xref: /linux/mm/readahead.c (revision 0dd9ac63ce26ec87b080ca9c3e6efed33c23ace6)
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/fs.h>
12 #include <linux/gfp.h>
13 #include <linux/mm.h>
14 #include <linux/module.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 
21 /*
22  * Initialise a struct file's readahead state.  Assumes that the caller has
23  * memset *ra to zero.
24  */
25 void
26 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
27 {
28 	ra->ra_pages = mapping->backing_dev_info->ra_pages;
29 	ra->prev_pos = -1;
30 }
31 EXPORT_SYMBOL_GPL(file_ra_state_init);
32 
33 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
34 
35 /*
36  * see if a page needs releasing upon read_cache_pages() failure
37  * - the caller of read_cache_pages() may have set PG_private or PG_fscache
38  *   before calling, such as the NFS fs marking pages that are cached locally
39  *   on disk, thus we need to give the fs a chance to clean up in the event of
40  *   an error
41  */
42 static void read_cache_pages_invalidate_page(struct address_space *mapping,
43 					     struct page *page)
44 {
45 	if (page_has_private(page)) {
46 		if (!trylock_page(page))
47 			BUG();
48 		page->mapping = mapping;
49 		do_invalidatepage(page, 0);
50 		page->mapping = NULL;
51 		unlock_page(page);
52 	}
53 	page_cache_release(page);
54 }
55 
56 /*
57  * release a list of pages, invalidating them first if need be
58  */
59 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
60 					      struct list_head *pages)
61 {
62 	struct page *victim;
63 
64 	while (!list_empty(pages)) {
65 		victim = list_to_page(pages);
66 		list_del(&victim->lru);
67 		read_cache_pages_invalidate_page(mapping, victim);
68 	}
69 }
70 
71 /**
72  * read_cache_pages - populate an address space with some pages & start reads against them
73  * @mapping: the address_space
74  * @pages: The address of a list_head which contains the target pages.  These
75  *   pages have their ->index populated and are otherwise uninitialised.
76  * @filler: callback routine for filling a single page.
77  * @data: private data for the callback routine.
78  *
79  * Hides the details of the LRU cache etc from the filesystems.
80  */
81 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
82 			int (*filler)(void *, struct page *), void *data)
83 {
84 	struct page *page;
85 	int ret = 0;
86 
87 	while (!list_empty(pages)) {
88 		page = list_to_page(pages);
89 		list_del(&page->lru);
90 		if (add_to_page_cache_lru(page, mapping,
91 					page->index, GFP_KERNEL)) {
92 			read_cache_pages_invalidate_page(mapping, page);
93 			continue;
94 		}
95 		page_cache_release(page);
96 
97 		ret = filler(data, page);
98 		if (unlikely(ret)) {
99 			read_cache_pages_invalidate_pages(mapping, pages);
100 			break;
101 		}
102 		task_io_account_read(PAGE_CACHE_SIZE);
103 	}
104 	return ret;
105 }
106 
107 EXPORT_SYMBOL(read_cache_pages);
108 
109 static int read_pages(struct address_space *mapping, struct file *filp,
110 		struct list_head *pages, unsigned nr_pages)
111 {
112 	unsigned page_idx;
113 	int ret;
114 
115 	if (mapping->a_ops->readpages) {
116 		ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
117 		/* Clean up the remaining pages */
118 		put_pages_list(pages);
119 		goto out;
120 	}
121 
122 	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
123 		struct page *page = list_to_page(pages);
124 		list_del(&page->lru);
125 		if (!add_to_page_cache_lru(page, mapping,
126 					page->index, GFP_KERNEL)) {
127 			mapping->a_ops->readpage(filp, page);
128 		}
129 		page_cache_release(page);
130 	}
131 	ret = 0;
132 out:
133 	return ret;
134 }
135 
136 /*
137  * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
138  * the pages first, then submits them all for I/O. This avoids the very bad
139  * behaviour which would occur if page allocations are causing VM writeback.
140  * We really don't want to intermingle reads and writes like that.
141  *
142  * Returns the number of pages requested, or the maximum amount of I/O allowed.
143  */
144 static int
145 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
146 			pgoff_t offset, unsigned long nr_to_read,
147 			unsigned long lookahead_size)
148 {
149 	struct inode *inode = mapping->host;
150 	struct page *page;
151 	unsigned long end_index;	/* The last page we want to read */
152 	LIST_HEAD(page_pool);
153 	int page_idx;
154 	int ret = 0;
155 	loff_t isize = i_size_read(inode);
156 
157 	if (isize == 0)
158 		goto out;
159 
160 	end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
161 
162 	/*
163 	 * Preallocate as many pages as we will need.
164 	 */
165 	for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
166 		pgoff_t page_offset = offset + page_idx;
167 
168 		if (page_offset > end_index)
169 			break;
170 
171 		rcu_read_lock();
172 		page = radix_tree_lookup(&mapping->page_tree, page_offset);
173 		rcu_read_unlock();
174 		if (page)
175 			continue;
176 
177 		page = page_cache_alloc_cold(mapping);
178 		if (!page)
179 			break;
180 		page->index = page_offset;
181 		list_add(&page->lru, &page_pool);
182 		if (page_idx == nr_to_read - lookahead_size)
183 			SetPageReadahead(page);
184 		ret++;
185 	}
186 
187 	/*
188 	 * Now start the IO.  We ignore I/O errors - if the page is not
189 	 * uptodate then the caller will launch readpage again, and
190 	 * will then handle the error.
191 	 */
192 	if (ret)
193 		read_pages(mapping, filp, &page_pool, ret);
194 	BUG_ON(!list_empty(&page_pool));
195 out:
196 	return ret;
197 }
198 
199 /*
200  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
201  * memory at once.
202  */
203 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
204 		pgoff_t offset, unsigned long nr_to_read)
205 {
206 	int ret = 0;
207 
208 	if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
209 		return -EINVAL;
210 
211 	nr_to_read = max_sane_readahead(nr_to_read);
212 	while (nr_to_read) {
213 		int err;
214 
215 		unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
216 
217 		if (this_chunk > nr_to_read)
218 			this_chunk = nr_to_read;
219 		err = __do_page_cache_readahead(mapping, filp,
220 						offset, this_chunk, 0);
221 		if (err < 0) {
222 			ret = err;
223 			break;
224 		}
225 		ret += err;
226 		offset += this_chunk;
227 		nr_to_read -= this_chunk;
228 	}
229 	return ret;
230 }
231 
232 /*
233  * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
234  * sensible upper limit.
235  */
236 unsigned long max_sane_readahead(unsigned long nr)
237 {
238 	return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
239 		+ node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
240 }
241 
242 /*
243  * Submit IO for the read-ahead request in file_ra_state.
244  */
245 unsigned long ra_submit(struct file_ra_state *ra,
246 		       struct address_space *mapping, struct file *filp)
247 {
248 	int actual;
249 
250 	actual = __do_page_cache_readahead(mapping, filp,
251 					ra->start, ra->size, ra->async_size);
252 
253 	return actual;
254 }
255 
256 /*
257  * Set the initial window size, round to next power of 2 and square
258  * for small size, x 4 for medium, and x 2 for large
259  * for 128k (32 page) max ra
260  * 1-8 page = 32k initial, > 8 page = 128k initial
261  */
262 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
263 {
264 	unsigned long newsize = roundup_pow_of_two(size);
265 
266 	if (newsize <= max / 32)
267 		newsize = newsize * 4;
268 	else if (newsize <= max / 4)
269 		newsize = newsize * 2;
270 	else
271 		newsize = max;
272 
273 	return newsize;
274 }
275 
276 /*
277  *  Get the previous window size, ramp it up, and
278  *  return it as the new window size.
279  */
280 static unsigned long get_next_ra_size(struct file_ra_state *ra,
281 						unsigned long max)
282 {
283 	unsigned long cur = ra->size;
284 	unsigned long newsize;
285 
286 	if (cur < max / 16)
287 		newsize = 4 * cur;
288 	else
289 		newsize = 2 * cur;
290 
291 	return min(newsize, max);
292 }
293 
294 /*
295  * On-demand readahead design.
296  *
297  * The fields in struct file_ra_state represent the most-recently-executed
298  * readahead attempt:
299  *
300  *                        |<----- async_size ---------|
301  *     |------------------- size -------------------->|
302  *     |==================#===========================|
303  *     ^start             ^page marked with PG_readahead
304  *
305  * To overlap application thinking time and disk I/O time, we do
306  * `readahead pipelining': Do not wait until the application consumed all
307  * readahead pages and stalled on the missing page at readahead_index;
308  * Instead, submit an asynchronous readahead I/O as soon as there are
309  * only async_size pages left in the readahead window. Normally async_size
310  * will be equal to size, for maximum pipelining.
311  *
312  * In interleaved sequential reads, concurrent streams on the same fd can
313  * be invalidating each other's readahead state. So we flag the new readahead
314  * page at (start+size-async_size) with PG_readahead, and use it as readahead
315  * indicator. The flag won't be set on already cached pages, to avoid the
316  * readahead-for-nothing fuss, saving pointless page cache lookups.
317  *
318  * prev_pos tracks the last visited byte in the _previous_ read request.
319  * It should be maintained by the caller, and will be used for detecting
320  * small random reads. Note that the readahead algorithm checks loosely
321  * for sequential patterns. Hence interleaved reads might be served as
322  * sequential ones.
323  *
324  * There is a special-case: if the first page which the application tries to
325  * read happens to be the first page of the file, it is assumed that a linear
326  * read is about to happen and the window is immediately set to the initial size
327  * based on I/O request size and the max_readahead.
328  *
329  * The code ramps up the readahead size aggressively at first, but slow down as
330  * it approaches max_readhead.
331  */
332 
333 /*
334  * Count contiguously cached pages from @offset-1 to @offset-@max,
335  * this count is a conservative estimation of
336  * 	- length of the sequential read sequence, or
337  * 	- thrashing threshold in memory tight systems
338  */
339 static pgoff_t count_history_pages(struct address_space *mapping,
340 				   struct file_ra_state *ra,
341 				   pgoff_t offset, unsigned long max)
342 {
343 	pgoff_t head;
344 
345 	rcu_read_lock();
346 	head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
347 	rcu_read_unlock();
348 
349 	return offset - 1 - head;
350 }
351 
352 /*
353  * page cache context based read-ahead
354  */
355 static int try_context_readahead(struct address_space *mapping,
356 				 struct file_ra_state *ra,
357 				 pgoff_t offset,
358 				 unsigned long req_size,
359 				 unsigned long max)
360 {
361 	pgoff_t size;
362 
363 	size = count_history_pages(mapping, ra, offset, max);
364 
365 	/*
366 	 * no history pages:
367 	 * it could be a random read
368 	 */
369 	if (!size)
370 		return 0;
371 
372 	/*
373 	 * starts from beginning of file:
374 	 * it is a strong indication of long-run stream (or whole-file-read)
375 	 */
376 	if (size >= offset)
377 		size *= 2;
378 
379 	ra->start = offset;
380 	ra->size = get_init_ra_size(size + req_size, max);
381 	ra->async_size = ra->size;
382 
383 	return 1;
384 }
385 
386 /*
387  * A minimal readahead algorithm for trivial sequential/random reads.
388  */
389 static unsigned long
390 ondemand_readahead(struct address_space *mapping,
391 		   struct file_ra_state *ra, struct file *filp,
392 		   bool hit_readahead_marker, pgoff_t offset,
393 		   unsigned long req_size)
394 {
395 	unsigned long max = max_sane_readahead(ra->ra_pages);
396 
397 	/*
398 	 * start of file
399 	 */
400 	if (!offset)
401 		goto initial_readahead;
402 
403 	/*
404 	 * It's the expected callback offset, assume sequential access.
405 	 * Ramp up sizes, and push forward the readahead window.
406 	 */
407 	if ((offset == (ra->start + ra->size - ra->async_size) ||
408 	     offset == (ra->start + ra->size))) {
409 		ra->start += ra->size;
410 		ra->size = get_next_ra_size(ra, max);
411 		ra->async_size = ra->size;
412 		goto readit;
413 	}
414 
415 	/*
416 	 * Hit a marked page without valid readahead state.
417 	 * E.g. interleaved reads.
418 	 * Query the pagecache for async_size, which normally equals to
419 	 * readahead size. Ramp it up and use it as the new readahead size.
420 	 */
421 	if (hit_readahead_marker) {
422 		pgoff_t start;
423 
424 		rcu_read_lock();
425 		start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
426 		rcu_read_unlock();
427 
428 		if (!start || start - offset > max)
429 			return 0;
430 
431 		ra->start = start;
432 		ra->size = start - offset;	/* old async_size */
433 		ra->size += req_size;
434 		ra->size = get_next_ra_size(ra, max);
435 		ra->async_size = ra->size;
436 		goto readit;
437 	}
438 
439 	/*
440 	 * oversize read
441 	 */
442 	if (req_size > max)
443 		goto initial_readahead;
444 
445 	/*
446 	 * sequential cache miss
447 	 */
448 	if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
449 		goto initial_readahead;
450 
451 	/*
452 	 * Query the page cache and look for the traces(cached history pages)
453 	 * that a sequential stream would leave behind.
454 	 */
455 	if (try_context_readahead(mapping, ra, offset, req_size, max))
456 		goto readit;
457 
458 	/*
459 	 * standalone, small random read
460 	 * Read as is, and do not pollute the readahead state.
461 	 */
462 	return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
463 
464 initial_readahead:
465 	ra->start = offset;
466 	ra->size = get_init_ra_size(req_size, max);
467 	ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
468 
469 readit:
470 	/*
471 	 * Will this read hit the readahead marker made by itself?
472 	 * If so, trigger the readahead marker hit now, and merge
473 	 * the resulted next readahead window into the current one.
474 	 */
475 	if (offset == ra->start && ra->size == ra->async_size) {
476 		ra->async_size = get_next_ra_size(ra, max);
477 		ra->size += ra->async_size;
478 	}
479 
480 	return ra_submit(ra, mapping, filp);
481 }
482 
483 /**
484  * page_cache_sync_readahead - generic file readahead
485  * @mapping: address_space which holds the pagecache and I/O vectors
486  * @ra: file_ra_state which holds the readahead state
487  * @filp: passed on to ->readpage() and ->readpages()
488  * @offset: start offset into @mapping, in pagecache page-sized units
489  * @req_size: hint: total size of the read which the caller is performing in
490  *            pagecache pages
491  *
492  * page_cache_sync_readahead() should be called when a cache miss happened:
493  * it will submit the read.  The readahead logic may decide to piggyback more
494  * pages onto the read request if access patterns suggest it will improve
495  * performance.
496  */
497 void page_cache_sync_readahead(struct address_space *mapping,
498 			       struct file_ra_state *ra, struct file *filp,
499 			       pgoff_t offset, unsigned long req_size)
500 {
501 	/* no read-ahead */
502 	if (!ra->ra_pages)
503 		return;
504 
505 	/* be dumb */
506 	if (filp && (filp->f_mode & FMODE_RANDOM)) {
507 		force_page_cache_readahead(mapping, filp, offset, req_size);
508 		return;
509 	}
510 
511 	/* do read-ahead */
512 	ondemand_readahead(mapping, ra, filp, false, offset, req_size);
513 }
514 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
515 
516 /**
517  * page_cache_async_readahead - file readahead for marked pages
518  * @mapping: address_space which holds the pagecache and I/O vectors
519  * @ra: file_ra_state which holds the readahead state
520  * @filp: passed on to ->readpage() and ->readpages()
521  * @page: the page at @offset which has the PG_readahead flag set
522  * @offset: start offset into @mapping, in pagecache page-sized units
523  * @req_size: hint: total size of the read which the caller is performing in
524  *            pagecache pages
525  *
526  * page_cache_async_readahead() should be called when a page is used which
527  * has the PG_readahead flag; this is a marker to suggest that the application
528  * has used up enough of the readahead window that we should start pulling in
529  * more pages.
530  */
531 void
532 page_cache_async_readahead(struct address_space *mapping,
533 			   struct file_ra_state *ra, struct file *filp,
534 			   struct page *page, pgoff_t offset,
535 			   unsigned long req_size)
536 {
537 	/* no read-ahead */
538 	if (!ra->ra_pages)
539 		return;
540 
541 	/*
542 	 * Same bit is used for PG_readahead and PG_reclaim.
543 	 */
544 	if (PageWriteback(page))
545 		return;
546 
547 	ClearPageReadahead(page);
548 
549 	/*
550 	 * Defer asynchronous read-ahead on IO congestion.
551 	 */
552 	if (bdi_read_congested(mapping->backing_dev_info))
553 		return;
554 
555 	/* do read-ahead */
556 	ondemand_readahead(mapping, ra, filp, true, offset, req_size);
557 
558 #ifdef CONFIG_BLOCK
559 	/*
560 	 * Normally the current page is !uptodate and lock_page() will be
561 	 * immediately called to implicitly unplug the device. However this
562 	 * is not always true for RAID conifgurations, where data arrives
563 	 * not strictly in their submission order. In this case we need to
564 	 * explicitly kick off the IO.
565 	 */
566 	if (PageUptodate(page))
567 		blk_run_backing_dev(mapping->backing_dev_info, NULL);
568 #endif
569 }
570 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
571