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