blob: 0f142a40984b1674228e067f491159d42b13f32d [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/**
121 * read_cache_pages - populate an address space with some pages, and
122 * start reads against them.
123 * @mapping: the address_space
124 * @pages: The address of a list_head which contains the target pages. These
125 * pages have their ->index populated and are otherwise uninitialised.
126 * @filler: callback routine for filling a single page.
127 * @data: private data for the callback routine.
128 *
129 * Hides the details of the LRU cache etc from the filesystems.
130 */
131int read_cache_pages(struct address_space *mapping, struct list_head *pages,
132 int (*filler)(void *, struct page *), void *data)
133{
134 struct page *page;
135 struct pagevec lru_pvec;
136 int ret = 0;
137
138 pagevec_init(&lru_pvec, 0);
139
140 while (!list_empty(pages)) {
141 page = list_to_page(pages);
142 list_del(&page->lru);
143 if (add_to_page_cache(page, mapping, page->index, GFP_KERNEL)) {
144 page_cache_release(page);
145 continue;
146 }
147 ret = filler(data, page);
148 if (!pagevec_add(&lru_pvec, page))
149 __pagevec_lru_add(&lru_pvec);
150 if (ret) {
151 while (!list_empty(pages)) {
152 struct page *victim;
153
154 victim = list_to_page(pages);
155 list_del(&victim->lru);
156 page_cache_release(victim);
157 }
158 break;
159 }
160 }
161 pagevec_lru_add(&lru_pvec);
162 return ret;
163}
164
165EXPORT_SYMBOL(read_cache_pages);
166
167static int read_pages(struct address_space *mapping, struct file *filp,
168 struct list_head *pages, unsigned nr_pages)
169{
170 unsigned page_idx;
171 struct pagevec lru_pvec;
Zach Brown994fc28c2005-12-15 14:28:17 -0800172 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173
174 if (mapping->a_ops->readpages) {
175 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
176 goto out;
177 }
178
179 pagevec_init(&lru_pvec, 0);
180 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
181 struct page *page = list_to_page(pages);
182 list_del(&page->lru);
183 if (!add_to_page_cache(page, mapping,
184 page->index, GFP_KERNEL)) {
Zach Brown994fc28c2005-12-15 14:28:17 -0800185 ret = mapping->a_ops->readpage(filp, page);
186 if (ret != AOP_TRUNCATED_PAGE) {
187 if (!pagevec_add(&lru_pvec, page))
188 __pagevec_lru_add(&lru_pvec);
189 continue;
190 } /* else fall through to release */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 }
Zach Brown994fc28c2005-12-15 14:28:17 -0800192 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 }
194 pagevec_lru_add(&lru_pvec);
Zach Brown994fc28c2005-12-15 14:28:17 -0800195 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196out:
197 return ret;
198}
199
200/*
201 * Readahead design.
202 *
203 * The fields in struct file_ra_state represent the most-recently-executed
204 * readahead attempt:
205 *
206 * start: Page index at which we started the readahead
207 * size: Number of pages in that read
208 * Together, these form the "current window".
209 * Together, start and size represent the `readahead window'.
210 * prev_page: The page which the readahead algorithm most-recently inspected.
211 * It is mainly used to detect sequential file reading.
212 * If page_cache_readahead sees that it is again being called for
213 * a page which it just looked at, it can return immediately without
214 * making any state changes.
215 * ahead_start,
216 * ahead_size: Together, these form the "ahead window".
217 * ra_pages: The externally controlled max readahead for this fd.
218 *
219 * When readahead is in the off state (size == 0), readahead is disabled.
220 * In this state, prev_page is used to detect the resumption of sequential I/O.
221 *
222 * The readahead code manages two windows - the "current" and the "ahead"
223 * windows. The intent is that while the application is walking the pages
224 * in the current window, I/O is underway on the ahead window. When the
225 * current window is fully traversed, it is replaced by the ahead window
226 * and the ahead window is invalidated. When this copying happens, the
227 * new current window's pages are probably still locked. So
228 * we submit a new batch of I/O immediately, creating a new ahead window.
229 *
230 * So:
231 *
232 * ----|----------------|----------------|-----
233 * ^start ^start+size
234 * ^ahead_start ^ahead_start+ahead_size
235 *
236 * ^ When this page is read, we submit I/O for the
237 * ahead window.
238 *
239 * A `readahead hit' occurs when a read request is made against a page which is
240 * the next sequential page. Ahead window calculations are done only when it
241 * is time to submit a new IO. The code ramps up the size agressively at first,
242 * but slow down as it approaches max_readhead.
243 *
244 * Any seek/ramdom IO will result in readahead being turned off. It will resume
245 * at the first sequential access.
246 *
247 * There is a special-case: if the first page which the application tries to
248 * read happens to be the first page of the file, it is assumed that a linear
249 * read is about to happen and the window is immediately set to the initial size
250 * based on I/O request size and the max_readahead.
251 *
252 * This function is to be called for every read request, rather than when
253 * it is time to perform readahead. It is called only once for the entire I/O
254 * regardless of size unless readahead is unable to start enough I/O to satisfy
255 * the request (I/O request > max_readahead).
256 */
257
258/*
259 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
260 * the pages first, then submits them all for I/O. This avoids the very bad
261 * behaviour which would occur if page allocations are causing VM writeback.
262 * We really don't want to intermingle reads and writes like that.
263 *
264 * Returns the number of pages requested, or the maximum amount of I/O allowed.
265 *
266 * do_page_cache_readahead() returns -1 if it encountered request queue
267 * congestion.
268 */
269static int
270__do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800271 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272{
273 struct inode *inode = mapping->host;
274 struct page *page;
275 unsigned long end_index; /* The last page we want to read */
276 LIST_HEAD(page_pool);
277 int page_idx;
278 int ret = 0;
279 loff_t isize = i_size_read(inode);
280
281 if (isize == 0)
282 goto out;
283
284 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
285
286 /*
287 * Preallocate as many pages as we will need.
288 */
289 read_lock_irq(&mapping->tree_lock);
290 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
Andrew Morton7361f4d2005-11-07 00:59:28 -0800291 pgoff_t page_offset = offset + page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
293 if (page_offset > end_index)
294 break;
295
296 page = radix_tree_lookup(&mapping->page_tree, page_offset);
297 if (page)
298 continue;
299
300 read_unlock_irq(&mapping->tree_lock);
301 page = page_cache_alloc_cold(mapping);
302 read_lock_irq(&mapping->tree_lock);
303 if (!page)
304 break;
305 page->index = page_offset;
306 list_add(&page->lru, &page_pool);
307 ret++;
308 }
309 read_unlock_irq(&mapping->tree_lock);
310
311 /*
312 * Now start the IO. We ignore I/O errors - if the page is not
313 * uptodate then the caller will launch readpage again, and
314 * will then handle the error.
315 */
316 if (ret)
317 read_pages(mapping, filp, &page_pool, ret);
318 BUG_ON(!list_empty(&page_pool));
319out:
320 return ret;
321}
322
323/*
324 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
325 * memory at once.
326 */
327int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800328 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329{
330 int ret = 0;
331
332 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
333 return -EINVAL;
334
335 while (nr_to_read) {
336 int err;
337
338 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
339
340 if (this_chunk > nr_to_read)
341 this_chunk = nr_to_read;
342 err = __do_page_cache_readahead(mapping, filp,
343 offset, this_chunk);
344 if (err < 0) {
345 ret = err;
346 break;
347 }
348 ret += err;
349 offset += this_chunk;
350 nr_to_read -= this_chunk;
351 }
352 return ret;
353}
354
355/*
356 * Check how effective readahead is being. If the amount of started IO is
357 * less than expected then the file is partly or fully in pagecache and
358 * readahead isn't helping.
359 *
360 */
361static inline int check_ra_success(struct file_ra_state *ra,
362 unsigned long nr_to_read, unsigned long actual)
363{
364 if (actual == 0) {
365 ra->cache_hit += nr_to_read;
366 if (ra->cache_hit >= VM_MAX_CACHE_HIT) {
367 ra_off(ra);
368 ra->flags |= RA_FLAG_INCACHE;
369 return 0;
370 }
371 } else {
372 ra->cache_hit=0;
373 }
374 return 1;
375}
376
377/*
378 * This version skips the IO if the queue is read-congested, and will tell the
379 * block layer to abandon the readahead if request allocation would block.
380 *
381 * force_page_cache_readahead() will ignore queue congestion and will block on
382 * request queues.
383 */
384int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800385 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386{
387 if (bdi_read_congested(mapping->backing_dev_info))
388 return -1;
389
390 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
391}
392
393/*
394 * Read 'nr_to_read' pages starting at page 'offset'. If the flag 'block'
395 * is set wait till the read completes. Otherwise attempt to read without
396 * blocking.
397 * Returns 1 meaning 'success' if read is succesfull without switching off
398 * readhaead mode. Otherwise return failure.
399 */
400static int
401blockable_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800402 pgoff_t offset, unsigned long nr_to_read,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 struct file_ra_state *ra, int block)
404{
405 int actual;
406
407 if (!block && bdi_read_congested(mapping->backing_dev_info))
408 return 0;
409
410 actual = __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
411
412 return check_ra_success(ra, nr_to_read, actual);
413}
414
415static int make_ahead_window(struct address_space *mapping, struct file *filp,
416 struct file_ra_state *ra, int force)
417{
418 int block, ret;
419
420 ra->ahead_size = get_next_ra_size(ra);
421 ra->ahead_start = ra->start + ra->size;
422
423 block = force || (ra->prev_page >= ra->ahead_start);
424 ret = blockable_page_cache_readahead(mapping, filp,
425 ra->ahead_start, ra->ahead_size, ra, block);
426
427 if (!ret && !force) {
428 /* A read failure in blocking mode, implies pages are
429 * all cached. So we can safely assume we have taken
430 * care of all the pages requested in this call.
431 * A read failure in non-blocking mode, implies we are
432 * reading more pages than requested in this call. So
433 * we safely assume we have taken care of all the pages
434 * requested in this call.
435 *
436 * Just reset the ahead window in case we failed due to
437 * congestion. The ahead window will any way be closed
438 * in case we failed due to excessive page cache hits.
439 */
Oleg Nesterova564da32006-03-22 00:08:47 -0800440 reset_ahead_window(ra);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 }
442
443 return ret;
444}
445
Andrew Morton7361f4d2005-11-07 00:59:28 -0800446/**
447 * page_cache_readahead - generic adaptive readahead
448 * @mapping: address_space which holds the pagecache and I/O vectors
449 * @ra: file_ra_state which holds the readahead state
450 * @filp: passed on to ->readpage() and ->readpages()
451 * @offset: start offset into @mapping, in PAGE_CACHE_SIZE units
452 * @req_size: hint: total size of the read which the caller is performing in
453 * PAGE_CACHE_SIZE units
454 *
455 * page_cache_readahead() is the main function. If performs the adaptive
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 * readahead window size management and submits the readahead I/O.
Andrew Morton7361f4d2005-11-07 00:59:28 -0800457 *
458 * Note that @filp is purely used for passing on to the ->readpage[s]()
459 * handler: it may refer to a different file from @mapping (so we may not use
460 * @filp->f_mapping or @filp->f_dentry->d_inode here).
461 * Also, @ra may not be equal to &@filp->f_ra.
462 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 */
464unsigned long
465page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800466 struct file *filp, pgoff_t offset, unsigned long req_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
468 unsigned long max, newsize;
469 int sequential;
470
471 /*
472 * We avoid doing extra work and bogusly perturbing the readahead
473 * window expansion logic.
474 */
475 if (offset == ra->prev_page && --req_size)
476 ++offset;
477
478 /* Note that prev_page == -1 if it is a first read */
479 sequential = (offset == ra->prev_page + 1);
480 ra->prev_page = offset;
481
482 max = get_max_readahead(ra);
483 newsize = min(req_size, max);
484
485 /* No readahead or sub-page sized read or file already in cache */
486 if (newsize == 0 || (ra->flags & RA_FLAG_INCACHE))
487 goto out;
488
489 ra->prev_page += newsize - 1;
490
491 /*
492 * Special case - first read at start of file. We'll assume it's
493 * a whole-file read and grow the window fast. Or detect first
494 * sequential access
495 */
496 if (sequential && ra->size == 0) {
497 ra->size = get_init_ra_size(newsize, max);
498 ra->start = offset;
499 if (!blockable_page_cache_readahead(mapping, filp, offset,
500 ra->size, ra, 1))
501 goto out;
502
503 /*
504 * If the request size is larger than our max readahead, we
505 * at least want to be sure that we get 2 IOs in flight and
506 * we know that we will definitly need the new I/O.
507 * once we do this, subsequent calls should be able to overlap
508 * IOs,* thus preventing stalls. so issue the ahead window
509 * immediately.
510 */
511 if (req_size >= max)
512 make_ahead_window(mapping, filp, ra, 1);
513
514 goto out;
515 }
516
517 /*
518 * Now handle the random case:
519 * partial page reads and first access were handled above,
520 * so this must be the next page otherwise it is random
521 */
522 if (!sequential) {
523 ra_off(ra);
524 blockable_page_cache_readahead(mapping, filp, offset,
525 newsize, ra, 1);
526 goto out;
527 }
528
529 /*
530 * If we get here we are doing sequential IO and this was not the first
531 * occurence (ie we have an existing window)
532 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 if (ra->ahead_start == 0) { /* no ahead window yet */
534 if (!make_ahead_window(mapping, filp, ra, 0))
Oleg Nesterova564da32006-03-22 00:08:47 -0800535 goto recheck;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 }
Oleg Nesterova564da32006-03-22 00:08:47 -0800537
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 /*
539 * Already have an ahead window, check if we crossed into it.
540 * If so, shift windows and issue a new ahead window.
541 * Only return the #pages that are in the current window, so that
542 * we get called back on the first page of the ahead window which
543 * will allow us to submit more IO.
544 */
545 if (ra->prev_page >= ra->ahead_start) {
546 ra->start = ra->ahead_start;
547 ra->size = ra->ahead_size;
548 make_ahead_window(mapping, filp, ra, 0);
Oleg Nesterova564da32006-03-22 00:08:47 -0800549recheck:
550 /* prev_page shouldn't overrun the ahead window */
551 ra->prev_page = min(ra->prev_page,
552 ra->ahead_start + ra->ahead_size - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 }
554
555out:
556 return ra->prev_page + 1;
557}
Andrew Mortond8733c22006-03-23 03:00:11 -0800558EXPORT_SYMBOL_GPL(page_cache_readahead);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559
560/*
561 * handle_ra_miss() is called when it is known that a page which should have
562 * been present in the pagecache (we just did some readahead there) was in fact
563 * not found. This will happen if it was evicted by the VM (readahead
564 * thrashing)
565 *
566 * Turn on the cache miss flag in the RA struct, this will cause the RA code
567 * to reduce the RA size on the next read.
568 */
569void handle_ra_miss(struct address_space *mapping,
570 struct file_ra_state *ra, pgoff_t offset)
571{
572 ra->flags |= RA_FLAG_MISS;
573 ra->flags &= ~RA_FLAG_INCACHE;
Steven Pratt3b30bbd2005-09-06 15:17:06 -0700574 ra->cache_hit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575}
576
577/*
578 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
579 * sensible upper limit.
580 */
581unsigned long max_sane_readahead(unsigned long nr)
582{
583 unsigned long active;
584 unsigned long inactive;
585 unsigned long free;
586
587 __get_zone_counts(&active, &inactive, &free, NODE_DATA(numa_node_id()));
588 return min(nr, (inactive + free) / 2);
589}