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