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 mapping_gfp_constraint(mapping, GFP_KERNEL))) {
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 nr_pages)
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,
130 mapping_gfp_constraint(mapping, GFP_KERNEL))) {
131 mapping->a_ops->readpage(filp, page);
138 blk_finish_plug(&plug);
144 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
145 * the pages first, then submits them all 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 int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
152 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);
161 loff_t isize = i_size_read(inode);
166 end_index = ((isize - 1) >> PAGE_SHIFT);
169 * Preallocate as many pages as we will need.
171 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
172 pgoff_t page_offset = offset + page_idx;
174 if (page_offset > end_index)
178 page = radix_tree_lookup(&mapping->page_tree, page_offset);
180 if (page && !radix_tree_exceptional_entry(page))
183 page = page_cache_alloc_readahead(mapping);
186 page->index = page_offset;
187 list_add(&page->lru, &page_pool);
188 if (page_idx == nr_to_read - lookahead_size)
189 SetPageReadahead(page);
194 * Now start the IO. We ignore I/O errors - if the page is not
195 * uptodate then the caller will launch readpage again, and
196 * will then handle the error.
199 read_pages(mapping, filp, &page_pool, ret);
200 BUG_ON(!list_empty(&page_pool));
206 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
209 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
210 pgoff_t offset, unsigned long nr_to_read)
212 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
215 nr_to_read = min(nr_to_read, inode_to_bdi(mapping->host)->ra_pages);
219 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
221 if (this_chunk > nr_to_read)
222 this_chunk = nr_to_read;
223 err = __do_page_cache_readahead(mapping, filp,
224 offset, this_chunk, 0);
228 offset += this_chunk;
229 nr_to_read -= this_chunk;
235 * Set the initial window size, round to next power of 2 and square
236 * for small size, x 4 for medium, and x 2 for large
237 * for 128k (32 page) max ra
238 * 1-8 page = 32k initial, > 8 page = 128k initial
240 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
242 unsigned long newsize = roundup_pow_of_two(size);
244 if (newsize <= max / 32)
245 newsize = newsize * 4;
246 else if (newsize <= max / 4)
247 newsize = newsize * 2;
255 * Get the previous window size, ramp it up, and
256 * return it as the new window size.
258 static unsigned long get_next_ra_size(struct file_ra_state *ra,
261 unsigned long cur = ra->size;
262 unsigned long newsize;
269 return min(newsize, max);
273 * On-demand readahead design.
275 * The fields in struct file_ra_state represent the most-recently-executed
278 * |<----- async_size ---------|
279 * |------------------- size -------------------->|
280 * |==================#===========================|
281 * ^start ^page marked with PG_readahead
283 * To overlap application thinking time and disk I/O time, we do
284 * `readahead pipelining': Do not wait until the application consumed all
285 * readahead pages and stalled on the missing page at readahead_index;
286 * Instead, submit an asynchronous readahead I/O as soon as there are
287 * only async_size pages left in the readahead window. Normally async_size
288 * will be equal to size, for maximum pipelining.
290 * In interleaved sequential reads, concurrent streams on the same fd can
291 * be invalidating each other's readahead state. So we flag the new readahead
292 * page at (start+size-async_size) with PG_readahead, and use it as readahead
293 * indicator. The flag won't be set on already cached pages, to avoid the
294 * readahead-for-nothing fuss, saving pointless page cache lookups.
296 * prev_pos tracks the last visited byte in the _previous_ read request.
297 * It should be maintained by the caller, and will be used for detecting
298 * small random reads. Note that the readahead algorithm checks loosely
299 * for sequential patterns. Hence interleaved reads might be served as
302 * There is a special-case: if the first page which the application tries to
303 * read happens to be the first page of the file, it is assumed that a linear
304 * read is about to happen and the window is immediately set to the initial size
305 * based on I/O request size and the max_readahead.
307 * The code ramps up the readahead size aggressively at first, but slow down as
308 * it approaches max_readhead.
312 * Count contiguously cached pages from @offset-1 to @offset-@max,
313 * this count is a conservative estimation of
314 * - length of the sequential read sequence, or
315 * - thrashing threshold in memory tight systems
317 static pgoff_t count_history_pages(struct address_space *mapping,
318 pgoff_t offset, unsigned long max)
323 head = page_cache_prev_hole(mapping, offset - 1, max);
326 return offset - 1 - head;
330 * page cache context based read-ahead
332 static int try_context_readahead(struct address_space *mapping,
333 struct file_ra_state *ra,
335 unsigned long req_size,
340 size = count_history_pages(mapping, offset, max);
343 * not enough history pages:
344 * it could be a random read
346 if (size <= req_size)
350 * starts from beginning of file:
351 * it is a strong indication of long-run stream (or whole-file-read)
357 ra->size = min(size + req_size, max);
364 * A minimal readahead algorithm for trivial sequential/random reads.
367 ondemand_readahead(struct address_space *mapping,
368 struct file_ra_state *ra, struct file *filp,
369 bool hit_readahead_marker, pgoff_t offset,
370 unsigned long req_size)
372 unsigned long max = ra->ra_pages;
379 goto initial_readahead;
382 * It's the expected callback offset, assume sequential access.
383 * Ramp up sizes, and push forward the readahead window.
385 if ((offset == (ra->start + ra->size - ra->async_size) ||
386 offset == (ra->start + ra->size))) {
387 ra->start += ra->size;
388 ra->size = get_next_ra_size(ra, max);
389 ra->async_size = ra->size;
394 * Hit a marked page without valid readahead state.
395 * E.g. interleaved reads.
396 * Query the pagecache for async_size, which normally equals to
397 * readahead size. Ramp it up and use it as the new readahead size.
399 if (hit_readahead_marker) {
403 start = page_cache_next_hole(mapping, offset + 1, max);
406 if (!start || start - offset > max)
410 ra->size = start - offset; /* old async_size */
411 ra->size += req_size;
412 ra->size = get_next_ra_size(ra, max);
413 ra->async_size = ra->size;
421 goto initial_readahead;
424 * sequential cache miss
425 * trivial case: (offset - prev_offset) == 1
426 * unaligned reads: (offset - prev_offset) == 0
428 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
429 if (offset - prev_offset <= 1UL)
430 goto initial_readahead;
433 * Query the page cache and look for the traces(cached history pages)
434 * that a sequential stream would leave behind.
436 if (try_context_readahead(mapping, ra, offset, req_size, max))
440 * standalone, small random read
441 * Read as is, and do not pollute the readahead state.
443 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
447 ra->size = get_init_ra_size(req_size, max);
448 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
452 * Will this read hit the readahead marker made by itself?
453 * If so, trigger the readahead marker hit now, and merge
454 * the resulted next readahead window into the current one.
456 if (offset == ra->start && ra->size == ra->async_size) {
457 ra->async_size = get_next_ra_size(ra, max);
458 ra->size += ra->async_size;
461 return ra_submit(ra, mapping, filp);
465 * page_cache_sync_readahead - generic file readahead
466 * @mapping: address_space which holds the pagecache and I/O vectors
467 * @ra: file_ra_state which holds the readahead state
468 * @filp: passed on to ->readpage() and ->readpages()
469 * @offset: start offset into @mapping, in pagecache page-sized units
470 * @req_size: hint: total size of the read which the caller is performing in
473 * page_cache_sync_readahead() should be called when a cache miss happened:
474 * it will submit the read. The readahead logic may decide to piggyback more
475 * pages onto the read request if access patterns suggest it will improve
478 void page_cache_sync_readahead(struct address_space *mapping,
479 struct file_ra_state *ra, struct file *filp,
480 pgoff_t offset, unsigned long req_size)
487 if (filp && (filp->f_mode & FMODE_RANDOM)) {
488 force_page_cache_readahead(mapping, filp, offset, req_size);
493 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
495 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
498 * page_cache_async_readahead - file readahead for marked pages
499 * @mapping: address_space which holds the pagecache and I/O vectors
500 * @ra: file_ra_state which holds the readahead state
501 * @filp: passed on to ->readpage() and ->readpages()
502 * @page: the page at @offset which has the PG_readahead flag set
503 * @offset: start offset into @mapping, in pagecache page-sized units
504 * @req_size: hint: total size of the read which the caller is performing in
507 * page_cache_async_readahead() should be called when a page is used which
508 * has the PG_readahead flag; this is a marker to suggest that the application
509 * has used up enough of the readahead window that we should start pulling in
513 page_cache_async_readahead(struct address_space *mapping,
514 struct file_ra_state *ra, struct file *filp,
515 struct page *page, pgoff_t offset,
516 unsigned long req_size)
523 * Same bit is used for PG_readahead and PG_reclaim.
525 if (PageWriteback(page))
528 ClearPageReadahead(page);
531 * Defer asynchronous read-ahead on IO congestion.
533 if (inode_read_congested(mapping->host))
537 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
539 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
542 do_readahead(struct address_space *mapping, struct file *filp,
543 pgoff_t index, unsigned long nr)
545 if (!mapping || !mapping->a_ops)
548 return force_page_cache_readahead(mapping, filp, index, nr);
551 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
559 if (f.file->f_mode & FMODE_READ) {
560 struct address_space *mapping = f.file->f_mapping;
561 pgoff_t start = offset >> PAGE_SHIFT;
562 pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
563 unsigned long len = end - start + 1;
564 ret = do_readahead(mapping, f.file, start, len);