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>
22 * Initialise a struct file's readahead state. Assumes that the caller has
26 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
28 ra->ra_pages = mapping->backing_dev_info->ra_pages;
31 EXPORT_SYMBOL_GPL(file_ra_state_init);
33 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
36 * see if a page needs releasing upon read_cache_pages() failure
37 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
38 * before calling, such as the NFS fs marking pages that are cached locally
39 * on disk, thus we need to give the fs a chance to clean up in the event of
42 static void read_cache_pages_invalidate_page(struct address_space *mapping,
45 if (page_has_private(page)) {
46 if (!trylock_page(page))
48 page->mapping = mapping;
49 do_invalidatepage(page, 0);
53 page_cache_release(page);
57 * release a list of pages, invalidating them first if need be
59 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
60 struct list_head *pages)
64 while (!list_empty(pages)) {
65 victim = list_to_page(pages);
66 list_del(&victim->lru);
67 read_cache_pages_invalidate_page(mapping, victim);
72 * read_cache_pages - populate an address space with some pages & start reads against them
73 * @mapping: the address_space
74 * @pages: The address of a list_head which contains the target pages. These
75 * pages have their ->index populated and are otherwise uninitialised.
76 * @filler: callback routine for filling a single page.
77 * @data: private data for the callback routine.
79 * Hides the details of the LRU cache etc from the filesystems.
81 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
82 int (*filler)(void *, struct page *), void *data)
87 while (!list_empty(pages)) {
88 page = list_to_page(pages);
90 if (add_to_page_cache_lru(page, mapping,
91 page->index, GFP_KERNEL)) {
92 read_cache_pages_invalidate_page(mapping, page);
95 page_cache_release(page);
97 ret = filler(data, page);
99 read_cache_pages_invalidate_pages(mapping, pages);
102 task_io_account_read(PAGE_CACHE_SIZE);
107 EXPORT_SYMBOL(read_cache_pages);
109 static int read_pages(struct address_space *mapping, struct file *filp,
110 struct list_head *pages, unsigned nr_pages)
112 struct blk_plug plug;
116 blk_start_plug(&plug);
118 if (mapping->a_ops->readpages) {
119 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
120 /* Clean up the remaining pages */
121 put_pages_list(pages);
125 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
126 struct page *page = list_to_page(pages);
127 list_del(&page->lru);
128 if (!add_to_page_cache_lru(page, mapping,
129 page->index, GFP_KERNEL)) {
130 mapping->a_ops->readpage(filp, page);
132 page_cache_release(page);
137 blk_finish_plug(&plug);
143 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
144 * the pages first, then submits them all for I/O. This avoids the very bad
145 * behaviour which would occur if page allocations are causing VM writeback.
146 * We really don't want to intermingle reads and writes like that.
148 * Returns the number of pages requested, or the maximum amount of I/O allowed.
151 __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_CACHE_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);
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);
190 SetPageReadaheadUnused(page);
195 * Now start the IO. We ignore I/O errors - if the page is not
196 * uptodate then the caller will launch readpage again, and
197 * will then handle the error.
200 read_pages(mapping, filp, &page_pool, ret);
201 BUG_ON(!list_empty(&page_pool));
207 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
210 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
211 pgoff_t offset, unsigned long nr_to_read)
215 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
218 nr_to_read = max_sane_readahead(nr_to_read);
222 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
224 if (this_chunk > nr_to_read)
225 this_chunk = nr_to_read;
226 err = __do_page_cache_readahead(mapping, filp,
227 offset, this_chunk, 0);
233 offset += this_chunk;
234 nr_to_read -= this_chunk;
240 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
241 * sensible upper limit.
243 unsigned long max_sane_readahead(unsigned long nr)
245 return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
246 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
250 * Submit IO for the read-ahead request in file_ra_state.
252 unsigned long ra_submit(struct file_ra_state *ra,
253 struct address_space *mapping, struct file *filp)
257 actual = __do_page_cache_readahead(mapping, filp,
258 ra->start, ra->size, ra->async_size);
264 * Set the initial window size, round to next power of 2 and square
265 * for small size, x 4 for medium, and x 2 for large
266 * for 128k (32 page) max ra
267 * 1-8 page = 32k initial, > 8 page = 128k initial
269 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
271 unsigned long newsize = roundup_pow_of_two(size);
273 if (newsize <= max / 32)
274 newsize = newsize * 4;
275 else if (newsize <= max / 4)
276 newsize = newsize * 2;
284 * Get the previous window size, ramp it up, and
285 * return it as the new window size.
287 static unsigned long get_next_ra_size(struct file_ra_state *ra,
290 unsigned long cur = ra->size;
291 unsigned long newsize;
298 return min(newsize, max);
302 * On-demand readahead design.
304 * The fields in struct file_ra_state represent the most-recently-executed
307 * |<----- async_size ---------|
308 * |------------------- size -------------------->|
309 * |==================#===========================|
310 * ^start ^page marked with PG_readahead
312 * To overlap application thinking time and disk I/O time, we do
313 * `readahead pipelining': Do not wait until the application consumed all
314 * readahead pages and stalled on the missing page at readahead_index;
315 * Instead, submit an asynchronous readahead I/O as soon as there are
316 * only async_size pages left in the readahead window. Normally async_size
317 * will be equal to size, for maximum pipelining.
319 * In interleaved sequential reads, concurrent streams on the same fd can
320 * be invalidating each other's readahead state. So we flag the new readahead
321 * page at (start+size-async_size) with PG_readahead, and use it as readahead
322 * indicator. The flag won't be set on already cached pages, to avoid the
323 * readahead-for-nothing fuss, saving pointless page cache lookups.
325 * prev_pos tracks the last visited byte in the _previous_ read request.
326 * It should be maintained by the caller, and will be used for detecting
327 * small random reads. Note that the readahead algorithm checks loosely
328 * for sequential patterns. Hence interleaved reads might be served as
331 * There is a special-case: if the first page which the application tries to
332 * read happens to be the first page of the file, it is assumed that a linear
333 * read is about to happen and the window is immediately set to the initial size
334 * based on I/O request size and the max_readahead.
336 * The code ramps up the readahead size aggressively at first, but slow down as
337 * it approaches max_readhead.
341 * Count contiguously cached pages from @offset-1 to @offset-@max,
342 * this count is a conservative estimation of
343 * - length of the sequential read sequence, or
344 * - thrashing threshold in memory tight systems
346 static pgoff_t count_history_pages(struct address_space *mapping,
347 struct file_ra_state *ra,
348 pgoff_t offset, unsigned long max)
353 head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
356 return offset - 1 - head;
360 * page cache context based read-ahead
362 static int try_context_readahead(struct address_space *mapping,
363 struct file_ra_state *ra,
365 unsigned long req_size,
370 size = count_history_pages(mapping, ra, offset, max);
374 * it could be a random read
380 * starts from beginning of file:
381 * it is a strong indication of long-run stream (or whole-file-read)
387 ra->size = get_init_ra_size(size + req_size, max);
388 ra->async_size = ra->size;
394 * A minimal readahead algorithm for trivial sequential/random reads.
397 ondemand_readahead(struct address_space *mapping,
398 struct file_ra_state *ra, struct file *filp,
399 bool hit_readahead_marker, pgoff_t offset,
400 unsigned long req_size)
402 unsigned long max = max_sane_readahead(ra->ra_pages);
408 goto initial_readahead;
411 * It's the expected callback offset, assume sequential access.
412 * Ramp up sizes, and push forward the readahead window.
414 if ((offset == (ra->start + ra->size - ra->async_size) ||
415 offset == (ra->start + ra->size))) {
416 ra->start += ra->size;
417 ra->size = get_next_ra_size(ra, max);
418 ra->async_size = ra->size;
423 * Hit a marked page without valid readahead state.
424 * E.g. interleaved reads.
425 * Query the pagecache for async_size, which normally equals to
426 * readahead size. Ramp it up and use it as the new readahead size.
428 if (hit_readahead_marker) {
432 start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
435 if (!start || start - offset > max)
439 ra->size = start - offset; /* old async_size */
440 ra->size += req_size;
441 ra->size = get_next_ra_size(ra, max);
442 ra->async_size = ra->size;
450 goto initial_readahead;
453 * sequential cache miss
455 if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
456 goto initial_readahead;
459 * Query the page cache and look for the traces(cached history pages)
460 * that a sequential stream would leave behind.
462 if (try_context_readahead(mapping, ra, offset, req_size, max))
466 * standalone, small random read
467 * Read as is, and do not pollute the readahead state.
469 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
473 ra->size = get_init_ra_size(req_size, max);
474 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
478 * Will this read hit the readahead marker made by itself?
479 * If so, trigger the readahead marker hit now, and merge
480 * the resulted next readahead window into the current one.
482 if (offset == ra->start && ra->size == ra->async_size) {
483 ra->async_size = get_next_ra_size(ra, max);
484 ra->size += ra->async_size;
487 return ra_submit(ra, mapping, filp);
491 * page_cache_sync_readahead - generic file readahead
492 * @mapping: address_space which holds the pagecache and I/O vectors
493 * @ra: file_ra_state which holds the readahead state
494 * @filp: passed on to ->readpage() and ->readpages()
495 * @offset: start offset into @mapping, in pagecache page-sized units
496 * @req_size: hint: total size of the read which the caller is performing in
499 * page_cache_sync_readahead() should be called when a cache miss happened:
500 * it will submit the read. The readahead logic may decide to piggyback more
501 * pages onto the read request if access patterns suggest it will improve
504 void page_cache_sync_readahead(struct address_space *mapping,
505 struct file_ra_state *ra, struct file *filp,
506 pgoff_t offset, unsigned long req_size)
513 if (filp && (filp->f_mode & FMODE_RANDOM)) {
514 force_page_cache_readahead(mapping, filp, offset, req_size);
519 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
521 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
524 * page_cache_async_readahead - file readahead for marked pages
525 * @mapping: address_space which holds the pagecache and I/O vectors
526 * @ra: file_ra_state which holds the readahead state
527 * @filp: passed on to ->readpage() and ->readpages()
528 * @page: the page at @offset which has the PG_readahead flag set
529 * @offset: start offset into @mapping, in pagecache page-sized units
530 * @req_size: hint: total size of the read which the caller is performing in
533 * page_cache_async_readahead() should be called when a page is used which
534 * has the PG_readahead flag; this is a marker to suggest that the application
535 * has used up enough of the readahead window that we should start pulling in
539 page_cache_async_readahead(struct address_space *mapping,
540 struct file_ra_state *ra, struct file *filp,
541 struct page *page, pgoff_t offset,
542 unsigned long req_size)
549 * Same bit is used for PG_readahead and PG_reclaim.
551 if (PageWriteback(page))
554 ClearPageReadahead(page);
557 * Defer asynchronous read-ahead on IO congestion.
559 if (bdi_read_congested(mapping->backing_dev_info))
563 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
565 EXPORT_SYMBOL_GPL(page_cache_async_readahead);