Loading...
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
23#include "internal.h"
24
25/*
26 * Initialise a struct file's readahead state. Assumes that the caller has
27 * memset *ra to zero.
28 */
29void
30file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
31{
32 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
33 ra->prev_pos = -1;
34}
35EXPORT_SYMBOL_GPL(file_ra_state_init);
36
37/*
38 * see if a page needs releasing upon read_cache_pages() failure
39 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
40 * before calling, such as the NFS fs marking pages that are cached locally
41 * on disk, thus we need to give the fs a chance to clean up in the event of
42 * an error
43 */
44static void read_cache_pages_invalidate_page(struct address_space *mapping,
45 struct page *page)
46{
47 if (page_has_private(page)) {
48 if (!trylock_page(page))
49 BUG();
50 page->mapping = mapping;
51 do_invalidatepage(page, 0, PAGE_SIZE);
52 page->mapping = NULL;
53 unlock_page(page);
54 }
55 put_page(page);
56}
57
58/*
59 * release a list of pages, invalidating them first if need be
60 */
61static void read_cache_pages_invalidate_pages(struct address_space *mapping,
62 struct list_head *pages)
63{
64 struct page *victim;
65
66 while (!list_empty(pages)) {
67 victim = lru_to_page(pages);
68 list_del(&victim->lru);
69 read_cache_pages_invalidate_page(mapping, victim);
70 }
71}
72
73/**
74 * read_cache_pages - populate an address space with some pages & start reads against them
75 * @mapping: the address_space
76 * @pages: The address of a list_head which contains the target pages. These
77 * pages have their ->index populated and are otherwise uninitialised.
78 * @filler: callback routine for filling a single page.
79 * @data: private data for the callback routine.
80 *
81 * Hides the details of the LRU cache etc from the filesystems.
82 */
83int read_cache_pages(struct address_space *mapping, struct list_head *pages,
84 int (*filler)(void *, struct page *), void *data)
85{
86 struct page *page;
87 int ret = 0;
88
89 while (!list_empty(pages)) {
90 page = lru_to_page(pages);
91 list_del(&page->lru);
92 if (add_to_page_cache_lru(page, mapping, page->index,
93 readahead_gfp_mask(mapping))) {
94 read_cache_pages_invalidate_page(mapping, page);
95 continue;
96 }
97 put_page(page);
98
99 ret = filler(data, page);
100 if (unlikely(ret)) {
101 read_cache_pages_invalidate_pages(mapping, pages);
102 break;
103 }
104 task_io_account_read(PAGE_SIZE);
105 }
106 return ret;
107}
108
109EXPORT_SYMBOL(read_cache_pages);
110
111static int read_pages(struct address_space *mapping, struct file *filp,
112 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
113{
114 struct blk_plug plug;
115 unsigned page_idx;
116 int ret;
117
118 blk_start_plug(&plug);
119
120 if (mapping->a_ops->readpages) {
121 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
122 /* Clean up the remaining pages */
123 put_pages_list(pages);
124 goto out;
125 }
126
127 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
128 struct page *page = lru_to_page(pages);
129 list_del(&page->lru);
130 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
131 mapping->a_ops->readpage(filp, page);
132 put_page(page);
133 }
134 ret = 0;
135
136out:
137 blk_finish_plug(&plug);
138
139 return ret;
140}
141
142/*
143 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
144 * the pages first, then submits them all for I/O. This avoids the very bad
145 * behaviour which would occur if page allocations are causing VM writeback.
146 * We really don't want to intermingle reads and writes like that.
147 *
148 * Returns the number of pages requested, or the maximum amount of I/O allowed.
149 */
150int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
151 pgoff_t offset, unsigned long nr_to_read,
152 unsigned long lookahead_size)
153{
154 struct inode *inode = mapping->host;
155 struct page *page;
156 unsigned long end_index; /* The last page we want to read */
157 LIST_HEAD(page_pool);
158 int page_idx;
159 int ret = 0;
160 loff_t isize = i_size_read(inode);
161 gfp_t gfp_mask = readahead_gfp_mask(mapping);
162
163 if (isize == 0)
164 goto out;
165
166 end_index = ((isize - 1) >> PAGE_SHIFT);
167
168 /*
169 * Preallocate as many pages as we will need.
170 */
171 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
172 pgoff_t page_offset = offset + page_idx;
173
174 if (page_offset > end_index)
175 break;
176
177 rcu_read_lock();
178 page = radix_tree_lookup(&mapping->i_pages, page_offset);
179 rcu_read_unlock();
180 if (page && !radix_tree_exceptional_entry(page))
181 continue;
182
183 page = __page_cache_alloc(gfp_mask);
184 if (!page)
185 break;
186 page->index = page_offset;
187 list_add(&page->lru, &page_pool);
188 if (page_idx == nr_to_read - lookahead_size)
189 SetPageReadahead(page);
190 ret++;
191 }
192
193 /*
194 * Now start the IO. We ignore I/O errors - if the page is not
195 * uptodate then the caller will launch readpage again, and
196 * will then handle the error.
197 */
198 if (ret)
199 read_pages(mapping, filp, &page_pool, ret, gfp_mask);
200 BUG_ON(!list_empty(&page_pool));
201out:
202 return ret;
203}
204
205/*
206 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
207 * memory at once.
208 */
209int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
210 pgoff_t offset, unsigned long nr_to_read)
211{
212 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
213 struct file_ra_state *ra = &filp->f_ra;
214 unsigned long max_pages;
215
216 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
217 return -EINVAL;
218
219 /*
220 * If the request exceeds the readahead window, allow the read to
221 * be up to the optimal hardware IO size
222 */
223 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
224 nr_to_read = min(nr_to_read, max_pages);
225 while (nr_to_read) {
226 int err;
227
228 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
229
230 if (this_chunk > nr_to_read)
231 this_chunk = nr_to_read;
232 err = __do_page_cache_readahead(mapping, filp,
233 offset, this_chunk, 0);
234 if (err < 0)
235 return err;
236
237 offset += this_chunk;
238 nr_to_read -= this_chunk;
239 }
240 return 0;
241}
242
243/*
244 * Set the initial window size, round to next power of 2 and square
245 * for small size, x 4 for medium, and x 2 for large
246 * for 128k (32 page) max ra
247 * 1-8 page = 32k initial, > 8 page = 128k initial
248 */
249static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
250{
251 unsigned long newsize = roundup_pow_of_two(size);
252
253 if (newsize <= max / 32)
254 newsize = newsize * 4;
255 else if (newsize <= max / 4)
256 newsize = newsize * 2;
257 else
258 newsize = max;
259
260 return newsize;
261}
262
263/*
264 * Get the previous window size, ramp it up, and
265 * return it as the new window size.
266 */
267static unsigned long get_next_ra_size(struct file_ra_state *ra,
268 unsigned long max)
269{
270 unsigned long cur = ra->size;
271 unsigned long newsize;
272
273 if (cur < max / 16)
274 newsize = 4 * cur;
275 else
276 newsize = 2 * cur;
277
278 return min(newsize, max);
279}
280
281/*
282 * On-demand readahead design.
283 *
284 * The fields in struct file_ra_state represent the most-recently-executed
285 * readahead attempt:
286 *
287 * |<----- async_size ---------|
288 * |------------------- size -------------------->|
289 * |==================#===========================|
290 * ^start ^page marked with PG_readahead
291 *
292 * To overlap application thinking time and disk I/O time, we do
293 * `readahead pipelining': Do not wait until the application consumed all
294 * readahead pages and stalled on the missing page at readahead_index;
295 * Instead, submit an asynchronous readahead I/O as soon as there are
296 * only async_size pages left in the readahead window. Normally async_size
297 * will be equal to size, for maximum pipelining.
298 *
299 * In interleaved sequential reads, concurrent streams on the same fd can
300 * be invalidating each other's readahead state. So we flag the new readahead
301 * page at (start+size-async_size) with PG_readahead, and use it as readahead
302 * indicator. The flag won't be set on already cached pages, to avoid the
303 * readahead-for-nothing fuss, saving pointless page cache lookups.
304 *
305 * prev_pos tracks the last visited byte in the _previous_ read request.
306 * It should be maintained by the caller, and will be used for detecting
307 * small random reads. Note that the readahead algorithm checks loosely
308 * for sequential patterns. Hence interleaved reads might be served as
309 * sequential ones.
310 *
311 * There is a special-case: if the first page which the application tries to
312 * read happens to be the first page of the file, it is assumed that a linear
313 * read is about to happen and the window is immediately set to the initial size
314 * based on I/O request size and the max_readahead.
315 *
316 * The code ramps up the readahead size aggressively at first, but slow down as
317 * it approaches max_readhead.
318 */
319
320/*
321 * Count contiguously cached pages from @offset-1 to @offset-@max,
322 * this count is a conservative estimation of
323 * - length of the sequential read sequence, or
324 * - thrashing threshold in memory tight systems
325 */
326static pgoff_t count_history_pages(struct address_space *mapping,
327 pgoff_t offset, unsigned long max)
328{
329 pgoff_t head;
330
331 rcu_read_lock();
332 head = page_cache_prev_hole(mapping, offset - 1, max);
333 rcu_read_unlock();
334
335 return offset - 1 - head;
336}
337
338/*
339 * page cache context based read-ahead
340 */
341static int try_context_readahead(struct address_space *mapping,
342 struct file_ra_state *ra,
343 pgoff_t offset,
344 unsigned long req_size,
345 unsigned long max)
346{
347 pgoff_t size;
348
349 size = count_history_pages(mapping, offset, max);
350
351 /*
352 * not enough history pages:
353 * it could be a random read
354 */
355 if (size <= req_size)
356 return 0;
357
358 /*
359 * starts from beginning of file:
360 * it is a strong indication of long-run stream (or whole-file-read)
361 */
362 if (size >= offset)
363 size *= 2;
364
365 ra->start = offset;
366 ra->size = min(size + req_size, max);
367 ra->async_size = 1;
368
369 return 1;
370}
371
372/*
373 * A minimal readahead algorithm for trivial sequential/random reads.
374 */
375static unsigned long
376ondemand_readahead(struct address_space *mapping,
377 struct file_ra_state *ra, struct file *filp,
378 bool hit_readahead_marker, pgoff_t offset,
379 unsigned long req_size)
380{
381 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
382 unsigned long max_pages = ra->ra_pages;
383 pgoff_t prev_offset;
384
385 /*
386 * If the request exceeds the readahead window, allow the read to
387 * be up to the optimal hardware IO size
388 */
389 if (req_size > max_pages && bdi->io_pages > max_pages)
390 max_pages = min(req_size, bdi->io_pages);
391
392 /*
393 * start of file
394 */
395 if (!offset)
396 goto initial_readahead;
397
398 /*
399 * It's the expected callback offset, assume sequential access.
400 * Ramp up sizes, and push forward the readahead window.
401 */
402 if ((offset == (ra->start + ra->size - ra->async_size) ||
403 offset == (ra->start + ra->size))) {
404 ra->start += ra->size;
405 ra->size = get_next_ra_size(ra, max_pages);
406 ra->async_size = ra->size;
407 goto readit;
408 }
409
410 /*
411 * Hit a marked page without valid readahead state.
412 * E.g. interleaved reads.
413 * Query the pagecache for async_size, which normally equals to
414 * readahead size. Ramp it up and use it as the new readahead size.
415 */
416 if (hit_readahead_marker) {
417 pgoff_t start;
418
419 rcu_read_lock();
420 start = page_cache_next_hole(mapping, offset + 1, max_pages);
421 rcu_read_unlock();
422
423 if (!start || start - offset > max_pages)
424 return 0;
425
426 ra->start = start;
427 ra->size = start - offset; /* old async_size */
428 ra->size += req_size;
429 ra->size = get_next_ra_size(ra, max_pages);
430 ra->async_size = ra->size;
431 goto readit;
432 }
433
434 /*
435 * oversize read
436 */
437 if (req_size > max_pages)
438 goto initial_readahead;
439
440 /*
441 * sequential cache miss
442 * trivial case: (offset - prev_offset) == 1
443 * unaligned reads: (offset - prev_offset) == 0
444 */
445 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
446 if (offset - prev_offset <= 1UL)
447 goto initial_readahead;
448
449 /*
450 * Query the page cache and look for the traces(cached history pages)
451 * that a sequential stream would leave behind.
452 */
453 if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
454 goto readit;
455
456 /*
457 * standalone, small random read
458 * Read as is, and do not pollute the readahead state.
459 */
460 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
461
462initial_readahead:
463 ra->start = offset;
464 ra->size = get_init_ra_size(req_size, max_pages);
465 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
466
467readit:
468 /*
469 * Will this read hit the readahead marker made by itself?
470 * If so, trigger the readahead marker hit now, and merge
471 * the resulted next readahead window into the current one.
472 */
473 if (offset == ra->start && ra->size == ra->async_size) {
474 ra->async_size = get_next_ra_size(ra, max_pages);
475 ra->size += ra->async_size;
476 }
477
478 return ra_submit(ra, mapping, filp);
479}
480
481/**
482 * page_cache_sync_readahead - generic file readahead
483 * @mapping: address_space which holds the pagecache and I/O vectors
484 * @ra: file_ra_state which holds the readahead state
485 * @filp: passed on to ->readpage() and ->readpages()
486 * @offset: start offset into @mapping, in pagecache page-sized units
487 * @req_size: hint: total size of the read which the caller is performing in
488 * pagecache pages
489 *
490 * page_cache_sync_readahead() should be called when a cache miss happened:
491 * it will submit the read. The readahead logic may decide to piggyback more
492 * pages onto the read request if access patterns suggest it will improve
493 * performance.
494 */
495void page_cache_sync_readahead(struct address_space *mapping,
496 struct file_ra_state *ra, struct file *filp,
497 pgoff_t offset, unsigned long req_size)
498{
499 /* no read-ahead */
500 if (!ra->ra_pages)
501 return;
502
503 /* be dumb */
504 if (filp && (filp->f_mode & FMODE_RANDOM)) {
505 force_page_cache_readahead(mapping, filp, offset, req_size);
506 return;
507 }
508
509 /* do read-ahead */
510 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
511}
512EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
513
514/**
515 * page_cache_async_readahead - file readahead for marked pages
516 * @mapping: address_space which holds the pagecache and I/O vectors
517 * @ra: file_ra_state which holds the readahead state
518 * @filp: passed on to ->readpage() and ->readpages()
519 * @page: the page at @offset which has the PG_readahead flag set
520 * @offset: start offset into @mapping, in pagecache page-sized units
521 * @req_size: hint: total size of the read which the caller is performing in
522 * pagecache pages
523 *
524 * page_cache_async_readahead() should be called when a page is used which
525 * has the PG_readahead flag; this is a marker to suggest that the application
526 * has used up enough of the readahead window that we should start pulling in
527 * more pages.
528 */
529void
530page_cache_async_readahead(struct address_space *mapping,
531 struct file_ra_state *ra, struct file *filp,
532 struct page *page, pgoff_t offset,
533 unsigned long req_size)
534{
535 /* no read-ahead */
536 if (!ra->ra_pages)
537 return;
538
539 /*
540 * Same bit is used for PG_readahead and PG_reclaim.
541 */
542 if (PageWriteback(page))
543 return;
544
545 ClearPageReadahead(page);
546
547 /*
548 * Defer asynchronous read-ahead on IO congestion.
549 */
550 if (inode_read_congested(mapping->host))
551 return;
552
553 /* do read-ahead */
554 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
555}
556EXPORT_SYMBOL_GPL(page_cache_async_readahead);
557
558static ssize_t
559do_readahead(struct address_space *mapping, struct file *filp,
560 pgoff_t index, unsigned long nr)
561{
562 if (!mapping || !mapping->a_ops)
563 return -EINVAL;
564
565 /*
566 * Readahead doesn't make sense for DAX inodes, but we don't want it
567 * to report a failure either. Instead, we just return success and
568 * don't do any work.
569 */
570 if (dax_mapping(mapping))
571 return 0;
572
573 return force_page_cache_readahead(mapping, filp, index, nr);
574}
575
576ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
577{
578 ssize_t ret;
579 struct fd f;
580
581 ret = -EBADF;
582 f = fdget(fd);
583 if (f.file) {
584 if (f.file->f_mode & FMODE_READ) {
585 struct address_space *mapping = f.file->f_mapping;
586 pgoff_t start = offset >> PAGE_SHIFT;
587 pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
588 unsigned long len = end - start + 1;
589 ret = do_readahead(mapping, f.file, start, len);
590 }
591 fdput(f);
592 }
593 return ret;
594}
595
596SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
597{
598 return ksys_readahead(fd, offset, count);
599}
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 mark = ULONG_MAX, i = 0;
212 unsigned int min_nrpages = mapping_min_folio_nrpages(mapping);
213
214 /*
215 * Partway through the readahead operation, we will have added
216 * locked pages to the page cache, but will not yet have submitted
217 * them for I/O. Adding another page may need to allocate memory,
218 * which can trigger memory reclaim. Telling the VM we're in
219 * the middle of a filesystem operation will cause it to not
220 * touch file-backed pages, preventing a deadlock. Most (all?)
221 * filesystems already specify __GFP_NOFS in their mapping's
222 * gfp_mask, but let's be explicit here.
223 */
224 unsigned int nofs = memalloc_nofs_save();
225
226 filemap_invalidate_lock_shared(mapping);
227 index = mapping_align_index(mapping, index);
228
229 /*
230 * As iterator `i` is aligned to min_nrpages, round_up the
231 * difference between nr_to_read and lookahead_size to mark the
232 * index that only has lookahead or "async_region" to set the
233 * readahead flag.
234 */
235 if (lookahead_size <= nr_to_read) {
236 unsigned long ra_folio_index;
237
238 ra_folio_index = round_up(readahead_index(ractl) +
239 nr_to_read - lookahead_size,
240 min_nrpages);
241 mark = ra_folio_index - index;
242 }
243 nr_to_read += readahead_index(ractl) - index;
244 ractl->_index = index;
245
246 /*
247 * Preallocate as many pages as we will need.
248 */
249 while (i < nr_to_read) {
250 struct folio *folio = xa_load(&mapping->i_pages, index + i);
251 int ret;
252
253 if (folio && !xa_is_value(folio)) {
254 /*
255 * Page already present? Kick off the current batch
256 * of contiguous pages before continuing with the
257 * next batch. This page may be the one we would
258 * have intended to mark as Readahead, but we don't
259 * have a stable reference to this page, and it's
260 * not worth getting one just for that.
261 */
262 read_pages(ractl);
263 ractl->_index += min_nrpages;
264 i = ractl->_index + ractl->_nr_pages - index;
265 continue;
266 }
267
268 folio = filemap_alloc_folio(gfp_mask,
269 mapping_min_folio_order(mapping));
270 if (!folio)
271 break;
272
273 ret = filemap_add_folio(mapping, folio, index + i, gfp_mask);
274 if (ret < 0) {
275 folio_put(folio);
276 if (ret == -ENOMEM)
277 break;
278 read_pages(ractl);
279 ractl->_index += min_nrpages;
280 i = ractl->_index + ractl->_nr_pages - index;
281 continue;
282 }
283 if (i == mark)
284 folio_set_readahead(folio);
285 ractl->_workingset |= folio_test_workingset(folio);
286 ractl->_nr_pages += min_nrpages;
287 i += min_nrpages;
288 }
289
290 /*
291 * Now start the IO. We ignore I/O errors - if the folio is not
292 * uptodate then the caller will launch read_folio again, and
293 * will then handle the error.
294 */
295 read_pages(ractl);
296 filemap_invalidate_unlock_shared(mapping);
297 memalloc_nofs_restore(nofs);
298}
299EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
300
301/*
302 * do_page_cache_ra() actually reads a chunk of disk. It allocates
303 * the pages first, then submits them for I/O. This avoids the very bad
304 * behaviour which would occur if page allocations are causing VM writeback.
305 * We really don't want to intermingle reads and writes like that.
306 */
307static void do_page_cache_ra(struct readahead_control *ractl,
308 unsigned long nr_to_read, unsigned long lookahead_size)
309{
310 struct inode *inode = ractl->mapping->host;
311 unsigned long index = readahead_index(ractl);
312 loff_t isize = i_size_read(inode);
313 pgoff_t end_index; /* The last page we want to read */
314
315 if (isize == 0)
316 return;
317
318 end_index = (isize - 1) >> PAGE_SHIFT;
319 if (index > end_index)
320 return;
321 /* Don't read past the page containing the last byte of the file */
322 if (nr_to_read > end_index - index)
323 nr_to_read = end_index - index + 1;
324
325 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
326}
327
328/*
329 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
330 * memory at once.
331 */
332void force_page_cache_ra(struct readahead_control *ractl,
333 unsigned long nr_to_read)
334{
335 struct address_space *mapping = ractl->mapping;
336 struct file_ra_state *ra = ractl->ra;
337 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
338 unsigned long max_pages;
339
340 if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead))
341 return;
342
343 /*
344 * If the request exceeds the readahead window, allow the read to
345 * be up to the optimal hardware IO size
346 */
347 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
348 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
349 while (nr_to_read) {
350 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
351
352 if (this_chunk > nr_to_read)
353 this_chunk = nr_to_read;
354 do_page_cache_ra(ractl, this_chunk, 0);
355
356 nr_to_read -= this_chunk;
357 }
358}
359
360/*
361 * Set the initial window size, round to next power of 2 and square
362 * for small size, x 4 for medium, and x 2 for large
363 * for 128k (32 page) max ra
364 * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
365 */
366static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
367{
368 unsigned long newsize = roundup_pow_of_two(size);
369
370 if (newsize <= max / 32)
371 newsize = newsize * 4;
372 else if (newsize <= max / 4)
373 newsize = newsize * 2;
374 else
375 newsize = max;
376
377 return newsize;
378}
379
380/*
381 * Get the previous window size, ramp it up, and
382 * return it as the new window size.
383 */
384static unsigned long get_next_ra_size(struct file_ra_state *ra,
385 unsigned long max)
386{
387 unsigned long cur = ra->size;
388
389 if (cur < max / 16)
390 return 4 * cur;
391 if (cur <= max / 2)
392 return 2 * cur;
393 return max;
394}
395
396/*
397 * On-demand readahead design.
398 *
399 * The fields in struct file_ra_state represent the most-recently-executed
400 * readahead attempt:
401 *
402 * |<----- async_size ---------|
403 * |------------------- size -------------------->|
404 * |==================#===========================|
405 * ^start ^page marked with PG_readahead
406 *
407 * To overlap application thinking time and disk I/O time, we do
408 * `readahead pipelining': Do not wait until the application consumed all
409 * readahead pages and stalled on the missing page at readahead_index;
410 * Instead, submit an asynchronous readahead I/O as soon as there are
411 * only async_size pages left in the readahead window. Normally async_size
412 * will be equal to size, for maximum pipelining.
413 *
414 * In interleaved sequential reads, concurrent streams on the same fd can
415 * be invalidating each other's readahead state. So we flag the new readahead
416 * page at (start+size-async_size) with PG_readahead, and use it as readahead
417 * indicator. The flag won't be set on already cached pages, to avoid the
418 * readahead-for-nothing fuss, saving pointless page cache lookups.
419 *
420 * prev_pos tracks the last visited byte in the _previous_ read request.
421 * It should be maintained by the caller, and will be used for detecting
422 * small random reads. Note that the readahead algorithm checks loosely
423 * for sequential patterns. Hence interleaved reads might be served as
424 * sequential ones.
425 *
426 * There is a special-case: if the first page which the application tries to
427 * read happens to be the first page of the file, it is assumed that a linear
428 * read is about to happen and the window is immediately set to the initial size
429 * based on I/O request size and the max_readahead.
430 *
431 * The code ramps up the readahead size aggressively at first, but slow down as
432 * it approaches max_readhead.
433 */
434
435static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
436 pgoff_t mark, unsigned int order, gfp_t gfp)
437{
438 int err;
439 struct folio *folio = filemap_alloc_folio(gfp, order);
440
441 if (!folio)
442 return -ENOMEM;
443 mark = round_down(mark, 1UL << order);
444 if (index == mark)
445 folio_set_readahead(folio);
446 err = filemap_add_folio(ractl->mapping, folio, index, gfp);
447 if (err) {
448 folio_put(folio);
449 return err;
450 }
451
452 ractl->_nr_pages += 1UL << order;
453 ractl->_workingset |= folio_test_workingset(folio);
454 return 0;
455}
456
457void page_cache_ra_order(struct readahead_control *ractl,
458 struct file_ra_state *ra, unsigned int new_order)
459{
460 struct address_space *mapping = ractl->mapping;
461 pgoff_t index = readahead_index(ractl);
462 unsigned int min_order = mapping_min_folio_order(mapping);
463 pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
464 pgoff_t mark = index + ra->size - ra->async_size;
465 unsigned int nofs;
466 int err = 0;
467 gfp_t gfp = readahead_gfp_mask(mapping);
468 unsigned int min_ra_size = max(4, mapping_min_folio_nrpages(mapping));
469
470 /*
471 * Fallback when size < min_nrpages as each folio should be
472 * at least min_nrpages anyway.
473 */
474 if (!mapping_large_folio_support(mapping) || ra->size < min_ra_size)
475 goto fallback;
476
477 limit = min(limit, index + ra->size - 1);
478
479 if (new_order < mapping_max_folio_order(mapping))
480 new_order += 2;
481
482 new_order = min(mapping_max_folio_order(mapping), new_order);
483 new_order = min_t(unsigned int, new_order, ilog2(ra->size));
484 new_order = max(new_order, min_order);
485
486 /* See comment in page_cache_ra_unbounded() */
487 nofs = memalloc_nofs_save();
488 filemap_invalidate_lock_shared(mapping);
489 /*
490 * If the new_order is greater than min_order and index is
491 * already aligned to new_order, then this will be noop as index
492 * aligned to new_order should also be aligned to min_order.
493 */
494 ractl->_index = mapping_align_index(mapping, index);
495 index = readahead_index(ractl);
496
497 while (index <= limit) {
498 unsigned int order = new_order;
499
500 /* Align with smaller pages if needed */
501 if (index & ((1UL << order) - 1))
502 order = __ffs(index);
503 /* Don't allocate pages past EOF */
504 while (order > min_order && index + (1UL << order) - 1 > limit)
505 order--;
506 err = ra_alloc_folio(ractl, index, mark, order, gfp);
507 if (err)
508 break;
509 index += 1UL << order;
510 }
511
512 read_pages(ractl);
513 filemap_invalidate_unlock_shared(mapping);
514 memalloc_nofs_restore(nofs);
515
516 /*
517 * If there were already pages in the page cache, then we may have
518 * left some gaps. Let the regular readahead code take care of this
519 * situation.
520 */
521 if (!err)
522 return;
523fallback:
524 do_page_cache_ra(ractl, ra->size, ra->async_size);
525}
526
527static unsigned long ractl_max_pages(struct readahead_control *ractl,
528 unsigned long req_size)
529{
530 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
531 unsigned long max_pages = ractl->ra->ra_pages;
532
533 /*
534 * If the request exceeds the readahead window, allow the read to
535 * be up to the optimal hardware IO size
536 */
537 if (req_size > max_pages && bdi->io_pages > max_pages)
538 max_pages = min(req_size, bdi->io_pages);
539 return max_pages;
540}
541
542void page_cache_sync_ra(struct readahead_control *ractl,
543 unsigned long req_count)
544{
545 pgoff_t index = readahead_index(ractl);
546 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
547 struct file_ra_state *ra = ractl->ra;
548 unsigned long max_pages, contig_count;
549 pgoff_t prev_index, miss;
550
551 /*
552 * Even if readahead is disabled, issue this request as readahead
553 * as we'll need it to satisfy the requested range. The forced
554 * readahead will do the right thing and limit the read to just the
555 * requested range, which we'll set to 1 page for this case.
556 */
557 if (!ra->ra_pages || blk_cgroup_congested()) {
558 if (!ractl->file)
559 return;
560 req_count = 1;
561 do_forced_ra = true;
562 }
563
564 /* be dumb */
565 if (do_forced_ra) {
566 force_page_cache_ra(ractl, req_count);
567 return;
568 }
569
570 max_pages = ractl_max_pages(ractl, req_count);
571 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
572 /*
573 * A start of file, oversized read, or sequential cache miss:
574 * trivial case: (index - prev_index) == 1
575 * unaligned reads: (index - prev_index) == 0
576 */
577 if (!index || req_count > max_pages || index - prev_index <= 1UL) {
578 ra->start = index;
579 ra->size = get_init_ra_size(req_count, max_pages);
580 ra->async_size = ra->size > req_count ? ra->size - req_count :
581 ra->size >> 1;
582 goto readit;
583 }
584
585 /*
586 * Query the page cache and look for the traces(cached history pages)
587 * that a sequential stream would leave behind.
588 */
589 rcu_read_lock();
590 miss = page_cache_prev_miss(ractl->mapping, index - 1, max_pages);
591 rcu_read_unlock();
592 contig_count = index - miss - 1;
593 /*
594 * Standalone, small random read. Read as is, and do not pollute the
595 * readahead state.
596 */
597 if (contig_count <= req_count) {
598 do_page_cache_ra(ractl, req_count, 0);
599 return;
600 }
601 /*
602 * File cached from the beginning:
603 * it is a strong indication of long-run stream (or whole-file-read)
604 */
605 if (miss == ULONG_MAX)
606 contig_count *= 2;
607 ra->start = index;
608 ra->size = min(contig_count + req_count, max_pages);
609 ra->async_size = 1;
610readit:
611 ractl->_index = ra->start;
612 page_cache_ra_order(ractl, ra, 0);
613}
614EXPORT_SYMBOL_GPL(page_cache_sync_ra);
615
616void page_cache_async_ra(struct readahead_control *ractl,
617 struct folio *folio, unsigned long req_count)
618{
619 unsigned long max_pages;
620 struct file_ra_state *ra = ractl->ra;
621 pgoff_t index = readahead_index(ractl);
622 pgoff_t expected, start;
623 unsigned int order = folio_order(folio);
624
625 /* no readahead */
626 if (!ra->ra_pages)
627 return;
628
629 /*
630 * Same bit is used for PG_readahead and PG_reclaim.
631 */
632 if (folio_test_writeback(folio))
633 return;
634
635 folio_clear_readahead(folio);
636
637 if (blk_cgroup_congested())
638 return;
639
640 max_pages = ractl_max_pages(ractl, req_count);
641 /*
642 * It's the expected callback index, assume sequential access.
643 * Ramp up sizes, and push forward the readahead window.
644 */
645 expected = round_down(ra->start + ra->size - ra->async_size,
646 1UL << order);
647 if (index == expected) {
648 ra->start += ra->size;
649 /*
650 * In the case of MADV_HUGEPAGE, the actual size might exceed
651 * the readahead window.
652 */
653 ra->size = max(ra->size, get_next_ra_size(ra, max_pages));
654 ra->async_size = ra->size;
655 goto readit;
656 }
657
658 /*
659 * Hit a marked folio without valid readahead state.
660 * E.g. interleaved reads.
661 * Query the pagecache for async_size, which normally equals to
662 * readahead size. Ramp it up and use it as the new readahead size.
663 */
664 rcu_read_lock();
665 start = page_cache_next_miss(ractl->mapping, index + 1, max_pages);
666 rcu_read_unlock();
667
668 if (!start || start - index > max_pages)
669 return;
670
671 ra->start = start;
672 ra->size = start - index; /* old async_size */
673 ra->size += req_count;
674 ra->size = get_next_ra_size(ra, max_pages);
675 ra->async_size = ra->size;
676readit:
677 ractl->_index = ra->start;
678 page_cache_ra_order(ractl, ra, order);
679}
680EXPORT_SYMBOL_GPL(page_cache_async_ra);
681
682ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
683{
684 CLASS(fd, f)(fd);
685
686 if (fd_empty(f) || !(fd_file(f)->f_mode & FMODE_READ))
687 return -EBADF;
688
689 /*
690 * The readahead() syscall is intended to run only on files
691 * that can execute readahead. If readahead is not possible
692 * on this file, then we must return -EINVAL.
693 */
694 if (!fd_file(f)->f_mapping || !fd_file(f)->f_mapping->a_ops ||
695 (!S_ISREG(file_inode(fd_file(f))->i_mode) &&
696 !S_ISBLK(file_inode(fd_file(f))->i_mode)))
697 return -EINVAL;
698
699 return vfs_fadvise(fd_file(f), offset, count, POSIX_FADV_WILLNEED);
700}
701
702SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
703{
704 return ksys_readahead(fd, offset, count);
705}
706
707#if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD)
708COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count)
709{
710 return ksys_readahead(fd, compat_arg_u64_glue(offset), count);
711}
712#endif
713
714/**
715 * readahead_expand - Expand a readahead request
716 * @ractl: The request to be expanded
717 * @new_start: The revised start
718 * @new_len: The revised size of the request
719 *
720 * Attempt to expand a readahead request outwards from the current size to the
721 * specified size by inserting locked pages before and after the current window
722 * to increase the size to the new window. This may involve the insertion of
723 * THPs, in which case the window may get expanded even beyond what was
724 * requested.
725 *
726 * The algorithm will stop if it encounters a conflicting page already in the
727 * pagecache and leave a smaller expansion than requested.
728 *
729 * The caller must check for this by examining the revised @ractl object for a
730 * different expansion than was requested.
731 */
732void readahead_expand(struct readahead_control *ractl,
733 loff_t new_start, size_t new_len)
734{
735 struct address_space *mapping = ractl->mapping;
736 struct file_ra_state *ra = ractl->ra;
737 pgoff_t new_index, new_nr_pages;
738 gfp_t gfp_mask = readahead_gfp_mask(mapping);
739 unsigned long min_nrpages = mapping_min_folio_nrpages(mapping);
740 unsigned int min_order = mapping_min_folio_order(mapping);
741
742 new_index = new_start / PAGE_SIZE;
743 /*
744 * Readahead code should have aligned the ractl->_index to
745 * min_nrpages before calling readahead aops.
746 */
747 VM_BUG_ON(!IS_ALIGNED(ractl->_index, min_nrpages));
748
749 /* Expand the leading edge downwards */
750 while (ractl->_index > new_index) {
751 unsigned long index = ractl->_index - 1;
752 struct folio *folio = xa_load(&mapping->i_pages, index);
753
754 if (folio && !xa_is_value(folio))
755 return; /* Folio apparently present */
756
757 folio = filemap_alloc_folio(gfp_mask, min_order);
758 if (!folio)
759 return;
760
761 index = mapping_align_index(mapping, index);
762 if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
763 folio_put(folio);
764 return;
765 }
766 if (unlikely(folio_test_workingset(folio)) &&
767 !ractl->_workingset) {
768 ractl->_workingset = true;
769 psi_memstall_enter(&ractl->_pflags);
770 }
771 ractl->_nr_pages += min_nrpages;
772 ractl->_index = folio->index;
773 }
774
775 new_len += new_start - readahead_pos(ractl);
776 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
777
778 /* Expand the trailing edge upwards */
779 while (ractl->_nr_pages < new_nr_pages) {
780 unsigned long index = ractl->_index + ractl->_nr_pages;
781 struct folio *folio = xa_load(&mapping->i_pages, index);
782
783 if (folio && !xa_is_value(folio))
784 return; /* Folio apparently present */
785
786 folio = filemap_alloc_folio(gfp_mask, min_order);
787 if (!folio)
788 return;
789
790 index = mapping_align_index(mapping, index);
791 if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
792 folio_put(folio);
793 return;
794 }
795 if (unlikely(folio_test_workingset(folio)) &&
796 !ractl->_workingset) {
797 ractl->_workingset = true;
798 psi_memstall_enter(&ractl->_pflags);
799 }
800 ractl->_nr_pages += min_nrpages;
801 if (ra) {
802 ra->size += min_nrpages;
803 ra->async_size += min_nrpages;
804 }
805 }
806}
807EXPORT_SYMBOL(readahead_expand);