Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/inode.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/inode.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * 64-bit file support on 64-bit platforms by Jakub Jelinek
17 * (jj@sunsite.ms.mff.cuni.cz)
18 *
19 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20 */
21
22#include <linux/fs.h>
23#include <linux/mount.h>
24#include <linux/time.h>
25#include <linux/highuid.h>
26#include <linux/pagemap.h>
27#include <linux/dax.h>
28#include <linux/quotaops.h>
29#include <linux/string.h>
30#include <linux/buffer_head.h>
31#include <linux/writeback.h>
32#include <linux/pagevec.h>
33#include <linux/mpage.h>
34#include <linux/namei.h>
35#include <linux/uio.h>
36#include <linux/bio.h>
37#include <linux/workqueue.h>
38#include <linux/kernel.h>
39#include <linux/printk.h>
40#include <linux/slab.h>
41#include <linux/bitops.h>
42#include <linux/iomap.h>
43#include <linux/iversion.h>
44
45#include "ext4_jbd2.h"
46#include "xattr.h"
47#include "acl.h"
48#include "truncate.h"
49
50#include <trace/events/ext4.h>
51
52static void ext4_journalled_zero_new_buffers(handle_t *handle,
53 struct inode *inode,
54 struct folio *folio,
55 unsigned from, unsigned to);
56
57static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
58 struct ext4_inode_info *ei)
59{
60 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
61 __u32 csum;
62 __u16 dummy_csum = 0;
63 int offset = offsetof(struct ext4_inode, i_checksum_lo);
64 unsigned int csum_size = sizeof(dummy_csum);
65
66 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
67 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
68 offset += csum_size;
69 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
70 EXT4_GOOD_OLD_INODE_SIZE - offset);
71
72 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
73 offset = offsetof(struct ext4_inode, i_checksum_hi);
74 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
75 EXT4_GOOD_OLD_INODE_SIZE,
76 offset - EXT4_GOOD_OLD_INODE_SIZE);
77 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
78 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
79 csum_size);
80 offset += csum_size;
81 }
82 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
83 EXT4_INODE_SIZE(inode->i_sb) - offset);
84 }
85
86 return csum;
87}
88
89static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
90 struct ext4_inode_info *ei)
91{
92 __u32 provided, calculated;
93
94 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
95 cpu_to_le32(EXT4_OS_LINUX) ||
96 !ext4_has_metadata_csum(inode->i_sb))
97 return 1;
98
99 provided = le16_to_cpu(raw->i_checksum_lo);
100 calculated = ext4_inode_csum(inode, raw, ei);
101 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
102 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
103 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
104 else
105 calculated &= 0xFFFF;
106
107 return provided == calculated;
108}
109
110void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
111 struct ext4_inode_info *ei)
112{
113 __u32 csum;
114
115 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
116 cpu_to_le32(EXT4_OS_LINUX) ||
117 !ext4_has_metadata_csum(inode->i_sb))
118 return;
119
120 csum = ext4_inode_csum(inode, raw, ei);
121 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
122 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
123 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
124 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
125}
126
127static inline int ext4_begin_ordered_truncate(struct inode *inode,
128 loff_t new_size)
129{
130 trace_ext4_begin_ordered_truncate(inode, new_size);
131 /*
132 * If jinode is zero, then we never opened the file for
133 * writing, so there's no need to call
134 * jbd2_journal_begin_ordered_truncate() since there's no
135 * outstanding writes we need to flush.
136 */
137 if (!EXT4_I(inode)->jinode)
138 return 0;
139 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
140 EXT4_I(inode)->jinode,
141 new_size);
142}
143
144static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
145 int pextents);
146
147/*
148 * Test whether an inode is a fast symlink.
149 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
150 */
151int ext4_inode_is_fast_symlink(struct inode *inode)
152{
153 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
154 int ea_blocks = EXT4_I(inode)->i_file_acl ?
155 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
156
157 if (ext4_has_inline_data(inode))
158 return 0;
159
160 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
161 }
162 return S_ISLNK(inode->i_mode) && inode->i_size &&
163 (inode->i_size < EXT4_N_BLOCKS * 4);
164}
165
166/*
167 * Called at the last iput() if i_nlink is zero.
168 */
169void ext4_evict_inode(struct inode *inode)
170{
171 handle_t *handle;
172 int err;
173 /*
174 * Credits for final inode cleanup and freeing:
175 * sb + inode (ext4_orphan_del()), block bitmap, group descriptor
176 * (xattr block freeing), bitmap, group descriptor (inode freeing)
177 */
178 int extra_credits = 6;
179 struct ext4_xattr_inode_array *ea_inode_array = NULL;
180 bool freeze_protected = false;
181
182 trace_ext4_evict_inode(inode);
183
184 if (EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)
185 ext4_evict_ea_inode(inode);
186 if (inode->i_nlink) {
187 truncate_inode_pages_final(&inode->i_data);
188
189 goto no_delete;
190 }
191
192 if (is_bad_inode(inode))
193 goto no_delete;
194 dquot_initialize(inode);
195
196 if (ext4_should_order_data(inode))
197 ext4_begin_ordered_truncate(inode, 0);
198 truncate_inode_pages_final(&inode->i_data);
199
200 /*
201 * For inodes with journalled data, transaction commit could have
202 * dirtied the inode. And for inodes with dioread_nolock, unwritten
203 * extents converting worker could merge extents and also have dirtied
204 * the inode. Flush worker is ignoring it because of I_FREEING flag but
205 * we still need to remove the inode from the writeback lists.
206 */
207 if (!list_empty_careful(&inode->i_io_list))
208 inode_io_list_del(inode);
209
210 /*
211 * Protect us against freezing - iput() caller didn't have to have any
212 * protection against it. When we are in a running transaction though,
213 * we are already protected against freezing and we cannot grab further
214 * protection due to lock ordering constraints.
215 */
216 if (!ext4_journal_current_handle()) {
217 sb_start_intwrite(inode->i_sb);
218 freeze_protected = true;
219 }
220
221 if (!IS_NOQUOTA(inode))
222 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
223
224 /*
225 * Block bitmap, group descriptor, and inode are accounted in both
226 * ext4_blocks_for_truncate() and extra_credits. So subtract 3.
227 */
228 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
229 ext4_blocks_for_truncate(inode) + extra_credits - 3);
230 if (IS_ERR(handle)) {
231 ext4_std_error(inode->i_sb, PTR_ERR(handle));
232 /*
233 * If we're going to skip the normal cleanup, we still need to
234 * make sure that the in-core orphan linked list is properly
235 * cleaned up.
236 */
237 ext4_orphan_del(NULL, inode);
238 if (freeze_protected)
239 sb_end_intwrite(inode->i_sb);
240 goto no_delete;
241 }
242
243 if (IS_SYNC(inode))
244 ext4_handle_sync(handle);
245
246 /*
247 * Set inode->i_size to 0 before calling ext4_truncate(). We need
248 * special handling of symlinks here because i_size is used to
249 * determine whether ext4_inode_info->i_data contains symlink data or
250 * block mappings. Setting i_size to 0 will remove its fast symlink
251 * status. Erase i_data so that it becomes a valid empty block map.
252 */
253 if (ext4_inode_is_fast_symlink(inode))
254 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
255 inode->i_size = 0;
256 err = ext4_mark_inode_dirty(handle, inode);
257 if (err) {
258 ext4_warning(inode->i_sb,
259 "couldn't mark inode dirty (err %d)", err);
260 goto stop_handle;
261 }
262 if (inode->i_blocks) {
263 err = ext4_truncate(inode);
264 if (err) {
265 ext4_error_err(inode->i_sb, -err,
266 "couldn't truncate inode %lu (err %d)",
267 inode->i_ino, err);
268 goto stop_handle;
269 }
270 }
271
272 /* Remove xattr references. */
273 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
274 extra_credits);
275 if (err) {
276 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
277stop_handle:
278 ext4_journal_stop(handle);
279 ext4_orphan_del(NULL, inode);
280 if (freeze_protected)
281 sb_end_intwrite(inode->i_sb);
282 ext4_xattr_inode_array_free(ea_inode_array);
283 goto no_delete;
284 }
285
286 /*
287 * Kill off the orphan record which ext4_truncate created.
288 * AKPM: I think this can be inside the above `if'.
289 * Note that ext4_orphan_del() has to be able to cope with the
290 * deletion of a non-existent orphan - this is because we don't
291 * know if ext4_truncate() actually created an orphan record.
292 * (Well, we could do this if we need to, but heck - it works)
293 */
294 ext4_orphan_del(handle, inode);
295 EXT4_I(inode)->i_dtime = (__u32)ktime_get_real_seconds();
296
297 /*
298 * One subtle ordering requirement: if anything has gone wrong
299 * (transaction abort, IO errors, whatever), then we can still
300 * do these next steps (the fs will already have been marked as
301 * having errors), but we can't free the inode if the mark_dirty
302 * fails.
303 */
304 if (ext4_mark_inode_dirty(handle, inode))
305 /* If that failed, just do the required in-core inode clear. */
306 ext4_clear_inode(inode);
307 else
308 ext4_free_inode(handle, inode);
309 ext4_journal_stop(handle);
310 if (freeze_protected)
311 sb_end_intwrite(inode->i_sb);
312 ext4_xattr_inode_array_free(ea_inode_array);
313 return;
314no_delete:
315 /*
316 * Check out some where else accidentally dirty the evicting inode,
317 * which may probably cause inode use-after-free issues later.
318 */
319 WARN_ON_ONCE(!list_empty_careful(&inode->i_io_list));
320
321 if (!list_empty(&EXT4_I(inode)->i_fc_list))
322 ext4_fc_mark_ineligible(inode->i_sb, EXT4_FC_REASON_NOMEM, NULL);
323 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
324}
325
326#ifdef CONFIG_QUOTA
327qsize_t *ext4_get_reserved_space(struct inode *inode)
328{
329 return &EXT4_I(inode)->i_reserved_quota;
330}
331#endif
332
333/*
334 * Called with i_data_sem down, which is important since we can call
335 * ext4_discard_preallocations() from here.
336 */
337void ext4_da_update_reserve_space(struct inode *inode,
338 int used, int quota_claim)
339{
340 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
341 struct ext4_inode_info *ei = EXT4_I(inode);
342
343 spin_lock(&ei->i_block_reservation_lock);
344 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
345 if (unlikely(used > ei->i_reserved_data_blocks)) {
346 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
347 "with only %d reserved data blocks",
348 __func__, inode->i_ino, used,
349 ei->i_reserved_data_blocks);
350 WARN_ON(1);
351 used = ei->i_reserved_data_blocks;
352 }
353
354 /* Update per-inode reservations */
355 ei->i_reserved_data_blocks -= used;
356 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
357
358 spin_unlock(&ei->i_block_reservation_lock);
359
360 /* Update quota subsystem for data blocks */
361 if (quota_claim)
362 dquot_claim_block(inode, EXT4_C2B(sbi, used));
363 else {
364 /*
365 * We did fallocate with an offset that is already delayed
366 * allocated. So on delayed allocated writeback we should
367 * not re-claim the quota for fallocated blocks.
368 */
369 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
370 }
371
372 /*
373 * If we have done all the pending block allocations and if
374 * there aren't any writers on the inode, we can discard the
375 * inode's preallocations.
376 */
377 if ((ei->i_reserved_data_blocks == 0) &&
378 !inode_is_open_for_write(inode))
379 ext4_discard_preallocations(inode);
380}
381
382static int __check_block_validity(struct inode *inode, const char *func,
383 unsigned int line,
384 struct ext4_map_blocks *map)
385{
386 if (ext4_has_feature_journal(inode->i_sb) &&
387 (inode->i_ino ==
388 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
389 return 0;
390 if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
391 ext4_error_inode(inode, func, line, map->m_pblk,
392 "lblock %lu mapped to illegal pblock %llu "
393 "(length %d)", (unsigned long) map->m_lblk,
394 map->m_pblk, map->m_len);
395 return -EFSCORRUPTED;
396 }
397 return 0;
398}
399
400int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
401 ext4_lblk_t len)
402{
403 int ret;
404
405 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
406 return fscrypt_zeroout_range(inode, lblk, pblk, len);
407
408 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
409 if (ret > 0)
410 ret = 0;
411
412 return ret;
413}
414
415#define check_block_validity(inode, map) \
416 __check_block_validity((inode), __func__, __LINE__, (map))
417
418#ifdef ES_AGGRESSIVE_TEST
419static void ext4_map_blocks_es_recheck(handle_t *handle,
420 struct inode *inode,
421 struct ext4_map_blocks *es_map,
422 struct ext4_map_blocks *map,
423 int flags)
424{
425 int retval;
426
427 map->m_flags = 0;
428 /*
429 * There is a race window that the result is not the same.
430 * e.g. xfstests #223 when dioread_nolock enables. The reason
431 * is that we lookup a block mapping in extent status tree with
432 * out taking i_data_sem. So at the time the unwritten extent
433 * could be converted.
434 */
435 down_read(&EXT4_I(inode)->i_data_sem);
436 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
437 retval = ext4_ext_map_blocks(handle, inode, map, 0);
438 } else {
439 retval = ext4_ind_map_blocks(handle, inode, map, 0);
440 }
441 up_read((&EXT4_I(inode)->i_data_sem));
442
443 /*
444 * We don't check m_len because extent will be collpased in status
445 * tree. So the m_len might not equal.
446 */
447 if (es_map->m_lblk != map->m_lblk ||
448 es_map->m_flags != map->m_flags ||
449 es_map->m_pblk != map->m_pblk) {
450 printk("ES cache assertion failed for inode: %lu "
451 "es_cached ex [%d/%d/%llu/%x] != "
452 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
453 inode->i_ino, es_map->m_lblk, es_map->m_len,
454 es_map->m_pblk, es_map->m_flags, map->m_lblk,
455 map->m_len, map->m_pblk, map->m_flags,
456 retval, flags);
457 }
458}
459#endif /* ES_AGGRESSIVE_TEST */
460
461static int ext4_map_query_blocks(handle_t *handle, struct inode *inode,
462 struct ext4_map_blocks *map)
463{
464 unsigned int status;
465 int retval;
466
467 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
468 retval = ext4_ext_map_blocks(handle, inode, map, 0);
469 else
470 retval = ext4_ind_map_blocks(handle, inode, map, 0);
471
472 if (retval <= 0)
473 return retval;
474
475 if (unlikely(retval != map->m_len)) {
476 ext4_warning(inode->i_sb,
477 "ES len assertion failed for inode "
478 "%lu: retval %d != map->m_len %d",
479 inode->i_ino, retval, map->m_len);
480 WARN_ON(1);
481 }
482
483 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
484 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
485 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
486 map->m_pblk, status, false);
487 return retval;
488}
489
490static int ext4_map_create_blocks(handle_t *handle, struct inode *inode,
491 struct ext4_map_blocks *map, int flags)
492{
493 struct extent_status es;
494 unsigned int status;
495 int err, retval = 0;
496
497 /*
498 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE
499 * indicates that the blocks and quotas has already been
500 * checked when the data was copied into the page cache.
501 */
502 if (map->m_flags & EXT4_MAP_DELAYED)
503 flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
504
505 /*
506 * Here we clear m_flags because after allocating an new extent,
507 * it will be set again.
508 */
509 map->m_flags &= ~EXT4_MAP_FLAGS;
510
511 /*
512 * We need to check for EXT4 here because migrate could have
513 * changed the inode type in between.
514 */
515 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
516 retval = ext4_ext_map_blocks(handle, inode, map, flags);
517 } else {
518 retval = ext4_ind_map_blocks(handle, inode, map, flags);
519
520 /*
521 * We allocated new blocks which will result in i_data's
522 * format changing. Force the migrate to fail by clearing
523 * migrate flags.
524 */
525 if (retval > 0 && map->m_flags & EXT4_MAP_NEW)
526 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
527 }
528 if (retval <= 0)
529 return retval;
530
531 if (unlikely(retval != map->m_len)) {
532 ext4_warning(inode->i_sb,
533 "ES len assertion failed for inode %lu: "
534 "retval %d != map->m_len %d",
535 inode->i_ino, retval, map->m_len);
536 WARN_ON(1);
537 }
538
539 /*
540 * We have to zeroout blocks before inserting them into extent
541 * status tree. Otherwise someone could look them up there and
542 * use them before they are really zeroed. We also have to
543 * unmap metadata before zeroing as otherwise writeback can
544 * overwrite zeros with stale data from block device.
545 */
546 if (flags & EXT4_GET_BLOCKS_ZERO &&
547 map->m_flags & EXT4_MAP_MAPPED && map->m_flags & EXT4_MAP_NEW) {
548 err = ext4_issue_zeroout(inode, map->m_lblk, map->m_pblk,
549 map->m_len);
550 if (err)
551 return err;
552 }
553
554 /*
555 * If the extent has been zeroed out, we don't need to update
556 * extent status tree.
557 */
558 if (flags & EXT4_GET_BLOCKS_PRE_IO &&
559 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
560 if (ext4_es_is_written(&es))
561 return retval;
562 }
563
564 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
565 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
566 ext4_es_insert_extent(inode, map->m_lblk, map->m_len, map->m_pblk,
567 status, flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE);
568
569 return retval;
570}
571
572/*
573 * The ext4_map_blocks() function tries to look up the requested blocks,
574 * and returns if the blocks are already mapped.
575 *
576 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
577 * and store the allocated blocks in the result buffer head and mark it
578 * mapped.
579 *
580 * If file type is extents based, it will call ext4_ext_map_blocks(),
581 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
582 * based files
583 *
584 * On success, it returns the number of blocks being mapped or allocated.
585 * If flags doesn't contain EXT4_GET_BLOCKS_CREATE the blocks are
586 * pre-allocated and unwritten, the resulting @map is marked as unwritten.
587 * If the flags contain EXT4_GET_BLOCKS_CREATE, it will mark @map as mapped.
588 *
589 * It returns 0 if plain look up failed (blocks have not been allocated), in
590 * that case, @map is returned as unmapped but we still do fill map->m_len to
591 * indicate the length of a hole starting at map->m_lblk.
592 *
593 * It returns the error in case of allocation failure.
594 */
595int ext4_map_blocks(handle_t *handle, struct inode *inode,
596 struct ext4_map_blocks *map, int flags)
597{
598 struct extent_status es;
599 int retval;
600 int ret = 0;
601#ifdef ES_AGGRESSIVE_TEST
602 struct ext4_map_blocks orig_map;
603
604 memcpy(&orig_map, map, sizeof(*map));
605#endif
606
607 map->m_flags = 0;
608 ext_debug(inode, "flag 0x%x, max_blocks %u, logical block %lu\n",
609 flags, map->m_len, (unsigned long) map->m_lblk);
610
611 /*
612 * ext4_map_blocks returns an int, and m_len is an unsigned int
613 */
614 if (unlikely(map->m_len > INT_MAX))
615 map->m_len = INT_MAX;
616
617 /* We can handle the block number less than EXT_MAX_BLOCKS */
618 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
619 return -EFSCORRUPTED;
620
621 /* Lookup extent status tree firstly */
622 if (!(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY) &&
623 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
624 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
625 map->m_pblk = ext4_es_pblock(&es) +
626 map->m_lblk - es.es_lblk;
627 map->m_flags |= ext4_es_is_written(&es) ?
628 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
629 retval = es.es_len - (map->m_lblk - es.es_lblk);
630 if (retval > map->m_len)
631 retval = map->m_len;
632 map->m_len = retval;
633 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
634 map->m_pblk = 0;
635 map->m_flags |= ext4_es_is_delayed(&es) ?
636 EXT4_MAP_DELAYED : 0;
637 retval = es.es_len - (map->m_lblk - es.es_lblk);
638 if (retval > map->m_len)
639 retval = map->m_len;
640 map->m_len = retval;
641 retval = 0;
642 } else {
643 BUG();
644 }
645
646 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
647 return retval;
648#ifdef ES_AGGRESSIVE_TEST
649 ext4_map_blocks_es_recheck(handle, inode, map,
650 &orig_map, flags);
651#endif
652 goto found;
653 }
654 /*
655 * In the query cache no-wait mode, nothing we can do more if we
656 * cannot find extent in the cache.
657 */
658 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
659 return 0;
660
661 /*
662 * Try to see if we can get the block without requesting a new
663 * file system block.
664 */
665 down_read(&EXT4_I(inode)->i_data_sem);
666 retval = ext4_map_query_blocks(handle, inode, map);
667 up_read((&EXT4_I(inode)->i_data_sem));
668
669found:
670 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
671 ret = check_block_validity(inode, map);
672 if (ret != 0)
673 return ret;
674 }
675
676 /* If it is only a block(s) look up */
677 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
678 return retval;
679
680 /*
681 * Returns if the blocks have already allocated
682 *
683 * Note that if blocks have been preallocated
684 * ext4_ext_map_blocks() returns with buffer head unmapped
685 */
686 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
687 /*
688 * If we need to convert extent to unwritten
689 * we continue and do the actual work in
690 * ext4_ext_map_blocks()
691 */
692 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
693 return retval;
694
695 /*
696 * New blocks allocate and/or writing to unwritten extent
697 * will possibly result in updating i_data, so we take
698 * the write lock of i_data_sem, and call get_block()
699 * with create == 1 flag.
700 */
701 down_write(&EXT4_I(inode)->i_data_sem);
702 retval = ext4_map_create_blocks(handle, inode, map, flags);
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 if (retval > 0 && (map->m_flags & EXT4_MAP_UNWRITTEN ||
734 map->m_flags & EXT4_MAP_MAPPED))
735 ext4_fc_track_range(handle, inode, map->m_lblk,
736 map->m_lblk + map->m_len - 1);
737 if (retval < 0)
738 ext_debug(inode, "failed with err %d\n", retval);
739 return retval;
740}
741
742/*
743 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
744 * we have to be careful as someone else may be manipulating b_state as well.
745 */
746static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
747{
748 unsigned long old_state;
749 unsigned long new_state;
750
751 flags &= EXT4_MAP_FLAGS;
752
753 /* Dummy buffer_head? Set non-atomically. */
754 if (!bh->b_page) {
755 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
756 return;
757 }
758 /*
759 * Someone else may be modifying b_state. Be careful! This is ugly but
760 * once we get rid of using bh as a container for mapping information
761 * to pass to / from get_block functions, this can go away.
762 */
763 old_state = READ_ONCE(bh->b_state);
764 do {
765 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
766 } while (unlikely(!try_cmpxchg(&bh->b_state, &old_state, new_state)));
767}
768
769static int _ext4_get_block(struct inode *inode, sector_t iblock,
770 struct buffer_head *bh, int flags)
771{
772 struct ext4_map_blocks map;
773 int ret = 0;
774
775 if (ext4_has_inline_data(inode))
776 return -ERANGE;
777
778 map.m_lblk = iblock;
779 map.m_len = bh->b_size >> inode->i_blkbits;
780
781 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
782 flags);
783 if (ret > 0) {
784 map_bh(bh, inode->i_sb, map.m_pblk);
785 ext4_update_bh_state(bh, map.m_flags);
786 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
787 ret = 0;
788 } else if (ret == 0) {
789 /* hole case, need to fill in bh->b_size */
790 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
791 }
792 return ret;
793}
794
795int ext4_get_block(struct inode *inode, sector_t iblock,
796 struct buffer_head *bh, int create)
797{
798 return _ext4_get_block(inode, iblock, bh,
799 create ? EXT4_GET_BLOCKS_CREATE : 0);
800}
801
802/*
803 * Get block function used when preparing for buffered write if we require
804 * creating an unwritten extent if blocks haven't been allocated. The extent
805 * will be converted to written after the IO is complete.
806 */
807int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
808 struct buffer_head *bh_result, int create)
809{
810 int ret = 0;
811
812 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
813 inode->i_ino, create);
814 ret = _ext4_get_block(inode, iblock, bh_result,
815 EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT);
816
817 /*
818 * If the buffer is marked unwritten, mark it as new to make sure it is
819 * zeroed out correctly in case of partial writes. Otherwise, there is
820 * a chance of stale data getting exposed.
821 */
822 if (ret == 0 && buffer_unwritten(bh_result))
823 set_buffer_new(bh_result);
824
825 return ret;
826}
827
828/* Maximum number of blocks we map for direct IO at once. */
829#define DIO_MAX_BLOCKS 4096
830
831/*
832 * `handle' can be NULL if create is zero
833 */
834struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
835 ext4_lblk_t block, int map_flags)
836{
837 struct ext4_map_blocks map;
838 struct buffer_head *bh;
839 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
840 bool nowait = map_flags & EXT4_GET_BLOCKS_CACHED_NOWAIT;
841 int err;
842
843 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
844 || handle != NULL || create == 0);
845 ASSERT(create == 0 || !nowait);
846
847 map.m_lblk = block;
848 map.m_len = 1;
849 err = ext4_map_blocks(handle, inode, &map, map_flags);
850
851 if (err == 0)
852 return create ? ERR_PTR(-ENOSPC) : NULL;
853 if (err < 0)
854 return ERR_PTR(err);
855
856 if (nowait)
857 return sb_find_get_block(inode->i_sb, map.m_pblk);
858
859 /*
860 * Since bh could introduce extra ref count such as referred by
861 * journal_head etc. Try to avoid using __GFP_MOVABLE here
862 * as it may fail the migration when journal_head remains.
863 */
864 bh = getblk_unmovable(inode->i_sb->s_bdev, map.m_pblk,
865 inode->i_sb->s_blocksize);
866
867 if (unlikely(!bh))
868 return ERR_PTR(-ENOMEM);
869 if (map.m_flags & EXT4_MAP_NEW) {
870 ASSERT(create != 0);
871 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
872 || (handle != NULL));
873
874 /*
875 * Now that we do not always journal data, we should
876 * keep in mind whether this should always journal the
877 * new buffer as metadata. For now, regular file
878 * writes use ext4_get_block instead, so it's not a
879 * problem.
880 */
881 lock_buffer(bh);
882 BUFFER_TRACE(bh, "call get_create_access");
883 err = ext4_journal_get_create_access(handle, inode->i_sb, bh,
884 EXT4_JTR_NONE);
885 if (unlikely(err)) {
886 unlock_buffer(bh);
887 goto errout;
888 }
889 if (!buffer_uptodate(bh)) {
890 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
891 set_buffer_uptodate(bh);
892 }
893 unlock_buffer(bh);
894 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
895 err = ext4_handle_dirty_metadata(handle, inode, bh);
896 if (unlikely(err))
897 goto errout;
898 } else
899 BUFFER_TRACE(bh, "not a new buffer");
900 return bh;
901errout:
902 brelse(bh);
903 return ERR_PTR(err);
904}
905
906struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
907 ext4_lblk_t block, int map_flags)
908{
909 struct buffer_head *bh;
910 int ret;
911
912 bh = ext4_getblk(handle, inode, block, map_flags);
913 if (IS_ERR(bh))
914 return bh;
915 if (!bh || ext4_buffer_uptodate(bh))
916 return bh;
917
918 ret = ext4_read_bh_lock(bh, REQ_META | REQ_PRIO, true);
919 if (ret) {
920 put_bh(bh);
921 return ERR_PTR(ret);
922 }
923 return bh;
924}
925
926/* Read a contiguous batch of blocks. */
927int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
928 bool wait, struct buffer_head **bhs)
929{
930 int i, err;
931
932 for (i = 0; i < bh_count; i++) {
933 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
934 if (IS_ERR(bhs[i])) {
935 err = PTR_ERR(bhs[i]);
936 bh_count = i;
937 goto out_brelse;
938 }
939 }
940
941 for (i = 0; i < bh_count; i++)
942 /* Note that NULL bhs[i] is valid because of holes. */
943 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
944 ext4_read_bh_lock(bhs[i], REQ_META | REQ_PRIO, false);
945
946 if (!wait)
947 return 0;
948
949 for (i = 0; i < bh_count; i++)
950 if (bhs[i])
951 wait_on_buffer(bhs[i]);
952
953 for (i = 0; i < bh_count; i++) {
954 if (bhs[i] && !buffer_uptodate(bhs[i])) {
955 err = -EIO;
956 goto out_brelse;
957 }
958 }
959 return 0;
960
961out_brelse:
962 for (i = 0; i < bh_count; i++) {
963 brelse(bhs[i]);
964 bhs[i] = NULL;
965 }
966 return err;
967}
968
969int ext4_walk_page_buffers(handle_t *handle, struct inode *inode,
970 struct buffer_head *head,
971 unsigned from,
972 unsigned to,
973 int *partial,
974 int (*fn)(handle_t *handle, struct inode *inode,
975 struct buffer_head *bh))
976{
977 struct buffer_head *bh;
978 unsigned block_start, block_end;
979 unsigned blocksize = head->b_size;
980 int err, ret = 0;
981 struct buffer_head *next;
982
983 for (bh = head, block_start = 0;
984 ret == 0 && (bh != head || !block_start);
985 block_start = block_end, bh = next) {
986 next = bh->b_this_page;
987 block_end = block_start + blocksize;
988 if (block_end <= from || block_start >= to) {
989 if (partial && !buffer_uptodate(bh))
990 *partial = 1;
991 continue;
992 }
993 err = (*fn)(handle, inode, bh);
994 if (!ret)
995 ret = err;
996 }
997 return ret;
998}
999
1000/*
1001 * Helper for handling dirtying of journalled data. We also mark the folio as
1002 * dirty so that writeback code knows about this page (and inode) contains
1003 * dirty data. ext4_writepages() then commits appropriate transaction to
1004 * make data stable.
1005 */
1006static int ext4_dirty_journalled_data(handle_t *handle, struct buffer_head *bh)
1007{
1008 folio_mark_dirty(bh->b_folio);
1009 return ext4_handle_dirty_metadata(handle, NULL, bh);
1010}
1011
1012int do_journal_get_write_access(handle_t *handle, struct inode *inode,
1013 struct buffer_head *bh)
1014{
1015 if (!buffer_mapped(bh) || buffer_freed(bh))
1016 return 0;
1017 BUFFER_TRACE(bh, "get write access");
1018 return ext4_journal_get_write_access(handle, inode->i_sb, bh,
1019 EXT4_JTR_NONE);
1020}
1021
1022int ext4_block_write_begin(handle_t *handle, struct folio *folio,
1023 loff_t pos, unsigned len,
1024 get_block_t *get_block)
1025{
1026 unsigned from = pos & (PAGE_SIZE - 1);
1027 unsigned to = from + len;
1028 struct inode *inode = folio->mapping->host;
1029 unsigned block_start, block_end;
1030 sector_t block;
1031 int err = 0;
1032 unsigned blocksize = inode->i_sb->s_blocksize;
1033 unsigned bbits;
1034 struct buffer_head *bh, *head, *wait[2];
1035 int nr_wait = 0;
1036 int i;
1037 bool should_journal_data = ext4_should_journal_data(inode);
1038
1039 BUG_ON(!folio_test_locked(folio));
1040 BUG_ON(from > PAGE_SIZE);
1041 BUG_ON(to > PAGE_SIZE);
1042 BUG_ON(from > to);
1043
1044 head = folio_buffers(folio);
1045 if (!head)
1046 head = create_empty_buffers(folio, blocksize, 0);
1047 bbits = ilog2(blocksize);
1048 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
1049
1050 for (bh = head, block_start = 0; bh != head || !block_start;
1051 block++, block_start = block_end, bh = bh->b_this_page) {
1052 block_end = block_start + blocksize;
1053 if (block_end <= from || block_start >= to) {
1054 if (folio_test_uptodate(folio)) {
1055 set_buffer_uptodate(bh);
1056 }
1057 continue;
1058 }
1059 if (buffer_new(bh))
1060 clear_buffer_new(bh);
1061 if (!buffer_mapped(bh)) {
1062 WARN_ON(bh->b_size != blocksize);
1063 err = get_block(inode, block, bh, 1);
1064 if (err)
1065 break;
1066 if (buffer_new(bh)) {
1067 /*
1068 * We may be zeroing partial buffers or all new
1069 * buffers in case of failure. Prepare JBD2 for
1070 * that.
1071 */
1072 if (should_journal_data)
1073 do_journal_get_write_access(handle,
1074 inode, bh);
1075 if (folio_test_uptodate(folio)) {
1076 /*
1077 * Unlike __block_write_begin() we leave
1078 * dirtying of new uptodate buffers to
1079 * ->write_end() time or
1080 * folio_zero_new_buffers().
1081 */
1082 set_buffer_uptodate(bh);
1083 continue;
1084 }
1085 if (block_end > to || block_start < from)
1086 folio_zero_segments(folio, to,
1087 block_end,
1088 block_start, from);
1089 continue;
1090 }
1091 }
1092 if (folio_test_uptodate(folio)) {
1093 set_buffer_uptodate(bh);
1094 continue;
1095 }
1096 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1097 !buffer_unwritten(bh) &&
1098 (block_start < from || block_end > to)) {
1099 ext4_read_bh_lock(bh, 0, false);
1100 wait[nr_wait++] = bh;
1101 }
1102 }
1103 /*
1104 * If we issued read requests, let them complete.
1105 */
1106 for (i = 0; i < nr_wait; i++) {
1107 wait_on_buffer(wait[i]);
1108 if (!buffer_uptodate(wait[i]))
1109 err = -EIO;
1110 }
1111 if (unlikely(err)) {
1112 if (should_journal_data)
1113 ext4_journalled_zero_new_buffers(handle, inode, folio,
1114 from, to);
1115 else
1116 folio_zero_new_buffers(folio, from, to);
1117 } else if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
1118 for (i = 0; i < nr_wait; i++) {
1119 int err2;
1120
1121 err2 = fscrypt_decrypt_pagecache_blocks(folio,
1122 blocksize, bh_offset(wait[i]));
1123 if (err2) {
1124 clear_buffer_uptodate(wait[i]);
1125 err = err2;
1126 }
1127 }
1128 }
1129
1130 return err;
1131}
1132
1133/*
1134 * To preserve ordering, it is essential that the hole instantiation and
1135 * the data write be encapsulated in a single transaction. We cannot
1136 * close off a transaction and start a new one between the ext4_get_block()
1137 * and the ext4_write_end(). So doing the jbd2_journal_start at the start of
1138 * ext4_write_begin() is the right place.
1139 */
1140static int ext4_write_begin(struct file *file, struct address_space *mapping,
1141 loff_t pos, unsigned len,
1142 struct folio **foliop, void **fsdata)
1143{
1144 struct inode *inode = mapping->host;
1145 int ret, needed_blocks;
1146 handle_t *handle;
1147 int retries = 0;
1148 struct folio *folio;
1149 pgoff_t index;
1150 unsigned from, to;
1151
1152 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
1153 return -EIO;
1154
1155 trace_ext4_write_begin(inode, pos, len);
1156 /*
1157 * Reserve one block more for addition to orphan list in case
1158 * we allocate blocks but write fails for some reason
1159 */
1160 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1161 index = pos >> PAGE_SHIFT;
1162 from = pos & (PAGE_SIZE - 1);
1163 to = from + len;
1164
1165 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1166 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1167 foliop);
1168 if (ret < 0)
1169 return ret;
1170 if (ret == 1)
1171 return 0;
1172 }
1173
1174 /*
1175 * __filemap_get_folio() can take a long time if the
1176 * system is thrashing due to memory pressure, or if the folio
1177 * is being written back. So grab it first before we start
1178 * the transaction handle. This also allows us to allocate
1179 * the folio (if needed) without using GFP_NOFS.
1180 */
1181retry_grab:
1182 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
1183 mapping_gfp_mask(mapping));
1184 if (IS_ERR(folio))
1185 return PTR_ERR(folio);
1186 /*
1187 * The same as page allocation, we prealloc buffer heads before
1188 * starting the handle.
1189 */
1190 if (!folio_buffers(folio))
1191 create_empty_buffers(folio, inode->i_sb->s_blocksize, 0);
1192
1193 folio_unlock(folio);
1194
1195retry_journal:
1196 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1197 if (IS_ERR(handle)) {
1198 folio_put(folio);
1199 return PTR_ERR(handle);
1200 }
1201
1202 folio_lock(folio);
1203 if (folio->mapping != mapping) {
1204 /* The folio got truncated from under us */
1205 folio_unlock(folio);
1206 folio_put(folio);
1207 ext4_journal_stop(handle);
1208 goto retry_grab;
1209 }
1210 /* In case writeback began while the folio was unlocked */
1211 folio_wait_stable(folio);
1212
1213 if (ext4_should_dioread_nolock(inode))
1214 ret = ext4_block_write_begin(handle, folio, pos, len,
1215 ext4_get_block_unwritten);
1216 else
1217 ret = ext4_block_write_begin(handle, folio, pos, len,
1218 ext4_get_block);
1219 if (!ret && ext4_should_journal_data(inode)) {
1220 ret = ext4_walk_page_buffers(handle, inode,
1221 folio_buffers(folio), from, to,
1222 NULL, do_journal_get_write_access);
1223 }
1224
1225 if (ret) {
1226 bool extended = (pos + len > inode->i_size) &&
1227 !ext4_verity_in_progress(inode);
1228
1229 folio_unlock(folio);
1230 /*
1231 * ext4_block_write_begin may have instantiated a few blocks
1232 * outside i_size. Trim these off again. Don't need
1233 * i_size_read because we hold i_rwsem.
1234 *
1235 * Add inode to orphan list in case we crash before
1236 * truncate finishes
1237 */
1238 if (extended && ext4_can_truncate(inode))
1239 ext4_orphan_add(handle, inode);
1240
1241 ext4_journal_stop(handle);
1242 if (extended) {
1243 ext4_truncate_failed_write(inode);
1244 /*
1245 * If truncate failed early the inode might
1246 * still be on the orphan list; we need to
1247 * make sure the inode is removed from the
1248 * orphan list in that case.
1249 */
1250 if (inode->i_nlink)
1251 ext4_orphan_del(NULL, inode);
1252 }
1253
1254 if (ret == -ENOSPC &&
1255 ext4_should_retry_alloc(inode->i_sb, &retries))
1256 goto retry_journal;
1257 folio_put(folio);
1258 return ret;
1259 }
1260 *foliop = folio;
1261 return ret;
1262}
1263
1264/* For write_end() in data=journal mode */
1265static int write_end_fn(handle_t *handle, struct inode *inode,
1266 struct buffer_head *bh)
1267{
1268 int ret;
1269 if (!buffer_mapped(bh) || buffer_freed(bh))
1270 return 0;
1271 set_buffer_uptodate(bh);
1272 ret = ext4_dirty_journalled_data(handle, bh);
1273 clear_buffer_meta(bh);
1274 clear_buffer_prio(bh);
1275 return ret;
1276}
1277
1278/*
1279 * We need to pick up the new inode size which generic_commit_write gave us
1280 * `file' can be NULL - eg, when called from page_symlink().
1281 *
1282 * ext4 never places buffers on inode->i_mapping->i_private_list. metadata
1283 * buffers are managed internally.
1284 */
1285static int ext4_write_end(struct file *file,
1286 struct address_space *mapping,
1287 loff_t pos, unsigned len, unsigned copied,
1288 struct folio *folio, void *fsdata)
1289{
1290 handle_t *handle = ext4_journal_current_handle();
1291 struct inode *inode = mapping->host;
1292 loff_t old_size = inode->i_size;
1293 int ret = 0, ret2;
1294 int i_size_changed = 0;
1295 bool verity = ext4_verity_in_progress(inode);
1296
1297 trace_ext4_write_end(inode, pos, len, copied);
1298
1299 if (ext4_has_inline_data(inode) &&
1300 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
1301 return ext4_write_inline_data_end(inode, pos, len, copied,
1302 folio);
1303
1304 copied = block_write_end(file, mapping, pos, len, copied, folio, fsdata);
1305 /*
1306 * it's important to update i_size while still holding folio lock:
1307 * page writeout could otherwise come in and zero beyond i_size.
1308 *
1309 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1310 * blocks are being written past EOF, so skip the i_size update.
1311 */
1312 if (!verity)
1313 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1314 folio_unlock(folio);
1315 folio_put(folio);
1316
1317 if (old_size < pos && !verity) {
1318 pagecache_isize_extended(inode, old_size, pos);
1319 ext4_zero_partial_blocks(handle, inode, old_size, pos - old_size);
1320 }
1321 /*
1322 * Don't mark the inode dirty under folio lock. First, it unnecessarily
1323 * makes the holding time of folio lock longer. Second, it forces lock
1324 * ordering of folio lock and transaction start for journaling
1325 * filesystems.
1326 */
1327 if (i_size_changed)
1328 ret = ext4_mark_inode_dirty(handle, inode);
1329
1330 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1331 /* if we have allocated more blocks and copied
1332 * less. We will have blocks allocated outside
1333 * inode->i_size. So truncate them
1334 */
1335 ext4_orphan_add(handle, inode);
1336
1337 ret2 = ext4_journal_stop(handle);
1338 if (!ret)
1339 ret = ret2;
1340
1341 if (pos + len > inode->i_size && !verity) {
1342 ext4_truncate_failed_write(inode);
1343 /*
1344 * If truncate failed early the inode might still be
1345 * on the orphan list; we need to make sure the inode
1346 * is removed from the orphan list in that case.
1347 */
1348 if (inode->i_nlink)
1349 ext4_orphan_del(NULL, inode);
1350 }
1351
1352 return ret ? ret : copied;
1353}
1354
1355/*
1356 * This is a private version of folio_zero_new_buffers() which doesn't
1357 * set the buffer to be dirty, since in data=journalled mode we need
1358 * to call ext4_dirty_journalled_data() instead.
1359 */
1360static void ext4_journalled_zero_new_buffers(handle_t *handle,
1361 struct inode *inode,
1362 struct folio *folio,
1363 unsigned from, unsigned to)
1364{
1365 unsigned int block_start = 0, block_end;
1366 struct buffer_head *head, *bh;
1367
1368 bh = head = folio_buffers(folio);
1369 do {
1370 block_end = block_start + bh->b_size;
1371 if (buffer_new(bh)) {
1372 if (block_end > from && block_start < to) {
1373 if (!folio_test_uptodate(folio)) {
1374 unsigned start, size;
1375
1376 start = max(from, block_start);
1377 size = min(to, block_end) - start;
1378
1379 folio_zero_range(folio, start, size);
1380 }
1381 clear_buffer_new(bh);
1382 write_end_fn(handle, inode, bh);
1383 }
1384 }
1385 block_start = block_end;
1386 bh = bh->b_this_page;
1387 } while (bh != head);
1388}
1389
1390static int ext4_journalled_write_end(struct file *file,
1391 struct address_space *mapping,
1392 loff_t pos, unsigned len, unsigned copied,
1393 struct folio *folio, void *fsdata)
1394{
1395 handle_t *handle = ext4_journal_current_handle();
1396 struct inode *inode = mapping->host;
1397 loff_t old_size = inode->i_size;
1398 int ret = 0, ret2;
1399 int partial = 0;
1400 unsigned from, to;
1401 int size_changed = 0;
1402 bool verity = ext4_verity_in_progress(inode);
1403
1404 trace_ext4_journalled_write_end(inode, pos, len, copied);
1405 from = pos & (PAGE_SIZE - 1);
1406 to = from + len;
1407
1408 BUG_ON(!ext4_handle_valid(handle));
1409
1410 if (ext4_has_inline_data(inode))
1411 return ext4_write_inline_data_end(inode, pos, len, copied,
1412 folio);
1413
1414 if (unlikely(copied < len) && !folio_test_uptodate(folio)) {
1415 copied = 0;
1416 ext4_journalled_zero_new_buffers(handle, inode, folio,
1417 from, to);
1418 } else {
1419 if (unlikely(copied < len))
1420 ext4_journalled_zero_new_buffers(handle, inode, folio,
1421 from + copied, to);
1422 ret = ext4_walk_page_buffers(handle, inode,
1423 folio_buffers(folio),
1424 from, from + copied, &partial,
1425 write_end_fn);
1426 if (!partial)
1427 folio_mark_uptodate(folio);
1428 }
1429 if (!verity)
1430 size_changed = ext4_update_inode_size(inode, pos + copied);
1431 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1432 folio_unlock(folio);
1433 folio_put(folio);
1434
1435 if (old_size < pos && !verity) {
1436 pagecache_isize_extended(inode, old_size, pos);
1437 ext4_zero_partial_blocks(handle, inode, old_size, pos - old_size);
1438 }
1439
1440 if (size_changed) {
1441 ret2 = ext4_mark_inode_dirty(handle, inode);
1442 if (!ret)
1443 ret = ret2;
1444 }
1445
1446 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1447 /* if we have allocated more blocks and copied
1448 * less. We will have blocks allocated outside
1449 * inode->i_size. So truncate them
1450 */
1451 ext4_orphan_add(handle, inode);
1452
1453 ret2 = ext4_journal_stop(handle);
1454 if (!ret)
1455 ret = ret2;
1456 if (pos + len > inode->i_size && !verity) {
1457 ext4_truncate_failed_write(inode);
1458 /*
1459 * If truncate failed early the inode might still be
1460 * on the orphan list; we need to make sure the inode
1461 * is removed from the orphan list in that case.
1462 */
1463 if (inode->i_nlink)
1464 ext4_orphan_del(NULL, inode);
1465 }
1466
1467 return ret ? ret : copied;
1468}
1469
1470/*
1471 * Reserve space for 'nr_resv' clusters
1472 */
1473static int ext4_da_reserve_space(struct inode *inode, int nr_resv)
1474{
1475 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1476 struct ext4_inode_info *ei = EXT4_I(inode);
1477 int ret;
1478
1479 /*
1480 * We will charge metadata quota at writeout time; this saves
1481 * us from metadata over-estimation, though we may go over by
1482 * a small amount in the end. Here we just reserve for data.
1483 */
1484 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, nr_resv));
1485 if (ret)
1486 return ret;
1487
1488 spin_lock(&ei->i_block_reservation_lock);
1489 if (ext4_claim_free_clusters(sbi, nr_resv, 0)) {
1490 spin_unlock(&ei->i_block_reservation_lock);
1491 dquot_release_reservation_block(inode, EXT4_C2B(sbi, nr_resv));
1492 return -ENOSPC;
1493 }
1494 ei->i_reserved_data_blocks += nr_resv;
1495 trace_ext4_da_reserve_space(inode, nr_resv);
1496 spin_unlock(&ei->i_block_reservation_lock);
1497
1498 return 0; /* success */
1499}
1500
1501void ext4_da_release_space(struct inode *inode, int to_free)
1502{
1503 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1504 struct ext4_inode_info *ei = EXT4_I(inode);
1505
1506 if (!to_free)
1507 return; /* Nothing to release, exit */
1508
1509 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1510
1511 trace_ext4_da_release_space(inode, to_free);
1512 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1513 /*
1514 * if there aren't enough reserved blocks, then the
1515 * counter is messed up somewhere. Since this
1516 * function is called from invalidate page, it's
1517 * harmless to return without any action.
1518 */
1519 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1520 "ino %lu, to_free %d with only %d reserved "
1521 "data blocks", inode->i_ino, to_free,
1522 ei->i_reserved_data_blocks);
1523 WARN_ON(1);
1524 to_free = ei->i_reserved_data_blocks;
1525 }
1526 ei->i_reserved_data_blocks -= to_free;
1527
1528 /* update fs dirty data blocks counter */
1529 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1530
1531 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1532
1533 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1534}
1535
1536/*
1537 * Delayed allocation stuff
1538 */
1539
1540struct mpage_da_data {
1541 /* These are input fields for ext4_do_writepages() */
1542 struct inode *inode;
1543 struct writeback_control *wbc;
1544 unsigned int can_map:1; /* Can writepages call map blocks? */
1545
1546 /* These are internal state of ext4_do_writepages() */
1547 pgoff_t first_page; /* The first page to write */
1548 pgoff_t next_page; /* Current page to examine */
1549 pgoff_t last_page; /* Last page to examine */
1550 /*
1551 * Extent to map - this can be after first_page because that can be
1552 * fully mapped. We somewhat abuse m_flags to store whether the extent
1553 * is delalloc or unwritten.
1554 */
1555 struct ext4_map_blocks map;
1556 struct ext4_io_submit io_submit; /* IO submission data */
1557 unsigned int do_map:1;
1558 unsigned int scanned_until_end:1;
1559 unsigned int journalled_more_data:1;
1560};
1561
1562static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1563 bool invalidate)
1564{
1565 unsigned nr, i;
1566 pgoff_t index, end;
1567 struct folio_batch fbatch;
1568 struct inode *inode = mpd->inode;
1569 struct address_space *mapping = inode->i_mapping;
1570
1571 /* This is necessary when next_page == 0. */
1572 if (mpd->first_page >= mpd->next_page)
1573 return;
1574
1575 mpd->scanned_until_end = 0;
1576 index = mpd->first_page;
1577 end = mpd->next_page - 1;
1578 if (invalidate) {
1579 ext4_lblk_t start, last;
1580 start = index << (PAGE_SHIFT - inode->i_blkbits);
1581 last = end << (PAGE_SHIFT - inode->i_blkbits);
1582
1583 /*
1584 * avoid racing with extent status tree scans made by
1585 * ext4_insert_delayed_block()
1586 */
1587 down_write(&EXT4_I(inode)->i_data_sem);
1588 ext4_es_remove_extent(inode, start, last - start + 1);
1589 up_write(&EXT4_I(inode)->i_data_sem);
1590 }
1591
1592 folio_batch_init(&fbatch);
1593 while (index <= end) {
1594 nr = filemap_get_folios(mapping, &index, end, &fbatch);
1595 if (nr == 0)
1596 break;
1597 for (i = 0; i < nr; i++) {
1598 struct folio *folio = fbatch.folios[i];
1599
1600 if (folio->index < mpd->first_page)
1601 continue;
1602 if (folio_next_index(folio) - 1 > end)
1603 continue;
1604 BUG_ON(!folio_test_locked(folio));
1605 BUG_ON(folio_test_writeback(folio));
1606 if (invalidate) {
1607 if (folio_mapped(folio))
1608 folio_clear_dirty_for_io(folio);
1609 block_invalidate_folio(folio, 0,
1610 folio_size(folio));
1611 folio_clear_uptodate(folio);
1612 }
1613 folio_unlock(folio);
1614 }
1615 folio_batch_release(&fbatch);
1616 }
1617}
1618
1619static void ext4_print_free_blocks(struct inode *inode)
1620{
1621 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1622 struct super_block *sb = inode->i_sb;
1623 struct ext4_inode_info *ei = EXT4_I(inode);
1624
1625 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1626 EXT4_C2B(EXT4_SB(inode->i_sb),
1627 ext4_count_free_clusters(sb)));
1628 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1629 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1630 (long long) EXT4_C2B(EXT4_SB(sb),
1631 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1632 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1633 (long long) EXT4_C2B(EXT4_SB(sb),
1634 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1635 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1636 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1637 ei->i_reserved_data_blocks);
1638 return;
1639}
1640
1641/*
1642 * Check whether the cluster containing lblk has been allocated or has
1643 * delalloc reservation.
1644 *
1645 * Returns 0 if the cluster doesn't have either, 1 if it has delalloc
1646 * reservation, 2 if it's already been allocated, negative error code on
1647 * failure.
1648 */
1649static int ext4_clu_alloc_state(struct inode *inode, ext4_lblk_t lblk)
1650{
1651 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1652 int ret;
1653
1654 /* Has delalloc reservation? */
1655 if (ext4_es_scan_clu(inode, &ext4_es_is_delayed, lblk))
1656 return 1;
1657
1658 /* Already been allocated? */
1659 if (ext4_es_scan_clu(inode, &ext4_es_is_mapped, lblk))
1660 return 2;
1661 ret = ext4_clu_mapped(inode, EXT4_B2C(sbi, lblk));
1662 if (ret < 0)
1663 return ret;
1664 if (ret > 0)
1665 return 2;
1666
1667 return 0;
1668}
1669
1670/*
1671 * ext4_insert_delayed_blocks - adds a multiple delayed blocks to the extents
1672 * status tree, incrementing the reserved
1673 * cluster/block count or making pending
1674 * reservations where needed
1675 *
1676 * @inode - file containing the newly added block
1677 * @lblk - start logical block to be added
1678 * @len - length of blocks to be added
1679 *
1680 * Returns 0 on success, negative error code on failure.
1681 */
1682static int ext4_insert_delayed_blocks(struct inode *inode, ext4_lblk_t lblk,
1683 ext4_lblk_t len)
1684{
1685 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1686 int ret;
1687 bool lclu_allocated = false;
1688 bool end_allocated = false;
1689 ext4_lblk_t resv_clu;
1690 ext4_lblk_t end = lblk + len - 1;
1691
1692 /*
1693 * If the cluster containing lblk or end is shared with a delayed,
1694 * written, or unwritten extent in a bigalloc file system, it's
1695 * already been accounted for and does not need to be reserved.
1696 * A pending reservation must be made for the cluster if it's
1697 * shared with a written or unwritten extent and doesn't already
1698 * have one. Written and unwritten extents can be purged from the
1699 * extents status tree if the system is under memory pressure, so
1700 * it's necessary to examine the extent tree if a search of the
1701 * extents status tree doesn't get a match.
1702 */
1703 if (sbi->s_cluster_ratio == 1) {
1704 ret = ext4_da_reserve_space(inode, len);
1705 if (ret != 0) /* ENOSPC */
1706 return ret;
1707 } else { /* bigalloc */
1708 resv_clu = EXT4_B2C(sbi, end) - EXT4_B2C(sbi, lblk) + 1;
1709
1710 ret = ext4_clu_alloc_state(inode, lblk);
1711 if (ret < 0)
1712 return ret;
1713 if (ret > 0) {
1714 resv_clu--;
1715 lclu_allocated = (ret == 2);
1716 }
1717
1718 if (EXT4_B2C(sbi, lblk) != EXT4_B2C(sbi, end)) {
1719 ret = ext4_clu_alloc_state(inode, end);
1720 if (ret < 0)
1721 return ret;
1722 if (ret > 0) {
1723 resv_clu--;
1724 end_allocated = (ret == 2);
1725 }
1726 }
1727
1728 if (resv_clu) {
1729 ret = ext4_da_reserve_space(inode, resv_clu);
1730 if (ret != 0) /* ENOSPC */
1731 return ret;
1732 }
1733 }
1734
1735 ext4_es_insert_delayed_extent(inode, lblk, len, lclu_allocated,
1736 end_allocated);
1737 return 0;
1738}
1739
1740/*
1741 * Looks up the requested blocks and sets the delalloc extent map.
1742 * First try to look up for the extent entry that contains the requested
1743 * blocks in the extent status tree without i_data_sem, then try to look
1744 * up for the ondisk extent mapping with i_data_sem in read mode,
1745 * finally hold i_data_sem in write mode, looks up again and add a
1746 * delalloc extent entry if it still couldn't find any extent. Pass out
1747 * the mapped extent through @map and return 0 on success.
1748 */
1749static int ext4_da_map_blocks(struct inode *inode, struct ext4_map_blocks *map)
1750{
1751 struct extent_status es;
1752 int retval;
1753#ifdef ES_AGGRESSIVE_TEST
1754 struct ext4_map_blocks orig_map;
1755
1756 memcpy(&orig_map, map, sizeof(*map));
1757#endif
1758
1759 map->m_flags = 0;
1760 ext_debug(inode, "max_blocks %u, logical block %lu\n", map->m_len,
1761 (unsigned long) map->m_lblk);
1762
1763 /* Lookup extent status tree firstly */
1764 if (ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
1765 map->m_len = min_t(unsigned int, map->m_len,
1766 es.es_len - (map->m_lblk - es.es_lblk));
1767
1768 if (ext4_es_is_hole(&es))
1769 goto add_delayed;
1770
1771found:
1772 /*
1773 * Delayed extent could be allocated by fallocate.
1774 * So we need to check it.
1775 */
1776 if (ext4_es_is_delayed(&es)) {
1777 map->m_flags |= EXT4_MAP_DELAYED;
1778 return 0;
1779 }
1780
1781 map->m_pblk = ext4_es_pblock(&es) + map->m_lblk - es.es_lblk;
1782 if (ext4_es_is_written(&es))
1783 map->m_flags |= EXT4_MAP_MAPPED;
1784 else if (ext4_es_is_unwritten(&es))
1785 map->m_flags |= EXT4_MAP_UNWRITTEN;
1786 else
1787 BUG();
1788
1789#ifdef ES_AGGRESSIVE_TEST
1790 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1791#endif
1792 return 0;
1793 }
1794
1795 /*
1796 * Try to see if we can get the block without requesting a new
1797 * file system block.
1798 */
1799 down_read(&EXT4_I(inode)->i_data_sem);
1800 if (ext4_has_inline_data(inode))
1801 retval = 0;
1802 else
1803 retval = ext4_map_query_blocks(NULL, inode, map);
1804 up_read(&EXT4_I(inode)->i_data_sem);
1805 if (retval)
1806 return retval < 0 ? retval : 0;
1807
1808add_delayed:
1809 down_write(&EXT4_I(inode)->i_data_sem);
1810 /*
1811 * Page fault path (ext4_page_mkwrite does not take i_rwsem)
1812 * and fallocate path (no folio lock) can race. Make sure we
1813 * lookup the extent status tree here again while i_data_sem
1814 * is held in write mode, before inserting a new da entry in
1815 * the extent status tree.
1816 */
1817 if (ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
1818 map->m_len = min_t(unsigned int, map->m_len,
1819 es.es_len - (map->m_lblk - es.es_lblk));
1820
1821 if (!ext4_es_is_hole(&es)) {
1822 up_write(&EXT4_I(inode)->i_data_sem);
1823 goto found;
1824 }
1825 } else if (!ext4_has_inline_data(inode)) {
1826 retval = ext4_map_query_blocks(NULL, inode, map);
1827 if (retval) {
1828 up_write(&EXT4_I(inode)->i_data_sem);
1829 return retval < 0 ? retval : 0;
1830 }
1831 }
1832
1833 map->m_flags |= EXT4_MAP_DELAYED;
1834 retval = ext4_insert_delayed_blocks(inode, map->m_lblk, map->m_len);
1835 up_write(&EXT4_I(inode)->i_data_sem);
1836
1837 return retval;
1838}
1839
1840/*
1841 * This is a special get_block_t callback which is used by
1842 * ext4_da_write_begin(). It will either return mapped block or
1843 * reserve space for a single block.
1844 *
1845 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1846 * We also have b_blocknr = -1 and b_bdev initialized properly
1847 *
1848 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1849 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1850 * initialized properly.
1851 */
1852int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1853 struct buffer_head *bh, int create)
1854{
1855 struct ext4_map_blocks map;
1856 sector_t invalid_block = ~((sector_t) 0xffff);
1857 int ret = 0;
1858
1859 BUG_ON(create == 0);
1860 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1861
1862 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1863 invalid_block = ~0;
1864
1865 map.m_lblk = iblock;
1866 map.m_len = 1;
1867
1868 /*
1869 * first, we need to know whether the block is allocated already
1870 * preallocated blocks are unmapped but should treated
1871 * the same as allocated blocks.
1872 */
1873 ret = ext4_da_map_blocks(inode, &map);
1874 if (ret < 0)
1875 return ret;
1876
1877 if (map.m_flags & EXT4_MAP_DELAYED) {
1878 map_bh(bh, inode->i_sb, invalid_block);
1879 set_buffer_new(bh);
1880 set_buffer_delay(bh);
1881 return 0;
1882 }
1883
1884 map_bh(bh, inode->i_sb, map.m_pblk);
1885 ext4_update_bh_state(bh, map.m_flags);
1886
1887 if (buffer_unwritten(bh)) {
1888 /* A delayed write to unwritten bh should be marked
1889 * new and mapped. Mapped ensures that we don't do
1890 * get_block multiple times when we write to the same
1891 * offset and new ensures that we do proper zero out
1892 * for partial write.
1893 */
1894 set_buffer_new(bh);
1895 set_buffer_mapped(bh);
1896 }
1897 return 0;
1898}
1899
1900static void mpage_folio_done(struct mpage_da_data *mpd, struct folio *folio)
1901{
1902 mpd->first_page += folio_nr_pages(folio);
1903 folio_unlock(folio);
1904}
1905
1906static int mpage_submit_folio(struct mpage_da_data *mpd, struct folio *folio)
1907{
1908 size_t len;
1909 loff_t size;
1910 int err;
1911
1912 BUG_ON(folio->index != mpd->first_page);
1913 folio_clear_dirty_for_io(folio);
1914 /*
1915 * We have to be very careful here! Nothing protects writeback path
1916 * against i_size changes and the page can be writeably mapped into
1917 * page tables. So an application can be growing i_size and writing
1918 * data through mmap while writeback runs. folio_clear_dirty_for_io()
1919 * write-protects our page in page tables and the page cannot get
1920 * written to again until we release folio lock. So only after
1921 * folio_clear_dirty_for_io() we are safe to sample i_size for
1922 * ext4_bio_write_folio() to zero-out tail of the written page. We rely
1923 * on the barrier provided by folio_test_clear_dirty() in
1924 * folio_clear_dirty_for_io() to make sure i_size is really sampled only
1925 * after page tables are updated.
1926 */
1927 size = i_size_read(mpd->inode);
1928 len = folio_size(folio);
1929 if (folio_pos(folio) + len > size &&
1930 !ext4_verity_in_progress(mpd->inode))
1931 len = size & (len - 1);
1932 err = ext4_bio_write_folio(&mpd->io_submit, folio, len);
1933 if (!err)
1934 mpd->wbc->nr_to_write--;
1935
1936 return err;
1937}
1938
1939#define BH_FLAGS (BIT(BH_Unwritten) | BIT(BH_Delay))
1940
1941/*
1942 * mballoc gives us at most this number of blocks...
1943 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
1944 * The rest of mballoc seems to handle chunks up to full group size.
1945 */
1946#define MAX_WRITEPAGES_EXTENT_LEN 2048
1947
1948/*
1949 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
1950 *
1951 * @mpd - extent of blocks
1952 * @lblk - logical number of the block in the file
1953 * @bh - buffer head we want to add to the extent
1954 *
1955 * The function is used to collect contig. blocks in the same state. If the
1956 * buffer doesn't require mapping for writeback and we haven't started the
1957 * extent of buffers to map yet, the function returns 'true' immediately - the
1958 * caller can write the buffer right away. Otherwise the function returns true
1959 * if the block has been added to the extent, false if the block couldn't be
1960 * added.
1961 */
1962static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
1963 struct buffer_head *bh)
1964{
1965 struct ext4_map_blocks *map = &mpd->map;
1966
1967 /* Buffer that doesn't need mapping for writeback? */
1968 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
1969 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
1970 /* So far no extent to map => we write the buffer right away */
1971 if (map->m_len == 0)
1972 return true;
1973 return false;
1974 }
1975
1976 /* First block in the extent? */
1977 if (map->m_len == 0) {
1978 /* We cannot map unless handle is started... */
1979 if (!mpd->do_map)
1980 return false;
1981 map->m_lblk = lblk;
1982 map->m_len = 1;
1983 map->m_flags = bh->b_state & BH_FLAGS;
1984 return true;
1985 }
1986
1987 /* Don't go larger than mballoc is willing to allocate */
1988 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
1989 return false;
1990
1991 /* Can we merge the block to our big extent? */
1992 if (lblk == map->m_lblk + map->m_len &&
1993 (bh->b_state & BH_FLAGS) == map->m_flags) {
1994 map->m_len++;
1995 return true;
1996 }
1997 return false;
1998}
1999
2000/*
2001 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2002 *
2003 * @mpd - extent of blocks for mapping
2004 * @head - the first buffer in the page
2005 * @bh - buffer we should start processing from
2006 * @lblk - logical number of the block in the file corresponding to @bh
2007 *
2008 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2009 * the page for IO if all buffers in this page were mapped and there's no
2010 * accumulated extent of buffers to map or add buffers in the page to the
2011 * extent of buffers to map. The function returns 1 if the caller can continue
2012 * by processing the next page, 0 if it should stop adding buffers to the
2013 * extent to map because we cannot extend it anymore. It can also return value
2014 * < 0 in case of error during IO submission.
2015 */
2016static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2017 struct buffer_head *head,
2018 struct buffer_head *bh,
2019 ext4_lblk_t lblk)
2020{
2021 struct inode *inode = mpd->inode;
2022 int err;
2023 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2024 >> inode->i_blkbits;
2025
2026 if (ext4_verity_in_progress(inode))
2027 blocks = EXT_MAX_BLOCKS;
2028
2029 do {
2030 BUG_ON(buffer_locked(bh));
2031
2032 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2033 /* Found extent to map? */
2034 if (mpd->map.m_len)
2035 return 0;
2036 /* Buffer needs mapping and handle is not started? */
2037 if (!mpd->do_map)
2038 return 0;
2039 /* Everything mapped so far and we hit EOF */
2040 break;
2041 }
2042 } while (lblk++, (bh = bh->b_this_page) != head);
2043 /* So far everything mapped? Submit the page for IO. */
2044 if (mpd->map.m_len == 0) {
2045 err = mpage_submit_folio(mpd, head->b_folio);
2046 if (err < 0)
2047 return err;
2048 mpage_folio_done(mpd, head->b_folio);
2049 }
2050 if (lblk >= blocks) {
2051 mpd->scanned_until_end = 1;
2052 return 0;
2053 }
2054 return 1;
2055}
2056
2057/*
2058 * mpage_process_folio - update folio buffers corresponding to changed extent
2059 * and may submit fully mapped page for IO
2060 * @mpd: description of extent to map, on return next extent to map
2061 * @folio: Contains these buffers.
2062 * @m_lblk: logical block mapping.
2063 * @m_pblk: corresponding physical mapping.
2064 * @map_bh: determines on return whether this page requires any further
2065 * mapping or not.
2066 *
2067 * Scan given folio buffers corresponding to changed extent and update buffer
2068 * state according to new extent state.
2069 * We map delalloc buffers to their physical location, clear unwritten bits.
2070 * If the given folio is not fully mapped, we update @mpd to the next extent in
2071 * the given folio that needs mapping & return @map_bh as true.
2072 */
2073static int mpage_process_folio(struct mpage_da_data *mpd, struct folio *folio,
2074 ext4_lblk_t *m_lblk, ext4_fsblk_t *m_pblk,
2075 bool *map_bh)
2076{
2077 struct buffer_head *head, *bh;
2078 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2079 ext4_lblk_t lblk = *m_lblk;
2080 ext4_fsblk_t pblock = *m_pblk;
2081 int err = 0;
2082 int blkbits = mpd->inode->i_blkbits;
2083 ssize_t io_end_size = 0;
2084 struct ext4_io_end_vec *io_end_vec = ext4_last_io_end_vec(io_end);
2085
2086 bh = head = folio_buffers(folio);
2087 do {
2088 if (lblk < mpd->map.m_lblk)
2089 continue;
2090 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2091 /*
2092 * Buffer after end of mapped extent.
2093 * Find next buffer in the folio to map.
2094 */
2095 mpd->map.m_len = 0;
2096 mpd->map.m_flags = 0;
2097 io_end_vec->size += io_end_size;
2098
2099 err = mpage_process_page_bufs(mpd, head, bh, lblk);
2100 if (err > 0)
2101 err = 0;
2102 if (!err && mpd->map.m_len && mpd->map.m_lblk > lblk) {
2103 io_end_vec = ext4_alloc_io_end_vec(io_end);
2104 if (IS_ERR(io_end_vec)) {
2105 err = PTR_ERR(io_end_vec);
2106 goto out;
2107 }
2108 io_end_vec->offset = (loff_t)mpd->map.m_lblk << blkbits;
2109 }
2110 *map_bh = true;
2111 goto out;
2112 }
2113 if (buffer_delay(bh)) {
2114 clear_buffer_delay(bh);
2115 bh->b_blocknr = pblock++;
2116 }
2117 clear_buffer_unwritten(bh);
2118 io_end_size += (1 << blkbits);
2119 } while (lblk++, (bh = bh->b_this_page) != head);
2120
2121 io_end_vec->size += io_end_size;
2122 *map_bh = false;
2123out:
2124 *m_lblk = lblk;
2125 *m_pblk = pblock;
2126 return err;
2127}
2128
2129/*
2130 * mpage_map_buffers - update buffers corresponding to changed extent and
2131 * submit fully mapped pages for IO
2132 *
2133 * @mpd - description of extent to map, on return next extent to map
2134 *
2135 * Scan buffers corresponding to changed extent (we expect corresponding pages
2136 * to be already locked) and update buffer state according to new extent state.
2137 * We map delalloc buffers to their physical location, clear unwritten bits,
2138 * and mark buffers as uninit when we perform writes to unwritten extents
2139 * and do extent conversion after IO is finished. If the last page is not fully
2140 * mapped, we update @map to the next extent in the last page that needs
2141 * mapping. Otherwise we submit the page for IO.
2142 */
2143static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2144{
2145 struct folio_batch fbatch;
2146 unsigned nr, i;
2147 struct inode *inode = mpd->inode;
2148 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2149 pgoff_t start, end;
2150 ext4_lblk_t lblk;
2151 ext4_fsblk_t pblock;
2152 int err;
2153 bool map_bh = false;
2154
2155 start = mpd->map.m_lblk >> bpp_bits;
2156 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2157 lblk = start << bpp_bits;
2158 pblock = mpd->map.m_pblk;
2159
2160 folio_batch_init(&fbatch);
2161 while (start <= end) {
2162 nr = filemap_get_folios(inode->i_mapping, &start, end, &fbatch);
2163 if (nr == 0)
2164 break;
2165 for (i = 0; i < nr; i++) {
2166 struct folio *folio = fbatch.folios[i];
2167
2168 err = mpage_process_folio(mpd, folio, &lblk, &pblock,
2169 &map_bh);
2170 /*
2171 * If map_bh is true, means page may require further bh
2172 * mapping, or maybe the page was submitted for IO.
2173 * So we return to call further extent mapping.
2174 */
2175 if (err < 0 || map_bh)
2176 goto out;
2177 /* Page fully mapped - let IO run! */
2178 err = mpage_submit_folio(mpd, folio);
2179 if (err < 0)
2180 goto out;
2181 mpage_folio_done(mpd, folio);
2182 }
2183 folio_batch_release(&fbatch);
2184 }
2185 /* Extent fully mapped and matches with page boundary. We are done. */
2186 mpd->map.m_len = 0;
2187 mpd->map.m_flags = 0;
2188 return 0;
2189out:
2190 folio_batch_release(&fbatch);
2191 return err;
2192}
2193
2194static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2195{
2196 struct inode *inode = mpd->inode;
2197 struct ext4_map_blocks *map = &mpd->map;
2198 int get_blocks_flags;
2199 int err, dioread_nolock;
2200
2201 trace_ext4_da_write_pages_extent(inode, map);
2202 /*
2203 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2204 * to convert an unwritten extent to be initialized (in the case
2205 * where we have written into one or more preallocated blocks). It is
2206 * possible that we're going to need more metadata blocks than
2207 * previously reserved. However we must not fail because we're in
2208 * writeback and there is nothing we can do about it so it might result
2209 * in data loss. So use reserved blocks to allocate metadata if
2210 * possible.
2211 */
2212 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2213 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2214 EXT4_GET_BLOCKS_IO_SUBMIT;
2215 dioread_nolock = ext4_should_dioread_nolock(inode);
2216 if (dioread_nolock)
2217 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2218
2219 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2220 if (err < 0)
2221 return err;
2222 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2223 if (!mpd->io_submit.io_end->handle &&
2224 ext4_handle_valid(handle)) {
2225 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2226 handle->h_rsv_handle = NULL;
2227 }
2228 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2229 }
2230
2231 BUG_ON(map->m_len == 0);
2232 return 0;
2233}
2234
2235/*
2236 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2237 * mpd->len and submit pages underlying it for IO
2238 *
2239 * @handle - handle for journal operations
2240 * @mpd - extent to map
2241 * @give_up_on_write - we set this to true iff there is a fatal error and there
2242 * is no hope of writing the data. The caller should discard
2243 * dirty pages to avoid infinite loops.
2244 *
2245 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2246 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2247 * them to initialized or split the described range from larger unwritten
2248 * extent. Note that we need not map all the described range since allocation
2249 * can return less blocks or the range is covered by more unwritten extents. We
2250 * cannot map more because we are limited by reserved transaction credits. On
2251 * the other hand we always make sure that the last touched page is fully
2252 * mapped so that it can be written out (and thus forward progress is
2253 * guaranteed). After mapping we submit all mapped pages for IO.
2254 */
2255static int mpage_map_and_submit_extent(handle_t *handle,
2256 struct mpage_da_data *mpd,
2257 bool *give_up_on_write)
2258{
2259 struct inode *inode = mpd->inode;
2260 struct ext4_map_blocks *map = &mpd->map;
2261 int err;
2262 loff_t disksize;
2263 int progress = 0;
2264 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2265 struct ext4_io_end_vec *io_end_vec;
2266
2267 io_end_vec = ext4_alloc_io_end_vec(io_end);
2268 if (IS_ERR(io_end_vec))
2269 return PTR_ERR(io_end_vec);
2270 io_end_vec->offset = ((loff_t)map->m_lblk) << inode->i_blkbits;
2271 do {
2272 err = mpage_map_one_extent(handle, mpd);
2273 if (err < 0) {
2274 struct super_block *sb = inode->i_sb;
2275
2276 if (ext4_forced_shutdown(sb))
2277 goto invalidate_dirty_pages;
2278 /*
2279 * Let the uper layers retry transient errors.
2280 * In the case of ENOSPC, if ext4_count_free_blocks()
2281 * is non-zero, a commit should free up blocks.
2282 */
2283 if ((err == -ENOMEM) ||
2284 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2285 if (progress)
2286 goto update_disksize;
2287 return err;
2288 }
2289 ext4_msg(sb, KERN_CRIT,
2290 "Delayed block allocation failed for "
2291 "inode %lu at logical offset %llu with"
2292 " max blocks %u with error %d",
2293 inode->i_ino,
2294 (unsigned long long)map->m_lblk,
2295 (unsigned)map->m_len, -err);
2296 ext4_msg(sb, KERN_CRIT,
2297 "This should not happen!! Data will "
2298 "be lost\n");
2299 if (err == -ENOSPC)
2300 ext4_print_free_blocks(inode);
2301 invalidate_dirty_pages:
2302 *give_up_on_write = true;
2303 return err;
2304 }
2305 progress = 1;
2306 /*
2307 * Update buffer state, submit mapped pages, and get us new
2308 * extent to map
2309 */
2310 err = mpage_map_and_submit_buffers(mpd);
2311 if (err < 0)
2312 goto update_disksize;
2313 } while (map->m_len);
2314
2315update_disksize:
2316 /*
2317 * Update on-disk size after IO is submitted. Races with
2318 * truncate are avoided by checking i_size under i_data_sem.
2319 */
2320 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2321 if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
2322 int err2;
2323 loff_t i_size;
2324
2325 down_write(&EXT4_I(inode)->i_data_sem);
2326 i_size = i_size_read(inode);
2327 if (disksize > i_size)
2328 disksize = i_size;
2329 if (disksize > EXT4_I(inode)->i_disksize)
2330 EXT4_I(inode)->i_disksize = disksize;
2331 up_write(&EXT4_I(inode)->i_data_sem);
2332 err2 = ext4_mark_inode_dirty(handle, inode);
2333 if (err2) {
2334 ext4_error_err(inode->i_sb, -err2,
2335 "Failed to mark inode %lu dirty",
2336 inode->i_ino);
2337 }
2338 if (!err)
2339 err = err2;
2340 }
2341 return err;
2342}
2343
2344/*
2345 * Calculate the total number of credits to reserve for one writepages
2346 * iteration. This is called from ext4_writepages(). We map an extent of
2347 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2348 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2349 * bpp - 1 blocks in bpp different extents.
2350 */
2351static int ext4_da_writepages_trans_blocks(struct inode *inode)
2352{
2353 int bpp = ext4_journal_blocks_per_page(inode);
2354
2355 return ext4_meta_trans_blocks(inode,
2356 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2357}
2358
2359static int ext4_journal_folio_buffers(handle_t *handle, struct folio *folio,
2360 size_t len)
2361{
2362 struct buffer_head *page_bufs = folio_buffers(folio);
2363 struct inode *inode = folio->mapping->host;
2364 int ret, err;
2365
2366 ret = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2367 NULL, do_journal_get_write_access);
2368 err = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2369 NULL, write_end_fn);
2370 if (ret == 0)
2371 ret = err;
2372 err = ext4_jbd2_inode_add_write(handle, inode, folio_pos(folio), len);
2373 if (ret == 0)
2374 ret = err;
2375 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2376
2377 return ret;
2378}
2379
2380static int mpage_journal_page_buffers(handle_t *handle,
2381 struct mpage_da_data *mpd,
2382 struct folio *folio)
2383{
2384 struct inode *inode = mpd->inode;
2385 loff_t size = i_size_read(inode);
2386 size_t len = folio_size(folio);
2387
2388 folio_clear_checked(folio);
2389 mpd->wbc->nr_to_write--;
2390
2391 if (folio_pos(folio) + len > size &&
2392 !ext4_verity_in_progress(inode))
2393 len = size & (len - 1);
2394
2395 return ext4_journal_folio_buffers(handle, folio, len);
2396}
2397
2398/*
2399 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2400 * needing mapping, submit mapped pages
2401 *
2402 * @mpd - where to look for pages
2403 *
2404 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2405 * IO immediately. If we cannot map blocks, we submit just already mapped
2406 * buffers in the page for IO and keep page dirty. When we can map blocks and
2407 * we find a page which isn't mapped we start accumulating extent of buffers
2408 * underlying these pages that needs mapping (formed by either delayed or
2409 * unwritten buffers). We also lock the pages containing these buffers. The
2410 * extent found is returned in @mpd structure (starting at mpd->lblk with
2411 * length mpd->len blocks).
2412 *
2413 * Note that this function can attach bios to one io_end structure which are
2414 * neither logically nor physically contiguous. Although it may seem as an
2415 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2416 * case as we need to track IO to all buffers underlying a page in one io_end.
2417 */
2418static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2419{
2420 struct address_space *mapping = mpd->inode->i_mapping;
2421 struct folio_batch fbatch;
2422 unsigned int nr_folios;
2423 pgoff_t index = mpd->first_page;
2424 pgoff_t end = mpd->last_page;
2425 xa_mark_t tag;
2426 int i, err = 0;
2427 int blkbits = mpd->inode->i_blkbits;
2428 ext4_lblk_t lblk;
2429 struct buffer_head *head;
2430 handle_t *handle = NULL;
2431 int bpp = ext4_journal_blocks_per_page(mpd->inode);
2432
2433 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2434 tag = PAGECACHE_TAG_TOWRITE;
2435 else
2436 tag = PAGECACHE_TAG_DIRTY;
2437
2438 mpd->map.m_len = 0;
2439 mpd->next_page = index;
2440 if (ext4_should_journal_data(mpd->inode)) {
2441 handle = ext4_journal_start(mpd->inode, EXT4_HT_WRITE_PAGE,
2442 bpp);
2443 if (IS_ERR(handle))
2444 return PTR_ERR(handle);
2445 }
2446 folio_batch_init(&fbatch);
2447 while (index <= end) {
2448 nr_folios = filemap_get_folios_tag(mapping, &index, end,
2449 tag, &fbatch);
2450 if (nr_folios == 0)
2451 break;
2452
2453 for (i = 0; i < nr_folios; i++) {
2454 struct folio *folio = fbatch.folios[i];
2455
2456 /*
2457 * Accumulated enough dirty pages? This doesn't apply
2458 * to WB_SYNC_ALL mode. For integrity sync we have to
2459 * keep going because someone may be concurrently
2460 * dirtying pages, and we might have synced a lot of
2461 * newly appeared dirty pages, but have not synced all
2462 * of the old dirty pages.
2463 */
2464 if (mpd->wbc->sync_mode == WB_SYNC_NONE &&
2465 mpd->wbc->nr_to_write <=
2466 mpd->map.m_len >> (PAGE_SHIFT - blkbits))
2467 goto out;
2468
2469 /* If we can't merge this page, we are done. */
2470 if (mpd->map.m_len > 0 && mpd->next_page != folio->index)
2471 goto out;
2472
2473 if (handle) {
2474 err = ext4_journal_ensure_credits(handle, bpp,
2475 0);
2476 if (err < 0)
2477 goto out;
2478 }
2479
2480 folio_lock(folio);
2481 /*
2482 * If the page is no longer dirty, or its mapping no
2483 * longer corresponds to inode we are writing (which
2484 * means it has been truncated or invalidated), or the
2485 * page is already under writeback and we are not doing
2486 * a data integrity writeback, skip the page
2487 */
2488 if (!folio_test_dirty(folio) ||
2489 (folio_test_writeback(folio) &&
2490 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2491 unlikely(folio->mapping != mapping)) {
2492 folio_unlock(folio);
2493 continue;
2494 }
2495
2496 folio_wait_writeback(folio);
2497 BUG_ON(folio_test_writeback(folio));
2498
2499 /*
2500 * Should never happen but for buggy code in
2501 * other subsystems that call
2502 * set_page_dirty() without properly warning
2503 * the file system first. See [1] for more
2504 * information.
2505 *
2506 * [1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@quack2.suse.cz
2507 */
2508 if (!folio_buffers(folio)) {
2509 ext4_warning_inode(mpd->inode, "page %lu does not have buffers attached", folio->index);
2510 folio_clear_dirty(folio);
2511 folio_unlock(folio);
2512 continue;
2513 }
2514
2515 if (mpd->map.m_len == 0)
2516 mpd->first_page = folio->index;
2517 mpd->next_page = folio_next_index(folio);
2518 /*
2519 * Writeout when we cannot modify metadata is simple.
2520 * Just submit the page. For data=journal mode we
2521 * first handle writeout of the page for checkpoint and
2522 * only after that handle delayed page dirtying. This
2523 * makes sure current data is checkpointed to the final
2524 * location before possibly journalling it again which
2525 * is desirable when the page is frequently dirtied
2526 * through a pin.
2527 */
2528 if (!mpd->can_map) {
2529 err = mpage_submit_folio(mpd, folio);
2530 if (err < 0)
2531 goto out;
2532 /* Pending dirtying of journalled data? */
2533 if (folio_test_checked(folio)) {
2534 err = mpage_journal_page_buffers(handle,
2535 mpd, folio);
2536 if (err < 0)
2537 goto out;
2538 mpd->journalled_more_data = 1;
2539 }
2540 mpage_folio_done(mpd, folio);
2541 } else {
2542 /* Add all dirty buffers to mpd */
2543 lblk = ((ext4_lblk_t)folio->index) <<
2544 (PAGE_SHIFT - blkbits);
2545 head = folio_buffers(folio);
2546 err = mpage_process_page_bufs(mpd, head, head,
2547 lblk);
2548 if (err <= 0)
2549 goto out;
2550 err = 0;
2551 }
2552 }
2553 folio_batch_release(&fbatch);
2554 cond_resched();
2555 }
2556 mpd->scanned_until_end = 1;
2557 if (handle)
2558 ext4_journal_stop(handle);
2559 return 0;
2560out:
2561 folio_batch_release(&fbatch);
2562 if (handle)
2563 ext4_journal_stop(handle);
2564 return err;
2565}
2566
2567static int ext4_do_writepages(struct mpage_da_data *mpd)
2568{
2569 struct writeback_control *wbc = mpd->wbc;
2570 pgoff_t writeback_index = 0;
2571 long nr_to_write = wbc->nr_to_write;
2572 int range_whole = 0;
2573 int cycled = 1;
2574 handle_t *handle = NULL;
2575 struct inode *inode = mpd->inode;
2576 struct address_space *mapping = inode->i_mapping;
2577 int needed_blocks, rsv_blocks = 0, ret = 0;
2578 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2579 struct blk_plug plug;
2580 bool give_up_on_write = false;
2581
2582 trace_ext4_writepages(inode, wbc);
2583
2584 /*
2585 * No pages to write? This is mainly a kludge to avoid starting
2586 * a transaction for special inodes like journal inode on last iput()
2587 * because that could violate lock ordering on umount
2588 */
2589 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2590 goto out_writepages;
2591
2592 /*
2593 * If the filesystem has aborted, it is read-only, so return
2594 * right away instead of dumping stack traces later on that
2595 * will obscure the real source of the problem. We test
2596 * fs shutdown state instead of sb->s_flag's SB_RDONLY because
2597 * the latter could be true if the filesystem is mounted
2598 * read-only, and in that case, ext4_writepages should
2599 * *never* be called, so if that ever happens, we would want
2600 * the stack trace.
2601 */
2602 if (unlikely(ext4_forced_shutdown(mapping->host->i_sb))) {
2603 ret = -EROFS;
2604 goto out_writepages;
2605 }
2606
2607 /*
2608 * If we have inline data and arrive here, it means that
2609 * we will soon create the block for the 1st page, so
2610 * we'd better clear the inline data here.
2611 */
2612 if (ext4_has_inline_data(inode)) {
2613 /* Just inode will be modified... */
2614 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2615 if (IS_ERR(handle)) {
2616 ret = PTR_ERR(handle);
2617 goto out_writepages;
2618 }
2619 BUG_ON(ext4_test_inode_state(inode,
2620 EXT4_STATE_MAY_INLINE_DATA));
2621 ext4_destroy_inline_data(handle, inode);
2622 ext4_journal_stop(handle);
2623 }
2624
2625 /*
2626 * data=journal mode does not do delalloc so we just need to writeout /
2627 * journal already mapped buffers. On the other hand we need to commit
2628 * transaction to make data stable. We expect all the data to be
2629 * already in the journal (the only exception are DMA pinned pages
2630 * dirtied behind our back) so we commit transaction here and run the
2631 * writeback loop to checkpoint them. The checkpointing is not actually
2632 * necessary to make data persistent *but* quite a few places (extent
2633 * shifting operations, fsverity, ...) depend on being able to drop
2634 * pagecache pages after calling filemap_write_and_wait() and for that
2635 * checkpointing needs to happen.
2636 */
2637 if (ext4_should_journal_data(inode)) {
2638 mpd->can_map = 0;
2639 if (wbc->sync_mode == WB_SYNC_ALL)
2640 ext4_fc_commit(sbi->s_journal,
2641 EXT4_I(inode)->i_datasync_tid);
2642 }
2643 mpd->journalled_more_data = 0;
2644
2645 if (ext4_should_dioread_nolock(inode)) {
2646 /*
2647 * We may need to convert up to one extent per block in
2648 * the page and we may dirty the inode.
2649 */
2650 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2651 PAGE_SIZE >> inode->i_blkbits);
2652 }
2653
2654 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2655 range_whole = 1;
2656
2657 if (wbc->range_cyclic) {
2658 writeback_index = mapping->writeback_index;
2659 if (writeback_index)
2660 cycled = 0;
2661 mpd->first_page = writeback_index;
2662 mpd->last_page = -1;
2663 } else {
2664 mpd->first_page = wbc->range_start >> PAGE_SHIFT;
2665 mpd->last_page = wbc->range_end >> PAGE_SHIFT;
2666 }
2667
2668 ext4_io_submit_init(&mpd->io_submit, wbc);
2669retry:
2670 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2671 tag_pages_for_writeback(mapping, mpd->first_page,
2672 mpd->last_page);
2673 blk_start_plug(&plug);
2674
2675 /*
2676 * First writeback pages that don't need mapping - we can avoid
2677 * starting a transaction unnecessarily and also avoid being blocked
2678 * in the block layer on device congestion while having transaction
2679 * started.
2680 */
2681 mpd->do_map = 0;
2682 mpd->scanned_until_end = 0;
2683 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2684 if (!mpd->io_submit.io_end) {
2685 ret = -ENOMEM;
2686 goto unplug;
2687 }
2688 ret = mpage_prepare_extent_to_map(mpd);
2689 /* Unlock pages we didn't use */
2690 mpage_release_unused_pages(mpd, false);
2691 /* Submit prepared bio */
2692 ext4_io_submit(&mpd->io_submit);
2693 ext4_put_io_end_defer(mpd->io_submit.io_end);
2694 mpd->io_submit.io_end = NULL;
2695 if (ret < 0)
2696 goto unplug;
2697
2698 while (!mpd->scanned_until_end && wbc->nr_to_write > 0) {
2699 /* For each extent of pages we use new io_end */
2700 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2701 if (!mpd->io_submit.io_end) {
2702 ret = -ENOMEM;
2703 break;
2704 }
2705
2706 WARN_ON_ONCE(!mpd->can_map);
2707 /*
2708 * We have two constraints: We find one extent to map and we
2709 * must always write out whole page (makes a difference when
2710 * blocksize < pagesize) so that we don't block on IO when we
2711 * try to write out the rest of the page. Journalled mode is
2712 * not supported by delalloc.
2713 */
2714 BUG_ON(ext4_should_journal_data(inode));
2715 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2716
2717 /* start a new transaction */
2718 handle = ext4_journal_start_with_reserve(inode,
2719 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2720 if (IS_ERR(handle)) {
2721 ret = PTR_ERR(handle);
2722 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2723 "%ld pages, ino %lu; err %d", __func__,
2724 wbc->nr_to_write, inode->i_ino, ret);
2725 /* Release allocated io_end */
2726 ext4_put_io_end(mpd->io_submit.io_end);
2727 mpd->io_submit.io_end = NULL;
2728 break;
2729 }
2730 mpd->do_map = 1;
2731
2732 trace_ext4_da_write_pages(inode, mpd->first_page, wbc);
2733 ret = mpage_prepare_extent_to_map(mpd);
2734 if (!ret && mpd->map.m_len)
2735 ret = mpage_map_and_submit_extent(handle, mpd,
2736 &give_up_on_write);
2737 /*
2738 * Caution: If the handle is synchronous,
2739 * ext4_journal_stop() can wait for transaction commit
2740 * to finish which may depend on writeback of pages to
2741 * complete or on page lock to be released. In that
2742 * case, we have to wait until after we have
2743 * submitted all the IO, released page locks we hold,
2744 * and dropped io_end reference (for extent conversion
2745 * to be able to complete) before stopping the handle.
2746 */
2747 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2748 ext4_journal_stop(handle);
2749 handle = NULL;
2750 mpd->do_map = 0;
2751 }
2752 /* Unlock pages we didn't use */
2753 mpage_release_unused_pages(mpd, give_up_on_write);
2754 /* Submit prepared bio */
2755 ext4_io_submit(&mpd->io_submit);
2756
2757 /*
2758 * Drop our io_end reference we got from init. We have
2759 * to be careful and use deferred io_end finishing if
2760 * we are still holding the transaction as we can
2761 * release the last reference to io_end which may end
2762 * up doing unwritten extent conversion.
2763 */
2764 if (handle) {
2765 ext4_put_io_end_defer(mpd->io_submit.io_end);
2766 ext4_journal_stop(handle);
2767 } else
2768 ext4_put_io_end(mpd->io_submit.io_end);
2769 mpd->io_submit.io_end = NULL;
2770
2771 if (ret == -ENOSPC && sbi->s_journal) {
2772 /*
2773 * Commit the transaction which would
2774 * free blocks released in the transaction
2775 * and try again
2776 */
2777 jbd2_journal_force_commit_nested(sbi->s_journal);
2778 ret = 0;
2779 continue;
2780 }
2781 /* Fatal error - ENOMEM, EIO... */
2782 if (ret)
2783 break;
2784 }
2785unplug:
2786 blk_finish_plug(&plug);
2787 if (!ret && !cycled && wbc->nr_to_write > 0) {
2788 cycled = 1;
2789 mpd->last_page = writeback_index - 1;
2790 mpd->first_page = 0;
2791 goto retry;
2792 }
2793
2794 /* Update index */
2795 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2796 /*
2797 * Set the writeback_index so that range_cyclic
2798 * mode will write it back later
2799 */
2800 mapping->writeback_index = mpd->first_page;
2801
2802out_writepages:
2803 trace_ext4_writepages_result(inode, wbc, ret,
2804 nr_to_write - wbc->nr_to_write);
2805 return ret;
2806}
2807
2808static int ext4_writepages(struct address_space *mapping,
2809 struct writeback_control *wbc)
2810{
2811 struct super_block *sb = mapping->host->i_sb;
2812 struct mpage_da_data mpd = {
2813 .inode = mapping->host,
2814 .wbc = wbc,
2815 .can_map = 1,
2816 };
2817 int ret;
2818 int alloc_ctx;
2819
2820 if (unlikely(ext4_forced_shutdown(sb)))
2821 return -EIO;
2822
2823 alloc_ctx = ext4_writepages_down_read(sb);
2824 ret = ext4_do_writepages(&mpd);
2825 /*
2826 * For data=journal writeback we could have come across pages marked
2827 * for delayed dirtying (PageChecked) which were just added to the
2828 * running transaction. Try once more to get them to stable storage.
2829 */
2830 if (!ret && mpd.journalled_more_data)
2831 ret = ext4_do_writepages(&mpd);
2832 ext4_writepages_up_read(sb, alloc_ctx);
2833
2834 return ret;
2835}
2836
2837int ext4_normal_submit_inode_data_buffers(struct jbd2_inode *jinode)
2838{
2839 struct writeback_control wbc = {
2840 .sync_mode = WB_SYNC_ALL,
2841 .nr_to_write = LONG_MAX,
2842 .range_start = jinode->i_dirty_start,
2843 .range_end = jinode->i_dirty_end,
2844 };
2845 struct mpage_da_data mpd = {
2846 .inode = jinode->i_vfs_inode,
2847 .wbc = &wbc,
2848 .can_map = 0,
2849 };
2850 return ext4_do_writepages(&mpd);
2851}
2852
2853static int ext4_dax_writepages(struct address_space *mapping,
2854 struct writeback_control *wbc)
2855{
2856 int ret;
2857 long nr_to_write = wbc->nr_to_write;
2858 struct inode *inode = mapping->host;
2859 int alloc_ctx;
2860
2861 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2862 return -EIO;
2863
2864 alloc_ctx = ext4_writepages_down_read(inode->i_sb);
2865 trace_ext4_writepages(inode, wbc);
2866
2867 ret = dax_writeback_mapping_range(mapping,
2868 EXT4_SB(inode->i_sb)->s_daxdev, wbc);
2869 trace_ext4_writepages_result(inode, wbc, ret,
2870 nr_to_write - wbc->nr_to_write);
2871 ext4_writepages_up_read(inode->i_sb, alloc_ctx);
2872 return ret;
2873}
2874
2875static int ext4_nonda_switch(struct super_block *sb)
2876{
2877 s64 free_clusters, dirty_clusters;
2878 struct ext4_sb_info *sbi = EXT4_SB(sb);
2879
2880 /*
2881 * switch to non delalloc mode if we are running low
2882 * on free block. The free block accounting via percpu
2883 * counters can get slightly wrong with percpu_counter_batch getting
2884 * accumulated on each CPU without updating global counters
2885 * Delalloc need an accurate free block accounting. So switch
2886 * to non delalloc when we are near to error range.
2887 */
2888 free_clusters =
2889 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2890 dirty_clusters =
2891 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2892 /*
2893 * Start pushing delalloc when 1/2 of free blocks are dirty.
2894 */
2895 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2896 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2897
2898 if (2 * free_clusters < 3 * dirty_clusters ||
2899 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2900 /*
2901 * free block count is less than 150% of dirty blocks
2902 * or free blocks is less than watermark
2903 */
2904 return 1;
2905 }
2906 return 0;
2907}
2908
2909static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2910 loff_t pos, unsigned len,
2911 struct folio **foliop, void **fsdata)
2912{
2913 int ret, retries = 0;
2914 struct folio *folio;
2915 pgoff_t index;
2916 struct inode *inode = mapping->host;
2917
2918 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2919 return -EIO;
2920
2921 index = pos >> PAGE_SHIFT;
2922
2923 if (ext4_nonda_switch(inode->i_sb) || ext4_verity_in_progress(inode)) {
2924 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2925 return ext4_write_begin(file, mapping, pos,
2926 len, foliop, fsdata);
2927 }
2928 *fsdata = (void *)0;
2929 trace_ext4_da_write_begin(inode, pos, len);
2930
2931 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2932 ret = ext4_da_write_inline_data_begin(mapping, inode, pos, len,
2933 foliop, fsdata);
2934 if (ret < 0)
2935 return ret;
2936 if (ret == 1)
2937 return 0;
2938 }
2939
2940retry:
2941 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
2942 mapping_gfp_mask(mapping));
2943 if (IS_ERR(folio))
2944 return PTR_ERR(folio);
2945
2946 ret = ext4_block_write_begin(NULL, folio, pos, len,
2947 ext4_da_get_block_prep);
2948 if (ret < 0) {
2949 folio_unlock(folio);
2950 folio_put(folio);
2951 /*
2952 * block_write_begin may have instantiated a few blocks
2953 * outside i_size. Trim these off again. Don't need
2954 * i_size_read because we hold inode lock.
2955 */
2956 if (pos + len > inode->i_size)
2957 ext4_truncate_failed_write(inode);
2958
2959 if (ret == -ENOSPC &&
2960 ext4_should_retry_alloc(inode->i_sb, &retries))
2961 goto retry;
2962 return ret;
2963 }
2964
2965 *foliop = folio;
2966 return ret;
2967}
2968
2969/*
2970 * Check if we should update i_disksize
2971 * when write to the end of file but not require block allocation
2972 */
2973static int ext4_da_should_update_i_disksize(struct folio *folio,
2974 unsigned long offset)
2975{
2976 struct buffer_head *bh;
2977 struct inode *inode = folio->mapping->host;
2978 unsigned int idx;
2979 int i;
2980
2981 bh = folio_buffers(folio);
2982 idx = offset >> inode->i_blkbits;
2983
2984 for (i = 0; i < idx; i++)
2985 bh = bh->b_this_page;
2986
2987 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2988 return 0;
2989 return 1;
2990}
2991
2992static int ext4_da_do_write_end(struct address_space *mapping,
2993 loff_t pos, unsigned len, unsigned copied,
2994 struct folio *folio)
2995{
2996 struct inode *inode = mapping->host;
2997 loff_t old_size = inode->i_size;
2998 bool disksize_changed = false;
2999 loff_t new_i_size, zero_len = 0;
3000 handle_t *handle;
3001
3002 if (unlikely(!folio_buffers(folio))) {
3003 folio_unlock(folio);
3004 folio_put(folio);
3005 return -EIO;
3006 }
3007 /*
3008 * block_write_end() will mark the inode as dirty with I_DIRTY_PAGES
3009 * flag, which all that's needed to trigger page writeback.
3010 */
3011 copied = block_write_end(NULL, mapping, pos, len, copied,
3012 folio, NULL);
3013 new_i_size = pos + copied;
3014
3015 /*
3016 * It's important to update i_size while still holding folio lock,
3017 * because folio writeout could otherwise come in and zero beyond
3018 * i_size.
3019 *
3020 * Since we are holding inode lock, we are sure i_disksize <=
3021 * i_size. We also know that if i_disksize < i_size, there are
3022 * delalloc writes pending in the range up to i_size. If the end of
3023 * the current write is <= i_size, there's no need to touch
3024 * i_disksize since writeback will push i_disksize up to i_size
3025 * eventually. If the end of the current write is > i_size and
3026 * inside an allocated block which ext4_da_should_update_i_disksize()
3027 * checked, we need to update i_disksize here as certain
3028 * ext4_writepages() paths not allocating blocks and update i_disksize.
3029 */
3030 if (new_i_size > inode->i_size) {
3031 unsigned long end;
3032
3033 i_size_write(inode, new_i_size);
3034 end = (new_i_size - 1) & (PAGE_SIZE - 1);
3035 if (copied && ext4_da_should_update_i_disksize(folio, end)) {
3036 ext4_update_i_disksize(inode, new_i_size);
3037 disksize_changed = true;
3038 }
3039 }
3040
3041 folio_unlock(folio);
3042 folio_put(folio);
3043
3044 if (pos > old_size) {
3045 pagecache_isize_extended(inode, old_size, pos);
3046 zero_len = pos - old_size;
3047 }
3048
3049 if (!disksize_changed && !zero_len)
3050 return copied;
3051
3052 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3053 if (IS_ERR(handle))
3054 return PTR_ERR(handle);
3055 if (zero_len)
3056 ext4_zero_partial_blocks(handle, inode, old_size, zero_len);
3057 ext4_mark_inode_dirty(handle, inode);
3058 ext4_journal_stop(handle);
3059
3060 return copied;
3061}
3062
3063static int ext4_da_write_end(struct file *file,
3064 struct address_space *mapping,
3065 loff_t pos, unsigned len, unsigned copied,
3066 struct folio *folio, void *fsdata)
3067{
3068 struct inode *inode = mapping->host;
3069 int write_mode = (int)(unsigned long)fsdata;
3070
3071 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3072 return ext4_write_end(file, mapping, pos,
3073 len, copied, folio, fsdata);
3074
3075 trace_ext4_da_write_end(inode, pos, len, copied);
3076
3077 if (write_mode != CONVERT_INLINE_DATA &&
3078 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3079 ext4_has_inline_data(inode))
3080 return ext4_write_inline_data_end(inode, pos, len, copied,
3081 folio);
3082
3083 if (unlikely(copied < len) && !folio_test_uptodate(folio))
3084 copied = 0;
3085
3086 return ext4_da_do_write_end(mapping, pos, len, copied, folio);
3087}
3088
3089/*
3090 * Force all delayed allocation blocks to be allocated for a given inode.
3091 */
3092int ext4_alloc_da_blocks(struct inode *inode)
3093{
3094 trace_ext4_alloc_da_blocks(inode);
3095
3096 if (!EXT4_I(inode)->i_reserved_data_blocks)
3097 return 0;
3098
3099 /*
3100 * We do something simple for now. The filemap_flush() will
3101 * also start triggering a write of the data blocks, which is
3102 * not strictly speaking necessary (and for users of
3103 * laptop_mode, not even desirable). However, to do otherwise
3104 * would require replicating code paths in:
3105 *
3106 * ext4_writepages() ->
3107 * write_cache_pages() ---> (via passed in callback function)
3108 * __mpage_da_writepage() -->
3109 * mpage_add_bh_to_extent()
3110 * mpage_da_map_blocks()
3111 *
3112 * The problem is that write_cache_pages(), located in
3113 * mm/page-writeback.c, marks pages clean in preparation for
3114 * doing I/O, which is not desirable if we're not planning on
3115 * doing I/O at all.
3116 *
3117 * We could call write_cache_pages(), and then redirty all of
3118 * the pages by calling redirty_page_for_writepage() but that
3119 * would be ugly in the extreme. So instead we would need to
3120 * replicate parts of the code in the above functions,
3121 * simplifying them because we wouldn't actually intend to
3122 * write out the pages, but rather only collect contiguous
3123 * logical block extents, call the multi-block allocator, and
3124 * then update the buffer heads with the block allocations.
3125 *
3126 * For now, though, we'll cheat by calling filemap_flush(),
3127 * which will map the blocks, and start the I/O, but not
3128 * actually wait for the I/O to complete.
3129 */
3130 return filemap_flush(inode->i_mapping);
3131}
3132
3133/*
3134 * bmap() is special. It gets used by applications such as lilo and by
3135 * the swapper to find the on-disk block of a specific piece of data.
3136 *
3137 * Naturally, this is dangerous if the block concerned is still in the
3138 * journal. If somebody makes a swapfile on an ext4 data-journaling
3139 * filesystem and enables swap, then they may get a nasty shock when the
3140 * data getting swapped to that swapfile suddenly gets overwritten by
3141 * the original zero's written out previously to the journal and
3142 * awaiting writeback in the kernel's buffer cache.
3143 *
3144 * So, if we see any bmap calls here on a modified, data-journaled file,
3145 * take extra steps to flush any blocks which might be in the cache.
3146 */
3147static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3148{
3149 struct inode *inode = mapping->host;
3150 sector_t ret = 0;
3151
3152 inode_lock_shared(inode);
3153 /*
3154 * We can get here for an inline file via the FIBMAP ioctl
3155 */
3156 if (ext4_has_inline_data(inode))
3157 goto out;
3158
3159 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3160 (test_opt(inode->i_sb, DELALLOC) ||
3161 ext4_should_journal_data(inode))) {
3162 /*
3163 * With delalloc or journalled data we want to sync the file so
3164 * that we can make sure we allocate blocks for file and data
3165 * is in place for the user to see it
3166 */
3167 filemap_write_and_wait(mapping);
3168 }
3169
3170 ret = iomap_bmap(mapping, block, &ext4_iomap_ops);
3171
3172out:
3173 inode_unlock_shared(inode);
3174 return ret;
3175}
3176
3177static int ext4_read_folio(struct file *file, struct folio *folio)
3178{
3179 int ret = -EAGAIN;
3180 struct inode *inode = folio->mapping->host;
3181
3182 trace_ext4_read_folio(inode, folio);
3183
3184 if (ext4_has_inline_data(inode))
3185 ret = ext4_readpage_inline(inode, folio);
3186
3187 if (ret == -EAGAIN)
3188 return ext4_mpage_readpages(inode, NULL, folio);
3189
3190 return ret;
3191}
3192
3193static void ext4_readahead(struct readahead_control *rac)
3194{
3195 struct inode *inode = rac->mapping->host;
3196
3197 /* If the file has inline data, no need to do readahead. */
3198 if (ext4_has_inline_data(inode))
3199 return;
3200
3201 ext4_mpage_readpages(inode, rac, NULL);
3202}
3203
3204static void ext4_invalidate_folio(struct folio *folio, size_t offset,
3205 size_t length)
3206{
3207 trace_ext4_invalidate_folio(folio, offset, length);
3208
3209 /* No journalling happens on data buffers when this function is used */
3210 WARN_ON(folio_buffers(folio) && buffer_jbd(folio_buffers(folio)));
3211
3212 block_invalidate_folio(folio, offset, length);
3213}
3214
3215static int __ext4_journalled_invalidate_folio(struct folio *folio,
3216 size_t offset, size_t length)
3217{
3218 journal_t *journal = EXT4_JOURNAL(folio->mapping->host);
3219
3220 trace_ext4_journalled_invalidate_folio(folio, offset, length);
3221
3222 /*
3223 * If it's a full truncate we just forget about the pending dirtying
3224 */
3225 if (offset == 0 && length == folio_size(folio))
3226 folio_clear_checked(folio);
3227
3228 return jbd2_journal_invalidate_folio(journal, folio, offset, length);
3229}
3230
3231/* Wrapper for aops... */
3232static void ext4_journalled_invalidate_folio(struct folio *folio,
3233 size_t offset,
3234 size_t length)
3235{
3236 WARN_ON(__ext4_journalled_invalidate_folio(folio, offset, length) < 0);
3237}
3238
3239static bool ext4_release_folio(struct folio *folio, gfp_t wait)
3240{
3241 struct inode *inode = folio->mapping->host;
3242 journal_t *journal = EXT4_JOURNAL(inode);
3243
3244 trace_ext4_release_folio(inode, folio);
3245
3246 /* Page has dirty journalled data -> cannot release */
3247 if (folio_test_checked(folio))
3248 return false;
3249 if (journal)
3250 return jbd2_journal_try_to_free_buffers(journal, folio);
3251 else
3252 return try_to_free_buffers(folio);
3253}
3254
3255static bool ext4_inode_datasync_dirty(struct inode *inode)
3256{
3257 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3258
3259 if (journal) {
3260 if (jbd2_transaction_committed(journal,
3261 EXT4_I(inode)->i_datasync_tid))
3262 return false;
3263 if (test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT))
3264 return !list_empty(&EXT4_I(inode)->i_fc_list);
3265 return true;
3266 }
3267
3268 /* Any metadata buffers to write? */
3269 if (!list_empty(&inode->i_mapping->i_private_list))
3270 return true;
3271 return inode->i_state & I_DIRTY_DATASYNC;
3272}
3273
3274static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
3275 struct ext4_map_blocks *map, loff_t offset,
3276 loff_t length, unsigned int flags)
3277{
3278 u8 blkbits = inode->i_blkbits;
3279
3280 /*
3281 * Writes that span EOF might trigger an I/O size update on completion,
3282 * so consider them to be dirty for the purpose of O_DSYNC, even if
3283 * there is no other metadata changes being made or are pending.
3284 */
3285 iomap->flags = 0;
3286 if (ext4_inode_datasync_dirty(inode) ||
3287 offset + length > i_size_read(inode))
3288 iomap->flags |= IOMAP_F_DIRTY;
3289
3290 if (map->m_flags & EXT4_MAP_NEW)
3291 iomap->flags |= IOMAP_F_NEW;
3292
3293 if (flags & IOMAP_DAX)
3294 iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
3295 else
3296 iomap->bdev = inode->i_sb->s_bdev;
3297 iomap->offset = (u64) map->m_lblk << blkbits;
3298 iomap->length = (u64) map->m_len << blkbits;
3299
3300 if ((map->m_flags & EXT4_MAP_MAPPED) &&
3301 !ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3302 iomap->flags |= IOMAP_F_MERGED;
3303
3304 /*
3305 * Flags passed to ext4_map_blocks() for direct I/O writes can result
3306 * in m_flags having both EXT4_MAP_MAPPED and EXT4_MAP_UNWRITTEN bits
3307 * set. In order for any allocated unwritten extents to be converted
3308 * into written extents correctly within the ->end_io() handler, we
3309 * need to ensure that the iomap->type is set appropriately. Hence, the
3310 * reason why we need to check whether the EXT4_MAP_UNWRITTEN bit has
3311 * been set first.
3312 */
3313 if (map->m_flags & EXT4_MAP_UNWRITTEN) {
3314 iomap->type = IOMAP_UNWRITTEN;
3315 iomap->addr = (u64) map->m_pblk << blkbits;
3316 if (flags & IOMAP_DAX)
3317 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3318 } else if (map->m_flags & EXT4_MAP_MAPPED) {
3319 iomap->type = IOMAP_MAPPED;
3320 iomap->addr = (u64) map->m_pblk << blkbits;
3321 if (flags & IOMAP_DAX)
3322 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3323 } else if (map->m_flags & EXT4_MAP_DELAYED) {
3324 iomap->type = IOMAP_DELALLOC;
3325 iomap->addr = IOMAP_NULL_ADDR;
3326 } else {
3327 iomap->type = IOMAP_HOLE;
3328 iomap->addr = IOMAP_NULL_ADDR;
3329 }
3330}
3331
3332static int ext4_iomap_alloc(struct inode *inode, struct ext4_map_blocks *map,
3333 unsigned int flags)
3334{
3335 handle_t *handle;
3336 u8 blkbits = inode->i_blkbits;
3337 int ret, dio_credits, m_flags = 0, retries = 0;
3338
3339 /*
3340 * Trim the mapping request to the maximum value that we can map at
3341 * once for direct I/O.
3342 */
3343 if (map->m_len > DIO_MAX_BLOCKS)
3344 map->m_len = DIO_MAX_BLOCKS;
3345 dio_credits = ext4_chunk_trans_blocks(inode, map->m_len);
3346
3347retry:
3348 /*
3349 * Either we allocate blocks and then don't get an unwritten extent, so
3350 * in that case we have reserved enough credits. Or, the blocks are
3351 * already allocated and unwritten. In that case, the extent conversion
3352 * fits into the credits as well.
3353 */
3354 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
3355 if (IS_ERR(handle))
3356 return PTR_ERR(handle);
3357
3358 /*
3359 * DAX and direct I/O are the only two operations that are currently
3360 * supported with IOMAP_WRITE.
3361 */
3362 WARN_ON(!(flags & (IOMAP_DAX | IOMAP_DIRECT)));
3363 if (flags & IOMAP_DAX)
3364 m_flags = EXT4_GET_BLOCKS_CREATE_ZERO;
3365 /*
3366 * We use i_size instead of i_disksize here because delalloc writeback
3367 * can complete at any point during the I/O and subsequently push the
3368 * i_disksize out to i_size. This could be beyond where direct I/O is
3369 * happening and thus expose allocated blocks to direct I/O reads.
3370 */
3371 else if (((loff_t)map->m_lblk << blkbits) >= i_size_read(inode))
3372 m_flags = EXT4_GET_BLOCKS_CREATE;
3373 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3374 m_flags = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3375
3376 ret = ext4_map_blocks(handle, inode, map, m_flags);
3377
3378 /*
3379 * We cannot fill holes in indirect tree based inodes as that could
3380 * expose stale data in the case of a crash. Use the magic error code
3381 * to fallback to buffered I/O.
3382 */
3383 if (!m_flags && !ret)
3384 ret = -ENOTBLK;
3385
3386 ext4_journal_stop(handle);
3387 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3388 goto retry;
3389
3390 return ret;
3391}
3392
3393
3394static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3395 unsigned flags, struct iomap *iomap, struct iomap *srcmap)
3396{
3397 int ret;
3398 struct ext4_map_blocks map;
3399 u8 blkbits = inode->i_blkbits;
3400
3401 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3402 return -EINVAL;
3403
3404 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3405 return -ERANGE;
3406
3407 /*
3408 * Calculate the first and last logical blocks respectively.
3409 */
3410 map.m_lblk = offset >> blkbits;
3411 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3412 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3413
3414 if (flags & IOMAP_WRITE) {
3415 /*
3416 * We check here if the blocks are already allocated, then we
3417 * don't need to start a journal txn and we can directly return
3418 * the mapping information. This could boost performance
3419 * especially in multi-threaded overwrite requests.
3420 */
3421 if (offset + length <= i_size_read(inode)) {
3422 ret = ext4_map_blocks(NULL, inode, &map, 0);
3423 if (ret > 0 && (map.m_flags & EXT4_MAP_MAPPED))
3424 goto out;
3425 }
3426 ret = ext4_iomap_alloc(inode, &map, flags);
3427 } else {
3428 ret = ext4_map_blocks(NULL, inode, &map, 0);
3429 }
3430
3431 if (ret < 0)
3432 return ret;
3433out:
3434 /*
3435 * When inline encryption is enabled, sometimes I/O to an encrypted file
3436 * has to be broken up to guarantee DUN contiguity. Handle this by
3437 * limiting the length of the mapping returned.
3438 */
3439 map.m_len = fscrypt_limit_io_blocks(inode, map.m_lblk, map.m_len);
3440
3441 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3442
3443 return 0;
3444}
3445
3446static int ext4_iomap_overwrite_begin(struct inode *inode, loff_t offset,
3447 loff_t length, unsigned flags, struct iomap *iomap,
3448 struct iomap *srcmap)
3449{
3450 int ret;
3451
3452 /*
3453 * Even for writes we don't need to allocate blocks, so just pretend
3454 * we are reading to save overhead of starting a transaction.
3455 */
3456 flags &= ~IOMAP_WRITE;
3457 ret = ext4_iomap_begin(inode, offset, length, flags, iomap, srcmap);
3458 WARN_ON_ONCE(!ret && iomap->type != IOMAP_MAPPED);
3459 return ret;
3460}
3461
3462static inline bool ext4_want_directio_fallback(unsigned flags, ssize_t written)
3463{
3464 /* must be a directio to fall back to buffered */
3465 if ((flags & (IOMAP_WRITE | IOMAP_DIRECT)) !=
3466 (IOMAP_WRITE | IOMAP_DIRECT))
3467 return false;
3468
3469 /* atomic writes are all-or-nothing */
3470 if (flags & IOMAP_ATOMIC)
3471 return false;
3472
3473 /* can only try again if we wrote nothing */
3474 return written == 0;
3475}
3476
3477static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3478 ssize_t written, unsigned flags, struct iomap *iomap)
3479{
3480 /*
3481 * Check to see whether an error occurred while writing out the data to
3482 * the allocated blocks. If so, return the magic error code for
3483 * non-atomic write so that we fallback to buffered I/O and attempt to
3484 * complete the remainder of the I/O.
3485 * For non-atomic writes, any blocks that may have been
3486 * allocated in preparation for the direct I/O will be reused during
3487 * buffered I/O. For atomic write, we never fallback to buffered-io.
3488 */
3489 if (ext4_want_directio_fallback(flags, written))
3490 return -ENOTBLK;
3491
3492 return 0;
3493}
3494
3495const struct iomap_ops ext4_iomap_ops = {
3496 .iomap_begin = ext4_iomap_begin,
3497 .iomap_end = ext4_iomap_end,
3498};
3499
3500const struct iomap_ops ext4_iomap_overwrite_ops = {
3501 .iomap_begin = ext4_iomap_overwrite_begin,
3502 .iomap_end = ext4_iomap_end,
3503};
3504
3505static int ext4_iomap_begin_report(struct inode *inode, loff_t offset,
3506 loff_t length, unsigned int flags,
3507 struct iomap *iomap, struct iomap *srcmap)
3508{
3509 int ret;
3510 struct ext4_map_blocks map;
3511 u8 blkbits = inode->i_blkbits;
3512
3513 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3514 return -EINVAL;
3515
3516 if (ext4_has_inline_data(inode)) {
3517 ret = ext4_inline_data_iomap(inode, iomap);
3518 if (ret != -EAGAIN) {
3519 if (ret == 0 && offset >= iomap->length)
3520 ret = -ENOENT;
3521 return ret;
3522 }
3523 }
3524
3525 /*
3526 * Calculate the first and last logical block respectively.
3527 */
3528 map.m_lblk = offset >> blkbits;
3529 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3530 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3531
3532 /*
3533 * Fiemap callers may call for offset beyond s_bitmap_maxbytes.
3534 * So handle it here itself instead of querying ext4_map_blocks().
3535 * Since ext4_map_blocks() will warn about it and will return
3536 * -EIO error.
3537 */
3538 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3539 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3540
3541 if (offset >= sbi->s_bitmap_maxbytes) {
3542 map.m_flags = 0;
3543 goto set_iomap;
3544 }
3545 }
3546
3547 ret = ext4_map_blocks(NULL, inode, &map, 0);
3548 if (ret < 0)
3549 return ret;
3550set_iomap:
3551 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3552
3553 return 0;
3554}
3555
3556const struct iomap_ops ext4_iomap_report_ops = {
3557 .iomap_begin = ext4_iomap_begin_report,
3558};
3559
3560/*
3561 * For data=journal mode, folio should be marked dirty only when it was
3562 * writeably mapped. When that happens, it was already attached to the
3563 * transaction and marked as jbddirty (we take care of this in
3564 * ext4_page_mkwrite()). On transaction commit, we writeprotect page mappings
3565 * so we should have nothing to do here, except for the case when someone
3566 * had the page pinned and dirtied the page through this pin (e.g. by doing
3567 * direct IO to it). In that case we'd need to attach buffers here to the
3568 * transaction but we cannot due to lock ordering. We cannot just dirty the
3569 * folio and leave attached buffers clean, because the buffers' dirty state is
3570 * "definitive". We cannot just set the buffers dirty or jbddirty because all
3571 * the journalling code will explode. So what we do is to mark the folio
3572 * "pending dirty" and next time ext4_writepages() is called, attach buffers
3573 * to the transaction appropriately.
3574 */
3575static bool ext4_journalled_dirty_folio(struct address_space *mapping,
3576 struct folio *folio)
3577{
3578 WARN_ON_ONCE(!folio_buffers(folio));
3579 if (folio_maybe_dma_pinned(folio))
3580 folio_set_checked(folio);
3581 return filemap_dirty_folio(mapping, folio);
3582}
3583
3584static bool ext4_dirty_folio(struct address_space *mapping, struct folio *folio)
3585{
3586 WARN_ON_ONCE(!folio_test_locked(folio) && !folio_test_dirty(folio));
3587 WARN_ON_ONCE(!folio_buffers(folio));
3588 return block_dirty_folio(mapping, folio);
3589}
3590
3591static int ext4_iomap_swap_activate(struct swap_info_struct *sis,
3592 struct file *file, sector_t *span)
3593{
3594 return iomap_swapfile_activate(sis, file, span,
3595 &ext4_iomap_report_ops);
3596}
3597
3598static const struct address_space_operations ext4_aops = {
3599 .read_folio = ext4_read_folio,
3600 .readahead = ext4_readahead,
3601 .writepages = ext4_writepages,
3602 .write_begin = ext4_write_begin,
3603 .write_end = ext4_write_end,
3604 .dirty_folio = ext4_dirty_folio,
3605 .bmap = ext4_bmap,
3606 .invalidate_folio = ext4_invalidate_folio,
3607 .release_folio = ext4_release_folio,
3608 .migrate_folio = buffer_migrate_folio,
3609 .is_partially_uptodate = block_is_partially_uptodate,
3610 .error_remove_folio = generic_error_remove_folio,
3611 .swap_activate = ext4_iomap_swap_activate,
3612};
3613
3614static const struct address_space_operations ext4_journalled_aops = {
3615 .read_folio = ext4_read_folio,
3616 .readahead = ext4_readahead,
3617 .writepages = ext4_writepages,
3618 .write_begin = ext4_write_begin,
3619 .write_end = ext4_journalled_write_end,
3620 .dirty_folio = ext4_journalled_dirty_folio,
3621 .bmap = ext4_bmap,
3622 .invalidate_folio = ext4_journalled_invalidate_folio,
3623 .release_folio = ext4_release_folio,
3624 .migrate_folio = buffer_migrate_folio_norefs,
3625 .is_partially_uptodate = block_is_partially_uptodate,
3626 .error_remove_folio = generic_error_remove_folio,
3627 .swap_activate = ext4_iomap_swap_activate,
3628};
3629
3630static const struct address_space_operations ext4_da_aops = {
3631 .read_folio = ext4_read_folio,
3632 .readahead = ext4_readahead,
3633 .writepages = ext4_writepages,
3634 .write_begin = ext4_da_write_begin,
3635 .write_end = ext4_da_write_end,
3636 .dirty_folio = ext4_dirty_folio,
3637 .bmap = ext4_bmap,
3638 .invalidate_folio = ext4_invalidate_folio,
3639 .release_folio = ext4_release_folio,
3640 .migrate_folio = buffer_migrate_folio,
3641 .is_partially_uptodate = block_is_partially_uptodate,
3642 .error_remove_folio = generic_error_remove_folio,
3643 .swap_activate = ext4_iomap_swap_activate,
3644};
3645
3646static const struct address_space_operations ext4_dax_aops = {
3647 .writepages = ext4_dax_writepages,
3648 .dirty_folio = noop_dirty_folio,
3649 .bmap = ext4_bmap,
3650 .swap_activate = ext4_iomap_swap_activate,
3651};
3652
3653void ext4_set_aops(struct inode *inode)
3654{
3655 switch (ext4_inode_journal_mode(inode)) {
3656 case EXT4_INODE_ORDERED_DATA_MODE:
3657 case EXT4_INODE_WRITEBACK_DATA_MODE:
3658 break;
3659 case EXT4_INODE_JOURNAL_DATA_MODE:
3660 inode->i_mapping->a_ops = &ext4_journalled_aops;
3661 return;
3662 default:
3663 BUG();
3664 }
3665 if (IS_DAX(inode))
3666 inode->i_mapping->a_ops = &ext4_dax_aops;
3667 else if (test_opt(inode->i_sb, DELALLOC))
3668 inode->i_mapping->a_ops = &ext4_da_aops;
3669 else
3670 inode->i_mapping->a_ops = &ext4_aops;
3671}
3672
3673/*
3674 * Here we can't skip an unwritten buffer even though it usually reads zero
3675 * because it might have data in pagecache (eg, if called from ext4_zero_range,
3676 * ext4_punch_hole, etc) which needs to be properly zeroed out. Otherwise a
3677 * racing writeback can come later and flush the stale pagecache to disk.
3678 */
3679static int __ext4_block_zero_page_range(handle_t *handle,
3680 struct address_space *mapping, loff_t from, loff_t length)
3681{
3682 ext4_fsblk_t index = from >> PAGE_SHIFT;
3683 unsigned offset = from & (PAGE_SIZE-1);
3684 unsigned blocksize, pos;
3685 ext4_lblk_t iblock;
3686 struct inode *inode = mapping->host;
3687 struct buffer_head *bh;
3688 struct folio *folio;
3689 int err = 0;
3690
3691 folio = __filemap_get_folio(mapping, from >> PAGE_SHIFT,
3692 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
3693 mapping_gfp_constraint(mapping, ~__GFP_FS));
3694 if (IS_ERR(folio))
3695 return PTR_ERR(folio);
3696
3697 blocksize = inode->i_sb->s_blocksize;
3698
3699 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3700
3701 bh = folio_buffers(folio);
3702 if (!bh)
3703 bh = create_empty_buffers(folio, blocksize, 0);
3704
3705 /* Find the buffer that contains "offset" */
3706 pos = blocksize;
3707 while (offset >= pos) {
3708 bh = bh->b_this_page;
3709 iblock++;
3710 pos += blocksize;
3711 }
3712 if (buffer_freed(bh)) {
3713 BUFFER_TRACE(bh, "freed: skip");
3714 goto unlock;
3715 }
3716 if (!buffer_mapped(bh)) {
3717 BUFFER_TRACE(bh, "unmapped");
3718 ext4_get_block(inode, iblock, bh, 0);
3719 /* unmapped? It's a hole - nothing to do */
3720 if (!buffer_mapped(bh)) {
3721 BUFFER_TRACE(bh, "still unmapped");
3722 goto unlock;
3723 }
3724 }
3725
3726 /* Ok, it's mapped. Make sure it's up-to-date */
3727 if (folio_test_uptodate(folio))
3728 set_buffer_uptodate(bh);
3729
3730 if (!buffer_uptodate(bh)) {
3731 err = ext4_read_bh_lock(bh, 0, true);
3732 if (err)
3733 goto unlock;
3734 if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
3735 /* We expect the key to be set. */
3736 BUG_ON(!fscrypt_has_encryption_key(inode));
3737 err = fscrypt_decrypt_pagecache_blocks(folio,
3738 blocksize,
3739 bh_offset(bh));
3740 if (err) {
3741 clear_buffer_uptodate(bh);
3742 goto unlock;
3743 }
3744 }
3745 }
3746 if (ext4_should_journal_data(inode)) {
3747 BUFFER_TRACE(bh, "get write access");
3748 err = ext4_journal_get_write_access(handle, inode->i_sb, bh,
3749 EXT4_JTR_NONE);
3750 if (err)
3751 goto unlock;
3752 }
3753 folio_zero_range(folio, offset, length);
3754 BUFFER_TRACE(bh, "zeroed end of block");
3755
3756 if (ext4_should_journal_data(inode)) {
3757 err = ext4_dirty_journalled_data(handle, bh);
3758 } else {
3759 err = 0;
3760 mark_buffer_dirty(bh);
3761 if (ext4_should_order_data(inode))
3762 err = ext4_jbd2_inode_add_write(handle, inode, from,
3763 length);
3764 }
3765
3766unlock:
3767 folio_unlock(folio);
3768 folio_put(folio);
3769 return err;
3770}
3771
3772/*
3773 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3774 * starting from file offset 'from'. The range to be zero'd must
3775 * be contained with in one block. If the specified range exceeds
3776 * the end of the block it will be shortened to end of the block
3777 * that corresponds to 'from'
3778 */
3779static int ext4_block_zero_page_range(handle_t *handle,
3780 struct address_space *mapping, loff_t from, loff_t length)
3781{
3782 struct inode *inode = mapping->host;
3783 unsigned offset = from & (PAGE_SIZE-1);
3784 unsigned blocksize = inode->i_sb->s_blocksize;
3785 unsigned max = blocksize - (offset & (blocksize - 1));
3786
3787 /*
3788 * correct length if it does not fall between
3789 * 'from' and the end of the block
3790 */
3791 if (length > max || length < 0)
3792 length = max;
3793
3794 if (IS_DAX(inode)) {
3795 return dax_zero_range(inode, from, length, NULL,
3796 &ext4_iomap_ops);
3797 }
3798 return __ext4_block_zero_page_range(handle, mapping, from, length);
3799}
3800
3801/*
3802 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3803 * up to the end of the block which corresponds to `from'.
3804 * This required during truncate. We need to physically zero the tail end
3805 * of that block so it doesn't yield old data if the file is later grown.
3806 */
3807static int ext4_block_truncate_page(handle_t *handle,
3808 struct address_space *mapping, loff_t from)
3809{
3810 unsigned offset = from & (PAGE_SIZE-1);
3811 unsigned length;
3812 unsigned blocksize;
3813 struct inode *inode = mapping->host;
3814
3815 /* If we are processing an encrypted inode during orphan list handling */
3816 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
3817 return 0;
3818
3819 blocksize = inode->i_sb->s_blocksize;
3820 length = blocksize - (offset & (blocksize - 1));
3821
3822 return ext4_block_zero_page_range(handle, mapping, from, length);
3823}
3824
3825int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3826 loff_t lstart, loff_t length)
3827{
3828 struct super_block *sb = inode->i_sb;
3829 struct address_space *mapping = inode->i_mapping;
3830 unsigned partial_start, partial_end;
3831 ext4_fsblk_t start, end;
3832 loff_t byte_end = (lstart + length - 1);
3833 int err = 0;
3834
3835 partial_start = lstart & (sb->s_blocksize - 1);
3836 partial_end = byte_end & (sb->s_blocksize - 1);
3837
3838 start = lstart >> sb->s_blocksize_bits;
3839 end = byte_end >> sb->s_blocksize_bits;
3840
3841 /* Handle partial zero within the single block */
3842 if (start == end &&
3843 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3844 err = ext4_block_zero_page_range(handle, mapping,
3845 lstart, length);
3846 return err;
3847 }
3848 /* Handle partial zero out on the start of the range */
3849 if (partial_start) {
3850 err = ext4_block_zero_page_range(handle, mapping,
3851 lstart, sb->s_blocksize);
3852 if (err)
3853 return err;
3854 }
3855 /* Handle partial zero out on the end of the range */
3856 if (partial_end != sb->s_blocksize - 1)
3857 err = ext4_block_zero_page_range(handle, mapping,
3858 byte_end - partial_end,
3859 partial_end + 1);
3860 return err;
3861}
3862
3863int ext4_can_truncate(struct inode *inode)
3864{
3865 if (S_ISREG(inode->i_mode))
3866 return 1;
3867 if (S_ISDIR(inode->i_mode))
3868 return 1;
3869 if (S_ISLNK(inode->i_mode))
3870 return !ext4_inode_is_fast_symlink(inode);
3871 return 0;
3872}
3873
3874/*
3875 * We have to make sure i_disksize gets properly updated before we truncate
3876 * page cache due to hole punching or zero range. Otherwise i_disksize update
3877 * can get lost as it may have been postponed to submission of writeback but
3878 * that will never happen after we truncate page cache.
3879 */
3880int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
3881 loff_t len)
3882{
3883 handle_t *handle;
3884 int ret;
3885
3886 loff_t size = i_size_read(inode);
3887
3888 WARN_ON(!inode_is_locked(inode));
3889 if (offset > size || offset + len < size)
3890 return 0;
3891
3892 if (EXT4_I(inode)->i_disksize >= size)
3893 return 0;
3894
3895 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
3896 if (IS_ERR(handle))
3897 return PTR_ERR(handle);
3898 ext4_update_i_disksize(inode, size);
3899 ret = ext4_mark_inode_dirty(handle, inode);
3900 ext4_journal_stop(handle);
3901
3902 return ret;
3903}
3904
3905static void ext4_wait_dax_page(struct inode *inode)
3906{
3907 filemap_invalidate_unlock(inode->i_mapping);
3908 schedule();
3909 filemap_invalidate_lock(inode->i_mapping);
3910}
3911
3912int ext4_break_layouts(struct inode *inode)
3913{
3914 struct page *page;
3915 int error;
3916
3917 if (WARN_ON_ONCE(!rwsem_is_locked(&inode->i_mapping->invalidate_lock)))
3918 return -EINVAL;
3919
3920 do {
3921 page = dax_layout_busy_page(inode->i_mapping);
3922 if (!page)
3923 return 0;
3924
3925 error = ___wait_var_event(&page->_refcount,
3926 atomic_read(&page->_refcount) == 1,
3927 TASK_INTERRUPTIBLE, 0, 0,
3928 ext4_wait_dax_page(inode));
3929 } while (error == 0);
3930
3931 return error;
3932}
3933
3934/*
3935 * ext4_punch_hole: punches a hole in a file by releasing the blocks
3936 * associated with the given offset and length
3937 *
3938 * @inode: File inode
3939 * @offset: The offset where the hole will begin
3940 * @len: The length of the hole
3941 *
3942 * Returns: 0 on success or negative on failure
3943 */
3944
3945int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3946{
3947 struct inode *inode = file_inode(file);
3948 struct super_block *sb = inode->i_sb;
3949 ext4_lblk_t first_block, stop_block;
3950 struct address_space *mapping = inode->i_mapping;
3951 loff_t first_block_offset, last_block_offset, max_length;
3952 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3953 handle_t *handle;
3954 unsigned int credits;
3955 int ret = 0, ret2 = 0;
3956
3957 trace_ext4_punch_hole(inode, offset, length, 0);
3958
3959 /*
3960 * Write out all dirty pages to avoid race conditions
3961 * Then release them.
3962 */
3963 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3964 ret = filemap_write_and_wait_range(mapping, offset,
3965 offset + length - 1);
3966 if (ret)
3967 return ret;
3968 }
3969
3970 inode_lock(inode);
3971
3972 /* No need to punch hole beyond i_size */
3973 if (offset >= inode->i_size)
3974 goto out_mutex;
3975
3976 /*
3977 * If the hole extends beyond i_size, set the hole
3978 * to end after the page that contains i_size
3979 */
3980 if (offset + length > inode->i_size) {
3981 length = inode->i_size +
3982 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
3983 offset;
3984 }
3985
3986 /*
3987 * For punch hole the length + offset needs to be within one block
3988 * before last range. Adjust the length if it goes beyond that limit.
3989 */
3990 max_length = sbi->s_bitmap_maxbytes - inode->i_sb->s_blocksize;
3991 if (offset + length > max_length)
3992 length = max_length - offset;
3993
3994 if (offset & (sb->s_blocksize - 1) ||
3995 (offset + length) & (sb->s_blocksize - 1)) {
3996 /*
3997 * Attach jinode to inode for jbd2 if we do any zeroing of
3998 * partial block
3999 */
4000 ret = ext4_inode_attach_jinode(inode);
4001 if (ret < 0)
4002 goto out_mutex;
4003
4004 }
4005
4006 /* Wait all existing dio workers, newcomers will block on i_rwsem */
4007 inode_dio_wait(inode);
4008
4009 ret = file_modified(file);
4010 if (ret)
4011 goto out_mutex;
4012
4013 /*
4014 * Prevent page faults from reinstantiating pages we have released from
4015 * page cache.
4016 */
4017 filemap_invalidate_lock(mapping);
4018
4019 ret = ext4_break_layouts(inode);
4020 if (ret)
4021 goto out_dio;
4022
4023 first_block_offset = round_up(offset, sb->s_blocksize);
4024 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4025
4026 /* Now release the pages and zero block aligned part of pages*/
4027 if (last_block_offset > first_block_offset) {
4028 ret = ext4_update_disksize_before_punch(inode, offset, length);
4029 if (ret)
4030 goto out_dio;
4031 truncate_pagecache_range(inode, first_block_offset,
4032 last_block_offset);
4033 }
4034
4035 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4036 credits = ext4_writepage_trans_blocks(inode);
4037 else
4038 credits = ext4_blocks_for_truncate(inode);
4039 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4040 if (IS_ERR(handle)) {
4041 ret = PTR_ERR(handle);
4042 ext4_std_error(sb, ret);
4043 goto out_dio;
4044 }
4045
4046 ret = ext4_zero_partial_blocks(handle, inode, offset,
4047 length);
4048 if (ret)
4049 goto out_stop;
4050
4051 first_block = (offset + sb->s_blocksize - 1) >>
4052 EXT4_BLOCK_SIZE_BITS(sb);
4053 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4054
4055 /* If there are blocks to remove, do it */
4056 if (stop_block > first_block) {
4057 ext4_lblk_t hole_len = stop_block - first_block;
4058
4059 down_write(&EXT4_I(inode)->i_data_sem);
4060 ext4_discard_preallocations(inode);
4061
4062 ext4_es_remove_extent(inode, first_block, hole_len);
4063
4064 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4065 ret = ext4_ext_remove_space(inode, first_block,
4066 stop_block - 1);
4067 else
4068 ret = ext4_ind_remove_space(handle, inode, first_block,
4069 stop_block);
4070
4071 ext4_es_insert_extent(inode, first_block, hole_len, ~0,
4072 EXTENT_STATUS_HOLE, 0);
4073 up_write(&EXT4_I(inode)->i_data_sem);
4074 }
4075 ext4_fc_track_range(handle, inode, first_block, stop_block);
4076 if (IS_SYNC(inode))
4077 ext4_handle_sync(handle);
4078
4079 inode_set_mtime_to_ts(inode, inode_set_ctime_current(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 filemap_invalidate_unlock(mapping);
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_rwsem 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 err = ext4_inode_attach_jinode(inode);
4181 if (err)
4182 goto out_trace;
4183 }
4184
4185 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4186 credits = ext4_writepage_trans_blocks(inode);
4187 else
4188 credits = ext4_blocks_for_truncate(inode);
4189
4190 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4191 if (IS_ERR(handle)) {
4192 err = PTR_ERR(handle);
4193 goto out_trace;
4194 }
4195
4196 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4197 ext4_block_truncate_page(handle, mapping, inode->i_size);
4198
4199 /*
4200 * We add the inode to the orphan list, so that if this
4201 * truncate spans multiple transactions, and we crash, we will
4202 * resume the truncate when the filesystem recovers. It also
4203 * marks the inode dirty, to catch the new size.
4204 *
4205 * Implication: the file must always be in a sane, consistent
4206 * truncatable state while each transaction commits.
4207 */
4208 err = ext4_orphan_add(handle, inode);
4209 if (err)
4210 goto out_stop;
4211
4212 down_write(&EXT4_I(inode)->i_data_sem);
4213
4214 ext4_discard_preallocations(inode);
4215
4216 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4217 err = ext4_ext_truncate(handle, inode);
4218 else
4219 ext4_ind_truncate(handle, inode);
4220
4221 up_write(&ei->i_data_sem);
4222 if (err)
4223 goto out_stop;
4224
4225 if (IS_SYNC(inode))
4226 ext4_handle_sync(handle);
4227
4228out_stop:
4229 /*
4230 * If this was a simple ftruncate() and the file will remain alive,
4231 * then we need to clear up the orphan record which we created above.
4232 * However, if this was a real unlink then we were called by
4233 * ext4_evict_inode(), and we allow that function to clean up the
4234 * orphan info for us.
4235 */
4236 if (inode->i_nlink)
4237 ext4_orphan_del(handle, inode);
4238
4239 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
4240 err2 = ext4_mark_inode_dirty(handle, inode);
4241 if (unlikely(err2 && !err))
4242 err = err2;
4243 ext4_journal_stop(handle);
4244
4245out_trace:
4246 trace_ext4_truncate_exit(inode);
4247 return err;
4248}
4249
4250static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4251{
4252 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4253 return inode_peek_iversion_raw(inode);
4254 else
4255 return inode_peek_iversion(inode);
4256}
4257
4258static int ext4_inode_blocks_set(struct ext4_inode *raw_inode,
4259 struct ext4_inode_info *ei)
4260{
4261 struct inode *inode = &(ei->vfs_inode);
4262 u64 i_blocks = READ_ONCE(inode->i_blocks);
4263 struct super_block *sb = inode->i_sb;
4264
4265 if (i_blocks <= ~0U) {
4266 /*
4267 * i_blocks can be represented in a 32 bit variable
4268 * as multiple of 512 bytes
4269 */
4270 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4271 raw_inode->i_blocks_high = 0;
4272 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4273 return 0;
4274 }
4275
4276 /*
4277 * This should never happen since sb->s_maxbytes should not have
4278 * allowed this, sb->s_maxbytes was set according to the huge_file
4279 * feature in ext4_fill_super().
4280 */
4281 if (!ext4_has_feature_huge_file(sb))
4282 return -EFSCORRUPTED;
4283
4284 if (i_blocks <= 0xffffffffffffULL) {
4285 /*
4286 * i_blocks can be represented in a 48 bit variable
4287 * as multiple of 512 bytes
4288 */
4289 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4290 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4291 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4292 } else {
4293 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4294 /* i_block is stored in file system block size */
4295 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4296 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4297 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4298 }
4299 return 0;
4300}
4301
4302static int ext4_fill_raw_inode(struct inode *inode, struct ext4_inode *raw_inode)
4303{
4304 struct ext4_inode_info *ei = EXT4_I(inode);
4305 uid_t i_uid;
4306 gid_t i_gid;
4307 projid_t i_projid;
4308 int block;
4309 int err;
4310
4311 err = ext4_inode_blocks_set(raw_inode, ei);
4312
4313 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4314 i_uid = i_uid_read(inode);
4315 i_gid = i_gid_read(inode);
4316 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4317 if (!(test_opt(inode->i_sb, NO_UID32))) {
4318 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4319 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4320 /*
4321 * Fix up interoperability with old kernels. Otherwise,
4322 * old inodes get re-used with the upper 16 bits of the
4323 * uid/gid intact.
4324 */
4325 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4326 raw_inode->i_uid_high = 0;
4327 raw_inode->i_gid_high = 0;
4328 } else {
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 }
4334 } else {
4335 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4336 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4337 raw_inode->i_uid_high = 0;
4338 raw_inode->i_gid_high = 0;
4339 }
4340 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4341
4342 EXT4_INODE_SET_CTIME(inode, raw_inode);
4343 EXT4_INODE_SET_MTIME(inode, raw_inode);
4344 EXT4_INODE_SET_ATIME(inode, raw_inode);
4345 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4346
4347 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4348 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4349 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4350 raw_inode->i_file_acl_high =
4351 cpu_to_le16(ei->i_file_acl >> 32);
4352 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4353 ext4_isize_set(raw_inode, ei->i_disksize);
4354
4355 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4356 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4357 if (old_valid_dev(inode->i_rdev)) {
4358 raw_inode->i_block[0] =
4359 cpu_to_le32(old_encode_dev(inode->i_rdev));
4360 raw_inode->i_block[1] = 0;
4361 } else {
4362 raw_inode->i_block[0] = 0;
4363 raw_inode->i_block[1] =
4364 cpu_to_le32(new_encode_dev(inode->i_rdev));
4365 raw_inode->i_block[2] = 0;
4366 }
4367 } else if (!ext4_has_inline_data(inode)) {
4368 for (block = 0; block < EXT4_N_BLOCKS; block++)
4369 raw_inode->i_block[block] = ei->i_data[block];
4370 }
4371
4372 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4373 u64 ivers = ext4_inode_peek_iversion(inode);
4374
4375 raw_inode->i_disk_version = cpu_to_le32(ivers);
4376 if (ei->i_extra_isize) {
4377 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4378 raw_inode->i_version_hi =
4379 cpu_to_le32(ivers >> 32);
4380 raw_inode->i_extra_isize =
4381 cpu_to_le16(ei->i_extra_isize);
4382 }
4383 }
4384
4385 if (i_projid != EXT4_DEF_PROJID &&
4386 !ext4_has_feature_project(inode->i_sb))
4387 err = err ?: -EFSCORRUPTED;
4388
4389 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4390 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4391 raw_inode->i_projid = cpu_to_le32(i_projid);
4392
4393 ext4_inode_csum_set(inode, raw_inode, ei);
4394 return err;
4395}
4396
4397/*
4398 * ext4_get_inode_loc returns with an extra refcount against the inode's
4399 * underlying buffer_head on success. If we pass 'inode' and it does not
4400 * have in-inode xattr, we have all inode data in memory that is needed
4401 * to recreate the on-disk version of this inode.
4402 */
4403static int __ext4_get_inode_loc(struct super_block *sb, unsigned long ino,
4404 struct inode *inode, struct ext4_iloc *iloc,
4405 ext4_fsblk_t *ret_block)
4406{
4407 struct ext4_group_desc *gdp;
4408 struct buffer_head *bh;
4409 ext4_fsblk_t block;
4410 struct blk_plug plug;
4411 int inodes_per_block, inode_offset;
4412
4413 iloc->bh = NULL;
4414 if (ino < EXT4_ROOT_INO ||
4415 ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4416 return -EFSCORRUPTED;
4417
4418 iloc->block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
4419 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4420 if (!gdp)
4421 return -EIO;
4422
4423 /*
4424 * Figure out the offset within the block group inode table
4425 */
4426 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4427 inode_offset = ((ino - 1) %
4428 EXT4_INODES_PER_GROUP(sb));
4429 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4430
4431 block = ext4_inode_table(sb, gdp);
4432 if ((block <= le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) ||
4433 (block >= ext4_blocks_count(EXT4_SB(sb)->s_es))) {
4434 ext4_error(sb, "Invalid inode table block %llu in "
4435 "block_group %u", block, iloc->block_group);
4436 return -EFSCORRUPTED;
4437 }
4438 block += (inode_offset / inodes_per_block);
4439
4440 bh = sb_getblk(sb, block);
4441 if (unlikely(!bh))
4442 return -ENOMEM;
4443 if (ext4_buffer_uptodate(bh))
4444 goto has_buffer;
4445
4446 lock_buffer(bh);
4447 if (ext4_buffer_uptodate(bh)) {
4448 /* Someone brought it uptodate while we waited */
4449 unlock_buffer(bh);
4450 goto has_buffer;
4451 }
4452
4453 /*
4454 * If we have all information of the inode in memory and this
4455 * is the only valid inode in the block, we need not read the
4456 * block.
4457 */
4458 if (inode && !ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4459 struct buffer_head *bitmap_bh;
4460 int i, start;
4461
4462 start = inode_offset & ~(inodes_per_block - 1);
4463
4464 /* Is the inode bitmap in cache? */
4465 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4466 if (unlikely(!bitmap_bh))
4467 goto make_io;
4468
4469 /*
4470 * If the inode bitmap isn't in cache then the
4471 * optimisation may end up performing two reads instead
4472 * of one, so skip it.
4473 */
4474 if (!buffer_uptodate(bitmap_bh)) {
4475 brelse(bitmap_bh);
4476 goto make_io;
4477 }
4478 for (i = start; i < start + inodes_per_block; i++) {
4479 if (i == inode_offset)
4480 continue;
4481 if (ext4_test_bit(i, bitmap_bh->b_data))
4482 break;
4483 }
4484 brelse(bitmap_bh);
4485 if (i == start + inodes_per_block) {
4486 struct ext4_inode *raw_inode =
4487 (struct ext4_inode *) (bh->b_data + iloc->offset);
4488
4489 /* all other inodes are free, so skip I/O */
4490 memset(bh->b_data, 0, bh->b_size);
4491 if (!ext4_test_inode_state(inode, EXT4_STATE_NEW))
4492 ext4_fill_raw_inode(inode, raw_inode);
4493 set_buffer_uptodate(bh);
4494 unlock_buffer(bh);
4495 goto has_buffer;
4496 }
4497 }
4498
4499make_io:
4500 /*
4501 * If we need to do any I/O, try to pre-readahead extra
4502 * blocks from the inode table.
4503 */
4504 blk_start_plug(&plug);
4505 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4506 ext4_fsblk_t b, end, table;
4507 unsigned num;
4508 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4509
4510 table = ext4_inode_table(sb, gdp);
4511 /* s_inode_readahead_blks is always a power of 2 */
4512 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4513 if (table > b)
4514 b = table;
4515 end = b + ra_blks;
4516 num = EXT4_INODES_PER_GROUP(sb);
4517 if (ext4_has_group_desc_csum(sb))
4518 num -= ext4_itable_unused_count(sb, gdp);
4519 table += num / inodes_per_block;
4520 if (end > table)
4521 end = table;
4522 while (b <= end)
4523 ext4_sb_breadahead_unmovable(sb, b++);
4524 }
4525
4526 /*
4527 * There are other valid inodes in the buffer, this inode
4528 * has in-inode xattrs, or we don't have this inode in memory.
4529 * Read the block from disk.
4530 */
4531 trace_ext4_load_inode(sb, ino);
4532 ext4_read_bh_nowait(bh, REQ_META | REQ_PRIO, NULL,
4533 ext4_simulate_fail(sb, EXT4_SIM_INODE_EIO));
4534 blk_finish_plug(&plug);
4535 wait_on_buffer(bh);
4536 if (!buffer_uptodate(bh)) {
4537 if (ret_block)
4538 *ret_block = block;
4539 brelse(bh);
4540 return -EIO;
4541 }
4542has_buffer:
4543 iloc->bh = bh;
4544 return 0;
4545}
4546
4547static int __ext4_get_inode_loc_noinmem(struct inode *inode,
4548 struct ext4_iloc *iloc)
4549{
4550 ext4_fsblk_t err_blk = 0;
4551 int ret;
4552
4553 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, NULL, iloc,
4554 &err_blk);
4555
4556 if (ret == -EIO)
4557 ext4_error_inode_block(inode, err_blk, EIO,
4558 "unable to read itable block");
4559
4560 return ret;
4561}
4562
4563int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4564{
4565 ext4_fsblk_t err_blk = 0;
4566 int ret;
4567
4568 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, inode, iloc,
4569 &err_blk);
4570
4571 if (ret == -EIO)
4572 ext4_error_inode_block(inode, err_blk, EIO,
4573 "unable to read itable block");
4574
4575 return ret;
4576}
4577
4578
4579int ext4_get_fc_inode_loc(struct super_block *sb, unsigned long ino,
4580 struct ext4_iloc *iloc)
4581{
4582 return __ext4_get_inode_loc(sb, ino, NULL, iloc, NULL);
4583}
4584
4585static bool ext4_should_enable_dax(struct inode *inode)
4586{
4587 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4588
4589 if (test_opt2(inode->i_sb, DAX_NEVER))
4590 return false;
4591 if (!S_ISREG(inode->i_mode))
4592 return false;
4593 if (ext4_should_journal_data(inode))
4594 return false;
4595 if (ext4_has_inline_data(inode))
4596 return false;
4597 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4598 return false;
4599 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4600 return false;
4601 if (!test_bit(EXT4_FLAGS_BDEV_IS_DAX, &sbi->s_ext4_flags))
4602 return false;
4603 if (test_opt(inode->i_sb, DAX_ALWAYS))
4604 return true;
4605
4606 return ext4_test_inode_flag(inode, EXT4_INODE_DAX);
4607}
4608
4609void ext4_set_inode_flags(struct inode *inode, bool init)
4610{
4611 unsigned int flags = EXT4_I(inode)->i_flags;
4612 unsigned int new_fl = 0;
4613
4614 WARN_ON_ONCE(IS_DAX(inode) && init);
4615
4616 if (flags & EXT4_SYNC_FL)
4617 new_fl |= S_SYNC;
4618 if (flags & EXT4_APPEND_FL)
4619 new_fl |= S_APPEND;
4620 if (flags & EXT4_IMMUTABLE_FL)
4621 new_fl |= S_IMMUTABLE;
4622 if (flags & EXT4_NOATIME_FL)
4623 new_fl |= S_NOATIME;
4624 if (flags & EXT4_DIRSYNC_FL)
4625 new_fl |= S_DIRSYNC;
4626
4627 /* Because of the way inode_set_flags() works we must preserve S_DAX
4628 * here if already set. */
4629 new_fl |= (inode->i_flags & S_DAX);
4630 if (init && ext4_should_enable_dax(inode))
4631 new_fl |= S_DAX;
4632
4633 if (flags & EXT4_ENCRYPT_FL)
4634 new_fl |= S_ENCRYPTED;
4635 if (flags & EXT4_CASEFOLD_FL)
4636 new_fl |= S_CASEFOLD;
4637 if (flags & EXT4_VERITY_FL)
4638 new_fl |= S_VERITY;
4639 inode_set_flags(inode, new_fl,
4640 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4641 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4642}
4643
4644static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4645 struct ext4_inode_info *ei)
4646{
4647 blkcnt_t i_blocks ;
4648 struct inode *inode = &(ei->vfs_inode);
4649 struct super_block *sb = inode->i_sb;
4650
4651 if (ext4_has_feature_huge_file(sb)) {
4652 /* we are using combined 48 bit field */
4653 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4654 le32_to_cpu(raw_inode->i_blocks_lo);
4655 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4656 /* i_blocks represent file system block size */
4657 return i_blocks << (inode->i_blkbits - 9);
4658 } else {
4659 return i_blocks;
4660 }
4661 } else {
4662 return le32_to_cpu(raw_inode->i_blocks_lo);
4663 }
4664}
4665
4666static inline int ext4_iget_extra_inode(struct inode *inode,
4667 struct ext4_inode *raw_inode,
4668 struct ext4_inode_info *ei)
4669{
4670 __le32 *magic = (void *)raw_inode +
4671 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4672
4673 if (EXT4_INODE_HAS_XATTR_SPACE(inode) &&
4674 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4675 int err;
4676
4677 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4678 err = ext4_find_inline_data_nolock(inode);
4679 if (!err && ext4_has_inline_data(inode))
4680 ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
4681 return err;
4682 } else
4683 EXT4_I(inode)->i_inline_off = 0;
4684 return 0;
4685}
4686
4687int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4688{
4689 if (!ext4_has_feature_project(inode->i_sb))
4690 return -EOPNOTSUPP;
4691 *projid = EXT4_I(inode)->i_projid;
4692 return 0;
4693}
4694
4695/*
4696 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4697 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4698 * set.
4699 */
4700static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4701{
4702 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4703 inode_set_iversion_raw(inode, val);
4704 else
4705 inode_set_iversion_queried(inode, val);
4706}
4707
4708static const char *check_igot_inode(struct inode *inode, ext4_iget_flags flags)
4709
4710{
4711 if (flags & EXT4_IGET_EA_INODE) {
4712 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4713 return "missing EA_INODE flag";
4714 if (ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4715 EXT4_I(inode)->i_file_acl)
4716 return "ea_inode with extended attributes";
4717 } else {
4718 if ((EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4719 return "unexpected EA_INODE flag";
4720 }
4721 if (is_bad_inode(inode) && !(flags & EXT4_IGET_BAD))
4722 return "unexpected bad inode w/o EXT4_IGET_BAD";
4723 return NULL;
4724}
4725
4726struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4727 ext4_iget_flags flags, const char *function,
4728 unsigned int line)
4729{
4730 struct ext4_iloc iloc;
4731 struct ext4_inode *raw_inode;
4732 struct ext4_inode_info *ei;
4733 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
4734 struct inode *inode;
4735 const char *err_str;
4736 journal_t *journal = EXT4_SB(sb)->s_journal;
4737 long ret;
4738 loff_t size;
4739 int block;
4740 uid_t i_uid;
4741 gid_t i_gid;
4742 projid_t i_projid;
4743
4744 if ((!(flags & EXT4_IGET_SPECIAL) &&
4745 ((ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO) ||
4746 ino == le32_to_cpu(es->s_usr_quota_inum) ||
4747 ino == le32_to_cpu(es->s_grp_quota_inum) ||
4748 ino == le32_to_cpu(es->s_prj_quota_inum) ||
4749 ino == le32_to_cpu(es->s_orphan_file_inum))) ||
4750 (ino < EXT4_ROOT_INO) ||
4751 (ino > le32_to_cpu(es->s_inodes_count))) {
4752 if (flags & EXT4_IGET_HANDLE)
4753 return ERR_PTR(-ESTALE);
4754 __ext4_error(sb, function, line, false, EFSCORRUPTED, 0,
4755 "inode #%lu: comm %s: iget: illegal inode #",
4756 ino, current->comm);
4757 return ERR_PTR(-EFSCORRUPTED);
4758 }
4759
4760 inode = iget_locked(sb, ino);
4761 if (!inode)
4762 return ERR_PTR(-ENOMEM);
4763 if (!(inode->i_state & I_NEW)) {
4764 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
4765 ext4_error_inode(inode, function, line, 0, err_str);
4766 iput(inode);
4767 return ERR_PTR(-EFSCORRUPTED);
4768 }
4769 return inode;
4770 }
4771
4772 ei = EXT4_I(inode);
4773 iloc.bh = NULL;
4774
4775 ret = __ext4_get_inode_loc_noinmem(inode, &iloc);
4776 if (ret < 0)
4777 goto bad_inode;
4778 raw_inode = ext4_raw_inode(&iloc);
4779
4780 if ((flags & EXT4_IGET_HANDLE) &&
4781 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4782 ret = -ESTALE;
4783 goto bad_inode;
4784 }
4785
4786 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4787 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4788 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4789 EXT4_INODE_SIZE(inode->i_sb) ||
4790 (ei->i_extra_isize & 3)) {
4791 ext4_error_inode(inode, function, line, 0,
4792 "iget: bad extra_isize %u "
4793 "(inode size %u)",
4794 ei->i_extra_isize,
4795 EXT4_INODE_SIZE(inode->i_sb));
4796 ret = -EFSCORRUPTED;
4797 goto bad_inode;
4798 }
4799 } else
4800 ei->i_extra_isize = 0;
4801
4802 /* Precompute checksum seed for inode metadata */
4803 if (ext4_has_metadata_csum(sb)) {
4804 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4805 __u32 csum;
4806 __le32 inum = cpu_to_le32(inode->i_ino);
4807 __le32 gen = raw_inode->i_generation;
4808 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4809 sizeof(inum));
4810 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4811 sizeof(gen));
4812 }
4813
4814 if ((!ext4_inode_csum_verify(inode, raw_inode, ei) ||
4815 ext4_simulate_fail(sb, EXT4_SIM_INODE_CRC)) &&
4816 (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))) {
4817 ext4_error_inode_err(inode, function, line, 0,
4818 EFSBADCRC, "iget: checksum invalid");
4819 ret = -EFSBADCRC;
4820 goto bad_inode;
4821 }
4822
4823 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4824 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4825 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4826 if (ext4_has_feature_project(sb) &&
4827 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4828 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4829 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4830 else
4831 i_projid = EXT4_DEF_PROJID;
4832
4833 if (!(test_opt(inode->i_sb, NO_UID32))) {
4834 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4835 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4836 }
4837 i_uid_write(inode, i_uid);
4838 i_gid_write(inode, i_gid);
4839 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4840 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4841
4842 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4843 ei->i_inline_off = 0;
4844 ei->i_dir_start_lookup = 0;
4845 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4846 /* We now have enough fields to check if the inode was active or not.
4847 * This is needed because nfsd might try to access dead inodes
4848 * the test is that same one that e2fsck uses
4849 * NeilBrown 1999oct15
4850 */
4851 if (inode->i_nlink == 0) {
4852 if ((inode->i_mode == 0 || flags & EXT4_IGET_SPECIAL ||
4853 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4854 ino != EXT4_BOOT_LOADER_INO) {
4855 /* this inode is deleted or unallocated */
4856 if (flags & EXT4_IGET_SPECIAL) {
4857 ext4_error_inode(inode, function, line, 0,
4858 "iget: special inode unallocated");
4859 ret = -EFSCORRUPTED;
4860 } else
4861 ret = -ESTALE;
4862 goto bad_inode;
4863 }
4864 /* The only unlinked inodes we let through here have
4865 * valid i_mode and are being read by the orphan
4866 * recovery code: that's fine, we're about to complete
4867 * the process of deleting those.
4868 * OR it is the EXT4_BOOT_LOADER_INO which is
4869 * not initialized on a new filesystem. */
4870 }
4871 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4872 ext4_set_inode_flags(inode, true);
4873 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4874 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4875 if (ext4_has_feature_64bit(sb))
4876 ei->i_file_acl |=
4877 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4878 inode->i_size = ext4_isize(sb, raw_inode);
4879 if ((size = i_size_read(inode)) < 0) {
4880 ext4_error_inode(inode, function, line, 0,
4881 "iget: bad i_size value: %lld", size);
4882 ret = -EFSCORRUPTED;
4883 goto bad_inode;
4884 }
4885 /*
4886 * If dir_index is not enabled but there's dir with INDEX flag set,
4887 * we'd normally treat htree data as empty space. But with metadata
4888 * checksumming that corrupts checksums so forbid that.
4889 */
4890 if (!ext4_has_feature_dir_index(sb) && ext4_has_metadata_csum(sb) &&
4891 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
4892 ext4_error_inode(inode, function, line, 0,
4893 "iget: Dir with htree data on filesystem without dir_index feature.");
4894 ret = -EFSCORRUPTED;
4895 goto bad_inode;
4896 }
4897 ei->i_disksize = inode->i_size;
4898#ifdef CONFIG_QUOTA
4899 ei->i_reserved_quota = 0;
4900#endif
4901 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4902 ei->i_block_group = iloc.block_group;
4903 ei->i_last_alloc_group = ~0;
4904 /*
4905 * NOTE! The in-memory inode i_data array is in little-endian order
4906 * even on big-endian machines: we do NOT byteswap the block numbers!
4907 */
4908 for (block = 0; block < EXT4_N_BLOCKS; block++)
4909 ei->i_data[block] = raw_inode->i_block[block];
4910 INIT_LIST_HEAD(&ei->i_orphan);
4911 ext4_fc_init_inode(&ei->vfs_inode);
4912
4913 /*
4914 * Set transaction id's of transactions that have to be committed
4915 * to finish f[data]sync. We set them to currently running transaction
4916 * as we cannot be sure that the inode or some of its metadata isn't
4917 * part of the transaction - the inode could have been reclaimed and
4918 * now it is reread from disk.
4919 */
4920 if (journal) {
4921 transaction_t *transaction;
4922 tid_t tid;
4923
4924 read_lock(&journal->j_state_lock);
4925 if (journal->j_running_transaction)
4926 transaction = journal->j_running_transaction;
4927 else
4928 transaction = journal->j_committing_transaction;
4929 if (transaction)
4930 tid = transaction->t_tid;
4931 else
4932 tid = journal->j_commit_sequence;
4933 read_unlock(&journal->j_state_lock);
4934 ei->i_sync_tid = tid;
4935 ei->i_datasync_tid = tid;
4936 }
4937
4938 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4939 if (ei->i_extra_isize == 0) {
4940 /* The extra space is currently unused. Use it. */
4941 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4942 ei->i_extra_isize = sizeof(struct ext4_inode) -
4943 EXT4_GOOD_OLD_INODE_SIZE;
4944 } else {
4945 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
4946 if (ret)
4947 goto bad_inode;
4948 }
4949 }
4950
4951 EXT4_INODE_GET_CTIME(inode, raw_inode);
4952 EXT4_INODE_GET_ATIME(inode, raw_inode);
4953 EXT4_INODE_GET_MTIME(inode, raw_inode);
4954 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4955
4956 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4957 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4958
4959 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4960 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4961 ivers |=
4962 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4963 }
4964 ext4_inode_set_iversion_queried(inode, ivers);
4965 }
4966
4967 ret = 0;
4968 if (ei->i_file_acl &&
4969 !ext4_inode_block_valid(inode, ei->i_file_acl, 1)) {
4970 ext4_error_inode(inode, function, line, 0,
4971 "iget: bad extended attribute block %llu",
4972 ei->i_file_acl);
4973 ret = -EFSCORRUPTED;
4974 goto bad_inode;
4975 } else if (!ext4_has_inline_data(inode)) {
4976 /* validate the block references in the inode */
4977 if (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) &&
4978 (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4979 (S_ISLNK(inode->i_mode) &&
4980 !ext4_inode_is_fast_symlink(inode)))) {
4981 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4982 ret = ext4_ext_check_inode(inode);
4983 else
4984 ret = ext4_ind_check_inode(inode);
4985 }
4986 }
4987 if (ret)
4988 goto bad_inode;
4989
4990 if (S_ISREG(inode->i_mode)) {
4991 inode->i_op = &ext4_file_inode_operations;
4992 inode->i_fop = &ext4_file_operations;
4993 ext4_set_aops(inode);
4994 } else if (S_ISDIR(inode->i_mode)) {
4995 inode->i_op = &ext4_dir_inode_operations;
4996 inode->i_fop = &ext4_dir_operations;
4997 } else if (S_ISLNK(inode->i_mode)) {
4998 /* VFS does not allow setting these so must be corruption */
4999 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
5000 ext4_error_inode(inode, function, line, 0,
5001 "iget: immutable or append flags "
5002 "not allowed on symlinks");
5003 ret = -EFSCORRUPTED;
5004 goto bad_inode;
5005 }
5006 if (IS_ENCRYPTED(inode)) {
5007 inode->i_op = &ext4_encrypted_symlink_inode_operations;
5008 } else if (ext4_inode_is_fast_symlink(inode)) {
5009 inode->i_link = (char *)ei->i_data;
5010 inode->i_op = &ext4_fast_symlink_inode_operations;
5011 nd_terminate_link(ei->i_data, inode->i_size,
5012 sizeof(ei->i_data) - 1);
5013 } else {
5014 inode->i_op = &ext4_symlink_inode_operations;
5015 }
5016 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5017 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5018 inode->i_op = &ext4_special_inode_operations;
5019 if (raw_inode->i_block[0])
5020 init_special_inode(inode, inode->i_mode,
5021 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5022 else
5023 init_special_inode(inode, inode->i_mode,
5024 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5025 } else if (ino == EXT4_BOOT_LOADER_INO) {
5026 make_bad_inode(inode);
5027 } else {
5028 ret = -EFSCORRUPTED;
5029 ext4_error_inode(inode, function, line, 0,
5030 "iget: bogus i_mode (%o)", inode->i_mode);
5031 goto bad_inode;
5032 }
5033 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb)) {
5034 ext4_error_inode(inode, function, line, 0,
5035 "casefold flag without casefold feature");
5036 ret = -EFSCORRUPTED;
5037 goto bad_inode;
5038 }
5039 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
5040 ext4_error_inode(inode, function, line, 0, err_str);
5041 ret = -EFSCORRUPTED;
5042 goto bad_inode;
5043 }
5044
5045 brelse(iloc.bh);
5046 unlock_new_inode(inode);
5047 return inode;
5048
5049bad_inode:
5050 brelse(iloc.bh);
5051 iget_failed(inode);
5052 return ERR_PTR(ret);
5053}
5054
5055static void __ext4_update_other_inode_time(struct super_block *sb,
5056 unsigned long orig_ino,
5057 unsigned long ino,
5058 struct ext4_inode *raw_inode)
5059{
5060 struct inode *inode;
5061
5062 inode = find_inode_by_ino_rcu(sb, ino);
5063 if (!inode)
5064 return;
5065
5066 if (!inode_is_dirtytime_only(inode))
5067 return;
5068
5069 spin_lock(&inode->i_lock);
5070 if (inode_is_dirtytime_only(inode)) {
5071 struct ext4_inode_info *ei = EXT4_I(inode);
5072
5073 inode->i_state &= ~I_DIRTY_TIME;
5074 spin_unlock(&inode->i_lock);
5075
5076 spin_lock(&ei->i_raw_lock);
5077 EXT4_INODE_SET_CTIME(inode, raw_inode);
5078 EXT4_INODE_SET_MTIME(inode, raw_inode);
5079 EXT4_INODE_SET_ATIME(inode, raw_inode);
5080 ext4_inode_csum_set(inode, raw_inode, ei);
5081 spin_unlock(&ei->i_raw_lock);
5082 trace_ext4_other_inode_update_time(inode, orig_ino);
5083 return;
5084 }
5085 spin_unlock(&inode->i_lock);
5086}
5087
5088/*
5089 * Opportunistically update the other time fields for other inodes in
5090 * the same inode table block.
5091 */
5092static void ext4_update_other_inodes_time(struct super_block *sb,
5093 unsigned long orig_ino, char *buf)
5094{
5095 unsigned long ino;
5096 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5097 int inode_size = EXT4_INODE_SIZE(sb);
5098
5099 /*
5100 * Calculate the first inode in the inode table block. Inode
5101 * numbers are one-based. That is, the first inode in a block
5102 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5103 */
5104 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5105 rcu_read_lock();
5106 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5107 if (ino == orig_ino)
5108 continue;
5109 __ext4_update_other_inode_time(sb, orig_ino, ino,
5110 (struct ext4_inode *)buf);
5111 }
5112 rcu_read_unlock();
5113}
5114
5115/*
5116 * Post the struct inode info into an on-disk inode location in the
5117 * buffer-cache. This gobbles the caller's reference to the
5118 * buffer_head in the inode location struct.
5119 *
5120 * The caller must have write access to iloc->bh.
5121 */
5122static int ext4_do_update_inode(handle_t *handle,
5123 struct inode *inode,
5124 struct ext4_iloc *iloc)
5125{
5126 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5127 struct ext4_inode_info *ei = EXT4_I(inode);
5128 struct buffer_head *bh = iloc->bh;
5129 struct super_block *sb = inode->i_sb;
5130 int err;
5131 int need_datasync = 0, set_large_file = 0;
5132
5133 spin_lock(&ei->i_raw_lock);
5134
5135 /*
5136 * For fields not tracked in the in-memory inode, initialise them
5137 * to zero for new inodes.
5138 */
5139 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5140 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5141
5142 if (READ_ONCE(ei->i_disksize) != ext4_isize(inode->i_sb, raw_inode))
5143 need_datasync = 1;
5144 if (ei->i_disksize > 0x7fffffffULL) {
5145 if (!ext4_has_feature_large_file(sb) ||
5146 EXT4_SB(sb)->s_es->s_rev_level == cpu_to_le32(EXT4_GOOD_OLD_REV))
5147 set_large_file = 1;
5148 }
5149
5150 err = ext4_fill_raw_inode(inode, raw_inode);
5151 spin_unlock(&ei->i_raw_lock);
5152 if (err) {
5153 EXT4_ERROR_INODE(inode, "corrupted inode contents");
5154 goto out_brelse;
5155 }
5156
5157 if (inode->i_sb->s_flags & SB_LAZYTIME)
5158 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5159 bh->b_data);
5160
5161 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5162 err = ext4_handle_dirty_metadata(handle, NULL, bh);
5163 if (err)
5164 goto out_error;
5165 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5166 if (set_large_file) {
5167 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5168 err = ext4_journal_get_write_access(handle, sb,
5169 EXT4_SB(sb)->s_sbh,
5170 EXT4_JTR_NONE);
5171 if (err)
5172 goto out_error;
5173 lock_buffer(EXT4_SB(sb)->s_sbh);
5174 ext4_set_feature_large_file(sb);
5175 ext4_superblock_csum_set(sb);
5176 unlock_buffer(EXT4_SB(sb)->s_sbh);
5177 ext4_handle_sync(handle);
5178 err = ext4_handle_dirty_metadata(handle, NULL,
5179 EXT4_SB(sb)->s_sbh);
5180 }
5181 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5182out_error:
5183 ext4_std_error(inode->i_sb, err);
5184out_brelse:
5185 brelse(bh);
5186 return err;
5187}
5188
5189/*
5190 * ext4_write_inode()
5191 *
5192 * We are called from a few places:
5193 *
5194 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5195 * Here, there will be no transaction running. We wait for any running
5196 * transaction to commit.
5197 *
5198 * - Within flush work (sys_sync(), kupdate and such).
5199 * We wait on commit, if told to.
5200 *
5201 * - Within iput_final() -> write_inode_now()
5202 * We wait on commit, if told to.
5203 *
5204 * In all cases it is actually safe for us to return without doing anything,
5205 * because the inode has been copied into a raw inode buffer in
5206 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5207 * writeback.
5208 *
5209 * Note that we are absolutely dependent upon all inode dirtiers doing the
5210 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5211 * which we are interested.
5212 *
5213 * It would be a bug for them to not do this. The code:
5214 *
5215 * mark_inode_dirty(inode)
5216 * stuff();
5217 * inode->i_size = expr;
5218 *
5219 * is in error because write_inode() could occur while `stuff()' is running,
5220 * and the new i_size will be lost. Plus the inode will no longer be on the
5221 * superblock's dirty inode list.
5222 */
5223int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5224{
5225 int err;
5226
5227 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5228 return 0;
5229
5230 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5231 return -EIO;
5232
5233 if (EXT4_SB(inode->i_sb)->s_journal) {
5234 if (ext4_journal_current_handle()) {
5235 ext4_debug("called recursively, non-PF_MEMALLOC!\n");
5236 dump_stack();
5237 return -EIO;
5238 }
5239
5240 /*
5241 * No need to force transaction in WB_SYNC_NONE mode. Also
5242 * ext4_sync_fs() will force the commit after everything is
5243 * written.
5244 */
5245 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5246 return 0;
5247
5248 err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
5249 EXT4_I(inode)->i_sync_tid);
5250 } else {
5251 struct ext4_iloc iloc;
5252
5253 err = __ext4_get_inode_loc_noinmem(inode, &iloc);
5254 if (err)
5255 return err;
5256 /*
5257 * sync(2) will flush the whole buffer cache. No need to do
5258 * it here separately for each inode.
5259 */
5260 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5261 sync_dirty_buffer(iloc.bh);
5262 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5263 ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
5264 "IO error syncing inode");
5265 err = -EIO;
5266 }
5267 brelse(iloc.bh);
5268 }
5269 return err;
5270}
5271
5272/*
5273 * In data=journal mode ext4_journalled_invalidate_folio() may fail to invalidate
5274 * buffers that are attached to a folio straddling i_size and are undergoing
5275 * commit. In that case we have to wait for commit to finish and try again.
5276 */
5277static void ext4_wait_for_tail_page_commit(struct inode *inode)
5278{
5279 unsigned offset;
5280 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5281 tid_t commit_tid;
5282 int ret;
5283 bool has_transaction;
5284
5285 offset = inode->i_size & (PAGE_SIZE - 1);
5286 /*
5287 * If the folio is fully truncated, we don't need to wait for any commit
5288 * (and we even should not as __ext4_journalled_invalidate_folio() may
5289 * strip all buffers from the folio but keep the folio dirty which can then
5290 * confuse e.g. concurrent ext4_writepages() seeing dirty folio without
5291 * buffers). Also we don't need to wait for any commit if all buffers in
5292 * the folio remain valid. This is most beneficial for the common case of
5293 * blocksize == PAGESIZE.
5294 */
5295 if (!offset || offset > (PAGE_SIZE - i_blocksize(inode)))
5296 return;
5297 while (1) {
5298 struct folio *folio = filemap_lock_folio(inode->i_mapping,
5299 inode->i_size >> PAGE_SHIFT);
5300 if (IS_ERR(folio))
5301 return;
5302 ret = __ext4_journalled_invalidate_folio(folio, offset,
5303 folio_size(folio) - offset);
5304 folio_unlock(folio);
5305 folio_put(folio);
5306 if (ret != -EBUSY)
5307 return;
5308 has_transaction = false;
5309 read_lock(&journal->j_state_lock);
5310 if (journal->j_committing_transaction) {
5311 commit_tid = journal->j_committing_transaction->t_tid;
5312 has_transaction = true;
5313 }
5314 read_unlock(&journal->j_state_lock);
5315 if (has_transaction)
5316 jbd2_log_wait_commit(journal, commit_tid);
5317 }
5318}
5319
5320/*
5321 * ext4_setattr()
5322 *
5323 * Called from notify_change.
5324 *
5325 * We want to trap VFS attempts to truncate the file as soon as
5326 * possible. In particular, we want to make sure that when the VFS
5327 * shrinks i_size, we put the inode on the orphan list and modify
5328 * i_disksize immediately, so that during the subsequent flushing of
5329 * dirty pages and freeing of disk blocks, we can guarantee that any
5330 * commit will leave the blocks being flushed in an unused state on
5331 * disk. (On recovery, the inode will get truncated and the blocks will
5332 * be freed, so we have a strong guarantee that no future commit will
5333 * leave these blocks visible to the user.)
5334 *
5335 * Another thing we have to assure is that if we are in ordered mode
5336 * and inode is still attached to the committing transaction, we must
5337 * we start writeout of all the dirty pages which are being truncated.
5338 * This way we are sure that all the data written in the previous
5339 * transaction are already on disk (truncate waits for pages under
5340 * writeback).
5341 *
5342 * Called with inode->i_rwsem down.
5343 */
5344int ext4_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
5345 struct iattr *attr)
5346{
5347 struct inode *inode = d_inode(dentry);
5348 int error, rc = 0;
5349 int orphan = 0;
5350 const unsigned int ia_valid = attr->ia_valid;
5351 bool inc_ivers = true;
5352
5353 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5354 return -EIO;
5355
5356 if (unlikely(IS_IMMUTABLE(inode)))
5357 return -EPERM;
5358
5359 if (unlikely(IS_APPEND(inode) &&
5360 (ia_valid & (ATTR_MODE | ATTR_UID |
5361 ATTR_GID | ATTR_TIMES_SET))))
5362 return -EPERM;
5363
5364 error = setattr_prepare(idmap, dentry, attr);
5365 if (error)
5366 return error;
5367
5368 error = fscrypt_prepare_setattr(dentry, attr);
5369 if (error)
5370 return error;
5371
5372 error = fsverity_prepare_setattr(dentry, attr);
5373 if (error)
5374 return error;
5375
5376 if (is_quota_modification(idmap, inode, attr)) {
5377 error = dquot_initialize(inode);
5378 if (error)
5379 return error;
5380 }
5381
5382 if (i_uid_needs_update(idmap, attr, inode) ||
5383 i_gid_needs_update(idmap, attr, inode)) {
5384 handle_t *handle;
5385
5386 /* (user+group)*(old+new) structure, inode write (sb,
5387 * inode block, ? - but truncate inode update has it) */
5388 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5389 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5390 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5391 if (IS_ERR(handle)) {
5392 error = PTR_ERR(handle);
5393 goto err_out;
5394 }
5395
5396 /* dquot_transfer() calls back ext4_get_inode_usage() which
5397 * counts xattr inode references.
5398 */
5399 down_read(&EXT4_I(inode)->xattr_sem);
5400 error = dquot_transfer(idmap, inode, attr);
5401 up_read(&EXT4_I(inode)->xattr_sem);
5402
5403 if (error) {
5404 ext4_journal_stop(handle);
5405 return error;
5406 }
5407 /* Update corresponding info in inode so that everything is in
5408 * one transaction */
5409 i_uid_update(idmap, attr, inode);
5410 i_gid_update(idmap, attr, inode);
5411 error = ext4_mark_inode_dirty(handle, inode);
5412 ext4_journal_stop(handle);
5413 if (unlikely(error)) {
5414 return error;
5415 }
5416 }
5417
5418 if (attr->ia_valid & ATTR_SIZE) {
5419 handle_t *handle;
5420 loff_t oldsize = inode->i_size;
5421 loff_t old_disksize;
5422 int shrink = (attr->ia_size < inode->i_size);
5423
5424 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5425 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5426
5427 if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5428 return -EFBIG;
5429 }
5430 }
5431 if (!S_ISREG(inode->i_mode)) {
5432 return -EINVAL;
5433 }
5434
5435 if (attr->ia_size == inode->i_size)
5436 inc_ivers = false;
5437
5438 if (shrink) {
5439 if (ext4_should_order_data(inode)) {
5440 error = ext4_begin_ordered_truncate(inode,
5441 attr->ia_size);
5442 if (error)
5443 goto err_out;
5444 }
5445 /*
5446 * Blocks are going to be removed from the inode. Wait
5447 * for dio in flight.
5448 */
5449 inode_dio_wait(inode);
5450 }
5451
5452 filemap_invalidate_lock(inode->i_mapping);
5453
5454 rc = ext4_break_layouts(inode);
5455 if (rc) {
5456 filemap_invalidate_unlock(inode->i_mapping);
5457 goto err_out;
5458 }
5459
5460 if (attr->ia_size != inode->i_size) {
5461 /* attach jbd2 jinode for EOF folio tail zeroing */
5462 if (attr->ia_size & (inode->i_sb->s_blocksize - 1) ||
5463 oldsize & (inode->i_sb->s_blocksize - 1)) {
5464 error = ext4_inode_attach_jinode(inode);
5465 if (error)
5466 goto err_out;
5467 }
5468
5469 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5470 if (IS_ERR(handle)) {
5471 error = PTR_ERR(handle);
5472 goto out_mmap_sem;
5473 }
5474 if (ext4_handle_valid(handle) && shrink) {
5475 error = ext4_orphan_add(handle, inode);
5476 orphan = 1;
5477 }
5478 /*
5479 * Update c/mtime and tail zero the EOF folio on
5480 * truncate up. ext4_truncate() handles the shrink case
5481 * below.
5482 */
5483 if (!shrink) {
5484 inode_set_mtime_to_ts(inode,
5485 inode_set_ctime_current(inode));
5486 if (oldsize & (inode->i_sb->s_blocksize - 1))
5487 ext4_block_truncate_page(handle,
5488 inode->i_mapping, oldsize);
5489 }
5490
5491 if (shrink)
5492 ext4_fc_track_range(handle, inode,
5493 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5494 inode->i_sb->s_blocksize_bits,
5495 EXT_MAX_BLOCKS - 1);
5496 else
5497 ext4_fc_track_range(
5498 handle, inode,
5499 (oldsize > 0 ? oldsize - 1 : oldsize) >>
5500 inode->i_sb->s_blocksize_bits,
5501 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5502 inode->i_sb->s_blocksize_bits);
5503
5504 down_write(&EXT4_I(inode)->i_data_sem);
5505 old_disksize = EXT4_I(inode)->i_disksize;
5506 EXT4_I(inode)->i_disksize = attr->ia_size;
5507 rc = ext4_mark_inode_dirty(handle, inode);
5508 if (!error)
5509 error = rc;
5510 /*
5511 * We have to update i_size under i_data_sem together
5512 * with i_disksize to avoid races with writeback code
5513 * running ext4_wb_update_i_disksize().
5514 */
5515 if (!error)
5516 i_size_write(inode, attr->ia_size);
5517 else
5518 EXT4_I(inode)->i_disksize = old_disksize;
5519 up_write(&EXT4_I(inode)->i_data_sem);
5520 ext4_journal_stop(handle);
5521 if (error)
5522 goto out_mmap_sem;
5523 if (!shrink) {
5524 pagecache_isize_extended(inode, oldsize,
5525 inode->i_size);
5526 } else if (ext4_should_journal_data(inode)) {
5527 ext4_wait_for_tail_page_commit(inode);
5528 }
5529 }
5530
5531 /*
5532 * Truncate pagecache after we've waited for commit
5533 * in data=journal mode to make pages freeable.
5534 */
5535 truncate_pagecache(inode, inode->i_size);
5536 /*
5537 * Call ext4_truncate() even if i_size didn't change to
5538 * truncate possible preallocated blocks.
5539 */
5540 if (attr->ia_size <= oldsize) {
5541 rc = ext4_truncate(inode);
5542 if (rc)
5543 error = rc;
5544 }
5545out_mmap_sem:
5546 filemap_invalidate_unlock(inode->i_mapping);
5547 }
5548
5549 if (!error) {
5550 if (inc_ivers)
5551 inode_inc_iversion(inode);
5552 setattr_copy(idmap, inode, attr);
5553 mark_inode_dirty(inode);
5554 }
5555
5556 /*
5557 * If the call to ext4_truncate failed to get a transaction handle at
5558 * all, we need to clean up the in-core orphan list manually.
5559 */
5560 if (orphan && inode->i_nlink)
5561 ext4_orphan_del(NULL, inode);
5562
5563 if (!error && (ia_valid & ATTR_MODE))
5564 rc = posix_acl_chmod(idmap, dentry, inode->i_mode);
5565
5566err_out:
5567 if (error)
5568 ext4_std_error(inode->i_sb, error);
5569 if (!error)
5570 error = rc;
5571 return error;
5572}
5573
5574u32 ext4_dio_alignment(struct inode *inode)
5575{
5576 if (fsverity_active(inode))
5577 return 0;
5578 if (ext4_should_journal_data(inode))
5579 return 0;
5580 if (ext4_has_inline_data(inode))
5581 return 0;
5582 if (IS_ENCRYPTED(inode)) {
5583 if (!fscrypt_dio_supported(inode))
5584 return 0;
5585 return i_blocksize(inode);
5586 }
5587 return 1; /* use the iomap defaults */
5588}
5589
5590int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
5591 struct kstat *stat, u32 request_mask, unsigned int query_flags)
5592{
5593 struct inode *inode = d_inode(path->dentry);
5594 struct ext4_inode *raw_inode;
5595 struct ext4_inode_info *ei = EXT4_I(inode);
5596 unsigned int flags;
5597
5598 if ((request_mask & STATX_BTIME) &&
5599 EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5600 stat->result_mask |= STATX_BTIME;
5601 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5602 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5603 }
5604
5605 /*
5606 * Return the DIO alignment restrictions if requested. We only return
5607 * this information when requested, since on encrypted files it might
5608 * take a fair bit of work to get if the file wasn't opened recently.
5609 */
5610 if ((request_mask & STATX_DIOALIGN) && S_ISREG(inode->i_mode)) {
5611 u32 dio_align = ext4_dio_alignment(inode);
5612
5613 stat->result_mask |= STATX_DIOALIGN;
5614 if (dio_align == 1) {
5615 struct block_device *bdev = inode->i_sb->s_bdev;
5616
5617 /* iomap defaults */
5618 stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
5619 stat->dio_offset_align = bdev_logical_block_size(bdev);
5620 } else {
5621 stat->dio_mem_align = dio_align;
5622 stat->dio_offset_align = dio_align;
5623 }
5624 }
5625
5626 if ((request_mask & STATX_WRITE_ATOMIC) && S_ISREG(inode->i_mode)) {
5627 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5628 unsigned int awu_min = 0, awu_max = 0;
5629
5630 if (ext4_inode_can_atomic_write(inode)) {
5631 awu_min = sbi->s_awu_min;
5632 awu_max = sbi->s_awu_max;
5633 }
5634
5635 generic_fill_statx_atomic_writes(stat, awu_min, awu_max);
5636 }
5637
5638 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5639 if (flags & EXT4_APPEND_FL)
5640 stat->attributes |= STATX_ATTR_APPEND;
5641 if (flags & EXT4_COMPR_FL)
5642 stat->attributes |= STATX_ATTR_COMPRESSED;
5643 if (flags & EXT4_ENCRYPT_FL)
5644 stat->attributes |= STATX_ATTR_ENCRYPTED;
5645 if (flags & EXT4_IMMUTABLE_FL)
5646 stat->attributes |= STATX_ATTR_IMMUTABLE;
5647 if (flags & EXT4_NODUMP_FL)
5648 stat->attributes |= STATX_ATTR_NODUMP;
5649 if (flags & EXT4_VERITY_FL)
5650 stat->attributes |= STATX_ATTR_VERITY;
5651
5652 stat->attributes_mask |= (STATX_ATTR_APPEND |
5653 STATX_ATTR_COMPRESSED |
5654 STATX_ATTR_ENCRYPTED |
5655 STATX_ATTR_IMMUTABLE |
5656 STATX_ATTR_NODUMP |
5657 STATX_ATTR_VERITY);
5658
5659 generic_fillattr(idmap, request_mask, inode, stat);
5660 return 0;
5661}
5662
5663int ext4_file_getattr(struct mnt_idmap *idmap,
5664 const struct path *path, struct kstat *stat,
5665 u32 request_mask, unsigned int query_flags)
5666{
5667 struct inode *inode = d_inode(path->dentry);
5668 u64 delalloc_blocks;
5669
5670 ext4_getattr(idmap, path, stat, request_mask, query_flags);
5671
5672 /*
5673 * If there is inline data in the inode, the inode will normally not
5674 * have data blocks allocated (it may have an external xattr block).
5675 * Report at least one sector for such files, so tools like tar, rsync,
5676 * others don't incorrectly think the file is completely sparse.
5677 */
5678 if (unlikely(ext4_has_inline_data(inode)))
5679 stat->blocks += (stat->size + 511) >> 9;
5680
5681 /*
5682 * We can't update i_blocks if the block allocation is delayed
5683 * otherwise in the case of system crash before the real block
5684 * allocation is done, we will have i_blocks inconsistent with
5685 * on-disk file blocks.
5686 * We always keep i_blocks updated together with real
5687 * allocation. But to not confuse with user, stat
5688 * will return the blocks that include the delayed allocation
5689 * blocks for this file.
5690 */
5691 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5692 EXT4_I(inode)->i_reserved_data_blocks);
5693 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5694 return 0;
5695}
5696
5697static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5698 int pextents)
5699{
5700 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5701 return ext4_ind_trans_blocks(inode, lblocks);
5702 return ext4_ext_index_trans_blocks(inode, pextents);
5703}
5704
5705/*
5706 * Account for index blocks, block groups bitmaps and block group
5707 * descriptor blocks if modify datablocks and index blocks
5708 * worse case, the indexs blocks spread over different block groups
5709 *
5710 * If datablocks are discontiguous, they are possible to spread over
5711 * different block groups too. If they are contiguous, with flexbg,
5712 * they could still across block group boundary.
5713 *
5714 * Also account for superblock, inode, quota and xattr blocks
5715 */
5716static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5717 int pextents)
5718{
5719 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5720 int gdpblocks;
5721 int idxblocks;
5722 int ret;
5723
5724 /*
5725 * How many index blocks need to touch to map @lblocks logical blocks
5726 * to @pextents physical extents?
5727 */
5728 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5729
5730 ret = idxblocks;
5731
5732 /*
5733 * Now let's see how many group bitmaps and group descriptors need
5734 * to account
5735 */
5736 groups = idxblocks + pextents;
5737 gdpblocks = groups;
5738 if (groups > ngroups)
5739 groups = ngroups;
5740 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5741 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5742
5743 /* bitmaps and block group descriptor blocks */
5744 ret += groups + gdpblocks;
5745
5746 /* Blocks for super block, inode, quota and xattr blocks */
5747 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5748
5749 return ret;
5750}
5751
5752/*
5753 * Calculate the total number of credits to reserve to fit
5754 * the modification of a single pages into a single transaction,
5755 * which may include multiple chunks of block allocations.
5756 *
5757 * This could be called via ext4_write_begin()
5758 *
5759 * We need to consider the worse case, when
5760 * one new block per extent.
5761 */
5762int ext4_writepage_trans_blocks(struct inode *inode)
5763{
5764 int bpp = ext4_journal_blocks_per_page(inode);
5765 int ret;
5766
5767 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5768
5769 /* Account for data blocks for journalled mode */
5770 if (ext4_should_journal_data(inode))
5771 ret += bpp;
5772 return ret;
5773}
5774
5775/*
5776 * Calculate the journal credits for a chunk of data modification.
5777 *
5778 * This is called from DIO, fallocate or whoever calling
5779 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5780 *
5781 * journal buffers for data blocks are not included here, as DIO
5782 * and fallocate do no need to journal data buffers.
5783 */
5784int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5785{
5786 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5787}
5788
5789/*
5790 * The caller must have previously called ext4_reserve_inode_write().
5791 * Give this, we know that the caller already has write access to iloc->bh.
5792 */
5793int ext4_mark_iloc_dirty(handle_t *handle,
5794 struct inode *inode, struct ext4_iloc *iloc)
5795{
5796 int err = 0;
5797
5798 if (unlikely(ext4_forced_shutdown(inode->i_sb))) {
5799 put_bh(iloc->bh);
5800 return -EIO;
5801 }
5802 ext4_fc_track_inode(handle, inode);
5803
5804 /* the do_update_inode consumes one bh->b_count */
5805 get_bh(iloc->bh);
5806
5807 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5808 err = ext4_do_update_inode(handle, inode, iloc);
5809 put_bh(iloc->bh);
5810 return err;
5811}
5812
5813/*
5814 * On success, We end up with an outstanding reference count against
5815 * iloc->bh. This _must_ be cleaned up later.
5816 */
5817
5818int
5819ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5820 struct ext4_iloc *iloc)
5821{
5822 int err;
5823
5824 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5825 return -EIO;
5826
5827 err = ext4_get_inode_loc(inode, iloc);
5828 if (!err) {
5829 BUFFER_TRACE(iloc->bh, "get_write_access");
5830 err = ext4_journal_get_write_access(handle, inode->i_sb,
5831 iloc->bh, EXT4_JTR_NONE);
5832 if (err) {
5833 brelse(iloc->bh);
5834 iloc->bh = NULL;
5835 }
5836 }
5837 ext4_std_error(inode->i_sb, err);
5838 return err;
5839}
5840
5841static int __ext4_expand_extra_isize(struct inode *inode,
5842 unsigned int new_extra_isize,
5843 struct ext4_iloc *iloc,
5844 handle_t *handle, int *no_expand)
5845{
5846 struct ext4_inode *raw_inode;
5847 struct ext4_xattr_ibody_header *header;
5848 unsigned int inode_size = EXT4_INODE_SIZE(inode->i_sb);
5849 struct ext4_inode_info *ei = EXT4_I(inode);
5850 int error;
5851
5852 /* this was checked at iget time, but double check for good measure */
5853 if ((EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize > inode_size) ||
5854 (ei->i_extra_isize & 3)) {
5855 EXT4_ERROR_INODE(inode, "bad extra_isize %u (inode size %u)",
5856 ei->i_extra_isize,
5857 EXT4_INODE_SIZE(inode->i_sb));
5858 return -EFSCORRUPTED;
5859 }
5860 if ((new_extra_isize < ei->i_extra_isize) ||
5861 (new_extra_isize < 4) ||
5862 (new_extra_isize > inode_size - EXT4_GOOD_OLD_INODE_SIZE))
5863 return -EINVAL; /* Should never happen */
5864
5865 raw_inode = ext4_raw_inode(iloc);
5866
5867 header = IHDR(inode, raw_inode);
5868
5869 /* No extended attributes present */
5870 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5871 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5872 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5873 EXT4_I(inode)->i_extra_isize, 0,
5874 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5875 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5876 return 0;
5877 }
5878
5879 /*
5880 * We may need to allocate external xattr block so we need quotas
5881 * initialized. Here we can be called with various locks held so we
5882 * cannot affort to initialize quotas ourselves. So just bail.
5883 */
5884 if (dquot_initialize_needed(inode))
5885 return -EAGAIN;
5886
5887 /* try to expand with EAs present */
5888 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5889 raw_inode, handle);
5890 if (error) {
5891 /*
5892 * Inode size expansion failed; don't try again
5893 */
5894 *no_expand = 1;
5895 }
5896
5897 return error;
5898}
5899
5900/*
5901 * Expand an inode by new_extra_isize bytes.
5902 * Returns 0 on success or negative error number on failure.
5903 */
5904static int ext4_try_to_expand_extra_isize(struct inode *inode,
5905 unsigned int new_extra_isize,
5906 struct ext4_iloc iloc,
5907 handle_t *handle)
5908{
5909 int no_expand;
5910 int error;
5911
5912 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5913 return -EOVERFLOW;
5914
5915 /*
5916 * In nojournal mode, we can immediately attempt to expand
5917 * the inode. When journaled, we first need to obtain extra
5918 * buffer credits since we may write into the EA block
5919 * with this same handle. If journal_extend fails, then it will
5920 * only result in a minor loss of functionality for that inode.
5921 * If this is felt to be critical, then e2fsck should be run to
5922 * force a large enough s_min_extra_isize.
5923 */
5924 if (ext4_journal_extend(handle,
5925 EXT4_DATA_TRANS_BLOCKS(inode->i_sb), 0) != 0)
5926 return -ENOSPC;
5927
5928 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5929 return -EBUSY;
5930
5931 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5932 handle, &no_expand);
5933 ext4_write_unlock_xattr(inode, &no_expand);
5934
5935 return error;
5936}
5937
5938int ext4_expand_extra_isize(struct inode *inode,
5939 unsigned int new_extra_isize,
5940 struct ext4_iloc *iloc)
5941{
5942 handle_t *handle;
5943 int no_expand;
5944 int error, rc;
5945
5946 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5947 brelse(iloc->bh);
5948 return -EOVERFLOW;
5949 }
5950
5951 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5952 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5953 if (IS_ERR(handle)) {
5954 error = PTR_ERR(handle);
5955 brelse(iloc->bh);
5956 return error;
5957 }
5958
5959 ext4_write_lock_xattr(inode, &no_expand);
5960
5961 BUFFER_TRACE(iloc->bh, "get_write_access");
5962 error = ext4_journal_get_write_access(handle, inode->i_sb, iloc->bh,
5963 EXT4_JTR_NONE);
5964 if (error) {
5965 brelse(iloc->bh);
5966 goto out_unlock;
5967 }
5968
5969 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5970 handle, &no_expand);
5971
5972 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5973 if (!error)
5974 error = rc;
5975
5976out_unlock:
5977 ext4_write_unlock_xattr(inode, &no_expand);
5978 ext4_journal_stop(handle);
5979 return error;
5980}
5981
5982/*
5983 * What we do here is to mark the in-core inode as clean with respect to inode
5984 * dirtiness (it may still be data-dirty).
5985 * This means that the in-core inode may be reaped by prune_icache
5986 * without having to perform any I/O. This is a very good thing,
5987 * because *any* task may call prune_icache - even ones which
5988 * have a transaction open against a different journal.
5989 *
5990 * Is this cheating? Not really. Sure, we haven't written the
5991 * inode out, but prune_icache isn't a user-visible syncing function.
5992 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5993 * we start and wait on commits.
5994 */
5995int __ext4_mark_inode_dirty(handle_t *handle, struct inode *inode,
5996 const char *func, unsigned int line)
5997{
5998 struct ext4_iloc iloc;
5999 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
6000 int err;
6001
6002 might_sleep();
6003 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
6004 err = ext4_reserve_inode_write(handle, inode, &iloc);
6005 if (err)
6006 goto out;
6007
6008 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
6009 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
6010 iloc, handle);
6011
6012 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
6013out:
6014 if (unlikely(err))
6015 ext4_error_inode_err(inode, func, line, 0, err,
6016 "mark_inode_dirty error");
6017 return err;
6018}
6019
6020/*
6021 * ext4_dirty_inode() is called from __mark_inode_dirty()
6022 *
6023 * We're really interested in the case where a file is being extended.
6024 * i_size has been changed by generic_commit_write() and we thus need
6025 * to include the updated inode in the current transaction.
6026 *
6027 * Also, dquot_alloc_block() will always dirty the inode when blocks
6028 * are allocated to the file.
6029 *
6030 * If the inode is marked synchronous, we don't honour that here - doing
6031 * so would cause a commit on atime updates, which we don't bother doing.
6032 * We handle synchronous inodes at the highest possible level.
6033 */
6034void ext4_dirty_inode(struct inode *inode, int flags)
6035{
6036 handle_t *handle;
6037
6038 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
6039 if (IS_ERR(handle))
6040 return;
6041 ext4_mark_inode_dirty(handle, inode);
6042 ext4_journal_stop(handle);
6043}
6044
6045int ext4_change_inode_journal_flag(struct inode *inode, int val)
6046{
6047 journal_t *journal;
6048 handle_t *handle;
6049 int err;
6050 int alloc_ctx;
6051
6052 /*
6053 * We have to be very careful here: changing a data block's
6054 * journaling status dynamically is dangerous. If we write a
6055 * data block to the journal, change the status and then delete
6056 * that block, we risk forgetting to revoke the old log record
6057 * from the journal and so a subsequent replay can corrupt data.
6058 * So, first we make sure that the journal is empty and that
6059 * nobody is changing anything.
6060 */
6061
6062 journal = EXT4_JOURNAL(inode);
6063 if (!journal)
6064 return 0;
6065 if (is_journal_aborted(journal))
6066 return -EROFS;
6067
6068 /* Wait for all existing dio workers */
6069 inode_dio_wait(inode);
6070
6071 /*
6072 * Before flushing the journal and switching inode's aops, we have
6073 * to flush all dirty data the inode has. There can be outstanding
6074 * delayed allocations, there can be unwritten extents created by
6075 * fallocate or buffered writes in dioread_nolock mode covered by
6076 * dirty data which can be converted only after flushing the dirty
6077 * data (and journalled aops don't know how to handle these cases).
6078 */
6079 if (val) {
6080 filemap_invalidate_lock(inode->i_mapping);
6081 err = filemap_write_and_wait(inode->i_mapping);
6082 if (err < 0) {
6083 filemap_invalidate_unlock(inode->i_mapping);
6084 return err;
6085 }
6086 }
6087
6088 alloc_ctx = ext4_writepages_down_write(inode->i_sb);
6089 jbd2_journal_lock_updates(journal);
6090
6091 /*
6092 * OK, there are no updates running now, and all cached data is
6093 * synced to disk. We are now in a completely consistent state
6094 * which doesn't have anything in the journal, and we know that
6095 * no filesystem updates are running, so it is safe to modify
6096 * the inode's in-core data-journaling state flag now.
6097 */
6098
6099 if (val)
6100 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6101 else {
6102 err = jbd2_journal_flush(journal, 0);
6103 if (err < 0) {
6104 jbd2_journal_unlock_updates(journal);
6105 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6106 return err;
6107 }
6108 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6109 }
6110 ext4_set_aops(inode);
6111
6112 jbd2_journal_unlock_updates(journal);
6113 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6114
6115 if (val)
6116 filemap_invalidate_unlock(inode->i_mapping);
6117
6118 /* Finally we can mark the inode as dirty. */
6119
6120 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6121 if (IS_ERR(handle))
6122 return PTR_ERR(handle);
6123
6124 ext4_fc_mark_ineligible(inode->i_sb,
6125 EXT4_FC_REASON_JOURNAL_FLAG_CHANGE, handle);
6126 err = ext4_mark_inode_dirty(handle, inode);
6127 ext4_handle_sync(handle);
6128 ext4_journal_stop(handle);
6129 ext4_std_error(inode->i_sb, err);
6130
6131 return err;
6132}
6133
6134static int ext4_bh_unmapped(handle_t *handle, struct inode *inode,
6135 struct buffer_head *bh)
6136{
6137 return !buffer_mapped(bh);
6138}
6139
6140vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
6141{
6142 struct vm_area_struct *vma = vmf->vma;
6143 struct folio *folio = page_folio(vmf->page);
6144 loff_t size;
6145 unsigned long len;
6146 int err;
6147 vm_fault_t ret;
6148 struct file *file = vma->vm_file;
6149 struct inode *inode = file_inode(file);
6150 struct address_space *mapping = inode->i_mapping;
6151 handle_t *handle;
6152 get_block_t *get_block;
6153 int retries = 0;
6154
6155 if (unlikely(IS_IMMUTABLE(inode)))
6156 return VM_FAULT_SIGBUS;
6157
6158 sb_start_pagefault(inode->i_sb);
6159 file_update_time(vma->vm_file);
6160
6161 filemap_invalidate_lock_shared(mapping);
6162
6163 err = ext4_convert_inline_data(inode);
6164 if (err)
6165 goto out_ret;
6166
6167 /*
6168 * On data journalling we skip straight to the transaction handle:
6169 * there's no delalloc; page truncated will be checked later; the
6170 * early return w/ all buffers mapped (calculates size/len) can't
6171 * be used; and there's no dioread_nolock, so only ext4_get_block.
6172 */
6173 if (ext4_should_journal_data(inode))
6174 goto retry_alloc;
6175
6176 /* Delalloc case is easy... */
6177 if (test_opt(inode->i_sb, DELALLOC) &&
6178 !ext4_nonda_switch(inode->i_sb)) {
6179 do {
6180 err = block_page_mkwrite(vma, vmf,
6181 ext4_da_get_block_prep);
6182 } while (err == -ENOSPC &&
6183 ext4_should_retry_alloc(inode->i_sb, &retries));
6184 goto out_ret;
6185 }
6186
6187 folio_lock(folio);
6188 size = i_size_read(inode);
6189 /* Page got truncated from under us? */
6190 if (folio->mapping != mapping || folio_pos(folio) > size) {
6191 folio_unlock(folio);
6192 ret = VM_FAULT_NOPAGE;
6193 goto out;
6194 }
6195
6196 len = folio_size(folio);
6197 if (folio_pos(folio) + len > size)
6198 len = size - folio_pos(folio);
6199 /*
6200 * Return if we have all the buffers mapped. This avoids the need to do
6201 * journal_start/journal_stop which can block and take a long time
6202 *
6203 * This cannot be done for data journalling, as we have to add the
6204 * inode to the transaction's list to writeprotect pages on commit.
6205 */
6206 if (folio_buffers(folio)) {
6207 if (!ext4_walk_page_buffers(NULL, inode, folio_buffers(folio),
6208 0, len, NULL,
6209 ext4_bh_unmapped)) {
6210 /* Wait so that we don't change page under IO */
6211 folio_wait_stable(folio);
6212 ret = VM_FAULT_LOCKED;
6213 goto out;
6214 }
6215 }
6216 folio_unlock(folio);
6217 /* OK, we need to fill the hole... */
6218 if (ext4_should_dioread_nolock(inode))
6219 get_block = ext4_get_block_unwritten;
6220 else
6221 get_block = ext4_get_block;
6222retry_alloc:
6223 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6224 ext4_writepage_trans_blocks(inode));
6225 if (IS_ERR(handle)) {
6226 ret = VM_FAULT_SIGBUS;
6227 goto out;
6228 }
6229 /*
6230 * Data journalling can't use block_page_mkwrite() because it
6231 * will set_buffer_dirty() before do_journal_get_write_access()
6232 * thus might hit warning messages for dirty metadata buffers.
6233 */
6234 if (!ext4_should_journal_data(inode)) {
6235 err = block_page_mkwrite(vma, vmf, get_block);
6236 } else {
6237 folio_lock(folio);
6238 size = i_size_read(inode);
6239 /* Page got truncated from under us? */
6240 if (folio->mapping != mapping || folio_pos(folio) > size) {
6241 ret = VM_FAULT_NOPAGE;
6242 goto out_error;
6243 }
6244
6245 len = folio_size(folio);
6246 if (folio_pos(folio) + len > size)
6247 len = size - folio_pos(folio);
6248
6249 err = ext4_block_write_begin(handle, folio, 0, len,
6250 ext4_get_block);
6251 if (!err) {
6252 ret = VM_FAULT_SIGBUS;
6253 if (ext4_journal_folio_buffers(handle, folio, len))
6254 goto out_error;
6255 } else {
6256 folio_unlock(folio);
6257 }
6258 }
6259 ext4_journal_stop(handle);
6260 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6261 goto retry_alloc;
6262out_ret:
6263 ret = vmf_fs_error(err);
6264out:
6265 filemap_invalidate_unlock_shared(mapping);
6266 sb_end_pagefault(inode->i_sb);
6267 return ret;
6268out_error:
6269 folio_unlock(folio);
6270 ext4_journal_stop(handle);
6271 goto out;
6272}
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}