blob: aa7ec424656ab5708307c7731b269f925982fa95 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/readahead.c - address_space-level file readahead.
3 *
4 * Copyright (C) 2002, Linus Torvalds
5 *
6 * 09Apr2002 akpm@zip.com.au
7 * Initial version.
8 */
9
10#include <linux/kernel.h>
11#include <linux/fs.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/blkdev.h>
15#include <linux/backing-dev.h>
16#include <linux/pagevec.h>
17
18void default_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
19{
20}
21EXPORT_SYMBOL(default_unplug_io_fn);
22
23struct backing_dev_info default_backing_dev_info = {
24 .ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE,
25 .state = 0,
26 .capabilities = BDI_CAP_MAP_COPY,
27 .unplug_io_fn = default_unplug_io_fn,
28};
29EXPORT_SYMBOL_GPL(default_backing_dev_info);
30
31/*
32 * Initialise a struct file's readahead state. Assumes that the caller has
33 * memset *ra to zero.
34 */
35void
36file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
37{
38 ra->ra_pages = mapping->backing_dev_info->ra_pages;
39 ra->prev_page = -1;
40}
41
42/*
43 * Return max readahead size for this inode in number-of-pages.
44 */
45static inline unsigned long get_max_readahead(struct file_ra_state *ra)
46{
47 return ra->ra_pages;
48}
49
50static inline unsigned long get_min_readahead(struct file_ra_state *ra)
51{
52 return (VM_MIN_READAHEAD * 1024) / PAGE_CACHE_SIZE;
53}
54
Oleg Nesterova564da32006-03-22 00:08:47 -080055static inline void reset_ahead_window(struct file_ra_state *ra)
56{
57 /*
58 * ... but preserve ahead_start + ahead_size value,
59 * see 'recheck:' label in page_cache_readahead().
60 * Note: We never use ->ahead_size as rvalue without
61 * checking ->ahead_start != 0 first.
62 */
63 ra->ahead_size += ra->ahead_start;
64 ra->ahead_start = 0;
65}
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067static inline void ra_off(struct file_ra_state *ra)
68{
69 ra->start = 0;
70 ra->flags = 0;
71 ra->size = 0;
Oleg Nesterova564da32006-03-22 00:08:47 -080072 reset_ahead_window(ra);
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 return;
74}
75
76/*
77 * Set the initial window size, round to next power of 2 and square
78 * for small size, x 4 for medium, and x 2 for large
79 * for 128k (32 page) max ra
80 * 1-8 page = 32k initial, > 8 page = 128k initial
81 */
82static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
83{
84 unsigned long newsize = roundup_pow_of_two(size);
85
Steven Prattaed75ff2006-03-22 00:08:48 -080086 if (newsize <= max / 32)
87 newsize = newsize * 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 else if (newsize <= max / 4)
Steven Prattaed75ff2006-03-22 00:08:48 -080089 newsize = newsize * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 else
91 newsize = max;
92 return newsize;
93}
94
95/*
96 * Set the new window size, this is called only when I/O is to be submitted,
97 * not for each call to readahead. If a cache miss occured, reduce next I/O
98 * size, else increase depending on how close to max we are.
99 */
100static inline unsigned long get_next_ra_size(struct file_ra_state *ra)
101{
102 unsigned long max = get_max_readahead(ra);
103 unsigned long min = get_min_readahead(ra);
104 unsigned long cur = ra->size;
105 unsigned long newsize;
106
107 if (ra->flags & RA_FLAG_MISS) {
108 ra->flags &= ~RA_FLAG_MISS;
109 newsize = max((cur - 2), min);
110 } else if (cur < max / 16) {
111 newsize = 4 * cur;
112 } else {
113 newsize = 2 * cur;
114 }
115 return min(newsize, max);
116}
117
118#define list_to_page(head) (list_entry((head)->prev, struct page, lru))
119
120/**
Randy Dunlapbd40cdd2006-06-25 05:48:08 -0700121 * read_cache_pages - populate an address space with some pages & start reads against them
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 * @mapping: the address_space
123 * @pages: The address of a list_head which contains the target pages. These
124 * pages have their ->index populated and are otherwise uninitialised.
125 * @filler: callback routine for filling a single page.
126 * @data: private data for the callback routine.
127 *
128 * Hides the details of the LRU cache etc from the filesystems.
129 */
130int read_cache_pages(struct address_space *mapping, struct list_head *pages,
131 int (*filler)(void *, struct page *), void *data)
132{
133 struct page *page;
134 struct pagevec lru_pvec;
135 int ret = 0;
136
137 pagevec_init(&lru_pvec, 0);
138
139 while (!list_empty(pages)) {
140 page = list_to_page(pages);
141 list_del(&page->lru);
142 if (add_to_page_cache(page, mapping, page->index, GFP_KERNEL)) {
143 page_cache_release(page);
144 continue;
145 }
146 ret = filler(data, page);
147 if (!pagevec_add(&lru_pvec, page))
148 __pagevec_lru_add(&lru_pvec);
149 if (ret) {
150 while (!list_empty(pages)) {
151 struct page *victim;
152
153 victim = list_to_page(pages);
154 list_del(&victim->lru);
155 page_cache_release(victim);
156 }
157 break;
158 }
159 }
160 pagevec_lru_add(&lru_pvec);
161 return ret;
162}
163
164EXPORT_SYMBOL(read_cache_pages);
165
166static int read_pages(struct address_space *mapping, struct file *filp,
167 struct list_head *pages, unsigned nr_pages)
168{
169 unsigned page_idx;
170 struct pagevec lru_pvec;
Zach Brown994fc28c2005-12-15 14:28:17 -0800171 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
173 if (mapping->a_ops->readpages) {
174 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
175 goto out;
176 }
177
178 pagevec_init(&lru_pvec, 0);
179 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
180 struct page *page = list_to_page(pages);
181 list_del(&page->lru);
182 if (!add_to_page_cache(page, mapping,
183 page->index, GFP_KERNEL)) {
Zach Brown9f1a3cf2006-06-25 05:46:46 -0700184 mapping->a_ops->readpage(filp, page);
185 if (!pagevec_add(&lru_pvec, page))
186 __pagevec_lru_add(&lru_pvec);
187 } else
188 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 }
190 pagevec_lru_add(&lru_pvec);
Zach Brown994fc28c2005-12-15 14:28:17 -0800191 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192out:
193 return ret;
194}
195
196/*
197 * Readahead design.
198 *
199 * The fields in struct file_ra_state represent the most-recently-executed
200 * readahead attempt:
201 *
202 * start: Page index at which we started the readahead
203 * size: Number of pages in that read
204 * Together, these form the "current window".
205 * Together, start and size represent the `readahead window'.
206 * prev_page: The page which the readahead algorithm most-recently inspected.
207 * It is mainly used to detect sequential file reading.
208 * If page_cache_readahead sees that it is again being called for
209 * a page which it just looked at, it can return immediately without
210 * making any state changes.
211 * ahead_start,
212 * ahead_size: Together, these form the "ahead window".
213 * ra_pages: The externally controlled max readahead for this fd.
214 *
215 * When readahead is in the off state (size == 0), readahead is disabled.
216 * In this state, prev_page is used to detect the resumption of sequential I/O.
217 *
218 * The readahead code manages two windows - the "current" and the "ahead"
219 * windows. The intent is that while the application is walking the pages
220 * in the current window, I/O is underway on the ahead window. When the
221 * current window is fully traversed, it is replaced by the ahead window
222 * and the ahead window is invalidated. When this copying happens, the
223 * new current window's pages are probably still locked. So
224 * we submit a new batch of I/O immediately, creating a new ahead window.
225 *
226 * So:
227 *
228 * ----|----------------|----------------|-----
229 * ^start ^start+size
230 * ^ahead_start ^ahead_start+ahead_size
231 *
232 * ^ When this page is read, we submit I/O for the
233 * ahead window.
234 *
235 * A `readahead hit' occurs when a read request is made against a page which is
236 * the next sequential page. Ahead window calculations are done only when it
237 * is time to submit a new IO. The code ramps up the size agressively at first,
238 * but slow down as it approaches max_readhead.
239 *
240 * Any seek/ramdom IO will result in readahead being turned off. It will resume
241 * at the first sequential access.
242 *
243 * There is a special-case: if the first page which the application tries to
244 * read happens to be the first page of the file, it is assumed that a linear
245 * read is about to happen and the window is immediately set to the initial size
246 * based on I/O request size and the max_readahead.
247 *
248 * This function is to be called for every read request, rather than when
249 * it is time to perform readahead. It is called only once for the entire I/O
250 * regardless of size unless readahead is unable to start enough I/O to satisfy
251 * the request (I/O request > max_readahead).
252 */
253
254/*
255 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
256 * the pages first, then submits them all for I/O. This avoids the very bad
257 * behaviour which would occur if page allocations are causing VM writeback.
258 * We really don't want to intermingle reads and writes like that.
259 *
260 * Returns the number of pages requested, or the maximum amount of I/O allowed.
261 *
262 * do_page_cache_readahead() returns -1 if it encountered request queue
263 * congestion.
264 */
265static int
266__do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800267 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
269 struct inode *inode = mapping->host;
270 struct page *page;
271 unsigned long end_index; /* The last page we want to read */
272 LIST_HEAD(page_pool);
273 int page_idx;
274 int ret = 0;
275 loff_t isize = i_size_read(inode);
276
277 if (isize == 0)
278 goto out;
279
280 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
281
282 /*
283 * Preallocate as many pages as we will need.
284 */
285 read_lock_irq(&mapping->tree_lock);
286 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
Andrew Morton7361f4d2005-11-07 00:59:28 -0800287 pgoff_t page_offset = offset + page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
289 if (page_offset > end_index)
290 break;
291
292 page = radix_tree_lookup(&mapping->page_tree, page_offset);
293 if (page)
294 continue;
295
296 read_unlock_irq(&mapping->tree_lock);
297 page = page_cache_alloc_cold(mapping);
298 read_lock_irq(&mapping->tree_lock);
299 if (!page)
300 break;
301 page->index = page_offset;
302 list_add(&page->lru, &page_pool);
303 ret++;
304 }
305 read_unlock_irq(&mapping->tree_lock);
306
307 /*
308 * Now start the IO. We ignore I/O errors - if the page is not
309 * uptodate then the caller will launch readpage again, and
310 * will then handle the error.
311 */
312 if (ret)
313 read_pages(mapping, filp, &page_pool, ret);
314 BUG_ON(!list_empty(&page_pool));
315out:
316 return ret;
317}
318
319/*
320 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
321 * memory at once.
322 */
323int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800324 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
326 int ret = 0;
327
328 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
329 return -EINVAL;
330
331 while (nr_to_read) {
332 int err;
333
334 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
335
336 if (this_chunk > nr_to_read)
337 this_chunk = nr_to_read;
338 err = __do_page_cache_readahead(mapping, filp,
339 offset, this_chunk);
340 if (err < 0) {
341 ret = err;
342 break;
343 }
344 ret += err;
345 offset += this_chunk;
346 nr_to_read -= this_chunk;
347 }
348 return ret;
349}
350
351/*
352 * Check how effective readahead is being. If the amount of started IO is
353 * less than expected then the file is partly or fully in pagecache and
354 * readahead isn't helping.
355 *
356 */
357static inline int check_ra_success(struct file_ra_state *ra,
358 unsigned long nr_to_read, unsigned long actual)
359{
360 if (actual == 0) {
361 ra->cache_hit += nr_to_read;
362 if (ra->cache_hit >= VM_MAX_CACHE_HIT) {
363 ra_off(ra);
364 ra->flags |= RA_FLAG_INCACHE;
365 return 0;
366 }
367 } else {
368 ra->cache_hit=0;
369 }
370 return 1;
371}
372
373/*
374 * This version skips the IO if the queue is read-congested, and will tell the
375 * block layer to abandon the readahead if request allocation would block.
376 *
377 * force_page_cache_readahead() will ignore queue congestion and will block on
378 * request queues.
379 */
380int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800381 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382{
383 if (bdi_read_congested(mapping->backing_dev_info))
384 return -1;
385
386 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
387}
388
389/*
390 * Read 'nr_to_read' pages starting at page 'offset'. If the flag 'block'
391 * is set wait till the read completes. Otherwise attempt to read without
392 * blocking.
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200393 * Returns 1 meaning 'success' if read is successful without switching off
394 * readahead mode. Otherwise return failure.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 */
396static int
397blockable_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800398 pgoff_t offset, unsigned long nr_to_read,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 struct file_ra_state *ra, int block)
400{
401 int actual;
402
403 if (!block && bdi_read_congested(mapping->backing_dev_info))
404 return 0;
405
406 actual = __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
407
408 return check_ra_success(ra, nr_to_read, actual);
409}
410
411static int make_ahead_window(struct address_space *mapping, struct file *filp,
412 struct file_ra_state *ra, int force)
413{
414 int block, ret;
415
416 ra->ahead_size = get_next_ra_size(ra);
417 ra->ahead_start = ra->start + ra->size;
418
419 block = force || (ra->prev_page >= ra->ahead_start);
420 ret = blockable_page_cache_readahead(mapping, filp,
421 ra->ahead_start, ra->ahead_size, ra, block);
422
423 if (!ret && !force) {
424 /* A read failure in blocking mode, implies pages are
425 * all cached. So we can safely assume we have taken
426 * care of all the pages requested in this call.
427 * A read failure in non-blocking mode, implies we are
428 * reading more pages than requested in this call. So
429 * we safely assume we have taken care of all the pages
430 * requested in this call.
431 *
432 * Just reset the ahead window in case we failed due to
433 * congestion. The ahead window will any way be closed
434 * in case we failed due to excessive page cache hits.
435 */
Oleg Nesterova564da32006-03-22 00:08:47 -0800436 reset_ahead_window(ra);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 }
438
439 return ret;
440}
441
Andrew Morton7361f4d2005-11-07 00:59:28 -0800442/**
443 * page_cache_readahead - generic adaptive readahead
444 * @mapping: address_space which holds the pagecache and I/O vectors
445 * @ra: file_ra_state which holds the readahead state
446 * @filp: passed on to ->readpage() and ->readpages()
447 * @offset: start offset into @mapping, in PAGE_CACHE_SIZE units
448 * @req_size: hint: total size of the read which the caller is performing in
449 * PAGE_CACHE_SIZE units
450 *
451 * page_cache_readahead() is the main function. If performs the adaptive
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 * readahead window size management and submits the readahead I/O.
Andrew Morton7361f4d2005-11-07 00:59:28 -0800453 *
454 * Note that @filp is purely used for passing on to the ->readpage[s]()
455 * handler: it may refer to a different file from @mapping (so we may not use
456 * @filp->f_mapping or @filp->f_dentry->d_inode here).
457 * Also, @ra may not be equal to &@filp->f_ra.
458 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 */
460unsigned long
461page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800462 struct file *filp, pgoff_t offset, unsigned long req_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
464 unsigned long max, newsize;
465 int sequential;
466
467 /*
468 * We avoid doing extra work and bogusly perturbing the readahead
469 * window expansion logic.
470 */
471 if (offset == ra->prev_page && --req_size)
472 ++offset;
473
474 /* Note that prev_page == -1 if it is a first read */
475 sequential = (offset == ra->prev_page + 1);
476 ra->prev_page = offset;
477
478 max = get_max_readahead(ra);
479 newsize = min(req_size, max);
480
481 /* No readahead or sub-page sized read or file already in cache */
482 if (newsize == 0 || (ra->flags & RA_FLAG_INCACHE))
483 goto out;
484
485 ra->prev_page += newsize - 1;
486
487 /*
488 * Special case - first read at start of file. We'll assume it's
489 * a whole-file read and grow the window fast. Or detect first
490 * sequential access
491 */
492 if (sequential && ra->size == 0) {
493 ra->size = get_init_ra_size(newsize, max);
494 ra->start = offset;
495 if (!blockable_page_cache_readahead(mapping, filp, offset,
496 ra->size, ra, 1))
497 goto out;
498
499 /*
500 * If the request size is larger than our max readahead, we
501 * at least want to be sure that we get 2 IOs in flight and
502 * we know that we will definitly need the new I/O.
503 * once we do this, subsequent calls should be able to overlap
504 * IOs,* thus preventing stalls. so issue the ahead window
505 * immediately.
506 */
507 if (req_size >= max)
508 make_ahead_window(mapping, filp, ra, 1);
509
510 goto out;
511 }
512
513 /*
514 * Now handle the random case:
515 * partial page reads and first access were handled above,
516 * so this must be the next page otherwise it is random
517 */
518 if (!sequential) {
519 ra_off(ra);
520 blockable_page_cache_readahead(mapping, filp, offset,
521 newsize, ra, 1);
522 goto out;
523 }
524
525 /*
526 * If we get here we are doing sequential IO and this was not the first
527 * occurence (ie we have an existing window)
528 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 if (ra->ahead_start == 0) { /* no ahead window yet */
530 if (!make_ahead_window(mapping, filp, ra, 0))
Oleg Nesterova564da32006-03-22 00:08:47 -0800531 goto recheck;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 }
Oleg Nesterova564da32006-03-22 00:08:47 -0800533
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 /*
535 * Already have an ahead window, check if we crossed into it.
536 * If so, shift windows and issue a new ahead window.
537 * Only return the #pages that are in the current window, so that
538 * we get called back on the first page of the ahead window which
539 * will allow us to submit more IO.
540 */
541 if (ra->prev_page >= ra->ahead_start) {
542 ra->start = ra->ahead_start;
543 ra->size = ra->ahead_size;
544 make_ahead_window(mapping, filp, ra, 0);
Oleg Nesterova564da32006-03-22 00:08:47 -0800545recheck:
546 /* prev_page shouldn't overrun the ahead window */
547 ra->prev_page = min(ra->prev_page,
548 ra->ahead_start + ra->ahead_size - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 }
550
551out:
552 return ra->prev_page + 1;
553}
Andrew Mortond8733c22006-03-23 03:00:11 -0800554EXPORT_SYMBOL_GPL(page_cache_readahead);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556/*
557 * handle_ra_miss() is called when it is known that a page which should have
558 * been present in the pagecache (we just did some readahead there) was in fact
559 * not found. This will happen if it was evicted by the VM (readahead
560 * thrashing)
561 *
562 * Turn on the cache miss flag in the RA struct, this will cause the RA code
563 * to reduce the RA size on the next read.
564 */
565void handle_ra_miss(struct address_space *mapping,
566 struct file_ra_state *ra, pgoff_t offset)
567{
568 ra->flags |= RA_FLAG_MISS;
569 ra->flags &= ~RA_FLAG_INCACHE;
Steven Pratt3b30bbd2005-09-06 15:17:06 -0700570 ra->cache_hit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571}
572
573/*
574 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
575 * sensible upper limit.
576 */
577unsigned long max_sane_readahead(unsigned long nr)
578{
579 unsigned long active;
580 unsigned long inactive;
581 unsigned long free;
582
583 __get_zone_counts(&active, &inactive, &free, NODE_DATA(numa_node_id()));
584 return min(nr, (inactive + free) / 2);
585}