2 * mm/readahead.c - address_space-level file readahead.
4 * Copyright (C) 2002, Linus Torvalds
6 * 09Apr2002 Andrew Morton
10 #include <linux/kernel.h>
12 #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>
24 * Initialise a struct file's readahead state. Assumes that the caller has
28 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
30 ra->ra_pages = mapping->backing_dev_info->ra_pages;
33 EXPORT_SYMBOL_GPL(file_ra_state_init);
35 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
38 * see if a page needs releasing upon read_cache_pages() failure
39 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
40 * before calling, such as the NFS fs marking pages that are cached locally
41 * on disk, thus we need to give the fs a chance to clean up in the event of
44 static void read_cache_pages_invalidate_page(struct address_space *mapping,
47 if (page_has_private(page)) {
48 if (!trylock_page(page))
50 page->mapping = mapping;
51 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
55 page_cache_release(page);
59 * release a list of pages, invalidating them first if need be
61 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
62 struct list_head *pages)
66 while (!list_empty(pages)) {
67 victim = list_to_page(pages);
68 list_del(&victim->lru);
69 read_cache_pages_invalidate_page(mapping, victim);
74 * read_cache_pages - populate an address space with some pages & start reads against them
75 * @mapping: the address_space
76 * @pages: The address of a list_head which contains the target pages. These
77 * pages have their ->index populated and are otherwise uninitialised.
78 * @filler: callback routine for filling a single page.
79 * @data: private data for the callback routine.
81 * Hides the details of the LRU cache etc from the filesystems.
83 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
84 int (*filler)(void *, struct page *), void *data)
89 while (!list_empty(pages)) {
90 page = list_to_page(pages);
92 if (add_to_page_cache_lru(page, mapping,
93 page->index, GFP_KERNEL)) {
94 read_cache_pages_invalidate_page(mapping, page);
97 page_cache_release(page);
99 ret = filler(data, page);
101 read_cache_pages_invalidate_pages(mapping, pages);
104 task_io_account_read(PAGE_CACHE_SIZE);
109 EXPORT_SYMBOL(read_cache_pages);
111 static int read_pages(struct address_space *mapping, struct file *filp,
112 struct list_head *pages, unsigned nr_pages)
114 struct blk_plug plug;
118 blk_start_plug(&plug);
120 if (mapping->a_ops->readpages) {
121 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
122 /* Clean up the remaining pages */
123 put_pages_list(pages);
127 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
128 struct page *page = list_to_page(pages);
129 list_del(&page->lru);
130 if (!add_to_page_cache_lru(page, mapping,
131 page->index, GFP_KERNEL)) {
132 mapping->a_ops->readpage(filp, page);
134 page_cache_release(page);
139 blk_finish_plug(&plug);
145 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
146 * the pages first, then submits them all 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.
153 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
154 pgoff_t offset, unsigned long nr_to_read,
155 unsigned long lookahead_size)
157 struct inode *inode = mapping->host;
159 unsigned long end_index; /* The last page we want to read */
160 LIST_HEAD(page_pool);
163 loff_t isize = i_size_read(inode);
168 end_index = ((isize - 1) >> PAGE_CACHE_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)
180 page = radix_tree_lookup(&mapping->page_tree, page_offset);
185 page = page_cache_alloc_readahead(mapping);
188 page->index = page_offset;
189 list_add(&page->lru, &page_pool);
190 if (page_idx == nr_to_read - lookahead_size)
191 SetPageReadahead(page);
196 * Now start the IO. We ignore I/O errors - if the page is not
197 * uptodate then the caller will launch readpage again, and
198 * will then handle the error.
201 read_pages(mapping, filp, &page_pool, ret);
202 BUG_ON(!list_empty(&page_pool));
208 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
211 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
212 pgoff_t offset, unsigned long nr_to_read)
214 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
217 nr_to_read = max_sane_readahead(nr_to_read);
221 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
223 if (this_chunk > nr_to_read)
224 this_chunk = nr_to_read;
225 err = __do_page_cache_readahead(mapping, filp,
226 offset, this_chunk, 0);
230 offset += this_chunk;
231 nr_to_read -= this_chunk;
237 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
238 * sensible upper limit.
240 unsigned long max_sane_readahead(unsigned long nr)
242 return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
243 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
247 * Submit IO for the read-ahead request in file_ra_state.
249 unsigned long ra_submit(struct file_ra_state *ra,
250 struct address_space *mapping, struct file *filp)
254 actual = __do_page_cache_readahead(mapping, filp,
255 ra->start, ra->size, ra->async_size);
261 * Set the initial window size, round to next power of 2 and square
262 * for small size, x 4 for medium, and x 2 for large
263 * for 128k (32 page) max ra
264 * 1-8 page = 32k initial, > 8 page = 128k initial
266 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
268 unsigned long newsize = roundup_pow_of_two(size);
270 if (newsize <= max / 32)
271 newsize = newsize * 4;
272 else if (newsize <= max / 4)
273 newsize = newsize * 2;
281 * Get the previous window size, ramp it up, and
282 * return it as the new window size.
284 static unsigned long get_next_ra_size(struct file_ra_state *ra,
287 unsigned long cur = ra->size;
288 unsigned long newsize;
295 return min(newsize, max);
299 * On-demand readahead design.
301 * The fields in struct file_ra_state represent the most-recently-executed
304 * |<----- async_size ---------|
305 * |------------------- size -------------------->|
306 * |==================#===========================|
307 * ^start ^page marked with PG_readahead
309 * To overlap application thinking time and disk I/O time, we do
310 * `readahead pipelining': Do not wait until the application consumed all
311 * readahead pages and stalled on the missing page at readahead_index;
312 * Instead, submit an asynchronous readahead I/O as soon as there are
313 * only async_size pages left in the readahead window. Normally async_size
314 * will be equal to size, for maximum pipelining.
316 * In interleaved sequential reads, concurrent streams on the same fd can
317 * be invalidating each other's readahead state. So we flag the new readahead
318 * page at (start+size-async_size) with PG_readahead, and use it as readahead
319 * indicator. The flag won't be set on already cached pages, to avoid the
320 * readahead-for-nothing fuss, saving pointless page cache lookups.
322 * prev_pos tracks the last visited byte in the _previous_ read request.
323 * It should be maintained by the caller, and will be used for detecting
324 * small random reads. Note that the readahead algorithm checks loosely
325 * for sequential patterns. Hence interleaved reads might be served as
328 * There is a special-case: if the first page which the application tries to
329 * read happens to be the first page of the file, it is assumed that a linear
330 * read is about to happen and the window is immediately set to the initial size
331 * based on I/O request size and the max_readahead.
333 * The code ramps up the readahead size aggressively at first, but slow down as
334 * it approaches max_readhead.
338 * Count contiguously cached pages from @offset-1 to @offset-@max,
339 * this count is a conservative estimation of
340 * - length of the sequential read sequence, or
341 * - thrashing threshold in memory tight systems
343 static pgoff_t count_history_pages(struct address_space *mapping,
344 struct file_ra_state *ra,
345 pgoff_t offset, unsigned long max)
350 head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
353 return offset - 1 - head;
357 * page cache context based read-ahead
359 static int try_context_readahead(struct address_space *mapping,
360 struct file_ra_state *ra,
362 unsigned long req_size,
367 size = count_history_pages(mapping, ra, offset, max);
370 * not enough history pages:
371 * it could be a random read
373 if (size <= req_size)
377 * starts from beginning of file:
378 * it is a strong indication of long-run stream (or whole-file-read)
384 ra->size = min(size + req_size, max);
391 * A minimal readahead algorithm for trivial sequential/random reads.
394 ondemand_readahead(struct address_space *mapping,
395 struct file_ra_state *ra, struct file *filp,
396 bool hit_readahead_marker, pgoff_t offset,
397 unsigned long req_size)
399 unsigned long max = max_sane_readahead(ra->ra_pages);
406 goto initial_readahead;
409 * It's the expected callback offset, assume sequential access.
410 * Ramp up sizes, and push forward the readahead window.
412 if ((offset == (ra->start + ra->size - ra->async_size) ||
413 offset == (ra->start + ra->size))) {
414 ra->start += ra->size;
415 ra->size = get_next_ra_size(ra, max);
416 ra->async_size = ra->size;
421 * Hit a marked page without valid readahead state.
422 * E.g. interleaved reads.
423 * Query the pagecache for async_size, which normally equals to
424 * readahead size. Ramp it up and use it as the new readahead size.
426 if (hit_readahead_marker) {
430 start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
433 if (!start || start - offset > max)
437 ra->size = start - offset; /* old async_size */
438 ra->size += req_size;
439 ra->size = get_next_ra_size(ra, max);
440 ra->async_size = ra->size;
448 goto initial_readahead;
451 * sequential cache miss
452 * trivial case: (offset - prev_offset) == 1
453 * unaligned reads: (offset - prev_offset) == 0
455 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_CACHE_SHIFT;
456 if (offset - prev_offset <= 1UL)
457 goto initial_readahead;
460 * Query the page cache and look for the traces(cached history pages)
461 * that a sequential stream would leave behind.
463 if (try_context_readahead(mapping, ra, offset, req_size, max))
467 * standalone, small random read
468 * Read as is, and do not pollute the readahead state.
470 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
474 ra->size = get_init_ra_size(req_size, max);
475 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
479 * Will this read hit the readahead marker made by itself?
480 * If so, trigger the readahead marker hit now, and merge
481 * the resulted next readahead window into the current one.
483 if (offset == ra->start && ra->size == ra->async_size) {
484 ra->async_size = get_next_ra_size(ra, max);
485 ra->size += ra->async_size;
488 return ra_submit(ra, mapping, filp);
492 * page_cache_sync_readahead - generic file readahead
493 * @mapping: address_space which holds the pagecache and I/O vectors
494 * @ra: file_ra_state which holds the readahead state
495 * @filp: passed on to ->readpage() and ->readpages()
496 * @offset: start offset into @mapping, in pagecache page-sized units
497 * @req_size: hint: total size of the read which the caller is performing in
500 * page_cache_sync_readahead() should be called when a cache miss happened:
501 * it will submit the read. The readahead logic may decide to piggyback more
502 * pages onto the read request if access patterns suggest it will improve
505 void page_cache_sync_readahead(struct address_space *mapping,
506 struct file_ra_state *ra, struct file *filp,
507 pgoff_t offset, unsigned long req_size)
514 if (filp && (filp->f_mode & FMODE_RANDOM)) {
515 force_page_cache_readahead(mapping, filp, offset, req_size);
520 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
522 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
525 * page_cache_async_readahead - file readahead for marked pages
526 * @mapping: address_space which holds the pagecache and I/O vectors
527 * @ra: file_ra_state which holds the readahead state
528 * @filp: passed on to ->readpage() and ->readpages()
529 * @page: the page at @offset which has the PG_readahead flag set
530 * @offset: start offset into @mapping, in pagecache page-sized units
531 * @req_size: hint: total size of the read which the caller is performing in
534 * page_cache_async_readahead() should be called when a page is used which
535 * has the PG_readahead flag; this is a marker to suggest that the application
536 * has used up enough of the readahead window that we should start pulling in
540 page_cache_async_readahead(struct address_space *mapping,
541 struct file_ra_state *ra, struct file *filp,
542 struct page *page, pgoff_t offset,
543 unsigned long req_size)
550 * Same bit is used for PG_readahead and PG_reclaim.
552 if (PageWriteback(page))
555 ClearPageReadahead(page);
558 * Defer asynchronous read-ahead on IO congestion.
560 if (bdi_read_congested(mapping->backing_dev_info))
564 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
566 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
569 do_readahead(struct address_space *mapping, struct file *filp,
570 pgoff_t index, unsigned long nr)
572 if (!mapping || !mapping->a_ops)
575 return force_page_cache_readahead(mapping, filp, index, nr);
578 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
586 if (f.file->f_mode & FMODE_READ) {
587 struct address_space *mapping = f.file->f_mapping;
588 pgoff_t start = offset >> PAGE_CACHE_SHIFT;
589 pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
590 unsigned long len = end - start + 1;
591 ret = do_readahead(mapping, f.file, start, len);