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/jbd2.h>
24#include <linux/highuid.h>
25#include <linux/pagemap.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/ratelimit.h>
40#include <linux/aio.h>
41#include <linux/bitops.h>
42
43#include "ext4_jbd2.h"
44#include "xattr.h"
45#include "acl.h"
46#include "truncate.h"
47
48#include <trace/events/ext4.h>
49
50#define MPAGE_DA_EXTENT_TAIL 0x01
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 __u16 csum_lo;
57 __u16 csum_hi = 0;
58 __u32 csum;
59
60 csum_lo = le16_to_cpu(raw->i_checksum_lo);
61 raw->i_checksum_lo = 0;
62 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
63 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
64 csum_hi = le16_to_cpu(raw->i_checksum_hi);
65 raw->i_checksum_hi = 0;
66 }
67
68 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw,
69 EXT4_INODE_SIZE(inode->i_sb));
70
71 raw->i_checksum_lo = cpu_to_le16(csum_lo);
72 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
73 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
74 raw->i_checksum_hi = cpu_to_le16(csum_hi);
75
76 return csum;
77}
78
79static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
80 struct ext4_inode_info *ei)
81{
82 __u32 provided, calculated;
83
84 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
85 cpu_to_le32(EXT4_OS_LINUX) ||
86 !EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
87 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
88 return 1;
89
90 provided = le16_to_cpu(raw->i_checksum_lo);
91 calculated = ext4_inode_csum(inode, raw, ei);
92 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
93 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
94 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
95 else
96 calculated &= 0xFFFF;
97
98 return provided == calculated;
99}
100
101static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
102 struct ext4_inode_info *ei)
103{
104 __u32 csum;
105
106 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
107 cpu_to_le32(EXT4_OS_LINUX) ||
108 !EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
109 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
110 return;
111
112 csum = ext4_inode_csum(inode, raw, ei);
113 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
114 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
115 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
116 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
117}
118
119static inline int ext4_begin_ordered_truncate(struct inode *inode,
120 loff_t new_size)
121{
122 trace_ext4_begin_ordered_truncate(inode, new_size);
123 /*
124 * If jinode is zero, then we never opened the file for
125 * writing, so there's no need to call
126 * jbd2_journal_begin_ordered_truncate() since there's no
127 * outstanding writes we need to flush.
128 */
129 if (!EXT4_I(inode)->jinode)
130 return 0;
131 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
132 EXT4_I(inode)->jinode,
133 new_size);
134}
135
136static void ext4_invalidatepage(struct page *page, unsigned int offset,
137 unsigned int length);
138static int __ext4_journalled_writepage(struct page *page, unsigned int len);
139static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
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 */
146static int ext4_inode_is_fast_symlink(struct inode *inode)
147{
148 int ea_blocks = EXT4_I(inode)->i_file_acl ?
149 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
150
151 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
152}
153
154/*
155 * Restart the transaction associated with *handle. This does a commit,
156 * so before we call here everything must be consistently dirtied against
157 * this transaction.
158 */
159int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
160 int nblocks)
161{
162 int ret;
163
164 /*
165 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
166 * moment, get_block can be called only for blocks inside i_size since
167 * page cache has been already dropped and writes are blocked by
168 * i_mutex. So we can safely drop the i_data_sem here.
169 */
170 BUG_ON(EXT4_JOURNAL(inode) == NULL);
171 jbd_debug(2, "restarting handle %p\n", handle);
172 up_write(&EXT4_I(inode)->i_data_sem);
173 ret = ext4_journal_restart(handle, nblocks);
174 down_write(&EXT4_I(inode)->i_data_sem);
175 ext4_discard_preallocations(inode);
176
177 return ret;
178}
179
180/*
181 * Called at the last iput() if i_nlink is zero.
182 */
183void ext4_evict_inode(struct inode *inode)
184{
185 handle_t *handle;
186 int err;
187
188 trace_ext4_evict_inode(inode);
189
190 if (inode->i_nlink) {
191 /*
192 * When journalling data dirty buffers are tracked only in the
193 * journal. So although mm thinks everything is clean and
194 * ready for reaping the inode might still have some pages to
195 * write in the running transaction or waiting to be
196 * checkpointed. Thus calling jbd2_journal_invalidatepage()
197 * (via truncate_inode_pages()) to discard these buffers can
198 * cause data loss. Also even if we did not discard these
199 * buffers, we would have no way to find them after the inode
200 * is reaped and thus user could see stale data if he tries to
201 * read them before the transaction is checkpointed. So be
202 * careful and force everything to disk here... We use
203 * ei->i_datasync_tid to store the newest transaction
204 * containing inode's data.
205 *
206 * Note that directories do not have this problem because they
207 * don't use page cache.
208 */
209 if (ext4_should_journal_data(inode) &&
210 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
211 inode->i_ino != EXT4_JOURNAL_INO) {
212 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
213 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
214
215 jbd2_complete_transaction(journal, commit_tid);
216 filemap_write_and_wait(&inode->i_data);
217 }
218 truncate_inode_pages_final(&inode->i_data);
219
220 WARN_ON(atomic_read(&EXT4_I(inode)->i_ioend_count));
221 goto no_delete;
222 }
223
224 if (!is_bad_inode(inode))
225 dquot_initialize(inode);
226
227 if (ext4_should_order_data(inode))
228 ext4_begin_ordered_truncate(inode, 0);
229 truncate_inode_pages_final(&inode->i_data);
230
231 WARN_ON(atomic_read(&EXT4_I(inode)->i_ioend_count));
232 if (is_bad_inode(inode))
233 goto no_delete;
234
235 /*
236 * Protect us against freezing - iput() caller didn't have to have any
237 * protection against it
238 */
239 sb_start_intwrite(inode->i_sb);
240 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
241 ext4_blocks_for_truncate(inode)+3);
242 if (IS_ERR(handle)) {
243 ext4_std_error(inode->i_sb, PTR_ERR(handle));
244 /*
245 * If we're going to skip the normal cleanup, we still need to
246 * make sure that the in-core orphan linked list is properly
247 * cleaned up.
248 */
249 ext4_orphan_del(NULL, inode);
250 sb_end_intwrite(inode->i_sb);
251 goto no_delete;
252 }
253
254 if (IS_SYNC(inode))
255 ext4_handle_sync(handle);
256 inode->i_size = 0;
257 err = ext4_mark_inode_dirty(handle, inode);
258 if (err) {
259 ext4_warning(inode->i_sb,
260 "couldn't mark inode dirty (err %d)", err);
261 goto stop_handle;
262 }
263 if (inode->i_blocks)
264 ext4_truncate(inode);
265
266 /*
267 * ext4_ext_truncate() doesn't reserve any slop when it
268 * restarts journal transactions; therefore there may not be
269 * enough credits left in the handle to remove the inode from
270 * the orphan list and set the dtime field.
271 */
272 if (!ext4_handle_has_enough_credits(handle, 3)) {
273 err = ext4_journal_extend(handle, 3);
274 if (err > 0)
275 err = ext4_journal_restart(handle, 3);
276 if (err != 0) {
277 ext4_warning(inode->i_sb,
278 "couldn't extend journal (err %d)", err);
279 stop_handle:
280 ext4_journal_stop(handle);
281 ext4_orphan_del(NULL, inode);
282 sb_end_intwrite(inode->i_sb);
283 goto no_delete;
284 }
285 }
286
287 /*
288 * Kill off the orphan record which ext4_truncate created.
289 * AKPM: I think this can be inside the above `if'.
290 * Note that ext4_orphan_del() has to be able to cope with the
291 * deletion of a non-existent orphan - this is because we don't
292 * know if ext4_truncate() actually created an orphan record.
293 * (Well, we could do this if we need to, but heck - it works)
294 */
295 ext4_orphan_del(handle, inode);
296 EXT4_I(inode)->i_dtime = get_seconds();
297
298 /*
299 * One subtle ordering requirement: if anything has gone wrong
300 * (transaction abort, IO errors, whatever), then we can still
301 * do these next steps (the fs will already have been marked as
302 * having errors), but we can't free the inode if the mark_dirty
303 * fails.
304 */
305 if (ext4_mark_inode_dirty(handle, inode))
306 /* If that failed, just do the required in-core inode clear. */
307 ext4_clear_inode(inode);
308 else
309 ext4_free_inode(handle, inode);
310 ext4_journal_stop(handle);
311 sb_end_intwrite(inode->i_sb);
312 return;
313no_delete:
314 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
315}
316
317#ifdef CONFIG_QUOTA
318qsize_t *ext4_get_reserved_space(struct inode *inode)
319{
320 return &EXT4_I(inode)->i_reserved_quota;
321}
322#endif
323
324/*
325 * Calculate the number of metadata blocks need to reserve
326 * to allocate a block located at @lblock
327 */
328static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
329{
330 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
331 return ext4_ext_calc_metadata_amount(inode, lblock);
332
333 return ext4_ind_calc_metadata_amount(inode, lblock);
334}
335
336/*
337 * Called with i_data_sem down, which is important since we can call
338 * ext4_discard_preallocations() from here.
339 */
340void ext4_da_update_reserve_space(struct inode *inode,
341 int used, int quota_claim)
342{
343 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
344 struct ext4_inode_info *ei = EXT4_I(inode);
345
346 spin_lock(&ei->i_block_reservation_lock);
347 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
348 if (unlikely(used > ei->i_reserved_data_blocks)) {
349 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
350 "with only %d reserved data blocks",
351 __func__, inode->i_ino, used,
352 ei->i_reserved_data_blocks);
353 WARN_ON(1);
354 used = ei->i_reserved_data_blocks;
355 }
356
357 if (unlikely(ei->i_allocated_meta_blocks > ei->i_reserved_meta_blocks)) {
358 ext4_warning(inode->i_sb, "ino %lu, allocated %d "
359 "with only %d reserved metadata blocks "
360 "(releasing %d blocks with reserved %d data blocks)",
361 inode->i_ino, ei->i_allocated_meta_blocks,
362 ei->i_reserved_meta_blocks, used,
363 ei->i_reserved_data_blocks);
364 WARN_ON(1);
365 ei->i_allocated_meta_blocks = ei->i_reserved_meta_blocks;
366 }
367
368 /* Update per-inode reservations */
369 ei->i_reserved_data_blocks -= used;
370 ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
371 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
372 used + ei->i_allocated_meta_blocks);
373 ei->i_allocated_meta_blocks = 0;
374
375 if (ei->i_reserved_data_blocks == 0) {
376 /*
377 * We can release all of the reserved metadata blocks
378 * only when we have written all of the delayed
379 * allocation blocks.
380 */
381 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
382 ei->i_reserved_meta_blocks);
383 ei->i_reserved_meta_blocks = 0;
384 ei->i_da_metadata_calc_len = 0;
385 }
386 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
387
388 /* Update quota subsystem for data blocks */
389 if (quota_claim)
390 dquot_claim_block(inode, EXT4_C2B(sbi, used));
391 else {
392 /*
393 * We did fallocate with an offset that is already delayed
394 * allocated. So on delayed allocated writeback we should
395 * not re-claim the quota for fallocated blocks.
396 */
397 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
398 }
399
400 /*
401 * If we have done all the pending block allocations and if
402 * there aren't any writers on the inode, we can discard the
403 * inode's preallocations.
404 */
405 if ((ei->i_reserved_data_blocks == 0) &&
406 (atomic_read(&inode->i_writecount) == 0))
407 ext4_discard_preallocations(inode);
408}
409
410static int __check_block_validity(struct inode *inode, const char *func,
411 unsigned int line,
412 struct ext4_map_blocks *map)
413{
414 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
415 map->m_len)) {
416 ext4_error_inode(inode, func, line, map->m_pblk,
417 "lblock %lu mapped to illegal pblock "
418 "(length %d)", (unsigned long) map->m_lblk,
419 map->m_len);
420 return -EIO;
421 }
422 return 0;
423}
424
425#define check_block_validity(inode, map) \
426 __check_block_validity((inode), __func__, __LINE__, (map))
427
428#ifdef ES_AGGRESSIVE_TEST
429static void ext4_map_blocks_es_recheck(handle_t *handle,
430 struct inode *inode,
431 struct ext4_map_blocks *es_map,
432 struct ext4_map_blocks *map,
433 int flags)
434{
435 int retval;
436
437 map->m_flags = 0;
438 /*
439 * There is a race window that the result is not the same.
440 * e.g. xfstests #223 when dioread_nolock enables. The reason
441 * is that we lookup a block mapping in extent status tree with
442 * out taking i_data_sem. So at the time the unwritten extent
443 * could be converted.
444 */
445 if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
446 down_read((&EXT4_I(inode)->i_data_sem));
447 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
448 retval = ext4_ext_map_blocks(handle, inode, map, flags &
449 EXT4_GET_BLOCKS_KEEP_SIZE);
450 } else {
451 retval = ext4_ind_map_blocks(handle, inode, map, flags &
452 EXT4_GET_BLOCKS_KEEP_SIZE);
453 }
454 if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
455 up_read((&EXT4_I(inode)->i_data_sem));
456 /*
457 * Clear EXT4_MAP_FROM_CLUSTER and EXT4_MAP_BOUNDARY flag
458 * because it shouldn't be marked in es_map->m_flags.
459 */
460 map->m_flags &= ~(EXT4_MAP_FROM_CLUSTER | EXT4_MAP_BOUNDARY);
461
462 /*
463 * We don't check m_len because extent will be collpased in status
464 * tree. So the m_len might not equal.
465 */
466 if (es_map->m_lblk != map->m_lblk ||
467 es_map->m_flags != map->m_flags ||
468 es_map->m_pblk != map->m_pblk) {
469 printk("ES cache assertion failed for inode: %lu "
470 "es_cached ex [%d/%d/%llu/%x] != "
471 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
472 inode->i_ino, es_map->m_lblk, es_map->m_len,
473 es_map->m_pblk, es_map->m_flags, map->m_lblk,
474 map->m_len, map->m_pblk, map->m_flags,
475 retval, flags);
476 }
477}
478#endif /* ES_AGGRESSIVE_TEST */
479
480/*
481 * The ext4_map_blocks() function tries to look up the requested blocks,
482 * and returns if the blocks are already mapped.
483 *
484 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
485 * and store the allocated blocks in the result buffer head and mark it
486 * mapped.
487 *
488 * If file type is extents based, it will call ext4_ext_map_blocks(),
489 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
490 * based files
491 *
492 * On success, it returns the number of blocks being mapped or allocate.
493 * if create==0 and the blocks are pre-allocated and uninitialized block,
494 * the result buffer head is unmapped. If the create ==1, it will make sure
495 * the buffer head is mapped.
496 *
497 * It returns 0 if plain look up failed (blocks have not been allocated), in
498 * that case, buffer head is unmapped
499 *
500 * It returns the error in case of allocation failure.
501 */
502int ext4_map_blocks(handle_t *handle, struct inode *inode,
503 struct ext4_map_blocks *map, int flags)
504{
505 struct extent_status es;
506 int retval;
507 int ret = 0;
508#ifdef ES_AGGRESSIVE_TEST
509 struct ext4_map_blocks orig_map;
510
511 memcpy(&orig_map, map, sizeof(*map));
512#endif
513
514 map->m_flags = 0;
515 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
516 "logical block %lu\n", inode->i_ino, flags, map->m_len,
517 (unsigned long) map->m_lblk);
518
519 /*
520 * ext4_map_blocks returns an int, and m_len is an unsigned int
521 */
522 if (unlikely(map->m_len > INT_MAX))
523 map->m_len = INT_MAX;
524
525 /* We can handle the block number less than EXT_MAX_BLOCKS */
526 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
527 return -EIO;
528
529 /* Lookup extent status tree firstly */
530 if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
531 ext4_es_lru_add(inode);
532 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
533 map->m_pblk = ext4_es_pblock(&es) +
534 map->m_lblk - es.es_lblk;
535 map->m_flags |= ext4_es_is_written(&es) ?
536 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
537 retval = es.es_len - (map->m_lblk - es.es_lblk);
538 if (retval > map->m_len)
539 retval = map->m_len;
540 map->m_len = retval;
541 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
542 retval = 0;
543 } else {
544 BUG_ON(1);
545 }
546#ifdef ES_AGGRESSIVE_TEST
547 ext4_map_blocks_es_recheck(handle, inode, map,
548 &orig_map, flags);
549#endif
550 goto found;
551 }
552
553 /*
554 * Try to see if we can get the block without requesting a new
555 * file system block.
556 */
557 if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
558 down_read((&EXT4_I(inode)->i_data_sem));
559 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
560 retval = ext4_ext_map_blocks(handle, inode, map, flags &
561 EXT4_GET_BLOCKS_KEEP_SIZE);
562 } else {
563 retval = ext4_ind_map_blocks(handle, inode, map, flags &
564 EXT4_GET_BLOCKS_KEEP_SIZE);
565 }
566 if (retval > 0) {
567 unsigned int status;
568
569 if (unlikely(retval != map->m_len)) {
570 ext4_warning(inode->i_sb,
571 "ES len assertion failed for inode "
572 "%lu: retval %d != map->m_len %d",
573 inode->i_ino, retval, map->m_len);
574 WARN_ON(1);
575 }
576
577 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
578 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
579 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
580 ext4_find_delalloc_range(inode, map->m_lblk,
581 map->m_lblk + map->m_len - 1))
582 status |= EXTENT_STATUS_DELAYED;
583 ret = ext4_es_insert_extent(inode, map->m_lblk,
584 map->m_len, map->m_pblk, status);
585 if (ret < 0)
586 retval = ret;
587 }
588 if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
589 up_read((&EXT4_I(inode)->i_data_sem));
590
591found:
592 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
593 ret = check_block_validity(inode, map);
594 if (ret != 0)
595 return ret;
596 }
597
598 /* If it is only a block(s) look up */
599 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
600 return retval;
601
602 /*
603 * Returns if the blocks have already allocated
604 *
605 * Note that if blocks have been preallocated
606 * ext4_ext_get_block() returns the create = 0
607 * with buffer head unmapped.
608 */
609 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
610 /*
611 * If we need to convert extent to unwritten
612 * we continue and do the actual work in
613 * ext4_ext_map_blocks()
614 */
615 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
616 return retval;
617
618 /*
619 * Here we clear m_flags because after allocating an new extent,
620 * it will be set again.
621 */
622 map->m_flags &= ~EXT4_MAP_FLAGS;
623
624 /*
625 * New blocks allocate and/or writing to uninitialized extent
626 * will possibly result in updating i_data, so we take
627 * the write lock of i_data_sem, and call get_blocks()
628 * with create == 1 flag.
629 */
630 down_write((&EXT4_I(inode)->i_data_sem));
631
632 /*
633 * if the caller is from delayed allocation writeout path
634 * we have already reserved fs blocks for allocation
635 * let the underlying get_block() function know to
636 * avoid double accounting
637 */
638 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
639 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
640 /*
641 * We need to check for EXT4 here because migrate
642 * could have changed the inode type in between
643 */
644 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
645 retval = ext4_ext_map_blocks(handle, inode, map, flags);
646 } else {
647 retval = ext4_ind_map_blocks(handle, inode, map, flags);
648
649 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
650 /*
651 * We allocated new blocks which will result in
652 * i_data's format changing. Force the migrate
653 * to fail by clearing migrate flags
654 */
655 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
656 }
657
658 /*
659 * Update reserved blocks/metadata blocks after successful
660 * block allocation which had been deferred till now. We don't
661 * support fallocate for non extent files. So we can update
662 * reserve space here.
663 */
664 if ((retval > 0) &&
665 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
666 ext4_da_update_reserve_space(inode, retval, 1);
667 }
668 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
669 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
670
671 if (retval > 0) {
672 unsigned int status;
673
674 if (unlikely(retval != map->m_len)) {
675 ext4_warning(inode->i_sb,
676 "ES len assertion failed for inode "
677 "%lu: retval %d != map->m_len %d",
678 inode->i_ino, retval, map->m_len);
679 WARN_ON(1);
680 }
681
682 /*
683 * If the extent has been zeroed out, we don't need to update
684 * extent status tree.
685 */
686 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
687 ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
688 if (ext4_es_is_written(&es))
689 goto has_zeroout;
690 }
691 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
692 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
693 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
694 ext4_find_delalloc_range(inode, map->m_lblk,
695 map->m_lblk + map->m_len - 1))
696 status |= EXTENT_STATUS_DELAYED;
697 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
698 map->m_pblk, status);
699 if (ret < 0)
700 retval = ret;
701 }
702
703has_zeroout:
704 up_write((&EXT4_I(inode)->i_data_sem));
705 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
706 ret = check_block_validity(inode, map);
707 if (ret != 0)
708 return ret;
709 }
710 return retval;
711}
712
713/* Maximum number of blocks we map for direct IO at once. */
714#define DIO_MAX_BLOCKS 4096
715
716static int _ext4_get_block(struct inode *inode, sector_t iblock,
717 struct buffer_head *bh, int flags)
718{
719 handle_t *handle = ext4_journal_current_handle();
720 struct ext4_map_blocks map;
721 int ret = 0, started = 0;
722 int dio_credits;
723
724 if (ext4_has_inline_data(inode))
725 return -ERANGE;
726
727 map.m_lblk = iblock;
728 map.m_len = bh->b_size >> inode->i_blkbits;
729
730 if (flags && !(flags & EXT4_GET_BLOCKS_NO_LOCK) && !handle) {
731 /* Direct IO write... */
732 if (map.m_len > DIO_MAX_BLOCKS)
733 map.m_len = DIO_MAX_BLOCKS;
734 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
735 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
736 dio_credits);
737 if (IS_ERR(handle)) {
738 ret = PTR_ERR(handle);
739 return ret;
740 }
741 started = 1;
742 }
743
744 ret = ext4_map_blocks(handle, inode, &map, flags);
745 if (ret > 0) {
746 ext4_io_end_t *io_end = ext4_inode_aio(inode);
747
748 map_bh(bh, inode->i_sb, map.m_pblk);
749 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
750 if (io_end && io_end->flag & EXT4_IO_END_UNWRITTEN)
751 set_buffer_defer_completion(bh);
752 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
753 ret = 0;
754 }
755 if (started)
756 ext4_journal_stop(handle);
757 return ret;
758}
759
760int ext4_get_block(struct inode *inode, sector_t iblock,
761 struct buffer_head *bh, int create)
762{
763 return _ext4_get_block(inode, iblock, bh,
764 create ? EXT4_GET_BLOCKS_CREATE : 0);
765}
766
767/*
768 * `handle' can be NULL if create is zero
769 */
770struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
771 ext4_lblk_t block, int create, int *errp)
772{
773 struct ext4_map_blocks map;
774 struct buffer_head *bh;
775 int fatal = 0, err;
776
777 J_ASSERT(handle != NULL || create == 0);
778
779 map.m_lblk = block;
780 map.m_len = 1;
781 err = ext4_map_blocks(handle, inode, &map,
782 create ? EXT4_GET_BLOCKS_CREATE : 0);
783
784 /* ensure we send some value back into *errp */
785 *errp = 0;
786
787 if (create && err == 0)
788 err = -ENOSPC; /* should never happen */
789 if (err < 0)
790 *errp = err;
791 if (err <= 0)
792 return NULL;
793
794 bh = sb_getblk(inode->i_sb, map.m_pblk);
795 if (unlikely(!bh)) {
796 *errp = -ENOMEM;
797 return NULL;
798 }
799 if (map.m_flags & EXT4_MAP_NEW) {
800 J_ASSERT(create != 0);
801 J_ASSERT(handle != NULL);
802
803 /*
804 * Now that we do not always journal data, we should
805 * keep in mind whether this should always journal the
806 * new buffer as metadata. For now, regular file
807 * writes use ext4_get_block instead, so it's not a
808 * problem.
809 */
810 lock_buffer(bh);
811 BUFFER_TRACE(bh, "call get_create_access");
812 fatal = ext4_journal_get_create_access(handle, bh);
813 if (!fatal && !buffer_uptodate(bh)) {
814 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
815 set_buffer_uptodate(bh);
816 }
817 unlock_buffer(bh);
818 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
819 err = ext4_handle_dirty_metadata(handle, inode, bh);
820 if (!fatal)
821 fatal = err;
822 } else {
823 BUFFER_TRACE(bh, "not a new buffer");
824 }
825 if (fatal) {
826 *errp = fatal;
827 brelse(bh);
828 bh = NULL;
829 }
830 return bh;
831}
832
833struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
834 ext4_lblk_t block, int create, int *err)
835{
836 struct buffer_head *bh;
837
838 bh = ext4_getblk(handle, inode, block, create, err);
839 if (!bh)
840 return bh;
841 if (buffer_uptodate(bh))
842 return bh;
843 ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
844 wait_on_buffer(bh);
845 if (buffer_uptodate(bh))
846 return bh;
847 put_bh(bh);
848 *err = -EIO;
849 return NULL;
850}
851
852int ext4_walk_page_buffers(handle_t *handle,
853 struct buffer_head *head,
854 unsigned from,
855 unsigned to,
856 int *partial,
857 int (*fn)(handle_t *handle,
858 struct buffer_head *bh))
859{
860 struct buffer_head *bh;
861 unsigned block_start, block_end;
862 unsigned blocksize = head->b_size;
863 int err, ret = 0;
864 struct buffer_head *next;
865
866 for (bh = head, block_start = 0;
867 ret == 0 && (bh != head || !block_start);
868 block_start = block_end, bh = next) {
869 next = bh->b_this_page;
870 block_end = block_start + blocksize;
871 if (block_end <= from || block_start >= to) {
872 if (partial && !buffer_uptodate(bh))
873 *partial = 1;
874 continue;
875 }
876 err = (*fn)(handle, bh);
877 if (!ret)
878 ret = err;
879 }
880 return ret;
881}
882
883/*
884 * To preserve ordering, it is essential that the hole instantiation and
885 * the data write be encapsulated in a single transaction. We cannot
886 * close off a transaction and start a new one between the ext4_get_block()
887 * and the commit_write(). So doing the jbd2_journal_start at the start of
888 * prepare_write() is the right place.
889 *
890 * Also, this function can nest inside ext4_writepage(). In that case, we
891 * *know* that ext4_writepage() has generated enough buffer credits to do the
892 * whole page. So we won't block on the journal in that case, which is good,
893 * because the caller may be PF_MEMALLOC.
894 *
895 * By accident, ext4 can be reentered when a transaction is open via
896 * quota file writes. If we were to commit the transaction while thus
897 * reentered, there can be a deadlock - we would be holding a quota
898 * lock, and the commit would never complete if another thread had a
899 * transaction open and was blocking on the quota lock - a ranking
900 * violation.
901 *
902 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
903 * will _not_ run commit under these circumstances because handle->h_ref
904 * is elevated. We'll still have enough credits for the tiny quotafile
905 * write.
906 */
907int do_journal_get_write_access(handle_t *handle,
908 struct buffer_head *bh)
909{
910 int dirty = buffer_dirty(bh);
911 int ret;
912
913 if (!buffer_mapped(bh) || buffer_freed(bh))
914 return 0;
915 /*
916 * __block_write_begin() could have dirtied some buffers. Clean
917 * the dirty bit as jbd2_journal_get_write_access() could complain
918 * otherwise about fs integrity issues. Setting of the dirty bit
919 * by __block_write_begin() isn't a real problem here as we clear
920 * the bit before releasing a page lock and thus writeback cannot
921 * ever write the buffer.
922 */
923 if (dirty)
924 clear_buffer_dirty(bh);
925 ret = ext4_journal_get_write_access(handle, bh);
926 if (!ret && dirty)
927 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
928 return ret;
929}
930
931static int ext4_get_block_write_nolock(struct inode *inode, sector_t iblock,
932 struct buffer_head *bh_result, int create);
933static int ext4_write_begin(struct file *file, struct address_space *mapping,
934 loff_t pos, unsigned len, unsigned flags,
935 struct page **pagep, void **fsdata)
936{
937 struct inode *inode = mapping->host;
938 int ret, needed_blocks;
939 handle_t *handle;
940 int retries = 0;
941 struct page *page;
942 pgoff_t index;
943 unsigned from, to;
944
945 trace_ext4_write_begin(inode, pos, len, flags);
946 /*
947 * Reserve one block more for addition to orphan list in case
948 * we allocate blocks but write fails for some reason
949 */
950 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
951 index = pos >> PAGE_CACHE_SHIFT;
952 from = pos & (PAGE_CACHE_SIZE - 1);
953 to = from + len;
954
955 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
956 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
957 flags, pagep);
958 if (ret < 0)
959 return ret;
960 if (ret == 1)
961 return 0;
962 }
963
964 /*
965 * grab_cache_page_write_begin() can take a long time if the
966 * system is thrashing due to memory pressure, or if the page
967 * is being written back. So grab it first before we start
968 * the transaction handle. This also allows us to allocate
969 * the page (if needed) without using GFP_NOFS.
970 */
971retry_grab:
972 page = grab_cache_page_write_begin(mapping, index, flags);
973 if (!page)
974 return -ENOMEM;
975 unlock_page(page);
976
977retry_journal:
978 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
979 if (IS_ERR(handle)) {
980 page_cache_release(page);
981 return PTR_ERR(handle);
982 }
983
984 lock_page(page);
985 if (page->mapping != mapping) {
986 /* The page got truncated from under us */
987 unlock_page(page);
988 page_cache_release(page);
989 ext4_journal_stop(handle);
990 goto retry_grab;
991 }
992 /* In case writeback began while the page was unlocked */
993 wait_for_stable_page(page);
994
995 if (ext4_should_dioread_nolock(inode))
996 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
997 else
998 ret = __block_write_begin(page, pos, len, ext4_get_block);
999
1000 if (!ret && ext4_should_journal_data(inode)) {
1001 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1002 from, to, NULL,
1003 do_journal_get_write_access);
1004 }
1005
1006 if (ret) {
1007 unlock_page(page);
1008 /*
1009 * __block_write_begin may have instantiated a few blocks
1010 * outside i_size. Trim these off again. Don't need
1011 * i_size_read because we hold i_mutex.
1012 *
1013 * Add inode to orphan list in case we crash before
1014 * truncate finishes
1015 */
1016 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1017 ext4_orphan_add(handle, inode);
1018
1019 ext4_journal_stop(handle);
1020 if (pos + len > inode->i_size) {
1021 ext4_truncate_failed_write(inode);
1022 /*
1023 * If truncate failed early the inode might
1024 * still be on the orphan list; we need to
1025 * make sure the inode is removed from the
1026 * orphan list in that case.
1027 */
1028 if (inode->i_nlink)
1029 ext4_orphan_del(NULL, inode);
1030 }
1031
1032 if (ret == -ENOSPC &&
1033 ext4_should_retry_alloc(inode->i_sb, &retries))
1034 goto retry_journal;
1035 page_cache_release(page);
1036 return ret;
1037 }
1038 *pagep = page;
1039 return ret;
1040}
1041
1042/* For write_end() in data=journal mode */
1043static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1044{
1045 int ret;
1046 if (!buffer_mapped(bh) || buffer_freed(bh))
1047 return 0;
1048 set_buffer_uptodate(bh);
1049 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1050 clear_buffer_meta(bh);
1051 clear_buffer_prio(bh);
1052 return ret;
1053}
1054
1055/*
1056 * We need to pick up the new inode size which generic_commit_write gave us
1057 * `file' can be NULL - eg, when called from page_symlink().
1058 *
1059 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1060 * buffers are managed internally.
1061 */
1062static int ext4_write_end(struct file *file,
1063 struct address_space *mapping,
1064 loff_t pos, unsigned len, unsigned copied,
1065 struct page *page, void *fsdata)
1066{
1067 handle_t *handle = ext4_journal_current_handle();
1068 struct inode *inode = mapping->host;
1069 int ret = 0, ret2;
1070 int i_size_changed = 0;
1071
1072 trace_ext4_write_end(inode, pos, len, copied);
1073 if (ext4_test_inode_state(inode, EXT4_STATE_ORDERED_MODE)) {
1074 ret = ext4_jbd2_file_inode(handle, inode);
1075 if (ret) {
1076 unlock_page(page);
1077 page_cache_release(page);
1078 goto errout;
1079 }
1080 }
1081
1082 if (ext4_has_inline_data(inode)) {
1083 ret = ext4_write_inline_data_end(inode, pos, len,
1084 copied, page);
1085 if (ret < 0)
1086 goto errout;
1087 copied = ret;
1088 } else
1089 copied = block_write_end(file, mapping, pos,
1090 len, copied, page, fsdata);
1091
1092 /*
1093 * No need to use i_size_read() here, the i_size
1094 * cannot change under us because we hole i_mutex.
1095 *
1096 * But it's important to update i_size while still holding page lock:
1097 * page writeout could otherwise come in and zero beyond i_size.
1098 */
1099 if (pos + copied > inode->i_size) {
1100 i_size_write(inode, pos + copied);
1101 i_size_changed = 1;
1102 }
1103
1104 if (pos + copied > EXT4_I(inode)->i_disksize) {
1105 /* We need to mark inode dirty even if
1106 * new_i_size is less that inode->i_size
1107 * but greater than i_disksize. (hint delalloc)
1108 */
1109 ext4_update_i_disksize(inode, (pos + copied));
1110 i_size_changed = 1;
1111 }
1112 unlock_page(page);
1113 page_cache_release(page);
1114
1115 /*
1116 * Don't mark the inode dirty under page lock. First, it unnecessarily
1117 * makes the holding time of page lock longer. Second, it forces lock
1118 * ordering of page lock and transaction start for journaling
1119 * filesystems.
1120 */
1121 if (i_size_changed)
1122 ext4_mark_inode_dirty(handle, inode);
1123
1124 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1125 /* if we have allocated more blocks and copied
1126 * less. We will have blocks allocated outside
1127 * inode->i_size. So truncate them
1128 */
1129 ext4_orphan_add(handle, inode);
1130errout:
1131 ret2 = ext4_journal_stop(handle);
1132 if (!ret)
1133 ret = ret2;
1134
1135 if (pos + len > inode->i_size) {
1136 ext4_truncate_failed_write(inode);
1137 /*
1138 * If truncate failed early the inode might still be
1139 * on the orphan list; we need to make sure the inode
1140 * is removed from the orphan list in that case.
1141 */
1142 if (inode->i_nlink)
1143 ext4_orphan_del(NULL, inode);
1144 }
1145
1146 return ret ? ret : copied;
1147}
1148
1149static int ext4_journalled_write_end(struct file *file,
1150 struct address_space *mapping,
1151 loff_t pos, unsigned len, unsigned copied,
1152 struct page *page, void *fsdata)
1153{
1154 handle_t *handle = ext4_journal_current_handle();
1155 struct inode *inode = mapping->host;
1156 int ret = 0, ret2;
1157 int partial = 0;
1158 unsigned from, to;
1159 loff_t new_i_size;
1160
1161 trace_ext4_journalled_write_end(inode, pos, len, copied);
1162 from = pos & (PAGE_CACHE_SIZE - 1);
1163 to = from + len;
1164
1165 BUG_ON(!ext4_handle_valid(handle));
1166
1167 if (ext4_has_inline_data(inode))
1168 copied = ext4_write_inline_data_end(inode, pos, len,
1169 copied, page);
1170 else {
1171 if (copied < len) {
1172 if (!PageUptodate(page))
1173 copied = 0;
1174 page_zero_new_buffers(page, from+copied, to);
1175 }
1176
1177 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1178 to, &partial, write_end_fn);
1179 if (!partial)
1180 SetPageUptodate(page);
1181 }
1182 new_i_size = pos + copied;
1183 if (new_i_size > inode->i_size)
1184 i_size_write(inode, pos+copied);
1185 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1186 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1187 if (new_i_size > EXT4_I(inode)->i_disksize) {
1188 ext4_update_i_disksize(inode, new_i_size);
1189 ret2 = ext4_mark_inode_dirty(handle, inode);
1190 if (!ret)
1191 ret = ret2;
1192 }
1193
1194 unlock_page(page);
1195 page_cache_release(page);
1196 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1197 /* if we have allocated more blocks and copied
1198 * less. We will have blocks allocated outside
1199 * inode->i_size. So truncate them
1200 */
1201 ext4_orphan_add(handle, inode);
1202
1203 ret2 = ext4_journal_stop(handle);
1204 if (!ret)
1205 ret = ret2;
1206 if (pos + len > inode->i_size) {
1207 ext4_truncate_failed_write(inode);
1208 /*
1209 * If truncate failed early the inode might still be
1210 * on the orphan list; we need to make sure the inode
1211 * is removed from the orphan list in that case.
1212 */
1213 if (inode->i_nlink)
1214 ext4_orphan_del(NULL, inode);
1215 }
1216
1217 return ret ? ret : copied;
1218}
1219
1220/*
1221 * Reserve a metadata for a single block located at lblock
1222 */
1223static int ext4_da_reserve_metadata(struct inode *inode, ext4_lblk_t lblock)
1224{
1225 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1226 struct ext4_inode_info *ei = EXT4_I(inode);
1227 unsigned int md_needed;
1228 ext4_lblk_t save_last_lblock;
1229 int save_len;
1230
1231 /*
1232 * recalculate the amount of metadata blocks to reserve
1233 * in order to allocate nrblocks
1234 * worse case is one extent per block
1235 */
1236 spin_lock(&ei->i_block_reservation_lock);
1237 /*
1238 * ext4_calc_metadata_amount() has side effects, which we have
1239 * to be prepared undo if we fail to claim space.
1240 */
1241 save_len = ei->i_da_metadata_calc_len;
1242 save_last_lblock = ei->i_da_metadata_calc_last_lblock;
1243 md_needed = EXT4_NUM_B2C(sbi,
1244 ext4_calc_metadata_amount(inode, lblock));
1245 trace_ext4_da_reserve_space(inode, md_needed);
1246
1247 /*
1248 * We do still charge estimated metadata to the sb though;
1249 * we cannot afford to run out of free blocks.
1250 */
1251 if (ext4_claim_free_clusters(sbi, md_needed, 0)) {
1252 ei->i_da_metadata_calc_len = save_len;
1253 ei->i_da_metadata_calc_last_lblock = save_last_lblock;
1254 spin_unlock(&ei->i_block_reservation_lock);
1255 return -ENOSPC;
1256 }
1257 ei->i_reserved_meta_blocks += md_needed;
1258 spin_unlock(&ei->i_block_reservation_lock);
1259
1260 return 0; /* success */
1261}
1262
1263/*
1264 * Reserve a single cluster located at lblock
1265 */
1266static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1267{
1268 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1269 struct ext4_inode_info *ei = EXT4_I(inode);
1270 unsigned int md_needed;
1271 int ret;
1272 ext4_lblk_t save_last_lblock;
1273 int save_len;
1274
1275 /*
1276 * We will charge metadata quota at writeout time; this saves
1277 * us from metadata over-estimation, though we may go over by
1278 * a small amount in the end. Here we just reserve for data.
1279 */
1280 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1281 if (ret)
1282 return ret;
1283
1284 /*
1285 * recalculate the amount of metadata blocks to reserve
1286 * in order to allocate nrblocks
1287 * worse case is one extent per block
1288 */
1289 spin_lock(&ei->i_block_reservation_lock);
1290 /*
1291 * ext4_calc_metadata_amount() has side effects, which we have
1292 * to be prepared undo if we fail to claim space.
1293 */
1294 save_len = ei->i_da_metadata_calc_len;
1295 save_last_lblock = ei->i_da_metadata_calc_last_lblock;
1296 md_needed = EXT4_NUM_B2C(sbi,
1297 ext4_calc_metadata_amount(inode, lblock));
1298 trace_ext4_da_reserve_space(inode, md_needed);
1299
1300 /*
1301 * We do still charge estimated metadata to the sb though;
1302 * we cannot afford to run out of free blocks.
1303 */
1304 if (ext4_claim_free_clusters(sbi, md_needed + 1, 0)) {
1305 ei->i_da_metadata_calc_len = save_len;
1306 ei->i_da_metadata_calc_last_lblock = save_last_lblock;
1307 spin_unlock(&ei->i_block_reservation_lock);
1308 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1309 return -ENOSPC;
1310 }
1311 ei->i_reserved_data_blocks++;
1312 ei->i_reserved_meta_blocks += md_needed;
1313 spin_unlock(&ei->i_block_reservation_lock);
1314
1315 return 0; /* success */
1316}
1317
1318static void ext4_da_release_space(struct inode *inode, int to_free)
1319{
1320 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1321 struct ext4_inode_info *ei = EXT4_I(inode);
1322
1323 if (!to_free)
1324 return; /* Nothing to release, exit */
1325
1326 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1327
1328 trace_ext4_da_release_space(inode, to_free);
1329 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1330 /*
1331 * if there aren't enough reserved blocks, then the
1332 * counter is messed up somewhere. Since this
1333 * function is called from invalidate page, it's
1334 * harmless to return without any action.
1335 */
1336 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1337 "ino %lu, to_free %d with only %d reserved "
1338 "data blocks", inode->i_ino, to_free,
1339 ei->i_reserved_data_blocks);
1340 WARN_ON(1);
1341 to_free = ei->i_reserved_data_blocks;
1342 }
1343 ei->i_reserved_data_blocks -= to_free;
1344
1345 if (ei->i_reserved_data_blocks == 0) {
1346 /*
1347 * We can release all of the reserved metadata blocks
1348 * only when we have written all of the delayed
1349 * allocation blocks.
1350 * Note that in case of bigalloc, i_reserved_meta_blocks,
1351 * i_reserved_data_blocks, etc. refer to number of clusters.
1352 */
1353 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
1354 ei->i_reserved_meta_blocks);
1355 ei->i_reserved_meta_blocks = 0;
1356 ei->i_da_metadata_calc_len = 0;
1357 }
1358
1359 /* update fs dirty data blocks counter */
1360 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1361
1362 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1363
1364 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1365}
1366
1367static void ext4_da_page_release_reservation(struct page *page,
1368 unsigned int offset,
1369 unsigned int length)
1370{
1371 int to_release = 0;
1372 struct buffer_head *head, *bh;
1373 unsigned int curr_off = 0;
1374 struct inode *inode = page->mapping->host;
1375 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1376 unsigned int stop = offset + length;
1377 int num_clusters;
1378 ext4_fsblk_t lblk;
1379
1380 BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1381
1382 head = page_buffers(page);
1383 bh = head;
1384 do {
1385 unsigned int next_off = curr_off + bh->b_size;
1386
1387 if (next_off > stop)
1388 break;
1389
1390 if ((offset <= curr_off) && (buffer_delay(bh))) {
1391 to_release++;
1392 clear_buffer_delay(bh);
1393 }
1394 curr_off = next_off;
1395 } while ((bh = bh->b_this_page) != head);
1396
1397 if (to_release) {
1398 lblk = page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1399 ext4_es_remove_extent(inode, lblk, to_release);
1400 }
1401
1402 /* If we have released all the blocks belonging to a cluster, then we
1403 * need to release the reserved space for that cluster. */
1404 num_clusters = EXT4_NUM_B2C(sbi, to_release);
1405 while (num_clusters > 0) {
1406 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1407 ((num_clusters - 1) << sbi->s_cluster_bits);
1408 if (sbi->s_cluster_ratio == 1 ||
1409 !ext4_find_delalloc_cluster(inode, lblk))
1410 ext4_da_release_space(inode, 1);
1411
1412 num_clusters--;
1413 }
1414}
1415
1416/*
1417 * Delayed allocation stuff
1418 */
1419
1420struct mpage_da_data {
1421 struct inode *inode;
1422 struct writeback_control *wbc;
1423
1424 pgoff_t first_page; /* The first page to write */
1425 pgoff_t next_page; /* Current page to examine */
1426 pgoff_t last_page; /* Last page to examine */
1427 /*
1428 * Extent to map - this can be after first_page because that can be
1429 * fully mapped. We somewhat abuse m_flags to store whether the extent
1430 * is delalloc or unwritten.
1431 */
1432 struct ext4_map_blocks map;
1433 struct ext4_io_submit io_submit; /* IO submission data */
1434};
1435
1436static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1437 bool invalidate)
1438{
1439 int nr_pages, i;
1440 pgoff_t index, end;
1441 struct pagevec pvec;
1442 struct inode *inode = mpd->inode;
1443 struct address_space *mapping = inode->i_mapping;
1444
1445 /* This is necessary when next_page == 0. */
1446 if (mpd->first_page >= mpd->next_page)
1447 return;
1448
1449 index = mpd->first_page;
1450 end = mpd->next_page - 1;
1451 if (invalidate) {
1452 ext4_lblk_t start, last;
1453 start = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1454 last = end << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1455 ext4_es_remove_extent(inode, start, last - start + 1);
1456 }
1457
1458 pagevec_init(&pvec, 0);
1459 while (index <= end) {
1460 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1461 if (nr_pages == 0)
1462 break;
1463 for (i = 0; i < nr_pages; i++) {
1464 struct page *page = pvec.pages[i];
1465 if (page->index > end)
1466 break;
1467 BUG_ON(!PageLocked(page));
1468 BUG_ON(PageWriteback(page));
1469 if (invalidate) {
1470 block_invalidatepage(page, 0, PAGE_CACHE_SIZE);
1471 ClearPageUptodate(page);
1472 }
1473 unlock_page(page);
1474 }
1475 index = pvec.pages[nr_pages - 1]->index + 1;
1476 pagevec_release(&pvec);
1477 }
1478}
1479
1480static void ext4_print_free_blocks(struct inode *inode)
1481{
1482 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1483 struct super_block *sb = inode->i_sb;
1484 struct ext4_inode_info *ei = EXT4_I(inode);
1485
1486 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1487 EXT4_C2B(EXT4_SB(inode->i_sb),
1488 ext4_count_free_clusters(sb)));
1489 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1490 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1491 (long long) EXT4_C2B(EXT4_SB(sb),
1492 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1493 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1494 (long long) EXT4_C2B(EXT4_SB(sb),
1495 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1496 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1497 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1498 ei->i_reserved_data_blocks);
1499 ext4_msg(sb, KERN_CRIT, "i_reserved_meta_blocks=%u",
1500 ei->i_reserved_meta_blocks);
1501 ext4_msg(sb, KERN_CRIT, "i_allocated_meta_blocks=%u",
1502 ei->i_allocated_meta_blocks);
1503 return;
1504}
1505
1506static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1507{
1508 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1509}
1510
1511/*
1512 * This function is grabs code from the very beginning of
1513 * ext4_map_blocks, but assumes that the caller is from delayed write
1514 * time. This function looks up the requested blocks and sets the
1515 * buffer delay bit under the protection of i_data_sem.
1516 */
1517static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1518 struct ext4_map_blocks *map,
1519 struct buffer_head *bh)
1520{
1521 struct extent_status es;
1522 int retval;
1523 sector_t invalid_block = ~((sector_t) 0xffff);
1524#ifdef ES_AGGRESSIVE_TEST
1525 struct ext4_map_blocks orig_map;
1526
1527 memcpy(&orig_map, map, sizeof(*map));
1528#endif
1529
1530 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1531 invalid_block = ~0;
1532
1533 map->m_flags = 0;
1534 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1535 "logical block %lu\n", inode->i_ino, map->m_len,
1536 (unsigned long) map->m_lblk);
1537
1538 /* Lookup extent status tree firstly */
1539 if (ext4_es_lookup_extent(inode, iblock, &es)) {
1540 ext4_es_lru_add(inode);
1541 if (ext4_es_is_hole(&es)) {
1542 retval = 0;
1543 down_read((&EXT4_I(inode)->i_data_sem));
1544 goto add_delayed;
1545 }
1546
1547 /*
1548 * Delayed extent could be allocated by fallocate.
1549 * So we need to check it.
1550 */
1551 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1552 map_bh(bh, inode->i_sb, invalid_block);
1553 set_buffer_new(bh);
1554 set_buffer_delay(bh);
1555 return 0;
1556 }
1557
1558 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1559 retval = es.es_len - (iblock - es.es_lblk);
1560 if (retval > map->m_len)
1561 retval = map->m_len;
1562 map->m_len = retval;
1563 if (ext4_es_is_written(&es))
1564 map->m_flags |= EXT4_MAP_MAPPED;
1565 else if (ext4_es_is_unwritten(&es))
1566 map->m_flags |= EXT4_MAP_UNWRITTEN;
1567 else
1568 BUG_ON(1);
1569
1570#ifdef ES_AGGRESSIVE_TEST
1571 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1572#endif
1573 return retval;
1574 }
1575
1576 /*
1577 * Try to see if we can get the block without requesting a new
1578 * file system block.
1579 */
1580 down_read((&EXT4_I(inode)->i_data_sem));
1581 if (ext4_has_inline_data(inode)) {
1582 /*
1583 * We will soon create blocks for this page, and let
1584 * us pretend as if the blocks aren't allocated yet.
1585 * In case of clusters, we have to handle the work
1586 * of mapping from cluster so that the reserved space
1587 * is calculated properly.
1588 */
1589 if ((EXT4_SB(inode->i_sb)->s_cluster_ratio > 1) &&
1590 ext4_find_delalloc_cluster(inode, map->m_lblk))
1591 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
1592 retval = 0;
1593 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1594 retval = ext4_ext_map_blocks(NULL, inode, map,
1595 EXT4_GET_BLOCKS_NO_PUT_HOLE);
1596 else
1597 retval = ext4_ind_map_blocks(NULL, inode, map,
1598 EXT4_GET_BLOCKS_NO_PUT_HOLE);
1599
1600add_delayed:
1601 if (retval == 0) {
1602 int ret;
1603 /*
1604 * XXX: __block_prepare_write() unmaps passed block,
1605 * is it OK?
1606 */
1607 /*
1608 * If the block was allocated from previously allocated cluster,
1609 * then we don't need to reserve it again. However we still need
1610 * to reserve metadata for every block we're going to write.
1611 */
1612 if (!(map->m_flags & EXT4_MAP_FROM_CLUSTER)) {
1613 ret = ext4_da_reserve_space(inode, iblock);
1614 if (ret) {
1615 /* not enough space to reserve */
1616 retval = ret;
1617 goto out_unlock;
1618 }
1619 } else {
1620 ret = ext4_da_reserve_metadata(inode, iblock);
1621 if (ret) {
1622 /* not enough space to reserve */
1623 retval = ret;
1624 goto out_unlock;
1625 }
1626 }
1627
1628 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1629 ~0, EXTENT_STATUS_DELAYED);
1630 if (ret) {
1631 retval = ret;
1632 goto out_unlock;
1633 }
1634
1635 /* Clear EXT4_MAP_FROM_CLUSTER flag since its purpose is served
1636 * and it should not appear on the bh->b_state.
1637 */
1638 map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
1639
1640 map_bh(bh, inode->i_sb, invalid_block);
1641 set_buffer_new(bh);
1642 set_buffer_delay(bh);
1643 } else if (retval > 0) {
1644 int ret;
1645 unsigned int status;
1646
1647 if (unlikely(retval != map->m_len)) {
1648 ext4_warning(inode->i_sb,
1649 "ES len assertion failed for inode "
1650 "%lu: retval %d != map->m_len %d",
1651 inode->i_ino, retval, map->m_len);
1652 WARN_ON(1);
1653 }
1654
1655 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1656 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1657 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1658 map->m_pblk, status);
1659 if (ret != 0)
1660 retval = ret;
1661 }
1662
1663out_unlock:
1664 up_read((&EXT4_I(inode)->i_data_sem));
1665
1666 return retval;
1667}
1668
1669/*
1670 * This is a special get_blocks_t callback which is used by
1671 * ext4_da_write_begin(). It will either return mapped block or
1672 * reserve space for a single block.
1673 *
1674 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1675 * We also have b_blocknr = -1 and b_bdev initialized properly
1676 *
1677 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1678 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1679 * initialized properly.
1680 */
1681int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1682 struct buffer_head *bh, int create)
1683{
1684 struct ext4_map_blocks map;
1685 int ret = 0;
1686
1687 BUG_ON(create == 0);
1688 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1689
1690 map.m_lblk = iblock;
1691 map.m_len = 1;
1692
1693 /*
1694 * first, we need to know whether the block is allocated already
1695 * preallocated blocks are unmapped but should treated
1696 * the same as allocated blocks.
1697 */
1698 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1699 if (ret <= 0)
1700 return ret;
1701
1702 map_bh(bh, inode->i_sb, map.m_pblk);
1703 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1704
1705 if (buffer_unwritten(bh)) {
1706 /* A delayed write to unwritten bh should be marked
1707 * new and mapped. Mapped ensures that we don't do
1708 * get_block multiple times when we write to the same
1709 * offset and new ensures that we do proper zero out
1710 * for partial write.
1711 */
1712 set_buffer_new(bh);
1713 set_buffer_mapped(bh);
1714 }
1715 return 0;
1716}
1717
1718static int bget_one(handle_t *handle, struct buffer_head *bh)
1719{
1720 get_bh(bh);
1721 return 0;
1722}
1723
1724static int bput_one(handle_t *handle, struct buffer_head *bh)
1725{
1726 put_bh(bh);
1727 return 0;
1728}
1729
1730static int __ext4_journalled_writepage(struct page *page,
1731 unsigned int len)
1732{
1733 struct address_space *mapping = page->mapping;
1734 struct inode *inode = mapping->host;
1735 struct buffer_head *page_bufs = NULL;
1736 handle_t *handle = NULL;
1737 int ret = 0, err = 0;
1738 int inline_data = ext4_has_inline_data(inode);
1739 struct buffer_head *inode_bh = NULL;
1740
1741 ClearPageChecked(page);
1742
1743 if (inline_data) {
1744 BUG_ON(page->index != 0);
1745 BUG_ON(len > ext4_get_max_inline_size(inode));
1746 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1747 if (inode_bh == NULL)
1748 goto out;
1749 } else {
1750 page_bufs = page_buffers(page);
1751 if (!page_bufs) {
1752 BUG();
1753 goto out;
1754 }
1755 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1756 NULL, bget_one);
1757 }
1758 /* As soon as we unlock the page, it can go away, but we have
1759 * references to buffers so we are safe */
1760 unlock_page(page);
1761
1762 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1763 ext4_writepage_trans_blocks(inode));
1764 if (IS_ERR(handle)) {
1765 ret = PTR_ERR(handle);
1766 goto out;
1767 }
1768
1769 BUG_ON(!ext4_handle_valid(handle));
1770
1771 if (inline_data) {
1772 ret = ext4_journal_get_write_access(handle, inode_bh);
1773
1774 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1775
1776 } else {
1777 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1778 do_journal_get_write_access);
1779
1780 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1781 write_end_fn);
1782 }
1783 if (ret == 0)
1784 ret = err;
1785 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1786 err = ext4_journal_stop(handle);
1787 if (!ret)
1788 ret = err;
1789
1790 if (!ext4_has_inline_data(inode))
1791 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1792 NULL, bput_one);
1793 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1794out:
1795 brelse(inode_bh);
1796 return ret;
1797}
1798
1799/*
1800 * Note that we don't need to start a transaction unless we're journaling data
1801 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1802 * need to file the inode to the transaction's list in ordered mode because if
1803 * we are writing back data added by write(), the inode is already there and if
1804 * we are writing back data modified via mmap(), no one guarantees in which
1805 * transaction the data will hit the disk. In case we are journaling data, we
1806 * cannot start transaction directly because transaction start ranks above page
1807 * lock so we have to do some magic.
1808 *
1809 * This function can get called via...
1810 * - ext4_writepages after taking page lock (have journal handle)
1811 * - journal_submit_inode_data_buffers (no journal handle)
1812 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1813 * - grab_page_cache when doing write_begin (have journal handle)
1814 *
1815 * We don't do any block allocation in this function. If we have page with
1816 * multiple blocks we need to write those buffer_heads that are mapped. This
1817 * is important for mmaped based write. So if we do with blocksize 1K
1818 * truncate(f, 1024);
1819 * a = mmap(f, 0, 4096);
1820 * a[0] = 'a';
1821 * truncate(f, 4096);
1822 * we have in the page first buffer_head mapped via page_mkwrite call back
1823 * but other buffer_heads would be unmapped but dirty (dirty done via the
1824 * do_wp_page). So writepage should write the first block. If we modify
1825 * the mmap area beyond 1024 we will again get a page_fault and the
1826 * page_mkwrite callback will do the block allocation and mark the
1827 * buffer_heads mapped.
1828 *
1829 * We redirty the page if we have any buffer_heads that is either delay or
1830 * unwritten in the page.
1831 *
1832 * We can get recursively called as show below.
1833 *
1834 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1835 * ext4_writepage()
1836 *
1837 * But since we don't do any block allocation we should not deadlock.
1838 * Page also have the dirty flag cleared so we don't get recurive page_lock.
1839 */
1840static int ext4_writepage(struct page *page,
1841 struct writeback_control *wbc)
1842{
1843 int ret = 0;
1844 loff_t size;
1845 unsigned int len;
1846 struct buffer_head *page_bufs = NULL;
1847 struct inode *inode = page->mapping->host;
1848 struct ext4_io_submit io_submit;
1849
1850 trace_ext4_writepage(page);
1851 size = i_size_read(inode);
1852 if (page->index == size >> PAGE_CACHE_SHIFT)
1853 len = size & ~PAGE_CACHE_MASK;
1854 else
1855 len = PAGE_CACHE_SIZE;
1856
1857 page_bufs = page_buffers(page);
1858 /*
1859 * We cannot do block allocation or other extent handling in this
1860 * function. If there are buffers needing that, we have to redirty
1861 * the page. But we may reach here when we do a journal commit via
1862 * journal_submit_inode_data_buffers() and in that case we must write
1863 * allocated buffers to achieve data=ordered mode guarantees.
1864 */
1865 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1866 ext4_bh_delay_or_unwritten)) {
1867 redirty_page_for_writepage(wbc, page);
1868 if (current->flags & PF_MEMALLOC) {
1869 /*
1870 * For memory cleaning there's no point in writing only
1871 * some buffers. So just bail out. Warn if we came here
1872 * from direct reclaim.
1873 */
1874 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
1875 == PF_MEMALLOC);
1876 unlock_page(page);
1877 return 0;
1878 }
1879 }
1880
1881 if (PageChecked(page) && ext4_should_journal_data(inode))
1882 /*
1883 * It's mmapped pagecache. Add buffers and journal it. There
1884 * doesn't seem much point in redirtying the page here.
1885 */
1886 return __ext4_journalled_writepage(page, len);
1887
1888 ext4_io_submit_init(&io_submit, wbc);
1889 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
1890 if (!io_submit.io_end) {
1891 redirty_page_for_writepage(wbc, page);
1892 unlock_page(page);
1893 return -ENOMEM;
1894 }
1895 ret = ext4_bio_write_page(&io_submit, page, len, wbc);
1896 ext4_io_submit(&io_submit);
1897 /* Drop io_end reference we got from init */
1898 ext4_put_io_end_defer(io_submit.io_end);
1899 return ret;
1900}
1901
1902static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
1903{
1904 int len;
1905 loff_t size = i_size_read(mpd->inode);
1906 int err;
1907
1908 BUG_ON(page->index != mpd->first_page);
1909 if (page->index == size >> PAGE_CACHE_SHIFT)
1910 len = size & ~PAGE_CACHE_MASK;
1911 else
1912 len = PAGE_CACHE_SIZE;
1913 clear_page_dirty_for_io(page);
1914 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc);
1915 if (!err)
1916 mpd->wbc->nr_to_write--;
1917 mpd->first_page++;
1918
1919 return err;
1920}
1921
1922#define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
1923
1924/*
1925 * mballoc gives us at most this number of blocks...
1926 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
1927 * The rest of mballoc seems to handle chunks up to full group size.
1928 */
1929#define MAX_WRITEPAGES_EXTENT_LEN 2048
1930
1931/*
1932 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
1933 *
1934 * @mpd - extent of blocks
1935 * @lblk - logical number of the block in the file
1936 * @bh - buffer head we want to add to the extent
1937 *
1938 * The function is used to collect contig. blocks in the same state. If the
1939 * buffer doesn't require mapping for writeback and we haven't started the
1940 * extent of buffers to map yet, the function returns 'true' immediately - the
1941 * caller can write the buffer right away. Otherwise the function returns true
1942 * if the block has been added to the extent, false if the block couldn't be
1943 * added.
1944 */
1945static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
1946 struct buffer_head *bh)
1947{
1948 struct ext4_map_blocks *map = &mpd->map;
1949
1950 /* Buffer that doesn't need mapping for writeback? */
1951 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
1952 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
1953 /* So far no extent to map => we write the buffer right away */
1954 if (map->m_len == 0)
1955 return true;
1956 return false;
1957 }
1958
1959 /* First block in the extent? */
1960 if (map->m_len == 0) {
1961 map->m_lblk = lblk;
1962 map->m_len = 1;
1963 map->m_flags = bh->b_state & BH_FLAGS;
1964 return true;
1965 }
1966
1967 /* Don't go larger than mballoc is willing to allocate */
1968 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
1969 return false;
1970
1971 /* Can we merge the block to our big extent? */
1972 if (lblk == map->m_lblk + map->m_len &&
1973 (bh->b_state & BH_FLAGS) == map->m_flags) {
1974 map->m_len++;
1975 return true;
1976 }
1977 return false;
1978}
1979
1980/*
1981 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
1982 *
1983 * @mpd - extent of blocks for mapping
1984 * @head - the first buffer in the page
1985 * @bh - buffer we should start processing from
1986 * @lblk - logical number of the block in the file corresponding to @bh
1987 *
1988 * Walk through page buffers from @bh upto @head (exclusive) and either submit
1989 * the page for IO if all buffers in this page were mapped and there's no
1990 * accumulated extent of buffers to map or add buffers in the page to the
1991 * extent of buffers to map. The function returns 1 if the caller can continue
1992 * by processing the next page, 0 if it should stop adding buffers to the
1993 * extent to map because we cannot extend it anymore. It can also return value
1994 * < 0 in case of error during IO submission.
1995 */
1996static int mpage_process_page_bufs(struct mpage_da_data *mpd,
1997 struct buffer_head *head,
1998 struct buffer_head *bh,
1999 ext4_lblk_t lblk)
2000{
2001 struct inode *inode = mpd->inode;
2002 int err;
2003 ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
2004 >> inode->i_blkbits;
2005
2006 do {
2007 BUG_ON(buffer_locked(bh));
2008
2009 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2010 /* Found extent to map? */
2011 if (mpd->map.m_len)
2012 return 0;
2013 /* Everything mapped so far and we hit EOF */
2014 break;
2015 }
2016 } while (lblk++, (bh = bh->b_this_page) != head);
2017 /* So far everything mapped? Submit the page for IO. */
2018 if (mpd->map.m_len == 0) {
2019 err = mpage_submit_page(mpd, head->b_page);
2020 if (err < 0)
2021 return err;
2022 }
2023 return lblk < blocks;
2024}
2025
2026/*
2027 * mpage_map_buffers - update buffers corresponding to changed extent and
2028 * submit fully mapped pages for IO
2029 *
2030 * @mpd - description of extent to map, on return next extent to map
2031 *
2032 * Scan buffers corresponding to changed extent (we expect corresponding pages
2033 * to be already locked) and update buffer state according to new extent state.
2034 * We map delalloc buffers to their physical location, clear unwritten bits,
2035 * and mark buffers as uninit when we perform writes to uninitialized extents
2036 * and do extent conversion after IO is finished. If the last page is not fully
2037 * mapped, we update @map to the next extent in the last page that needs
2038 * mapping. Otherwise we submit the page for IO.
2039 */
2040static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2041{
2042 struct pagevec pvec;
2043 int nr_pages, i;
2044 struct inode *inode = mpd->inode;
2045 struct buffer_head *head, *bh;
2046 int bpp_bits = PAGE_CACHE_SHIFT - inode->i_blkbits;
2047 pgoff_t start, end;
2048 ext4_lblk_t lblk;
2049 sector_t pblock;
2050 int err;
2051
2052 start = mpd->map.m_lblk >> bpp_bits;
2053 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2054 lblk = start << bpp_bits;
2055 pblock = mpd->map.m_pblk;
2056
2057 pagevec_init(&pvec, 0);
2058 while (start <= end) {
2059 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2060 PAGEVEC_SIZE);
2061 if (nr_pages == 0)
2062 break;
2063 for (i = 0; i < nr_pages; i++) {
2064 struct page *page = pvec.pages[i];
2065
2066 if (page->index > end)
2067 break;
2068 /* Up to 'end' pages must be contiguous */
2069 BUG_ON(page->index != start);
2070 bh = head = page_buffers(page);
2071 do {
2072 if (lblk < mpd->map.m_lblk)
2073 continue;
2074 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2075 /*
2076 * Buffer after end of mapped extent.
2077 * Find next buffer in the page to map.
2078 */
2079 mpd->map.m_len = 0;
2080 mpd->map.m_flags = 0;
2081 /*
2082 * FIXME: If dioread_nolock supports
2083 * blocksize < pagesize, we need to make
2084 * sure we add size mapped so far to
2085 * io_end->size as the following call
2086 * can submit the page for IO.
2087 */
2088 err = mpage_process_page_bufs(mpd, head,
2089 bh, lblk);
2090 pagevec_release(&pvec);
2091 if (err > 0)
2092 err = 0;
2093 return err;
2094 }
2095 if (buffer_delay(bh)) {
2096 clear_buffer_delay(bh);
2097 bh->b_blocknr = pblock++;
2098 }
2099 clear_buffer_unwritten(bh);
2100 } while (lblk++, (bh = bh->b_this_page) != head);
2101
2102 /*
2103 * FIXME: This is going to break if dioread_nolock
2104 * supports blocksize < pagesize as we will try to
2105 * convert potentially unmapped parts of inode.
2106 */
2107 mpd->io_submit.io_end->size += PAGE_CACHE_SIZE;
2108 /* Page fully mapped - let IO run! */
2109 err = mpage_submit_page(mpd, page);
2110 if (err < 0) {
2111 pagevec_release(&pvec);
2112 return err;
2113 }
2114 start++;
2115 }
2116 pagevec_release(&pvec);
2117 }
2118 /* Extent fully mapped and matches with page boundary. We are done. */
2119 mpd->map.m_len = 0;
2120 mpd->map.m_flags = 0;
2121 return 0;
2122}
2123
2124static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2125{
2126 struct inode *inode = mpd->inode;
2127 struct ext4_map_blocks *map = &mpd->map;
2128 int get_blocks_flags;
2129 int err;
2130
2131 trace_ext4_da_write_pages_extent(inode, map);
2132 /*
2133 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2134 * to convert an uninitialized extent to be initialized (in the case
2135 * where we have written into one or more preallocated blocks). It is
2136 * possible that we're going to need more metadata blocks than
2137 * previously reserved. However we must not fail because we're in
2138 * writeback and there is nothing we can do about it so it might result
2139 * in data loss. So use reserved blocks to allocate metadata if
2140 * possible.
2141 *
2142 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if the blocks
2143 * in question are delalloc blocks. This affects functions in many
2144 * different parts of the allocation call path. This flag exists
2145 * primarily because we don't want to change *many* call functions, so
2146 * ext4_map_blocks() will set the EXT4_STATE_DELALLOC_RESERVED flag
2147 * once the inode's allocation semaphore is taken.
2148 */
2149 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2150 EXT4_GET_BLOCKS_METADATA_NOFAIL;
2151 if (ext4_should_dioread_nolock(inode))
2152 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2153 if (map->m_flags & (1 << BH_Delay))
2154 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2155
2156 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2157 if (err < 0)
2158 return err;
2159 if (map->m_flags & EXT4_MAP_UNINIT) {
2160 if (!mpd->io_submit.io_end->handle &&
2161 ext4_handle_valid(handle)) {
2162 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2163 handle->h_rsv_handle = NULL;
2164 }
2165 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2166 }
2167
2168 BUG_ON(map->m_len == 0);
2169 if (map->m_flags & EXT4_MAP_NEW) {
2170 struct block_device *bdev = inode->i_sb->s_bdev;
2171 int i;
2172
2173 for (i = 0; i < map->m_len; i++)
2174 unmap_underlying_metadata(bdev, map->m_pblk + i);
2175 }
2176 return 0;
2177}
2178
2179/*
2180 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2181 * mpd->len and submit pages underlying it for IO
2182 *
2183 * @handle - handle for journal operations
2184 * @mpd - extent to map
2185 * @give_up_on_write - we set this to true iff there is a fatal error and there
2186 * is no hope of writing the data. The caller should discard
2187 * dirty pages to avoid infinite loops.
2188 *
2189 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2190 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2191 * them to initialized or split the described range from larger unwritten
2192 * extent. Note that we need not map all the described range since allocation
2193 * can return less blocks or the range is covered by more unwritten extents. We
2194 * cannot map more because we are limited by reserved transaction credits. On
2195 * the other hand we always make sure that the last touched page is fully
2196 * mapped so that it can be written out (and thus forward progress is
2197 * guaranteed). After mapping we submit all mapped pages for IO.
2198 */
2199static int mpage_map_and_submit_extent(handle_t *handle,
2200 struct mpage_da_data *mpd,
2201 bool *give_up_on_write)
2202{
2203 struct inode *inode = mpd->inode;
2204 struct ext4_map_blocks *map = &mpd->map;
2205 int err;
2206 loff_t disksize;
2207
2208 mpd->io_submit.io_end->offset =
2209 ((loff_t)map->m_lblk) << inode->i_blkbits;
2210 do {
2211 err = mpage_map_one_extent(handle, mpd);
2212 if (err < 0) {
2213 struct super_block *sb = inode->i_sb;
2214
2215 if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2216 goto invalidate_dirty_pages;
2217 /*
2218 * Let the uper layers retry transient errors.
2219 * In the case of ENOSPC, if ext4_count_free_blocks()
2220 * is non-zero, a commit should free up blocks.
2221 */
2222 if ((err == -ENOMEM) ||
2223 (err == -ENOSPC && ext4_count_free_clusters(sb)))
2224 return err;
2225 ext4_msg(sb, KERN_CRIT,
2226 "Delayed block allocation failed for "
2227 "inode %lu at logical offset %llu with"
2228 " max blocks %u with error %d",
2229 inode->i_ino,
2230 (unsigned long long)map->m_lblk,
2231 (unsigned)map->m_len, -err);
2232 ext4_msg(sb, KERN_CRIT,
2233 "This should not happen!! Data will "
2234 "be lost\n");
2235 if (err == -ENOSPC)
2236 ext4_print_free_blocks(inode);
2237 invalidate_dirty_pages:
2238 *give_up_on_write = true;
2239 return err;
2240 }
2241 /*
2242 * Update buffer state, submit mapped pages, and get us new
2243 * extent to map
2244 */
2245 err = mpage_map_and_submit_buffers(mpd);
2246 if (err < 0)
2247 return err;
2248 } while (map->m_len);
2249
2250 /*
2251 * Update on-disk size after IO is submitted. Races with
2252 * truncate are avoided by checking i_size under i_data_sem.
2253 */
2254 disksize = ((loff_t)mpd->first_page) << PAGE_CACHE_SHIFT;
2255 if (disksize > EXT4_I(inode)->i_disksize) {
2256 int err2;
2257 loff_t i_size;
2258
2259 down_write(&EXT4_I(inode)->i_data_sem);
2260 i_size = i_size_read(inode);
2261 if (disksize > i_size)
2262 disksize = i_size;
2263 if (disksize > EXT4_I(inode)->i_disksize)
2264 EXT4_I(inode)->i_disksize = disksize;
2265 err2 = ext4_mark_inode_dirty(handle, inode);
2266 up_write(&EXT4_I(inode)->i_data_sem);
2267 if (err2)
2268 ext4_error(inode->i_sb,
2269 "Failed to mark inode %lu dirty",
2270 inode->i_ino);
2271 if (!err)
2272 err = err2;
2273 }
2274 return err;
2275}
2276
2277/*
2278 * Calculate the total number of credits to reserve for one writepages
2279 * iteration. This is called from ext4_writepages(). We map an extent of
2280 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2281 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2282 * bpp - 1 blocks in bpp different extents.
2283 */
2284static int ext4_da_writepages_trans_blocks(struct inode *inode)
2285{
2286 int bpp = ext4_journal_blocks_per_page(inode);
2287
2288 return ext4_meta_trans_blocks(inode,
2289 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2290}
2291
2292/*
2293 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2294 * and underlying extent to map
2295 *
2296 * @mpd - where to look for pages
2297 *
2298 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2299 * IO immediately. When we find a page which isn't mapped we start accumulating
2300 * extent of buffers underlying these pages that needs mapping (formed by
2301 * either delayed or unwritten buffers). We also lock the pages containing
2302 * these buffers. The extent found is returned in @mpd structure (starting at
2303 * mpd->lblk with length mpd->len blocks).
2304 *
2305 * Note that this function can attach bios to one io_end structure which are
2306 * neither logically nor physically contiguous. Although it may seem as an
2307 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2308 * case as we need to track IO to all buffers underlying a page in one io_end.
2309 */
2310static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2311{
2312 struct address_space *mapping = mpd->inode->i_mapping;
2313 struct pagevec pvec;
2314 unsigned int nr_pages;
2315 long left = mpd->wbc->nr_to_write;
2316 pgoff_t index = mpd->first_page;
2317 pgoff_t end = mpd->last_page;
2318 int tag;
2319 int i, err = 0;
2320 int blkbits = mpd->inode->i_blkbits;
2321 ext4_lblk_t lblk;
2322 struct buffer_head *head;
2323
2324 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2325 tag = PAGECACHE_TAG_TOWRITE;
2326 else
2327 tag = PAGECACHE_TAG_DIRTY;
2328
2329 pagevec_init(&pvec, 0);
2330 mpd->map.m_len = 0;
2331 mpd->next_page = index;
2332 while (index <= end) {
2333 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2334 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2335 if (nr_pages == 0)
2336 goto out;
2337
2338 for (i = 0; i < nr_pages; i++) {
2339 struct page *page = pvec.pages[i];
2340
2341 /*
2342 * At this point, the page may be truncated or
2343 * invalidated (changing page->mapping to NULL), or
2344 * even swizzled back from swapper_space to tmpfs file
2345 * mapping. However, page->index will not change
2346 * because we have a reference on the page.
2347 */
2348 if (page->index > end)
2349 goto out;
2350
2351 /*
2352 * Accumulated enough dirty pages? This doesn't apply
2353 * to WB_SYNC_ALL mode. For integrity sync we have to
2354 * keep going because someone may be concurrently
2355 * dirtying pages, and we might have synced a lot of
2356 * newly appeared dirty pages, but have not synced all
2357 * of the old dirty pages.
2358 */
2359 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2360 goto out;
2361
2362 /* If we can't merge this page, we are done. */
2363 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2364 goto out;
2365
2366 lock_page(page);
2367 /*
2368 * If the page is no longer dirty, or its mapping no
2369 * longer corresponds to inode we are writing (which
2370 * means it has been truncated or invalidated), or the
2371 * page is already under writeback and we are not doing
2372 * a data integrity writeback, skip the page
2373 */
2374 if (!PageDirty(page) ||
2375 (PageWriteback(page) &&
2376 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2377 unlikely(page->mapping != mapping)) {
2378 unlock_page(page);
2379 continue;
2380 }
2381
2382 wait_on_page_writeback(page);
2383 BUG_ON(PageWriteback(page));
2384
2385 if (mpd->map.m_len == 0)
2386 mpd->first_page = page->index;
2387 mpd->next_page = page->index + 1;
2388 /* Add all dirty buffers to mpd */
2389 lblk = ((ext4_lblk_t)page->index) <<
2390 (PAGE_CACHE_SHIFT - blkbits);
2391 head = page_buffers(page);
2392 err = mpage_process_page_bufs(mpd, head, head, lblk);
2393 if (err <= 0)
2394 goto out;
2395 err = 0;
2396 left--;
2397 }
2398 pagevec_release(&pvec);
2399 cond_resched();
2400 }
2401 return 0;
2402out:
2403 pagevec_release(&pvec);
2404 return err;
2405}
2406
2407static int __writepage(struct page *page, struct writeback_control *wbc,
2408 void *data)
2409{
2410 struct address_space *mapping = data;
2411 int ret = ext4_writepage(page, wbc);
2412 mapping_set_error(mapping, ret);
2413 return ret;
2414}
2415
2416static int ext4_writepages(struct address_space *mapping,
2417 struct writeback_control *wbc)
2418{
2419 pgoff_t writeback_index = 0;
2420 long nr_to_write = wbc->nr_to_write;
2421 int range_whole = 0;
2422 int cycled = 1;
2423 handle_t *handle = NULL;
2424 struct mpage_da_data mpd;
2425 struct inode *inode = mapping->host;
2426 int needed_blocks, rsv_blocks = 0, ret = 0;
2427 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2428 bool done;
2429 struct blk_plug plug;
2430 bool give_up_on_write = false;
2431
2432 trace_ext4_writepages(inode, wbc);
2433
2434 /*
2435 * No pages to write? This is mainly a kludge to avoid starting
2436 * a transaction for special inodes like journal inode on last iput()
2437 * because that could violate lock ordering on umount
2438 */
2439 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2440 goto out_writepages;
2441
2442 if (ext4_should_journal_data(inode)) {
2443 struct blk_plug plug;
2444
2445 blk_start_plug(&plug);
2446 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2447 blk_finish_plug(&plug);
2448 goto out_writepages;
2449 }
2450
2451 /*
2452 * If the filesystem has aborted, it is read-only, so return
2453 * right away instead of dumping stack traces later on that
2454 * will obscure the real source of the problem. We test
2455 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2456 * the latter could be true if the filesystem is mounted
2457 * read-only, and in that case, ext4_writepages should
2458 * *never* be called, so if that ever happens, we would want
2459 * the stack trace.
2460 */
2461 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2462 ret = -EROFS;
2463 goto out_writepages;
2464 }
2465
2466 if (ext4_should_dioread_nolock(inode)) {
2467 /*
2468 * We may need to convert up to one extent per block in
2469 * the page and we may dirty the inode.
2470 */
2471 rsv_blocks = 1 + (PAGE_CACHE_SIZE >> inode->i_blkbits);
2472 }
2473
2474 /*
2475 * If we have inline data and arrive here, it means that
2476 * we will soon create the block for the 1st page, so
2477 * we'd better clear the inline data here.
2478 */
2479 if (ext4_has_inline_data(inode)) {
2480 /* Just inode will be modified... */
2481 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2482 if (IS_ERR(handle)) {
2483 ret = PTR_ERR(handle);
2484 goto out_writepages;
2485 }
2486 BUG_ON(ext4_test_inode_state(inode,
2487 EXT4_STATE_MAY_INLINE_DATA));
2488 ext4_destroy_inline_data(handle, inode);
2489 ext4_journal_stop(handle);
2490 }
2491
2492 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2493 range_whole = 1;
2494
2495 if (wbc->range_cyclic) {
2496 writeback_index = mapping->writeback_index;
2497 if (writeback_index)
2498 cycled = 0;
2499 mpd.first_page = writeback_index;
2500 mpd.last_page = -1;
2501 } else {
2502 mpd.first_page = wbc->range_start >> PAGE_CACHE_SHIFT;
2503 mpd.last_page = wbc->range_end >> PAGE_CACHE_SHIFT;
2504 }
2505
2506 mpd.inode = inode;
2507 mpd.wbc = wbc;
2508 ext4_io_submit_init(&mpd.io_submit, wbc);
2509retry:
2510 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2511 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2512 done = false;
2513 blk_start_plug(&plug);
2514 while (!done && mpd.first_page <= mpd.last_page) {
2515 /* For each extent of pages we use new io_end */
2516 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2517 if (!mpd.io_submit.io_end) {
2518 ret = -ENOMEM;
2519 break;
2520 }
2521
2522 /*
2523 * We have two constraints: We find one extent to map and we
2524 * must always write out whole page (makes a difference when
2525 * blocksize < pagesize) so that we don't block on IO when we
2526 * try to write out the rest of the page. Journalled mode is
2527 * not supported by delalloc.
2528 */
2529 BUG_ON(ext4_should_journal_data(inode));
2530 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2531
2532 /* start a new transaction */
2533 handle = ext4_journal_start_with_reserve(inode,
2534 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2535 if (IS_ERR(handle)) {
2536 ret = PTR_ERR(handle);
2537 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2538 "%ld pages, ino %lu; err %d", __func__,
2539 wbc->nr_to_write, inode->i_ino, ret);
2540 /* Release allocated io_end */
2541 ext4_put_io_end(mpd.io_submit.io_end);
2542 break;
2543 }
2544
2545 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2546 ret = mpage_prepare_extent_to_map(&mpd);
2547 if (!ret) {
2548 if (mpd.map.m_len)
2549 ret = mpage_map_and_submit_extent(handle, &mpd,
2550 &give_up_on_write);
2551 else {
2552 /*
2553 * We scanned the whole range (or exhausted
2554 * nr_to_write), submitted what was mapped and
2555 * didn't find anything needing mapping. We are
2556 * done.
2557 */
2558 done = true;
2559 }
2560 }
2561 ext4_journal_stop(handle);
2562 /* Submit prepared bio */
2563 ext4_io_submit(&mpd.io_submit);
2564 /* Unlock pages we didn't use */
2565 mpage_release_unused_pages(&mpd, give_up_on_write);
2566 /* Drop our io_end reference we got from init */
2567 ext4_put_io_end(mpd.io_submit.io_end);
2568
2569 if (ret == -ENOSPC && sbi->s_journal) {
2570 /*
2571 * Commit the transaction which would
2572 * free blocks released in the transaction
2573 * and try again
2574 */
2575 jbd2_journal_force_commit_nested(sbi->s_journal);
2576 ret = 0;
2577 continue;
2578 }
2579 /* Fatal error - ENOMEM, EIO... */
2580 if (ret)
2581 break;
2582 }
2583 blk_finish_plug(&plug);
2584 if (!ret && !cycled && wbc->nr_to_write > 0) {
2585 cycled = 1;
2586 mpd.last_page = writeback_index - 1;
2587 mpd.first_page = 0;
2588 goto retry;
2589 }
2590
2591 /* Update index */
2592 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2593 /*
2594 * Set the writeback_index so that range_cyclic
2595 * mode will write it back later
2596 */
2597 mapping->writeback_index = mpd.first_page;
2598
2599out_writepages:
2600 trace_ext4_writepages_result(inode, wbc, ret,
2601 nr_to_write - wbc->nr_to_write);
2602 return ret;
2603}
2604
2605static int ext4_nonda_switch(struct super_block *sb)
2606{
2607 s64 free_clusters, dirty_clusters;
2608 struct ext4_sb_info *sbi = EXT4_SB(sb);
2609
2610 /*
2611 * switch to non delalloc mode if we are running low
2612 * on free block. The free block accounting via percpu
2613 * counters can get slightly wrong with percpu_counter_batch getting
2614 * accumulated on each CPU without updating global counters
2615 * Delalloc need an accurate free block accounting. So switch
2616 * to non delalloc when we are near to error range.
2617 */
2618 free_clusters =
2619 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2620 dirty_clusters =
2621 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2622 /*
2623 * Start pushing delalloc when 1/2 of free blocks are dirty.
2624 */
2625 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2626 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2627
2628 if (2 * free_clusters < 3 * dirty_clusters ||
2629 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2630 /*
2631 * free block count is less than 150% of dirty blocks
2632 * or free blocks is less than watermark
2633 */
2634 return 1;
2635 }
2636 return 0;
2637}
2638
2639static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2640 loff_t pos, unsigned len, unsigned flags,
2641 struct page **pagep, void **fsdata)
2642{
2643 int ret, retries = 0;
2644 struct page *page;
2645 pgoff_t index;
2646 struct inode *inode = mapping->host;
2647 handle_t *handle;
2648
2649 index = pos >> PAGE_CACHE_SHIFT;
2650
2651 if (ext4_nonda_switch(inode->i_sb)) {
2652 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2653 return ext4_write_begin(file, mapping, pos,
2654 len, flags, pagep, fsdata);
2655 }
2656 *fsdata = (void *)0;
2657 trace_ext4_da_write_begin(inode, pos, len, flags);
2658
2659 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2660 ret = ext4_da_write_inline_data_begin(mapping, inode,
2661 pos, len, flags,
2662 pagep, fsdata);
2663 if (ret < 0)
2664 return ret;
2665 if (ret == 1)
2666 return 0;
2667 }
2668
2669 /*
2670 * grab_cache_page_write_begin() can take a long time if the
2671 * system is thrashing due to memory pressure, or if the page
2672 * is being written back. So grab it first before we start
2673 * the transaction handle. This also allows us to allocate
2674 * the page (if needed) without using GFP_NOFS.
2675 */
2676retry_grab:
2677 page = grab_cache_page_write_begin(mapping, index, flags);
2678 if (!page)
2679 return -ENOMEM;
2680 unlock_page(page);
2681
2682 /*
2683 * With delayed allocation, we don't log the i_disksize update
2684 * if there is delayed block allocation. But we still need
2685 * to journalling the i_disksize update if writes to the end
2686 * of file which has an already mapped buffer.
2687 */
2688retry_journal:
2689 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, 1);
2690 if (IS_ERR(handle)) {
2691 page_cache_release(page);
2692 return PTR_ERR(handle);
2693 }
2694
2695 lock_page(page);
2696 if (page->mapping != mapping) {
2697 /* The page got truncated from under us */
2698 unlock_page(page);
2699 page_cache_release(page);
2700 ext4_journal_stop(handle);
2701 goto retry_grab;
2702 }
2703 /* In case writeback began while the page was unlocked */
2704 wait_for_stable_page(page);
2705
2706 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2707 if (ret < 0) {
2708 unlock_page(page);
2709 ext4_journal_stop(handle);
2710 /*
2711 * block_write_begin may have instantiated a few blocks
2712 * outside i_size. Trim these off again. Don't need
2713 * i_size_read because we hold i_mutex.
2714 */
2715 if (pos + len > inode->i_size)
2716 ext4_truncate_failed_write(inode);
2717
2718 if (ret == -ENOSPC &&
2719 ext4_should_retry_alloc(inode->i_sb, &retries))
2720 goto retry_journal;
2721
2722 page_cache_release(page);
2723 return ret;
2724 }
2725
2726 *pagep = page;
2727 return ret;
2728}
2729
2730/*
2731 * Check if we should update i_disksize
2732 * when write to the end of file but not require block allocation
2733 */
2734static int ext4_da_should_update_i_disksize(struct page *page,
2735 unsigned long offset)
2736{
2737 struct buffer_head *bh;
2738 struct inode *inode = page->mapping->host;
2739 unsigned int idx;
2740 int i;
2741
2742 bh = page_buffers(page);
2743 idx = offset >> inode->i_blkbits;
2744
2745 for (i = 0; i < idx; i++)
2746 bh = bh->b_this_page;
2747
2748 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2749 return 0;
2750 return 1;
2751}
2752
2753static int ext4_da_write_end(struct file *file,
2754 struct address_space *mapping,
2755 loff_t pos, unsigned len, unsigned copied,
2756 struct page *page, void *fsdata)
2757{
2758 struct inode *inode = mapping->host;
2759 int ret = 0, ret2;
2760 handle_t *handle = ext4_journal_current_handle();
2761 loff_t new_i_size;
2762 unsigned long start, end;
2763 int write_mode = (int)(unsigned long)fsdata;
2764
2765 if (write_mode == FALL_BACK_TO_NONDELALLOC)
2766 return ext4_write_end(file, mapping, pos,
2767 len, copied, page, fsdata);
2768
2769 trace_ext4_da_write_end(inode, pos, len, copied);
2770 start = pos & (PAGE_CACHE_SIZE - 1);
2771 end = start + copied - 1;
2772
2773 /*
2774 * generic_write_end() will run mark_inode_dirty() if i_size
2775 * changes. So let's piggyback the i_disksize mark_inode_dirty
2776 * into that.
2777 */
2778 new_i_size = pos + copied;
2779 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2780 if (ext4_has_inline_data(inode) ||
2781 ext4_da_should_update_i_disksize(page, end)) {
2782 down_write(&EXT4_I(inode)->i_data_sem);
2783 if (new_i_size > EXT4_I(inode)->i_disksize)
2784 EXT4_I(inode)->i_disksize = new_i_size;
2785 up_write(&EXT4_I(inode)->i_data_sem);
2786 /* We need to mark inode dirty even if
2787 * new_i_size is less that inode->i_size
2788 * bu greater than i_disksize.(hint delalloc)
2789 */
2790 ext4_mark_inode_dirty(handle, inode);
2791 }
2792 }
2793
2794 if (write_mode != CONVERT_INLINE_DATA &&
2795 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
2796 ext4_has_inline_data(inode))
2797 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
2798 page);
2799 else
2800 ret2 = generic_write_end(file, mapping, pos, len, copied,
2801 page, fsdata);
2802
2803 copied = ret2;
2804 if (ret2 < 0)
2805 ret = ret2;
2806 ret2 = ext4_journal_stop(handle);
2807 if (!ret)
2808 ret = ret2;
2809
2810 return ret ? ret : copied;
2811}
2812
2813static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
2814 unsigned int length)
2815{
2816 /*
2817 * Drop reserved blocks
2818 */
2819 BUG_ON(!PageLocked(page));
2820 if (!page_has_buffers(page))
2821 goto out;
2822
2823 ext4_da_page_release_reservation(page, offset, length);
2824
2825out:
2826 ext4_invalidatepage(page, offset, length);
2827
2828 return;
2829}
2830
2831/*
2832 * Force all delayed allocation blocks to be allocated for a given inode.
2833 */
2834int ext4_alloc_da_blocks(struct inode *inode)
2835{
2836 trace_ext4_alloc_da_blocks(inode);
2837
2838 if (!EXT4_I(inode)->i_reserved_data_blocks &&
2839 !EXT4_I(inode)->i_reserved_meta_blocks)
2840 return 0;
2841
2842 /*
2843 * We do something simple for now. The filemap_flush() will
2844 * also start triggering a write of the data blocks, which is
2845 * not strictly speaking necessary (and for users of
2846 * laptop_mode, not even desirable). However, to do otherwise
2847 * would require replicating code paths in:
2848 *
2849 * ext4_writepages() ->
2850 * write_cache_pages() ---> (via passed in callback function)
2851 * __mpage_da_writepage() -->
2852 * mpage_add_bh_to_extent()
2853 * mpage_da_map_blocks()
2854 *
2855 * The problem is that write_cache_pages(), located in
2856 * mm/page-writeback.c, marks pages clean in preparation for
2857 * doing I/O, which is not desirable if we're not planning on
2858 * doing I/O at all.
2859 *
2860 * We could call write_cache_pages(), and then redirty all of
2861 * the pages by calling redirty_page_for_writepage() but that
2862 * would be ugly in the extreme. So instead we would need to
2863 * replicate parts of the code in the above functions,
2864 * simplifying them because we wouldn't actually intend to
2865 * write out the pages, but rather only collect contiguous
2866 * logical block extents, call the multi-block allocator, and
2867 * then update the buffer heads with the block allocations.
2868 *
2869 * For now, though, we'll cheat by calling filemap_flush(),
2870 * which will map the blocks, and start the I/O, but not
2871 * actually wait for the I/O to complete.
2872 */
2873 return filemap_flush(inode->i_mapping);
2874}
2875
2876/*
2877 * bmap() is special. It gets used by applications such as lilo and by
2878 * the swapper to find the on-disk block of a specific piece of data.
2879 *
2880 * Naturally, this is dangerous if the block concerned is still in the
2881 * journal. If somebody makes a swapfile on an ext4 data-journaling
2882 * filesystem and enables swap, then they may get a nasty shock when the
2883 * data getting swapped to that swapfile suddenly gets overwritten by
2884 * the original zero's written out previously to the journal and
2885 * awaiting writeback in the kernel's buffer cache.
2886 *
2887 * So, if we see any bmap calls here on a modified, data-journaled file,
2888 * take extra steps to flush any blocks which might be in the cache.
2889 */
2890static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2891{
2892 struct inode *inode = mapping->host;
2893 journal_t *journal;
2894 int err;
2895
2896 /*
2897 * We can get here for an inline file via the FIBMAP ioctl
2898 */
2899 if (ext4_has_inline_data(inode))
2900 return 0;
2901
2902 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2903 test_opt(inode->i_sb, DELALLOC)) {
2904 /*
2905 * With delalloc we want to sync the file
2906 * so that we can make sure we allocate
2907 * blocks for file
2908 */
2909 filemap_write_and_wait(mapping);
2910 }
2911
2912 if (EXT4_JOURNAL(inode) &&
2913 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2914 /*
2915 * This is a REALLY heavyweight approach, but the use of
2916 * bmap on dirty files is expected to be extremely rare:
2917 * only if we run lilo or swapon on a freshly made file
2918 * do we expect this to happen.
2919 *
2920 * (bmap requires CAP_SYS_RAWIO so this does not
2921 * represent an unprivileged user DOS attack --- we'd be
2922 * in trouble if mortal users could trigger this path at
2923 * will.)
2924 *
2925 * NB. EXT4_STATE_JDATA is not set on files other than
2926 * regular files. If somebody wants to bmap a directory
2927 * or symlink and gets confused because the buffer
2928 * hasn't yet been flushed to disk, they deserve
2929 * everything they get.
2930 */
2931
2932 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2933 journal = EXT4_JOURNAL(inode);
2934 jbd2_journal_lock_updates(journal);
2935 err = jbd2_journal_flush(journal);
2936 jbd2_journal_unlock_updates(journal);
2937
2938 if (err)
2939 return 0;
2940 }
2941
2942 return generic_block_bmap(mapping, block, ext4_get_block);
2943}
2944
2945static int ext4_readpage(struct file *file, struct page *page)
2946{
2947 int ret = -EAGAIN;
2948 struct inode *inode = page->mapping->host;
2949
2950 trace_ext4_readpage(page);
2951
2952 if (ext4_has_inline_data(inode))
2953 ret = ext4_readpage_inline(inode, page);
2954
2955 if (ret == -EAGAIN)
2956 return mpage_readpage(page, ext4_get_block);
2957
2958 return ret;
2959}
2960
2961static int
2962ext4_readpages(struct file *file, struct address_space *mapping,
2963 struct list_head *pages, unsigned nr_pages)
2964{
2965 struct inode *inode = mapping->host;
2966
2967 /* If the file has inline data, no need to do readpages. */
2968 if (ext4_has_inline_data(inode))
2969 return 0;
2970
2971 return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2972}
2973
2974static void ext4_invalidatepage(struct page *page, unsigned int offset,
2975 unsigned int length)
2976{
2977 trace_ext4_invalidatepage(page, offset, length);
2978
2979 /* No journalling happens on data buffers when this function is used */
2980 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
2981
2982 block_invalidatepage(page, offset, length);
2983}
2984
2985static int __ext4_journalled_invalidatepage(struct page *page,
2986 unsigned int offset,
2987 unsigned int length)
2988{
2989 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2990
2991 trace_ext4_journalled_invalidatepage(page, offset, length);
2992
2993 /*
2994 * If it's a full truncate we just forget about the pending dirtying
2995 */
2996 if (offset == 0 && length == PAGE_CACHE_SIZE)
2997 ClearPageChecked(page);
2998
2999 return jbd2_journal_invalidatepage(journal, page, offset, length);
3000}
3001
3002/* Wrapper for aops... */
3003static void ext4_journalled_invalidatepage(struct page *page,
3004 unsigned int offset,
3005 unsigned int length)
3006{
3007 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3008}
3009
3010static int ext4_releasepage(struct page *page, gfp_t wait)
3011{
3012 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3013
3014 trace_ext4_releasepage(page);
3015
3016 /* Page has dirty journalled data -> cannot release */
3017 if (PageChecked(page))
3018 return 0;
3019 if (journal)
3020 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3021 else
3022 return try_to_free_buffers(page);
3023}
3024
3025/*
3026 * ext4_get_block used when preparing for a DIO write or buffer write.
3027 * We allocate an uinitialized extent if blocks haven't been allocated.
3028 * The extent will be converted to initialized after the IO is complete.
3029 */
3030int ext4_get_block_write(struct inode *inode, sector_t iblock,
3031 struct buffer_head *bh_result, int create)
3032{
3033 ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3034 inode->i_ino, create);
3035 return _ext4_get_block(inode, iblock, bh_result,
3036 EXT4_GET_BLOCKS_IO_CREATE_EXT);
3037}
3038
3039static int ext4_get_block_write_nolock(struct inode *inode, sector_t iblock,
3040 struct buffer_head *bh_result, int create)
3041{
3042 ext4_debug("ext4_get_block_write_nolock: inode %lu, create flag %d\n",
3043 inode->i_ino, create);
3044 return _ext4_get_block(inode, iblock, bh_result,
3045 EXT4_GET_BLOCKS_NO_LOCK);
3046}
3047
3048static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3049 ssize_t size, void *private)
3050{
3051 ext4_io_end_t *io_end = iocb->private;
3052
3053 /* if not async direct IO just return */
3054 if (!io_end)
3055 return;
3056
3057 ext_debug("ext4_end_io_dio(): io_end 0x%p "
3058 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3059 iocb->private, io_end->inode->i_ino, iocb, offset,
3060 size);
3061
3062 iocb->private = NULL;
3063 io_end->offset = offset;
3064 io_end->size = size;
3065 ext4_put_io_end(io_end);
3066}
3067
3068/*
3069 * For ext4 extent files, ext4 will do direct-io write to holes,
3070 * preallocated extents, and those write extend the file, no need to
3071 * fall back to buffered IO.
3072 *
3073 * For holes, we fallocate those blocks, mark them as uninitialized
3074 * If those blocks were preallocated, we mark sure they are split, but
3075 * still keep the range to write as uninitialized.
3076 *
3077 * The unwritten extents will be converted to written when DIO is completed.
3078 * For async direct IO, since the IO may still pending when return, we
3079 * set up an end_io call back function, which will do the conversion
3080 * when async direct IO completed.
3081 *
3082 * If the O_DIRECT write will extend the file then add this inode to the
3083 * orphan list. So recovery will truncate it back to the original size
3084 * if the machine crashes during the write.
3085 *
3086 */
3087static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3088 const struct iovec *iov, loff_t offset,
3089 unsigned long nr_segs)
3090{
3091 struct file *file = iocb->ki_filp;
3092 struct inode *inode = file->f_mapping->host;
3093 ssize_t ret;
3094 size_t count = iov_length(iov, nr_segs);
3095 int overwrite = 0;
3096 get_block_t *get_block_func = NULL;
3097 int dio_flags = 0;
3098 loff_t final_size = offset + count;
3099 ext4_io_end_t *io_end = NULL;
3100
3101 /* Use the old path for reads and writes beyond i_size. */
3102 if (rw != WRITE || final_size > inode->i_size)
3103 return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3104
3105 BUG_ON(iocb->private == NULL);
3106
3107 /*
3108 * Make all waiters for direct IO properly wait also for extent
3109 * conversion. This also disallows race between truncate() and
3110 * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3111 */
3112 if (rw == WRITE)
3113 atomic_inc(&inode->i_dio_count);
3114
3115 /* If we do a overwrite dio, i_mutex locking can be released */
3116 overwrite = *((int *)iocb->private);
3117
3118 if (overwrite) {
3119 down_read(&EXT4_I(inode)->i_data_sem);
3120 mutex_unlock(&inode->i_mutex);
3121 }
3122
3123 /*
3124 * We could direct write to holes and fallocate.
3125 *
3126 * Allocated blocks to fill the hole are marked as
3127 * uninitialized to prevent parallel buffered read to expose
3128 * the stale data before DIO complete the data IO.
3129 *
3130 * As to previously fallocated extents, ext4 get_block will
3131 * just simply mark the buffer mapped but still keep the
3132 * extents uninitialized.
3133 *
3134 * For non AIO case, we will convert those unwritten extents
3135 * to written after return back from blockdev_direct_IO.
3136 *
3137 * For async DIO, the conversion needs to be deferred when the
3138 * IO is completed. The ext4 end_io callback function will be
3139 * called to take care of the conversion work. Here for async
3140 * case, we allocate an io_end structure to hook to the iocb.
3141 */
3142 iocb->private = NULL;
3143 ext4_inode_aio_set(inode, NULL);
3144 if (!is_sync_kiocb(iocb)) {
3145 io_end = ext4_init_io_end(inode, GFP_NOFS);
3146 if (!io_end) {
3147 ret = -ENOMEM;
3148 goto retake_lock;
3149 }
3150 /*
3151 * Grab reference for DIO. Will be dropped in ext4_end_io_dio()
3152 */
3153 iocb->private = ext4_get_io_end(io_end);
3154 /*
3155 * we save the io structure for current async direct
3156 * IO, so that later ext4_map_blocks() could flag the
3157 * io structure whether there is a unwritten extents
3158 * needs to be converted when IO is completed.
3159 */
3160 ext4_inode_aio_set(inode, io_end);
3161 }
3162
3163 if (overwrite) {
3164 get_block_func = ext4_get_block_write_nolock;
3165 } else {
3166 get_block_func = ext4_get_block_write;
3167 dio_flags = DIO_LOCKING;
3168 }
3169 ret = __blockdev_direct_IO(rw, iocb, inode,
3170 inode->i_sb->s_bdev, iov,
3171 offset, nr_segs,
3172 get_block_func,
3173 ext4_end_io_dio,
3174 NULL,
3175 dio_flags);
3176
3177 /*
3178 * Put our reference to io_end. This can free the io_end structure e.g.
3179 * in sync IO case or in case of error. It can even perform extent
3180 * conversion if all bios we submitted finished before we got here.
3181 * Note that in that case iocb->private can be already set to NULL
3182 * here.
3183 */
3184 if (io_end) {
3185 ext4_inode_aio_set(inode, NULL);
3186 ext4_put_io_end(io_end);
3187 /*
3188 * When no IO was submitted ext4_end_io_dio() was not
3189 * called so we have to put iocb's reference.
3190 */
3191 if (ret <= 0 && ret != -EIOCBQUEUED && iocb->private) {
3192 WARN_ON(iocb->private != io_end);
3193 WARN_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
3194 ext4_put_io_end(io_end);
3195 iocb->private = NULL;
3196 }
3197 }
3198 if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3199 EXT4_STATE_DIO_UNWRITTEN)) {
3200 int err;
3201 /*
3202 * for non AIO case, since the IO is already
3203 * completed, we could do the conversion right here
3204 */
3205 err = ext4_convert_unwritten_extents(NULL, inode,
3206 offset, ret);
3207 if (err < 0)
3208 ret = err;
3209 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3210 }
3211
3212retake_lock:
3213 if (rw == WRITE)
3214 inode_dio_done(inode);
3215 /* take i_mutex locking again if we do a ovewrite dio */
3216 if (overwrite) {
3217 up_read(&EXT4_I(inode)->i_data_sem);
3218 mutex_lock(&inode->i_mutex);
3219 }
3220
3221 return ret;
3222}
3223
3224static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3225 const struct iovec *iov, loff_t offset,
3226 unsigned long nr_segs)
3227{
3228 struct file *file = iocb->ki_filp;
3229 struct inode *inode = file->f_mapping->host;
3230 ssize_t ret;
3231
3232 /*
3233 * If we are doing data journalling we don't support O_DIRECT
3234 */
3235 if (ext4_should_journal_data(inode))
3236 return 0;
3237
3238 /* Let buffer I/O handle the inline data case. */
3239 if (ext4_has_inline_data(inode))
3240 return 0;
3241
3242 trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
3243 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3244 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3245 else
3246 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3247 trace_ext4_direct_IO_exit(inode, offset,
3248 iov_length(iov, nr_segs), rw, ret);
3249 return ret;
3250}
3251
3252/*
3253 * Pages can be marked dirty completely asynchronously from ext4's journalling
3254 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3255 * much here because ->set_page_dirty is called under VFS locks. The page is
3256 * not necessarily locked.
3257 *
3258 * We cannot just dirty the page and leave attached buffers clean, because the
3259 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3260 * or jbddirty because all the journalling code will explode.
3261 *
3262 * So what we do is to mark the page "pending dirty" and next time writepage
3263 * is called, propagate that into the buffers appropriately.
3264 */
3265static int ext4_journalled_set_page_dirty(struct page *page)
3266{
3267 SetPageChecked(page);
3268 return __set_page_dirty_nobuffers(page);
3269}
3270
3271static const struct address_space_operations ext4_aops = {
3272 .readpage = ext4_readpage,
3273 .readpages = ext4_readpages,
3274 .writepage = ext4_writepage,
3275 .writepages = ext4_writepages,
3276 .write_begin = ext4_write_begin,
3277 .write_end = ext4_write_end,
3278 .bmap = ext4_bmap,
3279 .invalidatepage = ext4_invalidatepage,
3280 .releasepage = ext4_releasepage,
3281 .direct_IO = ext4_direct_IO,
3282 .migratepage = buffer_migrate_page,
3283 .is_partially_uptodate = block_is_partially_uptodate,
3284 .error_remove_page = generic_error_remove_page,
3285};
3286
3287static const struct address_space_operations ext4_journalled_aops = {
3288 .readpage = ext4_readpage,
3289 .readpages = ext4_readpages,
3290 .writepage = ext4_writepage,
3291 .writepages = ext4_writepages,
3292 .write_begin = ext4_write_begin,
3293 .write_end = ext4_journalled_write_end,
3294 .set_page_dirty = ext4_journalled_set_page_dirty,
3295 .bmap = ext4_bmap,
3296 .invalidatepage = ext4_journalled_invalidatepage,
3297 .releasepage = ext4_releasepage,
3298 .direct_IO = ext4_direct_IO,
3299 .is_partially_uptodate = block_is_partially_uptodate,
3300 .error_remove_page = generic_error_remove_page,
3301};
3302
3303static const struct address_space_operations ext4_da_aops = {
3304 .readpage = ext4_readpage,
3305 .readpages = ext4_readpages,
3306 .writepage = ext4_writepage,
3307 .writepages = ext4_writepages,
3308 .write_begin = ext4_da_write_begin,
3309 .write_end = ext4_da_write_end,
3310 .bmap = ext4_bmap,
3311 .invalidatepage = ext4_da_invalidatepage,
3312 .releasepage = ext4_releasepage,
3313 .direct_IO = ext4_direct_IO,
3314 .migratepage = buffer_migrate_page,
3315 .is_partially_uptodate = block_is_partially_uptodate,
3316 .error_remove_page = generic_error_remove_page,
3317};
3318
3319void ext4_set_aops(struct inode *inode)
3320{
3321 switch (ext4_inode_journal_mode(inode)) {
3322 case EXT4_INODE_ORDERED_DATA_MODE:
3323 ext4_set_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3324 break;
3325 case EXT4_INODE_WRITEBACK_DATA_MODE:
3326 ext4_clear_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3327 break;
3328 case EXT4_INODE_JOURNAL_DATA_MODE:
3329 inode->i_mapping->a_ops = &ext4_journalled_aops;
3330 return;
3331 default:
3332 BUG();
3333 }
3334 if (test_opt(inode->i_sb, DELALLOC))
3335 inode->i_mapping->a_ops = &ext4_da_aops;
3336 else
3337 inode->i_mapping->a_ops = &ext4_aops;
3338}
3339
3340/*
3341 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3342 * starting from file offset 'from'. The range to be zero'd must
3343 * be contained with in one block. If the specified range exceeds
3344 * the end of the block it will be shortened to end of the block
3345 * that cooresponds to 'from'
3346 */
3347static int ext4_block_zero_page_range(handle_t *handle,
3348 struct address_space *mapping, loff_t from, loff_t length)
3349{
3350 ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3351 unsigned offset = from & (PAGE_CACHE_SIZE-1);
3352 unsigned blocksize, max, pos;
3353 ext4_lblk_t iblock;
3354 struct inode *inode = mapping->host;
3355 struct buffer_head *bh;
3356 struct page *page;
3357 int err = 0;
3358
3359 page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3360 mapping_gfp_mask(mapping) & ~__GFP_FS);
3361 if (!page)
3362 return -ENOMEM;
3363
3364 blocksize = inode->i_sb->s_blocksize;
3365 max = blocksize - (offset & (blocksize - 1));
3366
3367 /*
3368 * correct length if it does not fall between
3369 * 'from' and the end of the block
3370 */
3371 if (length > max || length < 0)
3372 length = max;
3373
3374 iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3375
3376 if (!page_has_buffers(page))
3377 create_empty_buffers(page, blocksize, 0);
3378
3379 /* Find the buffer that contains "offset" */
3380 bh = page_buffers(page);
3381 pos = blocksize;
3382 while (offset >= pos) {
3383 bh = bh->b_this_page;
3384 iblock++;
3385 pos += blocksize;
3386 }
3387 if (buffer_freed(bh)) {
3388 BUFFER_TRACE(bh, "freed: skip");
3389 goto unlock;
3390 }
3391 if (!buffer_mapped(bh)) {
3392 BUFFER_TRACE(bh, "unmapped");
3393 ext4_get_block(inode, iblock, bh, 0);
3394 /* unmapped? It's a hole - nothing to do */
3395 if (!buffer_mapped(bh)) {
3396 BUFFER_TRACE(bh, "still unmapped");
3397 goto unlock;
3398 }
3399 }
3400
3401 /* Ok, it's mapped. Make sure it's up-to-date */
3402 if (PageUptodate(page))
3403 set_buffer_uptodate(bh);
3404
3405 if (!buffer_uptodate(bh)) {
3406 err = -EIO;
3407 ll_rw_block(READ, 1, &bh);
3408 wait_on_buffer(bh);
3409 /* Uhhuh. Read error. Complain and punt. */
3410 if (!buffer_uptodate(bh))
3411 goto unlock;
3412 }
3413 if (ext4_should_journal_data(inode)) {
3414 BUFFER_TRACE(bh, "get write access");
3415 err = ext4_journal_get_write_access(handle, bh);
3416 if (err)
3417 goto unlock;
3418 }
3419 zero_user(page, offset, length);
3420 BUFFER_TRACE(bh, "zeroed end of block");
3421
3422 if (ext4_should_journal_data(inode)) {
3423 err = ext4_handle_dirty_metadata(handle, inode, bh);
3424 } else {
3425 err = 0;
3426 mark_buffer_dirty(bh);
3427 if (ext4_test_inode_state(inode, EXT4_STATE_ORDERED_MODE))
3428 err = ext4_jbd2_file_inode(handle, inode);
3429 }
3430
3431unlock:
3432 unlock_page(page);
3433 page_cache_release(page);
3434 return err;
3435}
3436
3437/*
3438 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3439 * up to the end of the block which corresponds to `from'.
3440 * This required during truncate. We need to physically zero the tail end
3441 * of that block so it doesn't yield old data if the file is later grown.
3442 */
3443int ext4_block_truncate_page(handle_t *handle,
3444 struct address_space *mapping, loff_t from)
3445{
3446 unsigned offset = from & (PAGE_CACHE_SIZE-1);
3447 unsigned length;
3448 unsigned blocksize;
3449 struct inode *inode = mapping->host;
3450
3451 blocksize = inode->i_sb->s_blocksize;
3452 length = blocksize - (offset & (blocksize - 1));
3453
3454 return ext4_block_zero_page_range(handle, mapping, from, length);
3455}
3456
3457int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3458 loff_t lstart, loff_t length)
3459{
3460 struct super_block *sb = inode->i_sb;
3461 struct address_space *mapping = inode->i_mapping;
3462 unsigned partial_start, partial_end;
3463 ext4_fsblk_t start, end;
3464 loff_t byte_end = (lstart + length - 1);
3465 int err = 0;
3466
3467 partial_start = lstart & (sb->s_blocksize - 1);
3468 partial_end = byte_end & (sb->s_blocksize - 1);
3469
3470 start = lstart >> sb->s_blocksize_bits;
3471 end = byte_end >> sb->s_blocksize_bits;
3472
3473 /* Handle partial zero within the single block */
3474 if (start == end &&
3475 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3476 err = ext4_block_zero_page_range(handle, mapping,
3477 lstart, length);
3478 return err;
3479 }
3480 /* Handle partial zero out on the start of the range */
3481 if (partial_start) {
3482 err = ext4_block_zero_page_range(handle, mapping,
3483 lstart, sb->s_blocksize);
3484 if (err)
3485 return err;
3486 }
3487 /* Handle partial zero out on the end of the range */
3488 if (partial_end != sb->s_blocksize - 1)
3489 err = ext4_block_zero_page_range(handle, mapping,
3490 byte_end - partial_end,
3491 partial_end + 1);
3492 return err;
3493}
3494
3495int ext4_can_truncate(struct inode *inode)
3496{
3497 if (S_ISREG(inode->i_mode))
3498 return 1;
3499 if (S_ISDIR(inode->i_mode))
3500 return 1;
3501 if (S_ISLNK(inode->i_mode))
3502 return !ext4_inode_is_fast_symlink(inode);
3503 return 0;
3504}
3505
3506/*
3507 * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3508 * associated with the given offset and length
3509 *
3510 * @inode: File inode
3511 * @offset: The offset where the hole will begin
3512 * @len: The length of the hole
3513 *
3514 * Returns: 0 on success or negative on failure
3515 */
3516
3517int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
3518{
3519 struct super_block *sb = inode->i_sb;
3520 ext4_lblk_t first_block, stop_block;
3521 struct address_space *mapping = inode->i_mapping;
3522 loff_t first_block_offset, last_block_offset;
3523 handle_t *handle;
3524 unsigned int credits;
3525 int ret = 0;
3526
3527 if (!S_ISREG(inode->i_mode))
3528 return -EOPNOTSUPP;
3529
3530 trace_ext4_punch_hole(inode, offset, length, 0);
3531
3532 /*
3533 * Write out all dirty pages to avoid race conditions
3534 * Then release them.
3535 */
3536 if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3537 ret = filemap_write_and_wait_range(mapping, offset,
3538 offset + length - 1);
3539 if (ret)
3540 return ret;
3541 }
3542
3543 mutex_lock(&inode->i_mutex);
3544
3545 /* No need to punch hole beyond i_size */
3546 if (offset >= inode->i_size)
3547 goto out_mutex;
3548
3549 /*
3550 * If the hole extends beyond i_size, set the hole
3551 * to end after the page that contains i_size
3552 */
3553 if (offset + length > inode->i_size) {
3554 length = inode->i_size +
3555 PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
3556 offset;
3557 }
3558
3559 if (offset & (sb->s_blocksize - 1) ||
3560 (offset + length) & (sb->s_blocksize - 1)) {
3561 /*
3562 * Attach jinode to inode for jbd2 if we do any zeroing of
3563 * partial block
3564 */
3565 ret = ext4_inode_attach_jinode(inode);
3566 if (ret < 0)
3567 goto out_mutex;
3568
3569 }
3570
3571 first_block_offset = round_up(offset, sb->s_blocksize);
3572 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
3573
3574 /* Now release the pages and zero block aligned part of pages*/
3575 if (last_block_offset > first_block_offset)
3576 truncate_pagecache_range(inode, first_block_offset,
3577 last_block_offset);
3578
3579 /* Wait all existing dio workers, newcomers will block on i_mutex */
3580 ext4_inode_block_unlocked_dio(inode);
3581 inode_dio_wait(inode);
3582
3583 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3584 credits = ext4_writepage_trans_blocks(inode);
3585 else
3586 credits = ext4_blocks_for_truncate(inode);
3587 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3588 if (IS_ERR(handle)) {
3589 ret = PTR_ERR(handle);
3590 ext4_std_error(sb, ret);
3591 goto out_dio;
3592 }
3593
3594 ret = ext4_zero_partial_blocks(handle, inode, offset,
3595 length);
3596 if (ret)
3597 goto out_stop;
3598
3599 first_block = (offset + sb->s_blocksize - 1) >>
3600 EXT4_BLOCK_SIZE_BITS(sb);
3601 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
3602
3603 /* If there are no blocks to remove, return now */
3604 if (first_block >= stop_block)
3605 goto out_stop;
3606
3607 down_write(&EXT4_I(inode)->i_data_sem);
3608 ext4_discard_preallocations(inode);
3609
3610 ret = ext4_es_remove_extent(inode, first_block,
3611 stop_block - first_block);
3612 if (ret) {
3613 up_write(&EXT4_I(inode)->i_data_sem);
3614 goto out_stop;
3615 }
3616
3617 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3618 ret = ext4_ext_remove_space(inode, first_block,
3619 stop_block - 1);
3620 else
3621 ret = ext4_free_hole_blocks(handle, inode, first_block,
3622 stop_block);
3623
3624 up_write(&EXT4_I(inode)->i_data_sem);
3625 if (IS_SYNC(inode))
3626 ext4_handle_sync(handle);
3627
3628 /* Now release the pages again to reduce race window */
3629 if (last_block_offset > first_block_offset)
3630 truncate_pagecache_range(inode, first_block_offset,
3631 last_block_offset);
3632
3633 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3634 ext4_mark_inode_dirty(handle, inode);
3635out_stop:
3636 ext4_journal_stop(handle);
3637out_dio:
3638 ext4_inode_resume_unlocked_dio(inode);
3639out_mutex:
3640 mutex_unlock(&inode->i_mutex);
3641 return ret;
3642}
3643
3644int ext4_inode_attach_jinode(struct inode *inode)
3645{
3646 struct ext4_inode_info *ei = EXT4_I(inode);
3647 struct jbd2_inode *jinode;
3648
3649 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
3650 return 0;
3651
3652 jinode = jbd2_alloc_inode(GFP_KERNEL);
3653 spin_lock(&inode->i_lock);
3654 if (!ei->jinode) {
3655 if (!jinode) {
3656 spin_unlock(&inode->i_lock);
3657 return -ENOMEM;
3658 }
3659 ei->jinode = jinode;
3660 jbd2_journal_init_jbd_inode(ei->jinode, inode);
3661 jinode = NULL;
3662 }
3663 spin_unlock(&inode->i_lock);
3664 if (unlikely(jinode != NULL))
3665 jbd2_free_inode(jinode);
3666 return 0;
3667}
3668
3669/*
3670 * ext4_truncate()
3671 *
3672 * We block out ext4_get_block() block instantiations across the entire
3673 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3674 * simultaneously on behalf of the same inode.
3675 *
3676 * As we work through the truncate and commit bits of it to the journal there
3677 * is one core, guiding principle: the file's tree must always be consistent on
3678 * disk. We must be able to restart the truncate after a crash.
3679 *
3680 * The file's tree may be transiently inconsistent in memory (although it
3681 * probably isn't), but whenever we close off and commit a journal transaction,
3682 * the contents of (the filesystem + the journal) must be consistent and
3683 * restartable. It's pretty simple, really: bottom up, right to left (although
3684 * left-to-right works OK too).
3685 *
3686 * Note that at recovery time, journal replay occurs *before* the restart of
3687 * truncate against the orphan inode list.
3688 *
3689 * The committed inode has the new, desired i_size (which is the same as
3690 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
3691 * that this inode's truncate did not complete and it will again call
3692 * ext4_truncate() to have another go. So there will be instantiated blocks
3693 * to the right of the truncation point in a crashed ext4 filesystem. But
3694 * that's fine - as long as they are linked from the inode, the post-crash
3695 * ext4_truncate() run will find them and release them.
3696 */
3697void ext4_truncate(struct inode *inode)
3698{
3699 struct ext4_inode_info *ei = EXT4_I(inode);
3700 unsigned int credits;
3701 handle_t *handle;
3702 struct address_space *mapping = inode->i_mapping;
3703
3704 /*
3705 * There is a possibility that we're either freeing the inode
3706 * or it's a completely new inode. In those cases we might not
3707 * have i_mutex locked because it's not necessary.
3708 */
3709 if (!(inode->i_state & (I_NEW|I_FREEING)))
3710 WARN_ON(!mutex_is_locked(&inode->i_mutex));
3711 trace_ext4_truncate_enter(inode);
3712
3713 if (!ext4_can_truncate(inode))
3714 return;
3715
3716 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3717
3718 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3719 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3720
3721 if (ext4_has_inline_data(inode)) {
3722 int has_inline = 1;
3723
3724 ext4_inline_data_truncate(inode, &has_inline);
3725 if (has_inline)
3726 return;
3727 }
3728
3729 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
3730 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
3731 if (ext4_inode_attach_jinode(inode) < 0)
3732 return;
3733 }
3734
3735 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3736 credits = ext4_writepage_trans_blocks(inode);
3737 else
3738 credits = ext4_blocks_for_truncate(inode);
3739
3740 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3741 if (IS_ERR(handle)) {
3742 ext4_std_error(inode->i_sb, PTR_ERR(handle));
3743 return;
3744 }
3745
3746 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
3747 ext4_block_truncate_page(handle, mapping, inode->i_size);
3748
3749 /*
3750 * We add the inode to the orphan list, so that if this
3751 * truncate spans multiple transactions, and we crash, we will
3752 * resume the truncate when the filesystem recovers. It also
3753 * marks the inode dirty, to catch the new size.
3754 *
3755 * Implication: the file must always be in a sane, consistent
3756 * truncatable state while each transaction commits.
3757 */
3758 if (ext4_orphan_add(handle, inode))
3759 goto out_stop;
3760
3761 down_write(&EXT4_I(inode)->i_data_sem);
3762
3763 ext4_discard_preallocations(inode);
3764
3765 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3766 ext4_ext_truncate(handle, inode);
3767 else
3768 ext4_ind_truncate(handle, inode);
3769
3770 up_write(&ei->i_data_sem);
3771
3772 if (IS_SYNC(inode))
3773 ext4_handle_sync(handle);
3774
3775out_stop:
3776 /*
3777 * If this was a simple ftruncate() and the file will remain alive,
3778 * then we need to clear up the orphan record which we created above.
3779 * However, if this was a real unlink then we were called by
3780 * ext4_delete_inode(), and we allow that function to clean up the
3781 * orphan info for us.
3782 */
3783 if (inode->i_nlink)
3784 ext4_orphan_del(handle, inode);
3785
3786 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3787 ext4_mark_inode_dirty(handle, inode);
3788 ext4_journal_stop(handle);
3789
3790 trace_ext4_truncate_exit(inode);
3791}
3792
3793/*
3794 * ext4_get_inode_loc returns with an extra refcount against the inode's
3795 * underlying buffer_head on success. If 'in_mem' is true, we have all
3796 * data in memory that is needed to recreate the on-disk version of this
3797 * inode.
3798 */
3799static int __ext4_get_inode_loc(struct inode *inode,
3800 struct ext4_iloc *iloc, int in_mem)
3801{
3802 struct ext4_group_desc *gdp;
3803 struct buffer_head *bh;
3804 struct super_block *sb = inode->i_sb;
3805 ext4_fsblk_t block;
3806 int inodes_per_block, inode_offset;
3807
3808 iloc->bh = NULL;
3809 if (!ext4_valid_inum(sb, inode->i_ino))
3810 return -EIO;
3811
3812 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3813 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3814 if (!gdp)
3815 return -EIO;
3816
3817 /*
3818 * Figure out the offset within the block group inode table
3819 */
3820 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3821 inode_offset = ((inode->i_ino - 1) %
3822 EXT4_INODES_PER_GROUP(sb));
3823 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3824 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3825
3826 bh = sb_getblk(sb, block);
3827 if (unlikely(!bh))
3828 return -ENOMEM;
3829 if (!buffer_uptodate(bh)) {
3830 lock_buffer(bh);
3831
3832 /*
3833 * If the buffer has the write error flag, we have failed
3834 * to write out another inode in the same block. In this
3835 * case, we don't have to read the block because we may
3836 * read the old inode data successfully.
3837 */
3838 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3839 set_buffer_uptodate(bh);
3840
3841 if (buffer_uptodate(bh)) {
3842 /* someone brought it uptodate while we waited */
3843 unlock_buffer(bh);
3844 goto has_buffer;
3845 }
3846
3847 /*
3848 * If we have all information of the inode in memory and this
3849 * is the only valid inode in the block, we need not read the
3850 * block.
3851 */
3852 if (in_mem) {
3853 struct buffer_head *bitmap_bh;
3854 int i, start;
3855
3856 start = inode_offset & ~(inodes_per_block - 1);
3857
3858 /* Is the inode bitmap in cache? */
3859 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3860 if (unlikely(!bitmap_bh))
3861 goto make_io;
3862
3863 /*
3864 * If the inode bitmap isn't in cache then the
3865 * optimisation may end up performing two reads instead
3866 * of one, so skip it.
3867 */
3868 if (!buffer_uptodate(bitmap_bh)) {
3869 brelse(bitmap_bh);
3870 goto make_io;
3871 }
3872 for (i = start; i < start + inodes_per_block; i++) {
3873 if (i == inode_offset)
3874 continue;
3875 if (ext4_test_bit(i, bitmap_bh->b_data))
3876 break;
3877 }
3878 brelse(bitmap_bh);
3879 if (i == start + inodes_per_block) {
3880 /* all other inodes are free, so skip I/O */
3881 memset(bh->b_data, 0, bh->b_size);
3882 set_buffer_uptodate(bh);
3883 unlock_buffer(bh);
3884 goto has_buffer;
3885 }
3886 }
3887
3888make_io:
3889 /*
3890 * If we need to do any I/O, try to pre-readahead extra
3891 * blocks from the inode table.
3892 */
3893 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3894 ext4_fsblk_t b, end, table;
3895 unsigned num;
3896 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
3897
3898 table = ext4_inode_table(sb, gdp);
3899 /* s_inode_readahead_blks is always a power of 2 */
3900 b = block & ~((ext4_fsblk_t) ra_blks - 1);
3901 if (table > b)
3902 b = table;
3903 end = b + ra_blks;
3904 num = EXT4_INODES_PER_GROUP(sb);
3905 if (ext4_has_group_desc_csum(sb))
3906 num -= ext4_itable_unused_count(sb, gdp);
3907 table += num / inodes_per_block;
3908 if (end > table)
3909 end = table;
3910 while (b <= end)
3911 sb_breadahead(sb, b++);
3912 }
3913
3914 /*
3915 * There are other valid inodes in the buffer, this inode
3916 * has in-inode xattrs, or we don't have this inode in memory.
3917 * Read the block from disk.
3918 */
3919 trace_ext4_load_inode(inode);
3920 get_bh(bh);
3921 bh->b_end_io = end_buffer_read_sync;
3922 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3923 wait_on_buffer(bh);
3924 if (!buffer_uptodate(bh)) {
3925 EXT4_ERROR_INODE_BLOCK(inode, block,
3926 "unable to read itable block");
3927 brelse(bh);
3928 return -EIO;
3929 }
3930 }
3931has_buffer:
3932 iloc->bh = bh;
3933 return 0;
3934}
3935
3936int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3937{
3938 /* We have all inode data except xattrs in memory here. */
3939 return __ext4_get_inode_loc(inode, iloc,
3940 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3941}
3942
3943void ext4_set_inode_flags(struct inode *inode)
3944{
3945 unsigned int flags = EXT4_I(inode)->i_flags;
3946 unsigned int new_fl = 0;
3947
3948 if (flags & EXT4_SYNC_FL)
3949 new_fl |= S_SYNC;
3950 if (flags & EXT4_APPEND_FL)
3951 new_fl |= S_APPEND;
3952 if (flags & EXT4_IMMUTABLE_FL)
3953 new_fl |= S_IMMUTABLE;
3954 if (flags & EXT4_NOATIME_FL)
3955 new_fl |= S_NOATIME;
3956 if (flags & EXT4_DIRSYNC_FL)
3957 new_fl |= S_DIRSYNC;
3958 inode_set_flags(inode, new_fl,
3959 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3960}
3961
3962/* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3963void ext4_get_inode_flags(struct ext4_inode_info *ei)
3964{
3965 unsigned int vfs_fl;
3966 unsigned long old_fl, new_fl;
3967
3968 do {
3969 vfs_fl = ei->vfs_inode.i_flags;
3970 old_fl = ei->i_flags;
3971 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3972 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3973 EXT4_DIRSYNC_FL);
3974 if (vfs_fl & S_SYNC)
3975 new_fl |= EXT4_SYNC_FL;
3976 if (vfs_fl & S_APPEND)
3977 new_fl |= EXT4_APPEND_FL;
3978 if (vfs_fl & S_IMMUTABLE)
3979 new_fl |= EXT4_IMMUTABLE_FL;
3980 if (vfs_fl & S_NOATIME)
3981 new_fl |= EXT4_NOATIME_FL;
3982 if (vfs_fl & S_DIRSYNC)
3983 new_fl |= EXT4_DIRSYNC_FL;
3984 } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3985}
3986
3987static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3988 struct ext4_inode_info *ei)
3989{
3990 blkcnt_t i_blocks ;
3991 struct inode *inode = &(ei->vfs_inode);
3992 struct super_block *sb = inode->i_sb;
3993
3994 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3995 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3996 /* we are using combined 48 bit field */
3997 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3998 le32_to_cpu(raw_inode->i_blocks_lo);
3999 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4000 /* i_blocks represent file system block size */
4001 return i_blocks << (inode->i_blkbits - 9);
4002 } else {
4003 return i_blocks;
4004 }
4005 } else {
4006 return le32_to_cpu(raw_inode->i_blocks_lo);
4007 }
4008}
4009
4010static inline void ext4_iget_extra_inode(struct inode *inode,
4011 struct ext4_inode *raw_inode,
4012 struct ext4_inode_info *ei)
4013{
4014 __le32 *magic = (void *)raw_inode +
4015 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4016 if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4017 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4018 ext4_find_inline_data_nolock(inode);
4019 } else
4020 EXT4_I(inode)->i_inline_off = 0;
4021}
4022
4023struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4024{
4025 struct ext4_iloc iloc;
4026 struct ext4_inode *raw_inode;
4027 struct ext4_inode_info *ei;
4028 struct inode *inode;
4029 journal_t *journal = EXT4_SB(sb)->s_journal;
4030 long ret;
4031 int block;
4032 uid_t i_uid;
4033 gid_t i_gid;
4034
4035 inode = iget_locked(sb, ino);
4036 if (!inode)
4037 return ERR_PTR(-ENOMEM);
4038 if (!(inode->i_state & I_NEW))
4039 return inode;
4040
4041 ei = EXT4_I(inode);
4042 iloc.bh = NULL;
4043
4044 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4045 if (ret < 0)
4046 goto bad_inode;
4047 raw_inode = ext4_raw_inode(&iloc);
4048
4049 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4050 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4051 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4052 EXT4_INODE_SIZE(inode->i_sb)) {
4053 EXT4_ERROR_INODE(inode, "bad extra_isize (%u != %u)",
4054 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize,
4055 EXT4_INODE_SIZE(inode->i_sb));
4056 ret = -EIO;
4057 goto bad_inode;
4058 }
4059 } else
4060 ei->i_extra_isize = 0;
4061
4062 /* Precompute checksum seed for inode metadata */
4063 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4064 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM)) {
4065 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4066 __u32 csum;
4067 __le32 inum = cpu_to_le32(inode->i_ino);
4068 __le32 gen = raw_inode->i_generation;
4069 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4070 sizeof(inum));
4071 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4072 sizeof(gen));
4073 }
4074
4075 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4076 EXT4_ERROR_INODE(inode, "checksum invalid");
4077 ret = -EIO;
4078 goto bad_inode;
4079 }
4080
4081 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4082 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4083 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4084 if (!(test_opt(inode->i_sb, NO_UID32))) {
4085 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4086 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4087 }
4088 i_uid_write(inode, i_uid);
4089 i_gid_write(inode, i_gid);
4090 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4091
4092 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4093 ei->i_inline_off = 0;
4094 ei->i_dir_start_lookup = 0;
4095 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4096 /* We now have enough fields to check if the inode was active or not.
4097 * This is needed because nfsd might try to access dead inodes
4098 * the test is that same one that e2fsck uses
4099 * NeilBrown 1999oct15
4100 */
4101 if (inode->i_nlink == 0) {
4102 if ((inode->i_mode == 0 ||
4103 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4104 ino != EXT4_BOOT_LOADER_INO) {
4105 /* this inode is deleted */
4106 ret = -ESTALE;
4107 goto bad_inode;
4108 }
4109 /* The only unlinked inodes we let through here have
4110 * valid i_mode and are being read by the orphan
4111 * recovery code: that's fine, we're about to complete
4112 * the process of deleting those.
4113 * OR it is the EXT4_BOOT_LOADER_INO which is
4114 * not initialized on a new filesystem. */
4115 }
4116 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4117 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4118 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4119 if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4120 ei->i_file_acl |=
4121 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4122 inode->i_size = ext4_isize(raw_inode);
4123 ei->i_disksize = inode->i_size;
4124#ifdef CONFIG_QUOTA
4125 ei->i_reserved_quota = 0;
4126#endif
4127 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4128 ei->i_block_group = iloc.block_group;
4129 ei->i_last_alloc_group = ~0;
4130 /*
4131 * NOTE! The in-memory inode i_data array is in little-endian order
4132 * even on big-endian machines: we do NOT byteswap the block numbers!
4133 */
4134 for (block = 0; block < EXT4_N_BLOCKS; block++)
4135 ei->i_data[block] = raw_inode->i_block[block];
4136 INIT_LIST_HEAD(&ei->i_orphan);
4137
4138 /*
4139 * Set transaction id's of transactions that have to be committed
4140 * to finish f[data]sync. We set them to currently running transaction
4141 * as we cannot be sure that the inode or some of its metadata isn't
4142 * part of the transaction - the inode could have been reclaimed and
4143 * now it is reread from disk.
4144 */
4145 if (journal) {
4146 transaction_t *transaction;
4147 tid_t tid;
4148
4149 read_lock(&journal->j_state_lock);
4150 if (journal->j_running_transaction)
4151 transaction = journal->j_running_transaction;
4152 else
4153 transaction = journal->j_committing_transaction;
4154 if (transaction)
4155 tid = transaction->t_tid;
4156 else
4157 tid = journal->j_commit_sequence;
4158 read_unlock(&journal->j_state_lock);
4159 ei->i_sync_tid = tid;
4160 ei->i_datasync_tid = tid;
4161 }
4162
4163 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4164 if (ei->i_extra_isize == 0) {
4165 /* The extra space is currently unused. Use it. */
4166 ei->i_extra_isize = sizeof(struct ext4_inode) -
4167 EXT4_GOOD_OLD_INODE_SIZE;
4168 } else {
4169 ext4_iget_extra_inode(inode, raw_inode, ei);
4170 }
4171 }
4172
4173 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4174 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4175 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4176 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4177
4178 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4179 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4180 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4181 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4182 inode->i_version |=
4183 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4184 }
4185 }
4186
4187 ret = 0;
4188 if (ei->i_file_acl &&
4189 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4190 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4191 ei->i_file_acl);
4192 ret = -EIO;
4193 goto bad_inode;
4194 } else if (!ext4_has_inline_data(inode)) {
4195 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4196 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4197 (S_ISLNK(inode->i_mode) &&
4198 !ext4_inode_is_fast_symlink(inode))))
4199 /* Validate extent which is part of inode */
4200 ret = ext4_ext_check_inode(inode);
4201 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4202 (S_ISLNK(inode->i_mode) &&
4203 !ext4_inode_is_fast_symlink(inode))) {
4204 /* Validate block references which are part of inode */
4205 ret = ext4_ind_check_inode(inode);
4206 }
4207 }
4208 if (ret)
4209 goto bad_inode;
4210
4211 if (S_ISREG(inode->i_mode)) {
4212 inode->i_op = &ext4_file_inode_operations;
4213 inode->i_fop = &ext4_file_operations;
4214 ext4_set_aops(inode);
4215 } else if (S_ISDIR(inode->i_mode)) {
4216 inode->i_op = &ext4_dir_inode_operations;
4217 inode->i_fop = &ext4_dir_operations;
4218 } else if (S_ISLNK(inode->i_mode)) {
4219 if (ext4_inode_is_fast_symlink(inode)) {
4220 inode->i_op = &ext4_fast_symlink_inode_operations;
4221 nd_terminate_link(ei->i_data, inode->i_size,
4222 sizeof(ei->i_data) - 1);
4223 } else {
4224 inode->i_op = &ext4_symlink_inode_operations;
4225 ext4_set_aops(inode);
4226 }
4227 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4228 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4229 inode->i_op = &ext4_special_inode_operations;
4230 if (raw_inode->i_block[0])
4231 init_special_inode(inode, inode->i_mode,
4232 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4233 else
4234 init_special_inode(inode, inode->i_mode,
4235 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4236 } else if (ino == EXT4_BOOT_LOADER_INO) {
4237 make_bad_inode(inode);
4238 } else {
4239 ret = -EIO;
4240 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
4241 goto bad_inode;
4242 }
4243 brelse(iloc.bh);
4244 ext4_set_inode_flags(inode);
4245 unlock_new_inode(inode);
4246 return inode;
4247
4248bad_inode:
4249 brelse(iloc.bh);
4250 iget_failed(inode);
4251 return ERR_PTR(ret);
4252}
4253
4254static int ext4_inode_blocks_set(handle_t *handle,
4255 struct ext4_inode *raw_inode,
4256 struct ext4_inode_info *ei)
4257{
4258 struct inode *inode = &(ei->vfs_inode);
4259 u64 i_blocks = inode->i_blocks;
4260 struct super_block *sb = inode->i_sb;
4261
4262 if (i_blocks <= ~0U) {
4263 /*
4264 * i_blocks can be represented in a 32 bit variable
4265 * as multiple of 512 bytes
4266 */
4267 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4268 raw_inode->i_blocks_high = 0;
4269 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4270 return 0;
4271 }
4272 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4273 return -EFBIG;
4274
4275 if (i_blocks <= 0xffffffffffffULL) {
4276 /*
4277 * i_blocks can be represented in a 48 bit variable
4278 * as multiple of 512 bytes
4279 */
4280 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4281 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4282 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4283 } else {
4284 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4285 /* i_block is stored in file system block size */
4286 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4287 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4288 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4289 }
4290 return 0;
4291}
4292
4293/*
4294 * Post the struct inode info into an on-disk inode location in the
4295 * buffer-cache. This gobbles the caller's reference to the
4296 * buffer_head in the inode location struct.
4297 *
4298 * The caller must have write access to iloc->bh.
4299 */
4300static int ext4_do_update_inode(handle_t *handle,
4301 struct inode *inode,
4302 struct ext4_iloc *iloc)
4303{
4304 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4305 struct ext4_inode_info *ei = EXT4_I(inode);
4306 struct buffer_head *bh = iloc->bh;
4307 int err = 0, rc, block;
4308 int need_datasync = 0;
4309 uid_t i_uid;
4310 gid_t i_gid;
4311
4312 /* For fields not not tracking in the in-memory inode,
4313 * initialise them to zero for new inodes. */
4314 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
4315 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4316
4317 ext4_get_inode_flags(ei);
4318 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4319 i_uid = i_uid_read(inode);
4320 i_gid = i_gid_read(inode);
4321 if (!(test_opt(inode->i_sb, NO_UID32))) {
4322 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4323 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4324/*
4325 * Fix up interoperability with old kernels. Otherwise, old inodes get
4326 * re-used with the upper 16 bits of the uid/gid intact
4327 */
4328 if (!ei->i_dtime) {
4329 raw_inode->i_uid_high =
4330 cpu_to_le16(high_16_bits(i_uid));
4331 raw_inode->i_gid_high =
4332 cpu_to_le16(high_16_bits(i_gid));
4333 } else {
4334 raw_inode->i_uid_high = 0;
4335 raw_inode->i_gid_high = 0;
4336 }
4337 } else {
4338 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4339 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4340 raw_inode->i_uid_high = 0;
4341 raw_inode->i_gid_high = 0;
4342 }
4343 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4344
4345 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4346 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4347 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4348 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4349
4350 if (ext4_inode_blocks_set(handle, raw_inode, ei))
4351 goto out_brelse;
4352 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4353 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4354 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4355 raw_inode->i_file_acl_high =
4356 cpu_to_le16(ei->i_file_acl >> 32);
4357 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4358 if (ei->i_disksize != ext4_isize(raw_inode)) {
4359 ext4_isize_set(raw_inode, ei->i_disksize);
4360 need_datasync = 1;
4361 }
4362 if (ei->i_disksize > 0x7fffffffULL) {
4363 struct super_block *sb = inode->i_sb;
4364 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4365 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4366 EXT4_SB(sb)->s_es->s_rev_level ==
4367 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4368 /* If this is the first large file
4369 * created, add a flag to the superblock.
4370 */
4371 err = ext4_journal_get_write_access(handle,
4372 EXT4_SB(sb)->s_sbh);
4373 if (err)
4374 goto out_brelse;
4375 ext4_update_dynamic_rev(sb);
4376 EXT4_SET_RO_COMPAT_FEATURE(sb,
4377 EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4378 ext4_handle_sync(handle);
4379 err = ext4_handle_dirty_super(handle, sb);
4380 }
4381 }
4382 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4383 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4384 if (old_valid_dev(inode->i_rdev)) {
4385 raw_inode->i_block[0] =
4386 cpu_to_le32(old_encode_dev(inode->i_rdev));
4387 raw_inode->i_block[1] = 0;
4388 } else {
4389 raw_inode->i_block[0] = 0;
4390 raw_inode->i_block[1] =
4391 cpu_to_le32(new_encode_dev(inode->i_rdev));
4392 raw_inode->i_block[2] = 0;
4393 }
4394 } else if (!ext4_has_inline_data(inode)) {
4395 for (block = 0; block < EXT4_N_BLOCKS; block++)
4396 raw_inode->i_block[block] = ei->i_data[block];
4397 }
4398
4399 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4400 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4401 if (ei->i_extra_isize) {
4402 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4403 raw_inode->i_version_hi =
4404 cpu_to_le32(inode->i_version >> 32);
4405 raw_inode->i_extra_isize =
4406 cpu_to_le16(ei->i_extra_isize);
4407 }
4408 }
4409
4410 ext4_inode_csum_set(inode, raw_inode, ei);
4411
4412 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4413 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
4414 if (!err)
4415 err = rc;
4416 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
4417
4418 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
4419out_brelse:
4420 brelse(bh);
4421 ext4_std_error(inode->i_sb, err);
4422 return err;
4423}
4424
4425/*
4426 * ext4_write_inode()
4427 *
4428 * We are called from a few places:
4429 *
4430 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
4431 * Here, there will be no transaction running. We wait for any running
4432 * transaction to commit.
4433 *
4434 * - Within flush work (sys_sync(), kupdate and such).
4435 * We wait on commit, if told to.
4436 *
4437 * - Within iput_final() -> write_inode_now()
4438 * We wait on commit, if told to.
4439 *
4440 * In all cases it is actually safe for us to return without doing anything,
4441 * because the inode has been copied into a raw inode buffer in
4442 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
4443 * writeback.
4444 *
4445 * Note that we are absolutely dependent upon all inode dirtiers doing the
4446 * right thing: they *must* call mark_inode_dirty() after dirtying info in
4447 * which we are interested.
4448 *
4449 * It would be a bug for them to not do this. The code:
4450 *
4451 * mark_inode_dirty(inode)
4452 * stuff();
4453 * inode->i_size = expr;
4454 *
4455 * is in error because write_inode() could occur while `stuff()' is running,
4456 * and the new i_size will be lost. Plus the inode will no longer be on the
4457 * superblock's dirty inode list.
4458 */
4459int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
4460{
4461 int err;
4462
4463 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
4464 return 0;
4465
4466 if (EXT4_SB(inode->i_sb)->s_journal) {
4467 if (ext4_journal_current_handle()) {
4468 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4469 dump_stack();
4470 return -EIO;
4471 }
4472
4473 /*
4474 * No need to force transaction in WB_SYNC_NONE mode. Also
4475 * ext4_sync_fs() will force the commit after everything is
4476 * written.
4477 */
4478 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
4479 return 0;
4480
4481 err = ext4_force_commit(inode->i_sb);
4482 } else {
4483 struct ext4_iloc iloc;
4484
4485 err = __ext4_get_inode_loc(inode, &iloc, 0);
4486 if (err)
4487 return err;
4488 /*
4489 * sync(2) will flush the whole buffer cache. No need to do
4490 * it here separately for each inode.
4491 */
4492 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
4493 sync_dirty_buffer(iloc.bh);
4494 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
4495 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
4496 "IO error syncing inode");
4497 err = -EIO;
4498 }
4499 brelse(iloc.bh);
4500 }
4501 return err;
4502}
4503
4504/*
4505 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
4506 * buffers that are attached to a page stradding i_size and are undergoing
4507 * commit. In that case we have to wait for commit to finish and try again.
4508 */
4509static void ext4_wait_for_tail_page_commit(struct inode *inode)
4510{
4511 struct page *page;
4512 unsigned offset;
4513 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
4514 tid_t commit_tid = 0;
4515 int ret;
4516
4517 offset = inode->i_size & (PAGE_CACHE_SIZE - 1);
4518 /*
4519 * All buffers in the last page remain valid? Then there's nothing to
4520 * do. We do the check mainly to optimize the common PAGE_CACHE_SIZE ==
4521 * blocksize case
4522 */
4523 if (offset > PAGE_CACHE_SIZE - (1 << inode->i_blkbits))
4524 return;
4525 while (1) {
4526 page = find_lock_page(inode->i_mapping,
4527 inode->i_size >> PAGE_CACHE_SHIFT);
4528 if (!page)
4529 return;
4530 ret = __ext4_journalled_invalidatepage(page, offset,
4531 PAGE_CACHE_SIZE - offset);
4532 unlock_page(page);
4533 page_cache_release(page);
4534 if (ret != -EBUSY)
4535 return;
4536 commit_tid = 0;
4537 read_lock(&journal->j_state_lock);
4538 if (journal->j_committing_transaction)
4539 commit_tid = journal->j_committing_transaction->t_tid;
4540 read_unlock(&journal->j_state_lock);
4541 if (commit_tid)
4542 jbd2_log_wait_commit(journal, commit_tid);
4543 }
4544}
4545
4546/*
4547 * ext4_setattr()
4548 *
4549 * Called from notify_change.
4550 *
4551 * We want to trap VFS attempts to truncate the file as soon as
4552 * possible. In particular, we want to make sure that when the VFS
4553 * shrinks i_size, we put the inode on the orphan list and modify
4554 * i_disksize immediately, so that during the subsequent flushing of
4555 * dirty pages and freeing of disk blocks, we can guarantee that any
4556 * commit will leave the blocks being flushed in an unused state on
4557 * disk. (On recovery, the inode will get truncated and the blocks will
4558 * be freed, so we have a strong guarantee that no future commit will
4559 * leave these blocks visible to the user.)
4560 *
4561 * Another thing we have to assure is that if we are in ordered mode
4562 * and inode is still attached to the committing transaction, we must
4563 * we start writeout of all the dirty pages which are being truncated.
4564 * This way we are sure that all the data written in the previous
4565 * transaction are already on disk (truncate waits for pages under
4566 * writeback).
4567 *
4568 * Called with inode->i_mutex down.
4569 */
4570int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4571{
4572 struct inode *inode = dentry->d_inode;
4573 int error, rc = 0;
4574 int orphan = 0;
4575 const unsigned int ia_valid = attr->ia_valid;
4576
4577 error = inode_change_ok(inode, attr);
4578 if (error)
4579 return error;
4580
4581 if (is_quota_modification(inode, attr))
4582 dquot_initialize(inode);
4583 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
4584 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
4585 handle_t *handle;
4586
4587 /* (user+group)*(old+new) structure, inode write (sb,
4588 * inode block, ? - but truncate inode update has it) */
4589 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
4590 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
4591 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
4592 if (IS_ERR(handle)) {
4593 error = PTR_ERR(handle);
4594 goto err_out;
4595 }
4596 error = dquot_transfer(inode, attr);
4597 if (error) {
4598 ext4_journal_stop(handle);
4599 return error;
4600 }
4601 /* Update corresponding info in inode so that everything is in
4602 * one transaction */
4603 if (attr->ia_valid & ATTR_UID)
4604 inode->i_uid = attr->ia_uid;
4605 if (attr->ia_valid & ATTR_GID)
4606 inode->i_gid = attr->ia_gid;
4607 error = ext4_mark_inode_dirty(handle, inode);
4608 ext4_journal_stop(handle);
4609 }
4610
4611 if (attr->ia_valid & ATTR_SIZE && attr->ia_size != inode->i_size) {
4612 handle_t *handle;
4613
4614 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4615 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4616
4617 if (attr->ia_size > sbi->s_bitmap_maxbytes)
4618 return -EFBIG;
4619 }
4620
4621 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
4622 inode_inc_iversion(inode);
4623
4624 if (S_ISREG(inode->i_mode) &&
4625 (attr->ia_size < inode->i_size)) {
4626 if (ext4_should_order_data(inode)) {
4627 error = ext4_begin_ordered_truncate(inode,
4628 attr->ia_size);
4629 if (error)
4630 goto err_out;
4631 }
4632 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
4633 if (IS_ERR(handle)) {
4634 error = PTR_ERR(handle);
4635 goto err_out;
4636 }
4637 if (ext4_handle_valid(handle)) {
4638 error = ext4_orphan_add(handle, inode);
4639 orphan = 1;
4640 }
4641 down_write(&EXT4_I(inode)->i_data_sem);
4642 EXT4_I(inode)->i_disksize = attr->ia_size;
4643 rc = ext4_mark_inode_dirty(handle, inode);
4644 if (!error)
4645 error = rc;
4646 /*
4647 * We have to update i_size under i_data_sem together
4648 * with i_disksize to avoid races with writeback code
4649 * running ext4_wb_update_i_disksize().
4650 */
4651 if (!error)
4652 i_size_write(inode, attr->ia_size);
4653 up_write(&EXT4_I(inode)->i_data_sem);
4654 ext4_journal_stop(handle);
4655 if (error) {
4656 ext4_orphan_del(NULL, inode);
4657 goto err_out;
4658 }
4659 } else
4660 i_size_write(inode, attr->ia_size);
4661
4662 /*
4663 * Blocks are going to be removed from the inode. Wait
4664 * for dio in flight. Temporarily disable
4665 * dioread_nolock to prevent livelock.
4666 */
4667 if (orphan) {
4668 if (!ext4_should_journal_data(inode)) {
4669 ext4_inode_block_unlocked_dio(inode);
4670 inode_dio_wait(inode);
4671 ext4_inode_resume_unlocked_dio(inode);
4672 } else
4673 ext4_wait_for_tail_page_commit(inode);
4674 }
4675 /*
4676 * Truncate pagecache after we've waited for commit
4677 * in data=journal mode to make pages freeable.
4678 */
4679 truncate_pagecache(inode, inode->i_size);
4680 }
4681 /*
4682 * We want to call ext4_truncate() even if attr->ia_size ==
4683 * inode->i_size for cases like truncation of fallocated space
4684 */
4685 if (attr->ia_valid & ATTR_SIZE)
4686 ext4_truncate(inode);
4687
4688 if (!rc) {
4689 setattr_copy(inode, attr);
4690 mark_inode_dirty(inode);
4691 }
4692
4693 /*
4694 * If the call to ext4_truncate failed to get a transaction handle at
4695 * all, we need to clean up the in-core orphan list manually.
4696 */
4697 if (orphan && inode->i_nlink)
4698 ext4_orphan_del(NULL, inode);
4699
4700 if (!rc && (ia_valid & ATTR_MODE))
4701 rc = posix_acl_chmod(inode, inode->i_mode);
4702
4703err_out:
4704 ext4_std_error(inode->i_sb, error);
4705 if (!error)
4706 error = rc;
4707 return error;
4708}
4709
4710int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4711 struct kstat *stat)
4712{
4713 struct inode *inode;
4714 unsigned long long delalloc_blocks;
4715
4716 inode = dentry->d_inode;
4717 generic_fillattr(inode, stat);
4718
4719 /*
4720 * If there is inline data in the inode, the inode will normally not
4721 * have data blocks allocated (it may have an external xattr block).
4722 * Report at least one sector for such files, so tools like tar, rsync,
4723 * others doen't incorrectly think the file is completely sparse.
4724 */
4725 if (unlikely(ext4_has_inline_data(inode)))
4726 stat->blocks += (stat->size + 511) >> 9;
4727
4728 /*
4729 * We can't update i_blocks if the block allocation is delayed
4730 * otherwise in the case of system crash before the real block
4731 * allocation is done, we will have i_blocks inconsistent with
4732 * on-disk file blocks.
4733 * We always keep i_blocks updated together with real
4734 * allocation. But to not confuse with user, stat
4735 * will return the blocks that include the delayed allocation
4736 * blocks for this file.
4737 */
4738 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
4739 EXT4_I(inode)->i_reserved_data_blocks);
4740 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
4741 return 0;
4742}
4743
4744static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
4745 int pextents)
4746{
4747 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4748 return ext4_ind_trans_blocks(inode, lblocks);
4749 return ext4_ext_index_trans_blocks(inode, pextents);
4750}
4751
4752/*
4753 * Account for index blocks, block groups bitmaps and block group
4754 * descriptor blocks if modify datablocks and index blocks
4755 * worse case, the indexs blocks spread over different block groups
4756 *
4757 * If datablocks are discontiguous, they are possible to spread over
4758 * different block groups too. If they are contiguous, with flexbg,
4759 * they could still across block group boundary.
4760 *
4761 * Also account for superblock, inode, quota and xattr blocks
4762 */
4763static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
4764 int pextents)
4765{
4766 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4767 int gdpblocks;
4768 int idxblocks;
4769 int ret = 0;
4770
4771 /*
4772 * How many index blocks need to touch to map @lblocks logical blocks
4773 * to @pextents physical extents?
4774 */
4775 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
4776
4777 ret = idxblocks;
4778
4779 /*
4780 * Now let's see how many group bitmaps and group descriptors need
4781 * to account
4782 */
4783 groups = idxblocks + pextents;
4784 gdpblocks = groups;
4785 if (groups > ngroups)
4786 groups = ngroups;
4787 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4788 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4789
4790 /* bitmaps and block group descriptor blocks */
4791 ret += groups + gdpblocks;
4792
4793 /* Blocks for super block, inode, quota and xattr blocks */
4794 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4795
4796 return ret;
4797}
4798
4799/*
4800 * Calculate the total number of credits to reserve to fit
4801 * the modification of a single pages into a single transaction,
4802 * which may include multiple chunks of block allocations.
4803 *
4804 * This could be called via ext4_write_begin()
4805 *
4806 * We need to consider the worse case, when
4807 * one new block per extent.
4808 */
4809int ext4_writepage_trans_blocks(struct inode *inode)
4810{
4811 int bpp = ext4_journal_blocks_per_page(inode);
4812 int ret;
4813
4814 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
4815
4816 /* Account for data blocks for journalled mode */
4817 if (ext4_should_journal_data(inode))
4818 ret += bpp;
4819 return ret;
4820}
4821
4822/*
4823 * Calculate the journal credits for a chunk of data modification.
4824 *
4825 * This is called from DIO, fallocate or whoever calling
4826 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4827 *
4828 * journal buffers for data blocks are not included here, as DIO
4829 * and fallocate do no need to journal data buffers.
4830 */
4831int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4832{
4833 return ext4_meta_trans_blocks(inode, nrblocks, 1);
4834}
4835
4836/*
4837 * The caller must have previously called ext4_reserve_inode_write().
4838 * Give this, we know that the caller already has write access to iloc->bh.
4839 */
4840int ext4_mark_iloc_dirty(handle_t *handle,
4841 struct inode *inode, struct ext4_iloc *iloc)
4842{
4843 int err = 0;
4844
4845 if (IS_I_VERSION(inode))
4846 inode_inc_iversion(inode);
4847
4848 /* the do_update_inode consumes one bh->b_count */
4849 get_bh(iloc->bh);
4850
4851 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4852 err = ext4_do_update_inode(handle, inode, iloc);
4853 put_bh(iloc->bh);
4854 return err;
4855}
4856
4857/*
4858 * On success, We end up with an outstanding reference count against
4859 * iloc->bh. This _must_ be cleaned up later.
4860 */
4861
4862int
4863ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4864 struct ext4_iloc *iloc)
4865{
4866 int err;
4867
4868 err = ext4_get_inode_loc(inode, iloc);
4869 if (!err) {
4870 BUFFER_TRACE(iloc->bh, "get_write_access");
4871 err = ext4_journal_get_write_access(handle, iloc->bh);
4872 if (err) {
4873 brelse(iloc->bh);
4874 iloc->bh = NULL;
4875 }
4876 }
4877 ext4_std_error(inode->i_sb, err);
4878 return err;
4879}
4880
4881/*
4882 * Expand an inode by new_extra_isize bytes.
4883 * Returns 0 on success or negative error number on failure.
4884 */
4885static int ext4_expand_extra_isize(struct inode *inode,
4886 unsigned int new_extra_isize,
4887 struct ext4_iloc iloc,
4888 handle_t *handle)
4889{
4890 struct ext4_inode *raw_inode;
4891 struct ext4_xattr_ibody_header *header;
4892
4893 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4894 return 0;
4895
4896 raw_inode = ext4_raw_inode(&iloc);
4897
4898 header = IHDR(inode, raw_inode);
4899
4900 /* No extended attributes present */
4901 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4902 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4903 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4904 new_extra_isize);
4905 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4906 return 0;
4907 }
4908
4909 /* try to expand with EAs present */
4910 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4911 raw_inode, handle);
4912}
4913
4914/*
4915 * What we do here is to mark the in-core inode as clean with respect to inode
4916 * dirtiness (it may still be data-dirty).
4917 * This means that the in-core inode may be reaped by prune_icache
4918 * without having to perform any I/O. This is a very good thing,
4919 * because *any* task may call prune_icache - even ones which
4920 * have a transaction open against a different journal.
4921 *
4922 * Is this cheating? Not really. Sure, we haven't written the
4923 * inode out, but prune_icache isn't a user-visible syncing function.
4924 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4925 * we start and wait on commits.
4926 */
4927int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4928{
4929 struct ext4_iloc iloc;
4930 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4931 static unsigned int mnt_count;
4932 int err, ret;
4933
4934 might_sleep();
4935 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4936 err = ext4_reserve_inode_write(handle, inode, &iloc);
4937 if (ext4_handle_valid(handle) &&
4938 EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4939 !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4940 /*
4941 * We need extra buffer credits since we may write into EA block
4942 * with this same handle. If journal_extend fails, then it will
4943 * only result in a minor loss of functionality for that inode.
4944 * If this is felt to be critical, then e2fsck should be run to
4945 * force a large enough s_min_extra_isize.
4946 */
4947 if ((jbd2_journal_extend(handle,
4948 EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4949 ret = ext4_expand_extra_isize(inode,
4950 sbi->s_want_extra_isize,
4951 iloc, handle);
4952 if (ret) {
4953 ext4_set_inode_state(inode,
4954 EXT4_STATE_NO_EXPAND);
4955 if (mnt_count !=
4956 le16_to_cpu(sbi->s_es->s_mnt_count)) {
4957 ext4_warning(inode->i_sb,
4958 "Unable to expand inode %lu. Delete"
4959 " some EAs or run e2fsck.",
4960 inode->i_ino);
4961 mnt_count =
4962 le16_to_cpu(sbi->s_es->s_mnt_count);
4963 }
4964 }
4965 }
4966 }
4967 if (!err)
4968 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4969 return err;
4970}
4971
4972/*
4973 * ext4_dirty_inode() is called from __mark_inode_dirty()
4974 *
4975 * We're really interested in the case where a file is being extended.
4976 * i_size has been changed by generic_commit_write() and we thus need
4977 * to include the updated inode in the current transaction.
4978 *
4979 * Also, dquot_alloc_block() will always dirty the inode when blocks
4980 * are allocated to the file.
4981 *
4982 * If the inode is marked synchronous, we don't honour that here - doing
4983 * so would cause a commit on atime updates, which we don't bother doing.
4984 * We handle synchronous inodes at the highest possible level.
4985 */
4986void ext4_dirty_inode(struct inode *inode, int flags)
4987{
4988 handle_t *handle;
4989
4990 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
4991 if (IS_ERR(handle))
4992 goto out;
4993
4994 ext4_mark_inode_dirty(handle, inode);
4995
4996 ext4_journal_stop(handle);
4997out:
4998 return;
4999}
5000
5001#if 0
5002/*
5003 * Bind an inode's backing buffer_head into this transaction, to prevent
5004 * it from being flushed to disk early. Unlike
5005 * ext4_reserve_inode_write, this leaves behind no bh reference and
5006 * returns no iloc structure, so the caller needs to repeat the iloc
5007 * lookup to mark the inode dirty later.
5008 */
5009static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5010{
5011 struct ext4_iloc iloc;
5012
5013 int err = 0;
5014 if (handle) {
5015 err = ext4_get_inode_loc(inode, &iloc);
5016 if (!err) {
5017 BUFFER_TRACE(iloc.bh, "get_write_access");
5018 err = jbd2_journal_get_write_access(handle, iloc.bh);
5019 if (!err)
5020 err = ext4_handle_dirty_metadata(handle,
5021 NULL,
5022 iloc.bh);
5023 brelse(iloc.bh);
5024 }
5025 }
5026 ext4_std_error(inode->i_sb, err);
5027 return err;
5028}
5029#endif
5030
5031int ext4_change_inode_journal_flag(struct inode *inode, int val)
5032{
5033 journal_t *journal;
5034 handle_t *handle;
5035 int err;
5036
5037 /*
5038 * We have to be very careful here: changing a data block's
5039 * journaling status dynamically is dangerous. If we write a
5040 * data block to the journal, change the status and then delete
5041 * that block, we risk forgetting to revoke the old log record
5042 * from the journal and so a subsequent replay can corrupt data.
5043 * So, first we make sure that the journal is empty and that
5044 * nobody is changing anything.
5045 */
5046
5047 journal = EXT4_JOURNAL(inode);
5048 if (!journal)
5049 return 0;
5050 if (is_journal_aborted(journal))
5051 return -EROFS;
5052 /* We have to allocate physical blocks for delalloc blocks
5053 * before flushing journal. otherwise delalloc blocks can not
5054 * be allocated any more. even more truncate on delalloc blocks
5055 * could trigger BUG by flushing delalloc blocks in journal.
5056 * There is no delalloc block in non-journal data mode.
5057 */
5058 if (val && test_opt(inode->i_sb, DELALLOC)) {
5059 err = ext4_alloc_da_blocks(inode);
5060 if (err < 0)
5061 return err;
5062 }
5063
5064 /* Wait for all existing dio workers */
5065 ext4_inode_block_unlocked_dio(inode);
5066 inode_dio_wait(inode);
5067
5068 jbd2_journal_lock_updates(journal);
5069
5070 /*
5071 * OK, there are no updates running now, and all cached data is
5072 * synced to disk. We are now in a completely consistent state
5073 * which doesn't have anything in the journal, and we know that
5074 * no filesystem updates are running, so it is safe to modify
5075 * the inode's in-core data-journaling state flag now.
5076 */
5077
5078 if (val)
5079 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5080 else {
5081 jbd2_journal_flush(journal);
5082 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5083 }
5084 ext4_set_aops(inode);
5085
5086 jbd2_journal_unlock_updates(journal);
5087 ext4_inode_resume_unlocked_dio(inode);
5088
5089 /* Finally we can mark the inode as dirty. */
5090
5091 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
5092 if (IS_ERR(handle))
5093 return PTR_ERR(handle);
5094
5095 err = ext4_mark_inode_dirty(handle, inode);
5096 ext4_handle_sync(handle);
5097 ext4_journal_stop(handle);
5098 ext4_std_error(inode->i_sb, err);
5099
5100 return err;
5101}
5102
5103static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5104{
5105 return !buffer_mapped(bh);
5106}
5107
5108int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5109{
5110 struct page *page = vmf->page;
5111 loff_t size;
5112 unsigned long len;
5113 int ret;
5114 struct file *file = vma->vm_file;
5115 struct inode *inode = file_inode(file);
5116 struct address_space *mapping = inode->i_mapping;
5117 handle_t *handle;
5118 get_block_t *get_block;
5119 int retries = 0;
5120
5121 sb_start_pagefault(inode->i_sb);
5122 file_update_time(vma->vm_file);
5123 /* Delalloc case is easy... */
5124 if (test_opt(inode->i_sb, DELALLOC) &&
5125 !ext4_should_journal_data(inode) &&
5126 !ext4_nonda_switch(inode->i_sb)) {
5127 do {
5128 ret = __block_page_mkwrite(vma, vmf,
5129 ext4_da_get_block_prep);
5130 } while (ret == -ENOSPC &&
5131 ext4_should_retry_alloc(inode->i_sb, &retries));
5132 goto out_ret;
5133 }
5134
5135 lock_page(page);
5136 size = i_size_read(inode);
5137 /* Page got truncated from under us? */
5138 if (page->mapping != mapping || page_offset(page) > size) {
5139 unlock_page(page);
5140 ret = VM_FAULT_NOPAGE;
5141 goto out;
5142 }
5143
5144 if (page->index == size >> PAGE_CACHE_SHIFT)
5145 len = size & ~PAGE_CACHE_MASK;
5146 else
5147 len = PAGE_CACHE_SIZE;
5148 /*
5149 * Return if we have all the buffers mapped. This avoids the need to do
5150 * journal_start/journal_stop which can block and take a long time
5151 */
5152 if (page_has_buffers(page)) {
5153 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
5154 0, len, NULL,
5155 ext4_bh_unmapped)) {
5156 /* Wait so that we don't change page under IO */
5157 wait_for_stable_page(page);
5158 ret = VM_FAULT_LOCKED;
5159 goto out;
5160 }
5161 }
5162 unlock_page(page);
5163 /* OK, we need to fill the hole... */
5164 if (ext4_should_dioread_nolock(inode))
5165 get_block = ext4_get_block_write;
5166 else
5167 get_block = ext4_get_block;
5168retry_alloc:
5169 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
5170 ext4_writepage_trans_blocks(inode));
5171 if (IS_ERR(handle)) {
5172 ret = VM_FAULT_SIGBUS;
5173 goto out;
5174 }
5175 ret = __block_page_mkwrite(vma, vmf, get_block);
5176 if (!ret && ext4_should_journal_data(inode)) {
5177 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
5178 PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
5179 unlock_page(page);
5180 ret = VM_FAULT_SIGBUS;
5181 ext4_journal_stop(handle);
5182 goto out;
5183 }
5184 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
5185 }
5186 ext4_journal_stop(handle);
5187 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
5188 goto retry_alloc;
5189out_ret:
5190 ret = block_page_mkwrite_return(ret);
5191out:
5192 sb_end_pagefault(inode->i_sb);
5193 return ret;
5194}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/inode.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/inode.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * 64-bit file support on 64-bit platforms by Jakub Jelinek
17 * (jj@sunsite.ms.mff.cuni.cz)
18 *
19 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20 */
21
22#include <linux/fs.h>
23#include <linux/time.h>
24#include <linux/highuid.h>
25#include <linux/pagemap.h>
26#include <linux/dax.h>
27#include <linux/quotaops.h>
28#include <linux/string.h>
29#include <linux/buffer_head.h>
30#include <linux/writeback.h>
31#include <linux/pagevec.h>
32#include <linux/mpage.h>
33#include <linux/namei.h>
34#include <linux/uio.h>
35#include <linux/bio.h>
36#include <linux/workqueue.h>
37#include <linux/kernel.h>
38#include <linux/printk.h>
39#include <linux/slab.h>
40#include <linux/bitops.h>
41#include <linux/iomap.h>
42#include <linux/iversion.h>
43
44#include "ext4_jbd2.h"
45#include "xattr.h"
46#include "acl.h"
47#include "truncate.h"
48
49#include <trace/events/ext4.h>
50
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 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
148 */
149int ext4_inode_is_fast_symlink(struct inode *inode)
150{
151 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
152 int ea_blocks = EXT4_I(inode)->i_file_acl ?
153 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
154
155 if (ext4_has_inline_data(inode))
156 return 0;
157
158 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
159 }
160 return S_ISLNK(inode->i_mode) && inode->i_size &&
161 (inode->i_size < EXT4_N_BLOCKS * 4);
162}
163
164/*
165 * Called at the last iput() if i_nlink is zero.
166 */
167void ext4_evict_inode(struct inode *inode)
168{
169 handle_t *handle;
170 int err;
171 /*
172 * Credits for final inode cleanup and freeing:
173 * sb + inode (ext4_orphan_del()), block bitmap, group descriptor
174 * (xattr block freeing), bitmap, group descriptor (inode freeing)
175 */
176 int extra_credits = 6;
177 struct ext4_xattr_inode_array *ea_inode_array = NULL;
178
179 trace_ext4_evict_inode(inode);
180
181 if (inode->i_nlink) {
182 /*
183 * When journalling data dirty buffers are tracked only in the
184 * journal. So although mm thinks everything is clean and
185 * ready for reaping the inode might still have some pages to
186 * write in the running transaction or waiting to be
187 * checkpointed. Thus calling jbd2_journal_invalidatepage()
188 * (via truncate_inode_pages()) to discard these buffers can
189 * cause data loss. Also even if we did not discard these
190 * buffers, we would have no way to find them after the inode
191 * is reaped and thus user could see stale data if he tries to
192 * read them before the transaction is checkpointed. So be
193 * careful and force everything to disk here... We use
194 * ei->i_datasync_tid to store the newest transaction
195 * containing inode's data.
196 *
197 * Note that directories do not have this problem because they
198 * don't use page cache.
199 */
200 if (inode->i_ino != EXT4_JOURNAL_INO &&
201 ext4_should_journal_data(inode) &&
202 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
203 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. Flush worker is ignoring it because of I_FREEING
226 * flag but we still need to remove the inode from the writeback lists.
227 */
228 if (!list_empty_careful(&inode->i_io_list)) {
229 WARN_ON_ONCE(!ext4_should_journal_data(inode));
230 inode_io_list_del(inode);
231 }
232
233 /*
234 * Protect us against freezing - iput() caller didn't have to have any
235 * protection against it
236 */
237 sb_start_intwrite(inode->i_sb);
238
239 if (!IS_NOQUOTA(inode))
240 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
241
242 /*
243 * Block bitmap, group descriptor, and inode are accounted in both
244 * ext4_blocks_for_truncate() and extra_credits. So subtract 3.
245 */
246 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
247 ext4_blocks_for_truncate(inode) + extra_credits - 3);
248 if (IS_ERR(handle)) {
249 ext4_std_error(inode->i_sb, PTR_ERR(handle));
250 /*
251 * If we're going to skip the normal cleanup, we still need to
252 * make sure that the in-core orphan linked list is properly
253 * cleaned up.
254 */
255 ext4_orphan_del(NULL, inode);
256 sb_end_intwrite(inode->i_sb);
257 goto no_delete;
258 }
259
260 if (IS_SYNC(inode))
261 ext4_handle_sync(handle);
262
263 /*
264 * Set inode->i_size to 0 before calling ext4_truncate(). We need
265 * special handling of symlinks here because i_size is used to
266 * determine whether ext4_inode_info->i_data contains symlink data or
267 * block mappings. Setting i_size to 0 will remove its fast symlink
268 * status. Erase i_data so that it becomes a valid empty block map.
269 */
270 if (ext4_inode_is_fast_symlink(inode))
271 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
272 inode->i_size = 0;
273 err = ext4_mark_inode_dirty(handle, inode);
274 if (err) {
275 ext4_warning(inode->i_sb,
276 "couldn't mark inode dirty (err %d)", err);
277 goto stop_handle;
278 }
279 if (inode->i_blocks) {
280 err = ext4_truncate(inode);
281 if (err) {
282 ext4_error_err(inode->i_sb, -err,
283 "couldn't truncate inode %lu (err %d)",
284 inode->i_ino, err);
285 goto stop_handle;
286 }
287 }
288
289 /* Remove xattr references. */
290 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
291 extra_credits);
292 if (err) {
293 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
294stop_handle:
295 ext4_journal_stop(handle);
296 ext4_orphan_del(NULL, inode);
297 sb_end_intwrite(inode->i_sb);
298 ext4_xattr_inode_array_free(ea_inode_array);
299 goto no_delete;
300 }
301
302 /*
303 * Kill off the orphan record which ext4_truncate created.
304 * AKPM: I think this can be inside the above `if'.
305 * Note that ext4_orphan_del() has to be able to cope with the
306 * deletion of a non-existent orphan - this is because we don't
307 * know if ext4_truncate() actually created an orphan record.
308 * (Well, we could do this if we need to, but heck - it works)
309 */
310 ext4_orphan_del(handle, inode);
311 EXT4_I(inode)->i_dtime = (__u32)ktime_get_real_seconds();
312
313 /*
314 * One subtle ordering requirement: if anything has gone wrong
315 * (transaction abort, IO errors, whatever), then we can still
316 * do these next steps (the fs will already have been marked as
317 * having errors), but we can't free the inode if the mark_dirty
318 * fails.
319 */
320 if (ext4_mark_inode_dirty(handle, inode))
321 /* If that failed, just do the required in-core inode clear. */
322 ext4_clear_inode(inode);
323 else
324 ext4_free_inode(handle, inode);
325 ext4_journal_stop(handle);
326 sb_end_intwrite(inode->i_sb);
327 ext4_xattr_inode_array_free(ea_inode_array);
328 return;
329no_delete:
330 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
331}
332
333#ifdef CONFIG_QUOTA
334qsize_t *ext4_get_reserved_space(struct inode *inode)
335{
336 return &EXT4_I(inode)->i_reserved_quota;
337}
338#endif
339
340/*
341 * Called with i_data_sem down, which is important since we can call
342 * ext4_discard_preallocations() from here.
343 */
344void ext4_da_update_reserve_space(struct inode *inode,
345 int used, int quota_claim)
346{
347 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
348 struct ext4_inode_info *ei = EXT4_I(inode);
349
350 spin_lock(&ei->i_block_reservation_lock);
351 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
352 if (unlikely(used > ei->i_reserved_data_blocks)) {
353 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
354 "with only %d reserved data blocks",
355 __func__, inode->i_ino, used,
356 ei->i_reserved_data_blocks);
357 WARN_ON(1);
358 used = ei->i_reserved_data_blocks;
359 }
360
361 /* Update per-inode reservations */
362 ei->i_reserved_data_blocks -= used;
363 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
364
365 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
366
367 /* Update quota subsystem for data blocks */
368 if (quota_claim)
369 dquot_claim_block(inode, EXT4_C2B(sbi, used));
370 else {
371 /*
372 * We did fallocate with an offset that is already delayed
373 * allocated. So on delayed allocated writeback we should
374 * not re-claim the quota for fallocated blocks.
375 */
376 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
377 }
378
379 /*
380 * If we have done all the pending block allocations and if
381 * there aren't any writers on the inode, we can discard the
382 * inode's preallocations.
383 */
384 if ((ei->i_reserved_data_blocks == 0) &&
385 !inode_is_open_for_write(inode))
386 ext4_discard_preallocations(inode, 0);
387}
388
389static int __check_block_validity(struct inode *inode, const char *func,
390 unsigned int line,
391 struct ext4_map_blocks *map)
392{
393 if (ext4_has_feature_journal(inode->i_sb) &&
394 (inode->i_ino ==
395 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
396 return 0;
397 if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
398 ext4_error_inode(inode, func, line, map->m_pblk,
399 "lblock %lu mapped to illegal pblock %llu "
400 "(length %d)", (unsigned long) map->m_lblk,
401 map->m_pblk, map->m_len);
402 return -EFSCORRUPTED;
403 }
404 return 0;
405}
406
407int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
408 ext4_lblk_t len)
409{
410 int ret;
411
412 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
413 return fscrypt_zeroout_range(inode, lblk, pblk, len);
414
415 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
416 if (ret > 0)
417 ret = 0;
418
419 return ret;
420}
421
422#define check_block_validity(inode, map) \
423 __check_block_validity((inode), __func__, __LINE__, (map))
424
425#ifdef ES_AGGRESSIVE_TEST
426static void ext4_map_blocks_es_recheck(handle_t *handle,
427 struct inode *inode,
428 struct ext4_map_blocks *es_map,
429 struct ext4_map_blocks *map,
430 int flags)
431{
432 int retval;
433
434 map->m_flags = 0;
435 /*
436 * There is a race window that the result is not the same.
437 * e.g. xfstests #223 when dioread_nolock enables. The reason
438 * is that we lookup a block mapping in extent status tree with
439 * out taking i_data_sem. So at the time the unwritten extent
440 * could be converted.
441 */
442 down_read(&EXT4_I(inode)->i_data_sem);
443 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
444 retval = ext4_ext_map_blocks(handle, inode, map, 0);
445 } else {
446 retval = ext4_ind_map_blocks(handle, inode, map, 0);
447 }
448 up_read((&EXT4_I(inode)->i_data_sem));
449
450 /*
451 * We don't check m_len because extent will be collpased in status
452 * tree. So the m_len might not equal.
453 */
454 if (es_map->m_lblk != map->m_lblk ||
455 es_map->m_flags != map->m_flags ||
456 es_map->m_pblk != map->m_pblk) {
457 printk("ES cache assertion failed for inode: %lu "
458 "es_cached ex [%d/%d/%llu/%x] != "
459 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
460 inode->i_ino, es_map->m_lblk, es_map->m_len,
461 es_map->m_pblk, es_map->m_flags, map->m_lblk,
462 map->m_len, map->m_pblk, map->m_flags,
463 retval, flags);
464 }
465}
466#endif /* ES_AGGRESSIVE_TEST */
467
468/*
469 * The ext4_map_blocks() function tries to look up the requested blocks,
470 * and returns if the blocks are already mapped.
471 *
472 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
473 * and store the allocated blocks in the result buffer head and mark it
474 * mapped.
475 *
476 * If file type is extents based, it will call ext4_ext_map_blocks(),
477 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
478 * based files
479 *
480 * On success, it returns the number of blocks being mapped or allocated. if
481 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
482 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
483 *
484 * It returns 0 if plain look up failed (blocks have not been allocated), in
485 * that case, @map is returned as unmapped but we still do fill map->m_len to
486 * indicate the length of a hole starting at map->m_lblk.
487 *
488 * It returns the error in case of allocation failure.
489 */
490int ext4_map_blocks(handle_t *handle, struct inode *inode,
491 struct ext4_map_blocks *map, int flags)
492{
493 struct extent_status es;
494 int retval;
495 int ret = 0;
496#ifdef ES_AGGRESSIVE_TEST
497 struct ext4_map_blocks orig_map;
498
499 memcpy(&orig_map, map, sizeof(*map));
500#endif
501
502 map->m_flags = 0;
503 ext_debug(inode, "flag 0x%x, max_blocks %u, logical block %lu\n",
504 flags, map->m_len, (unsigned long) map->m_lblk);
505
506 /*
507 * ext4_map_blocks returns an int, and m_len is an unsigned int
508 */
509 if (unlikely(map->m_len > INT_MAX))
510 map->m_len = INT_MAX;
511
512 /* We can handle the block number less than EXT_MAX_BLOCKS */
513 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
514 return -EFSCORRUPTED;
515
516 /* Lookup extent status tree firstly */
517 if (ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
518 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
519 map->m_pblk = ext4_es_pblock(&es) +
520 map->m_lblk - es.es_lblk;
521 map->m_flags |= ext4_es_is_written(&es) ?
522 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
523 retval = es.es_len - (map->m_lblk - es.es_lblk);
524 if (retval > map->m_len)
525 retval = map->m_len;
526 map->m_len = retval;
527 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
528 map->m_pblk = 0;
529 retval = es.es_len - (map->m_lblk - es.es_lblk);
530 if (retval > map->m_len)
531 retval = map->m_len;
532 map->m_len = retval;
533 retval = 0;
534 } else {
535 BUG();
536 }
537#ifdef ES_AGGRESSIVE_TEST
538 ext4_map_blocks_es_recheck(handle, inode, map,
539 &orig_map, flags);
540#endif
541 goto found;
542 }
543
544 /*
545 * Try to see if we can get the block without requesting a new
546 * file system block.
547 */
548 down_read(&EXT4_I(inode)->i_data_sem);
549 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
550 retval = ext4_ext_map_blocks(handle, inode, map, 0);
551 } else {
552 retval = ext4_ind_map_blocks(handle, inode, map, 0);
553 }
554 if (retval > 0) {
555 unsigned int status;
556
557 if (unlikely(retval != map->m_len)) {
558 ext4_warning(inode->i_sb,
559 "ES len assertion failed for inode "
560 "%lu: retval %d != map->m_len %d",
561 inode->i_ino, retval, map->m_len);
562 WARN_ON(1);
563 }
564
565 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
566 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
567 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
568 !(status & EXTENT_STATUS_WRITTEN) &&
569 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
570 map->m_lblk + map->m_len - 1))
571 status |= EXTENT_STATUS_DELAYED;
572 ret = ext4_es_insert_extent(inode, map->m_lblk,
573 map->m_len, map->m_pblk, status);
574 if (ret < 0)
575 retval = ret;
576 }
577 up_read((&EXT4_I(inode)->i_data_sem));
578
579found:
580 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
581 ret = check_block_validity(inode, map);
582 if (ret != 0)
583 return ret;
584 }
585
586 /* If it is only a block(s) look up */
587 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
588 return retval;
589
590 /*
591 * Returns if the blocks have already allocated
592 *
593 * Note that if blocks have been preallocated
594 * ext4_ext_get_block() returns the create = 0
595 * with buffer head unmapped.
596 */
597 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
598 /*
599 * If we need to convert extent to unwritten
600 * we continue and do the actual work in
601 * ext4_ext_map_blocks()
602 */
603 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
604 return retval;
605
606 /*
607 * Here we clear m_flags because after allocating an new extent,
608 * it will be set again.
609 */
610 map->m_flags &= ~EXT4_MAP_FLAGS;
611
612 /*
613 * New blocks allocate and/or writing to unwritten extent
614 * will possibly result in updating i_data, so we take
615 * the write lock of i_data_sem, and call get_block()
616 * with create == 1 flag.
617 */
618 down_write(&EXT4_I(inode)->i_data_sem);
619
620 /*
621 * We need to check for EXT4 here because migrate
622 * could have changed the inode type in between
623 */
624 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
625 retval = ext4_ext_map_blocks(handle, inode, map, flags);
626 } else {
627 retval = ext4_ind_map_blocks(handle, inode, map, flags);
628
629 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
630 /*
631 * We allocated new blocks which will result in
632 * i_data's format changing. Force the migrate
633 * to fail by clearing migrate flags
634 */
635 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
636 }
637
638 /*
639 * Update reserved blocks/metadata blocks after successful
640 * block allocation which had been deferred till now. We don't
641 * support fallocate for non extent files. So we can update
642 * reserve space here.
643 */
644 if ((retval > 0) &&
645 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
646 ext4_da_update_reserve_space(inode, retval, 1);
647 }
648
649 if (retval > 0) {
650 unsigned int status;
651
652 if (unlikely(retval != map->m_len)) {
653 ext4_warning(inode->i_sb,
654 "ES len assertion failed for inode "
655 "%lu: retval %d != map->m_len %d",
656 inode->i_ino, retval, map->m_len);
657 WARN_ON(1);
658 }
659
660 /*
661 * We have to zeroout blocks before inserting them into extent
662 * status tree. Otherwise someone could look them up there and
663 * use them before they are really zeroed. We also have to
664 * unmap metadata before zeroing as otherwise writeback can
665 * overwrite zeros with stale data from block device.
666 */
667 if (flags & EXT4_GET_BLOCKS_ZERO &&
668 map->m_flags & EXT4_MAP_MAPPED &&
669 map->m_flags & EXT4_MAP_NEW) {
670 ret = ext4_issue_zeroout(inode, map->m_lblk,
671 map->m_pblk, map->m_len);
672 if (ret) {
673 retval = ret;
674 goto out_sem;
675 }
676 }
677
678 /*
679 * If the extent has been zeroed out, we don't need to update
680 * extent status tree.
681 */
682 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
683 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
684 if (ext4_es_is_written(&es))
685 goto out_sem;
686 }
687 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
688 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
689 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
690 !(status & EXTENT_STATUS_WRITTEN) &&
691 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
692 map->m_lblk + map->m_len - 1))
693 status |= EXTENT_STATUS_DELAYED;
694 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
695 map->m_pblk, status);
696 if (ret < 0) {
697 retval = ret;
698 goto out_sem;
699 }
700 }
701
702out_sem:
703 up_write((&EXT4_I(inode)->i_data_sem));
704 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
705 ret = check_block_validity(inode, map);
706 if (ret != 0)
707 return ret;
708
709 /*
710 * Inodes with freshly allocated blocks where contents will be
711 * visible after transaction commit must be on transaction's
712 * ordered data list.
713 */
714 if (map->m_flags & EXT4_MAP_NEW &&
715 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
716 !(flags & EXT4_GET_BLOCKS_ZERO) &&
717 !ext4_is_quota_file(inode) &&
718 ext4_should_order_data(inode)) {
719 loff_t start_byte =
720 (loff_t)map->m_lblk << inode->i_blkbits;
721 loff_t length = (loff_t)map->m_len << inode->i_blkbits;
722
723 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
724 ret = ext4_jbd2_inode_add_wait(handle, inode,
725 start_byte, length);
726 else
727 ret = ext4_jbd2_inode_add_write(handle, inode,
728 start_byte, length);
729 if (ret)
730 return ret;
731 }
732 }
733
734 if (retval < 0)
735 ext_debug(inode, "failed with err %d\n", retval);
736 return retval;
737}
738
739/*
740 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
741 * we have to be careful as someone else may be manipulating b_state as well.
742 */
743static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
744{
745 unsigned long old_state;
746 unsigned long new_state;
747
748 flags &= EXT4_MAP_FLAGS;
749
750 /* Dummy buffer_head? Set non-atomically. */
751 if (!bh->b_page) {
752 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
753 return;
754 }
755 /*
756 * Someone else may be modifying b_state. Be careful! This is ugly but
757 * once we get rid of using bh as a container for mapping information
758 * to pass to / from get_block functions, this can go away.
759 */
760 do {
761 old_state = READ_ONCE(bh->b_state);
762 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
763 } while (unlikely(
764 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
765}
766
767static int _ext4_get_block(struct inode *inode, sector_t iblock,
768 struct buffer_head *bh, int flags)
769{
770 struct ext4_map_blocks map;
771 int ret = 0;
772
773 if (ext4_has_inline_data(inode))
774 return -ERANGE;
775
776 map.m_lblk = iblock;
777 map.m_len = bh->b_size >> inode->i_blkbits;
778
779 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
780 flags);
781 if (ret > 0) {
782 map_bh(bh, inode->i_sb, map.m_pblk);
783 ext4_update_bh_state(bh, map.m_flags);
784 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
785 ret = 0;
786 } else if (ret == 0) {
787 /* hole case, need to fill in bh->b_size */
788 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
789 }
790 return ret;
791}
792
793int ext4_get_block(struct inode *inode, sector_t iblock,
794 struct buffer_head *bh, int create)
795{
796 return _ext4_get_block(inode, iblock, bh,
797 create ? EXT4_GET_BLOCKS_CREATE : 0);
798}
799
800/*
801 * Get block function used when preparing for buffered write if we require
802 * creating an unwritten extent if blocks haven't been allocated. The extent
803 * will be converted to written after the IO is complete.
804 */
805int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
806 struct buffer_head *bh_result, int create)
807{
808 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
809 inode->i_ino, create);
810 return _ext4_get_block(inode, iblock, bh_result,
811 EXT4_GET_BLOCKS_IO_CREATE_EXT);
812}
813
814/* Maximum number of blocks we map for direct IO at once. */
815#define DIO_MAX_BLOCKS 4096
816
817/*
818 * `handle' can be NULL if create is zero
819 */
820struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
821 ext4_lblk_t block, int map_flags)
822{
823 struct ext4_map_blocks map;
824 struct buffer_head *bh;
825 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
826 int err;
827
828 J_ASSERT(handle != NULL || create == 0);
829
830 map.m_lblk = block;
831 map.m_len = 1;
832 err = ext4_map_blocks(handle, inode, &map, map_flags);
833
834 if (err == 0)
835 return create ? ERR_PTR(-ENOSPC) : NULL;
836 if (err < 0)
837 return ERR_PTR(err);
838
839 bh = sb_getblk(inode->i_sb, map.m_pblk);
840 if (unlikely(!bh))
841 return ERR_PTR(-ENOMEM);
842 if (map.m_flags & EXT4_MAP_NEW) {
843 J_ASSERT(create != 0);
844 J_ASSERT(handle != NULL);
845
846 /*
847 * Now that we do not always journal data, we should
848 * keep in mind whether this should always journal the
849 * new buffer as metadata. For now, regular file
850 * writes use ext4_get_block instead, so it's not a
851 * problem.
852 */
853 lock_buffer(bh);
854 BUFFER_TRACE(bh, "call get_create_access");
855 err = ext4_journal_get_create_access(handle, bh);
856 if (unlikely(err)) {
857 unlock_buffer(bh);
858 goto errout;
859 }
860 if (!buffer_uptodate(bh)) {
861 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
862 set_buffer_uptodate(bh);
863 }
864 unlock_buffer(bh);
865 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
866 err = ext4_handle_dirty_metadata(handle, inode, bh);
867 if (unlikely(err))
868 goto errout;
869 } else
870 BUFFER_TRACE(bh, "not a new buffer");
871 return bh;
872errout:
873 brelse(bh);
874 return ERR_PTR(err);
875}
876
877struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
878 ext4_lblk_t block, int map_flags)
879{
880 struct buffer_head *bh;
881
882 bh = ext4_getblk(handle, inode, block, map_flags);
883 if (IS_ERR(bh))
884 return bh;
885 if (!bh || ext4_buffer_uptodate(bh))
886 return bh;
887 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
888 wait_on_buffer(bh);
889 if (buffer_uptodate(bh))
890 return bh;
891 put_bh(bh);
892 return ERR_PTR(-EIO);
893}
894
895/* Read a contiguous batch of blocks. */
896int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
897 bool wait, struct buffer_head **bhs)
898{
899 int i, err;
900
901 for (i = 0; i < bh_count; i++) {
902 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
903 if (IS_ERR(bhs[i])) {
904 err = PTR_ERR(bhs[i]);
905 bh_count = i;
906 goto out_brelse;
907 }
908 }
909
910 for (i = 0; i < bh_count; i++)
911 /* Note that NULL bhs[i] is valid because of holes. */
912 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
913 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
914 &bhs[i]);
915
916 if (!wait)
917 return 0;
918
919 for (i = 0; i < bh_count; i++)
920 if (bhs[i])
921 wait_on_buffer(bhs[i]);
922
923 for (i = 0; i < bh_count; i++) {
924 if (bhs[i] && !buffer_uptodate(bhs[i])) {
925 err = -EIO;
926 goto out_brelse;
927 }
928 }
929 return 0;
930
931out_brelse:
932 for (i = 0; i < bh_count; i++) {
933 brelse(bhs[i]);
934 bhs[i] = NULL;
935 }
936 return err;
937}
938
939int ext4_walk_page_buffers(handle_t *handle,
940 struct buffer_head *head,
941 unsigned from,
942 unsigned to,
943 int *partial,
944 int (*fn)(handle_t *handle,
945 struct buffer_head *bh))
946{
947 struct buffer_head *bh;
948 unsigned block_start, block_end;
949 unsigned blocksize = head->b_size;
950 int err, ret = 0;
951 struct buffer_head *next;
952
953 for (bh = head, block_start = 0;
954 ret == 0 && (bh != head || !block_start);
955 block_start = block_end, bh = next) {
956 next = bh->b_this_page;
957 block_end = block_start + blocksize;
958 if (block_end <= from || block_start >= to) {
959 if (partial && !buffer_uptodate(bh))
960 *partial = 1;
961 continue;
962 }
963 err = (*fn)(handle, bh);
964 if (!ret)
965 ret = err;
966 }
967 return ret;
968}
969
970/*
971 * To preserve ordering, it is essential that the hole instantiation and
972 * the data write be encapsulated in a single transaction. We cannot
973 * close off a transaction and start a new one between the ext4_get_block()
974 * and the commit_write(). So doing the jbd2_journal_start at the start of
975 * prepare_write() is the right place.
976 *
977 * Also, this function can nest inside ext4_writepage(). In that case, we
978 * *know* that ext4_writepage() has generated enough buffer credits to do the
979 * whole page. So we won't block on the journal in that case, which is good,
980 * because the caller may be PF_MEMALLOC.
981 *
982 * By accident, ext4 can be reentered when a transaction is open via
983 * quota file writes. If we were to commit the transaction while thus
984 * reentered, there can be a deadlock - we would be holding a quota
985 * lock, and the commit would never complete if another thread had a
986 * transaction open and was blocking on the quota lock - a ranking
987 * violation.
988 *
989 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
990 * will _not_ run commit under these circumstances because handle->h_ref
991 * is elevated. We'll still have enough credits for the tiny quotafile
992 * write.
993 */
994int do_journal_get_write_access(handle_t *handle,
995 struct buffer_head *bh)
996{
997 int dirty = buffer_dirty(bh);
998 int ret;
999
1000 if (!buffer_mapped(bh) || buffer_freed(bh))
1001 return 0;
1002 /*
1003 * __block_write_begin() could have dirtied some buffers. Clean
1004 * the dirty bit as jbd2_journal_get_write_access() could complain
1005 * otherwise about fs integrity issues. Setting of the dirty bit
1006 * by __block_write_begin() isn't a real problem here as we clear
1007 * the bit before releasing a page lock and thus writeback cannot
1008 * ever write the buffer.
1009 */
1010 if (dirty)
1011 clear_buffer_dirty(bh);
1012 BUFFER_TRACE(bh, "get write access");
1013 ret = ext4_journal_get_write_access(handle, bh);
1014 if (!ret && dirty)
1015 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1016 return ret;
1017}
1018
1019#ifdef CONFIG_FS_ENCRYPTION
1020static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1021 get_block_t *get_block)
1022{
1023 unsigned from = pos & (PAGE_SIZE - 1);
1024 unsigned to = from + len;
1025 struct inode *inode = page->mapping->host;
1026 unsigned block_start, block_end;
1027 sector_t block;
1028 int err = 0;
1029 unsigned blocksize = inode->i_sb->s_blocksize;
1030 unsigned bbits;
1031 struct buffer_head *bh, *head, *wait[2];
1032 int nr_wait = 0;
1033 int i;
1034
1035 BUG_ON(!PageLocked(page));
1036 BUG_ON(from > PAGE_SIZE);
1037 BUG_ON(to > PAGE_SIZE);
1038 BUG_ON(from > to);
1039
1040 if (!page_has_buffers(page))
1041 create_empty_buffers(page, blocksize, 0);
1042 head = page_buffers(page);
1043 bbits = ilog2(blocksize);
1044 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1045
1046 for (bh = head, block_start = 0; bh != head || !block_start;
1047 block++, block_start = block_end, bh = bh->b_this_page) {
1048 block_end = block_start + blocksize;
1049 if (block_end <= from || block_start >= to) {
1050 if (PageUptodate(page)) {
1051 if (!buffer_uptodate(bh))
1052 set_buffer_uptodate(bh);
1053 }
1054 continue;
1055 }
1056 if (buffer_new(bh))
1057 clear_buffer_new(bh);
1058 if (!buffer_mapped(bh)) {
1059 WARN_ON(bh->b_size != blocksize);
1060 err = get_block(inode, block, bh, 1);
1061 if (err)
1062 break;
1063 if (buffer_new(bh)) {
1064 if (PageUptodate(page)) {
1065 clear_buffer_new(bh);
1066 set_buffer_uptodate(bh);
1067 mark_buffer_dirty(bh);
1068 continue;
1069 }
1070 if (block_end > to || block_start < from)
1071 zero_user_segments(page, to, block_end,
1072 block_start, from);
1073 continue;
1074 }
1075 }
1076 if (PageUptodate(page)) {
1077 if (!buffer_uptodate(bh))
1078 set_buffer_uptodate(bh);
1079 continue;
1080 }
1081 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1082 !buffer_unwritten(bh) &&
1083 (block_start < from || block_end > to)) {
1084 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1085 wait[nr_wait++] = bh;
1086 }
1087 }
1088 /*
1089 * If we issued read requests, let them complete.
1090 */
1091 for (i = 0; i < nr_wait; i++) {
1092 wait_on_buffer(wait[i]);
1093 if (!buffer_uptodate(wait[i]))
1094 err = -EIO;
1095 }
1096 if (unlikely(err)) {
1097 page_zero_new_buffers(page, from, to);
1098 } else if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
1099 for (i = 0; i < nr_wait; i++) {
1100 int err2;
1101
1102 err2 = fscrypt_decrypt_pagecache_blocks(page, blocksize,
1103 bh_offset(wait[i]));
1104 if (err2) {
1105 clear_buffer_uptodate(wait[i]);
1106 err = err2;
1107 }
1108 }
1109 }
1110
1111 return err;
1112}
1113#endif
1114
1115static int ext4_write_begin(struct file *file, struct address_space *mapping,
1116 loff_t pos, unsigned len, unsigned flags,
1117 struct page **pagep, void **fsdata)
1118{
1119 struct inode *inode = mapping->host;
1120 int ret, needed_blocks;
1121 handle_t *handle;
1122 int retries = 0;
1123 struct page *page;
1124 pgoff_t index;
1125 unsigned from, to;
1126
1127 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1128 return -EIO;
1129
1130 trace_ext4_write_begin(inode, pos, len, flags);
1131 /*
1132 * Reserve one block more for addition to orphan list in case
1133 * we allocate blocks but write fails for some reason
1134 */
1135 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1136 index = pos >> PAGE_SHIFT;
1137 from = pos & (PAGE_SIZE - 1);
1138 to = from + len;
1139
1140 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1141 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1142 flags, pagep);
1143 if (ret < 0)
1144 return ret;
1145 if (ret == 1)
1146 return 0;
1147 }
1148
1149 /*
1150 * grab_cache_page_write_begin() can take a long time if the
1151 * system is thrashing due to memory pressure, or if the page
1152 * is being written back. So grab it first before we start
1153 * the transaction handle. This also allows us to allocate
1154 * the page (if needed) without using GFP_NOFS.
1155 */
1156retry_grab:
1157 page = grab_cache_page_write_begin(mapping, index, flags);
1158 if (!page)
1159 return -ENOMEM;
1160 unlock_page(page);
1161
1162retry_journal:
1163 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1164 if (IS_ERR(handle)) {
1165 put_page(page);
1166 return PTR_ERR(handle);
1167 }
1168
1169 lock_page(page);
1170 if (page->mapping != mapping) {
1171 /* The page got truncated from under us */
1172 unlock_page(page);
1173 put_page(page);
1174 ext4_journal_stop(handle);
1175 goto retry_grab;
1176 }
1177 /* In case writeback began while the page was unlocked */
1178 wait_for_stable_page(page);
1179
1180#ifdef CONFIG_FS_ENCRYPTION
1181 if (ext4_should_dioread_nolock(inode))
1182 ret = ext4_block_write_begin(page, pos, len,
1183 ext4_get_block_unwritten);
1184 else
1185 ret = ext4_block_write_begin(page, pos, len,
1186 ext4_get_block);
1187#else
1188 if (ext4_should_dioread_nolock(inode))
1189 ret = __block_write_begin(page, pos, len,
1190 ext4_get_block_unwritten);
1191 else
1192 ret = __block_write_begin(page, pos, len, ext4_get_block);
1193#endif
1194 if (!ret && ext4_should_journal_data(inode)) {
1195 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1196 from, to, NULL,
1197 do_journal_get_write_access);
1198 }
1199
1200 if (ret) {
1201 bool extended = (pos + len > inode->i_size) &&
1202 !ext4_verity_in_progress(inode);
1203
1204 unlock_page(page);
1205 /*
1206 * __block_write_begin may have instantiated a few blocks
1207 * outside i_size. Trim these off again. Don't need
1208 * i_size_read because we hold i_mutex.
1209 *
1210 * Add inode to orphan list in case we crash before
1211 * truncate finishes
1212 */
1213 if (extended && ext4_can_truncate(inode))
1214 ext4_orphan_add(handle, inode);
1215
1216 ext4_journal_stop(handle);
1217 if (extended) {
1218 ext4_truncate_failed_write(inode);
1219 /*
1220 * If truncate failed early the inode might
1221 * still be on the orphan list; we need to
1222 * make sure the inode is removed from the
1223 * orphan list in that case.
1224 */
1225 if (inode->i_nlink)
1226 ext4_orphan_del(NULL, inode);
1227 }
1228
1229 if (ret == -ENOSPC &&
1230 ext4_should_retry_alloc(inode->i_sb, &retries))
1231 goto retry_journal;
1232 put_page(page);
1233 return ret;
1234 }
1235 *pagep = page;
1236 return ret;
1237}
1238
1239/* For write_end() in data=journal mode */
1240static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1241{
1242 int ret;
1243 if (!buffer_mapped(bh) || buffer_freed(bh))
1244 return 0;
1245 set_buffer_uptodate(bh);
1246 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1247 clear_buffer_meta(bh);
1248 clear_buffer_prio(bh);
1249 return ret;
1250}
1251
1252/*
1253 * We need to pick up the new inode size which generic_commit_write gave us
1254 * `file' can be NULL - eg, when called from page_symlink().
1255 *
1256 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1257 * buffers are managed internally.
1258 */
1259static int ext4_write_end(struct file *file,
1260 struct address_space *mapping,
1261 loff_t pos, unsigned len, unsigned copied,
1262 struct page *page, void *fsdata)
1263{
1264 handle_t *handle = ext4_journal_current_handle();
1265 struct inode *inode = mapping->host;
1266 loff_t old_size = inode->i_size;
1267 int ret = 0, ret2;
1268 int i_size_changed = 0;
1269 int inline_data = ext4_has_inline_data(inode);
1270 bool verity = ext4_verity_in_progress(inode);
1271
1272 trace_ext4_write_end(inode, pos, len, copied);
1273 if (inline_data) {
1274 ret = ext4_write_inline_data_end(inode, pos, len,
1275 copied, page);
1276 if (ret < 0) {
1277 unlock_page(page);
1278 put_page(page);
1279 goto errout;
1280 }
1281 copied = ret;
1282 } else
1283 copied = block_write_end(file, mapping, pos,
1284 len, copied, page, fsdata);
1285 /*
1286 * it's important to update i_size while still holding page lock:
1287 * page writeout could otherwise come in and zero beyond i_size.
1288 *
1289 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1290 * blocks are being written past EOF, so skip the i_size update.
1291 */
1292 if (!verity)
1293 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1294 unlock_page(page);
1295 put_page(page);
1296
1297 if (old_size < pos && !verity)
1298 pagecache_isize_extended(inode, old_size, pos);
1299 /*
1300 * Don't mark the inode dirty under page lock. First, it unnecessarily
1301 * makes the holding time of page lock longer. Second, it forces lock
1302 * ordering of page lock and transaction start for journaling
1303 * filesystems.
1304 */
1305 if (i_size_changed || inline_data)
1306 ret = ext4_mark_inode_dirty(handle, inode);
1307
1308 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1309 /* if we have allocated more blocks and copied
1310 * less. We will have blocks allocated outside
1311 * inode->i_size. So truncate them
1312 */
1313 ext4_orphan_add(handle, inode);
1314errout:
1315 ret2 = ext4_journal_stop(handle);
1316 if (!ret)
1317 ret = ret2;
1318
1319 if (pos + len > inode->i_size && !verity) {
1320 ext4_truncate_failed_write(inode);
1321 /*
1322 * If truncate failed early the inode might still be
1323 * on the orphan list; we need to make sure the inode
1324 * is removed from the orphan list in that case.
1325 */
1326 if (inode->i_nlink)
1327 ext4_orphan_del(NULL, inode);
1328 }
1329
1330 return ret ? ret : copied;
1331}
1332
1333/*
1334 * This is a private version of page_zero_new_buffers() which doesn't
1335 * set the buffer to be dirty, since in data=journalled mode we need
1336 * to call ext4_handle_dirty_metadata() instead.
1337 */
1338static void ext4_journalled_zero_new_buffers(handle_t *handle,
1339 struct page *page,
1340 unsigned from, unsigned to)
1341{
1342 unsigned int block_start = 0, block_end;
1343 struct buffer_head *head, *bh;
1344
1345 bh = head = page_buffers(page);
1346 do {
1347 block_end = block_start + bh->b_size;
1348 if (buffer_new(bh)) {
1349 if (block_end > from && block_start < to) {
1350 if (!PageUptodate(page)) {
1351 unsigned start, size;
1352
1353 start = max(from, block_start);
1354 size = min(to, block_end) - start;
1355
1356 zero_user(page, start, size);
1357 write_end_fn(handle, bh);
1358 }
1359 clear_buffer_new(bh);
1360 }
1361 }
1362 block_start = block_end;
1363 bh = bh->b_this_page;
1364 } while (bh != head);
1365}
1366
1367static int ext4_journalled_write_end(struct file *file,
1368 struct address_space *mapping,
1369 loff_t pos, unsigned len, unsigned copied,
1370 struct page *page, void *fsdata)
1371{
1372 handle_t *handle = ext4_journal_current_handle();
1373 struct inode *inode = mapping->host;
1374 loff_t old_size = inode->i_size;
1375 int ret = 0, ret2;
1376 int partial = 0;
1377 unsigned from, to;
1378 int size_changed = 0;
1379 int inline_data = ext4_has_inline_data(inode);
1380 bool verity = ext4_verity_in_progress(inode);
1381
1382 trace_ext4_journalled_write_end(inode, pos, len, copied);
1383 from = pos & (PAGE_SIZE - 1);
1384 to = from + len;
1385
1386 BUG_ON(!ext4_handle_valid(handle));
1387
1388 if (inline_data) {
1389 ret = ext4_write_inline_data_end(inode, pos, len,
1390 copied, page);
1391 if (ret < 0) {
1392 unlock_page(page);
1393 put_page(page);
1394 goto errout;
1395 }
1396 copied = ret;
1397 } else if (unlikely(copied < len) && !PageUptodate(page)) {
1398 copied = 0;
1399 ext4_journalled_zero_new_buffers(handle, page, from, to);
1400 } else {
1401 if (unlikely(copied < len))
1402 ext4_journalled_zero_new_buffers(handle, page,
1403 from + copied, to);
1404 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1405 from + copied, &partial,
1406 write_end_fn);
1407 if (!partial)
1408 SetPageUptodate(page);
1409 }
1410 if (!verity)
1411 size_changed = ext4_update_inode_size(inode, pos + copied);
1412 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1413 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1414 unlock_page(page);
1415 put_page(page);
1416
1417 if (old_size < pos && !verity)
1418 pagecache_isize_extended(inode, old_size, pos);
1419
1420 if (size_changed || inline_data) {
1421 ret2 = ext4_mark_inode_dirty(handle, inode);
1422 if (!ret)
1423 ret = ret2;
1424 }
1425
1426 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1427 /* if we have allocated more blocks and copied
1428 * less. We will have blocks allocated outside
1429 * inode->i_size. So truncate them
1430 */
1431 ext4_orphan_add(handle, inode);
1432
1433errout:
1434 ret2 = ext4_journal_stop(handle);
1435 if (!ret)
1436 ret = ret2;
1437 if (pos + len > inode->i_size && !verity) {
1438 ext4_truncate_failed_write(inode);
1439 /*
1440 * If truncate failed early the inode might still be
1441 * on the orphan list; we need to make sure the inode
1442 * is removed from the orphan list in that case.
1443 */
1444 if (inode->i_nlink)
1445 ext4_orphan_del(NULL, inode);
1446 }
1447
1448 return ret ? ret : copied;
1449}
1450
1451/*
1452 * Reserve space for a single cluster
1453 */
1454static int ext4_da_reserve_space(struct inode *inode)
1455{
1456 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1457 struct ext4_inode_info *ei = EXT4_I(inode);
1458 int ret;
1459
1460 /*
1461 * We will charge metadata quota at writeout time; this saves
1462 * us from metadata over-estimation, though we may go over by
1463 * a small amount in the end. Here we just reserve for data.
1464 */
1465 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1466 if (ret)
1467 return ret;
1468
1469 spin_lock(&ei->i_block_reservation_lock);
1470 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1471 spin_unlock(&ei->i_block_reservation_lock);
1472 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1473 return -ENOSPC;
1474 }
1475 ei->i_reserved_data_blocks++;
1476 trace_ext4_da_reserve_space(inode);
1477 spin_unlock(&ei->i_block_reservation_lock);
1478
1479 return 0; /* success */
1480}
1481
1482void ext4_da_release_space(struct inode *inode, int to_free)
1483{
1484 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1485 struct ext4_inode_info *ei = EXT4_I(inode);
1486
1487 if (!to_free)
1488 return; /* Nothing to release, exit */
1489
1490 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1491
1492 trace_ext4_da_release_space(inode, to_free);
1493 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1494 /*
1495 * if there aren't enough reserved blocks, then the
1496 * counter is messed up somewhere. Since this
1497 * function is called from invalidate page, it's
1498 * harmless to return without any action.
1499 */
1500 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1501 "ino %lu, to_free %d with only %d reserved "
1502 "data blocks", inode->i_ino, to_free,
1503 ei->i_reserved_data_blocks);
1504 WARN_ON(1);
1505 to_free = ei->i_reserved_data_blocks;
1506 }
1507 ei->i_reserved_data_blocks -= to_free;
1508
1509 /* update fs dirty data blocks counter */
1510 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1511
1512 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1513
1514 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1515}
1516
1517/*
1518 * Delayed allocation stuff
1519 */
1520
1521struct mpage_da_data {
1522 struct inode *inode;
1523 struct writeback_control *wbc;
1524
1525 pgoff_t first_page; /* The first page to write */
1526 pgoff_t next_page; /* Current page to examine */
1527 pgoff_t last_page; /* Last page to examine */
1528 /*
1529 * Extent to map - this can be after first_page because that can be
1530 * fully mapped. We somewhat abuse m_flags to store whether the extent
1531 * is delalloc or unwritten.
1532 */
1533 struct ext4_map_blocks map;
1534 struct ext4_io_submit io_submit; /* IO submission data */
1535 unsigned int do_map:1;
1536 unsigned int scanned_until_end:1;
1537};
1538
1539static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1540 bool invalidate)
1541{
1542 int nr_pages, i;
1543 pgoff_t index, end;
1544 struct pagevec pvec;
1545 struct inode *inode = mpd->inode;
1546 struct address_space *mapping = inode->i_mapping;
1547
1548 /* This is necessary when next_page == 0. */
1549 if (mpd->first_page >= mpd->next_page)
1550 return;
1551
1552 mpd->scanned_until_end = 0;
1553 index = mpd->first_page;
1554 end = mpd->next_page - 1;
1555 if (invalidate) {
1556 ext4_lblk_t start, last;
1557 start = index << (PAGE_SHIFT - inode->i_blkbits);
1558 last = end << (PAGE_SHIFT - inode->i_blkbits);
1559 ext4_es_remove_extent(inode, start, last - start + 1);
1560 }
1561
1562 pagevec_init(&pvec);
1563 while (index <= end) {
1564 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1565 if (nr_pages == 0)
1566 break;
1567 for (i = 0; i < nr_pages; i++) {
1568 struct page *page = pvec.pages[i];
1569
1570 BUG_ON(!PageLocked(page));
1571 BUG_ON(PageWriteback(page));
1572 if (invalidate) {
1573 if (page_mapped(page))
1574 clear_page_dirty_for_io(page);
1575 block_invalidatepage(page, 0, PAGE_SIZE);
1576 ClearPageUptodate(page);
1577 }
1578 unlock_page(page);
1579 }
1580 pagevec_release(&pvec);
1581 }
1582}
1583
1584static void ext4_print_free_blocks(struct inode *inode)
1585{
1586 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1587 struct super_block *sb = inode->i_sb;
1588 struct ext4_inode_info *ei = EXT4_I(inode);
1589
1590 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1591 EXT4_C2B(EXT4_SB(inode->i_sb),
1592 ext4_count_free_clusters(sb)));
1593 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1594 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1595 (long long) EXT4_C2B(EXT4_SB(sb),
1596 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1597 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1598 (long long) EXT4_C2B(EXT4_SB(sb),
1599 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1600 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1601 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1602 ei->i_reserved_data_blocks);
1603 return;
1604}
1605
1606static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1607{
1608 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1609}
1610
1611/*
1612 * ext4_insert_delayed_block - adds a delayed block to the extents status
1613 * tree, incrementing the reserved cluster/block
1614 * count or making a pending reservation
1615 * where needed
1616 *
1617 * @inode - file containing the newly added block
1618 * @lblk - logical block to be added
1619 *
1620 * Returns 0 on success, negative error code on failure.
1621 */
1622static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1623{
1624 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1625 int ret;
1626 bool allocated = false;
1627
1628 /*
1629 * If the cluster containing lblk is shared with a delayed,
1630 * written, or unwritten extent in a bigalloc file system, it's
1631 * already been accounted for and does not need to be reserved.
1632 * A pending reservation must be made for the cluster if it's
1633 * shared with a written or unwritten extent and doesn't already
1634 * have one. Written and unwritten extents can be purged from the
1635 * extents status tree if the system is under memory pressure, so
1636 * it's necessary to examine the extent tree if a search of the
1637 * extents status tree doesn't get a match.
1638 */
1639 if (sbi->s_cluster_ratio == 1) {
1640 ret = ext4_da_reserve_space(inode);
1641 if (ret != 0) /* ENOSPC */
1642 goto errout;
1643 } else { /* bigalloc */
1644 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1645 if (!ext4_es_scan_clu(inode,
1646 &ext4_es_is_mapped, lblk)) {
1647 ret = ext4_clu_mapped(inode,
1648 EXT4_B2C(sbi, lblk));
1649 if (ret < 0)
1650 goto errout;
1651 if (ret == 0) {
1652 ret = ext4_da_reserve_space(inode);
1653 if (ret != 0) /* ENOSPC */
1654 goto errout;
1655 } else {
1656 allocated = true;
1657 }
1658 } else {
1659 allocated = true;
1660 }
1661 }
1662 }
1663
1664 ret = ext4_es_insert_delayed_block(inode, lblk, allocated);
1665
1666errout:
1667 return ret;
1668}
1669
1670/*
1671 * This function is grabs code from the very beginning of
1672 * ext4_map_blocks, but assumes that the caller is from delayed write
1673 * time. This function looks up the requested blocks and sets the
1674 * buffer delay bit under the protection of i_data_sem.
1675 */
1676static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1677 struct ext4_map_blocks *map,
1678 struct buffer_head *bh)
1679{
1680 struct extent_status es;
1681 int retval;
1682 sector_t invalid_block = ~((sector_t) 0xffff);
1683#ifdef ES_AGGRESSIVE_TEST
1684 struct ext4_map_blocks orig_map;
1685
1686 memcpy(&orig_map, map, sizeof(*map));
1687#endif
1688
1689 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1690 invalid_block = ~0;
1691
1692 map->m_flags = 0;
1693 ext_debug(inode, "max_blocks %u, logical block %lu\n", map->m_len,
1694 (unsigned long) map->m_lblk);
1695
1696 /* Lookup extent status tree firstly */
1697 if (ext4_es_lookup_extent(inode, iblock, NULL, &es)) {
1698 if (ext4_es_is_hole(&es)) {
1699 retval = 0;
1700 down_read(&EXT4_I(inode)->i_data_sem);
1701 goto add_delayed;
1702 }
1703
1704 /*
1705 * Delayed extent could be allocated by fallocate.
1706 * So we need to check it.
1707 */
1708 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1709 map_bh(bh, inode->i_sb, invalid_block);
1710 set_buffer_new(bh);
1711 set_buffer_delay(bh);
1712 return 0;
1713 }
1714
1715 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1716 retval = es.es_len - (iblock - es.es_lblk);
1717 if (retval > map->m_len)
1718 retval = map->m_len;
1719 map->m_len = retval;
1720 if (ext4_es_is_written(&es))
1721 map->m_flags |= EXT4_MAP_MAPPED;
1722 else if (ext4_es_is_unwritten(&es))
1723 map->m_flags |= EXT4_MAP_UNWRITTEN;
1724 else
1725 BUG();
1726
1727#ifdef ES_AGGRESSIVE_TEST
1728 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1729#endif
1730 return retval;
1731 }
1732
1733 /*
1734 * Try to see if we can get the block without requesting a new
1735 * file system block.
1736 */
1737 down_read(&EXT4_I(inode)->i_data_sem);
1738 if (ext4_has_inline_data(inode))
1739 retval = 0;
1740 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1741 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1742 else
1743 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1744
1745add_delayed:
1746 if (retval == 0) {
1747 int ret;
1748
1749 /*
1750 * XXX: __block_prepare_write() unmaps passed block,
1751 * is it OK?
1752 */
1753
1754 ret = ext4_insert_delayed_block(inode, map->m_lblk);
1755 if (ret != 0) {
1756 retval = ret;
1757 goto out_unlock;
1758 }
1759
1760 map_bh(bh, inode->i_sb, invalid_block);
1761 set_buffer_new(bh);
1762 set_buffer_delay(bh);
1763 } else if (retval > 0) {
1764 int ret;
1765 unsigned int status;
1766
1767 if (unlikely(retval != map->m_len)) {
1768 ext4_warning(inode->i_sb,
1769 "ES len assertion failed for inode "
1770 "%lu: retval %d != map->m_len %d",
1771 inode->i_ino, retval, map->m_len);
1772 WARN_ON(1);
1773 }
1774
1775 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1776 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1777 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1778 map->m_pblk, status);
1779 if (ret != 0)
1780 retval = ret;
1781 }
1782
1783out_unlock:
1784 up_read((&EXT4_I(inode)->i_data_sem));
1785
1786 return retval;
1787}
1788
1789/*
1790 * This is a special get_block_t callback which is used by
1791 * ext4_da_write_begin(). It will either return mapped block or
1792 * reserve space for a single block.
1793 *
1794 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1795 * We also have b_blocknr = -1 and b_bdev initialized properly
1796 *
1797 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1798 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1799 * initialized properly.
1800 */
1801int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1802 struct buffer_head *bh, int create)
1803{
1804 struct ext4_map_blocks map;
1805 int ret = 0;
1806
1807 BUG_ON(create == 0);
1808 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1809
1810 map.m_lblk = iblock;
1811 map.m_len = 1;
1812
1813 /*
1814 * first, we need to know whether the block is allocated already
1815 * preallocated blocks are unmapped but should treated
1816 * the same as allocated blocks.
1817 */
1818 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1819 if (ret <= 0)
1820 return ret;
1821
1822 map_bh(bh, inode->i_sb, map.m_pblk);
1823 ext4_update_bh_state(bh, map.m_flags);
1824
1825 if (buffer_unwritten(bh)) {
1826 /* A delayed write to unwritten bh should be marked
1827 * new and mapped. Mapped ensures that we don't do
1828 * get_block multiple times when we write to the same
1829 * offset and new ensures that we do proper zero out
1830 * for partial write.
1831 */
1832 set_buffer_new(bh);
1833 set_buffer_mapped(bh);
1834 }
1835 return 0;
1836}
1837
1838static int bget_one(handle_t *handle, struct buffer_head *bh)
1839{
1840 get_bh(bh);
1841 return 0;
1842}
1843
1844static int bput_one(handle_t *handle, struct buffer_head *bh)
1845{
1846 put_bh(bh);
1847 return 0;
1848}
1849
1850static int __ext4_journalled_writepage(struct page *page,
1851 unsigned int len)
1852{
1853 struct address_space *mapping = page->mapping;
1854 struct inode *inode = mapping->host;
1855 struct buffer_head *page_bufs = NULL;
1856 handle_t *handle = NULL;
1857 int ret = 0, err = 0;
1858 int inline_data = ext4_has_inline_data(inode);
1859 struct buffer_head *inode_bh = NULL;
1860
1861 ClearPageChecked(page);
1862
1863 if (inline_data) {
1864 BUG_ON(page->index != 0);
1865 BUG_ON(len > ext4_get_max_inline_size(inode));
1866 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1867 if (inode_bh == NULL)
1868 goto out;
1869 } else {
1870 page_bufs = page_buffers(page);
1871 if (!page_bufs) {
1872 BUG();
1873 goto out;
1874 }
1875 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1876 NULL, bget_one);
1877 }
1878 /*
1879 * We need to release the page lock before we start the
1880 * journal, so grab a reference so the page won't disappear
1881 * out from under us.
1882 */
1883 get_page(page);
1884 unlock_page(page);
1885
1886 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1887 ext4_writepage_trans_blocks(inode));
1888 if (IS_ERR(handle)) {
1889 ret = PTR_ERR(handle);
1890 put_page(page);
1891 goto out_no_pagelock;
1892 }
1893 BUG_ON(!ext4_handle_valid(handle));
1894
1895 lock_page(page);
1896 put_page(page);
1897 if (page->mapping != mapping) {
1898 /* The page got truncated from under us */
1899 ext4_journal_stop(handle);
1900 ret = 0;
1901 goto out;
1902 }
1903
1904 if (inline_data) {
1905 ret = ext4_mark_inode_dirty(handle, inode);
1906 } else {
1907 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1908 do_journal_get_write_access);
1909
1910 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1911 write_end_fn);
1912 }
1913 if (ret == 0)
1914 ret = err;
1915 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1916 err = ext4_journal_stop(handle);
1917 if (!ret)
1918 ret = err;
1919
1920 if (!ext4_has_inline_data(inode))
1921 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1922 NULL, bput_one);
1923 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1924out:
1925 unlock_page(page);
1926out_no_pagelock:
1927 brelse(inode_bh);
1928 return ret;
1929}
1930
1931/*
1932 * Note that we don't need to start a transaction unless we're journaling data
1933 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1934 * need to file the inode to the transaction's list in ordered mode because if
1935 * we are writing back data added by write(), the inode is already there and if
1936 * we are writing back data modified via mmap(), no one guarantees in which
1937 * transaction the data will hit the disk. In case we are journaling data, we
1938 * cannot start transaction directly because transaction start ranks above page
1939 * lock so we have to do some magic.
1940 *
1941 * This function can get called via...
1942 * - ext4_writepages after taking page lock (have journal handle)
1943 * - journal_submit_inode_data_buffers (no journal handle)
1944 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1945 * - grab_page_cache when doing write_begin (have journal handle)
1946 *
1947 * We don't do any block allocation in this function. If we have page with
1948 * multiple blocks we need to write those buffer_heads that are mapped. This
1949 * is important for mmaped based write. So if we do with blocksize 1K
1950 * truncate(f, 1024);
1951 * a = mmap(f, 0, 4096);
1952 * a[0] = 'a';
1953 * truncate(f, 4096);
1954 * we have in the page first buffer_head mapped via page_mkwrite call back
1955 * but other buffer_heads would be unmapped but dirty (dirty done via the
1956 * do_wp_page). So writepage should write the first block. If we modify
1957 * the mmap area beyond 1024 we will again get a page_fault and the
1958 * page_mkwrite callback will do the block allocation and mark the
1959 * buffer_heads mapped.
1960 *
1961 * We redirty the page if we have any buffer_heads that is either delay or
1962 * unwritten in the page.
1963 *
1964 * We can get recursively called as show below.
1965 *
1966 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1967 * ext4_writepage()
1968 *
1969 * But since we don't do any block allocation we should not deadlock.
1970 * Page also have the dirty flag cleared so we don't get recurive page_lock.
1971 */
1972static int ext4_writepage(struct page *page,
1973 struct writeback_control *wbc)
1974{
1975 int ret = 0;
1976 loff_t size;
1977 unsigned int len;
1978 struct buffer_head *page_bufs = NULL;
1979 struct inode *inode = page->mapping->host;
1980 struct ext4_io_submit io_submit;
1981 bool keep_towrite = false;
1982
1983 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
1984 inode->i_mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
1985 unlock_page(page);
1986 return -EIO;
1987 }
1988
1989 trace_ext4_writepage(page);
1990 size = i_size_read(inode);
1991 if (page->index == size >> PAGE_SHIFT &&
1992 !ext4_verity_in_progress(inode))
1993 len = size & ~PAGE_MASK;
1994 else
1995 len = PAGE_SIZE;
1996
1997 page_bufs = page_buffers(page);
1998 /*
1999 * We cannot do block allocation or other extent handling in this
2000 * function. If there are buffers needing that, we have to redirty
2001 * the page. But we may reach here when we do a journal commit via
2002 * journal_submit_inode_data_buffers() and in that case we must write
2003 * allocated buffers to achieve data=ordered mode guarantees.
2004 *
2005 * Also, if there is only one buffer per page (the fs block
2006 * size == the page size), if one buffer needs block
2007 * allocation or needs to modify the extent tree to clear the
2008 * unwritten flag, we know that the page can't be written at
2009 * all, so we might as well refuse the write immediately.
2010 * Unfortunately if the block size != page size, we can't as
2011 * easily detect this case using ext4_walk_page_buffers(), but
2012 * for the extremely common case, this is an optimization that
2013 * skips a useless round trip through ext4_bio_write_page().
2014 */
2015 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2016 ext4_bh_delay_or_unwritten)) {
2017 redirty_page_for_writepage(wbc, page);
2018 if ((current->flags & PF_MEMALLOC) ||
2019 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2020 /*
2021 * For memory cleaning there's no point in writing only
2022 * some buffers. So just bail out. Warn if we came here
2023 * from direct reclaim.
2024 */
2025 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2026 == PF_MEMALLOC);
2027 unlock_page(page);
2028 return 0;
2029 }
2030 keep_towrite = true;
2031 }
2032
2033 if (PageChecked(page) && ext4_should_journal_data(inode))
2034 /*
2035 * It's mmapped pagecache. Add buffers and journal it. There
2036 * doesn't seem much point in redirtying the page here.
2037 */
2038 return __ext4_journalled_writepage(page, len);
2039
2040 ext4_io_submit_init(&io_submit, wbc);
2041 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2042 if (!io_submit.io_end) {
2043 redirty_page_for_writepage(wbc, page);
2044 unlock_page(page);
2045 return -ENOMEM;
2046 }
2047 ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2048 ext4_io_submit(&io_submit);
2049 /* Drop io_end reference we got from init */
2050 ext4_put_io_end_defer(io_submit.io_end);
2051 return ret;
2052}
2053
2054static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2055{
2056 int len;
2057 loff_t size;
2058 int err;
2059
2060 BUG_ON(page->index != mpd->first_page);
2061 clear_page_dirty_for_io(page);
2062 /*
2063 * We have to be very careful here! Nothing protects writeback path
2064 * against i_size changes and the page can be writeably mapped into
2065 * page tables. So an application can be growing i_size and writing
2066 * data through mmap while writeback runs. clear_page_dirty_for_io()
2067 * write-protects our page in page tables and the page cannot get
2068 * written to again until we release page lock. So only after
2069 * clear_page_dirty_for_io() we are safe to sample i_size for
2070 * ext4_bio_write_page() to zero-out tail of the written page. We rely
2071 * on the barrier provided by TestClearPageDirty in
2072 * clear_page_dirty_for_io() to make sure i_size is really sampled only
2073 * after page tables are updated.
2074 */
2075 size = i_size_read(mpd->inode);
2076 if (page->index == size >> PAGE_SHIFT &&
2077 !ext4_verity_in_progress(mpd->inode))
2078 len = size & ~PAGE_MASK;
2079 else
2080 len = PAGE_SIZE;
2081 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2082 if (!err)
2083 mpd->wbc->nr_to_write--;
2084 mpd->first_page++;
2085
2086 return err;
2087}
2088
2089#define BH_FLAGS (BIT(BH_Unwritten) | BIT(BH_Delay))
2090
2091/*
2092 * mballoc gives us at most this number of blocks...
2093 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2094 * The rest of mballoc seems to handle chunks up to full group size.
2095 */
2096#define MAX_WRITEPAGES_EXTENT_LEN 2048
2097
2098/*
2099 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2100 *
2101 * @mpd - extent of blocks
2102 * @lblk - logical number of the block in the file
2103 * @bh - buffer head we want to add to the extent
2104 *
2105 * The function is used to collect contig. blocks in the same state. If the
2106 * buffer doesn't require mapping for writeback and we haven't started the
2107 * extent of buffers to map yet, the function returns 'true' immediately - the
2108 * caller can write the buffer right away. Otherwise the function returns true
2109 * if the block has been added to the extent, false if the block couldn't be
2110 * added.
2111 */
2112static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2113 struct buffer_head *bh)
2114{
2115 struct ext4_map_blocks *map = &mpd->map;
2116
2117 /* Buffer that doesn't need mapping for writeback? */
2118 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2119 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2120 /* So far no extent to map => we write the buffer right away */
2121 if (map->m_len == 0)
2122 return true;
2123 return false;
2124 }
2125
2126 /* First block in the extent? */
2127 if (map->m_len == 0) {
2128 /* We cannot map unless handle is started... */
2129 if (!mpd->do_map)
2130 return false;
2131 map->m_lblk = lblk;
2132 map->m_len = 1;
2133 map->m_flags = bh->b_state & BH_FLAGS;
2134 return true;
2135 }
2136
2137 /* Don't go larger than mballoc is willing to allocate */
2138 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2139 return false;
2140
2141 /* Can we merge the block to our big extent? */
2142 if (lblk == map->m_lblk + map->m_len &&
2143 (bh->b_state & BH_FLAGS) == map->m_flags) {
2144 map->m_len++;
2145 return true;
2146 }
2147 return false;
2148}
2149
2150/*
2151 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2152 *
2153 * @mpd - extent of blocks for mapping
2154 * @head - the first buffer in the page
2155 * @bh - buffer we should start processing from
2156 * @lblk - logical number of the block in the file corresponding to @bh
2157 *
2158 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2159 * the page for IO if all buffers in this page were mapped and there's no
2160 * accumulated extent of buffers to map or add buffers in the page to the
2161 * extent of buffers to map. The function returns 1 if the caller can continue
2162 * by processing the next page, 0 if it should stop adding buffers to the
2163 * extent to map because we cannot extend it anymore. It can also return value
2164 * < 0 in case of error during IO submission.
2165 */
2166static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2167 struct buffer_head *head,
2168 struct buffer_head *bh,
2169 ext4_lblk_t lblk)
2170{
2171 struct inode *inode = mpd->inode;
2172 int err;
2173 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2174 >> inode->i_blkbits;
2175
2176 if (ext4_verity_in_progress(inode))
2177 blocks = EXT_MAX_BLOCKS;
2178
2179 do {
2180 BUG_ON(buffer_locked(bh));
2181
2182 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2183 /* Found extent to map? */
2184 if (mpd->map.m_len)
2185 return 0;
2186 /* Buffer needs mapping and handle is not started? */
2187 if (!mpd->do_map)
2188 return 0;
2189 /* Everything mapped so far and we hit EOF */
2190 break;
2191 }
2192 } while (lblk++, (bh = bh->b_this_page) != head);
2193 /* So far everything mapped? Submit the page for IO. */
2194 if (mpd->map.m_len == 0) {
2195 err = mpage_submit_page(mpd, head->b_page);
2196 if (err < 0)
2197 return err;
2198 }
2199 if (lblk >= blocks) {
2200 mpd->scanned_until_end = 1;
2201 return 0;
2202 }
2203 return 1;
2204}
2205
2206/*
2207 * mpage_process_page - update page buffers corresponding to changed extent and
2208 * may submit fully mapped page for IO
2209 *
2210 * @mpd - description of extent to map, on return next extent to map
2211 * @m_lblk - logical block mapping.
2212 * @m_pblk - corresponding physical mapping.
2213 * @map_bh - determines on return whether this page requires any further
2214 * mapping or not.
2215 * Scan given page buffers corresponding to changed extent and update buffer
2216 * state according to new extent state.
2217 * We map delalloc buffers to their physical location, clear unwritten bits.
2218 * If the given page is not fully mapped, we update @map to the next extent in
2219 * the given page that needs mapping & return @map_bh as true.
2220 */
2221static int mpage_process_page(struct mpage_da_data *mpd, struct page *page,
2222 ext4_lblk_t *m_lblk, ext4_fsblk_t *m_pblk,
2223 bool *map_bh)
2224{
2225 struct buffer_head *head, *bh;
2226 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2227 ext4_lblk_t lblk = *m_lblk;
2228 ext4_fsblk_t pblock = *m_pblk;
2229 int err = 0;
2230 int blkbits = mpd->inode->i_blkbits;
2231 ssize_t io_end_size = 0;
2232 struct ext4_io_end_vec *io_end_vec = ext4_last_io_end_vec(io_end);
2233
2234 bh = head = page_buffers(page);
2235 do {
2236 if (lblk < mpd->map.m_lblk)
2237 continue;
2238 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2239 /*
2240 * Buffer after end of mapped extent.
2241 * Find next buffer in the page to map.
2242 */
2243 mpd->map.m_len = 0;
2244 mpd->map.m_flags = 0;
2245 io_end_vec->size += io_end_size;
2246 io_end_size = 0;
2247
2248 err = mpage_process_page_bufs(mpd, head, bh, lblk);
2249 if (err > 0)
2250 err = 0;
2251 if (!err && mpd->map.m_len && mpd->map.m_lblk > lblk) {
2252 io_end_vec = ext4_alloc_io_end_vec(io_end);
2253 if (IS_ERR(io_end_vec)) {
2254 err = PTR_ERR(io_end_vec);
2255 goto out;
2256 }
2257 io_end_vec->offset = mpd->map.m_lblk << blkbits;
2258 }
2259 *map_bh = true;
2260 goto out;
2261 }
2262 if (buffer_delay(bh)) {
2263 clear_buffer_delay(bh);
2264 bh->b_blocknr = pblock++;
2265 }
2266 clear_buffer_unwritten(bh);
2267 io_end_size += (1 << blkbits);
2268 } while (lblk++, (bh = bh->b_this_page) != head);
2269
2270 io_end_vec->size += io_end_size;
2271 io_end_size = 0;
2272 *map_bh = false;
2273out:
2274 *m_lblk = lblk;
2275 *m_pblk = pblock;
2276 return err;
2277}
2278
2279/*
2280 * mpage_map_buffers - update buffers corresponding to changed extent and
2281 * submit fully mapped pages for IO
2282 *
2283 * @mpd - description of extent to map, on return next extent to map
2284 *
2285 * Scan buffers corresponding to changed extent (we expect corresponding pages
2286 * to be already locked) and update buffer state according to new extent state.
2287 * We map delalloc buffers to their physical location, clear unwritten bits,
2288 * and mark buffers as uninit when we perform writes to unwritten extents
2289 * and do extent conversion after IO is finished. If the last page is not fully
2290 * mapped, we update @map to the next extent in the last page that needs
2291 * mapping. Otherwise we submit the page for IO.
2292 */
2293static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2294{
2295 struct pagevec pvec;
2296 int nr_pages, i;
2297 struct inode *inode = mpd->inode;
2298 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2299 pgoff_t start, end;
2300 ext4_lblk_t lblk;
2301 ext4_fsblk_t pblock;
2302 int err;
2303 bool map_bh = false;
2304
2305 start = mpd->map.m_lblk >> bpp_bits;
2306 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2307 lblk = start << bpp_bits;
2308 pblock = mpd->map.m_pblk;
2309
2310 pagevec_init(&pvec);
2311 while (start <= end) {
2312 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2313 &start, end);
2314 if (nr_pages == 0)
2315 break;
2316 for (i = 0; i < nr_pages; i++) {
2317 struct page *page = pvec.pages[i];
2318
2319 err = mpage_process_page(mpd, page, &lblk, &pblock,
2320 &map_bh);
2321 /*
2322 * If map_bh is true, means page may require further bh
2323 * mapping, or maybe the page was submitted for IO.
2324 * So we return to call further extent mapping.
2325 */
2326 if (err < 0 || map_bh)
2327 goto out;
2328 /* Page fully mapped - let IO run! */
2329 err = mpage_submit_page(mpd, page);
2330 if (err < 0)
2331 goto out;
2332 }
2333 pagevec_release(&pvec);
2334 }
2335 /* Extent fully mapped and matches with page boundary. We are done. */
2336 mpd->map.m_len = 0;
2337 mpd->map.m_flags = 0;
2338 return 0;
2339out:
2340 pagevec_release(&pvec);
2341 return err;
2342}
2343
2344static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2345{
2346 struct inode *inode = mpd->inode;
2347 struct ext4_map_blocks *map = &mpd->map;
2348 int get_blocks_flags;
2349 int err, dioread_nolock;
2350
2351 trace_ext4_da_write_pages_extent(inode, map);
2352 /*
2353 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2354 * to convert an unwritten extent to be initialized (in the case
2355 * where we have written into one or more preallocated blocks). It is
2356 * possible that we're going to need more metadata blocks than
2357 * previously reserved. However we must not fail because we're in
2358 * writeback and there is nothing we can do about it so it might result
2359 * in data loss. So use reserved blocks to allocate metadata if
2360 * possible.
2361 *
2362 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2363 * the blocks in question are delalloc blocks. This indicates
2364 * that the blocks and quotas has already been checked when
2365 * the data was copied into the page cache.
2366 */
2367 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2368 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2369 EXT4_GET_BLOCKS_IO_SUBMIT;
2370 dioread_nolock = ext4_should_dioread_nolock(inode);
2371 if (dioread_nolock)
2372 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2373 if (map->m_flags & BIT(BH_Delay))
2374 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2375
2376 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2377 if (err < 0)
2378 return err;
2379 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2380 if (!mpd->io_submit.io_end->handle &&
2381 ext4_handle_valid(handle)) {
2382 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2383 handle->h_rsv_handle = NULL;
2384 }
2385 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2386 }
2387
2388 BUG_ON(map->m_len == 0);
2389 return 0;
2390}
2391
2392/*
2393 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2394 * mpd->len and submit pages underlying it for IO
2395 *
2396 * @handle - handle for journal operations
2397 * @mpd - extent to map
2398 * @give_up_on_write - we set this to true iff there is a fatal error and there
2399 * is no hope of writing the data. The caller should discard
2400 * dirty pages to avoid infinite loops.
2401 *
2402 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2403 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2404 * them to initialized or split the described range from larger unwritten
2405 * extent. Note that we need not map all the described range since allocation
2406 * can return less blocks or the range is covered by more unwritten extents. We
2407 * cannot map more because we are limited by reserved transaction credits. On
2408 * the other hand we always make sure that the last touched page is fully
2409 * mapped so that it can be written out (and thus forward progress is
2410 * guaranteed). After mapping we submit all mapped pages for IO.
2411 */
2412static int mpage_map_and_submit_extent(handle_t *handle,
2413 struct mpage_da_data *mpd,
2414 bool *give_up_on_write)
2415{
2416 struct inode *inode = mpd->inode;
2417 struct ext4_map_blocks *map = &mpd->map;
2418 int err;
2419 loff_t disksize;
2420 int progress = 0;
2421 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2422 struct ext4_io_end_vec *io_end_vec;
2423
2424 io_end_vec = ext4_alloc_io_end_vec(io_end);
2425 if (IS_ERR(io_end_vec))
2426 return PTR_ERR(io_end_vec);
2427 io_end_vec->offset = ((loff_t)map->m_lblk) << inode->i_blkbits;
2428 do {
2429 err = mpage_map_one_extent(handle, mpd);
2430 if (err < 0) {
2431 struct super_block *sb = inode->i_sb;
2432
2433 if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2434 EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2435 goto invalidate_dirty_pages;
2436 /*
2437 * Let the uper layers retry transient errors.
2438 * In the case of ENOSPC, if ext4_count_free_blocks()
2439 * is non-zero, a commit should free up blocks.
2440 */
2441 if ((err == -ENOMEM) ||
2442 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2443 if (progress)
2444 goto update_disksize;
2445 return err;
2446 }
2447 ext4_msg(sb, KERN_CRIT,
2448 "Delayed block allocation failed for "
2449 "inode %lu at logical offset %llu with"
2450 " max blocks %u with error %d",
2451 inode->i_ino,
2452 (unsigned long long)map->m_lblk,
2453 (unsigned)map->m_len, -err);
2454 ext4_msg(sb, KERN_CRIT,
2455 "This should not happen!! Data will "
2456 "be lost\n");
2457 if (err == -ENOSPC)
2458 ext4_print_free_blocks(inode);
2459 invalidate_dirty_pages:
2460 *give_up_on_write = true;
2461 return err;
2462 }
2463 progress = 1;
2464 /*
2465 * Update buffer state, submit mapped pages, and get us new
2466 * extent to map
2467 */
2468 err = mpage_map_and_submit_buffers(mpd);
2469 if (err < 0)
2470 goto update_disksize;
2471 } while (map->m_len);
2472
2473update_disksize:
2474 /*
2475 * Update on-disk size after IO is submitted. Races with
2476 * truncate are avoided by checking i_size under i_data_sem.
2477 */
2478 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2479 if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
2480 int err2;
2481 loff_t i_size;
2482
2483 down_write(&EXT4_I(inode)->i_data_sem);
2484 i_size = i_size_read(inode);
2485 if (disksize > i_size)
2486 disksize = i_size;
2487 if (disksize > EXT4_I(inode)->i_disksize)
2488 EXT4_I(inode)->i_disksize = disksize;
2489 up_write(&EXT4_I(inode)->i_data_sem);
2490 err2 = ext4_mark_inode_dirty(handle, inode);
2491 if (err2) {
2492 ext4_error_err(inode->i_sb, -err2,
2493 "Failed to mark inode %lu dirty",
2494 inode->i_ino);
2495 }
2496 if (!err)
2497 err = err2;
2498 }
2499 return err;
2500}
2501
2502/*
2503 * Calculate the total number of credits to reserve for one writepages
2504 * iteration. This is called from ext4_writepages(). We map an extent of
2505 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2506 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2507 * bpp - 1 blocks in bpp different extents.
2508 */
2509static int ext4_da_writepages_trans_blocks(struct inode *inode)
2510{
2511 int bpp = ext4_journal_blocks_per_page(inode);
2512
2513 return ext4_meta_trans_blocks(inode,
2514 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2515}
2516
2517/*
2518 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2519 * and underlying extent to map
2520 *
2521 * @mpd - where to look for pages
2522 *
2523 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2524 * IO immediately. When we find a page which isn't mapped we start accumulating
2525 * extent of buffers underlying these pages that needs mapping (formed by
2526 * either delayed or unwritten buffers). We also lock the pages containing
2527 * these buffers. The extent found is returned in @mpd structure (starting at
2528 * mpd->lblk with length mpd->len blocks).
2529 *
2530 * Note that this function can attach bios to one io_end structure which are
2531 * neither logically nor physically contiguous. Although it may seem as an
2532 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2533 * case as we need to track IO to all buffers underlying a page in one io_end.
2534 */
2535static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2536{
2537 struct address_space *mapping = mpd->inode->i_mapping;
2538 struct pagevec pvec;
2539 unsigned int nr_pages;
2540 long left = mpd->wbc->nr_to_write;
2541 pgoff_t index = mpd->first_page;
2542 pgoff_t end = mpd->last_page;
2543 xa_mark_t tag;
2544 int i, err = 0;
2545 int blkbits = mpd->inode->i_blkbits;
2546 ext4_lblk_t lblk;
2547 struct buffer_head *head;
2548
2549 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2550 tag = PAGECACHE_TAG_TOWRITE;
2551 else
2552 tag = PAGECACHE_TAG_DIRTY;
2553
2554 pagevec_init(&pvec);
2555 mpd->map.m_len = 0;
2556 mpd->next_page = index;
2557 while (index <= end) {
2558 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2559 tag);
2560 if (nr_pages == 0)
2561 break;
2562
2563 for (i = 0; i < nr_pages; i++) {
2564 struct page *page = pvec.pages[i];
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 mpd->scanned_until_end = 1;
2617 return 0;
2618out:
2619 pagevec_release(&pvec);
2620 return err;
2621}
2622
2623static int ext4_writepages(struct address_space *mapping,
2624 struct writeback_control *wbc)
2625{
2626 pgoff_t writeback_index = 0;
2627 long nr_to_write = wbc->nr_to_write;
2628 int range_whole = 0;
2629 int cycled = 1;
2630 handle_t *handle = NULL;
2631 struct mpage_da_data mpd;
2632 struct inode *inode = mapping->host;
2633 int needed_blocks, rsv_blocks = 0, ret = 0;
2634 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2635 struct blk_plug plug;
2636 bool give_up_on_write = false;
2637
2638 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2639 return -EIO;
2640
2641 percpu_down_read(&sbi->s_writepages_rwsem);
2642 trace_ext4_writepages(inode, wbc);
2643
2644 /*
2645 * No pages to write? This is mainly a kludge to avoid starting
2646 * a transaction for special inodes like journal inode on last iput()
2647 * because that could violate lock ordering on umount
2648 */
2649 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2650 goto out_writepages;
2651
2652 if (ext4_should_journal_data(inode)) {
2653 ret = generic_writepages(mapping, wbc);
2654 goto out_writepages;
2655 }
2656
2657 /*
2658 * If the filesystem has aborted, it is read-only, so return
2659 * right away instead of dumping stack traces later on that
2660 * will obscure the real source of the problem. We test
2661 * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2662 * the latter could be true if the filesystem is mounted
2663 * read-only, and in that case, ext4_writepages should
2664 * *never* be called, so if that ever happens, we would want
2665 * the stack trace.
2666 */
2667 if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2668 sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2669 ret = -EROFS;
2670 goto out_writepages;
2671 }
2672
2673 /*
2674 * If we have inline data and arrive here, it means that
2675 * we will soon create the block for the 1st page, so
2676 * we'd better clear the inline data here.
2677 */
2678 if (ext4_has_inline_data(inode)) {
2679 /* Just inode will be modified... */
2680 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2681 if (IS_ERR(handle)) {
2682 ret = PTR_ERR(handle);
2683 goto out_writepages;
2684 }
2685 BUG_ON(ext4_test_inode_state(inode,
2686 EXT4_STATE_MAY_INLINE_DATA));
2687 ext4_destroy_inline_data(handle, inode);
2688 ext4_journal_stop(handle);
2689 }
2690
2691 if (ext4_should_dioread_nolock(inode)) {
2692 /*
2693 * We may need to convert up to one extent per block in
2694 * the page and we may dirty the inode.
2695 */
2696 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2697 PAGE_SIZE >> inode->i_blkbits);
2698 }
2699
2700 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2701 range_whole = 1;
2702
2703 if (wbc->range_cyclic) {
2704 writeback_index = mapping->writeback_index;
2705 if (writeback_index)
2706 cycled = 0;
2707 mpd.first_page = writeback_index;
2708 mpd.last_page = -1;
2709 } else {
2710 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2711 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2712 }
2713
2714 mpd.inode = inode;
2715 mpd.wbc = wbc;
2716 ext4_io_submit_init(&mpd.io_submit, wbc);
2717retry:
2718 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2719 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2720 blk_start_plug(&plug);
2721
2722 /*
2723 * First writeback pages that don't need mapping - we can avoid
2724 * starting a transaction unnecessarily and also avoid being blocked
2725 * in the block layer on device congestion while having transaction
2726 * started.
2727 */
2728 mpd.do_map = 0;
2729 mpd.scanned_until_end = 0;
2730 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2731 if (!mpd.io_submit.io_end) {
2732 ret = -ENOMEM;
2733 goto unplug;
2734 }
2735 ret = mpage_prepare_extent_to_map(&mpd);
2736 /* Unlock pages we didn't use */
2737 mpage_release_unused_pages(&mpd, false);
2738 /* Submit prepared bio */
2739 ext4_io_submit(&mpd.io_submit);
2740 ext4_put_io_end_defer(mpd.io_submit.io_end);
2741 mpd.io_submit.io_end = NULL;
2742 if (ret < 0)
2743 goto unplug;
2744
2745 while (!mpd.scanned_until_end && wbc->nr_to_write > 0) {
2746 /* For each extent of pages we use new io_end */
2747 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2748 if (!mpd.io_submit.io_end) {
2749 ret = -ENOMEM;
2750 break;
2751 }
2752
2753 /*
2754 * We have two constraints: We find one extent to map and we
2755 * must always write out whole page (makes a difference when
2756 * blocksize < pagesize) so that we don't block on IO when we
2757 * try to write out the rest of the page. Journalled mode is
2758 * not supported by delalloc.
2759 */
2760 BUG_ON(ext4_should_journal_data(inode));
2761 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2762
2763 /* start a new transaction */
2764 handle = ext4_journal_start_with_reserve(inode,
2765 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2766 if (IS_ERR(handle)) {
2767 ret = PTR_ERR(handle);
2768 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2769 "%ld pages, ino %lu; err %d", __func__,
2770 wbc->nr_to_write, inode->i_ino, ret);
2771 /* Release allocated io_end */
2772 ext4_put_io_end(mpd.io_submit.io_end);
2773 mpd.io_submit.io_end = NULL;
2774 break;
2775 }
2776 mpd.do_map = 1;
2777
2778 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2779 ret = mpage_prepare_extent_to_map(&mpd);
2780 if (!ret && mpd.map.m_len)
2781 ret = mpage_map_and_submit_extent(handle, &mpd,
2782 &give_up_on_write);
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 mpd.do_map = 0;
2797 }
2798 /* Unlock pages we didn't use */
2799 mpage_release_unused_pages(&mpd, give_up_on_write);
2800 /* Submit prepared bio */
2801 ext4_io_submit(&mpd.io_submit);
2802
2803 /*
2804 * Drop our io_end reference we got from init. We have
2805 * to be careful and use deferred io_end finishing if
2806 * we are still holding the transaction as we can
2807 * release the last reference to io_end which may end
2808 * up doing unwritten extent conversion.
2809 */
2810 if (handle) {
2811 ext4_put_io_end_defer(mpd.io_submit.io_end);
2812 ext4_journal_stop(handle);
2813 } else
2814 ext4_put_io_end(mpd.io_submit.io_end);
2815 mpd.io_submit.io_end = NULL;
2816
2817 if (ret == -ENOSPC && sbi->s_journal) {
2818 /*
2819 * Commit the transaction which would
2820 * free blocks released in the transaction
2821 * and try again
2822 */
2823 jbd2_journal_force_commit_nested(sbi->s_journal);
2824 ret = 0;
2825 continue;
2826 }
2827 /* Fatal error - ENOMEM, EIO... */
2828 if (ret)
2829 break;
2830 }
2831unplug:
2832 blk_finish_plug(&plug);
2833 if (!ret && !cycled && wbc->nr_to_write > 0) {
2834 cycled = 1;
2835 mpd.last_page = writeback_index - 1;
2836 mpd.first_page = 0;
2837 goto retry;
2838 }
2839
2840 /* Update index */
2841 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2842 /*
2843 * Set the writeback_index so that range_cyclic
2844 * mode will write it back later
2845 */
2846 mapping->writeback_index = mpd.first_page;
2847
2848out_writepages:
2849 trace_ext4_writepages_result(inode, wbc, ret,
2850 nr_to_write - wbc->nr_to_write);
2851 percpu_up_read(&sbi->s_writepages_rwsem);
2852 return ret;
2853}
2854
2855static int ext4_dax_writepages(struct address_space *mapping,
2856 struct writeback_control *wbc)
2857{
2858 int ret;
2859 long nr_to_write = wbc->nr_to_write;
2860 struct inode *inode = mapping->host;
2861 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2862
2863 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2864 return -EIO;
2865
2866 percpu_down_read(&sbi->s_writepages_rwsem);
2867 trace_ext4_writepages(inode, wbc);
2868
2869 ret = dax_writeback_mapping_range(mapping, sbi->s_daxdev, wbc);
2870 trace_ext4_writepages_result(inode, wbc, ret,
2871 nr_to_write - wbc->nr_to_write);
2872 percpu_up_read(&sbi->s_writepages_rwsem);
2873 return ret;
2874}
2875
2876static int ext4_nonda_switch(struct super_block *sb)
2877{
2878 s64 free_clusters, dirty_clusters;
2879 struct ext4_sb_info *sbi = EXT4_SB(sb);
2880
2881 /*
2882 * switch to non delalloc mode if we are running low
2883 * on free block. The free block accounting via percpu
2884 * counters can get slightly wrong with percpu_counter_batch getting
2885 * accumulated on each CPU without updating global counters
2886 * Delalloc need an accurate free block accounting. So switch
2887 * to non delalloc when we are near to error range.
2888 */
2889 free_clusters =
2890 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2891 dirty_clusters =
2892 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2893 /*
2894 * Start pushing delalloc when 1/2 of free blocks are dirty.
2895 */
2896 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2897 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2898
2899 if (2 * free_clusters < 3 * dirty_clusters ||
2900 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2901 /*
2902 * free block count is less than 150% of dirty blocks
2903 * or free blocks is less than watermark
2904 */
2905 return 1;
2906 }
2907 return 0;
2908}
2909
2910/* We always reserve for an inode update; the superblock could be there too */
2911static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2912{
2913 if (likely(ext4_has_feature_large_file(inode->i_sb)))
2914 return 1;
2915
2916 if (pos + len <= 0x7fffffffULL)
2917 return 1;
2918
2919 /* We might need to update the superblock to set LARGE_FILE */
2920 return 2;
2921}
2922
2923static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2924 loff_t pos, unsigned len, unsigned flags,
2925 struct page **pagep, void **fsdata)
2926{
2927 int ret, retries = 0;
2928 struct page *page;
2929 pgoff_t index;
2930 struct inode *inode = mapping->host;
2931 handle_t *handle;
2932
2933 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2934 return -EIO;
2935
2936 index = pos >> PAGE_SHIFT;
2937
2938 if (ext4_nonda_switch(inode->i_sb) || S_ISLNK(inode->i_mode) ||
2939 ext4_verity_in_progress(inode)) {
2940 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2941 return ext4_write_begin(file, mapping, pos,
2942 len, flags, pagep, fsdata);
2943 }
2944 *fsdata = (void *)0;
2945 trace_ext4_da_write_begin(inode, pos, len, flags);
2946
2947 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2948 ret = ext4_da_write_inline_data_begin(mapping, inode,
2949 pos, len, flags,
2950 pagep, fsdata);
2951 if (ret < 0)
2952 return ret;
2953 if (ret == 1)
2954 return 0;
2955 }
2956
2957 /*
2958 * grab_cache_page_write_begin() can take a long time if the
2959 * system is thrashing due to memory pressure, or if the page
2960 * is being written back. So grab it first before we start
2961 * the transaction handle. This also allows us to allocate
2962 * the page (if needed) without using GFP_NOFS.
2963 */
2964retry_grab:
2965 page = grab_cache_page_write_begin(mapping, index, flags);
2966 if (!page)
2967 return -ENOMEM;
2968 unlock_page(page);
2969
2970 /*
2971 * With delayed allocation, we don't log the i_disksize update
2972 * if there is delayed block allocation. But we still need
2973 * to journalling the i_disksize update if writes to the end
2974 * of file which has an already mapped buffer.
2975 */
2976retry_journal:
2977 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2978 ext4_da_write_credits(inode, pos, len));
2979 if (IS_ERR(handle)) {
2980 put_page(page);
2981 return PTR_ERR(handle);
2982 }
2983
2984 lock_page(page);
2985 if (page->mapping != mapping) {
2986 /* The page got truncated from under us */
2987 unlock_page(page);
2988 put_page(page);
2989 ext4_journal_stop(handle);
2990 goto retry_grab;
2991 }
2992 /* In case writeback began while the page was unlocked */
2993 wait_for_stable_page(page);
2994
2995#ifdef CONFIG_FS_ENCRYPTION
2996 ret = ext4_block_write_begin(page, pos, len,
2997 ext4_da_get_block_prep);
2998#else
2999 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3000#endif
3001 if (ret < 0) {
3002 unlock_page(page);
3003 ext4_journal_stop(handle);
3004 /*
3005 * block_write_begin may have instantiated a few blocks
3006 * outside i_size. Trim these off again. Don't need
3007 * i_size_read because we hold i_mutex.
3008 */
3009 if (pos + len > inode->i_size)
3010 ext4_truncate_failed_write(inode);
3011
3012 if (ret == -ENOSPC &&
3013 ext4_should_retry_alloc(inode->i_sb, &retries))
3014 goto retry_journal;
3015
3016 put_page(page);
3017 return ret;
3018 }
3019
3020 *pagep = page;
3021 return ret;
3022}
3023
3024/*
3025 * Check if we should update i_disksize
3026 * when write to the end of file but not require block allocation
3027 */
3028static int ext4_da_should_update_i_disksize(struct page *page,
3029 unsigned long offset)
3030{
3031 struct buffer_head *bh;
3032 struct inode *inode = page->mapping->host;
3033 unsigned int idx;
3034 int i;
3035
3036 bh = page_buffers(page);
3037 idx = offset >> inode->i_blkbits;
3038
3039 for (i = 0; i < idx; i++)
3040 bh = bh->b_this_page;
3041
3042 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3043 return 0;
3044 return 1;
3045}
3046
3047static int ext4_da_write_end(struct file *file,
3048 struct address_space *mapping,
3049 loff_t pos, unsigned len, unsigned copied,
3050 struct page *page, void *fsdata)
3051{
3052 struct inode *inode = mapping->host;
3053 int ret = 0, ret2;
3054 handle_t *handle = ext4_journal_current_handle();
3055 loff_t new_i_size;
3056 unsigned long start, end;
3057 int write_mode = (int)(unsigned long)fsdata;
3058
3059 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3060 return ext4_write_end(file, mapping, pos,
3061 len, copied, page, fsdata);
3062
3063 trace_ext4_da_write_end(inode, pos, len, copied);
3064 start = pos & (PAGE_SIZE - 1);
3065 end = start + copied - 1;
3066
3067 /*
3068 * generic_write_end() will run mark_inode_dirty() if i_size
3069 * changes. So let's piggyback the i_disksize mark_inode_dirty
3070 * into that.
3071 */
3072 new_i_size = pos + copied;
3073 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3074 if (ext4_has_inline_data(inode) ||
3075 ext4_da_should_update_i_disksize(page, end)) {
3076 ext4_update_i_disksize(inode, new_i_size);
3077 /* We need to mark inode dirty even if
3078 * new_i_size is less that inode->i_size
3079 * bu greater than i_disksize.(hint delalloc)
3080 */
3081 ret = ext4_mark_inode_dirty(handle, inode);
3082 }
3083 }
3084
3085 if (write_mode != CONVERT_INLINE_DATA &&
3086 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3087 ext4_has_inline_data(inode))
3088 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3089 page);
3090 else
3091 ret2 = generic_write_end(file, mapping, pos, len, copied,
3092 page, fsdata);
3093
3094 copied = ret2;
3095 if (ret2 < 0)
3096 ret = ret2;
3097 ret2 = ext4_journal_stop(handle);
3098 if (unlikely(ret2 && !ret))
3099 ret = ret2;
3100
3101 return ret ? ret : copied;
3102}
3103
3104/*
3105 * Force all delayed allocation blocks to be allocated for a given inode.
3106 */
3107int ext4_alloc_da_blocks(struct inode *inode)
3108{
3109 trace_ext4_alloc_da_blocks(inode);
3110
3111 if (!EXT4_I(inode)->i_reserved_data_blocks)
3112 return 0;
3113
3114 /*
3115 * We do something simple for now. The filemap_flush() will
3116 * also start triggering a write of the data blocks, which is
3117 * not strictly speaking necessary (and for users of
3118 * laptop_mode, not even desirable). However, to do otherwise
3119 * would require replicating code paths in:
3120 *
3121 * ext4_writepages() ->
3122 * write_cache_pages() ---> (via passed in callback function)
3123 * __mpage_da_writepage() -->
3124 * mpage_add_bh_to_extent()
3125 * mpage_da_map_blocks()
3126 *
3127 * The problem is that write_cache_pages(), located in
3128 * mm/page-writeback.c, marks pages clean in preparation for
3129 * doing I/O, which is not desirable if we're not planning on
3130 * doing I/O at all.
3131 *
3132 * We could call write_cache_pages(), and then redirty all of
3133 * the pages by calling redirty_page_for_writepage() but that
3134 * would be ugly in the extreme. So instead we would need to
3135 * replicate parts of the code in the above functions,
3136 * simplifying them because we wouldn't actually intend to
3137 * write out the pages, but rather only collect contiguous
3138 * logical block extents, call the multi-block allocator, and
3139 * then update the buffer heads with the block allocations.
3140 *
3141 * For now, though, we'll cheat by calling filemap_flush(),
3142 * which will map the blocks, and start the I/O, but not
3143 * actually wait for the I/O to complete.
3144 */
3145 return filemap_flush(inode->i_mapping);
3146}
3147
3148/*
3149 * bmap() is special. It gets used by applications such as lilo and by
3150 * the swapper to find the on-disk block of a specific piece of data.
3151 *
3152 * Naturally, this is dangerous if the block concerned is still in the
3153 * journal. If somebody makes a swapfile on an ext4 data-journaling
3154 * filesystem and enables swap, then they may get a nasty shock when the
3155 * data getting swapped to that swapfile suddenly gets overwritten by
3156 * the original zero's written out previously to the journal and
3157 * awaiting writeback in the kernel's buffer cache.
3158 *
3159 * So, if we see any bmap calls here on a modified, data-journaled file,
3160 * take extra steps to flush any blocks which might be in the cache.
3161 */
3162static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3163{
3164 struct inode *inode = mapping->host;
3165 journal_t *journal;
3166 int err;
3167
3168 /*
3169 * We can get here for an inline file via the FIBMAP ioctl
3170 */
3171 if (ext4_has_inline_data(inode))
3172 return 0;
3173
3174 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3175 test_opt(inode->i_sb, DELALLOC)) {
3176 /*
3177 * With delalloc we want to sync the file
3178 * so that we can make sure we allocate
3179 * blocks for file
3180 */
3181 filemap_write_and_wait(mapping);
3182 }
3183
3184 if (EXT4_JOURNAL(inode) &&
3185 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3186 /*
3187 * This is a REALLY heavyweight approach, but the use of
3188 * bmap on dirty files is expected to be extremely rare:
3189 * only if we run lilo or swapon on a freshly made file
3190 * do we expect this to happen.
3191 *
3192 * (bmap requires CAP_SYS_RAWIO so this does not
3193 * represent an unprivileged user DOS attack --- we'd be
3194 * in trouble if mortal users could trigger this path at
3195 * will.)
3196 *
3197 * NB. EXT4_STATE_JDATA is not set on files other than
3198 * regular files. If somebody wants to bmap a directory
3199 * or symlink and gets confused because the buffer
3200 * hasn't yet been flushed to disk, they deserve
3201 * everything they get.
3202 */
3203
3204 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3205 journal = EXT4_JOURNAL(inode);
3206 jbd2_journal_lock_updates(journal);
3207 err = jbd2_journal_flush(journal);
3208 jbd2_journal_unlock_updates(journal);
3209
3210 if (err)
3211 return 0;
3212 }
3213
3214 return iomap_bmap(mapping, block, &ext4_iomap_ops);
3215}
3216
3217static int ext4_readpage(struct file *file, struct page *page)
3218{
3219 int ret = -EAGAIN;
3220 struct inode *inode = page->mapping->host;
3221
3222 trace_ext4_readpage(page);
3223
3224 if (ext4_has_inline_data(inode))
3225 ret = ext4_readpage_inline(inode, page);
3226
3227 if (ret == -EAGAIN)
3228 return ext4_mpage_readpages(inode, NULL, page);
3229
3230 return ret;
3231}
3232
3233static void ext4_readahead(struct readahead_control *rac)
3234{
3235 struct inode *inode = rac->mapping->host;
3236
3237 /* If the file has inline data, no need to do readahead. */
3238 if (ext4_has_inline_data(inode))
3239 return;
3240
3241 ext4_mpage_readpages(inode, rac, NULL);
3242}
3243
3244static void ext4_invalidatepage(struct page *page, unsigned int offset,
3245 unsigned int length)
3246{
3247 trace_ext4_invalidatepage(page, offset, length);
3248
3249 /* No journalling happens on data buffers when this function is used */
3250 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3251
3252 block_invalidatepage(page, offset, length);
3253}
3254
3255static int __ext4_journalled_invalidatepage(struct page *page,
3256 unsigned int offset,
3257 unsigned int length)
3258{
3259 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3260
3261 trace_ext4_journalled_invalidatepage(page, offset, length);
3262
3263 /*
3264 * If it's a full truncate we just forget about the pending dirtying
3265 */
3266 if (offset == 0 && length == PAGE_SIZE)
3267 ClearPageChecked(page);
3268
3269 return jbd2_journal_invalidatepage(journal, page, offset, length);
3270}
3271
3272/* Wrapper for aops... */
3273static void ext4_journalled_invalidatepage(struct page *page,
3274 unsigned int offset,
3275 unsigned int length)
3276{
3277 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3278}
3279
3280static int ext4_releasepage(struct page *page, gfp_t wait)
3281{
3282 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3283
3284 trace_ext4_releasepage(page);
3285
3286 /* Page has dirty journalled data -> cannot release */
3287 if (PageChecked(page))
3288 return 0;
3289 if (journal)
3290 return jbd2_journal_try_to_free_buffers(journal, page);
3291 else
3292 return try_to_free_buffers(page);
3293}
3294
3295static bool ext4_inode_datasync_dirty(struct inode *inode)
3296{
3297 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3298
3299 if (journal)
3300 return !jbd2_transaction_committed(journal,
3301 EXT4_I(inode)->i_datasync_tid);
3302 /* Any metadata buffers to write? */
3303 if (!list_empty(&inode->i_mapping->private_list))
3304 return true;
3305 return inode->i_state & I_DIRTY_DATASYNC;
3306}
3307
3308static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
3309 struct ext4_map_blocks *map, loff_t offset,
3310 loff_t length)
3311{
3312 u8 blkbits = inode->i_blkbits;
3313
3314 /*
3315 * Writes that span EOF might trigger an I/O size update on completion,
3316 * so consider them to be dirty for the purpose of O_DSYNC, even if
3317 * there is no other metadata changes being made or are pending.
3318 */
3319 iomap->flags = 0;
3320 if (ext4_inode_datasync_dirty(inode) ||
3321 offset + length > i_size_read(inode))
3322 iomap->flags |= IOMAP_F_DIRTY;
3323
3324 if (map->m_flags & EXT4_MAP_NEW)
3325 iomap->flags |= IOMAP_F_NEW;
3326
3327 iomap->bdev = inode->i_sb->s_bdev;
3328 iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
3329 iomap->offset = (u64) map->m_lblk << blkbits;
3330 iomap->length = (u64) map->m_len << blkbits;
3331
3332 if ((map->m_flags & EXT4_MAP_MAPPED) &&
3333 !ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3334 iomap->flags |= IOMAP_F_MERGED;
3335
3336 /*
3337 * Flags passed to ext4_map_blocks() for direct I/O writes can result
3338 * in m_flags having both EXT4_MAP_MAPPED and EXT4_MAP_UNWRITTEN bits
3339 * set. In order for any allocated unwritten extents to be converted
3340 * into written extents correctly within the ->end_io() handler, we
3341 * need to ensure that the iomap->type is set appropriately. Hence, the
3342 * reason why we need to check whether the EXT4_MAP_UNWRITTEN bit has
3343 * been set first.
3344 */
3345 if (map->m_flags & EXT4_MAP_UNWRITTEN) {
3346 iomap->type = IOMAP_UNWRITTEN;
3347 iomap->addr = (u64) map->m_pblk << blkbits;
3348 } else if (map->m_flags & EXT4_MAP_MAPPED) {
3349 iomap->type = IOMAP_MAPPED;
3350 iomap->addr = (u64) map->m_pblk << blkbits;
3351 } else {
3352 iomap->type = IOMAP_HOLE;
3353 iomap->addr = IOMAP_NULL_ADDR;
3354 }
3355}
3356
3357static int ext4_iomap_alloc(struct inode *inode, struct ext4_map_blocks *map,
3358 unsigned int flags)
3359{
3360 handle_t *handle;
3361 u8 blkbits = inode->i_blkbits;
3362 int ret, dio_credits, m_flags = 0, retries = 0;
3363
3364 /*
3365 * Trim the mapping request to the maximum value that we can map at
3366 * once for direct I/O.
3367 */
3368 if (map->m_len > DIO_MAX_BLOCKS)
3369 map->m_len = DIO_MAX_BLOCKS;
3370 dio_credits = ext4_chunk_trans_blocks(inode, map->m_len);
3371
3372retry:
3373 /*
3374 * Either we allocate blocks and then don't get an unwritten extent, so
3375 * in that case we have reserved enough credits. Or, the blocks are
3376 * already allocated and unwritten. In that case, the extent conversion
3377 * fits into the credits as well.
3378 */
3379 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
3380 if (IS_ERR(handle))
3381 return PTR_ERR(handle);
3382
3383 /*
3384 * DAX and direct I/O are the only two operations that are currently
3385 * supported with IOMAP_WRITE.
3386 */
3387 WARN_ON(!IS_DAX(inode) && !(flags & IOMAP_DIRECT));
3388 if (IS_DAX(inode))
3389 m_flags = EXT4_GET_BLOCKS_CREATE_ZERO;
3390 /*
3391 * We use i_size instead of i_disksize here because delalloc writeback
3392 * can complete at any point during the I/O and subsequently push the
3393 * i_disksize out to i_size. This could be beyond where direct I/O is
3394 * happening and thus expose allocated blocks to direct I/O reads.
3395 */
3396 else if ((map->m_lblk * (1 << blkbits)) >= i_size_read(inode))
3397 m_flags = EXT4_GET_BLOCKS_CREATE;
3398 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3399 m_flags = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3400
3401 ret = ext4_map_blocks(handle, inode, map, m_flags);
3402
3403 /*
3404 * We cannot fill holes in indirect tree based inodes as that could
3405 * expose stale data in the case of a crash. Use the magic error code
3406 * to fallback to buffered I/O.
3407 */
3408 if (!m_flags && !ret)
3409 ret = -ENOTBLK;
3410
3411 ext4_journal_stop(handle);
3412 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3413 goto retry;
3414
3415 return ret;
3416}
3417
3418
3419static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3420 unsigned flags, struct iomap *iomap, struct iomap *srcmap)
3421{
3422 int ret;
3423 struct ext4_map_blocks map;
3424 u8 blkbits = inode->i_blkbits;
3425
3426 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3427 return -EINVAL;
3428
3429 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3430 return -ERANGE;
3431
3432 /*
3433 * Calculate the first and last logical blocks respectively.
3434 */
3435 map.m_lblk = offset >> blkbits;
3436 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3437 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3438
3439 if (flags & IOMAP_WRITE)
3440 ret = ext4_iomap_alloc(inode, &map, flags);
3441 else
3442 ret = ext4_map_blocks(NULL, inode, &map, 0);
3443
3444 if (ret < 0)
3445 return ret;
3446
3447 ext4_set_iomap(inode, iomap, &map, offset, length);
3448
3449 return 0;
3450}
3451
3452static int ext4_iomap_overwrite_begin(struct inode *inode, loff_t offset,
3453 loff_t length, unsigned flags, struct iomap *iomap,
3454 struct iomap *srcmap)
3455{
3456 int ret;
3457
3458 /*
3459 * Even for writes we don't need to allocate blocks, so just pretend
3460 * we are reading to save overhead of starting a transaction.
3461 */
3462 flags &= ~IOMAP_WRITE;
3463 ret = ext4_iomap_begin(inode, offset, length, flags, iomap, srcmap);
3464 WARN_ON_ONCE(iomap->type != IOMAP_MAPPED);
3465 return ret;
3466}
3467
3468static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3469 ssize_t written, unsigned flags, struct iomap *iomap)
3470{
3471 /*
3472 * Check to see whether an error occurred while writing out the data to
3473 * the allocated blocks. If so, return the magic error code so that we
3474 * fallback to buffered I/O and attempt to complete the remainder of
3475 * the I/O. Any blocks that may have been allocated in preparation for
3476 * the direct I/O will be reused during buffered I/O.
3477 */
3478 if (flags & (IOMAP_WRITE | IOMAP_DIRECT) && written == 0)
3479 return -ENOTBLK;
3480
3481 return 0;
3482}
3483
3484const struct iomap_ops ext4_iomap_ops = {
3485 .iomap_begin = ext4_iomap_begin,
3486 .iomap_end = ext4_iomap_end,
3487};
3488
3489const struct iomap_ops ext4_iomap_overwrite_ops = {
3490 .iomap_begin = ext4_iomap_overwrite_begin,
3491 .iomap_end = ext4_iomap_end,
3492};
3493
3494static bool ext4_iomap_is_delalloc(struct inode *inode,
3495 struct ext4_map_blocks *map)
3496{
3497 struct extent_status es;
3498 ext4_lblk_t offset = 0, end = map->m_lblk + map->m_len - 1;
3499
3500 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
3501 map->m_lblk, end, &es);
3502
3503 if (!es.es_len || es.es_lblk > end)
3504 return false;
3505
3506 if (es.es_lblk > map->m_lblk) {
3507 map->m_len = es.es_lblk - map->m_lblk;
3508 return false;
3509 }
3510
3511 offset = map->m_lblk - es.es_lblk;
3512 map->m_len = es.es_len - offset;
3513
3514 return true;
3515}
3516
3517static int ext4_iomap_begin_report(struct inode *inode, loff_t offset,
3518 loff_t length, unsigned int flags,
3519 struct iomap *iomap, struct iomap *srcmap)
3520{
3521 int ret;
3522 bool delalloc = false;
3523 struct ext4_map_blocks map;
3524 u8 blkbits = inode->i_blkbits;
3525
3526 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3527 return -EINVAL;
3528
3529 if (ext4_has_inline_data(inode)) {
3530 ret = ext4_inline_data_iomap(inode, iomap);
3531 if (ret != -EAGAIN) {
3532 if (ret == 0 && offset >= iomap->length)
3533 ret = -ENOENT;
3534 return ret;
3535 }
3536 }
3537
3538 /*
3539 * Calculate the first and last logical block respectively.
3540 */
3541 map.m_lblk = offset >> blkbits;
3542 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3543 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3544
3545 /*
3546 * Fiemap callers may call for offset beyond s_bitmap_maxbytes.
3547 * So handle it here itself instead of querying ext4_map_blocks().
3548 * Since ext4_map_blocks() will warn about it and will return
3549 * -EIO error.
3550 */
3551 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3552 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3553
3554 if (offset >= sbi->s_bitmap_maxbytes) {
3555 map.m_flags = 0;
3556 goto set_iomap;
3557 }
3558 }
3559
3560 ret = ext4_map_blocks(NULL, inode, &map, 0);
3561 if (ret < 0)
3562 return ret;
3563 if (ret == 0)
3564 delalloc = ext4_iomap_is_delalloc(inode, &map);
3565
3566set_iomap:
3567 ext4_set_iomap(inode, iomap, &map, offset, length);
3568 if (delalloc && iomap->type == IOMAP_HOLE)
3569 iomap->type = IOMAP_DELALLOC;
3570
3571 return 0;
3572}
3573
3574const struct iomap_ops ext4_iomap_report_ops = {
3575 .iomap_begin = ext4_iomap_begin_report,
3576};
3577
3578/*
3579 * Pages can be marked dirty completely asynchronously from ext4's journalling
3580 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3581 * much here because ->set_page_dirty is called under VFS locks. The page is
3582 * not necessarily locked.
3583 *
3584 * We cannot just dirty the page and leave attached buffers clean, because the
3585 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3586 * or jbddirty because all the journalling code will explode.
3587 *
3588 * So what we do is to mark the page "pending dirty" and next time writepage
3589 * is called, propagate that into the buffers appropriately.
3590 */
3591static int ext4_journalled_set_page_dirty(struct page *page)
3592{
3593 SetPageChecked(page);
3594 return __set_page_dirty_nobuffers(page);
3595}
3596
3597static int ext4_set_page_dirty(struct page *page)
3598{
3599 WARN_ON_ONCE(!PageLocked(page) && !PageDirty(page));
3600 WARN_ON_ONCE(!page_has_buffers(page));
3601 return __set_page_dirty_buffers(page);
3602}
3603
3604static const struct address_space_operations ext4_aops = {
3605 .readpage = ext4_readpage,
3606 .readahead = ext4_readahead,
3607 .writepage = ext4_writepage,
3608 .writepages = ext4_writepages,
3609 .write_begin = ext4_write_begin,
3610 .write_end = ext4_write_end,
3611 .set_page_dirty = ext4_set_page_dirty,
3612 .bmap = ext4_bmap,
3613 .invalidatepage = ext4_invalidatepage,
3614 .releasepage = ext4_releasepage,
3615 .direct_IO = noop_direct_IO,
3616 .migratepage = buffer_migrate_page,
3617 .is_partially_uptodate = block_is_partially_uptodate,
3618 .error_remove_page = generic_error_remove_page,
3619};
3620
3621static const struct address_space_operations ext4_journalled_aops = {
3622 .readpage = ext4_readpage,
3623 .readahead = ext4_readahead,
3624 .writepage = ext4_writepage,
3625 .writepages = ext4_writepages,
3626 .write_begin = ext4_write_begin,
3627 .write_end = ext4_journalled_write_end,
3628 .set_page_dirty = ext4_journalled_set_page_dirty,
3629 .bmap = ext4_bmap,
3630 .invalidatepage = ext4_journalled_invalidatepage,
3631 .releasepage = ext4_releasepage,
3632 .direct_IO = noop_direct_IO,
3633 .is_partially_uptodate = block_is_partially_uptodate,
3634 .error_remove_page = generic_error_remove_page,
3635};
3636
3637static const struct address_space_operations ext4_da_aops = {
3638 .readpage = ext4_readpage,
3639 .readahead = ext4_readahead,
3640 .writepage = ext4_writepage,
3641 .writepages = ext4_writepages,
3642 .write_begin = ext4_da_write_begin,
3643 .write_end = ext4_da_write_end,
3644 .set_page_dirty = ext4_set_page_dirty,
3645 .bmap = ext4_bmap,
3646 .invalidatepage = ext4_invalidatepage,
3647 .releasepage = ext4_releasepage,
3648 .direct_IO = noop_direct_IO,
3649 .migratepage = buffer_migrate_page,
3650 .is_partially_uptodate = block_is_partially_uptodate,
3651 .error_remove_page = generic_error_remove_page,
3652};
3653
3654static const struct address_space_operations ext4_dax_aops = {
3655 .writepages = ext4_dax_writepages,
3656 .direct_IO = noop_direct_IO,
3657 .set_page_dirty = noop_set_page_dirty,
3658 .bmap = ext4_bmap,
3659 .invalidatepage = noop_invalidatepage,
3660};
3661
3662void ext4_set_aops(struct inode *inode)
3663{
3664 switch (ext4_inode_journal_mode(inode)) {
3665 case EXT4_INODE_ORDERED_DATA_MODE:
3666 case EXT4_INODE_WRITEBACK_DATA_MODE:
3667 break;
3668 case EXT4_INODE_JOURNAL_DATA_MODE:
3669 inode->i_mapping->a_ops = &ext4_journalled_aops;
3670 return;
3671 default:
3672 BUG();
3673 }
3674 if (IS_DAX(inode))
3675 inode->i_mapping->a_ops = &ext4_dax_aops;
3676 else if (test_opt(inode->i_sb, DELALLOC))
3677 inode->i_mapping->a_ops = &ext4_da_aops;
3678 else
3679 inode->i_mapping->a_ops = &ext4_aops;
3680}
3681
3682static int __ext4_block_zero_page_range(handle_t *handle,
3683 struct address_space *mapping, loff_t from, loff_t length)
3684{
3685 ext4_fsblk_t index = from >> PAGE_SHIFT;
3686 unsigned offset = from & (PAGE_SIZE-1);
3687 unsigned blocksize, pos;
3688 ext4_lblk_t iblock;
3689 struct inode *inode = mapping->host;
3690 struct buffer_head *bh;
3691 struct page *page;
3692 int err = 0;
3693
3694 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3695 mapping_gfp_constraint(mapping, ~__GFP_FS));
3696 if (!page)
3697 return -ENOMEM;
3698
3699 blocksize = inode->i_sb->s_blocksize;
3700
3701 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3702
3703 if (!page_has_buffers(page))
3704 create_empty_buffers(page, blocksize, 0);
3705
3706 /* Find the buffer that contains "offset" */
3707 bh = page_buffers(page);
3708 pos = blocksize;
3709 while (offset >= pos) {
3710 bh = bh->b_this_page;
3711 iblock++;
3712 pos += blocksize;
3713 }
3714 if (buffer_freed(bh)) {
3715 BUFFER_TRACE(bh, "freed: skip");
3716 goto unlock;
3717 }
3718 if (!buffer_mapped(bh)) {
3719 BUFFER_TRACE(bh, "unmapped");
3720 ext4_get_block(inode, iblock, bh, 0);
3721 /* unmapped? It's a hole - nothing to do */
3722 if (!buffer_mapped(bh)) {
3723 BUFFER_TRACE(bh, "still unmapped");
3724 goto unlock;
3725 }
3726 }
3727
3728 /* Ok, it's mapped. Make sure it's up-to-date */
3729 if (PageUptodate(page))
3730 set_buffer_uptodate(bh);
3731
3732 if (!buffer_uptodate(bh)) {
3733 err = -EIO;
3734 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
3735 wait_on_buffer(bh);
3736 /* Uhhuh. Read error. Complain and punt. */
3737 if (!buffer_uptodate(bh))
3738 goto unlock;
3739 if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
3740 /* We expect the key to be set. */
3741 BUG_ON(!fscrypt_has_encryption_key(inode));
3742 err = fscrypt_decrypt_pagecache_blocks(page, blocksize,
3743 bh_offset(bh));
3744 if (err) {
3745 clear_buffer_uptodate(bh);
3746 goto unlock;
3747 }
3748 }
3749 }
3750 if (ext4_should_journal_data(inode)) {
3751 BUFFER_TRACE(bh, "get write access");
3752 err = ext4_journal_get_write_access(handle, bh);
3753 if (err)
3754 goto unlock;
3755 }
3756 zero_user(page, offset, length);
3757 BUFFER_TRACE(bh, "zeroed end of block");
3758
3759 if (ext4_should_journal_data(inode)) {
3760 err = ext4_handle_dirty_metadata(handle, inode, bh);
3761 } else {
3762 err = 0;
3763 mark_buffer_dirty(bh);
3764 if (ext4_should_order_data(inode))
3765 err = ext4_jbd2_inode_add_write(handle, inode, from,
3766 length);
3767 }
3768
3769unlock:
3770 unlock_page(page);
3771 put_page(page);
3772 return err;
3773}
3774
3775/*
3776 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3777 * starting from file offset 'from'. The range to be zero'd must
3778 * be contained with in one block. If the specified range exceeds
3779 * the end of the block it will be shortened to end of the block
3780 * that cooresponds to 'from'
3781 */
3782static int ext4_block_zero_page_range(handle_t *handle,
3783 struct address_space *mapping, loff_t from, loff_t length)
3784{
3785 struct inode *inode = mapping->host;
3786 unsigned offset = from & (PAGE_SIZE-1);
3787 unsigned blocksize = inode->i_sb->s_blocksize;
3788 unsigned max = blocksize - (offset & (blocksize - 1));
3789
3790 /*
3791 * correct length if it does not fall between
3792 * 'from' and the end of the block
3793 */
3794 if (length > max || length < 0)
3795 length = max;
3796
3797 if (IS_DAX(inode)) {
3798 return iomap_zero_range(inode, from, length, NULL,
3799 &ext4_iomap_ops);
3800 }
3801 return __ext4_block_zero_page_range(handle, mapping, from, length);
3802}
3803
3804/*
3805 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3806 * up to the end of the block which corresponds to `from'.
3807 * This required during truncate. We need to physically zero the tail end
3808 * of that block so it doesn't yield old data if the file is later grown.
3809 */
3810static int ext4_block_truncate_page(handle_t *handle,
3811 struct address_space *mapping, loff_t from)
3812{
3813 unsigned offset = from & (PAGE_SIZE-1);
3814 unsigned length;
3815 unsigned blocksize;
3816 struct inode *inode = mapping->host;
3817
3818 /* If we are processing an encrypted inode during orphan list handling */
3819 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
3820 return 0;
3821
3822 blocksize = inode->i_sb->s_blocksize;
3823 length = blocksize - (offset & (blocksize - 1));
3824
3825 return ext4_block_zero_page_range(handle, mapping, from, length);
3826}
3827
3828int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3829 loff_t lstart, loff_t length)
3830{
3831 struct super_block *sb = inode->i_sb;
3832 struct address_space *mapping = inode->i_mapping;
3833 unsigned partial_start, partial_end;
3834 ext4_fsblk_t start, end;
3835 loff_t byte_end = (lstart + length - 1);
3836 int err = 0;
3837
3838 partial_start = lstart & (sb->s_blocksize - 1);
3839 partial_end = byte_end & (sb->s_blocksize - 1);
3840
3841 start = lstart >> sb->s_blocksize_bits;
3842 end = byte_end >> sb->s_blocksize_bits;
3843
3844 /* Handle partial zero within the single block */
3845 if (start == end &&
3846 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3847 err = ext4_block_zero_page_range(handle, mapping,
3848 lstart, length);
3849 return err;
3850 }
3851 /* Handle partial zero out on the start of the range */
3852 if (partial_start) {
3853 err = ext4_block_zero_page_range(handle, mapping,
3854 lstart, sb->s_blocksize);
3855 if (err)
3856 return err;
3857 }
3858 /* Handle partial zero out on the end of the range */
3859 if (partial_end != sb->s_blocksize - 1)
3860 err = ext4_block_zero_page_range(handle, mapping,
3861 byte_end - partial_end,
3862 partial_end + 1);
3863 return err;
3864}
3865
3866int ext4_can_truncate(struct inode *inode)
3867{
3868 if (S_ISREG(inode->i_mode))
3869 return 1;
3870 if (S_ISDIR(inode->i_mode))
3871 return 1;
3872 if (S_ISLNK(inode->i_mode))
3873 return !ext4_inode_is_fast_symlink(inode);
3874 return 0;
3875}
3876
3877/*
3878 * We have to make sure i_disksize gets properly updated before we truncate
3879 * page cache due to hole punching or zero range. Otherwise i_disksize update
3880 * can get lost as it may have been postponed to submission of writeback but
3881 * that will never happen after we truncate page cache.
3882 */
3883int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
3884 loff_t len)
3885{
3886 handle_t *handle;
3887 int ret;
3888
3889 loff_t size = i_size_read(inode);
3890
3891 WARN_ON(!inode_is_locked(inode));
3892 if (offset > size || offset + len < size)
3893 return 0;
3894
3895 if (EXT4_I(inode)->i_disksize >= size)
3896 return 0;
3897
3898 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
3899 if (IS_ERR(handle))
3900 return PTR_ERR(handle);
3901 ext4_update_i_disksize(inode, size);
3902 ret = ext4_mark_inode_dirty(handle, inode);
3903 ext4_journal_stop(handle);
3904
3905 return ret;
3906}
3907
3908static void ext4_wait_dax_page(struct ext4_inode_info *ei)
3909{
3910 up_write(&ei->i_mmap_sem);
3911 schedule();
3912 down_write(&ei->i_mmap_sem);
3913}
3914
3915int ext4_break_layouts(struct inode *inode)
3916{
3917 struct ext4_inode_info *ei = EXT4_I(inode);
3918 struct page *page;
3919 int error;
3920
3921 if (WARN_ON_ONCE(!rwsem_is_locked(&ei->i_mmap_sem)))
3922 return -EINVAL;
3923
3924 do {
3925 page = dax_layout_busy_page(inode->i_mapping);
3926 if (!page)
3927 return 0;
3928
3929 error = ___wait_var_event(&page->_refcount,
3930 atomic_read(&page->_refcount) == 1,
3931 TASK_INTERRUPTIBLE, 0, 0,
3932 ext4_wait_dax_page(ei));
3933 } while (error == 0);
3934
3935 return error;
3936}
3937
3938/*
3939 * ext4_punch_hole: punches a hole in a file by releasing the blocks
3940 * associated with the given offset and length
3941 *
3942 * @inode: File inode
3943 * @offset: The offset where the hole will begin
3944 * @len: The length of the hole
3945 *
3946 * Returns: 0 on success or negative on failure
3947 */
3948
3949int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
3950{
3951 struct super_block *sb = inode->i_sb;
3952 ext4_lblk_t first_block, stop_block;
3953 struct address_space *mapping = inode->i_mapping;
3954 loff_t first_block_offset, last_block_offset;
3955 handle_t *handle;
3956 unsigned int credits;
3957 int ret = 0, ret2 = 0;
3958
3959 trace_ext4_punch_hole(inode, offset, length, 0);
3960
3961 ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
3962 if (ext4_has_inline_data(inode)) {
3963 down_write(&EXT4_I(inode)->i_mmap_sem);
3964 ret = ext4_convert_inline_data(inode);
3965 up_write(&EXT4_I(inode)->i_mmap_sem);
3966 if (ret)
3967 return ret;
3968 }
3969
3970 /*
3971 * Write out all dirty pages to avoid race conditions
3972 * Then release them.
3973 */
3974 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3975 ret = filemap_write_and_wait_range(mapping, offset,
3976 offset + length - 1);
3977 if (ret)
3978 return ret;
3979 }
3980
3981 inode_lock(inode);
3982
3983 /* No need to punch hole beyond i_size */
3984 if (offset >= inode->i_size)
3985 goto out_mutex;
3986
3987 /*
3988 * If the hole extends beyond i_size, set the hole
3989 * to end after the page that contains i_size
3990 */
3991 if (offset + length > inode->i_size) {
3992 length = inode->i_size +
3993 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
3994 offset;
3995 }
3996
3997 if (offset & (sb->s_blocksize - 1) ||
3998 (offset + length) & (sb->s_blocksize - 1)) {
3999 /*
4000 * Attach jinode to inode for jbd2 if we do any zeroing of
4001 * partial block
4002 */
4003 ret = ext4_inode_attach_jinode(inode);
4004 if (ret < 0)
4005 goto out_mutex;
4006
4007 }
4008
4009 /* Wait all existing dio workers, newcomers will block on i_mutex */
4010 inode_dio_wait(inode);
4011
4012 /*
4013 * Prevent page faults from reinstantiating pages we have released from
4014 * page cache.
4015 */
4016 down_write(&EXT4_I(inode)->i_mmap_sem);
4017
4018 ret = ext4_break_layouts(inode);
4019 if (ret)
4020 goto out_dio;
4021
4022 first_block_offset = round_up(offset, sb->s_blocksize);
4023 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4024
4025 /* Now release the pages and zero block aligned part of pages*/
4026 if (last_block_offset > first_block_offset) {
4027 ret = ext4_update_disksize_before_punch(inode, offset, length);
4028 if (ret)
4029 goto out_dio;
4030 truncate_pagecache_range(inode, first_block_offset,
4031 last_block_offset);
4032 }
4033
4034 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4035 credits = ext4_writepage_trans_blocks(inode);
4036 else
4037 credits = ext4_blocks_for_truncate(inode);
4038 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4039 if (IS_ERR(handle)) {
4040 ret = PTR_ERR(handle);
4041 ext4_std_error(sb, ret);
4042 goto out_dio;
4043 }
4044
4045 ret = ext4_zero_partial_blocks(handle, inode, offset,
4046 length);
4047 if (ret)
4048 goto out_stop;
4049
4050 first_block = (offset + sb->s_blocksize - 1) >>
4051 EXT4_BLOCK_SIZE_BITS(sb);
4052 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4053
4054 /* If there are blocks to remove, do it */
4055 if (stop_block > first_block) {
4056
4057 down_write(&EXT4_I(inode)->i_data_sem);
4058 ext4_discard_preallocations(inode, 0);
4059
4060 ret = ext4_es_remove_extent(inode, first_block,
4061 stop_block - first_block);
4062 if (ret) {
4063 up_write(&EXT4_I(inode)->i_data_sem);
4064 goto out_stop;
4065 }
4066
4067 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4068 ret = ext4_ext_remove_space(inode, first_block,
4069 stop_block - 1);
4070 else
4071 ret = ext4_ind_remove_space(handle, inode, first_block,
4072 stop_block);
4073
4074 up_write(&EXT4_I(inode)->i_data_sem);
4075 }
4076 if (IS_SYNC(inode))
4077 ext4_handle_sync(handle);
4078
4079 inode->i_mtime = inode->i_ctime = current_time(inode);
4080 ret2 = ext4_mark_inode_dirty(handle, inode);
4081 if (unlikely(ret2))
4082 ret = ret2;
4083 if (ret >= 0)
4084 ext4_update_inode_fsync_trans(handle, inode, 1);
4085out_stop:
4086 ext4_journal_stop(handle);
4087out_dio:
4088 up_write(&EXT4_I(inode)->i_mmap_sem);
4089out_mutex:
4090 inode_unlock(inode);
4091 return ret;
4092}
4093
4094int ext4_inode_attach_jinode(struct inode *inode)
4095{
4096 struct ext4_inode_info *ei = EXT4_I(inode);
4097 struct jbd2_inode *jinode;
4098
4099 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4100 return 0;
4101
4102 jinode = jbd2_alloc_inode(GFP_KERNEL);
4103 spin_lock(&inode->i_lock);
4104 if (!ei->jinode) {
4105 if (!jinode) {
4106 spin_unlock(&inode->i_lock);
4107 return -ENOMEM;
4108 }
4109 ei->jinode = jinode;
4110 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4111 jinode = NULL;
4112 }
4113 spin_unlock(&inode->i_lock);
4114 if (unlikely(jinode != NULL))
4115 jbd2_free_inode(jinode);
4116 return 0;
4117}
4118
4119/*
4120 * ext4_truncate()
4121 *
4122 * We block out ext4_get_block() block instantiations across the entire
4123 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4124 * simultaneously on behalf of the same inode.
4125 *
4126 * As we work through the truncate and commit bits of it to the journal there
4127 * is one core, guiding principle: the file's tree must always be consistent on
4128 * disk. We must be able to restart the truncate after a crash.
4129 *
4130 * The file's tree may be transiently inconsistent in memory (although it
4131 * probably isn't), but whenever we close off and commit a journal transaction,
4132 * the contents of (the filesystem + the journal) must be consistent and
4133 * restartable. It's pretty simple, really: bottom up, right to left (although
4134 * left-to-right works OK too).
4135 *
4136 * Note that at recovery time, journal replay occurs *before* the restart of
4137 * truncate against the orphan inode list.
4138 *
4139 * The committed inode has the new, desired i_size (which is the same as
4140 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4141 * that this inode's truncate did not complete and it will again call
4142 * ext4_truncate() to have another go. So there will be instantiated blocks
4143 * to the right of the truncation point in a crashed ext4 filesystem. But
4144 * that's fine - as long as they are linked from the inode, the post-crash
4145 * ext4_truncate() run will find them and release them.
4146 */
4147int ext4_truncate(struct inode *inode)
4148{
4149 struct ext4_inode_info *ei = EXT4_I(inode);
4150 unsigned int credits;
4151 int err = 0, err2;
4152 handle_t *handle;
4153 struct address_space *mapping = inode->i_mapping;
4154
4155 /*
4156 * There is a possibility that we're either freeing the inode
4157 * or it's a completely new inode. In those cases we might not
4158 * have i_mutex locked because it's not necessary.
4159 */
4160 if (!(inode->i_state & (I_NEW|I_FREEING)))
4161 WARN_ON(!inode_is_locked(inode));
4162 trace_ext4_truncate_enter(inode);
4163
4164 if (!ext4_can_truncate(inode))
4165 goto out_trace;
4166
4167 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4168 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4169
4170 if (ext4_has_inline_data(inode)) {
4171 int has_inline = 1;
4172
4173 err = ext4_inline_data_truncate(inode, &has_inline);
4174 if (err || has_inline)
4175 goto out_trace;
4176 }
4177
4178 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4179 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4180 if (ext4_inode_attach_jinode(inode) < 0)
4181 goto out_trace;
4182 }
4183
4184 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4185 credits = ext4_writepage_trans_blocks(inode);
4186 else
4187 credits = ext4_blocks_for_truncate(inode);
4188
4189 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4190 if (IS_ERR(handle)) {
4191 err = PTR_ERR(handle);
4192 goto out_trace;
4193 }
4194
4195 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4196 ext4_block_truncate_page(handle, mapping, inode->i_size);
4197
4198 /*
4199 * We add the inode to the orphan list, so that if this
4200 * truncate spans multiple transactions, and we crash, we will
4201 * resume the truncate when the filesystem recovers. It also
4202 * marks the inode dirty, to catch the new size.
4203 *
4204 * Implication: the file must always be in a sane, consistent
4205 * truncatable state while each transaction commits.
4206 */
4207 err = ext4_orphan_add(handle, inode);
4208 if (err)
4209 goto out_stop;
4210
4211 down_write(&EXT4_I(inode)->i_data_sem);
4212
4213 ext4_discard_preallocations(inode, 0);
4214
4215 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4216 err = ext4_ext_truncate(handle, inode);
4217 else
4218 ext4_ind_truncate(handle, inode);
4219
4220 up_write(&ei->i_data_sem);
4221 if (err)
4222 goto out_stop;
4223
4224 if (IS_SYNC(inode))
4225 ext4_handle_sync(handle);
4226
4227out_stop:
4228 /*
4229 * If this was a simple ftruncate() and the file will remain alive,
4230 * then we need to clear up the orphan record which we created above.
4231 * However, if this was a real unlink then we were called by
4232 * ext4_evict_inode(), and we allow that function to clean up the
4233 * orphan info for us.
4234 */
4235 if (inode->i_nlink)
4236 ext4_orphan_del(handle, inode);
4237
4238 inode->i_mtime = inode->i_ctime = current_time(inode);
4239 err2 = ext4_mark_inode_dirty(handle, inode);
4240 if (unlikely(err2 && !err))
4241 err = err2;
4242 ext4_journal_stop(handle);
4243
4244out_trace:
4245 trace_ext4_truncate_exit(inode);
4246 return err;
4247}
4248
4249/*
4250 * ext4_get_inode_loc returns with an extra refcount against the inode's
4251 * underlying buffer_head on success. If 'in_mem' is true, we have all
4252 * data in memory that is needed to recreate the on-disk version of this
4253 * inode.
4254 */
4255static int __ext4_get_inode_loc(struct inode *inode,
4256 struct ext4_iloc *iloc, int in_mem)
4257{
4258 struct ext4_group_desc *gdp;
4259 struct buffer_head *bh;
4260 struct super_block *sb = inode->i_sb;
4261 ext4_fsblk_t block;
4262 struct blk_plug plug;
4263 int inodes_per_block, inode_offset;
4264
4265 iloc->bh = NULL;
4266 if (inode->i_ino < EXT4_ROOT_INO ||
4267 inode->i_ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4268 return -EFSCORRUPTED;
4269
4270 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4271 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4272 if (!gdp)
4273 return -EIO;
4274
4275 /*
4276 * Figure out the offset within the block group inode table
4277 */
4278 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4279 inode_offset = ((inode->i_ino - 1) %
4280 EXT4_INODES_PER_GROUP(sb));
4281 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4282 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4283
4284 bh = sb_getblk(sb, block);
4285 if (unlikely(!bh))
4286 return -ENOMEM;
4287 if (ext4_simulate_fail(sb, EXT4_SIM_INODE_EIO))
4288 goto simulate_eio;
4289 if (!buffer_uptodate(bh)) {
4290 lock_buffer(bh);
4291
4292 /*
4293 * If the buffer has the write error flag, we have failed
4294 * to write out another inode in the same block. In this
4295 * case, we don't have to read the block because we may
4296 * read the old inode data successfully.
4297 */
4298 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4299 set_buffer_uptodate(bh);
4300
4301 if (buffer_uptodate(bh)) {
4302 /* someone brought it uptodate while we waited */
4303 unlock_buffer(bh);
4304 goto has_buffer;
4305 }
4306
4307 /*
4308 * If we have all information of the inode in memory and this
4309 * is the only valid inode in the block, we need not read the
4310 * block.
4311 */
4312 if (in_mem) {
4313 struct buffer_head *bitmap_bh;
4314 int i, start;
4315
4316 start = inode_offset & ~(inodes_per_block - 1);
4317
4318 /* Is the inode bitmap in cache? */
4319 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4320 if (unlikely(!bitmap_bh))
4321 goto make_io;
4322
4323 /*
4324 * If the inode bitmap isn't in cache then the
4325 * optimisation may end up performing two reads instead
4326 * of one, so skip it.
4327 */
4328 if (!buffer_uptodate(bitmap_bh)) {
4329 brelse(bitmap_bh);
4330 goto make_io;
4331 }
4332 for (i = start; i < start + inodes_per_block; i++) {
4333 if (i == inode_offset)
4334 continue;
4335 if (ext4_test_bit(i, bitmap_bh->b_data))
4336 break;
4337 }
4338 brelse(bitmap_bh);
4339 if (i == start + inodes_per_block) {
4340 /* all other inodes are free, so skip I/O */
4341 memset(bh->b_data, 0, bh->b_size);
4342 set_buffer_uptodate(bh);
4343 unlock_buffer(bh);
4344 goto has_buffer;
4345 }
4346 }
4347
4348make_io:
4349 /*
4350 * If we need to do any I/O, try to pre-readahead extra
4351 * blocks from the inode table.
4352 */
4353 blk_start_plug(&plug);
4354 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4355 ext4_fsblk_t b, end, table;
4356 unsigned num;
4357 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4358
4359 table = ext4_inode_table(sb, gdp);
4360 /* s_inode_readahead_blks is always a power of 2 */
4361 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4362 if (table > b)
4363 b = table;
4364 end = b + ra_blks;
4365 num = EXT4_INODES_PER_GROUP(sb);
4366 if (ext4_has_group_desc_csum(sb))
4367 num -= ext4_itable_unused_count(sb, gdp);
4368 table += num / inodes_per_block;
4369 if (end > table)
4370 end = table;
4371 while (b <= end)
4372 sb_breadahead_unmovable(sb, b++);
4373 }
4374
4375 /*
4376 * There are other valid inodes in the buffer, this inode
4377 * has in-inode xattrs, or we don't have this inode in memory.
4378 * Read the block from disk.
4379 */
4380 trace_ext4_load_inode(inode);
4381 get_bh(bh);
4382 bh->b_end_io = end_buffer_read_sync;
4383 submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
4384 blk_finish_plug(&plug);
4385 wait_on_buffer(bh);
4386 if (!buffer_uptodate(bh)) {
4387 simulate_eio:
4388 ext4_error_inode_block(inode, block, EIO,
4389 "unable to read itable block");
4390 brelse(bh);
4391 return -EIO;
4392 }
4393 }
4394has_buffer:
4395 iloc->bh = bh;
4396 return 0;
4397}
4398
4399int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4400{
4401 /* We have all inode data except xattrs in memory here. */
4402 return __ext4_get_inode_loc(inode, iloc,
4403 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4404}
4405
4406static bool ext4_should_enable_dax(struct inode *inode)
4407{
4408 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4409
4410 if (test_opt2(inode->i_sb, DAX_NEVER))
4411 return false;
4412 if (!S_ISREG(inode->i_mode))
4413 return false;
4414 if (ext4_should_journal_data(inode))
4415 return false;
4416 if (ext4_has_inline_data(inode))
4417 return false;
4418 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4419 return false;
4420 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4421 return false;
4422 if (!test_bit(EXT4_FLAGS_BDEV_IS_DAX, &sbi->s_ext4_flags))
4423 return false;
4424 if (test_opt(inode->i_sb, DAX_ALWAYS))
4425 return true;
4426
4427 return ext4_test_inode_flag(inode, EXT4_INODE_DAX);
4428}
4429
4430void ext4_set_inode_flags(struct inode *inode, bool init)
4431{
4432 unsigned int flags = EXT4_I(inode)->i_flags;
4433 unsigned int new_fl = 0;
4434
4435 WARN_ON_ONCE(IS_DAX(inode) && init);
4436
4437 if (flags & EXT4_SYNC_FL)
4438 new_fl |= S_SYNC;
4439 if (flags & EXT4_APPEND_FL)
4440 new_fl |= S_APPEND;
4441 if (flags & EXT4_IMMUTABLE_FL)
4442 new_fl |= S_IMMUTABLE;
4443 if (flags & EXT4_NOATIME_FL)
4444 new_fl |= S_NOATIME;
4445 if (flags & EXT4_DIRSYNC_FL)
4446 new_fl |= S_DIRSYNC;
4447
4448 /* Because of the way inode_set_flags() works we must preserve S_DAX
4449 * here if already set. */
4450 new_fl |= (inode->i_flags & S_DAX);
4451 if (init && ext4_should_enable_dax(inode))
4452 new_fl |= S_DAX;
4453
4454 if (flags & EXT4_ENCRYPT_FL)
4455 new_fl |= S_ENCRYPTED;
4456 if (flags & EXT4_CASEFOLD_FL)
4457 new_fl |= S_CASEFOLD;
4458 if (flags & EXT4_VERITY_FL)
4459 new_fl |= S_VERITY;
4460 inode_set_flags(inode, new_fl,
4461 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4462 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4463}
4464
4465static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4466 struct ext4_inode_info *ei)
4467{
4468 blkcnt_t i_blocks ;
4469 struct inode *inode = &(ei->vfs_inode);
4470 struct super_block *sb = inode->i_sb;
4471
4472 if (ext4_has_feature_huge_file(sb)) {
4473 /* we are using combined 48 bit field */
4474 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4475 le32_to_cpu(raw_inode->i_blocks_lo);
4476 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4477 /* i_blocks represent file system block size */
4478 return i_blocks << (inode->i_blkbits - 9);
4479 } else {
4480 return i_blocks;
4481 }
4482 } else {
4483 return le32_to_cpu(raw_inode->i_blocks_lo);
4484 }
4485}
4486
4487static inline int ext4_iget_extra_inode(struct inode *inode,
4488 struct ext4_inode *raw_inode,
4489 struct ext4_inode_info *ei)
4490{
4491 __le32 *magic = (void *)raw_inode +
4492 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4493
4494 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize + sizeof(__le32) <=
4495 EXT4_INODE_SIZE(inode->i_sb) &&
4496 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4497 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4498 return ext4_find_inline_data_nolock(inode);
4499 } else
4500 EXT4_I(inode)->i_inline_off = 0;
4501 return 0;
4502}
4503
4504int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4505{
4506 if (!ext4_has_feature_project(inode->i_sb))
4507 return -EOPNOTSUPP;
4508 *projid = EXT4_I(inode)->i_projid;
4509 return 0;
4510}
4511
4512/*
4513 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4514 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4515 * set.
4516 */
4517static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4518{
4519 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4520 inode_set_iversion_raw(inode, val);
4521 else
4522 inode_set_iversion_queried(inode, val);
4523}
4524static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4525{
4526 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4527 return inode_peek_iversion_raw(inode);
4528 else
4529 return inode_peek_iversion(inode);
4530}
4531
4532struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4533 ext4_iget_flags flags, const char *function,
4534 unsigned int line)
4535{
4536 struct ext4_iloc iloc;
4537 struct ext4_inode *raw_inode;
4538 struct ext4_inode_info *ei;
4539 struct inode *inode;
4540 journal_t *journal = EXT4_SB(sb)->s_journal;
4541 long ret;
4542 loff_t size;
4543 int block;
4544 uid_t i_uid;
4545 gid_t i_gid;
4546 projid_t i_projid;
4547
4548 if ((!(flags & EXT4_IGET_SPECIAL) &&
4549 (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)) ||
4550 (ino < EXT4_ROOT_INO) ||
4551 (ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))) {
4552 if (flags & EXT4_IGET_HANDLE)
4553 return ERR_PTR(-ESTALE);
4554 __ext4_error(sb, function, line, EFSCORRUPTED, 0,
4555 "inode #%lu: comm %s: iget: illegal inode #",
4556 ino, current->comm);
4557 return ERR_PTR(-EFSCORRUPTED);
4558 }
4559
4560 inode = iget_locked(sb, ino);
4561 if (!inode)
4562 return ERR_PTR(-ENOMEM);
4563 if (!(inode->i_state & I_NEW))
4564 return inode;
4565
4566 ei = EXT4_I(inode);
4567 iloc.bh = NULL;
4568
4569 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4570 if (ret < 0)
4571 goto bad_inode;
4572 raw_inode = ext4_raw_inode(&iloc);
4573
4574 if ((ino == EXT4_ROOT_INO) && (raw_inode->i_links_count == 0)) {
4575 ext4_error_inode(inode, function, line, 0,
4576 "iget: root inode unallocated");
4577 ret = -EFSCORRUPTED;
4578 goto bad_inode;
4579 }
4580
4581 if ((flags & EXT4_IGET_HANDLE) &&
4582 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4583 ret = -ESTALE;
4584 goto bad_inode;
4585 }
4586
4587 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4588 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4589 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4590 EXT4_INODE_SIZE(inode->i_sb) ||
4591 (ei->i_extra_isize & 3)) {
4592 ext4_error_inode(inode, function, line, 0,
4593 "iget: bad extra_isize %u "
4594 "(inode size %u)",
4595 ei->i_extra_isize,
4596 EXT4_INODE_SIZE(inode->i_sb));
4597 ret = -EFSCORRUPTED;
4598 goto bad_inode;
4599 }
4600 } else
4601 ei->i_extra_isize = 0;
4602
4603 /* Precompute checksum seed for inode metadata */
4604 if (ext4_has_metadata_csum(sb)) {
4605 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4606 __u32 csum;
4607 __le32 inum = cpu_to_le32(inode->i_ino);
4608 __le32 gen = raw_inode->i_generation;
4609 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4610 sizeof(inum));
4611 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4612 sizeof(gen));
4613 }
4614
4615 if (!ext4_inode_csum_verify(inode, raw_inode, ei) ||
4616 ext4_simulate_fail(sb, EXT4_SIM_INODE_CRC)) {
4617 ext4_error_inode_err(inode, function, line, 0, EFSBADCRC,
4618 "iget: checksum invalid");
4619 ret = -EFSBADCRC;
4620 goto bad_inode;
4621 }
4622
4623 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4624 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4625 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4626 if (ext4_has_feature_project(sb) &&
4627 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4628 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4629 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4630 else
4631 i_projid = EXT4_DEF_PROJID;
4632
4633 if (!(test_opt(inode->i_sb, NO_UID32))) {
4634 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4635 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4636 }
4637 i_uid_write(inode, i_uid);
4638 i_gid_write(inode, i_gid);
4639 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4640 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4641
4642 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4643 ei->i_inline_off = 0;
4644 ei->i_dir_start_lookup = 0;
4645 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4646 /* We now have enough fields to check if the inode was active or not.
4647 * This is needed because nfsd might try to access dead inodes
4648 * the test is that same one that e2fsck uses
4649 * NeilBrown 1999oct15
4650 */
4651 if (inode->i_nlink == 0) {
4652 if ((inode->i_mode == 0 ||
4653 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4654 ino != EXT4_BOOT_LOADER_INO) {
4655 /* this inode is deleted */
4656 ret = -ESTALE;
4657 goto bad_inode;
4658 }
4659 /* The only unlinked inodes we let through here have
4660 * valid i_mode and are being read by the orphan
4661 * recovery code: that's fine, we're about to complete
4662 * the process of deleting those.
4663 * OR it is the EXT4_BOOT_LOADER_INO which is
4664 * not initialized on a new filesystem. */
4665 }
4666 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4667 ext4_set_inode_flags(inode, true);
4668 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4669 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4670 if (ext4_has_feature_64bit(sb))
4671 ei->i_file_acl |=
4672 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4673 inode->i_size = ext4_isize(sb, raw_inode);
4674 if ((size = i_size_read(inode)) < 0) {
4675 ext4_error_inode(inode, function, line, 0,
4676 "iget: bad i_size value: %lld", size);
4677 ret = -EFSCORRUPTED;
4678 goto bad_inode;
4679 }
4680 /*
4681 * If dir_index is not enabled but there's dir with INDEX flag set,
4682 * we'd normally treat htree data as empty space. But with metadata
4683 * checksumming that corrupts checksums so forbid that.
4684 */
4685 if (!ext4_has_feature_dir_index(sb) && ext4_has_metadata_csum(sb) &&
4686 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
4687 ext4_error_inode(inode, function, line, 0,
4688 "iget: Dir with htree data on filesystem without dir_index feature.");
4689 ret = -EFSCORRUPTED;
4690 goto bad_inode;
4691 }
4692 ei->i_disksize = inode->i_size;
4693#ifdef CONFIG_QUOTA
4694 ei->i_reserved_quota = 0;
4695#endif
4696 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4697 ei->i_block_group = iloc.block_group;
4698 ei->i_last_alloc_group = ~0;
4699 /*
4700 * NOTE! The in-memory inode i_data array is in little-endian order
4701 * even on big-endian machines: we do NOT byteswap the block numbers!
4702 */
4703 for (block = 0; block < EXT4_N_BLOCKS; block++)
4704 ei->i_data[block] = raw_inode->i_block[block];
4705 INIT_LIST_HEAD(&ei->i_orphan);
4706
4707 /*
4708 * Set transaction id's of transactions that have to be committed
4709 * to finish f[data]sync. We set them to currently running transaction
4710 * as we cannot be sure that the inode or some of its metadata isn't
4711 * part of the transaction - the inode could have been reclaimed and
4712 * now it is reread from disk.
4713 */
4714 if (journal) {
4715 transaction_t *transaction;
4716 tid_t tid;
4717
4718 read_lock(&journal->j_state_lock);
4719 if (journal->j_running_transaction)
4720 transaction = journal->j_running_transaction;
4721 else
4722 transaction = journal->j_committing_transaction;
4723 if (transaction)
4724 tid = transaction->t_tid;
4725 else
4726 tid = journal->j_commit_sequence;
4727 read_unlock(&journal->j_state_lock);
4728 ei->i_sync_tid = tid;
4729 ei->i_datasync_tid = tid;
4730 }
4731
4732 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4733 if (ei->i_extra_isize == 0) {
4734 /* The extra space is currently unused. Use it. */
4735 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4736 ei->i_extra_isize = sizeof(struct ext4_inode) -
4737 EXT4_GOOD_OLD_INODE_SIZE;
4738 } else {
4739 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
4740 if (ret)
4741 goto bad_inode;
4742 }
4743 }
4744
4745 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4746 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4747 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4748 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4749
4750 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4751 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4752
4753 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4754 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4755 ivers |=
4756 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4757 }
4758 ext4_inode_set_iversion_queried(inode, ivers);
4759 }
4760
4761 ret = 0;
4762 if (ei->i_file_acl &&
4763 !ext4_inode_block_valid(inode, ei->i_file_acl, 1)) {
4764 ext4_error_inode(inode, function, line, 0,
4765 "iget: bad extended attribute block %llu",
4766 ei->i_file_acl);
4767 ret = -EFSCORRUPTED;
4768 goto bad_inode;
4769 } else if (!ext4_has_inline_data(inode)) {
4770 /* validate the block references in the inode */
4771 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4772 (S_ISLNK(inode->i_mode) &&
4773 !ext4_inode_is_fast_symlink(inode))) {
4774 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4775 ret = ext4_ext_check_inode(inode);
4776 else
4777 ret = ext4_ind_check_inode(inode);
4778 }
4779 }
4780 if (ret)
4781 goto bad_inode;
4782
4783 if (S_ISREG(inode->i_mode)) {
4784 inode->i_op = &ext4_file_inode_operations;
4785 inode->i_fop = &ext4_file_operations;
4786 ext4_set_aops(inode);
4787 } else if (S_ISDIR(inode->i_mode)) {
4788 inode->i_op = &ext4_dir_inode_operations;
4789 inode->i_fop = &ext4_dir_operations;
4790 } else if (S_ISLNK(inode->i_mode)) {
4791 /* VFS does not allow setting these so must be corruption */
4792 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
4793 ext4_error_inode(inode, function, line, 0,
4794 "iget: immutable or append flags "
4795 "not allowed on symlinks");
4796 ret = -EFSCORRUPTED;
4797 goto bad_inode;
4798 }
4799 if (IS_ENCRYPTED(inode)) {
4800 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4801 ext4_set_aops(inode);
4802 } else if (ext4_inode_is_fast_symlink(inode)) {
4803 inode->i_link = (char *)ei->i_data;
4804 inode->i_op = &ext4_fast_symlink_inode_operations;
4805 nd_terminate_link(ei->i_data, inode->i_size,
4806 sizeof(ei->i_data) - 1);
4807 } else {
4808 inode->i_op = &ext4_symlink_inode_operations;
4809 ext4_set_aops(inode);
4810 }
4811 inode_nohighmem(inode);
4812 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4813 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4814 inode->i_op = &ext4_special_inode_operations;
4815 if (raw_inode->i_block[0])
4816 init_special_inode(inode, inode->i_mode,
4817 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4818 else
4819 init_special_inode(inode, inode->i_mode,
4820 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4821 } else if (ino == EXT4_BOOT_LOADER_INO) {
4822 make_bad_inode(inode);
4823 } else {
4824 ret = -EFSCORRUPTED;
4825 ext4_error_inode(inode, function, line, 0,
4826 "iget: bogus i_mode (%o)", inode->i_mode);
4827 goto bad_inode;
4828 }
4829 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb))
4830 ext4_error_inode(inode, function, line, 0,
4831 "casefold flag without casefold feature");
4832 brelse(iloc.bh);
4833
4834 unlock_new_inode(inode);
4835 return inode;
4836
4837bad_inode:
4838 brelse(iloc.bh);
4839 iget_failed(inode);
4840 return ERR_PTR(ret);
4841}
4842
4843static int ext4_inode_blocks_set(handle_t *handle,
4844 struct ext4_inode *raw_inode,
4845 struct ext4_inode_info *ei)
4846{
4847 struct inode *inode = &(ei->vfs_inode);
4848 u64 i_blocks = READ_ONCE(inode->i_blocks);
4849 struct super_block *sb = inode->i_sb;
4850
4851 if (i_blocks <= ~0U) {
4852 /*
4853 * i_blocks can be represented in a 32 bit variable
4854 * as multiple of 512 bytes
4855 */
4856 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4857 raw_inode->i_blocks_high = 0;
4858 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4859 return 0;
4860 }
4861 if (!ext4_has_feature_huge_file(sb))
4862 return -EFBIG;
4863
4864 if (i_blocks <= 0xffffffffffffULL) {
4865 /*
4866 * i_blocks can be represented in a 48 bit variable
4867 * as multiple of 512 bytes
4868 */
4869 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4870 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4871 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4872 } else {
4873 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4874 /* i_block is stored in file system block size */
4875 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4876 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4877 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4878 }
4879 return 0;
4880}
4881
4882static void __ext4_update_other_inode_time(struct super_block *sb,
4883 unsigned long orig_ino,
4884 unsigned long ino,
4885 struct ext4_inode *raw_inode)
4886{
4887 struct inode *inode;
4888
4889 inode = find_inode_by_ino_rcu(sb, ino);
4890 if (!inode)
4891 return;
4892
4893 if ((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4894 I_DIRTY_INODE)) ||
4895 ((inode->i_state & I_DIRTY_TIME) == 0))
4896 return;
4897
4898 spin_lock(&inode->i_lock);
4899 if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4900 I_DIRTY_INODE)) == 0) &&
4901 (inode->i_state & I_DIRTY_TIME)) {
4902 struct ext4_inode_info *ei = EXT4_I(inode);
4903
4904 inode->i_state &= ~I_DIRTY_TIME;
4905 spin_unlock(&inode->i_lock);
4906
4907 spin_lock(&ei->i_raw_lock);
4908 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4909 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4910 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4911 ext4_inode_csum_set(inode, raw_inode, ei);
4912 spin_unlock(&ei->i_raw_lock);
4913 trace_ext4_other_inode_update_time(inode, orig_ino);
4914 return;
4915 }
4916 spin_unlock(&inode->i_lock);
4917}
4918
4919/*
4920 * Opportunistically update the other time fields for other inodes in
4921 * the same inode table block.
4922 */
4923static void ext4_update_other_inodes_time(struct super_block *sb,
4924 unsigned long orig_ino, char *buf)
4925{
4926 unsigned long ino;
4927 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4928 int inode_size = EXT4_INODE_SIZE(sb);
4929
4930 /*
4931 * Calculate the first inode in the inode table block. Inode
4932 * numbers are one-based. That is, the first inode in a block
4933 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
4934 */
4935 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
4936 rcu_read_lock();
4937 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
4938 if (ino == orig_ino)
4939 continue;
4940 __ext4_update_other_inode_time(sb, orig_ino, ino,
4941 (struct ext4_inode *)buf);
4942 }
4943 rcu_read_unlock();
4944}
4945
4946/*
4947 * Post the struct inode info into an on-disk inode location in the
4948 * buffer-cache. This gobbles the caller's reference to the
4949 * buffer_head in the inode location struct.
4950 *
4951 * The caller must have write access to iloc->bh.
4952 */
4953static int ext4_do_update_inode(handle_t *handle,
4954 struct inode *inode,
4955 struct ext4_iloc *iloc)
4956{
4957 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4958 struct ext4_inode_info *ei = EXT4_I(inode);
4959 struct buffer_head *bh = iloc->bh;
4960 struct super_block *sb = inode->i_sb;
4961 int err = 0, rc, block;
4962 int need_datasync = 0, set_large_file = 0;
4963 uid_t i_uid;
4964 gid_t i_gid;
4965 projid_t i_projid;
4966
4967 spin_lock(&ei->i_raw_lock);
4968
4969 /* For fields not tracked in the in-memory inode,
4970 * initialise them to zero for new inodes. */
4971 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
4972 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4973
4974 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4975 i_uid = i_uid_read(inode);
4976 i_gid = i_gid_read(inode);
4977 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4978 if (!(test_opt(inode->i_sb, NO_UID32))) {
4979 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4980 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4981/*
4982 * Fix up interoperability with old kernels. Otherwise, old inodes get
4983 * re-used with the upper 16 bits of the uid/gid intact
4984 */
4985 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4986 raw_inode->i_uid_high = 0;
4987 raw_inode->i_gid_high = 0;
4988 } else {
4989 raw_inode->i_uid_high =
4990 cpu_to_le16(high_16_bits(i_uid));
4991 raw_inode->i_gid_high =
4992 cpu_to_le16(high_16_bits(i_gid));
4993 }
4994 } else {
4995 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4996 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4997 raw_inode->i_uid_high = 0;
4998 raw_inode->i_gid_high = 0;
4999 }
5000 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5001
5002 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5003 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5004 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5005 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5006
5007 err = ext4_inode_blocks_set(handle, raw_inode, ei);
5008 if (err) {
5009 spin_unlock(&ei->i_raw_lock);
5010 goto out_brelse;
5011 }
5012 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5013 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
5014 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
5015 raw_inode->i_file_acl_high =
5016 cpu_to_le16(ei->i_file_acl >> 32);
5017 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5018 if (READ_ONCE(ei->i_disksize) != ext4_isize(inode->i_sb, raw_inode)) {
5019 ext4_isize_set(raw_inode, ei->i_disksize);
5020 need_datasync = 1;
5021 }
5022 if (ei->i_disksize > 0x7fffffffULL) {
5023 if (!ext4_has_feature_large_file(sb) ||
5024 EXT4_SB(sb)->s_es->s_rev_level ==
5025 cpu_to_le32(EXT4_GOOD_OLD_REV))
5026 set_large_file = 1;
5027 }
5028 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5029 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5030 if (old_valid_dev(inode->i_rdev)) {
5031 raw_inode->i_block[0] =
5032 cpu_to_le32(old_encode_dev(inode->i_rdev));
5033 raw_inode->i_block[1] = 0;
5034 } else {
5035 raw_inode->i_block[0] = 0;
5036 raw_inode->i_block[1] =
5037 cpu_to_le32(new_encode_dev(inode->i_rdev));
5038 raw_inode->i_block[2] = 0;
5039 }
5040 } else if (!ext4_has_inline_data(inode)) {
5041 for (block = 0; block < EXT4_N_BLOCKS; block++)
5042 raw_inode->i_block[block] = ei->i_data[block];
5043 }
5044
5045 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5046 u64 ivers = ext4_inode_peek_iversion(inode);
5047
5048 raw_inode->i_disk_version = cpu_to_le32(ivers);
5049 if (ei->i_extra_isize) {
5050 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5051 raw_inode->i_version_hi =
5052 cpu_to_le32(ivers >> 32);
5053 raw_inode->i_extra_isize =
5054 cpu_to_le16(ei->i_extra_isize);
5055 }
5056 }
5057
5058 BUG_ON(!ext4_has_feature_project(inode->i_sb) &&
5059 i_projid != EXT4_DEF_PROJID);
5060
5061 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
5062 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
5063 raw_inode->i_projid = cpu_to_le32(i_projid);
5064
5065 ext4_inode_csum_set(inode, raw_inode, ei);
5066 spin_unlock(&ei->i_raw_lock);
5067 if (inode->i_sb->s_flags & SB_LAZYTIME)
5068 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5069 bh->b_data);
5070
5071 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5072 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5073 if (!err)
5074 err = rc;
5075 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5076 if (set_large_file) {
5077 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5078 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
5079 if (err)
5080 goto out_brelse;
5081 ext4_set_feature_large_file(sb);
5082 ext4_handle_sync(handle);
5083 err = ext4_handle_dirty_super(handle, sb);
5084 }
5085 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5086out_brelse:
5087 brelse(bh);
5088 ext4_std_error(inode->i_sb, err);
5089 return err;
5090}
5091
5092/*
5093 * ext4_write_inode()
5094 *
5095 * We are called from a few places:
5096 *
5097 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5098 * Here, there will be no transaction running. We wait for any running
5099 * transaction to commit.
5100 *
5101 * - Within flush work (sys_sync(), kupdate and such).
5102 * We wait on commit, if told to.
5103 *
5104 * - Within iput_final() -> write_inode_now()
5105 * We wait on commit, if told to.
5106 *
5107 * In all cases it is actually safe for us to return without doing anything,
5108 * because the inode has been copied into a raw inode buffer in
5109 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5110 * writeback.
5111 *
5112 * Note that we are absolutely dependent upon all inode dirtiers doing the
5113 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5114 * which we are interested.
5115 *
5116 * It would be a bug for them to not do this. The code:
5117 *
5118 * mark_inode_dirty(inode)
5119 * stuff();
5120 * inode->i_size = expr;
5121 *
5122 * is in error because write_inode() could occur while `stuff()' is running,
5123 * and the new i_size will be lost. Plus the inode will no longer be on the
5124 * superblock's dirty inode list.
5125 */
5126int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5127{
5128 int err;
5129
5130 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC) ||
5131 sb_rdonly(inode->i_sb))
5132 return 0;
5133
5134 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5135 return -EIO;
5136
5137 if (EXT4_SB(inode->i_sb)->s_journal) {
5138 if (ext4_journal_current_handle()) {
5139 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5140 dump_stack();
5141 return -EIO;
5142 }
5143
5144 /*
5145 * No need to force transaction in WB_SYNC_NONE mode. Also
5146 * ext4_sync_fs() will force the commit after everything is
5147 * written.
5148 */
5149 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5150 return 0;
5151
5152 err = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5153 EXT4_I(inode)->i_sync_tid);
5154 } else {
5155 struct ext4_iloc iloc;
5156
5157 err = __ext4_get_inode_loc(inode, &iloc, 0);
5158 if (err)
5159 return err;
5160 /*
5161 * sync(2) will flush the whole buffer cache. No need to do
5162 * it here separately for each inode.
5163 */
5164 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5165 sync_dirty_buffer(iloc.bh);
5166 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5167 ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
5168 "IO error syncing inode");
5169 err = -EIO;
5170 }
5171 brelse(iloc.bh);
5172 }
5173 return err;
5174}
5175
5176/*
5177 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
5178 * buffers that are attached to a page stradding i_size and are undergoing
5179 * commit. In that case we have to wait for commit to finish and try again.
5180 */
5181static void ext4_wait_for_tail_page_commit(struct inode *inode)
5182{
5183 struct page *page;
5184 unsigned offset;
5185 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5186 tid_t commit_tid = 0;
5187 int ret;
5188
5189 offset = inode->i_size & (PAGE_SIZE - 1);
5190 /*
5191 * If the page is fully truncated, we don't need to wait for any commit
5192 * (and we even should not as __ext4_journalled_invalidatepage() may
5193 * strip all buffers from the page but keep the page dirty which can then
5194 * confuse e.g. concurrent ext4_writepage() seeing dirty page without
5195 * buffers). Also we don't need to wait for any commit if all buffers in
5196 * the page remain valid. This is most beneficial for the common case of
5197 * blocksize == PAGESIZE.
5198 */
5199 if (!offset || offset > (PAGE_SIZE - i_blocksize(inode)))
5200 return;
5201 while (1) {
5202 page = find_lock_page(inode->i_mapping,
5203 inode->i_size >> PAGE_SHIFT);
5204 if (!page)
5205 return;
5206 ret = __ext4_journalled_invalidatepage(page, offset,
5207 PAGE_SIZE - offset);
5208 unlock_page(page);
5209 put_page(page);
5210 if (ret != -EBUSY)
5211 return;
5212 commit_tid = 0;
5213 read_lock(&journal->j_state_lock);
5214 if (journal->j_committing_transaction)
5215 commit_tid = journal->j_committing_transaction->t_tid;
5216 read_unlock(&journal->j_state_lock);
5217 if (commit_tid)
5218 jbd2_log_wait_commit(journal, commit_tid);
5219 }
5220}
5221
5222/*
5223 * ext4_setattr()
5224 *
5225 * Called from notify_change.
5226 *
5227 * We want to trap VFS attempts to truncate the file as soon as
5228 * possible. In particular, we want to make sure that when the VFS
5229 * shrinks i_size, we put the inode on the orphan list and modify
5230 * i_disksize immediately, so that during the subsequent flushing of
5231 * dirty pages and freeing of disk blocks, we can guarantee that any
5232 * commit will leave the blocks being flushed in an unused state on
5233 * disk. (On recovery, the inode will get truncated and the blocks will
5234 * be freed, so we have a strong guarantee that no future commit will
5235 * leave these blocks visible to the user.)
5236 *
5237 * Another thing we have to assure is that if we are in ordered mode
5238 * and inode is still attached to the committing transaction, we must
5239 * we start writeout of all the dirty pages which are being truncated.
5240 * This way we are sure that all the data written in the previous
5241 * transaction are already on disk (truncate waits for pages under
5242 * writeback).
5243 *
5244 * Called with inode->i_mutex down.
5245 */
5246int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5247{
5248 struct inode *inode = d_inode(dentry);
5249 int error, rc = 0;
5250 int orphan = 0;
5251 const unsigned int ia_valid = attr->ia_valid;
5252
5253 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5254 return -EIO;
5255
5256 if (unlikely(IS_IMMUTABLE(inode)))
5257 return -EPERM;
5258
5259 if (unlikely(IS_APPEND(inode) &&
5260 (ia_valid & (ATTR_MODE | ATTR_UID |
5261 ATTR_GID | ATTR_TIMES_SET))))
5262 return -EPERM;
5263
5264 error = setattr_prepare(dentry, attr);
5265 if (error)
5266 return error;
5267
5268 error = fscrypt_prepare_setattr(dentry, attr);
5269 if (error)
5270 return error;
5271
5272 error = fsverity_prepare_setattr(dentry, attr);
5273 if (error)
5274 return error;
5275
5276 if (is_quota_modification(inode, attr)) {
5277 error = dquot_initialize(inode);
5278 if (error)
5279 return error;
5280 }
5281 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
5282 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
5283 handle_t *handle;
5284
5285 /* (user+group)*(old+new) structure, inode write (sb,
5286 * inode block, ? - but truncate inode update has it) */
5287 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5288 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5289 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5290 if (IS_ERR(handle)) {
5291 error = PTR_ERR(handle);
5292 goto err_out;
5293 }
5294
5295 /* dquot_transfer() calls back ext4_get_inode_usage() which
5296 * counts xattr inode references.
5297 */
5298 down_read(&EXT4_I(inode)->xattr_sem);
5299 error = dquot_transfer(inode, attr);
5300 up_read(&EXT4_I(inode)->xattr_sem);
5301
5302 if (error) {
5303 ext4_journal_stop(handle);
5304 return error;
5305 }
5306 /* Update corresponding info in inode so that everything is in
5307 * one transaction */
5308 if (attr->ia_valid & ATTR_UID)
5309 inode->i_uid = attr->ia_uid;
5310 if (attr->ia_valid & ATTR_GID)
5311 inode->i_gid = attr->ia_gid;
5312 error = ext4_mark_inode_dirty(handle, inode);
5313 ext4_journal_stop(handle);
5314 if (unlikely(error))
5315 return error;
5316 }
5317
5318 if (attr->ia_valid & ATTR_SIZE) {
5319 handle_t *handle;
5320 loff_t oldsize = inode->i_size;
5321 int shrink = (attr->ia_size < inode->i_size);
5322
5323 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5324 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5325
5326 if (attr->ia_size > sbi->s_bitmap_maxbytes)
5327 return -EFBIG;
5328 }
5329 if (!S_ISREG(inode->i_mode))
5330 return -EINVAL;
5331
5332 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
5333 inode_inc_iversion(inode);
5334
5335 if (shrink) {
5336 if (ext4_should_order_data(inode)) {
5337 error = ext4_begin_ordered_truncate(inode,
5338 attr->ia_size);
5339 if (error)
5340 goto err_out;
5341 }
5342 /*
5343 * Blocks are going to be removed from the inode. Wait
5344 * for dio in flight.
5345 */
5346 inode_dio_wait(inode);
5347 }
5348
5349 down_write(&EXT4_I(inode)->i_mmap_sem);
5350
5351 rc = ext4_break_layouts(inode);
5352 if (rc) {
5353 up_write(&EXT4_I(inode)->i_mmap_sem);
5354 return rc;
5355 }
5356
5357 if (attr->ia_size != inode->i_size) {
5358 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5359 if (IS_ERR(handle)) {
5360 error = PTR_ERR(handle);
5361 goto out_mmap_sem;
5362 }
5363 if (ext4_handle_valid(handle) && shrink) {
5364 error = ext4_orphan_add(handle, inode);
5365 orphan = 1;
5366 }
5367 /*
5368 * Update c/mtime on truncate up, ext4_truncate() will
5369 * update c/mtime in shrink case below
5370 */
5371 if (!shrink) {
5372 inode->i_mtime = current_time(inode);
5373 inode->i_ctime = inode->i_mtime;
5374 }
5375 down_write(&EXT4_I(inode)->i_data_sem);
5376 EXT4_I(inode)->i_disksize = attr->ia_size;
5377 rc = ext4_mark_inode_dirty(handle, inode);
5378 if (!error)
5379 error = rc;
5380 /*
5381 * We have to update i_size under i_data_sem together
5382 * with i_disksize to avoid races with writeback code
5383 * running ext4_wb_update_i_disksize().
5384 */
5385 if (!error)
5386 i_size_write(inode, attr->ia_size);
5387 up_write(&EXT4_I(inode)->i_data_sem);
5388 ext4_journal_stop(handle);
5389 if (error)
5390 goto out_mmap_sem;
5391 if (!shrink) {
5392 pagecache_isize_extended(inode, oldsize,
5393 inode->i_size);
5394 } else if (ext4_should_journal_data(inode)) {
5395 ext4_wait_for_tail_page_commit(inode);
5396 }
5397 }
5398
5399 /*
5400 * Truncate pagecache after we've waited for commit
5401 * in data=journal mode to make pages freeable.
5402 */
5403 truncate_pagecache(inode, inode->i_size);
5404 /*
5405 * Call ext4_truncate() even if i_size didn't change to
5406 * truncate possible preallocated blocks.
5407 */
5408 if (attr->ia_size <= oldsize) {
5409 rc = ext4_truncate(inode);
5410 if (rc)
5411 error = rc;
5412 }
5413out_mmap_sem:
5414 up_write(&EXT4_I(inode)->i_mmap_sem);
5415 }
5416
5417 if (!error) {
5418 setattr_copy(inode, attr);
5419 mark_inode_dirty(inode);
5420 }
5421
5422 /*
5423 * If the call to ext4_truncate failed to get a transaction handle at
5424 * all, we need to clean up the in-core orphan list manually.
5425 */
5426 if (orphan && inode->i_nlink)
5427 ext4_orphan_del(NULL, inode);
5428
5429 if (!error && (ia_valid & ATTR_MODE))
5430 rc = posix_acl_chmod(inode, inode->i_mode);
5431
5432err_out:
5433 ext4_std_error(inode->i_sb, error);
5434 if (!error)
5435 error = rc;
5436 return error;
5437}
5438
5439int ext4_getattr(const struct path *path, struct kstat *stat,
5440 u32 request_mask, unsigned int query_flags)
5441{
5442 struct inode *inode = d_inode(path->dentry);
5443 struct ext4_inode *raw_inode;
5444 struct ext4_inode_info *ei = EXT4_I(inode);
5445 unsigned int flags;
5446
5447 if ((request_mask & STATX_BTIME) &&
5448 EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5449 stat->result_mask |= STATX_BTIME;
5450 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5451 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5452 }
5453
5454 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5455 if (flags & EXT4_APPEND_FL)
5456 stat->attributes |= STATX_ATTR_APPEND;
5457 if (flags & EXT4_COMPR_FL)
5458 stat->attributes |= STATX_ATTR_COMPRESSED;
5459 if (flags & EXT4_ENCRYPT_FL)
5460 stat->attributes |= STATX_ATTR_ENCRYPTED;
5461 if (flags & EXT4_IMMUTABLE_FL)
5462 stat->attributes |= STATX_ATTR_IMMUTABLE;
5463 if (flags & EXT4_NODUMP_FL)
5464 stat->attributes |= STATX_ATTR_NODUMP;
5465 if (flags & EXT4_VERITY_FL)
5466 stat->attributes |= STATX_ATTR_VERITY;
5467
5468 stat->attributes_mask |= (STATX_ATTR_APPEND |
5469 STATX_ATTR_COMPRESSED |
5470 STATX_ATTR_ENCRYPTED |
5471 STATX_ATTR_IMMUTABLE |
5472 STATX_ATTR_NODUMP |
5473 STATX_ATTR_VERITY);
5474
5475 generic_fillattr(inode, stat);
5476 return 0;
5477}
5478
5479int ext4_file_getattr(const struct path *path, struct kstat *stat,
5480 u32 request_mask, unsigned int query_flags)
5481{
5482 struct inode *inode = d_inode(path->dentry);
5483 u64 delalloc_blocks;
5484
5485 ext4_getattr(path, stat, request_mask, query_flags);
5486
5487 /*
5488 * If there is inline data in the inode, the inode will normally not
5489 * have data blocks allocated (it may have an external xattr block).
5490 * Report at least one sector for such files, so tools like tar, rsync,
5491 * others don't incorrectly think the file is completely sparse.
5492 */
5493 if (unlikely(ext4_has_inline_data(inode)))
5494 stat->blocks += (stat->size + 511) >> 9;
5495
5496 /*
5497 * We can't update i_blocks if the block allocation is delayed
5498 * otherwise in the case of system crash before the real block
5499 * allocation is done, we will have i_blocks inconsistent with
5500 * on-disk file blocks.
5501 * We always keep i_blocks updated together with real
5502 * allocation. But to not confuse with user, stat
5503 * will return the blocks that include the delayed allocation
5504 * blocks for this file.
5505 */
5506 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5507 EXT4_I(inode)->i_reserved_data_blocks);
5508 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5509 return 0;
5510}
5511
5512static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5513 int pextents)
5514{
5515 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5516 return ext4_ind_trans_blocks(inode, lblocks);
5517 return ext4_ext_index_trans_blocks(inode, pextents);
5518}
5519
5520/*
5521 * Account for index blocks, block groups bitmaps and block group
5522 * descriptor blocks if modify datablocks and index blocks
5523 * worse case, the indexs blocks spread over different block groups
5524 *
5525 * If datablocks are discontiguous, they are possible to spread over
5526 * different block groups too. If they are contiguous, with flexbg,
5527 * they could still across block group boundary.
5528 *
5529 * Also account for superblock, inode, quota and xattr blocks
5530 */
5531static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5532 int pextents)
5533{
5534 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5535 int gdpblocks;
5536 int idxblocks;
5537 int ret = 0;
5538
5539 /*
5540 * How many index blocks need to touch to map @lblocks logical blocks
5541 * to @pextents physical extents?
5542 */
5543 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5544
5545 ret = idxblocks;
5546
5547 /*
5548 * Now let's see how many group bitmaps and group descriptors need
5549 * to account
5550 */
5551 groups = idxblocks + pextents;
5552 gdpblocks = groups;
5553 if (groups > ngroups)
5554 groups = ngroups;
5555 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5556 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5557
5558 /* bitmaps and block group descriptor blocks */
5559 ret += groups + gdpblocks;
5560
5561 /* Blocks for super block, inode, quota and xattr blocks */
5562 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5563
5564 return ret;
5565}
5566
5567/*
5568 * Calculate the total number of credits to reserve to fit
5569 * the modification of a single pages into a single transaction,
5570 * which may include multiple chunks of block allocations.
5571 *
5572 * This could be called via ext4_write_begin()
5573 *
5574 * We need to consider the worse case, when
5575 * one new block per extent.
5576 */
5577int ext4_writepage_trans_blocks(struct inode *inode)
5578{
5579 int bpp = ext4_journal_blocks_per_page(inode);
5580 int ret;
5581
5582 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5583
5584 /* Account for data blocks for journalled mode */
5585 if (ext4_should_journal_data(inode))
5586 ret += bpp;
5587 return ret;
5588}
5589
5590/*
5591 * Calculate the journal credits for a chunk of data modification.
5592 *
5593 * This is called from DIO, fallocate or whoever calling
5594 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5595 *
5596 * journal buffers for data blocks are not included here, as DIO
5597 * and fallocate do no need to journal data buffers.
5598 */
5599int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5600{
5601 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5602}
5603
5604/*
5605 * The caller must have previously called ext4_reserve_inode_write().
5606 * Give this, we know that the caller already has write access to iloc->bh.
5607 */
5608int ext4_mark_iloc_dirty(handle_t *handle,
5609 struct inode *inode, struct ext4_iloc *iloc)
5610{
5611 int err = 0;
5612
5613 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
5614 put_bh(iloc->bh);
5615 return -EIO;
5616 }
5617 if (IS_I_VERSION(inode))
5618 inode_inc_iversion(inode);
5619
5620 /* the do_update_inode consumes one bh->b_count */
5621 get_bh(iloc->bh);
5622
5623 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5624 err = ext4_do_update_inode(handle, inode, iloc);
5625 put_bh(iloc->bh);
5626 return err;
5627}
5628
5629/*
5630 * On success, We end up with an outstanding reference count against
5631 * iloc->bh. This _must_ be cleaned up later.
5632 */
5633
5634int
5635ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5636 struct ext4_iloc *iloc)
5637{
5638 int err;
5639
5640 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5641 return -EIO;
5642
5643 err = ext4_get_inode_loc(inode, iloc);
5644 if (!err) {
5645 BUFFER_TRACE(iloc->bh, "get_write_access");
5646 err = ext4_journal_get_write_access(handle, iloc->bh);
5647 if (err) {
5648 brelse(iloc->bh);
5649 iloc->bh = NULL;
5650 }
5651 }
5652 ext4_std_error(inode->i_sb, err);
5653 return err;
5654}
5655
5656static int __ext4_expand_extra_isize(struct inode *inode,
5657 unsigned int new_extra_isize,
5658 struct ext4_iloc *iloc,
5659 handle_t *handle, int *no_expand)
5660{
5661 struct ext4_inode *raw_inode;
5662 struct ext4_xattr_ibody_header *header;
5663 unsigned int inode_size = EXT4_INODE_SIZE(inode->i_sb);
5664 struct ext4_inode_info *ei = EXT4_I(inode);
5665 int error;
5666
5667 /* this was checked at iget time, but double check for good measure */
5668 if ((EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize > inode_size) ||
5669 (ei->i_extra_isize & 3)) {
5670 EXT4_ERROR_INODE(inode, "bad extra_isize %u (inode size %u)",
5671 ei->i_extra_isize,
5672 EXT4_INODE_SIZE(inode->i_sb));
5673 return -EFSCORRUPTED;
5674 }
5675 if ((new_extra_isize < ei->i_extra_isize) ||
5676 (new_extra_isize < 4) ||
5677 (new_extra_isize > inode_size - EXT4_GOOD_OLD_INODE_SIZE))
5678 return -EINVAL; /* Should never happen */
5679
5680 raw_inode = ext4_raw_inode(iloc);
5681
5682 header = IHDR(inode, raw_inode);
5683
5684 /* No extended attributes present */
5685 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5686 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5687 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5688 EXT4_I(inode)->i_extra_isize, 0,
5689 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5690 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5691 return 0;
5692 }
5693
5694 /* try to expand with EAs present */
5695 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5696 raw_inode, handle);
5697 if (error) {
5698 /*
5699 * Inode size expansion failed; don't try again
5700 */
5701 *no_expand = 1;
5702 }
5703
5704 return error;
5705}
5706
5707/*
5708 * Expand an inode by new_extra_isize bytes.
5709 * Returns 0 on success or negative error number on failure.
5710 */
5711static int ext4_try_to_expand_extra_isize(struct inode *inode,
5712 unsigned int new_extra_isize,
5713 struct ext4_iloc iloc,
5714 handle_t *handle)
5715{
5716 int no_expand;
5717 int error;
5718
5719 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5720 return -EOVERFLOW;
5721
5722 /*
5723 * In nojournal mode, we can immediately attempt to expand
5724 * the inode. When journaled, we first need to obtain extra
5725 * buffer credits since we may write into the EA block
5726 * with this same handle. If journal_extend fails, then it will
5727 * only result in a minor loss of functionality for that inode.
5728 * If this is felt to be critical, then e2fsck should be run to
5729 * force a large enough s_min_extra_isize.
5730 */
5731 if (ext4_journal_extend(handle,
5732 EXT4_DATA_TRANS_BLOCKS(inode->i_sb), 0) != 0)
5733 return -ENOSPC;
5734
5735 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5736 return -EBUSY;
5737
5738 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5739 handle, &no_expand);
5740 ext4_write_unlock_xattr(inode, &no_expand);
5741
5742 return error;
5743}
5744
5745int ext4_expand_extra_isize(struct inode *inode,
5746 unsigned int new_extra_isize,
5747 struct ext4_iloc *iloc)
5748{
5749 handle_t *handle;
5750 int no_expand;
5751 int error, rc;
5752
5753 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5754 brelse(iloc->bh);
5755 return -EOVERFLOW;
5756 }
5757
5758 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5759 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5760 if (IS_ERR(handle)) {
5761 error = PTR_ERR(handle);
5762 brelse(iloc->bh);
5763 return error;
5764 }
5765
5766 ext4_write_lock_xattr(inode, &no_expand);
5767
5768 BUFFER_TRACE(iloc->bh, "get_write_access");
5769 error = ext4_journal_get_write_access(handle, iloc->bh);
5770 if (error) {
5771 brelse(iloc->bh);
5772 goto out_unlock;
5773 }
5774
5775 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5776 handle, &no_expand);
5777
5778 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5779 if (!error)
5780 error = rc;
5781
5782out_unlock:
5783 ext4_write_unlock_xattr(inode, &no_expand);
5784 ext4_journal_stop(handle);
5785 return error;
5786}
5787
5788/*
5789 * What we do here is to mark the in-core inode as clean with respect to inode
5790 * dirtiness (it may still be data-dirty).
5791 * This means that the in-core inode may be reaped by prune_icache
5792 * without having to perform any I/O. This is a very good thing,
5793 * because *any* task may call prune_icache - even ones which
5794 * have a transaction open against a different journal.
5795 *
5796 * Is this cheating? Not really. Sure, we haven't written the
5797 * inode out, but prune_icache isn't a user-visible syncing function.
5798 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5799 * we start and wait on commits.
5800 */
5801int __ext4_mark_inode_dirty(handle_t *handle, struct inode *inode,
5802 const char *func, unsigned int line)
5803{
5804 struct ext4_iloc iloc;
5805 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5806 int err;
5807
5808 might_sleep();
5809 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5810 err = ext4_reserve_inode_write(handle, inode, &iloc);
5811 if (err)
5812 goto out;
5813
5814 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
5815 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
5816 iloc, handle);
5817
5818 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5819out:
5820 if (unlikely(err))
5821 ext4_error_inode_err(inode, func, line, 0, err,
5822 "mark_inode_dirty error");
5823 return err;
5824}
5825
5826/*
5827 * ext4_dirty_inode() is called from __mark_inode_dirty()
5828 *
5829 * We're really interested in the case where a file is being extended.
5830 * i_size has been changed by generic_commit_write() and we thus need
5831 * to include the updated inode in the current transaction.
5832 *
5833 * Also, dquot_alloc_block() will always dirty the inode when blocks
5834 * are allocated to the file.
5835 *
5836 * If the inode is marked synchronous, we don't honour that here - doing
5837 * so would cause a commit on atime updates, which we don't bother doing.
5838 * We handle synchronous inodes at the highest possible level.
5839 *
5840 * If only the I_DIRTY_TIME flag is set, we can skip everything. If
5841 * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
5842 * to copy into the on-disk inode structure are the timestamp files.
5843 */
5844void ext4_dirty_inode(struct inode *inode, int flags)
5845{
5846 handle_t *handle;
5847
5848 if (flags == I_DIRTY_TIME)
5849 return;
5850 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5851 if (IS_ERR(handle))
5852 goto out;
5853
5854 ext4_mark_inode_dirty(handle, inode);
5855
5856 ext4_journal_stop(handle);
5857out:
5858 return;
5859}
5860
5861int ext4_change_inode_journal_flag(struct inode *inode, int val)
5862{
5863 journal_t *journal;
5864 handle_t *handle;
5865 int err;
5866 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5867
5868 /*
5869 * We have to be very careful here: changing a data block's
5870 * journaling status dynamically is dangerous. If we write a
5871 * data block to the journal, change the status and then delete
5872 * that block, we risk forgetting to revoke the old log record
5873 * from the journal and so a subsequent replay can corrupt data.
5874 * So, first we make sure that the journal is empty and that
5875 * nobody is changing anything.
5876 */
5877
5878 journal = EXT4_JOURNAL(inode);
5879 if (!journal)
5880 return 0;
5881 if (is_journal_aborted(journal))
5882 return -EROFS;
5883
5884 /* Wait for all existing dio workers */
5885 inode_dio_wait(inode);
5886
5887 /*
5888 * Before flushing the journal and switching inode's aops, we have
5889 * to flush all dirty data the inode has. There can be outstanding
5890 * delayed allocations, there can be unwritten extents created by
5891 * fallocate or buffered writes in dioread_nolock mode covered by
5892 * dirty data which can be converted only after flushing the dirty
5893 * data (and journalled aops don't know how to handle these cases).
5894 */
5895 if (val) {
5896 down_write(&EXT4_I(inode)->i_mmap_sem);
5897 err = filemap_write_and_wait(inode->i_mapping);
5898 if (err < 0) {
5899 up_write(&EXT4_I(inode)->i_mmap_sem);
5900 return err;
5901 }
5902 }
5903
5904 percpu_down_write(&sbi->s_writepages_rwsem);
5905 jbd2_journal_lock_updates(journal);
5906
5907 /*
5908 * OK, there are no updates running now, and all cached data is
5909 * synced to disk. We are now in a completely consistent state
5910 * which doesn't have anything in the journal, and we know that
5911 * no filesystem updates are running, so it is safe to modify
5912 * the inode's in-core data-journaling state flag now.
5913 */
5914
5915 if (val)
5916 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5917 else {
5918 err = jbd2_journal_flush(journal);
5919 if (err < 0) {
5920 jbd2_journal_unlock_updates(journal);
5921 percpu_up_write(&sbi->s_writepages_rwsem);
5922 return err;
5923 }
5924 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5925 }
5926 ext4_set_aops(inode);
5927
5928 jbd2_journal_unlock_updates(journal);
5929 percpu_up_write(&sbi->s_writepages_rwsem);
5930
5931 if (val)
5932 up_write(&EXT4_I(inode)->i_mmap_sem);
5933
5934 /* Finally we can mark the inode as dirty. */
5935
5936 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
5937 if (IS_ERR(handle))
5938 return PTR_ERR(handle);
5939
5940 err = ext4_mark_inode_dirty(handle, inode);
5941 ext4_handle_sync(handle);
5942 ext4_journal_stop(handle);
5943 ext4_std_error(inode->i_sb, err);
5944
5945 return err;
5946}
5947
5948static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5949{
5950 return !buffer_mapped(bh);
5951}
5952
5953vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
5954{
5955 struct vm_area_struct *vma = vmf->vma;
5956 struct page *page = vmf->page;
5957 loff_t size;
5958 unsigned long len;
5959 int err;
5960 vm_fault_t ret;
5961 struct file *file = vma->vm_file;
5962 struct inode *inode = file_inode(file);
5963 struct address_space *mapping = inode->i_mapping;
5964 handle_t *handle;
5965 get_block_t *get_block;
5966 int retries = 0;
5967
5968 if (unlikely(IS_IMMUTABLE(inode)))
5969 return VM_FAULT_SIGBUS;
5970
5971 sb_start_pagefault(inode->i_sb);
5972 file_update_time(vma->vm_file);
5973
5974 down_read(&EXT4_I(inode)->i_mmap_sem);
5975
5976 err = ext4_convert_inline_data(inode);
5977 if (err)
5978 goto out_ret;
5979
5980 /* Delalloc case is easy... */
5981 if (test_opt(inode->i_sb, DELALLOC) &&
5982 !ext4_should_journal_data(inode) &&
5983 !ext4_nonda_switch(inode->i_sb)) {
5984 do {
5985 err = block_page_mkwrite(vma, vmf,
5986 ext4_da_get_block_prep);
5987 } while (err == -ENOSPC &&
5988 ext4_should_retry_alloc(inode->i_sb, &retries));
5989 goto out_ret;
5990 }
5991
5992 lock_page(page);
5993 size = i_size_read(inode);
5994 /* Page got truncated from under us? */
5995 if (page->mapping != mapping || page_offset(page) > size) {
5996 unlock_page(page);
5997 ret = VM_FAULT_NOPAGE;
5998 goto out;
5999 }
6000
6001 if (page->index == size >> PAGE_SHIFT)
6002 len = size & ~PAGE_MASK;
6003 else
6004 len = PAGE_SIZE;
6005 /*
6006 * Return if we have all the buffers mapped. This avoids the need to do
6007 * journal_start/journal_stop which can block and take a long time
6008 */
6009 if (page_has_buffers(page)) {
6010 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
6011 0, len, NULL,
6012 ext4_bh_unmapped)) {
6013 /* Wait so that we don't change page under IO */
6014 wait_for_stable_page(page);
6015 ret = VM_FAULT_LOCKED;
6016 goto out;
6017 }
6018 }
6019 unlock_page(page);
6020 /* OK, we need to fill the hole... */
6021 if (ext4_should_dioread_nolock(inode))
6022 get_block = ext4_get_block_unwritten;
6023 else
6024 get_block = ext4_get_block;
6025retry_alloc:
6026 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6027 ext4_writepage_trans_blocks(inode));
6028 if (IS_ERR(handle)) {
6029 ret = VM_FAULT_SIGBUS;
6030 goto out;
6031 }
6032 err = block_page_mkwrite(vma, vmf, get_block);
6033 if (!err && ext4_should_journal_data(inode)) {
6034 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
6035 PAGE_SIZE, NULL, do_journal_get_write_access)) {
6036 unlock_page(page);
6037 ret = VM_FAULT_SIGBUS;
6038 ext4_journal_stop(handle);
6039 goto out;
6040 }
6041 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
6042 }
6043 ext4_journal_stop(handle);
6044 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6045 goto retry_alloc;
6046out_ret:
6047 ret = block_page_mkwrite_return(err);
6048out:
6049 up_read(&EXT4_I(inode)->i_mmap_sem);
6050 sb_end_pagefault(inode->i_sb);
6051 return ret;
6052}
6053
6054vm_fault_t ext4_filemap_fault(struct vm_fault *vmf)
6055{
6056 struct inode *inode = file_inode(vmf->vma->vm_file);
6057 vm_fault_t ret;
6058
6059 down_read(&EXT4_I(inode)->i_mmap_sem);
6060 ret = filemap_fault(vmf);
6061 up_read(&EXT4_I(inode)->i_mmap_sem);
6062
6063 return ret;
6064}