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/gfp.h>
12 #include <linux/export.h>
13 #include <linux/blkdev.h>
14 #include <linux/backing-dev.h>
15 #include <linux/task_io_accounting_ops.h>
16 #include <linux/pagevec.h>
17 #include <linux/pagemap.h>
18 #include <linux/syscalls.h>
19 #include <linux/file.h>
20 #include <linux/mm_inline.h>
25 * Initialise a struct file's readahead state. Assumes that the caller has
29 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
31 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
34 EXPORT_SYMBOL_GPL(file_ra_state_init);
37 * see if a page needs releasing upon read_cache_pages() failure
38 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
39 * before calling, such as the NFS fs marking pages that are cached locally
40 * on disk, thus we need to give the fs a chance to clean up in the event of
43 static void read_cache_pages_invalidate_page(struct address_space *mapping,
46 if (page_has_private(page)) {
47 if (!trylock_page(page))
49 page->mapping = mapping;
50 do_invalidatepage(page, 0, PAGE_SIZE);
58 * release a list of pages, invalidating them first if need be
60 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
61 struct list_head *pages)
65 while (!list_empty(pages)) {
66 victim = lru_to_page(pages);
67 list_del(&victim->lru);
68 read_cache_pages_invalidate_page(mapping, victim);
73 * read_cache_pages - populate an address space with some pages & start reads against them
74 * @mapping: the address_space
75 * @pages: The address of a list_head which contains the target pages. These
76 * pages have their ->index populated and are otherwise uninitialised.
77 * @filler: callback routine for filling a single page.
78 * @data: private data for the callback routine.
80 * Hides the details of the LRU cache etc from the filesystems.
82 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
83 int (*filler)(void *, struct page *), void *data)
88 while (!list_empty(pages)) {
89 page = lru_to_page(pages);
91 if (add_to_page_cache_lru(page, mapping, page->index,
92 readahead_gfp_mask(mapping))) {
93 read_cache_pages_invalidate_page(mapping, page);
98 ret = filler(data, page);
100 read_cache_pages_invalidate_pages(mapping, pages);
103 task_io_account_read(PAGE_SIZE);
108 EXPORT_SYMBOL(read_cache_pages);
110 static int read_pages(struct address_space *mapping, struct file *filp,
111 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
113 struct blk_plug plug;
117 blk_start_plug(&plug);
119 if (mapping->a_ops->readpages) {
120 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
121 /* Clean up the remaining pages */
122 put_pages_list(pages);
126 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
127 struct page *page = lru_to_page(pages);
128 list_del(&page->lru);
129 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
130 mapping->a_ops->readpage(filp, page);
136 blk_finish_plug(&plug);
142 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
143 * the pages first, then submits them all for I/O. This avoids the very bad
144 * behaviour which would occur if page allocations are causing VM writeback.
145 * We really don't want to intermingle reads and writes like that.
147 * Returns the number of pages requested, or the maximum amount of I/O allowed.
149 int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
150 pgoff_t offset, unsigned long nr_to_read,
151 unsigned long lookahead_size)
153 struct inode *inode = mapping->host;
155 unsigned long end_index; /* The last page we want to read */
156 LIST_HEAD(page_pool);
159 loff_t isize = i_size_read(inode);
160 gfp_t gfp_mask = readahead_gfp_mask(mapping);
165 end_index = ((isize - 1) >> PAGE_SHIFT);
168 * Preallocate as many pages as we will need.
170 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
171 pgoff_t page_offset = offset + page_idx;
173 if (page_offset > end_index)
177 page = radix_tree_lookup(&mapping->page_tree, page_offset);
179 if (page && !radix_tree_exceptional_entry(page))
182 page = __page_cache_alloc(gfp_mask);
185 page->index = page_offset;
186 list_add(&page->lru, &page_pool);
187 if (page_idx == nr_to_read - lookahead_size)
188 SetPageReadahead(page);
193 * Now start the IO. We ignore I/O errors - if the page is not
194 * uptodate then the caller will launch readpage again, and
195 * will then handle the error.
198 read_pages(mapping, filp, &page_pool, ret, gfp_mask);
199 BUG_ON(!list_empty(&page_pool));
205 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
208 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
209 pgoff_t offset, unsigned long nr_to_read)
211 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
214 nr_to_read = min(nr_to_read, inode_to_bdi(mapping->host)->ra_pages);
218 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
220 if (this_chunk > nr_to_read)
221 this_chunk = nr_to_read;
222 err = __do_page_cache_readahead(mapping, filp,
223 offset, this_chunk, 0);
227 offset += this_chunk;
228 nr_to_read -= this_chunk;
234 * Set the initial window size, round to next power of 2 and square
235 * for small size, x 4 for medium, and x 2 for large
236 * for 128k (32 page) max ra
237 * 1-8 page = 32k initial, > 8 page = 128k initial
239 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
241 unsigned long newsize = roundup_pow_of_two(size);
243 if (newsize <= max / 32)
244 newsize = newsize * 4;
245 else if (newsize <= max / 4)
246 newsize = newsize * 2;
254 * Get the previous window size, ramp it up, and
255 * return it as the new window size.
257 static unsigned long get_next_ra_size(struct file_ra_state *ra,
260 unsigned long cur = ra->size;
261 unsigned long newsize;
268 return min(newsize, max);
272 * On-demand readahead design.
274 * The fields in struct file_ra_state represent the most-recently-executed
277 * |<----- async_size ---------|
278 * |------------------- size -------------------->|
279 * |==================#===========================|
280 * ^start ^page marked with PG_readahead
282 * To overlap application thinking time and disk I/O time, we do
283 * `readahead pipelining': Do not wait until the application consumed all
284 * readahead pages and stalled on the missing page at readahead_index;
285 * Instead, submit an asynchronous readahead I/O as soon as there are
286 * only async_size pages left in the readahead window. Normally async_size
287 * will be equal to size, for maximum pipelining.
289 * In interleaved sequential reads, concurrent streams on the same fd can
290 * be invalidating each other's readahead state. So we flag the new readahead
291 * page at (start+size-async_size) with PG_readahead, and use it as readahead
292 * indicator. The flag won't be set on already cached pages, to avoid the
293 * readahead-for-nothing fuss, saving pointless page cache lookups.
295 * prev_pos tracks the last visited byte in the _previous_ read request.
296 * It should be maintained by the caller, and will be used for detecting
297 * small random reads. Note that the readahead algorithm checks loosely
298 * for sequential patterns. Hence interleaved reads might be served as
301 * There is a special-case: if the first page which the application tries to
302 * read happens to be the first page of the file, it is assumed that a linear
303 * read is about to happen and the window is immediately set to the initial size
304 * based on I/O request size and the max_readahead.
306 * The code ramps up the readahead size aggressively at first, but slow down as
307 * it approaches max_readhead.
311 * Count contiguously cached pages from @offset-1 to @offset-@max,
312 * this count is a conservative estimation of
313 * - length of the sequential read sequence, or
314 * - thrashing threshold in memory tight systems
316 static pgoff_t count_history_pages(struct address_space *mapping,
317 pgoff_t offset, unsigned long max)
322 head = page_cache_prev_hole(mapping, offset - 1, max);
325 return offset - 1 - head;
329 * page cache context based read-ahead
331 static int try_context_readahead(struct address_space *mapping,
332 struct file_ra_state *ra,
334 unsigned long req_size,
339 size = count_history_pages(mapping, offset, max);
342 * not enough history pages:
343 * it could be a random read
345 if (size <= req_size)
349 * starts from beginning of file:
350 * it is a strong indication of long-run stream (or whole-file-read)
356 ra->size = min(size + req_size, max);
363 * A minimal readahead algorithm for trivial sequential/random reads.
366 ondemand_readahead(struct address_space *mapping,
367 struct file_ra_state *ra, struct file *filp,
368 bool hit_readahead_marker, pgoff_t offset,
369 unsigned long req_size)
371 unsigned long max = ra->ra_pages;
378 goto initial_readahead;
381 * It's the expected callback offset, assume sequential access.
382 * Ramp up sizes, and push forward the readahead window.
384 if ((offset == (ra->start + ra->size - ra->async_size) ||
385 offset == (ra->start + ra->size))) {
386 ra->start += ra->size;
387 ra->size = get_next_ra_size(ra, max);
388 ra->async_size = ra->size;
393 * Hit a marked page without valid readahead state.
394 * E.g. interleaved reads.
395 * Query the pagecache for async_size, which normally equals to
396 * readahead size. Ramp it up and use it as the new readahead size.
398 if (hit_readahead_marker) {
402 start = page_cache_next_hole(mapping, offset + 1, max);
405 if (!start || start - offset > max)
409 ra->size = start - offset; /* old async_size */
410 ra->size += req_size;
411 ra->size = get_next_ra_size(ra, max);
412 ra->async_size = ra->size;
420 goto initial_readahead;
423 * sequential cache miss
424 * trivial case: (offset - prev_offset) == 1
425 * unaligned reads: (offset - prev_offset) == 0
427 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
428 if (offset - prev_offset <= 1UL)
429 goto initial_readahead;
432 * Query the page cache and look for the traces(cached history pages)
433 * that a sequential stream would leave behind.
435 if (try_context_readahead(mapping, ra, offset, req_size, max))
439 * standalone, small random read
440 * Read as is, and do not pollute the readahead state.
442 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
446 ra->size = get_init_ra_size(req_size, max);
447 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
451 * Will this read hit the readahead marker made by itself?
452 * If so, trigger the readahead marker hit now, and merge
453 * the resulted next readahead window into the current one.
455 if (offset == ra->start && ra->size == ra->async_size) {
456 ra->async_size = get_next_ra_size(ra, max);
457 ra->size += ra->async_size;
460 return ra_submit(ra, mapping, filp);
464 * page_cache_sync_readahead - generic file readahead
465 * @mapping: address_space which holds the pagecache and I/O vectors
466 * @ra: file_ra_state which holds the readahead state
467 * @filp: passed on to ->readpage() and ->readpages()
468 * @offset: start offset into @mapping, in pagecache page-sized units
469 * @req_size: hint: total size of the read which the caller is performing in
472 * page_cache_sync_readahead() should be called when a cache miss happened:
473 * it will submit the read. The readahead logic may decide to piggyback more
474 * pages onto the read request if access patterns suggest it will improve
477 void page_cache_sync_readahead(struct address_space *mapping,
478 struct file_ra_state *ra, struct file *filp,
479 pgoff_t offset, unsigned long req_size)
486 if (filp && (filp->f_mode & FMODE_RANDOM)) {
487 force_page_cache_readahead(mapping, filp, offset, req_size);
492 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
494 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
497 * page_cache_async_readahead - file readahead for marked pages
498 * @mapping: address_space which holds the pagecache and I/O vectors
499 * @ra: file_ra_state which holds the readahead state
500 * @filp: passed on to ->readpage() and ->readpages()
501 * @page: the page at @offset which has the PG_readahead flag set
502 * @offset: start offset into @mapping, in pagecache page-sized units
503 * @req_size: hint: total size of the read which the caller is performing in
506 * page_cache_async_readahead() should be called when a page is used which
507 * has the PG_readahead flag; this is a marker to suggest that the application
508 * has used up enough of the readahead window that we should start pulling in
512 page_cache_async_readahead(struct address_space *mapping,
513 struct file_ra_state *ra, struct file *filp,
514 struct page *page, pgoff_t offset,
515 unsigned long req_size)
522 * Same bit is used for PG_readahead and PG_reclaim.
524 if (PageWriteback(page))
527 ClearPageReadahead(page);
530 * Defer asynchronous read-ahead on IO congestion.
532 if (inode_read_congested(mapping->host))
536 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
538 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
541 do_readahead(struct address_space *mapping, struct file *filp,
542 pgoff_t index, unsigned long nr)
544 if (!mapping || !mapping->a_ops)
547 return force_page_cache_readahead(mapping, filp, index, nr);
550 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
558 if (f.file->f_mode & FMODE_READ) {
559 struct address_space *mapping = f.file->f_mapping;
560 pgoff_t start = offset >> PAGE_SHIFT;
561 pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
562 unsigned long len = end - start + 1;
563 ret = do_readahead(mapping, f.file, start, len);