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/blkdev.h>
16 #include <linux/backing-dev.h>
17 #include <linux/task_io_accounting_ops.h>
18 #include <linux/pagevec.h>
19 #include <linux/pagemap.h>
20 #include <linux/syscalls.h>
21 #include <linux/file.h>
22 #include <linux/mm_inline.h>
23 #include <linux/blk-cgroup.h>
24 #include <linux/fadvise.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 int read_pages(struct address_space *mapping, struct file *filp,
117 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
119 struct blk_plug plug;
123 blk_start_plug(&plug);
125 if (mapping->a_ops->readpages) {
126 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
127 /* Clean up the remaining pages */
128 put_pages_list(pages);
132 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
133 struct page *page = lru_to_page(pages);
134 list_del(&page->lru);
135 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
136 mapping->a_ops->readpage(filp, page);
142 blk_finish_plug(&plug);
148 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates
149 * the pages first, then submits them for I/O. This avoids the very bad
150 * behaviour which would occur if page allocations are causing VM writeback.
151 * We really don't want to intermingle reads and writes like that.
153 * Returns the number of pages requested, or the maximum amount of I/O allowed.
155 unsigned int __do_page_cache_readahead(struct address_space *mapping,
156 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
157 unsigned long lookahead_size)
159 struct inode *inode = mapping->host;
161 unsigned long end_index; /* The last page we want to read */
162 LIST_HEAD(page_pool);
164 unsigned int nr_pages = 0;
165 loff_t isize = i_size_read(inode);
166 gfp_t gfp_mask = readahead_gfp_mask(mapping);
171 end_index = ((isize - 1) >> PAGE_SHIFT);
174 * Preallocate as many pages as we will need.
176 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
177 pgoff_t page_offset = offset + page_idx;
179 if (page_offset > end_index)
182 page = xa_load(&mapping->i_pages, page_offset);
183 if (page && !xa_is_value(page)) {
185 * Page already present? Kick off the current batch of
186 * contiguous pages before continuing with the next
190 read_pages(mapping, filp, &page_pool, nr_pages,
196 page = __page_cache_alloc(gfp_mask);
199 page->index = page_offset;
200 list_add(&page->lru, &page_pool);
201 if (page_idx == nr_to_read - lookahead_size)
202 SetPageReadahead(page);
207 * Now start the IO. We ignore I/O errors - if the page is not
208 * uptodate then the caller will launch readpage again, and
209 * will then handle the error.
212 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
213 BUG_ON(!list_empty(&page_pool));
219 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
222 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
223 pgoff_t offset, unsigned long nr_to_read)
225 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
226 struct file_ra_state *ra = &filp->f_ra;
227 unsigned long max_pages;
229 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
233 * If the request exceeds the readahead window, allow the read to
234 * be up to the optimal hardware IO size
236 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
237 nr_to_read = min(nr_to_read, max_pages);
239 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
241 if (this_chunk > nr_to_read)
242 this_chunk = nr_to_read;
243 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
245 offset += this_chunk;
246 nr_to_read -= this_chunk;
252 * Set the initial window size, round to next power of 2 and square
253 * for small size, x 4 for medium, and x 2 for large
254 * for 128k (32 page) max ra
255 * 1-8 page = 32k initial, > 8 page = 128k initial
257 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
259 unsigned long newsize = roundup_pow_of_two(size);
261 if (newsize <= max / 32)
262 newsize = newsize * 4;
263 else if (newsize <= max / 4)
264 newsize = newsize * 2;
272 * Get the previous window size, ramp it up, and
273 * return it as the new window size.
275 static unsigned long get_next_ra_size(struct file_ra_state *ra,
278 unsigned long cur = ra->size;
288 * On-demand readahead design.
290 * The fields in struct file_ra_state represent the most-recently-executed
293 * |<----- async_size ---------|
294 * |------------------- size -------------------->|
295 * |==================#===========================|
296 * ^start ^page marked with PG_readahead
298 * To overlap application thinking time and disk I/O time, we do
299 * `readahead pipelining': Do not wait until the application consumed all
300 * readahead pages and stalled on the missing page at readahead_index;
301 * Instead, submit an asynchronous readahead I/O as soon as there are
302 * only async_size pages left in the readahead window. Normally async_size
303 * will be equal to size, for maximum pipelining.
305 * In interleaved sequential reads, concurrent streams on the same fd can
306 * be invalidating each other's readahead state. So we flag the new readahead
307 * page at (start+size-async_size) with PG_readahead, and use it as readahead
308 * indicator. The flag won't be set on already cached pages, to avoid the
309 * readahead-for-nothing fuss, saving pointless page cache lookups.
311 * prev_pos tracks the last visited byte in the _previous_ read request.
312 * It should be maintained by the caller, and will be used for detecting
313 * small random reads. Note that the readahead algorithm checks loosely
314 * for sequential patterns. Hence interleaved reads might be served as
317 * There is a special-case: if the first page which the application tries to
318 * read happens to be the first page of the file, it is assumed that a linear
319 * read is about to happen and the window is immediately set to the initial size
320 * based on I/O request size and the max_readahead.
322 * The code ramps up the readahead size aggressively at first, but slow down as
323 * it approaches max_readhead.
327 * Count contiguously cached pages from @offset-1 to @offset-@max,
328 * this count is a conservative estimation of
329 * - length of the sequential read sequence, or
330 * - thrashing threshold in memory tight systems
332 static pgoff_t count_history_pages(struct address_space *mapping,
333 pgoff_t offset, unsigned long max)
338 head = page_cache_prev_miss(mapping, offset - 1, max);
341 return offset - 1 - head;
345 * page cache context based read-ahead
347 static int try_context_readahead(struct address_space *mapping,
348 struct file_ra_state *ra,
350 unsigned long req_size,
355 size = count_history_pages(mapping, offset, max);
358 * not enough history pages:
359 * it could be a random read
361 if (size <= req_size)
365 * starts from beginning of file:
366 * it is a strong indication of long-run stream (or whole-file-read)
372 ra->size = min(size + req_size, max);
379 * A minimal readahead algorithm for trivial sequential/random reads.
382 ondemand_readahead(struct address_space *mapping,
383 struct file_ra_state *ra, struct file *filp,
384 bool hit_readahead_marker, pgoff_t offset,
385 unsigned long req_size)
387 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
388 unsigned long max_pages = ra->ra_pages;
389 unsigned long add_pages;
393 * If the request exceeds the readahead window, allow the read to
394 * be up to the optimal hardware IO size
396 if (req_size > max_pages && bdi->io_pages > max_pages)
397 max_pages = min(req_size, bdi->io_pages);
403 goto initial_readahead;
406 * It's the expected callback offset, assume sequential access.
407 * Ramp up sizes, and push forward the readahead window.
409 if ((offset == (ra->start + ra->size - ra->async_size) ||
410 offset == (ra->start + ra->size))) {
411 ra->start += ra->size;
412 ra->size = get_next_ra_size(ra, max_pages);
413 ra->async_size = ra->size;
418 * Hit a marked page without valid readahead state.
419 * E.g. interleaved reads.
420 * Query the pagecache for async_size, which normally equals to
421 * readahead size. Ramp it up and use it as the new readahead size.
423 if (hit_readahead_marker) {
427 start = page_cache_next_miss(mapping, offset + 1, max_pages);
430 if (!start || start - offset > max_pages)
434 ra->size = start - offset; /* old async_size */
435 ra->size += req_size;
436 ra->size = get_next_ra_size(ra, max_pages);
437 ra->async_size = ra->size;
444 if (req_size > max_pages)
445 goto initial_readahead;
448 * sequential cache miss
449 * trivial case: (offset - prev_offset) == 1
450 * unaligned reads: (offset - prev_offset) == 0
452 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
453 if (offset - prev_offset <= 1UL)
454 goto initial_readahead;
457 * Query the page cache and look for the traces(cached history pages)
458 * that a sequential stream would leave behind.
460 if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
464 * standalone, small random read
465 * Read as is, and do not pollute the readahead state.
467 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
471 ra->size = get_init_ra_size(req_size, max_pages);
472 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
476 * Will this read hit the readahead marker made by itself?
477 * If so, trigger the readahead marker hit now, and merge
478 * the resulted next readahead window into the current one.
479 * Take care of maximum IO pages as above.
481 if (offset == ra->start && ra->size == ra->async_size) {
482 add_pages = get_next_ra_size(ra, max_pages);
483 if (ra->size + add_pages <= max_pages) {
484 ra->async_size = add_pages;
485 ra->size += add_pages;
487 ra->size = max_pages;
488 ra->async_size = max_pages >> 1;
492 return ra_submit(ra, mapping, filp);
496 * page_cache_sync_readahead - generic file readahead
497 * @mapping: address_space which holds the pagecache and I/O vectors
498 * @ra: file_ra_state which holds the readahead state
499 * @filp: passed on to ->readpage() and ->readpages()
500 * @offset: start offset into @mapping, in pagecache page-sized units
501 * @req_size: hint: total size of the read which the caller is performing in
504 * page_cache_sync_readahead() should be called when a cache miss happened:
505 * it will submit the read. The readahead logic may decide to piggyback more
506 * pages onto the read request if access patterns suggest it will improve
509 void page_cache_sync_readahead(struct address_space *mapping,
510 struct file_ra_state *ra, struct file *filp,
511 pgoff_t offset, unsigned long req_size)
517 if (blk_cgroup_congested())
521 if (filp && (filp->f_mode & FMODE_RANDOM)) {
522 force_page_cache_readahead(mapping, filp, offset, req_size);
527 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
529 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
532 * page_cache_async_readahead - file readahead for marked pages
533 * @mapping: address_space which holds the pagecache and I/O vectors
534 * @ra: file_ra_state which holds the readahead state
535 * @filp: passed on to ->readpage() and ->readpages()
536 * @page: the page at @offset which has the PG_readahead flag set
537 * @offset: start offset into @mapping, in pagecache page-sized units
538 * @req_size: hint: total size of the read which the caller is performing in
541 * page_cache_async_readahead() should be called when a page is used which
542 * has the PG_readahead flag; this is a marker to suggest that the application
543 * has used up enough of the readahead window that we should start pulling in
547 page_cache_async_readahead(struct address_space *mapping,
548 struct file_ra_state *ra, struct file *filp,
549 struct page *page, pgoff_t offset,
550 unsigned long req_size)
557 * Same bit is used for PG_readahead and PG_reclaim.
559 if (PageWriteback(page))
562 ClearPageReadahead(page);
565 * Defer asynchronous read-ahead on IO congestion.
567 if (inode_read_congested(mapping->host))
570 if (blk_cgroup_congested())
574 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
576 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
578 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
585 if (!f.file || !(f.file->f_mode & FMODE_READ))
589 * The readahead() syscall is intended to run only on files
590 * that can execute readahead. If readahead is not possible
591 * on this file, then we must return -EINVAL.
594 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
595 !S_ISREG(file_inode(f.file)->i_mode))
598 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
604 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
606 return ksys_readahead(fd, offset, count);