1 // SPDX-License-Identifier: GPL-2.0-only
3 * mm/readahead.c - address_space-level file readahead.
5 * Copyright (C) 2002, Linus Torvalds
7 * 09Apr2002 Andrew Morton
11 #include <linux/kernel.h>
12 #include <linux/dax.h>
13 #include <linux/gfp.h>
14 #include <linux/export.h>
15 #include <linux/backing-dev.h>
16 #include <linux/task_io_accounting_ops.h>
17 #include <linux/pagevec.h>
18 #include <linux/pagemap.h>
19 #include <linux/syscalls.h>
20 #include <linux/file.h>
21 #include <linux/mm_inline.h>
22 #include <linux/blk-cgroup.h>
23 #include <linux/fadvise.h>
24 #include <linux/sched/mm.h>
29 * Initialise a struct file's readahead state. Assumes that the caller has
33 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
35 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
38 EXPORT_SYMBOL_GPL(file_ra_state_init);
41 * see if a page needs releasing upon read_cache_pages() failure
42 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
43 * before calling, such as the NFS fs marking pages that are cached locally
44 * on disk, thus we need to give the fs a chance to clean up in the event of
47 static void read_cache_pages_invalidate_page(struct address_space *mapping,
50 if (page_has_private(page)) {
51 if (!trylock_page(page))
53 page->mapping = mapping;
54 do_invalidatepage(page, 0, PAGE_SIZE);
62 * release a list of pages, invalidating them first if need be
64 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
65 struct list_head *pages)
69 while (!list_empty(pages)) {
70 victim = lru_to_page(pages);
71 list_del(&victim->lru);
72 read_cache_pages_invalidate_page(mapping, victim);
77 * read_cache_pages - populate an address space with some pages & start reads against them
78 * @mapping: the address_space
79 * @pages: The address of a list_head which contains the target pages. These
80 * pages have their ->index populated and are otherwise uninitialised.
81 * @filler: callback routine for filling a single page.
82 * @data: private data for the callback routine.
84 * Hides the details of the LRU cache etc from the filesystems.
86 * Returns: %0 on success, error return by @filler otherwise
88 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
89 int (*filler)(void *, struct page *), void *data)
94 while (!list_empty(pages)) {
95 page = lru_to_page(pages);
97 if (add_to_page_cache_lru(page, mapping, page->index,
98 readahead_gfp_mask(mapping))) {
99 read_cache_pages_invalidate_page(mapping, page);
104 ret = filler(data, page);
106 read_cache_pages_invalidate_pages(mapping, pages);
109 task_io_account_read(PAGE_SIZE);
114 EXPORT_SYMBOL(read_cache_pages);
116 static void read_pages(struct readahead_control *rac, struct list_head *pages,
119 const struct address_space_operations *aops = rac->mapping->a_ops;
121 struct blk_plug plug;
123 if (!readahead_count(rac))
126 blk_start_plug(&plug);
128 if (aops->readahead) {
129 aops->readahead(rac);
130 /* Clean up the remaining pages */
131 while ((page = readahead_page(rac))) {
135 } else if (aops->readpages) {
136 aops->readpages(rac->file, rac->mapping, pages,
137 readahead_count(rac));
138 /* Clean up the remaining pages */
139 put_pages_list(pages);
140 rac->_index += rac->_nr_pages;
143 while ((page = readahead_page(rac))) {
144 aops->readpage(rac->file, page);
149 blk_finish_plug(&plug);
151 BUG_ON(pages && !list_empty(pages));
152 BUG_ON(readahead_count(rac));
160 * page_cache_ra_unbounded - Start unchecked readahead.
161 * @ractl: Readahead control.
162 * @nr_to_read: The number of pages to read.
163 * @lookahead_size: Where to start the next readahead.
165 * This function is for filesystems to call when they want to start
166 * readahead beyond a file's stated i_size. This is almost certainly
167 * not the function you want to call. Use page_cache_async_readahead()
168 * or page_cache_sync_readahead() instead.
170 * Context: File is referenced by caller. Mutexes may be held by caller.
171 * May sleep, but will not reenter filesystem to reclaim memory.
173 void page_cache_ra_unbounded(struct readahead_control *ractl,
174 unsigned long nr_to_read, unsigned long lookahead_size)
176 struct address_space *mapping = ractl->mapping;
177 unsigned long index = readahead_index(ractl);
178 LIST_HEAD(page_pool);
179 gfp_t gfp_mask = readahead_gfp_mask(mapping);
183 * Partway through the readahead operation, we will have added
184 * locked pages to the page cache, but will not yet have submitted
185 * them for I/O. Adding another page may need to allocate memory,
186 * which can trigger memory reclaim. Telling the VM we're in
187 * the middle of a filesystem operation will cause it to not
188 * touch file-backed pages, preventing a deadlock. Most (all?)
189 * filesystems already specify __GFP_NOFS in their mapping's
190 * gfp_mask, but let's be explicit here.
192 unsigned int nofs = memalloc_nofs_save();
194 filemap_invalidate_lock_shared(mapping);
196 * Preallocate as many pages as we will need.
198 for (i = 0; i < nr_to_read; i++) {
199 struct folio *folio = xa_load(&mapping->i_pages, index + i);
201 if (folio && !xa_is_value(folio)) {
203 * Page already present? Kick off the current batch
204 * of contiguous pages before continuing with the
205 * next batch. This page may be the one we would
206 * have intended to mark as Readahead, but we don't
207 * have a stable reference to this page, and it's
208 * not worth getting one just for that.
210 read_pages(ractl, &page_pool, true);
211 i = ractl->_index + ractl->_nr_pages - index - 1;
215 folio = filemap_alloc_folio(gfp_mask, 0);
218 if (mapping->a_ops->readpages) {
219 folio->index = index + i;
220 list_add(&folio->lru, &page_pool);
221 } else if (filemap_add_folio(mapping, folio, index + i,
224 read_pages(ractl, &page_pool, true);
225 i = ractl->_index + ractl->_nr_pages - index - 1;
228 if (i == nr_to_read - lookahead_size)
229 folio_set_readahead(folio);
234 * Now start the IO. We ignore I/O errors - if the page is not
235 * uptodate then the caller will launch readpage again, and
236 * will then handle the error.
238 read_pages(ractl, &page_pool, false);
239 filemap_invalidate_unlock_shared(mapping);
240 memalloc_nofs_restore(nofs);
242 EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
245 * do_page_cache_ra() actually reads a chunk of disk. It allocates
246 * the pages first, then submits them for I/O. This avoids the very bad
247 * behaviour which would occur if page allocations are causing VM writeback.
248 * We really don't want to intermingle reads and writes like that.
250 void do_page_cache_ra(struct readahead_control *ractl,
251 unsigned long nr_to_read, unsigned long lookahead_size)
253 struct inode *inode = ractl->mapping->host;
254 unsigned long index = readahead_index(ractl);
255 loff_t isize = i_size_read(inode);
256 pgoff_t end_index; /* The last page we want to read */
261 end_index = (isize - 1) >> PAGE_SHIFT;
262 if (index > end_index)
264 /* Don't read past the page containing the last byte of the file */
265 if (nr_to_read > end_index - index)
266 nr_to_read = end_index - index + 1;
268 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
272 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
275 void force_page_cache_ra(struct readahead_control *ractl,
276 unsigned long nr_to_read)
278 struct address_space *mapping = ractl->mapping;
279 struct file_ra_state *ra = ractl->ra;
280 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
281 unsigned long max_pages, index;
283 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages &&
284 !mapping->a_ops->readahead))
288 * If the request exceeds the readahead window, allow the read to
289 * be up to the optimal hardware IO size
291 index = readahead_index(ractl);
292 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
293 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
295 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
297 if (this_chunk > nr_to_read)
298 this_chunk = nr_to_read;
299 ractl->_index = index;
300 do_page_cache_ra(ractl, this_chunk, 0);
303 nr_to_read -= this_chunk;
308 * Set the initial window size, round to next power of 2 and square
309 * for small size, x 4 for medium, and x 2 for large
310 * for 128k (32 page) max ra
311 * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
313 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
315 unsigned long newsize = roundup_pow_of_two(size);
317 if (newsize <= max / 32)
318 newsize = newsize * 4;
319 else if (newsize <= max / 4)
320 newsize = newsize * 2;
328 * Get the previous window size, ramp it up, and
329 * return it as the new window size.
331 static unsigned long get_next_ra_size(struct file_ra_state *ra,
334 unsigned long cur = ra->size;
344 * On-demand readahead design.
346 * The fields in struct file_ra_state represent the most-recently-executed
349 * |<----- async_size ---------|
350 * |------------------- size -------------------->|
351 * |==================#===========================|
352 * ^start ^page marked with PG_readahead
354 * To overlap application thinking time and disk I/O time, we do
355 * `readahead pipelining': Do not wait until the application consumed all
356 * readahead pages and stalled on the missing page at readahead_index;
357 * Instead, submit an asynchronous readahead I/O as soon as there are
358 * only async_size pages left in the readahead window. Normally async_size
359 * will be equal to size, for maximum pipelining.
361 * In interleaved sequential reads, concurrent streams on the same fd can
362 * be invalidating each other's readahead state. So we flag the new readahead
363 * page at (start+size-async_size) with PG_readahead, and use it as readahead
364 * indicator. The flag won't be set on already cached pages, to avoid the
365 * readahead-for-nothing fuss, saving pointless page cache lookups.
367 * prev_pos tracks the last visited byte in the _previous_ read request.
368 * It should be maintained by the caller, and will be used for detecting
369 * small random reads. Note that the readahead algorithm checks loosely
370 * for sequential patterns. Hence interleaved reads might be served as
373 * There is a special-case: if the first page which the application tries to
374 * read happens to be the first page of the file, it is assumed that a linear
375 * read is about to happen and the window is immediately set to the initial size
376 * based on I/O request size and the max_readahead.
378 * The code ramps up the readahead size aggressively at first, but slow down as
379 * it approaches max_readhead.
383 * Count contiguously cached pages from @index-1 to @index-@max,
384 * this count is a conservative estimation of
385 * - length of the sequential read sequence, or
386 * - thrashing threshold in memory tight systems
388 static pgoff_t count_history_pages(struct address_space *mapping,
389 pgoff_t index, unsigned long max)
394 head = page_cache_prev_miss(mapping, index - 1, max);
397 return index - 1 - head;
401 * page cache context based read-ahead
403 static int try_context_readahead(struct address_space *mapping,
404 struct file_ra_state *ra,
406 unsigned long req_size,
411 size = count_history_pages(mapping, index, max);
414 * not enough history pages:
415 * it could be a random read
417 if (size <= req_size)
421 * starts from beginning of file:
422 * it is a strong indication of long-run stream (or whole-file-read)
428 ra->size = min(size + req_size, max);
435 * There are some parts of the kernel which assume that PMD entries
436 * are exactly HPAGE_PMD_ORDER. Those should be fixed, but until then,
437 * limit the maximum allocation order to PMD size. I'm not aware of any
438 * assumptions about maximum order if THP are disabled, but 8 seems like
439 * a good order (that's 1MB if you're using 4kB pages)
441 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
442 #define MAX_PAGECACHE_ORDER HPAGE_PMD_ORDER
444 #define MAX_PAGECACHE_ORDER 8
447 static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
448 pgoff_t mark, unsigned int order, gfp_t gfp)
451 struct folio *folio = filemap_alloc_folio(gfp, order);
455 if (mark - index < (1UL << order))
456 folio_set_readahead(folio);
457 err = filemap_add_folio(ractl->mapping, folio, index, gfp);
461 ractl->_nr_pages += 1UL << order;
465 static void page_cache_ra_order(struct readahead_control *ractl,
466 struct file_ra_state *ra, unsigned int new_order)
468 struct address_space *mapping = ractl->mapping;
469 pgoff_t index = readahead_index(ractl);
470 pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
471 pgoff_t mark = index + ra->size - ra->async_size;
473 gfp_t gfp = readahead_gfp_mask(mapping);
475 if (!mapping_large_folio_support(mapping) || ra->size < 4)
478 limit = min(limit, index + ra->size - 1);
480 if (new_order < MAX_PAGECACHE_ORDER) {
482 if (new_order > MAX_PAGECACHE_ORDER)
483 new_order = MAX_PAGECACHE_ORDER;
484 while ((1 << new_order) > ra->size)
488 while (index <= limit) {
489 unsigned int order = new_order;
491 /* Align with smaller pages if needed */
492 if (index & ((1UL << order) - 1)) {
493 order = __ffs(index);
497 /* Don't allocate pages past EOF */
498 while (index + (1UL << order) - 1 > limit) {
502 err = ra_alloc_folio(ractl, index, mark, order, gfp);
505 index += 1UL << order;
509 ra->size += index - limit - 1;
510 ra->async_size += index - limit - 1;
513 read_pages(ractl, NULL, false);
516 * If there were already pages in the page cache, then we may have
517 * left some gaps. Let the regular readahead code take care of this
523 do_page_cache_ra(ractl, ra->size, ra->async_size);
527 * A minimal readahead algorithm for trivial sequential/random reads.
529 static void ondemand_readahead(struct readahead_control *ractl,
530 struct folio *folio, unsigned long req_size)
532 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
533 struct file_ra_state *ra = ractl->ra;
534 unsigned long max_pages = ra->ra_pages;
535 unsigned long add_pages;
536 unsigned long index = readahead_index(ractl);
540 * If the request exceeds the readahead window, allow the read to
541 * be up to the optimal hardware IO size
543 if (req_size > max_pages && bdi->io_pages > max_pages)
544 max_pages = min(req_size, bdi->io_pages);
550 goto initial_readahead;
553 * It's the expected callback index, assume sequential access.
554 * Ramp up sizes, and push forward the readahead window.
556 if ((index == (ra->start + ra->size - ra->async_size) ||
557 index == (ra->start + ra->size))) {
558 ra->start += ra->size;
559 ra->size = get_next_ra_size(ra, max_pages);
560 ra->async_size = ra->size;
565 * Hit a marked folio without valid readahead state.
566 * E.g. interleaved reads.
567 * Query the pagecache for async_size, which normally equals to
568 * readahead size. Ramp it up and use it as the new readahead size.
574 start = page_cache_next_miss(ractl->mapping, index + 1,
578 if (!start || start - index > max_pages)
582 ra->size = start - index; /* old async_size */
583 ra->size += req_size;
584 ra->size = get_next_ra_size(ra, max_pages);
585 ra->async_size = ra->size;
592 if (req_size > max_pages)
593 goto initial_readahead;
596 * sequential cache miss
597 * trivial case: (index - prev_index) == 1
598 * unaligned reads: (index - prev_index) == 0
600 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
601 if (index - prev_index <= 1UL)
602 goto initial_readahead;
605 * Query the page cache and look for the traces(cached history pages)
606 * that a sequential stream would leave behind.
608 if (try_context_readahead(ractl->mapping, ra, index, req_size,
613 * standalone, small random read
614 * Read as is, and do not pollute the readahead state.
616 do_page_cache_ra(ractl, req_size, 0);
621 ra->size = get_init_ra_size(req_size, max_pages);
622 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
626 * Will this read hit the readahead marker made by itself?
627 * If so, trigger the readahead marker hit now, and merge
628 * the resulted next readahead window into the current one.
629 * Take care of maximum IO pages as above.
631 if (index == ra->start && ra->size == ra->async_size) {
632 add_pages = get_next_ra_size(ra, max_pages);
633 if (ra->size + add_pages <= max_pages) {
634 ra->async_size = add_pages;
635 ra->size += add_pages;
637 ra->size = max_pages;
638 ra->async_size = max_pages >> 1;
642 ractl->_index = ra->start;
643 page_cache_ra_order(ractl, ra, folio ? folio_order(folio) : 0);
646 void page_cache_sync_ra(struct readahead_control *ractl,
647 unsigned long req_count)
649 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
652 * Even if read-ahead is disabled, issue this request as read-ahead
653 * as we'll need it to satisfy the requested range. The forced
654 * read-ahead will do the right thing and limit the read to just the
655 * requested range, which we'll set to 1 page for this case.
657 if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
666 force_page_cache_ra(ractl, req_count);
671 ondemand_readahead(ractl, NULL, req_count);
673 EXPORT_SYMBOL_GPL(page_cache_sync_ra);
675 void page_cache_async_ra(struct readahead_control *ractl,
676 struct folio *folio, unsigned long req_count)
679 if (!ractl->ra->ra_pages)
683 * Same bit is used for PG_readahead and PG_reclaim.
685 if (folio_test_writeback(folio))
688 folio_clear_readahead(folio);
691 * Defer asynchronous read-ahead on IO congestion.
693 if (inode_read_congested(ractl->mapping->host))
696 if (blk_cgroup_congested())
700 ondemand_readahead(ractl, folio, req_count);
702 EXPORT_SYMBOL_GPL(page_cache_async_ra);
704 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
711 if (!f.file || !(f.file->f_mode & FMODE_READ))
715 * The readahead() syscall is intended to run only on files
716 * that can execute readahead. If readahead is not possible
717 * on this file, then we must return -EINVAL.
720 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
721 !S_ISREG(file_inode(f.file)->i_mode))
724 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
730 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
732 return ksys_readahead(fd, offset, count);
736 * readahead_expand - Expand a readahead request
737 * @ractl: The request to be expanded
738 * @new_start: The revised start
739 * @new_len: The revised size of the request
741 * Attempt to expand a readahead request outwards from the current size to the
742 * specified size by inserting locked pages before and after the current window
743 * to increase the size to the new window. This may involve the insertion of
744 * THPs, in which case the window may get expanded even beyond what was
747 * The algorithm will stop if it encounters a conflicting page already in the
748 * pagecache and leave a smaller expansion than requested.
750 * The caller must check for this by examining the revised @ractl object for a
751 * different expansion than was requested.
753 void readahead_expand(struct readahead_control *ractl,
754 loff_t new_start, size_t new_len)
756 struct address_space *mapping = ractl->mapping;
757 struct file_ra_state *ra = ractl->ra;
758 pgoff_t new_index, new_nr_pages;
759 gfp_t gfp_mask = readahead_gfp_mask(mapping);
761 new_index = new_start / PAGE_SIZE;
763 /* Expand the leading edge downwards */
764 while (ractl->_index > new_index) {
765 unsigned long index = ractl->_index - 1;
766 struct page *page = xa_load(&mapping->i_pages, index);
768 if (page && !xa_is_value(page))
769 return; /* Page apparently present */
771 page = __page_cache_alloc(gfp_mask);
774 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
780 ractl->_index = page->index;
783 new_len += new_start - readahead_pos(ractl);
784 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
786 /* Expand the trailing edge upwards */
787 while (ractl->_nr_pages < new_nr_pages) {
788 unsigned long index = ractl->_index + ractl->_nr_pages;
789 struct page *page = xa_load(&mapping->i_pages, index);
791 if (page && !xa_is_value(page))
792 return; /* Page apparently present */
794 page = __page_cache_alloc(gfp_mask);
797 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
808 EXPORT_SYMBOL(readahead_expand);