]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * mm/readahead.c - address_space-level file readahead. | |
3 | * | |
4 | * Copyright (C) 2002, Linus Torvalds | |
5 | * | |
6 | * 09Apr2002 Andrew Morton | |
7 | * Initial version. | |
8 | */ | |
9 | ||
10 | #include <linux/kernel.h> | |
11 | #include <linux/gfp.h> | |
12 | #include <linux/export.h> | |
13 | #include <linux/blkdev.h> | |
14 | #include <linux/backing-dev.h> | |
15 | #include <linux/task_io_accounting_ops.h> | |
16 | #include <linux/pagevec.h> | |
17 | #include <linux/pagemap.h> | |
18 | #include <linux/syscalls.h> | |
19 | #include <linux/file.h> | |
20 | ||
21 | #include "internal.h" | |
22 | ||
23 | /* | |
24 | * Initialise a struct file's readahead state. Assumes that the caller has | |
25 | * memset *ra to zero. | |
26 | */ | |
27 | void | |
28 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) | |
29 | { | |
30 | ra->ra_pages = mapping->backing_dev_info->ra_pages; | |
31 | ra->prev_pos = -1; | |
32 | } | |
33 | EXPORT_SYMBOL_GPL(file_ra_state_init); | |
34 | ||
35 | #define list_to_page(head) (list_entry((head)->prev, struct page, lru)) | |
36 | ||
37 | /* | |
38 | * see if a page needs releasing upon read_cache_pages() failure | |
39 | * - the caller of read_cache_pages() may have set PG_private or PG_fscache | |
40 | * before calling, such as the NFS fs marking pages that are cached locally | |
41 | * on disk, thus we need to give the fs a chance to clean up in the event of | |
42 | * an error | |
43 | */ | |
44 | static void read_cache_pages_invalidate_page(struct address_space *mapping, | |
45 | struct page *page) | |
46 | { | |
47 | if (page_has_private(page)) { | |
48 | if (!trylock_page(page)) | |
49 | BUG(); | |
50 | page->mapping = mapping; | |
51 | do_invalidatepage(page, 0, PAGE_CACHE_SIZE); | |
52 | page->mapping = NULL; | |
53 | unlock_page(page); | |
54 | } | |
55 | page_cache_release(page); | |
56 | } | |
57 | ||
58 | /* | |
59 | * release a list of pages, invalidating them first if need be | |
60 | */ | |
61 | static void read_cache_pages_invalidate_pages(struct address_space *mapping, | |
62 | struct list_head *pages) | |
63 | { | |
64 | struct page *victim; | |
65 | ||
66 | while (!list_empty(pages)) { | |
67 | victim = list_to_page(pages); | |
68 | list_del(&victim->lru); | |
69 | read_cache_pages_invalidate_page(mapping, victim); | |
70 | } | |
71 | } | |
72 | ||
73 | /** | |
74 | * read_cache_pages - populate an address space with some pages & start reads against them | |
75 | * @mapping: the address_space | |
76 | * @pages: The address of a list_head which contains the target pages. These | |
77 | * pages have their ->index populated and are otherwise uninitialised. | |
78 | * @filler: callback routine for filling a single page. | |
79 | * @data: private data for the callback routine. | |
80 | * | |
81 | * Hides the details of the LRU cache etc from the filesystems. | |
82 | */ | |
83 | int read_cache_pages(struct address_space *mapping, struct list_head *pages, | |
84 | int (*filler)(void *, struct page *), void *data) | |
85 | { | |
86 | struct page *page; | |
87 | int ret = 0; | |
88 | ||
89 | while (!list_empty(pages)) { | |
90 | page = list_to_page(pages); | |
91 | list_del(&page->lru); | |
92 | if (add_to_page_cache_lru(page, mapping, | |
93 | page->index, GFP_KERNEL)) { | |
94 | read_cache_pages_invalidate_page(mapping, page); | |
95 | continue; | |
96 | } | |
97 | page_cache_release(page); | |
98 | ||
99 | ret = filler(data, page); | |
100 | if (unlikely(ret)) { | |
101 | read_cache_pages_invalidate_pages(mapping, pages); | |
102 | break; | |
103 | } | |
104 | task_io_account_read(PAGE_CACHE_SIZE); | |
105 | } | |
106 | return ret; | |
107 | } | |
108 | ||
109 | EXPORT_SYMBOL(read_cache_pages); | |
110 | ||
111 | static int read_pages(struct address_space *mapping, struct file *filp, | |
112 | struct list_head *pages, unsigned nr_pages) | |
113 | { | |
114 | struct blk_plug plug; | |
115 | unsigned page_idx; | |
116 | int ret; | |
117 | ||
118 | blk_start_plug(&plug); | |
119 | ||
120 | if (mapping->a_ops->readpages) { | |
121 | ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages); | |
122 | /* Clean up the remaining pages */ | |
123 | put_pages_list(pages); | |
124 | goto out; | |
125 | } | |
126 | ||
127 | for (page_idx = 0; page_idx < nr_pages; page_idx++) { | |
128 | struct page *page = list_to_page(pages); | |
129 | list_del(&page->lru); | |
130 | if (!add_to_page_cache_lru(page, mapping, | |
131 | page->index, GFP_KERNEL)) { | |
132 | mapping->a_ops->readpage(filp, page); | |
133 | } | |
134 | page_cache_release(page); | |
135 | } | |
136 | ret = 0; | |
137 | ||
138 | out: | |
139 | blk_finish_plug(&plug); | |
140 | ||
141 | return ret; | |
142 | } | |
143 | ||
144 | /* | |
145 | * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all | |
146 | * the pages first, then submits them all for I/O. This avoids the very bad | |
147 | * behaviour which would occur if page allocations are causing VM writeback. | |
148 | * We really don't want to intermingle reads and writes like that. | |
149 | * | |
150 | * Returns the number of pages requested, or the maximum amount of I/O allowed. | |
151 | */ | |
152 | int __do_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
153 | pgoff_t offset, unsigned long nr_to_read, | |
154 | unsigned long lookahead_size) | |
155 | { | |
156 | struct inode *inode = mapping->host; | |
157 | struct page *page; | |
158 | unsigned long end_index; /* The last page we want to read */ | |
159 | LIST_HEAD(page_pool); | |
160 | int page_idx; | |
161 | int ret = 0; | |
162 | loff_t isize = i_size_read(inode); | |
163 | ||
164 | if (isize == 0) | |
165 | goto out; | |
166 | ||
167 | end_index = ((isize - 1) >> PAGE_CACHE_SHIFT); | |
168 | ||
169 | /* | |
170 | * Preallocate as many pages as we will need. | |
171 | */ | |
172 | for (page_idx = 0; page_idx < nr_to_read; page_idx++) { | |
173 | pgoff_t page_offset = offset + page_idx; | |
174 | ||
175 | if (page_offset > end_index) | |
176 | break; | |
177 | ||
178 | rcu_read_lock(); | |
179 | page = radix_tree_lookup(&mapping->page_tree, page_offset); | |
180 | rcu_read_unlock(); | |
181 | if (page && !radix_tree_exceptional_entry(page)) | |
182 | continue; | |
183 | ||
184 | page = page_cache_alloc_readahead(mapping); | |
185 | if (!page) | |
186 | break; | |
187 | page->index = page_offset; | |
188 | list_add(&page->lru, &page_pool); | |
189 | if (page_idx == nr_to_read - lookahead_size) | |
190 | SetPageReadahead(page); | |
191 | ret++; | |
192 | } | |
193 | ||
194 | /* | |
195 | * Now start the IO. We ignore I/O errors - if the page is not | |
196 | * uptodate then the caller will launch readpage again, and | |
197 | * will then handle the error. | |
198 | */ | |
199 | if (ret) | |
200 | read_pages(mapping, filp, &page_pool, ret); | |
201 | BUG_ON(!list_empty(&page_pool)); | |
202 | out: | |
203 | return ret; | |
204 | } | |
205 | ||
206 | /* | |
207 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much | |
208 | * memory at once. | |
209 | */ | |
210 | int force_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
211 | pgoff_t offset, unsigned long nr_to_read) | |
212 | { | |
213 | if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages)) | |
214 | return -EINVAL; | |
215 | ||
216 | nr_to_read = max_sane_readahead(nr_to_read); | |
217 | while (nr_to_read) { | |
218 | int err; | |
219 | ||
220 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE; | |
221 | ||
222 | if (this_chunk > nr_to_read) | |
223 | this_chunk = nr_to_read; | |
224 | err = __do_page_cache_readahead(mapping, filp, | |
225 | offset, this_chunk, 0); | |
226 | if (err < 0) | |
227 | return err; | |
228 | ||
229 | offset += this_chunk; | |
230 | nr_to_read -= this_chunk; | |
231 | } | |
232 | return 0; | |
233 | } | |
234 | ||
235 | #define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE) | |
236 | /* | |
237 | * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a | |
238 | * sensible upper limit. | |
239 | */ | |
240 | unsigned long max_sane_readahead(unsigned long nr) | |
241 | { | |
242 | return min(nr, MAX_READAHEAD); | |
243 | } | |
244 | ||
245 | /* | |
246 | * Set the initial window size, round to next power of 2 and square | |
247 | * for small size, x 4 for medium, and x 2 for large | |
248 | * for 128k (32 page) max ra | |
249 | * 1-8 page = 32k initial, > 8 page = 128k initial | |
250 | */ | |
251 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max) | |
252 | { | |
253 | unsigned long newsize = roundup_pow_of_two(size); | |
254 | ||
255 | if (newsize <= max / 32) | |
256 | newsize = newsize * 4; | |
257 | else if (newsize <= max / 4) | |
258 | newsize = newsize * 2; | |
259 | else | |
260 | newsize = max; | |
261 | ||
262 | return newsize; | |
263 | } | |
264 | ||
265 | /* | |
266 | * Get the previous window size, ramp it up, and | |
267 | * return it as the new window size. | |
268 | */ | |
269 | static unsigned long get_next_ra_size(struct file_ra_state *ra, | |
270 | unsigned long max) | |
271 | { | |
272 | unsigned long cur = ra->size; | |
273 | unsigned long newsize; | |
274 | ||
275 | if (cur < max / 16) | |
276 | newsize = 4 * cur; | |
277 | else | |
278 | newsize = 2 * cur; | |
279 | ||
280 | return min(newsize, max); | |
281 | } | |
282 | ||
283 | /* | |
284 | * On-demand readahead design. | |
285 | * | |
286 | * The fields in struct file_ra_state represent the most-recently-executed | |
287 | * readahead attempt: | |
288 | * | |
289 | * |<----- async_size ---------| | |
290 | * |------------------- size -------------------->| | |
291 | * |==================#===========================| | |
292 | * ^start ^page marked with PG_readahead | |
293 | * | |
294 | * To overlap application thinking time and disk I/O time, we do | |
295 | * `readahead pipelining': Do not wait until the application consumed all | |
296 | * readahead pages and stalled on the missing page at readahead_index; | |
297 | * Instead, submit an asynchronous readahead I/O as soon as there are | |
298 | * only async_size pages left in the readahead window. Normally async_size | |
299 | * will be equal to size, for maximum pipelining. | |
300 | * | |
301 | * In interleaved sequential reads, concurrent streams on the same fd can | |
302 | * be invalidating each other's readahead state. So we flag the new readahead | |
303 | * page at (start+size-async_size) with PG_readahead, and use it as readahead | |
304 | * indicator. The flag won't be set on already cached pages, to avoid the | |
305 | * readahead-for-nothing fuss, saving pointless page cache lookups. | |
306 | * | |
307 | * prev_pos tracks the last visited byte in the _previous_ read request. | |
308 | * It should be maintained by the caller, and will be used for detecting | |
309 | * small random reads. Note that the readahead algorithm checks loosely | |
310 | * for sequential patterns. Hence interleaved reads might be served as | |
311 | * sequential ones. | |
312 | * | |
313 | * There is a special-case: if the first page which the application tries to | |
314 | * read happens to be the first page of the file, it is assumed that a linear | |
315 | * read is about to happen and the window is immediately set to the initial size | |
316 | * based on I/O request size and the max_readahead. | |
317 | * | |
318 | * The code ramps up the readahead size aggressively at first, but slow down as | |
319 | * it approaches max_readhead. | |
320 | */ | |
321 | ||
322 | /* | |
323 | * Count contiguously cached pages from @offset-1 to @offset-@max, | |
324 | * this count is a conservative estimation of | |
325 | * - length of the sequential read sequence, or | |
326 | * - thrashing threshold in memory tight systems | |
327 | */ | |
328 | static pgoff_t count_history_pages(struct address_space *mapping, | |
329 | struct file_ra_state *ra, | |
330 | pgoff_t offset, unsigned long max) | |
331 | { | |
332 | pgoff_t head; | |
333 | ||
334 | rcu_read_lock(); | |
335 | head = page_cache_prev_hole(mapping, offset - 1, max); | |
336 | rcu_read_unlock(); | |
337 | ||
338 | return offset - 1 - head; | |
339 | } | |
340 | ||
341 | /* | |
342 | * page cache context based read-ahead | |
343 | */ | |
344 | static int try_context_readahead(struct address_space *mapping, | |
345 | struct file_ra_state *ra, | |
346 | pgoff_t offset, | |
347 | unsigned long req_size, | |
348 | unsigned long max) | |
349 | { | |
350 | pgoff_t size; | |
351 | ||
352 | size = count_history_pages(mapping, ra, offset, max); | |
353 | ||
354 | /* | |
355 | * not enough history pages: | |
356 | * it could be a random read | |
357 | */ | |
358 | if (size <= req_size) | |
359 | return 0; | |
360 | ||
361 | /* | |
362 | * starts from beginning of file: | |
363 | * it is a strong indication of long-run stream (or whole-file-read) | |
364 | */ | |
365 | if (size >= offset) | |
366 | size *= 2; | |
367 | ||
368 | ra->start = offset; | |
369 | ra->size = min(size + req_size, max); | |
370 | ra->async_size = 1; | |
371 | ||
372 | return 1; | |
373 | } | |
374 | ||
375 | /* | |
376 | * A minimal readahead algorithm for trivial sequential/random reads. | |
377 | */ | |
378 | static unsigned long | |
379 | ondemand_readahead(struct address_space *mapping, | |
380 | struct file_ra_state *ra, struct file *filp, | |
381 | bool hit_readahead_marker, pgoff_t offset, | |
382 | unsigned long req_size) | |
383 | { | |
384 | unsigned long max = max_sane_readahead(ra->ra_pages); | |
385 | pgoff_t prev_offset; | |
386 | ||
387 | /* | |
388 | * start of file | |
389 | */ | |
390 | if (!offset) | |
391 | goto initial_readahead; | |
392 | ||
393 | /* | |
394 | * It's the expected callback offset, assume sequential access. | |
395 | * Ramp up sizes, and push forward the readahead window. | |
396 | */ | |
397 | if ((offset == (ra->start + ra->size - ra->async_size) || | |
398 | offset == (ra->start + ra->size))) { | |
399 | ra->start += ra->size; | |
400 | ra->size = get_next_ra_size(ra, max); | |
401 | ra->async_size = ra->size; | |
402 | goto readit; | |
403 | } | |
404 | ||
405 | /* | |
406 | * Hit a marked page without valid readahead state. | |
407 | * E.g. interleaved reads. | |
408 | * Query the pagecache for async_size, which normally equals to | |
409 | * readahead size. Ramp it up and use it as the new readahead size. | |
410 | */ | |
411 | if (hit_readahead_marker) { | |
412 | pgoff_t start; | |
413 | ||
414 | rcu_read_lock(); | |
415 | start = page_cache_next_hole(mapping, offset + 1, max); | |
416 | rcu_read_unlock(); | |
417 | ||
418 | if (!start || start - offset > max) | |
419 | return 0; | |
420 | ||
421 | ra->start = start; | |
422 | ra->size = start - offset; /* old async_size */ | |
423 | ra->size += req_size; | |
424 | ra->size = get_next_ra_size(ra, max); | |
425 | ra->async_size = ra->size; | |
426 | goto readit; | |
427 | } | |
428 | ||
429 | /* | |
430 | * oversize read | |
431 | */ | |
432 | if (req_size > max) | |
433 | goto initial_readahead; | |
434 | ||
435 | /* | |
436 | * sequential cache miss | |
437 | * trivial case: (offset - prev_offset) == 1 | |
438 | * unaligned reads: (offset - prev_offset) == 0 | |
439 | */ | |
440 | prev_offset = (unsigned long long)ra->prev_pos >> PAGE_CACHE_SHIFT; | |
441 | if (offset - prev_offset <= 1UL) | |
442 | goto initial_readahead; | |
443 | ||
444 | /* | |
445 | * Query the page cache and look for the traces(cached history pages) | |
446 | * that a sequential stream would leave behind. | |
447 | */ | |
448 | if (try_context_readahead(mapping, ra, offset, req_size, max)) | |
449 | goto readit; | |
450 | ||
451 | /* | |
452 | * standalone, small random read | |
453 | * Read as is, and do not pollute the readahead state. | |
454 | */ | |
455 | return __do_page_cache_readahead(mapping, filp, offset, req_size, 0); | |
456 | ||
457 | initial_readahead: | |
458 | ra->start = offset; | |
459 | ra->size = get_init_ra_size(req_size, max); | |
460 | ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; | |
461 | ||
462 | readit: | |
463 | /* | |
464 | * Will this read hit the readahead marker made by itself? | |
465 | * If so, trigger the readahead marker hit now, and merge | |
466 | * the resulted next readahead window into the current one. | |
467 | */ | |
468 | if (offset == ra->start && ra->size == ra->async_size) { | |
469 | ra->async_size = get_next_ra_size(ra, max); | |
470 | ra->size += ra->async_size; | |
471 | } | |
472 | ||
473 | return ra_submit(ra, mapping, filp); | |
474 | } | |
475 | ||
476 | /** | |
477 | * page_cache_sync_readahead - generic file readahead | |
478 | * @mapping: address_space which holds the pagecache and I/O vectors | |
479 | * @ra: file_ra_state which holds the readahead state | |
480 | * @filp: passed on to ->readpage() and ->readpages() | |
481 | * @offset: start offset into @mapping, in pagecache page-sized units | |
482 | * @req_size: hint: total size of the read which the caller is performing in | |
483 | * pagecache pages | |
484 | * | |
485 | * page_cache_sync_readahead() should be called when a cache miss happened: | |
486 | * it will submit the read. The readahead logic may decide to piggyback more | |
487 | * pages onto the read request if access patterns suggest it will improve | |
488 | * performance. | |
489 | */ | |
490 | void page_cache_sync_readahead(struct address_space *mapping, | |
491 | struct file_ra_state *ra, struct file *filp, | |
492 | pgoff_t offset, unsigned long req_size) | |
493 | { | |
494 | /* no read-ahead */ | |
495 | if (!ra->ra_pages) | |
496 | return; | |
497 | ||
498 | /* be dumb */ | |
499 | if (filp && (filp->f_mode & FMODE_RANDOM)) { | |
500 | force_page_cache_readahead(mapping, filp, offset, req_size); | |
501 | return; | |
502 | } | |
503 | ||
504 | /* do read-ahead */ | |
505 | ondemand_readahead(mapping, ra, filp, false, offset, req_size); | |
506 | } | |
507 | EXPORT_SYMBOL_GPL(page_cache_sync_readahead); | |
508 | ||
509 | /** | |
510 | * page_cache_async_readahead - file readahead for marked pages | |
511 | * @mapping: address_space which holds the pagecache and I/O vectors | |
512 | * @ra: file_ra_state which holds the readahead state | |
513 | * @filp: passed on to ->readpage() and ->readpages() | |
514 | * @page: the page at @offset which has the PG_readahead flag set | |
515 | * @offset: start offset into @mapping, in pagecache page-sized units | |
516 | * @req_size: hint: total size of the read which the caller is performing in | |
517 | * pagecache pages | |
518 | * | |
519 | * page_cache_async_readahead() should be called when a page is used which | |
520 | * has the PG_readahead flag; this is a marker to suggest that the application | |
521 | * has used up enough of the readahead window that we should start pulling in | |
522 | * more pages. | |
523 | */ | |
524 | void | |
525 | page_cache_async_readahead(struct address_space *mapping, | |
526 | struct file_ra_state *ra, struct file *filp, | |
527 | struct page *page, pgoff_t offset, | |
528 | unsigned long req_size) | |
529 | { | |
530 | /* no read-ahead */ | |
531 | if (!ra->ra_pages) | |
532 | return; | |
533 | ||
534 | /* | |
535 | * Same bit is used for PG_readahead and PG_reclaim. | |
536 | */ | |
537 | if (PageWriteback(page)) | |
538 | return; | |
539 | ||
540 | ClearPageReadahead(page); | |
541 | ||
542 | /* | |
543 | * Defer asynchronous read-ahead on IO congestion. | |
544 | */ | |
545 | if (bdi_read_congested(mapping->backing_dev_info)) | |
546 | return; | |
547 | ||
548 | /* do read-ahead */ | |
549 | ondemand_readahead(mapping, ra, filp, true, offset, req_size); | |
550 | } | |
551 | EXPORT_SYMBOL_GPL(page_cache_async_readahead); | |
552 | ||
553 | static ssize_t | |
554 | do_readahead(struct address_space *mapping, struct file *filp, | |
555 | pgoff_t index, unsigned long nr) | |
556 | { | |
557 | if (!mapping || !mapping->a_ops) | |
558 | return -EINVAL; | |
559 | ||
560 | return force_page_cache_readahead(mapping, filp, index, nr); | |
561 | } | |
562 | ||
563 | SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count) | |
564 | { | |
565 | ssize_t ret; | |
566 | struct fd f; | |
567 | ||
568 | ret = -EBADF; | |
569 | f = fdget(fd); | |
570 | if (f.file) { | |
571 | if (f.file->f_mode & FMODE_READ) { | |
572 | struct address_space *mapping = f.file->f_mapping; | |
573 | pgoff_t start = offset >> PAGE_CACHE_SHIFT; | |
574 | pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT; | |
575 | unsigned long len = end - start + 1; | |
576 | ret = do_readahead(mapping, f.file, start, len); | |
577 | } | |
578 | fdput(f); | |
579 | } | |
580 | return ret; | |
581 | } |