blob: 4ee52cadab9371f4d22522489b5d4cec1be2125b [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 Brown9f1a3cf2006-06-25 05:46:46 -0700185 mapping->a_ops->readpage(filp, page);
186 if (!pagevec_add(&lru_pvec, page))
187 __pagevec_lru_add(&lru_pvec);
188 } else
189 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 }
191 pagevec_lru_add(&lru_pvec);
Zach Brown994fc28c2005-12-15 14:28:17 -0800192 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193out:
194 return ret;
195}
196
197/*
198 * Readahead design.
199 *
200 * The fields in struct file_ra_state represent the most-recently-executed
201 * readahead attempt:
202 *
203 * start: Page index at which we started the readahead
204 * size: Number of pages in that read
205 * Together, these form the "current window".
206 * Together, start and size represent the `readahead window'.
207 * prev_page: The page which the readahead algorithm most-recently inspected.
208 * It is mainly used to detect sequential file reading.
209 * If page_cache_readahead sees that it is again being called for
210 * a page which it just looked at, it can return immediately without
211 * making any state changes.
212 * ahead_start,
213 * ahead_size: Together, these form the "ahead window".
214 * ra_pages: The externally controlled max readahead for this fd.
215 *
216 * When readahead is in the off state (size == 0), readahead is disabled.
217 * In this state, prev_page is used to detect the resumption of sequential I/O.
218 *
219 * The readahead code manages two windows - the "current" and the "ahead"
220 * windows. The intent is that while the application is walking the pages
221 * in the current window, I/O is underway on the ahead window. When the
222 * current window is fully traversed, it is replaced by the ahead window
223 * and the ahead window is invalidated. When this copying happens, the
224 * new current window's pages are probably still locked. So
225 * we submit a new batch of I/O immediately, creating a new ahead window.
226 *
227 * So:
228 *
229 * ----|----------------|----------------|-----
230 * ^start ^start+size
231 * ^ahead_start ^ahead_start+ahead_size
232 *
233 * ^ When this page is read, we submit I/O for the
234 * ahead window.
235 *
236 * A `readahead hit' occurs when a read request is made against a page which is
237 * the next sequential page. Ahead window calculations are done only when it
238 * is time to submit a new IO. The code ramps up the size agressively at first,
239 * but slow down as it approaches max_readhead.
240 *
241 * Any seek/ramdom IO will result in readahead being turned off. It will resume
242 * at the first sequential access.
243 *
244 * There is a special-case: if the first page which the application tries to
245 * read happens to be the first page of the file, it is assumed that a linear
246 * read is about to happen and the window is immediately set to the initial size
247 * based on I/O request size and the max_readahead.
248 *
249 * This function is to be called for every read request, rather than when
250 * it is time to perform readahead. It is called only once for the entire I/O
251 * regardless of size unless readahead is unable to start enough I/O to satisfy
252 * the request (I/O request > max_readahead).
253 */
254
255/*
256 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
257 * the pages first, then submits them all for I/O. This avoids the very bad
258 * behaviour which would occur if page allocations are causing VM writeback.
259 * We really don't want to intermingle reads and writes like that.
260 *
261 * Returns the number of pages requested, or the maximum amount of I/O allowed.
262 *
263 * do_page_cache_readahead() returns -1 if it encountered request queue
264 * congestion.
265 */
266static int
267__do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800268 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269{
270 struct inode *inode = mapping->host;
271 struct page *page;
272 unsigned long end_index; /* The last page we want to read */
273 LIST_HEAD(page_pool);
274 int page_idx;
275 int ret = 0;
276 loff_t isize = i_size_read(inode);
277
278 if (isize == 0)
279 goto out;
280
281 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
282
283 /*
284 * Preallocate as many pages as we will need.
285 */
286 read_lock_irq(&mapping->tree_lock);
287 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
Andrew Morton7361f4d2005-11-07 00:59:28 -0800288 pgoff_t page_offset = offset + page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
290 if (page_offset > end_index)
291 break;
292
293 page = radix_tree_lookup(&mapping->page_tree, page_offset);
294 if (page)
295 continue;
296
297 read_unlock_irq(&mapping->tree_lock);
298 page = page_cache_alloc_cold(mapping);
299 read_lock_irq(&mapping->tree_lock);
300 if (!page)
301 break;
302 page->index = page_offset;
303 list_add(&page->lru, &page_pool);
304 ret++;
305 }
306 read_unlock_irq(&mapping->tree_lock);
307
308 /*
309 * Now start the IO. We ignore I/O errors - if the page is not
310 * uptodate then the caller will launch readpage again, and
311 * will then handle the error.
312 */
313 if (ret)
314 read_pages(mapping, filp, &page_pool, ret);
315 BUG_ON(!list_empty(&page_pool));
316out:
317 return ret;
318}
319
320/*
321 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
322 * memory at once.
323 */
324int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800325 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326{
327 int ret = 0;
328
329 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
330 return -EINVAL;
331
332 while (nr_to_read) {
333 int err;
334
335 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
336
337 if (this_chunk > nr_to_read)
338 this_chunk = nr_to_read;
339 err = __do_page_cache_readahead(mapping, filp,
340 offset, this_chunk);
341 if (err < 0) {
342 ret = err;
343 break;
344 }
345 ret += err;
346 offset += this_chunk;
347 nr_to_read -= this_chunk;
348 }
349 return ret;
350}
351
352/*
353 * Check how effective readahead is being. If the amount of started IO is
354 * less than expected then the file is partly or fully in pagecache and
355 * readahead isn't helping.
356 *
357 */
358static inline int check_ra_success(struct file_ra_state *ra,
359 unsigned long nr_to_read, unsigned long actual)
360{
361 if (actual == 0) {
362 ra->cache_hit += nr_to_read;
363 if (ra->cache_hit >= VM_MAX_CACHE_HIT) {
364 ra_off(ra);
365 ra->flags |= RA_FLAG_INCACHE;
366 return 0;
367 }
368 } else {
369 ra->cache_hit=0;
370 }
371 return 1;
372}
373
374/*
375 * This version skips the IO if the queue is read-congested, and will tell the
376 * block layer to abandon the readahead if request allocation would block.
377 *
378 * force_page_cache_readahead() will ignore queue congestion and will block on
379 * request queues.
380 */
381int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800382 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383{
384 if (bdi_read_congested(mapping->backing_dev_info))
385 return -1;
386
387 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
388}
389
390/*
391 * Read 'nr_to_read' pages starting at page 'offset'. If the flag 'block'
392 * is set wait till the read completes. Otherwise attempt to read without
393 * blocking.
394 * Returns 1 meaning 'success' if read is succesfull without switching off
395 * readhaead mode. Otherwise return failure.
396 */
397static int
398blockable_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800399 pgoff_t offset, unsigned long nr_to_read,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 struct file_ra_state *ra, int block)
401{
402 int actual;
403
404 if (!block && bdi_read_congested(mapping->backing_dev_info))
405 return 0;
406
407 actual = __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
408
409 return check_ra_success(ra, nr_to_read, actual);
410}
411
412static int make_ahead_window(struct address_space *mapping, struct file *filp,
413 struct file_ra_state *ra, int force)
414{
415 int block, ret;
416
417 ra->ahead_size = get_next_ra_size(ra);
418 ra->ahead_start = ra->start + ra->size;
419
420 block = force || (ra->prev_page >= ra->ahead_start);
421 ret = blockable_page_cache_readahead(mapping, filp,
422 ra->ahead_start, ra->ahead_size, ra, block);
423
424 if (!ret && !force) {
425 /* A read failure in blocking mode, implies pages are
426 * all cached. So we can safely assume we have taken
427 * care of all the pages requested in this call.
428 * A read failure in non-blocking mode, implies we are
429 * reading more pages than requested in this call. So
430 * we safely assume we have taken care of all the pages
431 * requested in this call.
432 *
433 * Just reset the ahead window in case we failed due to
434 * congestion. The ahead window will any way be closed
435 * in case we failed due to excessive page cache hits.
436 */
Oleg Nesterova564da32006-03-22 00:08:47 -0800437 reset_ahead_window(ra);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 }
439
440 return ret;
441}
442
Andrew Morton7361f4d2005-11-07 00:59:28 -0800443/**
444 * page_cache_readahead - generic adaptive readahead
445 * @mapping: address_space which holds the pagecache and I/O vectors
446 * @ra: file_ra_state which holds the readahead state
447 * @filp: passed on to ->readpage() and ->readpages()
448 * @offset: start offset into @mapping, in PAGE_CACHE_SIZE units
449 * @req_size: hint: total size of the read which the caller is performing in
450 * PAGE_CACHE_SIZE units
451 *
452 * page_cache_readahead() is the main function. If performs the adaptive
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 * readahead window size management and submits the readahead I/O.
Andrew Morton7361f4d2005-11-07 00:59:28 -0800454 *
455 * Note that @filp is purely used for passing on to the ->readpage[s]()
456 * handler: it may refer to a different file from @mapping (so we may not use
457 * @filp->f_mapping or @filp->f_dentry->d_inode here).
458 * Also, @ra may not be equal to &@filp->f_ra.
459 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 */
461unsigned long
462page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800463 struct file *filp, pgoff_t offset, unsigned long req_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464{
465 unsigned long max, newsize;
466 int sequential;
467
468 /*
469 * We avoid doing extra work and bogusly perturbing the readahead
470 * window expansion logic.
471 */
472 if (offset == ra->prev_page && --req_size)
473 ++offset;
474
475 /* Note that prev_page == -1 if it is a first read */
476 sequential = (offset == ra->prev_page + 1);
477 ra->prev_page = offset;
478
479 max = get_max_readahead(ra);
480 newsize = min(req_size, max);
481
482 /* No readahead or sub-page sized read or file already in cache */
483 if (newsize == 0 || (ra->flags & RA_FLAG_INCACHE))
484 goto out;
485
486 ra->prev_page += newsize - 1;
487
488 /*
489 * Special case - first read at start of file. We'll assume it's
490 * a whole-file read and grow the window fast. Or detect first
491 * sequential access
492 */
493 if (sequential && ra->size == 0) {
494 ra->size = get_init_ra_size(newsize, max);
495 ra->start = offset;
496 if (!blockable_page_cache_readahead(mapping, filp, offset,
497 ra->size, ra, 1))
498 goto out;
499
500 /*
501 * If the request size is larger than our max readahead, we
502 * at least want to be sure that we get 2 IOs in flight and
503 * we know that we will definitly need the new I/O.
504 * once we do this, subsequent calls should be able to overlap
505 * IOs,* thus preventing stalls. so issue the ahead window
506 * immediately.
507 */
508 if (req_size >= max)
509 make_ahead_window(mapping, filp, ra, 1);
510
511 goto out;
512 }
513
514 /*
515 * Now handle the random case:
516 * partial page reads and first access were handled above,
517 * so this must be the next page otherwise it is random
518 */
519 if (!sequential) {
520 ra_off(ra);
521 blockable_page_cache_readahead(mapping, filp, offset,
522 newsize, ra, 1);
523 goto out;
524 }
525
526 /*
527 * If we get here we are doing sequential IO and this was not the first
528 * occurence (ie we have an existing window)
529 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 if (ra->ahead_start == 0) { /* no ahead window yet */
531 if (!make_ahead_window(mapping, filp, ra, 0))
Oleg Nesterova564da32006-03-22 00:08:47 -0800532 goto recheck;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 }
Oleg Nesterova564da32006-03-22 00:08:47 -0800534
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 /*
536 * Already have an ahead window, check if we crossed into it.
537 * If so, shift windows and issue a new ahead window.
538 * Only return the #pages that are in the current window, so that
539 * we get called back on the first page of the ahead window which
540 * will allow us to submit more IO.
541 */
542 if (ra->prev_page >= ra->ahead_start) {
543 ra->start = ra->ahead_start;
544 ra->size = ra->ahead_size;
545 make_ahead_window(mapping, filp, ra, 0);
Oleg Nesterova564da32006-03-22 00:08:47 -0800546recheck:
547 /* prev_page shouldn't overrun the ahead window */
548 ra->prev_page = min(ra->prev_page,
549 ra->ahead_start + ra->ahead_size - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 }
551
552out:
553 return ra->prev_page + 1;
554}
Andrew Mortond8733c22006-03-23 03:00:11 -0800555EXPORT_SYMBOL_GPL(page_cache_readahead);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
557/*
558 * handle_ra_miss() is called when it is known that a page which should have
559 * been present in the pagecache (we just did some readahead there) was in fact
560 * not found. This will happen if it was evicted by the VM (readahead
561 * thrashing)
562 *
563 * Turn on the cache miss flag in the RA struct, this will cause the RA code
564 * to reduce the RA size on the next read.
565 */
566void handle_ra_miss(struct address_space *mapping,
567 struct file_ra_state *ra, pgoff_t offset)
568{
569 ra->flags |= RA_FLAG_MISS;
570 ra->flags &= ~RA_FLAG_INCACHE;
Steven Pratt3b30bbd2005-09-06 15:17:06 -0700571 ra->cache_hit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572}
573
574/*
575 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
576 * sensible upper limit.
577 */
578unsigned long max_sane_readahead(unsigned long nr)
579{
580 unsigned long active;
581 unsigned long inactive;
582 unsigned long free;
583
584 __get_zone_counts(&active, &inactive, &free, NODE_DATA(numa_node_id()));
585 return min(nr, (inactive + free) / 2);
586}