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/**
12 * DOC: Readahead Overview
13 *
14 * Readahead is used to read content into the page cache before it is
15 * explicitly requested by the application. Readahead only ever
16 * attempts to read folios that are not yet in the page cache. If a
17 * folio is present but not up-to-date, readahead will not try to read
18 * it. In that case a simple ->read_folio() will be requested.
19 *
20 * Readahead is triggered when an application read request (whether a
21 * system call or a page fault) finds that the requested folio is not in
22 * the page cache, or that it is in the page cache and has the
23 * readahead flag set. This flag indicates that the folio was read
24 * as part of a previous readahead request and now that it has been
25 * accessed, it is time for the next readahead.
26 *
27 * Each readahead request is partly synchronous read, and partly async
28 * readahead. This is reflected in the struct file_ra_state which
29 * contains ->size being the total number of pages, and ->async_size
30 * which is the number of pages in the async section. The readahead
31 * flag will be set on the first folio in this async section to trigger
32 * a subsequent readahead. Once a series of sequential reads has been
33 * established, there should be no need for a synchronous component and
34 * all readahead request will be fully asynchronous.
35 *
36 * When either of the triggers causes a readahead, three numbers need
37 * to be determined: the start of the region to read, the size of the
38 * region, and the size of the async tail.
39 *
40 * The start of the region is simply the first page address at or after
41 * the accessed address, which is not currently populated in the page
42 * cache. This is found with a simple search in the page cache.
43 *
44 * The size of the async tail is determined by subtracting the size that
45 * was explicitly requested from the determined request size, unless
46 * this would be less than zero - then zero is used. NOTE THIS
47 * CALCULATION IS WRONG WHEN THE START OF THE REGION IS NOT THE ACCESSED
48 * PAGE. ALSO THIS CALCULATION IS NOT USED CONSISTENTLY.
49 *
50 * The size of the region is normally determined from the size of the
51 * previous readahead which loaded the preceding pages. This may be
52 * discovered from the struct file_ra_state for simple sequential reads,
53 * or from examining the state of the page cache when multiple
54 * sequential reads are interleaved. Specifically: where the readahead
55 * was triggered by the readahead flag, the size of the previous
56 * readahead is assumed to be the number of pages from the triggering
57 * page to the start of the new readahead. In these cases, the size of
58 * the previous readahead is scaled, often doubled, for the new
59 * readahead, though see get_next_ra_size() for details.
60 *
61 * If the size of the previous read cannot be determined, the number of
62 * preceding pages in the page cache is used to estimate the size of
63 * a previous read. This estimate could easily be misled by random
64 * reads being coincidentally adjacent, so it is ignored unless it is
65 * larger than the current request, and it is not scaled up, unless it
66 * is at the start of file.
67 *
68 * In general readahead is accelerated at the start of the file, as
69 * reads from there are often sequential. There are other minor
70 * adjustments to the readahead size in various special cases and these
71 * are best discovered by reading the code.
72 *
73 * The above calculation, based on the previous readahead size,
74 * determines the size of the readahead, to which any requested read
75 * size may be added.
76 *
77 * Readahead requests are sent to the filesystem using the ->readahead()
78 * address space operation, for which mpage_readahead() is a canonical
79 * implementation. ->readahead() should normally initiate reads on all
80 * folios, but may fail to read any or all folios without causing an I/O
81 * error. The page cache reading code will issue a ->read_folio() request
82 * for any folio which ->readahead() did not read, and only an error
83 * from this will be final.
84 *
85 * ->readahead() will generally call readahead_folio() repeatedly to get
86 * each folio from those prepared for readahead. It may fail to read a
87 * folio by:
88 *
89 * * not calling readahead_folio() sufficiently many times, effectively
90 * ignoring some folios, as might be appropriate if the path to
91 * storage is congested.
92 *
93 * * failing to actually submit a read request for a given folio,
94 * possibly due to insufficient resources, or
95 *
96 * * getting an error during subsequent processing of a request.
97 *
98 * In the last two cases, the folio should be unlocked by the filesystem
99 * to indicate that the read attempt has failed. In the first case the
100 * folio will be unlocked by the VFS.
101 *
102 * Those folios not in the final ``async_size`` of the request should be
103 * considered to be important and ->readahead() should not fail them due
104 * to congestion or temporary resource unavailability, but should wait
105 * for necessary resources (e.g. memory or indexing information) to
106 * become available. Folios in the final ``async_size`` may be
107 * considered less urgent and failure to read them is more acceptable.
108 * In this case it is best to use filemap_remove_folio() to remove the
109 * folios from the page cache as is automatically done for folios that
110 * were not fetched with readahead_folio(). This will allow a
111 * subsequent synchronous readahead request to try them again. If they
112 * are left in the page cache, then they will be read individually using
113 * ->read_folio() which may be less efficient.
114 */
115
116#include <linux/blkdev.h>
117#include <linux/kernel.h>
118#include <linux/dax.h>
119#include <linux/gfp.h>
120#include <linux/export.h>
121#include <linux/backing-dev.h>
122#include <linux/task_io_accounting_ops.h>
123#include <linux/pagemap.h>
124#include <linux/psi.h>
125#include <linux/syscalls.h>
126#include <linux/file.h>
127#include <linux/mm_inline.h>
128#include <linux/blk-cgroup.h>
129#include <linux/fadvise.h>
130#include <linux/sched/mm.h>
131
132#include "internal.h"
133
134/*
135 * Initialise a struct file's readahead state. Assumes that the caller has
136 * memset *ra to zero.
137 */
138void
139file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
140{
141 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
142 ra->prev_pos = -1;
143}
144EXPORT_SYMBOL_GPL(file_ra_state_init);
145
146static void read_pages(struct readahead_control *rac)
147{
148 const struct address_space_operations *aops = rac->mapping->a_ops;
149 struct folio *folio;
150 struct blk_plug plug;
151
152 if (!readahead_count(rac))
153 return;
154
155 if (unlikely(rac->_workingset))
156 psi_memstall_enter(&rac->_pflags);
157 blk_start_plug(&plug);
158
159 if (aops->readahead) {
160 aops->readahead(rac);
161 /*
162 * Clean up the remaining folios. The sizes in ->ra
163 * may be used to size the next readahead, so make sure
164 * they accurately reflect what happened.
165 */
166 while ((folio = readahead_folio(rac)) != NULL) {
167 unsigned long nr = folio_nr_pages(folio);
168
169 folio_get(folio);
170 rac->ra->size -= nr;
171 if (rac->ra->async_size >= nr) {
172 rac->ra->async_size -= nr;
173 filemap_remove_folio(folio);
174 }
175 folio_unlock(folio);
176 folio_put(folio);
177 }
178 } else {
179 while ((folio = readahead_folio(rac)) != NULL)
180 aops->read_folio(rac->file, folio);
181 }
182
183 blk_finish_plug(&plug);
184 if (unlikely(rac->_workingset))
185 psi_memstall_leave(&rac->_pflags);
186 rac->_workingset = false;
187
188 BUG_ON(readahead_count(rac));
189}
190
191/**
192 * page_cache_ra_unbounded - Start unchecked readahead.
193 * @ractl: Readahead control.
194 * @nr_to_read: The number of pages to read.
195 * @lookahead_size: Where to start the next readahead.
196 *
197 * This function is for filesystems to call when they want to start
198 * readahead beyond a file's stated i_size. This is almost certainly
199 * not the function you want to call. Use page_cache_async_readahead()
200 * or page_cache_sync_readahead() instead.
201 *
202 * Context: File is referenced by caller. Mutexes may be held by caller.
203 * May sleep, but will not reenter filesystem to reclaim memory.
204 */
205void page_cache_ra_unbounded(struct readahead_control *ractl,
206 unsigned long nr_to_read, unsigned long lookahead_size)
207{
208 struct address_space *mapping = ractl->mapping;
209 unsigned long index = readahead_index(ractl);
210 gfp_t gfp_mask = readahead_gfp_mask(mapping);
211 unsigned long i;
212
213 /*
214 * Partway through the readahead operation, we will have added
215 * locked pages to the page cache, but will not yet have submitted
216 * them for I/O. Adding another page may need to allocate memory,
217 * which can trigger memory reclaim. Telling the VM we're in
218 * the middle of a filesystem operation will cause it to not
219 * touch file-backed pages, preventing a deadlock. Most (all?)
220 * filesystems already specify __GFP_NOFS in their mapping's
221 * gfp_mask, but let's be explicit here.
222 */
223 unsigned int nofs = memalloc_nofs_save();
224
225 filemap_invalidate_lock_shared(mapping);
226 /*
227 * Preallocate as many pages as we will need.
228 */
229 for (i = 0; i < nr_to_read; i++) {
230 struct folio *folio = xa_load(&mapping->i_pages, index + i);
231
232 if (folio && !xa_is_value(folio)) {
233 /*
234 * Page already present? Kick off the current batch
235 * of contiguous pages before continuing with the
236 * next batch. This page may be the one we would
237 * have intended to mark as Readahead, but we don't
238 * have a stable reference to this page, and it's
239 * not worth getting one just for that.
240 */
241 read_pages(ractl);
242 ractl->_index++;
243 i = ractl->_index + ractl->_nr_pages - index - 1;
244 continue;
245 }
246
247 folio = filemap_alloc_folio(gfp_mask, 0);
248 if (!folio)
249 break;
250 if (filemap_add_folio(mapping, folio, index + i,
251 gfp_mask) < 0) {
252 folio_put(folio);
253 read_pages(ractl);
254 ractl->_index++;
255 i = ractl->_index + ractl->_nr_pages - index - 1;
256 continue;
257 }
258 if (i == nr_to_read - lookahead_size)
259 folio_set_readahead(folio);
260 ractl->_workingset |= folio_test_workingset(folio);
261 ractl->_nr_pages++;
262 }
263
264 /*
265 * Now start the IO. We ignore I/O errors - if the folio is not
266 * uptodate then the caller will launch read_folio again, and
267 * will then handle the error.
268 */
269 read_pages(ractl);
270 filemap_invalidate_unlock_shared(mapping);
271 memalloc_nofs_restore(nofs);
272}
273EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
274
275/*
276 * do_page_cache_ra() actually reads a chunk of disk. It allocates
277 * the pages first, then submits them for I/O. This avoids the very bad
278 * behaviour which would occur if page allocations are causing VM writeback.
279 * We really don't want to intermingle reads and writes like that.
280 */
281static void do_page_cache_ra(struct readahead_control *ractl,
282 unsigned long nr_to_read, unsigned long lookahead_size)
283{
284 struct inode *inode = ractl->mapping->host;
285 unsigned long index = readahead_index(ractl);
286 loff_t isize = i_size_read(inode);
287 pgoff_t end_index; /* The last page we want to read */
288
289 if (isize == 0)
290 return;
291
292 end_index = (isize - 1) >> PAGE_SHIFT;
293 if (index > end_index)
294 return;
295 /* Don't read past the page containing the last byte of the file */
296 if (nr_to_read > end_index - index)
297 nr_to_read = end_index - index + 1;
298
299 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
300}
301
302/*
303 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
304 * memory at once.
305 */
306void force_page_cache_ra(struct readahead_control *ractl,
307 unsigned long nr_to_read)
308{
309 struct address_space *mapping = ractl->mapping;
310 struct file_ra_state *ra = ractl->ra;
311 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
312 unsigned long max_pages, index;
313
314 if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead))
315 return;
316
317 /*
318 * If the request exceeds the readahead window, allow the read to
319 * be up to the optimal hardware IO size
320 */
321 index = readahead_index(ractl);
322 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
323 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
324 while (nr_to_read) {
325 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
326
327 if (this_chunk > nr_to_read)
328 this_chunk = nr_to_read;
329 ractl->_index = index;
330 do_page_cache_ra(ractl, this_chunk, 0);
331
332 index += this_chunk;
333 nr_to_read -= this_chunk;
334 }
335}
336
337/*
338 * Set the initial window size, round to next power of 2 and square
339 * for small size, x 4 for medium, and x 2 for large
340 * for 128k (32 page) max ra
341 * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
342 */
343static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
344{
345 unsigned long newsize = roundup_pow_of_two(size);
346
347 if (newsize <= max / 32)
348 newsize = newsize * 4;
349 else if (newsize <= max / 4)
350 newsize = newsize * 2;
351 else
352 newsize = max;
353
354 return newsize;
355}
356
357/*
358 * Get the previous window size, ramp it up, and
359 * return it as the new window size.
360 */
361static unsigned long get_next_ra_size(struct file_ra_state *ra,
362 unsigned long max)
363{
364 unsigned long cur = ra->size;
365
366 if (cur < max / 16)
367 return 4 * cur;
368 if (cur <= max / 2)
369 return 2 * cur;
370 return max;
371}
372
373/*
374 * On-demand readahead design.
375 *
376 * The fields in struct file_ra_state represent the most-recently-executed
377 * readahead attempt:
378 *
379 * |<----- async_size ---------|
380 * |------------------- size -------------------->|
381 * |==================#===========================|
382 * ^start ^page marked with PG_readahead
383 *
384 * To overlap application thinking time and disk I/O time, we do
385 * `readahead pipelining': Do not wait until the application consumed all
386 * readahead pages and stalled on the missing page at readahead_index;
387 * Instead, submit an asynchronous readahead I/O as soon as there are
388 * only async_size pages left in the readahead window. Normally async_size
389 * will be equal to size, for maximum pipelining.
390 *
391 * In interleaved sequential reads, concurrent streams on the same fd can
392 * be invalidating each other's readahead state. So we flag the new readahead
393 * page at (start+size-async_size) with PG_readahead, and use it as readahead
394 * indicator. The flag won't be set on already cached pages, to avoid the
395 * readahead-for-nothing fuss, saving pointless page cache lookups.
396 *
397 * prev_pos tracks the last visited byte in the _previous_ read request.
398 * It should be maintained by the caller, and will be used for detecting
399 * small random reads. Note that the readahead algorithm checks loosely
400 * for sequential patterns. Hence interleaved reads might be served as
401 * sequential ones.
402 *
403 * There is a special-case: if the first page which the application tries to
404 * read happens to be the first page of the file, it is assumed that a linear
405 * read is about to happen and the window is immediately set to the initial size
406 * based on I/O request size and the max_readahead.
407 *
408 * The code ramps up the readahead size aggressively at first, but slow down as
409 * it approaches max_readhead.
410 */
411
412/*
413 * Count contiguously cached pages from @index-1 to @index-@max,
414 * this count is a conservative estimation of
415 * - length of the sequential read sequence, or
416 * - thrashing threshold in memory tight systems
417 */
418static pgoff_t count_history_pages(struct address_space *mapping,
419 pgoff_t index, unsigned long max)
420{
421 pgoff_t head;
422
423 rcu_read_lock();
424 head = page_cache_prev_miss(mapping, index - 1, max);
425 rcu_read_unlock();
426
427 return index - 1 - head;
428}
429
430/*
431 * page cache context based readahead
432 */
433static int try_context_readahead(struct address_space *mapping,
434 struct file_ra_state *ra,
435 pgoff_t index,
436 unsigned long req_size,
437 unsigned long max)
438{
439 pgoff_t size;
440
441 size = count_history_pages(mapping, index, max);
442
443 /*
444 * not enough history pages:
445 * it could be a random read
446 */
447 if (size <= req_size)
448 return 0;
449
450 /*
451 * starts from beginning of file:
452 * it is a strong indication of long-run stream (or whole-file-read)
453 */
454 if (size >= index)
455 size *= 2;
456
457 ra->start = index;
458 ra->size = min(size + req_size, max);
459 ra->async_size = 1;
460
461 return 1;
462}
463
464static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
465 pgoff_t mark, unsigned int order, gfp_t gfp)
466{
467 int err;
468 struct folio *folio = filemap_alloc_folio(gfp, order);
469
470 if (!folio)
471 return -ENOMEM;
472 mark = round_down(mark, 1UL << order);
473 if (index == mark)
474 folio_set_readahead(folio);
475 err = filemap_add_folio(ractl->mapping, folio, index, gfp);
476 if (err) {
477 folio_put(folio);
478 return err;
479 }
480
481 ractl->_nr_pages += 1UL << order;
482 ractl->_workingset |= folio_test_workingset(folio);
483 return 0;
484}
485
486void page_cache_ra_order(struct readahead_control *ractl,
487 struct file_ra_state *ra, unsigned int new_order)
488{
489 struct address_space *mapping = ractl->mapping;
490 pgoff_t index = readahead_index(ractl);
491 pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
492 pgoff_t mark = index + ra->size - ra->async_size;
493 unsigned int nofs;
494 int err = 0;
495 gfp_t gfp = readahead_gfp_mask(mapping);
496
497 if (!mapping_large_folio_support(mapping) || ra->size < 4)
498 goto fallback;
499
500 limit = min(limit, index + ra->size - 1);
501
502 if (new_order < MAX_PAGECACHE_ORDER) {
503 new_order += 2;
504 new_order = min_t(unsigned int, MAX_PAGECACHE_ORDER, new_order);
505 new_order = min_t(unsigned int, new_order, ilog2(ra->size));
506 }
507
508 /* See comment in page_cache_ra_unbounded() */
509 nofs = memalloc_nofs_save();
510 filemap_invalidate_lock_shared(mapping);
511 while (index <= limit) {
512 unsigned int order = new_order;
513
514 /* Align with smaller pages if needed */
515 if (index & ((1UL << order) - 1))
516 order = __ffs(index);
517 /* Don't allocate pages past EOF */
518 while (index + (1UL << order) - 1 > limit)
519 order--;
520 err = ra_alloc_folio(ractl, index, mark, order, gfp);
521 if (err)
522 break;
523 index += 1UL << order;
524 }
525
526 if (index > limit) {
527 ra->size += index - limit - 1;
528 ra->async_size += index - limit - 1;
529 }
530
531 read_pages(ractl);
532 filemap_invalidate_unlock_shared(mapping);
533 memalloc_nofs_restore(nofs);
534
535 /*
536 * If there were already pages in the page cache, then we may have
537 * left some gaps. Let the regular readahead code take care of this
538 * situation.
539 */
540 if (!err)
541 return;
542fallback:
543 do_page_cache_ra(ractl, ra->size, ra->async_size);
544}
545
546/*
547 * A minimal readahead algorithm for trivial sequential/random reads.
548 */
549static void ondemand_readahead(struct readahead_control *ractl,
550 struct folio *folio, unsigned long req_size)
551{
552 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
553 struct file_ra_state *ra = ractl->ra;
554 unsigned long max_pages = ra->ra_pages;
555 unsigned long add_pages;
556 pgoff_t index = readahead_index(ractl);
557 pgoff_t expected, prev_index;
558 unsigned int order = folio ? folio_order(folio) : 0;
559
560 /*
561 * If the request exceeds the readahead window, allow the read to
562 * be up to the optimal hardware IO size
563 */
564 if (req_size > max_pages && bdi->io_pages > max_pages)
565 max_pages = min(req_size, bdi->io_pages);
566
567 /*
568 * start of file
569 */
570 if (!index)
571 goto initial_readahead;
572
573 /*
574 * It's the expected callback index, assume sequential access.
575 * Ramp up sizes, and push forward the readahead window.
576 */
577 expected = round_down(ra->start + ra->size - ra->async_size,
578 1UL << order);
579 if (index == expected || index == (ra->start + ra->size)) {
580 ra->start += ra->size;
581 ra->size = get_next_ra_size(ra, max_pages);
582 ra->async_size = ra->size;
583 goto readit;
584 }
585
586 /*
587 * Hit a marked folio without valid readahead state.
588 * E.g. interleaved reads.
589 * Query the pagecache for async_size, which normally equals to
590 * readahead size. Ramp it up and use it as the new readahead size.
591 */
592 if (folio) {
593 pgoff_t start;
594
595 rcu_read_lock();
596 start = page_cache_next_miss(ractl->mapping, index + 1,
597 max_pages);
598 rcu_read_unlock();
599
600 if (!start || start - index > max_pages)
601 return;
602
603 ra->start = start;
604 ra->size = start - index; /* old async_size */
605 ra->size += req_size;
606 ra->size = get_next_ra_size(ra, max_pages);
607 ra->async_size = ra->size;
608 goto readit;
609 }
610
611 /*
612 * oversize read
613 */
614 if (req_size > max_pages)
615 goto initial_readahead;
616
617 /*
618 * sequential cache miss
619 * trivial case: (index - prev_index) == 1
620 * unaligned reads: (index - prev_index) == 0
621 */
622 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
623 if (index - prev_index <= 1UL)
624 goto initial_readahead;
625
626 /*
627 * Query the page cache and look for the traces(cached history pages)
628 * that a sequential stream would leave behind.
629 */
630 if (try_context_readahead(ractl->mapping, ra, index, req_size,
631 max_pages))
632 goto readit;
633
634 /*
635 * standalone, small random read
636 * Read as is, and do not pollute the readahead state.
637 */
638 do_page_cache_ra(ractl, req_size, 0);
639 return;
640
641initial_readahead:
642 ra->start = index;
643 ra->size = get_init_ra_size(req_size, max_pages);
644 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
645
646readit:
647 /*
648 * Will this read hit the readahead marker made by itself?
649 * If so, trigger the readahead marker hit now, and merge
650 * the resulted next readahead window into the current one.
651 * Take care of maximum IO pages as above.
652 */
653 if (index == ra->start && ra->size == ra->async_size) {
654 add_pages = get_next_ra_size(ra, max_pages);
655 if (ra->size + add_pages <= max_pages) {
656 ra->async_size = add_pages;
657 ra->size += add_pages;
658 } else {
659 ra->size = max_pages;
660 ra->async_size = max_pages >> 1;
661 }
662 }
663
664 ractl->_index = ra->start;
665 page_cache_ra_order(ractl, ra, order);
666}
667
668void page_cache_sync_ra(struct readahead_control *ractl,
669 unsigned long req_count)
670{
671 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
672
673 /*
674 * Even if readahead is disabled, issue this request as readahead
675 * as we'll need it to satisfy the requested range. The forced
676 * readahead will do the right thing and limit the read to just the
677 * requested range, which we'll set to 1 page for this case.
678 */
679 if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
680 if (!ractl->file)
681 return;
682 req_count = 1;
683 do_forced_ra = true;
684 }
685
686 /* be dumb */
687 if (do_forced_ra) {
688 force_page_cache_ra(ractl, req_count);
689 return;
690 }
691
692 ondemand_readahead(ractl, NULL, req_count);
693}
694EXPORT_SYMBOL_GPL(page_cache_sync_ra);
695
696void page_cache_async_ra(struct readahead_control *ractl,
697 struct folio *folio, unsigned long req_count)
698{
699 /* no readahead */
700 if (!ractl->ra->ra_pages)
701 return;
702
703 /*
704 * Same bit is used for PG_readahead and PG_reclaim.
705 */
706 if (folio_test_writeback(folio))
707 return;
708
709 folio_clear_readahead(folio);
710
711 if (blk_cgroup_congested())
712 return;
713
714 ondemand_readahead(ractl, folio, req_count);
715}
716EXPORT_SYMBOL_GPL(page_cache_async_ra);
717
718ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
719{
720 ssize_t ret;
721 struct fd f;
722
723 ret = -EBADF;
724 f = fdget(fd);
725 if (!f.file || !(f.file->f_mode & FMODE_READ))
726 goto out;
727
728 /*
729 * The readahead() syscall is intended to run only on files
730 * that can execute readahead. If readahead is not possible
731 * on this file, then we must return -EINVAL.
732 */
733 ret = -EINVAL;
734 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
735 (!S_ISREG(file_inode(f.file)->i_mode) &&
736 !S_ISBLK(file_inode(f.file)->i_mode)))
737 goto out;
738
739 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
740out:
741 fdput(f);
742 return ret;
743}
744
745SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
746{
747 return ksys_readahead(fd, offset, count);
748}
749
750#if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD)
751COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count)
752{
753 return ksys_readahead(fd, compat_arg_u64_glue(offset), count);
754}
755#endif
756
757/**
758 * readahead_expand - Expand a readahead request
759 * @ractl: The request to be expanded
760 * @new_start: The revised start
761 * @new_len: The revised size of the request
762 *
763 * Attempt to expand a readahead request outwards from the current size to the
764 * specified size by inserting locked pages before and after the current window
765 * to increase the size to the new window. This may involve the insertion of
766 * THPs, in which case the window may get expanded even beyond what was
767 * requested.
768 *
769 * The algorithm will stop if it encounters a conflicting page already in the
770 * pagecache and leave a smaller expansion than requested.
771 *
772 * The caller must check for this by examining the revised @ractl object for a
773 * different expansion than was requested.
774 */
775void readahead_expand(struct readahead_control *ractl,
776 loff_t new_start, size_t new_len)
777{
778 struct address_space *mapping = ractl->mapping;
779 struct file_ra_state *ra = ractl->ra;
780 pgoff_t new_index, new_nr_pages;
781 gfp_t gfp_mask = readahead_gfp_mask(mapping);
782
783 new_index = new_start / PAGE_SIZE;
784
785 /* Expand the leading edge downwards */
786 while (ractl->_index > new_index) {
787 unsigned long index = ractl->_index - 1;
788 struct folio *folio = xa_load(&mapping->i_pages, index);
789
790 if (folio && !xa_is_value(folio))
791 return; /* Folio apparently present */
792
793 folio = filemap_alloc_folio(gfp_mask, 0);
794 if (!folio)
795 return;
796 if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
797 folio_put(folio);
798 return;
799 }
800 if (unlikely(folio_test_workingset(folio)) &&
801 !ractl->_workingset) {
802 ractl->_workingset = true;
803 psi_memstall_enter(&ractl->_pflags);
804 }
805 ractl->_nr_pages++;
806 ractl->_index = folio->index;
807 }
808
809 new_len += new_start - readahead_pos(ractl);
810 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
811
812 /* Expand the trailing edge upwards */
813 while (ractl->_nr_pages < new_nr_pages) {
814 unsigned long index = ractl->_index + ractl->_nr_pages;
815 struct folio *folio = xa_load(&mapping->i_pages, index);
816
817 if (folio && !xa_is_value(folio))
818 return; /* Folio apparently present */
819
820 folio = filemap_alloc_folio(gfp_mask, 0);
821 if (!folio)
822 return;
823 if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
824 folio_put(folio);
825 return;
826 }
827 if (unlikely(folio_test_workingset(folio)) &&
828 !ractl->_workingset) {
829 ractl->_workingset = true;
830 psi_memstall_enter(&ractl->_pflags);
831 }
832 ractl->_nr_pages++;
833 if (ra) {
834 ra->size++;
835 ra->async_size++;
836 }
837 }
838}
839EXPORT_SYMBOL(readahead_expand);