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>
27 * Initialise a struct file's readahead state. Assumes that the caller has
31 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
33 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
36 EXPORT_SYMBOL_GPL(file_ra_state_init);
39 * see if a page needs releasing upon read_cache_pages() failure
40 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
41 * before calling, such as the NFS fs marking pages that are cached locally
42 * on disk, thus we need to give the fs a chance to clean up in the event of
45 static void read_cache_pages_invalidate_page(struct address_space *mapping,
48 if (page_has_private(page)) {
49 if (!trylock_page(page))
51 page->mapping = mapping;
52 do_invalidatepage(page, 0, PAGE_SIZE);
60 * release a list of pages, invalidating them first if need be
62 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
63 struct list_head *pages)
67 while (!list_empty(pages)) {
68 victim = lru_to_page(pages);
69 list_del(&victim->lru);
70 read_cache_pages_invalidate_page(mapping, victim);
75 * read_cache_pages - populate an address space with some pages & start reads against them
76 * @mapping: the address_space
77 * @pages: The address of a list_head which contains the target pages. These
78 * pages have their ->index populated and are otherwise uninitialised.
79 * @filler: callback routine for filling a single page.
80 * @data: private data for the callback routine.
82 * Hides the details of the LRU cache etc from the filesystems.
84 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
85 int (*filler)(void *, struct page *), void *data)
90 while (!list_empty(pages)) {
91 page = lru_to_page(pages);
93 if (add_to_page_cache_lru(page, mapping, page->index,
94 readahead_gfp_mask(mapping))) {
95 read_cache_pages_invalidate_page(mapping, page);
100 ret = filler(data, page);
102 read_cache_pages_invalidate_pages(mapping, pages);
105 task_io_account_read(PAGE_SIZE);
110 EXPORT_SYMBOL(read_cache_pages);
112 static int read_pages(struct address_space *mapping, struct file *filp,
113 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
115 struct blk_plug plug;
119 blk_start_plug(&plug);
121 if (mapping->a_ops->readpages) {
122 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
123 /* Clean up the remaining pages */
124 put_pages_list(pages);
128 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
129 struct page *page = lru_to_page(pages);
130 list_del(&page->lru);
131 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
132 mapping->a_ops->readpage(filp, page);
138 blk_finish_plug(&plug);
144 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates
145 * the pages first, then submits them for I/O. This avoids the very bad
146 * behaviour which would occur if page allocations are causing VM writeback.
147 * We really don't want to intermingle reads and writes like that.
149 * Returns the number of pages requested, or the maximum amount of I/O allowed.
151 unsigned int __do_page_cache_readahead(struct address_space *mapping,
152 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
153 unsigned long lookahead_size)
155 struct inode *inode = mapping->host;
157 unsigned long end_index; /* The last page we want to read */
158 LIST_HEAD(page_pool);
160 unsigned int nr_pages = 0;
161 loff_t isize = i_size_read(inode);
162 gfp_t gfp_mask = readahead_gfp_mask(mapping);
167 end_index = ((isize - 1) >> PAGE_SHIFT);
170 * Preallocate as many pages as we will need.
172 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
173 pgoff_t page_offset = offset + page_idx;
175 if (page_offset > end_index)
179 page = radix_tree_lookup(&mapping->i_pages, page_offset);
181 if (page && !radix_tree_exceptional_entry(page)) {
183 * Page already present? Kick off the current batch of
184 * contiguous pages before continuing with the next
188 read_pages(mapping, filp, &page_pool, nr_pages,
194 page = __page_cache_alloc(gfp_mask);
197 page->index = page_offset;
198 list_add(&page->lru, &page_pool);
199 if (page_idx == nr_to_read - lookahead_size)
200 SetPageReadahead(page);
205 * Now start the IO. We ignore I/O errors - if the page is not
206 * uptodate then the caller will launch readpage again, and
207 * will then handle the error.
210 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
211 BUG_ON(!list_empty(&page_pool));
217 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
220 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
221 pgoff_t offset, unsigned long nr_to_read)
223 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
224 struct file_ra_state *ra = &filp->f_ra;
225 unsigned long max_pages;
227 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
231 * If the request exceeds the readahead window, allow the read to
232 * be up to the optimal hardware IO size
234 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
235 nr_to_read = min(nr_to_read, max_pages);
237 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
239 if (this_chunk > nr_to_read)
240 this_chunk = nr_to_read;
241 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
243 offset += this_chunk;
244 nr_to_read -= this_chunk;
250 * Set the initial window size, round to next power of 2 and square
251 * for small size, x 4 for medium, and x 2 for large
252 * for 128k (32 page) max ra
253 * 1-8 page = 32k initial, > 8 page = 128k initial
255 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
257 unsigned long newsize = roundup_pow_of_two(size);
259 if (newsize <= max / 32)
260 newsize = newsize * 4;
261 else if (newsize <= max / 4)
262 newsize = newsize * 2;
270 * Get the previous window size, ramp it up, and
271 * return it as the new window size.
273 static unsigned long get_next_ra_size(struct file_ra_state *ra,
276 unsigned long cur = ra->size;
277 unsigned long newsize;
284 return min(newsize, max);
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_hole(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_hole(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);
579 do_readahead(struct address_space *mapping, struct file *filp,
580 pgoff_t index, unsigned long nr)
582 if (!mapping || !mapping->a_ops)
586 * Readahead doesn't make sense for DAX inodes, but we don't want it
587 * to report a failure either. Instead, we just return success and
590 if (dax_mapping(mapping))
593 return force_page_cache_readahead(mapping, filp, index, nr);
596 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
604 if (f.file->f_mode & FMODE_READ) {
605 struct address_space *mapping = f.file->f_mapping;
606 pgoff_t start = offset >> PAGE_SHIFT;
607 pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
608 unsigned long len = end - start + 1;
609 ret = do_readahead(mapping, f.file, start, len);
616 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
618 return ksys_readahead(fd, offset, count);