Loading...
1/*
2 * linux/fs/ext4/inode.c
3 *
4 * Copyright (C) 1992, 1993, 1994, 1995
5 * Remy Card (card@masi.ibp.fr)
6 * Laboratoire MASI - Institut Blaise Pascal
7 * Universite Pierre et Marie Curie (Paris VI)
8 *
9 * from
10 *
11 * linux/fs/minix/inode.c
12 *
13 * Copyright (C) 1991, 1992 Linus Torvalds
14 *
15 * 64-bit file support on 64-bit platforms by Jakub Jelinek
16 * (jj@sunsite.ms.mff.cuni.cz)
17 *
18 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19 */
20
21#include <linux/fs.h>
22#include <linux/time.h>
23#include <linux/jbd2.h>
24#include <linux/highuid.h>
25#include <linux/pagemap.h>
26#include <linux/quotaops.h>
27#include <linux/string.h>
28#include <linux/buffer_head.h>
29#include <linux/writeback.h>
30#include <linux/pagevec.h>
31#include <linux/mpage.h>
32#include <linux/namei.h>
33#include <linux/uio.h>
34#include <linux/bio.h>
35#include <linux/workqueue.h>
36#include <linux/kernel.h>
37#include <linux/printk.h>
38#include <linux/slab.h>
39#include <linux/ratelimit.h>
40
41#include "ext4_jbd2.h"
42#include "xattr.h"
43#include "acl.h"
44#include "truncate.h"
45
46#include <trace/events/ext4.h>
47
48#define MPAGE_DA_EXTENT_TAIL 0x01
49
50static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
51 struct ext4_inode_info *ei)
52{
53 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
54 __u16 csum_lo;
55 __u16 csum_hi = 0;
56 __u32 csum;
57
58 csum_lo = raw->i_checksum_lo;
59 raw->i_checksum_lo = 0;
60 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
61 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
62 csum_hi = raw->i_checksum_hi;
63 raw->i_checksum_hi = 0;
64 }
65
66 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw,
67 EXT4_INODE_SIZE(inode->i_sb));
68
69 raw->i_checksum_lo = csum_lo;
70 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
71 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
72 raw->i_checksum_hi = csum_hi;
73
74 return csum;
75}
76
77static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
78 struct ext4_inode_info *ei)
79{
80 __u32 provided, calculated;
81
82 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
83 cpu_to_le32(EXT4_OS_LINUX) ||
84 !EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
85 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
86 return 1;
87
88 provided = le16_to_cpu(raw->i_checksum_lo);
89 calculated = ext4_inode_csum(inode, raw, ei);
90 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
91 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
92 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
93 else
94 calculated &= 0xFFFF;
95
96 return provided == calculated;
97}
98
99static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
100 struct ext4_inode_info *ei)
101{
102 __u32 csum;
103
104 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
105 cpu_to_le32(EXT4_OS_LINUX) ||
106 !EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
107 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
108 return;
109
110 csum = ext4_inode_csum(inode, raw, ei);
111 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
112 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
113 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
114 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
115}
116
117static inline int ext4_begin_ordered_truncate(struct inode *inode,
118 loff_t new_size)
119{
120 trace_ext4_begin_ordered_truncate(inode, new_size);
121 /*
122 * If jinode is zero, then we never opened the file for
123 * writing, so there's no need to call
124 * jbd2_journal_begin_ordered_truncate() since there's no
125 * outstanding writes we need to flush.
126 */
127 if (!EXT4_I(inode)->jinode)
128 return 0;
129 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
130 EXT4_I(inode)->jinode,
131 new_size);
132}
133
134static void ext4_invalidatepage(struct page *page, unsigned long offset);
135static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
136 struct buffer_head *bh_result, int create);
137static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
138static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
139static int __ext4_journalled_writepage(struct page *page, unsigned int len);
140static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
141static int ext4_discard_partial_page_buffers_no_lock(handle_t *handle,
142 struct inode *inode, struct page *page, loff_t from,
143 loff_t length, int flags);
144
145/*
146 * Test whether an inode is a fast symlink.
147 */
148static int ext4_inode_is_fast_symlink(struct inode *inode)
149{
150 int ea_blocks = EXT4_I(inode)->i_file_acl ?
151 (inode->i_sb->s_blocksize >> 9) : 0;
152
153 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
154}
155
156/*
157 * Restart the transaction associated with *handle. This does a commit,
158 * so before we call here everything must be consistently dirtied against
159 * this transaction.
160 */
161int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
162 int nblocks)
163{
164 int ret;
165
166 /*
167 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
168 * moment, get_block can be called only for blocks inside i_size since
169 * page cache has been already dropped and writes are blocked by
170 * i_mutex. So we can safely drop the i_data_sem here.
171 */
172 BUG_ON(EXT4_JOURNAL(inode) == NULL);
173 jbd_debug(2, "restarting handle %p\n", handle);
174 up_write(&EXT4_I(inode)->i_data_sem);
175 ret = ext4_journal_restart(handle, nblocks);
176 down_write(&EXT4_I(inode)->i_data_sem);
177 ext4_discard_preallocations(inode);
178
179 return ret;
180}
181
182/*
183 * Called at the last iput() if i_nlink is zero.
184 */
185void ext4_evict_inode(struct inode *inode)
186{
187 handle_t *handle;
188 int err;
189
190 trace_ext4_evict_inode(inode);
191
192 ext4_ioend_wait(inode);
193
194 if (inode->i_nlink) {
195 /*
196 * When journalling data dirty buffers are tracked only in the
197 * journal. So although mm thinks everything is clean and
198 * ready for reaping the inode might still have some pages to
199 * write in the running transaction or waiting to be
200 * checkpointed. Thus calling jbd2_journal_invalidatepage()
201 * (via truncate_inode_pages()) to discard these buffers can
202 * cause data loss. Also even if we did not discard these
203 * buffers, we would have no way to find them after the inode
204 * is reaped and thus user could see stale data if he tries to
205 * read them before the transaction is checkpointed. So be
206 * careful and force everything to disk here... We use
207 * ei->i_datasync_tid to store the newest transaction
208 * containing inode's data.
209 *
210 * Note that directories do not have this problem because they
211 * don't use page cache.
212 */
213 if (ext4_should_journal_data(inode) &&
214 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
215 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
216 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
217
218 jbd2_log_start_commit(journal, commit_tid);
219 jbd2_log_wait_commit(journal, commit_tid);
220 filemap_write_and_wait(&inode->i_data);
221 }
222 truncate_inode_pages(&inode->i_data, 0);
223 goto no_delete;
224 }
225
226 if (!is_bad_inode(inode))
227 dquot_initialize(inode);
228
229 if (ext4_should_order_data(inode))
230 ext4_begin_ordered_truncate(inode, 0);
231 truncate_inode_pages(&inode->i_data, 0);
232
233 if (is_bad_inode(inode))
234 goto no_delete;
235
236 handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
237 if (IS_ERR(handle)) {
238 ext4_std_error(inode->i_sb, PTR_ERR(handle));
239 /*
240 * If we're going to skip the normal cleanup, we still need to
241 * make sure that the in-core orphan linked list is properly
242 * cleaned up.
243 */
244 ext4_orphan_del(NULL, inode);
245 goto no_delete;
246 }
247
248 if (IS_SYNC(inode))
249 ext4_handle_sync(handle);
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 ext4_truncate(inode);
259
260 /*
261 * ext4_ext_truncate() doesn't reserve any slop when it
262 * restarts journal transactions; therefore there may not be
263 * enough credits left in the handle to remove the inode from
264 * the orphan list and set the dtime field.
265 */
266 if (!ext4_handle_has_enough_credits(handle, 3)) {
267 err = ext4_journal_extend(handle, 3);
268 if (err > 0)
269 err = ext4_journal_restart(handle, 3);
270 if (err != 0) {
271 ext4_warning(inode->i_sb,
272 "couldn't extend journal (err %d)", err);
273 stop_handle:
274 ext4_journal_stop(handle);
275 ext4_orphan_del(NULL, inode);
276 goto no_delete;
277 }
278 }
279
280 /*
281 * Kill off the orphan record which ext4_truncate created.
282 * AKPM: I think this can be inside the above `if'.
283 * Note that ext4_orphan_del() has to be able to cope with the
284 * deletion of a non-existent orphan - this is because we don't
285 * know if ext4_truncate() actually created an orphan record.
286 * (Well, we could do this if we need to, but heck - it works)
287 */
288 ext4_orphan_del(handle, inode);
289 EXT4_I(inode)->i_dtime = get_seconds();
290
291 /*
292 * One subtle ordering requirement: if anything has gone wrong
293 * (transaction abort, IO errors, whatever), then we can still
294 * do these next steps (the fs will already have been marked as
295 * having errors), but we can't free the inode if the mark_dirty
296 * fails.
297 */
298 if (ext4_mark_inode_dirty(handle, inode))
299 /* If that failed, just do the required in-core inode clear. */
300 ext4_clear_inode(inode);
301 else
302 ext4_free_inode(handle, inode);
303 ext4_journal_stop(handle);
304 return;
305no_delete:
306 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
307}
308
309#ifdef CONFIG_QUOTA
310qsize_t *ext4_get_reserved_space(struct inode *inode)
311{
312 return &EXT4_I(inode)->i_reserved_quota;
313}
314#endif
315
316/*
317 * Calculate the number of metadata blocks need to reserve
318 * to allocate a block located at @lblock
319 */
320static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
321{
322 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
323 return ext4_ext_calc_metadata_amount(inode, lblock);
324
325 return ext4_ind_calc_metadata_amount(inode, lblock);
326}
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_msg(inode->i_sb, KERN_NOTICE, "%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 if (unlikely(ei->i_allocated_meta_blocks > ei->i_reserved_meta_blocks)) {
350 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, allocated %d "
351 "with only %d reserved metadata blocks\n", __func__,
352 inode->i_ino, ei->i_allocated_meta_blocks,
353 ei->i_reserved_meta_blocks);
354 WARN_ON(1);
355 ei->i_allocated_meta_blocks = ei->i_reserved_meta_blocks;
356 }
357
358 /* Update per-inode reservations */
359 ei->i_reserved_data_blocks -= used;
360 ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
361 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
362 used + ei->i_allocated_meta_blocks);
363 ei->i_allocated_meta_blocks = 0;
364
365 if (ei->i_reserved_data_blocks == 0) {
366 /*
367 * We can release all of the reserved metadata blocks
368 * only when we have written all of the delayed
369 * allocation blocks.
370 */
371 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
372 ei->i_reserved_meta_blocks);
373 ei->i_reserved_meta_blocks = 0;
374 ei->i_da_metadata_calc_len = 0;
375 }
376 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
377
378 /* Update quota subsystem for data blocks */
379 if (quota_claim)
380 dquot_claim_block(inode, EXT4_C2B(sbi, used));
381 else {
382 /*
383 * We did fallocate with an offset that is already delayed
384 * allocated. So on delayed allocated writeback we should
385 * not re-claim the quota for fallocated blocks.
386 */
387 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
388 }
389
390 /*
391 * If we have done all the pending block allocations and if
392 * there aren't any writers on the inode, we can discard the
393 * inode's preallocations.
394 */
395 if ((ei->i_reserved_data_blocks == 0) &&
396 (atomic_read(&inode->i_writecount) == 0))
397 ext4_discard_preallocations(inode);
398}
399
400static int __check_block_validity(struct inode *inode, const char *func,
401 unsigned int line,
402 struct ext4_map_blocks *map)
403{
404 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
405 map->m_len)) {
406 ext4_error_inode(inode, func, line, map->m_pblk,
407 "lblock %lu mapped to illegal pblock "
408 "(length %d)", (unsigned long) map->m_lblk,
409 map->m_len);
410 return -EIO;
411 }
412 return 0;
413}
414
415#define check_block_validity(inode, map) \
416 __check_block_validity((inode), __func__, __LINE__, (map))
417
418/*
419 * Return the number of contiguous dirty pages in a given inode
420 * starting at page frame idx.
421 */
422static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
423 unsigned int max_pages)
424{
425 struct address_space *mapping = inode->i_mapping;
426 pgoff_t index;
427 struct pagevec pvec;
428 pgoff_t num = 0;
429 int i, nr_pages, done = 0;
430
431 if (max_pages == 0)
432 return 0;
433 pagevec_init(&pvec, 0);
434 while (!done) {
435 index = idx;
436 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
437 PAGECACHE_TAG_DIRTY,
438 (pgoff_t)PAGEVEC_SIZE);
439 if (nr_pages == 0)
440 break;
441 for (i = 0; i < nr_pages; i++) {
442 struct page *page = pvec.pages[i];
443 struct buffer_head *bh, *head;
444
445 lock_page(page);
446 if (unlikely(page->mapping != mapping) ||
447 !PageDirty(page) ||
448 PageWriteback(page) ||
449 page->index != idx) {
450 done = 1;
451 unlock_page(page);
452 break;
453 }
454 if (page_has_buffers(page)) {
455 bh = head = page_buffers(page);
456 do {
457 if (!buffer_delay(bh) &&
458 !buffer_unwritten(bh))
459 done = 1;
460 bh = bh->b_this_page;
461 } while (!done && (bh != head));
462 }
463 unlock_page(page);
464 if (done)
465 break;
466 idx++;
467 num++;
468 if (num >= max_pages) {
469 done = 1;
470 break;
471 }
472 }
473 pagevec_release(&pvec);
474 }
475 return num;
476}
477
478/*
479 * Sets the BH_Da_Mapped bit on the buffer heads corresponding to the given map.
480 */
481static void set_buffers_da_mapped(struct inode *inode,
482 struct ext4_map_blocks *map)
483{
484 struct address_space *mapping = inode->i_mapping;
485 struct pagevec pvec;
486 int i, nr_pages;
487 pgoff_t index, end;
488
489 index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
490 end = (map->m_lblk + map->m_len - 1) >>
491 (PAGE_CACHE_SHIFT - inode->i_blkbits);
492
493 pagevec_init(&pvec, 0);
494 while (index <= end) {
495 nr_pages = pagevec_lookup(&pvec, mapping, index,
496 min(end - index + 1,
497 (pgoff_t)PAGEVEC_SIZE));
498 if (nr_pages == 0)
499 break;
500 for (i = 0; i < nr_pages; i++) {
501 struct page *page = pvec.pages[i];
502 struct buffer_head *bh, *head;
503
504 if (unlikely(page->mapping != mapping) ||
505 !PageDirty(page))
506 break;
507
508 if (page_has_buffers(page)) {
509 bh = head = page_buffers(page);
510 do {
511 set_buffer_da_mapped(bh);
512 bh = bh->b_this_page;
513 } while (bh != head);
514 }
515 index++;
516 }
517 pagevec_release(&pvec);
518 }
519}
520
521/*
522 * The ext4_map_blocks() function tries to look up the requested blocks,
523 * and returns if the blocks are already mapped.
524 *
525 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
526 * and store the allocated blocks in the result buffer head and mark it
527 * mapped.
528 *
529 * If file type is extents based, it will call ext4_ext_map_blocks(),
530 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
531 * based files
532 *
533 * On success, it returns the number of blocks being mapped or allocate.
534 * if create==0 and the blocks are pre-allocated and uninitialized block,
535 * the result buffer head is unmapped. If the create ==1, it will make sure
536 * the buffer head is mapped.
537 *
538 * It returns 0 if plain look up failed (blocks have not been allocated), in
539 * that case, buffer head is unmapped
540 *
541 * It returns the error in case of allocation failure.
542 */
543int ext4_map_blocks(handle_t *handle, struct inode *inode,
544 struct ext4_map_blocks *map, int flags)
545{
546 int retval;
547
548 map->m_flags = 0;
549 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
550 "logical block %lu\n", inode->i_ino, flags, map->m_len,
551 (unsigned long) map->m_lblk);
552 /*
553 * Try to see if we can get the block without requesting a new
554 * file system block.
555 */
556 down_read((&EXT4_I(inode)->i_data_sem));
557 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
558 retval = ext4_ext_map_blocks(handle, inode, map, flags &
559 EXT4_GET_BLOCKS_KEEP_SIZE);
560 } else {
561 retval = ext4_ind_map_blocks(handle, inode, map, flags &
562 EXT4_GET_BLOCKS_KEEP_SIZE);
563 }
564 up_read((&EXT4_I(inode)->i_data_sem));
565
566 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
567 int ret = check_block_validity(inode, map);
568 if (ret != 0)
569 return ret;
570 }
571
572 /* If it is only a block(s) look up */
573 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
574 return retval;
575
576 /*
577 * Returns if the blocks have already allocated
578 *
579 * Note that if blocks have been preallocated
580 * ext4_ext_get_block() returns the create = 0
581 * with buffer head unmapped.
582 */
583 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
584 return retval;
585
586 /*
587 * When we call get_blocks without the create flag, the
588 * BH_Unwritten flag could have gotten set if the blocks
589 * requested were part of a uninitialized extent. We need to
590 * clear this flag now that we are committed to convert all or
591 * part of the uninitialized extent to be an initialized
592 * extent. This is because we need to avoid the combination
593 * of BH_Unwritten and BH_Mapped flags being simultaneously
594 * set on the buffer_head.
595 */
596 map->m_flags &= ~EXT4_MAP_UNWRITTEN;
597
598 /*
599 * New blocks allocate and/or writing to uninitialized extent
600 * will possibly result in updating i_data, so we take
601 * the write lock of i_data_sem, and call get_blocks()
602 * with create == 1 flag.
603 */
604 down_write((&EXT4_I(inode)->i_data_sem));
605
606 /*
607 * if the caller is from delayed allocation writeout path
608 * we have already reserved fs blocks for allocation
609 * let the underlying get_block() function know to
610 * avoid double accounting
611 */
612 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
613 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
614 /*
615 * We need to check for EXT4 here because migrate
616 * could have changed the inode type in between
617 */
618 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
619 retval = ext4_ext_map_blocks(handle, inode, map, flags);
620 } else {
621 retval = ext4_ind_map_blocks(handle, inode, map, flags);
622
623 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
624 /*
625 * We allocated new blocks which will result in
626 * i_data's format changing. Force the migrate
627 * to fail by clearing migrate flags
628 */
629 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
630 }
631
632 /*
633 * Update reserved blocks/metadata blocks after successful
634 * block allocation which had been deferred till now. We don't
635 * support fallocate for non extent files. So we can update
636 * reserve space here.
637 */
638 if ((retval > 0) &&
639 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
640 ext4_da_update_reserve_space(inode, retval, 1);
641 }
642 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
643 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
644
645 /* If we have successfully mapped the delayed allocated blocks,
646 * set the BH_Da_Mapped bit on them. Its important to do this
647 * under the protection of i_data_sem.
648 */
649 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
650 set_buffers_da_mapped(inode, map);
651 }
652
653 up_write((&EXT4_I(inode)->i_data_sem));
654 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
655 int ret = check_block_validity(inode, map);
656 if (ret != 0)
657 return ret;
658 }
659 return retval;
660}
661
662/* Maximum number of blocks we map for direct IO at once. */
663#define DIO_MAX_BLOCKS 4096
664
665static int _ext4_get_block(struct inode *inode, sector_t iblock,
666 struct buffer_head *bh, int flags)
667{
668 handle_t *handle = ext4_journal_current_handle();
669 struct ext4_map_blocks map;
670 int ret = 0, started = 0;
671 int dio_credits;
672
673 map.m_lblk = iblock;
674 map.m_len = bh->b_size >> inode->i_blkbits;
675
676 if (flags && !handle) {
677 /* Direct IO write... */
678 if (map.m_len > DIO_MAX_BLOCKS)
679 map.m_len = DIO_MAX_BLOCKS;
680 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
681 handle = ext4_journal_start(inode, dio_credits);
682 if (IS_ERR(handle)) {
683 ret = PTR_ERR(handle);
684 return ret;
685 }
686 started = 1;
687 }
688
689 ret = ext4_map_blocks(handle, inode, &map, flags);
690 if (ret > 0) {
691 map_bh(bh, inode->i_sb, map.m_pblk);
692 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
693 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
694 ret = 0;
695 }
696 if (started)
697 ext4_journal_stop(handle);
698 return ret;
699}
700
701int ext4_get_block(struct inode *inode, sector_t iblock,
702 struct buffer_head *bh, int create)
703{
704 return _ext4_get_block(inode, iblock, bh,
705 create ? EXT4_GET_BLOCKS_CREATE : 0);
706}
707
708/*
709 * `handle' can be NULL if create is zero
710 */
711struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
712 ext4_lblk_t block, int create, int *errp)
713{
714 struct ext4_map_blocks map;
715 struct buffer_head *bh;
716 int fatal = 0, err;
717
718 J_ASSERT(handle != NULL || create == 0);
719
720 map.m_lblk = block;
721 map.m_len = 1;
722 err = ext4_map_blocks(handle, inode, &map,
723 create ? EXT4_GET_BLOCKS_CREATE : 0);
724
725 if (err < 0)
726 *errp = err;
727 if (err <= 0)
728 return NULL;
729 *errp = 0;
730
731 bh = sb_getblk(inode->i_sb, map.m_pblk);
732 if (!bh) {
733 *errp = -EIO;
734 return NULL;
735 }
736 if (map.m_flags & EXT4_MAP_NEW) {
737 J_ASSERT(create != 0);
738 J_ASSERT(handle != NULL);
739
740 /*
741 * Now that we do not always journal data, we should
742 * keep in mind whether this should always journal the
743 * new buffer as metadata. For now, regular file
744 * writes use ext4_get_block instead, so it's not a
745 * problem.
746 */
747 lock_buffer(bh);
748 BUFFER_TRACE(bh, "call get_create_access");
749 fatal = ext4_journal_get_create_access(handle, bh);
750 if (!fatal && !buffer_uptodate(bh)) {
751 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
752 set_buffer_uptodate(bh);
753 }
754 unlock_buffer(bh);
755 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
756 err = ext4_handle_dirty_metadata(handle, inode, bh);
757 if (!fatal)
758 fatal = err;
759 } else {
760 BUFFER_TRACE(bh, "not a new buffer");
761 }
762 if (fatal) {
763 *errp = fatal;
764 brelse(bh);
765 bh = NULL;
766 }
767 return bh;
768}
769
770struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
771 ext4_lblk_t block, int create, int *err)
772{
773 struct buffer_head *bh;
774
775 bh = ext4_getblk(handle, inode, block, create, err);
776 if (!bh)
777 return bh;
778 if (buffer_uptodate(bh))
779 return bh;
780 ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
781 wait_on_buffer(bh);
782 if (buffer_uptodate(bh))
783 return bh;
784 put_bh(bh);
785 *err = -EIO;
786 return NULL;
787}
788
789static int walk_page_buffers(handle_t *handle,
790 struct buffer_head *head,
791 unsigned from,
792 unsigned to,
793 int *partial,
794 int (*fn)(handle_t *handle,
795 struct buffer_head *bh))
796{
797 struct buffer_head *bh;
798 unsigned block_start, block_end;
799 unsigned blocksize = head->b_size;
800 int err, ret = 0;
801 struct buffer_head *next;
802
803 for (bh = head, block_start = 0;
804 ret == 0 && (bh != head || !block_start);
805 block_start = block_end, bh = next) {
806 next = bh->b_this_page;
807 block_end = block_start + blocksize;
808 if (block_end <= from || block_start >= to) {
809 if (partial && !buffer_uptodate(bh))
810 *partial = 1;
811 continue;
812 }
813 err = (*fn)(handle, bh);
814 if (!ret)
815 ret = err;
816 }
817 return ret;
818}
819
820/*
821 * To preserve ordering, it is essential that the hole instantiation and
822 * the data write be encapsulated in a single transaction. We cannot
823 * close off a transaction and start a new one between the ext4_get_block()
824 * and the commit_write(). So doing the jbd2_journal_start at the start of
825 * prepare_write() is the right place.
826 *
827 * Also, this function can nest inside ext4_writepage() ->
828 * block_write_full_page(). In that case, we *know* that ext4_writepage()
829 * has generated enough buffer credits to do the whole page. So we won't
830 * block on the journal in that case, which is good, because the caller may
831 * be PF_MEMALLOC.
832 *
833 * By accident, ext4 can be reentered when a transaction is open via
834 * quota file writes. If we were to commit the transaction while thus
835 * reentered, there can be a deadlock - we would be holding a quota
836 * lock, and the commit would never complete if another thread had a
837 * transaction open and was blocking on the quota lock - a ranking
838 * violation.
839 *
840 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
841 * will _not_ run commit under these circumstances because handle->h_ref
842 * is elevated. We'll still have enough credits for the tiny quotafile
843 * write.
844 */
845static int do_journal_get_write_access(handle_t *handle,
846 struct buffer_head *bh)
847{
848 int dirty = buffer_dirty(bh);
849 int ret;
850
851 if (!buffer_mapped(bh) || buffer_freed(bh))
852 return 0;
853 /*
854 * __block_write_begin() could have dirtied some buffers. Clean
855 * the dirty bit as jbd2_journal_get_write_access() could complain
856 * otherwise about fs integrity issues. Setting of the dirty bit
857 * by __block_write_begin() isn't a real problem here as we clear
858 * the bit before releasing a page lock and thus writeback cannot
859 * ever write the buffer.
860 */
861 if (dirty)
862 clear_buffer_dirty(bh);
863 ret = ext4_journal_get_write_access(handle, bh);
864 if (!ret && dirty)
865 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
866 return ret;
867}
868
869static int ext4_get_block_write(struct inode *inode, sector_t iblock,
870 struct buffer_head *bh_result, int create);
871static int ext4_write_begin(struct file *file, struct address_space *mapping,
872 loff_t pos, unsigned len, unsigned flags,
873 struct page **pagep, void **fsdata)
874{
875 struct inode *inode = mapping->host;
876 int ret, needed_blocks;
877 handle_t *handle;
878 int retries = 0;
879 struct page *page;
880 pgoff_t index;
881 unsigned from, to;
882
883 trace_ext4_write_begin(inode, pos, len, flags);
884 /*
885 * Reserve one block more for addition to orphan list in case
886 * we allocate blocks but write fails for some reason
887 */
888 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
889 index = pos >> PAGE_CACHE_SHIFT;
890 from = pos & (PAGE_CACHE_SIZE - 1);
891 to = from + len;
892
893retry:
894 handle = ext4_journal_start(inode, needed_blocks);
895 if (IS_ERR(handle)) {
896 ret = PTR_ERR(handle);
897 goto out;
898 }
899
900 /* We cannot recurse into the filesystem as the transaction is already
901 * started */
902 flags |= AOP_FLAG_NOFS;
903
904 page = grab_cache_page_write_begin(mapping, index, flags);
905 if (!page) {
906 ext4_journal_stop(handle);
907 ret = -ENOMEM;
908 goto out;
909 }
910 *pagep = page;
911
912 if (ext4_should_dioread_nolock(inode))
913 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
914 else
915 ret = __block_write_begin(page, pos, len, ext4_get_block);
916
917 if (!ret && ext4_should_journal_data(inode)) {
918 ret = walk_page_buffers(handle, page_buffers(page),
919 from, to, NULL, do_journal_get_write_access);
920 }
921
922 if (ret) {
923 unlock_page(page);
924 page_cache_release(page);
925 /*
926 * __block_write_begin may have instantiated a few blocks
927 * outside i_size. Trim these off again. Don't need
928 * i_size_read because we hold i_mutex.
929 *
930 * Add inode to orphan list in case we crash before
931 * truncate finishes
932 */
933 if (pos + len > inode->i_size && ext4_can_truncate(inode))
934 ext4_orphan_add(handle, inode);
935
936 ext4_journal_stop(handle);
937 if (pos + len > inode->i_size) {
938 ext4_truncate_failed_write(inode);
939 /*
940 * If truncate failed early the inode might
941 * still be on the orphan list; we need to
942 * make sure the inode is removed from the
943 * orphan list in that case.
944 */
945 if (inode->i_nlink)
946 ext4_orphan_del(NULL, inode);
947 }
948 }
949
950 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
951 goto retry;
952out:
953 return ret;
954}
955
956/* For write_end() in data=journal mode */
957static int write_end_fn(handle_t *handle, struct buffer_head *bh)
958{
959 if (!buffer_mapped(bh) || buffer_freed(bh))
960 return 0;
961 set_buffer_uptodate(bh);
962 return ext4_handle_dirty_metadata(handle, NULL, bh);
963}
964
965static int ext4_generic_write_end(struct file *file,
966 struct address_space *mapping,
967 loff_t pos, unsigned len, unsigned copied,
968 struct page *page, void *fsdata)
969{
970 int i_size_changed = 0;
971 struct inode *inode = mapping->host;
972 handle_t *handle = ext4_journal_current_handle();
973
974 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
975
976 /*
977 * No need to use i_size_read() here, the i_size
978 * cannot change under us because we hold i_mutex.
979 *
980 * But it's important to update i_size while still holding page lock:
981 * page writeout could otherwise come in and zero beyond i_size.
982 */
983 if (pos + copied > inode->i_size) {
984 i_size_write(inode, pos + copied);
985 i_size_changed = 1;
986 }
987
988 if (pos + copied > EXT4_I(inode)->i_disksize) {
989 /* We need to mark inode dirty even if
990 * new_i_size is less that inode->i_size
991 * bu greater than i_disksize.(hint delalloc)
992 */
993 ext4_update_i_disksize(inode, (pos + copied));
994 i_size_changed = 1;
995 }
996 unlock_page(page);
997 page_cache_release(page);
998
999 /*
1000 * Don't mark the inode dirty under page lock. First, it unnecessarily
1001 * makes the holding time of page lock longer. Second, it forces lock
1002 * ordering of page lock and transaction start for journaling
1003 * filesystems.
1004 */
1005 if (i_size_changed)
1006 ext4_mark_inode_dirty(handle, inode);
1007
1008 return copied;
1009}
1010
1011/*
1012 * We need to pick up the new inode size which generic_commit_write gave us
1013 * `file' can be NULL - eg, when called from page_symlink().
1014 *
1015 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1016 * buffers are managed internally.
1017 */
1018static int ext4_ordered_write_end(struct file *file,
1019 struct address_space *mapping,
1020 loff_t pos, unsigned len, unsigned copied,
1021 struct page *page, void *fsdata)
1022{
1023 handle_t *handle = ext4_journal_current_handle();
1024 struct inode *inode = mapping->host;
1025 int ret = 0, ret2;
1026
1027 trace_ext4_ordered_write_end(inode, pos, len, copied);
1028 ret = ext4_jbd2_file_inode(handle, inode);
1029
1030 if (ret == 0) {
1031 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1032 page, fsdata);
1033 copied = ret2;
1034 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1035 /* if we have allocated more blocks and copied
1036 * less. We will have blocks allocated outside
1037 * inode->i_size. So truncate them
1038 */
1039 ext4_orphan_add(handle, inode);
1040 if (ret2 < 0)
1041 ret = ret2;
1042 } else {
1043 unlock_page(page);
1044 page_cache_release(page);
1045 }
1046
1047 ret2 = ext4_journal_stop(handle);
1048 if (!ret)
1049 ret = ret2;
1050
1051 if (pos + len > inode->i_size) {
1052 ext4_truncate_failed_write(inode);
1053 /*
1054 * If truncate failed early the inode might still be
1055 * on the orphan list; we need to make sure the inode
1056 * is removed from the orphan list in that case.
1057 */
1058 if (inode->i_nlink)
1059 ext4_orphan_del(NULL, inode);
1060 }
1061
1062
1063 return ret ? ret : copied;
1064}
1065
1066static int ext4_writeback_write_end(struct file *file,
1067 struct address_space *mapping,
1068 loff_t pos, unsigned len, unsigned copied,
1069 struct page *page, void *fsdata)
1070{
1071 handle_t *handle = ext4_journal_current_handle();
1072 struct inode *inode = mapping->host;
1073 int ret = 0, ret2;
1074
1075 trace_ext4_writeback_write_end(inode, pos, len, copied);
1076 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1077 page, fsdata);
1078 copied = ret2;
1079 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1080 /* if we have allocated more blocks and copied
1081 * less. We will have blocks allocated outside
1082 * inode->i_size. So truncate them
1083 */
1084 ext4_orphan_add(handle, inode);
1085
1086 if (ret2 < 0)
1087 ret = ret2;
1088
1089 ret2 = ext4_journal_stop(handle);
1090 if (!ret)
1091 ret = ret2;
1092
1093 if (pos + len > inode->i_size) {
1094 ext4_truncate_failed_write(inode);
1095 /*
1096 * If truncate failed early the inode might still be
1097 * on the orphan list; we need to make sure the inode
1098 * is removed from the orphan list in that case.
1099 */
1100 if (inode->i_nlink)
1101 ext4_orphan_del(NULL, inode);
1102 }
1103
1104 return ret ? ret : copied;
1105}
1106
1107static int ext4_journalled_write_end(struct file *file,
1108 struct address_space *mapping,
1109 loff_t pos, unsigned len, unsigned copied,
1110 struct page *page, void *fsdata)
1111{
1112 handle_t *handle = ext4_journal_current_handle();
1113 struct inode *inode = mapping->host;
1114 int ret = 0, ret2;
1115 int partial = 0;
1116 unsigned from, to;
1117 loff_t new_i_size;
1118
1119 trace_ext4_journalled_write_end(inode, pos, len, copied);
1120 from = pos & (PAGE_CACHE_SIZE - 1);
1121 to = from + len;
1122
1123 BUG_ON(!ext4_handle_valid(handle));
1124
1125 if (copied < len) {
1126 if (!PageUptodate(page))
1127 copied = 0;
1128 page_zero_new_buffers(page, from+copied, to);
1129 }
1130
1131 ret = walk_page_buffers(handle, page_buffers(page), from,
1132 to, &partial, write_end_fn);
1133 if (!partial)
1134 SetPageUptodate(page);
1135 new_i_size = pos + copied;
1136 if (new_i_size > inode->i_size)
1137 i_size_write(inode, pos+copied);
1138 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1139 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1140 if (new_i_size > EXT4_I(inode)->i_disksize) {
1141 ext4_update_i_disksize(inode, new_i_size);
1142 ret2 = ext4_mark_inode_dirty(handle, inode);
1143 if (!ret)
1144 ret = ret2;
1145 }
1146
1147 unlock_page(page);
1148 page_cache_release(page);
1149 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1150 /* if we have allocated more blocks and copied
1151 * less. We will have blocks allocated outside
1152 * inode->i_size. So truncate them
1153 */
1154 ext4_orphan_add(handle, inode);
1155
1156 ret2 = ext4_journal_stop(handle);
1157 if (!ret)
1158 ret = ret2;
1159 if (pos + len > inode->i_size) {
1160 ext4_truncate_failed_write(inode);
1161 /*
1162 * If truncate failed early the inode might still be
1163 * on the orphan list; we need to make sure the inode
1164 * is removed from the orphan list in that case.
1165 */
1166 if (inode->i_nlink)
1167 ext4_orphan_del(NULL, inode);
1168 }
1169
1170 return ret ? ret : copied;
1171}
1172
1173/*
1174 * Reserve a single cluster located at lblock
1175 */
1176static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1177{
1178 int retries = 0;
1179 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1180 struct ext4_inode_info *ei = EXT4_I(inode);
1181 unsigned int md_needed;
1182 int ret;
1183 ext4_lblk_t save_last_lblock;
1184 int save_len;
1185
1186 /*
1187 * We will charge metadata quota at writeout time; this saves
1188 * us from metadata over-estimation, though we may go over by
1189 * a small amount in the end. Here we just reserve for data.
1190 */
1191 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1192 if (ret)
1193 return ret;
1194
1195 /*
1196 * recalculate the amount of metadata blocks to reserve
1197 * in order to allocate nrblocks
1198 * worse case is one extent per block
1199 */
1200repeat:
1201 spin_lock(&ei->i_block_reservation_lock);
1202 /*
1203 * ext4_calc_metadata_amount() has side effects, which we have
1204 * to be prepared undo if we fail to claim space.
1205 */
1206 save_len = ei->i_da_metadata_calc_len;
1207 save_last_lblock = ei->i_da_metadata_calc_last_lblock;
1208 md_needed = EXT4_NUM_B2C(sbi,
1209 ext4_calc_metadata_amount(inode, lblock));
1210 trace_ext4_da_reserve_space(inode, md_needed);
1211
1212 /*
1213 * We do still charge estimated metadata to the sb though;
1214 * we cannot afford to run out of free blocks.
1215 */
1216 if (ext4_claim_free_clusters(sbi, md_needed + 1, 0)) {
1217 ei->i_da_metadata_calc_len = save_len;
1218 ei->i_da_metadata_calc_last_lblock = save_last_lblock;
1219 spin_unlock(&ei->i_block_reservation_lock);
1220 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1221 yield();
1222 goto repeat;
1223 }
1224 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1225 return -ENOSPC;
1226 }
1227 ei->i_reserved_data_blocks++;
1228 ei->i_reserved_meta_blocks += md_needed;
1229 spin_unlock(&ei->i_block_reservation_lock);
1230
1231 return 0; /* success */
1232}
1233
1234static void ext4_da_release_space(struct inode *inode, int to_free)
1235{
1236 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1237 struct ext4_inode_info *ei = EXT4_I(inode);
1238
1239 if (!to_free)
1240 return; /* Nothing to release, exit */
1241
1242 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1243
1244 trace_ext4_da_release_space(inode, to_free);
1245 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1246 /*
1247 * if there aren't enough reserved blocks, then the
1248 * counter is messed up somewhere. Since this
1249 * function is called from invalidate page, it's
1250 * harmless to return without any action.
1251 */
1252 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1253 "ino %lu, to_free %d with only %d reserved "
1254 "data blocks", inode->i_ino, to_free,
1255 ei->i_reserved_data_blocks);
1256 WARN_ON(1);
1257 to_free = ei->i_reserved_data_blocks;
1258 }
1259 ei->i_reserved_data_blocks -= to_free;
1260
1261 if (ei->i_reserved_data_blocks == 0) {
1262 /*
1263 * We can release all of the reserved metadata blocks
1264 * only when we have written all of the delayed
1265 * allocation blocks.
1266 * Note that in case of bigalloc, i_reserved_meta_blocks,
1267 * i_reserved_data_blocks, etc. refer to number of clusters.
1268 */
1269 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
1270 ei->i_reserved_meta_blocks);
1271 ei->i_reserved_meta_blocks = 0;
1272 ei->i_da_metadata_calc_len = 0;
1273 }
1274
1275 /* update fs dirty data blocks counter */
1276 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1277
1278 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1279
1280 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1281}
1282
1283static void ext4_da_page_release_reservation(struct page *page,
1284 unsigned long offset)
1285{
1286 int to_release = 0;
1287 struct buffer_head *head, *bh;
1288 unsigned int curr_off = 0;
1289 struct inode *inode = page->mapping->host;
1290 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1291 int num_clusters;
1292
1293 head = page_buffers(page);
1294 bh = head;
1295 do {
1296 unsigned int next_off = curr_off + bh->b_size;
1297
1298 if ((offset <= curr_off) && (buffer_delay(bh))) {
1299 to_release++;
1300 clear_buffer_delay(bh);
1301 clear_buffer_da_mapped(bh);
1302 }
1303 curr_off = next_off;
1304 } while ((bh = bh->b_this_page) != head);
1305
1306 /* If we have released all the blocks belonging to a cluster, then we
1307 * need to release the reserved space for that cluster. */
1308 num_clusters = EXT4_NUM_B2C(sbi, to_release);
1309 while (num_clusters > 0) {
1310 ext4_fsblk_t lblk;
1311 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1312 ((num_clusters - 1) << sbi->s_cluster_bits);
1313 if (sbi->s_cluster_ratio == 1 ||
1314 !ext4_find_delalloc_cluster(inode, lblk, 1))
1315 ext4_da_release_space(inode, 1);
1316
1317 num_clusters--;
1318 }
1319}
1320
1321/*
1322 * Delayed allocation stuff
1323 */
1324
1325/*
1326 * mpage_da_submit_io - walks through extent of pages and try to write
1327 * them with writepage() call back
1328 *
1329 * @mpd->inode: inode
1330 * @mpd->first_page: first page of the extent
1331 * @mpd->next_page: page after the last page of the extent
1332 *
1333 * By the time mpage_da_submit_io() is called we expect all blocks
1334 * to be allocated. this may be wrong if allocation failed.
1335 *
1336 * As pages are already locked by write_cache_pages(), we can't use it
1337 */
1338static int mpage_da_submit_io(struct mpage_da_data *mpd,
1339 struct ext4_map_blocks *map)
1340{
1341 struct pagevec pvec;
1342 unsigned long index, end;
1343 int ret = 0, err, nr_pages, i;
1344 struct inode *inode = mpd->inode;
1345 struct address_space *mapping = inode->i_mapping;
1346 loff_t size = i_size_read(inode);
1347 unsigned int len, block_start;
1348 struct buffer_head *bh, *page_bufs = NULL;
1349 int journal_data = ext4_should_journal_data(inode);
1350 sector_t pblock = 0, cur_logical = 0;
1351 struct ext4_io_submit io_submit;
1352
1353 BUG_ON(mpd->next_page <= mpd->first_page);
1354 memset(&io_submit, 0, sizeof(io_submit));
1355 /*
1356 * We need to start from the first_page to the next_page - 1
1357 * to make sure we also write the mapped dirty buffer_heads.
1358 * If we look at mpd->b_blocknr we would only be looking
1359 * at the currently mapped buffer_heads.
1360 */
1361 index = mpd->first_page;
1362 end = mpd->next_page - 1;
1363
1364 pagevec_init(&pvec, 0);
1365 while (index <= end) {
1366 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1367 if (nr_pages == 0)
1368 break;
1369 for (i = 0; i < nr_pages; i++) {
1370 int commit_write = 0, skip_page = 0;
1371 struct page *page = pvec.pages[i];
1372
1373 index = page->index;
1374 if (index > end)
1375 break;
1376
1377 if (index == size >> PAGE_CACHE_SHIFT)
1378 len = size & ~PAGE_CACHE_MASK;
1379 else
1380 len = PAGE_CACHE_SIZE;
1381 if (map) {
1382 cur_logical = index << (PAGE_CACHE_SHIFT -
1383 inode->i_blkbits);
1384 pblock = map->m_pblk + (cur_logical -
1385 map->m_lblk);
1386 }
1387 index++;
1388
1389 BUG_ON(!PageLocked(page));
1390 BUG_ON(PageWriteback(page));
1391
1392 /*
1393 * If the page does not have buffers (for
1394 * whatever reason), try to create them using
1395 * __block_write_begin. If this fails,
1396 * skip the page and move on.
1397 */
1398 if (!page_has_buffers(page)) {
1399 if (__block_write_begin(page, 0, len,
1400 noalloc_get_block_write)) {
1401 skip_page:
1402 unlock_page(page);
1403 continue;
1404 }
1405 commit_write = 1;
1406 }
1407
1408 bh = page_bufs = page_buffers(page);
1409 block_start = 0;
1410 do {
1411 if (!bh)
1412 goto skip_page;
1413 if (map && (cur_logical >= map->m_lblk) &&
1414 (cur_logical <= (map->m_lblk +
1415 (map->m_len - 1)))) {
1416 if (buffer_delay(bh)) {
1417 clear_buffer_delay(bh);
1418 bh->b_blocknr = pblock;
1419 }
1420 if (buffer_da_mapped(bh))
1421 clear_buffer_da_mapped(bh);
1422 if (buffer_unwritten(bh) ||
1423 buffer_mapped(bh))
1424 BUG_ON(bh->b_blocknr != pblock);
1425 if (map->m_flags & EXT4_MAP_UNINIT)
1426 set_buffer_uninit(bh);
1427 clear_buffer_unwritten(bh);
1428 }
1429
1430 /*
1431 * skip page if block allocation undone and
1432 * block is dirty
1433 */
1434 if (ext4_bh_delay_or_unwritten(NULL, bh))
1435 skip_page = 1;
1436 bh = bh->b_this_page;
1437 block_start += bh->b_size;
1438 cur_logical++;
1439 pblock++;
1440 } while (bh != page_bufs);
1441
1442 if (skip_page)
1443 goto skip_page;
1444
1445 if (commit_write)
1446 /* mark the buffer_heads as dirty & uptodate */
1447 block_commit_write(page, 0, len);
1448
1449 clear_page_dirty_for_io(page);
1450 /*
1451 * Delalloc doesn't support data journalling,
1452 * but eventually maybe we'll lift this
1453 * restriction.
1454 */
1455 if (unlikely(journal_data && PageChecked(page)))
1456 err = __ext4_journalled_writepage(page, len);
1457 else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1458 err = ext4_bio_write_page(&io_submit, page,
1459 len, mpd->wbc);
1460 else if (buffer_uninit(page_bufs)) {
1461 ext4_set_bh_endio(page_bufs, inode);
1462 err = block_write_full_page_endio(page,
1463 noalloc_get_block_write,
1464 mpd->wbc, ext4_end_io_buffer_write);
1465 } else
1466 err = block_write_full_page(page,
1467 noalloc_get_block_write, mpd->wbc);
1468
1469 if (!err)
1470 mpd->pages_written++;
1471 /*
1472 * In error case, we have to continue because
1473 * remaining pages are still locked
1474 */
1475 if (ret == 0)
1476 ret = err;
1477 }
1478 pagevec_release(&pvec);
1479 }
1480 ext4_io_submit(&io_submit);
1481 return ret;
1482}
1483
1484static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1485{
1486 int nr_pages, i;
1487 pgoff_t index, end;
1488 struct pagevec pvec;
1489 struct inode *inode = mpd->inode;
1490 struct address_space *mapping = inode->i_mapping;
1491
1492 index = mpd->first_page;
1493 end = mpd->next_page - 1;
1494 while (index <= end) {
1495 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1496 if (nr_pages == 0)
1497 break;
1498 for (i = 0; i < nr_pages; i++) {
1499 struct page *page = pvec.pages[i];
1500 if (page->index > end)
1501 break;
1502 BUG_ON(!PageLocked(page));
1503 BUG_ON(PageWriteback(page));
1504 block_invalidatepage(page, 0);
1505 ClearPageUptodate(page);
1506 unlock_page(page);
1507 }
1508 index = pvec.pages[nr_pages - 1]->index + 1;
1509 pagevec_release(&pvec);
1510 }
1511 return;
1512}
1513
1514static void ext4_print_free_blocks(struct inode *inode)
1515{
1516 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1517 struct super_block *sb = inode->i_sb;
1518
1519 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1520 EXT4_C2B(EXT4_SB(inode->i_sb),
1521 ext4_count_free_clusters(inode->i_sb)));
1522 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1523 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1524 (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1525 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1526 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1527 (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1528 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1529 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1530 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1531 EXT4_I(inode)->i_reserved_data_blocks);
1532 ext4_msg(sb, KERN_CRIT, "i_reserved_meta_blocks=%u",
1533 EXT4_I(inode)->i_reserved_meta_blocks);
1534 return;
1535}
1536
1537/*
1538 * mpage_da_map_and_submit - go through given space, map them
1539 * if necessary, and then submit them for I/O
1540 *
1541 * @mpd - bh describing space
1542 *
1543 * The function skips space we know is already mapped to disk blocks.
1544 *
1545 */
1546static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1547{
1548 int err, blks, get_blocks_flags;
1549 struct ext4_map_blocks map, *mapp = NULL;
1550 sector_t next = mpd->b_blocknr;
1551 unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1552 loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1553 handle_t *handle = NULL;
1554
1555 /*
1556 * If the blocks are mapped already, or we couldn't accumulate
1557 * any blocks, then proceed immediately to the submission stage.
1558 */
1559 if ((mpd->b_size == 0) ||
1560 ((mpd->b_state & (1 << BH_Mapped)) &&
1561 !(mpd->b_state & (1 << BH_Delay)) &&
1562 !(mpd->b_state & (1 << BH_Unwritten))))
1563 goto submit_io;
1564
1565 handle = ext4_journal_current_handle();
1566 BUG_ON(!handle);
1567
1568 /*
1569 * Call ext4_map_blocks() to allocate any delayed allocation
1570 * blocks, or to convert an uninitialized extent to be
1571 * initialized (in the case where we have written into
1572 * one or more preallocated blocks).
1573 *
1574 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1575 * indicate that we are on the delayed allocation path. This
1576 * affects functions in many different parts of the allocation
1577 * call path. This flag exists primarily because we don't
1578 * want to change *many* call functions, so ext4_map_blocks()
1579 * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1580 * inode's allocation semaphore is taken.
1581 *
1582 * If the blocks in questions were delalloc blocks, set
1583 * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1584 * variables are updated after the blocks have been allocated.
1585 */
1586 map.m_lblk = next;
1587 map.m_len = max_blocks;
1588 get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1589 if (ext4_should_dioread_nolock(mpd->inode))
1590 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1591 if (mpd->b_state & (1 << BH_Delay))
1592 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1593
1594 blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1595 if (blks < 0) {
1596 struct super_block *sb = mpd->inode->i_sb;
1597
1598 err = blks;
1599 /*
1600 * If get block returns EAGAIN or ENOSPC and there
1601 * appears to be free blocks we will just let
1602 * mpage_da_submit_io() unlock all of the pages.
1603 */
1604 if (err == -EAGAIN)
1605 goto submit_io;
1606
1607 if (err == -ENOSPC && ext4_count_free_clusters(sb)) {
1608 mpd->retval = err;
1609 goto submit_io;
1610 }
1611
1612 /*
1613 * get block failure will cause us to loop in
1614 * writepages, because a_ops->writepage won't be able
1615 * to make progress. The page will be redirtied by
1616 * writepage and writepages will again try to write
1617 * the same.
1618 */
1619 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1620 ext4_msg(sb, KERN_CRIT,
1621 "delayed block allocation failed for inode %lu "
1622 "at logical offset %llu with max blocks %zd "
1623 "with error %d", mpd->inode->i_ino,
1624 (unsigned long long) next,
1625 mpd->b_size >> mpd->inode->i_blkbits, err);
1626 ext4_msg(sb, KERN_CRIT,
1627 "This should not happen!! Data will be lost\n");
1628 if (err == -ENOSPC)
1629 ext4_print_free_blocks(mpd->inode);
1630 }
1631 /* invalidate all the pages */
1632 ext4_da_block_invalidatepages(mpd);
1633
1634 /* Mark this page range as having been completed */
1635 mpd->io_done = 1;
1636 return;
1637 }
1638 BUG_ON(blks == 0);
1639
1640 mapp = ↦
1641 if (map.m_flags & EXT4_MAP_NEW) {
1642 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1643 int i;
1644
1645 for (i = 0; i < map.m_len; i++)
1646 unmap_underlying_metadata(bdev, map.m_pblk + i);
1647
1648 if (ext4_should_order_data(mpd->inode)) {
1649 err = ext4_jbd2_file_inode(handle, mpd->inode);
1650 if (err) {
1651 /* Only if the journal is aborted */
1652 mpd->retval = err;
1653 goto submit_io;
1654 }
1655 }
1656 }
1657
1658 /*
1659 * Update on-disk size along with block allocation.
1660 */
1661 disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1662 if (disksize > i_size_read(mpd->inode))
1663 disksize = i_size_read(mpd->inode);
1664 if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1665 ext4_update_i_disksize(mpd->inode, disksize);
1666 err = ext4_mark_inode_dirty(handle, mpd->inode);
1667 if (err)
1668 ext4_error(mpd->inode->i_sb,
1669 "Failed to mark inode %lu dirty",
1670 mpd->inode->i_ino);
1671 }
1672
1673submit_io:
1674 mpage_da_submit_io(mpd, mapp);
1675 mpd->io_done = 1;
1676}
1677
1678#define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1679 (1 << BH_Delay) | (1 << BH_Unwritten))
1680
1681/*
1682 * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1683 *
1684 * @mpd->lbh - extent of blocks
1685 * @logical - logical number of the block in the file
1686 * @bh - bh of the block (used to access block's state)
1687 *
1688 * the function is used to collect contig. blocks in same state
1689 */
1690static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1691 sector_t logical, size_t b_size,
1692 unsigned long b_state)
1693{
1694 sector_t next;
1695 int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1696
1697 /*
1698 * XXX Don't go larger than mballoc is willing to allocate
1699 * This is a stopgap solution. We eventually need to fold
1700 * mpage_da_submit_io() into this function and then call
1701 * ext4_map_blocks() multiple times in a loop
1702 */
1703 if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1704 goto flush_it;
1705
1706 /* check if thereserved journal credits might overflow */
1707 if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1708 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1709 /*
1710 * With non-extent format we are limited by the journal
1711 * credit available. Total credit needed to insert
1712 * nrblocks contiguous blocks is dependent on the
1713 * nrblocks. So limit nrblocks.
1714 */
1715 goto flush_it;
1716 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1717 EXT4_MAX_TRANS_DATA) {
1718 /*
1719 * Adding the new buffer_head would make it cross the
1720 * allowed limit for which we have journal credit
1721 * reserved. So limit the new bh->b_size
1722 */
1723 b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1724 mpd->inode->i_blkbits;
1725 /* we will do mpage_da_submit_io in the next loop */
1726 }
1727 }
1728 /*
1729 * First block in the extent
1730 */
1731 if (mpd->b_size == 0) {
1732 mpd->b_blocknr = logical;
1733 mpd->b_size = b_size;
1734 mpd->b_state = b_state & BH_FLAGS;
1735 return;
1736 }
1737
1738 next = mpd->b_blocknr + nrblocks;
1739 /*
1740 * Can we merge the block to our big extent?
1741 */
1742 if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1743 mpd->b_size += b_size;
1744 return;
1745 }
1746
1747flush_it:
1748 /*
1749 * We couldn't merge the block to our extent, so we
1750 * need to flush current extent and start new one
1751 */
1752 mpage_da_map_and_submit(mpd);
1753 return;
1754}
1755
1756static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1757{
1758 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1759}
1760
1761/*
1762 * This function is grabs code from the very beginning of
1763 * ext4_map_blocks, but assumes that the caller is from delayed write
1764 * time. This function looks up the requested blocks and sets the
1765 * buffer delay bit under the protection of i_data_sem.
1766 */
1767static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1768 struct ext4_map_blocks *map,
1769 struct buffer_head *bh)
1770{
1771 int retval;
1772 sector_t invalid_block = ~((sector_t) 0xffff);
1773
1774 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1775 invalid_block = ~0;
1776
1777 map->m_flags = 0;
1778 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1779 "logical block %lu\n", inode->i_ino, map->m_len,
1780 (unsigned long) map->m_lblk);
1781 /*
1782 * Try to see if we can get the block without requesting a new
1783 * file system block.
1784 */
1785 down_read((&EXT4_I(inode)->i_data_sem));
1786 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1787 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1788 else
1789 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1790
1791 if (retval == 0) {
1792 /*
1793 * XXX: __block_prepare_write() unmaps passed block,
1794 * is it OK?
1795 */
1796 /* If the block was allocated from previously allocated cluster,
1797 * then we dont need to reserve it again. */
1798 if (!(map->m_flags & EXT4_MAP_FROM_CLUSTER)) {
1799 retval = ext4_da_reserve_space(inode, iblock);
1800 if (retval)
1801 /* not enough space to reserve */
1802 goto out_unlock;
1803 }
1804
1805 /* Clear EXT4_MAP_FROM_CLUSTER flag since its purpose is served
1806 * and it should not appear on the bh->b_state.
1807 */
1808 map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
1809
1810 map_bh(bh, inode->i_sb, invalid_block);
1811 set_buffer_new(bh);
1812 set_buffer_delay(bh);
1813 }
1814
1815out_unlock:
1816 up_read((&EXT4_I(inode)->i_data_sem));
1817
1818 return retval;
1819}
1820
1821/*
1822 * This is a special get_blocks_t callback which is used by
1823 * ext4_da_write_begin(). It will either return mapped block or
1824 * reserve space for a single block.
1825 *
1826 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1827 * We also have b_blocknr = -1 and b_bdev initialized properly
1828 *
1829 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1830 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1831 * initialized properly.
1832 */
1833static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1834 struct buffer_head *bh, int create)
1835{
1836 struct ext4_map_blocks map;
1837 int ret = 0;
1838
1839 BUG_ON(create == 0);
1840 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1841
1842 map.m_lblk = iblock;
1843 map.m_len = 1;
1844
1845 /*
1846 * first, we need to know whether the block is allocated already
1847 * preallocated blocks are unmapped but should treated
1848 * the same as allocated blocks.
1849 */
1850 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1851 if (ret <= 0)
1852 return ret;
1853
1854 map_bh(bh, inode->i_sb, map.m_pblk);
1855 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1856
1857 if (buffer_unwritten(bh)) {
1858 /* A delayed write to unwritten bh should be marked
1859 * new and mapped. Mapped ensures that we don't do
1860 * get_block multiple times when we write to the same
1861 * offset and new ensures that we do proper zero out
1862 * for partial write.
1863 */
1864 set_buffer_new(bh);
1865 set_buffer_mapped(bh);
1866 }
1867 return 0;
1868}
1869
1870/*
1871 * This function is used as a standard get_block_t calback function
1872 * when there is no desire to allocate any blocks. It is used as a
1873 * callback function for block_write_begin() and block_write_full_page().
1874 * These functions should only try to map a single block at a time.
1875 *
1876 * Since this function doesn't do block allocations even if the caller
1877 * requests it by passing in create=1, it is critically important that
1878 * any caller checks to make sure that any buffer heads are returned
1879 * by this function are either all already mapped or marked for
1880 * delayed allocation before calling block_write_full_page(). Otherwise,
1881 * b_blocknr could be left unitialized, and the page write functions will
1882 * be taken by surprise.
1883 */
1884static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1885 struct buffer_head *bh_result, int create)
1886{
1887 BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1888 return _ext4_get_block(inode, iblock, bh_result, 0);
1889}
1890
1891static int bget_one(handle_t *handle, struct buffer_head *bh)
1892{
1893 get_bh(bh);
1894 return 0;
1895}
1896
1897static int bput_one(handle_t *handle, struct buffer_head *bh)
1898{
1899 put_bh(bh);
1900 return 0;
1901}
1902
1903static int __ext4_journalled_writepage(struct page *page,
1904 unsigned int len)
1905{
1906 struct address_space *mapping = page->mapping;
1907 struct inode *inode = mapping->host;
1908 struct buffer_head *page_bufs;
1909 handle_t *handle = NULL;
1910 int ret = 0;
1911 int err;
1912
1913 ClearPageChecked(page);
1914 page_bufs = page_buffers(page);
1915 BUG_ON(!page_bufs);
1916 walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1917 /* As soon as we unlock the page, it can go away, but we have
1918 * references to buffers so we are safe */
1919 unlock_page(page);
1920
1921 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1922 if (IS_ERR(handle)) {
1923 ret = PTR_ERR(handle);
1924 goto out;
1925 }
1926
1927 BUG_ON(!ext4_handle_valid(handle));
1928
1929 ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1930 do_journal_get_write_access);
1931
1932 err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1933 write_end_fn);
1934 if (ret == 0)
1935 ret = err;
1936 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1937 err = ext4_journal_stop(handle);
1938 if (!ret)
1939 ret = err;
1940
1941 walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1942 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1943out:
1944 return ret;
1945}
1946
1947static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1948static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1949
1950/*
1951 * Note that we don't need to start a transaction unless we're journaling data
1952 * because we should have holes filled from ext4_page_mkwrite(). We even don't
1953 * need to file the inode to the transaction's list in ordered mode because if
1954 * we are writing back data added by write(), the inode is already there and if
1955 * we are writing back data modified via mmap(), no one guarantees in which
1956 * transaction the data will hit the disk. In case we are journaling data, we
1957 * cannot start transaction directly because transaction start ranks above page
1958 * lock so we have to do some magic.
1959 *
1960 * This function can get called via...
1961 * - ext4_da_writepages after taking page lock (have journal handle)
1962 * - journal_submit_inode_data_buffers (no journal handle)
1963 * - shrink_page_list via pdflush (no journal handle)
1964 * - grab_page_cache when doing write_begin (have journal handle)
1965 *
1966 * We don't do any block allocation in this function. If we have page with
1967 * multiple blocks we need to write those buffer_heads that are mapped. This
1968 * is important for mmaped based write. So if we do with blocksize 1K
1969 * truncate(f, 1024);
1970 * a = mmap(f, 0, 4096);
1971 * a[0] = 'a';
1972 * truncate(f, 4096);
1973 * we have in the page first buffer_head mapped via page_mkwrite call back
1974 * but other buffer_heads would be unmapped but dirty (dirty done via the
1975 * do_wp_page). So writepage should write the first block. If we modify
1976 * the mmap area beyond 1024 we will again get a page_fault and the
1977 * page_mkwrite callback will do the block allocation and mark the
1978 * buffer_heads mapped.
1979 *
1980 * We redirty the page if we have any buffer_heads that is either delay or
1981 * unwritten in the page.
1982 *
1983 * We can get recursively called as show below.
1984 *
1985 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1986 * ext4_writepage()
1987 *
1988 * But since we don't do any block allocation we should not deadlock.
1989 * Page also have the dirty flag cleared so we don't get recurive page_lock.
1990 */
1991static int ext4_writepage(struct page *page,
1992 struct writeback_control *wbc)
1993{
1994 int ret = 0, commit_write = 0;
1995 loff_t size;
1996 unsigned int len;
1997 struct buffer_head *page_bufs = NULL;
1998 struct inode *inode = page->mapping->host;
1999
2000 trace_ext4_writepage(page);
2001 size = i_size_read(inode);
2002 if (page->index == size >> PAGE_CACHE_SHIFT)
2003 len = size & ~PAGE_CACHE_MASK;
2004 else
2005 len = PAGE_CACHE_SIZE;
2006
2007 /*
2008 * If the page does not have buffers (for whatever reason),
2009 * try to create them using __block_write_begin. If this
2010 * fails, redirty the page and move on.
2011 */
2012 if (!page_has_buffers(page)) {
2013 if (__block_write_begin(page, 0, len,
2014 noalloc_get_block_write)) {
2015 redirty_page:
2016 redirty_page_for_writepage(wbc, page);
2017 unlock_page(page);
2018 return 0;
2019 }
2020 commit_write = 1;
2021 }
2022 page_bufs = page_buffers(page);
2023 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2024 ext4_bh_delay_or_unwritten)) {
2025 /*
2026 * We don't want to do block allocation, so redirty
2027 * the page and return. We may reach here when we do
2028 * a journal commit via journal_submit_inode_data_buffers.
2029 * We can also reach here via shrink_page_list but it
2030 * should never be for direct reclaim so warn if that
2031 * happens
2032 */
2033 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
2034 PF_MEMALLOC);
2035 goto redirty_page;
2036 }
2037 if (commit_write)
2038 /* now mark the buffer_heads as dirty and uptodate */
2039 block_commit_write(page, 0, len);
2040
2041 if (PageChecked(page) && ext4_should_journal_data(inode))
2042 /*
2043 * It's mmapped pagecache. Add buffers and journal it. There
2044 * doesn't seem much point in redirtying the page here.
2045 */
2046 return __ext4_journalled_writepage(page, len);
2047
2048 if (buffer_uninit(page_bufs)) {
2049 ext4_set_bh_endio(page_bufs, inode);
2050 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2051 wbc, ext4_end_io_buffer_write);
2052 } else
2053 ret = block_write_full_page(page, noalloc_get_block_write,
2054 wbc);
2055
2056 return ret;
2057}
2058
2059/*
2060 * This is called via ext4_da_writepages() to
2061 * calculate the total number of credits to reserve to fit
2062 * a single extent allocation into a single transaction,
2063 * ext4_da_writpeages() will loop calling this before
2064 * the block allocation.
2065 */
2066
2067static int ext4_da_writepages_trans_blocks(struct inode *inode)
2068{
2069 int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2070
2071 /*
2072 * With non-extent format the journal credit needed to
2073 * insert nrblocks contiguous block is dependent on
2074 * number of contiguous block. So we will limit
2075 * number of contiguous block to a sane value
2076 */
2077 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2078 (max_blocks > EXT4_MAX_TRANS_DATA))
2079 max_blocks = EXT4_MAX_TRANS_DATA;
2080
2081 return ext4_chunk_trans_blocks(inode, max_blocks);
2082}
2083
2084/*
2085 * write_cache_pages_da - walk the list of dirty pages of the given
2086 * address space and accumulate pages that need writing, and call
2087 * mpage_da_map_and_submit to map a single contiguous memory region
2088 * and then write them.
2089 */
2090static int write_cache_pages_da(struct address_space *mapping,
2091 struct writeback_control *wbc,
2092 struct mpage_da_data *mpd,
2093 pgoff_t *done_index)
2094{
2095 struct buffer_head *bh, *head;
2096 struct inode *inode = mapping->host;
2097 struct pagevec pvec;
2098 unsigned int nr_pages;
2099 sector_t logical;
2100 pgoff_t index, end;
2101 long nr_to_write = wbc->nr_to_write;
2102 int i, tag, ret = 0;
2103
2104 memset(mpd, 0, sizeof(struct mpage_da_data));
2105 mpd->wbc = wbc;
2106 mpd->inode = inode;
2107 pagevec_init(&pvec, 0);
2108 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2109 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2110
2111 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2112 tag = PAGECACHE_TAG_TOWRITE;
2113 else
2114 tag = PAGECACHE_TAG_DIRTY;
2115
2116 *done_index = index;
2117 while (index <= end) {
2118 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2119 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2120 if (nr_pages == 0)
2121 return 0;
2122
2123 for (i = 0; i < nr_pages; i++) {
2124 struct page *page = pvec.pages[i];
2125
2126 /*
2127 * At this point, the page may be truncated or
2128 * invalidated (changing page->mapping to NULL), or
2129 * even swizzled back from swapper_space to tmpfs file
2130 * mapping. However, page->index will not change
2131 * because we have a reference on the page.
2132 */
2133 if (page->index > end)
2134 goto out;
2135
2136 *done_index = page->index + 1;
2137
2138 /*
2139 * If we can't merge this page, and we have
2140 * accumulated an contiguous region, write it
2141 */
2142 if ((mpd->next_page != page->index) &&
2143 (mpd->next_page != mpd->first_page)) {
2144 mpage_da_map_and_submit(mpd);
2145 goto ret_extent_tail;
2146 }
2147
2148 lock_page(page);
2149
2150 /*
2151 * If the page is no longer dirty, or its
2152 * mapping no longer corresponds to inode we
2153 * are writing (which means it has been
2154 * truncated or invalidated), or the page is
2155 * already under writeback and we are not
2156 * doing a data integrity writeback, skip the page
2157 */
2158 if (!PageDirty(page) ||
2159 (PageWriteback(page) &&
2160 (wbc->sync_mode == WB_SYNC_NONE)) ||
2161 unlikely(page->mapping != mapping)) {
2162 unlock_page(page);
2163 continue;
2164 }
2165
2166 wait_on_page_writeback(page);
2167 BUG_ON(PageWriteback(page));
2168
2169 if (mpd->next_page != page->index)
2170 mpd->first_page = page->index;
2171 mpd->next_page = page->index + 1;
2172 logical = (sector_t) page->index <<
2173 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2174
2175 if (!page_has_buffers(page)) {
2176 mpage_add_bh_to_extent(mpd, logical,
2177 PAGE_CACHE_SIZE,
2178 (1 << BH_Dirty) | (1 << BH_Uptodate));
2179 if (mpd->io_done)
2180 goto ret_extent_tail;
2181 } else {
2182 /*
2183 * Page with regular buffer heads,
2184 * just add all dirty ones
2185 */
2186 head = page_buffers(page);
2187 bh = head;
2188 do {
2189 BUG_ON(buffer_locked(bh));
2190 /*
2191 * We need to try to allocate
2192 * unmapped blocks in the same page.
2193 * Otherwise we won't make progress
2194 * with the page in ext4_writepage
2195 */
2196 if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2197 mpage_add_bh_to_extent(mpd, logical,
2198 bh->b_size,
2199 bh->b_state);
2200 if (mpd->io_done)
2201 goto ret_extent_tail;
2202 } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2203 /*
2204 * mapped dirty buffer. We need
2205 * to update the b_state
2206 * because we look at b_state
2207 * in mpage_da_map_blocks. We
2208 * don't update b_size because
2209 * if we find an unmapped
2210 * buffer_head later we need to
2211 * use the b_state flag of that
2212 * buffer_head.
2213 */
2214 if (mpd->b_size == 0)
2215 mpd->b_state = bh->b_state & BH_FLAGS;
2216 }
2217 logical++;
2218 } while ((bh = bh->b_this_page) != head);
2219 }
2220
2221 if (nr_to_write > 0) {
2222 nr_to_write--;
2223 if (nr_to_write == 0 &&
2224 wbc->sync_mode == WB_SYNC_NONE)
2225 /*
2226 * We stop writing back only if we are
2227 * not doing integrity sync. In case of
2228 * integrity sync we have to keep going
2229 * because someone may be concurrently
2230 * dirtying pages, and we might have
2231 * synced a lot of newly appeared dirty
2232 * pages, but have not synced all of the
2233 * old dirty pages.
2234 */
2235 goto out;
2236 }
2237 }
2238 pagevec_release(&pvec);
2239 cond_resched();
2240 }
2241 return 0;
2242ret_extent_tail:
2243 ret = MPAGE_DA_EXTENT_TAIL;
2244out:
2245 pagevec_release(&pvec);
2246 cond_resched();
2247 return ret;
2248}
2249
2250
2251static int ext4_da_writepages(struct address_space *mapping,
2252 struct writeback_control *wbc)
2253{
2254 pgoff_t index;
2255 int range_whole = 0;
2256 handle_t *handle = NULL;
2257 struct mpage_da_data mpd;
2258 struct inode *inode = mapping->host;
2259 int pages_written = 0;
2260 unsigned int max_pages;
2261 int range_cyclic, cycled = 1, io_done = 0;
2262 int needed_blocks, ret = 0;
2263 long desired_nr_to_write, nr_to_writebump = 0;
2264 loff_t range_start = wbc->range_start;
2265 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2266 pgoff_t done_index = 0;
2267 pgoff_t end;
2268 struct blk_plug plug;
2269
2270 trace_ext4_da_writepages(inode, wbc);
2271
2272 /*
2273 * No pages to write? This is mainly a kludge to avoid starting
2274 * a transaction for special inodes like journal inode on last iput()
2275 * because that could violate lock ordering on umount
2276 */
2277 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2278 return 0;
2279
2280 /*
2281 * If the filesystem has aborted, it is read-only, so return
2282 * right away instead of dumping stack traces later on that
2283 * will obscure the real source of the problem. We test
2284 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2285 * the latter could be true if the filesystem is mounted
2286 * read-only, and in that case, ext4_da_writepages should
2287 * *never* be called, so if that ever happens, we would want
2288 * the stack trace.
2289 */
2290 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2291 return -EROFS;
2292
2293 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2294 range_whole = 1;
2295
2296 range_cyclic = wbc->range_cyclic;
2297 if (wbc->range_cyclic) {
2298 index = mapping->writeback_index;
2299 if (index)
2300 cycled = 0;
2301 wbc->range_start = index << PAGE_CACHE_SHIFT;
2302 wbc->range_end = LLONG_MAX;
2303 wbc->range_cyclic = 0;
2304 end = -1;
2305 } else {
2306 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2307 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2308 }
2309
2310 /*
2311 * This works around two forms of stupidity. The first is in
2312 * the writeback code, which caps the maximum number of pages
2313 * written to be 1024 pages. This is wrong on multiple
2314 * levels; different architectues have a different page size,
2315 * which changes the maximum amount of data which gets
2316 * written. Secondly, 4 megabytes is way too small. XFS
2317 * forces this value to be 16 megabytes by multiplying
2318 * nr_to_write parameter by four, and then relies on its
2319 * allocator to allocate larger extents to make them
2320 * contiguous. Unfortunately this brings us to the second
2321 * stupidity, which is that ext4's mballoc code only allocates
2322 * at most 2048 blocks. So we force contiguous writes up to
2323 * the number of dirty blocks in the inode, or
2324 * sbi->max_writeback_mb_bump whichever is smaller.
2325 */
2326 max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2327 if (!range_cyclic && range_whole) {
2328 if (wbc->nr_to_write == LONG_MAX)
2329 desired_nr_to_write = wbc->nr_to_write;
2330 else
2331 desired_nr_to_write = wbc->nr_to_write * 8;
2332 } else
2333 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2334 max_pages);
2335 if (desired_nr_to_write > max_pages)
2336 desired_nr_to_write = max_pages;
2337
2338 if (wbc->nr_to_write < desired_nr_to_write) {
2339 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2340 wbc->nr_to_write = desired_nr_to_write;
2341 }
2342
2343retry:
2344 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2345 tag_pages_for_writeback(mapping, index, end);
2346
2347 blk_start_plug(&plug);
2348 while (!ret && wbc->nr_to_write > 0) {
2349
2350 /*
2351 * we insert one extent at a time. So we need
2352 * credit needed for single extent allocation.
2353 * journalled mode is currently not supported
2354 * by delalloc
2355 */
2356 BUG_ON(ext4_should_journal_data(inode));
2357 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2358
2359 /* start a new transaction*/
2360 handle = ext4_journal_start(inode, needed_blocks);
2361 if (IS_ERR(handle)) {
2362 ret = PTR_ERR(handle);
2363 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2364 "%ld pages, ino %lu; err %d", __func__,
2365 wbc->nr_to_write, inode->i_ino, ret);
2366 blk_finish_plug(&plug);
2367 goto out_writepages;
2368 }
2369
2370 /*
2371 * Now call write_cache_pages_da() to find the next
2372 * contiguous region of logical blocks that need
2373 * blocks to be allocated by ext4 and submit them.
2374 */
2375 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2376 /*
2377 * If we have a contiguous extent of pages and we
2378 * haven't done the I/O yet, map the blocks and submit
2379 * them for I/O.
2380 */
2381 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2382 mpage_da_map_and_submit(&mpd);
2383 ret = MPAGE_DA_EXTENT_TAIL;
2384 }
2385 trace_ext4_da_write_pages(inode, &mpd);
2386 wbc->nr_to_write -= mpd.pages_written;
2387
2388 ext4_journal_stop(handle);
2389
2390 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2391 /* commit the transaction which would
2392 * free blocks released in the transaction
2393 * and try again
2394 */
2395 jbd2_journal_force_commit_nested(sbi->s_journal);
2396 ret = 0;
2397 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2398 /*
2399 * Got one extent now try with rest of the pages.
2400 * If mpd.retval is set -EIO, journal is aborted.
2401 * So we don't need to write any more.
2402 */
2403 pages_written += mpd.pages_written;
2404 ret = mpd.retval;
2405 io_done = 1;
2406 } else if (wbc->nr_to_write)
2407 /*
2408 * There is no more writeout needed
2409 * or we requested for a noblocking writeout
2410 * and we found the device congested
2411 */
2412 break;
2413 }
2414 blk_finish_plug(&plug);
2415 if (!io_done && !cycled) {
2416 cycled = 1;
2417 index = 0;
2418 wbc->range_start = index << PAGE_CACHE_SHIFT;
2419 wbc->range_end = mapping->writeback_index - 1;
2420 goto retry;
2421 }
2422
2423 /* Update index */
2424 wbc->range_cyclic = range_cyclic;
2425 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2426 /*
2427 * set the writeback_index so that range_cyclic
2428 * mode will write it back later
2429 */
2430 mapping->writeback_index = done_index;
2431
2432out_writepages:
2433 wbc->nr_to_write -= nr_to_writebump;
2434 wbc->range_start = range_start;
2435 trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2436 return ret;
2437}
2438
2439#define FALL_BACK_TO_NONDELALLOC 1
2440static int ext4_nonda_switch(struct super_block *sb)
2441{
2442 s64 free_blocks, dirty_blocks;
2443 struct ext4_sb_info *sbi = EXT4_SB(sb);
2444
2445 /*
2446 * switch to non delalloc mode if we are running low
2447 * on free block. The free block accounting via percpu
2448 * counters can get slightly wrong with percpu_counter_batch getting
2449 * accumulated on each CPU without updating global counters
2450 * Delalloc need an accurate free block accounting. So switch
2451 * to non delalloc when we are near to error range.
2452 */
2453 free_blocks = EXT4_C2B(sbi,
2454 percpu_counter_read_positive(&sbi->s_freeclusters_counter));
2455 dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2456 if (2 * free_blocks < 3 * dirty_blocks ||
2457 free_blocks < (dirty_blocks + EXT4_FREECLUSTERS_WATERMARK)) {
2458 /*
2459 * free block count is less than 150% of dirty blocks
2460 * or free blocks is less than watermark
2461 */
2462 return 1;
2463 }
2464 /*
2465 * Even if we don't switch but are nearing capacity,
2466 * start pushing delalloc when 1/2 of free blocks are dirty.
2467 */
2468 if (free_blocks < 2 * dirty_blocks)
2469 writeback_inodes_sb_if_idle(sb, WB_REASON_FS_FREE_SPACE);
2470
2471 return 0;
2472}
2473
2474static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2475 loff_t pos, unsigned len, unsigned flags,
2476 struct page **pagep, void **fsdata)
2477{
2478 int ret, retries = 0;
2479 struct page *page;
2480 pgoff_t index;
2481 struct inode *inode = mapping->host;
2482 handle_t *handle;
2483
2484 index = pos >> PAGE_CACHE_SHIFT;
2485
2486 if (ext4_nonda_switch(inode->i_sb)) {
2487 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2488 return ext4_write_begin(file, mapping, pos,
2489 len, flags, pagep, fsdata);
2490 }
2491 *fsdata = (void *)0;
2492 trace_ext4_da_write_begin(inode, pos, len, flags);
2493retry:
2494 /*
2495 * With delayed allocation, we don't log the i_disksize update
2496 * if there is delayed block allocation. But we still need
2497 * to journalling the i_disksize update if writes to the end
2498 * of file which has an already mapped buffer.
2499 */
2500 handle = ext4_journal_start(inode, 1);
2501 if (IS_ERR(handle)) {
2502 ret = PTR_ERR(handle);
2503 goto out;
2504 }
2505 /* We cannot recurse into the filesystem as the transaction is already
2506 * started */
2507 flags |= AOP_FLAG_NOFS;
2508
2509 page = grab_cache_page_write_begin(mapping, index, flags);
2510 if (!page) {
2511 ext4_journal_stop(handle);
2512 ret = -ENOMEM;
2513 goto out;
2514 }
2515 *pagep = page;
2516
2517 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2518 if (ret < 0) {
2519 unlock_page(page);
2520 ext4_journal_stop(handle);
2521 page_cache_release(page);
2522 /*
2523 * block_write_begin may have instantiated a few blocks
2524 * outside i_size. Trim these off again. Don't need
2525 * i_size_read because we hold i_mutex.
2526 */
2527 if (pos + len > inode->i_size)
2528 ext4_truncate_failed_write(inode);
2529 }
2530
2531 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2532 goto retry;
2533out:
2534 return ret;
2535}
2536
2537/*
2538 * Check if we should update i_disksize
2539 * when write to the end of file but not require block allocation
2540 */
2541static int ext4_da_should_update_i_disksize(struct page *page,
2542 unsigned long offset)
2543{
2544 struct buffer_head *bh;
2545 struct inode *inode = page->mapping->host;
2546 unsigned int idx;
2547 int i;
2548
2549 bh = page_buffers(page);
2550 idx = offset >> inode->i_blkbits;
2551
2552 for (i = 0; i < idx; i++)
2553 bh = bh->b_this_page;
2554
2555 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2556 return 0;
2557 return 1;
2558}
2559
2560static int ext4_da_write_end(struct file *file,
2561 struct address_space *mapping,
2562 loff_t pos, unsigned len, unsigned copied,
2563 struct page *page, void *fsdata)
2564{
2565 struct inode *inode = mapping->host;
2566 int ret = 0, ret2;
2567 handle_t *handle = ext4_journal_current_handle();
2568 loff_t new_i_size;
2569 unsigned long start, end;
2570 int write_mode = (int)(unsigned long)fsdata;
2571
2572 if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2573 switch (ext4_inode_journal_mode(inode)) {
2574 case EXT4_INODE_ORDERED_DATA_MODE:
2575 return ext4_ordered_write_end(file, mapping, pos,
2576 len, copied, page, fsdata);
2577 case EXT4_INODE_WRITEBACK_DATA_MODE:
2578 return ext4_writeback_write_end(file, mapping, pos,
2579 len, copied, page, fsdata);
2580 default:
2581 BUG();
2582 }
2583 }
2584
2585 trace_ext4_da_write_end(inode, pos, len, copied);
2586 start = pos & (PAGE_CACHE_SIZE - 1);
2587 end = start + copied - 1;
2588
2589 /*
2590 * generic_write_end() will run mark_inode_dirty() if i_size
2591 * changes. So let's piggyback the i_disksize mark_inode_dirty
2592 * into that.
2593 */
2594
2595 new_i_size = pos + copied;
2596 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2597 if (ext4_da_should_update_i_disksize(page, end)) {
2598 down_write(&EXT4_I(inode)->i_data_sem);
2599 if (new_i_size > EXT4_I(inode)->i_disksize) {
2600 /*
2601 * Updating i_disksize when extending file
2602 * without needing block allocation
2603 */
2604 if (ext4_should_order_data(inode))
2605 ret = ext4_jbd2_file_inode(handle,
2606 inode);
2607
2608 EXT4_I(inode)->i_disksize = new_i_size;
2609 }
2610 up_write(&EXT4_I(inode)->i_data_sem);
2611 /* We need to mark inode dirty even if
2612 * new_i_size is less that inode->i_size
2613 * bu greater than i_disksize.(hint delalloc)
2614 */
2615 ext4_mark_inode_dirty(handle, inode);
2616 }
2617 }
2618 ret2 = generic_write_end(file, mapping, pos, len, copied,
2619 page, fsdata);
2620 copied = ret2;
2621 if (ret2 < 0)
2622 ret = ret2;
2623 ret2 = ext4_journal_stop(handle);
2624 if (!ret)
2625 ret = ret2;
2626
2627 return ret ? ret : copied;
2628}
2629
2630static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2631{
2632 /*
2633 * Drop reserved blocks
2634 */
2635 BUG_ON(!PageLocked(page));
2636 if (!page_has_buffers(page))
2637 goto out;
2638
2639 ext4_da_page_release_reservation(page, offset);
2640
2641out:
2642 ext4_invalidatepage(page, offset);
2643
2644 return;
2645}
2646
2647/*
2648 * Force all delayed allocation blocks to be allocated for a given inode.
2649 */
2650int ext4_alloc_da_blocks(struct inode *inode)
2651{
2652 trace_ext4_alloc_da_blocks(inode);
2653
2654 if (!EXT4_I(inode)->i_reserved_data_blocks &&
2655 !EXT4_I(inode)->i_reserved_meta_blocks)
2656 return 0;
2657
2658 /*
2659 * We do something simple for now. The filemap_flush() will
2660 * also start triggering a write of the data blocks, which is
2661 * not strictly speaking necessary (and for users of
2662 * laptop_mode, not even desirable). However, to do otherwise
2663 * would require replicating code paths in:
2664 *
2665 * ext4_da_writepages() ->
2666 * write_cache_pages() ---> (via passed in callback function)
2667 * __mpage_da_writepage() -->
2668 * mpage_add_bh_to_extent()
2669 * mpage_da_map_blocks()
2670 *
2671 * The problem is that write_cache_pages(), located in
2672 * mm/page-writeback.c, marks pages clean in preparation for
2673 * doing I/O, which is not desirable if we're not planning on
2674 * doing I/O at all.
2675 *
2676 * We could call write_cache_pages(), and then redirty all of
2677 * the pages by calling redirty_page_for_writepage() but that
2678 * would be ugly in the extreme. So instead we would need to
2679 * replicate parts of the code in the above functions,
2680 * simplifying them because we wouldn't actually intend to
2681 * write out the pages, but rather only collect contiguous
2682 * logical block extents, call the multi-block allocator, and
2683 * then update the buffer heads with the block allocations.
2684 *
2685 * For now, though, we'll cheat by calling filemap_flush(),
2686 * which will map the blocks, and start the I/O, but not
2687 * actually wait for the I/O to complete.
2688 */
2689 return filemap_flush(inode->i_mapping);
2690}
2691
2692/*
2693 * bmap() is special. It gets used by applications such as lilo and by
2694 * the swapper to find the on-disk block of a specific piece of data.
2695 *
2696 * Naturally, this is dangerous if the block concerned is still in the
2697 * journal. If somebody makes a swapfile on an ext4 data-journaling
2698 * filesystem and enables swap, then they may get a nasty shock when the
2699 * data getting swapped to that swapfile suddenly gets overwritten by
2700 * the original zero's written out previously to the journal and
2701 * awaiting writeback in the kernel's buffer cache.
2702 *
2703 * So, if we see any bmap calls here on a modified, data-journaled file,
2704 * take extra steps to flush any blocks which might be in the cache.
2705 */
2706static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2707{
2708 struct inode *inode = mapping->host;
2709 journal_t *journal;
2710 int err;
2711
2712 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2713 test_opt(inode->i_sb, DELALLOC)) {
2714 /*
2715 * With delalloc we want to sync the file
2716 * so that we can make sure we allocate
2717 * blocks for file
2718 */
2719 filemap_write_and_wait(mapping);
2720 }
2721
2722 if (EXT4_JOURNAL(inode) &&
2723 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2724 /*
2725 * This is a REALLY heavyweight approach, but the use of
2726 * bmap on dirty files is expected to be extremely rare:
2727 * only if we run lilo or swapon on a freshly made file
2728 * do we expect this to happen.
2729 *
2730 * (bmap requires CAP_SYS_RAWIO so this does not
2731 * represent an unprivileged user DOS attack --- we'd be
2732 * in trouble if mortal users could trigger this path at
2733 * will.)
2734 *
2735 * NB. EXT4_STATE_JDATA is not set on files other than
2736 * regular files. If somebody wants to bmap a directory
2737 * or symlink and gets confused because the buffer
2738 * hasn't yet been flushed to disk, they deserve
2739 * everything they get.
2740 */
2741
2742 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2743 journal = EXT4_JOURNAL(inode);
2744 jbd2_journal_lock_updates(journal);
2745 err = jbd2_journal_flush(journal);
2746 jbd2_journal_unlock_updates(journal);
2747
2748 if (err)
2749 return 0;
2750 }
2751
2752 return generic_block_bmap(mapping, block, ext4_get_block);
2753}
2754
2755static int ext4_readpage(struct file *file, struct page *page)
2756{
2757 trace_ext4_readpage(page);
2758 return mpage_readpage(page, ext4_get_block);
2759}
2760
2761static int
2762ext4_readpages(struct file *file, struct address_space *mapping,
2763 struct list_head *pages, unsigned nr_pages)
2764{
2765 return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2766}
2767
2768static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2769{
2770 struct buffer_head *head, *bh;
2771 unsigned int curr_off = 0;
2772
2773 if (!page_has_buffers(page))
2774 return;
2775 head = bh = page_buffers(page);
2776 do {
2777 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2778 && bh->b_private) {
2779 ext4_free_io_end(bh->b_private);
2780 bh->b_private = NULL;
2781 bh->b_end_io = NULL;
2782 }
2783 curr_off = curr_off + bh->b_size;
2784 bh = bh->b_this_page;
2785 } while (bh != head);
2786}
2787
2788static void ext4_invalidatepage(struct page *page, unsigned long offset)
2789{
2790 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2791
2792 trace_ext4_invalidatepage(page, offset);
2793
2794 /*
2795 * free any io_end structure allocated for buffers to be discarded
2796 */
2797 if (ext4_should_dioread_nolock(page->mapping->host))
2798 ext4_invalidatepage_free_endio(page, offset);
2799 /*
2800 * If it's a full truncate we just forget about the pending dirtying
2801 */
2802 if (offset == 0)
2803 ClearPageChecked(page);
2804
2805 if (journal)
2806 jbd2_journal_invalidatepage(journal, page, offset);
2807 else
2808 block_invalidatepage(page, offset);
2809}
2810
2811static int ext4_releasepage(struct page *page, gfp_t wait)
2812{
2813 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2814
2815 trace_ext4_releasepage(page);
2816
2817 WARN_ON(PageChecked(page));
2818 if (!page_has_buffers(page))
2819 return 0;
2820 if (journal)
2821 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2822 else
2823 return try_to_free_buffers(page);
2824}
2825
2826/*
2827 * ext4_get_block used when preparing for a DIO write or buffer write.
2828 * We allocate an uinitialized extent if blocks haven't been allocated.
2829 * The extent will be converted to initialized after the IO is complete.
2830 */
2831static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2832 struct buffer_head *bh_result, int create)
2833{
2834 ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2835 inode->i_ino, create);
2836 return _ext4_get_block(inode, iblock, bh_result,
2837 EXT4_GET_BLOCKS_IO_CREATE_EXT);
2838}
2839
2840static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2841 ssize_t size, void *private, int ret,
2842 bool is_async)
2843{
2844 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2845 ext4_io_end_t *io_end = iocb->private;
2846 struct workqueue_struct *wq;
2847 unsigned long flags;
2848 struct ext4_inode_info *ei;
2849
2850 /* if not async direct IO or dio with 0 bytes write, just return */
2851 if (!io_end || !size)
2852 goto out;
2853
2854 ext_debug("ext4_end_io_dio(): io_end 0x%p "
2855 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
2856 iocb->private, io_end->inode->i_ino, iocb, offset,
2857 size);
2858
2859 iocb->private = NULL;
2860
2861 /* if not aio dio with unwritten extents, just free io and return */
2862 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2863 ext4_free_io_end(io_end);
2864out:
2865 if (is_async)
2866 aio_complete(iocb, ret, 0);
2867 inode_dio_done(inode);
2868 return;
2869 }
2870
2871 io_end->offset = offset;
2872 io_end->size = size;
2873 if (is_async) {
2874 io_end->iocb = iocb;
2875 io_end->result = ret;
2876 }
2877 wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2878
2879 /* Add the io_end to per-inode completed aio dio list*/
2880 ei = EXT4_I(io_end->inode);
2881 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2882 list_add_tail(&io_end->list, &ei->i_completed_io_list);
2883 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2884
2885 /* queue the work to convert unwritten extents to written */
2886 queue_work(wq, &io_end->work);
2887}
2888
2889static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2890{
2891 ext4_io_end_t *io_end = bh->b_private;
2892 struct workqueue_struct *wq;
2893 struct inode *inode;
2894 unsigned long flags;
2895
2896 if (!test_clear_buffer_uninit(bh) || !io_end)
2897 goto out;
2898
2899 if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2900 ext4_msg(io_end->inode->i_sb, KERN_INFO,
2901 "sb umounted, discard end_io request for inode %lu",
2902 io_end->inode->i_ino);
2903 ext4_free_io_end(io_end);
2904 goto out;
2905 }
2906
2907 /*
2908 * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2909 * but being more careful is always safe for the future change.
2910 */
2911 inode = io_end->inode;
2912 ext4_set_io_unwritten_flag(inode, io_end);
2913
2914 /* Add the io_end to per-inode completed io list*/
2915 spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2916 list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2917 spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2918
2919 wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2920 /* queue the work to convert unwritten extents to written */
2921 queue_work(wq, &io_end->work);
2922out:
2923 bh->b_private = NULL;
2924 bh->b_end_io = NULL;
2925 clear_buffer_uninit(bh);
2926 end_buffer_async_write(bh, uptodate);
2927}
2928
2929static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2930{
2931 ext4_io_end_t *io_end;
2932 struct page *page = bh->b_page;
2933 loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2934 size_t size = bh->b_size;
2935
2936retry:
2937 io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2938 if (!io_end) {
2939 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2940 schedule();
2941 goto retry;
2942 }
2943 io_end->offset = offset;
2944 io_end->size = size;
2945 /*
2946 * We need to hold a reference to the page to make sure it
2947 * doesn't get evicted before ext4_end_io_work() has a chance
2948 * to convert the extent from written to unwritten.
2949 */
2950 io_end->page = page;
2951 get_page(io_end->page);
2952
2953 bh->b_private = io_end;
2954 bh->b_end_io = ext4_end_io_buffer_write;
2955 return 0;
2956}
2957
2958/*
2959 * For ext4 extent files, ext4 will do direct-io write to holes,
2960 * preallocated extents, and those write extend the file, no need to
2961 * fall back to buffered IO.
2962 *
2963 * For holes, we fallocate those blocks, mark them as uninitialized
2964 * If those blocks were preallocated, we mark sure they are splited, but
2965 * still keep the range to write as uninitialized.
2966 *
2967 * The unwrritten extents will be converted to written when DIO is completed.
2968 * For async direct IO, since the IO may still pending when return, we
2969 * set up an end_io call back function, which will do the conversion
2970 * when async direct IO completed.
2971 *
2972 * If the O_DIRECT write will extend the file then add this inode to the
2973 * orphan list. So recovery will truncate it back to the original size
2974 * if the machine crashes during the write.
2975 *
2976 */
2977static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2978 const struct iovec *iov, loff_t offset,
2979 unsigned long nr_segs)
2980{
2981 struct file *file = iocb->ki_filp;
2982 struct inode *inode = file->f_mapping->host;
2983 ssize_t ret;
2984 size_t count = iov_length(iov, nr_segs);
2985
2986 loff_t final_size = offset + count;
2987 if (rw == WRITE && final_size <= inode->i_size) {
2988 /*
2989 * We could direct write to holes and fallocate.
2990 *
2991 * Allocated blocks to fill the hole are marked as uninitialized
2992 * to prevent parallel buffered read to expose the stale data
2993 * before DIO complete the data IO.
2994 *
2995 * As to previously fallocated extents, ext4 get_block
2996 * will just simply mark the buffer mapped but still
2997 * keep the extents uninitialized.
2998 *
2999 * for non AIO case, we will convert those unwritten extents
3000 * to written after return back from blockdev_direct_IO.
3001 *
3002 * for async DIO, the conversion needs to be defered when
3003 * the IO is completed. The ext4 end_io callback function
3004 * will be called to take care of the conversion work.
3005 * Here for async case, we allocate an io_end structure to
3006 * hook to the iocb.
3007 */
3008 iocb->private = NULL;
3009 EXT4_I(inode)->cur_aio_dio = NULL;
3010 if (!is_sync_kiocb(iocb)) {
3011 ext4_io_end_t *io_end =
3012 ext4_init_io_end(inode, GFP_NOFS);
3013 if (!io_end)
3014 return -ENOMEM;
3015 io_end->flag |= EXT4_IO_END_DIRECT;
3016 iocb->private = io_end;
3017 /*
3018 * we save the io structure for current async
3019 * direct IO, so that later ext4_map_blocks()
3020 * could flag the io structure whether there
3021 * is a unwritten extents needs to be converted
3022 * when IO is completed.
3023 */
3024 EXT4_I(inode)->cur_aio_dio = iocb->private;
3025 }
3026
3027 ret = __blockdev_direct_IO(rw, iocb, inode,
3028 inode->i_sb->s_bdev, iov,
3029 offset, nr_segs,
3030 ext4_get_block_write,
3031 ext4_end_io_dio,
3032 NULL,
3033 DIO_LOCKING);
3034 if (iocb->private)
3035 EXT4_I(inode)->cur_aio_dio = NULL;
3036 /*
3037 * The io_end structure takes a reference to the inode,
3038 * that structure needs to be destroyed and the
3039 * reference to the inode need to be dropped, when IO is
3040 * complete, even with 0 byte write, or failed.
3041 *
3042 * In the successful AIO DIO case, the io_end structure will be
3043 * desctroyed and the reference to the inode will be dropped
3044 * after the end_io call back function is called.
3045 *
3046 * In the case there is 0 byte write, or error case, since
3047 * VFS direct IO won't invoke the end_io call back function,
3048 * we need to free the end_io structure here.
3049 */
3050 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3051 ext4_free_io_end(iocb->private);
3052 iocb->private = NULL;
3053 } else if (ret > 0 && ext4_test_inode_state(inode,
3054 EXT4_STATE_DIO_UNWRITTEN)) {
3055 int err;
3056 /*
3057 * for non AIO case, since the IO is already
3058 * completed, we could do the conversion right here
3059 */
3060 err = ext4_convert_unwritten_extents(inode,
3061 offset, ret);
3062 if (err < 0)
3063 ret = err;
3064 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3065 }
3066 return ret;
3067 }
3068
3069 /* for write the the end of file case, we fall back to old way */
3070 return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3071}
3072
3073static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3074 const struct iovec *iov, loff_t offset,
3075 unsigned long nr_segs)
3076{
3077 struct file *file = iocb->ki_filp;
3078 struct inode *inode = file->f_mapping->host;
3079 ssize_t ret;
3080
3081 /*
3082 * If we are doing data journalling we don't support O_DIRECT
3083 */
3084 if (ext4_should_journal_data(inode))
3085 return 0;
3086
3087 trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
3088 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3089 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3090 else
3091 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3092 trace_ext4_direct_IO_exit(inode, offset,
3093 iov_length(iov, nr_segs), rw, ret);
3094 return ret;
3095}
3096
3097/*
3098 * Pages can be marked dirty completely asynchronously from ext4's journalling
3099 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3100 * much here because ->set_page_dirty is called under VFS locks. The page is
3101 * not necessarily locked.
3102 *
3103 * We cannot just dirty the page and leave attached buffers clean, because the
3104 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3105 * or jbddirty because all the journalling code will explode.
3106 *
3107 * So what we do is to mark the page "pending dirty" and next time writepage
3108 * is called, propagate that into the buffers appropriately.
3109 */
3110static int ext4_journalled_set_page_dirty(struct page *page)
3111{
3112 SetPageChecked(page);
3113 return __set_page_dirty_nobuffers(page);
3114}
3115
3116static const struct address_space_operations ext4_ordered_aops = {
3117 .readpage = ext4_readpage,
3118 .readpages = ext4_readpages,
3119 .writepage = ext4_writepage,
3120 .write_begin = ext4_write_begin,
3121 .write_end = ext4_ordered_write_end,
3122 .bmap = ext4_bmap,
3123 .invalidatepage = ext4_invalidatepage,
3124 .releasepage = ext4_releasepage,
3125 .direct_IO = ext4_direct_IO,
3126 .migratepage = buffer_migrate_page,
3127 .is_partially_uptodate = block_is_partially_uptodate,
3128 .error_remove_page = generic_error_remove_page,
3129};
3130
3131static const struct address_space_operations ext4_writeback_aops = {
3132 .readpage = ext4_readpage,
3133 .readpages = ext4_readpages,
3134 .writepage = ext4_writepage,
3135 .write_begin = ext4_write_begin,
3136 .write_end = ext4_writeback_write_end,
3137 .bmap = ext4_bmap,
3138 .invalidatepage = ext4_invalidatepage,
3139 .releasepage = ext4_releasepage,
3140 .direct_IO = ext4_direct_IO,
3141 .migratepage = buffer_migrate_page,
3142 .is_partially_uptodate = block_is_partially_uptodate,
3143 .error_remove_page = generic_error_remove_page,
3144};
3145
3146static const struct address_space_operations ext4_journalled_aops = {
3147 .readpage = ext4_readpage,
3148 .readpages = ext4_readpages,
3149 .writepage = ext4_writepage,
3150 .write_begin = ext4_write_begin,
3151 .write_end = ext4_journalled_write_end,
3152 .set_page_dirty = ext4_journalled_set_page_dirty,
3153 .bmap = ext4_bmap,
3154 .invalidatepage = ext4_invalidatepage,
3155 .releasepage = ext4_releasepage,
3156 .direct_IO = ext4_direct_IO,
3157 .is_partially_uptodate = block_is_partially_uptodate,
3158 .error_remove_page = generic_error_remove_page,
3159};
3160
3161static const struct address_space_operations ext4_da_aops = {
3162 .readpage = ext4_readpage,
3163 .readpages = ext4_readpages,
3164 .writepage = ext4_writepage,
3165 .writepages = ext4_da_writepages,
3166 .write_begin = ext4_da_write_begin,
3167 .write_end = ext4_da_write_end,
3168 .bmap = ext4_bmap,
3169 .invalidatepage = ext4_da_invalidatepage,
3170 .releasepage = ext4_releasepage,
3171 .direct_IO = ext4_direct_IO,
3172 .migratepage = buffer_migrate_page,
3173 .is_partially_uptodate = block_is_partially_uptodate,
3174 .error_remove_page = generic_error_remove_page,
3175};
3176
3177void ext4_set_aops(struct inode *inode)
3178{
3179 switch (ext4_inode_journal_mode(inode)) {
3180 case EXT4_INODE_ORDERED_DATA_MODE:
3181 if (test_opt(inode->i_sb, DELALLOC))
3182 inode->i_mapping->a_ops = &ext4_da_aops;
3183 else
3184 inode->i_mapping->a_ops = &ext4_ordered_aops;
3185 break;
3186 case EXT4_INODE_WRITEBACK_DATA_MODE:
3187 if (test_opt(inode->i_sb, DELALLOC))
3188 inode->i_mapping->a_ops = &ext4_da_aops;
3189 else
3190 inode->i_mapping->a_ops = &ext4_writeback_aops;
3191 break;
3192 case EXT4_INODE_JOURNAL_DATA_MODE:
3193 inode->i_mapping->a_ops = &ext4_journalled_aops;
3194 break;
3195 default:
3196 BUG();
3197 }
3198}
3199
3200
3201/*
3202 * ext4_discard_partial_page_buffers()
3203 * Wrapper function for ext4_discard_partial_page_buffers_no_lock.
3204 * This function finds and locks the page containing the offset
3205 * "from" and passes it to ext4_discard_partial_page_buffers_no_lock.
3206 * Calling functions that already have the page locked should call
3207 * ext4_discard_partial_page_buffers_no_lock directly.
3208 */
3209int ext4_discard_partial_page_buffers(handle_t *handle,
3210 struct address_space *mapping, loff_t from,
3211 loff_t length, int flags)
3212{
3213 struct inode *inode = mapping->host;
3214 struct page *page;
3215 int err = 0;
3216
3217 page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3218 mapping_gfp_mask(mapping) & ~__GFP_FS);
3219 if (!page)
3220 return -ENOMEM;
3221
3222 err = ext4_discard_partial_page_buffers_no_lock(handle, inode, page,
3223 from, length, flags);
3224
3225 unlock_page(page);
3226 page_cache_release(page);
3227 return err;
3228}
3229
3230/*
3231 * ext4_discard_partial_page_buffers_no_lock()
3232 * Zeros a page range of length 'length' starting from offset 'from'.
3233 * Buffer heads that correspond to the block aligned regions of the
3234 * zeroed range will be unmapped. Unblock aligned regions
3235 * will have the corresponding buffer head mapped if needed so that
3236 * that region of the page can be updated with the partial zero out.
3237 *
3238 * This function assumes that the page has already been locked. The
3239 * The range to be discarded must be contained with in the given page.
3240 * If the specified range exceeds the end of the page it will be shortened
3241 * to the end of the page that corresponds to 'from'. This function is
3242 * appropriate for updating a page and it buffer heads to be unmapped and
3243 * zeroed for blocks that have been either released, or are going to be
3244 * released.
3245 *
3246 * handle: The journal handle
3247 * inode: The files inode
3248 * page: A locked page that contains the offset "from"
3249 * from: The starting byte offset (from the begining of the file)
3250 * to begin discarding
3251 * len: The length of bytes to discard
3252 * flags: Optional flags that may be used:
3253 *
3254 * EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED
3255 * Only zero the regions of the page whose buffer heads
3256 * have already been unmapped. This flag is appropriate
3257 * for updateing the contents of a page whose blocks may
3258 * have already been released, and we only want to zero
3259 * out the regions that correspond to those released blocks.
3260 *
3261 * Returns zero on sucess or negative on failure.
3262 */
3263static int ext4_discard_partial_page_buffers_no_lock(handle_t *handle,
3264 struct inode *inode, struct page *page, loff_t from,
3265 loff_t length, int flags)
3266{
3267 ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3268 unsigned int offset = from & (PAGE_CACHE_SIZE-1);
3269 unsigned int blocksize, max, pos;
3270 ext4_lblk_t iblock;
3271 struct buffer_head *bh;
3272 int err = 0;
3273
3274 blocksize = inode->i_sb->s_blocksize;
3275 max = PAGE_CACHE_SIZE - offset;
3276
3277 if (index != page->index)
3278 return -EINVAL;
3279
3280 /*
3281 * correct length if it does not fall between
3282 * 'from' and the end of the page
3283 */
3284 if (length > max || length < 0)
3285 length = max;
3286
3287 iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3288
3289 if (!page_has_buffers(page))
3290 create_empty_buffers(page, blocksize, 0);
3291
3292 /* Find the buffer that contains "offset" */
3293 bh = page_buffers(page);
3294 pos = blocksize;
3295 while (offset >= pos) {
3296 bh = bh->b_this_page;
3297 iblock++;
3298 pos += blocksize;
3299 }
3300
3301 pos = offset;
3302 while (pos < offset + length) {
3303 unsigned int end_of_block, range_to_discard;
3304
3305 err = 0;
3306
3307 /* The length of space left to zero and unmap */
3308 range_to_discard = offset + length - pos;
3309
3310 /* The length of space until the end of the block */
3311 end_of_block = blocksize - (pos & (blocksize-1));
3312
3313 /*
3314 * Do not unmap or zero past end of block
3315 * for this buffer head
3316 */
3317 if (range_to_discard > end_of_block)
3318 range_to_discard = end_of_block;
3319
3320
3321 /*
3322 * Skip this buffer head if we are only zeroing unampped
3323 * regions of the page
3324 */
3325 if (flags & EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED &&
3326 buffer_mapped(bh))
3327 goto next;
3328
3329 /* If the range is block aligned, unmap */
3330 if (range_to_discard == blocksize) {
3331 clear_buffer_dirty(bh);
3332 bh->b_bdev = NULL;
3333 clear_buffer_mapped(bh);
3334 clear_buffer_req(bh);
3335 clear_buffer_new(bh);
3336 clear_buffer_delay(bh);
3337 clear_buffer_unwritten(bh);
3338 clear_buffer_uptodate(bh);
3339 zero_user(page, pos, range_to_discard);
3340 BUFFER_TRACE(bh, "Buffer discarded");
3341 goto next;
3342 }
3343
3344 /*
3345 * If this block is not completely contained in the range
3346 * to be discarded, then it is not going to be released. Because
3347 * we need to keep this block, we need to make sure this part
3348 * of the page is uptodate before we modify it by writeing
3349 * partial zeros on it.
3350 */
3351 if (!buffer_mapped(bh)) {
3352 /*
3353 * Buffer head must be mapped before we can read
3354 * from the block
3355 */
3356 BUFFER_TRACE(bh, "unmapped");
3357 ext4_get_block(inode, iblock, bh, 0);
3358 /* unmapped? It's a hole - nothing to do */
3359 if (!buffer_mapped(bh)) {
3360 BUFFER_TRACE(bh, "still unmapped");
3361 goto next;
3362 }
3363 }
3364
3365 /* Ok, it's mapped. Make sure it's up-to-date */
3366 if (PageUptodate(page))
3367 set_buffer_uptodate(bh);
3368
3369 if (!buffer_uptodate(bh)) {
3370 err = -EIO;
3371 ll_rw_block(READ, 1, &bh);
3372 wait_on_buffer(bh);
3373 /* Uhhuh. Read error. Complain and punt.*/
3374 if (!buffer_uptodate(bh))
3375 goto next;
3376 }
3377
3378 if (ext4_should_journal_data(inode)) {
3379 BUFFER_TRACE(bh, "get write access");
3380 err = ext4_journal_get_write_access(handle, bh);
3381 if (err)
3382 goto next;
3383 }
3384
3385 zero_user(page, pos, range_to_discard);
3386
3387 err = 0;
3388 if (ext4_should_journal_data(inode)) {
3389 err = ext4_handle_dirty_metadata(handle, inode, bh);
3390 } else
3391 mark_buffer_dirty(bh);
3392
3393 BUFFER_TRACE(bh, "Partial buffer zeroed");
3394next:
3395 bh = bh->b_this_page;
3396 iblock++;
3397 pos += range_to_discard;
3398 }
3399
3400 return err;
3401}
3402
3403int ext4_can_truncate(struct inode *inode)
3404{
3405 if (S_ISREG(inode->i_mode))
3406 return 1;
3407 if (S_ISDIR(inode->i_mode))
3408 return 1;
3409 if (S_ISLNK(inode->i_mode))
3410 return !ext4_inode_is_fast_symlink(inode);
3411 return 0;
3412}
3413
3414/*
3415 * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3416 * associated with the given offset and length
3417 *
3418 * @inode: File inode
3419 * @offset: The offset where the hole will begin
3420 * @len: The length of the hole
3421 *
3422 * Returns: 0 on sucess or negative on failure
3423 */
3424
3425int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3426{
3427 struct inode *inode = file->f_path.dentry->d_inode;
3428 if (!S_ISREG(inode->i_mode))
3429 return -EOPNOTSUPP;
3430
3431 if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3432 /* TODO: Add support for non extent hole punching */
3433 return -EOPNOTSUPP;
3434 }
3435
3436 if (EXT4_SB(inode->i_sb)->s_cluster_ratio > 1) {
3437 /* TODO: Add support for bigalloc file systems */
3438 return -EOPNOTSUPP;
3439 }
3440
3441 return ext4_ext_punch_hole(file, offset, length);
3442}
3443
3444/*
3445 * ext4_truncate()
3446 *
3447 * We block out ext4_get_block() block instantiations across the entire
3448 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3449 * simultaneously on behalf of the same inode.
3450 *
3451 * As we work through the truncate and commit bits of it to the journal there
3452 * is one core, guiding principle: the file's tree must always be consistent on
3453 * disk. We must be able to restart the truncate after a crash.
3454 *
3455 * The file's tree may be transiently inconsistent in memory (although it
3456 * probably isn't), but whenever we close off and commit a journal transaction,
3457 * the contents of (the filesystem + the journal) must be consistent and
3458 * restartable. It's pretty simple, really: bottom up, right to left (although
3459 * left-to-right works OK too).
3460 *
3461 * Note that at recovery time, journal replay occurs *before* the restart of
3462 * truncate against the orphan inode list.
3463 *
3464 * The committed inode has the new, desired i_size (which is the same as
3465 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
3466 * that this inode's truncate did not complete and it will again call
3467 * ext4_truncate() to have another go. So there will be instantiated blocks
3468 * to the right of the truncation point in a crashed ext4 filesystem. But
3469 * that's fine - as long as they are linked from the inode, the post-crash
3470 * ext4_truncate() run will find them and release them.
3471 */
3472void ext4_truncate(struct inode *inode)
3473{
3474 trace_ext4_truncate_enter(inode);
3475
3476 if (!ext4_can_truncate(inode))
3477 return;
3478
3479 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3480
3481 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3482 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3483
3484 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3485 ext4_ext_truncate(inode);
3486 else
3487 ext4_ind_truncate(inode);
3488
3489 trace_ext4_truncate_exit(inode);
3490}
3491
3492/*
3493 * ext4_get_inode_loc returns with an extra refcount against the inode's
3494 * underlying buffer_head on success. If 'in_mem' is true, we have all
3495 * data in memory that is needed to recreate the on-disk version of this
3496 * inode.
3497 */
3498static int __ext4_get_inode_loc(struct inode *inode,
3499 struct ext4_iloc *iloc, int in_mem)
3500{
3501 struct ext4_group_desc *gdp;
3502 struct buffer_head *bh;
3503 struct super_block *sb = inode->i_sb;
3504 ext4_fsblk_t block;
3505 int inodes_per_block, inode_offset;
3506
3507 iloc->bh = NULL;
3508 if (!ext4_valid_inum(sb, inode->i_ino))
3509 return -EIO;
3510
3511 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3512 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3513 if (!gdp)
3514 return -EIO;
3515
3516 /*
3517 * Figure out the offset within the block group inode table
3518 */
3519 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3520 inode_offset = ((inode->i_ino - 1) %
3521 EXT4_INODES_PER_GROUP(sb));
3522 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3523 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3524
3525 bh = sb_getblk(sb, block);
3526 if (!bh) {
3527 EXT4_ERROR_INODE_BLOCK(inode, block,
3528 "unable to read itable block");
3529 return -EIO;
3530 }
3531 if (!buffer_uptodate(bh)) {
3532 lock_buffer(bh);
3533
3534 /*
3535 * If the buffer has the write error flag, we have failed
3536 * to write out another inode in the same block. In this
3537 * case, we don't have to read the block because we may
3538 * read the old inode data successfully.
3539 */
3540 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3541 set_buffer_uptodate(bh);
3542
3543 if (buffer_uptodate(bh)) {
3544 /* someone brought it uptodate while we waited */
3545 unlock_buffer(bh);
3546 goto has_buffer;
3547 }
3548
3549 /*
3550 * If we have all information of the inode in memory and this
3551 * is the only valid inode in the block, we need not read the
3552 * block.
3553 */
3554 if (in_mem) {
3555 struct buffer_head *bitmap_bh;
3556 int i, start;
3557
3558 start = inode_offset & ~(inodes_per_block - 1);
3559
3560 /* Is the inode bitmap in cache? */
3561 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3562 if (!bitmap_bh)
3563 goto make_io;
3564
3565 /*
3566 * If the inode bitmap isn't in cache then the
3567 * optimisation may end up performing two reads instead
3568 * of one, so skip it.
3569 */
3570 if (!buffer_uptodate(bitmap_bh)) {
3571 brelse(bitmap_bh);
3572 goto make_io;
3573 }
3574 for (i = start; i < start + inodes_per_block; i++) {
3575 if (i == inode_offset)
3576 continue;
3577 if (ext4_test_bit(i, bitmap_bh->b_data))
3578 break;
3579 }
3580 brelse(bitmap_bh);
3581 if (i == start + inodes_per_block) {
3582 /* all other inodes are free, so skip I/O */
3583 memset(bh->b_data, 0, bh->b_size);
3584 set_buffer_uptodate(bh);
3585 unlock_buffer(bh);
3586 goto has_buffer;
3587 }
3588 }
3589
3590make_io:
3591 /*
3592 * If we need to do any I/O, try to pre-readahead extra
3593 * blocks from the inode table.
3594 */
3595 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3596 ext4_fsblk_t b, end, table;
3597 unsigned num;
3598
3599 table = ext4_inode_table(sb, gdp);
3600 /* s_inode_readahead_blks is always a power of 2 */
3601 b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3602 if (table > b)
3603 b = table;
3604 end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3605 num = EXT4_INODES_PER_GROUP(sb);
3606 if (ext4_has_group_desc_csum(sb))
3607 num -= ext4_itable_unused_count(sb, gdp);
3608 table += num / inodes_per_block;
3609 if (end > table)
3610 end = table;
3611 while (b <= end)
3612 sb_breadahead(sb, b++);
3613 }
3614
3615 /*
3616 * There are other valid inodes in the buffer, this inode
3617 * has in-inode xattrs, or we don't have this inode in memory.
3618 * Read the block from disk.
3619 */
3620 trace_ext4_load_inode(inode);
3621 get_bh(bh);
3622 bh->b_end_io = end_buffer_read_sync;
3623 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3624 wait_on_buffer(bh);
3625 if (!buffer_uptodate(bh)) {
3626 EXT4_ERROR_INODE_BLOCK(inode, block,
3627 "unable to read itable block");
3628 brelse(bh);
3629 return -EIO;
3630 }
3631 }
3632has_buffer:
3633 iloc->bh = bh;
3634 return 0;
3635}
3636
3637int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3638{
3639 /* We have all inode data except xattrs in memory here. */
3640 return __ext4_get_inode_loc(inode, iloc,
3641 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3642}
3643
3644void ext4_set_inode_flags(struct inode *inode)
3645{
3646 unsigned int flags = EXT4_I(inode)->i_flags;
3647
3648 inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3649 if (flags & EXT4_SYNC_FL)
3650 inode->i_flags |= S_SYNC;
3651 if (flags & EXT4_APPEND_FL)
3652 inode->i_flags |= S_APPEND;
3653 if (flags & EXT4_IMMUTABLE_FL)
3654 inode->i_flags |= S_IMMUTABLE;
3655 if (flags & EXT4_NOATIME_FL)
3656 inode->i_flags |= S_NOATIME;
3657 if (flags & EXT4_DIRSYNC_FL)
3658 inode->i_flags |= S_DIRSYNC;
3659}
3660
3661/* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3662void ext4_get_inode_flags(struct ext4_inode_info *ei)
3663{
3664 unsigned int vfs_fl;
3665 unsigned long old_fl, new_fl;
3666
3667 do {
3668 vfs_fl = ei->vfs_inode.i_flags;
3669 old_fl = ei->i_flags;
3670 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3671 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3672 EXT4_DIRSYNC_FL);
3673 if (vfs_fl & S_SYNC)
3674 new_fl |= EXT4_SYNC_FL;
3675 if (vfs_fl & S_APPEND)
3676 new_fl |= EXT4_APPEND_FL;
3677 if (vfs_fl & S_IMMUTABLE)
3678 new_fl |= EXT4_IMMUTABLE_FL;
3679 if (vfs_fl & S_NOATIME)
3680 new_fl |= EXT4_NOATIME_FL;
3681 if (vfs_fl & S_DIRSYNC)
3682 new_fl |= EXT4_DIRSYNC_FL;
3683 } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3684}
3685
3686static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3687 struct ext4_inode_info *ei)
3688{
3689 blkcnt_t i_blocks ;
3690 struct inode *inode = &(ei->vfs_inode);
3691 struct super_block *sb = inode->i_sb;
3692
3693 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3694 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3695 /* we are using combined 48 bit field */
3696 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3697 le32_to_cpu(raw_inode->i_blocks_lo);
3698 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
3699 /* i_blocks represent file system block size */
3700 return i_blocks << (inode->i_blkbits - 9);
3701 } else {
3702 return i_blocks;
3703 }
3704 } else {
3705 return le32_to_cpu(raw_inode->i_blocks_lo);
3706 }
3707}
3708
3709struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3710{
3711 struct ext4_iloc iloc;
3712 struct ext4_inode *raw_inode;
3713 struct ext4_inode_info *ei;
3714 struct inode *inode;
3715 journal_t *journal = EXT4_SB(sb)->s_journal;
3716 long ret;
3717 int block;
3718 uid_t i_uid;
3719 gid_t i_gid;
3720
3721 inode = iget_locked(sb, ino);
3722 if (!inode)
3723 return ERR_PTR(-ENOMEM);
3724 if (!(inode->i_state & I_NEW))
3725 return inode;
3726
3727 ei = EXT4_I(inode);
3728 iloc.bh = NULL;
3729
3730 ret = __ext4_get_inode_loc(inode, &iloc, 0);
3731 if (ret < 0)
3732 goto bad_inode;
3733 raw_inode = ext4_raw_inode(&iloc);
3734
3735 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3736 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3737 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3738 EXT4_INODE_SIZE(inode->i_sb)) {
3739 EXT4_ERROR_INODE(inode, "bad extra_isize (%u != %u)",
3740 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize,
3741 EXT4_INODE_SIZE(inode->i_sb));
3742 ret = -EIO;
3743 goto bad_inode;
3744 }
3745 } else
3746 ei->i_extra_isize = 0;
3747
3748 /* Precompute checksum seed for inode metadata */
3749 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3750 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM)) {
3751 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3752 __u32 csum;
3753 __le32 inum = cpu_to_le32(inode->i_ino);
3754 __le32 gen = raw_inode->i_generation;
3755 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
3756 sizeof(inum));
3757 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
3758 sizeof(gen));
3759 }
3760
3761 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
3762 EXT4_ERROR_INODE(inode, "checksum invalid");
3763 ret = -EIO;
3764 goto bad_inode;
3765 }
3766
3767 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3768 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3769 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3770 if (!(test_opt(inode->i_sb, NO_UID32))) {
3771 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3772 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3773 }
3774 i_uid_write(inode, i_uid);
3775 i_gid_write(inode, i_gid);
3776 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
3777
3778 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
3779 ei->i_dir_start_lookup = 0;
3780 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3781 /* We now have enough fields to check if the inode was active or not.
3782 * This is needed because nfsd might try to access dead inodes
3783 * the test is that same one that e2fsck uses
3784 * NeilBrown 1999oct15
3785 */
3786 if (inode->i_nlink == 0) {
3787 if (inode->i_mode == 0 ||
3788 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3789 /* this inode is deleted */
3790 ret = -ESTALE;
3791 goto bad_inode;
3792 }
3793 /* The only unlinked inodes we let through here have
3794 * valid i_mode and are being read by the orphan
3795 * recovery code: that's fine, we're about to complete
3796 * the process of deleting those. */
3797 }
3798 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3799 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3800 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3801 if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
3802 ei->i_file_acl |=
3803 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3804 inode->i_size = ext4_isize(raw_inode);
3805 ei->i_disksize = inode->i_size;
3806#ifdef CONFIG_QUOTA
3807 ei->i_reserved_quota = 0;
3808#endif
3809 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3810 ei->i_block_group = iloc.block_group;
3811 ei->i_last_alloc_group = ~0;
3812 /*
3813 * NOTE! The in-memory inode i_data array is in little-endian order
3814 * even on big-endian machines: we do NOT byteswap the block numbers!
3815 */
3816 for (block = 0; block < EXT4_N_BLOCKS; block++)
3817 ei->i_data[block] = raw_inode->i_block[block];
3818 INIT_LIST_HEAD(&ei->i_orphan);
3819
3820 /*
3821 * Set transaction id's of transactions that have to be committed
3822 * to finish f[data]sync. We set them to currently running transaction
3823 * as we cannot be sure that the inode or some of its metadata isn't
3824 * part of the transaction - the inode could have been reclaimed and
3825 * now it is reread from disk.
3826 */
3827 if (journal) {
3828 transaction_t *transaction;
3829 tid_t tid;
3830
3831 read_lock(&journal->j_state_lock);
3832 if (journal->j_running_transaction)
3833 transaction = journal->j_running_transaction;
3834 else
3835 transaction = journal->j_committing_transaction;
3836 if (transaction)
3837 tid = transaction->t_tid;
3838 else
3839 tid = journal->j_commit_sequence;
3840 read_unlock(&journal->j_state_lock);
3841 ei->i_sync_tid = tid;
3842 ei->i_datasync_tid = tid;
3843 }
3844
3845 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3846 if (ei->i_extra_isize == 0) {
3847 /* The extra space is currently unused. Use it. */
3848 ei->i_extra_isize = sizeof(struct ext4_inode) -
3849 EXT4_GOOD_OLD_INODE_SIZE;
3850 } else {
3851 __le32 *magic = (void *)raw_inode +
3852 EXT4_GOOD_OLD_INODE_SIZE +
3853 ei->i_extra_isize;
3854 if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3855 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
3856 }
3857 }
3858
3859 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
3860 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
3861 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
3862 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
3863
3864 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
3865 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3866 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3867 inode->i_version |=
3868 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
3869 }
3870
3871 ret = 0;
3872 if (ei->i_file_acl &&
3873 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
3874 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
3875 ei->i_file_acl);
3876 ret = -EIO;
3877 goto bad_inode;
3878 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3879 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3880 (S_ISLNK(inode->i_mode) &&
3881 !ext4_inode_is_fast_symlink(inode)))
3882 /* Validate extent which is part of inode */
3883 ret = ext4_ext_check_inode(inode);
3884 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3885 (S_ISLNK(inode->i_mode) &&
3886 !ext4_inode_is_fast_symlink(inode))) {
3887 /* Validate block references which are part of inode */
3888 ret = ext4_ind_check_inode(inode);
3889 }
3890 if (ret)
3891 goto bad_inode;
3892
3893 if (S_ISREG(inode->i_mode)) {
3894 inode->i_op = &ext4_file_inode_operations;
3895 inode->i_fop = &ext4_file_operations;
3896 ext4_set_aops(inode);
3897 } else if (S_ISDIR(inode->i_mode)) {
3898 inode->i_op = &ext4_dir_inode_operations;
3899 inode->i_fop = &ext4_dir_operations;
3900 } else if (S_ISLNK(inode->i_mode)) {
3901 if (ext4_inode_is_fast_symlink(inode)) {
3902 inode->i_op = &ext4_fast_symlink_inode_operations;
3903 nd_terminate_link(ei->i_data, inode->i_size,
3904 sizeof(ei->i_data) - 1);
3905 } else {
3906 inode->i_op = &ext4_symlink_inode_operations;
3907 ext4_set_aops(inode);
3908 }
3909 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
3910 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
3911 inode->i_op = &ext4_special_inode_operations;
3912 if (raw_inode->i_block[0])
3913 init_special_inode(inode, inode->i_mode,
3914 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3915 else
3916 init_special_inode(inode, inode->i_mode,
3917 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3918 } else {
3919 ret = -EIO;
3920 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
3921 goto bad_inode;
3922 }
3923 brelse(iloc.bh);
3924 ext4_set_inode_flags(inode);
3925 unlock_new_inode(inode);
3926 return inode;
3927
3928bad_inode:
3929 brelse(iloc.bh);
3930 iget_failed(inode);
3931 return ERR_PTR(ret);
3932}
3933
3934static int ext4_inode_blocks_set(handle_t *handle,
3935 struct ext4_inode *raw_inode,
3936 struct ext4_inode_info *ei)
3937{
3938 struct inode *inode = &(ei->vfs_inode);
3939 u64 i_blocks = inode->i_blocks;
3940 struct super_block *sb = inode->i_sb;
3941
3942 if (i_blocks <= ~0U) {
3943 /*
3944 * i_blocks can be represnted in a 32 bit variable
3945 * as multiple of 512 bytes
3946 */
3947 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3948 raw_inode->i_blocks_high = 0;
3949 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3950 return 0;
3951 }
3952 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
3953 return -EFBIG;
3954
3955 if (i_blocks <= 0xffffffffffffULL) {
3956 /*
3957 * i_blocks can be represented in a 48 bit variable
3958 * as multiple of 512 bytes
3959 */
3960 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3961 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3962 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3963 } else {
3964 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3965 /* i_block is stored in file system block size */
3966 i_blocks = i_blocks >> (inode->i_blkbits - 9);
3967 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
3968 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3969 }
3970 return 0;
3971}
3972
3973/*
3974 * Post the struct inode info into an on-disk inode location in the
3975 * buffer-cache. This gobbles the caller's reference to the
3976 * buffer_head in the inode location struct.
3977 *
3978 * The caller must have write access to iloc->bh.
3979 */
3980static int ext4_do_update_inode(handle_t *handle,
3981 struct inode *inode,
3982 struct ext4_iloc *iloc)
3983{
3984 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
3985 struct ext4_inode_info *ei = EXT4_I(inode);
3986 struct buffer_head *bh = iloc->bh;
3987 int err = 0, rc, block;
3988 uid_t i_uid;
3989 gid_t i_gid;
3990
3991 /* For fields not not tracking in the in-memory inode,
3992 * initialise them to zero for new inodes. */
3993 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
3994 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
3995
3996 ext4_get_inode_flags(ei);
3997 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3998 i_uid = i_uid_read(inode);
3999 i_gid = i_gid_read(inode);
4000 if (!(test_opt(inode->i_sb, NO_UID32))) {
4001 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4002 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4003/*
4004 * Fix up interoperability with old kernels. Otherwise, old inodes get
4005 * re-used with the upper 16 bits of the uid/gid intact
4006 */
4007 if (!ei->i_dtime) {
4008 raw_inode->i_uid_high =
4009 cpu_to_le16(high_16_bits(i_uid));
4010 raw_inode->i_gid_high =
4011 cpu_to_le16(high_16_bits(i_gid));
4012 } else {
4013 raw_inode->i_uid_high = 0;
4014 raw_inode->i_gid_high = 0;
4015 }
4016 } else {
4017 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4018 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4019 raw_inode->i_uid_high = 0;
4020 raw_inode->i_gid_high = 0;
4021 }
4022 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4023
4024 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4025 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4026 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4027 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4028
4029 if (ext4_inode_blocks_set(handle, raw_inode, ei))
4030 goto out_brelse;
4031 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4032 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4033 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4034 cpu_to_le32(EXT4_OS_HURD))
4035 raw_inode->i_file_acl_high =
4036 cpu_to_le16(ei->i_file_acl >> 32);
4037 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4038 ext4_isize_set(raw_inode, ei->i_disksize);
4039 if (ei->i_disksize > 0x7fffffffULL) {
4040 struct super_block *sb = inode->i_sb;
4041 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4042 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4043 EXT4_SB(sb)->s_es->s_rev_level ==
4044 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4045 /* If this is the first large file
4046 * created, add a flag to the superblock.
4047 */
4048 err = ext4_journal_get_write_access(handle,
4049 EXT4_SB(sb)->s_sbh);
4050 if (err)
4051 goto out_brelse;
4052 ext4_update_dynamic_rev(sb);
4053 EXT4_SET_RO_COMPAT_FEATURE(sb,
4054 EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4055 ext4_handle_sync(handle);
4056 err = ext4_handle_dirty_super_now(handle, sb);
4057 }
4058 }
4059 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4060 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4061 if (old_valid_dev(inode->i_rdev)) {
4062 raw_inode->i_block[0] =
4063 cpu_to_le32(old_encode_dev(inode->i_rdev));
4064 raw_inode->i_block[1] = 0;
4065 } else {
4066 raw_inode->i_block[0] = 0;
4067 raw_inode->i_block[1] =
4068 cpu_to_le32(new_encode_dev(inode->i_rdev));
4069 raw_inode->i_block[2] = 0;
4070 }
4071 } else
4072 for (block = 0; block < EXT4_N_BLOCKS; block++)
4073 raw_inode->i_block[block] = ei->i_data[block];
4074
4075 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4076 if (ei->i_extra_isize) {
4077 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4078 raw_inode->i_version_hi =
4079 cpu_to_le32(inode->i_version >> 32);
4080 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4081 }
4082
4083 ext4_inode_csum_set(inode, raw_inode, ei);
4084
4085 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4086 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
4087 if (!err)
4088 err = rc;
4089 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
4090
4091 ext4_update_inode_fsync_trans(handle, inode, 0);
4092out_brelse:
4093 brelse(bh);
4094 ext4_std_error(inode->i_sb, err);
4095 return err;
4096}
4097
4098/*
4099 * ext4_write_inode()
4100 *
4101 * We are called from a few places:
4102 *
4103 * - Within generic_file_write() for O_SYNC files.
4104 * Here, there will be no transaction running. We wait for any running
4105 * trasnaction to commit.
4106 *
4107 * - Within sys_sync(), kupdate and such.
4108 * We wait on commit, if tol to.
4109 *
4110 * - Within prune_icache() (PF_MEMALLOC == true)
4111 * Here we simply return. We can't afford to block kswapd on the
4112 * journal commit.
4113 *
4114 * In all cases it is actually safe for us to return without doing anything,
4115 * because the inode has been copied into a raw inode buffer in
4116 * ext4_mark_inode_dirty(). This is a correctness thing for O_SYNC and for
4117 * knfsd.
4118 *
4119 * Note that we are absolutely dependent upon all inode dirtiers doing the
4120 * right thing: they *must* call mark_inode_dirty() after dirtying info in
4121 * which we are interested.
4122 *
4123 * It would be a bug for them to not do this. The code:
4124 *
4125 * mark_inode_dirty(inode)
4126 * stuff();
4127 * inode->i_size = expr;
4128 *
4129 * is in error because a kswapd-driven write_inode() could occur while
4130 * `stuff()' is running, and the new i_size will be lost. Plus the inode
4131 * will no longer be on the superblock's dirty inode list.
4132 */
4133int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
4134{
4135 int err;
4136
4137 if (current->flags & PF_MEMALLOC)
4138 return 0;
4139
4140 if (EXT4_SB(inode->i_sb)->s_journal) {
4141 if (ext4_journal_current_handle()) {
4142 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4143 dump_stack();
4144 return -EIO;
4145 }
4146
4147 if (wbc->sync_mode != WB_SYNC_ALL)
4148 return 0;
4149
4150 err = ext4_force_commit(inode->i_sb);
4151 } else {
4152 struct ext4_iloc iloc;
4153
4154 err = __ext4_get_inode_loc(inode, &iloc, 0);
4155 if (err)
4156 return err;
4157 if (wbc->sync_mode == WB_SYNC_ALL)
4158 sync_dirty_buffer(iloc.bh);
4159 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
4160 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
4161 "IO error syncing inode");
4162 err = -EIO;
4163 }
4164 brelse(iloc.bh);
4165 }
4166 return err;
4167}
4168
4169/*
4170 * ext4_setattr()
4171 *
4172 * Called from notify_change.
4173 *
4174 * We want to trap VFS attempts to truncate the file as soon as
4175 * possible. In particular, we want to make sure that when the VFS
4176 * shrinks i_size, we put the inode on the orphan list and modify
4177 * i_disksize immediately, so that during the subsequent flushing of
4178 * dirty pages and freeing of disk blocks, we can guarantee that any
4179 * commit will leave the blocks being flushed in an unused state on
4180 * disk. (On recovery, the inode will get truncated and the blocks will
4181 * be freed, so we have a strong guarantee that no future commit will
4182 * leave these blocks visible to the user.)
4183 *
4184 * Another thing we have to assure is that if we are in ordered mode
4185 * and inode is still attached to the committing transaction, we must
4186 * we start writeout of all the dirty pages which are being truncated.
4187 * This way we are sure that all the data written in the previous
4188 * transaction are already on disk (truncate waits for pages under
4189 * writeback).
4190 *
4191 * Called with inode->i_mutex down.
4192 */
4193int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4194{
4195 struct inode *inode = dentry->d_inode;
4196 int error, rc = 0;
4197 int orphan = 0;
4198 const unsigned int ia_valid = attr->ia_valid;
4199
4200 error = inode_change_ok(inode, attr);
4201 if (error)
4202 return error;
4203
4204 if (is_quota_modification(inode, attr))
4205 dquot_initialize(inode);
4206 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
4207 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
4208 handle_t *handle;
4209
4210 /* (user+group)*(old+new) structure, inode write (sb,
4211 * inode block, ? - but truncate inode update has it) */
4212 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
4213 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
4214 if (IS_ERR(handle)) {
4215 error = PTR_ERR(handle);
4216 goto err_out;
4217 }
4218 error = dquot_transfer(inode, attr);
4219 if (error) {
4220 ext4_journal_stop(handle);
4221 return error;
4222 }
4223 /* Update corresponding info in inode so that everything is in
4224 * one transaction */
4225 if (attr->ia_valid & ATTR_UID)
4226 inode->i_uid = attr->ia_uid;
4227 if (attr->ia_valid & ATTR_GID)
4228 inode->i_gid = attr->ia_gid;
4229 error = ext4_mark_inode_dirty(handle, inode);
4230 ext4_journal_stop(handle);
4231 }
4232
4233 if (attr->ia_valid & ATTR_SIZE) {
4234 inode_dio_wait(inode);
4235
4236 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4237 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4238
4239 if (attr->ia_size > sbi->s_bitmap_maxbytes)
4240 return -EFBIG;
4241 }
4242 }
4243
4244 if (S_ISREG(inode->i_mode) &&
4245 attr->ia_valid & ATTR_SIZE &&
4246 (attr->ia_size < inode->i_size)) {
4247 handle_t *handle;
4248
4249 handle = ext4_journal_start(inode, 3);
4250 if (IS_ERR(handle)) {
4251 error = PTR_ERR(handle);
4252 goto err_out;
4253 }
4254 if (ext4_handle_valid(handle)) {
4255 error = ext4_orphan_add(handle, inode);
4256 orphan = 1;
4257 }
4258 EXT4_I(inode)->i_disksize = attr->ia_size;
4259 rc = ext4_mark_inode_dirty(handle, inode);
4260 if (!error)
4261 error = rc;
4262 ext4_journal_stop(handle);
4263
4264 if (ext4_should_order_data(inode)) {
4265 error = ext4_begin_ordered_truncate(inode,
4266 attr->ia_size);
4267 if (error) {
4268 /* Do as much error cleanup as possible */
4269 handle = ext4_journal_start(inode, 3);
4270 if (IS_ERR(handle)) {
4271 ext4_orphan_del(NULL, inode);
4272 goto err_out;
4273 }
4274 ext4_orphan_del(handle, inode);
4275 orphan = 0;
4276 ext4_journal_stop(handle);
4277 goto err_out;
4278 }
4279 }
4280 }
4281
4282 if (attr->ia_valid & ATTR_SIZE) {
4283 if (attr->ia_size != i_size_read(inode))
4284 truncate_setsize(inode, attr->ia_size);
4285 ext4_truncate(inode);
4286 }
4287
4288 if (!rc) {
4289 setattr_copy(inode, attr);
4290 mark_inode_dirty(inode);
4291 }
4292
4293 /*
4294 * If the call to ext4_truncate failed to get a transaction handle at
4295 * all, we need to clean up the in-core orphan list manually.
4296 */
4297 if (orphan && inode->i_nlink)
4298 ext4_orphan_del(NULL, inode);
4299
4300 if (!rc && (ia_valid & ATTR_MODE))
4301 rc = ext4_acl_chmod(inode);
4302
4303err_out:
4304 ext4_std_error(inode->i_sb, error);
4305 if (!error)
4306 error = rc;
4307 return error;
4308}
4309
4310int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4311 struct kstat *stat)
4312{
4313 struct inode *inode;
4314 unsigned long delalloc_blocks;
4315
4316 inode = dentry->d_inode;
4317 generic_fillattr(inode, stat);
4318
4319 /*
4320 * We can't update i_blocks if the block allocation is delayed
4321 * otherwise in the case of system crash before the real block
4322 * allocation is done, we will have i_blocks inconsistent with
4323 * on-disk file blocks.
4324 * We always keep i_blocks updated together with real
4325 * allocation. But to not confuse with user, stat
4326 * will return the blocks that include the delayed allocation
4327 * blocks for this file.
4328 */
4329 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
4330 EXT4_I(inode)->i_reserved_data_blocks);
4331
4332 stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4333 return 0;
4334}
4335
4336static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4337{
4338 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4339 return ext4_ind_trans_blocks(inode, nrblocks, chunk);
4340 return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4341}
4342
4343/*
4344 * Account for index blocks, block groups bitmaps and block group
4345 * descriptor blocks if modify datablocks and index blocks
4346 * worse case, the indexs blocks spread over different block groups
4347 *
4348 * If datablocks are discontiguous, they are possible to spread over
4349 * different block groups too. If they are contiuguous, with flexbg,
4350 * they could still across block group boundary.
4351 *
4352 * Also account for superblock, inode, quota and xattr blocks
4353 */
4354static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4355{
4356 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4357 int gdpblocks;
4358 int idxblocks;
4359 int ret = 0;
4360
4361 /*
4362 * How many index blocks need to touch to modify nrblocks?
4363 * The "Chunk" flag indicating whether the nrblocks is
4364 * physically contiguous on disk
4365 *
4366 * For Direct IO and fallocate, they calls get_block to allocate
4367 * one single extent at a time, so they could set the "Chunk" flag
4368 */
4369 idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4370
4371 ret = idxblocks;
4372
4373 /*
4374 * Now let's see how many group bitmaps and group descriptors need
4375 * to account
4376 */
4377 groups = idxblocks;
4378 if (chunk)
4379 groups += 1;
4380 else
4381 groups += nrblocks;
4382
4383 gdpblocks = groups;
4384 if (groups > ngroups)
4385 groups = ngroups;
4386 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4387 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4388
4389 /* bitmaps and block group descriptor blocks */
4390 ret += groups + gdpblocks;
4391
4392 /* Blocks for super block, inode, quota and xattr blocks */
4393 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4394
4395 return ret;
4396}
4397
4398/*
4399 * Calculate the total number of credits to reserve to fit
4400 * the modification of a single pages into a single transaction,
4401 * which may include multiple chunks of block allocations.
4402 *
4403 * This could be called via ext4_write_begin()
4404 *
4405 * We need to consider the worse case, when
4406 * one new block per extent.
4407 */
4408int ext4_writepage_trans_blocks(struct inode *inode)
4409{
4410 int bpp = ext4_journal_blocks_per_page(inode);
4411 int ret;
4412
4413 ret = ext4_meta_trans_blocks(inode, bpp, 0);
4414
4415 /* Account for data blocks for journalled mode */
4416 if (ext4_should_journal_data(inode))
4417 ret += bpp;
4418 return ret;
4419}
4420
4421/*
4422 * Calculate the journal credits for a chunk of data modification.
4423 *
4424 * This is called from DIO, fallocate or whoever calling
4425 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4426 *
4427 * journal buffers for data blocks are not included here, as DIO
4428 * and fallocate do no need to journal data buffers.
4429 */
4430int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4431{
4432 return ext4_meta_trans_blocks(inode, nrblocks, 1);
4433}
4434
4435/*
4436 * The caller must have previously called ext4_reserve_inode_write().
4437 * Give this, we know that the caller already has write access to iloc->bh.
4438 */
4439int ext4_mark_iloc_dirty(handle_t *handle,
4440 struct inode *inode, struct ext4_iloc *iloc)
4441{
4442 int err = 0;
4443
4444 if (IS_I_VERSION(inode))
4445 inode_inc_iversion(inode);
4446
4447 /* the do_update_inode consumes one bh->b_count */
4448 get_bh(iloc->bh);
4449
4450 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4451 err = ext4_do_update_inode(handle, inode, iloc);
4452 put_bh(iloc->bh);
4453 return err;
4454}
4455
4456/*
4457 * On success, We end up with an outstanding reference count against
4458 * iloc->bh. This _must_ be cleaned up later.
4459 */
4460
4461int
4462ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4463 struct ext4_iloc *iloc)
4464{
4465 int err;
4466
4467 err = ext4_get_inode_loc(inode, iloc);
4468 if (!err) {
4469 BUFFER_TRACE(iloc->bh, "get_write_access");
4470 err = ext4_journal_get_write_access(handle, iloc->bh);
4471 if (err) {
4472 brelse(iloc->bh);
4473 iloc->bh = NULL;
4474 }
4475 }
4476 ext4_std_error(inode->i_sb, err);
4477 return err;
4478}
4479
4480/*
4481 * Expand an inode by new_extra_isize bytes.
4482 * Returns 0 on success or negative error number on failure.
4483 */
4484static int ext4_expand_extra_isize(struct inode *inode,
4485 unsigned int new_extra_isize,
4486 struct ext4_iloc iloc,
4487 handle_t *handle)
4488{
4489 struct ext4_inode *raw_inode;
4490 struct ext4_xattr_ibody_header *header;
4491
4492 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4493 return 0;
4494
4495 raw_inode = ext4_raw_inode(&iloc);
4496
4497 header = IHDR(inode, raw_inode);
4498
4499 /* No extended attributes present */
4500 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4501 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4502 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4503 new_extra_isize);
4504 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4505 return 0;
4506 }
4507
4508 /* try to expand with EAs present */
4509 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4510 raw_inode, handle);
4511}
4512
4513/*
4514 * What we do here is to mark the in-core inode as clean with respect to inode
4515 * dirtiness (it may still be data-dirty).
4516 * This means that the in-core inode may be reaped by prune_icache
4517 * without having to perform any I/O. This is a very good thing,
4518 * because *any* task may call prune_icache - even ones which
4519 * have a transaction open against a different journal.
4520 *
4521 * Is this cheating? Not really. Sure, we haven't written the
4522 * inode out, but prune_icache isn't a user-visible syncing function.
4523 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4524 * we start and wait on commits.
4525 *
4526 * Is this efficient/effective? Well, we're being nice to the system
4527 * by cleaning up our inodes proactively so they can be reaped
4528 * without I/O. But we are potentially leaving up to five seconds'
4529 * worth of inodes floating about which prune_icache wants us to
4530 * write out. One way to fix that would be to get prune_icache()
4531 * to do a write_super() to free up some memory. It has the desired
4532 * effect.
4533 */
4534int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4535{
4536 struct ext4_iloc iloc;
4537 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4538 static unsigned int mnt_count;
4539 int err, ret;
4540
4541 might_sleep();
4542 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4543 err = ext4_reserve_inode_write(handle, inode, &iloc);
4544 if (ext4_handle_valid(handle) &&
4545 EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4546 !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4547 /*
4548 * We need extra buffer credits since we may write into EA block
4549 * with this same handle. If journal_extend fails, then it will
4550 * only result in a minor loss of functionality for that inode.
4551 * If this is felt to be critical, then e2fsck should be run to
4552 * force a large enough s_min_extra_isize.
4553 */
4554 if ((jbd2_journal_extend(handle,
4555 EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4556 ret = ext4_expand_extra_isize(inode,
4557 sbi->s_want_extra_isize,
4558 iloc, handle);
4559 if (ret) {
4560 ext4_set_inode_state(inode,
4561 EXT4_STATE_NO_EXPAND);
4562 if (mnt_count !=
4563 le16_to_cpu(sbi->s_es->s_mnt_count)) {
4564 ext4_warning(inode->i_sb,
4565 "Unable to expand inode %lu. Delete"
4566 " some EAs or run e2fsck.",
4567 inode->i_ino);
4568 mnt_count =
4569 le16_to_cpu(sbi->s_es->s_mnt_count);
4570 }
4571 }
4572 }
4573 }
4574 if (!err)
4575 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4576 return err;
4577}
4578
4579/*
4580 * ext4_dirty_inode() is called from __mark_inode_dirty()
4581 *
4582 * We're really interested in the case where a file is being extended.
4583 * i_size has been changed by generic_commit_write() and we thus need
4584 * to include the updated inode in the current transaction.
4585 *
4586 * Also, dquot_alloc_block() will always dirty the inode when blocks
4587 * are allocated to the file.
4588 *
4589 * If the inode is marked synchronous, we don't honour that here - doing
4590 * so would cause a commit on atime updates, which we don't bother doing.
4591 * We handle synchronous inodes at the highest possible level.
4592 */
4593void ext4_dirty_inode(struct inode *inode, int flags)
4594{
4595 handle_t *handle;
4596
4597 handle = ext4_journal_start(inode, 2);
4598 if (IS_ERR(handle))
4599 goto out;
4600
4601 ext4_mark_inode_dirty(handle, inode);
4602
4603 ext4_journal_stop(handle);
4604out:
4605 return;
4606}
4607
4608#if 0
4609/*
4610 * Bind an inode's backing buffer_head into this transaction, to prevent
4611 * it from being flushed to disk early. Unlike
4612 * ext4_reserve_inode_write, this leaves behind no bh reference and
4613 * returns no iloc structure, so the caller needs to repeat the iloc
4614 * lookup to mark the inode dirty later.
4615 */
4616static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4617{
4618 struct ext4_iloc iloc;
4619
4620 int err = 0;
4621 if (handle) {
4622 err = ext4_get_inode_loc(inode, &iloc);
4623 if (!err) {
4624 BUFFER_TRACE(iloc.bh, "get_write_access");
4625 err = jbd2_journal_get_write_access(handle, iloc.bh);
4626 if (!err)
4627 err = ext4_handle_dirty_metadata(handle,
4628 NULL,
4629 iloc.bh);
4630 brelse(iloc.bh);
4631 }
4632 }
4633 ext4_std_error(inode->i_sb, err);
4634 return err;
4635}
4636#endif
4637
4638int ext4_change_inode_journal_flag(struct inode *inode, int val)
4639{
4640 journal_t *journal;
4641 handle_t *handle;
4642 int err;
4643
4644 /*
4645 * We have to be very careful here: changing a data block's
4646 * journaling status dynamically is dangerous. If we write a
4647 * data block to the journal, change the status and then delete
4648 * that block, we risk forgetting to revoke the old log record
4649 * from the journal and so a subsequent replay can corrupt data.
4650 * So, first we make sure that the journal is empty and that
4651 * nobody is changing anything.
4652 */
4653
4654 journal = EXT4_JOURNAL(inode);
4655 if (!journal)
4656 return 0;
4657 if (is_journal_aborted(journal))
4658 return -EROFS;
4659 /* We have to allocate physical blocks for delalloc blocks
4660 * before flushing journal. otherwise delalloc blocks can not
4661 * be allocated any more. even more truncate on delalloc blocks
4662 * could trigger BUG by flushing delalloc blocks in journal.
4663 * There is no delalloc block in non-journal data mode.
4664 */
4665 if (val && test_opt(inode->i_sb, DELALLOC)) {
4666 err = ext4_alloc_da_blocks(inode);
4667 if (err < 0)
4668 return err;
4669 }
4670
4671 jbd2_journal_lock_updates(journal);
4672
4673 /*
4674 * OK, there are no updates running now, and all cached data is
4675 * synced to disk. We are now in a completely consistent state
4676 * which doesn't have anything in the journal, and we know that
4677 * no filesystem updates are running, so it is safe to modify
4678 * the inode's in-core data-journaling state flag now.
4679 */
4680
4681 if (val)
4682 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4683 else {
4684 jbd2_journal_flush(journal);
4685 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4686 }
4687 ext4_set_aops(inode);
4688
4689 jbd2_journal_unlock_updates(journal);
4690
4691 /* Finally we can mark the inode as dirty. */
4692
4693 handle = ext4_journal_start(inode, 1);
4694 if (IS_ERR(handle))
4695 return PTR_ERR(handle);
4696
4697 err = ext4_mark_inode_dirty(handle, inode);
4698 ext4_handle_sync(handle);
4699 ext4_journal_stop(handle);
4700 ext4_std_error(inode->i_sb, err);
4701
4702 return err;
4703}
4704
4705static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4706{
4707 return !buffer_mapped(bh);
4708}
4709
4710int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
4711{
4712 struct page *page = vmf->page;
4713 loff_t size;
4714 unsigned long len;
4715 int ret;
4716 struct file *file = vma->vm_file;
4717 struct inode *inode = file->f_path.dentry->d_inode;
4718 struct address_space *mapping = inode->i_mapping;
4719 handle_t *handle;
4720 get_block_t *get_block;
4721 int retries = 0;
4722
4723 /*
4724 * This check is racy but catches the common case. We rely on
4725 * __block_page_mkwrite() to do a reliable check.
4726 */
4727 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
4728 /* Delalloc case is easy... */
4729 if (test_opt(inode->i_sb, DELALLOC) &&
4730 !ext4_should_journal_data(inode) &&
4731 !ext4_nonda_switch(inode->i_sb)) {
4732 do {
4733 ret = __block_page_mkwrite(vma, vmf,
4734 ext4_da_get_block_prep);
4735 } while (ret == -ENOSPC &&
4736 ext4_should_retry_alloc(inode->i_sb, &retries));
4737 goto out_ret;
4738 }
4739
4740 lock_page(page);
4741 size = i_size_read(inode);
4742 /* Page got truncated from under us? */
4743 if (page->mapping != mapping || page_offset(page) > size) {
4744 unlock_page(page);
4745 ret = VM_FAULT_NOPAGE;
4746 goto out;
4747 }
4748
4749 if (page->index == size >> PAGE_CACHE_SHIFT)
4750 len = size & ~PAGE_CACHE_MASK;
4751 else
4752 len = PAGE_CACHE_SIZE;
4753 /*
4754 * Return if we have all the buffers mapped. This avoids the need to do
4755 * journal_start/journal_stop which can block and take a long time
4756 */
4757 if (page_has_buffers(page)) {
4758 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4759 ext4_bh_unmapped)) {
4760 /* Wait so that we don't change page under IO */
4761 wait_on_page_writeback(page);
4762 ret = VM_FAULT_LOCKED;
4763 goto out;
4764 }
4765 }
4766 unlock_page(page);
4767 /* OK, we need to fill the hole... */
4768 if (ext4_should_dioread_nolock(inode))
4769 get_block = ext4_get_block_write;
4770 else
4771 get_block = ext4_get_block;
4772retry_alloc:
4773 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
4774 if (IS_ERR(handle)) {
4775 ret = VM_FAULT_SIGBUS;
4776 goto out;
4777 }
4778 ret = __block_page_mkwrite(vma, vmf, get_block);
4779 if (!ret && ext4_should_journal_data(inode)) {
4780 if (walk_page_buffers(handle, page_buffers(page), 0,
4781 PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
4782 unlock_page(page);
4783 ret = VM_FAULT_SIGBUS;
4784 ext4_journal_stop(handle);
4785 goto out;
4786 }
4787 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
4788 }
4789 ext4_journal_stop(handle);
4790 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
4791 goto retry_alloc;
4792out_ret:
4793 ret = block_page_mkwrite_return(ret);
4794out:
4795 return ret;
4796}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/inode.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/inode.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * 64-bit file support on 64-bit platforms by Jakub Jelinek
17 * (jj@sunsite.ms.mff.cuni.cz)
18 *
19 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20 */
21
22#include <linux/fs.h>
23#include <linux/time.h>
24#include <linux/highuid.h>
25#include <linux/pagemap.h>
26#include <linux/dax.h>
27#include <linux/quotaops.h>
28#include <linux/string.h>
29#include <linux/buffer_head.h>
30#include <linux/writeback.h>
31#include <linux/pagevec.h>
32#include <linux/mpage.h>
33#include <linux/namei.h>
34#include <linux/uio.h>
35#include <linux/bio.h>
36#include <linux/workqueue.h>
37#include <linux/kernel.h>
38#include <linux/printk.h>
39#include <linux/slab.h>
40#include <linux/bitops.h>
41#include <linux/iomap.h>
42#include <linux/iversion.h>
43
44#include "ext4_jbd2.h"
45#include "xattr.h"
46#include "acl.h"
47#include "truncate.h"
48
49#include <trace/events/ext4.h>
50
51#define MPAGE_DA_EXTENT_TAIL 0x01
52
53static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
54 struct ext4_inode_info *ei)
55{
56 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
57 __u32 csum;
58 __u16 dummy_csum = 0;
59 int offset = offsetof(struct ext4_inode, i_checksum_lo);
60 unsigned int csum_size = sizeof(dummy_csum);
61
62 csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
63 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
64 offset += csum_size;
65 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
66 EXT4_GOOD_OLD_INODE_SIZE - offset);
67
68 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
69 offset = offsetof(struct ext4_inode, i_checksum_hi);
70 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
71 EXT4_GOOD_OLD_INODE_SIZE,
72 offset - EXT4_GOOD_OLD_INODE_SIZE);
73 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
74 csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
75 csum_size);
76 offset += csum_size;
77 }
78 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
79 EXT4_INODE_SIZE(inode->i_sb) - offset);
80 }
81
82 return csum;
83}
84
85static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
86 struct ext4_inode_info *ei)
87{
88 __u32 provided, calculated;
89
90 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
91 cpu_to_le32(EXT4_OS_LINUX) ||
92 !ext4_has_metadata_csum(inode->i_sb))
93 return 1;
94
95 provided = le16_to_cpu(raw->i_checksum_lo);
96 calculated = ext4_inode_csum(inode, raw, ei);
97 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
98 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
99 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
100 else
101 calculated &= 0xFFFF;
102
103 return provided == calculated;
104}
105
106static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
107 struct ext4_inode_info *ei)
108{
109 __u32 csum;
110
111 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
112 cpu_to_le32(EXT4_OS_LINUX) ||
113 !ext4_has_metadata_csum(inode->i_sb))
114 return;
115
116 csum = ext4_inode_csum(inode, raw, ei);
117 raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
118 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
119 EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
120 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
121}
122
123static inline int ext4_begin_ordered_truncate(struct inode *inode,
124 loff_t new_size)
125{
126 trace_ext4_begin_ordered_truncate(inode, new_size);
127 /*
128 * If jinode is zero, then we never opened the file for
129 * writing, so there's no need to call
130 * jbd2_journal_begin_ordered_truncate() since there's no
131 * outstanding writes we need to flush.
132 */
133 if (!EXT4_I(inode)->jinode)
134 return 0;
135 return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
136 EXT4_I(inode)->jinode,
137 new_size);
138}
139
140static void ext4_invalidatepage(struct page *page, unsigned int offset,
141 unsigned int length);
142static int __ext4_journalled_writepage(struct page *page, unsigned int len);
143static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
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 * Restart the transaction associated with *handle. This does a commit,
168 * so before we call here everything must be consistently dirtied against
169 * this transaction.
170 */
171int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
172 int nblocks)
173{
174 int ret;
175
176 /*
177 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
178 * moment, get_block can be called only for blocks inside i_size since
179 * page cache has been already dropped and writes are blocked by
180 * i_mutex. So we can safely drop the i_data_sem here.
181 */
182 BUG_ON(EXT4_JOURNAL(inode) == NULL);
183 jbd_debug(2, "restarting handle %p\n", handle);
184 up_write(&EXT4_I(inode)->i_data_sem);
185 ret = ext4_journal_restart(handle, nblocks);
186 down_write(&EXT4_I(inode)->i_data_sem);
187 ext4_discard_preallocations(inode);
188
189 return ret;
190}
191
192/*
193 * Called at the last iput() if i_nlink is zero.
194 */
195void ext4_evict_inode(struct inode *inode)
196{
197 handle_t *handle;
198 int err;
199 int extra_credits = 3;
200 struct ext4_xattr_inode_array *ea_inode_array = NULL;
201
202 trace_ext4_evict_inode(inode);
203
204 if (inode->i_nlink) {
205 /*
206 * When journalling data dirty buffers are tracked only in the
207 * journal. So although mm thinks everything is clean and
208 * ready for reaping the inode might still have some pages to
209 * write in the running transaction or waiting to be
210 * checkpointed. Thus calling jbd2_journal_invalidatepage()
211 * (via truncate_inode_pages()) to discard these buffers can
212 * cause data loss. Also even if we did not discard these
213 * buffers, we would have no way to find them after the inode
214 * is reaped and thus user could see stale data if he tries to
215 * read them before the transaction is checkpointed. So be
216 * careful and force everything to disk here... We use
217 * ei->i_datasync_tid to store the newest transaction
218 * containing inode's data.
219 *
220 * Note that directories do not have this problem because they
221 * don't use page cache.
222 */
223 if (inode->i_ino != EXT4_JOURNAL_INO &&
224 ext4_should_journal_data(inode) &&
225 (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
226 inode->i_data.nrpages) {
227 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
228 tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
229
230 jbd2_complete_transaction(journal, commit_tid);
231 filemap_write_and_wait(&inode->i_data);
232 }
233 truncate_inode_pages_final(&inode->i_data);
234
235 goto no_delete;
236 }
237
238 if (is_bad_inode(inode))
239 goto no_delete;
240 dquot_initialize(inode);
241
242 if (ext4_should_order_data(inode))
243 ext4_begin_ordered_truncate(inode, 0);
244 truncate_inode_pages_final(&inode->i_data);
245
246 /*
247 * Protect us against freezing - iput() caller didn't have to have any
248 * protection against it
249 */
250 sb_start_intwrite(inode->i_sb);
251
252 if (!IS_NOQUOTA(inode))
253 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
254
255 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
256 ext4_blocks_for_truncate(inode)+extra_credits);
257 if (IS_ERR(handle)) {
258 ext4_std_error(inode->i_sb, PTR_ERR(handle));
259 /*
260 * If we're going to skip the normal cleanup, we still need to
261 * make sure that the in-core orphan linked list is properly
262 * cleaned up.
263 */
264 ext4_orphan_del(NULL, inode);
265 sb_end_intwrite(inode->i_sb);
266 goto no_delete;
267 }
268
269 if (IS_SYNC(inode))
270 ext4_handle_sync(handle);
271
272 /*
273 * Set inode->i_size to 0 before calling ext4_truncate(). We need
274 * special handling of symlinks here because i_size is used to
275 * determine whether ext4_inode_info->i_data contains symlink data or
276 * block mappings. Setting i_size to 0 will remove its fast symlink
277 * status. Erase i_data so that it becomes a valid empty block map.
278 */
279 if (ext4_inode_is_fast_symlink(inode))
280 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
281 inode->i_size = 0;
282 err = ext4_mark_inode_dirty(handle, inode);
283 if (err) {
284 ext4_warning(inode->i_sb,
285 "couldn't mark inode dirty (err %d)", err);
286 goto stop_handle;
287 }
288 if (inode->i_blocks) {
289 err = ext4_truncate(inode);
290 if (err) {
291 ext4_error(inode->i_sb,
292 "couldn't truncate inode %lu (err %d)",
293 inode->i_ino, err);
294 goto stop_handle;
295 }
296 }
297
298 /* Remove xattr references. */
299 err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
300 extra_credits);
301 if (err) {
302 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
303stop_handle:
304 ext4_journal_stop(handle);
305 ext4_orphan_del(NULL, inode);
306 sb_end_intwrite(inode->i_sb);
307 ext4_xattr_inode_array_free(ea_inode_array);
308 goto no_delete;
309 }
310
311 /*
312 * Kill off the orphan record which ext4_truncate created.
313 * AKPM: I think this can be inside the above `if'.
314 * Note that ext4_orphan_del() has to be able to cope with the
315 * deletion of a non-existent orphan - this is because we don't
316 * know if ext4_truncate() actually created an orphan record.
317 * (Well, we could do this if we need to, but heck - it works)
318 */
319 ext4_orphan_del(handle, inode);
320 EXT4_I(inode)->i_dtime = get_seconds();
321
322 /*
323 * One subtle ordering requirement: if anything has gone wrong
324 * (transaction abort, IO errors, whatever), then we can still
325 * do these next steps (the fs will already have been marked as
326 * having errors), but we can't free the inode if the mark_dirty
327 * fails.
328 */
329 if (ext4_mark_inode_dirty(handle, inode))
330 /* If that failed, just do the required in-core inode clear. */
331 ext4_clear_inode(inode);
332 else
333 ext4_free_inode(handle, inode);
334 ext4_journal_stop(handle);
335 sb_end_intwrite(inode->i_sb);
336 ext4_xattr_inode_array_free(ea_inode_array);
337 return;
338no_delete:
339 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
340}
341
342#ifdef CONFIG_QUOTA
343qsize_t *ext4_get_reserved_space(struct inode *inode)
344{
345 return &EXT4_I(inode)->i_reserved_quota;
346}
347#endif
348
349/*
350 * Called with i_data_sem down, which is important since we can call
351 * ext4_discard_preallocations() from here.
352 */
353void ext4_da_update_reserve_space(struct inode *inode,
354 int used, int quota_claim)
355{
356 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
357 struct ext4_inode_info *ei = EXT4_I(inode);
358
359 spin_lock(&ei->i_block_reservation_lock);
360 trace_ext4_da_update_reserve_space(inode, used, quota_claim);
361 if (unlikely(used > ei->i_reserved_data_blocks)) {
362 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
363 "with only %d reserved data blocks",
364 __func__, inode->i_ino, used,
365 ei->i_reserved_data_blocks);
366 WARN_ON(1);
367 used = ei->i_reserved_data_blocks;
368 }
369
370 /* Update per-inode reservations */
371 ei->i_reserved_data_blocks -= used;
372 percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
373
374 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
375
376 /* Update quota subsystem for data blocks */
377 if (quota_claim)
378 dquot_claim_block(inode, EXT4_C2B(sbi, used));
379 else {
380 /*
381 * We did fallocate with an offset that is already delayed
382 * allocated. So on delayed allocated writeback we should
383 * not re-claim the quota for fallocated blocks.
384 */
385 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
386 }
387
388 /*
389 * If we have done all the pending block allocations and if
390 * there aren't any writers on the inode, we can discard the
391 * inode's preallocations.
392 */
393 if ((ei->i_reserved_data_blocks == 0) &&
394 (atomic_read(&inode->i_writecount) == 0))
395 ext4_discard_preallocations(inode);
396}
397
398static int __check_block_validity(struct inode *inode, const char *func,
399 unsigned int line,
400 struct ext4_map_blocks *map)
401{
402 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
403 map->m_len)) {
404 ext4_error_inode(inode, func, line, map->m_pblk,
405 "lblock %lu mapped to illegal pblock "
406 "(length %d)", (unsigned long) map->m_lblk,
407 map->m_len);
408 return -EFSCORRUPTED;
409 }
410 return 0;
411}
412
413int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
414 ext4_lblk_t len)
415{
416 int ret;
417
418 if (ext4_encrypted_inode(inode))
419 return fscrypt_zeroout_range(inode, lblk, pblk, len);
420
421 ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
422 if (ret > 0)
423 ret = 0;
424
425 return ret;
426}
427
428#define check_block_validity(inode, map) \
429 __check_block_validity((inode), __func__, __LINE__, (map))
430
431#ifdef ES_AGGRESSIVE_TEST
432static void ext4_map_blocks_es_recheck(handle_t *handle,
433 struct inode *inode,
434 struct ext4_map_blocks *es_map,
435 struct ext4_map_blocks *map,
436 int flags)
437{
438 int retval;
439
440 map->m_flags = 0;
441 /*
442 * There is a race window that the result is not the same.
443 * e.g. xfstests #223 when dioread_nolock enables. The reason
444 * is that we lookup a block mapping in extent status tree with
445 * out taking i_data_sem. So at the time the unwritten extent
446 * could be converted.
447 */
448 down_read(&EXT4_I(inode)->i_data_sem);
449 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
450 retval = ext4_ext_map_blocks(handle, inode, map, flags &
451 EXT4_GET_BLOCKS_KEEP_SIZE);
452 } else {
453 retval = ext4_ind_map_blocks(handle, inode, map, flags &
454 EXT4_GET_BLOCKS_KEEP_SIZE);
455 }
456 up_read((&EXT4_I(inode)->i_data_sem));
457
458 /*
459 * We don't check m_len because extent will be collpased in status
460 * tree. So the m_len might not equal.
461 */
462 if (es_map->m_lblk != map->m_lblk ||
463 es_map->m_flags != map->m_flags ||
464 es_map->m_pblk != map->m_pblk) {
465 printk("ES cache assertion failed for inode: %lu "
466 "es_cached ex [%d/%d/%llu/%x] != "
467 "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
468 inode->i_ino, es_map->m_lblk, es_map->m_len,
469 es_map->m_pblk, es_map->m_flags, map->m_lblk,
470 map->m_len, map->m_pblk, map->m_flags,
471 retval, flags);
472 }
473}
474#endif /* ES_AGGRESSIVE_TEST */
475
476/*
477 * The ext4_map_blocks() function tries to look up the requested blocks,
478 * and returns if the blocks are already mapped.
479 *
480 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
481 * and store the allocated blocks in the result buffer head and mark it
482 * mapped.
483 *
484 * If file type is extents based, it will call ext4_ext_map_blocks(),
485 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
486 * based files
487 *
488 * On success, it returns the number of blocks being mapped or allocated. if
489 * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
490 * is marked as unwritten. If the create == 1, it will mark @map as mapped.
491 *
492 * It returns 0 if plain look up failed (blocks have not been allocated), in
493 * that case, @map is returned as unmapped but we still do fill map->m_len to
494 * indicate the length of a hole starting at map->m_lblk.
495 *
496 * It returns the error in case of allocation failure.
497 */
498int ext4_map_blocks(handle_t *handle, struct inode *inode,
499 struct ext4_map_blocks *map, int flags)
500{
501 struct extent_status es;
502 int retval;
503 int ret = 0;
504#ifdef ES_AGGRESSIVE_TEST
505 struct ext4_map_blocks orig_map;
506
507 memcpy(&orig_map, map, sizeof(*map));
508#endif
509
510 map->m_flags = 0;
511 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
512 "logical block %lu\n", inode->i_ino, flags, map->m_len,
513 (unsigned long) map->m_lblk);
514
515 /*
516 * ext4_map_blocks returns an int, and m_len is an unsigned int
517 */
518 if (unlikely(map->m_len > INT_MAX))
519 map->m_len = INT_MAX;
520
521 /* We can handle the block number less than EXT_MAX_BLOCKS */
522 if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
523 return -EFSCORRUPTED;
524
525 /* Lookup extent status tree firstly */
526 if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
527 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
528 map->m_pblk = ext4_es_pblock(&es) +
529 map->m_lblk - es.es_lblk;
530 map->m_flags |= ext4_es_is_written(&es) ?
531 EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
532 retval = es.es_len - (map->m_lblk - es.es_lblk);
533 if (retval > map->m_len)
534 retval = map->m_len;
535 map->m_len = retval;
536 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
537 map->m_pblk = 0;
538 retval = es.es_len - (map->m_lblk - es.es_lblk);
539 if (retval > map->m_len)
540 retval = map->m_len;
541 map->m_len = retval;
542 retval = 0;
543 } else {
544 BUG_ON(1);
545 }
546#ifdef ES_AGGRESSIVE_TEST
547 ext4_map_blocks_es_recheck(handle, inode, map,
548 &orig_map, flags);
549#endif
550 goto found;
551 }
552
553 /*
554 * Try to see if we can get the block without requesting a new
555 * file system block.
556 */
557 down_read(&EXT4_I(inode)->i_data_sem);
558 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
559 retval = ext4_ext_map_blocks(handle, inode, map, flags &
560 EXT4_GET_BLOCKS_KEEP_SIZE);
561 } else {
562 retval = ext4_ind_map_blocks(handle, inode, map, flags &
563 EXT4_GET_BLOCKS_KEEP_SIZE);
564 }
565 if (retval > 0) {
566 unsigned int status;
567
568 if (unlikely(retval != map->m_len)) {
569 ext4_warning(inode->i_sb,
570 "ES len assertion failed for inode "
571 "%lu: retval %d != map->m_len %d",
572 inode->i_ino, retval, map->m_len);
573 WARN_ON(1);
574 }
575
576 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
577 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
578 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
579 !(status & EXTENT_STATUS_WRITTEN) &&
580 ext4_find_delalloc_range(inode, map->m_lblk,
581 map->m_lblk + map->m_len - 1))
582 status |= EXTENT_STATUS_DELAYED;
583 ret = ext4_es_insert_extent(inode, map->m_lblk,
584 map->m_len, map->m_pblk, status);
585 if (ret < 0)
586 retval = ret;
587 }
588 up_read((&EXT4_I(inode)->i_data_sem));
589
590found:
591 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
592 ret = check_block_validity(inode, map);
593 if (ret != 0)
594 return ret;
595 }
596
597 /* If it is only a block(s) look up */
598 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
599 return retval;
600
601 /*
602 * Returns if the blocks have already allocated
603 *
604 * Note that if blocks have been preallocated
605 * ext4_ext_get_block() returns the create = 0
606 * with buffer head unmapped.
607 */
608 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
609 /*
610 * If we need to convert extent to unwritten
611 * we continue and do the actual work in
612 * ext4_ext_map_blocks()
613 */
614 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
615 return retval;
616
617 /*
618 * Here we clear m_flags because after allocating an new extent,
619 * it will be set again.
620 */
621 map->m_flags &= ~EXT4_MAP_FLAGS;
622
623 /*
624 * New blocks allocate and/or writing to unwritten extent
625 * will possibly result in updating i_data, so we take
626 * the write lock of i_data_sem, and call get_block()
627 * with create == 1 flag.
628 */
629 down_write(&EXT4_I(inode)->i_data_sem);
630
631 /*
632 * We need to check for EXT4 here because migrate
633 * could have changed the inode type in between
634 */
635 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
636 retval = ext4_ext_map_blocks(handle, inode, map, flags);
637 } else {
638 retval = ext4_ind_map_blocks(handle, inode, map, flags);
639
640 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
641 /*
642 * We allocated new blocks which will result in
643 * i_data's format changing. Force the migrate
644 * to fail by clearing migrate flags
645 */
646 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
647 }
648
649 /*
650 * Update reserved blocks/metadata blocks after successful
651 * block allocation which had been deferred till now. We don't
652 * support fallocate for non extent files. So we can update
653 * reserve space here.
654 */
655 if ((retval > 0) &&
656 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
657 ext4_da_update_reserve_space(inode, retval, 1);
658 }
659
660 if (retval > 0) {
661 unsigned int status;
662
663 if (unlikely(retval != map->m_len)) {
664 ext4_warning(inode->i_sb,
665 "ES len assertion failed for inode "
666 "%lu: retval %d != map->m_len %d",
667 inode->i_ino, retval, map->m_len);
668 WARN_ON(1);
669 }
670
671 /*
672 * We have to zeroout blocks before inserting them into extent
673 * status tree. Otherwise someone could look them up there and
674 * use them before they are really zeroed. We also have to
675 * unmap metadata before zeroing as otherwise writeback can
676 * overwrite zeros with stale data from block device.
677 */
678 if (flags & EXT4_GET_BLOCKS_ZERO &&
679 map->m_flags & EXT4_MAP_MAPPED &&
680 map->m_flags & EXT4_MAP_NEW) {
681 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
682 map->m_len);
683 ret = ext4_issue_zeroout(inode, map->m_lblk,
684 map->m_pblk, map->m_len);
685 if (ret) {
686 retval = ret;
687 goto out_sem;
688 }
689 }
690
691 /*
692 * If the extent has been zeroed out, we don't need to update
693 * extent status tree.
694 */
695 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
696 ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
697 if (ext4_es_is_written(&es))
698 goto out_sem;
699 }
700 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
701 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
702 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
703 !(status & EXTENT_STATUS_WRITTEN) &&
704 ext4_find_delalloc_range(inode, map->m_lblk,
705 map->m_lblk + map->m_len - 1))
706 status |= EXTENT_STATUS_DELAYED;
707 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
708 map->m_pblk, status);
709 if (ret < 0) {
710 retval = ret;
711 goto out_sem;
712 }
713 }
714
715out_sem:
716 up_write((&EXT4_I(inode)->i_data_sem));
717 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
718 ret = check_block_validity(inode, map);
719 if (ret != 0)
720 return ret;
721
722 /*
723 * Inodes with freshly allocated blocks where contents will be
724 * visible after transaction commit must be on transaction's
725 * ordered data list.
726 */
727 if (map->m_flags & EXT4_MAP_NEW &&
728 !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
729 !(flags & EXT4_GET_BLOCKS_ZERO) &&
730 !ext4_is_quota_file(inode) &&
731 ext4_should_order_data(inode)) {
732 if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
733 ret = ext4_jbd2_inode_add_wait(handle, inode);
734 else
735 ret = ext4_jbd2_inode_add_write(handle, inode);
736 if (ret)
737 return ret;
738 }
739 }
740 return retval;
741}
742
743/*
744 * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
745 * we have to be careful as someone else may be manipulating b_state as well.
746 */
747static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
748{
749 unsigned long old_state;
750 unsigned long new_state;
751
752 flags &= EXT4_MAP_FLAGS;
753
754 /* Dummy buffer_head? Set non-atomically. */
755 if (!bh->b_page) {
756 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
757 return;
758 }
759 /*
760 * Someone else may be modifying b_state. Be careful! This is ugly but
761 * once we get rid of using bh as a container for mapping information
762 * to pass to / from get_block functions, this can go away.
763 */
764 do {
765 old_state = READ_ONCE(bh->b_state);
766 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
767 } while (unlikely(
768 cmpxchg(&bh->b_state, old_state, new_state) != old_state));
769}
770
771static int _ext4_get_block(struct inode *inode, sector_t iblock,
772 struct buffer_head *bh, int flags)
773{
774 struct ext4_map_blocks map;
775 int ret = 0;
776
777 if (ext4_has_inline_data(inode))
778 return -ERANGE;
779
780 map.m_lblk = iblock;
781 map.m_len = bh->b_size >> inode->i_blkbits;
782
783 ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
784 flags);
785 if (ret > 0) {
786 map_bh(bh, inode->i_sb, map.m_pblk);
787 ext4_update_bh_state(bh, map.m_flags);
788 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
789 ret = 0;
790 } else if (ret == 0) {
791 /* hole case, need to fill in bh->b_size */
792 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
793 }
794 return ret;
795}
796
797int ext4_get_block(struct inode *inode, sector_t iblock,
798 struct buffer_head *bh, int create)
799{
800 return _ext4_get_block(inode, iblock, bh,
801 create ? EXT4_GET_BLOCKS_CREATE : 0);
802}
803
804/*
805 * Get block function used when preparing for buffered write if we require
806 * creating an unwritten extent if blocks haven't been allocated. The extent
807 * will be converted to written after the IO is complete.
808 */
809int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
810 struct buffer_head *bh_result, int create)
811{
812 ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
813 inode->i_ino, create);
814 return _ext4_get_block(inode, iblock, bh_result,
815 EXT4_GET_BLOCKS_IO_CREATE_EXT);
816}
817
818/* Maximum number of blocks we map for direct IO at once. */
819#define DIO_MAX_BLOCKS 4096
820
821/*
822 * Get blocks function for the cases that need to start a transaction -
823 * generally difference cases of direct IO and DAX IO. It also handles retries
824 * in case of ENOSPC.
825 */
826static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
827 struct buffer_head *bh_result, int flags)
828{
829 int dio_credits;
830 handle_t *handle;
831 int retries = 0;
832 int ret;
833
834 /* Trim mapping request to maximum we can map at once for DIO */
835 if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
836 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
837 dio_credits = ext4_chunk_trans_blocks(inode,
838 bh_result->b_size >> inode->i_blkbits);
839retry:
840 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
841 if (IS_ERR(handle))
842 return PTR_ERR(handle);
843
844 ret = _ext4_get_block(inode, iblock, bh_result, flags);
845 ext4_journal_stop(handle);
846
847 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
848 goto retry;
849 return ret;
850}
851
852/* Get block function for DIO reads and writes to inodes without extents */
853int ext4_dio_get_block(struct inode *inode, sector_t iblock,
854 struct buffer_head *bh, int create)
855{
856 /* We don't expect handle for direct IO */
857 WARN_ON_ONCE(ext4_journal_current_handle());
858
859 if (!create)
860 return _ext4_get_block(inode, iblock, bh, 0);
861 return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
862}
863
864/*
865 * Get block function for AIO DIO writes when we create unwritten extent if
866 * blocks are not allocated yet. The extent will be converted to written
867 * after IO is complete.
868 */
869static int ext4_dio_get_block_unwritten_async(struct inode *inode,
870 sector_t iblock, struct buffer_head *bh_result, int create)
871{
872 int ret;
873
874 /* We don't expect handle for direct IO */
875 WARN_ON_ONCE(ext4_journal_current_handle());
876
877 ret = ext4_get_block_trans(inode, iblock, bh_result,
878 EXT4_GET_BLOCKS_IO_CREATE_EXT);
879
880 /*
881 * When doing DIO using unwritten extents, we need io_end to convert
882 * unwritten extents to written on IO completion. We allocate io_end
883 * once we spot unwritten extent and store it in b_private. Generic
884 * DIO code keeps b_private set and furthermore passes the value to
885 * our completion callback in 'private' argument.
886 */
887 if (!ret && buffer_unwritten(bh_result)) {
888 if (!bh_result->b_private) {
889 ext4_io_end_t *io_end;
890
891 io_end = ext4_init_io_end(inode, GFP_KERNEL);
892 if (!io_end)
893 return -ENOMEM;
894 bh_result->b_private = io_end;
895 ext4_set_io_unwritten_flag(inode, io_end);
896 }
897 set_buffer_defer_completion(bh_result);
898 }
899
900 return ret;
901}
902
903/*
904 * Get block function for non-AIO DIO writes when we create unwritten extent if
905 * blocks are not allocated yet. The extent will be converted to written
906 * after IO is complete by ext4_direct_IO_write().
907 */
908static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
909 sector_t iblock, struct buffer_head *bh_result, int create)
910{
911 int ret;
912
913 /* We don't expect handle for direct IO */
914 WARN_ON_ONCE(ext4_journal_current_handle());
915
916 ret = ext4_get_block_trans(inode, iblock, bh_result,
917 EXT4_GET_BLOCKS_IO_CREATE_EXT);
918
919 /*
920 * Mark inode as having pending DIO writes to unwritten extents.
921 * ext4_direct_IO_write() checks this flag and converts extents to
922 * written.
923 */
924 if (!ret && buffer_unwritten(bh_result))
925 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
926
927 return ret;
928}
929
930static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
931 struct buffer_head *bh_result, int create)
932{
933 int ret;
934
935 ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
936 inode->i_ino, create);
937 /* We don't expect handle for direct IO */
938 WARN_ON_ONCE(ext4_journal_current_handle());
939
940 ret = _ext4_get_block(inode, iblock, bh_result, 0);
941 /*
942 * Blocks should have been preallocated! ext4_file_write_iter() checks
943 * that.
944 */
945 WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
946
947 return ret;
948}
949
950
951/*
952 * `handle' can be NULL if create is zero
953 */
954struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
955 ext4_lblk_t block, int map_flags)
956{
957 struct ext4_map_blocks map;
958 struct buffer_head *bh;
959 int create = map_flags & EXT4_GET_BLOCKS_CREATE;
960 int err;
961
962 J_ASSERT(handle != NULL || create == 0);
963
964 map.m_lblk = block;
965 map.m_len = 1;
966 err = ext4_map_blocks(handle, inode, &map, map_flags);
967
968 if (err == 0)
969 return create ? ERR_PTR(-ENOSPC) : NULL;
970 if (err < 0)
971 return ERR_PTR(err);
972
973 bh = sb_getblk(inode->i_sb, map.m_pblk);
974 if (unlikely(!bh))
975 return ERR_PTR(-ENOMEM);
976 if (map.m_flags & EXT4_MAP_NEW) {
977 J_ASSERT(create != 0);
978 J_ASSERT(handle != NULL);
979
980 /*
981 * Now that we do not always journal data, we should
982 * keep in mind whether this should always journal the
983 * new buffer as metadata. For now, regular file
984 * writes use ext4_get_block instead, so it's not a
985 * problem.
986 */
987 lock_buffer(bh);
988 BUFFER_TRACE(bh, "call get_create_access");
989 err = ext4_journal_get_create_access(handle, bh);
990 if (unlikely(err)) {
991 unlock_buffer(bh);
992 goto errout;
993 }
994 if (!buffer_uptodate(bh)) {
995 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
996 set_buffer_uptodate(bh);
997 }
998 unlock_buffer(bh);
999 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1000 err = ext4_handle_dirty_metadata(handle, inode, bh);
1001 if (unlikely(err))
1002 goto errout;
1003 } else
1004 BUFFER_TRACE(bh, "not a new buffer");
1005 return bh;
1006errout:
1007 brelse(bh);
1008 return ERR_PTR(err);
1009}
1010
1011struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1012 ext4_lblk_t block, int map_flags)
1013{
1014 struct buffer_head *bh;
1015
1016 bh = ext4_getblk(handle, inode, block, map_flags);
1017 if (IS_ERR(bh))
1018 return bh;
1019 if (!bh || buffer_uptodate(bh))
1020 return bh;
1021 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1022 wait_on_buffer(bh);
1023 if (buffer_uptodate(bh))
1024 return bh;
1025 put_bh(bh);
1026 return ERR_PTR(-EIO);
1027}
1028
1029/* Read a contiguous batch of blocks. */
1030int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
1031 bool wait, struct buffer_head **bhs)
1032{
1033 int i, err;
1034
1035 for (i = 0; i < bh_count; i++) {
1036 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
1037 if (IS_ERR(bhs[i])) {
1038 err = PTR_ERR(bhs[i]);
1039 bh_count = i;
1040 goto out_brelse;
1041 }
1042 }
1043
1044 for (i = 0; i < bh_count; i++)
1045 /* Note that NULL bhs[i] is valid because of holes. */
1046 if (bhs[i] && !buffer_uptodate(bhs[i]))
1047 ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
1048 &bhs[i]);
1049
1050 if (!wait)
1051 return 0;
1052
1053 for (i = 0; i < bh_count; i++)
1054 if (bhs[i])
1055 wait_on_buffer(bhs[i]);
1056
1057 for (i = 0; i < bh_count; i++) {
1058 if (bhs[i] && !buffer_uptodate(bhs[i])) {
1059 err = -EIO;
1060 goto out_brelse;
1061 }
1062 }
1063 return 0;
1064
1065out_brelse:
1066 for (i = 0; i < bh_count; i++) {
1067 brelse(bhs[i]);
1068 bhs[i] = NULL;
1069 }
1070 return err;
1071}
1072
1073int ext4_walk_page_buffers(handle_t *handle,
1074 struct buffer_head *head,
1075 unsigned from,
1076 unsigned to,
1077 int *partial,
1078 int (*fn)(handle_t *handle,
1079 struct buffer_head *bh))
1080{
1081 struct buffer_head *bh;
1082 unsigned block_start, block_end;
1083 unsigned blocksize = head->b_size;
1084 int err, ret = 0;
1085 struct buffer_head *next;
1086
1087 for (bh = head, block_start = 0;
1088 ret == 0 && (bh != head || !block_start);
1089 block_start = block_end, bh = next) {
1090 next = bh->b_this_page;
1091 block_end = block_start + blocksize;
1092 if (block_end <= from || block_start >= to) {
1093 if (partial && !buffer_uptodate(bh))
1094 *partial = 1;
1095 continue;
1096 }
1097 err = (*fn)(handle, bh);
1098 if (!ret)
1099 ret = err;
1100 }
1101 return ret;
1102}
1103
1104/*
1105 * To preserve ordering, it is essential that the hole instantiation and
1106 * the data write be encapsulated in a single transaction. We cannot
1107 * close off a transaction and start a new one between the ext4_get_block()
1108 * and the commit_write(). So doing the jbd2_journal_start at the start of
1109 * prepare_write() is the right place.
1110 *
1111 * Also, this function can nest inside ext4_writepage(). In that case, we
1112 * *know* that ext4_writepage() has generated enough buffer credits to do the
1113 * whole page. So we won't block on the journal in that case, which is good,
1114 * because the caller may be PF_MEMALLOC.
1115 *
1116 * By accident, ext4 can be reentered when a transaction is open via
1117 * quota file writes. If we were to commit the transaction while thus
1118 * reentered, there can be a deadlock - we would be holding a quota
1119 * lock, and the commit would never complete if another thread had a
1120 * transaction open and was blocking on the quota lock - a ranking
1121 * violation.
1122 *
1123 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1124 * will _not_ run commit under these circumstances because handle->h_ref
1125 * is elevated. We'll still have enough credits for the tiny quotafile
1126 * write.
1127 */
1128int do_journal_get_write_access(handle_t *handle,
1129 struct buffer_head *bh)
1130{
1131 int dirty = buffer_dirty(bh);
1132 int ret;
1133
1134 if (!buffer_mapped(bh) || buffer_freed(bh))
1135 return 0;
1136 /*
1137 * __block_write_begin() could have dirtied some buffers. Clean
1138 * the dirty bit as jbd2_journal_get_write_access() could complain
1139 * otherwise about fs integrity issues. Setting of the dirty bit
1140 * by __block_write_begin() isn't a real problem here as we clear
1141 * the bit before releasing a page lock and thus writeback cannot
1142 * ever write the buffer.
1143 */
1144 if (dirty)
1145 clear_buffer_dirty(bh);
1146 BUFFER_TRACE(bh, "get write access");
1147 ret = ext4_journal_get_write_access(handle, bh);
1148 if (!ret && dirty)
1149 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1150 return ret;
1151}
1152
1153#ifdef CONFIG_EXT4_FS_ENCRYPTION
1154static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1155 get_block_t *get_block)
1156{
1157 unsigned from = pos & (PAGE_SIZE - 1);
1158 unsigned to = from + len;
1159 struct inode *inode = page->mapping->host;
1160 unsigned block_start, block_end;
1161 sector_t block;
1162 int err = 0;
1163 unsigned blocksize = inode->i_sb->s_blocksize;
1164 unsigned bbits;
1165 struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1166 bool decrypt = false;
1167
1168 BUG_ON(!PageLocked(page));
1169 BUG_ON(from > PAGE_SIZE);
1170 BUG_ON(to > PAGE_SIZE);
1171 BUG_ON(from > to);
1172
1173 if (!page_has_buffers(page))
1174 create_empty_buffers(page, blocksize, 0);
1175 head = page_buffers(page);
1176 bbits = ilog2(blocksize);
1177 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1178
1179 for (bh = head, block_start = 0; bh != head || !block_start;
1180 block++, block_start = block_end, bh = bh->b_this_page) {
1181 block_end = block_start + blocksize;
1182 if (block_end <= from || block_start >= to) {
1183 if (PageUptodate(page)) {
1184 if (!buffer_uptodate(bh))
1185 set_buffer_uptodate(bh);
1186 }
1187 continue;
1188 }
1189 if (buffer_new(bh))
1190 clear_buffer_new(bh);
1191 if (!buffer_mapped(bh)) {
1192 WARN_ON(bh->b_size != blocksize);
1193 err = get_block(inode, block, bh, 1);
1194 if (err)
1195 break;
1196 if (buffer_new(bh)) {
1197 clean_bdev_bh_alias(bh);
1198 if (PageUptodate(page)) {
1199 clear_buffer_new(bh);
1200 set_buffer_uptodate(bh);
1201 mark_buffer_dirty(bh);
1202 continue;
1203 }
1204 if (block_end > to || block_start < from)
1205 zero_user_segments(page, to, block_end,
1206 block_start, from);
1207 continue;
1208 }
1209 }
1210 if (PageUptodate(page)) {
1211 if (!buffer_uptodate(bh))
1212 set_buffer_uptodate(bh);
1213 continue;
1214 }
1215 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1216 !buffer_unwritten(bh) &&
1217 (block_start < from || block_end > to)) {
1218 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1219 *wait_bh++ = bh;
1220 decrypt = ext4_encrypted_inode(inode) &&
1221 S_ISREG(inode->i_mode);
1222 }
1223 }
1224 /*
1225 * If we issued read requests, let them complete.
1226 */
1227 while (wait_bh > wait) {
1228 wait_on_buffer(*--wait_bh);
1229 if (!buffer_uptodate(*wait_bh))
1230 err = -EIO;
1231 }
1232 if (unlikely(err))
1233 page_zero_new_buffers(page, from, to);
1234 else if (decrypt)
1235 err = fscrypt_decrypt_page(page->mapping->host, page,
1236 PAGE_SIZE, 0, page->index);
1237 return err;
1238}
1239#endif
1240
1241static int ext4_write_begin(struct file *file, struct address_space *mapping,
1242 loff_t pos, unsigned len, unsigned flags,
1243 struct page **pagep, void **fsdata)
1244{
1245 struct inode *inode = mapping->host;
1246 int ret, needed_blocks;
1247 handle_t *handle;
1248 int retries = 0;
1249 struct page *page;
1250 pgoff_t index;
1251 unsigned from, to;
1252
1253 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1254 return -EIO;
1255
1256 trace_ext4_write_begin(inode, pos, len, flags);
1257 /*
1258 * Reserve one block more for addition to orphan list in case
1259 * we allocate blocks but write fails for some reason
1260 */
1261 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1262 index = pos >> PAGE_SHIFT;
1263 from = pos & (PAGE_SIZE - 1);
1264 to = from + len;
1265
1266 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1267 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1268 flags, pagep);
1269 if (ret < 0)
1270 return ret;
1271 if (ret == 1)
1272 return 0;
1273 }
1274
1275 /*
1276 * grab_cache_page_write_begin() can take a long time if the
1277 * system is thrashing due to memory pressure, or if the page
1278 * is being written back. So grab it first before we start
1279 * the transaction handle. This also allows us to allocate
1280 * the page (if needed) without using GFP_NOFS.
1281 */
1282retry_grab:
1283 page = grab_cache_page_write_begin(mapping, index, flags);
1284 if (!page)
1285 return -ENOMEM;
1286 unlock_page(page);
1287
1288retry_journal:
1289 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1290 if (IS_ERR(handle)) {
1291 put_page(page);
1292 return PTR_ERR(handle);
1293 }
1294
1295 lock_page(page);
1296 if (page->mapping != mapping) {
1297 /* The page got truncated from under us */
1298 unlock_page(page);
1299 put_page(page);
1300 ext4_journal_stop(handle);
1301 goto retry_grab;
1302 }
1303 /* In case writeback began while the page was unlocked */
1304 wait_for_stable_page(page);
1305
1306#ifdef CONFIG_EXT4_FS_ENCRYPTION
1307 if (ext4_should_dioread_nolock(inode))
1308 ret = ext4_block_write_begin(page, pos, len,
1309 ext4_get_block_unwritten);
1310 else
1311 ret = ext4_block_write_begin(page, pos, len,
1312 ext4_get_block);
1313#else
1314 if (ext4_should_dioread_nolock(inode))
1315 ret = __block_write_begin(page, pos, len,
1316 ext4_get_block_unwritten);
1317 else
1318 ret = __block_write_begin(page, pos, len, ext4_get_block);
1319#endif
1320 if (!ret && ext4_should_journal_data(inode)) {
1321 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1322 from, to, NULL,
1323 do_journal_get_write_access);
1324 }
1325
1326 if (ret) {
1327 unlock_page(page);
1328 /*
1329 * __block_write_begin may have instantiated a few blocks
1330 * outside i_size. Trim these off again. Don't need
1331 * i_size_read because we hold i_mutex.
1332 *
1333 * Add inode to orphan list in case we crash before
1334 * truncate finishes
1335 */
1336 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1337 ext4_orphan_add(handle, inode);
1338
1339 ext4_journal_stop(handle);
1340 if (pos + len > inode->i_size) {
1341 ext4_truncate_failed_write(inode);
1342 /*
1343 * If truncate failed early the inode might
1344 * still be on the orphan list; we need to
1345 * make sure the inode is removed from the
1346 * orphan list in that case.
1347 */
1348 if (inode->i_nlink)
1349 ext4_orphan_del(NULL, inode);
1350 }
1351
1352 if (ret == -ENOSPC &&
1353 ext4_should_retry_alloc(inode->i_sb, &retries))
1354 goto retry_journal;
1355 put_page(page);
1356 return ret;
1357 }
1358 *pagep = page;
1359 return ret;
1360}
1361
1362/* For write_end() in data=journal mode */
1363static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1364{
1365 int ret;
1366 if (!buffer_mapped(bh) || buffer_freed(bh))
1367 return 0;
1368 set_buffer_uptodate(bh);
1369 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1370 clear_buffer_meta(bh);
1371 clear_buffer_prio(bh);
1372 return ret;
1373}
1374
1375/*
1376 * We need to pick up the new inode size which generic_commit_write gave us
1377 * `file' can be NULL - eg, when called from page_symlink().
1378 *
1379 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1380 * buffers are managed internally.
1381 */
1382static int ext4_write_end(struct file *file,
1383 struct address_space *mapping,
1384 loff_t pos, unsigned len, unsigned copied,
1385 struct page *page, void *fsdata)
1386{
1387 handle_t *handle = ext4_journal_current_handle();
1388 struct inode *inode = mapping->host;
1389 loff_t old_size = inode->i_size;
1390 int ret = 0, ret2;
1391 int i_size_changed = 0;
1392
1393 trace_ext4_write_end(inode, pos, len, copied);
1394 if (ext4_has_inline_data(inode)) {
1395 ret = ext4_write_inline_data_end(inode, pos, len,
1396 copied, page);
1397 if (ret < 0) {
1398 unlock_page(page);
1399 put_page(page);
1400 goto errout;
1401 }
1402 copied = ret;
1403 } else
1404 copied = block_write_end(file, mapping, pos,
1405 len, copied, page, fsdata);
1406 /*
1407 * it's important to update i_size while still holding page lock:
1408 * page writeout could otherwise come in and zero beyond i_size.
1409 */
1410 i_size_changed = ext4_update_inode_size(inode, pos + copied);
1411 unlock_page(page);
1412 put_page(page);
1413
1414 if (old_size < pos)
1415 pagecache_isize_extended(inode, old_size, pos);
1416 /*
1417 * Don't mark the inode dirty under page lock. First, it unnecessarily
1418 * makes the holding time of page lock longer. Second, it forces lock
1419 * ordering of page lock and transaction start for journaling
1420 * filesystems.
1421 */
1422 if (i_size_changed)
1423 ext4_mark_inode_dirty(handle, inode);
1424
1425 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1426 /* if we have allocated more blocks and copied
1427 * less. We will have blocks allocated outside
1428 * inode->i_size. So truncate them
1429 */
1430 ext4_orphan_add(handle, inode);
1431errout:
1432 ret2 = ext4_journal_stop(handle);
1433 if (!ret)
1434 ret = ret2;
1435
1436 if (pos + len > inode->i_size) {
1437 ext4_truncate_failed_write(inode);
1438 /*
1439 * If truncate failed early the inode might still be
1440 * on the orphan list; we need to make sure the inode
1441 * is removed from the orphan list in that case.
1442 */
1443 if (inode->i_nlink)
1444 ext4_orphan_del(NULL, inode);
1445 }
1446
1447 return ret ? ret : copied;
1448}
1449
1450/*
1451 * This is a private version of page_zero_new_buffers() which doesn't
1452 * set the buffer to be dirty, since in data=journalled mode we need
1453 * to call ext4_handle_dirty_metadata() instead.
1454 */
1455static void ext4_journalled_zero_new_buffers(handle_t *handle,
1456 struct page *page,
1457 unsigned from, unsigned to)
1458{
1459 unsigned int block_start = 0, block_end;
1460 struct buffer_head *head, *bh;
1461
1462 bh = head = page_buffers(page);
1463 do {
1464 block_end = block_start + bh->b_size;
1465 if (buffer_new(bh)) {
1466 if (block_end > from && block_start < to) {
1467 if (!PageUptodate(page)) {
1468 unsigned start, size;
1469
1470 start = max(from, block_start);
1471 size = min(to, block_end) - start;
1472
1473 zero_user(page, start, size);
1474 write_end_fn(handle, bh);
1475 }
1476 clear_buffer_new(bh);
1477 }
1478 }
1479 block_start = block_end;
1480 bh = bh->b_this_page;
1481 } while (bh != head);
1482}
1483
1484static int ext4_journalled_write_end(struct file *file,
1485 struct address_space *mapping,
1486 loff_t pos, unsigned len, unsigned copied,
1487 struct page *page, void *fsdata)
1488{
1489 handle_t *handle = ext4_journal_current_handle();
1490 struct inode *inode = mapping->host;
1491 loff_t old_size = inode->i_size;
1492 int ret = 0, ret2;
1493 int partial = 0;
1494 unsigned from, to;
1495 int size_changed = 0;
1496
1497 trace_ext4_journalled_write_end(inode, pos, len, copied);
1498 from = pos & (PAGE_SIZE - 1);
1499 to = from + len;
1500
1501 BUG_ON(!ext4_handle_valid(handle));
1502
1503 if (ext4_has_inline_data(inode)) {
1504 ret = ext4_write_inline_data_end(inode, pos, len,
1505 copied, page);
1506 if (ret < 0) {
1507 unlock_page(page);
1508 put_page(page);
1509 goto errout;
1510 }
1511 copied = ret;
1512 } else if (unlikely(copied < len) && !PageUptodate(page)) {
1513 copied = 0;
1514 ext4_journalled_zero_new_buffers(handle, page, from, to);
1515 } else {
1516 if (unlikely(copied < len))
1517 ext4_journalled_zero_new_buffers(handle, page,
1518 from + copied, to);
1519 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1520 from + copied, &partial,
1521 write_end_fn);
1522 if (!partial)
1523 SetPageUptodate(page);
1524 }
1525 size_changed = ext4_update_inode_size(inode, pos + copied);
1526 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1527 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1528 unlock_page(page);
1529 put_page(page);
1530
1531 if (old_size < pos)
1532 pagecache_isize_extended(inode, old_size, pos);
1533
1534 if (size_changed) {
1535 ret2 = ext4_mark_inode_dirty(handle, inode);
1536 if (!ret)
1537 ret = ret2;
1538 }
1539
1540 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1541 /* if we have allocated more blocks and copied
1542 * less. We will have blocks allocated outside
1543 * inode->i_size. So truncate them
1544 */
1545 ext4_orphan_add(handle, inode);
1546
1547errout:
1548 ret2 = ext4_journal_stop(handle);
1549 if (!ret)
1550 ret = ret2;
1551 if (pos + len > inode->i_size) {
1552 ext4_truncate_failed_write(inode);
1553 /*
1554 * If truncate failed early the inode might still be
1555 * on the orphan list; we need to make sure the inode
1556 * is removed from the orphan list in that case.
1557 */
1558 if (inode->i_nlink)
1559 ext4_orphan_del(NULL, inode);
1560 }
1561
1562 return ret ? ret : copied;
1563}
1564
1565/*
1566 * Reserve space for a single cluster
1567 */
1568static int ext4_da_reserve_space(struct inode *inode)
1569{
1570 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1571 struct ext4_inode_info *ei = EXT4_I(inode);
1572 int ret;
1573
1574 /*
1575 * We will charge metadata quota at writeout time; this saves
1576 * us from metadata over-estimation, though we may go over by
1577 * a small amount in the end. Here we just reserve for data.
1578 */
1579 ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1580 if (ret)
1581 return ret;
1582
1583 spin_lock(&ei->i_block_reservation_lock);
1584 if (ext4_claim_free_clusters(sbi, 1, 0)) {
1585 spin_unlock(&ei->i_block_reservation_lock);
1586 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1587 return -ENOSPC;
1588 }
1589 ei->i_reserved_data_blocks++;
1590 trace_ext4_da_reserve_space(inode);
1591 spin_unlock(&ei->i_block_reservation_lock);
1592
1593 return 0; /* success */
1594}
1595
1596static void ext4_da_release_space(struct inode *inode, int to_free)
1597{
1598 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1599 struct ext4_inode_info *ei = EXT4_I(inode);
1600
1601 if (!to_free)
1602 return; /* Nothing to release, exit */
1603
1604 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1605
1606 trace_ext4_da_release_space(inode, to_free);
1607 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1608 /*
1609 * if there aren't enough reserved blocks, then the
1610 * counter is messed up somewhere. Since this
1611 * function is called from invalidate page, it's
1612 * harmless to return without any action.
1613 */
1614 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1615 "ino %lu, to_free %d with only %d reserved "
1616 "data blocks", inode->i_ino, to_free,
1617 ei->i_reserved_data_blocks);
1618 WARN_ON(1);
1619 to_free = ei->i_reserved_data_blocks;
1620 }
1621 ei->i_reserved_data_blocks -= to_free;
1622
1623 /* update fs dirty data blocks counter */
1624 percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1625
1626 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1627
1628 dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1629}
1630
1631static void ext4_da_page_release_reservation(struct page *page,
1632 unsigned int offset,
1633 unsigned int length)
1634{
1635 int to_release = 0, contiguous_blks = 0;
1636 struct buffer_head *head, *bh;
1637 unsigned int curr_off = 0;
1638 struct inode *inode = page->mapping->host;
1639 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1640 unsigned int stop = offset + length;
1641 int num_clusters;
1642 ext4_fsblk_t lblk;
1643
1644 BUG_ON(stop > PAGE_SIZE || stop < length);
1645
1646 head = page_buffers(page);
1647 bh = head;
1648 do {
1649 unsigned int next_off = curr_off + bh->b_size;
1650
1651 if (next_off > stop)
1652 break;
1653
1654 if ((offset <= curr_off) && (buffer_delay(bh))) {
1655 to_release++;
1656 contiguous_blks++;
1657 clear_buffer_delay(bh);
1658 } else if (contiguous_blks) {
1659 lblk = page->index <<
1660 (PAGE_SHIFT - inode->i_blkbits);
1661 lblk += (curr_off >> inode->i_blkbits) -
1662 contiguous_blks;
1663 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1664 contiguous_blks = 0;
1665 }
1666 curr_off = next_off;
1667 } while ((bh = bh->b_this_page) != head);
1668
1669 if (contiguous_blks) {
1670 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1671 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1672 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1673 }
1674
1675 /* If we have released all the blocks belonging to a cluster, then we
1676 * need to release the reserved space for that cluster. */
1677 num_clusters = EXT4_NUM_B2C(sbi, to_release);
1678 while (num_clusters > 0) {
1679 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1680 ((num_clusters - 1) << sbi->s_cluster_bits);
1681 if (sbi->s_cluster_ratio == 1 ||
1682 !ext4_find_delalloc_cluster(inode, lblk))
1683 ext4_da_release_space(inode, 1);
1684
1685 num_clusters--;
1686 }
1687}
1688
1689/*
1690 * Delayed allocation stuff
1691 */
1692
1693struct mpage_da_data {
1694 struct inode *inode;
1695 struct writeback_control *wbc;
1696
1697 pgoff_t first_page; /* The first page to write */
1698 pgoff_t next_page; /* Current page to examine */
1699 pgoff_t last_page; /* Last page to examine */
1700 /*
1701 * Extent to map - this can be after first_page because that can be
1702 * fully mapped. We somewhat abuse m_flags to store whether the extent
1703 * is delalloc or unwritten.
1704 */
1705 struct ext4_map_blocks map;
1706 struct ext4_io_submit io_submit; /* IO submission data */
1707 unsigned int do_map:1;
1708};
1709
1710static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1711 bool invalidate)
1712{
1713 int nr_pages, i;
1714 pgoff_t index, end;
1715 struct pagevec pvec;
1716 struct inode *inode = mpd->inode;
1717 struct address_space *mapping = inode->i_mapping;
1718
1719 /* This is necessary when next_page == 0. */
1720 if (mpd->first_page >= mpd->next_page)
1721 return;
1722
1723 index = mpd->first_page;
1724 end = mpd->next_page - 1;
1725 if (invalidate) {
1726 ext4_lblk_t start, last;
1727 start = index << (PAGE_SHIFT - inode->i_blkbits);
1728 last = end << (PAGE_SHIFT - inode->i_blkbits);
1729 ext4_es_remove_extent(inode, start, last - start + 1);
1730 }
1731
1732 pagevec_init(&pvec);
1733 while (index <= end) {
1734 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1735 if (nr_pages == 0)
1736 break;
1737 for (i = 0; i < nr_pages; i++) {
1738 struct page *page = pvec.pages[i];
1739
1740 BUG_ON(!PageLocked(page));
1741 BUG_ON(PageWriteback(page));
1742 if (invalidate) {
1743 if (page_mapped(page))
1744 clear_page_dirty_for_io(page);
1745 block_invalidatepage(page, 0, PAGE_SIZE);
1746 ClearPageUptodate(page);
1747 }
1748 unlock_page(page);
1749 }
1750 pagevec_release(&pvec);
1751 }
1752}
1753
1754static void ext4_print_free_blocks(struct inode *inode)
1755{
1756 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1757 struct super_block *sb = inode->i_sb;
1758 struct ext4_inode_info *ei = EXT4_I(inode);
1759
1760 ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1761 EXT4_C2B(EXT4_SB(inode->i_sb),
1762 ext4_count_free_clusters(sb)));
1763 ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1764 ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1765 (long long) EXT4_C2B(EXT4_SB(sb),
1766 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1767 ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1768 (long long) EXT4_C2B(EXT4_SB(sb),
1769 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1770 ext4_msg(sb, KERN_CRIT, "Block reservation details");
1771 ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1772 ei->i_reserved_data_blocks);
1773 return;
1774}
1775
1776static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1777{
1778 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1779}
1780
1781/*
1782 * This function is grabs code from the very beginning of
1783 * ext4_map_blocks, but assumes that the caller is from delayed write
1784 * time. This function looks up the requested blocks and sets the
1785 * buffer delay bit under the protection of i_data_sem.
1786 */
1787static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1788 struct ext4_map_blocks *map,
1789 struct buffer_head *bh)
1790{
1791 struct extent_status es;
1792 int retval;
1793 sector_t invalid_block = ~((sector_t) 0xffff);
1794#ifdef ES_AGGRESSIVE_TEST
1795 struct ext4_map_blocks orig_map;
1796
1797 memcpy(&orig_map, map, sizeof(*map));
1798#endif
1799
1800 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1801 invalid_block = ~0;
1802
1803 map->m_flags = 0;
1804 ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1805 "logical block %lu\n", inode->i_ino, map->m_len,
1806 (unsigned long) map->m_lblk);
1807
1808 /* Lookup extent status tree firstly */
1809 if (ext4_es_lookup_extent(inode, iblock, &es)) {
1810 if (ext4_es_is_hole(&es)) {
1811 retval = 0;
1812 down_read(&EXT4_I(inode)->i_data_sem);
1813 goto add_delayed;
1814 }
1815
1816 /*
1817 * Delayed extent could be allocated by fallocate.
1818 * So we need to check it.
1819 */
1820 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1821 map_bh(bh, inode->i_sb, invalid_block);
1822 set_buffer_new(bh);
1823 set_buffer_delay(bh);
1824 return 0;
1825 }
1826
1827 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1828 retval = es.es_len - (iblock - es.es_lblk);
1829 if (retval > map->m_len)
1830 retval = map->m_len;
1831 map->m_len = retval;
1832 if (ext4_es_is_written(&es))
1833 map->m_flags |= EXT4_MAP_MAPPED;
1834 else if (ext4_es_is_unwritten(&es))
1835 map->m_flags |= EXT4_MAP_UNWRITTEN;
1836 else
1837 BUG_ON(1);
1838
1839#ifdef ES_AGGRESSIVE_TEST
1840 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1841#endif
1842 return retval;
1843 }
1844
1845 /*
1846 * Try to see if we can get the block without requesting a new
1847 * file system block.
1848 */
1849 down_read(&EXT4_I(inode)->i_data_sem);
1850 if (ext4_has_inline_data(inode))
1851 retval = 0;
1852 else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1853 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1854 else
1855 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1856
1857add_delayed:
1858 if (retval == 0) {
1859 int ret;
1860 /*
1861 * XXX: __block_prepare_write() unmaps passed block,
1862 * is it OK?
1863 */
1864 /*
1865 * If the block was allocated from previously allocated cluster,
1866 * then we don't need to reserve it again. However we still need
1867 * to reserve metadata for every block we're going to write.
1868 */
1869 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1870 !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1871 ret = ext4_da_reserve_space(inode);
1872 if (ret) {
1873 /* not enough space to reserve */
1874 retval = ret;
1875 goto out_unlock;
1876 }
1877 }
1878
1879 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1880 ~0, EXTENT_STATUS_DELAYED);
1881 if (ret) {
1882 retval = ret;
1883 goto out_unlock;
1884 }
1885
1886 map_bh(bh, inode->i_sb, invalid_block);
1887 set_buffer_new(bh);
1888 set_buffer_delay(bh);
1889 } else if (retval > 0) {
1890 int ret;
1891 unsigned int status;
1892
1893 if (unlikely(retval != map->m_len)) {
1894 ext4_warning(inode->i_sb,
1895 "ES len assertion failed for inode "
1896 "%lu: retval %d != map->m_len %d",
1897 inode->i_ino, retval, map->m_len);
1898 WARN_ON(1);
1899 }
1900
1901 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1902 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1903 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1904 map->m_pblk, status);
1905 if (ret != 0)
1906 retval = ret;
1907 }
1908
1909out_unlock:
1910 up_read((&EXT4_I(inode)->i_data_sem));
1911
1912 return retval;
1913}
1914
1915/*
1916 * This is a special get_block_t callback which is used by
1917 * ext4_da_write_begin(). It will either return mapped block or
1918 * reserve space for a single block.
1919 *
1920 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1921 * We also have b_blocknr = -1 and b_bdev initialized properly
1922 *
1923 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1924 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1925 * initialized properly.
1926 */
1927int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1928 struct buffer_head *bh, int create)
1929{
1930 struct ext4_map_blocks map;
1931 int ret = 0;
1932
1933 BUG_ON(create == 0);
1934 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1935
1936 map.m_lblk = iblock;
1937 map.m_len = 1;
1938
1939 /*
1940 * first, we need to know whether the block is allocated already
1941 * preallocated blocks are unmapped but should treated
1942 * the same as allocated blocks.
1943 */
1944 ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1945 if (ret <= 0)
1946 return ret;
1947
1948 map_bh(bh, inode->i_sb, map.m_pblk);
1949 ext4_update_bh_state(bh, map.m_flags);
1950
1951 if (buffer_unwritten(bh)) {
1952 /* A delayed write to unwritten bh should be marked
1953 * new and mapped. Mapped ensures that we don't do
1954 * get_block multiple times when we write to the same
1955 * offset and new ensures that we do proper zero out
1956 * for partial write.
1957 */
1958 set_buffer_new(bh);
1959 set_buffer_mapped(bh);
1960 }
1961 return 0;
1962}
1963
1964static int bget_one(handle_t *handle, struct buffer_head *bh)
1965{
1966 get_bh(bh);
1967 return 0;
1968}
1969
1970static int bput_one(handle_t *handle, struct buffer_head *bh)
1971{
1972 put_bh(bh);
1973 return 0;
1974}
1975
1976static int __ext4_journalled_writepage(struct page *page,
1977 unsigned int len)
1978{
1979 struct address_space *mapping = page->mapping;
1980 struct inode *inode = mapping->host;
1981 struct buffer_head *page_bufs = NULL;
1982 handle_t *handle = NULL;
1983 int ret = 0, err = 0;
1984 int inline_data = ext4_has_inline_data(inode);
1985 struct buffer_head *inode_bh = NULL;
1986
1987 ClearPageChecked(page);
1988
1989 if (inline_data) {
1990 BUG_ON(page->index != 0);
1991 BUG_ON(len > ext4_get_max_inline_size(inode));
1992 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1993 if (inode_bh == NULL)
1994 goto out;
1995 } else {
1996 page_bufs = page_buffers(page);
1997 if (!page_bufs) {
1998 BUG();
1999 goto out;
2000 }
2001 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2002 NULL, bget_one);
2003 }
2004 /*
2005 * We need to release the page lock before we start the
2006 * journal, so grab a reference so the page won't disappear
2007 * out from under us.
2008 */
2009 get_page(page);
2010 unlock_page(page);
2011
2012 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2013 ext4_writepage_trans_blocks(inode));
2014 if (IS_ERR(handle)) {
2015 ret = PTR_ERR(handle);
2016 put_page(page);
2017 goto out_no_pagelock;
2018 }
2019 BUG_ON(!ext4_handle_valid(handle));
2020
2021 lock_page(page);
2022 put_page(page);
2023 if (page->mapping != mapping) {
2024 /* The page got truncated from under us */
2025 ext4_journal_stop(handle);
2026 ret = 0;
2027 goto out;
2028 }
2029
2030 if (inline_data) {
2031 BUFFER_TRACE(inode_bh, "get write access");
2032 ret = ext4_journal_get_write_access(handle, inode_bh);
2033
2034 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
2035
2036 } else {
2037 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2038 do_journal_get_write_access);
2039
2040 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2041 write_end_fn);
2042 }
2043 if (ret == 0)
2044 ret = err;
2045 EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2046 err = ext4_journal_stop(handle);
2047 if (!ret)
2048 ret = err;
2049
2050 if (!ext4_has_inline_data(inode))
2051 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2052 NULL, bput_one);
2053 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2054out:
2055 unlock_page(page);
2056out_no_pagelock:
2057 brelse(inode_bh);
2058 return ret;
2059}
2060
2061/*
2062 * Note that we don't need to start a transaction unless we're journaling data
2063 * because we should have holes filled from ext4_page_mkwrite(). We even don't
2064 * need to file the inode to the transaction's list in ordered mode because if
2065 * we are writing back data added by write(), the inode is already there and if
2066 * we are writing back data modified via mmap(), no one guarantees in which
2067 * transaction the data will hit the disk. In case we are journaling data, we
2068 * cannot start transaction directly because transaction start ranks above page
2069 * lock so we have to do some magic.
2070 *
2071 * This function can get called via...
2072 * - ext4_writepages after taking page lock (have journal handle)
2073 * - journal_submit_inode_data_buffers (no journal handle)
2074 * - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2075 * - grab_page_cache when doing write_begin (have journal handle)
2076 *
2077 * We don't do any block allocation in this function. If we have page with
2078 * multiple blocks we need to write those buffer_heads that are mapped. This
2079 * is important for mmaped based write. So if we do with blocksize 1K
2080 * truncate(f, 1024);
2081 * a = mmap(f, 0, 4096);
2082 * a[0] = 'a';
2083 * truncate(f, 4096);
2084 * we have in the page first buffer_head mapped via page_mkwrite call back
2085 * but other buffer_heads would be unmapped but dirty (dirty done via the
2086 * do_wp_page). So writepage should write the first block. If we modify
2087 * the mmap area beyond 1024 we will again get a page_fault and the
2088 * page_mkwrite callback will do the block allocation and mark the
2089 * buffer_heads mapped.
2090 *
2091 * We redirty the page if we have any buffer_heads that is either delay or
2092 * unwritten in the page.
2093 *
2094 * We can get recursively called as show below.
2095 *
2096 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2097 * ext4_writepage()
2098 *
2099 * But since we don't do any block allocation we should not deadlock.
2100 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2101 */
2102static int ext4_writepage(struct page *page,
2103 struct writeback_control *wbc)
2104{
2105 int ret = 0;
2106 loff_t size;
2107 unsigned int len;
2108 struct buffer_head *page_bufs = NULL;
2109 struct inode *inode = page->mapping->host;
2110 struct ext4_io_submit io_submit;
2111 bool keep_towrite = false;
2112
2113 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2114 ext4_invalidatepage(page, 0, PAGE_SIZE);
2115 unlock_page(page);
2116 return -EIO;
2117 }
2118
2119 trace_ext4_writepage(page);
2120 size = i_size_read(inode);
2121 if (page->index == size >> PAGE_SHIFT)
2122 len = size & ~PAGE_MASK;
2123 else
2124 len = PAGE_SIZE;
2125
2126 page_bufs = page_buffers(page);
2127 /*
2128 * We cannot do block allocation or other extent handling in this
2129 * function. If there are buffers needing that, we have to redirty
2130 * the page. But we may reach here when we do a journal commit via
2131 * journal_submit_inode_data_buffers() and in that case we must write
2132 * allocated buffers to achieve data=ordered mode guarantees.
2133 *
2134 * Also, if there is only one buffer per page (the fs block
2135 * size == the page size), if one buffer needs block
2136 * allocation or needs to modify the extent tree to clear the
2137 * unwritten flag, we know that the page can't be written at
2138 * all, so we might as well refuse the write immediately.
2139 * Unfortunately if the block size != page size, we can't as
2140 * easily detect this case using ext4_walk_page_buffers(), but
2141 * for the extremely common case, this is an optimization that
2142 * skips a useless round trip through ext4_bio_write_page().
2143 */
2144 if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2145 ext4_bh_delay_or_unwritten)) {
2146 redirty_page_for_writepage(wbc, page);
2147 if ((current->flags & PF_MEMALLOC) ||
2148 (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2149 /*
2150 * For memory cleaning there's no point in writing only
2151 * some buffers. So just bail out. Warn if we came here
2152 * from direct reclaim.
2153 */
2154 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2155 == PF_MEMALLOC);
2156 unlock_page(page);
2157 return 0;
2158 }
2159 keep_towrite = true;
2160 }
2161
2162 if (PageChecked(page) && ext4_should_journal_data(inode))
2163 /*
2164 * It's mmapped pagecache. Add buffers and journal it. There
2165 * doesn't seem much point in redirtying the page here.
2166 */
2167 return __ext4_journalled_writepage(page, len);
2168
2169 ext4_io_submit_init(&io_submit, wbc);
2170 io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2171 if (!io_submit.io_end) {
2172 redirty_page_for_writepage(wbc, page);
2173 unlock_page(page);
2174 return -ENOMEM;
2175 }
2176 ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2177 ext4_io_submit(&io_submit);
2178 /* Drop io_end reference we got from init */
2179 ext4_put_io_end_defer(io_submit.io_end);
2180 return ret;
2181}
2182
2183static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2184{
2185 int len;
2186 loff_t size;
2187 int err;
2188
2189 BUG_ON(page->index != mpd->first_page);
2190 clear_page_dirty_for_io(page);
2191 /*
2192 * We have to be very careful here! Nothing protects writeback path
2193 * against i_size changes and the page can be writeably mapped into
2194 * page tables. So an application can be growing i_size and writing
2195 * data through mmap while writeback runs. clear_page_dirty_for_io()
2196 * write-protects our page in page tables and the page cannot get
2197 * written to again until we release page lock. So only after
2198 * clear_page_dirty_for_io() we are safe to sample i_size for
2199 * ext4_bio_write_page() to zero-out tail of the written page. We rely
2200 * on the barrier provided by TestClearPageDirty in
2201 * clear_page_dirty_for_io() to make sure i_size is really sampled only
2202 * after page tables are updated.
2203 */
2204 size = i_size_read(mpd->inode);
2205 if (page->index == size >> PAGE_SHIFT)
2206 len = size & ~PAGE_MASK;
2207 else
2208 len = PAGE_SIZE;
2209 err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2210 if (!err)
2211 mpd->wbc->nr_to_write--;
2212 mpd->first_page++;
2213
2214 return err;
2215}
2216
2217#define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2218
2219/*
2220 * mballoc gives us at most this number of blocks...
2221 * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2222 * The rest of mballoc seems to handle chunks up to full group size.
2223 */
2224#define MAX_WRITEPAGES_EXTENT_LEN 2048
2225
2226/*
2227 * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2228 *
2229 * @mpd - extent of blocks
2230 * @lblk - logical number of the block in the file
2231 * @bh - buffer head we want to add to the extent
2232 *
2233 * The function is used to collect contig. blocks in the same state. If the
2234 * buffer doesn't require mapping for writeback and we haven't started the
2235 * extent of buffers to map yet, the function returns 'true' immediately - the
2236 * caller can write the buffer right away. Otherwise the function returns true
2237 * if the block has been added to the extent, false if the block couldn't be
2238 * added.
2239 */
2240static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2241 struct buffer_head *bh)
2242{
2243 struct ext4_map_blocks *map = &mpd->map;
2244
2245 /* Buffer that doesn't need mapping for writeback? */
2246 if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2247 (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2248 /* So far no extent to map => we write the buffer right away */
2249 if (map->m_len == 0)
2250 return true;
2251 return false;
2252 }
2253
2254 /* First block in the extent? */
2255 if (map->m_len == 0) {
2256 /* We cannot map unless handle is started... */
2257 if (!mpd->do_map)
2258 return false;
2259 map->m_lblk = lblk;
2260 map->m_len = 1;
2261 map->m_flags = bh->b_state & BH_FLAGS;
2262 return true;
2263 }
2264
2265 /* Don't go larger than mballoc is willing to allocate */
2266 if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2267 return false;
2268
2269 /* Can we merge the block to our big extent? */
2270 if (lblk == map->m_lblk + map->m_len &&
2271 (bh->b_state & BH_FLAGS) == map->m_flags) {
2272 map->m_len++;
2273 return true;
2274 }
2275 return false;
2276}
2277
2278/*
2279 * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2280 *
2281 * @mpd - extent of blocks for mapping
2282 * @head - the first buffer in the page
2283 * @bh - buffer we should start processing from
2284 * @lblk - logical number of the block in the file corresponding to @bh
2285 *
2286 * Walk through page buffers from @bh upto @head (exclusive) and either submit
2287 * the page for IO if all buffers in this page were mapped and there's no
2288 * accumulated extent of buffers to map or add buffers in the page to the
2289 * extent of buffers to map. The function returns 1 if the caller can continue
2290 * by processing the next page, 0 if it should stop adding buffers to the
2291 * extent to map because we cannot extend it anymore. It can also return value
2292 * < 0 in case of error during IO submission.
2293 */
2294static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2295 struct buffer_head *head,
2296 struct buffer_head *bh,
2297 ext4_lblk_t lblk)
2298{
2299 struct inode *inode = mpd->inode;
2300 int err;
2301 ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2302 >> inode->i_blkbits;
2303
2304 do {
2305 BUG_ON(buffer_locked(bh));
2306
2307 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2308 /* Found extent to map? */
2309 if (mpd->map.m_len)
2310 return 0;
2311 /* Buffer needs mapping and handle is not started? */
2312 if (!mpd->do_map)
2313 return 0;
2314 /* Everything mapped so far and we hit EOF */
2315 break;
2316 }
2317 } while (lblk++, (bh = bh->b_this_page) != head);
2318 /* So far everything mapped? Submit the page for IO. */
2319 if (mpd->map.m_len == 0) {
2320 err = mpage_submit_page(mpd, head->b_page);
2321 if (err < 0)
2322 return err;
2323 }
2324 return lblk < blocks;
2325}
2326
2327/*
2328 * mpage_map_buffers - update buffers corresponding to changed extent and
2329 * submit fully mapped pages for IO
2330 *
2331 * @mpd - description of extent to map, on return next extent to map
2332 *
2333 * Scan buffers corresponding to changed extent (we expect corresponding pages
2334 * to be already locked) and update buffer state according to new extent state.
2335 * We map delalloc buffers to their physical location, clear unwritten bits,
2336 * and mark buffers as uninit when we perform writes to unwritten extents
2337 * and do extent conversion after IO is finished. If the last page is not fully
2338 * mapped, we update @map to the next extent in the last page that needs
2339 * mapping. Otherwise we submit the page for IO.
2340 */
2341static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2342{
2343 struct pagevec pvec;
2344 int nr_pages, i;
2345 struct inode *inode = mpd->inode;
2346 struct buffer_head *head, *bh;
2347 int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2348 pgoff_t start, end;
2349 ext4_lblk_t lblk;
2350 sector_t pblock;
2351 int err;
2352
2353 start = mpd->map.m_lblk >> bpp_bits;
2354 end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2355 lblk = start << bpp_bits;
2356 pblock = mpd->map.m_pblk;
2357
2358 pagevec_init(&pvec);
2359 while (start <= end) {
2360 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2361 &start, end);
2362 if (nr_pages == 0)
2363 break;
2364 for (i = 0; i < nr_pages; i++) {
2365 struct page *page = pvec.pages[i];
2366
2367 bh = head = page_buffers(page);
2368 do {
2369 if (lblk < mpd->map.m_lblk)
2370 continue;
2371 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2372 /*
2373 * Buffer after end of mapped extent.
2374 * Find next buffer in the page to map.
2375 */
2376 mpd->map.m_len = 0;
2377 mpd->map.m_flags = 0;
2378 /*
2379 * FIXME: If dioread_nolock supports
2380 * blocksize < pagesize, we need to make
2381 * sure we add size mapped so far to
2382 * io_end->size as the following call
2383 * can submit the page for IO.
2384 */
2385 err = mpage_process_page_bufs(mpd, head,
2386 bh, lblk);
2387 pagevec_release(&pvec);
2388 if (err > 0)
2389 err = 0;
2390 return err;
2391 }
2392 if (buffer_delay(bh)) {
2393 clear_buffer_delay(bh);
2394 bh->b_blocknr = pblock++;
2395 }
2396 clear_buffer_unwritten(bh);
2397 } while (lblk++, (bh = bh->b_this_page) != head);
2398
2399 /*
2400 * FIXME: This is going to break if dioread_nolock
2401 * supports blocksize < pagesize as we will try to
2402 * convert potentially unmapped parts of inode.
2403 */
2404 mpd->io_submit.io_end->size += PAGE_SIZE;
2405 /* Page fully mapped - let IO run! */
2406 err = mpage_submit_page(mpd, page);
2407 if (err < 0) {
2408 pagevec_release(&pvec);
2409 return err;
2410 }
2411 }
2412 pagevec_release(&pvec);
2413 }
2414 /* Extent fully mapped and matches with page boundary. We are done. */
2415 mpd->map.m_len = 0;
2416 mpd->map.m_flags = 0;
2417 return 0;
2418}
2419
2420static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2421{
2422 struct inode *inode = mpd->inode;
2423 struct ext4_map_blocks *map = &mpd->map;
2424 int get_blocks_flags;
2425 int err, dioread_nolock;
2426
2427 trace_ext4_da_write_pages_extent(inode, map);
2428 /*
2429 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2430 * to convert an unwritten extent to be initialized (in the case
2431 * where we have written into one or more preallocated blocks). It is
2432 * possible that we're going to need more metadata blocks than
2433 * previously reserved. However we must not fail because we're in
2434 * writeback and there is nothing we can do about it so it might result
2435 * in data loss. So use reserved blocks to allocate metadata if
2436 * possible.
2437 *
2438 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2439 * the blocks in question are delalloc blocks. This indicates
2440 * that the blocks and quotas has already been checked when
2441 * the data was copied into the page cache.
2442 */
2443 get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2444 EXT4_GET_BLOCKS_METADATA_NOFAIL |
2445 EXT4_GET_BLOCKS_IO_SUBMIT;
2446 dioread_nolock = ext4_should_dioread_nolock(inode);
2447 if (dioread_nolock)
2448 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2449 if (map->m_flags & (1 << BH_Delay))
2450 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2451
2452 err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2453 if (err < 0)
2454 return err;
2455 if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2456 if (!mpd->io_submit.io_end->handle &&
2457 ext4_handle_valid(handle)) {
2458 mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2459 handle->h_rsv_handle = NULL;
2460 }
2461 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2462 }
2463
2464 BUG_ON(map->m_len == 0);
2465 if (map->m_flags & EXT4_MAP_NEW) {
2466 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2467 map->m_len);
2468 }
2469 return 0;
2470}
2471
2472/*
2473 * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2474 * mpd->len and submit pages underlying it for IO
2475 *
2476 * @handle - handle for journal operations
2477 * @mpd - extent to map
2478 * @give_up_on_write - we set this to true iff there is a fatal error and there
2479 * is no hope of writing the data. The caller should discard
2480 * dirty pages to avoid infinite loops.
2481 *
2482 * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2483 * delayed, blocks are allocated, if it is unwritten, we may need to convert
2484 * them to initialized or split the described range from larger unwritten
2485 * extent. Note that we need not map all the described range since allocation
2486 * can return less blocks or the range is covered by more unwritten extents. We
2487 * cannot map more because we are limited by reserved transaction credits. On
2488 * the other hand we always make sure that the last touched page is fully
2489 * mapped so that it can be written out (and thus forward progress is
2490 * guaranteed). After mapping we submit all mapped pages for IO.
2491 */
2492static int mpage_map_and_submit_extent(handle_t *handle,
2493 struct mpage_da_data *mpd,
2494 bool *give_up_on_write)
2495{
2496 struct inode *inode = mpd->inode;
2497 struct ext4_map_blocks *map = &mpd->map;
2498 int err;
2499 loff_t disksize;
2500 int progress = 0;
2501
2502 mpd->io_submit.io_end->offset =
2503 ((loff_t)map->m_lblk) << inode->i_blkbits;
2504 do {
2505 err = mpage_map_one_extent(handle, mpd);
2506 if (err < 0) {
2507 struct super_block *sb = inode->i_sb;
2508
2509 if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2510 EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2511 goto invalidate_dirty_pages;
2512 /*
2513 * Let the uper layers retry transient errors.
2514 * In the case of ENOSPC, if ext4_count_free_blocks()
2515 * is non-zero, a commit should free up blocks.
2516 */
2517 if ((err == -ENOMEM) ||
2518 (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2519 if (progress)
2520 goto update_disksize;
2521 return err;
2522 }
2523 ext4_msg(sb, KERN_CRIT,
2524 "Delayed block allocation failed for "
2525 "inode %lu at logical offset %llu with"
2526 " max blocks %u with error %d",
2527 inode->i_ino,
2528 (unsigned long long)map->m_lblk,
2529 (unsigned)map->m_len, -err);
2530 ext4_msg(sb, KERN_CRIT,
2531 "This should not happen!! Data will "
2532 "be lost\n");
2533 if (err == -ENOSPC)
2534 ext4_print_free_blocks(inode);
2535 invalidate_dirty_pages:
2536 *give_up_on_write = true;
2537 return err;
2538 }
2539 progress = 1;
2540 /*
2541 * Update buffer state, submit mapped pages, and get us new
2542 * extent to map
2543 */
2544 err = mpage_map_and_submit_buffers(mpd);
2545 if (err < 0)
2546 goto update_disksize;
2547 } while (map->m_len);
2548
2549update_disksize:
2550 /*
2551 * Update on-disk size after IO is submitted. Races with
2552 * truncate are avoided by checking i_size under i_data_sem.
2553 */
2554 disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2555 if (disksize > EXT4_I(inode)->i_disksize) {
2556 int err2;
2557 loff_t i_size;
2558
2559 down_write(&EXT4_I(inode)->i_data_sem);
2560 i_size = i_size_read(inode);
2561 if (disksize > i_size)
2562 disksize = i_size;
2563 if (disksize > EXT4_I(inode)->i_disksize)
2564 EXT4_I(inode)->i_disksize = disksize;
2565 up_write(&EXT4_I(inode)->i_data_sem);
2566 err2 = ext4_mark_inode_dirty(handle, inode);
2567 if (err2)
2568 ext4_error(inode->i_sb,
2569 "Failed to mark inode %lu dirty",
2570 inode->i_ino);
2571 if (!err)
2572 err = err2;
2573 }
2574 return err;
2575}
2576
2577/*
2578 * Calculate the total number of credits to reserve for one writepages
2579 * iteration. This is called from ext4_writepages(). We map an extent of
2580 * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2581 * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2582 * bpp - 1 blocks in bpp different extents.
2583 */
2584static int ext4_da_writepages_trans_blocks(struct inode *inode)
2585{
2586 int bpp = ext4_journal_blocks_per_page(inode);
2587
2588 return ext4_meta_trans_blocks(inode,
2589 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2590}
2591
2592/*
2593 * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2594 * and underlying extent to map
2595 *
2596 * @mpd - where to look for pages
2597 *
2598 * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2599 * IO immediately. When we find a page which isn't mapped we start accumulating
2600 * extent of buffers underlying these pages that needs mapping (formed by
2601 * either delayed or unwritten buffers). We also lock the pages containing
2602 * these buffers. The extent found is returned in @mpd structure (starting at
2603 * mpd->lblk with length mpd->len blocks).
2604 *
2605 * Note that this function can attach bios to one io_end structure which are
2606 * neither logically nor physically contiguous. Although it may seem as an
2607 * unnecessary complication, it is actually inevitable in blocksize < pagesize
2608 * case as we need to track IO to all buffers underlying a page in one io_end.
2609 */
2610static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2611{
2612 struct address_space *mapping = mpd->inode->i_mapping;
2613 struct pagevec pvec;
2614 unsigned int nr_pages;
2615 long left = mpd->wbc->nr_to_write;
2616 pgoff_t index = mpd->first_page;
2617 pgoff_t end = mpd->last_page;
2618 int tag;
2619 int i, err = 0;
2620 int blkbits = mpd->inode->i_blkbits;
2621 ext4_lblk_t lblk;
2622 struct buffer_head *head;
2623
2624 if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2625 tag = PAGECACHE_TAG_TOWRITE;
2626 else
2627 tag = PAGECACHE_TAG_DIRTY;
2628
2629 pagevec_init(&pvec);
2630 mpd->map.m_len = 0;
2631 mpd->next_page = index;
2632 while (index <= end) {
2633 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2634 tag);
2635 if (nr_pages == 0)
2636 goto out;
2637
2638 for (i = 0; i < nr_pages; i++) {
2639 struct page *page = pvec.pages[i];
2640
2641 /*
2642 * Accumulated enough dirty pages? This doesn't apply
2643 * to WB_SYNC_ALL mode. For integrity sync we have to
2644 * keep going because someone may be concurrently
2645 * dirtying pages, and we might have synced a lot of
2646 * newly appeared dirty pages, but have not synced all
2647 * of the old dirty pages.
2648 */
2649 if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2650 goto out;
2651
2652 /* If we can't merge this page, we are done. */
2653 if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2654 goto out;
2655
2656 lock_page(page);
2657 /*
2658 * If the page is no longer dirty, or its mapping no
2659 * longer corresponds to inode we are writing (which
2660 * means it has been truncated or invalidated), or the
2661 * page is already under writeback and we are not doing
2662 * a data integrity writeback, skip the page
2663 */
2664 if (!PageDirty(page) ||
2665 (PageWriteback(page) &&
2666 (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2667 unlikely(page->mapping != mapping)) {
2668 unlock_page(page);
2669 continue;
2670 }
2671
2672 wait_on_page_writeback(page);
2673 BUG_ON(PageWriteback(page));
2674
2675 if (mpd->map.m_len == 0)
2676 mpd->first_page = page->index;
2677 mpd->next_page = page->index + 1;
2678 /* Add all dirty buffers to mpd */
2679 lblk = ((ext4_lblk_t)page->index) <<
2680 (PAGE_SHIFT - blkbits);
2681 head = page_buffers(page);
2682 err = mpage_process_page_bufs(mpd, head, head, lblk);
2683 if (err <= 0)
2684 goto out;
2685 err = 0;
2686 left--;
2687 }
2688 pagevec_release(&pvec);
2689 cond_resched();
2690 }
2691 return 0;
2692out:
2693 pagevec_release(&pvec);
2694 return err;
2695}
2696
2697static int ext4_writepages(struct address_space *mapping,
2698 struct writeback_control *wbc)
2699{
2700 pgoff_t writeback_index = 0;
2701 long nr_to_write = wbc->nr_to_write;
2702 int range_whole = 0;
2703 int cycled = 1;
2704 handle_t *handle = NULL;
2705 struct mpage_da_data mpd;
2706 struct inode *inode = mapping->host;
2707 int needed_blocks, rsv_blocks = 0, ret = 0;
2708 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2709 bool done;
2710 struct blk_plug plug;
2711 bool give_up_on_write = false;
2712
2713 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2714 return -EIO;
2715
2716 percpu_down_read(&sbi->s_journal_flag_rwsem);
2717 trace_ext4_writepages(inode, wbc);
2718
2719 /*
2720 * No pages to write? This is mainly a kludge to avoid starting
2721 * a transaction for special inodes like journal inode on last iput()
2722 * because that could violate lock ordering on umount
2723 */
2724 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2725 goto out_writepages;
2726
2727 if (ext4_should_journal_data(inode)) {
2728 ret = generic_writepages(mapping, wbc);
2729 goto out_writepages;
2730 }
2731
2732 /*
2733 * If the filesystem has aborted, it is read-only, so return
2734 * right away instead of dumping stack traces later on that
2735 * will obscure the real source of the problem. We test
2736 * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2737 * the latter could be true if the filesystem is mounted
2738 * read-only, and in that case, ext4_writepages should
2739 * *never* be called, so if that ever happens, we would want
2740 * the stack trace.
2741 */
2742 if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2743 sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2744 ret = -EROFS;
2745 goto out_writepages;
2746 }
2747
2748 if (ext4_should_dioread_nolock(inode)) {
2749 /*
2750 * We may need to convert up to one extent per block in
2751 * the page and we may dirty the inode.
2752 */
2753 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2754 }
2755
2756 /*
2757 * If we have inline data and arrive here, it means that
2758 * we will soon create the block for the 1st page, so
2759 * we'd better clear the inline data here.
2760 */
2761 if (ext4_has_inline_data(inode)) {
2762 /* Just inode will be modified... */
2763 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2764 if (IS_ERR(handle)) {
2765 ret = PTR_ERR(handle);
2766 goto out_writepages;
2767 }
2768 BUG_ON(ext4_test_inode_state(inode,
2769 EXT4_STATE_MAY_INLINE_DATA));
2770 ext4_destroy_inline_data(handle, inode);
2771 ext4_journal_stop(handle);
2772 }
2773
2774 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2775 range_whole = 1;
2776
2777 if (wbc->range_cyclic) {
2778 writeback_index = mapping->writeback_index;
2779 if (writeback_index)
2780 cycled = 0;
2781 mpd.first_page = writeback_index;
2782 mpd.last_page = -1;
2783 } else {
2784 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2785 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2786 }
2787
2788 mpd.inode = inode;
2789 mpd.wbc = wbc;
2790 ext4_io_submit_init(&mpd.io_submit, wbc);
2791retry:
2792 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2793 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2794 done = false;
2795 blk_start_plug(&plug);
2796
2797 /*
2798 * First writeback pages that don't need mapping - we can avoid
2799 * starting a transaction unnecessarily and also avoid being blocked
2800 * in the block layer on device congestion while having transaction
2801 * started.
2802 */
2803 mpd.do_map = 0;
2804 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2805 if (!mpd.io_submit.io_end) {
2806 ret = -ENOMEM;
2807 goto unplug;
2808 }
2809 ret = mpage_prepare_extent_to_map(&mpd);
2810 /* Submit prepared bio */
2811 ext4_io_submit(&mpd.io_submit);
2812 ext4_put_io_end_defer(mpd.io_submit.io_end);
2813 mpd.io_submit.io_end = NULL;
2814 /* Unlock pages we didn't use */
2815 mpage_release_unused_pages(&mpd, false);
2816 if (ret < 0)
2817 goto unplug;
2818
2819 while (!done && mpd.first_page <= mpd.last_page) {
2820 /* For each extent of pages we use new io_end */
2821 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2822 if (!mpd.io_submit.io_end) {
2823 ret = -ENOMEM;
2824 break;
2825 }
2826
2827 /*
2828 * We have two constraints: We find one extent to map and we
2829 * must always write out whole page (makes a difference when
2830 * blocksize < pagesize) so that we don't block on IO when we
2831 * try to write out the rest of the page. Journalled mode is
2832 * not supported by delalloc.
2833 */
2834 BUG_ON(ext4_should_journal_data(inode));
2835 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2836
2837 /* start a new transaction */
2838 handle = ext4_journal_start_with_reserve(inode,
2839 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2840 if (IS_ERR(handle)) {
2841 ret = PTR_ERR(handle);
2842 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2843 "%ld pages, ino %lu; err %d", __func__,
2844 wbc->nr_to_write, inode->i_ino, ret);
2845 /* Release allocated io_end */
2846 ext4_put_io_end(mpd.io_submit.io_end);
2847 mpd.io_submit.io_end = NULL;
2848 break;
2849 }
2850 mpd.do_map = 1;
2851
2852 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2853 ret = mpage_prepare_extent_to_map(&mpd);
2854 if (!ret) {
2855 if (mpd.map.m_len)
2856 ret = mpage_map_and_submit_extent(handle, &mpd,
2857 &give_up_on_write);
2858 else {
2859 /*
2860 * We scanned the whole range (or exhausted
2861 * nr_to_write), submitted what was mapped and
2862 * didn't find anything needing mapping. We are
2863 * done.
2864 */
2865 done = true;
2866 }
2867 }
2868 /*
2869 * Caution: If the handle is synchronous,
2870 * ext4_journal_stop() can wait for transaction commit
2871 * to finish which may depend on writeback of pages to
2872 * complete or on page lock to be released. In that
2873 * case, we have to wait until after after we have
2874 * submitted all the IO, released page locks we hold,
2875 * and dropped io_end reference (for extent conversion
2876 * to be able to complete) before stopping the handle.
2877 */
2878 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2879 ext4_journal_stop(handle);
2880 handle = NULL;
2881 mpd.do_map = 0;
2882 }
2883 /* Submit prepared bio */
2884 ext4_io_submit(&mpd.io_submit);
2885 /* Unlock pages we didn't use */
2886 mpage_release_unused_pages(&mpd, give_up_on_write);
2887 /*
2888 * Drop our io_end reference we got from init. We have
2889 * to be careful and use deferred io_end finishing if
2890 * we are still holding the transaction as we can
2891 * release the last reference to io_end which may end
2892 * up doing unwritten extent conversion.
2893 */
2894 if (handle) {
2895 ext4_put_io_end_defer(mpd.io_submit.io_end);
2896 ext4_journal_stop(handle);
2897 } else
2898 ext4_put_io_end(mpd.io_submit.io_end);
2899 mpd.io_submit.io_end = NULL;
2900
2901 if (ret == -ENOSPC && sbi->s_journal) {
2902 /*
2903 * Commit the transaction which would
2904 * free blocks released in the transaction
2905 * and try again
2906 */
2907 jbd2_journal_force_commit_nested(sbi->s_journal);
2908 ret = 0;
2909 continue;
2910 }
2911 /* Fatal error - ENOMEM, EIO... */
2912 if (ret)
2913 break;
2914 }
2915unplug:
2916 blk_finish_plug(&plug);
2917 if (!ret && !cycled && wbc->nr_to_write > 0) {
2918 cycled = 1;
2919 mpd.last_page = writeback_index - 1;
2920 mpd.first_page = 0;
2921 goto retry;
2922 }
2923
2924 /* Update index */
2925 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2926 /*
2927 * Set the writeback_index so that range_cyclic
2928 * mode will write it back later
2929 */
2930 mapping->writeback_index = mpd.first_page;
2931
2932out_writepages:
2933 trace_ext4_writepages_result(inode, wbc, ret,
2934 nr_to_write - wbc->nr_to_write);
2935 percpu_up_read(&sbi->s_journal_flag_rwsem);
2936 return ret;
2937}
2938
2939static int ext4_dax_writepages(struct address_space *mapping,
2940 struct writeback_control *wbc)
2941{
2942 int ret;
2943 long nr_to_write = wbc->nr_to_write;
2944 struct inode *inode = mapping->host;
2945 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2946
2947 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2948 return -EIO;
2949
2950 percpu_down_read(&sbi->s_journal_flag_rwsem);
2951 trace_ext4_writepages(inode, wbc);
2952
2953 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev, wbc);
2954 trace_ext4_writepages_result(inode, wbc, ret,
2955 nr_to_write - wbc->nr_to_write);
2956 percpu_up_read(&sbi->s_journal_flag_rwsem);
2957 return ret;
2958}
2959
2960static int ext4_nonda_switch(struct super_block *sb)
2961{
2962 s64 free_clusters, dirty_clusters;
2963 struct ext4_sb_info *sbi = EXT4_SB(sb);
2964
2965 /*
2966 * switch to non delalloc mode if we are running low
2967 * on free block. The free block accounting via percpu
2968 * counters can get slightly wrong with percpu_counter_batch getting
2969 * accumulated on each CPU without updating global counters
2970 * Delalloc need an accurate free block accounting. So switch
2971 * to non delalloc when we are near to error range.
2972 */
2973 free_clusters =
2974 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2975 dirty_clusters =
2976 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2977 /*
2978 * Start pushing delalloc when 1/2 of free blocks are dirty.
2979 */
2980 if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2981 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2982
2983 if (2 * free_clusters < 3 * dirty_clusters ||
2984 free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2985 /*
2986 * free block count is less than 150% of dirty blocks
2987 * or free blocks is less than watermark
2988 */
2989 return 1;
2990 }
2991 return 0;
2992}
2993
2994/* We always reserve for an inode update; the superblock could be there too */
2995static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2996{
2997 if (likely(ext4_has_feature_large_file(inode->i_sb)))
2998 return 1;
2999
3000 if (pos + len <= 0x7fffffffULL)
3001 return 1;
3002
3003 /* We might need to update the superblock to set LARGE_FILE */
3004 return 2;
3005}
3006
3007static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3008 loff_t pos, unsigned len, unsigned flags,
3009 struct page **pagep, void **fsdata)
3010{
3011 int ret, retries = 0;
3012 struct page *page;
3013 pgoff_t index;
3014 struct inode *inode = mapping->host;
3015 handle_t *handle;
3016
3017 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3018 return -EIO;
3019
3020 index = pos >> PAGE_SHIFT;
3021
3022 if (ext4_nonda_switch(inode->i_sb) ||
3023 S_ISLNK(inode->i_mode)) {
3024 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3025 return ext4_write_begin(file, mapping, pos,
3026 len, flags, pagep, fsdata);
3027 }
3028 *fsdata = (void *)0;
3029 trace_ext4_da_write_begin(inode, pos, len, flags);
3030
3031 if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3032 ret = ext4_da_write_inline_data_begin(mapping, inode,
3033 pos, len, flags,
3034 pagep, fsdata);
3035 if (ret < 0)
3036 return ret;
3037 if (ret == 1)
3038 return 0;
3039 }
3040
3041 /*
3042 * grab_cache_page_write_begin() can take a long time if the
3043 * system is thrashing due to memory pressure, or if the page
3044 * is being written back. So grab it first before we start
3045 * the transaction handle. This also allows us to allocate
3046 * the page (if needed) without using GFP_NOFS.
3047 */
3048retry_grab:
3049 page = grab_cache_page_write_begin(mapping, index, flags);
3050 if (!page)
3051 return -ENOMEM;
3052 unlock_page(page);
3053
3054 /*
3055 * With delayed allocation, we don't log the i_disksize update
3056 * if there is delayed block allocation. But we still need
3057 * to journalling the i_disksize update if writes to the end
3058 * of file which has an already mapped buffer.
3059 */
3060retry_journal:
3061 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3062 ext4_da_write_credits(inode, pos, len));
3063 if (IS_ERR(handle)) {
3064 put_page(page);
3065 return PTR_ERR(handle);
3066 }
3067
3068 lock_page(page);
3069 if (page->mapping != mapping) {
3070 /* The page got truncated from under us */
3071 unlock_page(page);
3072 put_page(page);
3073 ext4_journal_stop(handle);
3074 goto retry_grab;
3075 }
3076 /* In case writeback began while the page was unlocked */
3077 wait_for_stable_page(page);
3078
3079#ifdef CONFIG_EXT4_FS_ENCRYPTION
3080 ret = ext4_block_write_begin(page, pos, len,
3081 ext4_da_get_block_prep);
3082#else
3083 ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3084#endif
3085 if (ret < 0) {
3086 unlock_page(page);
3087 ext4_journal_stop(handle);
3088 /*
3089 * block_write_begin may have instantiated a few blocks
3090 * outside i_size. Trim these off again. Don't need
3091 * i_size_read because we hold i_mutex.
3092 */
3093 if (pos + len > inode->i_size)
3094 ext4_truncate_failed_write(inode);
3095
3096 if (ret == -ENOSPC &&
3097 ext4_should_retry_alloc(inode->i_sb, &retries))
3098 goto retry_journal;
3099
3100 put_page(page);
3101 return ret;
3102 }
3103
3104 *pagep = page;
3105 return ret;
3106}
3107
3108/*
3109 * Check if we should update i_disksize
3110 * when write to the end of file but not require block allocation
3111 */
3112static int ext4_da_should_update_i_disksize(struct page *page,
3113 unsigned long offset)
3114{
3115 struct buffer_head *bh;
3116 struct inode *inode = page->mapping->host;
3117 unsigned int idx;
3118 int i;
3119
3120 bh = page_buffers(page);
3121 idx = offset >> inode->i_blkbits;
3122
3123 for (i = 0; i < idx; i++)
3124 bh = bh->b_this_page;
3125
3126 if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3127 return 0;
3128 return 1;
3129}
3130
3131static int ext4_da_write_end(struct file *file,
3132 struct address_space *mapping,
3133 loff_t pos, unsigned len, unsigned copied,
3134 struct page *page, void *fsdata)
3135{
3136 struct inode *inode = mapping->host;
3137 int ret = 0, ret2;
3138 handle_t *handle = ext4_journal_current_handle();
3139 loff_t new_i_size;
3140 unsigned long start, end;
3141 int write_mode = (int)(unsigned long)fsdata;
3142
3143 if (write_mode == FALL_BACK_TO_NONDELALLOC)
3144 return ext4_write_end(file, mapping, pos,
3145 len, copied, page, fsdata);
3146
3147 trace_ext4_da_write_end(inode, pos, len, copied);
3148 start = pos & (PAGE_SIZE - 1);
3149 end = start + copied - 1;
3150
3151 /*
3152 * generic_write_end() will run mark_inode_dirty() if i_size
3153 * changes. So let's piggyback the i_disksize mark_inode_dirty
3154 * into that.
3155 */
3156 new_i_size = pos + copied;
3157 if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3158 if (ext4_has_inline_data(inode) ||
3159 ext4_da_should_update_i_disksize(page, end)) {
3160 ext4_update_i_disksize(inode, new_i_size);
3161 /* We need to mark inode dirty even if
3162 * new_i_size is less that inode->i_size
3163 * bu greater than i_disksize.(hint delalloc)
3164 */
3165 ext4_mark_inode_dirty(handle, inode);
3166 }
3167 }
3168
3169 if (write_mode != CONVERT_INLINE_DATA &&
3170 ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3171 ext4_has_inline_data(inode))
3172 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3173 page);
3174 else
3175 ret2 = generic_write_end(file, mapping, pos, len, copied,
3176 page, fsdata);
3177
3178 copied = ret2;
3179 if (ret2 < 0)
3180 ret = ret2;
3181 ret2 = ext4_journal_stop(handle);
3182 if (!ret)
3183 ret = ret2;
3184
3185 return ret ? ret : copied;
3186}
3187
3188static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3189 unsigned int length)
3190{
3191 /*
3192 * Drop reserved blocks
3193 */
3194 BUG_ON(!PageLocked(page));
3195 if (!page_has_buffers(page))
3196 goto out;
3197
3198 ext4_da_page_release_reservation(page, offset, length);
3199
3200out:
3201 ext4_invalidatepage(page, offset, length);
3202
3203 return;
3204}
3205
3206/*
3207 * Force all delayed allocation blocks to be allocated for a given inode.
3208 */
3209int ext4_alloc_da_blocks(struct inode *inode)
3210{
3211 trace_ext4_alloc_da_blocks(inode);
3212
3213 if (!EXT4_I(inode)->i_reserved_data_blocks)
3214 return 0;
3215
3216 /*
3217 * We do something simple for now. The filemap_flush() will
3218 * also start triggering a write of the data blocks, which is
3219 * not strictly speaking necessary (and for users of
3220 * laptop_mode, not even desirable). However, to do otherwise
3221 * would require replicating code paths in:
3222 *
3223 * ext4_writepages() ->
3224 * write_cache_pages() ---> (via passed in callback function)
3225 * __mpage_da_writepage() -->
3226 * mpage_add_bh_to_extent()
3227 * mpage_da_map_blocks()
3228 *
3229 * The problem is that write_cache_pages(), located in
3230 * mm/page-writeback.c, marks pages clean in preparation for
3231 * doing I/O, which is not desirable if we're not planning on
3232 * doing I/O at all.
3233 *
3234 * We could call write_cache_pages(), and then redirty all of
3235 * the pages by calling redirty_page_for_writepage() but that
3236 * would be ugly in the extreme. So instead we would need to
3237 * replicate parts of the code in the above functions,
3238 * simplifying them because we wouldn't actually intend to
3239 * write out the pages, but rather only collect contiguous
3240 * logical block extents, call the multi-block allocator, and
3241 * then update the buffer heads with the block allocations.
3242 *
3243 * For now, though, we'll cheat by calling filemap_flush(),
3244 * which will map the blocks, and start the I/O, but not
3245 * actually wait for the I/O to complete.
3246 */
3247 return filemap_flush(inode->i_mapping);
3248}
3249
3250/*
3251 * bmap() is special. It gets used by applications such as lilo and by
3252 * the swapper to find the on-disk block of a specific piece of data.
3253 *
3254 * Naturally, this is dangerous if the block concerned is still in the
3255 * journal. If somebody makes a swapfile on an ext4 data-journaling
3256 * filesystem and enables swap, then they may get a nasty shock when the
3257 * data getting swapped to that swapfile suddenly gets overwritten by
3258 * the original zero's written out previously to the journal and
3259 * awaiting writeback in the kernel's buffer cache.
3260 *
3261 * So, if we see any bmap calls here on a modified, data-journaled file,
3262 * take extra steps to flush any blocks which might be in the cache.
3263 */
3264static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3265{
3266 struct inode *inode = mapping->host;
3267 journal_t *journal;
3268 int err;
3269
3270 /*
3271 * We can get here for an inline file via the FIBMAP ioctl
3272 */
3273 if (ext4_has_inline_data(inode))
3274 return 0;
3275
3276 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3277 test_opt(inode->i_sb, DELALLOC)) {
3278 /*
3279 * With delalloc we want to sync the file
3280 * so that we can make sure we allocate
3281 * blocks for file
3282 */
3283 filemap_write_and_wait(mapping);
3284 }
3285
3286 if (EXT4_JOURNAL(inode) &&
3287 ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3288 /*
3289 * This is a REALLY heavyweight approach, but the use of
3290 * bmap on dirty files is expected to be extremely rare:
3291 * only if we run lilo or swapon on a freshly made file
3292 * do we expect this to happen.
3293 *
3294 * (bmap requires CAP_SYS_RAWIO so this does not
3295 * represent an unprivileged user DOS attack --- we'd be
3296 * in trouble if mortal users could trigger this path at
3297 * will.)
3298 *
3299 * NB. EXT4_STATE_JDATA is not set on files other than
3300 * regular files. If somebody wants to bmap a directory
3301 * or symlink and gets confused because the buffer
3302 * hasn't yet been flushed to disk, they deserve
3303 * everything they get.
3304 */
3305
3306 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3307 journal = EXT4_JOURNAL(inode);
3308 jbd2_journal_lock_updates(journal);
3309 err = jbd2_journal_flush(journal);
3310 jbd2_journal_unlock_updates(journal);
3311
3312 if (err)
3313 return 0;
3314 }
3315
3316 return generic_block_bmap(mapping, block, ext4_get_block);
3317}
3318
3319static int ext4_readpage(struct file *file, struct page *page)
3320{
3321 int ret = -EAGAIN;
3322 struct inode *inode = page->mapping->host;
3323
3324 trace_ext4_readpage(page);
3325
3326 if (ext4_has_inline_data(inode))
3327 ret = ext4_readpage_inline(inode, page);
3328
3329 if (ret == -EAGAIN)
3330 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3331
3332 return ret;
3333}
3334
3335static int
3336ext4_readpages(struct file *file, struct address_space *mapping,
3337 struct list_head *pages, unsigned nr_pages)
3338{
3339 struct inode *inode = mapping->host;
3340
3341 /* If the file has inline data, no need to do readpages. */
3342 if (ext4_has_inline_data(inode))
3343 return 0;
3344
3345 return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3346}
3347
3348static void ext4_invalidatepage(struct page *page, unsigned int offset,
3349 unsigned int length)
3350{
3351 trace_ext4_invalidatepage(page, offset, length);
3352
3353 /* No journalling happens on data buffers when this function is used */
3354 WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3355
3356 block_invalidatepage(page, offset, length);
3357}
3358
3359static int __ext4_journalled_invalidatepage(struct page *page,
3360 unsigned int offset,
3361 unsigned int length)
3362{
3363 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3364
3365 trace_ext4_journalled_invalidatepage(page, offset, length);
3366
3367 /*
3368 * If it's a full truncate we just forget about the pending dirtying
3369 */
3370 if (offset == 0 && length == PAGE_SIZE)
3371 ClearPageChecked(page);
3372
3373 return jbd2_journal_invalidatepage(journal, page, offset, length);
3374}
3375
3376/* Wrapper for aops... */
3377static void ext4_journalled_invalidatepage(struct page *page,
3378 unsigned int offset,
3379 unsigned int length)
3380{
3381 WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3382}
3383
3384static int ext4_releasepage(struct page *page, gfp_t wait)
3385{
3386 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3387
3388 trace_ext4_releasepage(page);
3389
3390 /* Page has dirty journalled data -> cannot release */
3391 if (PageChecked(page))
3392 return 0;
3393 if (journal)
3394 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3395 else
3396 return try_to_free_buffers(page);
3397}
3398
3399static bool ext4_inode_datasync_dirty(struct inode *inode)
3400{
3401 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
3402
3403 if (journal)
3404 return !jbd2_transaction_committed(journal,
3405 EXT4_I(inode)->i_datasync_tid);
3406 /* Any metadata buffers to write? */
3407 if (!list_empty(&inode->i_mapping->private_list))
3408 return true;
3409 return inode->i_state & I_DIRTY_DATASYNC;
3410}
3411
3412static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3413 unsigned flags, struct iomap *iomap)
3414{
3415 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3416 unsigned int blkbits = inode->i_blkbits;
3417 unsigned long first_block = offset >> blkbits;
3418 unsigned long last_block = (offset + length - 1) >> blkbits;
3419 struct ext4_map_blocks map;
3420 bool delalloc = false;
3421 int ret;
3422
3423
3424 if (flags & IOMAP_REPORT) {
3425 if (ext4_has_inline_data(inode)) {
3426 ret = ext4_inline_data_iomap(inode, iomap);
3427 if (ret != -EAGAIN) {
3428 if (ret == 0 && offset >= iomap->length)
3429 ret = -ENOENT;
3430 return ret;
3431 }
3432 }
3433 } else {
3434 if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3435 return -ERANGE;
3436 }
3437
3438 map.m_lblk = first_block;
3439 map.m_len = last_block - first_block + 1;
3440
3441 if (flags & IOMAP_REPORT) {
3442 ret = ext4_map_blocks(NULL, inode, &map, 0);
3443 if (ret < 0)
3444 return ret;
3445
3446 if (ret == 0) {
3447 ext4_lblk_t end = map.m_lblk + map.m_len - 1;
3448 struct extent_status es;
3449
3450 ext4_es_find_delayed_extent_range(inode, map.m_lblk, end, &es);
3451
3452 if (!es.es_len || es.es_lblk > end) {
3453 /* entire range is a hole */
3454 } else if (es.es_lblk > map.m_lblk) {
3455 /* range starts with a hole */
3456 map.m_len = es.es_lblk - map.m_lblk;
3457 } else {
3458 ext4_lblk_t offs = 0;
3459
3460 if (es.es_lblk < map.m_lblk)
3461 offs = map.m_lblk - es.es_lblk;
3462 map.m_lblk = es.es_lblk + offs;
3463 map.m_len = es.es_len - offs;
3464 delalloc = true;
3465 }
3466 }
3467 } else if (flags & IOMAP_WRITE) {
3468 int dio_credits;
3469 handle_t *handle;
3470 int retries = 0;
3471
3472 /* Trim mapping request to maximum we can map at once for DIO */
3473 if (map.m_len > DIO_MAX_BLOCKS)
3474 map.m_len = DIO_MAX_BLOCKS;
3475 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3476retry:
3477 /*
3478 * Either we allocate blocks and then we don't get unwritten
3479 * extent so we have reserved enough credits, or the blocks
3480 * are already allocated and unwritten and in that case
3481 * extent conversion fits in the credits as well.
3482 */
3483 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3484 dio_credits);
3485 if (IS_ERR(handle))
3486 return PTR_ERR(handle);
3487
3488 ret = ext4_map_blocks(handle, inode, &map,
3489 EXT4_GET_BLOCKS_CREATE_ZERO);
3490 if (ret < 0) {
3491 ext4_journal_stop(handle);
3492 if (ret == -ENOSPC &&
3493 ext4_should_retry_alloc(inode->i_sb, &retries))
3494 goto retry;
3495 return ret;
3496 }
3497
3498 /*
3499 * If we added blocks beyond i_size, we need to make sure they
3500 * will get truncated if we crash before updating i_size in
3501 * ext4_iomap_end(). For faults we don't need to do that (and
3502 * even cannot because for orphan list operations inode_lock is
3503 * required) - if we happen to instantiate block beyond i_size,
3504 * it is because we race with truncate which has already added
3505 * the inode to the orphan list.
3506 */
3507 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3508 (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3509 int err;
3510
3511 err = ext4_orphan_add(handle, inode);
3512 if (err < 0) {
3513 ext4_journal_stop(handle);
3514 return err;
3515 }
3516 }
3517 ext4_journal_stop(handle);
3518 } else {
3519 ret = ext4_map_blocks(NULL, inode, &map, 0);
3520 if (ret < 0)
3521 return ret;
3522 }
3523
3524 iomap->flags = 0;
3525 if (ext4_inode_datasync_dirty(inode))
3526 iomap->flags |= IOMAP_F_DIRTY;
3527 iomap->bdev = inode->i_sb->s_bdev;
3528 iomap->dax_dev = sbi->s_daxdev;
3529 iomap->offset = (u64)first_block << blkbits;
3530 iomap->length = (u64)map.m_len << blkbits;
3531
3532 if (ret == 0) {
3533 iomap->type = delalloc ? IOMAP_DELALLOC : IOMAP_HOLE;
3534 iomap->addr = IOMAP_NULL_ADDR;
3535 } else {
3536 if (map.m_flags & EXT4_MAP_MAPPED) {
3537 iomap->type = IOMAP_MAPPED;
3538 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3539 iomap->type = IOMAP_UNWRITTEN;
3540 } else {
3541 WARN_ON_ONCE(1);
3542 return -EIO;
3543 }
3544 iomap->addr = (u64)map.m_pblk << blkbits;
3545 }
3546
3547 if (map.m_flags & EXT4_MAP_NEW)
3548 iomap->flags |= IOMAP_F_NEW;
3549
3550 return 0;
3551}
3552
3553static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3554 ssize_t written, unsigned flags, struct iomap *iomap)
3555{
3556 int ret = 0;
3557 handle_t *handle;
3558 int blkbits = inode->i_blkbits;
3559 bool truncate = false;
3560
3561 if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3562 return 0;
3563
3564 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3565 if (IS_ERR(handle)) {
3566 ret = PTR_ERR(handle);
3567 goto orphan_del;
3568 }
3569 if (ext4_update_inode_size(inode, offset + written))
3570 ext4_mark_inode_dirty(handle, inode);
3571 /*
3572 * We may need to truncate allocated but not written blocks beyond EOF.
3573 */
3574 if (iomap->offset + iomap->length >
3575 ALIGN(inode->i_size, 1 << blkbits)) {
3576 ext4_lblk_t written_blk, end_blk;
3577
3578 written_blk = (offset + written) >> blkbits;
3579 end_blk = (offset + length) >> blkbits;
3580 if (written_blk < end_blk && ext4_can_truncate(inode))
3581 truncate = true;
3582 }
3583 /*
3584 * Remove inode from orphan list if we were extending a inode and
3585 * everything went fine.
3586 */
3587 if (!truncate && inode->i_nlink &&
3588 !list_empty(&EXT4_I(inode)->i_orphan))
3589 ext4_orphan_del(handle, inode);
3590 ext4_journal_stop(handle);
3591 if (truncate) {
3592 ext4_truncate_failed_write(inode);
3593orphan_del:
3594 /*
3595 * If truncate failed early the inode might still be on the
3596 * orphan list; we need to make sure the inode is removed from
3597 * the orphan list in that case.
3598 */
3599 if (inode->i_nlink)
3600 ext4_orphan_del(NULL, inode);
3601 }
3602 return ret;
3603}
3604
3605const struct iomap_ops ext4_iomap_ops = {
3606 .iomap_begin = ext4_iomap_begin,
3607 .iomap_end = ext4_iomap_end,
3608};
3609
3610static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3611 ssize_t size, void *private)
3612{
3613 ext4_io_end_t *io_end = private;
3614
3615 /* if not async direct IO just return */
3616 if (!io_end)
3617 return 0;
3618
3619 ext_debug("ext4_end_io_dio(): io_end 0x%p "
3620 "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3621 io_end, io_end->inode->i_ino, iocb, offset, size);
3622
3623 /*
3624 * Error during AIO DIO. We cannot convert unwritten extents as the
3625 * data was not written. Just clear the unwritten flag and drop io_end.
3626 */
3627 if (size <= 0) {
3628 ext4_clear_io_unwritten_flag(io_end);
3629 size = 0;
3630 }
3631 io_end->offset = offset;
3632 io_end->size = size;
3633 ext4_put_io_end(io_end);
3634
3635 return 0;
3636}
3637
3638/*
3639 * Handling of direct IO writes.
3640 *
3641 * For ext4 extent files, ext4 will do direct-io write even to holes,
3642 * preallocated extents, and those write extend the file, no need to
3643 * fall back to buffered IO.
3644 *
3645 * For holes, we fallocate those blocks, mark them as unwritten
3646 * If those blocks were preallocated, we mark sure they are split, but
3647 * still keep the range to write as unwritten.
3648 *
3649 * The unwritten extents will be converted to written when DIO is completed.
3650 * For async direct IO, since the IO may still pending when return, we
3651 * set up an end_io call back function, which will do the conversion
3652 * when async direct IO completed.
3653 *
3654 * If the O_DIRECT write will extend the file then add this inode to the
3655 * orphan list. So recovery will truncate it back to the original size
3656 * if the machine crashes during the write.
3657 *
3658 */
3659static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3660{
3661 struct file *file = iocb->ki_filp;
3662 struct inode *inode = file->f_mapping->host;
3663 struct ext4_inode_info *ei = EXT4_I(inode);
3664 ssize_t ret;
3665 loff_t offset = iocb->ki_pos;
3666 size_t count = iov_iter_count(iter);
3667 int overwrite = 0;
3668 get_block_t *get_block_func = NULL;
3669 int dio_flags = 0;
3670 loff_t final_size = offset + count;
3671 int orphan = 0;
3672 handle_t *handle;
3673
3674 if (final_size > inode->i_size || final_size > ei->i_disksize) {
3675 /* Credits for sb + inode write */
3676 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3677 if (IS_ERR(handle)) {
3678 ret = PTR_ERR(handle);
3679 goto out;
3680 }
3681 ret = ext4_orphan_add(handle, inode);
3682 if (ret) {
3683 ext4_journal_stop(handle);
3684 goto out;
3685 }
3686 orphan = 1;
3687 ext4_update_i_disksize(inode, inode->i_size);
3688 ext4_journal_stop(handle);
3689 }
3690
3691 BUG_ON(iocb->private == NULL);
3692
3693 /*
3694 * Make all waiters for direct IO properly wait also for extent
3695 * conversion. This also disallows race between truncate() and
3696 * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3697 */
3698 inode_dio_begin(inode);
3699
3700 /* If we do a overwrite dio, i_mutex locking can be released */
3701 overwrite = *((int *)iocb->private);
3702
3703 if (overwrite)
3704 inode_unlock(inode);
3705
3706 /*
3707 * For extent mapped files we could direct write to holes and fallocate.
3708 *
3709 * Allocated blocks to fill the hole are marked as unwritten to prevent
3710 * parallel buffered read to expose the stale data before DIO complete
3711 * the data IO.
3712 *
3713 * As to previously fallocated extents, ext4 get_block will just simply
3714 * mark the buffer mapped but still keep the extents unwritten.
3715 *
3716 * For non AIO case, we will convert those unwritten extents to written
3717 * after return back from blockdev_direct_IO. That way we save us from
3718 * allocating io_end structure and also the overhead of offloading
3719 * the extent convertion to a workqueue.
3720 *
3721 * For async DIO, the conversion needs to be deferred when the
3722 * IO is completed. The ext4 end_io callback function will be
3723 * called to take care of the conversion work. Here for async
3724 * case, we allocate an io_end structure to hook to the iocb.
3725 */
3726 iocb->private = NULL;
3727 if (overwrite)
3728 get_block_func = ext4_dio_get_block_overwrite;
3729 else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
3730 round_down(offset, i_blocksize(inode)) >= inode->i_size) {
3731 get_block_func = ext4_dio_get_block;
3732 dio_flags = DIO_LOCKING | DIO_SKIP_HOLES;
3733 } else if (is_sync_kiocb(iocb)) {
3734 get_block_func = ext4_dio_get_block_unwritten_sync;
3735 dio_flags = DIO_LOCKING;
3736 } else {
3737 get_block_func = ext4_dio_get_block_unwritten_async;
3738 dio_flags = DIO_LOCKING;
3739 }
3740 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
3741 get_block_func, ext4_end_io_dio, NULL,
3742 dio_flags);
3743
3744 if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3745 EXT4_STATE_DIO_UNWRITTEN)) {
3746 int err;
3747 /*
3748 * for non AIO case, since the IO is already
3749 * completed, we could do the conversion right here
3750 */
3751 err = ext4_convert_unwritten_extents(NULL, inode,
3752 offset, ret);
3753 if (err < 0)
3754 ret = err;
3755 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3756 }
3757
3758 inode_dio_end(inode);
3759 /* take i_mutex locking again if we do a ovewrite dio */
3760 if (overwrite)
3761 inode_lock(inode);
3762
3763 if (ret < 0 && final_size > inode->i_size)
3764 ext4_truncate_failed_write(inode);
3765
3766 /* Handle extending of i_size after direct IO write */
3767 if (orphan) {
3768 int err;
3769
3770 /* Credits for sb + inode write */
3771 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3772 if (IS_ERR(handle)) {
3773 /*
3774 * We wrote the data but cannot extend
3775 * i_size. Bail out. In async io case, we do
3776 * not return error here because we have
3777 * already submmitted the corresponding
3778 * bio. Returning error here makes the caller
3779 * think that this IO is done and failed
3780 * resulting in race with bio's completion
3781 * handler.
3782 */
3783 if (!ret)
3784 ret = PTR_ERR(handle);
3785 if (inode->i_nlink)
3786 ext4_orphan_del(NULL, inode);
3787
3788 goto out;
3789 }
3790 if (inode->i_nlink)
3791 ext4_orphan_del(handle, inode);
3792 if (ret > 0) {
3793 loff_t end = offset + ret;
3794 if (end > inode->i_size || end > ei->i_disksize) {
3795 ext4_update_i_disksize(inode, end);
3796 if (end > inode->i_size)
3797 i_size_write(inode, end);
3798 /*
3799 * We're going to return a positive `ret'
3800 * here due to non-zero-length I/O, so there's
3801 * no way of reporting error returns from
3802 * ext4_mark_inode_dirty() to userspace. So
3803 * ignore it.
3804 */
3805 ext4_mark_inode_dirty(handle, inode);
3806 }
3807 }
3808 err = ext4_journal_stop(handle);
3809 if (ret == 0)
3810 ret = err;
3811 }
3812out:
3813 return ret;
3814}
3815
3816static ssize_t ext4_direct_IO_read(struct kiocb *iocb, struct iov_iter *iter)
3817{
3818 struct address_space *mapping = iocb->ki_filp->f_mapping;
3819 struct inode *inode = mapping->host;
3820 size_t count = iov_iter_count(iter);
3821 ssize_t ret;
3822
3823 /*
3824 * Shared inode_lock is enough for us - it protects against concurrent
3825 * writes & truncates and since we take care of writing back page cache,
3826 * we are protected against page writeback as well.
3827 */
3828 inode_lock_shared(inode);
3829 ret = filemap_write_and_wait_range(mapping, iocb->ki_pos,
3830 iocb->ki_pos + count - 1);
3831 if (ret)
3832 goto out_unlock;
3833 ret = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
3834 iter, ext4_dio_get_block, NULL, NULL, 0);
3835out_unlock:
3836 inode_unlock_shared(inode);
3837 return ret;
3838}
3839
3840static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3841{
3842 struct file *file = iocb->ki_filp;
3843 struct inode *inode = file->f_mapping->host;
3844 size_t count = iov_iter_count(iter);
3845 loff_t offset = iocb->ki_pos;
3846 ssize_t ret;
3847
3848#ifdef CONFIG_EXT4_FS_ENCRYPTION
3849 if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode))
3850 return 0;
3851#endif
3852
3853 /*
3854 * If we are doing data journalling we don't support O_DIRECT
3855 */
3856 if (ext4_should_journal_data(inode))
3857 return 0;
3858
3859 /* Let buffer I/O handle the inline data case. */
3860 if (ext4_has_inline_data(inode))
3861 return 0;
3862
3863 trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3864 if (iov_iter_rw(iter) == READ)
3865 ret = ext4_direct_IO_read(iocb, iter);
3866 else
3867 ret = ext4_direct_IO_write(iocb, iter);
3868 trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3869 return ret;
3870}
3871
3872/*
3873 * Pages can be marked dirty completely asynchronously from ext4's journalling
3874 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3875 * much here because ->set_page_dirty is called under VFS locks. The page is
3876 * not necessarily locked.
3877 *
3878 * We cannot just dirty the page and leave attached buffers clean, because the
3879 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3880 * or jbddirty because all the journalling code will explode.
3881 *
3882 * So what we do is to mark the page "pending dirty" and next time writepage
3883 * is called, propagate that into the buffers appropriately.
3884 */
3885static int ext4_journalled_set_page_dirty(struct page *page)
3886{
3887 SetPageChecked(page);
3888 return __set_page_dirty_nobuffers(page);
3889}
3890
3891static int ext4_set_page_dirty(struct page *page)
3892{
3893 WARN_ON_ONCE(!PageLocked(page) && !PageDirty(page));
3894 WARN_ON_ONCE(!page_has_buffers(page));
3895 return __set_page_dirty_buffers(page);
3896}
3897
3898static const struct address_space_operations ext4_aops = {
3899 .readpage = ext4_readpage,
3900 .readpages = ext4_readpages,
3901 .writepage = ext4_writepage,
3902 .writepages = ext4_writepages,
3903 .write_begin = ext4_write_begin,
3904 .write_end = ext4_write_end,
3905 .set_page_dirty = ext4_set_page_dirty,
3906 .bmap = ext4_bmap,
3907 .invalidatepage = ext4_invalidatepage,
3908 .releasepage = ext4_releasepage,
3909 .direct_IO = ext4_direct_IO,
3910 .migratepage = buffer_migrate_page,
3911 .is_partially_uptodate = block_is_partially_uptodate,
3912 .error_remove_page = generic_error_remove_page,
3913};
3914
3915static const struct address_space_operations ext4_journalled_aops = {
3916 .readpage = ext4_readpage,
3917 .readpages = ext4_readpages,
3918 .writepage = ext4_writepage,
3919 .writepages = ext4_writepages,
3920 .write_begin = ext4_write_begin,
3921 .write_end = ext4_journalled_write_end,
3922 .set_page_dirty = ext4_journalled_set_page_dirty,
3923 .bmap = ext4_bmap,
3924 .invalidatepage = ext4_journalled_invalidatepage,
3925 .releasepage = ext4_releasepage,
3926 .direct_IO = ext4_direct_IO,
3927 .is_partially_uptodate = block_is_partially_uptodate,
3928 .error_remove_page = generic_error_remove_page,
3929};
3930
3931static const struct address_space_operations ext4_da_aops = {
3932 .readpage = ext4_readpage,
3933 .readpages = ext4_readpages,
3934 .writepage = ext4_writepage,
3935 .writepages = ext4_writepages,
3936 .write_begin = ext4_da_write_begin,
3937 .write_end = ext4_da_write_end,
3938 .set_page_dirty = ext4_set_page_dirty,
3939 .bmap = ext4_bmap,
3940 .invalidatepage = ext4_da_invalidatepage,
3941 .releasepage = ext4_releasepage,
3942 .direct_IO = ext4_direct_IO,
3943 .migratepage = buffer_migrate_page,
3944 .is_partially_uptodate = block_is_partially_uptodate,
3945 .error_remove_page = generic_error_remove_page,
3946};
3947
3948static const struct address_space_operations ext4_dax_aops = {
3949 .writepages = ext4_dax_writepages,
3950 .direct_IO = noop_direct_IO,
3951 .set_page_dirty = noop_set_page_dirty,
3952 .invalidatepage = noop_invalidatepage,
3953};
3954
3955void ext4_set_aops(struct inode *inode)
3956{
3957 switch (ext4_inode_journal_mode(inode)) {
3958 case EXT4_INODE_ORDERED_DATA_MODE:
3959 case EXT4_INODE_WRITEBACK_DATA_MODE:
3960 break;
3961 case EXT4_INODE_JOURNAL_DATA_MODE:
3962 inode->i_mapping->a_ops = &ext4_journalled_aops;
3963 return;
3964 default:
3965 BUG();
3966 }
3967 if (IS_DAX(inode))
3968 inode->i_mapping->a_ops = &ext4_dax_aops;
3969 else if (test_opt(inode->i_sb, DELALLOC))
3970 inode->i_mapping->a_ops = &ext4_da_aops;
3971 else
3972 inode->i_mapping->a_ops = &ext4_aops;
3973}
3974
3975static int __ext4_block_zero_page_range(handle_t *handle,
3976 struct address_space *mapping, loff_t from, loff_t length)
3977{
3978 ext4_fsblk_t index = from >> PAGE_SHIFT;
3979 unsigned offset = from & (PAGE_SIZE-1);
3980 unsigned blocksize, pos;
3981 ext4_lblk_t iblock;
3982 struct inode *inode = mapping->host;
3983 struct buffer_head *bh;
3984 struct page *page;
3985 int err = 0;
3986
3987 page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3988 mapping_gfp_constraint(mapping, ~__GFP_FS));
3989 if (!page)
3990 return -ENOMEM;
3991
3992 blocksize = inode->i_sb->s_blocksize;
3993
3994 iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3995
3996 if (!page_has_buffers(page))
3997 create_empty_buffers(page, blocksize, 0);
3998
3999 /* Find the buffer that contains "offset" */
4000 bh = page_buffers(page);
4001 pos = blocksize;
4002 while (offset >= pos) {
4003 bh = bh->b_this_page;
4004 iblock++;
4005 pos += blocksize;
4006 }
4007 if (buffer_freed(bh)) {
4008 BUFFER_TRACE(bh, "freed: skip");
4009 goto unlock;
4010 }
4011 if (!buffer_mapped(bh)) {
4012 BUFFER_TRACE(bh, "unmapped");
4013 ext4_get_block(inode, iblock, bh, 0);
4014 /* unmapped? It's a hole - nothing to do */
4015 if (!buffer_mapped(bh)) {
4016 BUFFER_TRACE(bh, "still unmapped");
4017 goto unlock;
4018 }
4019 }
4020
4021 /* Ok, it's mapped. Make sure it's up-to-date */
4022 if (PageUptodate(page))
4023 set_buffer_uptodate(bh);
4024
4025 if (!buffer_uptodate(bh)) {
4026 err = -EIO;
4027 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
4028 wait_on_buffer(bh);
4029 /* Uhhuh. Read error. Complain and punt. */
4030 if (!buffer_uptodate(bh))
4031 goto unlock;
4032 if (S_ISREG(inode->i_mode) &&
4033 ext4_encrypted_inode(inode)) {
4034 /* We expect the key to be set. */
4035 BUG_ON(!fscrypt_has_encryption_key(inode));
4036 BUG_ON(blocksize != PAGE_SIZE);
4037 WARN_ON_ONCE(fscrypt_decrypt_page(page->mapping->host,
4038 page, PAGE_SIZE, 0, page->index));
4039 }
4040 }
4041 if (ext4_should_journal_data(inode)) {
4042 BUFFER_TRACE(bh, "get write access");
4043 err = ext4_journal_get_write_access(handle, bh);
4044 if (err)
4045 goto unlock;
4046 }
4047 zero_user(page, offset, length);
4048 BUFFER_TRACE(bh, "zeroed end of block");
4049
4050 if (ext4_should_journal_data(inode)) {
4051 err = ext4_handle_dirty_metadata(handle, inode, bh);
4052 } else {
4053 err = 0;
4054 mark_buffer_dirty(bh);
4055 if (ext4_should_order_data(inode))
4056 err = ext4_jbd2_inode_add_write(handle, inode);
4057 }
4058
4059unlock:
4060 unlock_page(page);
4061 put_page(page);
4062 return err;
4063}
4064
4065/*
4066 * ext4_block_zero_page_range() zeros out a mapping of length 'length'
4067 * starting from file offset 'from'. The range to be zero'd must
4068 * be contained with in one block. If the specified range exceeds
4069 * the end of the block it will be shortened to end of the block
4070 * that cooresponds to 'from'
4071 */
4072static int ext4_block_zero_page_range(handle_t *handle,
4073 struct address_space *mapping, loff_t from, loff_t length)
4074{
4075 struct inode *inode = mapping->host;
4076 unsigned offset = from & (PAGE_SIZE-1);
4077 unsigned blocksize = inode->i_sb->s_blocksize;
4078 unsigned max = blocksize - (offset & (blocksize - 1));
4079
4080 /*
4081 * correct length if it does not fall between
4082 * 'from' and the end of the block
4083 */
4084 if (length > max || length < 0)
4085 length = max;
4086
4087 if (IS_DAX(inode)) {
4088 return iomap_zero_range(inode, from, length, NULL,
4089 &ext4_iomap_ops);
4090 }
4091 return __ext4_block_zero_page_range(handle, mapping, from, length);
4092}
4093
4094/*
4095 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4096 * up to the end of the block which corresponds to `from'.
4097 * This required during truncate. We need to physically zero the tail end
4098 * of that block so it doesn't yield old data if the file is later grown.
4099 */
4100static int ext4_block_truncate_page(handle_t *handle,
4101 struct address_space *mapping, loff_t from)
4102{
4103 unsigned offset = from & (PAGE_SIZE-1);
4104 unsigned length;
4105 unsigned blocksize;
4106 struct inode *inode = mapping->host;
4107
4108 /* If we are processing an encrypted inode during orphan list handling */
4109 if (ext4_encrypted_inode(inode) && !fscrypt_has_encryption_key(inode))
4110 return 0;
4111
4112 blocksize = inode->i_sb->s_blocksize;
4113 length = blocksize - (offset & (blocksize - 1));
4114
4115 return ext4_block_zero_page_range(handle, mapping, from, length);
4116}
4117
4118int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
4119 loff_t lstart, loff_t length)
4120{
4121 struct super_block *sb = inode->i_sb;
4122 struct address_space *mapping = inode->i_mapping;
4123 unsigned partial_start, partial_end;
4124 ext4_fsblk_t start, end;
4125 loff_t byte_end = (lstart + length - 1);
4126 int err = 0;
4127
4128 partial_start = lstart & (sb->s_blocksize - 1);
4129 partial_end = byte_end & (sb->s_blocksize - 1);
4130
4131 start = lstart >> sb->s_blocksize_bits;
4132 end = byte_end >> sb->s_blocksize_bits;
4133
4134 /* Handle partial zero within the single block */
4135 if (start == end &&
4136 (partial_start || (partial_end != sb->s_blocksize - 1))) {
4137 err = ext4_block_zero_page_range(handle, mapping,
4138 lstart, length);
4139 return err;
4140 }
4141 /* Handle partial zero out on the start of the range */
4142 if (partial_start) {
4143 err = ext4_block_zero_page_range(handle, mapping,
4144 lstart, sb->s_blocksize);
4145 if (err)
4146 return err;
4147 }
4148 /* Handle partial zero out on the end of the range */
4149 if (partial_end != sb->s_blocksize - 1)
4150 err = ext4_block_zero_page_range(handle, mapping,
4151 byte_end - partial_end,
4152 partial_end + 1);
4153 return err;
4154}
4155
4156int ext4_can_truncate(struct inode *inode)
4157{
4158 if (S_ISREG(inode->i_mode))
4159 return 1;
4160 if (S_ISDIR(inode->i_mode))
4161 return 1;
4162 if (S_ISLNK(inode->i_mode))
4163 return !ext4_inode_is_fast_symlink(inode);
4164 return 0;
4165}
4166
4167/*
4168 * We have to make sure i_disksize gets properly updated before we truncate
4169 * page cache due to hole punching or zero range. Otherwise i_disksize update
4170 * can get lost as it may have been postponed to submission of writeback but
4171 * that will never happen after we truncate page cache.
4172 */
4173int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
4174 loff_t len)
4175{
4176 handle_t *handle;
4177 loff_t size = i_size_read(inode);
4178
4179 WARN_ON(!inode_is_locked(inode));
4180 if (offset > size || offset + len < size)
4181 return 0;
4182
4183 if (EXT4_I(inode)->i_disksize >= size)
4184 return 0;
4185
4186 handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
4187 if (IS_ERR(handle))
4188 return PTR_ERR(handle);
4189 ext4_update_i_disksize(inode, size);
4190 ext4_mark_inode_dirty(handle, inode);
4191 ext4_journal_stop(handle);
4192
4193 return 0;
4194}
4195
4196/*
4197 * ext4_punch_hole: punches a hole in a file by releasing the blocks
4198 * associated with the given offset and length
4199 *
4200 * @inode: File inode
4201 * @offset: The offset where the hole will begin
4202 * @len: The length of the hole
4203 *
4204 * Returns: 0 on success or negative on failure
4205 */
4206
4207int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
4208{
4209 struct super_block *sb = inode->i_sb;
4210 ext4_lblk_t first_block, stop_block;
4211 struct address_space *mapping = inode->i_mapping;
4212 loff_t first_block_offset, last_block_offset;
4213 handle_t *handle;
4214 unsigned int credits;
4215 int ret = 0;
4216
4217 if (!S_ISREG(inode->i_mode))
4218 return -EOPNOTSUPP;
4219
4220 trace_ext4_punch_hole(inode, offset, length, 0);
4221
4222 /*
4223 * Write out all dirty pages to avoid race conditions
4224 * Then release them.
4225 */
4226 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4227 ret = filemap_write_and_wait_range(mapping, offset,
4228 offset + length - 1);
4229 if (ret)
4230 return ret;
4231 }
4232
4233 inode_lock(inode);
4234
4235 /* No need to punch hole beyond i_size */
4236 if (offset >= inode->i_size)
4237 goto out_mutex;
4238
4239 /*
4240 * If the hole extends beyond i_size, set the hole
4241 * to end after the page that contains i_size
4242 */
4243 if (offset + length > inode->i_size) {
4244 length = inode->i_size +
4245 PAGE_SIZE - (inode->i_size & (PAGE_SIZE - 1)) -
4246 offset;
4247 }
4248
4249 if (offset & (sb->s_blocksize - 1) ||
4250 (offset + length) & (sb->s_blocksize - 1)) {
4251 /*
4252 * Attach jinode to inode for jbd2 if we do any zeroing of
4253 * partial block
4254 */
4255 ret = ext4_inode_attach_jinode(inode);
4256 if (ret < 0)
4257 goto out_mutex;
4258
4259 }
4260
4261 /* Wait all existing dio workers, newcomers will block on i_mutex */
4262 inode_dio_wait(inode);
4263
4264 /*
4265 * Prevent page faults from reinstantiating pages we have released from
4266 * page cache.
4267 */
4268 down_write(&EXT4_I(inode)->i_mmap_sem);
4269 first_block_offset = round_up(offset, sb->s_blocksize);
4270 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
4271
4272 /* Now release the pages and zero block aligned part of pages*/
4273 if (last_block_offset > first_block_offset) {
4274 ret = ext4_update_disksize_before_punch(inode, offset, length);
4275 if (ret)
4276 goto out_dio;
4277 truncate_pagecache_range(inode, first_block_offset,
4278 last_block_offset);
4279 }
4280
4281 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4282 credits = ext4_writepage_trans_blocks(inode);
4283 else
4284 credits = ext4_blocks_for_truncate(inode);
4285 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4286 if (IS_ERR(handle)) {
4287 ret = PTR_ERR(handle);
4288 ext4_std_error(sb, ret);
4289 goto out_dio;
4290 }
4291
4292 ret = ext4_zero_partial_blocks(handle, inode, offset,
4293 length);
4294 if (ret)
4295 goto out_stop;
4296
4297 first_block = (offset + sb->s_blocksize - 1) >>
4298 EXT4_BLOCK_SIZE_BITS(sb);
4299 stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4300
4301 /* If there are no blocks to remove, return now */
4302 if (first_block >= stop_block)
4303 goto out_stop;
4304
4305 down_write(&EXT4_I(inode)->i_data_sem);
4306 ext4_discard_preallocations(inode);
4307
4308 ret = ext4_es_remove_extent(inode, first_block,
4309 stop_block - first_block);
4310 if (ret) {
4311 up_write(&EXT4_I(inode)->i_data_sem);
4312 goto out_stop;
4313 }
4314
4315 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4316 ret = ext4_ext_remove_space(inode, first_block,
4317 stop_block - 1);
4318 else
4319 ret = ext4_ind_remove_space(handle, inode, first_block,
4320 stop_block);
4321
4322 up_write(&EXT4_I(inode)->i_data_sem);
4323 if (IS_SYNC(inode))
4324 ext4_handle_sync(handle);
4325
4326 inode->i_mtime = inode->i_ctime = current_time(inode);
4327 ext4_mark_inode_dirty(handle, inode);
4328 if (ret >= 0)
4329 ext4_update_inode_fsync_trans(handle, inode, 1);
4330out_stop:
4331 ext4_journal_stop(handle);
4332out_dio:
4333 up_write(&EXT4_I(inode)->i_mmap_sem);
4334out_mutex:
4335 inode_unlock(inode);
4336 return ret;
4337}
4338
4339int ext4_inode_attach_jinode(struct inode *inode)
4340{
4341 struct ext4_inode_info *ei = EXT4_I(inode);
4342 struct jbd2_inode *jinode;
4343
4344 if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
4345 return 0;
4346
4347 jinode = jbd2_alloc_inode(GFP_KERNEL);
4348 spin_lock(&inode->i_lock);
4349 if (!ei->jinode) {
4350 if (!jinode) {
4351 spin_unlock(&inode->i_lock);
4352 return -ENOMEM;
4353 }
4354 ei->jinode = jinode;
4355 jbd2_journal_init_jbd_inode(ei->jinode, inode);
4356 jinode = NULL;
4357 }
4358 spin_unlock(&inode->i_lock);
4359 if (unlikely(jinode != NULL))
4360 jbd2_free_inode(jinode);
4361 return 0;
4362}
4363
4364/*
4365 * ext4_truncate()
4366 *
4367 * We block out ext4_get_block() block instantiations across the entire
4368 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4369 * simultaneously on behalf of the same inode.
4370 *
4371 * As we work through the truncate and commit bits of it to the journal there
4372 * is one core, guiding principle: the file's tree must always be consistent on
4373 * disk. We must be able to restart the truncate after a crash.
4374 *
4375 * The file's tree may be transiently inconsistent in memory (although it
4376 * probably isn't), but whenever we close off and commit a journal transaction,
4377 * the contents of (the filesystem + the journal) must be consistent and
4378 * restartable. It's pretty simple, really: bottom up, right to left (although
4379 * left-to-right works OK too).
4380 *
4381 * Note that at recovery time, journal replay occurs *before* the restart of
4382 * truncate against the orphan inode list.
4383 *
4384 * The committed inode has the new, desired i_size (which is the same as
4385 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
4386 * that this inode's truncate did not complete and it will again call
4387 * ext4_truncate() to have another go. So there will be instantiated blocks
4388 * to the right of the truncation point in a crashed ext4 filesystem. But
4389 * that's fine - as long as they are linked from the inode, the post-crash
4390 * ext4_truncate() run will find them and release them.
4391 */
4392int ext4_truncate(struct inode *inode)
4393{
4394 struct ext4_inode_info *ei = EXT4_I(inode);
4395 unsigned int credits;
4396 int err = 0;
4397 handle_t *handle;
4398 struct address_space *mapping = inode->i_mapping;
4399
4400 /*
4401 * There is a possibility that we're either freeing the inode
4402 * or it's a completely new inode. In those cases we might not
4403 * have i_mutex locked because it's not necessary.
4404 */
4405 if (!(inode->i_state & (I_NEW|I_FREEING)))
4406 WARN_ON(!inode_is_locked(inode));
4407 trace_ext4_truncate_enter(inode);
4408
4409 if (!ext4_can_truncate(inode))
4410 return 0;
4411
4412 ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4413
4414 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4415 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4416
4417 if (ext4_has_inline_data(inode)) {
4418 int has_inline = 1;
4419
4420 err = ext4_inline_data_truncate(inode, &has_inline);
4421 if (err)
4422 return err;
4423 if (has_inline)
4424 return 0;
4425 }
4426
4427 /* If we zero-out tail of the page, we have to create jinode for jbd2 */
4428 if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
4429 if (ext4_inode_attach_jinode(inode) < 0)
4430 return 0;
4431 }
4432
4433 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4434 credits = ext4_writepage_trans_blocks(inode);
4435 else
4436 credits = ext4_blocks_for_truncate(inode);
4437
4438 handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4439 if (IS_ERR(handle))
4440 return PTR_ERR(handle);
4441
4442 if (inode->i_size & (inode->i_sb->s_blocksize - 1))
4443 ext4_block_truncate_page(handle, mapping, inode->i_size);
4444
4445 /*
4446 * We add the inode to the orphan list, so that if this
4447 * truncate spans multiple transactions, and we crash, we will
4448 * resume the truncate when the filesystem recovers. It also
4449 * marks the inode dirty, to catch the new size.
4450 *
4451 * Implication: the file must always be in a sane, consistent
4452 * truncatable state while each transaction commits.
4453 */
4454 err = ext4_orphan_add(handle, inode);
4455 if (err)
4456 goto out_stop;
4457
4458 down_write(&EXT4_I(inode)->i_data_sem);
4459
4460 ext4_discard_preallocations(inode);
4461
4462 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4463 err = ext4_ext_truncate(handle, inode);
4464 else
4465 ext4_ind_truncate(handle, inode);
4466
4467 up_write(&ei->i_data_sem);
4468 if (err)
4469 goto out_stop;
4470
4471 if (IS_SYNC(inode))
4472 ext4_handle_sync(handle);
4473
4474out_stop:
4475 /*
4476 * If this was a simple ftruncate() and the file will remain alive,
4477 * then we need to clear up the orphan record which we created above.
4478 * However, if this was a real unlink then we were called by
4479 * ext4_evict_inode(), and we allow that function to clean up the
4480 * orphan info for us.
4481 */
4482 if (inode->i_nlink)
4483 ext4_orphan_del(handle, inode);
4484
4485 inode->i_mtime = inode->i_ctime = current_time(inode);
4486 ext4_mark_inode_dirty(handle, inode);
4487 ext4_journal_stop(handle);
4488
4489 trace_ext4_truncate_exit(inode);
4490 return err;
4491}
4492
4493/*
4494 * ext4_get_inode_loc returns with an extra refcount against the inode's
4495 * underlying buffer_head on success. If 'in_mem' is true, we have all
4496 * data in memory that is needed to recreate the on-disk version of this
4497 * inode.
4498 */
4499static int __ext4_get_inode_loc(struct inode *inode,
4500 struct ext4_iloc *iloc, int in_mem)
4501{
4502 struct ext4_group_desc *gdp;
4503 struct buffer_head *bh;
4504 struct super_block *sb = inode->i_sb;
4505 ext4_fsblk_t block;
4506 int inodes_per_block, inode_offset;
4507
4508 iloc->bh = NULL;
4509 if (!ext4_valid_inum(sb, inode->i_ino))
4510 return -EFSCORRUPTED;
4511
4512 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4513 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4514 if (!gdp)
4515 return -EIO;
4516
4517 /*
4518 * Figure out the offset within the block group inode table
4519 */
4520 inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4521 inode_offset = ((inode->i_ino - 1) %
4522 EXT4_INODES_PER_GROUP(sb));
4523 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4524 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4525
4526 bh = sb_getblk(sb, block);
4527 if (unlikely(!bh))
4528 return -ENOMEM;
4529 if (!buffer_uptodate(bh)) {
4530 lock_buffer(bh);
4531
4532 /*
4533 * If the buffer has the write error flag, we have failed
4534 * to write out another inode in the same block. In this
4535 * case, we don't have to read the block because we may
4536 * read the old inode data successfully.
4537 */
4538 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4539 set_buffer_uptodate(bh);
4540
4541 if (buffer_uptodate(bh)) {
4542 /* someone brought it uptodate while we waited */
4543 unlock_buffer(bh);
4544 goto has_buffer;
4545 }
4546
4547 /*
4548 * If we have all information of the inode in memory and this
4549 * is the only valid inode in the block, we need not read the
4550 * block.
4551 */
4552 if (in_mem) {
4553 struct buffer_head *bitmap_bh;
4554 int i, start;
4555
4556 start = inode_offset & ~(inodes_per_block - 1);
4557
4558 /* Is the inode bitmap in cache? */
4559 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4560 if (unlikely(!bitmap_bh))
4561 goto make_io;
4562
4563 /*
4564 * If the inode bitmap isn't in cache then the
4565 * optimisation may end up performing two reads instead
4566 * of one, so skip it.
4567 */
4568 if (!buffer_uptodate(bitmap_bh)) {
4569 brelse(bitmap_bh);
4570 goto make_io;
4571 }
4572 for (i = start; i < start + inodes_per_block; i++) {
4573 if (i == inode_offset)
4574 continue;
4575 if (ext4_test_bit(i, bitmap_bh->b_data))
4576 break;
4577 }
4578 brelse(bitmap_bh);
4579 if (i == start + inodes_per_block) {
4580 /* all other inodes are free, so skip I/O */
4581 memset(bh->b_data, 0, bh->b_size);
4582 set_buffer_uptodate(bh);
4583 unlock_buffer(bh);
4584 goto has_buffer;
4585 }
4586 }
4587
4588make_io:
4589 /*
4590 * If we need to do any I/O, try to pre-readahead extra
4591 * blocks from the inode table.
4592 */
4593 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4594 ext4_fsblk_t b, end, table;
4595 unsigned num;
4596 __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
4597
4598 table = ext4_inode_table(sb, gdp);
4599 /* s_inode_readahead_blks is always a power of 2 */
4600 b = block & ~((ext4_fsblk_t) ra_blks - 1);
4601 if (table > b)
4602 b = table;
4603 end = b + ra_blks;
4604 num = EXT4_INODES_PER_GROUP(sb);
4605 if (ext4_has_group_desc_csum(sb))
4606 num -= ext4_itable_unused_count(sb, gdp);
4607 table += num / inodes_per_block;
4608 if (end > table)
4609 end = table;
4610 while (b <= end)
4611 sb_breadahead(sb, b++);
4612 }
4613
4614 /*
4615 * There are other valid inodes in the buffer, this inode
4616 * has in-inode xattrs, or we don't have this inode in memory.
4617 * Read the block from disk.
4618 */
4619 trace_ext4_load_inode(inode);
4620 get_bh(bh);
4621 bh->b_end_io = end_buffer_read_sync;
4622 submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
4623 wait_on_buffer(bh);
4624 if (!buffer_uptodate(bh)) {
4625 EXT4_ERROR_INODE_BLOCK(inode, block,
4626 "unable to read itable block");
4627 brelse(bh);
4628 return -EIO;
4629 }
4630 }
4631has_buffer:
4632 iloc->bh = bh;
4633 return 0;
4634}
4635
4636int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4637{
4638 /* We have all inode data except xattrs in memory here. */
4639 return __ext4_get_inode_loc(inode, iloc,
4640 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4641}
4642
4643static bool ext4_should_use_dax(struct inode *inode)
4644{
4645 if (!test_opt(inode->i_sb, DAX))
4646 return false;
4647 if (!S_ISREG(inode->i_mode))
4648 return false;
4649 if (ext4_should_journal_data(inode))
4650 return false;
4651 if (ext4_has_inline_data(inode))
4652 return false;
4653 if (ext4_encrypted_inode(inode))
4654 return false;
4655 return true;
4656}
4657
4658void ext4_set_inode_flags(struct inode *inode)
4659{
4660 unsigned int flags = EXT4_I(inode)->i_flags;
4661 unsigned int new_fl = 0;
4662
4663 if (flags & EXT4_SYNC_FL)
4664 new_fl |= S_SYNC;
4665 if (flags & EXT4_APPEND_FL)
4666 new_fl |= S_APPEND;
4667 if (flags & EXT4_IMMUTABLE_FL)
4668 new_fl |= S_IMMUTABLE;
4669 if (flags & EXT4_NOATIME_FL)
4670 new_fl |= S_NOATIME;
4671 if (flags & EXT4_DIRSYNC_FL)
4672 new_fl |= S_DIRSYNC;
4673 if (ext4_should_use_dax(inode))
4674 new_fl |= S_DAX;
4675 if (flags & EXT4_ENCRYPT_FL)
4676 new_fl |= S_ENCRYPTED;
4677 inode_set_flags(inode, new_fl,
4678 S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX|
4679 S_ENCRYPTED);
4680}
4681
4682static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4683 struct ext4_inode_info *ei)
4684{
4685 blkcnt_t i_blocks ;
4686 struct inode *inode = &(ei->vfs_inode);
4687 struct super_block *sb = inode->i_sb;
4688
4689 if (ext4_has_feature_huge_file(sb)) {
4690 /* we are using combined 48 bit field */
4691 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4692 le32_to_cpu(raw_inode->i_blocks_lo);
4693 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4694 /* i_blocks represent file system block size */
4695 return i_blocks << (inode->i_blkbits - 9);
4696 } else {
4697 return i_blocks;
4698 }
4699 } else {
4700 return le32_to_cpu(raw_inode->i_blocks_lo);
4701 }
4702}
4703
4704static inline void ext4_iget_extra_inode(struct inode *inode,
4705 struct ext4_inode *raw_inode,
4706 struct ext4_inode_info *ei)
4707{
4708 __le32 *magic = (void *)raw_inode +
4709 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4710 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize + sizeof(__le32) <=
4711 EXT4_INODE_SIZE(inode->i_sb) &&
4712 *magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4713 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4714 ext4_find_inline_data_nolock(inode);
4715 } else
4716 EXT4_I(inode)->i_inline_off = 0;
4717}
4718
4719int ext4_get_projid(struct inode *inode, kprojid_t *projid)
4720{
4721 if (!ext4_has_feature_project(inode->i_sb))
4722 return -EOPNOTSUPP;
4723 *projid = EXT4_I(inode)->i_projid;
4724 return 0;
4725}
4726
4727struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4728{
4729 struct ext4_iloc iloc;
4730 struct ext4_inode *raw_inode;
4731 struct ext4_inode_info *ei;
4732 struct inode *inode;
4733 journal_t *journal = EXT4_SB(sb)->s_journal;
4734 long ret;
4735 loff_t size;
4736 int block;
4737 uid_t i_uid;
4738 gid_t i_gid;
4739 projid_t i_projid;
4740
4741 inode = iget_locked(sb, ino);
4742 if (!inode)
4743 return ERR_PTR(-ENOMEM);
4744 if (!(inode->i_state & I_NEW))
4745 return inode;
4746
4747 ei = EXT4_I(inode);
4748 iloc.bh = NULL;
4749
4750 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4751 if (ret < 0)
4752 goto bad_inode;
4753 raw_inode = ext4_raw_inode(&iloc);
4754
4755 if ((ino == EXT4_ROOT_INO) && (raw_inode->i_links_count == 0)) {
4756 EXT4_ERROR_INODE(inode, "root inode unallocated");
4757 ret = -EFSCORRUPTED;
4758 goto bad_inode;
4759 }
4760
4761 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4762 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4763 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4764 EXT4_INODE_SIZE(inode->i_sb) ||
4765 (ei->i_extra_isize & 3)) {
4766 EXT4_ERROR_INODE(inode,
4767 "bad extra_isize %u (inode size %u)",
4768 ei->i_extra_isize,
4769 EXT4_INODE_SIZE(inode->i_sb));
4770 ret = -EFSCORRUPTED;
4771 goto bad_inode;
4772 }
4773 } else
4774 ei->i_extra_isize = 0;
4775
4776 /* Precompute checksum seed for inode metadata */
4777 if (ext4_has_metadata_csum(sb)) {
4778 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4779 __u32 csum;
4780 __le32 inum = cpu_to_le32(inode->i_ino);
4781 __le32 gen = raw_inode->i_generation;
4782 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4783 sizeof(inum));
4784 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4785 sizeof(gen));
4786 }
4787
4788 if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4789 EXT4_ERROR_INODE(inode, "checksum invalid");
4790 ret = -EFSBADCRC;
4791 goto bad_inode;
4792 }
4793
4794 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4795 i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4796 i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4797 if (ext4_has_feature_project(sb) &&
4798 EXT4_INODE_SIZE(sb) > EXT4_GOOD_OLD_INODE_SIZE &&
4799 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
4800 i_projid = (projid_t)le32_to_cpu(raw_inode->i_projid);
4801 else
4802 i_projid = EXT4_DEF_PROJID;
4803
4804 if (!(test_opt(inode->i_sb, NO_UID32))) {
4805 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4806 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4807 }
4808 i_uid_write(inode, i_uid);
4809 i_gid_write(inode, i_gid);
4810 ei->i_projid = make_kprojid(&init_user_ns, i_projid);
4811 set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4812
4813 ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */
4814 ei->i_inline_off = 0;
4815 ei->i_dir_start_lookup = 0;
4816 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4817 /* We now have enough fields to check if the inode was active or not.
4818 * This is needed because nfsd might try to access dead inodes
4819 * the test is that same one that e2fsck uses
4820 * NeilBrown 1999oct15
4821 */
4822 if (inode->i_nlink == 0) {
4823 if ((inode->i_mode == 0 ||
4824 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4825 ino != EXT4_BOOT_LOADER_INO) {
4826 /* this inode is deleted */
4827 ret = -ESTALE;
4828 goto bad_inode;
4829 }
4830 /* The only unlinked inodes we let through here have
4831 * valid i_mode and are being read by the orphan
4832 * recovery code: that's fine, we're about to complete
4833 * the process of deleting those.
4834 * OR it is the EXT4_BOOT_LOADER_INO which is
4835 * not initialized on a new filesystem. */
4836 }
4837 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4838 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4839 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4840 if (ext4_has_feature_64bit(sb))
4841 ei->i_file_acl |=
4842 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4843 inode->i_size = ext4_isize(sb, raw_inode);
4844 if ((size = i_size_read(inode)) < 0) {
4845 EXT4_ERROR_INODE(inode, "bad i_size value: %lld", size);
4846 ret = -EFSCORRUPTED;
4847 goto bad_inode;
4848 }
4849 ei->i_disksize = inode->i_size;
4850#ifdef CONFIG_QUOTA
4851 ei->i_reserved_quota = 0;
4852#endif
4853 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4854 ei->i_block_group = iloc.block_group;
4855 ei->i_last_alloc_group = ~0;
4856 /*
4857 * NOTE! The in-memory inode i_data array is in little-endian order
4858 * even on big-endian machines: we do NOT byteswap the block numbers!
4859 */
4860 for (block = 0; block < EXT4_N_BLOCKS; block++)
4861 ei->i_data[block] = raw_inode->i_block[block];
4862 INIT_LIST_HEAD(&ei->i_orphan);
4863
4864 /*
4865 * Set transaction id's of transactions that have to be committed
4866 * to finish f[data]sync. We set them to currently running transaction
4867 * as we cannot be sure that the inode or some of its metadata isn't
4868 * part of the transaction - the inode could have been reclaimed and
4869 * now it is reread from disk.
4870 */
4871 if (journal) {
4872 transaction_t *transaction;
4873 tid_t tid;
4874
4875 read_lock(&journal->j_state_lock);
4876 if (journal->j_running_transaction)
4877 transaction = journal->j_running_transaction;
4878 else
4879 transaction = journal->j_committing_transaction;
4880 if (transaction)
4881 tid = transaction->t_tid;
4882 else
4883 tid = journal->j_commit_sequence;
4884 read_unlock(&journal->j_state_lock);
4885 ei->i_sync_tid = tid;
4886 ei->i_datasync_tid = tid;
4887 }
4888
4889 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4890 if (ei->i_extra_isize == 0) {
4891 /* The extra space is currently unused. Use it. */
4892 BUILD_BUG_ON(sizeof(struct ext4_inode) & 3);
4893 ei->i_extra_isize = sizeof(struct ext4_inode) -
4894 EXT4_GOOD_OLD_INODE_SIZE;
4895 } else {
4896 ext4_iget_extra_inode(inode, raw_inode, ei);
4897 }
4898 }
4899
4900 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4901 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4902 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4903 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4904
4905 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4906 u64 ivers = le32_to_cpu(raw_inode->i_disk_version);
4907
4908 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4909 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4910 ivers |=
4911 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4912 }
4913 inode_set_iversion_queried(inode, ivers);
4914 }
4915
4916 ret = 0;
4917 if (ei->i_file_acl &&
4918 !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4919 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4920 ei->i_file_acl);
4921 ret = -EFSCORRUPTED;
4922 goto bad_inode;
4923 } else if (!ext4_has_inline_data(inode)) {
4924 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4925 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4926 (S_ISLNK(inode->i_mode) &&
4927 !ext4_inode_is_fast_symlink(inode))))
4928 /* Validate extent which is part of inode */
4929 ret = ext4_ext_check_inode(inode);
4930 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4931 (S_ISLNK(inode->i_mode) &&
4932 !ext4_inode_is_fast_symlink(inode))) {
4933 /* Validate block references which are part of inode */
4934 ret = ext4_ind_check_inode(inode);
4935 }
4936 }
4937 if (ret)
4938 goto bad_inode;
4939
4940 if (S_ISREG(inode->i_mode)) {
4941 inode->i_op = &ext4_file_inode_operations;
4942 inode->i_fop = &ext4_file_operations;
4943 ext4_set_aops(inode);
4944 } else if (S_ISDIR(inode->i_mode)) {
4945 inode->i_op = &ext4_dir_inode_operations;
4946 inode->i_fop = &ext4_dir_operations;
4947 } else if (S_ISLNK(inode->i_mode)) {
4948 if (ext4_encrypted_inode(inode)) {
4949 inode->i_op = &ext4_encrypted_symlink_inode_operations;
4950 ext4_set_aops(inode);
4951 } else if (ext4_inode_is_fast_symlink(inode)) {
4952 inode->i_link = (char *)ei->i_data;
4953 inode->i_op = &ext4_fast_symlink_inode_operations;
4954 nd_terminate_link(ei->i_data, inode->i_size,
4955 sizeof(ei->i_data) - 1);
4956 } else {
4957 inode->i_op = &ext4_symlink_inode_operations;
4958 ext4_set_aops(inode);
4959 }
4960 inode_nohighmem(inode);
4961 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4962 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4963 inode->i_op = &ext4_special_inode_operations;
4964 if (raw_inode->i_block[0])
4965 init_special_inode(inode, inode->i_mode,
4966 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4967 else
4968 init_special_inode(inode, inode->i_mode,
4969 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4970 } else if (ino == EXT4_BOOT_LOADER_INO) {
4971 make_bad_inode(inode);
4972 } else {
4973 ret = -EFSCORRUPTED;
4974 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
4975 goto bad_inode;
4976 }
4977 brelse(iloc.bh);
4978 ext4_set_inode_flags(inode);
4979
4980 unlock_new_inode(inode);
4981 return inode;
4982
4983bad_inode:
4984 brelse(iloc.bh);
4985 iget_failed(inode);
4986 return ERR_PTR(ret);
4987}
4988
4989struct inode *ext4_iget_normal(struct super_block *sb, unsigned long ino)
4990{
4991 if (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)
4992 return ERR_PTR(-EFSCORRUPTED);
4993 return ext4_iget(sb, ino);
4994}
4995
4996static int ext4_inode_blocks_set(handle_t *handle,
4997 struct ext4_inode *raw_inode,
4998 struct ext4_inode_info *ei)
4999{
5000 struct inode *inode = &(ei->vfs_inode);
5001 u64 i_blocks = inode->i_blocks;
5002 struct super_block *sb = inode->i_sb;
5003
5004 if (i_blocks <= ~0U) {
5005 /*
5006 * i_blocks can be represented in a 32 bit variable
5007 * as multiple of 512 bytes
5008 */
5009 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5010 raw_inode->i_blocks_high = 0;
5011 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5012 return 0;
5013 }
5014 if (!ext4_has_feature_huge_file(sb))
5015 return -EFBIG;
5016
5017 if (i_blocks <= 0xffffffffffffULL) {
5018 /*
5019 * i_blocks can be represented in a 48 bit variable
5020 * as multiple of 512 bytes
5021 */
5022 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5023 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5024 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5025 } else {
5026 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5027 /* i_block is stored in file system block size */
5028 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5029 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
5030 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5031 }
5032 return 0;
5033}
5034
5035struct other_inode {
5036 unsigned long orig_ino;
5037 struct ext4_inode *raw_inode;
5038};
5039
5040static int other_inode_match(struct inode * inode, unsigned long ino,
5041 void *data)
5042{
5043 struct other_inode *oi = (struct other_inode *) data;
5044
5045 if ((inode->i_ino != ino) ||
5046 (inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5047 I_DIRTY_INODE)) ||
5048 ((inode->i_state & I_DIRTY_TIME) == 0))
5049 return 0;
5050 spin_lock(&inode->i_lock);
5051 if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
5052 I_DIRTY_INODE)) == 0) &&
5053 (inode->i_state & I_DIRTY_TIME)) {
5054 struct ext4_inode_info *ei = EXT4_I(inode);
5055
5056 inode->i_state &= ~(I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED);
5057 spin_unlock(&inode->i_lock);
5058
5059 spin_lock(&ei->i_raw_lock);
5060 EXT4_INODE_SET_XTIME(i_ctime, inode, oi->raw_inode);
5061 EXT4_INODE_SET_XTIME(i_mtime, inode, oi->raw_inode);
5062 EXT4_INODE_SET_XTIME(i_atime, inode, oi->raw_inode);
5063 ext4_inode_csum_set(inode, oi->raw_inode, ei);
5064 spin_unlock(&ei->i_raw_lock);
5065 trace_ext4_other_inode_update_time(inode, oi->orig_ino);
5066 return -1;
5067 }
5068 spin_unlock(&inode->i_lock);
5069 return -1;
5070}
5071
5072/*
5073 * Opportunistically update the other time fields for other inodes in
5074 * the same inode table block.
5075 */
5076static void ext4_update_other_inodes_time(struct super_block *sb,
5077 unsigned long orig_ino, char *buf)
5078{
5079 struct other_inode oi;
5080 unsigned long ino;
5081 int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
5082 int inode_size = EXT4_INODE_SIZE(sb);
5083
5084 oi.orig_ino = orig_ino;
5085 /*
5086 * Calculate the first inode in the inode table block. Inode
5087 * numbers are one-based. That is, the first inode in a block
5088 * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
5089 */
5090 ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
5091 for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
5092 if (ino == orig_ino)
5093 continue;
5094 oi.raw_inode = (struct ext4_inode *) buf;
5095 (void) find_inode_nowait(sb, ino, other_inode_match, &oi);
5096 }
5097}
5098
5099/*
5100 * Post the struct inode info into an on-disk inode location in the
5101 * buffer-cache. This gobbles the caller's reference to the
5102 * buffer_head in the inode location struct.
5103 *
5104 * The caller must have write access to iloc->bh.
5105 */
5106static int ext4_do_update_inode(handle_t *handle,
5107 struct inode *inode,
5108 struct ext4_iloc *iloc)
5109{
5110 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5111 struct ext4_inode_info *ei = EXT4_I(inode);
5112 struct buffer_head *bh = iloc->bh;
5113 struct super_block *sb = inode->i_sb;
5114 int err = 0, rc, block;
5115 int need_datasync = 0, set_large_file = 0;
5116 uid_t i_uid;
5117 gid_t i_gid;
5118 projid_t i_projid;
5119
5120 spin_lock(&ei->i_raw_lock);
5121
5122 /* For fields not tracked in the in-memory inode,
5123 * initialise them to zero for new inodes. */
5124 if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5125 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5126
5127 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5128 i_uid = i_uid_read(inode);
5129 i_gid = i_gid_read(inode);
5130 i_projid = from_kprojid(&init_user_ns, ei->i_projid);
5131 if (!(test_opt(inode->i_sb, NO_UID32))) {
5132 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
5133 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
5134/*
5135 * Fix up interoperability with old kernels. Otherwise, old inodes get
5136 * re-used with the upper 16 bits of the uid/gid intact
5137 */
5138 if (ei->i_dtime && list_empty(&ei->i_orphan)) {
5139 raw_inode->i_uid_high = 0;
5140 raw_inode->i_gid_high = 0;
5141 } else {
5142 raw_inode->i_uid_high =
5143 cpu_to_le16(high_16_bits(i_uid));
5144 raw_inode->i_gid_high =
5145 cpu_to_le16(high_16_bits(i_gid));
5146 }
5147 } else {
5148 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
5149 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
5150 raw_inode->i_uid_high = 0;
5151 raw_inode->i_gid_high = 0;
5152 }
5153 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5154
5155 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5156 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5157 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5158 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5159
5160 err = ext4_inode_blocks_set(handle, raw_inode, ei);
5161 if (err) {
5162 spin_unlock(&ei->i_raw_lock);
5163 goto out_brelse;
5164 }
5165 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5166 raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
5167 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
5168 raw_inode->i_file_acl_high =
5169 cpu_to_le16(ei->i_file_acl >> 32);
5170 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5171 if (ei->i_disksize != ext4_isize(inode->i_sb, raw_inode)) {
5172 ext4_isize_set(raw_inode, ei->i_disksize);
5173 need_datasync = 1;
5174 }
5175 if (ei->i_disksize > 0x7fffffffULL) {
5176 if (!ext4_has_feature_large_file(sb) ||
5177 EXT4_SB(sb)->s_es->s_rev_level ==
5178 cpu_to_le32(EXT4_GOOD_OLD_REV))
5179 set_large_file = 1;
5180 }
5181 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5182 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5183 if (old_valid_dev(inode->i_rdev)) {
5184 raw_inode->i_block[0] =
5185 cpu_to_le32(old_encode_dev(inode->i_rdev));
5186 raw_inode->i_block[1] = 0;
5187 } else {
5188 raw_inode->i_block[0] = 0;
5189 raw_inode->i_block[1] =
5190 cpu_to_le32(new_encode_dev(inode->i_rdev));
5191 raw_inode->i_block[2] = 0;
5192 }
5193 } else if (!ext4_has_inline_data(inode)) {
5194 for (block = 0; block < EXT4_N_BLOCKS; block++)
5195 raw_inode->i_block[block] = ei->i_data[block];
5196 }
5197
5198 if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
5199 u64 ivers = inode_peek_iversion(inode);
5200
5201 raw_inode->i_disk_version = cpu_to_le32(ivers);
5202 if (ei->i_extra_isize) {
5203 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5204 raw_inode->i_version_hi =
5205 cpu_to_le32(ivers >> 32);
5206 raw_inode->i_extra_isize =
5207 cpu_to_le16(ei->i_extra_isize);
5208 }
5209 }
5210
5211 BUG_ON(!ext4_has_feature_project(inode->i_sb) &&
5212 i_projid != EXT4_DEF_PROJID);
5213
5214 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
5215 EXT4_FITS_IN_INODE(raw_inode, ei, i_projid))
5216 raw_inode->i_projid = cpu_to_le32(i_projid);
5217
5218 ext4_inode_csum_set(inode, raw_inode, ei);
5219 spin_unlock(&ei->i_raw_lock);
5220 if (inode->i_sb->s_flags & SB_LAZYTIME)
5221 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
5222 bh->b_data);
5223
5224 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5225 rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5226 if (!err)
5227 err = rc;
5228 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5229 if (set_large_file) {
5230 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
5231 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
5232 if (err)
5233 goto out_brelse;
5234 ext4_update_dynamic_rev(sb);
5235 ext4_set_feature_large_file(sb);
5236 ext4_handle_sync(handle);
5237 err = ext4_handle_dirty_super(handle, sb);
5238 }
5239 ext4_update_inode_fsync_trans(handle, inode, need_datasync);
5240out_brelse:
5241 brelse(bh);
5242 ext4_std_error(inode->i_sb, err);
5243 return err;
5244}
5245
5246/*
5247 * ext4_write_inode()
5248 *
5249 * We are called from a few places:
5250 *
5251 * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
5252 * Here, there will be no transaction running. We wait for any running
5253 * transaction to commit.
5254 *
5255 * - Within flush work (sys_sync(), kupdate and such).
5256 * We wait on commit, if told to.
5257 *
5258 * - Within iput_final() -> write_inode_now()
5259 * We wait on commit, if told to.
5260 *
5261 * In all cases it is actually safe for us to return without doing anything,
5262 * because the inode has been copied into a raw inode buffer in
5263 * ext4_mark_inode_dirty(). This is a correctness thing for WB_SYNC_ALL
5264 * writeback.
5265 *
5266 * Note that we are absolutely dependent upon all inode dirtiers doing the
5267 * right thing: they *must* call mark_inode_dirty() after dirtying info in
5268 * which we are interested.
5269 *
5270 * It would be a bug for them to not do this. The code:
5271 *
5272 * mark_inode_dirty(inode)
5273 * stuff();
5274 * inode->i_size = expr;
5275 *
5276 * is in error because write_inode() could occur while `stuff()' is running,
5277 * and the new i_size will be lost. Plus the inode will no longer be on the
5278 * superblock's dirty inode list.
5279 */
5280int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5281{
5282 int err;
5283
5284 if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
5285 return 0;
5286
5287 if (EXT4_SB(inode->i_sb)->s_journal) {
5288 if (ext4_journal_current_handle()) {
5289 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5290 dump_stack();
5291 return -EIO;
5292 }
5293
5294 /*
5295 * No need to force transaction in WB_SYNC_NONE mode. Also
5296 * ext4_sync_fs() will force the commit after everything is
5297 * written.
5298 */
5299 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
5300 return 0;
5301
5302 err = ext4_force_commit(inode->i_sb);
5303 } else {
5304 struct ext4_iloc iloc;
5305
5306 err = __ext4_get_inode_loc(inode, &iloc, 0);
5307 if (err)
5308 return err;
5309 /*
5310 * sync(2) will flush the whole buffer cache. No need to do
5311 * it here separately for each inode.
5312 */
5313 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
5314 sync_dirty_buffer(iloc.bh);
5315 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5316 EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5317 "IO error syncing inode");
5318 err = -EIO;
5319 }
5320 brelse(iloc.bh);
5321 }
5322 return err;
5323}
5324
5325/*
5326 * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
5327 * buffers that are attached to a page stradding i_size and are undergoing
5328 * commit. In that case we have to wait for commit to finish and try again.
5329 */
5330static void ext4_wait_for_tail_page_commit(struct inode *inode)
5331{
5332 struct page *page;
5333 unsigned offset;
5334 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
5335 tid_t commit_tid = 0;
5336 int ret;
5337
5338 offset = inode->i_size & (PAGE_SIZE - 1);
5339 /*
5340 * All buffers in the last page remain valid? Then there's nothing to
5341 * do. We do the check mainly to optimize the common PAGE_SIZE ==
5342 * blocksize case
5343 */
5344 if (offset > PAGE_SIZE - i_blocksize(inode))
5345 return;
5346 while (1) {
5347 page = find_lock_page(inode->i_mapping,
5348 inode->i_size >> PAGE_SHIFT);
5349 if (!page)
5350 return;
5351 ret = __ext4_journalled_invalidatepage(page, offset,
5352 PAGE_SIZE - offset);
5353 unlock_page(page);
5354 put_page(page);
5355 if (ret != -EBUSY)
5356 return;
5357 commit_tid = 0;
5358 read_lock(&journal->j_state_lock);
5359 if (journal->j_committing_transaction)
5360 commit_tid = journal->j_committing_transaction->t_tid;
5361 read_unlock(&journal->j_state_lock);
5362 if (commit_tid)
5363 jbd2_log_wait_commit(journal, commit_tid);
5364 }
5365}
5366
5367/*
5368 * ext4_setattr()
5369 *
5370 * Called from notify_change.
5371 *
5372 * We want to trap VFS attempts to truncate the file as soon as
5373 * possible. In particular, we want to make sure that when the VFS
5374 * shrinks i_size, we put the inode on the orphan list and modify
5375 * i_disksize immediately, so that during the subsequent flushing of
5376 * dirty pages and freeing of disk blocks, we can guarantee that any
5377 * commit will leave the blocks being flushed in an unused state on
5378 * disk. (On recovery, the inode will get truncated and the blocks will
5379 * be freed, so we have a strong guarantee that no future commit will
5380 * leave these blocks visible to the user.)
5381 *
5382 * Another thing we have to assure is that if we are in ordered mode
5383 * and inode is still attached to the committing transaction, we must
5384 * we start writeout of all the dirty pages which are being truncated.
5385 * This way we are sure that all the data written in the previous
5386 * transaction are already on disk (truncate waits for pages under
5387 * writeback).
5388 *
5389 * Called with inode->i_mutex down.
5390 */
5391int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5392{
5393 struct inode *inode = d_inode(dentry);
5394 int error, rc = 0;
5395 int orphan = 0;
5396 const unsigned int ia_valid = attr->ia_valid;
5397
5398 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5399 return -EIO;
5400
5401 error = setattr_prepare(dentry, attr);
5402 if (error)
5403 return error;
5404
5405 error = fscrypt_prepare_setattr(dentry, attr);
5406 if (error)
5407 return error;
5408
5409 if (is_quota_modification(inode, attr)) {
5410 error = dquot_initialize(inode);
5411 if (error)
5412 return error;
5413 }
5414 if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
5415 (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
5416 handle_t *handle;
5417
5418 /* (user+group)*(old+new) structure, inode write (sb,
5419 * inode block, ? - but truncate inode update has it) */
5420 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
5421 (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
5422 EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
5423 if (IS_ERR(handle)) {
5424 error = PTR_ERR(handle);
5425 goto err_out;
5426 }
5427
5428 /* dquot_transfer() calls back ext4_get_inode_usage() which
5429 * counts xattr inode references.
5430 */
5431 down_read(&EXT4_I(inode)->xattr_sem);
5432 error = dquot_transfer(inode, attr);
5433 up_read(&EXT4_I(inode)->xattr_sem);
5434
5435 if (error) {
5436 ext4_journal_stop(handle);
5437 return error;
5438 }
5439 /* Update corresponding info in inode so that everything is in
5440 * one transaction */
5441 if (attr->ia_valid & ATTR_UID)
5442 inode->i_uid = attr->ia_uid;
5443 if (attr->ia_valid & ATTR_GID)
5444 inode->i_gid = attr->ia_gid;
5445 error = ext4_mark_inode_dirty(handle, inode);
5446 ext4_journal_stop(handle);
5447 }
5448
5449 if (attr->ia_valid & ATTR_SIZE) {
5450 handle_t *handle;
5451 loff_t oldsize = inode->i_size;
5452 int shrink = (attr->ia_size <= inode->i_size);
5453
5454 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5455 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5456
5457 if (attr->ia_size > sbi->s_bitmap_maxbytes)
5458 return -EFBIG;
5459 }
5460 if (!S_ISREG(inode->i_mode))
5461 return -EINVAL;
5462
5463 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
5464 inode_inc_iversion(inode);
5465
5466 if (ext4_should_order_data(inode) &&
5467 (attr->ia_size < inode->i_size)) {
5468 error = ext4_begin_ordered_truncate(inode,
5469 attr->ia_size);
5470 if (error)
5471 goto err_out;
5472 }
5473 if (attr->ia_size != inode->i_size) {
5474 handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
5475 if (IS_ERR(handle)) {
5476 error = PTR_ERR(handle);
5477 goto err_out;
5478 }
5479 if (ext4_handle_valid(handle) && shrink) {
5480 error = ext4_orphan_add(handle, inode);
5481 orphan = 1;
5482 }
5483 /*
5484 * Update c/mtime on truncate up, ext4_truncate() will
5485 * update c/mtime in shrink case below
5486 */
5487 if (!shrink) {
5488 inode->i_mtime = current_time(inode);
5489 inode->i_ctime = inode->i_mtime;
5490 }
5491 down_write(&EXT4_I(inode)->i_data_sem);
5492 EXT4_I(inode)->i_disksize = attr->ia_size;
5493 rc = ext4_mark_inode_dirty(handle, inode);
5494 if (!error)
5495 error = rc;
5496 /*
5497 * We have to update i_size under i_data_sem together
5498 * with i_disksize to avoid races with writeback code
5499 * running ext4_wb_update_i_disksize().
5500 */
5501 if (!error)
5502 i_size_write(inode, attr->ia_size);
5503 up_write(&EXT4_I(inode)->i_data_sem);
5504 ext4_journal_stop(handle);
5505 if (error) {
5506 if (orphan)
5507 ext4_orphan_del(NULL, inode);
5508 goto err_out;
5509 }
5510 }
5511 if (!shrink)
5512 pagecache_isize_extended(inode, oldsize, inode->i_size);
5513
5514 /*
5515 * Blocks are going to be removed from the inode. Wait
5516 * for dio in flight. Temporarily disable
5517 * dioread_nolock to prevent livelock.
5518 */
5519 if (orphan) {
5520 if (!ext4_should_journal_data(inode)) {
5521 inode_dio_wait(inode);
5522 } else
5523 ext4_wait_for_tail_page_commit(inode);
5524 }
5525 down_write(&EXT4_I(inode)->i_mmap_sem);
5526 /*
5527 * Truncate pagecache after we've waited for commit
5528 * in data=journal mode to make pages freeable.
5529 */
5530 truncate_pagecache(inode, inode->i_size);
5531 if (shrink) {
5532 rc = ext4_truncate(inode);
5533 if (rc)
5534 error = rc;
5535 }
5536 up_write(&EXT4_I(inode)->i_mmap_sem);
5537 }
5538
5539 if (!error) {
5540 setattr_copy(inode, attr);
5541 mark_inode_dirty(inode);
5542 }
5543
5544 /*
5545 * If the call to ext4_truncate failed to get a transaction handle at
5546 * all, we need to clean up the in-core orphan list manually.
5547 */
5548 if (orphan && inode->i_nlink)
5549 ext4_orphan_del(NULL, inode);
5550
5551 if (!error && (ia_valid & ATTR_MODE))
5552 rc = posix_acl_chmod(inode, inode->i_mode);
5553
5554err_out:
5555 ext4_std_error(inode->i_sb, error);
5556 if (!error)
5557 error = rc;
5558 return error;
5559}
5560
5561int ext4_getattr(const struct path *path, struct kstat *stat,
5562 u32 request_mask, unsigned int query_flags)
5563{
5564 struct inode *inode = d_inode(path->dentry);
5565 struct ext4_inode *raw_inode;
5566 struct ext4_inode_info *ei = EXT4_I(inode);
5567 unsigned int flags;
5568
5569 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_crtime)) {
5570 stat->result_mask |= STATX_BTIME;
5571 stat->btime.tv_sec = ei->i_crtime.tv_sec;
5572 stat->btime.tv_nsec = ei->i_crtime.tv_nsec;
5573 }
5574
5575 flags = ei->i_flags & EXT4_FL_USER_VISIBLE;
5576 if (flags & EXT4_APPEND_FL)
5577 stat->attributes |= STATX_ATTR_APPEND;
5578 if (flags & EXT4_COMPR_FL)
5579 stat->attributes |= STATX_ATTR_COMPRESSED;
5580 if (flags & EXT4_ENCRYPT_FL)
5581 stat->attributes |= STATX_ATTR_ENCRYPTED;
5582 if (flags & EXT4_IMMUTABLE_FL)
5583 stat->attributes |= STATX_ATTR_IMMUTABLE;
5584 if (flags & EXT4_NODUMP_FL)
5585 stat->attributes |= STATX_ATTR_NODUMP;
5586
5587 stat->attributes_mask |= (STATX_ATTR_APPEND |
5588 STATX_ATTR_COMPRESSED |
5589 STATX_ATTR_ENCRYPTED |
5590 STATX_ATTR_IMMUTABLE |
5591 STATX_ATTR_NODUMP);
5592
5593 generic_fillattr(inode, stat);
5594 return 0;
5595}
5596
5597int ext4_file_getattr(const struct path *path, struct kstat *stat,
5598 u32 request_mask, unsigned int query_flags)
5599{
5600 struct inode *inode = d_inode(path->dentry);
5601 u64 delalloc_blocks;
5602
5603 ext4_getattr(path, stat, request_mask, query_flags);
5604
5605 /*
5606 * If there is inline data in the inode, the inode will normally not
5607 * have data blocks allocated (it may have an external xattr block).
5608 * Report at least one sector for such files, so tools like tar, rsync,
5609 * others don't incorrectly think the file is completely sparse.
5610 */
5611 if (unlikely(ext4_has_inline_data(inode)))
5612 stat->blocks += (stat->size + 511) >> 9;
5613
5614 /*
5615 * We can't update i_blocks if the block allocation is delayed
5616 * otherwise in the case of system crash before the real block
5617 * allocation is done, we will have i_blocks inconsistent with
5618 * on-disk file blocks.
5619 * We always keep i_blocks updated together with real
5620 * allocation. But to not confuse with user, stat
5621 * will return the blocks that include the delayed allocation
5622 * blocks for this file.
5623 */
5624 delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
5625 EXT4_I(inode)->i_reserved_data_blocks);
5626 stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
5627 return 0;
5628}
5629
5630static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
5631 int pextents)
5632{
5633 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5634 return ext4_ind_trans_blocks(inode, lblocks);
5635 return ext4_ext_index_trans_blocks(inode, pextents);
5636}
5637
5638/*
5639 * Account for index blocks, block groups bitmaps and block group
5640 * descriptor blocks if modify datablocks and index blocks
5641 * worse case, the indexs blocks spread over different block groups
5642 *
5643 * If datablocks are discontiguous, they are possible to spread over
5644 * different block groups too. If they are contiguous, with flexbg,
5645 * they could still across block group boundary.
5646 *
5647 * Also account for superblock, inode, quota and xattr blocks
5648 */
5649static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
5650 int pextents)
5651{
5652 ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5653 int gdpblocks;
5654 int idxblocks;
5655 int ret = 0;
5656
5657 /*
5658 * How many index blocks need to touch to map @lblocks logical blocks
5659 * to @pextents physical extents?
5660 */
5661 idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
5662
5663 ret = idxblocks;
5664
5665 /*
5666 * Now let's see how many group bitmaps and group descriptors need
5667 * to account
5668 */
5669 groups = idxblocks + pextents;
5670 gdpblocks = groups;
5671 if (groups > ngroups)
5672 groups = ngroups;
5673 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5674 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5675
5676 /* bitmaps and block group descriptor blocks */
5677 ret += groups + gdpblocks;
5678
5679 /* Blocks for super block, inode, quota and xattr blocks */
5680 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5681
5682 return ret;
5683}
5684
5685/*
5686 * Calculate the total number of credits to reserve to fit
5687 * the modification of a single pages into a single transaction,
5688 * which may include multiple chunks of block allocations.
5689 *
5690 * This could be called via ext4_write_begin()
5691 *
5692 * We need to consider the worse case, when
5693 * one new block per extent.
5694 */
5695int ext4_writepage_trans_blocks(struct inode *inode)
5696{
5697 int bpp = ext4_journal_blocks_per_page(inode);
5698 int ret;
5699
5700 ret = ext4_meta_trans_blocks(inode, bpp, bpp);
5701
5702 /* Account for data blocks for journalled mode */
5703 if (ext4_should_journal_data(inode))
5704 ret += bpp;
5705 return ret;
5706}
5707
5708/*
5709 * Calculate the journal credits for a chunk of data modification.
5710 *
5711 * This is called from DIO, fallocate or whoever calling
5712 * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
5713 *
5714 * journal buffers for data blocks are not included here, as DIO
5715 * and fallocate do no need to journal data buffers.
5716 */
5717int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5718{
5719 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5720}
5721
5722/*
5723 * The caller must have previously called ext4_reserve_inode_write().
5724 * Give this, we know that the caller already has write access to iloc->bh.
5725 */
5726int ext4_mark_iloc_dirty(handle_t *handle,
5727 struct inode *inode, struct ext4_iloc *iloc)
5728{
5729 int err = 0;
5730
5731 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5732 return -EIO;
5733
5734 if (IS_I_VERSION(inode))
5735 inode_inc_iversion(inode);
5736
5737 /* the do_update_inode consumes one bh->b_count */
5738 get_bh(iloc->bh);
5739
5740 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5741 err = ext4_do_update_inode(handle, inode, iloc);
5742 put_bh(iloc->bh);
5743 return err;
5744}
5745
5746/*
5747 * On success, We end up with an outstanding reference count against
5748 * iloc->bh. This _must_ be cleaned up later.
5749 */
5750
5751int
5752ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5753 struct ext4_iloc *iloc)
5754{
5755 int err;
5756
5757 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
5758 return -EIO;
5759
5760 err = ext4_get_inode_loc(inode, iloc);
5761 if (!err) {
5762 BUFFER_TRACE(iloc->bh, "get_write_access");
5763 err = ext4_journal_get_write_access(handle, iloc->bh);
5764 if (err) {
5765 brelse(iloc->bh);
5766 iloc->bh = NULL;
5767 }
5768 }
5769 ext4_std_error(inode->i_sb, err);
5770 return err;
5771}
5772
5773static int __ext4_expand_extra_isize(struct inode *inode,
5774 unsigned int new_extra_isize,
5775 struct ext4_iloc *iloc,
5776 handle_t *handle, int *no_expand)
5777{
5778 struct ext4_inode *raw_inode;
5779 struct ext4_xattr_ibody_header *header;
5780 int error;
5781
5782 raw_inode = ext4_raw_inode(iloc);
5783
5784 header = IHDR(inode, raw_inode);
5785
5786 /* No extended attributes present */
5787 if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5788 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5789 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE +
5790 EXT4_I(inode)->i_extra_isize, 0,
5791 new_extra_isize - EXT4_I(inode)->i_extra_isize);
5792 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5793 return 0;
5794 }
5795
5796 /* try to expand with EAs present */
5797 error = ext4_expand_extra_isize_ea(inode, new_extra_isize,
5798 raw_inode, handle);
5799 if (error) {
5800 /*
5801 * Inode size expansion failed; don't try again
5802 */
5803 *no_expand = 1;
5804 }
5805
5806 return error;
5807}
5808
5809/*
5810 * Expand an inode by new_extra_isize bytes.
5811 * Returns 0 on success or negative error number on failure.
5812 */
5813static int ext4_try_to_expand_extra_isize(struct inode *inode,
5814 unsigned int new_extra_isize,
5815 struct ext4_iloc iloc,
5816 handle_t *handle)
5817{
5818 int no_expand;
5819 int error;
5820
5821 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND))
5822 return -EOVERFLOW;
5823
5824 /*
5825 * In nojournal mode, we can immediately attempt to expand
5826 * the inode. When journaled, we first need to obtain extra
5827 * buffer credits since we may write into the EA block
5828 * with this same handle. If journal_extend fails, then it will
5829 * only result in a minor loss of functionality for that inode.
5830 * If this is felt to be critical, then e2fsck should be run to
5831 * force a large enough s_min_extra_isize.
5832 */
5833 if (ext4_handle_valid(handle) &&
5834 jbd2_journal_extend(handle,
5835 EXT4_DATA_TRANS_BLOCKS(inode->i_sb)) != 0)
5836 return -ENOSPC;
5837
5838 if (ext4_write_trylock_xattr(inode, &no_expand) == 0)
5839 return -EBUSY;
5840
5841 error = __ext4_expand_extra_isize(inode, new_extra_isize, &iloc,
5842 handle, &no_expand);
5843 ext4_write_unlock_xattr(inode, &no_expand);
5844
5845 return error;
5846}
5847
5848int ext4_expand_extra_isize(struct inode *inode,
5849 unsigned int new_extra_isize,
5850 struct ext4_iloc *iloc)
5851{
5852 handle_t *handle;
5853 int no_expand;
5854 int error, rc;
5855
5856 if (ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5857 brelse(iloc->bh);
5858 return -EOVERFLOW;
5859 }
5860
5861 handle = ext4_journal_start(inode, EXT4_HT_INODE,
5862 EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
5863 if (IS_ERR(handle)) {
5864 error = PTR_ERR(handle);
5865 brelse(iloc->bh);
5866 return error;
5867 }
5868
5869 ext4_write_lock_xattr(inode, &no_expand);
5870
5871 BUFFER_TRACE(iloc.bh, "get_write_access");
5872 error = ext4_journal_get_write_access(handle, iloc->bh);
5873 if (error) {
5874 brelse(iloc->bh);
5875 goto out_stop;
5876 }
5877
5878 error = __ext4_expand_extra_isize(inode, new_extra_isize, iloc,
5879 handle, &no_expand);
5880
5881 rc = ext4_mark_iloc_dirty(handle, inode, iloc);
5882 if (!error)
5883 error = rc;
5884
5885 ext4_write_unlock_xattr(inode, &no_expand);
5886out_stop:
5887 ext4_journal_stop(handle);
5888 return error;
5889}
5890
5891/*
5892 * What we do here is to mark the in-core inode as clean with respect to inode
5893 * dirtiness (it may still be data-dirty).
5894 * This means that the in-core inode may be reaped by prune_icache
5895 * without having to perform any I/O. This is a very good thing,
5896 * because *any* task may call prune_icache - even ones which
5897 * have a transaction open against a different journal.
5898 *
5899 * Is this cheating? Not really. Sure, we haven't written the
5900 * inode out, but prune_icache isn't a user-visible syncing function.
5901 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5902 * we start and wait on commits.
5903 */
5904int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5905{
5906 struct ext4_iloc iloc;
5907 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5908 int err;
5909
5910 might_sleep();
5911 trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5912 err = ext4_reserve_inode_write(handle, inode, &iloc);
5913 if (err)
5914 return err;
5915
5916 if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize)
5917 ext4_try_to_expand_extra_isize(inode, sbi->s_want_extra_isize,
5918 iloc, handle);
5919
5920 return ext4_mark_iloc_dirty(handle, inode, &iloc);
5921}
5922
5923/*
5924 * ext4_dirty_inode() is called from __mark_inode_dirty()
5925 *
5926 * We're really interested in the case where a file is being extended.
5927 * i_size has been changed by generic_commit_write() and we thus need
5928 * to include the updated inode in the current transaction.
5929 *
5930 * Also, dquot_alloc_block() will always dirty the inode when blocks
5931 * are allocated to the file.
5932 *
5933 * If the inode is marked synchronous, we don't honour that here - doing
5934 * so would cause a commit on atime updates, which we don't bother doing.
5935 * We handle synchronous inodes at the highest possible level.
5936 *
5937 * If only the I_DIRTY_TIME flag is set, we can skip everything. If
5938 * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
5939 * to copy into the on-disk inode structure are the timestamp files.
5940 */
5941void ext4_dirty_inode(struct inode *inode, int flags)
5942{
5943 handle_t *handle;
5944
5945 if (flags == I_DIRTY_TIME)
5946 return;
5947 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5948 if (IS_ERR(handle))
5949 goto out;
5950
5951 ext4_mark_inode_dirty(handle, inode);
5952
5953 ext4_journal_stop(handle);
5954out:
5955 return;
5956}
5957
5958#if 0
5959/*
5960 * Bind an inode's backing buffer_head into this transaction, to prevent
5961 * it from being flushed to disk early. Unlike
5962 * ext4_reserve_inode_write, this leaves behind no bh reference and
5963 * returns no iloc structure, so the caller needs to repeat the iloc
5964 * lookup to mark the inode dirty later.
5965 */
5966static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5967{
5968 struct ext4_iloc iloc;
5969
5970 int err = 0;
5971 if (handle) {
5972 err = ext4_get_inode_loc(inode, &iloc);
5973 if (!err) {
5974 BUFFER_TRACE(iloc.bh, "get_write_access");
5975 err = jbd2_journal_get_write_access(handle, iloc.bh);
5976 if (!err)
5977 err = ext4_handle_dirty_metadata(handle,
5978 NULL,
5979 iloc.bh);
5980 brelse(iloc.bh);
5981 }
5982 }
5983 ext4_std_error(inode->i_sb, err);
5984 return err;
5985}
5986#endif
5987
5988int ext4_change_inode_journal_flag(struct inode *inode, int val)
5989{
5990 journal_t *journal;
5991 handle_t *handle;
5992 int err;
5993 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5994
5995 /*
5996 * We have to be very careful here: changing a data block's
5997 * journaling status dynamically is dangerous. If we write a
5998 * data block to the journal, change the status and then delete
5999 * that block, we risk forgetting to revoke the old log record
6000 * from the journal and so a subsequent replay can corrupt data.
6001 * So, first we make sure that the journal is empty and that
6002 * nobody is changing anything.
6003 */
6004
6005 journal = EXT4_JOURNAL(inode);
6006 if (!journal)
6007 return 0;
6008 if (is_journal_aborted(journal))
6009 return -EROFS;
6010
6011 /* Wait for all existing dio workers */
6012 inode_dio_wait(inode);
6013
6014 /*
6015 * Before flushing the journal and switching inode's aops, we have
6016 * to flush all dirty data the inode has. There can be outstanding
6017 * delayed allocations, there can be unwritten extents created by
6018 * fallocate or buffered writes in dioread_nolock mode covered by
6019 * dirty data which can be converted only after flushing the dirty
6020 * data (and journalled aops don't know how to handle these cases).
6021 */
6022 if (val) {
6023 down_write(&EXT4_I(inode)->i_mmap_sem);
6024 err = filemap_write_and_wait(inode->i_mapping);
6025 if (err < 0) {
6026 up_write(&EXT4_I(inode)->i_mmap_sem);
6027 return err;
6028 }
6029 }
6030
6031 percpu_down_write(&sbi->s_journal_flag_rwsem);
6032 jbd2_journal_lock_updates(journal);
6033
6034 /*
6035 * OK, there are no updates running now, and all cached data is
6036 * synced to disk. We are now in a completely consistent state
6037 * which doesn't have anything in the journal, and we know that
6038 * no filesystem updates are running, so it is safe to modify
6039 * the inode's in-core data-journaling state flag now.
6040 */
6041
6042 if (val)
6043 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6044 else {
6045 err = jbd2_journal_flush(journal);
6046 if (err < 0) {
6047 jbd2_journal_unlock_updates(journal);
6048 percpu_up_write(&sbi->s_journal_flag_rwsem);
6049 return err;
6050 }
6051 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
6052 }
6053 ext4_set_aops(inode);
6054
6055 jbd2_journal_unlock_updates(journal);
6056 percpu_up_write(&sbi->s_journal_flag_rwsem);
6057
6058 if (val)
6059 up_write(&EXT4_I(inode)->i_mmap_sem);
6060
6061 /* Finally we can mark the inode as dirty. */
6062
6063 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
6064 if (IS_ERR(handle))
6065 return PTR_ERR(handle);
6066
6067 err = ext4_mark_inode_dirty(handle, inode);
6068 ext4_handle_sync(handle);
6069 ext4_journal_stop(handle);
6070 ext4_std_error(inode->i_sb, err);
6071
6072 return err;
6073}
6074
6075static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
6076{
6077 return !buffer_mapped(bh);
6078}
6079
6080int ext4_page_mkwrite(struct vm_fault *vmf)
6081{
6082 struct vm_area_struct *vma = vmf->vma;
6083 struct page *page = vmf->page;
6084 loff_t size;
6085 unsigned long len;
6086 int ret;
6087 struct file *file = vma->vm_file;
6088 struct inode *inode = file_inode(file);
6089 struct address_space *mapping = inode->i_mapping;
6090 handle_t *handle;
6091 get_block_t *get_block;
6092 int retries = 0;
6093
6094 sb_start_pagefault(inode->i_sb);
6095 file_update_time(vma->vm_file);
6096
6097 down_read(&EXT4_I(inode)->i_mmap_sem);
6098
6099 ret = ext4_convert_inline_data(inode);
6100 if (ret)
6101 goto out_ret;
6102
6103 /* Delalloc case is easy... */
6104 if (test_opt(inode->i_sb, DELALLOC) &&
6105 !ext4_should_journal_data(inode) &&
6106 !ext4_nonda_switch(inode->i_sb)) {
6107 do {
6108 ret = block_page_mkwrite(vma, vmf,
6109 ext4_da_get_block_prep);
6110 } while (ret == -ENOSPC &&
6111 ext4_should_retry_alloc(inode->i_sb, &retries));
6112 goto out_ret;
6113 }
6114
6115 lock_page(page);
6116 size = i_size_read(inode);
6117 /* Page got truncated from under us? */
6118 if (page->mapping != mapping || page_offset(page) > size) {
6119 unlock_page(page);
6120 ret = VM_FAULT_NOPAGE;
6121 goto out;
6122 }
6123
6124 if (page->index == size >> PAGE_SHIFT)
6125 len = size & ~PAGE_MASK;
6126 else
6127 len = PAGE_SIZE;
6128 /*
6129 * Return if we have all the buffers mapped. This avoids the need to do
6130 * journal_start/journal_stop which can block and take a long time
6131 */
6132 if (page_has_buffers(page)) {
6133 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
6134 0, len, NULL,
6135 ext4_bh_unmapped)) {
6136 /* Wait so that we don't change page under IO */
6137 wait_for_stable_page(page);
6138 ret = VM_FAULT_LOCKED;
6139 goto out;
6140 }
6141 }
6142 unlock_page(page);
6143 /* OK, we need to fill the hole... */
6144 if (ext4_should_dioread_nolock(inode))
6145 get_block = ext4_get_block_unwritten;
6146 else
6147 get_block = ext4_get_block;
6148retry_alloc:
6149 handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
6150 ext4_writepage_trans_blocks(inode));
6151 if (IS_ERR(handle)) {
6152 ret = VM_FAULT_SIGBUS;
6153 goto out;
6154 }
6155 ret = block_page_mkwrite(vma, vmf, get_block);
6156 if (!ret && ext4_should_journal_data(inode)) {
6157 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
6158 PAGE_SIZE, NULL, do_journal_get_write_access)) {
6159 unlock_page(page);
6160 ret = VM_FAULT_SIGBUS;
6161 ext4_journal_stop(handle);
6162 goto out;
6163 }
6164 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
6165 }
6166 ext4_journal_stop(handle);
6167 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
6168 goto retry_alloc;
6169out_ret:
6170 ret = block_page_mkwrite_return(ret);
6171out:
6172 up_read(&EXT4_I(inode)->i_mmap_sem);
6173 sb_end_pagefault(inode->i_sb);
6174 return ret;
6175}
6176
6177int ext4_filemap_fault(struct vm_fault *vmf)
6178{
6179 struct inode *inode = file_inode(vmf->vma->vm_file);
6180 int err;
6181
6182 down_read(&EXT4_I(inode)->i_mmap_sem);
6183 err = filemap_fault(vmf);
6184 up_read(&EXT4_I(inode)->i_mmap_sem);
6185
6186 return err;
6187}