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/sched/signal.h>
17#include <linux/uaccess.h>
18#include <linux/capability.h>
19#include <linux/kernel_stat.h>
20#include <linux/gfp.h>
21#include <linux/mm.h>
22#include <linux/swap.h>
23#include <linux/mman.h>
24#include <linux/pagemap.h>
25#include <linux/file.h>
26#include <linux/uio.h>
27#include <linux/hash.h>
28#include <linux/writeback.h>
29#include <linux/backing-dev.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/security.h>
33#include <linux/cpuset.h>
34#include <linux/hugetlb.h>
35#include <linux/memcontrol.h>
36#include <linux/cleancache.h>
37#include <linux/shmem_fs.h>
38#include <linux/rmap.h>
39#include "internal.h"
40
41#define CREATE_TRACE_POINTS
42#include <trace/events/filemap.h>
43
44/*
45 * FIXME: remove all knowledge of the buffer layer from the core VM
46 */
47#include <linux/buffer_head.h> /* for try_to_free_buffers */
48
49#include <asm/mman.h>
50
51/*
52 * Shared mappings implemented 30.11.1994. It's not fully working yet,
53 * though.
54 *
55 * Shared mappings now work. 15.8.1995 Bruno.
56 *
57 * finished 'unifying' the page and buffer cache and SMP-threaded the
58 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
59 *
60 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
61 */
62
63/*
64 * Lock ordering:
65 *
66 * ->i_mmap_rwsem (truncate_pagecache)
67 * ->private_lock (__free_pte->__set_page_dirty_buffers)
68 * ->swap_lock (exclusive_swap_page, others)
69 * ->i_pages lock
70 *
71 * ->i_mutex
72 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
73 *
74 * ->mmap_sem
75 * ->i_mmap_rwsem
76 * ->page_table_lock or pte_lock (various, mainly in memory.c)
77 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
78 *
79 * ->mmap_sem
80 * ->lock_page (access_process_vm)
81 *
82 * ->i_mutex (generic_perform_write)
83 * ->mmap_sem (fault_in_pages_readable->do_page_fault)
84 *
85 * bdi->wb.list_lock
86 * sb_lock (fs/fs-writeback.c)
87 * ->i_pages lock (__sync_single_inode)
88 *
89 * ->i_mmap_rwsem
90 * ->anon_vma.lock (vma_adjust)
91 *
92 * ->anon_vma.lock
93 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
94 *
95 * ->page_table_lock or pte_lock
96 * ->swap_lock (try_to_unmap_one)
97 * ->private_lock (try_to_unmap_one)
98 * ->i_pages lock (try_to_unmap_one)
99 * ->zone_lru_lock(zone) (follow_page->mark_page_accessed)
100 * ->zone_lru_lock(zone) (check_pte_range->isolate_lru_page)
101 * ->private_lock (page_remove_rmap->set_page_dirty)
102 * ->i_pages lock (page_remove_rmap->set_page_dirty)
103 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
104 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
105 * ->memcg->move_lock (page_remove_rmap->lock_page_memcg)
106 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
107 * ->inode->i_lock (zap_pte_range->set_page_dirty)
108 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
109 *
110 * ->i_mmap_rwsem
111 * ->tasklist_lock (memory_failure, collect_procs_ao)
112 */
113
114static int page_cache_tree_insert(struct address_space *mapping,
115 struct page *page, void **shadowp)
116{
117 struct radix_tree_node *node;
118 void **slot;
119 int error;
120
121 error = __radix_tree_create(&mapping->i_pages, page->index, 0,
122 &node, &slot);
123 if (error)
124 return error;
125 if (*slot) {
126 void *p;
127
128 p = radix_tree_deref_slot_protected(slot,
129 &mapping->i_pages.xa_lock);
130 if (!radix_tree_exceptional_entry(p))
131 return -EEXIST;
132
133 mapping->nrexceptional--;
134 if (shadowp)
135 *shadowp = p;
136 }
137 __radix_tree_replace(&mapping->i_pages, node, slot, page,
138 workingset_lookup_update(mapping));
139 mapping->nrpages++;
140 return 0;
141}
142
143static void page_cache_tree_delete(struct address_space *mapping,
144 struct page *page, void *shadow)
145{
146 int i, nr;
147
148 /* hugetlb pages are represented by one entry in the radix tree */
149 nr = PageHuge(page) ? 1 : hpage_nr_pages(page);
150
151 VM_BUG_ON_PAGE(!PageLocked(page), page);
152 VM_BUG_ON_PAGE(PageTail(page), page);
153 VM_BUG_ON_PAGE(nr != 1 && shadow, page);
154
155 for (i = 0; i < nr; i++) {
156 struct radix_tree_node *node;
157 void **slot;
158
159 __radix_tree_lookup(&mapping->i_pages, page->index + i,
160 &node, &slot);
161
162 VM_BUG_ON_PAGE(!node && nr != 1, page);
163
164 radix_tree_clear_tags(&mapping->i_pages, node, slot);
165 __radix_tree_replace(&mapping->i_pages, node, slot, shadow,
166 workingset_lookup_update(mapping));
167 }
168
169 page->mapping = NULL;
170 /* Leave page->index set: truncation lookup relies upon it */
171
172 if (shadow) {
173 mapping->nrexceptional += nr;
174 /*
175 * Make sure the nrexceptional update is committed before
176 * the nrpages update so that final truncate racing
177 * with reclaim does not see both counters 0 at the
178 * same time and miss a shadow entry.
179 */
180 smp_wmb();
181 }
182 mapping->nrpages -= nr;
183}
184
185static void unaccount_page_cache_page(struct address_space *mapping,
186 struct page *page)
187{
188 int nr;
189
190 /*
191 * if we're uptodate, flush out into the cleancache, otherwise
192 * invalidate any existing cleancache entries. We can't leave
193 * stale data around in the cleancache once our page is gone
194 */
195 if (PageUptodate(page) && PageMappedToDisk(page))
196 cleancache_put_page(page);
197 else
198 cleancache_invalidate_page(mapping, page);
199
200 VM_BUG_ON_PAGE(PageTail(page), page);
201 VM_BUG_ON_PAGE(page_mapped(page), page);
202 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
203 int mapcount;
204
205 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
206 current->comm, page_to_pfn(page));
207 dump_page(page, "still mapped when deleted");
208 dump_stack();
209 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
210
211 mapcount = page_mapcount(page);
212 if (mapping_exiting(mapping) &&
213 page_count(page) >= mapcount + 2) {
214 /*
215 * All vmas have already been torn down, so it's
216 * a good bet that actually the page is unmapped,
217 * and we'd prefer not to leak it: if we're wrong,
218 * some other bad page check should catch it later.
219 */
220 page_mapcount_reset(page);
221 page_ref_sub(page, mapcount);
222 }
223 }
224
225 /* hugetlb pages do not participate in page cache accounting. */
226 if (PageHuge(page))
227 return;
228
229 nr = hpage_nr_pages(page);
230
231 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
232 if (PageSwapBacked(page)) {
233 __mod_node_page_state(page_pgdat(page), NR_SHMEM, -nr);
234 if (PageTransHuge(page))
235 __dec_node_page_state(page, NR_SHMEM_THPS);
236 } else {
237 VM_BUG_ON_PAGE(PageTransHuge(page), page);
238 }
239
240 /*
241 * At this point page must be either written or cleaned by
242 * truncate. Dirty page here signals a bug and loss of
243 * unwritten data.
244 *
245 * This fixes dirty accounting after removing the page entirely
246 * but leaves PageDirty set: it has no effect for truncated
247 * page and anyway will be cleared before returning page into
248 * buddy allocator.
249 */
250 if (WARN_ON_ONCE(PageDirty(page)))
251 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
252}
253
254/*
255 * Delete a page from the page cache and free it. Caller has to make
256 * sure the page is locked and that nobody else uses it - or that usage
257 * is safe. The caller must hold the i_pages lock.
258 */
259void __delete_from_page_cache(struct page *page, void *shadow)
260{
261 struct address_space *mapping = page->mapping;
262
263 trace_mm_filemap_delete_from_page_cache(page);
264
265 unaccount_page_cache_page(mapping, page);
266 page_cache_tree_delete(mapping, page, shadow);
267}
268
269static void page_cache_free_page(struct address_space *mapping,
270 struct page *page)
271{
272 void (*freepage)(struct page *);
273
274 freepage = mapping->a_ops->freepage;
275 if (freepage)
276 freepage(page);
277
278 if (PageTransHuge(page) && !PageHuge(page)) {
279 page_ref_sub(page, HPAGE_PMD_NR);
280 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
281 } else {
282 put_page(page);
283 }
284}
285
286/**
287 * delete_from_page_cache - delete page from page cache
288 * @page: the page which the kernel is trying to remove from page cache
289 *
290 * This must be called only on pages that have been verified to be in the page
291 * cache and locked. It will never put the page into the free list, the caller
292 * has a reference on the page.
293 */
294void delete_from_page_cache(struct page *page)
295{
296 struct address_space *mapping = page_mapping(page);
297 unsigned long flags;
298
299 BUG_ON(!PageLocked(page));
300 xa_lock_irqsave(&mapping->i_pages, flags);
301 __delete_from_page_cache(page, NULL);
302 xa_unlock_irqrestore(&mapping->i_pages, flags);
303
304 page_cache_free_page(mapping, page);
305}
306EXPORT_SYMBOL(delete_from_page_cache);
307
308/*
309 * page_cache_tree_delete_batch - delete several pages from page cache
310 * @mapping: the mapping to which pages belong
311 * @pvec: pagevec with pages to delete
312 *
313 * The function walks over mapping->i_pages and removes pages passed in @pvec
314 * from the mapping. The function expects @pvec to be sorted by page index.
315 * It tolerates holes in @pvec (mapping entries at those indices are not
316 * modified). The function expects only THP head pages to be present in the
317 * @pvec and takes care to delete all corresponding tail pages from the
318 * mapping as well.
319 *
320 * The function expects the i_pages lock to be held.
321 */
322static void
323page_cache_tree_delete_batch(struct address_space *mapping,
324 struct pagevec *pvec)
325{
326 struct radix_tree_iter iter;
327 void **slot;
328 int total_pages = 0;
329 int i = 0, tail_pages = 0;
330 struct page *page;
331 pgoff_t start;
332
333 start = pvec->pages[0]->index;
334 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
335 if (i >= pagevec_count(pvec) && !tail_pages)
336 break;
337 page = radix_tree_deref_slot_protected(slot,
338 &mapping->i_pages.xa_lock);
339 if (radix_tree_exceptional_entry(page))
340 continue;
341 if (!tail_pages) {
342 /*
343 * Some page got inserted in our range? Skip it. We
344 * have our pages locked so they are protected from
345 * being removed.
346 */
347 if (page != pvec->pages[i])
348 continue;
349 WARN_ON_ONCE(!PageLocked(page));
350 if (PageTransHuge(page) && !PageHuge(page))
351 tail_pages = HPAGE_PMD_NR - 1;
352 page->mapping = NULL;
353 /*
354 * Leave page->index set: truncation lookup relies
355 * upon it
356 */
357 i++;
358 } else {
359 tail_pages--;
360 }
361 radix_tree_clear_tags(&mapping->i_pages, iter.node, slot);
362 __radix_tree_replace(&mapping->i_pages, iter.node, slot, NULL,
363 workingset_lookup_update(mapping));
364 total_pages++;
365 }
366 mapping->nrpages -= total_pages;
367}
368
369void delete_from_page_cache_batch(struct address_space *mapping,
370 struct pagevec *pvec)
371{
372 int i;
373 unsigned long flags;
374
375 if (!pagevec_count(pvec))
376 return;
377
378 xa_lock_irqsave(&mapping->i_pages, flags);
379 for (i = 0; i < pagevec_count(pvec); i++) {
380 trace_mm_filemap_delete_from_page_cache(pvec->pages[i]);
381
382 unaccount_page_cache_page(mapping, pvec->pages[i]);
383 }
384 page_cache_tree_delete_batch(mapping, pvec);
385 xa_unlock_irqrestore(&mapping->i_pages, flags);
386
387 for (i = 0; i < pagevec_count(pvec); i++)
388 page_cache_free_page(mapping, pvec->pages[i]);
389}
390
391int filemap_check_errors(struct address_space *mapping)
392{
393 int ret = 0;
394 /* Check for outstanding write errors */
395 if (test_bit(AS_ENOSPC, &mapping->flags) &&
396 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
397 ret = -ENOSPC;
398 if (test_bit(AS_EIO, &mapping->flags) &&
399 test_and_clear_bit(AS_EIO, &mapping->flags))
400 ret = -EIO;
401 return ret;
402}
403EXPORT_SYMBOL(filemap_check_errors);
404
405static int filemap_check_and_keep_errors(struct address_space *mapping)
406{
407 /* Check for outstanding write errors */
408 if (test_bit(AS_EIO, &mapping->flags))
409 return -EIO;
410 if (test_bit(AS_ENOSPC, &mapping->flags))
411 return -ENOSPC;
412 return 0;
413}
414
415/**
416 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
417 * @mapping: address space structure to write
418 * @start: offset in bytes where the range starts
419 * @end: offset in bytes where the range ends (inclusive)
420 * @sync_mode: enable synchronous operation
421 *
422 * Start writeback against all of a mapping's dirty pages that lie
423 * within the byte offsets <start, end> inclusive.
424 *
425 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
426 * opposed to a regular memory cleansing writeback. The difference between
427 * these two operations is that if a dirty page/buffer is encountered, it must
428 * be waited upon, and not just skipped over.
429 */
430int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
431 loff_t end, int sync_mode)
432{
433 int ret;
434 struct writeback_control wbc = {
435 .sync_mode = sync_mode,
436 .nr_to_write = LONG_MAX,
437 .range_start = start,
438 .range_end = end,
439 };
440
441 if (!mapping_cap_writeback_dirty(mapping))
442 return 0;
443
444 wbc_attach_fdatawrite_inode(&wbc, mapping->host);
445 ret = do_writepages(mapping, &wbc);
446 wbc_detach_inode(&wbc);
447 return ret;
448}
449
450static inline int __filemap_fdatawrite(struct address_space *mapping,
451 int sync_mode)
452{
453 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
454}
455
456int filemap_fdatawrite(struct address_space *mapping)
457{
458 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
459}
460EXPORT_SYMBOL(filemap_fdatawrite);
461
462int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
463 loff_t end)
464{
465 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
466}
467EXPORT_SYMBOL(filemap_fdatawrite_range);
468
469/**
470 * filemap_flush - mostly a non-blocking flush
471 * @mapping: target address_space
472 *
473 * This is a mostly non-blocking flush. Not suitable for data-integrity
474 * purposes - I/O may not be started against all dirty pages.
475 */
476int filemap_flush(struct address_space *mapping)
477{
478 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
479}
480EXPORT_SYMBOL(filemap_flush);
481
482/**
483 * filemap_range_has_page - check if a page exists in range.
484 * @mapping: address space within which to check
485 * @start_byte: offset in bytes where the range starts
486 * @end_byte: offset in bytes where the range ends (inclusive)
487 *
488 * Find at least one page in the range supplied, usually used to check if
489 * direct writing in this range will trigger a writeback.
490 */
491bool filemap_range_has_page(struct address_space *mapping,
492 loff_t start_byte, loff_t end_byte)
493{
494 pgoff_t index = start_byte >> PAGE_SHIFT;
495 pgoff_t end = end_byte >> PAGE_SHIFT;
496 struct page *page;
497
498 if (end_byte < start_byte)
499 return false;
500
501 if (mapping->nrpages == 0)
502 return false;
503
504 if (!find_get_pages_range(mapping, &index, end, 1, &page))
505 return false;
506 put_page(page);
507 return true;
508}
509EXPORT_SYMBOL(filemap_range_has_page);
510
511static void __filemap_fdatawait_range(struct address_space *mapping,
512 loff_t start_byte, loff_t end_byte)
513{
514 pgoff_t index = start_byte >> PAGE_SHIFT;
515 pgoff_t end = end_byte >> PAGE_SHIFT;
516 struct pagevec pvec;
517 int nr_pages;
518
519 if (end_byte < start_byte)
520 return;
521
522 pagevec_init(&pvec);
523 while (index <= end) {
524 unsigned i;
525
526 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
527 end, PAGECACHE_TAG_WRITEBACK);
528 if (!nr_pages)
529 break;
530
531 for (i = 0; i < nr_pages; i++) {
532 struct page *page = pvec.pages[i];
533
534 wait_on_page_writeback(page);
535 ClearPageError(page);
536 }
537 pagevec_release(&pvec);
538 cond_resched();
539 }
540}
541
542/**
543 * filemap_fdatawait_range - wait for writeback to complete
544 * @mapping: address space structure to wait for
545 * @start_byte: offset in bytes where the range starts
546 * @end_byte: offset in bytes where the range ends (inclusive)
547 *
548 * Walk the list of under-writeback pages of the given address space
549 * in the given range and wait for all of them. Check error status of
550 * the address space and return it.
551 *
552 * Since the error status of the address space is cleared by this function,
553 * callers are responsible for checking the return value and handling and/or
554 * reporting the error.
555 */
556int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
557 loff_t end_byte)
558{
559 __filemap_fdatawait_range(mapping, start_byte, end_byte);
560 return filemap_check_errors(mapping);
561}
562EXPORT_SYMBOL(filemap_fdatawait_range);
563
564/**
565 * file_fdatawait_range - wait for writeback to complete
566 * @file: file pointing to address space structure to wait for
567 * @start_byte: offset in bytes where the range starts
568 * @end_byte: offset in bytes where the range ends (inclusive)
569 *
570 * Walk the list of under-writeback pages of the address space that file
571 * refers to, in the given range and wait for all of them. Check error
572 * status of the address space vs. the file->f_wb_err cursor and return it.
573 *
574 * Since the error status of the file is advanced by this function,
575 * callers are responsible for checking the return value and handling and/or
576 * reporting the error.
577 */
578int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
579{
580 struct address_space *mapping = file->f_mapping;
581
582 __filemap_fdatawait_range(mapping, start_byte, end_byte);
583 return file_check_and_advance_wb_err(file);
584}
585EXPORT_SYMBOL(file_fdatawait_range);
586
587/**
588 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
589 * @mapping: address space structure to wait for
590 *
591 * Walk the list of under-writeback pages of the given address space
592 * and wait for all of them. Unlike filemap_fdatawait(), this function
593 * does not clear error status of the address space.
594 *
595 * Use this function if callers don't handle errors themselves. Expected
596 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
597 * fsfreeze(8)
598 */
599int filemap_fdatawait_keep_errors(struct address_space *mapping)
600{
601 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
602 return filemap_check_and_keep_errors(mapping);
603}
604EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
605
606static bool mapping_needs_writeback(struct address_space *mapping)
607{
608 return (!dax_mapping(mapping) && mapping->nrpages) ||
609 (dax_mapping(mapping) && mapping->nrexceptional);
610}
611
612int filemap_write_and_wait(struct address_space *mapping)
613{
614 int err = 0;
615
616 if (mapping_needs_writeback(mapping)) {
617 err = filemap_fdatawrite(mapping);
618 /*
619 * Even if the above returned error, the pages may be
620 * written partially (e.g. -ENOSPC), so we wait for it.
621 * But the -EIO is special case, it may indicate the worst
622 * thing (e.g. bug) happened, so we avoid waiting for it.
623 */
624 if (err != -EIO) {
625 int err2 = filemap_fdatawait(mapping);
626 if (!err)
627 err = err2;
628 } else {
629 /* Clear any previously stored errors */
630 filemap_check_errors(mapping);
631 }
632 } else {
633 err = filemap_check_errors(mapping);
634 }
635 return err;
636}
637EXPORT_SYMBOL(filemap_write_and_wait);
638
639/**
640 * filemap_write_and_wait_range - write out & wait on a file range
641 * @mapping: the address_space for the pages
642 * @lstart: offset in bytes where the range starts
643 * @lend: offset in bytes where the range ends (inclusive)
644 *
645 * Write out and wait upon file offsets lstart->lend, inclusive.
646 *
647 * Note that @lend is inclusive (describes the last byte to be written) so
648 * that this function can be used to write to the very end-of-file (end = -1).
649 */
650int filemap_write_and_wait_range(struct address_space *mapping,
651 loff_t lstart, loff_t lend)
652{
653 int err = 0;
654
655 if (mapping_needs_writeback(mapping)) {
656 err = __filemap_fdatawrite_range(mapping, lstart, lend,
657 WB_SYNC_ALL);
658 /* See comment of filemap_write_and_wait() */
659 if (err != -EIO) {
660 int err2 = filemap_fdatawait_range(mapping,
661 lstart, lend);
662 if (!err)
663 err = err2;
664 } else {
665 /* Clear any previously stored errors */
666 filemap_check_errors(mapping);
667 }
668 } else {
669 err = filemap_check_errors(mapping);
670 }
671 return err;
672}
673EXPORT_SYMBOL(filemap_write_and_wait_range);
674
675void __filemap_set_wb_err(struct address_space *mapping, int err)
676{
677 errseq_t eseq = errseq_set(&mapping->wb_err, err);
678
679 trace_filemap_set_wb_err(mapping, eseq);
680}
681EXPORT_SYMBOL(__filemap_set_wb_err);
682
683/**
684 * file_check_and_advance_wb_err - report wb error (if any) that was previously
685 * and advance wb_err to current one
686 * @file: struct file on which the error is being reported
687 *
688 * When userland calls fsync (or something like nfsd does the equivalent), we
689 * want to report any writeback errors that occurred since the last fsync (or
690 * since the file was opened if there haven't been any).
691 *
692 * Grab the wb_err from the mapping. If it matches what we have in the file,
693 * then just quickly return 0. The file is all caught up.
694 *
695 * If it doesn't match, then take the mapping value, set the "seen" flag in
696 * it and try to swap it into place. If it works, or another task beat us
697 * to it with the new value, then update the f_wb_err and return the error
698 * portion. The error at this point must be reported via proper channels
699 * (a'la fsync, or NFS COMMIT operation, etc.).
700 *
701 * While we handle mapping->wb_err with atomic operations, the f_wb_err
702 * value is protected by the f_lock since we must ensure that it reflects
703 * the latest value swapped in for this file descriptor.
704 */
705int file_check_and_advance_wb_err(struct file *file)
706{
707 int err = 0;
708 errseq_t old = READ_ONCE(file->f_wb_err);
709 struct address_space *mapping = file->f_mapping;
710
711 /* Locklessly handle the common case where nothing has changed */
712 if (errseq_check(&mapping->wb_err, old)) {
713 /* Something changed, must use slow path */
714 spin_lock(&file->f_lock);
715 old = file->f_wb_err;
716 err = errseq_check_and_advance(&mapping->wb_err,
717 &file->f_wb_err);
718 trace_file_check_and_advance_wb_err(file, old);
719 spin_unlock(&file->f_lock);
720 }
721
722 /*
723 * We're mostly using this function as a drop in replacement for
724 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
725 * that the legacy code would have had on these flags.
726 */
727 clear_bit(AS_EIO, &mapping->flags);
728 clear_bit(AS_ENOSPC, &mapping->flags);
729 return err;
730}
731EXPORT_SYMBOL(file_check_and_advance_wb_err);
732
733/**
734 * file_write_and_wait_range - write out & wait on a file range
735 * @file: file pointing to address_space with pages
736 * @lstart: offset in bytes where the range starts
737 * @lend: offset in bytes where the range ends (inclusive)
738 *
739 * Write out and wait upon file offsets lstart->lend, inclusive.
740 *
741 * Note that @lend is inclusive (describes the last byte to be written) so
742 * that this function can be used to write to the very end-of-file (end = -1).
743 *
744 * After writing out and waiting on the data, we check and advance the
745 * f_wb_err cursor to the latest value, and return any errors detected there.
746 */
747int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
748{
749 int err = 0, err2;
750 struct address_space *mapping = file->f_mapping;
751
752 if (mapping_needs_writeback(mapping)) {
753 err = __filemap_fdatawrite_range(mapping, lstart, lend,
754 WB_SYNC_ALL);
755 /* See comment of filemap_write_and_wait() */
756 if (err != -EIO)
757 __filemap_fdatawait_range(mapping, lstart, lend);
758 }
759 err2 = file_check_and_advance_wb_err(file);
760 if (!err)
761 err = err2;
762 return err;
763}
764EXPORT_SYMBOL(file_write_and_wait_range);
765
766/**
767 * replace_page_cache_page - replace a pagecache page with a new one
768 * @old: page to be replaced
769 * @new: page to replace with
770 * @gfp_mask: allocation mode
771 *
772 * This function replaces a page in the pagecache with a new one. On
773 * success it acquires the pagecache reference for the new page and
774 * drops it for the old page. Both the old and new pages must be
775 * locked. This function does not add the new page to the LRU, the
776 * caller must do that.
777 *
778 * The remove + add is atomic. The only way this function can fail is
779 * memory allocation failure.
780 */
781int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
782{
783 int error;
784
785 VM_BUG_ON_PAGE(!PageLocked(old), old);
786 VM_BUG_ON_PAGE(!PageLocked(new), new);
787 VM_BUG_ON_PAGE(new->mapping, new);
788
789 error = radix_tree_preload(gfp_mask & GFP_RECLAIM_MASK);
790 if (!error) {
791 struct address_space *mapping = old->mapping;
792 void (*freepage)(struct page *);
793 unsigned long flags;
794
795 pgoff_t offset = old->index;
796 freepage = mapping->a_ops->freepage;
797
798 get_page(new);
799 new->mapping = mapping;
800 new->index = offset;
801
802 xa_lock_irqsave(&mapping->i_pages, flags);
803 __delete_from_page_cache(old, NULL);
804 error = page_cache_tree_insert(mapping, new, NULL);
805 BUG_ON(error);
806
807 /*
808 * hugetlb pages do not participate in page cache accounting.
809 */
810 if (!PageHuge(new))
811 __inc_node_page_state(new, NR_FILE_PAGES);
812 if (PageSwapBacked(new))
813 __inc_node_page_state(new, NR_SHMEM);
814 xa_unlock_irqrestore(&mapping->i_pages, flags);
815 mem_cgroup_migrate(old, new);
816 radix_tree_preload_end();
817 if (freepage)
818 freepage(old);
819 put_page(old);
820 }
821
822 return error;
823}
824EXPORT_SYMBOL_GPL(replace_page_cache_page);
825
826static int __add_to_page_cache_locked(struct page *page,
827 struct address_space *mapping,
828 pgoff_t offset, gfp_t gfp_mask,
829 void **shadowp)
830{
831 int huge = PageHuge(page);
832 struct mem_cgroup *memcg;
833 int error;
834
835 VM_BUG_ON_PAGE(!PageLocked(page), page);
836 VM_BUG_ON_PAGE(PageSwapBacked(page), page);
837
838 if (!huge) {
839 error = mem_cgroup_try_charge(page, current->mm,
840 gfp_mask, &memcg, false);
841 if (error)
842 return error;
843 }
844
845 error = radix_tree_maybe_preload(gfp_mask & GFP_RECLAIM_MASK);
846 if (error) {
847 if (!huge)
848 mem_cgroup_cancel_charge(page, memcg, false);
849 return error;
850 }
851
852 get_page(page);
853 page->mapping = mapping;
854 page->index = offset;
855
856 xa_lock_irq(&mapping->i_pages);
857 error = page_cache_tree_insert(mapping, page, shadowp);
858 radix_tree_preload_end();
859 if (unlikely(error))
860 goto err_insert;
861
862 /* hugetlb pages do not participate in page cache accounting. */
863 if (!huge)
864 __inc_node_page_state(page, NR_FILE_PAGES);
865 xa_unlock_irq(&mapping->i_pages);
866 if (!huge)
867 mem_cgroup_commit_charge(page, memcg, false, false);
868 trace_mm_filemap_add_to_page_cache(page);
869 return 0;
870err_insert:
871 page->mapping = NULL;
872 /* Leave page->index set: truncation relies upon it */
873 xa_unlock_irq(&mapping->i_pages);
874 if (!huge)
875 mem_cgroup_cancel_charge(page, memcg, false);
876 put_page(page);
877 return error;
878}
879
880/**
881 * add_to_page_cache_locked - add a locked page to the pagecache
882 * @page: page to add
883 * @mapping: the page's address_space
884 * @offset: page index
885 * @gfp_mask: page allocation mode
886 *
887 * This function is used to add a page to the pagecache. It must be locked.
888 * This function does not add the page to the LRU. The caller must do that.
889 */
890int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
891 pgoff_t offset, gfp_t gfp_mask)
892{
893 return __add_to_page_cache_locked(page, mapping, offset,
894 gfp_mask, NULL);
895}
896EXPORT_SYMBOL(add_to_page_cache_locked);
897
898int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
899 pgoff_t offset, gfp_t gfp_mask)
900{
901 void *shadow = NULL;
902 int ret;
903
904 __SetPageLocked(page);
905 ret = __add_to_page_cache_locked(page, mapping, offset,
906 gfp_mask, &shadow);
907 if (unlikely(ret))
908 __ClearPageLocked(page);
909 else {
910 /*
911 * The page might have been evicted from cache only
912 * recently, in which case it should be activated like
913 * any other repeatedly accessed page.
914 * The exception is pages getting rewritten; evicting other
915 * data from the working set, only to cache data that will
916 * get overwritten with something else, is a waste of memory.
917 */
918 if (!(gfp_mask & __GFP_WRITE) &&
919 shadow && workingset_refault(shadow)) {
920 SetPageActive(page);
921 workingset_activation(page);
922 } else
923 ClearPageActive(page);
924 lru_cache_add(page);
925 }
926 return ret;
927}
928EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
929
930#ifdef CONFIG_NUMA
931struct page *__page_cache_alloc(gfp_t gfp)
932{
933 int n;
934 struct page *page;
935
936 if (cpuset_do_page_mem_spread()) {
937 unsigned int cpuset_mems_cookie;
938 do {
939 cpuset_mems_cookie = read_mems_allowed_begin();
940 n = cpuset_mem_spread_node();
941 page = __alloc_pages_node(n, gfp, 0);
942 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
943
944 return page;
945 }
946 return alloc_pages(gfp, 0);
947}
948EXPORT_SYMBOL(__page_cache_alloc);
949#endif
950
951/*
952 * In order to wait for pages to become available there must be
953 * waitqueues associated with pages. By using a hash table of
954 * waitqueues where the bucket discipline is to maintain all
955 * waiters on the same queue and wake all when any of the pages
956 * become available, and for the woken contexts to check to be
957 * sure the appropriate page became available, this saves space
958 * at a cost of "thundering herd" phenomena during rare hash
959 * collisions.
960 */
961#define PAGE_WAIT_TABLE_BITS 8
962#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
963static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
964
965static wait_queue_head_t *page_waitqueue(struct page *page)
966{
967 return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
968}
969
970void __init pagecache_init(void)
971{
972 int i;
973
974 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
975 init_waitqueue_head(&page_wait_table[i]);
976
977 page_writeback_init();
978}
979
980/* This has the same layout as wait_bit_key - see fs/cachefiles/rdwr.c */
981struct wait_page_key {
982 struct page *page;
983 int bit_nr;
984 int page_match;
985};
986
987struct wait_page_queue {
988 struct page *page;
989 int bit_nr;
990 wait_queue_entry_t wait;
991};
992
993static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
994{
995 struct wait_page_key *key = arg;
996 struct wait_page_queue *wait_page
997 = container_of(wait, struct wait_page_queue, wait);
998
999 if (wait_page->page != key->page)
1000 return 0;
1001 key->page_match = 1;
1002
1003 if (wait_page->bit_nr != key->bit_nr)
1004 return 0;
1005
1006 /* Stop walking if it's locked */
1007 if (test_bit(key->bit_nr, &key->page->flags))
1008 return -1;
1009
1010 return autoremove_wake_function(wait, mode, sync, key);
1011}
1012
1013static void wake_up_page_bit(struct page *page, int bit_nr)
1014{
1015 wait_queue_head_t *q = page_waitqueue(page);
1016 struct wait_page_key key;
1017 unsigned long flags;
1018 wait_queue_entry_t bookmark;
1019
1020 key.page = page;
1021 key.bit_nr = bit_nr;
1022 key.page_match = 0;
1023
1024 bookmark.flags = 0;
1025 bookmark.private = NULL;
1026 bookmark.func = NULL;
1027 INIT_LIST_HEAD(&bookmark.entry);
1028
1029 spin_lock_irqsave(&q->lock, flags);
1030 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1031
1032 while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1033 /*
1034 * Take a breather from holding the lock,
1035 * allow pages that finish wake up asynchronously
1036 * to acquire the lock and remove themselves
1037 * from wait queue
1038 */
1039 spin_unlock_irqrestore(&q->lock, flags);
1040 cpu_relax();
1041 spin_lock_irqsave(&q->lock, flags);
1042 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1043 }
1044
1045 /*
1046 * It is possible for other pages to have collided on the waitqueue
1047 * hash, so in that case check for a page match. That prevents a long-
1048 * term waiter
1049 *
1050 * It is still possible to miss a case here, when we woke page waiters
1051 * and removed them from the waitqueue, but there are still other
1052 * page waiters.
1053 */
1054 if (!waitqueue_active(q) || !key.page_match) {
1055 ClearPageWaiters(page);
1056 /*
1057 * It's possible to miss clearing Waiters here, when we woke
1058 * our page waiters, but the hashed waitqueue has waiters for
1059 * other pages on it.
1060 *
1061 * That's okay, it's a rare case. The next waker will clear it.
1062 */
1063 }
1064 spin_unlock_irqrestore(&q->lock, flags);
1065}
1066
1067static void wake_up_page(struct page *page, int bit)
1068{
1069 if (!PageWaiters(page))
1070 return;
1071 wake_up_page_bit(page, bit);
1072}
1073
1074static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1075 struct page *page, int bit_nr, int state, bool lock)
1076{
1077 struct wait_page_queue wait_page;
1078 wait_queue_entry_t *wait = &wait_page.wait;
1079 int ret = 0;
1080
1081 init_wait(wait);
1082 wait->flags = lock ? WQ_FLAG_EXCLUSIVE : 0;
1083 wait->func = wake_page_function;
1084 wait_page.page = page;
1085 wait_page.bit_nr = bit_nr;
1086
1087 for (;;) {
1088 spin_lock_irq(&q->lock);
1089
1090 if (likely(list_empty(&wait->entry))) {
1091 __add_wait_queue_entry_tail(q, wait);
1092 SetPageWaiters(page);
1093 }
1094
1095 set_current_state(state);
1096
1097 spin_unlock_irq(&q->lock);
1098
1099 if (likely(test_bit(bit_nr, &page->flags))) {
1100 io_schedule();
1101 }
1102
1103 if (lock) {
1104 if (!test_and_set_bit_lock(bit_nr, &page->flags))
1105 break;
1106 } else {
1107 if (!test_bit(bit_nr, &page->flags))
1108 break;
1109 }
1110
1111 if (unlikely(signal_pending_state(state, current))) {
1112 ret = -EINTR;
1113 break;
1114 }
1115 }
1116
1117 finish_wait(q, wait);
1118
1119 /*
1120 * A signal could leave PageWaiters set. Clearing it here if
1121 * !waitqueue_active would be possible (by open-coding finish_wait),
1122 * but still fail to catch it in the case of wait hash collision. We
1123 * already can fail to clear wait hash collision cases, so don't
1124 * bother with signals either.
1125 */
1126
1127 return ret;
1128}
1129
1130void wait_on_page_bit(struct page *page, int bit_nr)
1131{
1132 wait_queue_head_t *q = page_waitqueue(page);
1133 wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, false);
1134}
1135EXPORT_SYMBOL(wait_on_page_bit);
1136
1137int wait_on_page_bit_killable(struct page *page, int bit_nr)
1138{
1139 wait_queue_head_t *q = page_waitqueue(page);
1140 return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, false);
1141}
1142EXPORT_SYMBOL(wait_on_page_bit_killable);
1143
1144/**
1145 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1146 * @page: Page defining the wait queue of interest
1147 * @waiter: Waiter to add to the queue
1148 *
1149 * Add an arbitrary @waiter to the wait queue for the nominated @page.
1150 */
1151void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1152{
1153 wait_queue_head_t *q = page_waitqueue(page);
1154 unsigned long flags;
1155
1156 spin_lock_irqsave(&q->lock, flags);
1157 __add_wait_queue_entry_tail(q, waiter);
1158 SetPageWaiters(page);
1159 spin_unlock_irqrestore(&q->lock, flags);
1160}
1161EXPORT_SYMBOL_GPL(add_page_wait_queue);
1162
1163#ifndef clear_bit_unlock_is_negative_byte
1164
1165/*
1166 * PG_waiters is the high bit in the same byte as PG_lock.
1167 *
1168 * On x86 (and on many other architectures), we can clear PG_lock and
1169 * test the sign bit at the same time. But if the architecture does
1170 * not support that special operation, we just do this all by hand
1171 * instead.
1172 *
1173 * The read of PG_waiters has to be after (or concurrently with) PG_locked
1174 * being cleared, but a memory barrier should be unneccssary since it is
1175 * in the same byte as PG_locked.
1176 */
1177static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1178{
1179 clear_bit_unlock(nr, mem);
1180 /* smp_mb__after_atomic(); */
1181 return test_bit(PG_waiters, mem);
1182}
1183
1184#endif
1185
1186/**
1187 * unlock_page - unlock a locked page
1188 * @page: the page
1189 *
1190 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1191 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1192 * mechanism between PageLocked pages and PageWriteback pages is shared.
1193 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1194 *
1195 * Note that this depends on PG_waiters being the sign bit in the byte
1196 * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1197 * clear the PG_locked bit and test PG_waiters at the same time fairly
1198 * portably (architectures that do LL/SC can test any bit, while x86 can
1199 * test the sign bit).
1200 */
1201void unlock_page(struct page *page)
1202{
1203 BUILD_BUG_ON(PG_waiters != 7);
1204 page = compound_head(page);
1205 VM_BUG_ON_PAGE(!PageLocked(page), page);
1206 if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1207 wake_up_page_bit(page, PG_locked);
1208}
1209EXPORT_SYMBOL(unlock_page);
1210
1211/**
1212 * end_page_writeback - end writeback against a page
1213 * @page: the page
1214 */
1215void end_page_writeback(struct page *page)
1216{
1217 /*
1218 * TestClearPageReclaim could be used here but it is an atomic
1219 * operation and overkill in this particular case. Failing to
1220 * shuffle a page marked for immediate reclaim is too mild to
1221 * justify taking an atomic operation penalty at the end of
1222 * ever page writeback.
1223 */
1224 if (PageReclaim(page)) {
1225 ClearPageReclaim(page);
1226 rotate_reclaimable_page(page);
1227 }
1228
1229 if (!test_clear_page_writeback(page))
1230 BUG();
1231
1232 smp_mb__after_atomic();
1233 wake_up_page(page, PG_writeback);
1234}
1235EXPORT_SYMBOL(end_page_writeback);
1236
1237/*
1238 * After completing I/O on a page, call this routine to update the page
1239 * flags appropriately
1240 */
1241void page_endio(struct page *page, bool is_write, int err)
1242{
1243 if (!is_write) {
1244 if (!err) {
1245 SetPageUptodate(page);
1246 } else {
1247 ClearPageUptodate(page);
1248 SetPageError(page);
1249 }
1250 unlock_page(page);
1251 } else {
1252 if (err) {
1253 struct address_space *mapping;
1254
1255 SetPageError(page);
1256 mapping = page_mapping(page);
1257 if (mapping)
1258 mapping_set_error(mapping, err);
1259 }
1260 end_page_writeback(page);
1261 }
1262}
1263EXPORT_SYMBOL_GPL(page_endio);
1264
1265/**
1266 * __lock_page - get a lock on the page, assuming we need to sleep to get it
1267 * @__page: the page to lock
1268 */
1269void __lock_page(struct page *__page)
1270{
1271 struct page *page = compound_head(__page);
1272 wait_queue_head_t *q = page_waitqueue(page);
1273 wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, true);
1274}
1275EXPORT_SYMBOL(__lock_page);
1276
1277int __lock_page_killable(struct page *__page)
1278{
1279 struct page *page = compound_head(__page);
1280 wait_queue_head_t *q = page_waitqueue(page);
1281 return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE, true);
1282}
1283EXPORT_SYMBOL_GPL(__lock_page_killable);
1284
1285/*
1286 * Return values:
1287 * 1 - page is locked; mmap_sem is still held.
1288 * 0 - page is not locked.
1289 * mmap_sem has been released (up_read()), unless flags had both
1290 * FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1291 * which case mmap_sem is still held.
1292 *
1293 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1294 * with the page locked and the mmap_sem unperturbed.
1295 */
1296int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1297 unsigned int flags)
1298{
1299 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1300 /*
1301 * CAUTION! In this case, mmap_sem is not released
1302 * even though return 0.
1303 */
1304 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1305 return 0;
1306
1307 up_read(&mm->mmap_sem);
1308 if (flags & FAULT_FLAG_KILLABLE)
1309 wait_on_page_locked_killable(page);
1310 else
1311 wait_on_page_locked(page);
1312 return 0;
1313 } else {
1314 if (flags & FAULT_FLAG_KILLABLE) {
1315 int ret;
1316
1317 ret = __lock_page_killable(page);
1318 if (ret) {
1319 up_read(&mm->mmap_sem);
1320 return 0;
1321 }
1322 } else
1323 __lock_page(page);
1324 return 1;
1325 }
1326}
1327
1328/**
1329 * page_cache_next_hole - find the next hole (not-present entry)
1330 * @mapping: mapping
1331 * @index: index
1332 * @max_scan: maximum range to search
1333 *
1334 * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
1335 * lowest indexed hole.
1336 *
1337 * Returns: the index of the hole if found, otherwise returns an index
1338 * outside of the set specified (in which case 'return - index >=
1339 * max_scan' will be true). In rare cases of index wrap-around, 0 will
1340 * be returned.
1341 *
1342 * page_cache_next_hole may be called under rcu_read_lock. However,
1343 * like radix_tree_gang_lookup, this will not atomically search a
1344 * snapshot of the tree at a single point in time. For example, if a
1345 * hole is created at index 5, then subsequently a hole is created at
1346 * index 10, page_cache_next_hole covering both indexes may return 10
1347 * if called under rcu_read_lock.
1348 */
1349pgoff_t page_cache_next_hole(struct address_space *mapping,
1350 pgoff_t index, unsigned long max_scan)
1351{
1352 unsigned long i;
1353
1354 for (i = 0; i < max_scan; i++) {
1355 struct page *page;
1356
1357 page = radix_tree_lookup(&mapping->i_pages, index);
1358 if (!page || radix_tree_exceptional_entry(page))
1359 break;
1360 index++;
1361 if (index == 0)
1362 break;
1363 }
1364
1365 return index;
1366}
1367EXPORT_SYMBOL(page_cache_next_hole);
1368
1369/**
1370 * page_cache_prev_hole - find the prev hole (not-present entry)
1371 * @mapping: mapping
1372 * @index: index
1373 * @max_scan: maximum range to search
1374 *
1375 * Search backwards in the range [max(index-max_scan+1, 0), index] for
1376 * the first hole.
1377 *
1378 * Returns: the index of the hole if found, otherwise returns an index
1379 * outside of the set specified (in which case 'index - return >=
1380 * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1381 * will be returned.
1382 *
1383 * page_cache_prev_hole may be called under rcu_read_lock. However,
1384 * like radix_tree_gang_lookup, this will not atomically search a
1385 * snapshot of the tree at a single point in time. For example, if a
1386 * hole is created at index 10, then subsequently a hole is created at
1387 * index 5, page_cache_prev_hole covering both indexes may return 5 if
1388 * called under rcu_read_lock.
1389 */
1390pgoff_t page_cache_prev_hole(struct address_space *mapping,
1391 pgoff_t index, unsigned long max_scan)
1392{
1393 unsigned long i;
1394
1395 for (i = 0; i < max_scan; i++) {
1396 struct page *page;
1397
1398 page = radix_tree_lookup(&mapping->i_pages, index);
1399 if (!page || radix_tree_exceptional_entry(page))
1400 break;
1401 index--;
1402 if (index == ULONG_MAX)
1403 break;
1404 }
1405
1406 return index;
1407}
1408EXPORT_SYMBOL(page_cache_prev_hole);
1409
1410/**
1411 * find_get_entry - find and get a page cache entry
1412 * @mapping: the address_space to search
1413 * @offset: the page cache index
1414 *
1415 * Looks up the page cache slot at @mapping & @offset. If there is a
1416 * page cache page, it is returned with an increased refcount.
1417 *
1418 * If the slot holds a shadow entry of a previously evicted page, or a
1419 * swap entry from shmem/tmpfs, it is returned.
1420 *
1421 * Otherwise, %NULL is returned.
1422 */
1423struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1424{
1425 void **pagep;
1426 struct page *head, *page;
1427
1428 rcu_read_lock();
1429repeat:
1430 page = NULL;
1431 pagep = radix_tree_lookup_slot(&mapping->i_pages, offset);
1432 if (pagep) {
1433 page = radix_tree_deref_slot(pagep);
1434 if (unlikely(!page))
1435 goto out;
1436 if (radix_tree_exception(page)) {
1437 if (radix_tree_deref_retry(page))
1438 goto repeat;
1439 /*
1440 * A shadow entry of a recently evicted page,
1441 * or a swap entry from shmem/tmpfs. Return
1442 * it without attempting to raise page count.
1443 */
1444 goto out;
1445 }
1446
1447 head = compound_head(page);
1448 if (!page_cache_get_speculative(head))
1449 goto repeat;
1450
1451 /* The page was split under us? */
1452 if (compound_head(page) != head) {
1453 put_page(head);
1454 goto repeat;
1455 }
1456
1457 /*
1458 * Has the page moved?
1459 * This is part of the lockless pagecache protocol. See
1460 * include/linux/pagemap.h for details.
1461 */
1462 if (unlikely(page != *pagep)) {
1463 put_page(head);
1464 goto repeat;
1465 }
1466 }
1467out:
1468 rcu_read_unlock();
1469
1470 return page;
1471}
1472EXPORT_SYMBOL(find_get_entry);
1473
1474/**
1475 * find_lock_entry - locate, pin and lock a page cache entry
1476 * @mapping: the address_space to search
1477 * @offset: the page cache index
1478 *
1479 * Looks up the page cache slot at @mapping & @offset. If there is a
1480 * page cache page, it is returned locked and with an increased
1481 * refcount.
1482 *
1483 * If the slot holds a shadow entry of a previously evicted page, or a
1484 * swap entry from shmem/tmpfs, it is returned.
1485 *
1486 * Otherwise, %NULL is returned.
1487 *
1488 * find_lock_entry() may sleep.
1489 */
1490struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1491{
1492 struct page *page;
1493
1494repeat:
1495 page = find_get_entry(mapping, offset);
1496 if (page && !radix_tree_exception(page)) {
1497 lock_page(page);
1498 /* Has the page been truncated? */
1499 if (unlikely(page_mapping(page) != mapping)) {
1500 unlock_page(page);
1501 put_page(page);
1502 goto repeat;
1503 }
1504 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1505 }
1506 return page;
1507}
1508EXPORT_SYMBOL(find_lock_entry);
1509
1510/**
1511 * pagecache_get_page - find and get a page reference
1512 * @mapping: the address_space to search
1513 * @offset: the page index
1514 * @fgp_flags: PCG flags
1515 * @gfp_mask: gfp mask to use for the page cache data page allocation
1516 *
1517 * Looks up the page cache slot at @mapping & @offset.
1518 *
1519 * PCG flags modify how the page is returned.
1520 *
1521 * @fgp_flags can be:
1522 *
1523 * - FGP_ACCESSED: the page will be marked accessed
1524 * - FGP_LOCK: Page is return locked
1525 * - FGP_CREAT: If page is not present then a new page is allocated using
1526 * @gfp_mask and added to the page cache and the VM's LRU
1527 * list. The page is returned locked and with an increased
1528 * refcount. Otherwise, NULL is returned.
1529 *
1530 * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1531 * if the GFP flags specified for FGP_CREAT are atomic.
1532 *
1533 * If there is a page cache page, it is returned with an increased refcount.
1534 */
1535struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1536 int fgp_flags, gfp_t gfp_mask)
1537{
1538 struct page *page;
1539
1540repeat:
1541 page = find_get_entry(mapping, offset);
1542 if (radix_tree_exceptional_entry(page))
1543 page = NULL;
1544 if (!page)
1545 goto no_page;
1546
1547 if (fgp_flags & FGP_LOCK) {
1548 if (fgp_flags & FGP_NOWAIT) {
1549 if (!trylock_page(page)) {
1550 put_page(page);
1551 return NULL;
1552 }
1553 } else {
1554 lock_page(page);
1555 }
1556
1557 /* Has the page been truncated? */
1558 if (unlikely(page->mapping != mapping)) {
1559 unlock_page(page);
1560 put_page(page);
1561 goto repeat;
1562 }
1563 VM_BUG_ON_PAGE(page->index != offset, page);
1564 }
1565
1566 if (page && (fgp_flags & FGP_ACCESSED))
1567 mark_page_accessed(page);
1568
1569no_page:
1570 if (!page && (fgp_flags & FGP_CREAT)) {
1571 int err;
1572 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1573 gfp_mask |= __GFP_WRITE;
1574 if (fgp_flags & FGP_NOFS)
1575 gfp_mask &= ~__GFP_FS;
1576
1577 page = __page_cache_alloc(gfp_mask);
1578 if (!page)
1579 return NULL;
1580
1581 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1582 fgp_flags |= FGP_LOCK;
1583
1584 /* Init accessed so avoid atomic mark_page_accessed later */
1585 if (fgp_flags & FGP_ACCESSED)
1586 __SetPageReferenced(page);
1587
1588 err = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
1589 if (unlikely(err)) {
1590 put_page(page);
1591 page = NULL;
1592 if (err == -EEXIST)
1593 goto repeat;
1594 }
1595 }
1596
1597 return page;
1598}
1599EXPORT_SYMBOL(pagecache_get_page);
1600
1601/**
1602 * find_get_entries - gang pagecache lookup
1603 * @mapping: The address_space to search
1604 * @start: The starting page cache index
1605 * @nr_entries: The maximum number of entries
1606 * @entries: Where the resulting entries are placed
1607 * @indices: The cache indices corresponding to the entries in @entries
1608 *
1609 * find_get_entries() will search for and return a group of up to
1610 * @nr_entries entries in the mapping. The entries are placed at
1611 * @entries. find_get_entries() takes a reference against any actual
1612 * pages it returns.
1613 *
1614 * The search returns a group of mapping-contiguous page cache entries
1615 * with ascending indexes. There may be holes in the indices due to
1616 * not-present pages.
1617 *
1618 * Any shadow entries of evicted pages, or swap entries from
1619 * shmem/tmpfs, are included in the returned array.
1620 *
1621 * find_get_entries() returns the number of pages and shadow entries
1622 * which were found.
1623 */
1624unsigned find_get_entries(struct address_space *mapping,
1625 pgoff_t start, unsigned int nr_entries,
1626 struct page **entries, pgoff_t *indices)
1627{
1628 void **slot;
1629 unsigned int ret = 0;
1630 struct radix_tree_iter iter;
1631
1632 if (!nr_entries)
1633 return 0;
1634
1635 rcu_read_lock();
1636 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
1637 struct page *head, *page;
1638repeat:
1639 page = radix_tree_deref_slot(slot);
1640 if (unlikely(!page))
1641 continue;
1642 if (radix_tree_exception(page)) {
1643 if (radix_tree_deref_retry(page)) {
1644 slot = radix_tree_iter_retry(&iter);
1645 continue;
1646 }
1647 /*
1648 * A shadow entry of a recently evicted page, a swap
1649 * entry from shmem/tmpfs or a DAX entry. Return it
1650 * without attempting to raise page count.
1651 */
1652 goto export;
1653 }
1654
1655 head = compound_head(page);
1656 if (!page_cache_get_speculative(head))
1657 goto repeat;
1658
1659 /* The page was split under us? */
1660 if (compound_head(page) != head) {
1661 put_page(head);
1662 goto repeat;
1663 }
1664
1665 /* Has the page moved? */
1666 if (unlikely(page != *slot)) {
1667 put_page(head);
1668 goto repeat;
1669 }
1670export:
1671 indices[ret] = iter.index;
1672 entries[ret] = page;
1673 if (++ret == nr_entries)
1674 break;
1675 }
1676 rcu_read_unlock();
1677 return ret;
1678}
1679
1680/**
1681 * find_get_pages_range - gang pagecache lookup
1682 * @mapping: The address_space to search
1683 * @start: The starting page index
1684 * @end: The final page index (inclusive)
1685 * @nr_pages: The maximum number of pages
1686 * @pages: Where the resulting pages are placed
1687 *
1688 * find_get_pages_range() will search for and return a group of up to @nr_pages
1689 * pages in the mapping starting at index @start and up to index @end
1690 * (inclusive). The pages are placed at @pages. find_get_pages_range() takes
1691 * a reference against the returned pages.
1692 *
1693 * The search returns a group of mapping-contiguous pages with ascending
1694 * indexes. There may be holes in the indices due to not-present pages.
1695 * We also update @start to index the next page for the traversal.
1696 *
1697 * find_get_pages_range() returns the number of pages which were found. If this
1698 * number is smaller than @nr_pages, the end of specified range has been
1699 * reached.
1700 */
1701unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1702 pgoff_t end, unsigned int nr_pages,
1703 struct page **pages)
1704{
1705 struct radix_tree_iter iter;
1706 void **slot;
1707 unsigned ret = 0;
1708
1709 if (unlikely(!nr_pages))
1710 return 0;
1711
1712 rcu_read_lock();
1713 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, *start) {
1714 struct page *head, *page;
1715
1716 if (iter.index > end)
1717 break;
1718repeat:
1719 page = radix_tree_deref_slot(slot);
1720 if (unlikely(!page))
1721 continue;
1722
1723 if (radix_tree_exception(page)) {
1724 if (radix_tree_deref_retry(page)) {
1725 slot = radix_tree_iter_retry(&iter);
1726 continue;
1727 }
1728 /*
1729 * A shadow entry of a recently evicted page,
1730 * or a swap entry from shmem/tmpfs. Skip
1731 * over it.
1732 */
1733 continue;
1734 }
1735
1736 head = compound_head(page);
1737 if (!page_cache_get_speculative(head))
1738 goto repeat;
1739
1740 /* The page was split under us? */
1741 if (compound_head(page) != head) {
1742 put_page(head);
1743 goto repeat;
1744 }
1745
1746 /* Has the page moved? */
1747 if (unlikely(page != *slot)) {
1748 put_page(head);
1749 goto repeat;
1750 }
1751
1752 pages[ret] = page;
1753 if (++ret == nr_pages) {
1754 *start = pages[ret - 1]->index + 1;
1755 goto out;
1756 }
1757 }
1758
1759 /*
1760 * We come here when there is no page beyond @end. We take care to not
1761 * overflow the index @start as it confuses some of the callers. This
1762 * breaks the iteration when there is page at index -1 but that is
1763 * already broken anyway.
1764 */
1765 if (end == (pgoff_t)-1)
1766 *start = (pgoff_t)-1;
1767 else
1768 *start = end + 1;
1769out:
1770 rcu_read_unlock();
1771
1772 return ret;
1773}
1774
1775/**
1776 * find_get_pages_contig - gang contiguous pagecache lookup
1777 * @mapping: The address_space to search
1778 * @index: The starting page index
1779 * @nr_pages: The maximum number of pages
1780 * @pages: Where the resulting pages are placed
1781 *
1782 * find_get_pages_contig() works exactly like find_get_pages(), except
1783 * that the returned number of pages are guaranteed to be contiguous.
1784 *
1785 * find_get_pages_contig() returns the number of pages which were found.
1786 */
1787unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1788 unsigned int nr_pages, struct page **pages)
1789{
1790 struct radix_tree_iter iter;
1791 void **slot;
1792 unsigned int ret = 0;
1793
1794 if (unlikely(!nr_pages))
1795 return 0;
1796
1797 rcu_read_lock();
1798 radix_tree_for_each_contig(slot, &mapping->i_pages, &iter, index) {
1799 struct page *head, *page;
1800repeat:
1801 page = radix_tree_deref_slot(slot);
1802 /* The hole, there no reason to continue */
1803 if (unlikely(!page))
1804 break;
1805
1806 if (radix_tree_exception(page)) {
1807 if (radix_tree_deref_retry(page)) {
1808 slot = radix_tree_iter_retry(&iter);
1809 continue;
1810 }
1811 /*
1812 * A shadow entry of a recently evicted page,
1813 * or a swap entry from shmem/tmpfs. Stop
1814 * looking for contiguous pages.
1815 */
1816 break;
1817 }
1818
1819 head = compound_head(page);
1820 if (!page_cache_get_speculative(head))
1821 goto repeat;
1822
1823 /* The page was split under us? */
1824 if (compound_head(page) != head) {
1825 put_page(head);
1826 goto repeat;
1827 }
1828
1829 /* Has the page moved? */
1830 if (unlikely(page != *slot)) {
1831 put_page(head);
1832 goto repeat;
1833 }
1834
1835 /*
1836 * must check mapping and index after taking the ref.
1837 * otherwise we can get both false positives and false
1838 * negatives, which is just confusing to the caller.
1839 */
1840 if (page->mapping == NULL || page_to_pgoff(page) != iter.index) {
1841 put_page(page);
1842 break;
1843 }
1844
1845 pages[ret] = page;
1846 if (++ret == nr_pages)
1847 break;
1848 }
1849 rcu_read_unlock();
1850 return ret;
1851}
1852EXPORT_SYMBOL(find_get_pages_contig);
1853
1854/**
1855 * find_get_pages_range_tag - find and return pages in given range matching @tag
1856 * @mapping: the address_space to search
1857 * @index: the starting page index
1858 * @end: The final page index (inclusive)
1859 * @tag: the tag index
1860 * @nr_pages: the maximum number of pages
1861 * @pages: where the resulting pages are placed
1862 *
1863 * Like find_get_pages, except we only return pages which are tagged with
1864 * @tag. We update @index to index the next page for the traversal.
1865 */
1866unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
1867 pgoff_t end, int tag, unsigned int nr_pages,
1868 struct page **pages)
1869{
1870 struct radix_tree_iter iter;
1871 void **slot;
1872 unsigned ret = 0;
1873
1874 if (unlikely(!nr_pages))
1875 return 0;
1876
1877 rcu_read_lock();
1878 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, *index, tag) {
1879 struct page *head, *page;
1880
1881 if (iter.index > end)
1882 break;
1883repeat:
1884 page = radix_tree_deref_slot(slot);
1885 if (unlikely(!page))
1886 continue;
1887
1888 if (radix_tree_exception(page)) {
1889 if (radix_tree_deref_retry(page)) {
1890 slot = radix_tree_iter_retry(&iter);
1891 continue;
1892 }
1893 /*
1894 * A shadow entry of a recently evicted page.
1895 *
1896 * Those entries should never be tagged, but
1897 * this tree walk is lockless and the tags are
1898 * looked up in bulk, one radix tree node at a
1899 * time, so there is a sizable window for page
1900 * reclaim to evict a page we saw tagged.
1901 *
1902 * Skip over it.
1903 */
1904 continue;
1905 }
1906
1907 head = compound_head(page);
1908 if (!page_cache_get_speculative(head))
1909 goto repeat;
1910
1911 /* The page was split under us? */
1912 if (compound_head(page) != head) {
1913 put_page(head);
1914 goto repeat;
1915 }
1916
1917 /* Has the page moved? */
1918 if (unlikely(page != *slot)) {
1919 put_page(head);
1920 goto repeat;
1921 }
1922
1923 pages[ret] = page;
1924 if (++ret == nr_pages) {
1925 *index = pages[ret - 1]->index + 1;
1926 goto out;
1927 }
1928 }
1929
1930 /*
1931 * We come here when we got at @end. We take care to not overflow the
1932 * index @index as it confuses some of the callers. This breaks the
1933 * iteration when there is page at index -1 but that is already broken
1934 * anyway.
1935 */
1936 if (end == (pgoff_t)-1)
1937 *index = (pgoff_t)-1;
1938 else
1939 *index = end + 1;
1940out:
1941 rcu_read_unlock();
1942
1943 return ret;
1944}
1945EXPORT_SYMBOL(find_get_pages_range_tag);
1946
1947/**
1948 * find_get_entries_tag - find and return entries that match @tag
1949 * @mapping: the address_space to search
1950 * @start: the starting page cache index
1951 * @tag: the tag index
1952 * @nr_entries: the maximum number of entries
1953 * @entries: where the resulting entries are placed
1954 * @indices: the cache indices corresponding to the entries in @entries
1955 *
1956 * Like find_get_entries, except we only return entries which are tagged with
1957 * @tag.
1958 */
1959unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1960 int tag, unsigned int nr_entries,
1961 struct page **entries, pgoff_t *indices)
1962{
1963 void **slot;
1964 unsigned int ret = 0;
1965 struct radix_tree_iter iter;
1966
1967 if (!nr_entries)
1968 return 0;
1969
1970 rcu_read_lock();
1971 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start, tag) {
1972 struct page *head, *page;
1973repeat:
1974 page = radix_tree_deref_slot(slot);
1975 if (unlikely(!page))
1976 continue;
1977 if (radix_tree_exception(page)) {
1978 if (radix_tree_deref_retry(page)) {
1979 slot = radix_tree_iter_retry(&iter);
1980 continue;
1981 }
1982
1983 /*
1984 * A shadow entry of a recently evicted page, a swap
1985 * entry from shmem/tmpfs or a DAX entry. Return it
1986 * without attempting to raise page count.
1987 */
1988 goto export;
1989 }
1990
1991 head = compound_head(page);
1992 if (!page_cache_get_speculative(head))
1993 goto repeat;
1994
1995 /* The page was split under us? */
1996 if (compound_head(page) != head) {
1997 put_page(head);
1998 goto repeat;
1999 }
2000
2001 /* Has the page moved? */
2002 if (unlikely(page != *slot)) {
2003 put_page(head);
2004 goto repeat;
2005 }
2006export:
2007 indices[ret] = iter.index;
2008 entries[ret] = page;
2009 if (++ret == nr_entries)
2010 break;
2011 }
2012 rcu_read_unlock();
2013 return ret;
2014}
2015EXPORT_SYMBOL(find_get_entries_tag);
2016
2017/*
2018 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2019 * a _large_ part of the i/o request. Imagine the worst scenario:
2020 *
2021 * ---R__________________________________________B__________
2022 * ^ reading here ^ bad block(assume 4k)
2023 *
2024 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2025 * => failing the whole request => read(R) => read(R+1) =>
2026 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2027 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2028 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2029 *
2030 * It is going insane. Fix it by quickly scaling down the readahead size.
2031 */
2032static void shrink_readahead_size_eio(struct file *filp,
2033 struct file_ra_state *ra)
2034{
2035 ra->ra_pages /= 4;
2036}
2037
2038/**
2039 * generic_file_buffered_read - generic file read routine
2040 * @iocb: the iocb to read
2041 * @iter: data destination
2042 * @written: already copied
2043 *
2044 * This is a generic file read routine, and uses the
2045 * mapping->a_ops->readpage() function for the actual low-level stuff.
2046 *
2047 * This is really ugly. But the goto's actually try to clarify some
2048 * of the logic when it comes to error handling etc.
2049 */
2050static ssize_t generic_file_buffered_read(struct kiocb *iocb,
2051 struct iov_iter *iter, ssize_t written)
2052{
2053 struct file *filp = iocb->ki_filp;
2054 struct address_space *mapping = filp->f_mapping;
2055 struct inode *inode = mapping->host;
2056 struct file_ra_state *ra = &filp->f_ra;
2057 loff_t *ppos = &iocb->ki_pos;
2058 pgoff_t index;
2059 pgoff_t last_index;
2060 pgoff_t prev_index;
2061 unsigned long offset; /* offset into pagecache page */
2062 unsigned int prev_offset;
2063 int error = 0;
2064
2065 if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
2066 return 0;
2067 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2068
2069 index = *ppos >> PAGE_SHIFT;
2070 prev_index = ra->prev_pos >> PAGE_SHIFT;
2071 prev_offset = ra->prev_pos & (PAGE_SIZE-1);
2072 last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
2073 offset = *ppos & ~PAGE_MASK;
2074
2075 for (;;) {
2076 struct page *page;
2077 pgoff_t end_index;
2078 loff_t isize;
2079 unsigned long nr, ret;
2080
2081 cond_resched();
2082find_page:
2083 if (fatal_signal_pending(current)) {
2084 error = -EINTR;
2085 goto out;
2086 }
2087
2088 page = find_get_page(mapping, index);
2089 if (!page) {
2090 if (iocb->ki_flags & IOCB_NOWAIT)
2091 goto would_block;
2092 page_cache_sync_readahead(mapping,
2093 ra, filp,
2094 index, last_index - index);
2095 page = find_get_page(mapping, index);
2096 if (unlikely(page == NULL))
2097 goto no_cached_page;
2098 }
2099 if (PageReadahead(page)) {
2100 page_cache_async_readahead(mapping,
2101 ra, filp, page,
2102 index, last_index - index);
2103 }
2104 if (!PageUptodate(page)) {
2105 if (iocb->ki_flags & IOCB_NOWAIT) {
2106 put_page(page);
2107 goto would_block;
2108 }
2109
2110 /*
2111 * See comment in do_read_cache_page on why
2112 * wait_on_page_locked is used to avoid unnecessarily
2113 * serialisations and why it's safe.
2114 */
2115 error = wait_on_page_locked_killable(page);
2116 if (unlikely(error))
2117 goto readpage_error;
2118 if (PageUptodate(page))
2119 goto page_ok;
2120
2121 if (inode->i_blkbits == PAGE_SHIFT ||
2122 !mapping->a_ops->is_partially_uptodate)
2123 goto page_not_up_to_date;
2124 /* pipes can't handle partially uptodate pages */
2125 if (unlikely(iter->type & ITER_PIPE))
2126 goto page_not_up_to_date;
2127 if (!trylock_page(page))
2128 goto page_not_up_to_date;
2129 /* Did it get truncated before we got the lock? */
2130 if (!page->mapping)
2131 goto page_not_up_to_date_locked;
2132 if (!mapping->a_ops->is_partially_uptodate(page,
2133 offset, iter->count))
2134 goto page_not_up_to_date_locked;
2135 unlock_page(page);
2136 }
2137page_ok:
2138 /*
2139 * i_size must be checked after we know the page is Uptodate.
2140 *
2141 * Checking i_size after the check allows us to calculate
2142 * the correct value for "nr", which means the zero-filled
2143 * part of the page is not copied back to userspace (unless
2144 * another truncate extends the file - this is desired though).
2145 */
2146
2147 isize = i_size_read(inode);
2148 end_index = (isize - 1) >> PAGE_SHIFT;
2149 if (unlikely(!isize || index > end_index)) {
2150 put_page(page);
2151 goto out;
2152 }
2153
2154 /* nr is the maximum number of bytes to copy from this page */
2155 nr = PAGE_SIZE;
2156 if (index == end_index) {
2157 nr = ((isize - 1) & ~PAGE_MASK) + 1;
2158 if (nr <= offset) {
2159 put_page(page);
2160 goto out;
2161 }
2162 }
2163 nr = nr - offset;
2164
2165 /* If users can be writing to this page using arbitrary
2166 * virtual addresses, take care about potential aliasing
2167 * before reading the page on the kernel side.
2168 */
2169 if (mapping_writably_mapped(mapping))
2170 flush_dcache_page(page);
2171
2172 /*
2173 * When a sequential read accesses a page several times,
2174 * only mark it as accessed the first time.
2175 */
2176 if (prev_index != index || offset != prev_offset)
2177 mark_page_accessed(page);
2178 prev_index = index;
2179
2180 /*
2181 * Ok, we have the page, and it's up-to-date, so
2182 * now we can copy it to user space...
2183 */
2184
2185 ret = copy_page_to_iter(page, offset, nr, iter);
2186 offset += ret;
2187 index += offset >> PAGE_SHIFT;
2188 offset &= ~PAGE_MASK;
2189 prev_offset = offset;
2190
2191 put_page(page);
2192 written += ret;
2193 if (!iov_iter_count(iter))
2194 goto out;
2195 if (ret < nr) {
2196 error = -EFAULT;
2197 goto out;
2198 }
2199 continue;
2200
2201page_not_up_to_date:
2202 /* Get exclusive access to the page ... */
2203 error = lock_page_killable(page);
2204 if (unlikely(error))
2205 goto readpage_error;
2206
2207page_not_up_to_date_locked:
2208 /* Did it get truncated before we got the lock? */
2209 if (!page->mapping) {
2210 unlock_page(page);
2211 put_page(page);
2212 continue;
2213 }
2214
2215 /* Did somebody else fill it already? */
2216 if (PageUptodate(page)) {
2217 unlock_page(page);
2218 goto page_ok;
2219 }
2220
2221readpage:
2222 /*
2223 * A previous I/O error may have been due to temporary
2224 * failures, eg. multipath errors.
2225 * PG_error will be set again if readpage fails.
2226 */
2227 ClearPageError(page);
2228 /* Start the actual read. The read will unlock the page. */
2229 error = mapping->a_ops->readpage(filp, page);
2230
2231 if (unlikely(error)) {
2232 if (error == AOP_TRUNCATED_PAGE) {
2233 put_page(page);
2234 error = 0;
2235 goto find_page;
2236 }
2237 goto readpage_error;
2238 }
2239
2240 if (!PageUptodate(page)) {
2241 error = lock_page_killable(page);
2242 if (unlikely(error))
2243 goto readpage_error;
2244 if (!PageUptodate(page)) {
2245 if (page->mapping == NULL) {
2246 /*
2247 * invalidate_mapping_pages got it
2248 */
2249 unlock_page(page);
2250 put_page(page);
2251 goto find_page;
2252 }
2253 unlock_page(page);
2254 shrink_readahead_size_eio(filp, ra);
2255 error = -EIO;
2256 goto readpage_error;
2257 }
2258 unlock_page(page);
2259 }
2260
2261 goto page_ok;
2262
2263readpage_error:
2264 /* UHHUH! A synchronous read error occurred. Report it */
2265 put_page(page);
2266 goto out;
2267
2268no_cached_page:
2269 /*
2270 * Ok, it wasn't cached, so we need to create a new
2271 * page..
2272 */
2273 page = page_cache_alloc(mapping);
2274 if (!page) {
2275 error = -ENOMEM;
2276 goto out;
2277 }
2278 error = add_to_page_cache_lru(page, mapping, index,
2279 mapping_gfp_constraint(mapping, GFP_KERNEL));
2280 if (error) {
2281 put_page(page);
2282 if (error == -EEXIST) {
2283 error = 0;
2284 goto find_page;
2285 }
2286 goto out;
2287 }
2288 goto readpage;
2289 }
2290
2291would_block:
2292 error = -EAGAIN;
2293out:
2294 ra->prev_pos = prev_index;
2295 ra->prev_pos <<= PAGE_SHIFT;
2296 ra->prev_pos |= prev_offset;
2297
2298 *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2299 file_accessed(filp);
2300 return written ? written : error;
2301}
2302
2303/**
2304 * generic_file_read_iter - generic filesystem read routine
2305 * @iocb: kernel I/O control block
2306 * @iter: destination for the data read
2307 *
2308 * This is the "read_iter()" routine for all filesystems
2309 * that can use the page cache directly.
2310 */
2311ssize_t
2312generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2313{
2314 size_t count = iov_iter_count(iter);
2315 ssize_t retval = 0;
2316
2317 if (!count)
2318 goto out; /* skip atime */
2319
2320 if (iocb->ki_flags & IOCB_DIRECT) {
2321 struct file *file = iocb->ki_filp;
2322 struct address_space *mapping = file->f_mapping;
2323 struct inode *inode = mapping->host;
2324 loff_t size;
2325
2326 size = i_size_read(inode);
2327 if (iocb->ki_flags & IOCB_NOWAIT) {
2328 if (filemap_range_has_page(mapping, iocb->ki_pos,
2329 iocb->ki_pos + count - 1))
2330 return -EAGAIN;
2331 } else {
2332 retval = filemap_write_and_wait_range(mapping,
2333 iocb->ki_pos,
2334 iocb->ki_pos + count - 1);
2335 if (retval < 0)
2336 goto out;
2337 }
2338
2339 file_accessed(file);
2340
2341 retval = mapping->a_ops->direct_IO(iocb, iter);
2342 if (retval >= 0) {
2343 iocb->ki_pos += retval;
2344 count -= retval;
2345 }
2346 iov_iter_revert(iter, count - iov_iter_count(iter));
2347
2348 /*
2349 * Btrfs can have a short DIO read if we encounter
2350 * compressed extents, so if there was an error, or if
2351 * we've already read everything we wanted to, or if
2352 * there was a short read because we hit EOF, go ahead
2353 * and return. Otherwise fallthrough to buffered io for
2354 * the rest of the read. Buffered reads will not work for
2355 * DAX files, so don't bother trying.
2356 */
2357 if (retval < 0 || !count || iocb->ki_pos >= size ||
2358 IS_DAX(inode))
2359 goto out;
2360 }
2361
2362 retval = generic_file_buffered_read(iocb, iter, retval);
2363out:
2364 return retval;
2365}
2366EXPORT_SYMBOL(generic_file_read_iter);
2367
2368#ifdef CONFIG_MMU
2369/**
2370 * page_cache_read - adds requested page to the page cache if not already there
2371 * @file: file to read
2372 * @offset: page index
2373 * @gfp_mask: memory allocation flags
2374 *
2375 * This adds the requested page to the page cache if it isn't already there,
2376 * and schedules an I/O to read in its contents from disk.
2377 */
2378static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
2379{
2380 struct address_space *mapping = file->f_mapping;
2381 struct page *page;
2382 int ret;
2383
2384 do {
2385 page = __page_cache_alloc(gfp_mask);
2386 if (!page)
2387 return -ENOMEM;
2388
2389 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask);
2390 if (ret == 0)
2391 ret = mapping->a_ops->readpage(file, page);
2392 else if (ret == -EEXIST)
2393 ret = 0; /* losing race to add is OK */
2394
2395 put_page(page);
2396
2397 } while (ret == AOP_TRUNCATED_PAGE);
2398
2399 return ret;
2400}
2401
2402#define MMAP_LOTSAMISS (100)
2403
2404/*
2405 * Synchronous readahead happens when we don't even find
2406 * a page in the page cache at all.
2407 */
2408static void do_sync_mmap_readahead(struct vm_area_struct *vma,
2409 struct file_ra_state *ra,
2410 struct file *file,
2411 pgoff_t offset)
2412{
2413 struct address_space *mapping = file->f_mapping;
2414
2415 /* If we don't want any read-ahead, don't bother */
2416 if (vma->vm_flags & VM_RAND_READ)
2417 return;
2418 if (!ra->ra_pages)
2419 return;
2420
2421 if (vma->vm_flags & VM_SEQ_READ) {
2422 page_cache_sync_readahead(mapping, ra, file, offset,
2423 ra->ra_pages);
2424 return;
2425 }
2426
2427 /* Avoid banging the cache line if not needed */
2428 if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
2429 ra->mmap_miss++;
2430
2431 /*
2432 * Do we miss much more than hit in this file? If so,
2433 * stop bothering with read-ahead. It will only hurt.
2434 */
2435 if (ra->mmap_miss > MMAP_LOTSAMISS)
2436 return;
2437
2438 /*
2439 * mmap read-around
2440 */
2441 ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2442 ra->size = ra->ra_pages;
2443 ra->async_size = ra->ra_pages / 4;
2444 ra_submit(ra, mapping, file);
2445}
2446
2447/*
2448 * Asynchronous readahead happens when we find the page and PG_readahead,
2449 * so we want to possibly extend the readahead further..
2450 */
2451static void do_async_mmap_readahead(struct vm_area_struct *vma,
2452 struct file_ra_state *ra,
2453 struct file *file,
2454 struct page *page,
2455 pgoff_t offset)
2456{
2457 struct address_space *mapping = file->f_mapping;
2458
2459 /* If we don't want any read-ahead, don't bother */
2460 if (vma->vm_flags & VM_RAND_READ)
2461 return;
2462 if (ra->mmap_miss > 0)
2463 ra->mmap_miss--;
2464 if (PageReadahead(page))
2465 page_cache_async_readahead(mapping, ra, file,
2466 page, offset, ra->ra_pages);
2467}
2468
2469/**
2470 * filemap_fault - read in file data for page fault handling
2471 * @vmf: struct vm_fault containing details of the fault
2472 *
2473 * filemap_fault() is invoked via the vma operations vector for a
2474 * mapped memory region to read in file data during a page fault.
2475 *
2476 * The goto's are kind of ugly, but this streamlines the normal case of having
2477 * it in the page cache, and handles the special cases reasonably without
2478 * having a lot of duplicated code.
2479 *
2480 * vma->vm_mm->mmap_sem must be held on entry.
2481 *
2482 * If our return value has VM_FAULT_RETRY set, it's because
2483 * lock_page_or_retry() returned 0.
2484 * The mmap_sem has usually been released in this case.
2485 * See __lock_page_or_retry() for the exception.
2486 *
2487 * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2488 * has not been released.
2489 *
2490 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2491 */
2492int filemap_fault(struct vm_fault *vmf)
2493{
2494 int error;
2495 struct file *file = vmf->vma->vm_file;
2496 struct address_space *mapping = file->f_mapping;
2497 struct file_ra_state *ra = &file->f_ra;
2498 struct inode *inode = mapping->host;
2499 pgoff_t offset = vmf->pgoff;
2500 pgoff_t max_off;
2501 struct page *page;
2502 int ret = 0;
2503
2504 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2505 if (unlikely(offset >= max_off))
2506 return VM_FAULT_SIGBUS;
2507
2508 /*
2509 * Do we have something in the page cache already?
2510 */
2511 page = find_get_page(mapping, offset);
2512 if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2513 /*
2514 * We found the page, so try async readahead before
2515 * waiting for the lock.
2516 */
2517 do_async_mmap_readahead(vmf->vma, ra, file, page, offset);
2518 } else if (!page) {
2519 /* No page in the page cache at all */
2520 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
2521 count_vm_event(PGMAJFAULT);
2522 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2523 ret = VM_FAULT_MAJOR;
2524retry_find:
2525 page = find_get_page(mapping, offset);
2526 if (!page)
2527 goto no_cached_page;
2528 }
2529
2530 if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
2531 put_page(page);
2532 return ret | VM_FAULT_RETRY;
2533 }
2534
2535 /* Did it get truncated? */
2536 if (unlikely(page->mapping != mapping)) {
2537 unlock_page(page);
2538 put_page(page);
2539 goto retry_find;
2540 }
2541 VM_BUG_ON_PAGE(page->index != offset, page);
2542
2543 /*
2544 * We have a locked page in the page cache, now we need to check
2545 * that it's up-to-date. If not, it is going to be due to an error.
2546 */
2547 if (unlikely(!PageUptodate(page)))
2548 goto page_not_uptodate;
2549
2550 /*
2551 * Found the page and have a reference on it.
2552 * We must recheck i_size under page lock.
2553 */
2554 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2555 if (unlikely(offset >= max_off)) {
2556 unlock_page(page);
2557 put_page(page);
2558 return VM_FAULT_SIGBUS;
2559 }
2560
2561 vmf->page = page;
2562 return ret | VM_FAULT_LOCKED;
2563
2564no_cached_page:
2565 /*
2566 * We're only likely to ever get here if MADV_RANDOM is in
2567 * effect.
2568 */
2569 error = page_cache_read(file, offset, vmf->gfp_mask);
2570
2571 /*
2572 * The page we want has now been added to the page cache.
2573 * In the unlikely event that someone removed it in the
2574 * meantime, we'll just come back here and read it again.
2575 */
2576 if (error >= 0)
2577 goto retry_find;
2578
2579 /*
2580 * An error return from page_cache_read can result if the
2581 * system is low on memory, or a problem occurs while trying
2582 * to schedule I/O.
2583 */
2584 if (error == -ENOMEM)
2585 return VM_FAULT_OOM;
2586 return VM_FAULT_SIGBUS;
2587
2588page_not_uptodate:
2589 /*
2590 * Umm, take care of errors if the page isn't up-to-date.
2591 * Try to re-read it _once_. We do this synchronously,
2592 * because there really aren't any performance issues here
2593 * and we need to check for errors.
2594 */
2595 ClearPageError(page);
2596 error = mapping->a_ops->readpage(file, page);
2597 if (!error) {
2598 wait_on_page_locked(page);
2599 if (!PageUptodate(page))
2600 error = -EIO;
2601 }
2602 put_page(page);
2603
2604 if (!error || error == AOP_TRUNCATED_PAGE)
2605 goto retry_find;
2606
2607 /* Things didn't work out. Return zero to tell the mm layer so. */
2608 shrink_readahead_size_eio(file, ra);
2609 return VM_FAULT_SIGBUS;
2610}
2611EXPORT_SYMBOL(filemap_fault);
2612
2613void filemap_map_pages(struct vm_fault *vmf,
2614 pgoff_t start_pgoff, pgoff_t end_pgoff)
2615{
2616 struct radix_tree_iter iter;
2617 void **slot;
2618 struct file *file = vmf->vma->vm_file;
2619 struct address_space *mapping = file->f_mapping;
2620 pgoff_t last_pgoff = start_pgoff;
2621 unsigned long max_idx;
2622 struct page *head, *page;
2623
2624 rcu_read_lock();
2625 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start_pgoff) {
2626 if (iter.index > end_pgoff)
2627 break;
2628repeat:
2629 page = radix_tree_deref_slot(slot);
2630 if (unlikely(!page))
2631 goto next;
2632 if (radix_tree_exception(page)) {
2633 if (radix_tree_deref_retry(page)) {
2634 slot = radix_tree_iter_retry(&iter);
2635 continue;
2636 }
2637 goto next;
2638 }
2639
2640 head = compound_head(page);
2641 if (!page_cache_get_speculative(head))
2642 goto repeat;
2643
2644 /* The page was split under us? */
2645 if (compound_head(page) != head) {
2646 put_page(head);
2647 goto repeat;
2648 }
2649
2650 /* Has the page moved? */
2651 if (unlikely(page != *slot)) {
2652 put_page(head);
2653 goto repeat;
2654 }
2655
2656 if (!PageUptodate(page) ||
2657 PageReadahead(page) ||
2658 PageHWPoison(page))
2659 goto skip;
2660 if (!trylock_page(page))
2661 goto skip;
2662
2663 if (page->mapping != mapping || !PageUptodate(page))
2664 goto unlock;
2665
2666 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2667 if (page->index >= max_idx)
2668 goto unlock;
2669
2670 if (file->f_ra.mmap_miss > 0)
2671 file->f_ra.mmap_miss--;
2672
2673 vmf->address += (iter.index - last_pgoff) << PAGE_SHIFT;
2674 if (vmf->pte)
2675 vmf->pte += iter.index - last_pgoff;
2676 last_pgoff = iter.index;
2677 if (alloc_set_pte(vmf, NULL, page))
2678 goto unlock;
2679 unlock_page(page);
2680 goto next;
2681unlock:
2682 unlock_page(page);
2683skip:
2684 put_page(page);
2685next:
2686 /* Huge page is mapped? No need to proceed. */
2687 if (pmd_trans_huge(*vmf->pmd))
2688 break;
2689 if (iter.index == end_pgoff)
2690 break;
2691 }
2692 rcu_read_unlock();
2693}
2694EXPORT_SYMBOL(filemap_map_pages);
2695
2696int filemap_page_mkwrite(struct vm_fault *vmf)
2697{
2698 struct page *page = vmf->page;
2699 struct inode *inode = file_inode(vmf->vma->vm_file);
2700 int ret = VM_FAULT_LOCKED;
2701
2702 sb_start_pagefault(inode->i_sb);
2703 file_update_time(vmf->vma->vm_file);
2704 lock_page(page);
2705 if (page->mapping != inode->i_mapping) {
2706 unlock_page(page);
2707 ret = VM_FAULT_NOPAGE;
2708 goto out;
2709 }
2710 /*
2711 * We mark the page dirty already here so that when freeze is in
2712 * progress, we are guaranteed that writeback during freezing will
2713 * see the dirty page and writeprotect it again.
2714 */
2715 set_page_dirty(page);
2716 wait_for_stable_page(page);
2717out:
2718 sb_end_pagefault(inode->i_sb);
2719 return ret;
2720}
2721
2722const struct vm_operations_struct generic_file_vm_ops = {
2723 .fault = filemap_fault,
2724 .map_pages = filemap_map_pages,
2725 .page_mkwrite = filemap_page_mkwrite,
2726};
2727
2728/* This is used for a general mmap of a disk file */
2729
2730int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2731{
2732 struct address_space *mapping = file->f_mapping;
2733
2734 if (!mapping->a_ops->readpage)
2735 return -ENOEXEC;
2736 file_accessed(file);
2737 vma->vm_ops = &generic_file_vm_ops;
2738 return 0;
2739}
2740
2741/*
2742 * This is for filesystems which do not implement ->writepage.
2743 */
2744int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2745{
2746 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2747 return -EINVAL;
2748 return generic_file_mmap(file, vma);
2749}
2750#else
2751int filemap_page_mkwrite(struct vm_fault *vmf)
2752{
2753 return -ENOSYS;
2754}
2755int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2756{
2757 return -ENOSYS;
2758}
2759int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2760{
2761 return -ENOSYS;
2762}
2763#endif /* CONFIG_MMU */
2764
2765EXPORT_SYMBOL(filemap_page_mkwrite);
2766EXPORT_SYMBOL(generic_file_mmap);
2767EXPORT_SYMBOL(generic_file_readonly_mmap);
2768
2769static struct page *wait_on_page_read(struct page *page)
2770{
2771 if (!IS_ERR(page)) {
2772 wait_on_page_locked(page);
2773 if (!PageUptodate(page)) {
2774 put_page(page);
2775 page = ERR_PTR(-EIO);
2776 }
2777 }
2778 return page;
2779}
2780
2781static struct page *do_read_cache_page(struct address_space *mapping,
2782 pgoff_t index,
2783 int (*filler)(void *, struct page *),
2784 void *data,
2785 gfp_t gfp)
2786{
2787 struct page *page;
2788 int err;
2789repeat:
2790 page = find_get_page(mapping, index);
2791 if (!page) {
2792 page = __page_cache_alloc(gfp);
2793 if (!page)
2794 return ERR_PTR(-ENOMEM);
2795 err = add_to_page_cache_lru(page, mapping, index, gfp);
2796 if (unlikely(err)) {
2797 put_page(page);
2798 if (err == -EEXIST)
2799 goto repeat;
2800 /* Presumably ENOMEM for radix tree node */
2801 return ERR_PTR(err);
2802 }
2803
2804filler:
2805 err = filler(data, page);
2806 if (err < 0) {
2807 put_page(page);
2808 return ERR_PTR(err);
2809 }
2810
2811 page = wait_on_page_read(page);
2812 if (IS_ERR(page))
2813 return page;
2814 goto out;
2815 }
2816 if (PageUptodate(page))
2817 goto out;
2818
2819 /*
2820 * Page is not up to date and may be locked due one of the following
2821 * case a: Page is being filled and the page lock is held
2822 * case b: Read/write error clearing the page uptodate status
2823 * case c: Truncation in progress (page locked)
2824 * case d: Reclaim in progress
2825 *
2826 * Case a, the page will be up to date when the page is unlocked.
2827 * There is no need to serialise on the page lock here as the page
2828 * is pinned so the lock gives no additional protection. Even if the
2829 * the page is truncated, the data is still valid if PageUptodate as
2830 * it's a race vs truncate race.
2831 * Case b, the page will not be up to date
2832 * Case c, the page may be truncated but in itself, the data may still
2833 * be valid after IO completes as it's a read vs truncate race. The
2834 * operation must restart if the page is not uptodate on unlock but
2835 * otherwise serialising on page lock to stabilise the mapping gives
2836 * no additional guarantees to the caller as the page lock is
2837 * released before return.
2838 * Case d, similar to truncation. If reclaim holds the page lock, it
2839 * will be a race with remove_mapping that determines if the mapping
2840 * is valid on unlock but otherwise the data is valid and there is
2841 * no need to serialise with page lock.
2842 *
2843 * As the page lock gives no additional guarantee, we optimistically
2844 * wait on the page to be unlocked and check if it's up to date and
2845 * use the page if it is. Otherwise, the page lock is required to
2846 * distinguish between the different cases. The motivation is that we
2847 * avoid spurious serialisations and wakeups when multiple processes
2848 * wait on the same page for IO to complete.
2849 */
2850 wait_on_page_locked(page);
2851 if (PageUptodate(page))
2852 goto out;
2853
2854 /* Distinguish between all the cases under the safety of the lock */
2855 lock_page(page);
2856
2857 /* Case c or d, restart the operation */
2858 if (!page->mapping) {
2859 unlock_page(page);
2860 put_page(page);
2861 goto repeat;
2862 }
2863
2864 /* Someone else locked and filled the page in a very small window */
2865 if (PageUptodate(page)) {
2866 unlock_page(page);
2867 goto out;
2868 }
2869 goto filler;
2870
2871out:
2872 mark_page_accessed(page);
2873 return page;
2874}
2875
2876/**
2877 * read_cache_page - read into page cache, fill it if needed
2878 * @mapping: the page's address_space
2879 * @index: the page index
2880 * @filler: function to perform the read
2881 * @data: first arg to filler(data, page) function, often left as NULL
2882 *
2883 * Read into the page cache. If a page already exists, and PageUptodate() is
2884 * not set, try to fill the page and wait for it to become unlocked.
2885 *
2886 * If the page does not get brought uptodate, return -EIO.
2887 */
2888struct page *read_cache_page(struct address_space *mapping,
2889 pgoff_t index,
2890 int (*filler)(void *, struct page *),
2891 void *data)
2892{
2893 return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2894}
2895EXPORT_SYMBOL(read_cache_page);
2896
2897/**
2898 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2899 * @mapping: the page's address_space
2900 * @index: the page index
2901 * @gfp: the page allocator flags to use if allocating
2902 *
2903 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2904 * any new page allocations done using the specified allocation flags.
2905 *
2906 * If the page does not get brought uptodate, return -EIO.
2907 */
2908struct page *read_cache_page_gfp(struct address_space *mapping,
2909 pgoff_t index,
2910 gfp_t gfp)
2911{
2912 filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2913
2914 return do_read_cache_page(mapping, index, filler, NULL, gfp);
2915}
2916EXPORT_SYMBOL(read_cache_page_gfp);
2917
2918/*
2919 * Performs necessary checks before doing a write
2920 *
2921 * Can adjust writing position or amount of bytes to write.
2922 * Returns appropriate error code that caller should return or
2923 * zero in case that write should be allowed.
2924 */
2925inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2926{
2927 struct file *file = iocb->ki_filp;
2928 struct inode *inode = file->f_mapping->host;
2929 unsigned long limit = rlimit(RLIMIT_FSIZE);
2930 loff_t pos;
2931
2932 if (!iov_iter_count(from))
2933 return 0;
2934
2935 /* FIXME: this is for backwards compatibility with 2.4 */
2936 if (iocb->ki_flags & IOCB_APPEND)
2937 iocb->ki_pos = i_size_read(inode);
2938
2939 pos = iocb->ki_pos;
2940
2941 if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2942 return -EINVAL;
2943
2944 if (limit != RLIM_INFINITY) {
2945 if (iocb->ki_pos >= limit) {
2946 send_sig(SIGXFSZ, current, 0);
2947 return -EFBIG;
2948 }
2949 iov_iter_truncate(from, limit - (unsigned long)pos);
2950 }
2951
2952 /*
2953 * LFS rule
2954 */
2955 if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2956 !(file->f_flags & O_LARGEFILE))) {
2957 if (pos >= MAX_NON_LFS)
2958 return -EFBIG;
2959 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2960 }
2961
2962 /*
2963 * Are we about to exceed the fs block limit ?
2964 *
2965 * If we have written data it becomes a short write. If we have
2966 * exceeded without writing data we send a signal and return EFBIG.
2967 * Linus frestrict idea will clean these up nicely..
2968 */
2969 if (unlikely(pos >= inode->i_sb->s_maxbytes))
2970 return -EFBIG;
2971
2972 iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2973 return iov_iter_count(from);
2974}
2975EXPORT_SYMBOL(generic_write_checks);
2976
2977int pagecache_write_begin(struct file *file, struct address_space *mapping,
2978 loff_t pos, unsigned len, unsigned flags,
2979 struct page **pagep, void **fsdata)
2980{
2981 const struct address_space_operations *aops = mapping->a_ops;
2982
2983 return aops->write_begin(file, mapping, pos, len, flags,
2984 pagep, fsdata);
2985}
2986EXPORT_SYMBOL(pagecache_write_begin);
2987
2988int pagecache_write_end(struct file *file, struct address_space *mapping,
2989 loff_t pos, unsigned len, unsigned copied,
2990 struct page *page, void *fsdata)
2991{
2992 const struct address_space_operations *aops = mapping->a_ops;
2993
2994 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2995}
2996EXPORT_SYMBOL(pagecache_write_end);
2997
2998ssize_t
2999generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3000{
3001 struct file *file = iocb->ki_filp;
3002 struct address_space *mapping = file->f_mapping;
3003 struct inode *inode = mapping->host;
3004 loff_t pos = iocb->ki_pos;
3005 ssize_t written;
3006 size_t write_len;
3007 pgoff_t end;
3008
3009 write_len = iov_iter_count(from);
3010 end = (pos + write_len - 1) >> PAGE_SHIFT;
3011
3012 if (iocb->ki_flags & IOCB_NOWAIT) {
3013 /* If there are pages to writeback, return */
3014 if (filemap_range_has_page(inode->i_mapping, pos,
3015 pos + iov_iter_count(from)))
3016 return -EAGAIN;
3017 } else {
3018 written = filemap_write_and_wait_range(mapping, pos,
3019 pos + write_len - 1);
3020 if (written)
3021 goto out;
3022 }
3023
3024 /*
3025 * After a write we want buffered reads to be sure to go to disk to get
3026 * the new data. We invalidate clean cached page from the region we're
3027 * about to write. We do this *before* the write so that we can return
3028 * without clobbering -EIOCBQUEUED from ->direct_IO().
3029 */
3030 written = invalidate_inode_pages2_range(mapping,
3031 pos >> PAGE_SHIFT, end);
3032 /*
3033 * If a page can not be invalidated, return 0 to fall back
3034 * to buffered write.
3035 */
3036 if (written) {
3037 if (written == -EBUSY)
3038 return 0;
3039 goto out;
3040 }
3041
3042 written = mapping->a_ops->direct_IO(iocb, from);
3043
3044 /*
3045 * Finally, try again to invalidate clean pages which might have been
3046 * cached by non-direct readahead, or faulted in by get_user_pages()
3047 * if the source of the write was an mmap'ed region of the file
3048 * we're writing. Either one is a pretty crazy thing to do,
3049 * so we don't support it 100%. If this invalidation
3050 * fails, tough, the write still worked...
3051 *
3052 * Most of the time we do not need this since dio_complete() will do
3053 * the invalidation for us. However there are some file systems that
3054 * do not end up with dio_complete() being called, so let's not break
3055 * them by removing it completely
3056 */
3057 if (mapping->nrpages)
3058 invalidate_inode_pages2_range(mapping,
3059 pos >> PAGE_SHIFT, end);
3060
3061 if (written > 0) {
3062 pos += written;
3063 write_len -= written;
3064 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3065 i_size_write(inode, pos);
3066 mark_inode_dirty(inode);
3067 }
3068 iocb->ki_pos = pos;
3069 }
3070 iov_iter_revert(from, write_len - iov_iter_count(from));
3071out:
3072 return written;
3073}
3074EXPORT_SYMBOL(generic_file_direct_write);
3075
3076/*
3077 * Find or create a page at the given pagecache position. Return the locked
3078 * page. This function is specifically for buffered writes.
3079 */
3080struct page *grab_cache_page_write_begin(struct address_space *mapping,
3081 pgoff_t index, unsigned flags)
3082{
3083 struct page *page;
3084 int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3085
3086 if (flags & AOP_FLAG_NOFS)
3087 fgp_flags |= FGP_NOFS;
3088
3089 page = pagecache_get_page(mapping, index, fgp_flags,
3090 mapping_gfp_mask(mapping));
3091 if (page)
3092 wait_for_stable_page(page);
3093
3094 return page;
3095}
3096EXPORT_SYMBOL(grab_cache_page_write_begin);
3097
3098ssize_t generic_perform_write(struct file *file,
3099 struct iov_iter *i, loff_t pos)
3100{
3101 struct address_space *mapping = file->f_mapping;
3102 const struct address_space_operations *a_ops = mapping->a_ops;
3103 long status = 0;
3104 ssize_t written = 0;
3105 unsigned int flags = 0;
3106
3107 do {
3108 struct page *page;
3109 unsigned long offset; /* Offset into pagecache page */
3110 unsigned long bytes; /* Bytes to write to page */
3111 size_t copied; /* Bytes copied from user */
3112 void *fsdata;
3113
3114 offset = (pos & (PAGE_SIZE - 1));
3115 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3116 iov_iter_count(i));
3117
3118again:
3119 /*
3120 * Bring in the user page that we will copy from _first_.
3121 * Otherwise there's a nasty deadlock on copying from the
3122 * same page as we're writing to, without it being marked
3123 * up-to-date.
3124 *
3125 * Not only is this an optimisation, but it is also required
3126 * to check that the address is actually valid, when atomic
3127 * usercopies are used, below.
3128 */
3129 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3130 status = -EFAULT;
3131 break;
3132 }
3133
3134 if (fatal_signal_pending(current)) {
3135 status = -EINTR;
3136 break;
3137 }
3138
3139 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3140 &page, &fsdata);
3141 if (unlikely(status < 0))
3142 break;
3143
3144 if (mapping_writably_mapped(mapping))
3145 flush_dcache_page(page);
3146
3147 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3148 flush_dcache_page(page);
3149
3150 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3151 page, fsdata);
3152 if (unlikely(status < 0))
3153 break;
3154 copied = status;
3155
3156 cond_resched();
3157
3158 iov_iter_advance(i, copied);
3159 if (unlikely(copied == 0)) {
3160 /*
3161 * If we were unable to copy any data at all, we must
3162 * fall back to a single segment length write.
3163 *
3164 * If we didn't fallback here, we could livelock
3165 * because not all segments in the iov can be copied at
3166 * once without a pagefault.
3167 */
3168 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3169 iov_iter_single_seg_count(i));
3170 goto again;
3171 }
3172 pos += copied;
3173 written += copied;
3174
3175 balance_dirty_pages_ratelimited(mapping);
3176 } while (iov_iter_count(i));
3177
3178 return written ? written : status;
3179}
3180EXPORT_SYMBOL(generic_perform_write);
3181
3182/**
3183 * __generic_file_write_iter - write data to a file
3184 * @iocb: IO state structure (file, offset, etc.)
3185 * @from: iov_iter with data to write
3186 *
3187 * This function does all the work needed for actually writing data to a
3188 * file. It does all basic checks, removes SUID from the file, updates
3189 * modification times and calls proper subroutines depending on whether we
3190 * do direct IO or a standard buffered write.
3191 *
3192 * It expects i_mutex to be grabbed unless we work on a block device or similar
3193 * object which does not need locking at all.
3194 *
3195 * This function does *not* take care of syncing data in case of O_SYNC write.
3196 * A caller has to handle it. This is mainly due to the fact that we want to
3197 * avoid syncing under i_mutex.
3198 */
3199ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3200{
3201 struct file *file = iocb->ki_filp;
3202 struct address_space * mapping = file->f_mapping;
3203 struct inode *inode = mapping->host;
3204 ssize_t written = 0;
3205 ssize_t err;
3206 ssize_t status;
3207
3208 /* We can write back this queue in page reclaim */
3209 current->backing_dev_info = inode_to_bdi(inode);
3210 err = file_remove_privs(file);
3211 if (err)
3212 goto out;
3213
3214 err = file_update_time(file);
3215 if (err)
3216 goto out;
3217
3218 if (iocb->ki_flags & IOCB_DIRECT) {
3219 loff_t pos, endbyte;
3220
3221 written = generic_file_direct_write(iocb, from);
3222 /*
3223 * If the write stopped short of completing, fall back to
3224 * buffered writes. Some filesystems do this for writes to
3225 * holes, for example. For DAX files, a buffered write will
3226 * not succeed (even if it did, DAX does not handle dirty
3227 * page-cache pages correctly).
3228 */
3229 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3230 goto out;
3231
3232 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3233 /*
3234 * If generic_perform_write() returned a synchronous error
3235 * then we want to return the number of bytes which were
3236 * direct-written, or the error code if that was zero. Note
3237 * that this differs from normal direct-io semantics, which
3238 * will return -EFOO even if some bytes were written.
3239 */
3240 if (unlikely(status < 0)) {
3241 err = status;
3242 goto out;
3243 }
3244 /*
3245 * We need to ensure that the page cache pages are written to
3246 * disk and invalidated to preserve the expected O_DIRECT
3247 * semantics.
3248 */
3249 endbyte = pos + status - 1;
3250 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3251 if (err == 0) {
3252 iocb->ki_pos = endbyte + 1;
3253 written += status;
3254 invalidate_mapping_pages(mapping,
3255 pos >> PAGE_SHIFT,
3256 endbyte >> PAGE_SHIFT);
3257 } else {
3258 /*
3259 * We don't know how much we wrote, so just return
3260 * the number of bytes which were direct-written
3261 */
3262 }
3263 } else {
3264 written = generic_perform_write(file, from, iocb->ki_pos);
3265 if (likely(written > 0))
3266 iocb->ki_pos += written;
3267 }
3268out:
3269 current->backing_dev_info = NULL;
3270 return written ? written : err;
3271}
3272EXPORT_SYMBOL(__generic_file_write_iter);
3273
3274/**
3275 * generic_file_write_iter - write data to a file
3276 * @iocb: IO state structure
3277 * @from: iov_iter with data to write
3278 *
3279 * This is a wrapper around __generic_file_write_iter() to be used by most
3280 * filesystems. It takes care of syncing the file in case of O_SYNC file
3281 * and acquires i_mutex as needed.
3282 */
3283ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3284{
3285 struct file *file = iocb->ki_filp;
3286 struct inode *inode = file->f_mapping->host;
3287 ssize_t ret;
3288
3289 inode_lock(inode);
3290 ret = generic_write_checks(iocb, from);
3291 if (ret > 0)
3292 ret = __generic_file_write_iter(iocb, from);
3293 inode_unlock(inode);
3294
3295 if (ret > 0)
3296 ret = generic_write_sync(iocb, ret);
3297 return ret;
3298}
3299EXPORT_SYMBOL(generic_file_write_iter);
3300
3301/**
3302 * try_to_release_page() - release old fs-specific metadata on a page
3303 *
3304 * @page: the page which the kernel is trying to free
3305 * @gfp_mask: memory allocation flags (and I/O mode)
3306 *
3307 * The address_space is to try to release any data against the page
3308 * (presumably at page->private). If the release was successful, return '1'.
3309 * Otherwise return zero.
3310 *
3311 * This may also be called if PG_fscache is set on a page, indicating that the
3312 * page is known to the local caching routines.
3313 *
3314 * The @gfp_mask argument specifies whether I/O may be performed to release
3315 * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3316 *
3317 */
3318int try_to_release_page(struct page *page, gfp_t gfp_mask)
3319{
3320 struct address_space * const mapping = page->mapping;
3321
3322 BUG_ON(!PageLocked(page));
3323 if (PageWriteback(page))
3324 return 0;
3325
3326 if (mapping && mapping->a_ops->releasepage)
3327 return mapping->a_ops->releasepage(page, gfp_mask);
3328 return try_to_free_buffers(page);
3329}
3330
3331EXPORT_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 "internal.h"
46
47#define CREATE_TRACE_POINTS
48#include <trace/events/filemap.h>
49
50/*
51 * FIXME: remove all knowledge of the buffer layer from the core VM
52 */
53#include <linux/buffer_head.h> /* for try_to_free_buffers */
54
55#include <asm/mman.h>
56
57/*
58 * Shared mappings implemented 30.11.1994. It's not fully working yet,
59 * though.
60 *
61 * Shared mappings now work. 15.8.1995 Bruno.
62 *
63 * finished 'unifying' the page and buffer cache and SMP-threaded the
64 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
65 *
66 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
67 */
68
69/*
70 * Lock ordering:
71 *
72 * ->i_mmap_rwsem (truncate_pagecache)
73 * ->private_lock (__free_pte->__set_page_dirty_buffers)
74 * ->swap_lock (exclusive_swap_page, others)
75 * ->i_pages lock
76 *
77 * ->i_mutex
78 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
79 *
80 * ->mmap_lock
81 * ->i_mmap_rwsem
82 * ->page_table_lock or pte_lock (various, mainly in memory.c)
83 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
84 *
85 * ->mmap_lock
86 * ->lock_page (access_process_vm)
87 *
88 * ->i_mutex (generic_perform_write)
89 * ->mmap_lock (fault_in_pages_readable->do_page_fault)
90 *
91 * bdi->wb.list_lock
92 * sb_lock (fs/fs-writeback.c)
93 * ->i_pages lock (__sync_single_inode)
94 *
95 * ->i_mmap_rwsem
96 * ->anon_vma.lock (vma_adjust)
97 *
98 * ->anon_vma.lock
99 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
100 *
101 * ->page_table_lock or pte_lock
102 * ->swap_lock (try_to_unmap_one)
103 * ->private_lock (try_to_unmap_one)
104 * ->i_pages lock (try_to_unmap_one)
105 * ->pgdat->lru_lock (follow_page->mark_page_accessed)
106 * ->pgdat->lru_lock (check_pte_range->isolate_lru_page)
107 * ->private_lock (page_remove_rmap->set_page_dirty)
108 * ->i_pages lock (page_remove_rmap->set_page_dirty)
109 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
110 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
111 * ->memcg->move_lock (page_remove_rmap->lock_page_memcg)
112 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
113 * ->inode->i_lock (zap_pte_range->set_page_dirty)
114 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
115 *
116 * ->i_mmap_rwsem
117 * ->tasklist_lock (memory_failure, collect_procs_ao)
118 */
119
120static void page_cache_delete(struct address_space *mapping,
121 struct page *page, void *shadow)
122{
123 XA_STATE(xas, &mapping->i_pages, page->index);
124 unsigned int nr = 1;
125
126 mapping_set_update(&xas, mapping);
127
128 /* hugetlb pages are represented by a single entry in the xarray */
129 if (!PageHuge(page)) {
130 xas_set_order(&xas, page->index, compound_order(page));
131 nr = compound_nr(page);
132 }
133
134 VM_BUG_ON_PAGE(!PageLocked(page), page);
135 VM_BUG_ON_PAGE(PageTail(page), page);
136 VM_BUG_ON_PAGE(nr != 1 && shadow, page);
137
138 xas_store(&xas, shadow);
139 xas_init_marks(&xas);
140
141 page->mapping = NULL;
142 /* Leave page->index set: truncation lookup relies upon it */
143
144 if (shadow) {
145 mapping->nrexceptional += nr;
146 /*
147 * Make sure the nrexceptional update is committed before
148 * the nrpages update so that final truncate racing
149 * with reclaim does not see both counters 0 at the
150 * same time and miss a shadow entry.
151 */
152 smp_wmb();
153 }
154 mapping->nrpages -= nr;
155}
156
157static void unaccount_page_cache_page(struct address_space *mapping,
158 struct page *page)
159{
160 int nr;
161
162 /*
163 * if we're uptodate, flush out into the cleancache, otherwise
164 * invalidate any existing cleancache entries. We can't leave
165 * stale data around in the cleancache once our page is gone
166 */
167 if (PageUptodate(page) && PageMappedToDisk(page))
168 cleancache_put_page(page);
169 else
170 cleancache_invalidate_page(mapping, page);
171
172 VM_BUG_ON_PAGE(PageTail(page), page);
173 VM_BUG_ON_PAGE(page_mapped(page), page);
174 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
175 int mapcount;
176
177 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
178 current->comm, page_to_pfn(page));
179 dump_page(page, "still mapped when deleted");
180 dump_stack();
181 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
182
183 mapcount = page_mapcount(page);
184 if (mapping_exiting(mapping) &&
185 page_count(page) >= mapcount + 2) {
186 /*
187 * All vmas have already been torn down, so it's
188 * a good bet that actually the page is unmapped,
189 * and we'd prefer not to leak it: if we're wrong,
190 * some other bad page check should catch it later.
191 */
192 page_mapcount_reset(page);
193 page_ref_sub(page, mapcount);
194 }
195 }
196
197 /* hugetlb pages do not participate in page cache accounting. */
198 if (PageHuge(page))
199 return;
200
201 nr = thp_nr_pages(page);
202
203 __mod_lruvec_page_state(page, NR_FILE_PAGES, -nr);
204 if (PageSwapBacked(page)) {
205 __mod_lruvec_page_state(page, NR_SHMEM, -nr);
206 if (PageTransHuge(page))
207 __dec_node_page_state(page, NR_SHMEM_THPS);
208 } else if (PageTransHuge(page)) {
209 __dec_node_page_state(page, NR_FILE_THPS);
210 filemap_nr_thps_dec(mapping);
211 }
212
213 /*
214 * At this point page must be either written or cleaned by
215 * truncate. Dirty page here signals a bug and loss of
216 * unwritten data.
217 *
218 * This fixes dirty accounting after removing the page entirely
219 * but leaves PageDirty set: it has no effect for truncated
220 * page and anyway will be cleared before returning page into
221 * buddy allocator.
222 */
223 if (WARN_ON_ONCE(PageDirty(page)))
224 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
225}
226
227/*
228 * Delete a page from the page cache and free it. Caller has to make
229 * sure the page is locked and that nobody else uses it - or that usage
230 * is safe. The caller must hold the i_pages lock.
231 */
232void __delete_from_page_cache(struct page *page, void *shadow)
233{
234 struct address_space *mapping = page->mapping;
235
236 trace_mm_filemap_delete_from_page_cache(page);
237
238 unaccount_page_cache_page(mapping, page);
239 page_cache_delete(mapping, page, shadow);
240}
241
242static void page_cache_free_page(struct address_space *mapping,
243 struct page *page)
244{
245 void (*freepage)(struct page *);
246
247 freepage = mapping->a_ops->freepage;
248 if (freepage)
249 freepage(page);
250
251 if (PageTransHuge(page) && !PageHuge(page)) {
252 page_ref_sub(page, HPAGE_PMD_NR);
253 VM_BUG_ON_PAGE(page_count(page) <= 0, page);
254 } else {
255 put_page(page);
256 }
257}
258
259/**
260 * delete_from_page_cache - delete page from page cache
261 * @page: the page which the kernel is trying to remove from page cache
262 *
263 * This must be called only on pages that have been verified to be in the page
264 * cache and locked. It will never put the page into the free list, the caller
265 * has a reference on the page.
266 */
267void delete_from_page_cache(struct page *page)
268{
269 struct address_space *mapping = page_mapping(page);
270 unsigned long flags;
271
272 BUG_ON(!PageLocked(page));
273 xa_lock_irqsave(&mapping->i_pages, flags);
274 __delete_from_page_cache(page, NULL);
275 xa_unlock_irqrestore(&mapping->i_pages, flags);
276
277 page_cache_free_page(mapping, page);
278}
279EXPORT_SYMBOL(delete_from_page_cache);
280
281/*
282 * page_cache_delete_batch - delete several pages from page cache
283 * @mapping: the mapping to which pages belong
284 * @pvec: pagevec with pages to delete
285 *
286 * The function walks over mapping->i_pages and removes pages passed in @pvec
287 * from the mapping. The function expects @pvec to be sorted by page index
288 * and is optimised for it to be dense.
289 * It tolerates holes in @pvec (mapping entries at those indices are not
290 * modified). The function expects only THP head pages to be present in the
291 * @pvec.
292 *
293 * The function expects the i_pages lock to be held.
294 */
295static void page_cache_delete_batch(struct address_space *mapping,
296 struct pagevec *pvec)
297{
298 XA_STATE(xas, &mapping->i_pages, pvec->pages[0]->index);
299 int total_pages = 0;
300 int i = 0;
301 struct page *page;
302
303 mapping_set_update(&xas, mapping);
304 xas_for_each(&xas, page, ULONG_MAX) {
305 if (i >= pagevec_count(pvec))
306 break;
307
308 /* A swap/dax/shadow entry got inserted? Skip it. */
309 if (xa_is_value(page))
310 continue;
311 /*
312 * A page got inserted in our range? Skip it. We have our
313 * pages locked so they are protected from being removed.
314 * If we see a page whose index is higher than ours, it
315 * means our page has been removed, which shouldn't be
316 * possible because we're holding the PageLock.
317 */
318 if (page != pvec->pages[i]) {
319 VM_BUG_ON_PAGE(page->index > pvec->pages[i]->index,
320 page);
321 continue;
322 }
323
324 WARN_ON_ONCE(!PageLocked(page));
325
326 if (page->index == xas.xa_index)
327 page->mapping = NULL;
328 /* Leave page->index set: truncation lookup relies on it */
329
330 /*
331 * Move to the next page in the vector if this is a regular
332 * page or the index is of the last sub-page of this compound
333 * page.
334 */
335 if (page->index + compound_nr(page) - 1 == xas.xa_index)
336 i++;
337 xas_store(&xas, NULL);
338 total_pages++;
339 }
340 mapping->nrpages -= total_pages;
341}
342
343void delete_from_page_cache_batch(struct address_space *mapping,
344 struct pagevec *pvec)
345{
346 int i;
347 unsigned long flags;
348
349 if (!pagevec_count(pvec))
350 return;
351
352 xa_lock_irqsave(&mapping->i_pages, flags);
353 for (i = 0; i < pagevec_count(pvec); i++) {
354 trace_mm_filemap_delete_from_page_cache(pvec->pages[i]);
355
356 unaccount_page_cache_page(mapping, pvec->pages[i]);
357 }
358 page_cache_delete_batch(mapping, pvec);
359 xa_unlock_irqrestore(&mapping->i_pages, flags);
360
361 for (i = 0; i < pagevec_count(pvec); i++)
362 page_cache_free_page(mapping, pvec->pages[i]);
363}
364
365int filemap_check_errors(struct address_space *mapping)
366{
367 int ret = 0;
368 /* Check for outstanding write errors */
369 if (test_bit(AS_ENOSPC, &mapping->flags) &&
370 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
371 ret = -ENOSPC;
372 if (test_bit(AS_EIO, &mapping->flags) &&
373 test_and_clear_bit(AS_EIO, &mapping->flags))
374 ret = -EIO;
375 return ret;
376}
377EXPORT_SYMBOL(filemap_check_errors);
378
379static int filemap_check_and_keep_errors(struct address_space *mapping)
380{
381 /* Check for outstanding write errors */
382 if (test_bit(AS_EIO, &mapping->flags))
383 return -EIO;
384 if (test_bit(AS_ENOSPC, &mapping->flags))
385 return -ENOSPC;
386 return 0;
387}
388
389/**
390 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
391 * @mapping: address space structure to write
392 * @start: offset in bytes where the range starts
393 * @end: offset in bytes where the range ends (inclusive)
394 * @sync_mode: enable synchronous operation
395 *
396 * Start writeback against all of a mapping's dirty pages that lie
397 * within the byte offsets <start, end> inclusive.
398 *
399 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
400 * opposed to a regular memory cleansing writeback. The difference between
401 * these two operations is that if a dirty page/buffer is encountered, it must
402 * be waited upon, and not just skipped over.
403 *
404 * Return: %0 on success, negative error code otherwise.
405 */
406int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
407 loff_t end, int sync_mode)
408{
409 int ret;
410 struct writeback_control wbc = {
411 .sync_mode = sync_mode,
412 .nr_to_write = LONG_MAX,
413 .range_start = start,
414 .range_end = end,
415 };
416
417 if (!mapping_cap_writeback_dirty(mapping) ||
418 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
419 return 0;
420
421 wbc_attach_fdatawrite_inode(&wbc, mapping->host);
422 ret = do_writepages(mapping, &wbc);
423 wbc_detach_inode(&wbc);
424 return ret;
425}
426
427static inline int __filemap_fdatawrite(struct address_space *mapping,
428 int sync_mode)
429{
430 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
431}
432
433int filemap_fdatawrite(struct address_space *mapping)
434{
435 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
436}
437EXPORT_SYMBOL(filemap_fdatawrite);
438
439int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
440 loff_t end)
441{
442 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
443}
444EXPORT_SYMBOL(filemap_fdatawrite_range);
445
446/**
447 * filemap_flush - mostly a non-blocking flush
448 * @mapping: target address_space
449 *
450 * This is a mostly non-blocking flush. Not suitable for data-integrity
451 * purposes - I/O may not be started against all dirty pages.
452 *
453 * Return: %0 on success, negative error code otherwise.
454 */
455int filemap_flush(struct address_space *mapping)
456{
457 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
458}
459EXPORT_SYMBOL(filemap_flush);
460
461/**
462 * filemap_range_has_page - check if a page exists in range.
463 * @mapping: address space within which to check
464 * @start_byte: offset in bytes where the range starts
465 * @end_byte: offset in bytes where the range ends (inclusive)
466 *
467 * Find at least one page in the range supplied, usually used to check if
468 * direct writing in this range will trigger a writeback.
469 *
470 * Return: %true if at least one page exists in the specified range,
471 * %false otherwise.
472 */
473bool filemap_range_has_page(struct address_space *mapping,
474 loff_t start_byte, loff_t end_byte)
475{
476 struct page *page;
477 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
478 pgoff_t max = end_byte >> PAGE_SHIFT;
479
480 if (end_byte < start_byte)
481 return false;
482
483 rcu_read_lock();
484 for (;;) {
485 page = xas_find(&xas, max);
486 if (xas_retry(&xas, page))
487 continue;
488 /* Shadow entries don't count */
489 if (xa_is_value(page))
490 continue;
491 /*
492 * We don't need to try to pin this page; we're about to
493 * release the RCU lock anyway. It is enough to know that
494 * there was a page here recently.
495 */
496 break;
497 }
498 rcu_read_unlock();
499
500 return page != NULL;
501}
502EXPORT_SYMBOL(filemap_range_has_page);
503
504static void __filemap_fdatawait_range(struct address_space *mapping,
505 loff_t start_byte, loff_t end_byte)
506{
507 pgoff_t index = start_byte >> PAGE_SHIFT;
508 pgoff_t end = end_byte >> PAGE_SHIFT;
509 struct pagevec pvec;
510 int nr_pages;
511
512 if (end_byte < start_byte)
513 return;
514
515 pagevec_init(&pvec);
516 while (index <= end) {
517 unsigned i;
518
519 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index,
520 end, PAGECACHE_TAG_WRITEBACK);
521 if (!nr_pages)
522 break;
523
524 for (i = 0; i < nr_pages; i++) {
525 struct page *page = pvec.pages[i];
526
527 wait_on_page_writeback(page);
528 ClearPageError(page);
529 }
530 pagevec_release(&pvec);
531 cond_resched();
532 }
533}
534
535/**
536 * filemap_fdatawait_range - wait for writeback to complete
537 * @mapping: address space structure to wait for
538 * @start_byte: offset in bytes where the range starts
539 * @end_byte: offset in bytes where the range ends (inclusive)
540 *
541 * Walk the list of under-writeback pages of the given address space
542 * in the given range and wait for all of them. Check error status of
543 * the address space and return it.
544 *
545 * Since the error status of the address space is cleared by this function,
546 * callers are responsible for checking the return value and handling and/or
547 * reporting the error.
548 *
549 * Return: error status of the address space.
550 */
551int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
552 loff_t end_byte)
553{
554 __filemap_fdatawait_range(mapping, start_byte, end_byte);
555 return filemap_check_errors(mapping);
556}
557EXPORT_SYMBOL(filemap_fdatawait_range);
558
559/**
560 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
561 * @mapping: address space structure to wait for
562 * @start_byte: offset in bytes where the range starts
563 * @end_byte: offset in bytes where the range ends (inclusive)
564 *
565 * Walk the list of under-writeback pages of the given address space in the
566 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
567 * this function does not clear error status of the address space.
568 *
569 * Use this function if callers don't handle errors themselves. Expected
570 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
571 * fsfreeze(8)
572 */
573int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
574 loff_t start_byte, loff_t end_byte)
575{
576 __filemap_fdatawait_range(mapping, start_byte, end_byte);
577 return filemap_check_and_keep_errors(mapping);
578}
579EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
580
581/**
582 * file_fdatawait_range - wait for writeback to complete
583 * @file: file pointing to address space structure to wait for
584 * @start_byte: offset in bytes where the range starts
585 * @end_byte: offset in bytes where the range ends (inclusive)
586 *
587 * Walk the list of under-writeback pages of the address space that file
588 * refers to, in the given range and wait for all of them. Check error
589 * status of the address space vs. the file->f_wb_err cursor and return it.
590 *
591 * Since the error status of the file is advanced by this function,
592 * callers are responsible for checking the return value and handling and/or
593 * reporting the error.
594 *
595 * Return: error status of the address space vs. the file->f_wb_err cursor.
596 */
597int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
598{
599 struct address_space *mapping = file->f_mapping;
600
601 __filemap_fdatawait_range(mapping, start_byte, end_byte);
602 return file_check_and_advance_wb_err(file);
603}
604EXPORT_SYMBOL(file_fdatawait_range);
605
606/**
607 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
608 * @mapping: address space structure to wait for
609 *
610 * Walk the list of under-writeback pages of the given address space
611 * and wait for all of them. Unlike filemap_fdatawait(), this function
612 * does not clear error status of the address space.
613 *
614 * Use this function if callers don't handle errors themselves. Expected
615 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
616 * fsfreeze(8)
617 *
618 * Return: error status of the address space.
619 */
620int filemap_fdatawait_keep_errors(struct address_space *mapping)
621{
622 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
623 return filemap_check_and_keep_errors(mapping);
624}
625EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
626
627/* Returns true if writeback might be needed or already in progress. */
628static bool mapping_needs_writeback(struct address_space *mapping)
629{
630 if (dax_mapping(mapping))
631 return mapping->nrexceptional;
632
633 return mapping->nrpages;
634}
635
636/**
637 * filemap_write_and_wait_range - write out & wait on a file range
638 * @mapping: the address_space for the pages
639 * @lstart: offset in bytes where the range starts
640 * @lend: offset in bytes where the range ends (inclusive)
641 *
642 * Write out and wait upon file offsets lstart->lend, inclusive.
643 *
644 * Note that @lend is inclusive (describes the last byte to be written) so
645 * that this function can be used to write to the very end-of-file (end = -1).
646 *
647 * Return: error status of the address space.
648 */
649int filemap_write_and_wait_range(struct address_space *mapping,
650 loff_t lstart, loff_t lend)
651{
652 int err = 0;
653
654 if (mapping_needs_writeback(mapping)) {
655 err = __filemap_fdatawrite_range(mapping, lstart, lend,
656 WB_SYNC_ALL);
657 /*
658 * Even if the above returned error, the pages may be
659 * written partially (e.g. -ENOSPC), so we wait for it.
660 * But the -EIO is special case, it may indicate the worst
661 * thing (e.g. bug) happened, so we avoid waiting for it.
662 */
663 if (err != -EIO) {
664 int err2 = filemap_fdatawait_range(mapping,
665 lstart, lend);
666 if (!err)
667 err = err2;
668 } else {
669 /* Clear any previously stored errors */
670 filemap_check_errors(mapping);
671 }
672 } else {
673 err = filemap_check_errors(mapping);
674 }
675 return err;
676}
677EXPORT_SYMBOL(filemap_write_and_wait_range);
678
679void __filemap_set_wb_err(struct address_space *mapping, int err)
680{
681 errseq_t eseq = errseq_set(&mapping->wb_err, err);
682
683 trace_filemap_set_wb_err(mapping, eseq);
684}
685EXPORT_SYMBOL(__filemap_set_wb_err);
686
687/**
688 * file_check_and_advance_wb_err - report wb error (if any) that was previously
689 * and advance wb_err to current one
690 * @file: struct file on which the error is being reported
691 *
692 * When userland calls fsync (or something like nfsd does the equivalent), we
693 * want to report any writeback errors that occurred since the last fsync (or
694 * since the file was opened if there haven't been any).
695 *
696 * Grab the wb_err from the mapping. If it matches what we have in the file,
697 * then just quickly return 0. The file is all caught up.
698 *
699 * If it doesn't match, then take the mapping value, set the "seen" flag in
700 * it and try to swap it into place. If it works, or another task beat us
701 * to it with the new value, then update the f_wb_err and return the error
702 * portion. The error at this point must be reported via proper channels
703 * (a'la fsync, or NFS COMMIT operation, etc.).
704 *
705 * While we handle mapping->wb_err with atomic operations, the f_wb_err
706 * value is protected by the f_lock since we must ensure that it reflects
707 * the latest value swapped in for this file descriptor.
708 *
709 * Return: %0 on success, negative error code otherwise.
710 */
711int file_check_and_advance_wb_err(struct file *file)
712{
713 int err = 0;
714 errseq_t old = READ_ONCE(file->f_wb_err);
715 struct address_space *mapping = file->f_mapping;
716
717 /* Locklessly handle the common case where nothing has changed */
718 if (errseq_check(&mapping->wb_err, old)) {
719 /* Something changed, must use slow path */
720 spin_lock(&file->f_lock);
721 old = file->f_wb_err;
722 err = errseq_check_and_advance(&mapping->wb_err,
723 &file->f_wb_err);
724 trace_file_check_and_advance_wb_err(file, old);
725 spin_unlock(&file->f_lock);
726 }
727
728 /*
729 * We're mostly using this function as a drop in replacement for
730 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
731 * that the legacy code would have had on these flags.
732 */
733 clear_bit(AS_EIO, &mapping->flags);
734 clear_bit(AS_ENOSPC, &mapping->flags);
735 return err;
736}
737EXPORT_SYMBOL(file_check_and_advance_wb_err);
738
739/**
740 * file_write_and_wait_range - write out & wait on a file range
741 * @file: file pointing to address_space with pages
742 * @lstart: offset in bytes where the range starts
743 * @lend: offset in bytes where the range ends (inclusive)
744 *
745 * Write out and wait upon file offsets lstart->lend, inclusive.
746 *
747 * Note that @lend is inclusive (describes the last byte to be written) so
748 * that this function can be used to write to the very end-of-file (end = -1).
749 *
750 * After writing out and waiting on the data, we check and advance the
751 * f_wb_err cursor to the latest value, and return any errors detected there.
752 *
753 * Return: %0 on success, negative error code otherwise.
754 */
755int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
756{
757 int err = 0, err2;
758 struct address_space *mapping = file->f_mapping;
759
760 if (mapping_needs_writeback(mapping)) {
761 err = __filemap_fdatawrite_range(mapping, lstart, lend,
762 WB_SYNC_ALL);
763 /* See comment of filemap_write_and_wait() */
764 if (err != -EIO)
765 __filemap_fdatawait_range(mapping, lstart, lend);
766 }
767 err2 = file_check_and_advance_wb_err(file);
768 if (!err)
769 err = err2;
770 return err;
771}
772EXPORT_SYMBOL(file_write_and_wait_range);
773
774/**
775 * replace_page_cache_page - replace a pagecache page with a new one
776 * @old: page to be replaced
777 * @new: page to replace with
778 * @gfp_mask: allocation mode
779 *
780 * This function replaces a page in the pagecache with a new one. On
781 * success it acquires the pagecache reference for the new page and
782 * drops it for the old page. Both the old and new pages must be
783 * locked. This function does not add the new page to the LRU, the
784 * caller must do that.
785 *
786 * The remove + add is atomic. This function cannot fail.
787 *
788 * Return: %0
789 */
790int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
791{
792 struct address_space *mapping = old->mapping;
793 void (*freepage)(struct page *) = mapping->a_ops->freepage;
794 pgoff_t offset = old->index;
795 XA_STATE(xas, &mapping->i_pages, offset);
796 unsigned long flags;
797
798 VM_BUG_ON_PAGE(!PageLocked(old), old);
799 VM_BUG_ON_PAGE(!PageLocked(new), new);
800 VM_BUG_ON_PAGE(new->mapping, new);
801
802 get_page(new);
803 new->mapping = mapping;
804 new->index = offset;
805
806 mem_cgroup_migrate(old, new);
807
808 xas_lock_irqsave(&xas, flags);
809 xas_store(&xas, new);
810
811 old->mapping = NULL;
812 /* hugetlb pages do not participate in page cache accounting. */
813 if (!PageHuge(old))
814 __dec_lruvec_page_state(old, NR_FILE_PAGES);
815 if (!PageHuge(new))
816 __inc_lruvec_page_state(new, NR_FILE_PAGES);
817 if (PageSwapBacked(old))
818 __dec_lruvec_page_state(old, NR_SHMEM);
819 if (PageSwapBacked(new))
820 __inc_lruvec_page_state(new, NR_SHMEM);
821 xas_unlock_irqrestore(&xas, flags);
822 if (freepage)
823 freepage(old);
824 put_page(old);
825
826 return 0;
827}
828EXPORT_SYMBOL_GPL(replace_page_cache_page);
829
830static int __add_to_page_cache_locked(struct page *page,
831 struct address_space *mapping,
832 pgoff_t offset, gfp_t gfp_mask,
833 void **shadowp)
834{
835 XA_STATE(xas, &mapping->i_pages, offset);
836 int huge = PageHuge(page);
837 int error;
838 void *old;
839
840 VM_BUG_ON_PAGE(!PageLocked(page), page);
841 VM_BUG_ON_PAGE(PageSwapBacked(page), page);
842 mapping_set_update(&xas, mapping);
843
844 get_page(page);
845 page->mapping = mapping;
846 page->index = offset;
847
848 if (!huge) {
849 error = mem_cgroup_charge(page, current->mm, gfp_mask);
850 if (error)
851 goto error;
852 }
853
854 do {
855 xas_lock_irq(&xas);
856 old = xas_load(&xas);
857 if (old && !xa_is_value(old))
858 xas_set_err(&xas, -EEXIST);
859 xas_store(&xas, page);
860 if (xas_error(&xas))
861 goto unlock;
862
863 if (xa_is_value(old)) {
864 mapping->nrexceptional--;
865 if (shadowp)
866 *shadowp = old;
867 }
868 mapping->nrpages++;
869
870 /* hugetlb pages do not participate in page cache accounting */
871 if (!huge)
872 __inc_lruvec_page_state(page, NR_FILE_PAGES);
873unlock:
874 xas_unlock_irq(&xas);
875 } while (xas_nomem(&xas, gfp_mask & GFP_RECLAIM_MASK));
876
877 if (xas_error(&xas)) {
878 error = xas_error(&xas);
879 goto error;
880 }
881
882 trace_mm_filemap_add_to_page_cache(page);
883 return 0;
884error:
885 page->mapping = NULL;
886 /* Leave page->index set: truncation relies upon it */
887 put_page(page);
888 return error;
889}
890ALLOW_ERROR_INJECTION(__add_to_page_cache_locked, ERRNO);
891
892/**
893 * add_to_page_cache_locked - add a locked page to the pagecache
894 * @page: page to add
895 * @mapping: the page's address_space
896 * @offset: page index
897 * @gfp_mask: page allocation mode
898 *
899 * This function is used to add a page to the pagecache. It must be locked.
900 * This function does not add the page to the LRU. The caller must do that.
901 *
902 * Return: %0 on success, negative error code otherwise.
903 */
904int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
905 pgoff_t offset, gfp_t gfp_mask)
906{
907 return __add_to_page_cache_locked(page, mapping, offset,
908 gfp_mask, NULL);
909}
910EXPORT_SYMBOL(add_to_page_cache_locked);
911
912int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
913 pgoff_t offset, gfp_t gfp_mask)
914{
915 void *shadow = NULL;
916 int ret;
917
918 __SetPageLocked(page);
919 ret = __add_to_page_cache_locked(page, mapping, offset,
920 gfp_mask, &shadow);
921 if (unlikely(ret))
922 __ClearPageLocked(page);
923 else {
924 /*
925 * The page might have been evicted from cache only
926 * recently, in which case it should be activated like
927 * any other repeatedly accessed page.
928 * The exception is pages getting rewritten; evicting other
929 * data from the working set, only to cache data that will
930 * get overwritten with something else, is a waste of memory.
931 */
932 WARN_ON_ONCE(PageActive(page));
933 if (!(gfp_mask & __GFP_WRITE) && shadow)
934 workingset_refault(page, shadow);
935 lru_cache_add(page);
936 }
937 return ret;
938}
939EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
940
941#ifdef CONFIG_NUMA
942struct page *__page_cache_alloc(gfp_t gfp)
943{
944 int n;
945 struct page *page;
946
947 if (cpuset_do_page_mem_spread()) {
948 unsigned int cpuset_mems_cookie;
949 do {
950 cpuset_mems_cookie = read_mems_allowed_begin();
951 n = cpuset_mem_spread_node();
952 page = __alloc_pages_node(n, gfp, 0);
953 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
954
955 return page;
956 }
957 return alloc_pages(gfp, 0);
958}
959EXPORT_SYMBOL(__page_cache_alloc);
960#endif
961
962/*
963 * In order to wait for pages to become available there must be
964 * waitqueues associated with pages. By using a hash table of
965 * waitqueues where the bucket discipline is to maintain all
966 * waiters on the same queue and wake all when any of the pages
967 * become available, and for the woken contexts to check to be
968 * sure the appropriate page became available, this saves space
969 * at a cost of "thundering herd" phenomena during rare hash
970 * collisions.
971 */
972#define PAGE_WAIT_TABLE_BITS 8
973#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
974static wait_queue_head_t page_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
975
976static wait_queue_head_t *page_waitqueue(struct page *page)
977{
978 return &page_wait_table[hash_ptr(page, PAGE_WAIT_TABLE_BITS)];
979}
980
981void __init pagecache_init(void)
982{
983 int i;
984
985 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
986 init_waitqueue_head(&page_wait_table[i]);
987
988 page_writeback_init();
989}
990
991/*
992 * The page wait code treats the "wait->flags" somewhat unusually, because
993 * we have multiple different kinds of waits, not just the usual "exclusive"
994 * one.
995 *
996 * We have:
997 *
998 * (a) no special bits set:
999 *
1000 * We're just waiting for the bit to be released, and when a waker
1001 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1002 * and remove it from the wait queue.
1003 *
1004 * Simple and straightforward.
1005 *
1006 * (b) WQ_FLAG_EXCLUSIVE:
1007 *
1008 * The waiter is waiting to get the lock, and only one waiter should
1009 * be woken up to avoid any thundering herd behavior. We'll set the
1010 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1011 *
1012 * This is the traditional exclusive wait.
1013 *
1014 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1015 *
1016 * The waiter is waiting to get the bit, and additionally wants the
1017 * lock to be transferred to it for fair lock behavior. If the lock
1018 * cannot be taken, we stop walking the wait queue without waking
1019 * the waiter.
1020 *
1021 * This is the "fair lock handoff" case, and in addition to setting
1022 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1023 * that it now has the lock.
1024 */
1025static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1026{
1027 unsigned int flags;
1028 struct wait_page_key *key = arg;
1029 struct wait_page_queue *wait_page
1030 = container_of(wait, struct wait_page_queue, wait);
1031
1032 if (!wake_page_match(wait_page, key))
1033 return 0;
1034
1035 /*
1036 * If it's a lock handoff wait, we get the bit for it, and
1037 * stop walking (and do not wake it up) if we can't.
1038 */
1039 flags = wait->flags;
1040 if (flags & WQ_FLAG_EXCLUSIVE) {
1041 if (test_bit(key->bit_nr, &key->page->flags))
1042 return -1;
1043 if (flags & WQ_FLAG_CUSTOM) {
1044 if (test_and_set_bit(key->bit_nr, &key->page->flags))
1045 return -1;
1046 flags |= WQ_FLAG_DONE;
1047 }
1048 }
1049
1050 /*
1051 * We are holding the wait-queue lock, but the waiter that
1052 * is waiting for this will be checking the flags without
1053 * any locking.
1054 *
1055 * So update the flags atomically, and wake up the waiter
1056 * afterwards to avoid any races. This store-release pairs
1057 * with the load-acquire in wait_on_page_bit_common().
1058 */
1059 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1060 wake_up_state(wait->private, mode);
1061
1062 /*
1063 * Ok, we have successfully done what we're waiting for,
1064 * and we can unconditionally remove the wait entry.
1065 *
1066 * Note that this pairs with the "finish_wait()" in the
1067 * waiter, and has to be the absolute last thing we do.
1068 * After this list_del_init(&wait->entry) the wait entry
1069 * might be de-allocated and the process might even have
1070 * exited.
1071 */
1072 list_del_init_careful(&wait->entry);
1073 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1074}
1075
1076static void wake_up_page_bit(struct page *page, int bit_nr)
1077{
1078 wait_queue_head_t *q = page_waitqueue(page);
1079 struct wait_page_key key;
1080 unsigned long flags;
1081 wait_queue_entry_t bookmark;
1082
1083 key.page = page;
1084 key.bit_nr = bit_nr;
1085 key.page_match = 0;
1086
1087 bookmark.flags = 0;
1088 bookmark.private = NULL;
1089 bookmark.func = NULL;
1090 INIT_LIST_HEAD(&bookmark.entry);
1091
1092 spin_lock_irqsave(&q->lock, flags);
1093 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1094
1095 while (bookmark.flags & WQ_FLAG_BOOKMARK) {
1096 /*
1097 * Take a breather from holding the lock,
1098 * allow pages that finish wake up asynchronously
1099 * to acquire the lock and remove themselves
1100 * from wait queue
1101 */
1102 spin_unlock_irqrestore(&q->lock, flags);
1103 cpu_relax();
1104 spin_lock_irqsave(&q->lock, flags);
1105 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark);
1106 }
1107
1108 /*
1109 * It is possible for other pages to have collided on the waitqueue
1110 * hash, so in that case check for a page match. That prevents a long-
1111 * term waiter
1112 *
1113 * It is still possible to miss a case here, when we woke page waiters
1114 * and removed them from the waitqueue, but there are still other
1115 * page waiters.
1116 */
1117 if (!waitqueue_active(q) || !key.page_match) {
1118 ClearPageWaiters(page);
1119 /*
1120 * It's possible to miss clearing Waiters here, when we woke
1121 * our page waiters, but the hashed waitqueue has waiters for
1122 * other pages on it.
1123 *
1124 * That's okay, it's a rare case. The next waker will clear it.
1125 */
1126 }
1127 spin_unlock_irqrestore(&q->lock, flags);
1128}
1129
1130static void wake_up_page(struct page *page, int bit)
1131{
1132 if (!PageWaiters(page))
1133 return;
1134 wake_up_page_bit(page, bit);
1135}
1136
1137/*
1138 * A choice of three behaviors for wait_on_page_bit_common():
1139 */
1140enum behavior {
1141 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1142 * __lock_page() waiting on then setting PG_locked.
1143 */
1144 SHARED, /* Hold ref to page and check the bit when woken, like
1145 * wait_on_page_writeback() waiting on PG_writeback.
1146 */
1147 DROP, /* Drop ref to page before wait, no check when woken,
1148 * like put_and_wait_on_page_locked() on PG_locked.
1149 */
1150};
1151
1152/*
1153 * Attempt to check (or get) the page bit, and mark us done
1154 * if successful.
1155 */
1156static inline bool trylock_page_bit_common(struct page *page, int bit_nr,
1157 struct wait_queue_entry *wait)
1158{
1159 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1160 if (test_and_set_bit(bit_nr, &page->flags))
1161 return false;
1162 } else if (test_bit(bit_nr, &page->flags))
1163 return false;
1164
1165 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1166 return true;
1167}
1168
1169/* How many times do we accept lock stealing from under a waiter? */
1170int sysctl_page_lock_unfairness = 5;
1171
1172static inline int wait_on_page_bit_common(wait_queue_head_t *q,
1173 struct page *page, int bit_nr, int state, enum behavior behavior)
1174{
1175 int unfairness = sysctl_page_lock_unfairness;
1176 struct wait_page_queue wait_page;
1177 wait_queue_entry_t *wait = &wait_page.wait;
1178 bool thrashing = false;
1179 bool delayacct = false;
1180 unsigned long pflags;
1181
1182 if (bit_nr == PG_locked &&
1183 !PageUptodate(page) && PageWorkingset(page)) {
1184 if (!PageSwapBacked(page)) {
1185 delayacct_thrashing_start();
1186 delayacct = true;
1187 }
1188 psi_memstall_enter(&pflags);
1189 thrashing = true;
1190 }
1191
1192 init_wait(wait);
1193 wait->func = wake_page_function;
1194 wait_page.page = page;
1195 wait_page.bit_nr = bit_nr;
1196
1197repeat:
1198 wait->flags = 0;
1199 if (behavior == EXCLUSIVE) {
1200 wait->flags = WQ_FLAG_EXCLUSIVE;
1201 if (--unfairness < 0)
1202 wait->flags |= WQ_FLAG_CUSTOM;
1203 }
1204
1205 /*
1206 * Do one last check whether we can get the
1207 * page bit synchronously.
1208 *
1209 * Do the SetPageWaiters() marking before that
1210 * to let any waker we _just_ missed know they
1211 * need to wake us up (otherwise they'll never
1212 * even go to the slow case that looks at the
1213 * page queue), and add ourselves to the wait
1214 * queue if we need to sleep.
1215 *
1216 * This part needs to be done under the queue
1217 * lock to avoid races.
1218 */
1219 spin_lock_irq(&q->lock);
1220 SetPageWaiters(page);
1221 if (!trylock_page_bit_common(page, bit_nr, wait))
1222 __add_wait_queue_entry_tail(q, wait);
1223 spin_unlock_irq(&q->lock);
1224
1225 /*
1226 * From now on, all the logic will be based on
1227 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1228 * see whether the page bit testing has already
1229 * been done by the wake function.
1230 *
1231 * We can drop our reference to the page.
1232 */
1233 if (behavior == DROP)
1234 put_page(page);
1235
1236 /*
1237 * Note that until the "finish_wait()", or until
1238 * we see the WQ_FLAG_WOKEN flag, we need to
1239 * be very careful with the 'wait->flags', because
1240 * we may race with a waker that sets them.
1241 */
1242 for (;;) {
1243 unsigned int flags;
1244
1245 set_current_state(state);
1246
1247 /* Loop until we've been woken or interrupted */
1248 flags = smp_load_acquire(&wait->flags);
1249 if (!(flags & WQ_FLAG_WOKEN)) {
1250 if (signal_pending_state(state, current))
1251 break;
1252
1253 io_schedule();
1254 continue;
1255 }
1256
1257 /* If we were non-exclusive, we're done */
1258 if (behavior != EXCLUSIVE)
1259 break;
1260
1261 /* If the waker got the lock for us, we're done */
1262 if (flags & WQ_FLAG_DONE)
1263 break;
1264
1265 /*
1266 * Otherwise, if we're getting the lock, we need to
1267 * try to get it ourselves.
1268 *
1269 * And if that fails, we'll have to retry this all.
1270 */
1271 if (unlikely(test_and_set_bit(bit_nr, &page->flags)))
1272 goto repeat;
1273
1274 wait->flags |= WQ_FLAG_DONE;
1275 break;
1276 }
1277
1278 /*
1279 * If a signal happened, this 'finish_wait()' may remove the last
1280 * waiter from the wait-queues, but the PageWaiters bit will remain
1281 * set. That's ok. The next wakeup will take care of it, and trying
1282 * to do it here would be difficult and prone to races.
1283 */
1284 finish_wait(q, wait);
1285
1286 if (thrashing) {
1287 if (delayacct)
1288 delayacct_thrashing_end();
1289 psi_memstall_leave(&pflags);
1290 }
1291
1292 /*
1293 * NOTE! The wait->flags weren't stable until we've done the
1294 * 'finish_wait()', and we could have exited the loop above due
1295 * to a signal, and had a wakeup event happen after the signal
1296 * test but before the 'finish_wait()'.
1297 *
1298 * So only after the finish_wait() can we reliably determine
1299 * if we got woken up or not, so we can now figure out the final
1300 * return value based on that state without races.
1301 *
1302 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1303 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1304 */
1305 if (behavior == EXCLUSIVE)
1306 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1307
1308 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1309}
1310
1311void wait_on_page_bit(struct page *page, int bit_nr)
1312{
1313 wait_queue_head_t *q = page_waitqueue(page);
1314 wait_on_page_bit_common(q, page, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1315}
1316EXPORT_SYMBOL(wait_on_page_bit);
1317
1318int wait_on_page_bit_killable(struct page *page, int bit_nr)
1319{
1320 wait_queue_head_t *q = page_waitqueue(page);
1321 return wait_on_page_bit_common(q, page, bit_nr, TASK_KILLABLE, SHARED);
1322}
1323EXPORT_SYMBOL(wait_on_page_bit_killable);
1324
1325static int __wait_on_page_locked_async(struct page *page,
1326 struct wait_page_queue *wait, bool set)
1327{
1328 struct wait_queue_head *q = page_waitqueue(page);
1329 int ret = 0;
1330
1331 wait->page = page;
1332 wait->bit_nr = PG_locked;
1333
1334 spin_lock_irq(&q->lock);
1335 __add_wait_queue_entry_tail(q, &wait->wait);
1336 SetPageWaiters(page);
1337 if (set)
1338 ret = !trylock_page(page);
1339 else
1340 ret = PageLocked(page);
1341 /*
1342 * If we were succesful now, we know we're still on the
1343 * waitqueue as we're still under the lock. This means it's
1344 * safe to remove and return success, we know the callback
1345 * isn't going to trigger.
1346 */
1347 if (!ret)
1348 __remove_wait_queue(q, &wait->wait);
1349 else
1350 ret = -EIOCBQUEUED;
1351 spin_unlock_irq(&q->lock);
1352 return ret;
1353}
1354
1355static int wait_on_page_locked_async(struct page *page,
1356 struct wait_page_queue *wait)
1357{
1358 if (!PageLocked(page))
1359 return 0;
1360 return __wait_on_page_locked_async(compound_head(page), wait, false);
1361}
1362
1363/**
1364 * put_and_wait_on_page_locked - Drop a reference and wait for it to be unlocked
1365 * @page: The page to wait for.
1366 *
1367 * The caller should hold a reference on @page. They expect the page to
1368 * become unlocked relatively soon, but do not wish to hold up migration
1369 * (for example) by holding the reference while waiting for the page to
1370 * come unlocked. After this function returns, the caller should not
1371 * dereference @page.
1372 */
1373void put_and_wait_on_page_locked(struct page *page)
1374{
1375 wait_queue_head_t *q;
1376
1377 page = compound_head(page);
1378 q = page_waitqueue(page);
1379 wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE, DROP);
1380}
1381
1382/**
1383 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
1384 * @page: Page defining the wait queue of interest
1385 * @waiter: Waiter to add to the queue
1386 *
1387 * Add an arbitrary @waiter to the wait queue for the nominated @page.
1388 */
1389void add_page_wait_queue(struct page *page, wait_queue_entry_t *waiter)
1390{
1391 wait_queue_head_t *q = page_waitqueue(page);
1392 unsigned long flags;
1393
1394 spin_lock_irqsave(&q->lock, flags);
1395 __add_wait_queue_entry_tail(q, waiter);
1396 SetPageWaiters(page);
1397 spin_unlock_irqrestore(&q->lock, flags);
1398}
1399EXPORT_SYMBOL_GPL(add_page_wait_queue);
1400
1401#ifndef clear_bit_unlock_is_negative_byte
1402
1403/*
1404 * PG_waiters is the high bit in the same byte as PG_lock.
1405 *
1406 * On x86 (and on many other architectures), we can clear PG_lock and
1407 * test the sign bit at the same time. But if the architecture does
1408 * not support that special operation, we just do this all by hand
1409 * instead.
1410 *
1411 * The read of PG_waiters has to be after (or concurrently with) PG_locked
1412 * being cleared, but a memory barrier should be unnecessary since it is
1413 * in the same byte as PG_locked.
1414 */
1415static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem)
1416{
1417 clear_bit_unlock(nr, mem);
1418 /* smp_mb__after_atomic(); */
1419 return test_bit(PG_waiters, mem);
1420}
1421
1422#endif
1423
1424/**
1425 * unlock_page - unlock a locked page
1426 * @page: the page
1427 *
1428 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
1429 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
1430 * mechanism between PageLocked pages and PageWriteback pages is shared.
1431 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
1432 *
1433 * Note that this depends on PG_waiters being the sign bit in the byte
1434 * that contains PG_locked - thus the BUILD_BUG_ON(). That allows us to
1435 * clear the PG_locked bit and test PG_waiters at the same time fairly
1436 * portably (architectures that do LL/SC can test any bit, while x86 can
1437 * test the sign bit).
1438 */
1439void unlock_page(struct page *page)
1440{
1441 BUILD_BUG_ON(PG_waiters != 7);
1442 page = compound_head(page);
1443 VM_BUG_ON_PAGE(!PageLocked(page), page);
1444 if (clear_bit_unlock_is_negative_byte(PG_locked, &page->flags))
1445 wake_up_page_bit(page, PG_locked);
1446}
1447EXPORT_SYMBOL(unlock_page);
1448
1449/**
1450 * end_page_writeback - end writeback against a page
1451 * @page: the page
1452 */
1453void end_page_writeback(struct page *page)
1454{
1455 /*
1456 * TestClearPageReclaim could be used here but it is an atomic
1457 * operation and overkill in this particular case. Failing to
1458 * shuffle a page marked for immediate reclaim is too mild to
1459 * justify taking an atomic operation penalty at the end of
1460 * ever page writeback.
1461 */
1462 if (PageReclaim(page)) {
1463 ClearPageReclaim(page);
1464 rotate_reclaimable_page(page);
1465 }
1466
1467 if (!test_clear_page_writeback(page))
1468 BUG();
1469
1470 smp_mb__after_atomic();
1471 wake_up_page(page, PG_writeback);
1472}
1473EXPORT_SYMBOL(end_page_writeback);
1474
1475/*
1476 * After completing I/O on a page, call this routine to update the page
1477 * flags appropriately
1478 */
1479void page_endio(struct page *page, bool is_write, int err)
1480{
1481 if (!is_write) {
1482 if (!err) {
1483 SetPageUptodate(page);
1484 } else {
1485 ClearPageUptodate(page);
1486 SetPageError(page);
1487 }
1488 unlock_page(page);
1489 } else {
1490 if (err) {
1491 struct address_space *mapping;
1492
1493 SetPageError(page);
1494 mapping = page_mapping(page);
1495 if (mapping)
1496 mapping_set_error(mapping, err);
1497 }
1498 end_page_writeback(page);
1499 }
1500}
1501EXPORT_SYMBOL_GPL(page_endio);
1502
1503/**
1504 * __lock_page - get a lock on the page, assuming we need to sleep to get it
1505 * @__page: the page to lock
1506 */
1507void __lock_page(struct page *__page)
1508{
1509 struct page *page = compound_head(__page);
1510 wait_queue_head_t *q = page_waitqueue(page);
1511 wait_on_page_bit_common(q, page, PG_locked, TASK_UNINTERRUPTIBLE,
1512 EXCLUSIVE);
1513}
1514EXPORT_SYMBOL(__lock_page);
1515
1516int __lock_page_killable(struct page *__page)
1517{
1518 struct page *page = compound_head(__page);
1519 wait_queue_head_t *q = page_waitqueue(page);
1520 return wait_on_page_bit_common(q, page, PG_locked, TASK_KILLABLE,
1521 EXCLUSIVE);
1522}
1523EXPORT_SYMBOL_GPL(__lock_page_killable);
1524
1525int __lock_page_async(struct page *page, struct wait_page_queue *wait)
1526{
1527 return __wait_on_page_locked_async(page, wait, true);
1528}
1529
1530/*
1531 * Return values:
1532 * 1 - page is locked; mmap_lock is still held.
1533 * 0 - page is not locked.
1534 * mmap_lock has been released (mmap_read_unlock(), unless flags had both
1535 * FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
1536 * which case mmap_lock is still held.
1537 *
1538 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
1539 * with the page locked and the mmap_lock unperturbed.
1540 */
1541int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
1542 unsigned int flags)
1543{
1544 if (fault_flag_allow_retry_first(flags)) {
1545 /*
1546 * CAUTION! In this case, mmap_lock is not released
1547 * even though return 0.
1548 */
1549 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1550 return 0;
1551
1552 mmap_read_unlock(mm);
1553 if (flags & FAULT_FLAG_KILLABLE)
1554 wait_on_page_locked_killable(page);
1555 else
1556 wait_on_page_locked(page);
1557 return 0;
1558 } else {
1559 if (flags & FAULT_FLAG_KILLABLE) {
1560 int ret;
1561
1562 ret = __lock_page_killable(page);
1563 if (ret) {
1564 mmap_read_unlock(mm);
1565 return 0;
1566 }
1567 } else
1568 __lock_page(page);
1569 return 1;
1570 }
1571}
1572
1573/**
1574 * page_cache_next_miss() - Find the next gap in the page cache.
1575 * @mapping: Mapping.
1576 * @index: Index.
1577 * @max_scan: Maximum range to search.
1578 *
1579 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1580 * gap with the lowest index.
1581 *
1582 * This function may be called under the rcu_read_lock. However, this will
1583 * not atomically search a snapshot of the cache at a single point in time.
1584 * For example, if a gap is created at index 5, then subsequently a gap is
1585 * created at index 10, page_cache_next_miss covering both indices may
1586 * return 10 if called under the rcu_read_lock.
1587 *
1588 * Return: The index of the gap if found, otherwise an index outside the
1589 * range specified (in which case 'return - index >= max_scan' will be true).
1590 * In the rare case of index wrap-around, 0 will be returned.
1591 */
1592pgoff_t page_cache_next_miss(struct address_space *mapping,
1593 pgoff_t index, unsigned long max_scan)
1594{
1595 XA_STATE(xas, &mapping->i_pages, index);
1596
1597 while (max_scan--) {
1598 void *entry = xas_next(&xas);
1599 if (!entry || xa_is_value(entry))
1600 break;
1601 if (xas.xa_index == 0)
1602 break;
1603 }
1604
1605 return xas.xa_index;
1606}
1607EXPORT_SYMBOL(page_cache_next_miss);
1608
1609/**
1610 * page_cache_prev_miss() - Find the previous gap in the page cache.
1611 * @mapping: Mapping.
1612 * @index: Index.
1613 * @max_scan: Maximum range to search.
1614 *
1615 * Search the range [max(index - max_scan + 1, 0), index] for the
1616 * gap with the highest index.
1617 *
1618 * This function may be called under the rcu_read_lock. However, this will
1619 * not atomically search a snapshot of the cache at a single point in time.
1620 * For example, if a gap is created at index 10, then subsequently a gap is
1621 * created at index 5, page_cache_prev_miss() covering both indices may
1622 * return 5 if called under the rcu_read_lock.
1623 *
1624 * Return: The index of the gap if found, otherwise an index outside the
1625 * range specified (in which case 'index - return >= max_scan' will be true).
1626 * In the rare case of wrap-around, ULONG_MAX will be returned.
1627 */
1628pgoff_t page_cache_prev_miss(struct address_space *mapping,
1629 pgoff_t index, unsigned long max_scan)
1630{
1631 XA_STATE(xas, &mapping->i_pages, index);
1632
1633 while (max_scan--) {
1634 void *entry = xas_prev(&xas);
1635 if (!entry || xa_is_value(entry))
1636 break;
1637 if (xas.xa_index == ULONG_MAX)
1638 break;
1639 }
1640
1641 return xas.xa_index;
1642}
1643EXPORT_SYMBOL(page_cache_prev_miss);
1644
1645/**
1646 * find_get_entry - find and get a page cache entry
1647 * @mapping: the address_space to search
1648 * @offset: the page cache index
1649 *
1650 * Looks up the page cache slot at @mapping & @offset. If there is a
1651 * page cache page, it is returned with an increased refcount.
1652 *
1653 * If the slot holds a shadow entry of a previously evicted page, or a
1654 * swap entry from shmem/tmpfs, it is returned.
1655 *
1656 * Return: the found page or shadow entry, %NULL if nothing is found.
1657 */
1658struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1659{
1660 XA_STATE(xas, &mapping->i_pages, offset);
1661 struct page *page;
1662
1663 rcu_read_lock();
1664repeat:
1665 xas_reset(&xas);
1666 page = xas_load(&xas);
1667 if (xas_retry(&xas, page))
1668 goto repeat;
1669 /*
1670 * A shadow entry of a recently evicted page, or a swap entry from
1671 * shmem/tmpfs. Return it without attempting to raise page count.
1672 */
1673 if (!page || xa_is_value(page))
1674 goto out;
1675
1676 if (!page_cache_get_speculative(page))
1677 goto repeat;
1678
1679 /*
1680 * Has the page moved or been split?
1681 * This is part of the lockless pagecache protocol. See
1682 * include/linux/pagemap.h for details.
1683 */
1684 if (unlikely(page != xas_reload(&xas))) {
1685 put_page(page);
1686 goto repeat;
1687 }
1688 page = find_subpage(page, offset);
1689out:
1690 rcu_read_unlock();
1691
1692 return page;
1693}
1694
1695/**
1696 * find_lock_entry - locate, pin and lock a page cache entry
1697 * @mapping: the address_space to search
1698 * @offset: the page cache index
1699 *
1700 * Looks up the page cache slot at @mapping & @offset. If there is a
1701 * page cache page, it is returned locked and with an increased
1702 * refcount.
1703 *
1704 * If the slot holds a shadow entry of a previously evicted page, or a
1705 * swap entry from shmem/tmpfs, it is returned.
1706 *
1707 * find_lock_entry() may sleep.
1708 *
1709 * Return: the found page or shadow entry, %NULL if nothing is found.
1710 */
1711struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1712{
1713 struct page *page;
1714
1715repeat:
1716 page = find_get_entry(mapping, offset);
1717 if (page && !xa_is_value(page)) {
1718 lock_page(page);
1719 /* Has the page been truncated? */
1720 if (unlikely(page_mapping(page) != mapping)) {
1721 unlock_page(page);
1722 put_page(page);
1723 goto repeat;
1724 }
1725 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
1726 }
1727 return page;
1728}
1729EXPORT_SYMBOL(find_lock_entry);
1730
1731/**
1732 * pagecache_get_page - Find and get a reference to a page.
1733 * @mapping: The address_space to search.
1734 * @index: The page index.
1735 * @fgp_flags: %FGP flags modify how the page is returned.
1736 * @gfp_mask: Memory allocation flags to use if %FGP_CREAT is specified.
1737 *
1738 * Looks up the page cache entry at @mapping & @index.
1739 *
1740 * @fgp_flags can be zero or more of these flags:
1741 *
1742 * * %FGP_ACCESSED - The page will be marked accessed.
1743 * * %FGP_LOCK - The page is returned locked.
1744 * * %FGP_CREAT - If no page is present then a new page is allocated using
1745 * @gfp_mask and added to the page cache and the VM's LRU list.
1746 * The page is returned locked and with an increased refcount.
1747 * * %FGP_FOR_MMAP - The caller wants to do its own locking dance if the
1748 * page is already in cache. If the page was allocated, unlock it before
1749 * returning so the caller can do the same dance.
1750 * * %FGP_WRITE - The page will be written
1751 * * %FGP_NOFS - __GFP_FS will get cleared in gfp mask
1752 * * %FGP_NOWAIT - Don't get blocked by page lock
1753 *
1754 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1755 * if the %GFP flags specified for %FGP_CREAT are atomic.
1756 *
1757 * If there is a page cache page, it is returned with an increased refcount.
1758 *
1759 * Return: The found page or %NULL otherwise.
1760 */
1761struct page *pagecache_get_page(struct address_space *mapping, pgoff_t index,
1762 int fgp_flags, gfp_t gfp_mask)
1763{
1764 struct page *page;
1765
1766repeat:
1767 page = find_get_entry(mapping, index);
1768 if (xa_is_value(page))
1769 page = NULL;
1770 if (!page)
1771 goto no_page;
1772
1773 if (fgp_flags & FGP_LOCK) {
1774 if (fgp_flags & FGP_NOWAIT) {
1775 if (!trylock_page(page)) {
1776 put_page(page);
1777 return NULL;
1778 }
1779 } else {
1780 lock_page(page);
1781 }
1782
1783 /* Has the page been truncated? */
1784 if (unlikely(compound_head(page)->mapping != mapping)) {
1785 unlock_page(page);
1786 put_page(page);
1787 goto repeat;
1788 }
1789 VM_BUG_ON_PAGE(page->index != index, page);
1790 }
1791
1792 if (fgp_flags & FGP_ACCESSED)
1793 mark_page_accessed(page);
1794 else if (fgp_flags & FGP_WRITE) {
1795 /* Clear idle flag for buffer write */
1796 if (page_is_idle(page))
1797 clear_page_idle(page);
1798 }
1799
1800no_page:
1801 if (!page && (fgp_flags & FGP_CREAT)) {
1802 int err;
1803 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1804 gfp_mask |= __GFP_WRITE;
1805 if (fgp_flags & FGP_NOFS)
1806 gfp_mask &= ~__GFP_FS;
1807
1808 page = __page_cache_alloc(gfp_mask);
1809 if (!page)
1810 return NULL;
1811
1812 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1813 fgp_flags |= FGP_LOCK;
1814
1815 /* Init accessed so avoid atomic mark_page_accessed later */
1816 if (fgp_flags & FGP_ACCESSED)
1817 __SetPageReferenced(page);
1818
1819 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
1820 if (unlikely(err)) {
1821 put_page(page);
1822 page = NULL;
1823 if (err == -EEXIST)
1824 goto repeat;
1825 }
1826
1827 /*
1828 * add_to_page_cache_lru locks the page, and for mmap we expect
1829 * an unlocked page.
1830 */
1831 if (page && (fgp_flags & FGP_FOR_MMAP))
1832 unlock_page(page);
1833 }
1834
1835 return page;
1836}
1837EXPORT_SYMBOL(pagecache_get_page);
1838
1839/**
1840 * find_get_entries - gang pagecache lookup
1841 * @mapping: The address_space to search
1842 * @start: The starting page cache index
1843 * @nr_entries: The maximum number of entries
1844 * @entries: Where the resulting entries are placed
1845 * @indices: The cache indices corresponding to the entries in @entries
1846 *
1847 * find_get_entries() will search for and return a group of up to
1848 * @nr_entries entries in the mapping. The entries are placed at
1849 * @entries. find_get_entries() takes a reference against any actual
1850 * pages it returns.
1851 *
1852 * The search returns a group of mapping-contiguous page cache entries
1853 * with ascending indexes. There may be holes in the indices due to
1854 * not-present pages.
1855 *
1856 * Any shadow entries of evicted pages, or swap entries from
1857 * shmem/tmpfs, are included in the returned array.
1858 *
1859 * If it finds a Transparent Huge Page, head or tail, find_get_entries()
1860 * stops at that page: the caller is likely to have a better way to handle
1861 * the compound page as a whole, and then skip its extent, than repeatedly
1862 * calling find_get_entries() to return all its tails.
1863 *
1864 * Return: the number of pages and shadow entries which were found.
1865 */
1866unsigned find_get_entries(struct address_space *mapping,
1867 pgoff_t start, unsigned int nr_entries,
1868 struct page **entries, pgoff_t *indices)
1869{
1870 XA_STATE(xas, &mapping->i_pages, start);
1871 struct page *page;
1872 unsigned int ret = 0;
1873
1874 if (!nr_entries)
1875 return 0;
1876
1877 rcu_read_lock();
1878 xas_for_each(&xas, page, ULONG_MAX) {
1879 if (xas_retry(&xas, page))
1880 continue;
1881 /*
1882 * A shadow entry of a recently evicted page, a swap
1883 * entry from shmem/tmpfs or a DAX entry. Return it
1884 * without attempting to raise page count.
1885 */
1886 if (xa_is_value(page))
1887 goto export;
1888
1889 if (!page_cache_get_speculative(page))
1890 goto retry;
1891
1892 /* Has the page moved or been split? */
1893 if (unlikely(page != xas_reload(&xas)))
1894 goto put_page;
1895
1896 /*
1897 * Terminate early on finding a THP, to allow the caller to
1898 * handle it all at once; but continue if this is hugetlbfs.
1899 */
1900 if (PageTransHuge(page) && !PageHuge(page)) {
1901 page = find_subpage(page, xas.xa_index);
1902 nr_entries = ret + 1;
1903 }
1904export:
1905 indices[ret] = xas.xa_index;
1906 entries[ret] = page;
1907 if (++ret == nr_entries)
1908 break;
1909 continue;
1910put_page:
1911 put_page(page);
1912retry:
1913 xas_reset(&xas);
1914 }
1915 rcu_read_unlock();
1916 return ret;
1917}
1918
1919/**
1920 * find_get_pages_range - gang pagecache lookup
1921 * @mapping: The address_space to search
1922 * @start: The starting page index
1923 * @end: The final page index (inclusive)
1924 * @nr_pages: The maximum number of pages
1925 * @pages: Where the resulting pages are placed
1926 *
1927 * find_get_pages_range() will search for and return a group of up to @nr_pages
1928 * pages in the mapping starting at index @start and up to index @end
1929 * (inclusive). The pages are placed at @pages. find_get_pages_range() takes
1930 * a reference against the returned pages.
1931 *
1932 * The search returns a group of mapping-contiguous pages with ascending
1933 * indexes. There may be holes in the indices due to not-present pages.
1934 * We also update @start to index the next page for the traversal.
1935 *
1936 * Return: the number of pages which were found. If this number is
1937 * smaller than @nr_pages, the end of specified range has been
1938 * reached.
1939 */
1940unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
1941 pgoff_t end, unsigned int nr_pages,
1942 struct page **pages)
1943{
1944 XA_STATE(xas, &mapping->i_pages, *start);
1945 struct page *page;
1946 unsigned ret = 0;
1947
1948 if (unlikely(!nr_pages))
1949 return 0;
1950
1951 rcu_read_lock();
1952 xas_for_each(&xas, page, end) {
1953 if (xas_retry(&xas, page))
1954 continue;
1955 /* Skip over shadow, swap and DAX entries */
1956 if (xa_is_value(page))
1957 continue;
1958
1959 if (!page_cache_get_speculative(page))
1960 goto retry;
1961
1962 /* Has the page moved or been split? */
1963 if (unlikely(page != xas_reload(&xas)))
1964 goto put_page;
1965
1966 pages[ret] = find_subpage(page, xas.xa_index);
1967 if (++ret == nr_pages) {
1968 *start = xas.xa_index + 1;
1969 goto out;
1970 }
1971 continue;
1972put_page:
1973 put_page(page);
1974retry:
1975 xas_reset(&xas);
1976 }
1977
1978 /*
1979 * We come here when there is no page beyond @end. We take care to not
1980 * overflow the index @start as it confuses some of the callers. This
1981 * breaks the iteration when there is a page at index -1 but that is
1982 * already broken anyway.
1983 */
1984 if (end == (pgoff_t)-1)
1985 *start = (pgoff_t)-1;
1986 else
1987 *start = end + 1;
1988out:
1989 rcu_read_unlock();
1990
1991 return ret;
1992}
1993
1994/**
1995 * find_get_pages_contig - gang contiguous pagecache lookup
1996 * @mapping: The address_space to search
1997 * @index: The starting page index
1998 * @nr_pages: The maximum number of pages
1999 * @pages: Where the resulting pages are placed
2000 *
2001 * find_get_pages_contig() works exactly like find_get_pages(), except
2002 * that the returned number of pages are guaranteed to be contiguous.
2003 *
2004 * Return: the number of pages which were found.
2005 */
2006unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
2007 unsigned int nr_pages, struct page **pages)
2008{
2009 XA_STATE(xas, &mapping->i_pages, index);
2010 struct page *page;
2011 unsigned int ret = 0;
2012
2013 if (unlikely(!nr_pages))
2014 return 0;
2015
2016 rcu_read_lock();
2017 for (page = xas_load(&xas); page; page = xas_next(&xas)) {
2018 if (xas_retry(&xas, page))
2019 continue;
2020 /*
2021 * If the entry has been swapped out, we can stop looking.
2022 * No current caller is looking for DAX entries.
2023 */
2024 if (xa_is_value(page))
2025 break;
2026
2027 if (!page_cache_get_speculative(page))
2028 goto retry;
2029
2030 /* Has the page moved or been split? */
2031 if (unlikely(page != xas_reload(&xas)))
2032 goto put_page;
2033
2034 pages[ret] = find_subpage(page, xas.xa_index);
2035 if (++ret == nr_pages)
2036 break;
2037 continue;
2038put_page:
2039 put_page(page);
2040retry:
2041 xas_reset(&xas);
2042 }
2043 rcu_read_unlock();
2044 return ret;
2045}
2046EXPORT_SYMBOL(find_get_pages_contig);
2047
2048/**
2049 * find_get_pages_range_tag - find and return pages in given range matching @tag
2050 * @mapping: the address_space to search
2051 * @index: the starting page index
2052 * @end: The final page index (inclusive)
2053 * @tag: the tag index
2054 * @nr_pages: the maximum number of pages
2055 * @pages: where the resulting pages are placed
2056 *
2057 * Like find_get_pages, except we only return pages which are tagged with
2058 * @tag. We update @index to index the next page for the traversal.
2059 *
2060 * Return: the number of pages which were found.
2061 */
2062unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
2063 pgoff_t end, xa_mark_t tag, unsigned int nr_pages,
2064 struct page **pages)
2065{
2066 XA_STATE(xas, &mapping->i_pages, *index);
2067 struct page *page;
2068 unsigned ret = 0;
2069
2070 if (unlikely(!nr_pages))
2071 return 0;
2072
2073 rcu_read_lock();
2074 xas_for_each_marked(&xas, page, end, tag) {
2075 if (xas_retry(&xas, page))
2076 continue;
2077 /*
2078 * Shadow entries should never be tagged, but this iteration
2079 * is lockless so there is a window for page reclaim to evict
2080 * a page we saw tagged. Skip over it.
2081 */
2082 if (xa_is_value(page))
2083 continue;
2084
2085 if (!page_cache_get_speculative(page))
2086 goto retry;
2087
2088 /* Has the page moved or been split? */
2089 if (unlikely(page != xas_reload(&xas)))
2090 goto put_page;
2091
2092 pages[ret] = find_subpage(page, xas.xa_index);
2093 if (++ret == nr_pages) {
2094 *index = xas.xa_index + 1;
2095 goto out;
2096 }
2097 continue;
2098put_page:
2099 put_page(page);
2100retry:
2101 xas_reset(&xas);
2102 }
2103
2104 /*
2105 * We come here when we got to @end. We take care to not overflow the
2106 * index @index as it confuses some of the callers. This breaks the
2107 * iteration when there is a page at index -1 but that is already
2108 * broken anyway.
2109 */
2110 if (end == (pgoff_t)-1)
2111 *index = (pgoff_t)-1;
2112 else
2113 *index = end + 1;
2114out:
2115 rcu_read_unlock();
2116
2117 return ret;
2118}
2119EXPORT_SYMBOL(find_get_pages_range_tag);
2120
2121/*
2122 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2123 * a _large_ part of the i/o request. Imagine the worst scenario:
2124 *
2125 * ---R__________________________________________B__________
2126 * ^ reading here ^ bad block(assume 4k)
2127 *
2128 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2129 * => failing the whole request => read(R) => read(R+1) =>
2130 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2131 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2132 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2133 *
2134 * It is going insane. Fix it by quickly scaling down the readahead size.
2135 */
2136static void shrink_readahead_size_eio(struct file_ra_state *ra)
2137{
2138 ra->ra_pages /= 4;
2139}
2140
2141/**
2142 * generic_file_buffered_read - generic file read routine
2143 * @iocb: the iocb to read
2144 * @iter: data destination
2145 * @written: already copied
2146 *
2147 * This is a generic file read routine, and uses the
2148 * mapping->a_ops->readpage() function for the actual low-level stuff.
2149 *
2150 * This is really ugly. But the goto's actually try to clarify some
2151 * of the logic when it comes to error handling etc.
2152 *
2153 * Return:
2154 * * total number of bytes copied, including those the were already @written
2155 * * negative error code if nothing was copied
2156 */
2157ssize_t generic_file_buffered_read(struct kiocb *iocb,
2158 struct iov_iter *iter, ssize_t written)
2159{
2160 struct file *filp = iocb->ki_filp;
2161 struct address_space *mapping = filp->f_mapping;
2162 struct inode *inode = mapping->host;
2163 struct file_ra_state *ra = &filp->f_ra;
2164 loff_t *ppos = &iocb->ki_pos;
2165 pgoff_t index;
2166 pgoff_t last_index;
2167 pgoff_t prev_index;
2168 unsigned long offset; /* offset into pagecache page */
2169 unsigned int prev_offset;
2170 int error = 0;
2171
2172 if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
2173 return 0;
2174 iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
2175
2176 index = *ppos >> PAGE_SHIFT;
2177 prev_index = ra->prev_pos >> PAGE_SHIFT;
2178 prev_offset = ra->prev_pos & (PAGE_SIZE-1);
2179 last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
2180 offset = *ppos & ~PAGE_MASK;
2181
2182 for (;;) {
2183 struct page *page;
2184 pgoff_t end_index;
2185 loff_t isize;
2186 unsigned long nr, ret;
2187
2188 cond_resched();
2189find_page:
2190 if (fatal_signal_pending(current)) {
2191 error = -EINTR;
2192 goto out;
2193 }
2194
2195 page = find_get_page(mapping, index);
2196 if (!page) {
2197 if (iocb->ki_flags & IOCB_NOIO)
2198 goto would_block;
2199 page_cache_sync_readahead(mapping,
2200 ra, filp,
2201 index, last_index - index);
2202 page = find_get_page(mapping, index);
2203 if (unlikely(page == NULL))
2204 goto no_cached_page;
2205 }
2206 if (PageReadahead(page)) {
2207 if (iocb->ki_flags & IOCB_NOIO) {
2208 put_page(page);
2209 goto out;
2210 }
2211 page_cache_async_readahead(mapping,
2212 ra, filp, page,
2213 index, last_index - index);
2214 }
2215 if (!PageUptodate(page)) {
2216 /*
2217 * See comment in do_read_cache_page on why
2218 * wait_on_page_locked is used to avoid unnecessarily
2219 * serialisations and why it's safe.
2220 */
2221 if (iocb->ki_flags & IOCB_WAITQ) {
2222 if (written) {
2223 put_page(page);
2224 goto out;
2225 }
2226 error = wait_on_page_locked_async(page,
2227 iocb->ki_waitq);
2228 } else {
2229 if (iocb->ki_flags & IOCB_NOWAIT) {
2230 put_page(page);
2231 goto would_block;
2232 }
2233 error = wait_on_page_locked_killable(page);
2234 }
2235 if (unlikely(error))
2236 goto readpage_error;
2237 if (PageUptodate(page))
2238 goto page_ok;
2239
2240 if (inode->i_blkbits == PAGE_SHIFT ||
2241 !mapping->a_ops->is_partially_uptodate)
2242 goto page_not_up_to_date;
2243 /* pipes can't handle partially uptodate pages */
2244 if (unlikely(iov_iter_is_pipe(iter)))
2245 goto page_not_up_to_date;
2246 if (!trylock_page(page))
2247 goto page_not_up_to_date;
2248 /* Did it get truncated before we got the lock? */
2249 if (!page->mapping)
2250 goto page_not_up_to_date_locked;
2251 if (!mapping->a_ops->is_partially_uptodate(page,
2252 offset, iter->count))
2253 goto page_not_up_to_date_locked;
2254 unlock_page(page);
2255 }
2256page_ok:
2257 /*
2258 * i_size must be checked after we know the page is Uptodate.
2259 *
2260 * Checking i_size after the check allows us to calculate
2261 * the correct value for "nr", which means the zero-filled
2262 * part of the page is not copied back to userspace (unless
2263 * another truncate extends the file - this is desired though).
2264 */
2265
2266 isize = i_size_read(inode);
2267 end_index = (isize - 1) >> PAGE_SHIFT;
2268 if (unlikely(!isize || index > end_index)) {
2269 put_page(page);
2270 goto out;
2271 }
2272
2273 /* nr is the maximum number of bytes to copy from this page */
2274 nr = PAGE_SIZE;
2275 if (index == end_index) {
2276 nr = ((isize - 1) & ~PAGE_MASK) + 1;
2277 if (nr <= offset) {
2278 put_page(page);
2279 goto out;
2280 }
2281 }
2282 nr = nr - offset;
2283
2284 /* If users can be writing to this page using arbitrary
2285 * virtual addresses, take care about potential aliasing
2286 * before reading the page on the kernel side.
2287 */
2288 if (mapping_writably_mapped(mapping))
2289 flush_dcache_page(page);
2290
2291 /*
2292 * When a sequential read accesses a page several times,
2293 * only mark it as accessed the first time.
2294 */
2295 if (prev_index != index || offset != prev_offset)
2296 mark_page_accessed(page);
2297 prev_index = index;
2298
2299 /*
2300 * Ok, we have the page, and it's up-to-date, so
2301 * now we can copy it to user space...
2302 */
2303
2304 ret = copy_page_to_iter(page, offset, nr, iter);
2305 offset += ret;
2306 index += offset >> PAGE_SHIFT;
2307 offset &= ~PAGE_MASK;
2308 prev_offset = offset;
2309
2310 put_page(page);
2311 written += ret;
2312 if (!iov_iter_count(iter))
2313 goto out;
2314 if (ret < nr) {
2315 error = -EFAULT;
2316 goto out;
2317 }
2318 continue;
2319
2320page_not_up_to_date:
2321 /* Get exclusive access to the page ... */
2322 if (iocb->ki_flags & IOCB_WAITQ)
2323 error = lock_page_async(page, iocb->ki_waitq);
2324 else
2325 error = lock_page_killable(page);
2326 if (unlikely(error))
2327 goto readpage_error;
2328
2329page_not_up_to_date_locked:
2330 /* Did it get truncated before we got the lock? */
2331 if (!page->mapping) {
2332 unlock_page(page);
2333 put_page(page);
2334 continue;
2335 }
2336
2337 /* Did somebody else fill it already? */
2338 if (PageUptodate(page)) {
2339 unlock_page(page);
2340 goto page_ok;
2341 }
2342
2343readpage:
2344 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT)) {
2345 unlock_page(page);
2346 put_page(page);
2347 goto would_block;
2348 }
2349 /*
2350 * A previous I/O error may have been due to temporary
2351 * failures, eg. multipath errors.
2352 * PG_error will be set again if readpage fails.
2353 */
2354 ClearPageError(page);
2355 /* Start the actual read. The read will unlock the page. */
2356 error = mapping->a_ops->readpage(filp, page);
2357
2358 if (unlikely(error)) {
2359 if (error == AOP_TRUNCATED_PAGE) {
2360 put_page(page);
2361 error = 0;
2362 goto find_page;
2363 }
2364 goto readpage_error;
2365 }
2366
2367 if (!PageUptodate(page)) {
2368 if (iocb->ki_flags & IOCB_WAITQ)
2369 error = lock_page_async(page, iocb->ki_waitq);
2370 else
2371 error = lock_page_killable(page);
2372
2373 if (unlikely(error))
2374 goto readpage_error;
2375 if (!PageUptodate(page)) {
2376 if (page->mapping == NULL) {
2377 /*
2378 * invalidate_mapping_pages got it
2379 */
2380 unlock_page(page);
2381 put_page(page);
2382 goto find_page;
2383 }
2384 unlock_page(page);
2385 shrink_readahead_size_eio(ra);
2386 error = -EIO;
2387 goto readpage_error;
2388 }
2389 unlock_page(page);
2390 }
2391
2392 goto page_ok;
2393
2394readpage_error:
2395 /* UHHUH! A synchronous read error occurred. Report it */
2396 put_page(page);
2397 goto out;
2398
2399no_cached_page:
2400 /*
2401 * Ok, it wasn't cached, so we need to create a new
2402 * page..
2403 */
2404 page = page_cache_alloc(mapping);
2405 if (!page) {
2406 error = -ENOMEM;
2407 goto out;
2408 }
2409 error = add_to_page_cache_lru(page, mapping, index,
2410 mapping_gfp_constraint(mapping, GFP_KERNEL));
2411 if (error) {
2412 put_page(page);
2413 if (error == -EEXIST) {
2414 error = 0;
2415 goto find_page;
2416 }
2417 goto out;
2418 }
2419 goto readpage;
2420 }
2421
2422would_block:
2423 error = -EAGAIN;
2424out:
2425 ra->prev_pos = prev_index;
2426 ra->prev_pos <<= PAGE_SHIFT;
2427 ra->prev_pos |= prev_offset;
2428
2429 *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
2430 file_accessed(filp);
2431 return written ? written : error;
2432}
2433EXPORT_SYMBOL_GPL(generic_file_buffered_read);
2434
2435/**
2436 * generic_file_read_iter - generic filesystem read routine
2437 * @iocb: kernel I/O control block
2438 * @iter: destination for the data read
2439 *
2440 * This is the "read_iter()" routine for all filesystems
2441 * that can use the page cache directly.
2442 *
2443 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2444 * be returned when no data can be read without waiting for I/O requests
2445 * to complete; it doesn't prevent readahead.
2446 *
2447 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2448 * requests shall be made for the read or for readahead. When no data
2449 * can be read, -EAGAIN shall be returned. When readahead would be
2450 * triggered, a partial, possibly empty read shall be returned.
2451 *
2452 * Return:
2453 * * number of bytes copied, even for partial reads
2454 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2455 */
2456ssize_t
2457generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2458{
2459 size_t count = iov_iter_count(iter);
2460 ssize_t retval = 0;
2461
2462 if (!count)
2463 goto out; /* skip atime */
2464
2465 if (iocb->ki_flags & IOCB_DIRECT) {
2466 struct file *file = iocb->ki_filp;
2467 struct address_space *mapping = file->f_mapping;
2468 struct inode *inode = mapping->host;
2469 loff_t size;
2470
2471 size = i_size_read(inode);
2472 if (iocb->ki_flags & IOCB_NOWAIT) {
2473 if (filemap_range_has_page(mapping, iocb->ki_pos,
2474 iocb->ki_pos + count - 1))
2475 return -EAGAIN;
2476 } else {
2477 retval = filemap_write_and_wait_range(mapping,
2478 iocb->ki_pos,
2479 iocb->ki_pos + count - 1);
2480 if (retval < 0)
2481 goto out;
2482 }
2483
2484 file_accessed(file);
2485
2486 retval = mapping->a_ops->direct_IO(iocb, iter);
2487 if (retval >= 0) {
2488 iocb->ki_pos += retval;
2489 count -= retval;
2490 }
2491 iov_iter_revert(iter, count - iov_iter_count(iter));
2492
2493 /*
2494 * Btrfs can have a short DIO read if we encounter
2495 * compressed extents, so if there was an error, or if
2496 * we've already read everything we wanted to, or if
2497 * there was a short read because we hit EOF, go ahead
2498 * and return. Otherwise fallthrough to buffered io for
2499 * the rest of the read. Buffered reads will not work for
2500 * DAX files, so don't bother trying.
2501 */
2502 if (retval < 0 || !count || iocb->ki_pos >= size ||
2503 IS_DAX(inode))
2504 goto out;
2505 }
2506
2507 retval = generic_file_buffered_read(iocb, iter, retval);
2508out:
2509 return retval;
2510}
2511EXPORT_SYMBOL(generic_file_read_iter);
2512
2513#ifdef CONFIG_MMU
2514#define MMAP_LOTSAMISS (100)
2515/*
2516 * lock_page_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
2517 * @vmf - the vm_fault for this fault.
2518 * @page - the page to lock.
2519 * @fpin - the pointer to the file we may pin (or is already pinned).
2520 *
2521 * This works similar to lock_page_or_retry in that it can drop the mmap_lock.
2522 * It differs in that it actually returns the page locked if it returns 1 and 0
2523 * if it couldn't lock the page. If we did have to drop the mmap_lock then fpin
2524 * will point to the pinned file and needs to be fput()'ed at a later point.
2525 */
2526static int lock_page_maybe_drop_mmap(struct vm_fault *vmf, struct page *page,
2527 struct file **fpin)
2528{
2529 if (trylock_page(page))
2530 return 1;
2531
2532 /*
2533 * NOTE! This will make us return with VM_FAULT_RETRY, but with
2534 * the mmap_lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
2535 * is supposed to work. We have way too many special cases..
2536 */
2537 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
2538 return 0;
2539
2540 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
2541 if (vmf->flags & FAULT_FLAG_KILLABLE) {
2542 if (__lock_page_killable(page)) {
2543 /*
2544 * We didn't have the right flags to drop the mmap_lock,
2545 * but all fault_handlers only check for fatal signals
2546 * if we return VM_FAULT_RETRY, so we need to drop the
2547 * mmap_lock here and return 0 if we don't have a fpin.
2548 */
2549 if (*fpin == NULL)
2550 mmap_read_unlock(vmf->vma->vm_mm);
2551 return 0;
2552 }
2553 } else
2554 __lock_page(page);
2555 return 1;
2556}
2557
2558
2559/*
2560 * Synchronous readahead happens when we don't even find a page in the page
2561 * cache at all. We don't want to perform IO under the mmap sem, so if we have
2562 * to drop the mmap sem we return the file that was pinned in order for us to do
2563 * that. If we didn't pin a file then we return NULL. The file that is
2564 * returned needs to be fput()'ed when we're done with it.
2565 */
2566static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
2567{
2568 struct file *file = vmf->vma->vm_file;
2569 struct file_ra_state *ra = &file->f_ra;
2570 struct address_space *mapping = file->f_mapping;
2571 struct file *fpin = NULL;
2572 pgoff_t offset = vmf->pgoff;
2573 unsigned int mmap_miss;
2574
2575 /* If we don't want any read-ahead, don't bother */
2576 if (vmf->vma->vm_flags & VM_RAND_READ)
2577 return fpin;
2578 if (!ra->ra_pages)
2579 return fpin;
2580
2581 if (vmf->vma->vm_flags & VM_SEQ_READ) {
2582 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2583 page_cache_sync_readahead(mapping, ra, file, offset,
2584 ra->ra_pages);
2585 return fpin;
2586 }
2587
2588 /* Avoid banging the cache line if not needed */
2589 mmap_miss = READ_ONCE(ra->mmap_miss);
2590 if (mmap_miss < MMAP_LOTSAMISS * 10)
2591 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
2592
2593 /*
2594 * Do we miss much more than hit in this file? If so,
2595 * stop bothering with read-ahead. It will only hurt.
2596 */
2597 if (mmap_miss > MMAP_LOTSAMISS)
2598 return fpin;
2599
2600 /*
2601 * mmap read-around
2602 */
2603 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2604 ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
2605 ra->size = ra->ra_pages;
2606 ra->async_size = ra->ra_pages / 4;
2607 ra_submit(ra, mapping, file);
2608 return fpin;
2609}
2610
2611/*
2612 * Asynchronous readahead happens when we find the page and PG_readahead,
2613 * so we want to possibly extend the readahead further. We return the file that
2614 * was pinned if we have to drop the mmap_lock in order to do IO.
2615 */
2616static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
2617 struct page *page)
2618{
2619 struct file *file = vmf->vma->vm_file;
2620 struct file_ra_state *ra = &file->f_ra;
2621 struct address_space *mapping = file->f_mapping;
2622 struct file *fpin = NULL;
2623 unsigned int mmap_miss;
2624 pgoff_t offset = vmf->pgoff;
2625
2626 /* If we don't want any read-ahead, don't bother */
2627 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
2628 return fpin;
2629 mmap_miss = READ_ONCE(ra->mmap_miss);
2630 if (mmap_miss)
2631 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
2632 if (PageReadahead(page)) {
2633 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2634 page_cache_async_readahead(mapping, ra, file,
2635 page, offset, ra->ra_pages);
2636 }
2637 return fpin;
2638}
2639
2640/**
2641 * filemap_fault - read in file data for page fault handling
2642 * @vmf: struct vm_fault containing details of the fault
2643 *
2644 * filemap_fault() is invoked via the vma operations vector for a
2645 * mapped memory region to read in file data during a page fault.
2646 *
2647 * The goto's are kind of ugly, but this streamlines the normal case of having
2648 * it in the page cache, and handles the special cases reasonably without
2649 * having a lot of duplicated code.
2650 *
2651 * vma->vm_mm->mmap_lock must be held on entry.
2652 *
2653 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
2654 * may be dropped before doing I/O or by lock_page_maybe_drop_mmap().
2655 *
2656 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
2657 * has not been released.
2658 *
2659 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2660 *
2661 * Return: bitwise-OR of %VM_FAULT_ codes.
2662 */
2663vm_fault_t filemap_fault(struct vm_fault *vmf)
2664{
2665 int error;
2666 struct file *file = vmf->vma->vm_file;
2667 struct file *fpin = NULL;
2668 struct address_space *mapping = file->f_mapping;
2669 struct file_ra_state *ra = &file->f_ra;
2670 struct inode *inode = mapping->host;
2671 pgoff_t offset = vmf->pgoff;
2672 pgoff_t max_off;
2673 struct page *page;
2674 vm_fault_t ret = 0;
2675
2676 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2677 if (unlikely(offset >= max_off))
2678 return VM_FAULT_SIGBUS;
2679
2680 /*
2681 * Do we have something in the page cache already?
2682 */
2683 page = find_get_page(mapping, offset);
2684 if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2685 /*
2686 * We found the page, so try async readahead before
2687 * waiting for the lock.
2688 */
2689 fpin = do_async_mmap_readahead(vmf, page);
2690 } else if (!page) {
2691 /* No page in the page cache at all */
2692 count_vm_event(PGMAJFAULT);
2693 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
2694 ret = VM_FAULT_MAJOR;
2695 fpin = do_sync_mmap_readahead(vmf);
2696retry_find:
2697 page = pagecache_get_page(mapping, offset,
2698 FGP_CREAT|FGP_FOR_MMAP,
2699 vmf->gfp_mask);
2700 if (!page) {
2701 if (fpin)
2702 goto out_retry;
2703 return VM_FAULT_OOM;
2704 }
2705 }
2706
2707 if (!lock_page_maybe_drop_mmap(vmf, page, &fpin))
2708 goto out_retry;
2709
2710 /* Did it get truncated? */
2711 if (unlikely(compound_head(page)->mapping != mapping)) {
2712 unlock_page(page);
2713 put_page(page);
2714 goto retry_find;
2715 }
2716 VM_BUG_ON_PAGE(page_to_pgoff(page) != offset, page);
2717
2718 /*
2719 * We have a locked page in the page cache, now we need to check
2720 * that it's up-to-date. If not, it is going to be due to an error.
2721 */
2722 if (unlikely(!PageUptodate(page)))
2723 goto page_not_uptodate;
2724
2725 /*
2726 * We've made it this far and we had to drop our mmap_lock, now is the
2727 * time to return to the upper layer and have it re-find the vma and
2728 * redo the fault.
2729 */
2730 if (fpin) {
2731 unlock_page(page);
2732 goto out_retry;
2733 }
2734
2735 /*
2736 * Found the page and have a reference on it.
2737 * We must recheck i_size under page lock.
2738 */
2739 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2740 if (unlikely(offset >= max_off)) {
2741 unlock_page(page);
2742 put_page(page);
2743 return VM_FAULT_SIGBUS;
2744 }
2745
2746 vmf->page = page;
2747 return ret | VM_FAULT_LOCKED;
2748
2749page_not_uptodate:
2750 /*
2751 * Umm, take care of errors if the page isn't up-to-date.
2752 * Try to re-read it _once_. We do this synchronously,
2753 * because there really aren't any performance issues here
2754 * and we need to check for errors.
2755 */
2756 ClearPageError(page);
2757 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
2758 error = mapping->a_ops->readpage(file, page);
2759 if (!error) {
2760 wait_on_page_locked(page);
2761 if (!PageUptodate(page))
2762 error = -EIO;
2763 }
2764 if (fpin)
2765 goto out_retry;
2766 put_page(page);
2767
2768 if (!error || error == AOP_TRUNCATED_PAGE)
2769 goto retry_find;
2770
2771 shrink_readahead_size_eio(ra);
2772 return VM_FAULT_SIGBUS;
2773
2774out_retry:
2775 /*
2776 * We dropped the mmap_lock, we need to return to the fault handler to
2777 * re-find the vma and come back and find our hopefully still populated
2778 * page.
2779 */
2780 if (page)
2781 put_page(page);
2782 if (fpin)
2783 fput(fpin);
2784 return ret | VM_FAULT_RETRY;
2785}
2786EXPORT_SYMBOL(filemap_fault);
2787
2788void filemap_map_pages(struct vm_fault *vmf,
2789 pgoff_t start_pgoff, pgoff_t end_pgoff)
2790{
2791 struct file *file = vmf->vma->vm_file;
2792 struct address_space *mapping = file->f_mapping;
2793 pgoff_t last_pgoff = start_pgoff;
2794 unsigned long max_idx;
2795 XA_STATE(xas, &mapping->i_pages, start_pgoff);
2796 struct page *page;
2797 unsigned int mmap_miss = READ_ONCE(file->f_ra.mmap_miss);
2798
2799 rcu_read_lock();
2800 xas_for_each(&xas, page, end_pgoff) {
2801 if (xas_retry(&xas, page))
2802 continue;
2803 if (xa_is_value(page))
2804 goto next;
2805
2806 /*
2807 * Check for a locked page first, as a speculative
2808 * reference may adversely influence page migration.
2809 */
2810 if (PageLocked(page))
2811 goto next;
2812 if (!page_cache_get_speculative(page))
2813 goto next;
2814
2815 /* Has the page moved or been split? */
2816 if (unlikely(page != xas_reload(&xas)))
2817 goto skip;
2818 page = find_subpage(page, xas.xa_index);
2819
2820 if (!PageUptodate(page) ||
2821 PageReadahead(page) ||
2822 PageHWPoison(page))
2823 goto skip;
2824 if (!trylock_page(page))
2825 goto skip;
2826
2827 if (page->mapping != mapping || !PageUptodate(page))
2828 goto unlock;
2829
2830 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2831 if (page->index >= max_idx)
2832 goto unlock;
2833
2834 if (mmap_miss > 0)
2835 mmap_miss--;
2836
2837 vmf->address += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
2838 if (vmf->pte)
2839 vmf->pte += xas.xa_index - last_pgoff;
2840 last_pgoff = xas.xa_index;
2841 if (alloc_set_pte(vmf, page))
2842 goto unlock;
2843 unlock_page(page);
2844 goto next;
2845unlock:
2846 unlock_page(page);
2847skip:
2848 put_page(page);
2849next:
2850 /* Huge page is mapped? No need to proceed. */
2851 if (pmd_trans_huge(*vmf->pmd))
2852 break;
2853 }
2854 rcu_read_unlock();
2855 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss);
2856}
2857EXPORT_SYMBOL(filemap_map_pages);
2858
2859vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2860{
2861 struct page *page = vmf->page;
2862 struct inode *inode = file_inode(vmf->vma->vm_file);
2863 vm_fault_t ret = VM_FAULT_LOCKED;
2864
2865 sb_start_pagefault(inode->i_sb);
2866 file_update_time(vmf->vma->vm_file);
2867 lock_page(page);
2868 if (page->mapping != inode->i_mapping) {
2869 unlock_page(page);
2870 ret = VM_FAULT_NOPAGE;
2871 goto out;
2872 }
2873 /*
2874 * We mark the page dirty already here so that when freeze is in
2875 * progress, we are guaranteed that writeback during freezing will
2876 * see the dirty page and writeprotect it again.
2877 */
2878 set_page_dirty(page);
2879 wait_for_stable_page(page);
2880out:
2881 sb_end_pagefault(inode->i_sb);
2882 return ret;
2883}
2884
2885const struct vm_operations_struct generic_file_vm_ops = {
2886 .fault = filemap_fault,
2887 .map_pages = filemap_map_pages,
2888 .page_mkwrite = filemap_page_mkwrite,
2889};
2890
2891/* This is used for a general mmap of a disk file */
2892
2893int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2894{
2895 struct address_space *mapping = file->f_mapping;
2896
2897 if (!mapping->a_ops->readpage)
2898 return -ENOEXEC;
2899 file_accessed(file);
2900 vma->vm_ops = &generic_file_vm_ops;
2901 return 0;
2902}
2903
2904/*
2905 * This is for filesystems which do not implement ->writepage.
2906 */
2907int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2908{
2909 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2910 return -EINVAL;
2911 return generic_file_mmap(file, vma);
2912}
2913#else
2914vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
2915{
2916 return VM_FAULT_SIGBUS;
2917}
2918int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2919{
2920 return -ENOSYS;
2921}
2922int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2923{
2924 return -ENOSYS;
2925}
2926#endif /* CONFIG_MMU */
2927
2928EXPORT_SYMBOL(filemap_page_mkwrite);
2929EXPORT_SYMBOL(generic_file_mmap);
2930EXPORT_SYMBOL(generic_file_readonly_mmap);
2931
2932static struct page *wait_on_page_read(struct page *page)
2933{
2934 if (!IS_ERR(page)) {
2935 wait_on_page_locked(page);
2936 if (!PageUptodate(page)) {
2937 put_page(page);
2938 page = ERR_PTR(-EIO);
2939 }
2940 }
2941 return page;
2942}
2943
2944static struct page *do_read_cache_page(struct address_space *mapping,
2945 pgoff_t index,
2946 int (*filler)(void *, struct page *),
2947 void *data,
2948 gfp_t gfp)
2949{
2950 struct page *page;
2951 int err;
2952repeat:
2953 page = find_get_page(mapping, index);
2954 if (!page) {
2955 page = __page_cache_alloc(gfp);
2956 if (!page)
2957 return ERR_PTR(-ENOMEM);
2958 err = add_to_page_cache_lru(page, mapping, index, gfp);
2959 if (unlikely(err)) {
2960 put_page(page);
2961 if (err == -EEXIST)
2962 goto repeat;
2963 /* Presumably ENOMEM for xarray node */
2964 return ERR_PTR(err);
2965 }
2966
2967filler:
2968 if (filler)
2969 err = filler(data, page);
2970 else
2971 err = mapping->a_ops->readpage(data, page);
2972
2973 if (err < 0) {
2974 put_page(page);
2975 return ERR_PTR(err);
2976 }
2977
2978 page = wait_on_page_read(page);
2979 if (IS_ERR(page))
2980 return page;
2981 goto out;
2982 }
2983 if (PageUptodate(page))
2984 goto out;
2985
2986 /*
2987 * Page is not up to date and may be locked due one of the following
2988 * case a: Page is being filled and the page lock is held
2989 * case b: Read/write error clearing the page uptodate status
2990 * case c: Truncation in progress (page locked)
2991 * case d: Reclaim in progress
2992 *
2993 * Case a, the page will be up to date when the page is unlocked.
2994 * There is no need to serialise on the page lock here as the page
2995 * is pinned so the lock gives no additional protection. Even if the
2996 * page is truncated, the data is still valid if PageUptodate as
2997 * it's a race vs truncate race.
2998 * Case b, the page will not be up to date
2999 * Case c, the page may be truncated but in itself, the data may still
3000 * be valid after IO completes as it's a read vs truncate race. The
3001 * operation must restart if the page is not uptodate on unlock but
3002 * otherwise serialising on page lock to stabilise the mapping gives
3003 * no additional guarantees to the caller as the page lock is
3004 * released before return.
3005 * Case d, similar to truncation. If reclaim holds the page lock, it
3006 * will be a race with remove_mapping that determines if the mapping
3007 * is valid on unlock but otherwise the data is valid and there is
3008 * no need to serialise with page lock.
3009 *
3010 * As the page lock gives no additional guarantee, we optimistically
3011 * wait on the page to be unlocked and check if it's up to date and
3012 * use the page if it is. Otherwise, the page lock is required to
3013 * distinguish between the different cases. The motivation is that we
3014 * avoid spurious serialisations and wakeups when multiple processes
3015 * wait on the same page for IO to complete.
3016 */
3017 wait_on_page_locked(page);
3018 if (PageUptodate(page))
3019 goto out;
3020
3021 /* Distinguish between all the cases under the safety of the lock */
3022 lock_page(page);
3023
3024 /* Case c or d, restart the operation */
3025 if (!page->mapping) {
3026 unlock_page(page);
3027 put_page(page);
3028 goto repeat;
3029 }
3030
3031 /* Someone else locked and filled the page in a very small window */
3032 if (PageUptodate(page)) {
3033 unlock_page(page);
3034 goto out;
3035 }
3036
3037 /*
3038 * A previous I/O error may have been due to temporary
3039 * failures.
3040 * Clear page error before actual read, PG_error will be
3041 * set again if read page fails.
3042 */
3043 ClearPageError(page);
3044 goto filler;
3045
3046out:
3047 mark_page_accessed(page);
3048 return page;
3049}
3050
3051/**
3052 * read_cache_page - read into page cache, fill it if needed
3053 * @mapping: the page's address_space
3054 * @index: the page index
3055 * @filler: function to perform the read
3056 * @data: first arg to filler(data, page) function, often left as NULL
3057 *
3058 * Read into the page cache. If a page already exists, and PageUptodate() is
3059 * not set, try to fill the page and wait for it to become unlocked.
3060 *
3061 * If the page does not get brought uptodate, return -EIO.
3062 *
3063 * Return: up to date page on success, ERR_PTR() on failure.
3064 */
3065struct page *read_cache_page(struct address_space *mapping,
3066 pgoff_t index,
3067 int (*filler)(void *, struct page *),
3068 void *data)
3069{
3070 return do_read_cache_page(mapping, index, filler, data,
3071 mapping_gfp_mask(mapping));
3072}
3073EXPORT_SYMBOL(read_cache_page);
3074
3075/**
3076 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3077 * @mapping: the page's address_space
3078 * @index: the page index
3079 * @gfp: the page allocator flags to use if allocating
3080 *
3081 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3082 * any new page allocations done using the specified allocation flags.
3083 *
3084 * If the page does not get brought uptodate, return -EIO.
3085 *
3086 * Return: up to date page on success, ERR_PTR() on failure.
3087 */
3088struct page *read_cache_page_gfp(struct address_space *mapping,
3089 pgoff_t index,
3090 gfp_t gfp)
3091{
3092 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3093}
3094EXPORT_SYMBOL(read_cache_page_gfp);
3095
3096/*
3097 * Don't operate on ranges the page cache doesn't support, and don't exceed the
3098 * LFS limits. If pos is under the limit it becomes a short access. If it
3099 * exceeds the limit we return -EFBIG.
3100 */
3101static int generic_write_check_limits(struct file *file, loff_t pos,
3102 loff_t *count)
3103{
3104 struct inode *inode = file->f_mapping->host;
3105 loff_t max_size = inode->i_sb->s_maxbytes;
3106 loff_t limit = rlimit(RLIMIT_FSIZE);
3107
3108 if (limit != RLIM_INFINITY) {
3109 if (pos >= limit) {
3110 send_sig(SIGXFSZ, current, 0);
3111 return -EFBIG;
3112 }
3113 *count = min(*count, limit - pos);
3114 }
3115
3116 if (!(file->f_flags & O_LARGEFILE))
3117 max_size = MAX_NON_LFS;
3118
3119 if (unlikely(pos >= max_size))
3120 return -EFBIG;
3121
3122 *count = min(*count, max_size - pos);
3123
3124 return 0;
3125}
3126
3127/*
3128 * Performs necessary checks before doing a write
3129 *
3130 * Can adjust writing position or amount of bytes to write.
3131 * Returns appropriate error code that caller should return or
3132 * zero in case that write should be allowed.
3133 */
3134inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
3135{
3136 struct file *file = iocb->ki_filp;
3137 struct inode *inode = file->f_mapping->host;
3138 loff_t count;
3139 int ret;
3140
3141 if (IS_SWAPFILE(inode))
3142 return -ETXTBSY;
3143
3144 if (!iov_iter_count(from))
3145 return 0;
3146
3147 /* FIXME: this is for backwards compatibility with 2.4 */
3148 if (iocb->ki_flags & IOCB_APPEND)
3149 iocb->ki_pos = i_size_read(inode);
3150
3151 if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
3152 return -EINVAL;
3153
3154 count = iov_iter_count(from);
3155 ret = generic_write_check_limits(file, iocb->ki_pos, &count);
3156 if (ret)
3157 return ret;
3158
3159 iov_iter_truncate(from, count);
3160 return iov_iter_count(from);
3161}
3162EXPORT_SYMBOL(generic_write_checks);
3163
3164/*
3165 * Performs necessary checks before doing a clone.
3166 *
3167 * Can adjust amount of bytes to clone via @req_count argument.
3168 * Returns appropriate error code that caller should return or
3169 * zero in case the clone should be allowed.
3170 */
3171int generic_remap_checks(struct file *file_in, loff_t pos_in,
3172 struct file *file_out, loff_t pos_out,
3173 loff_t *req_count, unsigned int remap_flags)
3174{
3175 struct inode *inode_in = file_in->f_mapping->host;
3176 struct inode *inode_out = file_out->f_mapping->host;
3177 uint64_t count = *req_count;
3178 uint64_t bcount;
3179 loff_t size_in, size_out;
3180 loff_t bs = inode_out->i_sb->s_blocksize;
3181 int ret;
3182
3183 /* The start of both ranges must be aligned to an fs block. */
3184 if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_out, bs))
3185 return -EINVAL;
3186
3187 /* Ensure offsets don't wrap. */
3188 if (pos_in + count < pos_in || pos_out + count < pos_out)
3189 return -EINVAL;
3190
3191 size_in = i_size_read(inode_in);
3192 size_out = i_size_read(inode_out);
3193
3194 /* Dedupe requires both ranges to be within EOF. */
3195 if ((remap_flags & REMAP_FILE_DEDUP) &&
3196 (pos_in >= size_in || pos_in + count > size_in ||
3197 pos_out >= size_out || pos_out + count > size_out))
3198 return -EINVAL;
3199
3200 /* Ensure the infile range is within the infile. */
3201 if (pos_in >= size_in)
3202 return -EINVAL;
3203 count = min(count, size_in - (uint64_t)pos_in);
3204
3205 ret = generic_write_check_limits(file_out, pos_out, &count);
3206 if (ret)
3207 return ret;
3208
3209 /*
3210 * If the user wanted us to link to the infile's EOF, round up to the
3211 * next block boundary for this check.
3212 *
3213 * Otherwise, make sure the count is also block-aligned, having
3214 * already confirmed the starting offsets' block alignment.
3215 */
3216 if (pos_in + count == size_in) {
3217 bcount = ALIGN(size_in, bs) - pos_in;
3218 } else {
3219 if (!IS_ALIGNED(count, bs))
3220 count = ALIGN_DOWN(count, bs);
3221 bcount = count;
3222 }
3223
3224 /* Don't allow overlapped cloning within the same file. */
3225 if (inode_in == inode_out &&
3226 pos_out + bcount > pos_in &&
3227 pos_out < pos_in + bcount)
3228 return -EINVAL;
3229
3230 /*
3231 * We shortened the request but the caller can't deal with that, so
3232 * bounce the request back to userspace.
3233 */
3234 if (*req_count != count && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
3235 return -EINVAL;
3236
3237 *req_count = count;
3238 return 0;
3239}
3240
3241
3242/*
3243 * Performs common checks before doing a file copy/clone
3244 * from @file_in to @file_out.
3245 */
3246int generic_file_rw_checks(struct file *file_in, struct file *file_out)
3247{
3248 struct inode *inode_in = file_inode(file_in);
3249 struct inode *inode_out = file_inode(file_out);
3250
3251 /* Don't copy dirs, pipes, sockets... */
3252 if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode))
3253 return -EISDIR;
3254 if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode))
3255 return -EINVAL;
3256
3257 if (!(file_in->f_mode & FMODE_READ) ||
3258 !(file_out->f_mode & FMODE_WRITE) ||
3259 (file_out->f_flags & O_APPEND))
3260 return -EBADF;
3261
3262 return 0;
3263}
3264
3265/*
3266 * Performs necessary checks before doing a file copy
3267 *
3268 * Can adjust amount of bytes to copy via @req_count argument.
3269 * Returns appropriate error code that caller should return or
3270 * zero in case the copy should be allowed.
3271 */
3272int generic_copy_file_checks(struct file *file_in, loff_t pos_in,
3273 struct file *file_out, loff_t pos_out,
3274 size_t *req_count, unsigned int flags)
3275{
3276 struct inode *inode_in = file_inode(file_in);
3277 struct inode *inode_out = file_inode(file_out);
3278 uint64_t count = *req_count;
3279 loff_t size_in;
3280 int ret;
3281
3282 ret = generic_file_rw_checks(file_in, file_out);
3283 if (ret)
3284 return ret;
3285
3286 /* Don't touch certain kinds of inodes */
3287 if (IS_IMMUTABLE(inode_out))
3288 return -EPERM;
3289
3290 if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out))
3291 return -ETXTBSY;
3292
3293 /* Ensure offsets don't wrap. */
3294 if (pos_in + count < pos_in || pos_out + count < pos_out)
3295 return -EOVERFLOW;
3296
3297 /* Shorten the copy to EOF */
3298 size_in = i_size_read(inode_in);
3299 if (pos_in >= size_in)
3300 count = 0;
3301 else
3302 count = min(count, size_in - (uint64_t)pos_in);
3303
3304 ret = generic_write_check_limits(file_out, pos_out, &count);
3305 if (ret)
3306 return ret;
3307
3308 /* Don't allow overlapped copying within the same file. */
3309 if (inode_in == inode_out &&
3310 pos_out + count > pos_in &&
3311 pos_out < pos_in + count)
3312 return -EINVAL;
3313
3314 *req_count = count;
3315 return 0;
3316}
3317
3318int pagecache_write_begin(struct file *file, struct address_space *mapping,
3319 loff_t pos, unsigned len, unsigned flags,
3320 struct page **pagep, void **fsdata)
3321{
3322 const struct address_space_operations *aops = mapping->a_ops;
3323
3324 return aops->write_begin(file, mapping, pos, len, flags,
3325 pagep, fsdata);
3326}
3327EXPORT_SYMBOL(pagecache_write_begin);
3328
3329int pagecache_write_end(struct file *file, struct address_space *mapping,
3330 loff_t pos, unsigned len, unsigned copied,
3331 struct page *page, void *fsdata)
3332{
3333 const struct address_space_operations *aops = mapping->a_ops;
3334
3335 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
3336}
3337EXPORT_SYMBOL(pagecache_write_end);
3338
3339/*
3340 * Warn about a page cache invalidation failure during a direct I/O write.
3341 */
3342void dio_warn_stale_pagecache(struct file *filp)
3343{
3344 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3345 char pathname[128];
3346 struct inode *inode = file_inode(filp);
3347 char *path;
3348
3349 errseq_set(&inode->i_mapping->wb_err, -EIO);
3350 if (__ratelimit(&_rs)) {
3351 path = file_path(filp, pathname, sizeof(pathname));
3352 if (IS_ERR(path))
3353 path = "(unknown)";
3354 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3355 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3356 current->comm);
3357 }
3358}
3359
3360ssize_t
3361generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3362{
3363 struct file *file = iocb->ki_filp;
3364 struct address_space *mapping = file->f_mapping;
3365 struct inode *inode = mapping->host;
3366 loff_t pos = iocb->ki_pos;
3367 ssize_t written;
3368 size_t write_len;
3369 pgoff_t end;
3370
3371 write_len = iov_iter_count(from);
3372 end = (pos + write_len - 1) >> PAGE_SHIFT;
3373
3374 if (iocb->ki_flags & IOCB_NOWAIT) {
3375 /* If there are pages to writeback, return */
3376 if (filemap_range_has_page(inode->i_mapping, pos,
3377 pos + write_len - 1))
3378 return -EAGAIN;
3379 } else {
3380 written = filemap_write_and_wait_range(mapping, pos,
3381 pos + write_len - 1);
3382 if (written)
3383 goto out;
3384 }
3385
3386 /*
3387 * After a write we want buffered reads to be sure to go to disk to get
3388 * the new data. We invalidate clean cached page from the region we're
3389 * about to write. We do this *before* the write so that we can return
3390 * without clobbering -EIOCBQUEUED from ->direct_IO().
3391 */
3392 written = invalidate_inode_pages2_range(mapping,
3393 pos >> PAGE_SHIFT, end);
3394 /*
3395 * If a page can not be invalidated, return 0 to fall back
3396 * to buffered write.
3397 */
3398 if (written) {
3399 if (written == -EBUSY)
3400 return 0;
3401 goto out;
3402 }
3403
3404 written = mapping->a_ops->direct_IO(iocb, from);
3405
3406 /*
3407 * Finally, try again to invalidate clean pages which might have been
3408 * cached by non-direct readahead, or faulted in by get_user_pages()
3409 * if the source of the write was an mmap'ed region of the file
3410 * we're writing. Either one is a pretty crazy thing to do,
3411 * so we don't support it 100%. If this invalidation
3412 * fails, tough, the write still worked...
3413 *
3414 * Most of the time we do not need this since dio_complete() will do
3415 * the invalidation for us. However there are some file systems that
3416 * do not end up with dio_complete() being called, so let's not break
3417 * them by removing it completely.
3418 *
3419 * Noticeable example is a blkdev_direct_IO().
3420 *
3421 * Skip invalidation for async writes or if mapping has no pages.
3422 */
3423 if (written > 0 && mapping->nrpages &&
3424 invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT, end))
3425 dio_warn_stale_pagecache(file);
3426
3427 if (written > 0) {
3428 pos += written;
3429 write_len -= written;
3430 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3431 i_size_write(inode, pos);
3432 mark_inode_dirty(inode);
3433 }
3434 iocb->ki_pos = pos;
3435 }
3436 iov_iter_revert(from, write_len - iov_iter_count(from));
3437out:
3438 return written;
3439}
3440EXPORT_SYMBOL(generic_file_direct_write);
3441
3442/*
3443 * Find or create a page at the given pagecache position. Return the locked
3444 * page. This function is specifically for buffered writes.
3445 */
3446struct page *grab_cache_page_write_begin(struct address_space *mapping,
3447 pgoff_t index, unsigned flags)
3448{
3449 struct page *page;
3450 int fgp_flags = FGP_LOCK|FGP_WRITE|FGP_CREAT;
3451
3452 if (flags & AOP_FLAG_NOFS)
3453 fgp_flags |= FGP_NOFS;
3454
3455 page = pagecache_get_page(mapping, index, fgp_flags,
3456 mapping_gfp_mask(mapping));
3457 if (page)
3458 wait_for_stable_page(page);
3459
3460 return page;
3461}
3462EXPORT_SYMBOL(grab_cache_page_write_begin);
3463
3464ssize_t generic_perform_write(struct file *file,
3465 struct iov_iter *i, loff_t pos)
3466{
3467 struct address_space *mapping = file->f_mapping;
3468 const struct address_space_operations *a_ops = mapping->a_ops;
3469 long status = 0;
3470 ssize_t written = 0;
3471 unsigned int flags = 0;
3472
3473 do {
3474 struct page *page;
3475 unsigned long offset; /* Offset into pagecache page */
3476 unsigned long bytes; /* Bytes to write to page */
3477 size_t copied; /* Bytes copied from user */
3478 void *fsdata;
3479
3480 offset = (pos & (PAGE_SIZE - 1));
3481 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3482 iov_iter_count(i));
3483
3484again:
3485 /*
3486 * Bring in the user page that we will copy from _first_.
3487 * Otherwise there's a nasty deadlock on copying from the
3488 * same page as we're writing to, without it being marked
3489 * up-to-date.
3490 *
3491 * Not only is this an optimisation, but it is also required
3492 * to check that the address is actually valid, when atomic
3493 * usercopies are used, below.
3494 */
3495 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
3496 status = -EFAULT;
3497 break;
3498 }
3499
3500 if (fatal_signal_pending(current)) {
3501 status = -EINTR;
3502 break;
3503 }
3504
3505 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
3506 &page, &fsdata);
3507 if (unlikely(status < 0))
3508 break;
3509
3510 if (mapping_writably_mapped(mapping))
3511 flush_dcache_page(page);
3512
3513 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
3514 flush_dcache_page(page);
3515
3516 status = a_ops->write_end(file, mapping, pos, bytes, copied,
3517 page, fsdata);
3518 if (unlikely(status < 0))
3519 break;
3520 copied = status;
3521
3522 cond_resched();
3523
3524 iov_iter_advance(i, copied);
3525 if (unlikely(copied == 0)) {
3526 /*
3527 * If we were unable to copy any data at all, we must
3528 * fall back to a single segment length write.
3529 *
3530 * If we didn't fallback here, we could livelock
3531 * because not all segments in the iov can be copied at
3532 * once without a pagefault.
3533 */
3534 bytes = min_t(unsigned long, PAGE_SIZE - offset,
3535 iov_iter_single_seg_count(i));
3536 goto again;
3537 }
3538 pos += copied;
3539 written += copied;
3540
3541 balance_dirty_pages_ratelimited(mapping);
3542 } while (iov_iter_count(i));
3543
3544 return written ? written : status;
3545}
3546EXPORT_SYMBOL(generic_perform_write);
3547
3548/**
3549 * __generic_file_write_iter - write data to a file
3550 * @iocb: IO state structure (file, offset, etc.)
3551 * @from: iov_iter with data to write
3552 *
3553 * This function does all the work needed for actually writing data to a
3554 * file. It does all basic checks, removes SUID from the file, updates
3555 * modification times and calls proper subroutines depending on whether we
3556 * do direct IO or a standard buffered write.
3557 *
3558 * It expects i_mutex to be grabbed unless we work on a block device or similar
3559 * object which does not need locking at all.
3560 *
3561 * This function does *not* take care of syncing data in case of O_SYNC write.
3562 * A caller has to handle it. This is mainly due to the fact that we want to
3563 * avoid syncing under i_mutex.
3564 *
3565 * Return:
3566 * * number of bytes written, even for truncated writes
3567 * * negative error code if no data has been written at all
3568 */
3569ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3570{
3571 struct file *file = iocb->ki_filp;
3572 struct address_space * mapping = file->f_mapping;
3573 struct inode *inode = mapping->host;
3574 ssize_t written = 0;
3575 ssize_t err;
3576 ssize_t status;
3577
3578 /* We can write back this queue in page reclaim */
3579 current->backing_dev_info = inode_to_bdi(inode);
3580 err = file_remove_privs(file);
3581 if (err)
3582 goto out;
3583
3584 err = file_update_time(file);
3585 if (err)
3586 goto out;
3587
3588 if (iocb->ki_flags & IOCB_DIRECT) {
3589 loff_t pos, endbyte;
3590
3591 written = generic_file_direct_write(iocb, from);
3592 /*
3593 * If the write stopped short of completing, fall back to
3594 * buffered writes. Some filesystems do this for writes to
3595 * holes, for example. For DAX files, a buffered write will
3596 * not succeed (even if it did, DAX does not handle dirty
3597 * page-cache pages correctly).
3598 */
3599 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
3600 goto out;
3601
3602 status = generic_perform_write(file, from, pos = iocb->ki_pos);
3603 /*
3604 * If generic_perform_write() returned a synchronous error
3605 * then we want to return the number of bytes which were
3606 * direct-written, or the error code if that was zero. Note
3607 * that this differs from normal direct-io semantics, which
3608 * will return -EFOO even if some bytes were written.
3609 */
3610 if (unlikely(status < 0)) {
3611 err = status;
3612 goto out;
3613 }
3614 /*
3615 * We need to ensure that the page cache pages are written to
3616 * disk and invalidated to preserve the expected O_DIRECT
3617 * semantics.
3618 */
3619 endbyte = pos + status - 1;
3620 err = filemap_write_and_wait_range(mapping, pos, endbyte);
3621 if (err == 0) {
3622 iocb->ki_pos = endbyte + 1;
3623 written += status;
3624 invalidate_mapping_pages(mapping,
3625 pos >> PAGE_SHIFT,
3626 endbyte >> PAGE_SHIFT);
3627 } else {
3628 /*
3629 * We don't know how much we wrote, so just return
3630 * the number of bytes which were direct-written
3631 */
3632 }
3633 } else {
3634 written = generic_perform_write(file, from, iocb->ki_pos);
3635 if (likely(written > 0))
3636 iocb->ki_pos += written;
3637 }
3638out:
3639 current->backing_dev_info = NULL;
3640 return written ? written : err;
3641}
3642EXPORT_SYMBOL(__generic_file_write_iter);
3643
3644/**
3645 * generic_file_write_iter - write data to a file
3646 * @iocb: IO state structure
3647 * @from: iov_iter with data to write
3648 *
3649 * This is a wrapper around __generic_file_write_iter() to be used by most
3650 * filesystems. It takes care of syncing the file in case of O_SYNC file
3651 * and acquires i_mutex as needed.
3652 * Return:
3653 * * negative error code if no data has been written at all of
3654 * vfs_fsync_range() failed for a synchronous write
3655 * * number of bytes written, even for truncated writes
3656 */
3657ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
3658{
3659 struct file *file = iocb->ki_filp;
3660 struct inode *inode = file->f_mapping->host;
3661 ssize_t ret;
3662
3663 inode_lock(inode);
3664 ret = generic_write_checks(iocb, from);
3665 if (ret > 0)
3666 ret = __generic_file_write_iter(iocb, from);
3667 inode_unlock(inode);
3668
3669 if (ret > 0)
3670 ret = generic_write_sync(iocb, ret);
3671 return ret;
3672}
3673EXPORT_SYMBOL(generic_file_write_iter);
3674
3675/**
3676 * try_to_release_page() - release old fs-specific metadata on a page
3677 *
3678 * @page: the page which the kernel is trying to free
3679 * @gfp_mask: memory allocation flags (and I/O mode)
3680 *
3681 * The address_space is to try to release any data against the page
3682 * (presumably at page->private).
3683 *
3684 * This may also be called if PG_fscache is set on a page, indicating that the
3685 * page is known to the local caching routines.
3686 *
3687 * The @gfp_mask argument specifies whether I/O may be performed to release
3688 * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
3689 *
3690 * Return: %1 if the release was successful, otherwise return zero.
3691 */
3692int try_to_release_page(struct page *page, gfp_t gfp_mask)
3693{
3694 struct address_space * const mapping = page->mapping;
3695
3696 BUG_ON(!PageLocked(page));
3697 if (PageWriteback(page))
3698 return 0;
3699
3700 if (mapping && mapping->a_ops->releasepage)
3701 return mapping->a_ops->releasepage(page, gfp_mask);
3702 return try_to_free_buffers(page);
3703}
3704
3705EXPORT_SYMBOL(try_to_release_page);