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 = (__u32)ktime_get_real_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 !inode_is_open_for_write(inode))
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_has_feature_journal(inode->i_sb) &&
403 (inode->i_ino ==
404 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
405 return 0;
406 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
407 map->m_len)) {
408 ext4_error_inode(inode, func, line, map->m_pblk,
409 "lblock %lu mapped to illegal pblock %llu "
410 "(length %d)", (unsigned long) map->m_lblk,
411 map->m_pblk, map->m_len);
412 return -EFSCORRUPTED;
413 }
414 return 0;
415}
416
417int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
418 ext4_lblk_t len)
419{
420 int ret;
421
422 if (IS_ENCRYPTED(inode))
423 return fscrypt_zeroout_range(inode, lblk, pblk, len);
424
425 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
426 if (ret > 0)
427 ret = 0;
428
429 return ret;
430}
431
432#define check_block_validity(inode, map) \
433 __check_block_validity((inode), __func__, __LINE__, (map))
434
435#ifdef ES_AGGRESSIVE_TEST
436static void ext4_map_blocks_es_recheck(handle_t *handle,
437 struct inode *inode,
438 struct ext4_map_blocks *es_map,
439 struct ext4_map_blocks *map,
440 int flags)
441{
442 int retval;
443
444 map->m_flags = 0;
445 /*
446 * There is a race window that the result is not the same.
447 * e.g. xfstests #223 when dioread_nolock enables. The reason
448 * is that we lookup a block mapping in extent status tree with
449 * out taking i_data_sem. So at the time the unwritten extent
450 * could be converted.
451 */
452 down_read(&EXT4_I(inode)->i_data_sem);
453 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
454 retval = ext4_ext_map_blocks(handle, inode, map, flags &
455 EXT4_GET_BLOCKS_KEEP_SIZE);
456 } else {
457 retval = ext4_ind_map_blocks(handle, inode, map, flags &
458 EXT4_GET_BLOCKS_KEEP_SIZE);
459 }
460 up_read((&EXT4_I(inode)->i_data_sem));
461
462 /*
463 * We don't check m_len because extent will be collpased in status
464 * tree. So the m_len might not equal.
465 */
466 if (es_map->m_lblk != map->m_lblk ||
467 es_map->m_flags != map->m_flags ||
468 es_map->m_pblk != map->m_pblk) {
469 printk("ES cache assertion failed for inode: %lu "
470 "es_cached ex [%d/%d/%llu/%x] != "
471 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
472 inode->i_ino, es_map->m_lblk, es_map->m_len,
473 es_map->m_pblk, es_map->m_flags, map->m_lblk,
474 map->m_len, map->m_pblk, map->m_flags,
475 retval, flags);
476 }
477}
478#endif /* ES_AGGRESSIVE_TEST */
479
480/*
481 * The ext4_map_blocks() function tries to look up the requested blocks,
482 * and returns if the blocks are already mapped.
483 *
484 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
485 * and store the allocated blocks in the result buffer head and mark it
486 * mapped.
487 *
488 * If file type is extents based, it will call ext4_ext_map_blocks(),
489 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
490 * based files
491 *
492 * On success, it returns the number of blocks being mapped or allocated. if
493 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
494 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
495 *
496 * It returns 0 if plain look up failed (blocks have not been allocated), in
497 * that case, @map is returned as unmapped but we still do fill map->m_len to
498 * indicate the length of a hole starting at map->m_lblk.
499 *
500 * It returns the error in case of allocation failure.
501 */
502int ext4_map_blocks(handle_t *handle, struct inode *inode,
503 struct ext4_map_blocks *map, int flags)
504{
505 struct extent_status es;
506 int retval;
507 int ret = 0;
508#ifdef ES_AGGRESSIVE_TEST
509 struct ext4_map_blocks orig_map;
510
511 memcpy(&orig_map, map, sizeof(*map));
512#endif
513
514 map->m_flags = 0;
515 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
516 "logical block %lu\n", inode->i_ino, flags, map->m_len,
517 (unsigned long) map->m_lblk);
518
519 /*
520 * ext4_map_blocks returns an int, and m_len is an unsigned int
521 */
522 if (unlikely(map->m_len > INT_MAX))
523 map->m_len = INT_MAX;
524
525 /* We can handle the block number less than EXT_MAX_BLOCKS */
526 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
527 return -EFSCORRUPTED;
528
529 /* Lookup extent status tree firstly */
530 if (ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
531 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
532 map->m_pblk = ext4_es_pblock(&es) +
533 map->m_lblk - es.es_lblk;
534 map->m_flags |= ext4_es_is_written(&es) ?
535 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
536 retval = es.es_len - (map->m_lblk - es.es_lblk);
537 if (retval > map->m_len)
538 retval = map->m_len;
539 map->m_len = retval;
540 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
541 map->m_pblk = 0;
542 retval = es.es_len - (map->m_lblk - es.es_lblk);
543 if (retval > map->m_len)
544 retval = map->m_len;
545 map->m_len = retval;
546 retval = 0;
547 } else {
548 BUG();
549 }
550#ifdef ES_AGGRESSIVE_TEST
551 ext4_map_blocks_es_recheck(handle, inode, map,
552 &orig_map, flags);
553#endif
554 goto found;
555 }
556
557 /*
558 * Try to see if we can get the block without requesting a new
559 * file system block.
560 */
561 down_read(&EXT4_I(inode)->i_data_sem);
562 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
563 retval = ext4_ext_map_blocks(handle, inode, map, flags &
564 EXT4_GET_BLOCKS_KEEP_SIZE);
565 } else {
566 retval = ext4_ind_map_blocks(handle, inode, map, flags &
567 EXT4_GET_BLOCKS_KEEP_SIZE);
568 }
569 if (retval > 0) {
570 unsigned int status;
571
572 if (unlikely(retval != map->m_len)) {
573 ext4_warning(inode->i_sb,
574 "ES len assertion failed for inode "
575 "%lu: retval %d != map->m_len %d",
576 inode->i_ino, retval, map->m_len);
577 WARN_ON(1);
578 }
579
580 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
581 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
582 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
583 !(status & EXTENT_STATUS_WRITTEN) &&
584 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
585 map->m_lblk + map->m_len - 1))
586 status |= EXTENT_STATUS_DELAYED;
587 ret = ext4_es_insert_extent(inode, map->m_lblk,
588 map->m_len, map->m_pblk, status);
589 if (ret < 0)
590 retval = ret;
591 }
592 up_read((&EXT4_I(inode)->i_data_sem));
593
594found:
595 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
596 ret = check_block_validity(inode, map);
597 if (ret != 0)
598 return ret;
599 }
600
601 /* If it is only a block(s) look up */
602 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
603 return retval;
604
605 /*
606 * Returns if the blocks have already allocated
607 *
608 * Note that if blocks have been preallocated
609 * ext4_ext_get_block() returns the create = 0
610 * with buffer head unmapped.
611 */
612 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
613 /*
614 * If we need to convert extent to unwritten
615 * we continue and do the actual work in
616 * ext4_ext_map_blocks()
617 */
618 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
619 return retval;
620
621 /*
622 * Here we clear m_flags because after allocating an new extent,
623 * it will be set again.
624 */
625 map->m_flags &= ~EXT4_MAP_FLAGS;
626
627 /*
628 * New blocks allocate and/or writing to unwritten extent
629 * will possibly result in updating i_data, so we take
630 * the write lock of i_data_sem, and call get_block()
631 * with create == 1 flag.
632 */
633 down_write(&EXT4_I(inode)->i_data_sem);
634
635 /*
636 * We need to check for EXT4 here because migrate
637 * could have changed the inode type in between
638 */
639 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
640 retval = ext4_ext_map_blocks(handle, inode, map, flags);
641 } else {
642 retval = ext4_ind_map_blocks(handle, inode, map, flags);
643
644 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
645 /*
646 * We allocated new blocks which will result in
647 * i_data's format changing. Force the migrate
648 * to fail by clearing migrate flags
649 */
650 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
651 }
652
653 /*
654 * Update reserved blocks/metadata blocks after successful
655 * block allocation which had been deferred till now. We don't
656 * support fallocate for non extent files. So we can update
657 * reserve space here.
658 */
659 if ((retval > 0) &&
660 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
661 ext4_da_update_reserve_space(inode, retval, 1);
662 }
663
664 if (retval > 0) {
665 unsigned int status;
666
667 if (unlikely(retval != map->m_len)) {
668 ext4_warning(inode->i_sb,
669 "ES len assertion failed for inode "
670 "%lu: retval %d != map->m_len %d",
671 inode->i_ino, retval, map->m_len);
672 WARN_ON(1);
673 }
674
675 /*
676 * We have to zeroout blocks before inserting them into extent
677 * status tree. Otherwise someone could look them up there and
678 * use them before they are really zeroed. We also have to
679 * unmap metadata before zeroing as otherwise writeback can
680 * overwrite zeros with stale data from block device.
681 */
682 if (flags & EXT4_GET_BLOCKS_ZERO &&
683 map->m_flags & EXT4_MAP_MAPPED &&
684 map->m_flags & EXT4_MAP_NEW) {
685 ret = ext4_issue_zeroout(inode, map->m_lblk,
686 map->m_pblk, map->m_len);
687 if (ret) {
688 retval = ret;
689 goto out_sem;
690 }
691 }
692
693 /*
694 * If the extent has been zeroed out, we don't need to update
695 * extent status tree.
696 */
697 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
698 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
699 if (ext4_es_is_written(&es))
700 goto out_sem;
701 }
702 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
703 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
704 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
705 !(status & EXTENT_STATUS_WRITTEN) &&
706 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
707 map->m_lblk + map->m_len - 1))
708 status |= EXTENT_STATUS_DELAYED;
709 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
710 map->m_pblk, status);
711 if (ret < 0) {
712 retval = ret;
713 goto out_sem;
714 }
715 }
716
717out_sem:
718 up_write((&EXT4_I(inode)->i_data_sem));
719 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
720 ret = check_block_validity(inode, map);
721 if (ret != 0)
722 return ret;
723
724 /*
725 * Inodes with freshly allocated blocks where contents will be
726 * visible after transaction commit must be on transaction's
727 * ordered data list.
728 */
729 if (map->m_flags & EXT4_MAP_NEW &&
730 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
731 !(flags & EXT4_GET_BLOCKS_ZERO) &&
732 !ext4_is_quota_file(inode) &&
733 ext4_should_order_data(inode)) {
734 loff_t start_byte =
735 (loff_t)map->m_lblk << inode->i_blkbits;
736 loff_t length = (loff_t)map->m_len << inode->i_blkbits;
737
738 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
739 ret = ext4_jbd2_inode_add_wait(handle, inode,
740 start_byte, length);
741 else
742 ret = ext4_jbd2_inode_add_write(handle, inode,
743 start_byte, length);
744 if (ret)
745 return ret;
746 }
747 }
748 return retval;
749}
750
751/*
752 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
753 * we have to be careful as someone else may be manipulating b_state as well.
754 */
755static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
756{
757 unsigned long old_state;
758 unsigned long new_state;
759
760 flags &= EXT4_MAP_FLAGS;
761
762 /* Dummy buffer_head? Set non-atomically. */
763 if (!bh->b_page) {
764 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
765 return;
766 }
767 /*
768 * Someone else may be modifying b_state. Be careful! This is ugly but
769 * once we get rid of using bh as a container for mapping information
770 * to pass to / from get_block functions, this can go away.
771 */
772 do {
773 old_state = READ_ONCE(bh->b_state);
774 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
775 } while (unlikely(
776 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
777}
778
779static int _ext4_get_block(struct inode *inode, sector_t iblock,
780 struct buffer_head *bh, int flags)
781{
782 struct ext4_map_blocks map;
783 int ret = 0;
784
785 if (ext4_has_inline_data(inode))
786 return -ERANGE;
787
788 map.m_lblk = iblock;
789 map.m_len = bh->b_size >> inode->i_blkbits;
790
791 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
792 flags);
793 if (ret > 0) {
794 map_bh(bh, inode->i_sb, map.m_pblk);
795 ext4_update_bh_state(bh, map.m_flags);
796 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
797 ret = 0;
798 } else if (ret == 0) {
799 /* hole case, need to fill in bh->b_size */
800 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
801 }
802 return ret;
803}
804
805int ext4_get_block(struct inode *inode, sector_t iblock,
806 struct buffer_head *bh, int create)
807{
808 return _ext4_get_block(inode, iblock, bh,
809 create ? EXT4_GET_BLOCKS_CREATE : 0);
810}
811
812/*
813 * Get block function used when preparing for buffered write if we require
814 * creating an unwritten extent if blocks haven't been allocated. The extent
815 * will be converted to written after the IO is complete.
816 */
817int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
818 struct buffer_head *bh_result, int create)
819{
820 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
821 inode->i_ino, create);
822 return _ext4_get_block(inode, iblock, bh_result,
823 EXT4_GET_BLOCKS_IO_CREATE_EXT);
824}
825
826/* Maximum number of blocks we map for direct IO at once. */
827#define DIO_MAX_BLOCKS 4096
828
829/*
830 * Get blocks function for the cases that need to start a transaction -
831 * generally difference cases of direct IO and DAX IO. It also handles retries
832 * in case of ENOSPC.
833 */
834static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
835 struct buffer_head *bh_result, int flags)
836{
837 int dio_credits;
838 handle_t *handle;
839 int retries = 0;
840 int ret;
841
842 /* Trim mapping request to maximum we can map at once for DIO */
843 if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
844 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
845 dio_credits = ext4_chunk_trans_blocks(inode,
846 bh_result->b_size >> inode->i_blkbits);
847retry:
848 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
849 if (IS_ERR(handle))
850 return PTR_ERR(handle);
851
852 ret = _ext4_get_block(inode, iblock, bh_result, flags);
853 ext4_journal_stop(handle);
854
855 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
856 goto retry;
857 return ret;
858}
859
860/* Get block function for DIO reads and writes to inodes without extents */
861int ext4_dio_get_block(struct inode *inode, sector_t iblock,
862 struct buffer_head *bh, int create)
863{
864 /* We don't expect handle for direct IO */
865 WARN_ON_ONCE(ext4_journal_current_handle());
866
867 if (!create)
868 return _ext4_get_block(inode, iblock, bh, 0);
869 return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
870}
871
872/*
873 * Get block function for AIO DIO writes when we create unwritten extent if
874 * blocks are not allocated yet. The extent will be converted to written
875 * after IO is complete.
876 */
877static int ext4_dio_get_block_unwritten_async(struct inode *inode,
878 sector_t iblock, struct buffer_head *bh_result, int create)
879{
880 int ret;
881
882 /* We don't expect handle for direct IO */
883 WARN_ON_ONCE(ext4_journal_current_handle());
884
885 ret = ext4_get_block_trans(inode, iblock, bh_result,
886 EXT4_GET_BLOCKS_IO_CREATE_EXT);
887
888 /*
889 * When doing DIO using unwritten extents, we need io_end to convert
890 * unwritten extents to written on IO completion. We allocate io_end
891 * once we spot unwritten extent and store it in b_private. Generic
892 * DIO code keeps b_private set and furthermore passes the value to
893 * our completion callback in 'private' argument.
894 */
895 if (!ret && buffer_unwritten(bh_result)) {
896 if (!bh_result->b_private) {
897 ext4_io_end_t *io_end;
898
899 io_end = ext4_init_io_end(inode, GFP_KERNEL);
900 if (!io_end)
901 return -ENOMEM;
902 bh_result->b_private = io_end;
903 ext4_set_io_unwritten_flag(inode, io_end);
904 }
905 set_buffer_defer_completion(bh_result);
906 }
907
908 return ret;
909}
910
911/*
912 * Get block function for non-AIO DIO writes when we create unwritten extent if
913 * blocks are not allocated yet. The extent will be converted to written
914 * after IO is complete by ext4_direct_IO_write().
915 */
916static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
917 sector_t iblock, struct buffer_head *bh_result, int create)
918{
919 int ret;
920
921 /* We don't expect handle for direct IO */
922 WARN_ON_ONCE(ext4_journal_current_handle());
923
924 ret = ext4_get_block_trans(inode, iblock, bh_result,
925 EXT4_GET_BLOCKS_IO_CREATE_EXT);
926
927 /*
928 * Mark inode as having pending DIO writes to unwritten extents.
929 * ext4_direct_IO_write() checks this flag and converts extents to
930 * written.
931 */
932 if (!ret && buffer_unwritten(bh_result))
933 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
934
935 return ret;
936}
937
938static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
939 struct buffer_head *bh_result, int create)
940{
941 int ret;
942
943 ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
944 inode->i_ino, create);
945 /* We don't expect handle for direct IO */
946 WARN_ON_ONCE(ext4_journal_current_handle());
947
948 ret = _ext4_get_block(inode, iblock, bh_result, 0);
949 /*
950 * Blocks should have been preallocated! ext4_file_write_iter() checks
951 * that.
952 */
953 WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
954
955 return ret;
956}
957
958
959/*
960 * `handle' can be NULL if create is zero
961 */
962struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
963 ext4_lblk_t block, int map_flags)
964{
965 struct ext4_map_blocks map;
966 struct buffer_head *bh;
967 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
968 int err;
969
970 J_ASSERT(handle != NULL || create == 0);
971
972 map.m_lblk = block;
973 map.m_len = 1;
974 err = ext4_map_blocks(handle, inode, &map, map_flags);
975
976 if (err == 0)
977 return create ? ERR_PTR(-ENOSPC) : NULL;
978 if (err < 0)
979 return ERR_PTR(err);
980
981 bh = sb_getblk(inode->i_sb, map.m_pblk);
982 if (unlikely(!bh))
983 return ERR_PTR(-ENOMEM);
984 if (map.m_flags & EXT4_MAP_NEW) {
985 J_ASSERT(create != 0);
986 J_ASSERT(handle != NULL);
987
988 /*
989 * Now that we do not always journal data, we should
990 * keep in mind whether this should always journal the
991 * new buffer as metadata. For now, regular file
992 * writes use ext4_get_block instead, so it's not a
993 * problem.
994 */
995 lock_buffer(bh);
996 BUFFER_TRACE(bh, "call get_create_access");
997 err = ext4_journal_get_create_access(handle, bh);
998 if (unlikely(err)) {
999 unlock_buffer(bh);
1000 goto errout;
1001 }
1002 if (!buffer_uptodate(bh)) {
1003 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1004 set_buffer_uptodate(bh);
1005 }
1006 unlock_buffer(bh);
1007 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1008 err = ext4_handle_dirty_metadata(handle, inode, bh);
1009 if (unlikely(err))
1010 goto errout;
1011 } else
1012 BUFFER_TRACE(bh, "not a new buffer");
1013 return bh;
1014errout:
1015 brelse(bh);
1016 return ERR_PTR(err);
1017}
1018
1019struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1020 ext4_lblk_t block, int map_flags)
1021{
1022 struct buffer_head *bh;
1023
1024 bh = ext4_getblk(handle, inode, block, map_flags);
1025 if (IS_ERR(bh))
1026 return bh;
1027 if (!bh || ext4_buffer_uptodate(bh))
1028 return bh;
1029 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1030 wait_on_buffer(bh);
1031 if (buffer_uptodate(bh))
1032 return bh;
1033 put_bh(bh);
1034 return ERR_PTR(-EIO);
1035}
1036
1037/* Read a contiguous batch of blocks. */
1038int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
1039 bool wait, struct buffer_head **bhs)
1040{
1041 int i, err;
1042
1043 for (i = 0; i < bh_count; i++) {
1044 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
1045 if (IS_ERR(bhs[i])) {
1046 err = PTR_ERR(bhs[i]);
1047 bh_count = i;
1048 goto out_brelse;
1049 }
1050 }
1051
1052 for (i = 0; i < bh_count; i++)
1053 /* Note that NULL bhs[i] is valid because of holes. */
1054 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
1055 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
1056 &bhs[i]);
1057
1058 if (!wait)
1059 return 0;
1060
1061 for (i = 0; i < bh_count; i++)
1062 if (bhs[i])
1063 wait_on_buffer(bhs[i]);
1064
1065 for (i = 0; i < bh_count; i++) {
1066 if (bhs[i] && !buffer_uptodate(bhs[i])) {
1067 err = -EIO;
1068 goto out_brelse;
1069 }
1070 }
1071 return 0;
1072
1073out_brelse:
1074 for (i = 0; i < bh_count; i++) {
1075 brelse(bhs[i]);
1076 bhs[i] = NULL;
1077 }
1078 return err;
1079}
1080
1081int ext4_walk_page_buffers(handle_t *handle,
1082 struct buffer_head *head,
1083 unsigned from,
1084 unsigned to,
1085 int *partial,
1086 int (*fn)(handle_t *handle,
1087 struct buffer_head *bh))
1088{
1089 struct buffer_head *bh;
1090 unsigned block_start, block_end;
1091 unsigned blocksize = head->b_size;
1092 int err, ret = 0;
1093 struct buffer_head *next;
1094
1095 for (bh = head, block_start = 0;
1096 ret == 0 && (bh != head || !block_start);
1097 block_start = block_end, bh = next) {
1098 next = bh->b_this_page;
1099 block_end = block_start + blocksize;
1100 if (block_end <= from || block_start >= to) {
1101 if (partial && !buffer_uptodate(bh))
1102 *partial = 1;
1103 continue;
1104 }
1105 err = (*fn)(handle, bh);
1106 if (!ret)
1107 ret = err;
1108 }
1109 return ret;
1110}
1111
1112/*
1113 * To preserve ordering, it is essential that the hole instantiation and
1114 * the data write be encapsulated in a single transaction. We cannot
1115 * close off a transaction and start a new one between the ext4_get_block()
1116 * and the commit_write(). So doing the jbd2_journal_start at the start of
1117 * prepare_write() is the right place.
1118 *
1119 * Also, this function can nest inside ext4_writepage(). In that case, we
1120 * *know* that ext4_writepage() has generated enough buffer credits to do the
1121 * whole page. So we won't block on the journal in that case, which is good,
1122 * because the caller may be PF_MEMALLOC.
1123 *
1124 * By accident, ext4 can be reentered when a transaction is open via
1125 * quota file writes. If we were to commit the transaction while thus
1126 * reentered, there can be a deadlock - we would be holding a quota
1127 * lock, and the commit would never complete if another thread had a
1128 * transaction open and was blocking on the quota lock - a ranking
1129 * violation.
1130 *
1131 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1132 * will _not_ run commit under these circumstances because handle->h_ref
1133 * is elevated. We'll still have enough credits for the tiny quotafile
1134 * write.
1135 */
1136int do_journal_get_write_access(handle_t *handle,
1137 struct buffer_head *bh)
1138{
1139 int dirty = buffer_dirty(bh);
1140 int ret;
1141
1142 if (!buffer_mapped(bh) || buffer_freed(bh))
1143 return 0;
1144 /*
1145 * __block_write_begin() could have dirtied some buffers. Clean
1146 * the dirty bit as jbd2_journal_get_write_access() could complain
1147 * otherwise about fs integrity issues. Setting of the dirty bit
1148 * by __block_write_begin() isn't a real problem here as we clear
1149 * the bit before releasing a page lock and thus writeback cannot
1150 * ever write the buffer.
1151 */
1152 if (dirty)
1153 clear_buffer_dirty(bh);
1154 BUFFER_TRACE(bh, "get write access");
1155 ret = ext4_journal_get_write_access(handle, bh);
1156 if (!ret && dirty)
1157 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1158 return ret;
1159}
1160
1161#ifdef CONFIG_FS_ENCRYPTION
1162static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1163 get_block_t *get_block)
1164{
1165 unsigned from = pos & (PAGE_SIZE - 1);
1166 unsigned to = from + len;
1167 struct inode *inode = page->mapping->host;
1168 unsigned block_start, block_end;
1169 sector_t block;
1170 int err = 0;
1171 unsigned blocksize = inode->i_sb->s_blocksize;
1172 unsigned bbits;
1173 struct buffer_head *bh, *head, *wait[2];
1174 int nr_wait = 0;
1175 int i;
1176
1177 BUG_ON(!PageLocked(page));
1178 BUG_ON(from > PAGE_SIZE);
1179 BUG_ON(to > PAGE_SIZE);
1180 BUG_ON(from > to);
1181
1182 if (!page_has_buffers(page))
1183 create_empty_buffers(page, blocksize, 0);
1184 head = page_buffers(page);
1185 bbits = ilog2(blocksize);
1186 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1187
1188 for (bh = head, block_start = 0; bh != head || !block_start;
1189 block++, block_start = block_end, bh = bh->b_this_page) {
1190 block_end = block_start + blocksize;
1191 if (block_end <= from || block_start >= to) {
1192 if (PageUptodate(page)) {
1193 if (!buffer_uptodate(bh))
1194 set_buffer_uptodate(bh);
1195 }
1196 continue;
1197 }
1198 if (buffer_new(bh))
1199 clear_buffer_new(bh);
1200 if (!buffer_mapped(bh)) {
1201 WARN_ON(bh->b_size != blocksize);
1202 err = get_block(inode, block, bh, 1);
1203 if (err)
1204 break;
1205 if (buffer_new(bh)) {
1206 if (PageUptodate(page)) {
1207 clear_buffer_new(bh);
1208 set_buffer_uptodate(bh);
1209 mark_buffer_dirty(bh);
1210 continue;
1211 }
1212 if (block_end > to || block_start < from)
1213 zero_user_segments(page, to, block_end,
1214 block_start, from);
1215 continue;
1216 }
1217 }
1218 if (PageUptodate(page)) {
1219 if (!buffer_uptodate(bh))
1220 set_buffer_uptodate(bh);
1221 continue;
1222 }
1223 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1224 !buffer_unwritten(bh) &&
1225 (block_start < from || block_end > to)) {
1226 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1227 wait[nr_wait++] = bh;
1228 }
1229 }
1230 /*
1231 * If we issued read requests, let them complete.
1232 */
1233 for (i = 0; i < nr_wait; i++) {
1234 wait_on_buffer(wait[i]);
1235 if (!buffer_uptodate(wait[i]))
1236 err = -EIO;
1237 }
1238 if (unlikely(err)) {
1239 page_zero_new_buffers(page, from, to);
1240 } else if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode)) {
1241 for (i = 0; i < nr_wait; i++) {
1242 int err2;
1243
1244 err2 = fscrypt_decrypt_pagecache_blocks(page, blocksize,
1245 bh_offset(wait[i]));
1246 if (err2) {
1247 clear_buffer_uptodate(wait[i]);
1248 err = err2;
1249 }
1250 }
1251 }
1252
1253 return err;
1254}
1255#endif
1256
1257static int ext4_write_begin(struct file *file, struct address_space *mapping,
1258 loff_t pos, unsigned len, unsigned flags,
1259 struct page **pagep, void **fsdata)
1260{
1261 struct inode *inode = mapping->host;
1262 int ret, needed_blocks;
1263 handle_t *handle;
1264 int retries = 0;
1265 struct page *page;
1266 pgoff_t index;
1267 unsigned from, to;
1268
1269 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1270 return -EIO;
1271
1272 trace_ext4_write_begin(inode, pos, len, flags);
1273 /*
1274 * Reserve one block more for addition to orphan list in case
1275 * we allocate blocks but write fails for some reason
1276 */
1277 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1278 index = pos >> PAGE_SHIFT;
1279 from = pos & (PAGE_SIZE - 1);
1280 to = from + len;
1281
1282 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1283 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1284 flags, pagep);
1285 if (ret < 0)
1286 return ret;
1287 if (ret == 1)
1288 return 0;
1289 }
1290
1291 /*
1292 * grab_cache_page_write_begin() can take a long time if the
1293 * system is thrashing due to memory pressure, or if the page
1294 * is being written back. So grab it first before we start
1295 * the transaction handle. This also allows us to allocate
1296 * the page (if needed) without using GFP_NOFS.
1297 */
1298retry_grab:
1299 page = grab_cache_page_write_begin(mapping, index, flags);
1300 if (!page)
1301 return -ENOMEM;
1302 unlock_page(page);
1303
1304retry_journal:
1305 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1306 if (IS_ERR(handle)) {
1307 put_page(page);
1308 return PTR_ERR(handle);
1309 }
1310
1311 lock_page(page);
1312 if (page->mapping != mapping) {
1313 /* The page got truncated from under us */
1314 unlock_page(page);
1315 put_page(page);
1316 ext4_journal_stop(handle);
1317 goto retry_grab;
1318 }
1319 /* In case writeback began while the page was unlocked */
1320 wait_for_stable_page(page);
1321
1322#ifdef CONFIG_FS_ENCRYPTION
1323 if (ext4_should_dioread_nolock(inode))
1324 ret = ext4_block_write_begin(page, pos, len,
1325 ext4_get_block_unwritten);
1326 else
1327 ret = ext4_block_write_begin(page, pos, len,
1328 ext4_get_block);
1329#else
1330 if (ext4_should_dioread_nolock(inode))
1331 ret = __block_write_begin(page, pos, len,
1332 ext4_get_block_unwritten);
1333 else
1334 ret = __block_write_begin(page, pos, len, ext4_get_block);
1335#endif
1336 if (!ret && ext4_should_journal_data(inode)) {
1337 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1338 from, to, NULL,
1339 do_journal_get_write_access);
1340 }
1341
1342 if (ret) {
1343 bool extended = (pos + len > inode->i_size) &&
1344 !ext4_verity_in_progress(inode);
1345
1346 unlock_page(page);
1347 /*
1348 * __block_write_begin may have instantiated a few blocks
1349 * outside i_size. Trim these off again. Don't need
1350 * i_size_read because we hold i_mutex.
1351 *
1352 * Add inode to orphan list in case we crash before
1353 * truncate finishes
1354 */
1355 if (extended && ext4_can_truncate(inode))
1356 ext4_orphan_add(handle, inode);
1357
1358 ext4_journal_stop(handle);
1359 if (extended) {
1360 ext4_truncate_failed_write(inode);
1361 /*
1362 * If truncate failed early the inode might
1363 * still be on the orphan list; we need to
1364 * make sure the inode is removed from the
1365 * orphan list in that case.
1366 */
1367 if (inode->i_nlink)
1368 ext4_orphan_del(NULL, inode);
1369 }
1370
1371 if (ret == -ENOSPC &&
1372 ext4_should_retry_alloc(inode->i_sb, &retries))
1373 goto retry_journal;
1374 put_page(page);
1375 return ret;
1376 }
1377 *pagep = page;
1378 return ret;
1379}
1380
1381/* For write_end() in data=journal mode */
1382static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1383{
1384 int ret;
1385 if (!buffer_mapped(bh) || buffer_freed(bh))
1386 return 0;
1387 set_buffer_uptodate(bh);
1388 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1389 clear_buffer_meta(bh);
1390 clear_buffer_prio(bh);
1391 return ret;
1392}
1393
1394/*
1395 * We need to pick up the new inode size which generic_commit_write gave us
1396 * `file' can be NULL - eg, when called from page_symlink().
1397 *
1398 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1399 * buffers are managed internally.
1400 */
1401static int ext4_write_end(struct file *file,
1402 struct address_space *mapping,
1403 loff_t pos, unsigned len, unsigned copied,
1404 struct page *page, void *fsdata)
1405{
1406 handle_t *handle = ext4_journal_current_handle();
1407 struct inode *inode = mapping->host;
1408 loff_t old_size = inode->i_size;
1409 int ret = 0, ret2;
1410 int i_size_changed = 0;
1411 int inline_data = ext4_has_inline_data(inode);
1412 bool verity = ext4_verity_in_progress(inode);
1413
1414 trace_ext4_write_end(inode, pos, len, copied);
1415 if (inline_data) {
1416 ret = ext4_write_inline_data_end(inode, pos, len,
1417 copied, page);
1418 if (ret < 0) {
1419 unlock_page(page);
1420 put_page(page);
1421 goto errout;
1422 }
1423 copied = ret;
1424 } else
1425 copied = block_write_end(file, mapping, pos,
1426 len, copied, page, fsdata);
1427 /*
1428 * it's important to update i_size while still holding page lock:
1429 * page writeout could otherwise come in and zero beyond i_size.
1430 *
1431 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1432 * blocks are being written past EOF, so skip the i_size update.
1433 */
1434 if (!verity)
1435 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1436 unlock_page(page);
1437 put_page(page);
1438
1439 if (old_size < pos && !verity)
1440 pagecache_isize_extended(inode, old_size, pos);
1441 /*
1442 * Don't mark the inode dirty under page lock. First, it unnecessarily
1443 * makes the holding time of page lock longer. Second, it forces lock
1444 * ordering of page lock and transaction start for journaling
1445 * filesystems.
1446 */
1447 if (i_size_changed || inline_data)
1448 ext4_mark_inode_dirty(handle, inode);
1449
1450 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1451 /* if we have allocated more blocks and copied
1452 * less. We will have blocks allocated outside
1453 * inode->i_size. So truncate them
1454 */
1455 ext4_orphan_add(handle, inode);
1456errout:
1457 ret2 = ext4_journal_stop(handle);
1458 if (!ret)
1459 ret = ret2;
1460
1461 if (pos + len > inode->i_size && !verity) {
1462 ext4_truncate_failed_write(inode);
1463 /*
1464 * If truncate failed early the inode might still be
1465 * on the orphan list; we need to make sure the inode
1466 * is removed from the orphan list in that case.
1467 */
1468 if (inode->i_nlink)
1469 ext4_orphan_del(NULL, inode);
1470 }
1471
1472 return ret ? ret : copied;
1473}
1474
1475/*
1476 * This is a private version of page_zero_new_buffers() which doesn't
1477 * set the buffer to be dirty, since in data=journalled mode we need
1478 * to call ext4_handle_dirty_metadata() instead.
1479 */
1480static void ext4_journalled_zero_new_buffers(handle_t *handle,
1481 struct page *page,
1482 unsigned from, unsigned to)
1483{
1484 unsigned int block_start = 0, block_end;
1485 struct buffer_head *head, *bh;
1486
1487 bh = head = page_buffers(page);
1488 do {
1489 block_end = block_start + bh->b_size;
1490 if (buffer_new(bh)) {
1491 if (block_end > from && block_start < to) {
1492 if (!PageUptodate(page)) {
1493 unsigned start, size;
1494
1495 start = max(from, block_start);
1496 size = min(to, block_end) - start;
1497
1498 zero_user(page, start, size);
1499 write_end_fn(handle, bh);
1500 }
1501 clear_buffer_new(bh);
1502 }
1503 }
1504 block_start = block_end;
1505 bh = bh->b_this_page;
1506 } while (bh != head);
1507}
1508
1509static int ext4_journalled_write_end(struct file *file,
1510 struct address_space *mapping,
1511 loff_t pos, unsigned len, unsigned copied,
1512 struct page *page, void *fsdata)
1513{
1514 handle_t *handle = ext4_journal_current_handle();
1515 struct inode *inode = mapping->host;
1516 loff_t old_size = inode->i_size;
1517 int ret = 0, ret2;
1518 int partial = 0;
1519 unsigned from, to;
1520 int size_changed = 0;
1521 int inline_data = ext4_has_inline_data(inode);
1522 bool verity = ext4_verity_in_progress(inode);
1523
1524 trace_ext4_journalled_write_end(inode, pos, len, copied);
1525 from = pos & (PAGE_SIZE - 1);
1526 to = from + len;
1527
1528 BUG_ON(!ext4_handle_valid(handle));
1529
1530 if (inline_data) {
1531 ret = ext4_write_inline_data_end(inode, pos, len,
1532 copied, page);
1533 if (ret < 0) {
1534 unlock_page(page);
1535 put_page(page);
1536 goto errout;
1537 }
1538 copied = ret;
1539 } else if (unlikely(copied < len) && !PageUptodate(page)) {
1540 copied = 0;
1541 ext4_journalled_zero_new_buffers(handle, page, from, to);
1542 } else {
1543 if (unlikely(copied < len))
1544 ext4_journalled_zero_new_buffers(handle, page,
1545 from + copied, to);
1546 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1547 from + copied, &partial,
1548 write_end_fn);
1549 if (!partial)
1550 SetPageUptodate(page);
1551 }
1552 if (!verity)
1553 size_changed = ext4_update_inode_size(inode, pos + copied);
1554 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1555 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1556 unlock_page(page);
1557 put_page(page);
1558
1559 if (old_size < pos && !verity)
1560 pagecache_isize_extended(inode, old_size, pos);
1561
1562 if (size_changed || inline_data) {
1563 ret2 = ext4_mark_inode_dirty(handle, inode);
1564 if (!ret)
1565 ret = ret2;
1566 }
1567
1568 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1569 /* if we have allocated more blocks and copied
1570 * less. We will have blocks allocated outside
1571 * inode->i_size. So truncate them
1572 */
1573 ext4_orphan_add(handle, inode);
1574
1575errout:
1576 ret2 = ext4_journal_stop(handle);
1577 if (!ret)
1578 ret = ret2;
1579 if (pos + len > inode->i_size && !verity) {
1580 ext4_truncate_failed_write(inode);
1581 /*
1582 * If truncate failed early the inode might still be
1583 * on the orphan list; we need to make sure the inode
1584 * is removed from the orphan list in that case.
1585 */
1586 if (inode->i_nlink)
1587 ext4_orphan_del(NULL, inode);
1588 }
1589
1590 return ret ? ret : copied;
1591}
1592
1593/*
1594 * Reserve space for a single cluster
1595 */
1596static int ext4_da_reserve_space(struct inode *inode)
1597{
1598 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1599 struct ext4_inode_info *ei = EXT4_I(inode);
1600 int ret;
1601
1602 /*
1603 * We will charge metadata quota at writeout time; this saves
1604 * us from metadata over-estimation, though we may go over by
1605 * a small amount in the end. Here we just reserve for data.
1606 */
1607 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1608 if (ret)
1609 return ret;
1610
1611 spin_lock(&ei->i_block_reservation_lock);
1612 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1613 spin_unlock(&ei->i_block_reservation_lock);
1614 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1615 return -ENOSPC;
1616 }
1617 ei->i_reserved_data_blocks++;
1618 trace_ext4_da_reserve_space(inode);
1619 spin_unlock(&ei->i_block_reservation_lock);
1620
1621 return 0; /* success */
1622}
1623
1624void ext4_da_release_space(struct inode *inode, int to_free)
1625{
1626 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1627 struct ext4_inode_info *ei = EXT4_I(inode);
1628
1629 if (!to_free)
1630 return; /* Nothing to release, exit */
1631
1632 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1633
1634 trace_ext4_da_release_space(inode, to_free);
1635 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1636 /*
1637 * if there aren't enough reserved blocks, then the
1638 * counter is messed up somewhere. Since this
1639 * function is called from invalidate page, it's
1640 * harmless to return without any action.
1641 */
1642 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1643 "ino %lu, to_free %d with only %d reserved "
1644 "data blocks", inode->i_ino, to_free,
1645 ei->i_reserved_data_blocks);
1646 WARN_ON(1);
1647 to_free = ei->i_reserved_data_blocks;
1648 }
1649 ei->i_reserved_data_blocks -= to_free;
1650
1651 /* update fs dirty data blocks counter */
1652 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1653
1654 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1655
1656 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1657}
1658
1659/*
1660 * Delayed allocation stuff
1661 */
1662
1663struct mpage_da_data {
1664 struct inode *inode;
1665 struct writeback_control *wbc;
1666
1667 pgoff_t first_page; /* The first page to write */
1668 pgoff_t next_page; /* Current page to examine */
1669 pgoff_t last_page; /* Last page to examine */
1670 /*
1671 * Extent to map - this can be after first_page because that can be
1672 * fully mapped. We somewhat abuse m_flags to store whether the extent
1673 * is delalloc or unwritten.
1674 */
1675 struct ext4_map_blocks map;
1676 struct ext4_io_submit io_submit; /* IO submission data */
1677 unsigned int do_map:1;
1678};
1679
1680static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1681 bool invalidate)
1682{
1683 int nr_pages, i;
1684 pgoff_t index, end;
1685 struct pagevec pvec;
1686 struct inode *inode = mpd->inode;
1687 struct address_space *mapping = inode->i_mapping;
1688
1689 /* This is necessary when next_page == 0. */
1690 if (mpd->first_page >= mpd->next_page)
1691 return;
1692
1693 index = mpd->first_page;
1694 end = mpd->next_page - 1;
1695 if (invalidate) {
1696 ext4_lblk_t start, last;
1697 start = index << (PAGE_SHIFT - inode->i_blkbits);
1698 last = end << (PAGE_SHIFT - inode->i_blkbits);
1699 ext4_es_remove_extent(inode, start, last - start + 1);
1700 }
1701
1702 pagevec_init(&pvec);
1703 while (index <= end) {
1704 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1705 if (nr_pages == 0)
1706 break;
1707 for (i = 0; i < nr_pages; i++) {
1708 struct page *page = pvec.pages[i];
1709
1710 BUG_ON(!PageLocked(page));
1711 BUG_ON(PageWriteback(page));
1712 if (invalidate) {
1713 if (page_mapped(page))
1714 clear_page_dirty_for_io(page);
1715 block_invalidatepage(page, 0, PAGE_SIZE);
1716 ClearPageUptodate(page);
1717 }
1718 unlock_page(page);
1719 }
1720 pagevec_release(&pvec);
1721 }
1722}
1723
1724static void ext4_print_free_blocks(struct inode *inode)
1725{
1726 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1727 struct super_block *sb = inode->i_sb;
1728 struct ext4_inode_info *ei = EXT4_I(inode);
1729
1730 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1731 EXT4_C2B(EXT4_SB(inode->i_sb),
1732 ext4_count_free_clusters(sb)));
1733 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1734 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1735 (long long) EXT4_C2B(EXT4_SB(sb),
1736 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1737 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1738 (long long) EXT4_C2B(EXT4_SB(sb),
1739 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1740 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1741 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1742 ei->i_reserved_data_blocks);
1743 return;
1744}
1745
1746static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1747{
1748 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1749}
1750
1751/*
1752 * ext4_insert_delayed_block - adds a delayed block to the extents status
1753 * tree, incrementing the reserved cluster/block
1754 * count or making a pending reservation
1755 * where needed
1756 *
1757 * @inode - file containing the newly added block
1758 * @lblk - logical block to be added
1759 *
1760 * Returns 0 on success, negative error code on failure.
1761 */
1762static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1763{
1764 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1765 int ret;
1766 bool allocated = false;
1767
1768 /*
1769 * If the cluster containing lblk is shared with a delayed,
1770 * written, or unwritten extent in a bigalloc file system, it's
1771 * already been accounted for and does not need to be reserved.
1772 * A pending reservation must be made for the cluster if it's
1773 * shared with a written or unwritten extent and doesn't already
1774 * have one. Written and unwritten extents can be purged from the
1775 * extents status tree if the system is under memory pressure, so
1776 * it's necessary to examine the extent tree if a search of the
1777 * extents status tree doesn't get a match.
1778 */
1779 if (sbi->s_cluster_ratio == 1) {
1780 ret = ext4_da_reserve_space(inode);
1781 if (ret != 0) /* ENOSPC */
1782 goto errout;
1783 } else { /* bigalloc */
1784 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1785 if (!ext4_es_scan_clu(inode,
1786 &ext4_es_is_mapped, lblk)) {
1787 ret = ext4_clu_mapped(inode,
1788 EXT4_B2C(sbi, lblk));
1789 if (ret < 0)
1790 goto errout;
1791 if (ret == 0) {
1792 ret = ext4_da_reserve_space(inode);
1793 if (ret != 0) /* ENOSPC */
1794 goto errout;
1795 } else {
1796 allocated = true;
1797 }
1798 } else {
1799 allocated = true;
1800 }
1801 }
1802 }
1803
1804 ret = ext4_es_insert_delayed_block(inode, lblk, allocated);
1805
1806errout:
1807 return ret;
1808}
1809
1810/*
1811 * This function is grabs code from the very beginning of
1812 * ext4_map_blocks, but assumes that the caller is from delayed write
1813 * time. This function looks up the requested blocks and sets the
1814 * buffer delay bit under the protection of i_data_sem.
1815 */
1816static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1817 struct ext4_map_blocks *map,
1818 struct buffer_head *bh)
1819{
1820 struct extent_status es;
1821 int retval;
1822 sector_t invalid_block = ~((sector_t) 0xffff);
1823#ifdef ES_AGGRESSIVE_TEST
1824 struct ext4_map_blocks orig_map;
1825
1826 memcpy(&orig_map, map, sizeof(*map));
1827#endif
1828
1829 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1830 invalid_block = ~0;
1831
1832 map->m_flags = 0;
1833 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1834 "logical block %lu\n", inode->i_ino, map->m_len,
1835 (unsigned long) map->m_lblk);
1836
1837 /* Lookup extent status tree firstly */
1838 if (ext4_es_lookup_extent(inode, iblock, NULL, &es)) {
1839 if (ext4_es_is_hole(&es)) {
1840 retval = 0;
1841 down_read(&EXT4_I(inode)->i_data_sem);
1842 goto add_delayed;
1843 }
1844
1845 /*
1846 * Delayed extent could be allocated by fallocate.
1847 * So we need to check it.
1848 */
1849 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1850 map_bh(bh, inode->i_sb, invalid_block);
1851 set_buffer_new(bh);
1852 set_buffer_delay(bh);
1853 return 0;
1854 }
1855
1856 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1857 retval = es.es_len - (iblock - es.es_lblk);
1858 if (retval > map->m_len)
1859 retval = map->m_len;
1860 map->m_len = retval;
1861 if (ext4_es_is_written(&es))
1862 map->m_flags |= EXT4_MAP_MAPPED;
1863 else if (ext4_es_is_unwritten(&es))
1864 map->m_flags |= EXT4_MAP_UNWRITTEN;
1865 else
1866 BUG();
1867
1868#ifdef ES_AGGRESSIVE_TEST
1869 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1870#endif
1871 return retval;
1872 }
1873
1874 /*
1875 * Try to see if we can get the block without requesting a new
1876 * file system block.
1877 */
1878 down_read(&EXT4_I(inode)->i_data_sem);
1879 if (ext4_has_inline_data(inode))
1880 retval = 0;
1881 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1882 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1883 else
1884 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1885
1886add_delayed:
1887 if (retval == 0) {
1888 int ret;
1889
1890 /*
1891 * XXX: __block_prepare_write() unmaps passed block,
1892 * is it OK?
1893 */
1894
1895 ret = ext4_insert_delayed_block(inode, map->m_lblk);
1896 if (ret != 0) {
1897 retval = ret;
1898 goto out_unlock;
1899 }
1900
1901 map_bh(bh, inode->i_sb, invalid_block);
1902 set_buffer_new(bh);
1903 set_buffer_delay(bh);
1904 } else if (retval > 0) {
1905 int ret;
1906 unsigned int status;
1907
1908 if (unlikely(retval != map->m_len)) {
1909 ext4_warning(inode->i_sb,
1910 "ES len assertion failed for inode "
1911 "%lu: retval %d != map->m_len %d",
1912 inode->i_ino, retval, map->m_len);
1913 WARN_ON(1);
1914 }
1915
1916 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1917 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1918 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1919 map->m_pblk, status);
1920 if (ret != 0)
1921 retval = ret;
1922 }
1923
1924out_unlock:
1925 up_read((&EXT4_I(inode)->i_data_sem));
1926
1927 return retval;
1928}
1929
1930/*
1931 * This is a special get_block_t callback which is used by
1932 * ext4_da_write_begin(). It will either return mapped block or
1933 * reserve space for a single block.
1934 *
1935 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1936 * We also have b_blocknr = -1 and b_bdev initialized properly
1937 *
1938 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1939 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1940 * initialized properly.
1941 */
1942int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1943 struct buffer_head *bh, int create)
1944{
1945 struct ext4_map_blocks map;
1946 int ret = 0;
1947
1948 BUG_ON(create == 0);
1949 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1950
1951 map.m_lblk = iblock;
1952 map.m_len = 1;
1953
1954 /*
1955 * first, we need to know whether the block is allocated already
1956 * preallocated blocks are unmapped but should treated
1957 * the same as allocated blocks.
1958 */
1959 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1960 if (ret <= 0)
1961 return ret;
1962
1963 map_bh(bh, inode->i_sb, map.m_pblk);
1964 ext4_update_bh_state(bh, map.m_flags);
1965
1966 if (buffer_unwritten(bh)) {
1967 /* A delayed write to unwritten bh should be marked
1968 * new and mapped. Mapped ensures that we don't do
1969 * get_block multiple times when we write to the same
1970 * offset and new ensures that we do proper zero out
1971 * for partial write.
1972 */
1973 set_buffer_new(bh);
1974 set_buffer_mapped(bh);
1975 }
1976 return 0;
1977}
1978
1979static int bget_one(handle_t *handle, struct buffer_head *bh)
1980{
1981 get_bh(bh);
1982 return 0;
1983}
1984
1985static int bput_one(handle_t *handle, struct buffer_head *bh)
1986{
1987 put_bh(bh);
1988 return 0;
1989}
1990
1991static int __ext4_journalled_writepage(struct page *page,
1992 unsigned int len)
1993{
1994 struct address_space *mapping = page->mapping;
1995 struct inode *inode = mapping->host;
1996 struct buffer_head *page_bufs = NULL;
1997 handle_t *handle = NULL;
1998 int ret = 0, err = 0;
1999 int inline_data = ext4_has_inline_data(inode);
2000 struct buffer_head *inode_bh = NULL;
2001
2002 ClearPageChecked(page);
2003
2004 if (inline_data) {
2005 BUG_ON(page->index != 0);
2006 BUG_ON(len > ext4_get_max_inline_size(inode));
2007 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
2008 if (inode_bh == NULL)
2009 goto out;
2010 } else {
2011 page_bufs = page_buffers(page);
2012 if (!page_bufs) {
2013 BUG();
2014 goto out;
2015 }
2016 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2017 NULL, bget_one);
2018 }
2019 /*
2020 * We need to release the page lock before we start the
2021 * journal, so grab a reference so the page won't disappear
2022 * out from under us.
2023 */
2024 get_page(page);
2025 unlock_page(page);
2026
2027 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2028 ext4_writepage_trans_blocks(inode));
2029 if (IS_ERR(handle)) {
2030 ret = PTR_ERR(handle);
2031 put_page(page);
2032 goto out_no_pagelock;
2033 }
2034 BUG_ON(!ext4_handle_valid(handle));
2035
2036 lock_page(page);
2037 put_page(page);
2038 if (page->mapping != mapping) {
2039 /* The page got truncated from under us */
2040 ext4_journal_stop(handle);
2041 ret = 0;
2042 goto out;
2043 }
2044
2045 if (inline_data) {
2046 ret = ext4_mark_inode_dirty(handle, inode);
2047 } else {
2048 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2049 do_journal_get_write_access);
2050
2051 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2052 write_end_fn);
2053 }
2054 if (ret == 0)
2055 ret = err;
2056 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2057 err = ext4_journal_stop(handle);
2058 if (!ret)
2059 ret = err;
2060
2061 if (!ext4_has_inline_data(inode))
2062 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2063 NULL, bput_one);
2064 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2065out:
2066 unlock_page(page);
2067out_no_pagelock:
2068 brelse(inode_bh);
2069 return ret;
2070}
2071
2072/*
2073 * Note that we don't need to start a transaction unless we're journaling data
2074 * because we should have holes filled from ext4_page_mkwrite(). We even don't
2075 * need to file the inode to the transaction's list in ordered mode because if
2076 * we are writing back data added by write(), the inode is already there and if
2077 * we are writing back data modified via mmap(), no one guarantees in which
2078 * transaction the data will hit the disk. In case we are journaling data, we
2079 * cannot start transaction directly because transaction start ranks above page
2080 * lock so we have to do some magic.
2081 *
2082 * This function can get called via...
2083 * - ext4_writepages after taking page lock (have journal handle)
2084 * - journal_submit_inode_data_buffers (no journal handle)
2085 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2086 * - grab_page_cache when doing write_begin (have journal handle)
2087 *
2088 * We don't do any block allocation in this function. If we have page with
2089 * multiple blocks we need to write those buffer_heads that are mapped. This
2090 * is important for mmaped based write. So if we do with blocksize 1K
2091 * truncate(f, 1024);
2092 * a = mmap(f, 0, 4096);
2093 * a[0] = 'a';
2094 * truncate(f, 4096);
2095 * we have in the page first buffer_head mapped via page_mkwrite call back
2096 * but other buffer_heads would be unmapped but dirty (dirty done via the
2097 * do_wp_page). So writepage should write the first block. If we modify
2098 * the mmap area beyond 1024 we will again get a page_fault and the
2099 * page_mkwrite callback will do the block allocation and mark the
2100 * buffer_heads mapped.
2101 *
2102 * We redirty the page if we have any buffer_heads that is either delay or
2103 * unwritten in the page.
2104 *
2105 * We can get recursively called as show below.
2106 *
2107 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2108 * ext4_writepage()
2109 *
2110 * But since we don't do any block allocation we should not deadlock.
2111 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2112 */
2113static int ext4_writepage(struct page *page,
2114 struct writeback_control *wbc)
2115{
2116 int ret = 0;
2117 loff_t size;
2118 unsigned int len;
2119 struct buffer_head *page_bufs = NULL;
2120 struct inode *inode = page->mapping->host;
2121 struct ext4_io_submit io_submit;
2122 bool keep_towrite = false;
2123
2124 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2125 ext4_invalidatepage(page, 0, PAGE_SIZE);
2126 unlock_page(page);
2127 return -EIO;
2128 }
2129
2130 trace_ext4_writepage(page);
2131 size = i_size_read(inode);
2132 if (page->index == size >> PAGE_SHIFT &&
2133 !ext4_verity_in_progress(inode))
2134 len = size & ~PAGE_MASK;
2135 else
2136 len = PAGE_SIZE;
2137
2138 page_bufs = page_buffers(page);
2139 /*
2140 * We cannot do block allocation or other extent handling in this
2141 * function. If there are buffers needing that, we have to redirty
2142 * the page. But we may reach here when we do a journal commit via
2143 * journal_submit_inode_data_buffers() and in that case we must write
2144 * allocated buffers to achieve data=ordered mode guarantees.
2145 *
2146 * Also, if there is only one buffer per page (the fs block
2147 * size == the page size), if one buffer needs block
2148 * allocation or needs to modify the extent tree to clear the
2149 * unwritten flag, we know that the page can't be written at
2150 * all, so we might as well refuse the write immediately.
2151 * Unfortunately if the block size != page size, we can't as
2152 * easily detect this case using ext4_walk_page_buffers(), but
2153 * for the extremely common case, this is an optimization that
2154 * skips a useless round trip through ext4_bio_write_page().
2155 */
2156 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2157 ext4_bh_delay_or_unwritten)) {
2158 redirty_page_for_writepage(wbc, page);
2159 if ((current->flags & PF_MEMALLOC) ||
2160 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2161 /*
2162 * For memory cleaning there's no point in writing only
2163 * some buffers. So just bail out. Warn if we came here
2164 * from direct reclaim.
2165 */
2166 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2167 == PF_MEMALLOC);
2168 unlock_page(page);
2169 return 0;
2170 }
2171 keep_towrite = true;
2172 }
2173
2174 if (PageChecked(page) && ext4_should_journal_data(inode))
2175 /*
2176 * It's mmapped pagecache. Add buffers and journal it. There
2177 * doesn't seem much point in redirtying the page here.
2178 */
2179 return __ext4_journalled_writepage(page, len);
2180
2181 ext4_io_submit_init(&io_submit, wbc);
2182 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2183 if (!io_submit.io_end) {
2184 redirty_page_for_writepage(wbc, page);
2185 unlock_page(page);
2186 return -ENOMEM;
2187 }
2188 ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2189 ext4_io_submit(&io_submit);
2190 /* Drop io_end reference we got from init */
2191 ext4_put_io_end_defer(io_submit.io_end);
2192 return ret;
2193}
2194
2195static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2196{
2197 int len;
2198 loff_t size;
2199 int err;
2200
2201 BUG_ON(page->index != mpd->first_page);
2202 clear_page_dirty_for_io(page);
2203 /*
2204 * We have to be very careful here! Nothing protects writeback path
2205 * against i_size changes and the page can be writeably mapped into
2206 * page tables. So an application can be growing i_size and writing
2207 * data through mmap while writeback runs. clear_page_dirty_for_io()
2208 * write-protects our page in page tables and the page cannot get
2209 * written to again until we release page lock. So only after
2210 * clear_page_dirty_for_io() we are safe to sample i_size for
2211 * ext4_bio_write_page() to zero-out tail of the written page. We rely
2212 * on the barrier provided by TestClearPageDirty in
2213 * clear_page_dirty_for_io() to make sure i_size is really sampled only
2214 * after page tables are updated.
2215 */
2216 size = i_size_read(mpd->inode);
2217 if (page->index == size >> PAGE_SHIFT &&
2218 !ext4_verity_in_progress(mpd->inode))
2219 len = size & ~PAGE_MASK;
2220 else
2221 len = PAGE_SIZE;
2222 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2223 if (!err)
2224 mpd->wbc->nr_to_write--;
2225 mpd->first_page++;
2226
2227 return err;
2228}
2229
2230#define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2231
2232/*
2233 * mballoc gives us at most this number of blocks...
2234 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2235 * The rest of mballoc seems to handle chunks up to full group size.
2236 */
2237#define MAX_WRITEPAGES_EXTENT_LEN 2048
2238
2239/*
2240 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2241 *
2242 * @mpd - extent of blocks
2243 * @lblk - logical number of the block in the file
2244 * @bh - buffer head we want to add to the extent
2245 *
2246 * The function is used to collect contig. blocks in the same state. If the
2247 * buffer doesn't require mapping for writeback and we haven't started the
2248 * extent of buffers to map yet, the function returns 'true' immediately - the
2249 * caller can write the buffer right away. Otherwise the function returns true
2250 * if the block has been added to the extent, false if the block couldn't be
2251 * added.
2252 */
2253static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2254 struct buffer_head *bh)
2255{
2256 struct ext4_map_blocks *map = &mpd->map;
2257
2258 /* Buffer that doesn't need mapping for writeback? */
2259 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2260 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2261 /* So far no extent to map => we write the buffer right away */
2262 if (map->m_len == 0)
2263 return true;
2264 return false;
2265 }
2266
2267 /* First block in the extent? */
2268 if (map->m_len == 0) {
2269 /* We cannot map unless handle is started... */
2270 if (!mpd->do_map)
2271 return false;
2272 map->m_lblk = lblk;
2273 map->m_len = 1;
2274 map->m_flags = bh->b_state & BH_FLAGS;
2275 return true;
2276 }
2277
2278 /* Don't go larger than mballoc is willing to allocate */
2279 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2280 return false;
2281
2282 /* Can we merge the block to our big extent? */
2283 if (lblk == map->m_lblk + map->m_len &&
2284 (bh->b_state & BH_FLAGS) == map->m_flags) {
2285 map->m_len++;
2286 return true;
2287 }
2288 return false;
2289}
2290
2291/*
2292 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2293 *
2294 * @mpd - extent of blocks for mapping
2295 * @head - the first buffer in the page
2296 * @bh - buffer we should start processing from
2297 * @lblk - logical number of the block in the file corresponding to @bh
2298 *
2299 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2300 * the page for IO if all buffers in this page were mapped and there's no
2301 * accumulated extent of buffers to map or add buffers in the page to the
2302 * extent of buffers to map. The function returns 1 if the caller can continue
2303 * by processing the next page, 0 if it should stop adding buffers to the
2304 * extent to map because we cannot extend it anymore. It can also return value
2305 * < 0 in case of error during IO submission.
2306 */
2307static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2308 struct buffer_head *head,
2309 struct buffer_head *bh,
2310 ext4_lblk_t lblk)
2311{
2312 struct inode *inode = mpd->inode;
2313 int err;
2314 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2315 >> inode->i_blkbits;
2316
2317 if (ext4_verity_in_progress(inode))
2318 blocks = EXT_MAX_BLOCKS;
2319
2320 do {
2321 BUG_ON(buffer_locked(bh));
2322
2323 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2324 /* Found extent to map? */
2325 if (mpd->map.m_len)
2326 return 0;
2327 /* Buffer needs mapping and handle is not started? */
2328 if (!mpd->do_map)
2329 return 0;
2330 /* Everything mapped so far and we hit EOF */
2331 break;
2332 }
2333 } while (lblk++, (bh = bh->b_this_page) != head);
2334 /* So far everything mapped? Submit the page for IO. */
2335 if (mpd->map.m_len == 0) {
2336 err = mpage_submit_page(mpd, head->b_page);
2337 if (err < 0)
2338 return err;
2339 }
2340 return lblk < blocks;
2341}
2342
2343/*
2344 * mpage_map_buffers - update buffers corresponding to changed extent and
2345 * submit fully mapped pages for IO
2346 *
2347 * @mpd - description of extent to map, on return next extent to map
2348 *
2349 * Scan buffers corresponding to changed extent (we expect corresponding pages
2350 * to be already locked) and update buffer state according to new extent state.
2351 * We map delalloc buffers to their physical location, clear unwritten bits,
2352 * and mark buffers as uninit when we perform writes to unwritten extents
2353 * and do extent conversion after IO is finished. If the last page is not fully
2354 * mapped, we update @map to the next extent in the last page that needs
2355 * mapping. Otherwise we submit the page for IO.
2356 */
2357static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2358{
2359 struct pagevec pvec;
2360 int nr_pages, i;
2361 struct inode *inode = mpd->inode;
2362 struct buffer_head *head, *bh;
2363 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2364 pgoff_t start, end;
2365 ext4_lblk_t lblk;
2366 sector_t pblock;
2367 int err;
2368
2369 start = mpd->map.m_lblk >> bpp_bits;
2370 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2371 lblk = start << bpp_bits;
2372 pblock = mpd->map.m_pblk;
2373
2374 pagevec_init(&pvec);
2375 while (start <= end) {
2376 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2377 &start, end);
2378 if (nr_pages == 0)
2379 break;
2380 for (i = 0; i < nr_pages; i++) {
2381 struct page *page = pvec.pages[i];
2382
2383 bh = head = page_buffers(page);
2384 do {
2385 if (lblk < mpd->map.m_lblk)
2386 continue;
2387 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2388 /*
2389 * Buffer after end of mapped extent.
2390 * Find next buffer in the page to map.
2391 */
2392 mpd->map.m_len = 0;
2393 mpd->map.m_flags = 0;
2394 /*
2395 * FIXME: If dioread_nolock supports
2396 * blocksize < pagesize, we need to make
2397 * sure we add size mapped so far to
2398 * io_end->size as the following call
2399 * can submit the page for IO.
2400 */
2401 err = mpage_process_page_bufs(mpd, head,
2402 bh, lblk);
2403 pagevec_release(&pvec);
2404 if (err > 0)
2405 err = 0;
2406 return err;
2407 }
2408 if (buffer_delay(bh)) {
2409 clear_buffer_delay(bh);
2410 bh->b_blocknr = pblock++;
2411 }
2412 clear_buffer_unwritten(bh);
2413 } while (lblk++, (bh = bh->b_this_page) != head);
2414
2415 /*
2416 * FIXME: This is going to break if dioread_nolock
2417 * supports blocksize < pagesize as we will try to
2418 * convert potentially unmapped parts of inode.
2419 */
2420 mpd->io_submit.io_end->size += PAGE_SIZE;
2421 /* Page fully mapped - let IO run! */
2422 err = mpage_submit_page(mpd, page);
2423 if (err < 0) {
2424 pagevec_release(&pvec);
2425 return err;
2426 }
2427 }
2428 pagevec_release(&pvec);
2429 }
2430 /* Extent fully mapped and matches with page boundary. We are done. */
2431 mpd->map.m_len = 0;
2432 mpd->map.m_flags = 0;
2433 return 0;
2434}
2435
2436static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2437{
2438 struct inode *inode = mpd->inode;
2439 struct ext4_map_blocks *map = &mpd->map;
2440 int get_blocks_flags;
2441 int err, dioread_nolock;
2442
2443 trace_ext4_da_write_pages_extent(inode, map);
2444 /*
2445 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2446 * to convert an unwritten extent to be initialized (in the case
2447 * where we have written into one or more preallocated blocks). It is
2448 * possible that we're going to need more metadata blocks than
2449 * previously reserved. However we must not fail because we're in
2450 * writeback and there is nothing we can do about it so it might result
2451 * in data loss. So use reserved blocks to allocate metadata if
2452 * possible.
2453 *
2454 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2455 * the blocks in question are delalloc blocks. This indicates
2456 * that the blocks and quotas has already been checked when
2457 * the data was copied into the page cache.
2458 */
2459 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2460 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2461 EXT4_GET_BLOCKS_IO_SUBMIT;
2462 dioread_nolock = ext4_should_dioread_nolock(inode);
2463 if (dioread_nolock)
2464 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2465 if (map->m_flags & (1 << BH_Delay))
2466 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2467
2468 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2469 if (err < 0)
2470 return err;
2471 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2472 if (!mpd->io_submit.io_end->handle &&
2473 ext4_handle_valid(handle)) {
2474 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2475 handle->h_rsv_handle = NULL;
2476 }
2477 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2478 }
2479
2480 BUG_ON(map->m_len == 0);
2481 return 0;
2482}
2483
2484/*
2485 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2486 * mpd->len and submit pages underlying it for IO
2487 *
2488 * @handle - handle for journal operations
2489 * @mpd - extent to map
2490 * @give_up_on_write - we set this to true iff there is a fatal error and there
2491 * is no hope of writing the data. The caller should discard
2492 * dirty pages to avoid infinite loops.
2493 *
2494 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2495 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2496 * them to initialized or split the described range from larger unwritten
2497 * extent. Note that we need not map all the described range since allocation
2498 * can return less blocks or the range is covered by more unwritten extents. We
2499 * cannot map more because we are limited by reserved transaction credits. On
2500 * the other hand we always make sure that the last touched page is fully
2501 * mapped so that it can be written out (and thus forward progress is
2502 * guaranteed). After mapping we submit all mapped pages for IO.
2503 */
2504static int mpage_map_and_submit_extent(handle_t *handle,
2505 struct mpage_da_data *mpd,
2506 bool *give_up_on_write)
2507{
2508 struct inode *inode = mpd->inode;
2509 struct ext4_map_blocks *map = &mpd->map;
2510 int err;
2511 loff_t disksize;
2512 int progress = 0;
2513
2514 mpd->io_submit.io_end->offset =
2515 ((loff_t)map->m_lblk) << inode->i_blkbits;
2516 do {
2517 err = mpage_map_one_extent(handle, mpd);
2518 if (err < 0) {
2519 struct super_block *sb = inode->i_sb;
2520
2521 if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2522 EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2523 goto invalidate_dirty_pages;
2524 /*
2525 * Let the uper layers retry transient errors.
2526 * In the case of ENOSPC, if ext4_count_free_blocks()
2527 * is non-zero, a commit should free up blocks.
2528 */
2529 if ((err == -ENOMEM) ||
2530 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2531 if (progress)
2532 goto update_disksize;
2533 return err;
2534 }
2535 ext4_msg(sb, KERN_CRIT,
2536 "Delayed block allocation failed for "
2537 "inode %lu at logical offset %llu with"
2538 " max blocks %u with error %d",
2539 inode->i_ino,
2540 (unsigned long long)map->m_lblk,
2541 (unsigned)map->m_len, -err);
2542 ext4_msg(sb, KERN_CRIT,
2543 "This should not happen!! Data will "
2544 "be lost\n");
2545 if (err == -ENOSPC)
2546 ext4_print_free_blocks(inode);
2547 invalidate_dirty_pages:
2548 *give_up_on_write = true;
2549 return err;
2550 }
2551 progress = 1;
2552 /*
2553 * Update buffer state, submit mapped pages, and get us new
2554 * extent to map
2555 */
2556 err = mpage_map_and_submit_buffers(mpd);
2557 if (err < 0)
2558 goto update_disksize;
2559 } while (map->m_len);
2560
2561update_disksize:
2562 /*
2563 * Update on-disk size after IO is submitted. Races with
2564 * truncate are avoided by checking i_size under i_data_sem.
2565 */
2566 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2567 if (disksize > EXT4_I(inode)->i_disksize) {
2568 int err2;
2569 loff_t i_size;
2570
2571 down_write(&EXT4_I(inode)->i_data_sem);
2572 i_size = i_size_read(inode);
2573 if (disksize > i_size)
2574 disksize = i_size;
2575 if (disksize > EXT4_I(inode)->i_disksize)
2576 EXT4_I(inode)->i_disksize = disksize;
2577 up_write(&EXT4_I(inode)->i_data_sem);
2578 err2 = ext4_mark_inode_dirty(handle, inode);
2579 if (err2)
2580 ext4_error(inode->i_sb,
2581 "Failed to mark inode %lu dirty",
2582 inode->i_ino);
2583 if (!err)
2584 err = err2;
2585 }
2586 return err;
2587}
2588
2589/*
2590 * Calculate the total number of credits to reserve for one writepages
2591 * iteration. This is called from ext4_writepages(). We map an extent of
2592 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2593 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2594 * bpp - 1 blocks in bpp different extents.
2595 */
2596static int ext4_da_writepages_trans_blocks(struct inode *inode)
2597{
2598 int bpp = ext4_journal_blocks_per_page(inode);
2599
2600 return ext4_meta_trans_blocks(inode,
2601 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2602}
2603
2604/*
2605 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2606 * and underlying extent to map
2607 *
2608 * @mpd - where to look for pages
2609 *
2610 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2611 * IO immediately. When we find a page which isn't mapped we start accumulating
2612 * extent of buffers underlying these pages that needs mapping (formed by
2613 * either delayed or unwritten buffers). We also lock the pages containing
2614 * these buffers. The extent found is returned in @mpd structure (starting at
2615 * mpd->lblk with length mpd->len blocks).
2616 *
2617 * Note that this function can attach bios to one io_end structure which are
2618 * neither logically nor physically contiguous. Although it may seem as an
2619 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2620 * case as we need to track IO to all buffers underlying a page in one io_end.
2621 */
2622static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2623{
2624 struct address_space *mapping = mpd->inode->i_mapping;
2625 struct pagevec pvec;
2626 unsigned int nr_pages;
2627 long left = mpd->wbc->nr_to_write;
2628 pgoff_t index = mpd->first_page;
2629 pgoff_t end = mpd->last_page;
2630 xa_mark_t tag;
2631 int i, err = 0;
2632 int blkbits = mpd->inode->i_blkbits;
2633 ext4_lblk_t lblk;
2634 struct buffer_head *head;
2635
2636 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2637 tag = PAGECACHE_TAG_TOWRITE;
2638 else
2639 tag = PAGECACHE_TAG_DIRTY;
2640
2641 pagevec_init(&pvec);
2642 mpd->map.m_len = 0;
2643 mpd->next_page = index;
2644 while (index <= end) {
2645 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2646 tag);
2647 if (nr_pages == 0)
2648 goto out;
2649
2650 for (i = 0; i < nr_pages; i++) {
2651 struct page *page = pvec.pages[i];
2652
2653 /*
2654 * Accumulated enough dirty pages? This doesn't apply
2655 * to WB_SYNC_ALL mode. For integrity sync we have to
2656 * keep going because someone may be concurrently
2657 * dirtying pages, and we might have synced a lot of
2658 * newly appeared dirty pages, but have not synced all
2659 * of the old dirty pages.
2660 */
2661 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2662 goto out;
2663
2664 /* If we can't merge this page, we are done. */
2665 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2666 goto out;
2667
2668 lock_page(page);
2669 /*
2670 * If the page is no longer dirty, or its mapping no
2671 * longer corresponds to inode we are writing (which
2672 * means it has been truncated or invalidated), or the
2673 * page is already under writeback and we are not doing
2674 * a data integrity writeback, skip the page
2675 */
2676 if (!PageDirty(page) ||
2677 (PageWriteback(page) &&
2678 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2679 unlikely(page->mapping != mapping)) {
2680 unlock_page(page);
2681 continue;
2682 }
2683
2684 wait_on_page_writeback(page);
2685 BUG_ON(PageWriteback(page));
2686
2687 if (mpd->map.m_len == 0)
2688 mpd->first_page = page->index;
2689 mpd->next_page = page->index + 1;
2690 /* Add all dirty buffers to mpd */
2691 lblk = ((ext4_lblk_t)page->index) <<
2692 (PAGE_SHIFT - blkbits);
2693 head = page_buffers(page);
2694 err = mpage_process_page_bufs(mpd, head, head, lblk);
2695 if (err <= 0)
2696 goto out;
2697 err = 0;
2698 left--;
2699 }
2700 pagevec_release(&pvec);
2701 cond_resched();
2702 }
2703 return 0;
2704out:
2705 pagevec_release(&pvec);
2706 return err;
2707}
2708
2709static int ext4_writepages(struct address_space *mapping,
2710 struct writeback_control *wbc)
2711{
2712 pgoff_t writeback_index = 0;
2713 long nr_to_write = wbc->nr_to_write;
2714 int range_whole = 0;
2715 int cycled = 1;
2716 handle_t *handle = NULL;
2717 struct mpage_da_data mpd;
2718 struct inode *inode = mapping->host;
2719 int needed_blocks, rsv_blocks = 0, ret = 0;
2720 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2721 bool done;
2722 struct blk_plug plug;
2723 bool give_up_on_write = false;
2724
2725 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2726 return -EIO;
2727
2728 percpu_down_read(&sbi->s_journal_flag_rwsem);
2729 trace_ext4_writepages(inode, wbc);
2730
2731 /*
2732 * No pages to write? This is mainly a kludge to avoid starting
2733 * a transaction for special inodes like journal inode on last iput()
2734 * because that could violate lock ordering on umount
2735 */
2736 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2737 goto out_writepages;
2738
2739 if (ext4_should_journal_data(inode)) {
2740 ret = generic_writepages(mapping, wbc);
2741 goto out_writepages;
2742 }
2743
2744 /*
2745 * If the filesystem has aborted, it is read-only, so return
2746 * right away instead of dumping stack traces later on that
2747 * will obscure the real source of the problem. We test
2748 * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2749 * the latter could be true if the filesystem is mounted
2750 * read-only, and in that case, ext4_writepages should
2751 * *never* be called, so if that ever happens, we would want
2752 * the stack trace.
2753 */
2754 if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2755 sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2756 ret = -EROFS;
2757 goto out_writepages;
2758 }
2759
2760 /*
2761 * If we have inline data and arrive here, it means that
2762 * we will soon create the block for the 1st page, so
2763 * we'd better clear the inline data here.
2764 */
2765 if (ext4_has_inline_data(inode)) {
2766 /* Just inode will be modified... */
2767 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2768 if (IS_ERR(handle)) {
2769 ret = PTR_ERR(handle);
2770 goto out_writepages;
2771 }
2772 BUG_ON(ext4_test_inode_state(inode,
2773 EXT4_STATE_MAY_INLINE_DATA));
2774 ext4_destroy_inline_data(handle, inode);
2775 ext4_journal_stop(handle);
2776 }
2777
2778 if (ext4_should_dioread_nolock(inode)) {
2779 /*
2780 * We may need to convert up to one extent per block in
2781 * the page and we may dirty the inode.
2782 */
2783 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2784 PAGE_SIZE >> inode->i_blkbits);
2785 }
2786
2787 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2788 range_whole = 1;
2789
2790 if (wbc->range_cyclic) {
2791 writeback_index = mapping->writeback_index;
2792 if (writeback_index)
2793 cycled = 0;
2794 mpd.first_page = writeback_index;
2795 mpd.last_page = -1;
2796 } else {
2797 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2798 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2799 }
2800
2801 mpd.inode = inode;
2802 mpd.wbc = wbc;
2803 ext4_io_submit_init(&mpd.io_submit, wbc);
2804retry:
2805 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2806 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2807 done = false;
2808 blk_start_plug(&plug);
2809
2810 /*
2811 * First writeback pages that don't need mapping - we can avoid
2812 * starting a transaction unnecessarily and also avoid being blocked
2813 * in the block layer on device congestion while having transaction
2814 * started.
2815 */
2816 mpd.do_map = 0;
2817 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2818 if (!mpd.io_submit.io_end) {
2819 ret = -ENOMEM;
2820 goto unplug;
2821 }
2822 ret = mpage_prepare_extent_to_map(&mpd);
2823 /* Unlock pages we didn't use */
2824 mpage_release_unused_pages(&mpd, false);
2825 /* Submit prepared bio */
2826 ext4_io_submit(&mpd.io_submit);
2827 ext4_put_io_end_defer(mpd.io_submit.io_end);
2828 mpd.io_submit.io_end = NULL;
2829 if (ret < 0)
2830 goto unplug;
2831
2832 while (!done && mpd.first_page <= mpd.last_page) {
2833 /* For each extent of pages we use new io_end */
2834 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2835 if (!mpd.io_submit.io_end) {
2836 ret = -ENOMEM;
2837 break;
2838 }
2839
2840 /*
2841 * We have two constraints: We find one extent to map and we
2842 * must always write out whole page (makes a difference when
2843 * blocksize < pagesize) so that we don't block on IO when we
2844 * try to write out the rest of the page. Journalled mode is
2845 * not supported by delalloc.
2846 */
2847 BUG_ON(ext4_should_journal_data(inode));
2848 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2849
2850 /* start a new transaction */
2851 handle = ext4_journal_start_with_reserve(inode,
2852 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2853 if (IS_ERR(handle)) {
2854 ret = PTR_ERR(handle);
2855 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2856 "%ld pages, ino %lu; err %d", __func__,
2857 wbc->nr_to_write, inode->i_ino, ret);
2858 /* Release allocated io_end */
2859 ext4_put_io_end(mpd.io_submit.io_end);
2860 mpd.io_submit.io_end = NULL;
2861 break;
2862 }
2863 mpd.do_map = 1;
2864
2865 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2866 ret = mpage_prepare_extent_to_map(&mpd);
2867 if (!ret) {
2868 if (mpd.map.m_len)
2869 ret = mpage_map_and_submit_extent(handle, &mpd,
2870 &give_up_on_write);
2871 else {
2872 /*
2873 * We scanned the whole range (or exhausted
2874 * nr_to_write), submitted what was mapped and
2875 * didn't find anything needing mapping. We are
2876 * done.
2877 */
2878 done = true;
2879 }
2880 }
2881 /*
2882 * Caution: If the handle is synchronous,
2883 * ext4_journal_stop() can wait for transaction commit
2884 * to finish which may depend on writeback of pages to
2885 * complete or on page lock to be released. In that
2886 * case, we have to wait until after after we have
2887 * submitted all the IO, released page locks we hold,
2888 * and dropped io_end reference (for extent conversion
2889 * to be able to complete) before stopping the handle.
2890 */
2891 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2892 ext4_journal_stop(handle);
2893 handle = NULL;
2894 mpd.do_map = 0;
2895 }
2896 /* Unlock pages we didn't use */
2897 mpage_release_unused_pages(&mpd, give_up_on_write);
2898 /* Submit prepared bio */
2899 ext4_io_submit(&mpd.io_submit);
2900
2901 /*
2902 * Drop our io_end reference we got from init. We have
2903 * to be careful and use deferred io_end finishing if
2904 * we are still holding the transaction as we can
2905 * release the last reference to io_end which may end
2906 * up doing unwritten extent conversion.
2907 */
2908 if (handle) {
2909 ext4_put_io_end_defer(mpd.io_submit.io_end);
2910 ext4_journal_stop(handle);
2911 } else
2912 ext4_put_io_end(mpd.io_submit.io_end);
2913 mpd.io_submit.io_end = NULL;
2914
2915 if (ret == -ENOSPC && sbi->s_journal) {
2916 /*
2917 * Commit the transaction which would
2918 * free blocks released in the transaction
2919 * and try again
2920 */
2921 jbd2_journal_force_commit_nested(sbi->s_journal);
2922 ret = 0;
2923 continue;
2924 }
2925 /* Fatal error - ENOMEM, EIO... */
2926 if (ret)
2927 break;
2928 }
2929unplug:
2930 blk_finish_plug(&plug);
2931 if (!ret && !cycled && wbc->nr_to_write > 0) {
2932 cycled = 1;
2933 mpd.last_page = writeback_index - 1;
2934 mpd.first_page = 0;
2935 goto retry;
2936 }
2937
2938 /* Update index */
2939 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2940 /*
2941 * Set the writeback_index so that range_cyclic
2942 * mode will write it back later
2943 */
2944 mapping->writeback_index = mpd.first_page;
2945
2946out_writepages:
2947 trace_ext4_writepages_result(inode, wbc, ret,
2948 nr_to_write - wbc->nr_to_write);
2949 percpu_up_read(&sbi->s_journal_flag_rwsem);
2950 return ret;
2951}
2952
2953static int ext4_dax_writepages(struct address_space *mapping,
2954 struct writeback_control *wbc)
2955{
2956 int ret;
2957 long nr_to_write = wbc->nr_to_write;
2958 struct inode *inode = mapping->host;
2959 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2960
2961 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2962 return -EIO;
2963
2964 percpu_down_read(&sbi->s_journal_flag_rwsem);
2965 trace_ext4_writepages(inode, wbc);
2966
2967 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev, wbc);
2968 trace_ext4_writepages_result(inode, wbc, ret,
2969 nr_to_write - wbc->nr_to_write);
2970 percpu_up_read(&sbi->s_journal_flag_rwsem);
2971 return ret;
2972}
2973
2974static int ext4_nonda_switch(struct super_block *sb)
2975{
2976 s64 free_clusters, dirty_clusters;
2977 struct ext4_sb_info *sbi = EXT4_SB(sb);
2978
2979 /*
2980 * switch to non delalloc mode if we are running low
2981 * on free block. The free block accounting via percpu
2982 * counters can get slightly wrong with percpu_counter_batch getting
2983 * accumulated on each CPU without updating global counters
2984 * Delalloc need an accurate free block accounting. So switch
2985 * to non delalloc when we are near to error range.
2986 */
2987 free_clusters =
2988 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2989 dirty_clusters =
2990 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2991 /*
2992 * Start pushing delalloc when 1/2 of free blocks are dirty.
2993 */
2994 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2995 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2996
2997 if (2 * free_clusters < 3 * dirty_clusters ||
2998 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2999 /*
3000 * free block count is less than 150% of dirty blocks
3001 * or free blocks is less than watermark
3002 */
3003 return 1;
3004 }
3005 return 0;
3006}
3007
3008/* We always reserve for an inode update; the superblock could be there too */
3009static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
3010{
3011 if (likely(ext4_has_feature_large_file(inode->i_sb)))
3012 return 1;
3013
3014 if (pos + len <= 0x7fffffffULL)
3015 return 1;
3016
3017 /* We might need to update the superblock to set LARGE_FILE */
3018 return 2;
3019}
3020
3021static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3022 loff_t pos, unsigned len, unsigned flags,
3023 struct page **pagep, void **fsdata)
3024{
3025 int ret, retries = 0;
3026 struct page *page;
3027 pgoff_t index;
3028 struct inode *inode = mapping->host;
3029 handle_t *handle;
3030
3031 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3032 return -EIO;
3033
3034 index = pos >> PAGE_SHIFT;
3035
3036 if (ext4_nonda_switch(inode->i_sb) || S_ISLNK(inode->i_mode) ||
3037 ext4_verity_in_progress(inode)) {
3038 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3039 return ext4_write_begin(file, mapping, pos,
3040 len, flags, pagep, fsdata);
3041 }
3042 *fsdata = (void *)0;
3043 trace_ext4_da_write_begin(inode, pos, len, flags);
3044
3045 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3046 ret = ext4_da_write_inline_data_begin(mapping, inode,
3047 pos, len, flags,
3048 pagep, fsdata);
3049 if (ret < 0)
3050 return ret;
3051 if (ret == 1)
3052 return 0;
3053 }
3054
3055 /*
3056 * grab_cache_page_write_begin() can take a long time if the
3057 * system is thrashing due to memory pressure, or if the page
3058 * is being written back. So grab it first before we start
3059 * the transaction handle. This also allows us to allocate
3060 * the page (if needed) without using GFP_NOFS.
3061 */
3062retry_grab:
3063 page = grab_cache_page_write_begin(mapping, index, flags);
3064 if (!page)
3065 return -ENOMEM;
3066 unlock_page(page);
3067
3068 /*
3069 * With delayed allocation, we don't log the i_disksize update
3070 * if there is delayed block allocation. But we still need
3071 * to journalling the i_disksize update if writes to the end
3072 * of file which has an already mapped buffer.
3073 */
3074retry_journal:
3075 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3076 ext4_da_write_credits(inode, pos, len));
3077 if (IS_ERR(handle)) {
3078 put_page(page);
3079 return PTR_ERR(handle);
3080 }
3081
3082 lock_page(page);
3083 if (page->mapping != mapping) {
3084 /* The page got truncated from under us */
3085 unlock_page(page);
3086 put_page(page);
3087 ext4_journal_stop(handle);
3088 goto retry_grab;
3089 }
3090 /* In case writeback began while the page was unlocked */
3091 wait_for_stable_page(page);
3092
3093#ifdef CONFIG_FS_ENCRYPTION
3094 ret = ext4_block_write_begin(page, pos, len,
3095 ext4_da_get_block_prep);
3096#else
3097 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3098#endif
3099 if (ret < 0) {
3100 unlock_page(page);
3101 ext4_journal_stop(handle);
3102 /*
3103 * block_write_begin may have instantiated a few blocks
3104 * outside i_size. Trim these off again. Don't need
3105 * i_size_read because we hold i_mutex.
3106 */
3107 if (pos + len > inode->i_size)
3108 ext4_truncate_failed_write(inode);
3109
3110 if (ret == -ENOSPC &&
3111 ext4_should_retry_alloc(inode->i_sb, &retries))
3112 goto retry_journal;
3113
3114 put_page(page);
3115 return ret;
3116 }
3117
3118 *pagep = page;
3119 return ret;
3120}
3121
3122/*
3123 * Check if we should update i_disksize
3124 * when write to the end of file but not require block allocation
3125 */
3126static int ext4_da_should_update_i_disksize(struct page *page,
3127 unsigned long offset)
3128{
3129 struct buffer_head *bh;
3130 struct inode *inode = page->mapping->host;
3131 unsigned int idx;
3132 int i;
3133
3134 bh = page_buffers(page);
3135 idx = offset >> inode->i_blkbits;
3136
3137 for (i = 0; i < idx; i++)
3138 bh = bh->b_this_page;
3139
3140 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3141 return 0;
3142 return 1;
3143}
3144
3145static int ext4_da_write_end(struct file *file,
3146 struct address_space *mapping,
3147 loff_t pos, unsigned len, unsigned copied,
3148 struct page *page, void *fsdata)
3149{
3150 struct inode *inode = mapping->host;
3151 int ret = 0, ret2;
3152 handle_t *handle = ext4_journal_current_handle();
3153 loff_t new_i_size;
3154 unsigned long start, end;
3155 int write_mode = (int)(unsigned long)fsdata;
3156
3157 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3158 return ext4_write_end(file, mapping, pos,
3159 len, copied, page, fsdata);
3160
3161 trace_ext4_da_write_end(inode, pos, len, copied);
3162 start = pos & (PAGE_SIZE - 1);
3163 end = start + copied - 1;
3164
3165 /*
3166 * generic_write_end() will run mark_inode_dirty() if i_size
3167 * changes. So let's piggyback the i_disksize mark_inode_dirty
3168 * into that.
3169 */
3170 new_i_size = pos + copied;
3171 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3172 if (ext4_has_inline_data(inode) ||
3173 ext4_da_should_update_i_disksize(page, end)) {
3174 ext4_update_i_disksize(inode, new_i_size);
3175 /* We need to mark inode dirty even if
3176 * new_i_size is less that inode->i_size
3177 * bu greater than i_disksize.(hint delalloc)
3178 */
3179 ext4_mark_inode_dirty(handle, inode);
3180 }
3181 }
3182
3183 if (write_mode != CONVERT_INLINE_DATA &&
3184 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3185 ext4_has_inline_data(inode))
3186 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3187 page);
3188 else
3189 ret2 = generic_write_end(file, mapping, pos, len, copied,
3190 page, fsdata);
3191
3192 copied = ret2;
3193 if (ret2 < 0)
3194 ret = ret2;
3195 ret2 = ext4_journal_stop(handle);
3196 if (!ret)
3197 ret = ret2;
3198
3199 return ret ? ret : copied;
3200}
3201
3202/*
3203 * Force all delayed allocation blocks to be allocated for a given inode.
3204 */
3205int ext4_alloc_da_blocks(struct inode *inode)
3206{
3207 trace_ext4_alloc_da_blocks(inode);
3208
3209 if (!EXT4_I(inode)->i_reserved_data_blocks)
3210 return 0;
3211
3212 /*
3213 * We do something simple for now. The filemap_flush() will
3214 * also start triggering a write of the data blocks, which is
3215 * not strictly speaking necessary (and for users of
3216 * laptop_mode, not even desirable). However, to do otherwise
3217 * would require replicating code paths in:
3218 *
3219 * ext4_writepages() ->
3220 * write_cache_pages() ---> (via passed in callback function)
3221 * __mpage_da_writepage() -->
3222 * mpage_add_bh_to_extent()
3223 * mpage_da_map_blocks()
3224 *
3225 * The problem is that write_cache_pages(), located in
3226 * mm/page-writeback.c, marks pages clean in preparation for
3227 * doing I/O, which is not desirable if we're not planning on
3228 * doing I/O at all.
3229 *
3230 * We could call write_cache_pages(), and then redirty all of
3231 * the pages by calling redirty_page_for_writepage() but that
3232 * would be ugly in the extreme. So instead we would need to
3233 * replicate parts of the code in the above functions,
3234 * simplifying them because we wouldn't actually intend to
3235 * write out the pages, but rather only collect contiguous
3236 * logical block extents, call the multi-block allocator, and
3237 * then update the buffer heads with the block allocations.
3238 *
3239 * For now, though, we'll cheat by calling filemap_flush(),
3240 * which will map the blocks, and start the I/O, but not
3241 * actually wait for the I/O to complete.
3242 */
3243 return filemap_flush(inode->i_mapping);
3244}
3245
3246/*
3247 * bmap() is special. It gets used by applications such as lilo and by
3248 * the swapper to find the on-disk block of a specific piece of data.
3249 *
3250 * Naturally, this is dangerous if the block concerned is still in the
3251 * journal. If somebody makes a swapfile on an ext4 data-journaling
3252 * filesystem and enables swap, then they may get a nasty shock when the
3253 * data getting swapped to that swapfile suddenly gets overwritten by
3254 * the original zero's written out previously to the journal and
3255 * awaiting writeback in the kernel's buffer cache.
3256 *
3257 * So, if we see any bmap calls here on a modified, data-journaled file,
3258 * take extra steps to flush any blocks which might be in the cache.
3259 */
3260static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3261{
3262 struct inode *inode = mapping->host;
3263 journal_t *journal;
3264 int err;
3265
3266 /*
3267 * We can get here for an inline file via the FIBMAP ioctl
3268 */
3269 if (ext4_has_inline_data(inode))
3270 return 0;
3271
3272 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3273 test_opt(inode->i_sb, DELALLOC)) {
3274 /*
3275 * With delalloc we want to sync the file
3276 * so that we can make sure we allocate
3277 * blocks for file
3278 */
3279 filemap_write_and_wait(mapping);
3280 }
3281
3282 if (EXT4_JOURNAL(inode) &&
3283 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3284 /*
3285 * This is a REALLY heavyweight approach, but the use of
3286 * bmap on dirty files is expected to be extremely rare:
3287 * only if we run lilo or swapon on a freshly made file
3288 * do we expect this to happen.
3289 *
3290 * (bmap requires CAP_SYS_RAWIO so this does not
3291 * represent an unprivileged user DOS attack --- we'd be
3292 * in trouble if mortal users could trigger this path at
3293 * will.)
3294 *
3295 * NB. EXT4_STATE_JDATA is not set on files other than
3296 * regular files. If somebody wants to bmap a directory
3297 * or symlink and gets confused because the buffer
3298 * hasn't yet been flushed to disk, they deserve
3299 * everything they get.
3300 */
3301
3302 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3303 journal = EXT4_JOURNAL(inode);
3304 jbd2_journal_lock_updates(journal);
3305 err = jbd2_journal_flush(journal);
3306 jbd2_journal_unlock_updates(journal);
3307
3308 if (err)
3309 return 0;
3310 }
3311
3312 return generic_block_bmap(mapping, block, ext4_get_block);
3313}
3314
3315static int ext4_readpage(struct file *file, struct page *page)
3316{
3317 int ret = -EAGAIN;
3318 struct inode *inode = page->mapping->host;
3319
3320 trace_ext4_readpage(page);
3321
3322 if (ext4_has_inline_data(inode))
3323 ret = ext4_readpage_inline(inode, page);
3324
3325 if (ret == -EAGAIN)
3326 return ext4_mpage_readpages(page->mapping, NULL, page, 1,
3327 false);
3328
3329 return ret;
3330}
3331
3332static int
3333ext4_readpages(struct file *file, struct address_space *mapping,
3334 struct list_head *pages, unsigned nr_pages)
3335{
3336 struct inode *inode = mapping->host;
3337
3338 /* If the file has inline data, no need to do readpages. */
3339 if (ext4_has_inline_data(inode))
3340 return 0;
3341
3342 return ext4_mpage_readpages(mapping, pages, NULL, nr_pages, true);
3343}
3344
3345static void ext4_invalidatepage(struct page *page, unsigned int offset,
3346 unsigned int length)
3347{
3348 trace_ext4_invalidatepage(page, offset, length);
3349
3350 /* No journalling happens on data buffers when this function is used */
3351 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3352
3353 block_invalidatepage(page, offset, length);
3354}
3355
3356static int __ext4_journalled_invalidatepage(struct page *page,
3357 unsigned int offset,
3358 unsigned int length)
3359{
3360 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3361
3362 trace_ext4_journalled_invalidatepage(page, offset, length);
3363
3364 /*
3365 * If it's a full truncate we just forget about the pending dirtying
3366 */
3367 if (offset == 0 && length == PAGE_SIZE)
3368 ClearPageChecked(page);
3369
3370 return jbd2_journal_invalidatepage(journal, page, offset, length);
3371}
3372
3373/* Wrapper for aops... */
3374static void ext4_journalled_invalidatepage(struct page *page,
3375 unsigned int offset,
3376 unsigned int length)
3377{
3378 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3379}
3380
3381static int ext4_releasepage(struct page *page, gfp_t wait)
3382{
3383 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3384
3385 trace_ext4_releasepage(page);
3386
3387 /* Page has dirty journalled data -> cannot release */
3388 if (PageChecked(page))
3389 return 0;
3390 if (journal)
3391 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3392 else
3393 return try_to_free_buffers(page);
3394}
3395
3396static bool ext4_inode_datasync_dirty(struct inode *inode)
3397{
3398 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3399
3400 if (journal)
3401 return !jbd2_transaction_committed(journal,
3402 EXT4_I(inode)->i_datasync_tid);
3403 /* Any metadata buffers to write? */
3404 if (!list_empty(&inode->i_mapping->private_list))
3405 return true;
3406 return inode->i_state & I_DIRTY_DATASYNC;
3407}
3408
3409static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3410 unsigned flags, struct iomap *iomap)
3411{
3412 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3413 unsigned int blkbits = inode->i_blkbits;
3414 unsigned long first_block, last_block;
3415 struct ext4_map_blocks map;
3416 bool delalloc = false;
3417 int ret;
3418
3419 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3420 return -EINVAL;
3421 first_block = offset >> blkbits;
3422 last_block = min_t(loff_t, (offset + length - 1) >> blkbits,
3423 EXT4_MAX_LOGICAL_BLOCK);
3424
3425 if (flags & IOMAP_REPORT) {
3426 if (ext4_has_inline_data(inode)) {
3427 ret = ext4_inline_data_iomap(inode, iomap);
3428 if (ret != -EAGAIN) {
3429 if (ret == 0 && offset >= iomap->length)
3430 ret = -ENOENT;
3431 return ret;
3432 }
3433 }
3434 } else {
3435 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3436 return -ERANGE;
3437 }
3438
3439 map.m_lblk = first_block;
3440 map.m_len = last_block - first_block + 1;
3441
3442 if (flags & IOMAP_REPORT) {
3443 ret = ext4_map_blocks(NULL, inode, &map, 0);
3444 if (ret < 0)
3445 return ret;
3446
3447 if (ret == 0) {
3448 ext4_lblk_t end = map.m_lblk + map.m_len - 1;
3449 struct extent_status es;
3450
3451 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
3452 map.m_lblk, end, &es);
3453
3454 if (!es.es_len || es.es_lblk > end) {
3455 /* entire range is a hole */
3456 } else if (es.es_lblk > map.m_lblk) {
3457 /* range starts with a hole */
3458 map.m_len = es.es_lblk - map.m_lblk;
3459 } else {
3460 ext4_lblk_t offs = 0;
3461
3462 if (es.es_lblk < map.m_lblk)
3463 offs = map.m_lblk - es.es_lblk;
3464 map.m_lblk = es.es_lblk + offs;
3465 map.m_len = es.es_len - offs;
3466 delalloc = true;
3467 }
3468 }
3469 } else if (flags & IOMAP_WRITE) {
3470 int dio_credits;
3471 handle_t *handle;
3472 int retries = 0;
3473
3474 /* Trim mapping request to maximum we can map at once for DIO */
3475 if (map.m_len > DIO_MAX_BLOCKS)
3476 map.m_len = DIO_MAX_BLOCKS;
3477 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3478retry:
3479 /*
3480 * Either we allocate blocks and then we don't get unwritten
3481 * extent so we have reserved enough credits, or the blocks
3482 * are already allocated and unwritten and in that case
3483 * extent conversion fits in the credits as well.
3484 */
3485 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3486 dio_credits);
3487 if (IS_ERR(handle))
3488 return PTR_ERR(handle);
3489
3490 ret = ext4_map_blocks(handle, inode, &map,
3491 EXT4_GET_BLOCKS_CREATE_ZERO);
3492 if (ret < 0) {
3493 ext4_journal_stop(handle);
3494 if (ret == -ENOSPC &&
3495 ext4_should_retry_alloc(inode->i_sb, &retries))
3496 goto retry;
3497 return ret;
3498 }
3499
3500 /*
3501 * If we added blocks beyond i_size, we need to make sure they
3502 * will get truncated if we crash before updating i_size in
3503 * ext4_iomap_end(). For faults we don't need to do that (and
3504 * even cannot because for orphan list operations inode_lock is
3505 * required) - if we happen to instantiate block beyond i_size,
3506 * it is because we race with truncate which has already added
3507 * the inode to the orphan list.
3508 */
3509 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3510 (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3511 int err;
3512
3513 err = ext4_orphan_add(handle, inode);
3514 if (err < 0) {
3515 ext4_journal_stop(handle);
3516 return err;
3517 }
3518 }
3519 ext4_journal_stop(handle);
3520 } else {
3521 ret = ext4_map_blocks(NULL, inode, &map, 0);
3522 if (ret < 0)
3523 return ret;
3524 }
3525
3526 iomap->flags = 0;
3527 if (ext4_inode_datasync_dirty(inode))
3528 iomap->flags |= IOMAP_F_DIRTY;
3529 iomap->bdev = inode->i_sb->s_bdev;
3530 iomap->dax_dev = sbi->s_daxdev;
3531 iomap->offset = (u64)first_block << blkbits;
3532 iomap->length = (u64)map.m_len << blkbits;
3533
3534 if (ret == 0) {
3535 iomap->type = delalloc ? IOMAP_DELALLOC : IOMAP_HOLE;
3536 iomap->addr = IOMAP_NULL_ADDR;
3537 } else {
3538 if (map.m_flags & EXT4_MAP_MAPPED) {
3539 iomap->type = IOMAP_MAPPED;
3540 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3541 iomap->type = IOMAP_UNWRITTEN;
3542 } else {
3543 WARN_ON_ONCE(1);
3544 return -EIO;
3545 }
3546 iomap->addr = (u64)map.m_pblk << blkbits;
3547 }
3548
3549 if (map.m_flags & EXT4_MAP_NEW)
3550 iomap->flags |= IOMAP_F_NEW;
3551
3552 return 0;
3553}
3554
3555static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3556 ssize_t written, unsigned flags, struct iomap *iomap)
3557{
3558 int ret = 0;
3559 handle_t *handle;
3560 int blkbits = inode->i_blkbits;
3561 bool truncate = false;
3562
3563 if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3564 return 0;
3565
3566 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3567 if (IS_ERR(handle)) {
3568 ret = PTR_ERR(handle);
3569 goto orphan_del;
3570 }
3571 if (ext4_update_inode_size(inode, offset + written))
3572 ext4_mark_inode_dirty(handle, inode);
3573 /*
3574 * We may need to truncate allocated but not written blocks beyond EOF.
3575 */
3576 if (iomap->offset + iomap->length >
3577 ALIGN(inode->i_size, 1 << blkbits)) {
3578 ext4_lblk_t written_blk, end_blk;
3579
3580 written_blk = (offset + written) >> blkbits;
3581 end_blk = (offset + length) >> blkbits;
3582 if (written_blk < end_blk && ext4_can_truncate(inode))
3583 truncate = true;
3584 }
3585 /*
3586 * Remove inode from orphan list if we were extending a inode and
3587 * everything went fine.
3588 */
3589 if (!truncate && inode->i_nlink &&
3590 !list_empty(&EXT4_I(inode)->i_orphan))
3591 ext4_orphan_del(handle, inode);
3592 ext4_journal_stop(handle);
3593 if (truncate) {
3594 ext4_truncate_failed_write(inode);
3595orphan_del:
3596 /*
3597 * If truncate failed early the inode might still be on the
3598 * orphan list; we need to make sure the inode is removed from
3599 * the orphan list in that case.
3600 */
3601 if (inode->i_nlink)
3602 ext4_orphan_del(NULL, inode);
3603 }
3604 return ret;
3605}
3606
3607const struct iomap_ops ext4_iomap_ops = {
3608 .iomap_begin = ext4_iomap_begin,
3609 .iomap_end = ext4_iomap_end,
3610};
3611
3612static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3613 ssize_t size, void *private)
3614{
3615 ext4_io_end_t *io_end = private;
3616
3617 /* if not async direct IO just return */
3618 if (!io_end)
3619 return 0;
3620
3621 ext_debug("ext4_end_io_dio(): io_end 0x%p "
3622 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3623 io_end, io_end->inode->i_ino, iocb, offset, size);
3624
3625 /*
3626 * Error during AIO DIO. We cannot convert unwritten extents as the
3627 * data was not written. Just clear the unwritten flag and drop io_end.
3628 */
3629 if (size <= 0) {
3630 ext4_clear_io_unwritten_flag(io_end);
3631 size = 0;
3632 }
3633 io_end->offset = offset;
3634 io_end->size = size;
3635 ext4_put_io_end(io_end);
3636
3637 return 0;
3638}
3639
3640/*
3641 * Handling of direct IO writes.
3642 *
3643 * For ext4 extent files, ext4 will do direct-io write even to holes,
3644 * preallocated extents, and those write extend the file, no need to
3645 * fall back to buffered IO.
3646 *
3647 * For holes, we fallocate those blocks, mark them as unwritten
3648 * If those blocks were preallocated, we mark sure they are split, but
3649 * still keep the range to write as unwritten.
3650 *
3651 * The unwritten extents will be converted to written when DIO is completed.
3652 * For async direct IO, since the IO may still pending when return, we
3653 * set up an end_io call back function, which will do the conversion
3654 * when async direct IO completed.
3655 *
3656 * If the O_DIRECT write will extend the file then add this inode to the
3657 * orphan list. So recovery will truncate it back to the original size
3658 * if the machine crashes during the write.
3659 *
3660 */
3661static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3662{
3663 struct file *file = iocb->ki_filp;
3664 struct inode *inode = file->f_mapping->host;
3665 struct ext4_inode_info *ei = EXT4_I(inode);
3666 ssize_t ret;
3667 loff_t offset = iocb->ki_pos;
3668 size_t count = iov_iter_count(iter);
3669 int overwrite = 0;
3670 get_block_t *get_block_func = NULL;
3671 int dio_flags = 0;
3672 loff_t final_size = offset + count;
3673 int orphan = 0;
3674 handle_t *handle;
3675
3676 if (final_size > inode->i_size || final_size > ei->i_disksize) {
3677 /* Credits for sb + inode write */
3678 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3679 if (IS_ERR(handle)) {
3680 ret = PTR_ERR(handle);
3681 goto out;
3682 }
3683 ret = ext4_orphan_add(handle, inode);
3684 if (ret) {
3685 ext4_journal_stop(handle);
3686 goto out;
3687 }
3688 orphan = 1;
3689 ext4_update_i_disksize(inode, inode->i_size);
3690 ext4_journal_stop(handle);
3691 }
3692
3693 BUG_ON(iocb->private == NULL);
3694
3695 /*
3696 * Make all waiters for direct IO properly wait also for extent
3697 * conversion. This also disallows race between truncate() and
3698 * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3699 */
3700 inode_dio_begin(inode);
3701
3702 /* If we do a overwrite dio, i_mutex locking can be released */
3703 overwrite = *((int *)iocb->private);
3704
3705 if (overwrite)
3706 inode_unlock(inode);
3707
3708 /*
3709 * For extent mapped files we could direct write to holes and fallocate.
3710 *
3711 * Allocated blocks to fill the hole are marked as unwritten to prevent
3712 * parallel buffered read to expose the stale data before DIO complete
3713 * the data IO.
3714 *
3715 * As to previously fallocated extents, ext4 get_block will just simply
3716 * mark the buffer mapped but still keep the extents unwritten.
3717 *
3718 * For non AIO case, we will convert those unwritten extents to written
3719 * after return back from blockdev_direct_IO. That way we save us from
3720 * allocating io_end structure and also the overhead of offloading
3721 * the extent convertion to a workqueue.
3722 *
3723 * For async DIO, the conversion needs to be deferred when the
3724 * IO is completed. The ext4 end_io callback function will be
3725 * called to take care of the conversion work. Here for async
3726 * case, we allocate an io_end structure to hook to the iocb.
3727 */
3728 iocb->private = NULL;
3729 if (overwrite)
3730 get_block_func = ext4_dio_get_block_overwrite;
3731 else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
3732 round_down(offset, i_blocksize(inode)) >= inode->i_size) {
3733 get_block_func = ext4_dio_get_block;
3734 dio_flags = DIO_LOCKING | DIO_SKIP_HOLES;
3735 } else if (is_sync_kiocb(iocb)) {
3736 get_block_func = ext4_dio_get_block_unwritten_sync;
3737 dio_flags = DIO_LOCKING;
3738 } else {
3739 get_block_func = ext4_dio_get_block_unwritten_async;
3740 dio_flags = DIO_LOCKING;
3741 }
3742 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
3743 get_block_func, ext4_end_io_dio, NULL,
3744 dio_flags);
3745
3746 if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3747 EXT4_STATE_DIO_UNWRITTEN)) {
3748 int err;
3749 /*
3750 * for non AIO case, since the IO is already
3751 * completed, we could do the conversion right here
3752 */
3753 err = ext4_convert_unwritten_extents(NULL, inode,
3754 offset, ret);
3755 if (err < 0)
3756 ret = err;
3757 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3758 }
3759
3760 inode_dio_end(inode);
3761 /* take i_mutex locking again if we do a ovewrite dio */
3762 if (overwrite)
3763 inode_lock(inode);
3764
3765 if (ret < 0 && final_size > inode->i_size)
3766 ext4_truncate_failed_write(inode);
3767
3768 /* Handle extending of i_size after direct IO write */
3769 if (orphan) {
3770 int err;
3771
3772 /* Credits for sb + inode write */
3773 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3774 if (IS_ERR(handle)) {
3775 /*
3776 * We wrote the data but cannot extend
3777 * i_size. Bail out. In async io case, we do
3778 * not return error here because we have
3779 * already submmitted the corresponding
3780 * bio. Returning error here makes the caller
3781 * think that this IO is done and failed
3782 * resulting in race with bio's completion
3783 * handler.
3784 */
3785 if (!ret)
3786 ret = PTR_ERR(handle);
3787 if (inode->i_nlink)
3788 ext4_orphan_del(NULL, inode);
3789
3790 goto out;
3791 }
3792 if (inode->i_nlink)
3793 ext4_orphan_del(handle, inode);
3794 if (ret > 0) {
3795 loff_t end = offset + ret;
3796 if (end > inode->i_size || end > ei->i_disksize) {
3797 ext4_update_i_disksize(inode, end);
3798 if (end > inode->i_size)
3799 i_size_write(inode, end);
3800 /*
3801 * We're going to return a positive `ret'
3802 * here due to non-zero-length I/O, so there's
3803 * no way of reporting error returns from
3804 * ext4_mark_inode_dirty() to userspace. So
3805 * ignore it.
3806 */
3807 ext4_mark_inode_dirty(handle, inode);
3808 }
3809 }
3810 err = ext4_journal_stop(handle);
3811 if (ret == 0)
3812 ret = err;
3813 }
3814out:
3815 return ret;
3816}
3817
3818static ssize_t ext4_direct_IO_read(struct kiocb *iocb, struct iov_iter *iter)
3819{
3820 struct address_space *mapping = iocb->ki_filp->f_mapping;
3821 struct inode *inode = mapping->host;
3822 size_t count = iov_iter_count(iter);
3823 ssize_t ret;
3824
3825 /*
3826 * Shared inode_lock is enough for us - it protects against concurrent
3827 * writes & truncates and since we take care of writing back page cache,
3828 * we are protected against page writeback as well.
3829 */
3830 inode_lock_shared(inode);
3831 ret = filemap_write_and_wait_range(mapping, iocb->ki_pos,
3832 iocb->ki_pos + count - 1);
3833 if (ret)
3834 goto out_unlock;
3835 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
3836 iter, ext4_dio_get_block, NULL, NULL, 0);
3837out_unlock:
3838 inode_unlock_shared(inode);
3839 return ret;
3840}
3841
3842static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3843{
3844 struct file *file = iocb->ki_filp;
3845 struct inode *inode = file->f_mapping->host;
3846 size_t count = iov_iter_count(iter);
3847 loff_t offset = iocb->ki_pos;
3848 ssize_t ret;
3849
3850#ifdef CONFIG_FS_ENCRYPTION
3851 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
3852 return 0;
3853#endif
3854 if (fsverity_active(inode))
3855 return 0;
3856
3857 /*
3858 * If we are doing data journalling we don't support O_DIRECT
3859 */
3860 if (ext4_should_journal_data(inode))
3861 return 0;
3862
3863 /* Let buffer I/O handle the inline data case. */
3864 if (ext4_has_inline_data(inode))
3865 return 0;
3866
3867 trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3868 if (iov_iter_rw(iter) == READ)
3869 ret = ext4_direct_IO_read(iocb, iter);
3870 else
3871 ret = ext4_direct_IO_write(iocb, iter);
3872 trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3873 return ret;
3874}
3875
3876/*
3877 * Pages can be marked dirty completely asynchronously from ext4's journalling
3878 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3879 * much here because ->set_page_dirty is called under VFS locks. The page is
3880 * not necessarily locked.
3881 *
3882 * We cannot just dirty the page and leave attached buffers clean, because the
3883 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3884 * or jbddirty because all the journalling code will explode.
3885 *
3886 * So what we do is to mark the page "pending dirty" and next time writepage
3887 * is called, propagate that into the buffers appropriately.
3888 */
3889static int ext4_journalled_set_page_dirty(struct page *page)
3890{
3891 SetPageChecked(page);
3892 return __set_page_dirty_nobuffers(page);
3893}
3894
3895static int ext4_set_page_dirty(struct page *page)
3896{
3897 WARN_ON_ONCE(!PageLocked(page) && !PageDirty(page));
3898 WARN_ON_ONCE(!page_has_buffers(page));
3899 return __set_page_dirty_buffers(page);
3900}
3901
3902static const struct address_space_operations ext4_aops = {
3903 .readpage = ext4_readpage,
3904 .readpages = ext4_readpages,
3905 .writepage = ext4_writepage,
3906 .writepages = ext4_writepages,
3907 .write_begin = ext4_write_begin,
3908 .write_end = ext4_write_end,
3909 .set_page_dirty = ext4_set_page_dirty,
3910 .bmap = ext4_bmap,
3911 .invalidatepage = ext4_invalidatepage,
3912 .releasepage = ext4_releasepage,
3913 .direct_IO = ext4_direct_IO,
3914 .migratepage = buffer_migrate_page,
3915 .is_partially_uptodate = block_is_partially_uptodate,
3916 .error_remove_page = generic_error_remove_page,
3917};
3918
3919static const struct address_space_operations ext4_journalled_aops = {
3920 .readpage = ext4_readpage,
3921 .readpages = ext4_readpages,
3922 .writepage = ext4_writepage,
3923 .writepages = ext4_writepages,
3924 .write_begin = ext4_write_begin,
3925 .write_end = ext4_journalled_write_end,
3926 .set_page_dirty = ext4_journalled_set_page_dirty,
3927 .bmap = ext4_bmap,
3928 .invalidatepage = ext4_journalled_invalidatepage,
3929 .releasepage = ext4_releasepage,
3930 .direct_IO = ext4_direct_IO,
3931 .is_partially_uptodate = block_is_partially_uptodate,
3932 .error_remove_page = generic_error_remove_page,
3933};
3934
3935static const struct address_space_operations ext4_da_aops = {
3936 .readpage = ext4_readpage,
3937 .readpages = ext4_readpages,
3938 .writepage = ext4_writepage,
3939 .writepages = ext4_writepages,
3940 .write_begin = ext4_da_write_begin,
3941 .write_end = ext4_da_write_end,
3942 .set_page_dirty = ext4_set_page_dirty,
3943 .bmap = ext4_bmap,
3944 .invalidatepage = ext4_invalidatepage,
3945 .releasepage = ext4_releasepage,
3946 .direct_IO = ext4_direct_IO,
3947 .migratepage = buffer_migrate_page,
3948 .is_partially_uptodate = block_is_partially_uptodate,
3949 .error_remove_page = generic_error_remove_page,
3950};
3951
3952static const struct address_space_operations ext4_dax_aops = {
3953 .writepages = ext4_dax_writepages,
3954 .direct_IO = noop_direct_IO,
3955 .set_page_dirty = noop_set_page_dirty,
3956 .bmap = ext4_bmap,
3957 .invalidatepage = noop_invalidatepage,
3958};
3959
3960void ext4_set_aops(struct inode *inode)
3961{
3962 switch (ext4_inode_journal_mode(inode)) {
3963 case EXT4_INODE_ORDERED_DATA_MODE:
3964 case EXT4_INODE_WRITEBACK_DATA_MODE:
3965 break;
3966 case EXT4_INODE_JOURNAL_DATA_MODE:
3967 inode->i_mapping->a_ops = &ext4_journalled_aops;
3968 return;
3969 default:
3970 BUG();
3971 }
3972 if (IS_DAX(inode))
3973 inode->i_mapping->a_ops = &ext4_dax_aops;
3974 else if (test_opt(inode->i_sb, DELALLOC))
3975 inode->i_mapping->a_ops = &ext4_da_aops;
3976 else
3977 inode->i_mapping->a_ops = &ext4_aops;
3978}
3979
3980static int __ext4_block_zero_page_range(handle_t *handle,
3981 struct address_space *mapping, loff_t from, loff_t length)
3982{
3983 ext4_fsblk_t index = from >> PAGE_SHIFT;
3984 unsigned offset = from & (PAGE_SIZE-1);
3985 unsigned blocksize, pos;
3986 ext4_lblk_t iblock;
3987 struct inode *inode = mapping->host;
3988 struct buffer_head *bh;
3989 struct page *page;
3990 int err = 0;
3991
3992 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3993 mapping_gfp_constraint(mapping, ~__GFP_FS));
3994 if (!page)
3995 return -ENOMEM;
3996
3997 blocksize = inode->i_sb->s_blocksize;
3998
3999 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
4000
4001 if (!page_has_buffers(page))
4002 create_empty_buffers(page, blocksize, 0);
4003
4004 /* Find the buffer that contains "offset" */
4005 bh = page_buffers(page);
4006 pos = blocksize;
4007 while (offset >= pos) {
4008 bh = bh->b_this_page;
4009 iblock++;
4010 pos += blocksize;
4011 }
4012 if (buffer_freed(bh)) {
4013 BUFFER_TRACE(bh, "freed: skip");
4014 goto unlock;
4015 }
4016 if (!buffer_mapped(bh)) {
4017 BUFFER_TRACE(bh, "unmapped");
4018 ext4_get_block(inode, iblock, bh, 0);
4019 /* unmapped? It's a hole - nothing to do */
4020 if (!buffer_mapped(bh)) {
4021 BUFFER_TRACE(bh, "still unmapped");
4022 goto unlock;
4023 }
4024 }
4025
4026 /* Ok, it's mapped. Make sure it's up-to-date */
4027 if (PageUptodate(page))
4028 set_buffer_uptodate(bh);
4029
4030 if (!buffer_uptodate(bh)) {
4031 err = -EIO;
4032 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
4033 wait_on_buffer(bh);
4034 /* Uhhuh. Read error. Complain and punt. */
4035 if (!buffer_uptodate(bh))
4036 goto unlock;
4037 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode)) {
4038 /* We expect the key to be set. */
4039 BUG_ON(!fscrypt_has_encryption_key(inode));
4040 WARN_ON_ONCE(fscrypt_decrypt_pagecache_blocks(
4041 page, blocksize, bh_offset(bh)));
4042 }
4043 }
4044 if (ext4_should_journal_data(inode)) {
4045 BUFFER_TRACE(bh, "get write access");
4046 err = ext4_journal_get_write_access(handle, bh);
4047 if (err)
4048 goto unlock;
4049 }
4050 zero_user(page, offset, length);
4051 BUFFER_TRACE(bh, "zeroed end of block");
4052
4053 if (ext4_should_journal_data(inode)) {
4054 err = ext4_handle_dirty_metadata(handle, inode, bh);
4055 } else {
4056 err = 0;
4057 mark_buffer_dirty(bh);
4058 if (ext4_should_order_data(inode))
4059 err = ext4_jbd2_inode_add_write(handle, inode, from,
4060 length);
4061 }
4062
4063unlock:
4064 unlock_page(page);
4065 put_page(page);
4066 return err;
4067}
4068
4069/*
4070 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
4071 * starting from file offset 'from'. The range to be zero'd must
4072 * be contained with in one block. If the specified range exceeds
4073 * the end of the block it will be shortened to end of the block
4074 * that cooresponds to 'from'
4075 */
4076static int ext4_block_zero_page_range(handle_t *handle,
4077 struct address_space *mapping, loff_t from, loff_t length)
4078{
4079 struct inode *inode = mapping->host;
4080 unsigned offset = from & (PAGE_SIZE-1);
4081 unsigned blocksize = inode->i_sb->s_blocksize;
4082 unsigned max = blocksize - (offset & (blocksize - 1));
4083
4084 /*
4085 * correct length if it does not fall between
4086 * 'from' and the end of the block
4087 */
4088 if (length > max || length < 0)
4089 length = max;
4090
4091 if (IS_DAX(inode)) {
4092 return iomap_zero_range(inode, from, length, NULL,
4093 &ext4_iomap_ops);
4094 }
4095 return __ext4_block_zero_page_range(handle, mapping, from, length);
4096}
4097
4098/*
4099 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4100 * up to the end of the block which corresponds to `from'.
4101 * This required during truncate. We need to physically zero the tail end
4102 * of that block so it doesn't yield old data if the file is later grown.
4103 */
4104static int ext4_block_truncate_page(handle_t *handle,
4105 struct address_space *mapping, loff_t from)
4106{
4107 unsigned offset = from & (PAGE_SIZE-1);
4108 unsigned length;
4109 unsigned blocksize;
4110 struct inode *inode = mapping->host;
4111
4112 /* If we are processing an encrypted inode during orphan list handling */
4113 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
4114 return 0;
4115
4116 blocksize = inode->i_sb->s_blocksize;
4117 length = blocksize - (offset & (blocksize - 1));
4118
4119 return ext4_block_zero_page_range(handle, mapping, from, length);
4120}
4121
4122int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
4123 loff_t lstart, loff_t length)
4124{
4125 struct super_block *sb = inode->i_sb;
4126 struct address_space *mapping = inode->i_mapping;
4127 unsigned partial_start, partial_end;
4128 ext4_fsblk_t start, end;
4129 loff_t byte_end = (lstart + length - 1);
4130 int err = 0;
4131
4132 partial_start = lstart & (sb->s_blocksize - 1);
4133 partial_end = byte_end & (sb->s_blocksize - 1);
4134
4135 start = lstart >> sb->s_blocksize_bits;
4136 end = byte_end >> sb->s_blocksize_bits;
4137
4138 /* Handle partial zero within the single block */
4139 if (start == end &&
4140 (partial_start || (partial_end != sb->s_blocksize - 1))) {
4141 err = ext4_block_zero_page_range(handle, mapping,
4142 lstart, length);
4143 return err;
4144 }
4145 /* Handle partial zero out on the start of the range */
4146 if (partial_start) {
4147 err = ext4_block_zero_page_range(handle, mapping,
4148 lstart, sb->s_blocksize);
4149 if (err)
4150 return err;
4151 }
4152 /* Handle partial zero out on the end of the range */
4153 if (partial_end != sb->s_blocksize - 1)
4154 err = ext4_block_zero_page_range(handle, mapping,
4155 byte_end - partial_end,
4156 partial_end + 1);
4157 return err;
4158}
4159
4160int ext4_can_truncate(struct inode *inode)
4161{
4162 if (S_ISREG(inode->i_mode))
4163 return 1;
4164 if (S_ISDIR(inode->i_mode))
4165 return 1;
4166 if (S_ISLNK(inode->i_mode))
4167 return !ext4_inode_is_fast_symlink(inode);
4168 return 0;
4169}
4170
4171/*
4172 * We have to make sure i_disksize gets properly updated before we truncate
4173 * page cache due to hole punching or zero range. Otherwise i_disksize update
4174 * can get lost as it may have been postponed to submission of writeback but
4175 * that will never happen after we truncate page cache.
4176 */
4177int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
4178 loff_t len)
4179{
4180 handle_t *handle;
4181 loff_t size = i_size_read(inode);
4182
4183 WARN_ON(!inode_is_locked(inode));
4184 if (offset > size || offset + len < size)
4185 return 0;
4186
4187 if (EXT4_I(inode)->i_disksize >= size)
4188 return 0;
4189
4190 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
4191 if (IS_ERR(handle))
4192 return PTR_ERR(handle);
4193 ext4_update_i_disksize(inode, size);
4194 ext4_mark_inode_dirty(handle, inode);
4195 ext4_journal_stop(handle);
4196
4197 return 0;
4198}
4199
4200static void ext4_wait_dax_page(struct ext4_inode_info *ei)
4201{
4202 up_write(&ei->i_mmap_sem);
4203 schedule();
4204 down_write(&ei->i_mmap_sem);
4205}
4206
4207int ext4_break_layouts(struct inode *inode)
4208{
4209 struct ext4_inode_info *ei = EXT4_I(inode);
4210 struct page *page;
4211 int error;
4212
4213 if (WARN_ON_ONCE(!rwsem_is_locked(&ei->i_mmap_sem)))
4214 return -EINVAL;
4215
4216 do {
4217 page = dax_layout_busy_page(inode->i_mapping);
4218 if (!page)
4219 return 0;
4220
4221 error = ___wait_var_event(&page->_refcount,
4222 atomic_read(&page->_refcount) == 1,
4223 TASK_INTERRUPTIBLE, 0, 0,
4224 ext4_wait_dax_page(ei));
4225 } while (error == 0);
4226
4227 return error;
4228}
4229
4230/*
4231 * ext4_punch_hole: punches a hole in a file by releasing the blocks
4232 * associated with the given offset and length
4233 *
4234 * @inode: File inode
4235 * @offset: The offset where the hole will begin
4236 * @len: The length of the hole
4237 *
4238 * Returns: 0 on success or negative on failure
4239 */
4240
4241int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
4242{
4243 struct super_block *sb = inode->i_sb;
4244 ext4_lblk_t first_block, stop_block;
4245 struct address_space *mapping = inode->i_mapping;
4246 loff_t first_block_offset, last_block_offset;
4247 handle_t *handle;
4248 unsigned int credits;
4249 int ret = 0;
4250
4251 if (!S_ISREG(inode->i_mode))
4252 return -EOPNOTSUPP;
4253
4254 trace_ext4_punch_hole(inode, offset, length, 0);
4255
4256 ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
4257 if (ext4_has_inline_data(inode)) {
4258 down_write(&EXT4_I(inode)->i_mmap_sem);
4259 ret = ext4_convert_inline_data(inode);
4260 up_write(&EXT4_I(inode)->i_mmap_sem);
4261 if (ret)
4262 return ret;
4263 }
4264
4265 /*
4266 * Write out all dirty pages to avoid race conditions
4267 * Then release them.
4268 */
4269 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4270 ret = filemap_write_and_wait_range(mapping, offset,
4271 offset + length - 1);
4272 if (ret)
4273 return ret;
4274 }
4275
4276 inode_lock(inode);
4277
4278 /* No need to punch hole beyond i_size */
4279 if (offset >= inode->i_size)
4280 goto out_mutex;
4281
4282 /*
4283 * If the hole extends beyond i_size, set the hole
4284 * to end after the page that contains i_size
4285 */
4286 if (offset + length > inode->i_size) {
4287 length = inode->i_size +
4288 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
4289 offset;
4290 }
4291
4292 if (offset & (sb->s_blocksize - 1) ||
4293 (offset + length) & (sb->s_blocksize - 1)) {
4294 /*
4295 * Attach jinode to inode for jbd2 if we do any zeroing of
4296 * partial block
4297 */
4298 ret = ext4_inode_attach_jinode(inode);
4299 if (ret < 0)
4300 goto out_mutex;
4301
4302 }
4303
4304 /* Wait all existing dio workers, newcomers will block on i_mutex */
4305 inode_dio_wait(inode);
4306
4307 /*
4308 * Prevent page faults from reinstantiating pages we have released from
4309 * page cache.
4310 */
4311 down_write(&EXT4_I(inode)->i_mmap_sem);
4312
4313 ret = ext4_break_layouts(inode);
4314 if (ret)
4315 goto out_dio;
4316
4317 first_block_offset = round_up(offset, sb->s_blocksize);
4318 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4319
4320 /* Now release the pages and zero block aligned part of pages*/
4321 if (last_block_offset > first_block_offset) {
4322 ret = ext4_update_disksize_before_punch(inode, offset, length);
4323 if (ret)
4324 goto out_dio;
4325 truncate_pagecache_range(inode, first_block_offset,
4326 last_block_offset);
4327 }
4328
4329 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4330 credits = ext4_writepage_trans_blocks(inode);
4331 else
4332 credits = ext4_blocks_for_truncate(inode);
4333 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4334 if (IS_ERR(handle)) {
4335 ret = PTR_ERR(handle);
4336 ext4_std_error(sb, ret);
4337 goto out_dio;
4338 }
4339
4340 ret = ext4_zero_partial_blocks(handle, inode, offset,
4341 length);
4342 if (ret)
4343 goto out_stop;
4344
4345 first_block = (offset + sb->s_blocksize - 1) >>
4346 EXT4_BLOCK_SIZE_BITS(sb);
4347 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4348
4349 /* If there are blocks to remove, do it */
4350 if (stop_block > first_block) {
4351
4352 down_write(&EXT4_I(inode)->i_data_sem);
4353 ext4_discard_preallocations(inode);
4354
4355 ret = ext4_es_remove_extent(inode, first_block,
4356 stop_block - first_block);
4357 if (ret) {
4358 up_write(&EXT4_I(inode)->i_data_sem);
4359 goto out_stop;
4360 }
4361
4362 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4363 ret = ext4_ext_remove_space(inode, first_block,
4364 stop_block - 1);
4365 else
4366 ret = ext4_ind_remove_space(handle, inode, first_block,
4367 stop_block);
4368
4369 up_write(&EXT4_I(inode)->i_data_sem);
4370 }
4371 if (IS_SYNC(inode))
4372 ext4_handle_sync(handle);
4373
4374 inode->i_mtime = inode->i_ctime = current_time(inode);
4375 ext4_mark_inode_dirty(handle, inode);
4376 if (ret >= 0)
4377 ext4_update_inode_fsync_trans(handle, inode, 1);
4378out_stop:
4379 ext4_journal_stop(handle);
4380out_dio:
4381 up_write(&EXT4_I(inode)->i_mmap_sem);
4382out_mutex:
4383 inode_unlock(inode);
4384 return ret;
4385}
4386
4387int ext4_inode_attach_jinode(struct inode *inode)
4388{
4389 struct ext4_inode_info *ei = EXT4_I(inode);
4390 struct jbd2_inode *jinode;
4391
4392 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4393 return 0;
4394
4395 jinode = jbd2_alloc_inode(GFP_KERNEL);
4396 spin_lock(&inode->i_lock);
4397 if (!ei->jinode) {
4398 if (!jinode) {
4399 spin_unlock(&inode->i_lock);
4400 return -ENOMEM;
4401 }
4402 ei->jinode = jinode;
4403 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4404 jinode = NULL;
4405 }
4406 spin_unlock(&inode->i_lock);
4407 if (unlikely(jinode != NULL))
4408 jbd2_free_inode(jinode);
4409 return 0;
4410}
4411
4412/*
4413 * ext4_truncate()
4414 *
4415 * We block out ext4_get_block() block instantiations across the entire
4416 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4417 * simultaneously on behalf of the same inode.
4418 *
4419 * As we work through the truncate and commit bits of it to the journal there
4420 * is one core, guiding principle: the file's tree must always be consistent on
4421 * disk. We must be able to restart the truncate after a crash.
4422 *
4423 * The file's tree may be transiently inconsistent in memory (although it
4424 * probably isn't), but whenever we close off and commit a journal transaction,
4425 * the contents of (the filesystem + the journal) must be consistent and
4426 * restartable. It's pretty simple, really: bottom up, right to left (although
4427 * left-to-right works OK too).
4428 *
4429 * Note that at recovery time, journal replay occurs *before* the restart of
4430 * truncate against the orphan inode list.
4431 *
4432 * The committed inode has the new, desired i_size (which is the same as
4433 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4434 * that this inode's truncate did not complete and it will again call
4435 * ext4_truncate() to have another go. So there will be instantiated blocks
4436 * to the right of the truncation point in a crashed ext4 filesystem. But
4437 * that's fine - as long as they are linked from the inode, the post-crash
4438 * ext4_truncate() run will find them and release them.
4439 */
4440int ext4_truncate(struct inode *inode)
4441{
4442 struct ext4_inode_info *ei = EXT4_I(inode);
4443 unsigned int credits;
4444 int err = 0;
4445 handle_t *handle;
4446 struct address_space *mapping = inode->i_mapping;
4447
4448 /*
4449 * There is a possibility that we're either freeing the inode
4450 * or it's a completely new inode. In those cases we might not
4451 * have i_mutex locked because it's not necessary.
4452 */
4453 if (!(inode->i_state & (I_NEW|I_FREEING)))
4454 WARN_ON(!inode_is_locked(inode));
4455 trace_ext4_truncate_enter(inode);
4456
4457 if (!ext4_can_truncate(inode))
4458 return 0;
4459
4460 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4461
4462 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4463 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4464
4465 if (ext4_has_inline_data(inode)) {
4466 int has_inline = 1;
4467
4468 err = ext4_inline_data_truncate(inode, &has_inline);
4469 if (err)
4470 return err;
4471 if (has_inline)
4472 return 0;
4473 }
4474
4475 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4476 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4477 if (ext4_inode_attach_jinode(inode) < 0)
4478 return 0;
4479 }
4480
4481 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4482 credits = ext4_writepage_trans_blocks(inode);
4483 else
4484 credits = ext4_blocks_for_truncate(inode);
4485
4486 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4487 if (IS_ERR(handle))
4488 return PTR_ERR(handle);
4489
4490 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4491 ext4_block_truncate_page(handle, mapping, inode->i_size);
4492
4493 /*
4494 * We add the inode to the orphan list, so that if this
4495 * truncate spans multiple transactions, and we crash, we will
4496 * resume the truncate when the filesystem recovers. It also
4497 * marks the inode dirty, to catch the new size.
4498 *
4499 * Implication: the file must always be in a sane, consistent
4500 * truncatable state while each transaction commits.
4501 */
4502 err = ext4_orphan_add(handle, inode);
4503 if (err)
4504 goto out_stop;
4505
4506 down_write(&EXT4_I(inode)->i_data_sem);
4507
4508 ext4_discard_preallocations(inode);
4509
4510 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4511 err = ext4_ext_truncate(handle, inode);
4512 else
4513 ext4_ind_truncate(handle, inode);
4514
4515 up_write(&ei->i_data_sem);
4516 if (err)
4517 goto out_stop;
4518
4519 if (IS_SYNC(inode))
4520 ext4_handle_sync(handle);
4521
4522out_stop:
4523 /*
4524 * If this was a simple ftruncate() and the file will remain alive,
4525 * then we need to clear up the orphan record which we created above.
4526 * However, if this was a real unlink then we were called by
4527 * ext4_evict_inode(), and we allow that function to clean up the
4528 * orphan info for us.
4529 */
4530 if (inode->i_nlink)
4531 ext4_orphan_del(handle, inode);
4532
4533 inode->i_mtime = inode->i_ctime = current_time(inode);
4534 ext4_mark_inode_dirty(handle, inode);
4535 ext4_journal_stop(handle);
4536
4537 trace_ext4_truncate_exit(inode);
4538 return err;
4539}
4540
4541/*
4542 * ext4_get_inode_loc returns with an extra refcount against the inode's
4543 * underlying buffer_head on success. If 'in_mem' is true, we have all
4544 * data in memory that is needed to recreate the on-disk version of this
4545 * inode.
4546 */
4547static int __ext4_get_inode_loc(struct inode *inode,
4548 struct ext4_iloc *iloc, int in_mem)
4549{
4550 struct ext4_group_desc *gdp;
4551 struct buffer_head *bh;
4552 struct super_block *sb = inode->i_sb;
4553 ext4_fsblk_t block;
4554 struct blk_plug plug;
4555 int inodes_per_block, inode_offset;
4556
4557 iloc->bh = NULL;
4558 if (inode->i_ino < EXT4_ROOT_INO ||
4559 inode->i_ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4560 return -EFSCORRUPTED;
4561
4562 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4563 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4564 if (!gdp)
4565 return -EIO;
4566
4567 /*
4568 * Figure out the offset within the block group inode table
4569 */
4570 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4571 inode_offset = ((inode->i_ino - 1) %
4572 EXT4_INODES_PER_GROUP(sb));
4573 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4574 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4575
4576 bh = sb_getblk(sb, block);
4577 if (unlikely(!bh))
4578 return -ENOMEM;
4579 if (!buffer_uptodate(bh)) {
4580 lock_buffer(bh);
4581
4582 /*
4583 * If the buffer has the write error flag, we have failed
4584 * to write out another inode in the same block. In this
4585 * case, we don't have to read the block because we may
4586 * read the old inode data successfully.
4587 */
4588 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4589 set_buffer_uptodate(bh);
4590
4591 if (buffer_uptodate(bh)) {
4592 /* someone brought it uptodate while we waited */
4593 unlock_buffer(bh);
4594 goto has_buffer;
4595 }
4596
4597 /*
4598 * If we have all information of the inode in memory and this
4599 * is the only valid inode in the block, we need not read the
4600 * block.
4601 */
4602 if (in_mem) {
4603 struct buffer_head *bitmap_bh;
4604 int i, start;
4605
4606 start = inode_offset & ~(inodes_per_block - 1);
4607
4608 /* Is the inode bitmap in cache? */
4609 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4610 if (unlikely(!bitmap_bh))
4611 goto make_io;
4612
4613 /*
4614 * If the inode bitmap isn't in cache then the
4615 * optimisation may end up performing two reads instead
4616 * of one, so skip it.
4617 */
4618 if (!buffer_uptodate(bitmap_bh)) {
4619 brelse(bitmap_bh);
4620 goto make_io;
4621 }
4622 for (i = start; i < start + inodes_per_block; i++) {
4623 if (i == inode_offset)
4624 continue;
4625 if (ext4_test_bit(i, bitmap_bh->b_data))
4626 break;
4627 }
4628 brelse(bitmap_bh);
4629 if (i == start + inodes_per_block) {
4630 /* all other inodes are free, so skip I/O */
4631 memset(bh->b_data, 0, bh->b_size);
4632 set_buffer_uptodate(bh);
4633 unlock_buffer(bh);
4634 goto has_buffer;
4635 }
4636 }
4637
4638make_io:
4639 /*
4640 * If we need to do any I/O, try to pre-readahead extra
4641 * blocks from the inode table.
4642 */
4643 blk_start_plug(&plug);
4644 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4645 ext4_fsblk_t b, end, table;
4646 unsigned num;
4647 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4648
4649 table = ext4_inode_table(sb, gdp);
4650 /* s_inode_readahead_blks is always a power of 2 */
4651 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4652 if (table > b)
4653 b = table;
4654 end = b + ra_blks;
4655 num = EXT4_INODES_PER_GROUP(sb);
4656 if (ext4_has_group_desc_csum(sb))
4657 num -= ext4_itable_unused_count(sb, gdp);
4658 table += num / inodes_per_block;
4659 if (end > table)
4660 end = table;
4661 while (b <= end)
4662 sb_breadahead(sb, b++);
4663 }
4664
4665 /*
4666 * There are other valid inodes in the buffer, this inode
4667 * has in-inode xattrs, or we don't have this inode in memory.
4668 * Read the block from disk.
4669 */
4670 trace_ext4_load_inode(inode);
4671 get_bh(bh);
4672 bh->b_end_io = end_buffer_read_sync;
4673 submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
4674 blk_finish_plug(&plug);
4675 wait_on_buffer(bh);
4676 if (!buffer_uptodate(bh)) {
4677 EXT4_ERROR_INODE_BLOCK(inode, block,
4678 "unable to read itable block");
4679 brelse(bh);
4680 return -EIO;
4681 }
4682 }
4683has_buffer:
4684 iloc->bh = bh;
4685 return 0;
4686}
4687
4688int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4689{
4690 /* We have all inode data except xattrs in memory here. */
4691 return __ext4_get_inode_loc(inode, iloc,
4692 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4693}
4694
4695static bool ext4_should_use_dax(struct inode *inode)
4696{
4697 if (!test_opt(inode->i_sb, DAX))
4698 return false;
4699 if (!S_ISREG(inode->i_mode))
4700 return false;
4701 if (ext4_should_journal_data(inode))
4702 return false;
4703 if (ext4_has_inline_data(inode))
4704 return false;
4705 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4706 return false;
4707 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4708 return false;
4709 return true;
4710}
4711
4712void ext4_set_inode_flags(struct inode *inode)
4713{
4714 unsigned int flags = EXT4_I(inode)->i_flags;
4715 unsigned int new_fl = 0;
4716
4717 if (flags & EXT4_SYNC_FL)
4718 new_fl |= S_SYNC;
4719 if (flags & EXT4_APPEND_FL)
4720 new_fl |= S_APPEND;
4721 if (flags & EXT4_IMMUTABLE_FL)
4722 new_fl |= S_IMMUTABLE;
4723 if (flags & EXT4_NOATIME_FL)
4724 new_fl |= S_NOATIME;
4725 if (flags & EXT4_DIRSYNC_FL)
4726 new_fl |= S_DIRSYNC;
4727 if (ext4_should_use_dax(inode))
4728 new_fl |= S_DAX;
4729 if (flags & EXT4_ENCRYPT_FL)
4730 new_fl |= S_ENCRYPTED;
4731 if (flags & EXT4_CASEFOLD_FL)
4732 new_fl |= S_CASEFOLD;
4733 if (flags & EXT4_VERITY_FL)
4734 new_fl |= S_VERITY;
4735 inode_set_flags(inode, new_fl,
4736 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4737 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4738}
4739
4740static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4741 struct ext4_inode_info *ei)
4742{
4743 blkcnt_t i_blocks ;
4744 struct inode *inode = &(ei->vfs_inode);
4745 struct super_block *sb = inode->i_sb;
4746
4747 if (ext4_has_feature_huge_file(sb)) {
4748 /* we are using combined 48 bit field */
4749 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4750 le32_to_cpu(raw_inode->i_blocks_lo);
4751 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4752 /* i_blocks represent file system block size */
4753 return i_blocks << (inode->i_blkbits - 9);
4754 } else {
4755 return i_blocks;
4756 }
4757 } else {
4758 return le32_to_cpu(raw_inode->i_blocks_lo);
4759 }
4760}
4761
4762static inline int ext4_iget_extra_inode(struct inode *inode,
4763 struct ext4_inode *raw_inode,
4764 struct ext4_inode_info *ei)
4765{
4766 __le32 *magic = (void *)raw_inode +
4767 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4768
4769 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize + sizeof(__le32) <=
4770 EXT4_INODE_SIZE(inode->i_sb) &&
4771 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4772 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4773 return ext4_find_inline_data_nolock(inode);
4774 } else
4775 EXT4_I(inode)->i_inline_off = 0;
4776 return 0;
4777}
4778
4779int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4780{
4781 if (!ext4_has_feature_project(inode->i_sb))
4782 return -EOPNOTSUPP;
4783 *projid = EXT4_I(inode)->i_projid;
4784 return 0;
4785}
4786
4787/*
4788 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4789 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4790 * set.
4791 */
4792static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4793{
4794 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4795 inode_set_iversion_raw(inode, val);
4796 else
4797 inode_set_iversion_queried(inode, val);
4798}
4799static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4800{
4801 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4802 return inode_peek_iversion_raw(inode);
4803 else
4804 return inode_peek_iversion(inode);
4805}
4806
4807struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4808 ext4_iget_flags flags, const char *function,
4809 unsigned int line)
4810{
4811 struct ext4_iloc iloc;
4812 struct ext4_inode *raw_inode;
4813 struct ext4_inode_info *ei;
4814 struct inode *inode;
4815 journal_t *journal = EXT4_SB(sb)->s_journal;
4816 long ret;
4817 loff_t size;
4818 int block;
4819 uid_t i_uid;
4820 gid_t i_gid;
4821 projid_t i_projid;
4822
4823 if ((!(flags & EXT4_IGET_SPECIAL) &&
4824 (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)) ||
4825 (ino < EXT4_ROOT_INO) ||
4826 (ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))) {
4827 if (flags & EXT4_IGET_HANDLE)
4828 return ERR_PTR(-ESTALE);
4829 __ext4_error(sb, function, line,
4830 "inode #%lu: comm %s: iget: illegal inode #",
4831 ino, current->comm);
4832 return ERR_PTR(-EFSCORRUPTED);
4833 }
4834
4835 inode = iget_locked(sb, ino);
4836 if (!inode)
4837 return ERR_PTR(-ENOMEM);
4838 if (!(inode->i_state & I_NEW))
4839 return inode;
4840
4841 ei = EXT4_I(inode);
4842 iloc.bh = NULL;
4843
4844 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4845 if (ret < 0)
4846 goto bad_inode;
4847 raw_inode = ext4_raw_inode(&iloc);
4848
4849 if ((ino == EXT4_ROOT_INO) && (raw_inode->i_links_count == 0)) {
4850 ext4_error_inode(inode, function, line, 0,
4851 "iget: root inode unallocated");
4852 ret = -EFSCORRUPTED;
4853 goto bad_inode;
4854 }
4855
4856 if ((flags & EXT4_IGET_HANDLE) &&
4857 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4858 ret = -ESTALE;
4859 goto bad_inode;
4860 }
4861
4862 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4863 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4864 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4865 EXT4_INODE_SIZE(inode->i_sb) ||
4866 (ei->i_extra_isize & 3)) {
4867 ext4_error_inode(inode, function, line, 0,
4868 "iget: bad extra_isize %u "
4869 "(inode size %u)",
4870 ei->i_extra_isize,
4871 EXT4_INODE_SIZE(inode->i_sb));
4872 ret = -EFSCORRUPTED;
4873 goto bad_inode;
4874 }
4875 } else
4876 ei->i_extra_isize = 0;
4877
4878 /* Precompute checksum seed for inode metadata */
4879 if (ext4_has_metadata_csum(sb)) {
4880 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4881 __u32 csum;
4882 __le32 inum = cpu_to_le32(inode->i_ino);
4883 __le32 gen = raw_inode->i_generation;
4884 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4885 sizeof(inum));
4886 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4887 sizeof(gen));
4888 }
4889
4890 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4891 ext4_error_inode(inode, function, line, 0,
4892 "iget: checksum invalid");
4893 ret = -EFSBADCRC;
4894 goto bad_inode;
4895 }
4896
4897 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4898 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4899 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4900 if (ext4_has_feature_project(sb) &&
4901 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4902 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4903 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4904 else
4905 i_projid = EXT4_DEF_PROJID;
4906
4907 if (!(test_opt(inode->i_sb, NO_UID32))) {
4908 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4909 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4910 }
4911 i_uid_write(inode, i_uid);
4912 i_gid_write(inode, i_gid);
4913 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4914 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4915
4916 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4917 ei->i_inline_off = 0;
4918 ei->i_dir_start_lookup = 0;
4919 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4920 /* We now have enough fields to check if the inode was active or not.
4921 * This is needed because nfsd might try to access dead inodes
4922 * the test is that same one that e2fsck uses
4923 * NeilBrown 1999oct15
4924 */
4925 if (inode->i_nlink == 0) {
4926 if ((inode->i_mode == 0 ||
4927 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4928 ino != EXT4_BOOT_LOADER_INO) {
4929 /* this inode is deleted */
4930 ret = -ESTALE;
4931 goto bad_inode;
4932 }
4933 /* The only unlinked inodes we let through here have
4934 * valid i_mode and are being read by the orphan
4935 * recovery code: that's fine, we're about to complete
4936 * the process of deleting those.
4937 * OR it is the EXT4_BOOT_LOADER_INO which is
4938 * not initialized on a new filesystem. */
4939 }
4940 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4941 ext4_set_inode_flags(inode);
4942 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4943 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4944 if (ext4_has_feature_64bit(sb))
4945 ei->i_file_acl |=
4946 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4947 inode->i_size = ext4_isize(sb, raw_inode);
4948 if ((size = i_size_read(inode)) < 0) {
4949 ext4_error_inode(inode, function, line, 0,
4950 "iget: bad i_size value: %lld", size);
4951 ret = -EFSCORRUPTED;
4952 goto bad_inode;
4953 }
4954 ei->i_disksize = inode->i_size;
4955#ifdef CONFIG_QUOTA
4956 ei->i_reserved_quota = 0;
4957#endif
4958 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4959 ei->i_block_group = iloc.block_group;
4960 ei->i_last_alloc_group = ~0;
4961 /*
4962 * NOTE! The in-memory inode i_data array is in little-endian order
4963 * even on big-endian machines: we do NOT byteswap the block numbers!
4964 */
4965 for (block = 0; block < EXT4_N_BLOCKS; block++)
4966 ei->i_data[block] = raw_inode->i_block[block];
4967 INIT_LIST_HEAD(&ei->i_orphan);
4968
4969 /*
4970 * Set transaction id's of transactions that have to be committed
4971 * to finish f[data]sync. We set them to currently running transaction
4972 * as we cannot be sure that the inode or some of its metadata isn't
4973 * part of the transaction - the inode could have been reclaimed and
4974 * now it is reread from disk.
4975 */
4976 if (journal) {
4977 transaction_t *transaction;
4978 tid_t tid;
4979
4980 read_lock(&journal->j_state_lock);
4981 if (journal->j_running_transaction)
4982 transaction = journal->j_running_transaction;
4983 else
4984 transaction = journal->j_committing_transaction;
4985 if (transaction)
4986 tid = transaction->t_tid;
4987 else
4988 tid = journal->j_commit_sequence;
4989 read_unlock(&journal->j_state_lock);
4990 ei->i_sync_tid = tid;
4991 ei->i_datasync_tid = tid;
4992 }
4993
4994 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4995 if (ei->i_extra_isize == 0) {
4996 /* The extra space is currently unused. Use it. */
4997 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4998 ei->i_extra_isize = sizeof(struct ext4_inode) -
4999 EXT4_GOOD_OLD_INODE_SIZE;
5000 } else {
5001 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
5002 if (ret)
5003 goto bad_inode;
5004 }
5005 }
5006
5007 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5008 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5009 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5010 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5011
5012 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5013 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
5014
5015 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5016 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5017 ivers |=
5018 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5019 }
5020 ext4_inode_set_iversion_queried(inode, ivers);
5021 }
5022
5023 ret = 0;
5024 if (ei->i_file_acl &&
5025 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5026 ext4_error_inode(inode, function, line, 0,
5027 "iget: bad extended attribute block %llu",
5028 ei->i_file_acl);
5029 ret = -EFSCORRUPTED;
5030 goto bad_inode;
5031 } else if (!ext4_has_inline_data(inode)) {
5032 /* validate the block references in the inode */
5033 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5034 (S_ISLNK(inode->i_mode) &&
5035 !ext4_inode_is_fast_symlink(inode))) {
5036 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5037 ret = ext4_ext_check_inode(inode);
5038 else
5039 ret = ext4_ind_check_inode(inode);
5040 }
5041 }
5042 if (ret)
5043 goto bad_inode;
5044
5045 if (S_ISREG(inode->i_mode)) {
5046 inode->i_op = &ext4_file_inode_operations;
5047 inode->i_fop = &ext4_file_operations;
5048 ext4_set_aops(inode);
5049 } else if (S_ISDIR(inode->i_mode)) {
5050 inode->i_op = &ext4_dir_inode_operations;
5051 inode->i_fop = &ext4_dir_operations;
5052 } else if (S_ISLNK(inode->i_mode)) {
5053 /* VFS does not allow setting these so must be corruption */
5054 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
5055 ext4_error_inode(inode, function, line, 0,
5056 "iget: immutable or append flags "
5057 "not allowed on symlinks");
5058 ret = -EFSCORRUPTED;
5059 goto bad_inode;
5060 }
5061 if (IS_ENCRYPTED(inode)) {
5062 inode->i_op = &ext4_encrypted_symlink_inode_operations;
5063 ext4_set_aops(inode);
5064 } else if (ext4_inode_is_fast_symlink(inode)) {
5065 inode->i_link = (char *)ei->i_data;
5066 inode->i_op = &ext4_fast_symlink_inode_operations;
5067 nd_terminate_link(ei->i_data, inode->i_size,
5068 sizeof(ei->i_data) - 1);
5069 } else {
5070 inode->i_op = &ext4_symlink_inode_operations;
5071 ext4_set_aops(inode);
5072 }
5073 inode_nohighmem(inode);
5074 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5075 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5076 inode->i_op = &ext4_special_inode_operations;
5077 if (raw_inode->i_block[0])
5078 init_special_inode(inode, inode->i_mode,
5079 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5080 else
5081 init_special_inode(inode, inode->i_mode,
5082 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5083 } else if (ino == EXT4_BOOT_LOADER_INO) {
5084 make_bad_inode(inode);
5085 } else {
5086 ret = -EFSCORRUPTED;
5087 ext4_error_inode(inode, function, line, 0,
5088 "iget: bogus i_mode (%o)", inode->i_mode);
5089 goto bad_inode;
5090 }
5091 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb))
5092 ext4_error_inode(inode, function, line, 0,
5093 "casefold flag without casefold feature");
5094 brelse(iloc.bh);
5095
5096 unlock_new_inode(inode);
5097 return inode;
5098
5099bad_inode:
5100 brelse(iloc.bh);
5101 iget_failed(inode);
5102 return ERR_PTR(ret);
5103}
5104
5105static int ext4_inode_blocks_set(handle_t *handle,
5106 struct ext4_inode *raw_inode,
5107 struct ext4_inode_info *ei)
5108{
5109 struct inode *inode = &(ei->vfs_inode);
5110 u64 i_blocks = inode->i_blocks;
5111 struct super_block *sb = inode->i_sb;
5112
5113 if (i_blocks <= ~0U) {
5114 /*
5115 * i_blocks can be represented in a 32 bit variable
5116 * as multiple of 512 bytes
5117 */
5118 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5119 raw_inode->i_blocks_high = 0;
5120 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5121 return 0;
5122 }
5123 if (!ext4_has_feature_huge_file(sb))
5124 return -EFBIG;
5125
5126 if (i_blocks <= 0xffffffffffffULL) {
5127 /*
5128 * i_blocks can be represented in a 48 bit variable
5129 * as multiple of 512 bytes
5130 */
5131 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5132 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5133 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5134 } else {
5135 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5136 /* i_block is stored in file system block size */
5137 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5138 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5139 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5140 }
5141 return 0;
5142}
5143
5144struct other_inode {
5145 unsigned long orig_ino;
5146 struct ext4_inode *raw_inode;
5147};
5148
5149static int other_inode_match(struct inode * inode, unsigned long ino,
5150 void *data)
5151{
5152 struct other_inode *oi = (struct other_inode *) data;
5153
5154 if ((inode->i_ino != ino) ||
5155 (inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5156 I_DIRTY_INODE)) ||
5157 ((inode->i_state & I_DIRTY_TIME) == 0))
5158 return 0;
5159 spin_lock(&inode->i_lock);
5160 if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5161 I_DIRTY_INODE)) == 0) &&
5162 (inode->i_state & I_DIRTY_TIME)) {
5163 struct ext4_inode_info *ei = EXT4_I(inode);
5164
5165 inode->i_state &= ~(I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED);
5166 spin_unlock(&inode->i_lock);
5167
5168 spin_lock(&ei->i_raw_lock);
5169 EXT4_INODE_SET_XTIME(i_ctime, inode, oi->raw_inode);
5170 EXT4_INODE_SET_XTIME(i_mtime, inode, oi->raw_inode);
5171 EXT4_INODE_SET_XTIME(i_atime, inode, oi->raw_inode);
5172 ext4_inode_csum_set(inode, oi->raw_inode, ei);
5173 spin_unlock(&ei->i_raw_lock);
5174 trace_ext4_other_inode_update_time(inode, oi->orig_ino);
5175 return -1;
5176 }
5177 spin_unlock(&inode->i_lock);
5178 return -1;
5179}
5180
5181/*
5182 * Opportunistically update the other time fields for other inodes in
5183 * the same inode table block.
5184 */
5185static void ext4_update_other_inodes_time(struct super_block *sb,
5186 unsigned long orig_ino, char *buf)
5187{
5188 struct other_inode oi;
5189 unsigned long ino;
5190 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5191 int inode_size = EXT4_INODE_SIZE(sb);
5192
5193 oi.orig_ino = orig_ino;
5194 /*
5195 * Calculate the first inode in the inode table block. Inode
5196 * numbers are one-based. That is, the first inode in a block
5197 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5198 */
5199 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5200 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5201 if (ino == orig_ino)
5202 continue;
5203 oi.raw_inode = (struct ext4_inode *) buf;
5204 (void) find_inode_nowait(sb, ino, other_inode_match, &oi);
5205 }
5206}
5207
5208/*
5209 * Post the struct inode info into an on-disk inode location in the
5210 * buffer-cache. This gobbles the caller's reference to the
5211 * buffer_head in the inode location struct.
5212 *
5213 * The caller must have write access to iloc->bh.
5214 */
5215static int ext4_do_update_inode(handle_t *handle,
5216 struct inode *inode,
5217 struct ext4_iloc *iloc)
5218{
5219 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5220 struct ext4_inode_info *ei = EXT4_I(inode);
5221 struct buffer_head *bh = iloc->bh;
5222 struct super_block *sb = inode->i_sb;
5223 int err = 0, rc, block;
5224 int need_datasync = 0, set_large_file = 0;
5225 uid_t i_uid;
5226 gid_t i_gid;
5227 projid_t i_projid;
5228
5229 spin_lock(&ei->i_raw_lock);
5230
5231 /* For fields not tracked in the in-memory inode,
5232 * initialise them to zero for new inodes. */
5233 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5234 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5235
5236 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5237 i_uid = i_uid_read(inode);
5238 i_gid = i_gid_read(inode);
5239 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
5240 if (!(test_opt(inode->i_sb, NO_UID32))) {
5241 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
5242 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
5243/*
5244 * Fix up interoperability with old kernels. Otherwise, old inodes get
5245 * re-used with the upper 16 bits of the uid/gid intact
5246 */
5247 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
5248 raw_inode->i_uid_high = 0;
5249 raw_inode->i_gid_high = 0;
5250 } else {
5251 raw_inode->i_uid_high =
5252 cpu_to_le16(high_16_bits(i_uid));
5253 raw_inode->i_gid_high =
5254 cpu_to_le16(high_16_bits(i_gid));
5255 }
5256 } else {
5257 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
5258 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
5259 raw_inode->i_uid_high = 0;
5260 raw_inode->i_gid_high = 0;
5261 }
5262 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5263
5264 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5265 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5266 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5267 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5268
5269 err = ext4_inode_blocks_set(handle, raw_inode, ei);
5270 if (err) {
5271 spin_unlock(&ei->i_raw_lock);
5272 goto out_brelse;
5273 }
5274 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5275 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
5276 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
5277 raw_inode->i_file_acl_high =
5278 cpu_to_le16(ei->i_file_acl >> 32);
5279 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5280 if (ei->i_disksize != ext4_isize(inode->i_sb, raw_inode)) {
5281 ext4_isize_set(raw_inode, ei->i_disksize);
5282 need_datasync = 1;
5283 }
5284 if (ei->i_disksize > 0x7fffffffULL) {
5285 if (!ext4_has_feature_large_file(sb) ||
5286 EXT4_SB(sb)->s_es->s_rev_level ==
5287 cpu_to_le32(EXT4_GOOD_OLD_REV))
5288 set_large_file = 1;
5289 }
5290 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5291 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5292 if (old_valid_dev(inode->i_rdev)) {
5293 raw_inode->i_block[0] =
5294 cpu_to_le32(old_encode_dev(inode->i_rdev));
5295 raw_inode->i_block[1] = 0;
5296 } else {
5297 raw_inode->i_block[0] = 0;
5298 raw_inode->i_block[1] =
5299 cpu_to_le32(new_encode_dev(inode->i_rdev));
5300 raw_inode->i_block[2] = 0;
5301 }
5302 } else if (!ext4_has_inline_data(inode)) {
5303 for (block = 0; block < EXT4_N_BLOCKS; block++)
5304 raw_inode->i_block[block] = ei->i_data[block];
5305 }
5306
5307 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5308 u64 ivers = ext4_inode_peek_iversion(inode);
5309
5310 raw_inode->i_disk_version = cpu_to_le32(ivers);
5311 if (ei->i_extra_isize) {
5312 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5313 raw_inode->i_version_hi =
5314 cpu_to_le32(ivers >> 32);
5315 raw_inode->i_extra_isize =
5316 cpu_to_le16(ei->i_extra_isize);
5317 }
5318 }
5319
5320 BUG_ON(!ext4_has_feature_project(inode->i_sb) &&
5321 i_projid != EXT4_DEF_PROJID);
5322
5323 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
5324 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
5325 raw_inode->i_projid = cpu_to_le32(i_projid);
5326
5327 ext4_inode_csum_set(inode, raw_inode, ei);
5328 spin_unlock(&ei->i_raw_lock);
5329 if (inode->i_sb->s_flags & SB_LAZYTIME)
5330 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5331 bh->b_data);
5332
5333 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5334 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5335 if (!err)
5336 err = rc;
5337 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5338 if (set_large_file) {
5339 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5340 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
5341 if (err)
5342 goto out_brelse;
5343 ext4_set_feature_large_file(sb);
5344 ext4_handle_sync(handle);
5345 err = ext4_handle_dirty_super(handle, sb);
5346 }
5347 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5348out_brelse:
5349 brelse(bh);
5350 ext4_std_error(inode->i_sb, err);
5351 return err;
5352}
5353
5354/*
5355 * ext4_write_inode()
5356 *
5357 * We are called from a few places:
5358 *
5359 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5360 * Here, there will be no transaction running. We wait for any running
5361 * transaction to commit.
5362 *
5363 * - Within flush work (sys_sync(), kupdate and such).
5364 * We wait on commit, if told to.
5365 *
5366 * - Within iput_final() -> write_inode_now()
5367 * We wait on commit, if told to.
5368 *
5369 * In all cases it is actually safe for us to return without doing anything,
5370 * because the inode has been copied into a raw inode buffer in
5371 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5372 * writeback.
5373 *
5374 * Note that we are absolutely dependent upon all inode dirtiers doing the
5375 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5376 * which we are interested.
5377 *
5378 * It would be a bug for them to not do this. The code:
5379 *
5380 * mark_inode_dirty(inode)
5381 * stuff();
5382 * inode->i_size = expr;
5383 *
5384 * is in error because write_inode() could occur while `stuff()' is running,
5385 * and the new i_size will be lost. Plus the inode will no longer be on the
5386 * superblock's dirty inode list.
5387 */
5388int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5389{
5390 int err;
5391
5392 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC) ||
5393 sb_rdonly(inode->i_sb))
5394 return 0;
5395
5396 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5397 return -EIO;
5398
5399 if (EXT4_SB(inode->i_sb)->s_journal) {
5400 if (ext4_journal_current_handle()) {
5401 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5402 dump_stack();
5403 return -EIO;
5404 }
5405
5406 /*
5407 * No need to force transaction in WB_SYNC_NONE mode. Also
5408 * ext4_sync_fs() will force the commit after everything is
5409 * written.
5410 */
5411 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5412 return 0;
5413
5414 err = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5415 EXT4_I(inode)->i_sync_tid);
5416 } else {
5417 struct ext4_iloc iloc;
5418
5419 err = __ext4_get_inode_loc(inode, &iloc, 0);
5420 if (err)
5421 return err;
5422 /*
5423 * sync(2) will flush the whole buffer cache. No need to do
5424 * it here separately for each inode.
5425 */
5426 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5427 sync_dirty_buffer(iloc.bh);
5428 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5429 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5430 "IO error syncing inode");
5431 err = -EIO;
5432 }
5433 brelse(iloc.bh);
5434 }
5435 return err;
5436}
5437
5438/*
5439 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
5440 * buffers that are attached to a page stradding i_size and are undergoing
5441 * commit. In that case we have to wait for commit to finish and try again.
5442 */
5443static void ext4_wait_for_tail_page_commit(struct inode *inode)
5444{
5445 struct page *page;
5446 unsigned offset;
5447 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5448 tid_t commit_tid = 0;
5449 int ret;
5450
5451 offset = inode->i_size & (PAGE_SIZE - 1);
5452 /*
5453 * All buffers in the last page remain valid? Then there's nothing to
5454 * do. We do the check mainly to optimize the common PAGE_SIZE ==
5455 * blocksize case
5456 */
5457 if (offset > PAGE_SIZE - i_blocksize(inode))
5458 return;
5459 while (1) {
5460 page = find_lock_page(inode->i_mapping,
5461 inode->i_size >> PAGE_SHIFT);
5462 if (!page)
5463 return;
5464 ret = __ext4_journalled_invalidatepage(page, offset,
5465 PAGE_SIZE - offset);
5466 unlock_page(page);
5467 put_page(page);
5468 if (ret != -EBUSY)
5469 return;
5470 commit_tid = 0;
5471 read_lock(&journal->j_state_lock);
5472 if (journal->j_committing_transaction)
5473 commit_tid = journal->j_committing_transaction->t_tid;
5474 read_unlock(&journal->j_state_lock);
5475 if (commit_tid)
5476 jbd2_log_wait_commit(journal, commit_tid);
5477 }
5478}
5479
5480/*
5481 * ext4_setattr()
5482 *
5483 * Called from notify_change.
5484 *
5485 * We want to trap VFS attempts to truncate the file as soon as
5486 * possible. In particular, we want to make sure that when the VFS
5487 * shrinks i_size, we put the inode on the orphan list and modify
5488 * i_disksize immediately, so that during the subsequent flushing of
5489 * dirty pages and freeing of disk blocks, we can guarantee that any
5490 * commit will leave the blocks being flushed in an unused state on
5491 * disk. (On recovery, the inode will get truncated and the blocks will
5492 * be freed, so we have a strong guarantee that no future commit will
5493 * leave these blocks visible to the user.)
5494 *
5495 * Another thing we have to assure is that if we are in ordered mode
5496 * and inode is still attached to the committing transaction, we must
5497 * we start writeout of all the dirty pages which are being truncated.
5498 * This way we are sure that all the data written in the previous
5499 * transaction are already on disk (truncate waits for pages under
5500 * writeback).
5501 *
5502 * Called with inode->i_mutex down.
5503 */
5504int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5505{
5506 struct inode *inode = d_inode(dentry);
5507 int error, rc = 0;
5508 int orphan = 0;
5509 const unsigned int ia_valid = attr->ia_valid;
5510
5511 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5512 return -EIO;
5513
5514 if (unlikely(IS_IMMUTABLE(inode)))
5515 return -EPERM;
5516
5517 if (unlikely(IS_APPEND(inode) &&
5518 (ia_valid & (ATTR_MODE | ATTR_UID |
5519 ATTR_GID | ATTR_TIMES_SET))))
5520 return -EPERM;
5521
5522 error = setattr_prepare(dentry, attr);
5523 if (error)
5524 return error;
5525
5526 error = fscrypt_prepare_setattr(dentry, attr);
5527 if (error)
5528 return error;
5529
5530 error = fsverity_prepare_setattr(dentry, attr);
5531 if (error)
5532 return error;
5533
5534 if (is_quota_modification(inode, attr)) {
5535 error = dquot_initialize(inode);
5536 if (error)
5537 return error;
5538 }
5539 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
5540 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
5541 handle_t *handle;
5542
5543 /* (user+group)*(old+new) structure, inode write (sb,
5544 * inode block, ? - but truncate inode update has it) */
5545 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5546 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5547 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5548 if (IS_ERR(handle)) {
5549 error = PTR_ERR(handle);
5550 goto err_out;
5551 }
5552
5553 /* dquot_transfer() calls back ext4_get_inode_usage() which
5554 * counts xattr inode references.
5555 */
5556 down_read(&EXT4_I(inode)->xattr_sem);
5557 error = dquot_transfer(inode, attr);
5558 up_read(&EXT4_I(inode)->xattr_sem);
5559
5560 if (error) {
5561 ext4_journal_stop(handle);
5562 return error;
5563 }
5564 /* Update corresponding info in inode so that everything is in
5565 * one transaction */
5566 if (attr->ia_valid & ATTR_UID)
5567 inode->i_uid = attr->ia_uid;
5568 if (attr->ia_valid & ATTR_GID)
5569 inode->i_gid = attr->ia_gid;
5570 error = ext4_mark_inode_dirty(handle, inode);
5571 ext4_journal_stop(handle);
5572 }
5573
5574 if (attr->ia_valid & ATTR_SIZE) {
5575 handle_t *handle;
5576 loff_t oldsize = inode->i_size;
5577 int shrink = (attr->ia_size < inode->i_size);
5578
5579 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5580 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5581
5582 if (attr->ia_size > sbi->s_bitmap_maxbytes)
5583 return -EFBIG;
5584 }
5585 if (!S_ISREG(inode->i_mode))
5586 return -EINVAL;
5587
5588 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
5589 inode_inc_iversion(inode);
5590
5591 if (shrink) {
5592 if (ext4_should_order_data(inode)) {
5593 error = ext4_begin_ordered_truncate(inode,
5594 attr->ia_size);
5595 if (error)
5596 goto err_out;
5597 }
5598 /*
5599 * Blocks are going to be removed from the inode. Wait
5600 * for dio in flight.
5601 */
5602 inode_dio_wait(inode);
5603 }
5604
5605 down_write(&EXT4_I(inode)->i_mmap_sem);
5606
5607 rc = ext4_break_layouts(inode);
5608 if (rc) {
5609 up_write(&EXT4_I(inode)->i_mmap_sem);
5610 return rc;
5611 }
5612
5613 if (attr->ia_size != inode->i_size) {
5614 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5615 if (IS_ERR(handle)) {
5616 error = PTR_ERR(handle);
5617 goto out_mmap_sem;
5618 }
5619 if (ext4_handle_valid(handle) && shrink) {
5620 error = ext4_orphan_add(handle, inode);
5621 orphan = 1;
5622 }
5623 /*
5624 * Update c/mtime on truncate up, ext4_truncate() will
5625 * update c/mtime in shrink case below
5626 */
5627 if (!shrink) {
5628 inode->i_mtime = current_time(inode);
5629 inode->i_ctime = inode->i_mtime;
5630 }
5631 down_write(&EXT4_I(inode)->i_data_sem);
5632 EXT4_I(inode)->i_disksize = attr->ia_size;
5633 rc = ext4_mark_inode_dirty(handle, inode);
5634 if (!error)
5635 error = rc;
5636 /*
5637 * We have to update i_size under i_data_sem together
5638 * with i_disksize to avoid races with writeback code
5639 * running ext4_wb_update_i_disksize().
5640 */
5641 if (!error)
5642 i_size_write(inode, attr->ia_size);
5643 up_write(&EXT4_I(inode)->i_data_sem);
5644 ext4_journal_stop(handle);
5645 if (error)
5646 goto out_mmap_sem;
5647 if (!shrink) {
5648 pagecache_isize_extended(inode, oldsize,
5649 inode->i_size);
5650 } else if (ext4_should_journal_data(inode)) {
5651 ext4_wait_for_tail_page_commit(inode);
5652 }
5653 }
5654
5655 /*
5656 * Truncate pagecache after we've waited for commit
5657 * in data=journal mode to make pages freeable.
5658 */
5659 truncate_pagecache(inode, inode->i_size);
5660 /*
5661 * Call ext4_truncate() even if i_size didn't change to
5662 * truncate possible preallocated blocks.
5663 */
5664 if (attr->ia_size <= oldsize) {
5665 rc = ext4_truncate(inode);
5666 if (rc)
5667 error = rc;
5668 }
5669out_mmap_sem:
5670 up_write(&EXT4_I(inode)->i_mmap_sem);
5671 }
5672
5673 if (!error) {
5674 setattr_copy(inode, attr);
5675 mark_inode_dirty(inode);
5676 }
5677
5678 /*
5679 * If the call to ext4_truncate failed to get a transaction handle at
5680 * all, we need to clean up the in-core orphan list manually.
5681 */
5682 if (orphan && inode->i_nlink)
5683 ext4_orphan_del(NULL, inode);
5684
5685 if (!error && (ia_valid & ATTR_MODE))
5686 rc = posix_acl_chmod(inode, inode->i_mode);
5687
5688err_out:
5689 ext4_std_error(inode->i_sb, error);
5690 if (!error)
5691 error = rc;
5692 return error;
5693}
5694
5695int ext4_getattr(const struct path *path, struct kstat *stat,
5696 u32 request_mask, unsigned int query_flags)
5697{
5698 struct inode *inode = d_inode(path->dentry);
5699 struct ext4_inode *raw_inode;
5700 struct ext4_inode_info *ei = EXT4_I(inode);
5701 unsigned int flags;
5702
5703 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5704 stat->result_mask |= STATX_BTIME;
5705 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5706 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5707 }
5708
5709 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5710 if (flags & EXT4_APPEND_FL)
5711 stat->attributes |= STATX_ATTR_APPEND;
5712 if (flags & EXT4_COMPR_FL)
5713 stat->attributes |= STATX_ATTR_COMPRESSED;
5714 if (flags & EXT4_ENCRYPT_FL)
5715 stat->attributes |= STATX_ATTR_ENCRYPTED;
5716 if (flags & EXT4_IMMUTABLE_FL)
5717 stat->attributes |= STATX_ATTR_IMMUTABLE;
5718 if (flags & EXT4_NODUMP_FL)
5719 stat->attributes |= STATX_ATTR_NODUMP;
5720
5721 stat->attributes_mask |= (STATX_ATTR_APPEND |
5722 STATX_ATTR_COMPRESSED |
5723 STATX_ATTR_ENCRYPTED |
5724 STATX_ATTR_IMMUTABLE |
5725 STATX_ATTR_NODUMP);
5726
5727 generic_fillattr(inode, stat);
5728 return 0;
5729}
5730
5731int ext4_file_getattr(const struct path *path, struct kstat *stat,
5732 u32 request_mask, unsigned int query_flags)
5733{
5734 struct inode *inode = d_inode(path->dentry);
5735 u64 delalloc_blocks;
5736
5737 ext4_getattr(path, stat, request_mask, query_flags);
5738
5739 /*
5740 * If there is inline data in the inode, the inode will normally not
5741 * have data blocks allocated (it may have an external xattr block).
5742 * Report at least one sector for such files, so tools like tar, rsync,
5743 * others don't incorrectly think the file is completely sparse.
5744 */
5745 if (unlikely(ext4_has_inline_data(inode)))
5746 stat->blocks += (stat->size + 511) >> 9;
5747
5748 /*
5749 * We can't update i_blocks if the block allocation is delayed
5750 * otherwise in the case of system crash before the real block
5751 * allocation is done, we will have i_blocks inconsistent with
5752 * on-disk file blocks.
5753 * We always keep i_blocks updated together with real
5754 * allocation. But to not confuse with user, stat
5755 * will return the blocks that include the delayed allocation
5756 * blocks for this file.
5757 */
5758 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5759 EXT4_I(inode)->i_reserved_data_blocks);
5760 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5761 return 0;
5762}
5763
5764static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5765 int pextents)
5766{
5767 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5768 return ext4_ind_trans_blocks(inode, lblocks);
5769 return ext4_ext_index_trans_blocks(inode, pextents);
5770}
5771
5772/*
5773 * Account for index blocks, block groups bitmaps and block group
5774 * descriptor blocks if modify datablocks and index blocks
5775 * worse case, the indexs blocks spread over different block groups
5776 *
5777 * If datablocks are discontiguous, they are possible to spread over
5778 * different block groups too. If they are contiguous, with flexbg,
5779 * they could still across block group boundary.
5780 *
5781 * Also account for superblock, inode, quota and xattr blocks
5782 */
5783static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5784 int pextents)
5785{
5786 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5787 int gdpblocks;
5788 int idxblocks;
5789 int ret = 0;
5790
5791 /*
5792 * How many index blocks need to touch to map @lblocks logical blocks
5793 * to @pextents physical extents?
5794 */
5795 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5796
5797 ret = idxblocks;
5798
5799 /*
5800 * Now let's see how many group bitmaps and group descriptors need
5801 * to account
5802 */
5803 groups = idxblocks + pextents;
5804 gdpblocks = groups;
5805 if (groups > ngroups)
5806 groups = ngroups;
5807 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5808 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5809
5810 /* bitmaps and block group descriptor blocks */
5811 ret += groups + gdpblocks;
5812
5813 /* Blocks for super block, inode, quota and xattr blocks */
5814 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5815
5816 return ret;
5817}
5818
5819/*
5820 * Calculate the total number of credits to reserve to fit
5821 * the modification of a single pages into a single transaction,
5822 * which may include multiple chunks of block allocations.
5823 *
5824 * This could be called via ext4_write_begin()
5825 *
5826 * We need to consider the worse case, when
5827 * one new block per extent.
5828 */
5829int ext4_writepage_trans_blocks(struct inode *inode)
5830{
5831 int bpp = ext4_journal_blocks_per_page(inode);
5832 int ret;
5833
5834 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5835
5836 /* Account for data blocks for journalled mode */
5837 if (ext4_should_journal_data(inode))
5838 ret += bpp;
5839 return ret;
5840}
5841
5842/*
5843 * Calculate the journal credits for a chunk of data modification.
5844 *
5845 * This is called from DIO, fallocate or whoever calling
5846 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5847 *
5848 * journal buffers for data blocks are not included here, as DIO
5849 * and fallocate do no need to journal data buffers.
5850 */
5851int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5852{
5853 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5854}
5855
5856/*
5857 * The caller must have previously called ext4_reserve_inode_write().
5858 * Give this, we know that the caller already has write access to iloc->bh.
5859 */
5860int ext4_mark_iloc_dirty(handle_t *handle,
5861 struct inode *inode, struct ext4_iloc *iloc)
5862{
5863 int err = 0;
5864
5865 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
5866 put_bh(iloc->bh);
5867 return -EIO;
5868 }
5869 if (IS_I_VERSION(inode))
5870 inode_inc_iversion(inode);
5871
5872 /* the do_update_inode consumes one bh->b_count */
5873 get_bh(iloc->bh);
5874
5875 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5876 err = ext4_do_update_inode(handle, inode, iloc);
5877 put_bh(iloc->bh);
5878 return err;
5879}
5880
5881/*
5882 * On success, We end up with an outstanding reference count against
5883 * iloc->bh. This _must_ be cleaned up later.
5884 */
5885
5886int
5887ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5888 struct ext4_iloc *iloc)
5889{
5890 int err;
5891
5892 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5893 return -EIO;
5894
5895 err = ext4_get_inode_loc(inode, iloc);
5896 if (!err) {
5897 BUFFER_TRACE(iloc->bh, "get_write_access");
5898 err = ext4_journal_get_write_access(handle, iloc->bh);
5899 if (err) {
5900 brelse(iloc->bh);
5901 iloc->bh = NULL;
5902 }
5903 }
5904 ext4_std_error(inode->i_sb, err);
5905 return err;
5906}
5907
5908static int __ext4_expand_extra_isize(struct inode *inode,
5909 unsigned int new_extra_isize,
5910 struct ext4_iloc *iloc,
5911 handle_t *handle, int *no_expand)
5912{
5913 struct ext4_inode *raw_inode;
5914 struct ext4_xattr_ibody_header *header;
5915 int error;
5916
5917 raw_inode = ext4_raw_inode(iloc);
5918
5919 header = IHDR(inode, raw_inode);
5920
5921 /* No extended attributes present */
5922 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5923 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5924 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5925 EXT4_I(inode)->i_extra_isize, 0,
5926 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5927 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5928 return 0;
5929 }
5930
5931 /* try to expand with EAs present */
5932 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5933 raw_inode, handle);
5934 if (error) {
5935 /*
5936 * Inode size expansion failed; don't try again
5937 */
5938 *no_expand = 1;
5939 }
5940
5941 return error;
5942}
5943
5944/*
5945 * Expand an inode by new_extra_isize bytes.
5946 * Returns 0 on success or negative error number on failure.
5947 */
5948static int ext4_try_to_expand_extra_isize(struct inode *inode,
5949 unsigned int new_extra_isize,
5950 struct ext4_iloc iloc,
5951 handle_t *handle)
5952{
5953 int no_expand;
5954 int error;
5955
5956 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5957 return -EOVERFLOW;
5958
5959 /*
5960 * In nojournal mode, we can immediately attempt to expand
5961 * the inode. When journaled, we first need to obtain extra
5962 * buffer credits since we may write into the EA block
5963 * with this same handle. If journal_extend fails, then it will
5964 * only result in a minor loss of functionality for that inode.
5965 * If this is felt to be critical, then e2fsck should be run to
5966 * force a large enough s_min_extra_isize.
5967 */
5968 if (ext4_handle_valid(handle) &&
5969 jbd2_journal_extend(handle,
5970 EXT4_DATA_TRANS_BLOCKS(inode->i_sb)) != 0)
5971 return -ENOSPC;
5972
5973 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5974 return -EBUSY;
5975
5976 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5977 handle, &no_expand);
5978 ext4_write_unlock_xattr(inode, &no_expand);
5979
5980 return error;
5981}
5982
5983int ext4_expand_extra_isize(struct inode *inode,
5984 unsigned int new_extra_isize,
5985 struct ext4_iloc *iloc)
5986{
5987 handle_t *handle;
5988 int no_expand;
5989 int error, rc;
5990
5991 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5992 brelse(iloc->bh);
5993 return -EOVERFLOW;
5994 }
5995
5996 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5997 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5998 if (IS_ERR(handle)) {
5999 error = PTR_ERR(handle);
6000 brelse(iloc->bh);
6001 return error;
6002 }
6003
6004 ext4_write_lock_xattr(inode, &no_expand);
6005
6006 BUFFER_TRACE(iloc->bh, "get_write_access");
6007 error = ext4_journal_get_write_access(handle, iloc->bh);
6008 if (error) {
6009 brelse(iloc->bh);
6010 goto out_stop;
6011 }
6012
6013 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
6014 handle, &no_expand);
6015
6016 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
6017 if (!error)
6018 error = rc;
6019
6020 ext4_write_unlock_xattr(inode, &no_expand);
6021out_stop:
6022 ext4_journal_stop(handle);
6023 return error;
6024}
6025
6026/*
6027 * What we do here is to mark the in-core inode as clean with respect to inode
6028 * dirtiness (it may still be data-dirty).
6029 * This means that the in-core inode may be reaped by prune_icache
6030 * without having to perform any I/O. This is a very good thing,
6031 * because *any* task may call prune_icache - even ones which
6032 * have a transaction open against a different journal.
6033 *
6034 * Is this cheating? Not really. Sure, we haven't written the
6035 * inode out, but prune_icache isn't a user-visible syncing function.
6036 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
6037 * we start and wait on commits.
6038 */
6039int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
6040{
6041 struct ext4_iloc iloc;
6042 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
6043 int err;
6044
6045 might_sleep();
6046 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
6047 err = ext4_reserve_inode_write(handle, inode, &iloc);
6048 if (err)
6049 return err;
6050
6051 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
6052 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
6053 iloc, handle);
6054
6055 return ext4_mark_iloc_dirty(handle, inode, &iloc);
6056}
6057
6058/*
6059 * ext4_dirty_inode() is called from __mark_inode_dirty()
6060 *
6061 * We're really interested in the case where a file is being extended.
6062 * i_size has been changed by generic_commit_write() and we thus need
6063 * to include the updated inode in the current transaction.
6064 *
6065 * Also, dquot_alloc_block() will always dirty the inode when blocks
6066 * are allocated to the file.
6067 *
6068 * If the inode is marked synchronous, we don't honour that here - doing
6069 * so would cause a commit on atime updates, which we don't bother doing.
6070 * We handle synchronous inodes at the highest possible level.
6071 *
6072 * If only the I_DIRTY_TIME flag is set, we can skip everything. If
6073 * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
6074 * to copy into the on-disk inode structure are the timestamp files.
6075 */
6076void ext4_dirty_inode(struct inode *inode, int flags)
6077{
6078 handle_t *handle;
6079
6080 if (flags == I_DIRTY_TIME)
6081 return;
6082 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
6083 if (IS_ERR(handle))
6084 goto out;
6085
6086 ext4_mark_inode_dirty(handle, inode);
6087
6088 ext4_journal_stop(handle);
6089out:
6090 return;
6091}
6092
6093int ext4_change_inode_journal_flag(struct inode *inode, int val)
6094{
6095 journal_t *journal;
6096 handle_t *handle;
6097 int err;
6098 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
6099
6100 /*
6101 * We have to be very careful here: changing a data block's
6102 * journaling status dynamically is dangerous. If we write a
6103 * data block to the journal, change the status and then delete
6104 * that block, we risk forgetting to revoke the old log record
6105 * from the journal and so a subsequent replay can corrupt data.
6106 * So, first we make sure that the journal is empty and that
6107 * nobody is changing anything.
6108 */
6109
6110 journal = EXT4_JOURNAL(inode);
6111 if (!journal)
6112 return 0;
6113 if (is_journal_aborted(journal))
6114 return -EROFS;
6115
6116 /* Wait for all existing dio workers */
6117 inode_dio_wait(inode);
6118
6119 /*
6120 * Before flushing the journal and switching inode's aops, we have
6121 * to flush all dirty data the inode has. There can be outstanding
6122 * delayed allocations, there can be unwritten extents created by
6123 * fallocate or buffered writes in dioread_nolock mode covered by
6124 * dirty data which can be converted only after flushing the dirty
6125 * data (and journalled aops don't know how to handle these cases).
6126 */
6127 if (val) {
6128 down_write(&EXT4_I(inode)->i_mmap_sem);
6129 err = filemap_write_and_wait(inode->i_mapping);
6130 if (err < 0) {
6131 up_write(&EXT4_I(inode)->i_mmap_sem);
6132 return err;
6133 }
6134 }
6135
6136 percpu_down_write(&sbi->s_journal_flag_rwsem);
6137 jbd2_journal_lock_updates(journal);
6138
6139 /*
6140 * OK, there are no updates running now, and all cached data is
6141 * synced to disk. We are now in a completely consistent state
6142 * which doesn't have anything in the journal, and we know that
6143 * no filesystem updates are running, so it is safe to modify
6144 * the inode's in-core data-journaling state flag now.
6145 */
6146
6147 if (val)
6148 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6149 else {
6150 err = jbd2_journal_flush(journal);
6151 if (err < 0) {
6152 jbd2_journal_unlock_updates(journal);
6153 percpu_up_write(&sbi->s_journal_flag_rwsem);
6154 return err;
6155 }
6156 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6157 }
6158 ext4_set_aops(inode);
6159
6160 jbd2_journal_unlock_updates(journal);
6161 percpu_up_write(&sbi->s_journal_flag_rwsem);
6162
6163 if (val)
6164 up_write(&EXT4_I(inode)->i_mmap_sem);
6165
6166 /* Finally we can mark the inode as dirty. */
6167
6168 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6169 if (IS_ERR(handle))
6170 return PTR_ERR(handle);
6171
6172 err = ext4_mark_inode_dirty(handle, inode);
6173 ext4_handle_sync(handle);
6174 ext4_journal_stop(handle);
6175 ext4_std_error(inode->i_sb, err);
6176
6177 return err;
6178}
6179
6180static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
6181{
6182 return !buffer_mapped(bh);
6183}
6184
6185vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
6186{
6187 struct vm_area_struct *vma = vmf->vma;
6188 struct page *page = vmf->page;
6189 loff_t size;
6190 unsigned long len;
6191 int err;
6192 vm_fault_t ret;
6193 struct file *file = vma->vm_file;
6194 struct inode *inode = file_inode(file);
6195 struct address_space *mapping = inode->i_mapping;
6196 handle_t *handle;
6197 get_block_t *get_block;
6198 int retries = 0;
6199
6200 if (unlikely(IS_IMMUTABLE(inode)))
6201 return VM_FAULT_SIGBUS;
6202
6203 sb_start_pagefault(inode->i_sb);
6204 file_update_time(vma->vm_file);
6205
6206 down_read(&EXT4_I(inode)->i_mmap_sem);
6207
6208 err = ext4_convert_inline_data(inode);
6209 if (err)
6210 goto out_ret;
6211
6212 /* Delalloc case is easy... */
6213 if (test_opt(inode->i_sb, DELALLOC) &&
6214 !ext4_should_journal_data(inode) &&
6215 !ext4_nonda_switch(inode->i_sb)) {
6216 do {
6217 err = block_page_mkwrite(vma, vmf,
6218 ext4_da_get_block_prep);
6219 } while (err == -ENOSPC &&
6220 ext4_should_retry_alloc(inode->i_sb, &retries));
6221 goto out_ret;
6222 }
6223
6224 lock_page(page);
6225 size = i_size_read(inode);
6226 /* Page got truncated from under us? */
6227 if (page->mapping != mapping || page_offset(page) > size) {
6228 unlock_page(page);
6229 ret = VM_FAULT_NOPAGE;
6230 goto out;
6231 }
6232
6233 if (page->index == size >> PAGE_SHIFT)
6234 len = size & ~PAGE_MASK;
6235 else
6236 len = PAGE_SIZE;
6237 /*
6238 * Return if we have all the buffers mapped. This avoids the need to do
6239 * journal_start/journal_stop which can block and take a long time
6240 */
6241 if (page_has_buffers(page)) {
6242 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
6243 0, len, NULL,
6244 ext4_bh_unmapped)) {
6245 /* Wait so that we don't change page under IO */
6246 wait_for_stable_page(page);
6247 ret = VM_FAULT_LOCKED;
6248 goto out;
6249 }
6250 }
6251 unlock_page(page);
6252 /* OK, we need to fill the hole... */
6253 if (ext4_should_dioread_nolock(inode))
6254 get_block = ext4_get_block_unwritten;
6255 else
6256 get_block = ext4_get_block;
6257retry_alloc:
6258 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6259 ext4_writepage_trans_blocks(inode));
6260 if (IS_ERR(handle)) {
6261 ret = VM_FAULT_SIGBUS;
6262 goto out;
6263 }
6264 err = block_page_mkwrite(vma, vmf, get_block);
6265 if (!err && ext4_should_journal_data(inode)) {
6266 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
6267 PAGE_SIZE, NULL, do_journal_get_write_access)) {
6268 unlock_page(page);
6269 ret = VM_FAULT_SIGBUS;
6270 ext4_journal_stop(handle);
6271 goto out;
6272 }
6273 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
6274 }
6275 ext4_journal_stop(handle);
6276 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6277 goto retry_alloc;
6278out_ret:
6279 ret = block_page_mkwrite_return(err);
6280out:
6281 up_read(&EXT4_I(inode)->i_mmap_sem);
6282 sb_end_pagefault(inode->i_sb);
6283 return ret;
6284}
6285
6286vm_fault_t ext4_filemap_fault(struct vm_fault *vmf)
6287{
6288 struct inode *inode = file_inode(vmf->vma->vm_file);
6289 vm_fault_t ret;
6290
6291 down_read(&EXT4_I(inode)->i_mmap_sem);
6292 ret = filemap_fault(vmf);
6293 up_read(&EXT4_I(inode)->i_mmap_sem);
6294
6295 return ret;
6296}