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