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/mount.h>
24#include <linux/time.h>
25#include <linux/highuid.h>
26#include <linux/pagemap.h>
27#include <linux/dax.h>
28#include <linux/quotaops.h>
29#include <linux/string.h>
30#include <linux/buffer_head.h>
31#include <linux/writeback.h>
32#include <linux/pagevec.h>
33#include <linux/mpage.h>
34#include <linux/namei.h>
35#include <linux/uio.h>
36#include <linux/bio.h>
37#include <linux/workqueue.h>
38#include <linux/kernel.h>
39#include <linux/printk.h>
40#include <linux/slab.h>
41#include <linux/bitops.h>
42#include <linux/iomap.h>
43#include <linux/iversion.h>
44
45#include "ext4_jbd2.h"
46#include "xattr.h"
47#include "acl.h"
48#include "truncate.h"
49
50#include <trace/events/ext4.h>
51
52static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
53 struct ext4_inode_info *ei)
54{
55 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
56 __u32 csum;
57 __u16 dummy_csum = 0;
58 int offset = offsetof(struct ext4_inode, i_checksum_lo);
59 unsigned int csum_size = sizeof(dummy_csum);
60
61 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
62 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
63 offset += csum_size;
64 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
65 EXT4_GOOD_OLD_INODE_SIZE - offset);
66
67 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
68 offset = offsetof(struct ext4_inode, i_checksum_hi);
69 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
70 EXT4_GOOD_OLD_INODE_SIZE,
71 offset - EXT4_GOOD_OLD_INODE_SIZE);
72 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
73 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
74 csum_size);
75 offset += csum_size;
76 }
77 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
78 EXT4_INODE_SIZE(inode->i_sb) - offset);
79 }
80
81 return csum;
82}
83
84static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
85 struct ext4_inode_info *ei)
86{
87 __u32 provided, calculated;
88
89 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
90 cpu_to_le32(EXT4_OS_LINUX) ||
91 !ext4_has_metadata_csum(inode->i_sb))
92 return 1;
93
94 provided = le16_to_cpu(raw->i_checksum_lo);
95 calculated = ext4_inode_csum(inode, raw, ei);
96 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
97 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
98 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
99 else
100 calculated &= 0xFFFF;
101
102 return provided == calculated;
103}
104
105void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
106 struct ext4_inode_info *ei)
107{
108 __u32 csum;
109
110 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
111 cpu_to_le32(EXT4_OS_LINUX) ||
112 !ext4_has_metadata_csum(inode->i_sb))
113 return;
114
115 csum = ext4_inode_csum(inode, raw, ei);
116 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
117 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
118 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
119 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
120}
121
122static inline int ext4_begin_ordered_truncate(struct inode *inode,
123 loff_t new_size)
124{
125 trace_ext4_begin_ordered_truncate(inode, new_size);
126 /*
127 * If jinode is zero, then we never opened the file for
128 * writing, so there's no need to call
129 * jbd2_journal_begin_ordered_truncate() since there's no
130 * outstanding writes we need to flush.
131 */
132 if (!EXT4_I(inode)->jinode)
133 return 0;
134 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
135 EXT4_I(inode)->jinode,
136 new_size);
137}
138
139static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
140 int pextents);
141
142/*
143 * Test whether an inode is a fast symlink.
144 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
145 */
146int ext4_inode_is_fast_symlink(struct inode *inode)
147{
148 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
149 int ea_blocks = EXT4_I(inode)->i_file_acl ?
150 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
151
152 if (ext4_has_inline_data(inode))
153 return 0;
154
155 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
156 }
157 return S_ISLNK(inode->i_mode) && inode->i_size &&
158 (inode->i_size < EXT4_N_BLOCKS * 4);
159}
160
161/*
162 * Called at the last iput() if i_nlink is zero.
163 */
164void ext4_evict_inode(struct inode *inode)
165{
166 handle_t *handle;
167 int err;
168 /*
169 * Credits for final inode cleanup and freeing:
170 * sb + inode (ext4_orphan_del()), block bitmap, group descriptor
171 * (xattr block freeing), bitmap, group descriptor (inode freeing)
172 */
173 int extra_credits = 6;
174 struct ext4_xattr_inode_array *ea_inode_array = NULL;
175 bool freeze_protected = false;
176
177 trace_ext4_evict_inode(inode);
178
179 if (EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)
180 ext4_evict_ea_inode(inode);
181 if (inode->i_nlink) {
182 truncate_inode_pages_final(&inode->i_data);
183
184 goto no_delete;
185 }
186
187 if (is_bad_inode(inode))
188 goto no_delete;
189 dquot_initialize(inode);
190
191 if (ext4_should_order_data(inode))
192 ext4_begin_ordered_truncate(inode, 0);
193 truncate_inode_pages_final(&inode->i_data);
194
195 /*
196 * For inodes with journalled data, transaction commit could have
197 * dirtied the inode. And for inodes with dioread_nolock, unwritten
198 * extents converting worker could merge extents and also have dirtied
199 * the inode. Flush worker is ignoring it because of I_FREEING flag but
200 * we still need to remove the inode from the writeback lists.
201 */
202 if (!list_empty_careful(&inode->i_io_list))
203 inode_io_list_del(inode);
204
205 /*
206 * Protect us against freezing - iput() caller didn't have to have any
207 * protection against it. When we are in a running transaction though,
208 * we are already protected against freezing and we cannot grab further
209 * protection due to lock ordering constraints.
210 */
211 if (!ext4_journal_current_handle()) {
212 sb_start_intwrite(inode->i_sb);
213 freeze_protected = true;
214 }
215
216 if (!IS_NOQUOTA(inode))
217 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
218
219 /*
220 * Block bitmap, group descriptor, and inode are accounted in both
221 * ext4_blocks_for_truncate() and extra_credits. So subtract 3.
222 */
223 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
224 ext4_blocks_for_truncate(inode) + extra_credits - 3);
225 if (IS_ERR(handle)) {
226 ext4_std_error(inode->i_sb, PTR_ERR(handle));
227 /*
228 * If we're going to skip the normal cleanup, we still need to
229 * make sure that the in-core orphan linked list is properly
230 * cleaned up.
231 */
232 ext4_orphan_del(NULL, inode);
233 if (freeze_protected)
234 sb_end_intwrite(inode->i_sb);
235 goto no_delete;
236 }
237
238 if (IS_SYNC(inode))
239 ext4_handle_sync(handle);
240
241 /*
242 * Set inode->i_size to 0 before calling ext4_truncate(). We need
243 * special handling of symlinks here because i_size is used to
244 * determine whether ext4_inode_info->i_data contains symlink data or
245 * block mappings. Setting i_size to 0 will remove its fast symlink
246 * status. Erase i_data so that it becomes a valid empty block map.
247 */
248 if (ext4_inode_is_fast_symlink(inode))
249 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
250 inode->i_size = 0;
251 err = ext4_mark_inode_dirty(handle, inode);
252 if (err) {
253 ext4_warning(inode->i_sb,
254 "couldn't mark inode dirty (err %d)", err);
255 goto stop_handle;
256 }
257 if (inode->i_blocks) {
258 err = ext4_truncate(inode);
259 if (err) {
260 ext4_error_err(inode->i_sb, -err,
261 "couldn't truncate inode %lu (err %d)",
262 inode->i_ino, err);
263 goto stop_handle;
264 }
265 }
266
267 /* Remove xattr references. */
268 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
269 extra_credits);
270 if (err) {
271 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
272stop_handle:
273 ext4_journal_stop(handle);
274 ext4_orphan_del(NULL, inode);
275 if (freeze_protected)
276 sb_end_intwrite(inode->i_sb);
277 ext4_xattr_inode_array_free(ea_inode_array);
278 goto no_delete;
279 }
280
281 /*
282 * Kill off the orphan record which ext4_truncate created.
283 * AKPM: I think this can be inside the above `if'.
284 * Note that ext4_orphan_del() has to be able to cope with the
285 * deletion of a non-existent orphan - this is because we don't
286 * know if ext4_truncate() actually created an orphan record.
287 * (Well, we could do this if we need to, but heck - it works)
288 */
289 ext4_orphan_del(handle, inode);
290 EXT4_I(inode)->i_dtime = (__u32)ktime_get_real_seconds();
291
292 /*
293 * One subtle ordering requirement: if anything has gone wrong
294 * (transaction abort, IO errors, whatever), then we can still
295 * do these next steps (the fs will already have been marked as
296 * having errors), but we can't free the inode if the mark_dirty
297 * fails.
298 */
299 if (ext4_mark_inode_dirty(handle, inode))
300 /* If that failed, just do the required in-core inode clear. */
301 ext4_clear_inode(inode);
302 else
303 ext4_free_inode(handle, inode);
304 ext4_journal_stop(handle);
305 if (freeze_protected)
306 sb_end_intwrite(inode->i_sb);
307 ext4_xattr_inode_array_free(ea_inode_array);
308 return;
309no_delete:
310 /*
311 * Check out some where else accidentally dirty the evicting inode,
312 * which may probably cause inode use-after-free issues later.
313 */
314 WARN_ON_ONCE(!list_empty_careful(&inode->i_io_list));
315
316 if (!list_empty(&EXT4_I(inode)->i_fc_list))
317 ext4_fc_mark_ineligible(inode->i_sb, EXT4_FC_REASON_NOMEM, NULL);
318 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
319}
320
321#ifdef CONFIG_QUOTA
322qsize_t *ext4_get_reserved_space(struct inode *inode)
323{
324 return &EXT4_I(inode)->i_reserved_quota;
325}
326#endif
327
328/*
329 * Called with i_data_sem down, which is important since we can call
330 * ext4_discard_preallocations() from here.
331 */
332void ext4_da_update_reserve_space(struct inode *inode,
333 int used, int quota_claim)
334{
335 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
336 struct ext4_inode_info *ei = EXT4_I(inode);
337
338 spin_lock(&ei->i_block_reservation_lock);
339 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
340 if (unlikely(used > ei->i_reserved_data_blocks)) {
341 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
342 "with only %d reserved data blocks",
343 __func__, inode->i_ino, used,
344 ei->i_reserved_data_blocks);
345 WARN_ON(1);
346 used = ei->i_reserved_data_blocks;
347 }
348
349 /* Update per-inode reservations */
350 ei->i_reserved_data_blocks -= used;
351 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
352
353 spin_unlock(&ei->i_block_reservation_lock);
354
355 /* Update quota subsystem for data blocks */
356 if (quota_claim)
357 dquot_claim_block(inode, EXT4_C2B(sbi, used));
358 else {
359 /*
360 * We did fallocate with an offset that is already delayed
361 * allocated. So on delayed allocated writeback we should
362 * not re-claim the quota for fallocated blocks.
363 */
364 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
365 }
366
367 /*
368 * If we have done all the pending block allocations and if
369 * there aren't any writers on the inode, we can discard the
370 * inode's preallocations.
371 */
372 if ((ei->i_reserved_data_blocks == 0) &&
373 !inode_is_open_for_write(inode))
374 ext4_discard_preallocations(inode);
375}
376
377static int __check_block_validity(struct inode *inode, const char *func,
378 unsigned int line,
379 struct ext4_map_blocks *map)
380{
381 if (ext4_has_feature_journal(inode->i_sb) &&
382 (inode->i_ino ==
383 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
384 return 0;
385 if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
386 ext4_error_inode(inode, func, line, map->m_pblk,
387 "lblock %lu mapped to illegal pblock %llu "
388 "(length %d)", (unsigned long) map->m_lblk,
389 map->m_pblk, map->m_len);
390 return -EFSCORRUPTED;
391 }
392 return 0;
393}
394
395int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
396 ext4_lblk_t len)
397{
398 int ret;
399
400 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
401 return fscrypt_zeroout_range(inode, lblk, pblk, len);
402
403 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
404 if (ret > 0)
405 ret = 0;
406
407 return ret;
408}
409
410#define check_block_validity(inode, map) \
411 __check_block_validity((inode), __func__, __LINE__, (map))
412
413#ifdef ES_AGGRESSIVE_TEST
414static void ext4_map_blocks_es_recheck(handle_t *handle,
415 struct inode *inode,
416 struct ext4_map_blocks *es_map,
417 struct ext4_map_blocks *map,
418 int flags)
419{
420 int retval;
421
422 map->m_flags = 0;
423 /*
424 * There is a race window that the result is not the same.
425 * e.g. xfstests #223 when dioread_nolock enables. The reason
426 * is that we lookup a block mapping in extent status tree with
427 * out taking i_data_sem. So at the time the unwritten extent
428 * could be converted.
429 */
430 down_read(&EXT4_I(inode)->i_data_sem);
431 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
432 retval = ext4_ext_map_blocks(handle, inode, map, 0);
433 } else {
434 retval = ext4_ind_map_blocks(handle, inode, map, 0);
435 }
436 up_read((&EXT4_I(inode)->i_data_sem));
437
438 /*
439 * We don't check m_len because extent will be collpased in status
440 * tree. So the m_len might not equal.
441 */
442 if (es_map->m_lblk != map->m_lblk ||
443 es_map->m_flags != map->m_flags ||
444 es_map->m_pblk != map->m_pblk) {
445 printk("ES cache assertion failed for inode: %lu "
446 "es_cached ex [%d/%d/%llu/%x] != "
447 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
448 inode->i_ino, es_map->m_lblk, es_map->m_len,
449 es_map->m_pblk, es_map->m_flags, map->m_lblk,
450 map->m_len, map->m_pblk, map->m_flags,
451 retval, flags);
452 }
453}
454#endif /* ES_AGGRESSIVE_TEST */
455
456/*
457 * The ext4_map_blocks() function tries to look up the requested blocks,
458 * and returns if the blocks are already mapped.
459 *
460 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
461 * and store the allocated blocks in the result buffer head and mark it
462 * mapped.
463 *
464 * If file type is extents based, it will call ext4_ext_map_blocks(),
465 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
466 * based files
467 *
468 * On success, it returns the number of blocks being mapped or allocated.
469 * If flags doesn't contain EXT4_GET_BLOCKS_CREATE the blocks are
470 * pre-allocated and unwritten, the resulting @map is marked as unwritten.
471 * If the flags contain EXT4_GET_BLOCKS_CREATE, it will mark @map as mapped.
472 *
473 * It returns 0 if plain look up failed (blocks have not been allocated), in
474 * that case, @map is returned as unmapped but we still do fill map->m_len to
475 * indicate the length of a hole starting at map->m_lblk.
476 *
477 * It returns the error in case of allocation failure.
478 */
479int ext4_map_blocks(handle_t *handle, struct inode *inode,
480 struct ext4_map_blocks *map, int flags)
481{
482 struct extent_status es;
483 int retval;
484 int ret = 0;
485#ifdef ES_AGGRESSIVE_TEST
486 struct ext4_map_blocks orig_map;
487
488 memcpy(&orig_map, map, sizeof(*map));
489#endif
490
491 map->m_flags = 0;
492 ext_debug(inode, "flag 0x%x, max_blocks %u, logical block %lu\n",
493 flags, map->m_len, (unsigned long) map->m_lblk);
494
495 /*
496 * ext4_map_blocks returns an int, and m_len is an unsigned int
497 */
498 if (unlikely(map->m_len > INT_MAX))
499 map->m_len = INT_MAX;
500
501 /* We can handle the block number less than EXT_MAX_BLOCKS */
502 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
503 return -EFSCORRUPTED;
504
505 /* Lookup extent status tree firstly */
506 if (!(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY) &&
507 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
508 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
509 map->m_pblk = ext4_es_pblock(&es) +
510 map->m_lblk - es.es_lblk;
511 map->m_flags |= ext4_es_is_written(&es) ?
512 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
513 retval = es.es_len - (map->m_lblk - es.es_lblk);
514 if (retval > map->m_len)
515 retval = map->m_len;
516 map->m_len = retval;
517 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
518 map->m_pblk = 0;
519 map->m_flags |= ext4_es_is_delayed(&es) ?
520 EXT4_MAP_DELAYED : 0;
521 retval = es.es_len - (map->m_lblk - es.es_lblk);
522 if (retval > map->m_len)
523 retval = map->m_len;
524 map->m_len = retval;
525 retval = 0;
526 } else {
527 BUG();
528 }
529
530 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
531 return retval;
532#ifdef ES_AGGRESSIVE_TEST
533 ext4_map_blocks_es_recheck(handle, inode, map,
534 &orig_map, flags);
535#endif
536 goto found;
537 }
538 /*
539 * In the query cache no-wait mode, nothing we can do more if we
540 * cannot find extent in the cache.
541 */
542 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
543 return 0;
544
545 /*
546 * Try to see if we can get the block without requesting a new
547 * file system block.
548 */
549 down_read(&EXT4_I(inode)->i_data_sem);
550 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
551 retval = ext4_ext_map_blocks(handle, inode, map, 0);
552 } else {
553 retval = ext4_ind_map_blocks(handle, inode, map, 0);
554 }
555 if (retval > 0) {
556 unsigned int status;
557
558 if (unlikely(retval != map->m_len)) {
559 ext4_warning(inode->i_sb,
560 "ES len assertion failed for inode "
561 "%lu: retval %d != map->m_len %d",
562 inode->i_ino, retval, map->m_len);
563 WARN_ON(1);
564 }
565
566 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
567 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
568 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
569 !(status & EXTENT_STATUS_WRITTEN) &&
570 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
571 map->m_lblk + map->m_len - 1))
572 status |= EXTENT_STATUS_DELAYED;
573 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
574 map->m_pblk, status);
575 }
576 up_read((&EXT4_I(inode)->i_data_sem));
577
578found:
579 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
580 ret = check_block_validity(inode, map);
581 if (ret != 0)
582 return ret;
583 }
584
585 /* If it is only a block(s) look up */
586 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
587 return retval;
588
589 /*
590 * Returns if the blocks have already allocated
591 *
592 * Note that if blocks have been preallocated
593 * ext4_ext_map_blocks() returns with buffer head unmapped
594 */
595 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
596 /*
597 * If we need to convert extent to unwritten
598 * we continue and do the actual work in
599 * ext4_ext_map_blocks()
600 */
601 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
602 return retval;
603
604 /*
605 * Here we clear m_flags because after allocating an new extent,
606 * it will be set again.
607 */
608 map->m_flags &= ~EXT4_MAP_FLAGS;
609
610 /*
611 * New blocks allocate and/or writing to unwritten extent
612 * will possibly result in updating i_data, so we take
613 * the write lock of i_data_sem, and call get_block()
614 * with create == 1 flag.
615 */
616 down_write(&EXT4_I(inode)->i_data_sem);
617
618 /*
619 * We need to check for EXT4 here because migrate
620 * could have changed the inode type in between
621 */
622 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
623 retval = ext4_ext_map_blocks(handle, inode, map, flags);
624 } else {
625 retval = ext4_ind_map_blocks(handle, inode, map, flags);
626
627 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
628 /*
629 * We allocated new blocks which will result in
630 * i_data's format changing. Force the migrate
631 * to fail by clearing migrate flags
632 */
633 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
634 }
635 }
636
637 if (retval > 0) {
638 unsigned int status;
639
640 if (unlikely(retval != map->m_len)) {
641 ext4_warning(inode->i_sb,
642 "ES len assertion failed for inode "
643 "%lu: retval %d != map->m_len %d",
644 inode->i_ino, retval, map->m_len);
645 WARN_ON(1);
646 }
647
648 /*
649 * We have to zeroout blocks before inserting them into extent
650 * status tree. Otherwise someone could look them up there and
651 * use them before they are really zeroed. We also have to
652 * unmap metadata before zeroing as otherwise writeback can
653 * overwrite zeros with stale data from block device.
654 */
655 if (flags & EXT4_GET_BLOCKS_ZERO &&
656 map->m_flags & EXT4_MAP_MAPPED &&
657 map->m_flags & EXT4_MAP_NEW) {
658 ret = ext4_issue_zeroout(inode, map->m_lblk,
659 map->m_pblk, map->m_len);
660 if (ret) {
661 retval = ret;
662 goto out_sem;
663 }
664 }
665
666 /*
667 * If the extent has been zeroed out, we don't need to update
668 * extent status tree.
669 */
670 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
671 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
672 if (ext4_es_is_written(&es))
673 goto out_sem;
674 }
675 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
676 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
677 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
678 !(status & EXTENT_STATUS_WRITTEN) &&
679 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
680 map->m_lblk + map->m_len - 1))
681 status |= EXTENT_STATUS_DELAYED;
682 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
683 map->m_pblk, status);
684 }
685
686out_sem:
687 up_write((&EXT4_I(inode)->i_data_sem));
688 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
689 ret = check_block_validity(inode, map);
690 if (ret != 0)
691 return ret;
692
693 /*
694 * Inodes with freshly allocated blocks where contents will be
695 * visible after transaction commit must be on transaction's
696 * ordered data list.
697 */
698 if (map->m_flags & EXT4_MAP_NEW &&
699 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
700 !(flags & EXT4_GET_BLOCKS_ZERO) &&
701 !ext4_is_quota_file(inode) &&
702 ext4_should_order_data(inode)) {
703 loff_t start_byte =
704 (loff_t)map->m_lblk << inode->i_blkbits;
705 loff_t length = (loff_t)map->m_len << inode->i_blkbits;
706
707 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
708 ret = ext4_jbd2_inode_add_wait(handle, inode,
709 start_byte, length);
710 else
711 ret = ext4_jbd2_inode_add_write(handle, inode,
712 start_byte, length);
713 if (ret)
714 return ret;
715 }
716 }
717 if (retval > 0 && (map->m_flags & EXT4_MAP_UNWRITTEN ||
718 map->m_flags & EXT4_MAP_MAPPED))
719 ext4_fc_track_range(handle, inode, map->m_lblk,
720 map->m_lblk + map->m_len - 1);
721 if (retval < 0)
722 ext_debug(inode, "failed with err %d\n", retval);
723 return retval;
724}
725
726/*
727 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
728 * we have to be careful as someone else may be manipulating b_state as well.
729 */
730static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
731{
732 unsigned long old_state;
733 unsigned long new_state;
734
735 flags &= EXT4_MAP_FLAGS;
736
737 /* Dummy buffer_head? Set non-atomically. */
738 if (!bh->b_page) {
739 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
740 return;
741 }
742 /*
743 * Someone else may be modifying b_state. Be careful! This is ugly but
744 * once we get rid of using bh as a container for mapping information
745 * to pass to / from get_block functions, this can go away.
746 */
747 old_state = READ_ONCE(bh->b_state);
748 do {
749 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
750 } while (unlikely(!try_cmpxchg(&bh->b_state, &old_state, new_state)));
751}
752
753static int _ext4_get_block(struct inode *inode, sector_t iblock,
754 struct buffer_head *bh, int flags)
755{
756 struct ext4_map_blocks map;
757 int ret = 0;
758
759 if (ext4_has_inline_data(inode))
760 return -ERANGE;
761
762 map.m_lblk = iblock;
763 map.m_len = bh->b_size >> inode->i_blkbits;
764
765 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
766 flags);
767 if (ret > 0) {
768 map_bh(bh, inode->i_sb, map.m_pblk);
769 ext4_update_bh_state(bh, map.m_flags);
770 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
771 ret = 0;
772 } else if (ret == 0) {
773 /* hole case, need to fill in bh->b_size */
774 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
775 }
776 return ret;
777}
778
779int ext4_get_block(struct inode *inode, sector_t iblock,
780 struct buffer_head *bh, int create)
781{
782 return _ext4_get_block(inode, iblock, bh,
783 create ? EXT4_GET_BLOCKS_CREATE : 0);
784}
785
786/*
787 * Get block function used when preparing for buffered write if we require
788 * creating an unwritten extent if blocks haven't been allocated. The extent
789 * will be converted to written after the IO is complete.
790 */
791int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
792 struct buffer_head *bh_result, int create)
793{
794 int ret = 0;
795
796 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
797 inode->i_ino, create);
798 ret = _ext4_get_block(inode, iblock, bh_result,
799 EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT);
800
801 /*
802 * If the buffer is marked unwritten, mark it as new to make sure it is
803 * zeroed out correctly in case of partial writes. Otherwise, there is
804 * a chance of stale data getting exposed.
805 */
806 if (ret == 0 && buffer_unwritten(bh_result))
807 set_buffer_new(bh_result);
808
809 return ret;
810}
811
812/* Maximum number of blocks we map for direct IO at once. */
813#define DIO_MAX_BLOCKS 4096
814
815/*
816 * `handle' can be NULL if create is zero
817 */
818struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
819 ext4_lblk_t block, int map_flags)
820{
821 struct ext4_map_blocks map;
822 struct buffer_head *bh;
823 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
824 bool nowait = map_flags & EXT4_GET_BLOCKS_CACHED_NOWAIT;
825 int err;
826
827 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
828 || handle != NULL || create == 0);
829 ASSERT(create == 0 || !nowait);
830
831 map.m_lblk = block;
832 map.m_len = 1;
833 err = ext4_map_blocks(handle, inode, &map, map_flags);
834
835 if (err == 0)
836 return create ? ERR_PTR(-ENOSPC) : NULL;
837 if (err < 0)
838 return ERR_PTR(err);
839
840 if (nowait)
841 return sb_find_get_block(inode->i_sb, map.m_pblk);
842
843 bh = sb_getblk(inode->i_sb, map.m_pblk);
844 if (unlikely(!bh))
845 return ERR_PTR(-ENOMEM);
846 if (map.m_flags & EXT4_MAP_NEW) {
847 ASSERT(create != 0);
848 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
849 || (handle != NULL));
850
851 /*
852 * Now that we do not always journal data, we should
853 * keep in mind whether this should always journal the
854 * new buffer as metadata. For now, regular file
855 * writes use ext4_get_block instead, so it's not a
856 * problem.
857 */
858 lock_buffer(bh);
859 BUFFER_TRACE(bh, "call get_create_access");
860 err = ext4_journal_get_create_access(handle, inode->i_sb, bh,
861 EXT4_JTR_NONE);
862 if (unlikely(err)) {
863 unlock_buffer(bh);
864 goto errout;
865 }
866 if (!buffer_uptodate(bh)) {
867 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
868 set_buffer_uptodate(bh);
869 }
870 unlock_buffer(bh);
871 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
872 err = ext4_handle_dirty_metadata(handle, inode, bh);
873 if (unlikely(err))
874 goto errout;
875 } else
876 BUFFER_TRACE(bh, "not a new buffer");
877 return bh;
878errout:
879 brelse(bh);
880 return ERR_PTR(err);
881}
882
883struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
884 ext4_lblk_t block, int map_flags)
885{
886 struct buffer_head *bh;
887 int ret;
888
889 bh = ext4_getblk(handle, inode, block, map_flags);
890 if (IS_ERR(bh))
891 return bh;
892 if (!bh || ext4_buffer_uptodate(bh))
893 return bh;
894
895 ret = ext4_read_bh_lock(bh, REQ_META | REQ_PRIO, true);
896 if (ret) {
897 put_bh(bh);
898 return ERR_PTR(ret);
899 }
900 return bh;
901}
902
903/* Read a contiguous batch of blocks. */
904int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
905 bool wait, struct buffer_head **bhs)
906{
907 int i, err;
908
909 for (i = 0; i < bh_count; i++) {
910 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
911 if (IS_ERR(bhs[i])) {
912 err = PTR_ERR(bhs[i]);
913 bh_count = i;
914 goto out_brelse;
915 }
916 }
917
918 for (i = 0; i < bh_count; i++)
919 /* Note that NULL bhs[i] is valid because of holes. */
920 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
921 ext4_read_bh_lock(bhs[i], REQ_META | REQ_PRIO, false);
922
923 if (!wait)
924 return 0;
925
926 for (i = 0; i < bh_count; i++)
927 if (bhs[i])
928 wait_on_buffer(bhs[i]);
929
930 for (i = 0; i < bh_count; i++) {
931 if (bhs[i] && !buffer_uptodate(bhs[i])) {
932 err = -EIO;
933 goto out_brelse;
934 }
935 }
936 return 0;
937
938out_brelse:
939 for (i = 0; i < bh_count; i++) {
940 brelse(bhs[i]);
941 bhs[i] = NULL;
942 }
943 return err;
944}
945
946int ext4_walk_page_buffers(handle_t *handle, struct inode *inode,
947 struct buffer_head *head,
948 unsigned from,
949 unsigned to,
950 int *partial,
951 int (*fn)(handle_t *handle, struct inode *inode,
952 struct buffer_head *bh))
953{
954 struct buffer_head *bh;
955 unsigned block_start, block_end;
956 unsigned blocksize = head->b_size;
957 int err, ret = 0;
958 struct buffer_head *next;
959
960 for (bh = head, block_start = 0;
961 ret == 0 && (bh != head || !block_start);
962 block_start = block_end, bh = next) {
963 next = bh->b_this_page;
964 block_end = block_start + blocksize;
965 if (block_end <= from || block_start >= to) {
966 if (partial && !buffer_uptodate(bh))
967 *partial = 1;
968 continue;
969 }
970 err = (*fn)(handle, inode, bh);
971 if (!ret)
972 ret = err;
973 }
974 return ret;
975}
976
977/*
978 * Helper for handling dirtying of journalled data. We also mark the folio as
979 * dirty so that writeback code knows about this page (and inode) contains
980 * dirty data. ext4_writepages() then commits appropriate transaction to
981 * make data stable.
982 */
983static int ext4_dirty_journalled_data(handle_t *handle, struct buffer_head *bh)
984{
985 folio_mark_dirty(bh->b_folio);
986 return ext4_handle_dirty_metadata(handle, NULL, bh);
987}
988
989int do_journal_get_write_access(handle_t *handle, struct inode *inode,
990 struct buffer_head *bh)
991{
992 int dirty = buffer_dirty(bh);
993 int ret;
994
995 if (!buffer_mapped(bh) || buffer_freed(bh))
996 return 0;
997 /*
998 * __block_write_begin() could have dirtied some buffers. Clean
999 * the dirty bit as jbd2_journal_get_write_access() could complain
1000 * otherwise about fs integrity issues. Setting of the dirty bit
1001 * by __block_write_begin() isn't a real problem here as we clear
1002 * the bit before releasing a page lock and thus writeback cannot
1003 * ever write the buffer.
1004 */
1005 if (dirty)
1006 clear_buffer_dirty(bh);
1007 BUFFER_TRACE(bh, "get write access");
1008 ret = ext4_journal_get_write_access(handle, inode->i_sb, bh,
1009 EXT4_JTR_NONE);
1010 if (!ret && dirty)
1011 ret = ext4_dirty_journalled_data(handle, bh);
1012 return ret;
1013}
1014
1015#ifdef CONFIG_FS_ENCRYPTION
1016static int ext4_block_write_begin(struct folio *folio, loff_t pos, unsigned len,
1017 get_block_t *get_block)
1018{
1019 unsigned from = pos & (PAGE_SIZE - 1);
1020 unsigned to = from + len;
1021 struct inode *inode = folio->mapping->host;
1022 unsigned block_start, block_end;
1023 sector_t block;
1024 int err = 0;
1025 unsigned blocksize = inode->i_sb->s_blocksize;
1026 unsigned bbits;
1027 struct buffer_head *bh, *head, *wait[2];
1028 int nr_wait = 0;
1029 int i;
1030
1031 BUG_ON(!folio_test_locked(folio));
1032 BUG_ON(from > PAGE_SIZE);
1033 BUG_ON(to > PAGE_SIZE);
1034 BUG_ON(from > to);
1035
1036 head = folio_buffers(folio);
1037 if (!head)
1038 head = create_empty_buffers(folio, blocksize, 0);
1039 bbits = ilog2(blocksize);
1040 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
1041
1042 for (bh = head, block_start = 0; bh != head || !block_start;
1043 block++, block_start = block_end, bh = bh->b_this_page) {
1044 block_end = block_start + blocksize;
1045 if (block_end <= from || block_start >= to) {
1046 if (folio_test_uptodate(folio)) {
1047 set_buffer_uptodate(bh);
1048 }
1049 continue;
1050 }
1051 if (buffer_new(bh))
1052 clear_buffer_new(bh);
1053 if (!buffer_mapped(bh)) {
1054 WARN_ON(bh->b_size != blocksize);
1055 err = get_block(inode, block, bh, 1);
1056 if (err)
1057 break;
1058 if (buffer_new(bh)) {
1059 if (folio_test_uptodate(folio)) {
1060 clear_buffer_new(bh);
1061 set_buffer_uptodate(bh);
1062 mark_buffer_dirty(bh);
1063 continue;
1064 }
1065 if (block_end > to || block_start < from)
1066 folio_zero_segments(folio, to,
1067 block_end,
1068 block_start, from);
1069 continue;
1070 }
1071 }
1072 if (folio_test_uptodate(folio)) {
1073 set_buffer_uptodate(bh);
1074 continue;
1075 }
1076 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1077 !buffer_unwritten(bh) &&
1078 (block_start < from || block_end > to)) {
1079 ext4_read_bh_lock(bh, 0, false);
1080 wait[nr_wait++] = bh;
1081 }
1082 }
1083 /*
1084 * If we issued read requests, let them complete.
1085 */
1086 for (i = 0; i < nr_wait; i++) {
1087 wait_on_buffer(wait[i]);
1088 if (!buffer_uptodate(wait[i]))
1089 err = -EIO;
1090 }
1091 if (unlikely(err)) {
1092 folio_zero_new_buffers(folio, from, to);
1093 } else if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
1094 for (i = 0; i < nr_wait; i++) {
1095 int err2;
1096
1097 err2 = fscrypt_decrypt_pagecache_blocks(folio,
1098 blocksize, bh_offset(wait[i]));
1099 if (err2) {
1100 clear_buffer_uptodate(wait[i]);
1101 err = err2;
1102 }
1103 }
1104 }
1105
1106 return err;
1107}
1108#endif
1109
1110/*
1111 * To preserve ordering, it is essential that the hole instantiation and
1112 * the data write be encapsulated in a single transaction. We cannot
1113 * close off a transaction and start a new one between the ext4_get_block()
1114 * and the ext4_write_end(). So doing the jbd2_journal_start at the start of
1115 * ext4_write_begin() is the right place.
1116 */
1117static int ext4_write_begin(struct file *file, struct address_space *mapping,
1118 loff_t pos, unsigned len,
1119 struct page **pagep, void **fsdata)
1120{
1121 struct inode *inode = mapping->host;
1122 int ret, needed_blocks;
1123 handle_t *handle;
1124 int retries = 0;
1125 struct folio *folio;
1126 pgoff_t index;
1127 unsigned from, to;
1128
1129 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
1130 return -EIO;
1131
1132 trace_ext4_write_begin(inode, pos, len);
1133 /*
1134 * Reserve one block more for addition to orphan list in case
1135 * we allocate blocks but write fails for some reason
1136 */
1137 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1138 index = pos >> PAGE_SHIFT;
1139 from = pos & (PAGE_SIZE - 1);
1140 to = from + len;
1141
1142 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1143 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1144 pagep);
1145 if (ret < 0)
1146 return ret;
1147 if (ret == 1)
1148 return 0;
1149 }
1150
1151 /*
1152 * __filemap_get_folio() can take a long time if the
1153 * system is thrashing due to memory pressure, or if the folio
1154 * is being written back. So grab it first before we start
1155 * the transaction handle. This also allows us to allocate
1156 * the folio (if needed) without using GFP_NOFS.
1157 */
1158retry_grab:
1159 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
1160 mapping_gfp_mask(mapping));
1161 if (IS_ERR(folio))
1162 return PTR_ERR(folio);
1163 /*
1164 * The same as page allocation, we prealloc buffer heads before
1165 * starting the handle.
1166 */
1167 if (!folio_buffers(folio))
1168 create_empty_buffers(folio, inode->i_sb->s_blocksize, 0);
1169
1170 folio_unlock(folio);
1171
1172retry_journal:
1173 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1174 if (IS_ERR(handle)) {
1175 folio_put(folio);
1176 return PTR_ERR(handle);
1177 }
1178
1179 folio_lock(folio);
1180 if (folio->mapping != mapping) {
1181 /* The folio got truncated from under us */
1182 folio_unlock(folio);
1183 folio_put(folio);
1184 ext4_journal_stop(handle);
1185 goto retry_grab;
1186 }
1187 /* In case writeback began while the folio was unlocked */
1188 folio_wait_stable(folio);
1189
1190#ifdef CONFIG_FS_ENCRYPTION
1191 if (ext4_should_dioread_nolock(inode))
1192 ret = ext4_block_write_begin(folio, pos, len,
1193 ext4_get_block_unwritten);
1194 else
1195 ret = ext4_block_write_begin(folio, pos, len, ext4_get_block);
1196#else
1197 if (ext4_should_dioread_nolock(inode))
1198 ret = __block_write_begin(&folio->page, pos, len,
1199 ext4_get_block_unwritten);
1200 else
1201 ret = __block_write_begin(&folio->page, pos, len, ext4_get_block);
1202#endif
1203 if (!ret && ext4_should_journal_data(inode)) {
1204 ret = ext4_walk_page_buffers(handle, inode,
1205 folio_buffers(folio), from, to,
1206 NULL, do_journal_get_write_access);
1207 }
1208
1209 if (ret) {
1210 bool extended = (pos + len > inode->i_size) &&
1211 !ext4_verity_in_progress(inode);
1212
1213 folio_unlock(folio);
1214 /*
1215 * __block_write_begin may have instantiated a few blocks
1216 * outside i_size. Trim these off again. Don't need
1217 * i_size_read because we hold i_rwsem.
1218 *
1219 * Add inode to orphan list in case we crash before
1220 * truncate finishes
1221 */
1222 if (extended && ext4_can_truncate(inode))
1223 ext4_orphan_add(handle, inode);
1224
1225 ext4_journal_stop(handle);
1226 if (extended) {
1227 ext4_truncate_failed_write(inode);
1228 /*
1229 * If truncate failed early the inode might
1230 * still be on the orphan list; we need to
1231 * make sure the inode is removed from the
1232 * orphan list in that case.
1233 */
1234 if (inode->i_nlink)
1235 ext4_orphan_del(NULL, inode);
1236 }
1237
1238 if (ret == -ENOSPC &&
1239 ext4_should_retry_alloc(inode->i_sb, &retries))
1240 goto retry_journal;
1241 folio_put(folio);
1242 return ret;
1243 }
1244 *pagep = &folio->page;
1245 return ret;
1246}
1247
1248/* For write_end() in data=journal mode */
1249static int write_end_fn(handle_t *handle, struct inode *inode,
1250 struct buffer_head *bh)
1251{
1252 int ret;
1253 if (!buffer_mapped(bh) || buffer_freed(bh))
1254 return 0;
1255 set_buffer_uptodate(bh);
1256 ret = ext4_dirty_journalled_data(handle, bh);
1257 clear_buffer_meta(bh);
1258 clear_buffer_prio(bh);
1259 return ret;
1260}
1261
1262/*
1263 * We need to pick up the new inode size which generic_commit_write gave us
1264 * `file' can be NULL - eg, when called from page_symlink().
1265 *
1266 * ext4 never places buffers on inode->i_mapping->i_private_list. metadata
1267 * buffers are managed internally.
1268 */
1269static int ext4_write_end(struct file *file,
1270 struct address_space *mapping,
1271 loff_t pos, unsigned len, unsigned copied,
1272 struct page *page, void *fsdata)
1273{
1274 struct folio *folio = page_folio(page);
1275 handle_t *handle = ext4_journal_current_handle();
1276 struct inode *inode = mapping->host;
1277 loff_t old_size = inode->i_size;
1278 int ret = 0, ret2;
1279 int i_size_changed = 0;
1280 bool verity = ext4_verity_in_progress(inode);
1281
1282 trace_ext4_write_end(inode, pos, len, copied);
1283
1284 if (ext4_has_inline_data(inode) &&
1285 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
1286 return ext4_write_inline_data_end(inode, pos, len, copied,
1287 folio);
1288
1289 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1290 /*
1291 * it's important to update i_size while still holding folio lock:
1292 * page writeout could otherwise come in and zero beyond i_size.
1293 *
1294 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1295 * blocks are being written past EOF, so skip the i_size update.
1296 */
1297 if (!verity)
1298 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1299 folio_unlock(folio);
1300 folio_put(folio);
1301
1302 if (old_size < pos && !verity)
1303 pagecache_isize_extended(inode, old_size, pos);
1304 /*
1305 * Don't mark the inode dirty under folio lock. First, it unnecessarily
1306 * makes the holding time of folio lock longer. Second, it forces lock
1307 * ordering of folio lock and transaction start for journaling
1308 * filesystems.
1309 */
1310 if (i_size_changed)
1311 ret = ext4_mark_inode_dirty(handle, inode);
1312
1313 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1314 /* if we have allocated more blocks and copied
1315 * less. We will have blocks allocated outside
1316 * inode->i_size. So truncate them
1317 */
1318 ext4_orphan_add(handle, inode);
1319
1320 ret2 = ext4_journal_stop(handle);
1321 if (!ret)
1322 ret = ret2;
1323
1324 if (pos + len > inode->i_size && !verity) {
1325 ext4_truncate_failed_write(inode);
1326 /*
1327 * If truncate failed early the inode might still be
1328 * on the orphan list; we need to make sure the inode
1329 * is removed from the orphan list in that case.
1330 */
1331 if (inode->i_nlink)
1332 ext4_orphan_del(NULL, inode);
1333 }
1334
1335 return ret ? ret : copied;
1336}
1337
1338/*
1339 * This is a private version of folio_zero_new_buffers() which doesn't
1340 * set the buffer to be dirty, since in data=journalled mode we need
1341 * to call ext4_dirty_journalled_data() instead.
1342 */
1343static void ext4_journalled_zero_new_buffers(handle_t *handle,
1344 struct inode *inode,
1345 struct folio *folio,
1346 unsigned from, unsigned to)
1347{
1348 unsigned int block_start = 0, block_end;
1349 struct buffer_head *head, *bh;
1350
1351 bh = head = folio_buffers(folio);
1352 do {
1353 block_end = block_start + bh->b_size;
1354 if (buffer_new(bh)) {
1355 if (block_end > from && block_start < to) {
1356 if (!folio_test_uptodate(folio)) {
1357 unsigned start, size;
1358
1359 start = max(from, block_start);
1360 size = min(to, block_end) - start;
1361
1362 folio_zero_range(folio, start, size);
1363 write_end_fn(handle, inode, bh);
1364 }
1365 clear_buffer_new(bh);
1366 }
1367 }
1368 block_start = block_end;
1369 bh = bh->b_this_page;
1370 } while (bh != head);
1371}
1372
1373static int ext4_journalled_write_end(struct file *file,
1374 struct address_space *mapping,
1375 loff_t pos, unsigned len, unsigned copied,
1376 struct page *page, void *fsdata)
1377{
1378 struct folio *folio = page_folio(page);
1379 handle_t *handle = ext4_journal_current_handle();
1380 struct inode *inode = mapping->host;
1381 loff_t old_size = inode->i_size;
1382 int ret = 0, ret2;
1383 int partial = 0;
1384 unsigned from, to;
1385 int size_changed = 0;
1386 bool verity = ext4_verity_in_progress(inode);
1387
1388 trace_ext4_journalled_write_end(inode, pos, len, copied);
1389 from = pos & (PAGE_SIZE - 1);
1390 to = from + len;
1391
1392 BUG_ON(!ext4_handle_valid(handle));
1393
1394 if (ext4_has_inline_data(inode))
1395 return ext4_write_inline_data_end(inode, pos, len, copied,
1396 folio);
1397
1398 if (unlikely(copied < len) && !folio_test_uptodate(folio)) {
1399 copied = 0;
1400 ext4_journalled_zero_new_buffers(handle, inode, folio,
1401 from, to);
1402 } else {
1403 if (unlikely(copied < len))
1404 ext4_journalled_zero_new_buffers(handle, inode, folio,
1405 from + copied, to);
1406 ret = ext4_walk_page_buffers(handle, inode,
1407 folio_buffers(folio),
1408 from, from + copied, &partial,
1409 write_end_fn);
1410 if (!partial)
1411 folio_mark_uptodate(folio);
1412 }
1413 if (!verity)
1414 size_changed = ext4_update_inode_size(inode, pos + copied);
1415 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1416 folio_unlock(folio);
1417 folio_put(folio);
1418
1419 if (old_size < pos && !verity)
1420 pagecache_isize_extended(inode, old_size, pos);
1421
1422 if (size_changed) {
1423 ret2 = ext4_mark_inode_dirty(handle, inode);
1424 if (!ret)
1425 ret = ret2;
1426 }
1427
1428 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1429 /* if we have allocated more blocks and copied
1430 * less. We will have blocks allocated outside
1431 * inode->i_size. So truncate them
1432 */
1433 ext4_orphan_add(handle, inode);
1434
1435 ret2 = ext4_journal_stop(handle);
1436 if (!ret)
1437 ret = ret2;
1438 if (pos + len > inode->i_size && !verity) {
1439 ext4_truncate_failed_write(inode);
1440 /*
1441 * If truncate failed early the inode might still be
1442 * on the orphan list; we need to make sure the inode
1443 * is removed from the orphan list in that case.
1444 */
1445 if (inode->i_nlink)
1446 ext4_orphan_del(NULL, inode);
1447 }
1448
1449 return ret ? ret : copied;
1450}
1451
1452/*
1453 * Reserve space for a single cluster
1454 */
1455static int ext4_da_reserve_space(struct inode *inode)
1456{
1457 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1458 struct ext4_inode_info *ei = EXT4_I(inode);
1459 int ret;
1460
1461 /*
1462 * We will charge metadata quota at writeout time; this saves
1463 * us from metadata over-estimation, though we may go over by
1464 * a small amount in the end. Here we just reserve for data.
1465 */
1466 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1467 if (ret)
1468 return ret;
1469
1470 spin_lock(&ei->i_block_reservation_lock);
1471 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1472 spin_unlock(&ei->i_block_reservation_lock);
1473 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1474 return -ENOSPC;
1475 }
1476 ei->i_reserved_data_blocks++;
1477 trace_ext4_da_reserve_space(inode);
1478 spin_unlock(&ei->i_block_reservation_lock);
1479
1480 return 0; /* success */
1481}
1482
1483void ext4_da_release_space(struct inode *inode, int to_free)
1484{
1485 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1486 struct ext4_inode_info *ei = EXT4_I(inode);
1487
1488 if (!to_free)
1489 return; /* Nothing to release, exit */
1490
1491 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1492
1493 trace_ext4_da_release_space(inode, to_free);
1494 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1495 /*
1496 * if there aren't enough reserved blocks, then the
1497 * counter is messed up somewhere. Since this
1498 * function is called from invalidate page, it's
1499 * harmless to return without any action.
1500 */
1501 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1502 "ino %lu, to_free %d with only %d reserved "
1503 "data blocks", inode->i_ino, to_free,
1504 ei->i_reserved_data_blocks);
1505 WARN_ON(1);
1506 to_free = ei->i_reserved_data_blocks;
1507 }
1508 ei->i_reserved_data_blocks -= to_free;
1509
1510 /* update fs dirty data blocks counter */
1511 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1512
1513 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1514
1515 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1516}
1517
1518/*
1519 * Delayed allocation stuff
1520 */
1521
1522struct mpage_da_data {
1523 /* These are input fields for ext4_do_writepages() */
1524 struct inode *inode;
1525 struct writeback_control *wbc;
1526 unsigned int can_map:1; /* Can writepages call map blocks? */
1527
1528 /* These are internal state of ext4_do_writepages() */
1529 pgoff_t first_page; /* The first page to write */
1530 pgoff_t next_page; /* Current page to examine */
1531 pgoff_t last_page; /* Last page to examine */
1532 /*
1533 * Extent to map - this can be after first_page because that can be
1534 * fully mapped. We somewhat abuse m_flags to store whether the extent
1535 * is delalloc or unwritten.
1536 */
1537 struct ext4_map_blocks map;
1538 struct ext4_io_submit io_submit; /* IO submission data */
1539 unsigned int do_map:1;
1540 unsigned int scanned_until_end:1;
1541 unsigned int journalled_more_data:1;
1542};
1543
1544static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1545 bool invalidate)
1546{
1547 unsigned nr, i;
1548 pgoff_t index, end;
1549 struct folio_batch fbatch;
1550 struct inode *inode = mpd->inode;
1551 struct address_space *mapping = inode->i_mapping;
1552
1553 /* This is necessary when next_page == 0. */
1554 if (mpd->first_page >= mpd->next_page)
1555 return;
1556
1557 mpd->scanned_until_end = 0;
1558 index = mpd->first_page;
1559 end = mpd->next_page - 1;
1560 if (invalidate) {
1561 ext4_lblk_t start, last;
1562 start = index << (PAGE_SHIFT - inode->i_blkbits);
1563 last = end << (PAGE_SHIFT - inode->i_blkbits);
1564
1565 /*
1566 * avoid racing with extent status tree scans made by
1567 * ext4_insert_delayed_block()
1568 */
1569 down_write(&EXT4_I(inode)->i_data_sem);
1570 ext4_es_remove_extent(inode, start, last - start + 1);
1571 up_write(&EXT4_I(inode)->i_data_sem);
1572 }
1573
1574 folio_batch_init(&fbatch);
1575 while (index <= end) {
1576 nr = filemap_get_folios(mapping, &index, end, &fbatch);
1577 if (nr == 0)
1578 break;
1579 for (i = 0; i < nr; i++) {
1580 struct folio *folio = fbatch.folios[i];
1581
1582 if (folio->index < mpd->first_page)
1583 continue;
1584 if (folio_next_index(folio) - 1 > end)
1585 continue;
1586 BUG_ON(!folio_test_locked(folio));
1587 BUG_ON(folio_test_writeback(folio));
1588 if (invalidate) {
1589 if (folio_mapped(folio))
1590 folio_clear_dirty_for_io(folio);
1591 block_invalidate_folio(folio, 0,
1592 folio_size(folio));
1593 folio_clear_uptodate(folio);
1594 }
1595 folio_unlock(folio);
1596 }
1597 folio_batch_release(&fbatch);
1598 }
1599}
1600
1601static void ext4_print_free_blocks(struct inode *inode)
1602{
1603 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1604 struct super_block *sb = inode->i_sb;
1605 struct ext4_inode_info *ei = EXT4_I(inode);
1606
1607 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1608 EXT4_C2B(EXT4_SB(inode->i_sb),
1609 ext4_count_free_clusters(sb)));
1610 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1611 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1612 (long long) EXT4_C2B(EXT4_SB(sb),
1613 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1614 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1615 (long long) EXT4_C2B(EXT4_SB(sb),
1616 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1617 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1618 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1619 ei->i_reserved_data_blocks);
1620 return;
1621}
1622
1623/*
1624 * ext4_insert_delayed_block - adds a delayed block to the extents status
1625 * tree, incrementing the reserved cluster/block
1626 * count or making a pending reservation
1627 * where needed
1628 *
1629 * @inode - file containing the newly added block
1630 * @lblk - logical block to be added
1631 *
1632 * Returns 0 on success, negative error code on failure.
1633 */
1634static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1635{
1636 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1637 int ret;
1638 bool allocated = false;
1639
1640 /*
1641 * If the cluster containing lblk is shared with a delayed,
1642 * written, or unwritten extent in a bigalloc file system, it's
1643 * already been accounted for and does not need to be reserved.
1644 * A pending reservation must be made for the cluster if it's
1645 * shared with a written or unwritten extent and doesn't already
1646 * have one. Written and unwritten extents can be purged from the
1647 * extents status tree if the system is under memory pressure, so
1648 * it's necessary to examine the extent tree if a search of the
1649 * extents status tree doesn't get a match.
1650 */
1651 if (sbi->s_cluster_ratio == 1) {
1652 ret = ext4_da_reserve_space(inode);
1653 if (ret != 0) /* ENOSPC */
1654 return ret;
1655 } else { /* bigalloc */
1656 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1657 if (!ext4_es_scan_clu(inode,
1658 &ext4_es_is_mapped, lblk)) {
1659 ret = ext4_clu_mapped(inode,
1660 EXT4_B2C(sbi, lblk));
1661 if (ret < 0)
1662 return ret;
1663 if (ret == 0) {
1664 ret = ext4_da_reserve_space(inode);
1665 if (ret != 0) /* ENOSPC */
1666 return ret;
1667 } else {
1668 allocated = true;
1669 }
1670 } else {
1671 allocated = true;
1672 }
1673 }
1674 }
1675
1676 ext4_es_insert_delayed_block(inode, lblk, allocated);
1677 return 0;
1678}
1679
1680/*
1681 * This function is grabs code from the very beginning of
1682 * ext4_map_blocks, but assumes that the caller is from delayed write
1683 * time. This function looks up the requested blocks and sets the
1684 * buffer delay bit under the protection of i_data_sem.
1685 */
1686static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1687 struct ext4_map_blocks *map,
1688 struct buffer_head *bh)
1689{
1690 struct extent_status es;
1691 int retval;
1692 sector_t invalid_block = ~((sector_t) 0xffff);
1693#ifdef ES_AGGRESSIVE_TEST
1694 struct ext4_map_blocks orig_map;
1695
1696 memcpy(&orig_map, map, sizeof(*map));
1697#endif
1698
1699 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1700 invalid_block = ~0;
1701
1702 map->m_flags = 0;
1703 ext_debug(inode, "max_blocks %u, logical block %lu\n", map->m_len,
1704 (unsigned long) map->m_lblk);
1705
1706 /* Lookup extent status tree firstly */
1707 if (ext4_es_lookup_extent(inode, iblock, NULL, &es)) {
1708 if (ext4_es_is_hole(&es))
1709 goto add_delayed;
1710
1711 /*
1712 * Delayed extent could be allocated by fallocate.
1713 * So we need to check it.
1714 */
1715 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1716 map_bh(bh, inode->i_sb, invalid_block);
1717 set_buffer_new(bh);
1718 set_buffer_delay(bh);
1719 return 0;
1720 }
1721
1722 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1723 retval = es.es_len - (iblock - es.es_lblk);
1724 if (retval > map->m_len)
1725 retval = map->m_len;
1726 map->m_len = retval;
1727 if (ext4_es_is_written(&es))
1728 map->m_flags |= EXT4_MAP_MAPPED;
1729 else if (ext4_es_is_unwritten(&es))
1730 map->m_flags |= EXT4_MAP_UNWRITTEN;
1731 else
1732 BUG();
1733
1734#ifdef ES_AGGRESSIVE_TEST
1735 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1736#endif
1737 return retval;
1738 }
1739
1740 /*
1741 * Try to see if we can get the block without requesting a new
1742 * file system block.
1743 */
1744 down_read(&EXT4_I(inode)->i_data_sem);
1745 if (ext4_has_inline_data(inode))
1746 retval = 0;
1747 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1748 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1749 else
1750 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1751 if (retval < 0) {
1752 up_read(&EXT4_I(inode)->i_data_sem);
1753 return retval;
1754 }
1755 if (retval > 0) {
1756 unsigned int status;
1757
1758 if (unlikely(retval != map->m_len)) {
1759 ext4_warning(inode->i_sb,
1760 "ES len assertion failed for inode "
1761 "%lu: retval %d != map->m_len %d",
1762 inode->i_ino, retval, map->m_len);
1763 WARN_ON(1);
1764 }
1765
1766 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1767 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1768 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1769 map->m_pblk, status);
1770 up_read(&EXT4_I(inode)->i_data_sem);
1771 return retval;
1772 }
1773 up_read(&EXT4_I(inode)->i_data_sem);
1774
1775add_delayed:
1776 down_write(&EXT4_I(inode)->i_data_sem);
1777 retval = ext4_insert_delayed_block(inode, map->m_lblk);
1778 up_write(&EXT4_I(inode)->i_data_sem);
1779 if (retval)
1780 return retval;
1781
1782 map_bh(bh, inode->i_sb, invalid_block);
1783 set_buffer_new(bh);
1784 set_buffer_delay(bh);
1785 return retval;
1786}
1787
1788/*
1789 * This is a special get_block_t callback which is used by
1790 * ext4_da_write_begin(). It will either return mapped block or
1791 * reserve space for a single block.
1792 *
1793 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1794 * We also have b_blocknr = -1 and b_bdev initialized properly
1795 *
1796 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1797 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1798 * initialized properly.
1799 */
1800int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1801 struct buffer_head *bh, int create)
1802{
1803 struct ext4_map_blocks map;
1804 int ret = 0;
1805
1806 BUG_ON(create == 0);
1807 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1808
1809 map.m_lblk = iblock;
1810 map.m_len = 1;
1811
1812 /*
1813 * first, we need to know whether the block is allocated already
1814 * preallocated blocks are unmapped but should treated
1815 * the same as allocated blocks.
1816 */
1817 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1818 if (ret <= 0)
1819 return ret;
1820
1821 map_bh(bh, inode->i_sb, map.m_pblk);
1822 ext4_update_bh_state(bh, map.m_flags);
1823
1824 if (buffer_unwritten(bh)) {
1825 /* A delayed write to unwritten bh should be marked
1826 * new and mapped. Mapped ensures that we don't do
1827 * get_block multiple times when we write to the same
1828 * offset and new ensures that we do proper zero out
1829 * for partial write.
1830 */
1831 set_buffer_new(bh);
1832 set_buffer_mapped(bh);
1833 }
1834 return 0;
1835}
1836
1837static void mpage_folio_done(struct mpage_da_data *mpd, struct folio *folio)
1838{
1839 mpd->first_page += folio_nr_pages(folio);
1840 folio_unlock(folio);
1841}
1842
1843static int mpage_submit_folio(struct mpage_da_data *mpd, struct folio *folio)
1844{
1845 size_t len;
1846 loff_t size;
1847 int err;
1848
1849 BUG_ON(folio->index != mpd->first_page);
1850 folio_clear_dirty_for_io(folio);
1851 /*
1852 * We have to be very careful here! Nothing protects writeback path
1853 * against i_size changes and the page can be writeably mapped into
1854 * page tables. So an application can be growing i_size and writing
1855 * data through mmap while writeback runs. folio_clear_dirty_for_io()
1856 * write-protects our page in page tables and the page cannot get
1857 * written to again until we release folio lock. So only after
1858 * folio_clear_dirty_for_io() we are safe to sample i_size for
1859 * ext4_bio_write_folio() to zero-out tail of the written page. We rely
1860 * on the barrier provided by folio_test_clear_dirty() in
1861 * folio_clear_dirty_for_io() to make sure i_size is really sampled only
1862 * after page tables are updated.
1863 */
1864 size = i_size_read(mpd->inode);
1865 len = folio_size(folio);
1866 if (folio_pos(folio) + len > size &&
1867 !ext4_verity_in_progress(mpd->inode))
1868 len = size & ~PAGE_MASK;
1869 err = ext4_bio_write_folio(&mpd->io_submit, folio, len);
1870 if (!err)
1871 mpd->wbc->nr_to_write--;
1872
1873 return err;
1874}
1875
1876#define BH_FLAGS (BIT(BH_Unwritten) | BIT(BH_Delay))
1877
1878/*
1879 * mballoc gives us at most this number of blocks...
1880 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
1881 * The rest of mballoc seems to handle chunks up to full group size.
1882 */
1883#define MAX_WRITEPAGES_EXTENT_LEN 2048
1884
1885/*
1886 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
1887 *
1888 * @mpd - extent of blocks
1889 * @lblk - logical number of the block in the file
1890 * @bh - buffer head we want to add to the extent
1891 *
1892 * The function is used to collect contig. blocks in the same state. If the
1893 * buffer doesn't require mapping for writeback and we haven't started the
1894 * extent of buffers to map yet, the function returns 'true' immediately - the
1895 * caller can write the buffer right away. Otherwise the function returns true
1896 * if the block has been added to the extent, false if the block couldn't be
1897 * added.
1898 */
1899static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
1900 struct buffer_head *bh)
1901{
1902 struct ext4_map_blocks *map = &mpd->map;
1903
1904 /* Buffer that doesn't need mapping for writeback? */
1905 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
1906 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
1907 /* So far no extent to map => we write the buffer right away */
1908 if (map->m_len == 0)
1909 return true;
1910 return false;
1911 }
1912
1913 /* First block in the extent? */
1914 if (map->m_len == 0) {
1915 /* We cannot map unless handle is started... */
1916 if (!mpd->do_map)
1917 return false;
1918 map->m_lblk = lblk;
1919 map->m_len = 1;
1920 map->m_flags = bh->b_state & BH_FLAGS;
1921 return true;
1922 }
1923
1924 /* Don't go larger than mballoc is willing to allocate */
1925 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
1926 return false;
1927
1928 /* Can we merge the block to our big extent? */
1929 if (lblk == map->m_lblk + map->m_len &&
1930 (bh->b_state & BH_FLAGS) == map->m_flags) {
1931 map->m_len++;
1932 return true;
1933 }
1934 return false;
1935}
1936
1937/*
1938 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
1939 *
1940 * @mpd - extent of blocks for mapping
1941 * @head - the first buffer in the page
1942 * @bh - buffer we should start processing from
1943 * @lblk - logical number of the block in the file corresponding to @bh
1944 *
1945 * Walk through page buffers from @bh upto @head (exclusive) and either submit
1946 * the page for IO if all buffers in this page were mapped and there's no
1947 * accumulated extent of buffers to map or add buffers in the page to the
1948 * extent of buffers to map. The function returns 1 if the caller can continue
1949 * by processing the next page, 0 if it should stop adding buffers to the
1950 * extent to map because we cannot extend it anymore. It can also return value
1951 * < 0 in case of error during IO submission.
1952 */
1953static int mpage_process_page_bufs(struct mpage_da_data *mpd,
1954 struct buffer_head *head,
1955 struct buffer_head *bh,
1956 ext4_lblk_t lblk)
1957{
1958 struct inode *inode = mpd->inode;
1959 int err;
1960 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
1961 >> inode->i_blkbits;
1962
1963 if (ext4_verity_in_progress(inode))
1964 blocks = EXT_MAX_BLOCKS;
1965
1966 do {
1967 BUG_ON(buffer_locked(bh));
1968
1969 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
1970 /* Found extent to map? */
1971 if (mpd->map.m_len)
1972 return 0;
1973 /* Buffer needs mapping and handle is not started? */
1974 if (!mpd->do_map)
1975 return 0;
1976 /* Everything mapped so far and we hit EOF */
1977 break;
1978 }
1979 } while (lblk++, (bh = bh->b_this_page) != head);
1980 /* So far everything mapped? Submit the page for IO. */
1981 if (mpd->map.m_len == 0) {
1982 err = mpage_submit_folio(mpd, head->b_folio);
1983 if (err < 0)
1984 return err;
1985 mpage_folio_done(mpd, head->b_folio);
1986 }
1987 if (lblk >= blocks) {
1988 mpd->scanned_until_end = 1;
1989 return 0;
1990 }
1991 return 1;
1992}
1993
1994/*
1995 * mpage_process_folio - update folio buffers corresponding to changed extent
1996 * and may submit fully mapped page for IO
1997 * @mpd: description of extent to map, on return next extent to map
1998 * @folio: Contains these buffers.
1999 * @m_lblk: logical block mapping.
2000 * @m_pblk: corresponding physical mapping.
2001 * @map_bh: determines on return whether this page requires any further
2002 * mapping or not.
2003 *
2004 * Scan given folio buffers corresponding to changed extent and update buffer
2005 * state according to new extent state.
2006 * We map delalloc buffers to their physical location, clear unwritten bits.
2007 * If the given folio is not fully mapped, we update @mpd to the next extent in
2008 * the given folio that needs mapping & return @map_bh as true.
2009 */
2010static int mpage_process_folio(struct mpage_da_data *mpd, struct folio *folio,
2011 ext4_lblk_t *m_lblk, ext4_fsblk_t *m_pblk,
2012 bool *map_bh)
2013{
2014 struct buffer_head *head, *bh;
2015 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2016 ext4_lblk_t lblk = *m_lblk;
2017 ext4_fsblk_t pblock = *m_pblk;
2018 int err = 0;
2019 int blkbits = mpd->inode->i_blkbits;
2020 ssize_t io_end_size = 0;
2021 struct ext4_io_end_vec *io_end_vec = ext4_last_io_end_vec(io_end);
2022
2023 bh = head = folio_buffers(folio);
2024 do {
2025 if (lblk < mpd->map.m_lblk)
2026 continue;
2027 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2028 /*
2029 * Buffer after end of mapped extent.
2030 * Find next buffer in the folio to map.
2031 */
2032 mpd->map.m_len = 0;
2033 mpd->map.m_flags = 0;
2034 io_end_vec->size += io_end_size;
2035
2036 err = mpage_process_page_bufs(mpd, head, bh, lblk);
2037 if (err > 0)
2038 err = 0;
2039 if (!err && mpd->map.m_len && mpd->map.m_lblk > lblk) {
2040 io_end_vec = ext4_alloc_io_end_vec(io_end);
2041 if (IS_ERR(io_end_vec)) {
2042 err = PTR_ERR(io_end_vec);
2043 goto out;
2044 }
2045 io_end_vec->offset = (loff_t)mpd->map.m_lblk << blkbits;
2046 }
2047 *map_bh = true;
2048 goto out;
2049 }
2050 if (buffer_delay(bh)) {
2051 clear_buffer_delay(bh);
2052 bh->b_blocknr = pblock++;
2053 }
2054 clear_buffer_unwritten(bh);
2055 io_end_size += (1 << blkbits);
2056 } while (lblk++, (bh = bh->b_this_page) != head);
2057
2058 io_end_vec->size += io_end_size;
2059 *map_bh = false;
2060out:
2061 *m_lblk = lblk;
2062 *m_pblk = pblock;
2063 return err;
2064}
2065
2066/*
2067 * mpage_map_buffers - update buffers corresponding to changed extent and
2068 * submit fully mapped pages for IO
2069 *
2070 * @mpd - description of extent to map, on return next extent to map
2071 *
2072 * Scan buffers corresponding to changed extent (we expect corresponding pages
2073 * to be already locked) and update buffer state according to new extent state.
2074 * We map delalloc buffers to their physical location, clear unwritten bits,
2075 * and mark buffers as uninit when we perform writes to unwritten extents
2076 * and do extent conversion after IO is finished. If the last page is not fully
2077 * mapped, we update @map to the next extent in the last page that needs
2078 * mapping. Otherwise we submit the page for IO.
2079 */
2080static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2081{
2082 struct folio_batch fbatch;
2083 unsigned nr, i;
2084 struct inode *inode = mpd->inode;
2085 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2086 pgoff_t start, end;
2087 ext4_lblk_t lblk;
2088 ext4_fsblk_t pblock;
2089 int err;
2090 bool map_bh = false;
2091
2092 start = mpd->map.m_lblk >> bpp_bits;
2093 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2094 lblk = start << bpp_bits;
2095 pblock = mpd->map.m_pblk;
2096
2097 folio_batch_init(&fbatch);
2098 while (start <= end) {
2099 nr = filemap_get_folios(inode->i_mapping, &start, end, &fbatch);
2100 if (nr == 0)
2101 break;
2102 for (i = 0; i < nr; i++) {
2103 struct folio *folio = fbatch.folios[i];
2104
2105 err = mpage_process_folio(mpd, folio, &lblk, &pblock,
2106 &map_bh);
2107 /*
2108 * If map_bh is true, means page may require further bh
2109 * mapping, or maybe the page was submitted for IO.
2110 * So we return to call further extent mapping.
2111 */
2112 if (err < 0 || map_bh)
2113 goto out;
2114 /* Page fully mapped - let IO run! */
2115 err = mpage_submit_folio(mpd, folio);
2116 if (err < 0)
2117 goto out;
2118 mpage_folio_done(mpd, folio);
2119 }
2120 folio_batch_release(&fbatch);
2121 }
2122 /* Extent fully mapped and matches with page boundary. We are done. */
2123 mpd->map.m_len = 0;
2124 mpd->map.m_flags = 0;
2125 return 0;
2126out:
2127 folio_batch_release(&fbatch);
2128 return err;
2129}
2130
2131static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2132{
2133 struct inode *inode = mpd->inode;
2134 struct ext4_map_blocks *map = &mpd->map;
2135 int get_blocks_flags;
2136 int err, dioread_nolock;
2137
2138 trace_ext4_da_write_pages_extent(inode, map);
2139 /*
2140 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2141 * to convert an unwritten extent to be initialized (in the case
2142 * where we have written into one or more preallocated blocks). It is
2143 * possible that we're going to need more metadata blocks than
2144 * previously reserved. However we must not fail because we're in
2145 * writeback and there is nothing we can do about it so it might result
2146 * in data loss. So use reserved blocks to allocate metadata if
2147 * possible.
2148 *
2149 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2150 * the blocks in question are delalloc blocks. This indicates
2151 * that the blocks and quotas has already been checked when
2152 * the data was copied into the page cache.
2153 */
2154 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2155 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2156 EXT4_GET_BLOCKS_IO_SUBMIT;
2157 dioread_nolock = ext4_should_dioread_nolock(inode);
2158 if (dioread_nolock)
2159 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2160 if (map->m_flags & BIT(BH_Delay))
2161 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2162
2163 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2164 if (err < 0)
2165 return err;
2166 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2167 if (!mpd->io_submit.io_end->handle &&
2168 ext4_handle_valid(handle)) {
2169 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2170 handle->h_rsv_handle = NULL;
2171 }
2172 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2173 }
2174
2175 BUG_ON(map->m_len == 0);
2176 return 0;
2177}
2178
2179/*
2180 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2181 * mpd->len and submit pages underlying it for IO
2182 *
2183 * @handle - handle for journal operations
2184 * @mpd - extent to map
2185 * @give_up_on_write - we set this to true iff there is a fatal error and there
2186 * is no hope of writing the data. The caller should discard
2187 * dirty pages to avoid infinite loops.
2188 *
2189 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2190 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2191 * them to initialized or split the described range from larger unwritten
2192 * extent. Note that we need not map all the described range since allocation
2193 * can return less blocks or the range is covered by more unwritten extents. We
2194 * cannot map more because we are limited by reserved transaction credits. On
2195 * the other hand we always make sure that the last touched page is fully
2196 * mapped so that it can be written out (and thus forward progress is
2197 * guaranteed). After mapping we submit all mapped pages for IO.
2198 */
2199static int mpage_map_and_submit_extent(handle_t *handle,
2200 struct mpage_da_data *mpd,
2201 bool *give_up_on_write)
2202{
2203 struct inode *inode = mpd->inode;
2204 struct ext4_map_blocks *map = &mpd->map;
2205 int err;
2206 loff_t disksize;
2207 int progress = 0;
2208 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2209 struct ext4_io_end_vec *io_end_vec;
2210
2211 io_end_vec = ext4_alloc_io_end_vec(io_end);
2212 if (IS_ERR(io_end_vec))
2213 return PTR_ERR(io_end_vec);
2214 io_end_vec->offset = ((loff_t)map->m_lblk) << inode->i_blkbits;
2215 do {
2216 err = mpage_map_one_extent(handle, mpd);
2217 if (err < 0) {
2218 struct super_block *sb = inode->i_sb;
2219
2220 if (ext4_forced_shutdown(sb))
2221 goto invalidate_dirty_pages;
2222 /*
2223 * Let the uper layers retry transient errors.
2224 * In the case of ENOSPC, if ext4_count_free_blocks()
2225 * is non-zero, a commit should free up blocks.
2226 */
2227 if ((err == -ENOMEM) ||
2228 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2229 if (progress)
2230 goto update_disksize;
2231 return err;
2232 }
2233 ext4_msg(sb, KERN_CRIT,
2234 "Delayed block allocation failed for "
2235 "inode %lu at logical offset %llu with"
2236 " max blocks %u with error %d",
2237 inode->i_ino,
2238 (unsigned long long)map->m_lblk,
2239 (unsigned)map->m_len, -err);
2240 ext4_msg(sb, KERN_CRIT,
2241 "This should not happen!! Data will "
2242 "be lost\n");
2243 if (err == -ENOSPC)
2244 ext4_print_free_blocks(inode);
2245 invalidate_dirty_pages:
2246 *give_up_on_write = true;
2247 return err;
2248 }
2249 progress = 1;
2250 /*
2251 * Update buffer state, submit mapped pages, and get us new
2252 * extent to map
2253 */
2254 err = mpage_map_and_submit_buffers(mpd);
2255 if (err < 0)
2256 goto update_disksize;
2257 } while (map->m_len);
2258
2259update_disksize:
2260 /*
2261 * Update on-disk size after IO is submitted. Races with
2262 * truncate are avoided by checking i_size under i_data_sem.
2263 */
2264 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2265 if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
2266 int err2;
2267 loff_t i_size;
2268
2269 down_write(&EXT4_I(inode)->i_data_sem);
2270 i_size = i_size_read(inode);
2271 if (disksize > i_size)
2272 disksize = i_size;
2273 if (disksize > EXT4_I(inode)->i_disksize)
2274 EXT4_I(inode)->i_disksize = disksize;
2275 up_write(&EXT4_I(inode)->i_data_sem);
2276 err2 = ext4_mark_inode_dirty(handle, inode);
2277 if (err2) {
2278 ext4_error_err(inode->i_sb, -err2,
2279 "Failed to mark inode %lu dirty",
2280 inode->i_ino);
2281 }
2282 if (!err)
2283 err = err2;
2284 }
2285 return err;
2286}
2287
2288/*
2289 * Calculate the total number of credits to reserve for one writepages
2290 * iteration. This is called from ext4_writepages(). We map an extent of
2291 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2292 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2293 * bpp - 1 blocks in bpp different extents.
2294 */
2295static int ext4_da_writepages_trans_blocks(struct inode *inode)
2296{
2297 int bpp = ext4_journal_blocks_per_page(inode);
2298
2299 return ext4_meta_trans_blocks(inode,
2300 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2301}
2302
2303static int ext4_journal_folio_buffers(handle_t *handle, struct folio *folio,
2304 size_t len)
2305{
2306 struct buffer_head *page_bufs = folio_buffers(folio);
2307 struct inode *inode = folio->mapping->host;
2308 int ret, err;
2309
2310 ret = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2311 NULL, do_journal_get_write_access);
2312 err = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2313 NULL, write_end_fn);
2314 if (ret == 0)
2315 ret = err;
2316 err = ext4_jbd2_inode_add_write(handle, inode, folio_pos(folio), len);
2317 if (ret == 0)
2318 ret = err;
2319 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2320
2321 return ret;
2322}
2323
2324static int mpage_journal_page_buffers(handle_t *handle,
2325 struct mpage_da_data *mpd,
2326 struct folio *folio)
2327{
2328 struct inode *inode = mpd->inode;
2329 loff_t size = i_size_read(inode);
2330 size_t len = folio_size(folio);
2331
2332 folio_clear_checked(folio);
2333 mpd->wbc->nr_to_write--;
2334
2335 if (folio_pos(folio) + len > size &&
2336 !ext4_verity_in_progress(inode))
2337 len = size - folio_pos(folio);
2338
2339 return ext4_journal_folio_buffers(handle, folio, len);
2340}
2341
2342/*
2343 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2344 * needing mapping, submit mapped pages
2345 *
2346 * @mpd - where to look for pages
2347 *
2348 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2349 * IO immediately. If we cannot map blocks, we submit just already mapped
2350 * buffers in the page for IO and keep page dirty. When we can map blocks and
2351 * we find a page which isn't mapped we start accumulating extent of buffers
2352 * underlying these pages that needs mapping (formed by either delayed or
2353 * unwritten buffers). We also lock the pages containing these buffers. The
2354 * extent found is returned in @mpd structure (starting at mpd->lblk with
2355 * length mpd->len blocks).
2356 *
2357 * Note that this function can attach bios to one io_end structure which are
2358 * neither logically nor physically contiguous. Although it may seem as an
2359 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2360 * case as we need to track IO to all buffers underlying a page in one io_end.
2361 */
2362static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2363{
2364 struct address_space *mapping = mpd->inode->i_mapping;
2365 struct folio_batch fbatch;
2366 unsigned int nr_folios;
2367 pgoff_t index = mpd->first_page;
2368 pgoff_t end = mpd->last_page;
2369 xa_mark_t tag;
2370 int i, err = 0;
2371 int blkbits = mpd->inode->i_blkbits;
2372 ext4_lblk_t lblk;
2373 struct buffer_head *head;
2374 handle_t *handle = NULL;
2375 int bpp = ext4_journal_blocks_per_page(mpd->inode);
2376
2377 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2378 tag = PAGECACHE_TAG_TOWRITE;
2379 else
2380 tag = PAGECACHE_TAG_DIRTY;
2381
2382 mpd->map.m_len = 0;
2383 mpd->next_page = index;
2384 if (ext4_should_journal_data(mpd->inode)) {
2385 handle = ext4_journal_start(mpd->inode, EXT4_HT_WRITE_PAGE,
2386 bpp);
2387 if (IS_ERR(handle))
2388 return PTR_ERR(handle);
2389 }
2390 folio_batch_init(&fbatch);
2391 while (index <= end) {
2392 nr_folios = filemap_get_folios_tag(mapping, &index, end,
2393 tag, &fbatch);
2394 if (nr_folios == 0)
2395 break;
2396
2397 for (i = 0; i < nr_folios; i++) {
2398 struct folio *folio = fbatch.folios[i];
2399
2400 /*
2401 * Accumulated enough dirty pages? This doesn't apply
2402 * to WB_SYNC_ALL mode. For integrity sync we have to
2403 * keep going because someone may be concurrently
2404 * dirtying pages, and we might have synced a lot of
2405 * newly appeared dirty pages, but have not synced all
2406 * of the old dirty pages.
2407 */
2408 if (mpd->wbc->sync_mode == WB_SYNC_NONE &&
2409 mpd->wbc->nr_to_write <=
2410 mpd->map.m_len >> (PAGE_SHIFT - blkbits))
2411 goto out;
2412
2413 /* If we can't merge this page, we are done. */
2414 if (mpd->map.m_len > 0 && mpd->next_page != folio->index)
2415 goto out;
2416
2417 if (handle) {
2418 err = ext4_journal_ensure_credits(handle, bpp,
2419 0);
2420 if (err < 0)
2421 goto out;
2422 }
2423
2424 folio_lock(folio);
2425 /*
2426 * If the page is no longer dirty, or its mapping no
2427 * longer corresponds to inode we are writing (which
2428 * means it has been truncated or invalidated), or the
2429 * page is already under writeback and we are not doing
2430 * a data integrity writeback, skip the page
2431 */
2432 if (!folio_test_dirty(folio) ||
2433 (folio_test_writeback(folio) &&
2434 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2435 unlikely(folio->mapping != mapping)) {
2436 folio_unlock(folio);
2437 continue;
2438 }
2439
2440 folio_wait_writeback(folio);
2441 BUG_ON(folio_test_writeback(folio));
2442
2443 /*
2444 * Should never happen but for buggy code in
2445 * other subsystems that call
2446 * set_page_dirty() without properly warning
2447 * the file system first. See [1] for more
2448 * information.
2449 *
2450 * [1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@quack2.suse.cz
2451 */
2452 if (!folio_buffers(folio)) {
2453 ext4_warning_inode(mpd->inode, "page %lu does not have buffers attached", folio->index);
2454 folio_clear_dirty(folio);
2455 folio_unlock(folio);
2456 continue;
2457 }
2458
2459 if (mpd->map.m_len == 0)
2460 mpd->first_page = folio->index;
2461 mpd->next_page = folio_next_index(folio);
2462 /*
2463 * Writeout when we cannot modify metadata is simple.
2464 * Just submit the page. For data=journal mode we
2465 * first handle writeout of the page for checkpoint and
2466 * only after that handle delayed page dirtying. This
2467 * makes sure current data is checkpointed to the final
2468 * location before possibly journalling it again which
2469 * is desirable when the page is frequently dirtied
2470 * through a pin.
2471 */
2472 if (!mpd->can_map) {
2473 err = mpage_submit_folio(mpd, folio);
2474 if (err < 0)
2475 goto out;
2476 /* Pending dirtying of journalled data? */
2477 if (folio_test_checked(folio)) {
2478 err = mpage_journal_page_buffers(handle,
2479 mpd, folio);
2480 if (err < 0)
2481 goto out;
2482 mpd->journalled_more_data = 1;
2483 }
2484 mpage_folio_done(mpd, folio);
2485 } else {
2486 /* Add all dirty buffers to mpd */
2487 lblk = ((ext4_lblk_t)folio->index) <<
2488 (PAGE_SHIFT - blkbits);
2489 head = folio_buffers(folio);
2490 err = mpage_process_page_bufs(mpd, head, head,
2491 lblk);
2492 if (err <= 0)
2493 goto out;
2494 err = 0;
2495 }
2496 }
2497 folio_batch_release(&fbatch);
2498 cond_resched();
2499 }
2500 mpd->scanned_until_end = 1;
2501 if (handle)
2502 ext4_journal_stop(handle);
2503 return 0;
2504out:
2505 folio_batch_release(&fbatch);
2506 if (handle)
2507 ext4_journal_stop(handle);
2508 return err;
2509}
2510
2511static int ext4_do_writepages(struct mpage_da_data *mpd)
2512{
2513 struct writeback_control *wbc = mpd->wbc;
2514 pgoff_t writeback_index = 0;
2515 long nr_to_write = wbc->nr_to_write;
2516 int range_whole = 0;
2517 int cycled = 1;
2518 handle_t *handle = NULL;
2519 struct inode *inode = mpd->inode;
2520 struct address_space *mapping = inode->i_mapping;
2521 int needed_blocks, rsv_blocks = 0, ret = 0;
2522 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2523 struct blk_plug plug;
2524 bool give_up_on_write = false;
2525
2526 trace_ext4_writepages(inode, wbc);
2527
2528 /*
2529 * No pages to write? This is mainly a kludge to avoid starting
2530 * a transaction for special inodes like journal inode on last iput()
2531 * because that could violate lock ordering on umount
2532 */
2533 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2534 goto out_writepages;
2535
2536 /*
2537 * If the filesystem has aborted, it is read-only, so return
2538 * right away instead of dumping stack traces later on that
2539 * will obscure the real source of the problem. We test
2540 * fs shutdown state instead of sb->s_flag's SB_RDONLY because
2541 * the latter could be true if the filesystem is mounted
2542 * read-only, and in that case, ext4_writepages should
2543 * *never* be called, so if that ever happens, we would want
2544 * the stack trace.
2545 */
2546 if (unlikely(ext4_forced_shutdown(mapping->host->i_sb))) {
2547 ret = -EROFS;
2548 goto out_writepages;
2549 }
2550
2551 /*
2552 * If we have inline data and arrive here, it means that
2553 * we will soon create the block for the 1st page, so
2554 * we'd better clear the inline data here.
2555 */
2556 if (ext4_has_inline_data(inode)) {
2557 /* Just inode will be modified... */
2558 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2559 if (IS_ERR(handle)) {
2560 ret = PTR_ERR(handle);
2561 goto out_writepages;
2562 }
2563 BUG_ON(ext4_test_inode_state(inode,
2564 EXT4_STATE_MAY_INLINE_DATA));
2565 ext4_destroy_inline_data(handle, inode);
2566 ext4_journal_stop(handle);
2567 }
2568
2569 /*
2570 * data=journal mode does not do delalloc so we just need to writeout /
2571 * journal already mapped buffers. On the other hand we need to commit
2572 * transaction to make data stable. We expect all the data to be
2573 * already in the journal (the only exception are DMA pinned pages
2574 * dirtied behind our back) so we commit transaction here and run the
2575 * writeback loop to checkpoint them. The checkpointing is not actually
2576 * necessary to make data persistent *but* quite a few places (extent
2577 * shifting operations, fsverity, ...) depend on being able to drop
2578 * pagecache pages after calling filemap_write_and_wait() and for that
2579 * checkpointing needs to happen.
2580 */
2581 if (ext4_should_journal_data(inode)) {
2582 mpd->can_map = 0;
2583 if (wbc->sync_mode == WB_SYNC_ALL)
2584 ext4_fc_commit(sbi->s_journal,
2585 EXT4_I(inode)->i_datasync_tid);
2586 }
2587 mpd->journalled_more_data = 0;
2588
2589 if (ext4_should_dioread_nolock(inode)) {
2590 /*
2591 * We may need to convert up to one extent per block in
2592 * the page and we may dirty the inode.
2593 */
2594 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2595 PAGE_SIZE >> inode->i_blkbits);
2596 }
2597
2598 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2599 range_whole = 1;
2600
2601 if (wbc->range_cyclic) {
2602 writeback_index = mapping->writeback_index;
2603 if (writeback_index)
2604 cycled = 0;
2605 mpd->first_page = writeback_index;
2606 mpd->last_page = -1;
2607 } else {
2608 mpd->first_page = wbc->range_start >> PAGE_SHIFT;
2609 mpd->last_page = wbc->range_end >> PAGE_SHIFT;
2610 }
2611
2612 ext4_io_submit_init(&mpd->io_submit, wbc);
2613retry:
2614 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2615 tag_pages_for_writeback(mapping, mpd->first_page,
2616 mpd->last_page);
2617 blk_start_plug(&plug);
2618
2619 /*
2620 * First writeback pages that don't need mapping - we can avoid
2621 * starting a transaction unnecessarily and also avoid being blocked
2622 * in the block layer on device congestion while having transaction
2623 * started.
2624 */
2625 mpd->do_map = 0;
2626 mpd->scanned_until_end = 0;
2627 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2628 if (!mpd->io_submit.io_end) {
2629 ret = -ENOMEM;
2630 goto unplug;
2631 }
2632 ret = mpage_prepare_extent_to_map(mpd);
2633 /* Unlock pages we didn't use */
2634 mpage_release_unused_pages(mpd, false);
2635 /* Submit prepared bio */
2636 ext4_io_submit(&mpd->io_submit);
2637 ext4_put_io_end_defer(mpd->io_submit.io_end);
2638 mpd->io_submit.io_end = NULL;
2639 if (ret < 0)
2640 goto unplug;
2641
2642 while (!mpd->scanned_until_end && wbc->nr_to_write > 0) {
2643 /* For each extent of pages we use new io_end */
2644 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2645 if (!mpd->io_submit.io_end) {
2646 ret = -ENOMEM;
2647 break;
2648 }
2649
2650 WARN_ON_ONCE(!mpd->can_map);
2651 /*
2652 * We have two constraints: We find one extent to map and we
2653 * must always write out whole page (makes a difference when
2654 * blocksize < pagesize) so that we don't block on IO when we
2655 * try to write out the rest of the page. Journalled mode is
2656 * not supported by delalloc.
2657 */
2658 BUG_ON(ext4_should_journal_data(inode));
2659 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2660
2661 /* start a new transaction */
2662 handle = ext4_journal_start_with_reserve(inode,
2663 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2664 if (IS_ERR(handle)) {
2665 ret = PTR_ERR(handle);
2666 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2667 "%ld pages, ino %lu; err %d", __func__,
2668 wbc->nr_to_write, inode->i_ino, ret);
2669 /* Release allocated io_end */
2670 ext4_put_io_end(mpd->io_submit.io_end);
2671 mpd->io_submit.io_end = NULL;
2672 break;
2673 }
2674 mpd->do_map = 1;
2675
2676 trace_ext4_da_write_pages(inode, mpd->first_page, wbc);
2677 ret = mpage_prepare_extent_to_map(mpd);
2678 if (!ret && mpd->map.m_len)
2679 ret = mpage_map_and_submit_extent(handle, mpd,
2680 &give_up_on_write);
2681 /*
2682 * Caution: If the handle is synchronous,
2683 * ext4_journal_stop() can wait for transaction commit
2684 * to finish which may depend on writeback of pages to
2685 * complete or on page lock to be released. In that
2686 * case, we have to wait until after we have
2687 * submitted all the IO, released page locks we hold,
2688 * and dropped io_end reference (for extent conversion
2689 * to be able to complete) before stopping the handle.
2690 */
2691 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2692 ext4_journal_stop(handle);
2693 handle = NULL;
2694 mpd->do_map = 0;
2695 }
2696 /* Unlock pages we didn't use */
2697 mpage_release_unused_pages(mpd, give_up_on_write);
2698 /* Submit prepared bio */
2699 ext4_io_submit(&mpd->io_submit);
2700
2701 /*
2702 * Drop our io_end reference we got from init. We have
2703 * to be careful and use deferred io_end finishing if
2704 * we are still holding the transaction as we can
2705 * release the last reference to io_end which may end
2706 * up doing unwritten extent conversion.
2707 */
2708 if (handle) {
2709 ext4_put_io_end_defer(mpd->io_submit.io_end);
2710 ext4_journal_stop(handle);
2711 } else
2712 ext4_put_io_end(mpd->io_submit.io_end);
2713 mpd->io_submit.io_end = NULL;
2714
2715 if (ret == -ENOSPC && sbi->s_journal) {
2716 /*
2717 * Commit the transaction which would
2718 * free blocks released in the transaction
2719 * and try again
2720 */
2721 jbd2_journal_force_commit_nested(sbi->s_journal);
2722 ret = 0;
2723 continue;
2724 }
2725 /* Fatal error - ENOMEM, EIO... */
2726 if (ret)
2727 break;
2728 }
2729unplug:
2730 blk_finish_plug(&plug);
2731 if (!ret && !cycled && wbc->nr_to_write > 0) {
2732 cycled = 1;
2733 mpd->last_page = writeback_index - 1;
2734 mpd->first_page = 0;
2735 goto retry;
2736 }
2737
2738 /* Update index */
2739 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2740 /*
2741 * Set the writeback_index so that range_cyclic
2742 * mode will write it back later
2743 */
2744 mapping->writeback_index = mpd->first_page;
2745
2746out_writepages:
2747 trace_ext4_writepages_result(inode, wbc, ret,
2748 nr_to_write - wbc->nr_to_write);
2749 return ret;
2750}
2751
2752static int ext4_writepages(struct address_space *mapping,
2753 struct writeback_control *wbc)
2754{
2755 struct super_block *sb = mapping->host->i_sb;
2756 struct mpage_da_data mpd = {
2757 .inode = mapping->host,
2758 .wbc = wbc,
2759 .can_map = 1,
2760 };
2761 int ret;
2762 int alloc_ctx;
2763
2764 if (unlikely(ext4_forced_shutdown(sb)))
2765 return -EIO;
2766
2767 alloc_ctx = ext4_writepages_down_read(sb);
2768 ret = ext4_do_writepages(&mpd);
2769 /*
2770 * For data=journal writeback we could have come across pages marked
2771 * for delayed dirtying (PageChecked) which were just added to the
2772 * running transaction. Try once more to get them to stable storage.
2773 */
2774 if (!ret && mpd.journalled_more_data)
2775 ret = ext4_do_writepages(&mpd);
2776 ext4_writepages_up_read(sb, alloc_ctx);
2777
2778 return ret;
2779}
2780
2781int ext4_normal_submit_inode_data_buffers(struct jbd2_inode *jinode)
2782{
2783 struct writeback_control wbc = {
2784 .sync_mode = WB_SYNC_ALL,
2785 .nr_to_write = LONG_MAX,
2786 .range_start = jinode->i_dirty_start,
2787 .range_end = jinode->i_dirty_end,
2788 };
2789 struct mpage_da_data mpd = {
2790 .inode = jinode->i_vfs_inode,
2791 .wbc = &wbc,
2792 .can_map = 0,
2793 };
2794 return ext4_do_writepages(&mpd);
2795}
2796
2797static int ext4_dax_writepages(struct address_space *mapping,
2798 struct writeback_control *wbc)
2799{
2800 int ret;
2801 long nr_to_write = wbc->nr_to_write;
2802 struct inode *inode = mapping->host;
2803 int alloc_ctx;
2804
2805 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2806 return -EIO;
2807
2808 alloc_ctx = ext4_writepages_down_read(inode->i_sb);
2809 trace_ext4_writepages(inode, wbc);
2810
2811 ret = dax_writeback_mapping_range(mapping,
2812 EXT4_SB(inode->i_sb)->s_daxdev, wbc);
2813 trace_ext4_writepages_result(inode, wbc, ret,
2814 nr_to_write - wbc->nr_to_write);
2815 ext4_writepages_up_read(inode->i_sb, alloc_ctx);
2816 return ret;
2817}
2818
2819static int ext4_nonda_switch(struct super_block *sb)
2820{
2821 s64 free_clusters, dirty_clusters;
2822 struct ext4_sb_info *sbi = EXT4_SB(sb);
2823
2824 /*
2825 * switch to non delalloc mode if we are running low
2826 * on free block. The free block accounting via percpu
2827 * counters can get slightly wrong with percpu_counter_batch getting
2828 * accumulated on each CPU without updating global counters
2829 * Delalloc need an accurate free block accounting. So switch
2830 * to non delalloc when we are near to error range.
2831 */
2832 free_clusters =
2833 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2834 dirty_clusters =
2835 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2836 /*
2837 * Start pushing delalloc when 1/2 of free blocks are dirty.
2838 */
2839 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2840 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2841
2842 if (2 * free_clusters < 3 * dirty_clusters ||
2843 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2844 /*
2845 * free block count is less than 150% of dirty blocks
2846 * or free blocks is less than watermark
2847 */
2848 return 1;
2849 }
2850 return 0;
2851}
2852
2853static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2854 loff_t pos, unsigned len,
2855 struct page **pagep, void **fsdata)
2856{
2857 int ret, retries = 0;
2858 struct folio *folio;
2859 pgoff_t index;
2860 struct inode *inode = mapping->host;
2861
2862 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2863 return -EIO;
2864
2865 index = pos >> PAGE_SHIFT;
2866
2867 if (ext4_nonda_switch(inode->i_sb) || ext4_verity_in_progress(inode)) {
2868 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2869 return ext4_write_begin(file, mapping, pos,
2870 len, pagep, fsdata);
2871 }
2872 *fsdata = (void *)0;
2873 trace_ext4_da_write_begin(inode, pos, len);
2874
2875 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2876 ret = ext4_da_write_inline_data_begin(mapping, inode, pos, len,
2877 pagep, fsdata);
2878 if (ret < 0)
2879 return ret;
2880 if (ret == 1)
2881 return 0;
2882 }
2883
2884retry:
2885 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
2886 mapping_gfp_mask(mapping));
2887 if (IS_ERR(folio))
2888 return PTR_ERR(folio);
2889
2890#ifdef CONFIG_FS_ENCRYPTION
2891 ret = ext4_block_write_begin(folio, pos, len, ext4_da_get_block_prep);
2892#else
2893 ret = __block_write_begin(&folio->page, pos, len, ext4_da_get_block_prep);
2894#endif
2895 if (ret < 0) {
2896 folio_unlock(folio);
2897 folio_put(folio);
2898 /*
2899 * block_write_begin may have instantiated a few blocks
2900 * outside i_size. Trim these off again. Don't need
2901 * i_size_read because we hold inode lock.
2902 */
2903 if (pos + len > inode->i_size)
2904 ext4_truncate_failed_write(inode);
2905
2906 if (ret == -ENOSPC &&
2907 ext4_should_retry_alloc(inode->i_sb, &retries))
2908 goto retry;
2909 return ret;
2910 }
2911
2912 *pagep = &folio->page;
2913 return ret;
2914}
2915
2916/*
2917 * Check if we should update i_disksize
2918 * when write to the end of file but not require block allocation
2919 */
2920static int ext4_da_should_update_i_disksize(struct folio *folio,
2921 unsigned long offset)
2922{
2923 struct buffer_head *bh;
2924 struct inode *inode = folio->mapping->host;
2925 unsigned int idx;
2926 int i;
2927
2928 bh = folio_buffers(folio);
2929 idx = offset >> inode->i_blkbits;
2930
2931 for (i = 0; i < idx; i++)
2932 bh = bh->b_this_page;
2933
2934 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2935 return 0;
2936 return 1;
2937}
2938
2939static int ext4_da_do_write_end(struct address_space *mapping,
2940 loff_t pos, unsigned len, unsigned copied,
2941 struct folio *folio)
2942{
2943 struct inode *inode = mapping->host;
2944 loff_t old_size = inode->i_size;
2945 bool disksize_changed = false;
2946 loff_t new_i_size;
2947
2948 /*
2949 * block_write_end() will mark the inode as dirty with I_DIRTY_PAGES
2950 * flag, which all that's needed to trigger page writeback.
2951 */
2952 copied = block_write_end(NULL, mapping, pos, len, copied,
2953 &folio->page, NULL);
2954 new_i_size = pos + copied;
2955
2956 /*
2957 * It's important to update i_size while still holding folio lock,
2958 * because folio writeout could otherwise come in and zero beyond
2959 * i_size.
2960 *
2961 * Since we are holding inode lock, we are sure i_disksize <=
2962 * i_size. We also know that if i_disksize < i_size, there are
2963 * delalloc writes pending in the range up to i_size. If the end of
2964 * the current write is <= i_size, there's no need to touch
2965 * i_disksize since writeback will push i_disksize up to i_size
2966 * eventually. If the end of the current write is > i_size and
2967 * inside an allocated block which ext4_da_should_update_i_disksize()
2968 * checked, we need to update i_disksize here as certain
2969 * ext4_writepages() paths not allocating blocks and update i_disksize.
2970 */
2971 if (new_i_size > inode->i_size) {
2972 unsigned long end;
2973
2974 i_size_write(inode, new_i_size);
2975 end = (new_i_size - 1) & (PAGE_SIZE - 1);
2976 if (copied && ext4_da_should_update_i_disksize(folio, end)) {
2977 ext4_update_i_disksize(inode, new_i_size);
2978 disksize_changed = true;
2979 }
2980 }
2981
2982 folio_unlock(folio);
2983 folio_put(folio);
2984
2985 if (old_size < pos)
2986 pagecache_isize_extended(inode, old_size, pos);
2987
2988 if (disksize_changed) {
2989 handle_t *handle;
2990
2991 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
2992 if (IS_ERR(handle))
2993 return PTR_ERR(handle);
2994 ext4_mark_inode_dirty(handle, inode);
2995 ext4_journal_stop(handle);
2996 }
2997
2998 return copied;
2999}
3000
3001static int ext4_da_write_end(struct file *file,
3002 struct address_space *mapping,
3003 loff_t pos, unsigned len, unsigned copied,
3004 struct page *page, void *fsdata)
3005{
3006 struct inode *inode = mapping->host;
3007 int write_mode = (int)(unsigned long)fsdata;
3008 struct folio *folio = page_folio(page);
3009
3010 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3011 return ext4_write_end(file, mapping, pos,
3012 len, copied, &folio->page, fsdata);
3013
3014 trace_ext4_da_write_end(inode, pos, len, copied);
3015
3016 if (write_mode != CONVERT_INLINE_DATA &&
3017 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3018 ext4_has_inline_data(inode))
3019 return ext4_write_inline_data_end(inode, pos, len, copied,
3020 folio);
3021
3022 if (unlikely(copied < len) && !folio_test_uptodate(folio))
3023 copied = 0;
3024
3025 return ext4_da_do_write_end(mapping, pos, len, copied, folio);
3026}
3027
3028/*
3029 * Force all delayed allocation blocks to be allocated for a given inode.
3030 */
3031int ext4_alloc_da_blocks(struct inode *inode)
3032{
3033 trace_ext4_alloc_da_blocks(inode);
3034
3035 if (!EXT4_I(inode)->i_reserved_data_blocks)
3036 return 0;
3037
3038 /*
3039 * We do something simple for now. The filemap_flush() will
3040 * also start triggering a write of the data blocks, which is
3041 * not strictly speaking necessary (and for users of
3042 * laptop_mode, not even desirable). However, to do otherwise
3043 * would require replicating code paths in:
3044 *
3045 * ext4_writepages() ->
3046 * write_cache_pages() ---> (via passed in callback function)
3047 * __mpage_da_writepage() -->
3048 * mpage_add_bh_to_extent()
3049 * mpage_da_map_blocks()
3050 *
3051 * The problem is that write_cache_pages(), located in
3052 * mm/page-writeback.c, marks pages clean in preparation for
3053 * doing I/O, which is not desirable if we're not planning on
3054 * doing I/O at all.
3055 *
3056 * We could call write_cache_pages(), and then redirty all of
3057 * the pages by calling redirty_page_for_writepage() but that
3058 * would be ugly in the extreme. So instead we would need to
3059 * replicate parts of the code in the above functions,
3060 * simplifying them because we wouldn't actually intend to
3061 * write out the pages, but rather only collect contiguous
3062 * logical block extents, call the multi-block allocator, and
3063 * then update the buffer heads with the block allocations.
3064 *
3065 * For now, though, we'll cheat by calling filemap_flush(),
3066 * which will map the blocks, and start the I/O, but not
3067 * actually wait for the I/O to complete.
3068 */
3069 return filemap_flush(inode->i_mapping);
3070}
3071
3072/*
3073 * bmap() is special. It gets used by applications such as lilo and by
3074 * the swapper to find the on-disk block of a specific piece of data.
3075 *
3076 * Naturally, this is dangerous if the block concerned is still in the
3077 * journal. If somebody makes a swapfile on an ext4 data-journaling
3078 * filesystem and enables swap, then they may get a nasty shock when the
3079 * data getting swapped to that swapfile suddenly gets overwritten by
3080 * the original zero's written out previously to the journal and
3081 * awaiting writeback in the kernel's buffer cache.
3082 *
3083 * So, if we see any bmap calls here on a modified, data-journaled file,
3084 * take extra steps to flush any blocks which might be in the cache.
3085 */
3086static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3087{
3088 struct inode *inode = mapping->host;
3089 sector_t ret = 0;
3090
3091 inode_lock_shared(inode);
3092 /*
3093 * We can get here for an inline file via the FIBMAP ioctl
3094 */
3095 if (ext4_has_inline_data(inode))
3096 goto out;
3097
3098 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3099 (test_opt(inode->i_sb, DELALLOC) ||
3100 ext4_should_journal_data(inode))) {
3101 /*
3102 * With delalloc or journalled data we want to sync the file so
3103 * that we can make sure we allocate blocks for file and data
3104 * is in place for the user to see it
3105 */
3106 filemap_write_and_wait(mapping);
3107 }
3108
3109 ret = iomap_bmap(mapping, block, &ext4_iomap_ops);
3110
3111out:
3112 inode_unlock_shared(inode);
3113 return ret;
3114}
3115
3116static int ext4_read_folio(struct file *file, struct folio *folio)
3117{
3118 int ret = -EAGAIN;
3119 struct inode *inode = folio->mapping->host;
3120
3121 trace_ext4_read_folio(inode, folio);
3122
3123 if (ext4_has_inline_data(inode))
3124 ret = ext4_readpage_inline(inode, folio);
3125
3126 if (ret == -EAGAIN)
3127 return ext4_mpage_readpages(inode, NULL, folio);
3128
3129 return ret;
3130}
3131
3132static void ext4_readahead(struct readahead_control *rac)
3133{
3134 struct inode *inode = rac->mapping->host;
3135
3136 /* If the file has inline data, no need to do readahead. */
3137 if (ext4_has_inline_data(inode))
3138 return;
3139
3140 ext4_mpage_readpages(inode, rac, NULL);
3141}
3142
3143static void ext4_invalidate_folio(struct folio *folio, size_t offset,
3144 size_t length)
3145{
3146 trace_ext4_invalidate_folio(folio, offset, length);
3147
3148 /* No journalling happens on data buffers when this function is used */
3149 WARN_ON(folio_buffers(folio) && buffer_jbd(folio_buffers(folio)));
3150
3151 block_invalidate_folio(folio, offset, length);
3152}
3153
3154static int __ext4_journalled_invalidate_folio(struct folio *folio,
3155 size_t offset, size_t length)
3156{
3157 journal_t *journal = EXT4_JOURNAL(folio->mapping->host);
3158
3159 trace_ext4_journalled_invalidate_folio(folio, offset, length);
3160
3161 /*
3162 * If it's a full truncate we just forget about the pending dirtying
3163 */
3164 if (offset == 0 && length == folio_size(folio))
3165 folio_clear_checked(folio);
3166
3167 return jbd2_journal_invalidate_folio(journal, folio, offset, length);
3168}
3169
3170/* Wrapper for aops... */
3171static void ext4_journalled_invalidate_folio(struct folio *folio,
3172 size_t offset,
3173 size_t length)
3174{
3175 WARN_ON(__ext4_journalled_invalidate_folio(folio, offset, length) < 0);
3176}
3177
3178static bool ext4_release_folio(struct folio *folio, gfp_t wait)
3179{
3180 struct inode *inode = folio->mapping->host;
3181 journal_t *journal = EXT4_JOURNAL(inode);
3182
3183 trace_ext4_release_folio(inode, folio);
3184
3185 /* Page has dirty journalled data -> cannot release */
3186 if (folio_test_checked(folio))
3187 return false;
3188 if (journal)
3189 return jbd2_journal_try_to_free_buffers(journal, folio);
3190 else
3191 return try_to_free_buffers(folio);
3192}
3193
3194static bool ext4_inode_datasync_dirty(struct inode *inode)
3195{
3196 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3197
3198 if (journal) {
3199 if (jbd2_transaction_committed(journal,
3200 EXT4_I(inode)->i_datasync_tid))
3201 return false;
3202 if (test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT))
3203 return !list_empty(&EXT4_I(inode)->i_fc_list);
3204 return true;
3205 }
3206
3207 /* Any metadata buffers to write? */
3208 if (!list_empty(&inode->i_mapping->i_private_list))
3209 return true;
3210 return inode->i_state & I_DIRTY_DATASYNC;
3211}
3212
3213static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
3214 struct ext4_map_blocks *map, loff_t offset,
3215 loff_t length, unsigned int flags)
3216{
3217 u8 blkbits = inode->i_blkbits;
3218
3219 /*
3220 * Writes that span EOF might trigger an I/O size update on completion,
3221 * so consider them to be dirty for the purpose of O_DSYNC, even if
3222 * there is no other metadata changes being made or are pending.
3223 */
3224 iomap->flags = 0;
3225 if (ext4_inode_datasync_dirty(inode) ||
3226 offset + length > i_size_read(inode))
3227 iomap->flags |= IOMAP_F_DIRTY;
3228
3229 if (map->m_flags & EXT4_MAP_NEW)
3230 iomap->flags |= IOMAP_F_NEW;
3231
3232 if (flags & IOMAP_DAX)
3233 iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
3234 else
3235 iomap->bdev = inode->i_sb->s_bdev;
3236 iomap->offset = (u64) map->m_lblk << blkbits;
3237 iomap->length = (u64) map->m_len << blkbits;
3238
3239 if ((map->m_flags & EXT4_MAP_MAPPED) &&
3240 !ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3241 iomap->flags |= IOMAP_F_MERGED;
3242
3243 /*
3244 * Flags passed to ext4_map_blocks() for direct I/O writes can result
3245 * in m_flags having both EXT4_MAP_MAPPED and EXT4_MAP_UNWRITTEN bits
3246 * set. In order for any allocated unwritten extents to be converted
3247 * into written extents correctly within the ->end_io() handler, we
3248 * need to ensure that the iomap->type is set appropriately. Hence, the
3249 * reason why we need to check whether the EXT4_MAP_UNWRITTEN bit has
3250 * been set first.
3251 */
3252 if (map->m_flags & EXT4_MAP_UNWRITTEN) {
3253 iomap->type = IOMAP_UNWRITTEN;
3254 iomap->addr = (u64) map->m_pblk << blkbits;
3255 if (flags & IOMAP_DAX)
3256 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3257 } else if (map->m_flags & EXT4_MAP_MAPPED) {
3258 iomap->type = IOMAP_MAPPED;
3259 iomap->addr = (u64) map->m_pblk << blkbits;
3260 if (flags & IOMAP_DAX)
3261 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3262 } else if (map->m_flags & EXT4_MAP_DELAYED) {
3263 iomap->type = IOMAP_DELALLOC;
3264 iomap->addr = IOMAP_NULL_ADDR;
3265 } else {
3266 iomap->type = IOMAP_HOLE;
3267 iomap->addr = IOMAP_NULL_ADDR;
3268 }
3269}
3270
3271static int ext4_iomap_alloc(struct inode *inode, struct ext4_map_blocks *map,
3272 unsigned int flags)
3273{
3274 handle_t *handle;
3275 u8 blkbits = inode->i_blkbits;
3276 int ret, dio_credits, m_flags = 0, retries = 0;
3277
3278 /*
3279 * Trim the mapping request to the maximum value that we can map at
3280 * once for direct I/O.
3281 */
3282 if (map->m_len > DIO_MAX_BLOCKS)
3283 map->m_len = DIO_MAX_BLOCKS;
3284 dio_credits = ext4_chunk_trans_blocks(inode, map->m_len);
3285
3286retry:
3287 /*
3288 * Either we allocate blocks and then don't get an unwritten extent, so
3289 * in that case we have reserved enough credits. Or, the blocks are
3290 * already allocated and unwritten. In that case, the extent conversion
3291 * fits into the credits as well.
3292 */
3293 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
3294 if (IS_ERR(handle))
3295 return PTR_ERR(handle);
3296
3297 /*
3298 * DAX and direct I/O are the only two operations that are currently
3299 * supported with IOMAP_WRITE.
3300 */
3301 WARN_ON(!(flags & (IOMAP_DAX | IOMAP_DIRECT)));
3302 if (flags & IOMAP_DAX)
3303 m_flags = EXT4_GET_BLOCKS_CREATE_ZERO;
3304 /*
3305 * We use i_size instead of i_disksize here because delalloc writeback
3306 * can complete at any point during the I/O and subsequently push the
3307 * i_disksize out to i_size. This could be beyond where direct I/O is
3308 * happening and thus expose allocated blocks to direct I/O reads.
3309 */
3310 else if (((loff_t)map->m_lblk << blkbits) >= i_size_read(inode))
3311 m_flags = EXT4_GET_BLOCKS_CREATE;
3312 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3313 m_flags = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3314
3315 ret = ext4_map_blocks(handle, inode, map, m_flags);
3316
3317 /*
3318 * We cannot fill holes in indirect tree based inodes as that could
3319 * expose stale data in the case of a crash. Use the magic error code
3320 * to fallback to buffered I/O.
3321 */
3322 if (!m_flags && !ret)
3323 ret = -ENOTBLK;
3324
3325 ext4_journal_stop(handle);
3326 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3327 goto retry;
3328
3329 return ret;
3330}
3331
3332
3333static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3334 unsigned flags, struct iomap *iomap, struct iomap *srcmap)
3335{
3336 int ret;
3337 struct ext4_map_blocks map;
3338 u8 blkbits = inode->i_blkbits;
3339
3340 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3341 return -EINVAL;
3342
3343 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3344 return -ERANGE;
3345
3346 /*
3347 * Calculate the first and last logical blocks respectively.
3348 */
3349 map.m_lblk = offset >> blkbits;
3350 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3351 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3352
3353 if (flags & IOMAP_WRITE) {
3354 /*
3355 * We check here if the blocks are already allocated, then we
3356 * don't need to start a journal txn and we can directly return
3357 * the mapping information. This could boost performance
3358 * especially in multi-threaded overwrite requests.
3359 */
3360 if (offset + length <= i_size_read(inode)) {
3361 ret = ext4_map_blocks(NULL, inode, &map, 0);
3362 if (ret > 0 && (map.m_flags & EXT4_MAP_MAPPED))
3363 goto out;
3364 }
3365 ret = ext4_iomap_alloc(inode, &map, flags);
3366 } else {
3367 ret = ext4_map_blocks(NULL, inode, &map, 0);
3368 }
3369
3370 if (ret < 0)
3371 return ret;
3372out:
3373 /*
3374 * When inline encryption is enabled, sometimes I/O to an encrypted file
3375 * has to be broken up to guarantee DUN contiguity. Handle this by
3376 * limiting the length of the mapping returned.
3377 */
3378 map.m_len = fscrypt_limit_io_blocks(inode, map.m_lblk, map.m_len);
3379
3380 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3381
3382 return 0;
3383}
3384
3385static int ext4_iomap_overwrite_begin(struct inode *inode, loff_t offset,
3386 loff_t length, unsigned flags, struct iomap *iomap,
3387 struct iomap *srcmap)
3388{
3389 int ret;
3390
3391 /*
3392 * Even for writes we don't need to allocate blocks, so just pretend
3393 * we are reading to save overhead of starting a transaction.
3394 */
3395 flags &= ~IOMAP_WRITE;
3396 ret = ext4_iomap_begin(inode, offset, length, flags, iomap, srcmap);
3397 WARN_ON_ONCE(!ret && iomap->type != IOMAP_MAPPED);
3398 return ret;
3399}
3400
3401static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3402 ssize_t written, unsigned flags, struct iomap *iomap)
3403{
3404 /*
3405 * Check to see whether an error occurred while writing out the data to
3406 * the allocated blocks. If so, return the magic error code so that we
3407 * fallback to buffered I/O and attempt to complete the remainder of
3408 * the I/O. Any blocks that may have been allocated in preparation for
3409 * the direct I/O will be reused during buffered I/O.
3410 */
3411 if (flags & (IOMAP_WRITE | IOMAP_DIRECT) && written == 0)
3412 return -ENOTBLK;
3413
3414 return 0;
3415}
3416
3417const struct iomap_ops ext4_iomap_ops = {
3418 .iomap_begin = ext4_iomap_begin,
3419 .iomap_end = ext4_iomap_end,
3420};
3421
3422const struct iomap_ops ext4_iomap_overwrite_ops = {
3423 .iomap_begin = ext4_iomap_overwrite_begin,
3424 .iomap_end = ext4_iomap_end,
3425};
3426
3427static int ext4_iomap_begin_report(struct inode *inode, loff_t offset,
3428 loff_t length, unsigned int flags,
3429 struct iomap *iomap, struct iomap *srcmap)
3430{
3431 int ret;
3432 struct ext4_map_blocks map;
3433 u8 blkbits = inode->i_blkbits;
3434
3435 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3436 return -EINVAL;
3437
3438 if (ext4_has_inline_data(inode)) {
3439 ret = ext4_inline_data_iomap(inode, iomap);
3440 if (ret != -EAGAIN) {
3441 if (ret == 0 && offset >= iomap->length)
3442 ret = -ENOENT;
3443 return ret;
3444 }
3445 }
3446
3447 /*
3448 * Calculate the first and last logical block respectively.
3449 */
3450 map.m_lblk = offset >> blkbits;
3451 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3452 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3453
3454 /*
3455 * Fiemap callers may call for offset beyond s_bitmap_maxbytes.
3456 * So handle it here itself instead of querying ext4_map_blocks().
3457 * Since ext4_map_blocks() will warn about it and will return
3458 * -EIO error.
3459 */
3460 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3461 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3462
3463 if (offset >= sbi->s_bitmap_maxbytes) {
3464 map.m_flags = 0;
3465 goto set_iomap;
3466 }
3467 }
3468
3469 ret = ext4_map_blocks(NULL, inode, &map, 0);
3470 if (ret < 0)
3471 return ret;
3472set_iomap:
3473 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3474
3475 return 0;
3476}
3477
3478const struct iomap_ops ext4_iomap_report_ops = {
3479 .iomap_begin = ext4_iomap_begin_report,
3480};
3481
3482/*
3483 * For data=journal mode, folio should be marked dirty only when it was
3484 * writeably mapped. When that happens, it was already attached to the
3485 * transaction and marked as jbddirty (we take care of this in
3486 * ext4_page_mkwrite()). On transaction commit, we writeprotect page mappings
3487 * so we should have nothing to do here, except for the case when someone
3488 * had the page pinned and dirtied the page through this pin (e.g. by doing
3489 * direct IO to it). In that case we'd need to attach buffers here to the
3490 * transaction but we cannot due to lock ordering. We cannot just dirty the
3491 * folio and leave attached buffers clean, because the buffers' dirty state is
3492 * "definitive". We cannot just set the buffers dirty or jbddirty because all
3493 * the journalling code will explode. So what we do is to mark the folio
3494 * "pending dirty" and next time ext4_writepages() is called, attach buffers
3495 * to the transaction appropriately.
3496 */
3497static bool ext4_journalled_dirty_folio(struct address_space *mapping,
3498 struct folio *folio)
3499{
3500 WARN_ON_ONCE(!folio_buffers(folio));
3501 if (folio_maybe_dma_pinned(folio))
3502 folio_set_checked(folio);
3503 return filemap_dirty_folio(mapping, folio);
3504}
3505
3506static bool ext4_dirty_folio(struct address_space *mapping, struct folio *folio)
3507{
3508 WARN_ON_ONCE(!folio_test_locked(folio) && !folio_test_dirty(folio));
3509 WARN_ON_ONCE(!folio_buffers(folio));
3510 return block_dirty_folio(mapping, folio);
3511}
3512
3513static int ext4_iomap_swap_activate(struct swap_info_struct *sis,
3514 struct file *file, sector_t *span)
3515{
3516 return iomap_swapfile_activate(sis, file, span,
3517 &ext4_iomap_report_ops);
3518}
3519
3520static const struct address_space_operations ext4_aops = {
3521 .read_folio = ext4_read_folio,
3522 .readahead = ext4_readahead,
3523 .writepages = ext4_writepages,
3524 .write_begin = ext4_write_begin,
3525 .write_end = ext4_write_end,
3526 .dirty_folio = ext4_dirty_folio,
3527 .bmap = ext4_bmap,
3528 .invalidate_folio = ext4_invalidate_folio,
3529 .release_folio = ext4_release_folio,
3530 .direct_IO = noop_direct_IO,
3531 .migrate_folio = buffer_migrate_folio,
3532 .is_partially_uptodate = block_is_partially_uptodate,
3533 .error_remove_folio = generic_error_remove_folio,
3534 .swap_activate = ext4_iomap_swap_activate,
3535};
3536
3537static const struct address_space_operations ext4_journalled_aops = {
3538 .read_folio = ext4_read_folio,
3539 .readahead = ext4_readahead,
3540 .writepages = ext4_writepages,
3541 .write_begin = ext4_write_begin,
3542 .write_end = ext4_journalled_write_end,
3543 .dirty_folio = ext4_journalled_dirty_folio,
3544 .bmap = ext4_bmap,
3545 .invalidate_folio = ext4_journalled_invalidate_folio,
3546 .release_folio = ext4_release_folio,
3547 .direct_IO = noop_direct_IO,
3548 .migrate_folio = buffer_migrate_folio_norefs,
3549 .is_partially_uptodate = block_is_partially_uptodate,
3550 .error_remove_folio = generic_error_remove_folio,
3551 .swap_activate = ext4_iomap_swap_activate,
3552};
3553
3554static const struct address_space_operations ext4_da_aops = {
3555 .read_folio = ext4_read_folio,
3556 .readahead = ext4_readahead,
3557 .writepages = ext4_writepages,
3558 .write_begin = ext4_da_write_begin,
3559 .write_end = ext4_da_write_end,
3560 .dirty_folio = ext4_dirty_folio,
3561 .bmap = ext4_bmap,
3562 .invalidate_folio = ext4_invalidate_folio,
3563 .release_folio = ext4_release_folio,
3564 .direct_IO = noop_direct_IO,
3565 .migrate_folio = buffer_migrate_folio,
3566 .is_partially_uptodate = block_is_partially_uptodate,
3567 .error_remove_folio = generic_error_remove_folio,
3568 .swap_activate = ext4_iomap_swap_activate,
3569};
3570
3571static const struct address_space_operations ext4_dax_aops = {
3572 .writepages = ext4_dax_writepages,
3573 .direct_IO = noop_direct_IO,
3574 .dirty_folio = noop_dirty_folio,
3575 .bmap = ext4_bmap,
3576 .swap_activate = ext4_iomap_swap_activate,
3577};
3578
3579void ext4_set_aops(struct inode *inode)
3580{
3581 switch (ext4_inode_journal_mode(inode)) {
3582 case EXT4_INODE_ORDERED_DATA_MODE:
3583 case EXT4_INODE_WRITEBACK_DATA_MODE:
3584 break;
3585 case EXT4_INODE_JOURNAL_DATA_MODE:
3586 inode->i_mapping->a_ops = &ext4_journalled_aops;
3587 return;
3588 default:
3589 BUG();
3590 }
3591 if (IS_DAX(inode))
3592 inode->i_mapping->a_ops = &ext4_dax_aops;
3593 else if (test_opt(inode->i_sb, DELALLOC))
3594 inode->i_mapping->a_ops = &ext4_da_aops;
3595 else
3596 inode->i_mapping->a_ops = &ext4_aops;
3597}
3598
3599/*
3600 * Here we can't skip an unwritten buffer even though it usually reads zero
3601 * because it might have data in pagecache (eg, if called from ext4_zero_range,
3602 * ext4_punch_hole, etc) which needs to be properly zeroed out. Otherwise a
3603 * racing writeback can come later and flush the stale pagecache to disk.
3604 */
3605static int __ext4_block_zero_page_range(handle_t *handle,
3606 struct address_space *mapping, loff_t from, loff_t length)
3607{
3608 ext4_fsblk_t index = from >> PAGE_SHIFT;
3609 unsigned offset = from & (PAGE_SIZE-1);
3610 unsigned blocksize, pos;
3611 ext4_lblk_t iblock;
3612 struct inode *inode = mapping->host;
3613 struct buffer_head *bh;
3614 struct folio *folio;
3615 int err = 0;
3616
3617 folio = __filemap_get_folio(mapping, from >> PAGE_SHIFT,
3618 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
3619 mapping_gfp_constraint(mapping, ~__GFP_FS));
3620 if (IS_ERR(folio))
3621 return PTR_ERR(folio);
3622
3623 blocksize = inode->i_sb->s_blocksize;
3624
3625 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3626
3627 bh = folio_buffers(folio);
3628 if (!bh)
3629 bh = create_empty_buffers(folio, blocksize, 0);
3630
3631 /* Find the buffer that contains "offset" */
3632 pos = blocksize;
3633 while (offset >= pos) {
3634 bh = bh->b_this_page;
3635 iblock++;
3636 pos += blocksize;
3637 }
3638 if (buffer_freed(bh)) {
3639 BUFFER_TRACE(bh, "freed: skip");
3640 goto unlock;
3641 }
3642 if (!buffer_mapped(bh)) {
3643 BUFFER_TRACE(bh, "unmapped");
3644 ext4_get_block(inode, iblock, bh, 0);
3645 /* unmapped? It's a hole - nothing to do */
3646 if (!buffer_mapped(bh)) {
3647 BUFFER_TRACE(bh, "still unmapped");
3648 goto unlock;
3649 }
3650 }
3651
3652 /* Ok, it's mapped. Make sure it's up-to-date */
3653 if (folio_test_uptodate(folio))
3654 set_buffer_uptodate(bh);
3655
3656 if (!buffer_uptodate(bh)) {
3657 err = ext4_read_bh_lock(bh, 0, true);
3658 if (err)
3659 goto unlock;
3660 if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
3661 /* We expect the key to be set. */
3662 BUG_ON(!fscrypt_has_encryption_key(inode));
3663 err = fscrypt_decrypt_pagecache_blocks(folio,
3664 blocksize,
3665 bh_offset(bh));
3666 if (err) {
3667 clear_buffer_uptodate(bh);
3668 goto unlock;
3669 }
3670 }
3671 }
3672 if (ext4_should_journal_data(inode)) {
3673 BUFFER_TRACE(bh, "get write access");
3674 err = ext4_journal_get_write_access(handle, inode->i_sb, bh,
3675 EXT4_JTR_NONE);
3676 if (err)
3677 goto unlock;
3678 }
3679 folio_zero_range(folio, offset, length);
3680 BUFFER_TRACE(bh, "zeroed end of block");
3681
3682 if (ext4_should_journal_data(inode)) {
3683 err = ext4_dirty_journalled_data(handle, bh);
3684 } else {
3685 err = 0;
3686 mark_buffer_dirty(bh);
3687 if (ext4_should_order_data(inode))
3688 err = ext4_jbd2_inode_add_write(handle, inode, from,
3689 length);
3690 }
3691
3692unlock:
3693 folio_unlock(folio);
3694 folio_put(folio);
3695 return err;
3696}
3697
3698/*
3699 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3700 * starting from file offset 'from'. The range to be zero'd must
3701 * be contained with in one block. If the specified range exceeds
3702 * the end of the block it will be shortened to end of the block
3703 * that corresponds to 'from'
3704 */
3705static int ext4_block_zero_page_range(handle_t *handle,
3706 struct address_space *mapping, loff_t from, loff_t length)
3707{
3708 struct inode *inode = mapping->host;
3709 unsigned offset = from & (PAGE_SIZE-1);
3710 unsigned blocksize = inode->i_sb->s_blocksize;
3711 unsigned max = blocksize - (offset & (blocksize - 1));
3712
3713 /*
3714 * correct length if it does not fall between
3715 * 'from' and the end of the block
3716 */
3717 if (length > max || length < 0)
3718 length = max;
3719
3720 if (IS_DAX(inode)) {
3721 return dax_zero_range(inode, from, length, NULL,
3722 &ext4_iomap_ops);
3723 }
3724 return __ext4_block_zero_page_range(handle, mapping, from, length);
3725}
3726
3727/*
3728 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3729 * up to the end of the block which corresponds to `from'.
3730 * This required during truncate. We need to physically zero the tail end
3731 * of that block so it doesn't yield old data if the file is later grown.
3732 */
3733static int ext4_block_truncate_page(handle_t *handle,
3734 struct address_space *mapping, loff_t from)
3735{
3736 unsigned offset = from & (PAGE_SIZE-1);
3737 unsigned length;
3738 unsigned blocksize;
3739 struct inode *inode = mapping->host;
3740
3741 /* If we are processing an encrypted inode during orphan list handling */
3742 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
3743 return 0;
3744
3745 blocksize = inode->i_sb->s_blocksize;
3746 length = blocksize - (offset & (blocksize - 1));
3747
3748 return ext4_block_zero_page_range(handle, mapping, from, length);
3749}
3750
3751int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3752 loff_t lstart, loff_t length)
3753{
3754 struct super_block *sb = inode->i_sb;
3755 struct address_space *mapping = inode->i_mapping;
3756 unsigned partial_start, partial_end;
3757 ext4_fsblk_t start, end;
3758 loff_t byte_end = (lstart + length - 1);
3759 int err = 0;
3760
3761 partial_start = lstart & (sb->s_blocksize - 1);
3762 partial_end = byte_end & (sb->s_blocksize - 1);
3763
3764 start = lstart >> sb->s_blocksize_bits;
3765 end = byte_end >> sb->s_blocksize_bits;
3766
3767 /* Handle partial zero within the single block */
3768 if (start == end &&
3769 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3770 err = ext4_block_zero_page_range(handle, mapping,
3771 lstart, length);
3772 return err;
3773 }
3774 /* Handle partial zero out on the start of the range */
3775 if (partial_start) {
3776 err = ext4_block_zero_page_range(handle, mapping,
3777 lstart, sb->s_blocksize);
3778 if (err)
3779 return err;
3780 }
3781 /* Handle partial zero out on the end of the range */
3782 if (partial_end != sb->s_blocksize - 1)
3783 err = ext4_block_zero_page_range(handle, mapping,
3784 byte_end - partial_end,
3785 partial_end + 1);
3786 return err;
3787}
3788
3789int ext4_can_truncate(struct inode *inode)
3790{
3791 if (S_ISREG(inode->i_mode))
3792 return 1;
3793 if (S_ISDIR(inode->i_mode))
3794 return 1;
3795 if (S_ISLNK(inode->i_mode))
3796 return !ext4_inode_is_fast_symlink(inode);
3797 return 0;
3798}
3799
3800/*
3801 * We have to make sure i_disksize gets properly updated before we truncate
3802 * page cache due to hole punching or zero range. Otherwise i_disksize update
3803 * can get lost as it may have been postponed to submission of writeback but
3804 * that will never happen after we truncate page cache.
3805 */
3806int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
3807 loff_t len)
3808{
3809 handle_t *handle;
3810 int ret;
3811
3812 loff_t size = i_size_read(inode);
3813
3814 WARN_ON(!inode_is_locked(inode));
3815 if (offset > size || offset + len < size)
3816 return 0;
3817
3818 if (EXT4_I(inode)->i_disksize >= size)
3819 return 0;
3820
3821 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
3822 if (IS_ERR(handle))
3823 return PTR_ERR(handle);
3824 ext4_update_i_disksize(inode, size);
3825 ret = ext4_mark_inode_dirty(handle, inode);
3826 ext4_journal_stop(handle);
3827
3828 return ret;
3829}
3830
3831static void ext4_wait_dax_page(struct inode *inode)
3832{
3833 filemap_invalidate_unlock(inode->i_mapping);
3834 schedule();
3835 filemap_invalidate_lock(inode->i_mapping);
3836}
3837
3838int ext4_break_layouts(struct inode *inode)
3839{
3840 struct page *page;
3841 int error;
3842
3843 if (WARN_ON_ONCE(!rwsem_is_locked(&inode->i_mapping->invalidate_lock)))
3844 return -EINVAL;
3845
3846 do {
3847 page = dax_layout_busy_page(inode->i_mapping);
3848 if (!page)
3849 return 0;
3850
3851 error = ___wait_var_event(&page->_refcount,
3852 atomic_read(&page->_refcount) == 1,
3853 TASK_INTERRUPTIBLE, 0, 0,
3854 ext4_wait_dax_page(inode));
3855 } while (error == 0);
3856
3857 return error;
3858}
3859
3860/*
3861 * ext4_punch_hole: punches a hole in a file by releasing the blocks
3862 * associated with the given offset and length
3863 *
3864 * @inode: File inode
3865 * @offset: The offset where the hole will begin
3866 * @len: The length of the hole
3867 *
3868 * Returns: 0 on success or negative on failure
3869 */
3870
3871int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3872{
3873 struct inode *inode = file_inode(file);
3874 struct super_block *sb = inode->i_sb;
3875 ext4_lblk_t first_block, stop_block;
3876 struct address_space *mapping = inode->i_mapping;
3877 loff_t first_block_offset, last_block_offset, max_length;
3878 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3879 handle_t *handle;
3880 unsigned int credits;
3881 int ret = 0, ret2 = 0;
3882
3883 trace_ext4_punch_hole(inode, offset, length, 0);
3884
3885 /*
3886 * Write out all dirty pages to avoid race conditions
3887 * Then release them.
3888 */
3889 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3890 ret = filemap_write_and_wait_range(mapping, offset,
3891 offset + length - 1);
3892 if (ret)
3893 return ret;
3894 }
3895
3896 inode_lock(inode);
3897
3898 /* No need to punch hole beyond i_size */
3899 if (offset >= inode->i_size)
3900 goto out_mutex;
3901
3902 /*
3903 * If the hole extends beyond i_size, set the hole
3904 * to end after the page that contains i_size
3905 */
3906 if (offset + length > inode->i_size) {
3907 length = inode->i_size +
3908 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
3909 offset;
3910 }
3911
3912 /*
3913 * For punch hole the length + offset needs to be within one block
3914 * before last range. Adjust the length if it goes beyond that limit.
3915 */
3916 max_length = sbi->s_bitmap_maxbytes - inode->i_sb->s_blocksize;
3917 if (offset + length > max_length)
3918 length = max_length - offset;
3919
3920 if (offset & (sb->s_blocksize - 1) ||
3921 (offset + length) & (sb->s_blocksize - 1)) {
3922 /*
3923 * Attach jinode to inode for jbd2 if we do any zeroing of
3924 * partial block
3925 */
3926 ret = ext4_inode_attach_jinode(inode);
3927 if (ret < 0)
3928 goto out_mutex;
3929
3930 }
3931
3932 /* Wait all existing dio workers, newcomers will block on i_rwsem */
3933 inode_dio_wait(inode);
3934
3935 ret = file_modified(file);
3936 if (ret)
3937 goto out_mutex;
3938
3939 /*
3940 * Prevent page faults from reinstantiating pages we have released from
3941 * page cache.
3942 */
3943 filemap_invalidate_lock(mapping);
3944
3945 ret = ext4_break_layouts(inode);
3946 if (ret)
3947 goto out_dio;
3948
3949 first_block_offset = round_up(offset, sb->s_blocksize);
3950 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
3951
3952 /* Now release the pages and zero block aligned part of pages*/
3953 if (last_block_offset > first_block_offset) {
3954 ret = ext4_update_disksize_before_punch(inode, offset, length);
3955 if (ret)
3956 goto out_dio;
3957 truncate_pagecache_range(inode, first_block_offset,
3958 last_block_offset);
3959 }
3960
3961 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3962 credits = ext4_writepage_trans_blocks(inode);
3963 else
3964 credits = ext4_blocks_for_truncate(inode);
3965 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3966 if (IS_ERR(handle)) {
3967 ret = PTR_ERR(handle);
3968 ext4_std_error(sb, ret);
3969 goto out_dio;
3970 }
3971
3972 ret = ext4_zero_partial_blocks(handle, inode, offset,
3973 length);
3974 if (ret)
3975 goto out_stop;
3976
3977 first_block = (offset + sb->s_blocksize - 1) >>
3978 EXT4_BLOCK_SIZE_BITS(sb);
3979 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
3980
3981 /* If there are blocks to remove, do it */
3982 if (stop_block > first_block) {
3983 ext4_lblk_t hole_len = stop_block - first_block;
3984
3985 down_write(&EXT4_I(inode)->i_data_sem);
3986 ext4_discard_preallocations(inode);
3987
3988 ext4_es_remove_extent(inode, first_block, hole_len);
3989
3990 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3991 ret = ext4_ext_remove_space(inode, first_block,
3992 stop_block - 1);
3993 else
3994 ret = ext4_ind_remove_space(handle, inode, first_block,
3995 stop_block);
3996
3997 ext4_es_insert_extent(inode, first_block, hole_len, ~0,
3998 EXTENT_STATUS_HOLE);
3999 up_write(&EXT4_I(inode)->i_data_sem);
4000 }
4001 ext4_fc_track_range(handle, inode, first_block, stop_block);
4002 if (IS_SYNC(inode))
4003 ext4_handle_sync(handle);
4004
4005 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
4006 ret2 = ext4_mark_inode_dirty(handle, inode);
4007 if (unlikely(ret2))
4008 ret = ret2;
4009 if (ret >= 0)
4010 ext4_update_inode_fsync_trans(handle, inode, 1);
4011out_stop:
4012 ext4_journal_stop(handle);
4013out_dio:
4014 filemap_invalidate_unlock(mapping);
4015out_mutex:
4016 inode_unlock(inode);
4017 return ret;
4018}
4019
4020int ext4_inode_attach_jinode(struct inode *inode)
4021{
4022 struct ext4_inode_info *ei = EXT4_I(inode);
4023 struct jbd2_inode *jinode;
4024
4025 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4026 return 0;
4027
4028 jinode = jbd2_alloc_inode(GFP_KERNEL);
4029 spin_lock(&inode->i_lock);
4030 if (!ei->jinode) {
4031 if (!jinode) {
4032 spin_unlock(&inode->i_lock);
4033 return -ENOMEM;
4034 }
4035 ei->jinode = jinode;
4036 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4037 jinode = NULL;
4038 }
4039 spin_unlock(&inode->i_lock);
4040 if (unlikely(jinode != NULL))
4041 jbd2_free_inode(jinode);
4042 return 0;
4043}
4044
4045/*
4046 * ext4_truncate()
4047 *
4048 * We block out ext4_get_block() block instantiations across the entire
4049 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4050 * simultaneously on behalf of the same inode.
4051 *
4052 * As we work through the truncate and commit bits of it to the journal there
4053 * is one core, guiding principle: the file's tree must always be consistent on
4054 * disk. We must be able to restart the truncate after a crash.
4055 *
4056 * The file's tree may be transiently inconsistent in memory (although it
4057 * probably isn't), but whenever we close off and commit a journal transaction,
4058 * the contents of (the filesystem + the journal) must be consistent and
4059 * restartable. It's pretty simple, really: bottom up, right to left (although
4060 * left-to-right works OK too).
4061 *
4062 * Note that at recovery time, journal replay occurs *before* the restart of
4063 * truncate against the orphan inode list.
4064 *
4065 * The committed inode has the new, desired i_size (which is the same as
4066 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4067 * that this inode's truncate did not complete and it will again call
4068 * ext4_truncate() to have another go. So there will be instantiated blocks
4069 * to the right of the truncation point in a crashed ext4 filesystem. But
4070 * that's fine - as long as they are linked from the inode, the post-crash
4071 * ext4_truncate() run will find them and release them.
4072 */
4073int ext4_truncate(struct inode *inode)
4074{
4075 struct ext4_inode_info *ei = EXT4_I(inode);
4076 unsigned int credits;
4077 int err = 0, err2;
4078 handle_t *handle;
4079 struct address_space *mapping = inode->i_mapping;
4080
4081 /*
4082 * There is a possibility that we're either freeing the inode
4083 * or it's a completely new inode. In those cases we might not
4084 * have i_rwsem locked because it's not necessary.
4085 */
4086 if (!(inode->i_state & (I_NEW|I_FREEING)))
4087 WARN_ON(!inode_is_locked(inode));
4088 trace_ext4_truncate_enter(inode);
4089
4090 if (!ext4_can_truncate(inode))
4091 goto out_trace;
4092
4093 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4094 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4095
4096 if (ext4_has_inline_data(inode)) {
4097 int has_inline = 1;
4098
4099 err = ext4_inline_data_truncate(inode, &has_inline);
4100 if (err || has_inline)
4101 goto out_trace;
4102 }
4103
4104 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4105 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4106 err = ext4_inode_attach_jinode(inode);
4107 if (err)
4108 goto out_trace;
4109 }
4110
4111 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4112 credits = ext4_writepage_trans_blocks(inode);
4113 else
4114 credits = ext4_blocks_for_truncate(inode);
4115
4116 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4117 if (IS_ERR(handle)) {
4118 err = PTR_ERR(handle);
4119 goto out_trace;
4120 }
4121
4122 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4123 ext4_block_truncate_page(handle, mapping, inode->i_size);
4124
4125 /*
4126 * We add the inode to the orphan list, so that if this
4127 * truncate spans multiple transactions, and we crash, we will
4128 * resume the truncate when the filesystem recovers. It also
4129 * marks the inode dirty, to catch the new size.
4130 *
4131 * Implication: the file must always be in a sane, consistent
4132 * truncatable state while each transaction commits.
4133 */
4134 err = ext4_orphan_add(handle, inode);
4135 if (err)
4136 goto out_stop;
4137
4138 down_write(&EXT4_I(inode)->i_data_sem);
4139
4140 ext4_discard_preallocations(inode);
4141
4142 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4143 err = ext4_ext_truncate(handle, inode);
4144 else
4145 ext4_ind_truncate(handle, inode);
4146
4147 up_write(&ei->i_data_sem);
4148 if (err)
4149 goto out_stop;
4150
4151 if (IS_SYNC(inode))
4152 ext4_handle_sync(handle);
4153
4154out_stop:
4155 /*
4156 * If this was a simple ftruncate() and the file will remain alive,
4157 * then we need to clear up the orphan record which we created above.
4158 * However, if this was a real unlink then we were called by
4159 * ext4_evict_inode(), and we allow that function to clean up the
4160 * orphan info for us.
4161 */
4162 if (inode->i_nlink)
4163 ext4_orphan_del(handle, inode);
4164
4165 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
4166 err2 = ext4_mark_inode_dirty(handle, inode);
4167 if (unlikely(err2 && !err))
4168 err = err2;
4169 ext4_journal_stop(handle);
4170
4171out_trace:
4172 trace_ext4_truncate_exit(inode);
4173 return err;
4174}
4175
4176static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4177{
4178 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4179 return inode_peek_iversion_raw(inode);
4180 else
4181 return inode_peek_iversion(inode);
4182}
4183
4184static int ext4_inode_blocks_set(struct ext4_inode *raw_inode,
4185 struct ext4_inode_info *ei)
4186{
4187 struct inode *inode = &(ei->vfs_inode);
4188 u64 i_blocks = READ_ONCE(inode->i_blocks);
4189 struct super_block *sb = inode->i_sb;
4190
4191 if (i_blocks <= ~0U) {
4192 /*
4193 * i_blocks can be represented in a 32 bit variable
4194 * as multiple of 512 bytes
4195 */
4196 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4197 raw_inode->i_blocks_high = 0;
4198 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4199 return 0;
4200 }
4201
4202 /*
4203 * This should never happen since sb->s_maxbytes should not have
4204 * allowed this, sb->s_maxbytes was set according to the huge_file
4205 * feature in ext4_fill_super().
4206 */
4207 if (!ext4_has_feature_huge_file(sb))
4208 return -EFSCORRUPTED;
4209
4210 if (i_blocks <= 0xffffffffffffULL) {
4211 /*
4212 * i_blocks can be represented in a 48 bit variable
4213 * as multiple of 512 bytes
4214 */
4215 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4216 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4217 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4218 } else {
4219 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4220 /* i_block is stored in file system block size */
4221 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4222 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4223 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4224 }
4225 return 0;
4226}
4227
4228static int ext4_fill_raw_inode(struct inode *inode, struct ext4_inode *raw_inode)
4229{
4230 struct ext4_inode_info *ei = EXT4_I(inode);
4231 uid_t i_uid;
4232 gid_t i_gid;
4233 projid_t i_projid;
4234 int block;
4235 int err;
4236
4237 err = ext4_inode_blocks_set(raw_inode, ei);
4238
4239 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4240 i_uid = i_uid_read(inode);
4241 i_gid = i_gid_read(inode);
4242 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4243 if (!(test_opt(inode->i_sb, NO_UID32))) {
4244 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4245 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4246 /*
4247 * Fix up interoperability with old kernels. Otherwise,
4248 * old inodes get re-used with the upper 16 bits of the
4249 * uid/gid intact.
4250 */
4251 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4252 raw_inode->i_uid_high = 0;
4253 raw_inode->i_gid_high = 0;
4254 } else {
4255 raw_inode->i_uid_high =
4256 cpu_to_le16(high_16_bits(i_uid));
4257 raw_inode->i_gid_high =
4258 cpu_to_le16(high_16_bits(i_gid));
4259 }
4260 } else {
4261 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4262 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4263 raw_inode->i_uid_high = 0;
4264 raw_inode->i_gid_high = 0;
4265 }
4266 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4267
4268 EXT4_INODE_SET_CTIME(inode, raw_inode);
4269 EXT4_INODE_SET_MTIME(inode, raw_inode);
4270 EXT4_INODE_SET_ATIME(inode, raw_inode);
4271 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4272
4273 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4274 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4275 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4276 raw_inode->i_file_acl_high =
4277 cpu_to_le16(ei->i_file_acl >> 32);
4278 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4279 ext4_isize_set(raw_inode, ei->i_disksize);
4280
4281 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4282 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4283 if (old_valid_dev(inode->i_rdev)) {
4284 raw_inode->i_block[0] =
4285 cpu_to_le32(old_encode_dev(inode->i_rdev));
4286 raw_inode->i_block[1] = 0;
4287 } else {
4288 raw_inode->i_block[0] = 0;
4289 raw_inode->i_block[1] =
4290 cpu_to_le32(new_encode_dev(inode->i_rdev));
4291 raw_inode->i_block[2] = 0;
4292 }
4293 } else if (!ext4_has_inline_data(inode)) {
4294 for (block = 0; block < EXT4_N_BLOCKS; block++)
4295 raw_inode->i_block[block] = ei->i_data[block];
4296 }
4297
4298 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4299 u64 ivers = ext4_inode_peek_iversion(inode);
4300
4301 raw_inode->i_disk_version = cpu_to_le32(ivers);
4302 if (ei->i_extra_isize) {
4303 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4304 raw_inode->i_version_hi =
4305 cpu_to_le32(ivers >> 32);
4306 raw_inode->i_extra_isize =
4307 cpu_to_le16(ei->i_extra_isize);
4308 }
4309 }
4310
4311 if (i_projid != EXT4_DEF_PROJID &&
4312 !ext4_has_feature_project(inode->i_sb))
4313 err = err ?: -EFSCORRUPTED;
4314
4315 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4316 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4317 raw_inode->i_projid = cpu_to_le32(i_projid);
4318
4319 ext4_inode_csum_set(inode, raw_inode, ei);
4320 return err;
4321}
4322
4323/*
4324 * ext4_get_inode_loc returns with an extra refcount against the inode's
4325 * underlying buffer_head on success. If we pass 'inode' and it does not
4326 * have in-inode xattr, we have all inode data in memory that is needed
4327 * to recreate the on-disk version of this inode.
4328 */
4329static int __ext4_get_inode_loc(struct super_block *sb, unsigned long ino,
4330 struct inode *inode, struct ext4_iloc *iloc,
4331 ext4_fsblk_t *ret_block)
4332{
4333 struct ext4_group_desc *gdp;
4334 struct buffer_head *bh;
4335 ext4_fsblk_t block;
4336 struct blk_plug plug;
4337 int inodes_per_block, inode_offset;
4338
4339 iloc->bh = NULL;
4340 if (ino < EXT4_ROOT_INO ||
4341 ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4342 return -EFSCORRUPTED;
4343
4344 iloc->block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
4345 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4346 if (!gdp)
4347 return -EIO;
4348
4349 /*
4350 * Figure out the offset within the block group inode table
4351 */
4352 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4353 inode_offset = ((ino - 1) %
4354 EXT4_INODES_PER_GROUP(sb));
4355 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4356
4357 block = ext4_inode_table(sb, gdp);
4358 if ((block <= le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) ||
4359 (block >= ext4_blocks_count(EXT4_SB(sb)->s_es))) {
4360 ext4_error(sb, "Invalid inode table block %llu in "
4361 "block_group %u", block, iloc->block_group);
4362 return -EFSCORRUPTED;
4363 }
4364 block += (inode_offset / inodes_per_block);
4365
4366 bh = sb_getblk(sb, block);
4367 if (unlikely(!bh))
4368 return -ENOMEM;
4369 if (ext4_buffer_uptodate(bh))
4370 goto has_buffer;
4371
4372 lock_buffer(bh);
4373 if (ext4_buffer_uptodate(bh)) {
4374 /* Someone brought it uptodate while we waited */
4375 unlock_buffer(bh);
4376 goto has_buffer;
4377 }
4378
4379 /*
4380 * If we have all information of the inode in memory and this
4381 * is the only valid inode in the block, we need not read the
4382 * block.
4383 */
4384 if (inode && !ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4385 struct buffer_head *bitmap_bh;
4386 int i, start;
4387
4388 start = inode_offset & ~(inodes_per_block - 1);
4389
4390 /* Is the inode bitmap in cache? */
4391 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4392 if (unlikely(!bitmap_bh))
4393 goto make_io;
4394
4395 /*
4396 * If the inode bitmap isn't in cache then the
4397 * optimisation may end up performing two reads instead
4398 * of one, so skip it.
4399 */
4400 if (!buffer_uptodate(bitmap_bh)) {
4401 brelse(bitmap_bh);
4402 goto make_io;
4403 }
4404 for (i = start; i < start + inodes_per_block; i++) {
4405 if (i == inode_offset)
4406 continue;
4407 if (ext4_test_bit(i, bitmap_bh->b_data))
4408 break;
4409 }
4410 brelse(bitmap_bh);
4411 if (i == start + inodes_per_block) {
4412 struct ext4_inode *raw_inode =
4413 (struct ext4_inode *) (bh->b_data + iloc->offset);
4414
4415 /* all other inodes are free, so skip I/O */
4416 memset(bh->b_data, 0, bh->b_size);
4417 if (!ext4_test_inode_state(inode, EXT4_STATE_NEW))
4418 ext4_fill_raw_inode(inode, raw_inode);
4419 set_buffer_uptodate(bh);
4420 unlock_buffer(bh);
4421 goto has_buffer;
4422 }
4423 }
4424
4425make_io:
4426 /*
4427 * If we need to do any I/O, try to pre-readahead extra
4428 * blocks from the inode table.
4429 */
4430 blk_start_plug(&plug);
4431 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4432 ext4_fsblk_t b, end, table;
4433 unsigned num;
4434 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4435
4436 table = ext4_inode_table(sb, gdp);
4437 /* s_inode_readahead_blks is always a power of 2 */
4438 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4439 if (table > b)
4440 b = table;
4441 end = b + ra_blks;
4442 num = EXT4_INODES_PER_GROUP(sb);
4443 if (ext4_has_group_desc_csum(sb))
4444 num -= ext4_itable_unused_count(sb, gdp);
4445 table += num / inodes_per_block;
4446 if (end > table)
4447 end = table;
4448 while (b <= end)
4449 ext4_sb_breadahead_unmovable(sb, b++);
4450 }
4451
4452 /*
4453 * There are other valid inodes in the buffer, this inode
4454 * has in-inode xattrs, or we don't have this inode in memory.
4455 * Read the block from disk.
4456 */
4457 trace_ext4_load_inode(sb, ino);
4458 ext4_read_bh_nowait(bh, REQ_META | REQ_PRIO, NULL);
4459 blk_finish_plug(&plug);
4460 wait_on_buffer(bh);
4461 ext4_simulate_fail_bh(sb, bh, EXT4_SIM_INODE_EIO);
4462 if (!buffer_uptodate(bh)) {
4463 if (ret_block)
4464 *ret_block = block;
4465 brelse(bh);
4466 return -EIO;
4467 }
4468has_buffer:
4469 iloc->bh = bh;
4470 return 0;
4471}
4472
4473static int __ext4_get_inode_loc_noinmem(struct inode *inode,
4474 struct ext4_iloc *iloc)
4475{
4476 ext4_fsblk_t err_blk = 0;
4477 int ret;
4478
4479 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, NULL, iloc,
4480 &err_blk);
4481
4482 if (ret == -EIO)
4483 ext4_error_inode_block(inode, err_blk, EIO,
4484 "unable to read itable block");
4485
4486 return ret;
4487}
4488
4489int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4490{
4491 ext4_fsblk_t err_blk = 0;
4492 int ret;
4493
4494 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, inode, iloc,
4495 &err_blk);
4496
4497 if (ret == -EIO)
4498 ext4_error_inode_block(inode, err_blk, EIO,
4499 "unable to read itable block");
4500
4501 return ret;
4502}
4503
4504
4505int ext4_get_fc_inode_loc(struct super_block *sb, unsigned long ino,
4506 struct ext4_iloc *iloc)
4507{
4508 return __ext4_get_inode_loc(sb, ino, NULL, iloc, NULL);
4509}
4510
4511static bool ext4_should_enable_dax(struct inode *inode)
4512{
4513 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4514
4515 if (test_opt2(inode->i_sb, DAX_NEVER))
4516 return false;
4517 if (!S_ISREG(inode->i_mode))
4518 return false;
4519 if (ext4_should_journal_data(inode))
4520 return false;
4521 if (ext4_has_inline_data(inode))
4522 return false;
4523 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4524 return false;
4525 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4526 return false;
4527 if (!test_bit(EXT4_FLAGS_BDEV_IS_DAX, &sbi->s_ext4_flags))
4528 return false;
4529 if (test_opt(inode->i_sb, DAX_ALWAYS))
4530 return true;
4531
4532 return ext4_test_inode_flag(inode, EXT4_INODE_DAX);
4533}
4534
4535void ext4_set_inode_flags(struct inode *inode, bool init)
4536{
4537 unsigned int flags = EXT4_I(inode)->i_flags;
4538 unsigned int new_fl = 0;
4539
4540 WARN_ON_ONCE(IS_DAX(inode) && init);
4541
4542 if (flags & EXT4_SYNC_FL)
4543 new_fl |= S_SYNC;
4544 if (flags & EXT4_APPEND_FL)
4545 new_fl |= S_APPEND;
4546 if (flags & EXT4_IMMUTABLE_FL)
4547 new_fl |= S_IMMUTABLE;
4548 if (flags & EXT4_NOATIME_FL)
4549 new_fl |= S_NOATIME;
4550 if (flags & EXT4_DIRSYNC_FL)
4551 new_fl |= S_DIRSYNC;
4552
4553 /* Because of the way inode_set_flags() works we must preserve S_DAX
4554 * here if already set. */
4555 new_fl |= (inode->i_flags & S_DAX);
4556 if (init && ext4_should_enable_dax(inode))
4557 new_fl |= S_DAX;
4558
4559 if (flags & EXT4_ENCRYPT_FL)
4560 new_fl |= S_ENCRYPTED;
4561 if (flags & EXT4_CASEFOLD_FL)
4562 new_fl |= S_CASEFOLD;
4563 if (flags & EXT4_VERITY_FL)
4564 new_fl |= S_VERITY;
4565 inode_set_flags(inode, new_fl,
4566 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4567 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4568}
4569
4570static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4571 struct ext4_inode_info *ei)
4572{
4573 blkcnt_t i_blocks ;
4574 struct inode *inode = &(ei->vfs_inode);
4575 struct super_block *sb = inode->i_sb;
4576
4577 if (ext4_has_feature_huge_file(sb)) {
4578 /* we are using combined 48 bit field */
4579 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4580 le32_to_cpu(raw_inode->i_blocks_lo);
4581 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4582 /* i_blocks represent file system block size */
4583 return i_blocks << (inode->i_blkbits - 9);
4584 } else {
4585 return i_blocks;
4586 }
4587 } else {
4588 return le32_to_cpu(raw_inode->i_blocks_lo);
4589 }
4590}
4591
4592static inline int ext4_iget_extra_inode(struct inode *inode,
4593 struct ext4_inode *raw_inode,
4594 struct ext4_inode_info *ei)
4595{
4596 __le32 *magic = (void *)raw_inode +
4597 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4598
4599 if (EXT4_INODE_HAS_XATTR_SPACE(inode) &&
4600 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4601 int err;
4602
4603 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4604 err = ext4_find_inline_data_nolock(inode);
4605 if (!err && ext4_has_inline_data(inode))
4606 ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
4607 return err;
4608 } else
4609 EXT4_I(inode)->i_inline_off = 0;
4610 return 0;
4611}
4612
4613int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4614{
4615 if (!ext4_has_feature_project(inode->i_sb))
4616 return -EOPNOTSUPP;
4617 *projid = EXT4_I(inode)->i_projid;
4618 return 0;
4619}
4620
4621/*
4622 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4623 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4624 * set.
4625 */
4626static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4627{
4628 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4629 inode_set_iversion_raw(inode, val);
4630 else
4631 inode_set_iversion_queried(inode, val);
4632}
4633
4634static const char *check_igot_inode(struct inode *inode, ext4_iget_flags flags)
4635
4636{
4637 if (flags & EXT4_IGET_EA_INODE) {
4638 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4639 return "missing EA_INODE flag";
4640 if (ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4641 EXT4_I(inode)->i_file_acl)
4642 return "ea_inode with extended attributes";
4643 } else {
4644 if ((EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4645 return "unexpected EA_INODE flag";
4646 }
4647 if (is_bad_inode(inode) && !(flags & EXT4_IGET_BAD))
4648 return "unexpected bad inode w/o EXT4_IGET_BAD";
4649 return NULL;
4650}
4651
4652struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4653 ext4_iget_flags flags, const char *function,
4654 unsigned int line)
4655{
4656 struct ext4_iloc iloc;
4657 struct ext4_inode *raw_inode;
4658 struct ext4_inode_info *ei;
4659 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
4660 struct inode *inode;
4661 const char *err_str;
4662 journal_t *journal = EXT4_SB(sb)->s_journal;
4663 long ret;
4664 loff_t size;
4665 int block;
4666 uid_t i_uid;
4667 gid_t i_gid;
4668 projid_t i_projid;
4669
4670 if ((!(flags & EXT4_IGET_SPECIAL) &&
4671 ((ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO) ||
4672 ino == le32_to_cpu(es->s_usr_quota_inum) ||
4673 ino == le32_to_cpu(es->s_grp_quota_inum) ||
4674 ino == le32_to_cpu(es->s_prj_quota_inum) ||
4675 ino == le32_to_cpu(es->s_orphan_file_inum))) ||
4676 (ino < EXT4_ROOT_INO) ||
4677 (ino > le32_to_cpu(es->s_inodes_count))) {
4678 if (flags & EXT4_IGET_HANDLE)
4679 return ERR_PTR(-ESTALE);
4680 __ext4_error(sb, function, line, false, EFSCORRUPTED, 0,
4681 "inode #%lu: comm %s: iget: illegal inode #",
4682 ino, current->comm);
4683 return ERR_PTR(-EFSCORRUPTED);
4684 }
4685
4686 inode = iget_locked(sb, ino);
4687 if (!inode)
4688 return ERR_PTR(-ENOMEM);
4689 if (!(inode->i_state & I_NEW)) {
4690 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
4691 ext4_error_inode(inode, function, line, 0, err_str);
4692 iput(inode);
4693 return ERR_PTR(-EFSCORRUPTED);
4694 }
4695 return inode;
4696 }
4697
4698 ei = EXT4_I(inode);
4699 iloc.bh = NULL;
4700
4701 ret = __ext4_get_inode_loc_noinmem(inode, &iloc);
4702 if (ret < 0)
4703 goto bad_inode;
4704 raw_inode = ext4_raw_inode(&iloc);
4705
4706 if ((flags & EXT4_IGET_HANDLE) &&
4707 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4708 ret = -ESTALE;
4709 goto bad_inode;
4710 }
4711
4712 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4713 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4714 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4715 EXT4_INODE_SIZE(inode->i_sb) ||
4716 (ei->i_extra_isize & 3)) {
4717 ext4_error_inode(inode, function, line, 0,
4718 "iget: bad extra_isize %u "
4719 "(inode size %u)",
4720 ei->i_extra_isize,
4721 EXT4_INODE_SIZE(inode->i_sb));
4722 ret = -EFSCORRUPTED;
4723 goto bad_inode;
4724 }
4725 } else
4726 ei->i_extra_isize = 0;
4727
4728 /* Precompute checksum seed for inode metadata */
4729 if (ext4_has_metadata_csum(sb)) {
4730 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4731 __u32 csum;
4732 __le32 inum = cpu_to_le32(inode->i_ino);
4733 __le32 gen = raw_inode->i_generation;
4734 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4735 sizeof(inum));
4736 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4737 sizeof(gen));
4738 }
4739
4740 if ((!ext4_inode_csum_verify(inode, raw_inode, ei) ||
4741 ext4_simulate_fail(sb, EXT4_SIM_INODE_CRC)) &&
4742 (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))) {
4743 ext4_error_inode_err(inode, function, line, 0,
4744 EFSBADCRC, "iget: checksum invalid");
4745 ret = -EFSBADCRC;
4746 goto bad_inode;
4747 }
4748
4749 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4750 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4751 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4752 if (ext4_has_feature_project(sb) &&
4753 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4754 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4755 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4756 else
4757 i_projid = EXT4_DEF_PROJID;
4758
4759 if (!(test_opt(inode->i_sb, NO_UID32))) {
4760 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4761 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4762 }
4763 i_uid_write(inode, i_uid);
4764 i_gid_write(inode, i_gid);
4765 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4766 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4767
4768 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4769 ei->i_inline_off = 0;
4770 ei->i_dir_start_lookup = 0;
4771 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4772 /* We now have enough fields to check if the inode was active or not.
4773 * This is needed because nfsd might try to access dead inodes
4774 * the test is that same one that e2fsck uses
4775 * NeilBrown 1999oct15
4776 */
4777 if (inode->i_nlink == 0) {
4778 if ((inode->i_mode == 0 || flags & EXT4_IGET_SPECIAL ||
4779 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4780 ino != EXT4_BOOT_LOADER_INO) {
4781 /* this inode is deleted or unallocated */
4782 if (flags & EXT4_IGET_SPECIAL) {
4783 ext4_error_inode(inode, function, line, 0,
4784 "iget: special inode unallocated");
4785 ret = -EFSCORRUPTED;
4786 } else
4787 ret = -ESTALE;
4788 goto bad_inode;
4789 }
4790 /* The only unlinked inodes we let through here have
4791 * valid i_mode and are being read by the orphan
4792 * recovery code: that's fine, we're about to complete
4793 * the process of deleting those.
4794 * OR it is the EXT4_BOOT_LOADER_INO which is
4795 * not initialized on a new filesystem. */
4796 }
4797 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4798 ext4_set_inode_flags(inode, true);
4799 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4800 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4801 if (ext4_has_feature_64bit(sb))
4802 ei->i_file_acl |=
4803 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4804 inode->i_size = ext4_isize(sb, raw_inode);
4805 if ((size = i_size_read(inode)) < 0) {
4806 ext4_error_inode(inode, function, line, 0,
4807 "iget: bad i_size value: %lld", size);
4808 ret = -EFSCORRUPTED;
4809 goto bad_inode;
4810 }
4811 /*
4812 * If dir_index is not enabled but there's dir with INDEX flag set,
4813 * we'd normally treat htree data as empty space. But with metadata
4814 * checksumming that corrupts checksums so forbid that.
4815 */
4816 if (!ext4_has_feature_dir_index(sb) && ext4_has_metadata_csum(sb) &&
4817 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
4818 ext4_error_inode(inode, function, line, 0,
4819 "iget: Dir with htree data on filesystem without dir_index feature.");
4820 ret = -EFSCORRUPTED;
4821 goto bad_inode;
4822 }
4823 ei->i_disksize = inode->i_size;
4824#ifdef CONFIG_QUOTA
4825 ei->i_reserved_quota = 0;
4826#endif
4827 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4828 ei->i_block_group = iloc.block_group;
4829 ei->i_last_alloc_group = ~0;
4830 /*
4831 * NOTE! The in-memory inode i_data array is in little-endian order
4832 * even on big-endian machines: we do NOT byteswap the block numbers!
4833 */
4834 for (block = 0; block < EXT4_N_BLOCKS; block++)
4835 ei->i_data[block] = raw_inode->i_block[block];
4836 INIT_LIST_HEAD(&ei->i_orphan);
4837 ext4_fc_init_inode(&ei->vfs_inode);
4838
4839 /*
4840 * Set transaction id's of transactions that have to be committed
4841 * to finish f[data]sync. We set them to currently running transaction
4842 * as we cannot be sure that the inode or some of its metadata isn't
4843 * part of the transaction - the inode could have been reclaimed and
4844 * now it is reread from disk.
4845 */
4846 if (journal) {
4847 transaction_t *transaction;
4848 tid_t tid;
4849
4850 read_lock(&journal->j_state_lock);
4851 if (journal->j_running_transaction)
4852 transaction = journal->j_running_transaction;
4853 else
4854 transaction = journal->j_committing_transaction;
4855 if (transaction)
4856 tid = transaction->t_tid;
4857 else
4858 tid = journal->j_commit_sequence;
4859 read_unlock(&journal->j_state_lock);
4860 ei->i_sync_tid = tid;
4861 ei->i_datasync_tid = tid;
4862 }
4863
4864 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4865 if (ei->i_extra_isize == 0) {
4866 /* The extra space is currently unused. Use it. */
4867 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4868 ei->i_extra_isize = sizeof(struct ext4_inode) -
4869 EXT4_GOOD_OLD_INODE_SIZE;
4870 } else {
4871 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
4872 if (ret)
4873 goto bad_inode;
4874 }
4875 }
4876
4877 EXT4_INODE_GET_CTIME(inode, raw_inode);
4878 EXT4_INODE_GET_ATIME(inode, raw_inode);
4879 EXT4_INODE_GET_MTIME(inode, raw_inode);
4880 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4881
4882 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4883 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4884
4885 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4886 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4887 ivers |=
4888 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4889 }
4890 ext4_inode_set_iversion_queried(inode, ivers);
4891 }
4892
4893 ret = 0;
4894 if (ei->i_file_acl &&
4895 !ext4_inode_block_valid(inode, ei->i_file_acl, 1)) {
4896 ext4_error_inode(inode, function, line, 0,
4897 "iget: bad extended attribute block %llu",
4898 ei->i_file_acl);
4899 ret = -EFSCORRUPTED;
4900 goto bad_inode;
4901 } else if (!ext4_has_inline_data(inode)) {
4902 /* validate the block references in the inode */
4903 if (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) &&
4904 (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4905 (S_ISLNK(inode->i_mode) &&
4906 !ext4_inode_is_fast_symlink(inode)))) {
4907 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4908 ret = ext4_ext_check_inode(inode);
4909 else
4910 ret = ext4_ind_check_inode(inode);
4911 }
4912 }
4913 if (ret)
4914 goto bad_inode;
4915
4916 if (S_ISREG(inode->i_mode)) {
4917 inode->i_op = &ext4_file_inode_operations;
4918 inode->i_fop = &ext4_file_operations;
4919 ext4_set_aops(inode);
4920 } else if (S_ISDIR(inode->i_mode)) {
4921 inode->i_op = &ext4_dir_inode_operations;
4922 inode->i_fop = &ext4_dir_operations;
4923 } else if (S_ISLNK(inode->i_mode)) {
4924 /* VFS does not allow setting these so must be corruption */
4925 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
4926 ext4_error_inode(inode, function, line, 0,
4927 "iget: immutable or append flags "
4928 "not allowed on symlinks");
4929 ret = -EFSCORRUPTED;
4930 goto bad_inode;
4931 }
4932 if (IS_ENCRYPTED(inode)) {
4933 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4934 } else if (ext4_inode_is_fast_symlink(inode)) {
4935 inode->i_link = (char *)ei->i_data;
4936 inode->i_op = &ext4_fast_symlink_inode_operations;
4937 nd_terminate_link(ei->i_data, inode->i_size,
4938 sizeof(ei->i_data) - 1);
4939 } else {
4940 inode->i_op = &ext4_symlink_inode_operations;
4941 }
4942 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4943 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4944 inode->i_op = &ext4_special_inode_operations;
4945 if (raw_inode->i_block[0])
4946 init_special_inode(inode, inode->i_mode,
4947 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4948 else
4949 init_special_inode(inode, inode->i_mode,
4950 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4951 } else if (ino == EXT4_BOOT_LOADER_INO) {
4952 make_bad_inode(inode);
4953 } else {
4954 ret = -EFSCORRUPTED;
4955 ext4_error_inode(inode, function, line, 0,
4956 "iget: bogus i_mode (%o)", inode->i_mode);
4957 goto bad_inode;
4958 }
4959 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb)) {
4960 ext4_error_inode(inode, function, line, 0,
4961 "casefold flag without casefold feature");
4962 ret = -EFSCORRUPTED;
4963 goto bad_inode;
4964 }
4965 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
4966 ext4_error_inode(inode, function, line, 0, err_str);
4967 ret = -EFSCORRUPTED;
4968 goto bad_inode;
4969 }
4970
4971 brelse(iloc.bh);
4972 unlock_new_inode(inode);
4973 return inode;
4974
4975bad_inode:
4976 brelse(iloc.bh);
4977 iget_failed(inode);
4978 return ERR_PTR(ret);
4979}
4980
4981static void __ext4_update_other_inode_time(struct super_block *sb,
4982 unsigned long orig_ino,
4983 unsigned long ino,
4984 struct ext4_inode *raw_inode)
4985{
4986 struct inode *inode;
4987
4988 inode = find_inode_by_ino_rcu(sb, ino);
4989 if (!inode)
4990 return;
4991
4992 if (!inode_is_dirtytime_only(inode))
4993 return;
4994
4995 spin_lock(&inode->i_lock);
4996 if (inode_is_dirtytime_only(inode)) {
4997 struct ext4_inode_info *ei = EXT4_I(inode);
4998
4999 inode->i_state &= ~I_DIRTY_TIME;
5000 spin_unlock(&inode->i_lock);
5001
5002 spin_lock(&ei->i_raw_lock);
5003 EXT4_INODE_SET_CTIME(inode, raw_inode);
5004 EXT4_INODE_SET_MTIME(inode, raw_inode);
5005 EXT4_INODE_SET_ATIME(inode, raw_inode);
5006 ext4_inode_csum_set(inode, raw_inode, ei);
5007 spin_unlock(&ei->i_raw_lock);
5008 trace_ext4_other_inode_update_time(inode, orig_ino);
5009 return;
5010 }
5011 spin_unlock(&inode->i_lock);
5012}
5013
5014/*
5015 * Opportunistically update the other time fields for other inodes in
5016 * the same inode table block.
5017 */
5018static void ext4_update_other_inodes_time(struct super_block *sb,
5019 unsigned long orig_ino, char *buf)
5020{
5021 unsigned long ino;
5022 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5023 int inode_size = EXT4_INODE_SIZE(sb);
5024
5025 /*
5026 * Calculate the first inode in the inode table block. Inode
5027 * numbers are one-based. That is, the first inode in a block
5028 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5029 */
5030 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5031 rcu_read_lock();
5032 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5033 if (ino == orig_ino)
5034 continue;
5035 __ext4_update_other_inode_time(sb, orig_ino, ino,
5036 (struct ext4_inode *)buf);
5037 }
5038 rcu_read_unlock();
5039}
5040
5041/*
5042 * Post the struct inode info into an on-disk inode location in the
5043 * buffer-cache. This gobbles the caller's reference to the
5044 * buffer_head in the inode location struct.
5045 *
5046 * The caller must have write access to iloc->bh.
5047 */
5048static int ext4_do_update_inode(handle_t *handle,
5049 struct inode *inode,
5050 struct ext4_iloc *iloc)
5051{
5052 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5053 struct ext4_inode_info *ei = EXT4_I(inode);
5054 struct buffer_head *bh = iloc->bh;
5055 struct super_block *sb = inode->i_sb;
5056 int err;
5057 int need_datasync = 0, set_large_file = 0;
5058
5059 spin_lock(&ei->i_raw_lock);
5060
5061 /*
5062 * For fields not tracked in the in-memory inode, initialise them
5063 * to zero for new inodes.
5064 */
5065 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5066 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5067
5068 if (READ_ONCE(ei->i_disksize) != ext4_isize(inode->i_sb, raw_inode))
5069 need_datasync = 1;
5070 if (ei->i_disksize > 0x7fffffffULL) {
5071 if (!ext4_has_feature_large_file(sb) ||
5072 EXT4_SB(sb)->s_es->s_rev_level == cpu_to_le32(EXT4_GOOD_OLD_REV))
5073 set_large_file = 1;
5074 }
5075
5076 err = ext4_fill_raw_inode(inode, raw_inode);
5077 spin_unlock(&ei->i_raw_lock);
5078 if (err) {
5079 EXT4_ERROR_INODE(inode, "corrupted inode contents");
5080 goto out_brelse;
5081 }
5082
5083 if (inode->i_sb->s_flags & SB_LAZYTIME)
5084 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5085 bh->b_data);
5086
5087 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5088 err = ext4_handle_dirty_metadata(handle, NULL, bh);
5089 if (err)
5090 goto out_error;
5091 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5092 if (set_large_file) {
5093 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5094 err = ext4_journal_get_write_access(handle, sb,
5095 EXT4_SB(sb)->s_sbh,
5096 EXT4_JTR_NONE);
5097 if (err)
5098 goto out_error;
5099 lock_buffer(EXT4_SB(sb)->s_sbh);
5100 ext4_set_feature_large_file(sb);
5101 ext4_superblock_csum_set(sb);
5102 unlock_buffer(EXT4_SB(sb)->s_sbh);
5103 ext4_handle_sync(handle);
5104 err = ext4_handle_dirty_metadata(handle, NULL,
5105 EXT4_SB(sb)->s_sbh);
5106 }
5107 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5108out_error:
5109 ext4_std_error(inode->i_sb, err);
5110out_brelse:
5111 brelse(bh);
5112 return err;
5113}
5114
5115/*
5116 * ext4_write_inode()
5117 *
5118 * We are called from a few places:
5119 *
5120 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5121 * Here, there will be no transaction running. We wait for any running
5122 * transaction to commit.
5123 *
5124 * - Within flush work (sys_sync(), kupdate and such).
5125 * We wait on commit, if told to.
5126 *
5127 * - Within iput_final() -> write_inode_now()
5128 * We wait on commit, if told to.
5129 *
5130 * In all cases it is actually safe for us to return without doing anything,
5131 * because the inode has been copied into a raw inode buffer in
5132 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5133 * writeback.
5134 *
5135 * Note that we are absolutely dependent upon all inode dirtiers doing the
5136 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5137 * which we are interested.
5138 *
5139 * It would be a bug for them to not do this. The code:
5140 *
5141 * mark_inode_dirty(inode)
5142 * stuff();
5143 * inode->i_size = expr;
5144 *
5145 * is in error because write_inode() could occur while `stuff()' is running,
5146 * and the new i_size will be lost. Plus the inode will no longer be on the
5147 * superblock's dirty inode list.
5148 */
5149int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5150{
5151 int err;
5152
5153 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5154 return 0;
5155
5156 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5157 return -EIO;
5158
5159 if (EXT4_SB(inode->i_sb)->s_journal) {
5160 if (ext4_journal_current_handle()) {
5161 ext4_debug("called recursively, non-PF_MEMALLOC!\n");
5162 dump_stack();
5163 return -EIO;
5164 }
5165
5166 /*
5167 * No need to force transaction in WB_SYNC_NONE mode. Also
5168 * ext4_sync_fs() will force the commit after everything is
5169 * written.
5170 */
5171 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5172 return 0;
5173
5174 err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
5175 EXT4_I(inode)->i_sync_tid);
5176 } else {
5177 struct ext4_iloc iloc;
5178
5179 err = __ext4_get_inode_loc_noinmem(inode, &iloc);
5180 if (err)
5181 return err;
5182 /*
5183 * sync(2) will flush the whole buffer cache. No need to do
5184 * it here separately for each inode.
5185 */
5186 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5187 sync_dirty_buffer(iloc.bh);
5188 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5189 ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
5190 "IO error syncing inode");
5191 err = -EIO;
5192 }
5193 brelse(iloc.bh);
5194 }
5195 return err;
5196}
5197
5198/*
5199 * In data=journal mode ext4_journalled_invalidate_folio() may fail to invalidate
5200 * buffers that are attached to a folio straddling i_size and are undergoing
5201 * commit. In that case we have to wait for commit to finish and try again.
5202 */
5203static void ext4_wait_for_tail_page_commit(struct inode *inode)
5204{
5205 unsigned offset;
5206 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5207 tid_t commit_tid = 0;
5208 int ret;
5209
5210 offset = inode->i_size & (PAGE_SIZE - 1);
5211 /*
5212 * If the folio is fully truncated, we don't need to wait for any commit
5213 * (and we even should not as __ext4_journalled_invalidate_folio() may
5214 * strip all buffers from the folio but keep the folio dirty which can then
5215 * confuse e.g. concurrent ext4_writepages() seeing dirty folio without
5216 * buffers). Also we don't need to wait for any commit if all buffers in
5217 * the folio remain valid. This is most beneficial for the common case of
5218 * blocksize == PAGESIZE.
5219 */
5220 if (!offset || offset > (PAGE_SIZE - i_blocksize(inode)))
5221 return;
5222 while (1) {
5223 struct folio *folio = filemap_lock_folio(inode->i_mapping,
5224 inode->i_size >> PAGE_SHIFT);
5225 if (IS_ERR(folio))
5226 return;
5227 ret = __ext4_journalled_invalidate_folio(folio, offset,
5228 folio_size(folio) - offset);
5229 folio_unlock(folio);
5230 folio_put(folio);
5231 if (ret != -EBUSY)
5232 return;
5233 commit_tid = 0;
5234 read_lock(&journal->j_state_lock);
5235 if (journal->j_committing_transaction)
5236 commit_tid = journal->j_committing_transaction->t_tid;
5237 read_unlock(&journal->j_state_lock);
5238 if (commit_tid)
5239 jbd2_log_wait_commit(journal, commit_tid);
5240 }
5241}
5242
5243/*
5244 * ext4_setattr()
5245 *
5246 * Called from notify_change.
5247 *
5248 * We want to trap VFS attempts to truncate the file as soon as
5249 * possible. In particular, we want to make sure that when the VFS
5250 * shrinks i_size, we put the inode on the orphan list and modify
5251 * i_disksize immediately, so that during the subsequent flushing of
5252 * dirty pages and freeing of disk blocks, we can guarantee that any
5253 * commit will leave the blocks being flushed in an unused state on
5254 * disk. (On recovery, the inode will get truncated and the blocks will
5255 * be freed, so we have a strong guarantee that no future commit will
5256 * leave these blocks visible to the user.)
5257 *
5258 * Another thing we have to assure is that if we are in ordered mode
5259 * and inode is still attached to the committing transaction, we must
5260 * we start writeout of all the dirty pages which are being truncated.
5261 * This way we are sure that all the data written in the previous
5262 * transaction are already on disk (truncate waits for pages under
5263 * writeback).
5264 *
5265 * Called with inode->i_rwsem down.
5266 */
5267int ext4_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
5268 struct iattr *attr)
5269{
5270 struct inode *inode = d_inode(dentry);
5271 int error, rc = 0;
5272 int orphan = 0;
5273 const unsigned int ia_valid = attr->ia_valid;
5274 bool inc_ivers = true;
5275
5276 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5277 return -EIO;
5278
5279 if (unlikely(IS_IMMUTABLE(inode)))
5280 return -EPERM;
5281
5282 if (unlikely(IS_APPEND(inode) &&
5283 (ia_valid & (ATTR_MODE | ATTR_UID |
5284 ATTR_GID | ATTR_TIMES_SET))))
5285 return -EPERM;
5286
5287 error = setattr_prepare(idmap, dentry, attr);
5288 if (error)
5289 return error;
5290
5291 error = fscrypt_prepare_setattr(dentry, attr);
5292 if (error)
5293 return error;
5294
5295 error = fsverity_prepare_setattr(dentry, attr);
5296 if (error)
5297 return error;
5298
5299 if (is_quota_modification(idmap, inode, attr)) {
5300 error = dquot_initialize(inode);
5301 if (error)
5302 return error;
5303 }
5304
5305 if (i_uid_needs_update(idmap, attr, inode) ||
5306 i_gid_needs_update(idmap, attr, inode)) {
5307 handle_t *handle;
5308
5309 /* (user+group)*(old+new) structure, inode write (sb,
5310 * inode block, ? - but truncate inode update has it) */
5311 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5312 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5313 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5314 if (IS_ERR(handle)) {
5315 error = PTR_ERR(handle);
5316 goto err_out;
5317 }
5318
5319 /* dquot_transfer() calls back ext4_get_inode_usage() which
5320 * counts xattr inode references.
5321 */
5322 down_read(&EXT4_I(inode)->xattr_sem);
5323 error = dquot_transfer(idmap, inode, attr);
5324 up_read(&EXT4_I(inode)->xattr_sem);
5325
5326 if (error) {
5327 ext4_journal_stop(handle);
5328 return error;
5329 }
5330 /* Update corresponding info in inode so that everything is in
5331 * one transaction */
5332 i_uid_update(idmap, attr, inode);
5333 i_gid_update(idmap, attr, inode);
5334 error = ext4_mark_inode_dirty(handle, inode);
5335 ext4_journal_stop(handle);
5336 if (unlikely(error)) {
5337 return error;
5338 }
5339 }
5340
5341 if (attr->ia_valid & ATTR_SIZE) {
5342 handle_t *handle;
5343 loff_t oldsize = inode->i_size;
5344 loff_t old_disksize;
5345 int shrink = (attr->ia_size < inode->i_size);
5346
5347 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5348 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5349
5350 if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5351 return -EFBIG;
5352 }
5353 }
5354 if (!S_ISREG(inode->i_mode)) {
5355 return -EINVAL;
5356 }
5357
5358 if (attr->ia_size == inode->i_size)
5359 inc_ivers = false;
5360
5361 if (shrink) {
5362 if (ext4_should_order_data(inode)) {
5363 error = ext4_begin_ordered_truncate(inode,
5364 attr->ia_size);
5365 if (error)
5366 goto err_out;
5367 }
5368 /*
5369 * Blocks are going to be removed from the inode. Wait
5370 * for dio in flight.
5371 */
5372 inode_dio_wait(inode);
5373 }
5374
5375 filemap_invalidate_lock(inode->i_mapping);
5376
5377 rc = ext4_break_layouts(inode);
5378 if (rc) {
5379 filemap_invalidate_unlock(inode->i_mapping);
5380 goto err_out;
5381 }
5382
5383 if (attr->ia_size != inode->i_size) {
5384 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5385 if (IS_ERR(handle)) {
5386 error = PTR_ERR(handle);
5387 goto out_mmap_sem;
5388 }
5389 if (ext4_handle_valid(handle) && shrink) {
5390 error = ext4_orphan_add(handle, inode);
5391 orphan = 1;
5392 }
5393 /*
5394 * Update c/mtime on truncate up, ext4_truncate() will
5395 * update c/mtime in shrink case below
5396 */
5397 if (!shrink)
5398 inode_set_mtime_to_ts(inode,
5399 inode_set_ctime_current(inode));
5400
5401 if (shrink)
5402 ext4_fc_track_range(handle, inode,
5403 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5404 inode->i_sb->s_blocksize_bits,
5405 EXT_MAX_BLOCKS - 1);
5406 else
5407 ext4_fc_track_range(
5408 handle, inode,
5409 (oldsize > 0 ? oldsize - 1 : oldsize) >>
5410 inode->i_sb->s_blocksize_bits,
5411 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5412 inode->i_sb->s_blocksize_bits);
5413
5414 down_write(&EXT4_I(inode)->i_data_sem);
5415 old_disksize = EXT4_I(inode)->i_disksize;
5416 EXT4_I(inode)->i_disksize = attr->ia_size;
5417 rc = ext4_mark_inode_dirty(handle, inode);
5418 if (!error)
5419 error = rc;
5420 /*
5421 * We have to update i_size under i_data_sem together
5422 * with i_disksize to avoid races with writeback code
5423 * running ext4_wb_update_i_disksize().
5424 */
5425 if (!error)
5426 i_size_write(inode, attr->ia_size);
5427 else
5428 EXT4_I(inode)->i_disksize = old_disksize;
5429 up_write(&EXT4_I(inode)->i_data_sem);
5430 ext4_journal_stop(handle);
5431 if (error)
5432 goto out_mmap_sem;
5433 if (!shrink) {
5434 pagecache_isize_extended(inode, oldsize,
5435 inode->i_size);
5436 } else if (ext4_should_journal_data(inode)) {
5437 ext4_wait_for_tail_page_commit(inode);
5438 }
5439 }
5440
5441 /*
5442 * Truncate pagecache after we've waited for commit
5443 * in data=journal mode to make pages freeable.
5444 */
5445 truncate_pagecache(inode, inode->i_size);
5446 /*
5447 * Call ext4_truncate() even if i_size didn't change to
5448 * truncate possible preallocated blocks.
5449 */
5450 if (attr->ia_size <= oldsize) {
5451 rc = ext4_truncate(inode);
5452 if (rc)
5453 error = rc;
5454 }
5455out_mmap_sem:
5456 filemap_invalidate_unlock(inode->i_mapping);
5457 }
5458
5459 if (!error) {
5460 if (inc_ivers)
5461 inode_inc_iversion(inode);
5462 setattr_copy(idmap, inode, attr);
5463 mark_inode_dirty(inode);
5464 }
5465
5466 /*
5467 * If the call to ext4_truncate failed to get a transaction handle at
5468 * all, we need to clean up the in-core orphan list manually.
5469 */
5470 if (orphan && inode->i_nlink)
5471 ext4_orphan_del(NULL, inode);
5472
5473 if (!error && (ia_valid & ATTR_MODE))
5474 rc = posix_acl_chmod(idmap, dentry, inode->i_mode);
5475
5476err_out:
5477 if (error)
5478 ext4_std_error(inode->i_sb, error);
5479 if (!error)
5480 error = rc;
5481 return error;
5482}
5483
5484u32 ext4_dio_alignment(struct inode *inode)
5485{
5486 if (fsverity_active(inode))
5487 return 0;
5488 if (ext4_should_journal_data(inode))
5489 return 0;
5490 if (ext4_has_inline_data(inode))
5491 return 0;
5492 if (IS_ENCRYPTED(inode)) {
5493 if (!fscrypt_dio_supported(inode))
5494 return 0;
5495 return i_blocksize(inode);
5496 }
5497 return 1; /* use the iomap defaults */
5498}
5499
5500int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
5501 struct kstat *stat, u32 request_mask, unsigned int query_flags)
5502{
5503 struct inode *inode = d_inode(path->dentry);
5504 struct ext4_inode *raw_inode;
5505 struct ext4_inode_info *ei = EXT4_I(inode);
5506 unsigned int flags;
5507
5508 if ((request_mask & STATX_BTIME) &&
5509 EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5510 stat->result_mask |= STATX_BTIME;
5511 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5512 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5513 }
5514
5515 /*
5516 * Return the DIO alignment restrictions if requested. We only return
5517 * this information when requested, since on encrypted files it might
5518 * take a fair bit of work to get if the file wasn't opened recently.
5519 */
5520 if ((request_mask & STATX_DIOALIGN) && S_ISREG(inode->i_mode)) {
5521 u32 dio_align = ext4_dio_alignment(inode);
5522
5523 stat->result_mask |= STATX_DIOALIGN;
5524 if (dio_align == 1) {
5525 struct block_device *bdev = inode->i_sb->s_bdev;
5526
5527 /* iomap defaults */
5528 stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
5529 stat->dio_offset_align = bdev_logical_block_size(bdev);
5530 } else {
5531 stat->dio_mem_align = dio_align;
5532 stat->dio_offset_align = dio_align;
5533 }
5534 }
5535
5536 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5537 if (flags & EXT4_APPEND_FL)
5538 stat->attributes |= STATX_ATTR_APPEND;
5539 if (flags & EXT4_COMPR_FL)
5540 stat->attributes |= STATX_ATTR_COMPRESSED;
5541 if (flags & EXT4_ENCRYPT_FL)
5542 stat->attributes |= STATX_ATTR_ENCRYPTED;
5543 if (flags & EXT4_IMMUTABLE_FL)
5544 stat->attributes |= STATX_ATTR_IMMUTABLE;
5545 if (flags & EXT4_NODUMP_FL)
5546 stat->attributes |= STATX_ATTR_NODUMP;
5547 if (flags & EXT4_VERITY_FL)
5548 stat->attributes |= STATX_ATTR_VERITY;
5549
5550 stat->attributes_mask |= (STATX_ATTR_APPEND |
5551 STATX_ATTR_COMPRESSED |
5552 STATX_ATTR_ENCRYPTED |
5553 STATX_ATTR_IMMUTABLE |
5554 STATX_ATTR_NODUMP |
5555 STATX_ATTR_VERITY);
5556
5557 generic_fillattr(idmap, request_mask, inode, stat);
5558 return 0;
5559}
5560
5561int ext4_file_getattr(struct mnt_idmap *idmap,
5562 const struct path *path, struct kstat *stat,
5563 u32 request_mask, unsigned int query_flags)
5564{
5565 struct inode *inode = d_inode(path->dentry);
5566 u64 delalloc_blocks;
5567
5568 ext4_getattr(idmap, path, stat, request_mask, query_flags);
5569
5570 /*
5571 * If there is inline data in the inode, the inode will normally not
5572 * have data blocks allocated (it may have an external xattr block).
5573 * Report at least one sector for such files, so tools like tar, rsync,
5574 * others don't incorrectly think the file is completely sparse.
5575 */
5576 if (unlikely(ext4_has_inline_data(inode)))
5577 stat->blocks += (stat->size + 511) >> 9;
5578
5579 /*
5580 * We can't update i_blocks if the block allocation is delayed
5581 * otherwise in the case of system crash before the real block
5582 * allocation is done, we will have i_blocks inconsistent with
5583 * on-disk file blocks.
5584 * We always keep i_blocks updated together with real
5585 * allocation. But to not confuse with user, stat
5586 * will return the blocks that include the delayed allocation
5587 * blocks for this file.
5588 */
5589 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5590 EXT4_I(inode)->i_reserved_data_blocks);
5591 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5592 return 0;
5593}
5594
5595static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5596 int pextents)
5597{
5598 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5599 return ext4_ind_trans_blocks(inode, lblocks);
5600 return ext4_ext_index_trans_blocks(inode, pextents);
5601}
5602
5603/*
5604 * Account for index blocks, block groups bitmaps and block group
5605 * descriptor blocks if modify datablocks and index blocks
5606 * worse case, the indexs blocks spread over different block groups
5607 *
5608 * If datablocks are discontiguous, they are possible to spread over
5609 * different block groups too. If they are contiguous, with flexbg,
5610 * they could still across block group boundary.
5611 *
5612 * Also account for superblock, inode, quota and xattr blocks
5613 */
5614static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5615 int pextents)
5616{
5617 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5618 int gdpblocks;
5619 int idxblocks;
5620 int ret;
5621
5622 /*
5623 * How many index blocks need to touch to map @lblocks logical blocks
5624 * to @pextents physical extents?
5625 */
5626 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5627
5628 ret = idxblocks;
5629
5630 /*
5631 * Now let's see how many group bitmaps and group descriptors need
5632 * to account
5633 */
5634 groups = idxblocks + pextents;
5635 gdpblocks = groups;
5636 if (groups > ngroups)
5637 groups = ngroups;
5638 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5639 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5640
5641 /* bitmaps and block group descriptor blocks */
5642 ret += groups + gdpblocks;
5643
5644 /* Blocks for super block, inode, quota and xattr blocks */
5645 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5646
5647 return ret;
5648}
5649
5650/*
5651 * Calculate the total number of credits to reserve to fit
5652 * the modification of a single pages into a single transaction,
5653 * which may include multiple chunks of block allocations.
5654 *
5655 * This could be called via ext4_write_begin()
5656 *
5657 * We need to consider the worse case, when
5658 * one new block per extent.
5659 */
5660int ext4_writepage_trans_blocks(struct inode *inode)
5661{
5662 int bpp = ext4_journal_blocks_per_page(inode);
5663 int ret;
5664
5665 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5666
5667 /* Account for data blocks for journalled mode */
5668 if (ext4_should_journal_data(inode))
5669 ret += bpp;
5670 return ret;
5671}
5672
5673/*
5674 * Calculate the journal credits for a chunk of data modification.
5675 *
5676 * This is called from DIO, fallocate or whoever calling
5677 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5678 *
5679 * journal buffers for data blocks are not included here, as DIO
5680 * and fallocate do no need to journal data buffers.
5681 */
5682int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5683{
5684 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5685}
5686
5687/*
5688 * The caller must have previously called ext4_reserve_inode_write().
5689 * Give this, we know that the caller already has write access to iloc->bh.
5690 */
5691int ext4_mark_iloc_dirty(handle_t *handle,
5692 struct inode *inode, struct ext4_iloc *iloc)
5693{
5694 int err = 0;
5695
5696 if (unlikely(ext4_forced_shutdown(inode->i_sb))) {
5697 put_bh(iloc->bh);
5698 return -EIO;
5699 }
5700 ext4_fc_track_inode(handle, inode);
5701
5702 /* the do_update_inode consumes one bh->b_count */
5703 get_bh(iloc->bh);
5704
5705 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5706 err = ext4_do_update_inode(handle, inode, iloc);
5707 put_bh(iloc->bh);
5708 return err;
5709}
5710
5711/*
5712 * On success, We end up with an outstanding reference count against
5713 * iloc->bh. This _must_ be cleaned up later.
5714 */
5715
5716int
5717ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5718 struct ext4_iloc *iloc)
5719{
5720 int err;
5721
5722 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5723 return -EIO;
5724
5725 err = ext4_get_inode_loc(inode, iloc);
5726 if (!err) {
5727 BUFFER_TRACE(iloc->bh, "get_write_access");
5728 err = ext4_journal_get_write_access(handle, inode->i_sb,
5729 iloc->bh, EXT4_JTR_NONE);
5730 if (err) {
5731 brelse(iloc->bh);
5732 iloc->bh = NULL;
5733 }
5734 }
5735 ext4_std_error(inode->i_sb, err);
5736 return err;
5737}
5738
5739static int __ext4_expand_extra_isize(struct inode *inode,
5740 unsigned int new_extra_isize,
5741 struct ext4_iloc *iloc,
5742 handle_t *handle, int *no_expand)
5743{
5744 struct ext4_inode *raw_inode;
5745 struct ext4_xattr_ibody_header *header;
5746 unsigned int inode_size = EXT4_INODE_SIZE(inode->i_sb);
5747 struct ext4_inode_info *ei = EXT4_I(inode);
5748 int error;
5749
5750 /* this was checked at iget time, but double check for good measure */
5751 if ((EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize > inode_size) ||
5752 (ei->i_extra_isize & 3)) {
5753 EXT4_ERROR_INODE(inode, "bad extra_isize %u (inode size %u)",
5754 ei->i_extra_isize,
5755 EXT4_INODE_SIZE(inode->i_sb));
5756 return -EFSCORRUPTED;
5757 }
5758 if ((new_extra_isize < ei->i_extra_isize) ||
5759 (new_extra_isize < 4) ||
5760 (new_extra_isize > inode_size - EXT4_GOOD_OLD_INODE_SIZE))
5761 return -EINVAL; /* Should never happen */
5762
5763 raw_inode = ext4_raw_inode(iloc);
5764
5765 header = IHDR(inode, raw_inode);
5766
5767 /* No extended attributes present */
5768 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5769 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5770 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5771 EXT4_I(inode)->i_extra_isize, 0,
5772 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5773 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5774 return 0;
5775 }
5776
5777 /*
5778 * We may need to allocate external xattr block so we need quotas
5779 * initialized. Here we can be called with various locks held so we
5780 * cannot affort to initialize quotas ourselves. So just bail.
5781 */
5782 if (dquot_initialize_needed(inode))
5783 return -EAGAIN;
5784
5785 /* try to expand with EAs present */
5786 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5787 raw_inode, handle);
5788 if (error) {
5789 /*
5790 * Inode size expansion failed; don't try again
5791 */
5792 *no_expand = 1;
5793 }
5794
5795 return error;
5796}
5797
5798/*
5799 * Expand an inode by new_extra_isize bytes.
5800 * Returns 0 on success or negative error number on failure.
5801 */
5802static int ext4_try_to_expand_extra_isize(struct inode *inode,
5803 unsigned int new_extra_isize,
5804 struct ext4_iloc iloc,
5805 handle_t *handle)
5806{
5807 int no_expand;
5808 int error;
5809
5810 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5811 return -EOVERFLOW;
5812
5813 /*
5814 * In nojournal mode, we can immediately attempt to expand
5815 * the inode. When journaled, we first need to obtain extra
5816 * buffer credits since we may write into the EA block
5817 * with this same handle. If journal_extend fails, then it will
5818 * only result in a minor loss of functionality for that inode.
5819 * If this is felt to be critical, then e2fsck should be run to
5820 * force a large enough s_min_extra_isize.
5821 */
5822 if (ext4_journal_extend(handle,
5823 EXT4_DATA_TRANS_BLOCKS(inode->i_sb), 0) != 0)
5824 return -ENOSPC;
5825
5826 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5827 return -EBUSY;
5828
5829 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5830 handle, &no_expand);
5831 ext4_write_unlock_xattr(inode, &no_expand);
5832
5833 return error;
5834}
5835
5836int ext4_expand_extra_isize(struct inode *inode,
5837 unsigned int new_extra_isize,
5838 struct ext4_iloc *iloc)
5839{
5840 handle_t *handle;
5841 int no_expand;
5842 int error, rc;
5843
5844 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5845 brelse(iloc->bh);
5846 return -EOVERFLOW;
5847 }
5848
5849 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5850 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5851 if (IS_ERR(handle)) {
5852 error = PTR_ERR(handle);
5853 brelse(iloc->bh);
5854 return error;
5855 }
5856
5857 ext4_write_lock_xattr(inode, &no_expand);
5858
5859 BUFFER_TRACE(iloc->bh, "get_write_access");
5860 error = ext4_journal_get_write_access(handle, inode->i_sb, iloc->bh,
5861 EXT4_JTR_NONE);
5862 if (error) {
5863 brelse(iloc->bh);
5864 goto out_unlock;
5865 }
5866
5867 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5868 handle, &no_expand);
5869
5870 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5871 if (!error)
5872 error = rc;
5873
5874out_unlock:
5875 ext4_write_unlock_xattr(inode, &no_expand);
5876 ext4_journal_stop(handle);
5877 return error;
5878}
5879
5880/*
5881 * What we do here is to mark the in-core inode as clean with respect to inode
5882 * dirtiness (it may still be data-dirty).
5883 * This means that the in-core inode may be reaped by prune_icache
5884 * without having to perform any I/O. This is a very good thing,
5885 * because *any* task may call prune_icache - even ones which
5886 * have a transaction open against a different journal.
5887 *
5888 * Is this cheating? Not really. Sure, we haven't written the
5889 * inode out, but prune_icache isn't a user-visible syncing function.
5890 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5891 * we start and wait on commits.
5892 */
5893int __ext4_mark_inode_dirty(handle_t *handle, struct inode *inode,
5894 const char *func, unsigned int line)
5895{
5896 struct ext4_iloc iloc;
5897 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5898 int err;
5899
5900 might_sleep();
5901 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5902 err = ext4_reserve_inode_write(handle, inode, &iloc);
5903 if (err)
5904 goto out;
5905
5906 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
5907 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
5908 iloc, handle);
5909
5910 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5911out:
5912 if (unlikely(err))
5913 ext4_error_inode_err(inode, func, line, 0, err,
5914 "mark_inode_dirty error");
5915 return err;
5916}
5917
5918/*
5919 * ext4_dirty_inode() is called from __mark_inode_dirty()
5920 *
5921 * We're really interested in the case where a file is being extended.
5922 * i_size has been changed by generic_commit_write() and we thus need
5923 * to include the updated inode in the current transaction.
5924 *
5925 * Also, dquot_alloc_block() will always dirty the inode when blocks
5926 * are allocated to the file.
5927 *
5928 * If the inode is marked synchronous, we don't honour that here - doing
5929 * so would cause a commit on atime updates, which we don't bother doing.
5930 * We handle synchronous inodes at the highest possible level.
5931 */
5932void ext4_dirty_inode(struct inode *inode, int flags)
5933{
5934 handle_t *handle;
5935
5936 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5937 if (IS_ERR(handle))
5938 return;
5939 ext4_mark_inode_dirty(handle, inode);
5940 ext4_journal_stop(handle);
5941}
5942
5943int ext4_change_inode_journal_flag(struct inode *inode, int val)
5944{
5945 journal_t *journal;
5946 handle_t *handle;
5947 int err;
5948 int alloc_ctx;
5949
5950 /*
5951 * We have to be very careful here: changing a data block's
5952 * journaling status dynamically is dangerous. If we write a
5953 * data block to the journal, change the status and then delete
5954 * that block, we risk forgetting to revoke the old log record
5955 * from the journal and so a subsequent replay can corrupt data.
5956 * So, first we make sure that the journal is empty and that
5957 * nobody is changing anything.
5958 */
5959
5960 journal = EXT4_JOURNAL(inode);
5961 if (!journal)
5962 return 0;
5963 if (is_journal_aborted(journal))
5964 return -EROFS;
5965
5966 /* Wait for all existing dio workers */
5967 inode_dio_wait(inode);
5968
5969 /*
5970 * Before flushing the journal and switching inode's aops, we have
5971 * to flush all dirty data the inode has. There can be outstanding
5972 * delayed allocations, there can be unwritten extents created by
5973 * fallocate or buffered writes in dioread_nolock mode covered by
5974 * dirty data which can be converted only after flushing the dirty
5975 * data (and journalled aops don't know how to handle these cases).
5976 */
5977 if (val) {
5978 filemap_invalidate_lock(inode->i_mapping);
5979 err = filemap_write_and_wait(inode->i_mapping);
5980 if (err < 0) {
5981 filemap_invalidate_unlock(inode->i_mapping);
5982 return err;
5983 }
5984 }
5985
5986 alloc_ctx = ext4_writepages_down_write(inode->i_sb);
5987 jbd2_journal_lock_updates(journal);
5988
5989 /*
5990 * OK, there are no updates running now, and all cached data is
5991 * synced to disk. We are now in a completely consistent state
5992 * which doesn't have anything in the journal, and we know that
5993 * no filesystem updates are running, so it is safe to modify
5994 * the inode's in-core data-journaling state flag now.
5995 */
5996
5997 if (val)
5998 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5999 else {
6000 err = jbd2_journal_flush(journal, 0);
6001 if (err < 0) {
6002 jbd2_journal_unlock_updates(journal);
6003 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6004 return err;
6005 }
6006 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6007 }
6008 ext4_set_aops(inode);
6009
6010 jbd2_journal_unlock_updates(journal);
6011 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6012
6013 if (val)
6014 filemap_invalidate_unlock(inode->i_mapping);
6015
6016 /* Finally we can mark the inode as dirty. */
6017
6018 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6019 if (IS_ERR(handle))
6020 return PTR_ERR(handle);
6021
6022 ext4_fc_mark_ineligible(inode->i_sb,
6023 EXT4_FC_REASON_JOURNAL_FLAG_CHANGE, handle);
6024 err = ext4_mark_inode_dirty(handle, inode);
6025 ext4_handle_sync(handle);
6026 ext4_journal_stop(handle);
6027 ext4_std_error(inode->i_sb, err);
6028
6029 return err;
6030}
6031
6032static int ext4_bh_unmapped(handle_t *handle, struct inode *inode,
6033 struct buffer_head *bh)
6034{
6035 return !buffer_mapped(bh);
6036}
6037
6038vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
6039{
6040 struct vm_area_struct *vma = vmf->vma;
6041 struct folio *folio = page_folio(vmf->page);
6042 loff_t size;
6043 unsigned long len;
6044 int err;
6045 vm_fault_t ret;
6046 struct file *file = vma->vm_file;
6047 struct inode *inode = file_inode(file);
6048 struct address_space *mapping = inode->i_mapping;
6049 handle_t *handle;
6050 get_block_t *get_block;
6051 int retries = 0;
6052
6053 if (unlikely(IS_IMMUTABLE(inode)))
6054 return VM_FAULT_SIGBUS;
6055
6056 sb_start_pagefault(inode->i_sb);
6057 file_update_time(vma->vm_file);
6058
6059 filemap_invalidate_lock_shared(mapping);
6060
6061 err = ext4_convert_inline_data(inode);
6062 if (err)
6063 goto out_ret;
6064
6065 /*
6066 * On data journalling we skip straight to the transaction handle:
6067 * there's no delalloc; page truncated will be checked later; the
6068 * early return w/ all buffers mapped (calculates size/len) can't
6069 * be used; and there's no dioread_nolock, so only ext4_get_block.
6070 */
6071 if (ext4_should_journal_data(inode))
6072 goto retry_alloc;
6073
6074 /* Delalloc case is easy... */
6075 if (test_opt(inode->i_sb, DELALLOC) &&
6076 !ext4_nonda_switch(inode->i_sb)) {
6077 do {
6078 err = block_page_mkwrite(vma, vmf,
6079 ext4_da_get_block_prep);
6080 } while (err == -ENOSPC &&
6081 ext4_should_retry_alloc(inode->i_sb, &retries));
6082 goto out_ret;
6083 }
6084
6085 folio_lock(folio);
6086 size = i_size_read(inode);
6087 /* Page got truncated from under us? */
6088 if (folio->mapping != mapping || folio_pos(folio) > size) {
6089 folio_unlock(folio);
6090 ret = VM_FAULT_NOPAGE;
6091 goto out;
6092 }
6093
6094 len = folio_size(folio);
6095 if (folio_pos(folio) + len > size)
6096 len = size - folio_pos(folio);
6097 /*
6098 * Return if we have all the buffers mapped. This avoids the need to do
6099 * journal_start/journal_stop which can block and take a long time
6100 *
6101 * This cannot be done for data journalling, as we have to add the
6102 * inode to the transaction's list to writeprotect pages on commit.
6103 */
6104 if (folio_buffers(folio)) {
6105 if (!ext4_walk_page_buffers(NULL, inode, folio_buffers(folio),
6106 0, len, NULL,
6107 ext4_bh_unmapped)) {
6108 /* Wait so that we don't change page under IO */
6109 folio_wait_stable(folio);
6110 ret = VM_FAULT_LOCKED;
6111 goto out;
6112 }
6113 }
6114 folio_unlock(folio);
6115 /* OK, we need to fill the hole... */
6116 if (ext4_should_dioread_nolock(inode))
6117 get_block = ext4_get_block_unwritten;
6118 else
6119 get_block = ext4_get_block;
6120retry_alloc:
6121 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6122 ext4_writepage_trans_blocks(inode));
6123 if (IS_ERR(handle)) {
6124 ret = VM_FAULT_SIGBUS;
6125 goto out;
6126 }
6127 /*
6128 * Data journalling can't use block_page_mkwrite() because it
6129 * will set_buffer_dirty() before do_journal_get_write_access()
6130 * thus might hit warning messages for dirty metadata buffers.
6131 */
6132 if (!ext4_should_journal_data(inode)) {
6133 err = block_page_mkwrite(vma, vmf, get_block);
6134 } else {
6135 folio_lock(folio);
6136 size = i_size_read(inode);
6137 /* Page got truncated from under us? */
6138 if (folio->mapping != mapping || folio_pos(folio) > size) {
6139 ret = VM_FAULT_NOPAGE;
6140 goto out_error;
6141 }
6142
6143 len = folio_size(folio);
6144 if (folio_pos(folio) + len > size)
6145 len = size - folio_pos(folio);
6146
6147 err = __block_write_begin(&folio->page, 0, len, ext4_get_block);
6148 if (!err) {
6149 ret = VM_FAULT_SIGBUS;
6150 if (ext4_journal_folio_buffers(handle, folio, len))
6151 goto out_error;
6152 } else {
6153 folio_unlock(folio);
6154 }
6155 }
6156 ext4_journal_stop(handle);
6157 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6158 goto retry_alloc;
6159out_ret:
6160 ret = vmf_fs_error(err);
6161out:
6162 filemap_invalidate_unlock_shared(mapping);
6163 sb_end_pagefault(inode->i_sb);
6164 return ret;
6165out_error:
6166 folio_unlock(folio);
6167 ext4_journal_stop(handle);
6168 goto out;
6169}