1 // SPDX-License-Identifier: GPL-2.0-only
3 * mm/readahead.c - address_space-level file readahead.
5 * Copyright (C) 2002, Linus Torvalds
7 * 09Apr2002 Andrew Morton
11 #include <linux/kernel.h>
12 #include <linux/dax.h>
13 #include <linux/gfp.h>
14 #include <linux/export.h>
15 #include <linux/blkdev.h>
16 #include <linux/backing-dev.h>
17 #include <linux/task_io_accounting_ops.h>
18 #include <linux/pagevec.h>
19 #include <linux/pagemap.h>
20 #include <linux/syscalls.h>
21 #include <linux/file.h>
22 #include <linux/mm_inline.h>
23 #include <linux/blk-cgroup.h>
24 #include <linux/fadvise.h>
25 #include <linux/sched/mm.h>
30 * Initialise a struct file's readahead state. Assumes that the caller has
34 file_ra_state_init(struct file_ra_state
*ra
, struct address_space
*mapping
)
36 ra
->ra_pages
= inode_to_bdi(mapping
->host
)->ra_pages
;
39 EXPORT_SYMBOL_GPL(file_ra_state_init
);
42 * see if a page needs releasing upon read_cache_pages() failure
43 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
44 * before calling, such as the NFS fs marking pages that are cached locally
45 * on disk, thus we need to give the fs a chance to clean up in the event of
48 static void read_cache_pages_invalidate_page(struct address_space
*mapping
,
51 if (page_has_private(page
)) {
52 if (!trylock_page(page
))
54 page
->mapping
= mapping
;
55 do_invalidatepage(page
, 0, PAGE_SIZE
);
63 * release a list of pages, invalidating them first if need be
65 static void read_cache_pages_invalidate_pages(struct address_space
*mapping
,
66 struct list_head
*pages
)
70 while (!list_empty(pages
)) {
71 victim
= lru_to_page(pages
);
72 list_del(&victim
->lru
);
73 read_cache_pages_invalidate_page(mapping
, victim
);
78 * read_cache_pages - populate an address space with some pages & start reads against them
79 * @mapping: the address_space
80 * @pages: The address of a list_head which contains the target pages. These
81 * pages have their ->index populated and are otherwise uninitialised.
82 * @filler: callback routine for filling a single page.
83 * @data: private data for the callback routine.
85 * Hides the details of the LRU cache etc from the filesystems.
87 * Returns: %0 on success, error return by @filler otherwise
89 int read_cache_pages(struct address_space
*mapping
, struct list_head
*pages
,
90 int (*filler
)(void *, struct page
*), void *data
)
95 while (!list_empty(pages
)) {
96 page
= lru_to_page(pages
);
98 if (add_to_page_cache_lru(page
, mapping
, page
->index
,
99 readahead_gfp_mask(mapping
))) {
100 read_cache_pages_invalidate_page(mapping
, page
);
105 ret
= filler(data
, page
);
107 read_cache_pages_invalidate_pages(mapping
, pages
);
110 task_io_account_read(PAGE_SIZE
);
115 EXPORT_SYMBOL(read_cache_pages
);
117 static void read_pages(struct readahead_control
*rac
, struct list_head
*pages
,
120 const struct address_space_operations
*aops
= rac
->mapping
->a_ops
;
122 struct blk_plug plug
;
124 if (!readahead_count(rac
))
127 blk_start_plug(&plug
);
129 if (aops
->readahead
) {
130 aops
->readahead(rac
);
131 /* Clean up the remaining pages */
132 while ((page
= readahead_page(rac
))) {
136 } else if (aops
->readpages
) {
137 aops
->readpages(rac
->file
, rac
->mapping
, pages
,
138 readahead_count(rac
));
139 /* Clean up the remaining pages */
140 put_pages_list(pages
);
141 rac
->_index
+= rac
->_nr_pages
;
144 while ((page
= readahead_page(rac
))) {
145 aops
->readpage(rac
->file
, page
);
150 blk_finish_plug(&plug
);
152 BUG_ON(!list_empty(pages
));
153 BUG_ON(readahead_count(rac
));
161 * page_cache_ra_unbounded - Start unchecked readahead.
162 * @ractl: Readahead control.
163 * @nr_to_read: The number of pages to read.
164 * @lookahead_size: Where to start the next readahead.
166 * This function is for filesystems to call when they want to start
167 * readahead beyond a file's stated i_size. This is almost certainly
168 * not the function you want to call. Use page_cache_async_readahead()
169 * or page_cache_sync_readahead() instead.
171 * Context: File is referenced by caller. Mutexes may be held by caller.
172 * May sleep, but will not reenter filesystem to reclaim memory.
174 void page_cache_ra_unbounded(struct readahead_control
*ractl
,
175 unsigned long nr_to_read
, unsigned long lookahead_size
)
177 struct address_space
*mapping
= ractl
->mapping
;
178 unsigned long index
= readahead_index(ractl
);
179 LIST_HEAD(page_pool
);
180 gfp_t gfp_mask
= readahead_gfp_mask(mapping
);
184 * Partway through the readahead operation, we will have added
185 * locked pages to the page cache, but will not yet have submitted
186 * them for I/O. Adding another page may need to allocate memory,
187 * which can trigger memory reclaim. Telling the VM we're in
188 * the middle of a filesystem operation will cause it to not
189 * touch file-backed pages, preventing a deadlock. Most (all?)
190 * filesystems already specify __GFP_NOFS in their mapping's
191 * gfp_mask, but let's be explicit here.
193 unsigned int nofs
= memalloc_nofs_save();
196 * Preallocate as many pages as we will need.
198 for (i
= 0; i
< nr_to_read
; i
++) {
199 struct page
*page
= xa_load(&mapping
->i_pages
, index
+ i
);
201 BUG_ON(index
+ i
!= ractl
->_index
+ ractl
->_nr_pages
);
203 if (page
&& !xa_is_value(page
)) {
205 * Page already present? Kick off the current batch
206 * of contiguous pages before continuing with the
207 * next batch. This page may be the one we would
208 * have intended to mark as Readahead, but we don't
209 * have a stable reference to this page, and it's
210 * not worth getting one just for that.
212 read_pages(ractl
, &page_pool
, true);
216 page
= __page_cache_alloc(gfp_mask
);
219 if (mapping
->a_ops
->readpages
) {
220 page
->index
= index
+ i
;
221 list_add(&page
->lru
, &page_pool
);
222 } else if (add_to_page_cache_lru(page
, mapping
, index
+ i
,
225 read_pages(ractl
, &page_pool
, true);
228 if (i
== nr_to_read
- lookahead_size
)
229 SetPageReadahead(page
);
234 * Now start the IO. We ignore I/O errors - if the page is not
235 * uptodate then the caller will launch readpage again, and
236 * will then handle the error.
238 read_pages(ractl
, &page_pool
, false);
239 memalloc_nofs_restore(nofs
);
241 EXPORT_SYMBOL_GPL(page_cache_ra_unbounded
);
244 * do_page_cache_ra() actually reads a chunk of disk. It allocates
245 * the pages first, then submits them for I/O. This avoids the very bad
246 * behaviour which would occur if page allocations are causing VM writeback.
247 * We really don't want to intermingle reads and writes like that.
249 void do_page_cache_ra(struct readahead_control
*ractl
,
250 unsigned long nr_to_read
, unsigned long lookahead_size
)
252 struct inode
*inode
= ractl
->mapping
->host
;
253 unsigned long index
= readahead_index(ractl
);
254 loff_t isize
= i_size_read(inode
);
255 pgoff_t end_index
; /* The last page we want to read */
260 end_index
= (isize
- 1) >> PAGE_SHIFT
;
261 if (index
> end_index
)
263 /* Don't read past the page containing the last byte of the file */
264 if (nr_to_read
> end_index
- index
)
265 nr_to_read
= end_index
- index
+ 1;
267 page_cache_ra_unbounded(ractl
, nr_to_read
, lookahead_size
);
271 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
274 void force_page_cache_ra(struct readahead_control
*ractl
,
275 struct file_ra_state
*ra
, unsigned long nr_to_read
)
277 struct address_space
*mapping
= ractl
->mapping
;
278 struct backing_dev_info
*bdi
= inode_to_bdi(mapping
->host
);
279 unsigned long max_pages
, index
;
281 if (unlikely(!mapping
->a_ops
->readpage
&& !mapping
->a_ops
->readpages
&&
282 !mapping
->a_ops
->readahead
))
286 * If the request exceeds the readahead window, allow the read to
287 * be up to the optimal hardware IO size
289 index
= readahead_index(ractl
);
290 max_pages
= max_t(unsigned long, bdi
->io_pages
, ra
->ra_pages
);
291 nr_to_read
= min_t(unsigned long, nr_to_read
, max_pages
);
293 unsigned long this_chunk
= (2 * 1024 * 1024) / PAGE_SIZE
;
295 if (this_chunk
> nr_to_read
)
296 this_chunk
= nr_to_read
;
297 ractl
->_index
= index
;
298 do_page_cache_ra(ractl
, this_chunk
, 0);
301 nr_to_read
-= this_chunk
;
306 * Set the initial window size, round to next power of 2 and square
307 * for small size, x 4 for medium, and x 2 for large
308 * for 128k (32 page) max ra
309 * 1-8 page = 32k initial, > 8 page = 128k initial
311 static unsigned long get_init_ra_size(unsigned long size
, unsigned long max
)
313 unsigned long newsize
= roundup_pow_of_two(size
);
315 if (newsize
<= max
/ 32)
316 newsize
= newsize
* 4;
317 else if (newsize
<= max
/ 4)
318 newsize
= newsize
* 2;
326 * Get the previous window size, ramp it up, and
327 * return it as the new window size.
329 static unsigned long get_next_ra_size(struct file_ra_state
*ra
,
332 unsigned long cur
= ra
->size
;
342 * On-demand readahead design.
344 * The fields in struct file_ra_state represent the most-recently-executed
347 * |<----- async_size ---------|
348 * |------------------- size -------------------->|
349 * |==================#===========================|
350 * ^start ^page marked with PG_readahead
352 * To overlap application thinking time and disk I/O time, we do
353 * `readahead pipelining': Do not wait until the application consumed all
354 * readahead pages and stalled on the missing page at readahead_index;
355 * Instead, submit an asynchronous readahead I/O as soon as there are
356 * only async_size pages left in the readahead window. Normally async_size
357 * will be equal to size, for maximum pipelining.
359 * In interleaved sequential reads, concurrent streams on the same fd can
360 * be invalidating each other's readahead state. So we flag the new readahead
361 * page at (start+size-async_size) with PG_readahead, and use it as readahead
362 * indicator. The flag won't be set on already cached pages, to avoid the
363 * readahead-for-nothing fuss, saving pointless page cache lookups.
365 * prev_pos tracks the last visited byte in the _previous_ read request.
366 * It should be maintained by the caller, and will be used for detecting
367 * small random reads. Note that the readahead algorithm checks loosely
368 * for sequential patterns. Hence interleaved reads might be served as
371 * There is a special-case: if the first page which the application tries to
372 * read happens to be the first page of the file, it is assumed that a linear
373 * read is about to happen and the window is immediately set to the initial size
374 * based on I/O request size and the max_readahead.
376 * The code ramps up the readahead size aggressively at first, but slow down as
377 * it approaches max_readhead.
381 * Count contiguously cached pages from @index-1 to @index-@max,
382 * this count is a conservative estimation of
383 * - length of the sequential read sequence, or
384 * - thrashing threshold in memory tight systems
386 static pgoff_t
count_history_pages(struct address_space
*mapping
,
387 pgoff_t index
, unsigned long max
)
392 head
= page_cache_prev_miss(mapping
, index
- 1, max
);
395 return index
- 1 - head
;
399 * page cache context based read-ahead
401 static int try_context_readahead(struct address_space
*mapping
,
402 struct file_ra_state
*ra
,
404 unsigned long req_size
,
409 size
= count_history_pages(mapping
, index
, max
);
412 * not enough history pages:
413 * it could be a random read
415 if (size
<= req_size
)
419 * starts from beginning of file:
420 * it is a strong indication of long-run stream (or whole-file-read)
426 ra
->size
= min(size
+ req_size
, max
);
433 * A minimal readahead algorithm for trivial sequential/random reads.
435 static void ondemand_readahead(struct readahead_control
*ractl
,
436 struct file_ra_state
*ra
, bool hit_readahead_marker
,
437 unsigned long req_size
)
439 struct backing_dev_info
*bdi
= inode_to_bdi(ractl
->mapping
->host
);
440 unsigned long max_pages
= ra
->ra_pages
;
441 unsigned long add_pages
;
442 unsigned long index
= readahead_index(ractl
);
446 * If the request exceeds the readahead window, allow the read to
447 * be up to the optimal hardware IO size
449 if (req_size
> max_pages
&& bdi
->io_pages
> max_pages
)
450 max_pages
= min(req_size
, bdi
->io_pages
);
456 goto initial_readahead
;
459 * It's the expected callback index, assume sequential access.
460 * Ramp up sizes, and push forward the readahead window.
462 if ((index
== (ra
->start
+ ra
->size
- ra
->async_size
) ||
463 index
== (ra
->start
+ ra
->size
))) {
464 ra
->start
+= ra
->size
;
465 ra
->size
= get_next_ra_size(ra
, max_pages
);
466 ra
->async_size
= ra
->size
;
471 * Hit a marked page without valid readahead state.
472 * E.g. interleaved reads.
473 * Query the pagecache for async_size, which normally equals to
474 * readahead size. Ramp it up and use it as the new readahead size.
476 if (hit_readahead_marker
) {
480 start
= page_cache_next_miss(ractl
->mapping
, index
+ 1,
484 if (!start
|| start
- index
> max_pages
)
488 ra
->size
= start
- index
; /* old async_size */
489 ra
->size
+= req_size
;
490 ra
->size
= get_next_ra_size(ra
, max_pages
);
491 ra
->async_size
= ra
->size
;
498 if (req_size
> max_pages
)
499 goto initial_readahead
;
502 * sequential cache miss
503 * trivial case: (index - prev_index) == 1
504 * unaligned reads: (index - prev_index) == 0
506 prev_index
= (unsigned long long)ra
->prev_pos
>> PAGE_SHIFT
;
507 if (index
- prev_index
<= 1UL)
508 goto initial_readahead
;
511 * Query the page cache and look for the traces(cached history pages)
512 * that a sequential stream would leave behind.
514 if (try_context_readahead(ractl
->mapping
, ra
, index
, req_size
,
519 * standalone, small random read
520 * Read as is, and do not pollute the readahead state.
522 do_page_cache_ra(ractl
, req_size
, 0);
527 ra
->size
= get_init_ra_size(req_size
, max_pages
);
528 ra
->async_size
= ra
->size
> req_size
? ra
->size
- req_size
: ra
->size
;
532 * Will this read hit the readahead marker made by itself?
533 * If so, trigger the readahead marker hit now, and merge
534 * the resulted next readahead window into the current one.
535 * Take care of maximum IO pages as above.
537 if (index
== ra
->start
&& ra
->size
== ra
->async_size
) {
538 add_pages
= get_next_ra_size(ra
, max_pages
);
539 if (ra
->size
+ add_pages
<= max_pages
) {
540 ra
->async_size
= add_pages
;
541 ra
->size
+= add_pages
;
543 ra
->size
= max_pages
;
544 ra
->async_size
= max_pages
>> 1;
548 ractl
->_index
= ra
->start
;
549 do_page_cache_ra(ractl
, ra
->size
, ra
->async_size
);
552 void page_cache_sync_ra(struct readahead_control
*ractl
,
553 struct file_ra_state
*ra
, unsigned long req_count
)
555 bool do_forced_ra
= ractl
->file
&& (ractl
->file
->f_mode
& FMODE_RANDOM
);
558 * Even if read-ahead is disabled, issue this request as read-ahead
559 * as we'll need it to satisfy the requested range. The forced
560 * read-ahead will do the right thing and limit the read to just the
561 * requested range, which we'll set to 1 page for this case.
563 if (!ra
->ra_pages
|| blk_cgroup_congested()) {
572 force_page_cache_ra(ractl
, ra
, req_count
);
577 ondemand_readahead(ractl
, ra
, false, req_count
);
579 EXPORT_SYMBOL_GPL(page_cache_sync_ra
);
581 void page_cache_async_ra(struct readahead_control
*ractl
,
582 struct file_ra_state
*ra
, struct page
*page
,
583 unsigned long req_count
)
590 * Same bit is used for PG_readahead and PG_reclaim.
592 if (PageWriteback(page
))
595 ClearPageReadahead(page
);
598 * Defer asynchronous read-ahead on IO congestion.
600 if (inode_read_congested(ractl
->mapping
->host
))
603 if (blk_cgroup_congested())
607 ondemand_readahead(ractl
, ra
, true, req_count
);
609 EXPORT_SYMBOL_GPL(page_cache_async_ra
);
611 ssize_t
ksys_readahead(int fd
, loff_t offset
, size_t count
)
618 if (!f
.file
|| !(f
.file
->f_mode
& FMODE_READ
))
622 * The readahead() syscall is intended to run only on files
623 * that can execute readahead. If readahead is not possible
624 * on this file, then we must return -EINVAL.
627 if (!f
.file
->f_mapping
|| !f
.file
->f_mapping
->a_ops
||
628 !S_ISREG(file_inode(f
.file
)->i_mode
))
631 ret
= vfs_fadvise(f
.file
, offset
, count
, POSIX_FADV_WILLNEED
);
637 SYSCALL_DEFINE3(readahead
, int, fd
, loff_t
, offset
, size_t, count
)
639 return ksys_readahead(fd
, offset
, count
);