Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * mm/readahead.c - address_space-level file readahead. | |
3 | * | |
4 | * Copyright (C) 2002, Linus Torvalds | |
5 | * | |
6 | * 09Apr2002 akpm@zip.com.au | |
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 | ||
f615bfca FW |
25 | /* |
26 | * Convienent macros for min/max read-ahead pages. | |
27 | * Note that MAX_RA_PAGES is rounded down, while MIN_RA_PAGES is rounded up. | |
28 | * The latter is necessary for systems with large page size(i.e. 64k). | |
29 | */ | |
30 | #define MAX_RA_PAGES (VM_MAX_READAHEAD*1024 / PAGE_CACHE_SIZE) | |
31 | #define MIN_RA_PAGES DIV_ROUND_UP(VM_MIN_READAHEAD*1024, PAGE_CACHE_SIZE) | |
32 | ||
1da177e4 | 33 | struct backing_dev_info default_backing_dev_info = { |
f615bfca | 34 | .ra_pages = MAX_RA_PAGES, |
1da177e4 LT |
35 | .state = 0, |
36 | .capabilities = BDI_CAP_MAP_COPY, | |
37 | .unplug_io_fn = default_unplug_io_fn, | |
38 | }; | |
39 | EXPORT_SYMBOL_GPL(default_backing_dev_info); | |
40 | ||
41 | /* | |
42 | * Initialise a struct file's readahead state. Assumes that the caller has | |
43 | * memset *ra to zero. | |
44 | */ | |
45 | void | |
46 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) | |
47 | { | |
48 | ra->ra_pages = mapping->backing_dev_info->ra_pages; | |
6ce745ed | 49 | ra->prev_index = -1; |
1da177e4 | 50 | } |
d41cc702 | 51 | EXPORT_SYMBOL_GPL(file_ra_state_init); |
1da177e4 | 52 | |
1da177e4 LT |
53 | #define list_to_page(head) (list_entry((head)->prev, struct page, lru)) |
54 | ||
55 | /** | |
bd40cdda | 56 | * read_cache_pages - populate an address space with some pages & start reads against them |
1da177e4 LT |
57 | * @mapping: the address_space |
58 | * @pages: The address of a list_head which contains the target pages. These | |
59 | * pages have their ->index populated and are otherwise uninitialised. | |
60 | * @filler: callback routine for filling a single page. | |
61 | * @data: private data for the callback routine. | |
62 | * | |
63 | * Hides the details of the LRU cache etc from the filesystems. | |
64 | */ | |
65 | int read_cache_pages(struct address_space *mapping, struct list_head *pages, | |
66 | int (*filler)(void *, struct page *), void *data) | |
67 | { | |
68 | struct page *page; | |
69 | struct pagevec lru_pvec; | |
70 | int ret = 0; | |
71 | ||
72 | pagevec_init(&lru_pvec, 0); | |
73 | ||
74 | while (!list_empty(pages)) { | |
75 | page = list_to_page(pages); | |
76 | list_del(&page->lru); | |
77 | if (add_to_page_cache(page, mapping, page->index, GFP_KERNEL)) { | |
78 | page_cache_release(page); | |
79 | continue; | |
80 | } | |
81 | ret = filler(data, page); | |
82 | if (!pagevec_add(&lru_pvec, page)) | |
83 | __pagevec_lru_add(&lru_pvec); | |
84 | if (ret) { | |
38da288b | 85 | put_pages_list(pages); |
1da177e4 LT |
86 | break; |
87 | } | |
8bde37f0 | 88 | task_io_account_read(PAGE_CACHE_SIZE); |
1da177e4 LT |
89 | } |
90 | pagevec_lru_add(&lru_pvec); | |
91 | return ret; | |
92 | } | |
93 | ||
94 | EXPORT_SYMBOL(read_cache_pages); | |
95 | ||
96 | static int read_pages(struct address_space *mapping, struct file *filp, | |
97 | struct list_head *pages, unsigned nr_pages) | |
98 | { | |
99 | unsigned page_idx; | |
100 | struct pagevec lru_pvec; | |
994fc28c | 101 | int ret; |
1da177e4 LT |
102 | |
103 | if (mapping->a_ops->readpages) { | |
104 | ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages); | |
029e332e OH |
105 | /* Clean up the remaining pages */ |
106 | put_pages_list(pages); | |
1da177e4 LT |
107 | goto out; |
108 | } | |
109 | ||
110 | pagevec_init(&lru_pvec, 0); | |
111 | for (page_idx = 0; page_idx < nr_pages; page_idx++) { | |
112 | struct page *page = list_to_page(pages); | |
113 | list_del(&page->lru); | |
114 | if (!add_to_page_cache(page, mapping, | |
115 | page->index, GFP_KERNEL)) { | |
9f1a3cfc ZB |
116 | mapping->a_ops->readpage(filp, page); |
117 | if (!pagevec_add(&lru_pvec, page)) | |
118 | __pagevec_lru_add(&lru_pvec); | |
119 | } else | |
120 | page_cache_release(page); | |
1da177e4 LT |
121 | } |
122 | pagevec_lru_add(&lru_pvec); | |
994fc28c | 123 | ret = 0; |
1da177e4 LT |
124 | out: |
125 | return ret; | |
126 | } | |
127 | ||
1da177e4 LT |
128 | /* |
129 | * do_page_cache_readahead actually reads a chunk of disk. It allocates all | |
130 | * the pages first, then submits them all for I/O. This avoids the very bad | |
131 | * behaviour which would occur if page allocations are causing VM writeback. | |
132 | * We really don't want to intermingle reads and writes like that. | |
133 | * | |
134 | * Returns the number of pages requested, or the maximum amount of I/O allowed. | |
135 | * | |
136 | * do_page_cache_readahead() returns -1 if it encountered request queue | |
137 | * congestion. | |
138 | */ | |
139 | static int | |
140 | __do_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
46fc3e7b FW |
141 | pgoff_t offset, unsigned long nr_to_read, |
142 | unsigned long lookahead_size) | |
1da177e4 LT |
143 | { |
144 | struct inode *inode = mapping->host; | |
145 | struct page *page; | |
146 | unsigned long end_index; /* The last page we want to read */ | |
147 | LIST_HEAD(page_pool); | |
148 | int page_idx; | |
149 | int ret = 0; | |
150 | loff_t isize = i_size_read(inode); | |
151 | ||
152 | if (isize == 0) | |
153 | goto out; | |
154 | ||
46fc3e7b | 155 | end_index = ((isize - 1) >> PAGE_CACHE_SHIFT); |
1da177e4 LT |
156 | |
157 | /* | |
158 | * Preallocate as many pages as we will need. | |
159 | */ | |
160 | read_lock_irq(&mapping->tree_lock); | |
161 | for (page_idx = 0; page_idx < nr_to_read; page_idx++) { | |
7361f4d8 | 162 | pgoff_t page_offset = offset + page_idx; |
c743d96b | 163 | |
1da177e4 LT |
164 | if (page_offset > end_index) |
165 | break; | |
166 | ||
167 | page = radix_tree_lookup(&mapping->page_tree, page_offset); | |
168 | if (page) | |
169 | continue; | |
170 | ||
171 | read_unlock_irq(&mapping->tree_lock); | |
172 | page = page_cache_alloc_cold(mapping); | |
173 | read_lock_irq(&mapping->tree_lock); | |
174 | if (!page) | |
175 | break; | |
176 | page->index = page_offset; | |
177 | list_add(&page->lru, &page_pool); | |
46fc3e7b FW |
178 | if (page_idx == nr_to_read - lookahead_size) |
179 | SetPageReadahead(page); | |
1da177e4 LT |
180 | ret++; |
181 | } | |
182 | read_unlock_irq(&mapping->tree_lock); | |
183 | ||
184 | /* | |
185 | * Now start the IO. We ignore I/O errors - if the page is not | |
186 | * uptodate then the caller will launch readpage again, and | |
187 | * will then handle the error. | |
188 | */ | |
189 | if (ret) | |
190 | read_pages(mapping, filp, &page_pool, ret); | |
191 | BUG_ON(!list_empty(&page_pool)); | |
192 | out: | |
193 | return ret; | |
194 | } | |
195 | ||
196 | /* | |
197 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much | |
198 | * memory at once. | |
199 | */ | |
200 | int force_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
7361f4d8 | 201 | pgoff_t offset, unsigned long nr_to_read) |
1da177e4 LT |
202 | { |
203 | int ret = 0; | |
204 | ||
205 | if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages)) | |
206 | return -EINVAL; | |
207 | ||
208 | while (nr_to_read) { | |
209 | int err; | |
210 | ||
211 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE; | |
212 | ||
213 | if (this_chunk > nr_to_read) | |
214 | this_chunk = nr_to_read; | |
215 | err = __do_page_cache_readahead(mapping, filp, | |
46fc3e7b | 216 | offset, this_chunk, 0); |
1da177e4 LT |
217 | if (err < 0) { |
218 | ret = err; | |
219 | break; | |
220 | } | |
221 | ret += err; | |
222 | offset += this_chunk; | |
223 | nr_to_read -= this_chunk; | |
224 | } | |
225 | return ret; | |
226 | } | |
227 | ||
1da177e4 LT |
228 | /* |
229 | * This version skips the IO if the queue is read-congested, and will tell the | |
230 | * block layer to abandon the readahead if request allocation would block. | |
231 | * | |
232 | * force_page_cache_readahead() will ignore queue congestion and will block on | |
233 | * request queues. | |
234 | */ | |
235 | int do_page_cache_readahead(struct address_space *mapping, struct file *filp, | |
7361f4d8 | 236 | pgoff_t offset, unsigned long nr_to_read) |
1da177e4 LT |
237 | { |
238 | if (bdi_read_congested(mapping->backing_dev_info)) | |
239 | return -1; | |
240 | ||
46fc3e7b | 241 | return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0); |
1da177e4 LT |
242 | } |
243 | ||
1da177e4 LT |
244 | /* |
245 | * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a | |
246 | * sensible upper limit. | |
247 | */ | |
248 | unsigned long max_sane_readahead(unsigned long nr) | |
249 | { | |
05a0416b CL |
250 | return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE) |
251 | + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2); | |
1da177e4 | 252 | } |
5ce1110b FW |
253 | |
254 | /* | |
255 | * Submit IO for the read-ahead request in file_ra_state. | |
256 | */ | |
f9acc8c7 | 257 | static unsigned long ra_submit(struct file_ra_state *ra, |
5ce1110b FW |
258 | struct address_space *mapping, struct file *filp) |
259 | { | |
5ce1110b FW |
260 | int actual; |
261 | ||
5ce1110b | 262 | actual = __do_page_cache_readahead(mapping, filp, |
f9acc8c7 | 263 | ra->start, ra->size, ra->async_size); |
5ce1110b FW |
264 | |
265 | return actual; | |
266 | } | |
122a21d1 | 267 | |
c743d96b FW |
268 | /* |
269 | * Set the initial window size, round to next power of 2 and square | |
270 | * for small size, x 4 for medium, and x 2 for large | |
271 | * for 128k (32 page) max ra | |
272 | * 1-8 page = 32k initial, > 8 page = 128k initial | |
273 | */ | |
274 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max) | |
275 | { | |
276 | unsigned long newsize = roundup_pow_of_two(size); | |
277 | ||
278 | if (newsize <= max / 32) | |
279 | newsize = newsize * 4; | |
280 | else if (newsize <= max / 4) | |
281 | newsize = newsize * 2; | |
282 | else | |
283 | newsize = max; | |
284 | ||
285 | return newsize; | |
286 | } | |
287 | ||
122a21d1 FW |
288 | /* |
289 | * Get the previous window size, ramp it up, and | |
290 | * return it as the new window size. | |
291 | */ | |
c743d96b | 292 | static unsigned long get_next_ra_size(struct file_ra_state *ra, |
122a21d1 FW |
293 | unsigned long max) |
294 | { | |
f9acc8c7 | 295 | unsigned long cur = ra->size; |
122a21d1 FW |
296 | unsigned long newsize; |
297 | ||
298 | if (cur < max / 16) | |
c743d96b | 299 | newsize = 4 * cur; |
122a21d1 | 300 | else |
c743d96b | 301 | newsize = 2 * cur; |
122a21d1 FW |
302 | |
303 | return min(newsize, max); | |
304 | } | |
305 | ||
306 | /* | |
307 | * On-demand readahead design. | |
308 | * | |
309 | * The fields in struct file_ra_state represent the most-recently-executed | |
310 | * readahead attempt: | |
311 | * | |
f9acc8c7 FW |
312 | * |<----- async_size ---------| |
313 | * |------------------- size -------------------->| | |
314 | * |==================#===========================| | |
315 | * ^start ^page marked with PG_readahead | |
122a21d1 FW |
316 | * |
317 | * To overlap application thinking time and disk I/O time, we do | |
318 | * `readahead pipelining': Do not wait until the application consumed all | |
319 | * readahead pages and stalled on the missing page at readahead_index; | |
f9acc8c7 FW |
320 | * Instead, submit an asynchronous readahead I/O as soon as there are |
321 | * only async_size pages left in the readahead window. Normally async_size | |
322 | * will be equal to size, for maximum pipelining. | |
122a21d1 FW |
323 | * |
324 | * In interleaved sequential reads, concurrent streams on the same fd can | |
325 | * be invalidating each other's readahead state. So we flag the new readahead | |
f9acc8c7 | 326 | * page at (start+size-async_size) with PG_readahead, and use it as readahead |
122a21d1 FW |
327 | * indicator. The flag won't be set on already cached pages, to avoid the |
328 | * readahead-for-nothing fuss, saving pointless page cache lookups. | |
329 | * | |
330 | * prev_index tracks the last visited page in the _previous_ read request. | |
331 | * It should be maintained by the caller, and will be used for detecting | |
332 | * small random reads. Note that the readahead algorithm checks loosely | |
333 | * for sequential patterns. Hence interleaved reads might be served as | |
334 | * sequential ones. | |
335 | * | |
336 | * There is a special-case: if the first page which the application tries to | |
337 | * read happens to be the first page of the file, it is assumed that a linear | |
338 | * read is about to happen and the window is immediately set to the initial size | |
339 | * based on I/O request size and the max_readahead. | |
340 | * | |
341 | * The code ramps up the readahead size aggressively at first, but slow down as | |
342 | * it approaches max_readhead. | |
343 | */ | |
344 | ||
345 | /* | |
346 | * A minimal readahead algorithm for trivial sequential/random reads. | |
347 | */ | |
348 | static unsigned long | |
349 | ondemand_readahead(struct address_space *mapping, | |
350 | struct file_ra_state *ra, struct file *filp, | |
cf914a7d | 351 | bool hit_readahead_marker, pgoff_t offset, |
122a21d1 FW |
352 | unsigned long req_size) |
353 | { | |
354 | unsigned long max; /* max readahead pages */ | |
122a21d1 FW |
355 | int sequential; |
356 | ||
357 | max = ra->ra_pages; | |
358 | sequential = (offset - ra->prev_index <= 1UL) || (req_size > max); | |
359 | ||
360 | /* | |
f9acc8c7 | 361 | * It's the expected callback offset, assume sequential access. |
122a21d1 FW |
362 | * Ramp up sizes, and push forward the readahead window. |
363 | */ | |
f9acc8c7 FW |
364 | if (offset && (offset == (ra->start + ra->size - ra->async_size) || |
365 | offset == (ra->start + ra->size))) { | |
366 | ra->start += ra->size; | |
367 | ra->size = get_next_ra_size(ra, max); | |
368 | ra->async_size = ra->size; | |
369 | goto readit; | |
122a21d1 FW |
370 | } |
371 | ||
372 | /* | |
373 | * Standalone, small read. | |
374 | * Read as is, and do not pollute the readahead state. | |
375 | */ | |
cf914a7d | 376 | if (!hit_readahead_marker && !sequential) { |
122a21d1 FW |
377 | return __do_page_cache_readahead(mapping, filp, |
378 | offset, req_size, 0); | |
379 | } | |
380 | ||
381 | /* | |
382 | * It may be one of | |
383 | * - first read on start of file | |
384 | * - sequential cache miss | |
385 | * - oversize random read | |
386 | * Start readahead for it. | |
387 | */ | |
f9acc8c7 FW |
388 | ra->start = offset; |
389 | ra->size = get_init_ra_size(req_size, max); | |
390 | ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; | |
122a21d1 FW |
391 | |
392 | /* | |
f9acc8c7 | 393 | * Hit on a marked page without valid readahead state. |
122a21d1 FW |
394 | * E.g. interleaved reads. |
395 | * Not knowing its readahead pos/size, bet on the minimal possible one. | |
396 | */ | |
cf914a7d | 397 | if (hit_readahead_marker) { |
f9acc8c7 FW |
398 | ra->start++; |
399 | ra->size = get_next_ra_size(ra, max); | |
122a21d1 FW |
400 | } |
401 | ||
f9acc8c7 | 402 | readit: |
122a21d1 FW |
403 | return ra_submit(ra, mapping, filp); |
404 | } | |
405 | ||
406 | /** | |
cf914a7d | 407 | * page_cache_sync_readahead - generic file readahead |
122a21d1 FW |
408 | * @mapping: address_space which holds the pagecache and I/O vectors |
409 | * @ra: file_ra_state which holds the readahead state | |
410 | * @filp: passed on to ->readpage() and ->readpages() | |
cf914a7d | 411 | * @offset: start offset into @mapping, in pagecache page-sized units |
122a21d1 | 412 | * @req_size: hint: total size of the read which the caller is performing in |
cf914a7d | 413 | * pagecache pages |
122a21d1 | 414 | * |
cf914a7d RR |
415 | * page_cache_sync_readahead() should be called when a cache miss happened: |
416 | * it will submit the read. The readahead logic may decide to piggyback more | |
417 | * pages onto the read request if access patterns suggest it will improve | |
418 | * performance. | |
122a21d1 | 419 | */ |
cf914a7d RR |
420 | void page_cache_sync_readahead(struct address_space *mapping, |
421 | struct file_ra_state *ra, struct file *filp, | |
422 | pgoff_t offset, unsigned long req_size) | |
122a21d1 FW |
423 | { |
424 | /* no read-ahead */ | |
425 | if (!ra->ra_pages) | |
cf914a7d RR |
426 | return; |
427 | ||
428 | /* do read-ahead */ | |
429 | ondemand_readahead(mapping, ra, filp, false, offset, req_size); | |
430 | } | |
431 | EXPORT_SYMBOL_GPL(page_cache_sync_readahead); | |
432 | ||
433 | /** | |
434 | * page_cache_async_readahead - file readahead for marked pages | |
435 | * @mapping: address_space which holds the pagecache and I/O vectors | |
436 | * @ra: file_ra_state which holds the readahead state | |
437 | * @filp: passed on to ->readpage() and ->readpages() | |
438 | * @page: the page at @offset which has the PG_readahead flag set | |
439 | * @offset: start offset into @mapping, in pagecache page-sized units | |
440 | * @req_size: hint: total size of the read which the caller is performing in | |
441 | * pagecache pages | |
442 | * | |
443 | * page_cache_async_ondemand() should be called when a page is used which | |
444 | * has the PG_readahead flag: this is a marker to suggest that the application | |
445 | * has used up enough of the readahead window that we should start pulling in | |
446 | * more pages. */ | |
447 | void | |
448 | page_cache_async_readahead(struct address_space *mapping, | |
449 | struct file_ra_state *ra, struct file *filp, | |
450 | struct page *page, pgoff_t offset, | |
451 | unsigned long req_size) | |
452 | { | |
453 | /* no read-ahead */ | |
454 | if (!ra->ra_pages) | |
455 | return; | |
456 | ||
457 | /* | |
458 | * Same bit is used for PG_readahead and PG_reclaim. | |
459 | */ | |
460 | if (PageWriteback(page)) | |
461 | return; | |
462 | ||
463 | ClearPageReadahead(page); | |
464 | ||
465 | /* | |
466 | * Defer asynchronous read-ahead on IO congestion. | |
467 | */ | |
468 | if (bdi_read_congested(mapping->backing_dev_info)) | |
469 | return; | |
122a21d1 FW |
470 | |
471 | /* do read-ahead */ | |
cf914a7d | 472 | ondemand_readahead(mapping, ra, filp, true, offset, req_size); |
122a21d1 | 473 | } |
cf914a7d | 474 | EXPORT_SYMBOL_GPL(page_cache_async_readahead); |