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