xref: /openbmc/linux/mm/readahead.c (revision b8bb76713ec50df2f11efee386e16f93d51e1076)
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/mm.h>
13 #include <linux/module.h>
14 #include <linux/blkdev.h>
15 #include <linux/backing-dev.h>
16 #include <linux/task_io_accounting_ops.h>
17 #include <linux/pagevec.h>
18 #include <linux/pagemap.h>
19 
20 /*
21  * Initialise a struct file's readahead state.  Assumes that the caller has
22  * memset *ra to zero.
23  */
24 void
25 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
26 {
27 	ra->ra_pages = mapping->backing_dev_info->ra_pages;
28 	ra->prev_pos = -1;
29 }
30 EXPORT_SYMBOL_GPL(file_ra_state_init);
31 
32 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
33 
34 /**
35  * read_cache_pages - populate an address space with some pages & start reads against them
36  * @mapping: the address_space
37  * @pages: The address of a list_head which contains the target pages.  These
38  *   pages have their ->index populated and are otherwise uninitialised.
39  * @filler: callback routine for filling a single page.
40  * @data: private data for the callback routine.
41  *
42  * Hides the details of the LRU cache etc from the filesystems.
43  */
44 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
45 			int (*filler)(void *, struct page *), void *data)
46 {
47 	struct page *page;
48 	int ret = 0;
49 
50 	while (!list_empty(pages)) {
51 		page = list_to_page(pages);
52 		list_del(&page->lru);
53 		if (add_to_page_cache_lru(page, mapping,
54 					page->index, GFP_KERNEL)) {
55 			page_cache_release(page);
56 			continue;
57 		}
58 		page_cache_release(page);
59 
60 		ret = filler(data, page);
61 		if (unlikely(ret)) {
62 			put_pages_list(pages);
63 			break;
64 		}
65 		task_io_account_read(PAGE_CACHE_SIZE);
66 	}
67 	return ret;
68 }
69 
70 EXPORT_SYMBOL(read_cache_pages);
71 
72 static int read_pages(struct address_space *mapping, struct file *filp,
73 		struct list_head *pages, unsigned nr_pages)
74 {
75 	unsigned page_idx;
76 	int ret;
77 
78 	if (mapping->a_ops->readpages) {
79 		ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
80 		/* Clean up the remaining pages */
81 		put_pages_list(pages);
82 		goto out;
83 	}
84 
85 	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
86 		struct page *page = list_to_page(pages);
87 		list_del(&page->lru);
88 		if (!add_to_page_cache_lru(page, mapping,
89 					page->index, GFP_KERNEL)) {
90 			mapping->a_ops->readpage(filp, page);
91 		}
92 		page_cache_release(page);
93 	}
94 	ret = 0;
95 out:
96 	return ret;
97 }
98 
99 /*
100  * do_page_cache_readahead actually reads a chunk of disk.  It allocates all
101  * the pages first, then submits them all for I/O. This avoids the very bad
102  * behaviour which would occur if page allocations are causing VM writeback.
103  * We really don't want to intermingle reads and writes like that.
104  *
105  * Returns the number of pages requested, or the maximum amount of I/O allowed.
106  *
107  * do_page_cache_readahead() returns -1 if it encountered request queue
108  * congestion.
109  */
110 static int
111 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
112 			pgoff_t offset, unsigned long nr_to_read,
113 			unsigned long lookahead_size)
114 {
115 	struct inode *inode = mapping->host;
116 	struct page *page;
117 	unsigned long end_index;	/* The last page we want to read */
118 	LIST_HEAD(page_pool);
119 	int page_idx;
120 	int ret = 0;
121 	loff_t isize = i_size_read(inode);
122 
123 	if (isize == 0)
124 		goto out;
125 
126 	end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
127 
128 	/*
129 	 * Preallocate as many pages as we will need.
130 	 */
131 	for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
132 		pgoff_t page_offset = offset + page_idx;
133 
134 		if (page_offset > end_index)
135 			break;
136 
137 		rcu_read_lock();
138 		page = radix_tree_lookup(&mapping->page_tree, page_offset);
139 		rcu_read_unlock();
140 		if (page)
141 			continue;
142 
143 		page = page_cache_alloc_cold(mapping);
144 		if (!page)
145 			break;
146 		page->index = page_offset;
147 		list_add(&page->lru, &page_pool);
148 		if (page_idx == nr_to_read - lookahead_size)
149 			SetPageReadahead(page);
150 		ret++;
151 	}
152 
153 	/*
154 	 * Now start the IO.  We ignore I/O errors - if the page is not
155 	 * uptodate then the caller will launch readpage again, and
156 	 * will then handle the error.
157 	 */
158 	if (ret)
159 		read_pages(mapping, filp, &page_pool, ret);
160 	BUG_ON(!list_empty(&page_pool));
161 out:
162 	return ret;
163 }
164 
165 /*
166  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
167  * memory at once.
168  */
169 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
170 		pgoff_t offset, unsigned long nr_to_read)
171 {
172 	int ret = 0;
173 
174 	if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
175 		return -EINVAL;
176 
177 	while (nr_to_read) {
178 		int err;
179 
180 		unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
181 
182 		if (this_chunk > nr_to_read)
183 			this_chunk = nr_to_read;
184 		err = __do_page_cache_readahead(mapping, filp,
185 						offset, this_chunk, 0);
186 		if (err < 0) {
187 			ret = err;
188 			break;
189 		}
190 		ret += err;
191 		offset += this_chunk;
192 		nr_to_read -= this_chunk;
193 	}
194 	return ret;
195 }
196 
197 /*
198  * This version skips the IO if the queue is read-congested, and will tell the
199  * block layer to abandon the readahead if request allocation would block.
200  *
201  * force_page_cache_readahead() will ignore queue congestion and will block on
202  * request queues.
203  */
204 int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
205 			pgoff_t offset, unsigned long nr_to_read)
206 {
207 	if (bdi_read_congested(mapping->backing_dev_info))
208 		return -1;
209 
210 	return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0);
211 }
212 
213 /*
214  * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
215  * sensible upper limit.
216  */
217 unsigned long max_sane_readahead(unsigned long nr)
218 {
219 	return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
220 		+ node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
221 }
222 
223 /*
224  * Submit IO for the read-ahead request in file_ra_state.
225  */
226 static unsigned long ra_submit(struct file_ra_state *ra,
227 		       struct address_space *mapping, struct file *filp)
228 {
229 	int actual;
230 
231 	actual = __do_page_cache_readahead(mapping, filp,
232 					ra->start, ra->size, ra->async_size);
233 
234 	return actual;
235 }
236 
237 /*
238  * Set the initial window size, round to next power of 2 and square
239  * for small size, x 4 for medium, and x 2 for large
240  * for 128k (32 page) max ra
241  * 1-8 page = 32k initial, > 8 page = 128k initial
242  */
243 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
244 {
245 	unsigned long newsize = roundup_pow_of_two(size);
246 
247 	if (newsize <= max / 32)
248 		newsize = newsize * 4;
249 	else if (newsize <= max / 4)
250 		newsize = newsize * 2;
251 	else
252 		newsize = max;
253 
254 	return newsize;
255 }
256 
257 /*
258  *  Get the previous window size, ramp it up, and
259  *  return it as the new window size.
260  */
261 static unsigned long get_next_ra_size(struct file_ra_state *ra,
262 						unsigned long max)
263 {
264 	unsigned long cur = ra->size;
265 	unsigned long newsize;
266 
267 	if (cur < max / 16)
268 		newsize = 4 * cur;
269 	else
270 		newsize = 2 * cur;
271 
272 	return min(newsize, max);
273 }
274 
275 /*
276  * On-demand readahead design.
277  *
278  * The fields in struct file_ra_state represent the most-recently-executed
279  * readahead attempt:
280  *
281  *                        |<----- async_size ---------|
282  *     |------------------- size -------------------->|
283  *     |==================#===========================|
284  *     ^start             ^page marked with PG_readahead
285  *
286  * To overlap application thinking time and disk I/O time, we do
287  * `readahead pipelining': Do not wait until the application consumed all
288  * readahead pages and stalled on the missing page at readahead_index;
289  * Instead, submit an asynchronous readahead I/O as soon as there are
290  * only async_size pages left in the readahead window. Normally async_size
291  * will be equal to size, for maximum pipelining.
292  *
293  * In interleaved sequential reads, concurrent streams on the same fd can
294  * be invalidating each other's readahead state. So we flag the new readahead
295  * page at (start+size-async_size) with PG_readahead, and use it as readahead
296  * indicator. The flag won't be set on already cached pages, to avoid the
297  * readahead-for-nothing fuss, saving pointless page cache lookups.
298  *
299  * prev_pos tracks the last visited byte in the _previous_ read request.
300  * It should be maintained by the caller, and will be used for detecting
301  * small random reads. Note that the readahead algorithm checks loosely
302  * for sequential patterns. Hence interleaved reads might be served as
303  * sequential ones.
304  *
305  * There is a special-case: if the first page which the application tries to
306  * read happens to be the first page of the file, it is assumed that a linear
307  * read is about to happen and the window is immediately set to the initial size
308  * based on I/O request size and the max_readahead.
309  *
310  * The code ramps up the readahead size aggressively at first, but slow down as
311  * it approaches max_readhead.
312  */
313 
314 /*
315  * A minimal readahead algorithm for trivial sequential/random reads.
316  */
317 static unsigned long
318 ondemand_readahead(struct address_space *mapping,
319 		   struct file_ra_state *ra, struct file *filp,
320 		   bool hit_readahead_marker, pgoff_t offset,
321 		   unsigned long req_size)
322 {
323 	int	max = ra->ra_pages;	/* max readahead pages */
324 	pgoff_t prev_offset;
325 	int	sequential;
326 
327 	/*
328 	 * It's the expected callback offset, assume sequential access.
329 	 * Ramp up sizes, and push forward the readahead window.
330 	 */
331 	if (offset && (offset == (ra->start + ra->size - ra->async_size) ||
332 			offset == (ra->start + ra->size))) {
333 		ra->start += ra->size;
334 		ra->size = get_next_ra_size(ra, max);
335 		ra->async_size = ra->size;
336 		goto readit;
337 	}
338 
339 	prev_offset = ra->prev_pos >> PAGE_CACHE_SHIFT;
340 	sequential = offset - prev_offset <= 1UL || req_size > max;
341 
342 	/*
343 	 * Standalone, small read.
344 	 * Read as is, and do not pollute the readahead state.
345 	 */
346 	if (!hit_readahead_marker && !sequential) {
347 		return __do_page_cache_readahead(mapping, filp,
348 						offset, req_size, 0);
349 	}
350 
351 	/*
352 	 * Hit a marked page without valid readahead state.
353 	 * E.g. interleaved reads.
354 	 * Query the pagecache for async_size, which normally equals to
355 	 * readahead size. Ramp it up and use it as the new readahead size.
356 	 */
357 	if (hit_readahead_marker) {
358 		pgoff_t start;
359 
360 		rcu_read_lock();
361 		start = radix_tree_next_hole(&mapping->page_tree, offset,max+1);
362 		rcu_read_unlock();
363 
364 		if (!start || start - offset > max)
365 			return 0;
366 
367 		ra->start = start;
368 		ra->size = start - offset;	/* old async_size */
369 		ra->size = get_next_ra_size(ra, max);
370 		ra->async_size = ra->size;
371 		goto readit;
372 	}
373 
374 	/*
375 	 * It may be one of
376 	 * 	- first read on start of file
377 	 * 	- sequential cache miss
378 	 * 	- oversize random read
379 	 * Start readahead for it.
380 	 */
381 	ra->start = offset;
382 	ra->size = get_init_ra_size(req_size, max);
383 	ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
384 
385 readit:
386 	return ra_submit(ra, mapping, filp);
387 }
388 
389 /**
390  * page_cache_sync_readahead - generic file readahead
391  * @mapping: address_space which holds the pagecache and I/O vectors
392  * @ra: file_ra_state which holds the readahead state
393  * @filp: passed on to ->readpage() and ->readpages()
394  * @offset: start offset into @mapping, in pagecache page-sized units
395  * @req_size: hint: total size of the read which the caller is performing in
396  *            pagecache pages
397  *
398  * page_cache_sync_readahead() should be called when a cache miss happened:
399  * it will submit the read.  The readahead logic may decide to piggyback more
400  * pages onto the read request if access patterns suggest it will improve
401  * performance.
402  */
403 void page_cache_sync_readahead(struct address_space *mapping,
404 			       struct file_ra_state *ra, struct file *filp,
405 			       pgoff_t offset, unsigned long req_size)
406 {
407 	/* no read-ahead */
408 	if (!ra->ra_pages)
409 		return;
410 
411 	/* do read-ahead */
412 	ondemand_readahead(mapping, ra, filp, false, offset, req_size);
413 }
414 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
415 
416 /**
417  * page_cache_async_readahead - file readahead for marked pages
418  * @mapping: address_space which holds the pagecache and I/O vectors
419  * @ra: file_ra_state which holds the readahead state
420  * @filp: passed on to ->readpage() and ->readpages()
421  * @page: the page at @offset which has the PG_readahead flag set
422  * @offset: start offset into @mapping, in pagecache page-sized units
423  * @req_size: hint: total size of the read which the caller is performing in
424  *            pagecache pages
425  *
426  * page_cache_async_ondemand() should be called when a page is used which
427  * has the PG_readahead flag; this is a marker to suggest that the application
428  * has used up enough of the readahead window that we should start pulling in
429  * more pages.
430  */
431 void
432 page_cache_async_readahead(struct address_space *mapping,
433 			   struct file_ra_state *ra, struct file *filp,
434 			   struct page *page, pgoff_t offset,
435 			   unsigned long req_size)
436 {
437 	/* no read-ahead */
438 	if (!ra->ra_pages)
439 		return;
440 
441 	/*
442 	 * Same bit is used for PG_readahead and PG_reclaim.
443 	 */
444 	if (PageWriteback(page))
445 		return;
446 
447 	ClearPageReadahead(page);
448 
449 	/*
450 	 * Defer asynchronous read-ahead on IO congestion.
451 	 */
452 	if (bdi_read_congested(mapping->backing_dev_info))
453 		return;
454 
455 	/* do read-ahead */
456 	ondemand_readahead(mapping, ra, filp, true, offset, req_size);
457 }
458 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
459