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/gfp.h>
12#include <linux/export.h>
13#include <linux/blkdev.h>
14#include <linux/backing-dev.h>
15#include <linux/task_io_accounting_ops.h>
16#include <linux/pagevec.h>
17#include <linux/pagemap.h>
18#include <linux/syscalls.h>
19#include <linux/file.h>
20#include <linux/mm_inline.h>
21
22#include "internal.h"
23
24/*
25 * Initialise a struct file's readahead state. Assumes that the caller has
26 * memset *ra to zero.
27 */
28void
29file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
30{
31 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
32 ra->prev_pos = -1;
33}
34EXPORT_SYMBOL_GPL(file_ra_state_init);
35
36/*
37 * see if a page needs releasing upon read_cache_pages() failure
38 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
39 * before calling, such as the NFS fs marking pages that are cached locally
40 * on disk, thus we need to give the fs a chance to clean up in the event of
41 * an error
42 */
43static void read_cache_pages_invalidate_page(struct address_space *mapping,
44 struct page *page)
45{
46 if (page_has_private(page)) {
47 if (!trylock_page(page))
48 BUG();
49 page->mapping = mapping;
50 do_invalidatepage(page, 0, PAGE_SIZE);
51 page->mapping = NULL;
52 unlock_page(page);
53 }
54 put_page(page);
55}
56
57/*
58 * release a list of pages, invalidating them first if need be
59 */
60static void read_cache_pages_invalidate_pages(struct address_space *mapping,
61 struct list_head *pages)
62{
63 struct page *victim;
64
65 while (!list_empty(pages)) {
66 victim = lru_to_page(pages);
67 list_del(&victim->lru);
68 read_cache_pages_invalidate_page(mapping, victim);
69 }
70}
71
72/**
73 * read_cache_pages - populate an address space with some pages & start reads against them
74 * @mapping: the address_space
75 * @pages: The address of a list_head which contains the target pages. These
76 * pages have their ->index populated and are otherwise uninitialised.
77 * @filler: callback routine for filling a single page.
78 * @data: private data for the callback routine.
79 *
80 * Hides the details of the LRU cache etc from the filesystems.
81 */
82int read_cache_pages(struct address_space *mapping, struct list_head *pages,
83 int (*filler)(void *, struct page *), void *data)
84{
85 struct page *page;
86 int ret = 0;
87
88 while (!list_empty(pages)) {
89 page = lru_to_page(pages);
90 list_del(&page->lru);
91 if (add_to_page_cache_lru(page, mapping, page->index,
92 mapping_gfp_constraint(mapping, GFP_KERNEL))) {
93 read_cache_pages_invalidate_page(mapping, page);
94 continue;
95 }
96 put_page(page);
97
98 ret = filler(data, page);
99 if (unlikely(ret)) {
100 read_cache_pages_invalidate_pages(mapping, pages);
101 break;
102 }
103 task_io_account_read(PAGE_SIZE);
104 }
105 return ret;
106}
107
108EXPORT_SYMBOL(read_cache_pages);
109
110static int read_pages(struct address_space *mapping, struct file *filp,
111 struct list_head *pages, unsigned nr_pages)
112{
113 struct blk_plug plug;
114 unsigned page_idx;
115 int ret;
116
117 blk_start_plug(&plug);
118
119 if (mapping->a_ops->readpages) {
120 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
121 /* Clean up the remaining pages */
122 put_pages_list(pages);
123 goto out;
124 }
125
126 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
127 struct page *page = lru_to_page(pages);
128 list_del(&page->lru);
129 if (!add_to_page_cache_lru(page, mapping, page->index,
130 mapping_gfp_constraint(mapping, GFP_KERNEL))) {
131 mapping->a_ops->readpage(filp, page);
132 }
133 put_page(page);
134 }
135 ret = 0;
136
137out:
138 blk_finish_plug(&plug);
139
140 return ret;
141}
142
143/*
144 * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all
145 * the pages first, then submits them all for I/O. This avoids the very bad
146 * behaviour which would occur if page allocations are causing VM writeback.
147 * We really don't want to intermingle reads and writes like that.
148 *
149 * Returns the number of pages requested, or the maximum amount of I/O allowed.
150 */
151int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
152 pgoff_t offset, unsigned long nr_to_read,
153 unsigned long lookahead_size)
154{
155 struct inode *inode = mapping->host;
156 struct page *page;
157 unsigned long end_index; /* The last page we want to read */
158 LIST_HEAD(page_pool);
159 int page_idx;
160 int ret = 0;
161 loff_t isize = i_size_read(inode);
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->page_tree, page_offset);
179 rcu_read_unlock();
180 if (page && !radix_tree_exceptional_entry(page))
181 continue;
182
183 page = page_cache_alloc_readahead(mapping);
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);
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 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
213 return -EINVAL;
214
215 nr_to_read = min(nr_to_read, inode_to_bdi(mapping->host)->ra_pages);
216 while (nr_to_read) {
217 int err;
218
219 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
220
221 if (this_chunk > nr_to_read)
222 this_chunk = nr_to_read;
223 err = __do_page_cache_readahead(mapping, filp,
224 offset, this_chunk, 0);
225 if (err < 0)
226 return err;
227
228 offset += this_chunk;
229 nr_to_read -= this_chunk;
230 }
231 return 0;
232}
233
234/*
235 * Set the initial window size, round to next power of 2 and square
236 * for small size, x 4 for medium, and x 2 for large
237 * for 128k (32 page) max ra
238 * 1-8 page = 32k initial, > 8 page = 128k initial
239 */
240static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
241{
242 unsigned long newsize = roundup_pow_of_two(size);
243
244 if (newsize <= max / 32)
245 newsize = newsize * 4;
246 else if (newsize <= max / 4)
247 newsize = newsize * 2;
248 else
249 newsize = max;
250
251 return newsize;
252}
253
254/*
255 * Get the previous window size, ramp it up, and
256 * return it as the new window size.
257 */
258static unsigned long get_next_ra_size(struct file_ra_state *ra,
259 unsigned long max)
260{
261 unsigned long cur = ra->size;
262 unsigned long newsize;
263
264 if (cur < max / 16)
265 newsize = 4 * cur;
266 else
267 newsize = 2 * cur;
268
269 return min(newsize, max);
270}
271
272/*
273 * On-demand readahead design.
274 *
275 * The fields in struct file_ra_state represent the most-recently-executed
276 * readahead attempt:
277 *
278 * |<----- async_size ---------|
279 * |------------------- size -------------------->|
280 * |==================#===========================|
281 * ^start ^page marked with PG_readahead
282 *
283 * To overlap application thinking time and disk I/O time, we do
284 * `readahead pipelining': Do not wait until the application consumed all
285 * readahead pages and stalled on the missing page at readahead_index;
286 * Instead, submit an asynchronous readahead I/O as soon as there are
287 * only async_size pages left in the readahead window. Normally async_size
288 * will be equal to size, for maximum pipelining.
289 *
290 * In interleaved sequential reads, concurrent streams on the same fd can
291 * be invalidating each other's readahead state. So we flag the new readahead
292 * page at (start+size-async_size) with PG_readahead, and use it as readahead
293 * indicator. The flag won't be set on already cached pages, to avoid the
294 * readahead-for-nothing fuss, saving pointless page cache lookups.
295 *
296 * prev_pos tracks the last visited byte in the _previous_ read request.
297 * It should be maintained by the caller, and will be used for detecting
298 * small random reads. Note that the readahead algorithm checks loosely
299 * for sequential patterns. Hence interleaved reads might be served as
300 * sequential ones.
301 *
302 * There is a special-case: if the first page which the application tries to
303 * read happens to be the first page of the file, it is assumed that a linear
304 * read is about to happen and the window is immediately set to the initial size
305 * based on I/O request size and the max_readahead.
306 *
307 * The code ramps up the readahead size aggressively at first, but slow down as
308 * it approaches max_readhead.
309 */
310
311/*
312 * Count contiguously cached pages from @offset-1 to @offset-@max,
313 * this count is a conservative estimation of
314 * - length of the sequential read sequence, or
315 * - thrashing threshold in memory tight systems
316 */
317static pgoff_t count_history_pages(struct address_space *mapping,
318 pgoff_t offset, unsigned long max)
319{
320 pgoff_t head;
321
322 rcu_read_lock();
323 head = page_cache_prev_hole(mapping, offset - 1, max);
324 rcu_read_unlock();
325
326 return offset - 1 - head;
327}
328
329/*
330 * page cache context based read-ahead
331 */
332static int try_context_readahead(struct address_space *mapping,
333 struct file_ra_state *ra,
334 pgoff_t offset,
335 unsigned long req_size,
336 unsigned long max)
337{
338 pgoff_t size;
339
340 size = count_history_pages(mapping, offset, max);
341
342 /*
343 * not enough history pages:
344 * it could be a random read
345 */
346 if (size <= req_size)
347 return 0;
348
349 /*
350 * starts from beginning of file:
351 * it is a strong indication of long-run stream (or whole-file-read)
352 */
353 if (size >= offset)
354 size *= 2;
355
356 ra->start = offset;
357 ra->size = min(size + req_size, max);
358 ra->async_size = 1;
359
360 return 1;
361}
362
363/*
364 * A minimal readahead algorithm for trivial sequential/random reads.
365 */
366static unsigned long
367ondemand_readahead(struct address_space *mapping,
368 struct file_ra_state *ra, struct file *filp,
369 bool hit_readahead_marker, pgoff_t offset,
370 unsigned long req_size)
371{
372 unsigned long max = ra->ra_pages;
373 pgoff_t prev_offset;
374
375 /*
376 * start of file
377 */
378 if (!offset)
379 goto initial_readahead;
380
381 /*
382 * It's the expected callback offset, assume sequential access.
383 * Ramp up sizes, and push forward the readahead window.
384 */
385 if ((offset == (ra->start + ra->size - ra->async_size) ||
386 offset == (ra->start + ra->size))) {
387 ra->start += ra->size;
388 ra->size = get_next_ra_size(ra, max);
389 ra->async_size = ra->size;
390 goto readit;
391 }
392
393 /*
394 * Hit a marked page without valid readahead state.
395 * E.g. interleaved reads.
396 * Query the pagecache for async_size, which normally equals to
397 * readahead size. Ramp it up and use it as the new readahead size.
398 */
399 if (hit_readahead_marker) {
400 pgoff_t start;
401
402 rcu_read_lock();
403 start = page_cache_next_hole(mapping, offset + 1, max);
404 rcu_read_unlock();
405
406 if (!start || start - offset > max)
407 return 0;
408
409 ra->start = start;
410 ra->size = start - offset; /* old async_size */
411 ra->size += req_size;
412 ra->size = get_next_ra_size(ra, max);
413 ra->async_size = ra->size;
414 goto readit;
415 }
416
417 /*
418 * oversize read
419 */
420 if (req_size > max)
421 goto initial_readahead;
422
423 /*
424 * sequential cache miss
425 * trivial case: (offset - prev_offset) == 1
426 * unaligned reads: (offset - prev_offset) == 0
427 */
428 prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
429 if (offset - prev_offset <= 1UL)
430 goto initial_readahead;
431
432 /*
433 * Query the page cache and look for the traces(cached history pages)
434 * that a sequential stream would leave behind.
435 */
436 if (try_context_readahead(mapping, ra, offset, req_size, max))
437 goto readit;
438
439 /*
440 * standalone, small random read
441 * Read as is, and do not pollute the readahead state.
442 */
443 return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
444
445initial_readahead:
446 ra->start = offset;
447 ra->size = get_init_ra_size(req_size, max);
448 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
449
450readit:
451 /*
452 * Will this read hit the readahead marker made by itself?
453 * If so, trigger the readahead marker hit now, and merge
454 * the resulted next readahead window into the current one.
455 */
456 if (offset == ra->start && ra->size == ra->async_size) {
457 ra->async_size = get_next_ra_size(ra, max);
458 ra->size += ra->async_size;
459 }
460
461 return ra_submit(ra, mapping, filp);
462}
463
464/**
465 * page_cache_sync_readahead - generic file readahead
466 * @mapping: address_space which holds the pagecache and I/O vectors
467 * @ra: file_ra_state which holds the readahead state
468 * @filp: passed on to ->readpage() and ->readpages()
469 * @offset: start offset into @mapping, in pagecache page-sized units
470 * @req_size: hint: total size of the read which the caller is performing in
471 * pagecache pages
472 *
473 * page_cache_sync_readahead() should be called when a cache miss happened:
474 * it will submit the read. The readahead logic may decide to piggyback more
475 * pages onto the read request if access patterns suggest it will improve
476 * performance.
477 */
478void page_cache_sync_readahead(struct address_space *mapping,
479 struct file_ra_state *ra, struct file *filp,
480 pgoff_t offset, unsigned long req_size)
481{
482 /* no read-ahead */
483 if (!ra->ra_pages)
484 return;
485
486 /* be dumb */
487 if (filp && (filp->f_mode & FMODE_RANDOM)) {
488 force_page_cache_readahead(mapping, filp, offset, req_size);
489 return;
490 }
491
492 /* do read-ahead */
493 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
494}
495EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
496
497/**
498 * page_cache_async_readahead - file readahead for marked pages
499 * @mapping: address_space which holds the pagecache and I/O vectors
500 * @ra: file_ra_state which holds the readahead state
501 * @filp: passed on to ->readpage() and ->readpages()
502 * @page: the page at @offset which has the PG_readahead flag set
503 * @offset: start offset into @mapping, in pagecache page-sized units
504 * @req_size: hint: total size of the read which the caller is performing in
505 * pagecache pages
506 *
507 * page_cache_async_readahead() should be called when a page is used which
508 * has the PG_readahead flag; this is a marker to suggest that the application
509 * has used up enough of the readahead window that we should start pulling in
510 * more pages.
511 */
512void
513page_cache_async_readahead(struct address_space *mapping,
514 struct file_ra_state *ra, struct file *filp,
515 struct page *page, pgoff_t offset,
516 unsigned long req_size)
517{
518 /* no read-ahead */
519 if (!ra->ra_pages)
520 return;
521
522 /*
523 * Same bit is used for PG_readahead and PG_reclaim.
524 */
525 if (PageWriteback(page))
526 return;
527
528 ClearPageReadahead(page);
529
530 /*
531 * Defer asynchronous read-ahead on IO congestion.
532 */
533 if (inode_read_congested(mapping->host))
534 return;
535
536 /* do read-ahead */
537 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
538}
539EXPORT_SYMBOL_GPL(page_cache_async_readahead);
540
541static ssize_t
542do_readahead(struct address_space *mapping, struct file *filp,
543 pgoff_t index, unsigned long nr)
544{
545 if (!mapping || !mapping->a_ops)
546 return -EINVAL;
547
548 return force_page_cache_readahead(mapping, filp, index, nr);
549}
550
551SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
552{
553 ssize_t ret;
554 struct fd f;
555
556 ret = -EBADF;
557 f = fdget(fd);
558 if (f.file) {
559 if (f.file->f_mode & FMODE_READ) {
560 struct address_space *mapping = f.file->f_mapping;
561 pgoff_t start = offset >> PAGE_SHIFT;
562 pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
563 unsigned long len = end - start + 1;
564 ret = do_readahead(mapping, f.file, start, len);
565 }
566 fdput(f);
567 }
568 return ret;
569}
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/pagevec.h>
124#include <linux/pagemap.h>
125#include <linux/psi.h>
126#include <linux/syscalls.h>
127#include <linux/file.h>
128#include <linux/mm_inline.h>
129#include <linux/blk-cgroup.h>
130#include <linux/fadvise.h>
131#include <linux/sched/mm.h>
132
133#include "internal.h"
134
135/*
136 * Initialise a struct file's readahead state. Assumes that the caller has
137 * memset *ra to zero.
138 */
139void
140file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
141{
142 ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
143 ra->prev_pos = -1;
144}
145EXPORT_SYMBOL_GPL(file_ra_state_init);
146
147static void read_pages(struct readahead_control *rac)
148{
149 const struct address_space_operations *aops = rac->mapping->a_ops;
150 struct folio *folio;
151 struct blk_plug plug;
152
153 if (!readahead_count(rac))
154 return;
155
156 if (unlikely(rac->_workingset))
157 psi_memstall_enter(&rac->_pflags);
158 blk_start_plug(&plug);
159
160 if (aops->readahead) {
161 aops->readahead(rac);
162 /*
163 * Clean up the remaining folios. The sizes in ->ra
164 * may be used to size the next readahead, so make sure
165 * they accurately reflect what happened.
166 */
167 while ((folio = readahead_folio(rac)) != NULL) {
168 unsigned long nr = folio_nr_pages(folio);
169
170 folio_get(folio);
171 rac->ra->size -= nr;
172 if (rac->ra->async_size >= nr) {
173 rac->ra->async_size -= nr;
174 filemap_remove_folio(folio);
175 }
176 folio_unlock(folio);
177 folio_put(folio);
178 }
179 } else {
180 while ((folio = readahead_folio(rac)) != NULL)
181 aops->read_folio(rac->file, folio);
182 }
183
184 blk_finish_plug(&plug);
185 if (unlikely(rac->_workingset))
186 psi_memstall_leave(&rac->_pflags);
187 rac->_workingset = false;
188
189 BUG_ON(readahead_count(rac));
190}
191
192/**
193 * page_cache_ra_unbounded - Start unchecked readahead.
194 * @ractl: Readahead control.
195 * @nr_to_read: The number of pages to read.
196 * @lookahead_size: Where to start the next readahead.
197 *
198 * This function is for filesystems to call when they want to start
199 * readahead beyond a file's stated i_size. This is almost certainly
200 * not the function you want to call. Use page_cache_async_readahead()
201 * or page_cache_sync_readahead() instead.
202 *
203 * Context: File is referenced by caller. Mutexes may be held by caller.
204 * May sleep, but will not reenter filesystem to reclaim memory.
205 */
206void page_cache_ra_unbounded(struct readahead_control *ractl,
207 unsigned long nr_to_read, unsigned long lookahead_size)
208{
209 struct address_space *mapping = ractl->mapping;
210 unsigned long index = readahead_index(ractl);
211 gfp_t gfp_mask = readahead_gfp_mask(mapping);
212 unsigned long i;
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 /*
228 * Preallocate as many pages as we will need.
229 */
230 for (i = 0; i < nr_to_read; i++) {
231 struct folio *folio = xa_load(&mapping->i_pages, index + i);
232
233 if (folio && !xa_is_value(folio)) {
234 /*
235 * Page already present? Kick off the current batch
236 * of contiguous pages before continuing with the
237 * next batch. This page may be the one we would
238 * have intended to mark as Readahead, but we don't
239 * have a stable reference to this page, and it's
240 * not worth getting one just for that.
241 */
242 read_pages(ractl);
243 ractl->_index++;
244 i = ractl->_index + ractl->_nr_pages - index - 1;
245 continue;
246 }
247
248 folio = filemap_alloc_folio(gfp_mask, 0);
249 if (!folio)
250 break;
251 if (filemap_add_folio(mapping, folio, index + i,
252 gfp_mask) < 0) {
253 folio_put(folio);
254 read_pages(ractl);
255 ractl->_index++;
256 i = ractl->_index + ractl->_nr_pages - index - 1;
257 continue;
258 }
259 if (i == nr_to_read - lookahead_size)
260 folio_set_readahead(folio);
261 ractl->_workingset |= folio_test_workingset(folio);
262 ractl->_nr_pages++;
263 }
264
265 /*
266 * Now start the IO. We ignore I/O errors - if the folio is not
267 * uptodate then the caller will launch read_folio again, and
268 * will then handle the error.
269 */
270 read_pages(ractl);
271 filemap_invalidate_unlock_shared(mapping);
272 memalloc_nofs_restore(nofs);
273}
274EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
275
276/*
277 * do_page_cache_ra() actually reads a chunk of disk. It allocates
278 * the pages first, then submits them for I/O. This avoids the very bad
279 * behaviour which would occur if page allocations are causing VM writeback.
280 * We really don't want to intermingle reads and writes like that.
281 */
282static void do_page_cache_ra(struct readahead_control *ractl,
283 unsigned long nr_to_read, unsigned long lookahead_size)
284{
285 struct inode *inode = ractl->mapping->host;
286 unsigned long index = readahead_index(ractl);
287 loff_t isize = i_size_read(inode);
288 pgoff_t end_index; /* The last page we want to read */
289
290 if (isize == 0)
291 return;
292
293 end_index = (isize - 1) >> PAGE_SHIFT;
294 if (index > end_index)
295 return;
296 /* Don't read past the page containing the last byte of the file */
297 if (nr_to_read > end_index - index)
298 nr_to_read = end_index - index + 1;
299
300 page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
301}
302
303/*
304 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
305 * memory at once.
306 */
307void force_page_cache_ra(struct readahead_control *ractl,
308 unsigned long nr_to_read)
309{
310 struct address_space *mapping = ractl->mapping;
311 struct file_ra_state *ra = ractl->ra;
312 struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
313 unsigned long max_pages, index;
314
315 if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead))
316 return;
317
318 /*
319 * If the request exceeds the readahead window, allow the read to
320 * be up to the optimal hardware IO size
321 */
322 index = readahead_index(ractl);
323 max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
324 nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
325 while (nr_to_read) {
326 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
327
328 if (this_chunk > nr_to_read)
329 this_chunk = nr_to_read;
330 ractl->_index = index;
331 do_page_cache_ra(ractl, this_chunk, 0);
332
333 index += this_chunk;
334 nr_to_read -= this_chunk;
335 }
336}
337
338/*
339 * Set the initial window size, round to next power of 2 and square
340 * for small size, x 4 for medium, and x 2 for large
341 * for 128k (32 page) max ra
342 * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
343 */
344static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
345{
346 unsigned long newsize = roundup_pow_of_two(size);
347
348 if (newsize <= max / 32)
349 newsize = newsize * 4;
350 else if (newsize <= max / 4)
351 newsize = newsize * 2;
352 else
353 newsize = max;
354
355 return newsize;
356}
357
358/*
359 * Get the previous window size, ramp it up, and
360 * return it as the new window size.
361 */
362static unsigned long get_next_ra_size(struct file_ra_state *ra,
363 unsigned long max)
364{
365 unsigned long cur = ra->size;
366
367 if (cur < max / 16)
368 return 4 * cur;
369 if (cur <= max / 2)
370 return 2 * cur;
371 return max;
372}
373
374/*
375 * On-demand readahead design.
376 *
377 * The fields in struct file_ra_state represent the most-recently-executed
378 * readahead attempt:
379 *
380 * |<----- async_size ---------|
381 * |------------------- size -------------------->|
382 * |==================#===========================|
383 * ^start ^page marked with PG_readahead
384 *
385 * To overlap application thinking time and disk I/O time, we do
386 * `readahead pipelining': Do not wait until the application consumed all
387 * readahead pages and stalled on the missing page at readahead_index;
388 * Instead, submit an asynchronous readahead I/O as soon as there are
389 * only async_size pages left in the readahead window. Normally async_size
390 * will be equal to size, for maximum pipelining.
391 *
392 * In interleaved sequential reads, concurrent streams on the same fd can
393 * be invalidating each other's readahead state. So we flag the new readahead
394 * page at (start+size-async_size) with PG_readahead, and use it as readahead
395 * indicator. The flag won't be set on already cached pages, to avoid the
396 * readahead-for-nothing fuss, saving pointless page cache lookups.
397 *
398 * prev_pos tracks the last visited byte in the _previous_ read request.
399 * It should be maintained by the caller, and will be used for detecting
400 * small random reads. Note that the readahead algorithm checks loosely
401 * for sequential patterns. Hence interleaved reads might be served as
402 * sequential ones.
403 *
404 * There is a special-case: if the first page which the application tries to
405 * read happens to be the first page of the file, it is assumed that a linear
406 * read is about to happen and the window is immediately set to the initial size
407 * based on I/O request size and the max_readahead.
408 *
409 * The code ramps up the readahead size aggressively at first, but slow down as
410 * it approaches max_readhead.
411 */
412
413/*
414 * Count contiguously cached pages from @index-1 to @index-@max,
415 * this count is a conservative estimation of
416 * - length of the sequential read sequence, or
417 * - thrashing threshold in memory tight systems
418 */
419static pgoff_t count_history_pages(struct address_space *mapping,
420 pgoff_t index, unsigned long max)
421{
422 pgoff_t head;
423
424 rcu_read_lock();
425 head = page_cache_prev_miss(mapping, index - 1, max);
426 rcu_read_unlock();
427
428 return index - 1 - head;
429}
430
431/*
432 * page cache context based readahead
433 */
434static int try_context_readahead(struct address_space *mapping,
435 struct file_ra_state *ra,
436 pgoff_t index,
437 unsigned long req_size,
438 unsigned long max)
439{
440 pgoff_t size;
441
442 size = count_history_pages(mapping, index, max);
443
444 /*
445 * not enough history pages:
446 * it could be a random read
447 */
448 if (size <= req_size)
449 return 0;
450
451 /*
452 * starts from beginning of file:
453 * it is a strong indication of long-run stream (or whole-file-read)
454 */
455 if (size >= index)
456 size *= 2;
457
458 ra->start = index;
459 ra->size = min(size + req_size, max);
460 ra->async_size = 1;
461
462 return 1;
463}
464
465/*
466 * There are some parts of the kernel which assume that PMD entries
467 * are exactly HPAGE_PMD_ORDER. Those should be fixed, but until then,
468 * limit the maximum allocation order to PMD size. I'm not aware of any
469 * assumptions about maximum order if THP are disabled, but 8 seems like
470 * a good order (that's 1MB if you're using 4kB pages)
471 */
472#ifdef CONFIG_TRANSPARENT_HUGEPAGE
473#define MAX_PAGECACHE_ORDER HPAGE_PMD_ORDER
474#else
475#define MAX_PAGECACHE_ORDER 8
476#endif
477
478static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
479 pgoff_t mark, unsigned int order, gfp_t gfp)
480{
481 int err;
482 struct folio *folio = filemap_alloc_folio(gfp, order);
483
484 if (!folio)
485 return -ENOMEM;
486 mark = round_up(mark, 1UL << order);
487 if (index == mark)
488 folio_set_readahead(folio);
489 err = filemap_add_folio(ractl->mapping, folio, index, gfp);
490 if (err) {
491 folio_put(folio);
492 return err;
493 }
494
495 ractl->_nr_pages += 1UL << order;
496 ractl->_workingset |= folio_test_workingset(folio);
497 return 0;
498}
499
500void page_cache_ra_order(struct readahead_control *ractl,
501 struct file_ra_state *ra, unsigned int new_order)
502{
503 struct address_space *mapping = ractl->mapping;
504 pgoff_t index = readahead_index(ractl);
505 pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
506 pgoff_t mark = index + ra->size - ra->async_size;
507 int err = 0;
508 gfp_t gfp = readahead_gfp_mask(mapping);
509
510 if (!mapping_large_folio_support(mapping) || ra->size < 4)
511 goto fallback;
512
513 limit = min(limit, index + ra->size - 1);
514
515 if (new_order < MAX_PAGECACHE_ORDER) {
516 new_order += 2;
517 if (new_order > MAX_PAGECACHE_ORDER)
518 new_order = MAX_PAGECACHE_ORDER;
519 while ((1 << new_order) > ra->size)
520 new_order--;
521 }
522
523 filemap_invalidate_lock_shared(mapping);
524 while (index <= limit) {
525 unsigned int order = new_order;
526
527 /* Align with smaller pages if needed */
528 if (index & ((1UL << order) - 1)) {
529 order = __ffs(index);
530 if (order == 1)
531 order = 0;
532 }
533 /* Don't allocate pages past EOF */
534 while (index + (1UL << order) - 1 > limit) {
535 if (--order == 1)
536 order = 0;
537 }
538 err = ra_alloc_folio(ractl, index, mark, order, gfp);
539 if (err)
540 break;
541 index += 1UL << order;
542 }
543
544 if (index > limit) {
545 ra->size += index - limit - 1;
546 ra->async_size += index - limit - 1;
547 }
548
549 read_pages(ractl);
550 filemap_invalidate_unlock_shared(mapping);
551
552 /*
553 * If there were already pages in the page cache, then we may have
554 * left some gaps. Let the regular readahead code take care of this
555 * situation.
556 */
557 if (!err)
558 return;
559fallback:
560 do_page_cache_ra(ractl, ra->size, ra->async_size);
561}
562
563/*
564 * A minimal readahead algorithm for trivial sequential/random reads.
565 */
566static void ondemand_readahead(struct readahead_control *ractl,
567 struct folio *folio, unsigned long req_size)
568{
569 struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
570 struct file_ra_state *ra = ractl->ra;
571 unsigned long max_pages = ra->ra_pages;
572 unsigned long add_pages;
573 pgoff_t index = readahead_index(ractl);
574 pgoff_t expected, prev_index;
575 unsigned int order = folio ? folio_order(folio) : 0;
576
577 /*
578 * If the request exceeds the readahead window, allow the read to
579 * be up to the optimal hardware IO size
580 */
581 if (req_size > max_pages && bdi->io_pages > max_pages)
582 max_pages = min(req_size, bdi->io_pages);
583
584 /*
585 * start of file
586 */
587 if (!index)
588 goto initial_readahead;
589
590 /*
591 * It's the expected callback index, assume sequential access.
592 * Ramp up sizes, and push forward the readahead window.
593 */
594 expected = round_up(ra->start + ra->size - ra->async_size,
595 1UL << order);
596 if (index == expected || index == (ra->start + ra->size)) {
597 ra->start += ra->size;
598 ra->size = get_next_ra_size(ra, max_pages);
599 ra->async_size = ra->size;
600 goto readit;
601 }
602
603 /*
604 * Hit a marked folio without valid readahead state.
605 * E.g. interleaved reads.
606 * Query the pagecache for async_size, which normally equals to
607 * readahead size. Ramp it up and use it as the new readahead size.
608 */
609 if (folio) {
610 pgoff_t start;
611
612 rcu_read_lock();
613 start = page_cache_next_miss(ractl->mapping, index + 1,
614 max_pages);
615 rcu_read_unlock();
616
617 if (!start || start - index > max_pages)
618 return;
619
620 ra->start = start;
621 ra->size = start - index; /* old async_size */
622 ra->size += req_size;
623 ra->size = get_next_ra_size(ra, max_pages);
624 ra->async_size = ra->size;
625 goto readit;
626 }
627
628 /*
629 * oversize read
630 */
631 if (req_size > max_pages)
632 goto initial_readahead;
633
634 /*
635 * sequential cache miss
636 * trivial case: (index - prev_index) == 1
637 * unaligned reads: (index - prev_index) == 0
638 */
639 prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
640 if (index - prev_index <= 1UL)
641 goto initial_readahead;
642
643 /*
644 * Query the page cache and look for the traces(cached history pages)
645 * that a sequential stream would leave behind.
646 */
647 if (try_context_readahead(ractl->mapping, ra, index, req_size,
648 max_pages))
649 goto readit;
650
651 /*
652 * standalone, small random read
653 * Read as is, and do not pollute the readahead state.
654 */
655 do_page_cache_ra(ractl, req_size, 0);
656 return;
657
658initial_readahead:
659 ra->start = index;
660 ra->size = get_init_ra_size(req_size, max_pages);
661 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
662
663readit:
664 /*
665 * Will this read hit the readahead marker made by itself?
666 * If so, trigger the readahead marker hit now, and merge
667 * the resulted next readahead window into the current one.
668 * Take care of maximum IO pages as above.
669 */
670 if (index == ra->start && ra->size == ra->async_size) {
671 add_pages = get_next_ra_size(ra, max_pages);
672 if (ra->size + add_pages <= max_pages) {
673 ra->async_size = add_pages;
674 ra->size += add_pages;
675 } else {
676 ra->size = max_pages;
677 ra->async_size = max_pages >> 1;
678 }
679 }
680
681 ractl->_index = ra->start;
682 page_cache_ra_order(ractl, ra, order);
683}
684
685void page_cache_sync_ra(struct readahead_control *ractl,
686 unsigned long req_count)
687{
688 bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
689
690 /*
691 * Even if readahead is disabled, issue this request as readahead
692 * as we'll need it to satisfy the requested range. The forced
693 * readahead will do the right thing and limit the read to just the
694 * requested range, which we'll set to 1 page for this case.
695 */
696 if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
697 if (!ractl->file)
698 return;
699 req_count = 1;
700 do_forced_ra = true;
701 }
702
703 /* be dumb */
704 if (do_forced_ra) {
705 force_page_cache_ra(ractl, req_count);
706 return;
707 }
708
709 ondemand_readahead(ractl, NULL, req_count);
710}
711EXPORT_SYMBOL_GPL(page_cache_sync_ra);
712
713void page_cache_async_ra(struct readahead_control *ractl,
714 struct folio *folio, unsigned long req_count)
715{
716 /* no readahead */
717 if (!ractl->ra->ra_pages)
718 return;
719
720 /*
721 * Same bit is used for PG_readahead and PG_reclaim.
722 */
723 if (folio_test_writeback(folio))
724 return;
725
726 folio_clear_readahead(folio);
727
728 if (blk_cgroup_congested())
729 return;
730
731 ondemand_readahead(ractl, folio, req_count);
732}
733EXPORT_SYMBOL_GPL(page_cache_async_ra);
734
735ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
736{
737 ssize_t ret;
738 struct fd f;
739
740 ret = -EBADF;
741 f = fdget(fd);
742 if (!f.file || !(f.file->f_mode & FMODE_READ))
743 goto out;
744
745 /*
746 * The readahead() syscall is intended to run only on files
747 * that can execute readahead. If readahead is not possible
748 * on this file, then we must return -EINVAL.
749 */
750 ret = -EINVAL;
751 if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
752 !S_ISREG(file_inode(f.file)->i_mode))
753 goto out;
754
755 ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
756out:
757 fdput(f);
758 return ret;
759}
760
761SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
762{
763 return ksys_readahead(fd, offset, count);
764}
765
766#if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD)
767COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count)
768{
769 return ksys_readahead(fd, compat_arg_u64_glue(offset), count);
770}
771#endif
772
773/**
774 * readahead_expand - Expand a readahead request
775 * @ractl: The request to be expanded
776 * @new_start: The revised start
777 * @new_len: The revised size of the request
778 *
779 * Attempt to expand a readahead request outwards from the current size to the
780 * specified size by inserting locked pages before and after the current window
781 * to increase the size to the new window. This may involve the insertion of
782 * THPs, in which case the window may get expanded even beyond what was
783 * requested.
784 *
785 * The algorithm will stop if it encounters a conflicting page already in the
786 * pagecache and leave a smaller expansion than requested.
787 *
788 * The caller must check for this by examining the revised @ractl object for a
789 * different expansion than was requested.
790 */
791void readahead_expand(struct readahead_control *ractl,
792 loff_t new_start, size_t new_len)
793{
794 struct address_space *mapping = ractl->mapping;
795 struct file_ra_state *ra = ractl->ra;
796 pgoff_t new_index, new_nr_pages;
797 gfp_t gfp_mask = readahead_gfp_mask(mapping);
798
799 new_index = new_start / PAGE_SIZE;
800
801 /* Expand the leading edge downwards */
802 while (ractl->_index > new_index) {
803 unsigned long index = ractl->_index - 1;
804 struct page *page = xa_load(&mapping->i_pages, index);
805
806 if (page && !xa_is_value(page))
807 return; /* Page apparently present */
808
809 page = __page_cache_alloc(gfp_mask);
810 if (!page)
811 return;
812 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
813 put_page(page);
814 return;
815 }
816
817 ractl->_nr_pages++;
818 ractl->_index = page->index;
819 }
820
821 new_len += new_start - readahead_pos(ractl);
822 new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
823
824 /* Expand the trailing edge upwards */
825 while (ractl->_nr_pages < new_nr_pages) {
826 unsigned long index = ractl->_index + ractl->_nr_pages;
827 struct page *page = xa_load(&mapping->i_pages, index);
828
829 if (page && !xa_is_value(page))
830 return; /* Page apparently present */
831
832 page = __page_cache_alloc(gfp_mask);
833 if (!page)
834 return;
835 if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
836 put_page(page);
837 return;
838 }
839 if (unlikely(PageWorkingset(page)) && !ractl->_workingset) {
840 ractl->_workingset = true;
841 psi_memstall_enter(&ractl->_pflags);
842 }
843 ractl->_nr_pages++;
844 if (ra) {
845 ra->size++;
846 ra->async_size++;
847 }
848 }
849}
850EXPORT_SYMBOL(readahead_expand);