_page))
1980 copied = 0;
1981
1982 new_isize = max_t(loff_t, i_size_read(inode), pos + copied);
1983 if (new_isize > page_offset(wc->w_target_page))
1984 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1985 start+len);
1986 else {
1987 /*
1988 * When page is fully beyond new isize (data copy
1989 * failed), do not bother zeroing the page. Invalidate
1990 * it instead so that writeback does not get confused
1991 * put page & buffer dirty bits into inconsistent
1992 * state.
1993 */
1994 block_invalidate_folio(page_folio(wc->w_target_page),
1995 0, PAGE_SIZE);
1996 }
1997 }
1998 if (wc->w_target_page)
1999 flush_dcache_page(wc->w_target_page);
2000
2001 for(i = 0; i < wc->w_num_pages; i++) {
2002 tmppage = wc->w_pages[i];
2003
2004 /* This is the direct io target page. */
2005 if (tmppage == NULL)
2006 continue;
2007
2008 if (tmppage == wc->w_target_page) {
2009 from = wc->w_target_from;
2010 to = wc->w_target_to;
2011
2012 BUG_ON(from > PAGE_SIZE ||
2013 to > PAGE_SIZE ||
2014 to < from);
2015 } else {
2016 /*
2017 * Pages adjacent to the target (if any) imply
2018 * a hole-filling write in which case we want
2019 * to flush their entire range.
2020 */
2021 from = 0;
2022 to = PAGE_SIZE;
2023 }
2024
2025 if (page_has_buffers(tmppage)) {
2026 if (handle && ocfs2_should_order_data(inode)) {
2027 loff_t start_byte =
2028 ((loff_t)tmppage->index << PAGE_SHIFT) +
2029 from;
2030 loff_t length = to - from;
2031 ocfs2_jbd2_inode_add_write(handle, inode,
2032 start_byte, length);
2033 }
2034 block_commit_write(tmppage, from, to);
2035 }
2036 }
2037
2038 out_write_size:
2039 /* Direct io do not update i_size here. */
2040 if (wc->w_type != OCFS2_WRITE_DIRECT) {
2041 pos += copied;
2042 if (pos > i_size_read(inode)) {
2043 i_size_write(inode, pos);
2044 mark_inode_dirty(inode);
2045 }
2046 inode->i_blocks = ocfs2_inode_sector_count(inode);
2047 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2048 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2049 di->i_mtime = di->i_ctime = cpu_to_le64(inode_get_mtime_sec(inode));
2050 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode_get_mtime_nsec(inode));
2051 if (handle)
2052 ocfs2_update_inode_fsync_trans(handle, inode, 1);
2053 }
2054 if (handle)
2055 ocfs2_journal_dirty(handle, wc->w_di_bh);
2056
2057 out:
2058 /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2059 * lock, or it will cause a deadlock since journal commit threads holds
2060 * this lock and will ask for the page lock when flushing the data.
2061 * put it here to preserve the unlock order.
2062 */
2063 ocfs2_unlock_pages(wc);
2064
2065 if (handle)
2066 ocfs2_commit_trans(osb, handle);
2067
2068 ocfs2_run_deallocs(osb, &wc->w_dealloc);
2069
2070 brelse(wc->w_di_bh);
2071 kfree(wc);
2072
2073 return copied;
2074 }
2075
2076 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2077 loff_t pos, unsigned len, unsigned copied,
2078 struct page *page, void *fsdata)
2079 {
2080 int ret;
2081 struct inode *inode = mapping->host;
2082
2083 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2084
2085 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2086 ocfs2_inode_unlock(inode, 1);
2087
2088 return ret;
2089 }
2090
2091 struct ocfs2_dio_write_ctxt {
2092 struct list_head dw_zero_list;
2093 unsigned dw_zero_count;
2094 int dw_orphaned;
2095 pid_t dw_writer_pid;
2096 };
2097
2098 static struct ocfs2_dio_write_ctxt *
2099 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2100 {
2101 struct ocfs2_dio_write_ctxt *dwc = NULL;
2102
2103 if (bh->b_private)
2104 return bh->b_private;
2105
2106 dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2107 if (dwc == NULL)
2108 return NULL;
2109 INIT_LIST_HEAD(&dwc->dw_zero_list);
2110 dwc->dw_zero_count = 0;
2111 dwc->dw_orphaned = 0;
2112 dwc->dw_writer_pid = task_pid_nr(current);
2113 bh->b_private = dwc;
2114 *alloc = 1;
2115
2116 return dwc;
2117 }
2118
2119 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2120 struct ocfs2_dio_write_ctxt *dwc)
2121 {
2122 ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2123 kfree(dwc);
2124 }
2125
2126 /*
2127 * TODO: Make this into a generic get_blocks function.
2128 *
2129 * From do_direct_io in direct-io.c:
2130 * "So what we do is to permit the ->get_blocks function to populate
2131 * bh.b_size with the size of IO which is permitted at this offset and
2132 * this i_blkbits."
2133 *
2134 * This function is called directly from get_more_blocks in direct-io.c.
2135 *
2136 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2137 * fs_count, map_bh, dio->rw == WRITE);
2138 */
2139 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2140 struct buffer_head *bh_result, int create)
2141 {
2142 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2143 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2144 struct ocfs2_write_ctxt *wc;
2145 struct ocfs2_write_cluster_desc *desc = NULL;
2146 struct ocfs2_dio_write_ctxt *dwc = NULL;
2147 struct buffer_head *di_bh = NULL;
2148 u64 p_blkno;
2149 unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2150 loff_t pos = iblock << i_blkbits;
2151 sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2152 unsigned len, total_len = bh_result->b_size;
2153 int ret = 0, first_get_block = 0;
2154
2155 len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2156 len = min(total_len, len);
2157
2158 /*
2159 * bh_result->b_size is count in get_more_blocks according to write
2160 * "pos" and "end", we need map twice to return different buffer state:
2161 * 1. area in file size, not set NEW;
2162 * 2. area out file size, set NEW.
2163 *
2164 * iblock endblk
2165 * |--------|---------|---------|---------
2166 * |<-------area in file------->|
2167 */
2168
2169 if ((iblock <= endblk) &&
2170 ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2171 len = (endblk - iblock + 1) << i_blkbits;
2172
2173 mlog(0, "get block of %lu at %llu:%u req %u\n",
2174 inode->i_ino, pos, len, total_len);
2175
2176 /*
2177 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2178 * we may need to add it to orphan dir. So can not fall to fast path
2179 * while file size will be changed.
2180 */
2181 if (pos + total_len <= i_size_read(inode)) {
2182
2183 /* This is the fast path for re-write. */
2184 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2185 if (buffer_mapped(bh_result) &&
2186 !buffer_new(bh_result) &&
2187 ret == 0)
2188 goto out;
2189
2190 /* Clear state set by ocfs2_get_block. */
2191 bh_result->b_state = 0;
2192 }
2193
2194 dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2195 if (unlikely(dwc == NULL)) {
2196 ret = -ENOMEM;
2197 mlog_errno(ret);
2198 goto out;
2199 }
2200
2201 if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2202 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2203 !dwc->dw_orphaned) {
2204 /*
2205 * when we are going to alloc extents beyond file size, add the
2206 * inode to orphan dir, so we can recall those spaces when
2207 * system crashed during write.
2208 */
2209 ret = ocfs2_add_inode_to_orphan(osb, inode);
2210 if (ret < 0) {
2211 mlog_errno(ret);
2212 goto out;
2213 }
2214 dwc->dw_orphaned = 1;
2215 }
2216
2217 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2218 if (ret) {
2219 mlog_errno(ret);
2220 goto out;
2221 }
2222
2223 down_write(&oi->ip_alloc_sem);
2224
2225 if (first_get_block) {
2226 if (ocfs2_sparse_alloc(osb))
2227 ret = ocfs2_zero_tail(inode, di_bh, pos);
2228 else
2229 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2230 total_len, NULL);
2231 if (ret < 0) {
2232 mlog_errno(ret);
2233 goto unlock;
2234 }
2235 }
2236
2237 ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2238 OCFS2_WRITE_DIRECT, NULL,
2239 (void **)&wc, di_bh, NULL);
2240 if (ret) {
2241 mlog_errno(ret);
2242 goto unlock;
2243 }
2244
2245 desc = &wc->w_desc[0];
2246
2247 p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2248 BUG_ON(p_blkno == 0);
2249 p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2250
2251 map_bh(bh_result, inode->i_sb, p_blkno);
2252 bh_result->b_size = len;
2253 if (desc->c_needs_zero)
2254 set_buffer_new(bh_result);
2255
2256 if (iblock > endblk)
2257 set_buffer_new(bh_result);
2258
2259 /* May sleep in end_io. It should not happen in a irq context. So defer
2260 * it to dio work queue. */
2261 set_buffer_defer_completion(bh_result);
2262
2263 if (!list_empty(&wc->w_unwritten_list)) {
2264 struct ocfs2_unwritten_extent *ue = NULL;
2265
2266 ue = list_first_entry(&wc->w_unwritten_list,
2267 struct ocfs2_unwritten_extent,
2268 ue_node);
2269 BUG_ON(ue->ue_cpos != desc->c_cpos);
2270 /* The physical address may be 0, fill it. */
2271 ue->ue_phys = desc->c_phys;
2272
2273 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2274 dwc->dw_zero_count += wc->w_unwritten_count;
2275 }
2276
2277 ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2278 BUG_ON(ret != len);
2279 ret = 0;
2280 unlock:
2281 up_write(&oi->ip_alloc_sem);
2282 ocfs2_inode_unlock(inode, 1);
2283 brelse(di_bh);
2284 out:
2285 return ret;
2286 }
2287
2288 static int ocfs2_dio_end_io_write(struct inode *inode,
2289 struct ocfs2_dio_write_ctxt *dwc,
2290 loff_t offset,
2291 ssize_t bytes)
2292 {
2293 struct ocfs2_cached_dealloc_ctxt dealloc;
2294 struct ocfs2_extent_tree et;
2295 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2296 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2297 struct ocfs2_unwritten_extent *ue = NULL;
2298 struct buffer_head *di_bh = NULL;
2299 struct ocfs2_dinode *di;
2300 struct ocfs2_alloc_context *data_ac = NULL;
2301 struct ocfs2_alloc_context *meta_ac = NULL;
2302 handle_t *handle = NULL;
2303 loff_t end = offset + bytes;
2304 int ret = 0, credits = 0;
2305
2306 ocfs2_init_dealloc_ctxt(&dealloc);
2307
2308 /* We do clear unwritten, delete orphan, change i_size here. If neither
2309 * of these happen, we can skip all this. */
2310 if (list_empty(&dwc->dw_zero_list) &&
2311 end <= i_size_read(inode) &&
2312 !dwc->dw_orphaned)
2313 goto out;
2314
2315 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2316 if (ret < 0) {
2317 mlog_errno(ret);
2318 goto out;
2319 }
2320
2321 down_write(&oi->ip_alloc_sem);
2322
2323 /* Delete orphan before acquire i_rwsem. */
2324 if (dwc->dw_orphaned) {
2325 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2326
2327 end = end > i_size_read(inode) ? end : 0;
2328
2329 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2330 !!end, end);
2331 if (ret < 0)
2332 mlog_errno(ret);
2333 }
2334
2335 di = (struct ocfs2_dinode *)di_bh->b_data;
2336
2337 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2338
2339 /* Attach dealloc with extent tree in case that we may reuse extents
2340 * which are already unlinked from current extent tree due to extent
2341 * rotation and merging.
2342 */
2343 et.et_dealloc = &dealloc;
2344
2345 ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2346 &data_ac, &meta_ac);
2347 if (ret) {
2348 mlog_errno(ret);
2349 goto unlock;
2350 }
2351
2352 credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2353
2354 handle = ocfs2_start_trans(osb, credits);
2355 if (IS_ERR(handle)) {
2356 ret = PTR_ERR(handle);
2357 mlog_errno(ret);
2358 goto unlock;
2359 }
2360 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2361 OCFS2_JOURNAL_ACCESS_WRITE);
2362 if (ret) {
2363 mlog_errno(ret);
2364 goto commit;
2365 }
2366
2367 list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2368 ret = ocfs2_assure_trans_credits(handle, credits);
2369 if (ret < 0) {
2370 mlog_errno(ret);
2371 break;
2372 }
2373 ret = ocfs2_mark_extent_written(inode, &et, handle,
2374 ue->ue_cpos, 1,
2375 ue->ue_phys,
2376 meta_ac, &dealloc);
2377 if (ret < 0) {
2378 mlog_errno(ret);
2379 break;
2380 }
2381 }
2382
2383 if (end > i_size_read(inode)) {
2384 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2385 if (ret < 0)
2386 mlog_errno(ret);
2387 }
2388 commit:
2389 ocfs2_commit_trans(osb, handle);
2390 unlock:
2391 up_write(&oi->ip_alloc_sem);
2392 ocfs2_inode_unlock(inode, 1);
2393 brelse(di_bh);
2394 out:
2395 if (data_ac)
2396 ocfs2_free_alloc_context(data_ac);
2397 if (meta_ac)
2398 ocfs2_free_alloc_context(meta_ac);
2399 ocfs2_run_deallocs(osb, &dealloc);
2400 ocfs2_dio_free_write_ctx(inode, dwc);
2401
2402 return ret;
2403 }
2404
2405 /*
2406 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
2407 * particularly interested in the aio/dio case. We use the rw_lock DLM lock
2408 * to protect io on one node from truncation on another.
2409 */
2410 static int ocfs2_dio_end_io(struct kiocb *iocb,
2411 loff_t offset,
2412 ssize_t bytes,
2413 void *private)
2414 {
2415 struct inode *inode = file_inode(iocb->ki_filp);
2416 int level;
2417 int ret = 0;
2418
2419 /* this io's submitter should not have unlocked this before we could */
2420 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2421
2422 if (bytes <= 0)
2423 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2424 (long long)bytes);
2425 if (private) {
2426 if (bytes > 0)
2427 ret = ocfs2_dio_end_io_write(inode, private, offset,
2428 bytes);
2429 else
2430 ocfs2_dio_free_write_ctx(inode, private);
2431 }
2432
2433 ocfs2_iocb_clear_rw_locked(iocb);
2434
2435 level = ocfs2_iocb_rw_locked_level(iocb);
2436 ocfs2_rw_unlock(inode, level);
2437 return ret;
2438 }
2439
2440 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2441 {
2442 struct file *file = iocb->ki_filp;
2443 struct inode *inode = file->f_mapping->host;
2444 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2445 get_block_t *get_block;
2446
2447 /*
2448 * Fallback to buffered I/O if we see an inode without
2449 * extents.
2450 */
2451 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2452 return 0;
2453
2454 /* Fallback to buffered I/O if we do not support append dio. */
2455 if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2456 !ocfs2_supports_append_dio(osb))
2457 return 0;
2458
2459 if (iov_iter_rw(iter) == READ)
2460 get_block = ocfs2_lock_get_block;
2461 else
2462 get_block = ocfs2_dio_wr_get_block;
2463
2464 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2465 iter, get_block,
2466 ocfs2_dio_end_io, 0);
2467 }
2468
2469 const struct address_space_operations ocfs2_aops = {
2470 .dirty_folio = block_dirty_folio,
2471 .read_folio = ocfs2_read_folio,
2472 .readahead = ocfs2_readahead,
2473 .writepages = ocfs2_writepages,
2474 .write_begin = ocfs2_write_begin,
2475 .write_end = ocfs2_write_end,
2476 .bmap = ocfs2_bmap,
2477 .direct_IO = ocfs2_direct_IO,
2478 .invalidate_folio = block_invalidate_folio,
2479 .release_folio = ocfs2_release_folio,
2480 .migrate_folio = buffer_migrate_folio,
2481 .is_partially_uptodate = block_is_partially_uptodate,
2482 .error_remove_folio = generic_error_remove_folio,
2483 };
2484
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.