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/mman.h>
26#include <linux/pagemap.h>
27#include <linux/file.h>
28#include <linux/uio.h>
29#include <linux/error-injection.h>
30#include <linux/hash.h>
31#include <linux/writeback.h>
32#include <linux/backing-dev.h>
33#include <linux/pagevec.h>
34#include <linux/security.h>
35#include <linux/cpuset.h>
36#include <linux/hugetlb.h>
37#include <linux/memcontrol.h>
38#include <linux/shmem_fs.h>
39#include <linux/rmap.h>
40#include <linux/delayacct.h>
41#include <linux/psi.h>
42#include <linux/ramfs.h>
43#include <linux/page_idle.h>
44#include <linux/migrate.h>
45#include <asm/pgalloc.h>
46#include <asm/tlbflush.h>
47#include "internal.h"
48
49#define CREATE_TRACE_POINTS
50#include <trace/events/filemap.h>
51
52/*
53 * FIXME: remove all knowledge of the buffer layer from the core VM
54 */
55#include <linux/buffer_head.h> /* for try_to_free_buffers */
56
57#include <asm/mman.h>
58
59/*
60 * Shared mappings implemented 30.11.1994. It's not fully working yet,
61 * though.
62 *
63 * Shared mappings now work. 15.8.1995 Bruno.
64 *
65 * finished 'unifying' the page and buffer cache and SMP-threaded the
66 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
67 *
68 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
69 */
70
71/*
72 * Lock ordering:
73 *
74 * ->i_mmap_rwsem (truncate_pagecache)
75 * ->private_lock (__free_pte->block_dirty_folio)
76 * ->swap_lock (exclusive_swap_page, others)
77 * ->i_pages lock
78 *
79 * ->i_rwsem
80 * ->invalidate_lock (acquired by fs in truncate path)
81 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
82 *
83 * ->mmap_lock
84 * ->i_mmap_rwsem
85 * ->page_table_lock or pte_lock (various, mainly in memory.c)
86 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
87 *
88 * ->mmap_lock
89 * ->invalidate_lock (filemap_fault)
90 * ->lock_page (filemap_fault, access_process_vm)
91 *
92 * ->i_rwsem (generic_perform_write)
93 * ->mmap_lock (fault_in_readable->do_page_fault)
94 *
95 * bdi->wb.list_lock
96 * sb_lock (fs/fs-writeback.c)
97 * ->i_pages lock (__sync_single_inode)
98 *
99 * ->i_mmap_rwsem
100 * ->anon_vma.lock (vma_adjust)
101 *
102 * ->anon_vma.lock
103 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
104 *
105 * ->page_table_lock or pte_lock
106 * ->swap_lock (try_to_unmap_one)
107 * ->private_lock (try_to_unmap_one)
108 * ->i_pages lock (try_to_unmap_one)
109 * ->lruvec->lru_lock (follow_page->mark_page_accessed)
110 * ->lruvec->lru_lock (check_pte_range->isolate_lru_page)
111 * ->private_lock (page_remove_rmap->set_page_dirty)
112 * ->i_pages lock (page_remove_rmap->set_page_dirty)
113 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
114 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
115 * ->memcg->move_lock (page_remove_rmap->lock_page_memcg)
116 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
117 * ->inode->i_lock (zap_pte_range->set_page_dirty)
118 * ->private_lock (zap_pte_range->block_dirty_folio)
119 *
120 * ->i_mmap_rwsem
121 * ->tasklist_lock (memory_failure, collect_procs_ao)
122 */
123
124static void page_cache_delete(struct address_space *mapping,
125 struct folio *folio, void *shadow)
126{
127 XA_STATE(xas, &mapping->i_pages, folio->index);
128 long nr = 1;
129
130 mapping_set_update(&xas, mapping);
131
132 /* hugetlb pages are represented by a single entry in the xarray */
133 if (!folio_test_hugetlb(folio)) {
134 xas_set_order(&xas, folio->index, folio_order(folio));
135 nr = folio_nr_pages(folio);
136 }
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) && !folio_test_hugetlb(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 page *page;
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 page = xas_find(&xas, max);
483 if (xas_retry(&xas, page))
484 continue;
485 /* Shadow entries don't count */
486 if (xa_is_value(page))
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 page != 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 pagevec pvec;
507 int nr_pages;
508
509 pagevec_init(&pvec);
510 while (index <= end) {
511 unsigned i;
512
513 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
514 end, PAGECACHE_TAG_WRITEBACK);
515 if (!nr_pages)
516 break;
517
518 for (i = 0; i < nr_pages; i++) {
519 struct page *page = pvec.pages[i];
520
521 wait_on_page_writeback(page);
522 ClearPageError(page);
523 }
524 pagevec_release(&pvec);
525 cond_resched();
526 }
527}
528
529/**
530 * filemap_fdatawait_range - wait for writeback to complete
531 * @mapping: address space structure to wait for
532 * @start_byte: offset in bytes where the range starts
533 * @end_byte: offset in bytes where the range ends (inclusive)
534 *
535 * Walk the list of under-writeback pages of the given address space
536 * in the given range and wait for all of them. Check error status of
537 * the address space and return it.
538 *
539 * Since the error status of the address space is cleared by this function,
540 * callers are responsible for checking the return value and handling and/or
541 * reporting the error.
542 *
543 * Return: error status of the address space.
544 */
545int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
546 loff_t end_byte)
547{
548 __filemap_fdatawait_range(mapping, start_byte, end_byte);
549 return filemap_check_errors(mapping);
550}
551EXPORT_SYMBOL(filemap_fdatawait_range);
552
553/**
554 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
555 * @mapping: address space structure to wait for
556 * @start_byte: offset in bytes where the range starts
557 * @end_byte: offset in bytes where the range ends (inclusive)
558 *
559 * Walk the list of under-writeback pages of the given address space in the
560 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
561 * this function does not clear error status of the address space.
562 *
563 * Use this function if callers don't handle errors themselves. Expected
564 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
565 * fsfreeze(8)
566 */
567int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
568 loff_t start_byte, loff_t end_byte)
569{
570 __filemap_fdatawait_range(mapping, start_byte, end_byte);
571 return filemap_check_and_keep_errors(mapping);
572}
573EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
574
575/**
576 * file_fdatawait_range - wait for writeback to complete
577 * @file: file pointing to address space structure to wait for
578 * @start_byte: offset in bytes where the range starts
579 * @end_byte: offset in bytes where the range ends (inclusive)
580 *
581 * Walk the list of under-writeback pages of the address space that file
582 * refers to, in the given range and wait for all of them. Check error
583 * status of the address space vs. the file->f_wb_err cursor and return it.
584 *
585 * Since the error status of the file is advanced by this function,
586 * callers are responsible for checking the return value and handling and/or
587 * reporting the error.
588 *
589 * Return: error status of the address space vs. the file->f_wb_err cursor.
590 */
591int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
592{
593 struct address_space *mapping = file->f_mapping;
594
595 __filemap_fdatawait_range(mapping, start_byte, end_byte);
596 return file_check_and_advance_wb_err(file);
597}
598EXPORT_SYMBOL(file_fdatawait_range);
599
600/**
601 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
602 * @mapping: address space structure to wait for
603 *
604 * Walk the list of under-writeback pages of the given address space
605 * and wait for all of them. Unlike filemap_fdatawait(), this function
606 * does not clear error status of the address space.
607 *
608 * Use this function if callers don't handle errors themselves. Expected
609 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
610 * fsfreeze(8)
611 *
612 * Return: error status of the address space.
613 */
614int filemap_fdatawait_keep_errors(struct address_space *mapping)
615{
616 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
617 return filemap_check_and_keep_errors(mapping);
618}
619EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
620
621/* Returns true if writeback might be needed or already in progress. */
622static bool mapping_needs_writeback(struct address_space *mapping)
623{
624 return mapping->nrpages;
625}
626
627bool filemap_range_has_writeback(struct address_space *mapping,
628 loff_t start_byte, loff_t end_byte)
629{
630 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
631 pgoff_t max = end_byte >> PAGE_SHIFT;
632 struct folio *folio;
633
634 if (end_byte < start_byte)
635 return false;
636
637 rcu_read_lock();
638 xas_for_each(&xas, folio, max) {
639 if (xas_retry(&xas, folio))
640 continue;
641 if (xa_is_value(folio))
642 continue;
643 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
644 folio_test_writeback(folio))
645 break;
646 }
647 rcu_read_unlock();
648 return folio != NULL;
649}
650EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
651
652/**
653 * filemap_write_and_wait_range - write out & wait on a file range
654 * @mapping: the address_space for the pages
655 * @lstart: offset in bytes where the range starts
656 * @lend: offset in bytes where the range ends (inclusive)
657 *
658 * Write out and wait upon file offsets lstart->lend, inclusive.
659 *
660 * Note that @lend is inclusive (describes the last byte to be written) so
661 * that this function can be used to write to the very end-of-file (end = -1).
662 *
663 * Return: error status of the address space.
664 */
665int filemap_write_and_wait_range(struct address_space *mapping,
666 loff_t lstart, loff_t lend)
667{
668 int err = 0, err2;
669
670 if (lend < lstart)
671 return 0;
672
673 if (mapping_needs_writeback(mapping)) {
674 err = __filemap_fdatawrite_range(mapping, lstart, lend,
675 WB_SYNC_ALL);
676 /*
677 * Even if the above returned error, the pages may be
678 * written partially (e.g. -ENOSPC), so we wait for it.
679 * But the -EIO is special case, it may indicate the worst
680 * thing (e.g. bug) happened, so we avoid waiting for it.
681 */
682 if (err != -EIO)
683 __filemap_fdatawait_range(mapping, lstart, lend);
684 }
685 err2 = filemap_check_errors(mapping);
686 if (!err)
687 err = err2;
688 return err;
689}
690EXPORT_SYMBOL(filemap_write_and_wait_range);
691
692void __filemap_set_wb_err(struct address_space *mapping, int err)
693{
694 errseq_t eseq = errseq_set(&mapping->wb_err, err);
695
696 trace_filemap_set_wb_err(mapping, eseq);
697}
698EXPORT_SYMBOL(__filemap_set_wb_err);
699
700/**
701 * file_check_and_advance_wb_err - report wb error (if any) that was previously
702 * and advance wb_err to current one
703 * @file: struct file on which the error is being reported
704 *
705 * When userland calls fsync (or something like nfsd does the equivalent), we
706 * want to report any writeback errors that occurred since the last fsync (or
707 * since the file was opened if there haven't been any).
708 *
709 * Grab the wb_err from the mapping. If it matches what we have in the file,
710 * then just quickly return 0. The file is all caught up.
711 *
712 * If it doesn't match, then take the mapping value, set the "seen" flag in
713 * it and try to swap it into place. If it works, or another task beat us
714 * to it with the new value, then update the f_wb_err and return the error
715 * portion. The error at this point must be reported via proper channels
716 * (a'la fsync, or NFS COMMIT operation, etc.).
717 *
718 * While we handle mapping->wb_err with atomic operations, the f_wb_err
719 * value is protected by the f_lock since we must ensure that it reflects
720 * the latest value swapped in for this file descriptor.
721 *
722 * Return: %0 on success, negative error code otherwise.
723 */
724int file_check_and_advance_wb_err(struct file *file)
725{
726 int err = 0;
727 errseq_t old = READ_ONCE(file->f_wb_err);
728 struct address_space *mapping = file->f_mapping;
729
730 /* Locklessly handle the common case where nothing has changed */
731 if (errseq_check(&mapping->wb_err, old)) {
732 /* Something changed, must use slow path */
733 spin_lock(&file->f_lock);
734 old = file->f_wb_err;
735 err = errseq_check_and_advance(&mapping->wb_err,
736 &file->f_wb_err);
737 trace_file_check_and_advance_wb_err(file, old);
738 spin_unlock(&file->f_lock);
739 }
740
741 /*
742 * We're mostly using this function as a drop in replacement for
743 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
744 * that the legacy code would have had on these flags.
745 */
746 clear_bit(AS_EIO, &mapping->flags);
747 clear_bit(AS_ENOSPC, &mapping->flags);
748 return err;
749}
750EXPORT_SYMBOL(file_check_and_advance_wb_err);
751
752/**
753 * file_write_and_wait_range - write out & wait on a file range
754 * @file: file pointing to address_space with pages
755 * @lstart: offset in bytes where the range starts
756 * @lend: offset in bytes where the range ends (inclusive)
757 *
758 * Write out and wait upon file offsets lstart->lend, inclusive.
759 *
760 * Note that @lend is inclusive (describes the last byte to be written) so
761 * that this function can be used to write to the very end-of-file (end = -1).
762 *
763 * After writing out and waiting on the data, we check and advance the
764 * f_wb_err cursor to the latest value, and return any errors detected there.
765 *
766 * Return: %0 on success, negative error code otherwise.
767 */
768int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
769{
770 int err = 0, err2;
771 struct address_space *mapping = file->f_mapping;
772
773 if (lend < lstart)
774 return 0;
775
776 if (mapping_needs_writeback(mapping)) {
777 err = __filemap_fdatawrite_range(mapping, lstart, lend,
778 WB_SYNC_ALL);
779 /* See comment of filemap_write_and_wait() */
780 if (err != -EIO)
781 __filemap_fdatawait_range(mapping, lstart, lend);
782 }
783 err2 = file_check_and_advance_wb_err(file);
784 if (!err)
785 err = err2;
786 return err;
787}
788EXPORT_SYMBOL(file_write_and_wait_range);
789
790/**
791 * replace_page_cache_folio - replace a pagecache folio with a new one
792 * @old: folio to be replaced
793 * @new: folio to replace with
794 *
795 * This function replaces a folio in the pagecache with a new one. On
796 * success it acquires the pagecache reference for the new folio and
797 * drops it for the old folio. Both the old and new folios must be
798 * locked. This function does not add the new folio to the LRU, the
799 * caller must do that.
800 *
801 * The remove + add is atomic. This function cannot fail.
802 */
803void replace_page_cache_folio(struct folio *old, struct folio *new)
804{
805 struct address_space *mapping = old->mapping;
806 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
807 pgoff_t offset = old->index;
808 XA_STATE(xas, &mapping->i_pages, offset);
809
810 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
811 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
812 VM_BUG_ON_FOLIO(new->mapping, new);
813
814 folio_get(new);
815 new->mapping = mapping;
816 new->index = offset;
817
818 mem_cgroup_migrate(old, new);
819
820 xas_lock_irq(&xas);
821 xas_store(&xas, new);
822
823 old->mapping = NULL;
824 /* hugetlb pages do not participate in page cache accounting. */
825 if (!folio_test_hugetlb(old))
826 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
827 if (!folio_test_hugetlb(new))
828 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
829 if (folio_test_swapbacked(old))
830 __lruvec_stat_sub_folio(old, NR_SHMEM);
831 if (folio_test_swapbacked(new))
832 __lruvec_stat_add_folio(new, NR_SHMEM);
833 xas_unlock_irq(&xas);
834 if (free_folio)
835 free_folio(old);
836 folio_put(old);
837}
838EXPORT_SYMBOL_GPL(replace_page_cache_folio);
839
840noinline int __filemap_add_folio(struct address_space *mapping,
841 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
842{
843 XA_STATE(xas, &mapping->i_pages, index);
844 int huge = folio_test_hugetlb(folio);
845 bool charged = false;
846 long nr = 1;
847
848 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
849 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
850 mapping_set_update(&xas, mapping);
851
852 if (!huge) {
853 int error = mem_cgroup_charge(folio, NULL, gfp);
854 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
855 if (error)
856 return error;
857 charged = true;
858 xas_set_order(&xas, index, folio_order(folio));
859 nr = folio_nr_pages(folio);
860 }
861
862 gfp &= GFP_RECLAIM_MASK;
863 folio_ref_add(folio, nr);
864 folio->mapping = mapping;
865 folio->index = xas.xa_index;
866
867 do {
868 unsigned int order = xa_get_order(xas.xa, xas.xa_index);
869 void *entry, *old = NULL;
870
871 if (order > folio_order(folio))
872 xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index),
873 order, gfp);
874 xas_lock_irq(&xas);
875 xas_for_each_conflict(&xas, entry) {
876 old = entry;
877 if (!xa_is_value(entry)) {
878 xas_set_err(&xas, -EEXIST);
879 goto unlock;
880 }
881 }
882
883 if (old) {
884 if (shadowp)
885 *shadowp = old;
886 /* entry may have been split before we acquired lock */
887 order = xa_get_order(xas.xa, xas.xa_index);
888 if (order > folio_order(folio)) {
889 /* How to handle large swap entries? */
890 BUG_ON(shmem_mapping(mapping));
891 xas_split(&xas, old, order);
892 xas_reset(&xas);
893 }
894 }
895
896 xas_store(&xas, folio);
897 if (xas_error(&xas))
898 goto unlock;
899
900 mapping->nrpages += nr;
901
902 /* hugetlb pages do not participate in page cache accounting */
903 if (!huge) {
904 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
905 if (folio_test_pmd_mappable(folio))
906 __lruvec_stat_mod_folio(folio,
907 NR_FILE_THPS, nr);
908 }
909unlock:
910 xas_unlock_irq(&xas);
911 } while (xas_nomem(&xas, gfp));
912
913 if (xas_error(&xas))
914 goto error;
915
916 trace_mm_filemap_add_to_page_cache(folio);
917 return 0;
918error:
919 if (charged)
920 mem_cgroup_uncharge(folio);
921 folio->mapping = NULL;
922 /* Leave page->index set: truncation relies upon it */
923 folio_put_refs(folio, nr);
924 return xas_error(&xas);
925}
926ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
927
928int filemap_add_folio(struct address_space *mapping, struct folio *folio,
929 pgoff_t index, gfp_t gfp)
930{
931 void *shadow = NULL;
932 int ret;
933
934 __folio_set_locked(folio);
935 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
936 if (unlikely(ret))
937 __folio_clear_locked(folio);
938 else {
939 /*
940 * The folio might have been evicted from cache only
941 * recently, in which case it should be activated like
942 * any other repeatedly accessed folio.
943 * The exception is folios getting rewritten; evicting other
944 * data from the working set, only to cache data that will
945 * get overwritten with something else, is a waste of memory.
946 */
947 WARN_ON_ONCE(folio_test_active(folio));
948 if (!(gfp & __GFP_WRITE) && shadow)
949 workingset_refault(folio, shadow);
950 folio_add_lru(folio);
951 }
952 return ret;
953}
954EXPORT_SYMBOL_GPL(filemap_add_folio);
955
956#ifdef CONFIG_NUMA
957struct folio *filemap_alloc_folio(gfp_t gfp, unsigned int order)
958{
959 int n;
960 struct folio *folio;
961
962 if (cpuset_do_page_mem_spread()) {
963 unsigned int cpuset_mems_cookie;
964 do {
965 cpuset_mems_cookie = read_mems_allowed_begin();
966 n = cpuset_mem_spread_node();
967 folio = __folio_alloc_node(gfp, order, n);
968 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
969
970 return folio;
971 }
972 return folio_alloc(gfp, order);
973}
974EXPORT_SYMBOL(filemap_alloc_folio);
975#endif
976
977/*
978 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
979 *
980 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
981 *
982 * @mapping1: the first mapping to lock
983 * @mapping2: the second mapping to lock
984 */
985void filemap_invalidate_lock_two(struct address_space *mapping1,
986 struct address_space *mapping2)
987{
988 if (mapping1 > mapping2)
989 swap(mapping1, mapping2);
990 if (mapping1)
991 down_write(&mapping1->invalidate_lock);
992 if (mapping2 && mapping1 != mapping2)
993 down_write_nested(&mapping2->invalidate_lock, 1);
994}
995EXPORT_SYMBOL(filemap_invalidate_lock_two);
996
997/*
998 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
999 *
1000 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1001 *
1002 * @mapping1: the first mapping to unlock
1003 * @mapping2: the second mapping to unlock
1004 */
1005void filemap_invalidate_unlock_two(struct address_space *mapping1,
1006 struct address_space *mapping2)
1007{
1008 if (mapping1)
1009 up_write(&mapping1->invalidate_lock);
1010 if (mapping2 && mapping1 != mapping2)
1011 up_write(&mapping2->invalidate_lock);
1012}
1013EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1014
1015/*
1016 * In order to wait for pages to become available there must be
1017 * waitqueues associated with pages. By using a hash table of
1018 * waitqueues where the bucket discipline is to maintain all
1019 * waiters on the same queue and wake all when any of the pages
1020 * become available, and for the woken contexts to check to be
1021 * sure the appropriate page became available, this saves space
1022 * at a cost of "thundering herd" phenomena during rare hash
1023 * collisions.
1024 */
1025#define PAGE_WAIT_TABLE_BITS 8
1026#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1027static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1028
1029static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1030{
1031 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1032}
1033
1034void __init pagecache_init(void)
1035{
1036 int i;
1037
1038 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1039 init_waitqueue_head(&folio_wait_table[i]);
1040
1041 page_writeback_init();
1042}
1043
1044/*
1045 * The page wait code treats the "wait->flags" somewhat unusually, because
1046 * we have multiple different kinds of waits, not just the usual "exclusive"
1047 * one.
1048 *
1049 * We have:
1050 *
1051 * (a) no special bits set:
1052 *
1053 * We're just waiting for the bit to be released, and when a waker
1054 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1055 * and remove it from the wait queue.
1056 *
1057 * Simple and straightforward.
1058 *
1059 * (b) WQ_FLAG_EXCLUSIVE:
1060 *
1061 * The waiter is waiting to get the lock, and only one waiter should
1062 * be woken up to avoid any thundering herd behavior. We'll set the
1063 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1064 *
1065 * This is the traditional exclusive wait.
1066 *
1067 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1068 *
1069 * The waiter is waiting to get the bit, and additionally wants the
1070 * lock to be transferred to it for fair lock behavior. If the lock
1071 * cannot be taken, we stop walking the wait queue without waking
1072 * the waiter.
1073 *
1074 * This is the "fair lock handoff" case, and in addition to setting
1075 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1076 * that it now has the lock.
1077 */
1078static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1079{
1080 unsigned int flags;
1081 struct wait_page_key *key = arg;
1082 struct wait_page_queue *wait_page
1083 = container_of(wait, struct wait_page_queue, wait);
1084
1085 if (!wake_page_match(wait_page, key))
1086 return 0;
1087
1088 /*
1089 * If it's a lock handoff wait, we get the bit for it, and
1090 * stop walking (and do not wake it up) if we can't.
1091 */
1092 flags = wait->flags;
1093 if (flags & WQ_FLAG_EXCLUSIVE) {
1094 if (test_bit(key->bit_nr, &key->folio->flags))
1095 return -1;
1096 if (flags & WQ_FLAG_CUSTOM) {
1097 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1098 return -1;
1099 flags |= WQ_FLAG_DONE;
1100 }
1101 }
1102
1103 /*
1104 * We are holding the wait-queue lock, but the waiter that
1105 * is waiting for this will be checking the flags without
1106 * any locking.
1107 *
1108 * So update the flags atomically, and wake up the waiter
1109 * afterwards to avoid any races. This store-release pairs
1110 * with the load-acquire in folio_wait_bit_common().
1111 */
1112 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1113 wake_up_state(wait->private, mode);
1114
1115 /*
1116 * Ok, we have successfully done what we're waiting for,
1117 * and we can unconditionally remove the wait entry.
1118 *
1119 * Note that this pairs with the "finish_wait()" in the
1120 * waiter, and has to be the absolute last thing we do.
1121 * After this list_del_init(&wait->entry) the wait entry
1122 * might be de-allocated and the process might even have
1123 * exited.
1124 */
1125 list_del_init_careful(&wait->entry);
1126 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1127}
1128
1129static void folio_wake_bit(struct folio *folio, int bit_nr)
1130{
1131 wait_queue_head_t *q = folio_waitqueue(folio);
1132 struct wait_page_key key;
1133 unsigned long flags;
1134 wait_queue_entry_t bookmark;
1135
1136 key.folio = folio;
1137 key.bit_nr = bit_nr;
1138 key.page_match = 0;
1139
1140 bookmark.flags = 0;
1141 bookmark.private = NULL;
1142 bookmark.func = NULL;
1143 INIT_LIST_HEAD(&bookmark.entry);
1144
1145 spin_lock_irqsave(&q->lock, flags);
1146 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1147
1148 while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1149 /*
1150 * Take a breather from holding the lock,
1151 * allow pages that finish wake up asynchronously
1152 * to acquire the lock and remove themselves
1153 * from wait queue
1154 */
1155 spin_unlock_irqrestore(&q->lock, flags);
1156 cpu_relax();
1157 spin_lock_irqsave(&q->lock, flags);
1158 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1159 }
1160
1161 /*
1162 * It's possible to miss clearing waiters here, when we woke our page
1163 * waiters, but the hashed waitqueue has waiters for other pages on it.
1164 * That's okay, it's a rare case. The next waker will clear it.
1165 *
1166 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1167 * other), the flag may be cleared in the course of freeing the page;
1168 * but that is not required for correctness.
1169 */
1170 if (!waitqueue_active(q) || !key.page_match)
1171 folio_clear_waiters(folio);
1172
1173 spin_unlock_irqrestore(&q->lock, flags);
1174}
1175
1176static void folio_wake(struct folio *folio, int bit)
1177{
1178 if (!folio_test_waiters(folio))
1179 return;
1180 folio_wake_bit(folio, bit);
1181}
1182
1183/*
1184 * A choice of three behaviors for folio_wait_bit_common():
1185 */
1186enum behavior {
1187 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1188 * __folio_lock() waiting on then setting PG_locked.
1189 */
1190 SHARED, /* Hold ref to page and check the bit when woken, like
1191 * folio_wait_writeback() waiting on PG_writeback.
1192 */
1193 DROP, /* Drop ref to page before wait, no check when woken,
1194 * like folio_put_wait_locked() on PG_locked.
1195 */
1196};
1197
1198/*
1199 * Attempt to check (or get) the folio flag, and mark us done
1200 * if successful.
1201 */
1202static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1203 struct wait_queue_entry *wait)
1204{
1205 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1206 if (test_and_set_bit(bit_nr, &folio->flags))
1207 return false;
1208 } else if (test_bit(bit_nr, &folio->flags))
1209 return false;
1210
1211 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1212 return true;
1213}
1214
1215/* How many times do we accept lock stealing from under a waiter? */
1216int sysctl_page_lock_unfairness = 5;
1217
1218static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1219 int state, enum behavior behavior)
1220{
1221 wait_queue_head_t *q = folio_waitqueue(folio);
1222 int unfairness = sysctl_page_lock_unfairness;
1223 struct wait_page_queue wait_page;
1224 wait_queue_entry_t *wait = &wait_page.wait;
1225 bool thrashing = false;
1226 unsigned long pflags;
1227 bool in_thrashing;
1228
1229 if (bit_nr == PG_locked &&
1230 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1231 delayacct_thrashing_start(&in_thrashing);
1232 psi_memstall_enter(&pflags);
1233 thrashing = true;
1234 }
1235
1236 init_wait(wait);
1237 wait->func = wake_page_function;
1238 wait_page.folio = folio;
1239 wait_page.bit_nr = bit_nr;
1240
1241repeat:
1242 wait->flags = 0;
1243 if (behavior == EXCLUSIVE) {
1244 wait->flags = WQ_FLAG_EXCLUSIVE;
1245 if (--unfairness < 0)
1246 wait->flags |= WQ_FLAG_CUSTOM;
1247 }
1248
1249 /*
1250 * Do one last check whether we can get the
1251 * page bit synchronously.
1252 *
1253 * Do the folio_set_waiters() marking before that
1254 * to let any waker we _just_ missed know they
1255 * need to wake us up (otherwise they'll never
1256 * even go to the slow case that looks at the
1257 * page queue), and add ourselves to the wait
1258 * queue if we need to sleep.
1259 *
1260 * This part needs to be done under the queue
1261 * lock to avoid races.
1262 */
1263 spin_lock_irq(&q->lock);
1264 folio_set_waiters(folio);
1265 if (!folio_trylock_flag(folio, bit_nr, wait))
1266 __add_wait_queue_entry_tail(q, wait);
1267 spin_unlock_irq(&q->lock);
1268
1269 /*
1270 * From now on, all the logic will be based on
1271 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1272 * see whether the page bit testing has already
1273 * been done by the wake function.
1274 *
1275 * We can drop our reference to the folio.
1276 */
1277 if (behavior == DROP)
1278 folio_put(folio);
1279
1280 /*
1281 * Note that until the "finish_wait()", or until
1282 * we see the WQ_FLAG_WOKEN flag, we need to
1283 * be very careful with the 'wait->flags', because
1284 * we may race with a waker that sets them.
1285 */
1286 for (;;) {
1287 unsigned int flags;
1288
1289 set_current_state(state);
1290
1291 /* Loop until we've been woken or interrupted */
1292 flags = smp_load_acquire(&wait->flags);
1293 if (!(flags & WQ_FLAG_WOKEN)) {
1294 if (signal_pending_state(state, current))
1295 break;
1296
1297 io_schedule();
1298 continue;
1299 }
1300
1301 /* If we were non-exclusive, we're done */
1302 if (behavior != EXCLUSIVE)
1303 break;
1304
1305 /* If the waker got the lock for us, we're done */
1306 if (flags & WQ_FLAG_DONE)
1307 break;
1308
1309 /*
1310 * Otherwise, if we're getting the lock, we need to
1311 * try to get it ourselves.
1312 *
1313 * And if that fails, we'll have to retry this all.
1314 */
1315 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1316 goto repeat;
1317
1318 wait->flags |= WQ_FLAG_DONE;
1319 break;
1320 }
1321
1322 /*
1323 * If a signal happened, this 'finish_wait()' may remove the last
1324 * waiter from the wait-queues, but the folio waiters bit will remain
1325 * set. That's ok. The next wakeup will take care of it, and trying
1326 * to do it here would be difficult and prone to races.
1327 */
1328 finish_wait(q, wait);
1329
1330 if (thrashing) {
1331 delayacct_thrashing_end(&in_thrashing);
1332 psi_memstall_leave(&pflags);
1333 }
1334
1335 /*
1336 * NOTE! The wait->flags weren't stable until we've done the
1337 * 'finish_wait()', and we could have exited the loop above due
1338 * to a signal, and had a wakeup event happen after the signal
1339 * test but before the 'finish_wait()'.
1340 *
1341 * So only after the finish_wait() can we reliably determine
1342 * if we got woken up or not, so we can now figure out the final
1343 * return value based on that state without races.
1344 *
1345 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1346 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1347 */
1348 if (behavior == EXCLUSIVE)
1349 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1350
1351 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1352}
1353
1354#ifdef CONFIG_MIGRATION
1355/**
1356 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1357 * @entry: migration swap entry.
1358 * @ptep: mapped pte pointer. Will return with the ptep unmapped. Only required
1359 * for pte entries, pass NULL for pmd entries.
1360 * @ptl: already locked ptl. This function will drop the lock.
1361 *
1362 * Wait for a migration entry referencing the given page to be removed. This is
1363 * equivalent to put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE) except
1364 * this can be called without taking a reference on the page. Instead this
1365 * should be called while holding the ptl for the migration entry referencing
1366 * the page.
1367 *
1368 * Returns after unmapping and unlocking the pte/ptl with pte_unmap_unlock().
1369 *
1370 * This follows the same logic as folio_wait_bit_common() so see the comments
1371 * there.
1372 */
1373void migration_entry_wait_on_locked(swp_entry_t entry, pte_t *ptep,
1374 spinlock_t *ptl)
1375{
1376 struct wait_page_queue wait_page;
1377 wait_queue_entry_t *wait = &wait_page.wait;
1378 bool thrashing = false;
1379 unsigned long pflags;
1380 bool in_thrashing;
1381 wait_queue_head_t *q;
1382 struct folio *folio = page_folio(pfn_swap_entry_to_page(entry));
1383
1384 q = folio_waitqueue(folio);
1385 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1386 delayacct_thrashing_start(&in_thrashing);
1387 psi_memstall_enter(&pflags);
1388 thrashing = true;
1389 }
1390
1391 init_wait(wait);
1392 wait->func = wake_page_function;
1393 wait_page.folio = folio;
1394 wait_page.bit_nr = PG_locked;
1395 wait->flags = 0;
1396
1397 spin_lock_irq(&q->lock);
1398 folio_set_waiters(folio);
1399 if (!folio_trylock_flag(folio, PG_locked, wait))
1400 __add_wait_queue_entry_tail(q, wait);
1401 spin_unlock_irq(&q->lock);
1402
1403 /*
1404 * If a migration entry exists for the page the migration path must hold
1405 * a valid reference to the page, and it must take the ptl to remove the
1406 * migration entry. So the page is valid until the ptl is dropped.
1407 */
1408 if (ptep)
1409 pte_unmap_unlock(ptep, ptl);
1410 else
1411 spin_unlock(ptl);
1412
1413 for (;;) {
1414 unsigned int flags;
1415
1416 set_current_state(TASK_UNINTERRUPTIBLE);
1417
1418 /* Loop until we've been woken or interrupted */
1419 flags = smp_load_acquire(&wait->flags);
1420 if (!(flags & WQ_FLAG_WOKEN)) {
1421 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1422 break;
1423
1424 io_schedule();
1425 continue;
1426 }
1427 break;
1428 }
1429
1430 finish_wait(q, wait);
1431
1432 if (thrashing) {
1433 delayacct_thrashing_end(&in_thrashing);
1434 psi_memstall_leave(&pflags);
1435 }
1436}
1437#endif
1438
1439void folio_wait_bit(struct folio *folio, int bit_nr)
1440{
1441 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1442}
1443EXPORT_SYMBOL(folio_wait_bit);
1444
1445int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1446{
1447 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1448}
1449EXPORT_SYMBOL(folio_wait_bit_killable);
1450
1451/**
1452 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1453 * @folio: The folio to wait for.
1454 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1455 *
1456 * The caller should hold a reference on @folio. They expect the page to
1457 * become unlocked relatively soon, but do not wish to hold up migration
1458 * (for example) by holding the reference while waiting for the folio to
1459 * come unlocked. After this function returns, the caller should not
1460 * dereference @folio.
1461 *
1462 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1463 */
1464static int folio_put_wait_locked(struct folio *folio, int state)
1465{
1466 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1467}
1468
1469/**
1470 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue
1471 * @folio: Folio defining the wait queue of interest
1472 * @waiter: Waiter to add to the queue
1473 *
1474 * Add an arbitrary @waiter to the wait queue for the nominated @folio.
1475 */
1476void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter)
1477{
1478 wait_queue_head_t *q = folio_waitqueue(folio);
1479 unsigned long flags;
1480
1481 spin_lock_irqsave(&q->lock, flags);
1482 __add_wait_queue_entry_tail(q, waiter);
1483 folio_set_waiters(folio);
1484 spin_unlock_irqrestore(&q->lock, flags);
1485}
1486EXPORT_SYMBOL_GPL(folio_add_wait_queue);
1487
1488#ifndef clear_bit_unlock_is_negative_byte
1489
1490/*
1491 * PG_waiters is the high bit in the same byte as PG_lock.
1492 *
1493 * On x86 (and on many other architectures), we can clear PG_lock and
1494 * test the sign bit at the same time. But if the architecture does
1495 * not support that special operation, we just do this all by hand
1496 * instead.
1497 *
1498 * The read of PG_waiters has to be after (or concurrently with) PG_locked
1499 * being cleared, but a memory barrier should be unnecessary since it is
1500 * in the same byte as PG_locked.
1501 */
1502static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1503{
1504 clear_bit_unlock(nr, mem);
1505 /* smp_mb__after_atomic(); */
1506 return test_bit(PG_waiters, mem);
1507}
1508
1509#endif
1510
1511/**
1512 * folio_unlock - Unlock a locked folio.
1513 * @folio: The folio.
1514 *
1515 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1516 *
1517 * Context: May be called from interrupt or process context. May not be
1518 * called from NMI context.
1519 */
1520void folio_unlock(struct folio *folio)
1521{
1522 /* Bit 7 allows x86 to check the byte's sign bit */
1523 BUILD_BUG_ON(PG_waiters != 7);
1524 BUILD_BUG_ON(PG_locked > 7);
1525 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1526 if (clear_bit_unlock_is_negative_byte(PG_locked, folio_flags(folio, 0)))
1527 folio_wake_bit(folio, PG_locked);
1528}
1529EXPORT_SYMBOL(folio_unlock);
1530
1531/**
1532 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1533 * @folio: The folio.
1534 *
1535 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1536 * it. The folio reference held for PG_private_2 being set is released.
1537 *
1538 * This is, for example, used when a netfs folio is being written to a local
1539 * disk cache, thereby allowing writes to the cache for the same folio to be
1540 * serialised.
1541 */
1542void folio_end_private_2(struct folio *folio)
1543{
1544 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1545 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1546 folio_wake_bit(folio, PG_private_2);
1547 folio_put(folio);
1548}
1549EXPORT_SYMBOL(folio_end_private_2);
1550
1551/**
1552 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1553 * @folio: The folio to wait on.
1554 *
1555 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio.
1556 */
1557void folio_wait_private_2(struct folio *folio)
1558{
1559 while (folio_test_private_2(folio))
1560 folio_wait_bit(folio, PG_private_2);
1561}
1562EXPORT_SYMBOL(folio_wait_private_2);
1563
1564/**
1565 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1566 * @folio: The folio to wait on.
1567 *
1568 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio or until a
1569 * fatal signal is received by the calling task.
1570 *
1571 * Return:
1572 * - 0 if successful.
1573 * - -EINTR if a fatal signal was encountered.
1574 */
1575int folio_wait_private_2_killable(struct folio *folio)
1576{
1577 int ret = 0;
1578
1579 while (folio_test_private_2(folio)) {
1580 ret = folio_wait_bit_killable(folio, PG_private_2);
1581 if (ret < 0)
1582 break;
1583 }
1584
1585 return ret;
1586}
1587EXPORT_SYMBOL(folio_wait_private_2_killable);
1588
1589/**
1590 * folio_end_writeback - End writeback against a folio.
1591 * @folio: The folio.
1592 */
1593void folio_end_writeback(struct folio *folio)
1594{
1595 /*
1596 * folio_test_clear_reclaim() could be used here but it is an
1597 * atomic operation and overkill in this particular case. Failing
1598 * to shuffle a folio marked for immediate reclaim is too mild
1599 * a gain to justify taking an atomic operation penalty at the
1600 * end of every folio writeback.
1601 */
1602 if (folio_test_reclaim(folio)) {
1603 folio_clear_reclaim(folio);
1604 folio_rotate_reclaimable(folio);
1605 }
1606
1607 /*
1608 * Writeback does not hold a folio reference of its own, relying
1609 * on truncation to wait for the clearing of PG_writeback.
1610 * But here we must make sure that the folio is not freed and
1611 * reused before the folio_wake().
1612 */
1613 folio_get(folio);
1614 if (!__folio_end_writeback(folio))
1615 BUG();
1616
1617 smp_mb__after_atomic();
1618 folio_wake(folio, PG_writeback);
1619 acct_reclaim_writeback(folio);
1620 folio_put(folio);
1621}
1622EXPORT_SYMBOL(folio_end_writeback);
1623
1624/*
1625 * After completing I/O on a page, call this routine to update the page
1626 * flags appropriately
1627 */
1628void page_endio(struct page *page, bool is_write, int err)
1629{
1630 struct folio *folio = page_folio(page);
1631
1632 if (!is_write) {
1633 if (!err) {
1634 folio_mark_uptodate(folio);
1635 } else {
1636 folio_clear_uptodate(folio);
1637 folio_set_error(folio);
1638 }
1639 folio_unlock(folio);
1640 } else {
1641 if (err) {
1642 struct address_space *mapping;
1643
1644 folio_set_error(folio);
1645 mapping = folio_mapping(folio);
1646 if (mapping)
1647 mapping_set_error(mapping, err);
1648 }
1649 folio_end_writeback(folio);
1650 }
1651}
1652EXPORT_SYMBOL_GPL(page_endio);
1653
1654/**
1655 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1656 * @folio: The folio to lock
1657 */
1658void __folio_lock(struct folio *folio)
1659{
1660 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1661 EXCLUSIVE);
1662}
1663EXPORT_SYMBOL(__folio_lock);
1664
1665int __folio_lock_killable(struct folio *folio)
1666{
1667 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1668 EXCLUSIVE);
1669}
1670EXPORT_SYMBOL_GPL(__folio_lock_killable);
1671
1672static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1673{
1674 struct wait_queue_head *q = folio_waitqueue(folio);
1675 int ret = 0;
1676
1677 wait->folio = folio;
1678 wait->bit_nr = PG_locked;
1679
1680 spin_lock_irq(&q->lock);
1681 __add_wait_queue_entry_tail(q, &wait->wait);
1682 folio_set_waiters(folio);
1683 ret = !folio_trylock(folio);
1684 /*
1685 * If we were successful now, we know we're still on the
1686 * waitqueue as we're still under the lock. This means it's
1687 * safe to remove and return success, we know the callback
1688 * isn't going to trigger.
1689 */
1690 if (!ret)
1691 __remove_wait_queue(q, &wait->wait);
1692 else
1693 ret = -EIOCBQUEUED;
1694 spin_unlock_irq(&q->lock);
1695 return ret;
1696}
1697
1698/*
1699 * Return values:
1700 * true - folio is locked; mmap_lock is still held.
1701 * false - folio is not locked.
1702 * mmap_lock has been released (mmap_read_unlock(), unless flags had both
1703 * FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1704 * which case mmap_lock is still held.
1705 *
1706 * If neither ALLOW_RETRY nor KILLABLE are set, will always return true
1707 * with the folio locked and the mmap_lock unperturbed.
1708 */
1709bool __folio_lock_or_retry(struct folio *folio, struct mm_struct *mm,
1710 unsigned int flags)
1711{
1712 if (fault_flag_allow_retry_first(flags)) {
1713 /*
1714 * CAUTION! In this case, mmap_lock is not released
1715 * even though return 0.
1716 */
1717 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1718 return false;
1719
1720 mmap_read_unlock(mm);
1721 if (flags & FAULT_FLAG_KILLABLE)
1722 folio_wait_locked_killable(folio);
1723 else
1724 folio_wait_locked(folio);
1725 return false;
1726 }
1727 if (flags & FAULT_FLAG_KILLABLE) {
1728 bool ret;
1729
1730 ret = __folio_lock_killable(folio);
1731 if (ret) {
1732 mmap_read_unlock(mm);
1733 return false;
1734 }
1735 } else {
1736 __folio_lock(folio);
1737 }
1738
1739 return true;
1740}
1741
1742/**
1743 * page_cache_next_miss() - Find the next gap in the page cache.
1744 * @mapping: Mapping.
1745 * @index: Index.
1746 * @max_scan: Maximum range to search.
1747 *
1748 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1749 * gap with the lowest index.
1750 *
1751 * This function may be called under the rcu_read_lock. However, this will
1752 * not atomically search a snapshot of the cache at a single point in time.
1753 * For example, if a gap is created at index 5, then subsequently a gap is
1754 * created at index 10, page_cache_next_miss covering both indices may
1755 * return 10 if called under the rcu_read_lock.
1756 *
1757 * Return: The index of the gap if found, otherwise an index outside the
1758 * range specified (in which case 'return - index >= max_scan' will be true).
1759 * In the rare case of index wrap-around, 0 will be returned.
1760 */
1761pgoff_t page_cache_next_miss(struct address_space *mapping,
1762 pgoff_t index, unsigned long max_scan)
1763{
1764 XA_STATE(xas, &mapping->i_pages, index);
1765
1766 while (max_scan--) {
1767 void *entry = xas_next(&xas);
1768 if (!entry || xa_is_value(entry))
1769 break;
1770 if (xas.xa_index == 0)
1771 break;
1772 }
1773
1774 return xas.xa_index;
1775}
1776EXPORT_SYMBOL(page_cache_next_miss);
1777
1778/**
1779 * page_cache_prev_miss() - Find the previous gap in the page cache.
1780 * @mapping: Mapping.
1781 * @index: Index.
1782 * @max_scan: Maximum range to search.
1783 *
1784 * Search the range [max(index - max_scan + 1, 0), index] for the
1785 * gap with the highest index.
1786 *
1787 * This function may be called under the rcu_read_lock. However, this will
1788 * not atomically search a snapshot of the cache at a single point in time.
1789 * For example, if a gap is created at index 10, then subsequently a gap is
1790 * created at index 5, page_cache_prev_miss() covering both indices may
1791 * return 5 if called under the rcu_read_lock.
1792 *
1793 * Return: The index of the gap if found, otherwise an index outside the
1794 * range specified (in which case 'index - return >= max_scan' will be true).
1795 * In the rare case of wrap-around, ULONG_MAX will be returned.
1796 */
1797pgoff_t page_cache_prev_miss(struct address_space *mapping,
1798 pgoff_t index, unsigned long max_scan)
1799{
1800 XA_STATE(xas, &mapping->i_pages, index);
1801
1802 while (max_scan--) {
1803 void *entry = xas_prev(&xas);
1804 if (!entry || xa_is_value(entry))
1805 break;
1806 if (xas.xa_index == ULONG_MAX)
1807 break;
1808 }
1809
1810 return xas.xa_index;
1811}
1812EXPORT_SYMBOL(page_cache_prev_miss);
1813
1814/*
1815 * Lockless page cache protocol:
1816 * On the lookup side:
1817 * 1. Load the folio from i_pages
1818 * 2. Increment the refcount if it's not zero
1819 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1820 *
1821 * On the removal side:
1822 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1823 * B. Remove the page from i_pages
1824 * C. Return the page to the page allocator
1825 *
1826 * This means that any page may have its reference count temporarily
1827 * increased by a speculative page cache (or fast GUP) lookup as it can
1828 * be allocated by another user before the RCU grace period expires.
1829 * Because the refcount temporarily acquired here may end up being the
1830 * last refcount on the page, any page allocation must be freeable by
1831 * folio_put().
1832 */
1833
1834/*
1835 * mapping_get_entry - Get a page cache entry.
1836 * @mapping: the address_space to search
1837 * @index: The page cache index.
1838 *
1839 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1840 * it is returned with an increased refcount. If it is a shadow entry
1841 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1842 * it is returned without further action.
1843 *
1844 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1845 */
1846static void *mapping_get_entry(struct address_space *mapping, pgoff_t index)
1847{
1848 XA_STATE(xas, &mapping->i_pages, index);
1849 struct folio *folio;
1850
1851 rcu_read_lock();
1852repeat:
1853 xas_reset(&xas);
1854 folio = xas_load(&xas);
1855 if (xas_retry(&xas, folio))
1856 goto repeat;
1857 /*
1858 * A shadow entry of a recently evicted page, or a swap entry from
1859 * shmem/tmpfs. Return it without attempting to raise page count.
1860 */
1861 if (!folio || xa_is_value(folio))
1862 goto out;
1863
1864 if (!folio_try_get_rcu(folio))
1865 goto repeat;
1866
1867 if (unlikely(folio != xas_reload(&xas))) {
1868 folio_put(folio);
1869 goto repeat;
1870 }
1871out:
1872 rcu_read_unlock();
1873
1874 return folio;
1875}
1876
1877/**
1878 * __filemap_get_folio - Find and get a reference to a folio.
1879 * @mapping: The address_space to search.
1880 * @index: The page index.
1881 * @fgp_flags: %FGP flags modify how the folio is returned.
1882 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1883 *
1884 * Looks up the page cache entry at @mapping & @index.
1885 *
1886 * @fgp_flags can be zero or more of these flags:
1887 *
1888 * * %FGP_ACCESSED - The folio will be marked accessed.
1889 * * %FGP_LOCK - The folio is returned locked.
1890 * * %FGP_ENTRY - If there is a shadow / swap / DAX entry, return it
1891 * instead of allocating a new folio to replace it.
1892 * * %FGP_CREAT - If no page is present then a new page is allocated using
1893 * @gfp and added to the page cache and the VM's LRU list.
1894 * The page is returned locked and with an increased refcount.
1895 * * %FGP_FOR_MMAP - The caller wants to do its own locking dance if the
1896 * page is already in cache. If the page was allocated, unlock it before
1897 * returning so the caller can do the same dance.
1898 * * %FGP_WRITE - The page will be written to by the caller.
1899 * * %FGP_NOFS - __GFP_FS will get cleared in gfp.
1900 * * %FGP_NOWAIT - Don't get blocked by page lock.
1901 * * %FGP_STABLE - Wait for the folio to be stable (finished writeback)
1902 *
1903 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1904 * if the %GFP flags specified for %FGP_CREAT are atomic.
1905 *
1906 * If there is a page cache page, it is returned with an increased refcount.
1907 *
1908 * Return: The found folio or %NULL otherwise.
1909 */
1910struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1911 int fgp_flags, gfp_t gfp)
1912{
1913 struct folio *folio;
1914
1915repeat:
1916 folio = mapping_get_entry(mapping, index);
1917 if (xa_is_value(folio)) {
1918 if (fgp_flags & FGP_ENTRY)
1919 return folio;
1920 folio = NULL;
1921 }
1922 if (!folio)
1923 goto no_page;
1924
1925 if (fgp_flags & FGP_LOCK) {
1926 if (fgp_flags & FGP_NOWAIT) {
1927 if (!folio_trylock(folio)) {
1928 folio_put(folio);
1929 return NULL;
1930 }
1931 } else {
1932 folio_lock(folio);
1933 }
1934
1935 /* Has the page been truncated? */
1936 if (unlikely(folio->mapping != mapping)) {
1937 folio_unlock(folio);
1938 folio_put(folio);
1939 goto repeat;
1940 }
1941 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1942 }
1943
1944 if (fgp_flags & FGP_ACCESSED)
1945 folio_mark_accessed(folio);
1946 else if (fgp_flags & FGP_WRITE) {
1947 /* Clear idle flag for buffer write */
1948 if (folio_test_idle(folio))
1949 folio_clear_idle(folio);
1950 }
1951
1952 if (fgp_flags & FGP_STABLE)
1953 folio_wait_stable(folio);
1954no_page:
1955 if (!folio && (fgp_flags & FGP_CREAT)) {
1956 int err;
1957 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1958 gfp |= __GFP_WRITE;
1959 if (fgp_flags & FGP_NOFS)
1960 gfp &= ~__GFP_FS;
1961 if (fgp_flags & FGP_NOWAIT) {
1962 gfp &= ~GFP_KERNEL;
1963 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1964 }
1965
1966 folio = filemap_alloc_folio(gfp, 0);
1967 if (!folio)
1968 return NULL;
1969
1970 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1971 fgp_flags |= FGP_LOCK;
1972
1973 /* Init accessed so avoid atomic mark_page_accessed later */
1974 if (fgp_flags & FGP_ACCESSED)
1975 __folio_set_referenced(folio);
1976
1977 err = filemap_add_folio(mapping, folio, index, gfp);
1978 if (unlikely(err)) {
1979 folio_put(folio);
1980 folio = NULL;
1981 if (err == -EEXIST)
1982 goto repeat;
1983 }
1984
1985 /*
1986 * filemap_add_folio locks the page, and for mmap
1987 * we expect an unlocked page.
1988 */
1989 if (folio && (fgp_flags & FGP_FOR_MMAP))
1990 folio_unlock(folio);
1991 }
1992
1993 return folio;
1994}
1995EXPORT_SYMBOL(__filemap_get_folio);
1996
1997static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
1998 xa_mark_t mark)
1999{
2000 struct folio *folio;
2001
2002retry:
2003 if (mark == XA_PRESENT)
2004 folio = xas_find(xas, max);
2005 else
2006 folio = xas_find_marked(xas, max, mark);
2007
2008 if (xas_retry(xas, folio))
2009 goto retry;
2010 /*
2011 * A shadow entry of a recently evicted page, a swap
2012 * entry from shmem/tmpfs or a DAX entry. Return it
2013 * without attempting to raise page count.
2014 */
2015 if (!folio || xa_is_value(folio))
2016 return folio;
2017
2018 if (!folio_try_get_rcu(folio))
2019 goto reset;
2020
2021 if (unlikely(folio != xas_reload(xas))) {
2022 folio_put(folio);
2023 goto reset;
2024 }
2025
2026 return folio;
2027reset:
2028 xas_reset(xas);
2029 goto retry;
2030}
2031
2032/**
2033 * find_get_entries - gang pagecache lookup
2034 * @mapping: The address_space to search
2035 * @start: The starting page cache index
2036 * @end: The final page index (inclusive).
2037 * @fbatch: Where the resulting entries are placed.
2038 * @indices: The cache indices corresponding to the entries in @entries
2039 *
2040 * find_get_entries() will search for and return a batch of entries in
2041 * the mapping. The entries are placed in @fbatch. find_get_entries()
2042 * takes a reference on any actual folios it returns.
2043 *
2044 * The entries have ascending indexes. The indices may not be consecutive
2045 * due to not-present entries or large folios.
2046 *
2047 * Any shadow entries of evicted folios, or swap entries from
2048 * shmem/tmpfs, are included in the returned array.
2049 *
2050 * Return: The number of entries which were found.
2051 */
2052unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2053 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2054{
2055 XA_STATE(xas, &mapping->i_pages, *start);
2056 struct folio *folio;
2057
2058 rcu_read_lock();
2059 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2060 indices[fbatch->nr] = xas.xa_index;
2061 if (!folio_batch_add(fbatch, folio))
2062 break;
2063 }
2064 rcu_read_unlock();
2065
2066 if (folio_batch_count(fbatch)) {
2067 unsigned long nr = 1;
2068 int idx = folio_batch_count(fbatch) - 1;
2069
2070 folio = fbatch->folios[idx];
2071 if (!xa_is_value(folio) && !folio_test_hugetlb(folio))
2072 nr = folio_nr_pages(folio);
2073 *start = indices[idx] + nr;
2074 }
2075 return folio_batch_count(fbatch);
2076}
2077
2078/**
2079 * find_lock_entries - Find a batch of pagecache entries.
2080 * @mapping: The address_space to search.
2081 * @start: The starting page cache index.
2082 * @end: The final page index (inclusive).
2083 * @fbatch: Where the resulting entries are placed.
2084 * @indices: The cache indices of the entries in @fbatch.
2085 *
2086 * find_lock_entries() will return a batch of entries from @mapping.
2087 * Swap, shadow and DAX entries are included. Folios are returned
2088 * locked and with an incremented refcount. Folios which are locked
2089 * by somebody else or under writeback are skipped. Folios which are
2090 * partially outside the range are not returned.
2091 *
2092 * The entries have ascending indexes. The indices may not be consecutive
2093 * due to not-present entries, large folios, folios which could not be
2094 * locked or folios under writeback.
2095 *
2096 * Return: The number of entries which were found.
2097 */
2098unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2099 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2100{
2101 XA_STATE(xas, &mapping->i_pages, *start);
2102 struct folio *folio;
2103
2104 rcu_read_lock();
2105 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2106 if (!xa_is_value(folio)) {
2107 if (folio->index < *start)
2108 goto put;
2109 if (folio->index + folio_nr_pages(folio) - 1 > end)
2110 goto put;
2111 if (!folio_trylock(folio))
2112 goto put;
2113 if (folio->mapping != mapping ||
2114 folio_test_writeback(folio))
2115 goto unlock;
2116 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2117 folio);
2118 }
2119 indices[fbatch->nr] = xas.xa_index;
2120 if (!folio_batch_add(fbatch, folio))
2121 break;
2122 continue;
2123unlock:
2124 folio_unlock(folio);
2125put:
2126 folio_put(folio);
2127 }
2128 rcu_read_unlock();
2129
2130 if (folio_batch_count(fbatch)) {
2131 unsigned long nr = 1;
2132 int idx = folio_batch_count(fbatch) - 1;
2133
2134 folio = fbatch->folios[idx];
2135 if (!xa_is_value(folio) && !folio_test_hugetlb(folio))
2136 nr = folio_nr_pages(folio);
2137 *start = indices[idx] + nr;
2138 }
2139 return folio_batch_count(fbatch);
2140}
2141
2142/**
2143 * filemap_get_folios - Get a batch of folios
2144 * @mapping: The address_space to search
2145 * @start: The starting page index
2146 * @end: The final page index (inclusive)
2147 * @fbatch: The batch to fill.
2148 *
2149 * Search for and return a batch of folios in the mapping starting at
2150 * index @start and up to index @end (inclusive). The folios are returned
2151 * in @fbatch with an elevated reference count.
2152 *
2153 * The first folio may start before @start; if it does, it will contain
2154 * @start. The final folio may extend beyond @end; if it does, it will
2155 * contain @end. The folios have ascending indices. There may be gaps
2156 * between the folios if there are indices which have no folio in the
2157 * page cache. If folios are added to or removed from the page cache
2158 * while this is running, they may or may not be found by this call.
2159 *
2160 * Return: The number of folios which were found.
2161 * We also update @start to index the next folio for the traversal.
2162 */
2163unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2164 pgoff_t end, struct folio_batch *fbatch)
2165{
2166 XA_STATE(xas, &mapping->i_pages, *start);
2167 struct folio *folio;
2168
2169 rcu_read_lock();
2170 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2171 /* Skip over shadow, swap and DAX entries */
2172 if (xa_is_value(folio))
2173 continue;
2174 if (!folio_batch_add(fbatch, folio)) {
2175 unsigned long nr = folio_nr_pages(folio);
2176
2177 if (folio_test_hugetlb(folio))
2178 nr = 1;
2179 *start = folio->index + nr;
2180 goto out;
2181 }
2182 }
2183
2184 /*
2185 * We come here when there is no page beyond @end. We take care to not
2186 * overflow the index @start as it confuses some of the callers. This
2187 * breaks the iteration when there is a page at index -1 but that is
2188 * already broken anyway.
2189 */
2190 if (end == (pgoff_t)-1)
2191 *start = (pgoff_t)-1;
2192 else
2193 *start = end + 1;
2194out:
2195 rcu_read_unlock();
2196
2197 return folio_batch_count(fbatch);
2198}
2199EXPORT_SYMBOL(filemap_get_folios);
2200
2201static inline
2202bool folio_more_pages(struct folio *folio, pgoff_t index, pgoff_t max)
2203{
2204 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
2205 return false;
2206 if (index >= max)
2207 return false;
2208 return index < folio->index + folio_nr_pages(folio) - 1;
2209}
2210
2211/**
2212 * filemap_get_folios_contig - Get a batch of contiguous folios
2213 * @mapping: The address_space to search
2214 * @start: The starting page index
2215 * @end: The final page index (inclusive)
2216 * @fbatch: The batch to fill
2217 *
2218 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2219 * except the returned folios are guaranteed to be contiguous. This may
2220 * not return all contiguous folios if the batch gets filled up.
2221 *
2222 * Return: The number of folios found.
2223 * Also update @start to be positioned for traversal of the next folio.
2224 */
2225
2226unsigned filemap_get_folios_contig(struct address_space *mapping,
2227 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2228{
2229 XA_STATE(xas, &mapping->i_pages, *start);
2230 unsigned long nr;
2231 struct folio *folio;
2232
2233 rcu_read_lock();
2234
2235 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2236 folio = xas_next(&xas)) {
2237 if (xas_retry(&xas, folio))
2238 continue;
2239 /*
2240 * If the entry has been swapped out, we can stop looking.
2241 * No current caller is looking for DAX entries.
2242 */
2243 if (xa_is_value(folio))
2244 goto update_start;
2245
2246 if (!folio_try_get_rcu(folio))
2247 goto retry;
2248
2249 if (unlikely(folio != xas_reload(&xas)))
2250 goto put_folio;
2251
2252 if (!folio_batch_add(fbatch, folio)) {
2253 nr = folio_nr_pages(folio);
2254
2255 if (folio_test_hugetlb(folio))
2256 nr = 1;
2257 *start = folio->index + nr;
2258 goto out;
2259 }
2260 continue;
2261put_folio:
2262 folio_put(folio);
2263
2264retry:
2265 xas_reset(&xas);
2266 }
2267
2268update_start:
2269 nr = folio_batch_count(fbatch);
2270
2271 if (nr) {
2272 folio = fbatch->folios[nr - 1];
2273 if (folio_test_hugetlb(folio))
2274 *start = folio->index + 1;
2275 else
2276 *start = folio->index + folio_nr_pages(folio);
2277 }
2278out:
2279 rcu_read_unlock();
2280 return folio_batch_count(fbatch);
2281}
2282EXPORT_SYMBOL(filemap_get_folios_contig);
2283
2284/**
2285 * find_get_pages_range_tag - Find and return head pages matching @tag.
2286 * @mapping: the address_space to search
2287 * @index: the starting page index
2288 * @end: The final page index (inclusive)
2289 * @tag: the tag index
2290 * @nr_pages: the maximum number of pages
2291 * @pages: where the resulting pages are placed
2292 *
2293 * Like find_get_pages_range(), except we only return head pages which are
2294 * tagged with @tag. @index is updated to the index immediately after the
2295 * last page we return, ready for the next iteration.
2296 *
2297 * Return: the number of pages which were found.
2298 */
2299unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
2300 pgoff_t end, xa_mark_t tag, unsigned int nr_pages,
2301 struct page **pages)
2302{
2303 XA_STATE(xas, &mapping->i_pages, *index);
2304 struct folio *folio;
2305 unsigned ret = 0;
2306
2307 if (unlikely(!nr_pages))
2308 return 0;
2309
2310 rcu_read_lock();
2311 while ((folio = find_get_entry(&xas, end, tag))) {
2312 /*
2313 * Shadow entries should never be tagged, but this iteration
2314 * is lockless so there is a window for page reclaim to evict
2315 * a page we saw tagged. Skip over it.
2316 */
2317 if (xa_is_value(folio))
2318 continue;
2319
2320 pages[ret] = &folio->page;
2321 if (++ret == nr_pages) {
2322 *index = folio->index + folio_nr_pages(folio);
2323 goto out;
2324 }
2325 }
2326
2327 /*
2328 * We come here when we got to @end. We take care to not overflow the
2329 * index @index as it confuses some of the callers. This breaks the
2330 * iteration when there is a page at index -1 but that is already
2331 * broken anyway.
2332 */
2333 if (end == (pgoff_t)-1)
2334 *index = (pgoff_t)-1;
2335 else
2336 *index = end + 1;
2337out:
2338 rcu_read_unlock();
2339
2340 return ret;
2341}
2342EXPORT_SYMBOL(find_get_pages_range_tag);
2343
2344/*
2345 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2346 * a _large_ part of the i/o request. Imagine the worst scenario:
2347 *
2348 * ---R__________________________________________B__________
2349 * ^ reading here ^ bad block(assume 4k)
2350 *
2351 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2352 * => failing the whole request => read(R) => read(R+1) =>
2353 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2354 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2355 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2356 *
2357 * It is going insane. Fix it by quickly scaling down the readahead size.
2358 */
2359static void shrink_readahead_size_eio(struct file_ra_state *ra)
2360{
2361 ra->ra_pages /= 4;
2362}
2363
2364/*
2365 * filemap_get_read_batch - Get a batch of folios for read
2366 *
2367 * Get a batch of folios which represent a contiguous range of bytes in
2368 * the file. No exceptional entries will be returned. If @index is in
2369 * the middle of a folio, the entire folio will be returned. The last
2370 * folio in the batch may have the readahead flag set or the uptodate flag
2371 * clear so that the caller can take the appropriate action.
2372 */
2373static void filemap_get_read_batch(struct address_space *mapping,
2374 pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
2375{
2376 XA_STATE(xas, &mapping->i_pages, index);
2377 struct folio *folio;
2378
2379 rcu_read_lock();
2380 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2381 if (xas_retry(&xas, folio))
2382 continue;
2383 if (xas.xa_index > max || xa_is_value(folio))
2384 break;
2385 if (xa_is_sibling(folio))
2386 break;
2387 if (!folio_try_get_rcu(folio))
2388 goto retry;
2389
2390 if (unlikely(folio != xas_reload(&xas)))
2391 goto put_folio;
2392
2393 if (!folio_batch_add(fbatch, folio))
2394 break;
2395 if (!folio_test_uptodate(folio))
2396 break;
2397 if (folio_test_readahead(folio))
2398 break;
2399 xas_advance(&xas, folio->index + folio_nr_pages(folio) - 1);
2400 continue;
2401put_folio:
2402 folio_put(folio);
2403retry:
2404 xas_reset(&xas);
2405 }
2406 rcu_read_unlock();
2407}
2408
2409static int filemap_read_folio(struct file *file, filler_t filler,
2410 struct folio *folio)
2411{
2412 bool workingset = folio_test_workingset(folio);
2413 unsigned long pflags;
2414 int error;
2415
2416 /*
2417 * A previous I/O error may have been due to temporary failures,
2418 * eg. multipath errors. PG_error will be set again if read_folio
2419 * fails.
2420 */
2421 folio_clear_error(folio);
2422
2423 /* Start the actual read. The read will unlock the page. */
2424 if (unlikely(workingset))
2425 psi_memstall_enter(&pflags);
2426 error = filler(file, folio);
2427 if (unlikely(workingset))
2428 psi_memstall_leave(&pflags);
2429 if (error)
2430 return error;
2431
2432 error = folio_wait_locked_killable(folio);
2433 if (error)
2434 return error;
2435 if (folio_test_uptodate(folio))
2436 return 0;
2437 if (file)
2438 shrink_readahead_size_eio(&file->f_ra);
2439 return -EIO;
2440}
2441
2442static bool filemap_range_uptodate(struct address_space *mapping,
2443 loff_t pos, struct iov_iter *iter, struct folio *folio)
2444{
2445 int count;
2446
2447 if (folio_test_uptodate(folio))
2448 return true;
2449 /* pipes can't handle partially uptodate pages */
2450 if (iov_iter_is_pipe(iter))
2451 return false;
2452 if (!mapping->a_ops->is_partially_uptodate)
2453 return false;
2454 if (mapping->host->i_blkbits >= folio_shift(folio))
2455 return false;
2456
2457 count = iter->count;
2458 if (folio_pos(folio) > pos) {
2459 count -= folio_pos(folio) - pos;
2460 pos = 0;
2461 } else {
2462 pos -= folio_pos(folio);
2463 }
2464
2465 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2466}
2467
2468static int filemap_update_page(struct kiocb *iocb,
2469 struct address_space *mapping, struct iov_iter *iter,
2470 struct folio *folio)
2471{
2472 int error;
2473
2474 if (iocb->ki_flags & IOCB_NOWAIT) {
2475 if (!filemap_invalidate_trylock_shared(mapping))
2476 return -EAGAIN;
2477 } else {
2478 filemap_invalidate_lock_shared(mapping);
2479 }
2480
2481 if (!folio_trylock(folio)) {
2482 error = -EAGAIN;
2483 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2484 goto unlock_mapping;
2485 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2486 filemap_invalidate_unlock_shared(mapping);
2487 /*
2488 * This is where we usually end up waiting for a
2489 * previously submitted readahead to finish.
2490 */
2491 folio_put_wait_locked(folio, TASK_KILLABLE);
2492 return AOP_TRUNCATED_PAGE;
2493 }
2494 error = __folio_lock_async(folio, iocb->ki_waitq);
2495 if (error)
2496 goto unlock_mapping;
2497 }
2498
2499 error = AOP_TRUNCATED_PAGE;
2500 if (!folio->mapping)
2501 goto unlock;
2502
2503 error = 0;
2504 if (filemap_range_uptodate(mapping, iocb->ki_pos, iter, folio))
2505 goto unlock;
2506
2507 error = -EAGAIN;
2508 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2509 goto unlock;
2510
2511 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2512 folio);
2513 goto unlock_mapping;
2514unlock:
2515 folio_unlock(folio);
2516unlock_mapping:
2517 filemap_invalidate_unlock_shared(mapping);
2518 if (error == AOP_TRUNCATED_PAGE)
2519 folio_put(folio);
2520 return error;
2521}
2522
2523static int filemap_create_folio(struct file *file,
2524 struct address_space *mapping, pgoff_t index,
2525 struct folio_batch *fbatch)
2526{
2527 struct folio *folio;
2528 int error;
2529
2530 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), 0);
2531 if (!folio)
2532 return -ENOMEM;
2533
2534 /*
2535 * Protect against truncate / hole punch. Grabbing invalidate_lock
2536 * here assures we cannot instantiate and bring uptodate new
2537 * pagecache folios after evicting page cache during truncate
2538 * and before actually freeing blocks. Note that we could
2539 * release invalidate_lock after inserting the folio into
2540 * the page cache as the locked folio would then be enough to
2541 * synchronize with hole punching. But there are code paths
2542 * such as filemap_update_page() filling in partially uptodate
2543 * pages or ->readahead() that need to hold invalidate_lock
2544 * while mapping blocks for IO so let's hold the lock here as
2545 * well to keep locking rules simple.
2546 */
2547 filemap_invalidate_lock_shared(mapping);
2548 error = filemap_add_folio(mapping, folio, index,
2549 mapping_gfp_constraint(mapping, GFP_KERNEL));
2550 if (error == -EEXIST)
2551 error = AOP_TRUNCATED_PAGE;
2552 if (error)
2553 goto error;
2554
2555 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
2556 if (error)
2557 goto error;
2558
2559 filemap_invalidate_unlock_shared(mapping);
2560 folio_batch_add(fbatch, folio);
2561 return 0;
2562error:
2563 filemap_invalidate_unlock_shared(mapping);
2564 folio_put(folio);
2565 return error;
2566}
2567
2568static int filemap_readahead(struct kiocb *iocb, struct file *file,
2569 struct address_space *mapping, struct folio *folio,
2570 pgoff_t last_index)
2571{
2572 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2573
2574 if (iocb->ki_flags & IOCB_NOIO)
2575 return -EAGAIN;
2576 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2577 return 0;
2578}
2579
2580static int filemap_get_pages(struct kiocb *iocb, struct iov_iter *iter,
2581 struct folio_batch *fbatch)
2582{
2583 struct file *filp = iocb->ki_filp;
2584 struct address_space *mapping = filp->f_mapping;
2585 struct file_ra_state *ra = &filp->f_ra;
2586 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2587 pgoff_t last_index;
2588 struct folio *folio;
2589 int err = 0;
2590
2591 /* "last_index" is the index of the page beyond the end of the read */
2592 last_index = DIV_ROUND_UP(iocb->ki_pos + iter->count, PAGE_SIZE);
2593retry:
2594 if (fatal_signal_pending(current))
2595 return -EINTR;
2596
2597 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2598 if (!folio_batch_count(fbatch)) {
2599 if (iocb->ki_flags & IOCB_NOIO)
2600 return -EAGAIN;
2601 page_cache_sync_readahead(mapping, ra, filp, index,
2602 last_index - index);
2603 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2604 }
2605 if (!folio_batch_count(fbatch)) {
2606 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2607 return -EAGAIN;
2608 err = filemap_create_folio(filp, mapping,
2609 iocb->ki_pos >> PAGE_SHIFT, fbatch);
2610 if (err == AOP_TRUNCATED_PAGE)
2611 goto retry;
2612 return err;
2613 }
2614
2615 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2616 if (folio_test_readahead(folio)) {
2617 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2618 if (err)
2619 goto err;
2620 }
2621 if (!folio_test_uptodate(folio)) {
2622 if ((iocb->ki_flags & IOCB_WAITQ) &&
2623 folio_batch_count(fbatch) > 1)
2624 iocb->ki_flags |= IOCB_NOWAIT;
2625 err = filemap_update_page(iocb, mapping, iter, folio);
2626 if (err)
2627 goto err;
2628 }
2629
2630 return 0;
2631err:
2632 if (err < 0)
2633 folio_put(folio);
2634 if (likely(--fbatch->nr))
2635 return 0;
2636 if (err == AOP_TRUNCATED_PAGE)
2637 goto retry;
2638 return err;
2639}
2640
2641static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2642{
2643 unsigned int shift = folio_shift(folio);
2644
2645 return (pos1 >> shift == pos2 >> shift);
2646}
2647
2648/**
2649 * filemap_read - Read data from the page cache.
2650 * @iocb: The iocb to read.
2651 * @iter: Destination for the data.
2652 * @already_read: Number of bytes already read by the caller.
2653 *
2654 * Copies data from the page cache. If the data is not currently present,
2655 * uses the readahead and read_folio address_space operations to fetch it.
2656 *
2657 * Return: Total number of bytes copied, including those already read by
2658 * the caller. If an error happens before any bytes are copied, returns
2659 * a negative error number.
2660 */
2661ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2662 ssize_t already_read)
2663{
2664 struct file *filp = iocb->ki_filp;
2665 struct file_ra_state *ra = &filp->f_ra;
2666 struct address_space *mapping = filp->f_mapping;
2667 struct inode *inode = mapping->host;
2668 struct folio_batch fbatch;
2669 int i, error = 0;
2670 bool writably_mapped;
2671 loff_t isize, end_offset;
2672
2673 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2674 return 0;
2675 if (unlikely(!iov_iter_count(iter)))
2676 return 0;
2677
2678 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2679 folio_batch_init(&fbatch);
2680
2681 do {
2682 cond_resched();
2683
2684 /*
2685 * If we've already successfully copied some data, then we
2686 * can no longer safely return -EIOCBQUEUED. Hence mark
2687 * an async read NOWAIT at that point.
2688 */
2689 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2690 iocb->ki_flags |= IOCB_NOWAIT;
2691
2692 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2693 break;
2694
2695 error = filemap_get_pages(iocb, iter, &fbatch);
2696 if (error < 0)
2697 break;
2698
2699 /*
2700 * i_size must be checked after we know the pages are Uptodate.
2701 *
2702 * Checking i_size after the check allows us to calculate
2703 * the correct value for "nr", which means the zero-filled
2704 * part of the page is not copied back to userspace (unless
2705 * another truncate extends the file - this is desired though).
2706 */
2707 isize = i_size_read(inode);
2708 if (unlikely(iocb->ki_pos >= isize))
2709 goto put_folios;
2710 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2711
2712 /*
2713 * Once we start copying data, we don't want to be touching any
2714 * cachelines that might be contended:
2715 */
2716 writably_mapped = mapping_writably_mapped(mapping);
2717
2718 /*
2719 * When a read accesses the same folio several times, only
2720 * mark it as accessed the first time.
2721 */
2722 if (!pos_same_folio(iocb->ki_pos, ra->prev_pos - 1,
2723 fbatch.folios[0]))
2724 folio_mark_accessed(fbatch.folios[0]);
2725
2726 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2727 struct folio *folio = fbatch.folios[i];
2728 size_t fsize = folio_size(folio);
2729 size_t offset = iocb->ki_pos & (fsize - 1);
2730 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2731 fsize - offset);
2732 size_t copied;
2733
2734 if (end_offset < folio_pos(folio))
2735 break;
2736 if (i > 0)
2737 folio_mark_accessed(folio);
2738 /*
2739 * If users can be writing to this folio using arbitrary
2740 * virtual addresses, take care of potential aliasing
2741 * before reading the folio on the kernel side.
2742 */
2743 if (writably_mapped)
2744 flush_dcache_folio(folio);
2745
2746 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2747
2748 already_read += copied;
2749 iocb->ki_pos += copied;
2750 ra->prev_pos = iocb->ki_pos;
2751
2752 if (copied < bytes) {
2753 error = -EFAULT;
2754 break;
2755 }
2756 }
2757put_folios:
2758 for (i = 0; i < folio_batch_count(&fbatch); i++)
2759 folio_put(fbatch.folios[i]);
2760 folio_batch_init(&fbatch);
2761 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2762
2763 file_accessed(filp);
2764
2765 return already_read ? already_read : error;
2766}
2767EXPORT_SYMBOL_GPL(filemap_read);
2768
2769/**
2770 * generic_file_read_iter - generic filesystem read routine
2771 * @iocb: kernel I/O control block
2772 * @iter: destination for the data read
2773 *
2774 * This is the "read_iter()" routine for all filesystems
2775 * that can use the page cache directly.
2776 *
2777 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2778 * be returned when no data can be read without waiting for I/O requests
2779 * to complete; it doesn't prevent readahead.
2780 *
2781 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2782 * requests shall be made for the read or for readahead. When no data
2783 * can be read, -EAGAIN shall be returned. When readahead would be
2784 * triggered, a partial, possibly empty read shall be returned.
2785 *
2786 * Return:
2787 * * number of bytes copied, even for partial reads
2788 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2789 */
2790ssize_t
2791generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2792{
2793 size_t count = iov_iter_count(iter);
2794 ssize_t retval = 0;
2795
2796 if (!count)
2797 return 0; /* skip atime */
2798
2799 if (iocb->ki_flags & IOCB_DIRECT) {
2800 struct file *file = iocb->ki_filp;
2801 struct address_space *mapping = file->f_mapping;
2802 struct inode *inode = mapping->host;
2803
2804 if (iocb->ki_flags & IOCB_NOWAIT) {
2805 if (filemap_range_needs_writeback(mapping, iocb->ki_pos,
2806 iocb->ki_pos + count - 1))
2807 return -EAGAIN;
2808 } else {
2809 retval = filemap_write_and_wait_range(mapping,
2810 iocb->ki_pos,
2811 iocb->ki_pos + count - 1);
2812 if (retval < 0)
2813 return retval;
2814 }
2815
2816 file_accessed(file);
2817
2818 retval = mapping->a_ops->direct_IO(iocb, iter);
2819 if (retval >= 0) {
2820 iocb->ki_pos += retval;
2821 count -= retval;
2822 }
2823 if (retval != -EIOCBQUEUED)
2824 iov_iter_revert(iter, count - iov_iter_count(iter));
2825
2826 /*
2827 * Btrfs can have a short DIO read if we encounter
2828 * compressed extents, so if there was an error, or if
2829 * we've already read everything we wanted to, or if
2830 * there was a short read because we hit EOF, go ahead
2831 * and return. Otherwise fallthrough to buffered io for
2832 * the rest of the read. Buffered reads will not work for
2833 * DAX files, so don't bother trying.
2834 */
2835 if (retval < 0 || !count || IS_DAX(inode))
2836 return retval;
2837 if (iocb->ki_pos >= i_size_read(inode))
2838 return retval;
2839 }
2840
2841 return filemap_read(iocb, iter, retval);
2842}
2843EXPORT_SYMBOL(generic_file_read_iter);
2844
2845static inline loff_t folio_seek_hole_data(struct xa_state *xas,
2846 struct address_space *mapping, struct folio *folio,
2847 loff_t start, loff_t end, bool seek_data)
2848{
2849 const struct address_space_operations *ops = mapping->a_ops;
2850 size_t offset, bsz = i_blocksize(mapping->host);
2851
2852 if (xa_is_value(folio) || folio_test_uptodate(folio))
2853 return seek_data ? start : end;
2854 if (!ops->is_partially_uptodate)
2855 return seek_data ? end : start;
2856
2857 xas_pause(xas);
2858 rcu_read_unlock();
2859 folio_lock(folio);
2860 if (unlikely(folio->mapping != mapping))
2861 goto unlock;
2862
2863 offset = offset_in_folio(folio, start) & ~(bsz - 1);
2864
2865 do {
2866 if (ops->is_partially_uptodate(folio, offset, bsz) ==
2867 seek_data)
2868 break;
2869 start = (start + bsz) & ~(bsz - 1);
2870 offset += bsz;
2871 } while (offset < folio_size(folio));
2872unlock:
2873 folio_unlock(folio);
2874 rcu_read_lock();
2875 return start;
2876}
2877
2878static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
2879{
2880 if (xa_is_value(folio))
2881 return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index);
2882 return folio_size(folio);
2883}
2884
2885/**
2886 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
2887 * @mapping: Address space to search.
2888 * @start: First byte to consider.
2889 * @end: Limit of search (exclusive).
2890 * @whence: Either SEEK_HOLE or SEEK_DATA.
2891 *
2892 * If the page cache knows which blocks contain holes and which blocks
2893 * contain data, your filesystem can use this function to implement
2894 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
2895 * entirely memory-based such as tmpfs, and filesystems which support
2896 * unwritten extents.
2897 *
2898 * Return: The requested offset on success, or -ENXIO if @whence specifies
2899 * SEEK_DATA and there is no data after @start. There is an implicit hole
2900 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
2901 * and @end contain data.
2902 */
2903loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
2904 loff_t end, int whence)
2905{
2906 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
2907 pgoff_t max = (end - 1) >> PAGE_SHIFT;
2908 bool seek_data = (whence == SEEK_DATA);
2909 struct folio *folio;
2910
2911 if (end <= start)
2912 return -ENXIO;
2913
2914 rcu_read_lock();
2915 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
2916 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
2917 size_t seek_size;
2918
2919 if (start < pos) {
2920 if (!seek_data)
2921 goto unlock;
2922 start = pos;
2923 }
2924
2925 seek_size = seek_folio_size(&xas, folio);
2926 pos = round_up((u64)pos + 1, seek_size);
2927 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
2928 seek_data);
2929 if (start < pos)
2930 goto unlock;
2931 if (start >= end)
2932 break;
2933 if (seek_size > PAGE_SIZE)
2934 xas_set(&xas, pos >> PAGE_SHIFT);
2935 if (!xa_is_value(folio))
2936 folio_put(folio);
2937 }
2938 if (seek_data)
2939 start = -ENXIO;
2940unlock:
2941 rcu_read_unlock();
2942 if (folio && !xa_is_value(folio))
2943 folio_put(folio);
2944 if (start > end)
2945 return end;
2946 return start;
2947}
2948
2949#ifdef CONFIG_MMU
2950#define MMAP_LOTSAMISS (100)
2951/*
2952 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
2953 * @vmf - the vm_fault for this fault.
2954 * @folio - the folio to lock.
2955 * @fpin - the pointer to the file we may pin (or is already pinned).
2956 *
2957 * This works similar to lock_folio_or_retry in that it can drop the
2958 * mmap_lock. It differs in that it actually returns the folio locked
2959 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
2960 * to drop the mmap_lock then fpin will point to the pinned file and
2961 * needs to be fput()'ed at a later point.
2962 */
2963static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
2964 struct file **fpin)
2965{
2966 if (folio_trylock(folio))
2967 return 1;
2968
2969 /*
2970 * NOTE! This will make us return with VM_FAULT_RETRY, but with
2971 * the mmap_lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
2972 * is supposed to work. We have way too many special cases..
2973 */
2974 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
2975 return 0;
2976
2977 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
2978 if (vmf->flags & FAULT_FLAG_KILLABLE) {
2979 if (__folio_lock_killable(folio)) {
2980 /*
2981 * We didn't have the right flags to drop the mmap_lock,
2982 * but all fault_handlers only check for fatal signals
2983 * if we return VM_FAULT_RETRY, so we need to drop the
2984 * mmap_lock here and return 0 if we don't have a fpin.
2985 */
2986 if (*fpin == NULL)
2987 mmap_read_unlock(vmf->vma->vm_mm);
2988 return 0;
2989 }
2990 } else
2991 __folio_lock(folio);
2992
2993 return 1;
2994}
2995
2996/*
2997 * Synchronous readahead happens when we don't even find a page in the page
2998 * cache at all. We don't want to perform IO under the mmap sem, so if we have
2999 * to drop the mmap sem we return the file that was pinned in order for us to do
3000 * that. If we didn't pin a file then we return NULL. The file that is
3001 * returned needs to be fput()'ed when we're done with it.
3002 */
3003static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3004{
3005 struct file *file = vmf->vma->vm_file;
3006 struct file_ra_state *ra = &file->f_ra;
3007 struct address_space *mapping = file->f_mapping;
3008 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3009 struct file *fpin = NULL;
3010 unsigned long vm_flags = vmf->vma->vm_flags;
3011 unsigned int mmap_miss;
3012
3013#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3014 /* Use the readahead code, even if readahead is disabled */
3015 if (vm_flags & VM_HUGEPAGE) {
3016 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3017 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3018 ra->size = HPAGE_PMD_NR;
3019 /*
3020 * Fetch two PMD folios, so we get the chance to actually
3021 * readahead, unless we've been told not to.
3022 */
3023 if (!(vm_flags & VM_RAND_READ))
3024 ra->size *= 2;
3025 ra->async_size = HPAGE_PMD_NR;
3026 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3027 return fpin;
3028 }
3029#endif
3030
3031 /* If we don't want any read-ahead, don't bother */
3032 if (vm_flags & VM_RAND_READ)
3033 return fpin;
3034 if (!ra->ra_pages)
3035 return fpin;
3036
3037 if (vm_flags & VM_SEQ_READ) {
3038 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3039 page_cache_sync_ra(&ractl, ra->ra_pages);
3040 return fpin;
3041 }
3042
3043 /* Avoid banging the cache line if not needed */
3044 mmap_miss = READ_ONCE(ra->mmap_miss);
3045 if (mmap_miss < MMAP_LOTSAMISS * 10)
3046 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3047
3048 /*
3049 * Do we miss much more than hit in this file? If so,
3050 * stop bothering with read-ahead. It will only hurt.
3051 */
3052 if (mmap_miss > MMAP_LOTSAMISS)
3053 return fpin;
3054
3055 /*
3056 * mmap read-around
3057 */
3058 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3059 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3060 ra->size = ra->ra_pages;
3061 ra->async_size = ra->ra_pages / 4;
3062 ractl._index = ra->start;
3063 page_cache_ra_order(&ractl, ra, 0);
3064 return fpin;
3065}
3066
3067/*
3068 * Asynchronous readahead happens when we find the page and PG_readahead,
3069 * so we want to possibly extend the readahead further. We return the file that
3070 * was pinned if we have to drop the mmap_lock in order to do IO.
3071 */
3072static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3073 struct folio *folio)
3074{
3075 struct file *file = vmf->vma->vm_file;
3076 struct file_ra_state *ra = &file->f_ra;
3077 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3078 struct file *fpin = NULL;
3079 unsigned int mmap_miss;
3080
3081 /* If we don't want any read-ahead, don't bother */
3082 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3083 return fpin;
3084
3085 mmap_miss = READ_ONCE(ra->mmap_miss);
3086 if (mmap_miss)
3087 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3088
3089 if (folio_test_readahead(folio)) {
3090 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3091 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3092 }
3093 return fpin;
3094}
3095
3096/**
3097 * filemap_fault - read in file data for page fault handling
3098 * @vmf: struct vm_fault containing details of the fault
3099 *
3100 * filemap_fault() is invoked via the vma operations vector for a
3101 * mapped memory region to read in file data during a page fault.
3102 *
3103 * The goto's are kind of ugly, but this streamlines the normal case of having
3104 * it in the page cache, and handles the special cases reasonably without
3105 * having a lot of duplicated code.
3106 *
3107 * vma->vm_mm->mmap_lock must be held on entry.
3108 *
3109 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3110 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3111 *
3112 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3113 * has not been released.
3114 *
3115 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3116 *
3117 * Return: bitwise-OR of %VM_FAULT_ codes.
3118 */
3119vm_fault_t filemap_fault(struct vm_fault *vmf)
3120{
3121 int error;
3122 struct file *file = vmf->vma->vm_file;
3123 struct file *fpin = NULL;
3124 struct address_space *mapping = file->f_mapping;
3125 struct inode *inode = mapping->host;
3126 pgoff_t max_idx, index = vmf->pgoff;
3127 struct folio *folio;
3128 vm_fault_t ret = 0;
3129 bool mapping_locked = false;
3130
3131 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3132 if (unlikely(index >= max_idx))
3133 return VM_FAULT_SIGBUS;
3134
3135 /*
3136 * Do we have something in the page cache already?
3137 */
3138 folio = filemap_get_folio(mapping, index);
3139 if (likely(folio)) {
3140 /*
3141 * We found the page, so try async readahead before waiting for
3142 * the lock.
3143 */
3144 if (!(vmf->flags & FAULT_FLAG_TRIED))
3145 fpin = do_async_mmap_readahead(vmf, folio);
3146 if (unlikely(!folio_test_uptodate(folio))) {
3147 filemap_invalidate_lock_shared(mapping);
3148 mapping_locked = true;
3149 }
3150 } else {
3151 /* No page in the page cache at all */
3152 count_vm_event(PGMAJFAULT);
3153 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3154 ret = VM_FAULT_MAJOR;
3155 fpin = do_sync_mmap_readahead(vmf);
3156retry_find:
3157 /*
3158 * See comment in filemap_create_folio() why we need
3159 * invalidate_lock
3160 */
3161 if (!mapping_locked) {
3162 filemap_invalidate_lock_shared(mapping);
3163 mapping_locked = true;
3164 }
3165 folio = __filemap_get_folio(mapping, index,
3166 FGP_CREAT|FGP_FOR_MMAP,
3167 vmf->gfp_mask);
3168 if (!folio) {
3169 if (fpin)
3170 goto out_retry;
3171 filemap_invalidate_unlock_shared(mapping);
3172 return VM_FAULT_OOM;
3173 }
3174 }
3175
3176 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3177 goto out_retry;
3178
3179 /* Did it get truncated? */
3180 if (unlikely(folio->mapping != mapping)) {
3181 folio_unlock(folio);
3182 folio_put(folio);
3183 goto retry_find;
3184 }
3185 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3186
3187 /*
3188 * We have a locked page in the page cache, now we need to check
3189 * that it's up-to-date. If not, it is going to be due to an error.
3190 */
3191 if (unlikely(!folio_test_uptodate(folio))) {
3192 /*
3193 * The page was in cache and uptodate and now it is not.
3194 * Strange but possible since we didn't hold the page lock all
3195 * the time. Let's drop everything get the invalidate lock and
3196 * try again.
3197 */
3198 if (!mapping_locked) {
3199 folio_unlock(folio);
3200 folio_put(folio);
3201 goto retry_find;
3202 }
3203 goto page_not_uptodate;
3204 }
3205
3206 /*
3207 * We've made it this far and we had to drop our mmap_lock, now is the
3208 * time to return to the upper layer and have it re-find the vma and
3209 * redo the fault.
3210 */
3211 if (fpin) {
3212 folio_unlock(folio);
3213 goto out_retry;
3214 }
3215 if (mapping_locked)
3216 filemap_invalidate_unlock_shared(mapping);
3217
3218 /*
3219 * Found the page and have a reference on it.
3220 * We must recheck i_size under page lock.
3221 */
3222 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3223 if (unlikely(index >= max_idx)) {
3224 folio_unlock(folio);
3225 folio_put(folio);
3226 return VM_FAULT_SIGBUS;
3227 }
3228
3229 vmf->page = folio_file_page(folio, index);
3230 return ret | VM_FAULT_LOCKED;
3231
3232page_not_uptodate:
3233 /*
3234 * Umm, take care of errors if the page isn't up-to-date.
3235 * Try to re-read it _once_. We do this synchronously,
3236 * because there really aren't any performance issues here
3237 * and we need to check for errors.
3238 */
3239 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3240 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3241 if (fpin)
3242 goto out_retry;
3243 folio_put(folio);
3244
3245 if (!error || error == AOP_TRUNCATED_PAGE)
3246 goto retry_find;
3247 filemap_invalidate_unlock_shared(mapping);
3248
3249 return VM_FAULT_SIGBUS;
3250
3251out_retry:
3252 /*
3253 * We dropped the mmap_lock, we need to return to the fault handler to
3254 * re-find the vma and come back and find our hopefully still populated
3255 * page.
3256 */
3257 if (folio)
3258 folio_put(folio);
3259 if (mapping_locked)
3260 filemap_invalidate_unlock_shared(mapping);
3261 if (fpin)
3262 fput(fpin);
3263 return ret | VM_FAULT_RETRY;
3264}
3265EXPORT_SYMBOL(filemap_fault);
3266
3267static bool filemap_map_pmd(struct vm_fault *vmf, struct page *page)
3268{
3269 struct mm_struct *mm = vmf->vma->vm_mm;
3270
3271 /* Huge page is mapped? No need to proceed. */
3272 if (pmd_trans_huge(*vmf->pmd)) {
3273 unlock_page(page);
3274 put_page(page);
3275 return true;
3276 }
3277
3278 if (pmd_none(*vmf->pmd) && PageTransHuge(page)) {
3279 vm_fault_t ret = do_set_pmd(vmf, page);
3280 if (!ret) {
3281 /* The page is mapped successfully, reference consumed. */
3282 unlock_page(page);
3283 return true;
3284 }
3285 }
3286
3287 if (pmd_none(*vmf->pmd))
3288 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3289
3290 /* See comment in handle_pte_fault() */
3291 if (pmd_devmap_trans_unstable(vmf->pmd)) {
3292 unlock_page(page);
3293 put_page(page);
3294 return true;
3295 }
3296
3297 return false;
3298}
3299
3300static struct folio *next_uptodate_page(struct folio *folio,
3301 struct address_space *mapping,
3302 struct xa_state *xas, pgoff_t end_pgoff)
3303{
3304 unsigned long max_idx;
3305
3306 do {
3307 if (!folio)
3308 return NULL;
3309 if (xas_retry(xas, folio))
3310 continue;
3311 if (xa_is_value(folio))
3312 continue;
3313 if (folio_test_locked(folio))
3314 continue;
3315 if (!folio_try_get_rcu(folio))
3316 continue;
3317 /* Has the page moved or been split? */
3318 if (unlikely(folio != xas_reload(xas)))
3319 goto skip;
3320 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3321 goto skip;
3322 if (!folio_trylock(folio))
3323 goto skip;
3324 if (folio->mapping != mapping)
3325 goto unlock;
3326 if (!folio_test_uptodate(folio))
3327 goto unlock;
3328 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3329 if (xas->xa_index >= max_idx)
3330 goto unlock;
3331 return folio;
3332unlock:
3333 folio_unlock(folio);
3334skip:
3335 folio_put(folio);
3336 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3337
3338 return NULL;
3339}
3340
3341static inline struct folio *first_map_page(struct address_space *mapping,
3342 struct xa_state *xas,
3343 pgoff_t end_pgoff)
3344{
3345 return next_uptodate_page(xas_find(xas, end_pgoff),
3346 mapping, xas, end_pgoff);
3347}
3348
3349static inline struct folio *next_map_page(struct address_space *mapping,
3350 struct xa_state *xas,
3351 pgoff_t end_pgoff)
3352{
3353 return next_uptodate_page(xas_next_entry(xas, end_pgoff),
3354 mapping, xas, end_pgoff);
3355}
3356
3357vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3358 pgoff_t start_pgoff, pgoff_t end_pgoff)
3359{
3360 struct vm_area_struct *vma = vmf->vma;
3361 struct file *file = vma->vm_file;
3362 struct address_space *mapping = file->f_mapping;
3363 pgoff_t last_pgoff = start_pgoff;
3364 unsigned long addr;
3365 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3366 struct folio *folio;
3367 struct page *page;
3368 unsigned int mmap_miss = READ_ONCE(file->f_ra.mmap_miss);
3369 vm_fault_t ret = 0;
3370
3371 rcu_read_lock();
3372 folio = first_map_page(mapping, &xas, end_pgoff);
3373 if (!folio)
3374 goto out;
3375
3376 if (filemap_map_pmd(vmf, &folio->page)) {
3377 ret = VM_FAULT_NOPAGE;
3378 goto out;
3379 }
3380
3381 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3382 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3383 do {
3384again:
3385 page = folio_file_page(folio, xas.xa_index);
3386 if (PageHWPoison(page))
3387 goto unlock;
3388
3389 if (mmap_miss > 0)
3390 mmap_miss--;
3391
3392 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3393 vmf->pte += xas.xa_index - last_pgoff;
3394 last_pgoff = xas.xa_index;
3395
3396 /*
3397 * NOTE: If there're PTE markers, we'll leave them to be
3398 * handled in the specific fault path, and it'll prohibit the
3399 * fault-around logic.
3400 */
3401 if (!pte_none(*vmf->pte))
3402 goto unlock;
3403
3404 /* We're about to handle the fault */
3405 if (vmf->address == addr)
3406 ret = VM_FAULT_NOPAGE;
3407
3408 do_set_pte(vmf, page, addr);
3409 /* no need to invalidate: a not-present page won't be cached */
3410 update_mmu_cache(vma, addr, vmf->pte);
3411 if (folio_more_pages(folio, xas.xa_index, end_pgoff)) {
3412 xas.xa_index++;
3413 folio_ref_inc(folio);
3414 goto again;
3415 }
3416 folio_unlock(folio);
3417 continue;
3418unlock:
3419 if (folio_more_pages(folio, xas.xa_index, end_pgoff)) {
3420 xas.xa_index++;
3421 goto again;
3422 }
3423 folio_unlock(folio);
3424 folio_put(folio);
3425 } while ((folio = next_map_page(mapping, &xas, end_pgoff)) != NULL);
3426 pte_unmap_unlock(vmf->pte, vmf->ptl);
3427out:
3428 rcu_read_unlock();
3429 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss);
3430 return ret;
3431}
3432EXPORT_SYMBOL(filemap_map_pages);
3433
3434vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3435{
3436 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3437 struct folio *folio = page_folio(vmf->page);
3438 vm_fault_t ret = VM_FAULT_LOCKED;
3439
3440 sb_start_pagefault(mapping->host->i_sb);
3441 file_update_time(vmf->vma->vm_file);
3442 folio_lock(folio);
3443 if (folio->mapping != mapping) {
3444 folio_unlock(folio);
3445 ret = VM_FAULT_NOPAGE;
3446 goto out;
3447 }
3448 /*
3449 * We mark the folio dirty already here so that when freeze is in
3450 * progress, we are guaranteed that writeback during freezing will
3451 * see the dirty folio and writeprotect it again.
3452 */
3453 folio_mark_dirty(folio);
3454 folio_wait_stable(folio);
3455out:
3456 sb_end_pagefault(mapping->host->i_sb);
3457 return ret;
3458}
3459
3460const struct vm_operations_struct generic_file_vm_ops = {
3461 .fault = filemap_fault,
3462 .map_pages = filemap_map_pages,
3463 .page_mkwrite = filemap_page_mkwrite,
3464};
3465
3466/* This is used for a general mmap of a disk file */
3467
3468int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3469{
3470 struct address_space *mapping = file->f_mapping;
3471
3472 if (!mapping->a_ops->read_folio)
3473 return -ENOEXEC;
3474 file_accessed(file);
3475 vma->vm_ops = &generic_file_vm_ops;
3476 return 0;
3477}
3478
3479/*
3480 * This is for filesystems which do not implement ->writepage.
3481 */
3482int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3483{
3484 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
3485 return -EINVAL;
3486 return generic_file_mmap(file, vma);
3487}
3488#else
3489vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3490{
3491 return VM_FAULT_SIGBUS;
3492}
3493int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3494{
3495 return -ENOSYS;
3496}
3497int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3498{
3499 return -ENOSYS;
3500}
3501#endif /* CONFIG_MMU */
3502
3503EXPORT_SYMBOL(filemap_page_mkwrite);
3504EXPORT_SYMBOL(generic_file_mmap);
3505EXPORT_SYMBOL(generic_file_readonly_mmap);
3506
3507static struct folio *do_read_cache_folio(struct address_space *mapping,
3508 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3509{
3510 struct folio *folio;
3511 int err;
3512
3513 if (!filler)
3514 filler = mapping->a_ops->read_folio;
3515repeat:
3516 folio = filemap_get_folio(mapping, index);
3517 if (!folio) {
3518 folio = filemap_alloc_folio(gfp, 0);
3519 if (!folio)
3520 return ERR_PTR(-ENOMEM);
3521 err = filemap_add_folio(mapping, folio, index, gfp);
3522 if (unlikely(err)) {
3523 folio_put(folio);
3524 if (err == -EEXIST)
3525 goto repeat;
3526 /* Presumably ENOMEM for xarray node */
3527 return ERR_PTR(err);
3528 }
3529
3530 goto filler;
3531 }
3532 if (folio_test_uptodate(folio))
3533 goto out;
3534
3535 if (!folio_trylock(folio)) {
3536 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3537 goto repeat;
3538 }
3539
3540 /* Folio was truncated from mapping */
3541 if (!folio->mapping) {
3542 folio_unlock(folio);
3543 folio_put(folio);
3544 goto repeat;
3545 }
3546
3547 /* Someone else locked and filled the page in a very small window */
3548 if (folio_test_uptodate(folio)) {
3549 folio_unlock(folio);
3550 goto out;
3551 }
3552
3553filler:
3554 err = filemap_read_folio(file, filler, folio);
3555 if (err) {
3556 folio_put(folio);
3557 if (err == AOP_TRUNCATED_PAGE)
3558 goto repeat;
3559 return ERR_PTR(err);
3560 }
3561
3562out:
3563 folio_mark_accessed(folio);
3564 return folio;
3565}
3566
3567/**
3568 * read_cache_folio - Read into page cache, fill it if needed.
3569 * @mapping: The address_space to read from.
3570 * @index: The index to read.
3571 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3572 * @file: Passed to filler function, may be NULL if not required.
3573 *
3574 * Read one page into the page cache. If it succeeds, the folio returned
3575 * will contain @index, but it may not be the first page of the folio.
3576 *
3577 * If the filler function returns an error, it will be returned to the
3578 * caller.
3579 *
3580 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3581 * Return: An uptodate folio on success, ERR_PTR() on failure.
3582 */
3583struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3584 filler_t filler, struct file *file)
3585{
3586 return do_read_cache_folio(mapping, index, filler, file,
3587 mapping_gfp_mask(mapping));
3588}
3589EXPORT_SYMBOL(read_cache_folio);
3590
3591static struct page *do_read_cache_page(struct address_space *mapping,
3592 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3593{
3594 struct folio *folio;
3595
3596 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3597 if (IS_ERR(folio))
3598 return &folio->page;
3599 return folio_file_page(folio, index);
3600}
3601
3602struct page *read_cache_page(struct address_space *mapping,
3603 pgoff_t index, filler_t *filler, struct file *file)
3604{
3605 return do_read_cache_page(mapping, index, filler, file,
3606 mapping_gfp_mask(mapping));
3607}
3608EXPORT_SYMBOL(read_cache_page);
3609
3610/**
3611 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3612 * @mapping: the page's address_space
3613 * @index: the page index
3614 * @gfp: the page allocator flags to use if allocating
3615 *
3616 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3617 * any new page allocations done using the specified allocation flags.
3618 *
3619 * If the page does not get brought uptodate, return -EIO.
3620 *
3621 * The function expects mapping->invalidate_lock to be already held.
3622 *
3623 * Return: up to date page on success, ERR_PTR() on failure.
3624 */
3625struct page *read_cache_page_gfp(struct address_space *mapping,
3626 pgoff_t index,
3627 gfp_t gfp)
3628{
3629 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3630}
3631EXPORT_SYMBOL(read_cache_page_gfp);
3632
3633/*
3634 * Warn about a page cache invalidation failure during a direct I/O write.
3635 */
3636void dio_warn_stale_pagecache(struct file *filp)
3637{
3638 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3639 char pathname[128];
3640 char *path;
3641
3642 errseq_set(&filp->f_mapping->wb_err, -EIO);
3643 if (__ratelimit(&_rs)) {
3644 path = file_path(filp, pathname, sizeof(pathname));
3645 if (IS_ERR(path))
3646 path = "(unknown)";
3647 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3648 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3649 current->comm);
3650 }
3651}
3652
3653ssize_t
3654generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3655{
3656 struct file *file = iocb->ki_filp;
3657 struct address_space *mapping = file->f_mapping;
3658 struct inode *inode = mapping->host;
3659 loff_t pos = iocb->ki_pos;
3660 ssize_t written;
3661 size_t write_len;
3662 pgoff_t end;
3663
3664 write_len = iov_iter_count(from);
3665 end = (pos + write_len - 1) >> PAGE_SHIFT;
3666
3667 if (iocb->ki_flags & IOCB_NOWAIT) {
3668 /* If there are pages to writeback, return */
3669 if (filemap_range_has_page(file->f_mapping, pos,
3670 pos + write_len - 1))
3671 return -EAGAIN;
3672 } else {
3673 written = filemap_write_and_wait_range(mapping, pos,
3674 pos + write_len - 1);
3675 if (written)
3676 goto out;
3677 }
3678
3679 /*
3680 * After a write we want buffered reads to be sure to go to disk to get
3681 * the new data. We invalidate clean cached page from the region we're
3682 * about to write. We do this *before* the write so that we can return
3683 * without clobbering -EIOCBQUEUED from ->direct_IO().
3684 */
3685 written = invalidate_inode_pages2_range(mapping,
3686 pos >> PAGE_SHIFT, end);
3687 /*
3688 * If a page can not be invalidated, return 0 to fall back
3689 * to buffered write.
3690 */
3691 if (written) {
3692 if (written == -EBUSY)
3693 return 0;
3694 goto out;
3695 }
3696
3697 written = mapping->a_ops->direct_IO(iocb, from);
3698
3699 /*
3700 * Finally, try again to invalidate clean pages which might have been
3701 * cached by non-direct readahead, or faulted in by get_user_pages()
3702 * if the source of the write was an mmap'ed region of the file
3703 * we're writing. Either one is a pretty crazy thing to do,
3704 * so we don't support it 100%. If this invalidation
3705 * fails, tough, the write still worked...
3706 *
3707 * Most of the time we do not need this since dio_complete() will do
3708 * the invalidation for us. However there are some file systems that
3709 * do not end up with dio_complete() being called, so let's not break
3710 * them by removing it completely.
3711 *
3712 * Noticeable example is a blkdev_direct_IO().
3713 *
3714 * Skip invalidation for async writes or if mapping has no pages.
3715 */
3716 if (written > 0 && mapping->nrpages &&
3717 invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT, end))
3718 dio_warn_stale_pagecache(file);
3719
3720 if (written > 0) {
3721 pos += written;
3722 write_len -= written;
3723 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3724 i_size_write(inode, pos);
3725 mark_inode_dirty(inode);
3726 }
3727 iocb->ki_pos = pos;
3728 }
3729 if (written != -EIOCBQUEUED)
3730 iov_iter_revert(from, write_len - iov_iter_count(from));
3731out:
3732 return written;
3733}
3734EXPORT_SYMBOL(generic_file_direct_write);
3735
3736ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
3737{
3738 struct file *file = iocb->ki_filp;
3739 loff_t pos = iocb->ki_pos;
3740 struct address_space *mapping = file->f_mapping;
3741 const struct address_space_operations *a_ops = mapping->a_ops;
3742 long status = 0;
3743 ssize_t written = 0;
3744
3745 do {
3746 struct page *page;
3747 unsigned long offset; /* Offset into pagecache page */
3748 unsigned long bytes; /* Bytes to write to page */
3749 size_t copied; /* Bytes copied from user */
3750 void *fsdata = NULL;
3751
3752 offset = (pos & (PAGE_SIZE - 1));
3753 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3754 iov_iter_count(i));
3755
3756again:
3757 /*
3758 * Bring in the user page that we will copy from _first_.
3759 * Otherwise there's a nasty deadlock on copying from the
3760 * same page as we're writing to, without it being marked
3761 * up-to-date.
3762 */
3763 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
3764 status = -EFAULT;
3765 break;
3766 }
3767
3768 if (fatal_signal_pending(current)) {
3769 status = -EINTR;
3770 break;
3771 }
3772
3773 status = a_ops->write_begin(file, mapping, pos, bytes,
3774 &page, &fsdata);
3775 if (unlikely(status < 0))
3776 break;
3777
3778 if (mapping_writably_mapped(mapping))
3779 flush_dcache_page(page);
3780
3781 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
3782 flush_dcache_page(page);
3783
3784 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3785 page, fsdata);
3786 if (unlikely(status != copied)) {
3787 iov_iter_revert(i, copied - max(status, 0L));
3788 if (unlikely(status < 0))
3789 break;
3790 }
3791 cond_resched();
3792
3793 if (unlikely(status == 0)) {
3794 /*
3795 * A short copy made ->write_end() reject the
3796 * thing entirely. Might be memory poisoning
3797 * halfway through, might be a race with munmap,
3798 * might be severe memory pressure.
3799 */
3800 if (copied)
3801 bytes = copied;
3802 goto again;
3803 }
3804 pos += status;
3805 written += status;
3806
3807 balance_dirty_pages_ratelimited(mapping);
3808 } while (iov_iter_count(i));
3809
3810 return written ? written : status;
3811}
3812EXPORT_SYMBOL(generic_perform_write);
3813
3814/**
3815 * __generic_file_write_iter - write data to a file
3816 * @iocb: IO state structure (file, offset, etc.)
3817 * @from: iov_iter with data to write
3818 *
3819 * This function does all the work needed for actually writing data to a
3820 * file. It does all basic checks, removes SUID from the file, updates
3821 * modification times and calls proper subroutines depending on whether we
3822 * do direct IO or a standard buffered write.
3823 *
3824 * It expects i_rwsem to be grabbed unless we work on a block device or similar
3825 * object which does not need locking at all.
3826 *
3827 * This function does *not* take care of syncing data in case of O_SYNC write.
3828 * A caller has to handle it. This is mainly due to the fact that we want to
3829 * avoid syncing under i_rwsem.
3830 *
3831 * Return:
3832 * * number of bytes written, even for truncated writes
3833 * * negative error code if no data has been written at all
3834 */
3835ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3836{
3837 struct file *file = iocb->ki_filp;
3838 struct address_space *mapping = file->f_mapping;
3839 struct inode *inode = mapping->host;
3840 ssize_t written = 0;
3841 ssize_t err;
3842 ssize_t status;
3843
3844 /* We can write back this queue in page reclaim */
3845 current->backing_dev_info = inode_to_bdi(inode);
3846 err = file_remove_privs(file);
3847 if (err)
3848 goto out;
3849
3850 err = file_update_time(file);
3851 if (err)
3852 goto out;
3853
3854 if (iocb->ki_flags & IOCB_DIRECT) {
3855 loff_t pos, endbyte;
3856
3857 written = generic_file_direct_write(iocb, from);
3858 /*
3859 * If the write stopped short of completing, fall back to
3860 * buffered writes. Some filesystems do this for writes to
3861 * holes, for example. For DAX files, a buffered write will
3862 * not succeed (even if it did, DAX does not handle dirty
3863 * page-cache pages correctly).
3864 */
3865 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3866 goto out;
3867
3868 pos = iocb->ki_pos;
3869 status = generic_perform_write(iocb, from);
3870 /*
3871 * If generic_perform_write() returned a synchronous error
3872 * then we want to return the number of bytes which were
3873 * direct-written, or the error code if that was zero. Note
3874 * that this differs from normal direct-io semantics, which
3875 * will return -EFOO even if some bytes were written.
3876 */
3877 if (unlikely(status < 0)) {
3878 err = status;
3879 goto out;
3880 }
3881 /*
3882 * We need to ensure that the page cache pages are written to
3883 * disk and invalidated to preserve the expected O_DIRECT
3884 * semantics.
3885 */
3886 endbyte = pos + status - 1;
3887 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3888 if (err == 0) {
3889 iocb->ki_pos = endbyte + 1;
3890 written += status;
3891 invalidate_mapping_pages(mapping,
3892 pos >> PAGE_SHIFT,
3893 endbyte >> PAGE_SHIFT);
3894 } else {
3895 /*
3896 * We don't know how much we wrote, so just return
3897 * the number of bytes which were direct-written
3898 */
3899 }
3900 } else {
3901 written = generic_perform_write(iocb, from);
3902 if (likely(written > 0))
3903 iocb->ki_pos += written;
3904 }
3905out:
3906 current->backing_dev_info = NULL;
3907 return written ? written : err;
3908}
3909EXPORT_SYMBOL(__generic_file_write_iter);
3910
3911/**
3912 * generic_file_write_iter - write data to a file
3913 * @iocb: IO state structure
3914 * @from: iov_iter with data to write
3915 *
3916 * This is a wrapper around __generic_file_write_iter() to be used by most
3917 * filesystems. It takes care of syncing the file in case of O_SYNC file
3918 * and acquires i_rwsem as needed.
3919 * Return:
3920 * * negative error code if no data has been written at all of
3921 * vfs_fsync_range() failed for a synchronous write
3922 * * number of bytes written, even for truncated writes
3923 */
3924ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3925{
3926 struct file *file = iocb->ki_filp;
3927 struct inode *inode = file->f_mapping->host;
3928 ssize_t ret;
3929
3930 inode_lock(inode);
3931 ret = generic_write_checks(iocb, from);
3932 if (ret > 0)
3933 ret = __generic_file_write_iter(iocb, from);
3934 inode_unlock(inode);
3935
3936 if (ret > 0)
3937 ret = generic_write_sync(iocb, ret);
3938 return ret;
3939}
3940EXPORT_SYMBOL(generic_file_write_iter);
3941
3942/**
3943 * filemap_release_folio() - Release fs-specific metadata on a folio.
3944 * @folio: The folio which the kernel is trying to free.
3945 * @gfp: Memory allocation flags (and I/O mode).
3946 *
3947 * The address_space is trying to release any data attached to a folio
3948 * (presumably at folio->private).
3949 *
3950 * This will also be called if the private_2 flag is set on a page,
3951 * indicating that the folio has other metadata associated with it.
3952 *
3953 * The @gfp argument specifies whether I/O may be performed to release
3954 * this page (__GFP_IO), and whether the call may block
3955 * (__GFP_RECLAIM & __GFP_FS).
3956 *
3957 * Return: %true if the release was successful, otherwise %false.
3958 */
3959bool filemap_release_folio(struct folio *folio, gfp_t gfp)
3960{
3961 struct address_space * const mapping = folio->mapping;
3962
3963 BUG_ON(!folio_test_locked(folio));
3964 if (folio_test_writeback(folio))
3965 return false;
3966
3967 if (mapping && mapping->a_ops->release_folio)
3968 return mapping->a_ops->release_folio(folio, gfp);
3969 return try_to_free_buffers(folio);
3970}
3971EXPORT_SYMBOL(filemap_release_folio);
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/mman.h>
25#include <linux/pagemap.h>
26#include <linux/file.h>
27#include <linux/uio.h>
28#include <linux/error-injection.h>
29#include <linux/hash.h>
30#include <linux/writeback.h>
31#include <linux/backing-dev.h>
32#include <linux/pagevec.h>
33#include <linux/blkdev.h>
34#include <linux/security.h>
35#include <linux/cpuset.h>
36#include <linux/hugetlb.h>
37#include <linux/memcontrol.h>
38#include <linux/cleancache.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 <asm/pgalloc.h>
46#include <asm/tlbflush.h>
47#include "internal.h"
48
49#define CREATE_TRACE_POINTS
50#include <trace/events/filemap.h>
51
52/*
53 * FIXME: remove all knowledge of the buffer layer from the core VM
54 */
55#include <linux/buffer_head.h> /* for try_to_free_buffers */
56
57#include <asm/mman.h>
58
59/*
60 * Shared mappings implemented 30.11.1994. It's not fully working yet,
61 * though.
62 *
63 * Shared mappings now work. 15.8.1995 Bruno.
64 *
65 * finished 'unifying' the page and buffer cache and SMP-threaded the
66 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
67 *
68 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
69 */
70
71/*
72 * Lock ordering:
73 *
74 * ->i_mmap_rwsem (truncate_pagecache)
75 * ->private_lock (__free_pte->__set_page_dirty_buffers)
76 * ->swap_lock (exclusive_swap_page, others)
77 * ->i_pages lock
78 *
79 * ->i_mutex
80 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
81 *
82 * ->mmap_lock
83 * ->i_mmap_rwsem
84 * ->page_table_lock or pte_lock (various, mainly in memory.c)
85 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
86 *
87 * ->mmap_lock
88 * ->lock_page (access_process_vm)
89 *
90 * ->i_mutex (generic_perform_write)
91 * ->mmap_lock (fault_in_pages_readable->do_page_fault)
92 *
93 * bdi->wb.list_lock
94 * sb_lock (fs/fs-writeback.c)
95 * ->i_pages lock (__sync_single_inode)
96 *
97 * ->i_mmap_rwsem
98 * ->anon_vma.lock (vma_adjust)
99 *
100 * ->anon_vma.lock
101 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
102 *
103 * ->page_table_lock or pte_lock
104 * ->swap_lock (try_to_unmap_one)
105 * ->private_lock (try_to_unmap_one)
106 * ->i_pages lock (try_to_unmap_one)
107 * ->lruvec->lru_lock (follow_page->mark_page_accessed)
108 * ->lruvec->lru_lock (check_pte_range->isolate_lru_page)
109 * ->private_lock (page_remove_rmap->set_page_dirty)
110 * ->i_pages lock (page_remove_rmap->set_page_dirty)
111 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
112 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
113 * ->memcg->move_lock (page_remove_rmap->lock_page_memcg)
114 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
115 * ->inode->i_lock (zap_pte_range->set_page_dirty)
116 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
117 *
118 * ->i_mmap_rwsem
119 * ->tasklist_lock (memory_failure, collect_procs_ao)
120 */
121
122static void page_cache_delete(struct address_space *mapping,
123 struct page *page, void *shadow)
124{
125 XA_STATE(xas, &mapping->i_pages, page->index);
126 unsigned int nr = 1;
127
128 mapping_set_update(&xas, mapping);
129
130 /* hugetlb pages are represented by a single entry in the xarray */
131 if (!PageHuge(page)) {
132 xas_set_order(&xas, page->index, compound_order(page));
133 nr = compound_nr(page);
134 }
135
136 VM_BUG_ON_PAGE(!PageLocked(page), page);
137 VM_BUG_ON_PAGE(PageTail(page), page);
138 VM_BUG_ON_PAGE(nr != 1 && shadow, page);
139
140 xas_store(&xas, shadow);
141 xas_init_marks(&xas);
142
143 page->mapping = NULL;
144 /* Leave page->index set: truncation lookup relies upon it */
145 mapping->nrpages -= nr;
146}
147
148static void unaccount_page_cache_page(struct address_space *mapping,
149 struct page *page)
150{
151 int nr;
152
153 /*
154 * if we're uptodate, flush out into the cleancache, otherwise
155 * invalidate any existing cleancache entries. We can't leave
156 * stale data around in the cleancache once our page is gone
157 */
158 if (PageUptodate(page) && PageMappedToDisk(page))
159 cleancache_put_page(page);
160 else
161 cleancache_invalidate_page(mapping, page);
162
163 VM_BUG_ON_PAGE(PageTail(page), page);
164 VM_BUG_ON_PAGE(page_mapped(page), page);
165 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
166 int mapcount;
167
168 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
169 current->comm, page_to_pfn(page));
170 dump_page(page, "still mapped when deleted");
171 dump_stack();
172 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
173
174 mapcount = page_mapcount(page);
175 if (mapping_exiting(mapping) &&
176 page_count(page) >= mapcount + 2) {
177 /*
178 * All vmas have already been torn down, so it's
179 * a good bet that actually the page is unmapped,
180 * and we'd prefer not to leak it: if we're wrong,
181 * some other bad page check should catch it later.
182 */
183 page_mapcount_reset(page);
184 page_ref_sub(page, mapcount);
185 }
186 }
187
188 /* hugetlb pages do not participate in page cache accounting. */
189 if (PageHuge(page))
190 return;
191
192 nr = thp_nr_pages(page);
193
194 __mod_lruvec_page_state(page, NR_FILE_PAGES, -nr);
195 if (PageSwapBacked(page)) {
196 __mod_lruvec_page_state(page, NR_SHMEM, -nr);
197 if (PageTransHuge(page))
198 __mod_lruvec_page_state(page, NR_SHMEM_THPS, -nr);
199 } else if (PageTransHuge(page)) {
200 __mod_lruvec_page_state(page, NR_FILE_THPS, -nr);
201 filemap_nr_thps_dec(mapping);
202 }
203
204 /*
205 * At this point page must be either written or cleaned by
206 * truncate. Dirty page here signals a bug and loss of
207 * unwritten data.
208 *
209 * This fixes dirty accounting after removing the page entirely
210 * but leaves PageDirty set: it has no effect for truncated
211 * page and anyway will be cleared before returning page into
212 * buddy allocator.
213 */
214 if (WARN_ON_ONCE(PageDirty(page)))
215 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
216}
217
218/*
219 * Delete a page from the page cache and free it. Caller has to make
220 * sure the page is locked and that nobody else uses it - or that usage
221 * is safe. The caller must hold the i_pages lock.
222 */
223void __delete_from_page_cache(struct page *page, void *shadow)
224{
225 struct address_space *mapping = page->mapping;
226
227 trace_mm_filemap_delete_from_page_cache(page);
228
229 unaccount_page_cache_page(mapping, page);
230 page_cache_delete(mapping, page, shadow);
231}
232
233static void page_cache_free_page(struct address_space *mapping,
234 struct page *page)
235{
236 void (*freepage)(struct page *);
237
238 freepage = mapping->a_ops->freepage;
239 if (freepage)
240 freepage(page);
241
242 if (PageTransHuge(page) && !PageHuge(page)) {
243 page_ref_sub(page, thp_nr_pages(page));
244 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
245 } else {
246 put_page(page);
247 }
248}
249
250/**
251 * delete_from_page_cache - delete page from page cache
252 * @page: the page which the kernel is trying to remove from page cache
253 *
254 * This must be called only on pages that have been verified to be in the page
255 * cache and locked. It will never put the page into the free list, the caller
256 * has a reference on the page.
257 */
258void delete_from_page_cache(struct page *page)
259{
260 struct address_space *mapping = page_mapping(page);
261 unsigned long flags;
262
263 BUG_ON(!PageLocked(page));
264 xa_lock_irqsave(&mapping->i_pages, flags);
265 __delete_from_page_cache(page, NULL);
266 xa_unlock_irqrestore(&mapping->i_pages, flags);
267
268 page_cache_free_page(mapping, page);
269}
270EXPORT_SYMBOL(delete_from_page_cache);
271
272/*
273 * page_cache_delete_batch - delete several pages from page cache
274 * @mapping: the mapping to which pages belong
275 * @pvec: pagevec with pages to delete
276 *
277 * The function walks over mapping->i_pages and removes pages passed in @pvec
278 * from the mapping. The function expects @pvec to be sorted by page index
279 * and is optimised for it to be dense.
280 * It tolerates holes in @pvec (mapping entries at those indices are not
281 * modified). The function expects only THP head pages to be present in the
282 * @pvec.
283 *
284 * The function expects the i_pages lock to be held.
285 */
286static void page_cache_delete_batch(struct address_space *mapping,
287 struct pagevec *pvec)
288{
289 XA_STATE(xas, &mapping->i_pages, pvec->pages[0]->index);
290 int total_pages = 0;
291 int i = 0;
292 struct page *page;
293
294 mapping_set_update(&xas, mapping);
295 xas_for_each(&xas, page, ULONG_MAX) {
296 if (i >= pagevec_count(pvec))
297 break;
298
299 /* A swap/dax/shadow entry got inserted? Skip it. */
300 if (xa_is_value(page))
301 continue;
302 /*
303 * A page got inserted in our range? Skip it. We have our
304 * pages locked so they are protected from being removed.
305 * If we see a page whose index is higher than ours, it
306 * means our page has been removed, which shouldn't be
307 * possible because we're holding the PageLock.
308 */
309 if (page != pvec->pages[i]) {
310 VM_BUG_ON_PAGE(page->index > pvec->pages[i]->index,
311 page);
312 continue;
313 }
314
315 WARN_ON_ONCE(!PageLocked(page));
316
317 if (page->index == xas.xa_index)
318 page->mapping = NULL;
319 /* Leave page->index set: truncation lookup relies on it */
320
321 /*
322 * Move to the next page in the vector if this is a regular
323 * page or the index is of the last sub-page of this compound
324 * page.
325 */
326 if (page->index + compound_nr(page) - 1 == xas.xa_index)
327 i++;
328 xas_store(&xas, NULL);
329 total_pages++;
330 }
331 mapping->nrpages -= total_pages;
332}
333
334void delete_from_page_cache_batch(struct address_space *mapping,
335 struct pagevec *pvec)
336{
337 int i;
338 unsigned long flags;
339
340 if (!pagevec_count(pvec))
341 return;
342
343 xa_lock_irqsave(&mapping->i_pages, flags);
344 for (i = 0; i < pagevec_count(pvec); i++) {
345 trace_mm_filemap_delete_from_page_cache(pvec->pages[i]);
346
347 unaccount_page_cache_page(mapping, pvec->pages[i]);
348 }
349 page_cache_delete_batch(mapping, pvec);
350 xa_unlock_irqrestore(&mapping->i_pages, flags);
351
352 for (i = 0; i < pagevec_count(pvec); i++)
353 page_cache_free_page(mapping, pvec->pages[i]);
354}
355
356int filemap_check_errors(struct address_space *mapping)
357{
358 int ret = 0;
359 /* Check for outstanding write errors */
360 if (test_bit(AS_ENOSPC, &mapping->flags) &&
361 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
362 ret = -ENOSPC;
363 if (test_bit(AS_EIO, &mapping->flags) &&
364 test_and_clear_bit(AS_EIO, &mapping->flags))
365 ret = -EIO;
366 return ret;
367}
368EXPORT_SYMBOL(filemap_check_errors);
369
370static int filemap_check_and_keep_errors(struct address_space *mapping)
371{
372 /* Check for outstanding write errors */
373 if (test_bit(AS_EIO, &mapping->flags))
374 return -EIO;
375 if (test_bit(AS_ENOSPC, &mapping->flags))
376 return -ENOSPC;
377 return 0;
378}
379
380/**
381 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
382 * @mapping: address space structure to write
383 * @start: offset in bytes where the range starts
384 * @end: offset in bytes where the range ends (inclusive)
385 * @sync_mode: enable synchronous operation
386 *
387 * Start writeback against all of a mapping's dirty pages that lie
388 * within the byte offsets <start, end> inclusive.
389 *
390 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
391 * opposed to a regular memory cleansing writeback. The difference between
392 * these two operations is that if a dirty page/buffer is encountered, it must
393 * be waited upon, and not just skipped over.
394 *
395 * Return: %0 on success, negative error code otherwise.
396 */
397int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
398 loff_t end, int sync_mode)
399{
400 int ret;
401 struct writeback_control wbc = {
402 .sync_mode = sync_mode,
403 .nr_to_write = LONG_MAX,
404 .range_start = start,
405 .range_end = end,
406 };
407
408 if (!mapping_can_writeback(mapping) ||
409 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
410 return 0;
411
412 wbc_attach_fdatawrite_inode(&wbc, mapping->host);
413 ret = do_writepages(mapping, &wbc);
414 wbc_detach_inode(&wbc);
415 return ret;
416}
417
418static inline int __filemap_fdatawrite(struct address_space *mapping,
419 int sync_mode)
420{
421 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
422}
423
424int filemap_fdatawrite(struct address_space *mapping)
425{
426 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
427}
428EXPORT_SYMBOL(filemap_fdatawrite);
429
430int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
431 loff_t end)
432{
433 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
434}
435EXPORT_SYMBOL(filemap_fdatawrite_range);
436
437/**
438 * filemap_flush - mostly a non-blocking flush
439 * @mapping: target address_space
440 *
441 * This is a mostly non-blocking flush. Not suitable for data-integrity
442 * purposes - I/O may not be started against all dirty pages.
443 *
444 * Return: %0 on success, negative error code otherwise.
445 */
446int filemap_flush(struct address_space *mapping)
447{
448 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
449}
450EXPORT_SYMBOL(filemap_flush);
451
452/**
453 * filemap_range_has_page - check if a page exists in range.
454 * @mapping: address space within which to check
455 * @start_byte: offset in bytes where the range starts
456 * @end_byte: offset in bytes where the range ends (inclusive)
457 *
458 * Find at least one page in the range supplied, usually used to check if
459 * direct writing in this range will trigger a writeback.
460 *
461 * Return: %true if at least one page exists in the specified range,
462 * %false otherwise.
463 */
464bool filemap_range_has_page(struct address_space *mapping,
465 loff_t start_byte, loff_t end_byte)
466{
467 struct page *page;
468 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
469 pgoff_t max = end_byte >> PAGE_SHIFT;
470
471 if (end_byte < start_byte)
472 return false;
473
474 rcu_read_lock();
475 for (;;) {
476 page = xas_find(&xas, max);
477 if (xas_retry(&xas, page))
478 continue;
479 /* Shadow entries don't count */
480 if (xa_is_value(page))
481 continue;
482 /*
483 * We don't need to try to pin this page; we're about to
484 * release the RCU lock anyway. It is enough to know that
485 * there was a page here recently.
486 */
487 break;
488 }
489 rcu_read_unlock();
490
491 return page != NULL;
492}
493EXPORT_SYMBOL(filemap_range_has_page);
494
495static void __filemap_fdatawait_range(struct address_space *mapping,
496 loff_t start_byte, loff_t end_byte)
497{
498 pgoff_t index = start_byte >> PAGE_SHIFT;
499 pgoff_t end = end_byte >> PAGE_SHIFT;
500 struct pagevec pvec;
501 int nr_pages;
502
503 if (end_byte < start_byte)
504 return;
505
506 pagevec_init(&pvec);
507 while (index <= end) {
508 unsigned i;
509
510 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
511 end, PAGECACHE_TAG_WRITEBACK);
512 if (!nr_pages)
513 break;
514
515 for (i = 0; i < nr_pages; i++) {
516 struct page *page = pvec.pages[i];
517
518 wait_on_page_writeback(page);
519 ClearPageError(page);
520 }
521 pagevec_release(&pvec);
522 cond_resched();
523 }
524}
525
526/**
527 * filemap_fdatawait_range - wait for writeback to complete
528 * @mapping: address space structure to wait for
529 * @start_byte: offset in bytes where the range starts
530 * @end_byte: offset in bytes where the range ends (inclusive)
531 *
532 * Walk the list of under-writeback pages of the given address space
533 * in the given range and wait for all of them. Check error status of
534 * the address space and return it.
535 *
536 * Since the error status of the address space is cleared by this function,
537 * callers are responsible for checking the return value and handling and/or
538 * reporting the error.
539 *
540 * Return: error status of the address space.
541 */
542int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
543 loff_t end_byte)
544{
545 __filemap_fdatawait_range(mapping, start_byte, end_byte);
546 return filemap_check_errors(mapping);
547}
548EXPORT_SYMBOL(filemap_fdatawait_range);
549
550/**
551 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
552 * @mapping: address space structure to wait for
553 * @start_byte: offset in bytes where the range starts
554 * @end_byte: offset in bytes where the range ends (inclusive)
555 *
556 * Walk the list of under-writeback pages of the given address space in the
557 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
558 * this function does not clear error status of the address space.
559 *
560 * Use this function if callers don't handle errors themselves. Expected
561 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
562 * fsfreeze(8)
563 */
564int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
565 loff_t start_byte, loff_t end_byte)
566{
567 __filemap_fdatawait_range(mapping, start_byte, end_byte);
568 return filemap_check_and_keep_errors(mapping);
569}
570EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
571
572/**
573 * file_fdatawait_range - wait for writeback to complete
574 * @file: file pointing to address space structure to wait for
575 * @start_byte: offset in bytes where the range starts
576 * @end_byte: offset in bytes where the range ends (inclusive)
577 *
578 * Walk the list of under-writeback pages of the address space that file
579 * refers to, in the given range and wait for all of them. Check error
580 * status of the address space vs. the file->f_wb_err cursor and return it.
581 *
582 * Since the error status of the file is advanced by this function,
583 * callers are responsible for checking the return value and handling and/or
584 * reporting the error.
585 *
586 * Return: error status of the address space vs. the file->f_wb_err cursor.
587 */
588int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
589{
590 struct address_space *mapping = file->f_mapping;
591
592 __filemap_fdatawait_range(mapping, start_byte, end_byte);
593 return file_check_and_advance_wb_err(file);
594}
595EXPORT_SYMBOL(file_fdatawait_range);
596
597/**
598 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
599 * @mapping: address space structure to wait for
600 *
601 * Walk the list of under-writeback pages of the given address space
602 * and wait for all of them. Unlike filemap_fdatawait(), this function
603 * does not clear error status of the address space.
604 *
605 * Use this function if callers don't handle errors themselves. Expected
606 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
607 * fsfreeze(8)
608 *
609 * Return: error status of the address space.
610 */
611int filemap_fdatawait_keep_errors(struct address_space *mapping)
612{
613 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
614 return filemap_check_and_keep_errors(mapping);
615}
616EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
617
618/* Returns true if writeback might be needed or already in progress. */
619static bool mapping_needs_writeback(struct address_space *mapping)
620{
621 return mapping->nrpages;
622}
623
624/**
625 * filemap_range_needs_writeback - check if range potentially needs writeback
626 * @mapping: address space within which to check
627 * @start_byte: offset in bytes where the range starts
628 * @end_byte: offset in bytes where the range ends (inclusive)
629 *
630 * Find at least one page in the range supplied, usually used to check if
631 * direct writing in this range will trigger a writeback. Used by O_DIRECT
632 * read/write with IOCB_NOWAIT, to see if the caller needs to do
633 * filemap_write_and_wait_range() before proceeding.
634 *
635 * Return: %true if the caller should do filemap_write_and_wait_range() before
636 * doing O_DIRECT to a page in this range, %false otherwise.
637 */
638bool filemap_range_needs_writeback(struct address_space *mapping,
639 loff_t start_byte, loff_t end_byte)
640{
641 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
642 pgoff_t max = end_byte >> PAGE_SHIFT;
643 struct page *page;
644
645 if (!mapping_needs_writeback(mapping))
646 return false;
647 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
648 !mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
649 return false;
650 if (end_byte < start_byte)
651 return false;
652
653 rcu_read_lock();
654 xas_for_each(&xas, page, max) {
655 if (xas_retry(&xas, page))
656 continue;
657 if (xa_is_value(page))
658 continue;
659 if (PageDirty(page) || PageLocked(page) || PageWriteback(page))
660 break;
661 }
662 rcu_read_unlock();
663 return page != NULL;
664}
665EXPORT_SYMBOL_GPL(filemap_range_needs_writeback);
666
667/**
668 * filemap_write_and_wait_range - write out & wait on a file range
669 * @mapping: the address_space for the pages
670 * @lstart: offset in bytes where the range starts
671 * @lend: offset in bytes where the range ends (inclusive)
672 *
673 * Write out and wait upon file offsets lstart->lend, inclusive.
674 *
675 * Note that @lend is inclusive (describes the last byte to be written) so
676 * that this function can be used to write to the very end-of-file (end = -1).
677 *
678 * Return: error status of the address space.
679 */
680int filemap_write_and_wait_range(struct address_space *mapping,
681 loff_t lstart, loff_t lend)
682{
683 int err = 0;
684
685 if (mapping_needs_writeback(mapping)) {
686 err = __filemap_fdatawrite_range(mapping, lstart, lend,
687 WB_SYNC_ALL);
688 /*
689 * Even if the above returned error, the pages may be
690 * written partially (e.g. -ENOSPC), so we wait for it.
691 * But the -EIO is special case, it may indicate the worst
692 * thing (e.g. bug) happened, so we avoid waiting for it.
693 */
694 if (err != -EIO) {
695 int err2 = filemap_fdatawait_range(mapping,
696 lstart, lend);
697 if (!err)
698 err = err2;
699 } else {
700 /* Clear any previously stored errors */
701 filemap_check_errors(mapping);
702 }
703 } else {
704 err = filemap_check_errors(mapping);
705 }
706 return err;
707}
708EXPORT_SYMBOL(filemap_write_and_wait_range);
709
710void __filemap_set_wb_err(struct address_space *mapping, int err)
711{
712 errseq_t eseq = errseq_set(&mapping->wb_err, err);
713
714 trace_filemap_set_wb_err(mapping, eseq);
715}
716EXPORT_SYMBOL(__filemap_set_wb_err);
717
718/**
719 * file_check_and_advance_wb_err - report wb error (if any) that was previously
720 * and advance wb_err to current one
721 * @file: struct file on which the error is being reported
722 *
723 * When userland calls fsync (or something like nfsd does the equivalent), we
724 * want to report any writeback errors that occurred since the last fsync (or
725 * since the file was opened if there haven't been any).
726 *
727 * Grab the wb_err from the mapping. If it matches what we have in the file,
728 * then just quickly return 0. The file is all caught up.
729 *
730 * If it doesn't match, then take the mapping value, set the "seen" flag in
731 * it and try to swap it into place. If it works, or another task beat us
732 * to it with the new value, then update the f_wb_err and return the error
733 * portion. The error at this point must be reported via proper channels
734 * (a'la fsync, or NFS COMMIT operation, etc.).
735 *
736 * While we handle mapping->wb_err with atomic operations, the f_wb_err
737 * value is protected by the f_lock since we must ensure that it reflects
738 * the latest value swapped in for this file descriptor.
739 *
740 * Return: %0 on success, negative error code otherwise.
741 */
742int file_check_and_advance_wb_err(struct file *file)
743{
744 int err = 0;
745 errseq_t old = READ_ONCE(file->f_wb_err);
746 struct address_space *mapping = file->f_mapping;
747
748 /* Locklessly handle the common case where nothing has changed */
749 if (errseq_check(&mapping->wb_err, old)) {
750 /* Something changed, must use slow path */
751 spin_lock(&file->f_lock);
752 old = file->f_wb_err;
753 err = errseq_check_and_advance(&mapping->wb_err,
754 &file->f_wb_err);
755 trace_file_check_and_advance_wb_err(file, old);
756 spin_unlock(&file->f_lock);
757 }
758
759 /*
760 * We're mostly using this function as a drop in replacement for
761 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
762 * that the legacy code would have had on these flags.
763 */
764 clear_bit(AS_EIO, &mapping->flags);
765 clear_bit(AS_ENOSPC, &mapping->flags);
766 return err;
767}
768EXPORT_SYMBOL(file_check_and_advance_wb_err);
769
770/**
771 * file_write_and_wait_range - write out & wait on a file range
772 * @file: file pointing to address_space with pages
773 * @lstart: offset in bytes where the range starts
774 * @lend: offset in bytes where the range ends (inclusive)
775 *
776 * Write out and wait upon file offsets lstart->lend, inclusive.
777 *
778 * Note that @lend is inclusive (describes the last byte to be written) so
779 * that this function can be used to write to the very end-of-file (end = -1).
780 *
781 * After writing out and waiting on the data, we check and advance the
782 * f_wb_err cursor to the latest value, and return any errors detected there.
783 *
784 * Return: %0 on success, negative error code otherwise.
785 */
786int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
787{
788 int err = 0, err2;
789 struct address_space *mapping = file->f_mapping;
790
791 if (mapping_needs_writeback(mapping)) {
792 err = __filemap_fdatawrite_range(mapping, lstart, lend,
793 WB_SYNC_ALL);
794 /* See comment of filemap_write_and_wait() */
795 if (err != -EIO)
796 __filemap_fdatawait_range(mapping, lstart, lend);
797 }
798 err2 = file_check_and_advance_wb_err(file);
799 if (!err)
800 err = err2;
801 return err;
802}
803EXPORT_SYMBOL(file_write_and_wait_range);
804
805/**
806 * replace_page_cache_page - replace a pagecache page with a new one
807 * @old: page to be replaced
808 * @new: page to replace with
809 *
810 * This function replaces a page in the pagecache with a new one. On
811 * success it acquires the pagecache reference for the new page and
812 * drops it for the old page. Both the old and new pages must be
813 * locked. This function does not add the new page to the LRU, the
814 * caller must do that.
815 *
816 * The remove + add is atomic. This function cannot fail.
817 */
818void replace_page_cache_page(struct page *old, struct page *new)
819{
820 struct address_space *mapping = old->mapping;
821 void (*freepage)(struct page *) = mapping->a_ops->freepage;
822 pgoff_t offset = old->index;
823 XA_STATE(xas, &mapping->i_pages, offset);
824 unsigned long flags;
825
826 VM_BUG_ON_PAGE(!PageLocked(old), old);
827 VM_BUG_ON_PAGE(!PageLocked(new), new);
828 VM_BUG_ON_PAGE(new->mapping, new);
829
830 get_page(new);
831 new->mapping = mapping;
832 new->index = offset;
833
834 mem_cgroup_migrate(old, new);
835
836 xas_lock_irqsave(&xas, flags);
837 xas_store(&xas, new);
838
839 old->mapping = NULL;
840 /* hugetlb pages do not participate in page cache accounting. */
841 if (!PageHuge(old))
842 __dec_lruvec_page_state(old, NR_FILE_PAGES);
843 if (!PageHuge(new))
844 __inc_lruvec_page_state(new, NR_FILE_PAGES);
845 if (PageSwapBacked(old))
846 __dec_lruvec_page_state(old, NR_SHMEM);
847 if (PageSwapBacked(new))
848 __inc_lruvec_page_state(new, NR_SHMEM);
849 xas_unlock_irqrestore(&xas, flags);
850 if (freepage)
851 freepage(old);
852 put_page(old);
853}
854EXPORT_SYMBOL_GPL(replace_page_cache_page);
855
856noinline int __add_to_page_cache_locked(struct page *page,
857 struct address_space *mapping,
858 pgoff_t offset, gfp_t gfp,
859 void **shadowp)
860{
861 XA_STATE(xas, &mapping->i_pages, offset);
862 int huge = PageHuge(page);
863 int error;
864 bool charged = false;
865
866 VM_BUG_ON_PAGE(!PageLocked(page), page);
867 VM_BUG_ON_PAGE(PageSwapBacked(page), page);
868 mapping_set_update(&xas, mapping);
869
870 get_page(page);
871 page->mapping = mapping;
872 page->index = offset;
873
874 if (!huge) {
875 error = mem_cgroup_charge(page, NULL, gfp);
876 if (error)
877 goto error;
878 charged = true;
879 }
880
881 gfp &= GFP_RECLAIM_MASK;
882
883 do {
884 unsigned int order = xa_get_order(xas.xa, xas.xa_index);
885 void *entry, *old = NULL;
886
887 if (order > thp_order(page))
888 xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index),
889 order, gfp);
890 xas_lock_irq(&xas);
891 xas_for_each_conflict(&xas, entry) {
892 old = entry;
893 if (!xa_is_value(entry)) {
894 xas_set_err(&xas, -EEXIST);
895 goto unlock;
896 }
897 }
898
899 if (old) {
900 if (shadowp)
901 *shadowp = old;
902 /* entry may have been split before we acquired lock */
903 order = xa_get_order(xas.xa, xas.xa_index);
904 if (order > thp_order(page)) {
905 xas_split(&xas, old, order);
906 xas_reset(&xas);
907 }
908 }
909
910 xas_store(&xas, page);
911 if (xas_error(&xas))
912 goto unlock;
913
914 mapping->nrpages++;
915
916 /* hugetlb pages do not participate in page cache accounting */
917 if (!huge)
918 __inc_lruvec_page_state(page, NR_FILE_PAGES);
919unlock:
920 xas_unlock_irq(&xas);
921 } while (xas_nomem(&xas, gfp));
922
923 if (xas_error(&xas)) {
924 error = xas_error(&xas);
925 if (charged)
926 mem_cgroup_uncharge(page);
927 goto error;
928 }
929
930 trace_mm_filemap_add_to_page_cache(page);
931 return 0;
932error:
933 page->mapping = NULL;
934 /* Leave page->index set: truncation relies upon it */
935 put_page(page);
936 return error;
937}
938ALLOW_ERROR_INJECTION(__add_to_page_cache_locked, ERRNO);
939
940/**
941 * add_to_page_cache_locked - add a locked page to the pagecache
942 * @page: page to add
943 * @mapping: the page's address_space
944 * @offset: page index
945 * @gfp_mask: page allocation mode
946 *
947 * This function is used to add a page to the pagecache. It must be locked.
948 * This function does not add the page to the LRU. The caller must do that.
949 *
950 * Return: %0 on success, negative error code otherwise.
951 */
952int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
953 pgoff_t offset, gfp_t gfp_mask)
954{
955 return __add_to_page_cache_locked(page, mapping, offset,
956 gfp_mask, NULL);
957}
958EXPORT_SYMBOL(add_to_page_cache_locked);
959
960int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
961 pgoff_t offset, gfp_t gfp_mask)
962{
963 void *shadow = NULL;
964 int ret;
965
966 __SetPageLocked(page);
967 ret = __add_to_page_cache_locked(page, mapping, offset,
968 gfp_mask, &shadow);
969 if (unlikely(ret))
970 __ClearPageLocked(page);
971 else {
972 /*
973 * The page might have been evicted from cache only
974 * recently, in which case it should be activated like
975 * any other repeatedly accessed page.
976 * The exception is pages getting rewritten; evicting other
977 * data from the working set, only to cache data that will
978 * get overwritten with something else, is a waste of memory.
979 */
980 WARN_ON_ONCE(PageActive(page));
981 if (!(gfp_mask & __GFP_WRITE) && shadow)
982 workingset_refault(page, shadow);
983 lru_cache_add(page);
984 }
985 return ret;
986}
987EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
988
989#ifdef CONFIG_NUMA
990struct page *__page_cache_alloc(gfp_t gfp)
991{
992 int n;
993 struct page *page;
994
995 if (cpuset_do_page_mem_spread()) {
996 unsigned int cpuset_mems_cookie;
997 do {
998 cpuset_mems_cookie = read_mems_allowed_begin();
999 n = cpuset_mem_spread_node();
1000 page = __alloc_pages_node(n, gfp, 0);
1001 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
1002
1003 return page;
1004 }
1005 return alloc_pages(gfp, 0);
1006}
1007EXPORT_SYMBOL(__page_cache_alloc);
1008#endif
1009
1010/*
1011 * In order to wait for pages to become available there must be
1012 * waitqueues associated with pages. By using a hash table of
1013 * waitqueues where the bucket discipline is to maintain all
1014 * waiters on the same queue and wake all when any of the pages
1015 * become available, and for the woken contexts to check to be
1016 * sure the appropriate page became available, this saves space
1017 * at a cost of "thundering herd" phenomena during rare hash
1018 * collisions.
1019 */
1020#define PAGE_WAIT_TABLE_BITS 8
1021#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1022static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1023
1024static wait_queue_head_t *page_waitqueue(struct page *page)
1025{
1026 return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
1027}
1028
1029void __init pagecache_init(void)
1030{
1031 int i;
1032
1033 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1034 init_waitqueue_head(&page_wait_table[i]);
1035
1036 page_writeback_init();
1037}
1038
1039/*
1040 * The page wait code treats the "wait->flags" somewhat unusually, because
1041 * we have multiple different kinds of waits, not just the usual "exclusive"
1042 * one.
1043 *
1044 * We have:
1045 *
1046 * (a) no special bits set:
1047 *
1048 * We're just waiting for the bit to be released, and when a waker
1049 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1050 * and remove it from the wait queue.
1051 *
1052 * Simple and straightforward.
1053 *
1054 * (b) WQ_FLAG_EXCLUSIVE:
1055 *
1056 * The waiter is waiting to get the lock, and only one waiter should
1057 * be woken up to avoid any thundering herd behavior. We'll set the
1058 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1059 *
1060 * This is the traditional exclusive wait.
1061 *
1062 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1063 *
1064 * The waiter is waiting to get the bit, and additionally wants the
1065 * lock to be transferred to it for fair lock behavior. If the lock
1066 * cannot be taken, we stop walking the wait queue without waking
1067 * the waiter.
1068 *
1069 * This is the "fair lock handoff" case, and in addition to setting
1070 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1071 * that it now has the lock.
1072 */
1073static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1074{
1075 unsigned int flags;
1076 struct wait_page_key *key = arg;
1077 struct wait_page_queue *wait_page
1078 = container_of(wait, struct wait_page_queue, wait);
1079
1080 if (!wake_page_match(wait_page, key))
1081 return 0;
1082
1083 /*
1084 * If it's a lock handoff wait, we get the bit for it, and
1085 * stop walking (and do not wake it up) if we can't.
1086 */
1087 flags = wait->flags;
1088 if (flags & WQ_FLAG_EXCLUSIVE) {
1089 if (test_bit(key->bit_nr, &key->page->flags))
1090 return -1;
1091 if (flags & WQ_FLAG_CUSTOM) {
1092 if (test_and_set_bit(key->bit_nr, &key->page->flags))
1093 return -1;
1094 flags |= WQ_FLAG_DONE;
1095 }
1096 }
1097
1098 /*
1099 * We are holding the wait-queue lock, but the waiter that
1100 * is waiting for this will be checking the flags without
1101 * any locking.
1102 *
1103 * So update the flags atomically, and wake up the waiter
1104 * afterwards to avoid any races. This store-release pairs
1105 * with the load-acquire in wait_on_page_bit_common().
1106 */
1107 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1108 wake_up_state(wait->private, mode);
1109
1110 /*
1111 * Ok, we have successfully done what we're waiting for,
1112 * and we can unconditionally remove the wait entry.
1113 *
1114 * Note that this pairs with the "finish_wait()" in the
1115 * waiter, and has to be the absolute last thing we do.
1116 * After this list_del_init(&wait->entry) the wait entry
1117 * might be de-allocated and the process might even have
1118 * exited.
1119 */
1120 list_del_init_careful(&wait->entry);
1121 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1122}
1123
1124static void wake_up_page_bit(struct page *page, int bit_nr)
1125{
1126 wait_queue_head_t *q = page_waitqueue(page);
1127 struct wait_page_key key;
1128 unsigned long flags;
1129 wait_queue_entry_t bookmark;
1130
1131 key.page = page;
1132 key.bit_nr = bit_nr;
1133 key.page_match = 0;
1134
1135 bookmark.flags = 0;
1136 bookmark.private = NULL;
1137 bookmark.func = NULL;
1138 INIT_LIST_HEAD(&bookmark.entry);
1139
1140 spin_lock_irqsave(&q->lock, flags);
1141 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1142
1143 while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1144 /*
1145 * Take a breather from holding the lock,
1146 * allow pages that finish wake up asynchronously
1147 * to acquire the lock and remove themselves
1148 * from wait queue
1149 */
1150 spin_unlock_irqrestore(&q->lock, flags);
1151 cpu_relax();
1152 spin_lock_irqsave(&q->lock, flags);
1153 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1154 }
1155
1156 /*
1157 * It is possible for other pages to have collided on the waitqueue
1158 * hash, so in that case check for a page match. That prevents a long-
1159 * term waiter
1160 *
1161 * It is still possible to miss a case here, when we woke page waiters
1162 * and removed them from the waitqueue, but there are still other
1163 * page waiters.
1164 */
1165 if (!waitqueue_active(q) || !key.page_match) {
1166 ClearPageWaiters(page);
1167 /*
1168 * It's possible to miss clearing Waiters here, when we woke
1169 * our page waiters, but the hashed waitqueue has waiters for
1170 * other pages on it.
1171 *
1172 * That's okay, it's a rare case. The next waker will clear it.
1173 */
1174 }
1175 spin_unlock_irqrestore(&q->lock, flags);
1176}
1177
1178static void wake_up_page(struct page *page, int bit)
1179{
1180 if (!PageWaiters(page))
1181 return;
1182 wake_up_page_bit(page, bit);
1183}
1184
1185/*
1186 * A choice of three behaviors for wait_on_page_bit_common():
1187 */
1188enum behavior {
1189 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1190 * __lock_page() waiting on then setting PG_locked.
1191 */
1192 SHARED, /* Hold ref to page and check the bit when woken, like
1193 * wait_on_page_writeback() waiting on PG_writeback.
1194 */
1195 DROP, /* Drop ref to page before wait, no check when woken,
1196 * like put_and_wait_on_page_locked() on PG_locked.
1197 */
1198};
1199
1200/*
1201 * Attempt to check (or get) the page bit, and mark us done
1202 * if successful.
1203 */
1204static inline bool trylock_page_bit_common(struct page *page, int bit_nr,
1205 struct wait_queue_entry *wait)
1206{
1207 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1208 if (test_and_set_bit(bit_nr, &page->flags))
1209 return false;
1210 } else if (test_bit(bit_nr, &page->flags))
1211 return false;
1212
1213 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1214 return true;
1215}
1216
1217/* How many times do we accept lock stealing from under a waiter? */
1218int sysctl_page_lock_unfairness = 5;
1219
1220static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1221 struct page *page, int bit_nr, int state, enum behavior behavior)
1222{
1223 int unfairness = sysctl_page_lock_unfairness;
1224 struct wait_page_queue wait_page;
1225 wait_queue_entry_t *wait = &wait_page.wait;
1226 bool thrashing = false;
1227 bool delayacct = false;
1228 unsigned long pflags;
1229
1230 if (bit_nr == PG_locked &&
1231 !PageUptodate(page) && PageWorkingset(page)) {
1232 if (!PageSwapBacked(page)) {
1233 delayacct_thrashing_start();
1234 delayacct = true;
1235 }
1236 psi_memstall_enter(&pflags);
1237 thrashing = true;
1238 }
1239
1240 init_wait(wait);
1241 wait->func = wake_page_function;
1242 wait_page.page = page;
1243 wait_page.bit_nr = bit_nr;
1244
1245repeat:
1246 wait->flags = 0;
1247 if (behavior == EXCLUSIVE) {
1248 wait->flags = WQ_FLAG_EXCLUSIVE;
1249 if (--unfairness < 0)
1250 wait->flags |= WQ_FLAG_CUSTOM;
1251 }
1252
1253 /*
1254 * Do one last check whether we can get the
1255 * page bit synchronously.
1256 *
1257 * Do the SetPageWaiters() marking before that
1258 * to let any waker we _just_ missed know they
1259 * need to wake us up (otherwise they'll never
1260 * even go to the slow case that looks at the
1261 * page queue), and add ourselves to the wait
1262 * queue if we need to sleep.
1263 *
1264 * This part needs to be done under the queue
1265 * lock to avoid races.
1266 */
1267 spin_lock_irq(&q->lock);
1268 SetPageWaiters(page);
1269 if (!trylock_page_bit_common(page, bit_nr, wait))
1270 __add_wait_queue_entry_tail(q, wait);
1271 spin_unlock_irq(&q->lock);
1272
1273 /*
1274 * From now on, all the logic will be based on
1275 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1276 * see whether the page bit testing has already
1277 * been done by the wake function.
1278 *
1279 * We can drop our reference to the page.
1280 */
1281 if (behavior == DROP)
1282 put_page(page);
1283
1284 /*
1285 * Note that until the "finish_wait()", or until
1286 * we see the WQ_FLAG_WOKEN flag, we need to
1287 * be very careful with the 'wait->flags', because
1288 * we may race with a waker that sets them.
1289 */
1290 for (;;) {
1291 unsigned int flags;
1292
1293 set_current_state(state);
1294
1295 /* Loop until we've been woken or interrupted */
1296 flags = smp_load_acquire(&wait->flags);
1297 if (!(flags & WQ_FLAG_WOKEN)) {
1298 if (signal_pending_state(state, current))
1299 break;
1300
1301 io_schedule();
1302 continue;
1303 }
1304
1305 /* If we were non-exclusive, we're done */
1306 if (behavior != EXCLUSIVE)
1307 break;
1308
1309 /* If the waker got the lock for us, we're done */
1310 if (flags & WQ_FLAG_DONE)
1311 break;
1312
1313 /*
1314 * Otherwise, if we're getting the lock, we need to
1315 * try to get it ourselves.
1316 *
1317 * And if that fails, we'll have to retry this all.
1318 */
1319 if (unlikely(test_and_set_bit(bit_nr, &page->flags)))
1320 goto repeat;
1321
1322 wait->flags |= WQ_FLAG_DONE;
1323 break;
1324 }
1325
1326 /*
1327 * If a signal happened, this 'finish_wait()' may remove the last
1328 * waiter from the wait-queues, but the PageWaiters bit will remain
1329 * set. That's ok. The next wakeup will take care of it, and trying
1330 * to do it here would be difficult and prone to races.
1331 */
1332 finish_wait(q, wait);
1333
1334 if (thrashing) {
1335 if (delayacct)
1336 delayacct_thrashing_end();
1337 psi_memstall_leave(&pflags);
1338 }
1339
1340 /*
1341 * NOTE! The wait->flags weren't stable until we've done the
1342 * 'finish_wait()', and we could have exited the loop above due
1343 * to a signal, and had a wakeup event happen after the signal
1344 * test but before the 'finish_wait()'.
1345 *
1346 * So only after the finish_wait() can we reliably determine
1347 * if we got woken up or not, so we can now figure out the final
1348 * return value based on that state without races.
1349 *
1350 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1351 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1352 */
1353 if (behavior == EXCLUSIVE)
1354 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1355
1356 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1357}
1358
1359void wait_on_page_bit(struct page *page, int bit_nr)
1360{
1361 wait_queue_head_t *q = page_waitqueue(page);
1362 wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1363}
1364EXPORT_SYMBOL(wait_on_page_bit);
1365
1366int wait_on_page_bit_killable(struct page *page, int bit_nr)
1367{
1368 wait_queue_head_t *q = page_waitqueue(page);
1369 return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, SHARED);
1370}
1371EXPORT_SYMBOL(wait_on_page_bit_killable);
1372
1373/**
1374 * put_and_wait_on_page_locked - Drop a reference and wait for it to be unlocked
1375 * @page: The page to wait for.
1376 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1377 *
1378 * The caller should hold a reference on @page. They expect the page to
1379 * become unlocked relatively soon, but do not wish to hold up migration
1380 * (for example) by holding the reference while waiting for the page to
1381 * come unlocked. After this function returns, the caller should not
1382 * dereference @page.
1383 *
1384 * Return: 0 if the page was unlocked or -EINTR if interrupted by a signal.
1385 */
1386int put_and_wait_on_page_locked(struct page *page, int state)
1387{
1388 wait_queue_head_t *q;
1389
1390 page = compound_head(page);
1391 q = page_waitqueue(page);
1392 return wait_on_page_bit_common(q, page, PG_locked, state, DROP);
1393}
1394
1395/**
1396 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1397 * @page: Page defining the wait queue of interest
1398 * @waiter: Waiter to add to the queue
1399 *
1400 * Add an arbitrary @waiter to the wait queue for the nominated @page.
1401 */
1402void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1403{
1404 wait_queue_head_t *q = page_waitqueue(page);
1405 unsigned long flags;
1406
1407 spin_lock_irqsave(&q->lock, flags);
1408 __add_wait_queue_entry_tail(q, waiter);
1409 SetPageWaiters(page);
1410 spin_unlock_irqrestore(&q->lock, flags);
1411}
1412EXPORT_SYMBOL_GPL(add_page_wait_queue);
1413
1414#ifndef clear_bit_unlock_is_negative_byte
1415
1416/*
1417 * PG_waiters is the high bit in the same byte as PG_lock.
1418 *
1419 * On x86 (and on many other architectures), we can clear PG_lock and
1420 * test the sign bit at the same time. But if the architecture does
1421 * not support that special operation, we just do this all by hand
1422 * instead.
1423 *
1424 * The read of PG_waiters has to be after (or concurrently with) PG_locked
1425 * being cleared, but a memory barrier should be unnecessary since it is
1426 * in the same byte as PG_locked.
1427 */
1428static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1429{
1430 clear_bit_unlock(nr, mem);
1431 /* smp_mb__after_atomic(); */
1432 return test_bit(PG_waiters, mem);
1433}
1434
1435#endif
1436
1437/**
1438 * unlock_page - unlock a locked page
1439 * @page: the page
1440 *
1441 * Unlocks the page and wakes up sleepers in wait_on_page_locked().
1442 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1443 * mechanism between PageLocked pages and PageWriteback pages is shared.
1444 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1445 *
1446 * Note that this depends on PG_waiters being the sign bit in the byte
1447 * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1448 * clear the PG_locked bit and test PG_waiters at the same time fairly
1449 * portably (architectures that do LL/SC can test any bit, while x86 can
1450 * test the sign bit).
1451 */
1452void unlock_page(struct page *page)
1453{
1454 BUILD_BUG_ON(PG_waiters != 7);
1455 page = compound_head(page);
1456 VM_BUG_ON_PAGE(!PageLocked(page), page);
1457 if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1458 wake_up_page_bit(page, PG_locked);
1459}
1460EXPORT_SYMBOL(unlock_page);
1461
1462/**
1463 * end_page_private_2 - Clear PG_private_2 and release any waiters
1464 * @page: The page
1465 *
1466 * Clear the PG_private_2 bit on a page and wake up any sleepers waiting for
1467 * this. The page ref held for PG_private_2 being set is released.
1468 *
1469 * This is, for example, used when a netfs page is being written to a local
1470 * disk cache, thereby allowing writes to the cache for the same page to be
1471 * serialised.
1472 */
1473void end_page_private_2(struct page *page)
1474{
1475 page = compound_head(page);
1476 VM_BUG_ON_PAGE(!PagePrivate2(page), page);
1477 clear_bit_unlock(PG_private_2, &page->flags);
1478 wake_up_page_bit(page, PG_private_2);
1479 put_page(page);
1480}
1481EXPORT_SYMBOL(end_page_private_2);
1482
1483/**
1484 * wait_on_page_private_2 - Wait for PG_private_2 to be cleared on a page
1485 * @page: The page to wait on
1486 *
1487 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a page.
1488 */
1489void wait_on_page_private_2(struct page *page)
1490{
1491 page = compound_head(page);
1492 while (PagePrivate2(page))
1493 wait_on_page_bit(page, PG_private_2);
1494}
1495EXPORT_SYMBOL(wait_on_page_private_2);
1496
1497/**
1498 * wait_on_page_private_2_killable - Wait for PG_private_2 to be cleared on a page
1499 * @page: The page to wait on
1500 *
1501 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a page or until a
1502 * fatal signal is received by the calling task.
1503 *
1504 * Return:
1505 * - 0 if successful.
1506 * - -EINTR if a fatal signal was encountered.
1507 */
1508int wait_on_page_private_2_killable(struct page *page)
1509{
1510 int ret = 0;
1511
1512 page = compound_head(page);
1513 while (PagePrivate2(page)) {
1514 ret = wait_on_page_bit_killable(page, PG_private_2);
1515 if (ret < 0)
1516 break;
1517 }
1518
1519 return ret;
1520}
1521EXPORT_SYMBOL(wait_on_page_private_2_killable);
1522
1523/**
1524 * end_page_writeback - end writeback against a page
1525 * @page: the page
1526 */
1527void end_page_writeback(struct page *page)
1528{
1529 /*
1530 * TestClearPageReclaim could be used here but it is an atomic
1531 * operation and overkill in this particular case. Failing to
1532 * shuffle a page marked for immediate reclaim is too mild to
1533 * justify taking an atomic operation penalty at the end of
1534 * ever page writeback.
1535 */
1536 if (PageReclaim(page)) {
1537 ClearPageReclaim(page);
1538 rotate_reclaimable_page(page);
1539 }
1540
1541 /*
1542 * Writeback does not hold a page reference of its own, relying
1543 * on truncation to wait for the clearing of PG_writeback.
1544 * But here we must make sure that the page is not freed and
1545 * reused before the wake_up_page().
1546 */
1547 get_page(page);
1548 if (!test_clear_page_writeback(page))
1549 BUG();
1550
1551 smp_mb__after_atomic();
1552 wake_up_page(page, PG_writeback);
1553 put_page(page);
1554}
1555EXPORT_SYMBOL(end_page_writeback);
1556
1557/*
1558 * After completing I/O on a page, call this routine to update the page
1559 * flags appropriately
1560 */
1561void page_endio(struct page *page, bool is_write, int err)
1562{
1563 if (!is_write) {
1564 if (!err) {
1565 SetPageUptodate(page);
1566 } else {
1567 ClearPageUptodate(page);
1568 SetPageError(page);
1569 }
1570 unlock_page(page);
1571 } else {
1572 if (err) {
1573 struct address_space *mapping;
1574
1575 SetPageError(page);
1576 mapping = page_mapping(page);
1577 if (mapping)
1578 mapping_set_error(mapping, err);
1579 }
1580 end_page_writeback(page);
1581 }
1582}
1583EXPORT_SYMBOL_GPL(page_endio);
1584
1585/**
1586 * __lock_page - get a lock on the page, assuming we need to sleep to get it
1587 * @__page: the page to lock
1588 */
1589void __lock_page(struct page *__page)
1590{
1591 struct page *page = compound_head(__page);
1592 wait_queue_head_t *q = page_waitqueue(page);
1593 wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE,
1594 EXCLUSIVE);
1595}
1596EXPORT_SYMBOL(__lock_page);
1597
1598int __lock_page_killable(struct page *__page)
1599{
1600 struct page *page = compound_head(__page);
1601 wait_queue_head_t *q = page_waitqueue(page);
1602 return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE,
1603 EXCLUSIVE);
1604}
1605EXPORT_SYMBOL_GPL(__lock_page_killable);
1606
1607int __lock_page_async(struct page *page, struct wait_page_queue *wait)
1608{
1609 struct wait_queue_head *q = page_waitqueue(page);
1610 int ret = 0;
1611
1612 wait->page = page;
1613 wait->bit_nr = PG_locked;
1614
1615 spin_lock_irq(&q->lock);
1616 __add_wait_queue_entry_tail(q, &wait->wait);
1617 SetPageWaiters(page);
1618 ret = !trylock_page(page);
1619 /*
1620 * If we were successful now, we know we're still on the
1621 * waitqueue as we're still under the lock. This means it's
1622 * safe to remove and return success, we know the callback
1623 * isn't going to trigger.
1624 */
1625 if (!ret)
1626 __remove_wait_queue(q, &wait->wait);
1627 else
1628 ret = -EIOCBQUEUED;
1629 spin_unlock_irq(&q->lock);
1630 return ret;
1631}
1632
1633/*
1634 * Return values:
1635 * 1 - page is locked; mmap_lock is still held.
1636 * 0 - page is not locked.
1637 * mmap_lock has been released (mmap_read_unlock(), unless flags had both
1638 * FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1639 * which case mmap_lock is still held.
1640 *
1641 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1642 * with the page locked and the mmap_lock unperturbed.
1643 */
1644int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1645 unsigned int flags)
1646{
1647 if (fault_flag_allow_retry_first(flags)) {
1648 /*
1649 * CAUTION! In this case, mmap_lock is not released
1650 * even though return 0.
1651 */
1652 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1653 return 0;
1654
1655 mmap_read_unlock(mm);
1656 if (flags & FAULT_FLAG_KILLABLE)
1657 wait_on_page_locked_killable(page);
1658 else
1659 wait_on_page_locked(page);
1660 return 0;
1661 }
1662 if (flags & FAULT_FLAG_KILLABLE) {
1663 int ret;
1664
1665 ret = __lock_page_killable(page);
1666 if (ret) {
1667 mmap_read_unlock(mm);
1668 return 0;
1669 }
1670 } else {
1671 __lock_page(page);
1672 }
1673 return 1;
1674
1675}
1676
1677/**
1678 * page_cache_next_miss() - Find the next gap in the page cache.
1679 * @mapping: Mapping.
1680 * @index: Index.
1681 * @max_scan: Maximum range to search.
1682 *
1683 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1684 * gap with the lowest index.
1685 *
1686 * This function may be called under the rcu_read_lock. However, this will
1687 * not atomically search a snapshot of the cache at a single point in time.
1688 * For example, if a gap is created at index 5, then subsequently a gap is
1689 * created at index 10, page_cache_next_miss covering both indices may
1690 * return 10 if called under the rcu_read_lock.
1691 *
1692 * Return: The index of the gap if found, otherwise an index outside the
1693 * range specified (in which case 'return - index >= max_scan' will be true).
1694 * In the rare case of index wrap-around, 0 will be returned.
1695 */
1696pgoff_t page_cache_next_miss(struct address_space *mapping,
1697 pgoff_t index, unsigned long max_scan)
1698{
1699 XA_STATE(xas, &mapping->i_pages, index);
1700
1701 while (max_scan--) {
1702 void *entry = xas_next(&xas);
1703 if (!entry || xa_is_value(entry))
1704 break;
1705 if (xas.xa_index == 0)
1706 break;
1707 }
1708
1709 return xas.xa_index;
1710}
1711EXPORT_SYMBOL(page_cache_next_miss);
1712
1713/**
1714 * page_cache_prev_miss() - Find the previous gap in the page cache.
1715 * @mapping: Mapping.
1716 * @index: Index.
1717 * @max_scan: Maximum range to search.
1718 *
1719 * Search the range [max(index - max_scan + 1, 0), index] for the
1720 * gap with the highest index.
1721 *
1722 * This function may be called under the rcu_read_lock. However, this will
1723 * not atomically search a snapshot of the cache at a single point in time.
1724 * For example, if a gap is created at index 10, then subsequently a gap is
1725 * created at index 5, page_cache_prev_miss() covering both indices may
1726 * return 5 if called under the rcu_read_lock.
1727 *
1728 * Return: The index of the gap if found, otherwise an index outside the
1729 * range specified (in which case 'index - return >= max_scan' will be true).
1730 * In the rare case of wrap-around, ULONG_MAX will be returned.
1731 */
1732pgoff_t page_cache_prev_miss(struct address_space *mapping,
1733 pgoff_t index, unsigned long max_scan)
1734{
1735 XA_STATE(xas, &mapping->i_pages, index);
1736
1737 while (max_scan--) {
1738 void *entry = xas_prev(&xas);
1739 if (!entry || xa_is_value(entry))
1740 break;
1741 if (xas.xa_index == ULONG_MAX)
1742 break;
1743 }
1744
1745 return xas.xa_index;
1746}
1747EXPORT_SYMBOL(page_cache_prev_miss);
1748
1749/*
1750 * mapping_get_entry - Get a page cache entry.
1751 * @mapping: the address_space to search
1752 * @index: The page cache index.
1753 *
1754 * Looks up the page cache slot at @mapping & @index. If there is a
1755 * page cache page, the head page is returned with an increased refcount.
1756 *
1757 * If the slot holds a shadow entry of a previously evicted page, or a
1758 * swap entry from shmem/tmpfs, it is returned.
1759 *
1760 * Return: The head page or shadow entry, %NULL if nothing is found.
1761 */
1762static struct page *mapping_get_entry(struct address_space *mapping,
1763 pgoff_t index)
1764{
1765 XA_STATE(xas, &mapping->i_pages, index);
1766 struct page *page;
1767
1768 rcu_read_lock();
1769repeat:
1770 xas_reset(&xas);
1771 page = xas_load(&xas);
1772 if (xas_retry(&xas, page))
1773 goto repeat;
1774 /*
1775 * A shadow entry of a recently evicted page, or a swap entry from
1776 * shmem/tmpfs. Return it without attempting to raise page count.
1777 */
1778 if (!page || xa_is_value(page))
1779 goto out;
1780
1781 if (!page_cache_get_speculative(page))
1782 goto repeat;
1783
1784 /*
1785 * Has the page moved or been split?
1786 * This is part of the lockless pagecache protocol. See
1787 * include/linux/pagemap.h for details.
1788 */
1789 if (unlikely(page != xas_reload(&xas))) {
1790 put_page(page);
1791 goto repeat;
1792 }
1793out:
1794 rcu_read_unlock();
1795
1796 return page;
1797}
1798
1799/**
1800 * pagecache_get_page - Find and get a reference to a page.
1801 * @mapping: The address_space to search.
1802 * @index: The page index.
1803 * @fgp_flags: %FGP flags modify how the page is returned.
1804 * @gfp_mask: Memory allocation flags to use if %FGP_CREAT is specified.
1805 *
1806 * Looks up the page cache entry at @mapping & @index.
1807 *
1808 * @fgp_flags can be zero or more of these flags:
1809 *
1810 * * %FGP_ACCESSED - The page will be marked accessed.
1811 * * %FGP_LOCK - The page is returned locked.
1812 * * %FGP_HEAD - If the page is present and a THP, return the head page
1813 * rather than the exact page specified by the index.
1814 * * %FGP_ENTRY - If there is a shadow / swap / DAX entry, return it
1815 * instead of allocating a new page to replace it.
1816 * * %FGP_CREAT - If no page is present then a new page is allocated using
1817 * @gfp_mask and added to the page cache and the VM's LRU list.
1818 * The page is returned locked and with an increased refcount.
1819 * * %FGP_FOR_MMAP - The caller wants to do its own locking dance if the
1820 * page is already in cache. If the page was allocated, unlock it before
1821 * returning so the caller can do the same dance.
1822 * * %FGP_WRITE - The page will be written
1823 * * %FGP_NOFS - __GFP_FS will get cleared in gfp mask
1824 * * %FGP_NOWAIT - Don't get blocked by page lock
1825 *
1826 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1827 * if the %GFP flags specified for %FGP_CREAT are atomic.
1828 *
1829 * If there is a page cache page, it is returned with an increased refcount.
1830 *
1831 * Return: The found page or %NULL otherwise.
1832 */
1833struct page *pagecache_get_page(struct address_space *mapping, pgoff_t index,
1834 int fgp_flags, gfp_t gfp_mask)
1835{
1836 struct page *page;
1837
1838repeat:
1839 page = mapping_get_entry(mapping, index);
1840 if (xa_is_value(page)) {
1841 if (fgp_flags & FGP_ENTRY)
1842 return page;
1843 page = NULL;
1844 }
1845 if (!page)
1846 goto no_page;
1847
1848 if (fgp_flags & FGP_LOCK) {
1849 if (fgp_flags & FGP_NOWAIT) {
1850 if (!trylock_page(page)) {
1851 put_page(page);
1852 return NULL;
1853 }
1854 } else {
1855 lock_page(page);
1856 }
1857
1858 /* Has the page been truncated? */
1859 if (unlikely(page->mapping != mapping)) {
1860 unlock_page(page);
1861 put_page(page);
1862 goto repeat;
1863 }
1864 VM_BUG_ON_PAGE(!thp_contains(page, index), page);
1865 }
1866
1867 if (fgp_flags & FGP_ACCESSED)
1868 mark_page_accessed(page);
1869 else if (fgp_flags & FGP_WRITE) {
1870 /* Clear idle flag for buffer write */
1871 if (page_is_idle(page))
1872 clear_page_idle(page);
1873 }
1874 if (!(fgp_flags & FGP_HEAD))
1875 page = find_subpage(page, index);
1876
1877no_page:
1878 if (!page && (fgp_flags & FGP_CREAT)) {
1879 int err;
1880 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1881 gfp_mask |= __GFP_WRITE;
1882 if (fgp_flags & FGP_NOFS)
1883 gfp_mask &= ~__GFP_FS;
1884
1885 page = __page_cache_alloc(gfp_mask);
1886 if (!page)
1887 return NULL;
1888
1889 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1890 fgp_flags |= FGP_LOCK;
1891
1892 /* Init accessed so avoid atomic mark_page_accessed later */
1893 if (fgp_flags & FGP_ACCESSED)
1894 __SetPageReferenced(page);
1895
1896 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
1897 if (unlikely(err)) {
1898 put_page(page);
1899 page = NULL;
1900 if (err == -EEXIST)
1901 goto repeat;
1902 }
1903
1904 /*
1905 * add_to_page_cache_lru locks the page, and for mmap we expect
1906 * an unlocked page.
1907 */
1908 if (page && (fgp_flags & FGP_FOR_MMAP))
1909 unlock_page(page);
1910 }
1911
1912 return page;
1913}
1914EXPORT_SYMBOL(pagecache_get_page);
1915
1916static inline struct page *find_get_entry(struct xa_state *xas, pgoff_t max,
1917 xa_mark_t mark)
1918{
1919 struct page *page;
1920
1921retry:
1922 if (mark == XA_PRESENT)
1923 page = xas_find(xas, max);
1924 else
1925 page = xas_find_marked(xas, max, mark);
1926
1927 if (xas_retry(xas, page))
1928 goto retry;
1929 /*
1930 * A shadow entry of a recently evicted page, a swap
1931 * entry from shmem/tmpfs or a DAX entry. Return it
1932 * without attempting to raise page count.
1933 */
1934 if (!page || xa_is_value(page))
1935 return page;
1936
1937 if (!page_cache_get_speculative(page))
1938 goto reset;
1939
1940 /* Has the page moved or been split? */
1941 if (unlikely(page != xas_reload(xas))) {
1942 put_page(page);
1943 goto reset;
1944 }
1945
1946 return page;
1947reset:
1948 xas_reset(xas);
1949 goto retry;
1950}
1951
1952/**
1953 * find_get_entries - gang pagecache lookup
1954 * @mapping: The address_space to search
1955 * @start: The starting page cache index
1956 * @end: The final page index (inclusive).
1957 * @pvec: Where the resulting entries are placed.
1958 * @indices: The cache indices corresponding to the entries in @entries
1959 *
1960 * find_get_entries() will search for and return a batch of entries in
1961 * the mapping. The entries are placed in @pvec. find_get_entries()
1962 * takes a reference on any actual pages it returns.
1963 *
1964 * The search returns a group of mapping-contiguous page cache entries
1965 * with ascending indexes. There may be holes in the indices due to
1966 * not-present pages.
1967 *
1968 * Any shadow entries of evicted pages, or swap entries from
1969 * shmem/tmpfs, are included in the returned array.
1970 *
1971 * If it finds a Transparent Huge Page, head or tail, find_get_entries()
1972 * stops at that page: the caller is likely to have a better way to handle
1973 * the compound page as a whole, and then skip its extent, than repeatedly
1974 * calling find_get_entries() to return all its tails.
1975 *
1976 * Return: the number of pages and shadow entries which were found.
1977 */
1978unsigned find_get_entries(struct address_space *mapping, pgoff_t start,
1979 pgoff_t end, struct pagevec *pvec, pgoff_t *indices)
1980{
1981 XA_STATE(xas, &mapping->i_pages, start);
1982 struct page *page;
1983 unsigned int ret = 0;
1984 unsigned nr_entries = PAGEVEC_SIZE;
1985
1986 rcu_read_lock();
1987 while ((page = find_get_entry(&xas, end, XA_PRESENT))) {
1988 /*
1989 * Terminate early on finding a THP, to allow the caller to
1990 * handle it all at once; but continue if this is hugetlbfs.
1991 */
1992 if (!xa_is_value(page) && PageTransHuge(page) &&
1993 !PageHuge(page)) {
1994 page = find_subpage(page, xas.xa_index);
1995 nr_entries = ret + 1;
1996 }
1997
1998 indices[ret] = xas.xa_index;
1999 pvec->pages[ret] = page;
2000 if (++ret == nr_entries)
2001 break;
2002 }
2003 rcu_read_unlock();
2004
2005 pvec->nr = ret;
2006 return ret;
2007}
2008
2009/**
2010 * find_lock_entries - Find a batch of pagecache entries.
2011 * @mapping: The address_space to search.
2012 * @start: The starting page cache index.
2013 * @end: The final page index (inclusive).
2014 * @pvec: Where the resulting entries are placed.
2015 * @indices: The cache indices of the entries in @pvec.
2016 *
2017 * find_lock_entries() will return a batch of entries from @mapping.
2018 * Swap, shadow and DAX entries are included. Pages are returned
2019 * locked and with an incremented refcount. Pages which are locked by
2020 * somebody else or under writeback are skipped. Only the head page of
2021 * a THP is returned. Pages which are partially outside the range are
2022 * not returned.
2023 *
2024 * The entries have ascending indexes. The indices may not be consecutive
2025 * due to not-present entries, THP pages, pages which could not be locked
2026 * or pages under writeback.
2027 *
2028 * Return: The number of entries which were found.
2029 */
2030unsigned find_lock_entries(struct address_space *mapping, pgoff_t start,
2031 pgoff_t end, struct pagevec *pvec, pgoff_t *indices)
2032{
2033 XA_STATE(xas, &mapping->i_pages, start);
2034 struct page *page;
2035
2036 rcu_read_lock();
2037 while ((page = find_get_entry(&xas, end, XA_PRESENT))) {
2038 if (!xa_is_value(page)) {
2039 if (page->index < start)
2040 goto put;
2041 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2042 if (page->index + thp_nr_pages(page) - 1 > end)
2043 goto put;
2044 if (!trylock_page(page))
2045 goto put;
2046 if (page->mapping != mapping || PageWriteback(page))
2047 goto unlock;
2048 VM_BUG_ON_PAGE(!thp_contains(page, xas.xa_index),
2049 page);
2050 }
2051 indices[pvec->nr] = xas.xa_index;
2052 if (!pagevec_add(pvec, page))
2053 break;
2054 goto next;
2055unlock:
2056 unlock_page(page);
2057put:
2058 put_page(page);
2059next:
2060 if (!xa_is_value(page) && PageTransHuge(page)) {
2061 unsigned int nr_pages = thp_nr_pages(page);
2062
2063 /* Final THP may cross MAX_LFS_FILESIZE on 32-bit */
2064 xas_set(&xas, page->index + nr_pages);
2065 if (xas.xa_index < nr_pages)
2066 break;
2067 }
2068 }
2069 rcu_read_unlock();
2070
2071 return pagevec_count(pvec);
2072}
2073
2074/**
2075 * find_get_pages_range - gang pagecache lookup
2076 * @mapping: The address_space to search
2077 * @start: The starting page index
2078 * @end: The final page index (inclusive)
2079 * @nr_pages: The maximum number of pages
2080 * @pages: Where the resulting pages are placed
2081 *
2082 * find_get_pages_range() will search for and return a group of up to @nr_pages
2083 * pages in the mapping starting at index @start and up to index @end
2084 * (inclusive). The pages are placed at @pages. find_get_pages_range() takes
2085 * a reference against the returned pages.
2086 *
2087 * The search returns a group of mapping-contiguous pages with ascending
2088 * indexes. There may be holes in the indices due to not-present pages.
2089 * We also update @start to index the next page for the traversal.
2090 *
2091 * Return: the number of pages which were found. If this number is
2092 * smaller than @nr_pages, the end of specified range has been
2093 * reached.
2094 */
2095unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
2096 pgoff_t end, unsigned int nr_pages,
2097 struct page **pages)
2098{
2099 XA_STATE(xas, &mapping->i_pages, *start);
2100 struct page *page;
2101 unsigned ret = 0;
2102
2103 if (unlikely(!nr_pages))
2104 return 0;
2105
2106 rcu_read_lock();
2107 while ((page = find_get_entry(&xas, end, XA_PRESENT))) {
2108 /* Skip over shadow, swap and DAX entries */
2109 if (xa_is_value(page))
2110 continue;
2111
2112 pages[ret] = find_subpage(page, xas.xa_index);
2113 if (++ret == nr_pages) {
2114 *start = xas.xa_index + 1;
2115 goto out;
2116 }
2117 }
2118
2119 /*
2120 * We come here when there is no page beyond @end. We take care to not
2121 * overflow the index @start as it confuses some of the callers. This
2122 * breaks the iteration when there is a page at index -1 but that is
2123 * already broken anyway.
2124 */
2125 if (end == (pgoff_t)-1)
2126 *start = (pgoff_t)-1;
2127 else
2128 *start = end + 1;
2129out:
2130 rcu_read_unlock();
2131
2132 return ret;
2133}
2134
2135/**
2136 * find_get_pages_contig - gang contiguous pagecache lookup
2137 * @mapping: The address_space to search
2138 * @index: The starting page index
2139 * @nr_pages: The maximum number of pages
2140 * @pages: Where the resulting pages are placed
2141 *
2142 * find_get_pages_contig() works exactly like find_get_pages(), except
2143 * that the returned number of pages are guaranteed to be contiguous.
2144 *
2145 * Return: the number of pages which were found.
2146 */
2147unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
2148 unsigned int nr_pages, struct page **pages)
2149{
2150 XA_STATE(xas, &mapping->i_pages, index);
2151 struct page *page;
2152 unsigned int ret = 0;
2153
2154 if (unlikely(!nr_pages))
2155 return 0;
2156
2157 rcu_read_lock();
2158 for (page = xas_load(&xas); page; page = xas_next(&xas)) {
2159 if (xas_retry(&xas, page))
2160 continue;
2161 /*
2162 * If the entry has been swapped out, we can stop looking.
2163 * No current caller is looking for DAX entries.
2164 */
2165 if (xa_is_value(page))
2166 break;
2167
2168 if (!page_cache_get_speculative(page))
2169 goto retry;
2170
2171 /* Has the page moved or been split? */
2172 if (unlikely(page != xas_reload(&xas)))
2173 goto put_page;
2174
2175 pages[ret] = find_subpage(page, xas.xa_index);
2176 if (++ret == nr_pages)
2177 break;
2178 continue;
2179put_page:
2180 put_page(page);
2181retry:
2182 xas_reset(&xas);
2183 }
2184 rcu_read_unlock();
2185 return ret;
2186}
2187EXPORT_SYMBOL(find_get_pages_contig);
2188
2189/**
2190 * find_get_pages_range_tag - Find and return head pages matching @tag.
2191 * @mapping: the address_space to search
2192 * @index: the starting page index
2193 * @end: The final page index (inclusive)
2194 * @tag: the tag index
2195 * @nr_pages: the maximum number of pages
2196 * @pages: where the resulting pages are placed
2197 *
2198 * Like find_get_pages(), except we only return head pages which are tagged
2199 * with @tag. @index is updated to the index immediately after the last
2200 * page we return, ready for the next iteration.
2201 *
2202 * Return: the number of pages which were found.
2203 */
2204unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
2205 pgoff_t end, xa_mark_t tag, unsigned int nr_pages,
2206 struct page **pages)
2207{
2208 XA_STATE(xas, &mapping->i_pages, *index);
2209 struct page *page;
2210 unsigned ret = 0;
2211
2212 if (unlikely(!nr_pages))
2213 return 0;
2214
2215 rcu_read_lock();
2216 while ((page = find_get_entry(&xas, end, tag))) {
2217 /*
2218 * Shadow entries should never be tagged, but this iteration
2219 * is lockless so there is a window for page reclaim to evict
2220 * a page we saw tagged. Skip over it.
2221 */
2222 if (xa_is_value(page))
2223 continue;
2224
2225 pages[ret] = page;
2226 if (++ret == nr_pages) {
2227 *index = page->index + thp_nr_pages(page);
2228 goto out;
2229 }
2230 }
2231
2232 /*
2233 * We come here when we got to @end. We take care to not overflow the
2234 * index @index as it confuses some of the callers. This breaks the
2235 * iteration when there is a page at index -1 but that is already
2236 * broken anyway.
2237 */
2238 if (end == (pgoff_t)-1)
2239 *index = (pgoff_t)-1;
2240 else
2241 *index = end + 1;
2242out:
2243 rcu_read_unlock();
2244
2245 return ret;
2246}
2247EXPORT_SYMBOL(find_get_pages_range_tag);
2248
2249/*
2250 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2251 * a _large_ part of the i/o request. Imagine the worst scenario:
2252 *
2253 * ---R__________________________________________B__________
2254 * ^ reading here ^ bad block(assume 4k)
2255 *
2256 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2257 * => failing the whole request => read(R) => read(R+1) =>
2258 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2259 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2260 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2261 *
2262 * It is going insane. Fix it by quickly scaling down the readahead size.
2263 */
2264static void shrink_readahead_size_eio(struct file_ra_state *ra)
2265{
2266 ra->ra_pages /= 4;
2267}
2268
2269/*
2270 * filemap_get_read_batch - Get a batch of pages for read
2271 *
2272 * Get a batch of pages which represent a contiguous range of bytes
2273 * in the file. No tail pages will be returned. If @index is in the
2274 * middle of a THP, the entire THP will be returned. The last page in
2275 * the batch may have Readahead set or be not Uptodate so that the
2276 * caller can take the appropriate action.
2277 */
2278static void filemap_get_read_batch(struct address_space *mapping,
2279 pgoff_t index, pgoff_t max, struct pagevec *pvec)
2280{
2281 XA_STATE(xas, &mapping->i_pages, index);
2282 struct page *head;
2283
2284 rcu_read_lock();
2285 for (head = xas_load(&xas); head; head = xas_next(&xas)) {
2286 if (xas_retry(&xas, head))
2287 continue;
2288 if (xas.xa_index > max || xa_is_value(head))
2289 break;
2290 if (!page_cache_get_speculative(head))
2291 goto retry;
2292
2293 /* Has the page moved or been split? */
2294 if (unlikely(head != xas_reload(&xas)))
2295 goto put_page;
2296
2297 if (!pagevec_add(pvec, head))
2298 break;
2299 if (!PageUptodate(head))
2300 break;
2301 if (PageReadahead(head))
2302 break;
2303 xas.xa_index = head->index + thp_nr_pages(head) - 1;
2304 xas.xa_offset = (xas.xa_index >> xas.xa_shift) & XA_CHUNK_MASK;
2305 continue;
2306put_page:
2307 put_page(head);
2308retry:
2309 xas_reset(&xas);
2310 }
2311 rcu_read_unlock();
2312}
2313
2314static int filemap_read_page(struct file *file, struct address_space *mapping,
2315 struct page *page)
2316{
2317 int error;
2318
2319 /*
2320 * A previous I/O error may have been due to temporary failures,
2321 * eg. multipath errors. PG_error will be set again if readpage
2322 * fails.
2323 */
2324 ClearPageError(page);
2325 /* Start the actual read. The read will unlock the page. */
2326 error = mapping->a_ops->readpage(file, page);
2327 if (error)
2328 return error;
2329
2330 error = wait_on_page_locked_killable(page);
2331 if (error)
2332 return error;
2333 if (PageUptodate(page))
2334 return 0;
2335 shrink_readahead_size_eio(&file->f_ra);
2336 return -EIO;
2337}
2338
2339static bool filemap_range_uptodate(struct address_space *mapping,
2340 loff_t pos, struct iov_iter *iter, struct page *page)
2341{
2342 int count;
2343
2344 if (PageUptodate(page))
2345 return true;
2346 /* pipes can't handle partially uptodate pages */
2347 if (iov_iter_is_pipe(iter))
2348 return false;
2349 if (!mapping->a_ops->is_partially_uptodate)
2350 return false;
2351 if (mapping->host->i_blkbits >= (PAGE_SHIFT + thp_order(page)))
2352 return false;
2353
2354 count = iter->count;
2355 if (page_offset(page) > pos) {
2356 count -= page_offset(page) - pos;
2357 pos = 0;
2358 } else {
2359 pos -= page_offset(page);
2360 }
2361
2362 return mapping->a_ops->is_partially_uptodate(page, pos, count);
2363}
2364
2365static int filemap_update_page(struct kiocb *iocb,
2366 struct address_space *mapping, struct iov_iter *iter,
2367 struct page *page)
2368{
2369 int error;
2370
2371 if (!trylock_page(page)) {
2372 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2373 return -EAGAIN;
2374 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2375 put_and_wait_on_page_locked(page, TASK_KILLABLE);
2376 return AOP_TRUNCATED_PAGE;
2377 }
2378 error = __lock_page_async(page, iocb->ki_waitq);
2379 if (error)
2380 return error;
2381 }
2382
2383 if (!page->mapping)
2384 goto truncated;
2385
2386 error = 0;
2387 if (filemap_range_uptodate(mapping, iocb->ki_pos, iter, page))
2388 goto unlock;
2389
2390 error = -EAGAIN;
2391 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2392 goto unlock;
2393
2394 error = filemap_read_page(iocb->ki_filp, mapping, page);
2395 if (error == AOP_TRUNCATED_PAGE)
2396 put_page(page);
2397 return error;
2398truncated:
2399 unlock_page(page);
2400 put_page(page);
2401 return AOP_TRUNCATED_PAGE;
2402unlock:
2403 unlock_page(page);
2404 return error;
2405}
2406
2407static int filemap_create_page(struct file *file,
2408 struct address_space *mapping, pgoff_t index,
2409 struct pagevec *pvec)
2410{
2411 struct page *page;
2412 int error;
2413
2414 page = page_cache_alloc(mapping);
2415 if (!page)
2416 return -ENOMEM;
2417
2418 error = add_to_page_cache_lru(page, mapping, index,
2419 mapping_gfp_constraint(mapping, GFP_KERNEL));
2420 if (error == -EEXIST)
2421 error = AOP_TRUNCATED_PAGE;
2422 if (error)
2423 goto error;
2424
2425 error = filemap_read_page(file, mapping, page);
2426 if (error)
2427 goto error;
2428
2429 pagevec_add(pvec, page);
2430 return 0;
2431error:
2432 put_page(page);
2433 return error;
2434}
2435
2436static int filemap_readahead(struct kiocb *iocb, struct file *file,
2437 struct address_space *mapping, struct page *page,
2438 pgoff_t last_index)
2439{
2440 if (iocb->ki_flags & IOCB_NOIO)
2441 return -EAGAIN;
2442 page_cache_async_readahead(mapping, &file->f_ra, file, page,
2443 page->index, last_index - page->index);
2444 return 0;
2445}
2446
2447static int filemap_get_pages(struct kiocb *iocb, struct iov_iter *iter,
2448 struct pagevec *pvec)
2449{
2450 struct file *filp = iocb->ki_filp;
2451 struct address_space *mapping = filp->f_mapping;
2452 struct file_ra_state *ra = &filp->f_ra;
2453 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2454 pgoff_t last_index;
2455 struct page *page;
2456 int err = 0;
2457
2458 last_index = DIV_ROUND_UP(iocb->ki_pos + iter->count, PAGE_SIZE);
2459retry:
2460 if (fatal_signal_pending(current))
2461 return -EINTR;
2462
2463 filemap_get_read_batch(mapping, index, last_index, pvec);
2464 if (!pagevec_count(pvec)) {
2465 if (iocb->ki_flags & IOCB_NOIO)
2466 return -EAGAIN;
2467 page_cache_sync_readahead(mapping, ra, filp, index,
2468 last_index - index);
2469 filemap_get_read_batch(mapping, index, last_index, pvec);
2470 }
2471 if (!pagevec_count(pvec)) {
2472 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2473 return -EAGAIN;
2474 err = filemap_create_page(filp, mapping,
2475 iocb->ki_pos >> PAGE_SHIFT, pvec);
2476 if (err == AOP_TRUNCATED_PAGE)
2477 goto retry;
2478 return err;
2479 }
2480
2481 page = pvec->pages[pagevec_count(pvec) - 1];
2482 if (PageReadahead(page)) {
2483 err = filemap_readahead(iocb, filp, mapping, page, last_index);
2484 if (err)
2485 goto err;
2486 }
2487 if (!PageUptodate(page)) {
2488 if ((iocb->ki_flags & IOCB_WAITQ) && pagevec_count(pvec) > 1)
2489 iocb->ki_flags |= IOCB_NOWAIT;
2490 err = filemap_update_page(iocb, mapping, iter, page);
2491 if (err)
2492 goto err;
2493 }
2494
2495 return 0;
2496err:
2497 if (err < 0)
2498 put_page(page);
2499 if (likely(--pvec->nr))
2500 return 0;
2501 if (err == AOP_TRUNCATED_PAGE)
2502 goto retry;
2503 return err;
2504}
2505
2506/**
2507 * filemap_read - Read data from the page cache.
2508 * @iocb: The iocb to read.
2509 * @iter: Destination for the data.
2510 * @already_read: Number of bytes already read by the caller.
2511 *
2512 * Copies data from the page cache. If the data is not currently present,
2513 * uses the readahead and readpage address_space operations to fetch it.
2514 *
2515 * Return: Total number of bytes copied, including those already read by
2516 * the caller. If an error happens before any bytes are copied, returns
2517 * a negative error number.
2518 */
2519ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2520 ssize_t already_read)
2521{
2522 struct file *filp = iocb->ki_filp;
2523 struct file_ra_state *ra = &filp->f_ra;
2524 struct address_space *mapping = filp->f_mapping;
2525 struct inode *inode = mapping->host;
2526 struct pagevec pvec;
2527 int i, error = 0;
2528 bool writably_mapped;
2529 loff_t isize, end_offset;
2530
2531 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2532 return 0;
2533 if (unlikely(!iov_iter_count(iter)))
2534 return 0;
2535
2536 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2537 pagevec_init(&pvec);
2538
2539 do {
2540 cond_resched();
2541
2542 /*
2543 * If we've already successfully copied some data, then we
2544 * can no longer safely return -EIOCBQUEUED. Hence mark
2545 * an async read NOWAIT at that point.
2546 */
2547 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2548 iocb->ki_flags |= IOCB_NOWAIT;
2549
2550 error = filemap_get_pages(iocb, iter, &pvec);
2551 if (error < 0)
2552 break;
2553
2554 /*
2555 * i_size must be checked after we know the pages are Uptodate.
2556 *
2557 * Checking i_size after the check allows us to calculate
2558 * the correct value for "nr", which means the zero-filled
2559 * part of the page is not copied back to userspace (unless
2560 * another truncate extends the file - this is desired though).
2561 */
2562 isize = i_size_read(inode);
2563 if (unlikely(iocb->ki_pos >= isize))
2564 goto put_pages;
2565 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2566
2567 /*
2568 * Once we start copying data, we don't want to be touching any
2569 * cachelines that might be contended:
2570 */
2571 writably_mapped = mapping_writably_mapped(mapping);
2572
2573 /*
2574 * When a sequential read accesses a page several times, only
2575 * mark it as accessed the first time.
2576 */
2577 if (iocb->ki_pos >> PAGE_SHIFT !=
2578 ra->prev_pos >> PAGE_SHIFT)
2579 mark_page_accessed(pvec.pages[0]);
2580
2581 for (i = 0; i < pagevec_count(&pvec); i++) {
2582 struct page *page = pvec.pages[i];
2583 size_t page_size = thp_size(page);
2584 size_t offset = iocb->ki_pos & (page_size - 1);
2585 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2586 page_size - offset);
2587 size_t copied;
2588
2589 if (end_offset < page_offset(page))
2590 break;
2591 if (i > 0)
2592 mark_page_accessed(page);
2593 /*
2594 * If users can be writing to this page using arbitrary
2595 * virtual addresses, take care about potential aliasing
2596 * before reading the page on the kernel side.
2597 */
2598 if (writably_mapped) {
2599 int j;
2600
2601 for (j = 0; j < thp_nr_pages(page); j++)
2602 flush_dcache_page(page + j);
2603 }
2604
2605 copied = copy_page_to_iter(page, offset, bytes, iter);
2606
2607 already_read += copied;
2608 iocb->ki_pos += copied;
2609 ra->prev_pos = iocb->ki_pos;
2610
2611 if (copied < bytes) {
2612 error = -EFAULT;
2613 break;
2614 }
2615 }
2616put_pages:
2617 for (i = 0; i < pagevec_count(&pvec); i++)
2618 put_page(pvec.pages[i]);
2619 pagevec_reinit(&pvec);
2620 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2621
2622 file_accessed(filp);
2623
2624 return already_read ? already_read : error;
2625}
2626EXPORT_SYMBOL_GPL(filemap_read);
2627
2628/**
2629 * generic_file_read_iter - generic filesystem read routine
2630 * @iocb: kernel I/O control block
2631 * @iter: destination for the data read
2632 *
2633 * This is the "read_iter()" routine for all filesystems
2634 * that can use the page cache directly.
2635 *
2636 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2637 * be returned when no data can be read without waiting for I/O requests
2638 * to complete; it doesn't prevent readahead.
2639 *
2640 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2641 * requests shall be made for the read or for readahead. When no data
2642 * can be read, -EAGAIN shall be returned. When readahead would be
2643 * triggered, a partial, possibly empty read shall be returned.
2644 *
2645 * Return:
2646 * * number of bytes copied, even for partial reads
2647 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2648 */
2649ssize_t
2650generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2651{
2652 size_t count = iov_iter_count(iter);
2653 ssize_t retval = 0;
2654
2655 if (!count)
2656 return 0; /* skip atime */
2657
2658 if (iocb->ki_flags & IOCB_DIRECT) {
2659 struct file *file = iocb->ki_filp;
2660 struct address_space *mapping = file->f_mapping;
2661 struct inode *inode = mapping->host;
2662 loff_t size;
2663
2664 size = i_size_read(inode);
2665 if (iocb->ki_flags & IOCB_NOWAIT) {
2666 if (filemap_range_needs_writeback(mapping, iocb->ki_pos,
2667 iocb->ki_pos + count - 1))
2668 return -EAGAIN;
2669 } else {
2670 retval = filemap_write_and_wait_range(mapping,
2671 iocb->ki_pos,
2672 iocb->ki_pos + count - 1);
2673 if (retval < 0)
2674 return retval;
2675 }
2676
2677 file_accessed(file);
2678
2679 retval = mapping->a_ops->direct_IO(iocb, iter);
2680 if (retval >= 0) {
2681 iocb->ki_pos += retval;
2682 count -= retval;
2683 }
2684 if (retval != -EIOCBQUEUED)
2685 iov_iter_revert(iter, count - iov_iter_count(iter));
2686
2687 /*
2688 * Btrfs can have a short DIO read if we encounter
2689 * compressed extents, so if there was an error, or if
2690 * we've already read everything we wanted to, or if
2691 * there was a short read because we hit EOF, go ahead
2692 * and return. Otherwise fallthrough to buffered io for
2693 * the rest of the read. Buffered reads will not work for
2694 * DAX files, so don't bother trying.
2695 */
2696 if (retval < 0 || !count || iocb->ki_pos >= size ||
2697 IS_DAX(inode))
2698 return retval;
2699 }
2700
2701 return filemap_read(iocb, iter, retval);
2702}
2703EXPORT_SYMBOL(generic_file_read_iter);
2704
2705static inline loff_t page_seek_hole_data(struct xa_state *xas,
2706 struct address_space *mapping, struct page *page,
2707 loff_t start, loff_t end, bool seek_data)
2708{
2709 const struct address_space_operations *ops = mapping->a_ops;
2710 size_t offset, bsz = i_blocksize(mapping->host);
2711
2712 if (xa_is_value(page) || PageUptodate(page))
2713 return seek_data ? start : end;
2714 if (!ops->is_partially_uptodate)
2715 return seek_data ? end : start;
2716
2717 xas_pause(xas);
2718 rcu_read_unlock();
2719 lock_page(page);
2720 if (unlikely(page->mapping != mapping))
2721 goto unlock;
2722
2723 offset = offset_in_thp(page, start) & ~(bsz - 1);
2724
2725 do {
2726 if (ops->is_partially_uptodate(page, offset, bsz) == seek_data)
2727 break;
2728 start = (start + bsz) & ~(bsz - 1);
2729 offset += bsz;
2730 } while (offset < thp_size(page));
2731unlock:
2732 unlock_page(page);
2733 rcu_read_lock();
2734 return start;
2735}
2736
2737static inline
2738unsigned int seek_page_size(struct xa_state *xas, struct page *page)
2739{
2740 if (xa_is_value(page))
2741 return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index);
2742 return thp_size(page);
2743}
2744
2745/**
2746 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
2747 * @mapping: Address space to search.
2748 * @start: First byte to consider.
2749 * @end: Limit of search (exclusive).
2750 * @whence: Either SEEK_HOLE or SEEK_DATA.
2751 *
2752 * If the page cache knows which blocks contain holes and which blocks
2753 * contain data, your filesystem can use this function to implement
2754 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
2755 * entirely memory-based such as tmpfs, and filesystems which support
2756 * unwritten extents.
2757 *
2758 * Return: The requested offset on success, or -ENXIO if @whence specifies
2759 * SEEK_DATA and there is no data after @start. There is an implicit hole
2760 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
2761 * and @end contain data.
2762 */
2763loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
2764 loff_t end, int whence)
2765{
2766 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
2767 pgoff_t max = (end - 1) >> PAGE_SHIFT;
2768 bool seek_data = (whence == SEEK_DATA);
2769 struct page *page;
2770
2771 if (end <= start)
2772 return -ENXIO;
2773
2774 rcu_read_lock();
2775 while ((page = find_get_entry(&xas, max, XA_PRESENT))) {
2776 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
2777 unsigned int seek_size;
2778
2779 if (start < pos) {
2780 if (!seek_data)
2781 goto unlock;
2782 start = pos;
2783 }
2784
2785 seek_size = seek_page_size(&xas, page);
2786 pos = round_up(pos + 1, seek_size);
2787 start = page_seek_hole_data(&xas, mapping, page, start, pos,
2788 seek_data);
2789 if (start < pos)
2790 goto unlock;
2791 if (start >= end)
2792 break;
2793 if (seek_size > PAGE_SIZE)
2794 xas_set(&xas, pos >> PAGE_SHIFT);
2795 if (!xa_is_value(page))
2796 put_page(page);
2797 }
2798 if (seek_data)
2799 start = -ENXIO;
2800unlock:
2801 rcu_read_unlock();
2802 if (page && !xa_is_value(page))
2803 put_page(page);
2804 if (start > end)
2805 return end;
2806 return start;
2807}
2808
2809#ifdef CONFIG_MMU
2810#define MMAP_LOTSAMISS (100)
2811/*
2812 * lock_page_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
2813 * @vmf - the vm_fault for this fault.
2814 * @page - the page to lock.
2815 * @fpin - the pointer to the file we may pin (or is already pinned).
2816 *
2817 * This works similar to lock_page_or_retry in that it can drop the mmap_lock.
2818 * It differs in that it actually returns the page locked if it returns 1 and 0
2819 * if it couldn't lock the page. If we did have to drop the mmap_lock then fpin
2820 * will point to the pinned file and needs to be fput()'ed at a later point.
2821 */
2822static int lock_page_maybe_drop_mmap(struct vm_fault *vmf, struct page *page,
2823 struct file **fpin)
2824{
2825 if (trylock_page(page))
2826 return 1;
2827
2828 /*
2829 * NOTE! This will make us return with VM_FAULT_RETRY, but with
2830 * the mmap_lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
2831 * is supposed to work. We have way too many special cases..
2832 */
2833 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
2834 return 0;
2835
2836 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
2837 if (vmf->flags & FAULT_FLAG_KILLABLE) {
2838 if (__lock_page_killable(page)) {
2839 /*
2840 * We didn't have the right flags to drop the mmap_lock,
2841 * but all fault_handlers only check for fatal signals
2842 * if we return VM_FAULT_RETRY, so we need to drop the
2843 * mmap_lock here and return 0 if we don't have a fpin.
2844 */
2845 if (*fpin == NULL)
2846 mmap_read_unlock(vmf->vma->vm_mm);
2847 return 0;
2848 }
2849 } else
2850 __lock_page(page);
2851 return 1;
2852}
2853
2854
2855/*
2856 * Synchronous readahead happens when we don't even find a page in the page
2857 * cache at all. We don't want to perform IO under the mmap sem, so if we have
2858 * to drop the mmap sem we return the file that was pinned in order for us to do
2859 * that. If we didn't pin a file then we return NULL. The file that is
2860 * returned needs to be fput()'ed when we're done with it.
2861 */
2862static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
2863{
2864 struct file *file = vmf->vma->vm_file;
2865 struct file_ra_state *ra = &file->f_ra;
2866 struct address_space *mapping = file->f_mapping;
2867 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
2868 struct file *fpin = NULL;
2869 unsigned int mmap_miss;
2870
2871 /* If we don't want any read-ahead, don't bother */
2872 if (vmf->vma->vm_flags & VM_RAND_READ)
2873 return fpin;
2874 if (!ra->ra_pages)
2875 return fpin;
2876
2877 if (vmf->vma->vm_flags & VM_SEQ_READ) {
2878 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2879 page_cache_sync_ra(&ractl, ra->ra_pages);
2880 return fpin;
2881 }
2882
2883 /* Avoid banging the cache line if not needed */
2884 mmap_miss = READ_ONCE(ra->mmap_miss);
2885 if (mmap_miss < MMAP_LOTSAMISS * 10)
2886 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
2887
2888 /*
2889 * Do we miss much more than hit in this file? If so,
2890 * stop bothering with read-ahead. It will only hurt.
2891 */
2892 if (mmap_miss > MMAP_LOTSAMISS)
2893 return fpin;
2894
2895 /*
2896 * mmap read-around
2897 */
2898 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2899 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
2900 ra->size = ra->ra_pages;
2901 ra->async_size = ra->ra_pages / 4;
2902 ractl._index = ra->start;
2903 do_page_cache_ra(&ractl, ra->size, ra->async_size);
2904 return fpin;
2905}
2906
2907/*
2908 * Asynchronous readahead happens when we find the page and PG_readahead,
2909 * so we want to possibly extend the readahead further. We return the file that
2910 * was pinned if we have to drop the mmap_lock in order to do IO.
2911 */
2912static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
2913 struct page *page)
2914{
2915 struct file *file = vmf->vma->vm_file;
2916 struct file_ra_state *ra = &file->f_ra;
2917 struct address_space *mapping = file->f_mapping;
2918 struct file *fpin = NULL;
2919 unsigned int mmap_miss;
2920 pgoff_t offset = vmf->pgoff;
2921
2922 /* If we don't want any read-ahead, don't bother */
2923 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
2924 return fpin;
2925 mmap_miss = READ_ONCE(ra->mmap_miss);
2926 if (mmap_miss)
2927 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
2928 if (PageReadahead(page)) {
2929 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2930 page_cache_async_readahead(mapping, ra, file,
2931 page, offset, ra->ra_pages);
2932 }
2933 return fpin;
2934}
2935
2936/**
2937 * filemap_fault - read in file data for page fault handling
2938 * @vmf: struct vm_fault containing details of the fault
2939 *
2940 * filemap_fault() is invoked via the vma operations vector for a
2941 * mapped memory region to read in file data during a page fault.
2942 *
2943 * The goto's are kind of ugly, but this streamlines the normal case of having
2944 * it in the page cache, and handles the special cases reasonably without
2945 * having a lot of duplicated code.
2946 *
2947 * vma->vm_mm->mmap_lock must be held on entry.
2948 *
2949 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
2950 * may be dropped before doing I/O or by lock_page_maybe_drop_mmap().
2951 *
2952 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
2953 * has not been released.
2954 *
2955 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2956 *
2957 * Return: bitwise-OR of %VM_FAULT_ codes.
2958 */
2959vm_fault_t filemap_fault(struct vm_fault *vmf)
2960{
2961 int error;
2962 struct file *file = vmf->vma->vm_file;
2963 struct file *fpin = NULL;
2964 struct address_space *mapping = file->f_mapping;
2965 struct inode *inode = mapping->host;
2966 pgoff_t offset = vmf->pgoff;
2967 pgoff_t max_off;
2968 struct page *page;
2969 vm_fault_t ret = 0;
2970
2971 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2972 if (unlikely(offset >= max_off))
2973 return VM_FAULT_SIGBUS;
2974
2975 /*
2976 * Do we have something in the page cache already?
2977 */
2978 page = find_get_page(mapping, offset);
2979 if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2980 /*
2981 * We found the page, so try async readahead before
2982 * waiting for the lock.
2983 */
2984 fpin = do_async_mmap_readahead(vmf, page);
2985 } else if (!page) {
2986 /* No page in the page cache at all */
2987 count_vm_event(PGMAJFAULT);
2988 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2989 ret = VM_FAULT_MAJOR;
2990 fpin = do_sync_mmap_readahead(vmf);
2991retry_find:
2992 page = pagecache_get_page(mapping, offset,
2993 FGP_CREAT|FGP_FOR_MMAP,
2994 vmf->gfp_mask);
2995 if (!page) {
2996 if (fpin)
2997 goto out_retry;
2998 return VM_FAULT_OOM;
2999 }
3000 }
3001
3002 if (!lock_page_maybe_drop_mmap(vmf, page, &fpin))
3003 goto out_retry;
3004
3005 /* Did it get truncated? */
3006 if (unlikely(compound_head(page)->mapping != mapping)) {
3007 unlock_page(page);
3008 put_page(page);
3009 goto retry_find;
3010 }
3011 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
3012
3013 /*
3014 * We have a locked page in the page cache, now we need to check
3015 * that it's up-to-date. If not, it is going to be due to an error.
3016 */
3017 if (unlikely(!PageUptodate(page)))
3018 goto page_not_uptodate;
3019
3020 /*
3021 * We've made it this far and we had to drop our mmap_lock, now is the
3022 * time to return to the upper layer and have it re-find the vma and
3023 * redo the fault.
3024 */
3025 if (fpin) {
3026 unlock_page(page);
3027 goto out_retry;
3028 }
3029
3030 /*
3031 * Found the page and have a reference on it.
3032 * We must recheck i_size under page lock.
3033 */
3034 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3035 if (unlikely(offset >= max_off)) {
3036 unlock_page(page);
3037 put_page(page);
3038 return VM_FAULT_SIGBUS;
3039 }
3040
3041 vmf->page = page;
3042 return ret | VM_FAULT_LOCKED;
3043
3044page_not_uptodate:
3045 /*
3046 * Umm, take care of errors if the page isn't up-to-date.
3047 * Try to re-read it _once_. We do this synchronously,
3048 * because there really aren't any performance issues here
3049 * and we need to check for errors.
3050 */
3051 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3052 error = filemap_read_page(file, mapping, page);
3053 if (fpin)
3054 goto out_retry;
3055 put_page(page);
3056
3057 if (!error || error == AOP_TRUNCATED_PAGE)
3058 goto retry_find;
3059
3060 return VM_FAULT_SIGBUS;
3061
3062out_retry:
3063 /*
3064 * We dropped the mmap_lock, we need to return to the fault handler to
3065 * re-find the vma and come back and find our hopefully still populated
3066 * page.
3067 */
3068 if (page)
3069 put_page(page);
3070 if (fpin)
3071 fput(fpin);
3072 return ret | VM_FAULT_RETRY;
3073}
3074EXPORT_SYMBOL(filemap_fault);
3075
3076static bool filemap_map_pmd(struct vm_fault *vmf, struct page *page)
3077{
3078 struct mm_struct *mm = vmf->vma->vm_mm;
3079
3080 /* Huge page is mapped? No need to proceed. */
3081 if (pmd_trans_huge(*vmf->pmd)) {
3082 unlock_page(page);
3083 put_page(page);
3084 return true;
3085 }
3086
3087 if (pmd_none(*vmf->pmd) && PageTransHuge(page)) {
3088 vm_fault_t ret = do_set_pmd(vmf, page);
3089 if (!ret) {
3090 /* The page is mapped successfully, reference consumed. */
3091 unlock_page(page);
3092 return true;
3093 }
3094 }
3095
3096 if (pmd_none(*vmf->pmd)) {
3097 vmf->ptl = pmd_lock(mm, vmf->pmd);
3098 if (likely(pmd_none(*vmf->pmd))) {
3099 mm_inc_nr_ptes(mm);
3100 pmd_populate(mm, vmf->pmd, vmf->prealloc_pte);
3101 vmf->prealloc_pte = NULL;
3102 }
3103 spin_unlock(vmf->ptl);
3104 }
3105
3106 /* See comment in handle_pte_fault() */
3107 if (pmd_devmap_trans_unstable(vmf->pmd)) {
3108 unlock_page(page);
3109 put_page(page);
3110 return true;
3111 }
3112
3113 return false;
3114}
3115
3116static struct page *next_uptodate_page(struct page *page,
3117 struct address_space *mapping,
3118 struct xa_state *xas, pgoff_t end_pgoff)
3119{
3120 unsigned long max_idx;
3121
3122 do {
3123 if (!page)
3124 return NULL;
3125 if (xas_retry(xas, page))
3126 continue;
3127 if (xa_is_value(page))
3128 continue;
3129 if (PageLocked(page))
3130 continue;
3131 if (!page_cache_get_speculative(page))
3132 continue;
3133 /* Has the page moved or been split? */
3134 if (unlikely(page != xas_reload(xas)))
3135 goto skip;
3136 if (!PageUptodate(page) || PageReadahead(page))
3137 goto skip;
3138 if (PageHWPoison(page))
3139 goto skip;
3140 if (!trylock_page(page))
3141 goto skip;
3142 if (page->mapping != mapping)
3143 goto unlock;
3144 if (!PageUptodate(page))
3145 goto unlock;
3146 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3147 if (xas->xa_index >= max_idx)
3148 goto unlock;
3149 return page;
3150unlock:
3151 unlock_page(page);
3152skip:
3153 put_page(page);
3154 } while ((page = xas_next_entry(xas, end_pgoff)) != NULL);
3155
3156 return NULL;
3157}
3158
3159static inline struct page *first_map_page(struct address_space *mapping,
3160 struct xa_state *xas,
3161 pgoff_t end_pgoff)
3162{
3163 return next_uptodate_page(xas_find(xas, end_pgoff),
3164 mapping, xas, end_pgoff);
3165}
3166
3167static inline struct page *next_map_page(struct address_space *mapping,
3168 struct xa_state *xas,
3169 pgoff_t end_pgoff)
3170{
3171 return next_uptodate_page(xas_next_entry(xas, end_pgoff),
3172 mapping, xas, end_pgoff);
3173}
3174
3175vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3176 pgoff_t start_pgoff, pgoff_t end_pgoff)
3177{
3178 struct vm_area_struct *vma = vmf->vma;
3179 struct file *file = vma->vm_file;
3180 struct address_space *mapping = file->f_mapping;
3181 pgoff_t last_pgoff = start_pgoff;
3182 unsigned long addr;
3183 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3184 struct page *head, *page;
3185 unsigned int mmap_miss = READ_ONCE(file->f_ra.mmap_miss);
3186 vm_fault_t ret = 0;
3187
3188 rcu_read_lock();
3189 head = first_map_page(mapping, &xas, end_pgoff);
3190 if (!head)
3191 goto out;
3192
3193 if (filemap_map_pmd(vmf, head)) {
3194 ret = VM_FAULT_NOPAGE;
3195 goto out;
3196 }
3197
3198 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3199 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3200 do {
3201 page = find_subpage(head, xas.xa_index);
3202 if (PageHWPoison(page))
3203 goto unlock;
3204
3205 if (mmap_miss > 0)
3206 mmap_miss--;
3207
3208 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3209 vmf->pte += xas.xa_index - last_pgoff;
3210 last_pgoff = xas.xa_index;
3211
3212 if (!pte_none(*vmf->pte))
3213 goto unlock;
3214
3215 /* We're about to handle the fault */
3216 if (vmf->address == addr)
3217 ret = VM_FAULT_NOPAGE;
3218
3219 do_set_pte(vmf, page, addr);
3220 /* no need to invalidate: a not-present page won't be cached */
3221 update_mmu_cache(vma, addr, vmf->pte);
3222 unlock_page(head);
3223 continue;
3224unlock:
3225 unlock_page(head);
3226 put_page(head);
3227 } while ((head = next_map_page(mapping, &xas, end_pgoff)) != NULL);
3228 pte_unmap_unlock(vmf->pte, vmf->ptl);
3229out:
3230 rcu_read_unlock();
3231 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss);
3232 return ret;
3233}
3234EXPORT_SYMBOL(filemap_map_pages);
3235
3236vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3237{
3238 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3239 struct page *page = vmf->page;
3240 vm_fault_t ret = VM_FAULT_LOCKED;
3241
3242 sb_start_pagefault(mapping->host->i_sb);
3243 file_update_time(vmf->vma->vm_file);
3244 lock_page(page);
3245 if (page->mapping != mapping) {
3246 unlock_page(page);
3247 ret = VM_FAULT_NOPAGE;
3248 goto out;
3249 }
3250 /*
3251 * We mark the page dirty already here so that when freeze is in
3252 * progress, we are guaranteed that writeback during freezing will
3253 * see the dirty page and writeprotect it again.
3254 */
3255 set_page_dirty(page);
3256 wait_for_stable_page(page);
3257out:
3258 sb_end_pagefault(mapping->host->i_sb);
3259 return ret;
3260}
3261
3262const struct vm_operations_struct generic_file_vm_ops = {
3263 .fault = filemap_fault,
3264 .map_pages = filemap_map_pages,
3265 .page_mkwrite = filemap_page_mkwrite,
3266};
3267
3268/* This is used for a general mmap of a disk file */
3269
3270int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3271{
3272 struct address_space *mapping = file->f_mapping;
3273
3274 if (!mapping->a_ops->readpage)
3275 return -ENOEXEC;
3276 file_accessed(file);
3277 vma->vm_ops = &generic_file_vm_ops;
3278 return 0;
3279}
3280
3281/*
3282 * This is for filesystems which do not implement ->writepage.
3283 */
3284int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3285{
3286 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
3287 return -EINVAL;
3288 return generic_file_mmap(file, vma);
3289}
3290#else
3291vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3292{
3293 return VM_FAULT_SIGBUS;
3294}
3295int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3296{
3297 return -ENOSYS;
3298}
3299int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3300{
3301 return -ENOSYS;
3302}
3303#endif /* CONFIG_MMU */
3304
3305EXPORT_SYMBOL(filemap_page_mkwrite);
3306EXPORT_SYMBOL(generic_file_mmap);
3307EXPORT_SYMBOL(generic_file_readonly_mmap);
3308
3309static struct page *wait_on_page_read(struct page *page)
3310{
3311 if (!IS_ERR(page)) {
3312 wait_on_page_locked(page);
3313 if (!PageUptodate(page)) {
3314 put_page(page);
3315 page = ERR_PTR(-EIO);
3316 }
3317 }
3318 return page;
3319}
3320
3321static struct page *do_read_cache_page(struct address_space *mapping,
3322 pgoff_t index,
3323 int (*filler)(void *, struct page *),
3324 void *data,
3325 gfp_t gfp)
3326{
3327 struct page *page;
3328 int err;
3329repeat:
3330 page = find_get_page(mapping, index);
3331 if (!page) {
3332 page = __page_cache_alloc(gfp);
3333 if (!page)
3334 return ERR_PTR(-ENOMEM);
3335 err = add_to_page_cache_lru(page, mapping, index, gfp);
3336 if (unlikely(err)) {
3337 put_page(page);
3338 if (err == -EEXIST)
3339 goto repeat;
3340 /* Presumably ENOMEM for xarray node */
3341 return ERR_PTR(err);
3342 }
3343
3344filler:
3345 if (filler)
3346 err = filler(data, page);
3347 else
3348 err = mapping->a_ops->readpage(data, page);
3349
3350 if (err < 0) {
3351 put_page(page);
3352 return ERR_PTR(err);
3353 }
3354
3355 page = wait_on_page_read(page);
3356 if (IS_ERR(page))
3357 return page;
3358 goto out;
3359 }
3360 if (PageUptodate(page))
3361 goto out;
3362
3363 /*
3364 * Page is not up to date and may be locked due to one of the following
3365 * case a: Page is being filled and the page lock is held
3366 * case b: Read/write error clearing the page uptodate status
3367 * case c: Truncation in progress (page locked)
3368 * case d: Reclaim in progress
3369 *
3370 * Case a, the page will be up to date when the page is unlocked.
3371 * There is no need to serialise on the page lock here as the page
3372 * is pinned so the lock gives no additional protection. Even if the
3373 * page is truncated, the data is still valid if PageUptodate as
3374 * it's a race vs truncate race.
3375 * Case b, the page will not be up to date
3376 * Case c, the page may be truncated but in itself, the data may still
3377 * be valid after IO completes as it's a read vs truncate race. The
3378 * operation must restart if the page is not uptodate on unlock but
3379 * otherwise serialising on page lock to stabilise the mapping gives
3380 * no additional guarantees to the caller as the page lock is
3381 * released before return.
3382 * Case d, similar to truncation. If reclaim holds the page lock, it
3383 * will be a race with remove_mapping that determines if the mapping
3384 * is valid on unlock but otherwise the data is valid and there is
3385 * no need to serialise with page lock.
3386 *
3387 * As the page lock gives no additional guarantee, we optimistically
3388 * wait on the page to be unlocked and check if it's up to date and
3389 * use the page if it is. Otherwise, the page lock is required to
3390 * distinguish between the different cases. The motivation is that we
3391 * avoid spurious serialisations and wakeups when multiple processes
3392 * wait on the same page for IO to complete.
3393 */
3394 wait_on_page_locked(page);
3395 if (PageUptodate(page))
3396 goto out;
3397
3398 /* Distinguish between all the cases under the safety of the lock */
3399 lock_page(page);
3400
3401 /* Case c or d, restart the operation */
3402 if (!page->mapping) {
3403 unlock_page(page);
3404 put_page(page);
3405 goto repeat;
3406 }
3407
3408 /* Someone else locked and filled the page in a very small window */
3409 if (PageUptodate(page)) {
3410 unlock_page(page);
3411 goto out;
3412 }
3413
3414 /*
3415 * A previous I/O error may have been due to temporary
3416 * failures.
3417 * Clear page error before actual read, PG_error will be
3418 * set again if read page fails.
3419 */
3420 ClearPageError(page);
3421 goto filler;
3422
3423out:
3424 mark_page_accessed(page);
3425 return page;
3426}
3427
3428/**
3429 * read_cache_page - read into page cache, fill it if needed
3430 * @mapping: the page's address_space
3431 * @index: the page index
3432 * @filler: function to perform the read
3433 * @data: first arg to filler(data, page) function, often left as NULL
3434 *
3435 * Read into the page cache. If a page already exists, and PageUptodate() is
3436 * not set, try to fill the page and wait for it to become unlocked.
3437 *
3438 * If the page does not get brought uptodate, return -EIO.
3439 *
3440 * Return: up to date page on success, ERR_PTR() on failure.
3441 */
3442struct page *read_cache_page(struct address_space *mapping,
3443 pgoff_t index,
3444 int (*filler)(void *, struct page *),
3445 void *data)
3446{
3447 return do_read_cache_page(mapping, index, filler, data,
3448 mapping_gfp_mask(mapping));
3449}
3450EXPORT_SYMBOL(read_cache_page);
3451
3452/**
3453 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3454 * @mapping: the page's address_space
3455 * @index: the page index
3456 * @gfp: the page allocator flags to use if allocating
3457 *
3458 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3459 * any new page allocations done using the specified allocation flags.
3460 *
3461 * If the page does not get brought uptodate, return -EIO.
3462 *
3463 * Return: up to date page on success, ERR_PTR() on failure.
3464 */
3465struct page *read_cache_page_gfp(struct address_space *mapping,
3466 pgoff_t index,
3467 gfp_t gfp)
3468{
3469 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3470}
3471EXPORT_SYMBOL(read_cache_page_gfp);
3472
3473int pagecache_write_begin(struct file *file, struct address_space *mapping,
3474 loff_t pos, unsigned len, unsigned flags,
3475 struct page **pagep, void **fsdata)
3476{
3477 const struct address_space_operations *aops = mapping->a_ops;
3478
3479 return aops->write_begin(file, mapping, pos, len, flags,
3480 pagep, fsdata);
3481}
3482EXPORT_SYMBOL(pagecache_write_begin);
3483
3484int pagecache_write_end(struct file *file, struct address_space *mapping,
3485 loff_t pos, unsigned len, unsigned copied,
3486 struct page *page, void *fsdata)
3487{
3488 const struct address_space_operations *aops = mapping->a_ops;
3489
3490 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
3491}
3492EXPORT_SYMBOL(pagecache_write_end);
3493
3494/*
3495 * Warn about a page cache invalidation failure during a direct I/O write.
3496 */
3497void dio_warn_stale_pagecache(struct file *filp)
3498{
3499 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3500 char pathname[128];
3501 char *path;
3502
3503 errseq_set(&filp->f_mapping->wb_err, -EIO);
3504 if (__ratelimit(&_rs)) {
3505 path = file_path(filp, pathname, sizeof(pathname));
3506 if (IS_ERR(path))
3507 path = "(unknown)";
3508 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3509 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3510 current->comm);
3511 }
3512}
3513
3514ssize_t
3515generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3516{
3517 struct file *file = iocb->ki_filp;
3518 struct address_space *mapping = file->f_mapping;
3519 struct inode *inode = mapping->host;
3520 loff_t pos = iocb->ki_pos;
3521 ssize_t written;
3522 size_t write_len;
3523 pgoff_t end;
3524
3525 write_len = iov_iter_count(from);
3526 end = (pos + write_len - 1) >> PAGE_SHIFT;
3527
3528 if (iocb->ki_flags & IOCB_NOWAIT) {
3529 /* If there are pages to writeback, return */
3530 if (filemap_range_has_page(file->f_mapping, pos,
3531 pos + write_len - 1))
3532 return -EAGAIN;
3533 } else {
3534 written = filemap_write_and_wait_range(mapping, pos,
3535 pos + write_len - 1);
3536 if (written)
3537 goto out;
3538 }
3539
3540 /*
3541 * After a write we want buffered reads to be sure to go to disk to get
3542 * the new data. We invalidate clean cached page from the region we're
3543 * about to write. We do this *before* the write so that we can return
3544 * without clobbering -EIOCBQUEUED from ->direct_IO().
3545 */
3546 written = invalidate_inode_pages2_range(mapping,
3547 pos >> PAGE_SHIFT, end);
3548 /*
3549 * If a page can not be invalidated, return 0 to fall back
3550 * to buffered write.
3551 */
3552 if (written) {
3553 if (written == -EBUSY)
3554 return 0;
3555 goto out;
3556 }
3557
3558 written = mapping->a_ops->direct_IO(iocb, from);
3559
3560 /*
3561 * Finally, try again to invalidate clean pages which might have been
3562 * cached by non-direct readahead, or faulted in by get_user_pages()
3563 * if the source of the write was an mmap'ed region of the file
3564 * we're writing. Either one is a pretty crazy thing to do,
3565 * so we don't support it 100%. If this invalidation
3566 * fails, tough, the write still worked...
3567 *
3568 * Most of the time we do not need this since dio_complete() will do
3569 * the invalidation for us. However there are some file systems that
3570 * do not end up with dio_complete() being called, so let's not break
3571 * them by removing it completely.
3572 *
3573 * Noticeable example is a blkdev_direct_IO().
3574 *
3575 * Skip invalidation for async writes or if mapping has no pages.
3576 */
3577 if (written > 0 && mapping->nrpages &&
3578 invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT, end))
3579 dio_warn_stale_pagecache(file);
3580
3581 if (written > 0) {
3582 pos += written;
3583 write_len -= written;
3584 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3585 i_size_write(inode, pos);
3586 mark_inode_dirty(inode);
3587 }
3588 iocb->ki_pos = pos;
3589 }
3590 if (written != -EIOCBQUEUED)
3591 iov_iter_revert(from, write_len - iov_iter_count(from));
3592out:
3593 return written;
3594}
3595EXPORT_SYMBOL(generic_file_direct_write);
3596
3597/*
3598 * Find or create a page at the given pagecache position. Return the locked
3599 * page. This function is specifically for buffered writes.
3600 */
3601struct page *grab_cache_page_write_begin(struct address_space *mapping,
3602 pgoff_t index, unsigned flags)
3603{
3604 struct page *page;
3605 int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3606
3607 if (flags & AOP_FLAG_NOFS)
3608 fgp_flags |= FGP_NOFS;
3609
3610 page = pagecache_get_page(mapping, index, fgp_flags,
3611 mapping_gfp_mask(mapping));
3612 if (page)
3613 wait_for_stable_page(page);
3614
3615 return page;
3616}
3617EXPORT_SYMBOL(grab_cache_page_write_begin);
3618
3619ssize_t generic_perform_write(struct file *file,
3620 struct iov_iter *i, loff_t pos)
3621{
3622 struct address_space *mapping = file->f_mapping;
3623 const struct address_space_operations *a_ops = mapping->a_ops;
3624 long status = 0;
3625 ssize_t written = 0;
3626 unsigned int flags = 0;
3627
3628 do {
3629 struct page *page;
3630 unsigned long offset; /* Offset into pagecache page */
3631 unsigned long bytes; /* Bytes to write to page */
3632 size_t copied; /* Bytes copied from user */
3633 void *fsdata;
3634
3635 offset = (pos & (PAGE_SIZE - 1));
3636 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3637 iov_iter_count(i));
3638
3639again:
3640 /*
3641 * Bring in the user page that we will copy from _first_.
3642 * Otherwise there's a nasty deadlock on copying from the
3643 * same page as we're writing to, without it being marked
3644 * up-to-date.
3645 */
3646 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3647 status = -EFAULT;
3648 break;
3649 }
3650
3651 if (fatal_signal_pending(current)) {
3652 status = -EINTR;
3653 break;
3654 }
3655
3656 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3657 &page, &fsdata);
3658 if (unlikely(status < 0))
3659 break;
3660
3661 if (mapping_writably_mapped(mapping))
3662 flush_dcache_page(page);
3663
3664 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
3665 flush_dcache_page(page);
3666
3667 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3668 page, fsdata);
3669 if (unlikely(status != copied)) {
3670 iov_iter_revert(i, copied - max(status, 0L));
3671 if (unlikely(status < 0))
3672 break;
3673 }
3674 cond_resched();
3675
3676 if (unlikely(status == 0)) {
3677 /*
3678 * A short copy made ->write_end() reject the
3679 * thing entirely. Might be memory poisoning
3680 * halfway through, might be a race with munmap,
3681 * might be severe memory pressure.
3682 */
3683 if (copied)
3684 bytes = copied;
3685 goto again;
3686 }
3687 pos += status;
3688 written += status;
3689
3690 balance_dirty_pages_ratelimited(mapping);
3691 } while (iov_iter_count(i));
3692
3693 return written ? written : status;
3694}
3695EXPORT_SYMBOL(generic_perform_write);
3696
3697/**
3698 * __generic_file_write_iter - write data to a file
3699 * @iocb: IO state structure (file, offset, etc.)
3700 * @from: iov_iter with data to write
3701 *
3702 * This function does all the work needed for actually writing data to a
3703 * file. It does all basic checks, removes SUID from the file, updates
3704 * modification times and calls proper subroutines depending on whether we
3705 * do direct IO or a standard buffered write.
3706 *
3707 * It expects i_mutex to be grabbed unless we work on a block device or similar
3708 * object which does not need locking at all.
3709 *
3710 * This function does *not* take care of syncing data in case of O_SYNC write.
3711 * A caller has to handle it. This is mainly due to the fact that we want to
3712 * avoid syncing under i_mutex.
3713 *
3714 * Return:
3715 * * number of bytes written, even for truncated writes
3716 * * negative error code if no data has been written at all
3717 */
3718ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3719{
3720 struct file *file = iocb->ki_filp;
3721 struct address_space *mapping = file->f_mapping;
3722 struct inode *inode = mapping->host;
3723 ssize_t written = 0;
3724 ssize_t err;
3725 ssize_t status;
3726
3727 /* We can write back this queue in page reclaim */
3728 current->backing_dev_info = inode_to_bdi(inode);
3729 err = file_remove_privs(file);
3730 if (err)
3731 goto out;
3732
3733 err = file_update_time(file);
3734 if (err)
3735 goto out;
3736
3737 if (iocb->ki_flags & IOCB_DIRECT) {
3738 loff_t pos, endbyte;
3739
3740 written = generic_file_direct_write(iocb, from);
3741 /*
3742 * If the write stopped short of completing, fall back to
3743 * buffered writes. Some filesystems do this for writes to
3744 * holes, for example. For DAX files, a buffered write will
3745 * not succeed (even if it did, DAX does not handle dirty
3746 * page-cache pages correctly).
3747 */
3748 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3749 goto out;
3750
3751 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3752 /*
3753 * If generic_perform_write() returned a synchronous error
3754 * then we want to return the number of bytes which were
3755 * direct-written, or the error code if that was zero. Note
3756 * that this differs from normal direct-io semantics, which
3757 * will return -EFOO even if some bytes were written.
3758 */
3759 if (unlikely(status < 0)) {
3760 err = status;
3761 goto out;
3762 }
3763 /*
3764 * We need to ensure that the page cache pages are written to
3765 * disk and invalidated to preserve the expected O_DIRECT
3766 * semantics.
3767 */
3768 endbyte = pos + status - 1;
3769 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3770 if (err == 0) {
3771 iocb->ki_pos = endbyte + 1;
3772 written += status;
3773 invalidate_mapping_pages(mapping,
3774 pos >> PAGE_SHIFT,
3775 endbyte >> PAGE_SHIFT);
3776 } else {
3777 /*
3778 * We don't know how much we wrote, so just return
3779 * the number of bytes which were direct-written
3780 */
3781 }
3782 } else {
3783 written = generic_perform_write(file, from, iocb->ki_pos);
3784 if (likely(written > 0))
3785 iocb->ki_pos += written;
3786 }
3787out:
3788 current->backing_dev_info = NULL;
3789 return written ? written : err;
3790}
3791EXPORT_SYMBOL(__generic_file_write_iter);
3792
3793/**
3794 * generic_file_write_iter - write data to a file
3795 * @iocb: IO state structure
3796 * @from: iov_iter with data to write
3797 *
3798 * This is a wrapper around __generic_file_write_iter() to be used by most
3799 * filesystems. It takes care of syncing the file in case of O_SYNC file
3800 * and acquires i_mutex as needed.
3801 * Return:
3802 * * negative error code if no data has been written at all of
3803 * vfs_fsync_range() failed for a synchronous write
3804 * * number of bytes written, even for truncated writes
3805 */
3806ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3807{
3808 struct file *file = iocb->ki_filp;
3809 struct inode *inode = file->f_mapping->host;
3810 ssize_t ret;
3811
3812 inode_lock(inode);
3813 ret = generic_write_checks(iocb, from);
3814 if (ret > 0)
3815 ret = __generic_file_write_iter(iocb, from);
3816 inode_unlock(inode);
3817
3818 if (ret > 0)
3819 ret = generic_write_sync(iocb, ret);
3820 return ret;
3821}
3822EXPORT_SYMBOL(generic_file_write_iter);
3823
3824/**
3825 * try_to_release_page() - release old fs-specific metadata on a page
3826 *
3827 * @page: the page which the kernel is trying to free
3828 * @gfp_mask: memory allocation flags (and I/O mode)
3829 *
3830 * The address_space is to try to release any data against the page
3831 * (presumably at page->private).
3832 *
3833 * This may also be called if PG_fscache is set on a page, indicating that the
3834 * page is known to the local caching routines.
3835 *
3836 * The @gfp_mask argument specifies whether I/O may be performed to release
3837 * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3838 *
3839 * Return: %1 if the release was successful, otherwise return zero.
3840 */
3841int try_to_release_page(struct page *page, gfp_t gfp_mask)
3842{
3843 struct address_space * const mapping = page->mapping;
3844
3845 BUG_ON(!PageLocked(page));
3846 if (PageWriteback(page))
3847 return 0;
3848
3849 if (mapping && mapping->a_ops->releasepage)
3850 return mapping->a_ops->releasepage(page, gfp_mask);
3851 return try_to_free_buffers(page);
3852}
3853
3854EXPORT_SYMBOL(try_to_release_page);