xref: /openbmc/linux/mm/readahead.c (revision fe4d0d5d)
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 ->readpage() 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 ->readpage() 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  * ->readpage() which may be less efficient.
114  */
115 
116 #include <linux/kernel.h>
117 #include <linux/dax.h>
118 #include <linux/gfp.h>
119 #include <linux/export.h>
120 #include <linux/backing-dev.h>
121 #include <linux/task_io_accounting_ops.h>
122 #include <linux/pagevec.h>
123 #include <linux/pagemap.h>
124 #include <linux/syscalls.h>
125 #include <linux/file.h>
126 #include <linux/mm_inline.h>
127 #include <linux/blk-cgroup.h>
128 #include <linux/fadvise.h>
129 #include <linux/sched/mm.h>
130 
131 #include "internal.h"
132 
133 /*
134  * Initialise a struct file's readahead state.  Assumes that the caller has
135  * memset *ra to zero.
136  */
137 void
138 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
139 {
140 	ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
141 	ra->prev_pos = -1;
142 }
143 EXPORT_SYMBOL_GPL(file_ra_state_init);
144 
145 static void read_pages(struct readahead_control *rac)
146 {
147 	const struct address_space_operations *aops = rac->mapping->a_ops;
148 	struct page *page;
149 	struct blk_plug plug;
150 
151 	if (!readahead_count(rac))
152 		return;
153 
154 	blk_start_plug(&plug);
155 
156 	if (aops->readahead) {
157 		aops->readahead(rac);
158 		/*
159 		 * Clean up the remaining pages.  The sizes in ->ra
160 		 * may be used to size the next readahead, so make sure
161 		 * they accurately reflect what happened.
162 		 */
163 		while ((page = readahead_page(rac))) {
164 			rac->ra->size -= 1;
165 			if (rac->ra->async_size > 0) {
166 				rac->ra->async_size -= 1;
167 				delete_from_page_cache(page);
168 			}
169 			unlock_page(page);
170 			put_page(page);
171 		}
172 	} else {
173 		while ((page = readahead_page(rac))) {
174 			aops->readpage(rac->file, page);
175 			put_page(page);
176 		}
177 	}
178 
179 	blk_finish_plug(&plug);
180 
181 	BUG_ON(readahead_count(rac));
182 }
183 
184 /**
185  * page_cache_ra_unbounded - Start unchecked readahead.
186  * @ractl: Readahead control.
187  * @nr_to_read: The number of pages to read.
188  * @lookahead_size: Where to start the next readahead.
189  *
190  * This function is for filesystems to call when they want to start
191  * readahead beyond a file's stated i_size.  This is almost certainly
192  * not the function you want to call.  Use page_cache_async_readahead()
193  * or page_cache_sync_readahead() instead.
194  *
195  * Context: File is referenced by caller.  Mutexes may be held by caller.
196  * May sleep, but will not reenter filesystem to reclaim memory.
197  */
198 void page_cache_ra_unbounded(struct readahead_control *ractl,
199 		unsigned long nr_to_read, unsigned long lookahead_size)
200 {
201 	struct address_space *mapping = ractl->mapping;
202 	unsigned long index = readahead_index(ractl);
203 	gfp_t gfp_mask = readahead_gfp_mask(mapping);
204 	unsigned long i;
205 
206 	/*
207 	 * Partway through the readahead operation, we will have added
208 	 * locked pages to the page cache, but will not yet have submitted
209 	 * them for I/O.  Adding another page may need to allocate memory,
210 	 * which can trigger memory reclaim.  Telling the VM we're in
211 	 * the middle of a filesystem operation will cause it to not
212 	 * touch file-backed pages, preventing a deadlock.  Most (all?)
213 	 * filesystems already specify __GFP_NOFS in their mapping's
214 	 * gfp_mask, but let's be explicit here.
215 	 */
216 	unsigned int nofs = memalloc_nofs_save();
217 
218 	filemap_invalidate_lock_shared(mapping);
219 	/*
220 	 * Preallocate as many pages as we will need.
221 	 */
222 	for (i = 0; i < nr_to_read; i++) {
223 		struct folio *folio = xa_load(&mapping->i_pages, index + i);
224 
225 		if (folio && !xa_is_value(folio)) {
226 			/*
227 			 * Page already present?  Kick off the current batch
228 			 * of contiguous pages before continuing with the
229 			 * next batch.  This page may be the one we would
230 			 * have intended to mark as Readahead, but we don't
231 			 * have a stable reference to this page, and it's
232 			 * not worth getting one just for that.
233 			 */
234 			read_pages(ractl);
235 			ractl->_index++;
236 			i = ractl->_index + ractl->_nr_pages - index - 1;
237 			continue;
238 		}
239 
240 		folio = filemap_alloc_folio(gfp_mask, 0);
241 		if (!folio)
242 			break;
243 		if (filemap_add_folio(mapping, folio, index + i,
244 					gfp_mask) < 0) {
245 			folio_put(folio);
246 			read_pages(ractl);
247 			ractl->_index++;
248 			i = ractl->_index + ractl->_nr_pages - index - 1;
249 			continue;
250 		}
251 		if (i == nr_to_read - lookahead_size)
252 			folio_set_readahead(folio);
253 		ractl->_nr_pages++;
254 	}
255 
256 	/*
257 	 * Now start the IO.  We ignore I/O errors - if the page is not
258 	 * uptodate then the caller will launch readpage again, and
259 	 * will then handle the error.
260 	 */
261 	read_pages(ractl);
262 	filemap_invalidate_unlock_shared(mapping);
263 	memalloc_nofs_restore(nofs);
264 }
265 EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
266 
267 /*
268  * do_page_cache_ra() actually reads a chunk of disk.  It allocates
269  * the pages first, then submits them for I/O. This avoids the very bad
270  * behaviour which would occur if page allocations are causing VM writeback.
271  * We really don't want to intermingle reads and writes like that.
272  */
273 static void do_page_cache_ra(struct readahead_control *ractl,
274 		unsigned long nr_to_read, unsigned long lookahead_size)
275 {
276 	struct inode *inode = ractl->mapping->host;
277 	unsigned long index = readahead_index(ractl);
278 	loff_t isize = i_size_read(inode);
279 	pgoff_t end_index;	/* The last page we want to read */
280 
281 	if (isize == 0)
282 		return;
283 
284 	end_index = (isize - 1) >> PAGE_SHIFT;
285 	if (index > end_index)
286 		return;
287 	/* Don't read past the page containing the last byte of the file */
288 	if (nr_to_read > end_index - index)
289 		nr_to_read = end_index - index + 1;
290 
291 	page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
292 }
293 
294 /*
295  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
296  * memory at once.
297  */
298 void force_page_cache_ra(struct readahead_control *ractl,
299 		unsigned long nr_to_read)
300 {
301 	struct address_space *mapping = ractl->mapping;
302 	struct file_ra_state *ra = ractl->ra;
303 	struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
304 	unsigned long max_pages, index;
305 
306 	if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readahead))
307 		return;
308 
309 	/*
310 	 * If the request exceeds the readahead window, allow the read to
311 	 * be up to the optimal hardware IO size
312 	 */
313 	index = readahead_index(ractl);
314 	max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
315 	nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
316 	while (nr_to_read) {
317 		unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
318 
319 		if (this_chunk > nr_to_read)
320 			this_chunk = nr_to_read;
321 		ractl->_index = index;
322 		do_page_cache_ra(ractl, this_chunk, 0);
323 
324 		index += this_chunk;
325 		nr_to_read -= this_chunk;
326 	}
327 }
328 
329 /*
330  * Set the initial window size, round to next power of 2 and square
331  * for small size, x 4 for medium, and x 2 for large
332  * for 128k (32 page) max ra
333  * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
334  */
335 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
336 {
337 	unsigned long newsize = roundup_pow_of_two(size);
338 
339 	if (newsize <= max / 32)
340 		newsize = newsize * 4;
341 	else if (newsize <= max / 4)
342 		newsize = newsize * 2;
343 	else
344 		newsize = max;
345 
346 	return newsize;
347 }
348 
349 /*
350  *  Get the previous window size, ramp it up, and
351  *  return it as the new window size.
352  */
353 static unsigned long get_next_ra_size(struct file_ra_state *ra,
354 				      unsigned long max)
355 {
356 	unsigned long cur = ra->size;
357 
358 	if (cur < max / 16)
359 		return 4 * cur;
360 	if (cur <= max / 2)
361 		return 2 * cur;
362 	return max;
363 }
364 
365 /*
366  * On-demand readahead design.
367  *
368  * The fields in struct file_ra_state represent the most-recently-executed
369  * readahead attempt:
370  *
371  *                        |<----- async_size ---------|
372  *     |------------------- size -------------------->|
373  *     |==================#===========================|
374  *     ^start             ^page marked with PG_readahead
375  *
376  * To overlap application thinking time and disk I/O time, we do
377  * `readahead pipelining': Do not wait until the application consumed all
378  * readahead pages and stalled on the missing page at readahead_index;
379  * Instead, submit an asynchronous readahead I/O as soon as there are
380  * only async_size pages left in the readahead window. Normally async_size
381  * will be equal to size, for maximum pipelining.
382  *
383  * In interleaved sequential reads, concurrent streams on the same fd can
384  * be invalidating each other's readahead state. So we flag the new readahead
385  * page at (start+size-async_size) with PG_readahead, and use it as readahead
386  * indicator. The flag won't be set on already cached pages, to avoid the
387  * readahead-for-nothing fuss, saving pointless page cache lookups.
388  *
389  * prev_pos tracks the last visited byte in the _previous_ read request.
390  * It should be maintained by the caller, and will be used for detecting
391  * small random reads. Note that the readahead algorithm checks loosely
392  * for sequential patterns. Hence interleaved reads might be served as
393  * sequential ones.
394  *
395  * There is a special-case: if the first page which the application tries to
396  * read happens to be the first page of the file, it is assumed that a linear
397  * read is about to happen and the window is immediately set to the initial size
398  * based on I/O request size and the max_readahead.
399  *
400  * The code ramps up the readahead size aggressively at first, but slow down as
401  * it approaches max_readhead.
402  */
403 
404 /*
405  * Count contiguously cached pages from @index-1 to @index-@max,
406  * this count is a conservative estimation of
407  * 	- length of the sequential read sequence, or
408  * 	- thrashing threshold in memory tight systems
409  */
410 static pgoff_t count_history_pages(struct address_space *mapping,
411 				   pgoff_t index, unsigned long max)
412 {
413 	pgoff_t head;
414 
415 	rcu_read_lock();
416 	head = page_cache_prev_miss(mapping, index - 1, max);
417 	rcu_read_unlock();
418 
419 	return index - 1 - head;
420 }
421 
422 /*
423  * page cache context based readahead
424  */
425 static int try_context_readahead(struct address_space *mapping,
426 				 struct file_ra_state *ra,
427 				 pgoff_t index,
428 				 unsigned long req_size,
429 				 unsigned long max)
430 {
431 	pgoff_t size;
432 
433 	size = count_history_pages(mapping, index, max);
434 
435 	/*
436 	 * not enough history pages:
437 	 * it could be a random read
438 	 */
439 	if (size <= req_size)
440 		return 0;
441 
442 	/*
443 	 * starts from beginning of file:
444 	 * it is a strong indication of long-run stream (or whole-file-read)
445 	 */
446 	if (size >= index)
447 		size *= 2;
448 
449 	ra->start = index;
450 	ra->size = min(size + req_size, max);
451 	ra->async_size = 1;
452 
453 	return 1;
454 }
455 
456 /*
457  * There are some parts of the kernel which assume that PMD entries
458  * are exactly HPAGE_PMD_ORDER.  Those should be fixed, but until then,
459  * limit the maximum allocation order to PMD size.  I'm not aware of any
460  * assumptions about maximum order if THP are disabled, but 8 seems like
461  * a good order (that's 1MB if you're using 4kB pages)
462  */
463 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
464 #define MAX_PAGECACHE_ORDER	HPAGE_PMD_ORDER
465 #else
466 #define MAX_PAGECACHE_ORDER	8
467 #endif
468 
469 static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
470 		pgoff_t mark, unsigned int order, gfp_t gfp)
471 {
472 	int err;
473 	struct folio *folio = filemap_alloc_folio(gfp, order);
474 
475 	if (!folio)
476 		return -ENOMEM;
477 	if (mark - index < (1UL << order))
478 		folio_set_readahead(folio);
479 	err = filemap_add_folio(ractl->mapping, folio, index, gfp);
480 	if (err)
481 		folio_put(folio);
482 	else
483 		ractl->_nr_pages += 1UL << order;
484 	return err;
485 }
486 
487 void page_cache_ra_order(struct readahead_control *ractl,
488 		struct file_ra_state *ra, unsigned int new_order)
489 {
490 	struct address_space *mapping = ractl->mapping;
491 	pgoff_t index = readahead_index(ractl);
492 	pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
493 	pgoff_t mark = index + ra->size - ra->async_size;
494 	int err = 0;
495 	gfp_t gfp = readahead_gfp_mask(mapping);
496 
497 	if (!mapping_large_folio_support(mapping) || ra->size < 4)
498 		goto fallback;
499 
500 	limit = min(limit, index + ra->size - 1);
501 
502 	if (new_order < MAX_PAGECACHE_ORDER) {
503 		new_order += 2;
504 		if (new_order > MAX_PAGECACHE_ORDER)
505 			new_order = MAX_PAGECACHE_ORDER;
506 		while ((1 << new_order) > ra->size)
507 			new_order--;
508 	}
509 
510 	while (index <= limit) {
511 		unsigned int order = new_order;
512 
513 		/* Align with smaller pages if needed */
514 		if (index & ((1UL << order) - 1)) {
515 			order = __ffs(index);
516 			if (order == 1)
517 				order = 0;
518 		}
519 		/* Don't allocate pages past EOF */
520 		while (index + (1UL << order) - 1 > limit) {
521 			if (--order == 1)
522 				order = 0;
523 		}
524 		err = ra_alloc_folio(ractl, index, mark, order, gfp);
525 		if (err)
526 			break;
527 		index += 1UL << order;
528 	}
529 
530 	if (index > limit) {
531 		ra->size += index - limit - 1;
532 		ra->async_size += index - limit - 1;
533 	}
534 
535 	read_pages(ractl);
536 
537 	/*
538 	 * If there were already pages in the page cache, then we may have
539 	 * left some gaps.  Let the regular readahead code take care of this
540 	 * situation.
541 	 */
542 	if (!err)
543 		return;
544 fallback:
545 	do_page_cache_ra(ractl, ra->size, ra->async_size);
546 }
547 
548 /*
549  * A minimal readahead algorithm for trivial sequential/random reads.
550  */
551 static void ondemand_readahead(struct readahead_control *ractl,
552 		struct folio *folio, unsigned long req_size)
553 {
554 	struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
555 	struct file_ra_state *ra = ractl->ra;
556 	unsigned long max_pages = ra->ra_pages;
557 	unsigned long add_pages;
558 	unsigned long index = readahead_index(ractl);
559 	pgoff_t prev_index;
560 
561 	/*
562 	 * If the request exceeds the readahead window, allow the read to
563 	 * be up to the optimal hardware IO size
564 	 */
565 	if (req_size > max_pages && bdi->io_pages > max_pages)
566 		max_pages = min(req_size, bdi->io_pages);
567 
568 	/*
569 	 * start of file
570 	 */
571 	if (!index)
572 		goto initial_readahead;
573 
574 	/*
575 	 * It's the expected callback index, assume sequential access.
576 	 * Ramp up sizes, and push forward the readahead window.
577 	 */
578 	if ((index == (ra->start + ra->size - ra->async_size) ||
579 	     index == (ra->start + ra->size))) {
580 		ra->start += ra->size;
581 		ra->size = get_next_ra_size(ra, max_pages);
582 		ra->async_size = ra->size;
583 		goto readit;
584 	}
585 
586 	/*
587 	 * Hit a marked folio without valid readahead state.
588 	 * E.g. interleaved reads.
589 	 * Query the pagecache for async_size, which normally equals to
590 	 * readahead size. Ramp it up and use it as the new readahead size.
591 	 */
592 	if (folio) {
593 		pgoff_t start;
594 
595 		rcu_read_lock();
596 		start = page_cache_next_miss(ractl->mapping, index + 1,
597 				max_pages);
598 		rcu_read_unlock();
599 
600 		if (!start || start - index > max_pages)
601 			return;
602 
603 		ra->start = start;
604 		ra->size = start - index;	/* old async_size */
605 		ra->size += req_size;
606 		ra->size = get_next_ra_size(ra, max_pages);
607 		ra->async_size = ra->size;
608 		goto readit;
609 	}
610 
611 	/*
612 	 * oversize read
613 	 */
614 	if (req_size > max_pages)
615 		goto initial_readahead;
616 
617 	/*
618 	 * sequential cache miss
619 	 * trivial case: (index - prev_index) == 1
620 	 * unaligned reads: (index - prev_index) == 0
621 	 */
622 	prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
623 	if (index - prev_index <= 1UL)
624 		goto initial_readahead;
625 
626 	/*
627 	 * Query the page cache and look for the traces(cached history pages)
628 	 * that a sequential stream would leave behind.
629 	 */
630 	if (try_context_readahead(ractl->mapping, ra, index, req_size,
631 			max_pages))
632 		goto readit;
633 
634 	/*
635 	 * standalone, small random read
636 	 * Read as is, and do not pollute the readahead state.
637 	 */
638 	do_page_cache_ra(ractl, req_size, 0);
639 	return;
640 
641 initial_readahead:
642 	ra->start = index;
643 	ra->size = get_init_ra_size(req_size, max_pages);
644 	ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
645 
646 readit:
647 	/*
648 	 * Will this read hit the readahead marker made by itself?
649 	 * If so, trigger the readahead marker hit now, and merge
650 	 * the resulted next readahead window into the current one.
651 	 * Take care of maximum IO pages as above.
652 	 */
653 	if (index == ra->start && ra->size == ra->async_size) {
654 		add_pages = get_next_ra_size(ra, max_pages);
655 		if (ra->size + add_pages <= max_pages) {
656 			ra->async_size = add_pages;
657 			ra->size += add_pages;
658 		} else {
659 			ra->size = max_pages;
660 			ra->async_size = max_pages >> 1;
661 		}
662 	}
663 
664 	ractl->_index = ra->start;
665 	page_cache_ra_order(ractl, ra, folio ? folio_order(folio) : 0);
666 }
667 
668 void page_cache_sync_ra(struct readahead_control *ractl,
669 		unsigned long req_count)
670 {
671 	bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
672 
673 	/*
674 	 * Even if readahead is disabled, issue this request as readahead
675 	 * as we'll need it to satisfy the requested range. The forced
676 	 * readahead will do the right thing and limit the read to just the
677 	 * requested range, which we'll set to 1 page for this case.
678 	 */
679 	if (!ractl->ra->ra_pages || blk_cgroup_congested()) {
680 		if (!ractl->file)
681 			return;
682 		req_count = 1;
683 		do_forced_ra = true;
684 	}
685 
686 	/* be dumb */
687 	if (do_forced_ra) {
688 		force_page_cache_ra(ractl, req_count);
689 		return;
690 	}
691 
692 	ondemand_readahead(ractl, NULL, req_count);
693 }
694 EXPORT_SYMBOL_GPL(page_cache_sync_ra);
695 
696 void page_cache_async_ra(struct readahead_control *ractl,
697 		struct folio *folio, unsigned long req_count)
698 {
699 	/* no readahead */
700 	if (!ractl->ra->ra_pages)
701 		return;
702 
703 	/*
704 	 * Same bit is used for PG_readahead and PG_reclaim.
705 	 */
706 	if (folio_test_writeback(folio))
707 		return;
708 
709 	folio_clear_readahead(folio);
710 
711 	if (blk_cgroup_congested())
712 		return;
713 
714 	ondemand_readahead(ractl, folio, req_count);
715 }
716 EXPORT_SYMBOL_GPL(page_cache_async_ra);
717 
718 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
719 {
720 	ssize_t ret;
721 	struct fd f;
722 
723 	ret = -EBADF;
724 	f = fdget(fd);
725 	if (!f.file || !(f.file->f_mode & FMODE_READ))
726 		goto out;
727 
728 	/*
729 	 * The readahead() syscall is intended to run only on files
730 	 * that can execute readahead. If readahead is not possible
731 	 * on this file, then we must return -EINVAL.
732 	 */
733 	ret = -EINVAL;
734 	if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
735 	    !S_ISREG(file_inode(f.file)->i_mode))
736 		goto out;
737 
738 	ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
739 out:
740 	fdput(f);
741 	return ret;
742 }
743 
744 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
745 {
746 	return ksys_readahead(fd, offset, count);
747 }
748 
749 /**
750  * readahead_expand - Expand a readahead request
751  * @ractl: The request to be expanded
752  * @new_start: The revised start
753  * @new_len: The revised size of the request
754  *
755  * Attempt to expand a readahead request outwards from the current size to the
756  * specified size by inserting locked pages before and after the current window
757  * to increase the size to the new window.  This may involve the insertion of
758  * THPs, in which case the window may get expanded even beyond what was
759  * requested.
760  *
761  * The algorithm will stop if it encounters a conflicting page already in the
762  * pagecache and leave a smaller expansion than requested.
763  *
764  * The caller must check for this by examining the revised @ractl object for a
765  * different expansion than was requested.
766  */
767 void readahead_expand(struct readahead_control *ractl,
768 		      loff_t new_start, size_t new_len)
769 {
770 	struct address_space *mapping = ractl->mapping;
771 	struct file_ra_state *ra = ractl->ra;
772 	pgoff_t new_index, new_nr_pages;
773 	gfp_t gfp_mask = readahead_gfp_mask(mapping);
774 
775 	new_index = new_start / PAGE_SIZE;
776 
777 	/* Expand the leading edge downwards */
778 	while (ractl->_index > new_index) {
779 		unsigned long index = ractl->_index - 1;
780 		struct page *page = xa_load(&mapping->i_pages, index);
781 
782 		if (page && !xa_is_value(page))
783 			return; /* Page apparently present */
784 
785 		page = __page_cache_alloc(gfp_mask);
786 		if (!page)
787 			return;
788 		if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
789 			put_page(page);
790 			return;
791 		}
792 
793 		ractl->_nr_pages++;
794 		ractl->_index = page->index;
795 	}
796 
797 	new_len += new_start - readahead_pos(ractl);
798 	new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
799 
800 	/* Expand the trailing edge upwards */
801 	while (ractl->_nr_pages < new_nr_pages) {
802 		unsigned long index = ractl->_index + ractl->_nr_pages;
803 		struct page *page = xa_load(&mapping->i_pages, index);
804 
805 		if (page && !xa_is_value(page))
806 			return; /* Page apparently present */
807 
808 		page = __page_cache_alloc(gfp_mask);
809 		if (!page)
810 			return;
811 		if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) {
812 			put_page(page);
813 			return;
814 		}
815 		ractl->_nr_pages++;
816 		if (ra) {
817 			ra->size++;
818 			ra->async_size++;
819 		}
820 	}
821 }
822 EXPORT_SYMBOL(readahead_expand);
823