xref: /openbmc/linux/mm/readahead.c (revision ca1cfc3f)
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 #include <linux/sched/mm.h>
26 
27 #include "internal.h"
28 
29 /*
30  * Initialise a struct file's readahead state.  Assumes that the caller has
31  * memset *ra to zero.
32  */
33 void
34 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
35 {
36 	ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
37 	ra->prev_pos = -1;
38 }
39 EXPORT_SYMBOL_GPL(file_ra_state_init);
40 
41 /*
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
46  *   an error
47  */
48 static void read_cache_pages_invalidate_page(struct address_space *mapping,
49 					     struct page *page)
50 {
51 	if (page_has_private(page)) {
52 		if (!trylock_page(page))
53 			BUG();
54 		page->mapping = mapping;
55 		do_invalidatepage(page, 0, PAGE_SIZE);
56 		page->mapping = NULL;
57 		unlock_page(page);
58 	}
59 	put_page(page);
60 }
61 
62 /*
63  * release a list of pages, invalidating them first if need be
64  */
65 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
66 					      struct list_head *pages)
67 {
68 	struct page *victim;
69 
70 	while (!list_empty(pages)) {
71 		victim = lru_to_page(pages);
72 		list_del(&victim->lru);
73 		read_cache_pages_invalidate_page(mapping, victim);
74 	}
75 }
76 
77 /**
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.
84  *
85  * Hides the details of the LRU cache etc from the filesystems.
86  *
87  * Returns: %0 on success, error return by @filler otherwise
88  */
89 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
90 			int (*filler)(void *, struct page *), void *data)
91 {
92 	struct page *page;
93 	int ret = 0;
94 
95 	while (!list_empty(pages)) {
96 		page = lru_to_page(pages);
97 		list_del(&page->lru);
98 		if (add_to_page_cache_lru(page, mapping, page->index,
99 				readahead_gfp_mask(mapping))) {
100 			read_cache_pages_invalidate_page(mapping, page);
101 			continue;
102 		}
103 		put_page(page);
104 
105 		ret = filler(data, page);
106 		if (unlikely(ret)) {
107 			read_cache_pages_invalidate_pages(mapping, pages);
108 			break;
109 		}
110 		task_io_account_read(PAGE_SIZE);
111 	}
112 	return ret;
113 }
114 
115 EXPORT_SYMBOL(read_cache_pages);
116 
117 static void read_pages(struct readahead_control *rac, struct list_head *pages,
118 		bool skip_page)
119 {
120 	const struct address_space_operations *aops = rac->mapping->a_ops;
121 	struct page *page;
122 	struct blk_plug plug;
123 
124 	if (!readahead_count(rac))
125 		goto out;
126 
127 	blk_start_plug(&plug);
128 
129 	if (aops->readahead) {
130 		aops->readahead(rac);
131 		/* Clean up the remaining pages */
132 		while ((page = readahead_page(rac))) {
133 			unlock_page(page);
134 			put_page(page);
135 		}
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;
142 		rac->_nr_pages = 0;
143 	} else {
144 		while ((page = readahead_page(rac))) {
145 			aops->readpage(rac->file, page);
146 			put_page(page);
147 		}
148 	}
149 
150 	blk_finish_plug(&plug);
151 
152 	BUG_ON(!list_empty(pages));
153 	BUG_ON(readahead_count(rac));
154 
155 out:
156 	if (skip_page)
157 		rac->_index++;
158 }
159 
160 /**
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.
165  *
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.
170  *
171  * Context: File is referenced by caller.  Mutexes may be held by caller.
172  * May sleep, but will not reenter filesystem to reclaim memory.
173  */
174 void page_cache_ra_unbounded(struct readahead_control *ractl,
175 		unsigned long nr_to_read, unsigned long lookahead_size)
176 {
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);
181 	unsigned long i;
182 
183 	/*
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.
192 	 */
193 	unsigned int nofs = memalloc_nofs_save();
194 
195 	/*
196 	 * Preallocate as many pages as we will need.
197 	 */
198 	for (i = 0; i < nr_to_read; i++) {
199 		struct page *page = xa_load(&mapping->i_pages, index + i);
200 
201 		BUG_ON(index + i != ractl->_index + ractl->_nr_pages);
202 
203 		if (page && !xa_is_value(page)) {
204 			/*
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.
211 			 */
212 			read_pages(ractl, &page_pool, true);
213 			continue;
214 		}
215 
216 		page = __page_cache_alloc(gfp_mask);
217 		if (!page)
218 			break;
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,
223 					gfp_mask) < 0) {
224 			put_page(page);
225 			read_pages(ractl, &page_pool, true);
226 			continue;
227 		}
228 		if (i == nr_to_read - lookahead_size)
229 			SetPageReadahead(page);
230 		ractl->_nr_pages++;
231 	}
232 
233 	/*
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.
237 	 */
238 	read_pages(ractl, &page_pool, false);
239 	memalloc_nofs_restore(nofs);
240 }
241 EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
242 
243 /*
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.
248  */
249 void do_page_cache_ra(struct readahead_control *ractl,
250 		unsigned long nr_to_read, unsigned long lookahead_size)
251 {
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 */
256 
257 	if (isize == 0)
258 		return;
259 
260 	end_index = (isize - 1) >> PAGE_SHIFT;
261 	if (index > end_index)
262 		return;
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;
266 
267 	page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
268 }
269 
270 /*
271  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
272  * memory at once.
273  */
274 void force_page_cache_ra(struct readahead_control *ractl,
275 		struct file_ra_state *ra, unsigned long nr_to_read)
276 {
277 	struct address_space *mapping = ractl->mapping;
278 	struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
279 	unsigned long max_pages, index;
280 
281 	if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages &&
282 			!mapping->a_ops->readahead))
283 		return;
284 
285 	/*
286 	 * If the request exceeds the readahead window, allow the read to
287 	 * be up to the optimal hardware IO size
288 	 */
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);
292 	while (nr_to_read) {
293 		unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
294 
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);
299 
300 		index += this_chunk;
301 		nr_to_read -= this_chunk;
302 	}
303 }
304 
305 /*
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
310  */
311 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
312 {
313 	unsigned long newsize = roundup_pow_of_two(size);
314 
315 	if (newsize <= max / 32)
316 		newsize = newsize * 4;
317 	else if (newsize <= max / 4)
318 		newsize = newsize * 2;
319 	else
320 		newsize = max;
321 
322 	return newsize;
323 }
324 
325 /*
326  *  Get the previous window size, ramp it up, and
327  *  return it as the new window size.
328  */
329 static unsigned long get_next_ra_size(struct file_ra_state *ra,
330 				      unsigned long max)
331 {
332 	unsigned long cur = ra->size;
333 
334 	if (cur < max / 16)
335 		return 4 * cur;
336 	if (cur <= max / 2)
337 		return 2 * cur;
338 	return max;
339 }
340 
341 /*
342  * On-demand readahead design.
343  *
344  * The fields in struct file_ra_state represent the most-recently-executed
345  * readahead attempt:
346  *
347  *                        |<----- async_size ---------|
348  *     |------------------- size -------------------->|
349  *     |==================#===========================|
350  *     ^start             ^page marked with PG_readahead
351  *
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.
358  *
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.
364  *
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
369  * sequential ones.
370  *
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.
375  *
376  * The code ramps up the readahead size aggressively at first, but slow down as
377  * it approaches max_readhead.
378  */
379 
380 /*
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
385  */
386 static pgoff_t count_history_pages(struct address_space *mapping,
387 				   pgoff_t index, unsigned long max)
388 {
389 	pgoff_t head;
390 
391 	rcu_read_lock();
392 	head = page_cache_prev_miss(mapping, index - 1, max);
393 	rcu_read_unlock();
394 
395 	return index - 1 - head;
396 }
397 
398 /*
399  * page cache context based read-ahead
400  */
401 static int try_context_readahead(struct address_space *mapping,
402 				 struct file_ra_state *ra,
403 				 pgoff_t index,
404 				 unsigned long req_size,
405 				 unsigned long max)
406 {
407 	pgoff_t size;
408 
409 	size = count_history_pages(mapping, index, max);
410 
411 	/*
412 	 * not enough history pages:
413 	 * it could be a random read
414 	 */
415 	if (size <= req_size)
416 		return 0;
417 
418 	/*
419 	 * starts from beginning of file:
420 	 * it is a strong indication of long-run stream (or whole-file-read)
421 	 */
422 	if (size >= index)
423 		size *= 2;
424 
425 	ra->start = index;
426 	ra->size = min(size + req_size, max);
427 	ra->async_size = 1;
428 
429 	return 1;
430 }
431 
432 /*
433  * A minimal readahead algorithm for trivial sequential/random reads.
434  */
435 static void ondemand_readahead(struct readahead_control *ractl,
436 		struct file_ra_state *ra, bool hit_readahead_marker,
437 		unsigned long req_size)
438 {
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);
443 	pgoff_t prev_index;
444 
445 	/*
446 	 * If the request exceeds the readahead window, allow the read to
447 	 * be up to the optimal hardware IO size
448 	 */
449 	if (req_size > max_pages && bdi->io_pages > max_pages)
450 		max_pages = min(req_size, bdi->io_pages);
451 
452 	/*
453 	 * start of file
454 	 */
455 	if (!index)
456 		goto initial_readahead;
457 
458 	/*
459 	 * It's the expected callback index, assume sequential access.
460 	 * Ramp up sizes, and push forward the readahead window.
461 	 */
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;
467 		goto readit;
468 	}
469 
470 	/*
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.
475 	 */
476 	if (hit_readahead_marker) {
477 		pgoff_t start;
478 
479 		rcu_read_lock();
480 		start = page_cache_next_miss(ractl->mapping, index + 1,
481 				max_pages);
482 		rcu_read_unlock();
483 
484 		if (!start || start - index > max_pages)
485 			return;
486 
487 		ra->start = start;
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;
492 		goto readit;
493 	}
494 
495 	/*
496 	 * oversize read
497 	 */
498 	if (req_size > max_pages)
499 		goto initial_readahead;
500 
501 	/*
502 	 * sequential cache miss
503 	 * trivial case: (index - prev_index) == 1
504 	 * unaligned reads: (index - prev_index) == 0
505 	 */
506 	prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
507 	if (index - prev_index <= 1UL)
508 		goto initial_readahead;
509 
510 	/*
511 	 * Query the page cache and look for the traces(cached history pages)
512 	 * that a sequential stream would leave behind.
513 	 */
514 	if (try_context_readahead(ractl->mapping, ra, index, req_size,
515 			max_pages))
516 		goto readit;
517 
518 	/*
519 	 * standalone, small random read
520 	 * Read as is, and do not pollute the readahead state.
521 	 */
522 	do_page_cache_ra(ractl, req_size, 0);
523 	return;
524 
525 initial_readahead:
526 	ra->start = index;
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;
529 
530 readit:
531 	/*
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.
536 	 */
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;
542 		} else {
543 			ra->size = max_pages;
544 			ra->async_size = max_pages >> 1;
545 		}
546 	}
547 
548 	ractl->_index = ra->start;
549 	do_page_cache_ra(ractl, ra->size, ra->async_size);
550 }
551 
552 void page_cache_sync_ra(struct readahead_control *ractl,
553 		struct file_ra_state *ra, unsigned long req_count)
554 {
555 	bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
556 
557 	/*
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.
562 	 */
563 	if (!ra->ra_pages || blk_cgroup_congested()) {
564 		if (!ractl->file)
565 			return;
566 		req_count = 1;
567 		do_forced_ra = true;
568 	}
569 
570 	/* be dumb */
571 	if (do_forced_ra) {
572 		force_page_cache_ra(ractl, ra, req_count);
573 		return;
574 	}
575 
576 	/* do read-ahead */
577 	ondemand_readahead(ractl, ra, false, req_count);
578 }
579 EXPORT_SYMBOL_GPL(page_cache_sync_ra);
580 
581 void page_cache_async_ra(struct readahead_control *ractl,
582 		struct file_ra_state *ra, struct page *page,
583 		unsigned long req_count)
584 {
585 	/* no read-ahead */
586 	if (!ra->ra_pages)
587 		return;
588 
589 	/*
590 	 * Same bit is used for PG_readahead and PG_reclaim.
591 	 */
592 	if (PageWriteback(page))
593 		return;
594 
595 	ClearPageReadahead(page);
596 
597 	/*
598 	 * Defer asynchronous read-ahead on IO congestion.
599 	 */
600 	if (inode_read_congested(ractl->mapping->host))
601 		return;
602 
603 	if (blk_cgroup_congested())
604 		return;
605 
606 	/* do read-ahead */
607 	ondemand_readahead(ractl, ra, true, req_count);
608 }
609 EXPORT_SYMBOL_GPL(page_cache_async_ra);
610 
611 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
612 {
613 	ssize_t ret;
614 	struct fd f;
615 
616 	ret = -EBADF;
617 	f = fdget(fd);
618 	if (!f.file || !(f.file->f_mode & FMODE_READ))
619 		goto out;
620 
621 	/*
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.
625 	 */
626 	ret = -EINVAL;
627 	if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
628 	    !S_ISREG(file_inode(f.file)->i_mode))
629 		goto out;
630 
631 	ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
632 out:
633 	fdput(f);
634 	return ret;
635 }
636 
637 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
638 {
639 	return ksys_readahead(fd, offset, count);
640 }
641