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/dax.h>
12#include <linux/gfp.h>
13#include <linux/export.h>
14#include <linux/blkdev.h>
15#include <linux/backing-dev.h>
16#include <linux/task_io_accounting_ops.h>
17#include <linux/pagevec.h>
18#include <linux/pagemap.h>
19#include <linux/syscalls.h>
20#include <linux/file.h>
21#include <linux/mm_inline.h>
22#include <linux/blk-cgroup.h>
23#include <linux/fadvise.h>
24
25#include "internal.h"
26
27/*
28 * Initialise a struct file's readahead state. Assumes that the caller has
29 * memset *ra to zero.
30 */
31void
32file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
33{
34 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
35 ra->prev_pos = -1;
36}
37EXPORT_SYMBOL_GPL(file_ra_state_init);
38
39/*
40 * see if a page needs releasing upon read_cache_pages() failure
41 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
42 * before calling, such as the NFS fs marking pages that are cached locally
43 * on disk, thus we need to give the fs a chance to clean up in the event of
44 * an error
45 */
46static void read_cache_pages_invalidate_page(struct address_space *mapping,
47 struct page *page)
48{
49 if (page_has_private(page)) {
50 if (!trylock_page(page))
51 BUG();
52 page->mapping = mapping;
53 do_invalidatepage(page, 0, PAGE_SIZE);
54 page->mapping = NULL;
55 unlock_page(page);
56 }
57 put_page(page);
58}
59
60/*
61 * release a list of pages, invalidating them first if need be
62 */
63static void read_cache_pages_invalidate_pages(struct address_space *mapping,
64 struct list_head *pages)
65{
66 struct page *victim;
67
68 while (!list_empty(pages)) {
69 victim = lru_to_page(pages);
70 list_del(&victim->lru);
71 read_cache_pages_invalidate_page(mapping, victim);
72 }
73}
74
75/**
76 * read_cache_pages - populate an address space with some pages & start reads against them
77 * @mapping: the address_space
78 * @pages: The address of a list_head which contains the target pages. These
79 * pages have their ->index populated and are otherwise uninitialised.
80 * @filler: callback routine for filling a single page.
81 * @data: private data for the callback routine.
82 *
83 * Hides the details of the LRU cache etc from the filesystems.
84 *
85 * Returns: %0 on success, error return by @filler otherwise
86 */
87int read_cache_pages(struct address_space *mapping, struct list_head *pages,
88 int (*filler)(void *, struct page *), void *data)
89{
90 struct page *page;
91 int ret = 0;
92
93 while (!list_empty(pages)) {
94 page = lru_to_page(pages);
95 list_del(&page->lru);
96 if (add_to_page_cache_lru(page, mapping, page->index,
97 readahead_gfp_mask(mapping))) {
98 read_cache_pages_invalidate_page(mapping, page);
99 continue;
100 }
101 put_page(page);
102
103 ret = filler(data, page);
104 if (unlikely(ret)) {
105 read_cache_pages_invalidate_pages(mapping, pages);
106 break;
107 }
108 task_io_account_read(PAGE_SIZE);
109 }
110 return ret;
111}
112
113EXPORT_SYMBOL(read_cache_pages);
114
115static int read_pages(struct address_space *mapping, struct file *filp,
116 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
117{
118 struct blk_plug plug;
119 unsigned page_idx;
120 int ret;
121
122 blk_start_plug(&plug);
123
124 if (mapping->a_ops->readpages) {
125 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
126 /* Clean up the remaining pages */
127 put_pages_list(pages);
128 goto out;
129 }
130
131 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
132 struct page *page = lru_to_page(pages);
133 list_del(&page->lru);
134 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
135 mapping->a_ops->readpage(filp, page);
136 put_page(page);
137 }
138 ret = 0;
139
140out:
141 blk_finish_plug(&plug);
142
143 return ret;
144}
145
146/*
147 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates
148 * the pages first, then submits them for I/O. This avoids the very bad
149 * behaviour which would occur if page allocations are causing VM writeback.
150 * We really don't want to intermingle reads and writes like that.
151 *
152 * Returns the number of pages requested, or the maximum amount of I/O allowed.
153 */
154unsigned int __do_page_cache_readahead(struct address_space *mapping,
155 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
156 unsigned long lookahead_size)
157{
158 struct inode *inode = mapping->host;
159 struct page *page;
160 unsigned long end_index; /* The last page we want to read */
161 LIST_HEAD(page_pool);
162 int page_idx;
163 unsigned int nr_pages = 0;
164 loff_t isize = i_size_read(inode);
165 gfp_t gfp_mask = readahead_gfp_mask(mapping);
166
167 if (isize == 0)
168 goto out;
169
170 end_index = ((isize - 1) >> PAGE_SHIFT);
171
172 /*
173 * Preallocate as many pages as we will need.
174 */
175 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
176 pgoff_t page_offset = offset + page_idx;
177
178 if (page_offset > end_index)
179 break;
180
181 page = xa_load(&mapping->i_pages, page_offset);
182 if (page && !xa_is_value(page)) {
183 /*
184 * Page already present? Kick off the current batch of
185 * contiguous pages before continuing with the next
186 * batch.
187 */
188 if (nr_pages)
189 read_pages(mapping, filp, &page_pool, nr_pages,
190 gfp_mask);
191 nr_pages = 0;
192 continue;
193 }
194
195 page = __page_cache_alloc(gfp_mask);
196 if (!page)
197 break;
198 page->index = page_offset;
199 list_add(&page->lru, &page_pool);
200 if (page_idx == nr_to_read - lookahead_size)
201 SetPageReadahead(page);
202 nr_pages++;
203 }
204
205 /*
206 * Now start the IO. We ignore I/O errors - if the page is not
207 * uptodate then the caller will launch readpage again, and
208 * will then handle the error.
209 */
210 if (nr_pages)
211 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
212 BUG_ON(!list_empty(&page_pool));
213out:
214 return nr_pages;
215}
216
217/*
218 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
219 * memory at once.
220 */
221int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
222 pgoff_t offset, unsigned long nr_to_read)
223{
224 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
225 struct file_ra_state *ra = &filp->f_ra;
226 unsigned long max_pages;
227
228 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
229 return -EINVAL;
230
231 /*
232 * If the request exceeds the readahead window, allow the read to
233 * be up to the optimal hardware IO size
234 */
235 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
236 nr_to_read = min(nr_to_read, max_pages);
237 while (nr_to_read) {
238 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
239
240 if (this_chunk > nr_to_read)
241 this_chunk = nr_to_read;
242 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
243
244 offset += this_chunk;
245 nr_to_read -= this_chunk;
246 }
247 return 0;
248}
249
250/*
251 * Set the initial window size, round to next power of 2 and square
252 * for small size, x 4 for medium, and x 2 for large
253 * for 128k (32 page) max ra
254 * 1-8 page = 32k initial, > 8 page = 128k initial
255 */
256static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
257{
258 unsigned long newsize = roundup_pow_of_two(size);
259
260 if (newsize <= max / 32)
261 newsize = newsize * 4;
262 else if (newsize <= max / 4)
263 newsize = newsize * 2;
264 else
265 newsize = max;
266
267 return newsize;
268}
269
270/*
271 * Get the previous window size, ramp it up, and
272 * return it as the new window size.
273 */
274static unsigned long get_next_ra_size(struct file_ra_state *ra,
275 unsigned long max)
276{
277 unsigned long cur = ra->size;
278
279 if (cur < max / 16)
280 return 4 * cur;
281 if (cur <= max / 2)
282 return 2 * cur;
283 return max;
284}
285
286/*
287 * On-demand readahead design.
288 *
289 * The fields in struct file_ra_state represent the most-recently-executed
290 * readahead attempt:
291 *
292 * |<----- async_size ---------|
293 * |------------------- size -------------------->|
294 * |==================#===========================|
295 * ^start ^page marked with PG_readahead
296 *
297 * To overlap application thinking time and disk I/O time, we do
298 * `readahead pipelining': Do not wait until the application consumed all
299 * readahead pages and stalled on the missing page at readahead_index;
300 * Instead, submit an asynchronous readahead I/O as soon as there are
301 * only async_size pages left in the readahead window. Normally async_size
302 * will be equal to size, for maximum pipelining.
303 *
304 * In interleaved sequential reads, concurrent streams on the same fd can
305 * be invalidating each other's readahead state. So we flag the new readahead
306 * page at (start+size-async_size) with PG_readahead, and use it as readahead
307 * indicator. The flag won't be set on already cached pages, to avoid the
308 * readahead-for-nothing fuss, saving pointless page cache lookups.
309 *
310 * prev_pos tracks the last visited byte in the _previous_ read request.
311 * It should be maintained by the caller, and will be used for detecting
312 * small random reads. Note that the readahead algorithm checks loosely
313 * for sequential patterns. Hence interleaved reads might be served as
314 * sequential ones.
315 *
316 * There is a special-case: if the first page which the application tries to
317 * read happens to be the first page of the file, it is assumed that a linear
318 * read is about to happen and the window is immediately set to the initial size
319 * based on I/O request size and the max_readahead.
320 *
321 * The code ramps up the readahead size aggressively at first, but slow down as
322 * it approaches max_readhead.
323 */
324
325/*
326 * Count contiguously cached pages from @offset-1 to @offset-@max,
327 * this count is a conservative estimation of
328 * - length of the sequential read sequence, or
329 * - thrashing threshold in memory tight systems
330 */
331static pgoff_t count_history_pages(struct address_space *mapping,
332 pgoff_t offset, unsigned long max)
333{
334 pgoff_t head;
335
336 rcu_read_lock();
337 head = page_cache_prev_miss(mapping, offset - 1, max);
338 rcu_read_unlock();
339
340 return offset - 1 - head;
341}
342
343/*
344 * page cache context based read-ahead
345 */
346static int try_context_readahead(struct address_space *mapping,
347 struct file_ra_state *ra,
348 pgoff_t offset,
349 unsigned long req_size,
350 unsigned long max)
351{
352 pgoff_t size;
353
354 size = count_history_pages(mapping, offset, max);
355
356 /*
357 * not enough history pages:
358 * it could be a random read
359 */
360 if (size <= req_size)
361 return 0;
362
363 /*
364 * starts from beginning of file:
365 * it is a strong indication of long-run stream (or whole-file-read)
366 */
367 if (size >= offset)
368 size *= 2;
369
370 ra->start = offset;
371 ra->size = min(size + req_size, max);
372 ra->async_size = 1;
373
374 return 1;
375}
376
377/*
378 * A minimal readahead algorithm for trivial sequential/random reads.
379 */
380static unsigned long
381ondemand_readahead(struct address_space *mapping,
382 struct file_ra_state *ra, struct file *filp,
383 bool hit_readahead_marker, pgoff_t offset,
384 unsigned long req_size)
385{
386 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
387 unsigned long max_pages = ra->ra_pages;
388 unsigned long add_pages;
389 pgoff_t prev_offset;
390
391 /*
392 * If the request exceeds the readahead window, allow the read to
393 * be up to the optimal hardware IO size
394 */
395 if (req_size > max_pages && bdi->io_pages > max_pages)
396 max_pages = min(req_size, bdi->io_pages);
397
398 /*
399 * start of file
400 */
401 if (!offset)
402 goto initial_readahead;
403
404 /*
405 * It's the expected callback offset, assume sequential access.
406 * Ramp up sizes, and push forward the readahead window.
407 */
408 if ((offset == (ra->start + ra->size - ra->async_size) ||
409 offset == (ra->start + ra->size))) {
410 ra->start += ra->size;
411 ra->size = get_next_ra_size(ra, max_pages);
412 ra->async_size = ra->size;
413 goto readit;
414 }
415
416 /*
417 * Hit a marked page without valid readahead state.
418 * E.g. interleaved reads.
419 * Query the pagecache for async_size, which normally equals to
420 * readahead size. Ramp it up and use it as the new readahead size.
421 */
422 if (hit_readahead_marker) {
423 pgoff_t start;
424
425 rcu_read_lock();
426 start = page_cache_next_miss(mapping, offset + 1, max_pages);
427 rcu_read_unlock();
428
429 if (!start || start - offset > max_pages)
430 return 0;
431
432 ra->start = start;
433 ra->size = start - offset; /* old async_size */
434 ra->size += req_size;
435 ra->size = get_next_ra_size(ra, max_pages);
436 ra->async_size = ra->size;
437 goto readit;
438 }
439
440 /*
441 * oversize read
442 */
443 if (req_size > max_pages)
444 goto initial_readahead;
445
446 /*
447 * sequential cache miss
448 * trivial case: (offset - prev_offset) == 1
449 * unaligned reads: (offset - prev_offset) == 0
450 */
451 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
452 if (offset - prev_offset <= 1UL)
453 goto initial_readahead;
454
455 /*
456 * Query the page cache and look for the traces(cached history pages)
457 * that a sequential stream would leave behind.
458 */
459 if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
460 goto readit;
461
462 /*
463 * standalone, small random read
464 * Read as is, and do not pollute the readahead state.
465 */
466 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
467
468initial_readahead:
469 ra->start = offset;
470 ra->size = get_init_ra_size(req_size, max_pages);
471 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
472
473readit:
474 /*
475 * Will this read hit the readahead marker made by itself?
476 * If so, trigger the readahead marker hit now, and merge
477 * the resulted next readahead window into the current one.
478 * Take care of maximum IO pages as above.
479 */
480 if (offset == ra->start && ra->size == ra->async_size) {
481 add_pages = get_next_ra_size(ra, max_pages);
482 if (ra->size + add_pages <= max_pages) {
483 ra->async_size = add_pages;
484 ra->size += add_pages;
485 } else {
486 ra->size = max_pages;
487 ra->async_size = max_pages >> 1;
488 }
489 }
490
491 return ra_submit(ra, mapping, filp);
492}
493
494/**
495 * page_cache_sync_readahead - generic file readahead
496 * @mapping: address_space which holds the pagecache and I/O vectors
497 * @ra: file_ra_state which holds the readahead state
498 * @filp: passed on to ->readpage() and ->readpages()
499 * @offset: start offset into @mapping, in pagecache page-sized units
500 * @req_size: hint: total size of the read which the caller is performing in
501 * pagecache pages
502 *
503 * page_cache_sync_readahead() should be called when a cache miss happened:
504 * it will submit the read. The readahead logic may decide to piggyback more
505 * pages onto the read request if access patterns suggest it will improve
506 * performance.
507 */
508void page_cache_sync_readahead(struct address_space *mapping,
509 struct file_ra_state *ra, struct file *filp,
510 pgoff_t offset, unsigned long req_size)
511{
512 /* no read-ahead */
513 if (!ra->ra_pages)
514 return;
515
516 if (blk_cgroup_congested())
517 return;
518
519 /* be dumb */
520 if (filp && (filp->f_mode & FMODE_RANDOM)) {
521 force_page_cache_readahead(mapping, filp, offset, req_size);
522 return;
523 }
524
525 /* do read-ahead */
526 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
527}
528EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
529
530/**
531 * page_cache_async_readahead - file readahead for marked pages
532 * @mapping: address_space which holds the pagecache and I/O vectors
533 * @ra: file_ra_state which holds the readahead state
534 * @filp: passed on to ->readpage() and ->readpages()
535 * @page: the page at @offset which has the PG_readahead flag set
536 * @offset: start offset into @mapping, in pagecache page-sized units
537 * @req_size: hint: total size of the read which the caller is performing in
538 * pagecache pages
539 *
540 * page_cache_async_readahead() should be called when a page is used which
541 * has the PG_readahead flag; this is a marker to suggest that the application
542 * has used up enough of the readahead window that we should start pulling in
543 * more pages.
544 */
545void
546page_cache_async_readahead(struct address_space *mapping,
547 struct file_ra_state *ra, struct file *filp,
548 struct page *page, pgoff_t offset,
549 unsigned long req_size)
550{
551 /* no read-ahead */
552 if (!ra->ra_pages)
553 return;
554
555 /*
556 * Same bit is used for PG_readahead and PG_reclaim.
557 */
558 if (PageWriteback(page))
559 return;
560
561 ClearPageReadahead(page);
562
563 /*
564 * Defer asynchronous read-ahead on IO congestion.
565 */
566 if (inode_read_congested(mapping->host))
567 return;
568
569 if (blk_cgroup_congested())
570 return;
571
572 /* do read-ahead */
573 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
574}
575EXPORT_SYMBOL_GPL(page_cache_async_readahead);
576
577ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
578{
579 ssize_t ret;
580 struct fd f;
581
582 ret = -EBADF;
583 f = fdget(fd);
584 if (!f.file || !(f.file->f_mode & FMODE_READ))
585 goto out;
586
587 /*
588 * The readahead() syscall is intended to run only on files
589 * that can execute readahead. If readahead is not possible
590 * on this file, then we must return -EINVAL.
591 */
592 ret = -EINVAL;
593 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
594 !S_ISREG(file_inode(f.file)->i_mode))
595 goto out;
596
597 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
598out:
599 fdput(f);
600 return ret;
601}
602
603SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
604{
605 return ksys_readahead(fd, offset, count);
606}
607