Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * mm/readahead.c - address_space-level file readahead. | |
3 | * | |
4 | * Copyright (C) 2002, Linus Torvalds | |
5 | * | |
e1f8e874 | 6 | * 09Apr2002 Andrew Morton |
1da177e4 LT |
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> | |
8bde37f0 | 16 | #include <linux/task_io_accounting_ops.h> |
1da177e4 | 17 | #include <linux/pagevec.h> |
f5ff8422 | 18 | #include <linux/pagemap.h> |
1da177e4 LT |
19 | |
20 | void default_unplug_io_fn(struct backing_dev_info *bdi, struct page *page) | |
21 | { | |
22 | } | |
23 | EXPORT_SYMBOL(default_unplug_io_fn); | |
24 | ||
25 | struct backing_dev_info default_backing_dev_info = { | |
535443f5 | 26 | .ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE, |
1da177e4 LT |
27 | .state = 0, |
28 | .capabilities = BDI_CAP_MAP_COPY, | |
29 | .unplug_io_fn = default_unplug_io_fn, | |
30 | }; | |
31 | EXPORT_SYMBOL_GPL(default_backing_dev_info); | |
32 | ||
33 | /* | |
34 | * Initialise a struct file's readahead state. Assumes that the caller has | |
35 | * memset *ra to zero. | |
36 | */ | |
37 | void | |
38 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) | |
39 | { | |
40 | ra->ra_pages = mapping->backing_dev_info->ra_pages; | |
f4e6b498 | 41 | ra->prev_pos = -1; |
1da177e4 | 42 | } |
d41cc702 | 43 | EXPORT_SYMBOL_GPL(file_ra_state_init); |
1da177e4 | 44 | |
1da177e4 LT |
45 | #define list_to_page(head) (list_entry((head)->prev, struct page, lru)) |
46 | ||
47 | /** | |
bd40cdda | 48 | * read_cache_pages - populate an address space with some pages & start reads against them |
1da177e4 LT |
49 | * @mapping: the address_space |
50 | * @pages: The address of a list_head which contains the target pages. These | |
51 | * pages have their ->index populated and are otherwise uninitialised. | |
52 | * @filler: callback routine for filling a single page. | |
53 | * @data: private data for the callback routine. | |
54 | * | |
55 | * Hides the details of the LRU cache etc from the filesystems. | |
56 | */ | |
57 | int read_cache_pages(struct address_space *mapping, struct list_head *pages, | |
58 | int (*filler)(void *, struct page *), void *data) | |
59 | { | |
60 | struct page *page; | |
1da177e4 LT |
61 | int ret = 0; |
62 | ||
1da177e4 LT |
63 | while (!list_empty(pages)) { |
64 | page = list_to_page(pages); | |
65 | list_del(&page->lru); | |
eb2be189 NP |
66 | if (add_to_page_cache_lru(page, mapping, |
67 | page->index, GFP_KERNEL)) { | |
1da177e4 LT |
68 | page_cache_release(page); |
69 | continue; | |
70 | } | |
eb2be189 NP |
71 | page_cache_release(page); |
72 | ||
1da177e4 | 73 | ret = filler(data, page); |
eb2be189 | 74 | if (unlikely(ret)) { |
38da288b | 75 | put_pages_list(pages); |
1da177e4 LT |
76 | break; |
77 | } | |
8bde37f0 | 78 | task_io_account_read(PAGE_CACHE_SIZE); |
1da177e4 | 79 | } |
1da177e4 LT |
80 | return ret; |
81 | } | |
82 | ||
83 | EXPORT_SYMBOL(read_cache_pages); | |
84 | ||
85 | static int read_pages(struct address_space *mapping, struct file *filp, | |
86 | struct list_head *pages, unsigned nr_pages) | |
87 | { | |
88 | unsigned page_idx; | |
994fc28c | 89 | int ret; |
1da177e4 LT |
90 | |
91 | if (mapping->a_ops->readpages) { | |
92 | ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages); | |
029e332e OH |
93 | /* Clean up the remaining pages */ |
94 | put_pages_list(pages); | |
1da177e4 LT |
95 | goto out; |
96 | } | |
97 | ||
1da177e4 LT |
98 | for (page_idx = 0; page_idx < nr_pages; page_idx++) { |
99 | struct page *page = list_to_page(pages); | |
100 | list_del(&page->lru); | |
eb2be189 | 101 | if (!add_to_page_cache_lru(page, mapping, |
1da177e4 | 102 | page->index, GFP_KERNEL)) { |
9f1a3cfc | 103 | mapping->a_ops->readpage(filp, page); |
eb2be189 NP |
104 | } |
105 | page_cache_release(page); | |
1da177e4 | 106 | } |
994fc28c | 107 | ret = 0; |
1da177e4 LT |
108 | out: |
109 | return ret; | |
110 | } | |
111 | ||
1da177e4 LT |
112 | /* |
113 | * do_page_cache_readahead actually reads a chunk of disk. It allocates all | |
114 | * the pages first, then submits them all for I/O. This avoids the very bad | |
115 | * behaviour which would occur if page allocations are causing VM writeback. | |
116 | * We really don't want to intermingle reads and writes like that. | |
117 | * | |
118 | * Returns the number of pages requested, or the maximum amount of I/O allowed. | |
119 | * | |
120 | * do_page_cache_readahead() returns -1 if it encountered request queue | |
121 | * congestion. | |
122 | */ | |
123 | static int | |
124 | __do_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
46fc3e7b FW |
125 | pgoff_t offset, unsigned long nr_to_read, |
126 | unsigned long lookahead_size) | |
1da177e4 LT |
127 | { |
128 | struct inode *inode = mapping->host; | |
129 | struct page *page; | |
130 | unsigned long end_index; /* The last page we want to read */ | |
131 | LIST_HEAD(page_pool); | |
132 | int page_idx; | |
133 | int ret = 0; | |
134 | loff_t isize = i_size_read(inode); | |
135 | ||
136 | if (isize == 0) | |
137 | goto out; | |
138 | ||
46fc3e7b | 139 | end_index = ((isize - 1) >> PAGE_CACHE_SHIFT); |
1da177e4 LT |
140 | |
141 | /* | |
142 | * Preallocate as many pages as we will need. | |
143 | */ | |
1da177e4 | 144 | for (page_idx = 0; page_idx < nr_to_read; page_idx++) { |
7361f4d8 | 145 | pgoff_t page_offset = offset + page_idx; |
c743d96b | 146 | |
1da177e4 LT |
147 | if (page_offset > end_index) |
148 | break; | |
149 | ||
00128188 | 150 | rcu_read_lock(); |
1da177e4 | 151 | page = radix_tree_lookup(&mapping->page_tree, page_offset); |
00128188 | 152 | rcu_read_unlock(); |
1da177e4 LT |
153 | if (page) |
154 | continue; | |
155 | ||
1da177e4 | 156 | page = page_cache_alloc_cold(mapping); |
1da177e4 LT |
157 | if (!page) |
158 | break; | |
159 | page->index = page_offset; | |
160 | list_add(&page->lru, &page_pool); | |
46fc3e7b FW |
161 | if (page_idx == nr_to_read - lookahead_size) |
162 | SetPageReadahead(page); | |
1da177e4 LT |
163 | ret++; |
164 | } | |
1da177e4 LT |
165 | |
166 | /* | |
167 | * Now start the IO. We ignore I/O errors - if the page is not | |
168 | * uptodate then the caller will launch readpage again, and | |
169 | * will then handle the error. | |
170 | */ | |
171 | if (ret) | |
172 | read_pages(mapping, filp, &page_pool, ret); | |
173 | BUG_ON(!list_empty(&page_pool)); | |
174 | out: | |
175 | return ret; | |
176 | } | |
177 | ||
178 | /* | |
179 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much | |
180 | * memory at once. | |
181 | */ | |
182 | int force_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
7361f4d8 | 183 | pgoff_t offset, unsigned long nr_to_read) |
1da177e4 LT |
184 | { |
185 | int ret = 0; | |
186 | ||
187 | if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages)) | |
188 | return -EINVAL; | |
189 | ||
190 | while (nr_to_read) { | |
191 | int err; | |
192 | ||
193 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE; | |
194 | ||
195 | if (this_chunk > nr_to_read) | |
196 | this_chunk = nr_to_read; | |
197 | err = __do_page_cache_readahead(mapping, filp, | |
46fc3e7b | 198 | offset, this_chunk, 0); |
1da177e4 LT |
199 | if (err < 0) { |
200 | ret = err; | |
201 | break; | |
202 | } | |
203 | ret += err; | |
204 | offset += this_chunk; | |
205 | nr_to_read -= this_chunk; | |
206 | } | |
207 | return ret; | |
208 | } | |
209 | ||
1da177e4 LT |
210 | /* |
211 | * This version skips the IO if the queue is read-congested, and will tell the | |
212 | * block layer to abandon the readahead if request allocation would block. | |
213 | * | |
214 | * force_page_cache_readahead() will ignore queue congestion and will block on | |
215 | * request queues. | |
216 | */ | |
217 | int do_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
7361f4d8 | 218 | pgoff_t offset, unsigned long nr_to_read) |
1da177e4 LT |
219 | { |
220 | if (bdi_read_congested(mapping->backing_dev_info)) | |
221 | return -1; | |
222 | ||
46fc3e7b | 223 | return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0); |
1da177e4 LT |
224 | } |
225 | ||
1da177e4 LT |
226 | /* |
227 | * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a | |
228 | * sensible upper limit. | |
229 | */ | |
230 | unsigned long max_sane_readahead(unsigned long nr) | |
231 | { | |
4f98a2fe | 232 | return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE) |
05a0416b | 233 | + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2); |
1da177e4 | 234 | } |
5ce1110b | 235 | |
e0bf68dd PZ |
236 | static int __init readahead_init(void) |
237 | { | |
cf0ca9fe PZ |
238 | int err; |
239 | ||
240 | err = bdi_init(&default_backing_dev_info); | |
241 | if (!err) | |
242 | bdi_register(&default_backing_dev_info, NULL, "default"); | |
243 | ||
244 | return err; | |
e0bf68dd PZ |
245 | } |
246 | subsys_initcall(readahead_init); | |
247 | ||
5ce1110b FW |
248 | /* |
249 | * Submit IO for the read-ahead request in file_ra_state. | |
250 | */ | |
f9acc8c7 | 251 | static unsigned long ra_submit(struct file_ra_state *ra, |
5ce1110b FW |
252 | struct address_space *mapping, struct file *filp) |
253 | { | |
5ce1110b FW |
254 | int actual; |
255 | ||
5ce1110b | 256 | actual = __do_page_cache_readahead(mapping, filp, |
f9acc8c7 | 257 | ra->start, ra->size, ra->async_size); |
5ce1110b FW |
258 | |
259 | return actual; | |
260 | } | |
122a21d1 | 261 | |
c743d96b FW |
262 | /* |
263 | * Set the initial window size, round to next power of 2 and square | |
264 | * for small size, x 4 for medium, and x 2 for large | |
265 | * for 128k (32 page) max ra | |
266 | * 1-8 page = 32k initial, > 8 page = 128k initial | |
267 | */ | |
268 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max) | |
269 | { | |
270 | unsigned long newsize = roundup_pow_of_two(size); | |
271 | ||
272 | if (newsize <= max / 32) | |
273 | newsize = newsize * 4; | |
274 | else if (newsize <= max / 4) | |
275 | newsize = newsize * 2; | |
276 | else | |
277 | newsize = max; | |
278 | ||
279 | return newsize; | |
280 | } | |
281 | ||
122a21d1 FW |
282 | /* |
283 | * Get the previous window size, ramp it up, and | |
284 | * return it as the new window size. | |
285 | */ | |
c743d96b | 286 | static unsigned long get_next_ra_size(struct file_ra_state *ra, |
122a21d1 FW |
287 | unsigned long max) |
288 | { | |
f9acc8c7 | 289 | unsigned long cur = ra->size; |
122a21d1 FW |
290 | unsigned long newsize; |
291 | ||
292 | if (cur < max / 16) | |
c743d96b | 293 | newsize = 4 * cur; |
122a21d1 | 294 | else |
c743d96b | 295 | newsize = 2 * cur; |
122a21d1 FW |
296 | |
297 | return min(newsize, max); | |
298 | } | |
299 | ||
300 | /* | |
301 | * On-demand readahead design. | |
302 | * | |
303 | * The fields in struct file_ra_state represent the most-recently-executed | |
304 | * readahead attempt: | |
305 | * | |
f9acc8c7 FW |
306 | * |<----- async_size ---------| |
307 | * |------------------- size -------------------->| | |
308 | * |==================#===========================| | |
309 | * ^start ^page marked with PG_readahead | |
122a21d1 FW |
310 | * |
311 | * To overlap application thinking time and disk I/O time, we do | |
312 | * `readahead pipelining': Do not wait until the application consumed all | |
313 | * readahead pages and stalled on the missing page at readahead_index; | |
f9acc8c7 FW |
314 | * Instead, submit an asynchronous readahead I/O as soon as there are |
315 | * only async_size pages left in the readahead window. Normally async_size | |
316 | * will be equal to size, for maximum pipelining. | |
122a21d1 FW |
317 | * |
318 | * In interleaved sequential reads, concurrent streams on the same fd can | |
319 | * be invalidating each other's readahead state. So we flag the new readahead | |
f9acc8c7 | 320 | * page at (start+size-async_size) with PG_readahead, and use it as readahead |
122a21d1 FW |
321 | * indicator. The flag won't be set on already cached pages, to avoid the |
322 | * readahead-for-nothing fuss, saving pointless page cache lookups. | |
323 | * | |
f4e6b498 | 324 | * prev_pos tracks the last visited byte in the _previous_ read request. |
122a21d1 FW |
325 | * It should be maintained by the caller, and will be used for detecting |
326 | * small random reads. Note that the readahead algorithm checks loosely | |
327 | * for sequential patterns. Hence interleaved reads might be served as | |
328 | * sequential ones. | |
329 | * | |
330 | * There is a special-case: if the first page which the application tries to | |
331 | * read happens to be the first page of the file, it is assumed that a linear | |
332 | * read is about to happen and the window is immediately set to the initial size | |
333 | * based on I/O request size and the max_readahead. | |
334 | * | |
335 | * The code ramps up the readahead size aggressively at first, but slow down as | |
336 | * it approaches max_readhead. | |
337 | */ | |
338 | ||
339 | /* | |
340 | * A minimal readahead algorithm for trivial sequential/random reads. | |
341 | */ | |
342 | static unsigned long | |
343 | ondemand_readahead(struct address_space *mapping, | |
344 | struct file_ra_state *ra, struct file *filp, | |
cf914a7d | 345 | bool hit_readahead_marker, pgoff_t offset, |
122a21d1 FW |
346 | unsigned long req_size) |
347 | { | |
f4e6b498 FW |
348 | int max = ra->ra_pages; /* max readahead pages */ |
349 | pgoff_t prev_offset; | |
350 | int sequential; | |
122a21d1 FW |
351 | |
352 | /* | |
f9acc8c7 | 353 | * It's the expected callback offset, assume sequential access. |
122a21d1 FW |
354 | * Ramp up sizes, and push forward the readahead window. |
355 | */ | |
f9acc8c7 FW |
356 | if (offset && (offset == (ra->start + ra->size - ra->async_size) || |
357 | offset == (ra->start + ra->size))) { | |
358 | ra->start += ra->size; | |
359 | ra->size = get_next_ra_size(ra, max); | |
360 | ra->async_size = ra->size; | |
361 | goto readit; | |
122a21d1 FW |
362 | } |
363 | ||
f4e6b498 FW |
364 | prev_offset = ra->prev_pos >> PAGE_CACHE_SHIFT; |
365 | sequential = offset - prev_offset <= 1UL || req_size > max; | |
366 | ||
122a21d1 FW |
367 | /* |
368 | * Standalone, small read. | |
369 | * Read as is, and do not pollute the readahead state. | |
370 | */ | |
cf914a7d | 371 | if (!hit_readahead_marker && !sequential) { |
122a21d1 FW |
372 | return __do_page_cache_readahead(mapping, filp, |
373 | offset, req_size, 0); | |
374 | } | |
375 | ||
6b10c6c9 FW |
376 | /* |
377 | * Hit a marked page without valid readahead state. | |
378 | * E.g. interleaved reads. | |
379 | * Query the pagecache for async_size, which normally equals to | |
380 | * readahead size. Ramp it up and use it as the new readahead size. | |
381 | */ | |
382 | if (hit_readahead_marker) { | |
383 | pgoff_t start; | |
384 | ||
30002ed2 NP |
385 | rcu_read_lock(); |
386 | start = radix_tree_next_hole(&mapping->page_tree, offset,max+1); | |
387 | rcu_read_unlock(); | |
6b10c6c9 FW |
388 | |
389 | if (!start || start - offset > max) | |
390 | return 0; | |
391 | ||
392 | ra->start = start; | |
393 | ra->size = start - offset; /* old async_size */ | |
394 | ra->size = get_next_ra_size(ra, max); | |
395 | ra->async_size = ra->size; | |
396 | goto readit; | |
397 | } | |
398 | ||
122a21d1 FW |
399 | /* |
400 | * It may be one of | |
401 | * - first read on start of file | |
402 | * - sequential cache miss | |
403 | * - oversize random read | |
404 | * Start readahead for it. | |
405 | */ | |
f9acc8c7 FW |
406 | ra->start = offset; |
407 | ra->size = get_init_ra_size(req_size, max); | |
408 | ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; | |
122a21d1 | 409 | |
f9acc8c7 | 410 | readit: |
122a21d1 FW |
411 | return ra_submit(ra, mapping, filp); |
412 | } | |
413 | ||
414 | /** | |
cf914a7d | 415 | * page_cache_sync_readahead - generic file readahead |
122a21d1 FW |
416 | * @mapping: address_space which holds the pagecache and I/O vectors |
417 | * @ra: file_ra_state which holds the readahead state | |
418 | * @filp: passed on to ->readpage() and ->readpages() | |
cf914a7d | 419 | * @offset: start offset into @mapping, in pagecache page-sized units |
122a21d1 | 420 | * @req_size: hint: total size of the read which the caller is performing in |
cf914a7d | 421 | * pagecache pages |
122a21d1 | 422 | * |
cf914a7d RR |
423 | * page_cache_sync_readahead() should be called when a cache miss happened: |
424 | * it will submit the read. The readahead logic may decide to piggyback more | |
425 | * pages onto the read request if access patterns suggest it will improve | |
426 | * performance. | |
122a21d1 | 427 | */ |
cf914a7d RR |
428 | void page_cache_sync_readahead(struct address_space *mapping, |
429 | struct file_ra_state *ra, struct file *filp, | |
430 | pgoff_t offset, unsigned long req_size) | |
122a21d1 FW |
431 | { |
432 | /* no read-ahead */ | |
433 | if (!ra->ra_pages) | |
cf914a7d RR |
434 | return; |
435 | ||
436 | /* do read-ahead */ | |
437 | ondemand_readahead(mapping, ra, filp, false, offset, req_size); | |
438 | } | |
439 | EXPORT_SYMBOL_GPL(page_cache_sync_readahead); | |
440 | ||
441 | /** | |
442 | * page_cache_async_readahead - file readahead for marked pages | |
443 | * @mapping: address_space which holds the pagecache and I/O vectors | |
444 | * @ra: file_ra_state which holds the readahead state | |
445 | * @filp: passed on to ->readpage() and ->readpages() | |
446 | * @page: the page at @offset which has the PG_readahead flag set | |
447 | * @offset: start offset into @mapping, in pagecache page-sized units | |
448 | * @req_size: hint: total size of the read which the caller is performing in | |
449 | * pagecache pages | |
450 | * | |
451 | * page_cache_async_ondemand() should be called when a page is used which | |
f7850d93 | 452 | * has the PG_readahead flag; this is a marker to suggest that the application |
cf914a7d | 453 | * has used up enough of the readahead window that we should start pulling in |
f7850d93 RD |
454 | * more pages. |
455 | */ | |
cf914a7d RR |
456 | void |
457 | page_cache_async_readahead(struct address_space *mapping, | |
458 | struct file_ra_state *ra, struct file *filp, | |
459 | struct page *page, pgoff_t offset, | |
460 | unsigned long req_size) | |
461 | { | |
462 | /* no read-ahead */ | |
463 | if (!ra->ra_pages) | |
464 | return; | |
465 | ||
466 | /* | |
467 | * Same bit is used for PG_readahead and PG_reclaim. | |
468 | */ | |
469 | if (PageWriteback(page)) | |
470 | return; | |
471 | ||
472 | ClearPageReadahead(page); | |
473 | ||
474 | /* | |
475 | * Defer asynchronous read-ahead on IO congestion. | |
476 | */ | |
477 | if (bdi_read_congested(mapping->backing_dev_info)) | |
478 | return; | |
122a21d1 FW |
479 | |
480 | /* do read-ahead */ | |
cf914a7d | 481 | ondemand_readahead(mapping, ra, filp, true, offset, req_size); |
122a21d1 | 482 | } |
cf914a7d | 483 | EXPORT_SYMBOL_GPL(page_cache_async_readahead); |