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// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/inode.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/inode.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * 64-bit file support on 64-bit platforms by Jakub Jelinek
17 * (jj@sunsite.ms.mff.cuni.cz)
18 *
19 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20 */
21
22#include <linux/fs.h>
23#include <linux/mount.h>
24#include <linux/time.h>
25#include <linux/highuid.h>
26#include <linux/pagemap.h>
27#include <linux/dax.h>
28#include <linux/quotaops.h>
29#include <linux/string.h>
30#include <linux/buffer_head.h>
31#include <linux/writeback.h>
32#include <linux/pagevec.h>
33#include <linux/mpage.h>
34#include <linux/namei.h>
35#include <linux/uio.h>
36#include <linux/bio.h>
37#include <linux/workqueue.h>
38#include <linux/kernel.h>
39#include <linux/printk.h>
40#include <linux/slab.h>
41#include <linux/bitops.h>
42#include <linux/iomap.h>
43#include <linux/iversion.h>
44
45#include "ext4_jbd2.h"
46#include "xattr.h"
47#include "acl.h"
48#include "truncate.h"
49
50#include <trace/events/ext4.h>
51
52static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
53 struct ext4_inode_info *ei)
54{
55 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
56 __u32 csum;
57 __u16 dummy_csum = 0;
58 int offset = offsetof(struct ext4_inode, i_checksum_lo);
59 unsigned int csum_size = sizeof(dummy_csum);
60
61 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
62 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
63 offset += csum_size;
64 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
65 EXT4_GOOD_OLD_INODE_SIZE - offset);
66
67 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
68 offset = offsetof(struct ext4_inode, i_checksum_hi);
69 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
70 EXT4_GOOD_OLD_INODE_SIZE,
71 offset - EXT4_GOOD_OLD_INODE_SIZE);
72 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
73 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
74 csum_size);
75 offset += csum_size;
76 }
77 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
78 EXT4_INODE_SIZE(inode->i_sb) - offset);
79 }
80
81 return csum;
82}
83
84static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
85 struct ext4_inode_info *ei)
86{
87 __u32 provided, calculated;
88
89 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
90 cpu_to_le32(EXT4_OS_LINUX) ||
91 !ext4_has_metadata_csum(inode->i_sb))
92 return 1;
93
94 provided = le16_to_cpu(raw->i_checksum_lo);
95 calculated = ext4_inode_csum(inode, raw, ei);
96 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
97 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
98 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
99 else
100 calculated &= 0xFFFF;
101
102 return provided == calculated;
103}
104
105void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
106 struct ext4_inode_info *ei)
107{
108 __u32 csum;
109
110 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
111 cpu_to_le32(EXT4_OS_LINUX) ||
112 !ext4_has_metadata_csum(inode->i_sb))
113 return;
114
115 csum = ext4_inode_csum(inode, raw, ei);
116 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
117 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
118 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
119 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
120}
121
122static inline int ext4_begin_ordered_truncate(struct inode *inode,
123 loff_t new_size)
124{
125 trace_ext4_begin_ordered_truncate(inode, new_size);
126 /*
127 * If jinode is zero, then we never opened the file for
128 * writing, so there's no need to call
129 * jbd2_journal_begin_ordered_truncate() since there's no
130 * outstanding writes we need to flush.
131 */
132 if (!EXT4_I(inode)->jinode)
133 return 0;
134 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
135 EXT4_I(inode)->jinode,
136 new_size);
137}
138
139static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
140 int pextents);
141
142/*
143 * Test whether an inode is a fast symlink.
144 * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
145 */
146int ext4_inode_is_fast_symlink(struct inode *inode)
147{
148 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
149 int ea_blocks = EXT4_I(inode)->i_file_acl ?
150 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
151
152 if (ext4_has_inline_data(inode))
153 return 0;
154
155 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
156 }
157 return S_ISLNK(inode->i_mode) && inode->i_size &&
158 (inode->i_size < EXT4_N_BLOCKS * 4);
159}
160
161/*
162 * Called at the last iput() if i_nlink is zero.
163 */
164void ext4_evict_inode(struct inode *inode)
165{
166 handle_t *handle;
167 int err;
168 /*
169 * Credits for final inode cleanup and freeing:
170 * sb + inode (ext4_orphan_del()), block bitmap, group descriptor
171 * (xattr block freeing), bitmap, group descriptor (inode freeing)
172 */
173 int extra_credits = 6;
174 struct ext4_xattr_inode_array *ea_inode_array = NULL;
175 bool freeze_protected = false;
176
177 trace_ext4_evict_inode(inode);
178
179 if (EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)
180 ext4_evict_ea_inode(inode);
181 if (inode->i_nlink) {
182 truncate_inode_pages_final(&inode->i_data);
183
184 goto no_delete;
185 }
186
187 if (is_bad_inode(inode))
188 goto no_delete;
189 dquot_initialize(inode);
190
191 if (ext4_should_order_data(inode))
192 ext4_begin_ordered_truncate(inode, 0);
193 truncate_inode_pages_final(&inode->i_data);
194
195 /*
196 * For inodes with journalled data, transaction commit could have
197 * dirtied the inode. And for inodes with dioread_nolock, unwritten
198 * extents converting worker could merge extents and also have dirtied
199 * the inode. Flush worker is ignoring it because of I_FREEING flag but
200 * we still need to remove the inode from the writeback lists.
201 */
202 if (!list_empty_careful(&inode->i_io_list))
203 inode_io_list_del(inode);
204
205 /*
206 * Protect us against freezing - iput() caller didn't have to have any
207 * protection against it. When we are in a running transaction though,
208 * we are already protected against freezing and we cannot grab further
209 * protection due to lock ordering constraints.
210 */
211 if (!ext4_journal_current_handle()) {
212 sb_start_intwrite(inode->i_sb);
213 freeze_protected = true;
214 }
215
216 if (!IS_NOQUOTA(inode))
217 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
218
219 /*
220 * Block bitmap, group descriptor, and inode are accounted in both
221 * ext4_blocks_for_truncate() and extra_credits. So subtract 3.
222 */
223 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
224 ext4_blocks_for_truncate(inode) + extra_credits - 3);
225 if (IS_ERR(handle)) {
226 ext4_std_error(inode->i_sb, PTR_ERR(handle));
227 /*
228 * If we're going to skip the normal cleanup, we still need to
229 * make sure that the in-core orphan linked list is properly
230 * cleaned up.
231 */
232 ext4_orphan_del(NULL, inode);
233 if (freeze_protected)
234 sb_end_intwrite(inode->i_sb);
235 goto no_delete;
236 }
237
238 if (IS_SYNC(inode))
239 ext4_handle_sync(handle);
240
241 /*
242 * Set inode->i_size to 0 before calling ext4_truncate(). We need
243 * special handling of symlinks here because i_size is used to
244 * determine whether ext4_inode_info->i_data contains symlink data or
245 * block mappings. Setting i_size to 0 will remove its fast symlink
246 * status. Erase i_data so that it becomes a valid empty block map.
247 */
248 if (ext4_inode_is_fast_symlink(inode))
249 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
250 inode->i_size = 0;
251 err = ext4_mark_inode_dirty(handle, inode);
252 if (err) {
253 ext4_warning(inode->i_sb,
254 "couldn't mark inode dirty (err %d)", err);
255 goto stop_handle;
256 }
257 if (inode->i_blocks) {
258 err = ext4_truncate(inode);
259 if (err) {
260 ext4_error_err(inode->i_sb, -err,
261 "couldn't truncate inode %lu (err %d)",
262 inode->i_ino, err);
263 goto stop_handle;
264 }
265 }
266
267 /* Remove xattr references. */
268 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
269 extra_credits);
270 if (err) {
271 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
272stop_handle:
273 ext4_journal_stop(handle);
274 ext4_orphan_del(NULL, inode);
275 if (freeze_protected)
276 sb_end_intwrite(inode->i_sb);
277 ext4_xattr_inode_array_free(ea_inode_array);
278 goto no_delete;
279 }
280
281 /*
282 * Kill off the orphan record which ext4_truncate created.
283 * AKPM: I think this can be inside the above `if'.
284 * Note that ext4_orphan_del() has to be able to cope with the
285 * deletion of a non-existent orphan - this is because we don't
286 * know if ext4_truncate() actually created an orphan record.
287 * (Well, we could do this if we need to, but heck - it works)
288 */
289 ext4_orphan_del(handle, inode);
290 EXT4_I(inode)->i_dtime = (__u32)ktime_get_real_seconds();
291
292 /*
293 * One subtle ordering requirement: if anything has gone wrong
294 * (transaction abort, IO errors, whatever), then we can still
295 * do these next steps (the fs will already have been marked as
296 * having errors), but we can't free the inode if the mark_dirty
297 * fails.
298 */
299 if (ext4_mark_inode_dirty(handle, inode))
300 /* If that failed, just do the required in-core inode clear. */
301 ext4_clear_inode(inode);
302 else
303 ext4_free_inode(handle, inode);
304 ext4_journal_stop(handle);
305 if (freeze_protected)
306 sb_end_intwrite(inode->i_sb);
307 ext4_xattr_inode_array_free(ea_inode_array);
308 return;
309no_delete:
310 /*
311 * Check out some where else accidentally dirty the evicting inode,
312 * which may probably cause inode use-after-free issues later.
313 */
314 WARN_ON_ONCE(!list_empty_careful(&inode->i_io_list));
315
316 if (!list_empty(&EXT4_I(inode)->i_fc_list))
317 ext4_fc_mark_ineligible(inode->i_sb, EXT4_FC_REASON_NOMEM, NULL);
318 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
319}
320
321#ifdef CONFIG_QUOTA
322qsize_t *ext4_get_reserved_space(struct inode *inode)
323{
324 return &EXT4_I(inode)->i_reserved_quota;
325}
326#endif
327
328/*
329 * Called with i_data_sem down, which is important since we can call
330 * ext4_discard_preallocations() from here.
331 */
332void ext4_da_update_reserve_space(struct inode *inode,
333 int used, int quota_claim)
334{
335 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
336 struct ext4_inode_info *ei = EXT4_I(inode);
337
338 spin_lock(&ei->i_block_reservation_lock);
339 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
340 if (unlikely(used > ei->i_reserved_data_blocks)) {
341 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
342 "with only %d reserved data blocks",
343 __func__, inode->i_ino, used,
344 ei->i_reserved_data_blocks);
345 WARN_ON(1);
346 used = ei->i_reserved_data_blocks;
347 }
348
349 /* Update per-inode reservations */
350 ei->i_reserved_data_blocks -= used;
351 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
352
353 spin_unlock(&ei->i_block_reservation_lock);
354
355 /* Update quota subsystem for data blocks */
356 if (quota_claim)
357 dquot_claim_block(inode, EXT4_C2B(sbi, used));
358 else {
359 /*
360 * We did fallocate with an offset that is already delayed
361 * allocated. So on delayed allocated writeback we should
362 * not re-claim the quota for fallocated blocks.
363 */
364 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
365 }
366
367 /*
368 * If we have done all the pending block allocations and if
369 * there aren't any writers on the inode, we can discard the
370 * inode's preallocations.
371 */
372 if ((ei->i_reserved_data_blocks == 0) &&
373 !inode_is_open_for_write(inode))
374 ext4_discard_preallocations(inode);
375}
376
377static int __check_block_validity(struct inode *inode, const char *func,
378 unsigned int line,
379 struct ext4_map_blocks *map)
380{
381 if (ext4_has_feature_journal(inode->i_sb) &&
382 (inode->i_ino ==
383 le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
384 return 0;
385 if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
386 ext4_error_inode(inode, func, line, map->m_pblk,
387 "lblock %lu mapped to illegal pblock %llu "
388 "(length %d)", (unsigned long) map->m_lblk,
389 map->m_pblk, map->m_len);
390 return -EFSCORRUPTED;
391 }
392 return 0;
393}
394
395int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
396 ext4_lblk_t len)
397{
398 int ret;
399
400 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
401 return fscrypt_zeroout_range(inode, lblk, pblk, len);
402
403 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
404 if (ret > 0)
405 ret = 0;
406
407 return ret;
408}
409
410#define check_block_validity(inode, map) \
411 __check_block_validity((inode), __func__, __LINE__, (map))
412
413#ifdef ES_AGGRESSIVE_TEST
414static void ext4_map_blocks_es_recheck(handle_t *handle,
415 struct inode *inode,
416 struct ext4_map_blocks *es_map,
417 struct ext4_map_blocks *map,
418 int flags)
419{
420 int retval;
421
422 map->m_flags = 0;
423 /*
424 * There is a race window that the result is not the same.
425 * e.g. xfstests #223 when dioread_nolock enables. The reason
426 * is that we lookup a block mapping in extent status tree with
427 * out taking i_data_sem. So at the time the unwritten extent
428 * could be converted.
429 */
430 down_read(&EXT4_I(inode)->i_data_sem);
431 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
432 retval = ext4_ext_map_blocks(handle, inode, map, 0);
433 } else {
434 retval = ext4_ind_map_blocks(handle, inode, map, 0);
435 }
436 up_read((&EXT4_I(inode)->i_data_sem));
437
438 /*
439 * We don't check m_len because extent will be collpased in status
440 * tree. So the m_len might not equal.
441 */
442 if (es_map->m_lblk != map->m_lblk ||
443 es_map->m_flags != map->m_flags ||
444 es_map->m_pblk != map->m_pblk) {
445 printk("ES cache assertion failed for inode: %lu "
446 "es_cached ex [%d/%d/%llu/%x] != "
447 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
448 inode->i_ino, es_map->m_lblk, es_map->m_len,
449 es_map->m_pblk, es_map->m_flags, map->m_lblk,
450 map->m_len, map->m_pblk, map->m_flags,
451 retval, flags);
452 }
453}
454#endif /* ES_AGGRESSIVE_TEST */
455
456/*
457 * The ext4_map_blocks() function tries to look up the requested blocks,
458 * and returns if the blocks are already mapped.
459 *
460 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
461 * and store the allocated blocks in the result buffer head and mark it
462 * mapped.
463 *
464 * If file type is extents based, it will call ext4_ext_map_blocks(),
465 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
466 * based files
467 *
468 * On success, it returns the number of blocks being mapped or allocated.
469 * If flags doesn't contain EXT4_GET_BLOCKS_CREATE the blocks are
470 * pre-allocated and unwritten, the resulting @map is marked as unwritten.
471 * If the flags contain EXT4_GET_BLOCKS_CREATE, it will mark @map as mapped.
472 *
473 * It returns 0 if plain look up failed (blocks have not been allocated), in
474 * that case, @map is returned as unmapped but we still do fill map->m_len to
475 * indicate the length of a hole starting at map->m_lblk.
476 *
477 * It returns the error in case of allocation failure.
478 */
479int ext4_map_blocks(handle_t *handle, struct inode *inode,
480 struct ext4_map_blocks *map, int flags)
481{
482 struct extent_status es;
483 int retval;
484 int ret = 0;
485#ifdef ES_AGGRESSIVE_TEST
486 struct ext4_map_blocks orig_map;
487
488 memcpy(&orig_map, map, sizeof(*map));
489#endif
490
491 map->m_flags = 0;
492 ext_debug(inode, "flag 0x%x, max_blocks %u, logical block %lu\n",
493 flags, map->m_len, (unsigned long) map->m_lblk);
494
495 /*
496 * ext4_map_blocks returns an int, and m_len is an unsigned int
497 */
498 if (unlikely(map->m_len > INT_MAX))
499 map->m_len = INT_MAX;
500
501 /* We can handle the block number less than EXT_MAX_BLOCKS */
502 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
503 return -EFSCORRUPTED;
504
505 /* Lookup extent status tree firstly */
506 if (!(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY) &&
507 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
508 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
509 map->m_pblk = ext4_es_pblock(&es) +
510 map->m_lblk - es.es_lblk;
511 map->m_flags |= ext4_es_is_written(&es) ?
512 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
513 retval = es.es_len - (map->m_lblk - es.es_lblk);
514 if (retval > map->m_len)
515 retval = map->m_len;
516 map->m_len = retval;
517 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
518 map->m_pblk = 0;
519 map->m_flags |= ext4_es_is_delayed(&es) ?
520 EXT4_MAP_DELAYED : 0;
521 retval = es.es_len - (map->m_lblk - es.es_lblk);
522 if (retval > map->m_len)
523 retval = map->m_len;
524 map->m_len = retval;
525 retval = 0;
526 } else {
527 BUG();
528 }
529
530 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
531 return retval;
532#ifdef ES_AGGRESSIVE_TEST
533 ext4_map_blocks_es_recheck(handle, inode, map,
534 &orig_map, flags);
535#endif
536 goto found;
537 }
538 /*
539 * In the query cache no-wait mode, nothing we can do more if we
540 * cannot find extent in the cache.
541 */
542 if (flags & EXT4_GET_BLOCKS_CACHED_NOWAIT)
543 return 0;
544
545 /*
546 * Try to see if we can get the block without requesting a new
547 * file system block.
548 */
549 down_read(&EXT4_I(inode)->i_data_sem);
550 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
551 retval = ext4_ext_map_blocks(handle, inode, map, 0);
552 } else {
553 retval = ext4_ind_map_blocks(handle, inode, map, 0);
554 }
555 if (retval > 0) {
556 unsigned int status;
557
558 if (unlikely(retval != map->m_len)) {
559 ext4_warning(inode->i_sb,
560 "ES len assertion failed for inode "
561 "%lu: retval %d != map->m_len %d",
562 inode->i_ino, retval, map->m_len);
563 WARN_ON(1);
564 }
565
566 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
567 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
568 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
569 !(status & EXTENT_STATUS_WRITTEN) &&
570 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
571 map->m_lblk + map->m_len - 1))
572 status |= EXTENT_STATUS_DELAYED;
573 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
574 map->m_pblk, status);
575 }
576 up_read((&EXT4_I(inode)->i_data_sem));
577
578found:
579 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
580 ret = check_block_validity(inode, map);
581 if (ret != 0)
582 return ret;
583 }
584
585 /* If it is only a block(s) look up */
586 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
587 return retval;
588
589 /*
590 * Returns if the blocks have already allocated
591 *
592 * Note that if blocks have been preallocated
593 * ext4_ext_map_blocks() returns with buffer head unmapped
594 */
595 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
596 /*
597 * If we need to convert extent to unwritten
598 * we continue and do the actual work in
599 * ext4_ext_map_blocks()
600 */
601 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
602 return retval;
603
604 /*
605 * Here we clear m_flags because after allocating an new extent,
606 * it will be set again.
607 */
608 map->m_flags &= ~EXT4_MAP_FLAGS;
609
610 /*
611 * New blocks allocate and/or writing to unwritten extent
612 * will possibly result in updating i_data, so we take
613 * the write lock of i_data_sem, and call get_block()
614 * with create == 1 flag.
615 */
616 down_write(&EXT4_I(inode)->i_data_sem);
617
618 /*
619 * We need to check for EXT4 here because migrate
620 * could have changed the inode type in between
621 */
622 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
623 retval = ext4_ext_map_blocks(handle, inode, map, flags);
624 } else {
625 retval = ext4_ind_map_blocks(handle, inode, map, flags);
626
627 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
628 /*
629 * We allocated new blocks which will result in
630 * i_data's format changing. Force the migrate
631 * to fail by clearing migrate flags
632 */
633 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
634 }
635 }
636
637 if (retval > 0) {
638 unsigned int status;
639
640 if (unlikely(retval != map->m_len)) {
641 ext4_warning(inode->i_sb,
642 "ES len assertion failed for inode "
643 "%lu: retval %d != map->m_len %d",
644 inode->i_ino, retval, map->m_len);
645 WARN_ON(1);
646 }
647
648 /*
649 * We have to zeroout blocks before inserting them into extent
650 * status tree. Otherwise someone could look them up there and
651 * use them before they are really zeroed. We also have to
652 * unmap metadata before zeroing as otherwise writeback can
653 * overwrite zeros with stale data from block device.
654 */
655 if (flags & EXT4_GET_BLOCKS_ZERO &&
656 map->m_flags & EXT4_MAP_MAPPED &&
657 map->m_flags & EXT4_MAP_NEW) {
658 ret = ext4_issue_zeroout(inode, map->m_lblk,
659 map->m_pblk, map->m_len);
660 if (ret) {
661 retval = ret;
662 goto out_sem;
663 }
664 }
665
666 /*
667 * If the extent has been zeroed out, we don't need to update
668 * extent status tree.
669 */
670 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
671 ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) {
672 if (ext4_es_is_written(&es))
673 goto out_sem;
674 }
675 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
676 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
677 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
678 !(status & EXTENT_STATUS_WRITTEN) &&
679 ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
680 map->m_lblk + map->m_len - 1))
681 status |= EXTENT_STATUS_DELAYED;
682 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
683 map->m_pblk, status);
684 }
685
686out_sem:
687 up_write((&EXT4_I(inode)->i_data_sem));
688 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
689 ret = check_block_validity(inode, map);
690 if (ret != 0)
691 return ret;
692
693 /*
694 * Inodes with freshly allocated blocks where contents will be
695 * visible after transaction commit must be on transaction's
696 * ordered data list.
697 */
698 if (map->m_flags & EXT4_MAP_NEW &&
699 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
700 !(flags & EXT4_GET_BLOCKS_ZERO) &&
701 !ext4_is_quota_file(inode) &&
702 ext4_should_order_data(inode)) {
703 loff_t start_byte =
704 (loff_t)map->m_lblk << inode->i_blkbits;
705 loff_t length = (loff_t)map->m_len << inode->i_blkbits;
706
707 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
708 ret = ext4_jbd2_inode_add_wait(handle, inode,
709 start_byte, length);
710 else
711 ret = ext4_jbd2_inode_add_write(handle, inode,
712 start_byte, length);
713 if (ret)
714 return ret;
715 }
716 }
717 if (retval > 0 && (map->m_flags & EXT4_MAP_UNWRITTEN ||
718 map->m_flags & EXT4_MAP_MAPPED))
719 ext4_fc_track_range(handle, inode, map->m_lblk,
720 map->m_lblk + map->m_len - 1);
721 if (retval < 0)
722 ext_debug(inode, "failed with err %d\n", retval);
723 return retval;
724}
725
726/*
727 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
728 * we have to be careful as someone else may be manipulating b_state as well.
729 */
730static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
731{
732 unsigned long old_state;
733 unsigned long new_state;
734
735 flags &= EXT4_MAP_FLAGS;
736
737 /* Dummy buffer_head? Set non-atomically. */
738 if (!bh->b_page) {
739 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
740 return;
741 }
742 /*
743 * Someone else may be modifying b_state. Be careful! This is ugly but
744 * once we get rid of using bh as a container for mapping information
745 * to pass to / from get_block functions, this can go away.
746 */
747 old_state = READ_ONCE(bh->b_state);
748 do {
749 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
750 } while (unlikely(!try_cmpxchg(&bh->b_state, &old_state, new_state)));
751}
752
753static int _ext4_get_block(struct inode *inode, sector_t iblock,
754 struct buffer_head *bh, int flags)
755{
756 struct ext4_map_blocks map;
757 int ret = 0;
758
759 if (ext4_has_inline_data(inode))
760 return -ERANGE;
761
762 map.m_lblk = iblock;
763 map.m_len = bh->b_size >> inode->i_blkbits;
764
765 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
766 flags);
767 if (ret > 0) {
768 map_bh(bh, inode->i_sb, map.m_pblk);
769 ext4_update_bh_state(bh, map.m_flags);
770 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
771 ret = 0;
772 } else if (ret == 0) {
773 /* hole case, need to fill in bh->b_size */
774 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
775 }
776 return ret;
777}
778
779int ext4_get_block(struct inode *inode, sector_t iblock,
780 struct buffer_head *bh, int create)
781{
782 return _ext4_get_block(inode, iblock, bh,
783 create ? EXT4_GET_BLOCKS_CREATE : 0);
784}
785
786/*
787 * Get block function used when preparing for buffered write if we require
788 * creating an unwritten extent if blocks haven't been allocated. The extent
789 * will be converted to written after the IO is complete.
790 */
791int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
792 struct buffer_head *bh_result, int create)
793{
794 int ret = 0;
795
796 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
797 inode->i_ino, create);
798 ret = _ext4_get_block(inode, iblock, bh_result,
799 EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT);
800
801 /*
802 * If the buffer is marked unwritten, mark it as new to make sure it is
803 * zeroed out correctly in case of partial writes. Otherwise, there is
804 * a chance of stale data getting exposed.
805 */
806 if (ret == 0 && buffer_unwritten(bh_result))
807 set_buffer_new(bh_result);
808
809 return ret;
810}
811
812/* Maximum number of blocks we map for direct IO at once. */
813#define DIO_MAX_BLOCKS 4096
814
815/*
816 * `handle' can be NULL if create is zero
817 */
818struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
819 ext4_lblk_t block, int map_flags)
820{
821 struct ext4_map_blocks map;
822 struct buffer_head *bh;
823 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
824 bool nowait = map_flags & EXT4_GET_BLOCKS_CACHED_NOWAIT;
825 int err;
826
827 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
828 || handle != NULL || create == 0);
829 ASSERT(create == 0 || !nowait);
830
831 map.m_lblk = block;
832 map.m_len = 1;
833 err = ext4_map_blocks(handle, inode, &map, map_flags);
834
835 if (err == 0)
836 return create ? ERR_PTR(-ENOSPC) : NULL;
837 if (err < 0)
838 return ERR_PTR(err);
839
840 if (nowait)
841 return sb_find_get_block(inode->i_sb, map.m_pblk);
842
843 bh = sb_getblk(inode->i_sb, map.m_pblk);
844 if (unlikely(!bh))
845 return ERR_PTR(-ENOMEM);
846 if (map.m_flags & EXT4_MAP_NEW) {
847 ASSERT(create != 0);
848 ASSERT((EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
849 || (handle != NULL));
850
851 /*
852 * Now that we do not always journal data, we should
853 * keep in mind whether this should always journal the
854 * new buffer as metadata. For now, regular file
855 * writes use ext4_get_block instead, so it's not a
856 * problem.
857 */
858 lock_buffer(bh);
859 BUFFER_TRACE(bh, "call get_create_access");
860 err = ext4_journal_get_create_access(handle, inode->i_sb, bh,
861 EXT4_JTR_NONE);
862 if (unlikely(err)) {
863 unlock_buffer(bh);
864 goto errout;
865 }
866 if (!buffer_uptodate(bh)) {
867 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
868 set_buffer_uptodate(bh);
869 }
870 unlock_buffer(bh);
871 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
872 err = ext4_handle_dirty_metadata(handle, inode, bh);
873 if (unlikely(err))
874 goto errout;
875 } else
876 BUFFER_TRACE(bh, "not a new buffer");
877 return bh;
878errout:
879 brelse(bh);
880 return ERR_PTR(err);
881}
882
883struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
884 ext4_lblk_t block, int map_flags)
885{
886 struct buffer_head *bh;
887 int ret;
888
889 bh = ext4_getblk(handle, inode, block, map_flags);
890 if (IS_ERR(bh))
891 return bh;
892 if (!bh || ext4_buffer_uptodate(bh))
893 return bh;
894
895 ret = ext4_read_bh_lock(bh, REQ_META | REQ_PRIO, true);
896 if (ret) {
897 put_bh(bh);
898 return ERR_PTR(ret);
899 }
900 return bh;
901}
902
903/* Read a contiguous batch of blocks. */
904int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
905 bool wait, struct buffer_head **bhs)
906{
907 int i, err;
908
909 for (i = 0; i < bh_count; i++) {
910 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
911 if (IS_ERR(bhs[i])) {
912 err = PTR_ERR(bhs[i]);
913 bh_count = i;
914 goto out_brelse;
915 }
916 }
917
918 for (i = 0; i < bh_count; i++)
919 /* Note that NULL bhs[i] is valid because of holes. */
920 if (bhs[i] && !ext4_buffer_uptodate(bhs[i]))
921 ext4_read_bh_lock(bhs[i], REQ_META | REQ_PRIO, false);
922
923 if (!wait)
924 return 0;
925
926 for (i = 0; i < bh_count; i++)
927 if (bhs[i])
928 wait_on_buffer(bhs[i]);
929
930 for (i = 0; i < bh_count; i++) {
931 if (bhs[i] && !buffer_uptodate(bhs[i])) {
932 err = -EIO;
933 goto out_brelse;
934 }
935 }
936 return 0;
937
938out_brelse:
939 for (i = 0; i < bh_count; i++) {
940 brelse(bhs[i]);
941 bhs[i] = NULL;
942 }
943 return err;
944}
945
946int ext4_walk_page_buffers(handle_t *handle, struct inode *inode,
947 struct buffer_head *head,
948 unsigned from,
949 unsigned to,
950 int *partial,
951 int (*fn)(handle_t *handle, struct inode *inode,
952 struct buffer_head *bh))
953{
954 struct buffer_head *bh;
955 unsigned block_start, block_end;
956 unsigned blocksize = head->b_size;
957 int err, ret = 0;
958 struct buffer_head *next;
959
960 for (bh = head, block_start = 0;
961 ret == 0 && (bh != head || !block_start);
962 block_start = block_end, bh = next) {
963 next = bh->b_this_page;
964 block_end = block_start + blocksize;
965 if (block_end <= from || block_start >= to) {
966 if (partial && !buffer_uptodate(bh))
967 *partial = 1;
968 continue;
969 }
970 err = (*fn)(handle, inode, bh);
971 if (!ret)
972 ret = err;
973 }
974 return ret;
975}
976
977/*
978 * Helper for handling dirtying of journalled data. We also mark the folio as
979 * dirty so that writeback code knows about this page (and inode) contains
980 * dirty data. ext4_writepages() then commits appropriate transaction to
981 * make data stable.
982 */
983static int ext4_dirty_journalled_data(handle_t *handle, struct buffer_head *bh)
984{
985 folio_mark_dirty(bh->b_folio);
986 return ext4_handle_dirty_metadata(handle, NULL, bh);
987}
988
989int do_journal_get_write_access(handle_t *handle, struct inode *inode,
990 struct buffer_head *bh)
991{
992 int dirty = buffer_dirty(bh);
993 int ret;
994
995 if (!buffer_mapped(bh) || buffer_freed(bh))
996 return 0;
997 /*
998 * __block_write_begin() could have dirtied some buffers. Clean
999 * the dirty bit as jbd2_journal_get_write_access() could complain
1000 * otherwise about fs integrity issues. Setting of the dirty bit
1001 * by __block_write_begin() isn't a real problem here as we clear
1002 * the bit before releasing a page lock and thus writeback cannot
1003 * ever write the buffer.
1004 */
1005 if (dirty)
1006 clear_buffer_dirty(bh);
1007 BUFFER_TRACE(bh, "get write access");
1008 ret = ext4_journal_get_write_access(handle, inode->i_sb, bh,
1009 EXT4_JTR_NONE);
1010 if (!ret && dirty)
1011 ret = ext4_dirty_journalled_data(handle, bh);
1012 return ret;
1013}
1014
1015#ifdef CONFIG_FS_ENCRYPTION
1016static int ext4_block_write_begin(struct folio *folio, loff_t pos, unsigned len,
1017 get_block_t *get_block)
1018{
1019 unsigned from = pos & (PAGE_SIZE - 1);
1020 unsigned to = from + len;
1021 struct inode *inode = folio->mapping->host;
1022 unsigned block_start, block_end;
1023 sector_t block;
1024 int err = 0;
1025 unsigned blocksize = inode->i_sb->s_blocksize;
1026 unsigned bbits;
1027 struct buffer_head *bh, *head, *wait[2];
1028 int nr_wait = 0;
1029 int i;
1030
1031 BUG_ON(!folio_test_locked(folio));
1032 BUG_ON(from > PAGE_SIZE);
1033 BUG_ON(to > PAGE_SIZE);
1034 BUG_ON(from > to);
1035
1036 head = folio_buffers(folio);
1037 if (!head)
1038 head = create_empty_buffers(folio, blocksize, 0);
1039 bbits = ilog2(blocksize);
1040 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
1041
1042 for (bh = head, block_start = 0; bh != head || !block_start;
1043 block++, block_start = block_end, bh = bh->b_this_page) {
1044 block_end = block_start + blocksize;
1045 if (block_end <= from || block_start >= to) {
1046 if (folio_test_uptodate(folio)) {
1047 set_buffer_uptodate(bh);
1048 }
1049 continue;
1050 }
1051 if (buffer_new(bh))
1052 clear_buffer_new(bh);
1053 if (!buffer_mapped(bh)) {
1054 WARN_ON(bh->b_size != blocksize);
1055 err = get_block(inode, block, bh, 1);
1056 if (err)
1057 break;
1058 if (buffer_new(bh)) {
1059 if (folio_test_uptodate(folio)) {
1060 clear_buffer_new(bh);
1061 set_buffer_uptodate(bh);
1062 mark_buffer_dirty(bh);
1063 continue;
1064 }
1065 if (block_end > to || block_start < from)
1066 folio_zero_segments(folio, to,
1067 block_end,
1068 block_start, from);
1069 continue;
1070 }
1071 }
1072 if (folio_test_uptodate(folio)) {
1073 set_buffer_uptodate(bh);
1074 continue;
1075 }
1076 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1077 !buffer_unwritten(bh) &&
1078 (block_start < from || block_end > to)) {
1079 ext4_read_bh_lock(bh, 0, false);
1080 wait[nr_wait++] = bh;
1081 }
1082 }
1083 /*
1084 * If we issued read requests, let them complete.
1085 */
1086 for (i = 0; i < nr_wait; i++) {
1087 wait_on_buffer(wait[i]);
1088 if (!buffer_uptodate(wait[i]))
1089 err = -EIO;
1090 }
1091 if (unlikely(err)) {
1092 folio_zero_new_buffers(folio, from, to);
1093 } else if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
1094 for (i = 0; i < nr_wait; i++) {
1095 int err2;
1096
1097 err2 = fscrypt_decrypt_pagecache_blocks(folio,
1098 blocksize, bh_offset(wait[i]));
1099 if (err2) {
1100 clear_buffer_uptodate(wait[i]);
1101 err = err2;
1102 }
1103 }
1104 }
1105
1106 return err;
1107}
1108#endif
1109
1110/*
1111 * To preserve ordering, it is essential that the hole instantiation and
1112 * the data write be encapsulated in a single transaction. We cannot
1113 * close off a transaction and start a new one between the ext4_get_block()
1114 * and the ext4_write_end(). So doing the jbd2_journal_start at the start of
1115 * ext4_write_begin() is the right place.
1116 */
1117static int ext4_write_begin(struct file *file, struct address_space *mapping,
1118 loff_t pos, unsigned len,
1119 struct page **pagep, void **fsdata)
1120{
1121 struct inode *inode = mapping->host;
1122 int ret, needed_blocks;
1123 handle_t *handle;
1124 int retries = 0;
1125 struct folio *folio;
1126 pgoff_t index;
1127 unsigned from, to;
1128
1129 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
1130 return -EIO;
1131
1132 trace_ext4_write_begin(inode, pos, len);
1133 /*
1134 * Reserve one block more for addition to orphan list in case
1135 * we allocate blocks but write fails for some reason
1136 */
1137 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1138 index = pos >> PAGE_SHIFT;
1139 from = pos & (PAGE_SIZE - 1);
1140 to = from + len;
1141
1142 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1143 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1144 pagep);
1145 if (ret < 0)
1146 return ret;
1147 if (ret == 1)
1148 return 0;
1149 }
1150
1151 /*
1152 * __filemap_get_folio() can take a long time if the
1153 * system is thrashing due to memory pressure, or if the folio
1154 * is being written back. So grab it first before we start
1155 * the transaction handle. This also allows us to allocate
1156 * the folio (if needed) without using GFP_NOFS.
1157 */
1158retry_grab:
1159 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
1160 mapping_gfp_mask(mapping));
1161 if (IS_ERR(folio))
1162 return PTR_ERR(folio);
1163 /*
1164 * The same as page allocation, we prealloc buffer heads before
1165 * starting the handle.
1166 */
1167 if (!folio_buffers(folio))
1168 create_empty_buffers(folio, inode->i_sb->s_blocksize, 0);
1169
1170 folio_unlock(folio);
1171
1172retry_journal:
1173 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1174 if (IS_ERR(handle)) {
1175 folio_put(folio);
1176 return PTR_ERR(handle);
1177 }
1178
1179 folio_lock(folio);
1180 if (folio->mapping != mapping) {
1181 /* The folio got truncated from under us */
1182 folio_unlock(folio);
1183 folio_put(folio);
1184 ext4_journal_stop(handle);
1185 goto retry_grab;
1186 }
1187 /* In case writeback began while the folio was unlocked */
1188 folio_wait_stable(folio);
1189
1190#ifdef CONFIG_FS_ENCRYPTION
1191 if (ext4_should_dioread_nolock(inode))
1192 ret = ext4_block_write_begin(folio, pos, len,
1193 ext4_get_block_unwritten);
1194 else
1195 ret = ext4_block_write_begin(folio, pos, len, ext4_get_block);
1196#else
1197 if (ext4_should_dioread_nolock(inode))
1198 ret = __block_write_begin(&folio->page, pos, len,
1199 ext4_get_block_unwritten);
1200 else
1201 ret = __block_write_begin(&folio->page, pos, len, ext4_get_block);
1202#endif
1203 if (!ret && ext4_should_journal_data(inode)) {
1204 ret = ext4_walk_page_buffers(handle, inode,
1205 folio_buffers(folio), from, to,
1206 NULL, do_journal_get_write_access);
1207 }
1208
1209 if (ret) {
1210 bool extended = (pos + len > inode->i_size) &&
1211 !ext4_verity_in_progress(inode);
1212
1213 folio_unlock(folio);
1214 /*
1215 * __block_write_begin may have instantiated a few blocks
1216 * outside i_size. Trim these off again. Don't need
1217 * i_size_read because we hold i_rwsem.
1218 *
1219 * Add inode to orphan list in case we crash before
1220 * truncate finishes
1221 */
1222 if (extended && ext4_can_truncate(inode))
1223 ext4_orphan_add(handle, inode);
1224
1225 ext4_journal_stop(handle);
1226 if (extended) {
1227 ext4_truncate_failed_write(inode);
1228 /*
1229 * If truncate failed early the inode might
1230 * still be on the orphan list; we need to
1231 * make sure the inode is removed from the
1232 * orphan list in that case.
1233 */
1234 if (inode->i_nlink)
1235 ext4_orphan_del(NULL, inode);
1236 }
1237
1238 if (ret == -ENOSPC &&
1239 ext4_should_retry_alloc(inode->i_sb, &retries))
1240 goto retry_journal;
1241 folio_put(folio);
1242 return ret;
1243 }
1244 *pagep = &folio->page;
1245 return ret;
1246}
1247
1248/* For write_end() in data=journal mode */
1249static int write_end_fn(handle_t *handle, struct inode *inode,
1250 struct buffer_head *bh)
1251{
1252 int ret;
1253 if (!buffer_mapped(bh) || buffer_freed(bh))
1254 return 0;
1255 set_buffer_uptodate(bh);
1256 ret = ext4_dirty_journalled_data(handle, bh);
1257 clear_buffer_meta(bh);
1258 clear_buffer_prio(bh);
1259 return ret;
1260}
1261
1262/*
1263 * We need to pick up the new inode size which generic_commit_write gave us
1264 * `file' can be NULL - eg, when called from page_symlink().
1265 *
1266 * ext4 never places buffers on inode->i_mapping->i_private_list. metadata
1267 * buffers are managed internally.
1268 */
1269static int ext4_write_end(struct file *file,
1270 struct address_space *mapping,
1271 loff_t pos, unsigned len, unsigned copied,
1272 struct page *page, void *fsdata)
1273{
1274 struct folio *folio = page_folio(page);
1275 handle_t *handle = ext4_journal_current_handle();
1276 struct inode *inode = mapping->host;
1277 loff_t old_size = inode->i_size;
1278 int ret = 0, ret2;
1279 int i_size_changed = 0;
1280 bool verity = ext4_verity_in_progress(inode);
1281
1282 trace_ext4_write_end(inode, pos, len, copied);
1283
1284 if (ext4_has_inline_data(inode) &&
1285 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA))
1286 return ext4_write_inline_data_end(inode, pos, len, copied,
1287 folio);
1288
1289 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1290 /*
1291 * it's important to update i_size while still holding folio lock:
1292 * page writeout could otherwise come in and zero beyond i_size.
1293 *
1294 * If FS_IOC_ENABLE_VERITY is running on this inode, then Merkle tree
1295 * blocks are being written past EOF, so skip the i_size update.
1296 */
1297 if (!verity)
1298 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1299 folio_unlock(folio);
1300 folio_put(folio);
1301
1302 if (old_size < pos && !verity)
1303 pagecache_isize_extended(inode, old_size, pos);
1304 /*
1305 * Don't mark the inode dirty under folio lock. First, it unnecessarily
1306 * makes the holding time of folio lock longer. Second, it forces lock
1307 * ordering of folio lock and transaction start for journaling
1308 * filesystems.
1309 */
1310 if (i_size_changed)
1311 ret = ext4_mark_inode_dirty(handle, inode);
1312
1313 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1314 /* if we have allocated more blocks and copied
1315 * less. We will have blocks allocated outside
1316 * inode->i_size. So truncate them
1317 */
1318 ext4_orphan_add(handle, inode);
1319
1320 ret2 = ext4_journal_stop(handle);
1321 if (!ret)
1322 ret = ret2;
1323
1324 if (pos + len > inode->i_size && !verity) {
1325 ext4_truncate_failed_write(inode);
1326 /*
1327 * If truncate failed early the inode might still be
1328 * on the orphan list; we need to make sure the inode
1329 * is removed from the orphan list in that case.
1330 */
1331 if (inode->i_nlink)
1332 ext4_orphan_del(NULL, inode);
1333 }
1334
1335 return ret ? ret : copied;
1336}
1337
1338/*
1339 * This is a private version of folio_zero_new_buffers() which doesn't
1340 * set the buffer to be dirty, since in data=journalled mode we need
1341 * to call ext4_dirty_journalled_data() instead.
1342 */
1343static void ext4_journalled_zero_new_buffers(handle_t *handle,
1344 struct inode *inode,
1345 struct folio *folio,
1346 unsigned from, unsigned to)
1347{
1348 unsigned int block_start = 0, block_end;
1349 struct buffer_head *head, *bh;
1350
1351 bh = head = folio_buffers(folio);
1352 do {
1353 block_end = block_start + bh->b_size;
1354 if (buffer_new(bh)) {
1355 if (block_end > from && block_start < to) {
1356 if (!folio_test_uptodate(folio)) {
1357 unsigned start, size;
1358
1359 start = max(from, block_start);
1360 size = min(to, block_end) - start;
1361
1362 folio_zero_range(folio, start, size);
1363 write_end_fn(handle, inode, bh);
1364 }
1365 clear_buffer_new(bh);
1366 }
1367 }
1368 block_start = block_end;
1369 bh = bh->b_this_page;
1370 } while (bh != head);
1371}
1372
1373static int ext4_journalled_write_end(struct file *file,
1374 struct address_space *mapping,
1375 loff_t pos, unsigned len, unsigned copied,
1376 struct page *page, void *fsdata)
1377{
1378 struct folio *folio = page_folio(page);
1379 handle_t *handle = ext4_journal_current_handle();
1380 struct inode *inode = mapping->host;
1381 loff_t old_size = inode->i_size;
1382 int ret = 0, ret2;
1383 int partial = 0;
1384 unsigned from, to;
1385 int size_changed = 0;
1386 bool verity = ext4_verity_in_progress(inode);
1387
1388 trace_ext4_journalled_write_end(inode, pos, len, copied);
1389 from = pos & (PAGE_SIZE - 1);
1390 to = from + len;
1391
1392 BUG_ON(!ext4_handle_valid(handle));
1393
1394 if (ext4_has_inline_data(inode))
1395 return ext4_write_inline_data_end(inode, pos, len, copied,
1396 folio);
1397
1398 if (unlikely(copied < len) && !folio_test_uptodate(folio)) {
1399 copied = 0;
1400 ext4_journalled_zero_new_buffers(handle, inode, folio,
1401 from, to);
1402 } else {
1403 if (unlikely(copied < len))
1404 ext4_journalled_zero_new_buffers(handle, inode, folio,
1405 from + copied, to);
1406 ret = ext4_walk_page_buffers(handle, inode,
1407 folio_buffers(folio),
1408 from, from + copied, &partial,
1409 write_end_fn);
1410 if (!partial)
1411 folio_mark_uptodate(folio);
1412 }
1413 if (!verity)
1414 size_changed = ext4_update_inode_size(inode, pos + copied);
1415 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1416 folio_unlock(folio);
1417 folio_put(folio);
1418
1419 if (old_size < pos && !verity)
1420 pagecache_isize_extended(inode, old_size, pos);
1421
1422 if (size_changed) {
1423 ret2 = ext4_mark_inode_dirty(handle, inode);
1424 if (!ret)
1425 ret = ret2;
1426 }
1427
1428 if (pos + len > inode->i_size && !verity && ext4_can_truncate(inode))
1429 /* if we have allocated more blocks and copied
1430 * less. We will have blocks allocated outside
1431 * inode->i_size. So truncate them
1432 */
1433 ext4_orphan_add(handle, inode);
1434
1435 ret2 = ext4_journal_stop(handle);
1436 if (!ret)
1437 ret = ret2;
1438 if (pos + len > inode->i_size && !verity) {
1439 ext4_truncate_failed_write(inode);
1440 /*
1441 * If truncate failed early the inode might still be
1442 * on the orphan list; we need to make sure the inode
1443 * is removed from the orphan list in that case.
1444 */
1445 if (inode->i_nlink)
1446 ext4_orphan_del(NULL, inode);
1447 }
1448
1449 return ret ? ret : copied;
1450}
1451
1452/*
1453 * Reserve space for a single cluster
1454 */
1455static int ext4_da_reserve_space(struct inode *inode)
1456{
1457 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1458 struct ext4_inode_info *ei = EXT4_I(inode);
1459 int ret;
1460
1461 /*
1462 * We will charge metadata quota at writeout time; this saves
1463 * us from metadata over-estimation, though we may go over by
1464 * a small amount in the end. Here we just reserve for data.
1465 */
1466 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1467 if (ret)
1468 return ret;
1469
1470 spin_lock(&ei->i_block_reservation_lock);
1471 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1472 spin_unlock(&ei->i_block_reservation_lock);
1473 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1474 return -ENOSPC;
1475 }
1476 ei->i_reserved_data_blocks++;
1477 trace_ext4_da_reserve_space(inode);
1478 spin_unlock(&ei->i_block_reservation_lock);
1479
1480 return 0; /* success */
1481}
1482
1483void ext4_da_release_space(struct inode *inode, int to_free)
1484{
1485 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1486 struct ext4_inode_info *ei = EXT4_I(inode);
1487
1488 if (!to_free)
1489 return; /* Nothing to release, exit */
1490
1491 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1492
1493 trace_ext4_da_release_space(inode, to_free);
1494 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1495 /*
1496 * if there aren't enough reserved blocks, then the
1497 * counter is messed up somewhere. Since this
1498 * function is called from invalidate page, it's
1499 * harmless to return without any action.
1500 */
1501 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1502 "ino %lu, to_free %d with only %d reserved "
1503 "data blocks", inode->i_ino, to_free,
1504 ei->i_reserved_data_blocks);
1505 WARN_ON(1);
1506 to_free = ei->i_reserved_data_blocks;
1507 }
1508 ei->i_reserved_data_blocks -= to_free;
1509
1510 /* update fs dirty data blocks counter */
1511 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1512
1513 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1514
1515 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1516}
1517
1518/*
1519 * Delayed allocation stuff
1520 */
1521
1522struct mpage_da_data {
1523 /* These are input fields for ext4_do_writepages() */
1524 struct inode *inode;
1525 struct writeback_control *wbc;
1526 unsigned int can_map:1; /* Can writepages call map blocks? */
1527
1528 /* These are internal state of ext4_do_writepages() */
1529 pgoff_t first_page; /* The first page to write */
1530 pgoff_t next_page; /* Current page to examine */
1531 pgoff_t last_page; /* Last page to examine */
1532 /*
1533 * Extent to map - this can be after first_page because that can be
1534 * fully mapped. We somewhat abuse m_flags to store whether the extent
1535 * is delalloc or unwritten.
1536 */
1537 struct ext4_map_blocks map;
1538 struct ext4_io_submit io_submit; /* IO submission data */
1539 unsigned int do_map:1;
1540 unsigned int scanned_until_end:1;
1541 unsigned int journalled_more_data:1;
1542};
1543
1544static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1545 bool invalidate)
1546{
1547 unsigned nr, i;
1548 pgoff_t index, end;
1549 struct folio_batch fbatch;
1550 struct inode *inode = mpd->inode;
1551 struct address_space *mapping = inode->i_mapping;
1552
1553 /* This is necessary when next_page == 0. */
1554 if (mpd->first_page >= mpd->next_page)
1555 return;
1556
1557 mpd->scanned_until_end = 0;
1558 index = mpd->first_page;
1559 end = mpd->next_page - 1;
1560 if (invalidate) {
1561 ext4_lblk_t start, last;
1562 start = index << (PAGE_SHIFT - inode->i_blkbits);
1563 last = end << (PAGE_SHIFT - inode->i_blkbits);
1564
1565 /*
1566 * avoid racing with extent status tree scans made by
1567 * ext4_insert_delayed_block()
1568 */
1569 down_write(&EXT4_I(inode)->i_data_sem);
1570 ext4_es_remove_extent(inode, start, last - start + 1);
1571 up_write(&EXT4_I(inode)->i_data_sem);
1572 }
1573
1574 folio_batch_init(&fbatch);
1575 while (index <= end) {
1576 nr = filemap_get_folios(mapping, &index, end, &fbatch);
1577 if (nr == 0)
1578 break;
1579 for (i = 0; i < nr; i++) {
1580 struct folio *folio = fbatch.folios[i];
1581
1582 if (folio->index < mpd->first_page)
1583 continue;
1584 if (folio_next_index(folio) - 1 > end)
1585 continue;
1586 BUG_ON(!folio_test_locked(folio));
1587 BUG_ON(folio_test_writeback(folio));
1588 if (invalidate) {
1589 if (folio_mapped(folio))
1590 folio_clear_dirty_for_io(folio);
1591 block_invalidate_folio(folio, 0,
1592 folio_size(folio));
1593 folio_clear_uptodate(folio);
1594 }
1595 folio_unlock(folio);
1596 }
1597 folio_batch_release(&fbatch);
1598 }
1599}
1600
1601static void ext4_print_free_blocks(struct inode *inode)
1602{
1603 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1604 struct super_block *sb = inode->i_sb;
1605 struct ext4_inode_info *ei = EXT4_I(inode);
1606
1607 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1608 EXT4_C2B(EXT4_SB(inode->i_sb),
1609 ext4_count_free_clusters(sb)));
1610 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1611 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1612 (long long) EXT4_C2B(EXT4_SB(sb),
1613 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1614 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1615 (long long) EXT4_C2B(EXT4_SB(sb),
1616 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1617 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1618 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1619 ei->i_reserved_data_blocks);
1620 return;
1621}
1622
1623/*
1624 * ext4_insert_delayed_block - adds a delayed block to the extents status
1625 * tree, incrementing the reserved cluster/block
1626 * count or making a pending reservation
1627 * where needed
1628 *
1629 * @inode - file containing the newly added block
1630 * @lblk - logical block to be added
1631 *
1632 * Returns 0 on success, negative error code on failure.
1633 */
1634static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1635{
1636 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1637 int ret;
1638 bool allocated = false;
1639
1640 /*
1641 * If the cluster containing lblk is shared with a delayed,
1642 * written, or unwritten extent in a bigalloc file system, it's
1643 * already been accounted for and does not need to be reserved.
1644 * A pending reservation must be made for the cluster if it's
1645 * shared with a written or unwritten extent and doesn't already
1646 * have one. Written and unwritten extents can be purged from the
1647 * extents status tree if the system is under memory pressure, so
1648 * it's necessary to examine the extent tree if a search of the
1649 * extents status tree doesn't get a match.
1650 */
1651 if (sbi->s_cluster_ratio == 1) {
1652 ret = ext4_da_reserve_space(inode);
1653 if (ret != 0) /* ENOSPC */
1654 return ret;
1655 } else { /* bigalloc */
1656 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1657 if (!ext4_es_scan_clu(inode,
1658 &ext4_es_is_mapped, lblk)) {
1659 ret = ext4_clu_mapped(inode,
1660 EXT4_B2C(sbi, lblk));
1661 if (ret < 0)
1662 return ret;
1663 if (ret == 0) {
1664 ret = ext4_da_reserve_space(inode);
1665 if (ret != 0) /* ENOSPC */
1666 return ret;
1667 } else {
1668 allocated = true;
1669 }
1670 } else {
1671 allocated = true;
1672 }
1673 }
1674 }
1675
1676 ext4_es_insert_delayed_block(inode, lblk, allocated);
1677 return 0;
1678}
1679
1680/*
1681 * This function is grabs code from the very beginning of
1682 * ext4_map_blocks, but assumes that the caller is from delayed write
1683 * time. This function looks up the requested blocks and sets the
1684 * buffer delay bit under the protection of i_data_sem.
1685 */
1686static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1687 struct ext4_map_blocks *map,
1688 struct buffer_head *bh)
1689{
1690 struct extent_status es;
1691 int retval;
1692 sector_t invalid_block = ~((sector_t) 0xffff);
1693#ifdef ES_AGGRESSIVE_TEST
1694 struct ext4_map_blocks orig_map;
1695
1696 memcpy(&orig_map, map, sizeof(*map));
1697#endif
1698
1699 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1700 invalid_block = ~0;
1701
1702 map->m_flags = 0;
1703 ext_debug(inode, "max_blocks %u, logical block %lu\n", map->m_len,
1704 (unsigned long) map->m_lblk);
1705
1706 /* Lookup extent status tree firstly */
1707 if (ext4_es_lookup_extent(inode, iblock, NULL, &es)) {
1708 if (ext4_es_is_hole(&es))
1709 goto add_delayed;
1710
1711 /*
1712 * Delayed extent could be allocated by fallocate.
1713 * So we need to check it.
1714 */
1715 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1716 map_bh(bh, inode->i_sb, invalid_block);
1717 set_buffer_new(bh);
1718 set_buffer_delay(bh);
1719 return 0;
1720 }
1721
1722 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1723 retval = es.es_len - (iblock - es.es_lblk);
1724 if (retval > map->m_len)
1725 retval = map->m_len;
1726 map->m_len = retval;
1727 if (ext4_es_is_written(&es))
1728 map->m_flags |= EXT4_MAP_MAPPED;
1729 else if (ext4_es_is_unwritten(&es))
1730 map->m_flags |= EXT4_MAP_UNWRITTEN;
1731 else
1732 BUG();
1733
1734#ifdef ES_AGGRESSIVE_TEST
1735 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1736#endif
1737 return retval;
1738 }
1739
1740 /*
1741 * Try to see if we can get the block without requesting a new
1742 * file system block.
1743 */
1744 down_read(&EXT4_I(inode)->i_data_sem);
1745 if (ext4_has_inline_data(inode))
1746 retval = 0;
1747 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1748 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1749 else
1750 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1751 if (retval < 0) {
1752 up_read(&EXT4_I(inode)->i_data_sem);
1753 return retval;
1754 }
1755 if (retval > 0) {
1756 unsigned int status;
1757
1758 if (unlikely(retval != map->m_len)) {
1759 ext4_warning(inode->i_sb,
1760 "ES len assertion failed for inode "
1761 "%lu: retval %d != map->m_len %d",
1762 inode->i_ino, retval, map->m_len);
1763 WARN_ON(1);
1764 }
1765
1766 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1767 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1768 ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1769 map->m_pblk, status);
1770 up_read(&EXT4_I(inode)->i_data_sem);
1771 return retval;
1772 }
1773 up_read(&EXT4_I(inode)->i_data_sem);
1774
1775add_delayed:
1776 down_write(&EXT4_I(inode)->i_data_sem);
1777 retval = ext4_insert_delayed_block(inode, map->m_lblk);
1778 up_write(&EXT4_I(inode)->i_data_sem);
1779 if (retval)
1780 return retval;
1781
1782 map_bh(bh, inode->i_sb, invalid_block);
1783 set_buffer_new(bh);
1784 set_buffer_delay(bh);
1785 return retval;
1786}
1787
1788/*
1789 * This is a special get_block_t callback which is used by
1790 * ext4_da_write_begin(). It will either return mapped block or
1791 * reserve space for a single block.
1792 *
1793 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1794 * We also have b_blocknr = -1 and b_bdev initialized properly
1795 *
1796 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1797 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1798 * initialized properly.
1799 */
1800int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1801 struct buffer_head *bh, int create)
1802{
1803 struct ext4_map_blocks map;
1804 int ret = 0;
1805
1806 BUG_ON(create == 0);
1807 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1808
1809 map.m_lblk = iblock;
1810 map.m_len = 1;
1811
1812 /*
1813 * first, we need to know whether the block is allocated already
1814 * preallocated blocks are unmapped but should treated
1815 * the same as allocated blocks.
1816 */
1817 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1818 if (ret <= 0)
1819 return ret;
1820
1821 map_bh(bh, inode->i_sb, map.m_pblk);
1822 ext4_update_bh_state(bh, map.m_flags);
1823
1824 if (buffer_unwritten(bh)) {
1825 /* A delayed write to unwritten bh should be marked
1826 * new and mapped. Mapped ensures that we don't do
1827 * get_block multiple times when we write to the same
1828 * offset and new ensures that we do proper zero out
1829 * for partial write.
1830 */
1831 set_buffer_new(bh);
1832 set_buffer_mapped(bh);
1833 }
1834 return 0;
1835}
1836
1837static void mpage_folio_done(struct mpage_da_data *mpd, struct folio *folio)
1838{
1839 mpd->first_page += folio_nr_pages(folio);
1840 folio_unlock(folio);
1841}
1842
1843static int mpage_submit_folio(struct mpage_da_data *mpd, struct folio *folio)
1844{
1845 size_t len;
1846 loff_t size;
1847 int err;
1848
1849 BUG_ON(folio->index != mpd->first_page);
1850 folio_clear_dirty_for_io(folio);
1851 /*
1852 * We have to be very careful here! Nothing protects writeback path
1853 * against i_size changes and the page can be writeably mapped into
1854 * page tables. So an application can be growing i_size and writing
1855 * data through mmap while writeback runs. folio_clear_dirty_for_io()
1856 * write-protects our page in page tables and the page cannot get
1857 * written to again until we release folio lock. So only after
1858 * folio_clear_dirty_for_io() we are safe to sample i_size for
1859 * ext4_bio_write_folio() to zero-out tail of the written page. We rely
1860 * on the barrier provided by folio_test_clear_dirty() in
1861 * folio_clear_dirty_for_io() to make sure i_size is really sampled only
1862 * after page tables are updated.
1863 */
1864 size = i_size_read(mpd->inode);
1865 len = folio_size(folio);
1866 if (folio_pos(folio) + len > size &&
1867 !ext4_verity_in_progress(mpd->inode))
1868 len = size & ~PAGE_MASK;
1869 err = ext4_bio_write_folio(&mpd->io_submit, folio, len);
1870 if (!err)
1871 mpd->wbc->nr_to_write--;
1872
1873 return err;
1874}
1875
1876#define BH_FLAGS (BIT(BH_Unwritten) | BIT(BH_Delay))
1877
1878/*
1879 * mballoc gives us at most this number of blocks...
1880 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
1881 * The rest of mballoc seems to handle chunks up to full group size.
1882 */
1883#define MAX_WRITEPAGES_EXTENT_LEN 2048
1884
1885/*
1886 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
1887 *
1888 * @mpd - extent of blocks
1889 * @lblk - logical number of the block in the file
1890 * @bh - buffer head we want to add to the extent
1891 *
1892 * The function is used to collect contig. blocks in the same state. If the
1893 * buffer doesn't require mapping for writeback and we haven't started the
1894 * extent of buffers to map yet, the function returns 'true' immediately - the
1895 * caller can write the buffer right away. Otherwise the function returns true
1896 * if the block has been added to the extent, false if the block couldn't be
1897 * added.
1898 */
1899static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
1900 struct buffer_head *bh)
1901{
1902 struct ext4_map_blocks *map = &mpd->map;
1903
1904 /* Buffer that doesn't need mapping for writeback? */
1905 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
1906 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
1907 /* So far no extent to map => we write the buffer right away */
1908 if (map->m_len == 0)
1909 return true;
1910 return false;
1911 }
1912
1913 /* First block in the extent? */
1914 if (map->m_len == 0) {
1915 /* We cannot map unless handle is started... */
1916 if (!mpd->do_map)
1917 return false;
1918 map->m_lblk = lblk;
1919 map->m_len = 1;
1920 map->m_flags = bh->b_state & BH_FLAGS;
1921 return true;
1922 }
1923
1924 /* Don't go larger than mballoc is willing to allocate */
1925 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
1926 return false;
1927
1928 /* Can we merge the block to our big extent? */
1929 if (lblk == map->m_lblk + map->m_len &&
1930 (bh->b_state & BH_FLAGS) == map->m_flags) {
1931 map->m_len++;
1932 return true;
1933 }
1934 return false;
1935}
1936
1937/*
1938 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
1939 *
1940 * @mpd - extent of blocks for mapping
1941 * @head - the first buffer in the page
1942 * @bh - buffer we should start processing from
1943 * @lblk - logical number of the block in the file corresponding to @bh
1944 *
1945 * Walk through page buffers from @bh upto @head (exclusive) and either submit
1946 * the page for IO if all buffers in this page were mapped and there's no
1947 * accumulated extent of buffers to map or add buffers in the page to the
1948 * extent of buffers to map. The function returns 1 if the caller can continue
1949 * by processing the next page, 0 if it should stop adding buffers to the
1950 * extent to map because we cannot extend it anymore. It can also return value
1951 * < 0 in case of error during IO submission.
1952 */
1953static int mpage_process_page_bufs(struct mpage_da_data *mpd,
1954 struct buffer_head *head,
1955 struct buffer_head *bh,
1956 ext4_lblk_t lblk)
1957{
1958 struct inode *inode = mpd->inode;
1959 int err;
1960 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
1961 >> inode->i_blkbits;
1962
1963 if (ext4_verity_in_progress(inode))
1964 blocks = EXT_MAX_BLOCKS;
1965
1966 do {
1967 BUG_ON(buffer_locked(bh));
1968
1969 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
1970 /* Found extent to map? */
1971 if (mpd->map.m_len)
1972 return 0;
1973 /* Buffer needs mapping and handle is not started? */
1974 if (!mpd->do_map)
1975 return 0;
1976 /* Everything mapped so far and we hit EOF */
1977 break;
1978 }
1979 } while (lblk++, (bh = bh->b_this_page) != head);
1980 /* So far everything mapped? Submit the page for IO. */
1981 if (mpd->map.m_len == 0) {
1982 err = mpage_submit_folio(mpd, head->b_folio);
1983 if (err < 0)
1984 return err;
1985 mpage_folio_done(mpd, head->b_folio);
1986 }
1987 if (lblk >= blocks) {
1988 mpd->scanned_until_end = 1;
1989 return 0;
1990 }
1991 return 1;
1992}
1993
1994/*
1995 * mpage_process_folio - update folio buffers corresponding to changed extent
1996 * and may submit fully mapped page for IO
1997 * @mpd: description of extent to map, on return next extent to map
1998 * @folio: Contains these buffers.
1999 * @m_lblk: logical block mapping.
2000 * @m_pblk: corresponding physical mapping.
2001 * @map_bh: determines on return whether this page requires any further
2002 * mapping or not.
2003 *
2004 * Scan given folio buffers corresponding to changed extent and update buffer
2005 * state according to new extent state.
2006 * We map delalloc buffers to their physical location, clear unwritten bits.
2007 * If the given folio is not fully mapped, we update @mpd to the next extent in
2008 * the given folio that needs mapping & return @map_bh as true.
2009 */
2010static int mpage_process_folio(struct mpage_da_data *mpd, struct folio *folio,
2011 ext4_lblk_t *m_lblk, ext4_fsblk_t *m_pblk,
2012 bool *map_bh)
2013{
2014 struct buffer_head *head, *bh;
2015 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2016 ext4_lblk_t lblk = *m_lblk;
2017 ext4_fsblk_t pblock = *m_pblk;
2018 int err = 0;
2019 int blkbits = mpd->inode->i_blkbits;
2020 ssize_t io_end_size = 0;
2021 struct ext4_io_end_vec *io_end_vec = ext4_last_io_end_vec(io_end);
2022
2023 bh = head = folio_buffers(folio);
2024 do {
2025 if (lblk < mpd->map.m_lblk)
2026 continue;
2027 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2028 /*
2029 * Buffer after end of mapped extent.
2030 * Find next buffer in the folio to map.
2031 */
2032 mpd->map.m_len = 0;
2033 mpd->map.m_flags = 0;
2034 io_end_vec->size += io_end_size;
2035
2036 err = mpage_process_page_bufs(mpd, head, bh, lblk);
2037 if (err > 0)
2038 err = 0;
2039 if (!err && mpd->map.m_len && mpd->map.m_lblk > lblk) {
2040 io_end_vec = ext4_alloc_io_end_vec(io_end);
2041 if (IS_ERR(io_end_vec)) {
2042 err = PTR_ERR(io_end_vec);
2043 goto out;
2044 }
2045 io_end_vec->offset = (loff_t)mpd->map.m_lblk << blkbits;
2046 }
2047 *map_bh = true;
2048 goto out;
2049 }
2050 if (buffer_delay(bh)) {
2051 clear_buffer_delay(bh);
2052 bh->b_blocknr = pblock++;
2053 }
2054 clear_buffer_unwritten(bh);
2055 io_end_size += (1 << blkbits);
2056 } while (lblk++, (bh = bh->b_this_page) != head);
2057
2058 io_end_vec->size += io_end_size;
2059 *map_bh = false;
2060out:
2061 *m_lblk = lblk;
2062 *m_pblk = pblock;
2063 return err;
2064}
2065
2066/*
2067 * mpage_map_buffers - update buffers corresponding to changed extent and
2068 * submit fully mapped pages for IO
2069 *
2070 * @mpd - description of extent to map, on return next extent to map
2071 *
2072 * Scan buffers corresponding to changed extent (we expect corresponding pages
2073 * to be already locked) and update buffer state according to new extent state.
2074 * We map delalloc buffers to their physical location, clear unwritten bits,
2075 * and mark buffers as uninit when we perform writes to unwritten extents
2076 * and do extent conversion after IO is finished. If the last page is not fully
2077 * mapped, we update @map to the next extent in the last page that needs
2078 * mapping. Otherwise we submit the page for IO.
2079 */
2080static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2081{
2082 struct folio_batch fbatch;
2083 unsigned nr, i;
2084 struct inode *inode = mpd->inode;
2085 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2086 pgoff_t start, end;
2087 ext4_lblk_t lblk;
2088 ext4_fsblk_t pblock;
2089 int err;
2090 bool map_bh = false;
2091
2092 start = mpd->map.m_lblk >> bpp_bits;
2093 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2094 lblk = start << bpp_bits;
2095 pblock = mpd->map.m_pblk;
2096
2097 folio_batch_init(&fbatch);
2098 while (start <= end) {
2099 nr = filemap_get_folios(inode->i_mapping, &start, end, &fbatch);
2100 if (nr == 0)
2101 break;
2102 for (i = 0; i < nr; i++) {
2103 struct folio *folio = fbatch.folios[i];
2104
2105 err = mpage_process_folio(mpd, folio, &lblk, &pblock,
2106 &map_bh);
2107 /*
2108 * If map_bh is true, means page may require further bh
2109 * mapping, or maybe the page was submitted for IO.
2110 * So we return to call further extent mapping.
2111 */
2112 if (err < 0 || map_bh)
2113 goto out;
2114 /* Page fully mapped - let IO run! */
2115 err = mpage_submit_folio(mpd, folio);
2116 if (err < 0)
2117 goto out;
2118 mpage_folio_done(mpd, folio);
2119 }
2120 folio_batch_release(&fbatch);
2121 }
2122 /* Extent fully mapped and matches with page boundary. We are done. */
2123 mpd->map.m_len = 0;
2124 mpd->map.m_flags = 0;
2125 return 0;
2126out:
2127 folio_batch_release(&fbatch);
2128 return err;
2129}
2130
2131static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2132{
2133 struct inode *inode = mpd->inode;
2134 struct ext4_map_blocks *map = &mpd->map;
2135 int get_blocks_flags;
2136 int err, dioread_nolock;
2137
2138 trace_ext4_da_write_pages_extent(inode, map);
2139 /*
2140 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2141 * to convert an unwritten extent to be initialized (in the case
2142 * where we have written into one or more preallocated blocks). It is
2143 * possible that we're going to need more metadata blocks than
2144 * previously reserved. However we must not fail because we're in
2145 * writeback and there is nothing we can do about it so it might result
2146 * in data loss. So use reserved blocks to allocate metadata if
2147 * possible.
2148 *
2149 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2150 * the blocks in question are delalloc blocks. This indicates
2151 * that the blocks and quotas has already been checked when
2152 * the data was copied into the page cache.
2153 */
2154 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2155 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2156 EXT4_GET_BLOCKS_IO_SUBMIT;
2157 dioread_nolock = ext4_should_dioread_nolock(inode);
2158 if (dioread_nolock)
2159 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2160 if (map->m_flags & BIT(BH_Delay))
2161 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2162
2163 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2164 if (err < 0)
2165 return err;
2166 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2167 if (!mpd->io_submit.io_end->handle &&
2168 ext4_handle_valid(handle)) {
2169 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2170 handle->h_rsv_handle = NULL;
2171 }
2172 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2173 }
2174
2175 BUG_ON(map->m_len == 0);
2176 return 0;
2177}
2178
2179/*
2180 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2181 * mpd->len and submit pages underlying it for IO
2182 *
2183 * @handle - handle for journal operations
2184 * @mpd - extent to map
2185 * @give_up_on_write - we set this to true iff there is a fatal error and there
2186 * is no hope of writing the data. The caller should discard
2187 * dirty pages to avoid infinite loops.
2188 *
2189 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2190 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2191 * them to initialized or split the described range from larger unwritten
2192 * extent. Note that we need not map all the described range since allocation
2193 * can return less blocks or the range is covered by more unwritten extents. We
2194 * cannot map more because we are limited by reserved transaction credits. On
2195 * the other hand we always make sure that the last touched page is fully
2196 * mapped so that it can be written out (and thus forward progress is
2197 * guaranteed). After mapping we submit all mapped pages for IO.
2198 */
2199static int mpage_map_and_submit_extent(handle_t *handle,
2200 struct mpage_da_data *mpd,
2201 bool *give_up_on_write)
2202{
2203 struct inode *inode = mpd->inode;
2204 struct ext4_map_blocks *map = &mpd->map;
2205 int err;
2206 loff_t disksize;
2207 int progress = 0;
2208 ext4_io_end_t *io_end = mpd->io_submit.io_end;
2209 struct ext4_io_end_vec *io_end_vec;
2210
2211 io_end_vec = ext4_alloc_io_end_vec(io_end);
2212 if (IS_ERR(io_end_vec))
2213 return PTR_ERR(io_end_vec);
2214 io_end_vec->offset = ((loff_t)map->m_lblk) << inode->i_blkbits;
2215 do {
2216 err = mpage_map_one_extent(handle, mpd);
2217 if (err < 0) {
2218 struct super_block *sb = inode->i_sb;
2219
2220 if (ext4_forced_shutdown(sb))
2221 goto invalidate_dirty_pages;
2222 /*
2223 * Let the uper layers retry transient errors.
2224 * In the case of ENOSPC, if ext4_count_free_blocks()
2225 * is non-zero, a commit should free up blocks.
2226 */
2227 if ((err == -ENOMEM) ||
2228 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2229 if (progress)
2230 goto update_disksize;
2231 return err;
2232 }
2233 ext4_msg(sb, KERN_CRIT,
2234 "Delayed block allocation failed for "
2235 "inode %lu at logical offset %llu with"
2236 " max blocks %u with error %d",
2237 inode->i_ino,
2238 (unsigned long long)map->m_lblk,
2239 (unsigned)map->m_len, -err);
2240 ext4_msg(sb, KERN_CRIT,
2241 "This should not happen!! Data will "
2242 "be lost\n");
2243 if (err == -ENOSPC)
2244 ext4_print_free_blocks(inode);
2245 invalidate_dirty_pages:
2246 *give_up_on_write = true;
2247 return err;
2248 }
2249 progress = 1;
2250 /*
2251 * Update buffer state, submit mapped pages, and get us new
2252 * extent to map
2253 */
2254 err = mpage_map_and_submit_buffers(mpd);
2255 if (err < 0)
2256 goto update_disksize;
2257 } while (map->m_len);
2258
2259update_disksize:
2260 /*
2261 * Update on-disk size after IO is submitted. Races with
2262 * truncate are avoided by checking i_size under i_data_sem.
2263 */
2264 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2265 if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
2266 int err2;
2267 loff_t i_size;
2268
2269 down_write(&EXT4_I(inode)->i_data_sem);
2270 i_size = i_size_read(inode);
2271 if (disksize > i_size)
2272 disksize = i_size;
2273 if (disksize > EXT4_I(inode)->i_disksize)
2274 EXT4_I(inode)->i_disksize = disksize;
2275 up_write(&EXT4_I(inode)->i_data_sem);
2276 err2 = ext4_mark_inode_dirty(handle, inode);
2277 if (err2) {
2278 ext4_error_err(inode->i_sb, -err2,
2279 "Failed to mark inode %lu dirty",
2280 inode->i_ino);
2281 }
2282 if (!err)
2283 err = err2;
2284 }
2285 return err;
2286}
2287
2288/*
2289 * Calculate the total number of credits to reserve for one writepages
2290 * iteration. This is called from ext4_writepages(). We map an extent of
2291 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2292 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2293 * bpp - 1 blocks in bpp different extents.
2294 */
2295static int ext4_da_writepages_trans_blocks(struct inode *inode)
2296{
2297 int bpp = ext4_journal_blocks_per_page(inode);
2298
2299 return ext4_meta_trans_blocks(inode,
2300 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2301}
2302
2303static int ext4_journal_folio_buffers(handle_t *handle, struct folio *folio,
2304 size_t len)
2305{
2306 struct buffer_head *page_bufs = folio_buffers(folio);
2307 struct inode *inode = folio->mapping->host;
2308 int ret, err;
2309
2310 ret = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2311 NULL, do_journal_get_write_access);
2312 err = ext4_walk_page_buffers(handle, inode, page_bufs, 0, len,
2313 NULL, write_end_fn);
2314 if (ret == 0)
2315 ret = err;
2316 err = ext4_jbd2_inode_add_write(handle, inode, folio_pos(folio), len);
2317 if (ret == 0)
2318 ret = err;
2319 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2320
2321 return ret;
2322}
2323
2324static int mpage_journal_page_buffers(handle_t *handle,
2325 struct mpage_da_data *mpd,
2326 struct folio *folio)
2327{
2328 struct inode *inode = mpd->inode;
2329 loff_t size = i_size_read(inode);
2330 size_t len = folio_size(folio);
2331
2332 folio_clear_checked(folio);
2333 mpd->wbc->nr_to_write--;
2334
2335 if (folio_pos(folio) + len > size &&
2336 !ext4_verity_in_progress(inode))
2337 len = size - folio_pos(folio);
2338
2339 return ext4_journal_folio_buffers(handle, folio, len);
2340}
2341
2342/*
2343 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2344 * needing mapping, submit mapped pages
2345 *
2346 * @mpd - where to look for pages
2347 *
2348 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2349 * IO immediately. If we cannot map blocks, we submit just already mapped
2350 * buffers in the page for IO and keep page dirty. When we can map blocks and
2351 * we find a page which isn't mapped we start accumulating extent of buffers
2352 * underlying these pages that needs mapping (formed by either delayed or
2353 * unwritten buffers). We also lock the pages containing these buffers. The
2354 * extent found is returned in @mpd structure (starting at mpd->lblk with
2355 * length mpd->len blocks).
2356 *
2357 * Note that this function can attach bios to one io_end structure which are
2358 * neither logically nor physically contiguous. Although it may seem as an
2359 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2360 * case as we need to track IO to all buffers underlying a page in one io_end.
2361 */
2362static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2363{
2364 struct address_space *mapping = mpd->inode->i_mapping;
2365 struct folio_batch fbatch;
2366 unsigned int nr_folios;
2367 pgoff_t index = mpd->first_page;
2368 pgoff_t end = mpd->last_page;
2369 xa_mark_t tag;
2370 int i, err = 0;
2371 int blkbits = mpd->inode->i_blkbits;
2372 ext4_lblk_t lblk;
2373 struct buffer_head *head;
2374 handle_t *handle = NULL;
2375 int bpp = ext4_journal_blocks_per_page(mpd->inode);
2376
2377 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2378 tag = PAGECACHE_TAG_TOWRITE;
2379 else
2380 tag = PAGECACHE_TAG_DIRTY;
2381
2382 mpd->map.m_len = 0;
2383 mpd->next_page = index;
2384 if (ext4_should_journal_data(mpd->inode)) {
2385 handle = ext4_journal_start(mpd->inode, EXT4_HT_WRITE_PAGE,
2386 bpp);
2387 if (IS_ERR(handle))
2388 return PTR_ERR(handle);
2389 }
2390 folio_batch_init(&fbatch);
2391 while (index <= end) {
2392 nr_folios = filemap_get_folios_tag(mapping, &index, end,
2393 tag, &fbatch);
2394 if (nr_folios == 0)
2395 break;
2396
2397 for (i = 0; i < nr_folios; i++) {
2398 struct folio *folio = fbatch.folios[i];
2399
2400 /*
2401 * Accumulated enough dirty pages? This doesn't apply
2402 * to WB_SYNC_ALL mode. For integrity sync we have to
2403 * keep going because someone may be concurrently
2404 * dirtying pages, and we might have synced a lot of
2405 * newly appeared dirty pages, but have not synced all
2406 * of the old dirty pages.
2407 */
2408 if (mpd->wbc->sync_mode == WB_SYNC_NONE &&
2409 mpd->wbc->nr_to_write <=
2410 mpd->map.m_len >> (PAGE_SHIFT - blkbits))
2411 goto out;
2412
2413 /* If we can't merge this page, we are done. */
2414 if (mpd->map.m_len > 0 && mpd->next_page != folio->index)
2415 goto out;
2416
2417 if (handle) {
2418 err = ext4_journal_ensure_credits(handle, bpp,
2419 0);
2420 if (err < 0)
2421 goto out;
2422 }
2423
2424 folio_lock(folio);
2425 /*
2426 * If the page is no longer dirty, or its mapping no
2427 * longer corresponds to inode we are writing (which
2428 * means it has been truncated or invalidated), or the
2429 * page is already under writeback and we are not doing
2430 * a data integrity writeback, skip the page
2431 */
2432 if (!folio_test_dirty(folio) ||
2433 (folio_test_writeback(folio) &&
2434 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2435 unlikely(folio->mapping != mapping)) {
2436 folio_unlock(folio);
2437 continue;
2438 }
2439
2440 folio_wait_writeback(folio);
2441 BUG_ON(folio_test_writeback(folio));
2442
2443 /*
2444 * Should never happen but for buggy code in
2445 * other subsystems that call
2446 * set_page_dirty() without properly warning
2447 * the file system first. See [1] for more
2448 * information.
2449 *
2450 * [1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@quack2.suse.cz
2451 */
2452 if (!folio_buffers(folio)) {
2453 ext4_warning_inode(mpd->inode, "page %lu does not have buffers attached", folio->index);
2454 folio_clear_dirty(folio);
2455 folio_unlock(folio);
2456 continue;
2457 }
2458
2459 if (mpd->map.m_len == 0)
2460 mpd->first_page = folio->index;
2461 mpd->next_page = folio_next_index(folio);
2462 /*
2463 * Writeout when we cannot modify metadata is simple.
2464 * Just submit the page. For data=journal mode we
2465 * first handle writeout of the page for checkpoint and
2466 * only after that handle delayed page dirtying. This
2467 * makes sure current data is checkpointed to the final
2468 * location before possibly journalling it again which
2469 * is desirable when the page is frequently dirtied
2470 * through a pin.
2471 */
2472 if (!mpd->can_map) {
2473 err = mpage_submit_folio(mpd, folio);
2474 if (err < 0)
2475 goto out;
2476 /* Pending dirtying of journalled data? */
2477 if (folio_test_checked(folio)) {
2478 err = mpage_journal_page_buffers(handle,
2479 mpd, folio);
2480 if (err < 0)
2481 goto out;
2482 mpd->journalled_more_data = 1;
2483 }
2484 mpage_folio_done(mpd, folio);
2485 } else {
2486 /* Add all dirty buffers to mpd */
2487 lblk = ((ext4_lblk_t)folio->index) <<
2488 (PAGE_SHIFT - blkbits);
2489 head = folio_buffers(folio);
2490 err = mpage_process_page_bufs(mpd, head, head,
2491 lblk);
2492 if (err <= 0)
2493 goto out;
2494 err = 0;
2495 }
2496 }
2497 folio_batch_release(&fbatch);
2498 cond_resched();
2499 }
2500 mpd->scanned_until_end = 1;
2501 if (handle)
2502 ext4_journal_stop(handle);
2503 return 0;
2504out:
2505 folio_batch_release(&fbatch);
2506 if (handle)
2507 ext4_journal_stop(handle);
2508 return err;
2509}
2510
2511static int ext4_do_writepages(struct mpage_da_data *mpd)
2512{
2513 struct writeback_control *wbc = mpd->wbc;
2514 pgoff_t writeback_index = 0;
2515 long nr_to_write = wbc->nr_to_write;
2516 int range_whole = 0;
2517 int cycled = 1;
2518 handle_t *handle = NULL;
2519 struct inode *inode = mpd->inode;
2520 struct address_space *mapping = inode->i_mapping;
2521 int needed_blocks, rsv_blocks = 0, ret = 0;
2522 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2523 struct blk_plug plug;
2524 bool give_up_on_write = false;
2525
2526 trace_ext4_writepages(inode, wbc);
2527
2528 /*
2529 * No pages to write? This is mainly a kludge to avoid starting
2530 * a transaction for special inodes like journal inode on last iput()
2531 * because that could violate lock ordering on umount
2532 */
2533 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2534 goto out_writepages;
2535
2536 /*
2537 * If the filesystem has aborted, it is read-only, so return
2538 * right away instead of dumping stack traces later on that
2539 * will obscure the real source of the problem. We test
2540 * fs shutdown state instead of sb->s_flag's SB_RDONLY because
2541 * the latter could be true if the filesystem is mounted
2542 * read-only, and in that case, ext4_writepages should
2543 * *never* be called, so if that ever happens, we would want
2544 * the stack trace.
2545 */
2546 if (unlikely(ext4_forced_shutdown(mapping->host->i_sb))) {
2547 ret = -EROFS;
2548 goto out_writepages;
2549 }
2550
2551 /*
2552 * If we have inline data and arrive here, it means that
2553 * we will soon create the block for the 1st page, so
2554 * we'd better clear the inline data here.
2555 */
2556 if (ext4_has_inline_data(inode)) {
2557 /* Just inode will be modified... */
2558 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2559 if (IS_ERR(handle)) {
2560 ret = PTR_ERR(handle);
2561 goto out_writepages;
2562 }
2563 BUG_ON(ext4_test_inode_state(inode,
2564 EXT4_STATE_MAY_INLINE_DATA));
2565 ext4_destroy_inline_data(handle, inode);
2566 ext4_journal_stop(handle);
2567 }
2568
2569 /*
2570 * data=journal mode does not do delalloc so we just need to writeout /
2571 * journal already mapped buffers. On the other hand we need to commit
2572 * transaction to make data stable. We expect all the data to be
2573 * already in the journal (the only exception are DMA pinned pages
2574 * dirtied behind our back) so we commit transaction here and run the
2575 * writeback loop to checkpoint them. The checkpointing is not actually
2576 * necessary to make data persistent *but* quite a few places (extent
2577 * shifting operations, fsverity, ...) depend on being able to drop
2578 * pagecache pages after calling filemap_write_and_wait() and for that
2579 * checkpointing needs to happen.
2580 */
2581 if (ext4_should_journal_data(inode)) {
2582 mpd->can_map = 0;
2583 if (wbc->sync_mode == WB_SYNC_ALL)
2584 ext4_fc_commit(sbi->s_journal,
2585 EXT4_I(inode)->i_datasync_tid);
2586 }
2587 mpd->journalled_more_data = 0;
2588
2589 if (ext4_should_dioread_nolock(inode)) {
2590 /*
2591 * We may need to convert up to one extent per block in
2592 * the page and we may dirty the inode.
2593 */
2594 rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
2595 PAGE_SIZE >> inode->i_blkbits);
2596 }
2597
2598 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2599 range_whole = 1;
2600
2601 if (wbc->range_cyclic) {
2602 writeback_index = mapping->writeback_index;
2603 if (writeback_index)
2604 cycled = 0;
2605 mpd->first_page = writeback_index;
2606 mpd->last_page = -1;
2607 } else {
2608 mpd->first_page = wbc->range_start >> PAGE_SHIFT;
2609 mpd->last_page = wbc->range_end >> PAGE_SHIFT;
2610 }
2611
2612 ext4_io_submit_init(&mpd->io_submit, wbc);
2613retry:
2614 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2615 tag_pages_for_writeback(mapping, mpd->first_page,
2616 mpd->last_page);
2617 blk_start_plug(&plug);
2618
2619 /*
2620 * First writeback pages that don't need mapping - we can avoid
2621 * starting a transaction unnecessarily and also avoid being blocked
2622 * in the block layer on device congestion while having transaction
2623 * started.
2624 */
2625 mpd->do_map = 0;
2626 mpd->scanned_until_end = 0;
2627 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2628 if (!mpd->io_submit.io_end) {
2629 ret = -ENOMEM;
2630 goto unplug;
2631 }
2632 ret = mpage_prepare_extent_to_map(mpd);
2633 /* Unlock pages we didn't use */
2634 mpage_release_unused_pages(mpd, false);
2635 /* Submit prepared bio */
2636 ext4_io_submit(&mpd->io_submit);
2637 ext4_put_io_end_defer(mpd->io_submit.io_end);
2638 mpd->io_submit.io_end = NULL;
2639 if (ret < 0)
2640 goto unplug;
2641
2642 while (!mpd->scanned_until_end && wbc->nr_to_write > 0) {
2643 /* For each extent of pages we use new io_end */
2644 mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2645 if (!mpd->io_submit.io_end) {
2646 ret = -ENOMEM;
2647 break;
2648 }
2649
2650 WARN_ON_ONCE(!mpd->can_map);
2651 /*
2652 * We have two constraints: We find one extent to map and we
2653 * must always write out whole page (makes a difference when
2654 * blocksize < pagesize) so that we don't block on IO when we
2655 * try to write out the rest of the page. Journalled mode is
2656 * not supported by delalloc.
2657 */
2658 BUG_ON(ext4_should_journal_data(inode));
2659 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2660
2661 /* start a new transaction */
2662 handle = ext4_journal_start_with_reserve(inode,
2663 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2664 if (IS_ERR(handle)) {
2665 ret = PTR_ERR(handle);
2666 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2667 "%ld pages, ino %lu; err %d", __func__,
2668 wbc->nr_to_write, inode->i_ino, ret);
2669 /* Release allocated io_end */
2670 ext4_put_io_end(mpd->io_submit.io_end);
2671 mpd->io_submit.io_end = NULL;
2672 break;
2673 }
2674 mpd->do_map = 1;
2675
2676 trace_ext4_da_write_pages(inode, mpd->first_page, wbc);
2677 ret = mpage_prepare_extent_to_map(mpd);
2678 if (!ret && mpd->map.m_len)
2679 ret = mpage_map_and_submit_extent(handle, mpd,
2680 &give_up_on_write);
2681 /*
2682 * Caution: If the handle is synchronous,
2683 * ext4_journal_stop() can wait for transaction commit
2684 * to finish which may depend on writeback of pages to
2685 * complete or on page lock to be released. In that
2686 * case, we have to wait until after we have
2687 * submitted all the IO, released page locks we hold,
2688 * and dropped io_end reference (for extent conversion
2689 * to be able to complete) before stopping the handle.
2690 */
2691 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2692 ext4_journal_stop(handle);
2693 handle = NULL;
2694 mpd->do_map = 0;
2695 }
2696 /* Unlock pages we didn't use */
2697 mpage_release_unused_pages(mpd, give_up_on_write);
2698 /* Submit prepared bio */
2699 ext4_io_submit(&mpd->io_submit);
2700
2701 /*
2702 * Drop our io_end reference we got from init. We have
2703 * to be careful and use deferred io_end finishing if
2704 * we are still holding the transaction as we can
2705 * release the last reference to io_end which may end
2706 * up doing unwritten extent conversion.
2707 */
2708 if (handle) {
2709 ext4_put_io_end_defer(mpd->io_submit.io_end);
2710 ext4_journal_stop(handle);
2711 } else
2712 ext4_put_io_end(mpd->io_submit.io_end);
2713 mpd->io_submit.io_end = NULL;
2714
2715 if (ret == -ENOSPC && sbi->s_journal) {
2716 /*
2717 * Commit the transaction which would
2718 * free blocks released in the transaction
2719 * and try again
2720 */
2721 jbd2_journal_force_commit_nested(sbi->s_journal);
2722 ret = 0;
2723 continue;
2724 }
2725 /* Fatal error - ENOMEM, EIO... */
2726 if (ret)
2727 break;
2728 }
2729unplug:
2730 blk_finish_plug(&plug);
2731 if (!ret && !cycled && wbc->nr_to_write > 0) {
2732 cycled = 1;
2733 mpd->last_page = writeback_index - 1;
2734 mpd->first_page = 0;
2735 goto retry;
2736 }
2737
2738 /* Update index */
2739 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2740 /*
2741 * Set the writeback_index so that range_cyclic
2742 * mode will write it back later
2743 */
2744 mapping->writeback_index = mpd->first_page;
2745
2746out_writepages:
2747 trace_ext4_writepages_result(inode, wbc, ret,
2748 nr_to_write - wbc->nr_to_write);
2749 return ret;
2750}
2751
2752static int ext4_writepages(struct address_space *mapping,
2753 struct writeback_control *wbc)
2754{
2755 struct super_block *sb = mapping->host->i_sb;
2756 struct mpage_da_data mpd = {
2757 .inode = mapping->host,
2758 .wbc = wbc,
2759 .can_map = 1,
2760 };
2761 int ret;
2762 int alloc_ctx;
2763
2764 if (unlikely(ext4_forced_shutdown(sb)))
2765 return -EIO;
2766
2767 alloc_ctx = ext4_writepages_down_read(sb);
2768 ret = ext4_do_writepages(&mpd);
2769 /*
2770 * For data=journal writeback we could have come across pages marked
2771 * for delayed dirtying (PageChecked) which were just added to the
2772 * running transaction. Try once more to get them to stable storage.
2773 */
2774 if (!ret && mpd.journalled_more_data)
2775 ret = ext4_do_writepages(&mpd);
2776 ext4_writepages_up_read(sb, alloc_ctx);
2777
2778 return ret;
2779}
2780
2781int ext4_normal_submit_inode_data_buffers(struct jbd2_inode *jinode)
2782{
2783 struct writeback_control wbc = {
2784 .sync_mode = WB_SYNC_ALL,
2785 .nr_to_write = LONG_MAX,
2786 .range_start = jinode->i_dirty_start,
2787 .range_end = jinode->i_dirty_end,
2788 };
2789 struct mpage_da_data mpd = {
2790 .inode = jinode->i_vfs_inode,
2791 .wbc = &wbc,
2792 .can_map = 0,
2793 };
2794 return ext4_do_writepages(&mpd);
2795}
2796
2797static int ext4_dax_writepages(struct address_space *mapping,
2798 struct writeback_control *wbc)
2799{
2800 int ret;
2801 long nr_to_write = wbc->nr_to_write;
2802 struct inode *inode = mapping->host;
2803 int alloc_ctx;
2804
2805 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2806 return -EIO;
2807
2808 alloc_ctx = ext4_writepages_down_read(inode->i_sb);
2809 trace_ext4_writepages(inode, wbc);
2810
2811 ret = dax_writeback_mapping_range(mapping,
2812 EXT4_SB(inode->i_sb)->s_daxdev, wbc);
2813 trace_ext4_writepages_result(inode, wbc, ret,
2814 nr_to_write - wbc->nr_to_write);
2815 ext4_writepages_up_read(inode->i_sb, alloc_ctx);
2816 return ret;
2817}
2818
2819static int ext4_nonda_switch(struct super_block *sb)
2820{
2821 s64 free_clusters, dirty_clusters;
2822 struct ext4_sb_info *sbi = EXT4_SB(sb);
2823
2824 /*
2825 * switch to non delalloc mode if we are running low
2826 * on free block. The free block accounting via percpu
2827 * counters can get slightly wrong with percpu_counter_batch getting
2828 * accumulated on each CPU without updating global counters
2829 * Delalloc need an accurate free block accounting. So switch
2830 * to non delalloc when we are near to error range.
2831 */
2832 free_clusters =
2833 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2834 dirty_clusters =
2835 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2836 /*
2837 * Start pushing delalloc when 1/2 of free blocks are dirty.
2838 */
2839 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2840 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2841
2842 if (2 * free_clusters < 3 * dirty_clusters ||
2843 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2844 /*
2845 * free block count is less than 150% of dirty blocks
2846 * or free blocks is less than watermark
2847 */
2848 return 1;
2849 }
2850 return 0;
2851}
2852
2853static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2854 loff_t pos, unsigned len,
2855 struct page **pagep, void **fsdata)
2856{
2857 int ret, retries = 0;
2858 struct folio *folio;
2859 pgoff_t index;
2860 struct inode *inode = mapping->host;
2861
2862 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
2863 return -EIO;
2864
2865 index = pos >> PAGE_SHIFT;
2866
2867 if (ext4_nonda_switch(inode->i_sb) || ext4_verity_in_progress(inode)) {
2868 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2869 return ext4_write_begin(file, mapping, pos,
2870 len, pagep, fsdata);
2871 }
2872 *fsdata = (void *)0;
2873 trace_ext4_da_write_begin(inode, pos, len);
2874
2875 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2876 ret = ext4_da_write_inline_data_begin(mapping, inode, pos, len,
2877 pagep, fsdata);
2878 if (ret < 0)
2879 return ret;
2880 if (ret == 1)
2881 return 0;
2882 }
2883
2884retry:
2885 folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
2886 mapping_gfp_mask(mapping));
2887 if (IS_ERR(folio))
2888 return PTR_ERR(folio);
2889
2890#ifdef CONFIG_FS_ENCRYPTION
2891 ret = ext4_block_write_begin(folio, pos, len, ext4_da_get_block_prep);
2892#else
2893 ret = __block_write_begin(&folio->page, pos, len, ext4_da_get_block_prep);
2894#endif
2895 if (ret < 0) {
2896 folio_unlock(folio);
2897 folio_put(folio);
2898 /*
2899 * block_write_begin may have instantiated a few blocks
2900 * outside i_size. Trim these off again. Don't need
2901 * i_size_read because we hold inode lock.
2902 */
2903 if (pos + len > inode->i_size)
2904 ext4_truncate_failed_write(inode);
2905
2906 if (ret == -ENOSPC &&
2907 ext4_should_retry_alloc(inode->i_sb, &retries))
2908 goto retry;
2909 return ret;
2910 }
2911
2912 *pagep = &folio->page;
2913 return ret;
2914}
2915
2916/*
2917 * Check if we should update i_disksize
2918 * when write to the end of file but not require block allocation
2919 */
2920static int ext4_da_should_update_i_disksize(struct folio *folio,
2921 unsigned long offset)
2922{
2923 struct buffer_head *bh;
2924 struct inode *inode = folio->mapping->host;
2925 unsigned int idx;
2926 int i;
2927
2928 bh = folio_buffers(folio);
2929 idx = offset >> inode->i_blkbits;
2930
2931 for (i = 0; i < idx; i++)
2932 bh = bh->b_this_page;
2933
2934 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2935 return 0;
2936 return 1;
2937}
2938
2939static int ext4_da_do_write_end(struct address_space *mapping,
2940 loff_t pos, unsigned len, unsigned copied,
2941 struct folio *folio)
2942{
2943 struct inode *inode = mapping->host;
2944 loff_t old_size = inode->i_size;
2945 bool disksize_changed = false;
2946 loff_t new_i_size;
2947
2948 /*
2949 * block_write_end() will mark the inode as dirty with I_DIRTY_PAGES
2950 * flag, which all that's needed to trigger page writeback.
2951 */
2952 copied = block_write_end(NULL, mapping, pos, len, copied,
2953 &folio->page, NULL);
2954 new_i_size = pos + copied;
2955
2956 /*
2957 * It's important to update i_size while still holding folio lock,
2958 * because folio writeout could otherwise come in and zero beyond
2959 * i_size.
2960 *
2961 * Since we are holding inode lock, we are sure i_disksize <=
2962 * i_size. We also know that if i_disksize < i_size, there are
2963 * delalloc writes pending in the range up to i_size. If the end of
2964 * the current write is <= i_size, there's no need to touch
2965 * i_disksize since writeback will push i_disksize up to i_size
2966 * eventually. If the end of the current write is > i_size and
2967 * inside an allocated block which ext4_da_should_update_i_disksize()
2968 * checked, we need to update i_disksize here as certain
2969 * ext4_writepages() paths not allocating blocks and update i_disksize.
2970 */
2971 if (new_i_size > inode->i_size) {
2972 unsigned long end;
2973
2974 i_size_write(inode, new_i_size);
2975 end = (new_i_size - 1) & (PAGE_SIZE - 1);
2976 if (copied && ext4_da_should_update_i_disksize(folio, end)) {
2977 ext4_update_i_disksize(inode, new_i_size);
2978 disksize_changed = true;
2979 }
2980 }
2981
2982 folio_unlock(folio);
2983 folio_put(folio);
2984
2985 if (old_size < pos)
2986 pagecache_isize_extended(inode, old_size, pos);
2987
2988 if (disksize_changed) {
2989 handle_t *handle;
2990
2991 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
2992 if (IS_ERR(handle))
2993 return PTR_ERR(handle);
2994 ext4_mark_inode_dirty(handle, inode);
2995 ext4_journal_stop(handle);
2996 }
2997
2998 return copied;
2999}
3000
3001static int ext4_da_write_end(struct file *file,
3002 struct address_space *mapping,
3003 loff_t pos, unsigned len, unsigned copied,
3004 struct page *page, void *fsdata)
3005{
3006 struct inode *inode = mapping->host;
3007 int write_mode = (int)(unsigned long)fsdata;
3008 struct folio *folio = page_folio(page);
3009
3010 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3011 return ext4_write_end(file, mapping, pos,
3012 len, copied, &folio->page, fsdata);
3013
3014 trace_ext4_da_write_end(inode, pos, len, copied);
3015
3016 if (write_mode != CONVERT_INLINE_DATA &&
3017 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3018 ext4_has_inline_data(inode))
3019 return ext4_write_inline_data_end(inode, pos, len, copied,
3020 folio);
3021
3022 if (unlikely(copied < len) && !folio_test_uptodate(folio))
3023 copied = 0;
3024
3025 return ext4_da_do_write_end(mapping, pos, len, copied, folio);
3026}
3027
3028/*
3029 * Force all delayed allocation blocks to be allocated for a given inode.
3030 */
3031int ext4_alloc_da_blocks(struct inode *inode)
3032{
3033 trace_ext4_alloc_da_blocks(inode);
3034
3035 if (!EXT4_I(inode)->i_reserved_data_blocks)
3036 return 0;
3037
3038 /*
3039 * We do something simple for now. The filemap_flush() will
3040 * also start triggering a write of the data blocks, which is
3041 * not strictly speaking necessary (and for users of
3042 * laptop_mode, not even desirable). However, to do otherwise
3043 * would require replicating code paths in:
3044 *
3045 * ext4_writepages() ->
3046 * write_cache_pages() ---> (via passed in callback function)
3047 * __mpage_da_writepage() -->
3048 * mpage_add_bh_to_extent()
3049 * mpage_da_map_blocks()
3050 *
3051 * The problem is that write_cache_pages(), located in
3052 * mm/page-writeback.c, marks pages clean in preparation for
3053 * doing I/O, which is not desirable if we're not planning on
3054 * doing I/O at all.
3055 *
3056 * We could call write_cache_pages(), and then redirty all of
3057 * the pages by calling redirty_page_for_writepage() but that
3058 * would be ugly in the extreme. So instead we would need to
3059 * replicate parts of the code in the above functions,
3060 * simplifying them because we wouldn't actually intend to
3061 * write out the pages, but rather only collect contiguous
3062 * logical block extents, call the multi-block allocator, and
3063 * then update the buffer heads with the block allocations.
3064 *
3065 * For now, though, we'll cheat by calling filemap_flush(),
3066 * which will map the blocks, and start the I/O, but not
3067 * actually wait for the I/O to complete.
3068 */
3069 return filemap_flush(inode->i_mapping);
3070}
3071
3072/*
3073 * bmap() is special. It gets used by applications such as lilo and by
3074 * the swapper to find the on-disk block of a specific piece of data.
3075 *
3076 * Naturally, this is dangerous if the block concerned is still in the
3077 * journal. If somebody makes a swapfile on an ext4 data-journaling
3078 * filesystem and enables swap, then they may get a nasty shock when the
3079 * data getting swapped to that swapfile suddenly gets overwritten by
3080 * the original zero's written out previously to the journal and
3081 * awaiting writeback in the kernel's buffer cache.
3082 *
3083 * So, if we see any bmap calls here on a modified, data-journaled file,
3084 * take extra steps to flush any blocks which might be in the cache.
3085 */
3086static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3087{
3088 struct inode *inode = mapping->host;
3089 sector_t ret = 0;
3090
3091 inode_lock_shared(inode);
3092 /*
3093 * We can get here for an inline file via the FIBMAP ioctl
3094 */
3095 if (ext4_has_inline_data(inode))
3096 goto out;
3097
3098 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3099 (test_opt(inode->i_sb, DELALLOC) ||
3100 ext4_should_journal_data(inode))) {
3101 /*
3102 * With delalloc or journalled data we want to sync the file so
3103 * that we can make sure we allocate blocks for file and data
3104 * is in place for the user to see it
3105 */
3106 filemap_write_and_wait(mapping);
3107 }
3108
3109 ret = iomap_bmap(mapping, block, &ext4_iomap_ops);
3110
3111out:
3112 inode_unlock_shared(inode);
3113 return ret;
3114}
3115
3116static int ext4_read_folio(struct file *file, struct folio *folio)
3117{
3118 int ret = -EAGAIN;
3119 struct inode *inode = folio->mapping->host;
3120
3121 trace_ext4_read_folio(inode, folio);
3122
3123 if (ext4_has_inline_data(inode))
3124 ret = ext4_readpage_inline(inode, folio);
3125
3126 if (ret == -EAGAIN)
3127 return ext4_mpage_readpages(inode, NULL, folio);
3128
3129 return ret;
3130}
3131
3132static void ext4_readahead(struct readahead_control *rac)
3133{
3134 struct inode *inode = rac->mapping->host;
3135
3136 /* If the file has inline data, no need to do readahead. */
3137 if (ext4_has_inline_data(inode))
3138 return;
3139
3140 ext4_mpage_readpages(inode, rac, NULL);
3141}
3142
3143static void ext4_invalidate_folio(struct folio *folio, size_t offset,
3144 size_t length)
3145{
3146 trace_ext4_invalidate_folio(folio, offset, length);
3147
3148 /* No journalling happens on data buffers when this function is used */
3149 WARN_ON(folio_buffers(folio) && buffer_jbd(folio_buffers(folio)));
3150
3151 block_invalidate_folio(folio, offset, length);
3152}
3153
3154static int __ext4_journalled_invalidate_folio(struct folio *folio,
3155 size_t offset, size_t length)
3156{
3157 journal_t *journal = EXT4_JOURNAL(folio->mapping->host);
3158
3159 trace_ext4_journalled_invalidate_folio(folio, offset, length);
3160
3161 /*
3162 * If it's a full truncate we just forget about the pending dirtying
3163 */
3164 if (offset == 0 && length == folio_size(folio))
3165 folio_clear_checked(folio);
3166
3167 return jbd2_journal_invalidate_folio(journal, folio, offset, length);
3168}
3169
3170/* Wrapper for aops... */
3171static void ext4_journalled_invalidate_folio(struct folio *folio,
3172 size_t offset,
3173 size_t length)
3174{
3175 WARN_ON(__ext4_journalled_invalidate_folio(folio, offset, length) < 0);
3176}
3177
3178static bool ext4_release_folio(struct folio *folio, gfp_t wait)
3179{
3180 struct inode *inode = folio->mapping->host;
3181 journal_t *journal = EXT4_JOURNAL(inode);
3182
3183 trace_ext4_release_folio(inode, folio);
3184
3185 /* Page has dirty journalled data -> cannot release */
3186 if (folio_test_checked(folio))
3187 return false;
3188 if (journal)
3189 return jbd2_journal_try_to_free_buffers(journal, folio);
3190 else
3191 return try_to_free_buffers(folio);
3192}
3193
3194static bool ext4_inode_datasync_dirty(struct inode *inode)
3195{
3196 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3197
3198 if (journal) {
3199 if (jbd2_transaction_committed(journal,
3200 EXT4_I(inode)->i_datasync_tid))
3201 return false;
3202 if (test_opt2(inode->i_sb, JOURNAL_FAST_COMMIT))
3203 return !list_empty(&EXT4_I(inode)->i_fc_list);
3204 return true;
3205 }
3206
3207 /* Any metadata buffers to write? */
3208 if (!list_empty(&inode->i_mapping->i_private_list))
3209 return true;
3210 return inode->i_state & I_DIRTY_DATASYNC;
3211}
3212
3213static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
3214 struct ext4_map_blocks *map, loff_t offset,
3215 loff_t length, unsigned int flags)
3216{
3217 u8 blkbits = inode->i_blkbits;
3218
3219 /*
3220 * Writes that span EOF might trigger an I/O size update on completion,
3221 * so consider them to be dirty for the purpose of O_DSYNC, even if
3222 * there is no other metadata changes being made or are pending.
3223 */
3224 iomap->flags = 0;
3225 if (ext4_inode_datasync_dirty(inode) ||
3226 offset + length > i_size_read(inode))
3227 iomap->flags |= IOMAP_F_DIRTY;
3228
3229 if (map->m_flags & EXT4_MAP_NEW)
3230 iomap->flags |= IOMAP_F_NEW;
3231
3232 if (flags & IOMAP_DAX)
3233 iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev;
3234 else
3235 iomap->bdev = inode->i_sb->s_bdev;
3236 iomap->offset = (u64) map->m_lblk << blkbits;
3237 iomap->length = (u64) map->m_len << blkbits;
3238
3239 if ((map->m_flags & EXT4_MAP_MAPPED) &&
3240 !ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3241 iomap->flags |= IOMAP_F_MERGED;
3242
3243 /*
3244 * Flags passed to ext4_map_blocks() for direct I/O writes can result
3245 * in m_flags having both EXT4_MAP_MAPPED and EXT4_MAP_UNWRITTEN bits
3246 * set. In order for any allocated unwritten extents to be converted
3247 * into written extents correctly within the ->end_io() handler, we
3248 * need to ensure that the iomap->type is set appropriately. Hence, the
3249 * reason why we need to check whether the EXT4_MAP_UNWRITTEN bit has
3250 * been set first.
3251 */
3252 if (map->m_flags & EXT4_MAP_UNWRITTEN) {
3253 iomap->type = IOMAP_UNWRITTEN;
3254 iomap->addr = (u64) map->m_pblk << blkbits;
3255 if (flags & IOMAP_DAX)
3256 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3257 } else if (map->m_flags & EXT4_MAP_MAPPED) {
3258 iomap->type = IOMAP_MAPPED;
3259 iomap->addr = (u64) map->m_pblk << blkbits;
3260 if (flags & IOMAP_DAX)
3261 iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off;
3262 } else if (map->m_flags & EXT4_MAP_DELAYED) {
3263 iomap->type = IOMAP_DELALLOC;
3264 iomap->addr = IOMAP_NULL_ADDR;
3265 } else {
3266 iomap->type = IOMAP_HOLE;
3267 iomap->addr = IOMAP_NULL_ADDR;
3268 }
3269}
3270
3271static int ext4_iomap_alloc(struct inode *inode, struct ext4_map_blocks *map,
3272 unsigned int flags)
3273{
3274 handle_t *handle;
3275 u8 blkbits = inode->i_blkbits;
3276 int ret, dio_credits, m_flags = 0, retries = 0;
3277
3278 /*
3279 * Trim the mapping request to the maximum value that we can map at
3280 * once for direct I/O.
3281 */
3282 if (map->m_len > DIO_MAX_BLOCKS)
3283 map->m_len = DIO_MAX_BLOCKS;
3284 dio_credits = ext4_chunk_trans_blocks(inode, map->m_len);
3285
3286retry:
3287 /*
3288 * Either we allocate blocks and then don't get an unwritten extent, so
3289 * in that case we have reserved enough credits. Or, the blocks are
3290 * already allocated and unwritten. In that case, the extent conversion
3291 * fits into the credits as well.
3292 */
3293 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
3294 if (IS_ERR(handle))
3295 return PTR_ERR(handle);
3296
3297 /*
3298 * DAX and direct I/O are the only two operations that are currently
3299 * supported with IOMAP_WRITE.
3300 */
3301 WARN_ON(!(flags & (IOMAP_DAX | IOMAP_DIRECT)));
3302 if (flags & IOMAP_DAX)
3303 m_flags = EXT4_GET_BLOCKS_CREATE_ZERO;
3304 /*
3305 * We use i_size instead of i_disksize here because delalloc writeback
3306 * can complete at any point during the I/O and subsequently push the
3307 * i_disksize out to i_size. This could be beyond where direct I/O is
3308 * happening and thus expose allocated blocks to direct I/O reads.
3309 */
3310 else if (((loff_t)map->m_lblk << blkbits) >= i_size_read(inode))
3311 m_flags = EXT4_GET_BLOCKS_CREATE;
3312 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3313 m_flags = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3314
3315 ret = ext4_map_blocks(handle, inode, map, m_flags);
3316
3317 /*
3318 * We cannot fill holes in indirect tree based inodes as that could
3319 * expose stale data in the case of a crash. Use the magic error code
3320 * to fallback to buffered I/O.
3321 */
3322 if (!m_flags && !ret)
3323 ret = -ENOTBLK;
3324
3325 ext4_journal_stop(handle);
3326 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3327 goto retry;
3328
3329 return ret;
3330}
3331
3332
3333static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3334 unsigned flags, struct iomap *iomap, struct iomap *srcmap)
3335{
3336 int ret;
3337 struct ext4_map_blocks map;
3338 u8 blkbits = inode->i_blkbits;
3339
3340 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3341 return -EINVAL;
3342
3343 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3344 return -ERANGE;
3345
3346 /*
3347 * Calculate the first and last logical blocks respectively.
3348 */
3349 map.m_lblk = offset >> blkbits;
3350 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3351 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3352
3353 if (flags & IOMAP_WRITE) {
3354 /*
3355 * We check here if the blocks are already allocated, then we
3356 * don't need to start a journal txn and we can directly return
3357 * the mapping information. This could boost performance
3358 * especially in multi-threaded overwrite requests.
3359 */
3360 if (offset + length <= i_size_read(inode)) {
3361 ret = ext4_map_blocks(NULL, inode, &map, 0);
3362 if (ret > 0 && (map.m_flags & EXT4_MAP_MAPPED))
3363 goto out;
3364 }
3365 ret = ext4_iomap_alloc(inode, &map, flags);
3366 } else {
3367 ret = ext4_map_blocks(NULL, inode, &map, 0);
3368 }
3369
3370 if (ret < 0)
3371 return ret;
3372out:
3373 /*
3374 * When inline encryption is enabled, sometimes I/O to an encrypted file
3375 * has to be broken up to guarantee DUN contiguity. Handle this by
3376 * limiting the length of the mapping returned.
3377 */
3378 map.m_len = fscrypt_limit_io_blocks(inode, map.m_lblk, map.m_len);
3379
3380 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3381
3382 return 0;
3383}
3384
3385static int ext4_iomap_overwrite_begin(struct inode *inode, loff_t offset,
3386 loff_t length, unsigned flags, struct iomap *iomap,
3387 struct iomap *srcmap)
3388{
3389 int ret;
3390
3391 /*
3392 * Even for writes we don't need to allocate blocks, so just pretend
3393 * we are reading to save overhead of starting a transaction.
3394 */
3395 flags &= ~IOMAP_WRITE;
3396 ret = ext4_iomap_begin(inode, offset, length, flags, iomap, srcmap);
3397 WARN_ON_ONCE(!ret && iomap->type != IOMAP_MAPPED);
3398 return ret;
3399}
3400
3401static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3402 ssize_t written, unsigned flags, struct iomap *iomap)
3403{
3404 /*
3405 * Check to see whether an error occurred while writing out the data to
3406 * the allocated blocks. If so, return the magic error code so that we
3407 * fallback to buffered I/O and attempt to complete the remainder of
3408 * the I/O. Any blocks that may have been allocated in preparation for
3409 * the direct I/O will be reused during buffered I/O.
3410 */
3411 if (flags & (IOMAP_WRITE | IOMAP_DIRECT) && written == 0)
3412 return -ENOTBLK;
3413
3414 return 0;
3415}
3416
3417const struct iomap_ops ext4_iomap_ops = {
3418 .iomap_begin = ext4_iomap_begin,
3419 .iomap_end = ext4_iomap_end,
3420};
3421
3422const struct iomap_ops ext4_iomap_overwrite_ops = {
3423 .iomap_begin = ext4_iomap_overwrite_begin,
3424 .iomap_end = ext4_iomap_end,
3425};
3426
3427static int ext4_iomap_begin_report(struct inode *inode, loff_t offset,
3428 loff_t length, unsigned int flags,
3429 struct iomap *iomap, struct iomap *srcmap)
3430{
3431 int ret;
3432 struct ext4_map_blocks map;
3433 u8 blkbits = inode->i_blkbits;
3434
3435 if ((offset >> blkbits) > EXT4_MAX_LOGICAL_BLOCK)
3436 return -EINVAL;
3437
3438 if (ext4_has_inline_data(inode)) {
3439 ret = ext4_inline_data_iomap(inode, iomap);
3440 if (ret != -EAGAIN) {
3441 if (ret == 0 && offset >= iomap->length)
3442 ret = -ENOENT;
3443 return ret;
3444 }
3445 }
3446
3447 /*
3448 * Calculate the first and last logical block respectively.
3449 */
3450 map.m_lblk = offset >> blkbits;
3451 map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits,
3452 EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1;
3453
3454 /*
3455 * Fiemap callers may call for offset beyond s_bitmap_maxbytes.
3456 * So handle it here itself instead of querying ext4_map_blocks().
3457 * Since ext4_map_blocks() will warn about it and will return
3458 * -EIO error.
3459 */
3460 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3461 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3462
3463 if (offset >= sbi->s_bitmap_maxbytes) {
3464 map.m_flags = 0;
3465 goto set_iomap;
3466 }
3467 }
3468
3469 ret = ext4_map_blocks(NULL, inode, &map, 0);
3470 if (ret < 0)
3471 return ret;
3472set_iomap:
3473 ext4_set_iomap(inode, iomap, &map, offset, length, flags);
3474
3475 return 0;
3476}
3477
3478const struct iomap_ops ext4_iomap_report_ops = {
3479 .iomap_begin = ext4_iomap_begin_report,
3480};
3481
3482/*
3483 * For data=journal mode, folio should be marked dirty only when it was
3484 * writeably mapped. When that happens, it was already attached to the
3485 * transaction and marked as jbddirty (we take care of this in
3486 * ext4_page_mkwrite()). On transaction commit, we writeprotect page mappings
3487 * so we should have nothing to do here, except for the case when someone
3488 * had the page pinned and dirtied the page through this pin (e.g. by doing
3489 * direct IO to it). In that case we'd need to attach buffers here to the
3490 * transaction but we cannot due to lock ordering. We cannot just dirty the
3491 * folio and leave attached buffers clean, because the buffers' dirty state is
3492 * "definitive". We cannot just set the buffers dirty or jbddirty because all
3493 * the journalling code will explode. So what we do is to mark the folio
3494 * "pending dirty" and next time ext4_writepages() is called, attach buffers
3495 * to the transaction appropriately.
3496 */
3497static bool ext4_journalled_dirty_folio(struct address_space *mapping,
3498 struct folio *folio)
3499{
3500 WARN_ON_ONCE(!folio_buffers(folio));
3501 if (folio_maybe_dma_pinned(folio))
3502 folio_set_checked(folio);
3503 return filemap_dirty_folio(mapping, folio);
3504}
3505
3506static bool ext4_dirty_folio(struct address_space *mapping, struct folio *folio)
3507{
3508 WARN_ON_ONCE(!folio_test_locked(folio) && !folio_test_dirty(folio));
3509 WARN_ON_ONCE(!folio_buffers(folio));
3510 return block_dirty_folio(mapping, folio);
3511}
3512
3513static int ext4_iomap_swap_activate(struct swap_info_struct *sis,
3514 struct file *file, sector_t *span)
3515{
3516 return iomap_swapfile_activate(sis, file, span,
3517 &ext4_iomap_report_ops);
3518}
3519
3520static const struct address_space_operations ext4_aops = {
3521 .read_folio = ext4_read_folio,
3522 .readahead = ext4_readahead,
3523 .writepages = ext4_writepages,
3524 .write_begin = ext4_write_begin,
3525 .write_end = ext4_write_end,
3526 .dirty_folio = ext4_dirty_folio,
3527 .bmap = ext4_bmap,
3528 .invalidate_folio = ext4_invalidate_folio,
3529 .release_folio = ext4_release_folio,
3530 .direct_IO = noop_direct_IO,
3531 .migrate_folio = buffer_migrate_folio,
3532 .is_partially_uptodate = block_is_partially_uptodate,
3533 .error_remove_folio = generic_error_remove_folio,
3534 .swap_activate = ext4_iomap_swap_activate,
3535};
3536
3537static const struct address_space_operations ext4_journalled_aops = {
3538 .read_folio = ext4_read_folio,
3539 .readahead = ext4_readahead,
3540 .writepages = ext4_writepages,
3541 .write_begin = ext4_write_begin,
3542 .write_end = ext4_journalled_write_end,
3543 .dirty_folio = ext4_journalled_dirty_folio,
3544 .bmap = ext4_bmap,
3545 .invalidate_folio = ext4_journalled_invalidate_folio,
3546 .release_folio = ext4_release_folio,
3547 .direct_IO = noop_direct_IO,
3548 .migrate_folio = buffer_migrate_folio_norefs,
3549 .is_partially_uptodate = block_is_partially_uptodate,
3550 .error_remove_folio = generic_error_remove_folio,
3551 .swap_activate = ext4_iomap_swap_activate,
3552};
3553
3554static const struct address_space_operations ext4_da_aops = {
3555 .read_folio = ext4_read_folio,
3556 .readahead = ext4_readahead,
3557 .writepages = ext4_writepages,
3558 .write_begin = ext4_da_write_begin,
3559 .write_end = ext4_da_write_end,
3560 .dirty_folio = ext4_dirty_folio,
3561 .bmap = ext4_bmap,
3562 .invalidate_folio = ext4_invalidate_folio,
3563 .release_folio = ext4_release_folio,
3564 .direct_IO = noop_direct_IO,
3565 .migrate_folio = buffer_migrate_folio,
3566 .is_partially_uptodate = block_is_partially_uptodate,
3567 .error_remove_folio = generic_error_remove_folio,
3568 .swap_activate = ext4_iomap_swap_activate,
3569};
3570
3571static const struct address_space_operations ext4_dax_aops = {
3572 .writepages = ext4_dax_writepages,
3573 .direct_IO = noop_direct_IO,
3574 .dirty_folio = noop_dirty_folio,
3575 .bmap = ext4_bmap,
3576 .swap_activate = ext4_iomap_swap_activate,
3577};
3578
3579void ext4_set_aops(struct inode *inode)
3580{
3581 switch (ext4_inode_journal_mode(inode)) {
3582 case EXT4_INODE_ORDERED_DATA_MODE:
3583 case EXT4_INODE_WRITEBACK_DATA_MODE:
3584 break;
3585 case EXT4_INODE_JOURNAL_DATA_MODE:
3586 inode->i_mapping->a_ops = &ext4_journalled_aops;
3587 return;
3588 default:
3589 BUG();
3590 }
3591 if (IS_DAX(inode))
3592 inode->i_mapping->a_ops = &ext4_dax_aops;
3593 else if (test_opt(inode->i_sb, DELALLOC))
3594 inode->i_mapping->a_ops = &ext4_da_aops;
3595 else
3596 inode->i_mapping->a_ops = &ext4_aops;
3597}
3598
3599/*
3600 * Here we can't skip an unwritten buffer even though it usually reads zero
3601 * because it might have data in pagecache (eg, if called from ext4_zero_range,
3602 * ext4_punch_hole, etc) which needs to be properly zeroed out. Otherwise a
3603 * racing writeback can come later and flush the stale pagecache to disk.
3604 */
3605static int __ext4_block_zero_page_range(handle_t *handle,
3606 struct address_space *mapping, loff_t from, loff_t length)
3607{
3608 ext4_fsblk_t index = from >> PAGE_SHIFT;
3609 unsigned offset = from & (PAGE_SIZE-1);
3610 unsigned blocksize, pos;
3611 ext4_lblk_t iblock;
3612 struct inode *inode = mapping->host;
3613 struct buffer_head *bh;
3614 struct folio *folio;
3615 int err = 0;
3616
3617 folio = __filemap_get_folio(mapping, from >> PAGE_SHIFT,
3618 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
3619 mapping_gfp_constraint(mapping, ~__GFP_FS));
3620 if (IS_ERR(folio))
3621 return PTR_ERR(folio);
3622
3623 blocksize = inode->i_sb->s_blocksize;
3624
3625 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3626
3627 bh = folio_buffers(folio);
3628 if (!bh)
3629 bh = create_empty_buffers(folio, blocksize, 0);
3630
3631 /* Find the buffer that contains "offset" */
3632 pos = blocksize;
3633 while (offset >= pos) {
3634 bh = bh->b_this_page;
3635 iblock++;
3636 pos += blocksize;
3637 }
3638 if (buffer_freed(bh)) {
3639 BUFFER_TRACE(bh, "freed: skip");
3640 goto unlock;
3641 }
3642 if (!buffer_mapped(bh)) {
3643 BUFFER_TRACE(bh, "unmapped");
3644 ext4_get_block(inode, iblock, bh, 0);
3645 /* unmapped? It's a hole - nothing to do */
3646 if (!buffer_mapped(bh)) {
3647 BUFFER_TRACE(bh, "still unmapped");
3648 goto unlock;
3649 }
3650 }
3651
3652 /* Ok, it's mapped. Make sure it's up-to-date */
3653 if (folio_test_uptodate(folio))
3654 set_buffer_uptodate(bh);
3655
3656 if (!buffer_uptodate(bh)) {
3657 err = ext4_read_bh_lock(bh, 0, true);
3658 if (err)
3659 goto unlock;
3660 if (fscrypt_inode_uses_fs_layer_crypto(inode)) {
3661 /* We expect the key to be set. */
3662 BUG_ON(!fscrypt_has_encryption_key(inode));
3663 err = fscrypt_decrypt_pagecache_blocks(folio,
3664 blocksize,
3665 bh_offset(bh));
3666 if (err) {
3667 clear_buffer_uptodate(bh);
3668 goto unlock;
3669 }
3670 }
3671 }
3672 if (ext4_should_journal_data(inode)) {
3673 BUFFER_TRACE(bh, "get write access");
3674 err = ext4_journal_get_write_access(handle, inode->i_sb, bh,
3675 EXT4_JTR_NONE);
3676 if (err)
3677 goto unlock;
3678 }
3679 folio_zero_range(folio, offset, length);
3680 BUFFER_TRACE(bh, "zeroed end of block");
3681
3682 if (ext4_should_journal_data(inode)) {
3683 err = ext4_dirty_journalled_data(handle, bh);
3684 } else {
3685 err = 0;
3686 mark_buffer_dirty(bh);
3687 if (ext4_should_order_data(inode))
3688 err = ext4_jbd2_inode_add_write(handle, inode, from,
3689 length);
3690 }
3691
3692unlock:
3693 folio_unlock(folio);
3694 folio_put(folio);
3695 return err;
3696}
3697
3698/*
3699 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3700 * starting from file offset 'from'. The range to be zero'd must
3701 * be contained with in one block. If the specified range exceeds
3702 * the end of the block it will be shortened to end of the block
3703 * that corresponds to 'from'
3704 */
3705static int ext4_block_zero_page_range(handle_t *handle,
3706 struct address_space *mapping, loff_t from, loff_t length)
3707{
3708 struct inode *inode = mapping->host;
3709 unsigned offset = from & (PAGE_SIZE-1);
3710 unsigned blocksize = inode->i_sb->s_blocksize;
3711 unsigned max = blocksize - (offset & (blocksize - 1));
3712
3713 /*
3714 * correct length if it does not fall between
3715 * 'from' and the end of the block
3716 */
3717 if (length > max || length < 0)
3718 length = max;
3719
3720 if (IS_DAX(inode)) {
3721 return dax_zero_range(inode, from, length, NULL,
3722 &ext4_iomap_ops);
3723 }
3724 return __ext4_block_zero_page_range(handle, mapping, from, length);
3725}
3726
3727/*
3728 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3729 * up to the end of the block which corresponds to `from'.
3730 * This required during truncate. We need to physically zero the tail end
3731 * of that block so it doesn't yield old data if the file is later grown.
3732 */
3733static int ext4_block_truncate_page(handle_t *handle,
3734 struct address_space *mapping, loff_t from)
3735{
3736 unsigned offset = from & (PAGE_SIZE-1);
3737 unsigned length;
3738 unsigned blocksize;
3739 struct inode *inode = mapping->host;
3740
3741 /* If we are processing an encrypted inode during orphan list handling */
3742 if (IS_ENCRYPTED(inode) && !fscrypt_has_encryption_key(inode))
3743 return 0;
3744
3745 blocksize = inode->i_sb->s_blocksize;
3746 length = blocksize - (offset & (blocksize - 1));
3747
3748 return ext4_block_zero_page_range(handle, mapping, from, length);
3749}
3750
3751int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3752 loff_t lstart, loff_t length)
3753{
3754 struct super_block *sb = inode->i_sb;
3755 struct address_space *mapping = inode->i_mapping;
3756 unsigned partial_start, partial_end;
3757 ext4_fsblk_t start, end;
3758 loff_t byte_end = (lstart + length - 1);
3759 int err = 0;
3760
3761 partial_start = lstart & (sb->s_blocksize - 1);
3762 partial_end = byte_end & (sb->s_blocksize - 1);
3763
3764 start = lstart >> sb->s_blocksize_bits;
3765 end = byte_end >> sb->s_blocksize_bits;
3766
3767 /* Handle partial zero within the single block */
3768 if (start == end &&
3769 (partial_start || (partial_end != sb->s_blocksize - 1))) {
3770 err = ext4_block_zero_page_range(handle, mapping,
3771 lstart, length);
3772 return err;
3773 }
3774 /* Handle partial zero out on the start of the range */
3775 if (partial_start) {
3776 err = ext4_block_zero_page_range(handle, mapping,
3777 lstart, sb->s_blocksize);
3778 if (err)
3779 return err;
3780 }
3781 /* Handle partial zero out on the end of the range */
3782 if (partial_end != sb->s_blocksize - 1)
3783 err = ext4_block_zero_page_range(handle, mapping,
3784 byte_end - partial_end,
3785 partial_end + 1);
3786 return err;
3787}
3788
3789int ext4_can_truncate(struct inode *inode)
3790{
3791 if (S_ISREG(inode->i_mode))
3792 return 1;
3793 if (S_ISDIR(inode->i_mode))
3794 return 1;
3795 if (S_ISLNK(inode->i_mode))
3796 return !ext4_inode_is_fast_symlink(inode);
3797 return 0;
3798}
3799
3800/*
3801 * We have to make sure i_disksize gets properly updated before we truncate
3802 * page cache due to hole punching or zero range. Otherwise i_disksize update
3803 * can get lost as it may have been postponed to submission of writeback but
3804 * that will never happen after we truncate page cache.
3805 */
3806int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
3807 loff_t len)
3808{
3809 handle_t *handle;
3810 int ret;
3811
3812 loff_t size = i_size_read(inode);
3813
3814 WARN_ON(!inode_is_locked(inode));
3815 if (offset > size || offset + len < size)
3816 return 0;
3817
3818 if (EXT4_I(inode)->i_disksize >= size)
3819 return 0;
3820
3821 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
3822 if (IS_ERR(handle))
3823 return PTR_ERR(handle);
3824 ext4_update_i_disksize(inode, size);
3825 ret = ext4_mark_inode_dirty(handle, inode);
3826 ext4_journal_stop(handle);
3827
3828 return ret;
3829}
3830
3831static void ext4_wait_dax_page(struct inode *inode)
3832{
3833 filemap_invalidate_unlock(inode->i_mapping);
3834 schedule();
3835 filemap_invalidate_lock(inode->i_mapping);
3836}
3837
3838int ext4_break_layouts(struct inode *inode)
3839{
3840 struct page *page;
3841 int error;
3842
3843 if (WARN_ON_ONCE(!rwsem_is_locked(&inode->i_mapping->invalidate_lock)))
3844 return -EINVAL;
3845
3846 do {
3847 page = dax_layout_busy_page(inode->i_mapping);
3848 if (!page)
3849 return 0;
3850
3851 error = ___wait_var_event(&page->_refcount,
3852 atomic_read(&page->_refcount) == 1,
3853 TASK_INTERRUPTIBLE, 0, 0,
3854 ext4_wait_dax_page(inode));
3855 } while (error == 0);
3856
3857 return error;
3858}
3859
3860/*
3861 * ext4_punch_hole: punches a hole in a file by releasing the blocks
3862 * associated with the given offset and length
3863 *
3864 * @inode: File inode
3865 * @offset: The offset where the hole will begin
3866 * @len: The length of the hole
3867 *
3868 * Returns: 0 on success or negative on failure
3869 */
3870
3871int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3872{
3873 struct inode *inode = file_inode(file);
3874 struct super_block *sb = inode->i_sb;
3875 ext4_lblk_t first_block, stop_block;
3876 struct address_space *mapping = inode->i_mapping;
3877 loff_t first_block_offset, last_block_offset, max_length;
3878 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3879 handle_t *handle;
3880 unsigned int credits;
3881 int ret = 0, ret2 = 0;
3882
3883 trace_ext4_punch_hole(inode, offset, length, 0);
3884
3885 /*
3886 * Write out all dirty pages to avoid race conditions
3887 * Then release them.
3888 */
3889 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3890 ret = filemap_write_and_wait_range(mapping, offset,
3891 offset + length - 1);
3892 if (ret)
3893 return ret;
3894 }
3895
3896 inode_lock(inode);
3897
3898 /* No need to punch hole beyond i_size */
3899 if (offset >= inode->i_size)
3900 goto out_mutex;
3901
3902 /*
3903 * If the hole extends beyond i_size, set the hole
3904 * to end after the page that contains i_size
3905 */
3906 if (offset + length > inode->i_size) {
3907 length = inode->i_size +
3908 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
3909 offset;
3910 }
3911
3912 /*
3913 * For punch hole the length + offset needs to be within one block
3914 * before last range. Adjust the length if it goes beyond that limit.
3915 */
3916 max_length = sbi->s_bitmap_maxbytes - inode->i_sb->s_blocksize;
3917 if (offset + length > max_length)
3918 length = max_length - offset;
3919
3920 if (offset & (sb->s_blocksize - 1) ||
3921 (offset + length) & (sb->s_blocksize - 1)) {
3922 /*
3923 * Attach jinode to inode for jbd2 if we do any zeroing of
3924 * partial block
3925 */
3926 ret = ext4_inode_attach_jinode(inode);
3927 if (ret < 0)
3928 goto out_mutex;
3929
3930 }
3931
3932 /* Wait all existing dio workers, newcomers will block on i_rwsem */
3933 inode_dio_wait(inode);
3934
3935 ret = file_modified(file);
3936 if (ret)
3937 goto out_mutex;
3938
3939 /*
3940 * Prevent page faults from reinstantiating pages we have released from
3941 * page cache.
3942 */
3943 filemap_invalidate_lock(mapping);
3944
3945 ret = ext4_break_layouts(inode);
3946 if (ret)
3947 goto out_dio;
3948
3949 first_block_offset = round_up(offset, sb->s_blocksize);
3950 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
3951
3952 /* Now release the pages and zero block aligned part of pages*/
3953 if (last_block_offset > first_block_offset) {
3954 ret = ext4_update_disksize_before_punch(inode, offset, length);
3955 if (ret)
3956 goto out_dio;
3957 truncate_pagecache_range(inode, first_block_offset,
3958 last_block_offset);
3959 }
3960
3961 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3962 credits = ext4_writepage_trans_blocks(inode);
3963 else
3964 credits = ext4_blocks_for_truncate(inode);
3965 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3966 if (IS_ERR(handle)) {
3967 ret = PTR_ERR(handle);
3968 ext4_std_error(sb, ret);
3969 goto out_dio;
3970 }
3971
3972 ret = ext4_zero_partial_blocks(handle, inode, offset,
3973 length);
3974 if (ret)
3975 goto out_stop;
3976
3977 first_block = (offset + sb->s_blocksize - 1) >>
3978 EXT4_BLOCK_SIZE_BITS(sb);
3979 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
3980
3981 /* If there are blocks to remove, do it */
3982 if (stop_block > first_block) {
3983 ext4_lblk_t hole_len = stop_block - first_block;
3984
3985 down_write(&EXT4_I(inode)->i_data_sem);
3986 ext4_discard_preallocations(inode);
3987
3988 ext4_es_remove_extent(inode, first_block, hole_len);
3989
3990 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3991 ret = ext4_ext_remove_space(inode, first_block,
3992 stop_block - 1);
3993 else
3994 ret = ext4_ind_remove_space(handle, inode, first_block,
3995 stop_block);
3996
3997 ext4_es_insert_extent(inode, first_block, hole_len, ~0,
3998 EXTENT_STATUS_HOLE);
3999 up_write(&EXT4_I(inode)->i_data_sem);
4000 }
4001 ext4_fc_track_range(handle, inode, first_block, stop_block);
4002 if (IS_SYNC(inode))
4003 ext4_handle_sync(handle);
4004
4005 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
4006 ret2 = ext4_mark_inode_dirty(handle, inode);
4007 if (unlikely(ret2))
4008 ret = ret2;
4009 if (ret >= 0)
4010 ext4_update_inode_fsync_trans(handle, inode, 1);
4011out_stop:
4012 ext4_journal_stop(handle);
4013out_dio:
4014 filemap_invalidate_unlock(mapping);
4015out_mutex:
4016 inode_unlock(inode);
4017 return ret;
4018}
4019
4020int ext4_inode_attach_jinode(struct inode *inode)
4021{
4022 struct ext4_inode_info *ei = EXT4_I(inode);
4023 struct jbd2_inode *jinode;
4024
4025 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4026 return 0;
4027
4028 jinode = jbd2_alloc_inode(GFP_KERNEL);
4029 spin_lock(&inode->i_lock);
4030 if (!ei->jinode) {
4031 if (!jinode) {
4032 spin_unlock(&inode->i_lock);
4033 return -ENOMEM;
4034 }
4035 ei->jinode = jinode;
4036 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4037 jinode = NULL;
4038 }
4039 spin_unlock(&inode->i_lock);
4040 if (unlikely(jinode != NULL))
4041 jbd2_free_inode(jinode);
4042 return 0;
4043}
4044
4045/*
4046 * ext4_truncate()
4047 *
4048 * We block out ext4_get_block() block instantiations across the entire
4049 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4050 * simultaneously on behalf of the same inode.
4051 *
4052 * As we work through the truncate and commit bits of it to the journal there
4053 * is one core, guiding principle: the file's tree must always be consistent on
4054 * disk. We must be able to restart the truncate after a crash.
4055 *
4056 * The file's tree may be transiently inconsistent in memory (although it
4057 * probably isn't), but whenever we close off and commit a journal transaction,
4058 * the contents of (the filesystem + the journal) must be consistent and
4059 * restartable. It's pretty simple, really: bottom up, right to left (although
4060 * left-to-right works OK too).
4061 *
4062 * Note that at recovery time, journal replay occurs *before* the restart of
4063 * truncate against the orphan inode list.
4064 *
4065 * The committed inode has the new, desired i_size (which is the same as
4066 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4067 * that this inode's truncate did not complete and it will again call
4068 * ext4_truncate() to have another go. So there will be instantiated blocks
4069 * to the right of the truncation point in a crashed ext4 filesystem. But
4070 * that's fine - as long as they are linked from the inode, the post-crash
4071 * ext4_truncate() run will find them and release them.
4072 */
4073int ext4_truncate(struct inode *inode)
4074{
4075 struct ext4_inode_info *ei = EXT4_I(inode);
4076 unsigned int credits;
4077 int err = 0, err2;
4078 handle_t *handle;
4079 struct address_space *mapping = inode->i_mapping;
4080
4081 /*
4082 * There is a possibility that we're either freeing the inode
4083 * or it's a completely new inode. In those cases we might not
4084 * have i_rwsem locked because it's not necessary.
4085 */
4086 if (!(inode->i_state & (I_NEW|I_FREEING)))
4087 WARN_ON(!inode_is_locked(inode));
4088 trace_ext4_truncate_enter(inode);
4089
4090 if (!ext4_can_truncate(inode))
4091 goto out_trace;
4092
4093 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4094 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4095
4096 if (ext4_has_inline_data(inode)) {
4097 int has_inline = 1;
4098
4099 err = ext4_inline_data_truncate(inode, &has_inline);
4100 if (err || has_inline)
4101 goto out_trace;
4102 }
4103
4104 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4105 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4106 err = ext4_inode_attach_jinode(inode);
4107 if (err)
4108 goto out_trace;
4109 }
4110
4111 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4112 credits = ext4_writepage_trans_blocks(inode);
4113 else
4114 credits = ext4_blocks_for_truncate(inode);
4115
4116 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4117 if (IS_ERR(handle)) {
4118 err = PTR_ERR(handle);
4119 goto out_trace;
4120 }
4121
4122 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4123 ext4_block_truncate_page(handle, mapping, inode->i_size);
4124
4125 /*
4126 * We add the inode to the orphan list, so that if this
4127 * truncate spans multiple transactions, and we crash, we will
4128 * resume the truncate when the filesystem recovers. It also
4129 * marks the inode dirty, to catch the new size.
4130 *
4131 * Implication: the file must always be in a sane, consistent
4132 * truncatable state while each transaction commits.
4133 */
4134 err = ext4_orphan_add(handle, inode);
4135 if (err)
4136 goto out_stop;
4137
4138 down_write(&EXT4_I(inode)->i_data_sem);
4139
4140 ext4_discard_preallocations(inode);
4141
4142 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4143 err = ext4_ext_truncate(handle, inode);
4144 else
4145 ext4_ind_truncate(handle, inode);
4146
4147 up_write(&ei->i_data_sem);
4148 if (err)
4149 goto out_stop;
4150
4151 if (IS_SYNC(inode))
4152 ext4_handle_sync(handle);
4153
4154out_stop:
4155 /*
4156 * If this was a simple ftruncate() and the file will remain alive,
4157 * then we need to clear up the orphan record which we created above.
4158 * However, if this was a real unlink then we were called by
4159 * ext4_evict_inode(), and we allow that function to clean up the
4160 * orphan info for us.
4161 */
4162 if (inode->i_nlink)
4163 ext4_orphan_del(handle, inode);
4164
4165 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
4166 err2 = ext4_mark_inode_dirty(handle, inode);
4167 if (unlikely(err2 && !err))
4168 err = err2;
4169 ext4_journal_stop(handle);
4170
4171out_trace:
4172 trace_ext4_truncate_exit(inode);
4173 return err;
4174}
4175
4176static inline u64 ext4_inode_peek_iversion(const struct inode *inode)
4177{
4178 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4179 return inode_peek_iversion_raw(inode);
4180 else
4181 return inode_peek_iversion(inode);
4182}
4183
4184static int ext4_inode_blocks_set(struct ext4_inode *raw_inode,
4185 struct ext4_inode_info *ei)
4186{
4187 struct inode *inode = &(ei->vfs_inode);
4188 u64 i_blocks = READ_ONCE(inode->i_blocks);
4189 struct super_block *sb = inode->i_sb;
4190
4191 if (i_blocks <= ~0U) {
4192 /*
4193 * i_blocks can be represented in a 32 bit variable
4194 * as multiple of 512 bytes
4195 */
4196 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4197 raw_inode->i_blocks_high = 0;
4198 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4199 return 0;
4200 }
4201
4202 /*
4203 * This should never happen since sb->s_maxbytes should not have
4204 * allowed this, sb->s_maxbytes was set according to the huge_file
4205 * feature in ext4_fill_super().
4206 */
4207 if (!ext4_has_feature_huge_file(sb))
4208 return -EFSCORRUPTED;
4209
4210 if (i_blocks <= 0xffffffffffffULL) {
4211 /*
4212 * i_blocks can be represented in a 48 bit variable
4213 * as multiple of 512 bytes
4214 */
4215 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4216 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4217 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4218 } else {
4219 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4220 /* i_block is stored in file system block size */
4221 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4222 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4223 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4224 }
4225 return 0;
4226}
4227
4228static int ext4_fill_raw_inode(struct inode *inode, struct ext4_inode *raw_inode)
4229{
4230 struct ext4_inode_info *ei = EXT4_I(inode);
4231 uid_t i_uid;
4232 gid_t i_gid;
4233 projid_t i_projid;
4234 int block;
4235 int err;
4236
4237 err = ext4_inode_blocks_set(raw_inode, ei);
4238
4239 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4240 i_uid = i_uid_read(inode);
4241 i_gid = i_gid_read(inode);
4242 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
4243 if (!(test_opt(inode->i_sb, NO_UID32))) {
4244 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4245 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4246 /*
4247 * Fix up interoperability with old kernels. Otherwise,
4248 * old inodes get re-used with the upper 16 bits of the
4249 * uid/gid intact.
4250 */
4251 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
4252 raw_inode->i_uid_high = 0;
4253 raw_inode->i_gid_high = 0;
4254 } else {
4255 raw_inode->i_uid_high =
4256 cpu_to_le16(high_16_bits(i_uid));
4257 raw_inode->i_gid_high =
4258 cpu_to_le16(high_16_bits(i_gid));
4259 }
4260 } else {
4261 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4262 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4263 raw_inode->i_uid_high = 0;
4264 raw_inode->i_gid_high = 0;
4265 }
4266 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4267
4268 EXT4_INODE_SET_CTIME(inode, raw_inode);
4269 EXT4_INODE_SET_MTIME(inode, raw_inode);
4270 EXT4_INODE_SET_ATIME(inode, raw_inode);
4271 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4272
4273 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4274 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4275 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4276 raw_inode->i_file_acl_high =
4277 cpu_to_le16(ei->i_file_acl >> 32);
4278 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4279 ext4_isize_set(raw_inode, ei->i_disksize);
4280
4281 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4282 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4283 if (old_valid_dev(inode->i_rdev)) {
4284 raw_inode->i_block[0] =
4285 cpu_to_le32(old_encode_dev(inode->i_rdev));
4286 raw_inode->i_block[1] = 0;
4287 } else {
4288 raw_inode->i_block[0] = 0;
4289 raw_inode->i_block[1] =
4290 cpu_to_le32(new_encode_dev(inode->i_rdev));
4291 raw_inode->i_block[2] = 0;
4292 }
4293 } else if (!ext4_has_inline_data(inode)) {
4294 for (block = 0; block < EXT4_N_BLOCKS; block++)
4295 raw_inode->i_block[block] = ei->i_data[block];
4296 }
4297
4298 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4299 u64 ivers = ext4_inode_peek_iversion(inode);
4300
4301 raw_inode->i_disk_version = cpu_to_le32(ivers);
4302 if (ei->i_extra_isize) {
4303 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4304 raw_inode->i_version_hi =
4305 cpu_to_le32(ivers >> 32);
4306 raw_inode->i_extra_isize =
4307 cpu_to_le16(ei->i_extra_isize);
4308 }
4309 }
4310
4311 if (i_projid != EXT4_DEF_PROJID &&
4312 !ext4_has_feature_project(inode->i_sb))
4313 err = err ?: -EFSCORRUPTED;
4314
4315 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4316 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4317 raw_inode->i_projid = cpu_to_le32(i_projid);
4318
4319 ext4_inode_csum_set(inode, raw_inode, ei);
4320 return err;
4321}
4322
4323/*
4324 * ext4_get_inode_loc returns with an extra refcount against the inode's
4325 * underlying buffer_head on success. If we pass 'inode' and it does not
4326 * have in-inode xattr, we have all inode data in memory that is needed
4327 * to recreate the on-disk version of this inode.
4328 */
4329static int __ext4_get_inode_loc(struct super_block *sb, unsigned long ino,
4330 struct inode *inode, struct ext4_iloc *iloc,
4331 ext4_fsblk_t *ret_block)
4332{
4333 struct ext4_group_desc *gdp;
4334 struct buffer_head *bh;
4335 ext4_fsblk_t block;
4336 struct blk_plug plug;
4337 int inodes_per_block, inode_offset;
4338
4339 iloc->bh = NULL;
4340 if (ino < EXT4_ROOT_INO ||
4341 ino > le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count))
4342 return -EFSCORRUPTED;
4343
4344 iloc->block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
4345 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4346 if (!gdp)
4347 return -EIO;
4348
4349 /*
4350 * Figure out the offset within the block group inode table
4351 */
4352 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4353 inode_offset = ((ino - 1) %
4354 EXT4_INODES_PER_GROUP(sb));
4355 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4356
4357 block = ext4_inode_table(sb, gdp);
4358 if ((block <= le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) ||
4359 (block >= ext4_blocks_count(EXT4_SB(sb)->s_es))) {
4360 ext4_error(sb, "Invalid inode table block %llu in "
4361 "block_group %u", block, iloc->block_group);
4362 return -EFSCORRUPTED;
4363 }
4364 block += (inode_offset / inodes_per_block);
4365
4366 bh = sb_getblk(sb, block);
4367 if (unlikely(!bh))
4368 return -ENOMEM;
4369 if (ext4_buffer_uptodate(bh))
4370 goto has_buffer;
4371
4372 lock_buffer(bh);
4373 if (ext4_buffer_uptodate(bh)) {
4374 /* Someone brought it uptodate while we waited */
4375 unlock_buffer(bh);
4376 goto has_buffer;
4377 }
4378
4379 /*
4380 * If we have all information of the inode in memory and this
4381 * is the only valid inode in the block, we need not read the
4382 * block.
4383 */
4384 if (inode && !ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4385 struct buffer_head *bitmap_bh;
4386 int i, start;
4387
4388 start = inode_offset & ~(inodes_per_block - 1);
4389
4390 /* Is the inode bitmap in cache? */
4391 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4392 if (unlikely(!bitmap_bh))
4393 goto make_io;
4394
4395 /*
4396 * If the inode bitmap isn't in cache then the
4397 * optimisation may end up performing two reads instead
4398 * of one, so skip it.
4399 */
4400 if (!buffer_uptodate(bitmap_bh)) {
4401 brelse(bitmap_bh);
4402 goto make_io;
4403 }
4404 for (i = start; i < start + inodes_per_block; i++) {
4405 if (i == inode_offset)
4406 continue;
4407 if (ext4_test_bit(i, bitmap_bh->b_data))
4408 break;
4409 }
4410 brelse(bitmap_bh);
4411 if (i == start + inodes_per_block) {
4412 struct ext4_inode *raw_inode =
4413 (struct ext4_inode *) (bh->b_data + iloc->offset);
4414
4415 /* all other inodes are free, so skip I/O */
4416 memset(bh->b_data, 0, bh->b_size);
4417 if (!ext4_test_inode_state(inode, EXT4_STATE_NEW))
4418 ext4_fill_raw_inode(inode, raw_inode);
4419 set_buffer_uptodate(bh);
4420 unlock_buffer(bh);
4421 goto has_buffer;
4422 }
4423 }
4424
4425make_io:
4426 /*
4427 * If we need to do any I/O, try to pre-readahead extra
4428 * blocks from the inode table.
4429 */
4430 blk_start_plug(&plug);
4431 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4432 ext4_fsblk_t b, end, table;
4433 unsigned num;
4434 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4435
4436 table = ext4_inode_table(sb, gdp);
4437 /* s_inode_readahead_blks is always a power of 2 */
4438 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4439 if (table > b)
4440 b = table;
4441 end = b + ra_blks;
4442 num = EXT4_INODES_PER_GROUP(sb);
4443 if (ext4_has_group_desc_csum(sb))
4444 num -= ext4_itable_unused_count(sb, gdp);
4445 table += num / inodes_per_block;
4446 if (end > table)
4447 end = table;
4448 while (b <= end)
4449 ext4_sb_breadahead_unmovable(sb, b++);
4450 }
4451
4452 /*
4453 * There are other valid inodes in the buffer, this inode
4454 * has in-inode xattrs, or we don't have this inode in memory.
4455 * Read the block from disk.
4456 */
4457 trace_ext4_load_inode(sb, ino);
4458 ext4_read_bh_nowait(bh, REQ_META | REQ_PRIO, NULL);
4459 blk_finish_plug(&plug);
4460 wait_on_buffer(bh);
4461 ext4_simulate_fail_bh(sb, bh, EXT4_SIM_INODE_EIO);
4462 if (!buffer_uptodate(bh)) {
4463 if (ret_block)
4464 *ret_block = block;
4465 brelse(bh);
4466 return -EIO;
4467 }
4468has_buffer:
4469 iloc->bh = bh;
4470 return 0;
4471}
4472
4473static int __ext4_get_inode_loc_noinmem(struct inode *inode,
4474 struct ext4_iloc *iloc)
4475{
4476 ext4_fsblk_t err_blk = 0;
4477 int ret;
4478
4479 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, NULL, iloc,
4480 &err_blk);
4481
4482 if (ret == -EIO)
4483 ext4_error_inode_block(inode, err_blk, EIO,
4484 "unable to read itable block");
4485
4486 return ret;
4487}
4488
4489int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4490{
4491 ext4_fsblk_t err_blk = 0;
4492 int ret;
4493
4494 ret = __ext4_get_inode_loc(inode->i_sb, inode->i_ino, inode, iloc,
4495 &err_blk);
4496
4497 if (ret == -EIO)
4498 ext4_error_inode_block(inode, err_blk, EIO,
4499 "unable to read itable block");
4500
4501 return ret;
4502}
4503
4504
4505int ext4_get_fc_inode_loc(struct super_block *sb, unsigned long ino,
4506 struct ext4_iloc *iloc)
4507{
4508 return __ext4_get_inode_loc(sb, ino, NULL, iloc, NULL);
4509}
4510
4511static bool ext4_should_enable_dax(struct inode *inode)
4512{
4513 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4514
4515 if (test_opt2(inode->i_sb, DAX_NEVER))
4516 return false;
4517 if (!S_ISREG(inode->i_mode))
4518 return false;
4519 if (ext4_should_journal_data(inode))
4520 return false;
4521 if (ext4_has_inline_data(inode))
4522 return false;
4523 if (ext4_test_inode_flag(inode, EXT4_INODE_ENCRYPT))
4524 return false;
4525 if (ext4_test_inode_flag(inode, EXT4_INODE_VERITY))
4526 return false;
4527 if (!test_bit(EXT4_FLAGS_BDEV_IS_DAX, &sbi->s_ext4_flags))
4528 return false;
4529 if (test_opt(inode->i_sb, DAX_ALWAYS))
4530 return true;
4531
4532 return ext4_test_inode_flag(inode, EXT4_INODE_DAX);
4533}
4534
4535void ext4_set_inode_flags(struct inode *inode, bool init)
4536{
4537 unsigned int flags = EXT4_I(inode)->i_flags;
4538 unsigned int new_fl = 0;
4539
4540 WARN_ON_ONCE(IS_DAX(inode) && init);
4541
4542 if (flags & EXT4_SYNC_FL)
4543 new_fl |= S_SYNC;
4544 if (flags & EXT4_APPEND_FL)
4545 new_fl |= S_APPEND;
4546 if (flags & EXT4_IMMUTABLE_FL)
4547 new_fl |= S_IMMUTABLE;
4548 if (flags & EXT4_NOATIME_FL)
4549 new_fl |= S_NOATIME;
4550 if (flags & EXT4_DIRSYNC_FL)
4551 new_fl |= S_DIRSYNC;
4552
4553 /* Because of the way inode_set_flags() works we must preserve S_DAX
4554 * here if already set. */
4555 new_fl |= (inode->i_flags & S_DAX);
4556 if (init && ext4_should_enable_dax(inode))
4557 new_fl |= S_DAX;
4558
4559 if (flags & EXT4_ENCRYPT_FL)
4560 new_fl |= S_ENCRYPTED;
4561 if (flags & EXT4_CASEFOLD_FL)
4562 new_fl |= S_CASEFOLD;
4563 if (flags & EXT4_VERITY_FL)
4564 new_fl |= S_VERITY;
4565 inode_set_flags(inode, new_fl,
4566 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4567 S_ENCRYPTED|S_CASEFOLD|S_VERITY);
4568}
4569
4570static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4571 struct ext4_inode_info *ei)
4572{
4573 blkcnt_t i_blocks ;
4574 struct inode *inode = &(ei->vfs_inode);
4575 struct super_block *sb = inode->i_sb;
4576
4577 if (ext4_has_feature_huge_file(sb)) {
4578 /* we are using combined 48 bit field */
4579 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4580 le32_to_cpu(raw_inode->i_blocks_lo);
4581 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4582 /* i_blocks represent file system block size */
4583 return i_blocks << (inode->i_blkbits - 9);
4584 } else {
4585 return i_blocks;
4586 }
4587 } else {
4588 return le32_to_cpu(raw_inode->i_blocks_lo);
4589 }
4590}
4591
4592static inline int ext4_iget_extra_inode(struct inode *inode,
4593 struct ext4_inode *raw_inode,
4594 struct ext4_inode_info *ei)
4595{
4596 __le32 *magic = (void *)raw_inode +
4597 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4598
4599 if (EXT4_INODE_HAS_XATTR_SPACE(inode) &&
4600 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4601 int err;
4602
4603 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4604 err = ext4_find_inline_data_nolock(inode);
4605 if (!err && ext4_has_inline_data(inode))
4606 ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
4607 return err;
4608 } else
4609 EXT4_I(inode)->i_inline_off = 0;
4610 return 0;
4611}
4612
4613int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4614{
4615 if (!ext4_has_feature_project(inode->i_sb))
4616 return -EOPNOTSUPP;
4617 *projid = EXT4_I(inode)->i_projid;
4618 return 0;
4619}
4620
4621/*
4622 * ext4 has self-managed i_version for ea inodes, it stores the lower 32bit of
4623 * refcount in i_version, so use raw values if inode has EXT4_EA_INODE_FL flag
4624 * set.
4625 */
4626static inline void ext4_inode_set_iversion_queried(struct inode *inode, u64 val)
4627{
4628 if (unlikely(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4629 inode_set_iversion_raw(inode, val);
4630 else
4631 inode_set_iversion_queried(inode, val);
4632}
4633
4634static const char *check_igot_inode(struct inode *inode, ext4_iget_flags flags)
4635
4636{
4637 if (flags & EXT4_IGET_EA_INODE) {
4638 if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4639 return "missing EA_INODE flag";
4640 if (ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4641 EXT4_I(inode)->i_file_acl)
4642 return "ea_inode with extended attributes";
4643 } else {
4644 if ((EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
4645 return "unexpected EA_INODE flag";
4646 }
4647 if (is_bad_inode(inode) && !(flags & EXT4_IGET_BAD))
4648 return "unexpected bad inode w/o EXT4_IGET_BAD";
4649 return NULL;
4650}
4651
4652struct inode *__ext4_iget(struct super_block *sb, unsigned long ino,
4653 ext4_iget_flags flags, const char *function,
4654 unsigned int line)
4655{
4656 struct ext4_iloc iloc;
4657 struct ext4_inode *raw_inode;
4658 struct ext4_inode_info *ei;
4659 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
4660 struct inode *inode;
4661 const char *err_str;
4662 journal_t *journal = EXT4_SB(sb)->s_journal;
4663 long ret;
4664 loff_t size;
4665 int block;
4666 uid_t i_uid;
4667 gid_t i_gid;
4668 projid_t i_projid;
4669
4670 if ((!(flags & EXT4_IGET_SPECIAL) &&
4671 ((ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO) ||
4672 ino == le32_to_cpu(es->s_usr_quota_inum) ||
4673 ino == le32_to_cpu(es->s_grp_quota_inum) ||
4674 ino == le32_to_cpu(es->s_prj_quota_inum) ||
4675 ino == le32_to_cpu(es->s_orphan_file_inum))) ||
4676 (ino < EXT4_ROOT_INO) ||
4677 (ino > le32_to_cpu(es->s_inodes_count))) {
4678 if (flags & EXT4_IGET_HANDLE)
4679 return ERR_PTR(-ESTALE);
4680 __ext4_error(sb, function, line, false, EFSCORRUPTED, 0,
4681 "inode #%lu: comm %s: iget: illegal inode #",
4682 ino, current->comm);
4683 return ERR_PTR(-EFSCORRUPTED);
4684 }
4685
4686 inode = iget_locked(sb, ino);
4687 if (!inode)
4688 return ERR_PTR(-ENOMEM);
4689 if (!(inode->i_state & I_NEW)) {
4690 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
4691 ext4_error_inode(inode, function, line, 0, err_str);
4692 iput(inode);
4693 return ERR_PTR(-EFSCORRUPTED);
4694 }
4695 return inode;
4696 }
4697
4698 ei = EXT4_I(inode);
4699 iloc.bh = NULL;
4700
4701 ret = __ext4_get_inode_loc_noinmem(inode, &iloc);
4702 if (ret < 0)
4703 goto bad_inode;
4704 raw_inode = ext4_raw_inode(&iloc);
4705
4706 if ((flags & EXT4_IGET_HANDLE) &&
4707 (raw_inode->i_links_count == 0) && (raw_inode->i_mode == 0)) {
4708 ret = -ESTALE;
4709 goto bad_inode;
4710 }
4711
4712 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4713 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4714 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4715 EXT4_INODE_SIZE(inode->i_sb) ||
4716 (ei->i_extra_isize & 3)) {
4717 ext4_error_inode(inode, function, line, 0,
4718 "iget: bad extra_isize %u "
4719 "(inode size %u)",
4720 ei->i_extra_isize,
4721 EXT4_INODE_SIZE(inode->i_sb));
4722 ret = -EFSCORRUPTED;
4723 goto bad_inode;
4724 }
4725 } else
4726 ei->i_extra_isize = 0;
4727
4728 /* Precompute checksum seed for inode metadata */
4729 if (ext4_has_metadata_csum(sb)) {
4730 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4731 __u32 csum;
4732 __le32 inum = cpu_to_le32(inode->i_ino);
4733 __le32 gen = raw_inode->i_generation;
4734 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4735 sizeof(inum));
4736 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4737 sizeof(gen));
4738 }
4739
4740 if ((!ext4_inode_csum_verify(inode, raw_inode, ei) ||
4741 ext4_simulate_fail(sb, EXT4_SIM_INODE_CRC)) &&
4742 (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))) {
4743 ext4_error_inode_err(inode, function, line, 0,
4744 EFSBADCRC, "iget: checksum invalid");
4745 ret = -EFSBADCRC;
4746 goto bad_inode;
4747 }
4748
4749 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4750 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4751 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4752 if (ext4_has_feature_project(sb) &&
4753 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4754 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4755 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4756 else
4757 i_projid = EXT4_DEF_PROJID;
4758
4759 if (!(test_opt(inode->i_sb, NO_UID32))) {
4760 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4761 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4762 }
4763 i_uid_write(inode, i_uid);
4764 i_gid_write(inode, i_gid);
4765 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4766 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4767
4768 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4769 ei->i_inline_off = 0;
4770 ei->i_dir_start_lookup = 0;
4771 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4772 /* We now have enough fields to check if the inode was active or not.
4773 * This is needed because nfsd might try to access dead inodes
4774 * the test is that same one that e2fsck uses
4775 * NeilBrown 1999oct15
4776 */
4777 if (inode->i_nlink == 0) {
4778 if ((inode->i_mode == 0 || flags & EXT4_IGET_SPECIAL ||
4779 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4780 ino != EXT4_BOOT_LOADER_INO) {
4781 /* this inode is deleted or unallocated */
4782 if (flags & EXT4_IGET_SPECIAL) {
4783 ext4_error_inode(inode, function, line, 0,
4784 "iget: special inode unallocated");
4785 ret = -EFSCORRUPTED;
4786 } else
4787 ret = -ESTALE;
4788 goto bad_inode;
4789 }
4790 /* The only unlinked inodes we let through here have
4791 * valid i_mode and are being read by the orphan
4792 * recovery code: that's fine, we're about to complete
4793 * the process of deleting those.
4794 * OR it is the EXT4_BOOT_LOADER_INO which is
4795 * not initialized on a new filesystem. */
4796 }
4797 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4798 ext4_set_inode_flags(inode, true);
4799 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4800 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4801 if (ext4_has_feature_64bit(sb))
4802 ei->i_file_acl |=
4803 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4804 inode->i_size = ext4_isize(sb, raw_inode);
4805 if ((size = i_size_read(inode)) < 0) {
4806 ext4_error_inode(inode, function, line, 0,
4807 "iget: bad i_size value: %lld", size);
4808 ret = -EFSCORRUPTED;
4809 goto bad_inode;
4810 }
4811 /*
4812 * If dir_index is not enabled but there's dir with INDEX flag set,
4813 * we'd normally treat htree data as empty space. But with metadata
4814 * checksumming that corrupts checksums so forbid that.
4815 */
4816 if (!ext4_has_feature_dir_index(sb) && ext4_has_metadata_csum(sb) &&
4817 ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) {
4818 ext4_error_inode(inode, function, line, 0,
4819 "iget: Dir with htree data on filesystem without dir_index feature.");
4820 ret = -EFSCORRUPTED;
4821 goto bad_inode;
4822 }
4823 ei->i_disksize = inode->i_size;
4824#ifdef CONFIG_QUOTA
4825 ei->i_reserved_quota = 0;
4826#endif
4827 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4828 ei->i_block_group = iloc.block_group;
4829 ei->i_last_alloc_group = ~0;
4830 /*
4831 * NOTE! The in-memory inode i_data array is in little-endian order
4832 * even on big-endian machines: we do NOT byteswap the block numbers!
4833 */
4834 for (block = 0; block < EXT4_N_BLOCKS; block++)
4835 ei->i_data[block] = raw_inode->i_block[block];
4836 INIT_LIST_HEAD(&ei->i_orphan);
4837 ext4_fc_init_inode(&ei->vfs_inode);
4838
4839 /*
4840 * Set transaction id's of transactions that have to be committed
4841 * to finish f[data]sync. We set them to currently running transaction
4842 * as we cannot be sure that the inode or some of its metadata isn't
4843 * part of the transaction - the inode could have been reclaimed and
4844 * now it is reread from disk.
4845 */
4846 if (journal) {
4847 transaction_t *transaction;
4848 tid_t tid;
4849
4850 read_lock(&journal->j_state_lock);
4851 if (journal->j_running_transaction)
4852 transaction = journal->j_running_transaction;
4853 else
4854 transaction = journal->j_committing_transaction;
4855 if (transaction)
4856 tid = transaction->t_tid;
4857 else
4858 tid = journal->j_commit_sequence;
4859 read_unlock(&journal->j_state_lock);
4860 ei->i_sync_tid = tid;
4861 ei->i_datasync_tid = tid;
4862 }
4863
4864 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4865 if (ei->i_extra_isize == 0) {
4866 /* The extra space is currently unused. Use it. */
4867 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4868 ei->i_extra_isize = sizeof(struct ext4_inode) -
4869 EXT4_GOOD_OLD_INODE_SIZE;
4870 } else {
4871 ret = ext4_iget_extra_inode(inode, raw_inode, ei);
4872 if (ret)
4873 goto bad_inode;
4874 }
4875 }
4876
4877 EXT4_INODE_GET_CTIME(inode, raw_inode);
4878 EXT4_INODE_GET_ATIME(inode, raw_inode);
4879 EXT4_INODE_GET_MTIME(inode, raw_inode);
4880 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4881
4882 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4883 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4884
4885 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4886 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4887 ivers |=
4888 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4889 }
4890 ext4_inode_set_iversion_queried(inode, ivers);
4891 }
4892
4893 ret = 0;
4894 if (ei->i_file_acl &&
4895 !ext4_inode_block_valid(inode, ei->i_file_acl, 1)) {
4896 ext4_error_inode(inode, function, line, 0,
4897 "iget: bad extended attribute block %llu",
4898 ei->i_file_acl);
4899 ret = -EFSCORRUPTED;
4900 goto bad_inode;
4901 } else if (!ext4_has_inline_data(inode)) {
4902 /* validate the block references in the inode */
4903 if (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) &&
4904 (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4905 (S_ISLNK(inode->i_mode) &&
4906 !ext4_inode_is_fast_symlink(inode)))) {
4907 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4908 ret = ext4_ext_check_inode(inode);
4909 else
4910 ret = ext4_ind_check_inode(inode);
4911 }
4912 }
4913 if (ret)
4914 goto bad_inode;
4915
4916 if (S_ISREG(inode->i_mode)) {
4917 inode->i_op = &ext4_file_inode_operations;
4918 inode->i_fop = &ext4_file_operations;
4919 ext4_set_aops(inode);
4920 } else if (S_ISDIR(inode->i_mode)) {
4921 inode->i_op = &ext4_dir_inode_operations;
4922 inode->i_fop = &ext4_dir_operations;
4923 } else if (S_ISLNK(inode->i_mode)) {
4924 /* VFS does not allow setting these so must be corruption */
4925 if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
4926 ext4_error_inode(inode, function, line, 0,
4927 "iget: immutable or append flags "
4928 "not allowed on symlinks");
4929 ret = -EFSCORRUPTED;
4930 goto bad_inode;
4931 }
4932 if (IS_ENCRYPTED(inode)) {
4933 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4934 } else if (ext4_inode_is_fast_symlink(inode)) {
4935 inode->i_link = (char *)ei->i_data;
4936 inode->i_op = &ext4_fast_symlink_inode_operations;
4937 nd_terminate_link(ei->i_data, inode->i_size,
4938 sizeof(ei->i_data) - 1);
4939 } else {
4940 inode->i_op = &ext4_symlink_inode_operations;
4941 }
4942 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4943 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4944 inode->i_op = &ext4_special_inode_operations;
4945 if (raw_inode->i_block[0])
4946 init_special_inode(inode, inode->i_mode,
4947 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4948 else
4949 init_special_inode(inode, inode->i_mode,
4950 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4951 } else if (ino == EXT4_BOOT_LOADER_INO) {
4952 make_bad_inode(inode);
4953 } else {
4954 ret = -EFSCORRUPTED;
4955 ext4_error_inode(inode, function, line, 0,
4956 "iget: bogus i_mode (%o)", inode->i_mode);
4957 goto bad_inode;
4958 }
4959 if (IS_CASEFOLDED(inode) && !ext4_has_feature_casefold(inode->i_sb)) {
4960 ext4_error_inode(inode, function, line, 0,
4961 "casefold flag without casefold feature");
4962 ret = -EFSCORRUPTED;
4963 goto bad_inode;
4964 }
4965 if ((err_str = check_igot_inode(inode, flags)) != NULL) {
4966 ext4_error_inode(inode, function, line, 0, err_str);
4967 ret = -EFSCORRUPTED;
4968 goto bad_inode;
4969 }
4970
4971 brelse(iloc.bh);
4972 unlock_new_inode(inode);
4973 return inode;
4974
4975bad_inode:
4976 brelse(iloc.bh);
4977 iget_failed(inode);
4978 return ERR_PTR(ret);
4979}
4980
4981static void __ext4_update_other_inode_time(struct super_block *sb,
4982 unsigned long orig_ino,
4983 unsigned long ino,
4984 struct ext4_inode *raw_inode)
4985{
4986 struct inode *inode;
4987
4988 inode = find_inode_by_ino_rcu(sb, ino);
4989 if (!inode)
4990 return;
4991
4992 if (!inode_is_dirtytime_only(inode))
4993 return;
4994
4995 spin_lock(&inode->i_lock);
4996 if (inode_is_dirtytime_only(inode)) {
4997 struct ext4_inode_info *ei = EXT4_I(inode);
4998
4999 inode->i_state &= ~I_DIRTY_TIME;
5000 spin_unlock(&inode->i_lock);
5001
5002 spin_lock(&ei->i_raw_lock);
5003 EXT4_INODE_SET_CTIME(inode, raw_inode);
5004 EXT4_INODE_SET_MTIME(inode, raw_inode);
5005 EXT4_INODE_SET_ATIME(inode, raw_inode);
5006 ext4_inode_csum_set(inode, raw_inode, ei);
5007 spin_unlock(&ei->i_raw_lock);
5008 trace_ext4_other_inode_update_time(inode, orig_ino);
5009 return;
5010 }
5011 spin_unlock(&inode->i_lock);
5012}
5013
5014/*
5015 * Opportunistically update the other time fields for other inodes in
5016 * the same inode table block.
5017 */
5018static void ext4_update_other_inodes_time(struct super_block *sb,
5019 unsigned long orig_ino, char *buf)
5020{
5021 unsigned long ino;
5022 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5023 int inode_size = EXT4_INODE_SIZE(sb);
5024
5025 /*
5026 * Calculate the first inode in the inode table block. Inode
5027 * numbers are one-based. That is, the first inode in a block
5028 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5029 */
5030 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5031 rcu_read_lock();
5032 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5033 if (ino == orig_ino)
5034 continue;
5035 __ext4_update_other_inode_time(sb, orig_ino, ino,
5036 (struct ext4_inode *)buf);
5037 }
5038 rcu_read_unlock();
5039}
5040
5041/*
5042 * Post the struct inode info into an on-disk inode location in the
5043 * buffer-cache. This gobbles the caller's reference to the
5044 * buffer_head in the inode location struct.
5045 *
5046 * The caller must have write access to iloc->bh.
5047 */
5048static int ext4_do_update_inode(handle_t *handle,
5049 struct inode *inode,
5050 struct ext4_iloc *iloc)
5051{
5052 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5053 struct ext4_inode_info *ei = EXT4_I(inode);
5054 struct buffer_head *bh = iloc->bh;
5055 struct super_block *sb = inode->i_sb;
5056 int err;
5057 int need_datasync = 0, set_large_file = 0;
5058
5059 spin_lock(&ei->i_raw_lock);
5060
5061 /*
5062 * For fields not tracked in the in-memory inode, initialise them
5063 * to zero for new inodes.
5064 */
5065 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5066 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5067
5068 if (READ_ONCE(ei->i_disksize) != ext4_isize(inode->i_sb, raw_inode))
5069 need_datasync = 1;
5070 if (ei->i_disksize > 0x7fffffffULL) {
5071 if (!ext4_has_feature_large_file(sb) ||
5072 EXT4_SB(sb)->s_es->s_rev_level == cpu_to_le32(EXT4_GOOD_OLD_REV))
5073 set_large_file = 1;
5074 }
5075
5076 err = ext4_fill_raw_inode(inode, raw_inode);
5077 spin_unlock(&ei->i_raw_lock);
5078 if (err) {
5079 EXT4_ERROR_INODE(inode, "corrupted inode contents");
5080 goto out_brelse;
5081 }
5082
5083 if (inode->i_sb->s_flags & SB_LAZYTIME)
5084 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5085 bh->b_data);
5086
5087 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5088 err = ext4_handle_dirty_metadata(handle, NULL, bh);
5089 if (err)
5090 goto out_error;
5091 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5092 if (set_large_file) {
5093 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5094 err = ext4_journal_get_write_access(handle, sb,
5095 EXT4_SB(sb)->s_sbh,
5096 EXT4_JTR_NONE);
5097 if (err)
5098 goto out_error;
5099 lock_buffer(EXT4_SB(sb)->s_sbh);
5100 ext4_set_feature_large_file(sb);
5101 ext4_superblock_csum_set(sb);
5102 unlock_buffer(EXT4_SB(sb)->s_sbh);
5103 ext4_handle_sync(handle);
5104 err = ext4_handle_dirty_metadata(handle, NULL,
5105 EXT4_SB(sb)->s_sbh);
5106 }
5107 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5108out_error:
5109 ext4_std_error(inode->i_sb, err);
5110out_brelse:
5111 brelse(bh);
5112 return err;
5113}
5114
5115/*
5116 * ext4_write_inode()
5117 *
5118 * We are called from a few places:
5119 *
5120 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5121 * Here, there will be no transaction running. We wait for any running
5122 * transaction to commit.
5123 *
5124 * - Within flush work (sys_sync(), kupdate and such).
5125 * We wait on commit, if told to.
5126 *
5127 * - Within iput_final() -> write_inode_now()
5128 * We wait on commit, if told to.
5129 *
5130 * In all cases it is actually safe for us to return without doing anything,
5131 * because the inode has been copied into a raw inode buffer in
5132 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5133 * writeback.
5134 *
5135 * Note that we are absolutely dependent upon all inode dirtiers doing the
5136 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5137 * which we are interested.
5138 *
5139 * It would be a bug for them to not do this. The code:
5140 *
5141 * mark_inode_dirty(inode)
5142 * stuff();
5143 * inode->i_size = expr;
5144 *
5145 * is in error because write_inode() could occur while `stuff()' is running,
5146 * and the new i_size will be lost. Plus the inode will no longer be on the
5147 * superblock's dirty inode list.
5148 */
5149int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5150{
5151 int err;
5152
5153 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5154 return 0;
5155
5156 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5157 return -EIO;
5158
5159 if (EXT4_SB(inode->i_sb)->s_journal) {
5160 if (ext4_journal_current_handle()) {
5161 ext4_debug("called recursively, non-PF_MEMALLOC!\n");
5162 dump_stack();
5163 return -EIO;
5164 }
5165
5166 /*
5167 * No need to force transaction in WB_SYNC_NONE mode. Also
5168 * ext4_sync_fs() will force the commit after everything is
5169 * written.
5170 */
5171 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5172 return 0;
5173
5174 err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
5175 EXT4_I(inode)->i_sync_tid);
5176 } else {
5177 struct ext4_iloc iloc;
5178
5179 err = __ext4_get_inode_loc_noinmem(inode, &iloc);
5180 if (err)
5181 return err;
5182 /*
5183 * sync(2) will flush the whole buffer cache. No need to do
5184 * it here separately for each inode.
5185 */
5186 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5187 sync_dirty_buffer(iloc.bh);
5188 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5189 ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
5190 "IO error syncing inode");
5191 err = -EIO;
5192 }
5193 brelse(iloc.bh);
5194 }
5195 return err;
5196}
5197
5198/*
5199 * In data=journal mode ext4_journalled_invalidate_folio() may fail to invalidate
5200 * buffers that are attached to a folio straddling i_size and are undergoing
5201 * commit. In that case we have to wait for commit to finish and try again.
5202 */
5203static void ext4_wait_for_tail_page_commit(struct inode *inode)
5204{
5205 unsigned offset;
5206 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5207 tid_t commit_tid = 0;
5208 int ret;
5209
5210 offset = inode->i_size & (PAGE_SIZE - 1);
5211 /*
5212 * If the folio is fully truncated, we don't need to wait for any commit
5213 * (and we even should not as __ext4_journalled_invalidate_folio() may
5214 * strip all buffers from the folio but keep the folio dirty which can then
5215 * confuse e.g. concurrent ext4_writepages() seeing dirty folio without
5216 * buffers). Also we don't need to wait for any commit if all buffers in
5217 * the folio remain valid. This is most beneficial for the common case of
5218 * blocksize == PAGESIZE.
5219 */
5220 if (!offset || offset > (PAGE_SIZE - i_blocksize(inode)))
5221 return;
5222 while (1) {
5223 struct folio *folio = filemap_lock_folio(inode->i_mapping,
5224 inode->i_size >> PAGE_SHIFT);
5225 if (IS_ERR(folio))
5226 return;
5227 ret = __ext4_journalled_invalidate_folio(folio, offset,
5228 folio_size(folio) - offset);
5229 folio_unlock(folio);
5230 folio_put(folio);
5231 if (ret != -EBUSY)
5232 return;
5233 commit_tid = 0;
5234 read_lock(&journal->j_state_lock);
5235 if (journal->j_committing_transaction)
5236 commit_tid = journal->j_committing_transaction->t_tid;
5237 read_unlock(&journal->j_state_lock);
5238 if (commit_tid)
5239 jbd2_log_wait_commit(journal, commit_tid);
5240 }
5241}
5242
5243/*
5244 * ext4_setattr()
5245 *
5246 * Called from notify_change.
5247 *
5248 * We want to trap VFS attempts to truncate the file as soon as
5249 * possible. In particular, we want to make sure that when the VFS
5250 * shrinks i_size, we put the inode on the orphan list and modify
5251 * i_disksize immediately, so that during the subsequent flushing of
5252 * dirty pages and freeing of disk blocks, we can guarantee that any
5253 * commit will leave the blocks being flushed in an unused state on
5254 * disk. (On recovery, the inode will get truncated and the blocks will
5255 * be freed, so we have a strong guarantee that no future commit will
5256 * leave these blocks visible to the user.)
5257 *
5258 * Another thing we have to assure is that if we are in ordered mode
5259 * and inode is still attached to the committing transaction, we must
5260 * we start writeout of all the dirty pages which are being truncated.
5261 * This way we are sure that all the data written in the previous
5262 * transaction are already on disk (truncate waits for pages under
5263 * writeback).
5264 *
5265 * Called with inode->i_rwsem down.
5266 */
5267int ext4_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
5268 struct iattr *attr)
5269{
5270 struct inode *inode = d_inode(dentry);
5271 int error, rc = 0;
5272 int orphan = 0;
5273 const unsigned int ia_valid = attr->ia_valid;
5274 bool inc_ivers = true;
5275
5276 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5277 return -EIO;
5278
5279 if (unlikely(IS_IMMUTABLE(inode)))
5280 return -EPERM;
5281
5282 if (unlikely(IS_APPEND(inode) &&
5283 (ia_valid & (ATTR_MODE | ATTR_UID |
5284 ATTR_GID | ATTR_TIMES_SET))))
5285 return -EPERM;
5286
5287 error = setattr_prepare(idmap, dentry, attr);
5288 if (error)
5289 return error;
5290
5291 error = fscrypt_prepare_setattr(dentry, attr);
5292 if (error)
5293 return error;
5294
5295 error = fsverity_prepare_setattr(dentry, attr);
5296 if (error)
5297 return error;
5298
5299 if (is_quota_modification(idmap, inode, attr)) {
5300 error = dquot_initialize(inode);
5301 if (error)
5302 return error;
5303 }
5304
5305 if (i_uid_needs_update(idmap, attr, inode) ||
5306 i_gid_needs_update(idmap, attr, inode)) {
5307 handle_t *handle;
5308
5309 /* (user+group)*(old+new) structure, inode write (sb,
5310 * inode block, ? - but truncate inode update has it) */
5311 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5312 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5313 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5314 if (IS_ERR(handle)) {
5315 error = PTR_ERR(handle);
5316 goto err_out;
5317 }
5318
5319 /* dquot_transfer() calls back ext4_get_inode_usage() which
5320 * counts xattr inode references.
5321 */
5322 down_read(&EXT4_I(inode)->xattr_sem);
5323 error = dquot_transfer(idmap, inode, attr);
5324 up_read(&EXT4_I(inode)->xattr_sem);
5325
5326 if (error) {
5327 ext4_journal_stop(handle);
5328 return error;
5329 }
5330 /* Update corresponding info in inode so that everything is in
5331 * one transaction */
5332 i_uid_update(idmap, attr, inode);
5333 i_gid_update(idmap, attr, inode);
5334 error = ext4_mark_inode_dirty(handle, inode);
5335 ext4_journal_stop(handle);
5336 if (unlikely(error)) {
5337 return error;
5338 }
5339 }
5340
5341 if (attr->ia_valid & ATTR_SIZE) {
5342 handle_t *handle;
5343 loff_t oldsize = inode->i_size;
5344 loff_t old_disksize;
5345 int shrink = (attr->ia_size < inode->i_size);
5346
5347 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5348 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5349
5350 if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5351 return -EFBIG;
5352 }
5353 }
5354 if (!S_ISREG(inode->i_mode)) {
5355 return -EINVAL;
5356 }
5357
5358 if (attr->ia_size == inode->i_size)
5359 inc_ivers = false;
5360
5361 if (shrink) {
5362 if (ext4_should_order_data(inode)) {
5363 error = ext4_begin_ordered_truncate(inode,
5364 attr->ia_size);
5365 if (error)
5366 goto err_out;
5367 }
5368 /*
5369 * Blocks are going to be removed from the inode. Wait
5370 * for dio in flight.
5371 */
5372 inode_dio_wait(inode);
5373 }
5374
5375 filemap_invalidate_lock(inode->i_mapping);
5376
5377 rc = ext4_break_layouts(inode);
5378 if (rc) {
5379 filemap_invalidate_unlock(inode->i_mapping);
5380 goto err_out;
5381 }
5382
5383 if (attr->ia_size != inode->i_size) {
5384 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5385 if (IS_ERR(handle)) {
5386 error = PTR_ERR(handle);
5387 goto out_mmap_sem;
5388 }
5389 if (ext4_handle_valid(handle) && shrink) {
5390 error = ext4_orphan_add(handle, inode);
5391 orphan = 1;
5392 }
5393 /*
5394 * Update c/mtime on truncate up, ext4_truncate() will
5395 * update c/mtime in shrink case below
5396 */
5397 if (!shrink)
5398 inode_set_mtime_to_ts(inode,
5399 inode_set_ctime_current(inode));
5400
5401 if (shrink)
5402 ext4_fc_track_range(handle, inode,
5403 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5404 inode->i_sb->s_blocksize_bits,
5405 EXT_MAX_BLOCKS - 1);
5406 else
5407 ext4_fc_track_range(
5408 handle, inode,
5409 (oldsize > 0 ? oldsize - 1 : oldsize) >>
5410 inode->i_sb->s_blocksize_bits,
5411 (attr->ia_size > 0 ? attr->ia_size - 1 : 0) >>
5412 inode->i_sb->s_blocksize_bits);
5413
5414 down_write(&EXT4_I(inode)->i_data_sem);
5415 old_disksize = EXT4_I(inode)->i_disksize;
5416 EXT4_I(inode)->i_disksize = attr->ia_size;
5417 rc = ext4_mark_inode_dirty(handle, inode);
5418 if (!error)
5419 error = rc;
5420 /*
5421 * We have to update i_size under i_data_sem together
5422 * with i_disksize to avoid races with writeback code
5423 * running ext4_wb_update_i_disksize().
5424 */
5425 if (!error)
5426 i_size_write(inode, attr->ia_size);
5427 else
5428 EXT4_I(inode)->i_disksize = old_disksize;
5429 up_write(&EXT4_I(inode)->i_data_sem);
5430 ext4_journal_stop(handle);
5431 if (error)
5432 goto out_mmap_sem;
5433 if (!shrink) {
5434 pagecache_isize_extended(inode, oldsize,
5435 inode->i_size);
5436 } else if (ext4_should_journal_data(inode)) {
5437 ext4_wait_for_tail_page_commit(inode);
5438 }
5439 }
5440
5441 /*
5442 * Truncate pagecache after we've waited for commit
5443 * in data=journal mode to make pages freeable.
5444 */
5445 truncate_pagecache(inode, inode->i_size);
5446 /*
5447 * Call ext4_truncate() even if i_size didn't change to
5448 * truncate possible preallocated blocks.
5449 */
5450 if (attr->ia_size <= oldsize) {
5451 rc = ext4_truncate(inode);
5452 if (rc)
5453 error = rc;
5454 }
5455out_mmap_sem:
5456 filemap_invalidate_unlock(inode->i_mapping);
5457 }
5458
5459 if (!error) {
5460 if (inc_ivers)
5461 inode_inc_iversion(inode);
5462 setattr_copy(idmap, inode, attr);
5463 mark_inode_dirty(inode);
5464 }
5465
5466 /*
5467 * If the call to ext4_truncate failed to get a transaction handle at
5468 * all, we need to clean up the in-core orphan list manually.
5469 */
5470 if (orphan && inode->i_nlink)
5471 ext4_orphan_del(NULL, inode);
5472
5473 if (!error && (ia_valid & ATTR_MODE))
5474 rc = posix_acl_chmod(idmap, dentry, inode->i_mode);
5475
5476err_out:
5477 if (error)
5478 ext4_std_error(inode->i_sb, error);
5479 if (!error)
5480 error = rc;
5481 return error;
5482}
5483
5484u32 ext4_dio_alignment(struct inode *inode)
5485{
5486 if (fsverity_active(inode))
5487 return 0;
5488 if (ext4_should_journal_data(inode))
5489 return 0;
5490 if (ext4_has_inline_data(inode))
5491 return 0;
5492 if (IS_ENCRYPTED(inode)) {
5493 if (!fscrypt_dio_supported(inode))
5494 return 0;
5495 return i_blocksize(inode);
5496 }
5497 return 1; /* use the iomap defaults */
5498}
5499
5500int ext4_getattr(struct mnt_idmap *idmap, const struct path *path,
5501 struct kstat *stat, u32 request_mask, unsigned int query_flags)
5502{
5503 struct inode *inode = d_inode(path->dentry);
5504 struct ext4_inode *raw_inode;
5505 struct ext4_inode_info *ei = EXT4_I(inode);
5506 unsigned int flags;
5507
5508 if ((request_mask & STATX_BTIME) &&
5509 EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5510 stat->result_mask |= STATX_BTIME;
5511 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5512 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5513 }
5514
5515 /*
5516 * Return the DIO alignment restrictions if requested. We only return
5517 * this information when requested, since on encrypted files it might
5518 * take a fair bit of work to get if the file wasn't opened recently.
5519 */
5520 if ((request_mask & STATX_DIOALIGN) && S_ISREG(inode->i_mode)) {
5521 u32 dio_align = ext4_dio_alignment(inode);
5522
5523 stat->result_mask |= STATX_DIOALIGN;
5524 if (dio_align == 1) {
5525 struct block_device *bdev = inode->i_sb->s_bdev;
5526
5527 /* iomap defaults */
5528 stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
5529 stat->dio_offset_align = bdev_logical_block_size(bdev);
5530 } else {
5531 stat->dio_mem_align = dio_align;
5532 stat->dio_offset_align = dio_align;
5533 }
5534 }
5535
5536 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5537 if (flags & EXT4_APPEND_FL)
5538 stat->attributes |= STATX_ATTR_APPEND;
5539 if (flags & EXT4_COMPR_FL)
5540 stat->attributes |= STATX_ATTR_COMPRESSED;
5541 if (flags & EXT4_ENCRYPT_FL)
5542 stat->attributes |= STATX_ATTR_ENCRYPTED;
5543 if (flags & EXT4_IMMUTABLE_FL)
5544 stat->attributes |= STATX_ATTR_IMMUTABLE;
5545 if (flags & EXT4_NODUMP_FL)
5546 stat->attributes |= STATX_ATTR_NODUMP;
5547 if (flags & EXT4_VERITY_FL)
5548 stat->attributes |= STATX_ATTR_VERITY;
5549
5550 stat->attributes_mask |= (STATX_ATTR_APPEND |
5551 STATX_ATTR_COMPRESSED |
5552 STATX_ATTR_ENCRYPTED |
5553 STATX_ATTR_IMMUTABLE |
5554 STATX_ATTR_NODUMP |
5555 STATX_ATTR_VERITY);
5556
5557 generic_fillattr(idmap, request_mask, inode, stat);
5558 return 0;
5559}
5560
5561int ext4_file_getattr(struct mnt_idmap *idmap,
5562 const struct path *path, struct kstat *stat,
5563 u32 request_mask, unsigned int query_flags)
5564{
5565 struct inode *inode = d_inode(path->dentry);
5566 u64 delalloc_blocks;
5567
5568 ext4_getattr(idmap, path, stat, request_mask, query_flags);
5569
5570 /*
5571 * If there is inline data in the inode, the inode will normally not
5572 * have data blocks allocated (it may have an external xattr block).
5573 * Report at least one sector for such files, so tools like tar, rsync,
5574 * others don't incorrectly think the file is completely sparse.
5575 */
5576 if (unlikely(ext4_has_inline_data(inode)))
5577 stat->blocks += (stat->size + 511) >> 9;
5578
5579 /*
5580 * We can't update i_blocks if the block allocation is delayed
5581 * otherwise in the case of system crash before the real block
5582 * allocation is done, we will have i_blocks inconsistent with
5583 * on-disk file blocks.
5584 * We always keep i_blocks updated together with real
5585 * allocation. But to not confuse with user, stat
5586 * will return the blocks that include the delayed allocation
5587 * blocks for this file.
5588 */
5589 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5590 EXT4_I(inode)->i_reserved_data_blocks);
5591 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5592 return 0;
5593}
5594
5595static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5596 int pextents)
5597{
5598 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5599 return ext4_ind_trans_blocks(inode, lblocks);
5600 return ext4_ext_index_trans_blocks(inode, pextents);
5601}
5602
5603/*
5604 * Account for index blocks, block groups bitmaps and block group
5605 * descriptor blocks if modify datablocks and index blocks
5606 * worse case, the indexs blocks spread over different block groups
5607 *
5608 * If datablocks are discontiguous, they are possible to spread over
5609 * different block groups too. If they are contiguous, with flexbg,
5610 * they could still across block group boundary.
5611 *
5612 * Also account for superblock, inode, quota and xattr blocks
5613 */
5614static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5615 int pextents)
5616{
5617 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5618 int gdpblocks;
5619 int idxblocks;
5620 int ret;
5621
5622 /*
5623 * How many index blocks need to touch to map @lblocks logical blocks
5624 * to @pextents physical extents?
5625 */
5626 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5627
5628 ret = idxblocks;
5629
5630 /*
5631 * Now let's see how many group bitmaps and group descriptors need
5632 * to account
5633 */
5634 groups = idxblocks + pextents;
5635 gdpblocks = groups;
5636 if (groups > ngroups)
5637 groups = ngroups;
5638 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5639 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5640
5641 /* bitmaps and block group descriptor blocks */
5642 ret += groups + gdpblocks;
5643
5644 /* Blocks for super block, inode, quota and xattr blocks */
5645 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5646
5647 return ret;
5648}
5649
5650/*
5651 * Calculate the total number of credits to reserve to fit
5652 * the modification of a single pages into a single transaction,
5653 * which may include multiple chunks of block allocations.
5654 *
5655 * This could be called via ext4_write_begin()
5656 *
5657 * We need to consider the worse case, when
5658 * one new block per extent.
5659 */
5660int ext4_writepage_trans_blocks(struct inode *inode)
5661{
5662 int bpp = ext4_journal_blocks_per_page(inode);
5663 int ret;
5664
5665 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5666
5667 /* Account for data blocks for journalled mode */
5668 if (ext4_should_journal_data(inode))
5669 ret += bpp;
5670 return ret;
5671}
5672
5673/*
5674 * Calculate the journal credits for a chunk of data modification.
5675 *
5676 * This is called from DIO, fallocate or whoever calling
5677 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5678 *
5679 * journal buffers for data blocks are not included here, as DIO
5680 * and fallocate do no need to journal data buffers.
5681 */
5682int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5683{
5684 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5685}
5686
5687/*
5688 * The caller must have previously called ext4_reserve_inode_write().
5689 * Give this, we know that the caller already has write access to iloc->bh.
5690 */
5691int ext4_mark_iloc_dirty(handle_t *handle,
5692 struct inode *inode, struct ext4_iloc *iloc)
5693{
5694 int err = 0;
5695
5696 if (unlikely(ext4_forced_shutdown(inode->i_sb))) {
5697 put_bh(iloc->bh);
5698 return -EIO;
5699 }
5700 ext4_fc_track_inode(handle, inode);
5701
5702 /* the do_update_inode consumes one bh->b_count */
5703 get_bh(iloc->bh);
5704
5705 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5706 err = ext4_do_update_inode(handle, inode, iloc);
5707 put_bh(iloc->bh);
5708 return err;
5709}
5710
5711/*
5712 * On success, We end up with an outstanding reference count against
5713 * iloc->bh. This _must_ be cleaned up later.
5714 */
5715
5716int
5717ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5718 struct ext4_iloc *iloc)
5719{
5720 int err;
5721
5722 if (unlikely(ext4_forced_shutdown(inode->i_sb)))
5723 return -EIO;
5724
5725 err = ext4_get_inode_loc(inode, iloc);
5726 if (!err) {
5727 BUFFER_TRACE(iloc->bh, "get_write_access");
5728 err = ext4_journal_get_write_access(handle, inode->i_sb,
5729 iloc->bh, EXT4_JTR_NONE);
5730 if (err) {
5731 brelse(iloc->bh);
5732 iloc->bh = NULL;
5733 }
5734 }
5735 ext4_std_error(inode->i_sb, err);
5736 return err;
5737}
5738
5739static int __ext4_expand_extra_isize(struct inode *inode,
5740 unsigned int new_extra_isize,
5741 struct ext4_iloc *iloc,
5742 handle_t *handle, int *no_expand)
5743{
5744 struct ext4_inode *raw_inode;
5745 struct ext4_xattr_ibody_header *header;
5746 unsigned int inode_size = EXT4_INODE_SIZE(inode->i_sb);
5747 struct ext4_inode_info *ei = EXT4_I(inode);
5748 int error;
5749
5750 /* this was checked at iget time, but double check for good measure */
5751 if ((EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize > inode_size) ||
5752 (ei->i_extra_isize & 3)) {
5753 EXT4_ERROR_INODE(inode, "bad extra_isize %u (inode size %u)",
5754 ei->i_extra_isize,
5755 EXT4_INODE_SIZE(inode->i_sb));
5756 return -EFSCORRUPTED;
5757 }
5758 if ((new_extra_isize < ei->i_extra_isize) ||
5759 (new_extra_isize < 4) ||
5760 (new_extra_isize > inode_size - EXT4_GOOD_OLD_INODE_SIZE))
5761 return -EINVAL; /* Should never happen */
5762
5763 raw_inode = ext4_raw_inode(iloc);
5764
5765 header = IHDR(inode, raw_inode);
5766
5767 /* No extended attributes present */
5768 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5769 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5770 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5771 EXT4_I(inode)->i_extra_isize, 0,
5772 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5773 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5774 return 0;
5775 }
5776
5777 /*
5778 * We may need to allocate external xattr block so we need quotas
5779 * initialized. Here we can be called with various locks held so we
5780 * cannot affort to initialize quotas ourselves. So just bail.
5781 */
5782 if (dquot_initialize_needed(inode))
5783 return -EAGAIN;
5784
5785 /* try to expand with EAs present */
5786 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5787 raw_inode, handle);
5788 if (error) {
5789 /*
5790 * Inode size expansion failed; don't try again
5791 */
5792 *no_expand = 1;
5793 }
5794
5795 return error;
5796}
5797
5798/*
5799 * Expand an inode by new_extra_isize bytes.
5800 * Returns 0 on success or negative error number on failure.
5801 */
5802static int ext4_try_to_expand_extra_isize(struct inode *inode,
5803 unsigned int new_extra_isize,
5804 struct ext4_iloc iloc,
5805 handle_t *handle)
5806{
5807 int no_expand;
5808 int error;
5809
5810 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5811 return -EOVERFLOW;
5812
5813 /*
5814 * In nojournal mode, we can immediately attempt to expand
5815 * the inode. When journaled, we first need to obtain extra
5816 * buffer credits since we may write into the EA block
5817 * with this same handle. If journal_extend fails, then it will
5818 * only result in a minor loss of functionality for that inode.
5819 * If this is felt to be critical, then e2fsck should be run to
5820 * force a large enough s_min_extra_isize.
5821 */
5822 if (ext4_journal_extend(handle,
5823 EXT4_DATA_TRANS_BLOCKS(inode->i_sb), 0) != 0)
5824 return -ENOSPC;
5825
5826 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5827 return -EBUSY;
5828
5829 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5830 handle, &no_expand);
5831 ext4_write_unlock_xattr(inode, &no_expand);
5832
5833 return error;
5834}
5835
5836int ext4_expand_extra_isize(struct inode *inode,
5837 unsigned int new_extra_isize,
5838 struct ext4_iloc *iloc)
5839{
5840 handle_t *handle;
5841 int no_expand;
5842 int error, rc;
5843
5844 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5845 brelse(iloc->bh);
5846 return -EOVERFLOW;
5847 }
5848
5849 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5850 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5851 if (IS_ERR(handle)) {
5852 error = PTR_ERR(handle);
5853 brelse(iloc->bh);
5854 return error;
5855 }
5856
5857 ext4_write_lock_xattr(inode, &no_expand);
5858
5859 BUFFER_TRACE(iloc->bh, "get_write_access");
5860 error = ext4_journal_get_write_access(handle, inode->i_sb, iloc->bh,
5861 EXT4_JTR_NONE);
5862 if (error) {
5863 brelse(iloc->bh);
5864 goto out_unlock;
5865 }
5866
5867 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5868 handle, &no_expand);
5869
5870 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5871 if (!error)
5872 error = rc;
5873
5874out_unlock:
5875 ext4_write_unlock_xattr(inode, &no_expand);
5876 ext4_journal_stop(handle);
5877 return error;
5878}
5879
5880/*
5881 * What we do here is to mark the in-core inode as clean with respect to inode
5882 * dirtiness (it may still be data-dirty).
5883 * This means that the in-core inode may be reaped by prune_icache
5884 * without having to perform any I/O. This is a very good thing,
5885 * because *any* task may call prune_icache - even ones which
5886 * have a transaction open against a different journal.
5887 *
5888 * Is this cheating? Not really. Sure, we haven't written the
5889 * inode out, but prune_icache isn't a user-visible syncing function.
5890 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5891 * we start and wait on commits.
5892 */
5893int __ext4_mark_inode_dirty(handle_t *handle, struct inode *inode,
5894 const char *func, unsigned int line)
5895{
5896 struct ext4_iloc iloc;
5897 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5898 int err;
5899
5900 might_sleep();
5901 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5902 err = ext4_reserve_inode_write(handle, inode, &iloc);
5903 if (err)
5904 goto out;
5905
5906 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
5907 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
5908 iloc, handle);
5909
5910 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5911out:
5912 if (unlikely(err))
5913 ext4_error_inode_err(inode, func, line, 0, err,
5914 "mark_inode_dirty error");
5915 return err;
5916}
5917
5918/*
5919 * ext4_dirty_inode() is called from __mark_inode_dirty()
5920 *
5921 * We're really interested in the case where a file is being extended.
5922 * i_size has been changed by generic_commit_write() and we thus need
5923 * to include the updated inode in the current transaction.
5924 *
5925 * Also, dquot_alloc_block() will always dirty the inode when blocks
5926 * are allocated to the file.
5927 *
5928 * If the inode is marked synchronous, we don't honour that here - doing
5929 * so would cause a commit on atime updates, which we don't bother doing.
5930 * We handle synchronous inodes at the highest possible level.
5931 */
5932void ext4_dirty_inode(struct inode *inode, int flags)
5933{
5934 handle_t *handle;
5935
5936 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5937 if (IS_ERR(handle))
5938 return;
5939 ext4_mark_inode_dirty(handle, inode);
5940 ext4_journal_stop(handle);
5941}
5942
5943int ext4_change_inode_journal_flag(struct inode *inode, int val)
5944{
5945 journal_t *journal;
5946 handle_t *handle;
5947 int err;
5948 int alloc_ctx;
5949
5950 /*
5951 * We have to be very careful here: changing a data block's
5952 * journaling status dynamically is dangerous. If we write a
5953 * data block to the journal, change the status and then delete
5954 * that block, we risk forgetting to revoke the old log record
5955 * from the journal and so a subsequent replay can corrupt data.
5956 * So, first we make sure that the journal is empty and that
5957 * nobody is changing anything.
5958 */
5959
5960 journal = EXT4_JOURNAL(inode);
5961 if (!journal)
5962 return 0;
5963 if (is_journal_aborted(journal))
5964 return -EROFS;
5965
5966 /* Wait for all existing dio workers */
5967 inode_dio_wait(inode);
5968
5969 /*
5970 * Before flushing the journal and switching inode's aops, we have
5971 * to flush all dirty data the inode has. There can be outstanding
5972 * delayed allocations, there can be unwritten extents created by
5973 * fallocate or buffered writes in dioread_nolock mode covered by
5974 * dirty data which can be converted only after flushing the dirty
5975 * data (and journalled aops don't know how to handle these cases).
5976 */
5977 if (val) {
5978 filemap_invalidate_lock(inode->i_mapping);
5979 err = filemap_write_and_wait(inode->i_mapping);
5980 if (err < 0) {
5981 filemap_invalidate_unlock(inode->i_mapping);
5982 return err;
5983 }
5984 }
5985
5986 alloc_ctx = ext4_writepages_down_write(inode->i_sb);
5987 jbd2_journal_lock_updates(journal);
5988
5989 /*
5990 * OK, there are no updates running now, and all cached data is
5991 * synced to disk. We are now in a completely consistent state
5992 * which doesn't have anything in the journal, and we know that
5993 * no filesystem updates are running, so it is safe to modify
5994 * the inode's in-core data-journaling state flag now.
5995 */
5996
5997 if (val)
5998 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5999 else {
6000 err = jbd2_journal_flush(journal, 0);
6001 if (err < 0) {
6002 jbd2_journal_unlock_updates(journal);
6003 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6004 return err;
6005 }
6006 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6007 }
6008 ext4_set_aops(inode);
6009
6010 jbd2_journal_unlock_updates(journal);
6011 ext4_writepages_up_write(inode->i_sb, alloc_ctx);
6012
6013 if (val)
6014 filemap_invalidate_unlock(inode->i_mapping);
6015
6016 /* Finally we can mark the inode as dirty. */
6017
6018 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6019 if (IS_ERR(handle))
6020 return PTR_ERR(handle);
6021
6022 ext4_fc_mark_ineligible(inode->i_sb,
6023 EXT4_FC_REASON_JOURNAL_FLAG_CHANGE, handle);
6024 err = ext4_mark_inode_dirty(handle, inode);
6025 ext4_handle_sync(handle);
6026 ext4_journal_stop(handle);
6027 ext4_std_error(inode->i_sb, err);
6028
6029 return err;
6030}
6031
6032static int ext4_bh_unmapped(handle_t *handle, struct inode *inode,
6033 struct buffer_head *bh)
6034{
6035 return !buffer_mapped(bh);
6036}
6037
6038vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
6039{
6040 struct vm_area_struct *vma = vmf->vma;
6041 struct folio *folio = page_folio(vmf->page);
6042 loff_t size;
6043 unsigned long len;
6044 int err;
6045 vm_fault_t ret;
6046 struct file *file = vma->vm_file;
6047 struct inode *inode = file_inode(file);
6048 struct address_space *mapping = inode->i_mapping;
6049 handle_t *handle;
6050 get_block_t *get_block;
6051 int retries = 0;
6052
6053 if (unlikely(IS_IMMUTABLE(inode)))
6054 return VM_FAULT_SIGBUS;
6055
6056 sb_start_pagefault(inode->i_sb);
6057 file_update_time(vma->vm_file);
6058
6059 filemap_invalidate_lock_shared(mapping);
6060
6061 err = ext4_convert_inline_data(inode);
6062 if (err)
6063 goto out_ret;
6064
6065 /*
6066 * On data journalling we skip straight to the transaction handle:
6067 * there's no delalloc; page truncated will be checked later; the
6068 * early return w/ all buffers mapped (calculates size/len) can't
6069 * be used; and there's no dioread_nolock, so only ext4_get_block.
6070 */
6071 if (ext4_should_journal_data(inode))
6072 goto retry_alloc;
6073
6074 /* Delalloc case is easy... */
6075 if (test_opt(inode->i_sb, DELALLOC) &&
6076 !ext4_nonda_switch(inode->i_sb)) {
6077 do {
6078 err = block_page_mkwrite(vma, vmf,
6079 ext4_da_get_block_prep);
6080 } while (err == -ENOSPC &&
6081 ext4_should_retry_alloc(inode->i_sb, &retries));
6082 goto out_ret;
6083 }
6084
6085 folio_lock(folio);
6086 size = i_size_read(inode);
6087 /* Page got truncated from under us? */
6088 if (folio->mapping != mapping || folio_pos(folio) > size) {
6089 folio_unlock(folio);
6090 ret = VM_FAULT_NOPAGE;
6091 goto out;
6092 }
6093
6094 len = folio_size(folio);
6095 if (folio_pos(folio) + len > size)
6096 len = size - folio_pos(folio);
6097 /*
6098 * Return if we have all the buffers mapped. This avoids the need to do
6099 * journal_start/journal_stop which can block and take a long time
6100 *
6101 * This cannot be done for data journalling, as we have to add the
6102 * inode to the transaction's list to writeprotect pages on commit.
6103 */
6104 if (folio_buffers(folio)) {
6105 if (!ext4_walk_page_buffers(NULL, inode, folio_buffers(folio),
6106 0, len, NULL,
6107 ext4_bh_unmapped)) {
6108 /* Wait so that we don't change page under IO */
6109 folio_wait_stable(folio);
6110 ret = VM_FAULT_LOCKED;
6111 goto out;
6112 }
6113 }
6114 folio_unlock(folio);
6115 /* OK, we need to fill the hole... */
6116 if (ext4_should_dioread_nolock(inode))
6117 get_block = ext4_get_block_unwritten;
6118 else
6119 get_block = ext4_get_block;
6120retry_alloc:
6121 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6122 ext4_writepage_trans_blocks(inode));
6123 if (IS_ERR(handle)) {
6124 ret = VM_FAULT_SIGBUS;
6125 goto out;
6126 }
6127 /*
6128 * Data journalling can't use block_page_mkwrite() because it
6129 * will set_buffer_dirty() before do_journal_get_write_access()
6130 * thus might hit warning messages for dirty metadata buffers.
6131 */
6132 if (!ext4_should_journal_data(inode)) {
6133 err = block_page_mkwrite(vma, vmf, get_block);
6134 } else {
6135 folio_lock(folio);
6136 size = i_size_read(inode);
6137 /* Page got truncated from under us? */
6138 if (folio->mapping != mapping || folio_pos(folio) > size) {
6139 ret = VM_FAULT_NOPAGE;
6140 goto out_error;
6141 }
6142
6143 len = folio_size(folio);
6144 if (folio_pos(folio) + len > size)
6145 len = size - folio_pos(folio);
6146
6147 err = __block_write_begin(&folio->page, 0, len, ext4_get_block);
6148 if (!err) {
6149 ret = VM_FAULT_SIGBUS;
6150 if (ext4_journal_folio_buffers(handle, folio, len))
6151 goto out_error;
6152 } else {
6153 folio_unlock(folio);
6154 }
6155 }
6156 ext4_journal_stop(handle);
6157 if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6158 goto retry_alloc;
6159out_ret:
6160 ret = vmf_fs_error(err);
6161out:
6162 filemap_invalidate_unlock_shared(mapping);
6163 sb_end_pagefault(inode->i_sb);
6164 return ret;
6165out_error:
6166 folio_unlock(folio);
6167 ext4_journal_stop(handle);
6168 goto out;
6169}