blob: 8d6eeaaa6296f9fb2372cff059138cbdd4b89723 [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
55static inline void ra_off(struct file_ra_state *ra)
56{
57 ra->start = 0;
58 ra->flags = 0;
59 ra->size = 0;
60 ra->ahead_start = 0;
61 ra->ahead_size = 0;
62 return;
63}
64
65/*
66 * Set the initial window size, round to next power of 2 and square
67 * for small size, x 4 for medium, and x 2 for large
68 * for 128k (32 page) max ra
69 * 1-8 page = 32k initial, > 8 page = 128k initial
70 */
71static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
72{
73 unsigned long newsize = roundup_pow_of_two(size);
74
75 if (newsize <= max / 64)
76 newsize = newsize * newsize;
77 else if (newsize <= max / 4)
78 newsize = max / 4;
79 else
80 newsize = max;
81 return newsize;
82}
83
84/*
85 * Set the new window size, this is called only when I/O is to be submitted,
86 * not for each call to readahead. If a cache miss occured, reduce next I/O
87 * size, else increase depending on how close to max we are.
88 */
89static inline unsigned long get_next_ra_size(struct file_ra_state *ra)
90{
91 unsigned long max = get_max_readahead(ra);
92 unsigned long min = get_min_readahead(ra);
93 unsigned long cur = ra->size;
94 unsigned long newsize;
95
96 if (ra->flags & RA_FLAG_MISS) {
97 ra->flags &= ~RA_FLAG_MISS;
98 newsize = max((cur - 2), min);
99 } else if (cur < max / 16) {
100 newsize = 4 * cur;
101 } else {
102 newsize = 2 * cur;
103 }
104 return min(newsize, max);
105}
106
107#define list_to_page(head) (list_entry((head)->prev, struct page, lru))
108
109/**
110 * read_cache_pages - populate an address space with some pages, and
111 * start reads against them.
112 * @mapping: the address_space
113 * @pages: The address of a list_head which contains the target pages. These
114 * pages have their ->index populated and are otherwise uninitialised.
115 * @filler: callback routine for filling a single page.
116 * @data: private data for the callback routine.
117 *
118 * Hides the details of the LRU cache etc from the filesystems.
119 */
120int read_cache_pages(struct address_space *mapping, struct list_head *pages,
121 int (*filler)(void *, struct page *), void *data)
122{
123 struct page *page;
124 struct pagevec lru_pvec;
125 int ret = 0;
126
127 pagevec_init(&lru_pvec, 0);
128
129 while (!list_empty(pages)) {
130 page = list_to_page(pages);
131 list_del(&page->lru);
132 if (add_to_page_cache(page, mapping, page->index, GFP_KERNEL)) {
133 page_cache_release(page);
134 continue;
135 }
136 ret = filler(data, page);
137 if (!pagevec_add(&lru_pvec, page))
138 __pagevec_lru_add(&lru_pvec);
139 if (ret) {
140 while (!list_empty(pages)) {
141 struct page *victim;
142
143 victim = list_to_page(pages);
144 list_del(&victim->lru);
145 page_cache_release(victim);
146 }
147 break;
148 }
149 }
150 pagevec_lru_add(&lru_pvec);
151 return ret;
152}
153
154EXPORT_SYMBOL(read_cache_pages);
155
156static int read_pages(struct address_space *mapping, struct file *filp,
157 struct list_head *pages, unsigned nr_pages)
158{
159 unsigned page_idx;
160 struct pagevec lru_pvec;
Zach Brown994fc28c2005-12-15 14:28:17 -0800161 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
163 if (mapping->a_ops->readpages) {
164 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
165 goto out;
166 }
167
168 pagevec_init(&lru_pvec, 0);
169 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
170 struct page *page = list_to_page(pages);
171 list_del(&page->lru);
172 if (!add_to_page_cache(page, mapping,
173 page->index, GFP_KERNEL)) {
Zach Brown994fc28c2005-12-15 14:28:17 -0800174 ret = mapping->a_ops->readpage(filp, page);
175 if (ret != AOP_TRUNCATED_PAGE) {
176 if (!pagevec_add(&lru_pvec, page))
177 __pagevec_lru_add(&lru_pvec);
178 continue;
179 } /* else fall through to release */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 }
Zach Brown994fc28c2005-12-15 14:28:17 -0800181 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 }
183 pagevec_lru_add(&lru_pvec);
Zach Brown994fc28c2005-12-15 14:28:17 -0800184 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185out:
186 return ret;
187}
188
189/*
190 * Readahead design.
191 *
192 * The fields in struct file_ra_state represent the most-recently-executed
193 * readahead attempt:
194 *
195 * start: Page index at which we started the readahead
196 * size: Number of pages in that read
197 * Together, these form the "current window".
198 * Together, start and size represent the `readahead window'.
199 * prev_page: The page which the readahead algorithm most-recently inspected.
200 * It is mainly used to detect sequential file reading.
201 * If page_cache_readahead sees that it is again being called for
202 * a page which it just looked at, it can return immediately without
203 * making any state changes.
204 * ahead_start,
205 * ahead_size: Together, these form the "ahead window".
206 * ra_pages: The externally controlled max readahead for this fd.
207 *
208 * When readahead is in the off state (size == 0), readahead is disabled.
209 * In this state, prev_page is used to detect the resumption of sequential I/O.
210 *
211 * The readahead code manages two windows - the "current" and the "ahead"
212 * windows. The intent is that while the application is walking the pages
213 * in the current window, I/O is underway on the ahead window. When the
214 * current window is fully traversed, it is replaced by the ahead window
215 * and the ahead window is invalidated. When this copying happens, the
216 * new current window's pages are probably still locked. So
217 * we submit a new batch of I/O immediately, creating a new ahead window.
218 *
219 * So:
220 *
221 * ----|----------------|----------------|-----
222 * ^start ^start+size
223 * ^ahead_start ^ahead_start+ahead_size
224 *
225 * ^ When this page is read, we submit I/O for the
226 * ahead window.
227 *
228 * A `readahead hit' occurs when a read request is made against a page which is
229 * the next sequential page. Ahead window calculations are done only when it
230 * is time to submit a new IO. The code ramps up the size agressively at first,
231 * but slow down as it approaches max_readhead.
232 *
233 * Any seek/ramdom IO will result in readahead being turned off. It will resume
234 * at the first sequential access.
235 *
236 * There is a special-case: if the first page which the application tries to
237 * read happens to be the first page of the file, it is assumed that a linear
238 * read is about to happen and the window is immediately set to the initial size
239 * based on I/O request size and the max_readahead.
240 *
241 * This function is to be called for every read request, rather than when
242 * it is time to perform readahead. It is called only once for the entire I/O
243 * regardless of size unless readahead is unable to start enough I/O to satisfy
244 * the request (I/O request > max_readahead).
245 */
246
247/*
248 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
249 * the pages first, then submits them all for I/O. This avoids the very bad
250 * behaviour which would occur if page allocations are causing VM writeback.
251 * We really don't want to intermingle reads and writes like that.
252 *
253 * Returns the number of pages requested, or the maximum amount of I/O allowed.
254 *
255 * do_page_cache_readahead() returns -1 if it encountered request queue
256 * congestion.
257 */
258static int
259__do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800260 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261{
262 struct inode *inode = mapping->host;
263 struct page *page;
264 unsigned long end_index; /* The last page we want to read */
265 LIST_HEAD(page_pool);
266 int page_idx;
267 int ret = 0;
268 loff_t isize = i_size_read(inode);
269
270 if (isize == 0)
271 goto out;
272
273 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
274
275 /*
276 * Preallocate as many pages as we will need.
277 */
278 read_lock_irq(&mapping->tree_lock);
279 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
Andrew Morton7361f4d2005-11-07 00:59:28 -0800280 pgoff_t page_offset = offset + page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
282 if (page_offset > end_index)
283 break;
284
285 page = radix_tree_lookup(&mapping->page_tree, page_offset);
286 if (page)
287 continue;
288
289 read_unlock_irq(&mapping->tree_lock);
290 page = page_cache_alloc_cold(mapping);
291 read_lock_irq(&mapping->tree_lock);
292 if (!page)
293 break;
294 page->index = page_offset;
295 list_add(&page->lru, &page_pool);
296 ret++;
297 }
298 read_unlock_irq(&mapping->tree_lock);
299
300 /*
301 * Now start the IO. We ignore I/O errors - if the page is not
302 * uptodate then the caller will launch readpage again, and
303 * will then handle the error.
304 */
305 if (ret)
306 read_pages(mapping, filp, &page_pool, ret);
307 BUG_ON(!list_empty(&page_pool));
308out:
309 return ret;
310}
311
312/*
313 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
314 * memory at once.
315 */
316int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800317 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
319 int ret = 0;
320
321 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
322 return -EINVAL;
323
324 while (nr_to_read) {
325 int err;
326
327 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
328
329 if (this_chunk > nr_to_read)
330 this_chunk = nr_to_read;
331 err = __do_page_cache_readahead(mapping, filp,
332 offset, this_chunk);
333 if (err < 0) {
334 ret = err;
335 break;
336 }
337 ret += err;
338 offset += this_chunk;
339 nr_to_read -= this_chunk;
340 }
341 return ret;
342}
343
344/*
345 * Check how effective readahead is being. If the amount of started IO is
346 * less than expected then the file is partly or fully in pagecache and
347 * readahead isn't helping.
348 *
349 */
350static inline int check_ra_success(struct file_ra_state *ra,
351 unsigned long nr_to_read, unsigned long actual)
352{
353 if (actual == 0) {
354 ra->cache_hit += nr_to_read;
355 if (ra->cache_hit >= VM_MAX_CACHE_HIT) {
356 ra_off(ra);
357 ra->flags |= RA_FLAG_INCACHE;
358 return 0;
359 }
360 } else {
361 ra->cache_hit=0;
362 }
363 return 1;
364}
365
366/*
367 * This version skips the IO if the queue is read-congested, and will tell the
368 * block layer to abandon the readahead if request allocation would block.
369 *
370 * force_page_cache_readahead() will ignore queue congestion and will block on
371 * request queues.
372 */
373int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800374 pgoff_t offset, unsigned long nr_to_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375{
376 if (bdi_read_congested(mapping->backing_dev_info))
377 return -1;
378
379 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
380}
381
382/*
383 * Read 'nr_to_read' pages starting at page 'offset'. If the flag 'block'
384 * is set wait till the read completes. Otherwise attempt to read without
385 * blocking.
386 * Returns 1 meaning 'success' if read is succesfull without switching off
387 * readhaead mode. Otherwise return failure.
388 */
389static int
390blockable_page_cache_readahead(struct address_space *mapping, struct file *filp,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800391 pgoff_t offset, unsigned long nr_to_read,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 struct file_ra_state *ra, int block)
393{
394 int actual;
395
396 if (!block && bdi_read_congested(mapping->backing_dev_info))
397 return 0;
398
399 actual = __do_page_cache_readahead(mapping, filp, offset, nr_to_read);
400
401 return check_ra_success(ra, nr_to_read, actual);
402}
403
404static int make_ahead_window(struct address_space *mapping, struct file *filp,
405 struct file_ra_state *ra, int force)
406{
407 int block, ret;
408
409 ra->ahead_size = get_next_ra_size(ra);
410 ra->ahead_start = ra->start + ra->size;
411
412 block = force || (ra->prev_page >= ra->ahead_start);
413 ret = blockable_page_cache_readahead(mapping, filp,
414 ra->ahead_start, ra->ahead_size, ra, block);
415
416 if (!ret && !force) {
417 /* A read failure in blocking mode, implies pages are
418 * all cached. So we can safely assume we have taken
419 * care of all the pages requested in this call.
420 * A read failure in non-blocking mode, implies we are
421 * reading more pages than requested in this call. So
422 * we safely assume we have taken care of all the pages
423 * requested in this call.
424 *
425 * Just reset the ahead window in case we failed due to
426 * congestion. The ahead window will any way be closed
427 * in case we failed due to excessive page cache hits.
428 */
429 ra->ahead_start = 0;
430 ra->ahead_size = 0;
431 }
432
433 return ret;
434}
435
Andrew Morton7361f4d2005-11-07 00:59:28 -0800436/**
437 * page_cache_readahead - generic adaptive readahead
438 * @mapping: address_space which holds the pagecache and I/O vectors
439 * @ra: file_ra_state which holds the readahead state
440 * @filp: passed on to ->readpage() and ->readpages()
441 * @offset: start offset into @mapping, in PAGE_CACHE_SIZE units
442 * @req_size: hint: total size of the read which the caller is performing in
443 * PAGE_CACHE_SIZE units
444 *
445 * page_cache_readahead() is the main function. If performs the adaptive
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 * readahead window size management and submits the readahead I/O.
Andrew Morton7361f4d2005-11-07 00:59:28 -0800447 *
448 * Note that @filp is purely used for passing on to the ->readpage[s]()
449 * handler: it may refer to a different file from @mapping (so we may not use
450 * @filp->f_mapping or @filp->f_dentry->d_inode here).
451 * Also, @ra may not be equal to &@filp->f_ra.
452 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 */
454unsigned long
455page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
Andrew Morton7361f4d2005-11-07 00:59:28 -0800456 struct file *filp, pgoff_t offset, unsigned long req_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457{
458 unsigned long max, newsize;
459 int sequential;
460
461 /*
462 * We avoid doing extra work and bogusly perturbing the readahead
463 * window expansion logic.
464 */
465 if (offset == ra->prev_page && --req_size)
466 ++offset;
467
468 /* Note that prev_page == -1 if it is a first read */
469 sequential = (offset == ra->prev_page + 1);
470 ra->prev_page = offset;
471
472 max = get_max_readahead(ra);
473 newsize = min(req_size, max);
474
475 /* No readahead or sub-page sized read or file already in cache */
476 if (newsize == 0 || (ra->flags & RA_FLAG_INCACHE))
477 goto out;
478
479 ra->prev_page += newsize - 1;
480
481 /*
482 * Special case - first read at start of file. We'll assume it's
483 * a whole-file read and grow the window fast. Or detect first
484 * sequential access
485 */
486 if (sequential && ra->size == 0) {
487 ra->size = get_init_ra_size(newsize, max);
488 ra->start = offset;
489 if (!blockable_page_cache_readahead(mapping, filp, offset,
490 ra->size, ra, 1))
491 goto out;
492
493 /*
494 * If the request size is larger than our max readahead, we
495 * at least want to be sure that we get 2 IOs in flight and
496 * we know that we will definitly need the new I/O.
497 * once we do this, subsequent calls should be able to overlap
498 * IOs,* thus preventing stalls. so issue the ahead window
499 * immediately.
500 */
501 if (req_size >= max)
502 make_ahead_window(mapping, filp, ra, 1);
503
504 goto out;
505 }
506
507 /*
508 * Now handle the random case:
509 * partial page reads and first access were handled above,
510 * so this must be the next page otherwise it is random
511 */
512 if (!sequential) {
513 ra_off(ra);
514 blockable_page_cache_readahead(mapping, filp, offset,
515 newsize, ra, 1);
516 goto out;
517 }
518
519 /*
520 * If we get here we are doing sequential IO and this was not the first
521 * occurence (ie we have an existing window)
522 */
523
524 if (ra->ahead_start == 0) { /* no ahead window yet */
525 if (!make_ahead_window(mapping, filp, ra, 0))
526 goto out;
527 }
528 /*
529 * Already have an ahead window, check if we crossed into it.
530 * If so, shift windows and issue a new ahead window.
531 * Only return the #pages that are in the current window, so that
532 * we get called back on the first page of the ahead window which
533 * will allow us to submit more IO.
534 */
535 if (ra->prev_page >= ra->ahead_start) {
536 ra->start = ra->ahead_start;
537 ra->size = ra->ahead_size;
538 make_ahead_window(mapping, filp, ra, 0);
539 }
540
541out:
542 return ra->prev_page + 1;
543}
544
545/*
546 * handle_ra_miss() is called when it is known that a page which should have
547 * been present in the pagecache (we just did some readahead there) was in fact
548 * not found. This will happen if it was evicted by the VM (readahead
549 * thrashing)
550 *
551 * Turn on the cache miss flag in the RA struct, this will cause the RA code
552 * to reduce the RA size on the next read.
553 */
554void handle_ra_miss(struct address_space *mapping,
555 struct file_ra_state *ra, pgoff_t offset)
556{
557 ra->flags |= RA_FLAG_MISS;
558 ra->flags &= ~RA_FLAG_INCACHE;
Steven Pratt3b30bbd2005-09-06 15:17:06 -0700559 ra->cache_hit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
562/*
563 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
564 * sensible upper limit.
565 */
566unsigned long max_sane_readahead(unsigned long nr)
567{
568 unsigned long active;
569 unsigned long inactive;
570 unsigned long free;
571
572 __get_zone_counts(&active, &inactive, &free, NODE_DATA(numa_node_id()));
573 return min(nr, (inactive + free) / 2);
574}