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
12 * DOC: Readahead Overview
14 * Readahead is used to read content into the page cache before it is
15 * explicitly requested by the application. Readahead only ever
16 * attempts to read folios that are not yet in the page cache. If a
17 * folio is present but not up-to-date, readahead will not try to read
18 * it. In that case a simple ->read_folio() will be requested.
20 * Readahead is triggered when an application read request (whether a
21 * system call or a page fault) finds that the requested folio is not in
22 * the page cache, or that it is in the page cache and has the
23 * readahead flag set. This flag indicates that the folio was read
24 * as part of a previous readahead request and now that it has been
25 * accessed, it is time for the next readahead.
27 * Each readahead request is partly synchronous read, and partly async
28 * readahead. This is reflected in the struct file_ra_state which
29 * contains ->size being the total number of pages, and ->async_size
30 * which is the number of pages in the async section. The readahead
31 * flag will be set on the first folio in this async section to trigger
32 * a subsequent readahead. Once a series of sequential reads has been
33 * established, there should be no need for a synchronous component and
34 * all readahead request will be fully asynchronous.
36 * When either of the triggers causes a readahead, three numbers need
37 * to be determined: the start of the region to read, the size of the
38 * region, and the size of the async tail.
40 * The start of the region is simply the first page address at or after
41 * the accessed address, which is not currently populated in the page
42 * cache. This is found with a simple search in the page cache.
44 * The size of the async tail is determined by subtracting the size that
45 * was explicitly requested from the determined request size, unless
46 * this would be less than zero - then zero is used. NOTE THIS
47 * CALCULATION IS WRONG WHEN THE START OF THE REGION IS NOT THE ACCESSED
48 * PAGE. ALSO THIS CALCULATION IS NOT USED CONSISTENTLY.
50 * The size of the region is normally determined from the size of the
51 * previous readahead which loaded the preceding pages. This may be
52 * discovered from the struct file_ra_state for simple sequential reads,
53 * or from examining the state of the page cache when multiple
54 * sequential reads are interleaved. Specifically: where the readahead
55 * was triggered by the readahead flag, the size of the previous
56 * readahead is assumed to be the number of pages from the triggering
57 * page to the start of the new readahead. In these cases, the size of
58 * the previous readahead is scaled, often doubled, for the new
59 * readahead, though see get_next_ra_size() for details.
61 * If the size of the previous read cannot be determined, the number of
62 * preceding pages in the page cache is used to estimate the size of
63 * a previous read. This estimate could easily be misled by random
64 * reads being coincidentally adjacent, so it is ignored unless it is
65 * larger than the current request, and it is not scaled up, unless it
66 * is at the start of file.
68 * In general readahead is accelerated at the start of the file, as
69 * reads from there are often sequential. There are other minor
70 * adjustments to the readahead size in various special cases and these
71 * are best discovered by reading the code.
73 * The above calculation, based on the previous readahead size,
74 * determines the size of the readahead, to which any requested read
77 * Readahead requests are sent to the filesystem using the ->readahead()
78 * address space operation, for which mpage_readahead() is a canonical
79 * implementation. ->readahead() should normally initiate reads on all
80 * folios, but may fail to read any or all folios without causing an I/O
81 * error. The page cache reading code will issue a ->read_folio() request
82 * for any folio which ->readahead() did not read, and only an error
83 * from this will be final.
85 * ->readahead() will generally call readahead_folio() repeatedly to get
86 * each folio from those prepared for readahead. It may fail to read a
89 * * not calling readahead_folio() sufficiently many times, effectively
90 * ignoring some folios, as might be appropriate if the path to
91 * storage is congested.
93 * * failing to actually submit a read request for a given folio,
94 * possibly due to insufficient resources, or
96 * * getting an error during subsequent processing of a request.
98 * In the last two cases, the folio should be unlocked by the filesystem
99 * to indicate that the read attempt has failed. In the first case the
100 * folio will be unlocked by the VFS.
102 * Those folios not in the final ``async_size`` of the request should be
103 * considered to be important and ->readahead() should not fail them due
104 * to congestion or temporary resource unavailability, but should wait
105 * for necessary resources (e.g. memory or indexing information) to
106 * become available. Folios in the final ``async_size`` may be
107 * considered less urgent and failure to read them is more acceptable.
108 * In this case it is best to use filemap_remove_folio() to remove the
109 * folios from the page cache as is automatically done for folios that
110 * were not fetched with readahead_folio(). This will allow a
111 * subsequent synchronous readahead request to try them again. If they
112 * are left in the page cache, then they will be read individually using
113 * ->read_folio() which may be less efficient.
116 #include <linux/blkdev.h>
117 #include <linux/kernel.h>
118 #include <linux/dax.h>
119 #include <linux/gfp.h>
120 #include <linux/export.h>
121 #include <linux/backing-dev.h>
122 #include <linux/task_io_accounting_ops.h>
123 #include <linux/pagevec.h>
124 #include <linux/pagemap.h>
125 #include <linux/syscalls.h>
126 #include <linux/file.h>
127 #include <linux/mm_inline.h>
128 #include <linux/blk-cgroup.h>
129 #include <linux/fadvise.h>
130 #include <linux/sched/mm.h>
132 #include "internal.h"
135 * Initialise a struct file's readahead state. Assumes that the caller has
136 * memset *ra to zero.
139 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
141 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
144 EXPORT_SYMBOL_GPL(file_ra_state_init);
146 static void read_pages(struct readahead_control *rac)
148 const struct address_space_operations *aops = rac->mapping->a_ops;
150 struct blk_plug plug;
152 if (!readahead_count(rac))
155 blk_start_plug(&plug);
157 if (aops->readahead) {
158 aops->readahead(rac);
160 * Clean up the remaining folios. The sizes in ->ra
161 * may be used to size the next readahead, so make sure
162 * they accurately reflect what happened.
164 while ((folio = readahead_folio(rac)) != NULL) {
165 unsigned long nr = folio_nr_pages(folio);
169 if (rac->ra->async_size >= nr) {
170 rac->ra->async_size -= nr;
171 filemap_remove_folio(folio);
177 while ((folio = readahead_folio(rac)) != NULL)
178 aops->read_folio(rac->file, folio);
181 blk_finish_plug(&plug);
183 BUG_ON(readahead_count(rac));
187 * page_cache_ra_unbounded - Start unchecked readahead.
188 * @ractl: Readahead control.
189 * @nr_to_read: The number of pages to read.
190 * @lookahead_size: Where to start the next readahead.
192 * This function is for filesystems to call when they want to start
193 * readahead beyond a file's stated i_size. This is almost certainly
194 * not the function you want to call. Use page_cache_async_readahead()
195 * or page_cache_sync_readahead() instead.
197 * Context: File is referenced by caller. Mutexes may be held by caller.
198 * May sleep, but will not reenter filesystem to reclaim memory.
200 void page_cache_ra_unbounded(struct readahead_control *ractl,
201 unsigned long nr_to_read, unsigned long lookahead_size)
203 struct address_space *mapping = ractl->mapping;
204 unsigned long index = readahead_index(ractl);
205 gfp_t gfp_mask = readahead_gfp_mask(mapping);
209 * Partway through the readahead operation, we will have added
210 * locked pages to the page cache, but will not yet have submitted
211 * them for I/O. Adding another page may need to allocate memory,
212 * which can trigger memory reclaim. Telling the VM we're in
213 * the middle of a filesystem operation will cause it to not
214 * touch file-backed pages, preventing a deadlock. Most (all?)
215 * filesystems already specify __GFP_NOFS in their mapping's
216 * gfp_mask, but let's be explicit here.
218 unsigned int nofs = memalloc_nofs_save();
220 filemap_invalidate_lock_shared(mapping);
222 * Preallocate as many pages as we will need.
224 for (i = 0; i < nr_to_read; i++) {
225 struct folio *folio = xa_load(&mapping->i_pages, index + i);
227 if (folio && !xa_is_value(folio)) {
229 * Page already present? Kick off the current batch
230 * of contiguous pages before continuing with the
231 * next batch. This page may be the one we would
232 * have intended to mark as Readahead, but we don't
233 * have a stable reference to this page, and it's
234 * not worth getting one just for that.
238 i = ractl->_index + ractl->_nr_pages - index - 1;
242 folio = filemap_alloc_folio(gfp_mask, 0);
245 if (filemap_add_folio(mapping, folio, index + i,
250 i = ractl->_index + ractl->_nr_pages - index - 1;
253 if (i == nr_to_read - lookahead_size)
254 folio_set_readahead(folio);
259 * Now start the IO. We ignore I/O errors - if the folio is not
260 * uptodate then the caller will launch read_folio again, and
261 * will then handle the error.
264 filemap_invalidate_unlock_shared(mapping);
265 memalloc_nofs_restore(nofs);
267 EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
270 * do_page_cache_ra() actually reads a chunk of disk. It allocates
271 * the pages first, then submits them for I/O. This avoids the very bad
272 * behaviour which would occur if page allocations are causing VM writeback.
273 * We really don't want to intermingle reads and writes like that.
275 static void do_page_cache_ra(struct readahead_control *ractl,
276 unsigned long nr_to_read, unsigned long lookahead_size)
278 struct inode *inode = ractl->mapping->host;
279 unsigned long index = readahead_index(ractl);
280 loff_t isize = i_size_read(inode);
281 pgoff_t end_index; /* The last page we want to read */
286 end_index = (isize - 1) >> PAGE_SHIFT;
287 if (index > end_index)
289 /* Don't read past the page containing the last byte of the file */
290 if (nr_to_read > end_index - index)
291 nr_to_read = end_index - index + 1;
293 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
297 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
300 void force_page_cache_ra(struct readahead_control *ractl,
301 unsigned long nr_to_read)
303 struct address_space *mapping = ractl->mapping;
304 struct file_ra_state *ra = ractl->ra;
305 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
306 unsigned long max_pages, index;
308 if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead))
312 * If the request exceeds the readahead window, allow the read to
313 * be up to the optimal hardware IO size
315 index = readahead_index(ractl);
316 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
317 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
319 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
321 if (this_chunk > nr_to_read)
322 this_chunk = nr_to_read;
323 ractl->_index = index;
324 do_page_cache_ra(ractl, this_chunk, 0);
327 nr_to_read -= this_chunk;
332 * Set the initial window size, round to next power of 2 and square
333 * for small size, x 4 for medium, and x 2 for large
334 * for 128k (32 page) max ra
335 * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
337 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
339 unsigned long newsize = roundup_pow_of_two(size);
341 if (newsize <= max / 32)
342 newsize = newsize * 4;
343 else if (newsize <= max / 4)
344 newsize = newsize * 2;
352 * Get the previous window size, ramp it up, and
353 * return it as the new window size.
355 static unsigned long get_next_ra_size(struct file_ra_state *ra,
358 unsigned long cur = ra->size;
368 * On-demand readahead design.
370 * The fields in struct file_ra_state represent the most-recently-executed
373 * |<----- async_size ---------|
374 * |------------------- size -------------------->|
375 * |==================#===========================|
376 * ^start ^page marked with PG_readahead
378 * To overlap application thinking time and disk I/O time, we do
379 * `readahead pipelining': Do not wait until the application consumed all
380 * readahead pages and stalled on the missing page at readahead_index;
381 * Instead, submit an asynchronous readahead I/O as soon as there are
382 * only async_size pages left in the readahead window. Normally async_size
383 * will be equal to size, for maximum pipelining.
385 * In interleaved sequential reads, concurrent streams on the same fd can
386 * be invalidating each other's readahead state. So we flag the new readahead
387 * page at (start+size-async_size) with PG_readahead, and use it as readahead
388 * indicator. The flag won't be set on already cached pages, to avoid the
389 * readahead-for-nothing fuss, saving pointless page cache lookups.
391 * prev_pos tracks the last visited byte in the _previous_ read request.
392 * It should be maintained by the caller, and will be used for detecting
393 * small random reads. Note that the readahead algorithm checks loosely
394 * for sequential patterns. Hence interleaved reads might be served as
397 * There is a special-case: if the first page which the application tries to
398 * read happens to be the first page of the file, it is assumed that a linear
399 * read is about to happen and the window is immediately set to the initial size
400 * based on I/O request size and the max_readahead.
402 * The code ramps up the readahead size aggressively at first, but slow down as
403 * it approaches max_readhead.
407 * Count contiguously cached pages from @index-1 to @index-@max,
408 * this count is a conservative estimation of
409 * - length of the sequential read sequence, or
410 * - thrashing threshold in memory tight systems
412 static pgoff_t count_history_pages(struct address_space *mapping,
413 pgoff_t index, unsigned long max)
418 head = page_cache_prev_miss(mapping, index - 1, max);
421 return index - 1 - head;
425 * page cache context based readahead
427 static int try_context_readahead(struct address_space *mapping,
428 struct file_ra_state *ra,
430 unsigned long req_size,
435 size = count_history_pages(mapping, index, max);
438 * not enough history pages:
439 * it could be a random read
441 if (size <= req_size)
445 * starts from beginning of file:
446 * it is a strong indication of long-run stream (or whole-file-read)
452 ra->size = min(size + req_size, max);
459 * There are some parts of the kernel which assume that PMD entries
460 * are exactly HPAGE_PMD_ORDER. Those should be fixed, but until then,
461 * limit the maximum allocation order to PMD size. I'm not aware of any
462 * assumptions about maximum order if THP are disabled, but 8 seems like
463 * a good order (that's 1MB if you're using 4kB pages)
465 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
466 #define MAX_PAGECACHE_ORDER HPAGE_PMD_ORDER
468 #define MAX_PAGECACHE_ORDER 8
471 static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
472 pgoff_t mark, unsigned int order, gfp_t gfp)
475 struct folio *folio = filemap_alloc_folio(gfp, order);
479 mark = round_up(mark, 1UL << order);
481 folio_set_readahead(folio);
482 err = filemap_add_folio(ractl->mapping, folio, index, gfp);
486 ractl->_nr_pages += 1UL << order;
490 void page_cache_ra_order(struct readahead_control *ractl,
491 struct file_ra_state *ra, unsigned int new_order)
493 struct address_space *mapping = ractl->mapping;
494 pgoff_t index = readahead_index(ractl);
495 pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
496 pgoff_t mark = index + ra->size - ra->async_size;
498 gfp_t gfp = readahead_gfp_mask(mapping);
500 if (!mapping_large_folio_support(mapping) || ra->size < 4)
503 limit = min(limit, index + ra->size - 1);
505 if (new_order < MAX_PAGECACHE_ORDER) {
507 if (new_order > MAX_PAGECACHE_ORDER)
508 new_order = MAX_PAGECACHE_ORDER;
509 while ((1 << new_order) > ra->size)
513 filemap_invalidate_lock_shared(mapping);
514 while (index <= limit) {
515 unsigned int order = new_order;
517 /* Align with smaller pages if needed */
518 if (index & ((1UL << order) - 1)) {
519 order = __ffs(index);
523 /* Don't allocate pages past EOF */
524 while (index + (1UL << order) - 1 > limit) {
528 err = ra_alloc_folio(ractl, index, mark, order, gfp);
531 index += 1UL << order;
535 ra->size += index - limit - 1;
536 ra->async_size += index - limit - 1;
540 filemap_invalidate_unlock_shared(mapping);
543 * If there were already pages in the page cache, then we may have
544 * left some gaps. Let the regular readahead code take care of this
550 do_page_cache_ra(ractl, ra->size, ra->async_size);
554 * A minimal readahead algorithm for trivial sequential/random reads.
556 static void ondemand_readahead(struct readahead_control *ractl,
557 struct folio *folio, unsigned long req_size)
559 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
560 struct file_ra_state *ra = ractl->ra;
561 unsigned long max_pages = ra->ra_pages;
562 unsigned long add_pages;
563 pgoff_t index = readahead_index(ractl);
564 pgoff_t expected, prev_index;
565 unsigned int order = folio ? folio_order(folio) : 0;
568 * If the request exceeds the readahead window, allow the read to
569 * be up to the optimal hardware IO size
571 if (req_size > max_pages && bdi->io_pages > max_pages)
572 max_pages = min(req_size, bdi->io_pages);
578 goto initial_readahead;
581 * It's the expected callback index, assume sequential access.
582 * Ramp up sizes, and push forward the readahead window.
584 expected = round_up(ra->start + ra->size - ra->async_size,
586 if (index == expected || index == (ra->start + ra->size)) {
587 ra->start += ra->size;
588 ra->size = get_next_ra_size(ra, max_pages);
589 ra->async_size = ra->size;
594 * Hit a marked folio without valid readahead state.
595 * E.g. interleaved reads.
596 * Query the pagecache for async_size, which normally equals to
597 * readahead size. Ramp it up and use it as the new readahead size.
603 start = page_cache_next_miss(ractl->mapping, index + 1,
607 if (!start || start - index > max_pages)
611 ra->size = start - index; /* old async_size */
612 ra->size += req_size;
613 ra->size = get_next_ra_size(ra, max_pages);
614 ra->async_size = ra->size;
621 if (req_size > max_pages)
622 goto initial_readahead;
625 * sequential cache miss
626 * trivial case: (index - prev_index) == 1
627 * unaligned reads: (index - prev_index) == 0
629 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
630 if (index - prev_index <= 1UL)
631 goto initial_readahead;
634 * Query the page cache and look for the traces(cached history pages)
635 * that a sequential stream would leave behind.
637 if (try_context_readahead(ractl->mapping, ra, index, req_size,
642 * standalone, small random read
643 * Read as is, and do not pollute the readahead state.
645 do_page_cache_ra(ractl, req_size, 0);
650 ra->size = get_init_ra_size(req_size, max_pages);
651 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
655 * Will this read hit the readahead marker made by itself?
656 * If so, trigger the readahead marker hit now, and merge
657 * the resulted next readahead window into the current one.
658 * Take care of maximum IO pages as above.
660 if (index == ra->start && ra->size == ra->async_size) {
661 add_pages = get_next_ra_size(ra, max_pages);
662 if (ra->size + add_pages <= max_pages) {
663 ra->async_size = add_pages;
664 ra->size += add_pages;
666 ra->size = max_pages;
667 ra->async_size = max_pages >> 1;
671 ractl->_index = ra->start;
672 page_cache_ra_order(ractl, ra, order);
675 void page_cache_sync_ra(struct readahead_control *ractl,
676 unsigned long req_count)
678 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
681 * Even if readahead is disabled, issue this request as readahead
682 * as we'll need it to satisfy the requested range. The forced
683 * readahead will do the right thing and limit the read to just the
684 * requested range, which we'll set to 1 page for this case.
686 if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
695 force_page_cache_ra(ractl, req_count);
699 ondemand_readahead(ractl, NULL, req_count);
701 EXPORT_SYMBOL_GPL(page_cache_sync_ra);
703 void page_cache_async_ra(struct readahead_control *ractl,
704 struct folio *folio, unsigned long req_count)
707 if (!ractl->ra->ra_pages)
711 * Same bit is used for PG_readahead and PG_reclaim.
713 if (folio_test_writeback(folio))
716 folio_clear_readahead(folio);
718 if (blk_cgroup_congested())
721 ondemand_readahead(ractl, folio, req_count);
723 EXPORT_SYMBOL_GPL(page_cache_async_ra);
725 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
732 if (!f.file || !(f.file->f_mode & FMODE_READ))
736 * The readahead() syscall is intended to run only on files
737 * that can execute readahead. If readahead is not possible
738 * on this file, then we must return -EINVAL.
741 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
742 !S_ISREG(file_inode(f.file)->i_mode))
745 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
751 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
753 return ksys_readahead(fd, offset, count);
756 #if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD)
757 COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count)
759 return ksys_readahead(fd, compat_arg_u64_glue(offset), count);
764 * readahead_expand - Expand a readahead request
765 * @ractl: The request to be expanded
766 * @new_start: The revised start
767 * @new_len: The revised size of the request
769 * Attempt to expand a readahead request outwards from the current size to the
770 * specified size by inserting locked pages before and after the current window
771 * to increase the size to the new window. This may involve the insertion of
772 * THPs, in which case the window may get expanded even beyond what was
775 * The algorithm will stop if it encounters a conflicting page already in the
776 * pagecache and leave a smaller expansion than requested.
778 * The caller must check for this by examining the revised @ractl object for a
779 * different expansion than was requested.
781 void readahead_expand(struct readahead_control *ractl,
782 loff_t new_start, size_t new_len)
784 struct address_space *mapping = ractl->mapping;
785 struct file_ra_state *ra = ractl->ra;
786 pgoff_t new_index, new_nr_pages;
787 gfp_t gfp_mask = readahead_gfp_mask(mapping);
789 new_index = new_start / PAGE_SIZE;
791 /* Expand the leading edge downwards */
792 while (ractl->_index > new_index) {
793 unsigned long index = ractl->_index - 1;
794 struct page *page = xa_load(&mapping->i_pages, index);
796 if (page && !xa_is_value(page))
797 return; /* Page apparently present */
799 page = __page_cache_alloc(gfp_mask);
802 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
808 ractl->_index = page->index;
811 new_len += new_start - readahead_pos(ractl);
812 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
814 /* Expand the trailing edge upwards */
815 while (ractl->_nr_pages < new_nr_pages) {
816 unsigned long index = ractl->_index + ractl->_nr_pages;
817 struct page *page = xa_load(&mapping->i_pages, index);
819 if (page && !xa_is_value(page))
820 return; /* Page apparently present */
822 page = __page_cache_alloc(gfp_mask);
825 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
836 EXPORT_SYMBOL(readahead_expand);