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/fs.h>
22#include <linux/time.h>
23#include <linux/highuid.h>
24#include <linux/pagemap.h>
25#include <linux/dax.h>
26#include <linux/quotaops.h>
27#include <linux/string.h>
28#include <linux/buffer_head.h>
29#include <linux/writeback.h>
30#include <linux/pagevec.h>
31#include <linux/mpage.h>
32#include <linux/namei.h>
33#include <linux/uio.h>
34#include <linux/bio.h>
35#include <linux/workqueue.h>
36#include <linux/kernel.h>
37#include <linux/printk.h>
38#include <linux/slab.h>
39#include <linux/bitops.h>
40#include <linux/iomap.h>
41
42#include "ext4_jbd2.h"
43#include "xattr.h"
44#include "acl.h"
45#include "truncate.h"
46
47#include <trace/events/ext4.h>
48
49#define MPAGE_DA_EXTENT_TAIL 0x01
50
51static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
52 struct ext4_inode_info *ei)
53{
54 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
55 __u32 csum;
56 __u16 dummy_csum = 0;
57 int offset = offsetof(struct ext4_inode, i_checksum_lo);
58 unsigned int csum_size = sizeof(dummy_csum);
59
60 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
61 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
62 offset += csum_size;
63 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
64 EXT4_GOOD_OLD_INODE_SIZE - offset);
65
66 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
67 offset = offsetof(struct ext4_inode, i_checksum_hi);
68 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
69 EXT4_GOOD_OLD_INODE_SIZE,
70 offset - EXT4_GOOD_OLD_INODE_SIZE);
71 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
72 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
73 csum_size);
74 offset += csum_size;
75 }
76 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
77 EXT4_INODE_SIZE(inode->i_sb) - offset);
78 }
79
80 return csum;
81}
82
83static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
84 struct ext4_inode_info *ei)
85{
86 __u32 provided, calculated;
87
88 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
89 cpu_to_le32(EXT4_OS_LINUX) ||
90 !ext4_has_metadata_csum(inode->i_sb))
91 return 1;
92
93 provided = le16_to_cpu(raw->i_checksum_lo);
94 calculated = ext4_inode_csum(inode, raw, ei);
95 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
96 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
97 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
98 else
99 calculated &= 0xFFFF;
100
101 return provided == calculated;
102}
103
104static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
105 struct ext4_inode_info *ei)
106{
107 __u32 csum;
108
109 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
110 cpu_to_le32(EXT4_OS_LINUX) ||
111 !ext4_has_metadata_csum(inode->i_sb))
112 return;
113
114 csum = ext4_inode_csum(inode, raw, ei);
115 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
116 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
117 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
118 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
119}
120
121static inline int ext4_begin_ordered_truncate(struct inode *inode,
122 loff_t new_size)
123{
124 trace_ext4_begin_ordered_truncate(inode, new_size);
125 /*
126 * If jinode is zero, then we never opened the file for
127 * writing, so there's no need to call
128 * jbd2_journal_begin_ordered_truncate() since there's no
129 * outstanding writes we need to flush.
130 */
131 if (!EXT4_I(inode)->jinode)
132 return 0;
133 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
134 EXT4_I(inode)->jinode,
135 new_size);
136}
137
138static void ext4_invalidatepage(struct page *page, unsigned int offset,
139 unsigned int length);
140static int __ext4_journalled_writepage(struct page *page, unsigned int len);
141static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
142static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
143 int pextents);
144
145/*
146 * Test whether an inode is a fast symlink.
147 */
148int ext4_inode_is_fast_symlink(struct inode *inode)
149{
150 int ea_blocks = EXT4_I(inode)->i_file_acl ?
151 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
152
153 if (ext4_has_inline_data(inode))
154 return 0;
155
156 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
157}
158
159/*
160 * Restart the transaction associated with *handle. This does a commit,
161 * so before we call here everything must be consistently dirtied against
162 * this transaction.
163 */
164int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
165 int nblocks)
166{
167 int ret;
168
169 /*
170 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
171 * moment, get_block can be called only for blocks inside i_size since
172 * page cache has been already dropped and writes are blocked by
173 * i_mutex. So we can safely drop the i_data_sem here.
174 */
175 BUG_ON(EXT4_JOURNAL(inode) == NULL);
176 jbd_debug(2, "restarting handle %p\n", handle);
177 up_write(&EXT4_I(inode)->i_data_sem);
178 ret = ext4_journal_restart(handle, nblocks);
179 down_write(&EXT4_I(inode)->i_data_sem);
180 ext4_discard_preallocations(inode);
181
182 return ret;
183}
184
185/*
186 * Called at the last iput() if i_nlink is zero.
187 */
188void ext4_evict_inode(struct inode *inode)
189{
190 handle_t *handle;
191 int err;
192
193 trace_ext4_evict_inode(inode);
194
195 if (inode->i_nlink) {
196 /*
197 * When journalling data dirty buffers are tracked only in the
198 * journal. So although mm thinks everything is clean and
199 * ready for reaping the inode might still have some pages to
200 * write in the running transaction or waiting to be
201 * checkpointed. Thus calling jbd2_journal_invalidatepage()
202 * (via truncate_inode_pages()) to discard these buffers can
203 * cause data loss. Also even if we did not discard these
204 * buffers, we would have no way to find them after the inode
205 * is reaped and thus user could see stale data if he tries to
206 * read them before the transaction is checkpointed. So be
207 * careful and force everything to disk here... We use
208 * ei->i_datasync_tid to store the newest transaction
209 * containing inode's data.
210 *
211 * Note that directories do not have this problem because they
212 * don't use page cache.
213 */
214 if (inode->i_ino != EXT4_JOURNAL_INO &&
215 ext4_should_journal_data(inode) &&
216 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
217 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
218 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
219
220 jbd2_complete_transaction(journal, commit_tid);
221 filemap_write_and_wait(&inode->i_data);
222 }
223 truncate_inode_pages_final(&inode->i_data);
224
225 goto no_delete;
226 }
227
228 if (is_bad_inode(inode))
229 goto no_delete;
230 dquot_initialize(inode);
231
232 if (ext4_should_order_data(inode))
233 ext4_begin_ordered_truncate(inode, 0);
234 truncate_inode_pages_final(&inode->i_data);
235
236 /*
237 * Protect us against freezing - iput() caller didn't have to have any
238 * protection against it
239 */
240 sb_start_intwrite(inode->i_sb);
241 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
242 ext4_blocks_for_truncate(inode)+3);
243 if (IS_ERR(handle)) {
244 ext4_std_error(inode->i_sb, PTR_ERR(handle));
245 /*
246 * If we're going to skip the normal cleanup, we still need to
247 * make sure that the in-core orphan linked list is properly
248 * cleaned up.
249 */
250 ext4_orphan_del(NULL, inode);
251 sb_end_intwrite(inode->i_sb);
252 goto no_delete;
253 }
254
255 if (IS_SYNC(inode))
256 ext4_handle_sync(handle);
257 inode->i_size = 0;
258 err = ext4_mark_inode_dirty(handle, inode);
259 if (err) {
260 ext4_warning(inode->i_sb,
261 "couldn't mark inode dirty (err %d)", err);
262 goto stop_handle;
263 }
264 if (inode->i_blocks) {
265 err = ext4_truncate(inode);
266 if (err) {
267 ext4_error(inode->i_sb,
268 "couldn't truncate inode %lu (err %d)",
269 inode->i_ino, err);
270 goto stop_handle;
271 }
272 }
273
274 /*
275 * ext4_ext_truncate() doesn't reserve any slop when it
276 * restarts journal transactions; therefore there may not be
277 * enough credits left in the handle to remove the inode from
278 * the orphan list and set the dtime field.
279 */
280 if (!ext4_handle_has_enough_credits(handle, 3)) {
281 err = ext4_journal_extend(handle, 3);
282 if (err > 0)
283 err = ext4_journal_restart(handle, 3);
284 if (err != 0) {
285 ext4_warning(inode->i_sb,
286 "couldn't extend journal (err %d)", err);
287 stop_handle:
288 ext4_journal_stop(handle);
289 ext4_orphan_del(NULL, inode);
290 sb_end_intwrite(inode->i_sb);
291 goto no_delete;
292 }
293 }
294
295 /*
296 * Kill off the orphan record which ext4_truncate created.
297 * AKPM: I think this can be inside the above `if'.
298 * Note that ext4_orphan_del() has to be able to cope with the
299 * deletion of a non-existent orphan - this is because we don't
300 * know if ext4_truncate() actually created an orphan record.
301 * (Well, we could do this if we need to, but heck - it works)
302 */
303 ext4_orphan_del(handle, inode);
304 EXT4_I(inode)->i_dtime = get_seconds();
305
306 /*
307 * One subtle ordering requirement: if anything has gone wrong
308 * (transaction abort, IO errors, whatever), then we can still
309 * do these next steps (the fs will already have been marked as
310 * having errors), but we can't free the inode if the mark_dirty
311 * fails.
312 */
313 if (ext4_mark_inode_dirty(handle, inode))
314 /* If that failed, just do the required in-core inode clear. */
315 ext4_clear_inode(inode);
316 else
317 ext4_free_inode(handle, inode);
318 ext4_journal_stop(handle);
319 sb_end_intwrite(inode->i_sb);
320 return;
321no_delete:
322 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
323}
324
325#ifdef CONFIG_QUOTA
326qsize_t *ext4_get_reserved_space(struct inode *inode)
327{
328 return &EXT4_I(inode)->i_reserved_quota;
329}
330#endif
331
332/*
333 * Called with i_data_sem down, which is important since we can call
334 * ext4_discard_preallocations() from here.
335 */
336void ext4_da_update_reserve_space(struct inode *inode,
337 int used, int quota_claim)
338{
339 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
340 struct ext4_inode_info *ei = EXT4_I(inode);
341
342 spin_lock(&ei->i_block_reservation_lock);
343 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
344 if (unlikely(used > ei->i_reserved_data_blocks)) {
345 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
346 "with only %d reserved data blocks",
347 __func__, inode->i_ino, used,
348 ei->i_reserved_data_blocks);
349 WARN_ON(1);
350 used = ei->i_reserved_data_blocks;
351 }
352
353 /* Update per-inode reservations */
354 ei->i_reserved_data_blocks -= used;
355 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
356
357 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
358
359 /* Update quota subsystem for data blocks */
360 if (quota_claim)
361 dquot_claim_block(inode, EXT4_C2B(sbi, used));
362 else {
363 /*
364 * We did fallocate with an offset that is already delayed
365 * allocated. So on delayed allocated writeback we should
366 * not re-claim the quota for fallocated blocks.
367 */
368 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
369 }
370
371 /*
372 * If we have done all the pending block allocations and if
373 * there aren't any writers on the inode, we can discard the
374 * inode's preallocations.
375 */
376 if ((ei->i_reserved_data_blocks == 0) &&
377 (atomic_read(&inode->i_writecount) == 0))
378 ext4_discard_preallocations(inode);
379}
380
381static int __check_block_validity(struct inode *inode, const char *func,
382 unsigned int line,
383 struct ext4_map_blocks *map)
384{
385 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
386 map->m_len)) {
387 ext4_error_inode(inode, func, line, map->m_pblk,
388 "lblock %lu mapped to illegal pblock "
389 "(length %d)", (unsigned long) map->m_lblk,
390 map->m_len);
391 return -EFSCORRUPTED;
392 }
393 return 0;
394}
395
396int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
397 ext4_lblk_t len)
398{
399 int ret;
400
401 if (ext4_encrypted_inode(inode))
402 return fscrypt_zeroout_range(inode, lblk, pblk, len);
403
404 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
405 if (ret > 0)
406 ret = 0;
407
408 return ret;
409}
410
411#define check_block_validity(inode, map) \
412 __check_block_validity((inode), __func__, __LINE__, (map))
413
414#ifdef ES_AGGRESSIVE_TEST
415static void ext4_map_blocks_es_recheck(handle_t *handle,
416 struct inode *inode,
417 struct ext4_map_blocks *es_map,
418 struct ext4_map_blocks *map,
419 int flags)
420{
421 int retval;
422
423 map->m_flags = 0;
424 /*
425 * There is a race window that the result is not the same.
426 * e.g. xfstests #223 when dioread_nolock enables. The reason
427 * is that we lookup a block mapping in extent status tree with
428 * out taking i_data_sem. So at the time the unwritten extent
429 * could be converted.
430 */
431 down_read(&EXT4_I(inode)->i_data_sem);
432 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
433 retval = ext4_ext_map_blocks(handle, inode, map, flags &
434 EXT4_GET_BLOCKS_KEEP_SIZE);
435 } else {
436 retval = ext4_ind_map_blocks(handle, inode, map, flags &
437 EXT4_GET_BLOCKS_KEEP_SIZE);
438 }
439 up_read((&EXT4_I(inode)->i_data_sem));
440
441 /*
442 * We don't check m_len because extent will be collpased in status
443 * tree. So the m_len might not equal.
444 */
445 if (es_map->m_lblk != map->m_lblk ||
446 es_map->m_flags != map->m_flags ||
447 es_map->m_pblk != map->m_pblk) {
448 printk("ES cache assertion failed for inode: %lu "
449 "es_cached ex [%d/%d/%llu/%x] != "
450 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
451 inode->i_ino, es_map->m_lblk, es_map->m_len,
452 es_map->m_pblk, es_map->m_flags, map->m_lblk,
453 map->m_len, map->m_pblk, map->m_flags,
454 retval, flags);
455 }
456}
457#endif /* ES_AGGRESSIVE_TEST */
458
459/*
460 * The ext4_map_blocks() function tries to look up the requested blocks,
461 * and returns if the blocks are already mapped.
462 *
463 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
464 * and store the allocated blocks in the result buffer head and mark it
465 * mapped.
466 *
467 * If file type is extents based, it will call ext4_ext_map_blocks(),
468 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
469 * based files
470 *
471 * On success, it returns the number of blocks being mapped or allocated. if
472 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
473 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
474 *
475 * It returns 0 if plain look up failed (blocks have not been allocated), in
476 * that case, @map is returned as unmapped but we still do fill map->m_len to
477 * indicate the length of a hole starting at map->m_lblk.
478 *
479 * It returns the error in case of allocation failure.
480 */
481int ext4_map_blocks(handle_t *handle, struct inode *inode,
482 struct ext4_map_blocks *map, int flags)
483{
484 struct extent_status es;
485 int retval;
486 int ret = 0;
487#ifdef ES_AGGRESSIVE_TEST
488 struct ext4_map_blocks orig_map;
489
490 memcpy(&orig_map, map, sizeof(*map));
491#endif
492
493 map->m_flags = 0;
494 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
495 "logical block %lu\n", inode->i_ino, flags, map->m_len,
496 (unsigned long) map->m_lblk);
497
498 /*
499 * ext4_map_blocks returns an int, and m_len is an unsigned int
500 */
501 if (unlikely(map->m_len > INT_MAX))
502 map->m_len = INT_MAX;
503
504 /* We can handle the block number less than EXT_MAX_BLOCKS */
505 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
506 return -EFSCORRUPTED;
507
508 /* Lookup extent status tree firstly */
509 if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
510 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
511 map->m_pblk = ext4_es_pblock(&es) +
512 map->m_lblk - es.es_lblk;
513 map->m_flags |= ext4_es_is_written(&es) ?
514 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
515 retval = es.es_len - (map->m_lblk - es.es_lblk);
516 if (retval > map->m_len)
517 retval = map->m_len;
518 map->m_len = retval;
519 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
520 map->m_pblk = 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_ON(1);
528 }
529#ifdef ES_AGGRESSIVE_TEST
530 ext4_map_blocks_es_recheck(handle, inode, map,
531 &orig_map, flags);
532#endif
533 goto found;
534 }
535
536 /*
537 * Try to see if we can get the block without requesting a new
538 * file system block.
539 */
540 down_read(&EXT4_I(inode)->i_data_sem);
541 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
542 retval = ext4_ext_map_blocks(handle, inode, map, flags &
543 EXT4_GET_BLOCKS_KEEP_SIZE);
544 } else {
545 retval = ext4_ind_map_blocks(handle, inode, map, flags &
546 EXT4_GET_BLOCKS_KEEP_SIZE);
547 }
548 if (retval > 0) {
549 unsigned int status;
550
551 if (unlikely(retval != map->m_len)) {
552 ext4_warning(inode->i_sb,
553 "ES len assertion failed for inode "
554 "%lu: retval %d != map->m_len %d",
555 inode->i_ino, retval, map->m_len);
556 WARN_ON(1);
557 }
558
559 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
560 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
561 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
562 !(status & EXTENT_STATUS_WRITTEN) &&
563 ext4_find_delalloc_range(inode, map->m_lblk,
564 map->m_lblk + map->m_len - 1))
565 status |= EXTENT_STATUS_DELAYED;
566 ret = ext4_es_insert_extent(inode, map->m_lblk,
567 map->m_len, map->m_pblk, status);
568 if (ret < 0)
569 retval = ret;
570 }
571 up_read((&EXT4_I(inode)->i_data_sem));
572
573found:
574 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
575 ret = check_block_validity(inode, map);
576 if (ret != 0)
577 return ret;
578 }
579
580 /* If it is only a block(s) look up */
581 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
582 return retval;
583
584 /*
585 * Returns if the blocks have already allocated
586 *
587 * Note that if blocks have been preallocated
588 * ext4_ext_get_block() returns the create = 0
589 * with buffer head unmapped.
590 */
591 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
592 /*
593 * If we need to convert extent to unwritten
594 * we continue and do the actual work in
595 * ext4_ext_map_blocks()
596 */
597 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
598 return retval;
599
600 /*
601 * Here we clear m_flags because after allocating an new extent,
602 * it will be set again.
603 */
604 map->m_flags &= ~EXT4_MAP_FLAGS;
605
606 /*
607 * New blocks allocate and/or writing to unwritten extent
608 * will possibly result in updating i_data, so we take
609 * the write lock of i_data_sem, and call get_block()
610 * with create == 1 flag.
611 */
612 down_write(&EXT4_I(inode)->i_data_sem);
613
614 /*
615 * We need to check for EXT4 here because migrate
616 * could have changed the inode type in between
617 */
618 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
619 retval = ext4_ext_map_blocks(handle, inode, map, flags);
620 } else {
621 retval = ext4_ind_map_blocks(handle, inode, map, flags);
622
623 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
624 /*
625 * We allocated new blocks which will result in
626 * i_data's format changing. Force the migrate
627 * to fail by clearing migrate flags
628 */
629 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
630 }
631
632 /*
633 * Update reserved blocks/metadata blocks after successful
634 * block allocation which had been deferred till now. We don't
635 * support fallocate for non extent files. So we can update
636 * reserve space here.
637 */
638 if ((retval > 0) &&
639 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
640 ext4_da_update_reserve_space(inode, retval, 1);
641 }
642
643 if (retval > 0) {
644 unsigned int status;
645
646 if (unlikely(retval != map->m_len)) {
647 ext4_warning(inode->i_sb,
648 "ES len assertion failed for inode "
649 "%lu: retval %d != map->m_len %d",
650 inode->i_ino, retval, map->m_len);
651 WARN_ON(1);
652 }
653
654 /*
655 * We have to zeroout blocks before inserting them into extent
656 * status tree. Otherwise someone could look them up there and
657 * use them before they are really zeroed. We also have to
658 * unmap metadata before zeroing as otherwise writeback can
659 * overwrite zeros with stale data from block device.
660 */
661 if (flags & EXT4_GET_BLOCKS_ZERO &&
662 map->m_flags & EXT4_MAP_MAPPED &&
663 map->m_flags & EXT4_MAP_NEW) {
664 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
665 map->m_len);
666 ret = ext4_issue_zeroout(inode, map->m_lblk,
667 map->m_pblk, map->m_len);
668 if (ret) {
669 retval = ret;
670 goto out_sem;
671 }
672 }
673
674 /*
675 * If the extent has been zeroed out, we don't need to update
676 * extent status tree.
677 */
678 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
679 ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
680 if (ext4_es_is_written(&es))
681 goto out_sem;
682 }
683 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
684 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
685 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
686 !(status & EXTENT_STATUS_WRITTEN) &&
687 ext4_find_delalloc_range(inode, map->m_lblk,
688 map->m_lblk + map->m_len - 1))
689 status |= EXTENT_STATUS_DELAYED;
690 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
691 map->m_pblk, status);
692 if (ret < 0) {
693 retval = ret;
694 goto out_sem;
695 }
696 }
697
698out_sem:
699 up_write((&EXT4_I(inode)->i_data_sem));
700 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
701 ret = check_block_validity(inode, map);
702 if (ret != 0)
703 return ret;
704
705 /*
706 * Inodes with freshly allocated blocks where contents will be
707 * visible after transaction commit must be on transaction's
708 * ordered data list.
709 */
710 if (map->m_flags & EXT4_MAP_NEW &&
711 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
712 !(flags & EXT4_GET_BLOCKS_ZERO) &&
713 !IS_NOQUOTA(inode) &&
714 ext4_should_order_data(inode)) {
715 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
716 ret = ext4_jbd2_inode_add_wait(handle, inode);
717 else
718 ret = ext4_jbd2_inode_add_write(handle, inode);
719 if (ret)
720 return ret;
721 }
722 }
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 do {
748 old_state = READ_ONCE(bh->b_state);
749 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
750 } while (unlikely(
751 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
752}
753
754static int _ext4_get_block(struct inode *inode, sector_t iblock,
755 struct buffer_head *bh, int flags)
756{
757 struct ext4_map_blocks map;
758 int ret = 0;
759
760 if (ext4_has_inline_data(inode))
761 return -ERANGE;
762
763 map.m_lblk = iblock;
764 map.m_len = bh->b_size >> inode->i_blkbits;
765
766 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
767 flags);
768 if (ret > 0) {
769 map_bh(bh, inode->i_sb, map.m_pblk);
770 ext4_update_bh_state(bh, map.m_flags);
771 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
772 ret = 0;
773 } else if (ret == 0) {
774 /* hole case, need to fill in bh->b_size */
775 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
776 }
777 return ret;
778}
779
780int ext4_get_block(struct inode *inode, sector_t iblock,
781 struct buffer_head *bh, int create)
782{
783 return _ext4_get_block(inode, iblock, bh,
784 create ? EXT4_GET_BLOCKS_CREATE : 0);
785}
786
787/*
788 * Get block function used when preparing for buffered write if we require
789 * creating an unwritten extent if blocks haven't been allocated. The extent
790 * will be converted to written after the IO is complete.
791 */
792int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
793 struct buffer_head *bh_result, int create)
794{
795 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
796 inode->i_ino, create);
797 return _ext4_get_block(inode, iblock, bh_result,
798 EXT4_GET_BLOCKS_IO_CREATE_EXT);
799}
800
801/* Maximum number of blocks we map for direct IO at once. */
802#define DIO_MAX_BLOCKS 4096
803
804/*
805 * Get blocks function for the cases that need to start a transaction -
806 * generally difference cases of direct IO and DAX IO. It also handles retries
807 * in case of ENOSPC.
808 */
809static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
810 struct buffer_head *bh_result, int flags)
811{
812 int dio_credits;
813 handle_t *handle;
814 int retries = 0;
815 int ret;
816
817 /* Trim mapping request to maximum we can map at once for DIO */
818 if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
819 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
820 dio_credits = ext4_chunk_trans_blocks(inode,
821 bh_result->b_size >> inode->i_blkbits);
822retry:
823 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
824 if (IS_ERR(handle))
825 return PTR_ERR(handle);
826
827 ret = _ext4_get_block(inode, iblock, bh_result, flags);
828 ext4_journal_stop(handle);
829
830 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
831 goto retry;
832 return ret;
833}
834
835/* Get block function for DIO reads and writes to inodes without extents */
836int ext4_dio_get_block(struct inode *inode, sector_t iblock,
837 struct buffer_head *bh, int create)
838{
839 /* We don't expect handle for direct IO */
840 WARN_ON_ONCE(ext4_journal_current_handle());
841
842 if (!create)
843 return _ext4_get_block(inode, iblock, bh, 0);
844 return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
845}
846
847/*
848 * Get block function for AIO DIO writes when we create unwritten extent if
849 * blocks are not allocated yet. The extent will be converted to written
850 * after IO is complete.
851 */
852static int ext4_dio_get_block_unwritten_async(struct inode *inode,
853 sector_t iblock, struct buffer_head *bh_result, int create)
854{
855 int ret;
856
857 /* We don't expect handle for direct IO */
858 WARN_ON_ONCE(ext4_journal_current_handle());
859
860 ret = ext4_get_block_trans(inode, iblock, bh_result,
861 EXT4_GET_BLOCKS_IO_CREATE_EXT);
862
863 /*
864 * When doing DIO using unwritten extents, we need io_end to convert
865 * unwritten extents to written on IO completion. We allocate io_end
866 * once we spot unwritten extent and store it in b_private. Generic
867 * DIO code keeps b_private set and furthermore passes the value to
868 * our completion callback in 'private' argument.
869 */
870 if (!ret && buffer_unwritten(bh_result)) {
871 if (!bh_result->b_private) {
872 ext4_io_end_t *io_end;
873
874 io_end = ext4_init_io_end(inode, GFP_KERNEL);
875 if (!io_end)
876 return -ENOMEM;
877 bh_result->b_private = io_end;
878 ext4_set_io_unwritten_flag(inode, io_end);
879 }
880 set_buffer_defer_completion(bh_result);
881 }
882
883 return ret;
884}
885
886/*
887 * Get block function for non-AIO DIO writes when we create unwritten extent if
888 * blocks are not allocated yet. The extent will be converted to written
889 * after IO is complete from ext4_ext_direct_IO() function.
890 */
891static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
892 sector_t iblock, struct buffer_head *bh_result, int create)
893{
894 int ret;
895
896 /* We don't expect handle for direct IO */
897 WARN_ON_ONCE(ext4_journal_current_handle());
898
899 ret = ext4_get_block_trans(inode, iblock, bh_result,
900 EXT4_GET_BLOCKS_IO_CREATE_EXT);
901
902 /*
903 * Mark inode as having pending DIO writes to unwritten extents.
904 * ext4_ext_direct_IO() checks this flag and converts extents to
905 * written.
906 */
907 if (!ret && buffer_unwritten(bh_result))
908 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
909
910 return ret;
911}
912
913static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
914 struct buffer_head *bh_result, int create)
915{
916 int ret;
917
918 ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
919 inode->i_ino, create);
920 /* We don't expect handle for direct IO */
921 WARN_ON_ONCE(ext4_journal_current_handle());
922
923 ret = _ext4_get_block(inode, iblock, bh_result, 0);
924 /*
925 * Blocks should have been preallocated! ext4_file_write_iter() checks
926 * that.
927 */
928 WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
929
930 return ret;
931}
932
933
934/*
935 * `handle' can be NULL if create is zero
936 */
937struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
938 ext4_lblk_t block, int map_flags)
939{
940 struct ext4_map_blocks map;
941 struct buffer_head *bh;
942 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
943 int err;
944
945 J_ASSERT(handle != NULL || create == 0);
946
947 map.m_lblk = block;
948 map.m_len = 1;
949 err = ext4_map_blocks(handle, inode, &map, map_flags);
950
951 if (err == 0)
952 return create ? ERR_PTR(-ENOSPC) : NULL;
953 if (err < 0)
954 return ERR_PTR(err);
955
956 bh = sb_getblk(inode->i_sb, map.m_pblk);
957 if (unlikely(!bh))
958 return ERR_PTR(-ENOMEM);
959 if (map.m_flags & EXT4_MAP_NEW) {
960 J_ASSERT(create != 0);
961 J_ASSERT(handle != NULL);
962
963 /*
964 * Now that we do not always journal data, we should
965 * keep in mind whether this should always journal the
966 * new buffer as metadata. For now, regular file
967 * writes use ext4_get_block instead, so it's not a
968 * problem.
969 */
970 lock_buffer(bh);
971 BUFFER_TRACE(bh, "call get_create_access");
972 err = ext4_journal_get_create_access(handle, bh);
973 if (unlikely(err)) {
974 unlock_buffer(bh);
975 goto errout;
976 }
977 if (!buffer_uptodate(bh)) {
978 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
979 set_buffer_uptodate(bh);
980 }
981 unlock_buffer(bh);
982 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
983 err = ext4_handle_dirty_metadata(handle, inode, bh);
984 if (unlikely(err))
985 goto errout;
986 } else
987 BUFFER_TRACE(bh, "not a new buffer");
988 return bh;
989errout:
990 brelse(bh);
991 return ERR_PTR(err);
992}
993
994struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
995 ext4_lblk_t block, int map_flags)
996{
997 struct buffer_head *bh;
998
999 bh = ext4_getblk(handle, inode, block, map_flags);
1000 if (IS_ERR(bh))
1001 return bh;
1002 if (!bh || buffer_uptodate(bh))
1003 return bh;
1004 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1005 wait_on_buffer(bh);
1006 if (buffer_uptodate(bh))
1007 return bh;
1008 put_bh(bh);
1009 return ERR_PTR(-EIO);
1010}
1011
1012int ext4_walk_page_buffers(handle_t *handle,
1013 struct buffer_head *head,
1014 unsigned from,
1015 unsigned to,
1016 int *partial,
1017 int (*fn)(handle_t *handle,
1018 struct buffer_head *bh))
1019{
1020 struct buffer_head *bh;
1021 unsigned block_start, block_end;
1022 unsigned blocksize = head->b_size;
1023 int err, ret = 0;
1024 struct buffer_head *next;
1025
1026 for (bh = head, block_start = 0;
1027 ret == 0 && (bh != head || !block_start);
1028 block_start = block_end, bh = next) {
1029 next = bh->b_this_page;
1030 block_end = block_start + blocksize;
1031 if (block_end <= from || block_start >= to) {
1032 if (partial && !buffer_uptodate(bh))
1033 *partial = 1;
1034 continue;
1035 }
1036 err = (*fn)(handle, bh);
1037 if (!ret)
1038 ret = err;
1039 }
1040 return ret;
1041}
1042
1043/*
1044 * To preserve ordering, it is essential that the hole instantiation and
1045 * the data write be encapsulated in a single transaction. We cannot
1046 * close off a transaction and start a new one between the ext4_get_block()
1047 * and the commit_write(). So doing the jbd2_journal_start at the start of
1048 * prepare_write() is the right place.
1049 *
1050 * Also, this function can nest inside ext4_writepage(). In that case, we
1051 * *know* that ext4_writepage() has generated enough buffer credits to do the
1052 * whole page. So we won't block on the journal in that case, which is good,
1053 * because the caller may be PF_MEMALLOC.
1054 *
1055 * By accident, ext4 can be reentered when a transaction is open via
1056 * quota file writes. If we were to commit the transaction while thus
1057 * reentered, there can be a deadlock - we would be holding a quota
1058 * lock, and the commit would never complete if another thread had a
1059 * transaction open and was blocking on the quota lock - a ranking
1060 * violation.
1061 *
1062 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1063 * will _not_ run commit under these circumstances because handle->h_ref
1064 * is elevated. We'll still have enough credits for the tiny quotafile
1065 * write.
1066 */
1067int do_journal_get_write_access(handle_t *handle,
1068 struct buffer_head *bh)
1069{
1070 int dirty = buffer_dirty(bh);
1071 int ret;
1072
1073 if (!buffer_mapped(bh) || buffer_freed(bh))
1074 return 0;
1075 /*
1076 * __block_write_begin() could have dirtied some buffers. Clean
1077 * the dirty bit as jbd2_journal_get_write_access() could complain
1078 * otherwise about fs integrity issues. Setting of the dirty bit
1079 * by __block_write_begin() isn't a real problem here as we clear
1080 * the bit before releasing a page lock and thus writeback cannot
1081 * ever write the buffer.
1082 */
1083 if (dirty)
1084 clear_buffer_dirty(bh);
1085 BUFFER_TRACE(bh, "get write access");
1086 ret = ext4_journal_get_write_access(handle, bh);
1087 if (!ret && dirty)
1088 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1089 return ret;
1090}
1091
1092#ifdef CONFIG_EXT4_FS_ENCRYPTION
1093static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1094 get_block_t *get_block)
1095{
1096 unsigned from = pos & (PAGE_SIZE - 1);
1097 unsigned to = from + len;
1098 struct inode *inode = page->mapping->host;
1099 unsigned block_start, block_end;
1100 sector_t block;
1101 int err = 0;
1102 unsigned blocksize = inode->i_sb->s_blocksize;
1103 unsigned bbits;
1104 struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1105 bool decrypt = false;
1106
1107 BUG_ON(!PageLocked(page));
1108 BUG_ON(from > PAGE_SIZE);
1109 BUG_ON(to > PAGE_SIZE);
1110 BUG_ON(from > to);
1111
1112 if (!page_has_buffers(page))
1113 create_empty_buffers(page, blocksize, 0);
1114 head = page_buffers(page);
1115 bbits = ilog2(blocksize);
1116 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1117
1118 for (bh = head, block_start = 0; bh != head || !block_start;
1119 block++, block_start = block_end, bh = bh->b_this_page) {
1120 block_end = block_start + blocksize;
1121 if (block_end <= from || block_start >= to) {
1122 if (PageUptodate(page)) {
1123 if (!buffer_uptodate(bh))
1124 set_buffer_uptodate(bh);
1125 }
1126 continue;
1127 }
1128 if (buffer_new(bh))
1129 clear_buffer_new(bh);
1130 if (!buffer_mapped(bh)) {
1131 WARN_ON(bh->b_size != blocksize);
1132 err = get_block(inode, block, bh, 1);
1133 if (err)
1134 break;
1135 if (buffer_new(bh)) {
1136 clean_bdev_bh_alias(bh);
1137 if (PageUptodate(page)) {
1138 clear_buffer_new(bh);
1139 set_buffer_uptodate(bh);
1140 mark_buffer_dirty(bh);
1141 continue;
1142 }
1143 if (block_end > to || block_start < from)
1144 zero_user_segments(page, to, block_end,
1145 block_start, from);
1146 continue;
1147 }
1148 }
1149 if (PageUptodate(page)) {
1150 if (!buffer_uptodate(bh))
1151 set_buffer_uptodate(bh);
1152 continue;
1153 }
1154 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1155 !buffer_unwritten(bh) &&
1156 (block_start < from || block_end > to)) {
1157 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1158 *wait_bh++ = bh;
1159 decrypt = ext4_encrypted_inode(inode) &&
1160 S_ISREG(inode->i_mode);
1161 }
1162 }
1163 /*
1164 * If we issued read requests, let them complete.
1165 */
1166 while (wait_bh > wait) {
1167 wait_on_buffer(*--wait_bh);
1168 if (!buffer_uptodate(*wait_bh))
1169 err = -EIO;
1170 }
1171 if (unlikely(err))
1172 page_zero_new_buffers(page, from, to);
1173 else if (decrypt)
1174 err = fscrypt_decrypt_page(page->mapping->host, page,
1175 PAGE_SIZE, 0, page->index);
1176 return err;
1177}
1178#endif
1179
1180static int ext4_write_begin(struct file *file, struct address_space *mapping,
1181 loff_t pos, unsigned len, unsigned flags,
1182 struct page **pagep, void **fsdata)
1183{
1184 struct inode *inode = mapping->host;
1185 int ret, needed_blocks;
1186 handle_t *handle;
1187 int retries = 0;
1188 struct page *page;
1189 pgoff_t index;
1190 unsigned from, to;
1191
1192 trace_ext4_write_begin(inode, pos, len, flags);
1193 /*
1194 * Reserve one block more for addition to orphan list in case
1195 * we allocate blocks but write fails for some reason
1196 */
1197 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1198 index = pos >> PAGE_SHIFT;
1199 from = pos & (PAGE_SIZE - 1);
1200 to = from + len;
1201
1202 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1203 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1204 flags, pagep);
1205 if (ret < 0)
1206 return ret;
1207 if (ret == 1)
1208 return 0;
1209 }
1210
1211 /*
1212 * grab_cache_page_write_begin() can take a long time if the
1213 * system is thrashing due to memory pressure, or if the page
1214 * is being written back. So grab it first before we start
1215 * the transaction handle. This also allows us to allocate
1216 * the page (if needed) without using GFP_NOFS.
1217 */
1218retry_grab:
1219 page = grab_cache_page_write_begin(mapping, index, flags);
1220 if (!page)
1221 return -ENOMEM;
1222 unlock_page(page);
1223
1224retry_journal:
1225 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1226 if (IS_ERR(handle)) {
1227 put_page(page);
1228 return PTR_ERR(handle);
1229 }
1230
1231 lock_page(page);
1232 if (page->mapping != mapping) {
1233 /* The page got truncated from under us */
1234 unlock_page(page);
1235 put_page(page);
1236 ext4_journal_stop(handle);
1237 goto retry_grab;
1238 }
1239 /* In case writeback began while the page was unlocked */
1240 wait_for_stable_page(page);
1241
1242#ifdef CONFIG_EXT4_FS_ENCRYPTION
1243 if (ext4_should_dioread_nolock(inode))
1244 ret = ext4_block_write_begin(page, pos, len,
1245 ext4_get_block_unwritten);
1246 else
1247 ret = ext4_block_write_begin(page, pos, len,
1248 ext4_get_block);
1249#else
1250 if (ext4_should_dioread_nolock(inode))
1251 ret = __block_write_begin(page, pos, len,
1252 ext4_get_block_unwritten);
1253 else
1254 ret = __block_write_begin(page, pos, len, ext4_get_block);
1255#endif
1256 if (!ret && ext4_should_journal_data(inode)) {
1257 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1258 from, to, NULL,
1259 do_journal_get_write_access);
1260 }
1261
1262 if (ret) {
1263 unlock_page(page);
1264 /*
1265 * __block_write_begin may have instantiated a few blocks
1266 * outside i_size. Trim these off again. Don't need
1267 * i_size_read because we hold i_mutex.
1268 *
1269 * Add inode to orphan list in case we crash before
1270 * truncate finishes
1271 */
1272 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1273 ext4_orphan_add(handle, inode);
1274
1275 ext4_journal_stop(handle);
1276 if (pos + len > inode->i_size) {
1277 ext4_truncate_failed_write(inode);
1278 /*
1279 * If truncate failed early the inode might
1280 * still be on the orphan list; we need to
1281 * make sure the inode is removed from the
1282 * orphan list in that case.
1283 */
1284 if (inode->i_nlink)
1285 ext4_orphan_del(NULL, inode);
1286 }
1287
1288 if (ret == -ENOSPC &&
1289 ext4_should_retry_alloc(inode->i_sb, &retries))
1290 goto retry_journal;
1291 put_page(page);
1292 return ret;
1293 }
1294 *pagep = page;
1295 return ret;
1296}
1297
1298/* For write_end() in data=journal mode */
1299static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1300{
1301 int ret;
1302 if (!buffer_mapped(bh) || buffer_freed(bh))
1303 return 0;
1304 set_buffer_uptodate(bh);
1305 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1306 clear_buffer_meta(bh);
1307 clear_buffer_prio(bh);
1308 return ret;
1309}
1310
1311/*
1312 * We need to pick up the new inode size which generic_commit_write gave us
1313 * `file' can be NULL - eg, when called from page_symlink().
1314 *
1315 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1316 * buffers are managed internally.
1317 */
1318static int ext4_write_end(struct file *file,
1319 struct address_space *mapping,
1320 loff_t pos, unsigned len, unsigned copied,
1321 struct page *page, void *fsdata)
1322{
1323 handle_t *handle = ext4_journal_current_handle();
1324 struct inode *inode = mapping->host;
1325 loff_t old_size = inode->i_size;
1326 int ret = 0, ret2;
1327 int i_size_changed = 0;
1328
1329 trace_ext4_write_end(inode, pos, len, copied);
1330 if (ext4_has_inline_data(inode)) {
1331 ret = ext4_write_inline_data_end(inode, pos, len,
1332 copied, page);
1333 if (ret < 0) {
1334 unlock_page(page);
1335 put_page(page);
1336 goto errout;
1337 }
1338 copied = ret;
1339 } else
1340 copied = block_write_end(file, mapping, pos,
1341 len, copied, page, fsdata);
1342 /*
1343 * it's important to update i_size while still holding page lock:
1344 * page writeout could otherwise come in and zero beyond i_size.
1345 */
1346 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1347 unlock_page(page);
1348 put_page(page);
1349
1350 if (old_size < pos)
1351 pagecache_isize_extended(inode, old_size, pos);
1352 /*
1353 * Don't mark the inode dirty under page lock. First, it unnecessarily
1354 * makes the holding time of page lock longer. Second, it forces lock
1355 * ordering of page lock and transaction start for journaling
1356 * filesystems.
1357 */
1358 if (i_size_changed)
1359 ext4_mark_inode_dirty(handle, inode);
1360
1361 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1362 /* if we have allocated more blocks and copied
1363 * less. We will have blocks allocated outside
1364 * inode->i_size. So truncate them
1365 */
1366 ext4_orphan_add(handle, inode);
1367errout:
1368 ret2 = ext4_journal_stop(handle);
1369 if (!ret)
1370 ret = ret2;
1371
1372 if (pos + len > inode->i_size) {
1373 ext4_truncate_failed_write(inode);
1374 /*
1375 * If truncate failed early the inode might still be
1376 * on the orphan list; we need to make sure the inode
1377 * is removed from the orphan list in that case.
1378 */
1379 if (inode->i_nlink)
1380 ext4_orphan_del(NULL, inode);
1381 }
1382
1383 return ret ? ret : copied;
1384}
1385
1386/*
1387 * This is a private version of page_zero_new_buffers() which doesn't
1388 * set the buffer to be dirty, since in data=journalled mode we need
1389 * to call ext4_handle_dirty_metadata() instead.
1390 */
1391static void ext4_journalled_zero_new_buffers(handle_t *handle,
1392 struct page *page,
1393 unsigned from, unsigned to)
1394{
1395 unsigned int block_start = 0, block_end;
1396 struct buffer_head *head, *bh;
1397
1398 bh = head = page_buffers(page);
1399 do {
1400 block_end = block_start + bh->b_size;
1401 if (buffer_new(bh)) {
1402 if (block_end > from && block_start < to) {
1403 if (!PageUptodate(page)) {
1404 unsigned start, size;
1405
1406 start = max(from, block_start);
1407 size = min(to, block_end) - start;
1408
1409 zero_user(page, start, size);
1410 write_end_fn(handle, bh);
1411 }
1412 clear_buffer_new(bh);
1413 }
1414 }
1415 block_start = block_end;
1416 bh = bh->b_this_page;
1417 } while (bh != head);
1418}
1419
1420static int ext4_journalled_write_end(struct file *file,
1421 struct address_space *mapping,
1422 loff_t pos, unsigned len, unsigned copied,
1423 struct page *page, void *fsdata)
1424{
1425 handle_t *handle = ext4_journal_current_handle();
1426 struct inode *inode = mapping->host;
1427 loff_t old_size = inode->i_size;
1428 int ret = 0, ret2;
1429 int partial = 0;
1430 unsigned from, to;
1431 int size_changed = 0;
1432
1433 trace_ext4_journalled_write_end(inode, pos, len, copied);
1434 from = pos & (PAGE_SIZE - 1);
1435 to = from + len;
1436
1437 BUG_ON(!ext4_handle_valid(handle));
1438
1439 if (ext4_has_inline_data(inode)) {
1440 ret = ext4_write_inline_data_end(inode, pos, len,
1441 copied, page);
1442 if (ret < 0) {
1443 unlock_page(page);
1444 put_page(page);
1445 goto errout;
1446 }
1447 copied = ret;
1448 } else if (unlikely(copied < len) && !PageUptodate(page)) {
1449 copied = 0;
1450 ext4_journalled_zero_new_buffers(handle, page, from, to);
1451 } else {
1452 if (unlikely(copied < len))
1453 ext4_journalled_zero_new_buffers(handle, page,
1454 from + copied, to);
1455 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1456 from + copied, &partial,
1457 write_end_fn);
1458 if (!partial)
1459 SetPageUptodate(page);
1460 }
1461 size_changed = ext4_update_inode_size(inode, pos + copied);
1462 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1463 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1464 unlock_page(page);
1465 put_page(page);
1466
1467 if (old_size < pos)
1468 pagecache_isize_extended(inode, old_size, pos);
1469
1470 if (size_changed) {
1471 ret2 = ext4_mark_inode_dirty(handle, inode);
1472 if (!ret)
1473 ret = ret2;
1474 }
1475
1476 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1477 /* if we have allocated more blocks and copied
1478 * less. We will have blocks allocated outside
1479 * inode->i_size. So truncate them
1480 */
1481 ext4_orphan_add(handle, inode);
1482
1483errout:
1484 ret2 = ext4_journal_stop(handle);
1485 if (!ret)
1486 ret = ret2;
1487 if (pos + len > inode->i_size) {
1488 ext4_truncate_failed_write(inode);
1489 /*
1490 * If truncate failed early the inode might still be
1491 * on the orphan list; we need to make sure the inode
1492 * is removed from the orphan list in that case.
1493 */
1494 if (inode->i_nlink)
1495 ext4_orphan_del(NULL, inode);
1496 }
1497
1498 return ret ? ret : copied;
1499}
1500
1501/*
1502 * Reserve space for a single cluster
1503 */
1504static int ext4_da_reserve_space(struct inode *inode)
1505{
1506 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1507 struct ext4_inode_info *ei = EXT4_I(inode);
1508 int ret;
1509
1510 /*
1511 * We will charge metadata quota at writeout time; this saves
1512 * us from metadata over-estimation, though we may go over by
1513 * a small amount in the end. Here we just reserve for data.
1514 */
1515 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1516 if (ret)
1517 return ret;
1518
1519 spin_lock(&ei->i_block_reservation_lock);
1520 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1521 spin_unlock(&ei->i_block_reservation_lock);
1522 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1523 return -ENOSPC;
1524 }
1525 ei->i_reserved_data_blocks++;
1526 trace_ext4_da_reserve_space(inode);
1527 spin_unlock(&ei->i_block_reservation_lock);
1528
1529 return 0; /* success */
1530}
1531
1532static void ext4_da_release_space(struct inode *inode, int to_free)
1533{
1534 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1535 struct ext4_inode_info *ei = EXT4_I(inode);
1536
1537 if (!to_free)
1538 return; /* Nothing to release, exit */
1539
1540 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1541
1542 trace_ext4_da_release_space(inode, to_free);
1543 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1544 /*
1545 * if there aren't enough reserved blocks, then the
1546 * counter is messed up somewhere. Since this
1547 * function is called from invalidate page, it's
1548 * harmless to return without any action.
1549 */
1550 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1551 "ino %lu, to_free %d with only %d reserved "
1552 "data blocks", inode->i_ino, to_free,
1553 ei->i_reserved_data_blocks);
1554 WARN_ON(1);
1555 to_free = ei->i_reserved_data_blocks;
1556 }
1557 ei->i_reserved_data_blocks -= to_free;
1558
1559 /* update fs dirty data blocks counter */
1560 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1561
1562 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1563
1564 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1565}
1566
1567static void ext4_da_page_release_reservation(struct page *page,
1568 unsigned int offset,
1569 unsigned int length)
1570{
1571 int to_release = 0, contiguous_blks = 0;
1572 struct buffer_head *head, *bh;
1573 unsigned int curr_off = 0;
1574 struct inode *inode = page->mapping->host;
1575 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1576 unsigned int stop = offset + length;
1577 int num_clusters;
1578 ext4_fsblk_t lblk;
1579
1580 BUG_ON(stop > PAGE_SIZE || stop < length);
1581
1582 head = page_buffers(page);
1583 bh = head;
1584 do {
1585 unsigned int next_off = curr_off + bh->b_size;
1586
1587 if (next_off > stop)
1588 break;
1589
1590 if ((offset <= curr_off) && (buffer_delay(bh))) {
1591 to_release++;
1592 contiguous_blks++;
1593 clear_buffer_delay(bh);
1594 } else if (contiguous_blks) {
1595 lblk = page->index <<
1596 (PAGE_SHIFT - inode->i_blkbits);
1597 lblk += (curr_off >> inode->i_blkbits) -
1598 contiguous_blks;
1599 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1600 contiguous_blks = 0;
1601 }
1602 curr_off = next_off;
1603 } while ((bh = bh->b_this_page) != head);
1604
1605 if (contiguous_blks) {
1606 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1607 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1608 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1609 }
1610
1611 /* If we have released all the blocks belonging to a cluster, then we
1612 * need to release the reserved space for that cluster. */
1613 num_clusters = EXT4_NUM_B2C(sbi, to_release);
1614 while (num_clusters > 0) {
1615 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1616 ((num_clusters - 1) << sbi->s_cluster_bits);
1617 if (sbi->s_cluster_ratio == 1 ||
1618 !ext4_find_delalloc_cluster(inode, lblk))
1619 ext4_da_release_space(inode, 1);
1620
1621 num_clusters--;
1622 }
1623}
1624
1625/*
1626 * Delayed allocation stuff
1627 */
1628
1629struct mpage_da_data {
1630 struct inode *inode;
1631 struct writeback_control *wbc;
1632
1633 pgoff_t first_page; /* The first page to write */
1634 pgoff_t next_page; /* Current page to examine */
1635 pgoff_t last_page; /* Last page to examine */
1636 /*
1637 * Extent to map - this can be after first_page because that can be
1638 * fully mapped. We somewhat abuse m_flags to store whether the extent
1639 * is delalloc or unwritten.
1640 */
1641 struct ext4_map_blocks map;
1642 struct ext4_io_submit io_submit; /* IO submission data */
1643};
1644
1645static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1646 bool invalidate)
1647{
1648 int nr_pages, i;
1649 pgoff_t index, end;
1650 struct pagevec pvec;
1651 struct inode *inode = mpd->inode;
1652 struct address_space *mapping = inode->i_mapping;
1653
1654 /* This is necessary when next_page == 0. */
1655 if (mpd->first_page >= mpd->next_page)
1656 return;
1657
1658 index = mpd->first_page;
1659 end = mpd->next_page - 1;
1660 if (invalidate) {
1661 ext4_lblk_t start, last;
1662 start = index << (PAGE_SHIFT - inode->i_blkbits);
1663 last = end << (PAGE_SHIFT - inode->i_blkbits);
1664 ext4_es_remove_extent(inode, start, last - start + 1);
1665 }
1666
1667 pagevec_init(&pvec, 0);
1668 while (index <= end) {
1669 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1670 if (nr_pages == 0)
1671 break;
1672 for (i = 0; i < nr_pages; i++) {
1673 struct page *page = pvec.pages[i];
1674 if (page->index > end)
1675 break;
1676 BUG_ON(!PageLocked(page));
1677 BUG_ON(PageWriteback(page));
1678 if (invalidate) {
1679 if (page_mapped(page))
1680 clear_page_dirty_for_io(page);
1681 block_invalidatepage(page, 0, PAGE_SIZE);
1682 ClearPageUptodate(page);
1683 }
1684 unlock_page(page);
1685 }
1686 index = pvec.pages[nr_pages - 1]->index + 1;
1687 pagevec_release(&pvec);
1688 }
1689}
1690
1691static void ext4_print_free_blocks(struct inode *inode)
1692{
1693 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1694 struct super_block *sb = inode->i_sb;
1695 struct ext4_inode_info *ei = EXT4_I(inode);
1696
1697 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1698 EXT4_C2B(EXT4_SB(inode->i_sb),
1699 ext4_count_free_clusters(sb)));
1700 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1701 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1702 (long long) EXT4_C2B(EXT4_SB(sb),
1703 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1704 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1705 (long long) EXT4_C2B(EXT4_SB(sb),
1706 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1707 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1708 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1709 ei->i_reserved_data_blocks);
1710 return;
1711}
1712
1713static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1714{
1715 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1716}
1717
1718/*
1719 * This function is grabs code from the very beginning of
1720 * ext4_map_blocks, but assumes that the caller is from delayed write
1721 * time. This function looks up the requested blocks and sets the
1722 * buffer delay bit under the protection of i_data_sem.
1723 */
1724static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1725 struct ext4_map_blocks *map,
1726 struct buffer_head *bh)
1727{
1728 struct extent_status es;
1729 int retval;
1730 sector_t invalid_block = ~((sector_t) 0xffff);
1731#ifdef ES_AGGRESSIVE_TEST
1732 struct ext4_map_blocks orig_map;
1733
1734 memcpy(&orig_map, map, sizeof(*map));
1735#endif
1736
1737 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1738 invalid_block = ~0;
1739
1740 map->m_flags = 0;
1741 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1742 "logical block %lu\n", inode->i_ino, map->m_len,
1743 (unsigned long) map->m_lblk);
1744
1745 /* Lookup extent status tree firstly */
1746 if (ext4_es_lookup_extent(inode, iblock, &es)) {
1747 if (ext4_es_is_hole(&es)) {
1748 retval = 0;
1749 down_read(&EXT4_I(inode)->i_data_sem);
1750 goto add_delayed;
1751 }
1752
1753 /*
1754 * Delayed extent could be allocated by fallocate.
1755 * So we need to check it.
1756 */
1757 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1758 map_bh(bh, inode->i_sb, invalid_block);
1759 set_buffer_new(bh);
1760 set_buffer_delay(bh);
1761 return 0;
1762 }
1763
1764 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1765 retval = es.es_len - (iblock - es.es_lblk);
1766 if (retval > map->m_len)
1767 retval = map->m_len;
1768 map->m_len = retval;
1769 if (ext4_es_is_written(&es))
1770 map->m_flags |= EXT4_MAP_MAPPED;
1771 else if (ext4_es_is_unwritten(&es))
1772 map->m_flags |= EXT4_MAP_UNWRITTEN;
1773 else
1774 BUG_ON(1);
1775
1776#ifdef ES_AGGRESSIVE_TEST
1777 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1778#endif
1779 return retval;
1780 }
1781
1782 /*
1783 * Try to see if we can get the block without requesting a new
1784 * file system block.
1785 */
1786 down_read(&EXT4_I(inode)->i_data_sem);
1787 if (ext4_has_inline_data(inode))
1788 retval = 0;
1789 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1790 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1791 else
1792 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1793
1794add_delayed:
1795 if (retval == 0) {
1796 int ret;
1797 /*
1798 * XXX: __block_prepare_write() unmaps passed block,
1799 * is it OK?
1800 */
1801 /*
1802 * If the block was allocated from previously allocated cluster,
1803 * then we don't need to reserve it again. However we still need
1804 * to reserve metadata for every block we're going to write.
1805 */
1806 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1807 !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1808 ret = ext4_da_reserve_space(inode);
1809 if (ret) {
1810 /* not enough space to reserve */
1811 retval = ret;
1812 goto out_unlock;
1813 }
1814 }
1815
1816 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1817 ~0, EXTENT_STATUS_DELAYED);
1818 if (ret) {
1819 retval = ret;
1820 goto out_unlock;
1821 }
1822
1823 map_bh(bh, inode->i_sb, invalid_block);
1824 set_buffer_new(bh);
1825 set_buffer_delay(bh);
1826 } else if (retval > 0) {
1827 int ret;
1828 unsigned int status;
1829
1830 if (unlikely(retval != map->m_len)) {
1831 ext4_warning(inode->i_sb,
1832 "ES len assertion failed for inode "
1833 "%lu: retval %d != map->m_len %d",
1834 inode->i_ino, retval, map->m_len);
1835 WARN_ON(1);
1836 }
1837
1838 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1839 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1840 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1841 map->m_pblk, status);
1842 if (ret != 0)
1843 retval = ret;
1844 }
1845
1846out_unlock:
1847 up_read((&EXT4_I(inode)->i_data_sem));
1848
1849 return retval;
1850}
1851
1852/*
1853 * This is a special get_block_t callback which is used by
1854 * ext4_da_write_begin(). It will either return mapped block or
1855 * reserve space for a single block.
1856 *
1857 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1858 * We also have b_blocknr = -1 and b_bdev initialized properly
1859 *
1860 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1861 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1862 * initialized properly.
1863 */
1864int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1865 struct buffer_head *bh, int create)
1866{
1867 struct ext4_map_blocks map;
1868 int ret = 0;
1869
1870 BUG_ON(create == 0);
1871 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1872
1873 map.m_lblk = iblock;
1874 map.m_len = 1;
1875
1876 /*
1877 * first, we need to know whether the block is allocated already
1878 * preallocated blocks are unmapped but should treated
1879 * the same as allocated blocks.
1880 */
1881 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1882 if (ret <= 0)
1883 return ret;
1884
1885 map_bh(bh, inode->i_sb, map.m_pblk);
1886 ext4_update_bh_state(bh, map.m_flags);
1887
1888 if (buffer_unwritten(bh)) {
1889 /* A delayed write to unwritten bh should be marked
1890 * new and mapped. Mapped ensures that we don't do
1891 * get_block multiple times when we write to the same
1892 * offset and new ensures that we do proper zero out
1893 * for partial write.
1894 */
1895 set_buffer_new(bh);
1896 set_buffer_mapped(bh);
1897 }
1898 return 0;
1899}
1900
1901static int bget_one(handle_t *handle, struct buffer_head *bh)
1902{
1903 get_bh(bh);
1904 return 0;
1905}
1906
1907static int bput_one(handle_t *handle, struct buffer_head *bh)
1908{
1909 put_bh(bh);
1910 return 0;
1911}
1912
1913static int __ext4_journalled_writepage(struct page *page,
1914 unsigned int len)
1915{
1916 struct address_space *mapping = page->mapping;
1917 struct inode *inode = mapping->host;
1918 struct buffer_head *page_bufs = NULL;
1919 handle_t *handle = NULL;
1920 int ret = 0, err = 0;
1921 int inline_data = ext4_has_inline_data(inode);
1922 struct buffer_head *inode_bh = NULL;
1923
1924 ClearPageChecked(page);
1925
1926 if (inline_data) {
1927 BUG_ON(page->index != 0);
1928 BUG_ON(len > ext4_get_max_inline_size(inode));
1929 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1930 if (inode_bh == NULL)
1931 goto out;
1932 } else {
1933 page_bufs = page_buffers(page);
1934 if (!page_bufs) {
1935 BUG();
1936 goto out;
1937 }
1938 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1939 NULL, bget_one);
1940 }
1941 /*
1942 * We need to release the page lock before we start the
1943 * journal, so grab a reference so the page won't disappear
1944 * out from under us.
1945 */
1946 get_page(page);
1947 unlock_page(page);
1948
1949 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1950 ext4_writepage_trans_blocks(inode));
1951 if (IS_ERR(handle)) {
1952 ret = PTR_ERR(handle);
1953 put_page(page);
1954 goto out_no_pagelock;
1955 }
1956 BUG_ON(!ext4_handle_valid(handle));
1957
1958 lock_page(page);
1959 put_page(page);
1960 if (page->mapping != mapping) {
1961 /* The page got truncated from under us */
1962 ext4_journal_stop(handle);
1963 ret = 0;
1964 goto out;
1965 }
1966
1967 if (inline_data) {
1968 BUFFER_TRACE(inode_bh, "get write access");
1969 ret = ext4_journal_get_write_access(handle, inode_bh);
1970
1971 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1972
1973 } else {
1974 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1975 do_journal_get_write_access);
1976
1977 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1978 write_end_fn);
1979 }
1980 if (ret == 0)
1981 ret = err;
1982 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1983 err = ext4_journal_stop(handle);
1984 if (!ret)
1985 ret = err;
1986
1987 if (!ext4_has_inline_data(inode))
1988 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1989 NULL, bput_one);
1990 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1991out:
1992 unlock_page(page);
1993out_no_pagelock:
1994 brelse(inode_bh);
1995 return ret;
1996}
1997
1998/*
1999 * Note that we don't need to start a transaction unless we're journaling data
2000 * because we should have holes filled from ext4_page_mkwrite(). We even don't
2001 * need to file the inode to the transaction's list in ordered mode because if
2002 * we are writing back data added by write(), the inode is already there and if
2003 * we are writing back data modified via mmap(), no one guarantees in which
2004 * transaction the data will hit the disk. In case we are journaling data, we
2005 * cannot start transaction directly because transaction start ranks above page
2006 * lock so we have to do some magic.
2007 *
2008 * This function can get called via...
2009 * - ext4_writepages after taking page lock (have journal handle)
2010 * - journal_submit_inode_data_buffers (no journal handle)
2011 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2012 * - grab_page_cache when doing write_begin (have journal handle)
2013 *
2014 * We don't do any block allocation in this function. If we have page with
2015 * multiple blocks we need to write those buffer_heads that are mapped. This
2016 * is important for mmaped based write. So if we do with blocksize 1K
2017 * truncate(f, 1024);
2018 * a = mmap(f, 0, 4096);
2019 * a[0] = 'a';
2020 * truncate(f, 4096);
2021 * we have in the page first buffer_head mapped via page_mkwrite call back
2022 * but other buffer_heads would be unmapped but dirty (dirty done via the
2023 * do_wp_page). So writepage should write the first block. If we modify
2024 * the mmap area beyond 1024 we will again get a page_fault and the
2025 * page_mkwrite callback will do the block allocation and mark the
2026 * buffer_heads mapped.
2027 *
2028 * We redirty the page if we have any buffer_heads that is either delay or
2029 * unwritten in the page.
2030 *
2031 * We can get recursively called as show below.
2032 *
2033 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2034 * ext4_writepage()
2035 *
2036 * But since we don't do any block allocation we should not deadlock.
2037 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2038 */
2039static int ext4_writepage(struct page *page,
2040 struct writeback_control *wbc)
2041{
2042 int ret = 0;
2043 loff_t size;
2044 unsigned int len;
2045 struct buffer_head *page_bufs = NULL;
2046 struct inode *inode = page->mapping->host;
2047 struct ext4_io_submit io_submit;
2048 bool keep_towrite = false;
2049
2050 trace_ext4_writepage(page);
2051 size = i_size_read(inode);
2052 if (page->index == size >> PAGE_SHIFT)
2053 len = size & ~PAGE_MASK;
2054 else
2055 len = PAGE_SIZE;
2056
2057 page_bufs = page_buffers(page);
2058 /*
2059 * We cannot do block allocation or other extent handling in this
2060 * function. If there are buffers needing that, we have to redirty
2061 * the page. But we may reach here when we do a journal commit via
2062 * journal_submit_inode_data_buffers() and in that case we must write
2063 * allocated buffers to achieve data=ordered mode guarantees.
2064 *
2065 * Also, if there is only one buffer per page (the fs block
2066 * size == the page size), if one buffer needs block
2067 * allocation or needs to modify the extent tree to clear the
2068 * unwritten flag, we know that the page can't be written at
2069 * all, so we might as well refuse the write immediately.
2070 * Unfortunately if the block size != page size, we can't as
2071 * easily detect this case using ext4_walk_page_buffers(), but
2072 * for the extremely common case, this is an optimization that
2073 * skips a useless round trip through ext4_bio_write_page().
2074 */
2075 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2076 ext4_bh_delay_or_unwritten)) {
2077 redirty_page_for_writepage(wbc, page);
2078 if ((current->flags & PF_MEMALLOC) ||
2079 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2080 /*
2081 * For memory cleaning there's no point in writing only
2082 * some buffers. So just bail out. Warn if we came here
2083 * from direct reclaim.
2084 */
2085 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2086 == PF_MEMALLOC);
2087 unlock_page(page);
2088 return 0;
2089 }
2090 keep_towrite = true;
2091 }
2092
2093 if (PageChecked(page) && ext4_should_journal_data(inode))
2094 /*
2095 * It's mmapped pagecache. Add buffers and journal it. There
2096 * doesn't seem much point in redirtying the page here.
2097 */
2098 return __ext4_journalled_writepage(page, len);
2099
2100 ext4_io_submit_init(&io_submit, wbc);
2101 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2102 if (!io_submit.io_end) {
2103 redirty_page_for_writepage(wbc, page);
2104 unlock_page(page);
2105 return -ENOMEM;
2106 }
2107 ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2108 ext4_io_submit(&io_submit);
2109 /* Drop io_end reference we got from init */
2110 ext4_put_io_end_defer(io_submit.io_end);
2111 return ret;
2112}
2113
2114static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2115{
2116 int len;
2117 loff_t size = i_size_read(mpd->inode);
2118 int err;
2119
2120 BUG_ON(page->index != mpd->first_page);
2121 if (page->index == size >> PAGE_SHIFT)
2122 len = size & ~PAGE_MASK;
2123 else
2124 len = PAGE_SIZE;
2125 clear_page_dirty_for_io(page);
2126 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2127 if (!err)
2128 mpd->wbc->nr_to_write--;
2129 mpd->first_page++;
2130
2131 return err;
2132}
2133
2134#define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2135
2136/*
2137 * mballoc gives us at most this number of blocks...
2138 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2139 * The rest of mballoc seems to handle chunks up to full group size.
2140 */
2141#define MAX_WRITEPAGES_EXTENT_LEN 2048
2142
2143/*
2144 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2145 *
2146 * @mpd - extent of blocks
2147 * @lblk - logical number of the block in the file
2148 * @bh - buffer head we want to add to the extent
2149 *
2150 * The function is used to collect contig. blocks in the same state. If the
2151 * buffer doesn't require mapping for writeback and we haven't started the
2152 * extent of buffers to map yet, the function returns 'true' immediately - the
2153 * caller can write the buffer right away. Otherwise the function returns true
2154 * if the block has been added to the extent, false if the block couldn't be
2155 * added.
2156 */
2157static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2158 struct buffer_head *bh)
2159{
2160 struct ext4_map_blocks *map = &mpd->map;
2161
2162 /* Buffer that doesn't need mapping for writeback? */
2163 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2164 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2165 /* So far no extent to map => we write the buffer right away */
2166 if (map->m_len == 0)
2167 return true;
2168 return false;
2169 }
2170
2171 /* First block in the extent? */
2172 if (map->m_len == 0) {
2173 map->m_lblk = lblk;
2174 map->m_len = 1;
2175 map->m_flags = bh->b_state & BH_FLAGS;
2176 return true;
2177 }
2178
2179 /* Don't go larger than mballoc is willing to allocate */
2180 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2181 return false;
2182
2183 /* Can we merge the block to our big extent? */
2184 if (lblk == map->m_lblk + map->m_len &&
2185 (bh->b_state & BH_FLAGS) == map->m_flags) {
2186 map->m_len++;
2187 return true;
2188 }
2189 return false;
2190}
2191
2192/*
2193 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2194 *
2195 * @mpd - extent of blocks for mapping
2196 * @head - the first buffer in the page
2197 * @bh - buffer we should start processing from
2198 * @lblk - logical number of the block in the file corresponding to @bh
2199 *
2200 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2201 * the page for IO if all buffers in this page were mapped and there's no
2202 * accumulated extent of buffers to map or add buffers in the page to the
2203 * extent of buffers to map. The function returns 1 if the caller can continue
2204 * by processing the next page, 0 if it should stop adding buffers to the
2205 * extent to map because we cannot extend it anymore. It can also return value
2206 * < 0 in case of error during IO submission.
2207 */
2208static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2209 struct buffer_head *head,
2210 struct buffer_head *bh,
2211 ext4_lblk_t lblk)
2212{
2213 struct inode *inode = mpd->inode;
2214 int err;
2215 ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
2216 >> inode->i_blkbits;
2217
2218 do {
2219 BUG_ON(buffer_locked(bh));
2220
2221 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2222 /* Found extent to map? */
2223 if (mpd->map.m_len)
2224 return 0;
2225 /* Everything mapped so far and we hit EOF */
2226 break;
2227 }
2228 } while (lblk++, (bh = bh->b_this_page) != head);
2229 /* So far everything mapped? Submit the page for IO. */
2230 if (mpd->map.m_len == 0) {
2231 err = mpage_submit_page(mpd, head->b_page);
2232 if (err < 0)
2233 return err;
2234 }
2235 return lblk < blocks;
2236}
2237
2238/*
2239 * mpage_map_buffers - update buffers corresponding to changed extent and
2240 * submit fully mapped pages for IO
2241 *
2242 * @mpd - description of extent to map, on return next extent to map
2243 *
2244 * Scan buffers corresponding to changed extent (we expect corresponding pages
2245 * to be already locked) and update buffer state according to new extent state.
2246 * We map delalloc buffers to their physical location, clear unwritten bits,
2247 * and mark buffers as uninit when we perform writes to unwritten extents
2248 * and do extent conversion after IO is finished. If the last page is not fully
2249 * mapped, we update @map to the next extent in the last page that needs
2250 * mapping. Otherwise we submit the page for IO.
2251 */
2252static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2253{
2254 struct pagevec pvec;
2255 int nr_pages, i;
2256 struct inode *inode = mpd->inode;
2257 struct buffer_head *head, *bh;
2258 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2259 pgoff_t start, end;
2260 ext4_lblk_t lblk;
2261 sector_t pblock;
2262 int err;
2263
2264 start = mpd->map.m_lblk >> bpp_bits;
2265 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2266 lblk = start << bpp_bits;
2267 pblock = mpd->map.m_pblk;
2268
2269 pagevec_init(&pvec, 0);
2270 while (start <= end) {
2271 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2272 PAGEVEC_SIZE);
2273 if (nr_pages == 0)
2274 break;
2275 for (i = 0; i < nr_pages; i++) {
2276 struct page *page = pvec.pages[i];
2277
2278 if (page->index > end)
2279 break;
2280 /* Up to 'end' pages must be contiguous */
2281 BUG_ON(page->index != start);
2282 bh = head = page_buffers(page);
2283 do {
2284 if (lblk < mpd->map.m_lblk)
2285 continue;
2286 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2287 /*
2288 * Buffer after end of mapped extent.
2289 * Find next buffer in the page to map.
2290 */
2291 mpd->map.m_len = 0;
2292 mpd->map.m_flags = 0;
2293 /*
2294 * FIXME: If dioread_nolock supports
2295 * blocksize < pagesize, we need to make
2296 * sure we add size mapped so far to
2297 * io_end->size as the following call
2298 * can submit the page for IO.
2299 */
2300 err = mpage_process_page_bufs(mpd, head,
2301 bh, lblk);
2302 pagevec_release(&pvec);
2303 if (err > 0)
2304 err = 0;
2305 return err;
2306 }
2307 if (buffer_delay(bh)) {
2308 clear_buffer_delay(bh);
2309 bh->b_blocknr = pblock++;
2310 }
2311 clear_buffer_unwritten(bh);
2312 } while (lblk++, (bh = bh->b_this_page) != head);
2313
2314 /*
2315 * FIXME: This is going to break if dioread_nolock
2316 * supports blocksize < pagesize as we will try to
2317 * convert potentially unmapped parts of inode.
2318 */
2319 mpd->io_submit.io_end->size += PAGE_SIZE;
2320 /* Page fully mapped - let IO run! */
2321 err = mpage_submit_page(mpd, page);
2322 if (err < 0) {
2323 pagevec_release(&pvec);
2324 return err;
2325 }
2326 start++;
2327 }
2328 pagevec_release(&pvec);
2329 }
2330 /* Extent fully mapped and matches with page boundary. We are done. */
2331 mpd->map.m_len = 0;
2332 mpd->map.m_flags = 0;
2333 return 0;
2334}
2335
2336static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2337{
2338 struct inode *inode = mpd->inode;
2339 struct ext4_map_blocks *map = &mpd->map;
2340 int get_blocks_flags;
2341 int err, dioread_nolock;
2342
2343 trace_ext4_da_write_pages_extent(inode, map);
2344 /*
2345 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2346 * to convert an unwritten extent to be initialized (in the case
2347 * where we have written into one or more preallocated blocks). It is
2348 * possible that we're going to need more metadata blocks than
2349 * previously reserved. However we must not fail because we're in
2350 * writeback and there is nothing we can do about it so it might result
2351 * in data loss. So use reserved blocks to allocate metadata if
2352 * possible.
2353 *
2354 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2355 * the blocks in question are delalloc blocks. This indicates
2356 * that the blocks and quotas has already been checked when
2357 * the data was copied into the page cache.
2358 */
2359 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2360 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2361 EXT4_GET_BLOCKS_IO_SUBMIT;
2362 dioread_nolock = ext4_should_dioread_nolock(inode);
2363 if (dioread_nolock)
2364 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2365 if (map->m_flags & (1 << BH_Delay))
2366 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2367
2368 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2369 if (err < 0)
2370 return err;
2371 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2372 if (!mpd->io_submit.io_end->handle &&
2373 ext4_handle_valid(handle)) {
2374 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2375 handle->h_rsv_handle = NULL;
2376 }
2377 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2378 }
2379
2380 BUG_ON(map->m_len == 0);
2381 if (map->m_flags & EXT4_MAP_NEW) {
2382 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2383 map->m_len);
2384 }
2385 return 0;
2386}
2387
2388/*
2389 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2390 * mpd->len and submit pages underlying it for IO
2391 *
2392 * @handle - handle for journal operations
2393 * @mpd - extent to map
2394 * @give_up_on_write - we set this to true iff there is a fatal error and there
2395 * is no hope of writing the data. The caller should discard
2396 * dirty pages to avoid infinite loops.
2397 *
2398 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2399 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2400 * them to initialized or split the described range from larger unwritten
2401 * extent. Note that we need not map all the described range since allocation
2402 * can return less blocks or the range is covered by more unwritten extents. We
2403 * cannot map more because we are limited by reserved transaction credits. On
2404 * the other hand we always make sure that the last touched page is fully
2405 * mapped so that it can be written out (and thus forward progress is
2406 * guaranteed). After mapping we submit all mapped pages for IO.
2407 */
2408static int mpage_map_and_submit_extent(handle_t *handle,
2409 struct mpage_da_data *mpd,
2410 bool *give_up_on_write)
2411{
2412 struct inode *inode = mpd->inode;
2413 struct ext4_map_blocks *map = &mpd->map;
2414 int err;
2415 loff_t disksize;
2416 int progress = 0;
2417
2418 mpd->io_submit.io_end->offset =
2419 ((loff_t)map->m_lblk) << inode->i_blkbits;
2420 do {
2421 err = mpage_map_one_extent(handle, mpd);
2422 if (err < 0) {
2423 struct super_block *sb = inode->i_sb;
2424
2425 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2426 goto invalidate_dirty_pages;
2427 /*
2428 * Let the uper layers retry transient errors.
2429 * In the case of ENOSPC, if ext4_count_free_blocks()
2430 * is non-zero, a commit should free up blocks.
2431 */
2432 if ((err == -ENOMEM) ||
2433 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2434 if (progress)
2435 goto update_disksize;
2436 return err;
2437 }
2438 ext4_msg(sb, KERN_CRIT,
2439 "Delayed block allocation failed for "
2440 "inode %lu at logical offset %llu with"
2441 " max blocks %u with error %d",
2442 inode->i_ino,
2443 (unsigned long long)map->m_lblk,
2444 (unsigned)map->m_len, -err);
2445 ext4_msg(sb, KERN_CRIT,
2446 "This should not happen!! Data will "
2447 "be lost\n");
2448 if (err == -ENOSPC)
2449 ext4_print_free_blocks(inode);
2450 invalidate_dirty_pages:
2451 *give_up_on_write = true;
2452 return err;
2453 }
2454 progress = 1;
2455 /*
2456 * Update buffer state, submit mapped pages, and get us new
2457 * extent to map
2458 */
2459 err = mpage_map_and_submit_buffers(mpd);
2460 if (err < 0)
2461 goto update_disksize;
2462 } while (map->m_len);
2463
2464update_disksize:
2465 /*
2466 * Update on-disk size after IO is submitted. Races with
2467 * truncate are avoided by checking i_size under i_data_sem.
2468 */
2469 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2470 if (disksize > EXT4_I(inode)->i_disksize) {
2471 int err2;
2472 loff_t i_size;
2473
2474 down_write(&EXT4_I(inode)->i_data_sem);
2475 i_size = i_size_read(inode);
2476 if (disksize > i_size)
2477 disksize = i_size;
2478 if (disksize > EXT4_I(inode)->i_disksize)
2479 EXT4_I(inode)->i_disksize = disksize;
2480 err2 = ext4_mark_inode_dirty(handle, inode);
2481 up_write(&EXT4_I(inode)->i_data_sem);
2482 if (err2)
2483 ext4_error(inode->i_sb,
2484 "Failed to mark inode %lu dirty",
2485 inode->i_ino);
2486 if (!err)
2487 err = err2;
2488 }
2489 return err;
2490}
2491
2492/*
2493 * Calculate the total number of credits to reserve for one writepages
2494 * iteration. This is called from ext4_writepages(). We map an extent of
2495 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2496 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2497 * bpp - 1 blocks in bpp different extents.
2498 */
2499static int ext4_da_writepages_trans_blocks(struct inode *inode)
2500{
2501 int bpp = ext4_journal_blocks_per_page(inode);
2502
2503 return ext4_meta_trans_blocks(inode,
2504 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2505}
2506
2507/*
2508 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2509 * and underlying extent to map
2510 *
2511 * @mpd - where to look for pages
2512 *
2513 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2514 * IO immediately. When we find a page which isn't mapped we start accumulating
2515 * extent of buffers underlying these pages that needs mapping (formed by
2516 * either delayed or unwritten buffers). We also lock the pages containing
2517 * these buffers. The extent found is returned in @mpd structure (starting at
2518 * mpd->lblk with length mpd->len blocks).
2519 *
2520 * Note that this function can attach bios to one io_end structure which are
2521 * neither logically nor physically contiguous. Although it may seem as an
2522 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2523 * case as we need to track IO to all buffers underlying a page in one io_end.
2524 */
2525static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2526{
2527 struct address_space *mapping = mpd->inode->i_mapping;
2528 struct pagevec pvec;
2529 unsigned int nr_pages;
2530 long left = mpd->wbc->nr_to_write;
2531 pgoff_t index = mpd->first_page;
2532 pgoff_t end = mpd->last_page;
2533 int tag;
2534 int i, err = 0;
2535 int blkbits = mpd->inode->i_blkbits;
2536 ext4_lblk_t lblk;
2537 struct buffer_head *head;
2538
2539 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2540 tag = PAGECACHE_TAG_TOWRITE;
2541 else
2542 tag = PAGECACHE_TAG_DIRTY;
2543
2544 pagevec_init(&pvec, 0);
2545 mpd->map.m_len = 0;
2546 mpd->next_page = index;
2547 while (index <= end) {
2548 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2549 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2550 if (nr_pages == 0)
2551 goto out;
2552
2553 for (i = 0; i < nr_pages; i++) {
2554 struct page *page = pvec.pages[i];
2555
2556 /*
2557 * At this point, the page may be truncated or
2558 * invalidated (changing page->mapping to NULL), or
2559 * even swizzled back from swapper_space to tmpfs file
2560 * mapping. However, page->index will not change
2561 * because we have a reference on the page.
2562 */
2563 if (page->index > end)
2564 goto out;
2565
2566 /*
2567 * Accumulated enough dirty pages? This doesn't apply
2568 * to WB_SYNC_ALL mode. For integrity sync we have to
2569 * keep going because someone may be concurrently
2570 * dirtying pages, and we might have synced a lot of
2571 * newly appeared dirty pages, but have not synced all
2572 * of the old dirty pages.
2573 */
2574 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2575 goto out;
2576
2577 /* If we can't merge this page, we are done. */
2578 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2579 goto out;
2580
2581 lock_page(page);
2582 /*
2583 * If the page is no longer dirty, or its mapping no
2584 * longer corresponds to inode we are writing (which
2585 * means it has been truncated or invalidated), or the
2586 * page is already under writeback and we are not doing
2587 * a data integrity writeback, skip the page
2588 */
2589 if (!PageDirty(page) ||
2590 (PageWriteback(page) &&
2591 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2592 unlikely(page->mapping != mapping)) {
2593 unlock_page(page);
2594 continue;
2595 }
2596
2597 wait_on_page_writeback(page);
2598 BUG_ON(PageWriteback(page));
2599
2600 if (mpd->map.m_len == 0)
2601 mpd->first_page = page->index;
2602 mpd->next_page = page->index + 1;
2603 /* Add all dirty buffers to mpd */
2604 lblk = ((ext4_lblk_t)page->index) <<
2605 (PAGE_SHIFT - blkbits);
2606 head = page_buffers(page);
2607 err = mpage_process_page_bufs(mpd, head, head, lblk);
2608 if (err <= 0)
2609 goto out;
2610 err = 0;
2611 left--;
2612 }
2613 pagevec_release(&pvec);
2614 cond_resched();
2615 }
2616 return 0;
2617out:
2618 pagevec_release(&pvec);
2619 return err;
2620}
2621
2622static int __writepage(struct page *page, struct writeback_control *wbc,
2623 void *data)
2624{
2625 struct address_space *mapping = data;
2626 int ret = ext4_writepage(page, wbc);
2627 mapping_set_error(mapping, ret);
2628 return ret;
2629}
2630
2631static int ext4_writepages(struct address_space *mapping,
2632 struct writeback_control *wbc)
2633{
2634 pgoff_t writeback_index = 0;
2635 long nr_to_write = wbc->nr_to_write;
2636 int range_whole = 0;
2637 int cycled = 1;
2638 handle_t *handle = NULL;
2639 struct mpage_da_data mpd;
2640 struct inode *inode = mapping->host;
2641 int needed_blocks, rsv_blocks = 0, ret = 0;
2642 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2643 bool done;
2644 struct blk_plug plug;
2645 bool give_up_on_write = false;
2646
2647 percpu_down_read(&sbi->s_journal_flag_rwsem);
2648 trace_ext4_writepages(inode, wbc);
2649
2650 if (dax_mapping(mapping)) {
2651 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2652 wbc);
2653 goto out_writepages;
2654 }
2655
2656 /*
2657 * No pages to write? This is mainly a kludge to avoid starting
2658 * a transaction for special inodes like journal inode on last iput()
2659 * because that could violate lock ordering on umount
2660 */
2661 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2662 goto out_writepages;
2663
2664 if (ext4_should_journal_data(inode)) {
2665 struct blk_plug plug;
2666
2667 blk_start_plug(&plug);
2668 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2669 blk_finish_plug(&plug);
2670 goto out_writepages;
2671 }
2672
2673 /*
2674 * If the filesystem has aborted, it is read-only, so return
2675 * right away instead of dumping stack traces later on that
2676 * will obscure the real source of the problem. We test
2677 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2678 * the latter could be true if the filesystem is mounted
2679 * read-only, and in that case, ext4_writepages should
2680 * *never* be called, so if that ever happens, we would want
2681 * the stack trace.
2682 */
2683 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2684 ret = -EROFS;
2685 goto out_writepages;
2686 }
2687
2688 if (ext4_should_dioread_nolock(inode)) {
2689 /*
2690 * We may need to convert up to one extent per block in
2691 * the page and we may dirty the inode.
2692 */
2693 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2694 }
2695
2696 /*
2697 * If we have inline data and arrive here, it means that
2698 * we will soon create the block for the 1st page, so
2699 * we'd better clear the inline data here.
2700 */
2701 if (ext4_has_inline_data(inode)) {
2702 /* Just inode will be modified... */
2703 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2704 if (IS_ERR(handle)) {
2705 ret = PTR_ERR(handle);
2706 goto out_writepages;
2707 }
2708 BUG_ON(ext4_test_inode_state(inode,
2709 EXT4_STATE_MAY_INLINE_DATA));
2710 ext4_destroy_inline_data(handle, inode);
2711 ext4_journal_stop(handle);
2712 }
2713
2714 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2715 range_whole = 1;
2716
2717 if (wbc->range_cyclic) {
2718 writeback_index = mapping->writeback_index;
2719 if (writeback_index)
2720 cycled = 0;
2721 mpd.first_page = writeback_index;
2722 mpd.last_page = -1;
2723 } else {
2724 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2725 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2726 }
2727
2728 mpd.inode = inode;
2729 mpd.wbc = wbc;
2730 ext4_io_submit_init(&mpd.io_submit, wbc);
2731retry:
2732 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2733 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2734 done = false;
2735 blk_start_plug(&plug);
2736 while (!done && mpd.first_page <= mpd.last_page) {
2737 /* For each extent of pages we use new io_end */
2738 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2739 if (!mpd.io_submit.io_end) {
2740 ret = -ENOMEM;
2741 break;
2742 }
2743
2744 /*
2745 * We have two constraints: We find one extent to map and we
2746 * must always write out whole page (makes a difference when
2747 * blocksize < pagesize) so that we don't block on IO when we
2748 * try to write out the rest of the page. Journalled mode is
2749 * not supported by delalloc.
2750 */
2751 BUG_ON(ext4_should_journal_data(inode));
2752 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2753
2754 /* start a new transaction */
2755 handle = ext4_journal_start_with_reserve(inode,
2756 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2757 if (IS_ERR(handle)) {
2758 ret = PTR_ERR(handle);
2759 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2760 "%ld pages, ino %lu; err %d", __func__,
2761 wbc->nr_to_write, inode->i_ino, ret);
2762 /* Release allocated io_end */
2763 ext4_put_io_end(mpd.io_submit.io_end);
2764 break;
2765 }
2766
2767 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2768 ret = mpage_prepare_extent_to_map(&mpd);
2769 if (!ret) {
2770 if (mpd.map.m_len)
2771 ret = mpage_map_and_submit_extent(handle, &mpd,
2772 &give_up_on_write);
2773 else {
2774 /*
2775 * We scanned the whole range (or exhausted
2776 * nr_to_write), submitted what was mapped and
2777 * didn't find anything needing mapping. We are
2778 * done.
2779 */
2780 done = true;
2781 }
2782 }
2783 /*
2784 * Caution: If the handle is synchronous,
2785 * ext4_journal_stop() can wait for transaction commit
2786 * to finish which may depend on writeback of pages to
2787 * complete or on page lock to be released. In that
2788 * case, we have to wait until after after we have
2789 * submitted all the IO, released page locks we hold,
2790 * and dropped io_end reference (for extent conversion
2791 * to be able to complete) before stopping the handle.
2792 */
2793 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2794 ext4_journal_stop(handle);
2795 handle = NULL;
2796 }
2797 /* Submit prepared bio */
2798 ext4_io_submit(&mpd.io_submit);
2799 /* Unlock pages we didn't use */
2800 mpage_release_unused_pages(&mpd, give_up_on_write);
2801 /*
2802 * Drop our io_end reference we got from init. We have
2803 * to be careful and use deferred io_end finishing if
2804 * we are still holding the transaction as we can
2805 * release the last reference to io_end which may end
2806 * up doing unwritten extent conversion.
2807 */
2808 if (handle) {
2809 ext4_put_io_end_defer(mpd.io_submit.io_end);
2810 ext4_journal_stop(handle);
2811 } else
2812 ext4_put_io_end(mpd.io_submit.io_end);
2813
2814 if (ret == -ENOSPC && sbi->s_journal) {
2815 /*
2816 * Commit the transaction which would
2817 * free blocks released in the transaction
2818 * and try again
2819 */
2820 jbd2_journal_force_commit_nested(sbi->s_journal);
2821 ret = 0;
2822 continue;
2823 }
2824 /* Fatal error - ENOMEM, EIO... */
2825 if (ret)
2826 break;
2827 }
2828 blk_finish_plug(&plug);
2829 if (!ret && !cycled && wbc->nr_to_write > 0) {
2830 cycled = 1;
2831 mpd.last_page = writeback_index - 1;
2832 mpd.first_page = 0;
2833 goto retry;
2834 }
2835
2836 /* Update index */
2837 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2838 /*
2839 * Set the writeback_index so that range_cyclic
2840 * mode will write it back later
2841 */
2842 mapping->writeback_index = mpd.first_page;
2843
2844out_writepages:
2845 trace_ext4_writepages_result(inode, wbc, ret,
2846 nr_to_write - wbc->nr_to_write);
2847 percpu_up_read(&sbi->s_journal_flag_rwsem);
2848 return ret;
2849}
2850
2851static int ext4_nonda_switch(struct super_block *sb)
2852{
2853 s64 free_clusters, dirty_clusters;
2854 struct ext4_sb_info *sbi = EXT4_SB(sb);
2855
2856 /*
2857 * switch to non delalloc mode if we are running low
2858 * on free block. The free block accounting via percpu
2859 * counters can get slightly wrong with percpu_counter_batch getting
2860 * accumulated on each CPU without updating global counters
2861 * Delalloc need an accurate free block accounting. So switch
2862 * to non delalloc when we are near to error range.
2863 */
2864 free_clusters =
2865 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2866 dirty_clusters =
2867 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2868 /*
2869 * Start pushing delalloc when 1/2 of free blocks are dirty.
2870 */
2871 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2872 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2873
2874 if (2 * free_clusters < 3 * dirty_clusters ||
2875 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2876 /*
2877 * free block count is less than 150% of dirty blocks
2878 * or free blocks is less than watermark
2879 */
2880 return 1;
2881 }
2882 return 0;
2883}
2884
2885/* We always reserve for an inode update; the superblock could be there too */
2886static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2887{
2888 if (likely(ext4_has_feature_large_file(inode->i_sb)))
2889 return 1;
2890
2891 if (pos + len <= 0x7fffffffULL)
2892 return 1;
2893
2894 /* We might need to update the superblock to set LARGE_FILE */
2895 return 2;
2896}
2897
2898static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2899 loff_t pos, unsigned len, unsigned flags,
2900 struct page **pagep, void **fsdata)
2901{
2902 int ret, retries = 0;
2903 struct page *page;
2904 pgoff_t index;
2905 struct inode *inode = mapping->host;
2906 handle_t *handle;
2907
2908 index = pos >> PAGE_SHIFT;
2909
2910 if (ext4_nonda_switch(inode->i_sb) ||
2911 S_ISLNK(inode->i_mode)) {
2912 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2913 return ext4_write_begin(file, mapping, pos,
2914 len, flags, pagep, fsdata);
2915 }
2916 *fsdata = (void *)0;
2917 trace_ext4_da_write_begin(inode, pos, len, flags);
2918
2919 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2920 ret = ext4_da_write_inline_data_begin(mapping, inode,
2921 pos, len, flags,
2922 pagep, fsdata);
2923 if (ret < 0)
2924 return ret;
2925 if (ret == 1)
2926 return 0;
2927 }
2928
2929 /*
2930 * grab_cache_page_write_begin() can take a long time if the
2931 * system is thrashing due to memory pressure, or if the page
2932 * is being written back. So grab it first before we start
2933 * the transaction handle. This also allows us to allocate
2934 * the page (if needed) without using GFP_NOFS.
2935 */
2936retry_grab:
2937 page = grab_cache_page_write_begin(mapping, index, flags);
2938 if (!page)
2939 return -ENOMEM;
2940 unlock_page(page);
2941
2942 /*
2943 * With delayed allocation, we don't log the i_disksize update
2944 * if there is delayed block allocation. But we still need
2945 * to journalling the i_disksize update if writes to the end
2946 * of file which has an already mapped buffer.
2947 */
2948retry_journal:
2949 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2950 ext4_da_write_credits(inode, pos, len));
2951 if (IS_ERR(handle)) {
2952 put_page(page);
2953 return PTR_ERR(handle);
2954 }
2955
2956 lock_page(page);
2957 if (page->mapping != mapping) {
2958 /* The page got truncated from under us */
2959 unlock_page(page);
2960 put_page(page);
2961 ext4_journal_stop(handle);
2962 goto retry_grab;
2963 }
2964 /* In case writeback began while the page was unlocked */
2965 wait_for_stable_page(page);
2966
2967#ifdef CONFIG_EXT4_FS_ENCRYPTION
2968 ret = ext4_block_write_begin(page, pos, len,
2969 ext4_da_get_block_prep);
2970#else
2971 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2972#endif
2973 if (ret < 0) {
2974 unlock_page(page);
2975 ext4_journal_stop(handle);
2976 /*
2977 * block_write_begin may have instantiated a few blocks
2978 * outside i_size. Trim these off again. Don't need
2979 * i_size_read because we hold i_mutex.
2980 */
2981 if (pos + len > inode->i_size)
2982 ext4_truncate_failed_write(inode);
2983
2984 if (ret == -ENOSPC &&
2985 ext4_should_retry_alloc(inode->i_sb, &retries))
2986 goto retry_journal;
2987
2988 put_page(page);
2989 return ret;
2990 }
2991
2992 *pagep = page;
2993 return ret;
2994}
2995
2996/*
2997 * Check if we should update i_disksize
2998 * when write to the end of file but not require block allocation
2999 */
3000static int ext4_da_should_update_i_disksize(struct page *page,
3001 unsigned long offset)
3002{
3003 struct buffer_head *bh;
3004 struct inode *inode = page->mapping->host;
3005 unsigned int idx;
3006 int i;
3007
3008 bh = page_buffers(page);
3009 idx = offset >> inode->i_blkbits;
3010
3011 for (i = 0; i < idx; i++)
3012 bh = bh->b_this_page;
3013
3014 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3015 return 0;
3016 return 1;
3017}
3018
3019static int ext4_da_write_end(struct file *file,
3020 struct address_space *mapping,
3021 loff_t pos, unsigned len, unsigned copied,
3022 struct page *page, void *fsdata)
3023{
3024 struct inode *inode = mapping->host;
3025 int ret = 0, ret2;
3026 handle_t *handle = ext4_journal_current_handle();
3027 loff_t new_i_size;
3028 unsigned long start, end;
3029 int write_mode = (int)(unsigned long)fsdata;
3030
3031 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3032 return ext4_write_end(file, mapping, pos,
3033 len, copied, page, fsdata);
3034
3035 trace_ext4_da_write_end(inode, pos, len, copied);
3036 start = pos & (PAGE_SIZE - 1);
3037 end = start + copied - 1;
3038
3039 /*
3040 * generic_write_end() will run mark_inode_dirty() if i_size
3041 * changes. So let's piggyback the i_disksize mark_inode_dirty
3042 * into that.
3043 */
3044 new_i_size = pos + copied;
3045 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3046 if (ext4_has_inline_data(inode) ||
3047 ext4_da_should_update_i_disksize(page, end)) {
3048 ext4_update_i_disksize(inode, new_i_size);
3049 /* We need to mark inode dirty even if
3050 * new_i_size is less that inode->i_size
3051 * bu greater than i_disksize.(hint delalloc)
3052 */
3053 ext4_mark_inode_dirty(handle, inode);
3054 }
3055 }
3056
3057 if (write_mode != CONVERT_INLINE_DATA &&
3058 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3059 ext4_has_inline_data(inode))
3060 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3061 page);
3062 else
3063 ret2 = generic_write_end(file, mapping, pos, len, copied,
3064 page, fsdata);
3065
3066 copied = ret2;
3067 if (ret2 < 0)
3068 ret = ret2;
3069 ret2 = ext4_journal_stop(handle);
3070 if (!ret)
3071 ret = ret2;
3072
3073 return ret ? ret : copied;
3074}
3075
3076static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3077 unsigned int length)
3078{
3079 /*
3080 * Drop reserved blocks
3081 */
3082 BUG_ON(!PageLocked(page));
3083 if (!page_has_buffers(page))
3084 goto out;
3085
3086 ext4_da_page_release_reservation(page, offset, length);
3087
3088out:
3089 ext4_invalidatepage(page, offset, length);
3090
3091 return;
3092}
3093
3094/*
3095 * Force all delayed allocation blocks to be allocated for a given inode.
3096 */
3097int ext4_alloc_da_blocks(struct inode *inode)
3098{
3099 trace_ext4_alloc_da_blocks(inode);
3100
3101 if (!EXT4_I(inode)->i_reserved_data_blocks)
3102 return 0;
3103
3104 /*
3105 * We do something simple for now. The filemap_flush() will
3106 * also start triggering a write of the data blocks, which is
3107 * not strictly speaking necessary (and for users of
3108 * laptop_mode, not even desirable). However, to do otherwise
3109 * would require replicating code paths in:
3110 *
3111 * ext4_writepages() ->
3112 * write_cache_pages() ---> (via passed in callback function)
3113 * __mpage_da_writepage() -->
3114 * mpage_add_bh_to_extent()
3115 * mpage_da_map_blocks()
3116 *
3117 * The problem is that write_cache_pages(), located in
3118 * mm/page-writeback.c, marks pages clean in preparation for
3119 * doing I/O, which is not desirable if we're not planning on
3120 * doing I/O at all.
3121 *
3122 * We could call write_cache_pages(), and then redirty all of
3123 * the pages by calling redirty_page_for_writepage() but that
3124 * would be ugly in the extreme. So instead we would need to
3125 * replicate parts of the code in the above functions,
3126 * simplifying them because we wouldn't actually intend to
3127 * write out the pages, but rather only collect contiguous
3128 * logical block extents, call the multi-block allocator, and
3129 * then update the buffer heads with the block allocations.
3130 *
3131 * For now, though, we'll cheat by calling filemap_flush(),
3132 * which will map the blocks, and start the I/O, but not
3133 * actually wait for the I/O to complete.
3134 */
3135 return filemap_flush(inode->i_mapping);
3136}
3137
3138/*
3139 * bmap() is special. It gets used by applications such as lilo and by
3140 * the swapper to find the on-disk block of a specific piece of data.
3141 *
3142 * Naturally, this is dangerous if the block concerned is still in the
3143 * journal. If somebody makes a swapfile on an ext4 data-journaling
3144 * filesystem and enables swap, then they may get a nasty shock when the
3145 * data getting swapped to that swapfile suddenly gets overwritten by
3146 * the original zero's written out previously to the journal and
3147 * awaiting writeback in the kernel's buffer cache.
3148 *
3149 * So, if we see any bmap calls here on a modified, data-journaled file,
3150 * take extra steps to flush any blocks which might be in the cache.
3151 */
3152static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3153{
3154 struct inode *inode = mapping->host;
3155 journal_t *journal;
3156 int err;
3157
3158 /*
3159 * We can get here for an inline file via the FIBMAP ioctl
3160 */
3161 if (ext4_has_inline_data(inode))
3162 return 0;
3163
3164 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3165 test_opt(inode->i_sb, DELALLOC)) {
3166 /*
3167 * With delalloc we want to sync the file
3168 * so that we can make sure we allocate
3169 * blocks for file
3170 */
3171 filemap_write_and_wait(mapping);
3172 }
3173
3174 if (EXT4_JOURNAL(inode) &&
3175 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3176 /*
3177 * This is a REALLY heavyweight approach, but the use of
3178 * bmap on dirty files is expected to be extremely rare:
3179 * only if we run lilo or swapon on a freshly made file
3180 * do we expect this to happen.
3181 *
3182 * (bmap requires CAP_SYS_RAWIO so this does not
3183 * represent an unprivileged user DOS attack --- we'd be
3184 * in trouble if mortal users could trigger this path at
3185 * will.)
3186 *
3187 * NB. EXT4_STATE_JDATA is not set on files other than
3188 * regular files. If somebody wants to bmap a directory
3189 * or symlink and gets confused because the buffer
3190 * hasn't yet been flushed to disk, they deserve
3191 * everything they get.
3192 */
3193
3194 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3195 journal = EXT4_JOURNAL(inode);
3196 jbd2_journal_lock_updates(journal);
3197 err = jbd2_journal_flush(journal);
3198 jbd2_journal_unlock_updates(journal);
3199
3200 if (err)
3201 return 0;
3202 }
3203
3204 return generic_block_bmap(mapping, block, ext4_get_block);
3205}
3206
3207static int ext4_readpage(struct file *file, struct page *page)
3208{
3209 int ret = -EAGAIN;
3210 struct inode *inode = page->mapping->host;
3211
3212 trace_ext4_readpage(page);
3213
3214 if (ext4_has_inline_data(inode))
3215 ret = ext4_readpage_inline(inode, page);
3216
3217 if (ret == -EAGAIN)
3218 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3219
3220 return ret;
3221}
3222
3223static int
3224ext4_readpages(struct file *file, struct address_space *mapping,
3225 struct list_head *pages, unsigned nr_pages)
3226{
3227 struct inode *inode = mapping->host;
3228
3229 /* If the file has inline data, no need to do readpages. */
3230 if (ext4_has_inline_data(inode))
3231 return 0;
3232
3233 return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3234}
3235
3236static void ext4_invalidatepage(struct page *page, unsigned int offset,
3237 unsigned int length)
3238{
3239 trace_ext4_invalidatepage(page, offset, length);
3240
3241 /* No journalling happens on data buffers when this function is used */
3242 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3243
3244 block_invalidatepage(page, offset, length);
3245}
3246
3247static int __ext4_journalled_invalidatepage(struct page *page,
3248 unsigned int offset,
3249 unsigned int length)
3250{
3251 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3252
3253 trace_ext4_journalled_invalidatepage(page, offset, length);
3254
3255 /*
3256 * If it's a full truncate we just forget about the pending dirtying
3257 */
3258 if (offset == 0 && length == PAGE_SIZE)
3259 ClearPageChecked(page);
3260
3261 return jbd2_journal_invalidatepage(journal, page, offset, length);
3262}
3263
3264/* Wrapper for aops... */
3265static void ext4_journalled_invalidatepage(struct page *page,
3266 unsigned int offset,
3267 unsigned int length)
3268{
3269 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3270}
3271
3272static int ext4_releasepage(struct page *page, gfp_t wait)
3273{
3274 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3275
3276 trace_ext4_releasepage(page);
3277
3278 /* Page has dirty journalled data -> cannot release */
3279 if (PageChecked(page))
3280 return 0;
3281 if (journal)
3282 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3283 else
3284 return try_to_free_buffers(page);
3285}
3286
3287#ifdef CONFIG_FS_DAX
3288static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3289 unsigned flags, struct iomap *iomap)
3290{
3291 unsigned int blkbits = inode->i_blkbits;
3292 unsigned long first_block = offset >> blkbits;
3293 unsigned long last_block = (offset + length - 1) >> blkbits;
3294 struct ext4_map_blocks map;
3295 int ret;
3296
3297 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3298 return -ERANGE;
3299
3300 map.m_lblk = first_block;
3301 map.m_len = last_block - first_block + 1;
3302
3303 if (!(flags & IOMAP_WRITE)) {
3304 ret = ext4_map_blocks(NULL, inode, &map, 0);
3305 } else {
3306 int dio_credits;
3307 handle_t *handle;
3308 int retries = 0;
3309
3310 /* Trim mapping request to maximum we can map at once for DIO */
3311 if (map.m_len > DIO_MAX_BLOCKS)
3312 map.m_len = DIO_MAX_BLOCKS;
3313 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3314retry:
3315 /*
3316 * Either we allocate blocks and then we don't get unwritten
3317 * extent so we have reserved enough credits, or the blocks
3318 * are already allocated and unwritten and in that case
3319 * extent conversion fits in the credits as well.
3320 */
3321 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3322 dio_credits);
3323 if (IS_ERR(handle))
3324 return PTR_ERR(handle);
3325
3326 ret = ext4_map_blocks(handle, inode, &map,
3327 EXT4_GET_BLOCKS_CREATE_ZERO);
3328 if (ret < 0) {
3329 ext4_journal_stop(handle);
3330 if (ret == -ENOSPC &&
3331 ext4_should_retry_alloc(inode->i_sb, &retries))
3332 goto retry;
3333 return ret;
3334 }
3335
3336 /*
3337 * If we added blocks beyond i_size, we need to make sure they
3338 * will get truncated if we crash before updating i_size in
3339 * ext4_iomap_end(). For faults we don't need to do that (and
3340 * even cannot because for orphan list operations inode_lock is
3341 * required) - if we happen to instantiate block beyond i_size,
3342 * it is because we race with truncate which has already added
3343 * the inode to the orphan list.
3344 */
3345 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3346 (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3347 int err;
3348
3349 err = ext4_orphan_add(handle, inode);
3350 if (err < 0) {
3351 ext4_journal_stop(handle);
3352 return err;
3353 }
3354 }
3355 ext4_journal_stop(handle);
3356 }
3357
3358 iomap->flags = 0;
3359 iomap->bdev = inode->i_sb->s_bdev;
3360 iomap->offset = first_block << blkbits;
3361
3362 if (ret == 0) {
3363 iomap->type = IOMAP_HOLE;
3364 iomap->blkno = IOMAP_NULL_BLOCK;
3365 iomap->length = (u64)map.m_len << blkbits;
3366 } else {
3367 if (map.m_flags & EXT4_MAP_MAPPED) {
3368 iomap->type = IOMAP_MAPPED;
3369 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3370 iomap->type = IOMAP_UNWRITTEN;
3371 } else {
3372 WARN_ON_ONCE(1);
3373 return -EIO;
3374 }
3375 iomap->blkno = (sector_t)map.m_pblk << (blkbits - 9);
3376 iomap->length = (u64)map.m_len << blkbits;
3377 }
3378
3379 if (map.m_flags & EXT4_MAP_NEW)
3380 iomap->flags |= IOMAP_F_NEW;
3381 return 0;
3382}
3383
3384static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3385 ssize_t written, unsigned flags, struct iomap *iomap)
3386{
3387 int ret = 0;
3388 handle_t *handle;
3389 int blkbits = inode->i_blkbits;
3390 bool truncate = false;
3391
3392 if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3393 return 0;
3394
3395 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3396 if (IS_ERR(handle)) {
3397 ret = PTR_ERR(handle);
3398 goto orphan_del;
3399 }
3400 if (ext4_update_inode_size(inode, offset + written))
3401 ext4_mark_inode_dirty(handle, inode);
3402 /*
3403 * We may need to truncate allocated but not written blocks beyond EOF.
3404 */
3405 if (iomap->offset + iomap->length >
3406 ALIGN(inode->i_size, 1 << blkbits)) {
3407 ext4_lblk_t written_blk, end_blk;
3408
3409 written_blk = (offset + written) >> blkbits;
3410 end_blk = (offset + length) >> blkbits;
3411 if (written_blk < end_blk && ext4_can_truncate(inode))
3412 truncate = true;
3413 }
3414 /*
3415 * Remove inode from orphan list if we were extending a inode and
3416 * everything went fine.
3417 */
3418 if (!truncate && inode->i_nlink &&
3419 !list_empty(&EXT4_I(inode)->i_orphan))
3420 ext4_orphan_del(handle, inode);
3421 ext4_journal_stop(handle);
3422 if (truncate) {
3423 ext4_truncate_failed_write(inode);
3424orphan_del:
3425 /*
3426 * If truncate failed early the inode might still be on the
3427 * orphan list; we need to make sure the inode is removed from
3428 * the orphan list in that case.
3429 */
3430 if (inode->i_nlink)
3431 ext4_orphan_del(NULL, inode);
3432 }
3433 return ret;
3434}
3435
3436struct iomap_ops ext4_iomap_ops = {
3437 .iomap_begin = ext4_iomap_begin,
3438 .iomap_end = ext4_iomap_end,
3439};
3440
3441#endif
3442
3443static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3444 ssize_t size, void *private)
3445{
3446 ext4_io_end_t *io_end = private;
3447
3448 /* if not async direct IO just return */
3449 if (!io_end)
3450 return 0;
3451
3452 ext_debug("ext4_end_io_dio(): io_end 0x%p "
3453 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3454 io_end, io_end->inode->i_ino, iocb, offset, size);
3455
3456 /*
3457 * Error during AIO DIO. We cannot convert unwritten extents as the
3458 * data was not written. Just clear the unwritten flag and drop io_end.
3459 */
3460 if (size <= 0) {
3461 ext4_clear_io_unwritten_flag(io_end);
3462 size = 0;
3463 }
3464 io_end->offset = offset;
3465 io_end->size = size;
3466 ext4_put_io_end(io_end);
3467
3468 return 0;
3469}
3470
3471/*
3472 * Handling of direct IO writes.
3473 *
3474 * For ext4 extent files, ext4 will do direct-io write even to holes,
3475 * preallocated extents, and those write extend the file, no need to
3476 * fall back to buffered IO.
3477 *
3478 * For holes, we fallocate those blocks, mark them as unwritten
3479 * If those blocks were preallocated, we mark sure they are split, but
3480 * still keep the range to write as unwritten.
3481 *
3482 * The unwritten extents will be converted to written when DIO is completed.
3483 * For async direct IO, since the IO may still pending when return, we
3484 * set up an end_io call back function, which will do the conversion
3485 * when async direct IO completed.
3486 *
3487 * If the O_DIRECT write will extend the file then add this inode to the
3488 * orphan list. So recovery will truncate it back to the original size
3489 * if the machine crashes during the write.
3490 *
3491 */
3492static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3493{
3494 struct file *file = iocb->ki_filp;
3495 struct inode *inode = file->f_mapping->host;
3496 struct ext4_inode_info *ei = EXT4_I(inode);
3497 ssize_t ret;
3498 loff_t offset = iocb->ki_pos;
3499 size_t count = iov_iter_count(iter);
3500 int overwrite = 0;
3501 get_block_t *get_block_func = NULL;
3502 int dio_flags = 0;
3503 loff_t final_size = offset + count;
3504 int orphan = 0;
3505 handle_t *handle;
3506
3507 if (final_size > inode->i_size) {
3508 /* Credits for sb + inode write */
3509 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3510 if (IS_ERR(handle)) {
3511 ret = PTR_ERR(handle);
3512 goto out;
3513 }
3514 ret = ext4_orphan_add(handle, inode);
3515 if (ret) {
3516 ext4_journal_stop(handle);
3517 goto out;
3518 }
3519 orphan = 1;
3520 ei->i_disksize = inode->i_size;
3521 ext4_journal_stop(handle);
3522 }
3523
3524 BUG_ON(iocb->private == NULL);
3525
3526 /*
3527 * Make all waiters for direct IO properly wait also for extent
3528 * conversion. This also disallows race between truncate() and
3529 * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3530 */
3531 inode_dio_begin(inode);
3532
3533 /* If we do a overwrite dio, i_mutex locking can be released */
3534 overwrite = *((int *)iocb->private);
3535
3536 if (overwrite)
3537 inode_unlock(inode);
3538
3539 /*
3540 * For extent mapped files we could direct write to holes and fallocate.
3541 *
3542 * Allocated blocks to fill the hole are marked as unwritten to prevent
3543 * parallel buffered read to expose the stale data before DIO complete
3544 * the data IO.
3545 *
3546 * As to previously fallocated extents, ext4 get_block will just simply
3547 * mark the buffer mapped but still keep the extents unwritten.
3548 *
3549 * For non AIO case, we will convert those unwritten extents to written
3550 * after return back from blockdev_direct_IO. That way we save us from
3551 * allocating io_end structure and also the overhead of offloading
3552 * the extent convertion to a workqueue.
3553 *
3554 * For async DIO, the conversion needs to be deferred when the
3555 * IO is completed. The ext4 end_io callback function will be
3556 * called to take care of the conversion work. Here for async
3557 * case, we allocate an io_end structure to hook to the iocb.
3558 */
3559 iocb->private = NULL;
3560 if (overwrite)
3561 get_block_func = ext4_dio_get_block_overwrite;
3562 else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
3563 round_down(offset, 1 << inode->i_blkbits) >= inode->i_size) {
3564 get_block_func = ext4_dio_get_block;
3565 dio_flags = DIO_LOCKING | DIO_SKIP_HOLES;
3566 } else if (is_sync_kiocb(iocb)) {
3567 get_block_func = ext4_dio_get_block_unwritten_sync;
3568 dio_flags = DIO_LOCKING;
3569 } else {
3570 get_block_func = ext4_dio_get_block_unwritten_async;
3571 dio_flags = DIO_LOCKING;
3572 }
3573#ifdef CONFIG_EXT4_FS_ENCRYPTION
3574 BUG_ON(ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode));
3575#endif
3576 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
3577 get_block_func, ext4_end_io_dio, NULL,
3578 dio_flags);
3579
3580 if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3581 EXT4_STATE_DIO_UNWRITTEN)) {
3582 int err;
3583 /*
3584 * for non AIO case, since the IO is already
3585 * completed, we could do the conversion right here
3586 */
3587 err = ext4_convert_unwritten_extents(NULL, inode,
3588 offset, ret);
3589 if (err < 0)
3590 ret = err;
3591 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3592 }
3593
3594 inode_dio_end(inode);
3595 /* take i_mutex locking again if we do a ovewrite dio */
3596 if (overwrite)
3597 inode_lock(inode);
3598
3599 if (ret < 0 && final_size > inode->i_size)
3600 ext4_truncate_failed_write(inode);
3601
3602 /* Handle extending of i_size after direct IO write */
3603 if (orphan) {
3604 int err;
3605
3606 /* Credits for sb + inode write */
3607 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3608 if (IS_ERR(handle)) {
3609 /* This is really bad luck. We've written the data
3610 * but cannot extend i_size. Bail out and pretend
3611 * the write failed... */
3612 ret = PTR_ERR(handle);
3613 if (inode->i_nlink)
3614 ext4_orphan_del(NULL, inode);
3615
3616 goto out;
3617 }
3618 if (inode->i_nlink)
3619 ext4_orphan_del(handle, inode);
3620 if (ret > 0) {
3621 loff_t end = offset + ret;
3622 if (end > inode->i_size) {
3623 ei->i_disksize = end;
3624 i_size_write(inode, end);
3625 /*
3626 * We're going to return a positive `ret'
3627 * here due to non-zero-length I/O, so there's
3628 * no way of reporting error returns from
3629 * ext4_mark_inode_dirty() to userspace. So
3630 * ignore it.
3631 */
3632 ext4_mark_inode_dirty(handle, inode);
3633 }
3634 }
3635 err = ext4_journal_stop(handle);
3636 if (ret == 0)
3637 ret = err;
3638 }
3639out:
3640 return ret;
3641}
3642
3643static ssize_t ext4_direct_IO_read(struct kiocb *iocb, struct iov_iter *iter)
3644{
3645 struct address_space *mapping = iocb->ki_filp->f_mapping;
3646 struct inode *inode = mapping->host;
3647 size_t count = iov_iter_count(iter);
3648 ssize_t ret;
3649
3650 /*
3651 * Shared inode_lock is enough for us - it protects against concurrent
3652 * writes & truncates and since we take care of writing back page cache,
3653 * we are protected against page writeback as well.
3654 */
3655 inode_lock_shared(inode);
3656 ret = filemap_write_and_wait_range(mapping, iocb->ki_pos,
3657 iocb->ki_pos + count);
3658 if (ret)
3659 goto out_unlock;
3660 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
3661 iter, ext4_dio_get_block, NULL, NULL, 0);
3662out_unlock:
3663 inode_unlock_shared(inode);
3664 return ret;
3665}
3666
3667static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3668{
3669 struct file *file = iocb->ki_filp;
3670 struct inode *inode = file->f_mapping->host;
3671 size_t count = iov_iter_count(iter);
3672 loff_t offset = iocb->ki_pos;
3673 ssize_t ret;
3674
3675#ifdef CONFIG_EXT4_FS_ENCRYPTION
3676 if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode))
3677 return 0;
3678#endif
3679
3680 /*
3681 * If we are doing data journalling we don't support O_DIRECT
3682 */
3683 if (ext4_should_journal_data(inode))
3684 return 0;
3685
3686 /* Let buffer I/O handle the inline data case. */
3687 if (ext4_has_inline_data(inode))
3688 return 0;
3689
3690 /* DAX uses iomap path now */
3691 if (WARN_ON_ONCE(IS_DAX(inode)))
3692 return 0;
3693
3694 trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3695 if (iov_iter_rw(iter) == READ)
3696 ret = ext4_direct_IO_read(iocb, iter);
3697 else
3698 ret = ext4_direct_IO_write(iocb, iter);
3699 trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3700 return ret;
3701}
3702
3703/*
3704 * Pages can be marked dirty completely asynchronously from ext4's journalling
3705 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3706 * much here because ->set_page_dirty is called under VFS locks. The page is
3707 * not necessarily locked.
3708 *
3709 * We cannot just dirty the page and leave attached buffers clean, because the
3710 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3711 * or jbddirty because all the journalling code will explode.
3712 *
3713 * So what we do is to mark the page "pending dirty" and next time writepage
3714 * is called, propagate that into the buffers appropriately.
3715 */
3716static int ext4_journalled_set_page_dirty(struct page *page)
3717{
3718 SetPageChecked(page);
3719 return __set_page_dirty_nobuffers(page);
3720}
3721
3722static int ext4_set_page_dirty(struct page *page)
3723{
3724 WARN_ON_ONCE(!PageLocked(page) && !PageDirty(page));
3725 WARN_ON_ONCE(!page_has_buffers(page));
3726 return __set_page_dirty_buffers(page);
3727}
3728
3729static const struct address_space_operations ext4_aops = {
3730 .readpage = ext4_readpage,
3731 .readpages = ext4_readpages,
3732 .writepage = ext4_writepage,
3733 .writepages = ext4_writepages,
3734 .write_begin = ext4_write_begin,
3735 .write_end = ext4_write_end,
3736 .set_page_dirty = ext4_set_page_dirty,
3737 .bmap = ext4_bmap,
3738 .invalidatepage = ext4_invalidatepage,
3739 .releasepage = ext4_releasepage,
3740 .direct_IO = ext4_direct_IO,
3741 .migratepage = buffer_migrate_page,
3742 .is_partially_uptodate = block_is_partially_uptodate,
3743 .error_remove_page = generic_error_remove_page,
3744};
3745
3746static const struct address_space_operations ext4_journalled_aops = {
3747 .readpage = ext4_readpage,
3748 .readpages = ext4_readpages,
3749 .writepage = ext4_writepage,
3750 .writepages = ext4_writepages,
3751 .write_begin = ext4_write_begin,
3752 .write_end = ext4_journalled_write_end,
3753 .set_page_dirty = ext4_journalled_set_page_dirty,
3754 .bmap = ext4_bmap,
3755 .invalidatepage = ext4_journalled_invalidatepage,
3756 .releasepage = ext4_releasepage,
3757 .direct_IO = ext4_direct_IO,
3758 .is_partially_uptodate = block_is_partially_uptodate,
3759 .error_remove_page = generic_error_remove_page,
3760};
3761
3762static const struct address_space_operations ext4_da_aops = {
3763 .readpage = ext4_readpage,
3764 .readpages = ext4_readpages,
3765 .writepage = ext4_writepage,
3766 .writepages = ext4_writepages,
3767 .write_begin = ext4_da_write_begin,
3768 .write_end = ext4_da_write_end,
3769 .set_page_dirty = ext4_set_page_dirty,
3770 .bmap = ext4_bmap,
3771 .invalidatepage = ext4_da_invalidatepage,
3772 .releasepage = ext4_releasepage,
3773 .direct_IO = ext4_direct_IO,
3774 .migratepage = buffer_migrate_page,
3775 .is_partially_uptodate = block_is_partially_uptodate,
3776 .error_remove_page = generic_error_remove_page,
3777};
3778
3779void ext4_set_aops(struct inode *inode)
3780{
3781 switch (ext4_inode_journal_mode(inode)) {
3782 case EXT4_INODE_ORDERED_DATA_MODE:
3783 case EXT4_INODE_WRITEBACK_DATA_MODE:
3784 break;
3785 case EXT4_INODE_JOURNAL_DATA_MODE:
3786 inode->i_mapping->a_ops = &ext4_journalled_aops;
3787 return;
3788 default:
3789 BUG();
3790 }
3791 if (test_opt(inode->i_sb, DELALLOC))
3792 inode->i_mapping->a_ops = &ext4_da_aops;
3793 else
3794 inode->i_mapping->a_ops = &ext4_aops;
3795}
3796
3797static int __ext4_block_zero_page_range(handle_t *handle,
3798 struct address_space *mapping, loff_t from, loff_t length)
3799{
3800 ext4_fsblk_t index = from >> PAGE_SHIFT;
3801 unsigned offset = from & (PAGE_SIZE-1);
3802 unsigned blocksize, pos;
3803 ext4_lblk_t iblock;
3804 struct inode *inode = mapping->host;
3805 struct buffer_head *bh;
3806 struct page *page;
3807 int err = 0;
3808
3809 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3810 mapping_gfp_constraint(mapping, ~__GFP_FS));
3811 if (!page)
3812 return -ENOMEM;
3813
3814 blocksize = inode->i_sb->s_blocksize;
3815
3816 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3817
3818 if (!page_has_buffers(page))
3819 create_empty_buffers(page, blocksize, 0);
3820
3821 /* Find the buffer that contains "offset" */
3822 bh = page_buffers(page);
3823 pos = blocksize;
3824 while (offset >= pos) {
3825 bh = bh->b_this_page;
3826 iblock++;
3827 pos += blocksize;
3828 }
3829 if (buffer_freed(bh)) {
3830 BUFFER_TRACE(bh, "freed: skip");
3831 goto unlock;
3832 }
3833 if (!buffer_mapped(bh)) {
3834 BUFFER_TRACE(bh, "unmapped");
3835 ext4_get_block(inode, iblock, bh, 0);
3836 /* unmapped? It's a hole - nothing to do */
3837 if (!buffer_mapped(bh)) {
3838 BUFFER_TRACE(bh, "still unmapped");
3839 goto unlock;
3840 }
3841 }
3842
3843 /* Ok, it's mapped. Make sure it's up-to-date */
3844 if (PageUptodate(page))
3845 set_buffer_uptodate(bh);
3846
3847 if (!buffer_uptodate(bh)) {
3848 err = -EIO;
3849 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
3850 wait_on_buffer(bh);
3851 /* Uhhuh. Read error. Complain and punt. */
3852 if (!buffer_uptodate(bh))
3853 goto unlock;
3854 if (S_ISREG(inode->i_mode) &&
3855 ext4_encrypted_inode(inode)) {
3856 /* We expect the key to be set. */
3857 BUG_ON(!fscrypt_has_encryption_key(inode));
3858 BUG_ON(blocksize != PAGE_SIZE);
3859 WARN_ON_ONCE(fscrypt_decrypt_page(page->mapping->host,
3860 page, PAGE_SIZE, 0, page->index));
3861 }
3862 }
3863 if (ext4_should_journal_data(inode)) {
3864 BUFFER_TRACE(bh, "get write access");
3865 err = ext4_journal_get_write_access(handle, bh);
3866 if (err)
3867 goto unlock;
3868 }
3869 zero_user(page, offset, length);
3870 BUFFER_TRACE(bh, "zeroed end of block");
3871
3872 if (ext4_should_journal_data(inode)) {
3873 err = ext4_handle_dirty_metadata(handle, inode, bh);
3874 } else {
3875 err = 0;
3876 mark_buffer_dirty(bh);
3877 if (ext4_should_order_data(inode))
3878 err = ext4_jbd2_inode_add_write(handle, inode);
3879 }
3880
3881unlock:
3882 unlock_page(page);
3883 put_page(page);
3884 return err;
3885}
3886
3887/*
3888 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3889 * starting from file offset 'from'. The range to be zero'd must
3890 * be contained with in one block. If the specified range exceeds
3891 * the end of the block it will be shortened to end of the block
3892 * that cooresponds to 'from'
3893 */
3894static int ext4_block_zero_page_range(handle_t *handle,
3895 struct address_space *mapping, loff_t from, loff_t length)
3896{
3897 struct inode *inode = mapping->host;
3898 unsigned offset = from & (PAGE_SIZE-1);
3899 unsigned blocksize = inode->i_sb->s_blocksize;
3900 unsigned max = blocksize - (offset & (blocksize - 1));
3901
3902 /*
3903 * correct length if it does not fall between
3904 * 'from' and the end of the block
3905 */
3906 if (length > max || length < 0)
3907 length = max;
3908
3909 if (IS_DAX(inode)) {
3910 return iomap_zero_range(inode, from, length, NULL,
3911 &ext4_iomap_ops);
3912 }
3913 return __ext4_block_zero_page_range(handle, mapping, from, length);
3914}
3915
3916/*
3917 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3918 * up to the end of the block which corresponds to `from'.
3919 * This required during truncate. We need to physically zero the tail end
3920 * of that block so it doesn't yield old data if the file is later grown.
3921 */
3922static int ext4_block_truncate_page(handle_t *handle,
3923 struct address_space *mapping, loff_t from)
3924{
3925 unsigned offset = from & (PAGE_SIZE-1);
3926 unsigned length;
3927 unsigned blocksize;
3928 struct inode *inode = mapping->host;
3929
3930 /* If we are processing an encrypted inode during orphan list handling */
3931 if (ext4_encrypted_inode(inode) && !fscrypt_has_encryption_key(inode))
3932 return 0;
3933
3934 blocksize = inode->i_sb->s_blocksize;
3935 length = blocksize - (offset & (blocksize - 1));
3936
3937 return ext4_block_zero_page_range(handle, mapping, from, length);
3938}
3939
3940int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3941 loff_t lstart, loff_t length)
3942{
3943 struct super_block *sb = inode->i_sb;
3944 struct address_space *mapping = inode->i_mapping;
3945 unsigned partial_start, partial_end;
3946 ext4_fsblk_t start, end;
3947 loff_t byte_end = (lstart + length - 1);
3948 int err = 0;
3949
3950 partial_start = lstart & (sb->s_blocksize - 1);
3951 partial_end = byte_end & (sb->s_blocksize - 1);
3952
3953 start = lstart >> sb->s_blocksize_bits;
3954 end = byte_end >> sb->s_blocksize_bits;
3955
3956 /* Handle partial zero within the single block */
3957 if (start == end &&
3958 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3959 err = ext4_block_zero_page_range(handle, mapping,
3960 lstart, length);
3961 return err;
3962 }
3963 /* Handle partial zero out on the start of the range */
3964 if (partial_start) {
3965 err = ext4_block_zero_page_range(handle, mapping,
3966 lstart, sb->s_blocksize);
3967 if (err)
3968 return err;
3969 }
3970 /* Handle partial zero out on the end of the range */
3971 if (partial_end != sb->s_blocksize - 1)
3972 err = ext4_block_zero_page_range(handle, mapping,
3973 byte_end - partial_end,
3974 partial_end + 1);
3975 return err;
3976}
3977
3978int ext4_can_truncate(struct inode *inode)
3979{
3980 if (S_ISREG(inode->i_mode))
3981 return 1;
3982 if (S_ISDIR(inode->i_mode))
3983 return 1;
3984 if (S_ISLNK(inode->i_mode))
3985 return !ext4_inode_is_fast_symlink(inode);
3986 return 0;
3987}
3988
3989/*
3990 * We have to make sure i_disksize gets properly updated before we truncate
3991 * page cache due to hole punching or zero range. Otherwise i_disksize update
3992 * can get lost as it may have been postponed to submission of writeback but
3993 * that will never happen after we truncate page cache.
3994 */
3995int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
3996 loff_t len)
3997{
3998 handle_t *handle;
3999 loff_t size = i_size_read(inode);
4000
4001 WARN_ON(!inode_is_locked(inode));
4002 if (offset > size || offset + len < size)
4003 return 0;
4004
4005 if (EXT4_I(inode)->i_disksize >= size)
4006 return 0;
4007
4008 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
4009 if (IS_ERR(handle))
4010 return PTR_ERR(handle);
4011 ext4_update_i_disksize(inode, size);
4012 ext4_mark_inode_dirty(handle, inode);
4013 ext4_journal_stop(handle);
4014
4015 return 0;
4016}
4017
4018/*
4019 * ext4_punch_hole: punches a hole in a file by releasing the blocks
4020 * associated with the given offset and length
4021 *
4022 * @inode: File inode
4023 * @offset: The offset where the hole will begin
4024 * @len: The length of the hole
4025 *
4026 * Returns: 0 on success or negative on failure
4027 */
4028
4029int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
4030{
4031 struct super_block *sb = inode->i_sb;
4032 ext4_lblk_t first_block, stop_block;
4033 struct address_space *mapping = inode->i_mapping;
4034 loff_t first_block_offset, last_block_offset;
4035 handle_t *handle;
4036 unsigned int credits;
4037 int ret = 0;
4038
4039 if (!S_ISREG(inode->i_mode))
4040 return -EOPNOTSUPP;
4041
4042 trace_ext4_punch_hole(inode, offset, length, 0);
4043
4044 /*
4045 * Write out all dirty pages to avoid race conditions
4046 * Then release them.
4047 */
4048 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4049 ret = filemap_write_and_wait_range(mapping, offset,
4050 offset + length - 1);
4051 if (ret)
4052 return ret;
4053 }
4054
4055 inode_lock(inode);
4056
4057 /* No need to punch hole beyond i_size */
4058 if (offset >= inode->i_size)
4059 goto out_mutex;
4060
4061 /*
4062 * If the hole extends beyond i_size, set the hole
4063 * to end after the page that contains i_size
4064 */
4065 if (offset + length > inode->i_size) {
4066 length = inode->i_size +
4067 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
4068 offset;
4069 }
4070
4071 if (offset & (sb->s_blocksize - 1) ||
4072 (offset + length) & (sb->s_blocksize - 1)) {
4073 /*
4074 * Attach jinode to inode for jbd2 if we do any zeroing of
4075 * partial block
4076 */
4077 ret = ext4_inode_attach_jinode(inode);
4078 if (ret < 0)
4079 goto out_mutex;
4080
4081 }
4082
4083 /* Wait all existing dio workers, newcomers will block on i_mutex */
4084 ext4_inode_block_unlocked_dio(inode);
4085 inode_dio_wait(inode);
4086
4087 /*
4088 * Prevent page faults from reinstantiating pages we have released from
4089 * page cache.
4090 */
4091 down_write(&EXT4_I(inode)->i_mmap_sem);
4092 first_block_offset = round_up(offset, sb->s_blocksize);
4093 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4094
4095 /* Now release the pages and zero block aligned part of pages*/
4096 if (last_block_offset > first_block_offset) {
4097 ret = ext4_update_disksize_before_punch(inode, offset, length);
4098 if (ret)
4099 goto out_dio;
4100 truncate_pagecache_range(inode, first_block_offset,
4101 last_block_offset);
4102 }
4103
4104 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4105 credits = ext4_writepage_trans_blocks(inode);
4106 else
4107 credits = ext4_blocks_for_truncate(inode);
4108 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4109 if (IS_ERR(handle)) {
4110 ret = PTR_ERR(handle);
4111 ext4_std_error(sb, ret);
4112 goto out_dio;
4113 }
4114
4115 ret = ext4_zero_partial_blocks(handle, inode, offset,
4116 length);
4117 if (ret)
4118 goto out_stop;
4119
4120 first_block = (offset + sb->s_blocksize - 1) >>
4121 EXT4_BLOCK_SIZE_BITS(sb);
4122 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4123
4124 /* If there are no blocks to remove, return now */
4125 if (first_block >= stop_block)
4126 goto out_stop;
4127
4128 down_write(&EXT4_I(inode)->i_data_sem);
4129 ext4_discard_preallocations(inode);
4130
4131 ret = ext4_es_remove_extent(inode, first_block,
4132 stop_block - first_block);
4133 if (ret) {
4134 up_write(&EXT4_I(inode)->i_data_sem);
4135 goto out_stop;
4136 }
4137
4138 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4139 ret = ext4_ext_remove_space(inode, first_block,
4140 stop_block - 1);
4141 else
4142 ret = ext4_ind_remove_space(handle, inode, first_block,
4143 stop_block);
4144
4145 up_write(&EXT4_I(inode)->i_data_sem);
4146 if (IS_SYNC(inode))
4147 ext4_handle_sync(handle);
4148
4149 inode->i_mtime = inode->i_ctime = current_time(inode);
4150 ext4_mark_inode_dirty(handle, inode);
4151out_stop:
4152 ext4_journal_stop(handle);
4153out_dio:
4154 up_write(&EXT4_I(inode)->i_mmap_sem);
4155 ext4_inode_resume_unlocked_dio(inode);
4156out_mutex:
4157 inode_unlock(inode);
4158 return ret;
4159}
4160
4161int ext4_inode_attach_jinode(struct inode *inode)
4162{
4163 struct ext4_inode_info *ei = EXT4_I(inode);
4164 struct jbd2_inode *jinode;
4165
4166 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4167 return 0;
4168
4169 jinode = jbd2_alloc_inode(GFP_KERNEL);
4170 spin_lock(&inode->i_lock);
4171 if (!ei->jinode) {
4172 if (!jinode) {
4173 spin_unlock(&inode->i_lock);
4174 return -ENOMEM;
4175 }
4176 ei->jinode = jinode;
4177 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4178 jinode = NULL;
4179 }
4180 spin_unlock(&inode->i_lock);
4181 if (unlikely(jinode != NULL))
4182 jbd2_free_inode(jinode);
4183 return 0;
4184}
4185
4186/*
4187 * ext4_truncate()
4188 *
4189 * We block out ext4_get_block() block instantiations across the entire
4190 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4191 * simultaneously on behalf of the same inode.
4192 *
4193 * As we work through the truncate and commit bits of it to the journal there
4194 * is one core, guiding principle: the file's tree must always be consistent on
4195 * disk. We must be able to restart the truncate after a crash.
4196 *
4197 * The file's tree may be transiently inconsistent in memory (although it
4198 * probably isn't), but whenever we close off and commit a journal transaction,
4199 * the contents of (the filesystem + the journal) must be consistent and
4200 * restartable. It's pretty simple, really: bottom up, right to left (although
4201 * left-to-right works OK too).
4202 *
4203 * Note that at recovery time, journal replay occurs *before* the restart of
4204 * truncate against the orphan inode list.
4205 *
4206 * The committed inode has the new, desired i_size (which is the same as
4207 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4208 * that this inode's truncate did not complete and it will again call
4209 * ext4_truncate() to have another go. So there will be instantiated blocks
4210 * to the right of the truncation point in a crashed ext4 filesystem. But
4211 * that's fine - as long as they are linked from the inode, the post-crash
4212 * ext4_truncate() run will find them and release them.
4213 */
4214int ext4_truncate(struct inode *inode)
4215{
4216 struct ext4_inode_info *ei = EXT4_I(inode);
4217 unsigned int credits;
4218 int err = 0;
4219 handle_t *handle;
4220 struct address_space *mapping = inode->i_mapping;
4221
4222 /*
4223 * There is a possibility that we're either freeing the inode
4224 * or it's a completely new inode. In those cases we might not
4225 * have i_mutex locked because it's not necessary.
4226 */
4227 if (!(inode->i_state & (I_NEW|I_FREEING)))
4228 WARN_ON(!inode_is_locked(inode));
4229 trace_ext4_truncate_enter(inode);
4230
4231 if (!ext4_can_truncate(inode))
4232 return 0;
4233
4234 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4235
4236 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4237 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4238
4239 if (ext4_has_inline_data(inode)) {
4240 int has_inline = 1;
4241
4242 ext4_inline_data_truncate(inode, &has_inline);
4243 if (has_inline)
4244 return 0;
4245 }
4246
4247 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4248 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4249 if (ext4_inode_attach_jinode(inode) < 0)
4250 return 0;
4251 }
4252
4253 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4254 credits = ext4_writepage_trans_blocks(inode);
4255 else
4256 credits = ext4_blocks_for_truncate(inode);
4257
4258 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4259 if (IS_ERR(handle))
4260 return PTR_ERR(handle);
4261
4262 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4263 ext4_block_truncate_page(handle, mapping, inode->i_size);
4264
4265 /*
4266 * We add the inode to the orphan list, so that if this
4267 * truncate spans multiple transactions, and we crash, we will
4268 * resume the truncate when the filesystem recovers. It also
4269 * marks the inode dirty, to catch the new size.
4270 *
4271 * Implication: the file must always be in a sane, consistent
4272 * truncatable state while each transaction commits.
4273 */
4274 err = ext4_orphan_add(handle, inode);
4275 if (err)
4276 goto out_stop;
4277
4278 down_write(&EXT4_I(inode)->i_data_sem);
4279
4280 ext4_discard_preallocations(inode);
4281
4282 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4283 err = ext4_ext_truncate(handle, inode);
4284 else
4285 ext4_ind_truncate(handle, inode);
4286
4287 up_write(&ei->i_data_sem);
4288 if (err)
4289 goto out_stop;
4290
4291 if (IS_SYNC(inode))
4292 ext4_handle_sync(handle);
4293
4294out_stop:
4295 /*
4296 * If this was a simple ftruncate() and the file will remain alive,
4297 * then we need to clear up the orphan record which we created above.
4298 * However, if this was a real unlink then we were called by
4299 * ext4_evict_inode(), and we allow that function to clean up the
4300 * orphan info for us.
4301 */
4302 if (inode->i_nlink)
4303 ext4_orphan_del(handle, inode);
4304
4305 inode->i_mtime = inode->i_ctime = current_time(inode);
4306 ext4_mark_inode_dirty(handle, inode);
4307 ext4_journal_stop(handle);
4308
4309 trace_ext4_truncate_exit(inode);
4310 return err;
4311}
4312
4313/*
4314 * ext4_get_inode_loc returns with an extra refcount against the inode's
4315 * underlying buffer_head on success. If 'in_mem' is true, we have all
4316 * data in memory that is needed to recreate the on-disk version of this
4317 * inode.
4318 */
4319static int __ext4_get_inode_loc(struct inode *inode,
4320 struct ext4_iloc *iloc, int in_mem)
4321{
4322 struct ext4_group_desc *gdp;
4323 struct buffer_head *bh;
4324 struct super_block *sb = inode->i_sb;
4325 ext4_fsblk_t block;
4326 int inodes_per_block, inode_offset;
4327
4328 iloc->bh = NULL;
4329 if (!ext4_valid_inum(sb, inode->i_ino))
4330 return -EFSCORRUPTED;
4331
4332 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4333 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4334 if (!gdp)
4335 return -EIO;
4336
4337 /*
4338 * Figure out the offset within the block group inode table
4339 */
4340 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4341 inode_offset = ((inode->i_ino - 1) %
4342 EXT4_INODES_PER_GROUP(sb));
4343 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4344 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4345
4346 bh = sb_getblk(sb, block);
4347 if (unlikely(!bh))
4348 return -ENOMEM;
4349 if (!buffer_uptodate(bh)) {
4350 lock_buffer(bh);
4351
4352 /*
4353 * If the buffer has the write error flag, we have failed
4354 * to write out another inode in the same block. In this
4355 * case, we don't have to read the block because we may
4356 * read the old inode data successfully.
4357 */
4358 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4359 set_buffer_uptodate(bh);
4360
4361 if (buffer_uptodate(bh)) {
4362 /* someone brought it uptodate while we waited */
4363 unlock_buffer(bh);
4364 goto has_buffer;
4365 }
4366
4367 /*
4368 * If we have all information of the inode in memory and this
4369 * is the only valid inode in the block, we need not read the
4370 * block.
4371 */
4372 if (in_mem) {
4373 struct buffer_head *bitmap_bh;
4374 int i, start;
4375
4376 start = inode_offset & ~(inodes_per_block - 1);
4377
4378 /* Is the inode bitmap in cache? */
4379 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4380 if (unlikely(!bitmap_bh))
4381 goto make_io;
4382
4383 /*
4384 * If the inode bitmap isn't in cache then the
4385 * optimisation may end up performing two reads instead
4386 * of one, so skip it.
4387 */
4388 if (!buffer_uptodate(bitmap_bh)) {
4389 brelse(bitmap_bh);
4390 goto make_io;
4391 }
4392 for (i = start; i < start + inodes_per_block; i++) {
4393 if (i == inode_offset)
4394 continue;
4395 if (ext4_test_bit(i, bitmap_bh->b_data))
4396 break;
4397 }
4398 brelse(bitmap_bh);
4399 if (i == start + inodes_per_block) {
4400 /* all other inodes are free, so skip I/O */
4401 memset(bh->b_data, 0, bh->b_size);
4402 set_buffer_uptodate(bh);
4403 unlock_buffer(bh);
4404 goto has_buffer;
4405 }
4406 }
4407
4408make_io:
4409 /*
4410 * If we need to do any I/O, try to pre-readahead extra
4411 * blocks from the inode table.
4412 */
4413 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4414 ext4_fsblk_t b, end, table;
4415 unsigned num;
4416 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4417
4418 table = ext4_inode_table(sb, gdp);
4419 /* s_inode_readahead_blks is always a power of 2 */
4420 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4421 if (table > b)
4422 b = table;
4423 end = b + ra_blks;
4424 num = EXT4_INODES_PER_GROUP(sb);
4425 if (ext4_has_group_desc_csum(sb))
4426 num -= ext4_itable_unused_count(sb, gdp);
4427 table += num / inodes_per_block;
4428 if (end > table)
4429 end = table;
4430 while (b <= end)
4431 sb_breadahead(sb, b++);
4432 }
4433
4434 /*
4435 * There are other valid inodes in the buffer, this inode
4436 * has in-inode xattrs, or we don't have this inode in memory.
4437 * Read the block from disk.
4438 */
4439 trace_ext4_load_inode(inode);
4440 get_bh(bh);
4441 bh->b_end_io = end_buffer_read_sync;
4442 submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
4443 wait_on_buffer(bh);
4444 if (!buffer_uptodate(bh)) {
4445 EXT4_ERROR_INODE_BLOCK(inode, block,
4446 "unable to read itable block");
4447 brelse(bh);
4448 return -EIO;
4449 }
4450 }
4451has_buffer:
4452 iloc->bh = bh;
4453 return 0;
4454}
4455
4456int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4457{
4458 /* We have all inode data except xattrs in memory here. */
4459 return __ext4_get_inode_loc(inode, iloc,
4460 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4461}
4462
4463void ext4_set_inode_flags(struct inode *inode)
4464{
4465 unsigned int flags = EXT4_I(inode)->i_flags;
4466 unsigned int new_fl = 0;
4467
4468 if (flags & EXT4_SYNC_FL)
4469 new_fl |= S_SYNC;
4470 if (flags & EXT4_APPEND_FL)
4471 new_fl |= S_APPEND;
4472 if (flags & EXT4_IMMUTABLE_FL)
4473 new_fl |= S_IMMUTABLE;
4474 if (flags & EXT4_NOATIME_FL)
4475 new_fl |= S_NOATIME;
4476 if (flags & EXT4_DIRSYNC_FL)
4477 new_fl |= S_DIRSYNC;
4478 if (test_opt(inode->i_sb, DAX) && S_ISREG(inode->i_mode) &&
4479 !ext4_should_journal_data(inode) && !ext4_has_inline_data(inode) &&
4480 !ext4_encrypted_inode(inode))
4481 new_fl |= S_DAX;
4482 inode_set_flags(inode, new_fl,
4483 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX);
4484}
4485
4486/* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4487void ext4_get_inode_flags(struct ext4_inode_info *ei)
4488{
4489 unsigned int vfs_fl;
4490 unsigned long old_fl, new_fl;
4491
4492 do {
4493 vfs_fl = ei->vfs_inode.i_flags;
4494 old_fl = ei->i_flags;
4495 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4496 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4497 EXT4_DIRSYNC_FL);
4498 if (vfs_fl & S_SYNC)
4499 new_fl |= EXT4_SYNC_FL;
4500 if (vfs_fl & S_APPEND)
4501 new_fl |= EXT4_APPEND_FL;
4502 if (vfs_fl & S_IMMUTABLE)
4503 new_fl |= EXT4_IMMUTABLE_FL;
4504 if (vfs_fl & S_NOATIME)
4505 new_fl |= EXT4_NOATIME_FL;
4506 if (vfs_fl & S_DIRSYNC)
4507 new_fl |= EXT4_DIRSYNC_FL;
4508 } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4509}
4510
4511static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4512 struct ext4_inode_info *ei)
4513{
4514 blkcnt_t i_blocks ;
4515 struct inode *inode = &(ei->vfs_inode);
4516 struct super_block *sb = inode->i_sb;
4517
4518 if (ext4_has_feature_huge_file(sb)) {
4519 /* we are using combined 48 bit field */
4520 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4521 le32_to_cpu(raw_inode->i_blocks_lo);
4522 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4523 /* i_blocks represent file system block size */
4524 return i_blocks << (inode->i_blkbits - 9);
4525 } else {
4526 return i_blocks;
4527 }
4528 } else {
4529 return le32_to_cpu(raw_inode->i_blocks_lo);
4530 }
4531}
4532
4533static inline void ext4_iget_extra_inode(struct inode *inode,
4534 struct ext4_inode *raw_inode,
4535 struct ext4_inode_info *ei)
4536{
4537 __le32 *magic = (void *)raw_inode +
4538 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4539 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize + sizeof(__le32) <=
4540 EXT4_INODE_SIZE(inode->i_sb) &&
4541 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4542 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4543 ext4_find_inline_data_nolock(inode);
4544 } else
4545 EXT4_I(inode)->i_inline_off = 0;
4546}
4547
4548int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4549{
4550 if (!ext4_has_feature_project(inode->i_sb))
4551 return -EOPNOTSUPP;
4552 *projid = EXT4_I(inode)->i_projid;
4553 return 0;
4554}
4555
4556struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4557{
4558 struct ext4_iloc iloc;
4559 struct ext4_inode *raw_inode;
4560 struct ext4_inode_info *ei;
4561 struct inode *inode;
4562 journal_t *journal = EXT4_SB(sb)->s_journal;
4563 long ret;
4564 loff_t size;
4565 int block;
4566 uid_t i_uid;
4567 gid_t i_gid;
4568 projid_t i_projid;
4569
4570 inode = iget_locked(sb, ino);
4571 if (!inode)
4572 return ERR_PTR(-ENOMEM);
4573 if (!(inode->i_state & I_NEW))
4574 return inode;
4575
4576 ei = EXT4_I(inode);
4577 iloc.bh = NULL;
4578
4579 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4580 if (ret < 0)
4581 goto bad_inode;
4582 raw_inode = ext4_raw_inode(&iloc);
4583
4584 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4585 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4586 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4587 EXT4_INODE_SIZE(inode->i_sb) ||
4588 (ei->i_extra_isize & 3)) {
4589 EXT4_ERROR_INODE(inode,
4590 "bad extra_isize %u (inode size %u)",
4591 ei->i_extra_isize,
4592 EXT4_INODE_SIZE(inode->i_sb));
4593 ret = -EFSCORRUPTED;
4594 goto bad_inode;
4595 }
4596 } else
4597 ei->i_extra_isize = 0;
4598
4599 /* Precompute checksum seed for inode metadata */
4600 if (ext4_has_metadata_csum(sb)) {
4601 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4602 __u32 csum;
4603 __le32 inum = cpu_to_le32(inode->i_ino);
4604 __le32 gen = raw_inode->i_generation;
4605 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4606 sizeof(inum));
4607 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4608 sizeof(gen));
4609 }
4610
4611 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4612 EXT4_ERROR_INODE(inode, "checksum invalid");
4613 ret = -EFSBADCRC;
4614 goto bad_inode;
4615 }
4616
4617 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4618 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4619 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4620 if (ext4_has_feature_project(sb) &&
4621 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4622 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4623 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4624 else
4625 i_projid = EXT4_DEF_PROJID;
4626
4627 if (!(test_opt(inode->i_sb, NO_UID32))) {
4628 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4629 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4630 }
4631 i_uid_write(inode, i_uid);
4632 i_gid_write(inode, i_gid);
4633 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4634 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4635
4636 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4637 ei->i_inline_off = 0;
4638 ei->i_dir_start_lookup = 0;
4639 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4640 /* We now have enough fields to check if the inode was active or not.
4641 * This is needed because nfsd might try to access dead inodes
4642 * the test is that same one that e2fsck uses
4643 * NeilBrown 1999oct15
4644 */
4645 if (inode->i_nlink == 0) {
4646 if ((inode->i_mode == 0 ||
4647 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4648 ino != EXT4_BOOT_LOADER_INO) {
4649 /* this inode is deleted */
4650 ret = -ESTALE;
4651 goto bad_inode;
4652 }
4653 /* The only unlinked inodes we let through here have
4654 * valid i_mode and are being read by the orphan
4655 * recovery code: that's fine, we're about to complete
4656 * the process of deleting those.
4657 * OR it is the EXT4_BOOT_LOADER_INO which is
4658 * not initialized on a new filesystem. */
4659 }
4660 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4661 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4662 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4663 if (ext4_has_feature_64bit(sb))
4664 ei->i_file_acl |=
4665 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4666 inode->i_size = ext4_isize(raw_inode);
4667 if ((size = i_size_read(inode)) < 0) {
4668 EXT4_ERROR_INODE(inode, "bad i_size value: %lld", size);
4669 ret = -EFSCORRUPTED;
4670 goto bad_inode;
4671 }
4672 ei->i_disksize = inode->i_size;
4673#ifdef CONFIG_QUOTA
4674 ei->i_reserved_quota = 0;
4675#endif
4676 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4677 ei->i_block_group = iloc.block_group;
4678 ei->i_last_alloc_group = ~0;
4679 /*
4680 * NOTE! The in-memory inode i_data array is in little-endian order
4681 * even on big-endian machines: we do NOT byteswap the block numbers!
4682 */
4683 for (block = 0; block < EXT4_N_BLOCKS; block++)
4684 ei->i_data[block] = raw_inode->i_block[block];
4685 INIT_LIST_HEAD(&ei->i_orphan);
4686
4687 /*
4688 * Set transaction id's of transactions that have to be committed
4689 * to finish f[data]sync. We set them to currently running transaction
4690 * as we cannot be sure that the inode or some of its metadata isn't
4691 * part of the transaction - the inode could have been reclaimed and
4692 * now it is reread from disk.
4693 */
4694 if (journal) {
4695 transaction_t *transaction;
4696 tid_t tid;
4697
4698 read_lock(&journal->j_state_lock);
4699 if (journal->j_running_transaction)
4700 transaction = journal->j_running_transaction;
4701 else
4702 transaction = journal->j_committing_transaction;
4703 if (transaction)
4704 tid = transaction->t_tid;
4705 else
4706 tid = journal->j_commit_sequence;
4707 read_unlock(&journal->j_state_lock);
4708 ei->i_sync_tid = tid;
4709 ei->i_datasync_tid = tid;
4710 }
4711
4712 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4713 if (ei->i_extra_isize == 0) {
4714 /* The extra space is currently unused. Use it. */
4715 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4716 ei->i_extra_isize = sizeof(struct ext4_inode) -
4717 EXT4_GOOD_OLD_INODE_SIZE;
4718 } else {
4719 ext4_iget_extra_inode(inode, raw_inode, ei);
4720 }
4721 }
4722
4723 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4724 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4725 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4726 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4727
4728 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4729 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4730 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4731 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4732 inode->i_version |=
4733 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4734 }
4735 }
4736
4737 ret = 0;
4738 if (ei->i_file_acl &&
4739 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4740 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4741 ei->i_file_acl);
4742 ret = -EFSCORRUPTED;
4743 goto bad_inode;
4744 } else if (!ext4_has_inline_data(inode)) {
4745 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4746 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4747 (S_ISLNK(inode->i_mode) &&
4748 !ext4_inode_is_fast_symlink(inode))))
4749 /* Validate extent which is part of inode */
4750 ret = ext4_ext_check_inode(inode);
4751 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4752 (S_ISLNK(inode->i_mode) &&
4753 !ext4_inode_is_fast_symlink(inode))) {
4754 /* Validate block references which are part of inode */
4755 ret = ext4_ind_check_inode(inode);
4756 }
4757 }
4758 if (ret)
4759 goto bad_inode;
4760
4761 if (S_ISREG(inode->i_mode)) {
4762 inode->i_op = &ext4_file_inode_operations;
4763 inode->i_fop = &ext4_file_operations;
4764 ext4_set_aops(inode);
4765 } else if (S_ISDIR(inode->i_mode)) {
4766 inode->i_op = &ext4_dir_inode_operations;
4767 inode->i_fop = &ext4_dir_operations;
4768 } else if (S_ISLNK(inode->i_mode)) {
4769 if (ext4_encrypted_inode(inode)) {
4770 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4771 ext4_set_aops(inode);
4772 } else if (ext4_inode_is_fast_symlink(inode)) {
4773 inode->i_link = (char *)ei->i_data;
4774 inode->i_op = &ext4_fast_symlink_inode_operations;
4775 nd_terminate_link(ei->i_data, inode->i_size,
4776 sizeof(ei->i_data) - 1);
4777 } else {
4778 inode->i_op = &ext4_symlink_inode_operations;
4779 ext4_set_aops(inode);
4780 }
4781 inode_nohighmem(inode);
4782 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4783 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4784 inode->i_op = &ext4_special_inode_operations;
4785 if (raw_inode->i_block[0])
4786 init_special_inode(inode, inode->i_mode,
4787 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4788 else
4789 init_special_inode(inode, inode->i_mode,
4790 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4791 } else if (ino == EXT4_BOOT_LOADER_INO) {
4792 make_bad_inode(inode);
4793 } else {
4794 ret = -EFSCORRUPTED;
4795 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
4796 goto bad_inode;
4797 }
4798 brelse(iloc.bh);
4799 ext4_set_inode_flags(inode);
4800 unlock_new_inode(inode);
4801 return inode;
4802
4803bad_inode:
4804 brelse(iloc.bh);
4805 iget_failed(inode);
4806 return ERR_PTR(ret);
4807}
4808
4809struct inode *ext4_iget_normal(struct super_block *sb, unsigned long ino)
4810{
4811 if (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)
4812 return ERR_PTR(-EFSCORRUPTED);
4813 return ext4_iget(sb, ino);
4814}
4815
4816static int ext4_inode_blocks_set(handle_t *handle,
4817 struct ext4_inode *raw_inode,
4818 struct ext4_inode_info *ei)
4819{
4820 struct inode *inode = &(ei->vfs_inode);
4821 u64 i_blocks = inode->i_blocks;
4822 struct super_block *sb = inode->i_sb;
4823
4824 if (i_blocks <= ~0U) {
4825 /*
4826 * i_blocks can be represented in a 32 bit variable
4827 * as multiple of 512 bytes
4828 */
4829 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4830 raw_inode->i_blocks_high = 0;
4831 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4832 return 0;
4833 }
4834 if (!ext4_has_feature_huge_file(sb))
4835 return -EFBIG;
4836
4837 if (i_blocks <= 0xffffffffffffULL) {
4838 /*
4839 * i_blocks can be represented in a 48 bit variable
4840 * as multiple of 512 bytes
4841 */
4842 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4843 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4844 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4845 } else {
4846 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4847 /* i_block is stored in file system block size */
4848 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4849 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4850 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4851 }
4852 return 0;
4853}
4854
4855struct other_inode {
4856 unsigned long orig_ino;
4857 struct ext4_inode *raw_inode;
4858};
4859
4860static int other_inode_match(struct inode * inode, unsigned long ino,
4861 void *data)
4862{
4863 struct other_inode *oi = (struct other_inode *) data;
4864
4865 if ((inode->i_ino != ino) ||
4866 (inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4867 I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
4868 ((inode->i_state & I_DIRTY_TIME) == 0))
4869 return 0;
4870 spin_lock(&inode->i_lock);
4871 if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4872 I_DIRTY_SYNC | I_DIRTY_DATASYNC)) == 0) &&
4873 (inode->i_state & I_DIRTY_TIME)) {
4874 struct ext4_inode_info *ei = EXT4_I(inode);
4875
4876 inode->i_state &= ~(I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED);
4877 spin_unlock(&inode->i_lock);
4878
4879 spin_lock(&ei->i_raw_lock);
4880 EXT4_INODE_SET_XTIME(i_ctime, inode, oi->raw_inode);
4881 EXT4_INODE_SET_XTIME(i_mtime, inode, oi->raw_inode);
4882 EXT4_INODE_SET_XTIME(i_atime, inode, oi->raw_inode);
4883 ext4_inode_csum_set(inode, oi->raw_inode, ei);
4884 spin_unlock(&ei->i_raw_lock);
4885 trace_ext4_other_inode_update_time(inode, oi->orig_ino);
4886 return -1;
4887 }
4888 spin_unlock(&inode->i_lock);
4889 return -1;
4890}
4891
4892/*
4893 * Opportunistically update the other time fields for other inodes in
4894 * the same inode table block.
4895 */
4896static void ext4_update_other_inodes_time(struct super_block *sb,
4897 unsigned long orig_ino, char *buf)
4898{
4899 struct other_inode oi;
4900 unsigned long ino;
4901 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4902 int inode_size = EXT4_INODE_SIZE(sb);
4903
4904 oi.orig_ino = orig_ino;
4905 /*
4906 * Calculate the first inode in the inode table block. Inode
4907 * numbers are one-based. That is, the first inode in a block
4908 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
4909 */
4910 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
4911 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
4912 if (ino == orig_ino)
4913 continue;
4914 oi.raw_inode = (struct ext4_inode *) buf;
4915 (void) find_inode_nowait(sb, ino, other_inode_match, &oi);
4916 }
4917}
4918
4919/*
4920 * Post the struct inode info into an on-disk inode location in the
4921 * buffer-cache. This gobbles the caller's reference to the
4922 * buffer_head in the inode location struct.
4923 *
4924 * The caller must have write access to iloc->bh.
4925 */
4926static int ext4_do_update_inode(handle_t *handle,
4927 struct inode *inode,
4928 struct ext4_iloc *iloc)
4929{
4930 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4931 struct ext4_inode_info *ei = EXT4_I(inode);
4932 struct buffer_head *bh = iloc->bh;
4933 struct super_block *sb = inode->i_sb;
4934 int err = 0, rc, block;
4935 int need_datasync = 0, set_large_file = 0;
4936 uid_t i_uid;
4937 gid_t i_gid;
4938 projid_t i_projid;
4939
4940 spin_lock(&ei->i_raw_lock);
4941
4942 /* For fields not tracked in the in-memory inode,
4943 * initialise them to zero for new inodes. */
4944 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
4945 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4946
4947 ext4_get_inode_flags(ei);
4948 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4949 i_uid = i_uid_read(inode);
4950 i_gid = i_gid_read(inode);
4951 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4952 if (!(test_opt(inode->i_sb, NO_UID32))) {
4953 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4954 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4955/*
4956 * Fix up interoperability with old kernels. Otherwise, old inodes get
4957 * re-used with the upper 16 bits of the uid/gid intact
4958 */
4959 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4960 raw_inode->i_uid_high = 0;
4961 raw_inode->i_gid_high = 0;
4962 } else {
4963 raw_inode->i_uid_high =
4964 cpu_to_le16(high_16_bits(i_uid));
4965 raw_inode->i_gid_high =
4966 cpu_to_le16(high_16_bits(i_gid));
4967 }
4968 } else {
4969 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4970 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4971 raw_inode->i_uid_high = 0;
4972 raw_inode->i_gid_high = 0;
4973 }
4974 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4975
4976 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4977 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4978 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4979 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4980
4981 err = ext4_inode_blocks_set(handle, raw_inode, ei);
4982 if (err) {
4983 spin_unlock(&ei->i_raw_lock);
4984 goto out_brelse;
4985 }
4986 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4987 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4988 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4989 raw_inode->i_file_acl_high =
4990 cpu_to_le16(ei->i_file_acl >> 32);
4991 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4992 if (ei->i_disksize != ext4_isize(raw_inode)) {
4993 ext4_isize_set(raw_inode, ei->i_disksize);
4994 need_datasync = 1;
4995 }
4996 if (ei->i_disksize > 0x7fffffffULL) {
4997 if (!ext4_has_feature_large_file(sb) ||
4998 EXT4_SB(sb)->s_es->s_rev_level ==
4999 cpu_to_le32(EXT4_GOOD_OLD_REV))
5000 set_large_file = 1;
5001 }
5002 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5003 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5004 if (old_valid_dev(inode->i_rdev)) {
5005 raw_inode->i_block[0] =
5006 cpu_to_le32(old_encode_dev(inode->i_rdev));
5007 raw_inode->i_block[1] = 0;
5008 } else {
5009 raw_inode->i_block[0] = 0;
5010 raw_inode->i_block[1] =
5011 cpu_to_le32(new_encode_dev(inode->i_rdev));
5012 raw_inode->i_block[2] = 0;
5013 }
5014 } else if (!ext4_has_inline_data(inode)) {
5015 for (block = 0; block < EXT4_N_BLOCKS; block++)
5016 raw_inode->i_block[block] = ei->i_data[block];
5017 }
5018
5019 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5020 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5021 if (ei->i_extra_isize) {
5022 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5023 raw_inode->i_version_hi =
5024 cpu_to_le32(inode->i_version >> 32);
5025 raw_inode->i_extra_isize =
5026 cpu_to_le16(ei->i_extra_isize);
5027 }
5028 }
5029
5030 BUG_ON(!ext4_has_feature_project(inode->i_sb) &&
5031 i_projid != EXT4_DEF_PROJID);
5032
5033 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
5034 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
5035 raw_inode->i_projid = cpu_to_le32(i_projid);
5036
5037 ext4_inode_csum_set(inode, raw_inode, ei);
5038 spin_unlock(&ei->i_raw_lock);
5039 if (inode->i_sb->s_flags & MS_LAZYTIME)
5040 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5041 bh->b_data);
5042
5043 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5044 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5045 if (!err)
5046 err = rc;
5047 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5048 if (set_large_file) {
5049 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5050 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
5051 if (err)
5052 goto out_brelse;
5053 ext4_update_dynamic_rev(sb);
5054 ext4_set_feature_large_file(sb);
5055 ext4_handle_sync(handle);
5056 err = ext4_handle_dirty_super(handle, sb);
5057 }
5058 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5059out_brelse:
5060 brelse(bh);
5061 ext4_std_error(inode->i_sb, err);
5062 return err;
5063}
5064
5065/*
5066 * ext4_write_inode()
5067 *
5068 * We are called from a few places:
5069 *
5070 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5071 * Here, there will be no transaction running. We wait for any running
5072 * transaction to commit.
5073 *
5074 * - Within flush work (sys_sync(), kupdate and such).
5075 * We wait on commit, if told to.
5076 *
5077 * - Within iput_final() -> write_inode_now()
5078 * We wait on commit, if told to.
5079 *
5080 * In all cases it is actually safe for us to return without doing anything,
5081 * because the inode has been copied into a raw inode buffer in
5082 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5083 * writeback.
5084 *
5085 * Note that we are absolutely dependent upon all inode dirtiers doing the
5086 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5087 * which we are interested.
5088 *
5089 * It would be a bug for them to not do this. The code:
5090 *
5091 * mark_inode_dirty(inode)
5092 * stuff();
5093 * inode->i_size = expr;
5094 *
5095 * is in error because write_inode() could occur while `stuff()' is running,
5096 * and the new i_size will be lost. Plus the inode will no longer be on the
5097 * superblock's dirty inode list.
5098 */
5099int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5100{
5101 int err;
5102
5103 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5104 return 0;
5105
5106 if (EXT4_SB(inode->i_sb)->s_journal) {
5107 if (ext4_journal_current_handle()) {
5108 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5109 dump_stack();
5110 return -EIO;
5111 }
5112
5113 /*
5114 * No need to force transaction in WB_SYNC_NONE mode. Also
5115 * ext4_sync_fs() will force the commit after everything is
5116 * written.
5117 */
5118 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5119 return 0;
5120
5121 err = ext4_force_commit(inode->i_sb);
5122 } else {
5123 struct ext4_iloc iloc;
5124
5125 err = __ext4_get_inode_loc(inode, &iloc, 0);
5126 if (err)
5127 return err;
5128 /*
5129 * sync(2) will flush the whole buffer cache. No need to do
5130 * it here separately for each inode.
5131 */
5132 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5133 sync_dirty_buffer(iloc.bh);
5134 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5135 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5136 "IO error syncing inode");
5137 err = -EIO;
5138 }
5139 brelse(iloc.bh);
5140 }
5141 return err;
5142}
5143
5144/*
5145 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
5146 * buffers that are attached to a page stradding i_size and are undergoing
5147 * commit. In that case we have to wait for commit to finish and try again.
5148 */
5149static void ext4_wait_for_tail_page_commit(struct inode *inode)
5150{
5151 struct page *page;
5152 unsigned offset;
5153 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5154 tid_t commit_tid = 0;
5155 int ret;
5156
5157 offset = inode->i_size & (PAGE_SIZE - 1);
5158 /*
5159 * All buffers in the last page remain valid? Then there's nothing to
5160 * do. We do the check mainly to optimize the common PAGE_SIZE ==
5161 * blocksize case
5162 */
5163 if (offset > PAGE_SIZE - (1 << inode->i_blkbits))
5164 return;
5165 while (1) {
5166 page = find_lock_page(inode->i_mapping,
5167 inode->i_size >> PAGE_SHIFT);
5168 if (!page)
5169 return;
5170 ret = __ext4_journalled_invalidatepage(page, offset,
5171 PAGE_SIZE - offset);
5172 unlock_page(page);
5173 put_page(page);
5174 if (ret != -EBUSY)
5175 return;
5176 commit_tid = 0;
5177 read_lock(&journal->j_state_lock);
5178 if (journal->j_committing_transaction)
5179 commit_tid = journal->j_committing_transaction->t_tid;
5180 read_unlock(&journal->j_state_lock);
5181 if (commit_tid)
5182 jbd2_log_wait_commit(journal, commit_tid);
5183 }
5184}
5185
5186/*
5187 * ext4_setattr()
5188 *
5189 * Called from notify_change.
5190 *
5191 * We want to trap VFS attempts to truncate the file as soon as
5192 * possible. In particular, we want to make sure that when the VFS
5193 * shrinks i_size, we put the inode on the orphan list and modify
5194 * i_disksize immediately, so that during the subsequent flushing of
5195 * dirty pages and freeing of disk blocks, we can guarantee that any
5196 * commit will leave the blocks being flushed in an unused state on
5197 * disk. (On recovery, the inode will get truncated and the blocks will
5198 * be freed, so we have a strong guarantee that no future commit will
5199 * leave these blocks visible to the user.)
5200 *
5201 * Another thing we have to assure is that if we are in ordered mode
5202 * and inode is still attached to the committing transaction, we must
5203 * we start writeout of all the dirty pages which are being truncated.
5204 * This way we are sure that all the data written in the previous
5205 * transaction are already on disk (truncate waits for pages under
5206 * writeback).
5207 *
5208 * Called with inode->i_mutex down.
5209 */
5210int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5211{
5212 struct inode *inode = d_inode(dentry);
5213 int error, rc = 0;
5214 int orphan = 0;
5215 const unsigned int ia_valid = attr->ia_valid;
5216
5217 error = setattr_prepare(dentry, attr);
5218 if (error)
5219 return error;
5220
5221 if (is_quota_modification(inode, attr)) {
5222 error = dquot_initialize(inode);
5223 if (error)
5224 return error;
5225 }
5226 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
5227 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
5228 handle_t *handle;
5229
5230 /* (user+group)*(old+new) structure, inode write (sb,
5231 * inode block, ? - but truncate inode update has it) */
5232 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5233 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5234 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5235 if (IS_ERR(handle)) {
5236 error = PTR_ERR(handle);
5237 goto err_out;
5238 }
5239 error = dquot_transfer(inode, attr);
5240 if (error) {
5241 ext4_journal_stop(handle);
5242 return error;
5243 }
5244 /* Update corresponding info in inode so that everything is in
5245 * one transaction */
5246 if (attr->ia_valid & ATTR_UID)
5247 inode->i_uid = attr->ia_uid;
5248 if (attr->ia_valid & ATTR_GID)
5249 inode->i_gid = attr->ia_gid;
5250 error = ext4_mark_inode_dirty(handle, inode);
5251 ext4_journal_stop(handle);
5252 }
5253
5254 if (attr->ia_valid & ATTR_SIZE) {
5255 handle_t *handle;
5256 loff_t oldsize = inode->i_size;
5257 int shrink = (attr->ia_size <= inode->i_size);
5258
5259 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5260 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5261
5262 if (attr->ia_size > sbi->s_bitmap_maxbytes)
5263 return -EFBIG;
5264 }
5265 if (!S_ISREG(inode->i_mode))
5266 return -EINVAL;
5267
5268 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
5269 inode_inc_iversion(inode);
5270
5271 if (ext4_should_order_data(inode) &&
5272 (attr->ia_size < inode->i_size)) {
5273 error = ext4_begin_ordered_truncate(inode,
5274 attr->ia_size);
5275 if (error)
5276 goto err_out;
5277 }
5278 if (attr->ia_size != inode->i_size) {
5279 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5280 if (IS_ERR(handle)) {
5281 error = PTR_ERR(handle);
5282 goto err_out;
5283 }
5284 if (ext4_handle_valid(handle) && shrink) {
5285 error = ext4_orphan_add(handle, inode);
5286 orphan = 1;
5287 }
5288 /*
5289 * Update c/mtime on truncate up, ext4_truncate() will
5290 * update c/mtime in shrink case below
5291 */
5292 if (!shrink) {
5293 inode->i_mtime = current_time(inode);
5294 inode->i_ctime = inode->i_mtime;
5295 }
5296 down_write(&EXT4_I(inode)->i_data_sem);
5297 EXT4_I(inode)->i_disksize = attr->ia_size;
5298 rc = ext4_mark_inode_dirty(handle, inode);
5299 if (!error)
5300 error = rc;
5301 /*
5302 * We have to update i_size under i_data_sem together
5303 * with i_disksize to avoid races with writeback code
5304 * running ext4_wb_update_i_disksize().
5305 */
5306 if (!error)
5307 i_size_write(inode, attr->ia_size);
5308 up_write(&EXT4_I(inode)->i_data_sem);
5309 ext4_journal_stop(handle);
5310 if (error) {
5311 if (orphan)
5312 ext4_orphan_del(NULL, inode);
5313 goto err_out;
5314 }
5315 }
5316 if (!shrink)
5317 pagecache_isize_extended(inode, oldsize, inode->i_size);
5318
5319 /*
5320 * Blocks are going to be removed from the inode. Wait
5321 * for dio in flight. Temporarily disable
5322 * dioread_nolock to prevent livelock.
5323 */
5324 if (orphan) {
5325 if (!ext4_should_journal_data(inode)) {
5326 ext4_inode_block_unlocked_dio(inode);
5327 inode_dio_wait(inode);
5328 ext4_inode_resume_unlocked_dio(inode);
5329 } else
5330 ext4_wait_for_tail_page_commit(inode);
5331 }
5332 down_write(&EXT4_I(inode)->i_mmap_sem);
5333 /*
5334 * Truncate pagecache after we've waited for commit
5335 * in data=journal mode to make pages freeable.
5336 */
5337 truncate_pagecache(inode, inode->i_size);
5338 if (shrink) {
5339 rc = ext4_truncate(inode);
5340 if (rc)
5341 error = rc;
5342 }
5343 up_write(&EXT4_I(inode)->i_mmap_sem);
5344 }
5345
5346 if (!error) {
5347 setattr_copy(inode, attr);
5348 mark_inode_dirty(inode);
5349 }
5350
5351 /*
5352 * If the call to ext4_truncate failed to get a transaction handle at
5353 * all, we need to clean up the in-core orphan list manually.
5354 */
5355 if (orphan && inode->i_nlink)
5356 ext4_orphan_del(NULL, inode);
5357
5358 if (!error && (ia_valid & ATTR_MODE))
5359 rc = posix_acl_chmod(inode, inode->i_mode);
5360
5361err_out:
5362 ext4_std_error(inode->i_sb, error);
5363 if (!error)
5364 error = rc;
5365 return error;
5366}
5367
5368int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5369 struct kstat *stat)
5370{
5371 struct inode *inode;
5372 unsigned long long delalloc_blocks;
5373
5374 inode = d_inode(dentry);
5375 generic_fillattr(inode, stat);
5376
5377 /*
5378 * If there is inline data in the inode, the inode will normally not
5379 * have data blocks allocated (it may have an external xattr block).
5380 * Report at least one sector for such files, so tools like tar, rsync,
5381 * others doen't incorrectly think the file is completely sparse.
5382 */
5383 if (unlikely(ext4_has_inline_data(inode)))
5384 stat->blocks += (stat->size + 511) >> 9;
5385
5386 /*
5387 * We can't update i_blocks if the block allocation is delayed
5388 * otherwise in the case of system crash before the real block
5389 * allocation is done, we will have i_blocks inconsistent with
5390 * on-disk file blocks.
5391 * We always keep i_blocks updated together with real
5392 * allocation. But to not confuse with user, stat
5393 * will return the blocks that include the delayed allocation
5394 * blocks for this file.
5395 */
5396 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5397 EXT4_I(inode)->i_reserved_data_blocks);
5398 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5399 return 0;
5400}
5401
5402static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5403 int pextents)
5404{
5405 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5406 return ext4_ind_trans_blocks(inode, lblocks);
5407 return ext4_ext_index_trans_blocks(inode, pextents);
5408}
5409
5410/*
5411 * Account for index blocks, block groups bitmaps and block group
5412 * descriptor blocks if modify datablocks and index blocks
5413 * worse case, the indexs blocks spread over different block groups
5414 *
5415 * If datablocks are discontiguous, they are possible to spread over
5416 * different block groups too. If they are contiguous, with flexbg,
5417 * they could still across block group boundary.
5418 *
5419 * Also account for superblock, inode, quota and xattr blocks
5420 */
5421static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5422 int pextents)
5423{
5424 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5425 int gdpblocks;
5426 int idxblocks;
5427 int ret = 0;
5428
5429 /*
5430 * How many index blocks need to touch to map @lblocks logical blocks
5431 * to @pextents physical extents?
5432 */
5433 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5434
5435 ret = idxblocks;
5436
5437 /*
5438 * Now let's see how many group bitmaps and group descriptors need
5439 * to account
5440 */
5441 groups = idxblocks + pextents;
5442 gdpblocks = groups;
5443 if (groups > ngroups)
5444 groups = ngroups;
5445 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5446 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5447
5448 /* bitmaps and block group descriptor blocks */
5449 ret += groups + gdpblocks;
5450
5451 /* Blocks for super block, inode, quota and xattr blocks */
5452 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5453
5454 return ret;
5455}
5456
5457/*
5458 * Calculate the total number of credits to reserve to fit
5459 * the modification of a single pages into a single transaction,
5460 * which may include multiple chunks of block allocations.
5461 *
5462 * This could be called via ext4_write_begin()
5463 *
5464 * We need to consider the worse case, when
5465 * one new block per extent.
5466 */
5467int ext4_writepage_trans_blocks(struct inode *inode)
5468{
5469 int bpp = ext4_journal_blocks_per_page(inode);
5470 int ret;
5471
5472 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5473
5474 /* Account for data blocks for journalled mode */
5475 if (ext4_should_journal_data(inode))
5476 ret += bpp;
5477 return ret;
5478}
5479
5480/*
5481 * Calculate the journal credits for a chunk of data modification.
5482 *
5483 * This is called from DIO, fallocate or whoever calling
5484 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5485 *
5486 * journal buffers for data blocks are not included here, as DIO
5487 * and fallocate do no need to journal data buffers.
5488 */
5489int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5490{
5491 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5492}
5493
5494/*
5495 * The caller must have previously called ext4_reserve_inode_write().
5496 * Give this, we know that the caller already has write access to iloc->bh.
5497 */
5498int ext4_mark_iloc_dirty(handle_t *handle,
5499 struct inode *inode, struct ext4_iloc *iloc)
5500{
5501 int err = 0;
5502
5503 if (IS_I_VERSION(inode))
5504 inode_inc_iversion(inode);
5505
5506 /* the do_update_inode consumes one bh->b_count */
5507 get_bh(iloc->bh);
5508
5509 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5510 err = ext4_do_update_inode(handle, inode, iloc);
5511 put_bh(iloc->bh);
5512 return err;
5513}
5514
5515/*
5516 * On success, We end up with an outstanding reference count against
5517 * iloc->bh. This _must_ be cleaned up later.
5518 */
5519
5520int
5521ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5522 struct ext4_iloc *iloc)
5523{
5524 int err;
5525
5526 err = ext4_get_inode_loc(inode, iloc);
5527 if (!err) {
5528 BUFFER_TRACE(iloc->bh, "get_write_access");
5529 err = ext4_journal_get_write_access(handle, iloc->bh);
5530 if (err) {
5531 brelse(iloc->bh);
5532 iloc->bh = NULL;
5533 }
5534 }
5535 ext4_std_error(inode->i_sb, err);
5536 return err;
5537}
5538
5539/*
5540 * Expand an inode by new_extra_isize bytes.
5541 * Returns 0 on success or negative error number on failure.
5542 */
5543static int ext4_expand_extra_isize(struct inode *inode,
5544 unsigned int new_extra_isize,
5545 struct ext4_iloc iloc,
5546 handle_t *handle)
5547{
5548 struct ext4_inode *raw_inode;
5549 struct ext4_xattr_ibody_header *header;
5550
5551 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5552 return 0;
5553
5554 raw_inode = ext4_raw_inode(&iloc);
5555
5556 header = IHDR(inode, raw_inode);
5557
5558 /* No extended attributes present */
5559 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5560 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5561 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5562 new_extra_isize);
5563 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5564 return 0;
5565 }
5566
5567 /* try to expand with EAs present */
5568 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5569 raw_inode, handle);
5570}
5571
5572/*
5573 * What we do here is to mark the in-core inode as clean with respect to inode
5574 * dirtiness (it may still be data-dirty).
5575 * This means that the in-core inode may be reaped by prune_icache
5576 * without having to perform any I/O. This is a very good thing,
5577 * because *any* task may call prune_icache - even ones which
5578 * have a transaction open against a different journal.
5579 *
5580 * Is this cheating? Not really. Sure, we haven't written the
5581 * inode out, but prune_icache isn't a user-visible syncing function.
5582 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5583 * we start and wait on commits.
5584 */
5585int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5586{
5587 struct ext4_iloc iloc;
5588 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5589 static unsigned int mnt_count;
5590 int err, ret;
5591
5592 might_sleep();
5593 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5594 err = ext4_reserve_inode_write(handle, inode, &iloc);
5595 if (err)
5596 return err;
5597 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5598 !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5599 /*
5600 * In nojournal mode, we can immediately attempt to expand
5601 * the inode. When journaled, we first need to obtain extra
5602 * buffer credits since we may write into the EA block
5603 * with this same handle. If journal_extend fails, then it will
5604 * only result in a minor loss of functionality for that inode.
5605 * If this is felt to be critical, then e2fsck should be run to
5606 * force a large enough s_min_extra_isize.
5607 */
5608 if (!ext4_handle_valid(handle) ||
5609 jbd2_journal_extend(handle,
5610 EXT4_DATA_TRANS_BLOCKS(inode->i_sb)) == 0) {
5611 ret = ext4_expand_extra_isize(inode,
5612 sbi->s_want_extra_isize,
5613 iloc, handle);
5614 if (ret) {
5615 if (mnt_count !=
5616 le16_to_cpu(sbi->s_es->s_mnt_count)) {
5617 ext4_warning(inode->i_sb,
5618 "Unable to expand inode %lu. Delete"
5619 " some EAs or run e2fsck.",
5620 inode->i_ino);
5621 mnt_count =
5622 le16_to_cpu(sbi->s_es->s_mnt_count);
5623 }
5624 }
5625 }
5626 }
5627 return ext4_mark_iloc_dirty(handle, inode, &iloc);
5628}
5629
5630/*
5631 * ext4_dirty_inode() is called from __mark_inode_dirty()
5632 *
5633 * We're really interested in the case where a file is being extended.
5634 * i_size has been changed by generic_commit_write() and we thus need
5635 * to include the updated inode in the current transaction.
5636 *
5637 * Also, dquot_alloc_block() will always dirty the inode when blocks
5638 * are allocated to the file.
5639 *
5640 * If the inode is marked synchronous, we don't honour that here - doing
5641 * so would cause a commit on atime updates, which we don't bother doing.
5642 * We handle synchronous inodes at the highest possible level.
5643 *
5644 * If only the I_DIRTY_TIME flag is set, we can skip everything. If
5645 * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
5646 * to copy into the on-disk inode structure are the timestamp files.
5647 */
5648void ext4_dirty_inode(struct inode *inode, int flags)
5649{
5650 handle_t *handle;
5651
5652 if (flags == I_DIRTY_TIME)
5653 return;
5654 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5655 if (IS_ERR(handle))
5656 goto out;
5657
5658 ext4_mark_inode_dirty(handle, inode);
5659
5660 ext4_journal_stop(handle);
5661out:
5662 return;
5663}
5664
5665#if 0
5666/*
5667 * Bind an inode's backing buffer_head into this transaction, to prevent
5668 * it from being flushed to disk early. Unlike
5669 * ext4_reserve_inode_write, this leaves behind no bh reference and
5670 * returns no iloc structure, so the caller needs to repeat the iloc
5671 * lookup to mark the inode dirty later.
5672 */
5673static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5674{
5675 struct ext4_iloc iloc;
5676
5677 int err = 0;
5678 if (handle) {
5679 err = ext4_get_inode_loc(inode, &iloc);
5680 if (!err) {
5681 BUFFER_TRACE(iloc.bh, "get_write_access");
5682 err = jbd2_journal_get_write_access(handle, iloc.bh);
5683 if (!err)
5684 err = ext4_handle_dirty_metadata(handle,
5685 NULL,
5686 iloc.bh);
5687 brelse(iloc.bh);
5688 }
5689 }
5690 ext4_std_error(inode->i_sb, err);
5691 return err;
5692}
5693#endif
5694
5695int ext4_change_inode_journal_flag(struct inode *inode, int val)
5696{
5697 journal_t *journal;
5698 handle_t *handle;
5699 int err;
5700 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5701
5702 /*
5703 * We have to be very careful here: changing a data block's
5704 * journaling status dynamically is dangerous. If we write a
5705 * data block to the journal, change the status and then delete
5706 * that block, we risk forgetting to revoke the old log record
5707 * from the journal and so a subsequent replay can corrupt data.
5708 * So, first we make sure that the journal is empty and that
5709 * nobody is changing anything.
5710 */
5711
5712 journal = EXT4_JOURNAL(inode);
5713 if (!journal)
5714 return 0;
5715 if (is_journal_aborted(journal))
5716 return -EROFS;
5717
5718 /* Wait for all existing dio workers */
5719 ext4_inode_block_unlocked_dio(inode);
5720 inode_dio_wait(inode);
5721
5722 /*
5723 * Before flushing the journal and switching inode's aops, we have
5724 * to flush all dirty data the inode has. There can be outstanding
5725 * delayed allocations, there can be unwritten extents created by
5726 * fallocate or buffered writes in dioread_nolock mode covered by
5727 * dirty data which can be converted only after flushing the dirty
5728 * data (and journalled aops don't know how to handle these cases).
5729 */
5730 if (val) {
5731 down_write(&EXT4_I(inode)->i_mmap_sem);
5732 err = filemap_write_and_wait(inode->i_mapping);
5733 if (err < 0) {
5734 up_write(&EXT4_I(inode)->i_mmap_sem);
5735 ext4_inode_resume_unlocked_dio(inode);
5736 return err;
5737 }
5738 }
5739
5740 percpu_down_write(&sbi->s_journal_flag_rwsem);
5741 jbd2_journal_lock_updates(journal);
5742
5743 /*
5744 * OK, there are no updates running now, and all cached data is
5745 * synced to disk. We are now in a completely consistent state
5746 * which doesn't have anything in the journal, and we know that
5747 * no filesystem updates are running, so it is safe to modify
5748 * the inode's in-core data-journaling state flag now.
5749 */
5750
5751 if (val)
5752 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5753 else {
5754 err = jbd2_journal_flush(journal);
5755 if (err < 0) {
5756 jbd2_journal_unlock_updates(journal);
5757 percpu_up_write(&sbi->s_journal_flag_rwsem);
5758 ext4_inode_resume_unlocked_dio(inode);
5759 return err;
5760 }
5761 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5762 }
5763 ext4_set_aops(inode);
5764 /*
5765 * Update inode->i_flags after EXT4_INODE_JOURNAL_DATA was updated.
5766 * E.g. S_DAX may get cleared / set.
5767 */
5768 ext4_set_inode_flags(inode);
5769
5770 jbd2_journal_unlock_updates(journal);
5771 percpu_up_write(&sbi->s_journal_flag_rwsem);
5772
5773 if (val)
5774 up_write(&EXT4_I(inode)->i_mmap_sem);
5775 ext4_inode_resume_unlocked_dio(inode);
5776
5777 /* Finally we can mark the inode as dirty. */
5778
5779 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
5780 if (IS_ERR(handle))
5781 return PTR_ERR(handle);
5782
5783 err = ext4_mark_inode_dirty(handle, inode);
5784 ext4_handle_sync(handle);
5785 ext4_journal_stop(handle);
5786 ext4_std_error(inode->i_sb, err);
5787
5788 return err;
5789}
5790
5791static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5792{
5793 return !buffer_mapped(bh);
5794}
5795
5796int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5797{
5798 struct page *page = vmf->page;
5799 loff_t size;
5800 unsigned long len;
5801 int ret;
5802 struct file *file = vma->vm_file;
5803 struct inode *inode = file_inode(file);
5804 struct address_space *mapping = inode->i_mapping;
5805 handle_t *handle;
5806 get_block_t *get_block;
5807 int retries = 0;
5808
5809 sb_start_pagefault(inode->i_sb);
5810 file_update_time(vma->vm_file);
5811
5812 down_read(&EXT4_I(inode)->i_mmap_sem);
5813 /* Delalloc case is easy... */
5814 if (test_opt(inode->i_sb, DELALLOC) &&
5815 !ext4_should_journal_data(inode) &&
5816 !ext4_nonda_switch(inode->i_sb)) {
5817 do {
5818 ret = block_page_mkwrite(vma, vmf,
5819 ext4_da_get_block_prep);
5820 } while (ret == -ENOSPC &&
5821 ext4_should_retry_alloc(inode->i_sb, &retries));
5822 goto out_ret;
5823 }
5824
5825 lock_page(page);
5826 size = i_size_read(inode);
5827 /* Page got truncated from under us? */
5828 if (page->mapping != mapping || page_offset(page) > size) {
5829 unlock_page(page);
5830 ret = VM_FAULT_NOPAGE;
5831 goto out;
5832 }
5833
5834 if (page->index == size >> PAGE_SHIFT)
5835 len = size & ~PAGE_MASK;
5836 else
5837 len = PAGE_SIZE;
5838 /*
5839 * Return if we have all the buffers mapped. This avoids the need to do
5840 * journal_start/journal_stop which can block and take a long time
5841 */
5842 if (page_has_buffers(page)) {
5843 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
5844 0, len, NULL,
5845 ext4_bh_unmapped)) {
5846 /* Wait so that we don't change page under IO */
5847 wait_for_stable_page(page);
5848 ret = VM_FAULT_LOCKED;
5849 goto out;
5850 }
5851 }
5852 unlock_page(page);
5853 /* OK, we need to fill the hole... */
5854 if (ext4_should_dioread_nolock(inode))
5855 get_block = ext4_get_block_unwritten;
5856 else
5857 get_block = ext4_get_block;
5858retry_alloc:
5859 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
5860 ext4_writepage_trans_blocks(inode));
5861 if (IS_ERR(handle)) {
5862 ret = VM_FAULT_SIGBUS;
5863 goto out;
5864 }
5865 ret = block_page_mkwrite(vma, vmf, get_block);
5866 if (!ret && ext4_should_journal_data(inode)) {
5867 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
5868 PAGE_SIZE, NULL, do_journal_get_write_access)) {
5869 unlock_page(page);
5870 ret = VM_FAULT_SIGBUS;
5871 ext4_journal_stop(handle);
5872 goto out;
5873 }
5874 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
5875 }
5876 ext4_journal_stop(handle);
5877 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
5878 goto retry_alloc;
5879out_ret:
5880 ret = block_page_mkwrite_return(ret);
5881out:
5882 up_read(&EXT4_I(inode)->i_mmap_sem);
5883 sb_end_pagefault(inode->i_sb);
5884 return ret;
5885}
5886
5887int ext4_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
5888{
5889 struct inode *inode = file_inode(vma->vm_file);
5890 int err;
5891
5892 down_read(&EXT4_I(inode)->i_mmap_sem);
5893 err = filemap_fault(vma, vmf);
5894 up_read(&EXT4_I(inode)->i_mmap_sem);
5895
5896 return err;
5897}
5898
5899/*
5900 * Find the first extent at or after @lblk in an inode that is not a hole.
5901 * Search for @map_len blocks at most. The extent is returned in @result.
5902 *
5903 * The function returns 1 if we found an extent. The function returns 0 in
5904 * case there is no extent at or after @lblk and in that case also sets
5905 * @result->es_len to 0. In case of error, the error code is returned.
5906 */
5907int ext4_get_next_extent(struct inode *inode, ext4_lblk_t lblk,
5908 unsigned int map_len, struct extent_status *result)
5909{
5910 struct ext4_map_blocks map;
5911 struct extent_status es = {};
5912 int ret;
5913
5914 map.m_lblk = lblk;
5915 map.m_len = map_len;
5916
5917 /*
5918 * For non-extent based files this loop may iterate several times since
5919 * we do not determine full hole size.
5920 */
5921 while (map.m_len > 0) {
5922 ret = ext4_map_blocks(NULL, inode, &map, 0);
5923 if (ret < 0)
5924 return ret;
5925 /* There's extent covering m_lblk? Just return it. */
5926 if (ret > 0) {
5927 int status;
5928
5929 ext4_es_store_pblock(result, map.m_pblk);
5930 result->es_lblk = map.m_lblk;
5931 result->es_len = map.m_len;
5932 if (map.m_flags & EXT4_MAP_UNWRITTEN)
5933 status = EXTENT_STATUS_UNWRITTEN;
5934 else
5935 status = EXTENT_STATUS_WRITTEN;
5936 ext4_es_store_status(result, status);
5937 return 1;
5938 }
5939 ext4_es_find_delayed_extent_range(inode, map.m_lblk,
5940 map.m_lblk + map.m_len - 1,
5941 &es);
5942 /* Is delalloc data before next block in extent tree? */
5943 if (es.es_len && es.es_lblk < map.m_lblk + map.m_len) {
5944 ext4_lblk_t offset = 0;
5945
5946 if (es.es_lblk < lblk)
5947 offset = lblk - es.es_lblk;
5948 result->es_lblk = es.es_lblk + offset;
5949 ext4_es_store_pblock(result,
5950 ext4_es_pblock(&es) + offset);
5951 result->es_len = es.es_len - offset;
5952 ext4_es_store_status(result, ext4_es_status(&es));
5953
5954 return 1;
5955 }
5956 /* There's a hole at m_lblk, advance us after it */
5957 map.m_lblk += map.m_len;
5958 map_len -= map.m_len;
5959 map.m_len = map_len;
5960 cond_resched();
5961 }
5962 result->es_len = 0;
5963 return 0;
5964}
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_journalled_writepage(struct page *page, unsigned int len);
140static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
141 int pextents);
142
143/*
144 * Test whether an inode is a fast symlink.
145 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
146 */
147int ext4_inode_is_fast_symlink(struct inode *inode)
148{
149 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
150 int ea_blocks = EXT4_I(inode)->i_file_acl ?
151 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
152
153 if (ext4_has_inline_data(inode))
154 return 0;
155
156 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
157 }
158 return S_ISLNK(inode->i_mode) && inode->i_size &&
159 (inode->i_size < EXT4_N_BLOCKS * 4);
160}
161
162/*
163 * Called at the last iput() if i_nlink is zero.
164 */
165void ext4_evict_inode(struct inode *inode)
166{
167 handle_t *handle;
168 int err;
169 /*
170 * Credits for final inode cleanup and freeing:
171 * sb + inode (ext4_orphan_del()), block bitmap, group descriptor
172 * (xattr block freeing), bitmap, group descriptor (inode freeing)
173 */
174 int extra_credits = 6;
175 struct ext4_xattr_inode_array *ea_inode_array = NULL;
176 bool freeze_protected = false;
177
178 trace_ext4_evict_inode(inode);
179
180 if (EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)
181 ext4_evict_ea_inode(inode);
182 if (inode->i_nlink) {
183 /*
184 * When journalling data dirty buffers are tracked only in the
185 * journal. So although mm thinks everything is clean and
186 * ready for reaping the inode might still have some pages to
187 * write in the running transaction or waiting to be
188 * checkpointed. Thus calling jbd2_journal_invalidate_folio()
189 * (via truncate_inode_pages()) to discard these buffers can
190 * cause data loss. Also even if we did not discard these
191 * buffers, we would have no way to find them after the inode
192 * is reaped and thus user could see stale data if he tries to
193 * read them before the transaction is checkpointed. So be
194 * careful and force everything to disk here... We use
195 * ei->i_datasync_tid to store the newest transaction
196 * containing inode's data.
197 *
198 * Note that directories do not have this problem because they
199 * don't use page cache.
200 */
201 if (inode->i_ino != EXT4_JOURNAL_INO &&
202 ext4_should_journal_data(inode) &&
203 S_ISREG(inode->i_mode) && inode->i_data.nrpages) {
204 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
205 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
206
207 jbd2_complete_transaction(journal, commit_tid);
208 filemap_write_and_wait(&inode->i_data);
209 }
210 truncate_inode_pages_final(&inode->i_data);
211
212 goto no_delete;
213 }
214
215 if (is_bad_inode(inode))
216 goto no_delete;
217 dquot_initialize(inode);
218
219 if (ext4_should_order_data(inode))
220 ext4_begin_ordered_truncate(inode, 0);
221 truncate_inode_pages_final(&inode->i_data);
222
223 /*
224 * For inodes with journalled data, transaction commit could have
225 * dirtied the inode. And for inodes with dioread_nolock, unwritten
226 * extents converting worker could merge extents and also have dirtied
227 * the inode. Flush worker is ignoring it because of I_FREEING flag but
228 * we still need to remove the inode from the writeback lists.
229 */
230 if (!list_empty_careful(&inode->i_io_list))
231 inode_io_list_del(inode);
232
233 /*
234 * Protect us against freezing - iput() caller didn't have to have any
235 * protection against it. When we are in a running transaction though,
236 * we are already protected against freezing and we cannot grab further
237 * protection due to lock ordering constraints.
238 */
239 if (!ext4_journal_current_handle()) {
240 sb_start_intwrite(inode->i_sb);
241 freeze_protected = true;
242 }
243
244 if (!IS_NOQUOTA(inode))
245 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
246
247 /*
248 * Block bitmap, group descriptor, and inode are accounted in both
249 * ext4_blocks_for_truncate() and extra_credits. So subtract 3.
250 */
251 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
252 ext4_blocks_for_truncate(inode) + extra_credits - 3);
253 if (IS_ERR(handle)) {
254 ext4_std_error(inode->i_sb, PTR_ERR(handle));
255 /*
256 * If we're going to skip the normal cleanup, we still need to
257 * make sure that the in-core orphan linked list is properly
258 * cleaned up.
259 */
260 ext4_orphan_del(NULL, inode);
261 if (freeze_protected)
262 sb_end_intwrite(inode->i_sb);
263 goto no_delete;
264 }
265
266 if (IS_SYNC(inode))
267 ext4_handle_sync(handle);
268
269 /*
270 * Set inode->i_size to 0 before calling ext4_truncate(). We need
271 * special handling of symlinks here because i_size is used to
272 * determine whether ext4_inode_info->i_data contains symlink data or
273 * block mappings. Setting i_size to 0 will remove its fast symlink
274 * status. Erase i_data so that it becomes a valid empty block map.
275 */
276 if (ext4_inode_is_fast_symlink(inode))
277 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
278 inode->i_size = 0;
279 err = ext4_mark_inode_dirty(handle, inode);
280 if (err) {
281 ext4_warning(inode->i_sb,
282 "couldn't mark inode dirty (err %d)", err);
283 goto stop_handle;
284 }
285 if (inode->i_blocks) {
286 err = ext4_truncate(inode);
287 if (err) {
288 ext4_error_err(inode->i_sb, -err,
289 "couldn't truncate inode %lu (err %d)",
290 inode->i_ino, err);
291 goto stop_handle;
292 }
293 }
294
295 /* Remove xattr references. */
296 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
297 extra_credits);
298 if (err) {
299 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
300stop_handle:
301 ext4_journal_stop(handle);
302 ext4_orphan_del(NULL, inode);
303 if (freeze_protected)
304 sb_end_intwrite(inode->i_sb);
305 ext4_xattr_inode_array_free(ea_inode_array);
306 goto no_delete;
307 }
308
309 /*
310 * Kill off the orphan record which ext4_truncate created.
311 * AKPM: I think this can be inside the above `if'.
312 * Note that ext4_orphan_del() has to be able to cope with the
313 * deletion of a non-existent orphan - this is because we don't
314 * know if ext4_truncate() actually created an orphan record.
315 * (Well, we could do this if we need to, but heck - it works)
316 */
317 ext4_orphan_del(handle, inode);
318 EXT4_I(inode)->i_dtime = (__u32)ktime_get_real_seconds();
319
320 /*
321 * One subtle ordering requirement: if anything has gone wrong
322 * (transaction abort, IO errors, whatever), then we can still
323 * do these next steps (the fs will already have been marked as
324 * having errors), but we can't free the inode if the mark_dirty
325 * fails.
326 */
327 if (ext4_mark_inode_dirty(handle, inode))
328 /* If that failed, just do the required in-core inode clear. */
329 ext4_clear_inode(inode);
330 else
331 ext4_free_inode(handle, inode);
332 ext4_journal_stop(handle);
333 if (freeze_protected)
334 sb_end_intwrite(inode->i_sb);
335 ext4_xattr_inode_array_free(ea_inode_array);
336 return;
337no_delete:
338 /*
339 * Check out some where else accidentally dirty the evicting inode,
340 * which may probably cause inode use-after-free issues later.
341 */
342 WARN_ON_ONCE(!list_empty_careful(&inode->i_io_list));
343
344 if (!list_empty(&EXT4_I(inode)->i_fc_list))
345 ext4_fc_mark_ineligible(inode->i_sb, EXT4_FC_REASON_NOMEM, NULL);
346 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
347}
348
349#ifdef CONFIG_QUOTA
350qsize_t *ext4_get_reserved_space(struct inode *inode)
351{
352 return &EXT4_I(inode)->i_reserved_quota;
353}
354#endif
355
356/*
357 * Called with i_data_sem down, which is important since we can call
358 * ext4_discard_preallocations() from here.
359 */
360void ext4_da_update_reserve_space(struct inode *inode,
361 int used, int quota_claim)
362{
363 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
364 struct ext4_inode_info *ei = EXT4_I(inode);
365
366 spin_lock(&ei->i_block_reservation_lock);
367 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
368 if (unlikely(used > ei->i_reserved_data_blocks)) {
369 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
370 "with only %d reserved data blocks",
371 __func__, inode->i_ino, used,
372 ei->i_reserved_data_blocks);
373 WARN_ON(1);
374 used = ei->i_reserved_data_blocks;
375 }
376
377 /* Update per-inode reservations */
378 ei->i_reserved_data_blocks -= used;
379 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
380
381 spin_unlock(&ei->i_block_reservation_lock);
382
383 /* Update quota subsystem for data blocks */
384 if (quota_claim)
385 dquot_claim_block(inode, EXT4_C2B(sbi, used));
386 else {
387 /*
388 * We did fallocate with an offset that is already delayed
389 * allocated. So on delayed allocated writeback we should
390 * not re-claim the quota for fallocated blocks.
391 */
392 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
393 }
394
395 /*
396 * If we have done all the pending block allocations and if
397 * there aren't any writers on the inode, we can discard the
398 * inode's preallocations.
399 */
400 if ((ei->i_reserved_data_blocks == 0) &&
401 !inode_is_open_for_write(inode))
402 ext4_discard_preallocations(inode, 0);
403}
404
405static int __check_block_validity(struct inode *inode, const char *func,
406 unsigned int line,
407 struct ext4_map_blocks *map)
408{
409 if (ext4_has_feature_journal(inode->i_sb) &&
410 (inode->i_ino ==
411 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
412 return 0;
413 if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
414 ext4_error_inode(inode, func, line, map->m_pblk,
415 "lblock %lu mapped to illegal pblock %llu "
416 "(length %d)", (unsigned long) map->m_lblk,
417 map->m_pblk, map->m_len);
418 return -EFSCORRUPTED;
419 }
420 return 0;
421}
422
423int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
424 ext4_lblk_t len)
425{
426 int ret;
427
428 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
429 return fscrypt_zeroout_range(inode, lblk, pblk, len);
430
431 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
432 if (ret > 0)
433 ret = 0;
434
435 return ret;
436}
437
438#define check_block_validity(inode, map) \
439 __check_block_validity((inode), __func__, __LINE__, (map))
440
441#ifdef ES_AGGRESSIVE_TEST
442static void ext4_map_blocks_es_recheck(handle_t *handle,
443 struct inode *inode,
444 struct ext4_map_blocks *es_map,
445 struct ext4_map_blocks *map,
446 int flags)
447{
448 int retval;
449
450 map->m_flags = 0;
451 /*
452 * There is a race window that the result is not the same.
453 * e.g. xfstests #223 when dioread_nolock enables. The reason
454 * is that we lookup a block mapping in extent status tree with
455 * out taking i_data_sem. So at the time the unwritten extent
456 * could be converted.
457 */
458 down_read(&EXT4_I(inode)->i_data_sem);
459 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
460 retval = ext4_ext_map_blocks(handle, inode, map, 0);
461 } else {
462 retval = ext4_ind_map_blocks(handle, inode, map, 0);
463 }
464 up_read((&EXT4_I(inode)->i_data_sem));
465
466 /*
467 * We don't check m_len because extent will be collpased in status
468 * tree. So the m_len might not equal.
469 */
470 if (es_map->m_lblk != map->m_lblk ||
471 es_map->m_flags != map->m_flags ||
472 es_map->m_pblk != map->m_pblk) {
473 printk("ES cache assertion failed for inode: %lu "
474 "es_cached ex [%d/%d/%llu/%x] != "
475 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
476 inode->i_ino, es_map->m_lblk, es_map->m_len,
477 es_map->m_pblk, es_map->m_flags, map->m_lblk,
478 map->m_len, map->m_pblk, map->m_flags,
479 retval, flags);
480 }
481}
482#endif /* ES_AGGRESSIVE_TEST */
483
484/*
485 * The ext4_map_blocks() function tries to look up the requested blocks,
486 * and returns if the blocks are already mapped.
487 *
488 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
489 * and store the allocated blocks in the result buffer head and mark it
490 * mapped.
491 *
492 * If file type is extents based, it will call ext4_ext_map_blocks(),
493 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
494 * based files
495 *
496 * On success, it returns the number of blocks being mapped or allocated. if
497 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
498 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
499 *
500 * It returns 0 if plain look up failed (blocks have not been allocated), in
501 * that case, @map is returned as unmapped but we still do fill map->m_len to
502 * indicate the length of a hole starting at map->m_lblk.
503 *
504 * It returns the error in case of allocation failure.
505 */
506int ext4_map_blocks(handle_t *handle, struct inode *inode,
507 struct ext4_map_blocks *map, int flags)
508{
509 struct extent_status es;
510 int retval;
511 int ret = 0;
512#ifdef ES_AGGRESSIVE_TEST
513 struct ext4_map_blocks orig_map;
514
515 memcpy(&orig_map, map, sizeof(*map));
516#endif
517
518 map->m_flags = 0;
519 ext_debug(inode, "flag 0x%x, max_blocks %u, logical block %lu\n",
520 flags, map->m_len, (unsigned long) map->m_lblk);
521
522 /*
523 * ext4_map_blocks returns an int, and m_len is an unsigned int
524 */
525 if (unlikely(map->m_len > INT_MAX))
526 map->m_len = INT_MAX;
527
528 /* We can handle the block number less than EXT_MAX_BLOCKS */
529 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
530 return -EFSCORRUPTED;
531
532 /* Lookup extent status tree firstly */
533 if (!(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY) &&
534 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
535 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
536 map->m_pblk = ext4_es_pblock(&es) +
537 map->m_lblk - es.es_lblk;
538 map->m_flags |= ext4_es_is_written(&es) ?
539 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
540 retval = es.es_len - (map->m_lblk - es.es_lblk);
541 if (retval > map->m_len)
542 retval = map->m_len;
543 map->m_len = retval;
544 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
545 map->m_pblk = 0;
546 retval = es.es_len - (map->m_lblk - es.es_lblk);
547 if (retval > map->m_len)
548 retval = map->m_len;
549 map->m_len = retval;
550 retval = 0;
551 } else {
552 BUG();
553 }
554
555 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
556 return retval;
557#ifdef ES_AGGRESSIVE_TEST
558 ext4_map_blocks_es_recheck(handle, inode, map,
559 &orig_map, flags);
560#endif
561 goto found;
562 }
563 /*
564 * In the query cache no-wait mode, nothing we can do more if we
565 * cannot find extent in the cache.
566 */
567 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
568 return 0;
569
570 /*
571 * Try to see if we can get the block without requesting a new
572 * file system block.
573 */
574 down_read(&EXT4_I(inode)->i_data_sem);
575 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
576 retval = ext4_ext_map_blocks(handle, inode, map, 0);
577 } else {
578 retval = ext4_ind_map_blocks(handle, inode, map, 0);
579 }
580 if (retval > 0) {
581 unsigned int status;
582
583 if (unlikely(retval != map->m_len)) {
584 ext4_warning(inode->i_sb,
585 "ES len assertion failed for inode "
586 "%lu: retval %d != map->m_len %d",
587 inode->i_ino, retval, map->m_len);
588 WARN_ON(1);
589 }
590
591 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
592 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
593 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
594 !(status & EXTENT_STATUS_WRITTEN) &&
595 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
596 map->m_lblk + map->m_len - 1))
597 status |= EXTENT_STATUS_DELAYED;
598 ret = ext4_es_insert_extent(inode, map->m_lblk,
599 map->m_len, map->m_pblk, status);
600 if (ret < 0)
601 retval = ret;
602 }
603 up_read((&EXT4_I(inode)->i_data_sem));
604
605found:
606 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
607 ret = check_block_validity(inode, map);
608 if (ret != 0)
609 return ret;
610 }
611
612 /* If it is only a block(s) look up */
613 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
614 return retval;
615
616 /*
617 * Returns if the blocks have already allocated
618 *
619 * Note that if blocks have been preallocated
620 * ext4_ext_get_block() returns the create = 0
621 * with buffer head unmapped.
622 */
623 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
624 /*
625 * If we need to convert extent to unwritten
626 * we continue and do the actual work in
627 * ext4_ext_map_blocks()
628 */
629 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
630 return retval;
631
632 /*
633 * Here we clear m_flags because after allocating an new extent,
634 * it will be set again.
635 */
636 map->m_flags &= ~EXT4_MAP_FLAGS;
637
638 /*
639 * New blocks allocate and/or writing to unwritten extent
640 * will possibly result in updating i_data, so we take
641 * the write lock of i_data_sem, and call get_block()
642 * with create == 1 flag.
643 */
644 down_write(&EXT4_I(inode)->i_data_sem);
645
646 /*
647 * We need to check for EXT4 here because migrate
648 * could have changed the inode type in between
649 */
650 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
651 retval = ext4_ext_map_blocks(handle, inode, map, flags);
652 } else {
653 retval = ext4_ind_map_blocks(handle, inode, map, flags);
654
655 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
656 /*
657 * We allocated new blocks which will result in
658 * i_data's format changing. Force the migrate
659 * to fail by clearing migrate flags
660 */
661 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
662 }
663
664 /*
665 * Update reserved blocks/metadata blocks after successful
666 * block allocation which had been deferred till now. We don't
667 * support fallocate for non extent files. So we can update
668 * reserve space here.
669 */
670 if ((retval > 0) &&
671 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
672 ext4_da_update_reserve_space(inode, retval, 1);
673 }
674
675 if (retval > 0) {
676 unsigned int status;
677
678 if (unlikely(retval != map->m_len)) {
679 ext4_warning(inode->i_sb,
680 "ES len assertion failed for inode "
681 "%lu: retval %d != map->m_len %d",
682 inode->i_ino, retval, map->m_len);
683 WARN_ON(1);
684 }
685
686 /*
687 * We have to zeroout blocks before inserting them into extent
688 * status tree. Otherwise someone could look them up there and
689 * use them before they are really zeroed. We also have to
690 * unmap metadata before zeroing as otherwise writeback can
691 * overwrite zeros with stale data from block device.
692 */
693 if (flags & EXT4_GET_BLOCKS_ZERO &&
694 map->m_flags & EXT4_MAP_MAPPED &&
695 map->m_flags & EXT4_MAP_NEW) {
696 ret = ext4_issue_zeroout(inode, map->m_lblk,
697 map->m_pblk, map->m_len);
698 if (ret) {
699 retval = ret;
700 goto out_sem;
701 }
702 }
703
704 /*
705 * If the extent has been zeroed out, we don't need to update
706 * extent status tree.
707 */
708 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
709 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
710 if (ext4_es_is_written(&es))
711 goto out_sem;
712 }
713 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
714 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
715 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
716 !(status & EXTENT_STATUS_WRITTEN) &&
717 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
718 map->m_lblk + map->m_len - 1))
719 status |= EXTENT_STATUS_DELAYED;
720 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
721 map->m_pblk, status);
722 if (ret < 0) {
723 retval = ret;
724 goto out_sem;
725 }
726 }
727
728out_sem:
729 up_write((&EXT4_I(inode)->i_data_sem));
730 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
731 ret = check_block_validity(inode, map);
732 if (ret != 0)
733 return ret;
734
735 /*
736 * Inodes with freshly allocated blocks where contents will be
737 * visible after transaction commit must be on transaction's
738 * ordered data list.
739 */
740 if (map->m_flags & EXT4_MAP_NEW &&
741 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
742 !(flags & EXT4_GET_BLOCKS_ZERO) &&
743 !ext4_is_quota_file(inode) &&
744 ext4_should_order_data(inode)) {
745 loff_t start_byte =
746 (loff_t)map->m_lblk << inode->i_blkbits;
747 loff_t length = (loff_t)map->m_len << inode->i_blkbits;
748
749 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
750 ret = ext4_jbd2_inode_add_wait(handle, inode,
751 start_byte, length);
752 else
753 ret = ext4_jbd2_inode_add_write(handle, inode,
754 start_byte, length);
755 if (ret)
756 return ret;
757 }
758 }
759 if (retval > 0 && (map->m_flags & EXT4_MAP_UNWRITTEN ||
760 map->m_flags & EXT4_MAP_MAPPED))
761 ext4_fc_track_range(handle, inode, map->m_lblk,
762 map->m_lblk + map->m_len - 1);
763 if (retval < 0)
764 ext_debug(inode, "failed with err %d\n", retval);
765 return retval;
766}
767
768/*
769 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
770 * we have to be careful as someone else may be manipulating b_state as well.
771 */
772static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
773{
774 unsigned long old_state;
775 unsigned long new_state;
776
777 flags &= EXT4_MAP_FLAGS;
778
779 /* Dummy buffer_head? Set non-atomically. */
780 if (!bh->b_page) {
781 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
782 return;
783 }
784 /*
785 * Someone else may be modifying b_state. Be careful! This is ugly but
786 * once we get rid of using bh as a container for mapping information
787 * to pass to / from get_block functions, this can go away.
788 */
789 do {
790 old_state = READ_ONCE(bh->b_state);
791 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
792 } while (unlikely(
793 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
794}
795
796static int _ext4_get_block(struct inode *inode, sector_t iblock,
797 struct buffer_head *bh, int flags)
798{
799 struct ext4_map_blocks map;
800 int ret = 0;
801
802 if (ext4_has_inline_data(inode))
803 return -ERANGE;
804
805 map.m_lblk = iblock;
806 map.m_len = bh->b_size >> inode->i_blkbits;
807
808 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
809 flags);
810 if (ret > 0) {
811 map_bh(bh, inode->i_sb, map.m_pblk);
812 ext4_update_bh_state(bh, map.m_flags);
813 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
814 ret = 0;
815 } else if (ret == 0) {
816 /* hole case, need to fill in bh->b_size */
817 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
818 }
819 return ret;
820}
821
822int ext4_get_block(struct inode *inode, sector_t iblock,
823 struct buffer_head *bh, int create)
824{
825 return _ext4_get_block(inode, iblock, bh,
826 create ? EXT4_GET_BLOCKS_CREATE : 0);
827}
828
829/*
830 * Get block function used when preparing for buffered write if we require
831 * creating an unwritten extent if blocks haven't been allocated. The extent
832 * will be converted to written after the IO is complete.
833 */
834int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
835 struct buffer_head *bh_result, int create)
836{
837 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
838 inode->i_ino, create);
839 return _ext4_get_block(inode, iblock, bh_result,
840 EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT);
841}
842
843/* Maximum number of blocks we map for direct IO at once. */
844#define DIO_MAX_BLOCKS 4096
845
846/*
847 * `handle' can be NULL if create is zero
848 */
849struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
850 ext4_lblk_t block, int map_flags)
851{
852 struct ext4_map_blocks map;
853 struct buffer_head *bh;
854 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
855 bool nowait = map_flags & EXT4_GET_BLOCKS_CACHED_NOWAIT;
856 int err;
857
858 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
859 || handle != NULL || create == 0);
860 ASSERT(create == 0 || !nowait);
861
862 map.m_lblk = block;
863 map.m_len = 1;
864 err = ext4_map_blocks(handle, inode, &map, map_flags);
865
866 if (err == 0)
867 return create ? ERR_PTR(-ENOSPC) : NULL;
868 if (err < 0)
869 return ERR_PTR(err);
870
871 if (nowait)
872 return sb_find_get_block(inode->i_sb, map.m_pblk);
873
874 bh = sb_getblk(inode->i_sb, map.m_pblk);
875 if (unlikely(!bh))
876 return ERR_PTR(-ENOMEM);
877 if (map.m_flags & EXT4_MAP_NEW) {
878 ASSERT(create != 0);
879 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
880 || (handle != NULL));
881
882 /*
883 * Now that we do not always journal data, we should
884 * keep in mind whether this should always journal the
885 * new buffer as metadata. For now, regular file
886 * writes use ext4_get_block instead, so it's not a
887 * problem.
888 */
889 lock_buffer(bh);
890 BUFFER_TRACE(bh, "call get_create_access");
891 err = ext4_journal_get_create_access(handle, inode->i_sb, bh,
892 EXT4_JTR_NONE);
893 if (unlikely(err)) {
894 unlock_buffer(bh);
895 goto errout;
896 }
897 if (!buffer_uptodate(bh)) {
898 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
899 set_buffer_uptodate(bh);
900 }
901 unlock_buffer(bh);
902 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
903 err = ext4_handle_dirty_metadata(handle, inode, bh);
904 if (unlikely(err))
905 goto errout;
906 } else
907 BUFFER_TRACE(bh, "not a new buffer");
908 return bh;
909errout:
910 brelse(bh);
911 return ERR_PTR(err);
912}
913
914struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
915 ext4_lblk_t block, int map_flags)
916{
917 struct buffer_head *bh;
918 int ret;
919
920 bh = ext4_getblk(handle, inode, block, map_flags);
921 if (IS_ERR(bh))
922 return bh;
923 if (!bh || ext4_buffer_uptodate(bh))
924 return bh;
925
926 ret = ext4_read_bh_lock(bh, REQ_META | REQ_PRIO, true);
927 if (ret) {
928 put_bh(bh);
929 return ERR_PTR(ret);
930 }
931 return bh;
932}
933
934/* Read a contiguous batch of blocks. */
935int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
936 bool wait, struct buffer_head **bhs)
937{
938 int i, err;
939
940 for (i = 0; i < bh_count; i++) {
941 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
942 if (IS_ERR(bhs[i])) {
943 err = PTR_ERR(bhs[i]);
944 bh_count = i;
945 goto out_brelse;
946 }
947 }
948
949 for (i = 0; i < bh_count; i++)
950 /* Note that NULL bhs[i] is valid because of holes. */
951 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
952 ext4_read_bh_lock(bhs[i], REQ_META | REQ_PRIO, false);
953
954 if (!wait)
955 return 0;
956
957 for (i = 0; i < bh_count; i++)
958 if (bhs[i])
959 wait_on_buffer(bhs[i]);
960
961 for (i = 0; i < bh_count; i++) {
962 if (bhs[i] && !buffer_uptodate(bhs[i])) {
963 err = -EIO;
964 goto out_brelse;
965 }
966 }
967 return 0;
968
969out_brelse:
970 for (i = 0; i < bh_count; i++) {
971 brelse(bhs[i]);
972 bhs[i] = NULL;
973 }
974 return err;
975}
976
977int ext4_walk_page_buffers(handle_t *handle, struct inode *inode,
978 struct buffer_head *head,
979 unsigned from,
980 unsigned to,
981 int *partial,
982 int (*fn)(handle_t *handle, struct inode *inode,
983 struct buffer_head *bh))
984{
985 struct buffer_head *bh;
986 unsigned block_start, block_end;
987 unsigned blocksize = head->b_size;
988 int err, ret = 0;
989 struct buffer_head *next;
990
991 for (bh = head, block_start = 0;
992 ret == 0 && (bh != head || !block_start);
993 block_start = block_end, bh = next) {
994 next = bh->b_this_page;
995 block_end = block_start + blocksize;
996 if (block_end <= from || block_start >= to) {
997 if (partial && !buffer_uptodate(bh))
998 *partial = 1;
999 continue;
1000 }
1001 err = (*fn)(handle, inode, bh);
1002 if (!ret)
1003 ret = err;
1004 }
1005 return ret;
1006}
1007
1008/*
1009 * To preserve ordering, it is essential that the hole instantiation and
1010 * the data write be encapsulated in a single transaction. We cannot
1011 * close off a transaction and start a new one between the ext4_get_block()
1012 * and the commit_write(). So doing the jbd2_journal_start at the start of
1013 * prepare_write() is the right place.
1014 *
1015 * Also, this function can nest inside ext4_writepage(). In that case, we
1016 * *know* that ext4_writepage() has generated enough buffer credits to do the
1017 * whole page. So we won't block on the journal in that case, which is good,
1018 * because the caller may be PF_MEMALLOC.
1019 *
1020 * By accident, ext4 can be reentered when a transaction is open via
1021 * quota file writes. If we were to commit the transaction while thus
1022 * reentered, there can be a deadlock - we would be holding a quota
1023 * lock, and the commit would never complete if another thread had a
1024 * transaction open and was blocking on the quota lock - a ranking
1025 * violation.
1026 *
1027 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1028 * will _not_ run commit under these circumstances because handle->h_ref
1029 * is elevated. We'll still have enough credits for the tiny quotafile
1030 * write.
1031 */
1032int do_journal_get_write_access(handle_t *handle, struct inode *inode,
1033 struct buffer_head *bh)
1034{
1035 int dirty = buffer_dirty(bh);
1036 int ret;
1037
1038 if (!buffer_mapped(bh) || buffer_freed(bh))
1039 return 0;
1040 /*
1041 * __block_write_begin() could have dirtied some buffers. Clean
1042 * the dirty bit as jbd2_journal_get_write_access() could complain
1043 * otherwise about fs integrity issues. Setting of the dirty bit
1044 * by __block_write_begin() isn't a real problem here as we clear
1045 * the bit before releasing a page lock and thus writeback cannot
1046 * ever write the buffer.
1047 */
1048 if (dirty)
1049 clear_buffer_dirty(bh);
1050 BUFFER_TRACE(bh, "get write access");
1051 ret = ext4_journal_get_write_access(handle, inode->i_sb, bh,
1052 EXT4_JTR_NONE);
1053 if (!ret && dirty)
1054 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1055 return ret;
1056}
1057
1058#ifdef CONFIG_FS_ENCRYPTION
1059static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1060 get_block_t *get_block)
1061{
1062 unsigned from = pos & (PAGE_SIZE - 1);
1063 unsigned to = from + len;
1064 struct inode *inode = page->mapping->host;
1065 unsigned block_start, block_end;
1066 sector_t block;
1067 int err = 0;
1068 unsigned blocksize = inode->i_sb->s_blocksize;
1069 unsigned bbits;
1070 struct buffer_head *bh, *head, *wait[2];
1071 int nr_wait = 0;
1072 int i;
1073
1074 BUG_ON(!PageLocked(page));
1075 BUG_ON(from > PAGE_SIZE);
1076 BUG_ON(to > PAGE_SIZE);
1077 BUG_ON(from > to);
1078
1079 if (!page_has_buffers(page))
1080 create_empty_buffers(page, blocksize, 0);
1081 head = page_buffers(page);
1082 bbits = ilog2(blocksize);
1083 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1084
1085 for (bh = head, block_start = 0; bh != head || !block_start;
1086 block++, block_start = block_end, bh = bh->b_this_page) {
1087 block_end = block_start + blocksize;
1088 if (block_end <= from || block_start >= to) {
1089 if (PageUptodate(page)) {
1090 set_buffer_uptodate(bh);
1091 }
1092 continue;
1093 }
1094 if (buffer_new(bh))
1095 clear_buffer_new(bh);
1096 if (!buffer_mapped(bh)) {
1097 WARN_ON(bh->b_size != blocksize);
1098 err = get_block(inode, block, bh, 1);
1099 if (err)
1100 break;
1101 if (buffer_new(bh)) {
1102 if (PageUptodate(page)) {
1103 clear_buffer_new(bh);
1104 set_buffer_uptodate(bh);
1105 mark_buffer_dirty(bh);
1106 continue;
1107 }
1108 if (block_end > to || block_start < from)
1109 zero_user_segments(page, to, block_end,
1110 block_start, from);
1111 continue;
1112 }
1113 }
1114 if (PageUptodate(page)) {
1115 set_buffer_uptodate(bh);
1116 continue;
1117 }
1118 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1119 !buffer_unwritten(bh) &&
1120 (block_start < from || block_end > to)) {
1121 ext4_read_bh_lock(bh, 0, false);
1122 wait[nr_wait++] = bh;
1123 }
1124 }
1125 /*
1126 * If we issued read requests, let them complete.
1127 */
1128 for (i = 0; i < nr_wait; i++) {
1129 wait_on_buffer(wait[i]);
1130 if (!buffer_uptodate(wait[i]))
1131 err = -EIO;
1132 }
1133 if (unlikely(err)) {
1134 page_zero_new_buffers(page, from, to);
1135 } else if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
1136 for (i = 0; i < nr_wait; i++) {
1137 int err2;
1138
1139 err2 = fscrypt_decrypt_pagecache_blocks(page, blocksize,
1140 bh_offset(wait[i]));
1141 if (err2) {
1142 clear_buffer_uptodate(wait[i]);
1143 err = err2;
1144 }
1145 }
1146 }
1147
1148 return err;
1149}
1150#endif
1151
1152static int ext4_write_begin(struct file *file, struct address_space *mapping,
1153 loff_t pos, unsigned len,
1154 struct page **pagep, void **fsdata)
1155{
1156 struct inode *inode = mapping->host;
1157 int ret, needed_blocks;
1158 handle_t *handle;
1159 int retries = 0;
1160 struct page *page;
1161 pgoff_t index;
1162 unsigned from, to;
1163
1164 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1165 return -EIO;
1166
1167 trace_ext4_write_begin(inode, pos, len);
1168 /*
1169 * Reserve one block more for addition to orphan list in case
1170 * we allocate blocks but write fails for some reason
1171 */
1172 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1173 index = pos >> PAGE_SHIFT;
1174 from = pos & (PAGE_SIZE - 1);
1175 to = from + len;
1176
1177 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1178 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1179 pagep);
1180 if (ret < 0)
1181 return ret;
1182 if (ret == 1)
1183 return 0;
1184 }
1185
1186 /*
1187 * grab_cache_page_write_begin() can take a long time if the
1188 * system is thrashing due to memory pressure, or if the page
1189 * is being written back. So grab it first before we start
1190 * the transaction handle. This also allows us to allocate
1191 * the page (if needed) without using GFP_NOFS.
1192 */
1193retry_grab:
1194 page = grab_cache_page_write_begin(mapping, index);
1195 if (!page)
1196 return -ENOMEM;
1197 /*
1198 * The same as page allocation, we prealloc buffer heads before
1199 * starting the handle.
1200 */
1201 if (!page_has_buffers(page))
1202 create_empty_buffers(page, inode->i_sb->s_blocksize, 0);
1203
1204 unlock_page(page);
1205
1206retry_journal:
1207 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1208 if (IS_ERR(handle)) {
1209 put_page(page);
1210 return PTR_ERR(handle);
1211 }
1212
1213 lock_page(page);
1214 if (page->mapping != mapping) {
1215 /* The page got truncated from under us */
1216 unlock_page(page);
1217 put_page(page);
1218 ext4_journal_stop(handle);
1219 goto retry_grab;
1220 }
1221 /* In case writeback began while the page was unlocked */
1222 wait_for_stable_page(page);
1223
1224#ifdef CONFIG_FS_ENCRYPTION
1225 if (ext4_should_dioread_nolock(inode))
1226 ret = ext4_block_write_begin(page, pos, len,
1227 ext4_get_block_unwritten);
1228 else
1229 ret = ext4_block_write_begin(page, pos, len,
1230 ext4_get_block);
1231#else
1232 if (ext4_should_dioread_nolock(inode))
1233 ret = __block_write_begin(page, pos, len,
1234 ext4_get_block_unwritten);
1235 else
1236 ret = __block_write_begin(page, pos, len, ext4_get_block);
1237#endif
1238 if (!ret && ext4_should_journal_data(inode)) {
1239 ret = ext4_walk_page_buffers(handle, inode,
1240 page_buffers(page), from, to, NULL,
1241 do_journal_get_write_access);
1242 }
1243
1244 if (ret) {
1245 bool extended = (pos + len > inode->i_size) &&
1246 !ext4_verity_in_progress(inode);
1247
1248 unlock_page(page);
1249 /*
1250 * __block_write_begin may have instantiated a few blocks
1251 * outside i_size. Trim these off again. Don't need
1252 * i_size_read because we hold i_rwsem.
1253 *
1254 * Add inode to orphan list in case we crash before
1255 * truncate finishes
1256 */
1257 if (extended && ext4_can_truncate(inode))
1258 ext4_orphan_add(handle, inode);
1259
1260 ext4_journal_stop(handle);
1261 if (extended) {
1262 ext4_truncate_failed_write(inode);
1263 /*
1264 * If truncate failed early the inode might
1265 * still be on the orphan list; we need to
1266 * make sure the inode is removed from the
1267 * orphan list in that case.
1268 */
1269 if (inode->i_nlink)
1270 ext4_orphan_del(NULL, inode);
1271 }
1272
1273 if (ret == -ENOSPC &&
1274 ext4_should_retry_alloc(inode->i_sb, &retries))
1275 goto retry_journal;
1276 put_page(page);
1277 return ret;
1278 }
1279 *pagep = page;
1280 return ret;
1281}
1282
1283/* For write_end() in data=journal mode */
1284static int write_end_fn(handle_t *handle, struct inode *inode,
1285 struct buffer_head *bh)
1286{
1287 int ret;
1288 if (!buffer_mapped(bh) || buffer_freed(bh))
1289 return 0;
1290 set_buffer_uptodate(bh);
1291 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1292 clear_buffer_meta(bh);
1293 clear_buffer_prio(bh);
1294 return ret;
1295}
1296
1297/*
1298 * We need to pick up the new inode size which generic_commit_write gave us
1299 * `file' can be NULL - eg, when called from page_symlink().
1300 *
1301 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1302 * buffers are managed internally.
1303 */
1304static int ext4_write_end(struct file *file,
1305 struct address_space *mapping,
1306 loff_t pos, unsigned len, unsigned copied,
1307 struct page *page, void *fsdata)
1308{
1309 handle_t *handle = ext4_journal_current_handle();
1310 struct inode *inode = mapping->host;
1311 loff_t old_size = inode->i_size;
1312 int ret = 0, ret2;
1313 int i_size_changed = 0;
1314 bool verity = ext4_verity_in_progress(inode);
1315
1316 trace_ext4_write_end(inode, pos, len, copied);
1317
1318 if (ext4_has_inline_data(inode) &&
1319 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
1320 return ext4_write_inline_data_end(inode, pos, len, copied, page);
1321
1322 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1323 /*
1324 * it's important to update i_size while still holding page lock:
1325 * page writeout could otherwise come in and zero beyond i_size.
1326 *
1327 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1328 * blocks are being written past EOF, so skip the i_size update.
1329 */
1330 if (!verity)
1331 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1332 unlock_page(page);
1333 put_page(page);
1334
1335 if (old_size < pos && !verity)
1336 pagecache_isize_extended(inode, old_size, pos);
1337 /*
1338 * Don't mark the inode dirty under page lock. First, it unnecessarily
1339 * makes the holding time of page lock longer. Second, it forces lock
1340 * ordering of page lock and transaction start for journaling
1341 * filesystems.
1342 */
1343 if (i_size_changed)
1344 ret = ext4_mark_inode_dirty(handle, inode);
1345
1346 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1347 /* if we have allocated more blocks and copied
1348 * less. We will have blocks allocated outside
1349 * inode->i_size. So truncate them
1350 */
1351 ext4_orphan_add(handle, inode);
1352
1353 ret2 = ext4_journal_stop(handle);
1354 if (!ret)
1355 ret = ret2;
1356
1357 if (pos + len > inode->i_size && !verity) {
1358 ext4_truncate_failed_write(inode);
1359 /*
1360 * If truncate failed early the inode might still be
1361 * on the orphan list; we need to make sure the inode
1362 * is removed from the orphan list in that case.
1363 */
1364 if (inode->i_nlink)
1365 ext4_orphan_del(NULL, inode);
1366 }
1367
1368 return ret ? ret : copied;
1369}
1370
1371/*
1372 * This is a private version of page_zero_new_buffers() which doesn't
1373 * set the buffer to be dirty, since in data=journalled mode we need
1374 * to call ext4_handle_dirty_metadata() instead.
1375 */
1376static void ext4_journalled_zero_new_buffers(handle_t *handle,
1377 struct inode *inode,
1378 struct page *page,
1379 unsigned from, unsigned to)
1380{
1381 unsigned int block_start = 0, block_end;
1382 struct buffer_head *head, *bh;
1383
1384 bh = head = page_buffers(page);
1385 do {
1386 block_end = block_start + bh->b_size;
1387 if (buffer_new(bh)) {
1388 if (block_end > from && block_start < to) {
1389 if (!PageUptodate(page)) {
1390 unsigned start, size;
1391
1392 start = max(from, block_start);
1393 size = min(to, block_end) - start;
1394
1395 zero_user(page, start, size);
1396 write_end_fn(handle, inode, bh);
1397 }
1398 clear_buffer_new(bh);
1399 }
1400 }
1401 block_start = block_end;
1402 bh = bh->b_this_page;
1403 } while (bh != head);
1404}
1405
1406static int ext4_journalled_write_end(struct file *file,
1407 struct address_space *mapping,
1408 loff_t pos, unsigned len, unsigned copied,
1409 struct page *page, void *fsdata)
1410{
1411 handle_t *handle = ext4_journal_current_handle();
1412 struct inode *inode = mapping->host;
1413 loff_t old_size = inode->i_size;
1414 int ret = 0, ret2;
1415 int partial = 0;
1416 unsigned from, to;
1417 int size_changed = 0;
1418 bool verity = ext4_verity_in_progress(inode);
1419
1420 trace_ext4_journalled_write_end(inode, pos, len, copied);
1421 from = pos & (PAGE_SIZE - 1);
1422 to = from + len;
1423
1424 BUG_ON(!ext4_handle_valid(handle));
1425
1426 if (ext4_has_inline_data(inode))
1427 return ext4_write_inline_data_end(inode, pos, len, copied, page);
1428
1429 if (unlikely(copied < len) && !PageUptodate(page)) {
1430 copied = 0;
1431 ext4_journalled_zero_new_buffers(handle, inode, page, from, to);
1432 } else {
1433 if (unlikely(copied < len))
1434 ext4_journalled_zero_new_buffers(handle, inode, page,
1435 from + copied, to);
1436 ret = ext4_walk_page_buffers(handle, inode, page_buffers(page),
1437 from, from + copied, &partial,
1438 write_end_fn);
1439 if (!partial)
1440 SetPageUptodate(page);
1441 }
1442 if (!verity)
1443 size_changed = ext4_update_inode_size(inode, pos + copied);
1444 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1445 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1446 unlock_page(page);
1447 put_page(page);
1448
1449 if (old_size < pos && !verity)
1450 pagecache_isize_extended(inode, old_size, pos);
1451
1452 if (size_changed) {
1453 ret2 = ext4_mark_inode_dirty(handle, inode);
1454 if (!ret)
1455 ret = ret2;
1456 }
1457
1458 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1459 /* if we have allocated more blocks and copied
1460 * less. We will have blocks allocated outside
1461 * inode->i_size. So truncate them
1462 */
1463 ext4_orphan_add(handle, inode);
1464
1465 ret2 = ext4_journal_stop(handle);
1466 if (!ret)
1467 ret = ret2;
1468 if (pos + len > inode->i_size && !verity) {
1469 ext4_truncate_failed_write(inode);
1470 /*
1471 * If truncate failed early the inode might still be
1472 * on the orphan list; we need to make sure the inode
1473 * is removed from the orphan list in that case.
1474 */
1475 if (inode->i_nlink)
1476 ext4_orphan_del(NULL, inode);
1477 }
1478
1479 return ret ? ret : copied;
1480}
1481
1482/*
1483 * Reserve space for a single cluster
1484 */
1485static int ext4_da_reserve_space(struct inode *inode)
1486{
1487 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1488 struct ext4_inode_info *ei = EXT4_I(inode);
1489 int ret;
1490
1491 /*
1492 * We will charge metadata quota at writeout time; this saves
1493 * us from metadata over-estimation, though we may go over by
1494 * a small amount in the end. Here we just reserve for data.
1495 */
1496 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1497 if (ret)
1498 return ret;
1499
1500 spin_lock(&ei->i_block_reservation_lock);
1501 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1502 spin_unlock(&ei->i_block_reservation_lock);
1503 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1504 return -ENOSPC;
1505 }
1506 ei->i_reserved_data_blocks++;
1507 trace_ext4_da_reserve_space(inode);
1508 spin_unlock(&ei->i_block_reservation_lock);
1509
1510 return 0; /* success */
1511}
1512
1513void ext4_da_release_space(struct inode *inode, int to_free)
1514{
1515 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1516 struct ext4_inode_info *ei = EXT4_I(inode);
1517
1518 if (!to_free)
1519 return; /* Nothing to release, exit */
1520
1521 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1522
1523 trace_ext4_da_release_space(inode, to_free);
1524 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1525 /*
1526 * if there aren't enough reserved blocks, then the
1527 * counter is messed up somewhere. Since this
1528 * function is called from invalidate page, it's
1529 * harmless to return without any action.
1530 */
1531 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1532 "ino %lu, to_free %d with only %d reserved "
1533 "data blocks", inode->i_ino, to_free,
1534 ei->i_reserved_data_blocks);
1535 WARN_ON(1);
1536 to_free = ei->i_reserved_data_blocks;
1537 }
1538 ei->i_reserved_data_blocks -= to_free;
1539
1540 /* update fs dirty data blocks counter */
1541 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1542
1543 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1544
1545 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1546}
1547
1548/*
1549 * Delayed allocation stuff
1550 */
1551
1552struct mpage_da_data {
1553 /* These are input fields for ext4_do_writepages() */
1554 struct inode *inode;
1555 struct writeback_control *wbc;
1556 unsigned int can_map:1; /* Can writepages call map blocks? */
1557
1558 /* These are internal state of ext4_do_writepages() */
1559 pgoff_t first_page; /* The first page to write */
1560 pgoff_t next_page; /* Current page to examine */
1561 pgoff_t last_page; /* Last page to examine */
1562 /*
1563 * Extent to map - this can be after first_page because that can be
1564 * fully mapped. We somewhat abuse m_flags to store whether the extent
1565 * is delalloc or unwritten.
1566 */
1567 struct ext4_map_blocks map;
1568 struct ext4_io_submit io_submit; /* IO submission data */
1569 unsigned int do_map:1;
1570 unsigned int scanned_until_end:1;
1571};
1572
1573static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1574 bool invalidate)
1575{
1576 unsigned nr, i;
1577 pgoff_t index, end;
1578 struct folio_batch fbatch;
1579 struct inode *inode = mpd->inode;
1580 struct address_space *mapping = inode->i_mapping;
1581
1582 /* This is necessary when next_page == 0. */
1583 if (mpd->first_page >= mpd->next_page)
1584 return;
1585
1586 mpd->scanned_until_end = 0;
1587 index = mpd->first_page;
1588 end = mpd->next_page - 1;
1589 if (invalidate) {
1590 ext4_lblk_t start, last;
1591 start = index << (PAGE_SHIFT - inode->i_blkbits);
1592 last = end << (PAGE_SHIFT - inode->i_blkbits);
1593
1594 /*
1595 * avoid racing with extent status tree scans made by
1596 * ext4_insert_delayed_block()
1597 */
1598 down_write(&EXT4_I(inode)->i_data_sem);
1599 ext4_es_remove_extent(inode, start, last - start + 1);
1600 up_write(&EXT4_I(inode)->i_data_sem);
1601 }
1602
1603 folio_batch_init(&fbatch);
1604 while (index <= end) {
1605 nr = filemap_get_folios(mapping, &index, end, &fbatch);
1606 if (nr == 0)
1607 break;
1608 for (i = 0; i < nr; i++) {
1609 struct folio *folio = fbatch.folios[i];
1610
1611 if (folio->index < mpd->first_page)
1612 continue;
1613 if (folio->index + folio_nr_pages(folio) - 1 > end)
1614 continue;
1615 BUG_ON(!folio_test_locked(folio));
1616 BUG_ON(folio_test_writeback(folio));
1617 if (invalidate) {
1618 if (folio_mapped(folio))
1619 folio_clear_dirty_for_io(folio);
1620 block_invalidate_folio(folio, 0,
1621 folio_size(folio));
1622 folio_clear_uptodate(folio);
1623 }
1624 folio_unlock(folio);
1625 }
1626 folio_batch_release(&fbatch);
1627 }
1628}
1629
1630static void ext4_print_free_blocks(struct inode *inode)
1631{
1632 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1633 struct super_block *sb = inode->i_sb;
1634 struct ext4_inode_info *ei = EXT4_I(inode);
1635
1636 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1637 EXT4_C2B(EXT4_SB(inode->i_sb),
1638 ext4_count_free_clusters(sb)));
1639 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1640 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1641 (long long) EXT4_C2B(EXT4_SB(sb),
1642 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1643 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1644 (long long) EXT4_C2B(EXT4_SB(sb),
1645 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1646 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1647 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1648 ei->i_reserved_data_blocks);
1649 return;
1650}
1651
1652static int ext4_bh_delay_or_unwritten(handle_t *handle, struct inode *inode,
1653 struct buffer_head *bh)
1654{
1655 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1656}
1657
1658/*
1659 * ext4_insert_delayed_block - adds a delayed block to the extents status
1660 * tree, incrementing the reserved cluster/block
1661 * count or making a pending reservation
1662 * where needed
1663 *
1664 * @inode - file containing the newly added block
1665 * @lblk - logical block to be added
1666 *
1667 * Returns 0 on success, negative error code on failure.
1668 */
1669static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1670{
1671 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1672 int ret;
1673 bool allocated = false;
1674 bool reserved = false;
1675
1676 /*
1677 * If the cluster containing lblk is shared with a delayed,
1678 * written, or unwritten extent in a bigalloc file system, it's
1679 * already been accounted for and does not need to be reserved.
1680 * A pending reservation must be made for the cluster if it's
1681 * shared with a written or unwritten extent and doesn't already
1682 * have one. Written and unwritten extents can be purged from the
1683 * extents status tree if the system is under memory pressure, so
1684 * it's necessary to examine the extent tree if a search of the
1685 * extents status tree doesn't get a match.
1686 */
1687 if (sbi->s_cluster_ratio == 1) {
1688 ret = ext4_da_reserve_space(inode);
1689 if (ret != 0) /* ENOSPC */
1690 goto errout;
1691 reserved = true;
1692 } else { /* bigalloc */
1693 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1694 if (!ext4_es_scan_clu(inode,
1695 &ext4_es_is_mapped, lblk)) {
1696 ret = ext4_clu_mapped(inode,
1697 EXT4_B2C(sbi, lblk));
1698 if (ret < 0)
1699 goto errout;
1700 if (ret == 0) {
1701 ret = ext4_da_reserve_space(inode);
1702 if (ret != 0) /* ENOSPC */
1703 goto errout;
1704 reserved = true;
1705 } else {
1706 allocated = true;
1707 }
1708 } else {
1709 allocated = true;
1710 }
1711 }
1712 }
1713
1714 ret = ext4_es_insert_delayed_block(inode, lblk, allocated);
1715 if (ret && reserved)
1716 ext4_da_release_space(inode, 1);
1717
1718errout:
1719 return ret;
1720}
1721
1722/*
1723 * This function is grabs code from the very beginning of
1724 * ext4_map_blocks, but assumes that the caller is from delayed write
1725 * time. This function looks up the requested blocks and sets the
1726 * buffer delay bit under the protection of i_data_sem.
1727 */
1728static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1729 struct ext4_map_blocks *map,
1730 struct buffer_head *bh)
1731{
1732 struct extent_status es;
1733 int retval;
1734 sector_t invalid_block = ~((sector_t) 0xffff);
1735#ifdef ES_AGGRESSIVE_TEST
1736 struct ext4_map_blocks orig_map;
1737
1738 memcpy(&orig_map, map, sizeof(*map));
1739#endif
1740
1741 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1742 invalid_block = ~0;
1743
1744 map->m_flags = 0;
1745 ext_debug(inode, "max_blocks %u, logical block %lu\n", map->m_len,
1746 (unsigned long) map->m_lblk);
1747
1748 /* Lookup extent status tree firstly */
1749 if (ext4_es_lookup_extent(inode, iblock, NULL, &es)) {
1750 if (ext4_es_is_hole(&es)) {
1751 retval = 0;
1752 down_read(&EXT4_I(inode)->i_data_sem);
1753 goto add_delayed;
1754 }
1755
1756 /*
1757 * Delayed extent could be allocated by fallocate.
1758 * So we need to check it.
1759 */
1760 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1761 map_bh(bh, inode->i_sb, invalid_block);
1762 set_buffer_new(bh);
1763 set_buffer_delay(bh);
1764 return 0;
1765 }
1766
1767 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1768 retval = es.es_len - (iblock - es.es_lblk);
1769 if (retval > map->m_len)
1770 retval = map->m_len;
1771 map->m_len = retval;
1772 if (ext4_es_is_written(&es))
1773 map->m_flags |= EXT4_MAP_MAPPED;
1774 else if (ext4_es_is_unwritten(&es))
1775 map->m_flags |= EXT4_MAP_UNWRITTEN;
1776 else
1777 BUG();
1778
1779#ifdef ES_AGGRESSIVE_TEST
1780 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1781#endif
1782 return retval;
1783 }
1784
1785 /*
1786 * Try to see if we can get the block without requesting a new
1787 * file system block.
1788 */
1789 down_read(&EXT4_I(inode)->i_data_sem);
1790 if (ext4_has_inline_data(inode))
1791 retval = 0;
1792 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1793 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1794 else
1795 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1796
1797add_delayed:
1798 if (retval == 0) {
1799 int ret;
1800
1801 /*
1802 * XXX: __block_prepare_write() unmaps passed block,
1803 * is it OK?
1804 */
1805
1806 ret = ext4_insert_delayed_block(inode, map->m_lblk);
1807 if (ret != 0) {
1808 retval = ret;
1809 goto out_unlock;
1810 }
1811
1812 map_bh(bh, inode->i_sb, invalid_block);
1813 set_buffer_new(bh);
1814 set_buffer_delay(bh);
1815 } else if (retval > 0) {
1816 int ret;
1817 unsigned int status;
1818
1819 if (unlikely(retval != map->m_len)) {
1820 ext4_warning(inode->i_sb,
1821 "ES len assertion failed for inode "
1822 "%lu: retval %d != map->m_len %d",
1823 inode->i_ino, retval, map->m_len);
1824 WARN_ON(1);
1825 }
1826
1827 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1828 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1829 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1830 map->m_pblk, status);
1831 if (ret != 0)
1832 retval = ret;
1833 }
1834
1835out_unlock:
1836 up_read((&EXT4_I(inode)->i_data_sem));
1837
1838 return retval;
1839}
1840
1841/*
1842 * This is a special get_block_t callback which is used by
1843 * ext4_da_write_begin(). It will either return mapped block or
1844 * reserve space for a single block.
1845 *
1846 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1847 * We also have b_blocknr = -1 and b_bdev initialized properly
1848 *
1849 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1850 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1851 * initialized properly.
1852 */
1853int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1854 struct buffer_head *bh, int create)
1855{
1856 struct ext4_map_blocks map;
1857 int ret = 0;
1858
1859 BUG_ON(create == 0);
1860 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1861
1862 map.m_lblk = iblock;
1863 map.m_len = 1;
1864
1865 /*
1866 * first, we need to know whether the block is allocated already
1867 * preallocated blocks are unmapped but should treated
1868 * the same as allocated blocks.
1869 */
1870 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1871 if (ret <= 0)
1872 return ret;
1873
1874 map_bh(bh, inode->i_sb, map.m_pblk);
1875 ext4_update_bh_state(bh, map.m_flags);
1876
1877 if (buffer_unwritten(bh)) {
1878 /* A delayed write to unwritten bh should be marked
1879 * new and mapped. Mapped ensures that we don't do
1880 * get_block multiple times when we write to the same
1881 * offset and new ensures that we do proper zero out
1882 * for partial write.
1883 */
1884 set_buffer_new(bh);
1885 set_buffer_mapped(bh);
1886 }
1887 return 0;
1888}
1889
1890static int __ext4_journalled_writepage(struct page *page,
1891 unsigned int len)
1892{
1893 struct address_space *mapping = page->mapping;
1894 struct inode *inode = mapping->host;
1895 handle_t *handle = NULL;
1896 int ret = 0, err = 0;
1897 int inline_data = ext4_has_inline_data(inode);
1898 struct buffer_head *inode_bh = NULL;
1899 loff_t size;
1900
1901 ClearPageChecked(page);
1902
1903 if (inline_data) {
1904 BUG_ON(page->index != 0);
1905 BUG_ON(len > ext4_get_max_inline_size(inode));
1906 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1907 if (inode_bh == NULL)
1908 goto out;
1909 }
1910 /*
1911 * We need to release the page lock before we start the
1912 * journal, so grab a reference so the page won't disappear
1913 * out from under us.
1914 */
1915 get_page(page);
1916 unlock_page(page);
1917
1918 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1919 ext4_writepage_trans_blocks(inode));
1920 if (IS_ERR(handle)) {
1921 ret = PTR_ERR(handle);
1922 put_page(page);
1923 goto out_no_pagelock;
1924 }
1925 BUG_ON(!ext4_handle_valid(handle));
1926
1927 lock_page(page);
1928 put_page(page);
1929 size = i_size_read(inode);
1930 if (page->mapping != mapping || page_offset(page) > size) {
1931 /* The page got truncated from under us */
1932 ext4_journal_stop(handle);
1933 ret = 0;
1934 goto out;
1935 }
1936
1937 if (inline_data) {
1938 ret = ext4_mark_inode_dirty(handle, inode);
1939 } else {
1940 struct buffer_head *page_bufs = page_buffers(page);
1941
1942 if (page->index == size >> PAGE_SHIFT)
1943 len = size & ~PAGE_MASK;
1944 else
1945 len = PAGE_SIZE;
1946
1947 ret = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
1948 NULL, do_journal_get_write_access);
1949
1950 err = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
1951 NULL, write_end_fn);
1952 }
1953 if (ret == 0)
1954 ret = err;
1955 err = ext4_jbd2_inode_add_write(handle, inode, page_offset(page), len);
1956 if (ret == 0)
1957 ret = err;
1958 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1959 err = ext4_journal_stop(handle);
1960 if (!ret)
1961 ret = err;
1962
1963 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1964out:
1965 unlock_page(page);
1966out_no_pagelock:
1967 brelse(inode_bh);
1968 return ret;
1969}
1970
1971/*
1972 * Note that we don't need to start a transaction unless we're journaling data
1973 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1974 * need to file the inode to the transaction's list in ordered mode because if
1975 * we are writing back data added by write(), the inode is already there and if
1976 * we are writing back data modified via mmap(), no one guarantees in which
1977 * transaction the data will hit the disk. In case we are journaling data, we
1978 * cannot start transaction directly because transaction start ranks above page
1979 * lock so we have to do some magic.
1980 *
1981 * This function can get called via...
1982 * - ext4_writepages after taking page lock (have journal handle)
1983 * - journal_submit_inode_data_buffers (no journal handle)
1984 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1985 * - grab_page_cache when doing write_begin (have journal handle)
1986 *
1987 * We don't do any block allocation in this function. If we have page with
1988 * multiple blocks we need to write those buffer_heads that are mapped. This
1989 * is important for mmaped based write. So if we do with blocksize 1K
1990 * truncate(f, 1024);
1991 * a = mmap(f, 0, 4096);
1992 * a[0] = 'a';
1993 * truncate(f, 4096);
1994 * we have in the page first buffer_head mapped via page_mkwrite call back
1995 * but other buffer_heads would be unmapped but dirty (dirty done via the
1996 * do_wp_page). So writepage should write the first block. If we modify
1997 * the mmap area beyond 1024 we will again get a page_fault and the
1998 * page_mkwrite callback will do the block allocation and mark the
1999 * buffer_heads mapped.
2000 *
2001 * We redirty the page if we have any buffer_heads that is either delay or
2002 * unwritten in the page.
2003 *
2004 * We can get recursively called as show below.
2005 *
2006 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2007 * ext4_writepage()
2008 *
2009 * But since we don't do any block allocation we should not deadlock.
2010 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2011 */
2012static int ext4_writepage(struct page *page,
2013 struct writeback_control *wbc)
2014{
2015 struct folio *folio = page_folio(page);
2016 int ret = 0;
2017 loff_t size;
2018 unsigned int len;
2019 struct buffer_head *page_bufs = NULL;
2020 struct inode *inode = page->mapping->host;
2021 struct ext4_io_submit io_submit;
2022
2023 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2024 folio_invalidate(folio, 0, folio_size(folio));
2025 folio_unlock(folio);
2026 return -EIO;
2027 }
2028
2029 trace_ext4_writepage(page);
2030 size = i_size_read(inode);
2031 if (page->index == size >> PAGE_SHIFT &&
2032 !ext4_verity_in_progress(inode))
2033 len = size & ~PAGE_MASK;
2034 else
2035 len = PAGE_SIZE;
2036
2037 /* Should never happen but for bugs in other kernel subsystems */
2038 if (!page_has_buffers(page)) {
2039 ext4_warning_inode(inode,
2040 "page %lu does not have buffers attached", page->index);
2041 ClearPageDirty(page);
2042 unlock_page(page);
2043 return 0;
2044 }
2045
2046 page_bufs = page_buffers(page);
2047 /*
2048 * We cannot do block allocation or other extent handling in this
2049 * function. If there are buffers needing that, we have to redirty
2050 * the page. But we may reach here when we do a journal commit via
2051 * journal_submit_inode_data_buffers() and in that case we must write
2052 * allocated buffers to achieve data=ordered mode guarantees.
2053 *
2054 * Also, if there is only one buffer per page (the fs block
2055 * size == the page size), if one buffer needs block
2056 * allocation or needs to modify the extent tree to clear the
2057 * unwritten flag, we know that the page can't be written at
2058 * all, so we might as well refuse the write immediately.
2059 * Unfortunately if the block size != page size, we can't as
2060 * easily detect this case using ext4_walk_page_buffers(), but
2061 * for the extremely common case, this is an optimization that
2062 * skips a useless round trip through ext4_bio_write_page().
2063 */
2064 if (ext4_walk_page_buffers(NULL, inode, page_bufs, 0, len, NULL,
2065 ext4_bh_delay_or_unwritten)) {
2066 redirty_page_for_writepage(wbc, page);
2067 if ((current->flags & PF_MEMALLOC) ||
2068 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2069 /*
2070 * For memory cleaning there's no point in writing only
2071 * some buffers. So just bail out. Warn if we came here
2072 * from direct reclaim.
2073 */
2074 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2075 == PF_MEMALLOC);
2076 unlock_page(page);
2077 return 0;
2078 }
2079 }
2080
2081 if (PageChecked(page) && ext4_should_journal_data(inode))
2082 /*
2083 * It's mmapped pagecache. Add buffers and journal it. There
2084 * doesn't seem much point in redirtying the page here.
2085 */
2086 return __ext4_journalled_writepage(page, len);
2087
2088 ext4_io_submit_init(&io_submit, wbc);
2089 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2090 if (!io_submit.io_end) {
2091 redirty_page_for_writepage(wbc, page);
2092 unlock_page(page);
2093 return -ENOMEM;
2094 }
2095 ret = ext4_bio_write_page(&io_submit, page, len);
2096 ext4_io_submit(&io_submit);
2097 /* Drop io_end reference we got from init */
2098 ext4_put_io_end_defer(io_submit.io_end);
2099 return ret;
2100}
2101
2102static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2103{
2104 int len;
2105 loff_t size;
2106 int err;
2107
2108 BUG_ON(page->index != mpd->first_page);
2109 clear_page_dirty_for_io(page);
2110 /*
2111 * We have to be very careful here! Nothing protects writeback path
2112 * against i_size changes and the page can be writeably mapped into
2113 * page tables. So an application can be growing i_size and writing
2114 * data through mmap while writeback runs. clear_page_dirty_for_io()
2115 * write-protects our page in page tables and the page cannot get
2116 * written to again until we release page lock. So only after
2117 * clear_page_dirty_for_io() we are safe to sample i_size for
2118 * ext4_bio_write_page() to zero-out tail of the written page. We rely
2119 * on the barrier provided by TestClearPageDirty in
2120 * clear_page_dirty_for_io() to make sure i_size is really sampled only
2121 * after page tables are updated.
2122 */
2123 size = i_size_read(mpd->inode);
2124 if (page->index == size >> PAGE_SHIFT &&
2125 !ext4_verity_in_progress(mpd->inode))
2126 len = size & ~PAGE_MASK;
2127 else
2128 len = PAGE_SIZE;
2129 err = ext4_bio_write_page(&mpd->io_submit, page, len);
2130 if (!err)
2131 mpd->wbc->nr_to_write--;
2132 mpd->first_page++;
2133
2134 return err;
2135}
2136
2137#define BH_FLAGS (BIT(BH_Unwritten) | BIT(BH_Delay))
2138
2139/*
2140 * mballoc gives us at most this number of blocks...
2141 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2142 * The rest of mballoc seems to handle chunks up to full group size.
2143 */
2144#define MAX_WRITEPAGES_EXTENT_LEN 2048
2145
2146/*
2147 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2148 *
2149 * @mpd - extent of blocks
2150 * @lblk - logical number of the block in the file
2151 * @bh - buffer head we want to add to the extent
2152 *
2153 * The function is used to collect contig. blocks in the same state. If the
2154 * buffer doesn't require mapping for writeback and we haven't started the
2155 * extent of buffers to map yet, the function returns 'true' immediately - the
2156 * caller can write the buffer right away. Otherwise the function returns true
2157 * if the block has been added to the extent, false if the block couldn't be
2158 * added.
2159 */
2160static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2161 struct buffer_head *bh)
2162{
2163 struct ext4_map_blocks *map = &mpd->map;
2164
2165 /* Buffer that doesn't need mapping for writeback? */
2166 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2167 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2168 /* So far no extent to map => we write the buffer right away */
2169 if (map->m_len == 0)
2170 return true;
2171 return false;
2172 }
2173
2174 /* First block in the extent? */
2175 if (map->m_len == 0) {
2176 /* We cannot map unless handle is started... */
2177 if (!mpd->do_map)
2178 return false;
2179 map->m_lblk = lblk;
2180 map->m_len = 1;
2181 map->m_flags = bh->b_state & BH_FLAGS;
2182 return true;
2183 }
2184
2185 /* Don't go larger than mballoc is willing to allocate */
2186 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2187 return false;
2188
2189 /* Can we merge the block to our big extent? */
2190 if (lblk == map->m_lblk + map->m_len &&
2191 (bh->b_state & BH_FLAGS) == map->m_flags) {
2192 map->m_len++;
2193 return true;
2194 }
2195 return false;
2196}
2197
2198/*
2199 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2200 *
2201 * @mpd - extent of blocks for mapping
2202 * @head - the first buffer in the page
2203 * @bh - buffer we should start processing from
2204 * @lblk - logical number of the block in the file corresponding to @bh
2205 *
2206 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2207 * the page for IO if all buffers in this page were mapped and there's no
2208 * accumulated extent of buffers to map or add buffers in the page to the
2209 * extent of buffers to map. The function returns 1 if the caller can continue
2210 * by processing the next page, 0 if it should stop adding buffers to the
2211 * extent to map because we cannot extend it anymore. It can also return value
2212 * < 0 in case of error during IO submission.
2213 */
2214static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2215 struct buffer_head *head,
2216 struct buffer_head *bh,
2217 ext4_lblk_t lblk)
2218{
2219 struct inode *inode = mpd->inode;
2220 int err;
2221 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2222 >> inode->i_blkbits;
2223
2224 if (ext4_verity_in_progress(inode))
2225 blocks = EXT_MAX_BLOCKS;
2226
2227 do {
2228 BUG_ON(buffer_locked(bh));
2229
2230 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2231 /* Found extent to map? */
2232 if (mpd->map.m_len)
2233 return 0;
2234 /* Buffer needs mapping and handle is not started? */
2235 if (!mpd->do_map)
2236 return 0;
2237 /* Everything mapped so far and we hit EOF */
2238 break;
2239 }
2240 } while (lblk++, (bh = bh->b_this_page) != head);
2241 /* So far everything mapped? Submit the page for IO. */
2242 if (mpd->map.m_len == 0) {
2243 err = mpage_submit_page(mpd, head->b_page);
2244 if (err < 0)
2245 return err;
2246 }
2247 if (lblk >= blocks) {
2248 mpd->scanned_until_end = 1;
2249 return 0;
2250 }
2251 return 1;
2252}
2253
2254/*
2255 * mpage_process_page - update page buffers corresponding to changed extent and
2256 * may submit fully mapped page for IO
2257 *
2258 * @mpd - description of extent to map, on return next extent to map
2259 * @m_lblk - logical block mapping.
2260 * @m_pblk - corresponding physical mapping.
2261 * @map_bh - determines on return whether this page requires any further
2262 * mapping or not.
2263 * Scan given page buffers corresponding to changed extent and update buffer
2264 * state according to new extent state.
2265 * We map delalloc buffers to their physical location, clear unwritten bits.
2266 * If the given page is not fully mapped, we update @map to the next extent in
2267 * the given page that needs mapping & return @map_bh as true.
2268 */
2269static int mpage_process_page(struct mpage_da_data *mpd, struct page *page,
2270 ext4_lblk_t *m_lblk, ext4_fsblk_t *m_pblk,
2271 bool *map_bh)
2272{
2273 struct buffer_head *head, *bh;
2274 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2275 ext4_lblk_t lblk = *m_lblk;
2276 ext4_fsblk_t pblock = *m_pblk;
2277 int err = 0;
2278 int blkbits = mpd->inode->i_blkbits;
2279 ssize_t io_end_size = 0;
2280 struct ext4_io_end_vec *io_end_vec = ext4_last_io_end_vec(io_end);
2281
2282 bh = head = page_buffers(page);
2283 do {
2284 if (lblk < mpd->map.m_lblk)
2285 continue;
2286 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2287 /*
2288 * Buffer after end of mapped extent.
2289 * Find next buffer in the page to map.
2290 */
2291 mpd->map.m_len = 0;
2292 mpd->map.m_flags = 0;
2293 io_end_vec->size += io_end_size;
2294
2295 err = mpage_process_page_bufs(mpd, head, bh, lblk);
2296 if (err > 0)
2297 err = 0;
2298 if (!err && mpd->map.m_len && mpd->map.m_lblk > lblk) {
2299 io_end_vec = ext4_alloc_io_end_vec(io_end);
2300 if (IS_ERR(io_end_vec)) {
2301 err = PTR_ERR(io_end_vec);
2302 goto out;
2303 }
2304 io_end_vec->offset = (loff_t)mpd->map.m_lblk << blkbits;
2305 }
2306 *map_bh = true;
2307 goto out;
2308 }
2309 if (buffer_delay(bh)) {
2310 clear_buffer_delay(bh);
2311 bh->b_blocknr = pblock++;
2312 }
2313 clear_buffer_unwritten(bh);
2314 io_end_size += (1 << blkbits);
2315 } while (lblk++, (bh = bh->b_this_page) != head);
2316
2317 io_end_vec->size += io_end_size;
2318 *map_bh = false;
2319out:
2320 *m_lblk = lblk;
2321 *m_pblk = pblock;
2322 return err;
2323}
2324
2325/*
2326 * mpage_map_buffers - update buffers corresponding to changed extent and
2327 * submit fully mapped pages for IO
2328 *
2329 * @mpd - description of extent to map, on return next extent to map
2330 *
2331 * Scan buffers corresponding to changed extent (we expect corresponding pages
2332 * to be already locked) and update buffer state according to new extent state.
2333 * We map delalloc buffers to their physical location, clear unwritten bits,
2334 * and mark buffers as uninit when we perform writes to unwritten extents
2335 * and do extent conversion after IO is finished. If the last page is not fully
2336 * mapped, we update @map to the next extent in the last page that needs
2337 * mapping. Otherwise we submit the page for IO.
2338 */
2339static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2340{
2341 struct folio_batch fbatch;
2342 unsigned nr, i;
2343 struct inode *inode = mpd->inode;
2344 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2345 pgoff_t start, end;
2346 ext4_lblk_t lblk;
2347 ext4_fsblk_t pblock;
2348 int err;
2349 bool map_bh = false;
2350
2351 start = mpd->map.m_lblk >> bpp_bits;
2352 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2353 lblk = start << bpp_bits;
2354 pblock = mpd->map.m_pblk;
2355
2356 folio_batch_init(&fbatch);
2357 while (start <= end) {
2358 nr = filemap_get_folios(inode->i_mapping, &start, end, &fbatch);
2359 if (nr == 0)
2360 break;
2361 for (i = 0; i < nr; i++) {
2362 struct page *page = &fbatch.folios[i]->page;
2363
2364 err = mpage_process_page(mpd, page, &lblk, &pblock,
2365 &map_bh);
2366 /*
2367 * If map_bh is true, means page may require further bh
2368 * mapping, or maybe the page was submitted for IO.
2369 * So we return to call further extent mapping.
2370 */
2371 if (err < 0 || map_bh)
2372 goto out;
2373 /* Page fully mapped - let IO run! */
2374 err = mpage_submit_page(mpd, page);
2375 if (err < 0)
2376 goto out;
2377 }
2378 folio_batch_release(&fbatch);
2379 }
2380 /* Extent fully mapped and matches with page boundary. We are done. */
2381 mpd->map.m_len = 0;
2382 mpd->map.m_flags = 0;
2383 return 0;
2384out:
2385 folio_batch_release(&fbatch);
2386 return err;
2387}
2388
2389static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2390{
2391 struct inode *inode = mpd->inode;
2392 struct ext4_map_blocks *map = &mpd->map;
2393 int get_blocks_flags;
2394 int err, dioread_nolock;
2395
2396 trace_ext4_da_write_pages_extent(inode, map);
2397 /*
2398 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2399 * to convert an unwritten extent to be initialized (in the case
2400 * where we have written into one or more preallocated blocks). It is
2401 * possible that we're going to need more metadata blocks than
2402 * previously reserved. However we must not fail because we're in
2403 * writeback and there is nothing we can do about it so it might result
2404 * in data loss. So use reserved blocks to allocate metadata if
2405 * possible.
2406 *
2407 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2408 * the blocks in question are delalloc blocks. This indicates
2409 * that the blocks and quotas has already been checked when
2410 * the data was copied into the page cache.
2411 */
2412 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2413 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2414 EXT4_GET_BLOCKS_IO_SUBMIT;
2415 dioread_nolock = ext4_should_dioread_nolock(inode);
2416 if (dioread_nolock)
2417 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2418 if (map->m_flags & BIT(BH_Delay))
2419 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2420
2421 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2422 if (err < 0)
2423 return err;
2424 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2425 if (!mpd->io_submit.io_end->handle &&
2426 ext4_handle_valid(handle)) {
2427 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2428 handle->h_rsv_handle = NULL;
2429 }
2430 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2431 }
2432
2433 BUG_ON(map->m_len == 0);
2434 return 0;
2435}
2436
2437/*
2438 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2439 * mpd->len and submit pages underlying it for IO
2440 *
2441 * @handle - handle for journal operations
2442 * @mpd - extent to map
2443 * @give_up_on_write - we set this to true iff there is a fatal error and there
2444 * is no hope of writing the data. The caller should discard
2445 * dirty pages to avoid infinite loops.
2446 *
2447 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2448 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2449 * them to initialized or split the described range from larger unwritten
2450 * extent. Note that we need not map all the described range since allocation
2451 * can return less blocks or the range is covered by more unwritten extents. We
2452 * cannot map more because we are limited by reserved transaction credits. On
2453 * the other hand we always make sure that the last touched page is fully
2454 * mapped so that it can be written out (and thus forward progress is
2455 * guaranteed). After mapping we submit all mapped pages for IO.
2456 */
2457static int mpage_map_and_submit_extent(handle_t *handle,
2458 struct mpage_da_data *mpd,
2459 bool *give_up_on_write)
2460{
2461 struct inode *inode = mpd->inode;
2462 struct ext4_map_blocks *map = &mpd->map;
2463 int err;
2464 loff_t disksize;
2465 int progress = 0;
2466 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2467 struct ext4_io_end_vec *io_end_vec;
2468
2469 io_end_vec = ext4_alloc_io_end_vec(io_end);
2470 if (IS_ERR(io_end_vec))
2471 return PTR_ERR(io_end_vec);
2472 io_end_vec->offset = ((loff_t)map->m_lblk) << inode->i_blkbits;
2473 do {
2474 err = mpage_map_one_extent(handle, mpd);
2475 if (err < 0) {
2476 struct super_block *sb = inode->i_sb;
2477
2478 if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2479 ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
2480 goto invalidate_dirty_pages;
2481 /*
2482 * Let the uper layers retry transient errors.
2483 * In the case of ENOSPC, if ext4_count_free_blocks()
2484 * is non-zero, a commit should free up blocks.
2485 */
2486 if ((err == -ENOMEM) ||
2487 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2488 if (progress)
2489 goto update_disksize;
2490 return err;
2491 }
2492 ext4_msg(sb, KERN_CRIT,
2493 "Delayed block allocation failed for "
2494 "inode %lu at logical offset %llu with"
2495 " max blocks %u with error %d",
2496 inode->i_ino,
2497 (unsigned long long)map->m_lblk,
2498 (unsigned)map->m_len, -err);
2499 ext4_msg(sb, KERN_CRIT,
2500 "This should not happen!! Data will "
2501 "be lost\n");
2502 if (err == -ENOSPC)
2503 ext4_print_free_blocks(inode);
2504 invalidate_dirty_pages:
2505 *give_up_on_write = true;
2506 return err;
2507 }
2508 progress = 1;
2509 /*
2510 * Update buffer state, submit mapped pages, and get us new
2511 * extent to map
2512 */
2513 err = mpage_map_and_submit_buffers(mpd);
2514 if (err < 0)
2515 goto update_disksize;
2516 } while (map->m_len);
2517
2518update_disksize:
2519 /*
2520 * Update on-disk size after IO is submitted. Races with
2521 * truncate are avoided by checking i_size under i_data_sem.
2522 */
2523 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2524 if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
2525 int err2;
2526 loff_t i_size;
2527
2528 down_write(&EXT4_I(inode)->i_data_sem);
2529 i_size = i_size_read(inode);
2530 if (disksize > i_size)
2531 disksize = i_size;
2532 if (disksize > EXT4_I(inode)->i_disksize)
2533 EXT4_I(inode)->i_disksize = disksize;
2534 up_write(&EXT4_I(inode)->i_data_sem);
2535 err2 = ext4_mark_inode_dirty(handle, inode);
2536 if (err2) {
2537 ext4_error_err(inode->i_sb, -err2,
2538 "Failed to mark inode %lu dirty",
2539 inode->i_ino);
2540 }
2541 if (!err)
2542 err = err2;
2543 }
2544 return err;
2545}
2546
2547/*
2548 * Calculate the total number of credits to reserve for one writepages
2549 * iteration. This is called from ext4_writepages(). We map an extent of
2550 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2551 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2552 * bpp - 1 blocks in bpp different extents.
2553 */
2554static int ext4_da_writepages_trans_blocks(struct inode *inode)
2555{
2556 int bpp = ext4_journal_blocks_per_page(inode);
2557
2558 return ext4_meta_trans_blocks(inode,
2559 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2560}
2561
2562/* Return true if the page needs to be written as part of transaction commit */
2563static bool ext4_page_nomap_can_writeout(struct page *page)
2564{
2565 struct buffer_head *bh, *head;
2566
2567 bh = head = page_buffers(page);
2568 do {
2569 if (buffer_dirty(bh) && buffer_mapped(bh) && !buffer_delay(bh))
2570 return true;
2571 } while ((bh = bh->b_this_page) != head);
2572 return false;
2573}
2574
2575/*
2576 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2577 * needing mapping, submit mapped pages
2578 *
2579 * @mpd - where to look for pages
2580 *
2581 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2582 * IO immediately. If we cannot map blocks, we submit just already mapped
2583 * buffers in the page for IO and keep page dirty. When we can map blocks and
2584 * we find a page which isn't mapped we start accumulating extent of buffers
2585 * underlying these pages that needs mapping (formed by either delayed or
2586 * unwritten buffers). We also lock the pages containing these buffers. The
2587 * extent found is returned in @mpd structure (starting at mpd->lblk with
2588 * length mpd->len blocks).
2589 *
2590 * Note that this function can attach bios to one io_end structure which are
2591 * neither logically nor physically contiguous. Although it may seem as an
2592 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2593 * case as we need to track IO to all buffers underlying a page in one io_end.
2594 */
2595static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2596{
2597 struct address_space *mapping = mpd->inode->i_mapping;
2598 struct pagevec pvec;
2599 unsigned int nr_pages;
2600 long left = mpd->wbc->nr_to_write;
2601 pgoff_t index = mpd->first_page;
2602 pgoff_t end = mpd->last_page;
2603 xa_mark_t tag;
2604 int i, err = 0;
2605 int blkbits = mpd->inode->i_blkbits;
2606 ext4_lblk_t lblk;
2607 struct buffer_head *head;
2608
2609 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2610 tag = PAGECACHE_TAG_TOWRITE;
2611 else
2612 tag = PAGECACHE_TAG_DIRTY;
2613
2614 pagevec_init(&pvec);
2615 mpd->map.m_len = 0;
2616 mpd->next_page = index;
2617 while (index <= end) {
2618 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2619 tag);
2620 if (nr_pages == 0)
2621 break;
2622
2623 for (i = 0; i < nr_pages; i++) {
2624 struct page *page = pvec.pages[i];
2625
2626 /*
2627 * Accumulated enough dirty pages? This doesn't apply
2628 * to WB_SYNC_ALL mode. For integrity sync we have to
2629 * keep going because someone may be concurrently
2630 * dirtying pages, and we might have synced a lot of
2631 * newly appeared dirty pages, but have not synced all
2632 * of the old dirty pages.
2633 */
2634 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2635 goto out;
2636
2637 /* If we can't merge this page, we are done. */
2638 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2639 goto out;
2640
2641 lock_page(page);
2642 /*
2643 * If the page is no longer dirty, or its mapping no
2644 * longer corresponds to inode we are writing (which
2645 * means it has been truncated or invalidated), or the
2646 * page is already under writeback and we are not doing
2647 * a data integrity writeback, skip the page
2648 */
2649 if (!PageDirty(page) ||
2650 (PageWriteback(page) &&
2651 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2652 unlikely(page->mapping != mapping)) {
2653 unlock_page(page);
2654 continue;
2655 }
2656
2657 wait_on_page_writeback(page);
2658 BUG_ON(PageWriteback(page));
2659
2660 /*
2661 * Should never happen but for buggy code in
2662 * other subsystems that call
2663 * set_page_dirty() without properly warning
2664 * the file system first. See [1] for more
2665 * information.
2666 *
2667 * [1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@quack2.suse.cz
2668 */
2669 if (!page_has_buffers(page)) {
2670 ext4_warning_inode(mpd->inode, "page %lu does not have buffers attached", page->index);
2671 ClearPageDirty(page);
2672 unlock_page(page);
2673 continue;
2674 }
2675
2676 if (mpd->map.m_len == 0)
2677 mpd->first_page = page->index;
2678 mpd->next_page = page->index + 1;
2679 /*
2680 * Writeout for transaction commit where we cannot
2681 * modify metadata is simple. Just submit the page.
2682 */
2683 if (!mpd->can_map) {
2684 if (ext4_page_nomap_can_writeout(page)) {
2685 err = mpage_submit_page(mpd, page);
2686 if (err < 0)
2687 goto out;
2688 } else {
2689 unlock_page(page);
2690 mpd->first_page++;
2691 }
2692 } else {
2693 /* Add all dirty buffers to mpd */
2694 lblk = ((ext4_lblk_t)page->index) <<
2695 (PAGE_SHIFT - blkbits);
2696 head = page_buffers(page);
2697 err = mpage_process_page_bufs(mpd, head, head,
2698 lblk);
2699 if (err <= 0)
2700 goto out;
2701 err = 0;
2702 }
2703 left--;
2704 }
2705 pagevec_release(&pvec);
2706 cond_resched();
2707 }
2708 mpd->scanned_until_end = 1;
2709 return 0;
2710out:
2711 pagevec_release(&pvec);
2712 return err;
2713}
2714
2715static int ext4_writepage_cb(struct page *page, struct writeback_control *wbc,
2716 void *data)
2717{
2718 return ext4_writepage(page, wbc);
2719}
2720
2721static int ext4_do_writepages(struct mpage_da_data *mpd)
2722{
2723 struct writeback_control *wbc = mpd->wbc;
2724 pgoff_t writeback_index = 0;
2725 long nr_to_write = wbc->nr_to_write;
2726 int range_whole = 0;
2727 int cycled = 1;
2728 handle_t *handle = NULL;
2729 struct inode *inode = mpd->inode;
2730 struct address_space *mapping = inode->i_mapping;
2731 int needed_blocks, rsv_blocks = 0, ret = 0;
2732 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2733 struct blk_plug plug;
2734 bool give_up_on_write = false;
2735
2736 trace_ext4_writepages(inode, wbc);
2737
2738 /*
2739 * No pages to write? This is mainly a kludge to avoid starting
2740 * a transaction for special inodes like journal inode on last iput()
2741 * because that could violate lock ordering on umount
2742 */
2743 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2744 goto out_writepages;
2745
2746 if (ext4_should_journal_data(inode)) {
2747 blk_start_plug(&plug);
2748 ret = write_cache_pages(mapping, wbc, ext4_writepage_cb, NULL);
2749 blk_finish_plug(&plug);
2750 goto out_writepages;
2751 }
2752
2753 /*
2754 * If the filesystem has aborted, it is read-only, so return
2755 * right away instead of dumping stack traces later on that
2756 * will obscure the real source of the problem. We test
2757 * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2758 * the latter could be true if the filesystem is mounted
2759 * read-only, and in that case, ext4_writepages should
2760 * *never* be called, so if that ever happens, we would want
2761 * the stack trace.
2762 */
2763 if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2764 ext4_test_mount_flag(inode->i_sb, EXT4_MF_FS_ABORTED))) {
2765 ret = -EROFS;
2766 goto out_writepages;
2767 }
2768
2769 /*
2770 * If we have inline data and arrive here, it means that
2771 * we will soon create the block for the 1st page, so
2772 * we'd better clear the inline data here.
2773 */
2774 if (ext4_has_inline_data(inode)) {
2775 /* Just inode will be modified... */
2776 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2777 if (IS_ERR(handle)) {
2778 ret = PTR_ERR(handle);
2779 goto out_writepages;
2780 }
2781 BUG_ON(ext4_test_inode_state(inode,
2782 EXT4_STATE_MAY_INLINE_DATA));
2783 ext4_destroy_inline_data(handle, inode);
2784 ext4_journal_stop(handle);
2785 }
2786
2787 if (ext4_should_dioread_nolock(inode)) {
2788 /*
2789 * We may need to convert up to one extent per block in
2790 * the page and we may dirty the inode.
2791 */
2792 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2793 PAGE_SIZE >> inode->i_blkbits);
2794 }
2795
2796 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2797 range_whole = 1;
2798
2799 if (wbc->range_cyclic) {
2800 writeback_index = mapping->writeback_index;
2801 if (writeback_index)
2802 cycled = 0;
2803 mpd->first_page = writeback_index;
2804 mpd->last_page = -1;
2805 } else {
2806 mpd->first_page = wbc->range_start >> PAGE_SHIFT;
2807 mpd->last_page = wbc->range_end >> PAGE_SHIFT;
2808 }
2809
2810 ext4_io_submit_init(&mpd->io_submit, wbc);
2811retry:
2812 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2813 tag_pages_for_writeback(mapping, mpd->first_page,
2814 mpd->last_page);
2815 blk_start_plug(&plug);
2816
2817 /*
2818 * First writeback pages that don't need mapping - we can avoid
2819 * starting a transaction unnecessarily and also avoid being blocked
2820 * in the block layer on device congestion while having transaction
2821 * started.
2822 */
2823 mpd->do_map = 0;
2824 mpd->scanned_until_end = 0;
2825 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2826 if (!mpd->io_submit.io_end) {
2827 ret = -ENOMEM;
2828 goto unplug;
2829 }
2830 ret = mpage_prepare_extent_to_map(mpd);
2831 /* Unlock pages we didn't use */
2832 mpage_release_unused_pages(mpd, false);
2833 /* Submit prepared bio */
2834 ext4_io_submit(&mpd->io_submit);
2835 ext4_put_io_end_defer(mpd->io_submit.io_end);
2836 mpd->io_submit.io_end = NULL;
2837 if (ret < 0)
2838 goto unplug;
2839
2840 while (!mpd->scanned_until_end && wbc->nr_to_write > 0) {
2841 /* For each extent of pages we use new io_end */
2842 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2843 if (!mpd->io_submit.io_end) {
2844 ret = -ENOMEM;
2845 break;
2846 }
2847
2848 WARN_ON_ONCE(!mpd->can_map);
2849 /*
2850 * We have two constraints: We find one extent to map and we
2851 * must always write out whole page (makes a difference when
2852 * blocksize < pagesize) so that we don't block on IO when we
2853 * try to write out the rest of the page. Journalled mode is
2854 * not supported by delalloc.
2855 */
2856 BUG_ON(ext4_should_journal_data(inode));
2857 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2858
2859 /* start a new transaction */
2860 handle = ext4_journal_start_with_reserve(inode,
2861 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2862 if (IS_ERR(handle)) {
2863 ret = PTR_ERR(handle);
2864 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2865 "%ld pages, ino %lu; err %d", __func__,
2866 wbc->nr_to_write, inode->i_ino, ret);
2867 /* Release allocated io_end */
2868 ext4_put_io_end(mpd->io_submit.io_end);
2869 mpd->io_submit.io_end = NULL;
2870 break;
2871 }
2872 mpd->do_map = 1;
2873
2874 trace_ext4_da_write_pages(inode, mpd->first_page, wbc);
2875 ret = mpage_prepare_extent_to_map(mpd);
2876 if (!ret && mpd->map.m_len)
2877 ret = mpage_map_and_submit_extent(handle, mpd,
2878 &give_up_on_write);
2879 /*
2880 * Caution: If the handle is synchronous,
2881 * ext4_journal_stop() can wait for transaction commit
2882 * to finish which may depend on writeback of pages to
2883 * complete or on page lock to be released. In that
2884 * case, we have to wait until after we have
2885 * submitted all the IO, released page locks we hold,
2886 * and dropped io_end reference (for extent conversion
2887 * to be able to complete) before stopping the handle.
2888 */
2889 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2890 ext4_journal_stop(handle);
2891 handle = NULL;
2892 mpd->do_map = 0;
2893 }
2894 /* Unlock pages we didn't use */
2895 mpage_release_unused_pages(mpd, give_up_on_write);
2896 /* Submit prepared bio */
2897 ext4_io_submit(&mpd->io_submit);
2898
2899 /*
2900 * Drop our io_end reference we got from init. We have
2901 * to be careful and use deferred io_end finishing if
2902 * we are still holding the transaction as we can
2903 * release the last reference to io_end which may end
2904 * up doing unwritten extent conversion.
2905 */
2906 if (handle) {
2907 ext4_put_io_end_defer(mpd->io_submit.io_end);
2908 ext4_journal_stop(handle);
2909 } else
2910 ext4_put_io_end(mpd->io_submit.io_end);
2911 mpd->io_submit.io_end = NULL;
2912
2913 if (ret == -ENOSPC && sbi->s_journal) {
2914 /*
2915 * Commit the transaction which would
2916 * free blocks released in the transaction
2917 * and try again
2918 */
2919 jbd2_journal_force_commit_nested(sbi->s_journal);
2920 ret = 0;
2921 continue;
2922 }
2923 /* Fatal error - ENOMEM, EIO... */
2924 if (ret)
2925 break;
2926 }
2927unplug:
2928 blk_finish_plug(&plug);
2929 if (!ret && !cycled && wbc->nr_to_write > 0) {
2930 cycled = 1;
2931 mpd->last_page = writeback_index - 1;
2932 mpd->first_page = 0;
2933 goto retry;
2934 }
2935
2936 /* Update index */
2937 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2938 /*
2939 * Set the writeback_index so that range_cyclic
2940 * mode will write it back later
2941 */
2942 mapping->writeback_index = mpd->first_page;
2943
2944out_writepages:
2945 trace_ext4_writepages_result(inode, wbc, ret,
2946 nr_to_write - wbc->nr_to_write);
2947 return ret;
2948}
2949
2950static int ext4_writepages(struct address_space *mapping,
2951 struct writeback_control *wbc)
2952{
2953 struct super_block *sb = mapping->host->i_sb;
2954 struct mpage_da_data mpd = {
2955 .inode = mapping->host,
2956 .wbc = wbc,
2957 .can_map = 1,
2958 };
2959 int ret;
2960
2961 if (unlikely(ext4_forced_shutdown(EXT4_SB(sb))))
2962 return -EIO;
2963
2964 percpu_down_read(&EXT4_SB(sb)->s_writepages_rwsem);
2965 ret = ext4_do_writepages(&mpd);
2966 percpu_up_read(&EXT4_SB(sb)->s_writepages_rwsem);
2967
2968 return ret;
2969}
2970
2971int ext4_normal_submit_inode_data_buffers(struct jbd2_inode *jinode)
2972{
2973 struct writeback_control wbc = {
2974 .sync_mode = WB_SYNC_ALL,
2975 .nr_to_write = LONG_MAX,
2976 .range_start = jinode->i_dirty_start,
2977 .range_end = jinode->i_dirty_end,
2978 };
2979 struct mpage_da_data mpd = {
2980 .inode = jinode->i_vfs_inode,
2981 .wbc = &wbc,
2982 .can_map = 0,
2983 };
2984 return ext4_do_writepages(&mpd);
2985}
2986
2987static int ext4_dax_writepages(struct address_space *mapping,
2988 struct writeback_control *wbc)
2989{
2990 int ret;
2991 long nr_to_write = wbc->nr_to_write;
2992 struct inode *inode = mapping->host;
2993 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2994
2995 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2996 return -EIO;
2997
2998 percpu_down_read(&sbi->s_writepages_rwsem);
2999 trace_ext4_writepages(inode, wbc);
3000
3001 ret = dax_writeback_mapping_range(mapping, sbi->s_daxdev, wbc);
3002 trace_ext4_writepages_result(inode, wbc, ret,
3003 nr_to_write - wbc->nr_to_write);
3004 percpu_up_read(&sbi->s_writepages_rwsem);
3005 return ret;
3006}
3007
3008static int ext4_nonda_switch(struct super_block *sb)
3009{
3010 s64 free_clusters, dirty_clusters;
3011 struct ext4_sb_info *sbi = EXT4_SB(sb);
3012
3013 /*
3014 * switch to non delalloc mode if we are running low
3015 * on free block. The free block accounting via percpu
3016 * counters can get slightly wrong with percpu_counter_batch getting
3017 * accumulated on each CPU without updating global counters
3018 * Delalloc need an accurate free block accounting. So switch
3019 * to non delalloc when we are near to error range.
3020 */
3021 free_clusters =
3022 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
3023 dirty_clusters =
3024 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
3025 /*
3026 * Start pushing delalloc when 1/2 of free blocks are dirty.
3027 */
3028 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
3029 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
3030
3031 if (2 * free_clusters < 3 * dirty_clusters ||
3032 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
3033 /*
3034 * free block count is less than 150% of dirty blocks
3035 * or free blocks is less than watermark
3036 */
3037 return 1;
3038 }
3039 return 0;
3040}
3041
3042static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3043 loff_t pos, unsigned len,
3044 struct page **pagep, void **fsdata)
3045{
3046 int ret, retries = 0;
3047 struct page *page;
3048 pgoff_t index;
3049 struct inode *inode = mapping->host;
3050
3051 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3052 return -EIO;
3053
3054 index = pos >> PAGE_SHIFT;
3055
3056 if (ext4_nonda_switch(inode->i_sb) || ext4_verity_in_progress(inode)) {
3057 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3058 return ext4_write_begin(file, mapping, pos,
3059 len, pagep, fsdata);
3060 }
3061 *fsdata = (void *)0;
3062 trace_ext4_da_write_begin(inode, pos, len);
3063
3064 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3065 ret = ext4_da_write_inline_data_begin(mapping, inode, pos, len,
3066 pagep, fsdata);
3067 if (ret < 0)
3068 return ret;
3069 if (ret == 1)
3070 return 0;
3071 }
3072
3073retry:
3074 page = grab_cache_page_write_begin(mapping, index);
3075 if (!page)
3076 return -ENOMEM;
3077
3078 /* In case writeback began while the page was unlocked */
3079 wait_for_stable_page(page);
3080
3081#ifdef CONFIG_FS_ENCRYPTION
3082 ret = ext4_block_write_begin(page, pos, len,
3083 ext4_da_get_block_prep);
3084#else
3085 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3086#endif
3087 if (ret < 0) {
3088 unlock_page(page);
3089 put_page(page);
3090 /*
3091 * block_write_begin may have instantiated a few blocks
3092 * outside i_size. Trim these off again. Don't need
3093 * i_size_read because we hold inode lock.
3094 */
3095 if (pos + len > inode->i_size)
3096 ext4_truncate_failed_write(inode);
3097
3098 if (ret == -ENOSPC &&
3099 ext4_should_retry_alloc(inode->i_sb, &retries))
3100 goto retry;
3101 return ret;
3102 }
3103
3104 *pagep = page;
3105 return ret;
3106}
3107
3108/*
3109 * Check if we should update i_disksize
3110 * when write to the end of file but not require block allocation
3111 */
3112static int ext4_da_should_update_i_disksize(struct page *page,
3113 unsigned long offset)
3114{
3115 struct buffer_head *bh;
3116 struct inode *inode = page->mapping->host;
3117 unsigned int idx;
3118 int i;
3119
3120 bh = page_buffers(page);
3121 idx = offset >> inode->i_blkbits;
3122
3123 for (i = 0; i < idx; i++)
3124 bh = bh->b_this_page;
3125
3126 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3127 return 0;
3128 return 1;
3129}
3130
3131static int ext4_da_write_end(struct file *file,
3132 struct address_space *mapping,
3133 loff_t pos, unsigned len, unsigned copied,
3134 struct page *page, void *fsdata)
3135{
3136 struct inode *inode = mapping->host;
3137 loff_t new_i_size;
3138 unsigned long start, end;
3139 int write_mode = (int)(unsigned long)fsdata;
3140
3141 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3142 return ext4_write_end(file, mapping, pos,
3143 len, copied, page, fsdata);
3144
3145 trace_ext4_da_write_end(inode, pos, len, copied);
3146
3147 if (write_mode != CONVERT_INLINE_DATA &&
3148 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3149 ext4_has_inline_data(inode))
3150 return ext4_write_inline_data_end(inode, pos, len, copied, page);
3151
3152 start = pos & (PAGE_SIZE - 1);
3153 end = start + copied - 1;
3154
3155 /*
3156 * Since we are holding inode lock, we are sure i_disksize <=
3157 * i_size. We also know that if i_disksize < i_size, there are
3158 * delalloc writes pending in the range upto i_size. If the end of
3159 * the current write is <= i_size, there's no need to touch
3160 * i_disksize since writeback will push i_disksize upto i_size
3161 * eventually. If the end of the current write is > i_size and
3162 * inside an allocated block (ext4_da_should_update_i_disksize()
3163 * check), we need to update i_disksize here as neither
3164 * ext4_writepage() nor certain ext4_writepages() paths not
3165 * allocating blocks update i_disksize.
3166 *
3167 * Note that we defer inode dirtying to generic_write_end() /
3168 * ext4_da_write_inline_data_end().
3169 */
3170 new_i_size = pos + copied;
3171 if (copied && new_i_size > inode->i_size &&
3172 ext4_da_should_update_i_disksize(page, end))
3173 ext4_update_i_disksize(inode, new_i_size);
3174
3175 return generic_write_end(file, mapping, pos, len, copied, page, fsdata);
3176}
3177
3178/*
3179 * Force all delayed allocation blocks to be allocated for a given inode.
3180 */
3181int ext4_alloc_da_blocks(struct inode *inode)
3182{
3183 trace_ext4_alloc_da_blocks(inode);
3184
3185 if (!EXT4_I(inode)->i_reserved_data_blocks)
3186 return 0;
3187
3188 /*
3189 * We do something simple for now. The filemap_flush() will
3190 * also start triggering a write of the data blocks, which is
3191 * not strictly speaking necessary (and for users of
3192 * laptop_mode, not even desirable). However, to do otherwise
3193 * would require replicating code paths in:
3194 *
3195 * ext4_writepages() ->
3196 * write_cache_pages() ---> (via passed in callback function)
3197 * __mpage_da_writepage() -->
3198 * mpage_add_bh_to_extent()
3199 * mpage_da_map_blocks()
3200 *
3201 * The problem is that write_cache_pages(), located in
3202 * mm/page-writeback.c, marks pages clean in preparation for
3203 * doing I/O, which is not desirable if we're not planning on
3204 * doing I/O at all.
3205 *
3206 * We could call write_cache_pages(), and then redirty all of
3207 * the pages by calling redirty_page_for_writepage() but that
3208 * would be ugly in the extreme. So instead we would need to
3209 * replicate parts of the code in the above functions,
3210 * simplifying them because we wouldn't actually intend to
3211 * write out the pages, but rather only collect contiguous
3212 * logical block extents, call the multi-block allocator, and
3213 * then update the buffer heads with the block allocations.
3214 *
3215 * For now, though, we'll cheat by calling filemap_flush(),
3216 * which will map the blocks, and start the I/O, but not
3217 * actually wait for the I/O to complete.
3218 */
3219 return filemap_flush(inode->i_mapping);
3220}
3221
3222/*
3223 * bmap() is special. It gets used by applications such as lilo and by
3224 * the swapper to find the on-disk block of a specific piece of data.
3225 *
3226 * Naturally, this is dangerous if the block concerned is still in the
3227 * journal. If somebody makes a swapfile on an ext4 data-journaling
3228 * filesystem and enables swap, then they may get a nasty shock when the
3229 * data getting swapped to that swapfile suddenly gets overwritten by
3230 * the original zero's written out previously to the journal and
3231 * awaiting writeback in the kernel's buffer cache.
3232 *
3233 * So, if we see any bmap calls here on a modified, data-journaled file,
3234 * take extra steps to flush any blocks which might be in the cache.
3235 */
3236static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3237{
3238 struct inode *inode = mapping->host;
3239 journal_t *journal;
3240 sector_t ret = 0;
3241 int err;
3242
3243 inode_lock_shared(inode);
3244 /*
3245 * We can get here for an inline file via the FIBMAP ioctl
3246 */
3247 if (ext4_has_inline_data(inode))
3248 goto out;
3249
3250 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3251 test_opt(inode->i_sb, DELALLOC)) {
3252 /*
3253 * With delalloc we want to sync the file
3254 * so that we can make sure we allocate
3255 * blocks for file
3256 */
3257 filemap_write_and_wait(mapping);
3258 }
3259
3260 if (EXT4_JOURNAL(inode) &&
3261 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3262 /*
3263 * This is a REALLY heavyweight approach, but the use of
3264 * bmap on dirty files is expected to be extremely rare:
3265 * only if we run lilo or swapon on a freshly made file
3266 * do we expect this to happen.
3267 *
3268 * (bmap requires CAP_SYS_RAWIO so this does not
3269 * represent an unprivileged user DOS attack --- we'd be
3270 * in trouble if mortal users could trigger this path at
3271 * will.)
3272 *
3273 * NB. EXT4_STATE_JDATA is not set on files other than
3274 * regular files. If somebody wants to bmap a directory
3275 * or symlink and gets confused because the buffer
3276 * hasn't yet been flushed to disk, they deserve
3277 * everything they get.
3278 */
3279
3280 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3281 journal = EXT4_JOURNAL(inode);
3282 jbd2_journal_lock_updates(journal);
3283 err = jbd2_journal_flush(journal, 0);
3284 jbd2_journal_unlock_updates(journal);
3285
3286 if (err)
3287 goto out;
3288 }
3289
3290 ret = iomap_bmap(mapping, block, &ext4_iomap_ops);
3291
3292out:
3293 inode_unlock_shared(inode);
3294 return ret;
3295}
3296
3297static int ext4_read_folio(struct file *file, struct folio *folio)
3298{
3299 struct page *page = &folio->page;
3300 int ret = -EAGAIN;
3301 struct inode *inode = page->mapping->host;
3302
3303 trace_ext4_readpage(page);
3304
3305 if (ext4_has_inline_data(inode))
3306 ret = ext4_readpage_inline(inode, page);
3307
3308 if (ret == -EAGAIN)
3309 return ext4_mpage_readpages(inode, NULL, page);
3310
3311 return ret;
3312}
3313
3314static void ext4_readahead(struct readahead_control *rac)
3315{
3316 struct inode *inode = rac->mapping->host;
3317
3318 /* If the file has inline data, no need to do readahead. */
3319 if (ext4_has_inline_data(inode))
3320 return;
3321
3322 ext4_mpage_readpages(inode, rac, NULL);
3323}
3324
3325static void ext4_invalidate_folio(struct folio *folio, size_t offset,
3326 size_t length)
3327{
3328 trace_ext4_invalidate_folio(folio, offset, length);
3329
3330 /* No journalling happens on data buffers when this function is used */
3331 WARN_ON(folio_buffers(folio) && buffer_jbd(folio_buffers(folio)));
3332
3333 block_invalidate_folio(folio, offset, length);
3334}
3335
3336static int __ext4_journalled_invalidate_folio(struct folio *folio,
3337 size_t offset, size_t length)
3338{
3339 journal_t *journal = EXT4_JOURNAL(folio->mapping->host);
3340
3341 trace_ext4_journalled_invalidate_folio(folio, offset, length);
3342
3343 /*
3344 * If it's a full truncate we just forget about the pending dirtying
3345 */
3346 if (offset == 0 && length == folio_size(folio))
3347 folio_clear_checked(folio);
3348
3349 return jbd2_journal_invalidate_folio(journal, folio, offset, length);
3350}
3351
3352/* Wrapper for aops... */
3353static void ext4_journalled_invalidate_folio(struct folio *folio,
3354 size_t offset,
3355 size_t length)
3356{
3357 WARN_ON(__ext4_journalled_invalidate_folio(folio, offset, length) < 0);
3358}
3359
3360static bool ext4_release_folio(struct folio *folio, gfp_t wait)
3361{
3362 journal_t *journal = EXT4_JOURNAL(folio->mapping->host);
3363
3364 trace_ext4_releasepage(&folio->page);
3365
3366 /* Page has dirty journalled data -> cannot release */
3367 if (folio_test_checked(folio))
3368 return false;
3369 if (journal)
3370 return jbd2_journal_try_to_free_buffers(journal, folio);
3371 else
3372 return try_to_free_buffers(folio);
3373}
3374
3375static bool ext4_inode_datasync_dirty(struct inode *inode)
3376{
3377 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3378
3379 if (journal) {
3380 if (jbd2_transaction_committed(journal,
3381 EXT4_I(inode)->i_datasync_tid))
3382 return false;
3383 if (test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT))
3384 return !list_empty(&EXT4_I(inode)->i_fc_list);
3385 return true;
3386 }
3387
3388 /* Any metadata buffers to write? */
3389 if (!list_empty(&inode->i_mapping->private_list))
3390 return true;
3391 return inode->i_state & I_DIRTY_DATASYNC;
3392}
3393
3394static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
3395 struct ext4_map_blocks *map, loff_t offset,
3396 loff_t length, unsigned int flags)
3397{
3398 u8 blkbits = inode->i_blkbits;
3399
3400 /*
3401 * Writes that span EOF might trigger an I/O size update on completion,
3402 * so consider them to be dirty for the purpose of O_DSYNC, even if
3403 * there is no other metadata changes being made or are pending.
3404 */
3405 iomap->flags = 0;
3406 if (ext4_inode_datasync_dirty(inode) ||
3407 offset + length > i_size_read(inode))
3408 iomap->flags |= IOMAP_F_DIRTY;
3409
3410 if (map->m_flags & EXT4_MAP_NEW)
3411 iomap->flags |= IOMAP_F_NEW;
3412
3413 if (flags & IOMAP_DAX)
3414 iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
3415 else
3416 iomap->bdev = inode->i_sb->s_bdev;
3417 iomap->offset = (u64) map->m_lblk << blkbits;
3418 iomap->length = (u64) map->m_len << blkbits;
3419
3420 if ((map->m_flags & EXT4_MAP_MAPPED) &&
3421 !ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3422 iomap->flags |= IOMAP_F_MERGED;
3423
3424 /*
3425 * Flags passed to ext4_map_blocks() for direct I/O writes can result
3426 * in m_flags having both EXT4_MAP_MAPPED and EXT4_MAP_UNWRITTEN bits
3427 * set. In order for any allocated unwritten extents to be converted
3428 * into written extents correctly within the ->end_io() handler, we
3429 * need to ensure that the iomap->type is set appropriately. Hence, the
3430 * reason why we need to check whether the EXT4_MAP_UNWRITTEN bit has
3431 * been set first.
3432 */
3433 if (map->m_flags & EXT4_MAP_UNWRITTEN) {
3434 iomap->type = IOMAP_UNWRITTEN;
3435 iomap->addr = (u64) map->m_pblk << blkbits;
3436 if (flags & IOMAP_DAX)
3437 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3438 } else if (map->m_flags & EXT4_MAP_MAPPED) {
3439 iomap->type = IOMAP_MAPPED;
3440 iomap->addr = (u64) map->m_pblk << blkbits;
3441 if (flags & IOMAP_DAX)
3442 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3443 } else {
3444 iomap->type = IOMAP_HOLE;
3445 iomap->addr = IOMAP_NULL_ADDR;
3446 }
3447}
3448
3449static int ext4_iomap_alloc(struct inode *inode, struct ext4_map_blocks *map,
3450 unsigned int flags)
3451{
3452 handle_t *handle;
3453 u8 blkbits = inode->i_blkbits;
3454 int ret, dio_credits, m_flags = 0, retries = 0;
3455
3456 /*
3457 * Trim the mapping request to the maximum value that we can map at
3458 * once for direct I/O.
3459 */
3460 if (map->m_len > DIO_MAX_BLOCKS)
3461 map->m_len = DIO_MAX_BLOCKS;
3462 dio_credits = ext4_chunk_trans_blocks(inode, map->m_len);
3463
3464retry:
3465 /*
3466 * Either we allocate blocks and then don't get an unwritten extent, so
3467 * in that case we have reserved enough credits. Or, the blocks are
3468 * already allocated and unwritten. In that case, the extent conversion
3469 * fits into the credits as well.
3470 */
3471 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
3472 if (IS_ERR(handle))
3473 return PTR_ERR(handle);
3474
3475 /*
3476 * DAX and direct I/O are the only two operations that are currently
3477 * supported with IOMAP_WRITE.
3478 */
3479 WARN_ON(!(flags & (IOMAP_DAX | IOMAP_DIRECT)));
3480 if (flags & IOMAP_DAX)
3481 m_flags = EXT4_GET_BLOCKS_CREATE_ZERO;
3482 /*
3483 * We use i_size instead of i_disksize here because delalloc writeback
3484 * can complete at any point during the I/O and subsequently push the
3485 * i_disksize out to i_size. This could be beyond where direct I/O is
3486 * happening and thus expose allocated blocks to direct I/O reads.
3487 */
3488 else if (((loff_t)map->m_lblk << blkbits) >= i_size_read(inode))
3489 m_flags = EXT4_GET_BLOCKS_CREATE;
3490 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3491 m_flags = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3492
3493 ret = ext4_map_blocks(handle, inode, map, m_flags);
3494
3495 /*
3496 * We cannot fill holes in indirect tree based inodes as that could
3497 * expose stale data in the case of a crash. Use the magic error code
3498 * to fallback to buffered I/O.
3499 */
3500 if (!m_flags && !ret)
3501 ret = -ENOTBLK;
3502
3503 ext4_journal_stop(handle);
3504 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3505 goto retry;
3506
3507 return ret;
3508}
3509
3510
3511static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3512 unsigned flags, struct iomap *iomap, struct iomap *srcmap)
3513{
3514 int ret;
3515 struct ext4_map_blocks map;
3516 u8 blkbits = inode->i_blkbits;
3517
3518 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3519 return -EINVAL;
3520
3521 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3522 return -ERANGE;
3523
3524 /*
3525 * Calculate the first and last logical blocks respectively.
3526 */
3527 map.m_lblk = offset >> blkbits;
3528 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3529 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3530
3531 if (flags & IOMAP_WRITE) {
3532 /*
3533 * We check here if the blocks are already allocated, then we
3534 * don't need to start a journal txn and we can directly return
3535 * the mapping information. This could boost performance
3536 * especially in multi-threaded overwrite requests.
3537 */
3538 if (offset + length <= i_size_read(inode)) {
3539 ret = ext4_map_blocks(NULL, inode, &map, 0);
3540 if (ret > 0 && (map.m_flags & EXT4_MAP_MAPPED))
3541 goto out;
3542 }
3543 ret = ext4_iomap_alloc(inode, &map, flags);
3544 } else {
3545 ret = ext4_map_blocks(NULL, inode, &map, 0);
3546 }
3547
3548 if (ret < 0)
3549 return ret;
3550out:
3551 /*
3552 * When inline encryption is enabled, sometimes I/O to an encrypted file
3553 * has to be broken up to guarantee DUN contiguity. Handle this by
3554 * limiting the length of the mapping returned.
3555 */
3556 map.m_len = fscrypt_limit_io_blocks(inode, map.m_lblk, map.m_len);
3557
3558 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3559
3560 return 0;
3561}
3562
3563static int ext4_iomap_overwrite_begin(struct inode *inode, loff_t offset,
3564 loff_t length, unsigned flags, struct iomap *iomap,
3565 struct iomap *srcmap)
3566{
3567 int ret;
3568
3569 /*
3570 * Even for writes we don't need to allocate blocks, so just pretend
3571 * we are reading to save overhead of starting a transaction.
3572 */
3573 flags &= ~IOMAP_WRITE;
3574 ret = ext4_iomap_begin(inode, offset, length, flags, iomap, srcmap);
3575 WARN_ON_ONCE(iomap->type != IOMAP_MAPPED);
3576 return ret;
3577}
3578
3579static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3580 ssize_t written, unsigned flags, struct iomap *iomap)
3581{
3582 /*
3583 * Check to see whether an error occurred while writing out the data to
3584 * the allocated blocks. If so, return the magic error code so that we
3585 * fallback to buffered I/O and attempt to complete the remainder of
3586 * the I/O. Any blocks that may have been allocated in preparation for
3587 * the direct I/O will be reused during buffered I/O.
3588 */
3589 if (flags & (IOMAP_WRITE | IOMAP_DIRECT) && written == 0)
3590 return -ENOTBLK;
3591
3592 return 0;
3593}
3594
3595const struct iomap_ops ext4_iomap_ops = {
3596 .iomap_begin = ext4_iomap_begin,
3597 .iomap_end = ext4_iomap_end,
3598};
3599
3600const struct iomap_ops ext4_iomap_overwrite_ops = {
3601 .iomap_begin = ext4_iomap_overwrite_begin,
3602 .iomap_end = ext4_iomap_end,
3603};
3604
3605static bool ext4_iomap_is_delalloc(struct inode *inode,
3606 struct ext4_map_blocks *map)
3607{
3608 struct extent_status es;
3609 ext4_lblk_t offset = 0, end = map->m_lblk + map->m_len - 1;
3610
3611 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
3612 map->m_lblk, end, &es);
3613
3614 if (!es.es_len || es.es_lblk > end)
3615 return false;
3616
3617 if (es.es_lblk > map->m_lblk) {
3618 map->m_len = es.es_lblk - map->m_lblk;
3619 return false;
3620 }
3621
3622 offset = map->m_lblk - es.es_lblk;
3623 map->m_len = es.es_len - offset;
3624
3625 return true;
3626}
3627
3628static int ext4_iomap_begin_report(struct inode *inode, loff_t offset,
3629 loff_t length, unsigned int flags,
3630 struct iomap *iomap, struct iomap *srcmap)
3631{
3632 int ret;
3633 bool delalloc = false;
3634 struct ext4_map_blocks map;
3635 u8 blkbits = inode->i_blkbits;
3636
3637 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3638 return -EINVAL;
3639
3640 if (ext4_has_inline_data(inode)) {
3641 ret = ext4_inline_data_iomap(inode, iomap);
3642 if (ret != -EAGAIN) {
3643 if (ret == 0 && offset >= iomap->length)
3644 ret = -ENOENT;
3645 return ret;
3646 }
3647 }
3648
3649 /*
3650 * Calculate the first and last logical block respectively.
3651 */
3652 map.m_lblk = offset >> blkbits;
3653 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3654 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3655
3656 /*
3657 * Fiemap callers may call for offset beyond s_bitmap_maxbytes.
3658 * So handle it here itself instead of querying ext4_map_blocks().
3659 * Since ext4_map_blocks() will warn about it and will return
3660 * -EIO error.
3661 */
3662 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3663 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3664
3665 if (offset >= sbi->s_bitmap_maxbytes) {
3666 map.m_flags = 0;
3667 goto set_iomap;
3668 }
3669 }
3670
3671 ret = ext4_map_blocks(NULL, inode, &map, 0);
3672 if (ret < 0)
3673 return ret;
3674 if (ret == 0)
3675 delalloc = ext4_iomap_is_delalloc(inode, &map);
3676
3677set_iomap:
3678 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3679 if (delalloc && iomap->type == IOMAP_HOLE)
3680 iomap->type = IOMAP_DELALLOC;
3681
3682 return 0;
3683}
3684
3685const struct iomap_ops ext4_iomap_report_ops = {
3686 .iomap_begin = ext4_iomap_begin_report,
3687};
3688
3689/*
3690 * Whenever the folio is being dirtied, corresponding buffers should already
3691 * be attached to the transaction (we take care of this in ext4_page_mkwrite()
3692 * and ext4_write_begin()). However we cannot move buffers to dirty transaction
3693 * lists here because ->dirty_folio is called under VFS locks and the folio
3694 * is not necessarily locked.
3695 *
3696 * We cannot just dirty the folio and leave attached buffers clean, because the
3697 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3698 * or jbddirty because all the journalling code will explode.
3699 *
3700 * So what we do is to mark the folio "pending dirty" and next time writepage
3701 * is called, propagate that into the buffers appropriately.
3702 */
3703static bool ext4_journalled_dirty_folio(struct address_space *mapping,
3704 struct folio *folio)
3705{
3706 WARN_ON_ONCE(!folio_buffers(folio));
3707 folio_set_checked(folio);
3708 return filemap_dirty_folio(mapping, folio);
3709}
3710
3711static bool ext4_dirty_folio(struct address_space *mapping, struct folio *folio)
3712{
3713 WARN_ON_ONCE(!folio_test_locked(folio) && !folio_test_dirty(folio));
3714 WARN_ON_ONCE(!folio_buffers(folio));
3715 return block_dirty_folio(mapping, folio);
3716}
3717
3718static int ext4_iomap_swap_activate(struct swap_info_struct *sis,
3719 struct file *file, sector_t *span)
3720{
3721 return iomap_swapfile_activate(sis, file, span,
3722 &ext4_iomap_report_ops);
3723}
3724
3725static const struct address_space_operations ext4_aops = {
3726 .read_folio = ext4_read_folio,
3727 .readahead = ext4_readahead,
3728 .writepages = ext4_writepages,
3729 .write_begin = ext4_write_begin,
3730 .write_end = ext4_write_end,
3731 .dirty_folio = ext4_dirty_folio,
3732 .bmap = ext4_bmap,
3733 .invalidate_folio = ext4_invalidate_folio,
3734 .release_folio = ext4_release_folio,
3735 .direct_IO = noop_direct_IO,
3736 .migrate_folio = buffer_migrate_folio,
3737 .is_partially_uptodate = block_is_partially_uptodate,
3738 .error_remove_page = generic_error_remove_page,
3739 .swap_activate = ext4_iomap_swap_activate,
3740};
3741
3742static const struct address_space_operations ext4_journalled_aops = {
3743 .read_folio = ext4_read_folio,
3744 .readahead = ext4_readahead,
3745 .writepages = ext4_writepages,
3746 .write_begin = ext4_write_begin,
3747 .write_end = ext4_journalled_write_end,
3748 .dirty_folio = ext4_journalled_dirty_folio,
3749 .bmap = ext4_bmap,
3750 .invalidate_folio = ext4_journalled_invalidate_folio,
3751 .release_folio = ext4_release_folio,
3752 .direct_IO = noop_direct_IO,
3753 .migrate_folio = buffer_migrate_folio_norefs,
3754 .is_partially_uptodate = block_is_partially_uptodate,
3755 .error_remove_page = generic_error_remove_page,
3756 .swap_activate = ext4_iomap_swap_activate,
3757};
3758
3759static const struct address_space_operations ext4_da_aops = {
3760 .read_folio = ext4_read_folio,
3761 .readahead = ext4_readahead,
3762 .writepages = ext4_writepages,
3763 .write_begin = ext4_da_write_begin,
3764 .write_end = ext4_da_write_end,
3765 .dirty_folio = ext4_dirty_folio,
3766 .bmap = ext4_bmap,
3767 .invalidate_folio = ext4_invalidate_folio,
3768 .release_folio = ext4_release_folio,
3769 .direct_IO = noop_direct_IO,
3770 .migrate_folio = buffer_migrate_folio,
3771 .is_partially_uptodate = block_is_partially_uptodate,
3772 .error_remove_page = generic_error_remove_page,
3773 .swap_activate = ext4_iomap_swap_activate,
3774};
3775
3776static const struct address_space_operations ext4_dax_aops = {
3777 .writepages = ext4_dax_writepages,
3778 .direct_IO = noop_direct_IO,
3779 .dirty_folio = noop_dirty_folio,
3780 .bmap = ext4_bmap,
3781 .swap_activate = ext4_iomap_swap_activate,
3782};
3783
3784void ext4_set_aops(struct inode *inode)
3785{
3786 switch (ext4_inode_journal_mode(inode)) {
3787 case EXT4_INODE_ORDERED_DATA_MODE:
3788 case EXT4_INODE_WRITEBACK_DATA_MODE:
3789 break;
3790 case EXT4_INODE_JOURNAL_DATA_MODE:
3791 inode->i_mapping->a_ops = &ext4_journalled_aops;
3792 return;
3793 default:
3794 BUG();
3795 }
3796 if (IS_DAX(inode))
3797 inode->i_mapping->a_ops = &ext4_dax_aops;
3798 else if (test_opt(inode->i_sb, DELALLOC))
3799 inode->i_mapping->a_ops = &ext4_da_aops;
3800 else
3801 inode->i_mapping->a_ops = &ext4_aops;
3802}
3803
3804static int __ext4_block_zero_page_range(handle_t *handle,
3805 struct address_space *mapping, loff_t from, loff_t length)
3806{
3807 ext4_fsblk_t index = from >> PAGE_SHIFT;
3808 unsigned offset = from & (PAGE_SIZE-1);
3809 unsigned blocksize, pos;
3810 ext4_lblk_t iblock;
3811 struct inode *inode = mapping->host;
3812 struct buffer_head *bh;
3813 struct page *page;
3814 int err = 0;
3815
3816 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3817 mapping_gfp_constraint(mapping, ~__GFP_FS));
3818 if (!page)
3819 return -ENOMEM;
3820
3821 blocksize = inode->i_sb->s_blocksize;
3822
3823 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3824
3825 if (!page_has_buffers(page))
3826 create_empty_buffers(page, blocksize, 0);
3827
3828 /* Find the buffer that contains "offset" */
3829 bh = page_buffers(page);
3830 pos = blocksize;
3831 while (offset >= pos) {
3832 bh = bh->b_this_page;
3833 iblock++;
3834 pos += blocksize;
3835 }
3836 if (buffer_freed(bh)) {
3837 BUFFER_TRACE(bh, "freed: skip");
3838 goto unlock;
3839 }
3840 if (!buffer_mapped(bh)) {
3841 BUFFER_TRACE(bh, "unmapped");
3842 ext4_get_block(inode, iblock, bh, 0);
3843 /* unmapped? It's a hole - nothing to do */
3844 if (!buffer_mapped(bh)) {
3845 BUFFER_TRACE(bh, "still unmapped");
3846 goto unlock;
3847 }
3848 }
3849
3850 /* Ok, it's mapped. Make sure it's up-to-date */
3851 if (PageUptodate(page))
3852 set_buffer_uptodate(bh);
3853
3854 if (!buffer_uptodate(bh)) {
3855 err = ext4_read_bh_lock(bh, 0, true);
3856 if (err)
3857 goto unlock;
3858 if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
3859 /* We expect the key to be set. */
3860 BUG_ON(!fscrypt_has_encryption_key(inode));
3861 err = fscrypt_decrypt_pagecache_blocks(page, blocksize,
3862 bh_offset(bh));
3863 if (err) {
3864 clear_buffer_uptodate(bh);
3865 goto unlock;
3866 }
3867 }
3868 }
3869 if (ext4_should_journal_data(inode)) {
3870 BUFFER_TRACE(bh, "get write access");
3871 err = ext4_journal_get_write_access(handle, inode->i_sb, bh,
3872 EXT4_JTR_NONE);
3873 if (err)
3874 goto unlock;
3875 }
3876 zero_user(page, offset, length);
3877 BUFFER_TRACE(bh, "zeroed end of block");
3878
3879 if (ext4_should_journal_data(inode)) {
3880 err = ext4_handle_dirty_metadata(handle, inode, bh);
3881 } else {
3882 err = 0;
3883 mark_buffer_dirty(bh);
3884 if (ext4_should_order_data(inode))
3885 err = ext4_jbd2_inode_add_write(handle, inode, from,
3886 length);
3887 }
3888
3889unlock:
3890 unlock_page(page);
3891 put_page(page);
3892 return err;
3893}
3894
3895/*
3896 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3897 * starting from file offset 'from'. The range to be zero'd must
3898 * be contained with in one block. If the specified range exceeds
3899 * the end of the block it will be shortened to end of the block
3900 * that corresponds to 'from'
3901 */
3902static int ext4_block_zero_page_range(handle_t *handle,
3903 struct address_space *mapping, loff_t from, loff_t length)
3904{
3905 struct inode *inode = mapping->host;
3906 unsigned offset = from & (PAGE_SIZE-1);
3907 unsigned blocksize = inode->i_sb->s_blocksize;
3908 unsigned max = blocksize - (offset & (blocksize - 1));
3909
3910 /*
3911 * correct length if it does not fall between
3912 * 'from' and the end of the block
3913 */
3914 if (length > max || length < 0)
3915 length = max;
3916
3917 if (IS_DAX(inode)) {
3918 return dax_zero_range(inode, from, length, NULL,
3919 &ext4_iomap_ops);
3920 }
3921 return __ext4_block_zero_page_range(handle, mapping, from, length);
3922}
3923
3924/*
3925 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3926 * up to the end of the block which corresponds to `from'.
3927 * This required during truncate. We need to physically zero the tail end
3928 * of that block so it doesn't yield old data if the file is later grown.
3929 */
3930static int ext4_block_truncate_page(handle_t *handle,
3931 struct address_space *mapping, loff_t from)
3932{
3933 unsigned offset = from & (PAGE_SIZE-1);
3934 unsigned length;
3935 unsigned blocksize;
3936 struct inode *inode = mapping->host;
3937
3938 /* If we are processing an encrypted inode during orphan list handling */
3939 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
3940 return 0;
3941
3942 blocksize = inode->i_sb->s_blocksize;
3943 length = blocksize - (offset & (blocksize - 1));
3944
3945 return ext4_block_zero_page_range(handle, mapping, from, length);
3946}
3947
3948int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3949 loff_t lstart, loff_t length)
3950{
3951 struct super_block *sb = inode->i_sb;
3952 struct address_space *mapping = inode->i_mapping;
3953 unsigned partial_start, partial_end;
3954 ext4_fsblk_t start, end;
3955 loff_t byte_end = (lstart + length - 1);
3956 int err = 0;
3957
3958 partial_start = lstart & (sb->s_blocksize - 1);
3959 partial_end = byte_end & (sb->s_blocksize - 1);
3960
3961 start = lstart >> sb->s_blocksize_bits;
3962 end = byte_end >> sb->s_blocksize_bits;
3963
3964 /* Handle partial zero within the single block */
3965 if (start == end &&
3966 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3967 err = ext4_block_zero_page_range(handle, mapping,
3968 lstart, length);
3969 return err;
3970 }
3971 /* Handle partial zero out on the start of the range */
3972 if (partial_start) {
3973 err = ext4_block_zero_page_range(handle, mapping,
3974 lstart, sb->s_blocksize);
3975 if (err)
3976 return err;
3977 }
3978 /* Handle partial zero out on the end of the range */
3979 if (partial_end != sb->s_blocksize - 1)
3980 err = ext4_block_zero_page_range(handle, mapping,
3981 byte_end - partial_end,
3982 partial_end + 1);
3983 return err;
3984}
3985
3986int ext4_can_truncate(struct inode *inode)
3987{
3988 if (S_ISREG(inode->i_mode))
3989 return 1;
3990 if (S_ISDIR(inode->i_mode))
3991 return 1;
3992 if (S_ISLNK(inode->i_mode))
3993 return !ext4_inode_is_fast_symlink(inode);
3994 return 0;
3995}
3996
3997/*
3998 * We have to make sure i_disksize gets properly updated before we truncate
3999 * page cache due to hole punching or zero range. Otherwise i_disksize update
4000 * can get lost as it may have been postponed to submission of writeback but
4001 * that will never happen after we truncate page cache.
4002 */
4003int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
4004 loff_t len)
4005{
4006 handle_t *handle;
4007 int ret;
4008
4009 loff_t size = i_size_read(inode);
4010
4011 WARN_ON(!inode_is_locked(inode));
4012 if (offset > size || offset + len < size)
4013 return 0;
4014
4015 if (EXT4_I(inode)->i_disksize >= size)
4016 return 0;
4017
4018 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
4019 if (IS_ERR(handle))
4020 return PTR_ERR(handle);
4021 ext4_update_i_disksize(inode, size);
4022 ret = ext4_mark_inode_dirty(handle, inode);
4023 ext4_journal_stop(handle);
4024
4025 return ret;
4026}
4027
4028static void ext4_wait_dax_page(struct inode *inode)
4029{
4030 filemap_invalidate_unlock(inode->i_mapping);
4031 schedule();
4032 filemap_invalidate_lock(inode->i_mapping);
4033}
4034
4035int ext4_break_layouts(struct inode *inode)
4036{
4037 struct page *page;
4038 int error;
4039
4040 if (WARN_ON_ONCE(!rwsem_is_locked(&inode->i_mapping->invalidate_lock)))
4041 return -EINVAL;
4042
4043 do {
4044 page = dax_layout_busy_page(inode->i_mapping);
4045 if (!page)
4046 return 0;
4047
4048 error = ___wait_var_event(&page->_refcount,
4049 atomic_read(&page->_refcount) == 1,
4050 TASK_INTERRUPTIBLE, 0, 0,
4051 ext4_wait_dax_page(inode));
4052 } while (error == 0);
4053
4054 return error;
4055}
4056
4057/*
4058 * ext4_punch_hole: punches a hole in a file by releasing the blocks
4059 * associated with the given offset and length
4060 *
4061 * @inode: File inode
4062 * @offset: The offset where the hole will begin
4063 * @len: The length of the hole
4064 *
4065 * Returns: 0 on success or negative on failure
4066 */
4067
4068int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
4069{
4070 struct inode *inode = file_inode(file);
4071 struct super_block *sb = inode->i_sb;
4072 ext4_lblk_t first_block, stop_block;
4073 struct address_space *mapping = inode->i_mapping;
4074 loff_t first_block_offset, last_block_offset, max_length;
4075 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4076 handle_t *handle;
4077 unsigned int credits;
4078 int ret = 0, ret2 = 0;
4079
4080 trace_ext4_punch_hole(inode, offset, length, 0);
4081
4082 /*
4083 * Write out all dirty pages to avoid race conditions
4084 * Then release them.
4085 */
4086 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4087 ret = filemap_write_and_wait_range(mapping, offset,
4088 offset + length - 1);
4089 if (ret)
4090 return ret;
4091 }
4092
4093 inode_lock(inode);
4094
4095 /* No need to punch hole beyond i_size */
4096 if (offset >= inode->i_size)
4097 goto out_mutex;
4098
4099 /*
4100 * If the hole extends beyond i_size, set the hole
4101 * to end after the page that contains i_size
4102 */
4103 if (offset + length > inode->i_size) {
4104 length = inode->i_size +
4105 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
4106 offset;
4107 }
4108
4109 /*
4110 * For punch hole the length + offset needs to be within one block
4111 * before last range. Adjust the length if it goes beyond that limit.
4112 */
4113 max_length = sbi->s_bitmap_maxbytes - inode->i_sb->s_blocksize;
4114 if (offset + length > max_length)
4115 length = max_length - offset;
4116
4117 if (offset & (sb->s_blocksize - 1) ||
4118 (offset + length) & (sb->s_blocksize - 1)) {
4119 /*
4120 * Attach jinode to inode for jbd2 if we do any zeroing of
4121 * partial block
4122 */
4123 ret = ext4_inode_attach_jinode(inode);
4124 if (ret < 0)
4125 goto out_mutex;
4126
4127 }
4128
4129 /* Wait all existing dio workers, newcomers will block on i_rwsem */
4130 inode_dio_wait(inode);
4131
4132 ret = file_modified(file);
4133 if (ret)
4134 goto out_mutex;
4135
4136 /*
4137 * Prevent page faults from reinstantiating pages we have released from
4138 * page cache.
4139 */
4140 filemap_invalidate_lock(mapping);
4141
4142 ret = ext4_break_layouts(inode);
4143 if (ret)
4144 goto out_dio;
4145
4146 first_block_offset = round_up(offset, sb->s_blocksize);
4147 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4148
4149 /* Now release the pages and zero block aligned part of pages*/
4150 if (last_block_offset > first_block_offset) {
4151 ret = ext4_update_disksize_before_punch(inode, offset, length);
4152 if (ret)
4153 goto out_dio;
4154 truncate_pagecache_range(inode, first_block_offset,
4155 last_block_offset);
4156 }
4157
4158 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4159 credits = ext4_writepage_trans_blocks(inode);
4160 else
4161 credits = ext4_blocks_for_truncate(inode);
4162 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4163 if (IS_ERR(handle)) {
4164 ret = PTR_ERR(handle);
4165 ext4_std_error(sb, ret);
4166 goto out_dio;
4167 }
4168
4169 ret = ext4_zero_partial_blocks(handle, inode, offset,
4170 length);
4171 if (ret)
4172 goto out_stop;
4173
4174 first_block = (offset + sb->s_blocksize - 1) >>
4175 EXT4_BLOCK_SIZE_BITS(sb);
4176 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4177
4178 /* If there are blocks to remove, do it */
4179 if (stop_block > first_block) {
4180
4181 down_write(&EXT4_I(inode)->i_data_sem);
4182 ext4_discard_preallocations(inode, 0);
4183
4184 ret = ext4_es_remove_extent(inode, first_block,
4185 stop_block - first_block);
4186 if (ret) {
4187 up_write(&EXT4_I(inode)->i_data_sem);
4188 goto out_stop;
4189 }
4190
4191 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4192 ret = ext4_ext_remove_space(inode, first_block,
4193 stop_block - 1);
4194 else
4195 ret = ext4_ind_remove_space(handle, inode, first_block,
4196 stop_block);
4197
4198 up_write(&EXT4_I(inode)->i_data_sem);
4199 }
4200 ext4_fc_track_range(handle, inode, first_block, stop_block);
4201 if (IS_SYNC(inode))
4202 ext4_handle_sync(handle);
4203
4204 inode->i_mtime = inode->i_ctime = current_time(inode);
4205 ret2 = ext4_mark_inode_dirty(handle, inode);
4206 if (unlikely(ret2))
4207 ret = ret2;
4208 if (ret >= 0)
4209 ext4_update_inode_fsync_trans(handle, inode, 1);
4210out_stop:
4211 ext4_journal_stop(handle);
4212out_dio:
4213 filemap_invalidate_unlock(mapping);
4214out_mutex:
4215 inode_unlock(inode);
4216 return ret;
4217}
4218
4219int ext4_inode_attach_jinode(struct inode *inode)
4220{
4221 struct ext4_inode_info *ei = EXT4_I(inode);
4222 struct jbd2_inode *jinode;
4223
4224 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4225 return 0;
4226
4227 jinode = jbd2_alloc_inode(GFP_KERNEL);
4228 spin_lock(&inode->i_lock);
4229 if (!ei->jinode) {
4230 if (!jinode) {
4231 spin_unlock(&inode->i_lock);
4232 return -ENOMEM;
4233 }
4234 ei->jinode = jinode;
4235 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4236 jinode = NULL;
4237 }
4238 spin_unlock(&inode->i_lock);
4239 if (unlikely(jinode != NULL))
4240 jbd2_free_inode(jinode);
4241 return 0;
4242}
4243
4244/*
4245 * ext4_truncate()
4246 *
4247 * We block out ext4_get_block() block instantiations across the entire
4248 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4249 * simultaneously on behalf of the same inode.
4250 *
4251 * As we work through the truncate and commit bits of it to the journal there
4252 * is one core, guiding principle: the file's tree must always be consistent on
4253 * disk. We must be able to restart the truncate after a crash.
4254 *
4255 * The file's tree may be transiently inconsistent in memory (although it
4256 * probably isn't), but whenever we close off and commit a journal transaction,
4257 * the contents of (the filesystem + the journal) must be consistent and
4258 * restartable. It's pretty simple, really: bottom up, right to left (although
4259 * left-to-right works OK too).
4260 *
4261 * Note that at recovery time, journal replay occurs *before* the restart of
4262 * truncate against the orphan inode list.
4263 *
4264 * The committed inode has the new, desired i_size (which is the same as
4265 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4266 * that this inode's truncate did not complete and it will again call
4267 * ext4_truncate() to have another go. So there will be instantiated blocks
4268 * to the right of the truncation point in a crashed ext4 filesystem. But
4269 * that's fine - as long as they are linked from the inode, the post-crash
4270 * ext4_truncate() run will find them and release them.
4271 */
4272int ext4_truncate(struct inode *inode)
4273{
4274 struct ext4_inode_info *ei = EXT4_I(inode);
4275 unsigned int credits;
4276 int err = 0, err2;
4277 handle_t *handle;
4278 struct address_space *mapping = inode->i_mapping;
4279
4280 /*
4281 * There is a possibility that we're either freeing the inode
4282 * or it's a completely new inode. In those cases we might not
4283 * have i_rwsem locked because it's not necessary.
4284 */
4285 if (!(inode->i_state & (I_NEW|I_FREEING)))
4286 WARN_ON(!inode_is_locked(inode));
4287 trace_ext4_truncate_enter(inode);
4288
4289 if (!ext4_can_truncate(inode))
4290 goto out_trace;
4291
4292 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4293 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4294
4295 if (ext4_has_inline_data(inode)) {
4296 int has_inline = 1;
4297
4298 err = ext4_inline_data_truncate(inode, &has_inline);
4299 if (err || has_inline)
4300 goto out_trace;
4301 }
4302
4303 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4304 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4305 err = ext4_inode_attach_jinode(inode);
4306 if (err)
4307 goto out_trace;
4308 }
4309
4310 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4311 credits = ext4_writepage_trans_blocks(inode);
4312 else
4313 credits = ext4_blocks_for_truncate(inode);
4314
4315 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4316 if (IS_ERR(handle)) {
4317 err = PTR_ERR(handle);
4318 goto out_trace;
4319 }
4320
4321 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4322 ext4_block_truncate_page(handle, mapping, inode->i_size);
4323
4324 /*
4325 * We add the inode to the orphan list, so that if this
4326 * truncate spans multiple transactions, and we crash, we will
4327 * resume the truncate when the filesystem recovers. It also
4328 * marks the inode dirty, to catch the new size.
4329 *
4330 * Implication: the file must always be in a sane, consistent
4331 * truncatable state while each transaction commits.
4332 */
4333 err = ext4_orphan_add(handle, inode);
4334 if (err)
4335 goto out_stop;
4336
4337 down_write(&EXT4_I(inode)->i_data_sem);
4338
4339 ext4_discard_preallocations(inode, 0);
4340
4341 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4342 err = ext4_ext_truncate(handle, inode);
4343 else
4344 ext4_ind_truncate(handle, inode);
4345
4346 up_write(&ei->i_data_sem);
4347 if (err)
4348 goto out_stop;
4349
4350 if (IS_SYNC(inode))
4351 ext4_handle_sync(handle);
4352
4353out_stop:
4354 /*
4355 * If this was a simple ftruncate() and the file will remain alive,
4356 * then we need to clear up the orphan record which we created above.
4357 * However, if this was a real unlink then we were called by
4358 * ext4_evict_inode(), and we allow that function to clean up the
4359 * orphan info for us.
4360 */
4361 if (inode->i_nlink)
4362 ext4_orphan_del(handle, inode);
4363
4364 inode->i_mtime = inode->i_ctime = current_time(inode);
4365 err2 = ext4_mark_inode_dirty(handle, inode);
4366 if (unlikely(err2 && !err))
4367 err = err2;
4368 ext4_journal_stop(handle);
4369
4370out_trace:
4371 trace_ext4_truncate_exit(inode);
4372 return err;
4373}
4374
4375static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4376{
4377 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4378 return inode_peek_iversion_raw(inode);
4379 else
4380 return inode_peek_iversion(inode);
4381}
4382
4383static int ext4_inode_blocks_set(struct ext4_inode *raw_inode,
4384 struct ext4_inode_info *ei)
4385{
4386 struct inode *inode = &(ei->vfs_inode);
4387 u64 i_blocks = READ_ONCE(inode->i_blocks);
4388 struct super_block *sb = inode->i_sb;
4389
4390 if (i_blocks <= ~0U) {
4391 /*
4392 * i_blocks can be represented in a 32 bit variable
4393 * as multiple of 512 bytes
4394 */
4395 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4396 raw_inode->i_blocks_high = 0;
4397 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4398 return 0;
4399 }
4400
4401 /*
4402 * This should never happen since sb->s_maxbytes should not have
4403 * allowed this, sb->s_maxbytes was set according to the huge_file
4404 * feature in ext4_fill_super().
4405 */
4406 if (!ext4_has_feature_huge_file(sb))
4407 return -EFSCORRUPTED;
4408
4409 if (i_blocks <= 0xffffffffffffULL) {
4410 /*
4411 * i_blocks can be represented in a 48 bit variable
4412 * as multiple of 512 bytes
4413 */
4414 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4415 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4416 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4417 } else {
4418 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4419 /* i_block is stored in file system block size */
4420 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4421 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4422 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4423 }
4424 return 0;
4425}
4426
4427static int ext4_fill_raw_inode(struct inode *inode, struct ext4_inode *raw_inode)
4428{
4429 struct ext4_inode_info *ei = EXT4_I(inode);
4430 uid_t i_uid;
4431 gid_t i_gid;
4432 projid_t i_projid;
4433 int block;
4434 int err;
4435
4436 err = ext4_inode_blocks_set(raw_inode, ei);
4437
4438 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4439 i_uid = i_uid_read(inode);
4440 i_gid = i_gid_read(inode);
4441 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4442 if (!(test_opt(inode->i_sb, NO_UID32))) {
4443 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4444 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4445 /*
4446 * Fix up interoperability with old kernels. Otherwise,
4447 * old inodes get re-used with the upper 16 bits of the
4448 * uid/gid intact.
4449 */
4450 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4451 raw_inode->i_uid_high = 0;
4452 raw_inode->i_gid_high = 0;
4453 } else {
4454 raw_inode->i_uid_high =
4455 cpu_to_le16(high_16_bits(i_uid));
4456 raw_inode->i_gid_high =
4457 cpu_to_le16(high_16_bits(i_gid));
4458 }
4459 } else {
4460 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4461 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4462 raw_inode->i_uid_high = 0;
4463 raw_inode->i_gid_high = 0;
4464 }
4465 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4466
4467 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4468 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4469 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4470 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4471
4472 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4473 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4474 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4475 raw_inode->i_file_acl_high =
4476 cpu_to_le16(ei->i_file_acl >> 32);
4477 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4478 ext4_isize_set(raw_inode, ei->i_disksize);
4479
4480 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4481 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4482 if (old_valid_dev(inode->i_rdev)) {
4483 raw_inode->i_block[0] =
4484 cpu_to_le32(old_encode_dev(inode->i_rdev));
4485 raw_inode->i_block[1] = 0;
4486 } else {
4487 raw_inode->i_block[0] = 0;
4488 raw_inode->i_block[1] =
4489 cpu_to_le32(new_encode_dev(inode->i_rdev));
4490 raw_inode->i_block[2] = 0;
4491 }
4492 } else if (!ext4_has_inline_data(inode)) {
4493 for (block = 0; block < EXT4_N_BLOCKS; block++)
4494 raw_inode->i_block[block] = ei->i_data[block];
4495 }
4496
4497 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4498 u64 ivers = ext4_inode_peek_iversion(inode);
4499
4500 raw_inode->i_disk_version = cpu_to_le32(ivers);
4501 if (ei->i_extra_isize) {
4502 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4503 raw_inode->i_version_hi =
4504 cpu_to_le32(ivers >> 32);
4505 raw_inode->i_extra_isize =
4506 cpu_to_le16(ei->i_extra_isize);
4507 }
4508 }
4509
4510 if (i_projid != EXT4_DEF_PROJID &&
4511 !ext4_has_feature_project(inode->i_sb))
4512 err = err ?: -EFSCORRUPTED;
4513
4514 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4515 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4516 raw_inode->i_projid = cpu_to_le32(i_projid);
4517
4518 ext4_inode_csum_set(inode, raw_inode, ei);
4519 return err;
4520}
4521
4522/*
4523 * ext4_get_inode_loc returns with an extra refcount against the inode's
4524 * underlying buffer_head on success. If we pass 'inode' and it does not
4525 * have in-inode xattr, we have all inode data in memory that is needed
4526 * to recreate the on-disk version of this inode.
4527 */
4528static int __ext4_get_inode_loc(struct super_block *sb, unsigned long ino,
4529 struct inode *inode, struct ext4_iloc *iloc,
4530 ext4_fsblk_t *ret_block)
4531{
4532 struct ext4_group_desc *gdp;
4533 struct buffer_head *bh;
4534 ext4_fsblk_t block;
4535 struct blk_plug plug;
4536 int inodes_per_block, inode_offset;
4537
4538 iloc->bh = NULL;
4539 if (ino < EXT4_ROOT_INO ||
4540 ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4541 return -EFSCORRUPTED;
4542
4543 iloc->block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
4544 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4545 if (!gdp)
4546 return -EIO;
4547
4548 /*
4549 * Figure out the offset within the block group inode table
4550 */
4551 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4552 inode_offset = ((ino - 1) %
4553 EXT4_INODES_PER_GROUP(sb));
4554 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4555
4556 block = ext4_inode_table(sb, gdp);
4557 if ((block <= le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) ||
4558 (block >= ext4_blocks_count(EXT4_SB(sb)->s_es))) {
4559 ext4_error(sb, "Invalid inode table block %llu in "
4560 "block_group %u", block, iloc->block_group);
4561 return -EFSCORRUPTED;
4562 }
4563 block += (inode_offset / inodes_per_block);
4564
4565 bh = sb_getblk(sb, block);
4566 if (unlikely(!bh))
4567 return -ENOMEM;
4568 if (ext4_buffer_uptodate(bh))
4569 goto has_buffer;
4570
4571 lock_buffer(bh);
4572 if (ext4_buffer_uptodate(bh)) {
4573 /* Someone brought it uptodate while we waited */
4574 unlock_buffer(bh);
4575 goto has_buffer;
4576 }
4577
4578 /*
4579 * If we have all information of the inode in memory and this
4580 * is the only valid inode in the block, we need not read the
4581 * block.
4582 */
4583 if (inode && !ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4584 struct buffer_head *bitmap_bh;
4585 int i, start;
4586
4587 start = inode_offset & ~(inodes_per_block - 1);
4588
4589 /* Is the inode bitmap in cache? */
4590 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4591 if (unlikely(!bitmap_bh))
4592 goto make_io;
4593
4594 /*
4595 * If the inode bitmap isn't in cache then the
4596 * optimisation may end up performing two reads instead
4597 * of one, so skip it.
4598 */
4599 if (!buffer_uptodate(bitmap_bh)) {
4600 brelse(bitmap_bh);
4601 goto make_io;
4602 }
4603 for (i = start; i < start + inodes_per_block; i++) {
4604 if (i == inode_offset)
4605 continue;
4606 if (ext4_test_bit(i, bitmap_bh->b_data))
4607 break;
4608 }
4609 brelse(bitmap_bh);
4610 if (i == start + inodes_per_block) {
4611 struct ext4_inode *raw_inode =
4612 (struct ext4_inode *) (bh->b_data + iloc->offset);
4613
4614 /* all other inodes are free, so skip I/O */
4615 memset(bh->b_data, 0, bh->b_size);
4616 if (!ext4_test_inode_state(inode, EXT4_STATE_NEW))
4617 ext4_fill_raw_inode(inode, raw_inode);
4618 set_buffer_uptodate(bh);
4619 unlock_buffer(bh);
4620 goto has_buffer;
4621 }
4622 }
4623
4624make_io:
4625 /*
4626 * If we need to do any I/O, try to pre-readahead extra
4627 * blocks from the inode table.
4628 */
4629 blk_start_plug(&plug);
4630 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4631 ext4_fsblk_t b, end, table;
4632 unsigned num;
4633 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4634
4635 table = ext4_inode_table(sb, gdp);
4636 /* s_inode_readahead_blks is always a power of 2 */
4637 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4638 if (table > b)
4639 b = table;
4640 end = b + ra_blks;
4641 num = EXT4_INODES_PER_GROUP(sb);
4642 if (ext4_has_group_desc_csum(sb))
4643 num -= ext4_itable_unused_count(sb, gdp);
4644 table += num / inodes_per_block;
4645 if (end > table)
4646 end = table;
4647 while (b <= end)
4648 ext4_sb_breadahead_unmovable(sb, b++);
4649 }
4650
4651 /*
4652 * There are other valid inodes in the buffer, this inode
4653 * has in-inode xattrs, or we don't have this inode in memory.
4654 * Read the block from disk.
4655 */
4656 trace_ext4_load_inode(sb, ino);
4657 ext4_read_bh_nowait(bh, REQ_META | REQ_PRIO, NULL);
4658 blk_finish_plug(&plug);
4659 wait_on_buffer(bh);
4660 ext4_simulate_fail_bh(sb, bh, EXT4_SIM_INODE_EIO);
4661 if (!buffer_uptodate(bh)) {
4662 if (ret_block)
4663 *ret_block = block;
4664 brelse(bh);
4665 return -EIO;
4666 }
4667has_buffer:
4668 iloc->bh = bh;
4669 return 0;
4670}
4671
4672static int __ext4_get_inode_loc_noinmem(struct inode *inode,
4673 struct ext4_iloc *iloc)
4674{
4675 ext4_fsblk_t err_blk = 0;
4676 int ret;
4677
4678 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, NULL, iloc,
4679 &err_blk);
4680
4681 if (ret == -EIO)
4682 ext4_error_inode_block(inode, err_blk, EIO,
4683 "unable to read itable block");
4684
4685 return ret;
4686}
4687
4688int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4689{
4690 ext4_fsblk_t err_blk = 0;
4691 int ret;
4692
4693 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, inode, iloc,
4694 &err_blk);
4695
4696 if (ret == -EIO)
4697 ext4_error_inode_block(inode, err_blk, EIO,
4698 "unable to read itable block");
4699
4700 return ret;
4701}
4702
4703
4704int ext4_get_fc_inode_loc(struct super_block *sb, unsigned long ino,
4705 struct ext4_iloc *iloc)
4706{
4707 return __ext4_get_inode_loc(sb, ino, NULL, iloc, NULL);
4708}
4709
4710static bool ext4_should_enable_dax(struct inode *inode)
4711{
4712 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4713
4714 if (test_opt2(inode->i_sb, DAX_NEVER))
4715 return false;
4716 if (!S_ISREG(inode->i_mode))
4717 return false;
4718 if (ext4_should_journal_data(inode))
4719 return false;
4720 if (ext4_has_inline_data(inode))
4721 return false;
4722 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4723 return false;
4724 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4725 return false;
4726 if (!test_bit(EXT4_FLAGS_BDEV_IS_DAX, &sbi->s_ext4_flags))
4727 return false;
4728 if (test_opt(inode->i_sb, DAX_ALWAYS))
4729 return true;
4730
4731 return ext4_test_inode_flag(inode, EXT4_INODE_DAX);
4732}
4733
4734void ext4_set_inode_flags(struct inode *inode, bool init)
4735{
4736 unsigned int flags = EXT4_I(inode)->i_flags;
4737 unsigned int new_fl = 0;
4738
4739 WARN_ON_ONCE(IS_DAX(inode) && init);
4740
4741 if (flags & EXT4_SYNC_FL)
4742 new_fl |= S_SYNC;
4743 if (flags & EXT4_APPEND_FL)
4744 new_fl |= S_APPEND;
4745 if (flags & EXT4_IMMUTABLE_FL)
4746 new_fl |= S_IMMUTABLE;
4747 if (flags & EXT4_NOATIME_FL)
4748 new_fl |= S_NOATIME;
4749 if (flags & EXT4_DIRSYNC_FL)
4750 new_fl |= S_DIRSYNC;
4751
4752 /* Because of the way inode_set_flags() works we must preserve S_DAX
4753 * here if already set. */
4754 new_fl |= (inode->i_flags & S_DAX);
4755 if (init && ext4_should_enable_dax(inode))
4756 new_fl |= S_DAX;
4757
4758 if (flags & EXT4_ENCRYPT_FL)
4759 new_fl |= S_ENCRYPTED;
4760 if (flags & EXT4_CASEFOLD_FL)
4761 new_fl |= S_CASEFOLD;
4762 if (flags & EXT4_VERITY_FL)
4763 new_fl |= S_VERITY;
4764 inode_set_flags(inode, new_fl,
4765 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4766 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4767}
4768
4769static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4770 struct ext4_inode_info *ei)
4771{
4772 blkcnt_t i_blocks ;
4773 struct inode *inode = &(ei->vfs_inode);
4774 struct super_block *sb = inode->i_sb;
4775
4776 if (ext4_has_feature_huge_file(sb)) {
4777 /* we are using combined 48 bit field */
4778 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4779 le32_to_cpu(raw_inode->i_blocks_lo);
4780 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4781 /* i_blocks represent file system block size */
4782 return i_blocks << (inode->i_blkbits - 9);
4783 } else {
4784 return i_blocks;
4785 }
4786 } else {
4787 return le32_to_cpu(raw_inode->i_blocks_lo);
4788 }
4789}
4790
4791static inline int ext4_iget_extra_inode(struct inode *inode,
4792 struct ext4_inode *raw_inode,
4793 struct ext4_inode_info *ei)
4794{
4795 __le32 *magic = (void *)raw_inode +
4796 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4797
4798 if (EXT4_INODE_HAS_XATTR_SPACE(inode) &&
4799 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4800 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4801 return ext4_find_inline_data_nolock(inode);
4802 } else
4803 EXT4_I(inode)->i_inline_off = 0;
4804 return 0;
4805}
4806
4807int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4808{
4809 if (!ext4_has_feature_project(inode->i_sb))
4810 return -EOPNOTSUPP;
4811 *projid = EXT4_I(inode)->i_projid;
4812 return 0;
4813}
4814
4815/*
4816 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4817 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4818 * set.
4819 */
4820static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4821{
4822 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4823 inode_set_iversion_raw(inode, val);
4824 else
4825 inode_set_iversion_queried(inode, val);
4826}
4827
4828struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4829 ext4_iget_flags flags, const char *function,
4830 unsigned int line)
4831{
4832 struct ext4_iloc iloc;
4833 struct ext4_inode *raw_inode;
4834 struct ext4_inode_info *ei;
4835 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
4836 struct inode *inode;
4837 journal_t *journal = EXT4_SB(sb)->s_journal;
4838 long ret;
4839 loff_t size;
4840 int block;
4841 uid_t i_uid;
4842 gid_t i_gid;
4843 projid_t i_projid;
4844
4845 if ((!(flags & EXT4_IGET_SPECIAL) &&
4846 ((ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO) ||
4847 ino == le32_to_cpu(es->s_usr_quota_inum) ||
4848 ino == le32_to_cpu(es->s_grp_quota_inum) ||
4849 ino == le32_to_cpu(es->s_prj_quota_inum) ||
4850 ino == le32_to_cpu(es->s_orphan_file_inum))) ||
4851 (ino < EXT4_ROOT_INO) ||
4852 (ino > le32_to_cpu(es->s_inodes_count))) {
4853 if (flags & EXT4_IGET_HANDLE)
4854 return ERR_PTR(-ESTALE);
4855 __ext4_error(sb, function, line, false, EFSCORRUPTED, 0,
4856 "inode #%lu: comm %s: iget: illegal inode #",
4857 ino, current->comm);
4858 return ERR_PTR(-EFSCORRUPTED);
4859 }
4860
4861 inode = iget_locked(sb, ino);
4862 if (!inode)
4863 return ERR_PTR(-ENOMEM);
4864 if (!(inode->i_state & I_NEW))
4865 return inode;
4866
4867 ei = EXT4_I(inode);
4868 iloc.bh = NULL;
4869
4870 ret = __ext4_get_inode_loc_noinmem(inode, &iloc);
4871 if (ret < 0)
4872 goto bad_inode;
4873 raw_inode = ext4_raw_inode(&iloc);
4874
4875 if ((ino == EXT4_ROOT_INO) && (raw_inode->i_links_count == 0)) {
4876 ext4_error_inode(inode, function, line, 0,
4877 "iget: root inode unallocated");
4878 ret = -EFSCORRUPTED;
4879 goto bad_inode;
4880 }
4881
4882 if ((flags & EXT4_IGET_HANDLE) &&
4883 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4884 ret = -ESTALE;
4885 goto bad_inode;
4886 }
4887
4888 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4889 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4890 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4891 EXT4_INODE_SIZE(inode->i_sb) ||
4892 (ei->i_extra_isize & 3)) {
4893 ext4_error_inode(inode, function, line, 0,
4894 "iget: bad extra_isize %u "
4895 "(inode size %u)",
4896 ei->i_extra_isize,
4897 EXT4_INODE_SIZE(inode->i_sb));
4898 ret = -EFSCORRUPTED;
4899 goto bad_inode;
4900 }
4901 } else
4902 ei->i_extra_isize = 0;
4903
4904 /* Precompute checksum seed for inode metadata */
4905 if (ext4_has_metadata_csum(sb)) {
4906 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4907 __u32 csum;
4908 __le32 inum = cpu_to_le32(inode->i_ino);
4909 __le32 gen = raw_inode->i_generation;
4910 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4911 sizeof(inum));
4912 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4913 sizeof(gen));
4914 }
4915
4916 if ((!ext4_inode_csum_verify(inode, raw_inode, ei) ||
4917 ext4_simulate_fail(sb, EXT4_SIM_INODE_CRC)) &&
4918 (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))) {
4919 ext4_error_inode_err(inode, function, line, 0,
4920 EFSBADCRC, "iget: checksum invalid");
4921 ret = -EFSBADCRC;
4922 goto bad_inode;
4923 }
4924
4925 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4926 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4927 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4928 if (ext4_has_feature_project(sb) &&
4929 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4930 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4931 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4932 else
4933 i_projid = EXT4_DEF_PROJID;
4934
4935 if (!(test_opt(inode->i_sb, NO_UID32))) {
4936 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4937 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4938 }
4939 i_uid_write(inode, i_uid);
4940 i_gid_write(inode, i_gid);
4941 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4942 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4943
4944 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4945 ei->i_inline_off = 0;
4946 ei->i_dir_start_lookup = 0;
4947 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4948 /* We now have enough fields to check if the inode was active or not.
4949 * This is needed because nfsd might try to access dead inodes
4950 * the test is that same one that e2fsck uses
4951 * NeilBrown 1999oct15
4952 */
4953 if (inode->i_nlink == 0) {
4954 if ((inode->i_mode == 0 ||
4955 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4956 ino != EXT4_BOOT_LOADER_INO) {
4957 /* this inode is deleted */
4958 ret = -ESTALE;
4959 goto bad_inode;
4960 }
4961 /* The only unlinked inodes we let through here have
4962 * valid i_mode and are being read by the orphan
4963 * recovery code: that's fine, we're about to complete
4964 * the process of deleting those.
4965 * OR it is the EXT4_BOOT_LOADER_INO which is
4966 * not initialized on a new filesystem. */
4967 }
4968 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4969 ext4_set_inode_flags(inode, true);
4970 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4971 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4972 if (ext4_has_feature_64bit(sb))
4973 ei->i_file_acl |=
4974 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4975 inode->i_size = ext4_isize(sb, raw_inode);
4976 if ((size = i_size_read(inode)) < 0) {
4977 ext4_error_inode(inode, function, line, 0,
4978 "iget: bad i_size value: %lld", size);
4979 ret = -EFSCORRUPTED;
4980 goto bad_inode;
4981 }
4982 /*
4983 * If dir_index is not enabled but there's dir with INDEX flag set,
4984 * we'd normally treat htree data as empty space. But with metadata
4985 * checksumming that corrupts checksums so forbid that.
4986 */
4987 if (!ext4_has_feature_dir_index(sb) && ext4_has_metadata_csum(sb) &&
4988 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
4989 ext4_error_inode(inode, function, line, 0,
4990 "iget: Dir with htree data on filesystem without dir_index feature.");
4991 ret = -EFSCORRUPTED;
4992 goto bad_inode;
4993 }
4994 ei->i_disksize = inode->i_size;
4995#ifdef CONFIG_QUOTA
4996 ei->i_reserved_quota = 0;
4997#endif
4998 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4999 ei->i_block_group = iloc.block_group;
5000 ei->i_last_alloc_group = ~0;
5001 /*
5002 * NOTE! The in-memory inode i_data array is in little-endian order
5003 * even on big-endian machines: we do NOT byteswap the block numbers!
5004 */
5005 for (block = 0; block < EXT4_N_BLOCKS; block++)
5006 ei->i_data[block] = raw_inode->i_block[block];
5007 INIT_LIST_HEAD(&ei->i_orphan);
5008 ext4_fc_init_inode(&ei->vfs_inode);
5009
5010 /*
5011 * Set transaction id's of transactions that have to be committed
5012 * to finish f[data]sync. We set them to currently running transaction
5013 * as we cannot be sure that the inode or some of its metadata isn't
5014 * part of the transaction - the inode could have been reclaimed and
5015 * now it is reread from disk.
5016 */
5017 if (journal) {
5018 transaction_t *transaction;
5019 tid_t tid;
5020
5021 read_lock(&journal->j_state_lock);
5022 if (journal->j_running_transaction)
5023 transaction = journal->j_running_transaction;
5024 else
5025 transaction = journal->j_committing_transaction;
5026 if (transaction)
5027 tid = transaction->t_tid;
5028 else
5029 tid = journal->j_commit_sequence;
5030 read_unlock(&journal->j_state_lock);
5031 ei->i_sync_tid = tid;
5032 ei->i_datasync_tid = tid;
5033 }
5034
5035 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5036 if (ei->i_extra_isize == 0) {
5037 /* The extra space is currently unused. Use it. */
5038 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
5039 ei->i_extra_isize = sizeof(struct ext4_inode) -
5040 EXT4_GOOD_OLD_INODE_SIZE;
5041 } else {
5042 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
5043 if (ret)
5044 goto bad_inode;
5045 }
5046 }
5047
5048 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5049 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5050 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5051 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5052
5053 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5054 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
5055
5056 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5057 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5058 ivers |=
5059 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5060 }
5061 ext4_inode_set_iversion_queried(inode, ivers);
5062 }
5063
5064 ret = 0;
5065 if (ei->i_file_acl &&
5066 !ext4_inode_block_valid(inode, ei->i_file_acl, 1)) {
5067 ext4_error_inode(inode, function, line, 0,
5068 "iget: bad extended attribute block %llu",
5069 ei->i_file_acl);
5070 ret = -EFSCORRUPTED;
5071 goto bad_inode;
5072 } else if (!ext4_has_inline_data(inode)) {
5073 /* validate the block references in the inode */
5074 if (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) &&
5075 (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5076 (S_ISLNK(inode->i_mode) &&
5077 !ext4_inode_is_fast_symlink(inode)))) {
5078 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5079 ret = ext4_ext_check_inode(inode);
5080 else
5081 ret = ext4_ind_check_inode(inode);
5082 }
5083 }
5084 if (ret)
5085 goto bad_inode;
5086
5087 if (S_ISREG(inode->i_mode)) {
5088 inode->i_op = &ext4_file_inode_operations;
5089 inode->i_fop = &ext4_file_operations;
5090 ext4_set_aops(inode);
5091 } else if (S_ISDIR(inode->i_mode)) {
5092 inode->i_op = &ext4_dir_inode_operations;
5093 inode->i_fop = &ext4_dir_operations;
5094 } else if (S_ISLNK(inode->i_mode)) {
5095 /* VFS does not allow setting these so must be corruption */
5096 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
5097 ext4_error_inode(inode, function, line, 0,
5098 "iget: immutable or append flags "
5099 "not allowed on symlinks");
5100 ret = -EFSCORRUPTED;
5101 goto bad_inode;
5102 }
5103 if (IS_ENCRYPTED(inode)) {
5104 inode->i_op = &ext4_encrypted_symlink_inode_operations;
5105 } else if (ext4_inode_is_fast_symlink(inode)) {
5106 inode->i_link = (char *)ei->i_data;
5107 inode->i_op = &ext4_fast_symlink_inode_operations;
5108 nd_terminate_link(ei->i_data, inode->i_size,
5109 sizeof(ei->i_data) - 1);
5110 } else {
5111 inode->i_op = &ext4_symlink_inode_operations;
5112 }
5113 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5114 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5115 inode->i_op = &ext4_special_inode_operations;
5116 if (raw_inode->i_block[0])
5117 init_special_inode(inode, inode->i_mode,
5118 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5119 else
5120 init_special_inode(inode, inode->i_mode,
5121 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5122 } else if (ino == EXT4_BOOT_LOADER_INO) {
5123 make_bad_inode(inode);
5124 } else {
5125 ret = -EFSCORRUPTED;
5126 ext4_error_inode(inode, function, line, 0,
5127 "iget: bogus i_mode (%o)", inode->i_mode);
5128 goto bad_inode;
5129 }
5130 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb))
5131 ext4_error_inode(inode, function, line, 0,
5132 "casefold flag without casefold feature");
5133 if (is_bad_inode(inode) && !(flags & EXT4_IGET_BAD)) {
5134 ext4_error_inode(inode, function, line, 0,
5135 "bad inode without EXT4_IGET_BAD flag");
5136 ret = -EUCLEAN;
5137 goto bad_inode;
5138 }
5139
5140 brelse(iloc.bh);
5141 unlock_new_inode(inode);
5142 return inode;
5143
5144bad_inode:
5145 brelse(iloc.bh);
5146 iget_failed(inode);
5147 return ERR_PTR(ret);
5148}
5149
5150static void __ext4_update_other_inode_time(struct super_block *sb,
5151 unsigned long orig_ino,
5152 unsigned long ino,
5153 struct ext4_inode *raw_inode)
5154{
5155 struct inode *inode;
5156
5157 inode = find_inode_by_ino_rcu(sb, ino);
5158 if (!inode)
5159 return;
5160
5161 if (!inode_is_dirtytime_only(inode))
5162 return;
5163
5164 spin_lock(&inode->i_lock);
5165 if (inode_is_dirtytime_only(inode)) {
5166 struct ext4_inode_info *ei = EXT4_I(inode);
5167
5168 inode->i_state &= ~I_DIRTY_TIME;
5169 spin_unlock(&inode->i_lock);
5170
5171 spin_lock(&ei->i_raw_lock);
5172 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5173 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5174 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5175 ext4_inode_csum_set(inode, raw_inode, ei);
5176 spin_unlock(&ei->i_raw_lock);
5177 trace_ext4_other_inode_update_time(inode, orig_ino);
5178 return;
5179 }
5180 spin_unlock(&inode->i_lock);
5181}
5182
5183/*
5184 * Opportunistically update the other time fields for other inodes in
5185 * the same inode table block.
5186 */
5187static void ext4_update_other_inodes_time(struct super_block *sb,
5188 unsigned long orig_ino, char *buf)
5189{
5190 unsigned long ino;
5191 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5192 int inode_size = EXT4_INODE_SIZE(sb);
5193
5194 /*
5195 * Calculate the first inode in the inode table block. Inode
5196 * numbers are one-based. That is, the first inode in a block
5197 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5198 */
5199 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5200 rcu_read_lock();
5201 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5202 if (ino == orig_ino)
5203 continue;
5204 __ext4_update_other_inode_time(sb, orig_ino, ino,
5205 (struct ext4_inode *)buf);
5206 }
5207 rcu_read_unlock();
5208}
5209
5210/*
5211 * Post the struct inode info into an on-disk inode location in the
5212 * buffer-cache. This gobbles the caller's reference to the
5213 * buffer_head in the inode location struct.
5214 *
5215 * The caller must have write access to iloc->bh.
5216 */
5217static int ext4_do_update_inode(handle_t *handle,
5218 struct inode *inode,
5219 struct ext4_iloc *iloc)
5220{
5221 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5222 struct ext4_inode_info *ei = EXT4_I(inode);
5223 struct buffer_head *bh = iloc->bh;
5224 struct super_block *sb = inode->i_sb;
5225 int err;
5226 int need_datasync = 0, set_large_file = 0;
5227
5228 spin_lock(&ei->i_raw_lock);
5229
5230 /*
5231 * For fields not tracked in the in-memory inode, initialise them
5232 * to zero for new inodes.
5233 */
5234 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5235 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5236
5237 if (READ_ONCE(ei->i_disksize) != ext4_isize(inode->i_sb, raw_inode))
5238 need_datasync = 1;
5239 if (ei->i_disksize > 0x7fffffffULL) {
5240 if (!ext4_has_feature_large_file(sb) ||
5241 EXT4_SB(sb)->s_es->s_rev_level == cpu_to_le32(EXT4_GOOD_OLD_REV))
5242 set_large_file = 1;
5243 }
5244
5245 err = ext4_fill_raw_inode(inode, raw_inode);
5246 spin_unlock(&ei->i_raw_lock);
5247 if (err) {
5248 EXT4_ERROR_INODE(inode, "corrupted inode contents");
5249 goto out_brelse;
5250 }
5251
5252 if (inode->i_sb->s_flags & SB_LAZYTIME)
5253 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5254 bh->b_data);
5255
5256 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5257 err = ext4_handle_dirty_metadata(handle, NULL, bh);
5258 if (err)
5259 goto out_error;
5260 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5261 if (set_large_file) {
5262 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5263 err = ext4_journal_get_write_access(handle, sb,
5264 EXT4_SB(sb)->s_sbh,
5265 EXT4_JTR_NONE);
5266 if (err)
5267 goto out_error;
5268 lock_buffer(EXT4_SB(sb)->s_sbh);
5269 ext4_set_feature_large_file(sb);
5270 ext4_superblock_csum_set(sb);
5271 unlock_buffer(EXT4_SB(sb)->s_sbh);
5272 ext4_handle_sync(handle);
5273 err = ext4_handle_dirty_metadata(handle, NULL,
5274 EXT4_SB(sb)->s_sbh);
5275 }
5276 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5277out_error:
5278 ext4_std_error(inode->i_sb, err);
5279out_brelse:
5280 brelse(bh);
5281 return err;
5282}
5283
5284/*
5285 * ext4_write_inode()
5286 *
5287 * We are called from a few places:
5288 *
5289 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5290 * Here, there will be no transaction running. We wait for any running
5291 * transaction to commit.
5292 *
5293 * - Within flush work (sys_sync(), kupdate and such).
5294 * We wait on commit, if told to.
5295 *
5296 * - Within iput_final() -> write_inode_now()
5297 * We wait on commit, if told to.
5298 *
5299 * In all cases it is actually safe for us to return without doing anything,
5300 * because the inode has been copied into a raw inode buffer in
5301 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5302 * writeback.
5303 *
5304 * Note that we are absolutely dependent upon all inode dirtiers doing the
5305 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5306 * which we are interested.
5307 *
5308 * It would be a bug for them to not do this. The code:
5309 *
5310 * mark_inode_dirty(inode)
5311 * stuff();
5312 * inode->i_size = expr;
5313 *
5314 * is in error because write_inode() could occur while `stuff()' is running,
5315 * and the new i_size will be lost. Plus the inode will no longer be on the
5316 * superblock's dirty inode list.
5317 */
5318int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5319{
5320 int err;
5321
5322 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC) ||
5323 sb_rdonly(inode->i_sb))
5324 return 0;
5325
5326 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5327 return -EIO;
5328
5329 if (EXT4_SB(inode->i_sb)->s_journal) {
5330 if (ext4_journal_current_handle()) {
5331 ext4_debug("called recursively, non-PF_MEMALLOC!\n");
5332 dump_stack();
5333 return -EIO;
5334 }
5335
5336 /*
5337 * No need to force transaction in WB_SYNC_NONE mode. Also
5338 * ext4_sync_fs() will force the commit after everything is
5339 * written.
5340 */
5341 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5342 return 0;
5343
5344 err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
5345 EXT4_I(inode)->i_sync_tid);
5346 } else {
5347 struct ext4_iloc iloc;
5348
5349 err = __ext4_get_inode_loc_noinmem(inode, &iloc);
5350 if (err)
5351 return err;
5352 /*
5353 * sync(2) will flush the whole buffer cache. No need to do
5354 * it here separately for each inode.
5355 */
5356 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5357 sync_dirty_buffer(iloc.bh);
5358 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5359 ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
5360 "IO error syncing inode");
5361 err = -EIO;
5362 }
5363 brelse(iloc.bh);
5364 }
5365 return err;
5366}
5367
5368/*
5369 * In data=journal mode ext4_journalled_invalidate_folio() may fail to invalidate
5370 * buffers that are attached to a folio straddling i_size and are undergoing
5371 * commit. In that case we have to wait for commit to finish and try again.
5372 */
5373static void ext4_wait_for_tail_page_commit(struct inode *inode)
5374{
5375 unsigned offset;
5376 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5377 tid_t commit_tid = 0;
5378 int ret;
5379
5380 offset = inode->i_size & (PAGE_SIZE - 1);
5381 /*
5382 * If the folio is fully truncated, we don't need to wait for any commit
5383 * (and we even should not as __ext4_journalled_invalidate_folio() may
5384 * strip all buffers from the folio but keep the folio dirty which can then
5385 * confuse e.g. concurrent ext4_writepage() seeing dirty folio without
5386 * buffers). Also we don't need to wait for any commit if all buffers in
5387 * the folio remain valid. This is most beneficial for the common case of
5388 * blocksize == PAGESIZE.
5389 */
5390 if (!offset || offset > (PAGE_SIZE - i_blocksize(inode)))
5391 return;
5392 while (1) {
5393 struct folio *folio = filemap_lock_folio(inode->i_mapping,
5394 inode->i_size >> PAGE_SHIFT);
5395 if (!folio)
5396 return;
5397 ret = __ext4_journalled_invalidate_folio(folio, offset,
5398 folio_size(folio) - offset);
5399 folio_unlock(folio);
5400 folio_put(folio);
5401 if (ret != -EBUSY)
5402 return;
5403 commit_tid = 0;
5404 read_lock(&journal->j_state_lock);
5405 if (journal->j_committing_transaction)
5406 commit_tid = journal->j_committing_transaction->t_tid;
5407 read_unlock(&journal->j_state_lock);
5408 if (commit_tid)
5409 jbd2_log_wait_commit(journal, commit_tid);
5410 }
5411}
5412
5413/*
5414 * ext4_setattr()
5415 *
5416 * Called from notify_change.
5417 *
5418 * We want to trap VFS attempts to truncate the file as soon as
5419 * possible. In particular, we want to make sure that when the VFS
5420 * shrinks i_size, we put the inode on the orphan list and modify
5421 * i_disksize immediately, so that during the subsequent flushing of
5422 * dirty pages and freeing of disk blocks, we can guarantee that any
5423 * commit will leave the blocks being flushed in an unused state on
5424 * disk. (On recovery, the inode will get truncated and the blocks will
5425 * be freed, so we have a strong guarantee that no future commit will
5426 * leave these blocks visible to the user.)
5427 *
5428 * Another thing we have to assure is that if we are in ordered mode
5429 * and inode is still attached to the committing transaction, we must
5430 * we start writeout of all the dirty pages which are being truncated.
5431 * This way we are sure that all the data written in the previous
5432 * transaction are already on disk (truncate waits for pages under
5433 * writeback).
5434 *
5435 * Called with inode->i_rwsem down.
5436 */
5437int ext4_setattr(struct user_namespace *mnt_userns, struct dentry *dentry,
5438 struct iattr *attr)
5439{
5440 struct inode *inode = d_inode(dentry);
5441 int error, rc = 0;
5442 int orphan = 0;
5443 const unsigned int ia_valid = attr->ia_valid;
5444 bool inc_ivers = true;
5445
5446 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5447 return -EIO;
5448
5449 if (unlikely(IS_IMMUTABLE(inode)))
5450 return -EPERM;
5451
5452 if (unlikely(IS_APPEND(inode) &&
5453 (ia_valid & (ATTR_MODE | ATTR_UID |
5454 ATTR_GID | ATTR_TIMES_SET))))
5455 return -EPERM;
5456
5457 error = setattr_prepare(mnt_userns, dentry, attr);
5458 if (error)
5459 return error;
5460
5461 error = fscrypt_prepare_setattr(dentry, attr);
5462 if (error)
5463 return error;
5464
5465 error = fsverity_prepare_setattr(dentry, attr);
5466 if (error)
5467 return error;
5468
5469 if (is_quota_modification(mnt_userns, inode, attr)) {
5470 error = dquot_initialize(inode);
5471 if (error)
5472 return error;
5473 }
5474
5475 if (i_uid_needs_update(mnt_userns, attr, inode) ||
5476 i_gid_needs_update(mnt_userns, attr, inode)) {
5477 handle_t *handle;
5478
5479 /* (user+group)*(old+new) structure, inode write (sb,
5480 * inode block, ? - but truncate inode update has it) */
5481 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5482 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5483 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5484 if (IS_ERR(handle)) {
5485 error = PTR_ERR(handle);
5486 goto err_out;
5487 }
5488
5489 /* dquot_transfer() calls back ext4_get_inode_usage() which
5490 * counts xattr inode references.
5491 */
5492 down_read(&EXT4_I(inode)->xattr_sem);
5493 error = dquot_transfer(mnt_userns, inode, attr);
5494 up_read(&EXT4_I(inode)->xattr_sem);
5495
5496 if (error) {
5497 ext4_journal_stop(handle);
5498 return error;
5499 }
5500 /* Update corresponding info in inode so that everything is in
5501 * one transaction */
5502 i_uid_update(mnt_userns, attr, inode);
5503 i_gid_update(mnt_userns, attr, inode);
5504 error = ext4_mark_inode_dirty(handle, inode);
5505 ext4_journal_stop(handle);
5506 if (unlikely(error)) {
5507 return error;
5508 }
5509 }
5510
5511 if (attr->ia_valid & ATTR_SIZE) {
5512 handle_t *handle;
5513 loff_t oldsize = inode->i_size;
5514 loff_t old_disksize;
5515 int shrink = (attr->ia_size < inode->i_size);
5516
5517 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5518 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5519
5520 if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5521 return -EFBIG;
5522 }
5523 }
5524 if (!S_ISREG(inode->i_mode)) {
5525 return -EINVAL;
5526 }
5527
5528 if (attr->ia_size == inode->i_size)
5529 inc_ivers = false;
5530
5531 if (shrink) {
5532 if (ext4_should_order_data(inode)) {
5533 error = ext4_begin_ordered_truncate(inode,
5534 attr->ia_size);
5535 if (error)
5536 goto err_out;
5537 }
5538 /*
5539 * Blocks are going to be removed from the inode. Wait
5540 * for dio in flight.
5541 */
5542 inode_dio_wait(inode);
5543 }
5544
5545 filemap_invalidate_lock(inode->i_mapping);
5546
5547 rc = ext4_break_layouts(inode);
5548 if (rc) {
5549 filemap_invalidate_unlock(inode->i_mapping);
5550 goto err_out;
5551 }
5552
5553 if (attr->ia_size != inode->i_size) {
5554 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5555 if (IS_ERR(handle)) {
5556 error = PTR_ERR(handle);
5557 goto out_mmap_sem;
5558 }
5559 if (ext4_handle_valid(handle) && shrink) {
5560 error = ext4_orphan_add(handle, inode);
5561 orphan = 1;
5562 }
5563 /*
5564 * Update c/mtime on truncate up, ext4_truncate() will
5565 * update c/mtime in shrink case below
5566 */
5567 if (!shrink) {
5568 inode->i_mtime = current_time(inode);
5569 inode->i_ctime = inode->i_mtime;
5570 }
5571
5572 if (shrink)
5573 ext4_fc_track_range(handle, inode,
5574 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5575 inode->i_sb->s_blocksize_bits,
5576 EXT_MAX_BLOCKS - 1);
5577 else
5578 ext4_fc_track_range(
5579 handle, inode,
5580 (oldsize > 0 ? oldsize - 1 : oldsize) >>
5581 inode->i_sb->s_blocksize_bits,
5582 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5583 inode->i_sb->s_blocksize_bits);
5584
5585 down_write(&EXT4_I(inode)->i_data_sem);
5586 old_disksize = EXT4_I(inode)->i_disksize;
5587 EXT4_I(inode)->i_disksize = attr->ia_size;
5588 rc = ext4_mark_inode_dirty(handle, inode);
5589 if (!error)
5590 error = rc;
5591 /*
5592 * We have to update i_size under i_data_sem together
5593 * with i_disksize to avoid races with writeback code
5594 * running ext4_wb_update_i_disksize().
5595 */
5596 if (!error)
5597 i_size_write(inode, attr->ia_size);
5598 else
5599 EXT4_I(inode)->i_disksize = old_disksize;
5600 up_write(&EXT4_I(inode)->i_data_sem);
5601 ext4_journal_stop(handle);
5602 if (error)
5603 goto out_mmap_sem;
5604 if (!shrink) {
5605 pagecache_isize_extended(inode, oldsize,
5606 inode->i_size);
5607 } else if (ext4_should_journal_data(inode)) {
5608 ext4_wait_for_tail_page_commit(inode);
5609 }
5610 }
5611
5612 /*
5613 * Truncate pagecache after we've waited for commit
5614 * in data=journal mode to make pages freeable.
5615 */
5616 truncate_pagecache(inode, inode->i_size);
5617 /*
5618 * Call ext4_truncate() even if i_size didn't change to
5619 * truncate possible preallocated blocks.
5620 */
5621 if (attr->ia_size <= oldsize) {
5622 rc = ext4_truncate(inode);
5623 if (rc)
5624 error = rc;
5625 }
5626out_mmap_sem:
5627 filemap_invalidate_unlock(inode->i_mapping);
5628 }
5629
5630 if (!error) {
5631 if (inc_ivers)
5632 inode_inc_iversion(inode);
5633 setattr_copy(mnt_userns, inode, attr);
5634 mark_inode_dirty(inode);
5635 }
5636
5637 /*
5638 * If the call to ext4_truncate failed to get a transaction handle at
5639 * all, we need to clean up the in-core orphan list manually.
5640 */
5641 if (orphan && inode->i_nlink)
5642 ext4_orphan_del(NULL, inode);
5643
5644 if (!error && (ia_valid & ATTR_MODE))
5645 rc = posix_acl_chmod(mnt_userns, dentry, inode->i_mode);
5646
5647err_out:
5648 if (error)
5649 ext4_std_error(inode->i_sb, error);
5650 if (!error)
5651 error = rc;
5652 return error;
5653}
5654
5655u32 ext4_dio_alignment(struct inode *inode)
5656{
5657 if (fsverity_active(inode))
5658 return 0;
5659 if (ext4_should_journal_data(inode))
5660 return 0;
5661 if (ext4_has_inline_data(inode))
5662 return 0;
5663 if (IS_ENCRYPTED(inode)) {
5664 if (!fscrypt_dio_supported(inode))
5665 return 0;
5666 return i_blocksize(inode);
5667 }
5668 return 1; /* use the iomap defaults */
5669}
5670
5671int ext4_getattr(struct user_namespace *mnt_userns, const struct path *path,
5672 struct kstat *stat, u32 request_mask, unsigned int query_flags)
5673{
5674 struct inode *inode = d_inode(path->dentry);
5675 struct ext4_inode *raw_inode;
5676 struct ext4_inode_info *ei = EXT4_I(inode);
5677 unsigned int flags;
5678
5679 if ((request_mask & STATX_BTIME) &&
5680 EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5681 stat->result_mask |= STATX_BTIME;
5682 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5683 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5684 }
5685
5686 /*
5687 * Return the DIO alignment restrictions if requested. We only return
5688 * this information when requested, since on encrypted files it might
5689 * take a fair bit of work to get if the file wasn't opened recently.
5690 */
5691 if ((request_mask & STATX_DIOALIGN) && S_ISREG(inode->i_mode)) {
5692 u32 dio_align = ext4_dio_alignment(inode);
5693
5694 stat->result_mask |= STATX_DIOALIGN;
5695 if (dio_align == 1) {
5696 struct block_device *bdev = inode->i_sb->s_bdev;
5697
5698 /* iomap defaults */
5699 stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
5700 stat->dio_offset_align = bdev_logical_block_size(bdev);
5701 } else {
5702 stat->dio_mem_align = dio_align;
5703 stat->dio_offset_align = dio_align;
5704 }
5705 }
5706
5707 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5708 if (flags & EXT4_APPEND_FL)
5709 stat->attributes |= STATX_ATTR_APPEND;
5710 if (flags & EXT4_COMPR_FL)
5711 stat->attributes |= STATX_ATTR_COMPRESSED;
5712 if (flags & EXT4_ENCRYPT_FL)
5713 stat->attributes |= STATX_ATTR_ENCRYPTED;
5714 if (flags & EXT4_IMMUTABLE_FL)
5715 stat->attributes |= STATX_ATTR_IMMUTABLE;
5716 if (flags & EXT4_NODUMP_FL)
5717 stat->attributes |= STATX_ATTR_NODUMP;
5718 if (flags & EXT4_VERITY_FL)
5719 stat->attributes |= STATX_ATTR_VERITY;
5720
5721 stat->attributes_mask |= (STATX_ATTR_APPEND |
5722 STATX_ATTR_COMPRESSED |
5723 STATX_ATTR_ENCRYPTED |
5724 STATX_ATTR_IMMUTABLE |
5725 STATX_ATTR_NODUMP |
5726 STATX_ATTR_VERITY);
5727
5728 generic_fillattr(mnt_userns, inode, stat);
5729 return 0;
5730}
5731
5732int ext4_file_getattr(struct user_namespace *mnt_userns,
5733 const struct path *path, struct kstat *stat,
5734 u32 request_mask, unsigned int query_flags)
5735{
5736 struct inode *inode = d_inode(path->dentry);
5737 u64 delalloc_blocks;
5738
5739 ext4_getattr(mnt_userns, path, stat, request_mask, query_flags);
5740
5741 /*
5742 * If there is inline data in the inode, the inode will normally not
5743 * have data blocks allocated (it may have an external xattr block).
5744 * Report at least one sector for such files, so tools like tar, rsync,
5745 * others don't incorrectly think the file is completely sparse.
5746 */
5747 if (unlikely(ext4_has_inline_data(inode)))
5748 stat->blocks += (stat->size + 511) >> 9;
5749
5750 /*
5751 * We can't update i_blocks if the block allocation is delayed
5752 * otherwise in the case of system crash before the real block
5753 * allocation is done, we will have i_blocks inconsistent with
5754 * on-disk file blocks.
5755 * We always keep i_blocks updated together with real
5756 * allocation. But to not confuse with user, stat
5757 * will return the blocks that include the delayed allocation
5758 * blocks for this file.
5759 */
5760 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5761 EXT4_I(inode)->i_reserved_data_blocks);
5762 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5763 return 0;
5764}
5765
5766static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5767 int pextents)
5768{
5769 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5770 return ext4_ind_trans_blocks(inode, lblocks);
5771 return ext4_ext_index_trans_blocks(inode, pextents);
5772}
5773
5774/*
5775 * Account for index blocks, block groups bitmaps and block group
5776 * descriptor blocks if modify datablocks and index blocks
5777 * worse case, the indexs blocks spread over different block groups
5778 *
5779 * If datablocks are discontiguous, they are possible to spread over
5780 * different block groups too. If they are contiguous, with flexbg,
5781 * they could still across block group boundary.
5782 *
5783 * Also account for superblock, inode, quota and xattr blocks
5784 */
5785static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5786 int pextents)
5787{
5788 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5789 int gdpblocks;
5790 int idxblocks;
5791 int ret = 0;
5792
5793 /*
5794 * How many index blocks need to touch to map @lblocks logical blocks
5795 * to @pextents physical extents?
5796 */
5797 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5798
5799 ret = idxblocks;
5800
5801 /*
5802 * Now let's see how many group bitmaps and group descriptors need
5803 * to account
5804 */
5805 groups = idxblocks + pextents;
5806 gdpblocks = groups;
5807 if (groups > ngroups)
5808 groups = ngroups;
5809 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5810 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5811
5812 /* bitmaps and block group descriptor blocks */
5813 ret += groups + gdpblocks;
5814
5815 /* Blocks for super block, inode, quota and xattr blocks */
5816 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5817
5818 return ret;
5819}
5820
5821/*
5822 * Calculate the total number of credits to reserve to fit
5823 * the modification of a single pages into a single transaction,
5824 * which may include multiple chunks of block allocations.
5825 *
5826 * This could be called via ext4_write_begin()
5827 *
5828 * We need to consider the worse case, when
5829 * one new block per extent.
5830 */
5831int ext4_writepage_trans_blocks(struct inode *inode)
5832{
5833 int bpp = ext4_journal_blocks_per_page(inode);
5834 int ret;
5835
5836 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5837
5838 /* Account for data blocks for journalled mode */
5839 if (ext4_should_journal_data(inode))
5840 ret += bpp;
5841 return ret;
5842}
5843
5844/*
5845 * Calculate the journal credits for a chunk of data modification.
5846 *
5847 * This is called from DIO, fallocate or whoever calling
5848 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5849 *
5850 * journal buffers for data blocks are not included here, as DIO
5851 * and fallocate do no need to journal data buffers.
5852 */
5853int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5854{
5855 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5856}
5857
5858/*
5859 * The caller must have previously called ext4_reserve_inode_write().
5860 * Give this, we know that the caller already has write access to iloc->bh.
5861 */
5862int ext4_mark_iloc_dirty(handle_t *handle,
5863 struct inode *inode, struct ext4_iloc *iloc)
5864{
5865 int err = 0;
5866
5867 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
5868 put_bh(iloc->bh);
5869 return -EIO;
5870 }
5871 ext4_fc_track_inode(handle, inode);
5872
5873 /* the do_update_inode consumes one bh->b_count */
5874 get_bh(iloc->bh);
5875
5876 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5877 err = ext4_do_update_inode(handle, inode, iloc);
5878 put_bh(iloc->bh);
5879 return err;
5880}
5881
5882/*
5883 * On success, We end up with an outstanding reference count against
5884 * iloc->bh. This _must_ be cleaned up later.
5885 */
5886
5887int
5888ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5889 struct ext4_iloc *iloc)
5890{
5891 int err;
5892
5893 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5894 return -EIO;
5895
5896 err = ext4_get_inode_loc(inode, iloc);
5897 if (!err) {
5898 BUFFER_TRACE(iloc->bh, "get_write_access");
5899 err = ext4_journal_get_write_access(handle, inode->i_sb,
5900 iloc->bh, EXT4_JTR_NONE);
5901 if (err) {
5902 brelse(iloc->bh);
5903 iloc->bh = NULL;
5904 }
5905 }
5906 ext4_std_error(inode->i_sb, err);
5907 return err;
5908}
5909
5910static int __ext4_expand_extra_isize(struct inode *inode,
5911 unsigned int new_extra_isize,
5912 struct ext4_iloc *iloc,
5913 handle_t *handle, int *no_expand)
5914{
5915 struct ext4_inode *raw_inode;
5916 struct ext4_xattr_ibody_header *header;
5917 unsigned int inode_size = EXT4_INODE_SIZE(inode->i_sb);
5918 struct ext4_inode_info *ei = EXT4_I(inode);
5919 int error;
5920
5921 /* this was checked at iget time, but double check for good measure */
5922 if ((EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize > inode_size) ||
5923 (ei->i_extra_isize & 3)) {
5924 EXT4_ERROR_INODE(inode, "bad extra_isize %u (inode size %u)",
5925 ei->i_extra_isize,
5926 EXT4_INODE_SIZE(inode->i_sb));
5927 return -EFSCORRUPTED;
5928 }
5929 if ((new_extra_isize < ei->i_extra_isize) ||
5930 (new_extra_isize < 4) ||
5931 (new_extra_isize > inode_size - EXT4_GOOD_OLD_INODE_SIZE))
5932 return -EINVAL; /* Should never happen */
5933
5934 raw_inode = ext4_raw_inode(iloc);
5935
5936 header = IHDR(inode, raw_inode);
5937
5938 /* No extended attributes present */
5939 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5940 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5941 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5942 EXT4_I(inode)->i_extra_isize, 0,
5943 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5944 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5945 return 0;
5946 }
5947
5948 /*
5949 * We may need to allocate external xattr block so we need quotas
5950 * initialized. Here we can be called with various locks held so we
5951 * cannot affort to initialize quotas ourselves. So just bail.
5952 */
5953 if (dquot_initialize_needed(inode))
5954 return -EAGAIN;
5955
5956 /* try to expand with EAs present */
5957 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5958 raw_inode, handle);
5959 if (error) {
5960 /*
5961 * Inode size expansion failed; don't try again
5962 */
5963 *no_expand = 1;
5964 }
5965
5966 return error;
5967}
5968
5969/*
5970 * Expand an inode by new_extra_isize bytes.
5971 * Returns 0 on success or negative error number on failure.
5972 */
5973static int ext4_try_to_expand_extra_isize(struct inode *inode,
5974 unsigned int new_extra_isize,
5975 struct ext4_iloc iloc,
5976 handle_t *handle)
5977{
5978 int no_expand;
5979 int error;
5980
5981 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5982 return -EOVERFLOW;
5983
5984 /*
5985 * In nojournal mode, we can immediately attempt to expand
5986 * the inode. When journaled, we first need to obtain extra
5987 * buffer credits since we may write into the EA block
5988 * with this same handle. If journal_extend fails, then it will
5989 * only result in a minor loss of functionality for that inode.
5990 * If this is felt to be critical, then e2fsck should be run to
5991 * force a large enough s_min_extra_isize.
5992 */
5993 if (ext4_journal_extend(handle,
5994 EXT4_DATA_TRANS_BLOCKS(inode->i_sb), 0) != 0)
5995 return -ENOSPC;
5996
5997 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5998 return -EBUSY;
5999
6000 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
6001 handle, &no_expand);
6002 ext4_write_unlock_xattr(inode, &no_expand);
6003
6004 return error;
6005}
6006
6007int ext4_expand_extra_isize(struct inode *inode,
6008 unsigned int new_extra_isize,
6009 struct ext4_iloc *iloc)
6010{
6011 handle_t *handle;
6012 int no_expand;
6013 int error, rc;
6014
6015 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
6016 brelse(iloc->bh);
6017 return -EOVERFLOW;
6018 }
6019
6020 handle = ext4_journal_start(inode, EXT4_HT_INODE,
6021 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
6022 if (IS_ERR(handle)) {
6023 error = PTR_ERR(handle);
6024 brelse(iloc->bh);
6025 return error;
6026 }
6027
6028 ext4_write_lock_xattr(inode, &no_expand);
6029
6030 BUFFER_TRACE(iloc->bh, "get_write_access");
6031 error = ext4_journal_get_write_access(handle, inode->i_sb, iloc->bh,
6032 EXT4_JTR_NONE);
6033 if (error) {
6034 brelse(iloc->bh);
6035 goto out_unlock;
6036 }
6037
6038 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
6039 handle, &no_expand);
6040
6041 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
6042 if (!error)
6043 error = rc;
6044
6045out_unlock:
6046 ext4_write_unlock_xattr(inode, &no_expand);
6047 ext4_journal_stop(handle);
6048 return error;
6049}
6050
6051/*
6052 * What we do here is to mark the in-core inode as clean with respect to inode
6053 * dirtiness (it may still be data-dirty).
6054 * This means that the in-core inode may be reaped by prune_icache
6055 * without having to perform any I/O. This is a very good thing,
6056 * because *any* task may call prune_icache - even ones which
6057 * have a transaction open against a different journal.
6058 *
6059 * Is this cheating? Not really. Sure, we haven't written the
6060 * inode out, but prune_icache isn't a user-visible syncing function.
6061 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
6062 * we start and wait on commits.
6063 */
6064int __ext4_mark_inode_dirty(handle_t *handle, struct inode *inode,
6065 const char *func, unsigned int line)
6066{
6067 struct ext4_iloc iloc;
6068 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
6069 int err;
6070
6071 might_sleep();
6072 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
6073 err = ext4_reserve_inode_write(handle, inode, &iloc);
6074 if (err)
6075 goto out;
6076
6077 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
6078 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
6079 iloc, handle);
6080
6081 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
6082out:
6083 if (unlikely(err))
6084 ext4_error_inode_err(inode, func, line, 0, err,
6085 "mark_inode_dirty error");
6086 return err;
6087}
6088
6089/*
6090 * ext4_dirty_inode() is called from __mark_inode_dirty()
6091 *
6092 * We're really interested in the case where a file is being extended.
6093 * i_size has been changed by generic_commit_write() and we thus need
6094 * to include the updated inode in the current transaction.
6095 *
6096 * Also, dquot_alloc_block() will always dirty the inode when blocks
6097 * are allocated to the file.
6098 *
6099 * If the inode is marked synchronous, we don't honour that here - doing
6100 * so would cause a commit on atime updates, which we don't bother doing.
6101 * We handle synchronous inodes at the highest possible level.
6102 */
6103void ext4_dirty_inode(struct inode *inode, int flags)
6104{
6105 handle_t *handle;
6106
6107 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
6108 if (IS_ERR(handle))
6109 return;
6110 ext4_mark_inode_dirty(handle, inode);
6111 ext4_journal_stop(handle);
6112}
6113
6114int ext4_change_inode_journal_flag(struct inode *inode, int val)
6115{
6116 journal_t *journal;
6117 handle_t *handle;
6118 int err;
6119 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
6120
6121 /*
6122 * We have to be very careful here: changing a data block's
6123 * journaling status dynamically is dangerous. If we write a
6124 * data block to the journal, change the status and then delete
6125 * that block, we risk forgetting to revoke the old log record
6126 * from the journal and so a subsequent replay can corrupt data.
6127 * So, first we make sure that the journal is empty and that
6128 * nobody is changing anything.
6129 */
6130
6131 journal = EXT4_JOURNAL(inode);
6132 if (!journal)
6133 return 0;
6134 if (is_journal_aborted(journal))
6135 return -EROFS;
6136
6137 /* Wait for all existing dio workers */
6138 inode_dio_wait(inode);
6139
6140 /*
6141 * Before flushing the journal and switching inode's aops, we have
6142 * to flush all dirty data the inode has. There can be outstanding
6143 * delayed allocations, there can be unwritten extents created by
6144 * fallocate or buffered writes in dioread_nolock mode covered by
6145 * dirty data which can be converted only after flushing the dirty
6146 * data (and journalled aops don't know how to handle these cases).
6147 */
6148 if (val) {
6149 filemap_invalidate_lock(inode->i_mapping);
6150 err = filemap_write_and_wait(inode->i_mapping);
6151 if (err < 0) {
6152 filemap_invalidate_unlock(inode->i_mapping);
6153 return err;
6154 }
6155 }
6156
6157 percpu_down_write(&sbi->s_writepages_rwsem);
6158 jbd2_journal_lock_updates(journal);
6159
6160 /*
6161 * OK, there are no updates running now, and all cached data is
6162 * synced to disk. We are now in a completely consistent state
6163 * which doesn't have anything in the journal, and we know that
6164 * no filesystem updates are running, so it is safe to modify
6165 * the inode's in-core data-journaling state flag now.
6166 */
6167
6168 if (val)
6169 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6170 else {
6171 err = jbd2_journal_flush(journal, 0);
6172 if (err < 0) {
6173 jbd2_journal_unlock_updates(journal);
6174 percpu_up_write(&sbi->s_writepages_rwsem);
6175 return err;
6176 }
6177 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6178 }
6179 ext4_set_aops(inode);
6180
6181 jbd2_journal_unlock_updates(journal);
6182 percpu_up_write(&sbi->s_writepages_rwsem);
6183
6184 if (val)
6185 filemap_invalidate_unlock(inode->i_mapping);
6186
6187 /* Finally we can mark the inode as dirty. */
6188
6189 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6190 if (IS_ERR(handle))
6191 return PTR_ERR(handle);
6192
6193 ext4_fc_mark_ineligible(inode->i_sb,
6194 EXT4_FC_REASON_JOURNAL_FLAG_CHANGE, handle);
6195 err = ext4_mark_inode_dirty(handle, inode);
6196 ext4_handle_sync(handle);
6197 ext4_journal_stop(handle);
6198 ext4_std_error(inode->i_sb, err);
6199
6200 return err;
6201}
6202
6203static int ext4_bh_unmapped(handle_t *handle, struct inode *inode,
6204 struct buffer_head *bh)
6205{
6206 return !buffer_mapped(bh);
6207}
6208
6209vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
6210{
6211 struct vm_area_struct *vma = vmf->vma;
6212 struct page *page = vmf->page;
6213 loff_t size;
6214 unsigned long len;
6215 int err;
6216 vm_fault_t ret;
6217 struct file *file = vma->vm_file;
6218 struct inode *inode = file_inode(file);
6219 struct address_space *mapping = inode->i_mapping;
6220 handle_t *handle;
6221 get_block_t *get_block;
6222 int retries = 0;
6223
6224 if (unlikely(IS_IMMUTABLE(inode)))
6225 return VM_FAULT_SIGBUS;
6226
6227 sb_start_pagefault(inode->i_sb);
6228 file_update_time(vma->vm_file);
6229
6230 filemap_invalidate_lock_shared(mapping);
6231
6232 err = ext4_convert_inline_data(inode);
6233 if (err)
6234 goto out_ret;
6235
6236 /*
6237 * On data journalling we skip straight to the transaction handle:
6238 * there's no delalloc; page truncated will be checked later; the
6239 * early return w/ all buffers mapped (calculates size/len) can't
6240 * be used; and there's no dioread_nolock, so only ext4_get_block.
6241 */
6242 if (ext4_should_journal_data(inode))
6243 goto retry_alloc;
6244
6245 /* Delalloc case is easy... */
6246 if (test_opt(inode->i_sb, DELALLOC) &&
6247 !ext4_nonda_switch(inode->i_sb)) {
6248 do {
6249 err = block_page_mkwrite(vma, vmf,
6250 ext4_da_get_block_prep);
6251 } while (err == -ENOSPC &&
6252 ext4_should_retry_alloc(inode->i_sb, &retries));
6253 goto out_ret;
6254 }
6255
6256 lock_page(page);
6257 size = i_size_read(inode);
6258 /* Page got truncated from under us? */
6259 if (page->mapping != mapping || page_offset(page) > size) {
6260 unlock_page(page);
6261 ret = VM_FAULT_NOPAGE;
6262 goto out;
6263 }
6264
6265 if (page->index == size >> PAGE_SHIFT)
6266 len = size & ~PAGE_MASK;
6267 else
6268 len = PAGE_SIZE;
6269 /*
6270 * Return if we have all the buffers mapped. This avoids the need to do
6271 * journal_start/journal_stop which can block and take a long time
6272 *
6273 * This cannot be done for data journalling, as we have to add the
6274 * inode to the transaction's list to writeprotect pages on commit.
6275 */
6276 if (page_has_buffers(page)) {
6277 if (!ext4_walk_page_buffers(NULL, inode, page_buffers(page),
6278 0, len, NULL,
6279 ext4_bh_unmapped)) {
6280 /* Wait so that we don't change page under IO */
6281 wait_for_stable_page(page);
6282 ret = VM_FAULT_LOCKED;
6283 goto out;
6284 }
6285 }
6286 unlock_page(page);
6287 /* OK, we need to fill the hole... */
6288 if (ext4_should_dioread_nolock(inode))
6289 get_block = ext4_get_block_unwritten;
6290 else
6291 get_block = ext4_get_block;
6292retry_alloc:
6293 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6294 ext4_writepage_trans_blocks(inode));
6295 if (IS_ERR(handle)) {
6296 ret = VM_FAULT_SIGBUS;
6297 goto out;
6298 }
6299 /*
6300 * Data journalling can't use block_page_mkwrite() because it
6301 * will set_buffer_dirty() before do_journal_get_write_access()
6302 * thus might hit warning messages for dirty metadata buffers.
6303 */
6304 if (!ext4_should_journal_data(inode)) {
6305 err = block_page_mkwrite(vma, vmf, get_block);
6306 } else {
6307 lock_page(page);
6308 size = i_size_read(inode);
6309 /* Page got truncated from under us? */
6310 if (page->mapping != mapping || page_offset(page) > size) {
6311 ret = VM_FAULT_NOPAGE;
6312 goto out_error;
6313 }
6314
6315 if (page->index == size >> PAGE_SHIFT)
6316 len = size & ~PAGE_MASK;
6317 else
6318 len = PAGE_SIZE;
6319
6320 err = __block_write_begin(page, 0, len, ext4_get_block);
6321 if (!err) {
6322 ret = VM_FAULT_SIGBUS;
6323 if (ext4_walk_page_buffers(handle, inode,
6324 page_buffers(page), 0, len, NULL,
6325 do_journal_get_write_access))
6326 goto out_error;
6327 if (ext4_walk_page_buffers(handle, inode,
6328 page_buffers(page), 0, len, NULL,
6329 write_end_fn))
6330 goto out_error;
6331 if (ext4_jbd2_inode_add_write(handle, inode,
6332 page_offset(page), len))
6333 goto out_error;
6334 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
6335 } else {
6336 unlock_page(page);
6337 }
6338 }
6339 ext4_journal_stop(handle);
6340 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6341 goto retry_alloc;
6342out_ret:
6343 ret = block_page_mkwrite_return(err);
6344out:
6345 filemap_invalidate_unlock_shared(mapping);
6346 sb_end_pagefault(inode->i_sb);
6347 return ret;
6348out_error:
6349 unlock_page(page);
6350 ext4_journal_stop(handle);
6351 goto out;
6352}