Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/readahead.c - address_space-level file readahead.
4 *
5 * Copyright (C) 2002, Linus Torvalds
6 *
7 * 09Apr2002 Andrew Morton
8 * Initial version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/dax.h>
13#include <linux/gfp.h>
14#include <linux/export.h>
15#include <linux/blkdev.h>
16#include <linux/backing-dev.h>
17#include <linux/task_io_accounting_ops.h>
18#include <linux/pagevec.h>
19#include <linux/pagemap.h>
20#include <linux/syscalls.h>
21#include <linux/file.h>
22#include <linux/mm_inline.h>
23#include <linux/blk-cgroup.h>
24#include <linux/fadvise.h>
25
26#include "internal.h"
27
28/*
29 * Initialise a struct file's readahead state. Assumes that the caller has
30 * memset *ra to zero.
31 */
32void
33file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
34{
35 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
36 ra->prev_pos = -1;
37}
38EXPORT_SYMBOL_GPL(file_ra_state_init);
39
40/*
41 * see if a page needs releasing upon read_cache_pages() failure
42 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
43 * before calling, such as the NFS fs marking pages that are cached locally
44 * on disk, thus we need to give the fs a chance to clean up in the event of
45 * an error
46 */
47static void read_cache_pages_invalidate_page(struct address_space *mapping,
48 struct page *page)
49{
50 if (page_has_private(page)) {
51 if (!trylock_page(page))
52 BUG();
53 page->mapping = mapping;
54 do_invalidatepage(page, 0, PAGE_SIZE);
55 page->mapping = NULL;
56 unlock_page(page);
57 }
58 put_page(page);
59}
60
61/*
62 * release a list of pages, invalidating them first if need be
63 */
64static void read_cache_pages_invalidate_pages(struct address_space *mapping,
65 struct list_head *pages)
66{
67 struct page *victim;
68
69 while (!list_empty(pages)) {
70 victim = lru_to_page(pages);
71 list_del(&victim->lru);
72 read_cache_pages_invalidate_page(mapping, victim);
73 }
74}
75
76/**
77 * read_cache_pages - populate an address space with some pages & start reads against them
78 * @mapping: the address_space
79 * @pages: The address of a list_head which contains the target pages. These
80 * pages have their ->index populated and are otherwise uninitialised.
81 * @filler: callback routine for filling a single page.
82 * @data: private data for the callback routine.
83 *
84 * Hides the details of the LRU cache etc from the filesystems.
85 *
86 * Returns: %0 on success, error return by @filler otherwise
87 */
88int read_cache_pages(struct address_space *mapping, struct list_head *pages,
89 int (*filler)(void *, struct page *), void *data)
90{
91 struct page *page;
92 int ret = 0;
93
94 while (!list_empty(pages)) {
95 page = lru_to_page(pages);
96 list_del(&page->lru);
97 if (add_to_page_cache_lru(page, mapping, page->index,
98 readahead_gfp_mask(mapping))) {
99 read_cache_pages_invalidate_page(mapping, page);
100 continue;
101 }
102 put_page(page);
103
104 ret = filler(data, page);
105 if (unlikely(ret)) {
106 read_cache_pages_invalidate_pages(mapping, pages);
107 break;
108 }
109 task_io_account_read(PAGE_SIZE);
110 }
111 return ret;
112}
113
114EXPORT_SYMBOL(read_cache_pages);
115
116static int read_pages(struct address_space *mapping, struct file *filp,
117 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
118{
119 struct blk_plug plug;
120 unsigned page_idx;
121 int ret;
122
123 blk_start_plug(&plug);
124
125 if (mapping->a_ops->readpages) {
126 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
127 /* Clean up the remaining pages */
128 put_pages_list(pages);
129 goto out;
130 }
131
132 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
133 struct page *page = lru_to_page(pages);
134 list_del(&page->lru);
135 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
136 mapping->a_ops->readpage(filp, page);
137 put_page(page);
138 }
139 ret = 0;
140
141out:
142 blk_finish_plug(&plug);
143
144 return ret;
145}
146
147/*
148 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates
149 * the pages first, then submits them for I/O. This avoids the very bad
150 * behaviour which would occur if page allocations are causing VM writeback.
151 * We really don't want to intermingle reads and writes like that.
152 *
153 * Returns the number of pages requested, or the maximum amount of I/O allowed.
154 */
155unsigned int __do_page_cache_readahead(struct address_space *mapping,
156 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
157 unsigned long lookahead_size)
158{
159 struct inode *inode = mapping->host;
160 struct page *page;
161 unsigned long end_index; /* The last page we want to read */
162 LIST_HEAD(page_pool);
163 int page_idx;
164 unsigned int nr_pages = 0;
165 loff_t isize = i_size_read(inode);
166 gfp_t gfp_mask = readahead_gfp_mask(mapping);
167
168 if (isize == 0)
169 goto out;
170
171 end_index = ((isize - 1) >> PAGE_SHIFT);
172
173 /*
174 * Preallocate as many pages as we will need.
175 */
176 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
177 pgoff_t page_offset = offset + page_idx;
178
179 if (page_offset > end_index)
180 break;
181
182 page = xa_load(&mapping->i_pages, page_offset);
183 if (page && !xa_is_value(page)) {
184 /*
185 * Page already present? Kick off the current batch of
186 * contiguous pages before continuing with the next
187 * batch.
188 */
189 if (nr_pages)
190 read_pages(mapping, filp, &page_pool, nr_pages,
191 gfp_mask);
192 nr_pages = 0;
193 continue;
194 }
195
196 page = __page_cache_alloc(gfp_mask);
197 if (!page)
198 break;
199 page->index = page_offset;
200 list_add(&page->lru, &page_pool);
201 if (page_idx == nr_to_read - lookahead_size)
202 SetPageReadahead(page);
203 nr_pages++;
204 }
205
206 /*
207 * Now start the IO. We ignore I/O errors - if the page is not
208 * uptodate then the caller will launch readpage again, and
209 * will then handle the error.
210 */
211 if (nr_pages)
212 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
213 BUG_ON(!list_empty(&page_pool));
214out:
215 return nr_pages;
216}
217
218/*
219 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
220 * memory at once.
221 */
222int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
223 pgoff_t offset, unsigned long nr_to_read)
224{
225 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
226 struct file_ra_state *ra = &filp->f_ra;
227 unsigned long max_pages;
228
229 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
230 return -EINVAL;
231
232 /*
233 * If the request exceeds the readahead window, allow the read to
234 * be up to the optimal hardware IO size
235 */
236 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
237 nr_to_read = min(nr_to_read, max_pages);
238 while (nr_to_read) {
239 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
240
241 if (this_chunk > nr_to_read)
242 this_chunk = nr_to_read;
243 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
244
245 offset += this_chunk;
246 nr_to_read -= this_chunk;
247 }
248 return 0;
249}
250
251/*
252 * Set the initial window size, round to next power of 2 and square
253 * for small size, x 4 for medium, and x 2 for large
254 * for 128k (32 page) max ra
255 * 1-8 page = 32k initial, > 8 page = 128k initial
256 */
257static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
258{
259 unsigned long newsize = roundup_pow_of_two(size);
260
261 if (newsize <= max / 32)
262 newsize = newsize * 4;
263 else if (newsize <= max / 4)
264 newsize = newsize * 2;
265 else
266 newsize = max;
267
268 return newsize;
269}
270
271/*
272 * Get the previous window size, ramp it up, and
273 * return it as the new window size.
274 */
275static unsigned long get_next_ra_size(struct file_ra_state *ra,
276 unsigned long max)
277{
278 unsigned long cur = ra->size;
279
280 if (cur < max / 16)
281 return 4 * cur;
282 if (cur <= max / 2)
283 return 2 * cur;
284 return max;
285}
286
287/*
288 * On-demand readahead design.
289 *
290 * The fields in struct file_ra_state represent the most-recently-executed
291 * readahead attempt:
292 *
293 * |<----- async_size ---------|
294 * |------------------- size -------------------->|
295 * |==================#===========================|
296 * ^start ^page marked with PG_readahead
297 *
298 * To overlap application thinking time and disk I/O time, we do
299 * `readahead pipelining': Do not wait until the application consumed all
300 * readahead pages and stalled on the missing page at readahead_index;
301 * Instead, submit an asynchronous readahead I/O as soon as there are
302 * only async_size pages left in the readahead window. Normally async_size
303 * will be equal to size, for maximum pipelining.
304 *
305 * In interleaved sequential reads, concurrent streams on the same fd can
306 * be invalidating each other's readahead state. So we flag the new readahead
307 * page at (start+size-async_size) with PG_readahead, and use it as readahead
308 * indicator. The flag won't be set on already cached pages, to avoid the
309 * readahead-for-nothing fuss, saving pointless page cache lookups.
310 *
311 * prev_pos tracks the last visited byte in the _previous_ read request.
312 * It should be maintained by the caller, and will be used for detecting
313 * small random reads. Note that the readahead algorithm checks loosely
314 * for sequential patterns. Hence interleaved reads might be served as
315 * sequential ones.
316 *
317 * There is a special-case: if the first page which the application tries to
318 * read happens to be the first page of the file, it is assumed that a linear
319 * read is about to happen and the window is immediately set to the initial size
320 * based on I/O request size and the max_readahead.
321 *
322 * The code ramps up the readahead size aggressively at first, but slow down as
323 * it approaches max_readhead.
324 */
325
326/*
327 * Count contiguously cached pages from @offset-1 to @offset-@max,
328 * this count is a conservative estimation of
329 * - length of the sequential read sequence, or
330 * - thrashing threshold in memory tight systems
331 */
332static pgoff_t count_history_pages(struct address_space *mapping,
333 pgoff_t offset, unsigned long max)
334{
335 pgoff_t head;
336
337 rcu_read_lock();
338 head = page_cache_prev_miss(mapping, offset - 1, max);
339 rcu_read_unlock();
340
341 return offset - 1 - head;
342}
343
344/*
345 * page cache context based read-ahead
346 */
347static int try_context_readahead(struct address_space *mapping,
348 struct file_ra_state *ra,
349 pgoff_t offset,
350 unsigned long req_size,
351 unsigned long max)
352{
353 pgoff_t size;
354
355 size = count_history_pages(mapping, offset, max);
356
357 /*
358 * not enough history pages:
359 * it could be a random read
360 */
361 if (size <= req_size)
362 return 0;
363
364 /*
365 * starts from beginning of file:
366 * it is a strong indication of long-run stream (or whole-file-read)
367 */
368 if (size >= offset)
369 size *= 2;
370
371 ra->start = offset;
372 ra->size = min(size + req_size, max);
373 ra->async_size = 1;
374
375 return 1;
376}
377
378/*
379 * A minimal readahead algorithm for trivial sequential/random reads.
380 */
381static unsigned long
382ondemand_readahead(struct address_space *mapping,
383 struct file_ra_state *ra, struct file *filp,
384 bool hit_readahead_marker, pgoff_t offset,
385 unsigned long req_size)
386{
387 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
388 unsigned long max_pages = ra->ra_pages;
389 unsigned long add_pages;
390 pgoff_t prev_offset;
391
392 /*
393 * If the request exceeds the readahead window, allow the read to
394 * be up to the optimal hardware IO size
395 */
396 if (req_size > max_pages && bdi->io_pages > max_pages)
397 max_pages = min(req_size, bdi->io_pages);
398
399 /*
400 * start of file
401 */
402 if (!offset)
403 goto initial_readahead;
404
405 /*
406 * It's the expected callback offset, assume sequential access.
407 * Ramp up sizes, and push forward the readahead window.
408 */
409 if ((offset == (ra->start + ra->size - ra->async_size) ||
410 offset == (ra->start + ra->size))) {
411 ra->start += ra->size;
412 ra->size = get_next_ra_size(ra, max_pages);
413 ra->async_size = ra->size;
414 goto readit;
415 }
416
417 /*
418 * Hit a marked page without valid readahead state.
419 * E.g. interleaved reads.
420 * Query the pagecache for async_size, which normally equals to
421 * readahead size. Ramp it up and use it as the new readahead size.
422 */
423 if (hit_readahead_marker) {
424 pgoff_t start;
425
426 rcu_read_lock();
427 start = page_cache_next_miss(mapping, offset + 1, max_pages);
428 rcu_read_unlock();
429
430 if (!start || start - offset > max_pages)
431 return 0;
432
433 ra->start = start;
434 ra->size = start - offset; /* old async_size */
435 ra->size += req_size;
436 ra->size = get_next_ra_size(ra, max_pages);
437 ra->async_size = ra->size;
438 goto readit;
439 }
440
441 /*
442 * oversize read
443 */
444 if (req_size > max_pages)
445 goto initial_readahead;
446
447 /*
448 * sequential cache miss
449 * trivial case: (offset - prev_offset) == 1
450 * unaligned reads: (offset - prev_offset) == 0
451 */
452 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
453 if (offset - prev_offset <= 1UL)
454 goto initial_readahead;
455
456 /*
457 * Query the page cache and look for the traces(cached history pages)
458 * that a sequential stream would leave behind.
459 */
460 if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
461 goto readit;
462
463 /*
464 * standalone, small random read
465 * Read as is, and do not pollute the readahead state.
466 */
467 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
468
469initial_readahead:
470 ra->start = offset;
471 ra->size = get_init_ra_size(req_size, max_pages);
472 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
473
474readit:
475 /*
476 * Will this read hit the readahead marker made by itself?
477 * If so, trigger the readahead marker hit now, and merge
478 * the resulted next readahead window into the current one.
479 * Take care of maximum IO pages as above.
480 */
481 if (offset == ra->start && ra->size == ra->async_size) {
482 add_pages = get_next_ra_size(ra, max_pages);
483 if (ra->size + add_pages <= max_pages) {
484 ra->async_size = add_pages;
485 ra->size += add_pages;
486 } else {
487 ra->size = max_pages;
488 ra->async_size = max_pages >> 1;
489 }
490 }
491
492 return ra_submit(ra, mapping, filp);
493}
494
495/**
496 * page_cache_sync_readahead - generic file readahead
497 * @mapping: address_space which holds the pagecache and I/O vectors
498 * @ra: file_ra_state which holds the readahead state
499 * @filp: passed on to ->readpage() and ->readpages()
500 * @offset: start offset into @mapping, in pagecache page-sized units
501 * @req_size: hint: total size of the read which the caller is performing in
502 * pagecache pages
503 *
504 * page_cache_sync_readahead() should be called when a cache miss happened:
505 * it will submit the read. The readahead logic may decide to piggyback more
506 * pages onto the read request if access patterns suggest it will improve
507 * performance.
508 */
509void page_cache_sync_readahead(struct address_space *mapping,
510 struct file_ra_state *ra, struct file *filp,
511 pgoff_t offset, unsigned long req_size)
512{
513 /* no read-ahead */
514 if (!ra->ra_pages)
515 return;
516
517 if (blk_cgroup_congested())
518 return;
519
520 /* be dumb */
521 if (filp && (filp->f_mode & FMODE_RANDOM)) {
522 force_page_cache_readahead(mapping, filp, offset, req_size);
523 return;
524 }
525
526 /* do read-ahead */
527 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
528}
529EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
530
531/**
532 * page_cache_async_readahead - file readahead for marked pages
533 * @mapping: address_space which holds the pagecache and I/O vectors
534 * @ra: file_ra_state which holds the readahead state
535 * @filp: passed on to ->readpage() and ->readpages()
536 * @page: the page at @offset which has the PG_readahead flag set
537 * @offset: start offset into @mapping, in pagecache page-sized units
538 * @req_size: hint: total size of the read which the caller is performing in
539 * pagecache pages
540 *
541 * page_cache_async_readahead() should be called when a page is used which
542 * has the PG_readahead flag; this is a marker to suggest that the application
543 * has used up enough of the readahead window that we should start pulling in
544 * more pages.
545 */
546void
547page_cache_async_readahead(struct address_space *mapping,
548 struct file_ra_state *ra, struct file *filp,
549 struct page *page, pgoff_t offset,
550 unsigned long req_size)
551{
552 /* no read-ahead */
553 if (!ra->ra_pages)
554 return;
555
556 /*
557 * Same bit is used for PG_readahead and PG_reclaim.
558 */
559 if (PageWriteback(page))
560 return;
561
562 ClearPageReadahead(page);
563
564 /*
565 * Defer asynchronous read-ahead on IO congestion.
566 */
567 if (inode_read_congested(mapping->host))
568 return;
569
570 if (blk_cgroup_congested())
571 return;
572
573 /* do read-ahead */
574 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
575}
576EXPORT_SYMBOL_GPL(page_cache_async_readahead);
577
578ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
579{
580 ssize_t ret;
581 struct fd f;
582
583 ret = -EBADF;
584 f = fdget(fd);
585 if (!f.file || !(f.file->f_mode & FMODE_READ))
586 goto out;
587
588 /*
589 * The readahead() syscall is intended to run only on files
590 * that can execute readahead. If readahead is not possible
591 * on this file, then we must return -EINVAL.
592 */
593 ret = -EINVAL;
594 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
595 !S_ISREG(file_inode(f.file)->i_mode))
596 goto out;
597
598 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
599out:
600 fdput(f);
601 return ret;
602}
603
604SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
605{
606 return ksys_readahead(fd, offset, count);
607}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/readahead.c - address_space-level file readahead.
4 *
5 * Copyright (C) 2002, Linus Torvalds
6 *
7 * 09Apr2002 Andrew Morton
8 * Initial version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/dax.h>
13#include <linux/gfp.h>
14#include <linux/export.h>
15#include <linux/blkdev.h>
16#include <linux/backing-dev.h>
17#include <linux/task_io_accounting_ops.h>
18#include <linux/pagevec.h>
19#include <linux/pagemap.h>
20#include <linux/syscalls.h>
21#include <linux/file.h>
22#include <linux/mm_inline.h>
23#include <linux/blk-cgroup.h>
24#include <linux/fadvise.h>
25#include <linux/sched/mm.h>
26
27#include "internal.h"
28
29/*
30 * Initialise a struct file's readahead state. Assumes that the caller has
31 * memset *ra to zero.
32 */
33void
34file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
35{
36 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
37 ra->prev_pos = -1;
38}
39EXPORT_SYMBOL_GPL(file_ra_state_init);
40
41/*
42 * see if a page needs releasing upon read_cache_pages() failure
43 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
44 * before calling, such as the NFS fs marking pages that are cached locally
45 * on disk, thus we need to give the fs a chance to clean up in the event of
46 * an error
47 */
48static void read_cache_pages_invalidate_page(struct address_space *mapping,
49 struct page *page)
50{
51 if (page_has_private(page)) {
52 if (!trylock_page(page))
53 BUG();
54 page->mapping = mapping;
55 do_invalidatepage(page, 0, PAGE_SIZE);
56 page->mapping = NULL;
57 unlock_page(page);
58 }
59 put_page(page);
60}
61
62/*
63 * release a list of pages, invalidating them first if need be
64 */
65static void read_cache_pages_invalidate_pages(struct address_space *mapping,
66 struct list_head *pages)
67{
68 struct page *victim;
69
70 while (!list_empty(pages)) {
71 victim = lru_to_page(pages);
72 list_del(&victim->lru);
73 read_cache_pages_invalidate_page(mapping, victim);
74 }
75}
76
77/**
78 * read_cache_pages - populate an address space with some pages & start reads against them
79 * @mapping: the address_space
80 * @pages: The address of a list_head which contains the target pages. These
81 * pages have their ->index populated and are otherwise uninitialised.
82 * @filler: callback routine for filling a single page.
83 * @data: private data for the callback routine.
84 *
85 * Hides the details of the LRU cache etc from the filesystems.
86 *
87 * Returns: %0 on success, error return by @filler otherwise
88 */
89int read_cache_pages(struct address_space *mapping, struct list_head *pages,
90 int (*filler)(void *, struct page *), void *data)
91{
92 struct page *page;
93 int ret = 0;
94
95 while (!list_empty(pages)) {
96 page = lru_to_page(pages);
97 list_del(&page->lru);
98 if (add_to_page_cache_lru(page, mapping, page->index,
99 readahead_gfp_mask(mapping))) {
100 read_cache_pages_invalidate_page(mapping, page);
101 continue;
102 }
103 put_page(page);
104
105 ret = filler(data, page);
106 if (unlikely(ret)) {
107 read_cache_pages_invalidate_pages(mapping, pages);
108 break;
109 }
110 task_io_account_read(PAGE_SIZE);
111 }
112 return ret;
113}
114
115EXPORT_SYMBOL(read_cache_pages);
116
117static void read_pages(struct readahead_control *rac, struct list_head *pages,
118 bool skip_page)
119{
120 const struct address_space_operations *aops = rac->mapping->a_ops;
121 struct page *page;
122 struct blk_plug plug;
123
124 if (!readahead_count(rac))
125 goto out;
126
127 blk_start_plug(&plug);
128
129 if (aops->readahead) {
130 aops->readahead(rac);
131 /* Clean up the remaining pages */
132 while ((page = readahead_page(rac))) {
133 unlock_page(page);
134 put_page(page);
135 }
136 } else if (aops->readpages) {
137 aops->readpages(rac->file, rac->mapping, pages,
138 readahead_count(rac));
139 /* Clean up the remaining pages */
140 put_pages_list(pages);
141 rac->_index += rac->_nr_pages;
142 rac->_nr_pages = 0;
143 } else {
144 while ((page = readahead_page(rac))) {
145 aops->readpage(rac->file, page);
146 put_page(page);
147 }
148 }
149
150 blk_finish_plug(&plug);
151
152 BUG_ON(!list_empty(pages));
153 BUG_ON(readahead_count(rac));
154
155out:
156 if (skip_page)
157 rac->_index++;
158}
159
160/**
161 * page_cache_ra_unbounded - Start unchecked readahead.
162 * @ractl: Readahead control.
163 * @nr_to_read: The number of pages to read.
164 * @lookahead_size: Where to start the next readahead.
165 *
166 * This function is for filesystems to call when they want to start
167 * readahead beyond a file's stated i_size. This is almost certainly
168 * not the function you want to call. Use page_cache_async_readahead()
169 * or page_cache_sync_readahead() instead.
170 *
171 * Context: File is referenced by caller. Mutexes may be held by caller.
172 * May sleep, but will not reenter filesystem to reclaim memory.
173 */
174void page_cache_ra_unbounded(struct readahead_control *ractl,
175 unsigned long nr_to_read, unsigned long lookahead_size)
176{
177 struct address_space *mapping = ractl->mapping;
178 unsigned long index = readahead_index(ractl);
179 LIST_HEAD(page_pool);
180 gfp_t gfp_mask = readahead_gfp_mask(mapping);
181 unsigned long i;
182
183 /*
184 * Partway through the readahead operation, we will have added
185 * locked pages to the page cache, but will not yet have submitted
186 * them for I/O. Adding another page may need to allocate memory,
187 * which can trigger memory reclaim. Telling the VM we're in
188 * the middle of a filesystem operation will cause it to not
189 * touch file-backed pages, preventing a deadlock. Most (all?)
190 * filesystems already specify __GFP_NOFS in their mapping's
191 * gfp_mask, but let's be explicit here.
192 */
193 unsigned int nofs = memalloc_nofs_save();
194
195 /*
196 * Preallocate as many pages as we will need.
197 */
198 for (i = 0; i < nr_to_read; i++) {
199 struct page *page = xa_load(&mapping->i_pages, index + i);
200
201 if (page && !xa_is_value(page)) {
202 /*
203 * Page already present? Kick off the current batch
204 * of contiguous pages before continuing with the
205 * next batch. This page may be the one we would
206 * have intended to mark as Readahead, but we don't
207 * have a stable reference to this page, and it's
208 * not worth getting one just for that.
209 */
210 read_pages(ractl, &page_pool, true);
211 i = ractl->_index + ractl->_nr_pages - index - 1;
212 continue;
213 }
214
215 page = __page_cache_alloc(gfp_mask);
216 if (!page)
217 break;
218 if (mapping->a_ops->readpages) {
219 page->index = index + i;
220 list_add(&page->lru, &page_pool);
221 } else if (add_to_page_cache_lru(page, mapping, index + i,
222 gfp_mask) < 0) {
223 put_page(page);
224 read_pages(ractl, &page_pool, true);
225 i = ractl->_index + ractl->_nr_pages - index - 1;
226 continue;
227 }
228 if (i == nr_to_read - lookahead_size)
229 SetPageReadahead(page);
230 ractl->_nr_pages++;
231 }
232
233 /*
234 * Now start the IO. We ignore I/O errors - if the page is not
235 * uptodate then the caller will launch readpage again, and
236 * will then handle the error.
237 */
238 read_pages(ractl, &page_pool, false);
239 memalloc_nofs_restore(nofs);
240}
241EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
242
243/*
244 * do_page_cache_ra() actually reads a chunk of disk. It allocates
245 * the pages first, then submits them for I/O. This avoids the very bad
246 * behaviour which would occur if page allocations are causing VM writeback.
247 * We really don't want to intermingle reads and writes like that.
248 */
249void do_page_cache_ra(struct readahead_control *ractl,
250 unsigned long nr_to_read, unsigned long lookahead_size)
251{
252 struct inode *inode = ractl->mapping->host;
253 unsigned long index = readahead_index(ractl);
254 loff_t isize = i_size_read(inode);
255 pgoff_t end_index; /* The last page we want to read */
256
257 if (isize == 0)
258 return;
259
260 end_index = (isize - 1) >> PAGE_SHIFT;
261 if (index > end_index)
262 return;
263 /* Don't read past the page containing the last byte of the file */
264 if (nr_to_read > end_index - index)
265 nr_to_read = end_index - index + 1;
266
267 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
268}
269
270/*
271 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
272 * memory at once.
273 */
274void force_page_cache_ra(struct readahead_control *ractl,
275 unsigned long nr_to_read)
276{
277 struct address_space *mapping = ractl->mapping;
278 struct file_ra_state *ra = ractl->ra;
279 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
280 unsigned long max_pages, index;
281
282 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages &&
283 !mapping->a_ops->readahead))
284 return;
285
286 /*
287 * If the request exceeds the readahead window, allow the read to
288 * be up to the optimal hardware IO size
289 */
290 index = readahead_index(ractl);
291 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
292 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
293 while (nr_to_read) {
294 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
295
296 if (this_chunk > nr_to_read)
297 this_chunk = nr_to_read;
298 ractl->_index = index;
299 do_page_cache_ra(ractl, this_chunk, 0);
300
301 index += this_chunk;
302 nr_to_read -= this_chunk;
303 }
304}
305
306/*
307 * Set the initial window size, round to next power of 2 and square
308 * for small size, x 4 for medium, and x 2 for large
309 * for 128k (32 page) max ra
310 * 1-8 page = 32k initial, > 8 page = 128k initial
311 */
312static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
313{
314 unsigned long newsize = roundup_pow_of_two(size);
315
316 if (newsize <= max / 32)
317 newsize = newsize * 4;
318 else if (newsize <= max / 4)
319 newsize = newsize * 2;
320 else
321 newsize = max;
322
323 return newsize;
324}
325
326/*
327 * Get the previous window size, ramp it up, and
328 * return it as the new window size.
329 */
330static unsigned long get_next_ra_size(struct file_ra_state *ra,
331 unsigned long max)
332{
333 unsigned long cur = ra->size;
334
335 if (cur < max / 16)
336 return 4 * cur;
337 if (cur <= max / 2)
338 return 2 * cur;
339 return max;
340}
341
342/*
343 * On-demand readahead design.
344 *
345 * The fields in struct file_ra_state represent the most-recently-executed
346 * readahead attempt:
347 *
348 * |<----- async_size ---------|
349 * |------------------- size -------------------->|
350 * |==================#===========================|
351 * ^start ^page marked with PG_readahead
352 *
353 * To overlap application thinking time and disk I/O time, we do
354 * `readahead pipelining': Do not wait until the application consumed all
355 * readahead pages and stalled on the missing page at readahead_index;
356 * Instead, submit an asynchronous readahead I/O as soon as there are
357 * only async_size pages left in the readahead window. Normally async_size
358 * will be equal to size, for maximum pipelining.
359 *
360 * In interleaved sequential reads, concurrent streams on the same fd can
361 * be invalidating each other's readahead state. So we flag the new readahead
362 * page at (start+size-async_size) with PG_readahead, and use it as readahead
363 * indicator. The flag won't be set on already cached pages, to avoid the
364 * readahead-for-nothing fuss, saving pointless page cache lookups.
365 *
366 * prev_pos tracks the last visited byte in the _previous_ read request.
367 * It should be maintained by the caller, and will be used for detecting
368 * small random reads. Note that the readahead algorithm checks loosely
369 * for sequential patterns. Hence interleaved reads might be served as
370 * sequential ones.
371 *
372 * There is a special-case: if the first page which the application tries to
373 * read happens to be the first page of the file, it is assumed that a linear
374 * read is about to happen and the window is immediately set to the initial size
375 * based on I/O request size and the max_readahead.
376 *
377 * The code ramps up the readahead size aggressively at first, but slow down as
378 * it approaches max_readhead.
379 */
380
381/*
382 * Count contiguously cached pages from @index-1 to @index-@max,
383 * this count is a conservative estimation of
384 * - length of the sequential read sequence, or
385 * - thrashing threshold in memory tight systems
386 */
387static pgoff_t count_history_pages(struct address_space *mapping,
388 pgoff_t index, unsigned long max)
389{
390 pgoff_t head;
391
392 rcu_read_lock();
393 head = page_cache_prev_miss(mapping, index - 1, max);
394 rcu_read_unlock();
395
396 return index - 1 - head;
397}
398
399/*
400 * page cache context based read-ahead
401 */
402static int try_context_readahead(struct address_space *mapping,
403 struct file_ra_state *ra,
404 pgoff_t index,
405 unsigned long req_size,
406 unsigned long max)
407{
408 pgoff_t size;
409
410 size = count_history_pages(mapping, index, max);
411
412 /*
413 * not enough history pages:
414 * it could be a random read
415 */
416 if (size <= req_size)
417 return 0;
418
419 /*
420 * starts from beginning of file:
421 * it is a strong indication of long-run stream (or whole-file-read)
422 */
423 if (size >= index)
424 size *= 2;
425
426 ra->start = index;
427 ra->size = min(size + req_size, max);
428 ra->async_size = 1;
429
430 return 1;
431}
432
433/*
434 * A minimal readahead algorithm for trivial sequential/random reads.
435 */
436static void ondemand_readahead(struct readahead_control *ractl,
437 bool hit_readahead_marker, unsigned long req_size)
438{
439 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
440 struct file_ra_state *ra = ractl->ra;
441 unsigned long max_pages = ra->ra_pages;
442 unsigned long add_pages;
443 unsigned long index = readahead_index(ractl);
444 pgoff_t prev_index;
445
446 /*
447 * If the request exceeds the readahead window, allow the read to
448 * be up to the optimal hardware IO size
449 */
450 if (req_size > max_pages && bdi->io_pages > max_pages)
451 max_pages = min(req_size, bdi->io_pages);
452
453 /*
454 * start of file
455 */
456 if (!index)
457 goto initial_readahead;
458
459 /*
460 * It's the expected callback index, assume sequential access.
461 * Ramp up sizes, and push forward the readahead window.
462 */
463 if ((index == (ra->start + ra->size - ra->async_size) ||
464 index == (ra->start + ra->size))) {
465 ra->start += ra->size;
466 ra->size = get_next_ra_size(ra, max_pages);
467 ra->async_size = ra->size;
468 goto readit;
469 }
470
471 /*
472 * Hit a marked page without valid readahead state.
473 * E.g. interleaved reads.
474 * Query the pagecache for async_size, which normally equals to
475 * readahead size. Ramp it up and use it as the new readahead size.
476 */
477 if (hit_readahead_marker) {
478 pgoff_t start;
479
480 rcu_read_lock();
481 start = page_cache_next_miss(ractl->mapping, index + 1,
482 max_pages);
483 rcu_read_unlock();
484
485 if (!start || start - index > max_pages)
486 return;
487
488 ra->start = start;
489 ra->size = start - index; /* old async_size */
490 ra->size += req_size;
491 ra->size = get_next_ra_size(ra, max_pages);
492 ra->async_size = ra->size;
493 goto readit;
494 }
495
496 /*
497 * oversize read
498 */
499 if (req_size > max_pages)
500 goto initial_readahead;
501
502 /*
503 * sequential cache miss
504 * trivial case: (index - prev_index) == 1
505 * unaligned reads: (index - prev_index) == 0
506 */
507 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
508 if (index - prev_index <= 1UL)
509 goto initial_readahead;
510
511 /*
512 * Query the page cache and look for the traces(cached history pages)
513 * that a sequential stream would leave behind.
514 */
515 if (try_context_readahead(ractl->mapping, ra, index, req_size,
516 max_pages))
517 goto readit;
518
519 /*
520 * standalone, small random read
521 * Read as is, and do not pollute the readahead state.
522 */
523 do_page_cache_ra(ractl, req_size, 0);
524 return;
525
526initial_readahead:
527 ra->start = index;
528 ra->size = get_init_ra_size(req_size, max_pages);
529 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
530
531readit:
532 /*
533 * Will this read hit the readahead marker made by itself?
534 * If so, trigger the readahead marker hit now, and merge
535 * the resulted next readahead window into the current one.
536 * Take care of maximum IO pages as above.
537 */
538 if (index == ra->start && ra->size == ra->async_size) {
539 add_pages = get_next_ra_size(ra, max_pages);
540 if (ra->size + add_pages <= max_pages) {
541 ra->async_size = add_pages;
542 ra->size += add_pages;
543 } else {
544 ra->size = max_pages;
545 ra->async_size = max_pages >> 1;
546 }
547 }
548
549 ractl->_index = ra->start;
550 do_page_cache_ra(ractl, ra->size, ra->async_size);
551}
552
553void page_cache_sync_ra(struct readahead_control *ractl,
554 unsigned long req_count)
555{
556 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
557
558 /*
559 * Even if read-ahead is disabled, issue this request as read-ahead
560 * as we'll need it to satisfy the requested range. The forced
561 * read-ahead will do the right thing and limit the read to just the
562 * requested range, which we'll set to 1 page for this case.
563 */
564 if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
565 if (!ractl->file)
566 return;
567 req_count = 1;
568 do_forced_ra = true;
569 }
570
571 /* be dumb */
572 if (do_forced_ra) {
573 force_page_cache_ra(ractl, req_count);
574 return;
575 }
576
577 /* do read-ahead */
578 ondemand_readahead(ractl, false, req_count);
579}
580EXPORT_SYMBOL_GPL(page_cache_sync_ra);
581
582void page_cache_async_ra(struct readahead_control *ractl,
583 struct page *page, unsigned long req_count)
584{
585 /* no read-ahead */
586 if (!ractl->ra->ra_pages)
587 return;
588
589 /*
590 * Same bit is used for PG_readahead and PG_reclaim.
591 */
592 if (PageWriteback(page))
593 return;
594
595 ClearPageReadahead(page);
596
597 /*
598 * Defer asynchronous read-ahead on IO congestion.
599 */
600 if (inode_read_congested(ractl->mapping->host))
601 return;
602
603 if (blk_cgroup_congested())
604 return;
605
606 /* do read-ahead */
607 ondemand_readahead(ractl, true, req_count);
608}
609EXPORT_SYMBOL_GPL(page_cache_async_ra);
610
611ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
612{
613 ssize_t ret;
614 struct fd f;
615
616 ret = -EBADF;
617 f = fdget(fd);
618 if (!f.file || !(f.file->f_mode & FMODE_READ))
619 goto out;
620
621 /*
622 * The readahead() syscall is intended to run only on files
623 * that can execute readahead. If readahead is not possible
624 * on this file, then we must return -EINVAL.
625 */
626 ret = -EINVAL;
627 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
628 !S_ISREG(file_inode(f.file)->i_mode))
629 goto out;
630
631 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
632out:
633 fdput(f);
634 return ret;
635}
636
637SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
638{
639 return ksys_readahead(fd, offset, count);
640}
641
642/**
643 * readahead_expand - Expand a readahead request
644 * @ractl: The request to be expanded
645 * @new_start: The revised start
646 * @new_len: The revised size of the request
647 *
648 * Attempt to expand a readahead request outwards from the current size to the
649 * specified size by inserting locked pages before and after the current window
650 * to increase the size to the new window. This may involve the insertion of
651 * THPs, in which case the window may get expanded even beyond what was
652 * requested.
653 *
654 * The algorithm will stop if it encounters a conflicting page already in the
655 * pagecache and leave a smaller expansion than requested.
656 *
657 * The caller must check for this by examining the revised @ractl object for a
658 * different expansion than was requested.
659 */
660void readahead_expand(struct readahead_control *ractl,
661 loff_t new_start, size_t new_len)
662{
663 struct address_space *mapping = ractl->mapping;
664 struct file_ra_state *ra = ractl->ra;
665 pgoff_t new_index, new_nr_pages;
666 gfp_t gfp_mask = readahead_gfp_mask(mapping);
667
668 new_index = new_start / PAGE_SIZE;
669
670 /* Expand the leading edge downwards */
671 while (ractl->_index > new_index) {
672 unsigned long index = ractl->_index - 1;
673 struct page *page = xa_load(&mapping->i_pages, index);
674
675 if (page && !xa_is_value(page))
676 return; /* Page apparently present */
677
678 page = __page_cache_alloc(gfp_mask);
679 if (!page)
680 return;
681 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
682 put_page(page);
683 return;
684 }
685
686 ractl->_nr_pages++;
687 ractl->_index = page->index;
688 }
689
690 new_len += new_start - readahead_pos(ractl);
691 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
692
693 /* Expand the trailing edge upwards */
694 while (ractl->_nr_pages < new_nr_pages) {
695 unsigned long index = ractl->_index + ractl->_nr_pages;
696 struct page *page = xa_load(&mapping->i_pages, index);
697
698 if (page && !xa_is_value(page))
699 return; /* Page apparently present */
700
701 page = __page_cache_alloc(gfp_mask);
702 if (!page)
703 return;
704 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
705 put_page(page);
706 return;
707 }
708 ractl->_nr_pages++;
709 if (ra) {
710 ra->size++;
711 ra->async_size++;
712 }
713 }
714}
715EXPORT_SYMBOL(readahead_expand);