]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - fs/ext4/inode.c
Merge branch 'for-linus' of git://git.open-osd.org/linux-open-osd
[mirror_ubuntu-artful-kernel.git] / fs / ext4 / inode.c
1 /*
2 * linux/fs/ext4/inode.c
3 *
4 * Copyright (C) 1992, 1993, 1994, 1995
5 * Remy Card (card@masi.ibp.fr)
6 * Laboratoire MASI - Institut Blaise Pascal
7 * Universite Pierre et Marie Curie (Paris VI)
8 *
9 * from
10 *
11 * linux/fs/minix/inode.c
12 *
13 * Copyright (C) 1991, 1992 Linus Torvalds
14 *
15 * 64-bit file support on 64-bit platforms by Jakub Jelinek
16 * (jj@sunsite.ms.mff.cuni.cz)
17 *
18 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19 */
20
21 #include <linux/module.h>
22 #include <linux/fs.h>
23 #include <linux/time.h>
24 #include <linux/jbd2.h>
25 #include <linux/highuid.h>
26 #include <linux/pagemap.h>
27 #include <linux/quotaops.h>
28 #include <linux/string.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/pagevec.h>
32 #include <linux/mpage.h>
33 #include <linux/namei.h>
34 #include <linux/uio.h>
35 #include <linux/bio.h>
36 #include <linux/workqueue.h>
37 #include <linux/kernel.h>
38 #include <linux/printk.h>
39 #include <linux/slab.h>
40 #include <linux/ratelimit.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46 #include "truncate.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53 loff_t new_size)
54 {
55 trace_ext4_begin_ordered_truncate(inode, new_size);
56 /*
57 * If jinode is zero, then we never opened the file for
58 * writing, so there's no need to call
59 * jbd2_journal_begin_ordered_truncate() since there's no
60 * outstanding writes we need to flush.
61 */
62 if (!EXT4_I(inode)->jinode)
63 return 0;
64 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
65 EXT4_I(inode)->jinode,
66 new_size);
67 }
68
69 static void ext4_invalidatepage(struct page *page, unsigned long offset);
70 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
71 struct buffer_head *bh_result, int create);
72 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
73 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
74 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
75 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
76
77 /*
78 * Test whether an inode is a fast symlink.
79 */
80 static int ext4_inode_is_fast_symlink(struct inode *inode)
81 {
82 int ea_blocks = EXT4_I(inode)->i_file_acl ?
83 (inode->i_sb->s_blocksize >> 9) : 0;
84
85 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
86 }
87
88 /*
89 * Restart the transaction associated with *handle. This does a commit,
90 * so before we call here everything must be consistently dirtied against
91 * this transaction.
92 */
93 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
94 int nblocks)
95 {
96 int ret;
97
98 /*
99 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
100 * moment, get_block can be called only for blocks inside i_size since
101 * page cache has been already dropped and writes are blocked by
102 * i_mutex. So we can safely drop the i_data_sem here.
103 */
104 BUG_ON(EXT4_JOURNAL(inode) == NULL);
105 jbd_debug(2, "restarting handle %p\n", handle);
106 up_write(&EXT4_I(inode)->i_data_sem);
107 ret = ext4_journal_restart(handle, nblocks);
108 down_write(&EXT4_I(inode)->i_data_sem);
109 ext4_discard_preallocations(inode);
110
111 return ret;
112 }
113
114 /*
115 * Called at the last iput() if i_nlink is zero.
116 */
117 void ext4_evict_inode(struct inode *inode)
118 {
119 handle_t *handle;
120 int err;
121
122 trace_ext4_evict_inode(inode);
123 if (inode->i_nlink) {
124 /*
125 * When journalling data dirty buffers are tracked only in the
126 * journal. So although mm thinks everything is clean and
127 * ready for reaping the inode might still have some pages to
128 * write in the running transaction or waiting to be
129 * checkpointed. Thus calling jbd2_journal_invalidatepage()
130 * (via truncate_inode_pages()) to discard these buffers can
131 * cause data loss. Also even if we did not discard these
132 * buffers, we would have no way to find them after the inode
133 * is reaped and thus user could see stale data if he tries to
134 * read them before the transaction is checkpointed. So be
135 * careful and force everything to disk here... We use
136 * ei->i_datasync_tid to store the newest transaction
137 * containing inode's data.
138 *
139 * Note that directories do not have this problem because they
140 * don't use page cache.
141 */
142 if (ext4_should_journal_data(inode) &&
143 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
144 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
145 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
146
147 jbd2_log_start_commit(journal, commit_tid);
148 jbd2_log_wait_commit(journal, commit_tid);
149 filemap_write_and_wait(&inode->i_data);
150 }
151 truncate_inode_pages(&inode->i_data, 0);
152 goto no_delete;
153 }
154
155 if (!is_bad_inode(inode))
156 dquot_initialize(inode);
157
158 if (ext4_should_order_data(inode))
159 ext4_begin_ordered_truncate(inode, 0);
160 truncate_inode_pages(&inode->i_data, 0);
161
162 if (is_bad_inode(inode))
163 goto no_delete;
164
165 handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
166 if (IS_ERR(handle)) {
167 ext4_std_error(inode->i_sb, PTR_ERR(handle));
168 /*
169 * If we're going to skip the normal cleanup, we still need to
170 * make sure that the in-core orphan linked list is properly
171 * cleaned up.
172 */
173 ext4_orphan_del(NULL, inode);
174 goto no_delete;
175 }
176
177 if (IS_SYNC(inode))
178 ext4_handle_sync(handle);
179 inode->i_size = 0;
180 err = ext4_mark_inode_dirty(handle, inode);
181 if (err) {
182 ext4_warning(inode->i_sb,
183 "couldn't mark inode dirty (err %d)", err);
184 goto stop_handle;
185 }
186 if (inode->i_blocks)
187 ext4_truncate(inode);
188
189 /*
190 * ext4_ext_truncate() doesn't reserve any slop when it
191 * restarts journal transactions; therefore there may not be
192 * enough credits left in the handle to remove the inode from
193 * the orphan list and set the dtime field.
194 */
195 if (!ext4_handle_has_enough_credits(handle, 3)) {
196 err = ext4_journal_extend(handle, 3);
197 if (err > 0)
198 err = ext4_journal_restart(handle, 3);
199 if (err != 0) {
200 ext4_warning(inode->i_sb,
201 "couldn't extend journal (err %d)", err);
202 stop_handle:
203 ext4_journal_stop(handle);
204 ext4_orphan_del(NULL, inode);
205 goto no_delete;
206 }
207 }
208
209 /*
210 * Kill off the orphan record which ext4_truncate created.
211 * AKPM: I think this can be inside the above `if'.
212 * Note that ext4_orphan_del() has to be able to cope with the
213 * deletion of a non-existent orphan - this is because we don't
214 * know if ext4_truncate() actually created an orphan record.
215 * (Well, we could do this if we need to, but heck - it works)
216 */
217 ext4_orphan_del(handle, inode);
218 EXT4_I(inode)->i_dtime = get_seconds();
219
220 /*
221 * One subtle ordering requirement: if anything has gone wrong
222 * (transaction abort, IO errors, whatever), then we can still
223 * do these next steps (the fs will already have been marked as
224 * having errors), but we can't free the inode if the mark_dirty
225 * fails.
226 */
227 if (ext4_mark_inode_dirty(handle, inode))
228 /* If that failed, just do the required in-core inode clear. */
229 ext4_clear_inode(inode);
230 else
231 ext4_free_inode(handle, inode);
232 ext4_journal_stop(handle);
233 return;
234 no_delete:
235 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
236 }
237
238 #ifdef CONFIG_QUOTA
239 qsize_t *ext4_get_reserved_space(struct inode *inode)
240 {
241 return &EXT4_I(inode)->i_reserved_quota;
242 }
243 #endif
244
245 /*
246 * Calculate the number of metadata blocks need to reserve
247 * to allocate a block located at @lblock
248 */
249 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
250 {
251 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
252 return ext4_ext_calc_metadata_amount(inode, lblock);
253
254 return ext4_ind_calc_metadata_amount(inode, lblock);
255 }
256
257 /*
258 * Called with i_data_sem down, which is important since we can call
259 * ext4_discard_preallocations() from here.
260 */
261 void ext4_da_update_reserve_space(struct inode *inode,
262 int used, int quota_claim)
263 {
264 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
265 struct ext4_inode_info *ei = EXT4_I(inode);
266
267 spin_lock(&ei->i_block_reservation_lock);
268 trace_ext4_da_update_reserve_space(inode, used);
269 if (unlikely(used > ei->i_reserved_data_blocks)) {
270 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
271 "with only %d reserved data blocks\n",
272 __func__, inode->i_ino, used,
273 ei->i_reserved_data_blocks);
274 WARN_ON(1);
275 used = ei->i_reserved_data_blocks;
276 }
277
278 /* Update per-inode reservations */
279 ei->i_reserved_data_blocks -= used;
280 ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
281 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
282 used + ei->i_allocated_meta_blocks);
283 ei->i_allocated_meta_blocks = 0;
284
285 if (ei->i_reserved_data_blocks == 0) {
286 /*
287 * We can release all of the reserved metadata blocks
288 * only when we have written all of the delayed
289 * allocation blocks.
290 */
291 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
292 ei->i_reserved_meta_blocks);
293 ei->i_reserved_meta_blocks = 0;
294 ei->i_da_metadata_calc_len = 0;
295 }
296 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
297
298 /* Update quota subsystem for data blocks */
299 if (quota_claim)
300 dquot_claim_block(inode, used);
301 else {
302 /*
303 * We did fallocate with an offset that is already delayed
304 * allocated. So on delayed allocated writeback we should
305 * not re-claim the quota for fallocated blocks.
306 */
307 dquot_release_reservation_block(inode, used);
308 }
309
310 /*
311 * If we have done all the pending block allocations and if
312 * there aren't any writers on the inode, we can discard the
313 * inode's preallocations.
314 */
315 if ((ei->i_reserved_data_blocks == 0) &&
316 (atomic_read(&inode->i_writecount) == 0))
317 ext4_discard_preallocations(inode);
318 }
319
320 static int __check_block_validity(struct inode *inode, const char *func,
321 unsigned int line,
322 struct ext4_map_blocks *map)
323 {
324 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
325 map->m_len)) {
326 ext4_error_inode(inode, func, line, map->m_pblk,
327 "lblock %lu mapped to illegal pblock "
328 "(length %d)", (unsigned long) map->m_lblk,
329 map->m_len);
330 return -EIO;
331 }
332 return 0;
333 }
334
335 #define check_block_validity(inode, map) \
336 __check_block_validity((inode), __func__, __LINE__, (map))
337
338 /*
339 * Return the number of contiguous dirty pages in a given inode
340 * starting at page frame idx.
341 */
342 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
343 unsigned int max_pages)
344 {
345 struct address_space *mapping = inode->i_mapping;
346 pgoff_t index;
347 struct pagevec pvec;
348 pgoff_t num = 0;
349 int i, nr_pages, done = 0;
350
351 if (max_pages == 0)
352 return 0;
353 pagevec_init(&pvec, 0);
354 while (!done) {
355 index = idx;
356 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
357 PAGECACHE_TAG_DIRTY,
358 (pgoff_t)PAGEVEC_SIZE);
359 if (nr_pages == 0)
360 break;
361 for (i = 0; i < nr_pages; i++) {
362 struct page *page = pvec.pages[i];
363 struct buffer_head *bh, *head;
364
365 lock_page(page);
366 if (unlikely(page->mapping != mapping) ||
367 !PageDirty(page) ||
368 PageWriteback(page) ||
369 page->index != idx) {
370 done = 1;
371 unlock_page(page);
372 break;
373 }
374 if (page_has_buffers(page)) {
375 bh = head = page_buffers(page);
376 do {
377 if (!buffer_delay(bh) &&
378 !buffer_unwritten(bh))
379 done = 1;
380 bh = bh->b_this_page;
381 } while (!done && (bh != head));
382 }
383 unlock_page(page);
384 if (done)
385 break;
386 idx++;
387 num++;
388 if (num >= max_pages) {
389 done = 1;
390 break;
391 }
392 }
393 pagevec_release(&pvec);
394 }
395 return num;
396 }
397
398 /*
399 * The ext4_map_blocks() function tries to look up the requested blocks,
400 * and returns if the blocks are already mapped.
401 *
402 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
403 * and store the allocated blocks in the result buffer head and mark it
404 * mapped.
405 *
406 * If file type is extents based, it will call ext4_ext_map_blocks(),
407 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
408 * based files
409 *
410 * On success, it returns the number of blocks being mapped or allocate.
411 * if create==0 and the blocks are pre-allocated and uninitialized block,
412 * the result buffer head is unmapped. If the create ==1, it will make sure
413 * the buffer head is mapped.
414 *
415 * It returns 0 if plain look up failed (blocks have not been allocated), in
416 * that casem, buffer head is unmapped
417 *
418 * It returns the error in case of allocation failure.
419 */
420 int ext4_map_blocks(handle_t *handle, struct inode *inode,
421 struct ext4_map_blocks *map, int flags)
422 {
423 int retval;
424
425 map->m_flags = 0;
426 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
427 "logical block %lu\n", inode->i_ino, flags, map->m_len,
428 (unsigned long) map->m_lblk);
429 /*
430 * Try to see if we can get the block without requesting a new
431 * file system block.
432 */
433 down_read((&EXT4_I(inode)->i_data_sem));
434 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
435 retval = ext4_ext_map_blocks(handle, inode, map, 0);
436 } else {
437 retval = ext4_ind_map_blocks(handle, inode, map, 0);
438 }
439 up_read((&EXT4_I(inode)->i_data_sem));
440
441 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
442 int ret = check_block_validity(inode, map);
443 if (ret != 0)
444 return ret;
445 }
446
447 /* If it is only a block(s) look up */
448 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
449 return retval;
450
451 /*
452 * Returns if the blocks have already allocated
453 *
454 * Note that if blocks have been preallocated
455 * ext4_ext_get_block() returns th create = 0
456 * with buffer head unmapped.
457 */
458 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
459 return retval;
460
461 /*
462 * When we call get_blocks without the create flag, the
463 * BH_Unwritten flag could have gotten set if the blocks
464 * requested were part of a uninitialized extent. We need to
465 * clear this flag now that we are committed to convert all or
466 * part of the uninitialized extent to be an initialized
467 * extent. This is because we need to avoid the combination
468 * of BH_Unwritten and BH_Mapped flags being simultaneously
469 * set on the buffer_head.
470 */
471 map->m_flags &= ~EXT4_MAP_UNWRITTEN;
472
473 /*
474 * New blocks allocate and/or writing to uninitialized extent
475 * will possibly result in updating i_data, so we take
476 * the write lock of i_data_sem, and call get_blocks()
477 * with create == 1 flag.
478 */
479 down_write((&EXT4_I(inode)->i_data_sem));
480
481 /*
482 * if the caller is from delayed allocation writeout path
483 * we have already reserved fs blocks for allocation
484 * let the underlying get_block() function know to
485 * avoid double accounting
486 */
487 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
488 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
489 /*
490 * We need to check for EXT4 here because migrate
491 * could have changed the inode type in between
492 */
493 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
494 retval = ext4_ext_map_blocks(handle, inode, map, flags);
495 } else {
496 retval = ext4_ind_map_blocks(handle, inode, map, flags);
497
498 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
499 /*
500 * We allocated new blocks which will result in
501 * i_data's format changing. Force the migrate
502 * to fail by clearing migrate flags
503 */
504 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
505 }
506
507 /*
508 * Update reserved blocks/metadata blocks after successful
509 * block allocation which had been deferred till now. We don't
510 * support fallocate for non extent files. So we can update
511 * reserve space here.
512 */
513 if ((retval > 0) &&
514 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
515 ext4_da_update_reserve_space(inode, retval, 1);
516 }
517 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
518 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
519
520 up_write((&EXT4_I(inode)->i_data_sem));
521 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
522 int ret = check_block_validity(inode, map);
523 if (ret != 0)
524 return ret;
525 }
526 return retval;
527 }
528
529 /* Maximum number of blocks we map for direct IO at once. */
530 #define DIO_MAX_BLOCKS 4096
531
532 static int _ext4_get_block(struct inode *inode, sector_t iblock,
533 struct buffer_head *bh, int flags)
534 {
535 handle_t *handle = ext4_journal_current_handle();
536 struct ext4_map_blocks map;
537 int ret = 0, started = 0;
538 int dio_credits;
539
540 map.m_lblk = iblock;
541 map.m_len = bh->b_size >> inode->i_blkbits;
542
543 if (flags && !handle) {
544 /* Direct IO write... */
545 if (map.m_len > DIO_MAX_BLOCKS)
546 map.m_len = DIO_MAX_BLOCKS;
547 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
548 handle = ext4_journal_start(inode, dio_credits);
549 if (IS_ERR(handle)) {
550 ret = PTR_ERR(handle);
551 return ret;
552 }
553 started = 1;
554 }
555
556 ret = ext4_map_blocks(handle, inode, &map, flags);
557 if (ret > 0) {
558 map_bh(bh, inode->i_sb, map.m_pblk);
559 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
560 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
561 ret = 0;
562 }
563 if (started)
564 ext4_journal_stop(handle);
565 return ret;
566 }
567
568 int ext4_get_block(struct inode *inode, sector_t iblock,
569 struct buffer_head *bh, int create)
570 {
571 return _ext4_get_block(inode, iblock, bh,
572 create ? EXT4_GET_BLOCKS_CREATE : 0);
573 }
574
575 /*
576 * `handle' can be NULL if create is zero
577 */
578 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
579 ext4_lblk_t block, int create, int *errp)
580 {
581 struct ext4_map_blocks map;
582 struct buffer_head *bh;
583 int fatal = 0, err;
584
585 J_ASSERT(handle != NULL || create == 0);
586
587 map.m_lblk = block;
588 map.m_len = 1;
589 err = ext4_map_blocks(handle, inode, &map,
590 create ? EXT4_GET_BLOCKS_CREATE : 0);
591
592 if (err < 0)
593 *errp = err;
594 if (err <= 0)
595 return NULL;
596 *errp = 0;
597
598 bh = sb_getblk(inode->i_sb, map.m_pblk);
599 if (!bh) {
600 *errp = -EIO;
601 return NULL;
602 }
603 if (map.m_flags & EXT4_MAP_NEW) {
604 J_ASSERT(create != 0);
605 J_ASSERT(handle != NULL);
606
607 /*
608 * Now that we do not always journal data, we should
609 * keep in mind whether this should always journal the
610 * new buffer as metadata. For now, regular file
611 * writes use ext4_get_block instead, so it's not a
612 * problem.
613 */
614 lock_buffer(bh);
615 BUFFER_TRACE(bh, "call get_create_access");
616 fatal = ext4_journal_get_create_access(handle, bh);
617 if (!fatal && !buffer_uptodate(bh)) {
618 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
619 set_buffer_uptodate(bh);
620 }
621 unlock_buffer(bh);
622 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
623 err = ext4_handle_dirty_metadata(handle, inode, bh);
624 if (!fatal)
625 fatal = err;
626 } else {
627 BUFFER_TRACE(bh, "not a new buffer");
628 }
629 if (fatal) {
630 *errp = fatal;
631 brelse(bh);
632 bh = NULL;
633 }
634 return bh;
635 }
636
637 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
638 ext4_lblk_t block, int create, int *err)
639 {
640 struct buffer_head *bh;
641
642 bh = ext4_getblk(handle, inode, block, create, err);
643 if (!bh)
644 return bh;
645 if (buffer_uptodate(bh))
646 return bh;
647 ll_rw_block(READ_META, 1, &bh);
648 wait_on_buffer(bh);
649 if (buffer_uptodate(bh))
650 return bh;
651 put_bh(bh);
652 *err = -EIO;
653 return NULL;
654 }
655
656 static int walk_page_buffers(handle_t *handle,
657 struct buffer_head *head,
658 unsigned from,
659 unsigned to,
660 int *partial,
661 int (*fn)(handle_t *handle,
662 struct buffer_head *bh))
663 {
664 struct buffer_head *bh;
665 unsigned block_start, block_end;
666 unsigned blocksize = head->b_size;
667 int err, ret = 0;
668 struct buffer_head *next;
669
670 for (bh = head, block_start = 0;
671 ret == 0 && (bh != head || !block_start);
672 block_start = block_end, bh = next) {
673 next = bh->b_this_page;
674 block_end = block_start + blocksize;
675 if (block_end <= from || block_start >= to) {
676 if (partial && !buffer_uptodate(bh))
677 *partial = 1;
678 continue;
679 }
680 err = (*fn)(handle, bh);
681 if (!ret)
682 ret = err;
683 }
684 return ret;
685 }
686
687 /*
688 * To preserve ordering, it is essential that the hole instantiation and
689 * the data write be encapsulated in a single transaction. We cannot
690 * close off a transaction and start a new one between the ext4_get_block()
691 * and the commit_write(). So doing the jbd2_journal_start at the start of
692 * prepare_write() is the right place.
693 *
694 * Also, this function can nest inside ext4_writepage() ->
695 * block_write_full_page(). In that case, we *know* that ext4_writepage()
696 * has generated enough buffer credits to do the whole page. So we won't
697 * block on the journal in that case, which is good, because the caller may
698 * be PF_MEMALLOC.
699 *
700 * By accident, ext4 can be reentered when a transaction is open via
701 * quota file writes. If we were to commit the transaction while thus
702 * reentered, there can be a deadlock - we would be holding a quota
703 * lock, and the commit would never complete if another thread had a
704 * transaction open and was blocking on the quota lock - a ranking
705 * violation.
706 *
707 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
708 * will _not_ run commit under these circumstances because handle->h_ref
709 * is elevated. We'll still have enough credits for the tiny quotafile
710 * write.
711 */
712 static int do_journal_get_write_access(handle_t *handle,
713 struct buffer_head *bh)
714 {
715 int dirty = buffer_dirty(bh);
716 int ret;
717
718 if (!buffer_mapped(bh) || buffer_freed(bh))
719 return 0;
720 /*
721 * __block_write_begin() could have dirtied some buffers. Clean
722 * the dirty bit as jbd2_journal_get_write_access() could complain
723 * otherwise about fs integrity issues. Setting of the dirty bit
724 * by __block_write_begin() isn't a real problem here as we clear
725 * the bit before releasing a page lock and thus writeback cannot
726 * ever write the buffer.
727 */
728 if (dirty)
729 clear_buffer_dirty(bh);
730 ret = ext4_journal_get_write_access(handle, bh);
731 if (!ret && dirty)
732 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
733 return ret;
734 }
735
736 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
737 struct buffer_head *bh_result, int create);
738 static int ext4_write_begin(struct file *file, struct address_space *mapping,
739 loff_t pos, unsigned len, unsigned flags,
740 struct page **pagep, void **fsdata)
741 {
742 struct inode *inode = mapping->host;
743 int ret, needed_blocks;
744 handle_t *handle;
745 int retries = 0;
746 struct page *page;
747 pgoff_t index;
748 unsigned from, to;
749
750 trace_ext4_write_begin(inode, pos, len, flags);
751 /*
752 * Reserve one block more for addition to orphan list in case
753 * we allocate blocks but write fails for some reason
754 */
755 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
756 index = pos >> PAGE_CACHE_SHIFT;
757 from = pos & (PAGE_CACHE_SIZE - 1);
758 to = from + len;
759
760 retry:
761 handle = ext4_journal_start(inode, needed_blocks);
762 if (IS_ERR(handle)) {
763 ret = PTR_ERR(handle);
764 goto out;
765 }
766
767 /* We cannot recurse into the filesystem as the transaction is already
768 * started */
769 flags |= AOP_FLAG_NOFS;
770
771 page = grab_cache_page_write_begin(mapping, index, flags);
772 if (!page) {
773 ext4_journal_stop(handle);
774 ret = -ENOMEM;
775 goto out;
776 }
777 *pagep = page;
778
779 if (ext4_should_dioread_nolock(inode))
780 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
781 else
782 ret = __block_write_begin(page, pos, len, ext4_get_block);
783
784 if (!ret && ext4_should_journal_data(inode)) {
785 ret = walk_page_buffers(handle, page_buffers(page),
786 from, to, NULL, do_journal_get_write_access);
787 }
788
789 if (ret) {
790 unlock_page(page);
791 page_cache_release(page);
792 /*
793 * __block_write_begin may have instantiated a few blocks
794 * outside i_size. Trim these off again. Don't need
795 * i_size_read because we hold i_mutex.
796 *
797 * Add inode to orphan list in case we crash before
798 * truncate finishes
799 */
800 if (pos + len > inode->i_size && ext4_can_truncate(inode))
801 ext4_orphan_add(handle, inode);
802
803 ext4_journal_stop(handle);
804 if (pos + len > inode->i_size) {
805 ext4_truncate_failed_write(inode);
806 /*
807 * If truncate failed early the inode might
808 * still be on the orphan list; we need to
809 * make sure the inode is removed from the
810 * orphan list in that case.
811 */
812 if (inode->i_nlink)
813 ext4_orphan_del(NULL, inode);
814 }
815 }
816
817 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
818 goto retry;
819 out:
820 return ret;
821 }
822
823 /* For write_end() in data=journal mode */
824 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
825 {
826 if (!buffer_mapped(bh) || buffer_freed(bh))
827 return 0;
828 set_buffer_uptodate(bh);
829 return ext4_handle_dirty_metadata(handle, NULL, bh);
830 }
831
832 static int ext4_generic_write_end(struct file *file,
833 struct address_space *mapping,
834 loff_t pos, unsigned len, unsigned copied,
835 struct page *page, void *fsdata)
836 {
837 int i_size_changed = 0;
838 struct inode *inode = mapping->host;
839 handle_t *handle = ext4_journal_current_handle();
840
841 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
842
843 /*
844 * No need to use i_size_read() here, the i_size
845 * cannot change under us because we hold i_mutex.
846 *
847 * But it's important to update i_size while still holding page lock:
848 * page writeout could otherwise come in and zero beyond i_size.
849 */
850 if (pos + copied > inode->i_size) {
851 i_size_write(inode, pos + copied);
852 i_size_changed = 1;
853 }
854
855 if (pos + copied > EXT4_I(inode)->i_disksize) {
856 /* We need to mark inode dirty even if
857 * new_i_size is less that inode->i_size
858 * bu greater than i_disksize.(hint delalloc)
859 */
860 ext4_update_i_disksize(inode, (pos + copied));
861 i_size_changed = 1;
862 }
863 unlock_page(page);
864 page_cache_release(page);
865
866 /*
867 * Don't mark the inode dirty under page lock. First, it unnecessarily
868 * makes the holding time of page lock longer. Second, it forces lock
869 * ordering of page lock and transaction start for journaling
870 * filesystems.
871 */
872 if (i_size_changed)
873 ext4_mark_inode_dirty(handle, inode);
874
875 return copied;
876 }
877
878 /*
879 * We need to pick up the new inode size which generic_commit_write gave us
880 * `file' can be NULL - eg, when called from page_symlink().
881 *
882 * ext4 never places buffers on inode->i_mapping->private_list. metadata
883 * buffers are managed internally.
884 */
885 static int ext4_ordered_write_end(struct file *file,
886 struct address_space *mapping,
887 loff_t pos, unsigned len, unsigned copied,
888 struct page *page, void *fsdata)
889 {
890 handle_t *handle = ext4_journal_current_handle();
891 struct inode *inode = mapping->host;
892 int ret = 0, ret2;
893
894 trace_ext4_ordered_write_end(inode, pos, len, copied);
895 ret = ext4_jbd2_file_inode(handle, inode);
896
897 if (ret == 0) {
898 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
899 page, fsdata);
900 copied = ret2;
901 if (pos + len > inode->i_size && ext4_can_truncate(inode))
902 /* if we have allocated more blocks and copied
903 * less. We will have blocks allocated outside
904 * inode->i_size. So truncate them
905 */
906 ext4_orphan_add(handle, inode);
907 if (ret2 < 0)
908 ret = ret2;
909 }
910 ret2 = ext4_journal_stop(handle);
911 if (!ret)
912 ret = ret2;
913
914 if (pos + len > inode->i_size) {
915 ext4_truncate_failed_write(inode);
916 /*
917 * If truncate failed early the inode might still be
918 * on the orphan list; we need to make sure the inode
919 * is removed from the orphan list in that case.
920 */
921 if (inode->i_nlink)
922 ext4_orphan_del(NULL, inode);
923 }
924
925
926 return ret ? ret : copied;
927 }
928
929 static int ext4_writeback_write_end(struct file *file,
930 struct address_space *mapping,
931 loff_t pos, unsigned len, unsigned copied,
932 struct page *page, void *fsdata)
933 {
934 handle_t *handle = ext4_journal_current_handle();
935 struct inode *inode = mapping->host;
936 int ret = 0, ret2;
937
938 trace_ext4_writeback_write_end(inode, pos, len, copied);
939 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
940 page, fsdata);
941 copied = ret2;
942 if (pos + len > inode->i_size && ext4_can_truncate(inode))
943 /* if we have allocated more blocks and copied
944 * less. We will have blocks allocated outside
945 * inode->i_size. So truncate them
946 */
947 ext4_orphan_add(handle, inode);
948
949 if (ret2 < 0)
950 ret = ret2;
951
952 ret2 = ext4_journal_stop(handle);
953 if (!ret)
954 ret = ret2;
955
956 if (pos + len > inode->i_size) {
957 ext4_truncate_failed_write(inode);
958 /*
959 * If truncate failed early the inode might still be
960 * on the orphan list; we need to make sure the inode
961 * is removed from the orphan list in that case.
962 */
963 if (inode->i_nlink)
964 ext4_orphan_del(NULL, inode);
965 }
966
967 return ret ? ret : copied;
968 }
969
970 static int ext4_journalled_write_end(struct file *file,
971 struct address_space *mapping,
972 loff_t pos, unsigned len, unsigned copied,
973 struct page *page, void *fsdata)
974 {
975 handle_t *handle = ext4_journal_current_handle();
976 struct inode *inode = mapping->host;
977 int ret = 0, ret2;
978 int partial = 0;
979 unsigned from, to;
980 loff_t new_i_size;
981
982 trace_ext4_journalled_write_end(inode, pos, len, copied);
983 from = pos & (PAGE_CACHE_SIZE - 1);
984 to = from + len;
985
986 if (copied < len) {
987 if (!PageUptodate(page))
988 copied = 0;
989 page_zero_new_buffers(page, from+copied, to);
990 }
991
992 ret = walk_page_buffers(handle, page_buffers(page), from,
993 to, &partial, write_end_fn);
994 if (!partial)
995 SetPageUptodate(page);
996 new_i_size = pos + copied;
997 if (new_i_size > inode->i_size)
998 i_size_write(inode, pos+copied);
999 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1000 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1001 if (new_i_size > EXT4_I(inode)->i_disksize) {
1002 ext4_update_i_disksize(inode, new_i_size);
1003 ret2 = ext4_mark_inode_dirty(handle, inode);
1004 if (!ret)
1005 ret = ret2;
1006 }
1007
1008 unlock_page(page);
1009 page_cache_release(page);
1010 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1011 /* if we have allocated more blocks and copied
1012 * less. We will have blocks allocated outside
1013 * inode->i_size. So truncate them
1014 */
1015 ext4_orphan_add(handle, inode);
1016
1017 ret2 = ext4_journal_stop(handle);
1018 if (!ret)
1019 ret = ret2;
1020 if (pos + len > inode->i_size) {
1021 ext4_truncate_failed_write(inode);
1022 /*
1023 * If truncate failed early the inode might still be
1024 * on the orphan list; we need to make sure the inode
1025 * is removed from the orphan list in that case.
1026 */
1027 if (inode->i_nlink)
1028 ext4_orphan_del(NULL, inode);
1029 }
1030
1031 return ret ? ret : copied;
1032 }
1033
1034 /*
1035 * Reserve a single block located at lblock
1036 */
1037 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1038 {
1039 int retries = 0;
1040 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1041 struct ext4_inode_info *ei = EXT4_I(inode);
1042 unsigned long md_needed;
1043 int ret;
1044
1045 /*
1046 * recalculate the amount of metadata blocks to reserve
1047 * in order to allocate nrblocks
1048 * worse case is one extent per block
1049 */
1050 repeat:
1051 spin_lock(&ei->i_block_reservation_lock);
1052 md_needed = ext4_calc_metadata_amount(inode, lblock);
1053 trace_ext4_da_reserve_space(inode, md_needed);
1054 spin_unlock(&ei->i_block_reservation_lock);
1055
1056 /*
1057 * We will charge metadata quota at writeout time; this saves
1058 * us from metadata over-estimation, though we may go over by
1059 * a small amount in the end. Here we just reserve for data.
1060 */
1061 ret = dquot_reserve_block(inode, 1);
1062 if (ret)
1063 return ret;
1064 /*
1065 * We do still charge estimated metadata to the sb though;
1066 * we cannot afford to run out of free blocks.
1067 */
1068 if (ext4_claim_free_blocks(sbi, md_needed + 1, 0)) {
1069 dquot_release_reservation_block(inode, 1);
1070 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1071 yield();
1072 goto repeat;
1073 }
1074 return -ENOSPC;
1075 }
1076 spin_lock(&ei->i_block_reservation_lock);
1077 ei->i_reserved_data_blocks++;
1078 ei->i_reserved_meta_blocks += md_needed;
1079 spin_unlock(&ei->i_block_reservation_lock);
1080
1081 return 0; /* success */
1082 }
1083
1084 static void ext4_da_release_space(struct inode *inode, int to_free)
1085 {
1086 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1087 struct ext4_inode_info *ei = EXT4_I(inode);
1088
1089 if (!to_free)
1090 return; /* Nothing to release, exit */
1091
1092 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1093
1094 trace_ext4_da_release_space(inode, to_free);
1095 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1096 /*
1097 * if there aren't enough reserved blocks, then the
1098 * counter is messed up somewhere. Since this
1099 * function is called from invalidate page, it's
1100 * harmless to return without any action.
1101 */
1102 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1103 "ino %lu, to_free %d with only %d reserved "
1104 "data blocks\n", inode->i_ino, to_free,
1105 ei->i_reserved_data_blocks);
1106 WARN_ON(1);
1107 to_free = ei->i_reserved_data_blocks;
1108 }
1109 ei->i_reserved_data_blocks -= to_free;
1110
1111 if (ei->i_reserved_data_blocks == 0) {
1112 /*
1113 * We can release all of the reserved metadata blocks
1114 * only when we have written all of the delayed
1115 * allocation blocks.
1116 */
1117 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1118 ei->i_reserved_meta_blocks);
1119 ei->i_reserved_meta_blocks = 0;
1120 ei->i_da_metadata_calc_len = 0;
1121 }
1122
1123 /* update fs dirty data blocks counter */
1124 percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1125
1126 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1127
1128 dquot_release_reservation_block(inode, to_free);
1129 }
1130
1131 static void ext4_da_page_release_reservation(struct page *page,
1132 unsigned long offset)
1133 {
1134 int to_release = 0;
1135 struct buffer_head *head, *bh;
1136 unsigned int curr_off = 0;
1137
1138 head = page_buffers(page);
1139 bh = head;
1140 do {
1141 unsigned int next_off = curr_off + bh->b_size;
1142
1143 if ((offset <= curr_off) && (buffer_delay(bh))) {
1144 to_release++;
1145 clear_buffer_delay(bh);
1146 }
1147 curr_off = next_off;
1148 } while ((bh = bh->b_this_page) != head);
1149 ext4_da_release_space(page->mapping->host, to_release);
1150 }
1151
1152 /*
1153 * Delayed allocation stuff
1154 */
1155
1156 /*
1157 * mpage_da_submit_io - walks through extent of pages and try to write
1158 * them with writepage() call back
1159 *
1160 * @mpd->inode: inode
1161 * @mpd->first_page: first page of the extent
1162 * @mpd->next_page: page after the last page of the extent
1163 *
1164 * By the time mpage_da_submit_io() is called we expect all blocks
1165 * to be allocated. this may be wrong if allocation failed.
1166 *
1167 * As pages are already locked by write_cache_pages(), we can't use it
1168 */
1169 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1170 struct ext4_map_blocks *map)
1171 {
1172 struct pagevec pvec;
1173 unsigned long index, end;
1174 int ret = 0, err, nr_pages, i;
1175 struct inode *inode = mpd->inode;
1176 struct address_space *mapping = inode->i_mapping;
1177 loff_t size = i_size_read(inode);
1178 unsigned int len, block_start;
1179 struct buffer_head *bh, *page_bufs = NULL;
1180 int journal_data = ext4_should_journal_data(inode);
1181 sector_t pblock = 0, cur_logical = 0;
1182 struct ext4_io_submit io_submit;
1183
1184 BUG_ON(mpd->next_page <= mpd->first_page);
1185 memset(&io_submit, 0, sizeof(io_submit));
1186 /*
1187 * We need to start from the first_page to the next_page - 1
1188 * to make sure we also write the mapped dirty buffer_heads.
1189 * If we look at mpd->b_blocknr we would only be looking
1190 * at the currently mapped buffer_heads.
1191 */
1192 index = mpd->first_page;
1193 end = mpd->next_page - 1;
1194
1195 pagevec_init(&pvec, 0);
1196 while (index <= end) {
1197 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1198 if (nr_pages == 0)
1199 break;
1200 for (i = 0; i < nr_pages; i++) {
1201 int commit_write = 0, skip_page = 0;
1202 struct page *page = pvec.pages[i];
1203
1204 index = page->index;
1205 if (index > end)
1206 break;
1207
1208 if (index == size >> PAGE_CACHE_SHIFT)
1209 len = size & ~PAGE_CACHE_MASK;
1210 else
1211 len = PAGE_CACHE_SIZE;
1212 if (map) {
1213 cur_logical = index << (PAGE_CACHE_SHIFT -
1214 inode->i_blkbits);
1215 pblock = map->m_pblk + (cur_logical -
1216 map->m_lblk);
1217 }
1218 index++;
1219
1220 BUG_ON(!PageLocked(page));
1221 BUG_ON(PageWriteback(page));
1222
1223 /*
1224 * If the page does not have buffers (for
1225 * whatever reason), try to create them using
1226 * __block_write_begin. If this fails,
1227 * skip the page and move on.
1228 */
1229 if (!page_has_buffers(page)) {
1230 if (__block_write_begin(page, 0, len,
1231 noalloc_get_block_write)) {
1232 skip_page:
1233 unlock_page(page);
1234 continue;
1235 }
1236 commit_write = 1;
1237 }
1238
1239 bh = page_bufs = page_buffers(page);
1240 block_start = 0;
1241 do {
1242 if (!bh)
1243 goto skip_page;
1244 if (map && (cur_logical >= map->m_lblk) &&
1245 (cur_logical <= (map->m_lblk +
1246 (map->m_len - 1)))) {
1247 if (buffer_delay(bh)) {
1248 clear_buffer_delay(bh);
1249 bh->b_blocknr = pblock;
1250 }
1251 if (buffer_unwritten(bh) ||
1252 buffer_mapped(bh))
1253 BUG_ON(bh->b_blocknr != pblock);
1254 if (map->m_flags & EXT4_MAP_UNINIT)
1255 set_buffer_uninit(bh);
1256 clear_buffer_unwritten(bh);
1257 }
1258
1259 /* skip page if block allocation undone */
1260 if (buffer_delay(bh) || buffer_unwritten(bh))
1261 skip_page = 1;
1262 bh = bh->b_this_page;
1263 block_start += bh->b_size;
1264 cur_logical++;
1265 pblock++;
1266 } while (bh != page_bufs);
1267
1268 if (skip_page)
1269 goto skip_page;
1270
1271 if (commit_write)
1272 /* mark the buffer_heads as dirty & uptodate */
1273 block_commit_write(page, 0, len);
1274
1275 clear_page_dirty_for_io(page);
1276 /*
1277 * Delalloc doesn't support data journalling,
1278 * but eventually maybe we'll lift this
1279 * restriction.
1280 */
1281 if (unlikely(journal_data && PageChecked(page)))
1282 err = __ext4_journalled_writepage(page, len);
1283 else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1284 err = ext4_bio_write_page(&io_submit, page,
1285 len, mpd->wbc);
1286 else
1287 err = block_write_full_page(page,
1288 noalloc_get_block_write, mpd->wbc);
1289
1290 if (!err)
1291 mpd->pages_written++;
1292 /*
1293 * In error case, we have to continue because
1294 * remaining pages are still locked
1295 */
1296 if (ret == 0)
1297 ret = err;
1298 }
1299 pagevec_release(&pvec);
1300 }
1301 ext4_io_submit(&io_submit);
1302 return ret;
1303 }
1304
1305 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1306 {
1307 int nr_pages, i;
1308 pgoff_t index, end;
1309 struct pagevec pvec;
1310 struct inode *inode = mpd->inode;
1311 struct address_space *mapping = inode->i_mapping;
1312
1313 index = mpd->first_page;
1314 end = mpd->next_page - 1;
1315 while (index <= end) {
1316 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1317 if (nr_pages == 0)
1318 break;
1319 for (i = 0; i < nr_pages; i++) {
1320 struct page *page = pvec.pages[i];
1321 if (page->index > end)
1322 break;
1323 BUG_ON(!PageLocked(page));
1324 BUG_ON(PageWriteback(page));
1325 block_invalidatepage(page, 0);
1326 ClearPageUptodate(page);
1327 unlock_page(page);
1328 }
1329 index = pvec.pages[nr_pages - 1]->index + 1;
1330 pagevec_release(&pvec);
1331 }
1332 return;
1333 }
1334
1335 static void ext4_print_free_blocks(struct inode *inode)
1336 {
1337 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1338 printk(KERN_CRIT "Total free blocks count %lld\n",
1339 ext4_count_free_blocks(inode->i_sb));
1340 printk(KERN_CRIT "Free/Dirty block details\n");
1341 printk(KERN_CRIT "free_blocks=%lld\n",
1342 (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
1343 printk(KERN_CRIT "dirty_blocks=%lld\n",
1344 (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1345 printk(KERN_CRIT "Block reservation details\n");
1346 printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1347 EXT4_I(inode)->i_reserved_data_blocks);
1348 printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1349 EXT4_I(inode)->i_reserved_meta_blocks);
1350 return;
1351 }
1352
1353 /*
1354 * mpage_da_map_and_submit - go through given space, map them
1355 * if necessary, and then submit them for I/O
1356 *
1357 * @mpd - bh describing space
1358 *
1359 * The function skips space we know is already mapped to disk blocks.
1360 *
1361 */
1362 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1363 {
1364 int err, blks, get_blocks_flags;
1365 struct ext4_map_blocks map, *mapp = NULL;
1366 sector_t next = mpd->b_blocknr;
1367 unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1368 loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1369 handle_t *handle = NULL;
1370
1371 /*
1372 * If the blocks are mapped already, or we couldn't accumulate
1373 * any blocks, then proceed immediately to the submission stage.
1374 */
1375 if ((mpd->b_size == 0) ||
1376 ((mpd->b_state & (1 << BH_Mapped)) &&
1377 !(mpd->b_state & (1 << BH_Delay)) &&
1378 !(mpd->b_state & (1 << BH_Unwritten))))
1379 goto submit_io;
1380
1381 handle = ext4_journal_current_handle();
1382 BUG_ON(!handle);
1383
1384 /*
1385 * Call ext4_map_blocks() to allocate any delayed allocation
1386 * blocks, or to convert an uninitialized extent to be
1387 * initialized (in the case where we have written into
1388 * one or more preallocated blocks).
1389 *
1390 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1391 * indicate that we are on the delayed allocation path. This
1392 * affects functions in many different parts of the allocation
1393 * call path. This flag exists primarily because we don't
1394 * want to change *many* call functions, so ext4_map_blocks()
1395 * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1396 * inode's allocation semaphore is taken.
1397 *
1398 * If the blocks in questions were delalloc blocks, set
1399 * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1400 * variables are updated after the blocks have been allocated.
1401 */
1402 map.m_lblk = next;
1403 map.m_len = max_blocks;
1404 get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1405 if (ext4_should_dioread_nolock(mpd->inode))
1406 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1407 if (mpd->b_state & (1 << BH_Delay))
1408 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1409
1410 blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1411 if (blks < 0) {
1412 struct super_block *sb = mpd->inode->i_sb;
1413
1414 err = blks;
1415 /*
1416 * If get block returns EAGAIN or ENOSPC and there
1417 * appears to be free blocks we will just let
1418 * mpage_da_submit_io() unlock all of the pages.
1419 */
1420 if (err == -EAGAIN)
1421 goto submit_io;
1422
1423 if (err == -ENOSPC &&
1424 ext4_count_free_blocks(sb)) {
1425 mpd->retval = err;
1426 goto submit_io;
1427 }
1428
1429 /*
1430 * get block failure will cause us to loop in
1431 * writepages, because a_ops->writepage won't be able
1432 * to make progress. The page will be redirtied by
1433 * writepage and writepages will again try to write
1434 * the same.
1435 */
1436 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1437 ext4_msg(sb, KERN_CRIT,
1438 "delayed block allocation failed for inode %lu "
1439 "at logical offset %llu with max blocks %zd "
1440 "with error %d", mpd->inode->i_ino,
1441 (unsigned long long) next,
1442 mpd->b_size >> mpd->inode->i_blkbits, err);
1443 ext4_msg(sb, KERN_CRIT,
1444 "This should not happen!! Data will be lost\n");
1445 if (err == -ENOSPC)
1446 ext4_print_free_blocks(mpd->inode);
1447 }
1448 /* invalidate all the pages */
1449 ext4_da_block_invalidatepages(mpd);
1450
1451 /* Mark this page range as having been completed */
1452 mpd->io_done = 1;
1453 return;
1454 }
1455 BUG_ON(blks == 0);
1456
1457 mapp = &map;
1458 if (map.m_flags & EXT4_MAP_NEW) {
1459 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1460 int i;
1461
1462 for (i = 0; i < map.m_len; i++)
1463 unmap_underlying_metadata(bdev, map.m_pblk + i);
1464 }
1465
1466 if (ext4_should_order_data(mpd->inode)) {
1467 err = ext4_jbd2_file_inode(handle, mpd->inode);
1468 if (err)
1469 /* This only happens if the journal is aborted */
1470 return;
1471 }
1472
1473 /*
1474 * Update on-disk size along with block allocation.
1475 */
1476 disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1477 if (disksize > i_size_read(mpd->inode))
1478 disksize = i_size_read(mpd->inode);
1479 if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1480 ext4_update_i_disksize(mpd->inode, disksize);
1481 err = ext4_mark_inode_dirty(handle, mpd->inode);
1482 if (err)
1483 ext4_error(mpd->inode->i_sb,
1484 "Failed to mark inode %lu dirty",
1485 mpd->inode->i_ino);
1486 }
1487
1488 submit_io:
1489 mpage_da_submit_io(mpd, mapp);
1490 mpd->io_done = 1;
1491 }
1492
1493 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1494 (1 << BH_Delay) | (1 << BH_Unwritten))
1495
1496 /*
1497 * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1498 *
1499 * @mpd->lbh - extent of blocks
1500 * @logical - logical number of the block in the file
1501 * @bh - bh of the block (used to access block's state)
1502 *
1503 * the function is used to collect contig. blocks in same state
1504 */
1505 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1506 sector_t logical, size_t b_size,
1507 unsigned long b_state)
1508 {
1509 sector_t next;
1510 int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1511
1512 /*
1513 * XXX Don't go larger than mballoc is willing to allocate
1514 * This is a stopgap solution. We eventually need to fold
1515 * mpage_da_submit_io() into this function and then call
1516 * ext4_map_blocks() multiple times in a loop
1517 */
1518 if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1519 goto flush_it;
1520
1521 /* check if thereserved journal credits might overflow */
1522 if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1523 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1524 /*
1525 * With non-extent format we are limited by the journal
1526 * credit available. Total credit needed to insert
1527 * nrblocks contiguous blocks is dependent on the
1528 * nrblocks. So limit nrblocks.
1529 */
1530 goto flush_it;
1531 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1532 EXT4_MAX_TRANS_DATA) {
1533 /*
1534 * Adding the new buffer_head would make it cross the
1535 * allowed limit for which we have journal credit
1536 * reserved. So limit the new bh->b_size
1537 */
1538 b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1539 mpd->inode->i_blkbits;
1540 /* we will do mpage_da_submit_io in the next loop */
1541 }
1542 }
1543 /*
1544 * First block in the extent
1545 */
1546 if (mpd->b_size == 0) {
1547 mpd->b_blocknr = logical;
1548 mpd->b_size = b_size;
1549 mpd->b_state = b_state & BH_FLAGS;
1550 return;
1551 }
1552
1553 next = mpd->b_blocknr + nrblocks;
1554 /*
1555 * Can we merge the block to our big extent?
1556 */
1557 if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1558 mpd->b_size += b_size;
1559 return;
1560 }
1561
1562 flush_it:
1563 /*
1564 * We couldn't merge the block to our extent, so we
1565 * need to flush current extent and start new one
1566 */
1567 mpage_da_map_and_submit(mpd);
1568 return;
1569 }
1570
1571 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1572 {
1573 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1574 }
1575
1576 /*
1577 * This is a special get_blocks_t callback which is used by
1578 * ext4_da_write_begin(). It will either return mapped block or
1579 * reserve space for a single block.
1580 *
1581 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1582 * We also have b_blocknr = -1 and b_bdev initialized properly
1583 *
1584 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1585 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1586 * initialized properly.
1587 */
1588 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1589 struct buffer_head *bh, int create)
1590 {
1591 struct ext4_map_blocks map;
1592 int ret = 0;
1593 sector_t invalid_block = ~((sector_t) 0xffff);
1594
1595 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1596 invalid_block = ~0;
1597
1598 BUG_ON(create == 0);
1599 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1600
1601 map.m_lblk = iblock;
1602 map.m_len = 1;
1603
1604 /*
1605 * first, we need to know whether the block is allocated already
1606 * preallocated blocks are unmapped but should treated
1607 * the same as allocated blocks.
1608 */
1609 ret = ext4_map_blocks(NULL, inode, &map, 0);
1610 if (ret < 0)
1611 return ret;
1612 if (ret == 0) {
1613 if (buffer_delay(bh))
1614 return 0; /* Not sure this could or should happen */
1615 /*
1616 * XXX: __block_write_begin() unmaps passed block, is it OK?
1617 */
1618 ret = ext4_da_reserve_space(inode, iblock);
1619 if (ret)
1620 /* not enough space to reserve */
1621 return ret;
1622
1623 map_bh(bh, inode->i_sb, invalid_block);
1624 set_buffer_new(bh);
1625 set_buffer_delay(bh);
1626 return 0;
1627 }
1628
1629 map_bh(bh, inode->i_sb, map.m_pblk);
1630 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1631
1632 if (buffer_unwritten(bh)) {
1633 /* A delayed write to unwritten bh should be marked
1634 * new and mapped. Mapped ensures that we don't do
1635 * get_block multiple times when we write to the same
1636 * offset and new ensures that we do proper zero out
1637 * for partial write.
1638 */
1639 set_buffer_new(bh);
1640 set_buffer_mapped(bh);
1641 }
1642 return 0;
1643 }
1644
1645 /*
1646 * This function is used as a standard get_block_t calback function
1647 * when there is no desire to allocate any blocks. It is used as a
1648 * callback function for block_write_begin() and block_write_full_page().
1649 * These functions should only try to map a single block at a time.
1650 *
1651 * Since this function doesn't do block allocations even if the caller
1652 * requests it by passing in create=1, it is critically important that
1653 * any caller checks to make sure that any buffer heads are returned
1654 * by this function are either all already mapped or marked for
1655 * delayed allocation before calling block_write_full_page(). Otherwise,
1656 * b_blocknr could be left unitialized, and the page write functions will
1657 * be taken by surprise.
1658 */
1659 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1660 struct buffer_head *bh_result, int create)
1661 {
1662 BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1663 return _ext4_get_block(inode, iblock, bh_result, 0);
1664 }
1665
1666 static int bget_one(handle_t *handle, struct buffer_head *bh)
1667 {
1668 get_bh(bh);
1669 return 0;
1670 }
1671
1672 static int bput_one(handle_t *handle, struct buffer_head *bh)
1673 {
1674 put_bh(bh);
1675 return 0;
1676 }
1677
1678 static int __ext4_journalled_writepage(struct page *page,
1679 unsigned int len)
1680 {
1681 struct address_space *mapping = page->mapping;
1682 struct inode *inode = mapping->host;
1683 struct buffer_head *page_bufs;
1684 handle_t *handle = NULL;
1685 int ret = 0;
1686 int err;
1687
1688 ClearPageChecked(page);
1689 page_bufs = page_buffers(page);
1690 BUG_ON(!page_bufs);
1691 walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1692 /* As soon as we unlock the page, it can go away, but we have
1693 * references to buffers so we are safe */
1694 unlock_page(page);
1695
1696 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1697 if (IS_ERR(handle)) {
1698 ret = PTR_ERR(handle);
1699 goto out;
1700 }
1701
1702 ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1703 do_journal_get_write_access);
1704
1705 err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1706 write_end_fn);
1707 if (ret == 0)
1708 ret = err;
1709 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1710 err = ext4_journal_stop(handle);
1711 if (!ret)
1712 ret = err;
1713
1714 walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1715 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1716 out:
1717 return ret;
1718 }
1719
1720 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1721 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1722
1723 /*
1724 * Note that we don't need to start a transaction unless we're journaling data
1725 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1726 * need to file the inode to the transaction's list in ordered mode because if
1727 * we are writing back data added by write(), the inode is already there and if
1728 * we are writing back data modified via mmap(), no one guarantees in which
1729 * transaction the data will hit the disk. In case we are journaling data, we
1730 * cannot start transaction directly because transaction start ranks above page
1731 * lock so we have to do some magic.
1732 *
1733 * This function can get called via...
1734 * - ext4_da_writepages after taking page lock (have journal handle)
1735 * - journal_submit_inode_data_buffers (no journal handle)
1736 * - shrink_page_list via pdflush (no journal handle)
1737 * - grab_page_cache when doing write_begin (have journal handle)
1738 *
1739 * We don't do any block allocation in this function. If we have page with
1740 * multiple blocks we need to write those buffer_heads that are mapped. This
1741 * is important for mmaped based write. So if we do with blocksize 1K
1742 * truncate(f, 1024);
1743 * a = mmap(f, 0, 4096);
1744 * a[0] = 'a';
1745 * truncate(f, 4096);
1746 * we have in the page first buffer_head mapped via page_mkwrite call back
1747 * but other bufer_heads would be unmapped but dirty(dirty done via the
1748 * do_wp_page). So writepage should write the first block. If we modify
1749 * the mmap area beyond 1024 we will again get a page_fault and the
1750 * page_mkwrite callback will do the block allocation and mark the
1751 * buffer_heads mapped.
1752 *
1753 * We redirty the page if we have any buffer_heads that is either delay or
1754 * unwritten in the page.
1755 *
1756 * We can get recursively called as show below.
1757 *
1758 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1759 * ext4_writepage()
1760 *
1761 * But since we don't do any block allocation we should not deadlock.
1762 * Page also have the dirty flag cleared so we don't get recurive page_lock.
1763 */
1764 static int ext4_writepage(struct page *page,
1765 struct writeback_control *wbc)
1766 {
1767 int ret = 0, commit_write = 0;
1768 loff_t size;
1769 unsigned int len;
1770 struct buffer_head *page_bufs = NULL;
1771 struct inode *inode = page->mapping->host;
1772
1773 trace_ext4_writepage(page);
1774 size = i_size_read(inode);
1775 if (page->index == size >> PAGE_CACHE_SHIFT)
1776 len = size & ~PAGE_CACHE_MASK;
1777 else
1778 len = PAGE_CACHE_SIZE;
1779
1780 /*
1781 * If the page does not have buffers (for whatever reason),
1782 * try to create them using __block_write_begin. If this
1783 * fails, redirty the page and move on.
1784 */
1785 if (!page_has_buffers(page)) {
1786 if (__block_write_begin(page, 0, len,
1787 noalloc_get_block_write)) {
1788 redirty_page:
1789 redirty_page_for_writepage(wbc, page);
1790 unlock_page(page);
1791 return 0;
1792 }
1793 commit_write = 1;
1794 }
1795 page_bufs = page_buffers(page);
1796 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1797 ext4_bh_delay_or_unwritten)) {
1798 /*
1799 * We don't want to do block allocation, so redirty
1800 * the page and return. We may reach here when we do
1801 * a journal commit via journal_submit_inode_data_buffers.
1802 * We can also reach here via shrink_page_list
1803 */
1804 goto redirty_page;
1805 }
1806 if (commit_write)
1807 /* now mark the buffer_heads as dirty and uptodate */
1808 block_commit_write(page, 0, len);
1809
1810 if (PageChecked(page) && ext4_should_journal_data(inode))
1811 /*
1812 * It's mmapped pagecache. Add buffers and journal it. There
1813 * doesn't seem much point in redirtying the page here.
1814 */
1815 return __ext4_journalled_writepage(page, len);
1816
1817 if (buffer_uninit(page_bufs)) {
1818 ext4_set_bh_endio(page_bufs, inode);
1819 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1820 wbc, ext4_end_io_buffer_write);
1821 } else
1822 ret = block_write_full_page(page, noalloc_get_block_write,
1823 wbc);
1824
1825 return ret;
1826 }
1827
1828 /*
1829 * This is called via ext4_da_writepages() to
1830 * calculate the total number of credits to reserve to fit
1831 * a single extent allocation into a single transaction,
1832 * ext4_da_writpeages() will loop calling this before
1833 * the block allocation.
1834 */
1835
1836 static int ext4_da_writepages_trans_blocks(struct inode *inode)
1837 {
1838 int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1839
1840 /*
1841 * With non-extent format the journal credit needed to
1842 * insert nrblocks contiguous block is dependent on
1843 * number of contiguous block. So we will limit
1844 * number of contiguous block to a sane value
1845 */
1846 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1847 (max_blocks > EXT4_MAX_TRANS_DATA))
1848 max_blocks = EXT4_MAX_TRANS_DATA;
1849
1850 return ext4_chunk_trans_blocks(inode, max_blocks);
1851 }
1852
1853 /*
1854 * write_cache_pages_da - walk the list of dirty pages of the given
1855 * address space and accumulate pages that need writing, and call
1856 * mpage_da_map_and_submit to map a single contiguous memory region
1857 * and then write them.
1858 */
1859 static int write_cache_pages_da(struct address_space *mapping,
1860 struct writeback_control *wbc,
1861 struct mpage_da_data *mpd,
1862 pgoff_t *done_index)
1863 {
1864 struct buffer_head *bh, *head;
1865 struct inode *inode = mapping->host;
1866 struct pagevec pvec;
1867 unsigned int nr_pages;
1868 sector_t logical;
1869 pgoff_t index, end;
1870 long nr_to_write = wbc->nr_to_write;
1871 int i, tag, ret = 0;
1872
1873 memset(mpd, 0, sizeof(struct mpage_da_data));
1874 mpd->wbc = wbc;
1875 mpd->inode = inode;
1876 pagevec_init(&pvec, 0);
1877 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1878 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1879
1880 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1881 tag = PAGECACHE_TAG_TOWRITE;
1882 else
1883 tag = PAGECACHE_TAG_DIRTY;
1884
1885 *done_index = index;
1886 while (index <= end) {
1887 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1888 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1889 if (nr_pages == 0)
1890 return 0;
1891
1892 for (i = 0; i < nr_pages; i++) {
1893 struct page *page = pvec.pages[i];
1894
1895 /*
1896 * At this point, the page may be truncated or
1897 * invalidated (changing page->mapping to NULL), or
1898 * even swizzled back from swapper_space to tmpfs file
1899 * mapping. However, page->index will not change
1900 * because we have a reference on the page.
1901 */
1902 if (page->index > end)
1903 goto out;
1904
1905 *done_index = page->index + 1;
1906
1907 /*
1908 * If we can't merge this page, and we have
1909 * accumulated an contiguous region, write it
1910 */
1911 if ((mpd->next_page != page->index) &&
1912 (mpd->next_page != mpd->first_page)) {
1913 mpage_da_map_and_submit(mpd);
1914 goto ret_extent_tail;
1915 }
1916
1917 lock_page(page);
1918
1919 /*
1920 * If the page is no longer dirty, or its
1921 * mapping no longer corresponds to inode we
1922 * are writing (which means it has been
1923 * truncated or invalidated), or the page is
1924 * already under writeback and we are not
1925 * doing a data integrity writeback, skip the page
1926 */
1927 if (!PageDirty(page) ||
1928 (PageWriteback(page) &&
1929 (wbc->sync_mode == WB_SYNC_NONE)) ||
1930 unlikely(page->mapping != mapping)) {
1931 unlock_page(page);
1932 continue;
1933 }
1934
1935 wait_on_page_writeback(page);
1936 BUG_ON(PageWriteback(page));
1937
1938 if (mpd->next_page != page->index)
1939 mpd->first_page = page->index;
1940 mpd->next_page = page->index + 1;
1941 logical = (sector_t) page->index <<
1942 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1943
1944 if (!page_has_buffers(page)) {
1945 mpage_add_bh_to_extent(mpd, logical,
1946 PAGE_CACHE_SIZE,
1947 (1 << BH_Dirty) | (1 << BH_Uptodate));
1948 if (mpd->io_done)
1949 goto ret_extent_tail;
1950 } else {
1951 /*
1952 * Page with regular buffer heads,
1953 * just add all dirty ones
1954 */
1955 head = page_buffers(page);
1956 bh = head;
1957 do {
1958 BUG_ON(buffer_locked(bh));
1959 /*
1960 * We need to try to allocate
1961 * unmapped blocks in the same page.
1962 * Otherwise we won't make progress
1963 * with the page in ext4_writepage
1964 */
1965 if (ext4_bh_delay_or_unwritten(NULL, bh)) {
1966 mpage_add_bh_to_extent(mpd, logical,
1967 bh->b_size,
1968 bh->b_state);
1969 if (mpd->io_done)
1970 goto ret_extent_tail;
1971 } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
1972 /*
1973 * mapped dirty buffer. We need
1974 * to update the b_state
1975 * because we look at b_state
1976 * in mpage_da_map_blocks. We
1977 * don't update b_size because
1978 * if we find an unmapped
1979 * buffer_head later we need to
1980 * use the b_state flag of that
1981 * buffer_head.
1982 */
1983 if (mpd->b_size == 0)
1984 mpd->b_state = bh->b_state & BH_FLAGS;
1985 }
1986 logical++;
1987 } while ((bh = bh->b_this_page) != head);
1988 }
1989
1990 if (nr_to_write > 0) {
1991 nr_to_write--;
1992 if (nr_to_write == 0 &&
1993 wbc->sync_mode == WB_SYNC_NONE)
1994 /*
1995 * We stop writing back only if we are
1996 * not doing integrity sync. In case of
1997 * integrity sync we have to keep going
1998 * because someone may be concurrently
1999 * dirtying pages, and we might have
2000 * synced a lot of newly appeared dirty
2001 * pages, but have not synced all of the
2002 * old dirty pages.
2003 */
2004 goto out;
2005 }
2006 }
2007 pagevec_release(&pvec);
2008 cond_resched();
2009 }
2010 return 0;
2011 ret_extent_tail:
2012 ret = MPAGE_DA_EXTENT_TAIL;
2013 out:
2014 pagevec_release(&pvec);
2015 cond_resched();
2016 return ret;
2017 }
2018
2019
2020 static int ext4_da_writepages(struct address_space *mapping,
2021 struct writeback_control *wbc)
2022 {
2023 pgoff_t index;
2024 int range_whole = 0;
2025 handle_t *handle = NULL;
2026 struct mpage_da_data mpd;
2027 struct inode *inode = mapping->host;
2028 int pages_written = 0;
2029 unsigned int max_pages;
2030 int range_cyclic, cycled = 1, io_done = 0;
2031 int needed_blocks, ret = 0;
2032 long desired_nr_to_write, nr_to_writebump = 0;
2033 loff_t range_start = wbc->range_start;
2034 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2035 pgoff_t done_index = 0;
2036 pgoff_t end;
2037
2038 trace_ext4_da_writepages(inode, wbc);
2039
2040 /*
2041 * No pages to write? This is mainly a kludge to avoid starting
2042 * a transaction for special inodes like journal inode on last iput()
2043 * because that could violate lock ordering on umount
2044 */
2045 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2046 return 0;
2047
2048 /*
2049 * If the filesystem has aborted, it is read-only, so return
2050 * right away instead of dumping stack traces later on that
2051 * will obscure the real source of the problem. We test
2052 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2053 * the latter could be true if the filesystem is mounted
2054 * read-only, and in that case, ext4_da_writepages should
2055 * *never* be called, so if that ever happens, we would want
2056 * the stack trace.
2057 */
2058 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2059 return -EROFS;
2060
2061 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2062 range_whole = 1;
2063
2064 range_cyclic = wbc->range_cyclic;
2065 if (wbc->range_cyclic) {
2066 index = mapping->writeback_index;
2067 if (index)
2068 cycled = 0;
2069 wbc->range_start = index << PAGE_CACHE_SHIFT;
2070 wbc->range_end = LLONG_MAX;
2071 wbc->range_cyclic = 0;
2072 end = -1;
2073 } else {
2074 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2075 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2076 }
2077
2078 /*
2079 * This works around two forms of stupidity. The first is in
2080 * the writeback code, which caps the maximum number of pages
2081 * written to be 1024 pages. This is wrong on multiple
2082 * levels; different architectues have a different page size,
2083 * which changes the maximum amount of data which gets
2084 * written. Secondly, 4 megabytes is way too small. XFS
2085 * forces this value to be 16 megabytes by multiplying
2086 * nr_to_write parameter by four, and then relies on its
2087 * allocator to allocate larger extents to make them
2088 * contiguous. Unfortunately this brings us to the second
2089 * stupidity, which is that ext4's mballoc code only allocates
2090 * at most 2048 blocks. So we force contiguous writes up to
2091 * the number of dirty blocks in the inode, or
2092 * sbi->max_writeback_mb_bump whichever is smaller.
2093 */
2094 max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2095 if (!range_cyclic && range_whole) {
2096 if (wbc->nr_to_write == LONG_MAX)
2097 desired_nr_to_write = wbc->nr_to_write;
2098 else
2099 desired_nr_to_write = wbc->nr_to_write * 8;
2100 } else
2101 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2102 max_pages);
2103 if (desired_nr_to_write > max_pages)
2104 desired_nr_to_write = max_pages;
2105
2106 if (wbc->nr_to_write < desired_nr_to_write) {
2107 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2108 wbc->nr_to_write = desired_nr_to_write;
2109 }
2110
2111 retry:
2112 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2113 tag_pages_for_writeback(mapping, index, end);
2114
2115 while (!ret && wbc->nr_to_write > 0) {
2116
2117 /*
2118 * we insert one extent at a time. So we need
2119 * credit needed for single extent allocation.
2120 * journalled mode is currently not supported
2121 * by delalloc
2122 */
2123 BUG_ON(ext4_should_journal_data(inode));
2124 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2125
2126 /* start a new transaction*/
2127 handle = ext4_journal_start(inode, needed_blocks);
2128 if (IS_ERR(handle)) {
2129 ret = PTR_ERR(handle);
2130 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2131 "%ld pages, ino %lu; err %d", __func__,
2132 wbc->nr_to_write, inode->i_ino, ret);
2133 goto out_writepages;
2134 }
2135
2136 /*
2137 * Now call write_cache_pages_da() to find the next
2138 * contiguous region of logical blocks that need
2139 * blocks to be allocated by ext4 and submit them.
2140 */
2141 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2142 /*
2143 * If we have a contiguous extent of pages and we
2144 * haven't done the I/O yet, map the blocks and submit
2145 * them for I/O.
2146 */
2147 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2148 mpage_da_map_and_submit(&mpd);
2149 ret = MPAGE_DA_EXTENT_TAIL;
2150 }
2151 trace_ext4_da_write_pages(inode, &mpd);
2152 wbc->nr_to_write -= mpd.pages_written;
2153
2154 ext4_journal_stop(handle);
2155
2156 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2157 /* commit the transaction which would
2158 * free blocks released in the transaction
2159 * and try again
2160 */
2161 jbd2_journal_force_commit_nested(sbi->s_journal);
2162 ret = 0;
2163 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2164 /*
2165 * got one extent now try with
2166 * rest of the pages
2167 */
2168 pages_written += mpd.pages_written;
2169 ret = 0;
2170 io_done = 1;
2171 } else if (wbc->nr_to_write)
2172 /*
2173 * There is no more writeout needed
2174 * or we requested for a noblocking writeout
2175 * and we found the device congested
2176 */
2177 break;
2178 }
2179 if (!io_done && !cycled) {
2180 cycled = 1;
2181 index = 0;
2182 wbc->range_start = index << PAGE_CACHE_SHIFT;
2183 wbc->range_end = mapping->writeback_index - 1;
2184 goto retry;
2185 }
2186
2187 /* Update index */
2188 wbc->range_cyclic = range_cyclic;
2189 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2190 /*
2191 * set the writeback_index so that range_cyclic
2192 * mode will write it back later
2193 */
2194 mapping->writeback_index = done_index;
2195
2196 out_writepages:
2197 wbc->nr_to_write -= nr_to_writebump;
2198 wbc->range_start = range_start;
2199 trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2200 return ret;
2201 }
2202
2203 #define FALL_BACK_TO_NONDELALLOC 1
2204 static int ext4_nonda_switch(struct super_block *sb)
2205 {
2206 s64 free_blocks, dirty_blocks;
2207 struct ext4_sb_info *sbi = EXT4_SB(sb);
2208
2209 /*
2210 * switch to non delalloc mode if we are running low
2211 * on free block. The free block accounting via percpu
2212 * counters can get slightly wrong with percpu_counter_batch getting
2213 * accumulated on each CPU without updating global counters
2214 * Delalloc need an accurate free block accounting. So switch
2215 * to non delalloc when we are near to error range.
2216 */
2217 free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2218 dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2219 if (2 * free_blocks < 3 * dirty_blocks ||
2220 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2221 /*
2222 * free block count is less than 150% of dirty blocks
2223 * or free blocks is less than watermark
2224 */
2225 return 1;
2226 }
2227 /*
2228 * Even if we don't switch but are nearing capacity,
2229 * start pushing delalloc when 1/2 of free blocks are dirty.
2230 */
2231 if (free_blocks < 2 * dirty_blocks)
2232 writeback_inodes_sb_if_idle(sb);
2233
2234 return 0;
2235 }
2236
2237 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2238 loff_t pos, unsigned len, unsigned flags,
2239 struct page **pagep, void **fsdata)
2240 {
2241 int ret, retries = 0;
2242 struct page *page;
2243 pgoff_t index;
2244 struct inode *inode = mapping->host;
2245 handle_t *handle;
2246
2247 index = pos >> PAGE_CACHE_SHIFT;
2248
2249 if (ext4_nonda_switch(inode->i_sb)) {
2250 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2251 return ext4_write_begin(file, mapping, pos,
2252 len, flags, pagep, fsdata);
2253 }
2254 *fsdata = (void *)0;
2255 trace_ext4_da_write_begin(inode, pos, len, flags);
2256 retry:
2257 /*
2258 * With delayed allocation, we don't log the i_disksize update
2259 * if there is delayed block allocation. But we still need
2260 * to journalling the i_disksize update if writes to the end
2261 * of file which has an already mapped buffer.
2262 */
2263 handle = ext4_journal_start(inode, 1);
2264 if (IS_ERR(handle)) {
2265 ret = PTR_ERR(handle);
2266 goto out;
2267 }
2268 /* We cannot recurse into the filesystem as the transaction is already
2269 * started */
2270 flags |= AOP_FLAG_NOFS;
2271
2272 page = grab_cache_page_write_begin(mapping, index, flags);
2273 if (!page) {
2274 ext4_journal_stop(handle);
2275 ret = -ENOMEM;
2276 goto out;
2277 }
2278 *pagep = page;
2279
2280 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2281 if (ret < 0) {
2282 unlock_page(page);
2283 ext4_journal_stop(handle);
2284 page_cache_release(page);
2285 /*
2286 * block_write_begin may have instantiated a few blocks
2287 * outside i_size. Trim these off again. Don't need
2288 * i_size_read because we hold i_mutex.
2289 */
2290 if (pos + len > inode->i_size)
2291 ext4_truncate_failed_write(inode);
2292 }
2293
2294 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2295 goto retry;
2296 out:
2297 return ret;
2298 }
2299
2300 /*
2301 * Check if we should update i_disksize
2302 * when write to the end of file but not require block allocation
2303 */
2304 static int ext4_da_should_update_i_disksize(struct page *page,
2305 unsigned long offset)
2306 {
2307 struct buffer_head *bh;
2308 struct inode *inode = page->mapping->host;
2309 unsigned int idx;
2310 int i;
2311
2312 bh = page_buffers(page);
2313 idx = offset >> inode->i_blkbits;
2314
2315 for (i = 0; i < idx; i++)
2316 bh = bh->b_this_page;
2317
2318 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2319 return 0;
2320 return 1;
2321 }
2322
2323 static int ext4_da_write_end(struct file *file,
2324 struct address_space *mapping,
2325 loff_t pos, unsigned len, unsigned copied,
2326 struct page *page, void *fsdata)
2327 {
2328 struct inode *inode = mapping->host;
2329 int ret = 0, ret2;
2330 handle_t *handle = ext4_journal_current_handle();
2331 loff_t new_i_size;
2332 unsigned long start, end;
2333 int write_mode = (int)(unsigned long)fsdata;
2334
2335 if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2336 if (ext4_should_order_data(inode)) {
2337 return ext4_ordered_write_end(file, mapping, pos,
2338 len, copied, page, fsdata);
2339 } else if (ext4_should_writeback_data(inode)) {
2340 return ext4_writeback_write_end(file, mapping, pos,
2341 len, copied, page, fsdata);
2342 } else {
2343 BUG();
2344 }
2345 }
2346
2347 trace_ext4_da_write_end(inode, pos, len, copied);
2348 start = pos & (PAGE_CACHE_SIZE - 1);
2349 end = start + copied - 1;
2350
2351 /*
2352 * generic_write_end() will run mark_inode_dirty() if i_size
2353 * changes. So let's piggyback the i_disksize mark_inode_dirty
2354 * into that.
2355 */
2356
2357 new_i_size = pos + copied;
2358 if (new_i_size > EXT4_I(inode)->i_disksize) {
2359 if (ext4_da_should_update_i_disksize(page, end)) {
2360 down_write(&EXT4_I(inode)->i_data_sem);
2361 if (new_i_size > EXT4_I(inode)->i_disksize) {
2362 /*
2363 * Updating i_disksize when extending file
2364 * without needing block allocation
2365 */
2366 if (ext4_should_order_data(inode))
2367 ret = ext4_jbd2_file_inode(handle,
2368 inode);
2369
2370 EXT4_I(inode)->i_disksize = new_i_size;
2371 }
2372 up_write(&EXT4_I(inode)->i_data_sem);
2373 /* We need to mark inode dirty even if
2374 * new_i_size is less that inode->i_size
2375 * bu greater than i_disksize.(hint delalloc)
2376 */
2377 ext4_mark_inode_dirty(handle, inode);
2378 }
2379 }
2380 ret2 = generic_write_end(file, mapping, pos, len, copied,
2381 page, fsdata);
2382 copied = ret2;
2383 if (ret2 < 0)
2384 ret = ret2;
2385 ret2 = ext4_journal_stop(handle);
2386 if (!ret)
2387 ret = ret2;
2388
2389 return ret ? ret : copied;
2390 }
2391
2392 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2393 {
2394 /*
2395 * Drop reserved blocks
2396 */
2397 BUG_ON(!PageLocked(page));
2398 if (!page_has_buffers(page))
2399 goto out;
2400
2401 ext4_da_page_release_reservation(page, offset);
2402
2403 out:
2404 ext4_invalidatepage(page, offset);
2405
2406 return;
2407 }
2408
2409 /*
2410 * Force all delayed allocation blocks to be allocated for a given inode.
2411 */
2412 int ext4_alloc_da_blocks(struct inode *inode)
2413 {
2414 trace_ext4_alloc_da_blocks(inode);
2415
2416 if (!EXT4_I(inode)->i_reserved_data_blocks &&
2417 !EXT4_I(inode)->i_reserved_meta_blocks)
2418 return 0;
2419
2420 /*
2421 * We do something simple for now. The filemap_flush() will
2422 * also start triggering a write of the data blocks, which is
2423 * not strictly speaking necessary (and for users of
2424 * laptop_mode, not even desirable). However, to do otherwise
2425 * would require replicating code paths in:
2426 *
2427 * ext4_da_writepages() ->
2428 * write_cache_pages() ---> (via passed in callback function)
2429 * __mpage_da_writepage() -->
2430 * mpage_add_bh_to_extent()
2431 * mpage_da_map_blocks()
2432 *
2433 * The problem is that write_cache_pages(), located in
2434 * mm/page-writeback.c, marks pages clean in preparation for
2435 * doing I/O, which is not desirable if we're not planning on
2436 * doing I/O at all.
2437 *
2438 * We could call write_cache_pages(), and then redirty all of
2439 * the pages by calling redirty_page_for_writepage() but that
2440 * would be ugly in the extreme. So instead we would need to
2441 * replicate parts of the code in the above functions,
2442 * simplifying them because we wouldn't actually intend to
2443 * write out the pages, but rather only collect contiguous
2444 * logical block extents, call the multi-block allocator, and
2445 * then update the buffer heads with the block allocations.
2446 *
2447 * For now, though, we'll cheat by calling filemap_flush(),
2448 * which will map the blocks, and start the I/O, but not
2449 * actually wait for the I/O to complete.
2450 */
2451 return filemap_flush(inode->i_mapping);
2452 }
2453
2454 /*
2455 * bmap() is special. It gets used by applications such as lilo and by
2456 * the swapper to find the on-disk block of a specific piece of data.
2457 *
2458 * Naturally, this is dangerous if the block concerned is still in the
2459 * journal. If somebody makes a swapfile on an ext4 data-journaling
2460 * filesystem and enables swap, then they may get a nasty shock when the
2461 * data getting swapped to that swapfile suddenly gets overwritten by
2462 * the original zero's written out previously to the journal and
2463 * awaiting writeback in the kernel's buffer cache.
2464 *
2465 * So, if we see any bmap calls here on a modified, data-journaled file,
2466 * take extra steps to flush any blocks which might be in the cache.
2467 */
2468 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2469 {
2470 struct inode *inode = mapping->host;
2471 journal_t *journal;
2472 int err;
2473
2474 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2475 test_opt(inode->i_sb, DELALLOC)) {
2476 /*
2477 * With delalloc we want to sync the file
2478 * so that we can make sure we allocate
2479 * blocks for file
2480 */
2481 filemap_write_and_wait(mapping);
2482 }
2483
2484 if (EXT4_JOURNAL(inode) &&
2485 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2486 /*
2487 * This is a REALLY heavyweight approach, but the use of
2488 * bmap on dirty files is expected to be extremely rare:
2489 * only if we run lilo or swapon on a freshly made file
2490 * do we expect this to happen.
2491 *
2492 * (bmap requires CAP_SYS_RAWIO so this does not
2493 * represent an unprivileged user DOS attack --- we'd be
2494 * in trouble if mortal users could trigger this path at
2495 * will.)
2496 *
2497 * NB. EXT4_STATE_JDATA is not set on files other than
2498 * regular files. If somebody wants to bmap a directory
2499 * or symlink and gets confused because the buffer
2500 * hasn't yet been flushed to disk, they deserve
2501 * everything they get.
2502 */
2503
2504 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2505 journal = EXT4_JOURNAL(inode);
2506 jbd2_journal_lock_updates(journal);
2507 err = jbd2_journal_flush(journal);
2508 jbd2_journal_unlock_updates(journal);
2509
2510 if (err)
2511 return 0;
2512 }
2513
2514 return generic_block_bmap(mapping, block, ext4_get_block);
2515 }
2516
2517 static int ext4_readpage(struct file *file, struct page *page)
2518 {
2519 trace_ext4_readpage(page);
2520 return mpage_readpage(page, ext4_get_block);
2521 }
2522
2523 static int
2524 ext4_readpages(struct file *file, struct address_space *mapping,
2525 struct list_head *pages, unsigned nr_pages)
2526 {
2527 return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2528 }
2529
2530 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2531 {
2532 struct buffer_head *head, *bh;
2533 unsigned int curr_off = 0;
2534
2535 if (!page_has_buffers(page))
2536 return;
2537 head = bh = page_buffers(page);
2538 do {
2539 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2540 && bh->b_private) {
2541 ext4_free_io_end(bh->b_private);
2542 bh->b_private = NULL;
2543 bh->b_end_io = NULL;
2544 }
2545 curr_off = curr_off + bh->b_size;
2546 bh = bh->b_this_page;
2547 } while (bh != head);
2548 }
2549
2550 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2551 {
2552 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2553
2554 trace_ext4_invalidatepage(page, offset);
2555
2556 /*
2557 * free any io_end structure allocated for buffers to be discarded
2558 */
2559 if (ext4_should_dioread_nolock(page->mapping->host))
2560 ext4_invalidatepage_free_endio(page, offset);
2561 /*
2562 * If it's a full truncate we just forget about the pending dirtying
2563 */
2564 if (offset == 0)
2565 ClearPageChecked(page);
2566
2567 if (journal)
2568 jbd2_journal_invalidatepage(journal, page, offset);
2569 else
2570 block_invalidatepage(page, offset);
2571 }
2572
2573 static int ext4_releasepage(struct page *page, gfp_t wait)
2574 {
2575 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2576
2577 trace_ext4_releasepage(page);
2578
2579 WARN_ON(PageChecked(page));
2580 if (!page_has_buffers(page))
2581 return 0;
2582 if (journal)
2583 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2584 else
2585 return try_to_free_buffers(page);
2586 }
2587
2588 /*
2589 * ext4_get_block used when preparing for a DIO write or buffer write.
2590 * We allocate an uinitialized extent if blocks haven't been allocated.
2591 * The extent will be converted to initialized after the IO is complete.
2592 */
2593 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2594 struct buffer_head *bh_result, int create)
2595 {
2596 ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2597 inode->i_ino, create);
2598 return _ext4_get_block(inode, iblock, bh_result,
2599 EXT4_GET_BLOCKS_IO_CREATE_EXT);
2600 }
2601
2602 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2603 ssize_t size, void *private, int ret,
2604 bool is_async)
2605 {
2606 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2607 ext4_io_end_t *io_end = iocb->private;
2608 struct workqueue_struct *wq;
2609 unsigned long flags;
2610 struct ext4_inode_info *ei;
2611
2612 /* if not async direct IO or dio with 0 bytes write, just return */
2613 if (!io_end || !size)
2614 goto out;
2615
2616 ext_debug("ext4_end_io_dio(): io_end 0x%p"
2617 "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2618 iocb->private, io_end->inode->i_ino, iocb, offset,
2619 size);
2620
2621 /* if not aio dio with unwritten extents, just free io and return */
2622 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2623 ext4_free_io_end(io_end);
2624 iocb->private = NULL;
2625 out:
2626 if (is_async)
2627 aio_complete(iocb, ret, 0);
2628 inode_dio_done(inode);
2629 return;
2630 }
2631
2632 io_end->offset = offset;
2633 io_end->size = size;
2634 if (is_async) {
2635 io_end->iocb = iocb;
2636 io_end->result = ret;
2637 }
2638 wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2639
2640 /* Add the io_end to per-inode completed aio dio list*/
2641 ei = EXT4_I(io_end->inode);
2642 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2643 list_add_tail(&io_end->list, &ei->i_completed_io_list);
2644 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2645
2646 /* queue the work to convert unwritten extents to written */
2647 queue_work(wq, &io_end->work);
2648 iocb->private = NULL;
2649
2650 /* XXX: probably should move into the real I/O completion handler */
2651 inode_dio_done(inode);
2652 }
2653
2654 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2655 {
2656 ext4_io_end_t *io_end = bh->b_private;
2657 struct workqueue_struct *wq;
2658 struct inode *inode;
2659 unsigned long flags;
2660
2661 if (!test_clear_buffer_uninit(bh) || !io_end)
2662 goto out;
2663
2664 if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2665 printk("sb umounted, discard end_io request for inode %lu\n",
2666 io_end->inode->i_ino);
2667 ext4_free_io_end(io_end);
2668 goto out;
2669 }
2670
2671 io_end->flag = EXT4_IO_END_UNWRITTEN;
2672 inode = io_end->inode;
2673
2674 /* Add the io_end to per-inode completed io list*/
2675 spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2676 list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2677 spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2678
2679 wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2680 /* queue the work to convert unwritten extents to written */
2681 queue_work(wq, &io_end->work);
2682 out:
2683 bh->b_private = NULL;
2684 bh->b_end_io = NULL;
2685 clear_buffer_uninit(bh);
2686 end_buffer_async_write(bh, uptodate);
2687 }
2688
2689 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2690 {
2691 ext4_io_end_t *io_end;
2692 struct page *page = bh->b_page;
2693 loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2694 size_t size = bh->b_size;
2695
2696 retry:
2697 io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2698 if (!io_end) {
2699 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2700 schedule();
2701 goto retry;
2702 }
2703 io_end->offset = offset;
2704 io_end->size = size;
2705 /*
2706 * We need to hold a reference to the page to make sure it
2707 * doesn't get evicted before ext4_end_io_work() has a chance
2708 * to convert the extent from written to unwritten.
2709 */
2710 io_end->page = page;
2711 get_page(io_end->page);
2712
2713 bh->b_private = io_end;
2714 bh->b_end_io = ext4_end_io_buffer_write;
2715 return 0;
2716 }
2717
2718 /*
2719 * For ext4 extent files, ext4 will do direct-io write to holes,
2720 * preallocated extents, and those write extend the file, no need to
2721 * fall back to buffered IO.
2722 *
2723 * For holes, we fallocate those blocks, mark them as uninitialized
2724 * If those blocks were preallocated, we mark sure they are splited, but
2725 * still keep the range to write as uninitialized.
2726 *
2727 * The unwrritten extents will be converted to written when DIO is completed.
2728 * For async direct IO, since the IO may still pending when return, we
2729 * set up an end_io call back function, which will do the conversion
2730 * when async direct IO completed.
2731 *
2732 * If the O_DIRECT write will extend the file then add this inode to the
2733 * orphan list. So recovery will truncate it back to the original size
2734 * if the machine crashes during the write.
2735 *
2736 */
2737 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2738 const struct iovec *iov, loff_t offset,
2739 unsigned long nr_segs)
2740 {
2741 struct file *file = iocb->ki_filp;
2742 struct inode *inode = file->f_mapping->host;
2743 ssize_t ret;
2744 size_t count = iov_length(iov, nr_segs);
2745
2746 loff_t final_size = offset + count;
2747 if (rw == WRITE && final_size <= inode->i_size) {
2748 /*
2749 * We could direct write to holes and fallocate.
2750 *
2751 * Allocated blocks to fill the hole are marked as uninitialized
2752 * to prevent parallel buffered read to expose the stale data
2753 * before DIO complete the data IO.
2754 *
2755 * As to previously fallocated extents, ext4 get_block
2756 * will just simply mark the buffer mapped but still
2757 * keep the extents uninitialized.
2758 *
2759 * for non AIO case, we will convert those unwritten extents
2760 * to written after return back from blockdev_direct_IO.
2761 *
2762 * for async DIO, the conversion needs to be defered when
2763 * the IO is completed. The ext4 end_io callback function
2764 * will be called to take care of the conversion work.
2765 * Here for async case, we allocate an io_end structure to
2766 * hook to the iocb.
2767 */
2768 iocb->private = NULL;
2769 EXT4_I(inode)->cur_aio_dio = NULL;
2770 if (!is_sync_kiocb(iocb)) {
2771 iocb->private = ext4_init_io_end(inode, GFP_NOFS);
2772 if (!iocb->private)
2773 return -ENOMEM;
2774 /*
2775 * we save the io structure for current async
2776 * direct IO, so that later ext4_map_blocks()
2777 * could flag the io structure whether there
2778 * is a unwritten extents needs to be converted
2779 * when IO is completed.
2780 */
2781 EXT4_I(inode)->cur_aio_dio = iocb->private;
2782 }
2783
2784 ret = __blockdev_direct_IO(rw, iocb, inode,
2785 inode->i_sb->s_bdev, iov,
2786 offset, nr_segs,
2787 ext4_get_block_write,
2788 ext4_end_io_dio,
2789 NULL,
2790 DIO_LOCKING | DIO_SKIP_HOLES);
2791 if (iocb->private)
2792 EXT4_I(inode)->cur_aio_dio = NULL;
2793 /*
2794 * The io_end structure takes a reference to the inode,
2795 * that structure needs to be destroyed and the
2796 * reference to the inode need to be dropped, when IO is
2797 * complete, even with 0 byte write, or failed.
2798 *
2799 * In the successful AIO DIO case, the io_end structure will be
2800 * desctroyed and the reference to the inode will be dropped
2801 * after the end_io call back function is called.
2802 *
2803 * In the case there is 0 byte write, or error case, since
2804 * VFS direct IO won't invoke the end_io call back function,
2805 * we need to free the end_io structure here.
2806 */
2807 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2808 ext4_free_io_end(iocb->private);
2809 iocb->private = NULL;
2810 } else if (ret > 0 && ext4_test_inode_state(inode,
2811 EXT4_STATE_DIO_UNWRITTEN)) {
2812 int err;
2813 /*
2814 * for non AIO case, since the IO is already
2815 * completed, we could do the conversion right here
2816 */
2817 err = ext4_convert_unwritten_extents(inode,
2818 offset, ret);
2819 if (err < 0)
2820 ret = err;
2821 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2822 }
2823 return ret;
2824 }
2825
2826 /* for write the the end of file case, we fall back to old way */
2827 return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2828 }
2829
2830 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2831 const struct iovec *iov, loff_t offset,
2832 unsigned long nr_segs)
2833 {
2834 struct file *file = iocb->ki_filp;
2835 struct inode *inode = file->f_mapping->host;
2836 ssize_t ret;
2837
2838 trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
2839 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
2840 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
2841 else
2842 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2843 trace_ext4_direct_IO_exit(inode, offset,
2844 iov_length(iov, nr_segs), rw, ret);
2845 return ret;
2846 }
2847
2848 /*
2849 * Pages can be marked dirty completely asynchronously from ext4's journalling
2850 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
2851 * much here because ->set_page_dirty is called under VFS locks. The page is
2852 * not necessarily locked.
2853 *
2854 * We cannot just dirty the page and leave attached buffers clean, because the
2855 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
2856 * or jbddirty because all the journalling code will explode.
2857 *
2858 * So what we do is to mark the page "pending dirty" and next time writepage
2859 * is called, propagate that into the buffers appropriately.
2860 */
2861 static int ext4_journalled_set_page_dirty(struct page *page)
2862 {
2863 SetPageChecked(page);
2864 return __set_page_dirty_nobuffers(page);
2865 }
2866
2867 static const struct address_space_operations ext4_ordered_aops = {
2868 .readpage = ext4_readpage,
2869 .readpages = ext4_readpages,
2870 .writepage = ext4_writepage,
2871 .write_begin = ext4_write_begin,
2872 .write_end = ext4_ordered_write_end,
2873 .bmap = ext4_bmap,
2874 .invalidatepage = ext4_invalidatepage,
2875 .releasepage = ext4_releasepage,
2876 .direct_IO = ext4_direct_IO,
2877 .migratepage = buffer_migrate_page,
2878 .is_partially_uptodate = block_is_partially_uptodate,
2879 .error_remove_page = generic_error_remove_page,
2880 };
2881
2882 static const struct address_space_operations ext4_writeback_aops = {
2883 .readpage = ext4_readpage,
2884 .readpages = ext4_readpages,
2885 .writepage = ext4_writepage,
2886 .write_begin = ext4_write_begin,
2887 .write_end = ext4_writeback_write_end,
2888 .bmap = ext4_bmap,
2889 .invalidatepage = ext4_invalidatepage,
2890 .releasepage = ext4_releasepage,
2891 .direct_IO = ext4_direct_IO,
2892 .migratepage = buffer_migrate_page,
2893 .is_partially_uptodate = block_is_partially_uptodate,
2894 .error_remove_page = generic_error_remove_page,
2895 };
2896
2897 static const struct address_space_operations ext4_journalled_aops = {
2898 .readpage = ext4_readpage,
2899 .readpages = ext4_readpages,
2900 .writepage = ext4_writepage,
2901 .write_begin = ext4_write_begin,
2902 .write_end = ext4_journalled_write_end,
2903 .set_page_dirty = ext4_journalled_set_page_dirty,
2904 .bmap = ext4_bmap,
2905 .invalidatepage = ext4_invalidatepage,
2906 .releasepage = ext4_releasepage,
2907 .is_partially_uptodate = block_is_partially_uptodate,
2908 .error_remove_page = generic_error_remove_page,
2909 };
2910
2911 static const struct address_space_operations ext4_da_aops = {
2912 .readpage = ext4_readpage,
2913 .readpages = ext4_readpages,
2914 .writepage = ext4_writepage,
2915 .writepages = ext4_da_writepages,
2916 .write_begin = ext4_da_write_begin,
2917 .write_end = ext4_da_write_end,
2918 .bmap = ext4_bmap,
2919 .invalidatepage = ext4_da_invalidatepage,
2920 .releasepage = ext4_releasepage,
2921 .direct_IO = ext4_direct_IO,
2922 .migratepage = buffer_migrate_page,
2923 .is_partially_uptodate = block_is_partially_uptodate,
2924 .error_remove_page = generic_error_remove_page,
2925 };
2926
2927 void ext4_set_aops(struct inode *inode)
2928 {
2929 if (ext4_should_order_data(inode) &&
2930 test_opt(inode->i_sb, DELALLOC))
2931 inode->i_mapping->a_ops = &ext4_da_aops;
2932 else if (ext4_should_order_data(inode))
2933 inode->i_mapping->a_ops = &ext4_ordered_aops;
2934 else if (ext4_should_writeback_data(inode) &&
2935 test_opt(inode->i_sb, DELALLOC))
2936 inode->i_mapping->a_ops = &ext4_da_aops;
2937 else if (ext4_should_writeback_data(inode))
2938 inode->i_mapping->a_ops = &ext4_writeback_aops;
2939 else
2940 inode->i_mapping->a_ops = &ext4_journalled_aops;
2941 }
2942
2943 /*
2944 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
2945 * up to the end of the block which corresponds to `from'.
2946 * This required during truncate. We need to physically zero the tail end
2947 * of that block so it doesn't yield old data if the file is later grown.
2948 */
2949 int ext4_block_truncate_page(handle_t *handle,
2950 struct address_space *mapping, loff_t from)
2951 {
2952 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2953 unsigned length;
2954 unsigned blocksize;
2955 struct inode *inode = mapping->host;
2956
2957 blocksize = inode->i_sb->s_blocksize;
2958 length = blocksize - (offset & (blocksize - 1));
2959
2960 return ext4_block_zero_page_range(handle, mapping, from, length);
2961 }
2962
2963 /*
2964 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
2965 * starting from file offset 'from'. The range to be zero'd must
2966 * be contained with in one block. If the specified range exceeds
2967 * the end of the block it will be shortened to end of the block
2968 * that cooresponds to 'from'
2969 */
2970 int ext4_block_zero_page_range(handle_t *handle,
2971 struct address_space *mapping, loff_t from, loff_t length)
2972 {
2973 ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2974 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2975 unsigned blocksize, max, pos;
2976 ext4_lblk_t iblock;
2977 struct inode *inode = mapping->host;
2978 struct buffer_head *bh;
2979 struct page *page;
2980 int err = 0;
2981
2982 page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
2983 mapping_gfp_mask(mapping) & ~__GFP_FS);
2984 if (!page)
2985 return -EINVAL;
2986
2987 blocksize = inode->i_sb->s_blocksize;
2988 max = blocksize - (offset & (blocksize - 1));
2989
2990 /*
2991 * correct length if it does not fall between
2992 * 'from' and the end of the block
2993 */
2994 if (length > max || length < 0)
2995 length = max;
2996
2997 iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
2998
2999 if (!page_has_buffers(page))
3000 create_empty_buffers(page, blocksize, 0);
3001
3002 /* Find the buffer that contains "offset" */
3003 bh = page_buffers(page);
3004 pos = blocksize;
3005 while (offset >= pos) {
3006 bh = bh->b_this_page;
3007 iblock++;
3008 pos += blocksize;
3009 }
3010
3011 err = 0;
3012 if (buffer_freed(bh)) {
3013 BUFFER_TRACE(bh, "freed: skip");
3014 goto unlock;
3015 }
3016
3017 if (!buffer_mapped(bh)) {
3018 BUFFER_TRACE(bh, "unmapped");
3019 ext4_get_block(inode, iblock, bh, 0);
3020 /* unmapped? It's a hole - nothing to do */
3021 if (!buffer_mapped(bh)) {
3022 BUFFER_TRACE(bh, "still unmapped");
3023 goto unlock;
3024 }
3025 }
3026
3027 /* Ok, it's mapped. Make sure it's up-to-date */
3028 if (PageUptodate(page))
3029 set_buffer_uptodate(bh);
3030
3031 if (!buffer_uptodate(bh)) {
3032 err = -EIO;
3033 ll_rw_block(READ, 1, &bh);
3034 wait_on_buffer(bh);
3035 /* Uhhuh. Read error. Complain and punt. */
3036 if (!buffer_uptodate(bh))
3037 goto unlock;
3038 }
3039
3040 if (ext4_should_journal_data(inode)) {
3041 BUFFER_TRACE(bh, "get write access");
3042 err = ext4_journal_get_write_access(handle, bh);
3043 if (err)
3044 goto unlock;
3045 }
3046
3047 zero_user(page, offset, length);
3048
3049 BUFFER_TRACE(bh, "zeroed end of block");
3050
3051 err = 0;
3052 if (ext4_should_journal_data(inode)) {
3053 err = ext4_handle_dirty_metadata(handle, inode, bh);
3054 } else {
3055 if (ext4_should_order_data(inode) && EXT4_I(inode)->jinode)
3056 err = ext4_jbd2_file_inode(handle, inode);
3057 mark_buffer_dirty(bh);
3058 }
3059
3060 unlock:
3061 unlock_page(page);
3062 page_cache_release(page);
3063 return err;
3064 }
3065
3066 int ext4_can_truncate(struct inode *inode)
3067 {
3068 if (S_ISREG(inode->i_mode))
3069 return 1;
3070 if (S_ISDIR(inode->i_mode))
3071 return 1;
3072 if (S_ISLNK(inode->i_mode))
3073 return !ext4_inode_is_fast_symlink(inode);
3074 return 0;
3075 }
3076
3077 /*
3078 * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3079 * associated with the given offset and length
3080 *
3081 * @inode: File inode
3082 * @offset: The offset where the hole will begin
3083 * @len: The length of the hole
3084 *
3085 * Returns: 0 on sucess or negative on failure
3086 */
3087
3088 int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3089 {
3090 struct inode *inode = file->f_path.dentry->d_inode;
3091 if (!S_ISREG(inode->i_mode))
3092 return -ENOTSUPP;
3093
3094 if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3095 /* TODO: Add support for non extent hole punching */
3096 return -ENOTSUPP;
3097 }
3098
3099 return ext4_ext_punch_hole(file, offset, length);
3100 }
3101
3102 /*
3103 * ext4_truncate()
3104 *
3105 * We block out ext4_get_block() block instantiations across the entire
3106 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3107 * simultaneously on behalf of the same inode.
3108 *
3109 * As we work through the truncate and commmit bits of it to the journal there
3110 * is one core, guiding principle: the file's tree must always be consistent on
3111 * disk. We must be able to restart the truncate after a crash.
3112 *
3113 * The file's tree may be transiently inconsistent in memory (although it
3114 * probably isn't), but whenever we close off and commit a journal transaction,
3115 * the contents of (the filesystem + the journal) must be consistent and
3116 * restartable. It's pretty simple, really: bottom up, right to left (although
3117 * left-to-right works OK too).
3118 *
3119 * Note that at recovery time, journal replay occurs *before* the restart of
3120 * truncate against the orphan inode list.
3121 *
3122 * The committed inode has the new, desired i_size (which is the same as
3123 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
3124 * that this inode's truncate did not complete and it will again call
3125 * ext4_truncate() to have another go. So there will be instantiated blocks
3126 * to the right of the truncation point in a crashed ext4 filesystem. But
3127 * that's fine - as long as they are linked from the inode, the post-crash
3128 * ext4_truncate() run will find them and release them.
3129 */
3130 void ext4_truncate(struct inode *inode)
3131 {
3132 trace_ext4_truncate_enter(inode);
3133
3134 if (!ext4_can_truncate(inode))
3135 return;
3136
3137 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3138
3139 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3140 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3141
3142 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3143 ext4_ext_truncate(inode);
3144 else
3145 ext4_ind_truncate(inode);
3146
3147 trace_ext4_truncate_exit(inode);
3148 }
3149
3150 /*
3151 * ext4_get_inode_loc returns with an extra refcount against the inode's
3152 * underlying buffer_head on success. If 'in_mem' is true, we have all
3153 * data in memory that is needed to recreate the on-disk version of this
3154 * inode.
3155 */
3156 static int __ext4_get_inode_loc(struct inode *inode,
3157 struct ext4_iloc *iloc, int in_mem)
3158 {
3159 struct ext4_group_desc *gdp;
3160 struct buffer_head *bh;
3161 struct super_block *sb = inode->i_sb;
3162 ext4_fsblk_t block;
3163 int inodes_per_block, inode_offset;
3164
3165 iloc->bh = NULL;
3166 if (!ext4_valid_inum(sb, inode->i_ino))
3167 return -EIO;
3168
3169 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3170 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3171 if (!gdp)
3172 return -EIO;
3173
3174 /*
3175 * Figure out the offset within the block group inode table
3176 */
3177 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3178 inode_offset = ((inode->i_ino - 1) %
3179 EXT4_INODES_PER_GROUP(sb));
3180 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3181 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3182
3183 bh = sb_getblk(sb, block);
3184 if (!bh) {
3185 EXT4_ERROR_INODE_BLOCK(inode, block,
3186 "unable to read itable block");
3187 return -EIO;
3188 }
3189 if (!buffer_uptodate(bh)) {
3190 lock_buffer(bh);
3191
3192 /*
3193 * If the buffer has the write error flag, we have failed
3194 * to write out another inode in the same block. In this
3195 * case, we don't have to read the block because we may
3196 * read the old inode data successfully.
3197 */
3198 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3199 set_buffer_uptodate(bh);
3200
3201 if (buffer_uptodate(bh)) {
3202 /* someone brought it uptodate while we waited */
3203 unlock_buffer(bh);
3204 goto has_buffer;
3205 }
3206
3207 /*
3208 * If we have all information of the inode in memory and this
3209 * is the only valid inode in the block, we need not read the
3210 * block.
3211 */
3212 if (in_mem) {
3213 struct buffer_head *bitmap_bh;
3214 int i, start;
3215
3216 start = inode_offset & ~(inodes_per_block - 1);
3217
3218 /* Is the inode bitmap in cache? */
3219 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3220 if (!bitmap_bh)
3221 goto make_io;
3222
3223 /*
3224 * If the inode bitmap isn't in cache then the
3225 * optimisation may end up performing two reads instead
3226 * of one, so skip it.
3227 */
3228 if (!buffer_uptodate(bitmap_bh)) {
3229 brelse(bitmap_bh);
3230 goto make_io;
3231 }
3232 for (i = start; i < start + inodes_per_block; i++) {
3233 if (i == inode_offset)
3234 continue;
3235 if (ext4_test_bit(i, bitmap_bh->b_data))
3236 break;
3237 }
3238 brelse(bitmap_bh);
3239 if (i == start + inodes_per_block) {
3240 /* all other inodes are free, so skip I/O */
3241 memset(bh->b_data, 0, bh->b_size);
3242 set_buffer_uptodate(bh);
3243 unlock_buffer(bh);
3244 goto has_buffer;
3245 }
3246 }
3247
3248 make_io:
3249 /*
3250 * If we need to do any I/O, try to pre-readahead extra
3251 * blocks from the inode table.
3252 */
3253 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3254 ext4_fsblk_t b, end, table;
3255 unsigned num;
3256
3257 table = ext4_inode_table(sb, gdp);
3258 /* s_inode_readahead_blks is always a power of 2 */
3259 b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3260 if (table > b)
3261 b = table;
3262 end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3263 num = EXT4_INODES_PER_GROUP(sb);
3264 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3265 EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3266 num -= ext4_itable_unused_count(sb, gdp);
3267 table += num / inodes_per_block;
3268 if (end > table)
3269 end = table;
3270 while (b <= end)
3271 sb_breadahead(sb, b++);
3272 }
3273
3274 /*
3275 * There are other valid inodes in the buffer, this inode
3276 * has in-inode xattrs, or we don't have this inode in memory.
3277 * Read the block from disk.
3278 */
3279 trace_ext4_load_inode(inode);
3280 get_bh(bh);
3281 bh->b_end_io = end_buffer_read_sync;
3282 submit_bh(READ_META, bh);
3283 wait_on_buffer(bh);
3284 if (!buffer_uptodate(bh)) {
3285 EXT4_ERROR_INODE_BLOCK(inode, block,
3286 "unable to read itable block");
3287 brelse(bh);
3288 return -EIO;
3289 }
3290 }
3291 has_buffer:
3292 iloc->bh = bh;
3293 return 0;
3294 }
3295
3296 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3297 {
3298 /* We have all inode data except xattrs in memory here. */
3299 return __ext4_get_inode_loc(inode, iloc,
3300 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3301 }
3302
3303 void ext4_set_inode_flags(struct inode *inode)
3304 {
3305 unsigned int flags = EXT4_I(inode)->i_flags;
3306
3307 inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3308 if (flags & EXT4_SYNC_FL)
3309 inode->i_flags |= S_SYNC;
3310 if (flags & EXT4_APPEND_FL)
3311 inode->i_flags |= S_APPEND;
3312 if (flags & EXT4_IMMUTABLE_FL)
3313 inode->i_flags |= S_IMMUTABLE;
3314 if (flags & EXT4_NOATIME_FL)
3315 inode->i_flags |= S_NOATIME;
3316 if (flags & EXT4_DIRSYNC_FL)
3317 inode->i_flags |= S_DIRSYNC;
3318 }
3319
3320 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3321 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3322 {
3323 unsigned int vfs_fl;
3324 unsigned long old_fl, new_fl;
3325
3326 do {
3327 vfs_fl = ei->vfs_inode.i_flags;
3328 old_fl = ei->i_flags;
3329 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3330 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3331 EXT4_DIRSYNC_FL);
3332 if (vfs_fl & S_SYNC)
3333 new_fl |= EXT4_SYNC_FL;
3334 if (vfs_fl & S_APPEND)
3335 new_fl |= EXT4_APPEND_FL;
3336 if (vfs_fl & S_IMMUTABLE)
3337 new_fl |= EXT4_IMMUTABLE_FL;
3338 if (vfs_fl & S_NOATIME)
3339 new_fl |= EXT4_NOATIME_FL;
3340 if (vfs_fl & S_DIRSYNC)
3341 new_fl |= EXT4_DIRSYNC_FL;
3342 } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3343 }
3344
3345 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3346 struct ext4_inode_info *ei)
3347 {
3348 blkcnt_t i_blocks ;
3349 struct inode *inode = &(ei->vfs_inode);
3350 struct super_block *sb = inode->i_sb;
3351
3352 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3353 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3354 /* we are using combined 48 bit field */
3355 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3356 le32_to_cpu(raw_inode->i_blocks_lo);
3357 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
3358 /* i_blocks represent file system block size */
3359 return i_blocks << (inode->i_blkbits - 9);
3360 } else {
3361 return i_blocks;
3362 }
3363 } else {
3364 return le32_to_cpu(raw_inode->i_blocks_lo);
3365 }
3366 }
3367
3368 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3369 {
3370 struct ext4_iloc iloc;
3371 struct ext4_inode *raw_inode;
3372 struct ext4_inode_info *ei;
3373 struct inode *inode;
3374 journal_t *journal = EXT4_SB(sb)->s_journal;
3375 long ret;
3376 int block;
3377
3378 inode = iget_locked(sb, ino);
3379 if (!inode)
3380 return ERR_PTR(-ENOMEM);
3381 if (!(inode->i_state & I_NEW))
3382 return inode;
3383
3384 ei = EXT4_I(inode);
3385 iloc.bh = NULL;
3386
3387 ret = __ext4_get_inode_loc(inode, &iloc, 0);
3388 if (ret < 0)
3389 goto bad_inode;
3390 raw_inode = ext4_raw_inode(&iloc);
3391 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3392 inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3393 inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3394 if (!(test_opt(inode->i_sb, NO_UID32))) {
3395 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3396 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3397 }
3398 inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
3399
3400 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
3401 ei->i_dir_start_lookup = 0;
3402 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3403 /* We now have enough fields to check if the inode was active or not.
3404 * This is needed because nfsd might try to access dead inodes
3405 * the test is that same one that e2fsck uses
3406 * NeilBrown 1999oct15
3407 */
3408 if (inode->i_nlink == 0) {
3409 if (inode->i_mode == 0 ||
3410 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3411 /* this inode is deleted */
3412 ret = -ESTALE;
3413 goto bad_inode;
3414 }
3415 /* The only unlinked inodes we let through here have
3416 * valid i_mode and are being read by the orphan
3417 * recovery code: that's fine, we're about to complete
3418 * the process of deleting those. */
3419 }
3420 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3421 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3422 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3423 if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
3424 ei->i_file_acl |=
3425 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3426 inode->i_size = ext4_isize(raw_inode);
3427 ei->i_disksize = inode->i_size;
3428 #ifdef CONFIG_QUOTA
3429 ei->i_reserved_quota = 0;
3430 #endif
3431 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3432 ei->i_block_group = iloc.block_group;
3433 ei->i_last_alloc_group = ~0;
3434 /*
3435 * NOTE! The in-memory inode i_data array is in little-endian order
3436 * even on big-endian machines: we do NOT byteswap the block numbers!
3437 */
3438 for (block = 0; block < EXT4_N_BLOCKS; block++)
3439 ei->i_data[block] = raw_inode->i_block[block];
3440 INIT_LIST_HEAD(&ei->i_orphan);
3441
3442 /*
3443 * Set transaction id's of transactions that have to be committed
3444 * to finish f[data]sync. We set them to currently running transaction
3445 * as we cannot be sure that the inode or some of its metadata isn't
3446 * part of the transaction - the inode could have been reclaimed and
3447 * now it is reread from disk.
3448 */
3449 if (journal) {
3450 transaction_t *transaction;
3451 tid_t tid;
3452
3453 read_lock(&journal->j_state_lock);
3454 if (journal->j_running_transaction)
3455 transaction = journal->j_running_transaction;
3456 else
3457 transaction = journal->j_committing_transaction;
3458 if (transaction)
3459 tid = transaction->t_tid;
3460 else
3461 tid = journal->j_commit_sequence;
3462 read_unlock(&journal->j_state_lock);
3463 ei->i_sync_tid = tid;
3464 ei->i_datasync_tid = tid;
3465 }
3466
3467 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3468 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3469 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3470 EXT4_INODE_SIZE(inode->i_sb)) {
3471 ret = -EIO;
3472 goto bad_inode;
3473 }
3474 if (ei->i_extra_isize == 0) {
3475 /* The extra space is currently unused. Use it. */
3476 ei->i_extra_isize = sizeof(struct ext4_inode) -
3477 EXT4_GOOD_OLD_INODE_SIZE;
3478 } else {
3479 __le32 *magic = (void *)raw_inode +
3480 EXT4_GOOD_OLD_INODE_SIZE +
3481 ei->i_extra_isize;
3482 if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3483 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
3484 }
3485 } else
3486 ei->i_extra_isize = 0;
3487
3488 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
3489 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
3490 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
3491 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
3492
3493 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
3494 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3495 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3496 inode->i_version |=
3497 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
3498 }
3499
3500 ret = 0;
3501 if (ei->i_file_acl &&
3502 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
3503 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
3504 ei->i_file_acl);
3505 ret = -EIO;
3506 goto bad_inode;
3507 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3508 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3509 (S_ISLNK(inode->i_mode) &&
3510 !ext4_inode_is_fast_symlink(inode)))
3511 /* Validate extent which is part of inode */
3512 ret = ext4_ext_check_inode(inode);
3513 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3514 (S_ISLNK(inode->i_mode) &&
3515 !ext4_inode_is_fast_symlink(inode))) {
3516 /* Validate block references which are part of inode */
3517 ret = ext4_ind_check_inode(inode);
3518 }
3519 if (ret)
3520 goto bad_inode;
3521
3522 if (S_ISREG(inode->i_mode)) {
3523 inode->i_op = &ext4_file_inode_operations;
3524 inode->i_fop = &ext4_file_operations;
3525 ext4_set_aops(inode);
3526 } else if (S_ISDIR(inode->i_mode)) {
3527 inode->i_op = &ext4_dir_inode_operations;
3528 inode->i_fop = &ext4_dir_operations;
3529 } else if (S_ISLNK(inode->i_mode)) {
3530 if (ext4_inode_is_fast_symlink(inode)) {
3531 inode->i_op = &ext4_fast_symlink_inode_operations;
3532 nd_terminate_link(ei->i_data, inode->i_size,
3533 sizeof(ei->i_data) - 1);
3534 } else {
3535 inode->i_op = &ext4_symlink_inode_operations;
3536 ext4_set_aops(inode);
3537 }
3538 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
3539 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
3540 inode->i_op = &ext4_special_inode_operations;
3541 if (raw_inode->i_block[0])
3542 init_special_inode(inode, inode->i_mode,
3543 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3544 else
3545 init_special_inode(inode, inode->i_mode,
3546 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3547 } else {
3548 ret = -EIO;
3549 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
3550 goto bad_inode;
3551 }
3552 brelse(iloc.bh);
3553 ext4_set_inode_flags(inode);
3554 unlock_new_inode(inode);
3555 return inode;
3556
3557 bad_inode:
3558 brelse(iloc.bh);
3559 iget_failed(inode);
3560 return ERR_PTR(ret);
3561 }
3562
3563 static int ext4_inode_blocks_set(handle_t *handle,
3564 struct ext4_inode *raw_inode,
3565 struct ext4_inode_info *ei)
3566 {
3567 struct inode *inode = &(ei->vfs_inode);
3568 u64 i_blocks = inode->i_blocks;
3569 struct super_block *sb = inode->i_sb;
3570
3571 if (i_blocks <= ~0U) {
3572 /*
3573 * i_blocks can be represnted in a 32 bit variable
3574 * as multiple of 512 bytes
3575 */
3576 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3577 raw_inode->i_blocks_high = 0;
3578 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3579 return 0;
3580 }
3581 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
3582 return -EFBIG;
3583
3584 if (i_blocks <= 0xffffffffffffULL) {
3585 /*
3586 * i_blocks can be represented in a 48 bit variable
3587 * as multiple of 512 bytes
3588 */
3589 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3590 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3591 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3592 } else {
3593 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3594 /* i_block is stored in file system block size */
3595 i_blocks = i_blocks >> (inode->i_blkbits - 9);
3596 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3597 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3598 }
3599 return 0;
3600 }
3601
3602 /*
3603 * Post the struct inode info into an on-disk inode location in the
3604 * buffer-cache. This gobbles the caller's reference to the
3605 * buffer_head in the inode location struct.
3606 *
3607 * The caller must have write access to iloc->bh.
3608 */
3609 static int ext4_do_update_inode(handle_t *handle,
3610 struct inode *inode,
3611 struct ext4_iloc *iloc)
3612 {
3613 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
3614 struct ext4_inode_info *ei = EXT4_I(inode);
3615 struct buffer_head *bh = iloc->bh;
3616 int err = 0, rc, block;
3617
3618 /* For fields not not tracking in the in-memory inode,
3619 * initialise them to zero for new inodes. */
3620 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
3621 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
3622
3623 ext4_get_inode_flags(ei);
3624 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3625 if (!(test_opt(inode->i_sb, NO_UID32))) {
3626 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3627 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3628 /*
3629 * Fix up interoperability with old kernels. Otherwise, old inodes get
3630 * re-used with the upper 16 bits of the uid/gid intact
3631 */
3632 if (!ei->i_dtime) {
3633 raw_inode->i_uid_high =
3634 cpu_to_le16(high_16_bits(inode->i_uid));
3635 raw_inode->i_gid_high =
3636 cpu_to_le16(high_16_bits(inode->i_gid));
3637 } else {
3638 raw_inode->i_uid_high = 0;
3639 raw_inode->i_gid_high = 0;
3640 }
3641 } else {
3642 raw_inode->i_uid_low =
3643 cpu_to_le16(fs_high2lowuid(inode->i_uid));
3644 raw_inode->i_gid_low =
3645 cpu_to_le16(fs_high2lowgid(inode->i_gid));
3646 raw_inode->i_uid_high = 0;
3647 raw_inode->i_gid_high = 0;
3648 }
3649 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3650
3651 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
3652 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
3653 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
3654 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
3655
3656 if (ext4_inode_blocks_set(handle, raw_inode, ei))
3657 goto out_brelse;
3658 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3659 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
3660 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
3661 cpu_to_le32(EXT4_OS_HURD))
3662 raw_inode->i_file_acl_high =
3663 cpu_to_le16(ei->i_file_acl >> 32);
3664 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
3665 ext4_isize_set(raw_inode, ei->i_disksize);
3666 if (ei->i_disksize > 0x7fffffffULL) {
3667 struct super_block *sb = inode->i_sb;
3668 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
3669 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
3670 EXT4_SB(sb)->s_es->s_rev_level ==
3671 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
3672 /* If this is the first large file
3673 * created, add a flag to the superblock.
3674 */
3675 err = ext4_journal_get_write_access(handle,
3676 EXT4_SB(sb)->s_sbh);
3677 if (err)
3678 goto out_brelse;
3679 ext4_update_dynamic_rev(sb);
3680 EXT4_SET_RO_COMPAT_FEATURE(sb,
3681 EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
3682 sb->s_dirt = 1;
3683 ext4_handle_sync(handle);
3684 err = ext4_handle_dirty_metadata(handle, NULL,
3685 EXT4_SB(sb)->s_sbh);
3686 }
3687 }
3688 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3689 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3690 if (old_valid_dev(inode->i_rdev)) {
3691 raw_inode->i_block[0] =
3692 cpu_to_le32(old_encode_dev(inode->i_rdev));
3693 raw_inode->i_block[1] = 0;
3694 } else {
3695 raw_inode->i_block[0] = 0;
3696 raw_inode->i_block[1] =
3697 cpu_to_le32(new_encode_dev(inode->i_rdev));
3698 raw_inode->i_block[2] = 0;
3699 }
3700 } else
3701 for (block = 0; block < EXT4_N_BLOCKS; block++)
3702 raw_inode->i_block[block] = ei->i_data[block];
3703
3704 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
3705 if (ei->i_extra_isize) {
3706 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3707 raw_inode->i_version_hi =
3708 cpu_to_le32(inode->i_version >> 32);
3709 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3710 }
3711
3712 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3713 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
3714 if (!err)
3715 err = rc;
3716 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
3717
3718 ext4_update_inode_fsync_trans(handle, inode, 0);
3719 out_brelse:
3720 brelse(bh);
3721 ext4_std_error(inode->i_sb, err);
3722 return err;
3723 }
3724
3725 /*
3726 * ext4_write_inode()
3727 *
3728 * We are called from a few places:
3729 *
3730 * - Within generic_file_write() for O_SYNC files.
3731 * Here, there will be no transaction running. We wait for any running
3732 * trasnaction to commit.
3733 *
3734 * - Within sys_sync(), kupdate and such.
3735 * We wait on commit, if tol to.
3736 *
3737 * - Within prune_icache() (PF_MEMALLOC == true)
3738 * Here we simply return. We can't afford to block kswapd on the
3739 * journal commit.
3740 *
3741 * In all cases it is actually safe for us to return without doing anything,
3742 * because the inode has been copied into a raw inode buffer in
3743 * ext4_mark_inode_dirty(). This is a correctness thing for O_SYNC and for
3744 * knfsd.
3745 *
3746 * Note that we are absolutely dependent upon all inode dirtiers doing the
3747 * right thing: they *must* call mark_inode_dirty() after dirtying info in
3748 * which we are interested.
3749 *
3750 * It would be a bug for them to not do this. The code:
3751 *
3752 * mark_inode_dirty(inode)
3753 * stuff();
3754 * inode->i_size = expr;
3755 *
3756 * is in error because a kswapd-driven write_inode() could occur while
3757 * `stuff()' is running, and the new i_size will be lost. Plus the inode
3758 * will no longer be on the superblock's dirty inode list.
3759 */
3760 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
3761 {
3762 int err;
3763
3764 if (current->flags & PF_MEMALLOC)
3765 return 0;
3766
3767 if (EXT4_SB(inode->i_sb)->s_journal) {
3768 if (ext4_journal_current_handle()) {
3769 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3770 dump_stack();
3771 return -EIO;
3772 }
3773
3774 if (wbc->sync_mode != WB_SYNC_ALL)
3775 return 0;
3776
3777 err = ext4_force_commit(inode->i_sb);
3778 } else {
3779 struct ext4_iloc iloc;
3780
3781 err = __ext4_get_inode_loc(inode, &iloc, 0);
3782 if (err)
3783 return err;
3784 if (wbc->sync_mode == WB_SYNC_ALL)
3785 sync_dirty_buffer(iloc.bh);
3786 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
3787 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
3788 "IO error syncing inode");
3789 err = -EIO;
3790 }
3791 brelse(iloc.bh);
3792 }
3793 return err;
3794 }
3795
3796 /*
3797 * ext4_setattr()
3798 *
3799 * Called from notify_change.
3800 *
3801 * We want to trap VFS attempts to truncate the file as soon as
3802 * possible. In particular, we want to make sure that when the VFS
3803 * shrinks i_size, we put the inode on the orphan list and modify
3804 * i_disksize immediately, so that during the subsequent flushing of
3805 * dirty pages and freeing of disk blocks, we can guarantee that any
3806 * commit will leave the blocks being flushed in an unused state on
3807 * disk. (On recovery, the inode will get truncated and the blocks will
3808 * be freed, so we have a strong guarantee that no future commit will
3809 * leave these blocks visible to the user.)
3810 *
3811 * Another thing we have to assure is that if we are in ordered mode
3812 * and inode is still attached to the committing transaction, we must
3813 * we start writeout of all the dirty pages which are being truncated.
3814 * This way we are sure that all the data written in the previous
3815 * transaction are already on disk (truncate waits for pages under
3816 * writeback).
3817 *
3818 * Called with inode->i_mutex down.
3819 */
3820 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
3821 {
3822 struct inode *inode = dentry->d_inode;
3823 int error, rc = 0;
3824 int orphan = 0;
3825 const unsigned int ia_valid = attr->ia_valid;
3826
3827 error = inode_change_ok(inode, attr);
3828 if (error)
3829 return error;
3830
3831 if (is_quota_modification(inode, attr))
3832 dquot_initialize(inode);
3833 if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3834 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3835 handle_t *handle;
3836
3837 /* (user+group)*(old+new) structure, inode write (sb,
3838 * inode block, ? - but truncate inode update has it) */
3839 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3840 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
3841 if (IS_ERR(handle)) {
3842 error = PTR_ERR(handle);
3843 goto err_out;
3844 }
3845 error = dquot_transfer(inode, attr);
3846 if (error) {
3847 ext4_journal_stop(handle);
3848 return error;
3849 }
3850 /* Update corresponding info in inode so that everything is in
3851 * one transaction */
3852 if (attr->ia_valid & ATTR_UID)
3853 inode->i_uid = attr->ia_uid;
3854 if (attr->ia_valid & ATTR_GID)
3855 inode->i_gid = attr->ia_gid;
3856 error = ext4_mark_inode_dirty(handle, inode);
3857 ext4_journal_stop(handle);
3858 }
3859
3860 if (attr->ia_valid & ATTR_SIZE) {
3861 inode_dio_wait(inode);
3862
3863 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3864 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3865
3866 if (attr->ia_size > sbi->s_bitmap_maxbytes)
3867 return -EFBIG;
3868 }
3869 }
3870
3871 if (S_ISREG(inode->i_mode) &&
3872 attr->ia_valid & ATTR_SIZE &&
3873 (attr->ia_size < inode->i_size)) {
3874 handle_t *handle;
3875
3876 handle = ext4_journal_start(inode, 3);
3877 if (IS_ERR(handle)) {
3878 error = PTR_ERR(handle);
3879 goto err_out;
3880 }
3881 if (ext4_handle_valid(handle)) {
3882 error = ext4_orphan_add(handle, inode);
3883 orphan = 1;
3884 }
3885 EXT4_I(inode)->i_disksize = attr->ia_size;
3886 rc = ext4_mark_inode_dirty(handle, inode);
3887 if (!error)
3888 error = rc;
3889 ext4_journal_stop(handle);
3890
3891 if (ext4_should_order_data(inode)) {
3892 error = ext4_begin_ordered_truncate(inode,
3893 attr->ia_size);
3894 if (error) {
3895 /* Do as much error cleanup as possible */
3896 handle = ext4_journal_start(inode, 3);
3897 if (IS_ERR(handle)) {
3898 ext4_orphan_del(NULL, inode);
3899 goto err_out;
3900 }
3901 ext4_orphan_del(handle, inode);
3902 orphan = 0;
3903 ext4_journal_stop(handle);
3904 goto err_out;
3905 }
3906 }
3907 }
3908
3909 if (attr->ia_valid & ATTR_SIZE) {
3910 if (attr->ia_size != i_size_read(inode)) {
3911 truncate_setsize(inode, attr->ia_size);
3912 ext4_truncate(inode);
3913 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3914 ext4_truncate(inode);
3915 }
3916
3917 if (!rc) {
3918 setattr_copy(inode, attr);
3919 mark_inode_dirty(inode);
3920 }
3921
3922 /*
3923 * If the call to ext4_truncate failed to get a transaction handle at
3924 * all, we need to clean up the in-core orphan list manually.
3925 */
3926 if (orphan && inode->i_nlink)
3927 ext4_orphan_del(NULL, inode);
3928
3929 if (!rc && (ia_valid & ATTR_MODE))
3930 rc = ext4_acl_chmod(inode);
3931
3932 err_out:
3933 ext4_std_error(inode->i_sb, error);
3934 if (!error)
3935 error = rc;
3936 return error;
3937 }
3938
3939 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
3940 struct kstat *stat)
3941 {
3942 struct inode *inode;
3943 unsigned long delalloc_blocks;
3944
3945 inode = dentry->d_inode;
3946 generic_fillattr(inode, stat);
3947
3948 /*
3949 * We can't update i_blocks if the block allocation is delayed
3950 * otherwise in the case of system crash before the real block
3951 * allocation is done, we will have i_blocks inconsistent with
3952 * on-disk file blocks.
3953 * We always keep i_blocks updated together with real
3954 * allocation. But to not confuse with user, stat
3955 * will return the blocks that include the delayed allocation
3956 * blocks for this file.
3957 */
3958 delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
3959
3960 stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
3961 return 0;
3962 }
3963
3964 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
3965 {
3966 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
3967 return ext4_ind_trans_blocks(inode, nrblocks, chunk);
3968 return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
3969 }
3970
3971 /*
3972 * Account for index blocks, block groups bitmaps and block group
3973 * descriptor blocks if modify datablocks and index blocks
3974 * worse case, the indexs blocks spread over different block groups
3975 *
3976 * If datablocks are discontiguous, they are possible to spread over
3977 * different block groups too. If they are contiuguous, with flexbg,
3978 * they could still across block group boundary.
3979 *
3980 * Also account for superblock, inode, quota and xattr blocks
3981 */
3982 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
3983 {
3984 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
3985 int gdpblocks;
3986 int idxblocks;
3987 int ret = 0;
3988
3989 /*
3990 * How many index blocks need to touch to modify nrblocks?
3991 * The "Chunk" flag indicating whether the nrblocks is
3992 * physically contiguous on disk
3993 *
3994 * For Direct IO and fallocate, they calls get_block to allocate
3995 * one single extent at a time, so they could set the "Chunk" flag
3996 */
3997 idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
3998
3999 ret = idxblocks;
4000
4001 /*
4002 * Now let's see how many group bitmaps and group descriptors need
4003 * to account
4004 */
4005 groups = idxblocks;
4006 if (chunk)
4007 groups += 1;
4008 else
4009 groups += nrblocks;
4010
4011 gdpblocks = groups;
4012 if (groups > ngroups)
4013 groups = ngroups;
4014 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4015 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4016
4017 /* bitmaps and block group descriptor blocks */
4018 ret += groups + gdpblocks;
4019
4020 /* Blocks for super block, inode, quota and xattr blocks */
4021 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4022
4023 return ret;
4024 }
4025
4026 /*
4027 * Calculate the total number of credits to reserve to fit
4028 * the modification of a single pages into a single transaction,
4029 * which may include multiple chunks of block allocations.
4030 *
4031 * This could be called via ext4_write_begin()
4032 *
4033 * We need to consider the worse case, when
4034 * one new block per extent.
4035 */
4036 int ext4_writepage_trans_blocks(struct inode *inode)
4037 {
4038 int bpp = ext4_journal_blocks_per_page(inode);
4039 int ret;
4040
4041 ret = ext4_meta_trans_blocks(inode, bpp, 0);
4042
4043 /* Account for data blocks for journalled mode */
4044 if (ext4_should_journal_data(inode))
4045 ret += bpp;
4046 return ret;
4047 }
4048
4049 /*
4050 * Calculate the journal credits for a chunk of data modification.
4051 *
4052 * This is called from DIO, fallocate or whoever calling
4053 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4054 *
4055 * journal buffers for data blocks are not included here, as DIO
4056 * and fallocate do no need to journal data buffers.
4057 */
4058 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4059 {
4060 return ext4_meta_trans_blocks(inode, nrblocks, 1);
4061 }
4062
4063 /*
4064 * The caller must have previously called ext4_reserve_inode_write().
4065 * Give this, we know that the caller already has write access to iloc->bh.
4066 */
4067 int ext4_mark_iloc_dirty(handle_t *handle,
4068 struct inode *inode, struct ext4_iloc *iloc)
4069 {
4070 int err = 0;
4071
4072 if (test_opt(inode->i_sb, I_VERSION))
4073 inode_inc_iversion(inode);
4074
4075 /* the do_update_inode consumes one bh->b_count */
4076 get_bh(iloc->bh);
4077
4078 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4079 err = ext4_do_update_inode(handle, inode, iloc);
4080 put_bh(iloc->bh);
4081 return err;
4082 }
4083
4084 /*
4085 * On success, We end up with an outstanding reference count against
4086 * iloc->bh. This _must_ be cleaned up later.
4087 */
4088
4089 int
4090 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4091 struct ext4_iloc *iloc)
4092 {
4093 int err;
4094
4095 err = ext4_get_inode_loc(inode, iloc);
4096 if (!err) {
4097 BUFFER_TRACE(iloc->bh, "get_write_access");
4098 err = ext4_journal_get_write_access(handle, iloc->bh);
4099 if (err) {
4100 brelse(iloc->bh);
4101 iloc->bh = NULL;
4102 }
4103 }
4104 ext4_std_error(inode->i_sb, err);
4105 return err;
4106 }
4107
4108 /*
4109 * Expand an inode by new_extra_isize bytes.
4110 * Returns 0 on success or negative error number on failure.
4111 */
4112 static int ext4_expand_extra_isize(struct inode *inode,
4113 unsigned int new_extra_isize,
4114 struct ext4_iloc iloc,
4115 handle_t *handle)
4116 {
4117 struct ext4_inode *raw_inode;
4118 struct ext4_xattr_ibody_header *header;
4119
4120 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4121 return 0;
4122
4123 raw_inode = ext4_raw_inode(&iloc);
4124
4125 header = IHDR(inode, raw_inode);
4126
4127 /* No extended attributes present */
4128 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4129 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4130 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4131 new_extra_isize);
4132 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4133 return 0;
4134 }
4135
4136 /* try to expand with EAs present */
4137 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4138 raw_inode, handle);
4139 }
4140
4141 /*
4142 * What we do here is to mark the in-core inode as clean with respect to inode
4143 * dirtiness (it may still be data-dirty).
4144 * This means that the in-core inode may be reaped by prune_icache
4145 * without having to perform any I/O. This is a very good thing,
4146 * because *any* task may call prune_icache - even ones which
4147 * have a transaction open against a different journal.
4148 *
4149 * Is this cheating? Not really. Sure, we haven't written the
4150 * inode out, but prune_icache isn't a user-visible syncing function.
4151 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4152 * we start and wait on commits.
4153 *
4154 * Is this efficient/effective? Well, we're being nice to the system
4155 * by cleaning up our inodes proactively so they can be reaped
4156 * without I/O. But we are potentially leaving up to five seconds'
4157 * worth of inodes floating about which prune_icache wants us to
4158 * write out. One way to fix that would be to get prune_icache()
4159 * to do a write_super() to free up some memory. It has the desired
4160 * effect.
4161 */
4162 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4163 {
4164 struct ext4_iloc iloc;
4165 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4166 static unsigned int mnt_count;
4167 int err, ret;
4168
4169 might_sleep();
4170 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4171 err = ext4_reserve_inode_write(handle, inode, &iloc);
4172 if (ext4_handle_valid(handle) &&
4173 EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4174 !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4175 /*
4176 * We need extra buffer credits since we may write into EA block
4177 * with this same handle. If journal_extend fails, then it will
4178 * only result in a minor loss of functionality for that inode.
4179 * If this is felt to be critical, then e2fsck should be run to
4180 * force a large enough s_min_extra_isize.
4181 */
4182 if ((jbd2_journal_extend(handle,
4183 EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4184 ret = ext4_expand_extra_isize(inode,
4185 sbi->s_want_extra_isize,
4186 iloc, handle);
4187 if (ret) {
4188 ext4_set_inode_state(inode,
4189 EXT4_STATE_NO_EXPAND);
4190 if (mnt_count !=
4191 le16_to_cpu(sbi->s_es->s_mnt_count)) {
4192 ext4_warning(inode->i_sb,
4193 "Unable to expand inode %lu. Delete"
4194 " some EAs or run e2fsck.",
4195 inode->i_ino);
4196 mnt_count =
4197 le16_to_cpu(sbi->s_es->s_mnt_count);
4198 }
4199 }
4200 }
4201 }
4202 if (!err)
4203 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4204 return err;
4205 }
4206
4207 /*
4208 * ext4_dirty_inode() is called from __mark_inode_dirty()
4209 *
4210 * We're really interested in the case where a file is being extended.
4211 * i_size has been changed by generic_commit_write() and we thus need
4212 * to include the updated inode in the current transaction.
4213 *
4214 * Also, dquot_alloc_block() will always dirty the inode when blocks
4215 * are allocated to the file.
4216 *
4217 * If the inode is marked synchronous, we don't honour that here - doing
4218 * so would cause a commit on atime updates, which we don't bother doing.
4219 * We handle synchronous inodes at the highest possible level.
4220 */
4221 void ext4_dirty_inode(struct inode *inode, int flags)
4222 {
4223 handle_t *handle;
4224
4225 handle = ext4_journal_start(inode, 2);
4226 if (IS_ERR(handle))
4227 goto out;
4228
4229 ext4_mark_inode_dirty(handle, inode);
4230
4231 ext4_journal_stop(handle);
4232 out:
4233 return;
4234 }
4235
4236 #if 0
4237 /*
4238 * Bind an inode's backing buffer_head into this transaction, to prevent
4239 * it from being flushed to disk early. Unlike
4240 * ext4_reserve_inode_write, this leaves behind no bh reference and
4241 * returns no iloc structure, so the caller needs to repeat the iloc
4242 * lookup to mark the inode dirty later.
4243 */
4244 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4245 {
4246 struct ext4_iloc iloc;
4247
4248 int err = 0;
4249 if (handle) {
4250 err = ext4_get_inode_loc(inode, &iloc);
4251 if (!err) {
4252 BUFFER_TRACE(iloc.bh, "get_write_access");
4253 err = jbd2_journal_get_write_access(handle, iloc.bh);
4254 if (!err)
4255 err = ext4_handle_dirty_metadata(handle,
4256 NULL,
4257 iloc.bh);
4258 brelse(iloc.bh);
4259 }
4260 }
4261 ext4_std_error(inode->i_sb, err);
4262 return err;
4263 }
4264 #endif
4265
4266 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4267 {
4268 journal_t *journal;
4269 handle_t *handle;
4270 int err;
4271
4272 /*
4273 * We have to be very careful here: changing a data block's
4274 * journaling status dynamically is dangerous. If we write a
4275 * data block to the journal, change the status and then delete
4276 * that block, we risk forgetting to revoke the old log record
4277 * from the journal and so a subsequent replay can corrupt data.
4278 * So, first we make sure that the journal is empty and that
4279 * nobody is changing anything.
4280 */
4281
4282 journal = EXT4_JOURNAL(inode);
4283 if (!journal)
4284 return 0;
4285 if (is_journal_aborted(journal))
4286 return -EROFS;
4287
4288 jbd2_journal_lock_updates(journal);
4289 jbd2_journal_flush(journal);
4290
4291 /*
4292 * OK, there are no updates running now, and all cached data is
4293 * synced to disk. We are now in a completely consistent state
4294 * which doesn't have anything in the journal, and we know that
4295 * no filesystem updates are running, so it is safe to modify
4296 * the inode's in-core data-journaling state flag now.
4297 */
4298
4299 if (val)
4300 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4301 else
4302 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4303 ext4_set_aops(inode);
4304
4305 jbd2_journal_unlock_updates(journal);
4306
4307 /* Finally we can mark the inode as dirty. */
4308
4309 handle = ext4_journal_start(inode, 1);
4310 if (IS_ERR(handle))
4311 return PTR_ERR(handle);
4312
4313 err = ext4_mark_inode_dirty(handle, inode);
4314 ext4_handle_sync(handle);
4315 ext4_journal_stop(handle);
4316 ext4_std_error(inode->i_sb, err);
4317
4318 return err;
4319 }
4320
4321 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4322 {
4323 return !buffer_mapped(bh);
4324 }
4325
4326 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
4327 {
4328 struct page *page = vmf->page;
4329 loff_t size;
4330 unsigned long len;
4331 int ret;
4332 struct file *file = vma->vm_file;
4333 struct inode *inode = file->f_path.dentry->d_inode;
4334 struct address_space *mapping = inode->i_mapping;
4335 handle_t *handle;
4336 get_block_t *get_block;
4337 int retries = 0;
4338
4339 /*
4340 * This check is racy but catches the common case. We rely on
4341 * __block_page_mkwrite() to do a reliable check.
4342 */
4343 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
4344 /* Delalloc case is easy... */
4345 if (test_opt(inode->i_sb, DELALLOC) &&
4346 !ext4_should_journal_data(inode) &&
4347 !ext4_nonda_switch(inode->i_sb)) {
4348 do {
4349 ret = __block_page_mkwrite(vma, vmf,
4350 ext4_da_get_block_prep);
4351 } while (ret == -ENOSPC &&
4352 ext4_should_retry_alloc(inode->i_sb, &retries));
4353 goto out_ret;
4354 }
4355
4356 lock_page(page);
4357 size = i_size_read(inode);
4358 /* Page got truncated from under us? */
4359 if (page->mapping != mapping || page_offset(page) > size) {
4360 unlock_page(page);
4361 ret = VM_FAULT_NOPAGE;
4362 goto out;
4363 }
4364
4365 if (page->index == size >> PAGE_CACHE_SHIFT)
4366 len = size & ~PAGE_CACHE_MASK;
4367 else
4368 len = PAGE_CACHE_SIZE;
4369 /*
4370 * Return if we have all the buffers mapped. This avoids the need to do
4371 * journal_start/journal_stop which can block and take a long time
4372 */
4373 if (page_has_buffers(page)) {
4374 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4375 ext4_bh_unmapped)) {
4376 /* Wait so that we don't change page under IO */
4377 wait_on_page_writeback(page);
4378 ret = VM_FAULT_LOCKED;
4379 goto out;
4380 }
4381 }
4382 unlock_page(page);
4383 /* OK, we need to fill the hole... */
4384 if (ext4_should_dioread_nolock(inode))
4385 get_block = ext4_get_block_write;
4386 else
4387 get_block = ext4_get_block;
4388 retry_alloc:
4389 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
4390 if (IS_ERR(handle)) {
4391 ret = VM_FAULT_SIGBUS;
4392 goto out;
4393 }
4394 ret = __block_page_mkwrite(vma, vmf, get_block);
4395 if (!ret && ext4_should_journal_data(inode)) {
4396 if (walk_page_buffers(handle, page_buffers(page), 0,
4397 PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
4398 unlock_page(page);
4399 ret = VM_FAULT_SIGBUS;
4400 goto out;
4401 }
4402 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
4403 }
4404 ext4_journal_stop(handle);
4405 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
4406 goto retry_alloc;
4407 out_ret:
4408 ret = block_page_mkwrite_return(ret);
4409 out:
4410 return ret;
4411 }