Loading...
1/*
2 * linux/mm/filemap.c
3 *
4 * Copyright (C) 1994-1999 Linus Torvalds
5 */
6
7/*
8 * This file handles the generic file mmap semantics used by
9 * most "normal" filesystems (but you don't /have/ to use this:
10 * the NFS filesystem used to do this differently, for example)
11 */
12#include <linux/export.h>
13#include <linux/compiler.h>
14#include <linux/fs.h>
15#include <linux/uaccess.h>
16#include <linux/aio.h>
17#include <linux/capability.h>
18#include <linux/kernel_stat.h>
19#include <linux/gfp.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/mman.h>
23#include <linux/pagemap.h>
24#include <linux/file.h>
25#include <linux/uio.h>
26#include <linux/hash.h>
27#include <linux/writeback.h>
28#include <linux/backing-dev.h>
29#include <linux/pagevec.h>
30#include <linux/blkdev.h>
31#include <linux/security.h>
32#include <linux/cpuset.h>
33#include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34#include <linux/memcontrol.h>
35#include <linux/cleancache.h>
36#include <linux/rmap.h>
37#include "internal.h"
38
39#define CREATE_TRACE_POINTS
40#include <trace/events/filemap.h>
41
42/*
43 * FIXME: remove all knowledge of the buffer layer from the core VM
44 */
45#include <linux/buffer_head.h> /* for try_to_free_buffers */
46
47#include <asm/mman.h>
48
49/*
50 * Shared mappings implemented 30.11.1994. It's not fully working yet,
51 * though.
52 *
53 * Shared mappings now work. 15.8.1995 Bruno.
54 *
55 * finished 'unifying' the page and buffer cache and SMP-threaded the
56 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57 *
58 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59 */
60
61/*
62 * Lock ordering:
63 *
64 * ->i_mmap_mutex (truncate_pagecache)
65 * ->private_lock (__free_pte->__set_page_dirty_buffers)
66 * ->swap_lock (exclusive_swap_page, others)
67 * ->mapping->tree_lock
68 *
69 * ->i_mutex
70 * ->i_mmap_mutex (truncate->unmap_mapping_range)
71 *
72 * ->mmap_sem
73 * ->i_mmap_mutex
74 * ->page_table_lock or pte_lock (various, mainly in memory.c)
75 * ->mapping->tree_lock (arch-dependent flush_dcache_mmap_lock)
76 *
77 * ->mmap_sem
78 * ->lock_page (access_process_vm)
79 *
80 * ->i_mutex (generic_perform_write)
81 * ->mmap_sem (fault_in_pages_readable->do_page_fault)
82 *
83 * bdi->wb.list_lock
84 * sb_lock (fs/fs-writeback.c)
85 * ->mapping->tree_lock (__sync_single_inode)
86 *
87 * ->i_mmap_mutex
88 * ->anon_vma.lock (vma_adjust)
89 *
90 * ->anon_vma.lock
91 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
92 *
93 * ->page_table_lock or pte_lock
94 * ->swap_lock (try_to_unmap_one)
95 * ->private_lock (try_to_unmap_one)
96 * ->tree_lock (try_to_unmap_one)
97 * ->zone.lru_lock (follow_page->mark_page_accessed)
98 * ->zone.lru_lock (check_pte_range->isolate_lru_page)
99 * ->private_lock (page_remove_rmap->set_page_dirty)
100 * ->tree_lock (page_remove_rmap->set_page_dirty)
101 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
102 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
103 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
104 * ->inode->i_lock (zap_pte_range->set_page_dirty)
105 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
106 *
107 * ->i_mmap_mutex
108 * ->tasklist_lock (memory_failure, collect_procs_ao)
109 */
110
111static void page_cache_tree_delete(struct address_space *mapping,
112 struct page *page, void *shadow)
113{
114 struct radix_tree_node *node;
115 unsigned long index;
116 unsigned int offset;
117 unsigned int tag;
118 void **slot;
119
120 VM_BUG_ON(!PageLocked(page));
121
122 __radix_tree_lookup(&mapping->page_tree, page->index, &node, &slot);
123
124 if (shadow) {
125 mapping->nrshadows++;
126 /*
127 * Make sure the nrshadows update is committed before
128 * the nrpages update so that final truncate racing
129 * with reclaim does not see both counters 0 at the
130 * same time and miss a shadow entry.
131 */
132 smp_wmb();
133 }
134 mapping->nrpages--;
135
136 if (!node) {
137 /* Clear direct pointer tags in root node */
138 mapping->page_tree.gfp_mask &= __GFP_BITS_MASK;
139 radix_tree_replace_slot(slot, shadow);
140 return;
141 }
142
143 /* Clear tree tags for the removed page */
144 index = page->index;
145 offset = index & RADIX_TREE_MAP_MASK;
146 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
147 if (test_bit(offset, node->tags[tag]))
148 radix_tree_tag_clear(&mapping->page_tree, index, tag);
149 }
150
151 /* Delete page, swap shadow entry */
152 radix_tree_replace_slot(slot, shadow);
153 workingset_node_pages_dec(node);
154 if (shadow)
155 workingset_node_shadows_inc(node);
156 else
157 if (__radix_tree_delete_node(&mapping->page_tree, node))
158 return;
159
160 /*
161 * Track node that only contains shadow entries.
162 *
163 * Avoid acquiring the list_lru lock if already tracked. The
164 * list_empty() test is safe as node->private_list is
165 * protected by mapping->tree_lock.
166 */
167 if (!workingset_node_pages(node) &&
168 list_empty(&node->private_list)) {
169 node->private_data = mapping;
170 list_lru_add(&workingset_shadow_nodes, &node->private_list);
171 }
172}
173
174/*
175 * Delete a page from the page cache and free it. Caller has to make
176 * sure the page is locked and that nobody else uses it - or that usage
177 * is safe. The caller must hold the mapping's tree_lock.
178 */
179void __delete_from_page_cache(struct page *page, void *shadow)
180{
181 struct address_space *mapping = page->mapping;
182
183 trace_mm_filemap_delete_from_page_cache(page);
184 /*
185 * if we're uptodate, flush out into the cleancache, otherwise
186 * invalidate any existing cleancache entries. We can't leave
187 * stale data around in the cleancache once our page is gone
188 */
189 if (PageUptodate(page) && PageMappedToDisk(page))
190 cleancache_put_page(page);
191 else
192 cleancache_invalidate_page(mapping, page);
193
194 page_cache_tree_delete(mapping, page, shadow);
195
196 page->mapping = NULL;
197 /* Leave page->index set: truncation lookup relies upon it */
198
199 __dec_zone_page_state(page, NR_FILE_PAGES);
200 if (PageSwapBacked(page))
201 __dec_zone_page_state(page, NR_SHMEM);
202 BUG_ON(page_mapped(page));
203
204 /*
205 * Some filesystems seem to re-dirty the page even after
206 * the VM has canceled the dirty bit (eg ext3 journaling).
207 *
208 * Fix it up by doing a final dirty accounting check after
209 * having removed the page entirely.
210 */
211 if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
212 dec_zone_page_state(page, NR_FILE_DIRTY);
213 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
214 }
215}
216
217/**
218 * delete_from_page_cache - delete page from page cache
219 * @page: the page which the kernel is trying to remove from page cache
220 *
221 * This must be called only on pages that have been verified to be in the page
222 * cache and locked. It will never put the page into the free list, the caller
223 * has a reference on the page.
224 */
225void delete_from_page_cache(struct page *page)
226{
227 struct address_space *mapping = page->mapping;
228 void (*freepage)(struct page *);
229
230 BUG_ON(!PageLocked(page));
231
232 freepage = mapping->a_ops->freepage;
233 spin_lock_irq(&mapping->tree_lock);
234 __delete_from_page_cache(page, NULL);
235 spin_unlock_irq(&mapping->tree_lock);
236 mem_cgroup_uncharge_cache_page(page);
237
238 if (freepage)
239 freepage(page);
240 page_cache_release(page);
241}
242EXPORT_SYMBOL(delete_from_page_cache);
243
244static int sleep_on_page(void *word)
245{
246 io_schedule();
247 return 0;
248}
249
250static int sleep_on_page_killable(void *word)
251{
252 sleep_on_page(word);
253 return fatal_signal_pending(current) ? -EINTR : 0;
254}
255
256static int filemap_check_errors(struct address_space *mapping)
257{
258 int ret = 0;
259 /* Check for outstanding write errors */
260 if (test_bit(AS_ENOSPC, &mapping->flags) &&
261 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
262 ret = -ENOSPC;
263 if (test_bit(AS_EIO, &mapping->flags) &&
264 test_and_clear_bit(AS_EIO, &mapping->flags))
265 ret = -EIO;
266 return ret;
267}
268
269/**
270 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
271 * @mapping: address space structure to write
272 * @start: offset in bytes where the range starts
273 * @end: offset in bytes where the range ends (inclusive)
274 * @sync_mode: enable synchronous operation
275 *
276 * Start writeback against all of a mapping's dirty pages that lie
277 * within the byte offsets <start, end> inclusive.
278 *
279 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
280 * opposed to a regular memory cleansing writeback. The difference between
281 * these two operations is that if a dirty page/buffer is encountered, it must
282 * be waited upon, and not just skipped over.
283 */
284int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
285 loff_t end, int sync_mode)
286{
287 int ret;
288 struct writeback_control wbc = {
289 .sync_mode = sync_mode,
290 .nr_to_write = LONG_MAX,
291 .range_start = start,
292 .range_end = end,
293 };
294
295 if (!mapping_cap_writeback_dirty(mapping))
296 return 0;
297
298 ret = do_writepages(mapping, &wbc);
299 return ret;
300}
301
302static inline int __filemap_fdatawrite(struct address_space *mapping,
303 int sync_mode)
304{
305 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
306}
307
308int filemap_fdatawrite(struct address_space *mapping)
309{
310 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
311}
312EXPORT_SYMBOL(filemap_fdatawrite);
313
314int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
315 loff_t end)
316{
317 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
318}
319EXPORT_SYMBOL(filemap_fdatawrite_range);
320
321/**
322 * filemap_flush - mostly a non-blocking flush
323 * @mapping: target address_space
324 *
325 * This is a mostly non-blocking flush. Not suitable for data-integrity
326 * purposes - I/O may not be started against all dirty pages.
327 */
328int filemap_flush(struct address_space *mapping)
329{
330 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
331}
332EXPORT_SYMBOL(filemap_flush);
333
334/**
335 * filemap_fdatawait_range - wait for writeback to complete
336 * @mapping: address space structure to wait for
337 * @start_byte: offset in bytes where the range starts
338 * @end_byte: offset in bytes where the range ends (inclusive)
339 *
340 * Walk the list of under-writeback pages of the given address space
341 * in the given range and wait for all of them.
342 */
343int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
344 loff_t end_byte)
345{
346 pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
347 pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
348 struct pagevec pvec;
349 int nr_pages;
350 int ret2, ret = 0;
351
352 if (end_byte < start_byte)
353 goto out;
354
355 pagevec_init(&pvec, 0);
356 while ((index <= end) &&
357 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
358 PAGECACHE_TAG_WRITEBACK,
359 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
360 unsigned i;
361
362 for (i = 0; i < nr_pages; i++) {
363 struct page *page = pvec.pages[i];
364
365 /* until radix tree lookup accepts end_index */
366 if (page->index > end)
367 continue;
368
369 wait_on_page_writeback(page);
370 if (TestClearPageError(page))
371 ret = -EIO;
372 }
373 pagevec_release(&pvec);
374 cond_resched();
375 }
376out:
377 ret2 = filemap_check_errors(mapping);
378 if (!ret)
379 ret = ret2;
380
381 return ret;
382}
383EXPORT_SYMBOL(filemap_fdatawait_range);
384
385/**
386 * filemap_fdatawait - wait for all under-writeback pages to complete
387 * @mapping: address space structure to wait for
388 *
389 * Walk the list of under-writeback pages of the given address space
390 * and wait for all of them.
391 */
392int filemap_fdatawait(struct address_space *mapping)
393{
394 loff_t i_size = i_size_read(mapping->host);
395
396 if (i_size == 0)
397 return 0;
398
399 return filemap_fdatawait_range(mapping, 0, i_size - 1);
400}
401EXPORT_SYMBOL(filemap_fdatawait);
402
403int filemap_write_and_wait(struct address_space *mapping)
404{
405 int err = 0;
406
407 if (mapping->nrpages) {
408 err = filemap_fdatawrite(mapping);
409 /*
410 * Even if the above returned error, the pages may be
411 * written partially (e.g. -ENOSPC), so we wait for it.
412 * But the -EIO is special case, it may indicate the worst
413 * thing (e.g. bug) happened, so we avoid waiting for it.
414 */
415 if (err != -EIO) {
416 int err2 = filemap_fdatawait(mapping);
417 if (!err)
418 err = err2;
419 }
420 } else {
421 err = filemap_check_errors(mapping);
422 }
423 return err;
424}
425EXPORT_SYMBOL(filemap_write_and_wait);
426
427/**
428 * filemap_write_and_wait_range - write out & wait on a file range
429 * @mapping: the address_space for the pages
430 * @lstart: offset in bytes where the range starts
431 * @lend: offset in bytes where the range ends (inclusive)
432 *
433 * Write out and wait upon file offsets lstart->lend, inclusive.
434 *
435 * Note that `lend' is inclusive (describes the last byte to be written) so
436 * that this function can be used to write to the very end-of-file (end = -1).
437 */
438int filemap_write_and_wait_range(struct address_space *mapping,
439 loff_t lstart, loff_t lend)
440{
441 int err = 0;
442
443 if (mapping->nrpages) {
444 err = __filemap_fdatawrite_range(mapping, lstart, lend,
445 WB_SYNC_ALL);
446 /* See comment of filemap_write_and_wait() */
447 if (err != -EIO) {
448 int err2 = filemap_fdatawait_range(mapping,
449 lstart, lend);
450 if (!err)
451 err = err2;
452 }
453 } else {
454 err = filemap_check_errors(mapping);
455 }
456 return err;
457}
458EXPORT_SYMBOL(filemap_write_and_wait_range);
459
460/**
461 * replace_page_cache_page - replace a pagecache page with a new one
462 * @old: page to be replaced
463 * @new: page to replace with
464 * @gfp_mask: allocation mode
465 *
466 * This function replaces a page in the pagecache with a new one. On
467 * success it acquires the pagecache reference for the new page and
468 * drops it for the old page. Both the old and new pages must be
469 * locked. This function does not add the new page to the LRU, the
470 * caller must do that.
471 *
472 * The remove + add is atomic. The only way this function can fail is
473 * memory allocation failure.
474 */
475int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
476{
477 int error;
478
479 VM_BUG_ON_PAGE(!PageLocked(old), old);
480 VM_BUG_ON_PAGE(!PageLocked(new), new);
481 VM_BUG_ON_PAGE(new->mapping, new);
482
483 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
484 if (!error) {
485 struct address_space *mapping = old->mapping;
486 void (*freepage)(struct page *);
487
488 pgoff_t offset = old->index;
489 freepage = mapping->a_ops->freepage;
490
491 page_cache_get(new);
492 new->mapping = mapping;
493 new->index = offset;
494
495 spin_lock_irq(&mapping->tree_lock);
496 __delete_from_page_cache(old, NULL);
497 error = radix_tree_insert(&mapping->page_tree, offset, new);
498 BUG_ON(error);
499 mapping->nrpages++;
500 __inc_zone_page_state(new, NR_FILE_PAGES);
501 if (PageSwapBacked(new))
502 __inc_zone_page_state(new, NR_SHMEM);
503 spin_unlock_irq(&mapping->tree_lock);
504 /* mem_cgroup codes must not be called under tree_lock */
505 mem_cgroup_replace_page_cache(old, new);
506 radix_tree_preload_end();
507 if (freepage)
508 freepage(old);
509 page_cache_release(old);
510 }
511
512 return error;
513}
514EXPORT_SYMBOL_GPL(replace_page_cache_page);
515
516static int page_cache_tree_insert(struct address_space *mapping,
517 struct page *page, void **shadowp)
518{
519 struct radix_tree_node *node;
520 void **slot;
521 int error;
522
523 error = __radix_tree_create(&mapping->page_tree, page->index,
524 &node, &slot);
525 if (error)
526 return error;
527 if (*slot) {
528 void *p;
529
530 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
531 if (!radix_tree_exceptional_entry(p))
532 return -EEXIST;
533 if (shadowp)
534 *shadowp = p;
535 mapping->nrshadows--;
536 if (node)
537 workingset_node_shadows_dec(node);
538 }
539 radix_tree_replace_slot(slot, page);
540 mapping->nrpages++;
541 if (node) {
542 workingset_node_pages_inc(node);
543 /*
544 * Don't track node that contains actual pages.
545 *
546 * Avoid acquiring the list_lru lock if already
547 * untracked. The list_empty() test is safe as
548 * node->private_list is protected by
549 * mapping->tree_lock.
550 */
551 if (!list_empty(&node->private_list))
552 list_lru_del(&workingset_shadow_nodes,
553 &node->private_list);
554 }
555 return 0;
556}
557
558static int __add_to_page_cache_locked(struct page *page,
559 struct address_space *mapping,
560 pgoff_t offset, gfp_t gfp_mask,
561 void **shadowp)
562{
563 int error;
564
565 VM_BUG_ON_PAGE(!PageLocked(page), page);
566 VM_BUG_ON_PAGE(PageSwapBacked(page), page);
567
568 error = mem_cgroup_charge_file(page, current->mm,
569 gfp_mask & GFP_RECLAIM_MASK);
570 if (error)
571 return error;
572
573 error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
574 if (error) {
575 mem_cgroup_uncharge_cache_page(page);
576 return error;
577 }
578
579 page_cache_get(page);
580 page->mapping = mapping;
581 page->index = offset;
582
583 spin_lock_irq(&mapping->tree_lock);
584 error = page_cache_tree_insert(mapping, page, shadowp);
585 radix_tree_preload_end();
586 if (unlikely(error))
587 goto err_insert;
588 __inc_zone_page_state(page, NR_FILE_PAGES);
589 spin_unlock_irq(&mapping->tree_lock);
590 trace_mm_filemap_add_to_page_cache(page);
591 return 0;
592err_insert:
593 page->mapping = NULL;
594 /* Leave page->index set: truncation relies upon it */
595 spin_unlock_irq(&mapping->tree_lock);
596 mem_cgroup_uncharge_cache_page(page);
597 page_cache_release(page);
598 return error;
599}
600
601/**
602 * add_to_page_cache_locked - add a locked page to the pagecache
603 * @page: page to add
604 * @mapping: the page's address_space
605 * @offset: page index
606 * @gfp_mask: page allocation mode
607 *
608 * This function is used to add a page to the pagecache. It must be locked.
609 * This function does not add the page to the LRU. The caller must do that.
610 */
611int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
612 pgoff_t offset, gfp_t gfp_mask)
613{
614 return __add_to_page_cache_locked(page, mapping, offset,
615 gfp_mask, NULL);
616}
617EXPORT_SYMBOL(add_to_page_cache_locked);
618
619int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
620 pgoff_t offset, gfp_t gfp_mask)
621{
622 void *shadow = NULL;
623 int ret;
624
625 __set_page_locked(page);
626 ret = __add_to_page_cache_locked(page, mapping, offset,
627 gfp_mask, &shadow);
628 if (unlikely(ret))
629 __clear_page_locked(page);
630 else {
631 /*
632 * The page might have been evicted from cache only
633 * recently, in which case it should be activated like
634 * any other repeatedly accessed page.
635 */
636 if (shadow && workingset_refault(shadow)) {
637 SetPageActive(page);
638 workingset_activation(page);
639 } else
640 ClearPageActive(page);
641 lru_cache_add(page);
642 }
643 return ret;
644}
645EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
646
647#ifdef CONFIG_NUMA
648struct page *__page_cache_alloc(gfp_t gfp)
649{
650 int n;
651 struct page *page;
652
653 if (cpuset_do_page_mem_spread()) {
654 unsigned int cpuset_mems_cookie;
655 do {
656 cpuset_mems_cookie = read_mems_allowed_begin();
657 n = cpuset_mem_spread_node();
658 page = alloc_pages_exact_node(n, gfp, 0);
659 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
660
661 return page;
662 }
663 return alloc_pages(gfp, 0);
664}
665EXPORT_SYMBOL(__page_cache_alloc);
666#endif
667
668/*
669 * In order to wait for pages to become available there must be
670 * waitqueues associated with pages. By using a hash table of
671 * waitqueues where the bucket discipline is to maintain all
672 * waiters on the same queue and wake all when any of the pages
673 * become available, and for the woken contexts to check to be
674 * sure the appropriate page became available, this saves space
675 * at a cost of "thundering herd" phenomena during rare hash
676 * collisions.
677 */
678static wait_queue_head_t *page_waitqueue(struct page *page)
679{
680 const struct zone *zone = page_zone(page);
681
682 return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
683}
684
685static inline void wake_up_page(struct page *page, int bit)
686{
687 __wake_up_bit(page_waitqueue(page), &page->flags, bit);
688}
689
690void wait_on_page_bit(struct page *page, int bit_nr)
691{
692 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
693
694 if (test_bit(bit_nr, &page->flags))
695 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
696 TASK_UNINTERRUPTIBLE);
697}
698EXPORT_SYMBOL(wait_on_page_bit);
699
700int wait_on_page_bit_killable(struct page *page, int bit_nr)
701{
702 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
703
704 if (!test_bit(bit_nr, &page->flags))
705 return 0;
706
707 return __wait_on_bit(page_waitqueue(page), &wait,
708 sleep_on_page_killable, TASK_KILLABLE);
709}
710
711/**
712 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
713 * @page: Page defining the wait queue of interest
714 * @waiter: Waiter to add to the queue
715 *
716 * Add an arbitrary @waiter to the wait queue for the nominated @page.
717 */
718void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
719{
720 wait_queue_head_t *q = page_waitqueue(page);
721 unsigned long flags;
722
723 spin_lock_irqsave(&q->lock, flags);
724 __add_wait_queue(q, waiter);
725 spin_unlock_irqrestore(&q->lock, flags);
726}
727EXPORT_SYMBOL_GPL(add_page_wait_queue);
728
729/**
730 * unlock_page - unlock a locked page
731 * @page: the page
732 *
733 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
734 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
735 * mechananism between PageLocked pages and PageWriteback pages is shared.
736 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
737 *
738 * The mb is necessary to enforce ordering between the clear_bit and the read
739 * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
740 */
741void unlock_page(struct page *page)
742{
743 VM_BUG_ON_PAGE(!PageLocked(page), page);
744 clear_bit_unlock(PG_locked, &page->flags);
745 smp_mb__after_clear_bit();
746 wake_up_page(page, PG_locked);
747}
748EXPORT_SYMBOL(unlock_page);
749
750/**
751 * end_page_writeback - end writeback against a page
752 * @page: the page
753 */
754void end_page_writeback(struct page *page)
755{
756 if (TestClearPageReclaim(page))
757 rotate_reclaimable_page(page);
758
759 if (!test_clear_page_writeback(page))
760 BUG();
761
762 smp_mb__after_clear_bit();
763 wake_up_page(page, PG_writeback);
764}
765EXPORT_SYMBOL(end_page_writeback);
766
767/**
768 * __lock_page - get a lock on the page, assuming we need to sleep to get it
769 * @page: the page to lock
770 */
771void __lock_page(struct page *page)
772{
773 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
774
775 __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
776 TASK_UNINTERRUPTIBLE);
777}
778EXPORT_SYMBOL(__lock_page);
779
780int __lock_page_killable(struct page *page)
781{
782 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
783
784 return __wait_on_bit_lock(page_waitqueue(page), &wait,
785 sleep_on_page_killable, TASK_KILLABLE);
786}
787EXPORT_SYMBOL_GPL(__lock_page_killable);
788
789int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
790 unsigned int flags)
791{
792 if (flags & FAULT_FLAG_ALLOW_RETRY) {
793 /*
794 * CAUTION! In this case, mmap_sem is not released
795 * even though return 0.
796 */
797 if (flags & FAULT_FLAG_RETRY_NOWAIT)
798 return 0;
799
800 up_read(&mm->mmap_sem);
801 if (flags & FAULT_FLAG_KILLABLE)
802 wait_on_page_locked_killable(page);
803 else
804 wait_on_page_locked(page);
805 return 0;
806 } else {
807 if (flags & FAULT_FLAG_KILLABLE) {
808 int ret;
809
810 ret = __lock_page_killable(page);
811 if (ret) {
812 up_read(&mm->mmap_sem);
813 return 0;
814 }
815 } else
816 __lock_page(page);
817 return 1;
818 }
819}
820
821/**
822 * page_cache_next_hole - find the next hole (not-present entry)
823 * @mapping: mapping
824 * @index: index
825 * @max_scan: maximum range to search
826 *
827 * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
828 * lowest indexed hole.
829 *
830 * Returns: the index of the hole if found, otherwise returns an index
831 * outside of the set specified (in which case 'return - index >=
832 * max_scan' will be true). In rare cases of index wrap-around, 0 will
833 * be returned.
834 *
835 * page_cache_next_hole may be called under rcu_read_lock. However,
836 * like radix_tree_gang_lookup, this will not atomically search a
837 * snapshot of the tree at a single point in time. For example, if a
838 * hole is created at index 5, then subsequently a hole is created at
839 * index 10, page_cache_next_hole covering both indexes may return 10
840 * if called under rcu_read_lock.
841 */
842pgoff_t page_cache_next_hole(struct address_space *mapping,
843 pgoff_t index, unsigned long max_scan)
844{
845 unsigned long i;
846
847 for (i = 0; i < max_scan; i++) {
848 struct page *page;
849
850 page = radix_tree_lookup(&mapping->page_tree, index);
851 if (!page || radix_tree_exceptional_entry(page))
852 break;
853 index++;
854 if (index == 0)
855 break;
856 }
857
858 return index;
859}
860EXPORT_SYMBOL(page_cache_next_hole);
861
862/**
863 * page_cache_prev_hole - find the prev hole (not-present entry)
864 * @mapping: mapping
865 * @index: index
866 * @max_scan: maximum range to search
867 *
868 * Search backwards in the range [max(index-max_scan+1, 0), index] for
869 * the first hole.
870 *
871 * Returns: the index of the hole if found, otherwise returns an index
872 * outside of the set specified (in which case 'index - return >=
873 * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
874 * will be returned.
875 *
876 * page_cache_prev_hole may be called under rcu_read_lock. However,
877 * like radix_tree_gang_lookup, this will not atomically search a
878 * snapshot of the tree at a single point in time. For example, if a
879 * hole is created at index 10, then subsequently a hole is created at
880 * index 5, page_cache_prev_hole covering both indexes may return 5 if
881 * called under rcu_read_lock.
882 */
883pgoff_t page_cache_prev_hole(struct address_space *mapping,
884 pgoff_t index, unsigned long max_scan)
885{
886 unsigned long i;
887
888 for (i = 0; i < max_scan; i++) {
889 struct page *page;
890
891 page = radix_tree_lookup(&mapping->page_tree, index);
892 if (!page || radix_tree_exceptional_entry(page))
893 break;
894 index--;
895 if (index == ULONG_MAX)
896 break;
897 }
898
899 return index;
900}
901EXPORT_SYMBOL(page_cache_prev_hole);
902
903/**
904 * find_get_entry - find and get a page cache entry
905 * @mapping: the address_space to search
906 * @offset: the page cache index
907 *
908 * Looks up the page cache slot at @mapping & @offset. If there is a
909 * page cache page, it is returned with an increased refcount.
910 *
911 * If the slot holds a shadow entry of a previously evicted page, or a
912 * swap entry from shmem/tmpfs, it is returned.
913 *
914 * Otherwise, %NULL is returned.
915 */
916struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
917{
918 void **pagep;
919 struct page *page;
920
921 rcu_read_lock();
922repeat:
923 page = NULL;
924 pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
925 if (pagep) {
926 page = radix_tree_deref_slot(pagep);
927 if (unlikely(!page))
928 goto out;
929 if (radix_tree_exception(page)) {
930 if (radix_tree_deref_retry(page))
931 goto repeat;
932 /*
933 * A shadow entry of a recently evicted page,
934 * or a swap entry from shmem/tmpfs. Return
935 * it without attempting to raise page count.
936 */
937 goto out;
938 }
939 if (!page_cache_get_speculative(page))
940 goto repeat;
941
942 /*
943 * Has the page moved?
944 * This is part of the lockless pagecache protocol. See
945 * include/linux/pagemap.h for details.
946 */
947 if (unlikely(page != *pagep)) {
948 page_cache_release(page);
949 goto repeat;
950 }
951 }
952out:
953 rcu_read_unlock();
954
955 return page;
956}
957EXPORT_SYMBOL(find_get_entry);
958
959/**
960 * find_get_page - find and get a page reference
961 * @mapping: the address_space to search
962 * @offset: the page index
963 *
964 * Looks up the page cache slot at @mapping & @offset. If there is a
965 * page cache page, it is returned with an increased refcount.
966 *
967 * Otherwise, %NULL is returned.
968 */
969struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
970{
971 struct page *page = find_get_entry(mapping, offset);
972
973 if (radix_tree_exceptional_entry(page))
974 page = NULL;
975 return page;
976}
977EXPORT_SYMBOL(find_get_page);
978
979/**
980 * find_lock_entry - locate, pin and lock a page cache entry
981 * @mapping: the address_space to search
982 * @offset: the page cache index
983 *
984 * Looks up the page cache slot at @mapping & @offset. If there is a
985 * page cache page, it is returned locked and with an increased
986 * refcount.
987 *
988 * If the slot holds a shadow entry of a previously evicted page, or a
989 * swap entry from shmem/tmpfs, it is returned.
990 *
991 * Otherwise, %NULL is returned.
992 *
993 * find_lock_entry() may sleep.
994 */
995struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
996{
997 struct page *page;
998
999repeat:
1000 page = find_get_entry(mapping, offset);
1001 if (page && !radix_tree_exception(page)) {
1002 lock_page(page);
1003 /* Has the page been truncated? */
1004 if (unlikely(page->mapping != mapping)) {
1005 unlock_page(page);
1006 page_cache_release(page);
1007 goto repeat;
1008 }
1009 VM_BUG_ON_PAGE(page->index != offset, page);
1010 }
1011 return page;
1012}
1013EXPORT_SYMBOL(find_lock_entry);
1014
1015/**
1016 * find_lock_page - locate, pin and lock a pagecache page
1017 * @mapping: the address_space to search
1018 * @offset: the page index
1019 *
1020 * Looks up the page cache slot at @mapping & @offset. If there is a
1021 * page cache page, it is returned locked and with an increased
1022 * refcount.
1023 *
1024 * Otherwise, %NULL is returned.
1025 *
1026 * find_lock_page() may sleep.
1027 */
1028struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
1029{
1030 struct page *page = find_lock_entry(mapping, offset);
1031
1032 if (radix_tree_exceptional_entry(page))
1033 page = NULL;
1034 return page;
1035}
1036EXPORT_SYMBOL(find_lock_page);
1037
1038/**
1039 * find_or_create_page - locate or add a pagecache page
1040 * @mapping: the page's address_space
1041 * @index: the page's index into the mapping
1042 * @gfp_mask: page allocation mode
1043 *
1044 * Looks up the page cache slot at @mapping & @offset. If there is a
1045 * page cache page, it is returned locked and with an increased
1046 * refcount.
1047 *
1048 * If the page is not present, a new page is allocated using @gfp_mask
1049 * and added to the page cache and the VM's LRU list. The page is
1050 * returned locked and with an increased refcount.
1051 *
1052 * On memory exhaustion, %NULL is returned.
1053 *
1054 * find_or_create_page() may sleep, even if @gfp_flags specifies an
1055 * atomic allocation!
1056 */
1057struct page *find_or_create_page(struct address_space *mapping,
1058 pgoff_t index, gfp_t gfp_mask)
1059{
1060 struct page *page;
1061 int err;
1062repeat:
1063 page = find_lock_page(mapping, index);
1064 if (!page) {
1065 page = __page_cache_alloc(gfp_mask);
1066 if (!page)
1067 return NULL;
1068 /*
1069 * We want a regular kernel memory (not highmem or DMA etc)
1070 * allocation for the radix tree nodes, but we need to honour
1071 * the context-specific requirements the caller has asked for.
1072 * GFP_RECLAIM_MASK collects those requirements.
1073 */
1074 err = add_to_page_cache_lru(page, mapping, index,
1075 (gfp_mask & GFP_RECLAIM_MASK));
1076 if (unlikely(err)) {
1077 page_cache_release(page);
1078 page = NULL;
1079 if (err == -EEXIST)
1080 goto repeat;
1081 }
1082 }
1083 return page;
1084}
1085EXPORT_SYMBOL(find_or_create_page);
1086
1087/**
1088 * find_get_entries - gang pagecache lookup
1089 * @mapping: The address_space to search
1090 * @start: The starting page cache index
1091 * @nr_entries: The maximum number of entries
1092 * @entries: Where the resulting entries are placed
1093 * @indices: The cache indices corresponding to the entries in @entries
1094 *
1095 * find_get_entries() will search for and return a group of up to
1096 * @nr_entries entries in the mapping. The entries are placed at
1097 * @entries. find_get_entries() takes a reference against any actual
1098 * pages it returns.
1099 *
1100 * The search returns a group of mapping-contiguous page cache entries
1101 * with ascending indexes. There may be holes in the indices due to
1102 * not-present pages.
1103 *
1104 * Any shadow entries of evicted pages, or swap entries from
1105 * shmem/tmpfs, are included in the returned array.
1106 *
1107 * find_get_entries() returns the number of pages and shadow entries
1108 * which were found.
1109 */
1110unsigned find_get_entries(struct address_space *mapping,
1111 pgoff_t start, unsigned int nr_entries,
1112 struct page **entries, pgoff_t *indices)
1113{
1114 void **slot;
1115 unsigned int ret = 0;
1116 struct radix_tree_iter iter;
1117
1118 if (!nr_entries)
1119 return 0;
1120
1121 rcu_read_lock();
1122restart:
1123 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1124 struct page *page;
1125repeat:
1126 page = radix_tree_deref_slot(slot);
1127 if (unlikely(!page))
1128 continue;
1129 if (radix_tree_exception(page)) {
1130 if (radix_tree_deref_retry(page))
1131 goto restart;
1132 /*
1133 * A shadow entry of a recently evicted page,
1134 * or a swap entry from shmem/tmpfs. Return
1135 * it without attempting to raise page count.
1136 */
1137 goto export;
1138 }
1139 if (!page_cache_get_speculative(page))
1140 goto repeat;
1141
1142 /* Has the page moved? */
1143 if (unlikely(page != *slot)) {
1144 page_cache_release(page);
1145 goto repeat;
1146 }
1147export:
1148 indices[ret] = iter.index;
1149 entries[ret] = page;
1150 if (++ret == nr_entries)
1151 break;
1152 }
1153 rcu_read_unlock();
1154 return ret;
1155}
1156
1157/**
1158 * find_get_pages - gang pagecache lookup
1159 * @mapping: The address_space to search
1160 * @start: The starting page index
1161 * @nr_pages: The maximum number of pages
1162 * @pages: Where the resulting pages are placed
1163 *
1164 * find_get_pages() will search for and return a group of up to
1165 * @nr_pages pages in the mapping. The pages are placed at @pages.
1166 * find_get_pages() takes a reference against the returned pages.
1167 *
1168 * The search returns a group of mapping-contiguous pages with ascending
1169 * indexes. There may be holes in the indices due to not-present pages.
1170 *
1171 * find_get_pages() returns the number of pages which were found.
1172 */
1173unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1174 unsigned int nr_pages, struct page **pages)
1175{
1176 struct radix_tree_iter iter;
1177 void **slot;
1178 unsigned ret = 0;
1179
1180 if (unlikely(!nr_pages))
1181 return 0;
1182
1183 rcu_read_lock();
1184restart:
1185 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1186 struct page *page;
1187repeat:
1188 page = radix_tree_deref_slot(slot);
1189 if (unlikely(!page))
1190 continue;
1191
1192 if (radix_tree_exception(page)) {
1193 if (radix_tree_deref_retry(page)) {
1194 /*
1195 * Transient condition which can only trigger
1196 * when entry at index 0 moves out of or back
1197 * to root: none yet gotten, safe to restart.
1198 */
1199 WARN_ON(iter.index);
1200 goto restart;
1201 }
1202 /*
1203 * A shadow entry of a recently evicted page,
1204 * or a swap entry from shmem/tmpfs. Skip
1205 * over it.
1206 */
1207 continue;
1208 }
1209
1210 if (!page_cache_get_speculative(page))
1211 goto repeat;
1212
1213 /* Has the page moved? */
1214 if (unlikely(page != *slot)) {
1215 page_cache_release(page);
1216 goto repeat;
1217 }
1218
1219 pages[ret] = page;
1220 if (++ret == nr_pages)
1221 break;
1222 }
1223
1224 rcu_read_unlock();
1225 return ret;
1226}
1227
1228/**
1229 * find_get_pages_contig - gang contiguous pagecache lookup
1230 * @mapping: The address_space to search
1231 * @index: The starting page index
1232 * @nr_pages: The maximum number of pages
1233 * @pages: Where the resulting pages are placed
1234 *
1235 * find_get_pages_contig() works exactly like find_get_pages(), except
1236 * that the returned number of pages are guaranteed to be contiguous.
1237 *
1238 * find_get_pages_contig() returns the number of pages which were found.
1239 */
1240unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1241 unsigned int nr_pages, struct page **pages)
1242{
1243 struct radix_tree_iter iter;
1244 void **slot;
1245 unsigned int ret = 0;
1246
1247 if (unlikely(!nr_pages))
1248 return 0;
1249
1250 rcu_read_lock();
1251restart:
1252 radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1253 struct page *page;
1254repeat:
1255 page = radix_tree_deref_slot(slot);
1256 /* The hole, there no reason to continue */
1257 if (unlikely(!page))
1258 break;
1259
1260 if (radix_tree_exception(page)) {
1261 if (radix_tree_deref_retry(page)) {
1262 /*
1263 * Transient condition which can only trigger
1264 * when entry at index 0 moves out of or back
1265 * to root: none yet gotten, safe to restart.
1266 */
1267 goto restart;
1268 }
1269 /*
1270 * A shadow entry of a recently evicted page,
1271 * or a swap entry from shmem/tmpfs. Stop
1272 * looking for contiguous pages.
1273 */
1274 break;
1275 }
1276
1277 if (!page_cache_get_speculative(page))
1278 goto repeat;
1279
1280 /* Has the page moved? */
1281 if (unlikely(page != *slot)) {
1282 page_cache_release(page);
1283 goto repeat;
1284 }
1285
1286 /*
1287 * must check mapping and index after taking the ref.
1288 * otherwise we can get both false positives and false
1289 * negatives, which is just confusing to the caller.
1290 */
1291 if (page->mapping == NULL || page->index != iter.index) {
1292 page_cache_release(page);
1293 break;
1294 }
1295
1296 pages[ret] = page;
1297 if (++ret == nr_pages)
1298 break;
1299 }
1300 rcu_read_unlock();
1301 return ret;
1302}
1303EXPORT_SYMBOL(find_get_pages_contig);
1304
1305/**
1306 * find_get_pages_tag - find and return pages that match @tag
1307 * @mapping: the address_space to search
1308 * @index: the starting page index
1309 * @tag: the tag index
1310 * @nr_pages: the maximum number of pages
1311 * @pages: where the resulting pages are placed
1312 *
1313 * Like find_get_pages, except we only return pages which are tagged with
1314 * @tag. We update @index to index the next page for the traversal.
1315 */
1316unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1317 int tag, unsigned int nr_pages, struct page **pages)
1318{
1319 struct radix_tree_iter iter;
1320 void **slot;
1321 unsigned ret = 0;
1322
1323 if (unlikely(!nr_pages))
1324 return 0;
1325
1326 rcu_read_lock();
1327restart:
1328 radix_tree_for_each_tagged(slot, &mapping->page_tree,
1329 &iter, *index, tag) {
1330 struct page *page;
1331repeat:
1332 page = radix_tree_deref_slot(slot);
1333 if (unlikely(!page))
1334 continue;
1335
1336 if (radix_tree_exception(page)) {
1337 if (radix_tree_deref_retry(page)) {
1338 /*
1339 * Transient condition which can only trigger
1340 * when entry at index 0 moves out of or back
1341 * to root: none yet gotten, safe to restart.
1342 */
1343 goto restart;
1344 }
1345 /*
1346 * A shadow entry of a recently evicted page.
1347 *
1348 * Those entries should never be tagged, but
1349 * this tree walk is lockless and the tags are
1350 * looked up in bulk, one radix tree node at a
1351 * time, so there is a sizable window for page
1352 * reclaim to evict a page we saw tagged.
1353 *
1354 * Skip over it.
1355 */
1356 continue;
1357 }
1358
1359 if (!page_cache_get_speculative(page))
1360 goto repeat;
1361
1362 /* Has the page moved? */
1363 if (unlikely(page != *slot)) {
1364 page_cache_release(page);
1365 goto repeat;
1366 }
1367
1368 pages[ret] = page;
1369 if (++ret == nr_pages)
1370 break;
1371 }
1372
1373 rcu_read_unlock();
1374
1375 if (ret)
1376 *index = pages[ret - 1]->index + 1;
1377
1378 return ret;
1379}
1380EXPORT_SYMBOL(find_get_pages_tag);
1381
1382/**
1383 * grab_cache_page_nowait - returns locked page at given index in given cache
1384 * @mapping: target address_space
1385 * @index: the page index
1386 *
1387 * Same as grab_cache_page(), but do not wait if the page is unavailable.
1388 * This is intended for speculative data generators, where the data can
1389 * be regenerated if the page couldn't be grabbed. This routine should
1390 * be safe to call while holding the lock for another page.
1391 *
1392 * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1393 * and deadlock against the caller's locked page.
1394 */
1395struct page *
1396grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1397{
1398 struct page *page = find_get_page(mapping, index);
1399
1400 if (page) {
1401 if (trylock_page(page))
1402 return page;
1403 page_cache_release(page);
1404 return NULL;
1405 }
1406 page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1407 if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1408 page_cache_release(page);
1409 page = NULL;
1410 }
1411 return page;
1412}
1413EXPORT_SYMBOL(grab_cache_page_nowait);
1414
1415/*
1416 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1417 * a _large_ part of the i/o request. Imagine the worst scenario:
1418 *
1419 * ---R__________________________________________B__________
1420 * ^ reading here ^ bad block(assume 4k)
1421 *
1422 * read(R) => miss => readahead(R...B) => media error => frustrating retries
1423 * => failing the whole request => read(R) => read(R+1) =>
1424 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1425 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1426 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1427 *
1428 * It is going insane. Fix it by quickly scaling down the readahead size.
1429 */
1430static void shrink_readahead_size_eio(struct file *filp,
1431 struct file_ra_state *ra)
1432{
1433 ra->ra_pages /= 4;
1434}
1435
1436/**
1437 * do_generic_file_read - generic file read routine
1438 * @filp: the file to read
1439 * @ppos: current file position
1440 * @iter: data destination
1441 * @written: already copied
1442 *
1443 * This is a generic file read routine, and uses the
1444 * mapping->a_ops->readpage() function for the actual low-level stuff.
1445 *
1446 * This is really ugly. But the goto's actually try to clarify some
1447 * of the logic when it comes to error handling etc.
1448 */
1449static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1450 struct iov_iter *iter, ssize_t written)
1451{
1452 struct address_space *mapping = filp->f_mapping;
1453 struct inode *inode = mapping->host;
1454 struct file_ra_state *ra = &filp->f_ra;
1455 pgoff_t index;
1456 pgoff_t last_index;
1457 pgoff_t prev_index;
1458 unsigned long offset; /* offset into pagecache page */
1459 unsigned int prev_offset;
1460 int error = 0;
1461
1462 index = *ppos >> PAGE_CACHE_SHIFT;
1463 prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1464 prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1465 last_index = (*ppos + iter->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1466 offset = *ppos & ~PAGE_CACHE_MASK;
1467
1468 for (;;) {
1469 struct page *page;
1470 pgoff_t end_index;
1471 loff_t isize;
1472 unsigned long nr, ret;
1473
1474 cond_resched();
1475find_page:
1476 page = find_get_page(mapping, index);
1477 if (!page) {
1478 page_cache_sync_readahead(mapping,
1479 ra, filp,
1480 index, last_index - index);
1481 page = find_get_page(mapping, index);
1482 if (unlikely(page == NULL))
1483 goto no_cached_page;
1484 }
1485 if (PageReadahead(page)) {
1486 page_cache_async_readahead(mapping,
1487 ra, filp, page,
1488 index, last_index - index);
1489 }
1490 if (!PageUptodate(page)) {
1491 if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1492 !mapping->a_ops->is_partially_uptodate)
1493 goto page_not_up_to_date;
1494 if (!trylock_page(page))
1495 goto page_not_up_to_date;
1496 /* Did it get truncated before we got the lock? */
1497 if (!page->mapping)
1498 goto page_not_up_to_date_locked;
1499 if (!mapping->a_ops->is_partially_uptodate(page,
1500 offset, iter->count))
1501 goto page_not_up_to_date_locked;
1502 unlock_page(page);
1503 }
1504page_ok:
1505 /*
1506 * i_size must be checked after we know the page is Uptodate.
1507 *
1508 * Checking i_size after the check allows us to calculate
1509 * the correct value for "nr", which means the zero-filled
1510 * part of the page is not copied back to userspace (unless
1511 * another truncate extends the file - this is desired though).
1512 */
1513
1514 isize = i_size_read(inode);
1515 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1516 if (unlikely(!isize || index > end_index)) {
1517 page_cache_release(page);
1518 goto out;
1519 }
1520
1521 /* nr is the maximum number of bytes to copy from this page */
1522 nr = PAGE_CACHE_SIZE;
1523 if (index == end_index) {
1524 nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1525 if (nr <= offset) {
1526 page_cache_release(page);
1527 goto out;
1528 }
1529 }
1530 nr = nr - offset;
1531
1532 /* If users can be writing to this page using arbitrary
1533 * virtual addresses, take care about potential aliasing
1534 * before reading the page on the kernel side.
1535 */
1536 if (mapping_writably_mapped(mapping))
1537 flush_dcache_page(page);
1538
1539 /*
1540 * When a sequential read accesses a page several times,
1541 * only mark it as accessed the first time.
1542 */
1543 if (prev_index != index || offset != prev_offset)
1544 mark_page_accessed(page);
1545 prev_index = index;
1546
1547 /*
1548 * Ok, we have the page, and it's up-to-date, so
1549 * now we can copy it to user space...
1550 */
1551
1552 ret = copy_page_to_iter(page, offset, nr, iter);
1553 offset += ret;
1554 index += offset >> PAGE_CACHE_SHIFT;
1555 offset &= ~PAGE_CACHE_MASK;
1556 prev_offset = offset;
1557
1558 page_cache_release(page);
1559 written += ret;
1560 if (!iov_iter_count(iter))
1561 goto out;
1562 if (ret < nr) {
1563 error = -EFAULT;
1564 goto out;
1565 }
1566 continue;
1567
1568page_not_up_to_date:
1569 /* Get exclusive access to the page ... */
1570 error = lock_page_killable(page);
1571 if (unlikely(error))
1572 goto readpage_error;
1573
1574page_not_up_to_date_locked:
1575 /* Did it get truncated before we got the lock? */
1576 if (!page->mapping) {
1577 unlock_page(page);
1578 page_cache_release(page);
1579 continue;
1580 }
1581
1582 /* Did somebody else fill it already? */
1583 if (PageUptodate(page)) {
1584 unlock_page(page);
1585 goto page_ok;
1586 }
1587
1588readpage:
1589 /*
1590 * A previous I/O error may have been due to temporary
1591 * failures, eg. multipath errors.
1592 * PG_error will be set again if readpage fails.
1593 */
1594 ClearPageError(page);
1595 /* Start the actual read. The read will unlock the page. */
1596 error = mapping->a_ops->readpage(filp, page);
1597
1598 if (unlikely(error)) {
1599 if (error == AOP_TRUNCATED_PAGE) {
1600 page_cache_release(page);
1601 error = 0;
1602 goto find_page;
1603 }
1604 goto readpage_error;
1605 }
1606
1607 if (!PageUptodate(page)) {
1608 error = lock_page_killable(page);
1609 if (unlikely(error))
1610 goto readpage_error;
1611 if (!PageUptodate(page)) {
1612 if (page->mapping == NULL) {
1613 /*
1614 * invalidate_mapping_pages got it
1615 */
1616 unlock_page(page);
1617 page_cache_release(page);
1618 goto find_page;
1619 }
1620 unlock_page(page);
1621 shrink_readahead_size_eio(filp, ra);
1622 error = -EIO;
1623 goto readpage_error;
1624 }
1625 unlock_page(page);
1626 }
1627
1628 goto page_ok;
1629
1630readpage_error:
1631 /* UHHUH! A synchronous read error occurred. Report it */
1632 page_cache_release(page);
1633 goto out;
1634
1635no_cached_page:
1636 /*
1637 * Ok, it wasn't cached, so we need to create a new
1638 * page..
1639 */
1640 page = page_cache_alloc_cold(mapping);
1641 if (!page) {
1642 error = -ENOMEM;
1643 goto out;
1644 }
1645 error = add_to_page_cache_lru(page, mapping,
1646 index, GFP_KERNEL);
1647 if (error) {
1648 page_cache_release(page);
1649 if (error == -EEXIST) {
1650 error = 0;
1651 goto find_page;
1652 }
1653 goto out;
1654 }
1655 goto readpage;
1656 }
1657
1658out:
1659 ra->prev_pos = prev_index;
1660 ra->prev_pos <<= PAGE_CACHE_SHIFT;
1661 ra->prev_pos |= prev_offset;
1662
1663 *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1664 file_accessed(filp);
1665 return written ? written : error;
1666}
1667
1668/*
1669 * Performs necessary checks before doing a write
1670 * @iov: io vector request
1671 * @nr_segs: number of segments in the iovec
1672 * @count: number of bytes to write
1673 * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1674 *
1675 * Adjust number of segments and amount of bytes to write (nr_segs should be
1676 * properly initialized first). Returns appropriate error code that caller
1677 * should return or zero in case that write should be allowed.
1678 */
1679int generic_segment_checks(const struct iovec *iov,
1680 unsigned long *nr_segs, size_t *count, int access_flags)
1681{
1682 unsigned long seg;
1683 size_t cnt = 0;
1684 for (seg = 0; seg < *nr_segs; seg++) {
1685 const struct iovec *iv = &iov[seg];
1686
1687 /*
1688 * If any segment has a negative length, or the cumulative
1689 * length ever wraps negative then return -EINVAL.
1690 */
1691 cnt += iv->iov_len;
1692 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1693 return -EINVAL;
1694 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1695 continue;
1696 if (seg == 0)
1697 return -EFAULT;
1698 *nr_segs = seg;
1699 cnt -= iv->iov_len; /* This segment is no good */
1700 break;
1701 }
1702 *count = cnt;
1703 return 0;
1704}
1705EXPORT_SYMBOL(generic_segment_checks);
1706
1707/**
1708 * generic_file_aio_read - generic filesystem read routine
1709 * @iocb: kernel I/O control block
1710 * @iov: io vector request
1711 * @nr_segs: number of segments in the iovec
1712 * @pos: current file position
1713 *
1714 * This is the "read()" routine for all filesystems
1715 * that can use the page cache directly.
1716 */
1717ssize_t
1718generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1719 unsigned long nr_segs, loff_t pos)
1720{
1721 struct file *filp = iocb->ki_filp;
1722 ssize_t retval;
1723 size_t count;
1724 loff_t *ppos = &iocb->ki_pos;
1725 struct iov_iter i;
1726
1727 count = 0;
1728 retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1729 if (retval)
1730 return retval;
1731 iov_iter_init(&i, iov, nr_segs, count, 0);
1732
1733 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1734 if (filp->f_flags & O_DIRECT) {
1735 loff_t size;
1736 struct address_space *mapping;
1737 struct inode *inode;
1738
1739 mapping = filp->f_mapping;
1740 inode = mapping->host;
1741 if (!count)
1742 goto out; /* skip atime */
1743 size = i_size_read(inode);
1744 retval = filemap_write_and_wait_range(mapping, pos,
1745 pos + iov_length(iov, nr_segs) - 1);
1746 if (!retval) {
1747 retval = mapping->a_ops->direct_IO(READ, iocb,
1748 iov, pos, nr_segs);
1749 }
1750 if (retval > 0) {
1751 *ppos = pos + retval;
1752 count -= retval;
1753 /*
1754 * If we did a short DIO read we need to skip the
1755 * section of the iov that we've already read data into.
1756 */
1757 iov_iter_advance(&i, retval);
1758 }
1759
1760 /*
1761 * Btrfs can have a short DIO read if we encounter
1762 * compressed extents, so if there was an error, or if
1763 * we've already read everything we wanted to, or if
1764 * there was a short read because we hit EOF, go ahead
1765 * and return. Otherwise fallthrough to buffered io for
1766 * the rest of the read.
1767 */
1768 if (retval < 0 || !count || *ppos >= size) {
1769 file_accessed(filp);
1770 goto out;
1771 }
1772 }
1773
1774 retval = do_generic_file_read(filp, ppos, &i, retval);
1775out:
1776 return retval;
1777}
1778EXPORT_SYMBOL(generic_file_aio_read);
1779
1780#ifdef CONFIG_MMU
1781/**
1782 * page_cache_read - adds requested page to the page cache if not already there
1783 * @file: file to read
1784 * @offset: page index
1785 *
1786 * This adds the requested page to the page cache if it isn't already there,
1787 * and schedules an I/O to read in its contents from disk.
1788 */
1789static int page_cache_read(struct file *file, pgoff_t offset)
1790{
1791 struct address_space *mapping = file->f_mapping;
1792 struct page *page;
1793 int ret;
1794
1795 do {
1796 page = page_cache_alloc_cold(mapping);
1797 if (!page)
1798 return -ENOMEM;
1799
1800 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1801 if (ret == 0)
1802 ret = mapping->a_ops->readpage(file, page);
1803 else if (ret == -EEXIST)
1804 ret = 0; /* losing race to add is OK */
1805
1806 page_cache_release(page);
1807
1808 } while (ret == AOP_TRUNCATED_PAGE);
1809
1810 return ret;
1811}
1812
1813#define MMAP_LOTSAMISS (100)
1814
1815/*
1816 * Synchronous readahead happens when we don't even find
1817 * a page in the page cache at all.
1818 */
1819static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1820 struct file_ra_state *ra,
1821 struct file *file,
1822 pgoff_t offset)
1823{
1824 unsigned long ra_pages;
1825 struct address_space *mapping = file->f_mapping;
1826
1827 /* If we don't want any read-ahead, don't bother */
1828 if (vma->vm_flags & VM_RAND_READ)
1829 return;
1830 if (!ra->ra_pages)
1831 return;
1832
1833 if (vma->vm_flags & VM_SEQ_READ) {
1834 page_cache_sync_readahead(mapping, ra, file, offset,
1835 ra->ra_pages);
1836 return;
1837 }
1838
1839 /* Avoid banging the cache line if not needed */
1840 if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1841 ra->mmap_miss++;
1842
1843 /*
1844 * Do we miss much more than hit in this file? If so,
1845 * stop bothering with read-ahead. It will only hurt.
1846 */
1847 if (ra->mmap_miss > MMAP_LOTSAMISS)
1848 return;
1849
1850 /*
1851 * mmap read-around
1852 */
1853 ra_pages = max_sane_readahead(ra->ra_pages);
1854 ra->start = max_t(long, 0, offset - ra_pages / 2);
1855 ra->size = ra_pages;
1856 ra->async_size = ra_pages / 4;
1857 ra_submit(ra, mapping, file);
1858}
1859
1860/*
1861 * Asynchronous readahead happens when we find the page and PG_readahead,
1862 * so we want to possibly extend the readahead further..
1863 */
1864static void do_async_mmap_readahead(struct vm_area_struct *vma,
1865 struct file_ra_state *ra,
1866 struct file *file,
1867 struct page *page,
1868 pgoff_t offset)
1869{
1870 struct address_space *mapping = file->f_mapping;
1871
1872 /* If we don't want any read-ahead, don't bother */
1873 if (vma->vm_flags & VM_RAND_READ)
1874 return;
1875 if (ra->mmap_miss > 0)
1876 ra->mmap_miss--;
1877 if (PageReadahead(page))
1878 page_cache_async_readahead(mapping, ra, file,
1879 page, offset, ra->ra_pages);
1880}
1881
1882/**
1883 * filemap_fault - read in file data for page fault handling
1884 * @vma: vma in which the fault was taken
1885 * @vmf: struct vm_fault containing details of the fault
1886 *
1887 * filemap_fault() is invoked via the vma operations vector for a
1888 * mapped memory region to read in file data during a page fault.
1889 *
1890 * The goto's are kind of ugly, but this streamlines the normal case of having
1891 * it in the page cache, and handles the special cases reasonably without
1892 * having a lot of duplicated code.
1893 */
1894int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1895{
1896 int error;
1897 struct file *file = vma->vm_file;
1898 struct address_space *mapping = file->f_mapping;
1899 struct file_ra_state *ra = &file->f_ra;
1900 struct inode *inode = mapping->host;
1901 pgoff_t offset = vmf->pgoff;
1902 struct page *page;
1903 loff_t size;
1904 int ret = 0;
1905
1906 size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
1907 if (offset >= size >> PAGE_CACHE_SHIFT)
1908 return VM_FAULT_SIGBUS;
1909
1910 /*
1911 * Do we have something in the page cache already?
1912 */
1913 page = find_get_page(mapping, offset);
1914 if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1915 /*
1916 * We found the page, so try async readahead before
1917 * waiting for the lock.
1918 */
1919 do_async_mmap_readahead(vma, ra, file, page, offset);
1920 } else if (!page) {
1921 /* No page in the page cache at all */
1922 do_sync_mmap_readahead(vma, ra, file, offset);
1923 count_vm_event(PGMAJFAULT);
1924 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1925 ret = VM_FAULT_MAJOR;
1926retry_find:
1927 page = find_get_page(mapping, offset);
1928 if (!page)
1929 goto no_cached_page;
1930 }
1931
1932 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1933 page_cache_release(page);
1934 return ret | VM_FAULT_RETRY;
1935 }
1936
1937 /* Did it get truncated? */
1938 if (unlikely(page->mapping != mapping)) {
1939 unlock_page(page);
1940 put_page(page);
1941 goto retry_find;
1942 }
1943 VM_BUG_ON_PAGE(page->index != offset, page);
1944
1945 /*
1946 * We have a locked page in the page cache, now we need to check
1947 * that it's up-to-date. If not, it is going to be due to an error.
1948 */
1949 if (unlikely(!PageUptodate(page)))
1950 goto page_not_uptodate;
1951
1952 /*
1953 * Found the page and have a reference on it.
1954 * We must recheck i_size under page lock.
1955 */
1956 size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
1957 if (unlikely(offset >= size >> PAGE_CACHE_SHIFT)) {
1958 unlock_page(page);
1959 page_cache_release(page);
1960 return VM_FAULT_SIGBUS;
1961 }
1962
1963 vmf->page = page;
1964 return ret | VM_FAULT_LOCKED;
1965
1966no_cached_page:
1967 /*
1968 * We're only likely to ever get here if MADV_RANDOM is in
1969 * effect.
1970 */
1971 error = page_cache_read(file, offset);
1972
1973 /*
1974 * The page we want has now been added to the page cache.
1975 * In the unlikely event that someone removed it in the
1976 * meantime, we'll just come back here and read it again.
1977 */
1978 if (error >= 0)
1979 goto retry_find;
1980
1981 /*
1982 * An error return from page_cache_read can result if the
1983 * system is low on memory, or a problem occurs while trying
1984 * to schedule I/O.
1985 */
1986 if (error == -ENOMEM)
1987 return VM_FAULT_OOM;
1988 return VM_FAULT_SIGBUS;
1989
1990page_not_uptodate:
1991 /*
1992 * Umm, take care of errors if the page isn't up-to-date.
1993 * Try to re-read it _once_. We do this synchronously,
1994 * because there really aren't any performance issues here
1995 * and we need to check for errors.
1996 */
1997 ClearPageError(page);
1998 error = mapping->a_ops->readpage(file, page);
1999 if (!error) {
2000 wait_on_page_locked(page);
2001 if (!PageUptodate(page))
2002 error = -EIO;
2003 }
2004 page_cache_release(page);
2005
2006 if (!error || error == AOP_TRUNCATED_PAGE)
2007 goto retry_find;
2008
2009 /* Things didn't work out. Return zero to tell the mm layer so. */
2010 shrink_readahead_size_eio(file, ra);
2011 return VM_FAULT_SIGBUS;
2012}
2013EXPORT_SYMBOL(filemap_fault);
2014
2015void filemap_map_pages(struct vm_area_struct *vma, struct vm_fault *vmf)
2016{
2017 struct radix_tree_iter iter;
2018 void **slot;
2019 struct file *file = vma->vm_file;
2020 struct address_space *mapping = file->f_mapping;
2021 loff_t size;
2022 struct page *page;
2023 unsigned long address = (unsigned long) vmf->virtual_address;
2024 unsigned long addr;
2025 pte_t *pte;
2026
2027 rcu_read_lock();
2028 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, vmf->pgoff) {
2029 if (iter.index > vmf->max_pgoff)
2030 break;
2031repeat:
2032 page = radix_tree_deref_slot(slot);
2033 if (unlikely(!page))
2034 goto next;
2035 if (radix_tree_exception(page)) {
2036 if (radix_tree_deref_retry(page))
2037 break;
2038 else
2039 goto next;
2040 }
2041
2042 if (!page_cache_get_speculative(page))
2043 goto repeat;
2044
2045 /* Has the page moved? */
2046 if (unlikely(page != *slot)) {
2047 page_cache_release(page);
2048 goto repeat;
2049 }
2050
2051 if (!PageUptodate(page) ||
2052 PageReadahead(page) ||
2053 PageHWPoison(page))
2054 goto skip;
2055 if (!trylock_page(page))
2056 goto skip;
2057
2058 if (page->mapping != mapping || !PageUptodate(page))
2059 goto unlock;
2060
2061 size = round_up(i_size_read(mapping->host), PAGE_CACHE_SIZE);
2062 if (page->index >= size >> PAGE_CACHE_SHIFT)
2063 goto unlock;
2064
2065 pte = vmf->pte + page->index - vmf->pgoff;
2066 if (!pte_none(*pte))
2067 goto unlock;
2068
2069 if (file->f_ra.mmap_miss > 0)
2070 file->f_ra.mmap_miss--;
2071 addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
2072 do_set_pte(vma, addr, page, pte, false, false);
2073 unlock_page(page);
2074 goto next;
2075unlock:
2076 unlock_page(page);
2077skip:
2078 page_cache_release(page);
2079next:
2080 if (iter.index == vmf->max_pgoff)
2081 break;
2082 }
2083 rcu_read_unlock();
2084}
2085EXPORT_SYMBOL(filemap_map_pages);
2086
2087int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2088{
2089 struct page *page = vmf->page;
2090 struct inode *inode = file_inode(vma->vm_file);
2091 int ret = VM_FAULT_LOCKED;
2092
2093 sb_start_pagefault(inode->i_sb);
2094 file_update_time(vma->vm_file);
2095 lock_page(page);
2096 if (page->mapping != inode->i_mapping) {
2097 unlock_page(page);
2098 ret = VM_FAULT_NOPAGE;
2099 goto out;
2100 }
2101 /*
2102 * We mark the page dirty already here so that when freeze is in
2103 * progress, we are guaranteed that writeback during freezing will
2104 * see the dirty page and writeprotect it again.
2105 */
2106 set_page_dirty(page);
2107 wait_for_stable_page(page);
2108out:
2109 sb_end_pagefault(inode->i_sb);
2110 return ret;
2111}
2112EXPORT_SYMBOL(filemap_page_mkwrite);
2113
2114const struct vm_operations_struct generic_file_vm_ops = {
2115 .fault = filemap_fault,
2116 .map_pages = filemap_map_pages,
2117 .page_mkwrite = filemap_page_mkwrite,
2118 .remap_pages = generic_file_remap_pages,
2119};
2120
2121/* This is used for a general mmap of a disk file */
2122
2123int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2124{
2125 struct address_space *mapping = file->f_mapping;
2126
2127 if (!mapping->a_ops->readpage)
2128 return -ENOEXEC;
2129 file_accessed(file);
2130 vma->vm_ops = &generic_file_vm_ops;
2131 return 0;
2132}
2133
2134/*
2135 * This is for filesystems which do not implement ->writepage.
2136 */
2137int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2138{
2139 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2140 return -EINVAL;
2141 return generic_file_mmap(file, vma);
2142}
2143#else
2144int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2145{
2146 return -ENOSYS;
2147}
2148int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2149{
2150 return -ENOSYS;
2151}
2152#endif /* CONFIG_MMU */
2153
2154EXPORT_SYMBOL(generic_file_mmap);
2155EXPORT_SYMBOL(generic_file_readonly_mmap);
2156
2157static struct page *wait_on_page_read(struct page *page)
2158{
2159 if (!IS_ERR(page)) {
2160 wait_on_page_locked(page);
2161 if (!PageUptodate(page)) {
2162 page_cache_release(page);
2163 page = ERR_PTR(-EIO);
2164 }
2165 }
2166 return page;
2167}
2168
2169static struct page *__read_cache_page(struct address_space *mapping,
2170 pgoff_t index,
2171 int (*filler)(void *, struct page *),
2172 void *data,
2173 gfp_t gfp)
2174{
2175 struct page *page;
2176 int err;
2177repeat:
2178 page = find_get_page(mapping, index);
2179 if (!page) {
2180 page = __page_cache_alloc(gfp | __GFP_COLD);
2181 if (!page)
2182 return ERR_PTR(-ENOMEM);
2183 err = add_to_page_cache_lru(page, mapping, index, gfp);
2184 if (unlikely(err)) {
2185 page_cache_release(page);
2186 if (err == -EEXIST)
2187 goto repeat;
2188 /* Presumably ENOMEM for radix tree node */
2189 return ERR_PTR(err);
2190 }
2191 err = filler(data, page);
2192 if (err < 0) {
2193 page_cache_release(page);
2194 page = ERR_PTR(err);
2195 } else {
2196 page = wait_on_page_read(page);
2197 }
2198 }
2199 return page;
2200}
2201
2202static struct page *do_read_cache_page(struct address_space *mapping,
2203 pgoff_t index,
2204 int (*filler)(void *, struct page *),
2205 void *data,
2206 gfp_t gfp)
2207
2208{
2209 struct page *page;
2210 int err;
2211
2212retry:
2213 page = __read_cache_page(mapping, index, filler, data, gfp);
2214 if (IS_ERR(page))
2215 return page;
2216 if (PageUptodate(page))
2217 goto out;
2218
2219 lock_page(page);
2220 if (!page->mapping) {
2221 unlock_page(page);
2222 page_cache_release(page);
2223 goto retry;
2224 }
2225 if (PageUptodate(page)) {
2226 unlock_page(page);
2227 goto out;
2228 }
2229 err = filler(data, page);
2230 if (err < 0) {
2231 page_cache_release(page);
2232 return ERR_PTR(err);
2233 } else {
2234 page = wait_on_page_read(page);
2235 if (IS_ERR(page))
2236 return page;
2237 }
2238out:
2239 mark_page_accessed(page);
2240 return page;
2241}
2242
2243/**
2244 * read_cache_page - read into page cache, fill it if needed
2245 * @mapping: the page's address_space
2246 * @index: the page index
2247 * @filler: function to perform the read
2248 * @data: first arg to filler(data, page) function, often left as NULL
2249 *
2250 * Read into the page cache. If a page already exists, and PageUptodate() is
2251 * not set, try to fill the page and wait for it to become unlocked.
2252 *
2253 * If the page does not get brought uptodate, return -EIO.
2254 */
2255struct page *read_cache_page(struct address_space *mapping,
2256 pgoff_t index,
2257 int (*filler)(void *, struct page *),
2258 void *data)
2259{
2260 return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2261}
2262EXPORT_SYMBOL(read_cache_page);
2263
2264/**
2265 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2266 * @mapping: the page's address_space
2267 * @index: the page index
2268 * @gfp: the page allocator flags to use if allocating
2269 *
2270 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2271 * any new page allocations done using the specified allocation flags.
2272 *
2273 * If the page does not get brought uptodate, return -EIO.
2274 */
2275struct page *read_cache_page_gfp(struct address_space *mapping,
2276 pgoff_t index,
2277 gfp_t gfp)
2278{
2279 filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2280
2281 return do_read_cache_page(mapping, index, filler, NULL, gfp);
2282}
2283EXPORT_SYMBOL(read_cache_page_gfp);
2284
2285/*
2286 * Performs necessary checks before doing a write
2287 *
2288 * Can adjust writing position or amount of bytes to write.
2289 * Returns appropriate error code that caller should return or
2290 * zero in case that write should be allowed.
2291 */
2292inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2293{
2294 struct inode *inode = file->f_mapping->host;
2295 unsigned long limit = rlimit(RLIMIT_FSIZE);
2296
2297 if (unlikely(*pos < 0))
2298 return -EINVAL;
2299
2300 if (!isblk) {
2301 /* FIXME: this is for backwards compatibility with 2.4 */
2302 if (file->f_flags & O_APPEND)
2303 *pos = i_size_read(inode);
2304
2305 if (limit != RLIM_INFINITY) {
2306 if (*pos >= limit) {
2307 send_sig(SIGXFSZ, current, 0);
2308 return -EFBIG;
2309 }
2310 if (*count > limit - (typeof(limit))*pos) {
2311 *count = limit - (typeof(limit))*pos;
2312 }
2313 }
2314 }
2315
2316 /*
2317 * LFS rule
2318 */
2319 if (unlikely(*pos + *count > MAX_NON_LFS &&
2320 !(file->f_flags & O_LARGEFILE))) {
2321 if (*pos >= MAX_NON_LFS) {
2322 return -EFBIG;
2323 }
2324 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2325 *count = MAX_NON_LFS - (unsigned long)*pos;
2326 }
2327 }
2328
2329 /*
2330 * Are we about to exceed the fs block limit ?
2331 *
2332 * If we have written data it becomes a short write. If we have
2333 * exceeded without writing data we send a signal and return EFBIG.
2334 * Linus frestrict idea will clean these up nicely..
2335 */
2336 if (likely(!isblk)) {
2337 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2338 if (*count || *pos > inode->i_sb->s_maxbytes) {
2339 return -EFBIG;
2340 }
2341 /* zero-length writes at ->s_maxbytes are OK */
2342 }
2343
2344 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2345 *count = inode->i_sb->s_maxbytes - *pos;
2346 } else {
2347#ifdef CONFIG_BLOCK
2348 loff_t isize;
2349 if (bdev_read_only(I_BDEV(inode)))
2350 return -EPERM;
2351 isize = i_size_read(inode);
2352 if (*pos >= isize) {
2353 if (*count || *pos > isize)
2354 return -ENOSPC;
2355 }
2356
2357 if (*pos + *count > isize)
2358 *count = isize - *pos;
2359#else
2360 return -EPERM;
2361#endif
2362 }
2363 return 0;
2364}
2365EXPORT_SYMBOL(generic_write_checks);
2366
2367int pagecache_write_begin(struct file *file, struct address_space *mapping,
2368 loff_t pos, unsigned len, unsigned flags,
2369 struct page **pagep, void **fsdata)
2370{
2371 const struct address_space_operations *aops = mapping->a_ops;
2372
2373 return aops->write_begin(file, mapping, pos, len, flags,
2374 pagep, fsdata);
2375}
2376EXPORT_SYMBOL(pagecache_write_begin);
2377
2378int pagecache_write_end(struct file *file, struct address_space *mapping,
2379 loff_t pos, unsigned len, unsigned copied,
2380 struct page *page, void *fsdata)
2381{
2382 const struct address_space_operations *aops = mapping->a_ops;
2383
2384 mark_page_accessed(page);
2385 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2386}
2387EXPORT_SYMBOL(pagecache_write_end);
2388
2389ssize_t
2390generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2391 unsigned long *nr_segs, loff_t pos,
2392 size_t count, size_t ocount)
2393{
2394 struct file *file = iocb->ki_filp;
2395 struct address_space *mapping = file->f_mapping;
2396 struct inode *inode = mapping->host;
2397 ssize_t written;
2398 size_t write_len;
2399 pgoff_t end;
2400
2401 if (count != ocount)
2402 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2403
2404 write_len = iov_length(iov, *nr_segs);
2405 end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2406
2407 written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2408 if (written)
2409 goto out;
2410
2411 /*
2412 * After a write we want buffered reads to be sure to go to disk to get
2413 * the new data. We invalidate clean cached page from the region we're
2414 * about to write. We do this *before* the write so that we can return
2415 * without clobbering -EIOCBQUEUED from ->direct_IO().
2416 */
2417 if (mapping->nrpages) {
2418 written = invalidate_inode_pages2_range(mapping,
2419 pos >> PAGE_CACHE_SHIFT, end);
2420 /*
2421 * If a page can not be invalidated, return 0 to fall back
2422 * to buffered write.
2423 */
2424 if (written) {
2425 if (written == -EBUSY)
2426 return 0;
2427 goto out;
2428 }
2429 }
2430
2431 written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2432
2433 /*
2434 * Finally, try again to invalidate clean pages which might have been
2435 * cached by non-direct readahead, or faulted in by get_user_pages()
2436 * if the source of the write was an mmap'ed region of the file
2437 * we're writing. Either one is a pretty crazy thing to do,
2438 * so we don't support it 100%. If this invalidation
2439 * fails, tough, the write still worked...
2440 */
2441 if (mapping->nrpages) {
2442 invalidate_inode_pages2_range(mapping,
2443 pos >> PAGE_CACHE_SHIFT, end);
2444 }
2445
2446 if (written > 0) {
2447 pos += written;
2448 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2449 i_size_write(inode, pos);
2450 mark_inode_dirty(inode);
2451 }
2452 iocb->ki_pos = pos;
2453 }
2454out:
2455 return written;
2456}
2457EXPORT_SYMBOL(generic_file_direct_write);
2458
2459/*
2460 * Find or create a page at the given pagecache position. Return the locked
2461 * page. This function is specifically for buffered writes.
2462 */
2463struct page *grab_cache_page_write_begin(struct address_space *mapping,
2464 pgoff_t index, unsigned flags)
2465{
2466 int status;
2467 gfp_t gfp_mask;
2468 struct page *page;
2469 gfp_t gfp_notmask = 0;
2470
2471 gfp_mask = mapping_gfp_mask(mapping);
2472 if (mapping_cap_account_dirty(mapping))
2473 gfp_mask |= __GFP_WRITE;
2474 if (flags & AOP_FLAG_NOFS)
2475 gfp_notmask = __GFP_FS;
2476repeat:
2477 page = find_lock_page(mapping, index);
2478 if (page)
2479 goto found;
2480
2481 page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2482 if (!page)
2483 return NULL;
2484 status = add_to_page_cache_lru(page, mapping, index,
2485 GFP_KERNEL & ~gfp_notmask);
2486 if (unlikely(status)) {
2487 page_cache_release(page);
2488 if (status == -EEXIST)
2489 goto repeat;
2490 return NULL;
2491 }
2492found:
2493 wait_for_stable_page(page);
2494 return page;
2495}
2496EXPORT_SYMBOL(grab_cache_page_write_begin);
2497
2498ssize_t generic_perform_write(struct file *file,
2499 struct iov_iter *i, loff_t pos)
2500{
2501 struct address_space *mapping = file->f_mapping;
2502 const struct address_space_operations *a_ops = mapping->a_ops;
2503 long status = 0;
2504 ssize_t written = 0;
2505 unsigned int flags = 0;
2506
2507 /*
2508 * Copies from kernel address space cannot fail (NFSD is a big user).
2509 */
2510 if (segment_eq(get_fs(), KERNEL_DS))
2511 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2512
2513 do {
2514 struct page *page;
2515 unsigned long offset; /* Offset into pagecache page */
2516 unsigned long bytes; /* Bytes to write to page */
2517 size_t copied; /* Bytes copied from user */
2518 void *fsdata;
2519
2520 offset = (pos & (PAGE_CACHE_SIZE - 1));
2521 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2522 iov_iter_count(i));
2523
2524again:
2525 /*
2526 * Bring in the user page that we will copy from _first_.
2527 * Otherwise there's a nasty deadlock on copying from the
2528 * same page as we're writing to, without it being marked
2529 * up-to-date.
2530 *
2531 * Not only is this an optimisation, but it is also required
2532 * to check that the address is actually valid, when atomic
2533 * usercopies are used, below.
2534 */
2535 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2536 status = -EFAULT;
2537 break;
2538 }
2539
2540 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2541 &page, &fsdata);
2542 if (unlikely(status))
2543 break;
2544
2545 if (mapping_writably_mapped(mapping))
2546 flush_dcache_page(page);
2547
2548 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2549 flush_dcache_page(page);
2550
2551 mark_page_accessed(page);
2552 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2553 page, fsdata);
2554 if (unlikely(status < 0))
2555 break;
2556 copied = status;
2557
2558 cond_resched();
2559
2560 iov_iter_advance(i, copied);
2561 if (unlikely(copied == 0)) {
2562 /*
2563 * If we were unable to copy any data at all, we must
2564 * fall back to a single segment length write.
2565 *
2566 * If we didn't fallback here, we could livelock
2567 * because not all segments in the iov can be copied at
2568 * once without a pagefault.
2569 */
2570 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2571 iov_iter_single_seg_count(i));
2572 goto again;
2573 }
2574 pos += copied;
2575 written += copied;
2576
2577 balance_dirty_pages_ratelimited(mapping);
2578 if (fatal_signal_pending(current)) {
2579 status = -EINTR;
2580 break;
2581 }
2582 } while (iov_iter_count(i));
2583
2584 return written ? written : status;
2585}
2586EXPORT_SYMBOL(generic_perform_write);
2587
2588/**
2589 * __generic_file_aio_write - write data to a file
2590 * @iocb: IO state structure (file, offset, etc.)
2591 * @iov: vector with data to write
2592 * @nr_segs: number of segments in the vector
2593 *
2594 * This function does all the work needed for actually writing data to a
2595 * file. It does all basic checks, removes SUID from the file, updates
2596 * modification times and calls proper subroutines depending on whether we
2597 * do direct IO or a standard buffered write.
2598 *
2599 * It expects i_mutex to be grabbed unless we work on a block device or similar
2600 * object which does not need locking at all.
2601 *
2602 * This function does *not* take care of syncing data in case of O_SYNC write.
2603 * A caller has to handle it. This is mainly due to the fact that we want to
2604 * avoid syncing under i_mutex.
2605 */
2606ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2607 unsigned long nr_segs)
2608{
2609 struct file *file = iocb->ki_filp;
2610 struct address_space * mapping = file->f_mapping;
2611 size_t ocount; /* original count */
2612 size_t count; /* after file limit checks */
2613 struct inode *inode = mapping->host;
2614 loff_t pos = iocb->ki_pos;
2615 ssize_t written = 0;
2616 ssize_t err;
2617 ssize_t status;
2618 struct iov_iter from;
2619
2620 ocount = 0;
2621 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2622 if (err)
2623 return err;
2624
2625 count = ocount;
2626
2627 /* We can write back this queue in page reclaim */
2628 current->backing_dev_info = mapping->backing_dev_info;
2629 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2630 if (err)
2631 goto out;
2632
2633 if (count == 0)
2634 goto out;
2635
2636 err = file_remove_suid(file);
2637 if (err)
2638 goto out;
2639
2640 err = file_update_time(file);
2641 if (err)
2642 goto out;
2643
2644 iov_iter_init(&from, iov, nr_segs, count, 0);
2645
2646 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2647 if (unlikely(file->f_flags & O_DIRECT)) {
2648 loff_t endbyte;
2649
2650 written = generic_file_direct_write(iocb, iov, &from.nr_segs, pos,
2651 count, ocount);
2652 if (written < 0 || written == count)
2653 goto out;
2654 iov_iter_advance(&from, written);
2655
2656 /*
2657 * direct-io write to a hole: fall through to buffered I/O
2658 * for completing the rest of the request.
2659 */
2660 pos += written;
2661 count -= written;
2662
2663 status = generic_perform_write(file, &from, pos);
2664 /*
2665 * If generic_perform_write() returned a synchronous error
2666 * then we want to return the number of bytes which were
2667 * direct-written, or the error code if that was zero. Note
2668 * that this differs from normal direct-io semantics, which
2669 * will return -EFOO even if some bytes were written.
2670 */
2671 if (unlikely(status < 0) && !written) {
2672 err = status;
2673 goto out;
2674 }
2675 iocb->ki_pos = pos + status;
2676 /*
2677 * We need to ensure that the page cache pages are written to
2678 * disk and invalidated to preserve the expected O_DIRECT
2679 * semantics.
2680 */
2681 endbyte = pos + status - 1;
2682 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2683 if (err == 0) {
2684 written += status;
2685 invalidate_mapping_pages(mapping,
2686 pos >> PAGE_CACHE_SHIFT,
2687 endbyte >> PAGE_CACHE_SHIFT);
2688 } else {
2689 /*
2690 * We don't know how much we wrote, so just return
2691 * the number of bytes which were direct-written
2692 */
2693 }
2694 } else {
2695 written = generic_perform_write(file, &from, pos);
2696 if (likely(written >= 0))
2697 iocb->ki_pos = pos + written;
2698 }
2699out:
2700 current->backing_dev_info = NULL;
2701 return written ? written : err;
2702}
2703EXPORT_SYMBOL(__generic_file_aio_write);
2704
2705/**
2706 * generic_file_aio_write - write data to a file
2707 * @iocb: IO state structure
2708 * @iov: vector with data to write
2709 * @nr_segs: number of segments in the vector
2710 * @pos: position in file where to write
2711 *
2712 * This is a wrapper around __generic_file_aio_write() to be used by most
2713 * filesystems. It takes care of syncing the file in case of O_SYNC file
2714 * and acquires i_mutex as needed.
2715 */
2716ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2717 unsigned long nr_segs, loff_t pos)
2718{
2719 struct file *file = iocb->ki_filp;
2720 struct inode *inode = file->f_mapping->host;
2721 ssize_t ret;
2722
2723 BUG_ON(iocb->ki_pos != pos);
2724
2725 mutex_lock(&inode->i_mutex);
2726 ret = __generic_file_aio_write(iocb, iov, nr_segs);
2727 mutex_unlock(&inode->i_mutex);
2728
2729 if (ret > 0) {
2730 ssize_t err;
2731
2732 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
2733 if (err < 0)
2734 ret = err;
2735 }
2736 return ret;
2737}
2738EXPORT_SYMBOL(generic_file_aio_write);
2739
2740/**
2741 * try_to_release_page() - release old fs-specific metadata on a page
2742 *
2743 * @page: the page which the kernel is trying to free
2744 * @gfp_mask: memory allocation flags (and I/O mode)
2745 *
2746 * The address_space is to try to release any data against the page
2747 * (presumably at page->private). If the release was successful, return `1'.
2748 * Otherwise return zero.
2749 *
2750 * This may also be called if PG_fscache is set on a page, indicating that the
2751 * page is known to the local caching routines.
2752 *
2753 * The @gfp_mask argument specifies whether I/O may be performed to release
2754 * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2755 *
2756 */
2757int try_to_release_page(struct page *page, gfp_t gfp_mask)
2758{
2759 struct address_space * const mapping = page->mapping;
2760
2761 BUG_ON(!PageLocked(page));
2762 if (PageWriteback(page))
2763 return 0;
2764
2765 if (mapping && mapping->a_ops->releasepage)
2766 return mapping->a_ops->releasepage(page, gfp_mask);
2767 return try_to_free_buffers(page);
2768}
2769
2770EXPORT_SYMBOL(try_to_release_page);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/filemap.c
4 *
5 * Copyright (C) 1994-1999 Linus Torvalds
6 */
7
8/*
9 * This file handles the generic file mmap semantics used by
10 * most "normal" filesystems (but you don't /have/ to use this:
11 * the NFS filesystem used to do this differently, for example)
12 */
13#include <linux/export.h>
14#include <linux/compiler.h>
15#include <linux/dax.h>
16#include <linux/fs.h>
17#include <linux/sched/signal.h>
18#include <linux/uaccess.h>
19#include <linux/capability.h>
20#include <linux/kernel_stat.h>
21#include <linux/gfp.h>
22#include <linux/mm.h>
23#include <linux/swap.h>
24#include <linux/swapops.h>
25#include <linux/syscalls.h>
26#include <linux/mman.h>
27#include <linux/pagemap.h>
28#include <linux/file.h>
29#include <linux/uio.h>
30#include <linux/error-injection.h>
31#include <linux/hash.h>
32#include <linux/writeback.h>
33#include <linux/backing-dev.h>
34#include <linux/pagevec.h>
35#include <linux/security.h>
36#include <linux/cpuset.h>
37#include <linux/hugetlb.h>
38#include <linux/memcontrol.h>
39#include <linux/shmem_fs.h>
40#include <linux/rmap.h>
41#include <linux/delayacct.h>
42#include <linux/psi.h>
43#include <linux/ramfs.h>
44#include <linux/page_idle.h>
45#include <linux/migrate.h>
46#include <linux/pipe_fs_i.h>
47#include <linux/splice.h>
48#include <linux/rcupdate_wait.h>
49#include <asm/pgalloc.h>
50#include <asm/tlbflush.h>
51#include "internal.h"
52
53#define CREATE_TRACE_POINTS
54#include <trace/events/filemap.h>
55
56/*
57 * FIXME: remove all knowledge of the buffer layer from the core VM
58 */
59#include <linux/buffer_head.h> /* for try_to_free_buffers */
60
61#include <asm/mman.h>
62
63#include "swap.h"
64
65/*
66 * Shared mappings implemented 30.11.1994. It's not fully working yet,
67 * though.
68 *
69 * Shared mappings now work. 15.8.1995 Bruno.
70 *
71 * finished 'unifying' the page and buffer cache and SMP-threaded the
72 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
73 *
74 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
75 */
76
77/*
78 * Lock ordering:
79 *
80 * ->i_mmap_rwsem (truncate_pagecache)
81 * ->private_lock (__free_pte->block_dirty_folio)
82 * ->swap_lock (exclusive_swap_page, others)
83 * ->i_pages lock
84 *
85 * ->i_rwsem
86 * ->invalidate_lock (acquired by fs in truncate path)
87 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
88 *
89 * ->mmap_lock
90 * ->i_mmap_rwsem
91 * ->page_table_lock or pte_lock (various, mainly in memory.c)
92 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
93 *
94 * ->mmap_lock
95 * ->invalidate_lock (filemap_fault)
96 * ->lock_page (filemap_fault, access_process_vm)
97 *
98 * ->i_rwsem (generic_perform_write)
99 * ->mmap_lock (fault_in_readable->do_page_fault)
100 *
101 * bdi->wb.list_lock
102 * sb_lock (fs/fs-writeback.c)
103 * ->i_pages lock (__sync_single_inode)
104 *
105 * ->i_mmap_rwsem
106 * ->anon_vma.lock (vma_merge)
107 *
108 * ->anon_vma.lock
109 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
110 *
111 * ->page_table_lock or pte_lock
112 * ->swap_lock (try_to_unmap_one)
113 * ->private_lock (try_to_unmap_one)
114 * ->i_pages lock (try_to_unmap_one)
115 * ->lruvec->lru_lock (follow_page->mark_page_accessed)
116 * ->lruvec->lru_lock (check_pte_range->isolate_lru_page)
117 * ->private_lock (folio_remove_rmap_pte->set_page_dirty)
118 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty)
119 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty)
120 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty)
121 * ->memcg->move_lock (folio_remove_rmap_pte->folio_memcg_lock)
122 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
123 * ->inode->i_lock (zap_pte_range->set_page_dirty)
124 * ->private_lock (zap_pte_range->block_dirty_folio)
125 */
126
127static void mapping_set_update(struct xa_state *xas,
128 struct address_space *mapping)
129{
130 if (dax_mapping(mapping) || shmem_mapping(mapping))
131 return;
132 xas_set_update(xas, workingset_update_node);
133 xas_set_lru(xas, &shadow_nodes);
134}
135
136static void page_cache_delete(struct address_space *mapping,
137 struct folio *folio, void *shadow)
138{
139 XA_STATE(xas, &mapping->i_pages, folio->index);
140 long nr = 1;
141
142 mapping_set_update(&xas, mapping);
143
144 xas_set_order(&xas, folio->index, folio_order(folio));
145 nr = folio_nr_pages(folio);
146
147 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
148
149 xas_store(&xas, shadow);
150 xas_init_marks(&xas);
151
152 folio->mapping = NULL;
153 /* Leave page->index set: truncation lookup relies upon it */
154 mapping->nrpages -= nr;
155}
156
157static void filemap_unaccount_folio(struct address_space *mapping,
158 struct folio *folio)
159{
160 long nr;
161
162 VM_BUG_ON_FOLIO(folio_mapped(folio), folio);
163 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) {
164 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
165 current->comm, folio_pfn(folio));
166 dump_page(&folio->page, "still mapped when deleted");
167 dump_stack();
168 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
169
170 if (mapping_exiting(mapping) && !folio_test_large(folio)) {
171 int mapcount = page_mapcount(&folio->page);
172
173 if (folio_ref_count(folio) >= mapcount + 2) {
174 /*
175 * All vmas have already been torn down, so it's
176 * a good bet that actually the page is unmapped
177 * and we'd rather not leak it: if we're wrong,
178 * another bad page check should catch it later.
179 */
180 page_mapcount_reset(&folio->page);
181 folio_ref_sub(folio, mapcount);
182 }
183 }
184 }
185
186 /* hugetlb folios do not participate in page cache accounting. */
187 if (folio_test_hugetlb(folio))
188 return;
189
190 nr = folio_nr_pages(folio);
191
192 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
193 if (folio_test_swapbacked(folio)) {
194 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
195 if (folio_test_pmd_mappable(folio))
196 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr);
197 } else if (folio_test_pmd_mappable(folio)) {
198 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr);
199 filemap_nr_thps_dec(mapping);
200 }
201
202 /*
203 * At this point folio must be either written or cleaned by
204 * truncate. Dirty folio here signals a bug and loss of
205 * unwritten data - on ordinary filesystems.
206 *
207 * But it's harmless on in-memory filesystems like tmpfs; and can
208 * occur when a driver which did get_user_pages() sets page dirty
209 * before putting it, while the inode is being finally evicted.
210 *
211 * Below fixes dirty accounting after removing the folio entirely
212 * but leaves the dirty flag set: it has no effect for truncated
213 * folio and anyway will be cleared before returning folio to
214 * buddy allocator.
215 */
216 if (WARN_ON_ONCE(folio_test_dirty(folio) &&
217 mapping_can_writeback(mapping)))
218 folio_account_cleaned(folio, inode_to_wb(mapping->host));
219}
220
221/*
222 * Delete a page from the page cache and free it. Caller has to make
223 * sure the page is locked and that nobody else uses it - or that usage
224 * is safe. The caller must hold the i_pages lock.
225 */
226void __filemap_remove_folio(struct folio *folio, void *shadow)
227{
228 struct address_space *mapping = folio->mapping;
229
230 trace_mm_filemap_delete_from_page_cache(folio);
231 filemap_unaccount_folio(mapping, folio);
232 page_cache_delete(mapping, folio, shadow);
233}
234
235void filemap_free_folio(struct address_space *mapping, struct folio *folio)
236{
237 void (*free_folio)(struct folio *);
238 int refs = 1;
239
240 free_folio = mapping->a_ops->free_folio;
241 if (free_folio)
242 free_folio(folio);
243
244 if (folio_test_large(folio))
245 refs = folio_nr_pages(folio);
246 folio_put_refs(folio, refs);
247}
248
249/**
250 * filemap_remove_folio - Remove folio from page cache.
251 * @folio: The folio.
252 *
253 * This must be called only on folios that are locked and have been
254 * verified to be in the page cache. It will never put the folio into
255 * the free list because the caller has a reference on the page.
256 */
257void filemap_remove_folio(struct folio *folio)
258{
259 struct address_space *mapping = folio->mapping;
260
261 BUG_ON(!folio_test_locked(folio));
262 spin_lock(&mapping->host->i_lock);
263 xa_lock_irq(&mapping->i_pages);
264 __filemap_remove_folio(folio, NULL);
265 xa_unlock_irq(&mapping->i_pages);
266 if (mapping_shrinkable(mapping))
267 inode_add_lru(mapping->host);
268 spin_unlock(&mapping->host->i_lock);
269
270 filemap_free_folio(mapping, folio);
271}
272
273/*
274 * page_cache_delete_batch - delete several folios from page cache
275 * @mapping: the mapping to which folios belong
276 * @fbatch: batch of folios to delete
277 *
278 * The function walks over mapping->i_pages and removes folios passed in
279 * @fbatch from the mapping. The function expects @fbatch to be sorted
280 * by page index and is optimised for it to be dense.
281 * It tolerates holes in @fbatch (mapping entries at those indices are not
282 * modified).
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 folio_batch *fbatch)
288{
289 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index);
290 long total_pages = 0;
291 int i = 0;
292 struct folio *folio;
293
294 mapping_set_update(&xas, mapping);
295 xas_for_each(&xas, folio, ULONG_MAX) {
296 if (i >= folio_batch_count(fbatch))
297 break;
298
299 /* A swap/dax/shadow entry got inserted? Skip it. */
300 if (xa_is_value(folio))
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 (folio != fbatch->folios[i]) {
310 VM_BUG_ON_FOLIO(folio->index >
311 fbatch->folios[i]->index, folio);
312 continue;
313 }
314
315 WARN_ON_ONCE(!folio_test_locked(folio));
316
317 folio->mapping = NULL;
318 /* Leave folio->index set: truncation lookup relies on it */
319
320 i++;
321 xas_store(&xas, NULL);
322 total_pages += folio_nr_pages(folio);
323 }
324 mapping->nrpages -= total_pages;
325}
326
327void delete_from_page_cache_batch(struct address_space *mapping,
328 struct folio_batch *fbatch)
329{
330 int i;
331
332 if (!folio_batch_count(fbatch))
333 return;
334
335 spin_lock(&mapping->host->i_lock);
336 xa_lock_irq(&mapping->i_pages);
337 for (i = 0; i < folio_batch_count(fbatch); i++) {
338 struct folio *folio = fbatch->folios[i];
339
340 trace_mm_filemap_delete_from_page_cache(folio);
341 filemap_unaccount_folio(mapping, folio);
342 }
343 page_cache_delete_batch(mapping, fbatch);
344 xa_unlock_irq(&mapping->i_pages);
345 if (mapping_shrinkable(mapping))
346 inode_add_lru(mapping->host);
347 spin_unlock(&mapping->host->i_lock);
348
349 for (i = 0; i < folio_batch_count(fbatch); i++)
350 filemap_free_folio(mapping, fbatch->folios[i]);
351}
352
353int filemap_check_errors(struct address_space *mapping)
354{
355 int ret = 0;
356 /* Check for outstanding write errors */
357 if (test_bit(AS_ENOSPC, &mapping->flags) &&
358 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
359 ret = -ENOSPC;
360 if (test_bit(AS_EIO, &mapping->flags) &&
361 test_and_clear_bit(AS_EIO, &mapping->flags))
362 ret = -EIO;
363 return ret;
364}
365EXPORT_SYMBOL(filemap_check_errors);
366
367static int filemap_check_and_keep_errors(struct address_space *mapping)
368{
369 /* Check for outstanding write errors */
370 if (test_bit(AS_EIO, &mapping->flags))
371 return -EIO;
372 if (test_bit(AS_ENOSPC, &mapping->flags))
373 return -ENOSPC;
374 return 0;
375}
376
377/**
378 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range
379 * @mapping: address space structure to write
380 * @wbc: the writeback_control controlling the writeout
381 *
382 * Call writepages on the mapping using the provided wbc to control the
383 * writeout.
384 *
385 * Return: %0 on success, negative error code otherwise.
386 */
387int filemap_fdatawrite_wbc(struct address_space *mapping,
388 struct writeback_control *wbc)
389{
390 int ret;
391
392 if (!mapping_can_writeback(mapping) ||
393 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
394 return 0;
395
396 wbc_attach_fdatawrite_inode(wbc, mapping->host);
397 ret = do_writepages(mapping, wbc);
398 wbc_detach_inode(wbc);
399 return ret;
400}
401EXPORT_SYMBOL(filemap_fdatawrite_wbc);
402
403/**
404 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
405 * @mapping: address space structure to write
406 * @start: offset in bytes where the range starts
407 * @end: offset in bytes where the range ends (inclusive)
408 * @sync_mode: enable synchronous operation
409 *
410 * Start writeback against all of a mapping's dirty pages that lie
411 * within the byte offsets <start, end> inclusive.
412 *
413 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
414 * opposed to a regular memory cleansing writeback. The difference between
415 * these two operations is that if a dirty page/buffer is encountered, it must
416 * be waited upon, and not just skipped over.
417 *
418 * Return: %0 on success, negative error code otherwise.
419 */
420int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
421 loff_t end, int sync_mode)
422{
423 struct writeback_control wbc = {
424 .sync_mode = sync_mode,
425 .nr_to_write = LONG_MAX,
426 .range_start = start,
427 .range_end = end,
428 };
429
430 return filemap_fdatawrite_wbc(mapping, &wbc);
431}
432
433static inline int __filemap_fdatawrite(struct address_space *mapping,
434 int sync_mode)
435{
436 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
437}
438
439int filemap_fdatawrite(struct address_space *mapping)
440{
441 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
442}
443EXPORT_SYMBOL(filemap_fdatawrite);
444
445int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
446 loff_t end)
447{
448 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
449}
450EXPORT_SYMBOL(filemap_fdatawrite_range);
451
452/**
453 * filemap_flush - mostly a non-blocking flush
454 * @mapping: target address_space
455 *
456 * This is a mostly non-blocking flush. Not suitable for data-integrity
457 * purposes - I/O may not be started against all dirty pages.
458 *
459 * Return: %0 on success, negative error code otherwise.
460 */
461int filemap_flush(struct address_space *mapping)
462{
463 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
464}
465EXPORT_SYMBOL(filemap_flush);
466
467/**
468 * filemap_range_has_page - check if a page exists in range.
469 * @mapping: address space within which to check
470 * @start_byte: offset in bytes where the range starts
471 * @end_byte: offset in bytes where the range ends (inclusive)
472 *
473 * Find at least one page in the range supplied, usually used to check if
474 * direct writing in this range will trigger a writeback.
475 *
476 * Return: %true if at least one page exists in the specified range,
477 * %false otherwise.
478 */
479bool filemap_range_has_page(struct address_space *mapping,
480 loff_t start_byte, loff_t end_byte)
481{
482 struct folio *folio;
483 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
484 pgoff_t max = end_byte >> PAGE_SHIFT;
485
486 if (end_byte < start_byte)
487 return false;
488
489 rcu_read_lock();
490 for (;;) {
491 folio = xas_find(&xas, max);
492 if (xas_retry(&xas, folio))
493 continue;
494 /* Shadow entries don't count */
495 if (xa_is_value(folio))
496 continue;
497 /*
498 * We don't need to try to pin this page; we're about to
499 * release the RCU lock anyway. It is enough to know that
500 * there was a page here recently.
501 */
502 break;
503 }
504 rcu_read_unlock();
505
506 return folio != NULL;
507}
508EXPORT_SYMBOL(filemap_range_has_page);
509
510static void __filemap_fdatawait_range(struct address_space *mapping,
511 loff_t start_byte, loff_t end_byte)
512{
513 pgoff_t index = start_byte >> PAGE_SHIFT;
514 pgoff_t end = end_byte >> PAGE_SHIFT;
515 struct folio_batch fbatch;
516 unsigned nr_folios;
517
518 folio_batch_init(&fbatch);
519
520 while (index <= end) {
521 unsigned i;
522
523 nr_folios = filemap_get_folios_tag(mapping, &index, end,
524 PAGECACHE_TAG_WRITEBACK, &fbatch);
525
526 if (!nr_folios)
527 break;
528
529 for (i = 0; i < nr_folios; i++) {
530 struct folio *folio = fbatch.folios[i];
531
532 folio_wait_writeback(folio);
533 folio_clear_error(folio);
534 }
535 folio_batch_release(&fbatch);
536 cond_resched();
537 }
538}
539
540/**
541 * filemap_fdatawait_range - wait for writeback to complete
542 * @mapping: address space structure to wait for
543 * @start_byte: offset in bytes where the range starts
544 * @end_byte: offset in bytes where the range ends (inclusive)
545 *
546 * Walk the list of under-writeback pages of the given address space
547 * in the given range and wait for all of them. Check error status of
548 * the address space and return it.
549 *
550 * Since the error status of the address space is cleared by this function,
551 * callers are responsible for checking the return value and handling and/or
552 * reporting the error.
553 *
554 * Return: error status of the address space.
555 */
556int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
557 loff_t end_byte)
558{
559 __filemap_fdatawait_range(mapping, start_byte, end_byte);
560 return filemap_check_errors(mapping);
561}
562EXPORT_SYMBOL(filemap_fdatawait_range);
563
564/**
565 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
566 * @mapping: address space structure to wait for
567 * @start_byte: offset in bytes where the range starts
568 * @end_byte: offset in bytes where the range ends (inclusive)
569 *
570 * Walk the list of under-writeback pages of the given address space in the
571 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
572 * this function does not clear error status of the address space.
573 *
574 * Use this function if callers don't handle errors themselves. Expected
575 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
576 * fsfreeze(8)
577 */
578int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
579 loff_t start_byte, loff_t end_byte)
580{
581 __filemap_fdatawait_range(mapping, start_byte, end_byte);
582 return filemap_check_and_keep_errors(mapping);
583}
584EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
585
586/**
587 * file_fdatawait_range - wait for writeback to complete
588 * @file: file pointing to address space structure to wait for
589 * @start_byte: offset in bytes where the range starts
590 * @end_byte: offset in bytes where the range ends (inclusive)
591 *
592 * Walk the list of under-writeback pages of the address space that file
593 * refers to, in the given range and wait for all of them. Check error
594 * status of the address space vs. the file->f_wb_err cursor and return it.
595 *
596 * Since the error status of the file is advanced by this function,
597 * callers are responsible for checking the return value and handling and/or
598 * reporting the error.
599 *
600 * Return: error status of the address space vs. the file->f_wb_err cursor.
601 */
602int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
603{
604 struct address_space *mapping = file->f_mapping;
605
606 __filemap_fdatawait_range(mapping, start_byte, end_byte);
607 return file_check_and_advance_wb_err(file);
608}
609EXPORT_SYMBOL(file_fdatawait_range);
610
611/**
612 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
613 * @mapping: address space structure to wait for
614 *
615 * Walk the list of under-writeback pages of the given address space
616 * and wait for all of them. Unlike filemap_fdatawait(), this function
617 * does not clear error status of the address space.
618 *
619 * Use this function if callers don't handle errors themselves. Expected
620 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
621 * fsfreeze(8)
622 *
623 * Return: error status of the address space.
624 */
625int filemap_fdatawait_keep_errors(struct address_space *mapping)
626{
627 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
628 return filemap_check_and_keep_errors(mapping);
629}
630EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
631
632/* Returns true if writeback might be needed or already in progress. */
633static bool mapping_needs_writeback(struct address_space *mapping)
634{
635 return mapping->nrpages;
636}
637
638bool filemap_range_has_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 folio *folio;
644
645 if (end_byte < start_byte)
646 return false;
647
648 rcu_read_lock();
649 xas_for_each(&xas, folio, max) {
650 if (xas_retry(&xas, folio))
651 continue;
652 if (xa_is_value(folio))
653 continue;
654 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
655 folio_test_writeback(folio))
656 break;
657 }
658 rcu_read_unlock();
659 return folio != NULL;
660}
661EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
662
663/**
664 * filemap_write_and_wait_range - write out & wait on a file range
665 * @mapping: the address_space for the pages
666 * @lstart: offset in bytes where the range starts
667 * @lend: offset in bytes where the range ends (inclusive)
668 *
669 * Write out and wait upon file offsets lstart->lend, inclusive.
670 *
671 * Note that @lend is inclusive (describes the last byte to be written) so
672 * that this function can be used to write to the very end-of-file (end = -1).
673 *
674 * Return: error status of the address space.
675 */
676int filemap_write_and_wait_range(struct address_space *mapping,
677 loff_t lstart, loff_t lend)
678{
679 int err = 0, err2;
680
681 if (lend < lstart)
682 return 0;
683
684 if (mapping_needs_writeback(mapping)) {
685 err = __filemap_fdatawrite_range(mapping, lstart, lend,
686 WB_SYNC_ALL);
687 /*
688 * Even if the above returned error, the pages may be
689 * written partially (e.g. -ENOSPC), so we wait for it.
690 * But the -EIO is special case, it may indicate the worst
691 * thing (e.g. bug) happened, so we avoid waiting for it.
692 */
693 if (err != -EIO)
694 __filemap_fdatawait_range(mapping, lstart, lend);
695 }
696 err2 = filemap_check_errors(mapping);
697 if (!err)
698 err = err2;
699 return err;
700}
701EXPORT_SYMBOL(filemap_write_and_wait_range);
702
703void __filemap_set_wb_err(struct address_space *mapping, int err)
704{
705 errseq_t eseq = errseq_set(&mapping->wb_err, err);
706
707 trace_filemap_set_wb_err(mapping, eseq);
708}
709EXPORT_SYMBOL(__filemap_set_wb_err);
710
711/**
712 * file_check_and_advance_wb_err - report wb error (if any) that was previously
713 * and advance wb_err to current one
714 * @file: struct file on which the error is being reported
715 *
716 * When userland calls fsync (or something like nfsd does the equivalent), we
717 * want to report any writeback errors that occurred since the last fsync (or
718 * since the file was opened if there haven't been any).
719 *
720 * Grab the wb_err from the mapping. If it matches what we have in the file,
721 * then just quickly return 0. The file is all caught up.
722 *
723 * If it doesn't match, then take the mapping value, set the "seen" flag in
724 * it and try to swap it into place. If it works, or another task beat us
725 * to it with the new value, then update the f_wb_err and return the error
726 * portion. The error at this point must be reported via proper channels
727 * (a'la fsync, or NFS COMMIT operation, etc.).
728 *
729 * While we handle mapping->wb_err with atomic operations, the f_wb_err
730 * value is protected by the f_lock since we must ensure that it reflects
731 * the latest value swapped in for this file descriptor.
732 *
733 * Return: %0 on success, negative error code otherwise.
734 */
735int file_check_and_advance_wb_err(struct file *file)
736{
737 int err = 0;
738 errseq_t old = READ_ONCE(file->f_wb_err);
739 struct address_space *mapping = file->f_mapping;
740
741 /* Locklessly handle the common case where nothing has changed */
742 if (errseq_check(&mapping->wb_err, old)) {
743 /* Something changed, must use slow path */
744 spin_lock(&file->f_lock);
745 old = file->f_wb_err;
746 err = errseq_check_and_advance(&mapping->wb_err,
747 &file->f_wb_err);
748 trace_file_check_and_advance_wb_err(file, old);
749 spin_unlock(&file->f_lock);
750 }
751
752 /*
753 * We're mostly using this function as a drop in replacement for
754 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
755 * that the legacy code would have had on these flags.
756 */
757 clear_bit(AS_EIO, &mapping->flags);
758 clear_bit(AS_ENOSPC, &mapping->flags);
759 return err;
760}
761EXPORT_SYMBOL(file_check_and_advance_wb_err);
762
763/**
764 * file_write_and_wait_range - write out & wait on a file range
765 * @file: file pointing to address_space with pages
766 * @lstart: offset in bytes where the range starts
767 * @lend: offset in bytes where the range ends (inclusive)
768 *
769 * Write out and wait upon file offsets lstart->lend, inclusive.
770 *
771 * Note that @lend is inclusive (describes the last byte to be written) so
772 * that this function can be used to write to the very end-of-file (end = -1).
773 *
774 * After writing out and waiting on the data, we check and advance the
775 * f_wb_err cursor to the latest value, and return any errors detected there.
776 *
777 * Return: %0 on success, negative error code otherwise.
778 */
779int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
780{
781 int err = 0, err2;
782 struct address_space *mapping = file->f_mapping;
783
784 if (lend < lstart)
785 return 0;
786
787 if (mapping_needs_writeback(mapping)) {
788 err = __filemap_fdatawrite_range(mapping, lstart, lend,
789 WB_SYNC_ALL);
790 /* See comment of filemap_write_and_wait() */
791 if (err != -EIO)
792 __filemap_fdatawait_range(mapping, lstart, lend);
793 }
794 err2 = file_check_and_advance_wb_err(file);
795 if (!err)
796 err = err2;
797 return err;
798}
799EXPORT_SYMBOL(file_write_and_wait_range);
800
801/**
802 * replace_page_cache_folio - replace a pagecache folio with a new one
803 * @old: folio to be replaced
804 * @new: folio to replace with
805 *
806 * This function replaces a folio in the pagecache with a new one. On
807 * success it acquires the pagecache reference for the new folio and
808 * drops it for the old folio. Both the old and new folios must be
809 * locked. This function does not add the new folio to the LRU, the
810 * caller must do that.
811 *
812 * The remove + add is atomic. This function cannot fail.
813 */
814void replace_page_cache_folio(struct folio *old, struct folio *new)
815{
816 struct address_space *mapping = old->mapping;
817 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
818 pgoff_t offset = old->index;
819 XA_STATE(xas, &mapping->i_pages, offset);
820
821 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
822 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
823 VM_BUG_ON_FOLIO(new->mapping, new);
824
825 folio_get(new);
826 new->mapping = mapping;
827 new->index = offset;
828
829 mem_cgroup_replace_folio(old, new);
830
831 xas_lock_irq(&xas);
832 xas_store(&xas, new);
833
834 old->mapping = NULL;
835 /* hugetlb pages do not participate in page cache accounting. */
836 if (!folio_test_hugetlb(old))
837 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
838 if (!folio_test_hugetlb(new))
839 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
840 if (folio_test_swapbacked(old))
841 __lruvec_stat_sub_folio(old, NR_SHMEM);
842 if (folio_test_swapbacked(new))
843 __lruvec_stat_add_folio(new, NR_SHMEM);
844 xas_unlock_irq(&xas);
845 if (free_folio)
846 free_folio(old);
847 folio_put(old);
848}
849EXPORT_SYMBOL_GPL(replace_page_cache_folio);
850
851noinline int __filemap_add_folio(struct address_space *mapping,
852 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
853{
854 XA_STATE(xas, &mapping->i_pages, index);
855 bool huge = folio_test_hugetlb(folio);
856 bool charged = false;
857 long nr = 1;
858
859 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
860 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
861 mapping_set_update(&xas, mapping);
862
863 if (!huge) {
864 int error = mem_cgroup_charge(folio, NULL, gfp);
865 if (error)
866 return error;
867 charged = true;
868 }
869
870 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
871 xas_set_order(&xas, index, folio_order(folio));
872 nr = folio_nr_pages(folio);
873
874 gfp &= GFP_RECLAIM_MASK;
875 folio_ref_add(folio, nr);
876 folio->mapping = mapping;
877 folio->index = xas.xa_index;
878
879 do {
880 unsigned int order = xa_get_order(xas.xa, xas.xa_index);
881 void *entry, *old = NULL;
882
883 if (order > folio_order(folio))
884 xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index),
885 order, gfp);
886 xas_lock_irq(&xas);
887 xas_for_each_conflict(&xas, entry) {
888 old = entry;
889 if (!xa_is_value(entry)) {
890 xas_set_err(&xas, -EEXIST);
891 goto unlock;
892 }
893 }
894
895 if (old) {
896 if (shadowp)
897 *shadowp = old;
898 /* entry may have been split before we acquired lock */
899 order = xa_get_order(xas.xa, xas.xa_index);
900 if (order > folio_order(folio)) {
901 /* How to handle large swap entries? */
902 BUG_ON(shmem_mapping(mapping));
903 xas_split(&xas, old, order);
904 xas_reset(&xas);
905 }
906 }
907
908 xas_store(&xas, folio);
909 if (xas_error(&xas))
910 goto unlock;
911
912 mapping->nrpages += nr;
913
914 /* hugetlb pages do not participate in page cache accounting */
915 if (!huge) {
916 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
917 if (folio_test_pmd_mappable(folio))
918 __lruvec_stat_mod_folio(folio,
919 NR_FILE_THPS, nr);
920 }
921unlock:
922 xas_unlock_irq(&xas);
923 } while (xas_nomem(&xas, gfp));
924
925 if (xas_error(&xas))
926 goto error;
927
928 trace_mm_filemap_add_to_page_cache(folio);
929 return 0;
930error:
931 if (charged)
932 mem_cgroup_uncharge(folio);
933 folio->mapping = NULL;
934 /* Leave page->index set: truncation relies upon it */
935 folio_put_refs(folio, nr);
936 return xas_error(&xas);
937}
938ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
939
940int filemap_add_folio(struct address_space *mapping, struct folio *folio,
941 pgoff_t index, gfp_t gfp)
942{
943 void *shadow = NULL;
944 int ret;
945
946 __folio_set_locked(folio);
947 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
948 if (unlikely(ret))
949 __folio_clear_locked(folio);
950 else {
951 /*
952 * The folio might have been evicted from cache only
953 * recently, in which case it should be activated like
954 * any other repeatedly accessed folio.
955 * The exception is folios getting rewritten; evicting other
956 * data from the working set, only to cache data that will
957 * get overwritten with something else, is a waste of memory.
958 */
959 WARN_ON_ONCE(folio_test_active(folio));
960 if (!(gfp & __GFP_WRITE) && shadow)
961 workingset_refault(folio, shadow);
962 folio_add_lru(folio);
963 }
964 return ret;
965}
966EXPORT_SYMBOL_GPL(filemap_add_folio);
967
968#ifdef CONFIG_NUMA
969struct folio *filemap_alloc_folio(gfp_t gfp, unsigned int order)
970{
971 int n;
972 struct folio *folio;
973
974 if (cpuset_do_page_mem_spread()) {
975 unsigned int cpuset_mems_cookie;
976 do {
977 cpuset_mems_cookie = read_mems_allowed_begin();
978 n = cpuset_mem_spread_node();
979 folio = __folio_alloc_node(gfp, order, n);
980 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
981
982 return folio;
983 }
984 return folio_alloc(gfp, order);
985}
986EXPORT_SYMBOL(filemap_alloc_folio);
987#endif
988
989/*
990 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
991 *
992 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
993 *
994 * @mapping1: the first mapping to lock
995 * @mapping2: the second mapping to lock
996 */
997void filemap_invalidate_lock_two(struct address_space *mapping1,
998 struct address_space *mapping2)
999{
1000 if (mapping1 > mapping2)
1001 swap(mapping1, mapping2);
1002 if (mapping1)
1003 down_write(&mapping1->invalidate_lock);
1004 if (mapping2 && mapping1 != mapping2)
1005 down_write_nested(&mapping2->invalidate_lock, 1);
1006}
1007EXPORT_SYMBOL(filemap_invalidate_lock_two);
1008
1009/*
1010 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
1011 *
1012 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1013 *
1014 * @mapping1: the first mapping to unlock
1015 * @mapping2: the second mapping to unlock
1016 */
1017void filemap_invalidate_unlock_two(struct address_space *mapping1,
1018 struct address_space *mapping2)
1019{
1020 if (mapping1)
1021 up_write(&mapping1->invalidate_lock);
1022 if (mapping2 && mapping1 != mapping2)
1023 up_write(&mapping2->invalidate_lock);
1024}
1025EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1026
1027/*
1028 * In order to wait for pages to become available there must be
1029 * waitqueues associated with pages. By using a hash table of
1030 * waitqueues where the bucket discipline is to maintain all
1031 * waiters on the same queue and wake all when any of the pages
1032 * become available, and for the woken contexts to check to be
1033 * sure the appropriate page became available, this saves space
1034 * at a cost of "thundering herd" phenomena during rare hash
1035 * collisions.
1036 */
1037#define PAGE_WAIT_TABLE_BITS 8
1038#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1039static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1040
1041static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1042{
1043 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1044}
1045
1046void __init pagecache_init(void)
1047{
1048 int i;
1049
1050 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1051 init_waitqueue_head(&folio_wait_table[i]);
1052
1053 page_writeback_init();
1054}
1055
1056/*
1057 * The page wait code treats the "wait->flags" somewhat unusually, because
1058 * we have multiple different kinds of waits, not just the usual "exclusive"
1059 * one.
1060 *
1061 * We have:
1062 *
1063 * (a) no special bits set:
1064 *
1065 * We're just waiting for the bit to be released, and when a waker
1066 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1067 * and remove it from the wait queue.
1068 *
1069 * Simple and straightforward.
1070 *
1071 * (b) WQ_FLAG_EXCLUSIVE:
1072 *
1073 * The waiter is waiting to get the lock, and only one waiter should
1074 * be woken up to avoid any thundering herd behavior. We'll set the
1075 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1076 *
1077 * This is the traditional exclusive wait.
1078 *
1079 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1080 *
1081 * The waiter is waiting to get the bit, and additionally wants the
1082 * lock to be transferred to it for fair lock behavior. If the lock
1083 * cannot be taken, we stop walking the wait queue without waking
1084 * the waiter.
1085 *
1086 * This is the "fair lock handoff" case, and in addition to setting
1087 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1088 * that it now has the lock.
1089 */
1090static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1091{
1092 unsigned int flags;
1093 struct wait_page_key *key = arg;
1094 struct wait_page_queue *wait_page
1095 = container_of(wait, struct wait_page_queue, wait);
1096
1097 if (!wake_page_match(wait_page, key))
1098 return 0;
1099
1100 /*
1101 * If it's a lock handoff wait, we get the bit for it, and
1102 * stop walking (and do not wake it up) if we can't.
1103 */
1104 flags = wait->flags;
1105 if (flags & WQ_FLAG_EXCLUSIVE) {
1106 if (test_bit(key->bit_nr, &key->folio->flags))
1107 return -1;
1108 if (flags & WQ_FLAG_CUSTOM) {
1109 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1110 return -1;
1111 flags |= WQ_FLAG_DONE;
1112 }
1113 }
1114
1115 /*
1116 * We are holding the wait-queue lock, but the waiter that
1117 * is waiting for this will be checking the flags without
1118 * any locking.
1119 *
1120 * So update the flags atomically, and wake up the waiter
1121 * afterwards to avoid any races. This store-release pairs
1122 * with the load-acquire in folio_wait_bit_common().
1123 */
1124 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1125 wake_up_state(wait->private, mode);
1126
1127 /*
1128 * Ok, we have successfully done what we're waiting for,
1129 * and we can unconditionally remove the wait entry.
1130 *
1131 * Note that this pairs with the "finish_wait()" in the
1132 * waiter, and has to be the absolute last thing we do.
1133 * After this list_del_init(&wait->entry) the wait entry
1134 * might be de-allocated and the process might even have
1135 * exited.
1136 */
1137 list_del_init_careful(&wait->entry);
1138 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1139}
1140
1141static void folio_wake_bit(struct folio *folio, int bit_nr)
1142{
1143 wait_queue_head_t *q = folio_waitqueue(folio);
1144 struct wait_page_key key;
1145 unsigned long flags;
1146
1147 key.folio = folio;
1148 key.bit_nr = bit_nr;
1149 key.page_match = 0;
1150
1151 spin_lock_irqsave(&q->lock, flags);
1152 __wake_up_locked_key(q, TASK_NORMAL, &key);
1153
1154 /*
1155 * It's possible to miss clearing waiters here, when we woke our page
1156 * waiters, but the hashed waitqueue has waiters for other pages on it.
1157 * That's okay, it's a rare case. The next waker will clear it.
1158 *
1159 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1160 * other), the flag may be cleared in the course of freeing the page;
1161 * but that is not required for correctness.
1162 */
1163 if (!waitqueue_active(q) || !key.page_match)
1164 folio_clear_waiters(folio);
1165
1166 spin_unlock_irqrestore(&q->lock, flags);
1167}
1168
1169/*
1170 * A choice of three behaviors for folio_wait_bit_common():
1171 */
1172enum behavior {
1173 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1174 * __folio_lock() waiting on then setting PG_locked.
1175 */
1176 SHARED, /* Hold ref to page and check the bit when woken, like
1177 * folio_wait_writeback() waiting on PG_writeback.
1178 */
1179 DROP, /* Drop ref to page before wait, no check when woken,
1180 * like folio_put_wait_locked() on PG_locked.
1181 */
1182};
1183
1184/*
1185 * Attempt to check (or get) the folio flag, and mark us done
1186 * if successful.
1187 */
1188static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1189 struct wait_queue_entry *wait)
1190{
1191 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1192 if (test_and_set_bit(bit_nr, &folio->flags))
1193 return false;
1194 } else if (test_bit(bit_nr, &folio->flags))
1195 return false;
1196
1197 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1198 return true;
1199}
1200
1201/* How many times do we accept lock stealing from under a waiter? */
1202int sysctl_page_lock_unfairness = 5;
1203
1204static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1205 int state, enum behavior behavior)
1206{
1207 wait_queue_head_t *q = folio_waitqueue(folio);
1208 int unfairness = sysctl_page_lock_unfairness;
1209 struct wait_page_queue wait_page;
1210 wait_queue_entry_t *wait = &wait_page.wait;
1211 bool thrashing = false;
1212 unsigned long pflags;
1213 bool in_thrashing;
1214
1215 if (bit_nr == PG_locked &&
1216 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1217 delayacct_thrashing_start(&in_thrashing);
1218 psi_memstall_enter(&pflags);
1219 thrashing = true;
1220 }
1221
1222 init_wait(wait);
1223 wait->func = wake_page_function;
1224 wait_page.folio = folio;
1225 wait_page.bit_nr = bit_nr;
1226
1227repeat:
1228 wait->flags = 0;
1229 if (behavior == EXCLUSIVE) {
1230 wait->flags = WQ_FLAG_EXCLUSIVE;
1231 if (--unfairness < 0)
1232 wait->flags |= WQ_FLAG_CUSTOM;
1233 }
1234
1235 /*
1236 * Do one last check whether we can get the
1237 * page bit synchronously.
1238 *
1239 * Do the folio_set_waiters() marking before that
1240 * to let any waker we _just_ missed know they
1241 * need to wake us up (otherwise they'll never
1242 * even go to the slow case that looks at the
1243 * page queue), and add ourselves to the wait
1244 * queue if we need to sleep.
1245 *
1246 * This part needs to be done under the queue
1247 * lock to avoid races.
1248 */
1249 spin_lock_irq(&q->lock);
1250 folio_set_waiters(folio);
1251 if (!folio_trylock_flag(folio, bit_nr, wait))
1252 __add_wait_queue_entry_tail(q, wait);
1253 spin_unlock_irq(&q->lock);
1254
1255 /*
1256 * From now on, all the logic will be based on
1257 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1258 * see whether the page bit testing has already
1259 * been done by the wake function.
1260 *
1261 * We can drop our reference to the folio.
1262 */
1263 if (behavior == DROP)
1264 folio_put(folio);
1265
1266 /*
1267 * Note that until the "finish_wait()", or until
1268 * we see the WQ_FLAG_WOKEN flag, we need to
1269 * be very careful with the 'wait->flags', because
1270 * we may race with a waker that sets them.
1271 */
1272 for (;;) {
1273 unsigned int flags;
1274
1275 set_current_state(state);
1276
1277 /* Loop until we've been woken or interrupted */
1278 flags = smp_load_acquire(&wait->flags);
1279 if (!(flags & WQ_FLAG_WOKEN)) {
1280 if (signal_pending_state(state, current))
1281 break;
1282
1283 io_schedule();
1284 continue;
1285 }
1286
1287 /* If we were non-exclusive, we're done */
1288 if (behavior != EXCLUSIVE)
1289 break;
1290
1291 /* If the waker got the lock for us, we're done */
1292 if (flags & WQ_FLAG_DONE)
1293 break;
1294
1295 /*
1296 * Otherwise, if we're getting the lock, we need to
1297 * try to get it ourselves.
1298 *
1299 * And if that fails, we'll have to retry this all.
1300 */
1301 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1302 goto repeat;
1303
1304 wait->flags |= WQ_FLAG_DONE;
1305 break;
1306 }
1307
1308 /*
1309 * If a signal happened, this 'finish_wait()' may remove the last
1310 * waiter from the wait-queues, but the folio waiters bit will remain
1311 * set. That's ok. The next wakeup will take care of it, and trying
1312 * to do it here would be difficult and prone to races.
1313 */
1314 finish_wait(q, wait);
1315
1316 if (thrashing) {
1317 delayacct_thrashing_end(&in_thrashing);
1318 psi_memstall_leave(&pflags);
1319 }
1320
1321 /*
1322 * NOTE! The wait->flags weren't stable until we've done the
1323 * 'finish_wait()', and we could have exited the loop above due
1324 * to a signal, and had a wakeup event happen after the signal
1325 * test but before the 'finish_wait()'.
1326 *
1327 * So only after the finish_wait() can we reliably determine
1328 * if we got woken up or not, so we can now figure out the final
1329 * return value based on that state without races.
1330 *
1331 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1332 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1333 */
1334 if (behavior == EXCLUSIVE)
1335 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1336
1337 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1338}
1339
1340#ifdef CONFIG_MIGRATION
1341/**
1342 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1343 * @entry: migration swap entry.
1344 * @ptl: already locked ptl. This function will drop the lock.
1345 *
1346 * Wait for a migration entry referencing the given page to be removed. This is
1347 * equivalent to put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE) except
1348 * this can be called without taking a reference on the page. Instead this
1349 * should be called while holding the ptl for the migration entry referencing
1350 * the page.
1351 *
1352 * Returns after unlocking the ptl.
1353 *
1354 * This follows the same logic as folio_wait_bit_common() so see the comments
1355 * there.
1356 */
1357void migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl)
1358 __releases(ptl)
1359{
1360 struct wait_page_queue wait_page;
1361 wait_queue_entry_t *wait = &wait_page.wait;
1362 bool thrashing = false;
1363 unsigned long pflags;
1364 bool in_thrashing;
1365 wait_queue_head_t *q;
1366 struct folio *folio = pfn_swap_entry_folio(entry);
1367
1368 q = folio_waitqueue(folio);
1369 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1370 delayacct_thrashing_start(&in_thrashing);
1371 psi_memstall_enter(&pflags);
1372 thrashing = true;
1373 }
1374
1375 init_wait(wait);
1376 wait->func = wake_page_function;
1377 wait_page.folio = folio;
1378 wait_page.bit_nr = PG_locked;
1379 wait->flags = 0;
1380
1381 spin_lock_irq(&q->lock);
1382 folio_set_waiters(folio);
1383 if (!folio_trylock_flag(folio, PG_locked, wait))
1384 __add_wait_queue_entry_tail(q, wait);
1385 spin_unlock_irq(&q->lock);
1386
1387 /*
1388 * If a migration entry exists for the page the migration path must hold
1389 * a valid reference to the page, and it must take the ptl to remove the
1390 * migration entry. So the page is valid until the ptl is dropped.
1391 */
1392 spin_unlock(ptl);
1393
1394 for (;;) {
1395 unsigned int flags;
1396
1397 set_current_state(TASK_UNINTERRUPTIBLE);
1398
1399 /* Loop until we've been woken or interrupted */
1400 flags = smp_load_acquire(&wait->flags);
1401 if (!(flags & WQ_FLAG_WOKEN)) {
1402 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1403 break;
1404
1405 io_schedule();
1406 continue;
1407 }
1408 break;
1409 }
1410
1411 finish_wait(q, wait);
1412
1413 if (thrashing) {
1414 delayacct_thrashing_end(&in_thrashing);
1415 psi_memstall_leave(&pflags);
1416 }
1417}
1418#endif
1419
1420void folio_wait_bit(struct folio *folio, int bit_nr)
1421{
1422 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1423}
1424EXPORT_SYMBOL(folio_wait_bit);
1425
1426int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1427{
1428 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1429}
1430EXPORT_SYMBOL(folio_wait_bit_killable);
1431
1432/**
1433 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1434 * @folio: The folio to wait for.
1435 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1436 *
1437 * The caller should hold a reference on @folio. They expect the page to
1438 * become unlocked relatively soon, but do not wish to hold up migration
1439 * (for example) by holding the reference while waiting for the folio to
1440 * come unlocked. After this function returns, the caller should not
1441 * dereference @folio.
1442 *
1443 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1444 */
1445static int folio_put_wait_locked(struct folio *folio, int state)
1446{
1447 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1448}
1449
1450/**
1451 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue
1452 * @folio: Folio defining the wait queue of interest
1453 * @waiter: Waiter to add to the queue
1454 *
1455 * Add an arbitrary @waiter to the wait queue for the nominated @folio.
1456 */
1457void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter)
1458{
1459 wait_queue_head_t *q = folio_waitqueue(folio);
1460 unsigned long flags;
1461
1462 spin_lock_irqsave(&q->lock, flags);
1463 __add_wait_queue_entry_tail(q, waiter);
1464 folio_set_waiters(folio);
1465 spin_unlock_irqrestore(&q->lock, flags);
1466}
1467EXPORT_SYMBOL_GPL(folio_add_wait_queue);
1468
1469/**
1470 * folio_unlock - Unlock a locked folio.
1471 * @folio: The folio.
1472 *
1473 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1474 *
1475 * Context: May be called from interrupt or process context. May not be
1476 * called from NMI context.
1477 */
1478void folio_unlock(struct folio *folio)
1479{
1480 /* Bit 7 allows x86 to check the byte's sign bit */
1481 BUILD_BUG_ON(PG_waiters != 7);
1482 BUILD_BUG_ON(PG_locked > 7);
1483 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1484 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked))
1485 folio_wake_bit(folio, PG_locked);
1486}
1487EXPORT_SYMBOL(folio_unlock);
1488
1489/**
1490 * folio_end_read - End read on a folio.
1491 * @folio: The folio.
1492 * @success: True if all reads completed successfully.
1493 *
1494 * When all reads against a folio have completed, filesystems should
1495 * call this function to let the pagecache know that no more reads
1496 * are outstanding. This will unlock the folio and wake up any thread
1497 * sleeping on the lock. The folio will also be marked uptodate if all
1498 * reads succeeded.
1499 *
1500 * Context: May be called from interrupt or process context. May not be
1501 * called from NMI context.
1502 */
1503void folio_end_read(struct folio *folio, bool success)
1504{
1505 unsigned long mask = 1 << PG_locked;
1506
1507 /* Must be in bottom byte for x86 to work */
1508 BUILD_BUG_ON(PG_uptodate > 7);
1509 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1510 VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
1511
1512 if (likely(success))
1513 mask |= 1 << PG_uptodate;
1514 if (folio_xor_flags_has_waiters(folio, mask))
1515 folio_wake_bit(folio, PG_locked);
1516}
1517EXPORT_SYMBOL(folio_end_read);
1518
1519/**
1520 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1521 * @folio: The folio.
1522 *
1523 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1524 * it. The folio reference held for PG_private_2 being set is released.
1525 *
1526 * This is, for example, used when a netfs folio is being written to a local
1527 * disk cache, thereby allowing writes to the cache for the same folio to be
1528 * serialised.
1529 */
1530void folio_end_private_2(struct folio *folio)
1531{
1532 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1533 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1534 folio_wake_bit(folio, PG_private_2);
1535 folio_put(folio);
1536}
1537EXPORT_SYMBOL(folio_end_private_2);
1538
1539/**
1540 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1541 * @folio: The folio to wait on.
1542 *
1543 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio.
1544 */
1545void folio_wait_private_2(struct folio *folio)
1546{
1547 while (folio_test_private_2(folio))
1548 folio_wait_bit(folio, PG_private_2);
1549}
1550EXPORT_SYMBOL(folio_wait_private_2);
1551
1552/**
1553 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1554 * @folio: The folio to wait on.
1555 *
1556 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio or until a
1557 * fatal signal is received by the calling task.
1558 *
1559 * Return:
1560 * - 0 if successful.
1561 * - -EINTR if a fatal signal was encountered.
1562 */
1563int folio_wait_private_2_killable(struct folio *folio)
1564{
1565 int ret = 0;
1566
1567 while (folio_test_private_2(folio)) {
1568 ret = folio_wait_bit_killable(folio, PG_private_2);
1569 if (ret < 0)
1570 break;
1571 }
1572
1573 return ret;
1574}
1575EXPORT_SYMBOL(folio_wait_private_2_killable);
1576
1577/**
1578 * folio_end_writeback - End writeback against a folio.
1579 * @folio: The folio.
1580 *
1581 * The folio must actually be under writeback.
1582 *
1583 * Context: May be called from process or interrupt context.
1584 */
1585void folio_end_writeback(struct folio *folio)
1586{
1587 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
1588
1589 /*
1590 * folio_test_clear_reclaim() could be used here but it is an
1591 * atomic operation and overkill in this particular case. Failing
1592 * to shuffle a folio marked for immediate reclaim is too mild
1593 * a gain to justify taking an atomic operation penalty at the
1594 * end of every folio writeback.
1595 */
1596 if (folio_test_reclaim(folio)) {
1597 folio_clear_reclaim(folio);
1598 folio_rotate_reclaimable(folio);
1599 }
1600
1601 /*
1602 * Writeback does not hold a folio reference of its own, relying
1603 * on truncation to wait for the clearing of PG_writeback.
1604 * But here we must make sure that the folio is not freed and
1605 * reused before the folio_wake_bit().
1606 */
1607 folio_get(folio);
1608 if (__folio_end_writeback(folio))
1609 folio_wake_bit(folio, PG_writeback);
1610 acct_reclaim_writeback(folio);
1611 folio_put(folio);
1612}
1613EXPORT_SYMBOL(folio_end_writeback);
1614
1615/**
1616 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1617 * @folio: The folio to lock
1618 */
1619void __folio_lock(struct folio *folio)
1620{
1621 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1622 EXCLUSIVE);
1623}
1624EXPORT_SYMBOL(__folio_lock);
1625
1626int __folio_lock_killable(struct folio *folio)
1627{
1628 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1629 EXCLUSIVE);
1630}
1631EXPORT_SYMBOL_GPL(__folio_lock_killable);
1632
1633static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1634{
1635 struct wait_queue_head *q = folio_waitqueue(folio);
1636 int ret;
1637
1638 wait->folio = folio;
1639 wait->bit_nr = PG_locked;
1640
1641 spin_lock_irq(&q->lock);
1642 __add_wait_queue_entry_tail(q, &wait->wait);
1643 folio_set_waiters(folio);
1644 ret = !folio_trylock(folio);
1645 /*
1646 * If we were successful now, we know we're still on the
1647 * waitqueue as we're still under the lock. This means it's
1648 * safe to remove and return success, we know the callback
1649 * isn't going to trigger.
1650 */
1651 if (!ret)
1652 __remove_wait_queue(q, &wait->wait);
1653 else
1654 ret = -EIOCBQUEUED;
1655 spin_unlock_irq(&q->lock);
1656 return ret;
1657}
1658
1659/*
1660 * Return values:
1661 * 0 - folio is locked.
1662 * non-zero - folio is not locked.
1663 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or
1664 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and
1665 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held.
1666 *
1667 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0
1668 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed.
1669 */
1670vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf)
1671{
1672 unsigned int flags = vmf->flags;
1673
1674 if (fault_flag_allow_retry_first(flags)) {
1675 /*
1676 * CAUTION! In this case, mmap_lock/per-VMA lock is not
1677 * released even though returning VM_FAULT_RETRY.
1678 */
1679 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1680 return VM_FAULT_RETRY;
1681
1682 release_fault_lock(vmf);
1683 if (flags & FAULT_FLAG_KILLABLE)
1684 folio_wait_locked_killable(folio);
1685 else
1686 folio_wait_locked(folio);
1687 return VM_FAULT_RETRY;
1688 }
1689 if (flags & FAULT_FLAG_KILLABLE) {
1690 bool ret;
1691
1692 ret = __folio_lock_killable(folio);
1693 if (ret) {
1694 release_fault_lock(vmf);
1695 return VM_FAULT_RETRY;
1696 }
1697 } else {
1698 __folio_lock(folio);
1699 }
1700
1701 return 0;
1702}
1703
1704/**
1705 * page_cache_next_miss() - Find the next gap in the page cache.
1706 * @mapping: Mapping.
1707 * @index: Index.
1708 * @max_scan: Maximum range to search.
1709 *
1710 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1711 * gap with the lowest index.
1712 *
1713 * This function may be called under the rcu_read_lock. However, this will
1714 * not atomically search a snapshot of the cache at a single point in time.
1715 * For example, if a gap is created at index 5, then subsequently a gap is
1716 * created at index 10, page_cache_next_miss covering both indices may
1717 * return 10 if called under the rcu_read_lock.
1718 *
1719 * Return: The index of the gap if found, otherwise an index outside the
1720 * range specified (in which case 'return - index >= max_scan' will be true).
1721 * In the rare case of index wrap-around, 0 will be returned.
1722 */
1723pgoff_t page_cache_next_miss(struct address_space *mapping,
1724 pgoff_t index, unsigned long max_scan)
1725{
1726 XA_STATE(xas, &mapping->i_pages, index);
1727
1728 while (max_scan--) {
1729 void *entry = xas_next(&xas);
1730 if (!entry || xa_is_value(entry))
1731 break;
1732 if (xas.xa_index == 0)
1733 break;
1734 }
1735
1736 return xas.xa_index;
1737}
1738EXPORT_SYMBOL(page_cache_next_miss);
1739
1740/**
1741 * page_cache_prev_miss() - Find the previous gap in the page cache.
1742 * @mapping: Mapping.
1743 * @index: Index.
1744 * @max_scan: Maximum range to search.
1745 *
1746 * Search the range [max(index - max_scan + 1, 0), index] for the
1747 * gap with the highest index.
1748 *
1749 * This function may be called under the rcu_read_lock. However, this will
1750 * not atomically search a snapshot of the cache at a single point in time.
1751 * For example, if a gap is created at index 10, then subsequently a gap is
1752 * created at index 5, page_cache_prev_miss() covering both indices may
1753 * return 5 if called under the rcu_read_lock.
1754 *
1755 * Return: The index of the gap if found, otherwise an index outside the
1756 * range specified (in which case 'index - return >= max_scan' will be true).
1757 * In the rare case of wrap-around, ULONG_MAX will be returned.
1758 */
1759pgoff_t page_cache_prev_miss(struct address_space *mapping,
1760 pgoff_t index, unsigned long max_scan)
1761{
1762 XA_STATE(xas, &mapping->i_pages, index);
1763
1764 while (max_scan--) {
1765 void *entry = xas_prev(&xas);
1766 if (!entry || xa_is_value(entry))
1767 break;
1768 if (xas.xa_index == ULONG_MAX)
1769 break;
1770 }
1771
1772 return xas.xa_index;
1773}
1774EXPORT_SYMBOL(page_cache_prev_miss);
1775
1776/*
1777 * Lockless page cache protocol:
1778 * On the lookup side:
1779 * 1. Load the folio from i_pages
1780 * 2. Increment the refcount if it's not zero
1781 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1782 *
1783 * On the removal side:
1784 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1785 * B. Remove the page from i_pages
1786 * C. Return the page to the page allocator
1787 *
1788 * This means that any page may have its reference count temporarily
1789 * increased by a speculative page cache (or fast GUP) lookup as it can
1790 * be allocated by another user before the RCU grace period expires.
1791 * Because the refcount temporarily acquired here may end up being the
1792 * last refcount on the page, any page allocation must be freeable by
1793 * folio_put().
1794 */
1795
1796/*
1797 * filemap_get_entry - Get a page cache entry.
1798 * @mapping: the address_space to search
1799 * @index: The page cache index.
1800 *
1801 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1802 * it is returned with an increased refcount. If it is a shadow entry
1803 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1804 * it is returned without further action.
1805 *
1806 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1807 */
1808void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
1809{
1810 XA_STATE(xas, &mapping->i_pages, index);
1811 struct folio *folio;
1812
1813 rcu_read_lock();
1814repeat:
1815 xas_reset(&xas);
1816 folio = xas_load(&xas);
1817 if (xas_retry(&xas, folio))
1818 goto repeat;
1819 /*
1820 * A shadow entry of a recently evicted page, or a swap entry from
1821 * shmem/tmpfs. Return it without attempting to raise page count.
1822 */
1823 if (!folio || xa_is_value(folio))
1824 goto out;
1825
1826 if (!folio_try_get_rcu(folio))
1827 goto repeat;
1828
1829 if (unlikely(folio != xas_reload(&xas))) {
1830 folio_put(folio);
1831 goto repeat;
1832 }
1833out:
1834 rcu_read_unlock();
1835
1836 return folio;
1837}
1838
1839/**
1840 * __filemap_get_folio - Find and get a reference to a folio.
1841 * @mapping: The address_space to search.
1842 * @index: The page index.
1843 * @fgp_flags: %FGP flags modify how the folio is returned.
1844 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1845 *
1846 * Looks up the page cache entry at @mapping & @index.
1847 *
1848 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1849 * if the %GFP flags specified for %FGP_CREAT are atomic.
1850 *
1851 * If this function returns a folio, it is returned with an increased refcount.
1852 *
1853 * Return: The found folio or an ERR_PTR() otherwise.
1854 */
1855struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1856 fgf_t fgp_flags, gfp_t gfp)
1857{
1858 struct folio *folio;
1859
1860repeat:
1861 folio = filemap_get_entry(mapping, index);
1862 if (xa_is_value(folio))
1863 folio = NULL;
1864 if (!folio)
1865 goto no_page;
1866
1867 if (fgp_flags & FGP_LOCK) {
1868 if (fgp_flags & FGP_NOWAIT) {
1869 if (!folio_trylock(folio)) {
1870 folio_put(folio);
1871 return ERR_PTR(-EAGAIN);
1872 }
1873 } else {
1874 folio_lock(folio);
1875 }
1876
1877 /* Has the page been truncated? */
1878 if (unlikely(folio->mapping != mapping)) {
1879 folio_unlock(folio);
1880 folio_put(folio);
1881 goto repeat;
1882 }
1883 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1884 }
1885
1886 if (fgp_flags & FGP_ACCESSED)
1887 folio_mark_accessed(folio);
1888 else if (fgp_flags & FGP_WRITE) {
1889 /* Clear idle flag for buffer write */
1890 if (folio_test_idle(folio))
1891 folio_clear_idle(folio);
1892 }
1893
1894 if (fgp_flags & FGP_STABLE)
1895 folio_wait_stable(folio);
1896no_page:
1897 if (!folio && (fgp_flags & FGP_CREAT)) {
1898 unsigned order = FGF_GET_ORDER(fgp_flags);
1899 int err;
1900
1901 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1902 gfp |= __GFP_WRITE;
1903 if (fgp_flags & FGP_NOFS)
1904 gfp &= ~__GFP_FS;
1905 if (fgp_flags & FGP_NOWAIT) {
1906 gfp &= ~GFP_KERNEL;
1907 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1908 }
1909 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1910 fgp_flags |= FGP_LOCK;
1911
1912 if (!mapping_large_folio_support(mapping))
1913 order = 0;
1914 if (order > MAX_PAGECACHE_ORDER)
1915 order = MAX_PAGECACHE_ORDER;
1916 /* If we're not aligned, allocate a smaller folio */
1917 if (index & ((1UL << order) - 1))
1918 order = __ffs(index);
1919
1920 do {
1921 gfp_t alloc_gfp = gfp;
1922
1923 err = -ENOMEM;
1924 if (order > 0)
1925 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN;
1926 folio = filemap_alloc_folio(alloc_gfp, order);
1927 if (!folio)
1928 continue;
1929
1930 /* Init accessed so avoid atomic mark_page_accessed later */
1931 if (fgp_flags & FGP_ACCESSED)
1932 __folio_set_referenced(folio);
1933
1934 err = filemap_add_folio(mapping, folio, index, gfp);
1935 if (!err)
1936 break;
1937 folio_put(folio);
1938 folio = NULL;
1939 } while (order-- > 0);
1940
1941 if (err == -EEXIST)
1942 goto repeat;
1943 if (err)
1944 return ERR_PTR(err);
1945 /*
1946 * filemap_add_folio locks the page, and for mmap
1947 * we expect an unlocked page.
1948 */
1949 if (folio && (fgp_flags & FGP_FOR_MMAP))
1950 folio_unlock(folio);
1951 }
1952
1953 if (!folio)
1954 return ERR_PTR(-ENOENT);
1955 return folio;
1956}
1957EXPORT_SYMBOL(__filemap_get_folio);
1958
1959static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
1960 xa_mark_t mark)
1961{
1962 struct folio *folio;
1963
1964retry:
1965 if (mark == XA_PRESENT)
1966 folio = xas_find(xas, max);
1967 else
1968 folio = xas_find_marked(xas, max, mark);
1969
1970 if (xas_retry(xas, folio))
1971 goto retry;
1972 /*
1973 * A shadow entry of a recently evicted page, a swap
1974 * entry from shmem/tmpfs or a DAX entry. Return it
1975 * without attempting to raise page count.
1976 */
1977 if (!folio || xa_is_value(folio))
1978 return folio;
1979
1980 if (!folio_try_get_rcu(folio))
1981 goto reset;
1982
1983 if (unlikely(folio != xas_reload(xas))) {
1984 folio_put(folio);
1985 goto reset;
1986 }
1987
1988 return folio;
1989reset:
1990 xas_reset(xas);
1991 goto retry;
1992}
1993
1994/**
1995 * find_get_entries - gang pagecache lookup
1996 * @mapping: The address_space to search
1997 * @start: The starting page cache index
1998 * @end: The final page index (inclusive).
1999 * @fbatch: Where the resulting entries are placed.
2000 * @indices: The cache indices corresponding to the entries in @entries
2001 *
2002 * find_get_entries() will search for and return a batch of entries in
2003 * the mapping. The entries are placed in @fbatch. find_get_entries()
2004 * takes a reference on any actual folios it returns.
2005 *
2006 * The entries have ascending indexes. The indices may not be consecutive
2007 * due to not-present entries or large folios.
2008 *
2009 * Any shadow entries of evicted folios, or swap entries from
2010 * shmem/tmpfs, are included in the returned array.
2011 *
2012 * Return: The number of entries which were found.
2013 */
2014unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2015 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2016{
2017 XA_STATE(xas, &mapping->i_pages, *start);
2018 struct folio *folio;
2019
2020 rcu_read_lock();
2021 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2022 indices[fbatch->nr] = xas.xa_index;
2023 if (!folio_batch_add(fbatch, folio))
2024 break;
2025 }
2026 rcu_read_unlock();
2027
2028 if (folio_batch_count(fbatch)) {
2029 unsigned long nr = 1;
2030 int idx = folio_batch_count(fbatch) - 1;
2031
2032 folio = fbatch->folios[idx];
2033 if (!xa_is_value(folio))
2034 nr = folio_nr_pages(folio);
2035 *start = indices[idx] + nr;
2036 }
2037 return folio_batch_count(fbatch);
2038}
2039
2040/**
2041 * find_lock_entries - Find a batch of pagecache entries.
2042 * @mapping: The address_space to search.
2043 * @start: The starting page cache index.
2044 * @end: The final page index (inclusive).
2045 * @fbatch: Where the resulting entries are placed.
2046 * @indices: The cache indices of the entries in @fbatch.
2047 *
2048 * find_lock_entries() will return a batch of entries from @mapping.
2049 * Swap, shadow and DAX entries are included. Folios are returned
2050 * locked and with an incremented refcount. Folios which are locked
2051 * by somebody else or under writeback are skipped. Folios which are
2052 * partially outside the range are not returned.
2053 *
2054 * The entries have ascending indexes. The indices may not be consecutive
2055 * due to not-present entries, large folios, folios which could not be
2056 * locked or folios under writeback.
2057 *
2058 * Return: The number of entries which were found.
2059 */
2060unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2061 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2062{
2063 XA_STATE(xas, &mapping->i_pages, *start);
2064 struct folio *folio;
2065
2066 rcu_read_lock();
2067 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2068 if (!xa_is_value(folio)) {
2069 if (folio->index < *start)
2070 goto put;
2071 if (folio_next_index(folio) - 1 > end)
2072 goto put;
2073 if (!folio_trylock(folio))
2074 goto put;
2075 if (folio->mapping != mapping ||
2076 folio_test_writeback(folio))
2077 goto unlock;
2078 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2079 folio);
2080 }
2081 indices[fbatch->nr] = xas.xa_index;
2082 if (!folio_batch_add(fbatch, folio))
2083 break;
2084 continue;
2085unlock:
2086 folio_unlock(folio);
2087put:
2088 folio_put(folio);
2089 }
2090 rcu_read_unlock();
2091
2092 if (folio_batch_count(fbatch)) {
2093 unsigned long nr = 1;
2094 int idx = folio_batch_count(fbatch) - 1;
2095
2096 folio = fbatch->folios[idx];
2097 if (!xa_is_value(folio))
2098 nr = folio_nr_pages(folio);
2099 *start = indices[idx] + nr;
2100 }
2101 return folio_batch_count(fbatch);
2102}
2103
2104/**
2105 * filemap_get_folios - Get a batch of folios
2106 * @mapping: The address_space to search
2107 * @start: The starting page index
2108 * @end: The final page index (inclusive)
2109 * @fbatch: The batch to fill.
2110 *
2111 * Search for and return a batch of folios in the mapping starting at
2112 * index @start and up to index @end (inclusive). The folios are returned
2113 * in @fbatch with an elevated reference count.
2114 *
2115 * Return: The number of folios which were found.
2116 * We also update @start to index the next folio for the traversal.
2117 */
2118unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2119 pgoff_t end, struct folio_batch *fbatch)
2120{
2121 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch);
2122}
2123EXPORT_SYMBOL(filemap_get_folios);
2124
2125/**
2126 * filemap_get_folios_contig - Get a batch of contiguous folios
2127 * @mapping: The address_space to search
2128 * @start: The starting page index
2129 * @end: The final page index (inclusive)
2130 * @fbatch: The batch to fill
2131 *
2132 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2133 * except the returned folios are guaranteed to be contiguous. This may
2134 * not return all contiguous folios if the batch gets filled up.
2135 *
2136 * Return: The number of folios found.
2137 * Also update @start to be positioned for traversal of the next folio.
2138 */
2139
2140unsigned filemap_get_folios_contig(struct address_space *mapping,
2141 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2142{
2143 XA_STATE(xas, &mapping->i_pages, *start);
2144 unsigned long nr;
2145 struct folio *folio;
2146
2147 rcu_read_lock();
2148
2149 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2150 folio = xas_next(&xas)) {
2151 if (xas_retry(&xas, folio))
2152 continue;
2153 /*
2154 * If the entry has been swapped out, we can stop looking.
2155 * No current caller is looking for DAX entries.
2156 */
2157 if (xa_is_value(folio))
2158 goto update_start;
2159
2160 if (!folio_try_get_rcu(folio))
2161 goto retry;
2162
2163 if (unlikely(folio != xas_reload(&xas)))
2164 goto put_folio;
2165
2166 if (!folio_batch_add(fbatch, folio)) {
2167 nr = folio_nr_pages(folio);
2168 *start = folio->index + nr;
2169 goto out;
2170 }
2171 continue;
2172put_folio:
2173 folio_put(folio);
2174
2175retry:
2176 xas_reset(&xas);
2177 }
2178
2179update_start:
2180 nr = folio_batch_count(fbatch);
2181
2182 if (nr) {
2183 folio = fbatch->folios[nr - 1];
2184 *start = folio_next_index(folio);
2185 }
2186out:
2187 rcu_read_unlock();
2188 return folio_batch_count(fbatch);
2189}
2190EXPORT_SYMBOL(filemap_get_folios_contig);
2191
2192/**
2193 * filemap_get_folios_tag - Get a batch of folios matching @tag
2194 * @mapping: The address_space to search
2195 * @start: The starting page index
2196 * @end: The final page index (inclusive)
2197 * @tag: The tag index
2198 * @fbatch: The batch to fill
2199 *
2200 * The first folio may start before @start; if it does, it will contain
2201 * @start. The final folio may extend beyond @end; if it does, it will
2202 * contain @end. The folios have ascending indices. There may be gaps
2203 * between the folios if there are indices which have no folio in the
2204 * page cache. If folios are added to or removed from the page cache
2205 * while this is running, they may or may not be found by this call.
2206 * Only returns folios that are tagged with @tag.
2207 *
2208 * Return: The number of folios found.
2209 * Also update @start to index the next folio for traversal.
2210 */
2211unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start,
2212 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch)
2213{
2214 XA_STATE(xas, &mapping->i_pages, *start);
2215 struct folio *folio;
2216
2217 rcu_read_lock();
2218 while ((folio = find_get_entry(&xas, end, tag)) != NULL) {
2219 /*
2220 * Shadow entries should never be tagged, but this iteration
2221 * is lockless so there is a window for page reclaim to evict
2222 * a page we saw tagged. Skip over it.
2223 */
2224 if (xa_is_value(folio))
2225 continue;
2226 if (!folio_batch_add(fbatch, folio)) {
2227 unsigned long nr = folio_nr_pages(folio);
2228 *start = folio->index + nr;
2229 goto out;
2230 }
2231 }
2232 /*
2233 * We come here when there is no page beyond @end. We take care to not
2234 * overflow the index @start as it confuses some of the callers. This
2235 * breaks the iteration when there is a page at index -1 but that is
2236 * already broke anyway.
2237 */
2238 if (end == (pgoff_t)-1)
2239 *start = (pgoff_t)-1;
2240 else
2241 *start = end + 1;
2242out:
2243 rcu_read_unlock();
2244
2245 return folio_batch_count(fbatch);
2246}
2247EXPORT_SYMBOL(filemap_get_folios_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 folios for read
2271 *
2272 * Get a batch of folios which represent a contiguous range of bytes in
2273 * the file. No exceptional entries will be returned. If @index is in
2274 * the middle of a folio, the entire folio will be returned. The last
2275 * folio in the batch may have the readahead flag set or the uptodate flag
2276 * clear so that the 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 folio_batch *fbatch)
2280{
2281 XA_STATE(xas, &mapping->i_pages, index);
2282 struct folio *folio;
2283
2284 rcu_read_lock();
2285 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2286 if (xas_retry(&xas, folio))
2287 continue;
2288 if (xas.xa_index > max || xa_is_value(folio))
2289 break;
2290 if (xa_is_sibling(folio))
2291 break;
2292 if (!folio_try_get_rcu(folio))
2293 goto retry;
2294
2295 if (unlikely(folio != xas_reload(&xas)))
2296 goto put_folio;
2297
2298 if (!folio_batch_add(fbatch, folio))
2299 break;
2300 if (!folio_test_uptodate(folio))
2301 break;
2302 if (folio_test_readahead(folio))
2303 break;
2304 xas_advance(&xas, folio_next_index(folio) - 1);
2305 continue;
2306put_folio:
2307 folio_put(folio);
2308retry:
2309 xas_reset(&xas);
2310 }
2311 rcu_read_unlock();
2312}
2313
2314static int filemap_read_folio(struct file *file, filler_t filler,
2315 struct folio *folio)
2316{
2317 bool workingset = folio_test_workingset(folio);
2318 unsigned long pflags;
2319 int error;
2320
2321 /*
2322 * A previous I/O error may have been due to temporary failures,
2323 * eg. multipath errors. PG_error will be set again if read_folio
2324 * fails.
2325 */
2326 folio_clear_error(folio);
2327
2328 /* Start the actual read. The read will unlock the page. */
2329 if (unlikely(workingset))
2330 psi_memstall_enter(&pflags);
2331 error = filler(file, folio);
2332 if (unlikely(workingset))
2333 psi_memstall_leave(&pflags);
2334 if (error)
2335 return error;
2336
2337 error = folio_wait_locked_killable(folio);
2338 if (error)
2339 return error;
2340 if (folio_test_uptodate(folio))
2341 return 0;
2342 if (file)
2343 shrink_readahead_size_eio(&file->f_ra);
2344 return -EIO;
2345}
2346
2347static bool filemap_range_uptodate(struct address_space *mapping,
2348 loff_t pos, size_t count, struct folio *folio,
2349 bool need_uptodate)
2350{
2351 if (folio_test_uptodate(folio))
2352 return true;
2353 /* pipes can't handle partially uptodate pages */
2354 if (need_uptodate)
2355 return false;
2356 if (!mapping->a_ops->is_partially_uptodate)
2357 return false;
2358 if (mapping->host->i_blkbits >= folio_shift(folio))
2359 return false;
2360
2361 if (folio_pos(folio) > pos) {
2362 count -= folio_pos(folio) - pos;
2363 pos = 0;
2364 } else {
2365 pos -= folio_pos(folio);
2366 }
2367
2368 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2369}
2370
2371static int filemap_update_page(struct kiocb *iocb,
2372 struct address_space *mapping, size_t count,
2373 struct folio *folio, bool need_uptodate)
2374{
2375 int error;
2376
2377 if (iocb->ki_flags & IOCB_NOWAIT) {
2378 if (!filemap_invalidate_trylock_shared(mapping))
2379 return -EAGAIN;
2380 } else {
2381 filemap_invalidate_lock_shared(mapping);
2382 }
2383
2384 if (!folio_trylock(folio)) {
2385 error = -EAGAIN;
2386 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2387 goto unlock_mapping;
2388 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2389 filemap_invalidate_unlock_shared(mapping);
2390 /*
2391 * This is where we usually end up waiting for a
2392 * previously submitted readahead to finish.
2393 */
2394 folio_put_wait_locked(folio, TASK_KILLABLE);
2395 return AOP_TRUNCATED_PAGE;
2396 }
2397 error = __folio_lock_async(folio, iocb->ki_waitq);
2398 if (error)
2399 goto unlock_mapping;
2400 }
2401
2402 error = AOP_TRUNCATED_PAGE;
2403 if (!folio->mapping)
2404 goto unlock;
2405
2406 error = 0;
2407 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio,
2408 need_uptodate))
2409 goto unlock;
2410
2411 error = -EAGAIN;
2412 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2413 goto unlock;
2414
2415 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2416 folio);
2417 goto unlock_mapping;
2418unlock:
2419 folio_unlock(folio);
2420unlock_mapping:
2421 filemap_invalidate_unlock_shared(mapping);
2422 if (error == AOP_TRUNCATED_PAGE)
2423 folio_put(folio);
2424 return error;
2425}
2426
2427static int filemap_create_folio(struct file *file,
2428 struct address_space *mapping, pgoff_t index,
2429 struct folio_batch *fbatch)
2430{
2431 struct folio *folio;
2432 int error;
2433
2434 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), 0);
2435 if (!folio)
2436 return -ENOMEM;
2437
2438 /*
2439 * Protect against truncate / hole punch. Grabbing invalidate_lock
2440 * here assures we cannot instantiate and bring uptodate new
2441 * pagecache folios after evicting page cache during truncate
2442 * and before actually freeing blocks. Note that we could
2443 * release invalidate_lock after inserting the folio into
2444 * the page cache as the locked folio would then be enough to
2445 * synchronize with hole punching. But there are code paths
2446 * such as filemap_update_page() filling in partially uptodate
2447 * pages or ->readahead() that need to hold invalidate_lock
2448 * while mapping blocks for IO so let's hold the lock here as
2449 * well to keep locking rules simple.
2450 */
2451 filemap_invalidate_lock_shared(mapping);
2452 error = filemap_add_folio(mapping, folio, index,
2453 mapping_gfp_constraint(mapping, GFP_KERNEL));
2454 if (error == -EEXIST)
2455 error = AOP_TRUNCATED_PAGE;
2456 if (error)
2457 goto error;
2458
2459 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
2460 if (error)
2461 goto error;
2462
2463 filemap_invalidate_unlock_shared(mapping);
2464 folio_batch_add(fbatch, folio);
2465 return 0;
2466error:
2467 filemap_invalidate_unlock_shared(mapping);
2468 folio_put(folio);
2469 return error;
2470}
2471
2472static int filemap_readahead(struct kiocb *iocb, struct file *file,
2473 struct address_space *mapping, struct folio *folio,
2474 pgoff_t last_index)
2475{
2476 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2477
2478 if (iocb->ki_flags & IOCB_NOIO)
2479 return -EAGAIN;
2480 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2481 return 0;
2482}
2483
2484static int filemap_get_pages(struct kiocb *iocb, size_t count,
2485 struct folio_batch *fbatch, bool need_uptodate)
2486{
2487 struct file *filp = iocb->ki_filp;
2488 struct address_space *mapping = filp->f_mapping;
2489 struct file_ra_state *ra = &filp->f_ra;
2490 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2491 pgoff_t last_index;
2492 struct folio *folio;
2493 int err = 0;
2494
2495 /* "last_index" is the index of the page beyond the end of the read */
2496 last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
2497retry:
2498 if (fatal_signal_pending(current))
2499 return -EINTR;
2500
2501 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2502 if (!folio_batch_count(fbatch)) {
2503 if (iocb->ki_flags & IOCB_NOIO)
2504 return -EAGAIN;
2505 page_cache_sync_readahead(mapping, ra, filp, index,
2506 last_index - index);
2507 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2508 }
2509 if (!folio_batch_count(fbatch)) {
2510 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2511 return -EAGAIN;
2512 err = filemap_create_folio(filp, mapping,
2513 iocb->ki_pos >> PAGE_SHIFT, fbatch);
2514 if (err == AOP_TRUNCATED_PAGE)
2515 goto retry;
2516 return err;
2517 }
2518
2519 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2520 if (folio_test_readahead(folio)) {
2521 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2522 if (err)
2523 goto err;
2524 }
2525 if (!folio_test_uptodate(folio)) {
2526 if ((iocb->ki_flags & IOCB_WAITQ) &&
2527 folio_batch_count(fbatch) > 1)
2528 iocb->ki_flags |= IOCB_NOWAIT;
2529 err = filemap_update_page(iocb, mapping, count, folio,
2530 need_uptodate);
2531 if (err)
2532 goto err;
2533 }
2534
2535 return 0;
2536err:
2537 if (err < 0)
2538 folio_put(folio);
2539 if (likely(--fbatch->nr))
2540 return 0;
2541 if (err == AOP_TRUNCATED_PAGE)
2542 goto retry;
2543 return err;
2544}
2545
2546static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2547{
2548 unsigned int shift = folio_shift(folio);
2549
2550 return (pos1 >> shift == pos2 >> shift);
2551}
2552
2553/**
2554 * filemap_read - Read data from the page cache.
2555 * @iocb: The iocb to read.
2556 * @iter: Destination for the data.
2557 * @already_read: Number of bytes already read by the caller.
2558 *
2559 * Copies data from the page cache. If the data is not currently present,
2560 * uses the readahead and read_folio address_space operations to fetch it.
2561 *
2562 * Return: Total number of bytes copied, including those already read by
2563 * the caller. If an error happens before any bytes are copied, returns
2564 * a negative error number.
2565 */
2566ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2567 ssize_t already_read)
2568{
2569 struct file *filp = iocb->ki_filp;
2570 struct file_ra_state *ra = &filp->f_ra;
2571 struct address_space *mapping = filp->f_mapping;
2572 struct inode *inode = mapping->host;
2573 struct folio_batch fbatch;
2574 int i, error = 0;
2575 bool writably_mapped;
2576 loff_t isize, end_offset;
2577 loff_t last_pos = ra->prev_pos;
2578
2579 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2580 return 0;
2581 if (unlikely(!iov_iter_count(iter)))
2582 return 0;
2583
2584 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2585 folio_batch_init(&fbatch);
2586
2587 do {
2588 cond_resched();
2589
2590 /*
2591 * If we've already successfully copied some data, then we
2592 * can no longer safely return -EIOCBQUEUED. Hence mark
2593 * an async read NOWAIT at that point.
2594 */
2595 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2596 iocb->ki_flags |= IOCB_NOWAIT;
2597
2598 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2599 break;
2600
2601 error = filemap_get_pages(iocb, iter->count, &fbatch, false);
2602 if (error < 0)
2603 break;
2604
2605 /*
2606 * i_size must be checked after we know the pages are Uptodate.
2607 *
2608 * Checking i_size after the check allows us to calculate
2609 * the correct value for "nr", which means the zero-filled
2610 * part of the page is not copied back to userspace (unless
2611 * another truncate extends the file - this is desired though).
2612 */
2613 isize = i_size_read(inode);
2614 if (unlikely(iocb->ki_pos >= isize))
2615 goto put_folios;
2616 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2617
2618 /*
2619 * Once we start copying data, we don't want to be touching any
2620 * cachelines that might be contended:
2621 */
2622 writably_mapped = mapping_writably_mapped(mapping);
2623
2624 /*
2625 * When a read accesses the same folio several times, only
2626 * mark it as accessed the first time.
2627 */
2628 if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
2629 fbatch.folios[0]))
2630 folio_mark_accessed(fbatch.folios[0]);
2631
2632 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2633 struct folio *folio = fbatch.folios[i];
2634 size_t fsize = folio_size(folio);
2635 size_t offset = iocb->ki_pos & (fsize - 1);
2636 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2637 fsize - offset);
2638 size_t copied;
2639
2640 if (end_offset < folio_pos(folio))
2641 break;
2642 if (i > 0)
2643 folio_mark_accessed(folio);
2644 /*
2645 * If users can be writing to this folio using arbitrary
2646 * virtual addresses, take care of potential aliasing
2647 * before reading the folio on the kernel side.
2648 */
2649 if (writably_mapped)
2650 flush_dcache_folio(folio);
2651
2652 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2653
2654 already_read += copied;
2655 iocb->ki_pos += copied;
2656 last_pos = iocb->ki_pos;
2657
2658 if (copied < bytes) {
2659 error = -EFAULT;
2660 break;
2661 }
2662 }
2663put_folios:
2664 for (i = 0; i < folio_batch_count(&fbatch); i++)
2665 folio_put(fbatch.folios[i]);
2666 folio_batch_init(&fbatch);
2667 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2668
2669 file_accessed(filp);
2670 ra->prev_pos = last_pos;
2671 return already_read ? already_read : error;
2672}
2673EXPORT_SYMBOL_GPL(filemap_read);
2674
2675int kiocb_write_and_wait(struct kiocb *iocb, size_t count)
2676{
2677 struct address_space *mapping = iocb->ki_filp->f_mapping;
2678 loff_t pos = iocb->ki_pos;
2679 loff_t end = pos + count - 1;
2680
2681 if (iocb->ki_flags & IOCB_NOWAIT) {
2682 if (filemap_range_needs_writeback(mapping, pos, end))
2683 return -EAGAIN;
2684 return 0;
2685 }
2686
2687 return filemap_write_and_wait_range(mapping, pos, end);
2688}
2689EXPORT_SYMBOL_GPL(kiocb_write_and_wait);
2690
2691int kiocb_invalidate_pages(struct kiocb *iocb, size_t count)
2692{
2693 struct address_space *mapping = iocb->ki_filp->f_mapping;
2694 loff_t pos = iocb->ki_pos;
2695 loff_t end = pos + count - 1;
2696 int ret;
2697
2698 if (iocb->ki_flags & IOCB_NOWAIT) {
2699 /* we could block if there are any pages in the range */
2700 if (filemap_range_has_page(mapping, pos, end))
2701 return -EAGAIN;
2702 } else {
2703 ret = filemap_write_and_wait_range(mapping, pos, end);
2704 if (ret)
2705 return ret;
2706 }
2707
2708 /*
2709 * After a write we want buffered reads to be sure to go to disk to get
2710 * the new data. We invalidate clean cached page from the region we're
2711 * about to write. We do this *before* the write so that we can return
2712 * without clobbering -EIOCBQUEUED from ->direct_IO().
2713 */
2714 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
2715 end >> PAGE_SHIFT);
2716}
2717EXPORT_SYMBOL_GPL(kiocb_invalidate_pages);
2718
2719/**
2720 * generic_file_read_iter - generic filesystem read routine
2721 * @iocb: kernel I/O control block
2722 * @iter: destination for the data read
2723 *
2724 * This is the "read_iter()" routine for all filesystems
2725 * that can use the page cache directly.
2726 *
2727 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2728 * be returned when no data can be read without waiting for I/O requests
2729 * to complete; it doesn't prevent readahead.
2730 *
2731 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2732 * requests shall be made for the read or for readahead. When no data
2733 * can be read, -EAGAIN shall be returned. When readahead would be
2734 * triggered, a partial, possibly empty read shall be returned.
2735 *
2736 * Return:
2737 * * number of bytes copied, even for partial reads
2738 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2739 */
2740ssize_t
2741generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2742{
2743 size_t count = iov_iter_count(iter);
2744 ssize_t retval = 0;
2745
2746 if (!count)
2747 return 0; /* skip atime */
2748
2749 if (iocb->ki_flags & IOCB_DIRECT) {
2750 struct file *file = iocb->ki_filp;
2751 struct address_space *mapping = file->f_mapping;
2752 struct inode *inode = mapping->host;
2753
2754 retval = kiocb_write_and_wait(iocb, count);
2755 if (retval < 0)
2756 return retval;
2757 file_accessed(file);
2758
2759 retval = mapping->a_ops->direct_IO(iocb, iter);
2760 if (retval >= 0) {
2761 iocb->ki_pos += retval;
2762 count -= retval;
2763 }
2764 if (retval != -EIOCBQUEUED)
2765 iov_iter_revert(iter, count - iov_iter_count(iter));
2766
2767 /*
2768 * Btrfs can have a short DIO read if we encounter
2769 * compressed extents, so if there was an error, or if
2770 * we've already read everything we wanted to, or if
2771 * there was a short read because we hit EOF, go ahead
2772 * and return. Otherwise fallthrough to buffered io for
2773 * the rest of the read. Buffered reads will not work for
2774 * DAX files, so don't bother trying.
2775 */
2776 if (retval < 0 || !count || IS_DAX(inode))
2777 return retval;
2778 if (iocb->ki_pos >= i_size_read(inode))
2779 return retval;
2780 }
2781
2782 return filemap_read(iocb, iter, retval);
2783}
2784EXPORT_SYMBOL(generic_file_read_iter);
2785
2786/*
2787 * Splice subpages from a folio into a pipe.
2788 */
2789size_t splice_folio_into_pipe(struct pipe_inode_info *pipe,
2790 struct folio *folio, loff_t fpos, size_t size)
2791{
2792 struct page *page;
2793 size_t spliced = 0, offset = offset_in_folio(folio, fpos);
2794
2795 page = folio_page(folio, offset / PAGE_SIZE);
2796 size = min(size, folio_size(folio) - offset);
2797 offset %= PAGE_SIZE;
2798
2799 while (spliced < size &&
2800 !pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
2801 struct pipe_buffer *buf = pipe_head_buf(pipe);
2802 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced);
2803
2804 *buf = (struct pipe_buffer) {
2805 .ops = &page_cache_pipe_buf_ops,
2806 .page = page,
2807 .offset = offset,
2808 .len = part,
2809 };
2810 folio_get(folio);
2811 pipe->head++;
2812 page++;
2813 spliced += part;
2814 offset = 0;
2815 }
2816
2817 return spliced;
2818}
2819
2820/**
2821 * filemap_splice_read - Splice data from a file's pagecache into a pipe
2822 * @in: The file to read from
2823 * @ppos: Pointer to the file position to read from
2824 * @pipe: The pipe to splice into
2825 * @len: The amount to splice
2826 * @flags: The SPLICE_F_* flags
2827 *
2828 * This function gets folios from a file's pagecache and splices them into the
2829 * pipe. Readahead will be called as necessary to fill more folios. This may
2830 * be used for blockdevs also.
2831 *
2832 * Return: On success, the number of bytes read will be returned and *@ppos
2833 * will be updated if appropriate; 0 will be returned if there is no more data
2834 * to be read; -EAGAIN will be returned if the pipe had no space, and some
2835 * other negative error code will be returned on error. A short read may occur
2836 * if the pipe has insufficient space, we reach the end of the data or we hit a
2837 * hole.
2838 */
2839ssize_t filemap_splice_read(struct file *in, loff_t *ppos,
2840 struct pipe_inode_info *pipe,
2841 size_t len, unsigned int flags)
2842{
2843 struct folio_batch fbatch;
2844 struct kiocb iocb;
2845 size_t total_spliced = 0, used, npages;
2846 loff_t isize, end_offset;
2847 bool writably_mapped;
2848 int i, error = 0;
2849
2850 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes))
2851 return 0;
2852
2853 init_sync_kiocb(&iocb, in);
2854 iocb.ki_pos = *ppos;
2855
2856 /* Work out how much data we can actually add into the pipe */
2857 used = pipe_occupancy(pipe->head, pipe->tail);
2858 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2859 len = min_t(size_t, len, npages * PAGE_SIZE);
2860
2861 folio_batch_init(&fbatch);
2862
2863 do {
2864 cond_resched();
2865
2866 if (*ppos >= i_size_read(in->f_mapping->host))
2867 break;
2868
2869 iocb.ki_pos = *ppos;
2870 error = filemap_get_pages(&iocb, len, &fbatch, true);
2871 if (error < 0)
2872 break;
2873
2874 /*
2875 * i_size must be checked after we know the pages are Uptodate.
2876 *
2877 * Checking i_size after the check allows us to calculate
2878 * the correct value for "nr", which means the zero-filled
2879 * part of the page is not copied back to userspace (unless
2880 * another truncate extends the file - this is desired though).
2881 */
2882 isize = i_size_read(in->f_mapping->host);
2883 if (unlikely(*ppos >= isize))
2884 break;
2885 end_offset = min_t(loff_t, isize, *ppos + len);
2886
2887 /*
2888 * Once we start copying data, we don't want to be touching any
2889 * cachelines that might be contended:
2890 */
2891 writably_mapped = mapping_writably_mapped(in->f_mapping);
2892
2893 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2894 struct folio *folio = fbatch.folios[i];
2895 size_t n;
2896
2897 if (folio_pos(folio) >= end_offset)
2898 goto out;
2899 folio_mark_accessed(folio);
2900
2901 /*
2902 * If users can be writing to this folio using arbitrary
2903 * virtual addresses, take care of potential aliasing
2904 * before reading the folio on the kernel side.
2905 */
2906 if (writably_mapped)
2907 flush_dcache_folio(folio);
2908
2909 n = min_t(loff_t, len, isize - *ppos);
2910 n = splice_folio_into_pipe(pipe, folio, *ppos, n);
2911 if (!n)
2912 goto out;
2913 len -= n;
2914 total_spliced += n;
2915 *ppos += n;
2916 in->f_ra.prev_pos = *ppos;
2917 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
2918 goto out;
2919 }
2920
2921 folio_batch_release(&fbatch);
2922 } while (len);
2923
2924out:
2925 folio_batch_release(&fbatch);
2926 file_accessed(in);
2927
2928 return total_spliced ? total_spliced : error;
2929}
2930EXPORT_SYMBOL(filemap_splice_read);
2931
2932static inline loff_t folio_seek_hole_data(struct xa_state *xas,
2933 struct address_space *mapping, struct folio *folio,
2934 loff_t start, loff_t end, bool seek_data)
2935{
2936 const struct address_space_operations *ops = mapping->a_ops;
2937 size_t offset, bsz = i_blocksize(mapping->host);
2938
2939 if (xa_is_value(folio) || folio_test_uptodate(folio))
2940 return seek_data ? start : end;
2941 if (!ops->is_partially_uptodate)
2942 return seek_data ? end : start;
2943
2944 xas_pause(xas);
2945 rcu_read_unlock();
2946 folio_lock(folio);
2947 if (unlikely(folio->mapping != mapping))
2948 goto unlock;
2949
2950 offset = offset_in_folio(folio, start) & ~(bsz - 1);
2951
2952 do {
2953 if (ops->is_partially_uptodate(folio, offset, bsz) ==
2954 seek_data)
2955 break;
2956 start = (start + bsz) & ~(bsz - 1);
2957 offset += bsz;
2958 } while (offset < folio_size(folio));
2959unlock:
2960 folio_unlock(folio);
2961 rcu_read_lock();
2962 return start;
2963}
2964
2965static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
2966{
2967 if (xa_is_value(folio))
2968 return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index);
2969 return folio_size(folio);
2970}
2971
2972/**
2973 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
2974 * @mapping: Address space to search.
2975 * @start: First byte to consider.
2976 * @end: Limit of search (exclusive).
2977 * @whence: Either SEEK_HOLE or SEEK_DATA.
2978 *
2979 * If the page cache knows which blocks contain holes and which blocks
2980 * contain data, your filesystem can use this function to implement
2981 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
2982 * entirely memory-based such as tmpfs, and filesystems which support
2983 * unwritten extents.
2984 *
2985 * Return: The requested offset on success, or -ENXIO if @whence specifies
2986 * SEEK_DATA and there is no data after @start. There is an implicit hole
2987 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
2988 * and @end contain data.
2989 */
2990loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
2991 loff_t end, int whence)
2992{
2993 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
2994 pgoff_t max = (end - 1) >> PAGE_SHIFT;
2995 bool seek_data = (whence == SEEK_DATA);
2996 struct folio *folio;
2997
2998 if (end <= start)
2999 return -ENXIO;
3000
3001 rcu_read_lock();
3002 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
3003 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
3004 size_t seek_size;
3005
3006 if (start < pos) {
3007 if (!seek_data)
3008 goto unlock;
3009 start = pos;
3010 }
3011
3012 seek_size = seek_folio_size(&xas, folio);
3013 pos = round_up((u64)pos + 1, seek_size);
3014 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
3015 seek_data);
3016 if (start < pos)
3017 goto unlock;
3018 if (start >= end)
3019 break;
3020 if (seek_size > PAGE_SIZE)
3021 xas_set(&xas, pos >> PAGE_SHIFT);
3022 if (!xa_is_value(folio))
3023 folio_put(folio);
3024 }
3025 if (seek_data)
3026 start = -ENXIO;
3027unlock:
3028 rcu_read_unlock();
3029 if (folio && !xa_is_value(folio))
3030 folio_put(folio);
3031 if (start > end)
3032 return end;
3033 return start;
3034}
3035
3036#ifdef CONFIG_MMU
3037#define MMAP_LOTSAMISS (100)
3038/*
3039 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
3040 * @vmf - the vm_fault for this fault.
3041 * @folio - the folio to lock.
3042 * @fpin - the pointer to the file we may pin (or is already pinned).
3043 *
3044 * This works similar to lock_folio_or_retry in that it can drop the
3045 * mmap_lock. It differs in that it actually returns the folio locked
3046 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
3047 * to drop the mmap_lock then fpin will point to the pinned file and
3048 * needs to be fput()'ed at a later point.
3049 */
3050static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
3051 struct file **fpin)
3052{
3053 if (folio_trylock(folio))
3054 return 1;
3055
3056 /*
3057 * NOTE! This will make us return with VM_FAULT_RETRY, but with
3058 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
3059 * is supposed to work. We have way too many special cases..
3060 */
3061 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
3062 return 0;
3063
3064 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
3065 if (vmf->flags & FAULT_FLAG_KILLABLE) {
3066 if (__folio_lock_killable(folio)) {
3067 /*
3068 * We didn't have the right flags to drop the
3069 * fault lock, but all fault_handlers only check
3070 * for fatal signals if we return VM_FAULT_RETRY,
3071 * so we need to drop the fault lock here and
3072 * return 0 if we don't have a fpin.
3073 */
3074 if (*fpin == NULL)
3075 release_fault_lock(vmf);
3076 return 0;
3077 }
3078 } else
3079 __folio_lock(folio);
3080
3081 return 1;
3082}
3083
3084/*
3085 * Synchronous readahead happens when we don't even find a page in the page
3086 * cache at all. We don't want to perform IO under the mmap sem, so if we have
3087 * to drop the mmap sem we return the file that was pinned in order for us to do
3088 * that. If we didn't pin a file then we return NULL. The file that is
3089 * returned needs to be fput()'ed when we're done with it.
3090 */
3091static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3092{
3093 struct file *file = vmf->vma->vm_file;
3094 struct file_ra_state *ra = &file->f_ra;
3095 struct address_space *mapping = file->f_mapping;
3096 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3097 struct file *fpin = NULL;
3098 unsigned long vm_flags = vmf->vma->vm_flags;
3099 unsigned int mmap_miss;
3100
3101#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3102 /* Use the readahead code, even if readahead is disabled */
3103 if (vm_flags & VM_HUGEPAGE) {
3104 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3105 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3106 ra->size = HPAGE_PMD_NR;
3107 /*
3108 * Fetch two PMD folios, so we get the chance to actually
3109 * readahead, unless we've been told not to.
3110 */
3111 if (!(vm_flags & VM_RAND_READ))
3112 ra->size *= 2;
3113 ra->async_size = HPAGE_PMD_NR;
3114 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3115 return fpin;
3116 }
3117#endif
3118
3119 /* If we don't want any read-ahead, don't bother */
3120 if (vm_flags & VM_RAND_READ)
3121 return fpin;
3122 if (!ra->ra_pages)
3123 return fpin;
3124
3125 if (vm_flags & VM_SEQ_READ) {
3126 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3127 page_cache_sync_ra(&ractl, ra->ra_pages);
3128 return fpin;
3129 }
3130
3131 /* Avoid banging the cache line if not needed */
3132 mmap_miss = READ_ONCE(ra->mmap_miss);
3133 if (mmap_miss < MMAP_LOTSAMISS * 10)
3134 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3135
3136 /*
3137 * Do we miss much more than hit in this file? If so,
3138 * stop bothering with read-ahead. It will only hurt.
3139 */
3140 if (mmap_miss > MMAP_LOTSAMISS)
3141 return fpin;
3142
3143 /*
3144 * mmap read-around
3145 */
3146 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3147 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3148 ra->size = ra->ra_pages;
3149 ra->async_size = ra->ra_pages / 4;
3150 ractl._index = ra->start;
3151 page_cache_ra_order(&ractl, ra, 0);
3152 return fpin;
3153}
3154
3155/*
3156 * Asynchronous readahead happens when we find the page and PG_readahead,
3157 * so we want to possibly extend the readahead further. We return the file that
3158 * was pinned if we have to drop the mmap_lock in order to do IO.
3159 */
3160static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3161 struct folio *folio)
3162{
3163 struct file *file = vmf->vma->vm_file;
3164 struct file_ra_state *ra = &file->f_ra;
3165 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3166 struct file *fpin = NULL;
3167 unsigned int mmap_miss;
3168
3169 /* If we don't want any read-ahead, don't bother */
3170 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3171 return fpin;
3172
3173 mmap_miss = READ_ONCE(ra->mmap_miss);
3174 if (mmap_miss)
3175 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3176
3177 if (folio_test_readahead(folio)) {
3178 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3179 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3180 }
3181 return fpin;
3182}
3183
3184static vm_fault_t filemap_fault_recheck_pte_none(struct vm_fault *vmf)
3185{
3186 struct vm_area_struct *vma = vmf->vma;
3187 vm_fault_t ret = 0;
3188 pte_t *ptep;
3189
3190 /*
3191 * We might have COW'ed a pagecache folio and might now have an mlocked
3192 * anon folio mapped. The original pagecache folio is not mlocked and
3193 * might have been evicted. During a read+clear/modify/write update of
3194 * the PTE, such as done in do_numa_page()/change_pte_range(), we
3195 * temporarily clear the PTE under PT lock and might detect it here as
3196 * "none" when not holding the PT lock.
3197 *
3198 * Not rechecking the PTE under PT lock could result in an unexpected
3199 * major fault in an mlock'ed region. Recheck only for this special
3200 * scenario while holding the PT lock, to not degrade non-mlocked
3201 * scenarios. Recheck the PTE without PT lock firstly, thereby reducing
3202 * the number of times we hold PT lock.
3203 */
3204 if (!(vma->vm_flags & VM_LOCKED))
3205 return 0;
3206
3207 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
3208 return 0;
3209
3210 ptep = pte_offset_map(vmf->pmd, vmf->address);
3211 if (unlikely(!ptep))
3212 return VM_FAULT_NOPAGE;
3213
3214 if (unlikely(!pte_none(ptep_get_lockless(ptep)))) {
3215 ret = VM_FAULT_NOPAGE;
3216 } else {
3217 spin_lock(vmf->ptl);
3218 if (unlikely(!pte_none(ptep_get(ptep))))
3219 ret = VM_FAULT_NOPAGE;
3220 spin_unlock(vmf->ptl);
3221 }
3222 pte_unmap(ptep);
3223 return ret;
3224}
3225
3226/**
3227 * filemap_fault - read in file data for page fault handling
3228 * @vmf: struct vm_fault containing details of the fault
3229 *
3230 * filemap_fault() is invoked via the vma operations vector for a
3231 * mapped memory region to read in file data during a page fault.
3232 *
3233 * The goto's are kind of ugly, but this streamlines the normal case of having
3234 * it in the page cache, and handles the special cases reasonably without
3235 * having a lot of duplicated code.
3236 *
3237 * vma->vm_mm->mmap_lock must be held on entry.
3238 *
3239 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3240 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3241 *
3242 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3243 * has not been released.
3244 *
3245 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3246 *
3247 * Return: bitwise-OR of %VM_FAULT_ codes.
3248 */
3249vm_fault_t filemap_fault(struct vm_fault *vmf)
3250{
3251 int error;
3252 struct file *file = vmf->vma->vm_file;
3253 struct file *fpin = NULL;
3254 struct address_space *mapping = file->f_mapping;
3255 struct inode *inode = mapping->host;
3256 pgoff_t max_idx, index = vmf->pgoff;
3257 struct folio *folio;
3258 vm_fault_t ret = 0;
3259 bool mapping_locked = false;
3260
3261 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3262 if (unlikely(index >= max_idx))
3263 return VM_FAULT_SIGBUS;
3264
3265 /*
3266 * Do we have something in the page cache already?
3267 */
3268 folio = filemap_get_folio(mapping, index);
3269 if (likely(!IS_ERR(folio))) {
3270 /*
3271 * We found the page, so try async readahead before waiting for
3272 * the lock.
3273 */
3274 if (!(vmf->flags & FAULT_FLAG_TRIED))
3275 fpin = do_async_mmap_readahead(vmf, folio);
3276 if (unlikely(!folio_test_uptodate(folio))) {
3277 filemap_invalidate_lock_shared(mapping);
3278 mapping_locked = true;
3279 }
3280 } else {
3281 ret = filemap_fault_recheck_pte_none(vmf);
3282 if (unlikely(ret))
3283 return ret;
3284
3285 /* No page in the page cache at all */
3286 count_vm_event(PGMAJFAULT);
3287 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3288 ret = VM_FAULT_MAJOR;
3289 fpin = do_sync_mmap_readahead(vmf);
3290retry_find:
3291 /*
3292 * See comment in filemap_create_folio() why we need
3293 * invalidate_lock
3294 */
3295 if (!mapping_locked) {
3296 filemap_invalidate_lock_shared(mapping);
3297 mapping_locked = true;
3298 }
3299 folio = __filemap_get_folio(mapping, index,
3300 FGP_CREAT|FGP_FOR_MMAP,
3301 vmf->gfp_mask);
3302 if (IS_ERR(folio)) {
3303 if (fpin)
3304 goto out_retry;
3305 filemap_invalidate_unlock_shared(mapping);
3306 return VM_FAULT_OOM;
3307 }
3308 }
3309
3310 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3311 goto out_retry;
3312
3313 /* Did it get truncated? */
3314 if (unlikely(folio->mapping != mapping)) {
3315 folio_unlock(folio);
3316 folio_put(folio);
3317 goto retry_find;
3318 }
3319 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3320
3321 /*
3322 * We have a locked folio in the page cache, now we need to check
3323 * that it's up-to-date. If not, it is going to be due to an error,
3324 * or because readahead was otherwise unable to retrieve it.
3325 */
3326 if (unlikely(!folio_test_uptodate(folio))) {
3327 /*
3328 * If the invalidate lock is not held, the folio was in cache
3329 * and uptodate and now it is not. Strange but possible since we
3330 * didn't hold the page lock all the time. Let's drop
3331 * everything, get the invalidate lock and try again.
3332 */
3333 if (!mapping_locked) {
3334 folio_unlock(folio);
3335 folio_put(folio);
3336 goto retry_find;
3337 }
3338
3339 /*
3340 * OK, the folio is really not uptodate. This can be because the
3341 * VMA has the VM_RAND_READ flag set, or because an error
3342 * arose. Let's read it in directly.
3343 */
3344 goto page_not_uptodate;
3345 }
3346
3347 /*
3348 * We've made it this far and we had to drop our mmap_lock, now is the
3349 * time to return to the upper layer and have it re-find the vma and
3350 * redo the fault.
3351 */
3352 if (fpin) {
3353 folio_unlock(folio);
3354 goto out_retry;
3355 }
3356 if (mapping_locked)
3357 filemap_invalidate_unlock_shared(mapping);
3358
3359 /*
3360 * Found the page and have a reference on it.
3361 * We must recheck i_size under page lock.
3362 */
3363 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3364 if (unlikely(index >= max_idx)) {
3365 folio_unlock(folio);
3366 folio_put(folio);
3367 return VM_FAULT_SIGBUS;
3368 }
3369
3370 vmf->page = folio_file_page(folio, index);
3371 return ret | VM_FAULT_LOCKED;
3372
3373page_not_uptodate:
3374 /*
3375 * Umm, take care of errors if the page isn't up-to-date.
3376 * Try to re-read it _once_. We do this synchronously,
3377 * because there really aren't any performance issues here
3378 * and we need to check for errors.
3379 */
3380 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3381 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3382 if (fpin)
3383 goto out_retry;
3384 folio_put(folio);
3385
3386 if (!error || error == AOP_TRUNCATED_PAGE)
3387 goto retry_find;
3388 filemap_invalidate_unlock_shared(mapping);
3389
3390 return VM_FAULT_SIGBUS;
3391
3392out_retry:
3393 /*
3394 * We dropped the mmap_lock, we need to return to the fault handler to
3395 * re-find the vma and come back and find our hopefully still populated
3396 * page.
3397 */
3398 if (!IS_ERR(folio))
3399 folio_put(folio);
3400 if (mapping_locked)
3401 filemap_invalidate_unlock_shared(mapping);
3402 if (fpin)
3403 fput(fpin);
3404 return ret | VM_FAULT_RETRY;
3405}
3406EXPORT_SYMBOL(filemap_fault);
3407
3408static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio,
3409 pgoff_t start)
3410{
3411 struct mm_struct *mm = vmf->vma->vm_mm;
3412
3413 /* Huge page is mapped? No need to proceed. */
3414 if (pmd_trans_huge(*vmf->pmd)) {
3415 folio_unlock(folio);
3416 folio_put(folio);
3417 return true;
3418 }
3419
3420 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) {
3421 struct page *page = folio_file_page(folio, start);
3422 vm_fault_t ret = do_set_pmd(vmf, page);
3423 if (!ret) {
3424 /* The page is mapped successfully, reference consumed. */
3425 folio_unlock(folio);
3426 return true;
3427 }
3428 }
3429
3430 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte)
3431 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3432
3433 return false;
3434}
3435
3436static struct folio *next_uptodate_folio(struct xa_state *xas,
3437 struct address_space *mapping, pgoff_t end_pgoff)
3438{
3439 struct folio *folio = xas_next_entry(xas, end_pgoff);
3440 unsigned long max_idx;
3441
3442 do {
3443 if (!folio)
3444 return NULL;
3445 if (xas_retry(xas, folio))
3446 continue;
3447 if (xa_is_value(folio))
3448 continue;
3449 if (folio_test_locked(folio))
3450 continue;
3451 if (!folio_try_get_rcu(folio))
3452 continue;
3453 /* Has the page moved or been split? */
3454 if (unlikely(folio != xas_reload(xas)))
3455 goto skip;
3456 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3457 goto skip;
3458 if (!folio_trylock(folio))
3459 goto skip;
3460 if (folio->mapping != mapping)
3461 goto unlock;
3462 if (!folio_test_uptodate(folio))
3463 goto unlock;
3464 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3465 if (xas->xa_index >= max_idx)
3466 goto unlock;
3467 return folio;
3468unlock:
3469 folio_unlock(folio);
3470skip:
3471 folio_put(folio);
3472 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3473
3474 return NULL;
3475}
3476
3477/*
3478 * Map page range [start_page, start_page + nr_pages) of folio.
3479 * start_page is gotten from start by folio_page(folio, start)
3480 */
3481static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
3482 struct folio *folio, unsigned long start,
3483 unsigned long addr, unsigned int nr_pages,
3484 unsigned int *mmap_miss)
3485{
3486 vm_fault_t ret = 0;
3487 struct page *page = folio_page(folio, start);
3488 unsigned int count = 0;
3489 pte_t *old_ptep = vmf->pte;
3490
3491 do {
3492 if (PageHWPoison(page + count))
3493 goto skip;
3494
3495 (*mmap_miss)++;
3496
3497 /*
3498 * NOTE: If there're PTE markers, we'll leave them to be
3499 * handled in the specific fault path, and it'll prohibit the
3500 * fault-around logic.
3501 */
3502 if (!pte_none(ptep_get(&vmf->pte[count])))
3503 goto skip;
3504
3505 count++;
3506 continue;
3507skip:
3508 if (count) {
3509 set_pte_range(vmf, folio, page, count, addr);
3510 folio_ref_add(folio, count);
3511 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3512 ret = VM_FAULT_NOPAGE;
3513 }
3514
3515 count++;
3516 page += count;
3517 vmf->pte += count;
3518 addr += count * PAGE_SIZE;
3519 count = 0;
3520 } while (--nr_pages > 0);
3521
3522 if (count) {
3523 set_pte_range(vmf, folio, page, count, addr);
3524 folio_ref_add(folio, count);
3525 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3526 ret = VM_FAULT_NOPAGE;
3527 }
3528
3529 vmf->pte = old_ptep;
3530
3531 return ret;
3532}
3533
3534static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
3535 struct folio *folio, unsigned long addr,
3536 unsigned int *mmap_miss)
3537{
3538 vm_fault_t ret = 0;
3539 struct page *page = &folio->page;
3540
3541 if (PageHWPoison(page))
3542 return ret;
3543
3544 (*mmap_miss)++;
3545
3546 /*
3547 * NOTE: If there're PTE markers, we'll leave them to be
3548 * handled in the specific fault path, and it'll prohibit
3549 * the fault-around logic.
3550 */
3551 if (!pte_none(ptep_get(vmf->pte)))
3552 return ret;
3553
3554 if (vmf->address == addr)
3555 ret = VM_FAULT_NOPAGE;
3556
3557 set_pte_range(vmf, folio, page, 1, addr);
3558 folio_ref_inc(folio);
3559
3560 return ret;
3561}
3562
3563vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3564 pgoff_t start_pgoff, pgoff_t end_pgoff)
3565{
3566 struct vm_area_struct *vma = vmf->vma;
3567 struct file *file = vma->vm_file;
3568 struct address_space *mapping = file->f_mapping;
3569 pgoff_t last_pgoff = start_pgoff;
3570 unsigned long addr;
3571 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3572 struct folio *folio;
3573 vm_fault_t ret = 0;
3574 unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved;
3575
3576 rcu_read_lock();
3577 folio = next_uptodate_folio(&xas, mapping, end_pgoff);
3578 if (!folio)
3579 goto out;
3580
3581 if (filemap_map_pmd(vmf, folio, start_pgoff)) {
3582 ret = VM_FAULT_NOPAGE;
3583 goto out;
3584 }
3585
3586 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3587 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3588 if (!vmf->pte) {
3589 folio_unlock(folio);
3590 folio_put(folio);
3591 goto out;
3592 }
3593 do {
3594 unsigned long end;
3595
3596 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3597 vmf->pte += xas.xa_index - last_pgoff;
3598 last_pgoff = xas.xa_index;
3599 end = folio_next_index(folio) - 1;
3600 nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
3601
3602 if (!folio_test_large(folio))
3603 ret |= filemap_map_order0_folio(vmf,
3604 folio, addr, &mmap_miss);
3605 else
3606 ret |= filemap_map_folio_range(vmf, folio,
3607 xas.xa_index - folio->index, addr,
3608 nr_pages, &mmap_miss);
3609
3610 folio_unlock(folio);
3611 folio_put(folio);
3612 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL);
3613 pte_unmap_unlock(vmf->pte, vmf->ptl);
3614out:
3615 rcu_read_unlock();
3616
3617 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss);
3618 if (mmap_miss >= mmap_miss_saved)
3619 WRITE_ONCE(file->f_ra.mmap_miss, 0);
3620 else
3621 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss);
3622
3623 return ret;
3624}
3625EXPORT_SYMBOL(filemap_map_pages);
3626
3627vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3628{
3629 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3630 struct folio *folio = page_folio(vmf->page);
3631 vm_fault_t ret = VM_FAULT_LOCKED;
3632
3633 sb_start_pagefault(mapping->host->i_sb);
3634 file_update_time(vmf->vma->vm_file);
3635 folio_lock(folio);
3636 if (folio->mapping != mapping) {
3637 folio_unlock(folio);
3638 ret = VM_FAULT_NOPAGE;
3639 goto out;
3640 }
3641 /*
3642 * We mark the folio dirty already here so that when freeze is in
3643 * progress, we are guaranteed that writeback during freezing will
3644 * see the dirty folio and writeprotect it again.
3645 */
3646 folio_mark_dirty(folio);
3647 folio_wait_stable(folio);
3648out:
3649 sb_end_pagefault(mapping->host->i_sb);
3650 return ret;
3651}
3652
3653const struct vm_operations_struct generic_file_vm_ops = {
3654 .fault = filemap_fault,
3655 .map_pages = filemap_map_pages,
3656 .page_mkwrite = filemap_page_mkwrite,
3657};
3658
3659/* This is used for a general mmap of a disk file */
3660
3661int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3662{
3663 struct address_space *mapping = file->f_mapping;
3664
3665 if (!mapping->a_ops->read_folio)
3666 return -ENOEXEC;
3667 file_accessed(file);
3668 vma->vm_ops = &generic_file_vm_ops;
3669 return 0;
3670}
3671
3672/*
3673 * This is for filesystems which do not implement ->writepage.
3674 */
3675int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3676{
3677 if (vma_is_shared_maywrite(vma))
3678 return -EINVAL;
3679 return generic_file_mmap(file, vma);
3680}
3681#else
3682vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3683{
3684 return VM_FAULT_SIGBUS;
3685}
3686int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3687{
3688 return -ENOSYS;
3689}
3690int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3691{
3692 return -ENOSYS;
3693}
3694#endif /* CONFIG_MMU */
3695
3696EXPORT_SYMBOL(filemap_page_mkwrite);
3697EXPORT_SYMBOL(generic_file_mmap);
3698EXPORT_SYMBOL(generic_file_readonly_mmap);
3699
3700static struct folio *do_read_cache_folio(struct address_space *mapping,
3701 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3702{
3703 struct folio *folio;
3704 int err;
3705
3706 if (!filler)
3707 filler = mapping->a_ops->read_folio;
3708repeat:
3709 folio = filemap_get_folio(mapping, index);
3710 if (IS_ERR(folio)) {
3711 folio = filemap_alloc_folio(gfp, 0);
3712 if (!folio)
3713 return ERR_PTR(-ENOMEM);
3714 err = filemap_add_folio(mapping, folio, index, gfp);
3715 if (unlikely(err)) {
3716 folio_put(folio);
3717 if (err == -EEXIST)
3718 goto repeat;
3719 /* Presumably ENOMEM for xarray node */
3720 return ERR_PTR(err);
3721 }
3722
3723 goto filler;
3724 }
3725 if (folio_test_uptodate(folio))
3726 goto out;
3727
3728 if (!folio_trylock(folio)) {
3729 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3730 goto repeat;
3731 }
3732
3733 /* Folio was truncated from mapping */
3734 if (!folio->mapping) {
3735 folio_unlock(folio);
3736 folio_put(folio);
3737 goto repeat;
3738 }
3739
3740 /* Someone else locked and filled the page in a very small window */
3741 if (folio_test_uptodate(folio)) {
3742 folio_unlock(folio);
3743 goto out;
3744 }
3745
3746filler:
3747 err = filemap_read_folio(file, filler, folio);
3748 if (err) {
3749 folio_put(folio);
3750 if (err == AOP_TRUNCATED_PAGE)
3751 goto repeat;
3752 return ERR_PTR(err);
3753 }
3754
3755out:
3756 folio_mark_accessed(folio);
3757 return folio;
3758}
3759
3760/**
3761 * read_cache_folio - Read into page cache, fill it if needed.
3762 * @mapping: The address_space to read from.
3763 * @index: The index to read.
3764 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3765 * @file: Passed to filler function, may be NULL if not required.
3766 *
3767 * Read one page into the page cache. If it succeeds, the folio returned
3768 * will contain @index, but it may not be the first page of the folio.
3769 *
3770 * If the filler function returns an error, it will be returned to the
3771 * caller.
3772 *
3773 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3774 * Return: An uptodate folio on success, ERR_PTR() on failure.
3775 */
3776struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3777 filler_t filler, struct file *file)
3778{
3779 return do_read_cache_folio(mapping, index, filler, file,
3780 mapping_gfp_mask(mapping));
3781}
3782EXPORT_SYMBOL(read_cache_folio);
3783
3784/**
3785 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags.
3786 * @mapping: The address_space for the folio.
3787 * @index: The index that the allocated folio will contain.
3788 * @gfp: The page allocator flags to use if allocating.
3789 *
3790 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with
3791 * any new memory allocations done using the specified allocation flags.
3792 *
3793 * The most likely error from this function is EIO, but ENOMEM is
3794 * possible and so is EINTR. If ->read_folio returns another error,
3795 * that will be returned to the caller.
3796 *
3797 * The function expects mapping->invalidate_lock to be already held.
3798 *
3799 * Return: Uptodate folio on success, ERR_PTR() on failure.
3800 */
3801struct folio *mapping_read_folio_gfp(struct address_space *mapping,
3802 pgoff_t index, gfp_t gfp)
3803{
3804 return do_read_cache_folio(mapping, index, NULL, NULL, gfp);
3805}
3806EXPORT_SYMBOL(mapping_read_folio_gfp);
3807
3808static struct page *do_read_cache_page(struct address_space *mapping,
3809 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3810{
3811 struct folio *folio;
3812
3813 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3814 if (IS_ERR(folio))
3815 return &folio->page;
3816 return folio_file_page(folio, index);
3817}
3818
3819struct page *read_cache_page(struct address_space *mapping,
3820 pgoff_t index, filler_t *filler, struct file *file)
3821{
3822 return do_read_cache_page(mapping, index, filler, file,
3823 mapping_gfp_mask(mapping));
3824}
3825EXPORT_SYMBOL(read_cache_page);
3826
3827/**
3828 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3829 * @mapping: the page's address_space
3830 * @index: the page index
3831 * @gfp: the page allocator flags to use if allocating
3832 *
3833 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3834 * any new page allocations done using the specified allocation flags.
3835 *
3836 * If the page does not get brought uptodate, return -EIO.
3837 *
3838 * The function expects mapping->invalidate_lock to be already held.
3839 *
3840 * Return: up to date page on success, ERR_PTR() on failure.
3841 */
3842struct page *read_cache_page_gfp(struct address_space *mapping,
3843 pgoff_t index,
3844 gfp_t gfp)
3845{
3846 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3847}
3848EXPORT_SYMBOL(read_cache_page_gfp);
3849
3850/*
3851 * Warn about a page cache invalidation failure during a direct I/O write.
3852 */
3853static void dio_warn_stale_pagecache(struct file *filp)
3854{
3855 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3856 char pathname[128];
3857 char *path;
3858
3859 errseq_set(&filp->f_mapping->wb_err, -EIO);
3860 if (__ratelimit(&_rs)) {
3861 path = file_path(filp, pathname, sizeof(pathname));
3862 if (IS_ERR(path))
3863 path = "(unknown)";
3864 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3865 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3866 current->comm);
3867 }
3868}
3869
3870void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count)
3871{
3872 struct address_space *mapping = iocb->ki_filp->f_mapping;
3873
3874 if (mapping->nrpages &&
3875 invalidate_inode_pages2_range(mapping,
3876 iocb->ki_pos >> PAGE_SHIFT,
3877 (iocb->ki_pos + count - 1) >> PAGE_SHIFT))
3878 dio_warn_stale_pagecache(iocb->ki_filp);
3879}
3880
3881ssize_t
3882generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3883{
3884 struct address_space *mapping = iocb->ki_filp->f_mapping;
3885 size_t write_len = iov_iter_count(from);
3886 ssize_t written;
3887
3888 /*
3889 * If a page can not be invalidated, return 0 to fall back
3890 * to buffered write.
3891 */
3892 written = kiocb_invalidate_pages(iocb, write_len);
3893 if (written) {
3894 if (written == -EBUSY)
3895 return 0;
3896 return written;
3897 }
3898
3899 written = mapping->a_ops->direct_IO(iocb, from);
3900
3901 /*
3902 * Finally, try again to invalidate clean pages which might have been
3903 * cached by non-direct readahead, or faulted in by get_user_pages()
3904 * if the source of the write was an mmap'ed region of the file
3905 * we're writing. Either one is a pretty crazy thing to do,
3906 * so we don't support it 100%. If this invalidation
3907 * fails, tough, the write still worked...
3908 *
3909 * Most of the time we do not need this since dio_complete() will do
3910 * the invalidation for us. However there are some file systems that
3911 * do not end up with dio_complete() being called, so let's not break
3912 * them by removing it completely.
3913 *
3914 * Noticeable example is a blkdev_direct_IO().
3915 *
3916 * Skip invalidation for async writes or if mapping has no pages.
3917 */
3918 if (written > 0) {
3919 struct inode *inode = mapping->host;
3920 loff_t pos = iocb->ki_pos;
3921
3922 kiocb_invalidate_post_direct_write(iocb, written);
3923 pos += written;
3924 write_len -= written;
3925 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3926 i_size_write(inode, pos);
3927 mark_inode_dirty(inode);
3928 }
3929 iocb->ki_pos = pos;
3930 }
3931 if (written != -EIOCBQUEUED)
3932 iov_iter_revert(from, write_len - iov_iter_count(from));
3933 return written;
3934}
3935EXPORT_SYMBOL(generic_file_direct_write);
3936
3937ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
3938{
3939 struct file *file = iocb->ki_filp;
3940 loff_t pos = iocb->ki_pos;
3941 struct address_space *mapping = file->f_mapping;
3942 const struct address_space_operations *a_ops = mapping->a_ops;
3943 long status = 0;
3944 ssize_t written = 0;
3945
3946 do {
3947 struct page *page;
3948 unsigned long offset; /* Offset into pagecache page */
3949 unsigned long bytes; /* Bytes to write to page */
3950 size_t copied; /* Bytes copied from user */
3951 void *fsdata = NULL;
3952
3953 offset = (pos & (PAGE_SIZE - 1));
3954 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3955 iov_iter_count(i));
3956
3957again:
3958 /*
3959 * Bring in the user page that we will copy from _first_.
3960 * Otherwise there's a nasty deadlock on copying from the
3961 * same page as we're writing to, without it being marked
3962 * up-to-date.
3963 */
3964 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
3965 status = -EFAULT;
3966 break;
3967 }
3968
3969 if (fatal_signal_pending(current)) {
3970 status = -EINTR;
3971 break;
3972 }
3973
3974 status = a_ops->write_begin(file, mapping, pos, bytes,
3975 &page, &fsdata);
3976 if (unlikely(status < 0))
3977 break;
3978
3979 if (mapping_writably_mapped(mapping))
3980 flush_dcache_page(page);
3981
3982 copied = copy_page_from_iter_atomic(page, offset, bytes, i);
3983 flush_dcache_page(page);
3984
3985 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3986 page, fsdata);
3987 if (unlikely(status != copied)) {
3988 iov_iter_revert(i, copied - max(status, 0L));
3989 if (unlikely(status < 0))
3990 break;
3991 }
3992 cond_resched();
3993
3994 if (unlikely(status == 0)) {
3995 /*
3996 * A short copy made ->write_end() reject the
3997 * thing entirely. Might be memory poisoning
3998 * halfway through, might be a race with munmap,
3999 * might be severe memory pressure.
4000 */
4001 if (copied)
4002 bytes = copied;
4003 goto again;
4004 }
4005 pos += status;
4006 written += status;
4007
4008 balance_dirty_pages_ratelimited(mapping);
4009 } while (iov_iter_count(i));
4010
4011 if (!written)
4012 return status;
4013 iocb->ki_pos += written;
4014 return written;
4015}
4016EXPORT_SYMBOL(generic_perform_write);
4017
4018/**
4019 * __generic_file_write_iter - write data to a file
4020 * @iocb: IO state structure (file, offset, etc.)
4021 * @from: iov_iter with data to write
4022 *
4023 * This function does all the work needed for actually writing data to a
4024 * file. It does all basic checks, removes SUID from the file, updates
4025 * modification times and calls proper subroutines depending on whether we
4026 * do direct IO or a standard buffered write.
4027 *
4028 * It expects i_rwsem to be grabbed unless we work on a block device or similar
4029 * object which does not need locking at all.
4030 *
4031 * This function does *not* take care of syncing data in case of O_SYNC write.
4032 * A caller has to handle it. This is mainly due to the fact that we want to
4033 * avoid syncing under i_rwsem.
4034 *
4035 * Return:
4036 * * number of bytes written, even for truncated writes
4037 * * negative error code if no data has been written at all
4038 */
4039ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4040{
4041 struct file *file = iocb->ki_filp;
4042 struct address_space *mapping = file->f_mapping;
4043 struct inode *inode = mapping->host;
4044 ssize_t ret;
4045
4046 ret = file_remove_privs(file);
4047 if (ret)
4048 return ret;
4049
4050 ret = file_update_time(file);
4051 if (ret)
4052 return ret;
4053
4054 if (iocb->ki_flags & IOCB_DIRECT) {
4055 ret = generic_file_direct_write(iocb, from);
4056 /*
4057 * If the write stopped short of completing, fall back to
4058 * buffered writes. Some filesystems do this for writes to
4059 * holes, for example. For DAX files, a buffered write will
4060 * not succeed (even if it did, DAX does not handle dirty
4061 * page-cache pages correctly).
4062 */
4063 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode))
4064 return ret;
4065 return direct_write_fallback(iocb, from, ret,
4066 generic_perform_write(iocb, from));
4067 }
4068
4069 return generic_perform_write(iocb, from);
4070}
4071EXPORT_SYMBOL(__generic_file_write_iter);
4072
4073/**
4074 * generic_file_write_iter - write data to a file
4075 * @iocb: IO state structure
4076 * @from: iov_iter with data to write
4077 *
4078 * This is a wrapper around __generic_file_write_iter() to be used by most
4079 * filesystems. It takes care of syncing the file in case of O_SYNC file
4080 * and acquires i_rwsem as needed.
4081 * Return:
4082 * * negative error code if no data has been written at all of
4083 * vfs_fsync_range() failed for a synchronous write
4084 * * number of bytes written, even for truncated writes
4085 */
4086ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4087{
4088 struct file *file = iocb->ki_filp;
4089 struct inode *inode = file->f_mapping->host;
4090 ssize_t ret;
4091
4092 inode_lock(inode);
4093 ret = generic_write_checks(iocb, from);
4094 if (ret > 0)
4095 ret = __generic_file_write_iter(iocb, from);
4096 inode_unlock(inode);
4097
4098 if (ret > 0)
4099 ret = generic_write_sync(iocb, ret);
4100 return ret;
4101}
4102EXPORT_SYMBOL(generic_file_write_iter);
4103
4104/**
4105 * filemap_release_folio() - Release fs-specific metadata on a folio.
4106 * @folio: The folio which the kernel is trying to free.
4107 * @gfp: Memory allocation flags (and I/O mode).
4108 *
4109 * The address_space is trying to release any data attached to a folio
4110 * (presumably at folio->private).
4111 *
4112 * This will also be called if the private_2 flag is set on a page,
4113 * indicating that the folio has other metadata associated with it.
4114 *
4115 * The @gfp argument specifies whether I/O may be performed to release
4116 * this page (__GFP_IO), and whether the call may block
4117 * (__GFP_RECLAIM & __GFP_FS).
4118 *
4119 * Return: %true if the release was successful, otherwise %false.
4120 */
4121bool filemap_release_folio(struct folio *folio, gfp_t gfp)
4122{
4123 struct address_space * const mapping = folio->mapping;
4124
4125 BUG_ON(!folio_test_locked(folio));
4126 if (!folio_needs_release(folio))
4127 return true;
4128 if (folio_test_writeback(folio))
4129 return false;
4130
4131 if (mapping && mapping->a_ops->release_folio)
4132 return mapping->a_ops->release_folio(folio, gfp);
4133 return try_to_free_buffers(folio);
4134}
4135EXPORT_SYMBOL(filemap_release_folio);
4136
4137#ifdef CONFIG_CACHESTAT_SYSCALL
4138/**
4139 * filemap_cachestat() - compute the page cache statistics of a mapping
4140 * @mapping: The mapping to compute the statistics for.
4141 * @first_index: The starting page cache index.
4142 * @last_index: The final page index (inclusive).
4143 * @cs: the cachestat struct to write the result to.
4144 *
4145 * This will query the page cache statistics of a mapping in the
4146 * page range of [first_index, last_index] (inclusive). The statistics
4147 * queried include: number of dirty pages, number of pages marked for
4148 * writeback, and the number of (recently) evicted pages.
4149 */
4150static void filemap_cachestat(struct address_space *mapping,
4151 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs)
4152{
4153 XA_STATE(xas, &mapping->i_pages, first_index);
4154 struct folio *folio;
4155
4156 rcu_read_lock();
4157 xas_for_each(&xas, folio, last_index) {
4158 int order;
4159 unsigned long nr_pages;
4160 pgoff_t folio_first_index, folio_last_index;
4161
4162 /*
4163 * Don't deref the folio. It is not pinned, and might
4164 * get freed (and reused) underneath us.
4165 *
4166 * We *could* pin it, but that would be expensive for
4167 * what should be a fast and lightweight syscall.
4168 *
4169 * Instead, derive all information of interest from
4170 * the rcu-protected xarray.
4171 */
4172
4173 if (xas_retry(&xas, folio))
4174 continue;
4175
4176 order = xa_get_order(xas.xa, xas.xa_index);
4177 nr_pages = 1 << order;
4178 folio_first_index = round_down(xas.xa_index, 1 << order);
4179 folio_last_index = folio_first_index + nr_pages - 1;
4180
4181 /* Folios might straddle the range boundaries, only count covered pages */
4182 if (folio_first_index < first_index)
4183 nr_pages -= first_index - folio_first_index;
4184
4185 if (folio_last_index > last_index)
4186 nr_pages -= folio_last_index - last_index;
4187
4188 if (xa_is_value(folio)) {
4189 /* page is evicted */
4190 void *shadow = (void *)folio;
4191 bool workingset; /* not used */
4192
4193 cs->nr_evicted += nr_pages;
4194
4195#ifdef CONFIG_SWAP /* implies CONFIG_MMU */
4196 if (shmem_mapping(mapping)) {
4197 /* shmem file - in swap cache */
4198 swp_entry_t swp = radix_to_swp_entry(folio);
4199
4200 /* swapin error results in poisoned entry */
4201 if (non_swap_entry(swp))
4202 goto resched;
4203
4204 /*
4205 * Getting a swap entry from the shmem
4206 * inode means we beat
4207 * shmem_unuse(). rcu_read_lock()
4208 * ensures swapoff waits for us before
4209 * freeing the swapper space. However,
4210 * we can race with swapping and
4211 * invalidation, so there might not be
4212 * a shadow in the swapcache (yet).
4213 */
4214 shadow = get_shadow_from_swap_cache(swp);
4215 if (!shadow)
4216 goto resched;
4217 }
4218#endif
4219 if (workingset_test_recent(shadow, true, &workingset))
4220 cs->nr_recently_evicted += nr_pages;
4221
4222 goto resched;
4223 }
4224
4225 /* page is in cache */
4226 cs->nr_cache += nr_pages;
4227
4228 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY))
4229 cs->nr_dirty += nr_pages;
4230
4231 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK))
4232 cs->nr_writeback += nr_pages;
4233
4234resched:
4235 if (need_resched()) {
4236 xas_pause(&xas);
4237 cond_resched_rcu();
4238 }
4239 }
4240 rcu_read_unlock();
4241}
4242
4243/*
4244 * The cachestat(2) system call.
4245 *
4246 * cachestat() returns the page cache statistics of a file in the
4247 * bytes range specified by `off` and `len`: number of cached pages,
4248 * number of dirty pages, number of pages marked for writeback,
4249 * number of evicted pages, and number of recently evicted pages.
4250 *
4251 * An evicted page is a page that is previously in the page cache
4252 * but has been evicted since. A page is recently evicted if its last
4253 * eviction was recent enough that its reentry to the cache would
4254 * indicate that it is actively being used by the system, and that
4255 * there is memory pressure on the system.
4256 *
4257 * `off` and `len` must be non-negative integers. If `len` > 0,
4258 * the queried range is [`off`, `off` + `len`]. If `len` == 0,
4259 * we will query in the range from `off` to the end of the file.
4260 *
4261 * The `flags` argument is unused for now, but is included for future
4262 * extensibility. User should pass 0 (i.e no flag specified).
4263 *
4264 * Currently, hugetlbfs is not supported.
4265 *
4266 * Because the status of a page can change after cachestat() checks it
4267 * but before it returns to the application, the returned values may
4268 * contain stale information.
4269 *
4270 * return values:
4271 * zero - success
4272 * -EFAULT - cstat or cstat_range points to an illegal address
4273 * -EINVAL - invalid flags
4274 * -EBADF - invalid file descriptor
4275 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file
4276 */
4277SYSCALL_DEFINE4(cachestat, unsigned int, fd,
4278 struct cachestat_range __user *, cstat_range,
4279 struct cachestat __user *, cstat, unsigned int, flags)
4280{
4281 struct fd f = fdget(fd);
4282 struct address_space *mapping;
4283 struct cachestat_range csr;
4284 struct cachestat cs;
4285 pgoff_t first_index, last_index;
4286
4287 if (!f.file)
4288 return -EBADF;
4289
4290 if (copy_from_user(&csr, cstat_range,
4291 sizeof(struct cachestat_range))) {
4292 fdput(f);
4293 return -EFAULT;
4294 }
4295
4296 /* hugetlbfs is not supported */
4297 if (is_file_hugepages(f.file)) {
4298 fdput(f);
4299 return -EOPNOTSUPP;
4300 }
4301
4302 if (flags != 0) {
4303 fdput(f);
4304 return -EINVAL;
4305 }
4306
4307 first_index = csr.off >> PAGE_SHIFT;
4308 last_index =
4309 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT;
4310 memset(&cs, 0, sizeof(struct cachestat));
4311 mapping = f.file->f_mapping;
4312 filemap_cachestat(mapping, first_index, last_index, &cs);
4313 fdput(f);
4314
4315 if (copy_to_user(cstat, &cs, sizeof(struct cachestat)))
4316 return -EFAULT;
4317
4318 return 0;
4319}
4320#endif /* CONFIG_CACHESTAT_SYSCALL */