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 <linux/sysctl.h>
51 #include <asm/pgalloc.h>
52 #include <asm/tlbflush.h>
53 #include "internal.h"
54
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/filemap.h>
57
58 /*
59 * FIXME: remove all knowledge of the buffer layer from the core VM
60 */
61 #include <linux/buffer_head.h> /* for try_to_free_buffers */
62
63 #include <asm/mman.h>
64
65 #include "swap.h"
66
67 /*
68 * Shared mappings implemented 30.11.1994. It's not fully working yet,
69 * though.
70 *
71 * Shared mappings now work. 15.8.1995 Bruno.
72 *
73 * finished 'unifying' the page and buffer cache and SMP-threaded the
74 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
75 *
76 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
77 */
78
79 /*
80 * Lock ordering:
81 *
82 * ->i_mmap_rwsem (truncate_pagecache)
83 * ->private_lock (__free_pte->block_dirty_folio)
84 * ->swap_lock (exclusive_swap_page, others)
85 * ->i_pages lock
86 *
87 * ->i_rwsem
88 * ->invalidate_lock (acquired by fs in truncate path)
89 * ->i_mmap_rwsem (truncate->unmap_mapping_range)
90 *
91 * ->mmap_lock
92 * ->i_mmap_rwsem
93 * ->page_table_lock or pte_lock (various, mainly in memory.c)
94 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock)
95 *
96 * ->mmap_lock
97 * ->invalidate_lock (filemap_fault)
98 * ->lock_page (filemap_fault, access_process_vm)
99 *
100 * ->i_rwsem (generic_perform_write)
101 * ->mmap_lock (fault_in_readable->do_page_fault)
102 *
103 * bdi->wb.list_lock
104 * sb_lock (fs/fs-writeback.c)
105 * ->i_pages lock (__sync_single_inode)
106 *
107 * ->i_mmap_rwsem
108 * ->anon_vma.lock (vma_merge)
109 *
110 * ->anon_vma.lock
111 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
112 *
113 * ->page_table_lock or pte_lock
114 * ->swap_lock (try_to_unmap_one)
115 * ->private_lock (try_to_unmap_one)
116 * ->i_pages lock (try_to_unmap_one)
117 * ->lruvec->lru_lock (follow_page_mask->mark_page_accessed)
118 * ->lruvec->lru_lock (check_pte_range->folio_isolate_lru)
119 * ->private_lock (folio_remove_rmap_pte->set_page_dirty)
120 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty)
121 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty)
122 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty)
123 * bdi.wb->list_lock (zap_pte_range->set_page_dirty)
124 * ->inode->i_lock (zap_pte_range->set_page_dirty)
125 * ->private_lock (zap_pte_range->block_dirty_folio)
126 */
127
page_cache_delete(struct address_space * mapping,struct folio * folio,void * shadow)128 static void page_cache_delete(struct address_space *mapping,
129 struct folio *folio, void *shadow)
130 {
131 XA_STATE(xas, &mapping->i_pages, folio->index);
132 long nr = 1;
133
134 mapping_set_update(&xas, mapping);
135
136 xas_set_order(&xas, folio->index, folio_order(folio));
137 nr = folio_nr_pages(folio);
138
139 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
140
141 xas_store(&xas, shadow);
142 xas_init_marks(&xas);
143
144 folio->mapping = NULL;
145 /* Leave page->index set: truncation lookup relies upon it */
146 mapping->nrpages -= nr;
147 }
148
filemap_unaccount_folio(struct address_space * mapping,struct folio * folio)149 static void filemap_unaccount_folio(struct address_space *mapping,
150 struct folio *folio)
151 {
152 long nr;
153
154 VM_BUG_ON_FOLIO(folio_mapped(folio), folio);
155 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) {
156 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n",
157 current->comm, folio_pfn(folio));
158 dump_page(&folio->page, "still mapped when deleted");
159 dump_stack();
160 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
161
162 if (mapping_exiting(mapping) && !folio_test_large(folio)) {
163 int mapcount = folio_mapcount(folio);
164
165 if (folio_ref_count(folio) >= mapcount + 2) {
166 /*
167 * All vmas have already been torn down, so it's
168 * a good bet that actually the page is unmapped
169 * and we'd rather not leak it: if we're wrong,
170 * another bad page check should catch it later.
171 */
172 atomic_set(&folio->_mapcount, -1);
173 folio_ref_sub(folio, mapcount);
174 }
175 }
176 }
177
178 /* hugetlb folios do not participate in page cache accounting. */
179 if (folio_test_hugetlb(folio))
180 return;
181
182 nr = folio_nr_pages(folio);
183
184 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
185 if (folio_test_swapbacked(folio)) {
186 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
187 if (folio_test_pmd_mappable(folio))
188 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr);
189 } else if (folio_test_pmd_mappable(folio)) {
190 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr);
191 filemap_nr_thps_dec(mapping);
192 }
193
194 /*
195 * At this point folio must be either written or cleaned by
196 * truncate. Dirty folio here signals a bug and loss of
197 * unwritten data - on ordinary filesystems.
198 *
199 * But it's harmless on in-memory filesystems like tmpfs; and can
200 * occur when a driver which did get_user_pages() sets page dirty
201 * before putting it, while the inode is being finally evicted.
202 *
203 * Below fixes dirty accounting after removing the folio entirely
204 * but leaves the dirty flag set: it has no effect for truncated
205 * folio and anyway will be cleared before returning folio to
206 * buddy allocator.
207 */
208 if (WARN_ON_ONCE(folio_test_dirty(folio) &&
209 mapping_can_writeback(mapping)))
210 folio_account_cleaned(folio, inode_to_wb(mapping->host));
211 }
212
213 /*
214 * Delete a page from the page cache and free it. Caller has to make
215 * sure the page is locked and that nobody else uses it - or that usage
216 * is safe. The caller must hold the i_pages lock.
217 */
__filemap_remove_folio(struct folio * folio,void * shadow)218 void __filemap_remove_folio(struct folio *folio, void *shadow)
219 {
220 struct address_space *mapping = folio->mapping;
221
222 trace_mm_filemap_delete_from_page_cache(folio);
223 filemap_unaccount_folio(mapping, folio);
224 page_cache_delete(mapping, folio, shadow);
225 }
226
filemap_free_folio(struct address_space * mapping,struct folio * folio)227 void filemap_free_folio(struct address_space *mapping, struct folio *folio)
228 {
229 void (*free_folio)(struct folio *);
230 int refs = 1;
231
232 free_folio = mapping->a_ops->free_folio;
233 if (free_folio)
234 free_folio(folio);
235
236 if (folio_test_large(folio))
237 refs = folio_nr_pages(folio);
238 folio_put_refs(folio, refs);
239 }
240
241 /**
242 * filemap_remove_folio - Remove folio from page cache.
243 * @folio: The folio.
244 *
245 * This must be called only on folios that are locked and have been
246 * verified to be in the page cache. It will never put the folio into
247 * the free list because the caller has a reference on the page.
248 */
filemap_remove_folio(struct folio * folio)249 void filemap_remove_folio(struct folio *folio)
250 {
251 struct address_space *mapping = folio->mapping;
252
253 BUG_ON(!folio_test_locked(folio));
254 spin_lock(&mapping->host->i_lock);
255 xa_lock_irq(&mapping->i_pages);
256 __filemap_remove_folio(folio, NULL);
257 xa_unlock_irq(&mapping->i_pages);
258 if (mapping_shrinkable(mapping))
259 inode_add_lru(mapping->host);
260 spin_unlock(&mapping->host->i_lock);
261
262 filemap_free_folio(mapping, folio);
263 }
264
265 /*
266 * page_cache_delete_batch - delete several folios from page cache
267 * @mapping: the mapping to which folios belong
268 * @fbatch: batch of folios to delete
269 *
270 * The function walks over mapping->i_pages and removes folios passed in
271 * @fbatch from the mapping. The function expects @fbatch to be sorted
272 * by page index and is optimised for it to be dense.
273 * It tolerates holes in @fbatch (mapping entries at those indices are not
274 * modified).
275 *
276 * The function expects the i_pages lock to be held.
277 */
page_cache_delete_batch(struct address_space * mapping,struct folio_batch * fbatch)278 static void page_cache_delete_batch(struct address_space *mapping,
279 struct folio_batch *fbatch)
280 {
281 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index);
282 long total_pages = 0;
283 int i = 0;
284 struct folio *folio;
285
286 mapping_set_update(&xas, mapping);
287 xas_for_each(&xas, folio, ULONG_MAX) {
288 if (i >= folio_batch_count(fbatch))
289 break;
290
291 /* A swap/dax/shadow entry got inserted? Skip it. */
292 if (xa_is_value(folio))
293 continue;
294 /*
295 * A page got inserted in our range? Skip it. We have our
296 * pages locked so they are protected from being removed.
297 * If we see a page whose index is higher than ours, it
298 * means our page has been removed, which shouldn't be
299 * possible because we're holding the PageLock.
300 */
301 if (folio != fbatch->folios[i]) {
302 VM_BUG_ON_FOLIO(folio->index >
303 fbatch->folios[i]->index, folio);
304 continue;
305 }
306
307 WARN_ON_ONCE(!folio_test_locked(folio));
308
309 folio->mapping = NULL;
310 /* Leave folio->index set: truncation lookup relies on it */
311
312 i++;
313 xas_store(&xas, NULL);
314 total_pages += folio_nr_pages(folio);
315 }
316 mapping->nrpages -= total_pages;
317 }
318
delete_from_page_cache_batch(struct address_space * mapping,struct folio_batch * fbatch)319 void delete_from_page_cache_batch(struct address_space *mapping,
320 struct folio_batch *fbatch)
321 {
322 int i;
323
324 if (!folio_batch_count(fbatch))
325 return;
326
327 spin_lock(&mapping->host->i_lock);
328 xa_lock_irq(&mapping->i_pages);
329 for (i = 0; i < folio_batch_count(fbatch); i++) {
330 struct folio *folio = fbatch->folios[i];
331
332 trace_mm_filemap_delete_from_page_cache(folio);
333 filemap_unaccount_folio(mapping, folio);
334 }
335 page_cache_delete_batch(mapping, fbatch);
336 xa_unlock_irq(&mapping->i_pages);
337 if (mapping_shrinkable(mapping))
338 inode_add_lru(mapping->host);
339 spin_unlock(&mapping->host->i_lock);
340
341 for (i = 0; i < folio_batch_count(fbatch); i++)
342 filemap_free_folio(mapping, fbatch->folios[i]);
343 }
344
filemap_check_errors(struct address_space * mapping)345 int filemap_check_errors(struct address_space *mapping)
346 {
347 int ret = 0;
348 /* Check for outstanding write errors */
349 if (test_bit(AS_ENOSPC, &mapping->flags) &&
350 test_and_clear_bit(AS_ENOSPC, &mapping->flags))
351 ret = -ENOSPC;
352 if (test_bit(AS_EIO, &mapping->flags) &&
353 test_and_clear_bit(AS_EIO, &mapping->flags))
354 ret = -EIO;
355 return ret;
356 }
357 EXPORT_SYMBOL(filemap_check_errors);
358
filemap_check_and_keep_errors(struct address_space * mapping)359 static int filemap_check_and_keep_errors(struct address_space *mapping)
360 {
361 /* Check for outstanding write errors */
362 if (test_bit(AS_EIO, &mapping->flags))
363 return -EIO;
364 if (test_bit(AS_ENOSPC, &mapping->flags))
365 return -ENOSPC;
366 return 0;
367 }
368
369 /**
370 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range
371 * @mapping: address space structure to write
372 * @wbc: the writeback_control controlling the writeout
373 *
374 * Call writepages on the mapping using the provided wbc to control the
375 * writeout.
376 *
377 * Return: %0 on success, negative error code otherwise.
378 */
filemap_fdatawrite_wbc(struct address_space * mapping,struct writeback_control * wbc)379 int filemap_fdatawrite_wbc(struct address_space *mapping,
380 struct writeback_control *wbc)
381 {
382 int ret;
383
384 if (!mapping_can_writeback(mapping) ||
385 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
386 return 0;
387
388 wbc_attach_fdatawrite_inode(wbc, mapping->host);
389 ret = do_writepages(mapping, wbc);
390 wbc_detach_inode(wbc);
391 return ret;
392 }
393 EXPORT_SYMBOL(filemap_fdatawrite_wbc);
394
395 /**
396 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
397 * @mapping: address space structure to write
398 * @start: offset in bytes where the range starts
399 * @end: offset in bytes where the range ends (inclusive)
400 * @sync_mode: enable synchronous operation
401 *
402 * Start writeback against all of a mapping's dirty pages that lie
403 * within the byte offsets <start, end> inclusive.
404 *
405 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
406 * opposed to a regular memory cleansing writeback. The difference between
407 * these two operations is that if a dirty page/buffer is encountered, it must
408 * be waited upon, and not just skipped over.
409 *
410 * Return: %0 on success, negative error code otherwise.
411 */
__filemap_fdatawrite_range(struct address_space * mapping,loff_t start,loff_t end,int sync_mode)412 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
413 loff_t end, int sync_mode)
414 {
415 struct writeback_control wbc = {
416 .sync_mode = sync_mode,
417 .nr_to_write = LONG_MAX,
418 .range_start = start,
419 .range_end = end,
420 };
421
422 return filemap_fdatawrite_wbc(mapping, &wbc);
423 }
424
__filemap_fdatawrite(struct address_space * mapping,int sync_mode)425 static inline int __filemap_fdatawrite(struct address_space *mapping,
426 int sync_mode)
427 {
428 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
429 }
430
filemap_fdatawrite(struct address_space * mapping)431 int filemap_fdatawrite(struct address_space *mapping)
432 {
433 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
434 }
435 EXPORT_SYMBOL(filemap_fdatawrite);
436
filemap_fdatawrite_range(struct address_space * mapping,loff_t start,loff_t end)437 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
438 loff_t end)
439 {
440 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
441 }
442 EXPORT_SYMBOL(filemap_fdatawrite_range);
443
444 /**
445 * filemap_fdatawrite_range_kick - start writeback on a range
446 * @mapping: target address_space
447 * @start: index to start writeback on
448 * @end: last (inclusive) index for writeback
449 *
450 * This is a non-integrity writeback helper, to start writing back folios
451 * for the indicated range.
452 *
453 * Return: %0 on success, negative error code otherwise.
454 */
filemap_fdatawrite_range_kick(struct address_space * mapping,loff_t start,loff_t end)455 int filemap_fdatawrite_range_kick(struct address_space *mapping, loff_t start,
456 loff_t end)
457 {
458 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_NONE);
459 }
460 EXPORT_SYMBOL_GPL(filemap_fdatawrite_range_kick);
461
462 /**
463 * filemap_flush - mostly a non-blocking flush
464 * @mapping: target address_space
465 *
466 * This is a mostly non-blocking flush. Not suitable for data-integrity
467 * purposes - I/O may not be started against all dirty pages.
468 *
469 * Return: %0 on success, negative error code otherwise.
470 */
filemap_flush(struct address_space * mapping)471 int filemap_flush(struct address_space *mapping)
472 {
473 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
474 }
475 EXPORT_SYMBOL(filemap_flush);
476
477 /**
478 * filemap_range_has_page - check if a page exists in range.
479 * @mapping: address space within which to check
480 * @start_byte: offset in bytes where the range starts
481 * @end_byte: offset in bytes where the range ends (inclusive)
482 *
483 * Find at least one page in the range supplied, usually used to check if
484 * direct writing in this range will trigger a writeback.
485 *
486 * Return: %true if at least one page exists in the specified range,
487 * %false otherwise.
488 */
filemap_range_has_page(struct address_space * mapping,loff_t start_byte,loff_t end_byte)489 bool filemap_range_has_page(struct address_space *mapping,
490 loff_t start_byte, loff_t end_byte)
491 {
492 struct folio *folio;
493 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
494 pgoff_t max = end_byte >> PAGE_SHIFT;
495
496 if (end_byte < start_byte)
497 return false;
498
499 rcu_read_lock();
500 for (;;) {
501 folio = xas_find(&xas, max);
502 if (xas_retry(&xas, folio))
503 continue;
504 /* Shadow entries don't count */
505 if (xa_is_value(folio))
506 continue;
507 /*
508 * We don't need to try to pin this page; we're about to
509 * release the RCU lock anyway. It is enough to know that
510 * there was a page here recently.
511 */
512 break;
513 }
514 rcu_read_unlock();
515
516 return folio != NULL;
517 }
518 EXPORT_SYMBOL(filemap_range_has_page);
519
__filemap_fdatawait_range(struct address_space * mapping,loff_t start_byte,loff_t end_byte)520 static void __filemap_fdatawait_range(struct address_space *mapping,
521 loff_t start_byte, loff_t end_byte)
522 {
523 pgoff_t index = start_byte >> PAGE_SHIFT;
524 pgoff_t end = end_byte >> PAGE_SHIFT;
525 struct folio_batch fbatch;
526 unsigned nr_folios;
527
528 folio_batch_init(&fbatch);
529
530 while (index <= end) {
531 unsigned i;
532
533 nr_folios = filemap_get_folios_tag(mapping, &index, end,
534 PAGECACHE_TAG_WRITEBACK, &fbatch);
535
536 if (!nr_folios)
537 break;
538
539 for (i = 0; i < nr_folios; i++) {
540 struct folio *folio = fbatch.folios[i];
541
542 folio_wait_writeback(folio);
543 }
544 folio_batch_release(&fbatch);
545 cond_resched();
546 }
547 }
548
549 /**
550 * filemap_fdatawait_range - wait for writeback to complete
551 * @mapping: address space structure to wait for
552 * @start_byte: offset in bytes where the range starts
553 * @end_byte: offset in bytes where the range ends (inclusive)
554 *
555 * Walk the list of under-writeback pages of the given address space
556 * in the given range and wait for all of them. Check error status of
557 * the address space and return it.
558 *
559 * Since the error status of the address space is cleared by this function,
560 * callers are responsible for checking the return value and handling and/or
561 * reporting the error.
562 *
563 * Return: error status of the address space.
564 */
filemap_fdatawait_range(struct address_space * mapping,loff_t start_byte,loff_t end_byte)565 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
566 loff_t end_byte)
567 {
568 __filemap_fdatawait_range(mapping, start_byte, end_byte);
569 return filemap_check_errors(mapping);
570 }
571 EXPORT_SYMBOL(filemap_fdatawait_range);
572
573 /**
574 * filemap_fdatawait_range_keep_errors - wait for writeback to complete
575 * @mapping: address space structure to wait for
576 * @start_byte: offset in bytes where the range starts
577 * @end_byte: offset in bytes where the range ends (inclusive)
578 *
579 * Walk the list of under-writeback pages of the given address space in the
580 * given range and wait for all of them. Unlike filemap_fdatawait_range(),
581 * this function does not clear error status of the address space.
582 *
583 * Use this function if callers don't handle errors themselves. Expected
584 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
585 * fsfreeze(8)
586 */
filemap_fdatawait_range_keep_errors(struct address_space * mapping,loff_t start_byte,loff_t end_byte)587 int filemap_fdatawait_range_keep_errors(struct address_space *mapping,
588 loff_t start_byte, loff_t end_byte)
589 {
590 __filemap_fdatawait_range(mapping, start_byte, end_byte);
591 return filemap_check_and_keep_errors(mapping);
592 }
593 EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors);
594
595 /**
596 * file_fdatawait_range - wait for writeback to complete
597 * @file: file pointing to address space structure to wait for
598 * @start_byte: offset in bytes where the range starts
599 * @end_byte: offset in bytes where the range ends (inclusive)
600 *
601 * Walk the list of under-writeback pages of the address space that file
602 * refers to, in the given range and wait for all of them. Check error
603 * status of the address space vs. the file->f_wb_err cursor and return it.
604 *
605 * Since the error status of the file is advanced by this function,
606 * callers are responsible for checking the return value and handling and/or
607 * reporting the error.
608 *
609 * Return: error status of the address space vs. the file->f_wb_err cursor.
610 */
file_fdatawait_range(struct file * file,loff_t start_byte,loff_t end_byte)611 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte)
612 {
613 struct address_space *mapping = file->f_mapping;
614
615 __filemap_fdatawait_range(mapping, start_byte, end_byte);
616 return file_check_and_advance_wb_err(file);
617 }
618 EXPORT_SYMBOL(file_fdatawait_range);
619
620 /**
621 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
622 * @mapping: address space structure to wait for
623 *
624 * Walk the list of under-writeback pages of the given address space
625 * and wait for all of them. Unlike filemap_fdatawait(), this function
626 * does not clear error status of the address space.
627 *
628 * Use this function if callers don't handle errors themselves. Expected
629 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
630 * fsfreeze(8)
631 *
632 * Return: error status of the address space.
633 */
filemap_fdatawait_keep_errors(struct address_space * mapping)634 int filemap_fdatawait_keep_errors(struct address_space *mapping)
635 {
636 __filemap_fdatawait_range(mapping, 0, LLONG_MAX);
637 return filemap_check_and_keep_errors(mapping);
638 }
639 EXPORT_SYMBOL(filemap_fdatawait_keep_errors);
640
641 /* Returns true if writeback might be needed or already in progress. */
mapping_needs_writeback(struct address_space * mapping)642 static bool mapping_needs_writeback(struct address_space *mapping)
643 {
644 return mapping->nrpages;
645 }
646
filemap_range_has_writeback(struct address_space * mapping,loff_t start_byte,loff_t end_byte)647 bool filemap_range_has_writeback(struct address_space *mapping,
648 loff_t start_byte, loff_t end_byte)
649 {
650 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT);
651 pgoff_t max = end_byte >> PAGE_SHIFT;
652 struct folio *folio;
653
654 if (end_byte < start_byte)
655 return false;
656
657 rcu_read_lock();
658 xas_for_each(&xas, folio, max) {
659 if (xas_retry(&xas, folio))
660 continue;
661 if (xa_is_value(folio))
662 continue;
663 if (folio_test_dirty(folio) || folio_test_locked(folio) ||
664 folio_test_writeback(folio))
665 break;
666 }
667 rcu_read_unlock();
668 return folio != NULL;
669 }
670 EXPORT_SYMBOL_GPL(filemap_range_has_writeback);
671
672 /**
673 * filemap_write_and_wait_range - write out & wait on a file range
674 * @mapping: the address_space for the pages
675 * @lstart: offset in bytes where the range starts
676 * @lend: offset in bytes where the range ends (inclusive)
677 *
678 * Write out and wait upon file offsets lstart->lend, inclusive.
679 *
680 * Note that @lend is inclusive (describes the last byte to be written) so
681 * that this function can be used to write to the very end-of-file (end = -1).
682 *
683 * Return: error status of the address space.
684 */
filemap_write_and_wait_range(struct address_space * mapping,loff_t lstart,loff_t lend)685 int filemap_write_and_wait_range(struct address_space *mapping,
686 loff_t lstart, loff_t lend)
687 {
688 int err = 0, err2;
689
690 if (lend < lstart)
691 return 0;
692
693 if (mapping_needs_writeback(mapping)) {
694 err = __filemap_fdatawrite_range(mapping, lstart, lend,
695 WB_SYNC_ALL);
696 /*
697 * Even if the above returned error, the pages may be
698 * written partially (e.g. -ENOSPC), so we wait for it.
699 * But the -EIO is special case, it may indicate the worst
700 * thing (e.g. bug) happened, so we avoid waiting for it.
701 */
702 if (err != -EIO)
703 __filemap_fdatawait_range(mapping, lstart, lend);
704 }
705 err2 = filemap_check_errors(mapping);
706 if (!err)
707 err = err2;
708 return err;
709 }
710 EXPORT_SYMBOL(filemap_write_and_wait_range);
711
__filemap_set_wb_err(struct address_space * mapping,int err)712 void __filemap_set_wb_err(struct address_space *mapping, int err)
713 {
714 errseq_t eseq = errseq_set(&mapping->wb_err, err);
715
716 trace_filemap_set_wb_err(mapping, eseq);
717 }
718 EXPORT_SYMBOL(__filemap_set_wb_err);
719
720 /**
721 * file_check_and_advance_wb_err - report wb error (if any) that was previously
722 * and advance wb_err to current one
723 * @file: struct file on which the error is being reported
724 *
725 * When userland calls fsync (or something like nfsd does the equivalent), we
726 * want to report any writeback errors that occurred since the last fsync (or
727 * since the file was opened if there haven't been any).
728 *
729 * Grab the wb_err from the mapping. If it matches what we have in the file,
730 * then just quickly return 0. The file is all caught up.
731 *
732 * If it doesn't match, then take the mapping value, set the "seen" flag in
733 * it and try to swap it into place. If it works, or another task beat us
734 * to it with the new value, then update the f_wb_err and return the error
735 * portion. The error at this point must be reported via proper channels
736 * (a'la fsync, or NFS COMMIT operation, etc.).
737 *
738 * While we handle mapping->wb_err with atomic operations, the f_wb_err
739 * value is protected by the f_lock since we must ensure that it reflects
740 * the latest value swapped in for this file descriptor.
741 *
742 * Return: %0 on success, negative error code otherwise.
743 */
file_check_and_advance_wb_err(struct file * file)744 int file_check_and_advance_wb_err(struct file *file)
745 {
746 int err = 0;
747 errseq_t old = READ_ONCE(file->f_wb_err);
748 struct address_space *mapping = file->f_mapping;
749
750 /* Locklessly handle the common case where nothing has changed */
751 if (errseq_check(&mapping->wb_err, old)) {
752 /* Something changed, must use slow path */
753 spin_lock(&file->f_lock);
754 old = file->f_wb_err;
755 err = errseq_check_and_advance(&mapping->wb_err,
756 &file->f_wb_err);
757 trace_file_check_and_advance_wb_err(file, old);
758 spin_unlock(&file->f_lock);
759 }
760
761 /*
762 * We're mostly using this function as a drop in replacement for
763 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect
764 * that the legacy code would have had on these flags.
765 */
766 clear_bit(AS_EIO, &mapping->flags);
767 clear_bit(AS_ENOSPC, &mapping->flags);
768 return err;
769 }
770 EXPORT_SYMBOL(file_check_and_advance_wb_err);
771
772 /**
773 * file_write_and_wait_range - write out & wait on a file range
774 * @file: file pointing to address_space with pages
775 * @lstart: offset in bytes where the range starts
776 * @lend: offset in bytes where the range ends (inclusive)
777 *
778 * Write out and wait upon file offsets lstart->lend, inclusive.
779 *
780 * Note that @lend is inclusive (describes the last byte to be written) so
781 * that this function can be used to write to the very end-of-file (end = -1).
782 *
783 * After writing out and waiting on the data, we check and advance the
784 * f_wb_err cursor to the latest value, and return any errors detected there.
785 *
786 * Return: %0 on success, negative error code otherwise.
787 */
file_write_and_wait_range(struct file * file,loff_t lstart,loff_t lend)788 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend)
789 {
790 int err = 0, err2;
791 struct address_space *mapping = file->f_mapping;
792
793 if (lend < lstart)
794 return 0;
795
796 if (mapping_needs_writeback(mapping)) {
797 err = __filemap_fdatawrite_range(mapping, lstart, lend,
798 WB_SYNC_ALL);
799 /* See comment of filemap_write_and_wait() */
800 if (err != -EIO)
801 __filemap_fdatawait_range(mapping, lstart, lend);
802 }
803 err2 = file_check_and_advance_wb_err(file);
804 if (!err)
805 err = err2;
806 return err;
807 }
808 EXPORT_SYMBOL(file_write_and_wait_range);
809
810 /**
811 * replace_page_cache_folio - replace a pagecache folio with a new one
812 * @old: folio to be replaced
813 * @new: folio to replace with
814 *
815 * This function replaces a folio in the pagecache with a new one. On
816 * success it acquires the pagecache reference for the new folio and
817 * drops it for the old folio. Both the old and new folios must be
818 * locked. This function does not add the new folio to the LRU, the
819 * caller must do that.
820 *
821 * The remove + add is atomic. This function cannot fail.
822 */
replace_page_cache_folio(struct folio * old,struct folio * new)823 void replace_page_cache_folio(struct folio *old, struct folio *new)
824 {
825 struct address_space *mapping = old->mapping;
826 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio;
827 pgoff_t offset = old->index;
828 XA_STATE(xas, &mapping->i_pages, offset);
829
830 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
831 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
832 VM_BUG_ON_FOLIO(new->mapping, new);
833
834 folio_get(new);
835 new->mapping = mapping;
836 new->index = offset;
837
838 mem_cgroup_replace_folio(old, new);
839
840 xas_lock_irq(&xas);
841 xas_store(&xas, new);
842
843 old->mapping = NULL;
844 /* hugetlb pages do not participate in page cache accounting. */
845 if (!folio_test_hugetlb(old))
846 __lruvec_stat_sub_folio(old, NR_FILE_PAGES);
847 if (!folio_test_hugetlb(new))
848 __lruvec_stat_add_folio(new, NR_FILE_PAGES);
849 if (folio_test_swapbacked(old))
850 __lruvec_stat_sub_folio(old, NR_SHMEM);
851 if (folio_test_swapbacked(new))
852 __lruvec_stat_add_folio(new, NR_SHMEM);
853 xas_unlock_irq(&xas);
854 if (free_folio)
855 free_folio(old);
856 folio_put(old);
857 }
858 EXPORT_SYMBOL_GPL(replace_page_cache_folio);
859
__filemap_add_folio(struct address_space * mapping,struct folio * folio,pgoff_t index,gfp_t gfp,void ** shadowp)860 noinline int __filemap_add_folio(struct address_space *mapping,
861 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
862 {
863 XA_STATE(xas, &mapping->i_pages, index);
864 void *alloced_shadow = NULL;
865 int alloced_order = 0;
866 bool huge;
867 long nr;
868
869 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
870 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio);
871 VM_BUG_ON_FOLIO(folio_order(folio) < mapping_min_folio_order(mapping),
872 folio);
873 mapping_set_update(&xas, mapping);
874
875 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
876 xas_set_order(&xas, index, folio_order(folio));
877 huge = folio_test_hugetlb(folio);
878 nr = folio_nr_pages(folio);
879
880 gfp &= GFP_RECLAIM_MASK;
881 folio_ref_add(folio, nr);
882 folio->mapping = mapping;
883 folio->index = xas.xa_index;
884
885 for (;;) {
886 int order = -1, split_order = 0;
887 void *entry, *old = NULL;
888
889 xas_lock_irq(&xas);
890 xas_for_each_conflict(&xas, entry) {
891 old = entry;
892 if (!xa_is_value(entry)) {
893 xas_set_err(&xas, -EEXIST);
894 goto unlock;
895 }
896 /*
897 * If a larger entry exists,
898 * it will be the first and only entry iterated.
899 */
900 if (order == -1)
901 order = xas_get_order(&xas);
902 }
903
904 /* entry may have changed before we re-acquire the lock */
905 if (alloced_order && (old != alloced_shadow || order != alloced_order)) {
906 xas_destroy(&xas);
907 alloced_order = 0;
908 }
909
910 if (old) {
911 if (order > 0 && order > folio_order(folio)) {
912 /* How to handle large swap entries? */
913 BUG_ON(shmem_mapping(mapping));
914 if (!alloced_order) {
915 split_order = order;
916 goto unlock;
917 }
918 xas_split(&xas, old, order);
919 xas_reset(&xas);
920 }
921 if (shadowp)
922 *shadowp = old;
923 }
924
925 xas_store(&xas, folio);
926 if (xas_error(&xas))
927 goto unlock;
928
929 mapping->nrpages += nr;
930
931 /* hugetlb pages do not participate in page cache accounting */
932 if (!huge) {
933 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
934 if (folio_test_pmd_mappable(folio))
935 __lruvec_stat_mod_folio(folio,
936 NR_FILE_THPS, nr);
937 }
938
939 unlock:
940 xas_unlock_irq(&xas);
941
942 /* split needed, alloc here and retry. */
943 if (split_order) {
944 xas_split_alloc(&xas, old, split_order, gfp);
945 if (xas_error(&xas))
946 goto error;
947 alloced_shadow = old;
948 alloced_order = split_order;
949 xas_reset(&xas);
950 continue;
951 }
952
953 if (!xas_nomem(&xas, gfp))
954 break;
955 }
956
957 if (xas_error(&xas))
958 goto error;
959
960 trace_mm_filemap_add_to_page_cache(folio);
961 return 0;
962 error:
963 folio->mapping = NULL;
964 /* Leave page->index set: truncation relies upon it */
965 folio_put_refs(folio, nr);
966 return xas_error(&xas);
967 }
968 ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO);
969
filemap_add_folio(struct address_space * mapping,struct folio * folio,pgoff_t index,gfp_t gfp)970 int filemap_add_folio(struct address_space *mapping, struct folio *folio,
971 pgoff_t index, gfp_t gfp)
972 {
973 void *shadow = NULL;
974 int ret;
975
976 ret = mem_cgroup_charge(folio, NULL, gfp);
977 if (ret)
978 return ret;
979
980 __folio_set_locked(folio);
981 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow);
982 if (unlikely(ret)) {
983 mem_cgroup_uncharge(folio);
984 __folio_clear_locked(folio);
985 } else {
986 /*
987 * The folio might have been evicted from cache only
988 * recently, in which case it should be activated like
989 * any other repeatedly accessed folio.
990 * The exception is folios getting rewritten; evicting other
991 * data from the working set, only to cache data that will
992 * get overwritten with something else, is a waste of memory.
993 */
994 WARN_ON_ONCE(folio_test_active(folio));
995 if (!(gfp & __GFP_WRITE) && shadow)
996 workingset_refault(folio, shadow);
997 folio_add_lru(folio);
998 }
999 return ret;
1000 }
1001 EXPORT_SYMBOL_GPL(filemap_add_folio);
1002
1003 #ifdef CONFIG_NUMA
filemap_alloc_folio_noprof(gfp_t gfp,unsigned int order)1004 struct folio *filemap_alloc_folio_noprof(gfp_t gfp, unsigned int order)
1005 {
1006 int n;
1007 struct folio *folio;
1008
1009 if (cpuset_do_page_mem_spread()) {
1010 unsigned int cpuset_mems_cookie;
1011 do {
1012 cpuset_mems_cookie = read_mems_allowed_begin();
1013 n = cpuset_mem_spread_node();
1014 folio = __folio_alloc_node_noprof(gfp, order, n);
1015 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie));
1016
1017 return folio;
1018 }
1019 return folio_alloc_noprof(gfp, order);
1020 }
1021 EXPORT_SYMBOL(filemap_alloc_folio_noprof);
1022 #endif
1023
1024 /*
1025 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings
1026 *
1027 * Lock exclusively invalidate_lock of any passed mapping that is not NULL.
1028 *
1029 * @mapping1: the first mapping to lock
1030 * @mapping2: the second mapping to lock
1031 */
filemap_invalidate_lock_two(struct address_space * mapping1,struct address_space * mapping2)1032 void filemap_invalidate_lock_two(struct address_space *mapping1,
1033 struct address_space *mapping2)
1034 {
1035 if (mapping1 > mapping2)
1036 swap(mapping1, mapping2);
1037 if (mapping1)
1038 down_write(&mapping1->invalidate_lock);
1039 if (mapping2 && mapping1 != mapping2)
1040 down_write_nested(&mapping2->invalidate_lock, 1);
1041 }
1042 EXPORT_SYMBOL(filemap_invalidate_lock_two);
1043
1044 /*
1045 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings
1046 *
1047 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL.
1048 *
1049 * @mapping1: the first mapping to unlock
1050 * @mapping2: the second mapping to unlock
1051 */
filemap_invalidate_unlock_two(struct address_space * mapping1,struct address_space * mapping2)1052 void filemap_invalidate_unlock_two(struct address_space *mapping1,
1053 struct address_space *mapping2)
1054 {
1055 if (mapping1)
1056 up_write(&mapping1->invalidate_lock);
1057 if (mapping2 && mapping1 != mapping2)
1058 up_write(&mapping2->invalidate_lock);
1059 }
1060 EXPORT_SYMBOL(filemap_invalidate_unlock_two);
1061
1062 /*
1063 * In order to wait for pages to become available there must be
1064 * waitqueues associated with pages. By using a hash table of
1065 * waitqueues where the bucket discipline is to maintain all
1066 * waiters on the same queue and wake all when any of the pages
1067 * become available, and for the woken contexts to check to be
1068 * sure the appropriate page became available, this saves space
1069 * at a cost of "thundering herd" phenomena during rare hash
1070 * collisions.
1071 */
1072 #define PAGE_WAIT_TABLE_BITS 8
1073 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS)
1074 static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned;
1075
folio_waitqueue(struct folio * folio)1076 static wait_queue_head_t *folio_waitqueue(struct folio *folio)
1077 {
1078 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)];
1079 }
1080
1081 /* How many times do we accept lock stealing from under a waiter? */
1082 static int sysctl_page_lock_unfairness = 5;
1083 static const struct ctl_table filemap_sysctl_table[] = {
1084 {
1085 .procname = "page_lock_unfairness",
1086 .data = &sysctl_page_lock_unfairness,
1087 .maxlen = sizeof(sysctl_page_lock_unfairness),
1088 .mode = 0644,
1089 .proc_handler = proc_dointvec_minmax,
1090 .extra1 = SYSCTL_ZERO,
1091 }
1092 };
1093
pagecache_init(void)1094 void __init pagecache_init(void)
1095 {
1096 int i;
1097
1098 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
1099 init_waitqueue_head(&folio_wait_table[i]);
1100
1101 page_writeback_init();
1102 register_sysctl_init("vm", filemap_sysctl_table);
1103 }
1104
1105 /*
1106 * The page wait code treats the "wait->flags" somewhat unusually, because
1107 * we have multiple different kinds of waits, not just the usual "exclusive"
1108 * one.
1109 *
1110 * We have:
1111 *
1112 * (a) no special bits set:
1113 *
1114 * We're just waiting for the bit to be released, and when a waker
1115 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up,
1116 * and remove it from the wait queue.
1117 *
1118 * Simple and straightforward.
1119 *
1120 * (b) WQ_FLAG_EXCLUSIVE:
1121 *
1122 * The waiter is waiting to get the lock, and only one waiter should
1123 * be woken up to avoid any thundering herd behavior. We'll set the
1124 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue.
1125 *
1126 * This is the traditional exclusive wait.
1127 *
1128 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM:
1129 *
1130 * The waiter is waiting to get the bit, and additionally wants the
1131 * lock to be transferred to it for fair lock behavior. If the lock
1132 * cannot be taken, we stop walking the wait queue without waking
1133 * the waiter.
1134 *
1135 * This is the "fair lock handoff" case, and in addition to setting
1136 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see
1137 * that it now has the lock.
1138 */
wake_page_function(wait_queue_entry_t * wait,unsigned mode,int sync,void * arg)1139 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg)
1140 {
1141 unsigned int flags;
1142 struct wait_page_key *key = arg;
1143 struct wait_page_queue *wait_page
1144 = container_of(wait, struct wait_page_queue, wait);
1145
1146 if (!wake_page_match(wait_page, key))
1147 return 0;
1148
1149 /*
1150 * If it's a lock handoff wait, we get the bit for it, and
1151 * stop walking (and do not wake it up) if we can't.
1152 */
1153 flags = wait->flags;
1154 if (flags & WQ_FLAG_EXCLUSIVE) {
1155 if (test_bit(key->bit_nr, &key->folio->flags))
1156 return -1;
1157 if (flags & WQ_FLAG_CUSTOM) {
1158 if (test_and_set_bit(key->bit_nr, &key->folio->flags))
1159 return -1;
1160 flags |= WQ_FLAG_DONE;
1161 }
1162 }
1163
1164 /*
1165 * We are holding the wait-queue lock, but the waiter that
1166 * is waiting for this will be checking the flags without
1167 * any locking.
1168 *
1169 * So update the flags atomically, and wake up the waiter
1170 * afterwards to avoid any races. This store-release pairs
1171 * with the load-acquire in folio_wait_bit_common().
1172 */
1173 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN);
1174 wake_up_state(wait->private, mode);
1175
1176 /*
1177 * Ok, we have successfully done what we're waiting for,
1178 * and we can unconditionally remove the wait entry.
1179 *
1180 * Note that this pairs with the "finish_wait()" in the
1181 * waiter, and has to be the absolute last thing we do.
1182 * After this list_del_init(&wait->entry) the wait entry
1183 * might be de-allocated and the process might even have
1184 * exited.
1185 */
1186 list_del_init_careful(&wait->entry);
1187 return (flags & WQ_FLAG_EXCLUSIVE) != 0;
1188 }
1189
folio_wake_bit(struct folio * folio,int bit_nr)1190 static void folio_wake_bit(struct folio *folio, int bit_nr)
1191 {
1192 wait_queue_head_t *q = folio_waitqueue(folio);
1193 struct wait_page_key key;
1194 unsigned long flags;
1195
1196 key.folio = folio;
1197 key.bit_nr = bit_nr;
1198 key.page_match = 0;
1199
1200 spin_lock_irqsave(&q->lock, flags);
1201 __wake_up_locked_key(q, TASK_NORMAL, &key);
1202
1203 /*
1204 * It's possible to miss clearing waiters here, when we woke our page
1205 * waiters, but the hashed waitqueue has waiters for other pages on it.
1206 * That's okay, it's a rare case. The next waker will clear it.
1207 *
1208 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE,
1209 * other), the flag may be cleared in the course of freeing the page;
1210 * but that is not required for correctness.
1211 */
1212 if (!waitqueue_active(q) || !key.page_match)
1213 folio_clear_waiters(folio);
1214
1215 spin_unlock_irqrestore(&q->lock, flags);
1216 }
1217
1218 /*
1219 * A choice of three behaviors for folio_wait_bit_common():
1220 */
1221 enum behavior {
1222 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like
1223 * __folio_lock() waiting on then setting PG_locked.
1224 */
1225 SHARED, /* Hold ref to page and check the bit when woken, like
1226 * folio_wait_writeback() waiting on PG_writeback.
1227 */
1228 DROP, /* Drop ref to page before wait, no check when woken,
1229 * like folio_put_wait_locked() on PG_locked.
1230 */
1231 };
1232
1233 /*
1234 * Attempt to check (or get) the folio flag, and mark us done
1235 * if successful.
1236 */
folio_trylock_flag(struct folio * folio,int bit_nr,struct wait_queue_entry * wait)1237 static inline bool folio_trylock_flag(struct folio *folio, int bit_nr,
1238 struct wait_queue_entry *wait)
1239 {
1240 if (wait->flags & WQ_FLAG_EXCLUSIVE) {
1241 if (test_and_set_bit(bit_nr, &folio->flags))
1242 return false;
1243 } else if (test_bit(bit_nr, &folio->flags))
1244 return false;
1245
1246 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE;
1247 return true;
1248 }
1249
folio_wait_bit_common(struct folio * folio,int bit_nr,int state,enum behavior behavior)1250 static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
1251 int state, enum behavior behavior)
1252 {
1253 wait_queue_head_t *q = folio_waitqueue(folio);
1254 int unfairness = sysctl_page_lock_unfairness;
1255 struct wait_page_queue wait_page;
1256 wait_queue_entry_t *wait = &wait_page.wait;
1257 bool thrashing = false;
1258 unsigned long pflags;
1259 bool in_thrashing;
1260
1261 if (bit_nr == PG_locked &&
1262 !folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1263 delayacct_thrashing_start(&in_thrashing);
1264 psi_memstall_enter(&pflags);
1265 thrashing = true;
1266 }
1267
1268 init_wait(wait);
1269 wait->func = wake_page_function;
1270 wait_page.folio = folio;
1271 wait_page.bit_nr = bit_nr;
1272
1273 repeat:
1274 wait->flags = 0;
1275 if (behavior == EXCLUSIVE) {
1276 wait->flags = WQ_FLAG_EXCLUSIVE;
1277 if (--unfairness < 0)
1278 wait->flags |= WQ_FLAG_CUSTOM;
1279 }
1280
1281 /*
1282 * Do one last check whether we can get the
1283 * page bit synchronously.
1284 *
1285 * Do the folio_set_waiters() marking before that
1286 * to let any waker we _just_ missed know they
1287 * need to wake us up (otherwise they'll never
1288 * even go to the slow case that looks at the
1289 * page queue), and add ourselves to the wait
1290 * queue if we need to sleep.
1291 *
1292 * This part needs to be done under the queue
1293 * lock to avoid races.
1294 */
1295 spin_lock_irq(&q->lock);
1296 folio_set_waiters(folio);
1297 if (!folio_trylock_flag(folio, bit_nr, wait))
1298 __add_wait_queue_entry_tail(q, wait);
1299 spin_unlock_irq(&q->lock);
1300
1301 /*
1302 * From now on, all the logic will be based on
1303 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to
1304 * see whether the page bit testing has already
1305 * been done by the wake function.
1306 *
1307 * We can drop our reference to the folio.
1308 */
1309 if (behavior == DROP)
1310 folio_put(folio);
1311
1312 /*
1313 * Note that until the "finish_wait()", or until
1314 * we see the WQ_FLAG_WOKEN flag, we need to
1315 * be very careful with the 'wait->flags', because
1316 * we may race with a waker that sets them.
1317 */
1318 for (;;) {
1319 unsigned int flags;
1320
1321 set_current_state(state);
1322
1323 /* Loop until we've been woken or interrupted */
1324 flags = smp_load_acquire(&wait->flags);
1325 if (!(flags & WQ_FLAG_WOKEN)) {
1326 if (signal_pending_state(state, current))
1327 break;
1328
1329 io_schedule();
1330 continue;
1331 }
1332
1333 /* If we were non-exclusive, we're done */
1334 if (behavior != EXCLUSIVE)
1335 break;
1336
1337 /* If the waker got the lock for us, we're done */
1338 if (flags & WQ_FLAG_DONE)
1339 break;
1340
1341 /*
1342 * Otherwise, if we're getting the lock, we need to
1343 * try to get it ourselves.
1344 *
1345 * And if that fails, we'll have to retry this all.
1346 */
1347 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0))))
1348 goto repeat;
1349
1350 wait->flags |= WQ_FLAG_DONE;
1351 break;
1352 }
1353
1354 /*
1355 * If a signal happened, this 'finish_wait()' may remove the last
1356 * waiter from the wait-queues, but the folio waiters bit will remain
1357 * set. That's ok. The next wakeup will take care of it, and trying
1358 * to do it here would be difficult and prone to races.
1359 */
1360 finish_wait(q, wait);
1361
1362 if (thrashing) {
1363 delayacct_thrashing_end(&in_thrashing);
1364 psi_memstall_leave(&pflags);
1365 }
1366
1367 /*
1368 * NOTE! The wait->flags weren't stable until we've done the
1369 * 'finish_wait()', and we could have exited the loop above due
1370 * to a signal, and had a wakeup event happen after the signal
1371 * test but before the 'finish_wait()'.
1372 *
1373 * So only after the finish_wait() can we reliably determine
1374 * if we got woken up or not, so we can now figure out the final
1375 * return value based on that state without races.
1376 *
1377 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive
1378 * waiter, but an exclusive one requires WQ_FLAG_DONE.
1379 */
1380 if (behavior == EXCLUSIVE)
1381 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR;
1382
1383 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR;
1384 }
1385
1386 #ifdef CONFIG_MIGRATION
1387 /**
1388 * migration_entry_wait_on_locked - Wait for a migration entry to be removed
1389 * @entry: migration swap entry.
1390 * @ptl: already locked ptl. This function will drop the lock.
1391 *
1392 * Wait for a migration entry referencing the given page to be removed. This is
1393 * equivalent to folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE) except
1394 * this can be called without taking a reference on the page. Instead this
1395 * should be called while holding the ptl for the migration entry referencing
1396 * the page.
1397 *
1398 * Returns after unlocking the ptl.
1399 *
1400 * This follows the same logic as folio_wait_bit_common() so see the comments
1401 * there.
1402 */
migration_entry_wait_on_locked(swp_entry_t entry,spinlock_t * ptl)1403 void migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl)
1404 __releases(ptl)
1405 {
1406 struct wait_page_queue wait_page;
1407 wait_queue_entry_t *wait = &wait_page.wait;
1408 bool thrashing = false;
1409 unsigned long pflags;
1410 bool in_thrashing;
1411 wait_queue_head_t *q;
1412 struct folio *folio = pfn_swap_entry_folio(entry);
1413
1414 q = folio_waitqueue(folio);
1415 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) {
1416 delayacct_thrashing_start(&in_thrashing);
1417 psi_memstall_enter(&pflags);
1418 thrashing = true;
1419 }
1420
1421 init_wait(wait);
1422 wait->func = wake_page_function;
1423 wait_page.folio = folio;
1424 wait_page.bit_nr = PG_locked;
1425 wait->flags = 0;
1426
1427 spin_lock_irq(&q->lock);
1428 folio_set_waiters(folio);
1429 if (!folio_trylock_flag(folio, PG_locked, wait))
1430 __add_wait_queue_entry_tail(q, wait);
1431 spin_unlock_irq(&q->lock);
1432
1433 /*
1434 * If a migration entry exists for the page the migration path must hold
1435 * a valid reference to the page, and it must take the ptl to remove the
1436 * migration entry. So the page is valid until the ptl is dropped.
1437 */
1438 spin_unlock(ptl);
1439
1440 for (;;) {
1441 unsigned int flags;
1442
1443 set_current_state(TASK_UNINTERRUPTIBLE);
1444
1445 /* Loop until we've been woken or interrupted */
1446 flags = smp_load_acquire(&wait->flags);
1447 if (!(flags & WQ_FLAG_WOKEN)) {
1448 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current))
1449 break;
1450
1451 io_schedule();
1452 continue;
1453 }
1454 break;
1455 }
1456
1457 finish_wait(q, wait);
1458
1459 if (thrashing) {
1460 delayacct_thrashing_end(&in_thrashing);
1461 psi_memstall_leave(&pflags);
1462 }
1463 }
1464 #endif
1465
folio_wait_bit(struct folio * folio,int bit_nr)1466 void folio_wait_bit(struct folio *folio, int bit_nr)
1467 {
1468 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED);
1469 }
1470 EXPORT_SYMBOL(folio_wait_bit);
1471
folio_wait_bit_killable(struct folio * folio,int bit_nr)1472 int folio_wait_bit_killable(struct folio *folio, int bit_nr)
1473 {
1474 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED);
1475 }
1476 EXPORT_SYMBOL(folio_wait_bit_killable);
1477
1478 /**
1479 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked
1480 * @folio: The folio to wait for.
1481 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc).
1482 *
1483 * The caller should hold a reference on @folio. They expect the page to
1484 * become unlocked relatively soon, but do not wish to hold up migration
1485 * (for example) by holding the reference while waiting for the folio to
1486 * come unlocked. After this function returns, the caller should not
1487 * dereference @folio.
1488 *
1489 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal.
1490 */
folio_put_wait_locked(struct folio * folio,int state)1491 static int folio_put_wait_locked(struct folio *folio, int state)
1492 {
1493 return folio_wait_bit_common(folio, PG_locked, state, DROP);
1494 }
1495
1496 /**
1497 * folio_unlock - Unlock a locked folio.
1498 * @folio: The folio.
1499 *
1500 * Unlocks the folio and wakes up any thread sleeping on the page lock.
1501 *
1502 * Context: May be called from interrupt or process context. May not be
1503 * called from NMI context.
1504 */
folio_unlock(struct folio * folio)1505 void folio_unlock(struct folio *folio)
1506 {
1507 /* Bit 7 allows x86 to check the byte's sign bit */
1508 BUILD_BUG_ON(PG_waiters != 7);
1509 BUILD_BUG_ON(PG_locked > 7);
1510 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1511 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked))
1512 folio_wake_bit(folio, PG_locked);
1513 }
1514 EXPORT_SYMBOL(folio_unlock);
1515
1516 /**
1517 * folio_end_read - End read on a folio.
1518 * @folio: The folio.
1519 * @success: True if all reads completed successfully.
1520 *
1521 * When all reads against a folio have completed, filesystems should
1522 * call this function to let the pagecache know that no more reads
1523 * are outstanding. This will unlock the folio and wake up any thread
1524 * sleeping on the lock. The folio will also be marked uptodate if all
1525 * reads succeeded.
1526 *
1527 * Context: May be called from interrupt or process context. May not be
1528 * called from NMI context.
1529 */
folio_end_read(struct folio * folio,bool success)1530 void folio_end_read(struct folio *folio, bool success)
1531 {
1532 unsigned long mask = 1 << PG_locked;
1533
1534 /* Must be in bottom byte for x86 to work */
1535 BUILD_BUG_ON(PG_uptodate > 7);
1536 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1537 VM_BUG_ON_FOLIO(success && folio_test_uptodate(folio), folio);
1538
1539 if (likely(success))
1540 mask |= 1 << PG_uptodate;
1541 if (folio_xor_flags_has_waiters(folio, mask))
1542 folio_wake_bit(folio, PG_locked);
1543 }
1544 EXPORT_SYMBOL(folio_end_read);
1545
1546 /**
1547 * folio_end_private_2 - Clear PG_private_2 and wake any waiters.
1548 * @folio: The folio.
1549 *
1550 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for
1551 * it. The folio reference held for PG_private_2 being set is released.
1552 *
1553 * This is, for example, used when a netfs folio is being written to a local
1554 * disk cache, thereby allowing writes to the cache for the same folio to be
1555 * serialised.
1556 */
folio_end_private_2(struct folio * folio)1557 void folio_end_private_2(struct folio *folio)
1558 {
1559 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio);
1560 clear_bit_unlock(PG_private_2, folio_flags(folio, 0));
1561 folio_wake_bit(folio, PG_private_2);
1562 folio_put(folio);
1563 }
1564 EXPORT_SYMBOL(folio_end_private_2);
1565
1566 /**
1567 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio.
1568 * @folio: The folio to wait on.
1569 *
1570 * Wait for PG_private_2 to be cleared on a folio.
1571 */
folio_wait_private_2(struct folio * folio)1572 void folio_wait_private_2(struct folio *folio)
1573 {
1574 while (folio_test_private_2(folio))
1575 folio_wait_bit(folio, PG_private_2);
1576 }
1577 EXPORT_SYMBOL(folio_wait_private_2);
1578
1579 /**
1580 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio.
1581 * @folio: The folio to wait on.
1582 *
1583 * Wait for PG_private_2 to be cleared on a folio or until a fatal signal is
1584 * received by the calling task.
1585 *
1586 * Return:
1587 * - 0 if successful.
1588 * - -EINTR if a fatal signal was encountered.
1589 */
folio_wait_private_2_killable(struct folio * folio)1590 int folio_wait_private_2_killable(struct folio *folio)
1591 {
1592 int ret = 0;
1593
1594 while (folio_test_private_2(folio)) {
1595 ret = folio_wait_bit_killable(folio, PG_private_2);
1596 if (ret < 0)
1597 break;
1598 }
1599
1600 return ret;
1601 }
1602 EXPORT_SYMBOL(folio_wait_private_2_killable);
1603
1604 /*
1605 * If folio was marked as dropbehind, then pages should be dropped when writeback
1606 * completes. Do that now. If we fail, it's likely because of a big folio -
1607 * just reset dropbehind for that case and latter completions should invalidate.
1608 */
folio_end_dropbehind_write(struct folio * folio)1609 static void folio_end_dropbehind_write(struct folio *folio)
1610 {
1611 /*
1612 * Hitting !in_task() should not happen off RWF_DONTCACHE writeback,
1613 * but can happen if normal writeback just happens to find dirty folios
1614 * that were created as part of uncached writeback, and that writeback
1615 * would otherwise not need non-IRQ handling. Just skip the
1616 * invalidation in that case.
1617 */
1618 if (in_task() && folio_trylock(folio)) {
1619 if (folio->mapping)
1620 folio_unmap_invalidate(folio->mapping, folio, 0);
1621 folio_unlock(folio);
1622 }
1623 }
1624
1625 /**
1626 * folio_end_writeback - End writeback against a folio.
1627 * @folio: The folio.
1628 *
1629 * The folio must actually be under writeback.
1630 *
1631 * Context: May be called from process or interrupt context.
1632 */
folio_end_writeback(struct folio * folio)1633 void folio_end_writeback(struct folio *folio)
1634 {
1635 bool folio_dropbehind = false;
1636
1637 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
1638
1639 /*
1640 * folio_test_clear_reclaim() could be used here but it is an
1641 * atomic operation and overkill in this particular case. Failing
1642 * to shuffle a folio marked for immediate reclaim is too mild
1643 * a gain to justify taking an atomic operation penalty at the
1644 * end of every folio writeback.
1645 */
1646 if (folio_test_reclaim(folio)) {
1647 folio_clear_reclaim(folio);
1648 folio_rotate_reclaimable(folio);
1649 }
1650
1651 /*
1652 * Writeback does not hold a folio reference of its own, relying
1653 * on truncation to wait for the clearing of PG_writeback.
1654 * But here we must make sure that the folio is not freed and
1655 * reused before the folio_wake_bit().
1656 */
1657 folio_get(folio);
1658 if (!folio_test_dirty(folio))
1659 folio_dropbehind = folio_test_clear_dropbehind(folio);
1660 if (__folio_end_writeback(folio))
1661 folio_wake_bit(folio, PG_writeback);
1662 acct_reclaim_writeback(folio);
1663
1664 if (folio_dropbehind)
1665 folio_end_dropbehind_write(folio);
1666 folio_put(folio);
1667 }
1668 EXPORT_SYMBOL(folio_end_writeback);
1669
1670 /**
1671 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
1672 * @folio: The folio to lock
1673 */
__folio_lock(struct folio * folio)1674 void __folio_lock(struct folio *folio)
1675 {
1676 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE,
1677 EXCLUSIVE);
1678 }
1679 EXPORT_SYMBOL(__folio_lock);
1680
__folio_lock_killable(struct folio * folio)1681 int __folio_lock_killable(struct folio *folio)
1682 {
1683 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE,
1684 EXCLUSIVE);
1685 }
1686 EXPORT_SYMBOL_GPL(__folio_lock_killable);
1687
__folio_lock_async(struct folio * folio,struct wait_page_queue * wait)1688 static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
1689 {
1690 struct wait_queue_head *q = folio_waitqueue(folio);
1691 int ret;
1692
1693 wait->folio = folio;
1694 wait->bit_nr = PG_locked;
1695
1696 spin_lock_irq(&q->lock);
1697 __add_wait_queue_entry_tail(q, &wait->wait);
1698 folio_set_waiters(folio);
1699 ret = !folio_trylock(folio);
1700 /*
1701 * If we were successful now, we know we're still on the
1702 * waitqueue as we're still under the lock. This means it's
1703 * safe to remove and return success, we know the callback
1704 * isn't going to trigger.
1705 */
1706 if (!ret)
1707 __remove_wait_queue(q, &wait->wait);
1708 else
1709 ret = -EIOCBQUEUED;
1710 spin_unlock_irq(&q->lock);
1711 return ret;
1712 }
1713
1714 /*
1715 * Return values:
1716 * 0 - folio is locked.
1717 * non-zero - folio is not locked.
1718 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or
1719 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and
1720 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held.
1721 *
1722 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0
1723 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed.
1724 */
__folio_lock_or_retry(struct folio * folio,struct vm_fault * vmf)1725 vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf)
1726 {
1727 unsigned int flags = vmf->flags;
1728
1729 if (fault_flag_allow_retry_first(flags)) {
1730 /*
1731 * CAUTION! In this case, mmap_lock/per-VMA lock is not
1732 * released even though returning VM_FAULT_RETRY.
1733 */
1734 if (flags & FAULT_FLAG_RETRY_NOWAIT)
1735 return VM_FAULT_RETRY;
1736
1737 release_fault_lock(vmf);
1738 if (flags & FAULT_FLAG_KILLABLE)
1739 folio_wait_locked_killable(folio);
1740 else
1741 folio_wait_locked(folio);
1742 return VM_FAULT_RETRY;
1743 }
1744 if (flags & FAULT_FLAG_KILLABLE) {
1745 bool ret;
1746
1747 ret = __folio_lock_killable(folio);
1748 if (ret) {
1749 release_fault_lock(vmf);
1750 return VM_FAULT_RETRY;
1751 }
1752 } else {
1753 __folio_lock(folio);
1754 }
1755
1756 return 0;
1757 }
1758
1759 /**
1760 * page_cache_next_miss() - Find the next gap in the page cache.
1761 * @mapping: Mapping.
1762 * @index: Index.
1763 * @max_scan: Maximum range to search.
1764 *
1765 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the
1766 * gap with the lowest index.
1767 *
1768 * This function may be called under the rcu_read_lock. However, this will
1769 * not atomically search a snapshot of the cache at a single point in time.
1770 * For example, if a gap is created at index 5, then subsequently a gap is
1771 * created at index 10, page_cache_next_miss covering both indices may
1772 * return 10 if called under the rcu_read_lock.
1773 *
1774 * Return: The index of the gap if found, otherwise an index outside the
1775 * range specified (in which case 'return - index >= max_scan' will be true).
1776 * In the rare case of index wrap-around, 0 will be returned.
1777 */
page_cache_next_miss(struct address_space * mapping,pgoff_t index,unsigned long max_scan)1778 pgoff_t page_cache_next_miss(struct address_space *mapping,
1779 pgoff_t index, unsigned long max_scan)
1780 {
1781 XA_STATE(xas, &mapping->i_pages, index);
1782
1783 while (max_scan--) {
1784 void *entry = xas_next(&xas);
1785 if (!entry || xa_is_value(entry))
1786 return xas.xa_index;
1787 if (xas.xa_index == 0)
1788 return 0;
1789 }
1790
1791 return index + max_scan;
1792 }
1793 EXPORT_SYMBOL(page_cache_next_miss);
1794
1795 /**
1796 * page_cache_prev_miss() - Find the previous gap in the page cache.
1797 * @mapping: Mapping.
1798 * @index: Index.
1799 * @max_scan: Maximum range to search.
1800 *
1801 * Search the range [max(index - max_scan + 1, 0), index] for the
1802 * gap with the highest index.
1803 *
1804 * This function may be called under the rcu_read_lock. However, this will
1805 * not atomically search a snapshot of the cache at a single point in time.
1806 * For example, if a gap is created at index 10, then subsequently a gap is
1807 * created at index 5, page_cache_prev_miss() covering both indices may
1808 * return 5 if called under the rcu_read_lock.
1809 *
1810 * Return: The index of the gap if found, otherwise an index outside the
1811 * range specified (in which case 'index - return >= max_scan' will be true).
1812 * In the rare case of wrap-around, ULONG_MAX will be returned.
1813 */
page_cache_prev_miss(struct address_space * mapping,pgoff_t index,unsigned long max_scan)1814 pgoff_t page_cache_prev_miss(struct address_space *mapping,
1815 pgoff_t index, unsigned long max_scan)
1816 {
1817 XA_STATE(xas, &mapping->i_pages, index);
1818
1819 while (max_scan--) {
1820 void *entry = xas_prev(&xas);
1821 if (!entry || xa_is_value(entry))
1822 break;
1823 if (xas.xa_index == ULONG_MAX)
1824 break;
1825 }
1826
1827 return xas.xa_index;
1828 }
1829 EXPORT_SYMBOL(page_cache_prev_miss);
1830
1831 /*
1832 * Lockless page cache protocol:
1833 * On the lookup side:
1834 * 1. Load the folio from i_pages
1835 * 2. Increment the refcount if it's not zero
1836 * 3. If the folio is not found by xas_reload(), put the refcount and retry
1837 *
1838 * On the removal side:
1839 * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
1840 * B. Remove the page from i_pages
1841 * C. Return the page to the page allocator
1842 *
1843 * This means that any page may have its reference count temporarily
1844 * increased by a speculative page cache (or GUP-fast) lookup as it can
1845 * be allocated by another user before the RCU grace period expires.
1846 * Because the refcount temporarily acquired here may end up being the
1847 * last refcount on the page, any page allocation must be freeable by
1848 * folio_put().
1849 */
1850
1851 /*
1852 * filemap_get_entry - Get a page cache entry.
1853 * @mapping: the address_space to search
1854 * @index: The page cache index.
1855 *
1856 * Looks up the page cache entry at @mapping & @index. If it is a folio,
1857 * it is returned with an increased refcount. If it is a shadow entry
1858 * of a previously evicted folio, or a swap entry from shmem/tmpfs,
1859 * it is returned without further action.
1860 *
1861 * Return: The folio, swap or shadow entry, %NULL if nothing is found.
1862 */
filemap_get_entry(struct address_space * mapping,pgoff_t index)1863 void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
1864 {
1865 XA_STATE(xas, &mapping->i_pages, index);
1866 struct folio *folio;
1867
1868 rcu_read_lock();
1869 repeat:
1870 xas_reset(&xas);
1871 folio = xas_load(&xas);
1872 if (xas_retry(&xas, folio))
1873 goto repeat;
1874 /*
1875 * A shadow entry of a recently evicted page, or a swap entry from
1876 * shmem/tmpfs. Return it without attempting to raise page count.
1877 */
1878 if (!folio || xa_is_value(folio))
1879 goto out;
1880
1881 if (!folio_try_get(folio))
1882 goto repeat;
1883
1884 if (unlikely(folio != xas_reload(&xas))) {
1885 folio_put(folio);
1886 goto repeat;
1887 }
1888 out:
1889 rcu_read_unlock();
1890
1891 return folio;
1892 }
1893
1894 /**
1895 * __filemap_get_folio - Find and get a reference to a folio.
1896 * @mapping: The address_space to search.
1897 * @index: The page index.
1898 * @fgp_flags: %FGP flags modify how the folio is returned.
1899 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified.
1900 *
1901 * Looks up the page cache entry at @mapping & @index.
1902 *
1903 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even
1904 * if the %GFP flags specified for %FGP_CREAT are atomic.
1905 *
1906 * If this function returns a folio, it is returned with an increased refcount.
1907 *
1908 * Return: The found folio or an ERR_PTR() otherwise.
1909 */
__filemap_get_folio(struct address_space * mapping,pgoff_t index,fgf_t fgp_flags,gfp_t gfp)1910 struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index,
1911 fgf_t fgp_flags, gfp_t gfp)
1912 {
1913 struct folio *folio;
1914
1915 repeat:
1916 folio = filemap_get_entry(mapping, index);
1917 if (xa_is_value(folio))
1918 folio = NULL;
1919 if (!folio)
1920 goto no_page;
1921
1922 if (fgp_flags & FGP_LOCK) {
1923 if (fgp_flags & FGP_NOWAIT) {
1924 if (!folio_trylock(folio)) {
1925 folio_put(folio);
1926 return ERR_PTR(-EAGAIN);
1927 }
1928 } else {
1929 folio_lock(folio);
1930 }
1931
1932 /* Has the page been truncated? */
1933 if (unlikely(folio->mapping != mapping)) {
1934 folio_unlock(folio);
1935 folio_put(folio);
1936 goto repeat;
1937 }
1938 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
1939 }
1940
1941 if (fgp_flags & FGP_ACCESSED)
1942 folio_mark_accessed(folio);
1943 else if (fgp_flags & FGP_WRITE) {
1944 /* Clear idle flag for buffer write */
1945 if (folio_test_idle(folio))
1946 folio_clear_idle(folio);
1947 }
1948
1949 if (fgp_flags & FGP_STABLE)
1950 folio_wait_stable(folio);
1951 no_page:
1952 if (!folio && (fgp_flags & FGP_CREAT)) {
1953 unsigned int min_order = mapping_min_folio_order(mapping);
1954 unsigned int order = max(min_order, FGF_GET_ORDER(fgp_flags));
1955 int err;
1956 index = mapping_align_index(mapping, index);
1957
1958 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping))
1959 gfp |= __GFP_WRITE;
1960 if (fgp_flags & FGP_NOFS)
1961 gfp &= ~__GFP_FS;
1962 if (fgp_flags & FGP_NOWAIT) {
1963 gfp &= ~GFP_KERNEL;
1964 gfp |= GFP_NOWAIT | __GFP_NOWARN;
1965 }
1966 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP))))
1967 fgp_flags |= FGP_LOCK;
1968
1969 if (order > mapping_max_folio_order(mapping))
1970 order = mapping_max_folio_order(mapping);
1971 /* If we're not aligned, allocate a smaller folio */
1972 if (index & ((1UL << order) - 1))
1973 order = __ffs(index);
1974
1975 do {
1976 gfp_t alloc_gfp = gfp;
1977
1978 err = -ENOMEM;
1979 if (order > min_order)
1980 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN;
1981 folio = filemap_alloc_folio(alloc_gfp, order);
1982 if (!folio)
1983 continue;
1984
1985 /* Init accessed so avoid atomic mark_page_accessed later */
1986 if (fgp_flags & FGP_ACCESSED)
1987 __folio_set_referenced(folio);
1988 if (fgp_flags & FGP_DONTCACHE)
1989 __folio_set_dropbehind(folio);
1990
1991 err = filemap_add_folio(mapping, folio, index, gfp);
1992 if (!err)
1993 break;
1994 folio_put(folio);
1995 folio = NULL;
1996 } while (order-- > min_order);
1997
1998 if (err == -EEXIST)
1999 goto repeat;
2000 if (err) {
2001 /*
2002 * When NOWAIT I/O fails to allocate folios this could
2003 * be due to a nonblocking memory allocation and not
2004 * because the system actually is out of memory.
2005 * Return -EAGAIN so that there caller retries in a
2006 * blocking fashion instead of propagating -ENOMEM
2007 * to the application.
2008 */
2009 if ((fgp_flags & FGP_NOWAIT) && err == -ENOMEM)
2010 err = -EAGAIN;
2011 return ERR_PTR(err);
2012 }
2013 /*
2014 * filemap_add_folio locks the page, and for mmap
2015 * we expect an unlocked page.
2016 */
2017 if (folio && (fgp_flags & FGP_FOR_MMAP))
2018 folio_unlock(folio);
2019 }
2020
2021 if (!folio)
2022 return ERR_PTR(-ENOENT);
2023 /* not an uncached lookup, clear uncached if set */
2024 if (folio_test_dropbehind(folio) && !(fgp_flags & FGP_DONTCACHE))
2025 folio_clear_dropbehind(folio);
2026 return folio;
2027 }
2028 EXPORT_SYMBOL(__filemap_get_folio);
2029
find_get_entry(struct xa_state * xas,pgoff_t max,xa_mark_t mark)2030 static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max,
2031 xa_mark_t mark)
2032 {
2033 struct folio *folio;
2034
2035 retry:
2036 if (mark == XA_PRESENT)
2037 folio = xas_find(xas, max);
2038 else
2039 folio = xas_find_marked(xas, max, mark);
2040
2041 if (xas_retry(xas, folio))
2042 goto retry;
2043 /*
2044 * A shadow entry of a recently evicted page, a swap
2045 * entry from shmem/tmpfs or a DAX entry. Return it
2046 * without attempting to raise page count.
2047 */
2048 if (!folio || xa_is_value(folio))
2049 return folio;
2050
2051 if (!folio_try_get(folio))
2052 goto reset;
2053
2054 if (unlikely(folio != xas_reload(xas))) {
2055 folio_put(folio);
2056 goto reset;
2057 }
2058
2059 return folio;
2060 reset:
2061 xas_reset(xas);
2062 goto retry;
2063 }
2064
2065 /**
2066 * find_get_entries - gang pagecache lookup
2067 * @mapping: The address_space to search
2068 * @start: The starting page cache index
2069 * @end: The final page index (inclusive).
2070 * @fbatch: Where the resulting entries are placed.
2071 * @indices: The cache indices corresponding to the entries in @entries
2072 *
2073 * find_get_entries() will search for and return a batch of entries in
2074 * the mapping. The entries are placed in @fbatch. find_get_entries()
2075 * takes a reference on any actual folios it returns.
2076 *
2077 * The entries have ascending indexes. The indices may not be consecutive
2078 * due to not-present entries or large folios.
2079 *
2080 * Any shadow entries of evicted folios, or swap entries from
2081 * shmem/tmpfs, are included in the returned array.
2082 *
2083 * Return: The number of entries which were found.
2084 */
find_get_entries(struct address_space * mapping,pgoff_t * start,pgoff_t end,struct folio_batch * fbatch,pgoff_t * indices)2085 unsigned find_get_entries(struct address_space *mapping, pgoff_t *start,
2086 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2087 {
2088 XA_STATE(xas, &mapping->i_pages, *start);
2089 struct folio *folio;
2090
2091 rcu_read_lock();
2092 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) {
2093 indices[fbatch->nr] = xas.xa_index;
2094 if (!folio_batch_add(fbatch, folio))
2095 break;
2096 }
2097
2098 if (folio_batch_count(fbatch)) {
2099 unsigned long nr;
2100 int idx = folio_batch_count(fbatch) - 1;
2101
2102 folio = fbatch->folios[idx];
2103 if (!xa_is_value(folio))
2104 nr = folio_nr_pages(folio);
2105 else
2106 nr = 1 << xa_get_order(&mapping->i_pages, indices[idx]);
2107 *start = round_down(indices[idx] + nr, nr);
2108 }
2109 rcu_read_unlock();
2110
2111 return folio_batch_count(fbatch);
2112 }
2113
2114 /**
2115 * find_lock_entries - Find a batch of pagecache entries.
2116 * @mapping: The address_space to search.
2117 * @start: The starting page cache index.
2118 * @end: The final page index (inclusive).
2119 * @fbatch: Where the resulting entries are placed.
2120 * @indices: The cache indices of the entries in @fbatch.
2121 *
2122 * find_lock_entries() will return a batch of entries from @mapping.
2123 * Swap, shadow and DAX entries are included. Folios are returned
2124 * locked and with an incremented refcount. Folios which are locked
2125 * by somebody else or under writeback are skipped. Folios which are
2126 * partially outside the range are not returned.
2127 *
2128 * The entries have ascending indexes. The indices may not be consecutive
2129 * due to not-present entries, large folios, folios which could not be
2130 * locked or folios under writeback.
2131 *
2132 * Return: The number of entries which were found.
2133 */
find_lock_entries(struct address_space * mapping,pgoff_t * start,pgoff_t end,struct folio_batch * fbatch,pgoff_t * indices)2134 unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start,
2135 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices)
2136 {
2137 XA_STATE(xas, &mapping->i_pages, *start);
2138 struct folio *folio;
2139
2140 rcu_read_lock();
2141 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) {
2142 unsigned long base;
2143 unsigned long nr;
2144
2145 if (!xa_is_value(folio)) {
2146 nr = folio_nr_pages(folio);
2147 base = folio->index;
2148 /* Omit large folio which begins before the start */
2149 if (base < *start)
2150 goto put;
2151 /* Omit large folio which extends beyond the end */
2152 if (base + nr - 1 > end)
2153 goto put;
2154 if (!folio_trylock(folio))
2155 goto put;
2156 if (folio->mapping != mapping ||
2157 folio_test_writeback(folio))
2158 goto unlock;
2159 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index),
2160 folio);
2161 } else {
2162 nr = 1 << xas_get_order(&xas);
2163 base = xas.xa_index & ~(nr - 1);
2164 /* Omit order>0 value which begins before the start */
2165 if (base < *start)
2166 continue;
2167 /* Omit order>0 value which extends beyond the end */
2168 if (base + nr - 1 > end)
2169 break;
2170 }
2171
2172 /* Update start now so that last update is correct on return */
2173 *start = base + nr;
2174 indices[fbatch->nr] = xas.xa_index;
2175 if (!folio_batch_add(fbatch, folio))
2176 break;
2177 continue;
2178 unlock:
2179 folio_unlock(folio);
2180 put:
2181 folio_put(folio);
2182 }
2183 rcu_read_unlock();
2184
2185 return folio_batch_count(fbatch);
2186 }
2187
2188 /**
2189 * filemap_get_folios - Get a batch of folios
2190 * @mapping: The address_space to search
2191 * @start: The starting page index
2192 * @end: The final page index (inclusive)
2193 * @fbatch: The batch to fill.
2194 *
2195 * Search for and return a batch of folios in the mapping starting at
2196 * index @start and up to index @end (inclusive). The folios are returned
2197 * in @fbatch with an elevated reference count.
2198 *
2199 * Return: The number of folios which were found.
2200 * We also update @start to index the next folio for the traversal.
2201 */
filemap_get_folios(struct address_space * mapping,pgoff_t * start,pgoff_t end,struct folio_batch * fbatch)2202 unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start,
2203 pgoff_t end, struct folio_batch *fbatch)
2204 {
2205 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch);
2206 }
2207 EXPORT_SYMBOL(filemap_get_folios);
2208
2209 /**
2210 * filemap_get_folios_contig - Get a batch of contiguous folios
2211 * @mapping: The address_space to search
2212 * @start: The starting page index
2213 * @end: The final page index (inclusive)
2214 * @fbatch: The batch to fill
2215 *
2216 * filemap_get_folios_contig() works exactly like filemap_get_folios(),
2217 * except the returned folios are guaranteed to be contiguous. This may
2218 * not return all contiguous folios if the batch gets filled up.
2219 *
2220 * Return: The number of folios found.
2221 * Also update @start to be positioned for traversal of the next folio.
2222 */
2223
filemap_get_folios_contig(struct address_space * mapping,pgoff_t * start,pgoff_t end,struct folio_batch * fbatch)2224 unsigned filemap_get_folios_contig(struct address_space *mapping,
2225 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch)
2226 {
2227 XA_STATE(xas, &mapping->i_pages, *start);
2228 unsigned long nr;
2229 struct folio *folio;
2230
2231 rcu_read_lock();
2232
2233 for (folio = xas_load(&xas); folio && xas.xa_index <= end;
2234 folio = xas_next(&xas)) {
2235 if (xas_retry(&xas, folio))
2236 continue;
2237 /*
2238 * If the entry has been swapped out, we can stop looking.
2239 * No current caller is looking for DAX entries.
2240 */
2241 if (xa_is_value(folio))
2242 goto update_start;
2243
2244 /* If we landed in the middle of a THP, continue at its end. */
2245 if (xa_is_sibling(folio))
2246 goto update_start;
2247
2248 if (!folio_try_get(folio))
2249 goto retry;
2250
2251 if (unlikely(folio != xas_reload(&xas)))
2252 goto put_folio;
2253
2254 if (!folio_batch_add(fbatch, folio)) {
2255 nr = folio_nr_pages(folio);
2256 *start = folio->index + nr;
2257 goto out;
2258 }
2259 continue;
2260 put_folio:
2261 folio_put(folio);
2262
2263 retry:
2264 xas_reset(&xas);
2265 }
2266
2267 update_start:
2268 nr = folio_batch_count(fbatch);
2269
2270 if (nr) {
2271 folio = fbatch->folios[nr - 1];
2272 *start = folio_next_index(folio);
2273 }
2274 out:
2275 rcu_read_unlock();
2276 return folio_batch_count(fbatch);
2277 }
2278 EXPORT_SYMBOL(filemap_get_folios_contig);
2279
2280 /**
2281 * filemap_get_folios_tag - Get a batch of folios matching @tag
2282 * @mapping: The address_space to search
2283 * @start: The starting page index
2284 * @end: The final page index (inclusive)
2285 * @tag: The tag index
2286 * @fbatch: The batch to fill
2287 *
2288 * The first folio may start before @start; if it does, it will contain
2289 * @start. The final folio may extend beyond @end; if it does, it will
2290 * contain @end. The folios have ascending indices. There may be gaps
2291 * between the folios if there are indices which have no folio in the
2292 * page cache. If folios are added to or removed from the page cache
2293 * while this is running, they may or may not be found by this call.
2294 * Only returns folios that are tagged with @tag.
2295 *
2296 * Return: The number of folios found.
2297 * Also update @start to index the next folio for traversal.
2298 */
filemap_get_folios_tag(struct address_space * mapping,pgoff_t * start,pgoff_t end,xa_mark_t tag,struct folio_batch * fbatch)2299 unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start,
2300 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch)
2301 {
2302 XA_STATE(xas, &mapping->i_pages, *start);
2303 struct folio *folio;
2304
2305 rcu_read_lock();
2306 while ((folio = find_get_entry(&xas, end, tag)) != NULL) {
2307 /*
2308 * Shadow entries should never be tagged, but this iteration
2309 * is lockless so there is a window for page reclaim to evict
2310 * a page we saw tagged. Skip over it.
2311 */
2312 if (xa_is_value(folio))
2313 continue;
2314 if (!folio_batch_add(fbatch, folio)) {
2315 unsigned long nr = folio_nr_pages(folio);
2316 *start = folio->index + nr;
2317 goto out;
2318 }
2319 }
2320 /*
2321 * We come here when there is no page beyond @end. We take care to not
2322 * overflow the index @start as it confuses some of the callers. This
2323 * breaks the iteration when there is a page at index -1 but that is
2324 * already broke anyway.
2325 */
2326 if (end == (pgoff_t)-1)
2327 *start = (pgoff_t)-1;
2328 else
2329 *start = end + 1;
2330 out:
2331 rcu_read_unlock();
2332
2333 return folio_batch_count(fbatch);
2334 }
2335 EXPORT_SYMBOL(filemap_get_folios_tag);
2336
2337 /*
2338 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
2339 * a _large_ part of the i/o request. Imagine the worst scenario:
2340 *
2341 * ---R__________________________________________B__________
2342 * ^ reading here ^ bad block(assume 4k)
2343 *
2344 * read(R) => miss => readahead(R...B) => media error => frustrating retries
2345 * => failing the whole request => read(R) => read(R+1) =>
2346 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
2347 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
2348 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
2349 *
2350 * It is going insane. Fix it by quickly scaling down the readahead size.
2351 */
shrink_readahead_size_eio(struct file_ra_state * ra)2352 static void shrink_readahead_size_eio(struct file_ra_state *ra)
2353 {
2354 ra->ra_pages /= 4;
2355 }
2356
2357 /*
2358 * filemap_get_read_batch - Get a batch of folios for read
2359 *
2360 * Get a batch of folios which represent a contiguous range of bytes in
2361 * the file. No exceptional entries will be returned. If @index is in
2362 * the middle of a folio, the entire folio will be returned. The last
2363 * folio in the batch may have the readahead flag set or the uptodate flag
2364 * clear so that the caller can take the appropriate action.
2365 */
filemap_get_read_batch(struct address_space * mapping,pgoff_t index,pgoff_t max,struct folio_batch * fbatch)2366 static void filemap_get_read_batch(struct address_space *mapping,
2367 pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
2368 {
2369 XA_STATE(xas, &mapping->i_pages, index);
2370 struct folio *folio;
2371
2372 rcu_read_lock();
2373 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) {
2374 if (xas_retry(&xas, folio))
2375 continue;
2376 if (xas.xa_index > max || xa_is_value(folio))
2377 break;
2378 if (xa_is_sibling(folio))
2379 break;
2380 if (!folio_try_get(folio))
2381 goto retry;
2382
2383 if (unlikely(folio != xas_reload(&xas)))
2384 goto put_folio;
2385
2386 if (!folio_batch_add(fbatch, folio))
2387 break;
2388 if (!folio_test_uptodate(folio))
2389 break;
2390 if (folio_test_readahead(folio))
2391 break;
2392 xas_advance(&xas, folio_next_index(folio) - 1);
2393 continue;
2394 put_folio:
2395 folio_put(folio);
2396 retry:
2397 xas_reset(&xas);
2398 }
2399 rcu_read_unlock();
2400 }
2401
filemap_read_folio(struct file * file,filler_t filler,struct folio * folio)2402 static int filemap_read_folio(struct file *file, filler_t filler,
2403 struct folio *folio)
2404 {
2405 bool workingset = folio_test_workingset(folio);
2406 unsigned long pflags;
2407 int error;
2408
2409 /* Start the actual read. The read will unlock the page. */
2410 if (unlikely(workingset))
2411 psi_memstall_enter(&pflags);
2412 error = filler(file, folio);
2413 if (unlikely(workingset))
2414 psi_memstall_leave(&pflags);
2415 if (error)
2416 return error;
2417
2418 error = folio_wait_locked_killable(folio);
2419 if (error)
2420 return error;
2421 if (folio_test_uptodate(folio))
2422 return 0;
2423 if (file)
2424 shrink_readahead_size_eio(&file->f_ra);
2425 return -EIO;
2426 }
2427
filemap_range_uptodate(struct address_space * mapping,loff_t pos,size_t count,struct folio * folio,bool need_uptodate)2428 static bool filemap_range_uptodate(struct address_space *mapping,
2429 loff_t pos, size_t count, struct folio *folio,
2430 bool need_uptodate)
2431 {
2432 if (folio_test_uptodate(folio))
2433 return true;
2434 /* pipes can't handle partially uptodate pages */
2435 if (need_uptodate)
2436 return false;
2437 if (!mapping->a_ops->is_partially_uptodate)
2438 return false;
2439 if (mapping->host->i_blkbits >= folio_shift(folio))
2440 return false;
2441
2442 if (folio_pos(folio) > pos) {
2443 count -= folio_pos(folio) - pos;
2444 pos = 0;
2445 } else {
2446 pos -= folio_pos(folio);
2447 }
2448
2449 return mapping->a_ops->is_partially_uptodate(folio, pos, count);
2450 }
2451
filemap_update_page(struct kiocb * iocb,struct address_space * mapping,size_t count,struct folio * folio,bool need_uptodate)2452 static int filemap_update_page(struct kiocb *iocb,
2453 struct address_space *mapping, size_t count,
2454 struct folio *folio, bool need_uptodate)
2455 {
2456 int error;
2457
2458 if (iocb->ki_flags & IOCB_NOWAIT) {
2459 if (!filemap_invalidate_trylock_shared(mapping))
2460 return -EAGAIN;
2461 } else {
2462 filemap_invalidate_lock_shared(mapping);
2463 }
2464
2465 if (!folio_trylock(folio)) {
2466 error = -EAGAIN;
2467 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO))
2468 goto unlock_mapping;
2469 if (!(iocb->ki_flags & IOCB_WAITQ)) {
2470 filemap_invalidate_unlock_shared(mapping);
2471 /*
2472 * This is where we usually end up waiting for a
2473 * previously submitted readahead to finish.
2474 */
2475 folio_put_wait_locked(folio, TASK_KILLABLE);
2476 return AOP_TRUNCATED_PAGE;
2477 }
2478 error = __folio_lock_async(folio, iocb->ki_waitq);
2479 if (error)
2480 goto unlock_mapping;
2481 }
2482
2483 error = AOP_TRUNCATED_PAGE;
2484 if (!folio->mapping)
2485 goto unlock;
2486
2487 error = 0;
2488 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio,
2489 need_uptodate))
2490 goto unlock;
2491
2492 error = -EAGAIN;
2493 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ))
2494 goto unlock;
2495
2496 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2497 folio);
2498 goto unlock_mapping;
2499 unlock:
2500 folio_unlock(folio);
2501 unlock_mapping:
2502 filemap_invalidate_unlock_shared(mapping);
2503 if (error == AOP_TRUNCATED_PAGE)
2504 folio_put(folio);
2505 return error;
2506 }
2507
filemap_create_folio(struct kiocb * iocb,struct folio_batch * fbatch)2508 static int filemap_create_folio(struct kiocb *iocb, struct folio_batch *fbatch)
2509 {
2510 struct address_space *mapping = iocb->ki_filp->f_mapping;
2511 struct folio *folio;
2512 int error;
2513 unsigned int min_order = mapping_min_folio_order(mapping);
2514 pgoff_t index;
2515
2516 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ))
2517 return -EAGAIN;
2518
2519 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), min_order);
2520 if (!folio)
2521 return -ENOMEM;
2522 if (iocb->ki_flags & IOCB_DONTCACHE)
2523 __folio_set_dropbehind(folio);
2524
2525 /*
2526 * Protect against truncate / hole punch. Grabbing invalidate_lock
2527 * here assures we cannot instantiate and bring uptodate new
2528 * pagecache folios after evicting page cache during truncate
2529 * and before actually freeing blocks. Note that we could
2530 * release invalidate_lock after inserting the folio into
2531 * the page cache as the locked folio would then be enough to
2532 * synchronize with hole punching. But there are code paths
2533 * such as filemap_update_page() filling in partially uptodate
2534 * pages or ->readahead() that need to hold invalidate_lock
2535 * while mapping blocks for IO so let's hold the lock here as
2536 * well to keep locking rules simple.
2537 */
2538 filemap_invalidate_lock_shared(mapping);
2539 index = (iocb->ki_pos >> (PAGE_SHIFT + min_order)) << min_order;
2540 error = filemap_add_folio(mapping, folio, index,
2541 mapping_gfp_constraint(mapping, GFP_KERNEL));
2542 if (error == -EEXIST)
2543 error = AOP_TRUNCATED_PAGE;
2544 if (error)
2545 goto error;
2546
2547 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio,
2548 folio);
2549 if (error)
2550 goto error;
2551
2552 filemap_invalidate_unlock_shared(mapping);
2553 folio_batch_add(fbatch, folio);
2554 return 0;
2555 error:
2556 filemap_invalidate_unlock_shared(mapping);
2557 folio_put(folio);
2558 return error;
2559 }
2560
filemap_readahead(struct kiocb * iocb,struct file * file,struct address_space * mapping,struct folio * folio,pgoff_t last_index)2561 static int filemap_readahead(struct kiocb *iocb, struct file *file,
2562 struct address_space *mapping, struct folio *folio,
2563 pgoff_t last_index)
2564 {
2565 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index);
2566
2567 if (iocb->ki_flags & IOCB_NOIO)
2568 return -EAGAIN;
2569 if (iocb->ki_flags & IOCB_DONTCACHE)
2570 ractl.dropbehind = 1;
2571 page_cache_async_ra(&ractl, folio, last_index - folio->index);
2572 return 0;
2573 }
2574
filemap_get_pages(struct kiocb * iocb,size_t count,struct folio_batch * fbatch,bool need_uptodate)2575 static int filemap_get_pages(struct kiocb *iocb, size_t count,
2576 struct folio_batch *fbatch, bool need_uptodate)
2577 {
2578 struct file *filp = iocb->ki_filp;
2579 struct address_space *mapping = filp->f_mapping;
2580 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT;
2581 pgoff_t last_index;
2582 struct folio *folio;
2583 unsigned int flags;
2584 int err = 0;
2585
2586 /* "last_index" is the index of the page beyond the end of the read */
2587 last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
2588 retry:
2589 if (fatal_signal_pending(current))
2590 return -EINTR;
2591
2592 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2593 if (!folio_batch_count(fbatch)) {
2594 DEFINE_READAHEAD(ractl, filp, &filp->f_ra, mapping, index);
2595
2596 if (iocb->ki_flags & IOCB_NOIO)
2597 return -EAGAIN;
2598 if (iocb->ki_flags & IOCB_NOWAIT)
2599 flags = memalloc_noio_save();
2600 if (iocb->ki_flags & IOCB_DONTCACHE)
2601 ractl.dropbehind = 1;
2602 page_cache_sync_ra(&ractl, last_index - index);
2603 if (iocb->ki_flags & IOCB_NOWAIT)
2604 memalloc_noio_restore(flags);
2605 filemap_get_read_batch(mapping, index, last_index - 1, fbatch);
2606 }
2607 if (!folio_batch_count(fbatch)) {
2608 err = filemap_create_folio(iocb, fbatch);
2609 if (err == AOP_TRUNCATED_PAGE)
2610 goto retry;
2611 return err;
2612 }
2613
2614 folio = fbatch->folios[folio_batch_count(fbatch) - 1];
2615 if (folio_test_readahead(folio)) {
2616 err = filemap_readahead(iocb, filp, mapping, folio, last_index);
2617 if (err)
2618 goto err;
2619 }
2620 if (!folio_test_uptodate(folio)) {
2621 if ((iocb->ki_flags & IOCB_WAITQ) &&
2622 folio_batch_count(fbatch) > 1)
2623 iocb->ki_flags |= IOCB_NOWAIT;
2624 err = filemap_update_page(iocb, mapping, count, folio,
2625 need_uptodate);
2626 if (err)
2627 goto err;
2628 }
2629
2630 trace_mm_filemap_get_pages(mapping, index, last_index - 1);
2631 return 0;
2632 err:
2633 if (err < 0)
2634 folio_put(folio);
2635 if (likely(--fbatch->nr))
2636 return 0;
2637 if (err == AOP_TRUNCATED_PAGE)
2638 goto retry;
2639 return err;
2640 }
2641
pos_same_folio(loff_t pos1,loff_t pos2,struct folio * folio)2642 static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio)
2643 {
2644 unsigned int shift = folio_shift(folio);
2645
2646 return (pos1 >> shift == pos2 >> shift);
2647 }
2648
filemap_end_dropbehind_read(struct address_space * mapping,struct folio * folio)2649 static void filemap_end_dropbehind_read(struct address_space *mapping,
2650 struct folio *folio)
2651 {
2652 if (!folio_test_dropbehind(folio))
2653 return;
2654 if (folio_test_writeback(folio) || folio_test_dirty(folio))
2655 return;
2656 if (folio_trylock(folio)) {
2657 if (folio_test_clear_dropbehind(folio))
2658 folio_unmap_invalidate(mapping, folio, 0);
2659 folio_unlock(folio);
2660 }
2661 }
2662
2663 /**
2664 * filemap_read - Read data from the page cache.
2665 * @iocb: The iocb to read.
2666 * @iter: Destination for the data.
2667 * @already_read: Number of bytes already read by the caller.
2668 *
2669 * Copies data from the page cache. If the data is not currently present,
2670 * uses the readahead and read_folio address_space operations to fetch it.
2671 *
2672 * Return: Total number of bytes copied, including those already read by
2673 * the caller. If an error happens before any bytes are copied, returns
2674 * a negative error number.
2675 */
filemap_read(struct kiocb * iocb,struct iov_iter * iter,ssize_t already_read)2676 ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
2677 ssize_t already_read)
2678 {
2679 struct file *filp = iocb->ki_filp;
2680 struct file_ra_state *ra = &filp->f_ra;
2681 struct address_space *mapping = filp->f_mapping;
2682 struct inode *inode = mapping->host;
2683 struct folio_batch fbatch;
2684 int i, error = 0;
2685 bool writably_mapped;
2686 loff_t isize, end_offset;
2687 loff_t last_pos = ra->prev_pos;
2688
2689 if (unlikely(iocb->ki_pos < 0))
2690 return -EINVAL;
2691 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
2692 return 0;
2693 if (unlikely(!iov_iter_count(iter)))
2694 return 0;
2695
2696 iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos);
2697 folio_batch_init(&fbatch);
2698
2699 do {
2700 cond_resched();
2701
2702 /*
2703 * If we've already successfully copied some data, then we
2704 * can no longer safely return -EIOCBQUEUED. Hence mark
2705 * an async read NOWAIT at that point.
2706 */
2707 if ((iocb->ki_flags & IOCB_WAITQ) && already_read)
2708 iocb->ki_flags |= IOCB_NOWAIT;
2709
2710 if (unlikely(iocb->ki_pos >= i_size_read(inode)))
2711 break;
2712
2713 error = filemap_get_pages(iocb, iter->count, &fbatch, false);
2714 if (error < 0)
2715 break;
2716
2717 /*
2718 * i_size must be checked after we know the pages are Uptodate.
2719 *
2720 * Checking i_size after the check allows us to calculate
2721 * the correct value for "nr", which means the zero-filled
2722 * part of the page is not copied back to userspace (unless
2723 * another truncate extends the file - this is desired though).
2724 */
2725 isize = i_size_read(inode);
2726 if (unlikely(iocb->ki_pos >= isize))
2727 goto put_folios;
2728 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count);
2729
2730 /*
2731 * Once we start copying data, we don't want to be touching any
2732 * cachelines that might be contended:
2733 */
2734 writably_mapped = mapping_writably_mapped(mapping);
2735
2736 /*
2737 * When a read accesses the same folio several times, only
2738 * mark it as accessed the first time.
2739 */
2740 if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
2741 fbatch.folios[0]))
2742 folio_mark_accessed(fbatch.folios[0]);
2743
2744 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2745 struct folio *folio = fbatch.folios[i];
2746 size_t fsize = folio_size(folio);
2747 size_t offset = iocb->ki_pos & (fsize - 1);
2748 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos,
2749 fsize - offset);
2750 size_t copied;
2751
2752 if (end_offset < folio_pos(folio))
2753 break;
2754 if (i > 0)
2755 folio_mark_accessed(folio);
2756 /*
2757 * If users can be writing to this folio using arbitrary
2758 * virtual addresses, take care of potential aliasing
2759 * before reading the folio on the kernel side.
2760 */
2761 if (writably_mapped)
2762 flush_dcache_folio(folio);
2763
2764 copied = copy_folio_to_iter(folio, offset, bytes, iter);
2765
2766 already_read += copied;
2767 iocb->ki_pos += copied;
2768 last_pos = iocb->ki_pos;
2769
2770 if (copied < bytes) {
2771 error = -EFAULT;
2772 break;
2773 }
2774 }
2775 put_folios:
2776 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2777 struct folio *folio = fbatch.folios[i];
2778
2779 filemap_end_dropbehind_read(mapping, folio);
2780 folio_put(folio);
2781 }
2782 folio_batch_init(&fbatch);
2783 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
2784
2785 file_accessed(filp);
2786 ra->prev_pos = last_pos;
2787 return already_read ? already_read : error;
2788 }
2789 EXPORT_SYMBOL_GPL(filemap_read);
2790
kiocb_write_and_wait(struct kiocb * iocb,size_t count)2791 int kiocb_write_and_wait(struct kiocb *iocb, size_t count)
2792 {
2793 struct address_space *mapping = iocb->ki_filp->f_mapping;
2794 loff_t pos = iocb->ki_pos;
2795 loff_t end = pos + count - 1;
2796
2797 if (iocb->ki_flags & IOCB_NOWAIT) {
2798 if (filemap_range_needs_writeback(mapping, pos, end))
2799 return -EAGAIN;
2800 return 0;
2801 }
2802
2803 return filemap_write_and_wait_range(mapping, pos, end);
2804 }
2805 EXPORT_SYMBOL_GPL(kiocb_write_and_wait);
2806
filemap_invalidate_pages(struct address_space * mapping,loff_t pos,loff_t end,bool nowait)2807 int filemap_invalidate_pages(struct address_space *mapping,
2808 loff_t pos, loff_t end, bool nowait)
2809 {
2810 int ret;
2811
2812 if (nowait) {
2813 /* we could block if there are any pages in the range */
2814 if (filemap_range_has_page(mapping, pos, end))
2815 return -EAGAIN;
2816 } else {
2817 ret = filemap_write_and_wait_range(mapping, pos, end);
2818 if (ret)
2819 return ret;
2820 }
2821
2822 /*
2823 * After a write we want buffered reads to be sure to go to disk to get
2824 * the new data. We invalidate clean cached page from the region we're
2825 * about to write. We do this *before* the write so that we can return
2826 * without clobbering -EIOCBQUEUED from ->direct_IO().
2827 */
2828 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
2829 end >> PAGE_SHIFT);
2830 }
2831
kiocb_invalidate_pages(struct kiocb * iocb,size_t count)2832 int kiocb_invalidate_pages(struct kiocb *iocb, size_t count)
2833 {
2834 struct address_space *mapping = iocb->ki_filp->f_mapping;
2835
2836 return filemap_invalidate_pages(mapping, iocb->ki_pos,
2837 iocb->ki_pos + count - 1,
2838 iocb->ki_flags & IOCB_NOWAIT);
2839 }
2840 EXPORT_SYMBOL_GPL(kiocb_invalidate_pages);
2841
2842 /**
2843 * generic_file_read_iter - generic filesystem read routine
2844 * @iocb: kernel I/O control block
2845 * @iter: destination for the data read
2846 *
2847 * This is the "read_iter()" routine for all filesystems
2848 * that can use the page cache directly.
2849 *
2850 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall
2851 * be returned when no data can be read without waiting for I/O requests
2852 * to complete; it doesn't prevent readahead.
2853 *
2854 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O
2855 * requests shall be made for the read or for readahead. When no data
2856 * can be read, -EAGAIN shall be returned. When readahead would be
2857 * triggered, a partial, possibly empty read shall be returned.
2858 *
2859 * Return:
2860 * * number of bytes copied, even for partial reads
2861 * * negative error code (or 0 if IOCB_NOIO) if nothing was read
2862 */
2863 ssize_t
generic_file_read_iter(struct kiocb * iocb,struct iov_iter * iter)2864 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2865 {
2866 size_t count = iov_iter_count(iter);
2867 ssize_t retval = 0;
2868
2869 if (!count)
2870 return 0; /* skip atime */
2871
2872 if (iocb->ki_flags & IOCB_DIRECT) {
2873 struct file *file = iocb->ki_filp;
2874 struct address_space *mapping = file->f_mapping;
2875 struct inode *inode = mapping->host;
2876
2877 retval = kiocb_write_and_wait(iocb, count);
2878 if (retval < 0)
2879 return retval;
2880 file_accessed(file);
2881
2882 retval = mapping->a_ops->direct_IO(iocb, iter);
2883 if (retval >= 0) {
2884 iocb->ki_pos += retval;
2885 count -= retval;
2886 }
2887 if (retval != -EIOCBQUEUED)
2888 iov_iter_revert(iter, count - iov_iter_count(iter));
2889
2890 /*
2891 * Btrfs can have a short DIO read if we encounter
2892 * compressed extents, so if there was an error, or if
2893 * we've already read everything we wanted to, or if
2894 * there was a short read because we hit EOF, go ahead
2895 * and return. Otherwise fallthrough to buffered io for
2896 * the rest of the read. Buffered reads will not work for
2897 * DAX files, so don't bother trying.
2898 */
2899 if (retval < 0 || !count || IS_DAX(inode))
2900 return retval;
2901 if (iocb->ki_pos >= i_size_read(inode))
2902 return retval;
2903 }
2904
2905 return filemap_read(iocb, iter, retval);
2906 }
2907 EXPORT_SYMBOL(generic_file_read_iter);
2908
2909 /*
2910 * Splice subpages from a folio into a pipe.
2911 */
splice_folio_into_pipe(struct pipe_inode_info * pipe,struct folio * folio,loff_t fpos,size_t size)2912 size_t splice_folio_into_pipe(struct pipe_inode_info *pipe,
2913 struct folio *folio, loff_t fpos, size_t size)
2914 {
2915 struct page *page;
2916 size_t spliced = 0, offset = offset_in_folio(folio, fpos);
2917
2918 page = folio_page(folio, offset / PAGE_SIZE);
2919 size = min(size, folio_size(folio) - offset);
2920 offset %= PAGE_SIZE;
2921
2922 while (spliced < size && !pipe_is_full(pipe)) {
2923 struct pipe_buffer *buf = pipe_head_buf(pipe);
2924 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced);
2925
2926 *buf = (struct pipe_buffer) {
2927 .ops = &page_cache_pipe_buf_ops,
2928 .page = page,
2929 .offset = offset,
2930 .len = part,
2931 };
2932 folio_get(folio);
2933 pipe->head++;
2934 page++;
2935 spliced += part;
2936 offset = 0;
2937 }
2938
2939 return spliced;
2940 }
2941
2942 /**
2943 * filemap_splice_read - Splice data from a file's pagecache into a pipe
2944 * @in: The file to read from
2945 * @ppos: Pointer to the file position to read from
2946 * @pipe: The pipe to splice into
2947 * @len: The amount to splice
2948 * @flags: The SPLICE_F_* flags
2949 *
2950 * This function gets folios from a file's pagecache and splices them into the
2951 * pipe. Readahead will be called as necessary to fill more folios. This may
2952 * be used for blockdevs also.
2953 *
2954 * Return: On success, the number of bytes read will be returned and *@ppos
2955 * will be updated if appropriate; 0 will be returned if there is no more data
2956 * to be read; -EAGAIN will be returned if the pipe had no space, and some
2957 * other negative error code will be returned on error. A short read may occur
2958 * if the pipe has insufficient space, we reach the end of the data or we hit a
2959 * hole.
2960 */
filemap_splice_read(struct file * in,loff_t * ppos,struct pipe_inode_info * pipe,size_t len,unsigned int flags)2961 ssize_t filemap_splice_read(struct file *in, loff_t *ppos,
2962 struct pipe_inode_info *pipe,
2963 size_t len, unsigned int flags)
2964 {
2965 struct folio_batch fbatch;
2966 struct kiocb iocb;
2967 size_t total_spliced = 0, used, npages;
2968 loff_t isize, end_offset;
2969 bool writably_mapped;
2970 int i, error = 0;
2971
2972 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes))
2973 return 0;
2974
2975 init_sync_kiocb(&iocb, in);
2976 iocb.ki_pos = *ppos;
2977
2978 /* Work out how much data we can actually add into the pipe */
2979 used = pipe_buf_usage(pipe);
2980 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2981 len = min_t(size_t, len, npages * PAGE_SIZE);
2982
2983 folio_batch_init(&fbatch);
2984
2985 do {
2986 cond_resched();
2987
2988 if (*ppos >= i_size_read(in->f_mapping->host))
2989 break;
2990
2991 iocb.ki_pos = *ppos;
2992 error = filemap_get_pages(&iocb, len, &fbatch, true);
2993 if (error < 0)
2994 break;
2995
2996 /*
2997 * i_size must be checked after we know the pages are Uptodate.
2998 *
2999 * Checking i_size after the check allows us to calculate
3000 * the correct value for "nr", which means the zero-filled
3001 * part of the page is not copied back to userspace (unless
3002 * another truncate extends the file - this is desired though).
3003 */
3004 isize = i_size_read(in->f_mapping->host);
3005 if (unlikely(*ppos >= isize))
3006 break;
3007 end_offset = min_t(loff_t, isize, *ppos + len);
3008
3009 /*
3010 * Once we start copying data, we don't want to be touching any
3011 * cachelines that might be contended:
3012 */
3013 writably_mapped = mapping_writably_mapped(in->f_mapping);
3014
3015 for (i = 0; i < folio_batch_count(&fbatch); i++) {
3016 struct folio *folio = fbatch.folios[i];
3017 size_t n;
3018
3019 if (folio_pos(folio) >= end_offset)
3020 goto out;
3021 folio_mark_accessed(folio);
3022
3023 /*
3024 * If users can be writing to this folio using arbitrary
3025 * virtual addresses, take care of potential aliasing
3026 * before reading the folio on the kernel side.
3027 */
3028 if (writably_mapped)
3029 flush_dcache_folio(folio);
3030
3031 n = min_t(loff_t, len, isize - *ppos);
3032 n = splice_folio_into_pipe(pipe, folio, *ppos, n);
3033 if (!n)
3034 goto out;
3035 len -= n;
3036 total_spliced += n;
3037 *ppos += n;
3038 in->f_ra.prev_pos = *ppos;
3039 if (pipe_is_full(pipe))
3040 goto out;
3041 }
3042
3043 folio_batch_release(&fbatch);
3044 } while (len);
3045
3046 out:
3047 folio_batch_release(&fbatch);
3048 file_accessed(in);
3049
3050 return total_spliced ? total_spliced : error;
3051 }
3052 EXPORT_SYMBOL(filemap_splice_read);
3053
folio_seek_hole_data(struct xa_state * xas,struct address_space * mapping,struct folio * folio,loff_t start,loff_t end,bool seek_data)3054 static inline loff_t folio_seek_hole_data(struct xa_state *xas,
3055 struct address_space *mapping, struct folio *folio,
3056 loff_t start, loff_t end, bool seek_data)
3057 {
3058 const struct address_space_operations *ops = mapping->a_ops;
3059 size_t offset, bsz = i_blocksize(mapping->host);
3060
3061 if (xa_is_value(folio) || folio_test_uptodate(folio))
3062 return seek_data ? start : end;
3063 if (!ops->is_partially_uptodate)
3064 return seek_data ? end : start;
3065
3066 xas_pause(xas);
3067 rcu_read_unlock();
3068 folio_lock(folio);
3069 if (unlikely(folio->mapping != mapping))
3070 goto unlock;
3071
3072 offset = offset_in_folio(folio, start) & ~(bsz - 1);
3073
3074 do {
3075 if (ops->is_partially_uptodate(folio, offset, bsz) ==
3076 seek_data)
3077 break;
3078 start = (start + bsz) & ~((u64)bsz - 1);
3079 offset += bsz;
3080 } while (offset < folio_size(folio));
3081 unlock:
3082 folio_unlock(folio);
3083 rcu_read_lock();
3084 return start;
3085 }
3086
seek_folio_size(struct xa_state * xas,struct folio * folio)3087 static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio)
3088 {
3089 if (xa_is_value(folio))
3090 return PAGE_SIZE << xas_get_order(xas);
3091 return folio_size(folio);
3092 }
3093
3094 /**
3095 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3096 * @mapping: Address space to search.
3097 * @start: First byte to consider.
3098 * @end: Limit of search (exclusive).
3099 * @whence: Either SEEK_HOLE or SEEK_DATA.
3100 *
3101 * If the page cache knows which blocks contain holes and which blocks
3102 * contain data, your filesystem can use this function to implement
3103 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are
3104 * entirely memory-based such as tmpfs, and filesystems which support
3105 * unwritten extents.
3106 *
3107 * Return: The requested offset on success, or -ENXIO if @whence specifies
3108 * SEEK_DATA and there is no data after @start. There is an implicit hole
3109 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start
3110 * and @end contain data.
3111 */
mapping_seek_hole_data(struct address_space * mapping,loff_t start,loff_t end,int whence)3112 loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start,
3113 loff_t end, int whence)
3114 {
3115 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT);
3116 pgoff_t max = (end - 1) >> PAGE_SHIFT;
3117 bool seek_data = (whence == SEEK_DATA);
3118 struct folio *folio;
3119
3120 if (end <= start)
3121 return -ENXIO;
3122
3123 rcu_read_lock();
3124 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) {
3125 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT;
3126 size_t seek_size;
3127
3128 if (start < pos) {
3129 if (!seek_data)
3130 goto unlock;
3131 start = pos;
3132 }
3133
3134 seek_size = seek_folio_size(&xas, folio);
3135 pos = round_up((u64)pos + 1, seek_size);
3136 start = folio_seek_hole_data(&xas, mapping, folio, start, pos,
3137 seek_data);
3138 if (start < pos)
3139 goto unlock;
3140 if (start >= end)
3141 break;
3142 if (seek_size > PAGE_SIZE)
3143 xas_set(&xas, pos >> PAGE_SHIFT);
3144 if (!xa_is_value(folio))
3145 folio_put(folio);
3146 }
3147 if (seek_data)
3148 start = -ENXIO;
3149 unlock:
3150 rcu_read_unlock();
3151 if (folio && !xa_is_value(folio))
3152 folio_put(folio);
3153 if (start > end)
3154 return end;
3155 return start;
3156 }
3157
3158 #ifdef CONFIG_MMU
3159 #define MMAP_LOTSAMISS (100)
3160 /*
3161 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock
3162 * @vmf - the vm_fault for this fault.
3163 * @folio - the folio to lock.
3164 * @fpin - the pointer to the file we may pin (or is already pinned).
3165 *
3166 * This works similar to lock_folio_or_retry in that it can drop the
3167 * mmap_lock. It differs in that it actually returns the folio locked
3168 * if it returns 1 and 0 if it couldn't lock the folio. If we did have
3169 * to drop the mmap_lock then fpin will point to the pinned file and
3170 * needs to be fput()'ed at a later point.
3171 */
lock_folio_maybe_drop_mmap(struct vm_fault * vmf,struct folio * folio,struct file ** fpin)3172 static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
3173 struct file **fpin)
3174 {
3175 if (folio_trylock(folio))
3176 return 1;
3177
3178 /*
3179 * NOTE! This will make us return with VM_FAULT_RETRY, but with
3180 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT
3181 * is supposed to work. We have way too many special cases..
3182 */
3183 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
3184 return 0;
3185
3186 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin);
3187 if (vmf->flags & FAULT_FLAG_KILLABLE) {
3188 if (__folio_lock_killable(folio)) {
3189 /*
3190 * We didn't have the right flags to drop the
3191 * fault lock, but all fault_handlers only check
3192 * for fatal signals if we return VM_FAULT_RETRY,
3193 * so we need to drop the fault lock here and
3194 * return 0 if we don't have a fpin.
3195 */
3196 if (*fpin == NULL)
3197 release_fault_lock(vmf);
3198 return 0;
3199 }
3200 } else
3201 __folio_lock(folio);
3202
3203 return 1;
3204 }
3205
3206 /*
3207 * Synchronous readahead happens when we don't even find a page in the page
3208 * cache at all. We don't want to perform IO under the mmap sem, so if we have
3209 * to drop the mmap sem we return the file that was pinned in order for us to do
3210 * that. If we didn't pin a file then we return NULL. The file that is
3211 * returned needs to be fput()'ed when we're done with it.
3212 */
do_sync_mmap_readahead(struct vm_fault * vmf)3213 static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
3214 {
3215 struct file *file = vmf->vma->vm_file;
3216 struct file_ra_state *ra = &file->f_ra;
3217 struct address_space *mapping = file->f_mapping;
3218 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
3219 struct file *fpin = NULL;
3220 unsigned long vm_flags = vmf->vma->vm_flags;
3221 unsigned int mmap_miss;
3222
3223 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3224 /* Use the readahead code, even if readahead is disabled */
3225 if ((vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER) {
3226 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3227 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
3228 ra->size = HPAGE_PMD_NR;
3229 /*
3230 * Fetch two PMD folios, so we get the chance to actually
3231 * readahead, unless we've been told not to.
3232 */
3233 if (!(vm_flags & VM_RAND_READ))
3234 ra->size *= 2;
3235 ra->async_size = HPAGE_PMD_NR;
3236 page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
3237 return fpin;
3238 }
3239 #endif
3240
3241 /* If we don't want any read-ahead, don't bother */
3242 if (vm_flags & VM_RAND_READ)
3243 return fpin;
3244 if (!ra->ra_pages)
3245 return fpin;
3246
3247 if (vm_flags & VM_SEQ_READ) {
3248 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3249 page_cache_sync_ra(&ractl, ra->ra_pages);
3250 return fpin;
3251 }
3252
3253 /* Avoid banging the cache line if not needed */
3254 mmap_miss = READ_ONCE(ra->mmap_miss);
3255 if (mmap_miss < MMAP_LOTSAMISS * 10)
3256 WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
3257
3258 /*
3259 * Do we miss much more than hit in this file? If so,
3260 * stop bothering with read-ahead. It will only hurt.
3261 */
3262 if (mmap_miss > MMAP_LOTSAMISS)
3263 return fpin;
3264
3265 /*
3266 * mmap read-around
3267 */
3268 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3269 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
3270 ra->size = ra->ra_pages;
3271 ra->async_size = ra->ra_pages / 4;
3272 ractl._index = ra->start;
3273 page_cache_ra_order(&ractl, ra, 0);
3274 return fpin;
3275 }
3276
3277 /*
3278 * Asynchronous readahead happens when we find the page and PG_readahead,
3279 * so we want to possibly extend the readahead further. We return the file that
3280 * was pinned if we have to drop the mmap_lock in order to do IO.
3281 */
do_async_mmap_readahead(struct vm_fault * vmf,struct folio * folio)3282 static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
3283 struct folio *folio)
3284 {
3285 struct file *file = vmf->vma->vm_file;
3286 struct file_ra_state *ra = &file->f_ra;
3287 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
3288 struct file *fpin = NULL;
3289 unsigned int mmap_miss;
3290
3291 /* If we don't want any read-ahead, don't bother */
3292 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages)
3293 return fpin;
3294
3295 mmap_miss = READ_ONCE(ra->mmap_miss);
3296 if (mmap_miss)
3297 WRITE_ONCE(ra->mmap_miss, --mmap_miss);
3298
3299 if (folio_test_readahead(folio)) {
3300 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3301 page_cache_async_ra(&ractl, folio, ra->ra_pages);
3302 }
3303 return fpin;
3304 }
3305
filemap_fault_recheck_pte_none(struct vm_fault * vmf)3306 static vm_fault_t filemap_fault_recheck_pte_none(struct vm_fault *vmf)
3307 {
3308 struct vm_area_struct *vma = vmf->vma;
3309 vm_fault_t ret = 0;
3310 pte_t *ptep;
3311
3312 /*
3313 * We might have COW'ed a pagecache folio and might now have an mlocked
3314 * anon folio mapped. The original pagecache folio is not mlocked and
3315 * might have been evicted. During a read+clear/modify/write update of
3316 * the PTE, such as done in do_numa_page()/change_pte_range(), we
3317 * temporarily clear the PTE under PT lock and might detect it here as
3318 * "none" when not holding the PT lock.
3319 *
3320 * Not rechecking the PTE under PT lock could result in an unexpected
3321 * major fault in an mlock'ed region. Recheck only for this special
3322 * scenario while holding the PT lock, to not degrade non-mlocked
3323 * scenarios. Recheck the PTE without PT lock firstly, thereby reducing
3324 * the number of times we hold PT lock.
3325 */
3326 if (!(vma->vm_flags & VM_LOCKED))
3327 return 0;
3328
3329 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
3330 return 0;
3331
3332 ptep = pte_offset_map_ro_nolock(vma->vm_mm, vmf->pmd, vmf->address,
3333 &vmf->ptl);
3334 if (unlikely(!ptep))
3335 return VM_FAULT_NOPAGE;
3336
3337 if (unlikely(!pte_none(ptep_get_lockless(ptep)))) {
3338 ret = VM_FAULT_NOPAGE;
3339 } else {
3340 spin_lock(vmf->ptl);
3341 if (unlikely(!pte_none(ptep_get(ptep))))
3342 ret = VM_FAULT_NOPAGE;
3343 spin_unlock(vmf->ptl);
3344 }
3345 pte_unmap(ptep);
3346 return ret;
3347 }
3348
3349 /**
3350 * filemap_fault - read in file data for page fault handling
3351 * @vmf: struct vm_fault containing details of the fault
3352 *
3353 * filemap_fault() is invoked via the vma operations vector for a
3354 * mapped memory region to read in file data during a page fault.
3355 *
3356 * The goto's are kind of ugly, but this streamlines the normal case of having
3357 * it in the page cache, and handles the special cases reasonably without
3358 * having a lot of duplicated code.
3359 *
3360 * vma->vm_mm->mmap_lock must be held on entry.
3361 *
3362 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock
3363 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap().
3364 *
3365 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock
3366 * has not been released.
3367 *
3368 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
3369 *
3370 * Return: bitwise-OR of %VM_FAULT_ codes.
3371 */
filemap_fault(struct vm_fault * vmf)3372 vm_fault_t filemap_fault(struct vm_fault *vmf)
3373 {
3374 int error;
3375 struct file *file = vmf->vma->vm_file;
3376 struct file *fpin = NULL;
3377 struct address_space *mapping = file->f_mapping;
3378 struct inode *inode = mapping->host;
3379 pgoff_t max_idx, index = vmf->pgoff;
3380 struct folio *folio;
3381 vm_fault_t ret = 0;
3382 bool mapping_locked = false;
3383
3384 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3385 if (unlikely(index >= max_idx))
3386 return VM_FAULT_SIGBUS;
3387
3388 trace_mm_filemap_fault(mapping, index);
3389
3390 /*
3391 * Do we have something in the page cache already?
3392 */
3393 folio = filemap_get_folio(mapping, index);
3394 if (likely(!IS_ERR(folio))) {
3395 /*
3396 * We found the page, so try async readahead before waiting for
3397 * the lock.
3398 */
3399 if (!(vmf->flags & FAULT_FLAG_TRIED))
3400 fpin = do_async_mmap_readahead(vmf, folio);
3401 if (unlikely(!folio_test_uptodate(folio))) {
3402 filemap_invalidate_lock_shared(mapping);
3403 mapping_locked = true;
3404 }
3405 } else {
3406 ret = filemap_fault_recheck_pte_none(vmf);
3407 if (unlikely(ret))
3408 return ret;
3409
3410 /* No page in the page cache at all */
3411 count_vm_event(PGMAJFAULT);
3412 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
3413 ret = VM_FAULT_MAJOR;
3414 fpin = do_sync_mmap_readahead(vmf);
3415 retry_find:
3416 /*
3417 * See comment in filemap_create_folio() why we need
3418 * invalidate_lock
3419 */
3420 if (!mapping_locked) {
3421 filemap_invalidate_lock_shared(mapping);
3422 mapping_locked = true;
3423 }
3424 folio = __filemap_get_folio(mapping, index,
3425 FGP_CREAT|FGP_FOR_MMAP,
3426 vmf->gfp_mask);
3427 if (IS_ERR(folio)) {
3428 if (fpin)
3429 goto out_retry;
3430 filemap_invalidate_unlock_shared(mapping);
3431 return VM_FAULT_OOM;
3432 }
3433 }
3434
3435 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin))
3436 goto out_retry;
3437
3438 /* Did it get truncated? */
3439 if (unlikely(folio->mapping != mapping)) {
3440 folio_unlock(folio);
3441 folio_put(folio);
3442 goto retry_find;
3443 }
3444 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio);
3445
3446 /*
3447 * We have a locked folio in the page cache, now we need to check
3448 * that it's up-to-date. If not, it is going to be due to an error,
3449 * or because readahead was otherwise unable to retrieve it.
3450 */
3451 if (unlikely(!folio_test_uptodate(folio))) {
3452 /*
3453 * If the invalidate lock is not held, the folio was in cache
3454 * and uptodate and now it is not. Strange but possible since we
3455 * didn't hold the page lock all the time. Let's drop
3456 * everything, get the invalidate lock and try again.
3457 */
3458 if (!mapping_locked) {
3459 folio_unlock(folio);
3460 folio_put(folio);
3461 goto retry_find;
3462 }
3463
3464 /*
3465 * OK, the folio is really not uptodate. This can be because the
3466 * VMA has the VM_RAND_READ flag set, or because an error
3467 * arose. Let's read it in directly.
3468 */
3469 goto page_not_uptodate;
3470 }
3471
3472 /*
3473 * We've made it this far and we had to drop our mmap_lock, now is the
3474 * time to return to the upper layer and have it re-find the vma and
3475 * redo the fault.
3476 */
3477 if (fpin) {
3478 folio_unlock(folio);
3479 goto out_retry;
3480 }
3481 if (mapping_locked)
3482 filemap_invalidate_unlock_shared(mapping);
3483
3484 /*
3485 * Found the page and have a reference on it.
3486 * We must recheck i_size under page lock.
3487 */
3488 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
3489 if (unlikely(index >= max_idx)) {
3490 folio_unlock(folio);
3491 folio_put(folio);
3492 return VM_FAULT_SIGBUS;
3493 }
3494
3495 vmf->page = folio_file_page(folio, index);
3496 return ret | VM_FAULT_LOCKED;
3497
3498 page_not_uptodate:
3499 /*
3500 * Umm, take care of errors if the page isn't up-to-date.
3501 * Try to re-read it _once_. We do this synchronously,
3502 * because there really aren't any performance issues here
3503 * and we need to check for errors.
3504 */
3505 fpin = maybe_unlock_mmap_for_io(vmf, fpin);
3506 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio);
3507 if (fpin)
3508 goto out_retry;
3509 folio_put(folio);
3510
3511 if (!error || error == AOP_TRUNCATED_PAGE)
3512 goto retry_find;
3513 filemap_invalidate_unlock_shared(mapping);
3514
3515 return VM_FAULT_SIGBUS;
3516
3517 out_retry:
3518 /*
3519 * We dropped the mmap_lock, we need to return to the fault handler to
3520 * re-find the vma and come back and find our hopefully still populated
3521 * page.
3522 */
3523 if (!IS_ERR(folio))
3524 folio_put(folio);
3525 if (mapping_locked)
3526 filemap_invalidate_unlock_shared(mapping);
3527 if (fpin)
3528 fput(fpin);
3529 return ret | VM_FAULT_RETRY;
3530 }
3531 EXPORT_SYMBOL(filemap_fault);
3532
filemap_map_pmd(struct vm_fault * vmf,struct folio * folio,pgoff_t start)3533 static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio,
3534 pgoff_t start)
3535 {
3536 struct mm_struct *mm = vmf->vma->vm_mm;
3537
3538 /* Huge page is mapped? No need to proceed. */
3539 if (pmd_trans_huge(*vmf->pmd)) {
3540 folio_unlock(folio);
3541 folio_put(folio);
3542 return true;
3543 }
3544
3545 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) {
3546 struct page *page = folio_file_page(folio, start);
3547 vm_fault_t ret = do_set_pmd(vmf, page);
3548 if (!ret) {
3549 /* The page is mapped successfully, reference consumed. */
3550 folio_unlock(folio);
3551 return true;
3552 }
3553 }
3554
3555 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte)
3556 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
3557
3558 return false;
3559 }
3560
next_uptodate_folio(struct xa_state * xas,struct address_space * mapping,pgoff_t end_pgoff)3561 static struct folio *next_uptodate_folio(struct xa_state *xas,
3562 struct address_space *mapping, pgoff_t end_pgoff)
3563 {
3564 struct folio *folio = xas_next_entry(xas, end_pgoff);
3565 unsigned long max_idx;
3566
3567 do {
3568 if (!folio)
3569 return NULL;
3570 if (xas_retry(xas, folio))
3571 continue;
3572 if (xa_is_value(folio))
3573 continue;
3574 if (!folio_try_get(folio))
3575 continue;
3576 if (folio_test_locked(folio))
3577 goto skip;
3578 /* Has the page moved or been split? */
3579 if (unlikely(folio != xas_reload(xas)))
3580 goto skip;
3581 if (!folio_test_uptodate(folio) || folio_test_readahead(folio))
3582 goto skip;
3583 if (!folio_trylock(folio))
3584 goto skip;
3585 if (folio->mapping != mapping)
3586 goto unlock;
3587 if (!folio_test_uptodate(folio))
3588 goto unlock;
3589 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3590 if (xas->xa_index >= max_idx)
3591 goto unlock;
3592 return folio;
3593 unlock:
3594 folio_unlock(folio);
3595 skip:
3596 folio_put(folio);
3597 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL);
3598
3599 return NULL;
3600 }
3601
3602 /*
3603 * Map page range [start_page, start_page + nr_pages) of folio.
3604 * start_page is gotten from start by folio_page(folio, start)
3605 */
filemap_map_folio_range(struct vm_fault * vmf,struct folio * folio,unsigned long start,unsigned long addr,unsigned int nr_pages,unsigned long * rss,unsigned int * mmap_miss)3606 static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
3607 struct folio *folio, unsigned long start,
3608 unsigned long addr, unsigned int nr_pages,
3609 unsigned long *rss, unsigned int *mmap_miss)
3610 {
3611 vm_fault_t ret = 0;
3612 struct page *page = folio_page(folio, start);
3613 unsigned int count = 0;
3614 pte_t *old_ptep = vmf->pte;
3615
3616 do {
3617 if (PageHWPoison(page + count))
3618 goto skip;
3619
3620 /*
3621 * If there are too many folios that are recently evicted
3622 * in a file, they will probably continue to be evicted.
3623 * In such situation, read-ahead is only a waste of IO.
3624 * Don't decrease mmap_miss in this scenario to make sure
3625 * we can stop read-ahead.
3626 */
3627 if (!folio_test_workingset(folio))
3628 (*mmap_miss)++;
3629
3630 /*
3631 * NOTE: If there're PTE markers, we'll leave them to be
3632 * handled in the specific fault path, and it'll prohibit the
3633 * fault-around logic.
3634 */
3635 if (!pte_none(ptep_get(&vmf->pte[count])))
3636 goto skip;
3637
3638 count++;
3639 continue;
3640 skip:
3641 if (count) {
3642 set_pte_range(vmf, folio, page, count, addr);
3643 *rss += count;
3644 folio_ref_add(folio, count);
3645 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3646 ret = VM_FAULT_NOPAGE;
3647 }
3648
3649 count++;
3650 page += count;
3651 vmf->pte += count;
3652 addr += count * PAGE_SIZE;
3653 count = 0;
3654 } while (--nr_pages > 0);
3655
3656 if (count) {
3657 set_pte_range(vmf, folio, page, count, addr);
3658 *rss += count;
3659 folio_ref_add(folio, count);
3660 if (in_range(vmf->address, addr, count * PAGE_SIZE))
3661 ret = VM_FAULT_NOPAGE;
3662 }
3663
3664 vmf->pte = old_ptep;
3665
3666 return ret;
3667 }
3668
filemap_map_order0_folio(struct vm_fault * vmf,struct folio * folio,unsigned long addr,unsigned long * rss,unsigned int * mmap_miss)3669 static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
3670 struct folio *folio, unsigned long addr,
3671 unsigned long *rss, unsigned int *mmap_miss)
3672 {
3673 vm_fault_t ret = 0;
3674 struct page *page = &folio->page;
3675
3676 if (PageHWPoison(page))
3677 return ret;
3678
3679 /* See comment of filemap_map_folio_range() */
3680 if (!folio_test_workingset(folio))
3681 (*mmap_miss)++;
3682
3683 /*
3684 * NOTE: If there're PTE markers, we'll leave them to be
3685 * handled in the specific fault path, and it'll prohibit
3686 * the fault-around logic.
3687 */
3688 if (!pte_none(ptep_get(vmf->pte)))
3689 return ret;
3690
3691 if (vmf->address == addr)
3692 ret = VM_FAULT_NOPAGE;
3693
3694 set_pte_range(vmf, folio, page, 1, addr);
3695 (*rss)++;
3696 folio_ref_inc(folio);
3697
3698 return ret;
3699 }
3700
filemap_map_pages(struct vm_fault * vmf,pgoff_t start_pgoff,pgoff_t end_pgoff)3701 vm_fault_t filemap_map_pages(struct vm_fault *vmf,
3702 pgoff_t start_pgoff, pgoff_t end_pgoff)
3703 {
3704 struct vm_area_struct *vma = vmf->vma;
3705 struct file *file = vma->vm_file;
3706 struct address_space *mapping = file->f_mapping;
3707 pgoff_t file_end, last_pgoff = start_pgoff;
3708 unsigned long addr;
3709 XA_STATE(xas, &mapping->i_pages, start_pgoff);
3710 struct folio *folio;
3711 vm_fault_t ret = 0;
3712 unsigned long rss = 0;
3713 unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type;
3714
3715 rcu_read_lock();
3716 folio = next_uptodate_folio(&xas, mapping, end_pgoff);
3717 if (!folio)
3718 goto out;
3719
3720 if (filemap_map_pmd(vmf, folio, start_pgoff)) {
3721 ret = VM_FAULT_NOPAGE;
3722 goto out;
3723 }
3724
3725 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
3726 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
3727 if (!vmf->pte) {
3728 folio_unlock(folio);
3729 folio_put(folio);
3730 goto out;
3731 }
3732
3733 file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
3734 if (end_pgoff > file_end)
3735 end_pgoff = file_end;
3736
3737 folio_type = mm_counter_file(folio);
3738 do {
3739 unsigned long end;
3740
3741 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
3742 vmf->pte += xas.xa_index - last_pgoff;
3743 last_pgoff = xas.xa_index;
3744 end = folio_next_index(folio) - 1;
3745 nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
3746
3747 if (!folio_test_large(folio))
3748 ret |= filemap_map_order0_folio(vmf,
3749 folio, addr, &rss, &mmap_miss);
3750 else
3751 ret |= filemap_map_folio_range(vmf, folio,
3752 xas.xa_index - folio->index, addr,
3753 nr_pages, &rss, &mmap_miss);
3754
3755 folio_unlock(folio);
3756 folio_put(folio);
3757 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL);
3758 add_mm_counter(vma->vm_mm, folio_type, rss);
3759 pte_unmap_unlock(vmf->pte, vmf->ptl);
3760 trace_mm_filemap_map_pages(mapping, start_pgoff, end_pgoff);
3761 out:
3762 rcu_read_unlock();
3763
3764 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss);
3765 if (mmap_miss >= mmap_miss_saved)
3766 WRITE_ONCE(file->f_ra.mmap_miss, 0);
3767 else
3768 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss);
3769
3770 return ret;
3771 }
3772 EXPORT_SYMBOL(filemap_map_pages);
3773
filemap_page_mkwrite(struct vm_fault * vmf)3774 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3775 {
3776 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
3777 struct folio *folio = page_folio(vmf->page);
3778 vm_fault_t ret = VM_FAULT_LOCKED;
3779
3780 sb_start_pagefault(mapping->host->i_sb);
3781 file_update_time(vmf->vma->vm_file);
3782 folio_lock(folio);
3783 if (folio->mapping != mapping) {
3784 folio_unlock(folio);
3785 ret = VM_FAULT_NOPAGE;
3786 goto out;
3787 }
3788 /*
3789 * We mark the folio dirty already here so that when freeze is in
3790 * progress, we are guaranteed that writeback during freezing will
3791 * see the dirty folio and writeprotect it again.
3792 */
3793 folio_mark_dirty(folio);
3794 folio_wait_stable(folio);
3795 out:
3796 sb_end_pagefault(mapping->host->i_sb);
3797 return ret;
3798 }
3799
3800 const struct vm_operations_struct generic_file_vm_ops = {
3801 .fault = filemap_fault,
3802 .map_pages = filemap_map_pages,
3803 .page_mkwrite = filemap_page_mkwrite,
3804 };
3805
3806 /* This is used for a general mmap of a disk file */
3807
generic_file_mmap(struct file * file,struct vm_area_struct * vma)3808 int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3809 {
3810 struct address_space *mapping = file->f_mapping;
3811
3812 if (!mapping->a_ops->read_folio)
3813 return -ENOEXEC;
3814 file_accessed(file);
3815 vma->vm_ops = &generic_file_vm_ops;
3816 return 0;
3817 }
3818
3819 /*
3820 * This is for filesystems which do not implement ->writepage.
3821 */
generic_file_readonly_mmap(struct file * file,struct vm_area_struct * vma)3822 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3823 {
3824 if (vma_is_shared_maywrite(vma))
3825 return -EINVAL;
3826 return generic_file_mmap(file, vma);
3827 }
3828 #else
filemap_page_mkwrite(struct vm_fault * vmf)3829 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf)
3830 {
3831 return VM_FAULT_SIGBUS;
3832 }
generic_file_mmap(struct file * file,struct vm_area_struct * vma)3833 int generic_file_mmap(struct file *file, struct vm_area_struct *vma)
3834 {
3835 return -ENOSYS;
3836 }
generic_file_readonly_mmap(struct file * file,struct vm_area_struct * vma)3837 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
3838 {
3839 return -ENOSYS;
3840 }
3841 #endif /* CONFIG_MMU */
3842
3843 EXPORT_SYMBOL(filemap_page_mkwrite);
3844 EXPORT_SYMBOL(generic_file_mmap);
3845 EXPORT_SYMBOL(generic_file_readonly_mmap);
3846
do_read_cache_folio(struct address_space * mapping,pgoff_t index,filler_t filler,struct file * file,gfp_t gfp)3847 static struct folio *do_read_cache_folio(struct address_space *mapping,
3848 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp)
3849 {
3850 struct folio *folio;
3851 int err;
3852
3853 if (!filler)
3854 filler = mapping->a_ops->read_folio;
3855 repeat:
3856 folio = filemap_get_folio(mapping, index);
3857 if (IS_ERR(folio)) {
3858 folio = filemap_alloc_folio(gfp,
3859 mapping_min_folio_order(mapping));
3860 if (!folio)
3861 return ERR_PTR(-ENOMEM);
3862 index = mapping_align_index(mapping, index);
3863 err = filemap_add_folio(mapping, folio, index, gfp);
3864 if (unlikely(err)) {
3865 folio_put(folio);
3866 if (err == -EEXIST)
3867 goto repeat;
3868 /* Presumably ENOMEM for xarray node */
3869 return ERR_PTR(err);
3870 }
3871
3872 goto filler;
3873 }
3874 if (folio_test_uptodate(folio))
3875 goto out;
3876
3877 if (!folio_trylock(folio)) {
3878 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
3879 goto repeat;
3880 }
3881
3882 /* Folio was truncated from mapping */
3883 if (!folio->mapping) {
3884 folio_unlock(folio);
3885 folio_put(folio);
3886 goto repeat;
3887 }
3888
3889 /* Someone else locked and filled the page in a very small window */
3890 if (folio_test_uptodate(folio)) {
3891 folio_unlock(folio);
3892 goto out;
3893 }
3894
3895 filler:
3896 err = filemap_read_folio(file, filler, folio);
3897 if (err) {
3898 folio_put(folio);
3899 if (err == AOP_TRUNCATED_PAGE)
3900 goto repeat;
3901 return ERR_PTR(err);
3902 }
3903
3904 out:
3905 folio_mark_accessed(folio);
3906 return folio;
3907 }
3908
3909 /**
3910 * read_cache_folio - Read into page cache, fill it if needed.
3911 * @mapping: The address_space to read from.
3912 * @index: The index to read.
3913 * @filler: Function to perform the read, or NULL to use aops->read_folio().
3914 * @file: Passed to filler function, may be NULL if not required.
3915 *
3916 * Read one page into the page cache. If it succeeds, the folio returned
3917 * will contain @index, but it may not be the first page of the folio.
3918 *
3919 * If the filler function returns an error, it will be returned to the
3920 * caller.
3921 *
3922 * Context: May sleep. Expects mapping->invalidate_lock to be held.
3923 * Return: An uptodate folio on success, ERR_PTR() on failure.
3924 */
read_cache_folio(struct address_space * mapping,pgoff_t index,filler_t filler,struct file * file)3925 struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index,
3926 filler_t filler, struct file *file)
3927 {
3928 return do_read_cache_folio(mapping, index, filler, file,
3929 mapping_gfp_mask(mapping));
3930 }
3931 EXPORT_SYMBOL(read_cache_folio);
3932
3933 /**
3934 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags.
3935 * @mapping: The address_space for the folio.
3936 * @index: The index that the allocated folio will contain.
3937 * @gfp: The page allocator flags to use if allocating.
3938 *
3939 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with
3940 * any new memory allocations done using the specified allocation flags.
3941 *
3942 * The most likely error from this function is EIO, but ENOMEM is
3943 * possible and so is EINTR. If ->read_folio returns another error,
3944 * that will be returned to the caller.
3945 *
3946 * The function expects mapping->invalidate_lock to be already held.
3947 *
3948 * Return: Uptodate folio on success, ERR_PTR() on failure.
3949 */
mapping_read_folio_gfp(struct address_space * mapping,pgoff_t index,gfp_t gfp)3950 struct folio *mapping_read_folio_gfp(struct address_space *mapping,
3951 pgoff_t index, gfp_t gfp)
3952 {
3953 return do_read_cache_folio(mapping, index, NULL, NULL, gfp);
3954 }
3955 EXPORT_SYMBOL(mapping_read_folio_gfp);
3956
do_read_cache_page(struct address_space * mapping,pgoff_t index,filler_t * filler,struct file * file,gfp_t gfp)3957 static struct page *do_read_cache_page(struct address_space *mapping,
3958 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp)
3959 {
3960 struct folio *folio;
3961
3962 folio = do_read_cache_folio(mapping, index, filler, file, gfp);
3963 if (IS_ERR(folio))
3964 return &folio->page;
3965 return folio_file_page(folio, index);
3966 }
3967
read_cache_page(struct address_space * mapping,pgoff_t index,filler_t * filler,struct file * file)3968 struct page *read_cache_page(struct address_space *mapping,
3969 pgoff_t index, filler_t *filler, struct file *file)
3970 {
3971 return do_read_cache_page(mapping, index, filler, file,
3972 mapping_gfp_mask(mapping));
3973 }
3974 EXPORT_SYMBOL(read_cache_page);
3975
3976 /**
3977 * read_cache_page_gfp - read into page cache, using specified page allocation flags.
3978 * @mapping: the page's address_space
3979 * @index: the page index
3980 * @gfp: the page allocator flags to use if allocating
3981 *
3982 * This is the same as "read_mapping_page(mapping, index, NULL)", but with
3983 * any new page allocations done using the specified allocation flags.
3984 *
3985 * If the page does not get brought uptodate, return -EIO.
3986 *
3987 * The function expects mapping->invalidate_lock to be already held.
3988 *
3989 * Return: up to date page on success, ERR_PTR() on failure.
3990 */
read_cache_page_gfp(struct address_space * mapping,pgoff_t index,gfp_t gfp)3991 struct page *read_cache_page_gfp(struct address_space *mapping,
3992 pgoff_t index,
3993 gfp_t gfp)
3994 {
3995 return do_read_cache_page(mapping, index, NULL, NULL, gfp);
3996 }
3997 EXPORT_SYMBOL(read_cache_page_gfp);
3998
3999 /*
4000 * Warn about a page cache invalidation failure during a direct I/O write.
4001 */
dio_warn_stale_pagecache(struct file * filp)4002 static void dio_warn_stale_pagecache(struct file *filp)
4003 {
4004 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
4005 char pathname[128];
4006 char *path;
4007
4008 errseq_set(&filp->f_mapping->wb_err, -EIO);
4009 if (__ratelimit(&_rs)) {
4010 path = file_path(filp, pathname, sizeof(pathname));
4011 if (IS_ERR(path))
4012 path = "(unknown)";
4013 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
4014 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
4015 current->comm);
4016 }
4017 }
4018
kiocb_invalidate_post_direct_write(struct kiocb * iocb,size_t count)4019 void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count)
4020 {
4021 struct address_space *mapping = iocb->ki_filp->f_mapping;
4022
4023 if (mapping->nrpages &&
4024 invalidate_inode_pages2_range(mapping,
4025 iocb->ki_pos >> PAGE_SHIFT,
4026 (iocb->ki_pos + count - 1) >> PAGE_SHIFT))
4027 dio_warn_stale_pagecache(iocb->ki_filp);
4028 }
4029
4030 ssize_t
generic_file_direct_write(struct kiocb * iocb,struct iov_iter * from)4031 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from)
4032 {
4033 struct address_space *mapping = iocb->ki_filp->f_mapping;
4034 size_t write_len = iov_iter_count(from);
4035 ssize_t written;
4036
4037 /*
4038 * If a page can not be invalidated, return 0 to fall back
4039 * to buffered write.
4040 */
4041 written = kiocb_invalidate_pages(iocb, write_len);
4042 if (written) {
4043 if (written == -EBUSY)
4044 return 0;
4045 return written;
4046 }
4047
4048 written = mapping->a_ops->direct_IO(iocb, from);
4049
4050 /*
4051 * Finally, try again to invalidate clean pages which might have been
4052 * cached by non-direct readahead, or faulted in by get_user_pages()
4053 * if the source of the write was an mmap'ed region of the file
4054 * we're writing. Either one is a pretty crazy thing to do,
4055 * so we don't support it 100%. If this invalidation
4056 * fails, tough, the write still worked...
4057 *
4058 * Most of the time we do not need this since dio_complete() will do
4059 * the invalidation for us. However there are some file systems that
4060 * do not end up with dio_complete() being called, so let's not break
4061 * them by removing it completely.
4062 *
4063 * Noticeable example is a blkdev_direct_IO().
4064 *
4065 * Skip invalidation for async writes or if mapping has no pages.
4066 */
4067 if (written > 0) {
4068 struct inode *inode = mapping->host;
4069 loff_t pos = iocb->ki_pos;
4070
4071 kiocb_invalidate_post_direct_write(iocb, written);
4072 pos += written;
4073 write_len -= written;
4074 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
4075 i_size_write(inode, pos);
4076 mark_inode_dirty(inode);
4077 }
4078 iocb->ki_pos = pos;
4079 }
4080 if (written != -EIOCBQUEUED)
4081 iov_iter_revert(from, write_len - iov_iter_count(from));
4082 return written;
4083 }
4084 EXPORT_SYMBOL(generic_file_direct_write);
4085
generic_perform_write(struct kiocb * iocb,struct iov_iter * i)4086 ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
4087 {
4088 struct file *file = iocb->ki_filp;
4089 loff_t pos = iocb->ki_pos;
4090 struct address_space *mapping = file->f_mapping;
4091 const struct address_space_operations *a_ops = mapping->a_ops;
4092 size_t chunk = mapping_max_folio_size(mapping);
4093 long status = 0;
4094 ssize_t written = 0;
4095
4096 do {
4097 struct folio *folio;
4098 size_t offset; /* Offset into folio */
4099 size_t bytes; /* Bytes to write to folio */
4100 size_t copied; /* Bytes copied from user */
4101 void *fsdata = NULL;
4102
4103 bytes = iov_iter_count(i);
4104 retry:
4105 offset = pos & (chunk - 1);
4106 bytes = min(chunk - offset, bytes);
4107 balance_dirty_pages_ratelimited(mapping);
4108
4109 if (fatal_signal_pending(current)) {
4110 status = -EINTR;
4111 break;
4112 }
4113
4114 status = a_ops->write_begin(file, mapping, pos, bytes,
4115 &folio, &fsdata);
4116 if (unlikely(status < 0))
4117 break;
4118
4119 offset = offset_in_folio(folio, pos);
4120 if (bytes > folio_size(folio) - offset)
4121 bytes = folio_size(folio) - offset;
4122
4123 if (mapping_writably_mapped(mapping))
4124 flush_dcache_folio(folio);
4125
4126 /*
4127 * Faults here on mmap()s can recurse into arbitrary
4128 * filesystem code. Lots of locks are held that can
4129 * deadlock. Use an atomic copy to avoid deadlocking
4130 * in page fault handling.
4131 */
4132 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
4133 flush_dcache_folio(folio);
4134
4135 status = a_ops->write_end(file, mapping, pos, bytes, copied,
4136 folio, fsdata);
4137 if (unlikely(status != copied)) {
4138 iov_iter_revert(i, copied - max(status, 0L));
4139 if (unlikely(status < 0))
4140 break;
4141 }
4142 cond_resched();
4143
4144 if (unlikely(status == 0)) {
4145 /*
4146 * A short copy made ->write_end() reject the
4147 * thing entirely. Might be memory poisoning
4148 * halfway through, might be a race with munmap,
4149 * might be severe memory pressure.
4150 */
4151 if (chunk > PAGE_SIZE)
4152 chunk /= 2;
4153 if (copied) {
4154 bytes = copied;
4155 goto retry;
4156 }
4157
4158 /*
4159 * 'folio' is now unlocked and faults on it can be
4160 * handled. Ensure forward progress by trying to
4161 * fault it in now.
4162 */
4163 if (fault_in_iov_iter_readable(i, bytes) == bytes) {
4164 status = -EFAULT;
4165 break;
4166 }
4167 } else {
4168 pos += status;
4169 written += status;
4170 }
4171 } while (iov_iter_count(i));
4172
4173 if (!written)
4174 return status;
4175 iocb->ki_pos += written;
4176 return written;
4177 }
4178 EXPORT_SYMBOL(generic_perform_write);
4179
4180 /**
4181 * __generic_file_write_iter - write data to a file
4182 * @iocb: IO state structure (file, offset, etc.)
4183 * @from: iov_iter with data to write
4184 *
4185 * This function does all the work needed for actually writing data to a
4186 * file. It does all basic checks, removes SUID from the file, updates
4187 * modification times and calls proper subroutines depending on whether we
4188 * do direct IO or a standard buffered write.
4189 *
4190 * It expects i_rwsem to be grabbed unless we work on a block device or similar
4191 * object which does not need locking at all.
4192 *
4193 * This function does *not* take care of syncing data in case of O_SYNC write.
4194 * A caller has to handle it. This is mainly due to the fact that we want to
4195 * avoid syncing under i_rwsem.
4196 *
4197 * Return:
4198 * * number of bytes written, even for truncated writes
4199 * * negative error code if no data has been written at all
4200 */
__generic_file_write_iter(struct kiocb * iocb,struct iov_iter * from)4201 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4202 {
4203 struct file *file = iocb->ki_filp;
4204 struct address_space *mapping = file->f_mapping;
4205 struct inode *inode = mapping->host;
4206 ssize_t ret;
4207
4208 ret = file_remove_privs(file);
4209 if (ret)
4210 return ret;
4211
4212 ret = file_update_time(file);
4213 if (ret)
4214 return ret;
4215
4216 if (iocb->ki_flags & IOCB_DIRECT) {
4217 ret = generic_file_direct_write(iocb, from);
4218 /*
4219 * If the write stopped short of completing, fall back to
4220 * buffered writes. Some filesystems do this for writes to
4221 * holes, for example. For DAX files, a buffered write will
4222 * not succeed (even if it did, DAX does not handle dirty
4223 * page-cache pages correctly).
4224 */
4225 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode))
4226 return ret;
4227 return direct_write_fallback(iocb, from, ret,
4228 generic_perform_write(iocb, from));
4229 }
4230
4231 return generic_perform_write(iocb, from);
4232 }
4233 EXPORT_SYMBOL(__generic_file_write_iter);
4234
4235 /**
4236 * generic_file_write_iter - write data to a file
4237 * @iocb: IO state structure
4238 * @from: iov_iter with data to write
4239 *
4240 * This is a wrapper around __generic_file_write_iter() to be used by most
4241 * filesystems. It takes care of syncing the file in case of O_SYNC file
4242 * and acquires i_rwsem as needed.
4243 * Return:
4244 * * negative error code if no data has been written at all of
4245 * vfs_fsync_range() failed for a synchronous write
4246 * * number of bytes written, even for truncated writes
4247 */
generic_file_write_iter(struct kiocb * iocb,struct iov_iter * from)4248 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
4249 {
4250 struct file *file = iocb->ki_filp;
4251 struct inode *inode = file->f_mapping->host;
4252 ssize_t ret;
4253
4254 inode_lock(inode);
4255 ret = generic_write_checks(iocb, from);
4256 if (ret > 0)
4257 ret = __generic_file_write_iter(iocb, from);
4258 inode_unlock(inode);
4259
4260 if (ret > 0)
4261 ret = generic_write_sync(iocb, ret);
4262 return ret;
4263 }
4264 EXPORT_SYMBOL(generic_file_write_iter);
4265
4266 /**
4267 * filemap_release_folio() - Release fs-specific metadata on a folio.
4268 * @folio: The folio which the kernel is trying to free.
4269 * @gfp: Memory allocation flags (and I/O mode).
4270 *
4271 * The address_space is trying to release any data attached to a folio
4272 * (presumably at folio->private).
4273 *
4274 * This will also be called if the private_2 flag is set on a page,
4275 * indicating that the folio has other metadata associated with it.
4276 *
4277 * The @gfp argument specifies whether I/O may be performed to release
4278 * this page (__GFP_IO), and whether the call may block
4279 * (__GFP_RECLAIM & __GFP_FS).
4280 *
4281 * Return: %true if the release was successful, otherwise %false.
4282 */
filemap_release_folio(struct folio * folio,gfp_t gfp)4283 bool filemap_release_folio(struct folio *folio, gfp_t gfp)
4284 {
4285 struct address_space * const mapping = folio->mapping;
4286
4287 BUG_ON(!folio_test_locked(folio));
4288 if (!folio_needs_release(folio))
4289 return true;
4290 if (folio_test_writeback(folio))
4291 return false;
4292
4293 if (mapping && mapping->a_ops->release_folio)
4294 return mapping->a_ops->release_folio(folio, gfp);
4295 return try_to_free_buffers(folio);
4296 }
4297 EXPORT_SYMBOL(filemap_release_folio);
4298
4299 /**
4300 * filemap_invalidate_inode - Invalidate/forcibly write back a range of an inode's pagecache
4301 * @inode: The inode to flush
4302 * @flush: Set to write back rather than simply invalidate.
4303 * @start: First byte to in range.
4304 * @end: Last byte in range (inclusive), or LLONG_MAX for everything from start
4305 * onwards.
4306 *
4307 * Invalidate all the folios on an inode that contribute to the specified
4308 * range, possibly writing them back first. Whilst the operation is
4309 * undertaken, the invalidate lock is held to prevent new folios from being
4310 * installed.
4311 */
filemap_invalidate_inode(struct inode * inode,bool flush,loff_t start,loff_t end)4312 int filemap_invalidate_inode(struct inode *inode, bool flush,
4313 loff_t start, loff_t end)
4314 {
4315 struct address_space *mapping = inode->i_mapping;
4316 pgoff_t first = start >> PAGE_SHIFT;
4317 pgoff_t last = end >> PAGE_SHIFT;
4318 pgoff_t nr = end == LLONG_MAX ? ULONG_MAX : last - first + 1;
4319
4320 if (!mapping || !mapping->nrpages || end < start)
4321 goto out;
4322
4323 /* Prevent new folios from being added to the inode. */
4324 filemap_invalidate_lock(mapping);
4325
4326 if (!mapping->nrpages)
4327 goto unlock;
4328
4329 unmap_mapping_pages(mapping, first, nr, false);
4330
4331 /* Write back the data if we're asked to. */
4332 if (flush) {
4333 struct writeback_control wbc = {
4334 .sync_mode = WB_SYNC_ALL,
4335 .nr_to_write = LONG_MAX,
4336 .range_start = start,
4337 .range_end = end,
4338 };
4339
4340 filemap_fdatawrite_wbc(mapping, &wbc);
4341 }
4342
4343 /* Wait for writeback to complete on all folios and discard. */
4344 invalidate_inode_pages2_range(mapping, start / PAGE_SIZE, end / PAGE_SIZE);
4345
4346 unlock:
4347 filemap_invalidate_unlock(mapping);
4348 out:
4349 return filemap_check_errors(mapping);
4350 }
4351 EXPORT_SYMBOL_GPL(filemap_invalidate_inode);
4352
4353 #ifdef CONFIG_CACHESTAT_SYSCALL
4354 /**
4355 * filemap_cachestat() - compute the page cache statistics of a mapping
4356 * @mapping: The mapping to compute the statistics for.
4357 * @first_index: The starting page cache index.
4358 * @last_index: The final page index (inclusive).
4359 * @cs: the cachestat struct to write the result to.
4360 *
4361 * This will query the page cache statistics of a mapping in the
4362 * page range of [first_index, last_index] (inclusive). The statistics
4363 * queried include: number of dirty pages, number of pages marked for
4364 * writeback, and the number of (recently) evicted pages.
4365 */
filemap_cachestat(struct address_space * mapping,pgoff_t first_index,pgoff_t last_index,struct cachestat * cs)4366 static void filemap_cachestat(struct address_space *mapping,
4367 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs)
4368 {
4369 XA_STATE(xas, &mapping->i_pages, first_index);
4370 struct folio *folio;
4371
4372 /* Flush stats (and potentially sleep) outside the RCU read section. */
4373 mem_cgroup_flush_stats_ratelimited(NULL);
4374
4375 rcu_read_lock();
4376 xas_for_each(&xas, folio, last_index) {
4377 int order;
4378 unsigned long nr_pages;
4379 pgoff_t folio_first_index, folio_last_index;
4380
4381 /*
4382 * Don't deref the folio. It is not pinned, and might
4383 * get freed (and reused) underneath us.
4384 *
4385 * We *could* pin it, but that would be expensive for
4386 * what should be a fast and lightweight syscall.
4387 *
4388 * Instead, derive all information of interest from
4389 * the rcu-protected xarray.
4390 */
4391
4392 if (xas_retry(&xas, folio))
4393 continue;
4394
4395 order = xas_get_order(&xas);
4396 nr_pages = 1 << order;
4397 folio_first_index = round_down(xas.xa_index, 1 << order);
4398 folio_last_index = folio_first_index + nr_pages - 1;
4399
4400 /* Folios might straddle the range boundaries, only count covered pages */
4401 if (folio_first_index < first_index)
4402 nr_pages -= first_index - folio_first_index;
4403
4404 if (folio_last_index > last_index)
4405 nr_pages -= folio_last_index - last_index;
4406
4407 if (xa_is_value(folio)) {
4408 /* page is evicted */
4409 void *shadow = (void *)folio;
4410 bool workingset; /* not used */
4411
4412 cs->nr_evicted += nr_pages;
4413
4414 #ifdef CONFIG_SWAP /* implies CONFIG_MMU */
4415 if (shmem_mapping(mapping)) {
4416 /* shmem file - in swap cache */
4417 swp_entry_t swp = radix_to_swp_entry(folio);
4418
4419 /* swapin error results in poisoned entry */
4420 if (non_swap_entry(swp))
4421 goto resched;
4422
4423 /*
4424 * Getting a swap entry from the shmem
4425 * inode means we beat
4426 * shmem_unuse(). rcu_read_lock()
4427 * ensures swapoff waits for us before
4428 * freeing the swapper space. However,
4429 * we can race with swapping and
4430 * invalidation, so there might not be
4431 * a shadow in the swapcache (yet).
4432 */
4433 shadow = get_shadow_from_swap_cache(swp);
4434 if (!shadow)
4435 goto resched;
4436 }
4437 #endif
4438 if (workingset_test_recent(shadow, true, &workingset, false))
4439 cs->nr_recently_evicted += nr_pages;
4440
4441 goto resched;
4442 }
4443
4444 /* page is in cache */
4445 cs->nr_cache += nr_pages;
4446
4447 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY))
4448 cs->nr_dirty += nr_pages;
4449
4450 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK))
4451 cs->nr_writeback += nr_pages;
4452
4453 resched:
4454 if (need_resched()) {
4455 xas_pause(&xas);
4456 cond_resched_rcu();
4457 }
4458 }
4459 rcu_read_unlock();
4460 }
4461
4462 /*
4463 * See mincore: reveal pagecache information only for files
4464 * that the calling process has write access to, or could (if
4465 * tried) open for writing.
4466 */
can_do_cachestat(struct file * f)4467 static inline bool can_do_cachestat(struct file *f)
4468 {
4469 if (f->f_mode & FMODE_WRITE)
4470 return true;
4471 if (inode_owner_or_capable(file_mnt_idmap(f), file_inode(f)))
4472 return true;
4473 return file_permission(f, MAY_WRITE) == 0;
4474 }
4475
4476 /*
4477 * The cachestat(2) system call.
4478 *
4479 * cachestat() returns the page cache statistics of a file in the
4480 * bytes range specified by `off` and `len`: number of cached pages,
4481 * number of dirty pages, number of pages marked for writeback,
4482 * number of evicted pages, and number of recently evicted pages.
4483 *
4484 * An evicted page is a page that is previously in the page cache
4485 * but has been evicted since. A page is recently evicted if its last
4486 * eviction was recent enough that its reentry to the cache would
4487 * indicate that it is actively being used by the system, and that
4488 * there is memory pressure on the system.
4489 *
4490 * `off` and `len` must be non-negative integers. If `len` > 0,
4491 * the queried range is [`off`, `off` + `len`]. If `len` == 0,
4492 * we will query in the range from `off` to the end of the file.
4493 *
4494 * The `flags` argument is unused for now, but is included for future
4495 * extensibility. User should pass 0 (i.e no flag specified).
4496 *
4497 * Currently, hugetlbfs is not supported.
4498 *
4499 * Because the status of a page can change after cachestat() checks it
4500 * but before it returns to the application, the returned values may
4501 * contain stale information.
4502 *
4503 * return values:
4504 * zero - success
4505 * -EFAULT - cstat or cstat_range points to an illegal address
4506 * -EINVAL - invalid flags
4507 * -EBADF - invalid file descriptor
4508 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file
4509 */
SYSCALL_DEFINE4(cachestat,unsigned int,fd,struct cachestat_range __user *,cstat_range,struct cachestat __user *,cstat,unsigned int,flags)4510 SYSCALL_DEFINE4(cachestat, unsigned int, fd,
4511 struct cachestat_range __user *, cstat_range,
4512 struct cachestat __user *, cstat, unsigned int, flags)
4513 {
4514 CLASS(fd, f)(fd);
4515 struct address_space *mapping;
4516 struct cachestat_range csr;
4517 struct cachestat cs;
4518 pgoff_t first_index, last_index;
4519
4520 if (fd_empty(f))
4521 return -EBADF;
4522
4523 if (copy_from_user(&csr, cstat_range,
4524 sizeof(struct cachestat_range)))
4525 return -EFAULT;
4526
4527 /* hugetlbfs is not supported */
4528 if (is_file_hugepages(fd_file(f)))
4529 return -EOPNOTSUPP;
4530
4531 if (!can_do_cachestat(fd_file(f)))
4532 return -EPERM;
4533
4534 if (flags != 0)
4535 return -EINVAL;
4536
4537 first_index = csr.off >> PAGE_SHIFT;
4538 last_index =
4539 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT;
4540 memset(&cs, 0, sizeof(struct cachestat));
4541 mapping = fd_file(f)->f_mapping;
4542 filemap_cachestat(mapping, first_index, last_index, &cs);
4543
4544 if (copy_to_user(cstat, &cs, sizeof(struct cachestat)))
4545 return -EFAULT;
4546
4547 return 0;
4548 }
4549 #endif /* CONFIG_CACHESTAT_SYSCALL */
4550