Loading...
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
52static 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
69static void ext4_invalidatepage(struct page *page, unsigned long offset);
70static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
71 struct buffer_head *bh_result, int create);
72static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
73static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
74static int __ext4_journalled_writepage(struct page *page, unsigned int len);
75static 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 */
80static 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 */
93int 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 */
117void ext4_evict_inode(struct inode *inode)
118{
119 handle_t *handle;
120 int err;
121
122 trace_ext4_evict_inode(inode);
123
124 ext4_ioend_wait(inode);
125
126 if (inode->i_nlink) {
127 /*
128 * When journalling data dirty buffers are tracked only in the
129 * journal. So although mm thinks everything is clean and
130 * ready for reaping the inode might still have some pages to
131 * write in the running transaction or waiting to be
132 * checkpointed. Thus calling jbd2_journal_invalidatepage()
133 * (via truncate_inode_pages()) to discard these buffers can
134 * cause data loss. Also even if we did not discard these
135 * buffers, we would have no way to find them after the inode
136 * is reaped and thus user could see stale data if he tries to
137 * read them before the transaction is checkpointed. So be
138 * careful and force everything to disk here... We use
139 * ei->i_datasync_tid to store the newest transaction
140 * containing inode's data.
141 *
142 * Note that directories do not have this problem because they
143 * don't use page cache.
144 */
145 if (ext4_should_journal_data(inode) &&
146 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
147 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
148 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
149
150 jbd2_log_start_commit(journal, commit_tid);
151 jbd2_log_wait_commit(journal, commit_tid);
152 filemap_write_and_wait(&inode->i_data);
153 }
154 truncate_inode_pages(&inode->i_data, 0);
155 goto no_delete;
156 }
157
158 if (!is_bad_inode(inode))
159 dquot_initialize(inode);
160
161 if (ext4_should_order_data(inode))
162 ext4_begin_ordered_truncate(inode, 0);
163 truncate_inode_pages(&inode->i_data, 0);
164
165 if (is_bad_inode(inode))
166 goto no_delete;
167
168 handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
169 if (IS_ERR(handle)) {
170 ext4_std_error(inode->i_sb, PTR_ERR(handle));
171 /*
172 * If we're going to skip the normal cleanup, we still need to
173 * make sure that the in-core orphan linked list is properly
174 * cleaned up.
175 */
176 ext4_orphan_del(NULL, inode);
177 goto no_delete;
178 }
179
180 if (IS_SYNC(inode))
181 ext4_handle_sync(handle);
182 inode->i_size = 0;
183 err = ext4_mark_inode_dirty(handle, inode);
184 if (err) {
185 ext4_warning(inode->i_sb,
186 "couldn't mark inode dirty (err %d)", err);
187 goto stop_handle;
188 }
189 if (inode->i_blocks)
190 ext4_truncate(inode);
191
192 /*
193 * ext4_ext_truncate() doesn't reserve any slop when it
194 * restarts journal transactions; therefore there may not be
195 * enough credits left in the handle to remove the inode from
196 * the orphan list and set the dtime field.
197 */
198 if (!ext4_handle_has_enough_credits(handle, 3)) {
199 err = ext4_journal_extend(handle, 3);
200 if (err > 0)
201 err = ext4_journal_restart(handle, 3);
202 if (err != 0) {
203 ext4_warning(inode->i_sb,
204 "couldn't extend journal (err %d)", err);
205 stop_handle:
206 ext4_journal_stop(handle);
207 ext4_orphan_del(NULL, inode);
208 goto no_delete;
209 }
210 }
211
212 /*
213 * Kill off the orphan record which ext4_truncate created.
214 * AKPM: I think this can be inside the above `if'.
215 * Note that ext4_orphan_del() has to be able to cope with the
216 * deletion of a non-existent orphan - this is because we don't
217 * know if ext4_truncate() actually created an orphan record.
218 * (Well, we could do this if we need to, but heck - it works)
219 */
220 ext4_orphan_del(handle, inode);
221 EXT4_I(inode)->i_dtime = get_seconds();
222
223 /*
224 * One subtle ordering requirement: if anything has gone wrong
225 * (transaction abort, IO errors, whatever), then we can still
226 * do these next steps (the fs will already have been marked as
227 * having errors), but we can't free the inode if the mark_dirty
228 * fails.
229 */
230 if (ext4_mark_inode_dirty(handle, inode))
231 /* If that failed, just do the required in-core inode clear. */
232 ext4_clear_inode(inode);
233 else
234 ext4_free_inode(handle, inode);
235 ext4_journal_stop(handle);
236 return;
237no_delete:
238 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
239}
240
241#ifdef CONFIG_QUOTA
242qsize_t *ext4_get_reserved_space(struct inode *inode)
243{
244 return &EXT4_I(inode)->i_reserved_quota;
245}
246#endif
247
248/*
249 * Calculate the number of metadata blocks need to reserve
250 * to allocate a block located at @lblock
251 */
252static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
253{
254 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
255 return ext4_ext_calc_metadata_amount(inode, lblock);
256
257 return ext4_ind_calc_metadata_amount(inode, lblock);
258}
259
260/*
261 * Called with i_data_sem down, which is important since we can call
262 * ext4_discard_preallocations() from here.
263 */
264void ext4_da_update_reserve_space(struct inode *inode,
265 int used, int quota_claim)
266{
267 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
268 struct ext4_inode_info *ei = EXT4_I(inode);
269
270 spin_lock(&ei->i_block_reservation_lock);
271 trace_ext4_da_update_reserve_space(inode, used);
272 if (unlikely(used > ei->i_reserved_data_blocks)) {
273 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
274 "with only %d reserved data blocks\n",
275 __func__, inode->i_ino, used,
276 ei->i_reserved_data_blocks);
277 WARN_ON(1);
278 used = ei->i_reserved_data_blocks;
279 }
280
281 /* Update per-inode reservations */
282 ei->i_reserved_data_blocks -= used;
283 ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
284 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
285 used + ei->i_allocated_meta_blocks);
286 ei->i_allocated_meta_blocks = 0;
287
288 if (ei->i_reserved_data_blocks == 0) {
289 /*
290 * We can release all of the reserved metadata blocks
291 * only when we have written all of the delayed
292 * allocation blocks.
293 */
294 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
295 ei->i_reserved_meta_blocks);
296 ei->i_reserved_meta_blocks = 0;
297 ei->i_da_metadata_calc_len = 0;
298 }
299 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
300
301 /* Update quota subsystem for data blocks */
302 if (quota_claim)
303 dquot_claim_block(inode, used);
304 else {
305 /*
306 * We did fallocate with an offset that is already delayed
307 * allocated. So on delayed allocated writeback we should
308 * not re-claim the quota for fallocated blocks.
309 */
310 dquot_release_reservation_block(inode, used);
311 }
312
313 /*
314 * If we have done all the pending block allocations and if
315 * there aren't any writers on the inode, we can discard the
316 * inode's preallocations.
317 */
318 if ((ei->i_reserved_data_blocks == 0) &&
319 (atomic_read(&inode->i_writecount) == 0))
320 ext4_discard_preallocations(inode);
321}
322
323static int __check_block_validity(struct inode *inode, const char *func,
324 unsigned int line,
325 struct ext4_map_blocks *map)
326{
327 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
328 map->m_len)) {
329 ext4_error_inode(inode, func, line, map->m_pblk,
330 "lblock %lu mapped to illegal pblock "
331 "(length %d)", (unsigned long) map->m_lblk,
332 map->m_len);
333 return -EIO;
334 }
335 return 0;
336}
337
338#define check_block_validity(inode, map) \
339 __check_block_validity((inode), __func__, __LINE__, (map))
340
341/*
342 * Return the number of contiguous dirty pages in a given inode
343 * starting at page frame idx.
344 */
345static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
346 unsigned int max_pages)
347{
348 struct address_space *mapping = inode->i_mapping;
349 pgoff_t index;
350 struct pagevec pvec;
351 pgoff_t num = 0;
352 int i, nr_pages, done = 0;
353
354 if (max_pages == 0)
355 return 0;
356 pagevec_init(&pvec, 0);
357 while (!done) {
358 index = idx;
359 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
360 PAGECACHE_TAG_DIRTY,
361 (pgoff_t)PAGEVEC_SIZE);
362 if (nr_pages == 0)
363 break;
364 for (i = 0; i < nr_pages; i++) {
365 struct page *page = pvec.pages[i];
366 struct buffer_head *bh, *head;
367
368 lock_page(page);
369 if (unlikely(page->mapping != mapping) ||
370 !PageDirty(page) ||
371 PageWriteback(page) ||
372 page->index != idx) {
373 done = 1;
374 unlock_page(page);
375 break;
376 }
377 if (page_has_buffers(page)) {
378 bh = head = page_buffers(page);
379 do {
380 if (!buffer_delay(bh) &&
381 !buffer_unwritten(bh))
382 done = 1;
383 bh = bh->b_this_page;
384 } while (!done && (bh != head));
385 }
386 unlock_page(page);
387 if (done)
388 break;
389 idx++;
390 num++;
391 if (num >= max_pages) {
392 done = 1;
393 break;
394 }
395 }
396 pagevec_release(&pvec);
397 }
398 return num;
399}
400
401/*
402 * The ext4_map_blocks() function tries to look up the requested blocks,
403 * and returns if the blocks are already mapped.
404 *
405 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
406 * and store the allocated blocks in the result buffer head and mark it
407 * mapped.
408 *
409 * If file type is extents based, it will call ext4_ext_map_blocks(),
410 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
411 * based files
412 *
413 * On success, it returns the number of blocks being mapped or allocate.
414 * if create==0 and the blocks are pre-allocated and uninitialized block,
415 * the result buffer head is unmapped. If the create ==1, it will make sure
416 * the buffer head is mapped.
417 *
418 * It returns 0 if plain look up failed (blocks have not been allocated), in
419 * that casem, buffer head is unmapped
420 *
421 * It returns the error in case of allocation failure.
422 */
423int ext4_map_blocks(handle_t *handle, struct inode *inode,
424 struct ext4_map_blocks *map, int flags)
425{
426 int retval;
427
428 map->m_flags = 0;
429 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
430 "logical block %lu\n", inode->i_ino, flags, map->m_len,
431 (unsigned long) map->m_lblk);
432 /*
433 * Try to see if we can get the block without requesting a new
434 * file system block.
435 */
436 down_read((&EXT4_I(inode)->i_data_sem));
437 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
438 retval = ext4_ext_map_blocks(handle, inode, map, 0);
439 } else {
440 retval = ext4_ind_map_blocks(handle, inode, map, 0);
441 }
442 up_read((&EXT4_I(inode)->i_data_sem));
443
444 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
445 int ret = check_block_validity(inode, map);
446 if (ret != 0)
447 return ret;
448 }
449
450 /* If it is only a block(s) look up */
451 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
452 return retval;
453
454 /*
455 * Returns if the blocks have already allocated
456 *
457 * Note that if blocks have been preallocated
458 * ext4_ext_get_block() returns th create = 0
459 * with buffer head unmapped.
460 */
461 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
462 return retval;
463
464 /*
465 * When we call get_blocks without the create flag, the
466 * BH_Unwritten flag could have gotten set if the blocks
467 * requested were part of a uninitialized extent. We need to
468 * clear this flag now that we are committed to convert all or
469 * part of the uninitialized extent to be an initialized
470 * extent. This is because we need to avoid the combination
471 * of BH_Unwritten and BH_Mapped flags being simultaneously
472 * set on the buffer_head.
473 */
474 map->m_flags &= ~EXT4_MAP_UNWRITTEN;
475
476 /*
477 * New blocks allocate and/or writing to uninitialized extent
478 * will possibly result in updating i_data, so we take
479 * the write lock of i_data_sem, and call get_blocks()
480 * with create == 1 flag.
481 */
482 down_write((&EXT4_I(inode)->i_data_sem));
483
484 /*
485 * if the caller is from delayed allocation writeout path
486 * we have already reserved fs blocks for allocation
487 * let the underlying get_block() function know to
488 * avoid double accounting
489 */
490 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
491 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
492 /*
493 * We need to check for EXT4 here because migrate
494 * could have changed the inode type in between
495 */
496 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
497 retval = ext4_ext_map_blocks(handle, inode, map, flags);
498 } else {
499 retval = ext4_ind_map_blocks(handle, inode, map, flags);
500
501 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
502 /*
503 * We allocated new blocks which will result in
504 * i_data's format changing. Force the migrate
505 * to fail by clearing migrate flags
506 */
507 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
508 }
509
510 /*
511 * Update reserved blocks/metadata blocks after successful
512 * block allocation which had been deferred till now. We don't
513 * support fallocate for non extent files. So we can update
514 * reserve space here.
515 */
516 if ((retval > 0) &&
517 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
518 ext4_da_update_reserve_space(inode, retval, 1);
519 }
520 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
521 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
522
523 up_write((&EXT4_I(inode)->i_data_sem));
524 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
525 int ret = check_block_validity(inode, map);
526 if (ret != 0)
527 return ret;
528 }
529 return retval;
530}
531
532/* Maximum number of blocks we map for direct IO at once. */
533#define DIO_MAX_BLOCKS 4096
534
535static int _ext4_get_block(struct inode *inode, sector_t iblock,
536 struct buffer_head *bh, int flags)
537{
538 handle_t *handle = ext4_journal_current_handle();
539 struct ext4_map_blocks map;
540 int ret = 0, started = 0;
541 int dio_credits;
542
543 map.m_lblk = iblock;
544 map.m_len = bh->b_size >> inode->i_blkbits;
545
546 if (flags && !handle) {
547 /* Direct IO write... */
548 if (map.m_len > DIO_MAX_BLOCKS)
549 map.m_len = DIO_MAX_BLOCKS;
550 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
551 handle = ext4_journal_start(inode, dio_credits);
552 if (IS_ERR(handle)) {
553 ret = PTR_ERR(handle);
554 return ret;
555 }
556 started = 1;
557 }
558
559 ret = ext4_map_blocks(handle, inode, &map, flags);
560 if (ret > 0) {
561 map_bh(bh, inode->i_sb, map.m_pblk);
562 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
563 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
564 ret = 0;
565 }
566 if (started)
567 ext4_journal_stop(handle);
568 return ret;
569}
570
571int ext4_get_block(struct inode *inode, sector_t iblock,
572 struct buffer_head *bh, int create)
573{
574 return _ext4_get_block(inode, iblock, bh,
575 create ? EXT4_GET_BLOCKS_CREATE : 0);
576}
577
578/*
579 * `handle' can be NULL if create is zero
580 */
581struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
582 ext4_lblk_t block, int create, int *errp)
583{
584 struct ext4_map_blocks map;
585 struct buffer_head *bh;
586 int fatal = 0, err;
587
588 J_ASSERT(handle != NULL || create == 0);
589
590 map.m_lblk = block;
591 map.m_len = 1;
592 err = ext4_map_blocks(handle, inode, &map,
593 create ? EXT4_GET_BLOCKS_CREATE : 0);
594
595 if (err < 0)
596 *errp = err;
597 if (err <= 0)
598 return NULL;
599 *errp = 0;
600
601 bh = sb_getblk(inode->i_sb, map.m_pblk);
602 if (!bh) {
603 *errp = -EIO;
604 return NULL;
605 }
606 if (map.m_flags & EXT4_MAP_NEW) {
607 J_ASSERT(create != 0);
608 J_ASSERT(handle != NULL);
609
610 /*
611 * Now that we do not always journal data, we should
612 * keep in mind whether this should always journal the
613 * new buffer as metadata. For now, regular file
614 * writes use ext4_get_block instead, so it's not a
615 * problem.
616 */
617 lock_buffer(bh);
618 BUFFER_TRACE(bh, "call get_create_access");
619 fatal = ext4_journal_get_create_access(handle, bh);
620 if (!fatal && !buffer_uptodate(bh)) {
621 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
622 set_buffer_uptodate(bh);
623 }
624 unlock_buffer(bh);
625 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
626 err = ext4_handle_dirty_metadata(handle, inode, bh);
627 if (!fatal)
628 fatal = err;
629 } else {
630 BUFFER_TRACE(bh, "not a new buffer");
631 }
632 if (fatal) {
633 *errp = fatal;
634 brelse(bh);
635 bh = NULL;
636 }
637 return bh;
638}
639
640struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
641 ext4_lblk_t block, int create, int *err)
642{
643 struct buffer_head *bh;
644
645 bh = ext4_getblk(handle, inode, block, create, err);
646 if (!bh)
647 return bh;
648 if (buffer_uptodate(bh))
649 return bh;
650 ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
651 wait_on_buffer(bh);
652 if (buffer_uptodate(bh))
653 return bh;
654 put_bh(bh);
655 *err = -EIO;
656 return NULL;
657}
658
659static int walk_page_buffers(handle_t *handle,
660 struct buffer_head *head,
661 unsigned from,
662 unsigned to,
663 int *partial,
664 int (*fn)(handle_t *handle,
665 struct buffer_head *bh))
666{
667 struct buffer_head *bh;
668 unsigned block_start, block_end;
669 unsigned blocksize = head->b_size;
670 int err, ret = 0;
671 struct buffer_head *next;
672
673 for (bh = head, block_start = 0;
674 ret == 0 && (bh != head || !block_start);
675 block_start = block_end, bh = next) {
676 next = bh->b_this_page;
677 block_end = block_start + blocksize;
678 if (block_end <= from || block_start >= to) {
679 if (partial && !buffer_uptodate(bh))
680 *partial = 1;
681 continue;
682 }
683 err = (*fn)(handle, bh);
684 if (!ret)
685 ret = err;
686 }
687 return ret;
688}
689
690/*
691 * To preserve ordering, it is essential that the hole instantiation and
692 * the data write be encapsulated in a single transaction. We cannot
693 * close off a transaction and start a new one between the ext4_get_block()
694 * and the commit_write(). So doing the jbd2_journal_start at the start of
695 * prepare_write() is the right place.
696 *
697 * Also, this function can nest inside ext4_writepage() ->
698 * block_write_full_page(). In that case, we *know* that ext4_writepage()
699 * has generated enough buffer credits to do the whole page. So we won't
700 * block on the journal in that case, which is good, because the caller may
701 * be PF_MEMALLOC.
702 *
703 * By accident, ext4 can be reentered when a transaction is open via
704 * quota file writes. If we were to commit the transaction while thus
705 * reentered, there can be a deadlock - we would be holding a quota
706 * lock, and the commit would never complete if another thread had a
707 * transaction open and was blocking on the quota lock - a ranking
708 * violation.
709 *
710 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
711 * will _not_ run commit under these circumstances because handle->h_ref
712 * is elevated. We'll still have enough credits for the tiny quotafile
713 * write.
714 */
715static int do_journal_get_write_access(handle_t *handle,
716 struct buffer_head *bh)
717{
718 int dirty = buffer_dirty(bh);
719 int ret;
720
721 if (!buffer_mapped(bh) || buffer_freed(bh))
722 return 0;
723 /*
724 * __block_write_begin() could have dirtied some buffers. Clean
725 * the dirty bit as jbd2_journal_get_write_access() could complain
726 * otherwise about fs integrity issues. Setting of the dirty bit
727 * by __block_write_begin() isn't a real problem here as we clear
728 * the bit before releasing a page lock and thus writeback cannot
729 * ever write the buffer.
730 */
731 if (dirty)
732 clear_buffer_dirty(bh);
733 ret = ext4_journal_get_write_access(handle, bh);
734 if (!ret && dirty)
735 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
736 return ret;
737}
738
739static int ext4_get_block_write(struct inode *inode, sector_t iblock,
740 struct buffer_head *bh_result, int create);
741static int ext4_write_begin(struct file *file, struct address_space *mapping,
742 loff_t pos, unsigned len, unsigned flags,
743 struct page **pagep, void **fsdata)
744{
745 struct inode *inode = mapping->host;
746 int ret, needed_blocks;
747 handle_t *handle;
748 int retries = 0;
749 struct page *page;
750 pgoff_t index;
751 unsigned from, to;
752
753 trace_ext4_write_begin(inode, pos, len, flags);
754 /*
755 * Reserve one block more for addition to orphan list in case
756 * we allocate blocks but write fails for some reason
757 */
758 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
759 index = pos >> PAGE_CACHE_SHIFT;
760 from = pos & (PAGE_CACHE_SIZE - 1);
761 to = from + len;
762
763retry:
764 handle = ext4_journal_start(inode, needed_blocks);
765 if (IS_ERR(handle)) {
766 ret = PTR_ERR(handle);
767 goto out;
768 }
769
770 /* We cannot recurse into the filesystem as the transaction is already
771 * started */
772 flags |= AOP_FLAG_NOFS;
773
774 page = grab_cache_page_write_begin(mapping, index, flags);
775 if (!page) {
776 ext4_journal_stop(handle);
777 ret = -ENOMEM;
778 goto out;
779 }
780 *pagep = page;
781
782 if (ext4_should_dioread_nolock(inode))
783 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
784 else
785 ret = __block_write_begin(page, pos, len, ext4_get_block);
786
787 if (!ret && ext4_should_journal_data(inode)) {
788 ret = walk_page_buffers(handle, page_buffers(page),
789 from, to, NULL, do_journal_get_write_access);
790 }
791
792 if (ret) {
793 unlock_page(page);
794 page_cache_release(page);
795 /*
796 * __block_write_begin may have instantiated a few blocks
797 * outside i_size. Trim these off again. Don't need
798 * i_size_read because we hold i_mutex.
799 *
800 * Add inode to orphan list in case we crash before
801 * truncate finishes
802 */
803 if (pos + len > inode->i_size && ext4_can_truncate(inode))
804 ext4_orphan_add(handle, inode);
805
806 ext4_journal_stop(handle);
807 if (pos + len > inode->i_size) {
808 ext4_truncate_failed_write(inode);
809 /*
810 * If truncate failed early the inode might
811 * still be on the orphan list; we need to
812 * make sure the inode is removed from the
813 * orphan list in that case.
814 */
815 if (inode->i_nlink)
816 ext4_orphan_del(NULL, inode);
817 }
818 }
819
820 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
821 goto retry;
822out:
823 return ret;
824}
825
826/* For write_end() in data=journal mode */
827static int write_end_fn(handle_t *handle, struct buffer_head *bh)
828{
829 if (!buffer_mapped(bh) || buffer_freed(bh))
830 return 0;
831 set_buffer_uptodate(bh);
832 return ext4_handle_dirty_metadata(handle, NULL, bh);
833}
834
835static int ext4_generic_write_end(struct file *file,
836 struct address_space *mapping,
837 loff_t pos, unsigned len, unsigned copied,
838 struct page *page, void *fsdata)
839{
840 int i_size_changed = 0;
841 struct inode *inode = mapping->host;
842 handle_t *handle = ext4_journal_current_handle();
843
844 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
845
846 /*
847 * No need to use i_size_read() here, the i_size
848 * cannot change under us because we hold i_mutex.
849 *
850 * But it's important to update i_size while still holding page lock:
851 * page writeout could otherwise come in and zero beyond i_size.
852 */
853 if (pos + copied > inode->i_size) {
854 i_size_write(inode, pos + copied);
855 i_size_changed = 1;
856 }
857
858 if (pos + copied > EXT4_I(inode)->i_disksize) {
859 /* We need to mark inode dirty even if
860 * new_i_size is less that inode->i_size
861 * bu greater than i_disksize.(hint delalloc)
862 */
863 ext4_update_i_disksize(inode, (pos + copied));
864 i_size_changed = 1;
865 }
866 unlock_page(page);
867 page_cache_release(page);
868
869 /*
870 * Don't mark the inode dirty under page lock. First, it unnecessarily
871 * makes the holding time of page lock longer. Second, it forces lock
872 * ordering of page lock and transaction start for journaling
873 * filesystems.
874 */
875 if (i_size_changed)
876 ext4_mark_inode_dirty(handle, inode);
877
878 return copied;
879}
880
881/*
882 * We need to pick up the new inode size which generic_commit_write gave us
883 * `file' can be NULL - eg, when called from page_symlink().
884 *
885 * ext4 never places buffers on inode->i_mapping->private_list. metadata
886 * buffers are managed internally.
887 */
888static int ext4_ordered_write_end(struct file *file,
889 struct address_space *mapping,
890 loff_t pos, unsigned len, unsigned copied,
891 struct page *page, void *fsdata)
892{
893 handle_t *handle = ext4_journal_current_handle();
894 struct inode *inode = mapping->host;
895 int ret = 0, ret2;
896
897 trace_ext4_ordered_write_end(inode, pos, len, copied);
898 ret = ext4_jbd2_file_inode(handle, inode);
899
900 if (ret == 0) {
901 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
902 page, fsdata);
903 copied = ret2;
904 if (pos + len > inode->i_size && ext4_can_truncate(inode))
905 /* if we have allocated more blocks and copied
906 * less. We will have blocks allocated outside
907 * inode->i_size. So truncate them
908 */
909 ext4_orphan_add(handle, inode);
910 if (ret2 < 0)
911 ret = ret2;
912 }
913 ret2 = ext4_journal_stop(handle);
914 if (!ret)
915 ret = ret2;
916
917 if (pos + len > inode->i_size) {
918 ext4_truncate_failed_write(inode);
919 /*
920 * If truncate failed early the inode might still be
921 * on the orphan list; we need to make sure the inode
922 * is removed from the orphan list in that case.
923 */
924 if (inode->i_nlink)
925 ext4_orphan_del(NULL, inode);
926 }
927
928
929 return ret ? ret : copied;
930}
931
932static int ext4_writeback_write_end(struct file *file,
933 struct address_space *mapping,
934 loff_t pos, unsigned len, unsigned copied,
935 struct page *page, void *fsdata)
936{
937 handle_t *handle = ext4_journal_current_handle();
938 struct inode *inode = mapping->host;
939 int ret = 0, ret2;
940
941 trace_ext4_writeback_write_end(inode, pos, len, copied);
942 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
943 page, fsdata);
944 copied = ret2;
945 if (pos + len > inode->i_size && ext4_can_truncate(inode))
946 /* if we have allocated more blocks and copied
947 * less. We will have blocks allocated outside
948 * inode->i_size. So truncate them
949 */
950 ext4_orphan_add(handle, inode);
951
952 if (ret2 < 0)
953 ret = ret2;
954
955 ret2 = ext4_journal_stop(handle);
956 if (!ret)
957 ret = ret2;
958
959 if (pos + len > inode->i_size) {
960 ext4_truncate_failed_write(inode);
961 /*
962 * If truncate failed early the inode might still be
963 * on the orphan list; we need to make sure the inode
964 * is removed from the orphan list in that case.
965 */
966 if (inode->i_nlink)
967 ext4_orphan_del(NULL, inode);
968 }
969
970 return ret ? ret : copied;
971}
972
973static int ext4_journalled_write_end(struct file *file,
974 struct address_space *mapping,
975 loff_t pos, unsigned len, unsigned copied,
976 struct page *page, void *fsdata)
977{
978 handle_t *handle = ext4_journal_current_handle();
979 struct inode *inode = mapping->host;
980 int ret = 0, ret2;
981 int partial = 0;
982 unsigned from, to;
983 loff_t new_i_size;
984
985 trace_ext4_journalled_write_end(inode, pos, len, copied);
986 from = pos & (PAGE_CACHE_SIZE - 1);
987 to = from + len;
988
989 BUG_ON(!ext4_handle_valid(handle));
990
991 if (copied < len) {
992 if (!PageUptodate(page))
993 copied = 0;
994 page_zero_new_buffers(page, from+copied, to);
995 }
996
997 ret = walk_page_buffers(handle, page_buffers(page), from,
998 to, &partial, write_end_fn);
999 if (!partial)
1000 SetPageUptodate(page);
1001 new_i_size = pos + copied;
1002 if (new_i_size > inode->i_size)
1003 i_size_write(inode, pos+copied);
1004 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1005 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1006 if (new_i_size > EXT4_I(inode)->i_disksize) {
1007 ext4_update_i_disksize(inode, new_i_size);
1008 ret2 = ext4_mark_inode_dirty(handle, inode);
1009 if (!ret)
1010 ret = ret2;
1011 }
1012
1013 unlock_page(page);
1014 page_cache_release(page);
1015 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1016 /* if we have allocated more blocks and copied
1017 * less. We will have blocks allocated outside
1018 * inode->i_size. So truncate them
1019 */
1020 ext4_orphan_add(handle, inode);
1021
1022 ret2 = ext4_journal_stop(handle);
1023 if (!ret)
1024 ret = ret2;
1025 if (pos + len > inode->i_size) {
1026 ext4_truncate_failed_write(inode);
1027 /*
1028 * If truncate failed early the inode might still be
1029 * on the orphan list; we need to make sure the inode
1030 * is removed from the orphan list in that case.
1031 */
1032 if (inode->i_nlink)
1033 ext4_orphan_del(NULL, inode);
1034 }
1035
1036 return ret ? ret : copied;
1037}
1038
1039/*
1040 * Reserve a single block located at lblock
1041 */
1042static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1043{
1044 int retries = 0;
1045 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1046 struct ext4_inode_info *ei = EXT4_I(inode);
1047 unsigned long md_needed;
1048 int ret;
1049
1050 /*
1051 * recalculate the amount of metadata blocks to reserve
1052 * in order to allocate nrblocks
1053 * worse case is one extent per block
1054 */
1055repeat:
1056 spin_lock(&ei->i_block_reservation_lock);
1057 md_needed = ext4_calc_metadata_amount(inode, lblock);
1058 trace_ext4_da_reserve_space(inode, md_needed);
1059 spin_unlock(&ei->i_block_reservation_lock);
1060
1061 /*
1062 * We will charge metadata quota at writeout time; this saves
1063 * us from metadata over-estimation, though we may go over by
1064 * a small amount in the end. Here we just reserve for data.
1065 */
1066 ret = dquot_reserve_block(inode, 1);
1067 if (ret)
1068 return ret;
1069 /*
1070 * We do still charge estimated metadata to the sb though;
1071 * we cannot afford to run out of free blocks.
1072 */
1073 if (ext4_claim_free_blocks(sbi, md_needed + 1, 0)) {
1074 dquot_release_reservation_block(inode, 1);
1075 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1076 yield();
1077 goto repeat;
1078 }
1079 return -ENOSPC;
1080 }
1081 spin_lock(&ei->i_block_reservation_lock);
1082 ei->i_reserved_data_blocks++;
1083 ei->i_reserved_meta_blocks += md_needed;
1084 spin_unlock(&ei->i_block_reservation_lock);
1085
1086 return 0; /* success */
1087}
1088
1089static void ext4_da_release_space(struct inode *inode, int to_free)
1090{
1091 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1092 struct ext4_inode_info *ei = EXT4_I(inode);
1093
1094 if (!to_free)
1095 return; /* Nothing to release, exit */
1096
1097 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1098
1099 trace_ext4_da_release_space(inode, to_free);
1100 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1101 /*
1102 * if there aren't enough reserved blocks, then the
1103 * counter is messed up somewhere. Since this
1104 * function is called from invalidate page, it's
1105 * harmless to return without any action.
1106 */
1107 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1108 "ino %lu, to_free %d with only %d reserved "
1109 "data blocks\n", inode->i_ino, to_free,
1110 ei->i_reserved_data_blocks);
1111 WARN_ON(1);
1112 to_free = ei->i_reserved_data_blocks;
1113 }
1114 ei->i_reserved_data_blocks -= to_free;
1115
1116 if (ei->i_reserved_data_blocks == 0) {
1117 /*
1118 * We can release all of the reserved metadata blocks
1119 * only when we have written all of the delayed
1120 * allocation blocks.
1121 */
1122 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1123 ei->i_reserved_meta_blocks);
1124 ei->i_reserved_meta_blocks = 0;
1125 ei->i_da_metadata_calc_len = 0;
1126 }
1127
1128 /* update fs dirty data blocks counter */
1129 percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1130
1131 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1132
1133 dquot_release_reservation_block(inode, to_free);
1134}
1135
1136static void ext4_da_page_release_reservation(struct page *page,
1137 unsigned long offset)
1138{
1139 int to_release = 0;
1140 struct buffer_head *head, *bh;
1141 unsigned int curr_off = 0;
1142
1143 head = page_buffers(page);
1144 bh = head;
1145 do {
1146 unsigned int next_off = curr_off + bh->b_size;
1147
1148 if ((offset <= curr_off) && (buffer_delay(bh))) {
1149 to_release++;
1150 clear_buffer_delay(bh);
1151 }
1152 curr_off = next_off;
1153 } while ((bh = bh->b_this_page) != head);
1154 ext4_da_release_space(page->mapping->host, to_release);
1155}
1156
1157/*
1158 * Delayed allocation stuff
1159 */
1160
1161/*
1162 * mpage_da_submit_io - walks through extent of pages and try to write
1163 * them with writepage() call back
1164 *
1165 * @mpd->inode: inode
1166 * @mpd->first_page: first page of the extent
1167 * @mpd->next_page: page after the last page of the extent
1168 *
1169 * By the time mpage_da_submit_io() is called we expect all blocks
1170 * to be allocated. this may be wrong if allocation failed.
1171 *
1172 * As pages are already locked by write_cache_pages(), we can't use it
1173 */
1174static int mpage_da_submit_io(struct mpage_da_data *mpd,
1175 struct ext4_map_blocks *map)
1176{
1177 struct pagevec pvec;
1178 unsigned long index, end;
1179 int ret = 0, err, nr_pages, i;
1180 struct inode *inode = mpd->inode;
1181 struct address_space *mapping = inode->i_mapping;
1182 loff_t size = i_size_read(inode);
1183 unsigned int len, block_start;
1184 struct buffer_head *bh, *page_bufs = NULL;
1185 int journal_data = ext4_should_journal_data(inode);
1186 sector_t pblock = 0, cur_logical = 0;
1187 struct ext4_io_submit io_submit;
1188
1189 BUG_ON(mpd->next_page <= mpd->first_page);
1190 memset(&io_submit, 0, sizeof(io_submit));
1191 /*
1192 * We need to start from the first_page to the next_page - 1
1193 * to make sure we also write the mapped dirty buffer_heads.
1194 * If we look at mpd->b_blocknr we would only be looking
1195 * at the currently mapped buffer_heads.
1196 */
1197 index = mpd->first_page;
1198 end = mpd->next_page - 1;
1199
1200 pagevec_init(&pvec, 0);
1201 while (index <= end) {
1202 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1203 if (nr_pages == 0)
1204 break;
1205 for (i = 0; i < nr_pages; i++) {
1206 int commit_write = 0, skip_page = 0;
1207 struct page *page = pvec.pages[i];
1208
1209 index = page->index;
1210 if (index > end)
1211 break;
1212
1213 if (index == size >> PAGE_CACHE_SHIFT)
1214 len = size & ~PAGE_CACHE_MASK;
1215 else
1216 len = PAGE_CACHE_SIZE;
1217 if (map) {
1218 cur_logical = index << (PAGE_CACHE_SHIFT -
1219 inode->i_blkbits);
1220 pblock = map->m_pblk + (cur_logical -
1221 map->m_lblk);
1222 }
1223 index++;
1224
1225 BUG_ON(!PageLocked(page));
1226 BUG_ON(PageWriteback(page));
1227
1228 /*
1229 * If the page does not have buffers (for
1230 * whatever reason), try to create them using
1231 * __block_write_begin. If this fails,
1232 * skip the page and move on.
1233 */
1234 if (!page_has_buffers(page)) {
1235 if (__block_write_begin(page, 0, len,
1236 noalloc_get_block_write)) {
1237 skip_page:
1238 unlock_page(page);
1239 continue;
1240 }
1241 commit_write = 1;
1242 }
1243
1244 bh = page_bufs = page_buffers(page);
1245 block_start = 0;
1246 do {
1247 if (!bh)
1248 goto skip_page;
1249 if (map && (cur_logical >= map->m_lblk) &&
1250 (cur_logical <= (map->m_lblk +
1251 (map->m_len - 1)))) {
1252 if (buffer_delay(bh)) {
1253 clear_buffer_delay(bh);
1254 bh->b_blocknr = pblock;
1255 }
1256 if (buffer_unwritten(bh) ||
1257 buffer_mapped(bh))
1258 BUG_ON(bh->b_blocknr != pblock);
1259 if (map->m_flags & EXT4_MAP_UNINIT)
1260 set_buffer_uninit(bh);
1261 clear_buffer_unwritten(bh);
1262 }
1263
1264 /* skip page if block allocation undone */
1265 if (buffer_delay(bh) || buffer_unwritten(bh))
1266 skip_page = 1;
1267 bh = bh->b_this_page;
1268 block_start += bh->b_size;
1269 cur_logical++;
1270 pblock++;
1271 } while (bh != page_bufs);
1272
1273 if (skip_page)
1274 goto skip_page;
1275
1276 if (commit_write)
1277 /* mark the buffer_heads as dirty & uptodate */
1278 block_commit_write(page, 0, len);
1279
1280 clear_page_dirty_for_io(page);
1281 /*
1282 * Delalloc doesn't support data journalling,
1283 * but eventually maybe we'll lift this
1284 * restriction.
1285 */
1286 if (unlikely(journal_data && PageChecked(page)))
1287 err = __ext4_journalled_writepage(page, len);
1288 else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1289 err = ext4_bio_write_page(&io_submit, page,
1290 len, mpd->wbc);
1291 else if (buffer_uninit(page_bufs)) {
1292 ext4_set_bh_endio(page_bufs, inode);
1293 err = block_write_full_page_endio(page,
1294 noalloc_get_block_write,
1295 mpd->wbc, ext4_end_io_buffer_write);
1296 } else
1297 err = block_write_full_page(page,
1298 noalloc_get_block_write, mpd->wbc);
1299
1300 if (!err)
1301 mpd->pages_written++;
1302 /*
1303 * In error case, we have to continue because
1304 * remaining pages are still locked
1305 */
1306 if (ret == 0)
1307 ret = err;
1308 }
1309 pagevec_release(&pvec);
1310 }
1311 ext4_io_submit(&io_submit);
1312 return ret;
1313}
1314
1315static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1316{
1317 int nr_pages, i;
1318 pgoff_t index, end;
1319 struct pagevec pvec;
1320 struct inode *inode = mpd->inode;
1321 struct address_space *mapping = inode->i_mapping;
1322
1323 index = mpd->first_page;
1324 end = mpd->next_page - 1;
1325 while (index <= end) {
1326 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1327 if (nr_pages == 0)
1328 break;
1329 for (i = 0; i < nr_pages; i++) {
1330 struct page *page = pvec.pages[i];
1331 if (page->index > end)
1332 break;
1333 BUG_ON(!PageLocked(page));
1334 BUG_ON(PageWriteback(page));
1335 block_invalidatepage(page, 0);
1336 ClearPageUptodate(page);
1337 unlock_page(page);
1338 }
1339 index = pvec.pages[nr_pages - 1]->index + 1;
1340 pagevec_release(&pvec);
1341 }
1342 return;
1343}
1344
1345static void ext4_print_free_blocks(struct inode *inode)
1346{
1347 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1348 printk(KERN_CRIT "Total free blocks count %lld\n",
1349 ext4_count_free_blocks(inode->i_sb));
1350 printk(KERN_CRIT "Free/Dirty block details\n");
1351 printk(KERN_CRIT "free_blocks=%lld\n",
1352 (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
1353 printk(KERN_CRIT "dirty_blocks=%lld\n",
1354 (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1355 printk(KERN_CRIT "Block reservation details\n");
1356 printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1357 EXT4_I(inode)->i_reserved_data_blocks);
1358 printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1359 EXT4_I(inode)->i_reserved_meta_blocks);
1360 return;
1361}
1362
1363/*
1364 * mpage_da_map_and_submit - go through given space, map them
1365 * if necessary, and then submit them for I/O
1366 *
1367 * @mpd - bh describing space
1368 *
1369 * The function skips space we know is already mapped to disk blocks.
1370 *
1371 */
1372static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1373{
1374 int err, blks, get_blocks_flags;
1375 struct ext4_map_blocks map, *mapp = NULL;
1376 sector_t next = mpd->b_blocknr;
1377 unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1378 loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1379 handle_t *handle = NULL;
1380
1381 /*
1382 * If the blocks are mapped already, or we couldn't accumulate
1383 * any blocks, then proceed immediately to the submission stage.
1384 */
1385 if ((mpd->b_size == 0) ||
1386 ((mpd->b_state & (1 << BH_Mapped)) &&
1387 !(mpd->b_state & (1 << BH_Delay)) &&
1388 !(mpd->b_state & (1 << BH_Unwritten))))
1389 goto submit_io;
1390
1391 handle = ext4_journal_current_handle();
1392 BUG_ON(!handle);
1393
1394 /*
1395 * Call ext4_map_blocks() to allocate any delayed allocation
1396 * blocks, or to convert an uninitialized extent to be
1397 * initialized (in the case where we have written into
1398 * one or more preallocated blocks).
1399 *
1400 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1401 * indicate that we are on the delayed allocation path. This
1402 * affects functions in many different parts of the allocation
1403 * call path. This flag exists primarily because we don't
1404 * want to change *many* call functions, so ext4_map_blocks()
1405 * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1406 * inode's allocation semaphore is taken.
1407 *
1408 * If the blocks in questions were delalloc blocks, set
1409 * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1410 * variables are updated after the blocks have been allocated.
1411 */
1412 map.m_lblk = next;
1413 map.m_len = max_blocks;
1414 get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1415 if (ext4_should_dioread_nolock(mpd->inode))
1416 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1417 if (mpd->b_state & (1 << BH_Delay))
1418 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1419
1420 blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1421 if (blks < 0) {
1422 struct super_block *sb = mpd->inode->i_sb;
1423
1424 err = blks;
1425 /*
1426 * If get block returns EAGAIN or ENOSPC and there
1427 * appears to be free blocks we will just let
1428 * mpage_da_submit_io() unlock all of the pages.
1429 */
1430 if (err == -EAGAIN)
1431 goto submit_io;
1432
1433 if (err == -ENOSPC &&
1434 ext4_count_free_blocks(sb)) {
1435 mpd->retval = err;
1436 goto submit_io;
1437 }
1438
1439 /*
1440 * get block failure will cause us to loop in
1441 * writepages, because a_ops->writepage won't be able
1442 * to make progress. The page will be redirtied by
1443 * writepage and writepages will again try to write
1444 * the same.
1445 */
1446 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1447 ext4_msg(sb, KERN_CRIT,
1448 "delayed block allocation failed for inode %lu "
1449 "at logical offset %llu with max blocks %zd "
1450 "with error %d", mpd->inode->i_ino,
1451 (unsigned long long) next,
1452 mpd->b_size >> mpd->inode->i_blkbits, err);
1453 ext4_msg(sb, KERN_CRIT,
1454 "This should not happen!! Data will be lost\n");
1455 if (err == -ENOSPC)
1456 ext4_print_free_blocks(mpd->inode);
1457 }
1458 /* invalidate all the pages */
1459 ext4_da_block_invalidatepages(mpd);
1460
1461 /* Mark this page range as having been completed */
1462 mpd->io_done = 1;
1463 return;
1464 }
1465 BUG_ON(blks == 0);
1466
1467 mapp = ↦
1468 if (map.m_flags & EXT4_MAP_NEW) {
1469 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1470 int i;
1471
1472 for (i = 0; i < map.m_len; i++)
1473 unmap_underlying_metadata(bdev, map.m_pblk + i);
1474 }
1475
1476 if (ext4_should_order_data(mpd->inode)) {
1477 err = ext4_jbd2_file_inode(handle, mpd->inode);
1478 if (err)
1479 /* This only happens if the journal is aborted */
1480 return;
1481 }
1482
1483 /*
1484 * Update on-disk size along with block allocation.
1485 */
1486 disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1487 if (disksize > i_size_read(mpd->inode))
1488 disksize = i_size_read(mpd->inode);
1489 if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1490 ext4_update_i_disksize(mpd->inode, disksize);
1491 err = ext4_mark_inode_dirty(handle, mpd->inode);
1492 if (err)
1493 ext4_error(mpd->inode->i_sb,
1494 "Failed to mark inode %lu dirty",
1495 mpd->inode->i_ino);
1496 }
1497
1498submit_io:
1499 mpage_da_submit_io(mpd, mapp);
1500 mpd->io_done = 1;
1501}
1502
1503#define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1504 (1 << BH_Delay) | (1 << BH_Unwritten))
1505
1506/*
1507 * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1508 *
1509 * @mpd->lbh - extent of blocks
1510 * @logical - logical number of the block in the file
1511 * @bh - bh of the block (used to access block's state)
1512 *
1513 * the function is used to collect contig. blocks in same state
1514 */
1515static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1516 sector_t logical, size_t b_size,
1517 unsigned long b_state)
1518{
1519 sector_t next;
1520 int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1521
1522 /*
1523 * XXX Don't go larger than mballoc is willing to allocate
1524 * This is a stopgap solution. We eventually need to fold
1525 * mpage_da_submit_io() into this function and then call
1526 * ext4_map_blocks() multiple times in a loop
1527 */
1528 if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1529 goto flush_it;
1530
1531 /* check if thereserved journal credits might overflow */
1532 if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1533 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1534 /*
1535 * With non-extent format we are limited by the journal
1536 * credit available. Total credit needed to insert
1537 * nrblocks contiguous blocks is dependent on the
1538 * nrblocks. So limit nrblocks.
1539 */
1540 goto flush_it;
1541 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1542 EXT4_MAX_TRANS_DATA) {
1543 /*
1544 * Adding the new buffer_head would make it cross the
1545 * allowed limit for which we have journal credit
1546 * reserved. So limit the new bh->b_size
1547 */
1548 b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1549 mpd->inode->i_blkbits;
1550 /* we will do mpage_da_submit_io in the next loop */
1551 }
1552 }
1553 /*
1554 * First block in the extent
1555 */
1556 if (mpd->b_size == 0) {
1557 mpd->b_blocknr = logical;
1558 mpd->b_size = b_size;
1559 mpd->b_state = b_state & BH_FLAGS;
1560 return;
1561 }
1562
1563 next = mpd->b_blocknr + nrblocks;
1564 /*
1565 * Can we merge the block to our big extent?
1566 */
1567 if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1568 mpd->b_size += b_size;
1569 return;
1570 }
1571
1572flush_it:
1573 /*
1574 * We couldn't merge the block to our extent, so we
1575 * need to flush current extent and start new one
1576 */
1577 mpage_da_map_and_submit(mpd);
1578 return;
1579}
1580
1581static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1582{
1583 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1584}
1585
1586/*
1587 * This is a special get_blocks_t callback which is used by
1588 * ext4_da_write_begin(). It will either return mapped block or
1589 * reserve space for a single block.
1590 *
1591 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1592 * We also have b_blocknr = -1 and b_bdev initialized properly
1593 *
1594 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1595 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1596 * initialized properly.
1597 */
1598static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1599 struct buffer_head *bh, int create)
1600{
1601 struct ext4_map_blocks map;
1602 int ret = 0;
1603 sector_t invalid_block = ~((sector_t) 0xffff);
1604
1605 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1606 invalid_block = ~0;
1607
1608 BUG_ON(create == 0);
1609 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1610
1611 map.m_lblk = iblock;
1612 map.m_len = 1;
1613
1614 /*
1615 * first, we need to know whether the block is allocated already
1616 * preallocated blocks are unmapped but should treated
1617 * the same as allocated blocks.
1618 */
1619 ret = ext4_map_blocks(NULL, inode, &map, 0);
1620 if (ret < 0)
1621 return ret;
1622 if (ret == 0) {
1623 if (buffer_delay(bh))
1624 return 0; /* Not sure this could or should happen */
1625 /*
1626 * XXX: __block_write_begin() unmaps passed block, is it OK?
1627 */
1628 ret = ext4_da_reserve_space(inode, iblock);
1629 if (ret)
1630 /* not enough space to reserve */
1631 return ret;
1632
1633 map_bh(bh, inode->i_sb, invalid_block);
1634 set_buffer_new(bh);
1635 set_buffer_delay(bh);
1636 return 0;
1637 }
1638
1639 map_bh(bh, inode->i_sb, map.m_pblk);
1640 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1641
1642 if (buffer_unwritten(bh)) {
1643 /* A delayed write to unwritten bh should be marked
1644 * new and mapped. Mapped ensures that we don't do
1645 * get_block multiple times when we write to the same
1646 * offset and new ensures that we do proper zero out
1647 * for partial write.
1648 */
1649 set_buffer_new(bh);
1650 set_buffer_mapped(bh);
1651 }
1652 return 0;
1653}
1654
1655/*
1656 * This function is used as a standard get_block_t calback function
1657 * when there is no desire to allocate any blocks. It is used as a
1658 * callback function for block_write_begin() and block_write_full_page().
1659 * These functions should only try to map a single block at a time.
1660 *
1661 * Since this function doesn't do block allocations even if the caller
1662 * requests it by passing in create=1, it is critically important that
1663 * any caller checks to make sure that any buffer heads are returned
1664 * by this function are either all already mapped or marked for
1665 * delayed allocation before calling block_write_full_page(). Otherwise,
1666 * b_blocknr could be left unitialized, and the page write functions will
1667 * be taken by surprise.
1668 */
1669static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1670 struct buffer_head *bh_result, int create)
1671{
1672 BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1673 return _ext4_get_block(inode, iblock, bh_result, 0);
1674}
1675
1676static int bget_one(handle_t *handle, struct buffer_head *bh)
1677{
1678 get_bh(bh);
1679 return 0;
1680}
1681
1682static int bput_one(handle_t *handle, struct buffer_head *bh)
1683{
1684 put_bh(bh);
1685 return 0;
1686}
1687
1688static int __ext4_journalled_writepage(struct page *page,
1689 unsigned int len)
1690{
1691 struct address_space *mapping = page->mapping;
1692 struct inode *inode = mapping->host;
1693 struct buffer_head *page_bufs;
1694 handle_t *handle = NULL;
1695 int ret = 0;
1696 int err;
1697
1698 ClearPageChecked(page);
1699 page_bufs = page_buffers(page);
1700 BUG_ON(!page_bufs);
1701 walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1702 /* As soon as we unlock the page, it can go away, but we have
1703 * references to buffers so we are safe */
1704 unlock_page(page);
1705
1706 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1707 if (IS_ERR(handle)) {
1708 ret = PTR_ERR(handle);
1709 goto out;
1710 }
1711
1712 BUG_ON(!ext4_handle_valid(handle));
1713
1714 ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1715 do_journal_get_write_access);
1716
1717 err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1718 write_end_fn);
1719 if (ret == 0)
1720 ret = err;
1721 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1722 err = ext4_journal_stop(handle);
1723 if (!ret)
1724 ret = err;
1725
1726 walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1727 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1728out:
1729 return ret;
1730}
1731
1732static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1733static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1734
1735/*
1736 * Note that we don't need to start a transaction unless we're journaling data
1737 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1738 * need to file the inode to the transaction's list in ordered mode because if
1739 * we are writing back data added by write(), the inode is already there and if
1740 * we are writing back data modified via mmap(), no one guarantees in which
1741 * transaction the data will hit the disk. In case we are journaling data, we
1742 * cannot start transaction directly because transaction start ranks above page
1743 * lock so we have to do some magic.
1744 *
1745 * This function can get called via...
1746 * - ext4_da_writepages after taking page lock (have journal handle)
1747 * - journal_submit_inode_data_buffers (no journal handle)
1748 * - shrink_page_list via pdflush (no journal handle)
1749 * - grab_page_cache when doing write_begin (have journal handle)
1750 *
1751 * We don't do any block allocation in this function. If we have page with
1752 * multiple blocks we need to write those buffer_heads that are mapped. This
1753 * is important for mmaped based write. So if we do with blocksize 1K
1754 * truncate(f, 1024);
1755 * a = mmap(f, 0, 4096);
1756 * a[0] = 'a';
1757 * truncate(f, 4096);
1758 * we have in the page first buffer_head mapped via page_mkwrite call back
1759 * but other bufer_heads would be unmapped but dirty(dirty done via the
1760 * do_wp_page). So writepage should write the first block. If we modify
1761 * the mmap area beyond 1024 we will again get a page_fault and the
1762 * page_mkwrite callback will do the block allocation and mark the
1763 * buffer_heads mapped.
1764 *
1765 * We redirty the page if we have any buffer_heads that is either delay or
1766 * unwritten in the page.
1767 *
1768 * We can get recursively called as show below.
1769 *
1770 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1771 * ext4_writepage()
1772 *
1773 * But since we don't do any block allocation we should not deadlock.
1774 * Page also have the dirty flag cleared so we don't get recurive page_lock.
1775 */
1776static int ext4_writepage(struct page *page,
1777 struct writeback_control *wbc)
1778{
1779 int ret = 0, commit_write = 0;
1780 loff_t size;
1781 unsigned int len;
1782 struct buffer_head *page_bufs = NULL;
1783 struct inode *inode = page->mapping->host;
1784
1785 trace_ext4_writepage(page);
1786 size = i_size_read(inode);
1787 if (page->index == size >> PAGE_CACHE_SHIFT)
1788 len = size & ~PAGE_CACHE_MASK;
1789 else
1790 len = PAGE_CACHE_SIZE;
1791
1792 /*
1793 * If the page does not have buffers (for whatever reason),
1794 * try to create them using __block_write_begin. If this
1795 * fails, redirty the page and move on.
1796 */
1797 if (!page_has_buffers(page)) {
1798 if (__block_write_begin(page, 0, len,
1799 noalloc_get_block_write)) {
1800 redirty_page:
1801 redirty_page_for_writepage(wbc, page);
1802 unlock_page(page);
1803 return 0;
1804 }
1805 commit_write = 1;
1806 }
1807 page_bufs = page_buffers(page);
1808 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1809 ext4_bh_delay_or_unwritten)) {
1810 /*
1811 * We don't want to do block allocation, so redirty
1812 * the page and return. We may reach here when we do
1813 * a journal commit via journal_submit_inode_data_buffers.
1814 * We can also reach here via shrink_page_list
1815 */
1816 goto redirty_page;
1817 }
1818 if (commit_write)
1819 /* now mark the buffer_heads as dirty and uptodate */
1820 block_commit_write(page, 0, len);
1821
1822 if (PageChecked(page) && ext4_should_journal_data(inode))
1823 /*
1824 * It's mmapped pagecache. Add buffers and journal it. There
1825 * doesn't seem much point in redirtying the page here.
1826 */
1827 return __ext4_journalled_writepage(page, len);
1828
1829 if (buffer_uninit(page_bufs)) {
1830 ext4_set_bh_endio(page_bufs, inode);
1831 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1832 wbc, ext4_end_io_buffer_write);
1833 } else
1834 ret = block_write_full_page(page, noalloc_get_block_write,
1835 wbc);
1836
1837 return ret;
1838}
1839
1840/*
1841 * This is called via ext4_da_writepages() to
1842 * calculate the total number of credits to reserve to fit
1843 * a single extent allocation into a single transaction,
1844 * ext4_da_writpeages() will loop calling this before
1845 * the block allocation.
1846 */
1847
1848static int ext4_da_writepages_trans_blocks(struct inode *inode)
1849{
1850 int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1851
1852 /*
1853 * With non-extent format the journal credit needed to
1854 * insert nrblocks contiguous block is dependent on
1855 * number of contiguous block. So we will limit
1856 * number of contiguous block to a sane value
1857 */
1858 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1859 (max_blocks > EXT4_MAX_TRANS_DATA))
1860 max_blocks = EXT4_MAX_TRANS_DATA;
1861
1862 return ext4_chunk_trans_blocks(inode, max_blocks);
1863}
1864
1865/*
1866 * write_cache_pages_da - walk the list of dirty pages of the given
1867 * address space and accumulate pages that need writing, and call
1868 * mpage_da_map_and_submit to map a single contiguous memory region
1869 * and then write them.
1870 */
1871static int write_cache_pages_da(struct address_space *mapping,
1872 struct writeback_control *wbc,
1873 struct mpage_da_data *mpd,
1874 pgoff_t *done_index)
1875{
1876 struct buffer_head *bh, *head;
1877 struct inode *inode = mapping->host;
1878 struct pagevec pvec;
1879 unsigned int nr_pages;
1880 sector_t logical;
1881 pgoff_t index, end;
1882 long nr_to_write = wbc->nr_to_write;
1883 int i, tag, ret = 0;
1884
1885 memset(mpd, 0, sizeof(struct mpage_da_data));
1886 mpd->wbc = wbc;
1887 mpd->inode = inode;
1888 pagevec_init(&pvec, 0);
1889 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1890 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1891
1892 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1893 tag = PAGECACHE_TAG_TOWRITE;
1894 else
1895 tag = PAGECACHE_TAG_DIRTY;
1896
1897 *done_index = index;
1898 while (index <= end) {
1899 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1900 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1901 if (nr_pages == 0)
1902 return 0;
1903
1904 for (i = 0; i < nr_pages; i++) {
1905 struct page *page = pvec.pages[i];
1906
1907 /*
1908 * At this point, the page may be truncated or
1909 * invalidated (changing page->mapping to NULL), or
1910 * even swizzled back from swapper_space to tmpfs file
1911 * mapping. However, page->index will not change
1912 * because we have a reference on the page.
1913 */
1914 if (page->index > end)
1915 goto out;
1916
1917 *done_index = page->index + 1;
1918
1919 /*
1920 * If we can't merge this page, and we have
1921 * accumulated an contiguous region, write it
1922 */
1923 if ((mpd->next_page != page->index) &&
1924 (mpd->next_page != mpd->first_page)) {
1925 mpage_da_map_and_submit(mpd);
1926 goto ret_extent_tail;
1927 }
1928
1929 lock_page(page);
1930
1931 /*
1932 * If the page is no longer dirty, or its
1933 * mapping no longer corresponds to inode we
1934 * are writing (which means it has been
1935 * truncated or invalidated), or the page is
1936 * already under writeback and we are not
1937 * doing a data integrity writeback, skip the page
1938 */
1939 if (!PageDirty(page) ||
1940 (PageWriteback(page) &&
1941 (wbc->sync_mode == WB_SYNC_NONE)) ||
1942 unlikely(page->mapping != mapping)) {
1943 unlock_page(page);
1944 continue;
1945 }
1946
1947 wait_on_page_writeback(page);
1948 BUG_ON(PageWriteback(page));
1949
1950 if (mpd->next_page != page->index)
1951 mpd->first_page = page->index;
1952 mpd->next_page = page->index + 1;
1953 logical = (sector_t) page->index <<
1954 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1955
1956 if (!page_has_buffers(page)) {
1957 mpage_add_bh_to_extent(mpd, logical,
1958 PAGE_CACHE_SIZE,
1959 (1 << BH_Dirty) | (1 << BH_Uptodate));
1960 if (mpd->io_done)
1961 goto ret_extent_tail;
1962 } else {
1963 /*
1964 * Page with regular buffer heads,
1965 * just add all dirty ones
1966 */
1967 head = page_buffers(page);
1968 bh = head;
1969 do {
1970 BUG_ON(buffer_locked(bh));
1971 /*
1972 * We need to try to allocate
1973 * unmapped blocks in the same page.
1974 * Otherwise we won't make progress
1975 * with the page in ext4_writepage
1976 */
1977 if (ext4_bh_delay_or_unwritten(NULL, bh)) {
1978 mpage_add_bh_to_extent(mpd, logical,
1979 bh->b_size,
1980 bh->b_state);
1981 if (mpd->io_done)
1982 goto ret_extent_tail;
1983 } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
1984 /*
1985 * mapped dirty buffer. We need
1986 * to update the b_state
1987 * because we look at b_state
1988 * in mpage_da_map_blocks. We
1989 * don't update b_size because
1990 * if we find an unmapped
1991 * buffer_head later we need to
1992 * use the b_state flag of that
1993 * buffer_head.
1994 */
1995 if (mpd->b_size == 0)
1996 mpd->b_state = bh->b_state & BH_FLAGS;
1997 }
1998 logical++;
1999 } while ((bh = bh->b_this_page) != head);
2000 }
2001
2002 if (nr_to_write > 0) {
2003 nr_to_write--;
2004 if (nr_to_write == 0 &&
2005 wbc->sync_mode == WB_SYNC_NONE)
2006 /*
2007 * We stop writing back only if we are
2008 * not doing integrity sync. In case of
2009 * integrity sync we have to keep going
2010 * because someone may be concurrently
2011 * dirtying pages, and we might have
2012 * synced a lot of newly appeared dirty
2013 * pages, but have not synced all of the
2014 * old dirty pages.
2015 */
2016 goto out;
2017 }
2018 }
2019 pagevec_release(&pvec);
2020 cond_resched();
2021 }
2022 return 0;
2023ret_extent_tail:
2024 ret = MPAGE_DA_EXTENT_TAIL;
2025out:
2026 pagevec_release(&pvec);
2027 cond_resched();
2028 return ret;
2029}
2030
2031
2032static int ext4_da_writepages(struct address_space *mapping,
2033 struct writeback_control *wbc)
2034{
2035 pgoff_t index;
2036 int range_whole = 0;
2037 handle_t *handle = NULL;
2038 struct mpage_da_data mpd;
2039 struct inode *inode = mapping->host;
2040 int pages_written = 0;
2041 unsigned int max_pages;
2042 int range_cyclic, cycled = 1, io_done = 0;
2043 int needed_blocks, ret = 0;
2044 long desired_nr_to_write, nr_to_writebump = 0;
2045 loff_t range_start = wbc->range_start;
2046 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2047 pgoff_t done_index = 0;
2048 pgoff_t end;
2049
2050 trace_ext4_da_writepages(inode, wbc);
2051
2052 /*
2053 * No pages to write? This is mainly a kludge to avoid starting
2054 * a transaction for special inodes like journal inode on last iput()
2055 * because that could violate lock ordering on umount
2056 */
2057 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2058 return 0;
2059
2060 /*
2061 * If the filesystem has aborted, it is read-only, so return
2062 * right away instead of dumping stack traces later on that
2063 * will obscure the real source of the problem. We test
2064 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2065 * the latter could be true if the filesystem is mounted
2066 * read-only, and in that case, ext4_da_writepages should
2067 * *never* be called, so if that ever happens, we would want
2068 * the stack trace.
2069 */
2070 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2071 return -EROFS;
2072
2073 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2074 range_whole = 1;
2075
2076 range_cyclic = wbc->range_cyclic;
2077 if (wbc->range_cyclic) {
2078 index = mapping->writeback_index;
2079 if (index)
2080 cycled = 0;
2081 wbc->range_start = index << PAGE_CACHE_SHIFT;
2082 wbc->range_end = LLONG_MAX;
2083 wbc->range_cyclic = 0;
2084 end = -1;
2085 } else {
2086 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2087 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2088 }
2089
2090 /*
2091 * This works around two forms of stupidity. The first is in
2092 * the writeback code, which caps the maximum number of pages
2093 * written to be 1024 pages. This is wrong on multiple
2094 * levels; different architectues have a different page size,
2095 * which changes the maximum amount of data which gets
2096 * written. Secondly, 4 megabytes is way too small. XFS
2097 * forces this value to be 16 megabytes by multiplying
2098 * nr_to_write parameter by four, and then relies on its
2099 * allocator to allocate larger extents to make them
2100 * contiguous. Unfortunately this brings us to the second
2101 * stupidity, which is that ext4's mballoc code only allocates
2102 * at most 2048 blocks. So we force contiguous writes up to
2103 * the number of dirty blocks in the inode, or
2104 * sbi->max_writeback_mb_bump whichever is smaller.
2105 */
2106 max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2107 if (!range_cyclic && range_whole) {
2108 if (wbc->nr_to_write == LONG_MAX)
2109 desired_nr_to_write = wbc->nr_to_write;
2110 else
2111 desired_nr_to_write = wbc->nr_to_write * 8;
2112 } else
2113 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2114 max_pages);
2115 if (desired_nr_to_write > max_pages)
2116 desired_nr_to_write = max_pages;
2117
2118 if (wbc->nr_to_write < desired_nr_to_write) {
2119 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2120 wbc->nr_to_write = desired_nr_to_write;
2121 }
2122
2123retry:
2124 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2125 tag_pages_for_writeback(mapping, index, end);
2126
2127 while (!ret && wbc->nr_to_write > 0) {
2128
2129 /*
2130 * we insert one extent at a time. So we need
2131 * credit needed for single extent allocation.
2132 * journalled mode is currently not supported
2133 * by delalloc
2134 */
2135 BUG_ON(ext4_should_journal_data(inode));
2136 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2137
2138 /* start a new transaction*/
2139 handle = ext4_journal_start(inode, needed_blocks);
2140 if (IS_ERR(handle)) {
2141 ret = PTR_ERR(handle);
2142 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2143 "%ld pages, ino %lu; err %d", __func__,
2144 wbc->nr_to_write, inode->i_ino, ret);
2145 goto out_writepages;
2146 }
2147
2148 /*
2149 * Now call write_cache_pages_da() to find the next
2150 * contiguous region of logical blocks that need
2151 * blocks to be allocated by ext4 and submit them.
2152 */
2153 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2154 /*
2155 * If we have a contiguous extent of pages and we
2156 * haven't done the I/O yet, map the blocks and submit
2157 * them for I/O.
2158 */
2159 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2160 mpage_da_map_and_submit(&mpd);
2161 ret = MPAGE_DA_EXTENT_TAIL;
2162 }
2163 trace_ext4_da_write_pages(inode, &mpd);
2164 wbc->nr_to_write -= mpd.pages_written;
2165
2166 ext4_journal_stop(handle);
2167
2168 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2169 /* commit the transaction which would
2170 * free blocks released in the transaction
2171 * and try again
2172 */
2173 jbd2_journal_force_commit_nested(sbi->s_journal);
2174 ret = 0;
2175 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2176 /*
2177 * got one extent now try with
2178 * rest of the pages
2179 */
2180 pages_written += mpd.pages_written;
2181 ret = 0;
2182 io_done = 1;
2183 } else if (wbc->nr_to_write)
2184 /*
2185 * There is no more writeout needed
2186 * or we requested for a noblocking writeout
2187 * and we found the device congested
2188 */
2189 break;
2190 }
2191 if (!io_done && !cycled) {
2192 cycled = 1;
2193 index = 0;
2194 wbc->range_start = index << PAGE_CACHE_SHIFT;
2195 wbc->range_end = mapping->writeback_index - 1;
2196 goto retry;
2197 }
2198
2199 /* Update index */
2200 wbc->range_cyclic = range_cyclic;
2201 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2202 /*
2203 * set the writeback_index so that range_cyclic
2204 * mode will write it back later
2205 */
2206 mapping->writeback_index = done_index;
2207
2208out_writepages:
2209 wbc->nr_to_write -= nr_to_writebump;
2210 wbc->range_start = range_start;
2211 trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2212 return ret;
2213}
2214
2215#define FALL_BACK_TO_NONDELALLOC 1
2216static int ext4_nonda_switch(struct super_block *sb)
2217{
2218 s64 free_blocks, dirty_blocks;
2219 struct ext4_sb_info *sbi = EXT4_SB(sb);
2220
2221 /*
2222 * switch to non delalloc mode if we are running low
2223 * on free block. The free block accounting via percpu
2224 * counters can get slightly wrong with percpu_counter_batch getting
2225 * accumulated on each CPU without updating global counters
2226 * Delalloc need an accurate free block accounting. So switch
2227 * to non delalloc when we are near to error range.
2228 */
2229 free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2230 dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2231 if (2 * free_blocks < 3 * dirty_blocks ||
2232 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2233 /*
2234 * free block count is less than 150% of dirty blocks
2235 * or free blocks is less than watermark
2236 */
2237 return 1;
2238 }
2239 /*
2240 * Even if we don't switch but are nearing capacity,
2241 * start pushing delalloc when 1/2 of free blocks are dirty.
2242 */
2243 if (free_blocks < 2 * dirty_blocks)
2244 writeback_inodes_sb_if_idle(sb);
2245
2246 return 0;
2247}
2248
2249static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2250 loff_t pos, unsigned len, unsigned flags,
2251 struct page **pagep, void **fsdata)
2252{
2253 int ret, retries = 0;
2254 struct page *page;
2255 pgoff_t index;
2256 struct inode *inode = mapping->host;
2257 handle_t *handle;
2258
2259 index = pos >> PAGE_CACHE_SHIFT;
2260
2261 if (ext4_nonda_switch(inode->i_sb)) {
2262 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2263 return ext4_write_begin(file, mapping, pos,
2264 len, flags, pagep, fsdata);
2265 }
2266 *fsdata = (void *)0;
2267 trace_ext4_da_write_begin(inode, pos, len, flags);
2268retry:
2269 /*
2270 * With delayed allocation, we don't log the i_disksize update
2271 * if there is delayed block allocation. But we still need
2272 * to journalling the i_disksize update if writes to the end
2273 * of file which has an already mapped buffer.
2274 */
2275 handle = ext4_journal_start(inode, 1);
2276 if (IS_ERR(handle)) {
2277 ret = PTR_ERR(handle);
2278 goto out;
2279 }
2280 /* We cannot recurse into the filesystem as the transaction is already
2281 * started */
2282 flags |= AOP_FLAG_NOFS;
2283
2284 page = grab_cache_page_write_begin(mapping, index, flags);
2285 if (!page) {
2286 ext4_journal_stop(handle);
2287 ret = -ENOMEM;
2288 goto out;
2289 }
2290 *pagep = page;
2291
2292 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2293 if (ret < 0) {
2294 unlock_page(page);
2295 ext4_journal_stop(handle);
2296 page_cache_release(page);
2297 /*
2298 * block_write_begin may have instantiated a few blocks
2299 * outside i_size. Trim these off again. Don't need
2300 * i_size_read because we hold i_mutex.
2301 */
2302 if (pos + len > inode->i_size)
2303 ext4_truncate_failed_write(inode);
2304 }
2305
2306 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2307 goto retry;
2308out:
2309 return ret;
2310}
2311
2312/*
2313 * Check if we should update i_disksize
2314 * when write to the end of file but not require block allocation
2315 */
2316static int ext4_da_should_update_i_disksize(struct page *page,
2317 unsigned long offset)
2318{
2319 struct buffer_head *bh;
2320 struct inode *inode = page->mapping->host;
2321 unsigned int idx;
2322 int i;
2323
2324 bh = page_buffers(page);
2325 idx = offset >> inode->i_blkbits;
2326
2327 for (i = 0; i < idx; i++)
2328 bh = bh->b_this_page;
2329
2330 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2331 return 0;
2332 return 1;
2333}
2334
2335static int ext4_da_write_end(struct file *file,
2336 struct address_space *mapping,
2337 loff_t pos, unsigned len, unsigned copied,
2338 struct page *page, void *fsdata)
2339{
2340 struct inode *inode = mapping->host;
2341 int ret = 0, ret2;
2342 handle_t *handle = ext4_journal_current_handle();
2343 loff_t new_i_size;
2344 unsigned long start, end;
2345 int write_mode = (int)(unsigned long)fsdata;
2346
2347 if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2348 if (ext4_should_order_data(inode)) {
2349 return ext4_ordered_write_end(file, mapping, pos,
2350 len, copied, page, fsdata);
2351 } else if (ext4_should_writeback_data(inode)) {
2352 return ext4_writeback_write_end(file, mapping, pos,
2353 len, copied, page, fsdata);
2354 } else {
2355 BUG();
2356 }
2357 }
2358
2359 trace_ext4_da_write_end(inode, pos, len, copied);
2360 start = pos & (PAGE_CACHE_SIZE - 1);
2361 end = start + copied - 1;
2362
2363 /*
2364 * generic_write_end() will run mark_inode_dirty() if i_size
2365 * changes. So let's piggyback the i_disksize mark_inode_dirty
2366 * into that.
2367 */
2368
2369 new_i_size = pos + copied;
2370 if (new_i_size > EXT4_I(inode)->i_disksize) {
2371 if (ext4_da_should_update_i_disksize(page, end)) {
2372 down_write(&EXT4_I(inode)->i_data_sem);
2373 if (new_i_size > EXT4_I(inode)->i_disksize) {
2374 /*
2375 * Updating i_disksize when extending file
2376 * without needing block allocation
2377 */
2378 if (ext4_should_order_data(inode))
2379 ret = ext4_jbd2_file_inode(handle,
2380 inode);
2381
2382 EXT4_I(inode)->i_disksize = new_i_size;
2383 }
2384 up_write(&EXT4_I(inode)->i_data_sem);
2385 /* We need to mark inode dirty even if
2386 * new_i_size is less that inode->i_size
2387 * bu greater than i_disksize.(hint delalloc)
2388 */
2389 ext4_mark_inode_dirty(handle, inode);
2390 }
2391 }
2392 ret2 = generic_write_end(file, mapping, pos, len, copied,
2393 page, fsdata);
2394 copied = ret2;
2395 if (ret2 < 0)
2396 ret = ret2;
2397 ret2 = ext4_journal_stop(handle);
2398 if (!ret)
2399 ret = ret2;
2400
2401 return ret ? ret : copied;
2402}
2403
2404static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2405{
2406 /*
2407 * Drop reserved blocks
2408 */
2409 BUG_ON(!PageLocked(page));
2410 if (!page_has_buffers(page))
2411 goto out;
2412
2413 ext4_da_page_release_reservation(page, offset);
2414
2415out:
2416 ext4_invalidatepage(page, offset);
2417
2418 return;
2419}
2420
2421/*
2422 * Force all delayed allocation blocks to be allocated for a given inode.
2423 */
2424int ext4_alloc_da_blocks(struct inode *inode)
2425{
2426 trace_ext4_alloc_da_blocks(inode);
2427
2428 if (!EXT4_I(inode)->i_reserved_data_blocks &&
2429 !EXT4_I(inode)->i_reserved_meta_blocks)
2430 return 0;
2431
2432 /*
2433 * We do something simple for now. The filemap_flush() will
2434 * also start triggering a write of the data blocks, which is
2435 * not strictly speaking necessary (and for users of
2436 * laptop_mode, not even desirable). However, to do otherwise
2437 * would require replicating code paths in:
2438 *
2439 * ext4_da_writepages() ->
2440 * write_cache_pages() ---> (via passed in callback function)
2441 * __mpage_da_writepage() -->
2442 * mpage_add_bh_to_extent()
2443 * mpage_da_map_blocks()
2444 *
2445 * The problem is that write_cache_pages(), located in
2446 * mm/page-writeback.c, marks pages clean in preparation for
2447 * doing I/O, which is not desirable if we're not planning on
2448 * doing I/O at all.
2449 *
2450 * We could call write_cache_pages(), and then redirty all of
2451 * the pages by calling redirty_page_for_writepage() but that
2452 * would be ugly in the extreme. So instead we would need to
2453 * replicate parts of the code in the above functions,
2454 * simplifying them because we wouldn't actually intend to
2455 * write out the pages, but rather only collect contiguous
2456 * logical block extents, call the multi-block allocator, and
2457 * then update the buffer heads with the block allocations.
2458 *
2459 * For now, though, we'll cheat by calling filemap_flush(),
2460 * which will map the blocks, and start the I/O, but not
2461 * actually wait for the I/O to complete.
2462 */
2463 return filemap_flush(inode->i_mapping);
2464}
2465
2466/*
2467 * bmap() is special. It gets used by applications such as lilo and by
2468 * the swapper to find the on-disk block of a specific piece of data.
2469 *
2470 * Naturally, this is dangerous if the block concerned is still in the
2471 * journal. If somebody makes a swapfile on an ext4 data-journaling
2472 * filesystem and enables swap, then they may get a nasty shock when the
2473 * data getting swapped to that swapfile suddenly gets overwritten by
2474 * the original zero's written out previously to the journal and
2475 * awaiting writeback in the kernel's buffer cache.
2476 *
2477 * So, if we see any bmap calls here on a modified, data-journaled file,
2478 * take extra steps to flush any blocks which might be in the cache.
2479 */
2480static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2481{
2482 struct inode *inode = mapping->host;
2483 journal_t *journal;
2484 int err;
2485
2486 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2487 test_opt(inode->i_sb, DELALLOC)) {
2488 /*
2489 * With delalloc we want to sync the file
2490 * so that we can make sure we allocate
2491 * blocks for file
2492 */
2493 filemap_write_and_wait(mapping);
2494 }
2495
2496 if (EXT4_JOURNAL(inode) &&
2497 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2498 /*
2499 * This is a REALLY heavyweight approach, but the use of
2500 * bmap on dirty files is expected to be extremely rare:
2501 * only if we run lilo or swapon on a freshly made file
2502 * do we expect this to happen.
2503 *
2504 * (bmap requires CAP_SYS_RAWIO so this does not
2505 * represent an unprivileged user DOS attack --- we'd be
2506 * in trouble if mortal users could trigger this path at
2507 * will.)
2508 *
2509 * NB. EXT4_STATE_JDATA is not set on files other than
2510 * regular files. If somebody wants to bmap a directory
2511 * or symlink and gets confused because the buffer
2512 * hasn't yet been flushed to disk, they deserve
2513 * everything they get.
2514 */
2515
2516 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2517 journal = EXT4_JOURNAL(inode);
2518 jbd2_journal_lock_updates(journal);
2519 err = jbd2_journal_flush(journal);
2520 jbd2_journal_unlock_updates(journal);
2521
2522 if (err)
2523 return 0;
2524 }
2525
2526 return generic_block_bmap(mapping, block, ext4_get_block);
2527}
2528
2529static int ext4_readpage(struct file *file, struct page *page)
2530{
2531 trace_ext4_readpage(page);
2532 return mpage_readpage(page, ext4_get_block);
2533}
2534
2535static int
2536ext4_readpages(struct file *file, struct address_space *mapping,
2537 struct list_head *pages, unsigned nr_pages)
2538{
2539 return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2540}
2541
2542static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2543{
2544 struct buffer_head *head, *bh;
2545 unsigned int curr_off = 0;
2546
2547 if (!page_has_buffers(page))
2548 return;
2549 head = bh = page_buffers(page);
2550 do {
2551 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2552 && bh->b_private) {
2553 ext4_free_io_end(bh->b_private);
2554 bh->b_private = NULL;
2555 bh->b_end_io = NULL;
2556 }
2557 curr_off = curr_off + bh->b_size;
2558 bh = bh->b_this_page;
2559 } while (bh != head);
2560}
2561
2562static void ext4_invalidatepage(struct page *page, unsigned long offset)
2563{
2564 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2565
2566 trace_ext4_invalidatepage(page, offset);
2567
2568 /*
2569 * free any io_end structure allocated for buffers to be discarded
2570 */
2571 if (ext4_should_dioread_nolock(page->mapping->host))
2572 ext4_invalidatepage_free_endio(page, offset);
2573 /*
2574 * If it's a full truncate we just forget about the pending dirtying
2575 */
2576 if (offset == 0)
2577 ClearPageChecked(page);
2578
2579 if (journal)
2580 jbd2_journal_invalidatepage(journal, page, offset);
2581 else
2582 block_invalidatepage(page, offset);
2583}
2584
2585static int ext4_releasepage(struct page *page, gfp_t wait)
2586{
2587 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2588
2589 trace_ext4_releasepage(page);
2590
2591 WARN_ON(PageChecked(page));
2592 if (!page_has_buffers(page))
2593 return 0;
2594 if (journal)
2595 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2596 else
2597 return try_to_free_buffers(page);
2598}
2599
2600/*
2601 * ext4_get_block used when preparing for a DIO write or buffer write.
2602 * We allocate an uinitialized extent if blocks haven't been allocated.
2603 * The extent will be converted to initialized after the IO is complete.
2604 */
2605static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2606 struct buffer_head *bh_result, int create)
2607{
2608 ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2609 inode->i_ino, create);
2610 return _ext4_get_block(inode, iblock, bh_result,
2611 EXT4_GET_BLOCKS_IO_CREATE_EXT);
2612}
2613
2614static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2615 ssize_t size, void *private, int ret,
2616 bool is_async)
2617{
2618 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2619 ext4_io_end_t *io_end = iocb->private;
2620 struct workqueue_struct *wq;
2621 unsigned long flags;
2622 struct ext4_inode_info *ei;
2623
2624 /* if not async direct IO or dio with 0 bytes write, just return */
2625 if (!io_end || !size)
2626 goto out;
2627
2628 ext_debug("ext4_end_io_dio(): io_end 0x%p"
2629 "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2630 iocb->private, io_end->inode->i_ino, iocb, offset,
2631 size);
2632
2633 /* if not aio dio with unwritten extents, just free io and return */
2634 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2635 ext4_free_io_end(io_end);
2636 iocb->private = NULL;
2637out:
2638 if (is_async)
2639 aio_complete(iocb, ret, 0);
2640 inode_dio_done(inode);
2641 return;
2642 }
2643
2644 io_end->offset = offset;
2645 io_end->size = size;
2646 if (is_async) {
2647 io_end->iocb = iocb;
2648 io_end->result = ret;
2649 }
2650 wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2651
2652 /* Add the io_end to per-inode completed aio dio list*/
2653 ei = EXT4_I(io_end->inode);
2654 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2655 list_add_tail(&io_end->list, &ei->i_completed_io_list);
2656 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2657
2658 /* queue the work to convert unwritten extents to written */
2659 queue_work(wq, &io_end->work);
2660 iocb->private = NULL;
2661
2662 /* XXX: probably should move into the real I/O completion handler */
2663 inode_dio_done(inode);
2664}
2665
2666static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2667{
2668 ext4_io_end_t *io_end = bh->b_private;
2669 struct workqueue_struct *wq;
2670 struct inode *inode;
2671 unsigned long flags;
2672
2673 if (!test_clear_buffer_uninit(bh) || !io_end)
2674 goto out;
2675
2676 if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2677 printk("sb umounted, discard end_io request for inode %lu\n",
2678 io_end->inode->i_ino);
2679 ext4_free_io_end(io_end);
2680 goto out;
2681 }
2682
2683 /*
2684 * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2685 * but being more careful is always safe for the future change.
2686 */
2687 inode = io_end->inode;
2688 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2689 io_end->flag |= EXT4_IO_END_UNWRITTEN;
2690 atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
2691 }
2692
2693 /* Add the io_end to per-inode completed io list*/
2694 spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2695 list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2696 spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2697
2698 wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2699 /* queue the work to convert unwritten extents to written */
2700 queue_work(wq, &io_end->work);
2701out:
2702 bh->b_private = NULL;
2703 bh->b_end_io = NULL;
2704 clear_buffer_uninit(bh);
2705 end_buffer_async_write(bh, uptodate);
2706}
2707
2708static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2709{
2710 ext4_io_end_t *io_end;
2711 struct page *page = bh->b_page;
2712 loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2713 size_t size = bh->b_size;
2714
2715retry:
2716 io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2717 if (!io_end) {
2718 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2719 schedule();
2720 goto retry;
2721 }
2722 io_end->offset = offset;
2723 io_end->size = size;
2724 /*
2725 * We need to hold a reference to the page to make sure it
2726 * doesn't get evicted before ext4_end_io_work() has a chance
2727 * to convert the extent from written to unwritten.
2728 */
2729 io_end->page = page;
2730 get_page(io_end->page);
2731
2732 bh->b_private = io_end;
2733 bh->b_end_io = ext4_end_io_buffer_write;
2734 return 0;
2735}
2736
2737/*
2738 * For ext4 extent files, ext4 will do direct-io write to holes,
2739 * preallocated extents, and those write extend the file, no need to
2740 * fall back to buffered IO.
2741 *
2742 * For holes, we fallocate those blocks, mark them as uninitialized
2743 * If those blocks were preallocated, we mark sure they are splited, but
2744 * still keep the range to write as uninitialized.
2745 *
2746 * The unwrritten extents will be converted to written when DIO is completed.
2747 * For async direct IO, since the IO may still pending when return, we
2748 * set up an end_io call back function, which will do the conversion
2749 * when async direct IO completed.
2750 *
2751 * If the O_DIRECT write will extend the file then add this inode to the
2752 * orphan list. So recovery will truncate it back to the original size
2753 * if the machine crashes during the write.
2754 *
2755 */
2756static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2757 const struct iovec *iov, loff_t offset,
2758 unsigned long nr_segs)
2759{
2760 struct file *file = iocb->ki_filp;
2761 struct inode *inode = file->f_mapping->host;
2762 ssize_t ret;
2763 size_t count = iov_length(iov, nr_segs);
2764
2765 loff_t final_size = offset + count;
2766 if (rw == WRITE && final_size <= inode->i_size) {
2767 /*
2768 * We could direct write to holes and fallocate.
2769 *
2770 * Allocated blocks to fill the hole are marked as uninitialized
2771 * to prevent parallel buffered read to expose the stale data
2772 * before DIO complete the data IO.
2773 *
2774 * As to previously fallocated extents, ext4 get_block
2775 * will just simply mark the buffer mapped but still
2776 * keep the extents uninitialized.
2777 *
2778 * for non AIO case, we will convert those unwritten extents
2779 * to written after return back from blockdev_direct_IO.
2780 *
2781 * for async DIO, the conversion needs to be defered when
2782 * the IO is completed. The ext4 end_io callback function
2783 * will be called to take care of the conversion work.
2784 * Here for async case, we allocate an io_end structure to
2785 * hook to the iocb.
2786 */
2787 iocb->private = NULL;
2788 EXT4_I(inode)->cur_aio_dio = NULL;
2789 if (!is_sync_kiocb(iocb)) {
2790 iocb->private = ext4_init_io_end(inode, GFP_NOFS);
2791 if (!iocb->private)
2792 return -ENOMEM;
2793 /*
2794 * we save the io structure for current async
2795 * direct IO, so that later ext4_map_blocks()
2796 * could flag the io structure whether there
2797 * is a unwritten extents needs to be converted
2798 * when IO is completed.
2799 */
2800 EXT4_I(inode)->cur_aio_dio = iocb->private;
2801 }
2802
2803 ret = __blockdev_direct_IO(rw, iocb, inode,
2804 inode->i_sb->s_bdev, iov,
2805 offset, nr_segs,
2806 ext4_get_block_write,
2807 ext4_end_io_dio,
2808 NULL,
2809 DIO_LOCKING | DIO_SKIP_HOLES);
2810 if (iocb->private)
2811 EXT4_I(inode)->cur_aio_dio = NULL;
2812 /*
2813 * The io_end structure takes a reference to the inode,
2814 * that structure needs to be destroyed and the
2815 * reference to the inode need to be dropped, when IO is
2816 * complete, even with 0 byte write, or failed.
2817 *
2818 * In the successful AIO DIO case, the io_end structure will be
2819 * desctroyed and the reference to the inode will be dropped
2820 * after the end_io call back function is called.
2821 *
2822 * In the case there is 0 byte write, or error case, since
2823 * VFS direct IO won't invoke the end_io call back function,
2824 * we need to free the end_io structure here.
2825 */
2826 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2827 ext4_free_io_end(iocb->private);
2828 iocb->private = NULL;
2829 } else if (ret > 0 && ext4_test_inode_state(inode,
2830 EXT4_STATE_DIO_UNWRITTEN)) {
2831 int err;
2832 /*
2833 * for non AIO case, since the IO is already
2834 * completed, we could do the conversion right here
2835 */
2836 err = ext4_convert_unwritten_extents(inode,
2837 offset, ret);
2838 if (err < 0)
2839 ret = err;
2840 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2841 }
2842 return ret;
2843 }
2844
2845 /* for write the the end of file case, we fall back to old way */
2846 return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2847}
2848
2849static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2850 const struct iovec *iov, loff_t offset,
2851 unsigned long nr_segs)
2852{
2853 struct file *file = iocb->ki_filp;
2854 struct inode *inode = file->f_mapping->host;
2855 ssize_t ret;
2856
2857 trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
2858 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
2859 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
2860 else
2861 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2862 trace_ext4_direct_IO_exit(inode, offset,
2863 iov_length(iov, nr_segs), rw, ret);
2864 return ret;
2865}
2866
2867/*
2868 * Pages can be marked dirty completely asynchronously from ext4's journalling
2869 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
2870 * much here because ->set_page_dirty is called under VFS locks. The page is
2871 * not necessarily locked.
2872 *
2873 * We cannot just dirty the page and leave attached buffers clean, because the
2874 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
2875 * or jbddirty because all the journalling code will explode.
2876 *
2877 * So what we do is to mark the page "pending dirty" and next time writepage
2878 * is called, propagate that into the buffers appropriately.
2879 */
2880static int ext4_journalled_set_page_dirty(struct page *page)
2881{
2882 SetPageChecked(page);
2883 return __set_page_dirty_nobuffers(page);
2884}
2885
2886static const struct address_space_operations ext4_ordered_aops = {
2887 .readpage = ext4_readpage,
2888 .readpages = ext4_readpages,
2889 .writepage = ext4_writepage,
2890 .write_begin = ext4_write_begin,
2891 .write_end = ext4_ordered_write_end,
2892 .bmap = ext4_bmap,
2893 .invalidatepage = ext4_invalidatepage,
2894 .releasepage = ext4_releasepage,
2895 .direct_IO = ext4_direct_IO,
2896 .migratepage = buffer_migrate_page,
2897 .is_partially_uptodate = block_is_partially_uptodate,
2898 .error_remove_page = generic_error_remove_page,
2899};
2900
2901static const struct address_space_operations ext4_writeback_aops = {
2902 .readpage = ext4_readpage,
2903 .readpages = ext4_readpages,
2904 .writepage = ext4_writepage,
2905 .write_begin = ext4_write_begin,
2906 .write_end = ext4_writeback_write_end,
2907 .bmap = ext4_bmap,
2908 .invalidatepage = ext4_invalidatepage,
2909 .releasepage = ext4_releasepage,
2910 .direct_IO = ext4_direct_IO,
2911 .migratepage = buffer_migrate_page,
2912 .is_partially_uptodate = block_is_partially_uptodate,
2913 .error_remove_page = generic_error_remove_page,
2914};
2915
2916static const struct address_space_operations ext4_journalled_aops = {
2917 .readpage = ext4_readpage,
2918 .readpages = ext4_readpages,
2919 .writepage = ext4_writepage,
2920 .write_begin = ext4_write_begin,
2921 .write_end = ext4_journalled_write_end,
2922 .set_page_dirty = ext4_journalled_set_page_dirty,
2923 .bmap = ext4_bmap,
2924 .invalidatepage = ext4_invalidatepage,
2925 .releasepage = ext4_releasepage,
2926 .is_partially_uptodate = block_is_partially_uptodate,
2927 .error_remove_page = generic_error_remove_page,
2928};
2929
2930static const struct address_space_operations ext4_da_aops = {
2931 .readpage = ext4_readpage,
2932 .readpages = ext4_readpages,
2933 .writepage = ext4_writepage,
2934 .writepages = ext4_da_writepages,
2935 .write_begin = ext4_da_write_begin,
2936 .write_end = ext4_da_write_end,
2937 .bmap = ext4_bmap,
2938 .invalidatepage = ext4_da_invalidatepage,
2939 .releasepage = ext4_releasepage,
2940 .direct_IO = ext4_direct_IO,
2941 .migratepage = buffer_migrate_page,
2942 .is_partially_uptodate = block_is_partially_uptodate,
2943 .error_remove_page = generic_error_remove_page,
2944};
2945
2946void ext4_set_aops(struct inode *inode)
2947{
2948 if (ext4_should_order_data(inode) &&
2949 test_opt(inode->i_sb, DELALLOC))
2950 inode->i_mapping->a_ops = &ext4_da_aops;
2951 else if (ext4_should_order_data(inode))
2952 inode->i_mapping->a_ops = &ext4_ordered_aops;
2953 else if (ext4_should_writeback_data(inode) &&
2954 test_opt(inode->i_sb, DELALLOC))
2955 inode->i_mapping->a_ops = &ext4_da_aops;
2956 else if (ext4_should_writeback_data(inode))
2957 inode->i_mapping->a_ops = &ext4_writeback_aops;
2958 else
2959 inode->i_mapping->a_ops = &ext4_journalled_aops;
2960}
2961
2962/*
2963 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
2964 * up to the end of the block which corresponds to `from'.
2965 * This required during truncate. We need to physically zero the tail end
2966 * of that block so it doesn't yield old data if the file is later grown.
2967 */
2968int ext4_block_truncate_page(handle_t *handle,
2969 struct address_space *mapping, loff_t from)
2970{
2971 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2972 unsigned length;
2973 unsigned blocksize;
2974 struct inode *inode = mapping->host;
2975
2976 blocksize = inode->i_sb->s_blocksize;
2977 length = blocksize - (offset & (blocksize - 1));
2978
2979 return ext4_block_zero_page_range(handle, mapping, from, length);
2980}
2981
2982/*
2983 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
2984 * starting from file offset 'from'. The range to be zero'd must
2985 * be contained with in one block. If the specified range exceeds
2986 * the end of the block it will be shortened to end of the block
2987 * that cooresponds to 'from'
2988 */
2989int ext4_block_zero_page_range(handle_t *handle,
2990 struct address_space *mapping, loff_t from, loff_t length)
2991{
2992 ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2993 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2994 unsigned blocksize, max, pos;
2995 ext4_lblk_t iblock;
2996 struct inode *inode = mapping->host;
2997 struct buffer_head *bh;
2998 struct page *page;
2999 int err = 0;
3000
3001 page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3002 mapping_gfp_mask(mapping) & ~__GFP_FS);
3003 if (!page)
3004 return -EINVAL;
3005
3006 blocksize = inode->i_sb->s_blocksize;
3007 max = blocksize - (offset & (blocksize - 1));
3008
3009 /*
3010 * correct length if it does not fall between
3011 * 'from' and the end of the block
3012 */
3013 if (length > max || length < 0)
3014 length = max;
3015
3016 iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3017
3018 if (!page_has_buffers(page))
3019 create_empty_buffers(page, blocksize, 0);
3020
3021 /* Find the buffer that contains "offset" */
3022 bh = page_buffers(page);
3023 pos = blocksize;
3024 while (offset >= pos) {
3025 bh = bh->b_this_page;
3026 iblock++;
3027 pos += blocksize;
3028 }
3029
3030 err = 0;
3031 if (buffer_freed(bh)) {
3032 BUFFER_TRACE(bh, "freed: skip");
3033 goto unlock;
3034 }
3035
3036 if (!buffer_mapped(bh)) {
3037 BUFFER_TRACE(bh, "unmapped");
3038 ext4_get_block(inode, iblock, bh, 0);
3039 /* unmapped? It's a hole - nothing to do */
3040 if (!buffer_mapped(bh)) {
3041 BUFFER_TRACE(bh, "still unmapped");
3042 goto unlock;
3043 }
3044 }
3045
3046 /* Ok, it's mapped. Make sure it's up-to-date */
3047 if (PageUptodate(page))
3048 set_buffer_uptodate(bh);
3049
3050 if (!buffer_uptodate(bh)) {
3051 err = -EIO;
3052 ll_rw_block(READ, 1, &bh);
3053 wait_on_buffer(bh);
3054 /* Uhhuh. Read error. Complain and punt. */
3055 if (!buffer_uptodate(bh))
3056 goto unlock;
3057 }
3058
3059 if (ext4_should_journal_data(inode)) {
3060 BUFFER_TRACE(bh, "get write access");
3061 err = ext4_journal_get_write_access(handle, bh);
3062 if (err)
3063 goto unlock;
3064 }
3065
3066 zero_user(page, offset, length);
3067
3068 BUFFER_TRACE(bh, "zeroed end of block");
3069
3070 err = 0;
3071 if (ext4_should_journal_data(inode)) {
3072 err = ext4_handle_dirty_metadata(handle, inode, bh);
3073 } else {
3074 if (ext4_should_order_data(inode) && EXT4_I(inode)->jinode)
3075 err = ext4_jbd2_file_inode(handle, inode);
3076 mark_buffer_dirty(bh);
3077 }
3078
3079unlock:
3080 unlock_page(page);
3081 page_cache_release(page);
3082 return err;
3083}
3084
3085int ext4_can_truncate(struct inode *inode)
3086{
3087 if (S_ISREG(inode->i_mode))
3088 return 1;
3089 if (S_ISDIR(inode->i_mode))
3090 return 1;
3091 if (S_ISLNK(inode->i_mode))
3092 return !ext4_inode_is_fast_symlink(inode);
3093 return 0;
3094}
3095
3096/*
3097 * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3098 * associated with the given offset and length
3099 *
3100 * @inode: File inode
3101 * @offset: The offset where the hole will begin
3102 * @len: The length of the hole
3103 *
3104 * Returns: 0 on sucess or negative on failure
3105 */
3106
3107int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3108{
3109 struct inode *inode = file->f_path.dentry->d_inode;
3110 if (!S_ISREG(inode->i_mode))
3111 return -ENOTSUPP;
3112
3113 if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3114 /* TODO: Add support for non extent hole punching */
3115 return -ENOTSUPP;
3116 }
3117
3118 return ext4_ext_punch_hole(file, offset, length);
3119}
3120
3121/*
3122 * ext4_truncate()
3123 *
3124 * We block out ext4_get_block() block instantiations across the entire
3125 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3126 * simultaneously on behalf of the same inode.
3127 *
3128 * As we work through the truncate and commmit bits of it to the journal there
3129 * is one core, guiding principle: the file's tree must always be consistent on
3130 * disk. We must be able to restart the truncate after a crash.
3131 *
3132 * The file's tree may be transiently inconsistent in memory (although it
3133 * probably isn't), but whenever we close off and commit a journal transaction,
3134 * the contents of (the filesystem + the journal) must be consistent and
3135 * restartable. It's pretty simple, really: bottom up, right to left (although
3136 * left-to-right works OK too).
3137 *
3138 * Note that at recovery time, journal replay occurs *before* the restart of
3139 * truncate against the orphan inode list.
3140 *
3141 * The committed inode has the new, desired i_size (which is the same as
3142 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
3143 * that this inode's truncate did not complete and it will again call
3144 * ext4_truncate() to have another go. So there will be instantiated blocks
3145 * to the right of the truncation point in a crashed ext4 filesystem. But
3146 * that's fine - as long as they are linked from the inode, the post-crash
3147 * ext4_truncate() run will find them and release them.
3148 */
3149void ext4_truncate(struct inode *inode)
3150{
3151 trace_ext4_truncate_enter(inode);
3152
3153 if (!ext4_can_truncate(inode))
3154 return;
3155
3156 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3157
3158 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3159 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3160
3161 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3162 ext4_ext_truncate(inode);
3163 else
3164 ext4_ind_truncate(inode);
3165
3166 trace_ext4_truncate_exit(inode);
3167}
3168
3169/*
3170 * ext4_get_inode_loc returns with an extra refcount against the inode's
3171 * underlying buffer_head on success. If 'in_mem' is true, we have all
3172 * data in memory that is needed to recreate the on-disk version of this
3173 * inode.
3174 */
3175static int __ext4_get_inode_loc(struct inode *inode,
3176 struct ext4_iloc *iloc, int in_mem)
3177{
3178 struct ext4_group_desc *gdp;
3179 struct buffer_head *bh;
3180 struct super_block *sb = inode->i_sb;
3181 ext4_fsblk_t block;
3182 int inodes_per_block, inode_offset;
3183
3184 iloc->bh = NULL;
3185 if (!ext4_valid_inum(sb, inode->i_ino))
3186 return -EIO;
3187
3188 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3189 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3190 if (!gdp)
3191 return -EIO;
3192
3193 /*
3194 * Figure out the offset within the block group inode table
3195 */
3196 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3197 inode_offset = ((inode->i_ino - 1) %
3198 EXT4_INODES_PER_GROUP(sb));
3199 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3200 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3201
3202 bh = sb_getblk(sb, block);
3203 if (!bh) {
3204 EXT4_ERROR_INODE_BLOCK(inode, block,
3205 "unable to read itable block");
3206 return -EIO;
3207 }
3208 if (!buffer_uptodate(bh)) {
3209 lock_buffer(bh);
3210
3211 /*
3212 * If the buffer has the write error flag, we have failed
3213 * to write out another inode in the same block. In this
3214 * case, we don't have to read the block because we may
3215 * read the old inode data successfully.
3216 */
3217 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3218 set_buffer_uptodate(bh);
3219
3220 if (buffer_uptodate(bh)) {
3221 /* someone brought it uptodate while we waited */
3222 unlock_buffer(bh);
3223 goto has_buffer;
3224 }
3225
3226 /*
3227 * If we have all information of the inode in memory and this
3228 * is the only valid inode in the block, we need not read the
3229 * block.
3230 */
3231 if (in_mem) {
3232 struct buffer_head *bitmap_bh;
3233 int i, start;
3234
3235 start = inode_offset & ~(inodes_per_block - 1);
3236
3237 /* Is the inode bitmap in cache? */
3238 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3239 if (!bitmap_bh)
3240 goto make_io;
3241
3242 /*
3243 * If the inode bitmap isn't in cache then the
3244 * optimisation may end up performing two reads instead
3245 * of one, so skip it.
3246 */
3247 if (!buffer_uptodate(bitmap_bh)) {
3248 brelse(bitmap_bh);
3249 goto make_io;
3250 }
3251 for (i = start; i < start + inodes_per_block; i++) {
3252 if (i == inode_offset)
3253 continue;
3254 if (ext4_test_bit(i, bitmap_bh->b_data))
3255 break;
3256 }
3257 brelse(bitmap_bh);
3258 if (i == start + inodes_per_block) {
3259 /* all other inodes are free, so skip I/O */
3260 memset(bh->b_data, 0, bh->b_size);
3261 set_buffer_uptodate(bh);
3262 unlock_buffer(bh);
3263 goto has_buffer;
3264 }
3265 }
3266
3267make_io:
3268 /*
3269 * If we need to do any I/O, try to pre-readahead extra
3270 * blocks from the inode table.
3271 */
3272 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3273 ext4_fsblk_t b, end, table;
3274 unsigned num;
3275
3276 table = ext4_inode_table(sb, gdp);
3277 /* s_inode_readahead_blks is always a power of 2 */
3278 b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3279 if (table > b)
3280 b = table;
3281 end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3282 num = EXT4_INODES_PER_GROUP(sb);
3283 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3284 EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3285 num -= ext4_itable_unused_count(sb, gdp);
3286 table += num / inodes_per_block;
3287 if (end > table)
3288 end = table;
3289 while (b <= end)
3290 sb_breadahead(sb, b++);
3291 }
3292
3293 /*
3294 * There are other valid inodes in the buffer, this inode
3295 * has in-inode xattrs, or we don't have this inode in memory.
3296 * Read the block from disk.
3297 */
3298 trace_ext4_load_inode(inode);
3299 get_bh(bh);
3300 bh->b_end_io = end_buffer_read_sync;
3301 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3302 wait_on_buffer(bh);
3303 if (!buffer_uptodate(bh)) {
3304 EXT4_ERROR_INODE_BLOCK(inode, block,
3305 "unable to read itable block");
3306 brelse(bh);
3307 return -EIO;
3308 }
3309 }
3310has_buffer:
3311 iloc->bh = bh;
3312 return 0;
3313}
3314
3315int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3316{
3317 /* We have all inode data except xattrs in memory here. */
3318 return __ext4_get_inode_loc(inode, iloc,
3319 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3320}
3321
3322void ext4_set_inode_flags(struct inode *inode)
3323{
3324 unsigned int flags = EXT4_I(inode)->i_flags;
3325
3326 inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3327 if (flags & EXT4_SYNC_FL)
3328 inode->i_flags |= S_SYNC;
3329 if (flags & EXT4_APPEND_FL)
3330 inode->i_flags |= S_APPEND;
3331 if (flags & EXT4_IMMUTABLE_FL)
3332 inode->i_flags |= S_IMMUTABLE;
3333 if (flags & EXT4_NOATIME_FL)
3334 inode->i_flags |= S_NOATIME;
3335 if (flags & EXT4_DIRSYNC_FL)
3336 inode->i_flags |= S_DIRSYNC;
3337}
3338
3339/* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3340void ext4_get_inode_flags(struct ext4_inode_info *ei)
3341{
3342 unsigned int vfs_fl;
3343 unsigned long old_fl, new_fl;
3344
3345 do {
3346 vfs_fl = ei->vfs_inode.i_flags;
3347 old_fl = ei->i_flags;
3348 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3349 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3350 EXT4_DIRSYNC_FL);
3351 if (vfs_fl & S_SYNC)
3352 new_fl |= EXT4_SYNC_FL;
3353 if (vfs_fl & S_APPEND)
3354 new_fl |= EXT4_APPEND_FL;
3355 if (vfs_fl & S_IMMUTABLE)
3356 new_fl |= EXT4_IMMUTABLE_FL;
3357 if (vfs_fl & S_NOATIME)
3358 new_fl |= EXT4_NOATIME_FL;
3359 if (vfs_fl & S_DIRSYNC)
3360 new_fl |= EXT4_DIRSYNC_FL;
3361 } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3362}
3363
3364static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3365 struct ext4_inode_info *ei)
3366{
3367 blkcnt_t i_blocks ;
3368 struct inode *inode = &(ei->vfs_inode);
3369 struct super_block *sb = inode->i_sb;
3370
3371 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3372 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3373 /* we are using combined 48 bit field */
3374 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3375 le32_to_cpu(raw_inode->i_blocks_lo);
3376 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
3377 /* i_blocks represent file system block size */
3378 return i_blocks << (inode->i_blkbits - 9);
3379 } else {
3380 return i_blocks;
3381 }
3382 } else {
3383 return le32_to_cpu(raw_inode->i_blocks_lo);
3384 }
3385}
3386
3387struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3388{
3389 struct ext4_iloc iloc;
3390 struct ext4_inode *raw_inode;
3391 struct ext4_inode_info *ei;
3392 struct inode *inode;
3393 journal_t *journal = EXT4_SB(sb)->s_journal;
3394 long ret;
3395 int block;
3396
3397 inode = iget_locked(sb, ino);
3398 if (!inode)
3399 return ERR_PTR(-ENOMEM);
3400 if (!(inode->i_state & I_NEW))
3401 return inode;
3402
3403 ei = EXT4_I(inode);
3404 iloc.bh = NULL;
3405
3406 ret = __ext4_get_inode_loc(inode, &iloc, 0);
3407 if (ret < 0)
3408 goto bad_inode;
3409 raw_inode = ext4_raw_inode(&iloc);
3410 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3411 inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3412 inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3413 if (!(test_opt(inode->i_sb, NO_UID32))) {
3414 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3415 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3416 }
3417 inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
3418
3419 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
3420 ei->i_dir_start_lookup = 0;
3421 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3422 /* We now have enough fields to check if the inode was active or not.
3423 * This is needed because nfsd might try to access dead inodes
3424 * the test is that same one that e2fsck uses
3425 * NeilBrown 1999oct15
3426 */
3427 if (inode->i_nlink == 0) {
3428 if (inode->i_mode == 0 ||
3429 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3430 /* this inode is deleted */
3431 ret = -ESTALE;
3432 goto bad_inode;
3433 }
3434 /* The only unlinked inodes we let through here have
3435 * valid i_mode and are being read by the orphan
3436 * recovery code: that's fine, we're about to complete
3437 * the process of deleting those. */
3438 }
3439 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3440 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3441 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3442 if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
3443 ei->i_file_acl |=
3444 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3445 inode->i_size = ext4_isize(raw_inode);
3446 ei->i_disksize = inode->i_size;
3447#ifdef CONFIG_QUOTA
3448 ei->i_reserved_quota = 0;
3449#endif
3450 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3451 ei->i_block_group = iloc.block_group;
3452 ei->i_last_alloc_group = ~0;
3453 /*
3454 * NOTE! The in-memory inode i_data array is in little-endian order
3455 * even on big-endian machines: we do NOT byteswap the block numbers!
3456 */
3457 for (block = 0; block < EXT4_N_BLOCKS; block++)
3458 ei->i_data[block] = raw_inode->i_block[block];
3459 INIT_LIST_HEAD(&ei->i_orphan);
3460
3461 /*
3462 * Set transaction id's of transactions that have to be committed
3463 * to finish f[data]sync. We set them to currently running transaction
3464 * as we cannot be sure that the inode or some of its metadata isn't
3465 * part of the transaction - the inode could have been reclaimed and
3466 * now it is reread from disk.
3467 */
3468 if (journal) {
3469 transaction_t *transaction;
3470 tid_t tid;
3471
3472 read_lock(&journal->j_state_lock);
3473 if (journal->j_running_transaction)
3474 transaction = journal->j_running_transaction;
3475 else
3476 transaction = journal->j_committing_transaction;
3477 if (transaction)
3478 tid = transaction->t_tid;
3479 else
3480 tid = journal->j_commit_sequence;
3481 read_unlock(&journal->j_state_lock);
3482 ei->i_sync_tid = tid;
3483 ei->i_datasync_tid = tid;
3484 }
3485
3486 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3487 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3488 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3489 EXT4_INODE_SIZE(inode->i_sb)) {
3490 ret = -EIO;
3491 goto bad_inode;
3492 }
3493 if (ei->i_extra_isize == 0) {
3494 /* The extra space is currently unused. Use it. */
3495 ei->i_extra_isize = sizeof(struct ext4_inode) -
3496 EXT4_GOOD_OLD_INODE_SIZE;
3497 } else {
3498 __le32 *magic = (void *)raw_inode +
3499 EXT4_GOOD_OLD_INODE_SIZE +
3500 ei->i_extra_isize;
3501 if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3502 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
3503 }
3504 } else
3505 ei->i_extra_isize = 0;
3506
3507 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
3508 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
3509 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
3510 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
3511
3512 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
3513 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3514 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3515 inode->i_version |=
3516 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
3517 }
3518
3519 ret = 0;
3520 if (ei->i_file_acl &&
3521 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
3522 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
3523 ei->i_file_acl);
3524 ret = -EIO;
3525 goto bad_inode;
3526 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3527 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3528 (S_ISLNK(inode->i_mode) &&
3529 !ext4_inode_is_fast_symlink(inode)))
3530 /* Validate extent which is part of inode */
3531 ret = ext4_ext_check_inode(inode);
3532 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3533 (S_ISLNK(inode->i_mode) &&
3534 !ext4_inode_is_fast_symlink(inode))) {
3535 /* Validate block references which are part of inode */
3536 ret = ext4_ind_check_inode(inode);
3537 }
3538 if (ret)
3539 goto bad_inode;
3540
3541 if (S_ISREG(inode->i_mode)) {
3542 inode->i_op = &ext4_file_inode_operations;
3543 inode->i_fop = &ext4_file_operations;
3544 ext4_set_aops(inode);
3545 } else if (S_ISDIR(inode->i_mode)) {
3546 inode->i_op = &ext4_dir_inode_operations;
3547 inode->i_fop = &ext4_dir_operations;
3548 } else if (S_ISLNK(inode->i_mode)) {
3549 if (ext4_inode_is_fast_symlink(inode)) {
3550 inode->i_op = &ext4_fast_symlink_inode_operations;
3551 nd_terminate_link(ei->i_data, inode->i_size,
3552 sizeof(ei->i_data) - 1);
3553 } else {
3554 inode->i_op = &ext4_symlink_inode_operations;
3555 ext4_set_aops(inode);
3556 }
3557 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
3558 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
3559 inode->i_op = &ext4_special_inode_operations;
3560 if (raw_inode->i_block[0])
3561 init_special_inode(inode, inode->i_mode,
3562 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3563 else
3564 init_special_inode(inode, inode->i_mode,
3565 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3566 } else {
3567 ret = -EIO;
3568 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
3569 goto bad_inode;
3570 }
3571 brelse(iloc.bh);
3572 ext4_set_inode_flags(inode);
3573 unlock_new_inode(inode);
3574 return inode;
3575
3576bad_inode:
3577 brelse(iloc.bh);
3578 iget_failed(inode);
3579 return ERR_PTR(ret);
3580}
3581
3582static int ext4_inode_blocks_set(handle_t *handle,
3583 struct ext4_inode *raw_inode,
3584 struct ext4_inode_info *ei)
3585{
3586 struct inode *inode = &(ei->vfs_inode);
3587 u64 i_blocks = inode->i_blocks;
3588 struct super_block *sb = inode->i_sb;
3589
3590 if (i_blocks <= ~0U) {
3591 /*
3592 * i_blocks can be represnted in a 32 bit variable
3593 * as multiple of 512 bytes
3594 */
3595 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3596 raw_inode->i_blocks_high = 0;
3597 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3598 return 0;
3599 }
3600 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
3601 return -EFBIG;
3602
3603 if (i_blocks <= 0xffffffffffffULL) {
3604 /*
3605 * i_blocks can be represented in a 48 bit variable
3606 * as multiple of 512 bytes
3607 */
3608 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3609 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3610 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3611 } else {
3612 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3613 /* i_block is stored in file system block size */
3614 i_blocks = i_blocks >> (inode->i_blkbits - 9);
3615 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3616 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3617 }
3618 return 0;
3619}
3620
3621/*
3622 * Post the struct inode info into an on-disk inode location in the
3623 * buffer-cache. This gobbles the caller's reference to the
3624 * buffer_head in the inode location struct.
3625 *
3626 * The caller must have write access to iloc->bh.
3627 */
3628static int ext4_do_update_inode(handle_t *handle,
3629 struct inode *inode,
3630 struct ext4_iloc *iloc)
3631{
3632 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
3633 struct ext4_inode_info *ei = EXT4_I(inode);
3634 struct buffer_head *bh = iloc->bh;
3635 int err = 0, rc, block;
3636
3637 /* For fields not not tracking in the in-memory inode,
3638 * initialise them to zero for new inodes. */
3639 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
3640 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
3641
3642 ext4_get_inode_flags(ei);
3643 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3644 if (!(test_opt(inode->i_sb, NO_UID32))) {
3645 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3646 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3647/*
3648 * Fix up interoperability with old kernels. Otherwise, old inodes get
3649 * re-used with the upper 16 bits of the uid/gid intact
3650 */
3651 if (!ei->i_dtime) {
3652 raw_inode->i_uid_high =
3653 cpu_to_le16(high_16_bits(inode->i_uid));
3654 raw_inode->i_gid_high =
3655 cpu_to_le16(high_16_bits(inode->i_gid));
3656 } else {
3657 raw_inode->i_uid_high = 0;
3658 raw_inode->i_gid_high = 0;
3659 }
3660 } else {
3661 raw_inode->i_uid_low =
3662 cpu_to_le16(fs_high2lowuid(inode->i_uid));
3663 raw_inode->i_gid_low =
3664 cpu_to_le16(fs_high2lowgid(inode->i_gid));
3665 raw_inode->i_uid_high = 0;
3666 raw_inode->i_gid_high = 0;
3667 }
3668 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3669
3670 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
3671 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
3672 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
3673 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
3674
3675 if (ext4_inode_blocks_set(handle, raw_inode, ei))
3676 goto out_brelse;
3677 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3678 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
3679 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
3680 cpu_to_le32(EXT4_OS_HURD))
3681 raw_inode->i_file_acl_high =
3682 cpu_to_le16(ei->i_file_acl >> 32);
3683 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
3684 ext4_isize_set(raw_inode, ei->i_disksize);
3685 if (ei->i_disksize > 0x7fffffffULL) {
3686 struct super_block *sb = inode->i_sb;
3687 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
3688 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
3689 EXT4_SB(sb)->s_es->s_rev_level ==
3690 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
3691 /* If this is the first large file
3692 * created, add a flag to the superblock.
3693 */
3694 err = ext4_journal_get_write_access(handle,
3695 EXT4_SB(sb)->s_sbh);
3696 if (err)
3697 goto out_brelse;
3698 ext4_update_dynamic_rev(sb);
3699 EXT4_SET_RO_COMPAT_FEATURE(sb,
3700 EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
3701 sb->s_dirt = 1;
3702 ext4_handle_sync(handle);
3703 err = ext4_handle_dirty_metadata(handle, NULL,
3704 EXT4_SB(sb)->s_sbh);
3705 }
3706 }
3707 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3708 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3709 if (old_valid_dev(inode->i_rdev)) {
3710 raw_inode->i_block[0] =
3711 cpu_to_le32(old_encode_dev(inode->i_rdev));
3712 raw_inode->i_block[1] = 0;
3713 } else {
3714 raw_inode->i_block[0] = 0;
3715 raw_inode->i_block[1] =
3716 cpu_to_le32(new_encode_dev(inode->i_rdev));
3717 raw_inode->i_block[2] = 0;
3718 }
3719 } else
3720 for (block = 0; block < EXT4_N_BLOCKS; block++)
3721 raw_inode->i_block[block] = ei->i_data[block];
3722
3723 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
3724 if (ei->i_extra_isize) {
3725 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3726 raw_inode->i_version_hi =
3727 cpu_to_le32(inode->i_version >> 32);
3728 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3729 }
3730
3731 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3732 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
3733 if (!err)
3734 err = rc;
3735 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
3736
3737 ext4_update_inode_fsync_trans(handle, inode, 0);
3738out_brelse:
3739 brelse(bh);
3740 ext4_std_error(inode->i_sb, err);
3741 return err;
3742}
3743
3744/*
3745 * ext4_write_inode()
3746 *
3747 * We are called from a few places:
3748 *
3749 * - Within generic_file_write() for O_SYNC files.
3750 * Here, there will be no transaction running. We wait for any running
3751 * trasnaction to commit.
3752 *
3753 * - Within sys_sync(), kupdate and such.
3754 * We wait on commit, if tol to.
3755 *
3756 * - Within prune_icache() (PF_MEMALLOC == true)
3757 * Here we simply return. We can't afford to block kswapd on the
3758 * journal commit.
3759 *
3760 * In all cases it is actually safe for us to return without doing anything,
3761 * because the inode has been copied into a raw inode buffer in
3762 * ext4_mark_inode_dirty(). This is a correctness thing for O_SYNC and for
3763 * knfsd.
3764 *
3765 * Note that we are absolutely dependent upon all inode dirtiers doing the
3766 * right thing: they *must* call mark_inode_dirty() after dirtying info in
3767 * which we are interested.
3768 *
3769 * It would be a bug for them to not do this. The code:
3770 *
3771 * mark_inode_dirty(inode)
3772 * stuff();
3773 * inode->i_size = expr;
3774 *
3775 * is in error because a kswapd-driven write_inode() could occur while
3776 * `stuff()' is running, and the new i_size will be lost. Plus the inode
3777 * will no longer be on the superblock's dirty inode list.
3778 */
3779int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
3780{
3781 int err;
3782
3783 if (current->flags & PF_MEMALLOC)
3784 return 0;
3785
3786 if (EXT4_SB(inode->i_sb)->s_journal) {
3787 if (ext4_journal_current_handle()) {
3788 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3789 dump_stack();
3790 return -EIO;
3791 }
3792
3793 if (wbc->sync_mode != WB_SYNC_ALL)
3794 return 0;
3795
3796 err = ext4_force_commit(inode->i_sb);
3797 } else {
3798 struct ext4_iloc iloc;
3799
3800 err = __ext4_get_inode_loc(inode, &iloc, 0);
3801 if (err)
3802 return err;
3803 if (wbc->sync_mode == WB_SYNC_ALL)
3804 sync_dirty_buffer(iloc.bh);
3805 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
3806 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
3807 "IO error syncing inode");
3808 err = -EIO;
3809 }
3810 brelse(iloc.bh);
3811 }
3812 return err;
3813}
3814
3815/*
3816 * ext4_setattr()
3817 *
3818 * Called from notify_change.
3819 *
3820 * We want to trap VFS attempts to truncate the file as soon as
3821 * possible. In particular, we want to make sure that when the VFS
3822 * shrinks i_size, we put the inode on the orphan list and modify
3823 * i_disksize immediately, so that during the subsequent flushing of
3824 * dirty pages and freeing of disk blocks, we can guarantee that any
3825 * commit will leave the blocks being flushed in an unused state on
3826 * disk. (On recovery, the inode will get truncated and the blocks will
3827 * be freed, so we have a strong guarantee that no future commit will
3828 * leave these blocks visible to the user.)
3829 *
3830 * Another thing we have to assure is that if we are in ordered mode
3831 * and inode is still attached to the committing transaction, we must
3832 * we start writeout of all the dirty pages which are being truncated.
3833 * This way we are sure that all the data written in the previous
3834 * transaction are already on disk (truncate waits for pages under
3835 * writeback).
3836 *
3837 * Called with inode->i_mutex down.
3838 */
3839int ext4_setattr(struct dentry *dentry, struct iattr *attr)
3840{
3841 struct inode *inode = dentry->d_inode;
3842 int error, rc = 0;
3843 int orphan = 0;
3844 const unsigned int ia_valid = attr->ia_valid;
3845
3846 error = inode_change_ok(inode, attr);
3847 if (error)
3848 return error;
3849
3850 if (is_quota_modification(inode, attr))
3851 dquot_initialize(inode);
3852 if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3853 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3854 handle_t *handle;
3855
3856 /* (user+group)*(old+new) structure, inode write (sb,
3857 * inode block, ? - but truncate inode update has it) */
3858 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3859 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
3860 if (IS_ERR(handle)) {
3861 error = PTR_ERR(handle);
3862 goto err_out;
3863 }
3864 error = dquot_transfer(inode, attr);
3865 if (error) {
3866 ext4_journal_stop(handle);
3867 return error;
3868 }
3869 /* Update corresponding info in inode so that everything is in
3870 * one transaction */
3871 if (attr->ia_valid & ATTR_UID)
3872 inode->i_uid = attr->ia_uid;
3873 if (attr->ia_valid & ATTR_GID)
3874 inode->i_gid = attr->ia_gid;
3875 error = ext4_mark_inode_dirty(handle, inode);
3876 ext4_journal_stop(handle);
3877 }
3878
3879 if (attr->ia_valid & ATTR_SIZE) {
3880 inode_dio_wait(inode);
3881
3882 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3883 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3884
3885 if (attr->ia_size > sbi->s_bitmap_maxbytes)
3886 return -EFBIG;
3887 }
3888 }
3889
3890 if (S_ISREG(inode->i_mode) &&
3891 attr->ia_valid & ATTR_SIZE &&
3892 (attr->ia_size < inode->i_size)) {
3893 handle_t *handle;
3894
3895 handle = ext4_journal_start(inode, 3);
3896 if (IS_ERR(handle)) {
3897 error = PTR_ERR(handle);
3898 goto err_out;
3899 }
3900 if (ext4_handle_valid(handle)) {
3901 error = ext4_orphan_add(handle, inode);
3902 orphan = 1;
3903 }
3904 EXT4_I(inode)->i_disksize = attr->ia_size;
3905 rc = ext4_mark_inode_dirty(handle, inode);
3906 if (!error)
3907 error = rc;
3908 ext4_journal_stop(handle);
3909
3910 if (ext4_should_order_data(inode)) {
3911 error = ext4_begin_ordered_truncate(inode,
3912 attr->ia_size);
3913 if (error) {
3914 /* Do as much error cleanup as possible */
3915 handle = ext4_journal_start(inode, 3);
3916 if (IS_ERR(handle)) {
3917 ext4_orphan_del(NULL, inode);
3918 goto err_out;
3919 }
3920 ext4_orphan_del(handle, inode);
3921 orphan = 0;
3922 ext4_journal_stop(handle);
3923 goto err_out;
3924 }
3925 }
3926 }
3927
3928 if (attr->ia_valid & ATTR_SIZE) {
3929 if (attr->ia_size != i_size_read(inode)) {
3930 truncate_setsize(inode, attr->ia_size);
3931 ext4_truncate(inode);
3932 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3933 ext4_truncate(inode);
3934 }
3935
3936 if (!rc) {
3937 setattr_copy(inode, attr);
3938 mark_inode_dirty(inode);
3939 }
3940
3941 /*
3942 * If the call to ext4_truncate failed to get a transaction handle at
3943 * all, we need to clean up the in-core orphan list manually.
3944 */
3945 if (orphan && inode->i_nlink)
3946 ext4_orphan_del(NULL, inode);
3947
3948 if (!rc && (ia_valid & ATTR_MODE))
3949 rc = ext4_acl_chmod(inode);
3950
3951err_out:
3952 ext4_std_error(inode->i_sb, error);
3953 if (!error)
3954 error = rc;
3955 return error;
3956}
3957
3958int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
3959 struct kstat *stat)
3960{
3961 struct inode *inode;
3962 unsigned long delalloc_blocks;
3963
3964 inode = dentry->d_inode;
3965 generic_fillattr(inode, stat);
3966
3967 /*
3968 * We can't update i_blocks if the block allocation is delayed
3969 * otherwise in the case of system crash before the real block
3970 * allocation is done, we will have i_blocks inconsistent with
3971 * on-disk file blocks.
3972 * We always keep i_blocks updated together with real
3973 * allocation. But to not confuse with user, stat
3974 * will return the blocks that include the delayed allocation
3975 * blocks for this file.
3976 */
3977 delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
3978
3979 stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
3980 return 0;
3981}
3982
3983static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
3984{
3985 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
3986 return ext4_ind_trans_blocks(inode, nrblocks, chunk);
3987 return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
3988}
3989
3990/*
3991 * Account for index blocks, block groups bitmaps and block group
3992 * descriptor blocks if modify datablocks and index blocks
3993 * worse case, the indexs blocks spread over different block groups
3994 *
3995 * If datablocks are discontiguous, they are possible to spread over
3996 * different block groups too. If they are contiuguous, with flexbg,
3997 * they could still across block group boundary.
3998 *
3999 * Also account for superblock, inode, quota and xattr blocks
4000 */
4001static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4002{
4003 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4004 int gdpblocks;
4005 int idxblocks;
4006 int ret = 0;
4007
4008 /*
4009 * How many index blocks need to touch to modify nrblocks?
4010 * The "Chunk" flag indicating whether the nrblocks is
4011 * physically contiguous on disk
4012 *
4013 * For Direct IO and fallocate, they calls get_block to allocate
4014 * one single extent at a time, so they could set the "Chunk" flag
4015 */
4016 idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4017
4018 ret = idxblocks;
4019
4020 /*
4021 * Now let's see how many group bitmaps and group descriptors need
4022 * to account
4023 */
4024 groups = idxblocks;
4025 if (chunk)
4026 groups += 1;
4027 else
4028 groups += nrblocks;
4029
4030 gdpblocks = groups;
4031 if (groups > ngroups)
4032 groups = ngroups;
4033 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4034 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4035
4036 /* bitmaps and block group descriptor blocks */
4037 ret += groups + gdpblocks;
4038
4039 /* Blocks for super block, inode, quota and xattr blocks */
4040 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4041
4042 return ret;
4043}
4044
4045/*
4046 * Calculate the total number of credits to reserve to fit
4047 * the modification of a single pages into a single transaction,
4048 * which may include multiple chunks of block allocations.
4049 *
4050 * This could be called via ext4_write_begin()
4051 *
4052 * We need to consider the worse case, when
4053 * one new block per extent.
4054 */
4055int ext4_writepage_trans_blocks(struct inode *inode)
4056{
4057 int bpp = ext4_journal_blocks_per_page(inode);
4058 int ret;
4059
4060 ret = ext4_meta_trans_blocks(inode, bpp, 0);
4061
4062 /* Account for data blocks for journalled mode */
4063 if (ext4_should_journal_data(inode))
4064 ret += bpp;
4065 return ret;
4066}
4067
4068/*
4069 * Calculate the journal credits for a chunk of data modification.
4070 *
4071 * This is called from DIO, fallocate or whoever calling
4072 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4073 *
4074 * journal buffers for data blocks are not included here, as DIO
4075 * and fallocate do no need to journal data buffers.
4076 */
4077int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4078{
4079 return ext4_meta_trans_blocks(inode, nrblocks, 1);
4080}
4081
4082/*
4083 * The caller must have previously called ext4_reserve_inode_write().
4084 * Give this, we know that the caller already has write access to iloc->bh.
4085 */
4086int ext4_mark_iloc_dirty(handle_t *handle,
4087 struct inode *inode, struct ext4_iloc *iloc)
4088{
4089 int err = 0;
4090
4091 if (test_opt(inode->i_sb, I_VERSION))
4092 inode_inc_iversion(inode);
4093
4094 /* the do_update_inode consumes one bh->b_count */
4095 get_bh(iloc->bh);
4096
4097 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4098 err = ext4_do_update_inode(handle, inode, iloc);
4099 put_bh(iloc->bh);
4100 return err;
4101}
4102
4103/*
4104 * On success, We end up with an outstanding reference count against
4105 * iloc->bh. This _must_ be cleaned up later.
4106 */
4107
4108int
4109ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4110 struct ext4_iloc *iloc)
4111{
4112 int err;
4113
4114 err = ext4_get_inode_loc(inode, iloc);
4115 if (!err) {
4116 BUFFER_TRACE(iloc->bh, "get_write_access");
4117 err = ext4_journal_get_write_access(handle, iloc->bh);
4118 if (err) {
4119 brelse(iloc->bh);
4120 iloc->bh = NULL;
4121 }
4122 }
4123 ext4_std_error(inode->i_sb, err);
4124 return err;
4125}
4126
4127/*
4128 * Expand an inode by new_extra_isize bytes.
4129 * Returns 0 on success or negative error number on failure.
4130 */
4131static int ext4_expand_extra_isize(struct inode *inode,
4132 unsigned int new_extra_isize,
4133 struct ext4_iloc iloc,
4134 handle_t *handle)
4135{
4136 struct ext4_inode *raw_inode;
4137 struct ext4_xattr_ibody_header *header;
4138
4139 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4140 return 0;
4141
4142 raw_inode = ext4_raw_inode(&iloc);
4143
4144 header = IHDR(inode, raw_inode);
4145
4146 /* No extended attributes present */
4147 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4148 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4149 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4150 new_extra_isize);
4151 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4152 return 0;
4153 }
4154
4155 /* try to expand with EAs present */
4156 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4157 raw_inode, handle);
4158}
4159
4160/*
4161 * What we do here is to mark the in-core inode as clean with respect to inode
4162 * dirtiness (it may still be data-dirty).
4163 * This means that the in-core inode may be reaped by prune_icache
4164 * without having to perform any I/O. This is a very good thing,
4165 * because *any* task may call prune_icache - even ones which
4166 * have a transaction open against a different journal.
4167 *
4168 * Is this cheating? Not really. Sure, we haven't written the
4169 * inode out, but prune_icache isn't a user-visible syncing function.
4170 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4171 * we start and wait on commits.
4172 *
4173 * Is this efficient/effective? Well, we're being nice to the system
4174 * by cleaning up our inodes proactively so they can be reaped
4175 * without I/O. But we are potentially leaving up to five seconds'
4176 * worth of inodes floating about which prune_icache wants us to
4177 * write out. One way to fix that would be to get prune_icache()
4178 * to do a write_super() to free up some memory. It has the desired
4179 * effect.
4180 */
4181int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4182{
4183 struct ext4_iloc iloc;
4184 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4185 static unsigned int mnt_count;
4186 int err, ret;
4187
4188 might_sleep();
4189 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4190 err = ext4_reserve_inode_write(handle, inode, &iloc);
4191 if (ext4_handle_valid(handle) &&
4192 EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4193 !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4194 /*
4195 * We need extra buffer credits since we may write into EA block
4196 * with this same handle. If journal_extend fails, then it will
4197 * only result in a minor loss of functionality for that inode.
4198 * If this is felt to be critical, then e2fsck should be run to
4199 * force a large enough s_min_extra_isize.
4200 */
4201 if ((jbd2_journal_extend(handle,
4202 EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4203 ret = ext4_expand_extra_isize(inode,
4204 sbi->s_want_extra_isize,
4205 iloc, handle);
4206 if (ret) {
4207 ext4_set_inode_state(inode,
4208 EXT4_STATE_NO_EXPAND);
4209 if (mnt_count !=
4210 le16_to_cpu(sbi->s_es->s_mnt_count)) {
4211 ext4_warning(inode->i_sb,
4212 "Unable to expand inode %lu. Delete"
4213 " some EAs or run e2fsck.",
4214 inode->i_ino);
4215 mnt_count =
4216 le16_to_cpu(sbi->s_es->s_mnt_count);
4217 }
4218 }
4219 }
4220 }
4221 if (!err)
4222 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4223 return err;
4224}
4225
4226/*
4227 * ext4_dirty_inode() is called from __mark_inode_dirty()
4228 *
4229 * We're really interested in the case where a file is being extended.
4230 * i_size has been changed by generic_commit_write() and we thus need
4231 * to include the updated inode in the current transaction.
4232 *
4233 * Also, dquot_alloc_block() will always dirty the inode when blocks
4234 * are allocated to the file.
4235 *
4236 * If the inode is marked synchronous, we don't honour that here - doing
4237 * so would cause a commit on atime updates, which we don't bother doing.
4238 * We handle synchronous inodes at the highest possible level.
4239 */
4240void ext4_dirty_inode(struct inode *inode, int flags)
4241{
4242 handle_t *handle;
4243
4244 handle = ext4_journal_start(inode, 2);
4245 if (IS_ERR(handle))
4246 goto out;
4247
4248 ext4_mark_inode_dirty(handle, inode);
4249
4250 ext4_journal_stop(handle);
4251out:
4252 return;
4253}
4254
4255#if 0
4256/*
4257 * Bind an inode's backing buffer_head into this transaction, to prevent
4258 * it from being flushed to disk early. Unlike
4259 * ext4_reserve_inode_write, this leaves behind no bh reference and
4260 * returns no iloc structure, so the caller needs to repeat the iloc
4261 * lookup to mark the inode dirty later.
4262 */
4263static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4264{
4265 struct ext4_iloc iloc;
4266
4267 int err = 0;
4268 if (handle) {
4269 err = ext4_get_inode_loc(inode, &iloc);
4270 if (!err) {
4271 BUFFER_TRACE(iloc.bh, "get_write_access");
4272 err = jbd2_journal_get_write_access(handle, iloc.bh);
4273 if (!err)
4274 err = ext4_handle_dirty_metadata(handle,
4275 NULL,
4276 iloc.bh);
4277 brelse(iloc.bh);
4278 }
4279 }
4280 ext4_std_error(inode->i_sb, err);
4281 return err;
4282}
4283#endif
4284
4285int ext4_change_inode_journal_flag(struct inode *inode, int val)
4286{
4287 journal_t *journal;
4288 handle_t *handle;
4289 int err;
4290
4291 /*
4292 * We have to be very careful here: changing a data block's
4293 * journaling status dynamically is dangerous. If we write a
4294 * data block to the journal, change the status and then delete
4295 * that block, we risk forgetting to revoke the old log record
4296 * from the journal and so a subsequent replay can corrupt data.
4297 * So, first we make sure that the journal is empty and that
4298 * nobody is changing anything.
4299 */
4300
4301 journal = EXT4_JOURNAL(inode);
4302 if (!journal)
4303 return 0;
4304 if (is_journal_aborted(journal))
4305 return -EROFS;
4306
4307 jbd2_journal_lock_updates(journal);
4308 jbd2_journal_flush(journal);
4309
4310 /*
4311 * OK, there are no updates running now, and all cached data is
4312 * synced to disk. We are now in a completely consistent state
4313 * which doesn't have anything in the journal, and we know that
4314 * no filesystem updates are running, so it is safe to modify
4315 * the inode's in-core data-journaling state flag now.
4316 */
4317
4318 if (val)
4319 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4320 else
4321 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4322 ext4_set_aops(inode);
4323
4324 jbd2_journal_unlock_updates(journal);
4325
4326 /* Finally we can mark the inode as dirty. */
4327
4328 handle = ext4_journal_start(inode, 1);
4329 if (IS_ERR(handle))
4330 return PTR_ERR(handle);
4331
4332 err = ext4_mark_inode_dirty(handle, inode);
4333 ext4_handle_sync(handle);
4334 ext4_journal_stop(handle);
4335 ext4_std_error(inode->i_sb, err);
4336
4337 return err;
4338}
4339
4340static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4341{
4342 return !buffer_mapped(bh);
4343}
4344
4345int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
4346{
4347 struct page *page = vmf->page;
4348 loff_t size;
4349 unsigned long len;
4350 int ret;
4351 struct file *file = vma->vm_file;
4352 struct inode *inode = file->f_path.dentry->d_inode;
4353 struct address_space *mapping = inode->i_mapping;
4354 handle_t *handle;
4355 get_block_t *get_block;
4356 int retries = 0;
4357
4358 /*
4359 * This check is racy but catches the common case. We rely on
4360 * __block_page_mkwrite() to do a reliable check.
4361 */
4362 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
4363 /* Delalloc case is easy... */
4364 if (test_opt(inode->i_sb, DELALLOC) &&
4365 !ext4_should_journal_data(inode) &&
4366 !ext4_nonda_switch(inode->i_sb)) {
4367 do {
4368 ret = __block_page_mkwrite(vma, vmf,
4369 ext4_da_get_block_prep);
4370 } while (ret == -ENOSPC &&
4371 ext4_should_retry_alloc(inode->i_sb, &retries));
4372 goto out_ret;
4373 }
4374
4375 lock_page(page);
4376 size = i_size_read(inode);
4377 /* Page got truncated from under us? */
4378 if (page->mapping != mapping || page_offset(page) > size) {
4379 unlock_page(page);
4380 ret = VM_FAULT_NOPAGE;
4381 goto out;
4382 }
4383
4384 if (page->index == size >> PAGE_CACHE_SHIFT)
4385 len = size & ~PAGE_CACHE_MASK;
4386 else
4387 len = PAGE_CACHE_SIZE;
4388 /*
4389 * Return if we have all the buffers mapped. This avoids the need to do
4390 * journal_start/journal_stop which can block and take a long time
4391 */
4392 if (page_has_buffers(page)) {
4393 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4394 ext4_bh_unmapped)) {
4395 /* Wait so that we don't change page under IO */
4396 wait_on_page_writeback(page);
4397 ret = VM_FAULT_LOCKED;
4398 goto out;
4399 }
4400 }
4401 unlock_page(page);
4402 /* OK, we need to fill the hole... */
4403 if (ext4_should_dioread_nolock(inode))
4404 get_block = ext4_get_block_write;
4405 else
4406 get_block = ext4_get_block;
4407retry_alloc:
4408 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
4409 if (IS_ERR(handle)) {
4410 ret = VM_FAULT_SIGBUS;
4411 goto out;
4412 }
4413 ret = __block_page_mkwrite(vma, vmf, get_block);
4414 if (!ret && ext4_should_journal_data(inode)) {
4415 if (walk_page_buffers(handle, page_buffers(page), 0,
4416 PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
4417 unlock_page(page);
4418 ret = VM_FAULT_SIGBUS;
4419 goto out;
4420 }
4421 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
4422 }
4423 ext4_journal_stop(handle);
4424 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
4425 goto retry_alloc;
4426out_ret:
4427 ret = block_page_mkwrite_return(ret);
4428out:
4429 return ret;
4430}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/inode.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/inode.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * 64-bit file support on 64-bit platforms by Jakub Jelinek
17 * (jj@sunsite.ms.mff.cuni.cz)
18 *
19 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20 */
21
22#include <linux/fs.h>
23#include <linux/time.h>
24#include <linux/highuid.h>
25#include <linux/pagemap.h>
26#include <linux/dax.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/bitops.h>
41#include <linux/iomap.h>
42#include <linux/iversion.h>
43
44#include "ext4_jbd2.h"
45#include "xattr.h"
46#include "acl.h"
47#include "truncate.h"
48
49#include <trace/events/ext4.h>
50
51#define MPAGE_DA_EXTENT_TAIL 0x01
52
53static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
54 struct ext4_inode_info *ei)
55{
56 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
57 __u32 csum;
58 __u16 dummy_csum = 0;
59 int offset = offsetof(struct ext4_inode, i_checksum_lo);
60 unsigned int csum_size = sizeof(dummy_csum);
61
62 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
63 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
64 offset += csum_size;
65 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
66 EXT4_GOOD_OLD_INODE_SIZE - offset);
67
68 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
69 offset = offsetof(struct ext4_inode, i_checksum_hi);
70 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
71 EXT4_GOOD_OLD_INODE_SIZE,
72 offset - EXT4_GOOD_OLD_INODE_SIZE);
73 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
74 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
75 csum_size);
76 offset += csum_size;
77 }
78 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
79 EXT4_INODE_SIZE(inode->i_sb) - offset);
80 }
81
82 return csum;
83}
84
85static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
86 struct ext4_inode_info *ei)
87{
88 __u32 provided, calculated;
89
90 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
91 cpu_to_le32(EXT4_OS_LINUX) ||
92 !ext4_has_metadata_csum(inode->i_sb))
93 return 1;
94
95 provided = le16_to_cpu(raw->i_checksum_lo);
96 calculated = ext4_inode_csum(inode, raw, ei);
97 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
98 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
99 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
100 else
101 calculated &= 0xFFFF;
102
103 return provided == calculated;
104}
105
106static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
107 struct ext4_inode_info *ei)
108{
109 __u32 csum;
110
111 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
112 cpu_to_le32(EXT4_OS_LINUX) ||
113 !ext4_has_metadata_csum(inode->i_sb))
114 return;
115
116 csum = ext4_inode_csum(inode, raw, ei);
117 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
118 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
119 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
120 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
121}
122
123static inline int ext4_begin_ordered_truncate(struct inode *inode,
124 loff_t new_size)
125{
126 trace_ext4_begin_ordered_truncate(inode, new_size);
127 /*
128 * If jinode is zero, then we never opened the file for
129 * writing, so there's no need to call
130 * jbd2_journal_begin_ordered_truncate() since there's no
131 * outstanding writes we need to flush.
132 */
133 if (!EXT4_I(inode)->jinode)
134 return 0;
135 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
136 EXT4_I(inode)->jinode,
137 new_size);
138}
139
140static void ext4_invalidatepage(struct page *page, unsigned int offset,
141 unsigned int length);
142static int __ext4_journalled_writepage(struct page *page, unsigned int len);
143static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
144static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
145 int pextents);
146
147/*
148 * Test whether an inode is a fast symlink.
149 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
150 */
151int ext4_inode_is_fast_symlink(struct inode *inode)
152{
153 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
154 int ea_blocks = EXT4_I(inode)->i_file_acl ?
155 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
156
157 if (ext4_has_inline_data(inode))
158 return 0;
159
160 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
161 }
162 return S_ISLNK(inode->i_mode) && inode->i_size &&
163 (inode->i_size < EXT4_N_BLOCKS * 4);
164}
165
166/*
167 * Restart the transaction associated with *handle. This does a commit,
168 * so before we call here everything must be consistently dirtied against
169 * this transaction.
170 */
171int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
172 int nblocks)
173{
174 int ret;
175
176 /*
177 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
178 * moment, get_block can be called only for blocks inside i_size since
179 * page cache has been already dropped and writes are blocked by
180 * i_mutex. So we can safely drop the i_data_sem here.
181 */
182 BUG_ON(EXT4_JOURNAL(inode) == NULL);
183 jbd_debug(2, "restarting handle %p\n", handle);
184 up_write(&EXT4_I(inode)->i_data_sem);
185 ret = ext4_journal_restart(handle, nblocks);
186 down_write(&EXT4_I(inode)->i_data_sem);
187 ext4_discard_preallocations(inode);
188
189 return ret;
190}
191
192/*
193 * Called at the last iput() if i_nlink is zero.
194 */
195void ext4_evict_inode(struct inode *inode)
196{
197 handle_t *handle;
198 int err;
199 int extra_credits = 3;
200 struct ext4_xattr_inode_array *ea_inode_array = NULL;
201
202 trace_ext4_evict_inode(inode);
203
204 if (inode->i_nlink) {
205 /*
206 * When journalling data dirty buffers are tracked only in the
207 * journal. So although mm thinks everything is clean and
208 * ready for reaping the inode might still have some pages to
209 * write in the running transaction or waiting to be
210 * checkpointed. Thus calling jbd2_journal_invalidatepage()
211 * (via truncate_inode_pages()) to discard these buffers can
212 * cause data loss. Also even if we did not discard these
213 * buffers, we would have no way to find them after the inode
214 * is reaped and thus user could see stale data if he tries to
215 * read them before the transaction is checkpointed. So be
216 * careful and force everything to disk here... We use
217 * ei->i_datasync_tid to store the newest transaction
218 * containing inode's data.
219 *
220 * Note that directories do not have this problem because they
221 * don't use page cache.
222 */
223 if (inode->i_ino != EXT4_JOURNAL_INO &&
224 ext4_should_journal_data(inode) &&
225 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
226 inode->i_data.nrpages) {
227 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
228 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
229
230 jbd2_complete_transaction(journal, commit_tid);
231 filemap_write_and_wait(&inode->i_data);
232 }
233 truncate_inode_pages_final(&inode->i_data);
234
235 goto no_delete;
236 }
237
238 if (is_bad_inode(inode))
239 goto no_delete;
240 dquot_initialize(inode);
241
242 if (ext4_should_order_data(inode))
243 ext4_begin_ordered_truncate(inode, 0);
244 truncate_inode_pages_final(&inode->i_data);
245
246 /*
247 * Protect us against freezing - iput() caller didn't have to have any
248 * protection against it
249 */
250 sb_start_intwrite(inode->i_sb);
251
252 if (!IS_NOQUOTA(inode))
253 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
254
255 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
256 ext4_blocks_for_truncate(inode)+extra_credits);
257 if (IS_ERR(handle)) {
258 ext4_std_error(inode->i_sb, PTR_ERR(handle));
259 /*
260 * If we're going to skip the normal cleanup, we still need to
261 * make sure that the in-core orphan linked list is properly
262 * cleaned up.
263 */
264 ext4_orphan_del(NULL, inode);
265 sb_end_intwrite(inode->i_sb);
266 goto no_delete;
267 }
268
269 if (IS_SYNC(inode))
270 ext4_handle_sync(handle);
271
272 /*
273 * Set inode->i_size to 0 before calling ext4_truncate(). We need
274 * special handling of symlinks here because i_size is used to
275 * determine whether ext4_inode_info->i_data contains symlink data or
276 * block mappings. Setting i_size to 0 will remove its fast symlink
277 * status. Erase i_data so that it becomes a valid empty block map.
278 */
279 if (ext4_inode_is_fast_symlink(inode))
280 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
281 inode->i_size = 0;
282 err = ext4_mark_inode_dirty(handle, inode);
283 if (err) {
284 ext4_warning(inode->i_sb,
285 "couldn't mark inode dirty (err %d)", err);
286 goto stop_handle;
287 }
288 if (inode->i_blocks) {
289 err = ext4_truncate(inode);
290 if (err) {
291 ext4_error(inode->i_sb,
292 "couldn't truncate inode %lu (err %d)",
293 inode->i_ino, err);
294 goto stop_handle;
295 }
296 }
297
298 /* Remove xattr references. */
299 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
300 extra_credits);
301 if (err) {
302 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
303stop_handle:
304 ext4_journal_stop(handle);
305 ext4_orphan_del(NULL, inode);
306 sb_end_intwrite(inode->i_sb);
307 ext4_xattr_inode_array_free(ea_inode_array);
308 goto no_delete;
309 }
310
311 /*
312 * Kill off the orphan record which ext4_truncate created.
313 * AKPM: I think this can be inside the above `if'.
314 * Note that ext4_orphan_del() has to be able to cope with the
315 * deletion of a non-existent orphan - this is because we don't
316 * know if ext4_truncate() actually created an orphan record.
317 * (Well, we could do this if we need to, but heck - it works)
318 */
319 ext4_orphan_del(handle, inode);
320 EXT4_I(inode)->i_dtime = get_seconds();
321
322 /*
323 * One subtle ordering requirement: if anything has gone wrong
324 * (transaction abort, IO errors, whatever), then we can still
325 * do these next steps (the fs will already have been marked as
326 * having errors), but we can't free the inode if the mark_dirty
327 * fails.
328 */
329 if (ext4_mark_inode_dirty(handle, inode))
330 /* If that failed, just do the required in-core inode clear. */
331 ext4_clear_inode(inode);
332 else
333 ext4_free_inode(handle, inode);
334 ext4_journal_stop(handle);
335 sb_end_intwrite(inode->i_sb);
336 ext4_xattr_inode_array_free(ea_inode_array);
337 return;
338no_delete:
339 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
340}
341
342#ifdef CONFIG_QUOTA
343qsize_t *ext4_get_reserved_space(struct inode *inode)
344{
345 return &EXT4_I(inode)->i_reserved_quota;
346}
347#endif
348
349/*
350 * Called with i_data_sem down, which is important since we can call
351 * ext4_discard_preallocations() from here.
352 */
353void ext4_da_update_reserve_space(struct inode *inode,
354 int used, int quota_claim)
355{
356 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
357 struct ext4_inode_info *ei = EXT4_I(inode);
358
359 spin_lock(&ei->i_block_reservation_lock);
360 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
361 if (unlikely(used > ei->i_reserved_data_blocks)) {
362 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
363 "with only %d reserved data blocks",
364 __func__, inode->i_ino, used,
365 ei->i_reserved_data_blocks);
366 WARN_ON(1);
367 used = ei->i_reserved_data_blocks;
368 }
369
370 /* Update per-inode reservations */
371 ei->i_reserved_data_blocks -= used;
372 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
373
374 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
375
376 /* Update quota subsystem for data blocks */
377 if (quota_claim)
378 dquot_claim_block(inode, EXT4_C2B(sbi, used));
379 else {
380 /*
381 * We did fallocate with an offset that is already delayed
382 * allocated. So on delayed allocated writeback we should
383 * not re-claim the quota for fallocated blocks.
384 */
385 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
386 }
387
388 /*
389 * If we have done all the pending block allocations and if
390 * there aren't any writers on the inode, we can discard the
391 * inode's preallocations.
392 */
393 if ((ei->i_reserved_data_blocks == 0) &&
394 (atomic_read(&inode->i_writecount) == 0))
395 ext4_discard_preallocations(inode);
396}
397
398static int __check_block_validity(struct inode *inode, const char *func,
399 unsigned int line,
400 struct ext4_map_blocks *map)
401{
402 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
403 map->m_len)) {
404 ext4_error_inode(inode, func, line, map->m_pblk,
405 "lblock %lu mapped to illegal pblock "
406 "(length %d)", (unsigned long) map->m_lblk,
407 map->m_len);
408 return -EFSCORRUPTED;
409 }
410 return 0;
411}
412
413int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
414 ext4_lblk_t len)
415{
416 int ret;
417
418 if (ext4_encrypted_inode(inode))
419 return fscrypt_zeroout_range(inode, lblk, pblk, len);
420
421 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
422 if (ret > 0)
423 ret = 0;
424
425 return ret;
426}
427
428#define check_block_validity(inode, map) \
429 __check_block_validity((inode), __func__, __LINE__, (map))
430
431#ifdef ES_AGGRESSIVE_TEST
432static void ext4_map_blocks_es_recheck(handle_t *handle,
433 struct inode *inode,
434 struct ext4_map_blocks *es_map,
435 struct ext4_map_blocks *map,
436 int flags)
437{
438 int retval;
439
440 map->m_flags = 0;
441 /*
442 * There is a race window that the result is not the same.
443 * e.g. xfstests #223 when dioread_nolock enables. The reason
444 * is that we lookup a block mapping in extent status tree with
445 * out taking i_data_sem. So at the time the unwritten extent
446 * could be converted.
447 */
448 down_read(&EXT4_I(inode)->i_data_sem);
449 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
450 retval = ext4_ext_map_blocks(handle, inode, map, flags &
451 EXT4_GET_BLOCKS_KEEP_SIZE);
452 } else {
453 retval = ext4_ind_map_blocks(handle, inode, map, flags &
454 EXT4_GET_BLOCKS_KEEP_SIZE);
455 }
456 up_read((&EXT4_I(inode)->i_data_sem));
457
458 /*
459 * We don't check m_len because extent will be collpased in status
460 * tree. So the m_len might not equal.
461 */
462 if (es_map->m_lblk != map->m_lblk ||
463 es_map->m_flags != map->m_flags ||
464 es_map->m_pblk != map->m_pblk) {
465 printk("ES cache assertion failed for inode: %lu "
466 "es_cached ex [%d/%d/%llu/%x] != "
467 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
468 inode->i_ino, es_map->m_lblk, es_map->m_len,
469 es_map->m_pblk, es_map->m_flags, map->m_lblk,
470 map->m_len, map->m_pblk, map->m_flags,
471 retval, flags);
472 }
473}
474#endif /* ES_AGGRESSIVE_TEST */
475
476/*
477 * The ext4_map_blocks() function tries to look up the requested blocks,
478 * and returns if the blocks are already mapped.
479 *
480 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
481 * and store the allocated blocks in the result buffer head and mark it
482 * mapped.
483 *
484 * If file type is extents based, it will call ext4_ext_map_blocks(),
485 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
486 * based files
487 *
488 * On success, it returns the number of blocks being mapped or allocated. if
489 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
490 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
491 *
492 * It returns 0 if plain look up failed (blocks have not been allocated), in
493 * that case, @map is returned as unmapped but we still do fill map->m_len to
494 * indicate the length of a hole starting at map->m_lblk.
495 *
496 * It returns the error in case of allocation failure.
497 */
498int ext4_map_blocks(handle_t *handle, struct inode *inode,
499 struct ext4_map_blocks *map, int flags)
500{
501 struct extent_status es;
502 int retval;
503 int ret = 0;
504#ifdef ES_AGGRESSIVE_TEST
505 struct ext4_map_blocks orig_map;
506
507 memcpy(&orig_map, map, sizeof(*map));
508#endif
509
510 map->m_flags = 0;
511 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
512 "logical block %lu\n", inode->i_ino, flags, map->m_len,
513 (unsigned long) map->m_lblk);
514
515 /*
516 * ext4_map_blocks returns an int, and m_len is an unsigned int
517 */
518 if (unlikely(map->m_len > INT_MAX))
519 map->m_len = INT_MAX;
520
521 /* We can handle the block number less than EXT_MAX_BLOCKS */
522 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
523 return -EFSCORRUPTED;
524
525 /* Lookup extent status tree firstly */
526 if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
527 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
528 map->m_pblk = ext4_es_pblock(&es) +
529 map->m_lblk - es.es_lblk;
530 map->m_flags |= ext4_es_is_written(&es) ?
531 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
532 retval = es.es_len - (map->m_lblk - es.es_lblk);
533 if (retval > map->m_len)
534 retval = map->m_len;
535 map->m_len = retval;
536 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
537 map->m_pblk = 0;
538 retval = es.es_len - (map->m_lblk - es.es_lblk);
539 if (retval > map->m_len)
540 retval = map->m_len;
541 map->m_len = retval;
542 retval = 0;
543 } else {
544 BUG_ON(1);
545 }
546#ifdef ES_AGGRESSIVE_TEST
547 ext4_map_blocks_es_recheck(handle, inode, map,
548 &orig_map, flags);
549#endif
550 goto found;
551 }
552
553 /*
554 * Try to see if we can get the block without requesting a new
555 * file system block.
556 */
557 down_read(&EXT4_I(inode)->i_data_sem);
558 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
559 retval = ext4_ext_map_blocks(handle, inode, map, flags &
560 EXT4_GET_BLOCKS_KEEP_SIZE);
561 } else {
562 retval = ext4_ind_map_blocks(handle, inode, map, flags &
563 EXT4_GET_BLOCKS_KEEP_SIZE);
564 }
565 if (retval > 0) {
566 unsigned int status;
567
568 if (unlikely(retval != map->m_len)) {
569 ext4_warning(inode->i_sb,
570 "ES len assertion failed for inode "
571 "%lu: retval %d != map->m_len %d",
572 inode->i_ino, retval, map->m_len);
573 WARN_ON(1);
574 }
575
576 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
577 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
578 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
579 !(status & EXTENT_STATUS_WRITTEN) &&
580 ext4_find_delalloc_range(inode, map->m_lblk,
581 map->m_lblk + map->m_len - 1))
582 status |= EXTENT_STATUS_DELAYED;
583 ret = ext4_es_insert_extent(inode, map->m_lblk,
584 map->m_len, map->m_pblk, status);
585 if (ret < 0)
586 retval = ret;
587 }
588 up_read((&EXT4_I(inode)->i_data_sem));
589
590found:
591 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
592 ret = check_block_validity(inode, map);
593 if (ret != 0)
594 return ret;
595 }
596
597 /* If it is only a block(s) look up */
598 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
599 return retval;
600
601 /*
602 * Returns if the blocks have already allocated
603 *
604 * Note that if blocks have been preallocated
605 * ext4_ext_get_block() returns the create = 0
606 * with buffer head unmapped.
607 */
608 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
609 /*
610 * If we need to convert extent to unwritten
611 * we continue and do the actual work in
612 * ext4_ext_map_blocks()
613 */
614 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
615 return retval;
616
617 /*
618 * Here we clear m_flags because after allocating an new extent,
619 * it will be set again.
620 */
621 map->m_flags &= ~EXT4_MAP_FLAGS;
622
623 /*
624 * New blocks allocate and/or writing to unwritten extent
625 * will possibly result in updating i_data, so we take
626 * the write lock of i_data_sem, and call get_block()
627 * with create == 1 flag.
628 */
629 down_write(&EXT4_I(inode)->i_data_sem);
630
631 /*
632 * We need to check for EXT4 here because migrate
633 * could have changed the inode type in between
634 */
635 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
636 retval = ext4_ext_map_blocks(handle, inode, map, flags);
637 } else {
638 retval = ext4_ind_map_blocks(handle, inode, map, flags);
639
640 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
641 /*
642 * We allocated new blocks which will result in
643 * i_data's format changing. Force the migrate
644 * to fail by clearing migrate flags
645 */
646 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
647 }
648
649 /*
650 * Update reserved blocks/metadata blocks after successful
651 * block allocation which had been deferred till now. We don't
652 * support fallocate for non extent files. So we can update
653 * reserve space here.
654 */
655 if ((retval > 0) &&
656 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
657 ext4_da_update_reserve_space(inode, retval, 1);
658 }
659
660 if (retval > 0) {
661 unsigned int status;
662
663 if (unlikely(retval != map->m_len)) {
664 ext4_warning(inode->i_sb,
665 "ES len assertion failed for inode "
666 "%lu: retval %d != map->m_len %d",
667 inode->i_ino, retval, map->m_len);
668 WARN_ON(1);
669 }
670
671 /*
672 * We have to zeroout blocks before inserting them into extent
673 * status tree. Otherwise someone could look them up there and
674 * use them before they are really zeroed. We also have to
675 * unmap metadata before zeroing as otherwise writeback can
676 * overwrite zeros with stale data from block device.
677 */
678 if (flags & EXT4_GET_BLOCKS_ZERO &&
679 map->m_flags & EXT4_MAP_MAPPED &&
680 map->m_flags & EXT4_MAP_NEW) {
681 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
682 map->m_len);
683 ret = ext4_issue_zeroout(inode, map->m_lblk,
684 map->m_pblk, map->m_len);
685 if (ret) {
686 retval = ret;
687 goto out_sem;
688 }
689 }
690
691 /*
692 * If the extent has been zeroed out, we don't need to update
693 * extent status tree.
694 */
695 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
696 ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
697 if (ext4_es_is_written(&es))
698 goto out_sem;
699 }
700 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
701 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
702 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
703 !(status & EXTENT_STATUS_WRITTEN) &&
704 ext4_find_delalloc_range(inode, map->m_lblk,
705 map->m_lblk + map->m_len - 1))
706 status |= EXTENT_STATUS_DELAYED;
707 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
708 map->m_pblk, status);
709 if (ret < 0) {
710 retval = ret;
711 goto out_sem;
712 }
713 }
714
715out_sem:
716 up_write((&EXT4_I(inode)->i_data_sem));
717 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
718 ret = check_block_validity(inode, map);
719 if (ret != 0)
720 return ret;
721
722 /*
723 * Inodes with freshly allocated blocks where contents will be
724 * visible after transaction commit must be on transaction's
725 * ordered data list.
726 */
727 if (map->m_flags & EXT4_MAP_NEW &&
728 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
729 !(flags & EXT4_GET_BLOCKS_ZERO) &&
730 !ext4_is_quota_file(inode) &&
731 ext4_should_order_data(inode)) {
732 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
733 ret = ext4_jbd2_inode_add_wait(handle, inode);
734 else
735 ret = ext4_jbd2_inode_add_write(handle, inode);
736 if (ret)
737 return ret;
738 }
739 }
740 return retval;
741}
742
743/*
744 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
745 * we have to be careful as someone else may be manipulating b_state as well.
746 */
747static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
748{
749 unsigned long old_state;
750 unsigned long new_state;
751
752 flags &= EXT4_MAP_FLAGS;
753
754 /* Dummy buffer_head? Set non-atomically. */
755 if (!bh->b_page) {
756 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
757 return;
758 }
759 /*
760 * Someone else may be modifying b_state. Be careful! This is ugly but
761 * once we get rid of using bh as a container for mapping information
762 * to pass to / from get_block functions, this can go away.
763 */
764 do {
765 old_state = READ_ONCE(bh->b_state);
766 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
767 } while (unlikely(
768 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
769}
770
771static int _ext4_get_block(struct inode *inode, sector_t iblock,
772 struct buffer_head *bh, int flags)
773{
774 struct ext4_map_blocks map;
775 int ret = 0;
776
777 if (ext4_has_inline_data(inode))
778 return -ERANGE;
779
780 map.m_lblk = iblock;
781 map.m_len = bh->b_size >> inode->i_blkbits;
782
783 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
784 flags);
785 if (ret > 0) {
786 map_bh(bh, inode->i_sb, map.m_pblk);
787 ext4_update_bh_state(bh, map.m_flags);
788 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
789 ret = 0;
790 } else if (ret == 0) {
791 /* hole case, need to fill in bh->b_size */
792 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
793 }
794 return ret;
795}
796
797int ext4_get_block(struct inode *inode, sector_t iblock,
798 struct buffer_head *bh, int create)
799{
800 return _ext4_get_block(inode, iblock, bh,
801 create ? EXT4_GET_BLOCKS_CREATE : 0);
802}
803
804/*
805 * Get block function used when preparing for buffered write if we require
806 * creating an unwritten extent if blocks haven't been allocated. The extent
807 * will be converted to written after the IO is complete.
808 */
809int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
810 struct buffer_head *bh_result, int create)
811{
812 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
813 inode->i_ino, create);
814 return _ext4_get_block(inode, iblock, bh_result,
815 EXT4_GET_BLOCKS_IO_CREATE_EXT);
816}
817
818/* Maximum number of blocks we map for direct IO at once. */
819#define DIO_MAX_BLOCKS 4096
820
821/*
822 * Get blocks function for the cases that need to start a transaction -
823 * generally difference cases of direct IO and DAX IO. It also handles retries
824 * in case of ENOSPC.
825 */
826static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
827 struct buffer_head *bh_result, int flags)
828{
829 int dio_credits;
830 handle_t *handle;
831 int retries = 0;
832 int ret;
833
834 /* Trim mapping request to maximum we can map at once for DIO */
835 if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
836 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
837 dio_credits = ext4_chunk_trans_blocks(inode,
838 bh_result->b_size >> inode->i_blkbits);
839retry:
840 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
841 if (IS_ERR(handle))
842 return PTR_ERR(handle);
843
844 ret = _ext4_get_block(inode, iblock, bh_result, flags);
845 ext4_journal_stop(handle);
846
847 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
848 goto retry;
849 return ret;
850}
851
852/* Get block function for DIO reads and writes to inodes without extents */
853int ext4_dio_get_block(struct inode *inode, sector_t iblock,
854 struct buffer_head *bh, int create)
855{
856 /* We don't expect handle for direct IO */
857 WARN_ON_ONCE(ext4_journal_current_handle());
858
859 if (!create)
860 return _ext4_get_block(inode, iblock, bh, 0);
861 return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
862}
863
864/*
865 * Get block function for AIO DIO writes when we create unwritten extent if
866 * blocks are not allocated yet. The extent will be converted to written
867 * after IO is complete.
868 */
869static int ext4_dio_get_block_unwritten_async(struct inode *inode,
870 sector_t iblock, struct buffer_head *bh_result, int create)
871{
872 int ret;
873
874 /* We don't expect handle for direct IO */
875 WARN_ON_ONCE(ext4_journal_current_handle());
876
877 ret = ext4_get_block_trans(inode, iblock, bh_result,
878 EXT4_GET_BLOCKS_IO_CREATE_EXT);
879
880 /*
881 * When doing DIO using unwritten extents, we need io_end to convert
882 * unwritten extents to written on IO completion. We allocate io_end
883 * once we spot unwritten extent and store it in b_private. Generic
884 * DIO code keeps b_private set and furthermore passes the value to
885 * our completion callback in 'private' argument.
886 */
887 if (!ret && buffer_unwritten(bh_result)) {
888 if (!bh_result->b_private) {
889 ext4_io_end_t *io_end;
890
891 io_end = ext4_init_io_end(inode, GFP_KERNEL);
892 if (!io_end)
893 return -ENOMEM;
894 bh_result->b_private = io_end;
895 ext4_set_io_unwritten_flag(inode, io_end);
896 }
897 set_buffer_defer_completion(bh_result);
898 }
899
900 return ret;
901}
902
903/*
904 * Get block function for non-AIO DIO writes when we create unwritten extent if
905 * blocks are not allocated yet. The extent will be converted to written
906 * after IO is complete by ext4_direct_IO_write().
907 */
908static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
909 sector_t iblock, struct buffer_head *bh_result, int create)
910{
911 int ret;
912
913 /* We don't expect handle for direct IO */
914 WARN_ON_ONCE(ext4_journal_current_handle());
915
916 ret = ext4_get_block_trans(inode, iblock, bh_result,
917 EXT4_GET_BLOCKS_IO_CREATE_EXT);
918
919 /*
920 * Mark inode as having pending DIO writes to unwritten extents.
921 * ext4_direct_IO_write() checks this flag and converts extents to
922 * written.
923 */
924 if (!ret && buffer_unwritten(bh_result))
925 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
926
927 return ret;
928}
929
930static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
931 struct buffer_head *bh_result, int create)
932{
933 int ret;
934
935 ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
936 inode->i_ino, create);
937 /* We don't expect handle for direct IO */
938 WARN_ON_ONCE(ext4_journal_current_handle());
939
940 ret = _ext4_get_block(inode, iblock, bh_result, 0);
941 /*
942 * Blocks should have been preallocated! ext4_file_write_iter() checks
943 * that.
944 */
945 WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
946
947 return ret;
948}
949
950
951/*
952 * `handle' can be NULL if create is zero
953 */
954struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
955 ext4_lblk_t block, int map_flags)
956{
957 struct ext4_map_blocks map;
958 struct buffer_head *bh;
959 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
960 int err;
961
962 J_ASSERT(handle != NULL || create == 0);
963
964 map.m_lblk = block;
965 map.m_len = 1;
966 err = ext4_map_blocks(handle, inode, &map, map_flags);
967
968 if (err == 0)
969 return create ? ERR_PTR(-ENOSPC) : NULL;
970 if (err < 0)
971 return ERR_PTR(err);
972
973 bh = sb_getblk(inode->i_sb, map.m_pblk);
974 if (unlikely(!bh))
975 return ERR_PTR(-ENOMEM);
976 if (map.m_flags & EXT4_MAP_NEW) {
977 J_ASSERT(create != 0);
978 J_ASSERT(handle != NULL);
979
980 /*
981 * Now that we do not always journal data, we should
982 * keep in mind whether this should always journal the
983 * new buffer as metadata. For now, regular file
984 * writes use ext4_get_block instead, so it's not a
985 * problem.
986 */
987 lock_buffer(bh);
988 BUFFER_TRACE(bh, "call get_create_access");
989 err = ext4_journal_get_create_access(handle, bh);
990 if (unlikely(err)) {
991 unlock_buffer(bh);
992 goto errout;
993 }
994 if (!buffer_uptodate(bh)) {
995 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
996 set_buffer_uptodate(bh);
997 }
998 unlock_buffer(bh);
999 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1000 err = ext4_handle_dirty_metadata(handle, inode, bh);
1001 if (unlikely(err))
1002 goto errout;
1003 } else
1004 BUFFER_TRACE(bh, "not a new buffer");
1005 return bh;
1006errout:
1007 brelse(bh);
1008 return ERR_PTR(err);
1009}
1010
1011struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1012 ext4_lblk_t block, int map_flags)
1013{
1014 struct buffer_head *bh;
1015
1016 bh = ext4_getblk(handle, inode, block, map_flags);
1017 if (IS_ERR(bh))
1018 return bh;
1019 if (!bh || buffer_uptodate(bh))
1020 return bh;
1021 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1022 wait_on_buffer(bh);
1023 if (buffer_uptodate(bh))
1024 return bh;
1025 put_bh(bh);
1026 return ERR_PTR(-EIO);
1027}
1028
1029/* Read a contiguous batch of blocks. */
1030int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
1031 bool wait, struct buffer_head **bhs)
1032{
1033 int i, err;
1034
1035 for (i = 0; i < bh_count; i++) {
1036 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
1037 if (IS_ERR(bhs[i])) {
1038 err = PTR_ERR(bhs[i]);
1039 bh_count = i;
1040 goto out_brelse;
1041 }
1042 }
1043
1044 for (i = 0; i < bh_count; i++)
1045 /* Note that NULL bhs[i] is valid because of holes. */
1046 if (bhs[i] && !buffer_uptodate(bhs[i]))
1047 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
1048 &bhs[i]);
1049
1050 if (!wait)
1051 return 0;
1052
1053 for (i = 0; i < bh_count; i++)
1054 if (bhs[i])
1055 wait_on_buffer(bhs[i]);
1056
1057 for (i = 0; i < bh_count; i++) {
1058 if (bhs[i] && !buffer_uptodate(bhs[i])) {
1059 err = -EIO;
1060 goto out_brelse;
1061 }
1062 }
1063 return 0;
1064
1065out_brelse:
1066 for (i = 0; i < bh_count; i++) {
1067 brelse(bhs[i]);
1068 bhs[i] = NULL;
1069 }
1070 return err;
1071}
1072
1073int ext4_walk_page_buffers(handle_t *handle,
1074 struct buffer_head *head,
1075 unsigned from,
1076 unsigned to,
1077 int *partial,
1078 int (*fn)(handle_t *handle,
1079 struct buffer_head *bh))
1080{
1081 struct buffer_head *bh;
1082 unsigned block_start, block_end;
1083 unsigned blocksize = head->b_size;
1084 int err, ret = 0;
1085 struct buffer_head *next;
1086
1087 for (bh = head, block_start = 0;
1088 ret == 0 && (bh != head || !block_start);
1089 block_start = block_end, bh = next) {
1090 next = bh->b_this_page;
1091 block_end = block_start + blocksize;
1092 if (block_end <= from || block_start >= to) {
1093 if (partial && !buffer_uptodate(bh))
1094 *partial = 1;
1095 continue;
1096 }
1097 err = (*fn)(handle, bh);
1098 if (!ret)
1099 ret = err;
1100 }
1101 return ret;
1102}
1103
1104/*
1105 * To preserve ordering, it is essential that the hole instantiation and
1106 * the data write be encapsulated in a single transaction. We cannot
1107 * close off a transaction and start a new one between the ext4_get_block()
1108 * and the commit_write(). So doing the jbd2_journal_start at the start of
1109 * prepare_write() is the right place.
1110 *
1111 * Also, this function can nest inside ext4_writepage(). In that case, we
1112 * *know* that ext4_writepage() has generated enough buffer credits to do the
1113 * whole page. So we won't block on the journal in that case, which is good,
1114 * because the caller may be PF_MEMALLOC.
1115 *
1116 * By accident, ext4 can be reentered when a transaction is open via
1117 * quota file writes. If we were to commit the transaction while thus
1118 * reentered, there can be a deadlock - we would be holding a quota
1119 * lock, and the commit would never complete if another thread had a
1120 * transaction open and was blocking on the quota lock - a ranking
1121 * violation.
1122 *
1123 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1124 * will _not_ run commit under these circumstances because handle->h_ref
1125 * is elevated. We'll still have enough credits for the tiny quotafile
1126 * write.
1127 */
1128int do_journal_get_write_access(handle_t *handle,
1129 struct buffer_head *bh)
1130{
1131 int dirty = buffer_dirty(bh);
1132 int ret;
1133
1134 if (!buffer_mapped(bh) || buffer_freed(bh))
1135 return 0;
1136 /*
1137 * __block_write_begin() could have dirtied some buffers. Clean
1138 * the dirty bit as jbd2_journal_get_write_access() could complain
1139 * otherwise about fs integrity issues. Setting of the dirty bit
1140 * by __block_write_begin() isn't a real problem here as we clear
1141 * the bit before releasing a page lock and thus writeback cannot
1142 * ever write the buffer.
1143 */
1144 if (dirty)
1145 clear_buffer_dirty(bh);
1146 BUFFER_TRACE(bh, "get write access");
1147 ret = ext4_journal_get_write_access(handle, bh);
1148 if (!ret && dirty)
1149 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1150 return ret;
1151}
1152
1153#ifdef CONFIG_EXT4_FS_ENCRYPTION
1154static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1155 get_block_t *get_block)
1156{
1157 unsigned from = pos & (PAGE_SIZE - 1);
1158 unsigned to = from + len;
1159 struct inode *inode = page->mapping->host;
1160 unsigned block_start, block_end;
1161 sector_t block;
1162 int err = 0;
1163 unsigned blocksize = inode->i_sb->s_blocksize;
1164 unsigned bbits;
1165 struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1166 bool decrypt = false;
1167
1168 BUG_ON(!PageLocked(page));
1169 BUG_ON(from > PAGE_SIZE);
1170 BUG_ON(to > PAGE_SIZE);
1171 BUG_ON(from > to);
1172
1173 if (!page_has_buffers(page))
1174 create_empty_buffers(page, blocksize, 0);
1175 head = page_buffers(page);
1176 bbits = ilog2(blocksize);
1177 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1178
1179 for (bh = head, block_start = 0; bh != head || !block_start;
1180 block++, block_start = block_end, bh = bh->b_this_page) {
1181 block_end = block_start + blocksize;
1182 if (block_end <= from || block_start >= to) {
1183 if (PageUptodate(page)) {
1184 if (!buffer_uptodate(bh))
1185 set_buffer_uptodate(bh);
1186 }
1187 continue;
1188 }
1189 if (buffer_new(bh))
1190 clear_buffer_new(bh);
1191 if (!buffer_mapped(bh)) {
1192 WARN_ON(bh->b_size != blocksize);
1193 err = get_block(inode, block, bh, 1);
1194 if (err)
1195 break;
1196 if (buffer_new(bh)) {
1197 clean_bdev_bh_alias(bh);
1198 if (PageUptodate(page)) {
1199 clear_buffer_new(bh);
1200 set_buffer_uptodate(bh);
1201 mark_buffer_dirty(bh);
1202 continue;
1203 }
1204 if (block_end > to || block_start < from)
1205 zero_user_segments(page, to, block_end,
1206 block_start, from);
1207 continue;
1208 }
1209 }
1210 if (PageUptodate(page)) {
1211 if (!buffer_uptodate(bh))
1212 set_buffer_uptodate(bh);
1213 continue;
1214 }
1215 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1216 !buffer_unwritten(bh) &&
1217 (block_start < from || block_end > to)) {
1218 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1219 *wait_bh++ = bh;
1220 decrypt = ext4_encrypted_inode(inode) &&
1221 S_ISREG(inode->i_mode);
1222 }
1223 }
1224 /*
1225 * If we issued read requests, let them complete.
1226 */
1227 while (wait_bh > wait) {
1228 wait_on_buffer(*--wait_bh);
1229 if (!buffer_uptodate(*wait_bh))
1230 err = -EIO;
1231 }
1232 if (unlikely(err))
1233 page_zero_new_buffers(page, from, to);
1234 else if (decrypt)
1235 err = fscrypt_decrypt_page(page->mapping->host, page,
1236 PAGE_SIZE, 0, page->index);
1237 return err;
1238}
1239#endif
1240
1241static int ext4_write_begin(struct file *file, struct address_space *mapping,
1242 loff_t pos, unsigned len, unsigned flags,
1243 struct page **pagep, void **fsdata)
1244{
1245 struct inode *inode = mapping->host;
1246 int ret, needed_blocks;
1247 handle_t *handle;
1248 int retries = 0;
1249 struct page *page;
1250 pgoff_t index;
1251 unsigned from, to;
1252
1253 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1254 return -EIO;
1255
1256 trace_ext4_write_begin(inode, pos, len, flags);
1257 /*
1258 * Reserve one block more for addition to orphan list in case
1259 * we allocate blocks but write fails for some reason
1260 */
1261 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1262 index = pos >> PAGE_SHIFT;
1263 from = pos & (PAGE_SIZE - 1);
1264 to = from + len;
1265
1266 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1267 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1268 flags, pagep);
1269 if (ret < 0)
1270 return ret;
1271 if (ret == 1)
1272 return 0;
1273 }
1274
1275 /*
1276 * grab_cache_page_write_begin() can take a long time if the
1277 * system is thrashing due to memory pressure, or if the page
1278 * is being written back. So grab it first before we start
1279 * the transaction handle. This also allows us to allocate
1280 * the page (if needed) without using GFP_NOFS.
1281 */
1282retry_grab:
1283 page = grab_cache_page_write_begin(mapping, index, flags);
1284 if (!page)
1285 return -ENOMEM;
1286 unlock_page(page);
1287
1288retry_journal:
1289 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1290 if (IS_ERR(handle)) {
1291 put_page(page);
1292 return PTR_ERR(handle);
1293 }
1294
1295 lock_page(page);
1296 if (page->mapping != mapping) {
1297 /* The page got truncated from under us */
1298 unlock_page(page);
1299 put_page(page);
1300 ext4_journal_stop(handle);
1301 goto retry_grab;
1302 }
1303 /* In case writeback began while the page was unlocked */
1304 wait_for_stable_page(page);
1305
1306#ifdef CONFIG_EXT4_FS_ENCRYPTION
1307 if (ext4_should_dioread_nolock(inode))
1308 ret = ext4_block_write_begin(page, pos, len,
1309 ext4_get_block_unwritten);
1310 else
1311 ret = ext4_block_write_begin(page, pos, len,
1312 ext4_get_block);
1313#else
1314 if (ext4_should_dioread_nolock(inode))
1315 ret = __block_write_begin(page, pos, len,
1316 ext4_get_block_unwritten);
1317 else
1318 ret = __block_write_begin(page, pos, len, ext4_get_block);
1319#endif
1320 if (!ret && ext4_should_journal_data(inode)) {
1321 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1322 from, to, NULL,
1323 do_journal_get_write_access);
1324 }
1325
1326 if (ret) {
1327 unlock_page(page);
1328 /*
1329 * __block_write_begin may have instantiated a few blocks
1330 * outside i_size. Trim these off again. Don't need
1331 * i_size_read because we hold i_mutex.
1332 *
1333 * Add inode to orphan list in case we crash before
1334 * truncate finishes
1335 */
1336 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1337 ext4_orphan_add(handle, inode);
1338
1339 ext4_journal_stop(handle);
1340 if (pos + len > inode->i_size) {
1341 ext4_truncate_failed_write(inode);
1342 /*
1343 * If truncate failed early the inode might
1344 * still be on the orphan list; we need to
1345 * make sure the inode is removed from the
1346 * orphan list in that case.
1347 */
1348 if (inode->i_nlink)
1349 ext4_orphan_del(NULL, inode);
1350 }
1351
1352 if (ret == -ENOSPC &&
1353 ext4_should_retry_alloc(inode->i_sb, &retries))
1354 goto retry_journal;
1355 put_page(page);
1356 return ret;
1357 }
1358 *pagep = page;
1359 return ret;
1360}
1361
1362/* For write_end() in data=journal mode */
1363static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1364{
1365 int ret;
1366 if (!buffer_mapped(bh) || buffer_freed(bh))
1367 return 0;
1368 set_buffer_uptodate(bh);
1369 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1370 clear_buffer_meta(bh);
1371 clear_buffer_prio(bh);
1372 return ret;
1373}
1374
1375/*
1376 * We need to pick up the new inode size which generic_commit_write gave us
1377 * `file' can be NULL - eg, when called from page_symlink().
1378 *
1379 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1380 * buffers are managed internally.
1381 */
1382static int ext4_write_end(struct file *file,
1383 struct address_space *mapping,
1384 loff_t pos, unsigned len, unsigned copied,
1385 struct page *page, void *fsdata)
1386{
1387 handle_t *handle = ext4_journal_current_handle();
1388 struct inode *inode = mapping->host;
1389 loff_t old_size = inode->i_size;
1390 int ret = 0, ret2;
1391 int i_size_changed = 0;
1392
1393 trace_ext4_write_end(inode, pos, len, copied);
1394 if (ext4_has_inline_data(inode)) {
1395 ret = ext4_write_inline_data_end(inode, pos, len,
1396 copied, page);
1397 if (ret < 0) {
1398 unlock_page(page);
1399 put_page(page);
1400 goto errout;
1401 }
1402 copied = ret;
1403 } else
1404 copied = block_write_end(file, mapping, pos,
1405 len, copied, page, fsdata);
1406 /*
1407 * it's important to update i_size while still holding page lock:
1408 * page writeout could otherwise come in and zero beyond i_size.
1409 */
1410 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1411 unlock_page(page);
1412 put_page(page);
1413
1414 if (old_size < pos)
1415 pagecache_isize_extended(inode, old_size, pos);
1416 /*
1417 * Don't mark the inode dirty under page lock. First, it unnecessarily
1418 * makes the holding time of page lock longer. Second, it forces lock
1419 * ordering of page lock and transaction start for journaling
1420 * filesystems.
1421 */
1422 if (i_size_changed)
1423 ext4_mark_inode_dirty(handle, inode);
1424
1425 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1426 /* if we have allocated more blocks and copied
1427 * less. We will have blocks allocated outside
1428 * inode->i_size. So truncate them
1429 */
1430 ext4_orphan_add(handle, inode);
1431errout:
1432 ret2 = ext4_journal_stop(handle);
1433 if (!ret)
1434 ret = ret2;
1435
1436 if (pos + len > inode->i_size) {
1437 ext4_truncate_failed_write(inode);
1438 /*
1439 * If truncate failed early the inode might still be
1440 * on the orphan list; we need to make sure the inode
1441 * is removed from the orphan list in that case.
1442 */
1443 if (inode->i_nlink)
1444 ext4_orphan_del(NULL, inode);
1445 }
1446
1447 return ret ? ret : copied;
1448}
1449
1450/*
1451 * This is a private version of page_zero_new_buffers() which doesn't
1452 * set the buffer to be dirty, since in data=journalled mode we need
1453 * to call ext4_handle_dirty_metadata() instead.
1454 */
1455static void ext4_journalled_zero_new_buffers(handle_t *handle,
1456 struct page *page,
1457 unsigned from, unsigned to)
1458{
1459 unsigned int block_start = 0, block_end;
1460 struct buffer_head *head, *bh;
1461
1462 bh = head = page_buffers(page);
1463 do {
1464 block_end = block_start + bh->b_size;
1465 if (buffer_new(bh)) {
1466 if (block_end > from && block_start < to) {
1467 if (!PageUptodate(page)) {
1468 unsigned start, size;
1469
1470 start = max(from, block_start);
1471 size = min(to, block_end) - start;
1472
1473 zero_user(page, start, size);
1474 write_end_fn(handle, bh);
1475 }
1476 clear_buffer_new(bh);
1477 }
1478 }
1479 block_start = block_end;
1480 bh = bh->b_this_page;
1481 } while (bh != head);
1482}
1483
1484static int ext4_journalled_write_end(struct file *file,
1485 struct address_space *mapping,
1486 loff_t pos, unsigned len, unsigned copied,
1487 struct page *page, void *fsdata)
1488{
1489 handle_t *handle = ext4_journal_current_handle();
1490 struct inode *inode = mapping->host;
1491 loff_t old_size = inode->i_size;
1492 int ret = 0, ret2;
1493 int partial = 0;
1494 unsigned from, to;
1495 int size_changed = 0;
1496
1497 trace_ext4_journalled_write_end(inode, pos, len, copied);
1498 from = pos & (PAGE_SIZE - 1);
1499 to = from + len;
1500
1501 BUG_ON(!ext4_handle_valid(handle));
1502
1503 if (ext4_has_inline_data(inode)) {
1504 ret = ext4_write_inline_data_end(inode, pos, len,
1505 copied, page);
1506 if (ret < 0) {
1507 unlock_page(page);
1508 put_page(page);
1509 goto errout;
1510 }
1511 copied = ret;
1512 } else if (unlikely(copied < len) && !PageUptodate(page)) {
1513 copied = 0;
1514 ext4_journalled_zero_new_buffers(handle, page, from, to);
1515 } else {
1516 if (unlikely(copied < len))
1517 ext4_journalled_zero_new_buffers(handle, page,
1518 from + copied, to);
1519 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1520 from + copied, &partial,
1521 write_end_fn);
1522 if (!partial)
1523 SetPageUptodate(page);
1524 }
1525 size_changed = ext4_update_inode_size(inode, pos + copied);
1526 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1527 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1528 unlock_page(page);
1529 put_page(page);
1530
1531 if (old_size < pos)
1532 pagecache_isize_extended(inode, old_size, pos);
1533
1534 if (size_changed) {
1535 ret2 = ext4_mark_inode_dirty(handle, inode);
1536 if (!ret)
1537 ret = ret2;
1538 }
1539
1540 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1541 /* if we have allocated more blocks and copied
1542 * less. We will have blocks allocated outside
1543 * inode->i_size. So truncate them
1544 */
1545 ext4_orphan_add(handle, inode);
1546
1547errout:
1548 ret2 = ext4_journal_stop(handle);
1549 if (!ret)
1550 ret = ret2;
1551 if (pos + len > inode->i_size) {
1552 ext4_truncate_failed_write(inode);
1553 /*
1554 * If truncate failed early the inode might still be
1555 * on the orphan list; we need to make sure the inode
1556 * is removed from the orphan list in that case.
1557 */
1558 if (inode->i_nlink)
1559 ext4_orphan_del(NULL, inode);
1560 }
1561
1562 return ret ? ret : copied;
1563}
1564
1565/*
1566 * Reserve space for a single cluster
1567 */
1568static int ext4_da_reserve_space(struct inode *inode)
1569{
1570 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1571 struct ext4_inode_info *ei = EXT4_I(inode);
1572 int ret;
1573
1574 /*
1575 * We will charge metadata quota at writeout time; this saves
1576 * us from metadata over-estimation, though we may go over by
1577 * a small amount in the end. Here we just reserve for data.
1578 */
1579 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1580 if (ret)
1581 return ret;
1582
1583 spin_lock(&ei->i_block_reservation_lock);
1584 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1585 spin_unlock(&ei->i_block_reservation_lock);
1586 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1587 return -ENOSPC;
1588 }
1589 ei->i_reserved_data_blocks++;
1590 trace_ext4_da_reserve_space(inode);
1591 spin_unlock(&ei->i_block_reservation_lock);
1592
1593 return 0; /* success */
1594}
1595
1596static void ext4_da_release_space(struct inode *inode, int to_free)
1597{
1598 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1599 struct ext4_inode_info *ei = EXT4_I(inode);
1600
1601 if (!to_free)
1602 return; /* Nothing to release, exit */
1603
1604 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1605
1606 trace_ext4_da_release_space(inode, to_free);
1607 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1608 /*
1609 * if there aren't enough reserved blocks, then the
1610 * counter is messed up somewhere. Since this
1611 * function is called from invalidate page, it's
1612 * harmless to return without any action.
1613 */
1614 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1615 "ino %lu, to_free %d with only %d reserved "
1616 "data blocks", inode->i_ino, to_free,
1617 ei->i_reserved_data_blocks);
1618 WARN_ON(1);
1619 to_free = ei->i_reserved_data_blocks;
1620 }
1621 ei->i_reserved_data_blocks -= to_free;
1622
1623 /* update fs dirty data blocks counter */
1624 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1625
1626 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1627
1628 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1629}
1630
1631static void ext4_da_page_release_reservation(struct page *page,
1632 unsigned int offset,
1633 unsigned int length)
1634{
1635 int to_release = 0, contiguous_blks = 0;
1636 struct buffer_head *head, *bh;
1637 unsigned int curr_off = 0;
1638 struct inode *inode = page->mapping->host;
1639 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1640 unsigned int stop = offset + length;
1641 int num_clusters;
1642 ext4_fsblk_t lblk;
1643
1644 BUG_ON(stop > PAGE_SIZE || stop < length);
1645
1646 head = page_buffers(page);
1647 bh = head;
1648 do {
1649 unsigned int next_off = curr_off + bh->b_size;
1650
1651 if (next_off > stop)
1652 break;
1653
1654 if ((offset <= curr_off) && (buffer_delay(bh))) {
1655 to_release++;
1656 contiguous_blks++;
1657 clear_buffer_delay(bh);
1658 } else if (contiguous_blks) {
1659 lblk = page->index <<
1660 (PAGE_SHIFT - inode->i_blkbits);
1661 lblk += (curr_off >> inode->i_blkbits) -
1662 contiguous_blks;
1663 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1664 contiguous_blks = 0;
1665 }
1666 curr_off = next_off;
1667 } while ((bh = bh->b_this_page) != head);
1668
1669 if (contiguous_blks) {
1670 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1671 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1672 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1673 }
1674
1675 /* If we have released all the blocks belonging to a cluster, then we
1676 * need to release the reserved space for that cluster. */
1677 num_clusters = EXT4_NUM_B2C(sbi, to_release);
1678 while (num_clusters > 0) {
1679 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1680 ((num_clusters - 1) << sbi->s_cluster_bits);
1681 if (sbi->s_cluster_ratio == 1 ||
1682 !ext4_find_delalloc_cluster(inode, lblk))
1683 ext4_da_release_space(inode, 1);
1684
1685 num_clusters--;
1686 }
1687}
1688
1689/*
1690 * Delayed allocation stuff
1691 */
1692
1693struct mpage_da_data {
1694 struct inode *inode;
1695 struct writeback_control *wbc;
1696
1697 pgoff_t first_page; /* The first page to write */
1698 pgoff_t next_page; /* Current page to examine */
1699 pgoff_t last_page; /* Last page to examine */
1700 /*
1701 * Extent to map - this can be after first_page because that can be
1702 * fully mapped. We somewhat abuse m_flags to store whether the extent
1703 * is delalloc or unwritten.
1704 */
1705 struct ext4_map_blocks map;
1706 struct ext4_io_submit io_submit; /* IO submission data */
1707 unsigned int do_map:1;
1708};
1709
1710static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1711 bool invalidate)
1712{
1713 int nr_pages, i;
1714 pgoff_t index, end;
1715 struct pagevec pvec;
1716 struct inode *inode = mpd->inode;
1717 struct address_space *mapping = inode->i_mapping;
1718
1719 /* This is necessary when next_page == 0. */
1720 if (mpd->first_page >= mpd->next_page)
1721 return;
1722
1723 index = mpd->first_page;
1724 end = mpd->next_page - 1;
1725 if (invalidate) {
1726 ext4_lblk_t start, last;
1727 start = index << (PAGE_SHIFT - inode->i_blkbits);
1728 last = end << (PAGE_SHIFT - inode->i_blkbits);
1729 ext4_es_remove_extent(inode, start, last - start + 1);
1730 }
1731
1732 pagevec_init(&pvec);
1733 while (index <= end) {
1734 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1735 if (nr_pages == 0)
1736 break;
1737 for (i = 0; i < nr_pages; i++) {
1738 struct page *page = pvec.pages[i];
1739
1740 BUG_ON(!PageLocked(page));
1741 BUG_ON(PageWriteback(page));
1742 if (invalidate) {
1743 if (page_mapped(page))
1744 clear_page_dirty_for_io(page);
1745 block_invalidatepage(page, 0, PAGE_SIZE);
1746 ClearPageUptodate(page);
1747 }
1748 unlock_page(page);
1749 }
1750 pagevec_release(&pvec);
1751 }
1752}
1753
1754static void ext4_print_free_blocks(struct inode *inode)
1755{
1756 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1757 struct super_block *sb = inode->i_sb;
1758 struct ext4_inode_info *ei = EXT4_I(inode);
1759
1760 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1761 EXT4_C2B(EXT4_SB(inode->i_sb),
1762 ext4_count_free_clusters(sb)));
1763 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1764 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1765 (long long) EXT4_C2B(EXT4_SB(sb),
1766 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1767 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1768 (long long) EXT4_C2B(EXT4_SB(sb),
1769 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1770 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1771 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1772 ei->i_reserved_data_blocks);
1773 return;
1774}
1775
1776static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1777{
1778 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1779}
1780
1781/*
1782 * This function is grabs code from the very beginning of
1783 * ext4_map_blocks, but assumes that the caller is from delayed write
1784 * time. This function looks up the requested blocks and sets the
1785 * buffer delay bit under the protection of i_data_sem.
1786 */
1787static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1788 struct ext4_map_blocks *map,
1789 struct buffer_head *bh)
1790{
1791 struct extent_status es;
1792 int retval;
1793 sector_t invalid_block = ~((sector_t) 0xffff);
1794#ifdef ES_AGGRESSIVE_TEST
1795 struct ext4_map_blocks orig_map;
1796
1797 memcpy(&orig_map, map, sizeof(*map));
1798#endif
1799
1800 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1801 invalid_block = ~0;
1802
1803 map->m_flags = 0;
1804 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1805 "logical block %lu\n", inode->i_ino, map->m_len,
1806 (unsigned long) map->m_lblk);
1807
1808 /* Lookup extent status tree firstly */
1809 if (ext4_es_lookup_extent(inode, iblock, &es)) {
1810 if (ext4_es_is_hole(&es)) {
1811 retval = 0;
1812 down_read(&EXT4_I(inode)->i_data_sem);
1813 goto add_delayed;
1814 }
1815
1816 /*
1817 * Delayed extent could be allocated by fallocate.
1818 * So we need to check it.
1819 */
1820 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1821 map_bh(bh, inode->i_sb, invalid_block);
1822 set_buffer_new(bh);
1823 set_buffer_delay(bh);
1824 return 0;
1825 }
1826
1827 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1828 retval = es.es_len - (iblock - es.es_lblk);
1829 if (retval > map->m_len)
1830 retval = map->m_len;
1831 map->m_len = retval;
1832 if (ext4_es_is_written(&es))
1833 map->m_flags |= EXT4_MAP_MAPPED;
1834 else if (ext4_es_is_unwritten(&es))
1835 map->m_flags |= EXT4_MAP_UNWRITTEN;
1836 else
1837 BUG_ON(1);
1838
1839#ifdef ES_AGGRESSIVE_TEST
1840 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1841#endif
1842 return retval;
1843 }
1844
1845 /*
1846 * Try to see if we can get the block without requesting a new
1847 * file system block.
1848 */
1849 down_read(&EXT4_I(inode)->i_data_sem);
1850 if (ext4_has_inline_data(inode))
1851 retval = 0;
1852 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1853 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1854 else
1855 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1856
1857add_delayed:
1858 if (retval == 0) {
1859 int ret;
1860 /*
1861 * XXX: __block_prepare_write() unmaps passed block,
1862 * is it OK?
1863 */
1864 /*
1865 * If the block was allocated from previously allocated cluster,
1866 * then we don't need to reserve it again. However we still need
1867 * to reserve metadata for every block we're going to write.
1868 */
1869 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1870 !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1871 ret = ext4_da_reserve_space(inode);
1872 if (ret) {
1873 /* not enough space to reserve */
1874 retval = ret;
1875 goto out_unlock;
1876 }
1877 }
1878
1879 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1880 ~0, EXTENT_STATUS_DELAYED);
1881 if (ret) {
1882 retval = ret;
1883 goto out_unlock;
1884 }
1885
1886 map_bh(bh, inode->i_sb, invalid_block);
1887 set_buffer_new(bh);
1888 set_buffer_delay(bh);
1889 } else if (retval > 0) {
1890 int ret;
1891 unsigned int status;
1892
1893 if (unlikely(retval != map->m_len)) {
1894 ext4_warning(inode->i_sb,
1895 "ES len assertion failed for inode "
1896 "%lu: retval %d != map->m_len %d",
1897 inode->i_ino, retval, map->m_len);
1898 WARN_ON(1);
1899 }
1900
1901 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1902 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1903 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1904 map->m_pblk, status);
1905 if (ret != 0)
1906 retval = ret;
1907 }
1908
1909out_unlock:
1910 up_read((&EXT4_I(inode)->i_data_sem));
1911
1912 return retval;
1913}
1914
1915/*
1916 * This is a special get_block_t callback which is used by
1917 * ext4_da_write_begin(). It will either return mapped block or
1918 * reserve space for a single block.
1919 *
1920 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1921 * We also have b_blocknr = -1 and b_bdev initialized properly
1922 *
1923 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1924 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1925 * initialized properly.
1926 */
1927int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1928 struct buffer_head *bh, int create)
1929{
1930 struct ext4_map_blocks map;
1931 int ret = 0;
1932
1933 BUG_ON(create == 0);
1934 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1935
1936 map.m_lblk = iblock;
1937 map.m_len = 1;
1938
1939 /*
1940 * first, we need to know whether the block is allocated already
1941 * preallocated blocks are unmapped but should treated
1942 * the same as allocated blocks.
1943 */
1944 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1945 if (ret <= 0)
1946 return ret;
1947
1948 map_bh(bh, inode->i_sb, map.m_pblk);
1949 ext4_update_bh_state(bh, map.m_flags);
1950
1951 if (buffer_unwritten(bh)) {
1952 /* A delayed write to unwritten bh should be marked
1953 * new and mapped. Mapped ensures that we don't do
1954 * get_block multiple times when we write to the same
1955 * offset and new ensures that we do proper zero out
1956 * for partial write.
1957 */
1958 set_buffer_new(bh);
1959 set_buffer_mapped(bh);
1960 }
1961 return 0;
1962}
1963
1964static int bget_one(handle_t *handle, struct buffer_head *bh)
1965{
1966 get_bh(bh);
1967 return 0;
1968}
1969
1970static int bput_one(handle_t *handle, struct buffer_head *bh)
1971{
1972 put_bh(bh);
1973 return 0;
1974}
1975
1976static int __ext4_journalled_writepage(struct page *page,
1977 unsigned int len)
1978{
1979 struct address_space *mapping = page->mapping;
1980 struct inode *inode = mapping->host;
1981 struct buffer_head *page_bufs = NULL;
1982 handle_t *handle = NULL;
1983 int ret = 0, err = 0;
1984 int inline_data = ext4_has_inline_data(inode);
1985 struct buffer_head *inode_bh = NULL;
1986
1987 ClearPageChecked(page);
1988
1989 if (inline_data) {
1990 BUG_ON(page->index != 0);
1991 BUG_ON(len > ext4_get_max_inline_size(inode));
1992 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1993 if (inode_bh == NULL)
1994 goto out;
1995 } else {
1996 page_bufs = page_buffers(page);
1997 if (!page_bufs) {
1998 BUG();
1999 goto out;
2000 }
2001 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2002 NULL, bget_one);
2003 }
2004 /*
2005 * We need to release the page lock before we start the
2006 * journal, so grab a reference so the page won't disappear
2007 * out from under us.
2008 */
2009 get_page(page);
2010 unlock_page(page);
2011
2012 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2013 ext4_writepage_trans_blocks(inode));
2014 if (IS_ERR(handle)) {
2015 ret = PTR_ERR(handle);
2016 put_page(page);
2017 goto out_no_pagelock;
2018 }
2019 BUG_ON(!ext4_handle_valid(handle));
2020
2021 lock_page(page);
2022 put_page(page);
2023 if (page->mapping != mapping) {
2024 /* The page got truncated from under us */
2025 ext4_journal_stop(handle);
2026 ret = 0;
2027 goto out;
2028 }
2029
2030 if (inline_data) {
2031 BUFFER_TRACE(inode_bh, "get write access");
2032 ret = ext4_journal_get_write_access(handle, inode_bh);
2033
2034 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
2035
2036 } else {
2037 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2038 do_journal_get_write_access);
2039
2040 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2041 write_end_fn);
2042 }
2043 if (ret == 0)
2044 ret = err;
2045 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2046 err = ext4_journal_stop(handle);
2047 if (!ret)
2048 ret = err;
2049
2050 if (!ext4_has_inline_data(inode))
2051 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2052 NULL, bput_one);
2053 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2054out:
2055 unlock_page(page);
2056out_no_pagelock:
2057 brelse(inode_bh);
2058 return ret;
2059}
2060
2061/*
2062 * Note that we don't need to start a transaction unless we're journaling data
2063 * because we should have holes filled from ext4_page_mkwrite(). We even don't
2064 * need to file the inode to the transaction's list in ordered mode because if
2065 * we are writing back data added by write(), the inode is already there and if
2066 * we are writing back data modified via mmap(), no one guarantees in which
2067 * transaction the data will hit the disk. In case we are journaling data, we
2068 * cannot start transaction directly because transaction start ranks above page
2069 * lock so we have to do some magic.
2070 *
2071 * This function can get called via...
2072 * - ext4_writepages after taking page lock (have journal handle)
2073 * - journal_submit_inode_data_buffers (no journal handle)
2074 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2075 * - grab_page_cache when doing write_begin (have journal handle)
2076 *
2077 * We don't do any block allocation in this function. If we have page with
2078 * multiple blocks we need to write those buffer_heads that are mapped. This
2079 * is important for mmaped based write. So if we do with blocksize 1K
2080 * truncate(f, 1024);
2081 * a = mmap(f, 0, 4096);
2082 * a[0] = 'a';
2083 * truncate(f, 4096);
2084 * we have in the page first buffer_head mapped via page_mkwrite call back
2085 * but other buffer_heads would be unmapped but dirty (dirty done via the
2086 * do_wp_page). So writepage should write the first block. If we modify
2087 * the mmap area beyond 1024 we will again get a page_fault and the
2088 * page_mkwrite callback will do the block allocation and mark the
2089 * buffer_heads mapped.
2090 *
2091 * We redirty the page if we have any buffer_heads that is either delay or
2092 * unwritten in the page.
2093 *
2094 * We can get recursively called as show below.
2095 *
2096 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2097 * ext4_writepage()
2098 *
2099 * But since we don't do any block allocation we should not deadlock.
2100 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2101 */
2102static int ext4_writepage(struct page *page,
2103 struct writeback_control *wbc)
2104{
2105 int ret = 0;
2106 loff_t size;
2107 unsigned int len;
2108 struct buffer_head *page_bufs = NULL;
2109 struct inode *inode = page->mapping->host;
2110 struct ext4_io_submit io_submit;
2111 bool keep_towrite = false;
2112
2113 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2114 ext4_invalidatepage(page, 0, PAGE_SIZE);
2115 unlock_page(page);
2116 return -EIO;
2117 }
2118
2119 trace_ext4_writepage(page);
2120 size = i_size_read(inode);
2121 if (page->index == size >> PAGE_SHIFT)
2122 len = size & ~PAGE_MASK;
2123 else
2124 len = PAGE_SIZE;
2125
2126 page_bufs = page_buffers(page);
2127 /*
2128 * We cannot do block allocation or other extent handling in this
2129 * function. If there are buffers needing that, we have to redirty
2130 * the page. But we may reach here when we do a journal commit via
2131 * journal_submit_inode_data_buffers() and in that case we must write
2132 * allocated buffers to achieve data=ordered mode guarantees.
2133 *
2134 * Also, if there is only one buffer per page (the fs block
2135 * size == the page size), if one buffer needs block
2136 * allocation or needs to modify the extent tree to clear the
2137 * unwritten flag, we know that the page can't be written at
2138 * all, so we might as well refuse the write immediately.
2139 * Unfortunately if the block size != page size, we can't as
2140 * easily detect this case using ext4_walk_page_buffers(), but
2141 * for the extremely common case, this is an optimization that
2142 * skips a useless round trip through ext4_bio_write_page().
2143 */
2144 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2145 ext4_bh_delay_or_unwritten)) {
2146 redirty_page_for_writepage(wbc, page);
2147 if ((current->flags & PF_MEMALLOC) ||
2148 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2149 /*
2150 * For memory cleaning there's no point in writing only
2151 * some buffers. So just bail out. Warn if we came here
2152 * from direct reclaim.
2153 */
2154 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2155 == PF_MEMALLOC);
2156 unlock_page(page);
2157 return 0;
2158 }
2159 keep_towrite = true;
2160 }
2161
2162 if (PageChecked(page) && ext4_should_journal_data(inode))
2163 /*
2164 * It's mmapped pagecache. Add buffers and journal it. There
2165 * doesn't seem much point in redirtying the page here.
2166 */
2167 return __ext4_journalled_writepage(page, len);
2168
2169 ext4_io_submit_init(&io_submit, wbc);
2170 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2171 if (!io_submit.io_end) {
2172 redirty_page_for_writepage(wbc, page);
2173 unlock_page(page);
2174 return -ENOMEM;
2175 }
2176 ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2177 ext4_io_submit(&io_submit);
2178 /* Drop io_end reference we got from init */
2179 ext4_put_io_end_defer(io_submit.io_end);
2180 return ret;
2181}
2182
2183static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2184{
2185 int len;
2186 loff_t size;
2187 int err;
2188
2189 BUG_ON(page->index != mpd->first_page);
2190 clear_page_dirty_for_io(page);
2191 /*
2192 * We have to be very careful here! Nothing protects writeback path
2193 * against i_size changes and the page can be writeably mapped into
2194 * page tables. So an application can be growing i_size and writing
2195 * data through mmap while writeback runs. clear_page_dirty_for_io()
2196 * write-protects our page in page tables and the page cannot get
2197 * written to again until we release page lock. So only after
2198 * clear_page_dirty_for_io() we are safe to sample i_size for
2199 * ext4_bio_write_page() to zero-out tail of the written page. We rely
2200 * on the barrier provided by TestClearPageDirty in
2201 * clear_page_dirty_for_io() to make sure i_size is really sampled only
2202 * after page tables are updated.
2203 */
2204 size = i_size_read(mpd->inode);
2205 if (page->index == size >> PAGE_SHIFT)
2206 len = size & ~PAGE_MASK;
2207 else
2208 len = PAGE_SIZE;
2209 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2210 if (!err)
2211 mpd->wbc->nr_to_write--;
2212 mpd->first_page++;
2213
2214 return err;
2215}
2216
2217#define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2218
2219/*
2220 * mballoc gives us at most this number of blocks...
2221 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2222 * The rest of mballoc seems to handle chunks up to full group size.
2223 */
2224#define MAX_WRITEPAGES_EXTENT_LEN 2048
2225
2226/*
2227 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2228 *
2229 * @mpd - extent of blocks
2230 * @lblk - logical number of the block in the file
2231 * @bh - buffer head we want to add to the extent
2232 *
2233 * The function is used to collect contig. blocks in the same state. If the
2234 * buffer doesn't require mapping for writeback and we haven't started the
2235 * extent of buffers to map yet, the function returns 'true' immediately - the
2236 * caller can write the buffer right away. Otherwise the function returns true
2237 * if the block has been added to the extent, false if the block couldn't be
2238 * added.
2239 */
2240static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2241 struct buffer_head *bh)
2242{
2243 struct ext4_map_blocks *map = &mpd->map;
2244
2245 /* Buffer that doesn't need mapping for writeback? */
2246 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2247 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2248 /* So far no extent to map => we write the buffer right away */
2249 if (map->m_len == 0)
2250 return true;
2251 return false;
2252 }
2253
2254 /* First block in the extent? */
2255 if (map->m_len == 0) {
2256 /* We cannot map unless handle is started... */
2257 if (!mpd->do_map)
2258 return false;
2259 map->m_lblk = lblk;
2260 map->m_len = 1;
2261 map->m_flags = bh->b_state & BH_FLAGS;
2262 return true;
2263 }
2264
2265 /* Don't go larger than mballoc is willing to allocate */
2266 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2267 return false;
2268
2269 /* Can we merge the block to our big extent? */
2270 if (lblk == map->m_lblk + map->m_len &&
2271 (bh->b_state & BH_FLAGS) == map->m_flags) {
2272 map->m_len++;
2273 return true;
2274 }
2275 return false;
2276}
2277
2278/*
2279 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2280 *
2281 * @mpd - extent of blocks for mapping
2282 * @head - the first buffer in the page
2283 * @bh - buffer we should start processing from
2284 * @lblk - logical number of the block in the file corresponding to @bh
2285 *
2286 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2287 * the page for IO if all buffers in this page were mapped and there's no
2288 * accumulated extent of buffers to map or add buffers in the page to the
2289 * extent of buffers to map. The function returns 1 if the caller can continue
2290 * by processing the next page, 0 if it should stop adding buffers to the
2291 * extent to map because we cannot extend it anymore. It can also return value
2292 * < 0 in case of error during IO submission.
2293 */
2294static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2295 struct buffer_head *head,
2296 struct buffer_head *bh,
2297 ext4_lblk_t lblk)
2298{
2299 struct inode *inode = mpd->inode;
2300 int err;
2301 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2302 >> inode->i_blkbits;
2303
2304 do {
2305 BUG_ON(buffer_locked(bh));
2306
2307 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2308 /* Found extent to map? */
2309 if (mpd->map.m_len)
2310 return 0;
2311 /* Buffer needs mapping and handle is not started? */
2312 if (!mpd->do_map)
2313 return 0;
2314 /* Everything mapped so far and we hit EOF */
2315 break;
2316 }
2317 } while (lblk++, (bh = bh->b_this_page) != head);
2318 /* So far everything mapped? Submit the page for IO. */
2319 if (mpd->map.m_len == 0) {
2320 err = mpage_submit_page(mpd, head->b_page);
2321 if (err < 0)
2322 return err;
2323 }
2324 return lblk < blocks;
2325}
2326
2327/*
2328 * mpage_map_buffers - update buffers corresponding to changed extent and
2329 * submit fully mapped pages for IO
2330 *
2331 * @mpd - description of extent to map, on return next extent to map
2332 *
2333 * Scan buffers corresponding to changed extent (we expect corresponding pages
2334 * to be already locked) and update buffer state according to new extent state.
2335 * We map delalloc buffers to their physical location, clear unwritten bits,
2336 * and mark buffers as uninit when we perform writes to unwritten extents
2337 * and do extent conversion after IO is finished. If the last page is not fully
2338 * mapped, we update @map to the next extent in the last page that needs
2339 * mapping. Otherwise we submit the page for IO.
2340 */
2341static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2342{
2343 struct pagevec pvec;
2344 int nr_pages, i;
2345 struct inode *inode = mpd->inode;
2346 struct buffer_head *head, *bh;
2347 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2348 pgoff_t start, end;
2349 ext4_lblk_t lblk;
2350 sector_t pblock;
2351 int err;
2352
2353 start = mpd->map.m_lblk >> bpp_bits;
2354 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2355 lblk = start << bpp_bits;
2356 pblock = mpd->map.m_pblk;
2357
2358 pagevec_init(&pvec);
2359 while (start <= end) {
2360 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2361 &start, end);
2362 if (nr_pages == 0)
2363 break;
2364 for (i = 0; i < nr_pages; i++) {
2365 struct page *page = pvec.pages[i];
2366
2367 bh = head = page_buffers(page);
2368 do {
2369 if (lblk < mpd->map.m_lblk)
2370 continue;
2371 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2372 /*
2373 * Buffer after end of mapped extent.
2374 * Find next buffer in the page to map.
2375 */
2376 mpd->map.m_len = 0;
2377 mpd->map.m_flags = 0;
2378 /*
2379 * FIXME: If dioread_nolock supports
2380 * blocksize < pagesize, we need to make
2381 * sure we add size mapped so far to
2382 * io_end->size as the following call
2383 * can submit the page for IO.
2384 */
2385 err = mpage_process_page_bufs(mpd, head,
2386 bh, lblk);
2387 pagevec_release(&pvec);
2388 if (err > 0)
2389 err = 0;
2390 return err;
2391 }
2392 if (buffer_delay(bh)) {
2393 clear_buffer_delay(bh);
2394 bh->b_blocknr = pblock++;
2395 }
2396 clear_buffer_unwritten(bh);
2397 } while (lblk++, (bh = bh->b_this_page) != head);
2398
2399 /*
2400 * FIXME: This is going to break if dioread_nolock
2401 * supports blocksize < pagesize as we will try to
2402 * convert potentially unmapped parts of inode.
2403 */
2404 mpd->io_submit.io_end->size += PAGE_SIZE;
2405 /* Page fully mapped - let IO run! */
2406 err = mpage_submit_page(mpd, page);
2407 if (err < 0) {
2408 pagevec_release(&pvec);
2409 return err;
2410 }
2411 }
2412 pagevec_release(&pvec);
2413 }
2414 /* Extent fully mapped and matches with page boundary. We are done. */
2415 mpd->map.m_len = 0;
2416 mpd->map.m_flags = 0;
2417 return 0;
2418}
2419
2420static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2421{
2422 struct inode *inode = mpd->inode;
2423 struct ext4_map_blocks *map = &mpd->map;
2424 int get_blocks_flags;
2425 int err, dioread_nolock;
2426
2427 trace_ext4_da_write_pages_extent(inode, map);
2428 /*
2429 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2430 * to convert an unwritten extent to be initialized (in the case
2431 * where we have written into one or more preallocated blocks). It is
2432 * possible that we're going to need more metadata blocks than
2433 * previously reserved. However we must not fail because we're in
2434 * writeback and there is nothing we can do about it so it might result
2435 * in data loss. So use reserved blocks to allocate metadata if
2436 * possible.
2437 *
2438 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2439 * the blocks in question are delalloc blocks. This indicates
2440 * that the blocks and quotas has already been checked when
2441 * the data was copied into the page cache.
2442 */
2443 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2444 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2445 EXT4_GET_BLOCKS_IO_SUBMIT;
2446 dioread_nolock = ext4_should_dioread_nolock(inode);
2447 if (dioread_nolock)
2448 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2449 if (map->m_flags & (1 << BH_Delay))
2450 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2451
2452 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2453 if (err < 0)
2454 return err;
2455 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2456 if (!mpd->io_submit.io_end->handle &&
2457 ext4_handle_valid(handle)) {
2458 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2459 handle->h_rsv_handle = NULL;
2460 }
2461 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2462 }
2463
2464 BUG_ON(map->m_len == 0);
2465 if (map->m_flags & EXT4_MAP_NEW) {
2466 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2467 map->m_len);
2468 }
2469 return 0;
2470}
2471
2472/*
2473 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2474 * mpd->len and submit pages underlying it for IO
2475 *
2476 * @handle - handle for journal operations
2477 * @mpd - extent to map
2478 * @give_up_on_write - we set this to true iff there is a fatal error and there
2479 * is no hope of writing the data. The caller should discard
2480 * dirty pages to avoid infinite loops.
2481 *
2482 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2483 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2484 * them to initialized or split the described range from larger unwritten
2485 * extent. Note that we need not map all the described range since allocation
2486 * can return less blocks or the range is covered by more unwritten extents. We
2487 * cannot map more because we are limited by reserved transaction credits. On
2488 * the other hand we always make sure that the last touched page is fully
2489 * mapped so that it can be written out (and thus forward progress is
2490 * guaranteed). After mapping we submit all mapped pages for IO.
2491 */
2492static int mpage_map_and_submit_extent(handle_t *handle,
2493 struct mpage_da_data *mpd,
2494 bool *give_up_on_write)
2495{
2496 struct inode *inode = mpd->inode;
2497 struct ext4_map_blocks *map = &mpd->map;
2498 int err;
2499 loff_t disksize;
2500 int progress = 0;
2501
2502 mpd->io_submit.io_end->offset =
2503 ((loff_t)map->m_lblk) << inode->i_blkbits;
2504 do {
2505 err = mpage_map_one_extent(handle, mpd);
2506 if (err < 0) {
2507 struct super_block *sb = inode->i_sb;
2508
2509 if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2510 EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2511 goto invalidate_dirty_pages;
2512 /*
2513 * Let the uper layers retry transient errors.
2514 * In the case of ENOSPC, if ext4_count_free_blocks()
2515 * is non-zero, a commit should free up blocks.
2516 */
2517 if ((err == -ENOMEM) ||
2518 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2519 if (progress)
2520 goto update_disksize;
2521 return err;
2522 }
2523 ext4_msg(sb, KERN_CRIT,
2524 "Delayed block allocation failed for "
2525 "inode %lu at logical offset %llu with"
2526 " max blocks %u with error %d",
2527 inode->i_ino,
2528 (unsigned long long)map->m_lblk,
2529 (unsigned)map->m_len, -err);
2530 ext4_msg(sb, KERN_CRIT,
2531 "This should not happen!! Data will "
2532 "be lost\n");
2533 if (err == -ENOSPC)
2534 ext4_print_free_blocks(inode);
2535 invalidate_dirty_pages:
2536 *give_up_on_write = true;
2537 return err;
2538 }
2539 progress = 1;
2540 /*
2541 * Update buffer state, submit mapped pages, and get us new
2542 * extent to map
2543 */
2544 err = mpage_map_and_submit_buffers(mpd);
2545 if (err < 0)
2546 goto update_disksize;
2547 } while (map->m_len);
2548
2549update_disksize:
2550 /*
2551 * Update on-disk size after IO is submitted. Races with
2552 * truncate are avoided by checking i_size under i_data_sem.
2553 */
2554 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2555 if (disksize > EXT4_I(inode)->i_disksize) {
2556 int err2;
2557 loff_t i_size;
2558
2559 down_write(&EXT4_I(inode)->i_data_sem);
2560 i_size = i_size_read(inode);
2561 if (disksize > i_size)
2562 disksize = i_size;
2563 if (disksize > EXT4_I(inode)->i_disksize)
2564 EXT4_I(inode)->i_disksize = disksize;
2565 up_write(&EXT4_I(inode)->i_data_sem);
2566 err2 = ext4_mark_inode_dirty(handle, inode);
2567 if (err2)
2568 ext4_error(inode->i_sb,
2569 "Failed to mark inode %lu dirty",
2570 inode->i_ino);
2571 if (!err)
2572 err = err2;
2573 }
2574 return err;
2575}
2576
2577/*
2578 * Calculate the total number of credits to reserve for one writepages
2579 * iteration. This is called from ext4_writepages(). We map an extent of
2580 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2581 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2582 * bpp - 1 blocks in bpp different extents.
2583 */
2584static int ext4_da_writepages_trans_blocks(struct inode *inode)
2585{
2586 int bpp = ext4_journal_blocks_per_page(inode);
2587
2588 return ext4_meta_trans_blocks(inode,
2589 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2590}
2591
2592/*
2593 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2594 * and underlying extent to map
2595 *
2596 * @mpd - where to look for pages
2597 *
2598 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2599 * IO immediately. When we find a page which isn't mapped we start accumulating
2600 * extent of buffers underlying these pages that needs mapping (formed by
2601 * either delayed or unwritten buffers). We also lock the pages containing
2602 * these buffers. The extent found is returned in @mpd structure (starting at
2603 * mpd->lblk with length mpd->len blocks).
2604 *
2605 * Note that this function can attach bios to one io_end structure which are
2606 * neither logically nor physically contiguous. Although it may seem as an
2607 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2608 * case as we need to track IO to all buffers underlying a page in one io_end.
2609 */
2610static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2611{
2612 struct address_space *mapping = mpd->inode->i_mapping;
2613 struct pagevec pvec;
2614 unsigned int nr_pages;
2615 long left = mpd->wbc->nr_to_write;
2616 pgoff_t index = mpd->first_page;
2617 pgoff_t end = mpd->last_page;
2618 int tag;
2619 int i, err = 0;
2620 int blkbits = mpd->inode->i_blkbits;
2621 ext4_lblk_t lblk;
2622 struct buffer_head *head;
2623
2624 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2625 tag = PAGECACHE_TAG_TOWRITE;
2626 else
2627 tag = PAGECACHE_TAG_DIRTY;
2628
2629 pagevec_init(&pvec);
2630 mpd->map.m_len = 0;
2631 mpd->next_page = index;
2632 while (index <= end) {
2633 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2634 tag);
2635 if (nr_pages == 0)
2636 goto out;
2637
2638 for (i = 0; i < nr_pages; i++) {
2639 struct page *page = pvec.pages[i];
2640
2641 /*
2642 * Accumulated enough dirty pages? This doesn't apply
2643 * to WB_SYNC_ALL mode. For integrity sync we have to
2644 * keep going because someone may be concurrently
2645 * dirtying pages, and we might have synced a lot of
2646 * newly appeared dirty pages, but have not synced all
2647 * of the old dirty pages.
2648 */
2649 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2650 goto out;
2651
2652 /* If we can't merge this page, we are done. */
2653 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2654 goto out;
2655
2656 lock_page(page);
2657 /*
2658 * If the page is no longer dirty, or its mapping no
2659 * longer corresponds to inode we are writing (which
2660 * means it has been truncated or invalidated), or the
2661 * page is already under writeback and we are not doing
2662 * a data integrity writeback, skip the page
2663 */
2664 if (!PageDirty(page) ||
2665 (PageWriteback(page) &&
2666 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2667 unlikely(page->mapping != mapping)) {
2668 unlock_page(page);
2669 continue;
2670 }
2671
2672 wait_on_page_writeback(page);
2673 BUG_ON(PageWriteback(page));
2674
2675 if (mpd->map.m_len == 0)
2676 mpd->first_page = page->index;
2677 mpd->next_page = page->index + 1;
2678 /* Add all dirty buffers to mpd */
2679 lblk = ((ext4_lblk_t)page->index) <<
2680 (PAGE_SHIFT - blkbits);
2681 head = page_buffers(page);
2682 err = mpage_process_page_bufs(mpd, head, head, lblk);
2683 if (err <= 0)
2684 goto out;
2685 err = 0;
2686 left--;
2687 }
2688 pagevec_release(&pvec);
2689 cond_resched();
2690 }
2691 return 0;
2692out:
2693 pagevec_release(&pvec);
2694 return err;
2695}
2696
2697static int ext4_writepages(struct address_space *mapping,
2698 struct writeback_control *wbc)
2699{
2700 pgoff_t writeback_index = 0;
2701 long nr_to_write = wbc->nr_to_write;
2702 int range_whole = 0;
2703 int cycled = 1;
2704 handle_t *handle = NULL;
2705 struct mpage_da_data mpd;
2706 struct inode *inode = mapping->host;
2707 int needed_blocks, rsv_blocks = 0, ret = 0;
2708 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2709 bool done;
2710 struct blk_plug plug;
2711 bool give_up_on_write = false;
2712
2713 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2714 return -EIO;
2715
2716 percpu_down_read(&sbi->s_journal_flag_rwsem);
2717 trace_ext4_writepages(inode, wbc);
2718
2719 /*
2720 * No pages to write? This is mainly a kludge to avoid starting
2721 * a transaction for special inodes like journal inode on last iput()
2722 * because that could violate lock ordering on umount
2723 */
2724 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2725 goto out_writepages;
2726
2727 if (ext4_should_journal_data(inode)) {
2728 ret = generic_writepages(mapping, wbc);
2729 goto out_writepages;
2730 }
2731
2732 /*
2733 * If the filesystem has aborted, it is read-only, so return
2734 * right away instead of dumping stack traces later on that
2735 * will obscure the real source of the problem. We test
2736 * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2737 * the latter could be true if the filesystem is mounted
2738 * read-only, and in that case, ext4_writepages should
2739 * *never* be called, so if that ever happens, we would want
2740 * the stack trace.
2741 */
2742 if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2743 sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2744 ret = -EROFS;
2745 goto out_writepages;
2746 }
2747
2748 if (ext4_should_dioread_nolock(inode)) {
2749 /*
2750 * We may need to convert up to one extent per block in
2751 * the page and we may dirty the inode.
2752 */
2753 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2754 }
2755
2756 /*
2757 * If we have inline data and arrive here, it means that
2758 * we will soon create the block for the 1st page, so
2759 * we'd better clear the inline data here.
2760 */
2761 if (ext4_has_inline_data(inode)) {
2762 /* Just inode will be modified... */
2763 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2764 if (IS_ERR(handle)) {
2765 ret = PTR_ERR(handle);
2766 goto out_writepages;
2767 }
2768 BUG_ON(ext4_test_inode_state(inode,
2769 EXT4_STATE_MAY_INLINE_DATA));
2770 ext4_destroy_inline_data(handle, inode);
2771 ext4_journal_stop(handle);
2772 }
2773
2774 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2775 range_whole = 1;
2776
2777 if (wbc->range_cyclic) {
2778 writeback_index = mapping->writeback_index;
2779 if (writeback_index)
2780 cycled = 0;
2781 mpd.first_page = writeback_index;
2782 mpd.last_page = -1;
2783 } else {
2784 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2785 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2786 }
2787
2788 mpd.inode = inode;
2789 mpd.wbc = wbc;
2790 ext4_io_submit_init(&mpd.io_submit, wbc);
2791retry:
2792 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2793 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2794 done = false;
2795 blk_start_plug(&plug);
2796
2797 /*
2798 * First writeback pages that don't need mapping - we can avoid
2799 * starting a transaction unnecessarily and also avoid being blocked
2800 * in the block layer on device congestion while having transaction
2801 * started.
2802 */
2803 mpd.do_map = 0;
2804 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2805 if (!mpd.io_submit.io_end) {
2806 ret = -ENOMEM;
2807 goto unplug;
2808 }
2809 ret = mpage_prepare_extent_to_map(&mpd);
2810 /* Submit prepared bio */
2811 ext4_io_submit(&mpd.io_submit);
2812 ext4_put_io_end_defer(mpd.io_submit.io_end);
2813 mpd.io_submit.io_end = NULL;
2814 /* Unlock pages we didn't use */
2815 mpage_release_unused_pages(&mpd, false);
2816 if (ret < 0)
2817 goto unplug;
2818
2819 while (!done && mpd.first_page <= mpd.last_page) {
2820 /* For each extent of pages we use new io_end */
2821 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2822 if (!mpd.io_submit.io_end) {
2823 ret = -ENOMEM;
2824 break;
2825 }
2826
2827 /*
2828 * We have two constraints: We find one extent to map and we
2829 * must always write out whole page (makes a difference when
2830 * blocksize < pagesize) so that we don't block on IO when we
2831 * try to write out the rest of the page. Journalled mode is
2832 * not supported by delalloc.
2833 */
2834 BUG_ON(ext4_should_journal_data(inode));
2835 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2836
2837 /* start a new transaction */
2838 handle = ext4_journal_start_with_reserve(inode,
2839 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2840 if (IS_ERR(handle)) {
2841 ret = PTR_ERR(handle);
2842 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2843 "%ld pages, ino %lu; err %d", __func__,
2844 wbc->nr_to_write, inode->i_ino, ret);
2845 /* Release allocated io_end */
2846 ext4_put_io_end(mpd.io_submit.io_end);
2847 mpd.io_submit.io_end = NULL;
2848 break;
2849 }
2850 mpd.do_map = 1;
2851
2852 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2853 ret = mpage_prepare_extent_to_map(&mpd);
2854 if (!ret) {
2855 if (mpd.map.m_len)
2856 ret = mpage_map_and_submit_extent(handle, &mpd,
2857 &give_up_on_write);
2858 else {
2859 /*
2860 * We scanned the whole range (or exhausted
2861 * nr_to_write), submitted what was mapped and
2862 * didn't find anything needing mapping. We are
2863 * done.
2864 */
2865 done = true;
2866 }
2867 }
2868 /*
2869 * Caution: If the handle is synchronous,
2870 * ext4_journal_stop() can wait for transaction commit
2871 * to finish which may depend on writeback of pages to
2872 * complete or on page lock to be released. In that
2873 * case, we have to wait until after after we have
2874 * submitted all the IO, released page locks we hold,
2875 * and dropped io_end reference (for extent conversion
2876 * to be able to complete) before stopping the handle.
2877 */
2878 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2879 ext4_journal_stop(handle);
2880 handle = NULL;
2881 mpd.do_map = 0;
2882 }
2883 /* Submit prepared bio */
2884 ext4_io_submit(&mpd.io_submit);
2885 /* Unlock pages we didn't use */
2886 mpage_release_unused_pages(&mpd, give_up_on_write);
2887 /*
2888 * Drop our io_end reference we got from init. We have
2889 * to be careful and use deferred io_end finishing if
2890 * we are still holding the transaction as we can
2891 * release the last reference to io_end which may end
2892 * up doing unwritten extent conversion.
2893 */
2894 if (handle) {
2895 ext4_put_io_end_defer(mpd.io_submit.io_end);
2896 ext4_journal_stop(handle);
2897 } else
2898 ext4_put_io_end(mpd.io_submit.io_end);
2899 mpd.io_submit.io_end = NULL;
2900
2901 if (ret == -ENOSPC && sbi->s_journal) {
2902 /*
2903 * Commit the transaction which would
2904 * free blocks released in the transaction
2905 * and try again
2906 */
2907 jbd2_journal_force_commit_nested(sbi->s_journal);
2908 ret = 0;
2909 continue;
2910 }
2911 /* Fatal error - ENOMEM, EIO... */
2912 if (ret)
2913 break;
2914 }
2915unplug:
2916 blk_finish_plug(&plug);
2917 if (!ret && !cycled && wbc->nr_to_write > 0) {
2918 cycled = 1;
2919 mpd.last_page = writeback_index - 1;
2920 mpd.first_page = 0;
2921 goto retry;
2922 }
2923
2924 /* Update index */
2925 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2926 /*
2927 * Set the writeback_index so that range_cyclic
2928 * mode will write it back later
2929 */
2930 mapping->writeback_index = mpd.first_page;
2931
2932out_writepages:
2933 trace_ext4_writepages_result(inode, wbc, ret,
2934 nr_to_write - wbc->nr_to_write);
2935 percpu_up_read(&sbi->s_journal_flag_rwsem);
2936 return ret;
2937}
2938
2939static int ext4_dax_writepages(struct address_space *mapping,
2940 struct writeback_control *wbc)
2941{
2942 int ret;
2943 long nr_to_write = wbc->nr_to_write;
2944 struct inode *inode = mapping->host;
2945 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2946
2947 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2948 return -EIO;
2949
2950 percpu_down_read(&sbi->s_journal_flag_rwsem);
2951 trace_ext4_writepages(inode, wbc);
2952
2953 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev, wbc);
2954 trace_ext4_writepages_result(inode, wbc, ret,
2955 nr_to_write - wbc->nr_to_write);
2956 percpu_up_read(&sbi->s_journal_flag_rwsem);
2957 return ret;
2958}
2959
2960static int ext4_nonda_switch(struct super_block *sb)
2961{
2962 s64 free_clusters, dirty_clusters;
2963 struct ext4_sb_info *sbi = EXT4_SB(sb);
2964
2965 /*
2966 * switch to non delalloc mode if we are running low
2967 * on free block. The free block accounting via percpu
2968 * counters can get slightly wrong with percpu_counter_batch getting
2969 * accumulated on each CPU without updating global counters
2970 * Delalloc need an accurate free block accounting. So switch
2971 * to non delalloc when we are near to error range.
2972 */
2973 free_clusters =
2974 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2975 dirty_clusters =
2976 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2977 /*
2978 * Start pushing delalloc when 1/2 of free blocks are dirty.
2979 */
2980 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2981 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2982
2983 if (2 * free_clusters < 3 * dirty_clusters ||
2984 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2985 /*
2986 * free block count is less than 150% of dirty blocks
2987 * or free blocks is less than watermark
2988 */
2989 return 1;
2990 }
2991 return 0;
2992}
2993
2994/* We always reserve for an inode update; the superblock could be there too */
2995static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2996{
2997 if (likely(ext4_has_feature_large_file(inode->i_sb)))
2998 return 1;
2999
3000 if (pos + len <= 0x7fffffffULL)
3001 return 1;
3002
3003 /* We might need to update the superblock to set LARGE_FILE */
3004 return 2;
3005}
3006
3007static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3008 loff_t pos, unsigned len, unsigned flags,
3009 struct page **pagep, void **fsdata)
3010{
3011 int ret, retries = 0;
3012 struct page *page;
3013 pgoff_t index;
3014 struct inode *inode = mapping->host;
3015 handle_t *handle;
3016
3017 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3018 return -EIO;
3019
3020 index = pos >> PAGE_SHIFT;
3021
3022 if (ext4_nonda_switch(inode->i_sb) ||
3023 S_ISLNK(inode->i_mode)) {
3024 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3025 return ext4_write_begin(file, mapping, pos,
3026 len, flags, pagep, fsdata);
3027 }
3028 *fsdata = (void *)0;
3029 trace_ext4_da_write_begin(inode, pos, len, flags);
3030
3031 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3032 ret = ext4_da_write_inline_data_begin(mapping, inode,
3033 pos, len, flags,
3034 pagep, fsdata);
3035 if (ret < 0)
3036 return ret;
3037 if (ret == 1)
3038 return 0;
3039 }
3040
3041 /*
3042 * grab_cache_page_write_begin() can take a long time if the
3043 * system is thrashing due to memory pressure, or if the page
3044 * is being written back. So grab it first before we start
3045 * the transaction handle. This also allows us to allocate
3046 * the page (if needed) without using GFP_NOFS.
3047 */
3048retry_grab:
3049 page = grab_cache_page_write_begin(mapping, index, flags);
3050 if (!page)
3051 return -ENOMEM;
3052 unlock_page(page);
3053
3054 /*
3055 * With delayed allocation, we don't log the i_disksize update
3056 * if there is delayed block allocation. But we still need
3057 * to journalling the i_disksize update if writes to the end
3058 * of file which has an already mapped buffer.
3059 */
3060retry_journal:
3061 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3062 ext4_da_write_credits(inode, pos, len));
3063 if (IS_ERR(handle)) {
3064 put_page(page);
3065 return PTR_ERR(handle);
3066 }
3067
3068 lock_page(page);
3069 if (page->mapping != mapping) {
3070 /* The page got truncated from under us */
3071 unlock_page(page);
3072 put_page(page);
3073 ext4_journal_stop(handle);
3074 goto retry_grab;
3075 }
3076 /* In case writeback began while the page was unlocked */
3077 wait_for_stable_page(page);
3078
3079#ifdef CONFIG_EXT4_FS_ENCRYPTION
3080 ret = ext4_block_write_begin(page, pos, len,
3081 ext4_da_get_block_prep);
3082#else
3083 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3084#endif
3085 if (ret < 0) {
3086 unlock_page(page);
3087 ext4_journal_stop(handle);
3088 /*
3089 * block_write_begin may have instantiated a few blocks
3090 * outside i_size. Trim these off again. Don't need
3091 * i_size_read because we hold i_mutex.
3092 */
3093 if (pos + len > inode->i_size)
3094 ext4_truncate_failed_write(inode);
3095
3096 if (ret == -ENOSPC &&
3097 ext4_should_retry_alloc(inode->i_sb, &retries))
3098 goto retry_journal;
3099
3100 put_page(page);
3101 return ret;
3102 }
3103
3104 *pagep = page;
3105 return ret;
3106}
3107
3108/*
3109 * Check if we should update i_disksize
3110 * when write to the end of file but not require block allocation
3111 */
3112static int ext4_da_should_update_i_disksize(struct page *page,
3113 unsigned long offset)
3114{
3115 struct buffer_head *bh;
3116 struct inode *inode = page->mapping->host;
3117 unsigned int idx;
3118 int i;
3119
3120 bh = page_buffers(page);
3121 idx = offset >> inode->i_blkbits;
3122
3123 for (i = 0; i < idx; i++)
3124 bh = bh->b_this_page;
3125
3126 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3127 return 0;
3128 return 1;
3129}
3130
3131static int ext4_da_write_end(struct file *file,
3132 struct address_space *mapping,
3133 loff_t pos, unsigned len, unsigned copied,
3134 struct page *page, void *fsdata)
3135{
3136 struct inode *inode = mapping->host;
3137 int ret = 0, ret2;
3138 handle_t *handle = ext4_journal_current_handle();
3139 loff_t new_i_size;
3140 unsigned long start, end;
3141 int write_mode = (int)(unsigned long)fsdata;
3142
3143 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3144 return ext4_write_end(file, mapping, pos,
3145 len, copied, page, fsdata);
3146
3147 trace_ext4_da_write_end(inode, pos, len, copied);
3148 start = pos & (PAGE_SIZE - 1);
3149 end = start + copied - 1;
3150
3151 /*
3152 * generic_write_end() will run mark_inode_dirty() if i_size
3153 * changes. So let's piggyback the i_disksize mark_inode_dirty
3154 * into that.
3155 */
3156 new_i_size = pos + copied;
3157 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3158 if (ext4_has_inline_data(inode) ||
3159 ext4_da_should_update_i_disksize(page, end)) {
3160 ext4_update_i_disksize(inode, new_i_size);
3161 /* We need to mark inode dirty even if
3162 * new_i_size is less that inode->i_size
3163 * bu greater than i_disksize.(hint delalloc)
3164 */
3165 ext4_mark_inode_dirty(handle, inode);
3166 }
3167 }
3168
3169 if (write_mode != CONVERT_INLINE_DATA &&
3170 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3171 ext4_has_inline_data(inode))
3172 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3173 page);
3174 else
3175 ret2 = generic_write_end(file, mapping, pos, len, copied,
3176 page, fsdata);
3177
3178 copied = ret2;
3179 if (ret2 < 0)
3180 ret = ret2;
3181 ret2 = ext4_journal_stop(handle);
3182 if (!ret)
3183 ret = ret2;
3184
3185 return ret ? ret : copied;
3186}
3187
3188static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3189 unsigned int length)
3190{
3191 /*
3192 * Drop reserved blocks
3193 */
3194 BUG_ON(!PageLocked(page));
3195 if (!page_has_buffers(page))
3196 goto out;
3197
3198 ext4_da_page_release_reservation(page, offset, length);
3199
3200out:
3201 ext4_invalidatepage(page, offset, length);
3202
3203 return;
3204}
3205
3206/*
3207 * Force all delayed allocation blocks to be allocated for a given inode.
3208 */
3209int ext4_alloc_da_blocks(struct inode *inode)
3210{
3211 trace_ext4_alloc_da_blocks(inode);
3212
3213 if (!EXT4_I(inode)->i_reserved_data_blocks)
3214 return 0;
3215
3216 /*
3217 * We do something simple for now. The filemap_flush() will
3218 * also start triggering a write of the data blocks, which is
3219 * not strictly speaking necessary (and for users of
3220 * laptop_mode, not even desirable). However, to do otherwise
3221 * would require replicating code paths in:
3222 *
3223 * ext4_writepages() ->
3224 * write_cache_pages() ---> (via passed in callback function)
3225 * __mpage_da_writepage() -->
3226 * mpage_add_bh_to_extent()
3227 * mpage_da_map_blocks()
3228 *
3229 * The problem is that write_cache_pages(), located in
3230 * mm/page-writeback.c, marks pages clean in preparation for
3231 * doing I/O, which is not desirable if we're not planning on
3232 * doing I/O at all.
3233 *
3234 * We could call write_cache_pages(), and then redirty all of
3235 * the pages by calling redirty_page_for_writepage() but that
3236 * would be ugly in the extreme. So instead we would need to
3237 * replicate parts of the code in the above functions,
3238 * simplifying them because we wouldn't actually intend to
3239 * write out the pages, but rather only collect contiguous
3240 * logical block extents, call the multi-block allocator, and
3241 * then update the buffer heads with the block allocations.
3242 *
3243 * For now, though, we'll cheat by calling filemap_flush(),
3244 * which will map the blocks, and start the I/O, but not
3245 * actually wait for the I/O to complete.
3246 */
3247 return filemap_flush(inode->i_mapping);
3248}
3249
3250/*
3251 * bmap() is special. It gets used by applications such as lilo and by
3252 * the swapper to find the on-disk block of a specific piece of data.
3253 *
3254 * Naturally, this is dangerous if the block concerned is still in the
3255 * journal. If somebody makes a swapfile on an ext4 data-journaling
3256 * filesystem and enables swap, then they may get a nasty shock when the
3257 * data getting swapped to that swapfile suddenly gets overwritten by
3258 * the original zero's written out previously to the journal and
3259 * awaiting writeback in the kernel's buffer cache.
3260 *
3261 * So, if we see any bmap calls here on a modified, data-journaled file,
3262 * take extra steps to flush any blocks which might be in the cache.
3263 */
3264static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3265{
3266 struct inode *inode = mapping->host;
3267 journal_t *journal;
3268 int err;
3269
3270 /*
3271 * We can get here for an inline file via the FIBMAP ioctl
3272 */
3273 if (ext4_has_inline_data(inode))
3274 return 0;
3275
3276 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3277 test_opt(inode->i_sb, DELALLOC)) {
3278 /*
3279 * With delalloc we want to sync the file
3280 * so that we can make sure we allocate
3281 * blocks for file
3282 */
3283 filemap_write_and_wait(mapping);
3284 }
3285
3286 if (EXT4_JOURNAL(inode) &&
3287 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3288 /*
3289 * This is a REALLY heavyweight approach, but the use of
3290 * bmap on dirty files is expected to be extremely rare:
3291 * only if we run lilo or swapon on a freshly made file
3292 * do we expect this to happen.
3293 *
3294 * (bmap requires CAP_SYS_RAWIO so this does not
3295 * represent an unprivileged user DOS attack --- we'd be
3296 * in trouble if mortal users could trigger this path at
3297 * will.)
3298 *
3299 * NB. EXT4_STATE_JDATA is not set on files other than
3300 * regular files. If somebody wants to bmap a directory
3301 * or symlink and gets confused because the buffer
3302 * hasn't yet been flushed to disk, they deserve
3303 * everything they get.
3304 */
3305
3306 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3307 journal = EXT4_JOURNAL(inode);
3308 jbd2_journal_lock_updates(journal);
3309 err = jbd2_journal_flush(journal);
3310 jbd2_journal_unlock_updates(journal);
3311
3312 if (err)
3313 return 0;
3314 }
3315
3316 return generic_block_bmap(mapping, block, ext4_get_block);
3317}
3318
3319static int ext4_readpage(struct file *file, struct page *page)
3320{
3321 int ret = -EAGAIN;
3322 struct inode *inode = page->mapping->host;
3323
3324 trace_ext4_readpage(page);
3325
3326 if (ext4_has_inline_data(inode))
3327 ret = ext4_readpage_inline(inode, page);
3328
3329 if (ret == -EAGAIN)
3330 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3331
3332 return ret;
3333}
3334
3335static int
3336ext4_readpages(struct file *file, struct address_space *mapping,
3337 struct list_head *pages, unsigned nr_pages)
3338{
3339 struct inode *inode = mapping->host;
3340
3341 /* If the file has inline data, no need to do readpages. */
3342 if (ext4_has_inline_data(inode))
3343 return 0;
3344
3345 return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3346}
3347
3348static void ext4_invalidatepage(struct page *page, unsigned int offset,
3349 unsigned int length)
3350{
3351 trace_ext4_invalidatepage(page, offset, length);
3352
3353 /* No journalling happens on data buffers when this function is used */
3354 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3355
3356 block_invalidatepage(page, offset, length);
3357}
3358
3359static int __ext4_journalled_invalidatepage(struct page *page,
3360 unsigned int offset,
3361 unsigned int length)
3362{
3363 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3364
3365 trace_ext4_journalled_invalidatepage(page, offset, length);
3366
3367 /*
3368 * If it's a full truncate we just forget about the pending dirtying
3369 */
3370 if (offset == 0 && length == PAGE_SIZE)
3371 ClearPageChecked(page);
3372
3373 return jbd2_journal_invalidatepage(journal, page, offset, length);
3374}
3375
3376/* Wrapper for aops... */
3377static void ext4_journalled_invalidatepage(struct page *page,
3378 unsigned int offset,
3379 unsigned int length)
3380{
3381 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3382}
3383
3384static int ext4_releasepage(struct page *page, gfp_t wait)
3385{
3386 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3387
3388 trace_ext4_releasepage(page);
3389
3390 /* Page has dirty journalled data -> cannot release */
3391 if (PageChecked(page))
3392 return 0;
3393 if (journal)
3394 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3395 else
3396 return try_to_free_buffers(page);
3397}
3398
3399static bool ext4_inode_datasync_dirty(struct inode *inode)
3400{
3401 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3402
3403 if (journal)
3404 return !jbd2_transaction_committed(journal,
3405 EXT4_I(inode)->i_datasync_tid);
3406 /* Any metadata buffers to write? */
3407 if (!list_empty(&inode->i_mapping->private_list))
3408 return true;
3409 return inode->i_state & I_DIRTY_DATASYNC;
3410}
3411
3412static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3413 unsigned flags, struct iomap *iomap)
3414{
3415 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3416 unsigned int blkbits = inode->i_blkbits;
3417 unsigned long first_block = offset >> blkbits;
3418 unsigned long last_block = (offset + length - 1) >> blkbits;
3419 struct ext4_map_blocks map;
3420 bool delalloc = false;
3421 int ret;
3422
3423
3424 if (flags & IOMAP_REPORT) {
3425 if (ext4_has_inline_data(inode)) {
3426 ret = ext4_inline_data_iomap(inode, iomap);
3427 if (ret != -EAGAIN) {
3428 if (ret == 0 && offset >= iomap->length)
3429 ret = -ENOENT;
3430 return ret;
3431 }
3432 }
3433 } else {
3434 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3435 return -ERANGE;
3436 }
3437
3438 map.m_lblk = first_block;
3439 map.m_len = last_block - first_block + 1;
3440
3441 if (flags & IOMAP_REPORT) {
3442 ret = ext4_map_blocks(NULL, inode, &map, 0);
3443 if (ret < 0)
3444 return ret;
3445
3446 if (ret == 0) {
3447 ext4_lblk_t end = map.m_lblk + map.m_len - 1;
3448 struct extent_status es;
3449
3450 ext4_es_find_delayed_extent_range(inode, map.m_lblk, end, &es);
3451
3452 if (!es.es_len || es.es_lblk > end) {
3453 /* entire range is a hole */
3454 } else if (es.es_lblk > map.m_lblk) {
3455 /* range starts with a hole */
3456 map.m_len = es.es_lblk - map.m_lblk;
3457 } else {
3458 ext4_lblk_t offs = 0;
3459
3460 if (es.es_lblk < map.m_lblk)
3461 offs = map.m_lblk - es.es_lblk;
3462 map.m_lblk = es.es_lblk + offs;
3463 map.m_len = es.es_len - offs;
3464 delalloc = true;
3465 }
3466 }
3467 } else if (flags & IOMAP_WRITE) {
3468 int dio_credits;
3469 handle_t *handle;
3470 int retries = 0;
3471
3472 /* Trim mapping request to maximum we can map at once for DIO */
3473 if (map.m_len > DIO_MAX_BLOCKS)
3474 map.m_len = DIO_MAX_BLOCKS;
3475 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3476retry:
3477 /*
3478 * Either we allocate blocks and then we don't get unwritten
3479 * extent so we have reserved enough credits, or the blocks
3480 * are already allocated and unwritten and in that case
3481 * extent conversion fits in the credits as well.
3482 */
3483 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3484 dio_credits);
3485 if (IS_ERR(handle))
3486 return PTR_ERR(handle);
3487
3488 ret = ext4_map_blocks(handle, inode, &map,
3489 EXT4_GET_BLOCKS_CREATE_ZERO);
3490 if (ret < 0) {
3491 ext4_journal_stop(handle);
3492 if (ret == -ENOSPC &&
3493 ext4_should_retry_alloc(inode->i_sb, &retries))
3494 goto retry;
3495 return ret;
3496 }
3497
3498 /*
3499 * If we added blocks beyond i_size, we need to make sure they
3500 * will get truncated if we crash before updating i_size in
3501 * ext4_iomap_end(). For faults we don't need to do that (and
3502 * even cannot because for orphan list operations inode_lock is
3503 * required) - if we happen to instantiate block beyond i_size,
3504 * it is because we race with truncate which has already added
3505 * the inode to the orphan list.
3506 */
3507 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3508 (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3509 int err;
3510
3511 err = ext4_orphan_add(handle, inode);
3512 if (err < 0) {
3513 ext4_journal_stop(handle);
3514 return err;
3515 }
3516 }
3517 ext4_journal_stop(handle);
3518 } else {
3519 ret = ext4_map_blocks(NULL, inode, &map, 0);
3520 if (ret < 0)
3521 return ret;
3522 }
3523
3524 iomap->flags = 0;
3525 if (ext4_inode_datasync_dirty(inode))
3526 iomap->flags |= IOMAP_F_DIRTY;
3527 iomap->bdev = inode->i_sb->s_bdev;
3528 iomap->dax_dev = sbi->s_daxdev;
3529 iomap->offset = (u64)first_block << blkbits;
3530 iomap->length = (u64)map.m_len << blkbits;
3531
3532 if (ret == 0) {
3533 iomap->type = delalloc ? IOMAP_DELALLOC : IOMAP_HOLE;
3534 iomap->addr = IOMAP_NULL_ADDR;
3535 } else {
3536 if (map.m_flags & EXT4_MAP_MAPPED) {
3537 iomap->type = IOMAP_MAPPED;
3538 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3539 iomap->type = IOMAP_UNWRITTEN;
3540 } else {
3541 WARN_ON_ONCE(1);
3542 return -EIO;
3543 }
3544 iomap->addr = (u64)map.m_pblk << blkbits;
3545 }
3546
3547 if (map.m_flags & EXT4_MAP_NEW)
3548 iomap->flags |= IOMAP_F_NEW;
3549
3550 return 0;
3551}
3552
3553static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3554 ssize_t written, unsigned flags, struct iomap *iomap)
3555{
3556 int ret = 0;
3557 handle_t *handle;
3558 int blkbits = inode->i_blkbits;
3559 bool truncate = false;
3560
3561 if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3562 return 0;
3563
3564 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3565 if (IS_ERR(handle)) {
3566 ret = PTR_ERR(handle);
3567 goto orphan_del;
3568 }
3569 if (ext4_update_inode_size(inode, offset + written))
3570 ext4_mark_inode_dirty(handle, inode);
3571 /*
3572 * We may need to truncate allocated but not written blocks beyond EOF.
3573 */
3574 if (iomap->offset + iomap->length >
3575 ALIGN(inode->i_size, 1 << blkbits)) {
3576 ext4_lblk_t written_blk, end_blk;
3577
3578 written_blk = (offset + written) >> blkbits;
3579 end_blk = (offset + length) >> blkbits;
3580 if (written_blk < end_blk && ext4_can_truncate(inode))
3581 truncate = true;
3582 }
3583 /*
3584 * Remove inode from orphan list if we were extending a inode and
3585 * everything went fine.
3586 */
3587 if (!truncate && inode->i_nlink &&
3588 !list_empty(&EXT4_I(inode)->i_orphan))
3589 ext4_orphan_del(handle, inode);
3590 ext4_journal_stop(handle);
3591 if (truncate) {
3592 ext4_truncate_failed_write(inode);
3593orphan_del:
3594 /*
3595 * If truncate failed early the inode might still be on the
3596 * orphan list; we need to make sure the inode is removed from
3597 * the orphan list in that case.
3598 */
3599 if (inode->i_nlink)
3600 ext4_orphan_del(NULL, inode);
3601 }
3602 return ret;
3603}
3604
3605const struct iomap_ops ext4_iomap_ops = {
3606 .iomap_begin = ext4_iomap_begin,
3607 .iomap_end = ext4_iomap_end,
3608};
3609
3610static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3611 ssize_t size, void *private)
3612{
3613 ext4_io_end_t *io_end = private;
3614
3615 /* if not async direct IO just return */
3616 if (!io_end)
3617 return 0;
3618
3619 ext_debug("ext4_end_io_dio(): io_end 0x%p "
3620 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3621 io_end, io_end->inode->i_ino, iocb, offset, size);
3622
3623 /*
3624 * Error during AIO DIO. We cannot convert unwritten extents as the
3625 * data was not written. Just clear the unwritten flag and drop io_end.
3626 */
3627 if (size <= 0) {
3628 ext4_clear_io_unwritten_flag(io_end);
3629 size = 0;
3630 }
3631 io_end->offset = offset;
3632 io_end->size = size;
3633 ext4_put_io_end(io_end);
3634
3635 return 0;
3636}
3637
3638/*
3639 * Handling of direct IO writes.
3640 *
3641 * For ext4 extent files, ext4 will do direct-io write even to holes,
3642 * preallocated extents, and those write extend the file, no need to
3643 * fall back to buffered IO.
3644 *
3645 * For holes, we fallocate those blocks, mark them as unwritten
3646 * If those blocks were preallocated, we mark sure they are split, but
3647 * still keep the range to write as unwritten.
3648 *
3649 * The unwritten extents will be converted to written when DIO is completed.
3650 * For async direct IO, since the IO may still pending when return, we
3651 * set up an end_io call back function, which will do the conversion
3652 * when async direct IO completed.
3653 *
3654 * If the O_DIRECT write will extend the file then add this inode to the
3655 * orphan list. So recovery will truncate it back to the original size
3656 * if the machine crashes during the write.
3657 *
3658 */
3659static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3660{
3661 struct file *file = iocb->ki_filp;
3662 struct inode *inode = file->f_mapping->host;
3663 struct ext4_inode_info *ei = EXT4_I(inode);
3664 ssize_t ret;
3665 loff_t offset = iocb->ki_pos;
3666 size_t count = iov_iter_count(iter);
3667 int overwrite = 0;
3668 get_block_t *get_block_func = NULL;
3669 int dio_flags = 0;
3670 loff_t final_size = offset + count;
3671 int orphan = 0;
3672 handle_t *handle;
3673
3674 if (final_size > inode->i_size || final_size > ei->i_disksize) {
3675 /* Credits for sb + inode write */
3676 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3677 if (IS_ERR(handle)) {
3678 ret = PTR_ERR(handle);
3679 goto out;
3680 }
3681 ret = ext4_orphan_add(handle, inode);
3682 if (ret) {
3683 ext4_journal_stop(handle);
3684 goto out;
3685 }
3686 orphan = 1;
3687 ext4_update_i_disksize(inode, inode->i_size);
3688 ext4_journal_stop(handle);
3689 }
3690
3691 BUG_ON(iocb->private == NULL);
3692
3693 /*
3694 * Make all waiters for direct IO properly wait also for extent
3695 * conversion. This also disallows race between truncate() and
3696 * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3697 */
3698 inode_dio_begin(inode);
3699
3700 /* If we do a overwrite dio, i_mutex locking can be released */
3701 overwrite = *((int *)iocb->private);
3702
3703 if (overwrite)
3704 inode_unlock(inode);
3705
3706 /*
3707 * For extent mapped files we could direct write to holes and fallocate.
3708 *
3709 * Allocated blocks to fill the hole are marked as unwritten to prevent
3710 * parallel buffered read to expose the stale data before DIO complete
3711 * the data IO.
3712 *
3713 * As to previously fallocated extents, ext4 get_block will just simply
3714 * mark the buffer mapped but still keep the extents unwritten.
3715 *
3716 * For non AIO case, we will convert those unwritten extents to written
3717 * after return back from blockdev_direct_IO. That way we save us from
3718 * allocating io_end structure and also the overhead of offloading
3719 * the extent convertion to a workqueue.
3720 *
3721 * For async DIO, the conversion needs to be deferred when the
3722 * IO is completed. The ext4 end_io callback function will be
3723 * called to take care of the conversion work. Here for async
3724 * case, we allocate an io_end structure to hook to the iocb.
3725 */
3726 iocb->private = NULL;
3727 if (overwrite)
3728 get_block_func = ext4_dio_get_block_overwrite;
3729 else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
3730 round_down(offset, i_blocksize(inode)) >= inode->i_size) {
3731 get_block_func = ext4_dio_get_block;
3732 dio_flags = DIO_LOCKING | DIO_SKIP_HOLES;
3733 } else if (is_sync_kiocb(iocb)) {
3734 get_block_func = ext4_dio_get_block_unwritten_sync;
3735 dio_flags = DIO_LOCKING;
3736 } else {
3737 get_block_func = ext4_dio_get_block_unwritten_async;
3738 dio_flags = DIO_LOCKING;
3739 }
3740 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
3741 get_block_func, ext4_end_io_dio, NULL,
3742 dio_flags);
3743
3744 if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3745 EXT4_STATE_DIO_UNWRITTEN)) {
3746 int err;
3747 /*
3748 * for non AIO case, since the IO is already
3749 * completed, we could do the conversion right here
3750 */
3751 err = ext4_convert_unwritten_extents(NULL, inode,
3752 offset, ret);
3753 if (err < 0)
3754 ret = err;
3755 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3756 }
3757
3758 inode_dio_end(inode);
3759 /* take i_mutex locking again if we do a ovewrite dio */
3760 if (overwrite)
3761 inode_lock(inode);
3762
3763 if (ret < 0 && final_size > inode->i_size)
3764 ext4_truncate_failed_write(inode);
3765
3766 /* Handle extending of i_size after direct IO write */
3767 if (orphan) {
3768 int err;
3769
3770 /* Credits for sb + inode write */
3771 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3772 if (IS_ERR(handle)) {
3773 /*
3774 * We wrote the data but cannot extend
3775 * i_size. Bail out. In async io case, we do
3776 * not return error here because we have
3777 * already submmitted the corresponding
3778 * bio. Returning error here makes the caller
3779 * think that this IO is done and failed
3780 * resulting in race with bio's completion
3781 * handler.
3782 */
3783 if (!ret)
3784 ret = PTR_ERR(handle);
3785 if (inode->i_nlink)
3786 ext4_orphan_del(NULL, inode);
3787
3788 goto out;
3789 }
3790 if (inode->i_nlink)
3791 ext4_orphan_del(handle, inode);
3792 if (ret > 0) {
3793 loff_t end = offset + ret;
3794 if (end > inode->i_size || end > ei->i_disksize) {
3795 ext4_update_i_disksize(inode, end);
3796 if (end > inode->i_size)
3797 i_size_write(inode, end);
3798 /*
3799 * We're going to return a positive `ret'
3800 * here due to non-zero-length I/O, so there's
3801 * no way of reporting error returns from
3802 * ext4_mark_inode_dirty() to userspace. So
3803 * ignore it.
3804 */
3805 ext4_mark_inode_dirty(handle, inode);
3806 }
3807 }
3808 err = ext4_journal_stop(handle);
3809 if (ret == 0)
3810 ret = err;
3811 }
3812out:
3813 return ret;
3814}
3815
3816static ssize_t ext4_direct_IO_read(struct kiocb *iocb, struct iov_iter *iter)
3817{
3818 struct address_space *mapping = iocb->ki_filp->f_mapping;
3819 struct inode *inode = mapping->host;
3820 size_t count = iov_iter_count(iter);
3821 ssize_t ret;
3822
3823 /*
3824 * Shared inode_lock is enough for us - it protects against concurrent
3825 * writes & truncates and since we take care of writing back page cache,
3826 * we are protected against page writeback as well.
3827 */
3828 inode_lock_shared(inode);
3829 ret = filemap_write_and_wait_range(mapping, iocb->ki_pos,
3830 iocb->ki_pos + count - 1);
3831 if (ret)
3832 goto out_unlock;
3833 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
3834 iter, ext4_dio_get_block, NULL, NULL, 0);
3835out_unlock:
3836 inode_unlock_shared(inode);
3837 return ret;
3838}
3839
3840static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3841{
3842 struct file *file = iocb->ki_filp;
3843 struct inode *inode = file->f_mapping->host;
3844 size_t count = iov_iter_count(iter);
3845 loff_t offset = iocb->ki_pos;
3846 ssize_t ret;
3847
3848#ifdef CONFIG_EXT4_FS_ENCRYPTION
3849 if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode))
3850 return 0;
3851#endif
3852
3853 /*
3854 * If we are doing data journalling we don't support O_DIRECT
3855 */
3856 if (ext4_should_journal_data(inode))
3857 return 0;
3858
3859 /* Let buffer I/O handle the inline data case. */
3860 if (ext4_has_inline_data(inode))
3861 return 0;
3862
3863 trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3864 if (iov_iter_rw(iter) == READ)
3865 ret = ext4_direct_IO_read(iocb, iter);
3866 else
3867 ret = ext4_direct_IO_write(iocb, iter);
3868 trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3869 return ret;
3870}
3871
3872/*
3873 * Pages can be marked dirty completely asynchronously from ext4's journalling
3874 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3875 * much here because ->set_page_dirty is called under VFS locks. The page is
3876 * not necessarily locked.
3877 *
3878 * We cannot just dirty the page and leave attached buffers clean, because the
3879 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3880 * or jbddirty because all the journalling code will explode.
3881 *
3882 * So what we do is to mark the page "pending dirty" and next time writepage
3883 * is called, propagate that into the buffers appropriately.
3884 */
3885static int ext4_journalled_set_page_dirty(struct page *page)
3886{
3887 SetPageChecked(page);
3888 return __set_page_dirty_nobuffers(page);
3889}
3890
3891static int ext4_set_page_dirty(struct page *page)
3892{
3893 WARN_ON_ONCE(!PageLocked(page) && !PageDirty(page));
3894 WARN_ON_ONCE(!page_has_buffers(page));
3895 return __set_page_dirty_buffers(page);
3896}
3897
3898static const struct address_space_operations ext4_aops = {
3899 .readpage = ext4_readpage,
3900 .readpages = ext4_readpages,
3901 .writepage = ext4_writepage,
3902 .writepages = ext4_writepages,
3903 .write_begin = ext4_write_begin,
3904 .write_end = ext4_write_end,
3905 .set_page_dirty = ext4_set_page_dirty,
3906 .bmap = ext4_bmap,
3907 .invalidatepage = ext4_invalidatepage,
3908 .releasepage = ext4_releasepage,
3909 .direct_IO = ext4_direct_IO,
3910 .migratepage = buffer_migrate_page,
3911 .is_partially_uptodate = block_is_partially_uptodate,
3912 .error_remove_page = generic_error_remove_page,
3913};
3914
3915static const struct address_space_operations ext4_journalled_aops = {
3916 .readpage = ext4_readpage,
3917 .readpages = ext4_readpages,
3918 .writepage = ext4_writepage,
3919 .writepages = ext4_writepages,
3920 .write_begin = ext4_write_begin,
3921 .write_end = ext4_journalled_write_end,
3922 .set_page_dirty = ext4_journalled_set_page_dirty,
3923 .bmap = ext4_bmap,
3924 .invalidatepage = ext4_journalled_invalidatepage,
3925 .releasepage = ext4_releasepage,
3926 .direct_IO = ext4_direct_IO,
3927 .is_partially_uptodate = block_is_partially_uptodate,
3928 .error_remove_page = generic_error_remove_page,
3929};
3930
3931static const struct address_space_operations ext4_da_aops = {
3932 .readpage = ext4_readpage,
3933 .readpages = ext4_readpages,
3934 .writepage = ext4_writepage,
3935 .writepages = ext4_writepages,
3936 .write_begin = ext4_da_write_begin,
3937 .write_end = ext4_da_write_end,
3938 .set_page_dirty = ext4_set_page_dirty,
3939 .bmap = ext4_bmap,
3940 .invalidatepage = ext4_da_invalidatepage,
3941 .releasepage = ext4_releasepage,
3942 .direct_IO = ext4_direct_IO,
3943 .migratepage = buffer_migrate_page,
3944 .is_partially_uptodate = block_is_partially_uptodate,
3945 .error_remove_page = generic_error_remove_page,
3946};
3947
3948static const struct address_space_operations ext4_dax_aops = {
3949 .writepages = ext4_dax_writepages,
3950 .direct_IO = noop_direct_IO,
3951 .set_page_dirty = noop_set_page_dirty,
3952 .invalidatepage = noop_invalidatepage,
3953};
3954
3955void ext4_set_aops(struct inode *inode)
3956{
3957 switch (ext4_inode_journal_mode(inode)) {
3958 case EXT4_INODE_ORDERED_DATA_MODE:
3959 case EXT4_INODE_WRITEBACK_DATA_MODE:
3960 break;
3961 case EXT4_INODE_JOURNAL_DATA_MODE:
3962 inode->i_mapping->a_ops = &ext4_journalled_aops;
3963 return;
3964 default:
3965 BUG();
3966 }
3967 if (IS_DAX(inode))
3968 inode->i_mapping->a_ops = &ext4_dax_aops;
3969 else if (test_opt(inode->i_sb, DELALLOC))
3970 inode->i_mapping->a_ops = &ext4_da_aops;
3971 else
3972 inode->i_mapping->a_ops = &ext4_aops;
3973}
3974
3975static int __ext4_block_zero_page_range(handle_t *handle,
3976 struct address_space *mapping, loff_t from, loff_t length)
3977{
3978 ext4_fsblk_t index = from >> PAGE_SHIFT;
3979 unsigned offset = from & (PAGE_SIZE-1);
3980 unsigned blocksize, pos;
3981 ext4_lblk_t iblock;
3982 struct inode *inode = mapping->host;
3983 struct buffer_head *bh;
3984 struct page *page;
3985 int err = 0;
3986
3987 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3988 mapping_gfp_constraint(mapping, ~__GFP_FS));
3989 if (!page)
3990 return -ENOMEM;
3991
3992 blocksize = inode->i_sb->s_blocksize;
3993
3994 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3995
3996 if (!page_has_buffers(page))
3997 create_empty_buffers(page, blocksize, 0);
3998
3999 /* Find the buffer that contains "offset" */
4000 bh = page_buffers(page);
4001 pos = blocksize;
4002 while (offset >= pos) {
4003 bh = bh->b_this_page;
4004 iblock++;
4005 pos += blocksize;
4006 }
4007 if (buffer_freed(bh)) {
4008 BUFFER_TRACE(bh, "freed: skip");
4009 goto unlock;
4010 }
4011 if (!buffer_mapped(bh)) {
4012 BUFFER_TRACE(bh, "unmapped");
4013 ext4_get_block(inode, iblock, bh, 0);
4014 /* unmapped? It's a hole - nothing to do */
4015 if (!buffer_mapped(bh)) {
4016 BUFFER_TRACE(bh, "still unmapped");
4017 goto unlock;
4018 }
4019 }
4020
4021 /* Ok, it's mapped. Make sure it's up-to-date */
4022 if (PageUptodate(page))
4023 set_buffer_uptodate(bh);
4024
4025 if (!buffer_uptodate(bh)) {
4026 err = -EIO;
4027 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
4028 wait_on_buffer(bh);
4029 /* Uhhuh. Read error. Complain and punt. */
4030 if (!buffer_uptodate(bh))
4031 goto unlock;
4032 if (S_ISREG(inode->i_mode) &&
4033 ext4_encrypted_inode(inode)) {
4034 /* We expect the key to be set. */
4035 BUG_ON(!fscrypt_has_encryption_key(inode));
4036 BUG_ON(blocksize != PAGE_SIZE);
4037 WARN_ON_ONCE(fscrypt_decrypt_page(page->mapping->host,
4038 page, PAGE_SIZE, 0, page->index));
4039 }
4040 }
4041 if (ext4_should_journal_data(inode)) {
4042 BUFFER_TRACE(bh, "get write access");
4043 err = ext4_journal_get_write_access(handle, bh);
4044 if (err)
4045 goto unlock;
4046 }
4047 zero_user(page, offset, length);
4048 BUFFER_TRACE(bh, "zeroed end of block");
4049
4050 if (ext4_should_journal_data(inode)) {
4051 err = ext4_handle_dirty_metadata(handle, inode, bh);
4052 } else {
4053 err = 0;
4054 mark_buffer_dirty(bh);
4055 if (ext4_should_order_data(inode))
4056 err = ext4_jbd2_inode_add_write(handle, inode);
4057 }
4058
4059unlock:
4060 unlock_page(page);
4061 put_page(page);
4062 return err;
4063}
4064
4065/*
4066 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
4067 * starting from file offset 'from'. The range to be zero'd must
4068 * be contained with in one block. If the specified range exceeds
4069 * the end of the block it will be shortened to end of the block
4070 * that cooresponds to 'from'
4071 */
4072static int ext4_block_zero_page_range(handle_t *handle,
4073 struct address_space *mapping, loff_t from, loff_t length)
4074{
4075 struct inode *inode = mapping->host;
4076 unsigned offset = from & (PAGE_SIZE-1);
4077 unsigned blocksize = inode->i_sb->s_blocksize;
4078 unsigned max = blocksize - (offset & (blocksize - 1));
4079
4080 /*
4081 * correct length if it does not fall between
4082 * 'from' and the end of the block
4083 */
4084 if (length > max || length < 0)
4085 length = max;
4086
4087 if (IS_DAX(inode)) {
4088 return iomap_zero_range(inode, from, length, NULL,
4089 &ext4_iomap_ops);
4090 }
4091 return __ext4_block_zero_page_range(handle, mapping, from, length);
4092}
4093
4094/*
4095 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4096 * up to the end of the block which corresponds to `from'.
4097 * This required during truncate. We need to physically zero the tail end
4098 * of that block so it doesn't yield old data if the file is later grown.
4099 */
4100static int ext4_block_truncate_page(handle_t *handle,
4101 struct address_space *mapping, loff_t from)
4102{
4103 unsigned offset = from & (PAGE_SIZE-1);
4104 unsigned length;
4105 unsigned blocksize;
4106 struct inode *inode = mapping->host;
4107
4108 /* If we are processing an encrypted inode during orphan list handling */
4109 if (ext4_encrypted_inode(inode) && !fscrypt_has_encryption_key(inode))
4110 return 0;
4111
4112 blocksize = inode->i_sb->s_blocksize;
4113 length = blocksize - (offset & (blocksize - 1));
4114
4115 return ext4_block_zero_page_range(handle, mapping, from, length);
4116}
4117
4118int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
4119 loff_t lstart, loff_t length)
4120{
4121 struct super_block *sb = inode->i_sb;
4122 struct address_space *mapping = inode->i_mapping;
4123 unsigned partial_start, partial_end;
4124 ext4_fsblk_t start, end;
4125 loff_t byte_end = (lstart + length - 1);
4126 int err = 0;
4127
4128 partial_start = lstart & (sb->s_blocksize - 1);
4129 partial_end = byte_end & (sb->s_blocksize - 1);
4130
4131 start = lstart >> sb->s_blocksize_bits;
4132 end = byte_end >> sb->s_blocksize_bits;
4133
4134 /* Handle partial zero within the single block */
4135 if (start == end &&
4136 (partial_start || (partial_end != sb->s_blocksize - 1))) {
4137 err = ext4_block_zero_page_range(handle, mapping,
4138 lstart, length);
4139 return err;
4140 }
4141 /* Handle partial zero out on the start of the range */
4142 if (partial_start) {
4143 err = ext4_block_zero_page_range(handle, mapping,
4144 lstart, sb->s_blocksize);
4145 if (err)
4146 return err;
4147 }
4148 /* Handle partial zero out on the end of the range */
4149 if (partial_end != sb->s_blocksize - 1)
4150 err = ext4_block_zero_page_range(handle, mapping,
4151 byte_end - partial_end,
4152 partial_end + 1);
4153 return err;
4154}
4155
4156int ext4_can_truncate(struct inode *inode)
4157{
4158 if (S_ISREG(inode->i_mode))
4159 return 1;
4160 if (S_ISDIR(inode->i_mode))
4161 return 1;
4162 if (S_ISLNK(inode->i_mode))
4163 return !ext4_inode_is_fast_symlink(inode);
4164 return 0;
4165}
4166
4167/*
4168 * We have to make sure i_disksize gets properly updated before we truncate
4169 * page cache due to hole punching or zero range. Otherwise i_disksize update
4170 * can get lost as it may have been postponed to submission of writeback but
4171 * that will never happen after we truncate page cache.
4172 */
4173int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
4174 loff_t len)
4175{
4176 handle_t *handle;
4177 loff_t size = i_size_read(inode);
4178
4179 WARN_ON(!inode_is_locked(inode));
4180 if (offset > size || offset + len < size)
4181 return 0;
4182
4183 if (EXT4_I(inode)->i_disksize >= size)
4184 return 0;
4185
4186 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
4187 if (IS_ERR(handle))
4188 return PTR_ERR(handle);
4189 ext4_update_i_disksize(inode, size);
4190 ext4_mark_inode_dirty(handle, inode);
4191 ext4_journal_stop(handle);
4192
4193 return 0;
4194}
4195
4196/*
4197 * ext4_punch_hole: punches a hole in a file by releasing the blocks
4198 * associated with the given offset and length
4199 *
4200 * @inode: File inode
4201 * @offset: The offset where the hole will begin
4202 * @len: The length of the hole
4203 *
4204 * Returns: 0 on success or negative on failure
4205 */
4206
4207int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
4208{
4209 struct super_block *sb = inode->i_sb;
4210 ext4_lblk_t first_block, stop_block;
4211 struct address_space *mapping = inode->i_mapping;
4212 loff_t first_block_offset, last_block_offset;
4213 handle_t *handle;
4214 unsigned int credits;
4215 int ret = 0;
4216
4217 if (!S_ISREG(inode->i_mode))
4218 return -EOPNOTSUPP;
4219
4220 trace_ext4_punch_hole(inode, offset, length, 0);
4221
4222 /*
4223 * Write out all dirty pages to avoid race conditions
4224 * Then release them.
4225 */
4226 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4227 ret = filemap_write_and_wait_range(mapping, offset,
4228 offset + length - 1);
4229 if (ret)
4230 return ret;
4231 }
4232
4233 inode_lock(inode);
4234
4235 /* No need to punch hole beyond i_size */
4236 if (offset >= inode->i_size)
4237 goto out_mutex;
4238
4239 /*
4240 * If the hole extends beyond i_size, set the hole
4241 * to end after the page that contains i_size
4242 */
4243 if (offset + length > inode->i_size) {
4244 length = inode->i_size +
4245 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
4246 offset;
4247 }
4248
4249 if (offset & (sb->s_blocksize - 1) ||
4250 (offset + length) & (sb->s_blocksize - 1)) {
4251 /*
4252 * Attach jinode to inode for jbd2 if we do any zeroing of
4253 * partial block
4254 */
4255 ret = ext4_inode_attach_jinode(inode);
4256 if (ret < 0)
4257 goto out_mutex;
4258
4259 }
4260
4261 /* Wait all existing dio workers, newcomers will block on i_mutex */
4262 inode_dio_wait(inode);
4263
4264 /*
4265 * Prevent page faults from reinstantiating pages we have released from
4266 * page cache.
4267 */
4268 down_write(&EXT4_I(inode)->i_mmap_sem);
4269 first_block_offset = round_up(offset, sb->s_blocksize);
4270 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4271
4272 /* Now release the pages and zero block aligned part of pages*/
4273 if (last_block_offset > first_block_offset) {
4274 ret = ext4_update_disksize_before_punch(inode, offset, length);
4275 if (ret)
4276 goto out_dio;
4277 truncate_pagecache_range(inode, first_block_offset,
4278 last_block_offset);
4279 }
4280
4281 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4282 credits = ext4_writepage_trans_blocks(inode);
4283 else
4284 credits = ext4_blocks_for_truncate(inode);
4285 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4286 if (IS_ERR(handle)) {
4287 ret = PTR_ERR(handle);
4288 ext4_std_error(sb, ret);
4289 goto out_dio;
4290 }
4291
4292 ret = ext4_zero_partial_blocks(handle, inode, offset,
4293 length);
4294 if (ret)
4295 goto out_stop;
4296
4297 first_block = (offset + sb->s_blocksize - 1) >>
4298 EXT4_BLOCK_SIZE_BITS(sb);
4299 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4300
4301 /* If there are no blocks to remove, return now */
4302 if (first_block >= stop_block)
4303 goto out_stop;
4304
4305 down_write(&EXT4_I(inode)->i_data_sem);
4306 ext4_discard_preallocations(inode);
4307
4308 ret = ext4_es_remove_extent(inode, first_block,
4309 stop_block - first_block);
4310 if (ret) {
4311 up_write(&EXT4_I(inode)->i_data_sem);
4312 goto out_stop;
4313 }
4314
4315 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4316 ret = ext4_ext_remove_space(inode, first_block,
4317 stop_block - 1);
4318 else
4319 ret = ext4_ind_remove_space(handle, inode, first_block,
4320 stop_block);
4321
4322 up_write(&EXT4_I(inode)->i_data_sem);
4323 if (IS_SYNC(inode))
4324 ext4_handle_sync(handle);
4325
4326 inode->i_mtime = inode->i_ctime = current_time(inode);
4327 ext4_mark_inode_dirty(handle, inode);
4328 if (ret >= 0)
4329 ext4_update_inode_fsync_trans(handle, inode, 1);
4330out_stop:
4331 ext4_journal_stop(handle);
4332out_dio:
4333 up_write(&EXT4_I(inode)->i_mmap_sem);
4334out_mutex:
4335 inode_unlock(inode);
4336 return ret;
4337}
4338
4339int ext4_inode_attach_jinode(struct inode *inode)
4340{
4341 struct ext4_inode_info *ei = EXT4_I(inode);
4342 struct jbd2_inode *jinode;
4343
4344 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4345 return 0;
4346
4347 jinode = jbd2_alloc_inode(GFP_KERNEL);
4348 spin_lock(&inode->i_lock);
4349 if (!ei->jinode) {
4350 if (!jinode) {
4351 spin_unlock(&inode->i_lock);
4352 return -ENOMEM;
4353 }
4354 ei->jinode = jinode;
4355 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4356 jinode = NULL;
4357 }
4358 spin_unlock(&inode->i_lock);
4359 if (unlikely(jinode != NULL))
4360 jbd2_free_inode(jinode);
4361 return 0;
4362}
4363
4364/*
4365 * ext4_truncate()
4366 *
4367 * We block out ext4_get_block() block instantiations across the entire
4368 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4369 * simultaneously on behalf of the same inode.
4370 *
4371 * As we work through the truncate and commit bits of it to the journal there
4372 * is one core, guiding principle: the file's tree must always be consistent on
4373 * disk. We must be able to restart the truncate after a crash.
4374 *
4375 * The file's tree may be transiently inconsistent in memory (although it
4376 * probably isn't), but whenever we close off and commit a journal transaction,
4377 * the contents of (the filesystem + the journal) must be consistent and
4378 * restartable. It's pretty simple, really: bottom up, right to left (although
4379 * left-to-right works OK too).
4380 *
4381 * Note that at recovery time, journal replay occurs *before* the restart of
4382 * truncate against the orphan inode list.
4383 *
4384 * The committed inode has the new, desired i_size (which is the same as
4385 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4386 * that this inode's truncate did not complete and it will again call
4387 * ext4_truncate() to have another go. So there will be instantiated blocks
4388 * to the right of the truncation point in a crashed ext4 filesystem. But
4389 * that's fine - as long as they are linked from the inode, the post-crash
4390 * ext4_truncate() run will find them and release them.
4391 */
4392int ext4_truncate(struct inode *inode)
4393{
4394 struct ext4_inode_info *ei = EXT4_I(inode);
4395 unsigned int credits;
4396 int err = 0;
4397 handle_t *handle;
4398 struct address_space *mapping = inode->i_mapping;
4399
4400 /*
4401 * There is a possibility that we're either freeing the inode
4402 * or it's a completely new inode. In those cases we might not
4403 * have i_mutex locked because it's not necessary.
4404 */
4405 if (!(inode->i_state & (I_NEW|I_FREEING)))
4406 WARN_ON(!inode_is_locked(inode));
4407 trace_ext4_truncate_enter(inode);
4408
4409 if (!ext4_can_truncate(inode))
4410 return 0;
4411
4412 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4413
4414 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4415 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4416
4417 if (ext4_has_inline_data(inode)) {
4418 int has_inline = 1;
4419
4420 err = ext4_inline_data_truncate(inode, &has_inline);
4421 if (err)
4422 return err;
4423 if (has_inline)
4424 return 0;
4425 }
4426
4427 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4428 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4429 if (ext4_inode_attach_jinode(inode) < 0)
4430 return 0;
4431 }
4432
4433 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4434 credits = ext4_writepage_trans_blocks(inode);
4435 else
4436 credits = ext4_blocks_for_truncate(inode);
4437
4438 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4439 if (IS_ERR(handle))
4440 return PTR_ERR(handle);
4441
4442 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4443 ext4_block_truncate_page(handle, mapping, inode->i_size);
4444
4445 /*
4446 * We add the inode to the orphan list, so that if this
4447 * truncate spans multiple transactions, and we crash, we will
4448 * resume the truncate when the filesystem recovers. It also
4449 * marks the inode dirty, to catch the new size.
4450 *
4451 * Implication: the file must always be in a sane, consistent
4452 * truncatable state while each transaction commits.
4453 */
4454 err = ext4_orphan_add(handle, inode);
4455 if (err)
4456 goto out_stop;
4457
4458 down_write(&EXT4_I(inode)->i_data_sem);
4459
4460 ext4_discard_preallocations(inode);
4461
4462 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4463 err = ext4_ext_truncate(handle, inode);
4464 else
4465 ext4_ind_truncate(handle, inode);
4466
4467 up_write(&ei->i_data_sem);
4468 if (err)
4469 goto out_stop;
4470
4471 if (IS_SYNC(inode))
4472 ext4_handle_sync(handle);
4473
4474out_stop:
4475 /*
4476 * If this was a simple ftruncate() and the file will remain alive,
4477 * then we need to clear up the orphan record which we created above.
4478 * However, if this was a real unlink then we were called by
4479 * ext4_evict_inode(), and we allow that function to clean up the
4480 * orphan info for us.
4481 */
4482 if (inode->i_nlink)
4483 ext4_orphan_del(handle, inode);
4484
4485 inode->i_mtime = inode->i_ctime = current_time(inode);
4486 ext4_mark_inode_dirty(handle, inode);
4487 ext4_journal_stop(handle);
4488
4489 trace_ext4_truncate_exit(inode);
4490 return err;
4491}
4492
4493/*
4494 * ext4_get_inode_loc returns with an extra refcount against the inode's
4495 * underlying buffer_head on success. If 'in_mem' is true, we have all
4496 * data in memory that is needed to recreate the on-disk version of this
4497 * inode.
4498 */
4499static int __ext4_get_inode_loc(struct inode *inode,
4500 struct ext4_iloc *iloc, int in_mem)
4501{
4502 struct ext4_group_desc *gdp;
4503 struct buffer_head *bh;
4504 struct super_block *sb = inode->i_sb;
4505 ext4_fsblk_t block;
4506 int inodes_per_block, inode_offset;
4507
4508 iloc->bh = NULL;
4509 if (!ext4_valid_inum(sb, inode->i_ino))
4510 return -EFSCORRUPTED;
4511
4512 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4513 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4514 if (!gdp)
4515 return -EIO;
4516
4517 /*
4518 * Figure out the offset within the block group inode table
4519 */
4520 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4521 inode_offset = ((inode->i_ino - 1) %
4522 EXT4_INODES_PER_GROUP(sb));
4523 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4524 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4525
4526 bh = sb_getblk(sb, block);
4527 if (unlikely(!bh))
4528 return -ENOMEM;
4529 if (!buffer_uptodate(bh)) {
4530 lock_buffer(bh);
4531
4532 /*
4533 * If the buffer has the write error flag, we have failed
4534 * to write out another inode in the same block. In this
4535 * case, we don't have to read the block because we may
4536 * read the old inode data successfully.
4537 */
4538 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4539 set_buffer_uptodate(bh);
4540
4541 if (buffer_uptodate(bh)) {
4542 /* someone brought it uptodate while we waited */
4543 unlock_buffer(bh);
4544 goto has_buffer;
4545 }
4546
4547 /*
4548 * If we have all information of the inode in memory and this
4549 * is the only valid inode in the block, we need not read the
4550 * block.
4551 */
4552 if (in_mem) {
4553 struct buffer_head *bitmap_bh;
4554 int i, start;
4555
4556 start = inode_offset & ~(inodes_per_block - 1);
4557
4558 /* Is the inode bitmap in cache? */
4559 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4560 if (unlikely(!bitmap_bh))
4561 goto make_io;
4562
4563 /*
4564 * If the inode bitmap isn't in cache then the
4565 * optimisation may end up performing two reads instead
4566 * of one, so skip it.
4567 */
4568 if (!buffer_uptodate(bitmap_bh)) {
4569 brelse(bitmap_bh);
4570 goto make_io;
4571 }
4572 for (i = start; i < start + inodes_per_block; i++) {
4573 if (i == inode_offset)
4574 continue;
4575 if (ext4_test_bit(i, bitmap_bh->b_data))
4576 break;
4577 }
4578 brelse(bitmap_bh);
4579 if (i == start + inodes_per_block) {
4580 /* all other inodes are free, so skip I/O */
4581 memset(bh->b_data, 0, bh->b_size);
4582 set_buffer_uptodate(bh);
4583 unlock_buffer(bh);
4584 goto has_buffer;
4585 }
4586 }
4587
4588make_io:
4589 /*
4590 * If we need to do any I/O, try to pre-readahead extra
4591 * blocks from the inode table.
4592 */
4593 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4594 ext4_fsblk_t b, end, table;
4595 unsigned num;
4596 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4597
4598 table = ext4_inode_table(sb, gdp);
4599 /* s_inode_readahead_blks is always a power of 2 */
4600 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4601 if (table > b)
4602 b = table;
4603 end = b + ra_blks;
4604 num = EXT4_INODES_PER_GROUP(sb);
4605 if (ext4_has_group_desc_csum(sb))
4606 num -= ext4_itable_unused_count(sb, gdp);
4607 table += num / inodes_per_block;
4608 if (end > table)
4609 end = table;
4610 while (b <= end)
4611 sb_breadahead(sb, b++);
4612 }
4613
4614 /*
4615 * There are other valid inodes in the buffer, this inode
4616 * has in-inode xattrs, or we don't have this inode in memory.
4617 * Read the block from disk.
4618 */
4619 trace_ext4_load_inode(inode);
4620 get_bh(bh);
4621 bh->b_end_io = end_buffer_read_sync;
4622 submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
4623 wait_on_buffer(bh);
4624 if (!buffer_uptodate(bh)) {
4625 EXT4_ERROR_INODE_BLOCK(inode, block,
4626 "unable to read itable block");
4627 brelse(bh);
4628 return -EIO;
4629 }
4630 }
4631has_buffer:
4632 iloc->bh = bh;
4633 return 0;
4634}
4635
4636int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4637{
4638 /* We have all inode data except xattrs in memory here. */
4639 return __ext4_get_inode_loc(inode, iloc,
4640 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4641}
4642
4643static bool ext4_should_use_dax(struct inode *inode)
4644{
4645 if (!test_opt(inode->i_sb, DAX))
4646 return false;
4647 if (!S_ISREG(inode->i_mode))
4648 return false;
4649 if (ext4_should_journal_data(inode))
4650 return false;
4651 if (ext4_has_inline_data(inode))
4652 return false;
4653 if (ext4_encrypted_inode(inode))
4654 return false;
4655 return true;
4656}
4657
4658void ext4_set_inode_flags(struct inode *inode)
4659{
4660 unsigned int flags = EXT4_I(inode)->i_flags;
4661 unsigned int new_fl = 0;
4662
4663 if (flags & EXT4_SYNC_FL)
4664 new_fl |= S_SYNC;
4665 if (flags & EXT4_APPEND_FL)
4666 new_fl |= S_APPEND;
4667 if (flags & EXT4_IMMUTABLE_FL)
4668 new_fl |= S_IMMUTABLE;
4669 if (flags & EXT4_NOATIME_FL)
4670 new_fl |= S_NOATIME;
4671 if (flags & EXT4_DIRSYNC_FL)
4672 new_fl |= S_DIRSYNC;
4673 if (ext4_should_use_dax(inode))
4674 new_fl |= S_DAX;
4675 if (flags & EXT4_ENCRYPT_FL)
4676 new_fl |= S_ENCRYPTED;
4677 inode_set_flags(inode, new_fl,
4678 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4679 S_ENCRYPTED);
4680}
4681
4682static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4683 struct ext4_inode_info *ei)
4684{
4685 blkcnt_t i_blocks ;
4686 struct inode *inode = &(ei->vfs_inode);
4687 struct super_block *sb = inode->i_sb;
4688
4689 if (ext4_has_feature_huge_file(sb)) {
4690 /* we are using combined 48 bit field */
4691 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4692 le32_to_cpu(raw_inode->i_blocks_lo);
4693 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4694 /* i_blocks represent file system block size */
4695 return i_blocks << (inode->i_blkbits - 9);
4696 } else {
4697 return i_blocks;
4698 }
4699 } else {
4700 return le32_to_cpu(raw_inode->i_blocks_lo);
4701 }
4702}
4703
4704static inline void ext4_iget_extra_inode(struct inode *inode,
4705 struct ext4_inode *raw_inode,
4706 struct ext4_inode_info *ei)
4707{
4708 __le32 *magic = (void *)raw_inode +
4709 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4710 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize + sizeof(__le32) <=
4711 EXT4_INODE_SIZE(inode->i_sb) &&
4712 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4713 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4714 ext4_find_inline_data_nolock(inode);
4715 } else
4716 EXT4_I(inode)->i_inline_off = 0;
4717}
4718
4719int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4720{
4721 if (!ext4_has_feature_project(inode->i_sb))
4722 return -EOPNOTSUPP;
4723 *projid = EXT4_I(inode)->i_projid;
4724 return 0;
4725}
4726
4727struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4728{
4729 struct ext4_iloc iloc;
4730 struct ext4_inode *raw_inode;
4731 struct ext4_inode_info *ei;
4732 struct inode *inode;
4733 journal_t *journal = EXT4_SB(sb)->s_journal;
4734 long ret;
4735 loff_t size;
4736 int block;
4737 uid_t i_uid;
4738 gid_t i_gid;
4739 projid_t i_projid;
4740
4741 inode = iget_locked(sb, ino);
4742 if (!inode)
4743 return ERR_PTR(-ENOMEM);
4744 if (!(inode->i_state & I_NEW))
4745 return inode;
4746
4747 ei = EXT4_I(inode);
4748 iloc.bh = NULL;
4749
4750 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4751 if (ret < 0)
4752 goto bad_inode;
4753 raw_inode = ext4_raw_inode(&iloc);
4754
4755 if ((ino == EXT4_ROOT_INO) && (raw_inode->i_links_count == 0)) {
4756 EXT4_ERROR_INODE(inode, "root inode unallocated");
4757 ret = -EFSCORRUPTED;
4758 goto bad_inode;
4759 }
4760
4761 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4762 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4763 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4764 EXT4_INODE_SIZE(inode->i_sb) ||
4765 (ei->i_extra_isize & 3)) {
4766 EXT4_ERROR_INODE(inode,
4767 "bad extra_isize %u (inode size %u)",
4768 ei->i_extra_isize,
4769 EXT4_INODE_SIZE(inode->i_sb));
4770 ret = -EFSCORRUPTED;
4771 goto bad_inode;
4772 }
4773 } else
4774 ei->i_extra_isize = 0;
4775
4776 /* Precompute checksum seed for inode metadata */
4777 if (ext4_has_metadata_csum(sb)) {
4778 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4779 __u32 csum;
4780 __le32 inum = cpu_to_le32(inode->i_ino);
4781 __le32 gen = raw_inode->i_generation;
4782 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4783 sizeof(inum));
4784 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4785 sizeof(gen));
4786 }
4787
4788 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4789 EXT4_ERROR_INODE(inode, "checksum invalid");
4790 ret = -EFSBADCRC;
4791 goto bad_inode;
4792 }
4793
4794 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4795 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4796 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4797 if (ext4_has_feature_project(sb) &&
4798 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4799 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4800 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4801 else
4802 i_projid = EXT4_DEF_PROJID;
4803
4804 if (!(test_opt(inode->i_sb, NO_UID32))) {
4805 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4806 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4807 }
4808 i_uid_write(inode, i_uid);
4809 i_gid_write(inode, i_gid);
4810 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4811 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4812
4813 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4814 ei->i_inline_off = 0;
4815 ei->i_dir_start_lookup = 0;
4816 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4817 /* We now have enough fields to check if the inode was active or not.
4818 * This is needed because nfsd might try to access dead inodes
4819 * the test is that same one that e2fsck uses
4820 * NeilBrown 1999oct15
4821 */
4822 if (inode->i_nlink == 0) {
4823 if ((inode->i_mode == 0 ||
4824 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4825 ino != EXT4_BOOT_LOADER_INO) {
4826 /* this inode is deleted */
4827 ret = -ESTALE;
4828 goto bad_inode;
4829 }
4830 /* The only unlinked inodes we let through here have
4831 * valid i_mode and are being read by the orphan
4832 * recovery code: that's fine, we're about to complete
4833 * the process of deleting those.
4834 * OR it is the EXT4_BOOT_LOADER_INO which is
4835 * not initialized on a new filesystem. */
4836 }
4837 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4838 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4839 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4840 if (ext4_has_feature_64bit(sb))
4841 ei->i_file_acl |=
4842 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4843 inode->i_size = ext4_isize(sb, raw_inode);
4844 if ((size = i_size_read(inode)) < 0) {
4845 EXT4_ERROR_INODE(inode, "bad i_size value: %lld", size);
4846 ret = -EFSCORRUPTED;
4847 goto bad_inode;
4848 }
4849 ei->i_disksize = inode->i_size;
4850#ifdef CONFIG_QUOTA
4851 ei->i_reserved_quota = 0;
4852#endif
4853 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4854 ei->i_block_group = iloc.block_group;
4855 ei->i_last_alloc_group = ~0;
4856 /*
4857 * NOTE! The in-memory inode i_data array is in little-endian order
4858 * even on big-endian machines: we do NOT byteswap the block numbers!
4859 */
4860 for (block = 0; block < EXT4_N_BLOCKS; block++)
4861 ei->i_data[block] = raw_inode->i_block[block];
4862 INIT_LIST_HEAD(&ei->i_orphan);
4863
4864 /*
4865 * Set transaction id's of transactions that have to be committed
4866 * to finish f[data]sync. We set them to currently running transaction
4867 * as we cannot be sure that the inode or some of its metadata isn't
4868 * part of the transaction - the inode could have been reclaimed and
4869 * now it is reread from disk.
4870 */
4871 if (journal) {
4872 transaction_t *transaction;
4873 tid_t tid;
4874
4875 read_lock(&journal->j_state_lock);
4876 if (journal->j_running_transaction)
4877 transaction = journal->j_running_transaction;
4878 else
4879 transaction = journal->j_committing_transaction;
4880 if (transaction)
4881 tid = transaction->t_tid;
4882 else
4883 tid = journal->j_commit_sequence;
4884 read_unlock(&journal->j_state_lock);
4885 ei->i_sync_tid = tid;
4886 ei->i_datasync_tid = tid;
4887 }
4888
4889 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4890 if (ei->i_extra_isize == 0) {
4891 /* The extra space is currently unused. Use it. */
4892 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4893 ei->i_extra_isize = sizeof(struct ext4_inode) -
4894 EXT4_GOOD_OLD_INODE_SIZE;
4895 } else {
4896 ext4_iget_extra_inode(inode, raw_inode, ei);
4897 }
4898 }
4899
4900 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4901 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4902 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4903 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4904
4905 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4906 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4907
4908 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4909 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4910 ivers |=
4911 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4912 }
4913 inode_set_iversion_queried(inode, ivers);
4914 }
4915
4916 ret = 0;
4917 if (ei->i_file_acl &&
4918 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4919 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4920 ei->i_file_acl);
4921 ret = -EFSCORRUPTED;
4922 goto bad_inode;
4923 } else if (!ext4_has_inline_data(inode)) {
4924 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4925 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4926 (S_ISLNK(inode->i_mode) &&
4927 !ext4_inode_is_fast_symlink(inode))))
4928 /* Validate extent which is part of inode */
4929 ret = ext4_ext_check_inode(inode);
4930 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4931 (S_ISLNK(inode->i_mode) &&
4932 !ext4_inode_is_fast_symlink(inode))) {
4933 /* Validate block references which are part of inode */
4934 ret = ext4_ind_check_inode(inode);
4935 }
4936 }
4937 if (ret)
4938 goto bad_inode;
4939
4940 if (S_ISREG(inode->i_mode)) {
4941 inode->i_op = &ext4_file_inode_operations;
4942 inode->i_fop = &ext4_file_operations;
4943 ext4_set_aops(inode);
4944 } else if (S_ISDIR(inode->i_mode)) {
4945 inode->i_op = &ext4_dir_inode_operations;
4946 inode->i_fop = &ext4_dir_operations;
4947 } else if (S_ISLNK(inode->i_mode)) {
4948 if (ext4_encrypted_inode(inode)) {
4949 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4950 ext4_set_aops(inode);
4951 } else if (ext4_inode_is_fast_symlink(inode)) {
4952 inode->i_link = (char *)ei->i_data;
4953 inode->i_op = &ext4_fast_symlink_inode_operations;
4954 nd_terminate_link(ei->i_data, inode->i_size,
4955 sizeof(ei->i_data) - 1);
4956 } else {
4957 inode->i_op = &ext4_symlink_inode_operations;
4958 ext4_set_aops(inode);
4959 }
4960 inode_nohighmem(inode);
4961 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4962 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4963 inode->i_op = &ext4_special_inode_operations;
4964 if (raw_inode->i_block[0])
4965 init_special_inode(inode, inode->i_mode,
4966 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4967 else
4968 init_special_inode(inode, inode->i_mode,
4969 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4970 } else if (ino == EXT4_BOOT_LOADER_INO) {
4971 make_bad_inode(inode);
4972 } else {
4973 ret = -EFSCORRUPTED;
4974 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
4975 goto bad_inode;
4976 }
4977 brelse(iloc.bh);
4978 ext4_set_inode_flags(inode);
4979
4980 unlock_new_inode(inode);
4981 return inode;
4982
4983bad_inode:
4984 brelse(iloc.bh);
4985 iget_failed(inode);
4986 return ERR_PTR(ret);
4987}
4988
4989struct inode *ext4_iget_normal(struct super_block *sb, unsigned long ino)
4990{
4991 if (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)
4992 return ERR_PTR(-EFSCORRUPTED);
4993 return ext4_iget(sb, ino);
4994}
4995
4996static int ext4_inode_blocks_set(handle_t *handle,
4997 struct ext4_inode *raw_inode,
4998 struct ext4_inode_info *ei)
4999{
5000 struct inode *inode = &(ei->vfs_inode);
5001 u64 i_blocks = inode->i_blocks;
5002 struct super_block *sb = inode->i_sb;
5003
5004 if (i_blocks <= ~0U) {
5005 /*
5006 * i_blocks can be represented in a 32 bit variable
5007 * as multiple of 512 bytes
5008 */
5009 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5010 raw_inode->i_blocks_high = 0;
5011 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5012 return 0;
5013 }
5014 if (!ext4_has_feature_huge_file(sb))
5015 return -EFBIG;
5016
5017 if (i_blocks <= 0xffffffffffffULL) {
5018 /*
5019 * i_blocks can be represented in a 48 bit variable
5020 * as multiple of 512 bytes
5021 */
5022 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5023 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5024 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5025 } else {
5026 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5027 /* i_block is stored in file system block size */
5028 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5029 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5030 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5031 }
5032 return 0;
5033}
5034
5035struct other_inode {
5036 unsigned long orig_ino;
5037 struct ext4_inode *raw_inode;
5038};
5039
5040static int other_inode_match(struct inode * inode, unsigned long ino,
5041 void *data)
5042{
5043 struct other_inode *oi = (struct other_inode *) data;
5044
5045 if ((inode->i_ino != ino) ||
5046 (inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5047 I_DIRTY_INODE)) ||
5048 ((inode->i_state & I_DIRTY_TIME) == 0))
5049 return 0;
5050 spin_lock(&inode->i_lock);
5051 if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5052 I_DIRTY_INODE)) == 0) &&
5053 (inode->i_state & I_DIRTY_TIME)) {
5054 struct ext4_inode_info *ei = EXT4_I(inode);
5055
5056 inode->i_state &= ~(I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED);
5057 spin_unlock(&inode->i_lock);
5058
5059 spin_lock(&ei->i_raw_lock);
5060 EXT4_INODE_SET_XTIME(i_ctime, inode, oi->raw_inode);
5061 EXT4_INODE_SET_XTIME(i_mtime, inode, oi->raw_inode);
5062 EXT4_INODE_SET_XTIME(i_atime, inode, oi->raw_inode);
5063 ext4_inode_csum_set(inode, oi->raw_inode, ei);
5064 spin_unlock(&ei->i_raw_lock);
5065 trace_ext4_other_inode_update_time(inode, oi->orig_ino);
5066 return -1;
5067 }
5068 spin_unlock(&inode->i_lock);
5069 return -1;
5070}
5071
5072/*
5073 * Opportunistically update the other time fields for other inodes in
5074 * the same inode table block.
5075 */
5076static void ext4_update_other_inodes_time(struct super_block *sb,
5077 unsigned long orig_ino, char *buf)
5078{
5079 struct other_inode oi;
5080 unsigned long ino;
5081 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5082 int inode_size = EXT4_INODE_SIZE(sb);
5083
5084 oi.orig_ino = orig_ino;
5085 /*
5086 * Calculate the first inode in the inode table block. Inode
5087 * numbers are one-based. That is, the first inode in a block
5088 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5089 */
5090 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5091 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5092 if (ino == orig_ino)
5093 continue;
5094 oi.raw_inode = (struct ext4_inode *) buf;
5095 (void) find_inode_nowait(sb, ino, other_inode_match, &oi);
5096 }
5097}
5098
5099/*
5100 * Post the struct inode info into an on-disk inode location in the
5101 * buffer-cache. This gobbles the caller's reference to the
5102 * buffer_head in the inode location struct.
5103 *
5104 * The caller must have write access to iloc->bh.
5105 */
5106static int ext4_do_update_inode(handle_t *handle,
5107 struct inode *inode,
5108 struct ext4_iloc *iloc)
5109{
5110 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5111 struct ext4_inode_info *ei = EXT4_I(inode);
5112 struct buffer_head *bh = iloc->bh;
5113 struct super_block *sb = inode->i_sb;
5114 int err = 0, rc, block;
5115 int need_datasync = 0, set_large_file = 0;
5116 uid_t i_uid;
5117 gid_t i_gid;
5118 projid_t i_projid;
5119
5120 spin_lock(&ei->i_raw_lock);
5121
5122 /* For fields not tracked in the in-memory inode,
5123 * initialise them to zero for new inodes. */
5124 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5125 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5126
5127 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5128 i_uid = i_uid_read(inode);
5129 i_gid = i_gid_read(inode);
5130 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
5131 if (!(test_opt(inode->i_sb, NO_UID32))) {
5132 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
5133 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
5134/*
5135 * Fix up interoperability with old kernels. Otherwise, old inodes get
5136 * re-used with the upper 16 bits of the uid/gid intact
5137 */
5138 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
5139 raw_inode->i_uid_high = 0;
5140 raw_inode->i_gid_high = 0;
5141 } else {
5142 raw_inode->i_uid_high =
5143 cpu_to_le16(high_16_bits(i_uid));
5144 raw_inode->i_gid_high =
5145 cpu_to_le16(high_16_bits(i_gid));
5146 }
5147 } else {
5148 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
5149 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
5150 raw_inode->i_uid_high = 0;
5151 raw_inode->i_gid_high = 0;
5152 }
5153 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5154
5155 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5156 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5157 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5158 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5159
5160 err = ext4_inode_blocks_set(handle, raw_inode, ei);
5161 if (err) {
5162 spin_unlock(&ei->i_raw_lock);
5163 goto out_brelse;
5164 }
5165 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5166 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
5167 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
5168 raw_inode->i_file_acl_high =
5169 cpu_to_le16(ei->i_file_acl >> 32);
5170 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5171 if (ei->i_disksize != ext4_isize(inode->i_sb, raw_inode)) {
5172 ext4_isize_set(raw_inode, ei->i_disksize);
5173 need_datasync = 1;
5174 }
5175 if (ei->i_disksize > 0x7fffffffULL) {
5176 if (!ext4_has_feature_large_file(sb) ||
5177 EXT4_SB(sb)->s_es->s_rev_level ==
5178 cpu_to_le32(EXT4_GOOD_OLD_REV))
5179 set_large_file = 1;
5180 }
5181 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5182 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5183 if (old_valid_dev(inode->i_rdev)) {
5184 raw_inode->i_block[0] =
5185 cpu_to_le32(old_encode_dev(inode->i_rdev));
5186 raw_inode->i_block[1] = 0;
5187 } else {
5188 raw_inode->i_block[0] = 0;
5189 raw_inode->i_block[1] =
5190 cpu_to_le32(new_encode_dev(inode->i_rdev));
5191 raw_inode->i_block[2] = 0;
5192 }
5193 } else if (!ext4_has_inline_data(inode)) {
5194 for (block = 0; block < EXT4_N_BLOCKS; block++)
5195 raw_inode->i_block[block] = ei->i_data[block];
5196 }
5197
5198 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5199 u64 ivers = inode_peek_iversion(inode);
5200
5201 raw_inode->i_disk_version = cpu_to_le32(ivers);
5202 if (ei->i_extra_isize) {
5203 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5204 raw_inode->i_version_hi =
5205 cpu_to_le32(ivers >> 32);
5206 raw_inode->i_extra_isize =
5207 cpu_to_le16(ei->i_extra_isize);
5208 }
5209 }
5210
5211 BUG_ON(!ext4_has_feature_project(inode->i_sb) &&
5212 i_projid != EXT4_DEF_PROJID);
5213
5214 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
5215 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
5216 raw_inode->i_projid = cpu_to_le32(i_projid);
5217
5218 ext4_inode_csum_set(inode, raw_inode, ei);
5219 spin_unlock(&ei->i_raw_lock);
5220 if (inode->i_sb->s_flags & SB_LAZYTIME)
5221 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5222 bh->b_data);
5223
5224 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5225 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5226 if (!err)
5227 err = rc;
5228 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5229 if (set_large_file) {
5230 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5231 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
5232 if (err)
5233 goto out_brelse;
5234 ext4_update_dynamic_rev(sb);
5235 ext4_set_feature_large_file(sb);
5236 ext4_handle_sync(handle);
5237 err = ext4_handle_dirty_super(handle, sb);
5238 }
5239 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5240out_brelse:
5241 brelse(bh);
5242 ext4_std_error(inode->i_sb, err);
5243 return err;
5244}
5245
5246/*
5247 * ext4_write_inode()
5248 *
5249 * We are called from a few places:
5250 *
5251 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5252 * Here, there will be no transaction running. We wait for any running
5253 * transaction to commit.
5254 *
5255 * - Within flush work (sys_sync(), kupdate and such).
5256 * We wait on commit, if told to.
5257 *
5258 * - Within iput_final() -> write_inode_now()
5259 * We wait on commit, if told to.
5260 *
5261 * In all cases it is actually safe for us to return without doing anything,
5262 * because the inode has been copied into a raw inode buffer in
5263 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5264 * writeback.
5265 *
5266 * Note that we are absolutely dependent upon all inode dirtiers doing the
5267 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5268 * which we are interested.
5269 *
5270 * It would be a bug for them to not do this. The code:
5271 *
5272 * mark_inode_dirty(inode)
5273 * stuff();
5274 * inode->i_size = expr;
5275 *
5276 * is in error because write_inode() could occur while `stuff()' is running,
5277 * and the new i_size will be lost. Plus the inode will no longer be on the
5278 * superblock's dirty inode list.
5279 */
5280int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5281{
5282 int err;
5283
5284 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5285 return 0;
5286
5287 if (EXT4_SB(inode->i_sb)->s_journal) {
5288 if (ext4_journal_current_handle()) {
5289 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5290 dump_stack();
5291 return -EIO;
5292 }
5293
5294 /*
5295 * No need to force transaction in WB_SYNC_NONE mode. Also
5296 * ext4_sync_fs() will force the commit after everything is
5297 * written.
5298 */
5299 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5300 return 0;
5301
5302 err = ext4_force_commit(inode->i_sb);
5303 } else {
5304 struct ext4_iloc iloc;
5305
5306 err = __ext4_get_inode_loc(inode, &iloc, 0);
5307 if (err)
5308 return err;
5309 /*
5310 * sync(2) will flush the whole buffer cache. No need to do
5311 * it here separately for each inode.
5312 */
5313 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5314 sync_dirty_buffer(iloc.bh);
5315 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5316 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5317 "IO error syncing inode");
5318 err = -EIO;
5319 }
5320 brelse(iloc.bh);
5321 }
5322 return err;
5323}
5324
5325/*
5326 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
5327 * buffers that are attached to a page stradding i_size and are undergoing
5328 * commit. In that case we have to wait for commit to finish and try again.
5329 */
5330static void ext4_wait_for_tail_page_commit(struct inode *inode)
5331{
5332 struct page *page;
5333 unsigned offset;
5334 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5335 tid_t commit_tid = 0;
5336 int ret;
5337
5338 offset = inode->i_size & (PAGE_SIZE - 1);
5339 /*
5340 * All buffers in the last page remain valid? Then there's nothing to
5341 * do. We do the check mainly to optimize the common PAGE_SIZE ==
5342 * blocksize case
5343 */
5344 if (offset > PAGE_SIZE - i_blocksize(inode))
5345 return;
5346 while (1) {
5347 page = find_lock_page(inode->i_mapping,
5348 inode->i_size >> PAGE_SHIFT);
5349 if (!page)
5350 return;
5351 ret = __ext4_journalled_invalidatepage(page, offset,
5352 PAGE_SIZE - offset);
5353 unlock_page(page);
5354 put_page(page);
5355 if (ret != -EBUSY)
5356 return;
5357 commit_tid = 0;
5358 read_lock(&journal->j_state_lock);
5359 if (journal->j_committing_transaction)
5360 commit_tid = journal->j_committing_transaction->t_tid;
5361 read_unlock(&journal->j_state_lock);
5362 if (commit_tid)
5363 jbd2_log_wait_commit(journal, commit_tid);
5364 }
5365}
5366
5367/*
5368 * ext4_setattr()
5369 *
5370 * Called from notify_change.
5371 *
5372 * We want to trap VFS attempts to truncate the file as soon as
5373 * possible. In particular, we want to make sure that when the VFS
5374 * shrinks i_size, we put the inode on the orphan list and modify
5375 * i_disksize immediately, so that during the subsequent flushing of
5376 * dirty pages and freeing of disk blocks, we can guarantee that any
5377 * commit will leave the blocks being flushed in an unused state on
5378 * disk. (On recovery, the inode will get truncated and the blocks will
5379 * be freed, so we have a strong guarantee that no future commit will
5380 * leave these blocks visible to the user.)
5381 *
5382 * Another thing we have to assure is that if we are in ordered mode
5383 * and inode is still attached to the committing transaction, we must
5384 * we start writeout of all the dirty pages which are being truncated.
5385 * This way we are sure that all the data written in the previous
5386 * transaction are already on disk (truncate waits for pages under
5387 * writeback).
5388 *
5389 * Called with inode->i_mutex down.
5390 */
5391int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5392{
5393 struct inode *inode = d_inode(dentry);
5394 int error, rc = 0;
5395 int orphan = 0;
5396 const unsigned int ia_valid = attr->ia_valid;
5397
5398 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5399 return -EIO;
5400
5401 error = setattr_prepare(dentry, attr);
5402 if (error)
5403 return error;
5404
5405 error = fscrypt_prepare_setattr(dentry, attr);
5406 if (error)
5407 return error;
5408
5409 if (is_quota_modification(inode, attr)) {
5410 error = dquot_initialize(inode);
5411 if (error)
5412 return error;
5413 }
5414 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
5415 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
5416 handle_t *handle;
5417
5418 /* (user+group)*(old+new) structure, inode write (sb,
5419 * inode block, ? - but truncate inode update has it) */
5420 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5421 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5422 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5423 if (IS_ERR(handle)) {
5424 error = PTR_ERR(handle);
5425 goto err_out;
5426 }
5427
5428 /* dquot_transfer() calls back ext4_get_inode_usage() which
5429 * counts xattr inode references.
5430 */
5431 down_read(&EXT4_I(inode)->xattr_sem);
5432 error = dquot_transfer(inode, attr);
5433 up_read(&EXT4_I(inode)->xattr_sem);
5434
5435 if (error) {
5436 ext4_journal_stop(handle);
5437 return error;
5438 }
5439 /* Update corresponding info in inode so that everything is in
5440 * one transaction */
5441 if (attr->ia_valid & ATTR_UID)
5442 inode->i_uid = attr->ia_uid;
5443 if (attr->ia_valid & ATTR_GID)
5444 inode->i_gid = attr->ia_gid;
5445 error = ext4_mark_inode_dirty(handle, inode);
5446 ext4_journal_stop(handle);
5447 }
5448
5449 if (attr->ia_valid & ATTR_SIZE) {
5450 handle_t *handle;
5451 loff_t oldsize = inode->i_size;
5452 int shrink = (attr->ia_size <= inode->i_size);
5453
5454 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5455 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5456
5457 if (attr->ia_size > sbi->s_bitmap_maxbytes)
5458 return -EFBIG;
5459 }
5460 if (!S_ISREG(inode->i_mode))
5461 return -EINVAL;
5462
5463 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
5464 inode_inc_iversion(inode);
5465
5466 if (ext4_should_order_data(inode) &&
5467 (attr->ia_size < inode->i_size)) {
5468 error = ext4_begin_ordered_truncate(inode,
5469 attr->ia_size);
5470 if (error)
5471 goto err_out;
5472 }
5473 if (attr->ia_size != inode->i_size) {
5474 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5475 if (IS_ERR(handle)) {
5476 error = PTR_ERR(handle);
5477 goto err_out;
5478 }
5479 if (ext4_handle_valid(handle) && shrink) {
5480 error = ext4_orphan_add(handle, inode);
5481 orphan = 1;
5482 }
5483 /*
5484 * Update c/mtime on truncate up, ext4_truncate() will
5485 * update c/mtime in shrink case below
5486 */
5487 if (!shrink) {
5488 inode->i_mtime = current_time(inode);
5489 inode->i_ctime = inode->i_mtime;
5490 }
5491 down_write(&EXT4_I(inode)->i_data_sem);
5492 EXT4_I(inode)->i_disksize = attr->ia_size;
5493 rc = ext4_mark_inode_dirty(handle, inode);
5494 if (!error)
5495 error = rc;
5496 /*
5497 * We have to update i_size under i_data_sem together
5498 * with i_disksize to avoid races with writeback code
5499 * running ext4_wb_update_i_disksize().
5500 */
5501 if (!error)
5502 i_size_write(inode, attr->ia_size);
5503 up_write(&EXT4_I(inode)->i_data_sem);
5504 ext4_journal_stop(handle);
5505 if (error) {
5506 if (orphan)
5507 ext4_orphan_del(NULL, inode);
5508 goto err_out;
5509 }
5510 }
5511 if (!shrink)
5512 pagecache_isize_extended(inode, oldsize, inode->i_size);
5513
5514 /*
5515 * Blocks are going to be removed from the inode. Wait
5516 * for dio in flight. Temporarily disable
5517 * dioread_nolock to prevent livelock.
5518 */
5519 if (orphan) {
5520 if (!ext4_should_journal_data(inode)) {
5521 inode_dio_wait(inode);
5522 } else
5523 ext4_wait_for_tail_page_commit(inode);
5524 }
5525 down_write(&EXT4_I(inode)->i_mmap_sem);
5526 /*
5527 * Truncate pagecache after we've waited for commit
5528 * in data=journal mode to make pages freeable.
5529 */
5530 truncate_pagecache(inode, inode->i_size);
5531 if (shrink) {
5532 rc = ext4_truncate(inode);
5533 if (rc)
5534 error = rc;
5535 }
5536 up_write(&EXT4_I(inode)->i_mmap_sem);
5537 }
5538
5539 if (!error) {
5540 setattr_copy(inode, attr);
5541 mark_inode_dirty(inode);
5542 }
5543
5544 /*
5545 * If the call to ext4_truncate failed to get a transaction handle at
5546 * all, we need to clean up the in-core orphan list manually.
5547 */
5548 if (orphan && inode->i_nlink)
5549 ext4_orphan_del(NULL, inode);
5550
5551 if (!error && (ia_valid & ATTR_MODE))
5552 rc = posix_acl_chmod(inode, inode->i_mode);
5553
5554err_out:
5555 ext4_std_error(inode->i_sb, error);
5556 if (!error)
5557 error = rc;
5558 return error;
5559}
5560
5561int ext4_getattr(const struct path *path, struct kstat *stat,
5562 u32 request_mask, unsigned int query_flags)
5563{
5564 struct inode *inode = d_inode(path->dentry);
5565 struct ext4_inode *raw_inode;
5566 struct ext4_inode_info *ei = EXT4_I(inode);
5567 unsigned int flags;
5568
5569 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5570 stat->result_mask |= STATX_BTIME;
5571 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5572 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5573 }
5574
5575 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5576 if (flags & EXT4_APPEND_FL)
5577 stat->attributes |= STATX_ATTR_APPEND;
5578 if (flags & EXT4_COMPR_FL)
5579 stat->attributes |= STATX_ATTR_COMPRESSED;
5580 if (flags & EXT4_ENCRYPT_FL)
5581 stat->attributes |= STATX_ATTR_ENCRYPTED;
5582 if (flags & EXT4_IMMUTABLE_FL)
5583 stat->attributes |= STATX_ATTR_IMMUTABLE;
5584 if (flags & EXT4_NODUMP_FL)
5585 stat->attributes |= STATX_ATTR_NODUMP;
5586
5587 stat->attributes_mask |= (STATX_ATTR_APPEND |
5588 STATX_ATTR_COMPRESSED |
5589 STATX_ATTR_ENCRYPTED |
5590 STATX_ATTR_IMMUTABLE |
5591 STATX_ATTR_NODUMP);
5592
5593 generic_fillattr(inode, stat);
5594 return 0;
5595}
5596
5597int ext4_file_getattr(const struct path *path, struct kstat *stat,
5598 u32 request_mask, unsigned int query_flags)
5599{
5600 struct inode *inode = d_inode(path->dentry);
5601 u64 delalloc_blocks;
5602
5603 ext4_getattr(path, stat, request_mask, query_flags);
5604
5605 /*
5606 * If there is inline data in the inode, the inode will normally not
5607 * have data blocks allocated (it may have an external xattr block).
5608 * Report at least one sector for such files, so tools like tar, rsync,
5609 * others don't incorrectly think the file is completely sparse.
5610 */
5611 if (unlikely(ext4_has_inline_data(inode)))
5612 stat->blocks += (stat->size + 511) >> 9;
5613
5614 /*
5615 * We can't update i_blocks if the block allocation is delayed
5616 * otherwise in the case of system crash before the real block
5617 * allocation is done, we will have i_blocks inconsistent with
5618 * on-disk file blocks.
5619 * We always keep i_blocks updated together with real
5620 * allocation. But to not confuse with user, stat
5621 * will return the blocks that include the delayed allocation
5622 * blocks for this file.
5623 */
5624 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5625 EXT4_I(inode)->i_reserved_data_blocks);
5626 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5627 return 0;
5628}
5629
5630static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5631 int pextents)
5632{
5633 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5634 return ext4_ind_trans_blocks(inode, lblocks);
5635 return ext4_ext_index_trans_blocks(inode, pextents);
5636}
5637
5638/*
5639 * Account for index blocks, block groups bitmaps and block group
5640 * descriptor blocks if modify datablocks and index blocks
5641 * worse case, the indexs blocks spread over different block groups
5642 *
5643 * If datablocks are discontiguous, they are possible to spread over
5644 * different block groups too. If they are contiguous, with flexbg,
5645 * they could still across block group boundary.
5646 *
5647 * Also account for superblock, inode, quota and xattr blocks
5648 */
5649static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5650 int pextents)
5651{
5652 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5653 int gdpblocks;
5654 int idxblocks;
5655 int ret = 0;
5656
5657 /*
5658 * How many index blocks need to touch to map @lblocks logical blocks
5659 * to @pextents physical extents?
5660 */
5661 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5662
5663 ret = idxblocks;
5664
5665 /*
5666 * Now let's see how many group bitmaps and group descriptors need
5667 * to account
5668 */
5669 groups = idxblocks + pextents;
5670 gdpblocks = groups;
5671 if (groups > ngroups)
5672 groups = ngroups;
5673 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5674 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5675
5676 /* bitmaps and block group descriptor blocks */
5677 ret += groups + gdpblocks;
5678
5679 /* Blocks for super block, inode, quota and xattr blocks */
5680 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5681
5682 return ret;
5683}
5684
5685/*
5686 * Calculate the total number of credits to reserve to fit
5687 * the modification of a single pages into a single transaction,
5688 * which may include multiple chunks of block allocations.
5689 *
5690 * This could be called via ext4_write_begin()
5691 *
5692 * We need to consider the worse case, when
5693 * one new block per extent.
5694 */
5695int ext4_writepage_trans_blocks(struct inode *inode)
5696{
5697 int bpp = ext4_journal_blocks_per_page(inode);
5698 int ret;
5699
5700 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5701
5702 /* Account for data blocks for journalled mode */
5703 if (ext4_should_journal_data(inode))
5704 ret += bpp;
5705 return ret;
5706}
5707
5708/*
5709 * Calculate the journal credits for a chunk of data modification.
5710 *
5711 * This is called from DIO, fallocate or whoever calling
5712 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5713 *
5714 * journal buffers for data blocks are not included here, as DIO
5715 * and fallocate do no need to journal data buffers.
5716 */
5717int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5718{
5719 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5720}
5721
5722/*
5723 * The caller must have previously called ext4_reserve_inode_write().
5724 * Give this, we know that the caller already has write access to iloc->bh.
5725 */
5726int ext4_mark_iloc_dirty(handle_t *handle,
5727 struct inode *inode, struct ext4_iloc *iloc)
5728{
5729 int err = 0;
5730
5731 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5732 return -EIO;
5733
5734 if (IS_I_VERSION(inode))
5735 inode_inc_iversion(inode);
5736
5737 /* the do_update_inode consumes one bh->b_count */
5738 get_bh(iloc->bh);
5739
5740 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5741 err = ext4_do_update_inode(handle, inode, iloc);
5742 put_bh(iloc->bh);
5743 return err;
5744}
5745
5746/*
5747 * On success, We end up with an outstanding reference count against
5748 * iloc->bh. This _must_ be cleaned up later.
5749 */
5750
5751int
5752ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5753 struct ext4_iloc *iloc)
5754{
5755 int err;
5756
5757 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5758 return -EIO;
5759
5760 err = ext4_get_inode_loc(inode, iloc);
5761 if (!err) {
5762 BUFFER_TRACE(iloc->bh, "get_write_access");
5763 err = ext4_journal_get_write_access(handle, iloc->bh);
5764 if (err) {
5765 brelse(iloc->bh);
5766 iloc->bh = NULL;
5767 }
5768 }
5769 ext4_std_error(inode->i_sb, err);
5770 return err;
5771}
5772
5773static int __ext4_expand_extra_isize(struct inode *inode,
5774 unsigned int new_extra_isize,
5775 struct ext4_iloc *iloc,
5776 handle_t *handle, int *no_expand)
5777{
5778 struct ext4_inode *raw_inode;
5779 struct ext4_xattr_ibody_header *header;
5780 int error;
5781
5782 raw_inode = ext4_raw_inode(iloc);
5783
5784 header = IHDR(inode, raw_inode);
5785
5786 /* No extended attributes present */
5787 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5788 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5789 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5790 EXT4_I(inode)->i_extra_isize, 0,
5791 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5792 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5793 return 0;
5794 }
5795
5796 /* try to expand with EAs present */
5797 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5798 raw_inode, handle);
5799 if (error) {
5800 /*
5801 * Inode size expansion failed; don't try again
5802 */
5803 *no_expand = 1;
5804 }
5805
5806 return error;
5807}
5808
5809/*
5810 * Expand an inode by new_extra_isize bytes.
5811 * Returns 0 on success or negative error number on failure.
5812 */
5813static int ext4_try_to_expand_extra_isize(struct inode *inode,
5814 unsigned int new_extra_isize,
5815 struct ext4_iloc iloc,
5816 handle_t *handle)
5817{
5818 int no_expand;
5819 int error;
5820
5821 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5822 return -EOVERFLOW;
5823
5824 /*
5825 * In nojournal mode, we can immediately attempt to expand
5826 * the inode. When journaled, we first need to obtain extra
5827 * buffer credits since we may write into the EA block
5828 * with this same handle. If journal_extend fails, then it will
5829 * only result in a minor loss of functionality for that inode.
5830 * If this is felt to be critical, then e2fsck should be run to
5831 * force a large enough s_min_extra_isize.
5832 */
5833 if (ext4_handle_valid(handle) &&
5834 jbd2_journal_extend(handle,
5835 EXT4_DATA_TRANS_BLOCKS(inode->i_sb)) != 0)
5836 return -ENOSPC;
5837
5838 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5839 return -EBUSY;
5840
5841 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5842 handle, &no_expand);
5843 ext4_write_unlock_xattr(inode, &no_expand);
5844
5845 return error;
5846}
5847
5848int ext4_expand_extra_isize(struct inode *inode,
5849 unsigned int new_extra_isize,
5850 struct ext4_iloc *iloc)
5851{
5852 handle_t *handle;
5853 int no_expand;
5854 int error, rc;
5855
5856 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5857 brelse(iloc->bh);
5858 return -EOVERFLOW;
5859 }
5860
5861 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5862 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5863 if (IS_ERR(handle)) {
5864 error = PTR_ERR(handle);
5865 brelse(iloc->bh);
5866 return error;
5867 }
5868
5869 ext4_write_lock_xattr(inode, &no_expand);
5870
5871 BUFFER_TRACE(iloc.bh, "get_write_access");
5872 error = ext4_journal_get_write_access(handle, iloc->bh);
5873 if (error) {
5874 brelse(iloc->bh);
5875 goto out_stop;
5876 }
5877
5878 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5879 handle, &no_expand);
5880
5881 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5882 if (!error)
5883 error = rc;
5884
5885 ext4_write_unlock_xattr(inode, &no_expand);
5886out_stop:
5887 ext4_journal_stop(handle);
5888 return error;
5889}
5890
5891/*
5892 * What we do here is to mark the in-core inode as clean with respect to inode
5893 * dirtiness (it may still be data-dirty).
5894 * This means that the in-core inode may be reaped by prune_icache
5895 * without having to perform any I/O. This is a very good thing,
5896 * because *any* task may call prune_icache - even ones which
5897 * have a transaction open against a different journal.
5898 *
5899 * Is this cheating? Not really. Sure, we haven't written the
5900 * inode out, but prune_icache isn't a user-visible syncing function.
5901 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5902 * we start and wait on commits.
5903 */
5904int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5905{
5906 struct ext4_iloc iloc;
5907 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5908 int err;
5909
5910 might_sleep();
5911 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5912 err = ext4_reserve_inode_write(handle, inode, &iloc);
5913 if (err)
5914 return err;
5915
5916 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
5917 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
5918 iloc, handle);
5919
5920 return ext4_mark_iloc_dirty(handle, inode, &iloc);
5921}
5922
5923/*
5924 * ext4_dirty_inode() is called from __mark_inode_dirty()
5925 *
5926 * We're really interested in the case where a file is being extended.
5927 * i_size has been changed by generic_commit_write() and we thus need
5928 * to include the updated inode in the current transaction.
5929 *
5930 * Also, dquot_alloc_block() will always dirty the inode when blocks
5931 * are allocated to the file.
5932 *
5933 * If the inode is marked synchronous, we don't honour that here - doing
5934 * so would cause a commit on atime updates, which we don't bother doing.
5935 * We handle synchronous inodes at the highest possible level.
5936 *
5937 * If only the I_DIRTY_TIME flag is set, we can skip everything. If
5938 * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
5939 * to copy into the on-disk inode structure are the timestamp files.
5940 */
5941void ext4_dirty_inode(struct inode *inode, int flags)
5942{
5943 handle_t *handle;
5944
5945 if (flags == I_DIRTY_TIME)
5946 return;
5947 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5948 if (IS_ERR(handle))
5949 goto out;
5950
5951 ext4_mark_inode_dirty(handle, inode);
5952
5953 ext4_journal_stop(handle);
5954out:
5955 return;
5956}
5957
5958#if 0
5959/*
5960 * Bind an inode's backing buffer_head into this transaction, to prevent
5961 * it from being flushed to disk early. Unlike
5962 * ext4_reserve_inode_write, this leaves behind no bh reference and
5963 * returns no iloc structure, so the caller needs to repeat the iloc
5964 * lookup to mark the inode dirty later.
5965 */
5966static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5967{
5968 struct ext4_iloc iloc;
5969
5970 int err = 0;
5971 if (handle) {
5972 err = ext4_get_inode_loc(inode, &iloc);
5973 if (!err) {
5974 BUFFER_TRACE(iloc.bh, "get_write_access");
5975 err = jbd2_journal_get_write_access(handle, iloc.bh);
5976 if (!err)
5977 err = ext4_handle_dirty_metadata(handle,
5978 NULL,
5979 iloc.bh);
5980 brelse(iloc.bh);
5981 }
5982 }
5983 ext4_std_error(inode->i_sb, err);
5984 return err;
5985}
5986#endif
5987
5988int ext4_change_inode_journal_flag(struct inode *inode, int val)
5989{
5990 journal_t *journal;
5991 handle_t *handle;
5992 int err;
5993 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5994
5995 /*
5996 * We have to be very careful here: changing a data block's
5997 * journaling status dynamically is dangerous. If we write a
5998 * data block to the journal, change the status and then delete
5999 * that block, we risk forgetting to revoke the old log record
6000 * from the journal and so a subsequent replay can corrupt data.
6001 * So, first we make sure that the journal is empty and that
6002 * nobody is changing anything.
6003 */
6004
6005 journal = EXT4_JOURNAL(inode);
6006 if (!journal)
6007 return 0;
6008 if (is_journal_aborted(journal))
6009 return -EROFS;
6010
6011 /* Wait for all existing dio workers */
6012 inode_dio_wait(inode);
6013
6014 /*
6015 * Before flushing the journal and switching inode's aops, we have
6016 * to flush all dirty data the inode has. There can be outstanding
6017 * delayed allocations, there can be unwritten extents created by
6018 * fallocate or buffered writes in dioread_nolock mode covered by
6019 * dirty data which can be converted only after flushing the dirty
6020 * data (and journalled aops don't know how to handle these cases).
6021 */
6022 if (val) {
6023 down_write(&EXT4_I(inode)->i_mmap_sem);
6024 err = filemap_write_and_wait(inode->i_mapping);
6025 if (err < 0) {
6026 up_write(&EXT4_I(inode)->i_mmap_sem);
6027 return err;
6028 }
6029 }
6030
6031 percpu_down_write(&sbi->s_journal_flag_rwsem);
6032 jbd2_journal_lock_updates(journal);
6033
6034 /*
6035 * OK, there are no updates running now, and all cached data is
6036 * synced to disk. We are now in a completely consistent state
6037 * which doesn't have anything in the journal, and we know that
6038 * no filesystem updates are running, so it is safe to modify
6039 * the inode's in-core data-journaling state flag now.
6040 */
6041
6042 if (val)
6043 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6044 else {
6045 err = jbd2_journal_flush(journal);
6046 if (err < 0) {
6047 jbd2_journal_unlock_updates(journal);
6048 percpu_up_write(&sbi->s_journal_flag_rwsem);
6049 return err;
6050 }
6051 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6052 }
6053 ext4_set_aops(inode);
6054
6055 jbd2_journal_unlock_updates(journal);
6056 percpu_up_write(&sbi->s_journal_flag_rwsem);
6057
6058 if (val)
6059 up_write(&EXT4_I(inode)->i_mmap_sem);
6060
6061 /* Finally we can mark the inode as dirty. */
6062
6063 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6064 if (IS_ERR(handle))
6065 return PTR_ERR(handle);
6066
6067 err = ext4_mark_inode_dirty(handle, inode);
6068 ext4_handle_sync(handle);
6069 ext4_journal_stop(handle);
6070 ext4_std_error(inode->i_sb, err);
6071
6072 return err;
6073}
6074
6075static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
6076{
6077 return !buffer_mapped(bh);
6078}
6079
6080int ext4_page_mkwrite(struct vm_fault *vmf)
6081{
6082 struct vm_area_struct *vma = vmf->vma;
6083 struct page *page = vmf->page;
6084 loff_t size;
6085 unsigned long len;
6086 int ret;
6087 struct file *file = vma->vm_file;
6088 struct inode *inode = file_inode(file);
6089 struct address_space *mapping = inode->i_mapping;
6090 handle_t *handle;
6091 get_block_t *get_block;
6092 int retries = 0;
6093
6094 sb_start_pagefault(inode->i_sb);
6095 file_update_time(vma->vm_file);
6096
6097 down_read(&EXT4_I(inode)->i_mmap_sem);
6098
6099 ret = ext4_convert_inline_data(inode);
6100 if (ret)
6101 goto out_ret;
6102
6103 /* Delalloc case is easy... */
6104 if (test_opt(inode->i_sb, DELALLOC) &&
6105 !ext4_should_journal_data(inode) &&
6106 !ext4_nonda_switch(inode->i_sb)) {
6107 do {
6108 ret = block_page_mkwrite(vma, vmf,
6109 ext4_da_get_block_prep);
6110 } while (ret == -ENOSPC &&
6111 ext4_should_retry_alloc(inode->i_sb, &retries));
6112 goto out_ret;
6113 }
6114
6115 lock_page(page);
6116 size = i_size_read(inode);
6117 /* Page got truncated from under us? */
6118 if (page->mapping != mapping || page_offset(page) > size) {
6119 unlock_page(page);
6120 ret = VM_FAULT_NOPAGE;
6121 goto out;
6122 }
6123
6124 if (page->index == size >> PAGE_SHIFT)
6125 len = size & ~PAGE_MASK;
6126 else
6127 len = PAGE_SIZE;
6128 /*
6129 * Return if we have all the buffers mapped. This avoids the need to do
6130 * journal_start/journal_stop which can block and take a long time
6131 */
6132 if (page_has_buffers(page)) {
6133 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
6134 0, len, NULL,
6135 ext4_bh_unmapped)) {
6136 /* Wait so that we don't change page under IO */
6137 wait_for_stable_page(page);
6138 ret = VM_FAULT_LOCKED;
6139 goto out;
6140 }
6141 }
6142 unlock_page(page);
6143 /* OK, we need to fill the hole... */
6144 if (ext4_should_dioread_nolock(inode))
6145 get_block = ext4_get_block_unwritten;
6146 else
6147 get_block = ext4_get_block;
6148retry_alloc:
6149 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6150 ext4_writepage_trans_blocks(inode));
6151 if (IS_ERR(handle)) {
6152 ret = VM_FAULT_SIGBUS;
6153 goto out;
6154 }
6155 ret = block_page_mkwrite(vma, vmf, get_block);
6156 if (!ret && ext4_should_journal_data(inode)) {
6157 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
6158 PAGE_SIZE, NULL, do_journal_get_write_access)) {
6159 unlock_page(page);
6160 ret = VM_FAULT_SIGBUS;
6161 ext4_journal_stop(handle);
6162 goto out;
6163 }
6164 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
6165 }
6166 ext4_journal_stop(handle);
6167 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6168 goto retry_alloc;
6169out_ret:
6170 ret = block_page_mkwrite_return(ret);
6171out:
6172 up_read(&EXT4_I(inode)->i_mmap_sem);
6173 sb_end_pagefault(inode->i_sb);
6174 return ret;
6175}
6176
6177int ext4_filemap_fault(struct vm_fault *vmf)
6178{
6179 struct inode *inode = file_inode(vmf->vma->vm_file);
6180 int err;
6181
6182 down_read(&EXT4_I(inode)->i_mmap_sem);
6183 err = filemap_fault(vmf);
6184 up_read(&EXT4_I(inode)->i_mmap_sem);
6185
6186 return err;
6187}