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/module.h>
13#include <linux/compiler.h>
14#include <linux/fs.h>
15#include <linux/uaccess.h>
16#include <linux/aio.h>
17#include <linux/capability.h>
18#include <linux/kernel_stat.h>
19#include <linux/gfp.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/mman.h>
23#include <linux/pagemap.h>
24#include <linux/file.h>
25#include <linux/uio.h>
26#include <linux/hash.h>
27#include <linux/writeback.h>
28#include <linux/backing-dev.h>
29#include <linux/pagevec.h>
30#include <linux/blkdev.h>
31#include <linux/security.h>
32#include <linux/syscalls.h>
33#include <linux/cpuset.h>
34#include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35#include <linux/memcontrol.h>
36#include <linux/cleancache.h>
37#include "internal.h"
38
39/*
40 * FIXME: remove all knowledge of the buffer layer from the core VM
41 */
42#include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44#include <asm/mman.h>
45
46/*
47 * Shared mappings implemented 30.11.1994. It's not fully working yet,
48 * though.
49 *
50 * Shared mappings now work. 15.8.1995 Bruno.
51 *
52 * finished 'unifying' the page and buffer cache and SMP-threaded the
53 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54 *
55 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56 */
57
58/*
59 * Lock ordering:
60 *
61 * ->i_mmap_mutex (truncate_pagecache)
62 * ->private_lock (__free_pte->__set_page_dirty_buffers)
63 * ->swap_lock (exclusive_swap_page, others)
64 * ->mapping->tree_lock
65 *
66 * ->i_mutex
67 * ->i_mmap_mutex (truncate->unmap_mapping_range)
68 *
69 * ->mmap_sem
70 * ->i_mmap_mutex
71 * ->page_table_lock or pte_lock (various, mainly in memory.c)
72 * ->mapping->tree_lock (arch-dependent flush_dcache_mmap_lock)
73 *
74 * ->mmap_sem
75 * ->lock_page (access_process_vm)
76 *
77 * ->i_mutex (generic_file_buffered_write)
78 * ->mmap_sem (fault_in_pages_readable->do_page_fault)
79 *
80 * bdi->wb.list_lock
81 * sb_lock (fs/fs-writeback.c)
82 * ->mapping->tree_lock (__sync_single_inode)
83 *
84 * ->i_mmap_mutex
85 * ->anon_vma.lock (vma_adjust)
86 *
87 * ->anon_vma.lock
88 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
89 *
90 * ->page_table_lock or pte_lock
91 * ->swap_lock (try_to_unmap_one)
92 * ->private_lock (try_to_unmap_one)
93 * ->tree_lock (try_to_unmap_one)
94 * ->zone.lru_lock (follow_page->mark_page_accessed)
95 * ->zone.lru_lock (check_pte_range->isolate_lru_page)
96 * ->private_lock (page_remove_rmap->set_page_dirty)
97 * ->tree_lock (page_remove_rmap->set_page_dirty)
98 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty)
99 * ->inode->i_lock (page_remove_rmap->set_page_dirty)
100 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
101 * ->inode->i_lock (zap_pte_range->set_page_dirty)
102 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
103 *
104 * (code doesn't rely on that order, so you could switch it around)
105 * ->tasklist_lock (memory_failure, collect_procs_ao)
106 * ->i_mmap_mutex
107 */
108
109/*
110 * Delete a page from the page cache and free it. Caller has to make
111 * sure the page is locked and that nobody else uses it - or that usage
112 * is safe. The caller must hold the mapping's tree_lock.
113 */
114void __delete_from_page_cache(struct page *page)
115{
116 struct address_space *mapping = page->mapping;
117
118 /*
119 * if we're uptodate, flush out into the cleancache, otherwise
120 * invalidate any existing cleancache entries. We can't leave
121 * stale data around in the cleancache once our page is gone
122 */
123 if (PageUptodate(page) && PageMappedToDisk(page))
124 cleancache_put_page(page);
125 else
126 cleancache_flush_page(mapping, page);
127
128 radix_tree_delete(&mapping->page_tree, page->index);
129 page->mapping = NULL;
130 /* Leave page->index set: truncation lookup relies upon it */
131 mapping->nrpages--;
132 __dec_zone_page_state(page, NR_FILE_PAGES);
133 if (PageSwapBacked(page))
134 __dec_zone_page_state(page, NR_SHMEM);
135 BUG_ON(page_mapped(page));
136
137 /*
138 * Some filesystems seem to re-dirty the page even after
139 * the VM has canceled the dirty bit (eg ext3 journaling).
140 *
141 * Fix it up by doing a final dirty accounting check after
142 * having removed the page entirely.
143 */
144 if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
145 dec_zone_page_state(page, NR_FILE_DIRTY);
146 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
147 }
148}
149
150/**
151 * delete_from_page_cache - delete page from page cache
152 * @page: the page which the kernel is trying to remove from page cache
153 *
154 * This must be called only on pages that have been verified to be in the page
155 * cache and locked. It will never put the page into the free list, the caller
156 * has a reference on the page.
157 */
158void delete_from_page_cache(struct page *page)
159{
160 struct address_space *mapping = page->mapping;
161 void (*freepage)(struct page *);
162
163 BUG_ON(!PageLocked(page));
164
165 freepage = mapping->a_ops->freepage;
166 spin_lock_irq(&mapping->tree_lock);
167 __delete_from_page_cache(page);
168 spin_unlock_irq(&mapping->tree_lock);
169 mem_cgroup_uncharge_cache_page(page);
170
171 if (freepage)
172 freepage(page);
173 page_cache_release(page);
174}
175EXPORT_SYMBOL(delete_from_page_cache);
176
177static int sleep_on_page(void *word)
178{
179 io_schedule();
180 return 0;
181}
182
183static int sleep_on_page_killable(void *word)
184{
185 sleep_on_page(word);
186 return fatal_signal_pending(current) ? -EINTR : 0;
187}
188
189/**
190 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
191 * @mapping: address space structure to write
192 * @start: offset in bytes where the range starts
193 * @end: offset in bytes where the range ends (inclusive)
194 * @sync_mode: enable synchronous operation
195 *
196 * Start writeback against all of a mapping's dirty pages that lie
197 * within the byte offsets <start, end> inclusive.
198 *
199 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
200 * opposed to a regular memory cleansing writeback. The difference between
201 * these two operations is that if a dirty page/buffer is encountered, it must
202 * be waited upon, and not just skipped over.
203 */
204int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
205 loff_t end, int sync_mode)
206{
207 int ret;
208 struct writeback_control wbc = {
209 .sync_mode = sync_mode,
210 .nr_to_write = LONG_MAX,
211 .range_start = start,
212 .range_end = end,
213 };
214
215 if (!mapping_cap_writeback_dirty(mapping))
216 return 0;
217
218 ret = do_writepages(mapping, &wbc);
219 return ret;
220}
221
222static inline int __filemap_fdatawrite(struct address_space *mapping,
223 int sync_mode)
224{
225 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
226}
227
228int filemap_fdatawrite(struct address_space *mapping)
229{
230 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
231}
232EXPORT_SYMBOL(filemap_fdatawrite);
233
234int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
235 loff_t end)
236{
237 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
238}
239EXPORT_SYMBOL(filemap_fdatawrite_range);
240
241/**
242 * filemap_flush - mostly a non-blocking flush
243 * @mapping: target address_space
244 *
245 * This is a mostly non-blocking flush. Not suitable for data-integrity
246 * purposes - I/O may not be started against all dirty pages.
247 */
248int filemap_flush(struct address_space *mapping)
249{
250 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
251}
252EXPORT_SYMBOL(filemap_flush);
253
254/**
255 * filemap_fdatawait_range - wait for writeback to complete
256 * @mapping: address space structure to wait for
257 * @start_byte: offset in bytes where the range starts
258 * @end_byte: offset in bytes where the range ends (inclusive)
259 *
260 * Walk the list of under-writeback pages of the given address space
261 * in the given range and wait for all of them.
262 */
263int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
264 loff_t end_byte)
265{
266 pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
267 pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
268 struct pagevec pvec;
269 int nr_pages;
270 int ret = 0;
271
272 if (end_byte < start_byte)
273 return 0;
274
275 pagevec_init(&pvec, 0);
276 while ((index <= end) &&
277 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
278 PAGECACHE_TAG_WRITEBACK,
279 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
280 unsigned i;
281
282 for (i = 0; i < nr_pages; i++) {
283 struct page *page = pvec.pages[i];
284
285 /* until radix tree lookup accepts end_index */
286 if (page->index > end)
287 continue;
288
289 wait_on_page_writeback(page);
290 if (TestClearPageError(page))
291 ret = -EIO;
292 }
293 pagevec_release(&pvec);
294 cond_resched();
295 }
296
297 /* Check for outstanding write errors */
298 if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
299 ret = -ENOSPC;
300 if (test_and_clear_bit(AS_EIO, &mapping->flags))
301 ret = -EIO;
302
303 return ret;
304}
305EXPORT_SYMBOL(filemap_fdatawait_range);
306
307/**
308 * filemap_fdatawait - wait for all under-writeback pages to complete
309 * @mapping: address space structure to wait for
310 *
311 * Walk the list of under-writeback pages of the given address space
312 * and wait for all of them.
313 */
314int filemap_fdatawait(struct address_space *mapping)
315{
316 loff_t i_size = i_size_read(mapping->host);
317
318 if (i_size == 0)
319 return 0;
320
321 return filemap_fdatawait_range(mapping, 0, i_size - 1);
322}
323EXPORT_SYMBOL(filemap_fdatawait);
324
325int filemap_write_and_wait(struct address_space *mapping)
326{
327 int err = 0;
328
329 if (mapping->nrpages) {
330 err = filemap_fdatawrite(mapping);
331 /*
332 * Even if the above returned error, the pages may be
333 * written partially (e.g. -ENOSPC), so we wait for it.
334 * But the -EIO is special case, it may indicate the worst
335 * thing (e.g. bug) happened, so we avoid waiting for it.
336 */
337 if (err != -EIO) {
338 int err2 = filemap_fdatawait(mapping);
339 if (!err)
340 err = err2;
341 }
342 }
343 return err;
344}
345EXPORT_SYMBOL(filemap_write_and_wait);
346
347/**
348 * filemap_write_and_wait_range - write out & wait on a file range
349 * @mapping: the address_space for the pages
350 * @lstart: offset in bytes where the range starts
351 * @lend: offset in bytes where the range ends (inclusive)
352 *
353 * Write out and wait upon file offsets lstart->lend, inclusive.
354 *
355 * Note that `lend' is inclusive (describes the last byte to be written) so
356 * that this function can be used to write to the very end-of-file (end = -1).
357 */
358int filemap_write_and_wait_range(struct address_space *mapping,
359 loff_t lstart, loff_t lend)
360{
361 int err = 0;
362
363 if (mapping->nrpages) {
364 err = __filemap_fdatawrite_range(mapping, lstart, lend,
365 WB_SYNC_ALL);
366 /* See comment of filemap_write_and_wait() */
367 if (err != -EIO) {
368 int err2 = filemap_fdatawait_range(mapping,
369 lstart, lend);
370 if (!err)
371 err = err2;
372 }
373 }
374 return err;
375}
376EXPORT_SYMBOL(filemap_write_and_wait_range);
377
378/**
379 * replace_page_cache_page - replace a pagecache page with a new one
380 * @old: page to be replaced
381 * @new: page to replace with
382 * @gfp_mask: allocation mode
383 *
384 * This function replaces a page in the pagecache with a new one. On
385 * success it acquires the pagecache reference for the new page and
386 * drops it for the old page. Both the old and new pages must be
387 * locked. This function does not add the new page to the LRU, the
388 * caller must do that.
389 *
390 * The remove + add is atomic. The only way this function can fail is
391 * memory allocation failure.
392 */
393int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
394{
395 int error;
396 struct mem_cgroup *memcg = NULL;
397
398 VM_BUG_ON(!PageLocked(old));
399 VM_BUG_ON(!PageLocked(new));
400 VM_BUG_ON(new->mapping);
401
402 /*
403 * This is not page migration, but prepare_migration and
404 * end_migration does enough work for charge replacement.
405 *
406 * In the longer term we probably want a specialized function
407 * for moving the charge from old to new in a more efficient
408 * manner.
409 */
410 error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
411 if (error)
412 return error;
413
414 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
415 if (!error) {
416 struct address_space *mapping = old->mapping;
417 void (*freepage)(struct page *);
418
419 pgoff_t offset = old->index;
420 freepage = mapping->a_ops->freepage;
421
422 page_cache_get(new);
423 new->mapping = mapping;
424 new->index = offset;
425
426 spin_lock_irq(&mapping->tree_lock);
427 __delete_from_page_cache(old);
428 error = radix_tree_insert(&mapping->page_tree, offset, new);
429 BUG_ON(error);
430 mapping->nrpages++;
431 __inc_zone_page_state(new, NR_FILE_PAGES);
432 if (PageSwapBacked(new))
433 __inc_zone_page_state(new, NR_SHMEM);
434 spin_unlock_irq(&mapping->tree_lock);
435 radix_tree_preload_end();
436 if (freepage)
437 freepage(old);
438 page_cache_release(old);
439 mem_cgroup_end_migration(memcg, old, new, true);
440 } else {
441 mem_cgroup_end_migration(memcg, old, new, false);
442 }
443
444 return error;
445}
446EXPORT_SYMBOL_GPL(replace_page_cache_page);
447
448/**
449 * add_to_page_cache_locked - add a locked page to the pagecache
450 * @page: page to add
451 * @mapping: the page's address_space
452 * @offset: page index
453 * @gfp_mask: page allocation mode
454 *
455 * This function is used to add a page to the pagecache. It must be locked.
456 * This function does not add the page to the LRU. The caller must do that.
457 */
458int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
459 pgoff_t offset, gfp_t gfp_mask)
460{
461 int error;
462
463 VM_BUG_ON(!PageLocked(page));
464 VM_BUG_ON(PageSwapBacked(page));
465
466 error = mem_cgroup_cache_charge(page, current->mm,
467 gfp_mask & GFP_RECLAIM_MASK);
468 if (error)
469 goto out;
470
471 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
472 if (error == 0) {
473 page_cache_get(page);
474 page->mapping = mapping;
475 page->index = offset;
476
477 spin_lock_irq(&mapping->tree_lock);
478 error = radix_tree_insert(&mapping->page_tree, offset, page);
479 if (likely(!error)) {
480 mapping->nrpages++;
481 __inc_zone_page_state(page, NR_FILE_PAGES);
482 spin_unlock_irq(&mapping->tree_lock);
483 } else {
484 page->mapping = NULL;
485 /* Leave page->index set: truncation relies upon it */
486 spin_unlock_irq(&mapping->tree_lock);
487 mem_cgroup_uncharge_cache_page(page);
488 page_cache_release(page);
489 }
490 radix_tree_preload_end();
491 } else
492 mem_cgroup_uncharge_cache_page(page);
493out:
494 return error;
495}
496EXPORT_SYMBOL(add_to_page_cache_locked);
497
498int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
499 pgoff_t offset, gfp_t gfp_mask)
500{
501 int ret;
502
503 ret = add_to_page_cache(page, mapping, offset, gfp_mask);
504 if (ret == 0)
505 lru_cache_add_file(page);
506 return ret;
507}
508EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
509
510#ifdef CONFIG_NUMA
511struct page *__page_cache_alloc(gfp_t gfp)
512{
513 int n;
514 struct page *page;
515
516 if (cpuset_do_page_mem_spread()) {
517 get_mems_allowed();
518 n = cpuset_mem_spread_node();
519 page = alloc_pages_exact_node(n, gfp, 0);
520 put_mems_allowed();
521 return page;
522 }
523 return alloc_pages(gfp, 0);
524}
525EXPORT_SYMBOL(__page_cache_alloc);
526#endif
527
528/*
529 * In order to wait for pages to become available there must be
530 * waitqueues associated with pages. By using a hash table of
531 * waitqueues where the bucket discipline is to maintain all
532 * waiters on the same queue and wake all when any of the pages
533 * become available, and for the woken contexts to check to be
534 * sure the appropriate page became available, this saves space
535 * at a cost of "thundering herd" phenomena during rare hash
536 * collisions.
537 */
538static wait_queue_head_t *page_waitqueue(struct page *page)
539{
540 const struct zone *zone = page_zone(page);
541
542 return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
543}
544
545static inline void wake_up_page(struct page *page, int bit)
546{
547 __wake_up_bit(page_waitqueue(page), &page->flags, bit);
548}
549
550void wait_on_page_bit(struct page *page, int bit_nr)
551{
552 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
553
554 if (test_bit(bit_nr, &page->flags))
555 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
556 TASK_UNINTERRUPTIBLE);
557}
558EXPORT_SYMBOL(wait_on_page_bit);
559
560int wait_on_page_bit_killable(struct page *page, int bit_nr)
561{
562 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
563
564 if (!test_bit(bit_nr, &page->flags))
565 return 0;
566
567 return __wait_on_bit(page_waitqueue(page), &wait,
568 sleep_on_page_killable, TASK_KILLABLE);
569}
570
571/**
572 * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
573 * @page: Page defining the wait queue of interest
574 * @waiter: Waiter to add to the queue
575 *
576 * Add an arbitrary @waiter to the wait queue for the nominated @page.
577 */
578void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
579{
580 wait_queue_head_t *q = page_waitqueue(page);
581 unsigned long flags;
582
583 spin_lock_irqsave(&q->lock, flags);
584 __add_wait_queue(q, waiter);
585 spin_unlock_irqrestore(&q->lock, flags);
586}
587EXPORT_SYMBOL_GPL(add_page_wait_queue);
588
589/**
590 * unlock_page - unlock a locked page
591 * @page: the page
592 *
593 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
594 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
595 * mechananism between PageLocked pages and PageWriteback pages is shared.
596 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
597 *
598 * The mb is necessary to enforce ordering between the clear_bit and the read
599 * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
600 */
601void unlock_page(struct page *page)
602{
603 VM_BUG_ON(!PageLocked(page));
604 clear_bit_unlock(PG_locked, &page->flags);
605 smp_mb__after_clear_bit();
606 wake_up_page(page, PG_locked);
607}
608EXPORT_SYMBOL(unlock_page);
609
610/**
611 * end_page_writeback - end writeback against a page
612 * @page: the page
613 */
614void end_page_writeback(struct page *page)
615{
616 if (TestClearPageReclaim(page))
617 rotate_reclaimable_page(page);
618
619 if (!test_clear_page_writeback(page))
620 BUG();
621
622 smp_mb__after_clear_bit();
623 wake_up_page(page, PG_writeback);
624}
625EXPORT_SYMBOL(end_page_writeback);
626
627/**
628 * __lock_page - get a lock on the page, assuming we need to sleep to get it
629 * @page: the page to lock
630 */
631void __lock_page(struct page *page)
632{
633 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
634
635 __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
636 TASK_UNINTERRUPTIBLE);
637}
638EXPORT_SYMBOL(__lock_page);
639
640int __lock_page_killable(struct page *page)
641{
642 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
643
644 return __wait_on_bit_lock(page_waitqueue(page), &wait,
645 sleep_on_page_killable, TASK_KILLABLE);
646}
647EXPORT_SYMBOL_GPL(__lock_page_killable);
648
649int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
650 unsigned int flags)
651{
652 if (flags & FAULT_FLAG_ALLOW_RETRY) {
653 /*
654 * CAUTION! In this case, mmap_sem is not released
655 * even though return 0.
656 */
657 if (flags & FAULT_FLAG_RETRY_NOWAIT)
658 return 0;
659
660 up_read(&mm->mmap_sem);
661 if (flags & FAULT_FLAG_KILLABLE)
662 wait_on_page_locked_killable(page);
663 else
664 wait_on_page_locked(page);
665 return 0;
666 } else {
667 if (flags & FAULT_FLAG_KILLABLE) {
668 int ret;
669
670 ret = __lock_page_killable(page);
671 if (ret) {
672 up_read(&mm->mmap_sem);
673 return 0;
674 }
675 } else
676 __lock_page(page);
677 return 1;
678 }
679}
680
681/**
682 * find_get_page - find and get a page reference
683 * @mapping: the address_space to search
684 * @offset: the page index
685 *
686 * Is there a pagecache struct page at the given (mapping, offset) tuple?
687 * If yes, increment its refcount and return it; if no, return NULL.
688 */
689struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
690{
691 void **pagep;
692 struct page *page;
693
694 rcu_read_lock();
695repeat:
696 page = NULL;
697 pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
698 if (pagep) {
699 page = radix_tree_deref_slot(pagep);
700 if (unlikely(!page))
701 goto out;
702 if (radix_tree_exception(page)) {
703 if (radix_tree_deref_retry(page))
704 goto repeat;
705 /*
706 * Otherwise, shmem/tmpfs must be storing a swap entry
707 * here as an exceptional entry: so return it without
708 * attempting to raise page count.
709 */
710 goto out;
711 }
712 if (!page_cache_get_speculative(page))
713 goto repeat;
714
715 /*
716 * Has the page moved?
717 * This is part of the lockless pagecache protocol. See
718 * include/linux/pagemap.h for details.
719 */
720 if (unlikely(page != *pagep)) {
721 page_cache_release(page);
722 goto repeat;
723 }
724 }
725out:
726 rcu_read_unlock();
727
728 return page;
729}
730EXPORT_SYMBOL(find_get_page);
731
732/**
733 * find_lock_page - locate, pin and lock a pagecache page
734 * @mapping: the address_space to search
735 * @offset: the page index
736 *
737 * Locates the desired pagecache page, locks it, increments its reference
738 * count and returns its address.
739 *
740 * Returns zero if the page was not present. find_lock_page() may sleep.
741 */
742struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
743{
744 struct page *page;
745
746repeat:
747 page = find_get_page(mapping, offset);
748 if (page && !radix_tree_exception(page)) {
749 lock_page(page);
750 /* Has the page been truncated? */
751 if (unlikely(page->mapping != mapping)) {
752 unlock_page(page);
753 page_cache_release(page);
754 goto repeat;
755 }
756 VM_BUG_ON(page->index != offset);
757 }
758 return page;
759}
760EXPORT_SYMBOL(find_lock_page);
761
762/**
763 * find_or_create_page - locate or add a pagecache page
764 * @mapping: the page's address_space
765 * @index: the page's index into the mapping
766 * @gfp_mask: page allocation mode
767 *
768 * Locates a page in the pagecache. If the page is not present, a new page
769 * is allocated using @gfp_mask and is added to the pagecache and to the VM's
770 * LRU list. The returned page is locked and has its reference count
771 * incremented.
772 *
773 * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
774 * allocation!
775 *
776 * find_or_create_page() returns the desired page's address, or zero on
777 * memory exhaustion.
778 */
779struct page *find_or_create_page(struct address_space *mapping,
780 pgoff_t index, gfp_t gfp_mask)
781{
782 struct page *page;
783 int err;
784repeat:
785 page = find_lock_page(mapping, index);
786 if (!page) {
787 page = __page_cache_alloc(gfp_mask);
788 if (!page)
789 return NULL;
790 /*
791 * We want a regular kernel memory (not highmem or DMA etc)
792 * allocation for the radix tree nodes, but we need to honour
793 * the context-specific requirements the caller has asked for.
794 * GFP_RECLAIM_MASK collects those requirements.
795 */
796 err = add_to_page_cache_lru(page, mapping, index,
797 (gfp_mask & GFP_RECLAIM_MASK));
798 if (unlikely(err)) {
799 page_cache_release(page);
800 page = NULL;
801 if (err == -EEXIST)
802 goto repeat;
803 }
804 }
805 return page;
806}
807EXPORT_SYMBOL(find_or_create_page);
808
809/**
810 * find_get_pages - gang pagecache lookup
811 * @mapping: The address_space to search
812 * @start: The starting page index
813 * @nr_pages: The maximum number of pages
814 * @pages: Where the resulting pages are placed
815 *
816 * find_get_pages() will search for and return a group of up to
817 * @nr_pages pages in the mapping. The pages are placed at @pages.
818 * find_get_pages() takes a reference against the returned pages.
819 *
820 * The search returns a group of mapping-contiguous pages with ascending
821 * indexes. There may be holes in the indices due to not-present pages.
822 *
823 * find_get_pages() returns the number of pages which were found.
824 */
825unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
826 unsigned int nr_pages, struct page **pages)
827{
828 unsigned int i;
829 unsigned int ret;
830 unsigned int nr_found, nr_skip;
831
832 rcu_read_lock();
833restart:
834 nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
835 (void ***)pages, NULL, start, nr_pages);
836 ret = 0;
837 nr_skip = 0;
838 for (i = 0; i < nr_found; i++) {
839 struct page *page;
840repeat:
841 page = radix_tree_deref_slot((void **)pages[i]);
842 if (unlikely(!page))
843 continue;
844
845 if (radix_tree_exception(page)) {
846 if (radix_tree_deref_retry(page)) {
847 /*
848 * Transient condition which can only trigger
849 * when entry at index 0 moves out of or back
850 * to root: none yet gotten, safe to restart.
851 */
852 WARN_ON(start | i);
853 goto restart;
854 }
855 /*
856 * Otherwise, shmem/tmpfs must be storing a swap entry
857 * here as an exceptional entry: so skip over it -
858 * we only reach this from invalidate_mapping_pages().
859 */
860 nr_skip++;
861 continue;
862 }
863
864 if (!page_cache_get_speculative(page))
865 goto repeat;
866
867 /* Has the page moved? */
868 if (unlikely(page != *((void **)pages[i]))) {
869 page_cache_release(page);
870 goto repeat;
871 }
872
873 pages[ret] = page;
874 ret++;
875 }
876
877 /*
878 * If all entries were removed before we could secure them,
879 * try again, because callers stop trying once 0 is returned.
880 */
881 if (unlikely(!ret && nr_found > nr_skip))
882 goto restart;
883 rcu_read_unlock();
884 return ret;
885}
886
887/**
888 * find_get_pages_contig - gang contiguous pagecache lookup
889 * @mapping: The address_space to search
890 * @index: The starting page index
891 * @nr_pages: The maximum number of pages
892 * @pages: Where the resulting pages are placed
893 *
894 * find_get_pages_contig() works exactly like find_get_pages(), except
895 * that the returned number of pages are guaranteed to be contiguous.
896 *
897 * find_get_pages_contig() returns the number of pages which were found.
898 */
899unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
900 unsigned int nr_pages, struct page **pages)
901{
902 unsigned int i;
903 unsigned int ret;
904 unsigned int nr_found;
905
906 rcu_read_lock();
907restart:
908 nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
909 (void ***)pages, NULL, index, nr_pages);
910 ret = 0;
911 for (i = 0; i < nr_found; i++) {
912 struct page *page;
913repeat:
914 page = radix_tree_deref_slot((void **)pages[i]);
915 if (unlikely(!page))
916 continue;
917
918 if (radix_tree_exception(page)) {
919 if (radix_tree_deref_retry(page)) {
920 /*
921 * Transient condition which can only trigger
922 * when entry at index 0 moves out of or back
923 * to root: none yet gotten, safe to restart.
924 */
925 goto restart;
926 }
927 /*
928 * Otherwise, shmem/tmpfs must be storing a swap entry
929 * here as an exceptional entry: so stop looking for
930 * contiguous pages.
931 */
932 break;
933 }
934
935 if (!page_cache_get_speculative(page))
936 goto repeat;
937
938 /* Has the page moved? */
939 if (unlikely(page != *((void **)pages[i]))) {
940 page_cache_release(page);
941 goto repeat;
942 }
943
944 /*
945 * must check mapping and index after taking the ref.
946 * otherwise we can get both false positives and false
947 * negatives, which is just confusing to the caller.
948 */
949 if (page->mapping == NULL || page->index != index) {
950 page_cache_release(page);
951 break;
952 }
953
954 pages[ret] = page;
955 ret++;
956 index++;
957 }
958 rcu_read_unlock();
959 return ret;
960}
961EXPORT_SYMBOL(find_get_pages_contig);
962
963/**
964 * find_get_pages_tag - find and return pages that match @tag
965 * @mapping: the address_space to search
966 * @index: the starting page index
967 * @tag: the tag index
968 * @nr_pages: the maximum number of pages
969 * @pages: where the resulting pages are placed
970 *
971 * Like find_get_pages, except we only return pages which are tagged with
972 * @tag. We update @index to index the next page for the traversal.
973 */
974unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
975 int tag, unsigned int nr_pages, struct page **pages)
976{
977 unsigned int i;
978 unsigned int ret;
979 unsigned int nr_found;
980
981 rcu_read_lock();
982restart:
983 nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
984 (void ***)pages, *index, nr_pages, tag);
985 ret = 0;
986 for (i = 0; i < nr_found; i++) {
987 struct page *page;
988repeat:
989 page = radix_tree_deref_slot((void **)pages[i]);
990 if (unlikely(!page))
991 continue;
992
993 if (radix_tree_exception(page)) {
994 if (radix_tree_deref_retry(page)) {
995 /*
996 * Transient condition which can only trigger
997 * when entry at index 0 moves out of or back
998 * to root: none yet gotten, safe to restart.
999 */
1000 goto restart;
1001 }
1002 /*
1003 * This function is never used on a shmem/tmpfs
1004 * mapping, so a swap entry won't be found here.
1005 */
1006 BUG();
1007 }
1008
1009 if (!page_cache_get_speculative(page))
1010 goto repeat;
1011
1012 /* Has the page moved? */
1013 if (unlikely(page != *((void **)pages[i]))) {
1014 page_cache_release(page);
1015 goto repeat;
1016 }
1017
1018 pages[ret] = page;
1019 ret++;
1020 }
1021
1022 /*
1023 * If all entries were removed before we could secure them,
1024 * try again, because callers stop trying once 0 is returned.
1025 */
1026 if (unlikely(!ret && nr_found))
1027 goto restart;
1028 rcu_read_unlock();
1029
1030 if (ret)
1031 *index = pages[ret - 1]->index + 1;
1032
1033 return ret;
1034}
1035EXPORT_SYMBOL(find_get_pages_tag);
1036
1037/**
1038 * grab_cache_page_nowait - returns locked page at given index in given cache
1039 * @mapping: target address_space
1040 * @index: the page index
1041 *
1042 * Same as grab_cache_page(), but do not wait if the page is unavailable.
1043 * This is intended for speculative data generators, where the data can
1044 * be regenerated if the page couldn't be grabbed. This routine should
1045 * be safe to call while holding the lock for another page.
1046 *
1047 * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1048 * and deadlock against the caller's locked page.
1049 */
1050struct page *
1051grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1052{
1053 struct page *page = find_get_page(mapping, index);
1054
1055 if (page) {
1056 if (trylock_page(page))
1057 return page;
1058 page_cache_release(page);
1059 return NULL;
1060 }
1061 page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1062 if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1063 page_cache_release(page);
1064 page = NULL;
1065 }
1066 return page;
1067}
1068EXPORT_SYMBOL(grab_cache_page_nowait);
1069
1070/*
1071 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1072 * a _large_ part of the i/o request. Imagine the worst scenario:
1073 *
1074 * ---R__________________________________________B__________
1075 * ^ reading here ^ bad block(assume 4k)
1076 *
1077 * read(R) => miss => readahead(R...B) => media error => frustrating retries
1078 * => failing the whole request => read(R) => read(R+1) =>
1079 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1080 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1081 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1082 *
1083 * It is going insane. Fix it by quickly scaling down the readahead size.
1084 */
1085static void shrink_readahead_size_eio(struct file *filp,
1086 struct file_ra_state *ra)
1087{
1088 ra->ra_pages /= 4;
1089}
1090
1091/**
1092 * do_generic_file_read - generic file read routine
1093 * @filp: the file to read
1094 * @ppos: current file position
1095 * @desc: read_descriptor
1096 * @actor: read method
1097 *
1098 * This is a generic file read routine, and uses the
1099 * mapping->a_ops->readpage() function for the actual low-level stuff.
1100 *
1101 * This is really ugly. But the goto's actually try to clarify some
1102 * of the logic when it comes to error handling etc.
1103 */
1104static void do_generic_file_read(struct file *filp, loff_t *ppos,
1105 read_descriptor_t *desc, read_actor_t actor)
1106{
1107 struct address_space *mapping = filp->f_mapping;
1108 struct inode *inode = mapping->host;
1109 struct file_ra_state *ra = &filp->f_ra;
1110 pgoff_t index;
1111 pgoff_t last_index;
1112 pgoff_t prev_index;
1113 unsigned long offset; /* offset into pagecache page */
1114 unsigned int prev_offset;
1115 int error;
1116
1117 index = *ppos >> PAGE_CACHE_SHIFT;
1118 prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1119 prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1120 last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1121 offset = *ppos & ~PAGE_CACHE_MASK;
1122
1123 for (;;) {
1124 struct page *page;
1125 pgoff_t end_index;
1126 loff_t isize;
1127 unsigned long nr, ret;
1128
1129 cond_resched();
1130find_page:
1131 page = find_get_page(mapping, index);
1132 if (!page) {
1133 page_cache_sync_readahead(mapping,
1134 ra, filp,
1135 index, last_index - index);
1136 page = find_get_page(mapping, index);
1137 if (unlikely(page == NULL))
1138 goto no_cached_page;
1139 }
1140 if (PageReadahead(page)) {
1141 page_cache_async_readahead(mapping,
1142 ra, filp, page,
1143 index, last_index - index);
1144 }
1145 if (!PageUptodate(page)) {
1146 if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1147 !mapping->a_ops->is_partially_uptodate)
1148 goto page_not_up_to_date;
1149 if (!trylock_page(page))
1150 goto page_not_up_to_date;
1151 /* Did it get truncated before we got the lock? */
1152 if (!page->mapping)
1153 goto page_not_up_to_date_locked;
1154 if (!mapping->a_ops->is_partially_uptodate(page,
1155 desc, offset))
1156 goto page_not_up_to_date_locked;
1157 unlock_page(page);
1158 }
1159page_ok:
1160 /*
1161 * i_size must be checked after we know the page is Uptodate.
1162 *
1163 * Checking i_size after the check allows us to calculate
1164 * the correct value for "nr", which means the zero-filled
1165 * part of the page is not copied back to userspace (unless
1166 * another truncate extends the file - this is desired though).
1167 */
1168
1169 isize = i_size_read(inode);
1170 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1171 if (unlikely(!isize || index > end_index)) {
1172 page_cache_release(page);
1173 goto out;
1174 }
1175
1176 /* nr is the maximum number of bytes to copy from this page */
1177 nr = PAGE_CACHE_SIZE;
1178 if (index == end_index) {
1179 nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1180 if (nr <= offset) {
1181 page_cache_release(page);
1182 goto out;
1183 }
1184 }
1185 nr = nr - offset;
1186
1187 /* If users can be writing to this page using arbitrary
1188 * virtual addresses, take care about potential aliasing
1189 * before reading the page on the kernel side.
1190 */
1191 if (mapping_writably_mapped(mapping))
1192 flush_dcache_page(page);
1193
1194 /*
1195 * When a sequential read accesses a page several times,
1196 * only mark it as accessed the first time.
1197 */
1198 if (prev_index != index || offset != prev_offset)
1199 mark_page_accessed(page);
1200 prev_index = index;
1201
1202 /*
1203 * Ok, we have the page, and it's up-to-date, so
1204 * now we can copy it to user space...
1205 *
1206 * The actor routine returns how many bytes were actually used..
1207 * NOTE! This may not be the same as how much of a user buffer
1208 * we filled up (we may be padding etc), so we can only update
1209 * "pos" here (the actor routine has to update the user buffer
1210 * pointers and the remaining count).
1211 */
1212 ret = actor(desc, page, offset, nr);
1213 offset += ret;
1214 index += offset >> PAGE_CACHE_SHIFT;
1215 offset &= ~PAGE_CACHE_MASK;
1216 prev_offset = offset;
1217
1218 page_cache_release(page);
1219 if (ret == nr && desc->count)
1220 continue;
1221 goto out;
1222
1223page_not_up_to_date:
1224 /* Get exclusive access to the page ... */
1225 error = lock_page_killable(page);
1226 if (unlikely(error))
1227 goto readpage_error;
1228
1229page_not_up_to_date_locked:
1230 /* Did it get truncated before we got the lock? */
1231 if (!page->mapping) {
1232 unlock_page(page);
1233 page_cache_release(page);
1234 continue;
1235 }
1236
1237 /* Did somebody else fill it already? */
1238 if (PageUptodate(page)) {
1239 unlock_page(page);
1240 goto page_ok;
1241 }
1242
1243readpage:
1244 /*
1245 * A previous I/O error may have been due to temporary
1246 * failures, eg. multipath errors.
1247 * PG_error will be set again if readpage fails.
1248 */
1249 ClearPageError(page);
1250 /* Start the actual read. The read will unlock the page. */
1251 error = mapping->a_ops->readpage(filp, page);
1252
1253 if (unlikely(error)) {
1254 if (error == AOP_TRUNCATED_PAGE) {
1255 page_cache_release(page);
1256 goto find_page;
1257 }
1258 goto readpage_error;
1259 }
1260
1261 if (!PageUptodate(page)) {
1262 error = lock_page_killable(page);
1263 if (unlikely(error))
1264 goto readpage_error;
1265 if (!PageUptodate(page)) {
1266 if (page->mapping == NULL) {
1267 /*
1268 * invalidate_mapping_pages got it
1269 */
1270 unlock_page(page);
1271 page_cache_release(page);
1272 goto find_page;
1273 }
1274 unlock_page(page);
1275 shrink_readahead_size_eio(filp, ra);
1276 error = -EIO;
1277 goto readpage_error;
1278 }
1279 unlock_page(page);
1280 }
1281
1282 goto page_ok;
1283
1284readpage_error:
1285 /* UHHUH! A synchronous read error occurred. Report it */
1286 desc->error = error;
1287 page_cache_release(page);
1288 goto out;
1289
1290no_cached_page:
1291 /*
1292 * Ok, it wasn't cached, so we need to create a new
1293 * page..
1294 */
1295 page = page_cache_alloc_cold(mapping);
1296 if (!page) {
1297 desc->error = -ENOMEM;
1298 goto out;
1299 }
1300 error = add_to_page_cache_lru(page, mapping,
1301 index, GFP_KERNEL);
1302 if (error) {
1303 page_cache_release(page);
1304 if (error == -EEXIST)
1305 goto find_page;
1306 desc->error = error;
1307 goto out;
1308 }
1309 goto readpage;
1310 }
1311
1312out:
1313 ra->prev_pos = prev_index;
1314 ra->prev_pos <<= PAGE_CACHE_SHIFT;
1315 ra->prev_pos |= prev_offset;
1316
1317 *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1318 file_accessed(filp);
1319}
1320
1321int file_read_actor(read_descriptor_t *desc, struct page *page,
1322 unsigned long offset, unsigned long size)
1323{
1324 char *kaddr;
1325 unsigned long left, count = desc->count;
1326
1327 if (size > count)
1328 size = count;
1329
1330 /*
1331 * Faults on the destination of a read are common, so do it before
1332 * taking the kmap.
1333 */
1334 if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1335 kaddr = kmap_atomic(page, KM_USER0);
1336 left = __copy_to_user_inatomic(desc->arg.buf,
1337 kaddr + offset, size);
1338 kunmap_atomic(kaddr, KM_USER0);
1339 if (left == 0)
1340 goto success;
1341 }
1342
1343 /* Do it the slow way */
1344 kaddr = kmap(page);
1345 left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1346 kunmap(page);
1347
1348 if (left) {
1349 size -= left;
1350 desc->error = -EFAULT;
1351 }
1352success:
1353 desc->count = count - size;
1354 desc->written += size;
1355 desc->arg.buf += size;
1356 return size;
1357}
1358
1359/*
1360 * Performs necessary checks before doing a write
1361 * @iov: io vector request
1362 * @nr_segs: number of segments in the iovec
1363 * @count: number of bytes to write
1364 * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1365 *
1366 * Adjust number of segments and amount of bytes to write (nr_segs should be
1367 * properly initialized first). Returns appropriate error code that caller
1368 * should return or zero in case that write should be allowed.
1369 */
1370int generic_segment_checks(const struct iovec *iov,
1371 unsigned long *nr_segs, size_t *count, int access_flags)
1372{
1373 unsigned long seg;
1374 size_t cnt = 0;
1375 for (seg = 0; seg < *nr_segs; seg++) {
1376 const struct iovec *iv = &iov[seg];
1377
1378 /*
1379 * If any segment has a negative length, or the cumulative
1380 * length ever wraps negative then return -EINVAL.
1381 */
1382 cnt += iv->iov_len;
1383 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1384 return -EINVAL;
1385 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1386 continue;
1387 if (seg == 0)
1388 return -EFAULT;
1389 *nr_segs = seg;
1390 cnt -= iv->iov_len; /* This segment is no good */
1391 break;
1392 }
1393 *count = cnt;
1394 return 0;
1395}
1396EXPORT_SYMBOL(generic_segment_checks);
1397
1398/**
1399 * generic_file_aio_read - generic filesystem read routine
1400 * @iocb: kernel I/O control block
1401 * @iov: io vector request
1402 * @nr_segs: number of segments in the iovec
1403 * @pos: current file position
1404 *
1405 * This is the "read()" routine for all filesystems
1406 * that can use the page cache directly.
1407 */
1408ssize_t
1409generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1410 unsigned long nr_segs, loff_t pos)
1411{
1412 struct file *filp = iocb->ki_filp;
1413 ssize_t retval;
1414 unsigned long seg = 0;
1415 size_t count;
1416 loff_t *ppos = &iocb->ki_pos;
1417 struct blk_plug plug;
1418
1419 count = 0;
1420 retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1421 if (retval)
1422 return retval;
1423
1424 blk_start_plug(&plug);
1425
1426 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1427 if (filp->f_flags & O_DIRECT) {
1428 loff_t size;
1429 struct address_space *mapping;
1430 struct inode *inode;
1431
1432 mapping = filp->f_mapping;
1433 inode = mapping->host;
1434 if (!count)
1435 goto out; /* skip atime */
1436 size = i_size_read(inode);
1437 if (pos < size) {
1438 retval = filemap_write_and_wait_range(mapping, pos,
1439 pos + iov_length(iov, nr_segs) - 1);
1440 if (!retval) {
1441 retval = mapping->a_ops->direct_IO(READ, iocb,
1442 iov, pos, nr_segs);
1443 }
1444 if (retval > 0) {
1445 *ppos = pos + retval;
1446 count -= retval;
1447 }
1448
1449 /*
1450 * Btrfs can have a short DIO read if we encounter
1451 * compressed extents, so if there was an error, or if
1452 * we've already read everything we wanted to, or if
1453 * there was a short read because we hit EOF, go ahead
1454 * and return. Otherwise fallthrough to buffered io for
1455 * the rest of the read.
1456 */
1457 if (retval < 0 || !count || *ppos >= size) {
1458 file_accessed(filp);
1459 goto out;
1460 }
1461 }
1462 }
1463
1464 count = retval;
1465 for (seg = 0; seg < nr_segs; seg++) {
1466 read_descriptor_t desc;
1467 loff_t offset = 0;
1468
1469 /*
1470 * If we did a short DIO read we need to skip the section of the
1471 * iov that we've already read data into.
1472 */
1473 if (count) {
1474 if (count > iov[seg].iov_len) {
1475 count -= iov[seg].iov_len;
1476 continue;
1477 }
1478 offset = count;
1479 count = 0;
1480 }
1481
1482 desc.written = 0;
1483 desc.arg.buf = iov[seg].iov_base + offset;
1484 desc.count = iov[seg].iov_len - offset;
1485 if (desc.count == 0)
1486 continue;
1487 desc.error = 0;
1488 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1489 retval += desc.written;
1490 if (desc.error) {
1491 retval = retval ?: desc.error;
1492 break;
1493 }
1494 if (desc.count > 0)
1495 break;
1496 }
1497out:
1498 blk_finish_plug(&plug);
1499 return retval;
1500}
1501EXPORT_SYMBOL(generic_file_aio_read);
1502
1503static ssize_t
1504do_readahead(struct address_space *mapping, struct file *filp,
1505 pgoff_t index, unsigned long nr)
1506{
1507 if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1508 return -EINVAL;
1509
1510 force_page_cache_readahead(mapping, filp, index, nr);
1511 return 0;
1512}
1513
1514SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1515{
1516 ssize_t ret;
1517 struct file *file;
1518
1519 ret = -EBADF;
1520 file = fget(fd);
1521 if (file) {
1522 if (file->f_mode & FMODE_READ) {
1523 struct address_space *mapping = file->f_mapping;
1524 pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1525 pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1526 unsigned long len = end - start + 1;
1527 ret = do_readahead(mapping, file, start, len);
1528 }
1529 fput(file);
1530 }
1531 return ret;
1532}
1533#ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1534asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1535{
1536 return SYSC_readahead((int) fd, offset, (size_t) count);
1537}
1538SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1539#endif
1540
1541#ifdef CONFIG_MMU
1542/**
1543 * page_cache_read - adds requested page to the page cache if not already there
1544 * @file: file to read
1545 * @offset: page index
1546 *
1547 * This adds the requested page to the page cache if it isn't already there,
1548 * and schedules an I/O to read in its contents from disk.
1549 */
1550static int page_cache_read(struct file *file, pgoff_t offset)
1551{
1552 struct address_space *mapping = file->f_mapping;
1553 struct page *page;
1554 int ret;
1555
1556 do {
1557 page = page_cache_alloc_cold(mapping);
1558 if (!page)
1559 return -ENOMEM;
1560
1561 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1562 if (ret == 0)
1563 ret = mapping->a_ops->readpage(file, page);
1564 else if (ret == -EEXIST)
1565 ret = 0; /* losing race to add is OK */
1566
1567 page_cache_release(page);
1568
1569 } while (ret == AOP_TRUNCATED_PAGE);
1570
1571 return ret;
1572}
1573
1574#define MMAP_LOTSAMISS (100)
1575
1576/*
1577 * Synchronous readahead happens when we don't even find
1578 * a page in the page cache at all.
1579 */
1580static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1581 struct file_ra_state *ra,
1582 struct file *file,
1583 pgoff_t offset)
1584{
1585 unsigned long ra_pages;
1586 struct address_space *mapping = file->f_mapping;
1587
1588 /* If we don't want any read-ahead, don't bother */
1589 if (VM_RandomReadHint(vma))
1590 return;
1591 if (!ra->ra_pages)
1592 return;
1593
1594 if (VM_SequentialReadHint(vma)) {
1595 page_cache_sync_readahead(mapping, ra, file, offset,
1596 ra->ra_pages);
1597 return;
1598 }
1599
1600 /* Avoid banging the cache line if not needed */
1601 if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1602 ra->mmap_miss++;
1603
1604 /*
1605 * Do we miss much more than hit in this file? If so,
1606 * stop bothering with read-ahead. It will only hurt.
1607 */
1608 if (ra->mmap_miss > MMAP_LOTSAMISS)
1609 return;
1610
1611 /*
1612 * mmap read-around
1613 */
1614 ra_pages = max_sane_readahead(ra->ra_pages);
1615 ra->start = max_t(long, 0, offset - ra_pages / 2);
1616 ra->size = ra_pages;
1617 ra->async_size = ra_pages / 4;
1618 ra_submit(ra, mapping, file);
1619}
1620
1621/*
1622 * Asynchronous readahead happens when we find the page and PG_readahead,
1623 * so we want to possibly extend the readahead further..
1624 */
1625static void do_async_mmap_readahead(struct vm_area_struct *vma,
1626 struct file_ra_state *ra,
1627 struct file *file,
1628 struct page *page,
1629 pgoff_t offset)
1630{
1631 struct address_space *mapping = file->f_mapping;
1632
1633 /* If we don't want any read-ahead, don't bother */
1634 if (VM_RandomReadHint(vma))
1635 return;
1636 if (ra->mmap_miss > 0)
1637 ra->mmap_miss--;
1638 if (PageReadahead(page))
1639 page_cache_async_readahead(mapping, ra, file,
1640 page, offset, ra->ra_pages);
1641}
1642
1643/**
1644 * filemap_fault - read in file data for page fault handling
1645 * @vma: vma in which the fault was taken
1646 * @vmf: struct vm_fault containing details of the fault
1647 *
1648 * filemap_fault() is invoked via the vma operations vector for a
1649 * mapped memory region to read in file data during a page fault.
1650 *
1651 * The goto's are kind of ugly, but this streamlines the normal case of having
1652 * it in the page cache, and handles the special cases reasonably without
1653 * having a lot of duplicated code.
1654 */
1655int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1656{
1657 int error;
1658 struct file *file = vma->vm_file;
1659 struct address_space *mapping = file->f_mapping;
1660 struct file_ra_state *ra = &file->f_ra;
1661 struct inode *inode = mapping->host;
1662 pgoff_t offset = vmf->pgoff;
1663 struct page *page;
1664 pgoff_t size;
1665 int ret = 0;
1666
1667 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1668 if (offset >= size)
1669 return VM_FAULT_SIGBUS;
1670
1671 /*
1672 * Do we have something in the page cache already?
1673 */
1674 page = find_get_page(mapping, offset);
1675 if (likely(page)) {
1676 /*
1677 * We found the page, so try async readahead before
1678 * waiting for the lock.
1679 */
1680 do_async_mmap_readahead(vma, ra, file, page, offset);
1681 } else {
1682 /* No page in the page cache at all */
1683 do_sync_mmap_readahead(vma, ra, file, offset);
1684 count_vm_event(PGMAJFAULT);
1685 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1686 ret = VM_FAULT_MAJOR;
1687retry_find:
1688 page = find_get_page(mapping, offset);
1689 if (!page)
1690 goto no_cached_page;
1691 }
1692
1693 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1694 page_cache_release(page);
1695 return ret | VM_FAULT_RETRY;
1696 }
1697
1698 /* Did it get truncated? */
1699 if (unlikely(page->mapping != mapping)) {
1700 unlock_page(page);
1701 put_page(page);
1702 goto retry_find;
1703 }
1704 VM_BUG_ON(page->index != offset);
1705
1706 /*
1707 * We have a locked page in the page cache, now we need to check
1708 * that it's up-to-date. If not, it is going to be due to an error.
1709 */
1710 if (unlikely(!PageUptodate(page)))
1711 goto page_not_uptodate;
1712
1713 /*
1714 * Found the page and have a reference on it.
1715 * We must recheck i_size under page lock.
1716 */
1717 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1718 if (unlikely(offset >= size)) {
1719 unlock_page(page);
1720 page_cache_release(page);
1721 return VM_FAULT_SIGBUS;
1722 }
1723
1724 vmf->page = page;
1725 return ret | VM_FAULT_LOCKED;
1726
1727no_cached_page:
1728 /*
1729 * We're only likely to ever get here if MADV_RANDOM is in
1730 * effect.
1731 */
1732 error = page_cache_read(file, offset);
1733
1734 /*
1735 * The page we want has now been added to the page cache.
1736 * In the unlikely event that someone removed it in the
1737 * meantime, we'll just come back here and read it again.
1738 */
1739 if (error >= 0)
1740 goto retry_find;
1741
1742 /*
1743 * An error return from page_cache_read can result if the
1744 * system is low on memory, or a problem occurs while trying
1745 * to schedule I/O.
1746 */
1747 if (error == -ENOMEM)
1748 return VM_FAULT_OOM;
1749 return VM_FAULT_SIGBUS;
1750
1751page_not_uptodate:
1752 /*
1753 * Umm, take care of errors if the page isn't up-to-date.
1754 * Try to re-read it _once_. We do this synchronously,
1755 * because there really aren't any performance issues here
1756 * and we need to check for errors.
1757 */
1758 ClearPageError(page);
1759 error = mapping->a_ops->readpage(file, page);
1760 if (!error) {
1761 wait_on_page_locked(page);
1762 if (!PageUptodate(page))
1763 error = -EIO;
1764 }
1765 page_cache_release(page);
1766
1767 if (!error || error == AOP_TRUNCATED_PAGE)
1768 goto retry_find;
1769
1770 /* Things didn't work out. Return zero to tell the mm layer so. */
1771 shrink_readahead_size_eio(file, ra);
1772 return VM_FAULT_SIGBUS;
1773}
1774EXPORT_SYMBOL(filemap_fault);
1775
1776const struct vm_operations_struct generic_file_vm_ops = {
1777 .fault = filemap_fault,
1778};
1779
1780/* This is used for a general mmap of a disk file */
1781
1782int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1783{
1784 struct address_space *mapping = file->f_mapping;
1785
1786 if (!mapping->a_ops->readpage)
1787 return -ENOEXEC;
1788 file_accessed(file);
1789 vma->vm_ops = &generic_file_vm_ops;
1790 vma->vm_flags |= VM_CAN_NONLINEAR;
1791 return 0;
1792}
1793
1794/*
1795 * This is for filesystems which do not implement ->writepage.
1796 */
1797int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1798{
1799 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1800 return -EINVAL;
1801 return generic_file_mmap(file, vma);
1802}
1803#else
1804int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1805{
1806 return -ENOSYS;
1807}
1808int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1809{
1810 return -ENOSYS;
1811}
1812#endif /* CONFIG_MMU */
1813
1814EXPORT_SYMBOL(generic_file_mmap);
1815EXPORT_SYMBOL(generic_file_readonly_mmap);
1816
1817static struct page *__read_cache_page(struct address_space *mapping,
1818 pgoff_t index,
1819 int (*filler)(void *, struct page *),
1820 void *data,
1821 gfp_t gfp)
1822{
1823 struct page *page;
1824 int err;
1825repeat:
1826 page = find_get_page(mapping, index);
1827 if (!page) {
1828 page = __page_cache_alloc(gfp | __GFP_COLD);
1829 if (!page)
1830 return ERR_PTR(-ENOMEM);
1831 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1832 if (unlikely(err)) {
1833 page_cache_release(page);
1834 if (err == -EEXIST)
1835 goto repeat;
1836 /* Presumably ENOMEM for radix tree node */
1837 return ERR_PTR(err);
1838 }
1839 err = filler(data, page);
1840 if (err < 0) {
1841 page_cache_release(page);
1842 page = ERR_PTR(err);
1843 }
1844 }
1845 return page;
1846}
1847
1848static struct page *do_read_cache_page(struct address_space *mapping,
1849 pgoff_t index,
1850 int (*filler)(void *, struct page *),
1851 void *data,
1852 gfp_t gfp)
1853
1854{
1855 struct page *page;
1856 int err;
1857
1858retry:
1859 page = __read_cache_page(mapping, index, filler, data, gfp);
1860 if (IS_ERR(page))
1861 return page;
1862 if (PageUptodate(page))
1863 goto out;
1864
1865 lock_page(page);
1866 if (!page->mapping) {
1867 unlock_page(page);
1868 page_cache_release(page);
1869 goto retry;
1870 }
1871 if (PageUptodate(page)) {
1872 unlock_page(page);
1873 goto out;
1874 }
1875 err = filler(data, page);
1876 if (err < 0) {
1877 page_cache_release(page);
1878 return ERR_PTR(err);
1879 }
1880out:
1881 mark_page_accessed(page);
1882 return page;
1883}
1884
1885/**
1886 * read_cache_page_async - read into page cache, fill it if needed
1887 * @mapping: the page's address_space
1888 * @index: the page index
1889 * @filler: function to perform the read
1890 * @data: first arg to filler(data, page) function, often left as NULL
1891 *
1892 * Same as read_cache_page, but don't wait for page to become unlocked
1893 * after submitting it to the filler.
1894 *
1895 * Read into the page cache. If a page already exists, and PageUptodate() is
1896 * not set, try to fill the page but don't wait for it to become unlocked.
1897 *
1898 * If the page does not get brought uptodate, return -EIO.
1899 */
1900struct page *read_cache_page_async(struct address_space *mapping,
1901 pgoff_t index,
1902 int (*filler)(void *, struct page *),
1903 void *data)
1904{
1905 return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1906}
1907EXPORT_SYMBOL(read_cache_page_async);
1908
1909static struct page *wait_on_page_read(struct page *page)
1910{
1911 if (!IS_ERR(page)) {
1912 wait_on_page_locked(page);
1913 if (!PageUptodate(page)) {
1914 page_cache_release(page);
1915 page = ERR_PTR(-EIO);
1916 }
1917 }
1918 return page;
1919}
1920
1921/**
1922 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1923 * @mapping: the page's address_space
1924 * @index: the page index
1925 * @gfp: the page allocator flags to use if allocating
1926 *
1927 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1928 * any new page allocations done using the specified allocation flags. Note
1929 * that the Radix tree operations will still use GFP_KERNEL, so you can't
1930 * expect to do this atomically or anything like that - but you can pass in
1931 * other page requirements.
1932 *
1933 * If the page does not get brought uptodate, return -EIO.
1934 */
1935struct page *read_cache_page_gfp(struct address_space *mapping,
1936 pgoff_t index,
1937 gfp_t gfp)
1938{
1939 filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1940
1941 return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1942}
1943EXPORT_SYMBOL(read_cache_page_gfp);
1944
1945/**
1946 * read_cache_page - read into page cache, fill it if needed
1947 * @mapping: the page's address_space
1948 * @index: the page index
1949 * @filler: function to perform the read
1950 * @data: first arg to filler(data, page) function, often left as NULL
1951 *
1952 * Read into the page cache. If a page already exists, and PageUptodate() is
1953 * not set, try to fill the page then wait for it to become unlocked.
1954 *
1955 * If the page does not get brought uptodate, return -EIO.
1956 */
1957struct page *read_cache_page(struct address_space *mapping,
1958 pgoff_t index,
1959 int (*filler)(void *, struct page *),
1960 void *data)
1961{
1962 return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1963}
1964EXPORT_SYMBOL(read_cache_page);
1965
1966/*
1967 * The logic we want is
1968 *
1969 * if suid or (sgid and xgrp)
1970 * remove privs
1971 */
1972int should_remove_suid(struct dentry *dentry)
1973{
1974 mode_t mode = dentry->d_inode->i_mode;
1975 int kill = 0;
1976
1977 /* suid always must be killed */
1978 if (unlikely(mode & S_ISUID))
1979 kill = ATTR_KILL_SUID;
1980
1981 /*
1982 * sgid without any exec bits is just a mandatory locking mark; leave
1983 * it alone. If some exec bits are set, it's a real sgid; kill it.
1984 */
1985 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1986 kill |= ATTR_KILL_SGID;
1987
1988 if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1989 return kill;
1990
1991 return 0;
1992}
1993EXPORT_SYMBOL(should_remove_suid);
1994
1995static int __remove_suid(struct dentry *dentry, int kill)
1996{
1997 struct iattr newattrs;
1998
1999 newattrs.ia_valid = ATTR_FORCE | kill;
2000 return notify_change(dentry, &newattrs);
2001}
2002
2003int file_remove_suid(struct file *file)
2004{
2005 struct dentry *dentry = file->f_path.dentry;
2006 struct inode *inode = dentry->d_inode;
2007 int killsuid;
2008 int killpriv;
2009 int error = 0;
2010
2011 /* Fast path for nothing security related */
2012 if (IS_NOSEC(inode))
2013 return 0;
2014
2015 killsuid = should_remove_suid(dentry);
2016 killpriv = security_inode_need_killpriv(dentry);
2017
2018 if (killpriv < 0)
2019 return killpriv;
2020 if (killpriv)
2021 error = security_inode_killpriv(dentry);
2022 if (!error && killsuid)
2023 error = __remove_suid(dentry, killsuid);
2024 if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2025 inode->i_flags |= S_NOSEC;
2026
2027 return error;
2028}
2029EXPORT_SYMBOL(file_remove_suid);
2030
2031static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2032 const struct iovec *iov, size_t base, size_t bytes)
2033{
2034 size_t copied = 0, left = 0;
2035
2036 while (bytes) {
2037 char __user *buf = iov->iov_base + base;
2038 int copy = min(bytes, iov->iov_len - base);
2039
2040 base = 0;
2041 left = __copy_from_user_inatomic(vaddr, buf, copy);
2042 copied += copy;
2043 bytes -= copy;
2044 vaddr += copy;
2045 iov++;
2046
2047 if (unlikely(left))
2048 break;
2049 }
2050 return copied - left;
2051}
2052
2053/*
2054 * Copy as much as we can into the page and return the number of bytes which
2055 * were successfully copied. If a fault is encountered then return the number of
2056 * bytes which were copied.
2057 */
2058size_t iov_iter_copy_from_user_atomic(struct page *page,
2059 struct iov_iter *i, unsigned long offset, size_t bytes)
2060{
2061 char *kaddr;
2062 size_t copied;
2063
2064 BUG_ON(!in_atomic());
2065 kaddr = kmap_atomic(page, KM_USER0);
2066 if (likely(i->nr_segs == 1)) {
2067 int left;
2068 char __user *buf = i->iov->iov_base + i->iov_offset;
2069 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2070 copied = bytes - left;
2071 } else {
2072 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2073 i->iov, i->iov_offset, bytes);
2074 }
2075 kunmap_atomic(kaddr, KM_USER0);
2076
2077 return copied;
2078}
2079EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2080
2081/*
2082 * This has the same sideeffects and return value as
2083 * iov_iter_copy_from_user_atomic().
2084 * The difference is that it attempts to resolve faults.
2085 * Page must not be locked.
2086 */
2087size_t iov_iter_copy_from_user(struct page *page,
2088 struct iov_iter *i, unsigned long offset, size_t bytes)
2089{
2090 char *kaddr;
2091 size_t copied;
2092
2093 kaddr = kmap(page);
2094 if (likely(i->nr_segs == 1)) {
2095 int left;
2096 char __user *buf = i->iov->iov_base + i->iov_offset;
2097 left = __copy_from_user(kaddr + offset, buf, bytes);
2098 copied = bytes - left;
2099 } else {
2100 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2101 i->iov, i->iov_offset, bytes);
2102 }
2103 kunmap(page);
2104 return copied;
2105}
2106EXPORT_SYMBOL(iov_iter_copy_from_user);
2107
2108void iov_iter_advance(struct iov_iter *i, size_t bytes)
2109{
2110 BUG_ON(i->count < bytes);
2111
2112 if (likely(i->nr_segs == 1)) {
2113 i->iov_offset += bytes;
2114 i->count -= bytes;
2115 } else {
2116 const struct iovec *iov = i->iov;
2117 size_t base = i->iov_offset;
2118
2119 /*
2120 * The !iov->iov_len check ensures we skip over unlikely
2121 * zero-length segments (without overruning the iovec).
2122 */
2123 while (bytes || unlikely(i->count && !iov->iov_len)) {
2124 int copy;
2125
2126 copy = min(bytes, iov->iov_len - base);
2127 BUG_ON(!i->count || i->count < copy);
2128 i->count -= copy;
2129 bytes -= copy;
2130 base += copy;
2131 if (iov->iov_len == base) {
2132 iov++;
2133 base = 0;
2134 }
2135 }
2136 i->iov = iov;
2137 i->iov_offset = base;
2138 }
2139}
2140EXPORT_SYMBOL(iov_iter_advance);
2141
2142/*
2143 * Fault in the first iovec of the given iov_iter, to a maximum length
2144 * of bytes. Returns 0 on success, or non-zero if the memory could not be
2145 * accessed (ie. because it is an invalid address).
2146 *
2147 * writev-intensive code may want this to prefault several iovecs -- that
2148 * would be possible (callers must not rely on the fact that _only_ the
2149 * first iovec will be faulted with the current implementation).
2150 */
2151int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2152{
2153 char __user *buf = i->iov->iov_base + i->iov_offset;
2154 bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2155 return fault_in_pages_readable(buf, bytes);
2156}
2157EXPORT_SYMBOL(iov_iter_fault_in_readable);
2158
2159/*
2160 * Return the count of just the current iov_iter segment.
2161 */
2162size_t iov_iter_single_seg_count(struct iov_iter *i)
2163{
2164 const struct iovec *iov = i->iov;
2165 if (i->nr_segs == 1)
2166 return i->count;
2167 else
2168 return min(i->count, iov->iov_len - i->iov_offset);
2169}
2170EXPORT_SYMBOL(iov_iter_single_seg_count);
2171
2172/*
2173 * Performs necessary checks before doing a write
2174 *
2175 * Can adjust writing position or amount of bytes to write.
2176 * Returns appropriate error code that caller should return or
2177 * zero in case that write should be allowed.
2178 */
2179inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2180{
2181 struct inode *inode = file->f_mapping->host;
2182 unsigned long limit = rlimit(RLIMIT_FSIZE);
2183
2184 if (unlikely(*pos < 0))
2185 return -EINVAL;
2186
2187 if (!isblk) {
2188 /* FIXME: this is for backwards compatibility with 2.4 */
2189 if (file->f_flags & O_APPEND)
2190 *pos = i_size_read(inode);
2191
2192 if (limit != RLIM_INFINITY) {
2193 if (*pos >= limit) {
2194 send_sig(SIGXFSZ, current, 0);
2195 return -EFBIG;
2196 }
2197 if (*count > limit - (typeof(limit))*pos) {
2198 *count = limit - (typeof(limit))*pos;
2199 }
2200 }
2201 }
2202
2203 /*
2204 * LFS rule
2205 */
2206 if (unlikely(*pos + *count > MAX_NON_LFS &&
2207 !(file->f_flags & O_LARGEFILE))) {
2208 if (*pos >= MAX_NON_LFS) {
2209 return -EFBIG;
2210 }
2211 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2212 *count = MAX_NON_LFS - (unsigned long)*pos;
2213 }
2214 }
2215
2216 /*
2217 * Are we about to exceed the fs block limit ?
2218 *
2219 * If we have written data it becomes a short write. If we have
2220 * exceeded without writing data we send a signal and return EFBIG.
2221 * Linus frestrict idea will clean these up nicely..
2222 */
2223 if (likely(!isblk)) {
2224 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2225 if (*count || *pos > inode->i_sb->s_maxbytes) {
2226 return -EFBIG;
2227 }
2228 /* zero-length writes at ->s_maxbytes are OK */
2229 }
2230
2231 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2232 *count = inode->i_sb->s_maxbytes - *pos;
2233 } else {
2234#ifdef CONFIG_BLOCK
2235 loff_t isize;
2236 if (bdev_read_only(I_BDEV(inode)))
2237 return -EPERM;
2238 isize = i_size_read(inode);
2239 if (*pos >= isize) {
2240 if (*count || *pos > isize)
2241 return -ENOSPC;
2242 }
2243
2244 if (*pos + *count > isize)
2245 *count = isize - *pos;
2246#else
2247 return -EPERM;
2248#endif
2249 }
2250 return 0;
2251}
2252EXPORT_SYMBOL(generic_write_checks);
2253
2254int pagecache_write_begin(struct file *file, struct address_space *mapping,
2255 loff_t pos, unsigned len, unsigned flags,
2256 struct page **pagep, void **fsdata)
2257{
2258 const struct address_space_operations *aops = mapping->a_ops;
2259
2260 return aops->write_begin(file, mapping, pos, len, flags,
2261 pagep, fsdata);
2262}
2263EXPORT_SYMBOL(pagecache_write_begin);
2264
2265int pagecache_write_end(struct file *file, struct address_space *mapping,
2266 loff_t pos, unsigned len, unsigned copied,
2267 struct page *page, void *fsdata)
2268{
2269 const struct address_space_operations *aops = mapping->a_ops;
2270
2271 mark_page_accessed(page);
2272 return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2273}
2274EXPORT_SYMBOL(pagecache_write_end);
2275
2276ssize_t
2277generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2278 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2279 size_t count, size_t ocount)
2280{
2281 struct file *file = iocb->ki_filp;
2282 struct address_space *mapping = file->f_mapping;
2283 struct inode *inode = mapping->host;
2284 ssize_t written;
2285 size_t write_len;
2286 pgoff_t end;
2287
2288 if (count != ocount)
2289 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2290
2291 write_len = iov_length(iov, *nr_segs);
2292 end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2293
2294 written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2295 if (written)
2296 goto out;
2297
2298 /*
2299 * After a write we want buffered reads to be sure to go to disk to get
2300 * the new data. We invalidate clean cached page from the region we're
2301 * about to write. We do this *before* the write so that we can return
2302 * without clobbering -EIOCBQUEUED from ->direct_IO().
2303 */
2304 if (mapping->nrpages) {
2305 written = invalidate_inode_pages2_range(mapping,
2306 pos >> PAGE_CACHE_SHIFT, end);
2307 /*
2308 * If a page can not be invalidated, return 0 to fall back
2309 * to buffered write.
2310 */
2311 if (written) {
2312 if (written == -EBUSY)
2313 return 0;
2314 goto out;
2315 }
2316 }
2317
2318 written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2319
2320 /*
2321 * Finally, try again to invalidate clean pages which might have been
2322 * cached by non-direct readahead, or faulted in by get_user_pages()
2323 * if the source of the write was an mmap'ed region of the file
2324 * we're writing. Either one is a pretty crazy thing to do,
2325 * so we don't support it 100%. If this invalidation
2326 * fails, tough, the write still worked...
2327 */
2328 if (mapping->nrpages) {
2329 invalidate_inode_pages2_range(mapping,
2330 pos >> PAGE_CACHE_SHIFT, end);
2331 }
2332
2333 if (written > 0) {
2334 pos += written;
2335 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2336 i_size_write(inode, pos);
2337 mark_inode_dirty(inode);
2338 }
2339 *ppos = pos;
2340 }
2341out:
2342 return written;
2343}
2344EXPORT_SYMBOL(generic_file_direct_write);
2345
2346/*
2347 * Find or create a page at the given pagecache position. Return the locked
2348 * page. This function is specifically for buffered writes.
2349 */
2350struct page *grab_cache_page_write_begin(struct address_space *mapping,
2351 pgoff_t index, unsigned flags)
2352{
2353 int status;
2354 struct page *page;
2355 gfp_t gfp_notmask = 0;
2356 if (flags & AOP_FLAG_NOFS)
2357 gfp_notmask = __GFP_FS;
2358repeat:
2359 page = find_lock_page(mapping, index);
2360 if (page)
2361 goto found;
2362
2363 page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2364 if (!page)
2365 return NULL;
2366 status = add_to_page_cache_lru(page, mapping, index,
2367 GFP_KERNEL & ~gfp_notmask);
2368 if (unlikely(status)) {
2369 page_cache_release(page);
2370 if (status == -EEXIST)
2371 goto repeat;
2372 return NULL;
2373 }
2374found:
2375 wait_on_page_writeback(page);
2376 return page;
2377}
2378EXPORT_SYMBOL(grab_cache_page_write_begin);
2379
2380static ssize_t generic_perform_write(struct file *file,
2381 struct iov_iter *i, loff_t pos)
2382{
2383 struct address_space *mapping = file->f_mapping;
2384 const struct address_space_operations *a_ops = mapping->a_ops;
2385 long status = 0;
2386 ssize_t written = 0;
2387 unsigned int flags = 0;
2388
2389 /*
2390 * Copies from kernel address space cannot fail (NFSD is a big user).
2391 */
2392 if (segment_eq(get_fs(), KERNEL_DS))
2393 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2394
2395 do {
2396 struct page *page;
2397 unsigned long offset; /* Offset into pagecache page */
2398 unsigned long bytes; /* Bytes to write to page */
2399 size_t copied; /* Bytes copied from user */
2400 void *fsdata;
2401
2402 offset = (pos & (PAGE_CACHE_SIZE - 1));
2403 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2404 iov_iter_count(i));
2405
2406again:
2407
2408 /*
2409 * Bring in the user page that we will copy from _first_.
2410 * Otherwise there's a nasty deadlock on copying from the
2411 * same page as we're writing to, without it being marked
2412 * up-to-date.
2413 *
2414 * Not only is this an optimisation, but it is also required
2415 * to check that the address is actually valid, when atomic
2416 * usercopies are used, below.
2417 */
2418 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2419 status = -EFAULT;
2420 break;
2421 }
2422
2423 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2424 &page, &fsdata);
2425 if (unlikely(status))
2426 break;
2427
2428 if (mapping_writably_mapped(mapping))
2429 flush_dcache_page(page);
2430
2431 pagefault_disable();
2432 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2433 pagefault_enable();
2434 flush_dcache_page(page);
2435
2436 mark_page_accessed(page);
2437 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2438 page, fsdata);
2439 if (unlikely(status < 0))
2440 break;
2441 copied = status;
2442
2443 cond_resched();
2444
2445 iov_iter_advance(i, copied);
2446 if (unlikely(copied == 0)) {
2447 /*
2448 * If we were unable to copy any data at all, we must
2449 * fall back to a single segment length write.
2450 *
2451 * If we didn't fallback here, we could livelock
2452 * because not all segments in the iov can be copied at
2453 * once without a pagefault.
2454 */
2455 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2456 iov_iter_single_seg_count(i));
2457 goto again;
2458 }
2459 pos += copied;
2460 written += copied;
2461
2462 balance_dirty_pages_ratelimited(mapping);
2463
2464 } while (iov_iter_count(i));
2465
2466 return written ? written : status;
2467}
2468
2469ssize_t
2470generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2471 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2472 size_t count, ssize_t written)
2473{
2474 struct file *file = iocb->ki_filp;
2475 ssize_t status;
2476 struct iov_iter i;
2477
2478 iov_iter_init(&i, iov, nr_segs, count, written);
2479 status = generic_perform_write(file, &i, pos);
2480
2481 if (likely(status >= 0)) {
2482 written += status;
2483 *ppos = pos + status;
2484 }
2485
2486 return written ? written : status;
2487}
2488EXPORT_SYMBOL(generic_file_buffered_write);
2489
2490/**
2491 * __generic_file_aio_write - write data to a file
2492 * @iocb: IO state structure (file, offset, etc.)
2493 * @iov: vector with data to write
2494 * @nr_segs: number of segments in the vector
2495 * @ppos: position where to write
2496 *
2497 * This function does all the work needed for actually writing data to a
2498 * file. It does all basic checks, removes SUID from the file, updates
2499 * modification times and calls proper subroutines depending on whether we
2500 * do direct IO or a standard buffered write.
2501 *
2502 * It expects i_mutex to be grabbed unless we work on a block device or similar
2503 * object which does not need locking at all.
2504 *
2505 * This function does *not* take care of syncing data in case of O_SYNC write.
2506 * A caller has to handle it. This is mainly due to the fact that we want to
2507 * avoid syncing under i_mutex.
2508 */
2509ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2510 unsigned long nr_segs, loff_t *ppos)
2511{
2512 struct file *file = iocb->ki_filp;
2513 struct address_space * mapping = file->f_mapping;
2514 size_t ocount; /* original count */
2515 size_t count; /* after file limit checks */
2516 struct inode *inode = mapping->host;
2517 loff_t pos;
2518 ssize_t written;
2519 ssize_t err;
2520
2521 ocount = 0;
2522 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2523 if (err)
2524 return err;
2525
2526 count = ocount;
2527 pos = *ppos;
2528
2529 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2530
2531 /* We can write back this queue in page reclaim */
2532 current->backing_dev_info = mapping->backing_dev_info;
2533 written = 0;
2534
2535 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2536 if (err)
2537 goto out;
2538
2539 if (count == 0)
2540 goto out;
2541
2542 err = file_remove_suid(file);
2543 if (err)
2544 goto out;
2545
2546 file_update_time(file);
2547
2548 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2549 if (unlikely(file->f_flags & O_DIRECT)) {
2550 loff_t endbyte;
2551 ssize_t written_buffered;
2552
2553 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2554 ppos, count, ocount);
2555 if (written < 0 || written == count)
2556 goto out;
2557 /*
2558 * direct-io write to a hole: fall through to buffered I/O
2559 * for completing the rest of the request.
2560 */
2561 pos += written;
2562 count -= written;
2563 written_buffered = generic_file_buffered_write(iocb, iov,
2564 nr_segs, pos, ppos, count,
2565 written);
2566 /*
2567 * If generic_file_buffered_write() retuned a synchronous error
2568 * then we want to return the number of bytes which were
2569 * direct-written, or the error code if that was zero. Note
2570 * that this differs from normal direct-io semantics, which
2571 * will return -EFOO even if some bytes were written.
2572 */
2573 if (written_buffered < 0) {
2574 err = written_buffered;
2575 goto out;
2576 }
2577
2578 /*
2579 * We need to ensure that the page cache pages are written to
2580 * disk and invalidated to preserve the expected O_DIRECT
2581 * semantics.
2582 */
2583 endbyte = pos + written_buffered - written - 1;
2584 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2585 if (err == 0) {
2586 written = written_buffered;
2587 invalidate_mapping_pages(mapping,
2588 pos >> PAGE_CACHE_SHIFT,
2589 endbyte >> PAGE_CACHE_SHIFT);
2590 } else {
2591 /*
2592 * We don't know how much we wrote, so just return
2593 * the number of bytes which were direct-written
2594 */
2595 }
2596 } else {
2597 written = generic_file_buffered_write(iocb, iov, nr_segs,
2598 pos, ppos, count, written);
2599 }
2600out:
2601 current->backing_dev_info = NULL;
2602 return written ? written : err;
2603}
2604EXPORT_SYMBOL(__generic_file_aio_write);
2605
2606/**
2607 * generic_file_aio_write - write data to a file
2608 * @iocb: IO state structure
2609 * @iov: vector with data to write
2610 * @nr_segs: number of segments in the vector
2611 * @pos: position in file where to write
2612 *
2613 * This is a wrapper around __generic_file_aio_write() to be used by most
2614 * filesystems. It takes care of syncing the file in case of O_SYNC file
2615 * and acquires i_mutex as needed.
2616 */
2617ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2618 unsigned long nr_segs, loff_t pos)
2619{
2620 struct file *file = iocb->ki_filp;
2621 struct inode *inode = file->f_mapping->host;
2622 struct blk_plug plug;
2623 ssize_t ret;
2624
2625 BUG_ON(iocb->ki_pos != pos);
2626
2627 mutex_lock(&inode->i_mutex);
2628 blk_start_plug(&plug);
2629 ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2630 mutex_unlock(&inode->i_mutex);
2631
2632 if (ret > 0 || ret == -EIOCBQUEUED) {
2633 ssize_t err;
2634
2635 err = generic_write_sync(file, pos, ret);
2636 if (err < 0 && ret > 0)
2637 ret = err;
2638 }
2639 blk_finish_plug(&plug);
2640 return ret;
2641}
2642EXPORT_SYMBOL(generic_file_aio_write);
2643
2644/**
2645 * try_to_release_page() - release old fs-specific metadata on a page
2646 *
2647 * @page: the page which the kernel is trying to free
2648 * @gfp_mask: memory allocation flags (and I/O mode)
2649 *
2650 * The address_space is to try to release any data against the page
2651 * (presumably at page->private). If the release was successful, return `1'.
2652 * Otherwise return zero.
2653 *
2654 * This may also be called if PG_fscache is set on a page, indicating that the
2655 * page is known to the local caching routines.
2656 *
2657 * The @gfp_mask argument specifies whether I/O may be performed to release
2658 * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2659 *
2660 */
2661int try_to_release_page(struct page *page, gfp_t gfp_mask)
2662{
2663 struct address_space * const mapping = page->mapping;
2664
2665 BUG_ON(!PageLocked(page));
2666 if (PageWriteback(page))
2667 return 0;
2668
2669 if (mapping && mapping->a_ops->releasepage)
2670 return mapping->a_ops->releasepage(page, gfp_mask);
2671 return try_to_free_buffers(page);
2672}
2673
2674EXPORT_SYMBOL(try_to_release_page);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/filemap.c
4 *
5 * Copyright (C) 1994-1999 Linus Torvalds
6 */
7
8/*
9 * This file handles the generic file mmap semantics used by
10 * most "normal" filesystems (but you don't /have/ to use this:
11 * the NFS filesystem used to do this differently, for example)
12 */
13#include <linux/export.h>
14#include <linux/compiler.h>
15#include <linux/dax.h>
16#include <linux/fs.h>
17#include <linux/sched/signal.h>
18#include <linux/uaccess.h>
19#include <linux/capability.h>
20#include <linux/kernel_stat.h>
21#include <linux/gfp.h>
22#include <linux/mm.h>
23#include <linux/swap.h>
24#include <linux/swapops.h>
25#include <linux/syscalls.h>
26#include <linux/mman.h>
27#include <linux/pagemap.h>
28#include <linux/file.h>
29#include <linux/uio.h>
30#include <linux/error-injection.h>
31#include <linux/hash.h>
32#include <linux/writeback.h>
33#include <linux/backing-dev.h>
34#include <linux/pagevec.h>
35#include <linux/security.h>
36#include <linux/cpuset.h>
37#include <linux/hugetlb.h>
38#include <linux/memcontrol.h>
39#include <linux/shmem_fs.h>
40#include <linux/rmap.h>
41#include <linux/delayacct.h>
42#include <linux/psi.h>
43#include <linux/ramfs.h>
44#include <linux/page_idle.h>
45#include <linux/migrate.h>
46#include <linux/pipe_fs_i.h>
47#include <linux/splice.h>
48#include <linux/rcupdate_wait.h>
49#include <linux/sched/mm.h>
50#include <asm/pgalloc.h>
51#include <asm/tlbflush.h>
52#include "internal.h"
53
54#define CREATE_TRACE_POINTS
55#include <trace/events/filemap.h>
56
57/*
58 * FIXME: remove all knowledge of the buffer layer from the core VM
59 */
60#include <linux/buffer_head.h> /* for try_to_free_buffers */
61
62#include <asm/mman.h>
63
64#include "swap.h"
65
66/*
67 * Shared mappings implemented 30.11.1994. It's not fully working yet,
68 * though.
69 *
70 * Shared mappings now work. 15.8.1995 Bruno.
71 *
72 * finished 'unifying' the page and buffer cache and SMP-threaded the
73 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
74 *
75 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
76 */
77
78/*
79 * Lock ordering:
80 *
81 * ->i_mmap_rwsem (truncate_pagecache)
82 * ->private_lock (__free_pte->block_dirty_folio)
83 * ->swap_lock (exclusive_swap_page, others)
84 * ->i_pages lock
85 *
86 * ->i_rwsem
87 * ->invalidate_lock (acquired by fs in truncate path)
88 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
89 *
90 * ->mmap_lock
91 * ->i_mmap_rwsem
92 * ->page_table_lock or pte_lock (various, mainly in memory.c)
93 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
94 *
95 * ->mmap_lock
96 * ->invalidate_lock (filemap_fault)
97 * ->lock_page (filemap_fault, access_process_vm)
98 *
99 * ->i_rwsem (generic_perform_write)
100 * ->mmap_lock (fault_in_readable->do_page_fault)
101 *
102 * bdi->wb.list_lock
103 * sb_lock (fs/fs-writeback.c)
104 * ->i_pages lock (__sync_single_inode)
105 *
106 * ->i_mmap_rwsem
107 * ->anon_vma.lock (vma_merge)
108 *
109 * ->anon_vma.lock
110 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
111 *
112 * ->page_table_lock or pte_lock
113 * ->swap_lock (try_to_unmap_one)
114 * ->private_lock (try_to_unmap_one)
115 * ->i_pages lock (try_to_unmap_one)
116 * ->lruvec->lru_lock (follow_page_mask->mark_page_accessed)
117 * ->lruvec->lru_lock (check_pte_range->folio_isolate_lru)
118 * ->private_lock (folio_remove_rmap_pte->set_page_dirty)
119 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty)
120 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty)
121 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty)
122 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
123 * ->inode->i_lock (zap_pte_range->set_page_dirty)
124 * ->private_lock (zap_pte_range->block_dirty_folio)
125 */
126
127static void page_cache_delete(struct address_space *mapping,
128 struct folio *folio, void *shadow)
129{
130 XA_STATE(xas, &mapping->i_pages, folio->index);
131 long nr = 1;
132
133 mapping_set_update(&xas, mapping);
134
135 xas_set_order(&xas, folio->index, folio_order(folio));
136 nr = folio_nr_pages(folio);
137
138 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
139
140 xas_store(&xas, shadow);
141 xas_init_marks(&xas);
142
143 folio->mapping = NULL;
144 /* Leave page->index set: truncation lookup relies upon it */
145 mapping->nrpages -= nr;
146}
147
148static void filemap_unaccount_folio(struct address_space *mapping,
149 struct folio *folio)
150{
151 long nr;
152
153 VM_BUG_ON_FOLIO(folio_mapped(folio), folio);
154 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) {
155 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
156 current->comm, folio_pfn(folio));
157 dump_page(&folio->page, "still mapped when deleted");
158 dump_stack();
159 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
160
161 if (mapping_exiting(mapping) && !folio_test_large(folio)) {
162 int mapcount = folio_mapcount(folio);
163
164 if (folio_ref_count(folio) >= mapcount + 2) {
165 /*
166 * All vmas have already been torn down, so it's
167 * a good bet that actually the page is unmapped
168 * and we'd rather not leak it: if we're wrong,
169 * another bad page check should catch it later.
170 */
171 atomic_set(&folio->_mapcount, -1);
172 folio_ref_sub(folio, mapcount);
173 }
174 }
175 }
176
177 /* hugetlb folios do not participate in page cache accounting. */
178 if (folio_test_hugetlb(folio))
179 return;
180
181 nr = folio_nr_pages(folio);
182
183 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
184 if (folio_test_swapbacked(folio)) {
185 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
186 if (folio_test_pmd_mappable(folio))
187 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr);
188 } else if (folio_test_pmd_mappable(folio)) {
189 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr);
190 filemap_nr_thps_dec(mapping);
191 }
192
193 /*
194 * At this point folio must be either written or cleaned by
195 * truncate. Dirty folio here signals a bug and loss of
196 * unwritten data - on ordinary filesystems.
197 *
198 * But it's harmless on in-memory filesystems like tmpfs; and can
199 * occur when a driver which did get_user_pages() sets page dirty
200 * before putting it, while the inode is being finally evicted.
201 *
202 * Below fixes dirty accounting after removing the folio entirely
203 * but leaves the dirty flag set: it has no effect for truncated
204 * folio and anyway will be cleared before returning folio to
205 * buddy allocator.
206 */
207 if (WARN_ON_ONCE(folio_test_dirty(folio) &&
208 mapping_can_writeback(mapping)))
209 folio_account_cleaned(folio, inode_to_wb(mapping->host));
210}
211
212/*
213 * Delete a page from the page cache and free it. Caller has to make
214 * sure the page is locked and that nobody else uses it - or that usage
215 * is safe. The caller must hold the i_pages lock.
216 */
217void __filemap_remove_folio(struct folio *folio, void *shadow)
218{
219 struct address_space *mapping = folio->mapping;
220
221 trace_mm_filemap_delete_from_page_cache(folio);
222 filemap_unaccount_folio(mapping, folio);
223 page_cache_delete(mapping, folio, shadow);
224}
225
226void filemap_free_folio(struct address_space *mapping, struct folio *folio)
227{
228 void (*free_folio)(struct folio *);
229 int refs = 1;
230
231 free_folio = mapping->a_ops->free_folio;
232 if (free_folio)
233 free_folio(folio);
234
235 if (folio_test_large(folio))
236 refs = folio_nr_pages(folio);
237 folio_put_refs(folio, refs);
238}
239
240/**
241 * filemap_remove_folio - Remove folio from page cache.
242 * @folio: The folio.
243 *
244 * This must be called only on folios that are locked and have been
245 * verified to be in the page cache. It will never put the folio into
246 * the free list because the caller has a reference on the page.
247 */
248void filemap_remove_folio(struct folio *folio)
249{
250 struct address_space *mapping = folio->mapping;
251
252 BUG_ON(!folio_test_locked(folio));
253 spin_lock(&mapping->host->i_lock);
254 xa_lock_irq(&mapping->i_pages);
255 __filemap_remove_folio(folio, NULL);
256 xa_unlock_irq(&mapping->i_pages);
257 if (mapping_shrinkable(mapping))
258 inode_add_lru(mapping->host);
259 spin_unlock(&mapping->host->i_lock);
260
261 filemap_free_folio(mapping, folio);
262}
263
264/*
265 * page_cache_delete_batch - delete several folios from page cache
266 * @mapping: the mapping to which folios belong
267 * @fbatch: batch of folios to delete
268 *
269 * The function walks over mapping->i_pages and removes folios passed in
270 * @fbatch from the mapping. The function expects @fbatch to be sorted
271 * by page index and is optimised for it to be dense.
272 * It tolerates holes in @fbatch (mapping entries at those indices are not
273 * modified).
274 *
275 * The function expects the i_pages lock to be held.
276 */
277static void page_cache_delete_batch(struct address_space *mapping,
278 struct folio_batch *fbatch)
279{
280 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index);
281 long total_pages = 0;
282 int i = 0;
283 struct folio *folio;
284
285 mapping_set_update(&xas, mapping);
286 xas_for_each(&xas, folio, ULONG_MAX) {
287 if (i >= folio_batch_count(fbatch))
288 break;
289
290 /* A swap/dax/shadow entry got inserted? Skip it. */
291 if (xa_is_value(folio))
292 continue;
293 /*
294 * A page got inserted in our range? Skip it. We have our
295 * pages locked so they are protected from being removed.
296 * If we see a page whose index is higher than ours, it
297 * means our page has been removed, which shouldn't be
298 * possible because we're holding the PageLock.
299 */
300 if (folio != fbatch->folios[i]) {
301 VM_BUG_ON_FOLIO(folio->index >
302 fbatch->folios[i]->index, folio);
303 continue;
304 }
305
306 WARN_ON_ONCE(!folio_test_locked(folio));
307
308 folio->mapping = NULL;
309 /* Leave folio->index set: truncation lookup relies on it */
310
311 i++;
312 xas_store(&xas, NULL);
313 total_pages += folio_nr_pages(folio);
314 }
315 mapping->nrpages -= total_pages;
316}
317
318void delete_from_page_cache_batch(struct address_space *mapping,
319 struct folio_batch *fbatch)
320{
321 int i;
322
323 if (!folio_batch_count(fbatch))
324 return;
325
326 spin_lock(&mapping->host->i_lock);
327 xa_lock_irq(&mapping->i_pages);
328 for (i = 0; i < folio_batch_count(fbatch); i++) {
329 struct folio *folio = fbatch->folios[i];
330
331 trace_mm_filemap_delete_from_page_cache(folio);
332 filemap_unaccount_folio(mapping, folio);
333 }
334 page_cache_delete_batch(mapping, fbatch);
335 xa_unlock_irq(&mapping->i_pages);
336 if (mapping_shrinkable(mapping))
337 inode_add_lru(mapping->host);
338 spin_unlock(&mapping->host->i_lock);
339
340 for (i = 0; i < folio_batch_count(fbatch); i++)
341 filemap_free_folio(mapping, fbatch->folios[i]);
342}
343
344int filemap_check_errors(struct address_space *mapping)
345{
346 int ret = 0;
347 /* Check for outstanding write errors */
348 if (test_bit(AS_ENOSPC, &mapping->flags) &&
349 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
350 ret = -ENOSPC;
351 if (test_bit(AS_EIO, &mapping->flags) &&
352 test_and_clear_bit(AS_EIO, &mapping->flags))
353 ret = -EIO;
354 return ret;
355}
356EXPORT_SYMBOL(filemap_check_errors);
357
358static int filemap_check_and_keep_errors(struct address_space *mapping)
359{
360 /* Check for outstanding write errors */
361 if (test_bit(AS_EIO, &mapping->flags))
362 return -EIO;
363 if (test_bit(AS_ENOSPC, &mapping->flags))
364 return -ENOSPC;
365 return 0;
366}
367
368/**
369 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range
370 * @mapping: address space structure to write
371 * @wbc: the writeback_control controlling the writeout
372 *
373 * Call writepages on the mapping using the provided wbc to control the
374 * writeout.
375 *
376 * Return: %0 on success, negative error code otherwise.
377 */
378int filemap_fdatawrite_wbc(struct address_space *mapping,
379 struct writeback_control *wbc)
380{
381 int ret;
382
383 if (!mapping_can_writeback(mapping) ||
384 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
385 return 0;
386
387 wbc_attach_fdatawrite_inode(wbc, mapping->host);
388 ret = do_writepages(mapping, wbc);
389 wbc_detach_inode(wbc);
390 return ret;
391}
392EXPORT_SYMBOL(filemap_fdatawrite_wbc);
393
394/**
395 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
396 * @mapping: address space structure to write
397 * @start: offset in bytes where the range starts
398 * @end: offset in bytes where the range ends (inclusive)
399 * @sync_mode: enable synchronous operation
400 *
401 * Start writeback against all of a mapping's dirty pages that lie
402 * within the byte offsets <start, end> inclusive.
403 *
404 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
405 * opposed to a regular memory cleansing writeback. The difference between
406 * these two operations is that if a dirty page/buffer is encountered, it must
407 * be waited upon, and not just skipped over.
408 *
409 * Return: %0 on success, negative error code otherwise.
410 */
411int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
412 loff_t end, int sync_mode)
413{
414 struct writeback_control wbc = {
415 .sync_mode = sync_mode,
416 .nr_to_write = LONG_MAX,
417 .range_start = start,
418 .range_end = end,
419 };
420
421 return filemap_fdatawrite_wbc(mapping, &wbc);
422}
423
424static inline int __filemap_fdatawrite(struct address_space *mapping,
425 int sync_mode)
426{
427 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
428}
429
430int filemap_fdatawrite(struct address_space *mapping)
431{
432 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
433}
434EXPORT_SYMBOL(filemap_fdatawrite);
435
436int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
437 loff_t end)
438{
439 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
440}
441EXPORT_SYMBOL(filemap_fdatawrite_range);
442
443/**
444 * filemap_flush - mostly a non-blocking flush
445 * @mapping: target address_space
446 *
447 * This is a mostly non-blocking flush. Not suitable for data-integrity
448 * purposes - I/O may not be started against all dirty pages.
449 *
450 * Return: %0 on success, negative error code otherwise.
451 */
452int filemap_flush(struct address_space *mapping)
453{
454 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
455}
456EXPORT_SYMBOL(filemap_flush);
457
458/**
459 * filemap_range_has_page - check if a page exists in range.
460 * @mapping: address space within which to check
461 * @start_byte: offset in bytes where the range starts
462 * @end_byte: offset in bytes where the range ends (inclusive)
463 *
464 * Find at least one page in the range supplied, usually used to check if
465 * direct writing in this range will trigger a writeback.
466 *
467 * Return: %true if at least one page exists in the specified range,
468 * %false otherwise.
469 */
470bool filemap_range_has_page(struct address_space *mapping,
471 loff_t start_byte, loff_t end_byte)
472{
473 struct folio *folio;
474 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
475 pgoff_t max = end_byte >> PAGE_SHIFT;
476
477 if (end_byte < start_byte)
478 return false;
479
480 rcu_read_lock();
481 for (;;) {
482 folio = xas_find(&xas, max);
483 if (xas_retry(&xas, folio))
484 continue;
485 /* Shadow entries don't count */
486 if (xa_is_value(folio))
487 continue;
488 /*
489 * We don't need to try to pin this page; we're about to
490 * release the RCU lock anyway. It is enough to know that
491 * there was a page here recently.
492 */
493 break;
494 }
495 rcu_read_unlock();
496
497 return folio != NULL;
498}
499EXPORT_SYMBOL(filemap_range_has_page);
500
501static void __filemap_fdatawait_range(struct address_space *mapping,
502 loff_t start_byte, loff_t end_byte)
503{
504 pgoff_t index = start_byte >> PAGE_SHIFT;
505 pgoff_t end = end_byte >> PAGE_SHIFT;
506 struct folio_batch fbatch;
507 unsigned nr_folios;
508
509 folio_batch_init(&fbatch);
510
511 while (index <= end) {
512 unsigned i;
513
514 nr_folios = filemap_get_folios_tag(mapping, &index, end,
515 PAGECACHE_TAG_WRITEBACK, &fbatch);
516
517 if (!nr_folios)
518 break;
519
520 for (i = 0; i < nr_folios; i++) {
521 struct folio *folio = fbatch.folios[i];
522
523 folio_wait_writeback(folio);
524 }
525 folio_batch_release(&fbatch);
526 cond_resched();
527 }
528}
529
530/**
531 * filemap_fdatawait_range - wait for writeback to complete
532 * @mapping: address space structure to wait for
533 * @start_byte: offset in bytes where the range starts
534 * @end_byte: offset in bytes where the range ends (inclusive)
535 *
536 * Walk the list of under-writeback pages of the given address space
537 * in the given range and wait for all of them. Check error status of
538 * the address space and return it.
539 *
540 * Since the error status of the address space is cleared by this function,
541 * callers are responsible for checking the return value and handling and/or
542 * reporting the error.
543 *
544 * Return: error status of the address space.
545 */
546int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
547 loff_t end_byte)
548{
549 __filemap_fdatawait_range(mapping, start_byte, end_byte);
550 return filemap_check_errors(mapping);
551}
552EXPORT_SYMBOL(filemap_fdatawait_range);
553
554/**
555 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
556 * @mapping: address space structure to wait for
557 * @start_byte: offset in bytes where the range starts
558 * @end_byte: offset in bytes where the range ends (inclusive)
559 *
560 * Walk the list of under-writeback pages of the given address space in the
561 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
562 * this function does not clear error status of the address space.
563 *
564 * Use this function if callers don't handle errors themselves. Expected
565 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
566 * fsfreeze(8)
567 */
568int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
569 loff_t start_byte, loff_t end_byte)
570{
571 __filemap_fdatawait_range(mapping, start_byte, end_byte);
572 return filemap_check_and_keep_errors(mapping);
573}
574EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
575
576/**
577 * file_fdatawait_range - wait for writeback to complete
578 * @file: file pointing to address space structure to wait for
579 * @start_byte: offset in bytes where the range starts
580 * @end_byte: offset in bytes where the range ends (inclusive)
581 *
582 * Walk the list of under-writeback pages of the address space that file
583 * refers to, in the given range and wait for all of them. Check error
584 * status of the address space vs. the file->f_wb_err cursor and return it.
585 *
586 * Since the error status of the file is advanced by this function,
587 * callers are responsible for checking the return value and handling and/or
588 * reporting the error.
589 *
590 * Return: error status of the address space vs. the file->f_wb_err cursor.
591 */
592int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
593{
594 struct address_space *mapping = file->f_mapping;
595
596 __filemap_fdatawait_range(mapping, start_byte, end_byte);
597 return file_check_and_advance_wb_err(file);
598}
599EXPORT_SYMBOL(file_fdatawait_range);
600
601/**
602 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
603 * @mapping: address space structure to wait for
604 *
605 * Walk the list of under-writeback pages of the given address space
606 * and wait for all of them. Unlike filemap_fdatawait(), this function
607 * does not clear error status of the address space.
608 *
609 * Use this function if callers don't handle errors themselves. Expected
610 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
611 * fsfreeze(8)
612 *
613 * Return: error status of the address space.
614 */
615int filemap_fdatawait_keep_errors(struct address_space *mapping)
616{
617 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
618 return filemap_check_and_keep_errors(mapping);
619}
620EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
621
622/* Returns true if writeback might be needed or already in progress. */
623static bool mapping_needs_writeback(struct address_space *mapping)
624{
625 return mapping->nrpages;
626}
627
628bool filemap_range_has_writeback(struct address_space *mapping,
629 loff_t start_byte, loff_t end_byte)
630{
631 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
632 pgoff_t max = end_byte >> PAGE_SHIFT;
633 struct folio *folio;
634
635 if (end_byte < start_byte)
636 return false;
637
638 rcu_read_lock();
639 xas_for_each(&xas, folio, max) {
640 if (xas_retry(&xas, folio))
641 continue;
642 if (xa_is_value(folio))
643 continue;
644 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
645 folio_test_writeback(folio))
646 break;
647 }
648 rcu_read_unlock();
649 return folio != NULL;
650}
651EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
652
653/**
654 * filemap_write_and_wait_range - write out & wait on a file range
655 * @mapping: the address_space for the pages
656 * @lstart: offset in bytes where the range starts
657 * @lend: offset in bytes where the range ends (inclusive)
658 *
659 * Write out and wait upon file offsets lstart->lend, inclusive.
660 *
661 * Note that @lend is inclusive (describes the last byte to be written) so
662 * that this function can be used to write to the very end-of-file (end = -1).
663 *
664 * Return: error status of the address space.
665 */
666int filemap_write_and_wait_range(struct address_space *mapping,
667 loff_t lstart, loff_t lend)
668{
669 int err = 0, err2;
670
671 if (lend < lstart)
672 return 0;
673
674 if (mapping_needs_writeback(mapping)) {
675 err = __filemap_fdatawrite_range(mapping, lstart, lend,
676 WB_SYNC_ALL);
677 /*
678 * Even if the above returned error, the pages may be
679 * written partially (e.g. -ENOSPC), so we wait for it.
680 * But the -EIO is special case, it may indicate the worst
681 * thing (e.g. bug) happened, so we avoid waiting for it.
682 */
683 if (err != -EIO)
684 __filemap_fdatawait_range(mapping, lstart, lend);
685 }
686 err2 = filemap_check_errors(mapping);
687 if (!err)
688 err = err2;
689 return err;
690}
691EXPORT_SYMBOL(filemap_write_and_wait_range);
692
693void __filemap_set_wb_err(struct address_space *mapping, int err)
694{
695 errseq_t eseq = errseq_set(&mapping->wb_err, err);
696
697 trace_filemap_set_wb_err(mapping, eseq);
698}
699EXPORT_SYMBOL(__filemap_set_wb_err);
700
701/**
702 * file_check_and_advance_wb_err - report wb error (if any) that was previously
703 * and advance wb_err to current one
704 * @file: struct file on which the error is being reported
705 *
706 * When userland calls fsync (or something like nfsd does the equivalent), we
707 * want to report any writeback errors that occurred since the last fsync (or
708 * since the file was opened if there haven't been any).
709 *
710 * Grab the wb_err from the mapping. If it matches what we have in the file,
711 * then just quickly return 0. The file is all caught up.
712 *
713 * If it doesn't match, then take the mapping value, set the "seen" flag in
714 * it and try to swap it into place. If it works, or another task beat us
715 * to it with the new value, then update the f_wb_err and return the error
716 * portion. The error at this point must be reported via proper channels
717 * (a'la fsync, or NFS COMMIT operation, etc.).
718 *
719 * While we handle mapping->wb_err with atomic operations, the f_wb_err
720 * value is protected by the f_lock since we must ensure that it reflects
721 * the latest value swapped in for this file descriptor.
722 *
723 * Return: %0 on success, negative error code otherwise.
724 */
725int file_check_and_advance_wb_err(struct file *file)
726{
727 int err = 0;
728 errseq_t old = READ_ONCE(file->f_wb_err);
729 struct address_space *mapping = file->f_mapping;
730
731 /* Locklessly handle the common case where nothing has changed */
732 if (errseq_check(&mapping->wb_err, old)) {
733 /* Something changed, must use slow path */
734 spin_lock(&file->f_lock);
735 old = file->f_wb_err;
736 err = errseq_check_and_advance(&mapping->wb_err,
737 &file->f_wb_err);
738 trace_file_check_and_advance_wb_err(file, old);
739 spin_unlock(&file->f_lock);
740 }
741
742 /*
743 * We're mostly using this function as a drop in replacement for
744 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
745 * that the legacy code would have had on these flags.
746 */
747 clear_bit(AS_EIO, &mapping->flags);
748 clear_bit(AS_ENOSPC, &mapping->flags);
749 return err;
750}
751EXPORT_SYMBOL(file_check_and_advance_wb_err);
752
753/**
754 * file_write_and_wait_range - write out & wait on a file range
755 * @file: file pointing to address_space with pages
756 * @lstart: offset in bytes where the range starts
757 * @lend: offset in bytes where the range ends (inclusive)
758 *
759 * Write out and wait upon file offsets lstart->lend, inclusive.
760 *
761 * Note that @lend is inclusive (describes the last byte to be written) so
762 * that this function can be used to write to the very end-of-file (end = -1).
763 *
764 * After writing out and waiting on the data, we check and advance the
765 * f_wb_err cursor to the latest value, and return any errors detected there.
766 *
767 * Return: %0 on success, negative error code otherwise.
768 */
769int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
770{
771 int err = 0, err2;
772 struct address_space *mapping = file->f_mapping;
773
774 if (lend < lstart)
775 return 0;
776
777 if (mapping_needs_writeback(mapping)) {
778 err = __filemap_fdatawrite_range(mapping, lstart, lend,
779 WB_SYNC_ALL);
780 /* See comment of filemap_write_and_wait() */
781 if (err != -EIO)
782 __filemap_fdatawait_range(mapping, lstart, lend);
783 }
784 err2 = file_check_and_advance_wb_err(file);
785 if (!err)
786 err = err2;
787 return err;
788}
789EXPORT_SYMBOL(file_write_and_wait_range);
790
791/**
792 * replace_page_cache_folio - replace a pagecache folio with a new one
793 * @old: folio to be replaced
794 * @new: folio to replace with
795 *
796 * This function replaces a folio in the pagecache with a new one. On
797 * success it acquires the pagecache reference for the new folio and
798 * drops it for the old folio. Both the old and new folios must be
799 * locked. This function does not add the new folio to the LRU, the
800 * caller must do that.
801 *
802 * The remove + add is atomic. This function cannot fail.
803 */
804void replace_page_cache_folio(struct folio *old, struct folio *new)
805{
806 struct address_space *mapping = old->mapping;
807 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
808 pgoff_t offset = old->index;
809 XA_STATE(xas, &mapping->i_pages, offset);
810
811 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
812 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
813 VM_BUG_ON_FOLIO(new->mapping, new);
814
815 folio_get(new);
816 new->mapping = mapping;
817 new->index = offset;
818
819 mem_cgroup_replace_folio(old, new);
820
821 xas_lock_irq(&xas);
822 xas_store(&xas, new);
823
824 old->mapping = NULL;
825 /* hugetlb pages do not participate in page cache accounting. */
826 if (!folio_test_hugetlb(old))
827 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
828 if (!folio_test_hugetlb(new))
829 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
830 if (folio_test_swapbacked(old))
831 __lruvec_stat_sub_folio(old, NR_SHMEM);
832 if (folio_test_swapbacked(new))
833 __lruvec_stat_add_folio(new, NR_SHMEM);
834 xas_unlock_irq(&xas);
835 if (free_folio)
836 free_folio(old);
837 folio_put(old);
838}
839EXPORT_SYMBOL_GPL(replace_page_cache_folio);
840
841noinline int __filemap_add_folio(struct address_space *mapping,
842 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
843{
844 XA_STATE(xas, &mapping->i_pages, index);
845 void *alloced_shadow = NULL;
846 int alloced_order = 0;
847 bool huge;
848 long nr;
849
850 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
851 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
852 VM_BUG_ON_FOLIO(folio_order(folio) < mapping_min_folio_order(mapping),
853 folio);
854 mapping_set_update(&xas, mapping);
855
856 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
857 xas_set_order(&xas, index, folio_order(folio));
858 huge = folio_test_hugetlb(folio);
859 nr = folio_nr_pages(folio);
860
861 gfp &= GFP_RECLAIM_MASK;
862 folio_ref_add(folio, nr);
863 folio->mapping = mapping;
864 folio->index = xas.xa_index;
865
866 for (;;) {
867 int order = -1, split_order = 0;
868 void *entry, *old = NULL;
869
870 xas_lock_irq(&xas);
871 xas_for_each_conflict(&xas, entry) {
872 old = entry;
873 if (!xa_is_value(entry)) {
874 xas_set_err(&xas, -EEXIST);
875 goto unlock;
876 }
877 /*
878 * If a larger entry exists,
879 * it will be the first and only entry iterated.
880 */
881 if (order == -1)
882 order = xas_get_order(&xas);
883 }
884
885 /* entry may have changed before we re-acquire the lock */
886 if (alloced_order && (old != alloced_shadow || order != alloced_order)) {
887 xas_destroy(&xas);
888 alloced_order = 0;
889 }
890
891 if (old) {
892 if (order > 0 && order > folio_order(folio)) {
893 /* How to handle large swap entries? */
894 BUG_ON(shmem_mapping(mapping));
895 if (!alloced_order) {
896 split_order = order;
897 goto unlock;
898 }
899 xas_split(&xas, old, order);
900 xas_reset(&xas);
901 }
902 if (shadowp)
903 *shadowp = old;
904 }
905
906 xas_store(&xas, folio);
907 if (xas_error(&xas))
908 goto unlock;
909
910 mapping->nrpages += nr;
911
912 /* hugetlb pages do not participate in page cache accounting */
913 if (!huge) {
914 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
915 if (folio_test_pmd_mappable(folio))
916 __lruvec_stat_mod_folio(folio,
917 NR_FILE_THPS, nr);
918 }
919
920unlock:
921 xas_unlock_irq(&xas);
922
923 /* split needed, alloc here and retry. */
924 if (split_order) {
925 xas_split_alloc(&xas, old, split_order, gfp);
926 if (xas_error(&xas))
927 goto error;
928 alloced_shadow = old;
929 alloced_order = split_order;
930 xas_reset(&xas);
931 continue;
932 }
933
934 if (!xas_nomem(&xas, gfp))
935 break;
936 }
937
938 if (xas_error(&xas))
939 goto error;
940
941 trace_mm_filemap_add_to_page_cache(folio);
942 return 0;
943error:
944 folio->mapping = NULL;
945 /* Leave page->index set: truncation relies upon it */
946 folio_put_refs(folio, nr);
947 return xas_error(&xas);
948}
949ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
950
951int filemap_add_folio(struct address_space *mapping, struct folio *folio,
952 pgoff_t index, gfp_t gfp)
953{
954 void *shadow = NULL;
955 int ret;
956
957 ret = mem_cgroup_charge(folio, NULL, gfp);
958 if (ret)
959 return ret;
960
961 __folio_set_locked(folio);
962 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
963 if (unlikely(ret)) {
964 mem_cgroup_uncharge(folio);
965 __folio_clear_locked(folio);
966 } else {
967 /*
968 * The folio might have been evicted from cache only
969 * recently, in which case it should be activated like
970 * any other repeatedly accessed folio.
971 * The exception is folios getting rewritten; evicting other
972 * data from the working set, only to cache data that will
973 * get overwritten with something else, is a waste of memory.
974 */
975 WARN_ON_ONCE(folio_test_active(folio));
976 if (!(gfp & __GFP_WRITE) && shadow)
977 workingset_refault(folio, shadow);
978 folio_add_lru(folio);
979 }
980 return ret;
981}
982EXPORT_SYMBOL_GPL(filemap_add_folio);
983
984#ifdef CONFIG_NUMA
985struct folio *filemap_alloc_folio_noprof(gfp_t gfp, unsigned int order)
986{
987 int n;
988 struct folio *folio;
989
990 if (cpuset_do_page_mem_spread()) {
991 unsigned int cpuset_mems_cookie;
992 do {
993 cpuset_mems_cookie = read_mems_allowed_begin();
994 n = cpuset_mem_spread_node();
995 folio = __folio_alloc_node_noprof(gfp, order, n);
996 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
997
998 return folio;
999 }
1000 return folio_alloc_noprof(gfp, order);
1001}
1002EXPORT_SYMBOL(filemap_alloc_folio_noprof);
1003#endif
1004
1005/*
1006 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
1007 *
1008 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
1009 *
1010 * @mapping1: the first mapping to lock
1011 * @mapping2: the second mapping to lock
1012 */
1013void filemap_invalidate_lock_two(struct address_space *mapping1,
1014 struct address_space *mapping2)
1015{
1016 if (mapping1 > mapping2)
1017 swap(mapping1, mapping2);
1018 if (mapping1)
1019 down_write(&mapping1->invalidate_lock);
1020 if (mapping2 && mapping1 != mapping2)
1021 down_write_nested(&mapping2->invalidate_lock, 1);
1022}
1023EXPORT_SYMBOL(filemap_invalidate_lock_two);
1024
1025/*
1026 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
1027 *
1028 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1029 *
1030 * @mapping1: the first mapping to unlock
1031 * @mapping2: the second mapping to unlock
1032 */
1033void filemap_invalidate_unlock_two(struct address_space *mapping1,
1034 struct address_space *mapping2)
1035{
1036 if (mapping1)
1037 up_write(&mapping1->invalidate_lock);
1038 if (mapping2 && mapping1 != mapping2)
1039 up_write(&mapping2->invalidate_lock);
1040}
1041EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1042
1043/*
1044 * In order to wait for pages to become available there must be
1045 * waitqueues associated with pages. By using a hash table of
1046 * waitqueues where the bucket discipline is to maintain all
1047 * waiters on the same queue and wake all when any of the pages
1048 * become available, and for the woken contexts to check to be
1049 * sure the appropriate page became available, this saves space
1050 * at a cost of "thundering herd" phenomena during rare hash
1051 * collisions.
1052 */
1053#define PAGE_WAIT_TABLE_BITS 8
1054#define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1055static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1056
1057static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1058{
1059 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1060}
1061
1062void __init pagecache_init(void)
1063{
1064 int i;
1065
1066 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1067 init_waitqueue_head(&folio_wait_table[i]);
1068
1069 page_writeback_init();
1070}
1071
1072/*
1073 * The page wait code treats the "wait->flags" somewhat unusually, because
1074 * we have multiple different kinds of waits, not just the usual "exclusive"
1075 * one.
1076 *
1077 * We have:
1078 *
1079 * (a) no special bits set:
1080 *
1081 * We're just waiting for the bit to be released, and when a waker
1082 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1083 * and remove it from the wait queue.
1084 *
1085 * Simple and straightforward.
1086 *
1087 * (b) WQ_FLAG_EXCLUSIVE:
1088 *
1089 * The waiter is waiting to get the lock, and only one waiter should
1090 * be woken up to avoid any thundering herd behavior. We'll set the
1091 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1092 *
1093 * This is the traditional exclusive wait.
1094 *
1095 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1096 *
1097 * The waiter is waiting to get the bit, and additionally wants the
1098 * lock to be transferred to it for fair lock behavior. If the lock
1099 * cannot be taken, we stop walking the wait queue without waking
1100 * the waiter.
1101 *
1102 * This is the "fair lock handoff" case, and in addition to setting
1103 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1104 * that it now has the lock.
1105 */
1106static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1107{
1108 unsigned int flags;
1109 struct wait_page_key *key = arg;
1110 struct wait_page_queue *wait_page
1111 = container_of(wait, struct wait_page_queue, wait);
1112
1113 if (!wake_page_match(wait_page, key))
1114 return 0;
1115
1116 /*
1117 * If it's a lock handoff wait, we get the bit for it, and
1118 * stop walking (and do not wake it up) if we can't.
1119 */
1120 flags = wait->flags;
1121 if (flags & WQ_FLAG_EXCLUSIVE) {
1122 if (test_bit(key->bit_nr, &key->folio->flags))
1123 return -1;
1124 if (flags & WQ_FLAG_CUSTOM) {
1125 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1126 return -1;
1127 flags |= WQ_FLAG_DONE;
1128 }
1129 }
1130
1131 /*
1132 * We are holding the wait-queue lock, but the waiter that
1133 * is waiting for this will be checking the flags without
1134 * any locking.
1135 *
1136 * So update the flags atomically, and wake up the waiter
1137 * afterwards to avoid any races. This store-release pairs
1138 * with the load-acquire in folio_wait_bit_common().
1139 */
1140 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1141 wake_up_state(wait->private, mode);
1142
1143 /*
1144 * Ok, we have successfully done what we're waiting for,
1145 * and we can unconditionally remove the wait entry.
1146 *
1147 * Note that this pairs with the "finish_wait()" in the
1148 * waiter, and has to be the absolute last thing we do.
1149 * After this list_del_init(&wait->entry) the wait entry
1150 * might be de-allocated and the process might even have
1151 * exited.
1152 */
1153 list_del_init_careful(&wait->entry);
1154 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1155}
1156
1157static void folio_wake_bit(struct folio *folio, int bit_nr)
1158{
1159 wait_queue_head_t *q = folio_waitqueue(folio);
1160 struct wait_page_key key;
1161 unsigned long flags;
1162
1163 key.folio = folio;
1164 key.bit_nr = bit_nr;
1165 key.page_match = 0;
1166
1167 spin_lock_irqsave(&q->lock, flags);
1168 __wake_up_locked_key(q, TASK_NORMAL, &key);
1169
1170 /*
1171 * It's possible to miss clearing waiters here, when we woke our page
1172 * waiters, but the hashed waitqueue has waiters for other pages on it.
1173 * That's okay, it's a rare case. The next waker will clear it.
1174 *
1175 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1176 * other), the flag may be cleared in the course of freeing the page;
1177 * but that is not required for correctness.
1178 */
1179 if (!waitqueue_active(q) || !key.page_match)
1180 folio_clear_waiters(folio);
1181
1182 spin_unlock_irqrestore(&q->lock, flags);
1183}
1184
1185/*
1186 * A choice of three behaviors for folio_wait_bit_common():
1187 */
1188enum behavior {
1189 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1190 * __folio_lock() waiting on then setting PG_locked.
1191 */
1192 SHARED, /* Hold ref to page and check the bit when woken, like
1193 * folio_wait_writeback() waiting on PG_writeback.
1194 */
1195 DROP, /* Drop ref to page before wait, no check when woken,
1196 * like folio_put_wait_locked() on PG_locked.
1197 */
1198};
1199
1200/*
1201 * Attempt to check (or get) the folio flag, and mark us done
1202 * if successful.
1203 */
1204static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1205 struct wait_queue_entry *wait)
1206{
1207 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1208 if (test_and_set_bit(bit_nr, &folio->flags))
1209 return false;
1210 } else if (test_bit(bit_nr, &folio->flags))
1211 return false;
1212
1213 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1214 return true;
1215}
1216
1217/* How many times do we accept lock stealing from under a waiter? */
1218int sysctl_page_lock_unfairness = 5;
1219
1220static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1221 int state, enum behavior behavior)
1222{
1223 wait_queue_head_t *q = folio_waitqueue(folio);
1224 int unfairness = sysctl_page_lock_unfairness;
1225 struct wait_page_queue wait_page;
1226 wait_queue_entry_t *wait = &wait_page.wait;
1227 bool thrashing = false;
1228 unsigned long pflags;
1229 bool in_thrashing;
1230
1231 if (bit_nr == PG_locked &&
1232 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1233 delayacct_thrashing_start(&in_thrashing);
1234 psi_memstall_enter(&pflags);
1235 thrashing = true;
1236 }
1237
1238 init_wait(wait);
1239 wait->func = wake_page_function;
1240 wait_page.folio = folio;
1241 wait_page.bit_nr = bit_nr;
1242
1243repeat:
1244 wait->flags = 0;
1245 if (behavior == EXCLUSIVE) {
1246 wait->flags = WQ_FLAG_EXCLUSIVE;
1247 if (--unfairness < 0)
1248 wait->flags |= WQ_FLAG_CUSTOM;
1249 }
1250
1251 /*
1252 * Do one last check whether we can get the
1253 * page bit synchronously.
1254 *
1255 * Do the folio_set_waiters() marking before that
1256 * to let any waker we _just_ missed know they
1257 * need to wake us up (otherwise they'll never
1258 * even go to the slow case that looks at the
1259 * page queue), and add ourselves to the wait
1260 * queue if we need to sleep.
1261 *
1262 * This part needs to be done under the queue
1263 * lock to avoid races.
1264 */
1265 spin_lock_irq(&q->lock);
1266 folio_set_waiters(folio);
1267 if (!folio_trylock_flag(folio, bit_nr, wait))
1268 __add_wait_queue_entry_tail(q, wait);
1269 spin_unlock_irq(&q->lock);
1270
1271 /*
1272 * From now on, all the logic will be based on
1273 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1274 * see whether the page bit testing has already
1275 * been done by the wake function.
1276 *
1277 * We can drop our reference to the folio.
1278 */
1279 if (behavior == DROP)
1280 folio_put(folio);
1281
1282 /*
1283 * Note that until the "finish_wait()", or until
1284 * we see the WQ_FLAG_WOKEN flag, we need to
1285 * be very careful with the 'wait->flags', because
1286 * we may race with a waker that sets them.
1287 */
1288 for (;;) {
1289 unsigned int flags;
1290
1291 set_current_state(state);
1292
1293 /* Loop until we've been woken or interrupted */
1294 flags = smp_load_acquire(&wait->flags);
1295 if (!(flags & WQ_FLAG_WOKEN)) {
1296 if (signal_pending_state(state, current))
1297 break;
1298
1299 io_schedule();
1300 continue;
1301 }
1302
1303 /* If we were non-exclusive, we're done */
1304 if (behavior != EXCLUSIVE)
1305 break;
1306
1307 /* If the waker got the lock for us, we're done */
1308 if (flags & WQ_FLAG_DONE)
1309 break;
1310
1311 /*
1312 * Otherwise, if we're getting the lock, we need to
1313 * try to get it ourselves.
1314 *
1315 * And if that fails, we'll have to retry this all.
1316 */
1317 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1318 goto repeat;
1319
1320 wait->flags |= WQ_FLAG_DONE;
1321 break;
1322 }
1323
1324 /*
1325 * If a signal happened, this 'finish_wait()' may remove the last
1326 * waiter from the wait-queues, but the folio waiters bit will remain
1327 * set. That's ok. The next wakeup will take care of it, and trying
1328 * to do it here would be difficult and prone to races.
1329 */
1330 finish_wait(q, wait);
1331
1332 if (thrashing) {
1333 delayacct_thrashing_end(&in_thrashing);
1334 psi_memstall_leave(&pflags);
1335 }
1336
1337 /*
1338 * NOTE! The wait->flags weren't stable until we've done the
1339 * 'finish_wait()', and we could have exited the loop above due
1340 * to a signal, and had a wakeup event happen after the signal
1341 * test but before the 'finish_wait()'.
1342 *
1343 * So only after the finish_wait() can we reliably determine
1344 * if we got woken up or not, so we can now figure out the final
1345 * return value based on that state without races.
1346 *
1347 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1348 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1349 */
1350 if (behavior == EXCLUSIVE)
1351 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1352
1353 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1354}
1355
1356#ifdef CONFIG_MIGRATION
1357/**
1358 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1359 * @entry: migration swap entry.
1360 * @ptl: already locked ptl. This function will drop the lock.
1361 *
1362 * Wait for a migration entry referencing the given page to be removed. This is
1363 * equivalent to put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE) except
1364 * this can be called without taking a reference on the page. Instead this
1365 * should be called while holding the ptl for the migration entry referencing
1366 * the page.
1367 *
1368 * Returns after unlocking the ptl.
1369 *
1370 * This follows the same logic as folio_wait_bit_common() so see the comments
1371 * there.
1372 */
1373void migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl)
1374 __releases(ptl)
1375{
1376 struct wait_page_queue wait_page;
1377 wait_queue_entry_t *wait = &wait_page.wait;
1378 bool thrashing = false;
1379 unsigned long pflags;
1380 bool in_thrashing;
1381 wait_queue_head_t *q;
1382 struct folio *folio = pfn_swap_entry_folio(entry);
1383
1384 q = folio_waitqueue(folio);
1385 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1386 delayacct_thrashing_start(&in_thrashing);
1387 psi_memstall_enter(&pflags);
1388 thrashing = true;
1389 }
1390
1391 init_wait(wait);
1392 wait->func = wake_page_function;
1393 wait_page.folio = folio;
1394 wait_page.bit_nr = PG_locked;
1395 wait->flags = 0;
1396
1397 spin_lock_irq(&q->lock);
1398 folio_set_waiters(folio);
1399 if (!folio_trylock_flag(folio, PG_locked, wait))
1400 __add_wait_queue_entry_tail(q, wait);
1401 spin_unlock_irq(&q->lock);
1402
1403 /*
1404 * If a migration entry exists for the page the migration path must hold
1405 * a valid reference to the page, and it must take the ptl to remove the
1406 * migration entry. So the page is valid until the ptl is dropped.
1407 */
1408 spin_unlock(ptl);
1409
1410 for (;;) {
1411 unsigned int flags;
1412
1413 set_current_state(TASK_UNINTERRUPTIBLE);
1414
1415 /* Loop until we've been woken or interrupted */
1416 flags = smp_load_acquire(&wait->flags);
1417 if (!(flags & WQ_FLAG_WOKEN)) {
1418 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1419 break;
1420
1421 io_schedule();
1422 continue;
1423 }
1424 break;
1425 }
1426
1427 finish_wait(q, wait);
1428
1429 if (thrashing) {
1430 delayacct_thrashing_end(&in_thrashing);
1431 psi_memstall_leave(&pflags);
1432 }
1433}
1434#endif
1435
1436void folio_wait_bit(struct folio *folio, int bit_nr)
1437{
1438 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1439}
1440EXPORT_SYMBOL(folio_wait_bit);
1441
1442int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1443{
1444 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1445}
1446EXPORT_SYMBOL(folio_wait_bit_killable);
1447
1448/**
1449 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1450 * @folio: The folio to wait for.
1451 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1452 *
1453 * The caller should hold a reference on @folio. They expect the page to
1454 * become unlocked relatively soon, but do not wish to hold up migration
1455 * (for example) by holding the reference while waiting for the folio to
1456 * come unlocked. After this function returns, the caller should not
1457 * dereference @folio.
1458 *
1459 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1460 */
1461static int folio_put_wait_locked(struct folio *folio, int state)
1462{
1463 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1464}
1465
1466/**
1467 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue
1468 * @folio: Folio defining the wait queue of interest
1469 * @waiter: Waiter to add to the queue
1470 *
1471 * Add an arbitrary @waiter to the wait queue for the nominated @folio.
1472 */
1473void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter)
1474{
1475 wait_queue_head_t *q = folio_waitqueue(folio);
1476 unsigned long flags;
1477
1478 spin_lock_irqsave(&q->lock, flags);
1479 __add_wait_queue_entry_tail(q, waiter);
1480 folio_set_waiters(folio);
1481 spin_unlock_irqrestore(&q->lock, flags);
1482}
1483EXPORT_SYMBOL_GPL(folio_add_wait_queue);
1484
1485/**
1486 * folio_unlock - Unlock a locked folio.
1487 * @folio: The folio.
1488 *
1489 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1490 *
1491 * Context: May be called from interrupt or process context. May not be
1492 * called from NMI context.
1493 */
1494void folio_unlock(struct folio *folio)
1495{
1496 /* Bit 7 allows x86 to check the byte's sign bit */
1497 BUILD_BUG_ON(PG_waiters != 7);
1498 BUILD_BUG_ON(PG_locked > 7);
1499 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1500 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked))
1501 folio_wake_bit(folio, PG_locked);
1502}
1503EXPORT_SYMBOL(folio_unlock);
1504
1505/**
1506 * folio_end_read - End read on a folio.
1507 * @folio: The folio.
1508 * @success: True if all reads completed successfully.
1509 *
1510 * When all reads against a folio have completed, filesystems should
1511 * call this function to let the pagecache know that no more reads
1512 * are outstanding. This will unlock the folio and wake up any thread
1513 * sleeping on the lock. The folio will also be marked uptodate if all
1514 * reads succeeded.
1515 *
1516 * Context: May be called from interrupt or process context. May not be
1517 * called from NMI context.
1518 */
1519void folio_end_read(struct folio *folio, bool success)
1520{
1521 unsigned long mask = 1 << PG_locked;
1522
1523 /* Must be in bottom byte for x86 to work */
1524 BUILD_BUG_ON(PG_uptodate > 7);
1525 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1526 VM_BUG_ON_FOLIO(success && folio_test_uptodate(folio), folio);
1527
1528 if (likely(success))
1529 mask |= 1 << PG_uptodate;
1530 if (folio_xor_flags_has_waiters(folio, mask))
1531 folio_wake_bit(folio, PG_locked);
1532}
1533EXPORT_SYMBOL(folio_end_read);
1534
1535/**
1536 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1537 * @folio: The folio.
1538 *
1539 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1540 * it. The folio reference held for PG_private_2 being set is released.
1541 *
1542 * This is, for example, used when a netfs folio is being written to a local
1543 * disk cache, thereby allowing writes to the cache for the same folio to be
1544 * serialised.
1545 */
1546void folio_end_private_2(struct folio *folio)
1547{
1548 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1549 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1550 folio_wake_bit(folio, PG_private_2);
1551 folio_put(folio);
1552}
1553EXPORT_SYMBOL(folio_end_private_2);
1554
1555/**
1556 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1557 * @folio: The folio to wait on.
1558 *
1559 * Wait for PG_private_2 to be cleared on a folio.
1560 */
1561void folio_wait_private_2(struct folio *folio)
1562{
1563 while (folio_test_private_2(folio))
1564 folio_wait_bit(folio, PG_private_2);
1565}
1566EXPORT_SYMBOL(folio_wait_private_2);
1567
1568/**
1569 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1570 * @folio: The folio to wait on.
1571 *
1572 * Wait for PG_private_2 to be cleared on a folio or until a fatal signal is
1573 * received by the calling task.
1574 *
1575 * Return:
1576 * - 0 if successful.
1577 * - -EINTR if a fatal signal was encountered.
1578 */
1579int folio_wait_private_2_killable(struct folio *folio)
1580{
1581 int ret = 0;
1582
1583 while (folio_test_private_2(folio)) {
1584 ret = folio_wait_bit_killable(folio, PG_private_2);
1585 if (ret < 0)
1586 break;
1587 }
1588
1589 return ret;
1590}
1591EXPORT_SYMBOL(folio_wait_private_2_killable);
1592
1593/**
1594 * folio_end_writeback - End writeback against a folio.
1595 * @folio: The folio.
1596 *
1597 * The folio must actually be under writeback.
1598 *
1599 * Context: May be called from process or interrupt context.
1600 */
1601void folio_end_writeback(struct folio *folio)
1602{
1603 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
1604
1605 /*
1606 * folio_test_clear_reclaim() could be used here but it is an
1607 * atomic operation and overkill in this particular case. Failing
1608 * to shuffle a folio marked for immediate reclaim is too mild
1609 * a gain to justify taking an atomic operation penalty at the
1610 * end of every folio writeback.
1611 */
1612 if (folio_test_reclaim(folio)) {
1613 folio_clear_reclaim(folio);
1614 folio_rotate_reclaimable(folio);
1615 }
1616
1617 /*
1618 * Writeback does not hold a folio reference of its own, relying
1619 * on truncation to wait for the clearing of PG_writeback.
1620 * But here we must make sure that the folio is not freed and
1621 * reused before the folio_wake_bit().
1622 */
1623 folio_get(folio);
1624 if (__folio_end_writeback(folio))
1625 folio_wake_bit(folio, PG_writeback);
1626 acct_reclaim_writeback(folio);
1627 folio_put(folio);
1628}
1629EXPORT_SYMBOL(folio_end_writeback);
1630
1631/**
1632 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1633 * @folio: The folio to lock
1634 */
1635void __folio_lock(struct folio *folio)
1636{
1637 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1638 EXCLUSIVE);
1639}
1640EXPORT_SYMBOL(__folio_lock);
1641
1642int __folio_lock_killable(struct folio *folio)
1643{
1644 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1645 EXCLUSIVE);
1646}
1647EXPORT_SYMBOL_GPL(__folio_lock_killable);
1648
1649static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1650{
1651 struct wait_queue_head *q = folio_waitqueue(folio);
1652 int ret;
1653
1654 wait->folio = folio;
1655 wait->bit_nr = PG_locked;
1656
1657 spin_lock_irq(&q->lock);
1658 __add_wait_queue_entry_tail(q, &wait->wait);
1659 folio_set_waiters(folio);
1660 ret = !folio_trylock(folio);
1661 /*
1662 * If we were successful now, we know we're still on the
1663 * waitqueue as we're still under the lock. This means it's
1664 * safe to remove and return success, we know the callback
1665 * isn't going to trigger.
1666 */
1667 if (!ret)
1668 __remove_wait_queue(q, &wait->wait);
1669 else
1670 ret = -EIOCBQUEUED;
1671 spin_unlock_irq(&q->lock);
1672 return ret;
1673}
1674
1675/*
1676 * Return values:
1677 * 0 - folio is locked.
1678 * non-zero - folio is not locked.
1679 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or
1680 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and
1681 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held.
1682 *
1683 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0
1684 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed.
1685 */
1686vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf)
1687{
1688 unsigned int flags = vmf->flags;
1689
1690 if (fault_flag_allow_retry_first(flags)) {
1691 /*
1692 * CAUTION! In this case, mmap_lock/per-VMA lock is not
1693 * released even though returning VM_FAULT_RETRY.
1694 */
1695 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1696 return VM_FAULT_RETRY;
1697
1698 release_fault_lock(vmf);
1699 if (flags & FAULT_FLAG_KILLABLE)
1700 folio_wait_locked_killable(folio);
1701 else
1702 folio_wait_locked(folio);
1703 return VM_FAULT_RETRY;
1704 }
1705 if (flags & FAULT_FLAG_KILLABLE) {
1706 bool ret;
1707
1708 ret = __folio_lock_killable(folio);
1709 if (ret) {
1710 release_fault_lock(vmf);
1711 return VM_FAULT_RETRY;
1712 }
1713 } else {
1714 __folio_lock(folio);
1715 }
1716
1717 return 0;
1718}
1719
1720/**
1721 * page_cache_next_miss() - Find the next gap in the page cache.
1722 * @mapping: Mapping.
1723 * @index: Index.
1724 * @max_scan: Maximum range to search.
1725 *
1726 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1727 * gap with the lowest index.
1728 *
1729 * This function may be called under the rcu_read_lock. However, this will
1730 * not atomically search a snapshot of the cache at a single point in time.
1731 * For example, if a gap is created at index 5, then subsequently a gap is
1732 * created at index 10, page_cache_next_miss covering both indices may
1733 * return 10 if called under the rcu_read_lock.
1734 *
1735 * Return: The index of the gap if found, otherwise an index outside the
1736 * range specified (in which case 'return - index >= max_scan' will be true).
1737 * In the rare case of index wrap-around, 0 will be returned.
1738 */
1739pgoff_t page_cache_next_miss(struct address_space *mapping,
1740 pgoff_t index, unsigned long max_scan)
1741{
1742 XA_STATE(xas, &mapping->i_pages, index);
1743
1744 while (max_scan--) {
1745 void *entry = xas_next(&xas);
1746 if (!entry || xa_is_value(entry))
1747 return xas.xa_index;
1748 if (xas.xa_index == 0)
1749 return 0;
1750 }
1751
1752 return index + max_scan;
1753}
1754EXPORT_SYMBOL(page_cache_next_miss);
1755
1756/**
1757 * page_cache_prev_miss() - Find the previous gap in the page cache.
1758 * @mapping: Mapping.
1759 * @index: Index.
1760 * @max_scan: Maximum range to search.
1761 *
1762 * Search the range [max(index - max_scan + 1, 0), index] for the
1763 * gap with the highest index.
1764 *
1765 * This function may be called under the rcu_read_lock. However, this will
1766 * not atomically search a snapshot of the cache at a single point in time.
1767 * For example, if a gap is created at index 10, then subsequently a gap is
1768 * created at index 5, page_cache_prev_miss() covering both indices may
1769 * return 5 if called under the rcu_read_lock.
1770 *
1771 * Return: The index of the gap if found, otherwise an index outside the
1772 * range specified (in which case 'index - return >= max_scan' will be true).
1773 * In the rare case of wrap-around, ULONG_MAX will be returned.
1774 */
1775pgoff_t page_cache_prev_miss(struct address_space *mapping,
1776 pgoff_t index, unsigned long max_scan)
1777{
1778 XA_STATE(xas, &mapping->i_pages, index);
1779
1780 while (max_scan--) {
1781 void *entry = xas_prev(&xas);
1782 if (!entry || xa_is_value(entry))
1783 break;
1784 if (xas.xa_index == ULONG_MAX)
1785 break;
1786 }
1787
1788 return xas.xa_index;
1789}
1790EXPORT_SYMBOL(page_cache_prev_miss);
1791
1792/*
1793 * Lockless page cache protocol:
1794 * On the lookup side:
1795 * 1. Load the folio from i_pages
1796 * 2. Increment the refcount if it's not zero
1797 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1798 *
1799 * On the removal side:
1800 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1801 * B. Remove the page from i_pages
1802 * C. Return the page to the page allocator
1803 *
1804 * This means that any page may have its reference count temporarily
1805 * increased by a speculative page cache (or GUP-fast) lookup as it can
1806 * be allocated by another user before the RCU grace period expires.
1807 * Because the refcount temporarily acquired here may end up being the
1808 * last refcount on the page, any page allocation must be freeable by
1809 * folio_put().
1810 */
1811
1812/*
1813 * filemap_get_entry - Get a page cache entry.
1814 * @mapping: the address_space to search
1815 * @index: The page cache index.
1816 *
1817 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1818 * it is returned with an increased refcount. If it is a shadow entry
1819 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1820 * it is returned without further action.
1821 *
1822 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1823 */
1824void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
1825{
1826 XA_STATE(xas, &mapping->i_pages, index);
1827 struct folio *folio;
1828
1829 rcu_read_lock();
1830repeat:
1831 xas_reset(&xas);
1832 folio = xas_load(&xas);
1833 if (xas_retry(&xas, folio))
1834 goto repeat;
1835 /*
1836 * A shadow entry of a recently evicted page, or a swap entry from
1837 * shmem/tmpfs. Return it without attempting to raise page count.
1838 */
1839 if (!folio || xa_is_value(folio))
1840 goto out;
1841
1842 if (!folio_try_get(folio))
1843 goto repeat;
1844
1845 if (unlikely(folio != xas_reload(&xas))) {
1846 folio_put(folio);
1847 goto repeat;
1848 }
1849out:
1850 rcu_read_unlock();
1851
1852 return folio;
1853}
1854
1855/**
1856 * __filemap_get_folio - Find and get a reference to a folio.
1857 * @mapping: The address_space to search.
1858 * @index: The page index.
1859 * @fgp_flags: %FGP flags modify how the folio is returned.
1860 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1861 *
1862 * Looks up the page cache entry at @mapping & @index.
1863 *
1864 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1865 * if the %GFP flags specified for %FGP_CREAT are atomic.
1866 *
1867 * If this function returns a folio, it is returned with an increased refcount.
1868 *
1869 * Return: The found folio or an ERR_PTR() otherwise.
1870 */
1871struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1872 fgf_t fgp_flags, gfp_t gfp)
1873{
1874 struct folio *folio;
1875
1876repeat:
1877 folio = filemap_get_entry(mapping, index);
1878 if (xa_is_value(folio))
1879 folio = NULL;
1880 if (!folio)
1881 goto no_page;
1882
1883 if (fgp_flags & FGP_LOCK) {
1884 if (fgp_flags & FGP_NOWAIT) {
1885 if (!folio_trylock(folio)) {
1886 folio_put(folio);
1887 return ERR_PTR(-EAGAIN);
1888 }
1889 } else {
1890 folio_lock(folio);
1891 }
1892
1893 /* Has the page been truncated? */
1894 if (unlikely(folio->mapping != mapping)) {
1895 folio_unlock(folio);
1896 folio_put(folio);
1897 goto repeat;
1898 }
1899 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1900 }
1901
1902 if (fgp_flags & FGP_ACCESSED)
1903 folio_mark_accessed(folio);
1904 else if (fgp_flags & FGP_WRITE) {
1905 /* Clear idle flag for buffer write */
1906 if (folio_test_idle(folio))
1907 folio_clear_idle(folio);
1908 }
1909
1910 if (fgp_flags & FGP_STABLE)
1911 folio_wait_stable(folio);
1912no_page:
1913 if (!folio && (fgp_flags & FGP_CREAT)) {
1914 unsigned int min_order = mapping_min_folio_order(mapping);
1915 unsigned int order = max(min_order, FGF_GET_ORDER(fgp_flags));
1916 int err;
1917 index = mapping_align_index(mapping, index);
1918
1919 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1920 gfp |= __GFP_WRITE;
1921 if (fgp_flags & FGP_NOFS)
1922 gfp &= ~__GFP_FS;
1923 if (fgp_flags & FGP_NOWAIT) {
1924 gfp &= ~GFP_KERNEL;
1925 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1926 }
1927 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1928 fgp_flags |= FGP_LOCK;
1929
1930 if (order > mapping_max_folio_order(mapping))
1931 order = mapping_max_folio_order(mapping);
1932 /* If we're not aligned, allocate a smaller folio */
1933 if (index & ((1UL << order) - 1))
1934 order = __ffs(index);
1935
1936 do {
1937 gfp_t alloc_gfp = gfp;
1938
1939 err = -ENOMEM;
1940 if (order > min_order)
1941 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN;
1942 folio = filemap_alloc_folio(alloc_gfp, order);
1943 if (!folio)
1944 continue;
1945
1946 /* Init accessed so avoid atomic mark_page_accessed later */
1947 if (fgp_flags & FGP_ACCESSED)
1948 __folio_set_referenced(folio);
1949
1950 err = filemap_add_folio(mapping, folio, index, gfp);
1951 if (!err)
1952 break;
1953 folio_put(folio);
1954 folio = NULL;
1955 } while (order-- > min_order);
1956
1957 if (err == -EEXIST)
1958 goto repeat;
1959 if (err)
1960 return ERR_PTR(err);
1961 /*
1962 * filemap_add_folio locks the page, and for mmap
1963 * we expect an unlocked page.
1964 */
1965 if (folio && (fgp_flags & FGP_FOR_MMAP))
1966 folio_unlock(folio);
1967 }
1968
1969 if (!folio)
1970 return ERR_PTR(-ENOENT);
1971 return folio;
1972}
1973EXPORT_SYMBOL(__filemap_get_folio);
1974
1975static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
1976 xa_mark_t mark)
1977{
1978 struct folio *folio;
1979
1980retry:
1981 if (mark == XA_PRESENT)
1982 folio = xas_find(xas, max);
1983 else
1984 folio = xas_find_marked(xas, max, mark);
1985
1986 if (xas_retry(xas, folio))
1987 goto retry;
1988 /*
1989 * A shadow entry of a recently evicted page, a swap
1990 * entry from shmem/tmpfs or a DAX entry. Return it
1991 * without attempting to raise page count.
1992 */
1993 if (!folio || xa_is_value(folio))
1994 return folio;
1995
1996 if (!folio_try_get(folio))
1997 goto reset;
1998
1999 if (unlikely(folio != xas_reload(xas))) {
2000 folio_put(folio);
2001 goto reset;
2002 }
2003
2004 return folio;
2005reset:
2006 xas_reset(xas);
2007 goto retry;
2008}
2009
2010/**
2011 * find_get_entries - gang pagecache lookup
2012 * @mapping: The address_space to search
2013 * @start: The starting page cache index
2014 * @end: The final page index (inclusive).
2015 * @fbatch: Where the resulting entries are placed.
2016 * @indices: The cache indices corresponding to the entries in @entries
2017 *
2018 * find_get_entries() will search for and return a batch of entries in
2019 * the mapping. The entries are placed in @fbatch. find_get_entries()
2020 * takes a reference on any actual folios it returns.
2021 *
2022 * The entries have ascending indexes. The indices may not be consecutive
2023 * due to not-present entries or large folios.
2024 *
2025 * Any shadow entries of evicted folios, or swap entries from
2026 * shmem/tmpfs, are included in the returned array.
2027 *
2028 * Return: The number of entries which were found.
2029 */
2030unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2031 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2032{
2033 XA_STATE(xas, &mapping->i_pages, *start);
2034 struct folio *folio;
2035
2036 rcu_read_lock();
2037 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2038 indices[fbatch->nr] = xas.xa_index;
2039 if (!folio_batch_add(fbatch, folio))
2040 break;
2041 }
2042
2043 if (folio_batch_count(fbatch)) {
2044 unsigned long nr;
2045 int idx = folio_batch_count(fbatch) - 1;
2046
2047 folio = fbatch->folios[idx];
2048 if (!xa_is_value(folio))
2049 nr = folio_nr_pages(folio);
2050 else
2051 nr = 1 << xa_get_order(&mapping->i_pages, indices[idx]);
2052 *start = round_down(indices[idx] + nr, nr);
2053 }
2054 rcu_read_unlock();
2055
2056 return folio_batch_count(fbatch);
2057}
2058
2059/**
2060 * find_lock_entries - Find a batch of pagecache entries.
2061 * @mapping: The address_space to search.
2062 * @start: The starting page cache index.
2063 * @end: The final page index (inclusive).
2064 * @fbatch: Where the resulting entries are placed.
2065 * @indices: The cache indices of the entries in @fbatch.
2066 *
2067 * find_lock_entries() will return a batch of entries from @mapping.
2068 * Swap, shadow and DAX entries are included. Folios are returned
2069 * locked and with an incremented refcount. Folios which are locked
2070 * by somebody else or under writeback are skipped. Folios which are
2071 * partially outside the range are not returned.
2072 *
2073 * The entries have ascending indexes. The indices may not be consecutive
2074 * due to not-present entries, large folios, folios which could not be
2075 * locked or folios under writeback.
2076 *
2077 * Return: The number of entries which were found.
2078 */
2079unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2080 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2081{
2082 XA_STATE(xas, &mapping->i_pages, *start);
2083 struct folio *folio;
2084
2085 rcu_read_lock();
2086 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2087 unsigned long base;
2088 unsigned long nr;
2089
2090 if (!xa_is_value(folio)) {
2091 nr = folio_nr_pages(folio);
2092 base = folio->index;
2093 /* Omit large folio which begins before the start */
2094 if (base < *start)
2095 goto put;
2096 /* Omit large folio which extends beyond the end */
2097 if (base + nr - 1 > end)
2098 goto put;
2099 if (!folio_trylock(folio))
2100 goto put;
2101 if (folio->mapping != mapping ||
2102 folio_test_writeback(folio))
2103 goto unlock;
2104 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2105 folio);
2106 } else {
2107 nr = 1 << xas_get_order(&xas);
2108 base = xas.xa_index & ~(nr - 1);
2109 /* Omit order>0 value which begins before the start */
2110 if (base < *start)
2111 continue;
2112 /* Omit order>0 value which extends beyond the end */
2113 if (base + nr - 1 > end)
2114 break;
2115 }
2116
2117 /* Update start now so that last update is correct on return */
2118 *start = base + nr;
2119 indices[fbatch->nr] = xas.xa_index;
2120 if (!folio_batch_add(fbatch, folio))
2121 break;
2122 continue;
2123unlock:
2124 folio_unlock(folio);
2125put:
2126 folio_put(folio);
2127 }
2128 rcu_read_unlock();
2129
2130 return folio_batch_count(fbatch);
2131}
2132
2133/**
2134 * filemap_get_folios - Get a batch of folios
2135 * @mapping: The address_space to search
2136 * @start: The starting page index
2137 * @end: The final page index (inclusive)
2138 * @fbatch: The batch to fill.
2139 *
2140 * Search for and return a batch of folios in the mapping starting at
2141 * index @start and up to index @end (inclusive). The folios are returned
2142 * in @fbatch with an elevated reference count.
2143 *
2144 * Return: The number of folios which were found.
2145 * We also update @start to index the next folio for the traversal.
2146 */
2147unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2148 pgoff_t end, struct folio_batch *fbatch)
2149{
2150 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch);
2151}
2152EXPORT_SYMBOL(filemap_get_folios);
2153
2154/**
2155 * filemap_get_folios_contig - Get a batch of contiguous folios
2156 * @mapping: The address_space to search
2157 * @start: The starting page index
2158 * @end: The final page index (inclusive)
2159 * @fbatch: The batch to fill
2160 *
2161 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2162 * except the returned folios are guaranteed to be contiguous. This may
2163 * not return all contiguous folios if the batch gets filled up.
2164 *
2165 * Return: The number of folios found.
2166 * Also update @start to be positioned for traversal of the next folio.
2167 */
2168
2169unsigned filemap_get_folios_contig(struct address_space *mapping,
2170 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2171{
2172 XA_STATE(xas, &mapping->i_pages, *start);
2173 unsigned long nr;
2174 struct folio *folio;
2175
2176 rcu_read_lock();
2177
2178 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2179 folio = xas_next(&xas)) {
2180 if (xas_retry(&xas, folio))
2181 continue;
2182 /*
2183 * If the entry has been swapped out, we can stop looking.
2184 * No current caller is looking for DAX entries.
2185 */
2186 if (xa_is_value(folio))
2187 goto update_start;
2188
2189 /* If we landed in the middle of a THP, continue at its end. */
2190 if (xa_is_sibling(folio))
2191 goto update_start;
2192
2193 if (!folio_try_get(folio))
2194 goto retry;
2195
2196 if (unlikely(folio != xas_reload(&xas)))
2197 goto put_folio;
2198
2199 if (!folio_batch_add(fbatch, folio)) {
2200 nr = folio_nr_pages(folio);
2201 *start = folio->index + nr;
2202 goto out;
2203 }
2204 continue;
2205put_folio:
2206 folio_put(folio);
2207
2208retry:
2209 xas_reset(&xas);
2210 }
2211
2212update_start:
2213 nr = folio_batch_count(fbatch);
2214
2215 if (nr) {
2216 folio = fbatch->folios[nr - 1];
2217 *start = folio_next_index(folio);
2218 }
2219out:
2220 rcu_read_unlock();
2221 return folio_batch_count(fbatch);
2222}
2223EXPORT_SYMBOL(filemap_get_folios_contig);
2224
2225/**
2226 * filemap_get_folios_tag - Get a batch of folios matching @tag
2227 * @mapping: The address_space to search
2228 * @start: The starting page index
2229 * @end: The final page index (inclusive)
2230 * @tag: The tag index
2231 * @fbatch: The batch to fill
2232 *
2233 * The first folio may start before @start; if it does, it will contain
2234 * @start. The final folio may extend beyond @end; if it does, it will
2235 * contain @end. The folios have ascending indices. There may be gaps
2236 * between the folios if there are indices which have no folio in the
2237 * page cache. If folios are added to or removed from the page cache
2238 * while this is running, they may or may not be found by this call.
2239 * Only returns folios that are tagged with @tag.
2240 *
2241 * Return: The number of folios found.
2242 * Also update @start to index the next folio for traversal.
2243 */
2244unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start,
2245 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch)
2246{
2247 XA_STATE(xas, &mapping->i_pages, *start);
2248 struct folio *folio;
2249
2250 rcu_read_lock();
2251 while ((folio = find_get_entry(&xas, end, tag)) != NULL) {
2252 /*
2253 * Shadow entries should never be tagged, but this iteration
2254 * is lockless so there is a window for page reclaim to evict
2255 * a page we saw tagged. Skip over it.
2256 */
2257 if (xa_is_value(folio))
2258 continue;
2259 if (!folio_batch_add(fbatch, folio)) {
2260 unsigned long nr = folio_nr_pages(folio);
2261 *start = folio->index + nr;
2262 goto out;
2263 }
2264 }
2265 /*
2266 * We come here when there is no page beyond @end. We take care to not
2267 * overflow the index @start as it confuses some of the callers. This
2268 * breaks the iteration when there is a page at index -1 but that is
2269 * already broke anyway.
2270 */
2271 if (end == (pgoff_t)-1)
2272 *start = (pgoff_t)-1;
2273 else
2274 *start = end + 1;
2275out:
2276 rcu_read_unlock();
2277
2278 return folio_batch_count(fbatch);
2279}
2280EXPORT_SYMBOL(filemap_get_folios_tag);
2281
2282/*
2283 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2284 * a _large_ part of the i/o request. Imagine the worst scenario:
2285 *
2286 * ---R__________________________________________B__________
2287 * ^ reading here ^ bad block(assume 4k)
2288 *
2289 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2290 * => failing the whole request => read(R) => read(R+1) =>
2291 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2292 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2293 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2294 *
2295 * It is going insane. Fix it by quickly scaling down the readahead size.
2296 */
2297static void shrink_readahead_size_eio(struct file_ra_state *ra)
2298{
2299 ra->ra_pages /= 4;
2300}
2301
2302/*
2303 * filemap_get_read_batch - Get a batch of folios for read
2304 *
2305 * Get a batch of folios which represent a contiguous range of bytes in
2306 * the file. No exceptional entries will be returned. If @index is in
2307 * the middle of a folio, the entire folio will be returned. The last
2308 * folio in the batch may have the readahead flag set or the uptodate flag
2309 * clear so that the caller can take the appropriate action.
2310 */
2311static void filemap_get_read_batch(struct address_space *mapping,
2312 pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
2313{
2314 XA_STATE(xas, &mapping->i_pages, index);
2315 struct folio *folio;
2316
2317 rcu_read_lock();
2318 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2319 if (xas_retry(&xas, folio))
2320 continue;
2321 if (xas.xa_index > max || xa_is_value(folio))
2322 break;
2323 if (xa_is_sibling(folio))
2324 break;
2325 if (!folio_try_get(folio))
2326 goto retry;
2327
2328 if (unlikely(folio != xas_reload(&xas)))
2329 goto put_folio;
2330
2331 if (!folio_batch_add(fbatch, folio))
2332 break;
2333 if (!folio_test_uptodate(folio))
2334 break;
2335 if (folio_test_readahead(folio))
2336 break;
2337 xas_advance(&xas, folio_next_index(folio) - 1);
2338 continue;
2339put_folio:
2340 folio_put(folio);
2341retry:
2342 xas_reset(&xas);
2343 }
2344 rcu_read_unlock();
2345}
2346
2347static int filemap_read_folio(struct file *file, filler_t filler,
2348 struct folio *folio)
2349{
2350 bool workingset = folio_test_workingset(folio);
2351 unsigned long pflags;
2352 int error;
2353
2354 /* Start the actual read. The read will unlock the page. */
2355 if (unlikely(workingset))
2356 psi_memstall_enter(&pflags);
2357 error = filler(file, folio);
2358 if (unlikely(workingset))
2359 psi_memstall_leave(&pflags);
2360 if (error)
2361 return error;
2362
2363 error = folio_wait_locked_killable(folio);
2364 if (error)
2365 return error;
2366 if (folio_test_uptodate(folio))
2367 return 0;
2368 if (file)
2369 shrink_readahead_size_eio(&file->f_ra);
2370 return -EIO;
2371}
2372
2373static bool filemap_range_uptodate(struct address_space *mapping,
2374 loff_t pos, size_t count, struct folio *folio,
2375 bool need_uptodate)
2376{
2377 if (folio_test_uptodate(folio))
2378 return true;
2379 /* pipes can't handle partially uptodate pages */
2380 if (need_uptodate)
2381 return false;
2382 if (!mapping->a_ops->is_partially_uptodate)
2383 return false;
2384 if (mapping->host->i_blkbits >= folio_shift(folio))
2385 return false;
2386
2387 if (folio_pos(folio) > pos) {
2388 count -= folio_pos(folio) - pos;
2389 pos = 0;
2390 } else {
2391 pos -= folio_pos(folio);
2392 }
2393
2394 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2395}
2396
2397static int filemap_update_page(struct kiocb *iocb,
2398 struct address_space *mapping, size_t count,
2399 struct folio *folio, bool need_uptodate)
2400{
2401 int error;
2402
2403 if (iocb->ki_flags & IOCB_NOWAIT) {
2404 if (!filemap_invalidate_trylock_shared(mapping))
2405 return -EAGAIN;
2406 } else {
2407 filemap_invalidate_lock_shared(mapping);
2408 }
2409
2410 if (!folio_trylock(folio)) {
2411 error = -EAGAIN;
2412 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2413 goto unlock_mapping;
2414 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2415 filemap_invalidate_unlock_shared(mapping);
2416 /*
2417 * This is where we usually end up waiting for a
2418 * previously submitted readahead to finish.
2419 */
2420 folio_put_wait_locked(folio, TASK_KILLABLE);
2421 return AOP_TRUNCATED_PAGE;
2422 }
2423 error = __folio_lock_async(folio, iocb->ki_waitq);
2424 if (error)
2425 goto unlock_mapping;
2426 }
2427
2428 error = AOP_TRUNCATED_PAGE;
2429 if (!folio->mapping)
2430 goto unlock;
2431
2432 error = 0;
2433 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio,
2434 need_uptodate))
2435 goto unlock;
2436
2437 error = -EAGAIN;
2438 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2439 goto unlock;
2440
2441 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2442 folio);
2443 goto unlock_mapping;
2444unlock:
2445 folio_unlock(folio);
2446unlock_mapping:
2447 filemap_invalidate_unlock_shared(mapping);
2448 if (error == AOP_TRUNCATED_PAGE)
2449 folio_put(folio);
2450 return error;
2451}
2452
2453static int filemap_create_folio(struct file *file,
2454 struct address_space *mapping, loff_t pos,
2455 struct folio_batch *fbatch)
2456{
2457 struct folio *folio;
2458 int error;
2459 unsigned int min_order = mapping_min_folio_order(mapping);
2460 pgoff_t index;
2461
2462 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), min_order);
2463 if (!folio)
2464 return -ENOMEM;
2465
2466 /*
2467 * Protect against truncate / hole punch. Grabbing invalidate_lock
2468 * here assures we cannot instantiate and bring uptodate new
2469 * pagecache folios after evicting page cache during truncate
2470 * and before actually freeing blocks. Note that we could
2471 * release invalidate_lock after inserting the folio into
2472 * the page cache as the locked folio would then be enough to
2473 * synchronize with hole punching. But there are code paths
2474 * such as filemap_update_page() filling in partially uptodate
2475 * pages or ->readahead() that need to hold invalidate_lock
2476 * while mapping blocks for IO so let's hold the lock here as
2477 * well to keep locking rules simple.
2478 */
2479 filemap_invalidate_lock_shared(mapping);
2480 index = (pos >> (PAGE_SHIFT + min_order)) << min_order;
2481 error = filemap_add_folio(mapping, folio, index,
2482 mapping_gfp_constraint(mapping, GFP_KERNEL));
2483 if (error == -EEXIST)
2484 error = AOP_TRUNCATED_PAGE;
2485 if (error)
2486 goto error;
2487
2488 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
2489 if (error)
2490 goto error;
2491
2492 filemap_invalidate_unlock_shared(mapping);
2493 folio_batch_add(fbatch, folio);
2494 return 0;
2495error:
2496 filemap_invalidate_unlock_shared(mapping);
2497 folio_put(folio);
2498 return error;
2499}
2500
2501static int filemap_readahead(struct kiocb *iocb, struct file *file,
2502 struct address_space *mapping, struct folio *folio,
2503 pgoff_t last_index)
2504{
2505 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2506
2507 if (iocb->ki_flags & IOCB_NOIO)
2508 return -EAGAIN;
2509 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2510 return 0;
2511}
2512
2513static int filemap_get_pages(struct kiocb *iocb, size_t count,
2514 struct folio_batch *fbatch, bool need_uptodate)
2515{
2516 struct file *filp = iocb->ki_filp;
2517 struct address_space *mapping = filp->f_mapping;
2518 struct file_ra_state *ra = &filp->f_ra;
2519 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2520 pgoff_t last_index;
2521 struct folio *folio;
2522 unsigned int flags;
2523 int err = 0;
2524
2525 /* "last_index" is the index of the page beyond the end of the read */
2526 last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
2527retry:
2528 if (fatal_signal_pending(current))
2529 return -EINTR;
2530
2531 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2532 if (!folio_batch_count(fbatch)) {
2533 if (iocb->ki_flags & IOCB_NOIO)
2534 return -EAGAIN;
2535 if (iocb->ki_flags & IOCB_NOWAIT)
2536 flags = memalloc_noio_save();
2537 page_cache_sync_readahead(mapping, ra, filp, index,
2538 last_index - index);
2539 if (iocb->ki_flags & IOCB_NOWAIT)
2540 memalloc_noio_restore(flags);
2541 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2542 }
2543 if (!folio_batch_count(fbatch)) {
2544 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2545 return -EAGAIN;
2546 err = filemap_create_folio(filp, mapping, iocb->ki_pos, fbatch);
2547 if (err == AOP_TRUNCATED_PAGE)
2548 goto retry;
2549 return err;
2550 }
2551
2552 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2553 if (folio_test_readahead(folio)) {
2554 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2555 if (err)
2556 goto err;
2557 }
2558 if (!folio_test_uptodate(folio)) {
2559 if ((iocb->ki_flags & IOCB_WAITQ) &&
2560 folio_batch_count(fbatch) > 1)
2561 iocb->ki_flags |= IOCB_NOWAIT;
2562 err = filemap_update_page(iocb, mapping, count, folio,
2563 need_uptodate);
2564 if (err)
2565 goto err;
2566 }
2567
2568 trace_mm_filemap_get_pages(mapping, index, last_index - 1);
2569 return 0;
2570err:
2571 if (err < 0)
2572 folio_put(folio);
2573 if (likely(--fbatch->nr))
2574 return 0;
2575 if (err == AOP_TRUNCATED_PAGE)
2576 goto retry;
2577 return err;
2578}
2579
2580static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2581{
2582 unsigned int shift = folio_shift(folio);
2583
2584 return (pos1 >> shift == pos2 >> shift);
2585}
2586
2587/**
2588 * filemap_read - Read data from the page cache.
2589 * @iocb: The iocb to read.
2590 * @iter: Destination for the data.
2591 * @already_read: Number of bytes already read by the caller.
2592 *
2593 * Copies data from the page cache. If the data is not currently present,
2594 * uses the readahead and read_folio address_space operations to fetch it.
2595 *
2596 * Return: Total number of bytes copied, including those already read by
2597 * the caller. If an error happens before any bytes are copied, returns
2598 * a negative error number.
2599 */
2600ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2601 ssize_t already_read)
2602{
2603 struct file *filp = iocb->ki_filp;
2604 struct file_ra_state *ra = &filp->f_ra;
2605 struct address_space *mapping = filp->f_mapping;
2606 struct inode *inode = mapping->host;
2607 struct folio_batch fbatch;
2608 int i, error = 0;
2609 bool writably_mapped;
2610 loff_t isize, end_offset;
2611 loff_t last_pos = ra->prev_pos;
2612
2613 if (unlikely(iocb->ki_pos < 0))
2614 return -EINVAL;
2615 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2616 return 0;
2617 if (unlikely(!iov_iter_count(iter)))
2618 return 0;
2619
2620 iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos);
2621 folio_batch_init(&fbatch);
2622
2623 do {
2624 cond_resched();
2625
2626 /*
2627 * If we've already successfully copied some data, then we
2628 * can no longer safely return -EIOCBQUEUED. Hence mark
2629 * an async read NOWAIT at that point.
2630 */
2631 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2632 iocb->ki_flags |= IOCB_NOWAIT;
2633
2634 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2635 break;
2636
2637 error = filemap_get_pages(iocb, iter->count, &fbatch, false);
2638 if (error < 0)
2639 break;
2640
2641 /*
2642 * i_size must be checked after we know the pages are Uptodate.
2643 *
2644 * Checking i_size after the check allows us to calculate
2645 * the correct value for "nr", which means the zero-filled
2646 * part of the page is not copied back to userspace (unless
2647 * another truncate extends the file - this is desired though).
2648 */
2649 isize = i_size_read(inode);
2650 if (unlikely(iocb->ki_pos >= isize))
2651 goto put_folios;
2652 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2653
2654 /*
2655 * Once we start copying data, we don't want to be touching any
2656 * cachelines that might be contended:
2657 */
2658 writably_mapped = mapping_writably_mapped(mapping);
2659
2660 /*
2661 * When a read accesses the same folio several times, only
2662 * mark it as accessed the first time.
2663 */
2664 if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
2665 fbatch.folios[0]))
2666 folio_mark_accessed(fbatch.folios[0]);
2667
2668 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2669 struct folio *folio = fbatch.folios[i];
2670 size_t fsize = folio_size(folio);
2671 size_t offset = iocb->ki_pos & (fsize - 1);
2672 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2673 fsize - offset);
2674 size_t copied;
2675
2676 if (end_offset < folio_pos(folio))
2677 break;
2678 if (i > 0)
2679 folio_mark_accessed(folio);
2680 /*
2681 * If users can be writing to this folio using arbitrary
2682 * virtual addresses, take care of potential aliasing
2683 * before reading the folio on the kernel side.
2684 */
2685 if (writably_mapped)
2686 flush_dcache_folio(folio);
2687
2688 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2689
2690 already_read += copied;
2691 iocb->ki_pos += copied;
2692 last_pos = iocb->ki_pos;
2693
2694 if (copied < bytes) {
2695 error = -EFAULT;
2696 break;
2697 }
2698 }
2699put_folios:
2700 for (i = 0; i < folio_batch_count(&fbatch); i++)
2701 folio_put(fbatch.folios[i]);
2702 folio_batch_init(&fbatch);
2703 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2704
2705 file_accessed(filp);
2706 ra->prev_pos = last_pos;
2707 return already_read ? already_read : error;
2708}
2709EXPORT_SYMBOL_GPL(filemap_read);
2710
2711int kiocb_write_and_wait(struct kiocb *iocb, size_t count)
2712{
2713 struct address_space *mapping = iocb->ki_filp->f_mapping;
2714 loff_t pos = iocb->ki_pos;
2715 loff_t end = pos + count - 1;
2716
2717 if (iocb->ki_flags & IOCB_NOWAIT) {
2718 if (filemap_range_needs_writeback(mapping, pos, end))
2719 return -EAGAIN;
2720 return 0;
2721 }
2722
2723 return filemap_write_and_wait_range(mapping, pos, end);
2724}
2725EXPORT_SYMBOL_GPL(kiocb_write_and_wait);
2726
2727int filemap_invalidate_pages(struct address_space *mapping,
2728 loff_t pos, loff_t end, bool nowait)
2729{
2730 int ret;
2731
2732 if (nowait) {
2733 /* we could block if there are any pages in the range */
2734 if (filemap_range_has_page(mapping, pos, end))
2735 return -EAGAIN;
2736 } else {
2737 ret = filemap_write_and_wait_range(mapping, pos, end);
2738 if (ret)
2739 return ret;
2740 }
2741
2742 /*
2743 * After a write we want buffered reads to be sure to go to disk to get
2744 * the new data. We invalidate clean cached page from the region we're
2745 * about to write. We do this *before* the write so that we can return
2746 * without clobbering -EIOCBQUEUED from ->direct_IO().
2747 */
2748 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
2749 end >> PAGE_SHIFT);
2750}
2751
2752int kiocb_invalidate_pages(struct kiocb *iocb, size_t count)
2753{
2754 struct address_space *mapping = iocb->ki_filp->f_mapping;
2755
2756 return filemap_invalidate_pages(mapping, iocb->ki_pos,
2757 iocb->ki_pos + count - 1,
2758 iocb->ki_flags & IOCB_NOWAIT);
2759}
2760EXPORT_SYMBOL_GPL(kiocb_invalidate_pages);
2761
2762/**
2763 * generic_file_read_iter - generic filesystem read routine
2764 * @iocb: kernel I/O control block
2765 * @iter: destination for the data read
2766 *
2767 * This is the "read_iter()" routine for all filesystems
2768 * that can use the page cache directly.
2769 *
2770 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2771 * be returned when no data can be read without waiting for I/O requests
2772 * to complete; it doesn't prevent readahead.
2773 *
2774 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2775 * requests shall be made for the read or for readahead. When no data
2776 * can be read, -EAGAIN shall be returned. When readahead would be
2777 * triggered, a partial, possibly empty read shall be returned.
2778 *
2779 * Return:
2780 * * number of bytes copied, even for partial reads
2781 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2782 */
2783ssize_t
2784generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2785{
2786 size_t count = iov_iter_count(iter);
2787 ssize_t retval = 0;
2788
2789 if (!count)
2790 return 0; /* skip atime */
2791
2792 if (iocb->ki_flags & IOCB_DIRECT) {
2793 struct file *file = iocb->ki_filp;
2794 struct address_space *mapping = file->f_mapping;
2795 struct inode *inode = mapping->host;
2796
2797 retval = kiocb_write_and_wait(iocb, count);
2798 if (retval < 0)
2799 return retval;
2800 file_accessed(file);
2801
2802 retval = mapping->a_ops->direct_IO(iocb, iter);
2803 if (retval >= 0) {
2804 iocb->ki_pos += retval;
2805 count -= retval;
2806 }
2807 if (retval != -EIOCBQUEUED)
2808 iov_iter_revert(iter, count - iov_iter_count(iter));
2809
2810 /*
2811 * Btrfs can have a short DIO read if we encounter
2812 * compressed extents, so if there was an error, or if
2813 * we've already read everything we wanted to, or if
2814 * there was a short read because we hit EOF, go ahead
2815 * and return. Otherwise fallthrough to buffered io for
2816 * the rest of the read. Buffered reads will not work for
2817 * DAX files, so don't bother trying.
2818 */
2819 if (retval < 0 || !count || IS_DAX(inode))
2820 return retval;
2821 if (iocb->ki_pos >= i_size_read(inode))
2822 return retval;
2823 }
2824
2825 return filemap_read(iocb, iter, retval);
2826}
2827EXPORT_SYMBOL(generic_file_read_iter);
2828
2829/*
2830 * Splice subpages from a folio into a pipe.
2831 */
2832size_t splice_folio_into_pipe(struct pipe_inode_info *pipe,
2833 struct folio *folio, loff_t fpos, size_t size)
2834{
2835 struct page *page;
2836 size_t spliced = 0, offset = offset_in_folio(folio, fpos);
2837
2838 page = folio_page(folio, offset / PAGE_SIZE);
2839 size = min(size, folio_size(folio) - offset);
2840 offset %= PAGE_SIZE;
2841
2842 while (spliced < size &&
2843 !pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
2844 struct pipe_buffer *buf = pipe_head_buf(pipe);
2845 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced);
2846
2847 *buf = (struct pipe_buffer) {
2848 .ops = &page_cache_pipe_buf_ops,
2849 .page = page,
2850 .offset = offset,
2851 .len = part,
2852 };
2853 folio_get(folio);
2854 pipe->head++;
2855 page++;
2856 spliced += part;
2857 offset = 0;
2858 }
2859
2860 return spliced;
2861}
2862
2863/**
2864 * filemap_splice_read - Splice data from a file's pagecache into a pipe
2865 * @in: The file to read from
2866 * @ppos: Pointer to the file position to read from
2867 * @pipe: The pipe to splice into
2868 * @len: The amount to splice
2869 * @flags: The SPLICE_F_* flags
2870 *
2871 * This function gets folios from a file's pagecache and splices them into the
2872 * pipe. Readahead will be called as necessary to fill more folios. This may
2873 * be used for blockdevs also.
2874 *
2875 * Return: On success, the number of bytes read will be returned and *@ppos
2876 * will be updated if appropriate; 0 will be returned if there is no more data
2877 * to be read; -EAGAIN will be returned if the pipe had no space, and some
2878 * other negative error code will be returned on error. A short read may occur
2879 * if the pipe has insufficient space, we reach the end of the data or we hit a
2880 * hole.
2881 */
2882ssize_t filemap_splice_read(struct file *in, loff_t *ppos,
2883 struct pipe_inode_info *pipe,
2884 size_t len, unsigned int flags)
2885{
2886 struct folio_batch fbatch;
2887 struct kiocb iocb;
2888 size_t total_spliced = 0, used, npages;
2889 loff_t isize, end_offset;
2890 bool writably_mapped;
2891 int i, error = 0;
2892
2893 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes))
2894 return 0;
2895
2896 init_sync_kiocb(&iocb, in);
2897 iocb.ki_pos = *ppos;
2898
2899 /* Work out how much data we can actually add into the pipe */
2900 used = pipe_occupancy(pipe->head, pipe->tail);
2901 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2902 len = min_t(size_t, len, npages * PAGE_SIZE);
2903
2904 folio_batch_init(&fbatch);
2905
2906 do {
2907 cond_resched();
2908
2909 if (*ppos >= i_size_read(in->f_mapping->host))
2910 break;
2911
2912 iocb.ki_pos = *ppos;
2913 error = filemap_get_pages(&iocb, len, &fbatch, true);
2914 if (error < 0)
2915 break;
2916
2917 /*
2918 * i_size must be checked after we know the pages are Uptodate.
2919 *
2920 * Checking i_size after the check allows us to calculate
2921 * the correct value for "nr", which means the zero-filled
2922 * part of the page is not copied back to userspace (unless
2923 * another truncate extends the file - this is desired though).
2924 */
2925 isize = i_size_read(in->f_mapping->host);
2926 if (unlikely(*ppos >= isize))
2927 break;
2928 end_offset = min_t(loff_t, isize, *ppos + len);
2929
2930 /*
2931 * Once we start copying data, we don't want to be touching any
2932 * cachelines that might be contended:
2933 */
2934 writably_mapped = mapping_writably_mapped(in->f_mapping);
2935
2936 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2937 struct folio *folio = fbatch.folios[i];
2938 size_t n;
2939
2940 if (folio_pos(folio) >= end_offset)
2941 goto out;
2942 folio_mark_accessed(folio);
2943
2944 /*
2945 * If users can be writing to this folio using arbitrary
2946 * virtual addresses, take care of potential aliasing
2947 * before reading the folio on the kernel side.
2948 */
2949 if (writably_mapped)
2950 flush_dcache_folio(folio);
2951
2952 n = min_t(loff_t, len, isize - *ppos);
2953 n = splice_folio_into_pipe(pipe, folio, *ppos, n);
2954 if (!n)
2955 goto out;
2956 len -= n;
2957 total_spliced += n;
2958 *ppos += n;
2959 in->f_ra.prev_pos = *ppos;
2960 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
2961 goto out;
2962 }
2963
2964 folio_batch_release(&fbatch);
2965 } while (len);
2966
2967out:
2968 folio_batch_release(&fbatch);
2969 file_accessed(in);
2970
2971 return total_spliced ? total_spliced : error;
2972}
2973EXPORT_SYMBOL(filemap_splice_read);
2974
2975static inline loff_t folio_seek_hole_data(struct xa_state *xas,
2976 struct address_space *mapping, struct folio *folio,
2977 loff_t start, loff_t end, bool seek_data)
2978{
2979 const struct address_space_operations *ops = mapping->a_ops;
2980 size_t offset, bsz = i_blocksize(mapping->host);
2981
2982 if (xa_is_value(folio) || folio_test_uptodate(folio))
2983 return seek_data ? start : end;
2984 if (!ops->is_partially_uptodate)
2985 return seek_data ? end : start;
2986
2987 xas_pause(xas);
2988 rcu_read_unlock();
2989 folio_lock(folio);
2990 if (unlikely(folio->mapping != mapping))
2991 goto unlock;
2992
2993 offset = offset_in_folio(folio, start) & ~(bsz - 1);
2994
2995 do {
2996 if (ops->is_partially_uptodate(folio, offset, bsz) ==
2997 seek_data)
2998 break;
2999 start = (start + bsz) & ~((u64)bsz - 1);
3000 offset += bsz;
3001 } while (offset < folio_size(folio));
3002unlock:
3003 folio_unlock(folio);
3004 rcu_read_lock();
3005 return start;
3006}
3007
3008static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
3009{
3010 if (xa_is_value(folio))
3011 return PAGE_SIZE << xas_get_order(xas);
3012 return folio_size(folio);
3013}
3014
3015/**
3016 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3017 * @mapping: Address space to search.
3018 * @start: First byte to consider.
3019 * @end: Limit of search (exclusive).
3020 * @whence: Either SEEK_HOLE or SEEK_DATA.
3021 *
3022 * If the page cache knows which blocks contain holes and which blocks
3023 * contain data, your filesystem can use this function to implement
3024 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
3025 * entirely memory-based such as tmpfs, and filesystems which support
3026 * unwritten extents.
3027 *
3028 * Return: The requested offset on success, or -ENXIO if @whence specifies
3029 * SEEK_DATA and there is no data after @start. There is an implicit hole
3030 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
3031 * and @end contain data.
3032 */
3033loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
3034 loff_t end, int whence)
3035{
3036 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
3037 pgoff_t max = (end - 1) >> PAGE_SHIFT;
3038 bool seek_data = (whence == SEEK_DATA);
3039 struct folio *folio;
3040
3041 if (end <= start)
3042 return -ENXIO;
3043
3044 rcu_read_lock();
3045 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
3046 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
3047 size_t seek_size;
3048
3049 if (start < pos) {
3050 if (!seek_data)
3051 goto unlock;
3052 start = pos;
3053 }
3054
3055 seek_size = seek_folio_size(&xas, folio);
3056 pos = round_up((u64)pos + 1, seek_size);
3057 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
3058 seek_data);
3059 if (start < pos)
3060 goto unlock;
3061 if (start >= end)
3062 break;
3063 if (seek_size > PAGE_SIZE)
3064 xas_set(&xas, pos >> PAGE_SHIFT);
3065 if (!xa_is_value(folio))
3066 folio_put(folio);
3067 }
3068 if (seek_data)
3069 start = -ENXIO;
3070unlock:
3071 rcu_read_unlock();
3072 if (folio && !xa_is_value(folio))
3073 folio_put(folio);
3074 if (start > end)
3075 return end;
3076 return start;
3077}
3078
3079#ifdef CONFIG_MMU
3080#define MMAP_LOTSAMISS (100)
3081/*
3082 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
3083 * @vmf - the vm_fault for this fault.
3084 * @folio - the folio to lock.
3085 * @fpin - the pointer to the file we may pin (or is already pinned).
3086 *
3087 * This works similar to lock_folio_or_retry in that it can drop the
3088 * mmap_lock. It differs in that it actually returns the folio locked
3089 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
3090 * to drop the mmap_lock then fpin will point to the pinned file and
3091 * needs to be fput()'ed at a later point.
3092 */
3093static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
3094 struct file **fpin)
3095{
3096 if (folio_trylock(folio))
3097 return 1;
3098
3099 /*
3100 * NOTE! This will make us return with VM_FAULT_RETRY, but with
3101 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
3102 * is supposed to work. We have way too many special cases..
3103 */
3104 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
3105 return 0;
3106
3107 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
3108 if (vmf->flags & FAULT_FLAG_KILLABLE) {
3109 if (__folio_lock_killable(folio)) {
3110 /*
3111 * We didn't have the right flags to drop the
3112 * fault lock, but all fault_handlers only check
3113 * for fatal signals if we return VM_FAULT_RETRY,
3114 * so we need to drop the fault lock here and
3115 * return 0 if we don't have a fpin.
3116 */
3117 if (*fpin == NULL)
3118 release_fault_lock(vmf);
3119 return 0;
3120 }
3121 } else
3122 __folio_lock(folio);
3123
3124 return 1;
3125}
3126
3127/*
3128 * Synchronous readahead happens when we don't even find a page in the page
3129 * cache at all. We don't want to perform IO under the mmap sem, so if we have
3130 * to drop the mmap sem we return the file that was pinned in order for us to do
3131 * that. If we didn't pin a file then we return NULL. The file that is
3132 * returned needs to be fput()'ed when we're done with it.
3133 */
3134static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3135{
3136 struct file *file = vmf->vma->vm_file;
3137 struct file_ra_state *ra = &file->f_ra;
3138 struct address_space *mapping = file->f_mapping;
3139 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3140 struct file *fpin = NULL;
3141 unsigned long vm_flags = vmf->vma->vm_flags;
3142 unsigned int mmap_miss;
3143
3144#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3145 /* Use the readahead code, even if readahead is disabled */
3146 if ((vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER) {
3147 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3148 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3149 ra->size = HPAGE_PMD_NR;
3150 /*
3151 * Fetch two PMD folios, so we get the chance to actually
3152 * readahead, unless we've been told not to.
3153 */
3154 if (!(vm_flags & VM_RAND_READ))
3155 ra->size *= 2;
3156 ra->async_size = HPAGE_PMD_NR;
3157 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3158 return fpin;
3159 }
3160#endif
3161
3162 /* If we don't want any read-ahead, don't bother */
3163 if (vm_flags & VM_RAND_READ)
3164 return fpin;
3165 if (!ra->ra_pages)
3166 return fpin;
3167
3168 if (vm_flags & VM_SEQ_READ) {
3169 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3170 page_cache_sync_ra(&ractl, ra->ra_pages);
3171 return fpin;
3172 }
3173
3174 /* Avoid banging the cache line if not needed */
3175 mmap_miss = READ_ONCE(ra->mmap_miss);
3176 if (mmap_miss < MMAP_LOTSAMISS * 10)
3177 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3178
3179 /*
3180 * Do we miss much more than hit in this file? If so,
3181 * stop bothering with read-ahead. It will only hurt.
3182 */
3183 if (mmap_miss > MMAP_LOTSAMISS)
3184 return fpin;
3185
3186 /*
3187 * mmap read-around
3188 */
3189 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3190 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3191 ra->size = ra->ra_pages;
3192 ra->async_size = ra->ra_pages / 4;
3193 ractl._index = ra->start;
3194 page_cache_ra_order(&ractl, ra, 0);
3195 return fpin;
3196}
3197
3198/*
3199 * Asynchronous readahead happens when we find the page and PG_readahead,
3200 * so we want to possibly extend the readahead further. We return the file that
3201 * was pinned if we have to drop the mmap_lock in order to do IO.
3202 */
3203static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3204 struct folio *folio)
3205{
3206 struct file *file = vmf->vma->vm_file;
3207 struct file_ra_state *ra = &file->f_ra;
3208 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3209 struct file *fpin = NULL;
3210 unsigned int mmap_miss;
3211
3212 /* If we don't want any read-ahead, don't bother */
3213 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3214 return fpin;
3215
3216 mmap_miss = READ_ONCE(ra->mmap_miss);
3217 if (mmap_miss)
3218 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3219
3220 if (folio_test_readahead(folio)) {
3221 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3222 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3223 }
3224 return fpin;
3225}
3226
3227static vm_fault_t filemap_fault_recheck_pte_none(struct vm_fault *vmf)
3228{
3229 struct vm_area_struct *vma = vmf->vma;
3230 vm_fault_t ret = 0;
3231 pte_t *ptep;
3232
3233 /*
3234 * We might have COW'ed a pagecache folio and might now have an mlocked
3235 * anon folio mapped. The original pagecache folio is not mlocked and
3236 * might have been evicted. During a read+clear/modify/write update of
3237 * the PTE, such as done in do_numa_page()/change_pte_range(), we
3238 * temporarily clear the PTE under PT lock and might detect it here as
3239 * "none" when not holding the PT lock.
3240 *
3241 * Not rechecking the PTE under PT lock could result in an unexpected
3242 * major fault in an mlock'ed region. Recheck only for this special
3243 * scenario while holding the PT lock, to not degrade non-mlocked
3244 * scenarios. Recheck the PTE without PT lock firstly, thereby reducing
3245 * the number of times we hold PT lock.
3246 */
3247 if (!(vma->vm_flags & VM_LOCKED))
3248 return 0;
3249
3250 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
3251 return 0;
3252
3253 ptep = pte_offset_map_ro_nolock(vma->vm_mm, vmf->pmd, vmf->address,
3254 &vmf->ptl);
3255 if (unlikely(!ptep))
3256 return VM_FAULT_NOPAGE;
3257
3258 if (unlikely(!pte_none(ptep_get_lockless(ptep)))) {
3259 ret = VM_FAULT_NOPAGE;
3260 } else {
3261 spin_lock(vmf->ptl);
3262 if (unlikely(!pte_none(ptep_get(ptep))))
3263 ret = VM_FAULT_NOPAGE;
3264 spin_unlock(vmf->ptl);
3265 }
3266 pte_unmap(ptep);
3267 return ret;
3268}
3269
3270/**
3271 * filemap_fault - read in file data for page fault handling
3272 * @vmf: struct vm_fault containing details of the fault
3273 *
3274 * filemap_fault() is invoked via the vma operations vector for a
3275 * mapped memory region to read in file data during a page fault.
3276 *
3277 * The goto's are kind of ugly, but this streamlines the normal case of having
3278 * it in the page cache, and handles the special cases reasonably without
3279 * having a lot of duplicated code.
3280 *
3281 * vma->vm_mm->mmap_lock must be held on entry.
3282 *
3283 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3284 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3285 *
3286 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3287 * has not been released.
3288 *
3289 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3290 *
3291 * Return: bitwise-OR of %VM_FAULT_ codes.
3292 */
3293vm_fault_t filemap_fault(struct vm_fault *vmf)
3294{
3295 int error;
3296 struct file *file = vmf->vma->vm_file;
3297 struct file *fpin = NULL;
3298 struct address_space *mapping = file->f_mapping;
3299 struct inode *inode = mapping->host;
3300 pgoff_t max_idx, index = vmf->pgoff;
3301 struct folio *folio;
3302 vm_fault_t ret = 0;
3303 bool mapping_locked = false;
3304
3305 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3306 if (unlikely(index >= max_idx))
3307 return VM_FAULT_SIGBUS;
3308
3309 trace_mm_filemap_fault(mapping, index);
3310
3311 /*
3312 * Do we have something in the page cache already?
3313 */
3314 folio = filemap_get_folio(mapping, index);
3315 if (likely(!IS_ERR(folio))) {
3316 /*
3317 * We found the page, so try async readahead before waiting for
3318 * the lock.
3319 */
3320 if (!(vmf->flags & FAULT_FLAG_TRIED))
3321 fpin = do_async_mmap_readahead(vmf, folio);
3322 if (unlikely(!folio_test_uptodate(folio))) {
3323 filemap_invalidate_lock_shared(mapping);
3324 mapping_locked = true;
3325 }
3326 } else {
3327 ret = filemap_fault_recheck_pte_none(vmf);
3328 if (unlikely(ret))
3329 return ret;
3330
3331 /* No page in the page cache at all */
3332 count_vm_event(PGMAJFAULT);
3333 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3334 ret = VM_FAULT_MAJOR;
3335 fpin = do_sync_mmap_readahead(vmf);
3336retry_find:
3337 /*
3338 * See comment in filemap_create_folio() why we need
3339 * invalidate_lock
3340 */
3341 if (!mapping_locked) {
3342 filemap_invalidate_lock_shared(mapping);
3343 mapping_locked = true;
3344 }
3345 folio = __filemap_get_folio(mapping, index,
3346 FGP_CREAT|FGP_FOR_MMAP,
3347 vmf->gfp_mask);
3348 if (IS_ERR(folio)) {
3349 if (fpin)
3350 goto out_retry;
3351 filemap_invalidate_unlock_shared(mapping);
3352 return VM_FAULT_OOM;
3353 }
3354 }
3355
3356 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3357 goto out_retry;
3358
3359 /* Did it get truncated? */
3360 if (unlikely(folio->mapping != mapping)) {
3361 folio_unlock(folio);
3362 folio_put(folio);
3363 goto retry_find;
3364 }
3365 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3366
3367 /*
3368 * We have a locked folio in the page cache, now we need to check
3369 * that it's up-to-date. If not, it is going to be due to an error,
3370 * or because readahead was otherwise unable to retrieve it.
3371 */
3372 if (unlikely(!folio_test_uptodate(folio))) {
3373 /*
3374 * If the invalidate lock is not held, the folio was in cache
3375 * and uptodate and now it is not. Strange but possible since we
3376 * didn't hold the page lock all the time. Let's drop
3377 * everything, get the invalidate lock and try again.
3378 */
3379 if (!mapping_locked) {
3380 folio_unlock(folio);
3381 folio_put(folio);
3382 goto retry_find;
3383 }
3384
3385 /*
3386 * OK, the folio is really not uptodate. This can be because the
3387 * VMA has the VM_RAND_READ flag set, or because an error
3388 * arose. Let's read it in directly.
3389 */
3390 goto page_not_uptodate;
3391 }
3392
3393 /*
3394 * We've made it this far and we had to drop our mmap_lock, now is the
3395 * time to return to the upper layer and have it re-find the vma and
3396 * redo the fault.
3397 */
3398 if (fpin) {
3399 folio_unlock(folio);
3400 goto out_retry;
3401 }
3402 if (mapping_locked)
3403 filemap_invalidate_unlock_shared(mapping);
3404
3405 /*
3406 * Found the page and have a reference on it.
3407 * We must recheck i_size under page lock.
3408 */
3409 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3410 if (unlikely(index >= max_idx)) {
3411 folio_unlock(folio);
3412 folio_put(folio);
3413 return VM_FAULT_SIGBUS;
3414 }
3415
3416 vmf->page = folio_file_page(folio, index);
3417 return ret | VM_FAULT_LOCKED;
3418
3419page_not_uptodate:
3420 /*
3421 * Umm, take care of errors if the page isn't up-to-date.
3422 * Try to re-read it _once_. We do this synchronously,
3423 * because there really aren't any performance issues here
3424 * and we need to check for errors.
3425 */
3426 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3427 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3428 if (fpin)
3429 goto out_retry;
3430 folio_put(folio);
3431
3432 if (!error || error == AOP_TRUNCATED_PAGE)
3433 goto retry_find;
3434 filemap_invalidate_unlock_shared(mapping);
3435
3436 return VM_FAULT_SIGBUS;
3437
3438out_retry:
3439 /*
3440 * We dropped the mmap_lock, we need to return to the fault handler to
3441 * re-find the vma and come back and find our hopefully still populated
3442 * page.
3443 */
3444 if (!IS_ERR(folio))
3445 folio_put(folio);
3446 if (mapping_locked)
3447 filemap_invalidate_unlock_shared(mapping);
3448 if (fpin)
3449 fput(fpin);
3450 return ret | VM_FAULT_RETRY;
3451}
3452EXPORT_SYMBOL(filemap_fault);
3453
3454static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio,
3455 pgoff_t start)
3456{
3457 struct mm_struct *mm = vmf->vma->vm_mm;
3458
3459 /* Huge page is mapped? No need to proceed. */
3460 if (pmd_trans_huge(*vmf->pmd)) {
3461 folio_unlock(folio);
3462 folio_put(folio);
3463 return true;
3464 }
3465
3466 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) {
3467 struct page *page = folio_file_page(folio, start);
3468 vm_fault_t ret = do_set_pmd(vmf, page);
3469 if (!ret) {
3470 /* The page is mapped successfully, reference consumed. */
3471 folio_unlock(folio);
3472 return true;
3473 }
3474 }
3475
3476 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte)
3477 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3478
3479 return false;
3480}
3481
3482static struct folio *next_uptodate_folio(struct xa_state *xas,
3483 struct address_space *mapping, pgoff_t end_pgoff)
3484{
3485 struct folio *folio = xas_next_entry(xas, end_pgoff);
3486 unsigned long max_idx;
3487
3488 do {
3489 if (!folio)
3490 return NULL;
3491 if (xas_retry(xas, folio))
3492 continue;
3493 if (xa_is_value(folio))
3494 continue;
3495 if (!folio_try_get(folio))
3496 continue;
3497 if (folio_test_locked(folio))
3498 goto skip;
3499 /* Has the page moved or been split? */
3500 if (unlikely(folio != xas_reload(xas)))
3501 goto skip;
3502 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3503 goto skip;
3504 if (!folio_trylock(folio))
3505 goto skip;
3506 if (folio->mapping != mapping)
3507 goto unlock;
3508 if (!folio_test_uptodate(folio))
3509 goto unlock;
3510 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3511 if (xas->xa_index >= max_idx)
3512 goto unlock;
3513 return folio;
3514unlock:
3515 folio_unlock(folio);
3516skip:
3517 folio_put(folio);
3518 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3519
3520 return NULL;
3521}
3522
3523/*
3524 * Map page range [start_page, start_page + nr_pages) of folio.
3525 * start_page is gotten from start by folio_page(folio, start)
3526 */
3527static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
3528 struct folio *folio, unsigned long start,
3529 unsigned long addr, unsigned int nr_pages,
3530 unsigned long *rss, unsigned int *mmap_miss)
3531{
3532 vm_fault_t ret = 0;
3533 struct page *page = folio_page(folio, start);
3534 unsigned int count = 0;
3535 pte_t *old_ptep = vmf->pte;
3536
3537 do {
3538 if (PageHWPoison(page + count))
3539 goto skip;
3540
3541 /*
3542 * If there are too many folios that are recently evicted
3543 * in a file, they will probably continue to be evicted.
3544 * In such situation, read-ahead is only a waste of IO.
3545 * Don't decrease mmap_miss in this scenario to make sure
3546 * we can stop read-ahead.
3547 */
3548 if (!folio_test_workingset(folio))
3549 (*mmap_miss)++;
3550
3551 /*
3552 * NOTE: If there're PTE markers, we'll leave them to be
3553 * handled in the specific fault path, and it'll prohibit the
3554 * fault-around logic.
3555 */
3556 if (!pte_none(ptep_get(&vmf->pte[count])))
3557 goto skip;
3558
3559 count++;
3560 continue;
3561skip:
3562 if (count) {
3563 set_pte_range(vmf, folio, page, count, addr);
3564 *rss += count;
3565 folio_ref_add(folio, count);
3566 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3567 ret = VM_FAULT_NOPAGE;
3568 }
3569
3570 count++;
3571 page += count;
3572 vmf->pte += count;
3573 addr += count * PAGE_SIZE;
3574 count = 0;
3575 } while (--nr_pages > 0);
3576
3577 if (count) {
3578 set_pte_range(vmf, folio, page, count, addr);
3579 *rss += count;
3580 folio_ref_add(folio, count);
3581 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3582 ret = VM_FAULT_NOPAGE;
3583 }
3584
3585 vmf->pte = old_ptep;
3586
3587 return ret;
3588}
3589
3590static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
3591 struct folio *folio, unsigned long addr,
3592 unsigned long *rss, unsigned int *mmap_miss)
3593{
3594 vm_fault_t ret = 0;
3595 struct page *page = &folio->page;
3596
3597 if (PageHWPoison(page))
3598 return ret;
3599
3600 /* See comment of filemap_map_folio_range() */
3601 if (!folio_test_workingset(folio))
3602 (*mmap_miss)++;
3603
3604 /*
3605 * NOTE: If there're PTE markers, we'll leave them to be
3606 * handled in the specific fault path, and it'll prohibit
3607 * the fault-around logic.
3608 */
3609 if (!pte_none(ptep_get(vmf->pte)))
3610 return ret;
3611
3612 if (vmf->address == addr)
3613 ret = VM_FAULT_NOPAGE;
3614
3615 set_pte_range(vmf, folio, page, 1, addr);
3616 (*rss)++;
3617 folio_ref_inc(folio);
3618
3619 return ret;
3620}
3621
3622vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3623 pgoff_t start_pgoff, pgoff_t end_pgoff)
3624{
3625 struct vm_area_struct *vma = vmf->vma;
3626 struct file *file = vma->vm_file;
3627 struct address_space *mapping = file->f_mapping;
3628 pgoff_t file_end, last_pgoff = start_pgoff;
3629 unsigned long addr;
3630 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3631 struct folio *folio;
3632 vm_fault_t ret = 0;
3633 unsigned long rss = 0;
3634 unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type;
3635
3636 rcu_read_lock();
3637 folio = next_uptodate_folio(&xas, mapping, end_pgoff);
3638 if (!folio)
3639 goto out;
3640
3641 if (filemap_map_pmd(vmf, folio, start_pgoff)) {
3642 ret = VM_FAULT_NOPAGE;
3643 goto out;
3644 }
3645
3646 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3647 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3648 if (!vmf->pte) {
3649 folio_unlock(folio);
3650 folio_put(folio);
3651 goto out;
3652 }
3653
3654 file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
3655 if (end_pgoff > file_end)
3656 end_pgoff = file_end;
3657
3658 folio_type = mm_counter_file(folio);
3659 do {
3660 unsigned long end;
3661
3662 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3663 vmf->pte += xas.xa_index - last_pgoff;
3664 last_pgoff = xas.xa_index;
3665 end = folio_next_index(folio) - 1;
3666 nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
3667
3668 if (!folio_test_large(folio))
3669 ret |= filemap_map_order0_folio(vmf,
3670 folio, addr, &rss, &mmap_miss);
3671 else
3672 ret |= filemap_map_folio_range(vmf, folio,
3673 xas.xa_index - folio->index, addr,
3674 nr_pages, &rss, &mmap_miss);
3675
3676 folio_unlock(folio);
3677 folio_put(folio);
3678 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL);
3679 add_mm_counter(vma->vm_mm, folio_type, rss);
3680 pte_unmap_unlock(vmf->pte, vmf->ptl);
3681 trace_mm_filemap_map_pages(mapping, start_pgoff, end_pgoff);
3682out:
3683 rcu_read_unlock();
3684
3685 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss);
3686 if (mmap_miss >= mmap_miss_saved)
3687 WRITE_ONCE(file->f_ra.mmap_miss, 0);
3688 else
3689 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss);
3690
3691 return ret;
3692}
3693EXPORT_SYMBOL(filemap_map_pages);
3694
3695vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3696{
3697 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3698 struct folio *folio = page_folio(vmf->page);
3699 vm_fault_t ret = VM_FAULT_LOCKED;
3700
3701 sb_start_pagefault(mapping->host->i_sb);
3702 file_update_time(vmf->vma->vm_file);
3703 folio_lock(folio);
3704 if (folio->mapping != mapping) {
3705 folio_unlock(folio);
3706 ret = VM_FAULT_NOPAGE;
3707 goto out;
3708 }
3709 /*
3710 * We mark the folio dirty already here so that when freeze is in
3711 * progress, we are guaranteed that writeback during freezing will
3712 * see the dirty folio and writeprotect it again.
3713 */
3714 folio_mark_dirty(folio);
3715 folio_wait_stable(folio);
3716out:
3717 sb_end_pagefault(mapping->host->i_sb);
3718 return ret;
3719}
3720
3721const struct vm_operations_struct generic_file_vm_ops = {
3722 .fault = filemap_fault,
3723 .map_pages = filemap_map_pages,
3724 .page_mkwrite = filemap_page_mkwrite,
3725};
3726
3727/* This is used for a general mmap of a disk file */
3728
3729int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3730{
3731 struct address_space *mapping = file->f_mapping;
3732
3733 if (!mapping->a_ops->read_folio)
3734 return -ENOEXEC;
3735 file_accessed(file);
3736 vma->vm_ops = &generic_file_vm_ops;
3737 return 0;
3738}
3739
3740/*
3741 * This is for filesystems which do not implement ->writepage.
3742 */
3743int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3744{
3745 if (vma_is_shared_maywrite(vma))
3746 return -EINVAL;
3747 return generic_file_mmap(file, vma);
3748}
3749#else
3750vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3751{
3752 return VM_FAULT_SIGBUS;
3753}
3754int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3755{
3756 return -ENOSYS;
3757}
3758int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3759{
3760 return -ENOSYS;
3761}
3762#endif /* CONFIG_MMU */
3763
3764EXPORT_SYMBOL(filemap_page_mkwrite);
3765EXPORT_SYMBOL(generic_file_mmap);
3766EXPORT_SYMBOL(generic_file_readonly_mmap);
3767
3768static struct folio *do_read_cache_folio(struct address_space *mapping,
3769 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3770{
3771 struct folio *folio;
3772 int err;
3773
3774 if (!filler)
3775 filler = mapping->a_ops->read_folio;
3776repeat:
3777 folio = filemap_get_folio(mapping, index);
3778 if (IS_ERR(folio)) {
3779 folio = filemap_alloc_folio(gfp,
3780 mapping_min_folio_order(mapping));
3781 if (!folio)
3782 return ERR_PTR(-ENOMEM);
3783 index = mapping_align_index(mapping, index);
3784 err = filemap_add_folio(mapping, folio, index, gfp);
3785 if (unlikely(err)) {
3786 folio_put(folio);
3787 if (err == -EEXIST)
3788 goto repeat;
3789 /* Presumably ENOMEM for xarray node */
3790 return ERR_PTR(err);
3791 }
3792
3793 goto filler;
3794 }
3795 if (folio_test_uptodate(folio))
3796 goto out;
3797
3798 if (!folio_trylock(folio)) {
3799 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3800 goto repeat;
3801 }
3802
3803 /* Folio was truncated from mapping */
3804 if (!folio->mapping) {
3805 folio_unlock(folio);
3806 folio_put(folio);
3807 goto repeat;
3808 }
3809
3810 /* Someone else locked and filled the page in a very small window */
3811 if (folio_test_uptodate(folio)) {
3812 folio_unlock(folio);
3813 goto out;
3814 }
3815
3816filler:
3817 err = filemap_read_folio(file, filler, folio);
3818 if (err) {
3819 folio_put(folio);
3820 if (err == AOP_TRUNCATED_PAGE)
3821 goto repeat;
3822 return ERR_PTR(err);
3823 }
3824
3825out:
3826 folio_mark_accessed(folio);
3827 return folio;
3828}
3829
3830/**
3831 * read_cache_folio - Read into page cache, fill it if needed.
3832 * @mapping: The address_space to read from.
3833 * @index: The index to read.
3834 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3835 * @file: Passed to filler function, may be NULL if not required.
3836 *
3837 * Read one page into the page cache. If it succeeds, the folio returned
3838 * will contain @index, but it may not be the first page of the folio.
3839 *
3840 * If the filler function returns an error, it will be returned to the
3841 * caller.
3842 *
3843 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3844 * Return: An uptodate folio on success, ERR_PTR() on failure.
3845 */
3846struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3847 filler_t filler, struct file *file)
3848{
3849 return do_read_cache_folio(mapping, index, filler, file,
3850 mapping_gfp_mask(mapping));
3851}
3852EXPORT_SYMBOL(read_cache_folio);
3853
3854/**
3855 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags.
3856 * @mapping: The address_space for the folio.
3857 * @index: The index that the allocated folio will contain.
3858 * @gfp: The page allocator flags to use if allocating.
3859 *
3860 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with
3861 * any new memory allocations done using the specified allocation flags.
3862 *
3863 * The most likely error from this function is EIO, but ENOMEM is
3864 * possible and so is EINTR. If ->read_folio returns another error,
3865 * that will be returned to the caller.
3866 *
3867 * The function expects mapping->invalidate_lock to be already held.
3868 *
3869 * Return: Uptodate folio on success, ERR_PTR() on failure.
3870 */
3871struct folio *mapping_read_folio_gfp(struct address_space *mapping,
3872 pgoff_t index, gfp_t gfp)
3873{
3874 return do_read_cache_folio(mapping, index, NULL, NULL, gfp);
3875}
3876EXPORT_SYMBOL(mapping_read_folio_gfp);
3877
3878static struct page *do_read_cache_page(struct address_space *mapping,
3879 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3880{
3881 struct folio *folio;
3882
3883 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3884 if (IS_ERR(folio))
3885 return &folio->page;
3886 return folio_file_page(folio, index);
3887}
3888
3889struct page *read_cache_page(struct address_space *mapping,
3890 pgoff_t index, filler_t *filler, struct file *file)
3891{
3892 return do_read_cache_page(mapping, index, filler, file,
3893 mapping_gfp_mask(mapping));
3894}
3895EXPORT_SYMBOL(read_cache_page);
3896
3897/**
3898 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3899 * @mapping: the page's address_space
3900 * @index: the page index
3901 * @gfp: the page allocator flags to use if allocating
3902 *
3903 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3904 * any new page allocations done using the specified allocation flags.
3905 *
3906 * If the page does not get brought uptodate, return -EIO.
3907 *
3908 * The function expects mapping->invalidate_lock to be already held.
3909 *
3910 * Return: up to date page on success, ERR_PTR() on failure.
3911 */
3912struct page *read_cache_page_gfp(struct address_space *mapping,
3913 pgoff_t index,
3914 gfp_t gfp)
3915{
3916 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3917}
3918EXPORT_SYMBOL(read_cache_page_gfp);
3919
3920/*
3921 * Warn about a page cache invalidation failure during a direct I/O write.
3922 */
3923static void dio_warn_stale_pagecache(struct file *filp)
3924{
3925 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
3926 char pathname[128];
3927 char *path;
3928
3929 errseq_set(&filp->f_mapping->wb_err, -EIO);
3930 if (__ratelimit(&_rs)) {
3931 path = file_path(filp, pathname, sizeof(pathname));
3932 if (IS_ERR(path))
3933 path = "(unknown)";
3934 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
3935 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
3936 current->comm);
3937 }
3938}
3939
3940void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count)
3941{
3942 struct address_space *mapping = iocb->ki_filp->f_mapping;
3943
3944 if (mapping->nrpages &&
3945 invalidate_inode_pages2_range(mapping,
3946 iocb->ki_pos >> PAGE_SHIFT,
3947 (iocb->ki_pos + count - 1) >> PAGE_SHIFT))
3948 dio_warn_stale_pagecache(iocb->ki_filp);
3949}
3950
3951ssize_t
3952generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
3953{
3954 struct address_space *mapping = iocb->ki_filp->f_mapping;
3955 size_t write_len = iov_iter_count(from);
3956 ssize_t written;
3957
3958 /*
3959 * If a page can not be invalidated, return 0 to fall back
3960 * to buffered write.
3961 */
3962 written = kiocb_invalidate_pages(iocb, write_len);
3963 if (written) {
3964 if (written == -EBUSY)
3965 return 0;
3966 return written;
3967 }
3968
3969 written = mapping->a_ops->direct_IO(iocb, from);
3970
3971 /*
3972 * Finally, try again to invalidate clean pages which might have been
3973 * cached by non-direct readahead, or faulted in by get_user_pages()
3974 * if the source of the write was an mmap'ed region of the file
3975 * we're writing. Either one is a pretty crazy thing to do,
3976 * so we don't support it 100%. If this invalidation
3977 * fails, tough, the write still worked...
3978 *
3979 * Most of the time we do not need this since dio_complete() will do
3980 * the invalidation for us. However there are some file systems that
3981 * do not end up with dio_complete() being called, so let's not break
3982 * them by removing it completely.
3983 *
3984 * Noticeable example is a blkdev_direct_IO().
3985 *
3986 * Skip invalidation for async writes or if mapping has no pages.
3987 */
3988 if (written > 0) {
3989 struct inode *inode = mapping->host;
3990 loff_t pos = iocb->ki_pos;
3991
3992 kiocb_invalidate_post_direct_write(iocb, written);
3993 pos += written;
3994 write_len -= written;
3995 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
3996 i_size_write(inode, pos);
3997 mark_inode_dirty(inode);
3998 }
3999 iocb->ki_pos = pos;
4000 }
4001 if (written != -EIOCBQUEUED)
4002 iov_iter_revert(from, write_len - iov_iter_count(from));
4003 return written;
4004}
4005EXPORT_SYMBOL(generic_file_direct_write);
4006
4007ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
4008{
4009 struct file *file = iocb->ki_filp;
4010 loff_t pos = iocb->ki_pos;
4011 struct address_space *mapping = file->f_mapping;
4012 const struct address_space_operations *a_ops = mapping->a_ops;
4013 size_t chunk = mapping_max_folio_size(mapping);
4014 long status = 0;
4015 ssize_t written = 0;
4016
4017 do {
4018 struct folio *folio;
4019 size_t offset; /* Offset into folio */
4020 size_t bytes; /* Bytes to write to folio */
4021 size_t copied; /* Bytes copied from user */
4022 void *fsdata = NULL;
4023
4024 bytes = iov_iter_count(i);
4025retry:
4026 offset = pos & (chunk - 1);
4027 bytes = min(chunk - offset, bytes);
4028 balance_dirty_pages_ratelimited(mapping);
4029
4030 /*
4031 * Bring in the user page that we will copy from _first_.
4032 * Otherwise there's a nasty deadlock on copying from the
4033 * same page as we're writing to, without it being marked
4034 * up-to-date.
4035 */
4036 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
4037 status = -EFAULT;
4038 break;
4039 }
4040
4041 if (fatal_signal_pending(current)) {
4042 status = -EINTR;
4043 break;
4044 }
4045
4046 status = a_ops->write_begin(file, mapping, pos, bytes,
4047 &folio, &fsdata);
4048 if (unlikely(status < 0))
4049 break;
4050
4051 offset = offset_in_folio(folio, pos);
4052 if (bytes > folio_size(folio) - offset)
4053 bytes = folio_size(folio) - offset;
4054
4055 if (mapping_writably_mapped(mapping))
4056 flush_dcache_folio(folio);
4057
4058 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
4059 flush_dcache_folio(folio);
4060
4061 status = a_ops->write_end(file, mapping, pos, bytes, copied,
4062 folio, fsdata);
4063 if (unlikely(status != copied)) {
4064 iov_iter_revert(i, copied - max(status, 0L));
4065 if (unlikely(status < 0))
4066 break;
4067 }
4068 cond_resched();
4069
4070 if (unlikely(status == 0)) {
4071 /*
4072 * A short copy made ->write_end() reject the
4073 * thing entirely. Might be memory poisoning
4074 * halfway through, might be a race with munmap,
4075 * might be severe memory pressure.
4076 */
4077 if (chunk > PAGE_SIZE)
4078 chunk /= 2;
4079 if (copied) {
4080 bytes = copied;
4081 goto retry;
4082 }
4083 } else {
4084 pos += status;
4085 written += status;
4086 }
4087 } while (iov_iter_count(i));
4088
4089 if (!written)
4090 return status;
4091 iocb->ki_pos += written;
4092 return written;
4093}
4094EXPORT_SYMBOL(generic_perform_write);
4095
4096/**
4097 * __generic_file_write_iter - write data to a file
4098 * @iocb: IO state structure (file, offset, etc.)
4099 * @from: iov_iter with data to write
4100 *
4101 * This function does all the work needed for actually writing data to a
4102 * file. It does all basic checks, removes SUID from the file, updates
4103 * modification times and calls proper subroutines depending on whether we
4104 * do direct IO or a standard buffered write.
4105 *
4106 * It expects i_rwsem to be grabbed unless we work on a block device or similar
4107 * object which does not need locking at all.
4108 *
4109 * This function does *not* take care of syncing data in case of O_SYNC write.
4110 * A caller has to handle it. This is mainly due to the fact that we want to
4111 * avoid syncing under i_rwsem.
4112 *
4113 * Return:
4114 * * number of bytes written, even for truncated writes
4115 * * negative error code if no data has been written at all
4116 */
4117ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4118{
4119 struct file *file = iocb->ki_filp;
4120 struct address_space *mapping = file->f_mapping;
4121 struct inode *inode = mapping->host;
4122 ssize_t ret;
4123
4124 ret = file_remove_privs(file);
4125 if (ret)
4126 return ret;
4127
4128 ret = file_update_time(file);
4129 if (ret)
4130 return ret;
4131
4132 if (iocb->ki_flags & IOCB_DIRECT) {
4133 ret = generic_file_direct_write(iocb, from);
4134 /*
4135 * If the write stopped short of completing, fall back to
4136 * buffered writes. Some filesystems do this for writes to
4137 * holes, for example. For DAX files, a buffered write will
4138 * not succeed (even if it did, DAX does not handle dirty
4139 * page-cache pages correctly).
4140 */
4141 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode))
4142 return ret;
4143 return direct_write_fallback(iocb, from, ret,
4144 generic_perform_write(iocb, from));
4145 }
4146
4147 return generic_perform_write(iocb, from);
4148}
4149EXPORT_SYMBOL(__generic_file_write_iter);
4150
4151/**
4152 * generic_file_write_iter - write data to a file
4153 * @iocb: IO state structure
4154 * @from: iov_iter with data to write
4155 *
4156 * This is a wrapper around __generic_file_write_iter() to be used by most
4157 * filesystems. It takes care of syncing the file in case of O_SYNC file
4158 * and acquires i_rwsem as needed.
4159 * Return:
4160 * * negative error code if no data has been written at all of
4161 * vfs_fsync_range() failed for a synchronous write
4162 * * number of bytes written, even for truncated writes
4163 */
4164ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4165{
4166 struct file *file = iocb->ki_filp;
4167 struct inode *inode = file->f_mapping->host;
4168 ssize_t ret;
4169
4170 inode_lock(inode);
4171 ret = generic_write_checks(iocb, from);
4172 if (ret > 0)
4173 ret = __generic_file_write_iter(iocb, from);
4174 inode_unlock(inode);
4175
4176 if (ret > 0)
4177 ret = generic_write_sync(iocb, ret);
4178 return ret;
4179}
4180EXPORT_SYMBOL(generic_file_write_iter);
4181
4182/**
4183 * filemap_release_folio() - Release fs-specific metadata on a folio.
4184 * @folio: The folio which the kernel is trying to free.
4185 * @gfp: Memory allocation flags (and I/O mode).
4186 *
4187 * The address_space is trying to release any data attached to a folio
4188 * (presumably at folio->private).
4189 *
4190 * This will also be called if the private_2 flag is set on a page,
4191 * indicating that the folio has other metadata associated with it.
4192 *
4193 * The @gfp argument specifies whether I/O may be performed to release
4194 * this page (__GFP_IO), and whether the call may block
4195 * (__GFP_RECLAIM & __GFP_FS).
4196 *
4197 * Return: %true if the release was successful, otherwise %false.
4198 */
4199bool filemap_release_folio(struct folio *folio, gfp_t gfp)
4200{
4201 struct address_space * const mapping = folio->mapping;
4202
4203 BUG_ON(!folio_test_locked(folio));
4204 if (!folio_needs_release(folio))
4205 return true;
4206 if (folio_test_writeback(folio))
4207 return false;
4208
4209 if (mapping && mapping->a_ops->release_folio)
4210 return mapping->a_ops->release_folio(folio, gfp);
4211 return try_to_free_buffers(folio);
4212}
4213EXPORT_SYMBOL(filemap_release_folio);
4214
4215/**
4216 * filemap_invalidate_inode - Invalidate/forcibly write back a range of an inode's pagecache
4217 * @inode: The inode to flush
4218 * @flush: Set to write back rather than simply invalidate.
4219 * @start: First byte to in range.
4220 * @end: Last byte in range (inclusive), or LLONG_MAX for everything from start
4221 * onwards.
4222 *
4223 * Invalidate all the folios on an inode that contribute to the specified
4224 * range, possibly writing them back first. Whilst the operation is
4225 * undertaken, the invalidate lock is held to prevent new folios from being
4226 * installed.
4227 */
4228int filemap_invalidate_inode(struct inode *inode, bool flush,
4229 loff_t start, loff_t end)
4230{
4231 struct address_space *mapping = inode->i_mapping;
4232 pgoff_t first = start >> PAGE_SHIFT;
4233 pgoff_t last = end >> PAGE_SHIFT;
4234 pgoff_t nr = end == LLONG_MAX ? ULONG_MAX : last - first + 1;
4235
4236 if (!mapping || !mapping->nrpages || end < start)
4237 goto out;
4238
4239 /* Prevent new folios from being added to the inode. */
4240 filemap_invalidate_lock(mapping);
4241
4242 if (!mapping->nrpages)
4243 goto unlock;
4244
4245 unmap_mapping_pages(mapping, first, nr, false);
4246
4247 /* Write back the data if we're asked to. */
4248 if (flush) {
4249 struct writeback_control wbc = {
4250 .sync_mode = WB_SYNC_ALL,
4251 .nr_to_write = LONG_MAX,
4252 .range_start = start,
4253 .range_end = end,
4254 };
4255
4256 filemap_fdatawrite_wbc(mapping, &wbc);
4257 }
4258
4259 /* Wait for writeback to complete on all folios and discard. */
4260 invalidate_inode_pages2_range(mapping, start / PAGE_SIZE, end / PAGE_SIZE);
4261
4262unlock:
4263 filemap_invalidate_unlock(mapping);
4264out:
4265 return filemap_check_errors(mapping);
4266}
4267EXPORT_SYMBOL_GPL(filemap_invalidate_inode);
4268
4269#ifdef CONFIG_CACHESTAT_SYSCALL
4270/**
4271 * filemap_cachestat() - compute the page cache statistics of a mapping
4272 * @mapping: The mapping to compute the statistics for.
4273 * @first_index: The starting page cache index.
4274 * @last_index: The final page index (inclusive).
4275 * @cs: the cachestat struct to write the result to.
4276 *
4277 * This will query the page cache statistics of a mapping in the
4278 * page range of [first_index, last_index] (inclusive). The statistics
4279 * queried include: number of dirty pages, number of pages marked for
4280 * writeback, and the number of (recently) evicted pages.
4281 */
4282static void filemap_cachestat(struct address_space *mapping,
4283 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs)
4284{
4285 XA_STATE(xas, &mapping->i_pages, first_index);
4286 struct folio *folio;
4287
4288 /* Flush stats (and potentially sleep) outside the RCU read section. */
4289 mem_cgroup_flush_stats_ratelimited(NULL);
4290
4291 rcu_read_lock();
4292 xas_for_each(&xas, folio, last_index) {
4293 int order;
4294 unsigned long nr_pages;
4295 pgoff_t folio_first_index, folio_last_index;
4296
4297 /*
4298 * Don't deref the folio. It is not pinned, and might
4299 * get freed (and reused) underneath us.
4300 *
4301 * We *could* pin it, but that would be expensive for
4302 * what should be a fast and lightweight syscall.
4303 *
4304 * Instead, derive all information of interest from
4305 * the rcu-protected xarray.
4306 */
4307
4308 if (xas_retry(&xas, folio))
4309 continue;
4310
4311 order = xas_get_order(&xas);
4312 nr_pages = 1 << order;
4313 folio_first_index = round_down(xas.xa_index, 1 << order);
4314 folio_last_index = folio_first_index + nr_pages - 1;
4315
4316 /* Folios might straddle the range boundaries, only count covered pages */
4317 if (folio_first_index < first_index)
4318 nr_pages -= first_index - folio_first_index;
4319
4320 if (folio_last_index > last_index)
4321 nr_pages -= folio_last_index - last_index;
4322
4323 if (xa_is_value(folio)) {
4324 /* page is evicted */
4325 void *shadow = (void *)folio;
4326 bool workingset; /* not used */
4327
4328 cs->nr_evicted += nr_pages;
4329
4330#ifdef CONFIG_SWAP /* implies CONFIG_MMU */
4331 if (shmem_mapping(mapping)) {
4332 /* shmem file - in swap cache */
4333 swp_entry_t swp = radix_to_swp_entry(folio);
4334
4335 /* swapin error results in poisoned entry */
4336 if (non_swap_entry(swp))
4337 goto resched;
4338
4339 /*
4340 * Getting a swap entry from the shmem
4341 * inode means we beat
4342 * shmem_unuse(). rcu_read_lock()
4343 * ensures swapoff waits for us before
4344 * freeing the swapper space. However,
4345 * we can race with swapping and
4346 * invalidation, so there might not be
4347 * a shadow in the swapcache (yet).
4348 */
4349 shadow = get_shadow_from_swap_cache(swp);
4350 if (!shadow)
4351 goto resched;
4352 }
4353#endif
4354 if (workingset_test_recent(shadow, true, &workingset, false))
4355 cs->nr_recently_evicted += nr_pages;
4356
4357 goto resched;
4358 }
4359
4360 /* page is in cache */
4361 cs->nr_cache += nr_pages;
4362
4363 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY))
4364 cs->nr_dirty += nr_pages;
4365
4366 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK))
4367 cs->nr_writeback += nr_pages;
4368
4369resched:
4370 if (need_resched()) {
4371 xas_pause(&xas);
4372 cond_resched_rcu();
4373 }
4374 }
4375 rcu_read_unlock();
4376}
4377
4378/*
4379 * See mincore: reveal pagecache information only for files
4380 * that the calling process has write access to, or could (if
4381 * tried) open for writing.
4382 */
4383static inline bool can_do_cachestat(struct file *f)
4384{
4385 if (f->f_mode & FMODE_WRITE)
4386 return true;
4387 if (inode_owner_or_capable(file_mnt_idmap(f), file_inode(f)))
4388 return true;
4389 return file_permission(f, MAY_WRITE) == 0;
4390}
4391
4392/*
4393 * The cachestat(2) system call.
4394 *
4395 * cachestat() returns the page cache statistics of a file in the
4396 * bytes range specified by `off` and `len`: number of cached pages,
4397 * number of dirty pages, number of pages marked for writeback,
4398 * number of evicted pages, and number of recently evicted pages.
4399 *
4400 * An evicted page is a page that is previously in the page cache
4401 * but has been evicted since. A page is recently evicted if its last
4402 * eviction was recent enough that its reentry to the cache would
4403 * indicate that it is actively being used by the system, and that
4404 * there is memory pressure on the system.
4405 *
4406 * `off` and `len` must be non-negative integers. If `len` > 0,
4407 * the queried range is [`off`, `off` + `len`]. If `len` == 0,
4408 * we will query in the range from `off` to the end of the file.
4409 *
4410 * The `flags` argument is unused for now, but is included for future
4411 * extensibility. User should pass 0 (i.e no flag specified).
4412 *
4413 * Currently, hugetlbfs is not supported.
4414 *
4415 * Because the status of a page can change after cachestat() checks it
4416 * but before it returns to the application, the returned values may
4417 * contain stale information.
4418 *
4419 * return values:
4420 * zero - success
4421 * -EFAULT - cstat or cstat_range points to an illegal address
4422 * -EINVAL - invalid flags
4423 * -EBADF - invalid file descriptor
4424 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file
4425 */
4426SYSCALL_DEFINE4(cachestat, unsigned int, fd,
4427 struct cachestat_range __user *, cstat_range,
4428 struct cachestat __user *, cstat, unsigned int, flags)
4429{
4430 CLASS(fd, f)(fd);
4431 struct address_space *mapping;
4432 struct cachestat_range csr;
4433 struct cachestat cs;
4434 pgoff_t first_index, last_index;
4435
4436 if (fd_empty(f))
4437 return -EBADF;
4438
4439 if (copy_from_user(&csr, cstat_range,
4440 sizeof(struct cachestat_range)))
4441 return -EFAULT;
4442
4443 /* hugetlbfs is not supported */
4444 if (is_file_hugepages(fd_file(f)))
4445 return -EOPNOTSUPP;
4446
4447 if (!can_do_cachestat(fd_file(f)))
4448 return -EPERM;
4449
4450 if (flags != 0)
4451 return -EINVAL;
4452
4453 first_index = csr.off >> PAGE_SHIFT;
4454 last_index =
4455 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT;
4456 memset(&cs, 0, sizeof(struct cachestat));
4457 mapping = fd_file(f)->f_mapping;
4458 filemap_cachestat(mapping, first_index, last_index, &cs);
4459
4460 if (copy_to_user(cstat, &cs, sizeof(struct cachestat)))
4461 return -EFAULT;
4462
4463 return 0;
4464}
4465#endif /* CONFIG_CACHESTAT_SYSCALL */