2 * mm/readahead.c - address_space-level file readahead.
4 * Copyright (C) 2002, Linus Torvalds
6 * 09Apr2002 akpm@zip.com.au
10 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/blkdev.h>
15 #include <linux/backing-dev.h>
16 #include <linux/pagevec.h>
18 void default_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
21 EXPORT_SYMBOL(default_unplug_io_fn);
23 struct backing_dev_info default_backing_dev_info = {
24 .ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE,
26 .capabilities = BDI_CAP_MAP_COPY,
27 .unplug_io_fn = default_unplug_io_fn,
29 EXPORT_SYMBOL_GPL(default_backing_dev_info);
32 * Initialise a struct file's readahead state. Assumes that the caller has
36 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
38 ra->ra_pages = mapping->backing_dev_info->ra_pages;
41 EXPORT_SYMBOL_GPL(file_ra_state_init);
44 * Return max readahead size for this inode in number-of-pages.
46 static inline unsigned long get_max_readahead(struct file_ra_state *ra)
51 static inline unsigned long get_min_readahead(struct file_ra_state *ra)
53 return (VM_MIN_READAHEAD * 1024) / PAGE_CACHE_SIZE;
56 static inline void ra_off(struct file_ra_state *ra)
67 * Set the initial window size, round to next power of 2 and square
68 * for small size, x 4 for medium, and x 2 for large
69 * for 128k (32 page) max ra
70 * 1-8 page = 32k initial, > 8 page = 128k initial
72 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
74 unsigned long newsize = roundup_pow_of_two(size);
76 if (newsize <= max / 64)
77 newsize = newsize * newsize;
78 else if (newsize <= max / 4)
86 * Set the new window size, this is called only when I/O is to be submitted,
87 * not for each call to readahead. If a cache miss occured, reduce next I/O
88 * size, else increase depending on how close to max we are.
90 static inline unsigned long get_next_ra_size(struct file_ra_state *ra)
92 unsigned long max = get_max_readahead(ra);
93 unsigned long min = get_min_readahead(ra);
94 unsigned long cur = ra->size;
95 unsigned long newsize;
97 if (ra->flags & RA_FLAG_MISS) {
98 ra->flags &= ~RA_FLAG_MISS;
99 newsize = max((cur - 2), min);
100 } else if (cur < max / 16) {
105 return min(newsize, max);
108 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
111 * read_cache_pages - populate an address space with some pages, and
112 * start reads against them.
113 * @mapping: the address_space
114 * @pages: The address of a list_head which contains the target pages. These
115 * pages have their ->index populated and are otherwise uninitialised.
116 * @filler: callback routine for filling a single page.
117 * @data: private data for the callback routine.
119 * Hides the details of the LRU cache etc from the filesystems.
121 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
122 int (*filler)(void *, struct page *), void *data)
125 struct pagevec lru_pvec;
128 pagevec_init(&lru_pvec, 0);
130 while (!list_empty(pages)) {
131 page = list_to_page(pages);
132 list_del(&page->lru);
133 if (add_to_page_cache(page, mapping, page->index, GFP_KERNEL)) {
134 page_cache_release(page);
137 ret = filler(data, page);
138 if (!pagevec_add(&lru_pvec, page))
139 __pagevec_lru_add(&lru_pvec);
141 while (!list_empty(pages)) {
144 victim = list_to_page(pages);
145 list_del(&victim->lru);
146 page_cache_release(victim);
151 pagevec_lru_add(&lru_pvec);
155 EXPORT_SYMBOL(read_cache_pages);
157 static int read_pages(struct address_space *mapping, struct file *filp,
158 struct list_head *pages, unsigned nr_pages)
161 struct pagevec lru_pvec;
164 if (mapping->a_ops->readpages) {
165 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
169 pagevec_init(&lru_pvec, 0);
170 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
171 struct page *page = list_to_page(pages);
172 list_del(&page->lru);
173 if (!add_to_page_cache(page, mapping,
174 page->index, GFP_KERNEL)) {
175 ret = mapping->a_ops->readpage(filp, page);
176 if (ret != AOP_TRUNCATED_PAGE) {
177 if (!pagevec_add(&lru_pvec, page))
178 __pagevec_lru_add(&lru_pvec);
180 } /* else fall through to release */
182 page_cache_release(page);
184 pagevec_lru_add(&lru_pvec);
193 * The fields in struct file_ra_state represent the most-recently-executed
196 * start: Page index at which we started the readahead
197 * size: Number of pages in that read
198 * Together, these form the "current window".
199 * Together, start and size represent the `readahead window'.
200 * prev_page: The page which the readahead algorithm most-recently inspected.
201 * It is mainly used to detect sequential file reading.
202 * If page_cache_readahead sees that it is again being called for
203 * a page which it just looked at, it can return immediately without
204 * making any state changes.
206 * ahead_size: Together, these form the "ahead window".
207 * ra_pages: The externally controlled max readahead for this fd.
209 * When readahead is in the off state (size == 0), readahead is disabled.
210 * In this state, prev_page is used to detect the resumption of sequential I/O.
212 * The readahead code manages two windows - the "current" and the "ahead"
213 * windows. The intent is that while the application is walking the pages
214 * in the current window, I/O is underway on the ahead window. When the
215 * current window is fully traversed, it is replaced by the ahead window
216 * and the ahead window is invalidated. When this copying happens, the
217 * new current window's pages are probably still locked. So
218 * we submit a new batch of I/O immediately, creating a new ahead window.
222 * ----|----------------|----------------|-----
224 * ^ahead_start ^ahead_start+ahead_size
226 * ^ When this page is read, we submit I/O for the
229 * A `readahead hit' occurs when a read request is made against a page which is
230 * the next sequential page. Ahead window calculations are done only when it
231 * is time to submit a new IO. The code ramps up the size agressively at first,
232 * but slow down as it approaches max_readhead.
234 * Any seek/ramdom IO will result in readahead being turned off. It will resume
235 * at the first sequential access.
237 * There is a special-case: if the first page which the application tries to
238 * read happens to be the first page of the file, it is assumed that a linear
239 * read is about to happen and the window is immediately set to the initial size
240 * based on I/O request size and the max_readahead.
242 * This function is to be called for every read request, rather than when
243 * it is time to perform readahead. It is called only once for the entire I/O
244 * regardless of size unless readahead is unable to start enough I/O to satisfy
245 * the request (I/O request > max_readahead).
249 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
250 * the pages first, then submits them all for I/O. This avoids the very bad
251 * behaviour which would occur if page allocations are causing VM writeback.
252 * We really don't want to intermingle reads and writes like that.
254 * Returns the number of pages requested, or the maximum amount of I/O allowed.
256 * do_page_cache_readahead() returns -1 if it encountered request queue
260 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
261 pgoff_t offset, unsigned long nr_to_read)
263 struct inode *inode = mapping->host;
265 unsigned long end_index; /* The last page we want to read */
266 LIST_HEAD(page_pool);
269 loff_t isize = i_size_read(inode);
274 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
277 * Preallocate as many pages as we will need.
279 read_lock_irq(&mapping->tree_lock);
280 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
281 pgoff_t page_offset = offset + page_idx;
283 if (page_offset > end_index)
286 page = radix_tree_lookup(&mapping->page_tree, page_offset);
290 read_unlock_irq(&mapping->tree_lock);
291 page = page_cache_alloc_cold(mapping);
292 read_lock_irq(&mapping->tree_lock);
295 page->index = page_offset;
296 list_add(&page->lru, &page_pool);
299 read_unlock_irq(&mapping->tree_lock);
302 * Now start the IO. We ignore I/O errors - if the page is not
303 * uptodate then the caller will launch readpage again, and
304 * will then handle the error.
307 read_pages(mapping, filp, &page_pool, ret);
308 BUG_ON(!list_empty(&page_pool));
314 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
317 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
318 pgoff_t offset, unsigned long nr_to_read)
322 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
328 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
330 if (this_chunk > nr_to_read)
331 this_chunk = nr_to_read;
332 err = __do_page_cache_readahead(mapping, filp,
339 offset += this_chunk;
340 nr_to_read -= this_chunk;
346 * Check how effective readahead is being. If the amount of started IO is
347 * less than expected then the file is partly or fully in pagecache and
348 * readahead isn't helping.
351 static inline int check_ra_success(struct file_ra_state *ra,
352 unsigned long nr_to_read, unsigned long actual)
355 ra->cache_hit += nr_to_read;
356 if (ra->cache_hit >= VM_MAX_CACHE_HIT) {
358 ra->flags |= RA_FLAG_INCACHE;
368 * This version skips the IO if the queue is read-congested, and will tell the
369 * block layer to abandon the readahead if request allocation would block.
371 * force_page_cache_readahead() will ignore queue congestion and will block on
374 int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
375 pgoff_t offset, unsigned long nr_to_read)
377 if (bdi_read_congested(mapping->backing_dev_info))
380 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
384 * Read 'nr_to_read' pages starting at page 'offset'. If the flag 'block'
385 * is set wait till the read completes. Otherwise attempt to read without
387 * Returns 1 meaning 'success' if read is succesfull without switching off
388 * readhaead mode. Otherwise return failure.
391 blockable_page_cache_readahead(struct address_space *mapping, struct file *filp,
392 pgoff_t offset, unsigned long nr_to_read,
393 struct file_ra_state *ra, int block)
397 if (!block && bdi_read_congested(mapping->backing_dev_info))
400 actual = __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
402 return check_ra_success(ra, nr_to_read, actual);
405 static int make_ahead_window(struct address_space *mapping, struct file *filp,
406 struct file_ra_state *ra, int force)
410 ra->ahead_size = get_next_ra_size(ra);
411 ra->ahead_start = ra->start + ra->size;
413 block = force || (ra->prev_page >= ra->ahead_start);
414 ret = blockable_page_cache_readahead(mapping, filp,
415 ra->ahead_start, ra->ahead_size, ra, block);
417 if (!ret && !force) {
418 /* A read failure in blocking mode, implies pages are
419 * all cached. So we can safely assume we have taken
420 * care of all the pages requested in this call.
421 * A read failure in non-blocking mode, implies we are
422 * reading more pages than requested in this call. So
423 * we safely assume we have taken care of all the pages
424 * requested in this call.
426 * Just reset the ahead window in case we failed due to
427 * congestion. The ahead window will any way be closed
428 * in case we failed due to excessive page cache hits.
438 * page_cache_readahead - generic adaptive readahead
439 * @mapping: address_space which holds the pagecache and I/O vectors
440 * @ra: file_ra_state which holds the readahead state
441 * @filp: passed on to ->readpage() and ->readpages()
442 * @offset: start offset into @mapping, in PAGE_CACHE_SIZE units
443 * @req_size: hint: total size of the read which the caller is performing in
444 * PAGE_CACHE_SIZE units
446 * page_cache_readahead() is the main function. If performs the adaptive
447 * readahead window size management and submits the readahead I/O.
449 * Note that @filp is purely used for passing on to the ->readpage[s]()
450 * handler: it may refer to a different file from @mapping (so we may not use
451 * @filp->f_mapping or @filp->f_dentry->d_inode here).
452 * Also, @ra may not be equal to &@filp->f_ra.
456 page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
457 struct file *filp, pgoff_t offset, unsigned long req_size)
459 unsigned long max, newsize;
463 * We avoid doing extra work and bogusly perturbing the readahead
464 * window expansion logic.
466 if (offset == ra->prev_page && --req_size)
469 /* Note that prev_page == -1 if it is a first read */
470 sequential = (offset == ra->prev_page + 1);
471 ra->prev_page = offset;
473 max = get_max_readahead(ra);
474 newsize = min(req_size, max);
476 /* No readahead or sub-page sized read or file already in cache */
477 if (newsize == 0 || (ra->flags & RA_FLAG_INCACHE))
480 ra->prev_page += newsize - 1;
483 * Special case - first read at start of file. We'll assume it's
484 * a whole-file read and grow the window fast. Or detect first
487 if (sequential && ra->size == 0) {
488 ra->size = get_init_ra_size(newsize, max);
490 if (!blockable_page_cache_readahead(mapping, filp, offset,
495 * If the request size is larger than our max readahead, we
496 * at least want to be sure that we get 2 IOs in flight and
497 * we know that we will definitly need the new I/O.
498 * once we do this, subsequent calls should be able to overlap
499 * IOs,* thus preventing stalls. so issue the ahead window
503 make_ahead_window(mapping, filp, ra, 1);
509 * Now handle the random case:
510 * partial page reads and first access were handled above,
511 * so this must be the next page otherwise it is random
515 blockable_page_cache_readahead(mapping, filp, offset,
521 * If we get here we are doing sequential IO and this was not the first
522 * occurence (ie we have an existing window)
525 if (ra->ahead_start == 0) { /* no ahead window yet */
526 if (!make_ahead_window(mapping, filp, ra, 0))
530 * Already have an ahead window, check if we crossed into it.
531 * If so, shift windows and issue a new ahead window.
532 * Only return the #pages that are in the current window, so that
533 * we get called back on the first page of the ahead window which
534 * will allow us to submit more IO.
536 if (ra->prev_page >= ra->ahead_start) {
537 ra->start = ra->ahead_start;
538 ra->size = ra->ahead_size;
539 make_ahead_window(mapping, filp, ra, 0);
543 return ra->prev_page + 1;
547 * handle_ra_miss() is called when it is known that a page which should have
548 * been present in the pagecache (we just did some readahead there) was in fact
549 * not found. This will happen if it was evicted by the VM (readahead
552 * Turn on the cache miss flag in the RA struct, this will cause the RA code
553 * to reduce the RA size on the next read.
555 void handle_ra_miss(struct address_space *mapping,
556 struct file_ra_state *ra, pgoff_t offset)
558 ra->flags |= RA_FLAG_MISS;
559 ra->flags &= ~RA_FLAG_INCACHE;
564 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
565 * sensible upper limit.
567 unsigned long max_sane_readahead(unsigned long nr)
569 unsigned long active;
570 unsigned long inactive;
573 __get_zone_counts(&active, &inactive, &free, NODE_DATA(numa_node_id()));
574 return min(nr, (inactive + free) / 2);