2 * mm/readahead.c - address_space-level file readahead.
4 * Copyright (C) 2002, Linus Torvalds
6 * 09Apr2002 Andrew Morton
10 #include <linux/kernel.h>
11 #include <linux/dax.h>
12 #include <linux/gfp.h>
13 #include <linux/export.h>
14 #include <linux/blkdev.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>
28 * Initialise a struct file's readahead state. Assumes that the caller has
32 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
34 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
37 EXPORT_SYMBOL_GPL(file_ra_state_init);
40 * see if a page needs releasing upon read_cache_pages() failure
41 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
42 * before calling, such as the NFS fs marking pages that are cached locally
43 * on disk, thus we need to give the fs a chance to clean up in the event of
46 static void read_cache_pages_invalidate_page(struct address_space *mapping,
49 if (page_has_private(page)) {
50 if (!trylock_page(page))
52 page->mapping = mapping;
53 do_invalidatepage(page, 0, PAGE_SIZE);
61 * release a list of pages, invalidating them first if need be
63 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
64 struct list_head *pages)
68 while (!list_empty(pages)) {
69 victim = lru_to_page(pages);
70 list_del(&victim->lru);
71 read_cache_pages_invalidate_page(mapping, victim);
76 * read_cache_pages - populate an address space with some pages & start reads against them
77 * @mapping: the address_space
78 * @pages: The address of a list_head which contains the target pages. These
79 * pages have their ->index populated and are otherwise uninitialised.
80 * @filler: callback routine for filling a single page.
81 * @data: private data for the callback routine.
83 * Hides the details of the LRU cache etc from the filesystems.
85 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
86 int (*filler)(void *, struct page *), void *data)
91 while (!list_empty(pages)) {
92 page = lru_to_page(pages);
94 if (add_to_page_cache_lru(page, mapping, page->index,
95 readahead_gfp_mask(mapping))) {
96 read_cache_pages_invalidate_page(mapping, page);
101 ret = filler(data, page);
103 read_cache_pages_invalidate_pages(mapping, pages);
106 task_io_account_read(PAGE_SIZE);
111 EXPORT_SYMBOL(read_cache_pages);
113 static int read_pages(struct address_space *mapping, struct file *filp,
114 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
116 struct blk_plug plug;
120 blk_start_plug(&plug);
122 if (mapping->a_ops->readpages) {
123 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
124 /* Clean up the remaining pages */
125 put_pages_list(pages);
129 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
130 struct page *page = lru_to_page(pages);
131 list_del(&page->lru);
132 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
133 mapping->a_ops->readpage(filp, page);
139 blk_finish_plug(&plug);
145 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates
146 * the pages first, then submits them for I/O. This avoids the very bad
147 * behaviour which would occur if page allocations are causing VM writeback.
148 * We really don't want to intermingle reads and writes like that.
150 * Returns the number of pages requested, or the maximum amount of I/O allowed.
152 unsigned int __do_page_cache_readahead(struct address_space *mapping,
153 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
154 unsigned long lookahead_size)
156 struct inode *inode = mapping->host;
158 unsigned long end_index; /* The last page we want to read */
159 LIST_HEAD(page_pool);
161 unsigned int nr_pages = 0;
162 loff_t isize = i_size_read(inode);
163 gfp_t gfp_mask = readahead_gfp_mask(mapping);
168 end_index = ((isize - 1) >> PAGE_SHIFT);
171 * Preallocate as many pages as we will need.
173 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
174 pgoff_t page_offset = offset + page_idx;
176 if (page_offset > end_index)
179 page = xa_load(&mapping->i_pages, page_offset);
180 if (page && !xa_is_value(page)) {
182 * Page already present? Kick off the current batch of
183 * contiguous pages before continuing with the next
187 read_pages(mapping, filp, &page_pool, nr_pages,
193 page = __page_cache_alloc(gfp_mask);
196 page->index = page_offset;
197 list_add(&page->lru, &page_pool);
198 if (page_idx == nr_to_read - lookahead_size)
199 SetPageReadahead(page);
204 * Now start the IO. We ignore I/O errors - if the page is not
205 * uptodate then the caller will launch readpage again, and
206 * will then handle the error.
209 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
210 BUG_ON(!list_empty(&page_pool));
216 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
219 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
220 pgoff_t offset, unsigned long nr_to_read)
222 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
223 struct file_ra_state *ra = &filp->f_ra;
224 unsigned long max_pages;
226 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
230 * If the request exceeds the readahead window, allow the read to
231 * be up to the optimal hardware IO size
233 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
234 nr_to_read = min(nr_to_read, max_pages);
236 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
238 if (this_chunk > nr_to_read)
239 this_chunk = nr_to_read;
240 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
242 offset += this_chunk;
243 nr_to_read -= this_chunk;
249 * Set the initial window size, round to next power of 2 and square
250 * for small size, x 4 for medium, and x 2 for large
251 * for 128k (32 page) max ra
252 * 1-8 page = 32k initial, > 8 page = 128k initial
254 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
256 unsigned long newsize = roundup_pow_of_two(size);
258 if (newsize <= max / 32)
259 newsize = newsize * 4;
260 else if (newsize <= max / 4)
261 newsize = newsize * 2;
269 * Get the previous window size, ramp it up, and
270 * return it as the new window size.
272 static unsigned long get_next_ra_size(struct file_ra_state *ra,
275 unsigned long cur = ra->size;
285 * On-demand readahead design.
287 * The fields in struct file_ra_state represent the most-recently-executed
290 * |<----- async_size ---------|
291 * |------------------- size -------------------->|
292 * |==================#===========================|
293 * ^start ^page marked with PG_readahead
295 * To overlap application thinking time and disk I/O time, we do
296 * `readahead pipelining': Do not wait until the application consumed all
297 * readahead pages and stalled on the missing page at readahead_index;
298 * Instead, submit an asynchronous readahead I/O as soon as there are
299 * only async_size pages left in the readahead window. Normally async_size
300 * will be equal to size, for maximum pipelining.
302 * In interleaved sequential reads, concurrent streams on the same fd can
303 * be invalidating each other's readahead state. So we flag the new readahead
304 * page at (start+size-async_size) with PG_readahead, and use it as readahead
305 * indicator. The flag won't be set on already cached pages, to avoid the
306 * readahead-for-nothing fuss, saving pointless page cache lookups.
308 * prev_pos tracks the last visited byte in the _previous_ read request.
309 * It should be maintained by the caller, and will be used for detecting
310 * small random reads. Note that the readahead algorithm checks loosely
311 * for sequential patterns. Hence interleaved reads might be served as
314 * There is a special-case: if the first page which the application tries to
315 * read happens to be the first page of the file, it is assumed that a linear
316 * read is about to happen and the window is immediately set to the initial size
317 * based on I/O request size and the max_readahead.
319 * The code ramps up the readahead size aggressively at first, but slow down as
320 * it approaches max_readhead.
324 * Count contiguously cached pages from @offset-1 to @offset-@max,
325 * this count is a conservative estimation of
326 * - length of the sequential read sequence, or
327 * - thrashing threshold in memory tight systems
329 static pgoff_t count_history_pages(struct address_space *mapping,
330 pgoff_t offset, unsigned long max)
335 head = page_cache_prev_miss(mapping, offset - 1, max);
338 return offset - 1 - head;
342 * page cache context based read-ahead
344 static int try_context_readahead(struct address_space *mapping,
345 struct file_ra_state *ra,
347 unsigned long req_size,
352 size = count_history_pages(mapping, offset, max);
355 * not enough history pages:
356 * it could be a random read
358 if (size <= req_size)
362 * starts from beginning of file:
363 * it is a strong indication of long-run stream (or whole-file-read)
369 ra->size = min(size + req_size, max);
376 * A minimal readahead algorithm for trivial sequential/random reads.
379 ondemand_readahead(struct address_space *mapping,
380 struct file_ra_state *ra, struct file *filp,
381 bool hit_readahead_marker, pgoff_t offset,
382 unsigned long req_size)
384 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
385 unsigned long max_pages = ra->ra_pages;
386 unsigned long add_pages;
390 * If the request exceeds the readahead window, allow the read to
391 * be up to the optimal hardware IO size
393 if (req_size > max_pages && bdi->io_pages > max_pages)
394 max_pages = min(req_size, bdi->io_pages);
400 goto initial_readahead;
403 * It's the expected callback offset, assume sequential access.
404 * Ramp up sizes, and push forward the readahead window.
406 if ((offset == (ra->start + ra->size - ra->async_size) ||
407 offset == (ra->start + ra->size))) {
408 ra->start += ra->size;
409 ra->size = get_next_ra_size(ra, max_pages);
410 ra->async_size = ra->size;
415 * Hit a marked page without valid readahead state.
416 * E.g. interleaved reads.
417 * Query the pagecache for async_size, which normally equals to
418 * readahead size. Ramp it up and use it as the new readahead size.
420 if (hit_readahead_marker) {
424 start = page_cache_next_miss(mapping, offset + 1, max_pages);
427 if (!start || start - offset > max_pages)
431 ra->size = start - offset; /* old async_size */
432 ra->size += req_size;
433 ra->size = get_next_ra_size(ra, max_pages);
434 ra->async_size = ra->size;
441 if (req_size > max_pages)
442 goto initial_readahead;
445 * sequential cache miss
446 * trivial case: (offset - prev_offset) == 1
447 * unaligned reads: (offset - prev_offset) == 0
449 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
450 if (offset - prev_offset <= 1UL)
451 goto initial_readahead;
454 * Query the page cache and look for the traces(cached history pages)
455 * that a sequential stream would leave behind.
457 if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
461 * standalone, small random read
462 * Read as is, and do not pollute the readahead state.
464 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
468 ra->size = get_init_ra_size(req_size, max_pages);
469 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
473 * Will this read hit the readahead marker made by itself?
474 * If so, trigger the readahead marker hit now, and merge
475 * the resulted next readahead window into the current one.
476 * Take care of maximum IO pages as above.
478 if (offset == ra->start && ra->size == ra->async_size) {
479 add_pages = get_next_ra_size(ra, max_pages);
480 if (ra->size + add_pages <= max_pages) {
481 ra->async_size = add_pages;
482 ra->size += add_pages;
484 ra->size = max_pages;
485 ra->async_size = max_pages >> 1;
489 return ra_submit(ra, mapping, filp);
493 * page_cache_sync_readahead - generic file readahead
494 * @mapping: address_space which holds the pagecache and I/O vectors
495 * @ra: file_ra_state which holds the readahead state
496 * @filp: passed on to ->readpage() and ->readpages()
497 * @offset: start offset into @mapping, in pagecache page-sized units
498 * @req_size: hint: total size of the read which the caller is performing in
501 * page_cache_sync_readahead() should be called when a cache miss happened:
502 * it will submit the read. The readahead logic may decide to piggyback more
503 * pages onto the read request if access patterns suggest it will improve
506 void page_cache_sync_readahead(struct address_space *mapping,
507 struct file_ra_state *ra, struct file *filp,
508 pgoff_t offset, unsigned long req_size)
514 if (blk_cgroup_congested())
518 if (filp && (filp->f_mode & FMODE_RANDOM)) {
519 force_page_cache_readahead(mapping, filp, offset, req_size);
524 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
526 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
529 * page_cache_async_readahead - file readahead for marked pages
530 * @mapping: address_space which holds the pagecache and I/O vectors
531 * @ra: file_ra_state which holds the readahead state
532 * @filp: passed on to ->readpage() and ->readpages()
533 * @page: the page at @offset which has the PG_readahead flag set
534 * @offset: start offset into @mapping, in pagecache page-sized units
535 * @req_size: hint: total size of the read which the caller is performing in
538 * page_cache_async_readahead() should be called when a page is used which
539 * has the PG_readahead flag; this is a marker to suggest that the application
540 * has used up enough of the readahead window that we should start pulling in
544 page_cache_async_readahead(struct address_space *mapping,
545 struct file_ra_state *ra, struct file *filp,
546 struct page *page, pgoff_t offset,
547 unsigned long req_size)
554 * Same bit is used for PG_readahead and PG_reclaim.
556 if (PageWriteback(page))
559 ClearPageReadahead(page);
562 * Defer asynchronous read-ahead on IO congestion.
564 if (inode_read_congested(mapping->host))
567 if (blk_cgroup_congested())
571 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
573 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
575 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
582 if (!f.file || !(f.file->f_mode & FMODE_READ))
586 * The readahead() syscall is intended to run only on files
587 * that can execute readahead. If readahead is not possible
588 * on this file, then we must return -EINVAL.
591 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
592 !S_ISREG(file_inode(f.file)->i_mode))
595 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
601 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
603 return ksys_readahead(fd, offset, count);