Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/filemap.c
4 *
5 * Copyright (C) 1994-1999 Linus Torvalds
6 */
7
8/*
9 * This file handles the generic file mmap semantics used by
10 * most "normal" filesystems (but you don't /have/ to use this:
11 * the NFS filesystem used to do this differently, for example)
12 */
13#include <linux/export.h>
14#include <linux/compiler.h>
15#include <linux/dax.h>
16#include <linux/fs.h>
17#include <linux/sched/signal.h>
18#include <linux/uaccess.h>
19#include <linux/capability.h>
20#include <linux/kernel_stat.h>
21#include <linux/gfp.h>
22#include <linux/mm.h>
23#include <linux/swap.h>
24#include <linux/swapops.h>
25#include <linux/syscalls.h>
26#include <linux/mman.h>
27#include <linux/pagemap.h>
28#include <linux/file.h>
29#include <linux/uio.h>
30#include <linux/error-injection.h>
31#include <linux/hash.h>
32#include <linux/writeback.h>
33#include <linux/backing-dev.h>
34#include <linux/pagevec.h>
35#include <linux/security.h>
36#include <linux/cpuset.h>
37#include <linux/hugetlb.h>
38#include <linux/memcontrol.h>
39#include <linux/shmem_fs.h>
40#include <linux/rmap.h>
41#include <linux/delayacct.h>
42#include <linux/psi.h>
43#include <linux/ramfs.h>
44#include <linux/page_idle.h>
45#include <linux/migrate.h>
46#include <linux/pipe_fs_i.h>
47#include <linux/splice.h>
48#include <linux/rcupdate_wait.h>
49#include <asm/pgalloc.h>
50#include <asm/tlbflush.h>
51#include "internal.h"
52
53#define CREATE_TRACE_POINTS
54#include <trace/events/filemap.h>
55
56/*
57 * FIXME: remove all knowledge of the buffer layer from the core VM
58 */
59#include <linux/buffer_head.h> /* for try_to_free_buffers */
60
61#include <asm/mman.h>
62
63#include "swap.h"
64
65/*
66 * Shared mappings implemented 30.11.1994. It's not fully working yet,
67 * though.
68 *
69 * Shared mappings now work. 15.8.1995 Bruno.
70 *
71 * finished 'unifying' the page and buffer cache and SMP-threaded the
72 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
73 *
74 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
75 */
76
77/*
78 * Lock ordering:
79 *
80 * ->i_mmap_rwsem (truncate_pagecache)
81 * ->private_lock (__free_pte->block_dirty_folio)
82 * ->swap_lock (exclusive_swap_page, others)
83 * ->i_pages lock
84 *
85 * ->i_rwsem
86 * ->invalidate_lock (acquired by fs in truncate path)
87 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
88 *
89 * ->mmap_lock
90 * ->i_mmap_rwsem
91 * ->page_table_lock or pte_lock (various, mainly in memory.c)
92 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
93 *
94 * ->mmap_lock
95 * ->invalidate_lock (filemap_fault)
96 * ->lock_page (filemap_fault, access_process_vm)
97 *
98 * ->i_rwsem (generic_perform_write)
99 * ->mmap_lock (fault_in_readable->do_page_fault)
100 *
101 * bdi->wb.list_lock
102 * sb_lock (fs/fs-writeback.c)
103 * ->i_pages lock (__sync_single_inode)
104 *
105 * ->i_mmap_rwsem
106 * ->anon_vma.lock (vma_merge)
107 *
108 * ->anon_vma.lock
109 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
110 *
111 * ->page_table_lock or pte_lock
112 * ->swap_lock (try_to_unmap_one)
113 * ->private_lock (try_to_unmap_one)
114 * ->i_pages lock (try_to_unmap_one)
115 * ->lruvec->lru_lock (follow_page->mark_page_accessed)
116 * ->lruvec->lru_lock (check_pte_range->isolate_lru_page)
117 * ->private_lock (folio_remove_rmap_pte->set_page_dirty)
118 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty)
119 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty)
120 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty)
121 * ->memcg->move_lock (folio_remove_rmap_pte->folio_memcg_lock)
122 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
123 * ->inode->i_lock (zap_pte_range->set_page_dirty)
124 * ->private_lock (zap_pte_range->block_dirty_folio)
125 */
126
127static void page_cache_delete(struct address_space *mapping,
128 struct folio *folio, void *shadow)
129{
130 XA_STATE(xas, &mapping->i_pages, folio->index);
131 long nr = 1;
132
133 mapping_set_update(&xas, mapping);
134
135 xas_set_order(&xas, folio->index, folio_order(folio));
136 nr = folio_nr_pages(folio);
137
138 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
139
140 xas_store(&xas, shadow);
141 xas_init_marks(&xas);
142
143 folio->mapping = NULL;
144 /* Leave page->index set: truncation lookup relies upon it */
145 mapping->nrpages -= nr;
146}
147
148static void filemap_unaccount_folio(struct address_space *mapping,
149 struct folio *folio)
150{
151 long nr;
152
153 VM_BUG_ON_FOLIO(folio_mapped(folio), folio);
154 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) {
155 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
156 current->comm, folio_pfn(folio));
157 dump_page(&folio->page, "still mapped when deleted");
158 dump_stack();
159 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
160
161 if (mapping_exiting(mapping) && !folio_test_large(folio)) {
162 int mapcount = page_mapcount(&folio->page);
163
164 if (folio_ref_count(folio) >= mapcount + 2) {
165 /*
166 * All vmas have already been torn down, so it's
167 * a good bet that actually the page is unmapped
168 * and we'd rather not leak it: if we're wrong,
169 * another bad page check should catch it later.
170 */
171 page_mapcount_reset(&folio->page);
172 folio_ref_sub(folio, mapcount);
173 }
174 }
175 }
176
177 /* hugetlb folios do not participate in page cache accounting. */
178 if (folio_test_hugetlb(folio))
179 return;
180
181 nr = folio_nr_pages(folio);
182
183 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
184 if (folio_test_swapbacked(folio)) {
185 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
186 if (folio_test_pmd_mappable(folio))
187 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr);
188 } else if (folio_test_pmd_mappable(folio)) {
189 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr);
190 filemap_nr_thps_dec(mapping);
191 }
192
193 /*
194 * At this point folio must be either written or cleaned by
195 * truncate. Dirty folio here signals a bug and loss of
196 * unwritten data - on ordinary filesystems.
197 *
198 * But it's harmless on in-memory filesystems like tmpfs; and can
199 * occur when a driver which did get_user_pages() sets page dirty
200 * before putting it, while the inode is being finally evicted.
201 *
202 * Below fixes dirty accounting after removing the folio entirely
203 * but leaves the dirty flag set: it has no effect for truncated
204 * folio and anyway will be cleared before returning folio to
205 * buddy allocator.
206 */
207 if (WARN_ON_ONCE(folio_test_dirty(folio) &&
208 mapping_can_writeback(mapping)))
209 folio_account_cleaned(folio, inode_to_wb(mapping->host));
210}
211
212/*
213 * Delete a page from the page cache and free it. Caller has to make
214 * sure the page is locked and that nobody else uses it - or that usage
215 * is safe. The caller must hold the i_pages lock.
216 */
217void __filemap_remove_folio(struct folio *folio, void *shadow)
218{
219 struct address_space *mapping = folio->mapping;
220
221 trace_mm_filemap_delete_from_page_cache(folio);
222 filemap_unaccount_folio(mapping, folio);
223 page_cache_delete(mapping, folio, shadow);
224}
225
226void filemap_free_folio(struct address_space *mapping, struct folio *folio)
227{
228 void (*free_folio)(struct folio *);
229 int refs = 1;
230
231 free_folio = mapping->a_ops->free_folio;
232 if (free_folio)
233 free_folio(folio);
234
235 if (folio_test_large(folio))
236 refs = folio_nr_pages(folio);
237 folio_put_refs(folio, refs);
238}
239
240/**
241 * filemap_remove_folio - Remove folio from page cache.
242 * @folio: The folio.
243 *
244 * This must be called only on folios that are locked and have been
245 * verified to be in the page cache. It will never put the folio into
246 * the free list because the caller has a reference on the page.
247 */
248void filemap_remove_folio(struct folio *folio)
249{
250 struct address_space *mapping = folio->mapping;
251
252 BUG_ON(!folio_test_locked(folio));
253 spin_lock(&mapping->host->i_lock);
254 xa_lock_irq(&mapping->i_pages);
255 __filemap_remove_folio(folio, NULL);
256 xa_unlock_irq(&mapping->i_pages);
257 if (mapping_shrinkable(mapping))
258 inode_add_lru(mapping->host);
259 spin_unlock(&mapping->host->i_lock);
260
261 filemap_free_folio(mapping, folio);
262}
263
264/*
265 * page_cache_delete_batch - delete several folios from page cache
266 * @mapping: the mapping to which folios belong
267 * @fbatch: batch of folios to delete
268 *
269 * The function walks over mapping->i_pages and removes folios passed in
270 * @fbatch from the mapping. The function expects @fbatch to be sorted
271 * by page index and is optimised for it to be dense.
272 * It tolerates holes in @fbatch (mapping entries at those indices are not
273 * modified).
274 *
275 * The function expects the i_pages lock to be held.
276 */
277static void page_cache_delete_batch(struct address_space *mapping,
278 struct folio_batch *fbatch)
279{
280 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index);
281 long total_pages = 0;
282 int i = 0;
283 struct folio *folio;
284
285 mapping_set_update(&xas, mapping);
286 xas_for_each(&xas, folio, ULONG_MAX) {
287 if (i >= folio_batch_count(fbatch))
288 break;
289
290 /* A swap/dax/shadow entry got inserted? Skip it. */
291 if (xa_is_value(folio))
292 continue;
293 /*
294 * A page got inserted in our range? Skip it. We have our
295 * pages locked so they are protected from being removed.
296 * If we see a page whose index is higher than ours, it
297 * means our page has been removed, which shouldn't be
298 * possible because we're holding the PageLock.
299 */
300 if (folio != fbatch->folios[i]) {
301 VM_BUG_ON_FOLIO(folio->index >
302 fbatch->folios[i]->index, folio);
303 continue;
304 }
305
306 WARN_ON_ONCE(!folio_test_locked(folio));
307
308 folio->mapping = NULL;
309 /* Leave folio->index set: truncation lookup relies on it */
310
311 i++;
312 xas_store(&xas, NULL);
313 total_pages += folio_nr_pages(folio);
314 }
315 mapping->nrpages -= total_pages;
316}
317
318void delete_from_page_cache_batch(struct address_space *mapping,
319 struct folio_batch *fbatch)
320{
321 int i;
322
323 if (!folio_batch_count(fbatch))
324 return;
325
326 spin_lock(&mapping->host->i_lock);
327 xa_lock_irq(&mapping->i_pages);
328 for (i = 0; i < folio_batch_count(fbatch); i++) {
329 struct folio *folio = fbatch->folios[i];
330
331 trace_mm_filemap_delete_from_page_cache(folio);
332 filemap_unaccount_folio(mapping, folio);
333 }
334 page_cache_delete_batch(mapping, fbatch);
335 xa_unlock_irq(&mapping->i_pages);
336 if (mapping_shrinkable(mapping))
337 inode_add_lru(mapping->host);
338 spin_unlock(&mapping->host->i_lock);
339
340 for (i = 0; i < folio_batch_count(fbatch); i++)
341 filemap_free_folio(mapping, fbatch->folios[i]);
342}
343
344int filemap_check_errors(struct address_space *mapping)
345{
346 int ret = 0;
347 /* Check for outstanding write errors */
348 if (test_bit(AS_ENOSPC, &mapping->flags) &&
349 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
350 ret = -ENOSPC;
351 if (test_bit(AS_EIO, &mapping->flags) &&
352 test_and_clear_bit(AS_EIO, &mapping->flags))
353 ret = -EIO;
354 return ret;
355}
356EXPORT_SYMBOL(filemap_check_errors);
357
358static int filemap_check_and_keep_errors(struct address_space *mapping)
359{
360 /* Check for outstanding write errors */
361 if (test_bit(AS_EIO, &mapping->flags))
362 return -EIO;
363 if (test_bit(AS_ENOSPC, &mapping->flags))
364 return -ENOSPC;
365 return 0;
366}
367
368/**
369 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range
370 * @mapping: address space structure to write
371 * @wbc: the writeback_control controlling the writeout
372 *
373 * Call writepages on the mapping using the provided wbc to control the
374 * writeout.
375 *
376 * Return: %0 on success, negative error code otherwise.
377 */
378int filemap_fdatawrite_wbc(struct address_space *mapping,
379 struct writeback_control *wbc)
380{
381 int ret;
382
383 if (!mapping_can_writeback(mapping) ||
384 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
385 return 0;
386
387 wbc_attach_fdatawrite_inode(wbc, mapping->host);
388 ret = do_writepages(mapping, wbc);
389 wbc_detach_inode(wbc);
390 return ret;
391}
392EXPORT_SYMBOL(filemap_fdatawrite_wbc);
393
394/**
395 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
396 * @mapping: address space structure to write
397 * @start: offset in bytes where the range starts
398 * @end: offset in bytes where the range ends (inclusive)
399 * @sync_mode: enable synchronous operation
400 *
401 * Start writeback against all of a mapping's dirty pages that lie
402 * within the byte offsets <start, end> inclusive.
403 *
404 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
405 * opposed to a regular memory cleansing writeback. The difference between
406 * these two operations is that if a dirty page/buffer is encountered, it must
407 * be waited upon, and not just skipped over.
408 *
409 * Return: %0 on success, negative error code otherwise.
410 */
411int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
412 loff_t end, int sync_mode)
413{
414 struct writeback_control wbc = {
415 .sync_mode = sync_mode,
416 .nr_to_write = LONG_MAX,
417 .range_start = start,
418 .range_end = end,
419 };
420
421 return filemap_fdatawrite_wbc(mapping, &wbc);
422}
423
424static inline int __filemap_fdatawrite(struct address_space *mapping,
425 int sync_mode)
426{
427 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
428}
429
430int filemap_fdatawrite(struct address_space *mapping)
431{
432 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
433}
434EXPORT_SYMBOL(filemap_fdatawrite);
435
436int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
437 loff_t end)
438{
439 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
440}
441EXPORT_SYMBOL(filemap_fdatawrite_range);
442
443/**
444 * filemap_flush - mostly a non-blocking flush
445 * @mapping: target address_space
446 *
447 * This is a mostly non-blocking flush. Not suitable for data-integrity
448 * purposes - I/O may not be started against all dirty pages.
449 *
450 * Return: %0 on success, negative error code otherwise.
451 */
452int filemap_flush(struct address_space *mapping)
453{
454 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
455}
456EXPORT_SYMBOL(filemap_flush);
457
458/**
459 * filemap_range_has_page - check if a page exists in range.
460 * @mapping: address space within which to check
461 * @start_byte: offset in bytes where the range starts
462 * @end_byte: offset in bytes where the range ends (inclusive)
463 *
464 * Find at least one page in the range supplied, usually used to check if
465 * direct writing in this range will trigger a writeback.
466 *
467 * Return: %true if at least one page exists in the specified range,
468 * %false otherwise.
469 */
470bool filemap_range_has_page(struct address_space *mapping,
471 loff_t start_byte, loff_t end_byte)
472{
473 struct folio *folio;
474 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
475 pgoff_t max = end_byte >> PAGE_SHIFT;
476
477 if (end_byte < start_byte)
478 return false;
479
480 rcu_read_lock();
481 for (;;) {
482 folio = xas_find(&xas, max);
483 if (xas_retry(&xas, folio))
484 continue;
485 /* Shadow entries don't count */
486 if (xa_is_value(folio))
487 continue;
488 /*
489 * We don't need to try to pin this page; we're about to
490 * release the RCU lock anyway. It is enough to know that
491 * there was a page here recently.
492 */
493 break;
494 }
495 rcu_read_unlock();
496
497 return folio != NULL;
498}
499EXPORT_SYMBOL(filemap_range_has_page);
500
501static void __filemap_fdatawait_range(struct address_space *mapping,
502 loff_t start_byte, loff_t end_byte)
503{
504 pgoff_t index = start_byte >> PAGE_SHIFT;
505 pgoff_t end = end_byte >> PAGE_SHIFT;
506 struct folio_batch fbatch;
507 unsigned nr_folios;
508
509 folio_batch_init(&fbatch);
510
511 while (index <= end) {
512 unsigned i;
513
514 nr_folios = filemap_get_folios_tag(mapping, &index, end,
515 PAGECACHE_TAG_WRITEBACK, &fbatch);
516
517 if (!nr_folios)
518 break;
519
520 for (i = 0; i < nr_folios; i++) {
521 struct folio *folio = fbatch.folios[i];
522
523 folio_wait_writeback(folio);
524 folio_clear_error(folio);
525 }
526 folio_batch_release(&fbatch);
527 cond_resched();
528 }
529}
530
531/**
532 * filemap_fdatawait_range - wait for writeback to complete
533 * @mapping: address space structure to wait for
534 * @start_byte: offset in bytes where the range starts
535 * @end_byte: offset in bytes where the range ends (inclusive)
536 *
537 * Walk the list of under-writeback pages of the given address space
538 * in the given range and wait for all of them. Check error status of
539 * the address space and return it.
540 *
541 * Since the error status of the address space is cleared by this function,
542 * callers are responsible for checking the return value and handling and/or
543 * reporting the error.
544 *
545 * Return: error status of the address space.
546 */
547int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
548 loff_t end_byte)
549{
550 __filemap_fdatawait_range(mapping, start_byte, end_byte);
551 return filemap_check_errors(mapping);
552}
553EXPORT_SYMBOL(filemap_fdatawait_range);
554
555/**
556 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
557 * @mapping: address space structure to wait for
558 * @start_byte: offset in bytes where the range starts
559 * @end_byte: offset in bytes where the range ends (inclusive)
560 *
561 * Walk the list of under-writeback pages of the given address space in the
562 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
563 * this function does not clear error status of the address space.
564 *
565 * Use this function if callers don't handle errors themselves. Expected
566 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
567 * fsfreeze(8)
568 */
569int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
570 loff_t start_byte, loff_t end_byte)
571{
572 __filemap_fdatawait_range(mapping, start_byte, end_byte);
573 return filemap_check_and_keep_errors(mapping);
574}
575EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
576
577/**
578 * file_fdatawait_range - wait for writeback to complete
579 * @file: file pointing to address space structure to wait for
580 * @start_byte: offset in bytes where the range starts
581 * @end_byte: offset in bytes where the range ends (inclusive)
582 *
583 * Walk the list of under-writeback pages of the address space that file
584 * refers to, in the given range and wait for all of them. Check error
585 * status of the address space vs. the file->f_wb_err cursor and return it.
586 *
587 * Since the error status of the file is advanced by this function,
588 * callers are responsible for checking the return value and handling and/or
589 * reporting the error.
590 *
591 * Return: error status of the address space vs. the file->f_wb_err cursor.
592 */
593int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
594{
595 struct address_space *mapping = file->f_mapping;
596
597 __filemap_fdatawait_range(mapping, start_byte, end_byte);
598 return file_check_and_advance_wb_err(file);
599}
600EXPORT_SYMBOL(file_fdatawait_range);
601
602/**
603 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
604 * @mapping: address space structure to wait for
605 *
606 * Walk the list of under-writeback pages of the given address space
607 * and wait for all of them. Unlike filemap_fdatawait(), this function
608 * does not clear error status of the address space.
609 *
610 * Use this function if callers don't handle errors themselves. Expected
611 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
612 * fsfreeze(8)
613 *
614 * Return: error status of the address space.
615 */
616int filemap_fdatawait_keep_errors(struct address_space *mapping)
617{
618 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
619 return filemap_check_and_keep_errors(mapping);
620}
621EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
622
623/* Returns true if writeback might be needed or already in progress. */
624static bool mapping_needs_writeback(struct address_space *mapping)
625{
626 return mapping->nrpages;
627}
628
629bool filemap_range_has_writeback(struct address_space *mapping,
630 loff_t start_byte, loff_t end_byte)
631{
632 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
633 pgoff_t max = end_byte >> PAGE_SHIFT;
634 struct folio *folio;
635
636 if (end_byte < start_byte)
637 return false;
638
639 rcu_read_lock();
640 xas_for_each(&xas, folio, max) {
641 if (xas_retry(&xas, folio))
642 continue;
643 if (xa_is_value(folio))
644 continue;
645 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
646 folio_test_writeback(folio))
647 break;
648 }
649 rcu_read_unlock();
650 return folio != NULL;
651}
652EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
653
654/**
655 * filemap_write_and_wait_range - write out & wait on a file range
656 * @mapping: the address_space for the pages
657 * @lstart: offset in bytes where the range starts
658 * @lend: offset in bytes where the range ends (inclusive)
659 *
660 * Write out and wait upon file offsets lstart->lend, inclusive.
661 *
662 * Note that @lend is inclusive (describes the last byte to be written) so
663 * that this function can be used to write to the very end-of-file (end = -1).
664 *
665 * Return: error status of the address space.
666 */
667int filemap_write_and_wait_range(struct address_space *mapping,
668 loff_t lstart, loff_t lend)
669{
670 int err = 0, err2;
671
672 if (lend < lstart)
673 return 0;
674
675 if (mapping_needs_writeback(mapping)) {
676 err = __filemap_fdatawrite_range(mapping, lstart, lend,
677 WB_SYNC_ALL);
678 /*
679 * Even if the above returned error, the pages may be
680 * written partially (e.g. -ENOSPC), so we wait for it.
681 * But the -EIO is special case, it may indicate the worst
682 * thing (e.g. bug) happened, so we avoid waiting for it.
683 */
684 if (err != -EIO)
685 __filemap_fdatawait_range(mapping, lstart, lend);
686 }
687 err2 = filemap_check_errors(mapping);
688 if (!err)
689 err = err2;
690 return err;
691}
692EXPORT_SYMBOL(filemap_write_and_wait_range);
693
694void __filemap_set_wb_err(struct address_space *mapping, int err)
695{
696 errseq_t eseq = errseq_set(&mapping->wb_err, err);
697
698 trace_filemap_set_wb_err(mapping, eseq);
699}
700EXPORT_SYMBOL(__filemap_set_wb_err);
701
702/**
703 * file_check_and_advance_wb_err - report wb error (if any) that was previously
704 * and advance wb_err to current one
705 * @file: struct file on which the error is being reported
706 *
707 * When userland calls fsync (or something like nfsd does the equivalent), we
708 * want to report any writeback errors that occurred since the last fsync (or
709 * since the file was opened if there haven't been any).
710 *
711 * Grab the wb_err from the mapping. If it matches what we have in the file,
712 * then just quickly return 0. The file is all caught up.
713 *
714 * If it doesn't match, then take the mapping value, set the "seen" flag in
715 * it and try to swap it into place. If it works, or another task beat us
716 * to it with the new value, then update the f_wb_err and return the error
717 * portion. The error at this point must be reported via proper channels
718 * (a'la fsync, or NFS COMMIT operation, etc.).
719 *
720 * While we handle mapping->wb_err with atomic operations, the f_wb_err
721 * value is protected by the f_lock since we must ensure that it reflects
722 * the latest value swapped in for this file descriptor.
723 *
724 * Return: %0 on success, negative error code otherwise.
725 */
726int file_check_and_advance_wb_err(struct file *file)
727{
728 int err = 0;
729 errseq_t old = READ_ONCE(file->f_wb_err);
730 struct address_space *mapping = file->f_mapping;
731
732 /* Locklessly handle the common case where nothing has changed */
733 if (errseq_check(&mapping->wb_err, old)) {
734 /* Something changed, must use slow path */
735 spin_lock(&file->f_lock);
736 old = file->f_wb_err;
737 err = errseq_check_and_advance(&mapping->wb_err,
738 &file->f_wb_err);
739 trace_file_check_and_advance_wb_err(file, old);
740 spin_unlock(&file->f_lock);
741 }
742
743 /*
744 * We're mostly using this function as a drop in replacement for
745 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
746 * that the legacy code would have had on these flags.
747 */
748 clear_bit(AS_EIO, &mapping->flags);
749 clear_bit(AS_ENOSPC, &mapping->flags);
750 return err;
751}
752EXPORT_SYMBOL(file_check_and_advance_wb_err);
753
754/**
755 * file_write_and_wait_range - write out & wait on a file range
756 * @file: file pointing to address_space with pages
757 * @lstart: offset in bytes where the range starts
758 * @lend: offset in bytes where the range ends (inclusive)
759 *
760 * Write out and wait upon file offsets lstart->lend, inclusive.
761 *
762 * Note that @lend is inclusive (describes the last byte to be written) so
763 * that this function can be used to write to the very end-of-file (end = -1).
764 *
765 * After writing out and waiting on the data, we check and advance the
766 * f_wb_err cursor to the latest value, and return any errors detected there.
767 *
768 * Return: %0 on success, negative error code otherwise.
769 */
770int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
771{
772 int err = 0, err2;
773 struct address_space *mapping = file->f_mapping;
774
775 if (lend < lstart)
776 return 0;
777
778 if (mapping_needs_writeback(mapping)) {
779 err = __filemap_fdatawrite_range(mapping, lstart, lend,
780 WB_SYNC_ALL);
781 /* See comment of filemap_write_and_wait() */
782 if (err != -EIO)
783 __filemap_fdatawait_range(mapping, lstart, lend);
784 }
785 err2 = file_check_and_advance_wb_err(file);
786 if (!err)
787 err = err2;
788 return err;
789}
790EXPORT_SYMBOL(file_write_and_wait_range);
791
792/**
793 * replace_page_cache_folio - replace a pagecache folio with a new one
794 * @old: folio to be replaced
795 * @new: folio to replace with
796 *
797 * This function replaces a folio in the pagecache with a new one. On
798 * success it acquires the pagecache reference for the new folio and
799 * drops it for the old folio. Both the old and new folios must be
800 * locked. This function does not add the new folio to the LRU, the
801 * caller must do that.
802 *
803 * The remove + add is atomic. This function cannot fail.
804 */
805void replace_page_cache_folio(struct folio *old, struct folio *new)
806{
807 struct address_space *mapping = old->mapping;
808 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
809 pgoff_t offset = old->index;
810 XA_STATE(xas, &mapping->i_pages, offset);
811
812 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
813 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
814 VM_BUG_ON_FOLIO(new->mapping, new);
815
816 folio_get(new);
817 new->mapping = mapping;
818 new->index = offset;
819
820 mem_cgroup_replace_folio(old, new);
821
822 xas_lock_irq(&xas);
823 xas_store(&xas, new);
824
825 old->mapping = NULL;
826 /* hugetlb pages do not participate in page cache accounting. */
827 if (!folio_test_hugetlb(old))
828 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
829 if (!folio_test_hugetlb(new))
830 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
831 if (folio_test_swapbacked(old))
832 __lruvec_stat_sub_folio(old, NR_SHMEM);
833 if (folio_test_swapbacked(new))
834 __lruvec_stat_add_folio(new, NR_SHMEM);
835 xas_unlock_irq(&xas);
836 if (free_folio)
837 free_folio(old);
838 folio_put(old);
839}
840EXPORT_SYMBOL_GPL(replace_page_cache_folio);
841
842noinline int __filemap_add_folio(struct address_space *mapping,
843 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
844{
845 XA_STATE(xas, &mapping->i_pages, index);
846 int huge = folio_test_hugetlb(folio);
847 bool charged = false;
848 long nr = 1;
849
850 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
851 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
852 mapping_set_update(&xas, mapping);
853
854 if (!huge) {
855 int error = mem_cgroup_charge(folio, NULL, gfp);
856 if (error)
857 return error;
858 charged = true;
859 }
860
861 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
862 xas_set_order(&xas, index, folio_order(folio));
863 nr = folio_nr_pages(folio);
864
865 gfp &= GFP_RECLAIM_MASK;
866 folio_ref_add(folio, nr);
867 folio->mapping = mapping;
868 folio->index = xas.xa_index;
869
870 do {
871 unsigned int order = xa_get_order(xas.xa, xas.xa_index);
872 void *entry, *old = NULL;
873
874 if (order > folio_order(folio))
875 xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index),
876 order, gfp);
877 xas_lock_irq(&xas);
878 xas_for_each_conflict(&xas, entry) {
879 old = entry;
880 if (!xa_is_value(entry)) {
881 xas_set_err(&xas, -EEXIST);
882 goto unlock;
883 }
884 }
885
886 if (old) {
887 if (shadowp)
888 *shadowp = old;
889 /* entry may have been split before we acquired lock */
890 order = xa_get_order(xas.xa, xas.xa_index);
891 if (order > folio_order(folio)) {
892 /* How to handle large swap entries? */
893 BUG_ON(shmem_mapping(mapping));
894 xas_split(&xas, old, order);
895 xas_reset(&xas);
896 }
897 }
898
899 xas_store(&xas, folio);
900 if (xas_error(&xas))
901 goto unlock;
902
903 mapping->nrpages += nr;
904
905 /* hugetlb pages do not participate in page cache accounting */
906 if (!huge) {
907 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
908 if (folio_test_pmd_mappable(folio))
909 __lruvec_stat_mod_folio(folio,
910 NR_FILE_THPS, nr);
911 }
912unlock:
913 xas_unlock_irq(&xas);
914 } while (xas_nomem(&xas, gfp));
915
916 if (xas_error(&xas))
917 goto error;
918
919 trace_mm_filemap_add_to_page_cache(folio);
920 return 0;
921error:
922 if (charged)
923 mem_cgroup_uncharge(folio);
924 folio->mapping = NULL;
925 /* Leave page->index set: truncation relies upon it */
926 folio_put_refs(folio, nr);
927 return xas_error(&xas);
928}
929ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
930
931int filemap_add_folio(struct address_space *mapping, struct folio *folio,
932 pgoff_t index, gfp_t gfp)
933{
934 void *shadow = NULL;
935 int ret;
936
937 __folio_set_locked(folio);
938 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
939 if (unlikely(ret))
940 __folio_clear_locked(folio);
941 else {
942 /*
943 * The folio might have been evicted from cache only
944 * recently, in which case it should be activated like
945 * any other repeatedly accessed folio.
946 * The exception is folios getting rewritten; evicting other
947 * data from the working set, only to cache data that will
948 * get overwritten with something else, is a waste of memory.
949 */
950 WARN_ON_ONCE(folio_test_active(folio));
951 if (!(gfp & __GFP_WRITE) && shadow)
952 workingset_refault(folio, shadow);
953 folio_add_lru(folio);
954 }
955 return ret;
956}
957EXPORT_SYMBOL_GPL(filemap_add_folio);
958
959#ifdef CONFIG_NUMA
960struct folio *filemap_alloc_folio(gfp_t gfp, unsigned int order)
961{
962 int n;
963 struct folio *folio;
964
965 if (cpuset_do_page_mem_spread()) {
966 unsigned int cpuset_mems_cookie;
967 do {
968 cpuset_mems_cookie = read_mems_allowed_begin();
969 n = cpuset_mem_spread_node();
970 folio = __folio_alloc_node(gfp, order, n);
971 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
972
973 return folio;
974 }
975 return folio_alloc(gfp, order);
976}
977EXPORT_SYMBOL(filemap_alloc_folio);
978#endif
979
980/*
981 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
982 *
983 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
984 *
985 * @mapping1: the first mapping to lock
986 * @mapping2: the second mapping to lock
987 */
988void filemap_invalidate_lock_two(struct address_space *mapping1,
989 struct address_space *mapping2)
990{
991 if (mapping1 > mapping2)
992 swap(mapping1, mapping2);
993 if (mapping1)
994 down_write(&mapping1->invalidate_lock);
995 if (mapping2 && mapping1 != mapping2)
996 down_write_nested(&mapping2->invalidate_lock, 1);
997}
998EXPORT_SYMBOL(filemap_invalidate_lock_two);
999
1000/*
1001 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
1002 *
1003 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1004 *
1005 * @mapping1: the first mapping to unlock
1006 * @mapping2: the second mapping to unlock
1007 */
1008void filemap_invalidate_unlock_two(struct address_space *mapping1,
1009 struct address_space *mapping2)
1010{
1011 if (mapping1)
1012 up_write(&mapping1->invalidate_lock);
1013 if (mapping2 && mapping1 != mapping2)
1014 up_write(&mapping2->invalidate_lock);
1015}
1016EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1017
1018/*
1019 * In order to wait for pages to become available there must be
1020 * waitqueues associated with pages. By using a hash table of
1021 * waitqueues where the bucket discipline is to maintain all
1022 * waiters on the same queue and wake all when any of the pages
1023 * become available, and for the woken contexts to check to be
1024 * sure the appropriate page became available, this saves space
1025 * at a cost of "thundering herd" phenomena during rare hash
1026 * collisions.
1027 */
1028#define PAGE_WAIT_TABLE_BITS 8
1029#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1030static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1031
1032static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1033{
1034 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1035}
1036
1037void __init pagecache_init(void)
1038{
1039 int i;
1040
1041 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1042 init_waitqueue_head(&folio_wait_table[i]);
1043
1044 page_writeback_init();
1045}
1046
1047/*
1048 * The page wait code treats the "wait->flags" somewhat unusually, because
1049 * we have multiple different kinds of waits, not just the usual "exclusive"
1050 * one.
1051 *
1052 * We have:
1053 *
1054 * (a) no special bits set:
1055 *
1056 * We're just waiting for the bit to be released, and when a waker
1057 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1058 * and remove it from the wait queue.
1059 *
1060 * Simple and straightforward.
1061 *
1062 * (b) WQ_FLAG_EXCLUSIVE:
1063 *
1064 * The waiter is waiting to get the lock, and only one waiter should
1065 * be woken up to avoid any thundering herd behavior. We'll set the
1066 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1067 *
1068 * This is the traditional exclusive wait.
1069 *
1070 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1071 *
1072 * The waiter is waiting to get the bit, and additionally wants the
1073 * lock to be transferred to it for fair lock behavior. If the lock
1074 * cannot be taken, we stop walking the wait queue without waking
1075 * the waiter.
1076 *
1077 * This is the "fair lock handoff" case, and in addition to setting
1078 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1079 * that it now has the lock.
1080 */
1081static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1082{
1083 unsigned int flags;
1084 struct wait_page_key *key = arg;
1085 struct wait_page_queue *wait_page
1086 = container_of(wait, struct wait_page_queue, wait);
1087
1088 if (!wake_page_match(wait_page, key))
1089 return 0;
1090
1091 /*
1092 * If it's a lock handoff wait, we get the bit for it, and
1093 * stop walking (and do not wake it up) if we can't.
1094 */
1095 flags = wait->flags;
1096 if (flags & WQ_FLAG_EXCLUSIVE) {
1097 if (test_bit(key->bit_nr, &key->folio->flags))
1098 return -1;
1099 if (flags & WQ_FLAG_CUSTOM) {
1100 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1101 return -1;
1102 flags |= WQ_FLAG_DONE;
1103 }
1104 }
1105
1106 /*
1107 * We are holding the wait-queue lock, but the waiter that
1108 * is waiting for this will be checking the flags without
1109 * any locking.
1110 *
1111 * So update the flags atomically, and wake up the waiter
1112 * afterwards to avoid any races. This store-release pairs
1113 * with the load-acquire in folio_wait_bit_common().
1114 */
1115 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1116 wake_up_state(wait->private, mode);
1117
1118 /*
1119 * Ok, we have successfully done what we're waiting for,
1120 * and we can unconditionally remove the wait entry.
1121 *
1122 * Note that this pairs with the "finish_wait()" in the
1123 * waiter, and has to be the absolute last thing we do.
1124 * After this list_del_init(&wait->entry) the wait entry
1125 * might be de-allocated and the process might even have
1126 * exited.
1127 */
1128 list_del_init_careful(&wait->entry);
1129 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1130}
1131
1132static void folio_wake_bit(struct folio *folio, int bit_nr)
1133{
1134 wait_queue_head_t *q = folio_waitqueue(folio);
1135 struct wait_page_key key;
1136 unsigned long flags;
1137
1138 key.folio = folio;
1139 key.bit_nr = bit_nr;
1140 key.page_match = 0;
1141
1142 spin_lock_irqsave(&q->lock, flags);
1143 __wake_up_locked_key(q, TASK_NORMAL, &key);
1144
1145 /*
1146 * It's possible to miss clearing waiters here, when we woke our page
1147 * waiters, but the hashed waitqueue has waiters for other pages on it.
1148 * That's okay, it's a rare case. The next waker will clear it.
1149 *
1150 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1151 * other), the flag may be cleared in the course of freeing the page;
1152 * but that is not required for correctness.
1153 */
1154 if (!waitqueue_active(q) || !key.page_match)
1155 folio_clear_waiters(folio);
1156
1157 spin_unlock_irqrestore(&q->lock, flags);
1158}
1159
1160/*
1161 * A choice of three behaviors for folio_wait_bit_common():
1162 */
1163enum behavior {
1164 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1165 * __folio_lock() waiting on then setting PG_locked.
1166 */
1167 SHARED, /* Hold ref to page and check the bit when woken, like
1168 * folio_wait_writeback() waiting on PG_writeback.
1169 */
1170 DROP, /* Drop ref to page before wait, no check when woken,
1171 * like folio_put_wait_locked() on PG_locked.
1172 */
1173};
1174
1175/*
1176 * Attempt to check (or get) the folio flag, and mark us done
1177 * if successful.
1178 */
1179static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1180 struct wait_queue_entry *wait)
1181{
1182 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1183 if (test_and_set_bit(bit_nr, &folio->flags))
1184 return false;
1185 } else if (test_bit(bit_nr, &folio->flags))
1186 return false;
1187
1188 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1189 return true;
1190}
1191
1192/* How many times do we accept lock stealing from under a waiter? */
1193int sysctl_page_lock_unfairness = 5;
1194
1195static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1196 int state, enum behavior behavior)
1197{
1198 wait_queue_head_t *q = folio_waitqueue(folio);
1199 int unfairness = sysctl_page_lock_unfairness;
1200 struct wait_page_queue wait_page;
1201 wait_queue_entry_t *wait = &wait_page.wait;
1202 bool thrashing = false;
1203 unsigned long pflags;
1204 bool in_thrashing;
1205
1206 if (bit_nr == PG_locked &&
1207 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1208 delayacct_thrashing_start(&in_thrashing);
1209 psi_memstall_enter(&pflags);
1210 thrashing = true;
1211 }
1212
1213 init_wait(wait);
1214 wait->func = wake_page_function;
1215 wait_page.folio = folio;
1216 wait_page.bit_nr = bit_nr;
1217
1218repeat:
1219 wait->flags = 0;
1220 if (behavior == EXCLUSIVE) {
1221 wait->flags = WQ_FLAG_EXCLUSIVE;
1222 if (--unfairness < 0)
1223 wait->flags |= WQ_FLAG_CUSTOM;
1224 }
1225
1226 /*
1227 * Do one last check whether we can get the
1228 * page bit synchronously.
1229 *
1230 * Do the folio_set_waiters() marking before that
1231 * to let any waker we _just_ missed know they
1232 * need to wake us up (otherwise they'll never
1233 * even go to the slow case that looks at the
1234 * page queue), and add ourselves to the wait
1235 * queue if we need to sleep.
1236 *
1237 * This part needs to be done under the queue
1238 * lock to avoid races.
1239 */
1240 spin_lock_irq(&q->lock);
1241 folio_set_waiters(folio);
1242 if (!folio_trylock_flag(folio, bit_nr, wait))
1243 __add_wait_queue_entry_tail(q, wait);
1244 spin_unlock_irq(&q->lock);
1245
1246 /*
1247 * From now on, all the logic will be based on
1248 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1249 * see whether the page bit testing has already
1250 * been done by the wake function.
1251 *
1252 * We can drop our reference to the folio.
1253 */
1254 if (behavior == DROP)
1255 folio_put(folio);
1256
1257 /*
1258 * Note that until the "finish_wait()", or until
1259 * we see the WQ_FLAG_WOKEN flag, we need to
1260 * be very careful with the 'wait->flags', because
1261 * we may race with a waker that sets them.
1262 */
1263 for (;;) {
1264 unsigned int flags;
1265
1266 set_current_state(state);
1267
1268 /* Loop until we've been woken or interrupted */
1269 flags = smp_load_acquire(&wait->flags);
1270 if (!(flags & WQ_FLAG_WOKEN)) {
1271 if (signal_pending_state(state, current))
1272 break;
1273
1274 io_schedule();
1275 continue;
1276 }
1277
1278 /* If we were non-exclusive, we're done */
1279 if (behavior != EXCLUSIVE)
1280 break;
1281
1282 /* If the waker got the lock for us, we're done */
1283 if (flags & WQ_FLAG_DONE)
1284 break;
1285
1286 /*
1287 * Otherwise, if we're getting the lock, we need to
1288 * try to get it ourselves.
1289 *
1290 * And if that fails, we'll have to retry this all.
1291 */
1292 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1293 goto repeat;
1294
1295 wait->flags |= WQ_FLAG_DONE;
1296 break;
1297 }
1298
1299 /*
1300 * If a signal happened, this 'finish_wait()' may remove the last
1301 * waiter from the wait-queues, but the folio waiters bit will remain
1302 * set. That's ok. The next wakeup will take care of it, and trying
1303 * to do it here would be difficult and prone to races.
1304 */
1305 finish_wait(q, wait);
1306
1307 if (thrashing) {
1308 delayacct_thrashing_end(&in_thrashing);
1309 psi_memstall_leave(&pflags);
1310 }
1311
1312 /*
1313 * NOTE! The wait->flags weren't stable until we've done the
1314 * 'finish_wait()', and we could have exited the loop above due
1315 * to a signal, and had a wakeup event happen after the signal
1316 * test but before the 'finish_wait()'.
1317 *
1318 * So only after the finish_wait() can we reliably determine
1319 * if we got woken up or not, so we can now figure out the final
1320 * return value based on that state without races.
1321 *
1322 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1323 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1324 */
1325 if (behavior == EXCLUSIVE)
1326 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1327
1328 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1329}
1330
1331#ifdef CONFIG_MIGRATION
1332/**
1333 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1334 * @entry: migration swap entry.
1335 * @ptl: already locked ptl. This function will drop the lock.
1336 *
1337 * Wait for a migration entry referencing the given page to be removed. This is
1338 * equivalent to put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE) except
1339 * this can be called without taking a reference on the page. Instead this
1340 * should be called while holding the ptl for the migration entry referencing
1341 * the page.
1342 *
1343 * Returns after unlocking the ptl.
1344 *
1345 * This follows the same logic as folio_wait_bit_common() so see the comments
1346 * there.
1347 */
1348void migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl)
1349 __releases(ptl)
1350{
1351 struct wait_page_queue wait_page;
1352 wait_queue_entry_t *wait = &wait_page.wait;
1353 bool thrashing = false;
1354 unsigned long pflags;
1355 bool in_thrashing;
1356 wait_queue_head_t *q;
1357 struct folio *folio = page_folio(pfn_swap_entry_to_page(entry));
1358
1359 q = folio_waitqueue(folio);
1360 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1361 delayacct_thrashing_start(&in_thrashing);
1362 psi_memstall_enter(&pflags);
1363 thrashing = true;
1364 }
1365
1366 init_wait(wait);
1367 wait->func = wake_page_function;
1368 wait_page.folio = folio;
1369 wait_page.bit_nr = PG_locked;
1370 wait->flags = 0;
1371
1372 spin_lock_irq(&q->lock);
1373 folio_set_waiters(folio);
1374 if (!folio_trylock_flag(folio, PG_locked, wait))
1375 __add_wait_queue_entry_tail(q, wait);
1376 spin_unlock_irq(&q->lock);
1377
1378 /*
1379 * If a migration entry exists for the page the migration path must hold
1380 * a valid reference to the page, and it must take the ptl to remove the
1381 * migration entry. So the page is valid until the ptl is dropped.
1382 */
1383 spin_unlock(ptl);
1384
1385 for (;;) {
1386 unsigned int flags;
1387
1388 set_current_state(TASK_UNINTERRUPTIBLE);
1389
1390 /* Loop until we've been woken or interrupted */
1391 flags = smp_load_acquire(&wait->flags);
1392 if (!(flags & WQ_FLAG_WOKEN)) {
1393 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1394 break;
1395
1396 io_schedule();
1397 continue;
1398 }
1399 break;
1400 }
1401
1402 finish_wait(q, wait);
1403
1404 if (thrashing) {
1405 delayacct_thrashing_end(&in_thrashing);
1406 psi_memstall_leave(&pflags);
1407 }
1408}
1409#endif
1410
1411void folio_wait_bit(struct folio *folio, int bit_nr)
1412{
1413 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1414}
1415EXPORT_SYMBOL(folio_wait_bit);
1416
1417int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1418{
1419 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1420}
1421EXPORT_SYMBOL(folio_wait_bit_killable);
1422
1423/**
1424 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1425 * @folio: The folio to wait for.
1426 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1427 *
1428 * The caller should hold a reference on @folio. They expect the page to
1429 * become unlocked relatively soon, but do not wish to hold up migration
1430 * (for example) by holding the reference while waiting for the folio to
1431 * come unlocked. After this function returns, the caller should not
1432 * dereference @folio.
1433 *
1434 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1435 */
1436static int folio_put_wait_locked(struct folio *folio, int state)
1437{
1438 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1439}
1440
1441/**
1442 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue
1443 * @folio: Folio defining the wait queue of interest
1444 * @waiter: Waiter to add to the queue
1445 *
1446 * Add an arbitrary @waiter to the wait queue for the nominated @folio.
1447 */
1448void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter)
1449{
1450 wait_queue_head_t *q = folio_waitqueue(folio);
1451 unsigned long flags;
1452
1453 spin_lock_irqsave(&q->lock, flags);
1454 __add_wait_queue_entry_tail(q, waiter);
1455 folio_set_waiters(folio);
1456 spin_unlock_irqrestore(&q->lock, flags);
1457}
1458EXPORT_SYMBOL_GPL(folio_add_wait_queue);
1459
1460/**
1461 * folio_unlock - Unlock a locked folio.
1462 * @folio: The folio.
1463 *
1464 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1465 *
1466 * Context: May be called from interrupt or process context. May not be
1467 * called from NMI context.
1468 */
1469void folio_unlock(struct folio *folio)
1470{
1471 /* Bit 7 allows x86 to check the byte's sign bit */
1472 BUILD_BUG_ON(PG_waiters != 7);
1473 BUILD_BUG_ON(PG_locked > 7);
1474 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1475 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked))
1476 folio_wake_bit(folio, PG_locked);
1477}
1478EXPORT_SYMBOL(folio_unlock);
1479
1480/**
1481 * folio_end_read - End read on a folio.
1482 * @folio: The folio.
1483 * @success: True if all reads completed successfully.
1484 *
1485 * When all reads against a folio have completed, filesystems should
1486 * call this function to let the pagecache know that no more reads
1487 * are outstanding. This will unlock the folio and wake up any thread
1488 * sleeping on the lock. The folio will also be marked uptodate if all
1489 * reads succeeded.
1490 *
1491 * Context: May be called from interrupt or process context. May not be
1492 * called from NMI context.
1493 */
1494void folio_end_read(struct folio *folio, bool success)
1495{
1496 unsigned long mask = 1 << PG_locked;
1497
1498 /* Must be in bottom byte for x86 to work */
1499 BUILD_BUG_ON(PG_uptodate > 7);
1500 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1501 VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
1502
1503 if (likely(success))
1504 mask |= 1 << PG_uptodate;
1505 if (folio_xor_flags_has_waiters(folio, mask))
1506 folio_wake_bit(folio, PG_locked);
1507}
1508EXPORT_SYMBOL(folio_end_read);
1509
1510/**
1511 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1512 * @folio: The folio.
1513 *
1514 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1515 * it. The folio reference held for PG_private_2 being set is released.
1516 *
1517 * This is, for example, used when a netfs folio is being written to a local
1518 * disk cache, thereby allowing writes to the cache for the same folio to be
1519 * serialised.
1520 */
1521void folio_end_private_2(struct folio *folio)
1522{
1523 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1524 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1525 folio_wake_bit(folio, PG_private_2);
1526 folio_put(folio);
1527}
1528EXPORT_SYMBOL(folio_end_private_2);
1529
1530/**
1531 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1532 * @folio: The folio to wait on.
1533 *
1534 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio.
1535 */
1536void folio_wait_private_2(struct folio *folio)
1537{
1538 while (folio_test_private_2(folio))
1539 folio_wait_bit(folio, PG_private_2);
1540}
1541EXPORT_SYMBOL(folio_wait_private_2);
1542
1543/**
1544 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1545 * @folio: The folio to wait on.
1546 *
1547 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio or until a
1548 * fatal signal is received by the calling task.
1549 *
1550 * Return:
1551 * - 0 if successful.
1552 * - -EINTR if a fatal signal was encountered.
1553 */
1554int folio_wait_private_2_killable(struct folio *folio)
1555{
1556 int ret = 0;
1557
1558 while (folio_test_private_2(folio)) {
1559 ret = folio_wait_bit_killable(folio, PG_private_2);
1560 if (ret < 0)
1561 break;
1562 }
1563
1564 return ret;
1565}
1566EXPORT_SYMBOL(folio_wait_private_2_killable);
1567
1568/**
1569 * folio_end_writeback - End writeback against a folio.
1570 * @folio: The folio.
1571 *
1572 * The folio must actually be under writeback.
1573 *
1574 * Context: May be called from process or interrupt context.
1575 */
1576void folio_end_writeback(struct folio *folio)
1577{
1578 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
1579
1580 /*
1581 * folio_test_clear_reclaim() could be used here but it is an
1582 * atomic operation and overkill in this particular case. Failing
1583 * to shuffle a folio marked for immediate reclaim is too mild
1584 * a gain to justify taking an atomic operation penalty at the
1585 * end of every folio writeback.
1586 */
1587 if (folio_test_reclaim(folio)) {
1588 folio_clear_reclaim(folio);
1589 folio_rotate_reclaimable(folio);
1590 }
1591
1592 /*
1593 * Writeback does not hold a folio reference of its own, relying
1594 * on truncation to wait for the clearing of PG_writeback.
1595 * But here we must make sure that the folio is not freed and
1596 * reused before the folio_wake_bit().
1597 */
1598 folio_get(folio);
1599 if (__folio_end_writeback(folio))
1600 folio_wake_bit(folio, PG_writeback);
1601 acct_reclaim_writeback(folio);
1602 folio_put(folio);
1603}
1604EXPORT_SYMBOL(folio_end_writeback);
1605
1606/**
1607 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1608 * @folio: The folio to lock
1609 */
1610void __folio_lock(struct folio *folio)
1611{
1612 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1613 EXCLUSIVE);
1614}
1615EXPORT_SYMBOL(__folio_lock);
1616
1617int __folio_lock_killable(struct folio *folio)
1618{
1619 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1620 EXCLUSIVE);
1621}
1622EXPORT_SYMBOL_GPL(__folio_lock_killable);
1623
1624static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1625{
1626 struct wait_queue_head *q = folio_waitqueue(folio);
1627 int ret;
1628
1629 wait->folio = folio;
1630 wait->bit_nr = PG_locked;
1631
1632 spin_lock_irq(&q->lock);
1633 __add_wait_queue_entry_tail(q, &wait->wait);
1634 folio_set_waiters(folio);
1635 ret = !folio_trylock(folio);
1636 /*
1637 * If we were successful now, we know we're still on the
1638 * waitqueue as we're still under the lock. This means it's
1639 * safe to remove and return success, we know the callback
1640 * isn't going to trigger.
1641 */
1642 if (!ret)
1643 __remove_wait_queue(q, &wait->wait);
1644 else
1645 ret = -EIOCBQUEUED;
1646 spin_unlock_irq(&q->lock);
1647 return ret;
1648}
1649
1650/*
1651 * Return values:
1652 * 0 - folio is locked.
1653 * non-zero - folio is not locked.
1654 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or
1655 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and
1656 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held.
1657 *
1658 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0
1659 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed.
1660 */
1661vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf)
1662{
1663 unsigned int flags = vmf->flags;
1664
1665 if (fault_flag_allow_retry_first(flags)) {
1666 /*
1667 * CAUTION! In this case, mmap_lock/per-VMA lock is not
1668 * released even though returning VM_FAULT_RETRY.
1669 */
1670 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1671 return VM_FAULT_RETRY;
1672
1673 release_fault_lock(vmf);
1674 if (flags & FAULT_FLAG_KILLABLE)
1675 folio_wait_locked_killable(folio);
1676 else
1677 folio_wait_locked(folio);
1678 return VM_FAULT_RETRY;
1679 }
1680 if (flags & FAULT_FLAG_KILLABLE) {
1681 bool ret;
1682
1683 ret = __folio_lock_killable(folio);
1684 if (ret) {
1685 release_fault_lock(vmf);
1686 return VM_FAULT_RETRY;
1687 }
1688 } else {
1689 __folio_lock(folio);
1690 }
1691
1692 return 0;
1693}
1694
1695/**
1696 * page_cache_next_miss() - Find the next gap in the page cache.
1697 * @mapping: Mapping.
1698 * @index: Index.
1699 * @max_scan: Maximum range to search.
1700 *
1701 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1702 * gap with the lowest index.
1703 *
1704 * This function may be called under the rcu_read_lock. However, this will
1705 * not atomically search a snapshot of the cache at a single point in time.
1706 * For example, if a gap is created at index 5, then subsequently a gap is
1707 * created at index 10, page_cache_next_miss covering both indices may
1708 * return 10 if called under the rcu_read_lock.
1709 *
1710 * Return: The index of the gap if found, otherwise an index outside the
1711 * range specified (in which case 'return - index >= max_scan' will be true).
1712 * In the rare case of index wrap-around, 0 will be returned.
1713 */
1714pgoff_t page_cache_next_miss(struct address_space *mapping,
1715 pgoff_t index, unsigned long max_scan)
1716{
1717 XA_STATE(xas, &mapping->i_pages, index);
1718
1719 while (max_scan--) {
1720 void *entry = xas_next(&xas);
1721 if (!entry || xa_is_value(entry))
1722 break;
1723 if (xas.xa_index == 0)
1724 break;
1725 }
1726
1727 return xas.xa_index;
1728}
1729EXPORT_SYMBOL(page_cache_next_miss);
1730
1731/**
1732 * page_cache_prev_miss() - Find the previous gap in the page cache.
1733 * @mapping: Mapping.
1734 * @index: Index.
1735 * @max_scan: Maximum range to search.
1736 *
1737 * Search the range [max(index - max_scan + 1, 0), index] for the
1738 * gap with the highest index.
1739 *
1740 * This function may be called under the rcu_read_lock. However, this will
1741 * not atomically search a snapshot of the cache at a single point in time.
1742 * For example, if a gap is created at index 10, then subsequently a gap is
1743 * created at index 5, page_cache_prev_miss() covering both indices may
1744 * return 5 if called under the rcu_read_lock.
1745 *
1746 * Return: The index of the gap if found, otherwise an index outside the
1747 * range specified (in which case 'index - return >= max_scan' will be true).
1748 * In the rare case of wrap-around, ULONG_MAX will be returned.
1749 */
1750pgoff_t page_cache_prev_miss(struct address_space *mapping,
1751 pgoff_t index, unsigned long max_scan)
1752{
1753 XA_STATE(xas, &mapping->i_pages, index);
1754
1755 while (max_scan--) {
1756 void *entry = xas_prev(&xas);
1757 if (!entry || xa_is_value(entry))
1758 break;
1759 if (xas.xa_index == ULONG_MAX)
1760 break;
1761 }
1762
1763 return xas.xa_index;
1764}
1765EXPORT_SYMBOL(page_cache_prev_miss);
1766
1767/*
1768 * Lockless page cache protocol:
1769 * On the lookup side:
1770 * 1. Load the folio from i_pages
1771 * 2. Increment the refcount if it's not zero
1772 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1773 *
1774 * On the removal side:
1775 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1776 * B. Remove the page from i_pages
1777 * C. Return the page to the page allocator
1778 *
1779 * This means that any page may have its reference count temporarily
1780 * increased by a speculative page cache (or fast GUP) lookup as it can
1781 * be allocated by another user before the RCU grace period expires.
1782 * Because the refcount temporarily acquired here may end up being the
1783 * last refcount on the page, any page allocation must be freeable by
1784 * folio_put().
1785 */
1786
1787/*
1788 * filemap_get_entry - Get a page cache entry.
1789 * @mapping: the address_space to search
1790 * @index: The page cache index.
1791 *
1792 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1793 * it is returned with an increased refcount. If it is a shadow entry
1794 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1795 * it is returned without further action.
1796 *
1797 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1798 */
1799void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
1800{
1801 XA_STATE(xas, &mapping->i_pages, index);
1802 struct folio *folio;
1803
1804 rcu_read_lock();
1805repeat:
1806 xas_reset(&xas);
1807 folio = xas_load(&xas);
1808 if (xas_retry(&xas, folio))
1809 goto repeat;
1810 /*
1811 * A shadow entry of a recently evicted page, or a swap entry from
1812 * shmem/tmpfs. Return it without attempting to raise page count.
1813 */
1814 if (!folio || xa_is_value(folio))
1815 goto out;
1816
1817 if (!folio_try_get_rcu(folio))
1818 goto repeat;
1819
1820 if (unlikely(folio != xas_reload(&xas))) {
1821 folio_put(folio);
1822 goto repeat;
1823 }
1824out:
1825 rcu_read_unlock();
1826
1827 return folio;
1828}
1829
1830/**
1831 * __filemap_get_folio - Find and get a reference to a folio.
1832 * @mapping: The address_space to search.
1833 * @index: The page index.
1834 * @fgp_flags: %FGP flags modify how the folio is returned.
1835 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1836 *
1837 * Looks up the page cache entry at @mapping & @index.
1838 *
1839 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1840 * if the %GFP flags specified for %FGP_CREAT are atomic.
1841 *
1842 * If this function returns a folio, it is returned with an increased refcount.
1843 *
1844 * Return: The found folio or an ERR_PTR() otherwise.
1845 */
1846struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1847 fgf_t fgp_flags, gfp_t gfp)
1848{
1849 struct folio *folio;
1850
1851repeat:
1852 folio = filemap_get_entry(mapping, index);
1853 if (xa_is_value(folio))
1854 folio = NULL;
1855 if (!folio)
1856 goto no_page;
1857
1858 if (fgp_flags & FGP_LOCK) {
1859 if (fgp_flags & FGP_NOWAIT) {
1860 if (!folio_trylock(folio)) {
1861 folio_put(folio);
1862 return ERR_PTR(-EAGAIN);
1863 }
1864 } else {
1865 folio_lock(folio);
1866 }
1867
1868 /* Has the page been truncated? */
1869 if (unlikely(folio->mapping != mapping)) {
1870 folio_unlock(folio);
1871 folio_put(folio);
1872 goto repeat;
1873 }
1874 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1875 }
1876
1877 if (fgp_flags & FGP_ACCESSED)
1878 folio_mark_accessed(folio);
1879 else if (fgp_flags & FGP_WRITE) {
1880 /* Clear idle flag for buffer write */
1881 if (folio_test_idle(folio))
1882 folio_clear_idle(folio);
1883 }
1884
1885 if (fgp_flags & FGP_STABLE)
1886 folio_wait_stable(folio);
1887no_page:
1888 if (!folio && (fgp_flags & FGP_CREAT)) {
1889 unsigned order = FGF_GET_ORDER(fgp_flags);
1890 int err;
1891
1892 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1893 gfp |= __GFP_WRITE;
1894 if (fgp_flags & FGP_NOFS)
1895 gfp &= ~__GFP_FS;
1896 if (fgp_flags & FGP_NOWAIT) {
1897 gfp &= ~GFP_KERNEL;
1898 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1899 }
1900 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1901 fgp_flags |= FGP_LOCK;
1902
1903 if (!mapping_large_folio_support(mapping))
1904 order = 0;
1905 if (order > MAX_PAGECACHE_ORDER)
1906 order = MAX_PAGECACHE_ORDER;
1907 /* If we're not aligned, allocate a smaller folio */
1908 if (index & ((1UL << order) - 1))
1909 order = __ffs(index);
1910
1911 do {
1912 gfp_t alloc_gfp = gfp;
1913
1914 err = -ENOMEM;
1915 if (order == 1)
1916 order = 0;
1917 if (order > 0)
1918 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN;
1919 folio = filemap_alloc_folio(alloc_gfp, order);
1920 if (!folio)
1921 continue;
1922
1923 /* Init accessed so avoid atomic mark_page_accessed later */
1924 if (fgp_flags & FGP_ACCESSED)
1925 __folio_set_referenced(folio);
1926
1927 err = filemap_add_folio(mapping, folio, index, gfp);
1928 if (!err)
1929 break;
1930 folio_put(folio);
1931 folio = NULL;
1932 } while (order-- > 0);
1933
1934 if (err == -EEXIST)
1935 goto repeat;
1936 if (err)
1937 return ERR_PTR(err);
1938 /*
1939 * filemap_add_folio locks the page, and for mmap
1940 * we expect an unlocked page.
1941 */
1942 if (folio && (fgp_flags & FGP_FOR_MMAP))
1943 folio_unlock(folio);
1944 }
1945
1946 if (!folio)
1947 return ERR_PTR(-ENOENT);
1948 return folio;
1949}
1950EXPORT_SYMBOL(__filemap_get_folio);
1951
1952static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
1953 xa_mark_t mark)
1954{
1955 struct folio *folio;
1956
1957retry:
1958 if (mark == XA_PRESENT)
1959 folio = xas_find(xas, max);
1960 else
1961 folio = xas_find_marked(xas, max, mark);
1962
1963 if (xas_retry(xas, folio))
1964 goto retry;
1965 /*
1966 * A shadow entry of a recently evicted page, a swap
1967 * entry from shmem/tmpfs or a DAX entry. Return it
1968 * without attempting to raise page count.
1969 */
1970 if (!folio || xa_is_value(folio))
1971 return folio;
1972
1973 if (!folio_try_get_rcu(folio))
1974 goto reset;
1975
1976 if (unlikely(folio != xas_reload(xas))) {
1977 folio_put(folio);
1978 goto reset;
1979 }
1980
1981 return folio;
1982reset:
1983 xas_reset(xas);
1984 goto retry;
1985}
1986
1987/**
1988 * find_get_entries - gang pagecache lookup
1989 * @mapping: The address_space to search
1990 * @start: The starting page cache index
1991 * @end: The final page index (inclusive).
1992 * @fbatch: Where the resulting entries are placed.
1993 * @indices: The cache indices corresponding to the entries in @entries
1994 *
1995 * find_get_entries() will search for and return a batch of entries in
1996 * the mapping. The entries are placed in @fbatch. find_get_entries()
1997 * takes a reference on any actual folios it returns.
1998 *
1999 * The entries have ascending indexes. The indices may not be consecutive
2000 * due to not-present entries or large folios.
2001 *
2002 * Any shadow entries of evicted folios, or swap entries from
2003 * shmem/tmpfs, are included in the returned array.
2004 *
2005 * Return: The number of entries which were found.
2006 */
2007unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2008 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2009{
2010 XA_STATE(xas, &mapping->i_pages, *start);
2011 struct folio *folio;
2012
2013 rcu_read_lock();
2014 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2015 indices[fbatch->nr] = xas.xa_index;
2016 if (!folio_batch_add(fbatch, folio))
2017 break;
2018 }
2019 rcu_read_unlock();
2020
2021 if (folio_batch_count(fbatch)) {
2022 unsigned long nr = 1;
2023 int idx = folio_batch_count(fbatch) - 1;
2024
2025 folio = fbatch->folios[idx];
2026 if (!xa_is_value(folio))
2027 nr = folio_nr_pages(folio);
2028 *start = indices[idx] + nr;
2029 }
2030 return folio_batch_count(fbatch);
2031}
2032
2033/**
2034 * find_lock_entries - Find a batch of pagecache entries.
2035 * @mapping: The address_space to search.
2036 * @start: The starting page cache index.
2037 * @end: The final page index (inclusive).
2038 * @fbatch: Where the resulting entries are placed.
2039 * @indices: The cache indices of the entries in @fbatch.
2040 *
2041 * find_lock_entries() will return a batch of entries from @mapping.
2042 * Swap, shadow and DAX entries are included. Folios are returned
2043 * locked and with an incremented refcount. Folios which are locked
2044 * by somebody else or under writeback are skipped. Folios which are
2045 * partially outside the range are not returned.
2046 *
2047 * The entries have ascending indexes. The indices may not be consecutive
2048 * due to not-present entries, large folios, folios which could not be
2049 * locked or folios under writeback.
2050 *
2051 * Return: The number of entries which were found.
2052 */
2053unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2054 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2055{
2056 XA_STATE(xas, &mapping->i_pages, *start);
2057 struct folio *folio;
2058
2059 rcu_read_lock();
2060 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2061 if (!xa_is_value(folio)) {
2062 if (folio->index < *start)
2063 goto put;
2064 if (folio_next_index(folio) - 1 > end)
2065 goto put;
2066 if (!folio_trylock(folio))
2067 goto put;
2068 if (folio->mapping != mapping ||
2069 folio_test_writeback(folio))
2070 goto unlock;
2071 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2072 folio);
2073 }
2074 indices[fbatch->nr] = xas.xa_index;
2075 if (!folio_batch_add(fbatch, folio))
2076 break;
2077 continue;
2078unlock:
2079 folio_unlock(folio);
2080put:
2081 folio_put(folio);
2082 }
2083 rcu_read_unlock();
2084
2085 if (folio_batch_count(fbatch)) {
2086 unsigned long nr = 1;
2087 int idx = folio_batch_count(fbatch) - 1;
2088
2089 folio = fbatch->folios[idx];
2090 if (!xa_is_value(folio))
2091 nr = folio_nr_pages(folio);
2092 *start = indices[idx] + nr;
2093 }
2094 return folio_batch_count(fbatch);
2095}
2096
2097/**
2098 * filemap_get_folios - Get a batch of folios
2099 * @mapping: The address_space to search
2100 * @start: The starting page index
2101 * @end: The final page index (inclusive)
2102 * @fbatch: The batch to fill.
2103 *
2104 * Search for and return a batch of folios in the mapping starting at
2105 * index @start and up to index @end (inclusive). The folios are returned
2106 * in @fbatch with an elevated reference count.
2107 *
2108 * Return: The number of folios which were found.
2109 * We also update @start to index the next folio for the traversal.
2110 */
2111unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2112 pgoff_t end, struct folio_batch *fbatch)
2113{
2114 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch);
2115}
2116EXPORT_SYMBOL(filemap_get_folios);
2117
2118/**
2119 * filemap_get_folios_contig - Get a batch of contiguous folios
2120 * @mapping: The address_space to search
2121 * @start: The starting page index
2122 * @end: The final page index (inclusive)
2123 * @fbatch: The batch to fill
2124 *
2125 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2126 * except the returned folios are guaranteed to be contiguous. This may
2127 * not return all contiguous folios if the batch gets filled up.
2128 *
2129 * Return: The number of folios found.
2130 * Also update @start to be positioned for traversal of the next folio.
2131 */
2132
2133unsigned filemap_get_folios_contig(struct address_space *mapping,
2134 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2135{
2136 XA_STATE(xas, &mapping->i_pages, *start);
2137 unsigned long nr;
2138 struct folio *folio;
2139
2140 rcu_read_lock();
2141
2142 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2143 folio = xas_next(&xas)) {
2144 if (xas_retry(&xas, folio))
2145 continue;
2146 /*
2147 * If the entry has been swapped out, we can stop looking.
2148 * No current caller is looking for DAX entries.
2149 */
2150 if (xa_is_value(folio))
2151 goto update_start;
2152
2153 if (!folio_try_get_rcu(folio))
2154 goto retry;
2155
2156 if (unlikely(folio != xas_reload(&xas)))
2157 goto put_folio;
2158
2159 if (!folio_batch_add(fbatch, folio)) {
2160 nr = folio_nr_pages(folio);
2161 *start = folio->index + nr;
2162 goto out;
2163 }
2164 continue;
2165put_folio:
2166 folio_put(folio);
2167
2168retry:
2169 xas_reset(&xas);
2170 }
2171
2172update_start:
2173 nr = folio_batch_count(fbatch);
2174
2175 if (nr) {
2176 folio = fbatch->folios[nr - 1];
2177 *start = folio_next_index(folio);
2178 }
2179out:
2180 rcu_read_unlock();
2181 return folio_batch_count(fbatch);
2182}
2183EXPORT_SYMBOL(filemap_get_folios_contig);
2184
2185/**
2186 * filemap_get_folios_tag - Get a batch of folios matching @tag
2187 * @mapping: The address_space to search
2188 * @start: The starting page index
2189 * @end: The final page index (inclusive)
2190 * @tag: The tag index
2191 * @fbatch: The batch to fill
2192 *
2193 * The first folio may start before @start; if it does, it will contain
2194 * @start. The final folio may extend beyond @end; if it does, it will
2195 * contain @end. The folios have ascending indices. There may be gaps
2196 * between the folios if there are indices which have no folio in the
2197 * page cache. If folios are added to or removed from the page cache
2198 * while this is running, they may or may not be found by this call.
2199 * Only returns folios that are tagged with @tag.
2200 *
2201 * Return: The number of folios found.
2202 * Also update @start to index the next folio for traversal.
2203 */
2204unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start,
2205 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch)
2206{
2207 XA_STATE(xas, &mapping->i_pages, *start);
2208 struct folio *folio;
2209
2210 rcu_read_lock();
2211 while ((folio = find_get_entry(&xas, end, tag)) != NULL) {
2212 /*
2213 * Shadow entries should never be tagged, but this iteration
2214 * is lockless so there is a window for page reclaim to evict
2215 * a page we saw tagged. Skip over it.
2216 */
2217 if (xa_is_value(folio))
2218 continue;
2219 if (!folio_batch_add(fbatch, folio)) {
2220 unsigned long nr = folio_nr_pages(folio);
2221 *start = folio->index + nr;
2222 goto out;
2223 }
2224 }
2225 /*
2226 * We come here when there is no page beyond @end. We take care to not
2227 * overflow the index @start as it confuses some of the callers. This
2228 * breaks the iteration when there is a page at index -1 but that is
2229 * already broke anyway.
2230 */
2231 if (end == (pgoff_t)-1)
2232 *start = (pgoff_t)-1;
2233 else
2234 *start = end + 1;
2235out:
2236 rcu_read_unlock();
2237
2238 return folio_batch_count(fbatch);
2239}
2240EXPORT_SYMBOL(filemap_get_folios_tag);
2241
2242/*
2243 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2244 * a _large_ part of the i/o request. Imagine the worst scenario:
2245 *
2246 * ---R__________________________________________B__________
2247 * ^ reading here ^ bad block(assume 4k)
2248 *
2249 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2250 * => failing the whole request => read(R) => read(R+1) =>
2251 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2252 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2253 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2254 *
2255 * It is going insane. Fix it by quickly scaling down the readahead size.
2256 */
2257static void shrink_readahead_size_eio(struct file_ra_state *ra)
2258{
2259 ra->ra_pages /= 4;
2260}
2261
2262/*
2263 * filemap_get_read_batch - Get a batch of folios for read
2264 *
2265 * Get a batch of folios which represent a contiguous range of bytes in
2266 * the file. No exceptional entries will be returned. If @index is in
2267 * the middle of a folio, the entire folio will be returned. The last
2268 * folio in the batch may have the readahead flag set or the uptodate flag
2269 * clear so that the caller can take the appropriate action.
2270 */
2271static void filemap_get_read_batch(struct address_space *mapping,
2272 pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
2273{
2274 XA_STATE(xas, &mapping->i_pages, index);
2275 struct folio *folio;
2276
2277 rcu_read_lock();
2278 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2279 if (xas_retry(&xas, folio))
2280 continue;
2281 if (xas.xa_index > max || xa_is_value(folio))
2282 break;
2283 if (xa_is_sibling(folio))
2284 break;
2285 if (!folio_try_get_rcu(folio))
2286 goto retry;
2287
2288 if (unlikely(folio != xas_reload(&xas)))
2289 goto put_folio;
2290
2291 if (!folio_batch_add(fbatch, folio))
2292 break;
2293 if (!folio_test_uptodate(folio))
2294 break;
2295 if (folio_test_readahead(folio))
2296 break;
2297 xas_advance(&xas, folio_next_index(folio) - 1);
2298 continue;
2299put_folio:
2300 folio_put(folio);
2301retry:
2302 xas_reset(&xas);
2303 }
2304 rcu_read_unlock();
2305}
2306
2307static int filemap_read_folio(struct file *file, filler_t filler,
2308 struct folio *folio)
2309{
2310 bool workingset = folio_test_workingset(folio);
2311 unsigned long pflags;
2312 int error;
2313
2314 /*
2315 * A previous I/O error may have been due to temporary failures,
2316 * eg. multipath errors. PG_error will be set again if read_folio
2317 * fails.
2318 */
2319 folio_clear_error(folio);
2320
2321 /* Start the actual read. The read will unlock the page. */
2322 if (unlikely(workingset))
2323 psi_memstall_enter(&pflags);
2324 error = filler(file, folio);
2325 if (unlikely(workingset))
2326 psi_memstall_leave(&pflags);
2327 if (error)
2328 return error;
2329
2330 error = folio_wait_locked_killable(folio);
2331 if (error)
2332 return error;
2333 if (folio_test_uptodate(folio))
2334 return 0;
2335 if (file)
2336 shrink_readahead_size_eio(&file->f_ra);
2337 return -EIO;
2338}
2339
2340static bool filemap_range_uptodate(struct address_space *mapping,
2341 loff_t pos, size_t count, struct folio *folio,
2342 bool need_uptodate)
2343{
2344 if (folio_test_uptodate(folio))
2345 return true;
2346 /* pipes can't handle partially uptodate pages */
2347 if (need_uptodate)
2348 return false;
2349 if (!mapping->a_ops->is_partially_uptodate)
2350 return false;
2351 if (mapping->host->i_blkbits >= folio_shift(folio))
2352 return false;
2353
2354 if (folio_pos(folio) > pos) {
2355 count -= folio_pos(folio) - pos;
2356 pos = 0;
2357 } else {
2358 pos -= folio_pos(folio);
2359 }
2360
2361 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2362}
2363
2364static int filemap_update_page(struct kiocb *iocb,
2365 struct address_space *mapping, size_t count,
2366 struct folio *folio, bool need_uptodate)
2367{
2368 int error;
2369
2370 if (iocb->ki_flags & IOCB_NOWAIT) {
2371 if (!filemap_invalidate_trylock_shared(mapping))
2372 return -EAGAIN;
2373 } else {
2374 filemap_invalidate_lock_shared(mapping);
2375 }
2376
2377 if (!folio_trylock(folio)) {
2378 error = -EAGAIN;
2379 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2380 goto unlock_mapping;
2381 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2382 filemap_invalidate_unlock_shared(mapping);
2383 /*
2384 * This is where we usually end up waiting for a
2385 * previously submitted readahead to finish.
2386 */
2387 folio_put_wait_locked(folio, TASK_KILLABLE);
2388 return AOP_TRUNCATED_PAGE;
2389 }
2390 error = __folio_lock_async(folio, iocb->ki_waitq);
2391 if (error)
2392 goto unlock_mapping;
2393 }
2394
2395 error = AOP_TRUNCATED_PAGE;
2396 if (!folio->mapping)
2397 goto unlock;
2398
2399 error = 0;
2400 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio,
2401 need_uptodate))
2402 goto unlock;
2403
2404 error = -EAGAIN;
2405 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2406 goto unlock;
2407
2408 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2409 folio);
2410 goto unlock_mapping;
2411unlock:
2412 folio_unlock(folio);
2413unlock_mapping:
2414 filemap_invalidate_unlock_shared(mapping);
2415 if (error == AOP_TRUNCATED_PAGE)
2416 folio_put(folio);
2417 return error;
2418}
2419
2420static int filemap_create_folio(struct file *file,
2421 struct address_space *mapping, pgoff_t index,
2422 struct folio_batch *fbatch)
2423{
2424 struct folio *folio;
2425 int error;
2426
2427 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), 0);
2428 if (!folio)
2429 return -ENOMEM;
2430
2431 /*
2432 * Protect against truncate / hole punch. Grabbing invalidate_lock
2433 * here assures we cannot instantiate and bring uptodate new
2434 * pagecache folios after evicting page cache during truncate
2435 * and before actually freeing blocks. Note that we could
2436 * release invalidate_lock after inserting the folio into
2437 * the page cache as the locked folio would then be enough to
2438 * synchronize with hole punching. But there are code paths
2439 * such as filemap_update_page() filling in partially uptodate
2440 * pages or ->readahead() that need to hold invalidate_lock
2441 * while mapping blocks for IO so let's hold the lock here as
2442 * well to keep locking rules simple.
2443 */
2444 filemap_invalidate_lock_shared(mapping);
2445 error = filemap_add_folio(mapping, folio, index,
2446 mapping_gfp_constraint(mapping, GFP_KERNEL));
2447 if (error == -EEXIST)
2448 error = AOP_TRUNCATED_PAGE;
2449 if (error)
2450 goto error;
2451
2452 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
2453 if (error)
2454 goto error;
2455
2456 filemap_invalidate_unlock_shared(mapping);
2457 folio_batch_add(fbatch, folio);
2458 return 0;
2459error:
2460 filemap_invalidate_unlock_shared(mapping);
2461 folio_put(folio);
2462 return error;
2463}
2464
2465static int filemap_readahead(struct kiocb *iocb, struct file *file,
2466 struct address_space *mapping, struct folio *folio,
2467 pgoff_t last_index)
2468{
2469 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2470
2471 if (iocb->ki_flags & IOCB_NOIO)
2472 return -EAGAIN;
2473 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2474 return 0;
2475}
2476
2477static int filemap_get_pages(struct kiocb *iocb, size_t count,
2478 struct folio_batch *fbatch, bool need_uptodate)
2479{
2480 struct file *filp = iocb->ki_filp;
2481 struct address_space *mapping = filp->f_mapping;
2482 struct file_ra_state *ra = &filp->f_ra;
2483 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2484 pgoff_t last_index;
2485 struct folio *folio;
2486 int err = 0;
2487
2488 /* "last_index" is the index of the page beyond the end of the read */
2489 last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
2490retry:
2491 if (fatal_signal_pending(current))
2492 return -EINTR;
2493
2494 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2495 if (!folio_batch_count(fbatch)) {
2496 if (iocb->ki_flags & IOCB_NOIO)
2497 return -EAGAIN;
2498 page_cache_sync_readahead(mapping, ra, filp, index,
2499 last_index - index);
2500 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2501 }
2502 if (!folio_batch_count(fbatch)) {
2503 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2504 return -EAGAIN;
2505 err = filemap_create_folio(filp, mapping,
2506 iocb->ki_pos >> PAGE_SHIFT, fbatch);
2507 if (err == AOP_TRUNCATED_PAGE)
2508 goto retry;
2509 return err;
2510 }
2511
2512 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2513 if (folio_test_readahead(folio)) {
2514 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2515 if (err)
2516 goto err;
2517 }
2518 if (!folio_test_uptodate(folio)) {
2519 if ((iocb->ki_flags & IOCB_WAITQ) &&
2520 folio_batch_count(fbatch) > 1)
2521 iocb->ki_flags |= IOCB_NOWAIT;
2522 err = filemap_update_page(iocb, mapping, count, folio,
2523 need_uptodate);
2524 if (err)
2525 goto err;
2526 }
2527
2528 return 0;
2529err:
2530 if (err < 0)
2531 folio_put(folio);
2532 if (likely(--fbatch->nr))
2533 return 0;
2534 if (err == AOP_TRUNCATED_PAGE)
2535 goto retry;
2536 return err;
2537}
2538
2539static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2540{
2541 unsigned int shift = folio_shift(folio);
2542
2543 return (pos1 >> shift == pos2 >> shift);
2544}
2545
2546/**
2547 * filemap_read - Read data from the page cache.
2548 * @iocb: The iocb to read.
2549 * @iter: Destination for the data.
2550 * @already_read: Number of bytes already read by the caller.
2551 *
2552 * Copies data from the page cache. If the data is not currently present,
2553 * uses the readahead and read_folio address_space operations to fetch it.
2554 *
2555 * Return: Total number of bytes copied, including those already read by
2556 * the caller. If an error happens before any bytes are copied, returns
2557 * a negative error number.
2558 */
2559ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2560 ssize_t already_read)
2561{
2562 struct file *filp = iocb->ki_filp;
2563 struct file_ra_state *ra = &filp->f_ra;
2564 struct address_space *mapping = filp->f_mapping;
2565 struct inode *inode = mapping->host;
2566 struct folio_batch fbatch;
2567 int i, error = 0;
2568 bool writably_mapped;
2569 loff_t isize, end_offset;
2570 loff_t last_pos = ra->prev_pos;
2571
2572 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2573 return 0;
2574 if (unlikely(!iov_iter_count(iter)))
2575 return 0;
2576
2577 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2578 folio_batch_init(&fbatch);
2579
2580 do {
2581 cond_resched();
2582
2583 /*
2584 * If we've already successfully copied some data, then we
2585 * can no longer safely return -EIOCBQUEUED. Hence mark
2586 * an async read NOWAIT at that point.
2587 */
2588 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2589 iocb->ki_flags |= IOCB_NOWAIT;
2590
2591 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2592 break;
2593
2594 error = filemap_get_pages(iocb, iter->count, &fbatch, false);
2595 if (error < 0)
2596 break;
2597
2598 /*
2599 * i_size must be checked after we know the pages are Uptodate.
2600 *
2601 * Checking i_size after the check allows us to calculate
2602 * the correct value for "nr", which means the zero-filled
2603 * part of the page is not copied back to userspace (unless
2604 * another truncate extends the file - this is desired though).
2605 */
2606 isize = i_size_read(inode);
2607 if (unlikely(iocb->ki_pos >= isize))
2608 goto put_folios;
2609 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2610
2611 /*
2612 * Pairs with a barrier in
2613 * block_write_end()->mark_buffer_dirty() or other page
2614 * dirtying routines like iomap_write_end() to ensure
2615 * changes to page contents are visible before we see
2616 * increased inode size.
2617 */
2618 smp_rmb();
2619
2620 /*
2621 * Once we start copying data, we don't want to be touching any
2622 * cachelines that might be contended:
2623 */
2624 writably_mapped = mapping_writably_mapped(mapping);
2625
2626 /*
2627 * When a read accesses the same folio several times, only
2628 * mark it as accessed the first time.
2629 */
2630 if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
2631 fbatch.folios[0]))
2632 folio_mark_accessed(fbatch.folios[0]);
2633
2634 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2635 struct folio *folio = fbatch.folios[i];
2636 size_t fsize = folio_size(folio);
2637 size_t offset = iocb->ki_pos & (fsize - 1);
2638 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2639 fsize - offset);
2640 size_t copied;
2641
2642 if (end_offset < folio_pos(folio))
2643 break;
2644 if (i > 0)
2645 folio_mark_accessed(folio);
2646 /*
2647 * If users can be writing to this folio using arbitrary
2648 * virtual addresses, take care of potential aliasing
2649 * before reading the folio on the kernel side.
2650 */
2651 if (writably_mapped)
2652 flush_dcache_folio(folio);
2653
2654 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2655
2656 already_read += copied;
2657 iocb->ki_pos += copied;
2658 last_pos = iocb->ki_pos;
2659
2660 if (copied < bytes) {
2661 error = -EFAULT;
2662 break;
2663 }
2664 }
2665put_folios:
2666 for (i = 0; i < folio_batch_count(&fbatch); i++)
2667 folio_put(fbatch.folios[i]);
2668 folio_batch_init(&fbatch);
2669 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2670
2671 file_accessed(filp);
2672 ra->prev_pos = last_pos;
2673 return already_read ? already_read : error;
2674}
2675EXPORT_SYMBOL_GPL(filemap_read);
2676
2677int kiocb_write_and_wait(struct kiocb *iocb, size_t count)
2678{
2679 struct address_space *mapping = iocb->ki_filp->f_mapping;
2680 loff_t pos = iocb->ki_pos;
2681 loff_t end = pos + count - 1;
2682
2683 if (iocb->ki_flags & IOCB_NOWAIT) {
2684 if (filemap_range_needs_writeback(mapping, pos, end))
2685 return -EAGAIN;
2686 return 0;
2687 }
2688
2689 return filemap_write_and_wait_range(mapping, pos, end);
2690}
2691EXPORT_SYMBOL_GPL(kiocb_write_and_wait);
2692
2693int kiocb_invalidate_pages(struct kiocb *iocb, size_t count)
2694{
2695 struct address_space *mapping = iocb->ki_filp->f_mapping;
2696 loff_t pos = iocb->ki_pos;
2697 loff_t end = pos + count - 1;
2698 int ret;
2699
2700 if (iocb->ki_flags & IOCB_NOWAIT) {
2701 /* we could block if there are any pages in the range */
2702 if (filemap_range_has_page(mapping, pos, end))
2703 return -EAGAIN;
2704 } else {
2705 ret = filemap_write_and_wait_range(mapping, pos, end);
2706 if (ret)
2707 return ret;
2708 }
2709
2710 /*
2711 * After a write we want buffered reads to be sure to go to disk to get
2712 * the new data. We invalidate clean cached page from the region we're
2713 * about to write. We do this *before* the write so that we can return
2714 * without clobbering -EIOCBQUEUED from ->direct_IO().
2715 */
2716 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
2717 end >> PAGE_SHIFT);
2718}
2719EXPORT_SYMBOL_GPL(kiocb_invalidate_pages);
2720
2721/**
2722 * generic_file_read_iter - generic filesystem read routine
2723 * @iocb: kernel I/O control block
2724 * @iter: destination for the data read
2725 *
2726 * This is the "read_iter()" routine for all filesystems
2727 * that can use the page cache directly.
2728 *
2729 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2730 * be returned when no data can be read without waiting for I/O requests
2731 * to complete; it doesn't prevent readahead.
2732 *
2733 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2734 * requests shall be made for the read or for readahead. When no data
2735 * can be read, -EAGAIN shall be returned. When readahead would be
2736 * triggered, a partial, possibly empty read shall be returned.
2737 *
2738 * Return:
2739 * * number of bytes copied, even for partial reads
2740 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2741 */
2742ssize_t
2743generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2744{
2745 size_t count = iov_iter_count(iter);
2746 ssize_t retval = 0;
2747
2748 if (!count)
2749 return 0; /* skip atime */
2750
2751 if (iocb->ki_flags & IOCB_DIRECT) {
2752 struct file *file = iocb->ki_filp;
2753 struct address_space *mapping = file->f_mapping;
2754 struct inode *inode = mapping->host;
2755
2756 retval = kiocb_write_and_wait(iocb, count);
2757 if (retval < 0)
2758 return retval;
2759 file_accessed(file);
2760
2761 retval = mapping->a_ops->direct_IO(iocb, iter);
2762 if (retval >= 0) {
2763 iocb->ki_pos += retval;
2764 count -= retval;
2765 }
2766 if (retval != -EIOCBQUEUED)
2767 iov_iter_revert(iter, count - iov_iter_count(iter));
2768
2769 /*
2770 * Btrfs can have a short DIO read if we encounter
2771 * compressed extents, so if there was an error, or if
2772 * we've already read everything we wanted to, or if
2773 * there was a short read because we hit EOF, go ahead
2774 * and return. Otherwise fallthrough to buffered io for
2775 * the rest of the read. Buffered reads will not work for
2776 * DAX files, so don't bother trying.
2777 */
2778 if (retval < 0 || !count || IS_DAX(inode))
2779 return retval;
2780 if (iocb->ki_pos >= i_size_read(inode))
2781 return retval;
2782 }
2783
2784 return filemap_read(iocb, iter, retval);
2785}
2786EXPORT_SYMBOL(generic_file_read_iter);
2787
2788/*
2789 * Splice subpages from a folio into a pipe.
2790 */
2791size_t splice_folio_into_pipe(struct pipe_inode_info *pipe,
2792 struct folio *folio, loff_t fpos, size_t size)
2793{
2794 struct page *page;
2795 size_t spliced = 0, offset = offset_in_folio(folio, fpos);
2796
2797 page = folio_page(folio, offset / PAGE_SIZE);
2798 size = min(size, folio_size(folio) - offset);
2799 offset %= PAGE_SIZE;
2800
2801 while (spliced < size &&
2802 !pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
2803 struct pipe_buffer *buf = pipe_head_buf(pipe);
2804 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced);
2805
2806 *buf = (struct pipe_buffer) {
2807 .ops = &page_cache_pipe_buf_ops,
2808 .page = page,
2809 .offset = offset,
2810 .len = part,
2811 };
2812 folio_get(folio);
2813 pipe->head++;
2814 page++;
2815 spliced += part;
2816 offset = 0;
2817 }
2818
2819 return spliced;
2820}
2821
2822/**
2823 * filemap_splice_read - Splice data from a file's pagecache into a pipe
2824 * @in: The file to read from
2825 * @ppos: Pointer to the file position to read from
2826 * @pipe: The pipe to splice into
2827 * @len: The amount to splice
2828 * @flags: The SPLICE_F_* flags
2829 *
2830 * This function gets folios from a file's pagecache and splices them into the
2831 * pipe. Readahead will be called as necessary to fill more folios. This may
2832 * be used for blockdevs also.
2833 *
2834 * Return: On success, the number of bytes read will be returned and *@ppos
2835 * will be updated if appropriate; 0 will be returned if there is no more data
2836 * to be read; -EAGAIN will be returned if the pipe had no space, and some
2837 * other negative error code will be returned on error. A short read may occur
2838 * if the pipe has insufficient space, we reach the end of the data or we hit a
2839 * hole.
2840 */
2841ssize_t filemap_splice_read(struct file *in, loff_t *ppos,
2842 struct pipe_inode_info *pipe,
2843 size_t len, unsigned int flags)
2844{
2845 struct folio_batch fbatch;
2846 struct kiocb iocb;
2847 size_t total_spliced = 0, used, npages;
2848 loff_t isize, end_offset;
2849 bool writably_mapped;
2850 int i, error = 0;
2851
2852 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes))
2853 return 0;
2854
2855 init_sync_kiocb(&iocb, in);
2856 iocb.ki_pos = *ppos;
2857
2858 /* Work out how much data we can actually add into the pipe */
2859 used = pipe_occupancy(pipe->head, pipe->tail);
2860 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2861 len = min_t(size_t, len, npages * PAGE_SIZE);
2862
2863 folio_batch_init(&fbatch);
2864
2865 do {
2866 cond_resched();
2867
2868 if (*ppos >= i_size_read(in->f_mapping->host))
2869 break;
2870
2871 iocb.ki_pos = *ppos;
2872 error = filemap_get_pages(&iocb, len, &fbatch, true);
2873 if (error < 0)
2874 break;
2875
2876 /*
2877 * i_size must be checked after we know the pages are Uptodate.
2878 *
2879 * Checking i_size after the check allows us to calculate
2880 * the correct value for "nr", which means the zero-filled
2881 * part of the page is not copied back to userspace (unless
2882 * another truncate extends the file - this is desired though).
2883 */
2884 isize = i_size_read(in->f_mapping->host);
2885 if (unlikely(*ppos >= isize))
2886 break;
2887 end_offset = min_t(loff_t, isize, *ppos + len);
2888
2889 /*
2890 * Once we start copying data, we don't want to be touching any
2891 * cachelines that might be contended:
2892 */
2893 writably_mapped = mapping_writably_mapped(in->f_mapping);
2894
2895 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2896 struct folio *folio = fbatch.folios[i];
2897 size_t n;
2898
2899 if (folio_pos(folio) >= end_offset)
2900 goto out;
2901 folio_mark_accessed(folio);
2902
2903 /*
2904 * If users can be writing to this folio using arbitrary
2905 * virtual addresses, take care of potential aliasing
2906 * before reading the folio on the kernel side.
2907 */
2908 if (writably_mapped)
2909 flush_dcache_folio(folio);
2910
2911 n = min_t(loff_t, len, isize - *ppos);
2912 n = splice_folio_into_pipe(pipe, folio, *ppos, n);
2913 if (!n)
2914 goto out;
2915 len -= n;
2916 total_spliced += n;
2917 *ppos += n;
2918 in->f_ra.prev_pos = *ppos;
2919 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
2920 goto out;
2921 }
2922
2923 folio_batch_release(&fbatch);
2924 } while (len);
2925
2926out:
2927 folio_batch_release(&fbatch);
2928 file_accessed(in);
2929
2930 return total_spliced ? total_spliced : error;
2931}
2932EXPORT_SYMBOL(filemap_splice_read);
2933
2934static inline loff_t folio_seek_hole_data(struct xa_state *xas,
2935 struct address_space *mapping, struct folio *folio,
2936 loff_t start, loff_t end, bool seek_data)
2937{
2938 const struct address_space_operations *ops = mapping->a_ops;
2939 size_t offset, bsz = i_blocksize(mapping->host);
2940
2941 if (xa_is_value(folio) || folio_test_uptodate(folio))
2942 return seek_data ? start : end;
2943 if (!ops->is_partially_uptodate)
2944 return seek_data ? end : start;
2945
2946 xas_pause(xas);
2947 rcu_read_unlock();
2948 folio_lock(folio);
2949 if (unlikely(folio->mapping != mapping))
2950 goto unlock;
2951
2952 offset = offset_in_folio(folio, start) & ~(bsz - 1);
2953
2954 do {
2955 if (ops->is_partially_uptodate(folio, offset, bsz) ==
2956 seek_data)
2957 break;
2958 start = (start + bsz) & ~(bsz - 1);
2959 offset += bsz;
2960 } while (offset < folio_size(folio));
2961unlock:
2962 folio_unlock(folio);
2963 rcu_read_lock();
2964 return start;
2965}
2966
2967static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
2968{
2969 if (xa_is_value(folio))
2970 return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index);
2971 return folio_size(folio);
2972}
2973
2974/**
2975 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
2976 * @mapping: Address space to search.
2977 * @start: First byte to consider.
2978 * @end: Limit of search (exclusive).
2979 * @whence: Either SEEK_HOLE or SEEK_DATA.
2980 *
2981 * If the page cache knows which blocks contain holes and which blocks
2982 * contain data, your filesystem can use this function to implement
2983 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
2984 * entirely memory-based such as tmpfs, and filesystems which support
2985 * unwritten extents.
2986 *
2987 * Return: The requested offset on success, or -ENXIO if @whence specifies
2988 * SEEK_DATA and there is no data after @start. There is an implicit hole
2989 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
2990 * and @end contain data.
2991 */
2992loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
2993 loff_t end, int whence)
2994{
2995 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
2996 pgoff_t max = (end - 1) >> PAGE_SHIFT;
2997 bool seek_data = (whence == SEEK_DATA);
2998 struct folio *folio;
2999
3000 if (end <= start)
3001 return -ENXIO;
3002
3003 rcu_read_lock();
3004 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
3005 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
3006 size_t seek_size;
3007
3008 if (start < pos) {
3009 if (!seek_data)
3010 goto unlock;
3011 start = pos;
3012 }
3013
3014 seek_size = seek_folio_size(&xas, folio);
3015 pos = round_up((u64)pos + 1, seek_size);
3016 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
3017 seek_data);
3018 if (start < pos)
3019 goto unlock;
3020 if (start >= end)
3021 break;
3022 if (seek_size > PAGE_SIZE)
3023 xas_set(&xas, pos >> PAGE_SHIFT);
3024 if (!xa_is_value(folio))
3025 folio_put(folio);
3026 }
3027 if (seek_data)
3028 start = -ENXIO;
3029unlock:
3030 rcu_read_unlock();
3031 if (folio && !xa_is_value(folio))
3032 folio_put(folio);
3033 if (start > end)
3034 return end;
3035 return start;
3036}
3037
3038#ifdef CONFIG_MMU
3039#define MMAP_LOTSAMISS (100)
3040/*
3041 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
3042 * @vmf - the vm_fault for this fault.
3043 * @folio - the folio to lock.
3044 * @fpin - the pointer to the file we may pin (or is already pinned).
3045 *
3046 * This works similar to lock_folio_or_retry in that it can drop the
3047 * mmap_lock. It differs in that it actually returns the folio locked
3048 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
3049 * to drop the mmap_lock then fpin will point to the pinned file and
3050 * needs to be fput()'ed at a later point.
3051 */
3052static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
3053 struct file **fpin)
3054{
3055 if (folio_trylock(folio))
3056 return 1;
3057
3058 /*
3059 * NOTE! This will make us return with VM_FAULT_RETRY, but with
3060 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
3061 * is supposed to work. We have way too many special cases..
3062 */
3063 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
3064 return 0;
3065
3066 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
3067 if (vmf->flags & FAULT_FLAG_KILLABLE) {
3068 if (__folio_lock_killable(folio)) {
3069 /*
3070 * We didn't have the right flags to drop the
3071 * fault lock, but all fault_handlers only check
3072 * for fatal signals if we return VM_FAULT_RETRY,
3073 * so we need to drop the fault lock here and
3074 * return 0 if we don't have a fpin.
3075 */
3076 if (*fpin == NULL)
3077 release_fault_lock(vmf);
3078 return 0;
3079 }
3080 } else
3081 __folio_lock(folio);
3082
3083 return 1;
3084}
3085
3086/*
3087 * Synchronous readahead happens when we don't even find a page in the page
3088 * cache at all. We don't want to perform IO under the mmap sem, so if we have
3089 * to drop the mmap sem we return the file that was pinned in order for us to do
3090 * that. If we didn't pin a file then we return NULL. The file that is
3091 * returned needs to be fput()'ed when we're done with it.
3092 */
3093static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3094{
3095 struct file *file = vmf->vma->vm_file;
3096 struct file_ra_state *ra = &file->f_ra;
3097 struct address_space *mapping = file->f_mapping;
3098 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3099 struct file *fpin = NULL;
3100 unsigned long vm_flags = vmf->vma->vm_flags;
3101 unsigned int mmap_miss;
3102
3103#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3104 /* Use the readahead code, even if readahead is disabled */
3105 if (vm_flags & VM_HUGEPAGE) {
3106 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3107 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3108 ra->size = HPAGE_PMD_NR;
3109 /*
3110 * Fetch two PMD folios, so we get the chance to actually
3111 * readahead, unless we've been told not to.
3112 */
3113 if (!(vm_flags & VM_RAND_READ))
3114 ra->size *= 2;
3115 ra->async_size = HPAGE_PMD_NR;
3116 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3117 return fpin;
3118 }
3119#endif
3120
3121 /* If we don't want any read-ahead, don't bother */
3122 if (vm_flags & VM_RAND_READ)
3123 return fpin;
3124 if (!ra->ra_pages)
3125 return fpin;
3126
3127 if (vm_flags & VM_SEQ_READ) {
3128 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3129 page_cache_sync_ra(&ractl, ra->ra_pages);
3130 return fpin;
3131 }
3132
3133 /* Avoid banging the cache line if not needed */
3134 mmap_miss = READ_ONCE(ra->mmap_miss);
3135 if (mmap_miss < MMAP_LOTSAMISS * 10)
3136 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3137
3138 /*
3139 * Do we miss much more than hit in this file? If so,
3140 * stop bothering with read-ahead. It will only hurt.
3141 */
3142 if (mmap_miss > MMAP_LOTSAMISS)
3143 return fpin;
3144
3145 /*
3146 * mmap read-around
3147 */
3148 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3149 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3150 ra->size = ra->ra_pages;
3151 ra->async_size = ra->ra_pages / 4;
3152 ractl._index = ra->start;
3153 page_cache_ra_order(&ractl, ra, 0);
3154 return fpin;
3155}
3156
3157/*
3158 * Asynchronous readahead happens when we find the page and PG_readahead,
3159 * so we want to possibly extend the readahead further. We return the file that
3160 * was pinned if we have to drop the mmap_lock in order to do IO.
3161 */
3162static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3163 struct folio *folio)
3164{
3165 struct file *file = vmf->vma->vm_file;
3166 struct file_ra_state *ra = &file->f_ra;
3167 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3168 struct file *fpin = NULL;
3169 unsigned int mmap_miss;
3170
3171 /* If we don't want any read-ahead, don't bother */
3172 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3173 return fpin;
3174
3175 mmap_miss = READ_ONCE(ra->mmap_miss);
3176 if (mmap_miss)
3177 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3178
3179 if (folio_test_readahead(folio)) {
3180 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3181 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3182 }
3183 return fpin;
3184}
3185
3186/**
3187 * filemap_fault - read in file data for page fault handling
3188 * @vmf: struct vm_fault containing details of the fault
3189 *
3190 * filemap_fault() is invoked via the vma operations vector for a
3191 * mapped memory region to read in file data during a page fault.
3192 *
3193 * The goto's are kind of ugly, but this streamlines the normal case of having
3194 * it in the page cache, and handles the special cases reasonably without
3195 * having a lot of duplicated code.
3196 *
3197 * vma->vm_mm->mmap_lock must be held on entry.
3198 *
3199 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3200 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3201 *
3202 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3203 * has not been released.
3204 *
3205 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3206 *
3207 * Return: bitwise-OR of %VM_FAULT_ codes.
3208 */
3209vm_fault_t filemap_fault(struct vm_fault *vmf)
3210{
3211 int error;
3212 struct file *file = vmf->vma->vm_file;
3213 struct file *fpin = NULL;
3214 struct address_space *mapping = file->f_mapping;
3215 struct inode *inode = mapping->host;
3216 pgoff_t max_idx, index = vmf->pgoff;
3217 struct folio *folio;
3218 vm_fault_t ret = 0;
3219 bool mapping_locked = false;
3220
3221 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3222 if (unlikely(index >= max_idx))
3223 return VM_FAULT_SIGBUS;
3224
3225 /*
3226 * Do we have something in the page cache already?
3227 */
3228 folio = filemap_get_folio(mapping, index);
3229 if (likely(!IS_ERR(folio))) {
3230 /*
3231 * We found the page, so try async readahead before waiting for
3232 * the lock.
3233 */
3234 if (!(vmf->flags & FAULT_FLAG_TRIED))
3235 fpin = do_async_mmap_readahead(vmf, folio);
3236 if (unlikely(!folio_test_uptodate(folio))) {
3237 filemap_invalidate_lock_shared(mapping);
3238 mapping_locked = true;
3239 }
3240 } else {
3241 /* No page in the page cache at all */
3242 count_vm_event(PGMAJFAULT);
3243 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3244 ret = VM_FAULT_MAJOR;
3245 fpin = do_sync_mmap_readahead(vmf);
3246retry_find:
3247 /*
3248 * See comment in filemap_create_folio() why we need
3249 * invalidate_lock
3250 */
3251 if (!mapping_locked) {
3252 filemap_invalidate_lock_shared(mapping);
3253 mapping_locked = true;
3254 }
3255 folio = __filemap_get_folio(mapping, index,
3256 FGP_CREAT|FGP_FOR_MMAP,
3257 vmf->gfp_mask);
3258 if (IS_ERR(folio)) {
3259 if (fpin)
3260 goto out_retry;
3261 filemap_invalidate_unlock_shared(mapping);
3262 return VM_FAULT_OOM;
3263 }
3264 }
3265
3266 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3267 goto out_retry;
3268
3269 /* Did it get truncated? */
3270 if (unlikely(folio->mapping != mapping)) {
3271 folio_unlock(folio);
3272 folio_put(folio);
3273 goto retry_find;
3274 }
3275 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3276
3277 /*
3278 * We have a locked folio in the page cache, now we need to check
3279 * that it's up-to-date. If not, it is going to be due to an error,
3280 * or because readahead was otherwise unable to retrieve it.
3281 */
3282 if (unlikely(!folio_test_uptodate(folio))) {
3283 /*
3284 * If the invalidate lock is not held, the folio was in cache
3285 * and uptodate and now it is not. Strange but possible since we
3286 * didn't hold the page lock all the time. Let's drop
3287 * everything, get the invalidate lock and try again.
3288 */
3289 if (!mapping_locked) {
3290 folio_unlock(folio);
3291 folio_put(folio);
3292 goto retry_find;
3293 }
3294
3295 /*
3296 * OK, the folio is really not uptodate. This can be because the
3297 * VMA has the VM_RAND_READ flag set, or because an error
3298 * arose. Let's read it in directly.
3299 */
3300 goto page_not_uptodate;
3301 }
3302
3303 /*
3304 * We've made it this far and we had to drop our mmap_lock, now is the
3305 * time to return to the upper layer and have it re-find the vma and
3306 * redo the fault.
3307 */
3308 if (fpin) {
3309 folio_unlock(folio);
3310 goto out_retry;
3311 }
3312 if (mapping_locked)
3313 filemap_invalidate_unlock_shared(mapping);
3314
3315 /*
3316 * Found the page and have a reference on it.
3317 * We must recheck i_size under page lock.
3318 */
3319 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3320 if (unlikely(index >= max_idx)) {
3321 folio_unlock(folio);
3322 folio_put(folio);
3323 return VM_FAULT_SIGBUS;
3324 }
3325
3326 vmf->page = folio_file_page(folio, index);
3327 return ret | VM_FAULT_LOCKED;
3328
3329page_not_uptodate:
3330 /*
3331 * Umm, take care of errors if the page isn't up-to-date.
3332 * Try to re-read it _once_. We do this synchronously,
3333 * because there really aren't any performance issues here
3334 * and we need to check for errors.
3335 */
3336 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3337 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3338 if (fpin)
3339 goto out_retry;
3340 folio_put(folio);
3341
3342 if (!error || error == AOP_TRUNCATED_PAGE)
3343 goto retry_find;
3344 filemap_invalidate_unlock_shared(mapping);
3345
3346 return VM_FAULT_SIGBUS;
3347
3348out_retry:
3349 /*
3350 * We dropped the mmap_lock, we need to return to the fault handler to
3351 * re-find the vma and come back and find our hopefully still populated
3352 * page.
3353 */
3354 if (!IS_ERR(folio))
3355 folio_put(folio);
3356 if (mapping_locked)
3357 filemap_invalidate_unlock_shared(mapping);
3358 if (fpin)
3359 fput(fpin);
3360 return ret | VM_FAULT_RETRY;
3361}
3362EXPORT_SYMBOL(filemap_fault);
3363
3364static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio,
3365 pgoff_t start)
3366{
3367 struct mm_struct *mm = vmf->vma->vm_mm;
3368
3369 /* Huge page is mapped? No need to proceed. */
3370 if (pmd_trans_huge(*vmf->pmd)) {
3371 folio_unlock(folio);
3372 folio_put(folio);
3373 return true;
3374 }
3375
3376 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) {
3377 struct page *page = folio_file_page(folio, start);
3378 vm_fault_t ret = do_set_pmd(vmf, page);
3379 if (!ret) {
3380 /* The page is mapped successfully, reference consumed. */
3381 folio_unlock(folio);
3382 return true;
3383 }
3384 }
3385
3386 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte)
3387 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3388
3389 return false;
3390}
3391
3392static struct folio *next_uptodate_folio(struct xa_state *xas,
3393 struct address_space *mapping, pgoff_t end_pgoff)
3394{
3395 struct folio *folio = xas_next_entry(xas, end_pgoff);
3396 unsigned long max_idx;
3397
3398 do {
3399 if (!folio)
3400 return NULL;
3401 if (xas_retry(xas, folio))
3402 continue;
3403 if (xa_is_value(folio))
3404 continue;
3405 if (folio_test_locked(folio))
3406 continue;
3407 if (!folio_try_get_rcu(folio))
3408 continue;
3409 /* Has the page moved or been split? */
3410 if (unlikely(folio != xas_reload(xas)))
3411 goto skip;
3412 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3413 goto skip;
3414 if (!folio_trylock(folio))
3415 goto skip;
3416 if (folio->mapping != mapping)
3417 goto unlock;
3418 if (!folio_test_uptodate(folio))
3419 goto unlock;
3420 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3421 if (xas->xa_index >= max_idx)
3422 goto unlock;
3423 return folio;
3424unlock:
3425 folio_unlock(folio);
3426skip:
3427 folio_put(folio);
3428 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3429
3430 return NULL;
3431}
3432
3433/*
3434 * Map page range [start_page, start_page + nr_pages) of folio.
3435 * start_page is gotten from start by folio_page(folio, start)
3436 */
3437static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
3438 struct folio *folio, unsigned long start,
3439 unsigned long addr, unsigned int nr_pages,
3440 unsigned int *mmap_miss)
3441{
3442 vm_fault_t ret = 0;
3443 struct page *page = folio_page(folio, start);
3444 unsigned int count = 0;
3445 pte_t *old_ptep = vmf->pte;
3446
3447 do {
3448 if (PageHWPoison(page + count))
3449 goto skip;
3450
3451 (*mmap_miss)++;
3452
3453 /*
3454 * NOTE: If there're PTE markers, we'll leave them to be
3455 * handled in the specific fault path, and it'll prohibit the
3456 * fault-around logic.
3457 */
3458 if (!pte_none(ptep_get(&vmf->pte[count])))
3459 goto skip;
3460
3461 count++;
3462 continue;
3463skip:
3464 if (count) {
3465 set_pte_range(vmf, folio, page, count, addr);
3466 folio_ref_add(folio, count);
3467 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3468 ret = VM_FAULT_NOPAGE;
3469 }
3470
3471 count++;
3472 page += count;
3473 vmf->pte += count;
3474 addr += count * PAGE_SIZE;
3475 count = 0;
3476 } while (--nr_pages > 0);
3477
3478 if (count) {
3479 set_pte_range(vmf, folio, page, count, addr);
3480 folio_ref_add(folio, count);
3481 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3482 ret = VM_FAULT_NOPAGE;
3483 }
3484
3485 vmf->pte = old_ptep;
3486
3487 return ret;
3488}
3489
3490static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
3491 struct folio *folio, unsigned long addr,
3492 unsigned int *mmap_miss)
3493{
3494 vm_fault_t ret = 0;
3495 struct page *page = &folio->page;
3496
3497 if (PageHWPoison(page))
3498 return ret;
3499
3500 (*mmap_miss)++;
3501
3502 /*
3503 * NOTE: If there're PTE markers, we'll leave them to be
3504 * handled in the specific fault path, and it'll prohibit
3505 * the fault-around logic.
3506 */
3507 if (!pte_none(ptep_get(vmf->pte)))
3508 return ret;
3509
3510 if (vmf->address == addr)
3511 ret = VM_FAULT_NOPAGE;
3512
3513 set_pte_range(vmf, folio, page, 1, addr);
3514 folio_ref_inc(folio);
3515
3516 return ret;
3517}
3518
3519vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3520 pgoff_t start_pgoff, pgoff_t end_pgoff)
3521{
3522 struct vm_area_struct *vma = vmf->vma;
3523 struct file *file = vma->vm_file;
3524 struct address_space *mapping = file->f_mapping;
3525 pgoff_t last_pgoff = start_pgoff;
3526 unsigned long addr;
3527 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3528 struct folio *folio;
3529 vm_fault_t ret = 0;
3530 unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved;
3531
3532 rcu_read_lock();
3533 folio = next_uptodate_folio(&xas, mapping, end_pgoff);
3534 if (!folio)
3535 goto out;
3536
3537 if (filemap_map_pmd(vmf, folio, start_pgoff)) {
3538 ret = VM_FAULT_NOPAGE;
3539 goto out;
3540 }
3541
3542 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3543 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3544 if (!vmf->pte) {
3545 folio_unlock(folio);
3546 folio_put(folio);
3547 goto out;
3548 }
3549 do {
3550 unsigned long end;
3551
3552 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3553 vmf->pte += xas.xa_index - last_pgoff;
3554 last_pgoff = xas.xa_index;
3555 end = folio_next_index(folio) - 1;
3556 nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
3557
3558 if (!folio_test_large(folio))
3559 ret |= filemap_map_order0_folio(vmf,
3560 folio, addr, &mmap_miss);
3561 else
3562 ret |= filemap_map_folio_range(vmf, folio,
3563 xas.xa_index - folio->index, addr,
3564 nr_pages, &mmap_miss);
3565
3566 folio_unlock(folio);
3567 folio_put(folio);
3568 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL);
3569 pte_unmap_unlock(vmf->pte, vmf->ptl);
3570out:
3571 rcu_read_unlock();
3572
3573 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss);
3574 if (mmap_miss >= mmap_miss_saved)
3575 WRITE_ONCE(file->f_ra.mmap_miss, 0);
3576 else
3577 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss);
3578
3579 return ret;
3580}
3581EXPORT_SYMBOL(filemap_map_pages);
3582
3583vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3584{
3585 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3586 struct folio *folio = page_folio(vmf->page);
3587 vm_fault_t ret = VM_FAULT_LOCKED;
3588
3589 sb_start_pagefault(mapping->host->i_sb);
3590 file_update_time(vmf->vma->vm_file);
3591 folio_lock(folio);
3592 if (folio->mapping != mapping) {
3593 folio_unlock(folio);
3594 ret = VM_FAULT_NOPAGE;
3595 goto out;
3596 }
3597 /*
3598 * We mark the folio dirty already here so that when freeze is in
3599 * progress, we are guaranteed that writeback during freezing will
3600 * see the dirty folio and writeprotect it again.
3601 */
3602 folio_mark_dirty(folio);
3603 folio_wait_stable(folio);
3604out:
3605 sb_end_pagefault(mapping->host->i_sb);
3606 return ret;
3607}
3608
3609const struct vm_operations_struct generic_file_vm_ops = {
3610 .fault = filemap_fault,
3611 .map_pages = filemap_map_pages,
3612 .page_mkwrite = filemap_page_mkwrite,
3613};
3614
3615/* This is used for a general mmap of a disk file */
3616
3617int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3618{
3619 struct address_space *mapping = file->f_mapping;
3620
3621 if (!mapping->a_ops->read_folio)
3622 return -ENOEXEC;
3623 file_accessed(file);
3624 vma->vm_ops = &generic_file_vm_ops;
3625 return 0;
3626}
3627
3628/*
3629 * This is for filesystems which do not implement ->writepage.
3630 */
3631int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3632{
3633 if (vma_is_shared_maywrite(vma))
3634 return -EINVAL;
3635 return generic_file_mmap(file, vma);
3636}
3637#else
3638vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3639{
3640 return VM_FAULT_SIGBUS;
3641}
3642int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3643{
3644 return -ENOSYS;
3645}
3646int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3647{
3648 return -ENOSYS;
3649}
3650#endif /* CONFIG_MMU */
3651
3652EXPORT_SYMBOL(filemap_page_mkwrite);
3653EXPORT_SYMBOL(generic_file_mmap);
3654EXPORT_SYMBOL(generic_file_readonly_mmap);
3655
3656static struct folio *do_read_cache_folio(struct address_space *mapping,
3657 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3658{
3659 struct folio *folio;
3660 int err;
3661
3662 if (!filler)
3663 filler = mapping->a_ops->read_folio;
3664repeat:
3665 folio = filemap_get_folio(mapping, index);
3666 if (IS_ERR(folio)) {
3667 folio = filemap_alloc_folio(gfp, 0);
3668 if (!folio)
3669 return ERR_PTR(-ENOMEM);
3670 err = filemap_add_folio(mapping, folio, index, gfp);
3671 if (unlikely(err)) {
3672 folio_put(folio);
3673 if (err == -EEXIST)
3674 goto repeat;
3675 /* Presumably ENOMEM for xarray node */
3676 return ERR_PTR(err);
3677 }
3678
3679 goto filler;
3680 }
3681 if (folio_test_uptodate(folio))
3682 goto out;
3683
3684 if (!folio_trylock(folio)) {
3685 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3686 goto repeat;
3687 }
3688
3689 /* Folio was truncated from mapping */
3690 if (!folio->mapping) {
3691 folio_unlock(folio);
3692 folio_put(folio);
3693 goto repeat;
3694 }
3695
3696 /* Someone else locked and filled the page in a very small window */
3697 if (folio_test_uptodate(folio)) {
3698 folio_unlock(folio);
3699 goto out;
3700 }
3701
3702filler:
3703 err = filemap_read_folio(file, filler, folio);
3704 if (err) {
3705 folio_put(folio);
3706 if (err == AOP_TRUNCATED_PAGE)
3707 goto repeat;
3708 return ERR_PTR(err);
3709 }
3710
3711out:
3712 folio_mark_accessed(folio);
3713 return folio;
3714}
3715
3716/**
3717 * read_cache_folio - Read into page cache, fill it if needed.
3718 * @mapping: The address_space to read from.
3719 * @index: The index to read.
3720 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3721 * @file: Passed to filler function, may be NULL if not required.
3722 *
3723 * Read one page into the page cache. If it succeeds, the folio returned
3724 * will contain @index, but it may not be the first page of the folio.
3725 *
3726 * If the filler function returns an error, it will be returned to the
3727 * caller.
3728 *
3729 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3730 * Return: An uptodate folio on success, ERR_PTR() on failure.
3731 */
3732struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3733 filler_t filler, struct file *file)
3734{
3735 return do_read_cache_folio(mapping, index, filler, file,
3736 mapping_gfp_mask(mapping));
3737}
3738EXPORT_SYMBOL(read_cache_folio);
3739
3740/**
3741 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags.
3742 * @mapping: The address_space for the folio.
3743 * @index: The index that the allocated folio will contain.
3744 * @gfp: The page allocator flags to use if allocating.
3745 *
3746 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with
3747 * any new memory allocations done using the specified allocation flags.
3748 *
3749 * The most likely error from this function is EIO, but ENOMEM is
3750 * possible and so is EINTR. If ->read_folio returns another error,
3751 * that will be returned to the caller.
3752 *
3753 * The function expects mapping->invalidate_lock to be already held.
3754 *
3755 * Return: Uptodate folio on success, ERR_PTR() on failure.
3756 */
3757struct folio *mapping_read_folio_gfp(struct address_space *mapping,
3758 pgoff_t index, gfp_t gfp)
3759{
3760 return do_read_cache_folio(mapping, index, NULL, NULL, gfp);
3761}
3762EXPORT_SYMBOL(mapping_read_folio_gfp);
3763
3764static struct page *do_read_cache_page(struct address_space *mapping,
3765 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3766{
3767 struct folio *folio;
3768
3769 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3770 if (IS_ERR(folio))
3771 return &folio->page;
3772 return folio_file_page(folio, index);
3773}
3774
3775struct page *read_cache_page(struct address_space *mapping,
3776 pgoff_t index, filler_t *filler, struct file *file)
3777{
3778 return do_read_cache_page(mapping, index, filler, file,
3779 mapping_gfp_mask(mapping));
3780}
3781EXPORT_SYMBOL(read_cache_page);
3782
3783/**
3784 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3785 * @mapping: the page's address_space
3786 * @index: the page index
3787 * @gfp: the page allocator flags to use if allocating
3788 *
3789 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3790 * any new page allocations done using the specified allocation flags.
3791 *
3792 * If the page does not get brought uptodate, return -EIO.
3793 *
3794 * The function expects mapping->invalidate_lock to be already held.
3795 *
3796 * Return: up to date page on success, ERR_PTR() on failure.
3797 */
3798struct page *read_cache_page_gfp(struct address_space *mapping,
3799 pgoff_t index,
3800 gfp_t gfp)
3801{
3802 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3803}
3804EXPORT_SYMBOL(read_cache_page_gfp);
3805
3806/*
3807 * Warn about a page cache invalidation failure during a direct I/O write.
3808 */
3809static void dio_warn_stale_pagecache(struct file *filp)
3810{
3811 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3812 char pathname[128];
3813 char *path;
3814
3815 errseq_set(&filp->f_mapping->wb_err, -EIO);
3816 if (__ratelimit(&_rs)) {
3817 path = file_path(filp, pathname, sizeof(pathname));
3818 if (IS_ERR(path))
3819 path = "(unknown)";
3820 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3821 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3822 current->comm);
3823 }
3824}
3825
3826void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count)
3827{
3828 struct address_space *mapping = iocb->ki_filp->f_mapping;
3829
3830 if (mapping->nrpages &&
3831 invalidate_inode_pages2_range(mapping,
3832 iocb->ki_pos >> PAGE_SHIFT,
3833 (iocb->ki_pos + count - 1) >> PAGE_SHIFT))
3834 dio_warn_stale_pagecache(iocb->ki_filp);
3835}
3836
3837ssize_t
3838generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3839{
3840 struct address_space *mapping = iocb->ki_filp->f_mapping;
3841 size_t write_len = iov_iter_count(from);
3842 ssize_t written;
3843
3844 /*
3845 * If a page can not be invalidated, return 0 to fall back
3846 * to buffered write.
3847 */
3848 written = kiocb_invalidate_pages(iocb, write_len);
3849 if (written) {
3850 if (written == -EBUSY)
3851 return 0;
3852 return written;
3853 }
3854
3855 written = mapping->a_ops->direct_IO(iocb, from);
3856
3857 /*
3858 * Finally, try again to invalidate clean pages which might have been
3859 * cached by non-direct readahead, or faulted in by get_user_pages()
3860 * if the source of the write was an mmap'ed region of the file
3861 * we're writing. Either one is a pretty crazy thing to do,
3862 * so we don't support it 100%. If this invalidation
3863 * fails, tough, the write still worked...
3864 *
3865 * Most of the time we do not need this since dio_complete() will do
3866 * the invalidation for us. However there are some file systems that
3867 * do not end up with dio_complete() being called, so let's not break
3868 * them by removing it completely.
3869 *
3870 * Noticeable example is a blkdev_direct_IO().
3871 *
3872 * Skip invalidation for async writes or if mapping has no pages.
3873 */
3874 if (written > 0) {
3875 struct inode *inode = mapping->host;
3876 loff_t pos = iocb->ki_pos;
3877
3878 kiocb_invalidate_post_direct_write(iocb, written);
3879 pos += written;
3880 write_len -= written;
3881 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3882 i_size_write(inode, pos);
3883 mark_inode_dirty(inode);
3884 }
3885 iocb->ki_pos = pos;
3886 }
3887 if (written != -EIOCBQUEUED)
3888 iov_iter_revert(from, write_len - iov_iter_count(from));
3889 return written;
3890}
3891EXPORT_SYMBOL(generic_file_direct_write);
3892
3893ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
3894{
3895 struct file *file = iocb->ki_filp;
3896 loff_t pos = iocb->ki_pos;
3897 struct address_space *mapping = file->f_mapping;
3898 const struct address_space_operations *a_ops = mapping->a_ops;
3899 long status = 0;
3900 ssize_t written = 0;
3901
3902 do {
3903 struct page *page;
3904 unsigned long offset; /* Offset into pagecache page */
3905 unsigned long bytes; /* Bytes to write to page */
3906 size_t copied; /* Bytes copied from user */
3907 void *fsdata = NULL;
3908
3909 offset = (pos & (PAGE_SIZE - 1));
3910 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3911 iov_iter_count(i));
3912
3913again:
3914 /*
3915 * Bring in the user page that we will copy from _first_.
3916 * Otherwise there's a nasty deadlock on copying from the
3917 * same page as we're writing to, without it being marked
3918 * up-to-date.
3919 */
3920 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
3921 status = -EFAULT;
3922 break;
3923 }
3924
3925 if (fatal_signal_pending(current)) {
3926 status = -EINTR;
3927 break;
3928 }
3929
3930 status = a_ops->write_begin(file, mapping, pos, bytes,
3931 &page, &fsdata);
3932 if (unlikely(status < 0))
3933 break;
3934
3935 if (mapping_writably_mapped(mapping))
3936 flush_dcache_page(page);
3937
3938 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
3939 flush_dcache_page(page);
3940
3941 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3942 page, fsdata);
3943 if (unlikely(status != copied)) {
3944 iov_iter_revert(i, copied - max(status, 0L));
3945 if (unlikely(status < 0))
3946 break;
3947 }
3948 cond_resched();
3949
3950 if (unlikely(status == 0)) {
3951 /*
3952 * A short copy made ->write_end() reject the
3953 * thing entirely. Might be memory poisoning
3954 * halfway through, might be a race with munmap,
3955 * might be severe memory pressure.
3956 */
3957 if (copied)
3958 bytes = copied;
3959 goto again;
3960 }
3961 pos += status;
3962 written += status;
3963
3964 balance_dirty_pages_ratelimited(mapping);
3965 } while (iov_iter_count(i));
3966
3967 if (!written)
3968 return status;
3969 iocb->ki_pos += written;
3970 return written;
3971}
3972EXPORT_SYMBOL(generic_perform_write);
3973
3974/**
3975 * __generic_file_write_iter - write data to a file
3976 * @iocb: IO state structure (file, offset, etc.)
3977 * @from: iov_iter with data to write
3978 *
3979 * This function does all the work needed for actually writing data to a
3980 * file. It does all basic checks, removes SUID from the file, updates
3981 * modification times and calls proper subroutines depending on whether we
3982 * do direct IO or a standard buffered write.
3983 *
3984 * It expects i_rwsem to be grabbed unless we work on a block device or similar
3985 * object which does not need locking at all.
3986 *
3987 * This function does *not* take care of syncing data in case of O_SYNC write.
3988 * A caller has to handle it. This is mainly due to the fact that we want to
3989 * avoid syncing under i_rwsem.
3990 *
3991 * Return:
3992 * * number of bytes written, even for truncated writes
3993 * * negative error code if no data has been written at all
3994 */
3995ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3996{
3997 struct file *file = iocb->ki_filp;
3998 struct address_space *mapping = file->f_mapping;
3999 struct inode *inode = mapping->host;
4000 ssize_t ret;
4001
4002 ret = file_remove_privs(file);
4003 if (ret)
4004 return ret;
4005
4006 ret = file_update_time(file);
4007 if (ret)
4008 return ret;
4009
4010 if (iocb->ki_flags & IOCB_DIRECT) {
4011 ret = generic_file_direct_write(iocb, from);
4012 /*
4013 * If the write stopped short of completing, fall back to
4014 * buffered writes. Some filesystems do this for writes to
4015 * holes, for example. For DAX files, a buffered write will
4016 * not succeed (even if it did, DAX does not handle dirty
4017 * page-cache pages correctly).
4018 */
4019 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode))
4020 return ret;
4021 return direct_write_fallback(iocb, from, ret,
4022 generic_perform_write(iocb, from));
4023 }
4024
4025 return generic_perform_write(iocb, from);
4026}
4027EXPORT_SYMBOL(__generic_file_write_iter);
4028
4029/**
4030 * generic_file_write_iter - write data to a file
4031 * @iocb: IO state structure
4032 * @from: iov_iter with data to write
4033 *
4034 * This is a wrapper around __generic_file_write_iter() to be used by most
4035 * filesystems. It takes care of syncing the file in case of O_SYNC file
4036 * and acquires i_rwsem as needed.
4037 * Return:
4038 * * negative error code if no data has been written at all of
4039 * vfs_fsync_range() failed for a synchronous write
4040 * * number of bytes written, even for truncated writes
4041 */
4042ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4043{
4044 struct file *file = iocb->ki_filp;
4045 struct inode *inode = file->f_mapping->host;
4046 ssize_t ret;
4047
4048 inode_lock(inode);
4049 ret = generic_write_checks(iocb, from);
4050 if (ret > 0)
4051 ret = __generic_file_write_iter(iocb, from);
4052 inode_unlock(inode);
4053
4054 if (ret > 0)
4055 ret = generic_write_sync(iocb, ret);
4056 return ret;
4057}
4058EXPORT_SYMBOL(generic_file_write_iter);
4059
4060/**
4061 * filemap_release_folio() - Release fs-specific metadata on a folio.
4062 * @folio: The folio which the kernel is trying to free.
4063 * @gfp: Memory allocation flags (and I/O mode).
4064 *
4065 * The address_space is trying to release any data attached to a folio
4066 * (presumably at folio->private).
4067 *
4068 * This will also be called if the private_2 flag is set on a page,
4069 * indicating that the folio has other metadata associated with it.
4070 *
4071 * The @gfp argument specifies whether I/O may be performed to release
4072 * this page (__GFP_IO), and whether the call may block
4073 * (__GFP_RECLAIM & __GFP_FS).
4074 *
4075 * Return: %true if the release was successful, otherwise %false.
4076 */
4077bool filemap_release_folio(struct folio *folio, gfp_t gfp)
4078{
4079 struct address_space * const mapping = folio->mapping;
4080
4081 BUG_ON(!folio_test_locked(folio));
4082 if (!folio_needs_release(folio))
4083 return true;
4084 if (folio_test_writeback(folio))
4085 return false;
4086
4087 if (mapping && mapping->a_ops->release_folio)
4088 return mapping->a_ops->release_folio(folio, gfp);
4089 return try_to_free_buffers(folio);
4090}
4091EXPORT_SYMBOL(filemap_release_folio);
4092
4093#ifdef CONFIG_CACHESTAT_SYSCALL
4094/**
4095 * filemap_cachestat() - compute the page cache statistics of a mapping
4096 * @mapping: The mapping to compute the statistics for.
4097 * @first_index: The starting page cache index.
4098 * @last_index: The final page index (inclusive).
4099 * @cs: the cachestat struct to write the result to.
4100 *
4101 * This will query the page cache statistics of a mapping in the
4102 * page range of [first_index, last_index] (inclusive). The statistics
4103 * queried include: number of dirty pages, number of pages marked for
4104 * writeback, and the number of (recently) evicted pages.
4105 */
4106static void filemap_cachestat(struct address_space *mapping,
4107 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs)
4108{
4109 XA_STATE(xas, &mapping->i_pages, first_index);
4110 struct folio *folio;
4111
4112 rcu_read_lock();
4113 xas_for_each(&xas, folio, last_index) {
4114 int order;
4115 unsigned long nr_pages;
4116 pgoff_t folio_first_index, folio_last_index;
4117
4118 /*
4119 * Don't deref the folio. It is not pinned, and might
4120 * get freed (and reused) underneath us.
4121 *
4122 * We *could* pin it, but that would be expensive for
4123 * what should be a fast and lightweight syscall.
4124 *
4125 * Instead, derive all information of interest from
4126 * the rcu-protected xarray.
4127 */
4128
4129 if (xas_retry(&xas, folio))
4130 continue;
4131
4132 order = xa_get_order(xas.xa, xas.xa_index);
4133 nr_pages = 1 << order;
4134 folio_first_index = round_down(xas.xa_index, 1 << order);
4135 folio_last_index = folio_first_index + nr_pages - 1;
4136
4137 /* Folios might straddle the range boundaries, only count covered pages */
4138 if (folio_first_index < first_index)
4139 nr_pages -= first_index - folio_first_index;
4140
4141 if (folio_last_index > last_index)
4142 nr_pages -= folio_last_index - last_index;
4143
4144 if (xa_is_value(folio)) {
4145 /* page is evicted */
4146 void *shadow = (void *)folio;
4147 bool workingset; /* not used */
4148
4149 cs->nr_evicted += nr_pages;
4150
4151#ifdef CONFIG_SWAP /* implies CONFIG_MMU */
4152 if (shmem_mapping(mapping)) {
4153 /* shmem file - in swap cache */
4154 swp_entry_t swp = radix_to_swp_entry(folio);
4155
4156 shadow = get_shadow_from_swap_cache(swp);
4157 }
4158#endif
4159 if (workingset_test_recent(shadow, true, &workingset))
4160 cs->nr_recently_evicted += nr_pages;
4161
4162 goto resched;
4163 }
4164
4165 /* page is in cache */
4166 cs->nr_cache += nr_pages;
4167
4168 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY))
4169 cs->nr_dirty += nr_pages;
4170
4171 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK))
4172 cs->nr_writeback += nr_pages;
4173
4174resched:
4175 if (need_resched()) {
4176 xas_pause(&xas);
4177 cond_resched_rcu();
4178 }
4179 }
4180 rcu_read_unlock();
4181}
4182
4183/*
4184 * The cachestat(2) system call.
4185 *
4186 * cachestat() returns the page cache statistics of a file in the
4187 * bytes range specified by `off` and `len`: number of cached pages,
4188 * number of dirty pages, number of pages marked for writeback,
4189 * number of evicted pages, and number of recently evicted pages.
4190 *
4191 * An evicted page is a page that is previously in the page cache
4192 * but has been evicted since. A page is recently evicted if its last
4193 * eviction was recent enough that its reentry to the cache would
4194 * indicate that it is actively being used by the system, and that
4195 * there is memory pressure on the system.
4196 *
4197 * `off` and `len` must be non-negative integers. If `len` > 0,
4198 * the queried range is [`off`, `off` + `len`]. If `len` == 0,
4199 * we will query in the range from `off` to the end of the file.
4200 *
4201 * The `flags` argument is unused for now, but is included for future
4202 * extensibility. User should pass 0 (i.e no flag specified).
4203 *
4204 * Currently, hugetlbfs is not supported.
4205 *
4206 * Because the status of a page can change after cachestat() checks it
4207 * but before it returns to the application, the returned values may
4208 * contain stale information.
4209 *
4210 * return values:
4211 * zero - success
4212 * -EFAULT - cstat or cstat_range points to an illegal address
4213 * -EINVAL - invalid flags
4214 * -EBADF - invalid file descriptor
4215 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file
4216 */
4217SYSCALL_DEFINE4(cachestat, unsigned int, fd,
4218 struct cachestat_range __user *, cstat_range,
4219 struct cachestat __user *, cstat, unsigned int, flags)
4220{
4221 struct fd f = fdget(fd);
4222 struct address_space *mapping;
4223 struct cachestat_range csr;
4224 struct cachestat cs;
4225 pgoff_t first_index, last_index;
4226
4227 if (!f.file)
4228 return -EBADF;
4229
4230 if (copy_from_user(&csr, cstat_range,
4231 sizeof(struct cachestat_range))) {
4232 fdput(f);
4233 return -EFAULT;
4234 }
4235
4236 /* hugetlbfs is not supported */
4237 if (is_file_hugepages(f.file)) {
4238 fdput(f);
4239 return -EOPNOTSUPP;
4240 }
4241
4242 if (flags != 0) {
4243 fdput(f);
4244 return -EINVAL;
4245 }
4246
4247 first_index = csr.off >> PAGE_SHIFT;
4248 last_index =
4249 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT;
4250 memset(&cs, 0, sizeof(struct cachestat));
4251 mapping = f.file->f_mapping;
4252 filemap_cachestat(mapping, first_index, last_index, &cs);
4253 fdput(f);
4254
4255 if (copy_to_user(cstat, &cs, sizeof(struct cachestat)))
4256 return -EFAULT;
4257
4258 return 0;
4259}
4260#endif /* CONFIG_CACHESTAT_SYSCALL */
1/*
2 * linux/mm/filemap.c
3 *
4 * Copyright (C) 1994-1999 Linus Torvalds
5 */
6
7/*
8 * This file handles the generic file mmap semantics used by
9 * most "normal" filesystems (but you don't /have/ to use this:
10 * the NFS filesystem used to do this differently, for example)
11 */
12#include <linux/export.h>
13#include <linux/compiler.h>
14#include <linux/fs.h>
15#include <linux/uaccess.h>
16#include <linux/aio.h>
17#include <linux/capability.h>
18#include <linux/kernel_stat.h>
19#include <linux/gfp.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/mman.h>
23#include <linux/pagemap.h>
24#include <linux/file.h>
25#include <linux/uio.h>
26#include <linux/hash.h>
27#include <linux/writeback.h>
28#include <linux/backing-dev.h>
29#include <linux/pagevec.h>
30#include <linux/blkdev.h>
31#include <linux/security.h>
32#include <linux/cpuset.h>
33#include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34#include <linux/memcontrol.h>
35#include <linux/cleancache.h>
36#include "internal.h"
37
38/*
39 * FIXME: remove all knowledge of the buffer layer from the core VM
40 */
41#include <linux/buffer_head.h> /* for try_to_free_buffers */
42
43#include <asm/mman.h>
44
45/*
46 * Shared mappings implemented 30.11.1994. It's not fully working yet,
47 * though.
48 *
49 * Shared mappings now work. 15.8.1995 Bruno.
50 *
51 * finished 'unifying' the page and buffer cache and SMP-threaded the
52 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
53 *
54 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
55 */
56
57/*
58 * Lock ordering:
59 *
60 * ->i_mmap_mutex (truncate_pagecache)
61 * ->private_lock (__free_pte->__set_page_dirty_buffers)
62 * ->swap_lock (exclusive_swap_page, others)
63 * ->mapping->tree_lock
64 *
65 * ->i_mutex
66 * ->i_mmap_mutex (truncate->unmap_mapping_range)
67 *
68 * ->mmap_sem
69 * ->i_mmap_mutex
70 * ->page_table_lock or pte_lock (various, mainly in memory.c)
71 * ->mapping->tree_lock (arch-dependent flush_dcache_mmap_lock)
72 *
73 * ->mmap_sem
74 * ->lock_page (access_process_vm)
75 *
76 * ->i_mutex (generic_file_buffered_write)
77 * ->mmap_sem (fault_in_pages_readable->do_page_fault)
78 *
79 * bdi->wb.list_lock
80 * sb_lock (fs/fs-writeback.c)
81 * ->mapping->tree_lock (__sync_single_inode)
82 *
83 * ->i_mmap_mutex
84 * ->anon_vma.lock (vma_adjust)
85 *
86 * ->anon_vma.lock
87 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
88 *
89 * ->page_table_lock or pte_lock
90 * ->swap_lock (try_to_unmap_one)
91 * ->private_lock (try_to_unmap_one)
92 * ->tree_lock (try_to_unmap_one)
93 * ->zone.lru_lock (follow_page->mark_page_accessed)
94 * ->zone.lru_lock (check_pte_range->isolate_lru_page)
95 * ->private_lock (page_remove_rmap->set_page_dirty)
96 * ->tree_lock (page_remove_rmap->set_page_dirty)
97 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
98 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
99 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
100 * ->inode->i_lock (zap_pte_range->set_page_dirty)
101 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
102 *
103 * ->i_mmap_mutex
104 * ->tasklist_lock (memory_failure, collect_procs_ao)
105 */
106
107/*
108 * Delete a page from the page cache and free it. Caller has to make
109 * sure the page is locked and that nobody else uses it - or that usage
110 * is safe. The caller must hold the mapping's tree_lock.
111 */
112void __delete_from_page_cache(struct page *page)
113{
114 struct address_space *mapping = page->mapping;
115
116 /*
117 * if we're uptodate, flush out into the cleancache, otherwise
118 * invalidate any existing cleancache entries. We can't leave
119 * stale data around in the cleancache once our page is gone
120 */
121 if (PageUptodate(page) && PageMappedToDisk(page))
122 cleancache_put_page(page);
123 else
124 cleancache_invalidate_page(mapping, page);
125
126 radix_tree_delete(&mapping->page_tree, page->index);
127 page->mapping = NULL;
128 /* Leave page->index set: truncation lookup relies upon it */
129 mapping->nrpages--;
130 __dec_zone_page_state(page, NR_FILE_PAGES);
131 if (PageSwapBacked(page))
132 __dec_zone_page_state(page, NR_SHMEM);
133 BUG_ON(page_mapped(page));
134
135 /*
136 * Some filesystems seem to re-dirty the page even after
137 * the VM has canceled the dirty bit (eg ext3 journaling).
138 *
139 * Fix it up by doing a final dirty accounting check after
140 * having removed the page entirely.
141 */
142 if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
143 dec_zone_page_state(page, NR_FILE_DIRTY);
144 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
145 }
146}
147
148/**
149 * delete_from_page_cache - delete page from page cache
150 * @page: the page which the kernel is trying to remove from page cache
151 *
152 * This must be called only on pages that have been verified to be in the page
153 * cache and locked. It will never put the page into the free list, the caller
154 * has a reference on the page.
155 */
156void delete_from_page_cache(struct page *page)
157{
158 struct address_space *mapping = page->mapping;
159 void (*freepage)(struct page *);
160
161 BUG_ON(!PageLocked(page));
162
163 freepage = mapping->a_ops->freepage;
164 spin_lock_irq(&mapping->tree_lock);
165 __delete_from_page_cache(page);
166 spin_unlock_irq(&mapping->tree_lock);
167 mem_cgroup_uncharge_cache_page(page);
168
169 if (freepage)
170 freepage(page);
171 page_cache_release(page);
172}
173EXPORT_SYMBOL(delete_from_page_cache);
174
175static int sleep_on_page(void *word)
176{
177 io_schedule();
178 return 0;
179}
180
181static int sleep_on_page_killable(void *word)
182{
183 sleep_on_page(word);
184 return fatal_signal_pending(current) ? -EINTR : 0;
185}
186
187/**
188 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
189 * @mapping: address space structure to write
190 * @start: offset in bytes where the range starts
191 * @end: offset in bytes where the range ends (inclusive)
192 * @sync_mode: enable synchronous operation
193 *
194 * Start writeback against all of a mapping's dirty pages that lie
195 * within the byte offsets <start, end> inclusive.
196 *
197 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
198 * opposed to a regular memory cleansing writeback. The difference between
199 * these two operations is that if a dirty page/buffer is encountered, it must
200 * be waited upon, and not just skipped over.
201 */
202int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
203 loff_t end, int sync_mode)
204{
205 int ret;
206 struct writeback_control wbc = {
207 .sync_mode = sync_mode,
208 .nr_to_write = LONG_MAX,
209 .range_start = start,
210 .range_end = end,
211 };
212
213 if (!mapping_cap_writeback_dirty(mapping))
214 return 0;
215
216 ret = do_writepages(mapping, &wbc);
217 return ret;
218}
219
220static inline int __filemap_fdatawrite(struct address_space *mapping,
221 int sync_mode)
222{
223 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
224}
225
226int filemap_fdatawrite(struct address_space *mapping)
227{
228 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
229}
230EXPORT_SYMBOL(filemap_fdatawrite);
231
232int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
233 loff_t end)
234{
235 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
236}
237EXPORT_SYMBOL(filemap_fdatawrite_range);
238
239/**
240 * filemap_flush - mostly a non-blocking flush
241 * @mapping: target address_space
242 *
243 * This is a mostly non-blocking flush. Not suitable for data-integrity
244 * purposes - I/O may not be started against all dirty pages.
245 */
246int filemap_flush(struct address_space *mapping)
247{
248 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
249}
250EXPORT_SYMBOL(filemap_flush);
251
252/**
253 * filemap_fdatawait_range - wait for writeback to complete
254 * @mapping: address space structure to wait for
255 * @start_byte: offset in bytes where the range starts
256 * @end_byte: offset in bytes where the range ends (inclusive)
257 *
258 * Walk the list of under-writeback pages of the given address space
259 * in the given range and wait for all of them.
260 */
261int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
262 loff_t end_byte)
263{
264 pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
265 pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
266 struct pagevec pvec;
267 int nr_pages;
268 int ret = 0;
269
270 if (end_byte < start_byte)
271 return 0;
272
273 pagevec_init(&pvec, 0);
274 while ((index <= end) &&
275 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
276 PAGECACHE_TAG_WRITEBACK,
277 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
278 unsigned i;
279
280 for (i = 0; i < nr_pages; i++) {
281 struct page *page = pvec.pages[i];
282
283 /* until radix tree lookup accepts end_index */
284 if (page->index > end)
285 continue;
286
287 wait_on_page_writeback(page);
288 if (TestClearPageError(page))
289 ret = -EIO;
290 }
291 pagevec_release(&pvec);
292 cond_resched();
293 }
294
295 /* Check for outstanding write errors */
296 if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297 ret = -ENOSPC;
298 if (test_and_clear_bit(AS_EIO, &mapping->flags))
299 ret = -EIO;
300
301 return ret;
302}
303EXPORT_SYMBOL(filemap_fdatawait_range);
304
305/**
306 * filemap_fdatawait - wait for all under-writeback pages to complete
307 * @mapping: address space structure to wait for
308 *
309 * Walk the list of under-writeback pages of the given address space
310 * and wait for all of them.
311 */
312int filemap_fdatawait(struct address_space *mapping)
313{
314 loff_t i_size = i_size_read(mapping->host);
315
316 if (i_size == 0)
317 return 0;
318
319 return filemap_fdatawait_range(mapping, 0, i_size - 1);
320}
321EXPORT_SYMBOL(filemap_fdatawait);
322
323int filemap_write_and_wait(struct address_space *mapping)
324{
325 int err = 0;
326
327 if (mapping->nrpages) {
328 err = filemap_fdatawrite(mapping);
329 /*
330 * Even if the above returned error, the pages may be
331 * written partially (e.g. -ENOSPC), so we wait for it.
332 * But the -EIO is special case, it may indicate the worst
333 * thing (e.g. bug) happened, so we avoid waiting for it.
334 */
335 if (err != -EIO) {
336 int err2 = filemap_fdatawait(mapping);
337 if (!err)
338 err = err2;
339 }
340 }
341 return err;
342}
343EXPORT_SYMBOL(filemap_write_and_wait);
344
345/**
346 * filemap_write_and_wait_range - write out & wait on a file range
347 * @mapping: the address_space for the pages
348 * @lstart: offset in bytes where the range starts
349 * @lend: offset in bytes where the range ends (inclusive)
350 *
351 * Write out and wait upon file offsets lstart->lend, inclusive.
352 *
353 * Note that `lend' is inclusive (describes the last byte to be written) so
354 * that this function can be used to write to the very end-of-file (end = -1).
355 */
356int filemap_write_and_wait_range(struct address_space *mapping,
357 loff_t lstart, loff_t lend)
358{
359 int err = 0;
360
361 if (mapping->nrpages) {
362 err = __filemap_fdatawrite_range(mapping, lstart, lend,
363 WB_SYNC_ALL);
364 /* See comment of filemap_write_and_wait() */
365 if (err != -EIO) {
366 int err2 = filemap_fdatawait_range(mapping,
367 lstart, lend);
368 if (!err)
369 err = err2;
370 }
371 }
372 return err;
373}
374EXPORT_SYMBOL(filemap_write_and_wait_range);
375
376/**
377 * replace_page_cache_page - replace a pagecache page with a new one
378 * @old: page to be replaced
379 * @new: page to replace with
380 * @gfp_mask: allocation mode
381 *
382 * This function replaces a page in the pagecache with a new one. On
383 * success it acquires the pagecache reference for the new page and
384 * drops it for the old page. Both the old and new pages must be
385 * locked. This function does not add the new page to the LRU, the
386 * caller must do that.
387 *
388 * The remove + add is atomic. The only way this function can fail is
389 * memory allocation failure.
390 */
391int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
392{
393 int error;
394
395 VM_BUG_ON(!PageLocked(old));
396 VM_BUG_ON(!PageLocked(new));
397 VM_BUG_ON(new->mapping);
398
399 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
400 if (!error) {
401 struct address_space *mapping = old->mapping;
402 void (*freepage)(struct page *);
403
404 pgoff_t offset = old->index;
405 freepage = mapping->a_ops->freepage;
406
407 page_cache_get(new);
408 new->mapping = mapping;
409 new->index = offset;
410
411 spin_lock_irq(&mapping->tree_lock);
412 __delete_from_page_cache(old);
413 error = radix_tree_insert(&mapping->page_tree, offset, new);
414 BUG_ON(error);
415 mapping->nrpages++;
416 __inc_zone_page_state(new, NR_FILE_PAGES);
417 if (PageSwapBacked(new))
418 __inc_zone_page_state(new, NR_SHMEM);
419 spin_unlock_irq(&mapping->tree_lock);
420 /* mem_cgroup codes must not be called under tree_lock */
421 mem_cgroup_replace_page_cache(old, new);
422 radix_tree_preload_end();
423 if (freepage)
424 freepage(old);
425 page_cache_release(old);
426 }
427
428 return error;
429}
430EXPORT_SYMBOL_GPL(replace_page_cache_page);
431
432/**
433 * add_to_page_cache_locked - add a locked page to the pagecache
434 * @page: page to add
435 * @mapping: the page's address_space
436 * @offset: page index
437 * @gfp_mask: page allocation mode
438 *
439 * This function is used to add a page to the pagecache. It must be locked.
440 * This function does not add the page to the LRU. The caller must do that.
441 */
442int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
443 pgoff_t offset, gfp_t gfp_mask)
444{
445 int error;
446
447 VM_BUG_ON(!PageLocked(page));
448 VM_BUG_ON(PageSwapBacked(page));
449
450 error = mem_cgroup_cache_charge(page, current->mm,
451 gfp_mask & GFP_RECLAIM_MASK);
452 if (error)
453 goto out;
454
455 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
456 if (error == 0) {
457 page_cache_get(page);
458 page->mapping = mapping;
459 page->index = offset;
460
461 spin_lock_irq(&mapping->tree_lock);
462 error = radix_tree_insert(&mapping->page_tree, offset, page);
463 if (likely(!error)) {
464 mapping->nrpages++;
465 __inc_zone_page_state(page, NR_FILE_PAGES);
466 spin_unlock_irq(&mapping->tree_lock);
467 } else {
468 page->mapping = NULL;
469 /* Leave page->index set: truncation relies upon it */
470 spin_unlock_irq(&mapping->tree_lock);
471 mem_cgroup_uncharge_cache_page(page);
472 page_cache_release(page);
473 }
474 radix_tree_preload_end();
475 } else
476 mem_cgroup_uncharge_cache_page(page);
477out:
478 return error;
479}
480EXPORT_SYMBOL(add_to_page_cache_locked);
481
482int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
483 pgoff_t offset, gfp_t gfp_mask)
484{
485 int ret;
486
487 ret = add_to_page_cache(page, mapping, offset, gfp_mask);
488 if (ret == 0)
489 lru_cache_add_file(page);
490 return ret;
491}
492EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
493
494#ifdef CONFIG_NUMA
495struct page *__page_cache_alloc(gfp_t gfp)
496{
497 int n;
498 struct page *page;
499
500 if (cpuset_do_page_mem_spread()) {
501 unsigned int cpuset_mems_cookie;
502 do {
503 cpuset_mems_cookie = get_mems_allowed();
504 n = cpuset_mem_spread_node();
505 page = alloc_pages_exact_node(n, gfp, 0);
506 } while (!put_mems_allowed(cpuset_mems_cookie) && !page);
507
508 return page;
509 }
510 return alloc_pages(gfp, 0);
511}
512EXPORT_SYMBOL(__page_cache_alloc);
513#endif
514
515/*
516 * In order to wait for pages to become available there must be
517 * waitqueues associated with pages. By using a hash table of
518 * waitqueues where the bucket discipline is to maintain all
519 * waiters on the same queue and wake all when any of the pages
520 * become available, and for the woken contexts to check to be
521 * sure the appropriate page became available, this saves space
522 * at a cost of "thundering herd" phenomena during rare hash
523 * collisions.
524 */
525static wait_queue_head_t *page_waitqueue(struct page *page)
526{
527 const struct zone *zone = page_zone(page);
528
529 return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
530}
531
532static inline void wake_up_page(struct page *page, int bit)
533{
534 __wake_up_bit(page_waitqueue(page), &page->flags, bit);
535}
536
537void wait_on_page_bit(struct page *page, int bit_nr)
538{
539 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
540
541 if (test_bit(bit_nr, &page->flags))
542 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
543 TASK_UNINTERRUPTIBLE);
544}
545EXPORT_SYMBOL(wait_on_page_bit);
546
547int wait_on_page_bit_killable(struct page *page, int bit_nr)
548{
549 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
550
551 if (!test_bit(bit_nr, &page->flags))
552 return 0;
553
554 return __wait_on_bit(page_waitqueue(page), &wait,
555 sleep_on_page_killable, TASK_KILLABLE);
556}
557
558/**
559 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
560 * @page: Page defining the wait queue of interest
561 * @waiter: Waiter to add to the queue
562 *
563 * Add an arbitrary @waiter to the wait queue for the nominated @page.
564 */
565void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
566{
567 wait_queue_head_t *q = page_waitqueue(page);
568 unsigned long flags;
569
570 spin_lock_irqsave(&q->lock, flags);
571 __add_wait_queue(q, waiter);
572 spin_unlock_irqrestore(&q->lock, flags);
573}
574EXPORT_SYMBOL_GPL(add_page_wait_queue);
575
576/**
577 * unlock_page - unlock a locked page
578 * @page: the page
579 *
580 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
581 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
582 * mechananism between PageLocked pages and PageWriteback pages is shared.
583 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
584 *
585 * The mb is necessary to enforce ordering between the clear_bit and the read
586 * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
587 */
588void unlock_page(struct page *page)
589{
590 VM_BUG_ON(!PageLocked(page));
591 clear_bit_unlock(PG_locked, &page->flags);
592 smp_mb__after_clear_bit();
593 wake_up_page(page, PG_locked);
594}
595EXPORT_SYMBOL(unlock_page);
596
597/**
598 * end_page_writeback - end writeback against a page
599 * @page: the page
600 */
601void end_page_writeback(struct page *page)
602{
603 if (TestClearPageReclaim(page))
604 rotate_reclaimable_page(page);
605
606 if (!test_clear_page_writeback(page))
607 BUG();
608
609 smp_mb__after_clear_bit();
610 wake_up_page(page, PG_writeback);
611}
612EXPORT_SYMBOL(end_page_writeback);
613
614/**
615 * __lock_page - get a lock on the page, assuming we need to sleep to get it
616 * @page: the page to lock
617 */
618void __lock_page(struct page *page)
619{
620 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
621
622 __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
623 TASK_UNINTERRUPTIBLE);
624}
625EXPORT_SYMBOL(__lock_page);
626
627int __lock_page_killable(struct page *page)
628{
629 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
630
631 return __wait_on_bit_lock(page_waitqueue(page), &wait,
632 sleep_on_page_killable, TASK_KILLABLE);
633}
634EXPORT_SYMBOL_GPL(__lock_page_killable);
635
636int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
637 unsigned int flags)
638{
639 if (flags & FAULT_FLAG_ALLOW_RETRY) {
640 /*
641 * CAUTION! In this case, mmap_sem is not released
642 * even though return 0.
643 */
644 if (flags & FAULT_FLAG_RETRY_NOWAIT)
645 return 0;
646
647 up_read(&mm->mmap_sem);
648 if (flags & FAULT_FLAG_KILLABLE)
649 wait_on_page_locked_killable(page);
650 else
651 wait_on_page_locked(page);
652 return 0;
653 } else {
654 if (flags & FAULT_FLAG_KILLABLE) {
655 int ret;
656
657 ret = __lock_page_killable(page);
658 if (ret) {
659 up_read(&mm->mmap_sem);
660 return 0;
661 }
662 } else
663 __lock_page(page);
664 return 1;
665 }
666}
667
668/**
669 * find_get_page - find and get a page reference
670 * @mapping: the address_space to search
671 * @offset: the page index
672 *
673 * Is there a pagecache struct page at the given (mapping, offset) tuple?
674 * If yes, increment its refcount and return it; if no, return NULL.
675 */
676struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
677{
678 void **pagep;
679 struct page *page;
680
681 rcu_read_lock();
682repeat:
683 page = NULL;
684 pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
685 if (pagep) {
686 page = radix_tree_deref_slot(pagep);
687 if (unlikely(!page))
688 goto out;
689 if (radix_tree_exception(page)) {
690 if (radix_tree_deref_retry(page))
691 goto repeat;
692 /*
693 * Otherwise, shmem/tmpfs must be storing a swap entry
694 * here as an exceptional entry: so return it without
695 * attempting to raise page count.
696 */
697 goto out;
698 }
699 if (!page_cache_get_speculative(page))
700 goto repeat;
701
702 /*
703 * Has the page moved?
704 * This is part of the lockless pagecache protocol. See
705 * include/linux/pagemap.h for details.
706 */
707 if (unlikely(page != *pagep)) {
708 page_cache_release(page);
709 goto repeat;
710 }
711 }
712out:
713 rcu_read_unlock();
714
715 return page;
716}
717EXPORT_SYMBOL(find_get_page);
718
719/**
720 * find_lock_page - locate, pin and lock a pagecache page
721 * @mapping: the address_space to search
722 * @offset: the page index
723 *
724 * Locates the desired pagecache page, locks it, increments its reference
725 * count and returns its address.
726 *
727 * Returns zero if the page was not present. find_lock_page() may sleep.
728 */
729struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
730{
731 struct page *page;
732
733repeat:
734 page = find_get_page(mapping, offset);
735 if (page && !radix_tree_exception(page)) {
736 lock_page(page);
737 /* Has the page been truncated? */
738 if (unlikely(page->mapping != mapping)) {
739 unlock_page(page);
740 page_cache_release(page);
741 goto repeat;
742 }
743 VM_BUG_ON(page->index != offset);
744 }
745 return page;
746}
747EXPORT_SYMBOL(find_lock_page);
748
749/**
750 * find_or_create_page - locate or add a pagecache page
751 * @mapping: the page's address_space
752 * @index: the page's index into the mapping
753 * @gfp_mask: page allocation mode
754 *
755 * Locates a page in the pagecache. If the page is not present, a new page
756 * is allocated using @gfp_mask and is added to the pagecache and to the VM's
757 * LRU list. The returned page is locked and has its reference count
758 * incremented.
759 *
760 * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
761 * allocation!
762 *
763 * find_or_create_page() returns the desired page's address, or zero on
764 * memory exhaustion.
765 */
766struct page *find_or_create_page(struct address_space *mapping,
767 pgoff_t index, gfp_t gfp_mask)
768{
769 struct page *page;
770 int err;
771repeat:
772 page = find_lock_page(mapping, index);
773 if (!page) {
774 page = __page_cache_alloc(gfp_mask);
775 if (!page)
776 return NULL;
777 /*
778 * We want a regular kernel memory (not highmem or DMA etc)
779 * allocation for the radix tree nodes, but we need to honour
780 * the context-specific requirements the caller has asked for.
781 * GFP_RECLAIM_MASK collects those requirements.
782 */
783 err = add_to_page_cache_lru(page, mapping, index,
784 (gfp_mask & GFP_RECLAIM_MASK));
785 if (unlikely(err)) {
786 page_cache_release(page);
787 page = NULL;
788 if (err == -EEXIST)
789 goto repeat;
790 }
791 }
792 return page;
793}
794EXPORT_SYMBOL(find_or_create_page);
795
796/**
797 * find_get_pages - gang pagecache lookup
798 * @mapping: The address_space to search
799 * @start: The starting page index
800 * @nr_pages: The maximum number of pages
801 * @pages: Where the resulting pages are placed
802 *
803 * find_get_pages() will search for and return a group of up to
804 * @nr_pages pages in the mapping. The pages are placed at @pages.
805 * find_get_pages() takes a reference against the returned pages.
806 *
807 * The search returns a group of mapping-contiguous pages with ascending
808 * indexes. There may be holes in the indices due to not-present pages.
809 *
810 * find_get_pages() returns the number of pages which were found.
811 */
812unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
813 unsigned int nr_pages, struct page **pages)
814{
815 struct radix_tree_iter iter;
816 void **slot;
817 unsigned ret = 0;
818
819 if (unlikely(!nr_pages))
820 return 0;
821
822 rcu_read_lock();
823restart:
824 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
825 struct page *page;
826repeat:
827 page = radix_tree_deref_slot(slot);
828 if (unlikely(!page))
829 continue;
830
831 if (radix_tree_exception(page)) {
832 if (radix_tree_deref_retry(page)) {
833 /*
834 * Transient condition which can only trigger
835 * when entry at index 0 moves out of or back
836 * to root: none yet gotten, safe to restart.
837 */
838 WARN_ON(iter.index);
839 goto restart;
840 }
841 /*
842 * Otherwise, shmem/tmpfs must be storing a swap entry
843 * here as an exceptional entry: so skip over it -
844 * we only reach this from invalidate_mapping_pages().
845 */
846 continue;
847 }
848
849 if (!page_cache_get_speculative(page))
850 goto repeat;
851
852 /* Has the page moved? */
853 if (unlikely(page != *slot)) {
854 page_cache_release(page);
855 goto repeat;
856 }
857
858 pages[ret] = page;
859 if (++ret == nr_pages)
860 break;
861 }
862
863 rcu_read_unlock();
864 return ret;
865}
866
867/**
868 * find_get_pages_contig - gang contiguous pagecache lookup
869 * @mapping: The address_space to search
870 * @index: The starting page index
871 * @nr_pages: The maximum number of pages
872 * @pages: Where the resulting pages are placed
873 *
874 * find_get_pages_contig() works exactly like find_get_pages(), except
875 * that the returned number of pages are guaranteed to be contiguous.
876 *
877 * find_get_pages_contig() returns the number of pages which were found.
878 */
879unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
880 unsigned int nr_pages, struct page **pages)
881{
882 struct radix_tree_iter iter;
883 void **slot;
884 unsigned int ret = 0;
885
886 if (unlikely(!nr_pages))
887 return 0;
888
889 rcu_read_lock();
890restart:
891 radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
892 struct page *page;
893repeat:
894 page = radix_tree_deref_slot(slot);
895 /* The hole, there no reason to continue */
896 if (unlikely(!page))
897 break;
898
899 if (radix_tree_exception(page)) {
900 if (radix_tree_deref_retry(page)) {
901 /*
902 * Transient condition which can only trigger
903 * when entry at index 0 moves out of or back
904 * to root: none yet gotten, safe to restart.
905 */
906 goto restart;
907 }
908 /*
909 * Otherwise, shmem/tmpfs must be storing a swap entry
910 * here as an exceptional entry: so stop looking for
911 * contiguous pages.
912 */
913 break;
914 }
915
916 if (!page_cache_get_speculative(page))
917 goto repeat;
918
919 /* Has the page moved? */
920 if (unlikely(page != *slot)) {
921 page_cache_release(page);
922 goto repeat;
923 }
924
925 /*
926 * must check mapping and index after taking the ref.
927 * otherwise we can get both false positives and false
928 * negatives, which is just confusing to the caller.
929 */
930 if (page->mapping == NULL || page->index != iter.index) {
931 page_cache_release(page);
932 break;
933 }
934
935 pages[ret] = page;
936 if (++ret == nr_pages)
937 break;
938 }
939 rcu_read_unlock();
940 return ret;
941}
942EXPORT_SYMBOL(find_get_pages_contig);
943
944/**
945 * find_get_pages_tag - find and return pages that match @tag
946 * @mapping: the address_space to search
947 * @index: the starting page index
948 * @tag: the tag index
949 * @nr_pages: the maximum number of pages
950 * @pages: where the resulting pages are placed
951 *
952 * Like find_get_pages, except we only return pages which are tagged with
953 * @tag. We update @index to index the next page for the traversal.
954 */
955unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
956 int tag, unsigned int nr_pages, struct page **pages)
957{
958 struct radix_tree_iter iter;
959 void **slot;
960 unsigned ret = 0;
961
962 if (unlikely(!nr_pages))
963 return 0;
964
965 rcu_read_lock();
966restart:
967 radix_tree_for_each_tagged(slot, &mapping->page_tree,
968 &iter, *index, tag) {
969 struct page *page;
970repeat:
971 page = radix_tree_deref_slot(slot);
972 if (unlikely(!page))
973 continue;
974
975 if (radix_tree_exception(page)) {
976 if (radix_tree_deref_retry(page)) {
977 /*
978 * Transient condition which can only trigger
979 * when entry at index 0 moves out of or back
980 * to root: none yet gotten, safe to restart.
981 */
982 goto restart;
983 }
984 /*
985 * This function is never used on a shmem/tmpfs
986 * mapping, so a swap entry won't be found here.
987 */
988 BUG();
989 }
990
991 if (!page_cache_get_speculative(page))
992 goto repeat;
993
994 /* Has the page moved? */
995 if (unlikely(page != *slot)) {
996 page_cache_release(page);
997 goto repeat;
998 }
999
1000 pages[ret] = page;
1001 if (++ret == nr_pages)
1002 break;
1003 }
1004
1005 rcu_read_unlock();
1006
1007 if (ret)
1008 *index = pages[ret - 1]->index + 1;
1009
1010 return ret;
1011}
1012EXPORT_SYMBOL(find_get_pages_tag);
1013
1014/**
1015 * grab_cache_page_nowait - returns locked page at given index in given cache
1016 * @mapping: target address_space
1017 * @index: the page index
1018 *
1019 * Same as grab_cache_page(), but do not wait if the page is unavailable.
1020 * This is intended for speculative data generators, where the data can
1021 * be regenerated if the page couldn't be grabbed. This routine should
1022 * be safe to call while holding the lock for another page.
1023 *
1024 * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1025 * and deadlock against the caller's locked page.
1026 */
1027struct page *
1028grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1029{
1030 struct page *page = find_get_page(mapping, index);
1031
1032 if (page) {
1033 if (trylock_page(page))
1034 return page;
1035 page_cache_release(page);
1036 return NULL;
1037 }
1038 page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1039 if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1040 page_cache_release(page);
1041 page = NULL;
1042 }
1043 return page;
1044}
1045EXPORT_SYMBOL(grab_cache_page_nowait);
1046
1047/*
1048 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1049 * a _large_ part of the i/o request. Imagine the worst scenario:
1050 *
1051 * ---R__________________________________________B__________
1052 * ^ reading here ^ bad block(assume 4k)
1053 *
1054 * read(R) => miss => readahead(R...B) => media error => frustrating retries
1055 * => failing the whole request => read(R) => read(R+1) =>
1056 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1057 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1058 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1059 *
1060 * It is going insane. Fix it by quickly scaling down the readahead size.
1061 */
1062static void shrink_readahead_size_eio(struct file *filp,
1063 struct file_ra_state *ra)
1064{
1065 ra->ra_pages /= 4;
1066}
1067
1068/**
1069 * do_generic_file_read - generic file read routine
1070 * @filp: the file to read
1071 * @ppos: current file position
1072 * @desc: read_descriptor
1073 * @actor: read method
1074 *
1075 * This is a generic file read routine, and uses the
1076 * mapping->a_ops->readpage() function for the actual low-level stuff.
1077 *
1078 * This is really ugly. But the goto's actually try to clarify some
1079 * of the logic when it comes to error handling etc.
1080 */
1081static void do_generic_file_read(struct file *filp, loff_t *ppos,
1082 read_descriptor_t *desc, read_actor_t actor)
1083{
1084 struct address_space *mapping = filp->f_mapping;
1085 struct inode *inode = mapping->host;
1086 struct file_ra_state *ra = &filp->f_ra;
1087 pgoff_t index;
1088 pgoff_t last_index;
1089 pgoff_t prev_index;
1090 unsigned long offset; /* offset into pagecache page */
1091 unsigned int prev_offset;
1092 int error;
1093
1094 index = *ppos >> PAGE_CACHE_SHIFT;
1095 prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1096 prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1097 last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1098 offset = *ppos & ~PAGE_CACHE_MASK;
1099
1100 for (;;) {
1101 struct page *page;
1102 pgoff_t end_index;
1103 loff_t isize;
1104 unsigned long nr, ret;
1105
1106 cond_resched();
1107find_page:
1108 page = find_get_page(mapping, index);
1109 if (!page) {
1110 page_cache_sync_readahead(mapping,
1111 ra, filp,
1112 index, last_index - index);
1113 page = find_get_page(mapping, index);
1114 if (unlikely(page == NULL))
1115 goto no_cached_page;
1116 }
1117 if (PageReadahead(page)) {
1118 page_cache_async_readahead(mapping,
1119 ra, filp, page,
1120 index, last_index - index);
1121 }
1122 if (!PageUptodate(page)) {
1123 if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1124 !mapping->a_ops->is_partially_uptodate)
1125 goto page_not_up_to_date;
1126 if (!trylock_page(page))
1127 goto page_not_up_to_date;
1128 /* Did it get truncated before we got the lock? */
1129 if (!page->mapping)
1130 goto page_not_up_to_date_locked;
1131 if (!mapping->a_ops->is_partially_uptodate(page,
1132 desc, offset))
1133 goto page_not_up_to_date_locked;
1134 unlock_page(page);
1135 }
1136page_ok:
1137 /*
1138 * i_size must be checked after we know the page is Uptodate.
1139 *
1140 * Checking i_size after the check allows us to calculate
1141 * the correct value for "nr", which means the zero-filled
1142 * part of the page is not copied back to userspace (unless
1143 * another truncate extends the file - this is desired though).
1144 */
1145
1146 isize = i_size_read(inode);
1147 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1148 if (unlikely(!isize || index > end_index)) {
1149 page_cache_release(page);
1150 goto out;
1151 }
1152
1153 /* nr is the maximum number of bytes to copy from this page */
1154 nr = PAGE_CACHE_SIZE;
1155 if (index == end_index) {
1156 nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1157 if (nr <= offset) {
1158 page_cache_release(page);
1159 goto out;
1160 }
1161 }
1162 nr = nr - offset;
1163
1164 /* If users can be writing to this page using arbitrary
1165 * virtual addresses, take care about potential aliasing
1166 * before reading the page on the kernel side.
1167 */
1168 if (mapping_writably_mapped(mapping))
1169 flush_dcache_page(page);
1170
1171 /*
1172 * When a sequential read accesses a page several times,
1173 * only mark it as accessed the first time.
1174 */
1175 if (prev_index != index || offset != prev_offset)
1176 mark_page_accessed(page);
1177 prev_index = index;
1178
1179 /*
1180 * Ok, we have the page, and it's up-to-date, so
1181 * now we can copy it to user space...
1182 *
1183 * The actor routine returns how many bytes were actually used..
1184 * NOTE! This may not be the same as how much of a user buffer
1185 * we filled up (we may be padding etc), so we can only update
1186 * "pos" here (the actor routine has to update the user buffer
1187 * pointers and the remaining count).
1188 */
1189 ret = actor(desc, page, offset, nr);
1190 offset += ret;
1191 index += offset >> PAGE_CACHE_SHIFT;
1192 offset &= ~PAGE_CACHE_MASK;
1193 prev_offset = offset;
1194
1195 page_cache_release(page);
1196 if (ret == nr && desc->count)
1197 continue;
1198 goto out;
1199
1200page_not_up_to_date:
1201 /* Get exclusive access to the page ... */
1202 error = lock_page_killable(page);
1203 if (unlikely(error))
1204 goto readpage_error;
1205
1206page_not_up_to_date_locked:
1207 /* Did it get truncated before we got the lock? */
1208 if (!page->mapping) {
1209 unlock_page(page);
1210 page_cache_release(page);
1211 continue;
1212 }
1213
1214 /* Did somebody else fill it already? */
1215 if (PageUptodate(page)) {
1216 unlock_page(page);
1217 goto page_ok;
1218 }
1219
1220readpage:
1221 /*
1222 * A previous I/O error may have been due to temporary
1223 * failures, eg. multipath errors.
1224 * PG_error will be set again if readpage fails.
1225 */
1226 ClearPageError(page);
1227 /* Start the actual read. The read will unlock the page. */
1228 error = mapping->a_ops->readpage(filp, page);
1229
1230 if (unlikely(error)) {
1231 if (error == AOP_TRUNCATED_PAGE) {
1232 page_cache_release(page);
1233 goto find_page;
1234 }
1235 goto readpage_error;
1236 }
1237
1238 if (!PageUptodate(page)) {
1239 error = lock_page_killable(page);
1240 if (unlikely(error))
1241 goto readpage_error;
1242 if (!PageUptodate(page)) {
1243 if (page->mapping == NULL) {
1244 /*
1245 * invalidate_mapping_pages got it
1246 */
1247 unlock_page(page);
1248 page_cache_release(page);
1249 goto find_page;
1250 }
1251 unlock_page(page);
1252 shrink_readahead_size_eio(filp, ra);
1253 error = -EIO;
1254 goto readpage_error;
1255 }
1256 unlock_page(page);
1257 }
1258
1259 goto page_ok;
1260
1261readpage_error:
1262 /* UHHUH! A synchronous read error occurred. Report it */
1263 desc->error = error;
1264 page_cache_release(page);
1265 goto out;
1266
1267no_cached_page:
1268 /*
1269 * Ok, it wasn't cached, so we need to create a new
1270 * page..
1271 */
1272 page = page_cache_alloc_cold(mapping);
1273 if (!page) {
1274 desc->error = -ENOMEM;
1275 goto out;
1276 }
1277 error = add_to_page_cache_lru(page, mapping,
1278 index, GFP_KERNEL);
1279 if (error) {
1280 page_cache_release(page);
1281 if (error == -EEXIST)
1282 goto find_page;
1283 desc->error = error;
1284 goto out;
1285 }
1286 goto readpage;
1287 }
1288
1289out:
1290 ra->prev_pos = prev_index;
1291 ra->prev_pos <<= PAGE_CACHE_SHIFT;
1292 ra->prev_pos |= prev_offset;
1293
1294 *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1295 file_accessed(filp);
1296}
1297
1298int file_read_actor(read_descriptor_t *desc, struct page *page,
1299 unsigned long offset, unsigned long size)
1300{
1301 char *kaddr;
1302 unsigned long left, count = desc->count;
1303
1304 if (size > count)
1305 size = count;
1306
1307 /*
1308 * Faults on the destination of a read are common, so do it before
1309 * taking the kmap.
1310 */
1311 if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1312 kaddr = kmap_atomic(page);
1313 left = __copy_to_user_inatomic(desc->arg.buf,
1314 kaddr + offset, size);
1315 kunmap_atomic(kaddr);
1316 if (left == 0)
1317 goto success;
1318 }
1319
1320 /* Do it the slow way */
1321 kaddr = kmap(page);
1322 left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1323 kunmap(page);
1324
1325 if (left) {
1326 size -= left;
1327 desc->error = -EFAULT;
1328 }
1329success:
1330 desc->count = count - size;
1331 desc->written += size;
1332 desc->arg.buf += size;
1333 return size;
1334}
1335
1336/*
1337 * Performs necessary checks before doing a write
1338 * @iov: io vector request
1339 * @nr_segs: number of segments in the iovec
1340 * @count: number of bytes to write
1341 * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1342 *
1343 * Adjust number of segments and amount of bytes to write (nr_segs should be
1344 * properly initialized first). Returns appropriate error code that caller
1345 * should return or zero in case that write should be allowed.
1346 */
1347int generic_segment_checks(const struct iovec *iov,
1348 unsigned long *nr_segs, size_t *count, int access_flags)
1349{
1350 unsigned long seg;
1351 size_t cnt = 0;
1352 for (seg = 0; seg < *nr_segs; seg++) {
1353 const struct iovec *iv = &iov[seg];
1354
1355 /*
1356 * If any segment has a negative length, or the cumulative
1357 * length ever wraps negative then return -EINVAL.
1358 */
1359 cnt += iv->iov_len;
1360 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1361 return -EINVAL;
1362 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1363 continue;
1364 if (seg == 0)
1365 return -EFAULT;
1366 *nr_segs = seg;
1367 cnt -= iv->iov_len; /* This segment is no good */
1368 break;
1369 }
1370 *count = cnt;
1371 return 0;
1372}
1373EXPORT_SYMBOL(generic_segment_checks);
1374
1375/**
1376 * generic_file_aio_read - generic filesystem read routine
1377 * @iocb: kernel I/O control block
1378 * @iov: io vector request
1379 * @nr_segs: number of segments in the iovec
1380 * @pos: current file position
1381 *
1382 * This is the "read()" routine for all filesystems
1383 * that can use the page cache directly.
1384 */
1385ssize_t
1386generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1387 unsigned long nr_segs, loff_t pos)
1388{
1389 struct file *filp = iocb->ki_filp;
1390 ssize_t retval;
1391 unsigned long seg = 0;
1392 size_t count;
1393 loff_t *ppos = &iocb->ki_pos;
1394
1395 count = 0;
1396 retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1397 if (retval)
1398 return retval;
1399
1400 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1401 if (filp->f_flags & O_DIRECT) {
1402 loff_t size;
1403 struct address_space *mapping;
1404 struct inode *inode;
1405
1406 mapping = filp->f_mapping;
1407 inode = mapping->host;
1408 if (!count)
1409 goto out; /* skip atime */
1410 size = i_size_read(inode);
1411 if (pos < size) {
1412 retval = filemap_write_and_wait_range(mapping, pos,
1413 pos + iov_length(iov, nr_segs) - 1);
1414 if (!retval) {
1415 struct blk_plug plug;
1416
1417 blk_start_plug(&plug);
1418 retval = mapping->a_ops->direct_IO(READ, iocb,
1419 iov, pos, nr_segs);
1420 blk_finish_plug(&plug);
1421 }
1422 if (retval > 0) {
1423 *ppos = pos + retval;
1424 count -= retval;
1425 }
1426
1427 /*
1428 * Btrfs can have a short DIO read if we encounter
1429 * compressed extents, so if there was an error, or if
1430 * we've already read everything we wanted to, or if
1431 * there was a short read because we hit EOF, go ahead
1432 * and return. Otherwise fallthrough to buffered io for
1433 * the rest of the read.
1434 */
1435 if (retval < 0 || !count || *ppos >= size) {
1436 file_accessed(filp);
1437 goto out;
1438 }
1439 }
1440 }
1441
1442 count = retval;
1443 for (seg = 0; seg < nr_segs; seg++) {
1444 read_descriptor_t desc;
1445 loff_t offset = 0;
1446
1447 /*
1448 * If we did a short DIO read we need to skip the section of the
1449 * iov that we've already read data into.
1450 */
1451 if (count) {
1452 if (count > iov[seg].iov_len) {
1453 count -= iov[seg].iov_len;
1454 continue;
1455 }
1456 offset = count;
1457 count = 0;
1458 }
1459
1460 desc.written = 0;
1461 desc.arg.buf = iov[seg].iov_base + offset;
1462 desc.count = iov[seg].iov_len - offset;
1463 if (desc.count == 0)
1464 continue;
1465 desc.error = 0;
1466 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1467 retval += desc.written;
1468 if (desc.error) {
1469 retval = retval ?: desc.error;
1470 break;
1471 }
1472 if (desc.count > 0)
1473 break;
1474 }
1475out:
1476 return retval;
1477}
1478EXPORT_SYMBOL(generic_file_aio_read);
1479
1480#ifdef CONFIG_MMU
1481/**
1482 * page_cache_read - adds requested page to the page cache if not already there
1483 * @file: file to read
1484 * @offset: page index
1485 *
1486 * This adds the requested page to the page cache if it isn't already there,
1487 * and schedules an I/O to read in its contents from disk.
1488 */
1489static int page_cache_read(struct file *file, pgoff_t offset)
1490{
1491 struct address_space *mapping = file->f_mapping;
1492 struct page *page;
1493 int ret;
1494
1495 do {
1496 page = page_cache_alloc_cold(mapping);
1497 if (!page)
1498 return -ENOMEM;
1499
1500 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1501 if (ret == 0)
1502 ret = mapping->a_ops->readpage(file, page);
1503 else if (ret == -EEXIST)
1504 ret = 0; /* losing race to add is OK */
1505
1506 page_cache_release(page);
1507
1508 } while (ret == AOP_TRUNCATED_PAGE);
1509
1510 return ret;
1511}
1512
1513#define MMAP_LOTSAMISS (100)
1514
1515/*
1516 * Synchronous readahead happens when we don't even find
1517 * a page in the page cache at all.
1518 */
1519static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1520 struct file_ra_state *ra,
1521 struct file *file,
1522 pgoff_t offset)
1523{
1524 unsigned long ra_pages;
1525 struct address_space *mapping = file->f_mapping;
1526
1527 /* If we don't want any read-ahead, don't bother */
1528 if (VM_RandomReadHint(vma))
1529 return;
1530 if (!ra->ra_pages)
1531 return;
1532
1533 if (VM_SequentialReadHint(vma)) {
1534 page_cache_sync_readahead(mapping, ra, file, offset,
1535 ra->ra_pages);
1536 return;
1537 }
1538
1539 /* Avoid banging the cache line if not needed */
1540 if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1541 ra->mmap_miss++;
1542
1543 /*
1544 * Do we miss much more than hit in this file? If so,
1545 * stop bothering with read-ahead. It will only hurt.
1546 */
1547 if (ra->mmap_miss > MMAP_LOTSAMISS)
1548 return;
1549
1550 /*
1551 * mmap read-around
1552 */
1553 ra_pages = max_sane_readahead(ra->ra_pages);
1554 ra->start = max_t(long, 0, offset - ra_pages / 2);
1555 ra->size = ra_pages;
1556 ra->async_size = ra_pages / 4;
1557 ra_submit(ra, mapping, file);
1558}
1559
1560/*
1561 * Asynchronous readahead happens when we find the page and PG_readahead,
1562 * so we want to possibly extend the readahead further..
1563 */
1564static void do_async_mmap_readahead(struct vm_area_struct *vma,
1565 struct file_ra_state *ra,
1566 struct file *file,
1567 struct page *page,
1568 pgoff_t offset)
1569{
1570 struct address_space *mapping = file->f_mapping;
1571
1572 /* If we don't want any read-ahead, don't bother */
1573 if (VM_RandomReadHint(vma))
1574 return;
1575 if (ra->mmap_miss > 0)
1576 ra->mmap_miss--;
1577 if (PageReadahead(page))
1578 page_cache_async_readahead(mapping, ra, file,
1579 page, offset, ra->ra_pages);
1580}
1581
1582/**
1583 * filemap_fault - read in file data for page fault handling
1584 * @vma: vma in which the fault was taken
1585 * @vmf: struct vm_fault containing details of the fault
1586 *
1587 * filemap_fault() is invoked via the vma operations vector for a
1588 * mapped memory region to read in file data during a page fault.
1589 *
1590 * The goto's are kind of ugly, but this streamlines the normal case of having
1591 * it in the page cache, and handles the special cases reasonably without
1592 * having a lot of duplicated code.
1593 */
1594int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1595{
1596 int error;
1597 struct file *file = vma->vm_file;
1598 struct address_space *mapping = file->f_mapping;
1599 struct file_ra_state *ra = &file->f_ra;
1600 struct inode *inode = mapping->host;
1601 pgoff_t offset = vmf->pgoff;
1602 struct page *page;
1603 pgoff_t size;
1604 int ret = 0;
1605
1606 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1607 if (offset >= size)
1608 return VM_FAULT_SIGBUS;
1609
1610 /*
1611 * Do we have something in the page cache already?
1612 */
1613 page = find_get_page(mapping, offset);
1614 if (likely(page)) {
1615 /*
1616 * We found the page, so try async readahead before
1617 * waiting for the lock.
1618 */
1619 do_async_mmap_readahead(vma, ra, file, page, offset);
1620 } else {
1621 /* No page in the page cache at all */
1622 do_sync_mmap_readahead(vma, ra, file, offset);
1623 count_vm_event(PGMAJFAULT);
1624 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1625 ret = VM_FAULT_MAJOR;
1626retry_find:
1627 page = find_get_page(mapping, offset);
1628 if (!page)
1629 goto no_cached_page;
1630 }
1631
1632 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1633 page_cache_release(page);
1634 return ret | VM_FAULT_RETRY;
1635 }
1636
1637 /* Did it get truncated? */
1638 if (unlikely(page->mapping != mapping)) {
1639 unlock_page(page);
1640 put_page(page);
1641 goto retry_find;
1642 }
1643 VM_BUG_ON(page->index != offset);
1644
1645 /*
1646 * We have a locked page in the page cache, now we need to check
1647 * that it's up-to-date. If not, it is going to be due to an error.
1648 */
1649 if (unlikely(!PageUptodate(page)))
1650 goto page_not_uptodate;
1651
1652 /*
1653 * Found the page and have a reference on it.
1654 * We must recheck i_size under page lock.
1655 */
1656 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1657 if (unlikely(offset >= size)) {
1658 unlock_page(page);
1659 page_cache_release(page);
1660 return VM_FAULT_SIGBUS;
1661 }
1662
1663 vmf->page = page;
1664 return ret | VM_FAULT_LOCKED;
1665
1666no_cached_page:
1667 /*
1668 * We're only likely to ever get here if MADV_RANDOM is in
1669 * effect.
1670 */
1671 error = page_cache_read(file, offset);
1672
1673 /*
1674 * The page we want has now been added to the page cache.
1675 * In the unlikely event that someone removed it in the
1676 * meantime, we'll just come back here and read it again.
1677 */
1678 if (error >= 0)
1679 goto retry_find;
1680
1681 /*
1682 * An error return from page_cache_read can result if the
1683 * system is low on memory, or a problem occurs while trying
1684 * to schedule I/O.
1685 */
1686 if (error == -ENOMEM)
1687 return VM_FAULT_OOM;
1688 return VM_FAULT_SIGBUS;
1689
1690page_not_uptodate:
1691 /*
1692 * Umm, take care of errors if the page isn't up-to-date.
1693 * Try to re-read it _once_. We do this synchronously,
1694 * because there really aren't any performance issues here
1695 * and we need to check for errors.
1696 */
1697 ClearPageError(page);
1698 error = mapping->a_ops->readpage(file, page);
1699 if (!error) {
1700 wait_on_page_locked(page);
1701 if (!PageUptodate(page))
1702 error = -EIO;
1703 }
1704 page_cache_release(page);
1705
1706 if (!error || error == AOP_TRUNCATED_PAGE)
1707 goto retry_find;
1708
1709 /* Things didn't work out. Return zero to tell the mm layer so. */
1710 shrink_readahead_size_eio(file, ra);
1711 return VM_FAULT_SIGBUS;
1712}
1713EXPORT_SYMBOL(filemap_fault);
1714
1715const struct vm_operations_struct generic_file_vm_ops = {
1716 .fault = filemap_fault,
1717};
1718
1719/* This is used for a general mmap of a disk file */
1720
1721int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1722{
1723 struct address_space *mapping = file->f_mapping;
1724
1725 if (!mapping->a_ops->readpage)
1726 return -ENOEXEC;
1727 file_accessed(file);
1728 vma->vm_ops = &generic_file_vm_ops;
1729 vma->vm_flags |= VM_CAN_NONLINEAR;
1730 return 0;
1731}
1732
1733/*
1734 * This is for filesystems which do not implement ->writepage.
1735 */
1736int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1737{
1738 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1739 return -EINVAL;
1740 return generic_file_mmap(file, vma);
1741}
1742#else
1743int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1744{
1745 return -ENOSYS;
1746}
1747int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1748{
1749 return -ENOSYS;
1750}
1751#endif /* CONFIG_MMU */
1752
1753EXPORT_SYMBOL(generic_file_mmap);
1754EXPORT_SYMBOL(generic_file_readonly_mmap);
1755
1756static struct page *__read_cache_page(struct address_space *mapping,
1757 pgoff_t index,
1758 int (*filler)(void *, struct page *),
1759 void *data,
1760 gfp_t gfp)
1761{
1762 struct page *page;
1763 int err;
1764repeat:
1765 page = find_get_page(mapping, index);
1766 if (!page) {
1767 page = __page_cache_alloc(gfp | __GFP_COLD);
1768 if (!page)
1769 return ERR_PTR(-ENOMEM);
1770 err = add_to_page_cache_lru(page, mapping, index, gfp);
1771 if (unlikely(err)) {
1772 page_cache_release(page);
1773 if (err == -EEXIST)
1774 goto repeat;
1775 /* Presumably ENOMEM for radix tree node */
1776 return ERR_PTR(err);
1777 }
1778 err = filler(data, page);
1779 if (err < 0) {
1780 page_cache_release(page);
1781 page = ERR_PTR(err);
1782 }
1783 }
1784 return page;
1785}
1786
1787static struct page *do_read_cache_page(struct address_space *mapping,
1788 pgoff_t index,
1789 int (*filler)(void *, struct page *),
1790 void *data,
1791 gfp_t gfp)
1792
1793{
1794 struct page *page;
1795 int err;
1796
1797retry:
1798 page = __read_cache_page(mapping, index, filler, data, gfp);
1799 if (IS_ERR(page))
1800 return page;
1801 if (PageUptodate(page))
1802 goto out;
1803
1804 lock_page(page);
1805 if (!page->mapping) {
1806 unlock_page(page);
1807 page_cache_release(page);
1808 goto retry;
1809 }
1810 if (PageUptodate(page)) {
1811 unlock_page(page);
1812 goto out;
1813 }
1814 err = filler(data, page);
1815 if (err < 0) {
1816 page_cache_release(page);
1817 return ERR_PTR(err);
1818 }
1819out:
1820 mark_page_accessed(page);
1821 return page;
1822}
1823
1824/**
1825 * read_cache_page_async - read into page cache, fill it if needed
1826 * @mapping: the page's address_space
1827 * @index: the page index
1828 * @filler: function to perform the read
1829 * @data: first arg to filler(data, page) function, often left as NULL
1830 *
1831 * Same as read_cache_page, but don't wait for page to become unlocked
1832 * after submitting it to the filler.
1833 *
1834 * Read into the page cache. If a page already exists, and PageUptodate() is
1835 * not set, try to fill the page but don't wait for it to become unlocked.
1836 *
1837 * If the page does not get brought uptodate, return -EIO.
1838 */
1839struct page *read_cache_page_async(struct address_space *mapping,
1840 pgoff_t index,
1841 int (*filler)(void *, struct page *),
1842 void *data)
1843{
1844 return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1845}
1846EXPORT_SYMBOL(read_cache_page_async);
1847
1848static struct page *wait_on_page_read(struct page *page)
1849{
1850 if (!IS_ERR(page)) {
1851 wait_on_page_locked(page);
1852 if (!PageUptodate(page)) {
1853 page_cache_release(page);
1854 page = ERR_PTR(-EIO);
1855 }
1856 }
1857 return page;
1858}
1859
1860/**
1861 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1862 * @mapping: the page's address_space
1863 * @index: the page index
1864 * @gfp: the page allocator flags to use if allocating
1865 *
1866 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1867 * any new page allocations done using the specified allocation flags.
1868 *
1869 * If the page does not get brought uptodate, return -EIO.
1870 */
1871struct page *read_cache_page_gfp(struct address_space *mapping,
1872 pgoff_t index,
1873 gfp_t gfp)
1874{
1875 filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1876
1877 return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1878}
1879EXPORT_SYMBOL(read_cache_page_gfp);
1880
1881/**
1882 * read_cache_page - read into page cache, fill it if needed
1883 * @mapping: the page's address_space
1884 * @index: the page index
1885 * @filler: function to perform the read
1886 * @data: first arg to filler(data, page) function, often left as NULL
1887 *
1888 * Read into the page cache. If a page already exists, and PageUptodate() is
1889 * not set, try to fill the page then wait for it to become unlocked.
1890 *
1891 * If the page does not get brought uptodate, return -EIO.
1892 */
1893struct page *read_cache_page(struct address_space *mapping,
1894 pgoff_t index,
1895 int (*filler)(void *, struct page *),
1896 void *data)
1897{
1898 return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1899}
1900EXPORT_SYMBOL(read_cache_page);
1901
1902static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1903 const struct iovec *iov, size_t base, size_t bytes)
1904{
1905 size_t copied = 0, left = 0;
1906
1907 while (bytes) {
1908 char __user *buf = iov->iov_base + base;
1909 int copy = min(bytes, iov->iov_len - base);
1910
1911 base = 0;
1912 left = __copy_from_user_inatomic(vaddr, buf, copy);
1913 copied += copy;
1914 bytes -= copy;
1915 vaddr += copy;
1916 iov++;
1917
1918 if (unlikely(left))
1919 break;
1920 }
1921 return copied - left;
1922}
1923
1924/*
1925 * Copy as much as we can into the page and return the number of bytes which
1926 * were successfully copied. If a fault is encountered then return the number of
1927 * bytes which were copied.
1928 */
1929size_t iov_iter_copy_from_user_atomic(struct page *page,
1930 struct iov_iter *i, unsigned long offset, size_t bytes)
1931{
1932 char *kaddr;
1933 size_t copied;
1934
1935 BUG_ON(!in_atomic());
1936 kaddr = kmap_atomic(page);
1937 if (likely(i->nr_segs == 1)) {
1938 int left;
1939 char __user *buf = i->iov->iov_base + i->iov_offset;
1940 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1941 copied = bytes - left;
1942 } else {
1943 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1944 i->iov, i->iov_offset, bytes);
1945 }
1946 kunmap_atomic(kaddr);
1947
1948 return copied;
1949}
1950EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1951
1952/*
1953 * This has the same sideeffects and return value as
1954 * iov_iter_copy_from_user_atomic().
1955 * The difference is that it attempts to resolve faults.
1956 * Page must not be locked.
1957 */
1958size_t iov_iter_copy_from_user(struct page *page,
1959 struct iov_iter *i, unsigned long offset, size_t bytes)
1960{
1961 char *kaddr;
1962 size_t copied;
1963
1964 kaddr = kmap(page);
1965 if (likely(i->nr_segs == 1)) {
1966 int left;
1967 char __user *buf = i->iov->iov_base + i->iov_offset;
1968 left = __copy_from_user(kaddr + offset, buf, bytes);
1969 copied = bytes - left;
1970 } else {
1971 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1972 i->iov, i->iov_offset, bytes);
1973 }
1974 kunmap(page);
1975 return copied;
1976}
1977EXPORT_SYMBOL(iov_iter_copy_from_user);
1978
1979void iov_iter_advance(struct iov_iter *i, size_t bytes)
1980{
1981 BUG_ON(i->count < bytes);
1982
1983 if (likely(i->nr_segs == 1)) {
1984 i->iov_offset += bytes;
1985 i->count -= bytes;
1986 } else {
1987 const struct iovec *iov = i->iov;
1988 size_t base = i->iov_offset;
1989 unsigned long nr_segs = i->nr_segs;
1990
1991 /*
1992 * The !iov->iov_len check ensures we skip over unlikely
1993 * zero-length segments (without overruning the iovec).
1994 */
1995 while (bytes || unlikely(i->count && !iov->iov_len)) {
1996 int copy;
1997
1998 copy = min(bytes, iov->iov_len - base);
1999 BUG_ON(!i->count || i->count < copy);
2000 i->count -= copy;
2001 bytes -= copy;
2002 base += copy;
2003 if (iov->iov_len == base) {
2004 iov++;
2005 nr_segs--;
2006 base = 0;
2007 }
2008 }
2009 i->iov = iov;
2010 i->iov_offset = base;
2011 i->nr_segs = nr_segs;
2012 }
2013}
2014EXPORT_SYMBOL(iov_iter_advance);
2015
2016/*
2017 * Fault in the first iovec of the given iov_iter, to a maximum length
2018 * of bytes. Returns 0 on success, or non-zero if the memory could not be
2019 * accessed (ie. because it is an invalid address).
2020 *
2021 * writev-intensive code may want this to prefault several iovecs -- that
2022 * would be possible (callers must not rely on the fact that _only_ the
2023 * first iovec will be faulted with the current implementation).
2024 */
2025int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2026{
2027 char __user *buf = i->iov->iov_base + i->iov_offset;
2028 bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2029 return fault_in_pages_readable(buf, bytes);
2030}
2031EXPORT_SYMBOL(iov_iter_fault_in_readable);
2032
2033/*
2034 * Return the count of just the current iov_iter segment.
2035 */
2036size_t iov_iter_single_seg_count(struct iov_iter *i)
2037{
2038 const struct iovec *iov = i->iov;
2039 if (i->nr_segs == 1)
2040 return i->count;
2041 else
2042 return min(i->count, iov->iov_len - i->iov_offset);
2043}
2044EXPORT_SYMBOL(iov_iter_single_seg_count);
2045
2046/*
2047 * Performs necessary checks before doing a write
2048 *
2049 * Can adjust writing position or amount of bytes to write.
2050 * Returns appropriate error code that caller should return or
2051 * zero in case that write should be allowed.
2052 */
2053inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2054{
2055 struct inode *inode = file->f_mapping->host;
2056 unsigned long limit = rlimit(RLIMIT_FSIZE);
2057
2058 if (unlikely(*pos < 0))
2059 return -EINVAL;
2060
2061 if (!isblk) {
2062 /* FIXME: this is for backwards compatibility with 2.4 */
2063 if (file->f_flags & O_APPEND)
2064 *pos = i_size_read(inode);
2065
2066 if (limit != RLIM_INFINITY) {
2067 if (*pos >= limit) {
2068 send_sig(SIGXFSZ, current, 0);
2069 return -EFBIG;
2070 }
2071 if (*count > limit - (typeof(limit))*pos) {
2072 *count = limit - (typeof(limit))*pos;
2073 }
2074 }
2075 }
2076
2077 /*
2078 * LFS rule
2079 */
2080 if (unlikely(*pos + *count > MAX_NON_LFS &&
2081 !(file->f_flags & O_LARGEFILE))) {
2082 if (*pos >= MAX_NON_LFS) {
2083 return -EFBIG;
2084 }
2085 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2086 *count = MAX_NON_LFS - (unsigned long)*pos;
2087 }
2088 }
2089
2090 /*
2091 * Are we about to exceed the fs block limit ?
2092 *
2093 * If we have written data it becomes a short write. If we have
2094 * exceeded without writing data we send a signal and return EFBIG.
2095 * Linus frestrict idea will clean these up nicely..
2096 */
2097 if (likely(!isblk)) {
2098 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2099 if (*count || *pos > inode->i_sb->s_maxbytes) {
2100 return -EFBIG;
2101 }
2102 /* zero-length writes at ->s_maxbytes are OK */
2103 }
2104
2105 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2106 *count = inode->i_sb->s_maxbytes - *pos;
2107 } else {
2108#ifdef CONFIG_BLOCK
2109 loff_t isize;
2110 if (bdev_read_only(I_BDEV(inode)))
2111 return -EPERM;
2112 isize = i_size_read(inode);
2113 if (*pos >= isize) {
2114 if (*count || *pos > isize)
2115 return -ENOSPC;
2116 }
2117
2118 if (*pos + *count > isize)
2119 *count = isize - *pos;
2120#else
2121 return -EPERM;
2122#endif
2123 }
2124 return 0;
2125}
2126EXPORT_SYMBOL(generic_write_checks);
2127
2128int pagecache_write_begin(struct file *file, struct address_space *mapping,
2129 loff_t pos, unsigned len, unsigned flags,
2130 struct page **pagep, void **fsdata)
2131{
2132 const struct address_space_operations *aops = mapping->a_ops;
2133
2134 return aops->write_begin(file, mapping, pos, len, flags,
2135 pagep, fsdata);
2136}
2137EXPORT_SYMBOL(pagecache_write_begin);
2138
2139int pagecache_write_end(struct file *file, struct address_space *mapping,
2140 loff_t pos, unsigned len, unsigned copied,
2141 struct page *page, void *fsdata)
2142{
2143 const struct address_space_operations *aops = mapping->a_ops;
2144
2145 mark_page_accessed(page);
2146 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2147}
2148EXPORT_SYMBOL(pagecache_write_end);
2149
2150ssize_t
2151generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2152 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2153 size_t count, size_t ocount)
2154{
2155 struct file *file = iocb->ki_filp;
2156 struct address_space *mapping = file->f_mapping;
2157 struct inode *inode = mapping->host;
2158 ssize_t written;
2159 size_t write_len;
2160 pgoff_t end;
2161
2162 if (count != ocount)
2163 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2164
2165 write_len = iov_length(iov, *nr_segs);
2166 end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2167
2168 written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2169 if (written)
2170 goto out;
2171
2172 /*
2173 * After a write we want buffered reads to be sure to go to disk to get
2174 * the new data. We invalidate clean cached page from the region we're
2175 * about to write. We do this *before* the write so that we can return
2176 * without clobbering -EIOCBQUEUED from ->direct_IO().
2177 */
2178 if (mapping->nrpages) {
2179 written = invalidate_inode_pages2_range(mapping,
2180 pos >> PAGE_CACHE_SHIFT, end);
2181 /*
2182 * If a page can not be invalidated, return 0 to fall back
2183 * to buffered write.
2184 */
2185 if (written) {
2186 if (written == -EBUSY)
2187 return 0;
2188 goto out;
2189 }
2190 }
2191
2192 written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2193
2194 /*
2195 * Finally, try again to invalidate clean pages which might have been
2196 * cached by non-direct readahead, or faulted in by get_user_pages()
2197 * if the source of the write was an mmap'ed region of the file
2198 * we're writing. Either one is a pretty crazy thing to do,
2199 * so we don't support it 100%. If this invalidation
2200 * fails, tough, the write still worked...
2201 */
2202 if (mapping->nrpages) {
2203 invalidate_inode_pages2_range(mapping,
2204 pos >> PAGE_CACHE_SHIFT, end);
2205 }
2206
2207 if (written > 0) {
2208 pos += written;
2209 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2210 i_size_write(inode, pos);
2211 mark_inode_dirty(inode);
2212 }
2213 *ppos = pos;
2214 }
2215out:
2216 return written;
2217}
2218EXPORT_SYMBOL(generic_file_direct_write);
2219
2220/*
2221 * Find or create a page at the given pagecache position. Return the locked
2222 * page. This function is specifically for buffered writes.
2223 */
2224struct page *grab_cache_page_write_begin(struct address_space *mapping,
2225 pgoff_t index, unsigned flags)
2226{
2227 int status;
2228 gfp_t gfp_mask;
2229 struct page *page;
2230 gfp_t gfp_notmask = 0;
2231
2232 gfp_mask = mapping_gfp_mask(mapping);
2233 if (mapping_cap_account_dirty(mapping))
2234 gfp_mask |= __GFP_WRITE;
2235 if (flags & AOP_FLAG_NOFS)
2236 gfp_notmask = __GFP_FS;
2237repeat:
2238 page = find_lock_page(mapping, index);
2239 if (page)
2240 goto found;
2241
2242 page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2243 if (!page)
2244 return NULL;
2245 status = add_to_page_cache_lru(page, mapping, index,
2246 GFP_KERNEL & ~gfp_notmask);
2247 if (unlikely(status)) {
2248 page_cache_release(page);
2249 if (status == -EEXIST)
2250 goto repeat;
2251 return NULL;
2252 }
2253found:
2254 wait_on_page_writeback(page);
2255 return page;
2256}
2257EXPORT_SYMBOL(grab_cache_page_write_begin);
2258
2259static ssize_t generic_perform_write(struct file *file,
2260 struct iov_iter *i, loff_t pos)
2261{
2262 struct address_space *mapping = file->f_mapping;
2263 const struct address_space_operations *a_ops = mapping->a_ops;
2264 long status = 0;
2265 ssize_t written = 0;
2266 unsigned int flags = 0;
2267
2268 /*
2269 * Copies from kernel address space cannot fail (NFSD is a big user).
2270 */
2271 if (segment_eq(get_fs(), KERNEL_DS))
2272 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2273
2274 do {
2275 struct page *page;
2276 unsigned long offset; /* Offset into pagecache page */
2277 unsigned long bytes; /* Bytes to write to page */
2278 size_t copied; /* Bytes copied from user */
2279 void *fsdata;
2280
2281 offset = (pos & (PAGE_CACHE_SIZE - 1));
2282 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2283 iov_iter_count(i));
2284
2285again:
2286 /*
2287 * Bring in the user page that we will copy from _first_.
2288 * Otherwise there's a nasty deadlock on copying from the
2289 * same page as we're writing to, without it being marked
2290 * up-to-date.
2291 *
2292 * Not only is this an optimisation, but it is also required
2293 * to check that the address is actually valid, when atomic
2294 * usercopies are used, below.
2295 */
2296 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2297 status = -EFAULT;
2298 break;
2299 }
2300
2301 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2302 &page, &fsdata);
2303 if (unlikely(status))
2304 break;
2305
2306 if (mapping_writably_mapped(mapping))
2307 flush_dcache_page(page);
2308
2309 pagefault_disable();
2310 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2311 pagefault_enable();
2312 flush_dcache_page(page);
2313
2314 mark_page_accessed(page);
2315 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2316 page, fsdata);
2317 if (unlikely(status < 0))
2318 break;
2319 copied = status;
2320
2321 cond_resched();
2322
2323 iov_iter_advance(i, copied);
2324 if (unlikely(copied == 0)) {
2325 /*
2326 * If we were unable to copy any data at all, we must
2327 * fall back to a single segment length write.
2328 *
2329 * If we didn't fallback here, we could livelock
2330 * because not all segments in the iov can be copied at
2331 * once without a pagefault.
2332 */
2333 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2334 iov_iter_single_seg_count(i));
2335 goto again;
2336 }
2337 pos += copied;
2338 written += copied;
2339
2340 balance_dirty_pages_ratelimited(mapping);
2341 if (fatal_signal_pending(current)) {
2342 status = -EINTR;
2343 break;
2344 }
2345 } while (iov_iter_count(i));
2346
2347 return written ? written : status;
2348}
2349
2350ssize_t
2351generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2352 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2353 size_t count, ssize_t written)
2354{
2355 struct file *file = iocb->ki_filp;
2356 ssize_t status;
2357 struct iov_iter i;
2358
2359 iov_iter_init(&i, iov, nr_segs, count, written);
2360 status = generic_perform_write(file, &i, pos);
2361
2362 if (likely(status >= 0)) {
2363 written += status;
2364 *ppos = pos + status;
2365 }
2366
2367 return written ? written : status;
2368}
2369EXPORT_SYMBOL(generic_file_buffered_write);
2370
2371/**
2372 * __generic_file_aio_write - write data to a file
2373 * @iocb: IO state structure (file, offset, etc.)
2374 * @iov: vector with data to write
2375 * @nr_segs: number of segments in the vector
2376 * @ppos: position where to write
2377 *
2378 * This function does all the work needed for actually writing data to a
2379 * file. It does all basic checks, removes SUID from the file, updates
2380 * modification times and calls proper subroutines depending on whether we
2381 * do direct IO or a standard buffered write.
2382 *
2383 * It expects i_mutex to be grabbed unless we work on a block device or similar
2384 * object which does not need locking at all.
2385 *
2386 * This function does *not* take care of syncing data in case of O_SYNC write.
2387 * A caller has to handle it. This is mainly due to the fact that we want to
2388 * avoid syncing under i_mutex.
2389 */
2390ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2391 unsigned long nr_segs, loff_t *ppos)
2392{
2393 struct file *file = iocb->ki_filp;
2394 struct address_space * mapping = file->f_mapping;
2395 size_t ocount; /* original count */
2396 size_t count; /* after file limit checks */
2397 struct inode *inode = mapping->host;
2398 loff_t pos;
2399 ssize_t written;
2400 ssize_t err;
2401
2402 ocount = 0;
2403 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2404 if (err)
2405 return err;
2406
2407 count = ocount;
2408 pos = *ppos;
2409
2410 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2411
2412 /* We can write back this queue in page reclaim */
2413 current->backing_dev_info = mapping->backing_dev_info;
2414 written = 0;
2415
2416 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2417 if (err)
2418 goto out;
2419
2420 if (count == 0)
2421 goto out;
2422
2423 err = file_remove_suid(file);
2424 if (err)
2425 goto out;
2426
2427 err = file_update_time(file);
2428 if (err)
2429 goto out;
2430
2431 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2432 if (unlikely(file->f_flags & O_DIRECT)) {
2433 loff_t endbyte;
2434 ssize_t written_buffered;
2435
2436 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2437 ppos, count, ocount);
2438 if (written < 0 || written == count)
2439 goto out;
2440 /*
2441 * direct-io write to a hole: fall through to buffered I/O
2442 * for completing the rest of the request.
2443 */
2444 pos += written;
2445 count -= written;
2446 written_buffered = generic_file_buffered_write(iocb, iov,
2447 nr_segs, pos, ppos, count,
2448 written);
2449 /*
2450 * If generic_file_buffered_write() retuned a synchronous error
2451 * then we want to return the number of bytes which were
2452 * direct-written, or the error code if that was zero. Note
2453 * that this differs from normal direct-io semantics, which
2454 * will return -EFOO even if some bytes were written.
2455 */
2456 if (written_buffered < 0) {
2457 err = written_buffered;
2458 goto out;
2459 }
2460
2461 /*
2462 * We need to ensure that the page cache pages are written to
2463 * disk and invalidated to preserve the expected O_DIRECT
2464 * semantics.
2465 */
2466 endbyte = pos + written_buffered - written - 1;
2467 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2468 if (err == 0) {
2469 written = written_buffered;
2470 invalidate_mapping_pages(mapping,
2471 pos >> PAGE_CACHE_SHIFT,
2472 endbyte >> PAGE_CACHE_SHIFT);
2473 } else {
2474 /*
2475 * We don't know how much we wrote, so just return
2476 * the number of bytes which were direct-written
2477 */
2478 }
2479 } else {
2480 written = generic_file_buffered_write(iocb, iov, nr_segs,
2481 pos, ppos, count, written);
2482 }
2483out:
2484 current->backing_dev_info = NULL;
2485 return written ? written : err;
2486}
2487EXPORT_SYMBOL(__generic_file_aio_write);
2488
2489/**
2490 * generic_file_aio_write - write data to a file
2491 * @iocb: IO state structure
2492 * @iov: vector with data to write
2493 * @nr_segs: number of segments in the vector
2494 * @pos: position in file where to write
2495 *
2496 * This is a wrapper around __generic_file_aio_write() to be used by most
2497 * filesystems. It takes care of syncing the file in case of O_SYNC file
2498 * and acquires i_mutex as needed.
2499 */
2500ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2501 unsigned long nr_segs, loff_t pos)
2502{
2503 struct file *file = iocb->ki_filp;
2504 struct inode *inode = file->f_mapping->host;
2505 struct blk_plug plug;
2506 ssize_t ret;
2507
2508 BUG_ON(iocb->ki_pos != pos);
2509
2510 mutex_lock(&inode->i_mutex);
2511 blk_start_plug(&plug);
2512 ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2513 mutex_unlock(&inode->i_mutex);
2514
2515 if (ret > 0 || ret == -EIOCBQUEUED) {
2516 ssize_t err;
2517
2518 err = generic_write_sync(file, pos, ret);
2519 if (err < 0 && ret > 0)
2520 ret = err;
2521 }
2522 blk_finish_plug(&plug);
2523 return ret;
2524}
2525EXPORT_SYMBOL(generic_file_aio_write);
2526
2527/**
2528 * try_to_release_page() - release old fs-specific metadata on a page
2529 *
2530 * @page: the page which the kernel is trying to free
2531 * @gfp_mask: memory allocation flags (and I/O mode)
2532 *
2533 * The address_space is to try to release any data against the page
2534 * (presumably at page->private). If the release was successful, return `1'.
2535 * Otherwise return zero.
2536 *
2537 * This may also be called if PG_fscache is set on a page, indicating that the
2538 * page is known to the local caching routines.
2539 *
2540 * The @gfp_mask argument specifies whether I/O may be performed to release
2541 * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2542 *
2543 */
2544int try_to_release_page(struct page *page, gfp_t gfp_mask)
2545{
2546 struct address_space * const mapping = page->mapping;
2547
2548 BUG_ON(!PageLocked(page));
2549 if (PageWriteback(page))
2550 return 0;
2551
2552 if (mapping && mapping->a_ops->releasepage)
2553 return mapping->a_ops->releasepage(page, gfp_mask);
2554 return try_to_free_buffers(page);
2555}
2556
2557EXPORT_SYMBOL(try_to_release_page);