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/syscalls.h>
26#include <linux/mman.h>
27#include <linux/pagemap.h>
28#include <linux/file.h>
29#include <linux/uio.h>
30#include <linux/error-injection.h>
31#include <linux/hash.h>
32#include <linux/writeback.h>
33#include <linux/backing-dev.h>
34#include <linux/pagevec.h>
35#include <linux/security.h>
36#include <linux/cpuset.h>
37#include <linux/hugetlb.h>
38#include <linux/memcontrol.h>
39#include <linux/shmem_fs.h>
40#include <linux/rmap.h>
41#include <linux/delayacct.h>
42#include <linux/psi.h>
43#include <linux/ramfs.h>
44#include <linux/page_idle.h>
45#include <linux/migrate.h>
46#include <linux/pipe_fs_i.h>
47#include <linux/splice.h>
48#include <linux/rcupdate_wait.h>
49#include <linux/sched/mm.h>
50#include <asm/pgalloc.h>
51#include <asm/tlbflush.h>
52#include "internal.h"
53
54#define CREATE_TRACE_POINTS
55#include <trace/events/filemap.h>
56
57/*
58 * FIXME: remove all knowledge of the buffer layer from the core VM
59 */
60#include <linux/buffer_head.h> /* for try_to_free_buffers */
61
62#include <asm/mman.h>
63
64#include "swap.h"
65
66/*
67 * Shared mappings implemented 30.11.1994. It's not fully working yet,
68 * though.
69 *
70 * Shared mappings now work. 15.8.1995 Bruno.
71 *
72 * finished 'unifying' the page and buffer cache and SMP-threaded the
73 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
74 *
75 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
76 */
77
78/*
79 * Lock ordering:
80 *
81 * ->i_mmap_rwsem (truncate_pagecache)
82 * ->private_lock (__free_pte->block_dirty_folio)
83 * ->swap_lock (exclusive_swap_page, others)
84 * ->i_pages lock
85 *
86 * ->i_rwsem
87 * ->invalidate_lock (acquired by fs in truncate path)
88 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
89 *
90 * ->mmap_lock
91 * ->i_mmap_rwsem
92 * ->page_table_lock or pte_lock (various, mainly in memory.c)
93 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
94 *
95 * ->mmap_lock
96 * ->invalidate_lock (filemap_fault)
97 * ->lock_page (filemap_fault, access_process_vm)
98 *
99 * ->i_rwsem (generic_perform_write)
100 * ->mmap_lock (fault_in_readable->do_page_fault)
101 *
102 * bdi->wb.list_lock
103 * sb_lock (fs/fs-writeback.c)
104 * ->i_pages lock (__sync_single_inode)
105 *
106 * ->i_mmap_rwsem
107 * ->anon_vma.lock (vma_merge)
108 *
109 * ->anon_vma.lock
110 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
111 *
112 * ->page_table_lock or pte_lock
113 * ->swap_lock (try_to_unmap_one)
114 * ->private_lock (try_to_unmap_one)
115 * ->i_pages lock (try_to_unmap_one)
116 * ->lruvec->lru_lock (follow_page_mask->mark_page_accessed)
117 * ->lruvec->lru_lock (check_pte_range->folio_isolate_lru)
118 * ->private_lock (folio_remove_rmap_pte->set_page_dirty)
119 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty)
120 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty)
121 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty)
122 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
123 * ->inode->i_lock (zap_pte_range->set_page_dirty)
124 * ->private_lock (zap_pte_range->block_dirty_folio)
125 */
126
127static void page_cache_delete(struct address_space *mapping,
128 struct folio *folio, void *shadow)
129{
130 XA_STATE(xas, &mapping->i_pages, folio->index);
131 long nr = 1;
132
133 mapping_set_update(&xas, mapping);
134
135 xas_set_order(&xas, folio->index, folio_order(folio));
136 nr = folio_nr_pages(folio);
137
138 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
139
140 xas_store(&xas, shadow);
141 xas_init_marks(&xas);
142
143 folio->mapping = NULL;
144 /* Leave page->index set: truncation lookup relies upon it */
145 mapping->nrpages -= nr;
146}
147
148static void filemap_unaccount_folio(struct address_space *mapping,
149 struct folio *folio)
150{
151 long nr;
152
153 VM_BUG_ON_FOLIO(folio_mapped(folio), folio);
154 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) {
155 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
156 current->comm, folio_pfn(folio));
157 dump_page(&folio->page, "still mapped when deleted");
158 dump_stack();
159 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
160
161 if (mapping_exiting(mapping) && !folio_test_large(folio)) {
162 int mapcount = folio_mapcount(folio);
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 atomic_set(&folio->_mapcount, -1);
172 folio_ref_sub(folio, mapcount);
173 }
174 }
175 }
176
177 /* hugetlb folios do not participate in page cache accounting. */
178 if (folio_test_hugetlb(folio))
179 return;
180
181 nr = folio_nr_pages(folio);
182
183 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
184 if (folio_test_swapbacked(folio)) {
185 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
186 if (folio_test_pmd_mappable(folio))
187 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr);
188 } else if (folio_test_pmd_mappable(folio)) {
189 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr);
190 filemap_nr_thps_dec(mapping);
191 }
192
193 /*
194 * At this point folio must be either written or cleaned by
195 * truncate. Dirty folio here signals a bug and loss of
196 * unwritten data - on ordinary filesystems.
197 *
198 * But it's harmless on in-memory filesystems like tmpfs; and can
199 * occur when a driver which did get_user_pages() sets page dirty
200 * before putting it, while the inode is being finally evicted.
201 *
202 * Below fixes dirty accounting after removing the folio entirely
203 * but leaves the dirty flag set: it has no effect for truncated
204 * folio and anyway will be cleared before returning folio to
205 * buddy allocator.
206 */
207 if (WARN_ON_ONCE(folio_test_dirty(folio) &&
208 mapping_can_writeback(mapping)))
209 folio_account_cleaned(folio, inode_to_wb(mapping->host));
210}
211
212/*
213 * Delete a page from the page cache and free it. Caller has to make
214 * sure the page is locked and that nobody else uses it - or that usage
215 * is safe. The caller must hold the i_pages lock.
216 */
217void __filemap_remove_folio(struct folio *folio, void *shadow)
218{
219 struct address_space *mapping = folio->mapping;
220
221 trace_mm_filemap_delete_from_page_cache(folio);
222 filemap_unaccount_folio(mapping, folio);
223 page_cache_delete(mapping, folio, shadow);
224}
225
226void filemap_free_folio(struct address_space *mapping, struct folio *folio)
227{
228 void (*free_folio)(struct folio *);
229 int refs = 1;
230
231 free_folio = mapping->a_ops->free_folio;
232 if (free_folio)
233 free_folio(folio);
234
235 if (folio_test_large(folio))
236 refs = folio_nr_pages(folio);
237 folio_put_refs(folio, refs);
238}
239
240/**
241 * filemap_remove_folio - Remove folio from page cache.
242 * @folio: The folio.
243 *
244 * This must be called only on folios that are locked and have been
245 * verified to be in the page cache. It will never put the folio into
246 * the free list because the caller has a reference on the page.
247 */
248void filemap_remove_folio(struct folio *folio)
249{
250 struct address_space *mapping = folio->mapping;
251
252 BUG_ON(!folio_test_locked(folio));
253 spin_lock(&mapping->host->i_lock);
254 xa_lock_irq(&mapping->i_pages);
255 __filemap_remove_folio(folio, NULL);
256 xa_unlock_irq(&mapping->i_pages);
257 if (mapping_shrinkable(mapping))
258 inode_add_lru(mapping->host);
259 spin_unlock(&mapping->host->i_lock);
260
261 filemap_free_folio(mapping, folio);
262}
263
264/*
265 * page_cache_delete_batch - delete several folios from page cache
266 * @mapping: the mapping to which folios belong
267 * @fbatch: batch of folios to delete
268 *
269 * The function walks over mapping->i_pages and removes folios passed in
270 * @fbatch from the mapping. The function expects @fbatch to be sorted
271 * by page index and is optimised for it to be dense.
272 * It tolerates holes in @fbatch (mapping entries at those indices are not
273 * modified).
274 *
275 * The function expects the i_pages lock to be held.
276 */
277static void page_cache_delete_batch(struct address_space *mapping,
278 struct folio_batch *fbatch)
279{
280 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index);
281 long total_pages = 0;
282 int i = 0;
283 struct folio *folio;
284
285 mapping_set_update(&xas, mapping);
286 xas_for_each(&xas, folio, ULONG_MAX) {
287 if (i >= folio_batch_count(fbatch))
288 break;
289
290 /* A swap/dax/shadow entry got inserted? Skip it. */
291 if (xa_is_value(folio))
292 continue;
293 /*
294 * A page got inserted in our range? Skip it. We have our
295 * pages locked so they are protected from being removed.
296 * If we see a page whose index is higher than ours, it
297 * means our page has been removed, which shouldn't be
298 * possible because we're holding the PageLock.
299 */
300 if (folio != fbatch->folios[i]) {
301 VM_BUG_ON_FOLIO(folio->index >
302 fbatch->folios[i]->index, folio);
303 continue;
304 }
305
306 WARN_ON_ONCE(!folio_test_locked(folio));
307
308 folio->mapping = NULL;
309 /* Leave folio->index set: truncation lookup relies on it */
310
311 i++;
312 xas_store(&xas, NULL);
313 total_pages += folio_nr_pages(folio);
314 }
315 mapping->nrpages -= total_pages;
316}
317
318void delete_from_page_cache_batch(struct address_space *mapping,
319 struct folio_batch *fbatch)
320{
321 int i;
322
323 if (!folio_batch_count(fbatch))
324 return;
325
326 spin_lock(&mapping->host->i_lock);
327 xa_lock_irq(&mapping->i_pages);
328 for (i = 0; i < folio_batch_count(fbatch); i++) {
329 struct folio *folio = fbatch->folios[i];
330
331 trace_mm_filemap_delete_from_page_cache(folio);
332 filemap_unaccount_folio(mapping, folio);
333 }
334 page_cache_delete_batch(mapping, fbatch);
335 xa_unlock_irq(&mapping->i_pages);
336 if (mapping_shrinkable(mapping))
337 inode_add_lru(mapping->host);
338 spin_unlock(&mapping->host->i_lock);
339
340 for (i = 0; i < folio_batch_count(fbatch); i++)
341 filemap_free_folio(mapping, fbatch->folios[i]);
342}
343
344int filemap_check_errors(struct address_space *mapping)
345{
346 int ret = 0;
347 /* Check for outstanding write errors */
348 if (test_bit(AS_ENOSPC, &mapping->flags) &&
349 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
350 ret = -ENOSPC;
351 if (test_bit(AS_EIO, &mapping->flags) &&
352 test_and_clear_bit(AS_EIO, &mapping->flags))
353 ret = -EIO;
354 return ret;
355}
356EXPORT_SYMBOL(filemap_check_errors);
357
358static int filemap_check_and_keep_errors(struct address_space *mapping)
359{
360 /* Check for outstanding write errors */
361 if (test_bit(AS_EIO, &mapping->flags))
362 return -EIO;
363 if (test_bit(AS_ENOSPC, &mapping->flags))
364 return -ENOSPC;
365 return 0;
366}
367
368/**
369 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range
370 * @mapping: address space structure to write
371 * @wbc: the writeback_control controlling the writeout
372 *
373 * Call writepages on the mapping using the provided wbc to control the
374 * writeout.
375 *
376 * Return: %0 on success, negative error code otherwise.
377 */
378int filemap_fdatawrite_wbc(struct address_space *mapping,
379 struct writeback_control *wbc)
380{
381 int ret;
382
383 if (!mapping_can_writeback(mapping) ||
384 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
385 return 0;
386
387 wbc_attach_fdatawrite_inode(wbc, mapping->host);
388 ret = do_writepages(mapping, wbc);
389 wbc_detach_inode(wbc);
390 return ret;
391}
392EXPORT_SYMBOL(filemap_fdatawrite_wbc);
393
394/**
395 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
396 * @mapping: address space structure to write
397 * @start: offset in bytes where the range starts
398 * @end: offset in bytes where the range ends (inclusive)
399 * @sync_mode: enable synchronous operation
400 *
401 * Start writeback against all of a mapping's dirty pages that lie
402 * within the byte offsets <start, end> inclusive.
403 *
404 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
405 * opposed to a regular memory cleansing writeback. The difference between
406 * these two operations is that if a dirty page/buffer is encountered, it must
407 * be waited upon, and not just skipped over.
408 *
409 * Return: %0 on success, negative error code otherwise.
410 */
411int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
412 loff_t end, int sync_mode)
413{
414 struct writeback_control wbc = {
415 .sync_mode = sync_mode,
416 .nr_to_write = LONG_MAX,
417 .range_start = start,
418 .range_end = end,
419 };
420
421 return filemap_fdatawrite_wbc(mapping, &wbc);
422}
423
424static inline int __filemap_fdatawrite(struct address_space *mapping,
425 int sync_mode)
426{
427 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
428}
429
430int filemap_fdatawrite(struct address_space *mapping)
431{
432 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
433}
434EXPORT_SYMBOL(filemap_fdatawrite);
435
436int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
437 loff_t end)
438{
439 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
440}
441EXPORT_SYMBOL(filemap_fdatawrite_range);
442
443/**
444 * filemap_flush - mostly a non-blocking flush
445 * @mapping: target address_space
446 *
447 * This is a mostly non-blocking flush. Not suitable for data-integrity
448 * purposes - I/O may not be started against all dirty pages.
449 *
450 * Return: %0 on success, negative error code otherwise.
451 */
452int filemap_flush(struct address_space *mapping)
453{
454 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
455}
456EXPORT_SYMBOL(filemap_flush);
457
458/**
459 * filemap_range_has_page - check if a page exists in range.
460 * @mapping: address space within which to check
461 * @start_byte: offset in bytes where the range starts
462 * @end_byte: offset in bytes where the range ends (inclusive)
463 *
464 * Find at least one page in the range supplied, usually used to check if
465 * direct writing in this range will trigger a writeback.
466 *
467 * Return: %true if at least one page exists in the specified range,
468 * %false otherwise.
469 */
470bool filemap_range_has_page(struct address_space *mapping,
471 loff_t start_byte, loff_t end_byte)
472{
473 struct folio *folio;
474 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
475 pgoff_t max = end_byte >> PAGE_SHIFT;
476
477 if (end_byte < start_byte)
478 return false;
479
480 rcu_read_lock();
481 for (;;) {
482 folio = xas_find(&xas, max);
483 if (xas_retry(&xas, folio))
484 continue;
485 /* Shadow entries don't count */
486 if (xa_is_value(folio))
487 continue;
488 /*
489 * We don't need to try to pin this page; we're about to
490 * release the RCU lock anyway. It is enough to know that
491 * there was a page here recently.
492 */
493 break;
494 }
495 rcu_read_unlock();
496
497 return folio != NULL;
498}
499EXPORT_SYMBOL(filemap_range_has_page);
500
501static void __filemap_fdatawait_range(struct address_space *mapping,
502 loff_t start_byte, loff_t end_byte)
503{
504 pgoff_t index = start_byte >> PAGE_SHIFT;
505 pgoff_t end = end_byte >> PAGE_SHIFT;
506 struct folio_batch fbatch;
507 unsigned nr_folios;
508
509 folio_batch_init(&fbatch);
510
511 while (index <= end) {
512 unsigned i;
513
514 nr_folios = filemap_get_folios_tag(mapping, &index, end,
515 PAGECACHE_TAG_WRITEBACK, &fbatch);
516
517 if (!nr_folios)
518 break;
519
520 for (i = 0; i < nr_folios; i++) {
521 struct folio *folio = fbatch.folios[i];
522
523 folio_wait_writeback(folio);
524 }
525 folio_batch_release(&fbatch);
526 cond_resched();
527 }
528}
529
530/**
531 * filemap_fdatawait_range - wait for writeback to complete
532 * @mapping: address space structure to wait for
533 * @start_byte: offset in bytes where the range starts
534 * @end_byte: offset in bytes where the range ends (inclusive)
535 *
536 * Walk the list of under-writeback pages of the given address space
537 * in the given range and wait for all of them. Check error status of
538 * the address space and return it.
539 *
540 * Since the error status of the address space is cleared by this function,
541 * callers are responsible for checking the return value and handling and/or
542 * reporting the error.
543 *
544 * Return: error status of the address space.
545 */
546int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
547 loff_t end_byte)
548{
549 __filemap_fdatawait_range(mapping, start_byte, end_byte);
550 return filemap_check_errors(mapping);
551}
552EXPORT_SYMBOL(filemap_fdatawait_range);
553
554/**
555 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
556 * @mapping: address space structure to wait for
557 * @start_byte: offset in bytes where the range starts
558 * @end_byte: offset in bytes where the range ends (inclusive)
559 *
560 * Walk the list of under-writeback pages of the given address space in the
561 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
562 * this function does not clear error status of the address space.
563 *
564 * Use this function if callers don't handle errors themselves. Expected
565 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
566 * fsfreeze(8)
567 */
568int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
569 loff_t start_byte, loff_t end_byte)
570{
571 __filemap_fdatawait_range(mapping, start_byte, end_byte);
572 return filemap_check_and_keep_errors(mapping);
573}
574EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
575
576/**
577 * file_fdatawait_range - wait for writeback to complete
578 * @file: file pointing to address space structure to wait for
579 * @start_byte: offset in bytes where the range starts
580 * @end_byte: offset in bytes where the range ends (inclusive)
581 *
582 * Walk the list of under-writeback pages of the address space that file
583 * refers to, in the given range and wait for all of them. Check error
584 * status of the address space vs. the file->f_wb_err cursor and return it.
585 *
586 * Since the error status of the file is advanced by this function,
587 * callers are responsible for checking the return value and handling and/or
588 * reporting the error.
589 *
590 * Return: error status of the address space vs. the file->f_wb_err cursor.
591 */
592int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
593{
594 struct address_space *mapping = file->f_mapping;
595
596 __filemap_fdatawait_range(mapping, start_byte, end_byte);
597 return file_check_and_advance_wb_err(file);
598}
599EXPORT_SYMBOL(file_fdatawait_range);
600
601/**
602 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
603 * @mapping: address space structure to wait for
604 *
605 * Walk the list of under-writeback pages of the given address space
606 * and wait for all of them. Unlike filemap_fdatawait(), this function
607 * does not clear error status of the address space.
608 *
609 * Use this function if callers don't handle errors themselves. Expected
610 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
611 * fsfreeze(8)
612 *
613 * Return: error status of the address space.
614 */
615int filemap_fdatawait_keep_errors(struct address_space *mapping)
616{
617 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
618 return filemap_check_and_keep_errors(mapping);
619}
620EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
621
622/* Returns true if writeback might be needed or already in progress. */
623static bool mapping_needs_writeback(struct address_space *mapping)
624{
625 return mapping->nrpages;
626}
627
628bool filemap_range_has_writeback(struct address_space *mapping,
629 loff_t start_byte, loff_t end_byte)
630{
631 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
632 pgoff_t max = end_byte >> PAGE_SHIFT;
633 struct folio *folio;
634
635 if (end_byte < start_byte)
636 return false;
637
638 rcu_read_lock();
639 xas_for_each(&xas, folio, max) {
640 if (xas_retry(&xas, folio))
641 continue;
642 if (xa_is_value(folio))
643 continue;
644 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
645 folio_test_writeback(folio))
646 break;
647 }
648 rcu_read_unlock();
649 return folio != NULL;
650}
651EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
652
653/**
654 * filemap_write_and_wait_range - write out & wait on a file range
655 * @mapping: the address_space for the pages
656 * @lstart: offset in bytes where the range starts
657 * @lend: offset in bytes where the range ends (inclusive)
658 *
659 * Write out and wait upon file offsets lstart->lend, inclusive.
660 *
661 * Note that @lend is inclusive (describes the last byte to be written) so
662 * that this function can be used to write to the very end-of-file (end = -1).
663 *
664 * Return: error status of the address space.
665 */
666int filemap_write_and_wait_range(struct address_space *mapping,
667 loff_t lstart, loff_t lend)
668{
669 int err = 0, err2;
670
671 if (lend < lstart)
672 return 0;
673
674 if (mapping_needs_writeback(mapping)) {
675 err = __filemap_fdatawrite_range(mapping, lstart, lend,
676 WB_SYNC_ALL);
677 /*
678 * Even if the above returned error, the pages may be
679 * written partially (e.g. -ENOSPC), so we wait for it.
680 * But the -EIO is special case, it may indicate the worst
681 * thing (e.g. bug) happened, so we avoid waiting for it.
682 */
683 if (err != -EIO)
684 __filemap_fdatawait_range(mapping, lstart, lend);
685 }
686 err2 = filemap_check_errors(mapping);
687 if (!err)
688 err = err2;
689 return err;
690}
691EXPORT_SYMBOL(filemap_write_and_wait_range);
692
693void __filemap_set_wb_err(struct address_space *mapping, int err)
694{
695 errseq_t eseq = errseq_set(&mapping->wb_err, err);
696
697 trace_filemap_set_wb_err(mapping, eseq);
698}
699EXPORT_SYMBOL(__filemap_set_wb_err);
700
701/**
702 * file_check_and_advance_wb_err - report wb error (if any) that was previously
703 * and advance wb_err to current one
704 * @file: struct file on which the error is being reported
705 *
706 * When userland calls fsync (or something like nfsd does the equivalent), we
707 * want to report any writeback errors that occurred since the last fsync (or
708 * since the file was opened if there haven't been any).
709 *
710 * Grab the wb_err from the mapping. If it matches what we have in the file,
711 * then just quickly return 0. The file is all caught up.
712 *
713 * If it doesn't match, then take the mapping value, set the "seen" flag in
714 * it and try to swap it into place. If it works, or another task beat us
715 * to it with the new value, then update the f_wb_err and return the error
716 * portion. The error at this point must be reported via proper channels
717 * (a'la fsync, or NFS COMMIT operation, etc.).
718 *
719 * While we handle mapping->wb_err with atomic operations, the f_wb_err
720 * value is protected by the f_lock since we must ensure that it reflects
721 * the latest value swapped in for this file descriptor.
722 *
723 * Return: %0 on success, negative error code otherwise.
724 */
725int file_check_and_advance_wb_err(struct file *file)
726{
727 int err = 0;
728 errseq_t old = READ_ONCE(file->f_wb_err);
729 struct address_space *mapping = file->f_mapping;
730
731 /* Locklessly handle the common case where nothing has changed */
732 if (errseq_check(&mapping->wb_err, old)) {
733 /* Something changed, must use slow path */
734 spin_lock(&file->f_lock);
735 old = file->f_wb_err;
736 err = errseq_check_and_advance(&mapping->wb_err,
737 &file->f_wb_err);
738 trace_file_check_and_advance_wb_err(file, old);
739 spin_unlock(&file->f_lock);
740 }
741
742 /*
743 * We're mostly using this function as a drop in replacement for
744 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
745 * that the legacy code would have had on these flags.
746 */
747 clear_bit(AS_EIO, &mapping->flags);
748 clear_bit(AS_ENOSPC, &mapping->flags);
749 return err;
750}
751EXPORT_SYMBOL(file_check_and_advance_wb_err);
752
753/**
754 * file_write_and_wait_range - write out & wait on a file range
755 * @file: file pointing to address_space with pages
756 * @lstart: offset in bytes where the range starts
757 * @lend: offset in bytes where the range ends (inclusive)
758 *
759 * Write out and wait upon file offsets lstart->lend, inclusive.
760 *
761 * Note that @lend is inclusive (describes the last byte to be written) so
762 * that this function can be used to write to the very end-of-file (end = -1).
763 *
764 * After writing out and waiting on the data, we check and advance the
765 * f_wb_err cursor to the latest value, and return any errors detected there.
766 *
767 * Return: %0 on success, negative error code otherwise.
768 */
769int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
770{
771 int err = 0, err2;
772 struct address_space *mapping = file->f_mapping;
773
774 if (lend < lstart)
775 return 0;
776
777 if (mapping_needs_writeback(mapping)) {
778 err = __filemap_fdatawrite_range(mapping, lstart, lend,
779 WB_SYNC_ALL);
780 /* See comment of filemap_write_and_wait() */
781 if (err != -EIO)
782 __filemap_fdatawait_range(mapping, lstart, lend);
783 }
784 err2 = file_check_and_advance_wb_err(file);
785 if (!err)
786 err = err2;
787 return err;
788}
789EXPORT_SYMBOL(file_write_and_wait_range);
790
791/**
792 * replace_page_cache_folio - replace a pagecache folio with a new one
793 * @old: folio to be replaced
794 * @new: folio to replace with
795 *
796 * This function replaces a folio in the pagecache with a new one. On
797 * success it acquires the pagecache reference for the new folio and
798 * drops it for the old folio. Both the old and new folios must be
799 * locked. This function does not add the new folio to the LRU, the
800 * caller must do that.
801 *
802 * The remove + add is atomic. This function cannot fail.
803 */
804void replace_page_cache_folio(struct folio *old, struct folio *new)
805{
806 struct address_space *mapping = old->mapping;
807 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
808 pgoff_t offset = old->index;
809 XA_STATE(xas, &mapping->i_pages, offset);
810
811 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
812 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
813 VM_BUG_ON_FOLIO(new->mapping, new);
814
815 folio_get(new);
816 new->mapping = mapping;
817 new->index = offset;
818
819 mem_cgroup_replace_folio(old, new);
820
821 xas_lock_irq(&xas);
822 xas_store(&xas, new);
823
824 old->mapping = NULL;
825 /* hugetlb pages do not participate in page cache accounting. */
826 if (!folio_test_hugetlb(old))
827 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
828 if (!folio_test_hugetlb(new))
829 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
830 if (folio_test_swapbacked(old))
831 __lruvec_stat_sub_folio(old, NR_SHMEM);
832 if (folio_test_swapbacked(new))
833 __lruvec_stat_add_folio(new, NR_SHMEM);
834 xas_unlock_irq(&xas);
835 if (free_folio)
836 free_folio(old);
837 folio_put(old);
838}
839EXPORT_SYMBOL_GPL(replace_page_cache_folio);
840
841noinline int __filemap_add_folio(struct address_space *mapping,
842 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
843{
844 XA_STATE(xas, &mapping->i_pages, index);
845 void *alloced_shadow = NULL;
846 int alloced_order = 0;
847 bool huge;
848 long nr;
849
850 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
851 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
852 VM_BUG_ON_FOLIO(folio_order(folio) < mapping_min_folio_order(mapping),
853 folio);
854 mapping_set_update(&xas, mapping);
855
856 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
857 xas_set_order(&xas, index, folio_order(folio));
858 huge = folio_test_hugetlb(folio);
859 nr = folio_nr_pages(folio);
860
861 gfp &= GFP_RECLAIM_MASK;
862 folio_ref_add(folio, nr);
863 folio->mapping = mapping;
864 folio->index = xas.xa_index;
865
866 for (;;) {
867 int order = -1, split_order = 0;
868 void *entry, *old = NULL;
869
870 xas_lock_irq(&xas);
871 xas_for_each_conflict(&xas, entry) {
872 old = entry;
873 if (!xa_is_value(entry)) {
874 xas_set_err(&xas, -EEXIST);
875 goto unlock;
876 }
877 /*
878 * If a larger entry exists,
879 * it will be the first and only entry iterated.
880 */
881 if (order == -1)
882 order = xas_get_order(&xas);
883 }
884
885 /* entry may have changed before we re-acquire the lock */
886 if (alloced_order && (old != alloced_shadow || order != alloced_order)) {
887 xas_destroy(&xas);
888 alloced_order = 0;
889 }
890
891 if (old) {
892 if (order > 0 && order > folio_order(folio)) {
893 /* How to handle large swap entries? */
894 BUG_ON(shmem_mapping(mapping));
895 if (!alloced_order) {
896 split_order = order;
897 goto unlock;
898 }
899 xas_split(&xas, old, order);
900 xas_reset(&xas);
901 }
902 if (shadowp)
903 *shadowp = old;
904 }
905
906 xas_store(&xas, folio);
907 if (xas_error(&xas))
908 goto unlock;
909
910 mapping->nrpages += nr;
911
912 /* hugetlb pages do not participate in page cache accounting */
913 if (!huge) {
914 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
915 if (folio_test_pmd_mappable(folio))
916 __lruvec_stat_mod_folio(folio,
917 NR_FILE_THPS, nr);
918 }
919
920unlock:
921 xas_unlock_irq(&xas);
922
923 /* split needed, alloc here and retry. */
924 if (split_order) {
925 xas_split_alloc(&xas, old, split_order, gfp);
926 if (xas_error(&xas))
927 goto error;
928 alloced_shadow = old;
929 alloced_order = split_order;
930 xas_reset(&xas);
931 continue;
932 }
933
934 if (!xas_nomem(&xas, gfp))
935 break;
936 }
937
938 if (xas_error(&xas))
939 goto error;
940
941 trace_mm_filemap_add_to_page_cache(folio);
942 return 0;
943error:
944 folio->mapping = NULL;
945 /* Leave page->index set: truncation relies upon it */
946 folio_put_refs(folio, nr);
947 return xas_error(&xas);
948}
949ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
950
951int filemap_add_folio(struct address_space *mapping, struct folio *folio,
952 pgoff_t index, gfp_t gfp)
953{
954 void *shadow = NULL;
955 int ret;
956
957 ret = mem_cgroup_charge(folio, NULL, gfp);
958 if (ret)
959 return ret;
960
961 __folio_set_locked(folio);
962 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
963 if (unlikely(ret)) {
964 mem_cgroup_uncharge(folio);
965 __folio_clear_locked(folio);
966 } else {
967 /*
968 * The folio might have been evicted from cache only
969 * recently, in which case it should be activated like
970 * any other repeatedly accessed folio.
971 * The exception is folios getting rewritten; evicting other
972 * data from the working set, only to cache data that will
973 * get overwritten with something else, is a waste of memory.
974 */
975 WARN_ON_ONCE(folio_test_active(folio));
976 if (!(gfp & __GFP_WRITE) && shadow)
977 workingset_refault(folio, shadow);
978 folio_add_lru(folio);
979 }
980 return ret;
981}
982EXPORT_SYMBOL_GPL(filemap_add_folio);
983
984#ifdef CONFIG_NUMA
985struct folio *filemap_alloc_folio_noprof(gfp_t gfp, unsigned int order)
986{
987 int n;
988 struct folio *folio;
989
990 if (cpuset_do_page_mem_spread()) {
991 unsigned int cpuset_mems_cookie;
992 do {
993 cpuset_mems_cookie = read_mems_allowed_begin();
994 n = cpuset_mem_spread_node();
995 folio = __folio_alloc_node_noprof(gfp, order, n);
996 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
997
998 return folio;
999 }
1000 return folio_alloc_noprof(gfp, order);
1001}
1002EXPORT_SYMBOL(filemap_alloc_folio_noprof);
1003#endif
1004
1005/*
1006 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
1007 *
1008 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
1009 *
1010 * @mapping1: the first mapping to lock
1011 * @mapping2: the second mapping to lock
1012 */
1013void filemap_invalidate_lock_two(struct address_space *mapping1,
1014 struct address_space *mapping2)
1015{
1016 if (mapping1 > mapping2)
1017 swap(mapping1, mapping2);
1018 if (mapping1)
1019 down_write(&mapping1->invalidate_lock);
1020 if (mapping2 && mapping1 != mapping2)
1021 down_write_nested(&mapping2->invalidate_lock, 1);
1022}
1023EXPORT_SYMBOL(filemap_invalidate_lock_two);
1024
1025/*
1026 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
1027 *
1028 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1029 *
1030 * @mapping1: the first mapping to unlock
1031 * @mapping2: the second mapping to unlock
1032 */
1033void filemap_invalidate_unlock_two(struct address_space *mapping1,
1034 struct address_space *mapping2)
1035{
1036 if (mapping1)
1037 up_write(&mapping1->invalidate_lock);
1038 if (mapping2 && mapping1 != mapping2)
1039 up_write(&mapping2->invalidate_lock);
1040}
1041EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1042
1043/*
1044 * In order to wait for pages to become available there must be
1045 * waitqueues associated with pages. By using a hash table of
1046 * waitqueues where the bucket discipline is to maintain all
1047 * waiters on the same queue and wake all when any of the pages
1048 * become available, and for the woken contexts to check to be
1049 * sure the appropriate page became available, this saves space
1050 * at a cost of "thundering herd" phenomena during rare hash
1051 * collisions.
1052 */
1053#define PAGE_WAIT_TABLE_BITS 8
1054#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1055static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1056
1057static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1058{
1059 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1060}
1061
1062void __init pagecache_init(void)
1063{
1064 int i;
1065
1066 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1067 init_waitqueue_head(&folio_wait_table[i]);
1068
1069 page_writeback_init();
1070}
1071
1072/*
1073 * The page wait code treats the "wait->flags" somewhat unusually, because
1074 * we have multiple different kinds of waits, not just the usual "exclusive"
1075 * one.
1076 *
1077 * We have:
1078 *
1079 * (a) no special bits set:
1080 *
1081 * We're just waiting for the bit to be released, and when a waker
1082 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1083 * and remove it from the wait queue.
1084 *
1085 * Simple and straightforward.
1086 *
1087 * (b) WQ_FLAG_EXCLUSIVE:
1088 *
1089 * The waiter is waiting to get the lock, and only one waiter should
1090 * be woken up to avoid any thundering herd behavior. We'll set the
1091 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1092 *
1093 * This is the traditional exclusive wait.
1094 *
1095 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1096 *
1097 * The waiter is waiting to get the bit, and additionally wants the
1098 * lock to be transferred to it for fair lock behavior. If the lock
1099 * cannot be taken, we stop walking the wait queue without waking
1100 * the waiter.
1101 *
1102 * This is the "fair lock handoff" case, and in addition to setting
1103 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1104 * that it now has the lock.
1105 */
1106static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1107{
1108 unsigned int flags;
1109 struct wait_page_key *key = arg;
1110 struct wait_page_queue *wait_page
1111 = container_of(wait, struct wait_page_queue, wait);
1112
1113 if (!wake_page_match(wait_page, key))
1114 return 0;
1115
1116 /*
1117 * If it's a lock handoff wait, we get the bit for it, and
1118 * stop walking (and do not wake it up) if we can't.
1119 */
1120 flags = wait->flags;
1121 if (flags & WQ_FLAG_EXCLUSIVE) {
1122 if (test_bit(key->bit_nr, &key->folio->flags))
1123 return -1;
1124 if (flags & WQ_FLAG_CUSTOM) {
1125 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1126 return -1;
1127 flags |= WQ_FLAG_DONE;
1128 }
1129 }
1130
1131 /*
1132 * We are holding the wait-queue lock, but the waiter that
1133 * is waiting for this will be checking the flags without
1134 * any locking.
1135 *
1136 * So update the flags atomically, and wake up the waiter
1137 * afterwards to avoid any races. This store-release pairs
1138 * with the load-acquire in folio_wait_bit_common().
1139 */
1140 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1141 wake_up_state(wait->private, mode);
1142
1143 /*
1144 * Ok, we have successfully done what we're waiting for,
1145 * and we can unconditionally remove the wait entry.
1146 *
1147 * Note that this pairs with the "finish_wait()" in the
1148 * waiter, and has to be the absolute last thing we do.
1149 * After this list_del_init(&wait->entry) the wait entry
1150 * might be de-allocated and the process might even have
1151 * exited.
1152 */
1153 list_del_init_careful(&wait->entry);
1154 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1155}
1156
1157static void folio_wake_bit(struct folio *folio, int bit_nr)
1158{
1159 wait_queue_head_t *q = folio_waitqueue(folio);
1160 struct wait_page_key key;
1161 unsigned long flags;
1162
1163 key.folio = folio;
1164 key.bit_nr = bit_nr;
1165 key.page_match = 0;
1166
1167 spin_lock_irqsave(&q->lock, flags);
1168 __wake_up_locked_key(q, TASK_NORMAL, &key);
1169
1170 /*
1171 * It's possible to miss clearing waiters here, when we woke our page
1172 * waiters, but the hashed waitqueue has waiters for other pages on it.
1173 * That's okay, it's a rare case. The next waker will clear it.
1174 *
1175 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1176 * other), the flag may be cleared in the course of freeing the page;
1177 * but that is not required for correctness.
1178 */
1179 if (!waitqueue_active(q) || !key.page_match)
1180 folio_clear_waiters(folio);
1181
1182 spin_unlock_irqrestore(&q->lock, flags);
1183}
1184
1185/*
1186 * A choice of three behaviors for folio_wait_bit_common():
1187 */
1188enum behavior {
1189 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1190 * __folio_lock() waiting on then setting PG_locked.
1191 */
1192 SHARED, /* Hold ref to page and check the bit when woken, like
1193 * folio_wait_writeback() waiting on PG_writeback.
1194 */
1195 DROP, /* Drop ref to page before wait, no check when woken,
1196 * like folio_put_wait_locked() on PG_locked.
1197 */
1198};
1199
1200/*
1201 * Attempt to check (or get) the folio flag, and mark us done
1202 * if successful.
1203 */
1204static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1205 struct wait_queue_entry *wait)
1206{
1207 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1208 if (test_and_set_bit(bit_nr, &folio->flags))
1209 return false;
1210 } else if (test_bit(bit_nr, &folio->flags))
1211 return false;
1212
1213 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1214 return true;
1215}
1216
1217/* How many times do we accept lock stealing from under a waiter? */
1218int sysctl_page_lock_unfairness = 5;
1219
1220static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1221 int state, enum behavior behavior)
1222{
1223 wait_queue_head_t *q = folio_waitqueue(folio);
1224 int unfairness = sysctl_page_lock_unfairness;
1225 struct wait_page_queue wait_page;
1226 wait_queue_entry_t *wait = &wait_page.wait;
1227 bool thrashing = false;
1228 unsigned long pflags;
1229 bool in_thrashing;
1230
1231 if (bit_nr == PG_locked &&
1232 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1233 delayacct_thrashing_start(&in_thrashing);
1234 psi_memstall_enter(&pflags);
1235 thrashing = true;
1236 }
1237
1238 init_wait(wait);
1239 wait->func = wake_page_function;
1240 wait_page.folio = folio;
1241 wait_page.bit_nr = bit_nr;
1242
1243repeat:
1244 wait->flags = 0;
1245 if (behavior == EXCLUSIVE) {
1246 wait->flags = WQ_FLAG_EXCLUSIVE;
1247 if (--unfairness < 0)
1248 wait->flags |= WQ_FLAG_CUSTOM;
1249 }
1250
1251 /*
1252 * Do one last check whether we can get the
1253 * page bit synchronously.
1254 *
1255 * Do the folio_set_waiters() marking before that
1256 * to let any waker we _just_ missed know they
1257 * need to wake us up (otherwise they'll never
1258 * even go to the slow case that looks at the
1259 * page queue), and add ourselves to the wait
1260 * queue if we need to sleep.
1261 *
1262 * This part needs to be done under the queue
1263 * lock to avoid races.
1264 */
1265 spin_lock_irq(&q->lock);
1266 folio_set_waiters(folio);
1267 if (!folio_trylock_flag(folio, bit_nr, wait))
1268 __add_wait_queue_entry_tail(q, wait);
1269 spin_unlock_irq(&q->lock);
1270
1271 /*
1272 * From now on, all the logic will be based on
1273 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1274 * see whether the page bit testing has already
1275 * been done by the wake function.
1276 *
1277 * We can drop our reference to the folio.
1278 */
1279 if (behavior == DROP)
1280 folio_put(folio);
1281
1282 /*
1283 * Note that until the "finish_wait()", or until
1284 * we see the WQ_FLAG_WOKEN flag, we need to
1285 * be very careful with the 'wait->flags', because
1286 * we may race with a waker that sets them.
1287 */
1288 for (;;) {
1289 unsigned int flags;
1290
1291 set_current_state(state);
1292
1293 /* Loop until we've been woken or interrupted */
1294 flags = smp_load_acquire(&wait->flags);
1295 if (!(flags & WQ_FLAG_WOKEN)) {
1296 if (signal_pending_state(state, current))
1297 break;
1298
1299 io_schedule();
1300 continue;
1301 }
1302
1303 /* If we were non-exclusive, we're done */
1304 if (behavior != EXCLUSIVE)
1305 break;
1306
1307 /* If the waker got the lock for us, we're done */
1308 if (flags & WQ_FLAG_DONE)
1309 break;
1310
1311 /*
1312 * Otherwise, if we're getting the lock, we need to
1313 * try to get it ourselves.
1314 *
1315 * And if that fails, we'll have to retry this all.
1316 */
1317 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1318 goto repeat;
1319
1320 wait->flags |= WQ_FLAG_DONE;
1321 break;
1322 }
1323
1324 /*
1325 * If a signal happened, this 'finish_wait()' may remove the last
1326 * waiter from the wait-queues, but the folio waiters bit will remain
1327 * set. That's ok. The next wakeup will take care of it, and trying
1328 * to do it here would be difficult and prone to races.
1329 */
1330 finish_wait(q, wait);
1331
1332 if (thrashing) {
1333 delayacct_thrashing_end(&in_thrashing);
1334 psi_memstall_leave(&pflags);
1335 }
1336
1337 /*
1338 * NOTE! The wait->flags weren't stable until we've done the
1339 * 'finish_wait()', and we could have exited the loop above due
1340 * to a signal, and had a wakeup event happen after the signal
1341 * test but before the 'finish_wait()'.
1342 *
1343 * So only after the finish_wait() can we reliably determine
1344 * if we got woken up or not, so we can now figure out the final
1345 * return value based on that state without races.
1346 *
1347 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1348 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1349 */
1350 if (behavior == EXCLUSIVE)
1351 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1352
1353 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1354}
1355
1356#ifdef CONFIG_MIGRATION
1357/**
1358 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1359 * @entry: migration swap entry.
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 unlocking the ptl.
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, spinlock_t *ptl)
1374 __releases(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 = pfn_swap_entry_folio(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 spin_unlock(ptl);
1409
1410 for (;;) {
1411 unsigned int flags;
1412
1413 set_current_state(TASK_UNINTERRUPTIBLE);
1414
1415 /* Loop until we've been woken or interrupted */
1416 flags = smp_load_acquire(&wait->flags);
1417 if (!(flags & WQ_FLAG_WOKEN)) {
1418 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1419 break;
1420
1421 io_schedule();
1422 continue;
1423 }
1424 break;
1425 }
1426
1427 finish_wait(q, wait);
1428
1429 if (thrashing) {
1430 delayacct_thrashing_end(&in_thrashing);
1431 psi_memstall_leave(&pflags);
1432 }
1433}
1434#endif
1435
1436void folio_wait_bit(struct folio *folio, int bit_nr)
1437{
1438 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1439}
1440EXPORT_SYMBOL(folio_wait_bit);
1441
1442int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1443{
1444 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1445}
1446EXPORT_SYMBOL(folio_wait_bit_killable);
1447
1448/**
1449 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1450 * @folio: The folio to wait for.
1451 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1452 *
1453 * The caller should hold a reference on @folio. They expect the page to
1454 * become unlocked relatively soon, but do not wish to hold up migration
1455 * (for example) by holding the reference while waiting for the folio to
1456 * come unlocked. After this function returns, the caller should not
1457 * dereference @folio.
1458 *
1459 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1460 */
1461static int folio_put_wait_locked(struct folio *folio, int state)
1462{
1463 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1464}
1465
1466/**
1467 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue
1468 * @folio: Folio defining the wait queue of interest
1469 * @waiter: Waiter to add to the queue
1470 *
1471 * Add an arbitrary @waiter to the wait queue for the nominated @folio.
1472 */
1473void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter)
1474{
1475 wait_queue_head_t *q = folio_waitqueue(folio);
1476 unsigned long flags;
1477
1478 spin_lock_irqsave(&q->lock, flags);
1479 __add_wait_queue_entry_tail(q, waiter);
1480 folio_set_waiters(folio);
1481 spin_unlock_irqrestore(&q->lock, flags);
1482}
1483EXPORT_SYMBOL_GPL(folio_add_wait_queue);
1484
1485/**
1486 * folio_unlock - Unlock a locked folio.
1487 * @folio: The folio.
1488 *
1489 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1490 *
1491 * Context: May be called from interrupt or process context. May not be
1492 * called from NMI context.
1493 */
1494void folio_unlock(struct folio *folio)
1495{
1496 /* Bit 7 allows x86 to check the byte's sign bit */
1497 BUILD_BUG_ON(PG_waiters != 7);
1498 BUILD_BUG_ON(PG_locked > 7);
1499 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1500 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked))
1501 folio_wake_bit(folio, PG_locked);
1502}
1503EXPORT_SYMBOL(folio_unlock);
1504
1505/**
1506 * folio_end_read - End read on a folio.
1507 * @folio: The folio.
1508 * @success: True if all reads completed successfully.
1509 *
1510 * When all reads against a folio have completed, filesystems should
1511 * call this function to let the pagecache know that no more reads
1512 * are outstanding. This will unlock the folio and wake up any thread
1513 * sleeping on the lock. The folio will also be marked uptodate if all
1514 * reads succeeded.
1515 *
1516 * Context: May be called from interrupt or process context. May not be
1517 * called from NMI context.
1518 */
1519void folio_end_read(struct folio *folio, bool success)
1520{
1521 unsigned long mask = 1 << PG_locked;
1522
1523 /* Must be in bottom byte for x86 to work */
1524 BUILD_BUG_ON(PG_uptodate > 7);
1525 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1526 VM_BUG_ON_FOLIO(success && folio_test_uptodate(folio), folio);
1527
1528 if (likely(success))
1529 mask |= 1 << PG_uptodate;
1530 if (folio_xor_flags_has_waiters(folio, mask))
1531 folio_wake_bit(folio, PG_locked);
1532}
1533EXPORT_SYMBOL(folio_end_read);
1534
1535/**
1536 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1537 * @folio: The folio.
1538 *
1539 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1540 * it. The folio reference held for PG_private_2 being set is released.
1541 *
1542 * This is, for example, used when a netfs folio is being written to a local
1543 * disk cache, thereby allowing writes to the cache for the same folio to be
1544 * serialised.
1545 */
1546void folio_end_private_2(struct folio *folio)
1547{
1548 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1549 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1550 folio_wake_bit(folio, PG_private_2);
1551 folio_put(folio);
1552}
1553EXPORT_SYMBOL(folio_end_private_2);
1554
1555/**
1556 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1557 * @folio: The folio to wait on.
1558 *
1559 * Wait for PG_private_2 to be cleared on a folio.
1560 */
1561void folio_wait_private_2(struct folio *folio)
1562{
1563 while (folio_test_private_2(folio))
1564 folio_wait_bit(folio, PG_private_2);
1565}
1566EXPORT_SYMBOL(folio_wait_private_2);
1567
1568/**
1569 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1570 * @folio: The folio to wait on.
1571 *
1572 * Wait for PG_private_2 to be cleared on a folio or until a fatal signal is
1573 * received by the calling task.
1574 *
1575 * Return:
1576 * - 0 if successful.
1577 * - -EINTR if a fatal signal was encountered.
1578 */
1579int folio_wait_private_2_killable(struct folio *folio)
1580{
1581 int ret = 0;
1582
1583 while (folio_test_private_2(folio)) {
1584 ret = folio_wait_bit_killable(folio, PG_private_2);
1585 if (ret < 0)
1586 break;
1587 }
1588
1589 return ret;
1590}
1591EXPORT_SYMBOL(folio_wait_private_2_killable);
1592
1593/**
1594 * folio_end_writeback - End writeback against a folio.
1595 * @folio: The folio.
1596 *
1597 * The folio must actually be under writeback.
1598 *
1599 * Context: May be called from process or interrupt context.
1600 */
1601void folio_end_writeback(struct folio *folio)
1602{
1603 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
1604
1605 /*
1606 * folio_test_clear_reclaim() could be used here but it is an
1607 * atomic operation and overkill in this particular case. Failing
1608 * to shuffle a folio marked for immediate reclaim is too mild
1609 * a gain to justify taking an atomic operation penalty at the
1610 * end of every folio writeback.
1611 */
1612 if (folio_test_reclaim(folio)) {
1613 folio_clear_reclaim(folio);
1614 folio_rotate_reclaimable(folio);
1615 }
1616
1617 /*
1618 * Writeback does not hold a folio reference of its own, relying
1619 * on truncation to wait for the clearing of PG_writeback.
1620 * But here we must make sure that the folio is not freed and
1621 * reused before the folio_wake_bit().
1622 */
1623 folio_get(folio);
1624 if (__folio_end_writeback(folio))
1625 folio_wake_bit(folio, PG_writeback);
1626 acct_reclaim_writeback(folio);
1627 folio_put(folio);
1628}
1629EXPORT_SYMBOL(folio_end_writeback);
1630
1631/**
1632 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1633 * @folio: The folio to lock
1634 */
1635void __folio_lock(struct folio *folio)
1636{
1637 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1638 EXCLUSIVE);
1639}
1640EXPORT_SYMBOL(__folio_lock);
1641
1642int __folio_lock_killable(struct folio *folio)
1643{
1644 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1645 EXCLUSIVE);
1646}
1647EXPORT_SYMBOL_GPL(__folio_lock_killable);
1648
1649static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1650{
1651 struct wait_queue_head *q = folio_waitqueue(folio);
1652 int ret;
1653
1654 wait->folio = folio;
1655 wait->bit_nr = PG_locked;
1656
1657 spin_lock_irq(&q->lock);
1658 __add_wait_queue_entry_tail(q, &wait->wait);
1659 folio_set_waiters(folio);
1660 ret = !folio_trylock(folio);
1661 /*
1662 * If we were successful now, we know we're still on the
1663 * waitqueue as we're still under the lock. This means it's
1664 * safe to remove and return success, we know the callback
1665 * isn't going to trigger.
1666 */
1667 if (!ret)
1668 __remove_wait_queue(q, &wait->wait);
1669 else
1670 ret = -EIOCBQUEUED;
1671 spin_unlock_irq(&q->lock);
1672 return ret;
1673}
1674
1675/*
1676 * Return values:
1677 * 0 - folio is locked.
1678 * non-zero - folio is not locked.
1679 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or
1680 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and
1681 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held.
1682 *
1683 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0
1684 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed.
1685 */
1686vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf)
1687{
1688 unsigned int flags = vmf->flags;
1689
1690 if (fault_flag_allow_retry_first(flags)) {
1691 /*
1692 * CAUTION! In this case, mmap_lock/per-VMA lock is not
1693 * released even though returning VM_FAULT_RETRY.
1694 */
1695 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1696 return VM_FAULT_RETRY;
1697
1698 release_fault_lock(vmf);
1699 if (flags & FAULT_FLAG_KILLABLE)
1700 folio_wait_locked_killable(folio);
1701 else
1702 folio_wait_locked(folio);
1703 return VM_FAULT_RETRY;
1704 }
1705 if (flags & FAULT_FLAG_KILLABLE) {
1706 bool ret;
1707
1708 ret = __folio_lock_killable(folio);
1709 if (ret) {
1710 release_fault_lock(vmf);
1711 return VM_FAULT_RETRY;
1712 }
1713 } else {
1714 __folio_lock(folio);
1715 }
1716
1717 return 0;
1718}
1719
1720/**
1721 * page_cache_next_miss() - Find the next gap in the page cache.
1722 * @mapping: Mapping.
1723 * @index: Index.
1724 * @max_scan: Maximum range to search.
1725 *
1726 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1727 * gap with the lowest index.
1728 *
1729 * This function may be called under the rcu_read_lock. However, this will
1730 * not atomically search a snapshot of the cache at a single point in time.
1731 * For example, if a gap is created at index 5, then subsequently a gap is
1732 * created at index 10, page_cache_next_miss covering both indices may
1733 * return 10 if called under the rcu_read_lock.
1734 *
1735 * Return: The index of the gap if found, otherwise an index outside the
1736 * range specified (in which case 'return - index >= max_scan' will be true).
1737 * In the rare case of index wrap-around, 0 will be returned.
1738 */
1739pgoff_t page_cache_next_miss(struct address_space *mapping,
1740 pgoff_t index, unsigned long max_scan)
1741{
1742 XA_STATE(xas, &mapping->i_pages, index);
1743
1744 while (max_scan--) {
1745 void *entry = xas_next(&xas);
1746 if (!entry || xa_is_value(entry))
1747 return xas.xa_index;
1748 if (xas.xa_index == 0)
1749 return 0;
1750 }
1751
1752 return index + max_scan;
1753}
1754EXPORT_SYMBOL(page_cache_next_miss);
1755
1756/**
1757 * page_cache_prev_miss() - Find the previous gap in the page cache.
1758 * @mapping: Mapping.
1759 * @index: Index.
1760 * @max_scan: Maximum range to search.
1761 *
1762 * Search the range [max(index - max_scan + 1, 0), index] for the
1763 * gap with the highest index.
1764 *
1765 * This function may be called under the rcu_read_lock. However, this will
1766 * not atomically search a snapshot of the cache at a single point in time.
1767 * For example, if a gap is created at index 10, then subsequently a gap is
1768 * created at index 5, page_cache_prev_miss() covering both indices may
1769 * return 5 if called under the rcu_read_lock.
1770 *
1771 * Return: The index of the gap if found, otherwise an index outside the
1772 * range specified (in which case 'index - return >= max_scan' will be true).
1773 * In the rare case of wrap-around, ULONG_MAX will be returned.
1774 */
1775pgoff_t page_cache_prev_miss(struct address_space *mapping,
1776 pgoff_t index, unsigned long max_scan)
1777{
1778 XA_STATE(xas, &mapping->i_pages, index);
1779
1780 while (max_scan--) {
1781 void *entry = xas_prev(&xas);
1782 if (!entry || xa_is_value(entry))
1783 break;
1784 if (xas.xa_index == ULONG_MAX)
1785 break;
1786 }
1787
1788 return xas.xa_index;
1789}
1790EXPORT_SYMBOL(page_cache_prev_miss);
1791
1792/*
1793 * Lockless page cache protocol:
1794 * On the lookup side:
1795 * 1. Load the folio from i_pages
1796 * 2. Increment the refcount if it's not zero
1797 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1798 *
1799 * On the removal side:
1800 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1801 * B. Remove the page from i_pages
1802 * C. Return the page to the page allocator
1803 *
1804 * This means that any page may have its reference count temporarily
1805 * increased by a speculative page cache (or GUP-fast) lookup as it can
1806 * be allocated by another user before the RCU grace period expires.
1807 * Because the refcount temporarily acquired here may end up being the
1808 * last refcount on the page, any page allocation must be freeable by
1809 * folio_put().
1810 */
1811
1812/*
1813 * filemap_get_entry - Get a page cache entry.
1814 * @mapping: the address_space to search
1815 * @index: The page cache index.
1816 *
1817 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1818 * it is returned with an increased refcount. If it is a shadow entry
1819 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1820 * it is returned without further action.
1821 *
1822 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1823 */
1824void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
1825{
1826 XA_STATE(xas, &mapping->i_pages, index);
1827 struct folio *folio;
1828
1829 rcu_read_lock();
1830repeat:
1831 xas_reset(&xas);
1832 folio = xas_load(&xas);
1833 if (xas_retry(&xas, folio))
1834 goto repeat;
1835 /*
1836 * A shadow entry of a recently evicted page, or a swap entry from
1837 * shmem/tmpfs. Return it without attempting to raise page count.
1838 */
1839 if (!folio || xa_is_value(folio))
1840 goto out;
1841
1842 if (!folio_try_get(folio))
1843 goto repeat;
1844
1845 if (unlikely(folio != xas_reload(&xas))) {
1846 folio_put(folio);
1847 goto repeat;
1848 }
1849out:
1850 rcu_read_unlock();
1851
1852 return folio;
1853}
1854
1855/**
1856 * __filemap_get_folio - Find and get a reference to a folio.
1857 * @mapping: The address_space to search.
1858 * @index: The page index.
1859 * @fgp_flags: %FGP flags modify how the folio is returned.
1860 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1861 *
1862 * Looks up the page cache entry at @mapping & @index.
1863 *
1864 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1865 * if the %GFP flags specified for %FGP_CREAT are atomic.
1866 *
1867 * If this function returns a folio, it is returned with an increased refcount.
1868 *
1869 * Return: The found folio or an ERR_PTR() otherwise.
1870 */
1871struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1872 fgf_t fgp_flags, gfp_t gfp)
1873{
1874 struct folio *folio;
1875
1876repeat:
1877 folio = filemap_get_entry(mapping, index);
1878 if (xa_is_value(folio))
1879 folio = NULL;
1880 if (!folio)
1881 goto no_page;
1882
1883 if (fgp_flags & FGP_LOCK) {
1884 if (fgp_flags & FGP_NOWAIT) {
1885 if (!folio_trylock(folio)) {
1886 folio_put(folio);
1887 return ERR_PTR(-EAGAIN);
1888 }
1889 } else {
1890 folio_lock(folio);
1891 }
1892
1893 /* Has the page been truncated? */
1894 if (unlikely(folio->mapping != mapping)) {
1895 folio_unlock(folio);
1896 folio_put(folio);
1897 goto repeat;
1898 }
1899 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1900 }
1901
1902 if (fgp_flags & FGP_ACCESSED)
1903 folio_mark_accessed(folio);
1904 else if (fgp_flags & FGP_WRITE) {
1905 /* Clear idle flag for buffer write */
1906 if (folio_test_idle(folio))
1907 folio_clear_idle(folio);
1908 }
1909
1910 if (fgp_flags & FGP_STABLE)
1911 folio_wait_stable(folio);
1912no_page:
1913 if (!folio && (fgp_flags & FGP_CREAT)) {
1914 unsigned int min_order = mapping_min_folio_order(mapping);
1915 unsigned int order = max(min_order, FGF_GET_ORDER(fgp_flags));
1916 int err;
1917 index = mapping_align_index(mapping, index);
1918
1919 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1920 gfp |= __GFP_WRITE;
1921 if (fgp_flags & FGP_NOFS)
1922 gfp &= ~__GFP_FS;
1923 if (fgp_flags & FGP_NOWAIT) {
1924 gfp &= ~GFP_KERNEL;
1925 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1926 }
1927 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1928 fgp_flags |= FGP_LOCK;
1929
1930 if (order > mapping_max_folio_order(mapping))
1931 order = mapping_max_folio_order(mapping);
1932 /* If we're not aligned, allocate a smaller folio */
1933 if (index & ((1UL << order) - 1))
1934 order = __ffs(index);
1935
1936 do {
1937 gfp_t alloc_gfp = gfp;
1938
1939 err = -ENOMEM;
1940 if (order > min_order)
1941 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN;
1942 folio = filemap_alloc_folio(alloc_gfp, order);
1943 if (!folio)
1944 continue;
1945
1946 /* Init accessed so avoid atomic mark_page_accessed later */
1947 if (fgp_flags & FGP_ACCESSED)
1948 __folio_set_referenced(folio);
1949
1950 err = filemap_add_folio(mapping, folio, index, gfp);
1951 if (!err)
1952 break;
1953 folio_put(folio);
1954 folio = NULL;
1955 } while (order-- > min_order);
1956
1957 if (err == -EEXIST)
1958 goto repeat;
1959 if (err)
1960 return ERR_PTR(err);
1961 /*
1962 * filemap_add_folio locks the page, and for mmap
1963 * we expect an unlocked page.
1964 */
1965 if (folio && (fgp_flags & FGP_FOR_MMAP))
1966 folio_unlock(folio);
1967 }
1968
1969 if (!folio)
1970 return ERR_PTR(-ENOENT);
1971 return folio;
1972}
1973EXPORT_SYMBOL(__filemap_get_folio);
1974
1975static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
1976 xa_mark_t mark)
1977{
1978 struct folio *folio;
1979
1980retry:
1981 if (mark == XA_PRESENT)
1982 folio = xas_find(xas, max);
1983 else
1984 folio = xas_find_marked(xas, max, mark);
1985
1986 if (xas_retry(xas, folio))
1987 goto retry;
1988 /*
1989 * A shadow entry of a recently evicted page, a swap
1990 * entry from shmem/tmpfs or a DAX entry. Return it
1991 * without attempting to raise page count.
1992 */
1993 if (!folio || xa_is_value(folio))
1994 return folio;
1995
1996 if (!folio_try_get(folio))
1997 goto reset;
1998
1999 if (unlikely(folio != xas_reload(xas))) {
2000 folio_put(folio);
2001 goto reset;
2002 }
2003
2004 return folio;
2005reset:
2006 xas_reset(xas);
2007 goto retry;
2008}
2009
2010/**
2011 * find_get_entries - gang pagecache lookup
2012 * @mapping: The address_space to search
2013 * @start: The starting page cache index
2014 * @end: The final page index (inclusive).
2015 * @fbatch: Where the resulting entries are placed.
2016 * @indices: The cache indices corresponding to the entries in @entries
2017 *
2018 * find_get_entries() will search for and return a batch of entries in
2019 * the mapping. The entries are placed in @fbatch. find_get_entries()
2020 * takes a reference on any actual folios it returns.
2021 *
2022 * The entries have ascending indexes. The indices may not be consecutive
2023 * due to not-present entries or large folios.
2024 *
2025 * Any shadow entries of evicted folios, or swap entries from
2026 * shmem/tmpfs, are included in the returned array.
2027 *
2028 * Return: The number of entries which were found.
2029 */
2030unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2031 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2032{
2033 XA_STATE(xas, &mapping->i_pages, *start);
2034 struct folio *folio;
2035
2036 rcu_read_lock();
2037 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2038 indices[fbatch->nr] = xas.xa_index;
2039 if (!folio_batch_add(fbatch, folio))
2040 break;
2041 }
2042
2043 if (folio_batch_count(fbatch)) {
2044 unsigned long nr;
2045 int idx = folio_batch_count(fbatch) - 1;
2046
2047 folio = fbatch->folios[idx];
2048 if (!xa_is_value(folio))
2049 nr = folio_nr_pages(folio);
2050 else
2051 nr = 1 << xa_get_order(&mapping->i_pages, indices[idx]);
2052 *start = round_down(indices[idx] + nr, nr);
2053 }
2054 rcu_read_unlock();
2055
2056 return folio_batch_count(fbatch);
2057}
2058
2059/**
2060 * find_lock_entries - Find a batch of pagecache entries.
2061 * @mapping: The address_space to search.
2062 * @start: The starting page cache index.
2063 * @end: The final page index (inclusive).
2064 * @fbatch: Where the resulting entries are placed.
2065 * @indices: The cache indices of the entries in @fbatch.
2066 *
2067 * find_lock_entries() will return a batch of entries from @mapping.
2068 * Swap, shadow and DAX entries are included. Folios are returned
2069 * locked and with an incremented refcount. Folios which are locked
2070 * by somebody else or under writeback are skipped. Folios which are
2071 * partially outside the range are not returned.
2072 *
2073 * The entries have ascending indexes. The indices may not be consecutive
2074 * due to not-present entries, large folios, folios which could not be
2075 * locked or folios under writeback.
2076 *
2077 * Return: The number of entries which were found.
2078 */
2079unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2080 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2081{
2082 XA_STATE(xas, &mapping->i_pages, *start);
2083 struct folio *folio;
2084
2085 rcu_read_lock();
2086 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2087 unsigned long base;
2088 unsigned long nr;
2089
2090 if (!xa_is_value(folio)) {
2091 nr = folio_nr_pages(folio);
2092 base = folio->index;
2093 /* Omit large folio which begins before the start */
2094 if (base < *start)
2095 goto put;
2096 /* Omit large folio which extends beyond the end */
2097 if (base + nr - 1 > end)
2098 goto put;
2099 if (!folio_trylock(folio))
2100 goto put;
2101 if (folio->mapping != mapping ||
2102 folio_test_writeback(folio))
2103 goto unlock;
2104 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2105 folio);
2106 } else {
2107 nr = 1 << xas_get_order(&xas);
2108 base = xas.xa_index & ~(nr - 1);
2109 /* Omit order>0 value which begins before the start */
2110 if (base < *start)
2111 continue;
2112 /* Omit order>0 value which extends beyond the end */
2113 if (base + nr - 1 > end)
2114 break;
2115 }
2116
2117 /* Update start now so that last update is correct on return */
2118 *start = base + nr;
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 return folio_batch_count(fbatch);
2131}
2132
2133/**
2134 * filemap_get_folios - Get a batch of folios
2135 * @mapping: The address_space to search
2136 * @start: The starting page index
2137 * @end: The final page index (inclusive)
2138 * @fbatch: The batch to fill.
2139 *
2140 * Search for and return a batch of folios in the mapping starting at
2141 * index @start and up to index @end (inclusive). The folios are returned
2142 * in @fbatch with an elevated reference count.
2143 *
2144 * Return: The number of folios which were found.
2145 * We also update @start to index the next folio for the traversal.
2146 */
2147unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2148 pgoff_t end, struct folio_batch *fbatch)
2149{
2150 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch);
2151}
2152EXPORT_SYMBOL(filemap_get_folios);
2153
2154/**
2155 * filemap_get_folios_contig - Get a batch of contiguous folios
2156 * @mapping: The address_space to search
2157 * @start: The starting page index
2158 * @end: The final page index (inclusive)
2159 * @fbatch: The batch to fill
2160 *
2161 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2162 * except the returned folios are guaranteed to be contiguous. This may
2163 * not return all contiguous folios if the batch gets filled up.
2164 *
2165 * Return: The number of folios found.
2166 * Also update @start to be positioned for traversal of the next folio.
2167 */
2168
2169unsigned filemap_get_folios_contig(struct address_space *mapping,
2170 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2171{
2172 XA_STATE(xas, &mapping->i_pages, *start);
2173 unsigned long nr;
2174 struct folio *folio;
2175
2176 rcu_read_lock();
2177
2178 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2179 folio = xas_next(&xas)) {
2180 if (xas_retry(&xas, folio))
2181 continue;
2182 /*
2183 * If the entry has been swapped out, we can stop looking.
2184 * No current caller is looking for DAX entries.
2185 */
2186 if (xa_is_value(folio))
2187 goto update_start;
2188
2189 /* If we landed in the middle of a THP, continue at its end. */
2190 if (xa_is_sibling(folio))
2191 goto update_start;
2192
2193 if (!folio_try_get(folio))
2194 goto retry;
2195
2196 if (unlikely(folio != xas_reload(&xas)))
2197 goto put_folio;
2198
2199 if (!folio_batch_add(fbatch, folio)) {
2200 nr = folio_nr_pages(folio);
2201 *start = folio->index + nr;
2202 goto out;
2203 }
2204 continue;
2205put_folio:
2206 folio_put(folio);
2207
2208retry:
2209 xas_reset(&xas);
2210 }
2211
2212update_start:
2213 nr = folio_batch_count(fbatch);
2214
2215 if (nr) {
2216 folio = fbatch->folios[nr - 1];
2217 *start = folio_next_index(folio);
2218 }
2219out:
2220 rcu_read_unlock();
2221 return folio_batch_count(fbatch);
2222}
2223EXPORT_SYMBOL(filemap_get_folios_contig);
2224
2225/**
2226 * filemap_get_folios_tag - Get a batch of folios matching @tag
2227 * @mapping: The address_space to search
2228 * @start: The starting page index
2229 * @end: The final page index (inclusive)
2230 * @tag: The tag index
2231 * @fbatch: The batch to fill
2232 *
2233 * The first folio may start before @start; if it does, it will contain
2234 * @start. The final folio may extend beyond @end; if it does, it will
2235 * contain @end. The folios have ascending indices. There may be gaps
2236 * between the folios if there are indices which have no folio in the
2237 * page cache. If folios are added to or removed from the page cache
2238 * while this is running, they may or may not be found by this call.
2239 * Only returns folios that are tagged with @tag.
2240 *
2241 * Return: The number of folios found.
2242 * Also update @start to index the next folio for traversal.
2243 */
2244unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start,
2245 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch)
2246{
2247 XA_STATE(xas, &mapping->i_pages, *start);
2248 struct folio *folio;
2249
2250 rcu_read_lock();
2251 while ((folio = find_get_entry(&xas, end, tag)) != NULL) {
2252 /*
2253 * Shadow entries should never be tagged, but this iteration
2254 * is lockless so there is a window for page reclaim to evict
2255 * a page we saw tagged. Skip over it.
2256 */
2257 if (xa_is_value(folio))
2258 continue;
2259 if (!folio_batch_add(fbatch, folio)) {
2260 unsigned long nr = folio_nr_pages(folio);
2261 *start = folio->index + nr;
2262 goto out;
2263 }
2264 }
2265 /*
2266 * We come here when there is no page beyond @end. We take care to not
2267 * overflow the index @start as it confuses some of the callers. This
2268 * breaks the iteration when there is a page at index -1 but that is
2269 * already broke anyway.
2270 */
2271 if (end == (pgoff_t)-1)
2272 *start = (pgoff_t)-1;
2273 else
2274 *start = end + 1;
2275out:
2276 rcu_read_unlock();
2277
2278 return folio_batch_count(fbatch);
2279}
2280EXPORT_SYMBOL(filemap_get_folios_tag);
2281
2282/*
2283 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2284 * a _large_ part of the i/o request. Imagine the worst scenario:
2285 *
2286 * ---R__________________________________________B__________
2287 * ^ reading here ^ bad block(assume 4k)
2288 *
2289 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2290 * => failing the whole request => read(R) => read(R+1) =>
2291 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2292 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2293 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2294 *
2295 * It is going insane. Fix it by quickly scaling down the readahead size.
2296 */
2297static void shrink_readahead_size_eio(struct file_ra_state *ra)
2298{
2299 ra->ra_pages /= 4;
2300}
2301
2302/*
2303 * filemap_get_read_batch - Get a batch of folios for read
2304 *
2305 * Get a batch of folios which represent a contiguous range of bytes in
2306 * the file. No exceptional entries will be returned. If @index is in
2307 * the middle of a folio, the entire folio will be returned. The last
2308 * folio in the batch may have the readahead flag set or the uptodate flag
2309 * clear so that the caller can take the appropriate action.
2310 */
2311static void filemap_get_read_batch(struct address_space *mapping,
2312 pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
2313{
2314 XA_STATE(xas, &mapping->i_pages, index);
2315 struct folio *folio;
2316
2317 rcu_read_lock();
2318 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2319 if (xas_retry(&xas, folio))
2320 continue;
2321 if (xas.xa_index > max || xa_is_value(folio))
2322 break;
2323 if (xa_is_sibling(folio))
2324 break;
2325 if (!folio_try_get(folio))
2326 goto retry;
2327
2328 if (unlikely(folio != xas_reload(&xas)))
2329 goto put_folio;
2330
2331 if (!folio_batch_add(fbatch, folio))
2332 break;
2333 if (!folio_test_uptodate(folio))
2334 break;
2335 if (folio_test_readahead(folio))
2336 break;
2337 xas_advance(&xas, folio_next_index(folio) - 1);
2338 continue;
2339put_folio:
2340 folio_put(folio);
2341retry:
2342 xas_reset(&xas);
2343 }
2344 rcu_read_unlock();
2345}
2346
2347static int filemap_read_folio(struct file *file, filler_t filler,
2348 struct folio *folio)
2349{
2350 bool workingset = folio_test_workingset(folio);
2351 unsigned long pflags;
2352 int error;
2353
2354 /* Start the actual read. The read will unlock the page. */
2355 if (unlikely(workingset))
2356 psi_memstall_enter(&pflags);
2357 error = filler(file, folio);
2358 if (unlikely(workingset))
2359 psi_memstall_leave(&pflags);
2360 if (error)
2361 return error;
2362
2363 error = folio_wait_locked_killable(folio);
2364 if (error)
2365 return error;
2366 if (folio_test_uptodate(folio))
2367 return 0;
2368 if (file)
2369 shrink_readahead_size_eio(&file->f_ra);
2370 return -EIO;
2371}
2372
2373static bool filemap_range_uptodate(struct address_space *mapping,
2374 loff_t pos, size_t count, struct folio *folio,
2375 bool need_uptodate)
2376{
2377 if (folio_test_uptodate(folio))
2378 return true;
2379 /* pipes can't handle partially uptodate pages */
2380 if (need_uptodate)
2381 return false;
2382 if (!mapping->a_ops->is_partially_uptodate)
2383 return false;
2384 if (mapping->host->i_blkbits >= folio_shift(folio))
2385 return false;
2386
2387 if (folio_pos(folio) > pos) {
2388 count -= folio_pos(folio) - pos;
2389 pos = 0;
2390 } else {
2391 pos -= folio_pos(folio);
2392 }
2393
2394 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2395}
2396
2397static int filemap_update_page(struct kiocb *iocb,
2398 struct address_space *mapping, size_t count,
2399 struct folio *folio, bool need_uptodate)
2400{
2401 int error;
2402
2403 if (iocb->ki_flags & IOCB_NOWAIT) {
2404 if (!filemap_invalidate_trylock_shared(mapping))
2405 return -EAGAIN;
2406 } else {
2407 filemap_invalidate_lock_shared(mapping);
2408 }
2409
2410 if (!folio_trylock(folio)) {
2411 error = -EAGAIN;
2412 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2413 goto unlock_mapping;
2414 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2415 filemap_invalidate_unlock_shared(mapping);
2416 /*
2417 * This is where we usually end up waiting for a
2418 * previously submitted readahead to finish.
2419 */
2420 folio_put_wait_locked(folio, TASK_KILLABLE);
2421 return AOP_TRUNCATED_PAGE;
2422 }
2423 error = __folio_lock_async(folio, iocb->ki_waitq);
2424 if (error)
2425 goto unlock_mapping;
2426 }
2427
2428 error = AOP_TRUNCATED_PAGE;
2429 if (!folio->mapping)
2430 goto unlock;
2431
2432 error = 0;
2433 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio,
2434 need_uptodate))
2435 goto unlock;
2436
2437 error = -EAGAIN;
2438 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2439 goto unlock;
2440
2441 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2442 folio);
2443 goto unlock_mapping;
2444unlock:
2445 folio_unlock(folio);
2446unlock_mapping:
2447 filemap_invalidate_unlock_shared(mapping);
2448 if (error == AOP_TRUNCATED_PAGE)
2449 folio_put(folio);
2450 return error;
2451}
2452
2453static int filemap_create_folio(struct file *file,
2454 struct address_space *mapping, loff_t pos,
2455 struct folio_batch *fbatch)
2456{
2457 struct folio *folio;
2458 int error;
2459 unsigned int min_order = mapping_min_folio_order(mapping);
2460 pgoff_t index;
2461
2462 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), min_order);
2463 if (!folio)
2464 return -ENOMEM;
2465
2466 /*
2467 * Protect against truncate / hole punch. Grabbing invalidate_lock
2468 * here assures we cannot instantiate and bring uptodate new
2469 * pagecache folios after evicting page cache during truncate
2470 * and before actually freeing blocks. Note that we could
2471 * release invalidate_lock after inserting the folio into
2472 * the page cache as the locked folio would then be enough to
2473 * synchronize with hole punching. But there are code paths
2474 * such as filemap_update_page() filling in partially uptodate
2475 * pages or ->readahead() that need to hold invalidate_lock
2476 * while mapping blocks for IO so let's hold the lock here as
2477 * well to keep locking rules simple.
2478 */
2479 filemap_invalidate_lock_shared(mapping);
2480 index = (pos >> (PAGE_SHIFT + min_order)) << min_order;
2481 error = filemap_add_folio(mapping, folio, index,
2482 mapping_gfp_constraint(mapping, GFP_KERNEL));
2483 if (error == -EEXIST)
2484 error = AOP_TRUNCATED_PAGE;
2485 if (error)
2486 goto error;
2487
2488 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
2489 if (error)
2490 goto error;
2491
2492 filemap_invalidate_unlock_shared(mapping);
2493 folio_batch_add(fbatch, folio);
2494 return 0;
2495error:
2496 filemap_invalidate_unlock_shared(mapping);
2497 folio_put(folio);
2498 return error;
2499}
2500
2501static int filemap_readahead(struct kiocb *iocb, struct file *file,
2502 struct address_space *mapping, struct folio *folio,
2503 pgoff_t last_index)
2504{
2505 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2506
2507 if (iocb->ki_flags & IOCB_NOIO)
2508 return -EAGAIN;
2509 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2510 return 0;
2511}
2512
2513static int filemap_get_pages(struct kiocb *iocb, size_t count,
2514 struct folio_batch *fbatch, bool need_uptodate)
2515{
2516 struct file *filp = iocb->ki_filp;
2517 struct address_space *mapping = filp->f_mapping;
2518 struct file_ra_state *ra = &filp->f_ra;
2519 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2520 pgoff_t last_index;
2521 struct folio *folio;
2522 unsigned int flags;
2523 int err = 0;
2524
2525 /* "last_index" is the index of the page beyond the end of the read */
2526 last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
2527retry:
2528 if (fatal_signal_pending(current))
2529 return -EINTR;
2530
2531 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2532 if (!folio_batch_count(fbatch)) {
2533 if (iocb->ki_flags & IOCB_NOIO)
2534 return -EAGAIN;
2535 if (iocb->ki_flags & IOCB_NOWAIT)
2536 flags = memalloc_noio_save();
2537 page_cache_sync_readahead(mapping, ra, filp, index,
2538 last_index - index);
2539 if (iocb->ki_flags & IOCB_NOWAIT)
2540 memalloc_noio_restore(flags);
2541 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2542 }
2543 if (!folio_batch_count(fbatch)) {
2544 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2545 return -EAGAIN;
2546 err = filemap_create_folio(filp, mapping, iocb->ki_pos, fbatch);
2547 if (err == AOP_TRUNCATED_PAGE)
2548 goto retry;
2549 return err;
2550 }
2551
2552 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2553 if (folio_test_readahead(folio)) {
2554 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2555 if (err)
2556 goto err;
2557 }
2558 if (!folio_test_uptodate(folio)) {
2559 if ((iocb->ki_flags & IOCB_WAITQ) &&
2560 folio_batch_count(fbatch) > 1)
2561 iocb->ki_flags |= IOCB_NOWAIT;
2562 err = filemap_update_page(iocb, mapping, count, folio,
2563 need_uptodate);
2564 if (err)
2565 goto err;
2566 }
2567
2568 trace_mm_filemap_get_pages(mapping, index, last_index - 1);
2569 return 0;
2570err:
2571 if (err < 0)
2572 folio_put(folio);
2573 if (likely(--fbatch->nr))
2574 return 0;
2575 if (err == AOP_TRUNCATED_PAGE)
2576 goto retry;
2577 return err;
2578}
2579
2580static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2581{
2582 unsigned int shift = folio_shift(folio);
2583
2584 return (pos1 >> shift == pos2 >> shift);
2585}
2586
2587/**
2588 * filemap_read - Read data from the page cache.
2589 * @iocb: The iocb to read.
2590 * @iter: Destination for the data.
2591 * @already_read: Number of bytes already read by the caller.
2592 *
2593 * Copies data from the page cache. If the data is not currently present,
2594 * uses the readahead and read_folio address_space operations to fetch it.
2595 *
2596 * Return: Total number of bytes copied, including those already read by
2597 * the caller. If an error happens before any bytes are copied, returns
2598 * a negative error number.
2599 */
2600ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2601 ssize_t already_read)
2602{
2603 struct file *filp = iocb->ki_filp;
2604 struct file_ra_state *ra = &filp->f_ra;
2605 struct address_space *mapping = filp->f_mapping;
2606 struct inode *inode = mapping->host;
2607 struct folio_batch fbatch;
2608 int i, error = 0;
2609 bool writably_mapped;
2610 loff_t isize, end_offset;
2611 loff_t last_pos = ra->prev_pos;
2612
2613 if (unlikely(iocb->ki_pos < 0))
2614 return -EINVAL;
2615 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2616 return 0;
2617 if (unlikely(!iov_iter_count(iter)))
2618 return 0;
2619
2620 iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos);
2621 folio_batch_init(&fbatch);
2622
2623 do {
2624 cond_resched();
2625
2626 /*
2627 * If we've already successfully copied some data, then we
2628 * can no longer safely return -EIOCBQUEUED. Hence mark
2629 * an async read NOWAIT at that point.
2630 */
2631 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2632 iocb->ki_flags |= IOCB_NOWAIT;
2633
2634 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2635 break;
2636
2637 error = filemap_get_pages(iocb, iter->count, &fbatch, false);
2638 if (error < 0)
2639 break;
2640
2641 /*
2642 * i_size must be checked after we know the pages are Uptodate.
2643 *
2644 * Checking i_size after the check allows us to calculate
2645 * the correct value for "nr", which means the zero-filled
2646 * part of the page is not copied back to userspace (unless
2647 * another truncate extends the file - this is desired though).
2648 */
2649 isize = i_size_read(inode);
2650 if (unlikely(iocb->ki_pos >= isize))
2651 goto put_folios;
2652 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2653
2654 /*
2655 * Once we start copying data, we don't want to be touching any
2656 * cachelines that might be contended:
2657 */
2658 writably_mapped = mapping_writably_mapped(mapping);
2659
2660 /*
2661 * When a read accesses the same folio several times, only
2662 * mark it as accessed the first time.
2663 */
2664 if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
2665 fbatch.folios[0]))
2666 folio_mark_accessed(fbatch.folios[0]);
2667
2668 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2669 struct folio *folio = fbatch.folios[i];
2670 size_t fsize = folio_size(folio);
2671 size_t offset = iocb->ki_pos & (fsize - 1);
2672 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2673 fsize - offset);
2674 size_t copied;
2675
2676 if (end_offset < folio_pos(folio))
2677 break;
2678 if (i > 0)
2679 folio_mark_accessed(folio);
2680 /*
2681 * If users can be writing to this folio using arbitrary
2682 * virtual addresses, take care of potential aliasing
2683 * before reading the folio on the kernel side.
2684 */
2685 if (writably_mapped)
2686 flush_dcache_folio(folio);
2687
2688 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2689
2690 already_read += copied;
2691 iocb->ki_pos += copied;
2692 last_pos = iocb->ki_pos;
2693
2694 if (copied < bytes) {
2695 error = -EFAULT;
2696 break;
2697 }
2698 }
2699put_folios:
2700 for (i = 0; i < folio_batch_count(&fbatch); i++)
2701 folio_put(fbatch.folios[i]);
2702 folio_batch_init(&fbatch);
2703 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2704
2705 file_accessed(filp);
2706 ra->prev_pos = last_pos;
2707 return already_read ? already_read : error;
2708}
2709EXPORT_SYMBOL_GPL(filemap_read);
2710
2711int kiocb_write_and_wait(struct kiocb *iocb, size_t count)
2712{
2713 struct address_space *mapping = iocb->ki_filp->f_mapping;
2714 loff_t pos = iocb->ki_pos;
2715 loff_t end = pos + count - 1;
2716
2717 if (iocb->ki_flags & IOCB_NOWAIT) {
2718 if (filemap_range_needs_writeback(mapping, pos, end))
2719 return -EAGAIN;
2720 return 0;
2721 }
2722
2723 return filemap_write_and_wait_range(mapping, pos, end);
2724}
2725EXPORT_SYMBOL_GPL(kiocb_write_and_wait);
2726
2727int filemap_invalidate_pages(struct address_space *mapping,
2728 loff_t pos, loff_t end, bool nowait)
2729{
2730 int ret;
2731
2732 if (nowait) {
2733 /* we could block if there are any pages in the range */
2734 if (filemap_range_has_page(mapping, pos, end))
2735 return -EAGAIN;
2736 } else {
2737 ret = filemap_write_and_wait_range(mapping, pos, end);
2738 if (ret)
2739 return ret;
2740 }
2741
2742 /*
2743 * After a write we want buffered reads to be sure to go to disk to get
2744 * the new data. We invalidate clean cached page from the region we're
2745 * about to write. We do this *before* the write so that we can return
2746 * without clobbering -EIOCBQUEUED from ->direct_IO().
2747 */
2748 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
2749 end >> PAGE_SHIFT);
2750}
2751
2752int kiocb_invalidate_pages(struct kiocb *iocb, size_t count)
2753{
2754 struct address_space *mapping = iocb->ki_filp->f_mapping;
2755
2756 return filemap_invalidate_pages(mapping, iocb->ki_pos,
2757 iocb->ki_pos + count - 1,
2758 iocb->ki_flags & IOCB_NOWAIT);
2759}
2760EXPORT_SYMBOL_GPL(kiocb_invalidate_pages);
2761
2762/**
2763 * generic_file_read_iter - generic filesystem read routine
2764 * @iocb: kernel I/O control block
2765 * @iter: destination for the data read
2766 *
2767 * This is the "read_iter()" routine for all filesystems
2768 * that can use the page cache directly.
2769 *
2770 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2771 * be returned when no data can be read without waiting for I/O requests
2772 * to complete; it doesn't prevent readahead.
2773 *
2774 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2775 * requests shall be made for the read or for readahead. When no data
2776 * can be read, -EAGAIN shall be returned. When readahead would be
2777 * triggered, a partial, possibly empty read shall be returned.
2778 *
2779 * Return:
2780 * * number of bytes copied, even for partial reads
2781 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2782 */
2783ssize_t
2784generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2785{
2786 size_t count = iov_iter_count(iter);
2787 ssize_t retval = 0;
2788
2789 if (!count)
2790 return 0; /* skip atime */
2791
2792 if (iocb->ki_flags & IOCB_DIRECT) {
2793 struct file *file = iocb->ki_filp;
2794 struct address_space *mapping = file->f_mapping;
2795 struct inode *inode = mapping->host;
2796
2797 retval = kiocb_write_and_wait(iocb, count);
2798 if (retval < 0)
2799 return retval;
2800 file_accessed(file);
2801
2802 retval = mapping->a_ops->direct_IO(iocb, iter);
2803 if (retval >= 0) {
2804 iocb->ki_pos += retval;
2805 count -= retval;
2806 }
2807 if (retval != -EIOCBQUEUED)
2808 iov_iter_revert(iter, count - iov_iter_count(iter));
2809
2810 /*
2811 * Btrfs can have a short DIO read if we encounter
2812 * compressed extents, so if there was an error, or if
2813 * we've already read everything we wanted to, or if
2814 * there was a short read because we hit EOF, go ahead
2815 * and return. Otherwise fallthrough to buffered io for
2816 * the rest of the read. Buffered reads will not work for
2817 * DAX files, so don't bother trying.
2818 */
2819 if (retval < 0 || !count || IS_DAX(inode))
2820 return retval;
2821 if (iocb->ki_pos >= i_size_read(inode))
2822 return retval;
2823 }
2824
2825 return filemap_read(iocb, iter, retval);
2826}
2827EXPORT_SYMBOL(generic_file_read_iter);
2828
2829/*
2830 * Splice subpages from a folio into a pipe.
2831 */
2832size_t splice_folio_into_pipe(struct pipe_inode_info *pipe,
2833 struct folio *folio, loff_t fpos, size_t size)
2834{
2835 struct page *page;
2836 size_t spliced = 0, offset = offset_in_folio(folio, fpos);
2837
2838 page = folio_page(folio, offset / PAGE_SIZE);
2839 size = min(size, folio_size(folio) - offset);
2840 offset %= PAGE_SIZE;
2841
2842 while (spliced < size &&
2843 !pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
2844 struct pipe_buffer *buf = pipe_head_buf(pipe);
2845 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced);
2846
2847 *buf = (struct pipe_buffer) {
2848 .ops = &page_cache_pipe_buf_ops,
2849 .page = page,
2850 .offset = offset,
2851 .len = part,
2852 };
2853 folio_get(folio);
2854 pipe->head++;
2855 page++;
2856 spliced += part;
2857 offset = 0;
2858 }
2859
2860 return spliced;
2861}
2862
2863/**
2864 * filemap_splice_read - Splice data from a file's pagecache into a pipe
2865 * @in: The file to read from
2866 * @ppos: Pointer to the file position to read from
2867 * @pipe: The pipe to splice into
2868 * @len: The amount to splice
2869 * @flags: The SPLICE_F_* flags
2870 *
2871 * This function gets folios from a file's pagecache and splices them into the
2872 * pipe. Readahead will be called as necessary to fill more folios. This may
2873 * be used for blockdevs also.
2874 *
2875 * Return: On success, the number of bytes read will be returned and *@ppos
2876 * will be updated if appropriate; 0 will be returned if there is no more data
2877 * to be read; -EAGAIN will be returned if the pipe had no space, and some
2878 * other negative error code will be returned on error. A short read may occur
2879 * if the pipe has insufficient space, we reach the end of the data or we hit a
2880 * hole.
2881 */
2882ssize_t filemap_splice_read(struct file *in, loff_t *ppos,
2883 struct pipe_inode_info *pipe,
2884 size_t len, unsigned int flags)
2885{
2886 struct folio_batch fbatch;
2887 struct kiocb iocb;
2888 size_t total_spliced = 0, used, npages;
2889 loff_t isize, end_offset;
2890 bool writably_mapped;
2891 int i, error = 0;
2892
2893 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes))
2894 return 0;
2895
2896 init_sync_kiocb(&iocb, in);
2897 iocb.ki_pos = *ppos;
2898
2899 /* Work out how much data we can actually add into the pipe */
2900 used = pipe_occupancy(pipe->head, pipe->tail);
2901 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2902 len = min_t(size_t, len, npages * PAGE_SIZE);
2903
2904 folio_batch_init(&fbatch);
2905
2906 do {
2907 cond_resched();
2908
2909 if (*ppos >= i_size_read(in->f_mapping->host))
2910 break;
2911
2912 iocb.ki_pos = *ppos;
2913 error = filemap_get_pages(&iocb, len, &fbatch, true);
2914 if (error < 0)
2915 break;
2916
2917 /*
2918 * i_size must be checked after we know the pages are Uptodate.
2919 *
2920 * Checking i_size after the check allows us to calculate
2921 * the correct value for "nr", which means the zero-filled
2922 * part of the page is not copied back to userspace (unless
2923 * another truncate extends the file - this is desired though).
2924 */
2925 isize = i_size_read(in->f_mapping->host);
2926 if (unlikely(*ppos >= isize))
2927 break;
2928 end_offset = min_t(loff_t, isize, *ppos + len);
2929
2930 /*
2931 * Once we start copying data, we don't want to be touching any
2932 * cachelines that might be contended:
2933 */
2934 writably_mapped = mapping_writably_mapped(in->f_mapping);
2935
2936 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2937 struct folio *folio = fbatch.folios[i];
2938 size_t n;
2939
2940 if (folio_pos(folio) >= end_offset)
2941 goto out;
2942 folio_mark_accessed(folio);
2943
2944 /*
2945 * If users can be writing to this folio using arbitrary
2946 * virtual addresses, take care of potential aliasing
2947 * before reading the folio on the kernel side.
2948 */
2949 if (writably_mapped)
2950 flush_dcache_folio(folio);
2951
2952 n = min_t(loff_t, len, isize - *ppos);
2953 n = splice_folio_into_pipe(pipe, folio, *ppos, n);
2954 if (!n)
2955 goto out;
2956 len -= n;
2957 total_spliced += n;
2958 *ppos += n;
2959 in->f_ra.prev_pos = *ppos;
2960 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
2961 goto out;
2962 }
2963
2964 folio_batch_release(&fbatch);
2965 } while (len);
2966
2967out:
2968 folio_batch_release(&fbatch);
2969 file_accessed(in);
2970
2971 return total_spliced ? total_spliced : error;
2972}
2973EXPORT_SYMBOL(filemap_splice_read);
2974
2975static inline loff_t folio_seek_hole_data(struct xa_state *xas,
2976 struct address_space *mapping, struct folio *folio,
2977 loff_t start, loff_t end, bool seek_data)
2978{
2979 const struct address_space_operations *ops = mapping->a_ops;
2980 size_t offset, bsz = i_blocksize(mapping->host);
2981
2982 if (xa_is_value(folio) || folio_test_uptodate(folio))
2983 return seek_data ? start : end;
2984 if (!ops->is_partially_uptodate)
2985 return seek_data ? end : start;
2986
2987 xas_pause(xas);
2988 rcu_read_unlock();
2989 folio_lock(folio);
2990 if (unlikely(folio->mapping != mapping))
2991 goto unlock;
2992
2993 offset = offset_in_folio(folio, start) & ~(bsz - 1);
2994
2995 do {
2996 if (ops->is_partially_uptodate(folio, offset, bsz) ==
2997 seek_data)
2998 break;
2999 start = (start + bsz) & ~((u64)bsz - 1);
3000 offset += bsz;
3001 } while (offset < folio_size(folio));
3002unlock:
3003 folio_unlock(folio);
3004 rcu_read_lock();
3005 return start;
3006}
3007
3008static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
3009{
3010 if (xa_is_value(folio))
3011 return PAGE_SIZE << xas_get_order(xas);
3012 return folio_size(folio);
3013}
3014
3015/**
3016 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3017 * @mapping: Address space to search.
3018 * @start: First byte to consider.
3019 * @end: Limit of search (exclusive).
3020 * @whence: Either SEEK_HOLE or SEEK_DATA.
3021 *
3022 * If the page cache knows which blocks contain holes and which blocks
3023 * contain data, your filesystem can use this function to implement
3024 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
3025 * entirely memory-based such as tmpfs, and filesystems which support
3026 * unwritten extents.
3027 *
3028 * Return: The requested offset on success, or -ENXIO if @whence specifies
3029 * SEEK_DATA and there is no data after @start. There is an implicit hole
3030 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
3031 * and @end contain data.
3032 */
3033loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
3034 loff_t end, int whence)
3035{
3036 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
3037 pgoff_t max = (end - 1) >> PAGE_SHIFT;
3038 bool seek_data = (whence == SEEK_DATA);
3039 struct folio *folio;
3040
3041 if (end <= start)
3042 return -ENXIO;
3043
3044 rcu_read_lock();
3045 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
3046 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
3047 size_t seek_size;
3048
3049 if (start < pos) {
3050 if (!seek_data)
3051 goto unlock;
3052 start = pos;
3053 }
3054
3055 seek_size = seek_folio_size(&xas, folio);
3056 pos = round_up((u64)pos + 1, seek_size);
3057 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
3058 seek_data);
3059 if (start < pos)
3060 goto unlock;
3061 if (start >= end)
3062 break;
3063 if (seek_size > PAGE_SIZE)
3064 xas_set(&xas, pos >> PAGE_SHIFT);
3065 if (!xa_is_value(folio))
3066 folio_put(folio);
3067 }
3068 if (seek_data)
3069 start = -ENXIO;
3070unlock:
3071 rcu_read_unlock();
3072 if (folio && !xa_is_value(folio))
3073 folio_put(folio);
3074 if (start > end)
3075 return end;
3076 return start;
3077}
3078
3079#ifdef CONFIG_MMU
3080#define MMAP_LOTSAMISS (100)
3081/*
3082 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
3083 * @vmf - the vm_fault for this fault.
3084 * @folio - the folio to lock.
3085 * @fpin - the pointer to the file we may pin (or is already pinned).
3086 *
3087 * This works similar to lock_folio_or_retry in that it can drop the
3088 * mmap_lock. It differs in that it actually returns the folio locked
3089 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
3090 * to drop the mmap_lock then fpin will point to the pinned file and
3091 * needs to be fput()'ed at a later point.
3092 */
3093static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
3094 struct file **fpin)
3095{
3096 if (folio_trylock(folio))
3097 return 1;
3098
3099 /*
3100 * NOTE! This will make us return with VM_FAULT_RETRY, but with
3101 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
3102 * is supposed to work. We have way too many special cases..
3103 */
3104 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
3105 return 0;
3106
3107 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
3108 if (vmf->flags & FAULT_FLAG_KILLABLE) {
3109 if (__folio_lock_killable(folio)) {
3110 /*
3111 * We didn't have the right flags to drop the
3112 * fault lock, but all fault_handlers only check
3113 * for fatal signals if we return VM_FAULT_RETRY,
3114 * so we need to drop the fault lock here and
3115 * return 0 if we don't have a fpin.
3116 */
3117 if (*fpin == NULL)
3118 release_fault_lock(vmf);
3119 return 0;
3120 }
3121 } else
3122 __folio_lock(folio);
3123
3124 return 1;
3125}
3126
3127/*
3128 * Synchronous readahead happens when we don't even find a page in the page
3129 * cache at all. We don't want to perform IO under the mmap sem, so if we have
3130 * to drop the mmap sem we return the file that was pinned in order for us to do
3131 * that. If we didn't pin a file then we return NULL. The file that is
3132 * returned needs to be fput()'ed when we're done with it.
3133 */
3134static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3135{
3136 struct file *file = vmf->vma->vm_file;
3137 struct file_ra_state *ra = &file->f_ra;
3138 struct address_space *mapping = file->f_mapping;
3139 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3140 struct file *fpin = NULL;
3141 unsigned long vm_flags = vmf->vma->vm_flags;
3142 unsigned int mmap_miss;
3143
3144#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3145 /* Use the readahead code, even if readahead is disabled */
3146 if ((vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER) {
3147 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3148 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3149 ra->size = HPAGE_PMD_NR;
3150 /*
3151 * Fetch two PMD folios, so we get the chance to actually
3152 * readahead, unless we've been told not to.
3153 */
3154 if (!(vm_flags & VM_RAND_READ))
3155 ra->size *= 2;
3156 ra->async_size = HPAGE_PMD_NR;
3157 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3158 return fpin;
3159 }
3160#endif
3161
3162 /* If we don't want any read-ahead, don't bother */
3163 if (vm_flags & VM_RAND_READ)
3164 return fpin;
3165 if (!ra->ra_pages)
3166 return fpin;
3167
3168 if (vm_flags & VM_SEQ_READ) {
3169 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3170 page_cache_sync_ra(&ractl, ra->ra_pages);
3171 return fpin;
3172 }
3173
3174 /* Avoid banging the cache line if not needed */
3175 mmap_miss = READ_ONCE(ra->mmap_miss);
3176 if (mmap_miss < MMAP_LOTSAMISS * 10)
3177 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3178
3179 /*
3180 * Do we miss much more than hit in this file? If so,
3181 * stop bothering with read-ahead. It will only hurt.
3182 */
3183 if (mmap_miss > MMAP_LOTSAMISS)
3184 return fpin;
3185
3186 /*
3187 * mmap read-around
3188 */
3189 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3190 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3191 ra->size = ra->ra_pages;
3192 ra->async_size = ra->ra_pages / 4;
3193 ractl._index = ra->start;
3194 page_cache_ra_order(&ractl, ra, 0);
3195 return fpin;
3196}
3197
3198/*
3199 * Asynchronous readahead happens when we find the page and PG_readahead,
3200 * so we want to possibly extend the readahead further. We return the file that
3201 * was pinned if we have to drop the mmap_lock in order to do IO.
3202 */
3203static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3204 struct folio *folio)
3205{
3206 struct file *file = vmf->vma->vm_file;
3207 struct file_ra_state *ra = &file->f_ra;
3208 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3209 struct file *fpin = NULL;
3210 unsigned int mmap_miss;
3211
3212 /* If we don't want any read-ahead, don't bother */
3213 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3214 return fpin;
3215
3216 mmap_miss = READ_ONCE(ra->mmap_miss);
3217 if (mmap_miss)
3218 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3219
3220 if (folio_test_readahead(folio)) {
3221 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3222 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3223 }
3224 return fpin;
3225}
3226
3227static vm_fault_t filemap_fault_recheck_pte_none(struct vm_fault *vmf)
3228{
3229 struct vm_area_struct *vma = vmf->vma;
3230 vm_fault_t ret = 0;
3231 pte_t *ptep;
3232
3233 /*
3234 * We might have COW'ed a pagecache folio and might now have an mlocked
3235 * anon folio mapped. The original pagecache folio is not mlocked and
3236 * might have been evicted. During a read+clear/modify/write update of
3237 * the PTE, such as done in do_numa_page()/change_pte_range(), we
3238 * temporarily clear the PTE under PT lock and might detect it here as
3239 * "none" when not holding the PT lock.
3240 *
3241 * Not rechecking the PTE under PT lock could result in an unexpected
3242 * major fault in an mlock'ed region. Recheck only for this special
3243 * scenario while holding the PT lock, to not degrade non-mlocked
3244 * scenarios. Recheck the PTE without PT lock firstly, thereby reducing
3245 * the number of times we hold PT lock.
3246 */
3247 if (!(vma->vm_flags & VM_LOCKED))
3248 return 0;
3249
3250 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
3251 return 0;
3252
3253 ptep = pte_offset_map_ro_nolock(vma->vm_mm, vmf->pmd, vmf->address,
3254 &vmf->ptl);
3255 if (unlikely(!ptep))
3256 return VM_FAULT_NOPAGE;
3257
3258 if (unlikely(!pte_none(ptep_get_lockless(ptep)))) {
3259 ret = VM_FAULT_NOPAGE;
3260 } else {
3261 spin_lock(vmf->ptl);
3262 if (unlikely(!pte_none(ptep_get(ptep))))
3263 ret = VM_FAULT_NOPAGE;
3264 spin_unlock(vmf->ptl);
3265 }
3266 pte_unmap(ptep);
3267 return ret;
3268}
3269
3270/**
3271 * filemap_fault - read in file data for page fault handling
3272 * @vmf: struct vm_fault containing details of the fault
3273 *
3274 * filemap_fault() is invoked via the vma operations vector for a
3275 * mapped memory region to read in file data during a page fault.
3276 *
3277 * The goto's are kind of ugly, but this streamlines the normal case of having
3278 * it in the page cache, and handles the special cases reasonably without
3279 * having a lot of duplicated code.
3280 *
3281 * vma->vm_mm->mmap_lock must be held on entry.
3282 *
3283 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3284 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3285 *
3286 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3287 * has not been released.
3288 *
3289 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3290 *
3291 * Return: bitwise-OR of %VM_FAULT_ codes.
3292 */
3293vm_fault_t filemap_fault(struct vm_fault *vmf)
3294{
3295 int error;
3296 struct file *file = vmf->vma->vm_file;
3297 struct file *fpin = NULL;
3298 struct address_space *mapping = file->f_mapping;
3299 struct inode *inode = mapping->host;
3300 pgoff_t max_idx, index = vmf->pgoff;
3301 struct folio *folio;
3302 vm_fault_t ret = 0;
3303 bool mapping_locked = false;
3304
3305 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3306 if (unlikely(index >= max_idx))
3307 return VM_FAULT_SIGBUS;
3308
3309 trace_mm_filemap_fault(mapping, index);
3310
3311 /*
3312 * Do we have something in the page cache already?
3313 */
3314 folio = filemap_get_folio(mapping, index);
3315 if (likely(!IS_ERR(folio))) {
3316 /*
3317 * We found the page, so try async readahead before waiting for
3318 * the lock.
3319 */
3320 if (!(vmf->flags & FAULT_FLAG_TRIED))
3321 fpin = do_async_mmap_readahead(vmf, folio);
3322 if (unlikely(!folio_test_uptodate(folio))) {
3323 filemap_invalidate_lock_shared(mapping);
3324 mapping_locked = true;
3325 }
3326 } else {
3327 ret = filemap_fault_recheck_pte_none(vmf);
3328 if (unlikely(ret))
3329 return ret;
3330
3331 /* No page in the page cache at all */
3332 count_vm_event(PGMAJFAULT);
3333 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3334 ret = VM_FAULT_MAJOR;
3335 fpin = do_sync_mmap_readahead(vmf);
3336retry_find:
3337 /*
3338 * See comment in filemap_create_folio() why we need
3339 * invalidate_lock
3340 */
3341 if (!mapping_locked) {
3342 filemap_invalidate_lock_shared(mapping);
3343 mapping_locked = true;
3344 }
3345 folio = __filemap_get_folio(mapping, index,
3346 FGP_CREAT|FGP_FOR_MMAP,
3347 vmf->gfp_mask);
3348 if (IS_ERR(folio)) {
3349 if (fpin)
3350 goto out_retry;
3351 filemap_invalidate_unlock_shared(mapping);
3352 return VM_FAULT_OOM;
3353 }
3354 }
3355
3356 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3357 goto out_retry;
3358
3359 /* Did it get truncated? */
3360 if (unlikely(folio->mapping != mapping)) {
3361 folio_unlock(folio);
3362 folio_put(folio);
3363 goto retry_find;
3364 }
3365 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3366
3367 /*
3368 * We have a locked folio in the page cache, now we need to check
3369 * that it's up-to-date. If not, it is going to be due to an error,
3370 * or because readahead was otherwise unable to retrieve it.
3371 */
3372 if (unlikely(!folio_test_uptodate(folio))) {
3373 /*
3374 * If the invalidate lock is not held, the folio was in cache
3375 * and uptodate and now it is not. Strange but possible since we
3376 * didn't hold the page lock all the time. Let's drop
3377 * everything, get the invalidate lock and try again.
3378 */
3379 if (!mapping_locked) {
3380 folio_unlock(folio);
3381 folio_put(folio);
3382 goto retry_find;
3383 }
3384
3385 /*
3386 * OK, the folio is really not uptodate. This can be because the
3387 * VMA has the VM_RAND_READ flag set, or because an error
3388 * arose. Let's read it in directly.
3389 */
3390 goto page_not_uptodate;
3391 }
3392
3393 /*
3394 * We've made it this far and we had to drop our mmap_lock, now is the
3395 * time to return to the upper layer and have it re-find the vma and
3396 * redo the fault.
3397 */
3398 if (fpin) {
3399 folio_unlock(folio);
3400 goto out_retry;
3401 }
3402 if (mapping_locked)
3403 filemap_invalidate_unlock_shared(mapping);
3404
3405 /*
3406 * Found the page and have a reference on it.
3407 * We must recheck i_size under page lock.
3408 */
3409 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3410 if (unlikely(index >= max_idx)) {
3411 folio_unlock(folio);
3412 folio_put(folio);
3413 return VM_FAULT_SIGBUS;
3414 }
3415
3416 vmf->page = folio_file_page(folio, index);
3417 return ret | VM_FAULT_LOCKED;
3418
3419page_not_uptodate:
3420 /*
3421 * Umm, take care of errors if the page isn't up-to-date.
3422 * Try to re-read it _once_. We do this synchronously,
3423 * because there really aren't any performance issues here
3424 * and we need to check for errors.
3425 */
3426 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3427 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3428 if (fpin)
3429 goto out_retry;
3430 folio_put(folio);
3431
3432 if (!error || error == AOP_TRUNCATED_PAGE)
3433 goto retry_find;
3434 filemap_invalidate_unlock_shared(mapping);
3435
3436 return VM_FAULT_SIGBUS;
3437
3438out_retry:
3439 /*
3440 * We dropped the mmap_lock, we need to return to the fault handler to
3441 * re-find the vma and come back and find our hopefully still populated
3442 * page.
3443 */
3444 if (!IS_ERR(folio))
3445 folio_put(folio);
3446 if (mapping_locked)
3447 filemap_invalidate_unlock_shared(mapping);
3448 if (fpin)
3449 fput(fpin);
3450 return ret | VM_FAULT_RETRY;
3451}
3452EXPORT_SYMBOL(filemap_fault);
3453
3454static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio,
3455 pgoff_t start)
3456{
3457 struct mm_struct *mm = vmf->vma->vm_mm;
3458
3459 /* Huge page is mapped? No need to proceed. */
3460 if (pmd_trans_huge(*vmf->pmd)) {
3461 folio_unlock(folio);
3462 folio_put(folio);
3463 return true;
3464 }
3465
3466 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) {
3467 struct page *page = folio_file_page(folio, start);
3468 vm_fault_t ret = do_set_pmd(vmf, page);
3469 if (!ret) {
3470 /* The page is mapped successfully, reference consumed. */
3471 folio_unlock(folio);
3472 return true;
3473 }
3474 }
3475
3476 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte)
3477 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3478
3479 return false;
3480}
3481
3482static struct folio *next_uptodate_folio(struct xa_state *xas,
3483 struct address_space *mapping, pgoff_t end_pgoff)
3484{
3485 struct folio *folio = xas_next_entry(xas, end_pgoff);
3486 unsigned long max_idx;
3487
3488 do {
3489 if (!folio)
3490 return NULL;
3491 if (xas_retry(xas, folio))
3492 continue;
3493 if (xa_is_value(folio))
3494 continue;
3495 if (!folio_try_get(folio))
3496 continue;
3497 if (folio_test_locked(folio))
3498 goto skip;
3499 /* Has the page moved or been split? */
3500 if (unlikely(folio != xas_reload(xas)))
3501 goto skip;
3502 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3503 goto skip;
3504 if (!folio_trylock(folio))
3505 goto skip;
3506 if (folio->mapping != mapping)
3507 goto unlock;
3508 if (!folio_test_uptodate(folio))
3509 goto unlock;
3510 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3511 if (xas->xa_index >= max_idx)
3512 goto unlock;
3513 return folio;
3514unlock:
3515 folio_unlock(folio);
3516skip:
3517 folio_put(folio);
3518 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3519
3520 return NULL;
3521}
3522
3523/*
3524 * Map page range [start_page, start_page + nr_pages) of folio.
3525 * start_page is gotten from start by folio_page(folio, start)
3526 */
3527static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
3528 struct folio *folio, unsigned long start,
3529 unsigned long addr, unsigned int nr_pages,
3530 unsigned long *rss, unsigned int *mmap_miss)
3531{
3532 vm_fault_t ret = 0;
3533 struct page *page = folio_page(folio, start);
3534 unsigned int count = 0;
3535 pte_t *old_ptep = vmf->pte;
3536
3537 do {
3538 if (PageHWPoison(page + count))
3539 goto skip;
3540
3541 /*
3542 * If there are too many folios that are recently evicted
3543 * in a file, they will probably continue to be evicted.
3544 * In such situation, read-ahead is only a waste of IO.
3545 * Don't decrease mmap_miss in this scenario to make sure
3546 * we can stop read-ahead.
3547 */
3548 if (!folio_test_workingset(folio))
3549 (*mmap_miss)++;
3550
3551 /*
3552 * NOTE: If there're PTE markers, we'll leave them to be
3553 * handled in the specific fault path, and it'll prohibit the
3554 * fault-around logic.
3555 */
3556 if (!pte_none(ptep_get(&vmf->pte[count])))
3557 goto skip;
3558
3559 count++;
3560 continue;
3561skip:
3562 if (count) {
3563 set_pte_range(vmf, folio, page, count, addr);
3564 *rss += count;
3565 folio_ref_add(folio, count);
3566 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3567 ret = VM_FAULT_NOPAGE;
3568 }
3569
3570 count++;
3571 page += count;
3572 vmf->pte += count;
3573 addr += count * PAGE_SIZE;
3574 count = 0;
3575 } while (--nr_pages > 0);
3576
3577 if (count) {
3578 set_pte_range(vmf, folio, page, count, addr);
3579 *rss += count;
3580 folio_ref_add(folio, count);
3581 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3582 ret = VM_FAULT_NOPAGE;
3583 }
3584
3585 vmf->pte = old_ptep;
3586
3587 return ret;
3588}
3589
3590static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
3591 struct folio *folio, unsigned long addr,
3592 unsigned long *rss, unsigned int *mmap_miss)
3593{
3594 vm_fault_t ret = 0;
3595 struct page *page = &folio->page;
3596
3597 if (PageHWPoison(page))
3598 return ret;
3599
3600 /* See comment of filemap_map_folio_range() */
3601 if (!folio_test_workingset(folio))
3602 (*mmap_miss)++;
3603
3604 /*
3605 * NOTE: If there're PTE markers, we'll leave them to be
3606 * handled in the specific fault path, and it'll prohibit
3607 * the fault-around logic.
3608 */
3609 if (!pte_none(ptep_get(vmf->pte)))
3610 return ret;
3611
3612 if (vmf->address == addr)
3613 ret = VM_FAULT_NOPAGE;
3614
3615 set_pte_range(vmf, folio, page, 1, addr);
3616 (*rss)++;
3617 folio_ref_inc(folio);
3618
3619 return ret;
3620}
3621
3622vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3623 pgoff_t start_pgoff, pgoff_t end_pgoff)
3624{
3625 struct vm_area_struct *vma = vmf->vma;
3626 struct file *file = vma->vm_file;
3627 struct address_space *mapping = file->f_mapping;
3628 pgoff_t file_end, last_pgoff = start_pgoff;
3629 unsigned long addr;
3630 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3631 struct folio *folio;
3632 vm_fault_t ret = 0;
3633 unsigned long rss = 0;
3634 unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type;
3635
3636 rcu_read_lock();
3637 folio = next_uptodate_folio(&xas, mapping, end_pgoff);
3638 if (!folio)
3639 goto out;
3640
3641 if (filemap_map_pmd(vmf, folio, start_pgoff)) {
3642 ret = VM_FAULT_NOPAGE;
3643 goto out;
3644 }
3645
3646 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3647 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3648 if (!vmf->pte) {
3649 folio_unlock(folio);
3650 folio_put(folio);
3651 goto out;
3652 }
3653
3654 file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
3655 if (end_pgoff > file_end)
3656 end_pgoff = file_end;
3657
3658 folio_type = mm_counter_file(folio);
3659 do {
3660 unsigned long end;
3661
3662 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3663 vmf->pte += xas.xa_index - last_pgoff;
3664 last_pgoff = xas.xa_index;
3665 end = folio_next_index(folio) - 1;
3666 nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
3667
3668 if (!folio_test_large(folio))
3669 ret |= filemap_map_order0_folio(vmf,
3670 folio, addr, &rss, &mmap_miss);
3671 else
3672 ret |= filemap_map_folio_range(vmf, folio,
3673 xas.xa_index - folio->index, addr,
3674 nr_pages, &rss, &mmap_miss);
3675
3676 folio_unlock(folio);
3677 folio_put(folio);
3678 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL);
3679 add_mm_counter(vma->vm_mm, folio_type, rss);
3680 pte_unmap_unlock(vmf->pte, vmf->ptl);
3681 trace_mm_filemap_map_pages(mapping, start_pgoff, end_pgoff);
3682out:
3683 rcu_read_unlock();
3684
3685 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss);
3686 if (mmap_miss >= mmap_miss_saved)
3687 WRITE_ONCE(file->f_ra.mmap_miss, 0);
3688 else
3689 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss);
3690
3691 return ret;
3692}
3693EXPORT_SYMBOL(filemap_map_pages);
3694
3695vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3696{
3697 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3698 struct folio *folio = page_folio(vmf->page);
3699 vm_fault_t ret = VM_FAULT_LOCKED;
3700
3701 sb_start_pagefault(mapping->host->i_sb);
3702 file_update_time(vmf->vma->vm_file);
3703 folio_lock(folio);
3704 if (folio->mapping != mapping) {
3705 folio_unlock(folio);
3706 ret = VM_FAULT_NOPAGE;
3707 goto out;
3708 }
3709 /*
3710 * We mark the folio dirty already here so that when freeze is in
3711 * progress, we are guaranteed that writeback during freezing will
3712 * see the dirty folio and writeprotect it again.
3713 */
3714 folio_mark_dirty(folio);
3715 folio_wait_stable(folio);
3716out:
3717 sb_end_pagefault(mapping->host->i_sb);
3718 return ret;
3719}
3720
3721const struct vm_operations_struct generic_file_vm_ops = {
3722 .fault = filemap_fault,
3723 .map_pages = filemap_map_pages,
3724 .page_mkwrite = filemap_page_mkwrite,
3725};
3726
3727/* This is used for a general mmap of a disk file */
3728
3729int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3730{
3731 struct address_space *mapping = file->f_mapping;
3732
3733 if (!mapping->a_ops->read_folio)
3734 return -ENOEXEC;
3735 file_accessed(file);
3736 vma->vm_ops = &generic_file_vm_ops;
3737 return 0;
3738}
3739
3740/*
3741 * This is for filesystems which do not implement ->writepage.
3742 */
3743int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3744{
3745 if (vma_is_shared_maywrite(vma))
3746 return -EINVAL;
3747 return generic_file_mmap(file, vma);
3748}
3749#else
3750vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3751{
3752 return VM_FAULT_SIGBUS;
3753}
3754int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3755{
3756 return -ENOSYS;
3757}
3758int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3759{
3760 return -ENOSYS;
3761}
3762#endif /* CONFIG_MMU */
3763
3764EXPORT_SYMBOL(filemap_page_mkwrite);
3765EXPORT_SYMBOL(generic_file_mmap);
3766EXPORT_SYMBOL(generic_file_readonly_mmap);
3767
3768static struct folio *do_read_cache_folio(struct address_space *mapping,
3769 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3770{
3771 struct folio *folio;
3772 int err;
3773
3774 if (!filler)
3775 filler = mapping->a_ops->read_folio;
3776repeat:
3777 folio = filemap_get_folio(mapping, index);
3778 if (IS_ERR(folio)) {
3779 folio = filemap_alloc_folio(gfp,
3780 mapping_min_folio_order(mapping));
3781 if (!folio)
3782 return ERR_PTR(-ENOMEM);
3783 index = mapping_align_index(mapping, index);
3784 err = filemap_add_folio(mapping, folio, index, gfp);
3785 if (unlikely(err)) {
3786 folio_put(folio);
3787 if (err == -EEXIST)
3788 goto repeat;
3789 /* Presumably ENOMEM for xarray node */
3790 return ERR_PTR(err);
3791 }
3792
3793 goto filler;
3794 }
3795 if (folio_test_uptodate(folio))
3796 goto out;
3797
3798 if (!folio_trylock(folio)) {
3799 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3800 goto repeat;
3801 }
3802
3803 /* Folio was truncated from mapping */
3804 if (!folio->mapping) {
3805 folio_unlock(folio);
3806 folio_put(folio);
3807 goto repeat;
3808 }
3809
3810 /* Someone else locked and filled the page in a very small window */
3811 if (folio_test_uptodate(folio)) {
3812 folio_unlock(folio);
3813 goto out;
3814 }
3815
3816filler:
3817 err = filemap_read_folio(file, filler, folio);
3818 if (err) {
3819 folio_put(folio);
3820 if (err == AOP_TRUNCATED_PAGE)
3821 goto repeat;
3822 return ERR_PTR(err);
3823 }
3824
3825out:
3826 folio_mark_accessed(folio);
3827 return folio;
3828}
3829
3830/**
3831 * read_cache_folio - Read into page cache, fill it if needed.
3832 * @mapping: The address_space to read from.
3833 * @index: The index to read.
3834 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3835 * @file: Passed to filler function, may be NULL if not required.
3836 *
3837 * Read one page into the page cache. If it succeeds, the folio returned
3838 * will contain @index, but it may not be the first page of the folio.
3839 *
3840 * If the filler function returns an error, it will be returned to the
3841 * caller.
3842 *
3843 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3844 * Return: An uptodate folio on success, ERR_PTR() on failure.
3845 */
3846struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3847 filler_t filler, struct file *file)
3848{
3849 return do_read_cache_folio(mapping, index, filler, file,
3850 mapping_gfp_mask(mapping));
3851}
3852EXPORT_SYMBOL(read_cache_folio);
3853
3854/**
3855 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags.
3856 * @mapping: The address_space for the folio.
3857 * @index: The index that the allocated folio will contain.
3858 * @gfp: The page allocator flags to use if allocating.
3859 *
3860 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with
3861 * any new memory allocations done using the specified allocation flags.
3862 *
3863 * The most likely error from this function is EIO, but ENOMEM is
3864 * possible and so is EINTR. If ->read_folio returns another error,
3865 * that will be returned to the caller.
3866 *
3867 * The function expects mapping->invalidate_lock to be already held.
3868 *
3869 * Return: Uptodate folio on success, ERR_PTR() on failure.
3870 */
3871struct folio *mapping_read_folio_gfp(struct address_space *mapping,
3872 pgoff_t index, gfp_t gfp)
3873{
3874 return do_read_cache_folio(mapping, index, NULL, NULL, gfp);
3875}
3876EXPORT_SYMBOL(mapping_read_folio_gfp);
3877
3878static struct page *do_read_cache_page(struct address_space *mapping,
3879 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3880{
3881 struct folio *folio;
3882
3883 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3884 if (IS_ERR(folio))
3885 return &folio->page;
3886 return folio_file_page(folio, index);
3887}
3888
3889struct page *read_cache_page(struct address_space *mapping,
3890 pgoff_t index, filler_t *filler, struct file *file)
3891{
3892 return do_read_cache_page(mapping, index, filler, file,
3893 mapping_gfp_mask(mapping));
3894}
3895EXPORT_SYMBOL(read_cache_page);
3896
3897/**
3898 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3899 * @mapping: the page's address_space
3900 * @index: the page index
3901 * @gfp: the page allocator flags to use if allocating
3902 *
3903 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3904 * any new page allocations done using the specified allocation flags.
3905 *
3906 * If the page does not get brought uptodate, return -EIO.
3907 *
3908 * The function expects mapping->invalidate_lock to be already held.
3909 *
3910 * Return: up to date page on success, ERR_PTR() on failure.
3911 */
3912struct page *read_cache_page_gfp(struct address_space *mapping,
3913 pgoff_t index,
3914 gfp_t gfp)
3915{
3916 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3917}
3918EXPORT_SYMBOL(read_cache_page_gfp);
3919
3920/*
3921 * Warn about a page cache invalidation failure during a direct I/O write.
3922 */
3923static void dio_warn_stale_pagecache(struct file *filp)
3924{
3925 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3926 char pathname[128];
3927 char *path;
3928
3929 errseq_set(&filp->f_mapping->wb_err, -EIO);
3930 if (__ratelimit(&_rs)) {
3931 path = file_path(filp, pathname, sizeof(pathname));
3932 if (IS_ERR(path))
3933 path = "(unknown)";
3934 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3935 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3936 current->comm);
3937 }
3938}
3939
3940void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count)
3941{
3942 struct address_space *mapping = iocb->ki_filp->f_mapping;
3943
3944 if (mapping->nrpages &&
3945 invalidate_inode_pages2_range(mapping,
3946 iocb->ki_pos >> PAGE_SHIFT,
3947 (iocb->ki_pos + count - 1) >> PAGE_SHIFT))
3948 dio_warn_stale_pagecache(iocb->ki_filp);
3949}
3950
3951ssize_t
3952generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3953{
3954 struct address_space *mapping = iocb->ki_filp->f_mapping;
3955 size_t write_len = iov_iter_count(from);
3956 ssize_t written;
3957
3958 /*
3959 * If a page can not be invalidated, return 0 to fall back
3960 * to buffered write.
3961 */
3962 written = kiocb_invalidate_pages(iocb, write_len);
3963 if (written) {
3964 if (written == -EBUSY)
3965 return 0;
3966 return written;
3967 }
3968
3969 written = mapping->a_ops->direct_IO(iocb, from);
3970
3971 /*
3972 * Finally, try again to invalidate clean pages which might have been
3973 * cached by non-direct readahead, or faulted in by get_user_pages()
3974 * if the source of the write was an mmap'ed region of the file
3975 * we're writing. Either one is a pretty crazy thing to do,
3976 * so we don't support it 100%. If this invalidation
3977 * fails, tough, the write still worked...
3978 *
3979 * Most of the time we do not need this since dio_complete() will do
3980 * the invalidation for us. However there are some file systems that
3981 * do not end up with dio_complete() being called, so let's not break
3982 * them by removing it completely.
3983 *
3984 * Noticeable example is a blkdev_direct_IO().
3985 *
3986 * Skip invalidation for async writes or if mapping has no pages.
3987 */
3988 if (written > 0) {
3989 struct inode *inode = mapping->host;
3990 loff_t pos = iocb->ki_pos;
3991
3992 kiocb_invalidate_post_direct_write(iocb, written);
3993 pos += written;
3994 write_len -= written;
3995 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3996 i_size_write(inode, pos);
3997 mark_inode_dirty(inode);
3998 }
3999 iocb->ki_pos = pos;
4000 }
4001 if (written != -EIOCBQUEUED)
4002 iov_iter_revert(from, write_len - iov_iter_count(from));
4003 return written;
4004}
4005EXPORT_SYMBOL(generic_file_direct_write);
4006
4007ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
4008{
4009 struct file *file = iocb->ki_filp;
4010 loff_t pos = iocb->ki_pos;
4011 struct address_space *mapping = file->f_mapping;
4012 const struct address_space_operations *a_ops = mapping->a_ops;
4013 size_t chunk = mapping_max_folio_size(mapping);
4014 long status = 0;
4015 ssize_t written = 0;
4016
4017 do {
4018 struct folio *folio;
4019 size_t offset; /* Offset into folio */
4020 size_t bytes; /* Bytes to write to folio */
4021 size_t copied; /* Bytes copied from user */
4022 void *fsdata = NULL;
4023
4024 bytes = iov_iter_count(i);
4025retry:
4026 offset = pos & (chunk - 1);
4027 bytes = min(chunk - offset, bytes);
4028 balance_dirty_pages_ratelimited(mapping);
4029
4030 /*
4031 * Bring in the user page that we will copy from _first_.
4032 * Otherwise there's a nasty deadlock on copying from the
4033 * same page as we're writing to, without it being marked
4034 * up-to-date.
4035 */
4036 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
4037 status = -EFAULT;
4038 break;
4039 }
4040
4041 if (fatal_signal_pending(current)) {
4042 status = -EINTR;
4043 break;
4044 }
4045
4046 status = a_ops->write_begin(file, mapping, pos, bytes,
4047 &folio, &fsdata);
4048 if (unlikely(status < 0))
4049 break;
4050
4051 offset = offset_in_folio(folio, pos);
4052 if (bytes > folio_size(folio) - offset)
4053 bytes = folio_size(folio) - offset;
4054
4055 if (mapping_writably_mapped(mapping))
4056 flush_dcache_folio(folio);
4057
4058 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
4059 flush_dcache_folio(folio);
4060
4061 status = a_ops->write_end(file, mapping, pos, bytes, copied,
4062 folio, fsdata);
4063 if (unlikely(status != copied)) {
4064 iov_iter_revert(i, copied - max(status, 0L));
4065 if (unlikely(status < 0))
4066 break;
4067 }
4068 cond_resched();
4069
4070 if (unlikely(status == 0)) {
4071 /*
4072 * A short copy made ->write_end() reject the
4073 * thing entirely. Might be memory poisoning
4074 * halfway through, might be a race with munmap,
4075 * might be severe memory pressure.
4076 */
4077 if (chunk > PAGE_SIZE)
4078 chunk /= 2;
4079 if (copied) {
4080 bytes = copied;
4081 goto retry;
4082 }
4083 } else {
4084 pos += status;
4085 written += status;
4086 }
4087 } while (iov_iter_count(i));
4088
4089 if (!written)
4090 return status;
4091 iocb->ki_pos += written;
4092 return written;
4093}
4094EXPORT_SYMBOL(generic_perform_write);
4095
4096/**
4097 * __generic_file_write_iter - write data to a file
4098 * @iocb: IO state structure (file, offset, etc.)
4099 * @from: iov_iter with data to write
4100 *
4101 * This function does all the work needed for actually writing data to a
4102 * file. It does all basic checks, removes SUID from the file, updates
4103 * modification times and calls proper subroutines depending on whether we
4104 * do direct IO or a standard buffered write.
4105 *
4106 * It expects i_rwsem to be grabbed unless we work on a block device or similar
4107 * object which does not need locking at all.
4108 *
4109 * This function does *not* take care of syncing data in case of O_SYNC write.
4110 * A caller has to handle it. This is mainly due to the fact that we want to
4111 * avoid syncing under i_rwsem.
4112 *
4113 * Return:
4114 * * number of bytes written, even for truncated writes
4115 * * negative error code if no data has been written at all
4116 */
4117ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4118{
4119 struct file *file = iocb->ki_filp;
4120 struct address_space *mapping = file->f_mapping;
4121 struct inode *inode = mapping->host;
4122 ssize_t ret;
4123
4124 ret = file_remove_privs(file);
4125 if (ret)
4126 return ret;
4127
4128 ret = file_update_time(file);
4129 if (ret)
4130 return ret;
4131
4132 if (iocb->ki_flags & IOCB_DIRECT) {
4133 ret = generic_file_direct_write(iocb, from);
4134 /*
4135 * If the write stopped short of completing, fall back to
4136 * buffered writes. Some filesystems do this for writes to
4137 * holes, for example. For DAX files, a buffered write will
4138 * not succeed (even if it did, DAX does not handle dirty
4139 * page-cache pages correctly).
4140 */
4141 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode))
4142 return ret;
4143 return direct_write_fallback(iocb, from, ret,
4144 generic_perform_write(iocb, from));
4145 }
4146
4147 return generic_perform_write(iocb, from);
4148}
4149EXPORT_SYMBOL(__generic_file_write_iter);
4150
4151/**
4152 * generic_file_write_iter - write data to a file
4153 * @iocb: IO state structure
4154 * @from: iov_iter with data to write
4155 *
4156 * This is a wrapper around __generic_file_write_iter() to be used by most
4157 * filesystems. It takes care of syncing the file in case of O_SYNC file
4158 * and acquires i_rwsem as needed.
4159 * Return:
4160 * * negative error code if no data has been written at all of
4161 * vfs_fsync_range() failed for a synchronous write
4162 * * number of bytes written, even for truncated writes
4163 */
4164ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4165{
4166 struct file *file = iocb->ki_filp;
4167 struct inode *inode = file->f_mapping->host;
4168 ssize_t ret;
4169
4170 inode_lock(inode);
4171 ret = generic_write_checks(iocb, from);
4172 if (ret > 0)
4173 ret = __generic_file_write_iter(iocb, from);
4174 inode_unlock(inode);
4175
4176 if (ret > 0)
4177 ret = generic_write_sync(iocb, ret);
4178 return ret;
4179}
4180EXPORT_SYMBOL(generic_file_write_iter);
4181
4182/**
4183 * filemap_release_folio() - Release fs-specific metadata on a folio.
4184 * @folio: The folio which the kernel is trying to free.
4185 * @gfp: Memory allocation flags (and I/O mode).
4186 *
4187 * The address_space is trying to release any data attached to a folio
4188 * (presumably at folio->private).
4189 *
4190 * This will also be called if the private_2 flag is set on a page,
4191 * indicating that the folio has other metadata associated with it.
4192 *
4193 * The @gfp argument specifies whether I/O may be performed to release
4194 * this page (__GFP_IO), and whether the call may block
4195 * (__GFP_RECLAIM & __GFP_FS).
4196 *
4197 * Return: %true if the release was successful, otherwise %false.
4198 */
4199bool filemap_release_folio(struct folio *folio, gfp_t gfp)
4200{
4201 struct address_space * const mapping = folio->mapping;
4202
4203 BUG_ON(!folio_test_locked(folio));
4204 if (!folio_needs_release(folio))
4205 return true;
4206 if (folio_test_writeback(folio))
4207 return false;
4208
4209 if (mapping && mapping->a_ops->release_folio)
4210 return mapping->a_ops->release_folio(folio, gfp);
4211 return try_to_free_buffers(folio);
4212}
4213EXPORT_SYMBOL(filemap_release_folio);
4214
4215/**
4216 * filemap_invalidate_inode - Invalidate/forcibly write back a range of an inode's pagecache
4217 * @inode: The inode to flush
4218 * @flush: Set to write back rather than simply invalidate.
4219 * @start: First byte to in range.
4220 * @end: Last byte in range (inclusive), or LLONG_MAX for everything from start
4221 * onwards.
4222 *
4223 * Invalidate all the folios on an inode that contribute to the specified
4224 * range, possibly writing them back first. Whilst the operation is
4225 * undertaken, the invalidate lock is held to prevent new folios from being
4226 * installed.
4227 */
4228int filemap_invalidate_inode(struct inode *inode, bool flush,
4229 loff_t start, loff_t end)
4230{
4231 struct address_space *mapping = inode->i_mapping;
4232 pgoff_t first = start >> PAGE_SHIFT;
4233 pgoff_t last = end >> PAGE_SHIFT;
4234 pgoff_t nr = end == LLONG_MAX ? ULONG_MAX : last - first + 1;
4235
4236 if (!mapping || !mapping->nrpages || end < start)
4237 goto out;
4238
4239 /* Prevent new folios from being added to the inode. */
4240 filemap_invalidate_lock(mapping);
4241
4242 if (!mapping->nrpages)
4243 goto unlock;
4244
4245 unmap_mapping_pages(mapping, first, nr, false);
4246
4247 /* Write back the data if we're asked to. */
4248 if (flush) {
4249 struct writeback_control wbc = {
4250 .sync_mode = WB_SYNC_ALL,
4251 .nr_to_write = LONG_MAX,
4252 .range_start = start,
4253 .range_end = end,
4254 };
4255
4256 filemap_fdatawrite_wbc(mapping, &wbc);
4257 }
4258
4259 /* Wait for writeback to complete on all folios and discard. */
4260 invalidate_inode_pages2_range(mapping, start / PAGE_SIZE, end / PAGE_SIZE);
4261
4262unlock:
4263 filemap_invalidate_unlock(mapping);
4264out:
4265 return filemap_check_errors(mapping);
4266}
4267EXPORT_SYMBOL_GPL(filemap_invalidate_inode);
4268
4269#ifdef CONFIG_CACHESTAT_SYSCALL
4270/**
4271 * filemap_cachestat() - compute the page cache statistics of a mapping
4272 * @mapping: The mapping to compute the statistics for.
4273 * @first_index: The starting page cache index.
4274 * @last_index: The final page index (inclusive).
4275 * @cs: the cachestat struct to write the result to.
4276 *
4277 * This will query the page cache statistics of a mapping in the
4278 * page range of [first_index, last_index] (inclusive). The statistics
4279 * queried include: number of dirty pages, number of pages marked for
4280 * writeback, and the number of (recently) evicted pages.
4281 */
4282static void filemap_cachestat(struct address_space *mapping,
4283 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs)
4284{
4285 XA_STATE(xas, &mapping->i_pages, first_index);
4286 struct folio *folio;
4287
4288 /* Flush stats (and potentially sleep) outside the RCU read section. */
4289 mem_cgroup_flush_stats_ratelimited(NULL);
4290
4291 rcu_read_lock();
4292 xas_for_each(&xas, folio, last_index) {
4293 int order;
4294 unsigned long nr_pages;
4295 pgoff_t folio_first_index, folio_last_index;
4296
4297 /*
4298 * Don't deref the folio. It is not pinned, and might
4299 * get freed (and reused) underneath us.
4300 *
4301 * We *could* pin it, but that would be expensive for
4302 * what should be a fast and lightweight syscall.
4303 *
4304 * Instead, derive all information of interest from
4305 * the rcu-protected xarray.
4306 */
4307
4308 if (xas_retry(&xas, folio))
4309 continue;
4310
4311 order = xas_get_order(&xas);
4312 nr_pages = 1 << order;
4313 folio_first_index = round_down(xas.xa_index, 1 << order);
4314 folio_last_index = folio_first_index + nr_pages - 1;
4315
4316 /* Folios might straddle the range boundaries, only count covered pages */
4317 if (folio_first_index < first_index)
4318 nr_pages -= first_index - folio_first_index;
4319
4320 if (folio_last_index > last_index)
4321 nr_pages -= folio_last_index - last_index;
4322
4323 if (xa_is_value(folio)) {
4324 /* page is evicted */
4325 void *shadow = (void *)folio;
4326 bool workingset; /* not used */
4327
4328 cs->nr_evicted += nr_pages;
4329
4330#ifdef CONFIG_SWAP /* implies CONFIG_MMU */
4331 if (shmem_mapping(mapping)) {
4332 /* shmem file - in swap cache */
4333 swp_entry_t swp = radix_to_swp_entry(folio);
4334
4335 /* swapin error results in poisoned entry */
4336 if (non_swap_entry(swp))
4337 goto resched;
4338
4339 /*
4340 * Getting a swap entry from the shmem
4341 * inode means we beat
4342 * shmem_unuse(). rcu_read_lock()
4343 * ensures swapoff waits for us before
4344 * freeing the swapper space. However,
4345 * we can race with swapping and
4346 * invalidation, so there might not be
4347 * a shadow in the swapcache (yet).
4348 */
4349 shadow = get_shadow_from_swap_cache(swp);
4350 if (!shadow)
4351 goto resched;
4352 }
4353#endif
4354 if (workingset_test_recent(shadow, true, &workingset, false))
4355 cs->nr_recently_evicted += nr_pages;
4356
4357 goto resched;
4358 }
4359
4360 /* page is in cache */
4361 cs->nr_cache += nr_pages;
4362
4363 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY))
4364 cs->nr_dirty += nr_pages;
4365
4366 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK))
4367 cs->nr_writeback += nr_pages;
4368
4369resched:
4370 if (need_resched()) {
4371 xas_pause(&xas);
4372 cond_resched_rcu();
4373 }
4374 }
4375 rcu_read_unlock();
4376}
4377
4378/*
4379 * See mincore: reveal pagecache information only for files
4380 * that the calling process has write access to, or could (if
4381 * tried) open for writing.
4382 */
4383static inline bool can_do_cachestat(struct file *f)
4384{
4385 if (f->f_mode & FMODE_WRITE)
4386 return true;
4387 if (inode_owner_or_capable(file_mnt_idmap(f), file_inode(f)))
4388 return true;
4389 return file_permission(f, MAY_WRITE) == 0;
4390}
4391
4392/*
4393 * The cachestat(2) system call.
4394 *
4395 * cachestat() returns the page cache statistics of a file in the
4396 * bytes range specified by `off` and `len`: number of cached pages,
4397 * number of dirty pages, number of pages marked for writeback,
4398 * number of evicted pages, and number of recently evicted pages.
4399 *
4400 * An evicted page is a page that is previously in the page cache
4401 * but has been evicted since. A page is recently evicted if its last
4402 * eviction was recent enough that its reentry to the cache would
4403 * indicate that it is actively being used by the system, and that
4404 * there is memory pressure on the system.
4405 *
4406 * `off` and `len` must be non-negative integers. If `len` > 0,
4407 * the queried range is [`off`, `off` + `len`]. If `len` == 0,
4408 * we will query in the range from `off` to the end of the file.
4409 *
4410 * The `flags` argument is unused for now, but is included for future
4411 * extensibility. User should pass 0 (i.e no flag specified).
4412 *
4413 * Currently, hugetlbfs is not supported.
4414 *
4415 * Because the status of a page can change after cachestat() checks it
4416 * but before it returns to the application, the returned values may
4417 * contain stale information.
4418 *
4419 * return values:
4420 * zero - success
4421 * -EFAULT - cstat or cstat_range points to an illegal address
4422 * -EINVAL - invalid flags
4423 * -EBADF - invalid file descriptor
4424 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file
4425 */
4426SYSCALL_DEFINE4(cachestat, unsigned int, fd,
4427 struct cachestat_range __user *, cstat_range,
4428 struct cachestat __user *, cstat, unsigned int, flags)
4429{
4430 CLASS(fd, f)(fd);
4431 struct address_space *mapping;
4432 struct cachestat_range csr;
4433 struct cachestat cs;
4434 pgoff_t first_index, last_index;
4435
4436 if (fd_empty(f))
4437 return -EBADF;
4438
4439 if (copy_from_user(&csr, cstat_range,
4440 sizeof(struct cachestat_range)))
4441 return -EFAULT;
4442
4443 /* hugetlbfs is not supported */
4444 if (is_file_hugepages(fd_file(f)))
4445 return -EOPNOTSUPP;
4446
4447 if (!can_do_cachestat(fd_file(f)))
4448 return -EPERM;
4449
4450 if (flags != 0)
4451 return -EINVAL;
4452
4453 first_index = csr.off >> PAGE_SHIFT;
4454 last_index =
4455 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT;
4456 memset(&cs, 0, sizeof(struct cachestat));
4457 mapping = fd_file(f)->f_mapping;
4458 filemap_cachestat(mapping, first_index, last_index, &cs);
4459
4460 if (copy_to_user(cstat, &cs, sizeof(struct cachestat)))
4461 return -EFAULT;
4462
4463 return 0;
4464}
4465#endif /* CONFIG_CACHESTAT_SYSCALL */