Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * mm/readahead.c - address_space-level file readahead. | |
4 | * | |
5 | * Copyright (C) 2002, Linus Torvalds | |
6 | * | |
e1f8e874 | 7 | * 09Apr2002 Andrew Morton |
1da177e4 LT |
8 | * Initial version. |
9 | */ | |
10 | ||
84dacdbd N |
11 | /** |
12 | * DOC: Readahead Overview | |
13 | * | |
14 | * Readahead is used to read content into the page cache before it is | |
15 | * explicitly requested by the application. Readahead only ever | |
1e470280 MWO |
16 | * attempts to read folios that are not yet in the page cache. If a |
17 | * folio is present but not up-to-date, readahead will not try to read | |
5efe7448 | 18 | * it. In that case a simple ->read_folio() will be requested. |
84dacdbd N |
19 | * |
20 | * Readahead is triggered when an application read request (whether a | |
1e470280 | 21 | * system call or a page fault) finds that the requested folio is not in |
84dacdbd | 22 | * the page cache, or that it is in the page cache and has the |
1e470280 MWO |
23 | * readahead flag set. This flag indicates that the folio was read |
24 | * as part of a previous readahead request and now that it has been | |
25 | * accessed, it is time for the next readahead. | |
84dacdbd N |
26 | * |
27 | * Each readahead request is partly synchronous read, and partly async | |
1e470280 MWO |
28 | * readahead. This is reflected in the struct file_ra_state which |
29 | * contains ->size being the total number of pages, and ->async_size | |
30 | * which is the number of pages in the async section. The readahead | |
31 | * flag will be set on the first folio in this async section to trigger | |
32 | * a subsequent readahead. Once a series of sequential reads has been | |
84dacdbd | 33 | * established, there should be no need for a synchronous component and |
1e470280 | 34 | * all readahead request will be fully asynchronous. |
84dacdbd | 35 | * |
1e470280 MWO |
36 | * When either of the triggers causes a readahead, three numbers need |
37 | * to be determined: the start of the region to read, the size of the | |
38 | * region, and the size of the async tail. | |
84dacdbd N |
39 | * |
40 | * The start of the region is simply the first page address at or after | |
41 | * the accessed address, which is not currently populated in the page | |
42 | * cache. This is found with a simple search in the page cache. | |
43 | * | |
44 | * The size of the async tail is determined by subtracting the size that | |
45 | * was explicitly requested from the determined request size, unless | |
46 | * this would be less than zero - then zero is used. NOTE THIS | |
47 | * CALCULATION IS WRONG WHEN THE START OF THE REGION IS NOT THE ACCESSED | |
1e470280 | 48 | * PAGE. ALSO THIS CALCULATION IS NOT USED CONSISTENTLY. |
84dacdbd N |
49 | * |
50 | * The size of the region is normally determined from the size of the | |
51 | * previous readahead which loaded the preceding pages. This may be | |
52 | * discovered from the struct file_ra_state for simple sequential reads, | |
53 | * or from examining the state of the page cache when multiple | |
54 | * sequential reads are interleaved. Specifically: where the readahead | |
1e470280 | 55 | * was triggered by the readahead flag, the size of the previous |
84dacdbd N |
56 | * readahead is assumed to be the number of pages from the triggering |
57 | * page to the start of the new readahead. In these cases, the size of | |
58 | * the previous readahead is scaled, often doubled, for the new | |
59 | * readahead, though see get_next_ra_size() for details. | |
60 | * | |
61 | * If the size of the previous read cannot be determined, the number of | |
62 | * preceding pages in the page cache is used to estimate the size of | |
63 | * a previous read. This estimate could easily be misled by random | |
64 | * reads being coincidentally adjacent, so it is ignored unless it is | |
65 | * larger than the current request, and it is not scaled up, unless it | |
66 | * is at the start of file. | |
67 | * | |
1e470280 | 68 | * In general readahead is accelerated at the start of the file, as |
84dacdbd | 69 | * reads from there are often sequential. There are other minor |
1e470280 | 70 | * adjustments to the readahead size in various special cases and these |
84dacdbd N |
71 | * are best discovered by reading the code. |
72 | * | |
1e470280 MWO |
73 | * The above calculation, based on the previous readahead size, |
74 | * determines the size of the readahead, to which any requested read | |
75 | * size may be added. | |
84dacdbd N |
76 | * |
77 | * Readahead requests are sent to the filesystem using the ->readahead() | |
78 | * address space operation, for which mpage_readahead() is a canonical | |
79 | * implementation. ->readahead() should normally initiate reads on all | |
1e470280 | 80 | * folios, but may fail to read any or all folios without causing an I/O |
5efe7448 | 81 | * error. The page cache reading code will issue a ->read_folio() request |
1e470280 | 82 | * for any folio which ->readahead() did not read, and only an error |
84dacdbd N |
83 | * from this will be final. |
84 | * | |
1e470280 MWO |
85 | * ->readahead() will generally call readahead_folio() repeatedly to get |
86 | * each folio from those prepared for readahead. It may fail to read a | |
87 | * folio by: | |
84dacdbd | 88 | * |
1e470280 MWO |
89 | * * not calling readahead_folio() sufficiently many times, effectively |
90 | * ignoring some folios, as might be appropriate if the path to | |
84dacdbd N |
91 | * storage is congested. |
92 | * | |
1e470280 | 93 | * * failing to actually submit a read request for a given folio, |
84dacdbd N |
94 | * possibly due to insufficient resources, or |
95 | * | |
96 | * * getting an error during subsequent processing of a request. | |
97 | * | |
1e470280 MWO |
98 | * In the last two cases, the folio should be unlocked by the filesystem |
99 | * to indicate that the read attempt has failed. In the first case the | |
100 | * folio will be unlocked by the VFS. | |
84dacdbd | 101 | * |
1e470280 | 102 | * Those folios not in the final ``async_size`` of the request should be |
84dacdbd N |
103 | * considered to be important and ->readahead() should not fail them due |
104 | * to congestion or temporary resource unavailability, but should wait | |
105 | * for necessary resources (e.g. memory or indexing information) to | |
1e470280 | 106 | * become available. Folios in the final ``async_size`` may be |
84dacdbd | 107 | * considered less urgent and failure to read them is more acceptable. |
1e470280 MWO |
108 | * In this case it is best to use filemap_remove_folio() to remove the |
109 | * folios from the page cache as is automatically done for folios that | |
110 | * were not fetched with readahead_folio(). This will allow a | |
111 | * subsequent synchronous readahead request to try them again. If they | |
9fd472af | 112 | * are left in the page cache, then they will be read individually using |
5efe7448 | 113 | * ->read_folio() which may be less efficient. |
84dacdbd N |
114 | */ |
115 | ||
c97ab271 | 116 | #include <linux/blkdev.h> |
1da177e4 | 117 | #include <linux/kernel.h> |
11bd969f | 118 | #include <linux/dax.h> |
5a0e3ad6 | 119 | #include <linux/gfp.h> |
b95f1b31 | 120 | #include <linux/export.h> |
1da177e4 | 121 | #include <linux/backing-dev.h> |
8bde37f0 | 122 | #include <linux/task_io_accounting_ops.h> |
1da177e4 | 123 | #include <linux/pagevec.h> |
f5ff8422 | 124 | #include <linux/pagemap.h> |
782182e5 CW |
125 | #include <linux/syscalls.h> |
126 | #include <linux/file.h> | |
d72ee911 | 127 | #include <linux/mm_inline.h> |
ca47e8c7 | 128 | #include <linux/blk-cgroup.h> |
3d8f7615 | 129 | #include <linux/fadvise.h> |
f2c817be | 130 | #include <linux/sched/mm.h> |
1da177e4 | 131 | |
29f175d1 FF |
132 | #include "internal.h" |
133 | ||
1da177e4 LT |
134 | /* |
135 | * Initialise a struct file's readahead state. Assumes that the caller has | |
136 | * memset *ra to zero. | |
137 | */ | |
138 | void | |
139 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) | |
140 | { | |
de1414a6 | 141 | ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages; |
f4e6b498 | 142 | ra->prev_pos = -1; |
1da177e4 | 143 | } |
d41cc702 | 144 | EXPORT_SYMBOL_GPL(file_ra_state_init); |
1da177e4 | 145 | |
b4e089d7 | 146 | static void read_pages(struct readahead_control *rac) |
1da177e4 | 147 | { |
a4d96536 | 148 | const struct address_space_operations *aops = rac->mapping->a_ops; |
a42634a6 | 149 | struct folio *folio; |
5b417b18 | 150 | struct blk_plug plug; |
1da177e4 | 151 | |
a4d96536 | 152 | if (!readahead_count(rac)) |
b4e089d7 | 153 | return; |
ad4ae1c7 | 154 | |
5b417b18 JA |
155 | blk_start_plug(&plug); |
156 | ||
8151b4c8 MWO |
157 | if (aops->readahead) { |
158 | aops->readahead(rac); | |
9fd472af | 159 | /* |
a42634a6 | 160 | * Clean up the remaining folios. The sizes in ->ra |
1e470280 | 161 | * may be used to size the next readahead, so make sure |
9fd472af N |
162 | * they accurately reflect what happened. |
163 | */ | |
a42634a6 MWO |
164 | while ((folio = readahead_folio(rac)) != NULL) { |
165 | unsigned long nr = folio_nr_pages(folio); | |
166 | ||
167 | rac->ra->size -= nr; | |
168 | if (rac->ra->async_size >= nr) { | |
169 | rac->ra->async_size -= nr; | |
170 | filemap_remove_folio(folio); | |
9fd472af | 171 | } |
a42634a6 | 172 | folio_unlock(folio); |
8151b4c8 | 173 | } |
c1f6925e | 174 | } else { |
5efe7448 | 175 | while ((folio = readahead_folio(rac)) != NULL) |
7e0a1265 | 176 | aops->read_folio(rac->file, folio); |
1da177e4 | 177 | } |
5b417b18 | 178 | |
5b417b18 | 179 | blk_finish_plug(&plug); |
ad4ae1c7 | 180 | |
c1f6925e | 181 | BUG_ON(readahead_count(rac)); |
1da177e4 LT |
182 | } |
183 | ||
2c684234 | 184 | /** |
73bb49da MWO |
185 | * page_cache_ra_unbounded - Start unchecked readahead. |
186 | * @ractl: Readahead control. | |
2c684234 MWO |
187 | * @nr_to_read: The number of pages to read. |
188 | * @lookahead_size: Where to start the next readahead. | |
189 | * | |
190 | * This function is for filesystems to call when they want to start | |
191 | * readahead beyond a file's stated i_size. This is almost certainly | |
192 | * not the function you want to call. Use page_cache_async_readahead() | |
193 | * or page_cache_sync_readahead() instead. | |
194 | * | |
195 | * Context: File is referenced by caller. Mutexes may be held by caller. | |
196 | * May sleep, but will not reenter filesystem to reclaim memory. | |
1da177e4 | 197 | */ |
73bb49da MWO |
198 | void page_cache_ra_unbounded(struct readahead_control *ractl, |
199 | unsigned long nr_to_read, unsigned long lookahead_size) | |
1da177e4 | 200 | { |
73bb49da MWO |
201 | struct address_space *mapping = ractl->mapping; |
202 | unsigned long index = readahead_index(ractl); | |
8a5c743e | 203 | gfp_t gfp_mask = readahead_gfp_mask(mapping); |
c2c7ad74 | 204 | unsigned long i; |
1da177e4 | 205 | |
f2c817be MWO |
206 | /* |
207 | * Partway through the readahead operation, we will have added | |
208 | * locked pages to the page cache, but will not yet have submitted | |
209 | * them for I/O. Adding another page may need to allocate memory, | |
210 | * which can trigger memory reclaim. Telling the VM we're in | |
211 | * the middle of a filesystem operation will cause it to not | |
212 | * touch file-backed pages, preventing a deadlock. Most (all?) | |
213 | * filesystems already specify __GFP_NOFS in their mapping's | |
214 | * gfp_mask, but let's be explicit here. | |
215 | */ | |
216 | unsigned int nofs = memalloc_nofs_save(); | |
217 | ||
730633f0 | 218 | filemap_invalidate_lock_shared(mapping); |
1da177e4 LT |
219 | /* |
220 | * Preallocate as many pages as we will need. | |
221 | */ | |
c2c7ad74 | 222 | for (i = 0; i < nr_to_read; i++) { |
0387df1d | 223 | struct folio *folio = xa_load(&mapping->i_pages, index + i); |
1da177e4 | 224 | |
0387df1d | 225 | if (folio && !xa_is_value(folio)) { |
b3751e6a | 226 | /* |
2d8163e4 MWO |
227 | * Page already present? Kick off the current batch |
228 | * of contiguous pages before continuing with the | |
229 | * next batch. This page may be the one we would | |
230 | * have intended to mark as Readahead, but we don't | |
231 | * have a stable reference to this page, and it's | |
232 | * not worth getting one just for that. | |
b3751e6a | 233 | */ |
b4e089d7 CH |
234 | read_pages(ractl); |
235 | ractl->_index++; | |
f615bd5c | 236 | i = ractl->_index + ractl->_nr_pages - index - 1; |
1da177e4 | 237 | continue; |
b3751e6a | 238 | } |
1da177e4 | 239 | |
0387df1d MWO |
240 | folio = filemap_alloc_folio(gfp_mask, 0); |
241 | if (!folio) | |
1da177e4 | 242 | break; |
704528d8 | 243 | if (filemap_add_folio(mapping, folio, index + i, |
c1f6925e | 244 | gfp_mask) < 0) { |
0387df1d | 245 | folio_put(folio); |
b4e089d7 CH |
246 | read_pages(ractl); |
247 | ractl->_index++; | |
f615bd5c | 248 | i = ractl->_index + ractl->_nr_pages - index - 1; |
c1f6925e MWO |
249 | continue; |
250 | } | |
c2c7ad74 | 251 | if (i == nr_to_read - lookahead_size) |
0387df1d | 252 | folio_set_readahead(folio); |
73bb49da | 253 | ractl->_nr_pages++; |
1da177e4 | 254 | } |
1da177e4 LT |
255 | |
256 | /* | |
7e0a1265 MWO |
257 | * Now start the IO. We ignore I/O errors - if the folio is not |
258 | * uptodate then the caller will launch read_folio again, and | |
1da177e4 LT |
259 | * will then handle the error. |
260 | */ | |
b4e089d7 | 261 | read_pages(ractl); |
730633f0 | 262 | filemap_invalidate_unlock_shared(mapping); |
f2c817be | 263 | memalloc_nofs_restore(nofs); |
1da177e4 | 264 | } |
73bb49da | 265 | EXPORT_SYMBOL_GPL(page_cache_ra_unbounded); |
2c684234 MWO |
266 | |
267 | /* | |
8238287e | 268 | * do_page_cache_ra() actually reads a chunk of disk. It allocates |
2c684234 MWO |
269 | * the pages first, then submits them for I/O. This avoids the very bad |
270 | * behaviour which would occur if page allocations are causing VM writeback. | |
271 | * We really don't want to intermingle reads and writes like that. | |
272 | */ | |
56a4d67c | 273 | static void do_page_cache_ra(struct readahead_control *ractl, |
8238287e | 274 | unsigned long nr_to_read, unsigned long lookahead_size) |
2c684234 | 275 | { |
8238287e MWO |
276 | struct inode *inode = ractl->mapping->host; |
277 | unsigned long index = readahead_index(ractl); | |
2c684234 MWO |
278 | loff_t isize = i_size_read(inode); |
279 | pgoff_t end_index; /* The last page we want to read */ | |
280 | ||
281 | if (isize == 0) | |
282 | return; | |
283 | ||
284 | end_index = (isize - 1) >> PAGE_SHIFT; | |
285 | if (index > end_index) | |
286 | return; | |
287 | /* Don't read past the page containing the last byte of the file */ | |
288 | if (nr_to_read > end_index - index) | |
289 | nr_to_read = end_index - index + 1; | |
290 | ||
8238287e | 291 | page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size); |
2c684234 | 292 | } |
1da177e4 LT |
293 | |
294 | /* | |
295 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much | |
296 | * memory at once. | |
297 | */ | |
7b3df3b9 | 298 | void force_page_cache_ra(struct readahead_control *ractl, |
fcd9ae4f | 299 | unsigned long nr_to_read) |
1da177e4 | 300 | { |
7b3df3b9 | 301 | struct address_space *mapping = ractl->mapping; |
fcd9ae4f | 302 | struct file_ra_state *ra = ractl->ra; |
9491ae4a | 303 | struct backing_dev_info *bdi = inode_to_bdi(mapping->host); |
7b3df3b9 | 304 | unsigned long max_pages, index; |
9491ae4a | 305 | |
7e0a1265 | 306 | if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead)) |
9a42823a | 307 | return; |
1da177e4 | 308 | |
9491ae4a JA |
309 | /* |
310 | * If the request exceeds the readahead window, allow the read to | |
311 | * be up to the optimal hardware IO size | |
312 | */ | |
7b3df3b9 | 313 | index = readahead_index(ractl); |
9491ae4a | 314 | max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages); |
7b3df3b9 | 315 | nr_to_read = min_t(unsigned long, nr_to_read, max_pages); |
1da177e4 | 316 | while (nr_to_read) { |
09cbfeaf | 317 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE; |
1da177e4 LT |
318 | |
319 | if (this_chunk > nr_to_read) | |
320 | this_chunk = nr_to_read; | |
7b3df3b9 DH |
321 | ractl->_index = index; |
322 | do_page_cache_ra(ractl, this_chunk, 0); | |
58d5640e | 323 | |
08eb9658 | 324 | index += this_chunk; |
1da177e4 LT |
325 | nr_to_read -= this_chunk; |
326 | } | |
1da177e4 LT |
327 | } |
328 | ||
c743d96b FW |
329 | /* |
330 | * Set the initial window size, round to next power of 2 and square | |
331 | * for small size, x 4 for medium, and x 2 for large | |
332 | * for 128k (32 page) max ra | |
fb25a77d | 333 | * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial |
c743d96b FW |
334 | */ |
335 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max) | |
336 | { | |
337 | unsigned long newsize = roundup_pow_of_two(size); | |
338 | ||
339 | if (newsize <= max / 32) | |
340 | newsize = newsize * 4; | |
341 | else if (newsize <= max / 4) | |
342 | newsize = newsize * 2; | |
343 | else | |
344 | newsize = max; | |
345 | ||
346 | return newsize; | |
347 | } | |
348 | ||
122a21d1 FW |
349 | /* |
350 | * Get the previous window size, ramp it up, and | |
351 | * return it as the new window size. | |
352 | */ | |
c743d96b | 353 | static unsigned long get_next_ra_size(struct file_ra_state *ra, |
20ff1c95 | 354 | unsigned long max) |
122a21d1 | 355 | { |
f9acc8c7 | 356 | unsigned long cur = ra->size; |
122a21d1 FW |
357 | |
358 | if (cur < max / 16) | |
20ff1c95 GX |
359 | return 4 * cur; |
360 | if (cur <= max / 2) | |
361 | return 2 * cur; | |
362 | return max; | |
122a21d1 FW |
363 | } |
364 | ||
365 | /* | |
366 | * On-demand readahead design. | |
367 | * | |
368 | * The fields in struct file_ra_state represent the most-recently-executed | |
369 | * readahead attempt: | |
370 | * | |
f9acc8c7 FW |
371 | * |<----- async_size ---------| |
372 | * |------------------- size -------------------->| | |
373 | * |==================#===========================| | |
374 | * ^start ^page marked with PG_readahead | |
122a21d1 FW |
375 | * |
376 | * To overlap application thinking time and disk I/O time, we do | |
377 | * `readahead pipelining': Do not wait until the application consumed all | |
378 | * readahead pages and stalled on the missing page at readahead_index; | |
f9acc8c7 FW |
379 | * Instead, submit an asynchronous readahead I/O as soon as there are |
380 | * only async_size pages left in the readahead window. Normally async_size | |
381 | * will be equal to size, for maximum pipelining. | |
122a21d1 FW |
382 | * |
383 | * In interleaved sequential reads, concurrent streams on the same fd can | |
384 | * be invalidating each other's readahead state. So we flag the new readahead | |
f9acc8c7 | 385 | * page at (start+size-async_size) with PG_readahead, and use it as readahead |
122a21d1 FW |
386 | * indicator. The flag won't be set on already cached pages, to avoid the |
387 | * readahead-for-nothing fuss, saving pointless page cache lookups. | |
388 | * | |
f4e6b498 | 389 | * prev_pos tracks the last visited byte in the _previous_ read request. |
122a21d1 FW |
390 | * It should be maintained by the caller, and will be used for detecting |
391 | * small random reads. Note that the readahead algorithm checks loosely | |
392 | * for sequential patterns. Hence interleaved reads might be served as | |
393 | * sequential ones. | |
394 | * | |
395 | * There is a special-case: if the first page which the application tries to | |
396 | * read happens to be the first page of the file, it is assumed that a linear | |
397 | * read is about to happen and the window is immediately set to the initial size | |
398 | * based on I/O request size and the max_readahead. | |
399 | * | |
400 | * The code ramps up the readahead size aggressively at first, but slow down as | |
401 | * it approaches max_readhead. | |
402 | */ | |
403 | ||
10be0b37 | 404 | /* |
08eb9658 | 405 | * Count contiguously cached pages from @index-1 to @index-@max, |
10be0b37 WF |
406 | * this count is a conservative estimation of |
407 | * - length of the sequential read sequence, or | |
408 | * - thrashing threshold in memory tight systems | |
409 | */ | |
410 | static pgoff_t count_history_pages(struct address_space *mapping, | |
08eb9658 | 411 | pgoff_t index, unsigned long max) |
10be0b37 WF |
412 | { |
413 | pgoff_t head; | |
414 | ||
415 | rcu_read_lock(); | |
08eb9658 | 416 | head = page_cache_prev_miss(mapping, index - 1, max); |
10be0b37 WF |
417 | rcu_read_unlock(); |
418 | ||
08eb9658 | 419 | return index - 1 - head; |
10be0b37 WF |
420 | } |
421 | ||
422 | /* | |
1e470280 | 423 | * page cache context based readahead |
10be0b37 WF |
424 | */ |
425 | static int try_context_readahead(struct address_space *mapping, | |
426 | struct file_ra_state *ra, | |
08eb9658 | 427 | pgoff_t index, |
10be0b37 WF |
428 | unsigned long req_size, |
429 | unsigned long max) | |
430 | { | |
431 | pgoff_t size; | |
432 | ||
08eb9658 | 433 | size = count_history_pages(mapping, index, max); |
10be0b37 WF |
434 | |
435 | /* | |
2cad4018 | 436 | * not enough history pages: |
10be0b37 WF |
437 | * it could be a random read |
438 | */ | |
2cad4018 | 439 | if (size <= req_size) |
10be0b37 WF |
440 | return 0; |
441 | ||
442 | /* | |
443 | * starts from beginning of file: | |
444 | * it is a strong indication of long-run stream (or whole-file-read) | |
445 | */ | |
08eb9658 | 446 | if (size >= index) |
10be0b37 WF |
447 | size *= 2; |
448 | ||
08eb9658 | 449 | ra->start = index; |
2cad4018 FW |
450 | ra->size = min(size + req_size, max); |
451 | ra->async_size = 1; | |
10be0b37 WF |
452 | |
453 | return 1; | |
454 | } | |
455 | ||
793917d9 MWO |
456 | /* |
457 | * There are some parts of the kernel which assume that PMD entries | |
458 | * are exactly HPAGE_PMD_ORDER. Those should be fixed, but until then, | |
459 | * limit the maximum allocation order to PMD size. I'm not aware of any | |
460 | * assumptions about maximum order if THP are disabled, but 8 seems like | |
461 | * a good order (that's 1MB if you're using 4kB pages) | |
462 | */ | |
463 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
464 | #define MAX_PAGECACHE_ORDER HPAGE_PMD_ORDER | |
465 | #else | |
466 | #define MAX_PAGECACHE_ORDER 8 | |
467 | #endif | |
468 | ||
469 | static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index, | |
470 | pgoff_t mark, unsigned int order, gfp_t gfp) | |
471 | { | |
472 | int err; | |
473 | struct folio *folio = filemap_alloc_folio(gfp, order); | |
474 | ||
475 | if (!folio) | |
476 | return -ENOMEM; | |
b9ff43dd MWO |
477 | mark = round_up(mark, 1UL << order); |
478 | if (index == mark) | |
793917d9 MWO |
479 | folio_set_readahead(folio); |
480 | err = filemap_add_folio(ractl->mapping, folio, index, gfp); | |
481 | if (err) | |
482 | folio_put(folio); | |
483 | else | |
484 | ractl->_nr_pages += 1UL << order; | |
485 | return err; | |
486 | } | |
487 | ||
56a4d67c | 488 | void page_cache_ra_order(struct readahead_control *ractl, |
793917d9 MWO |
489 | struct file_ra_state *ra, unsigned int new_order) |
490 | { | |
491 | struct address_space *mapping = ractl->mapping; | |
492 | pgoff_t index = readahead_index(ractl); | |
493 | pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT; | |
494 | pgoff_t mark = index + ra->size - ra->async_size; | |
495 | int err = 0; | |
496 | gfp_t gfp = readahead_gfp_mask(mapping); | |
497 | ||
498 | if (!mapping_large_folio_support(mapping) || ra->size < 4) | |
499 | goto fallback; | |
500 | ||
501 | limit = min(limit, index + ra->size - 1); | |
502 | ||
503 | if (new_order < MAX_PAGECACHE_ORDER) { | |
504 | new_order += 2; | |
505 | if (new_order > MAX_PAGECACHE_ORDER) | |
506 | new_order = MAX_PAGECACHE_ORDER; | |
507 | while ((1 << new_order) > ra->size) | |
508 | new_order--; | |
509 | } | |
510 | ||
511 | while (index <= limit) { | |
512 | unsigned int order = new_order; | |
513 | ||
514 | /* Align with smaller pages if needed */ | |
515 | if (index & ((1UL << order) - 1)) { | |
516 | order = __ffs(index); | |
517 | if (order == 1) | |
518 | order = 0; | |
519 | } | |
520 | /* Don't allocate pages past EOF */ | |
521 | while (index + (1UL << order) - 1 > limit) { | |
522 | if (--order == 1) | |
523 | order = 0; | |
524 | } | |
525 | err = ra_alloc_folio(ractl, index, mark, order, gfp); | |
526 | if (err) | |
527 | break; | |
528 | index += 1UL << order; | |
529 | } | |
530 | ||
531 | if (index > limit) { | |
532 | ra->size += index - limit - 1; | |
533 | ra->async_size += index - limit - 1; | |
534 | } | |
535 | ||
b4e089d7 | 536 | read_pages(ractl); |
793917d9 MWO |
537 | |
538 | /* | |
539 | * If there were already pages in the page cache, then we may have | |
540 | * left some gaps. Let the regular readahead code take care of this | |
541 | * situation. | |
542 | */ | |
543 | if (!err) | |
544 | return; | |
545 | fallback: | |
546 | do_page_cache_ra(ractl, ra->size, ra->async_size); | |
547 | } | |
548 | ||
122a21d1 FW |
549 | /* |
550 | * A minimal readahead algorithm for trivial sequential/random reads. | |
551 | */ | |
6e4af69a | 552 | static void ondemand_readahead(struct readahead_control *ractl, |
793917d9 | 553 | struct folio *folio, unsigned long req_size) |
122a21d1 | 554 | { |
6e4af69a | 555 | struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host); |
fcd9ae4f | 556 | struct file_ra_state *ra = ractl->ra; |
9491ae4a | 557 | unsigned long max_pages = ra->ra_pages; |
dc30b96a | 558 | unsigned long add_pages; |
b9ff43dd MWO |
559 | pgoff_t index = readahead_index(ractl); |
560 | pgoff_t expected, prev_index; | |
561 | unsigned int order = folio ? folio_order(folio) : 0; | |
045a2529 | 562 | |
9491ae4a JA |
563 | /* |
564 | * If the request exceeds the readahead window, allow the read to | |
565 | * be up to the optimal hardware IO size | |
566 | */ | |
567 | if (req_size > max_pages && bdi->io_pages > max_pages) | |
568 | max_pages = min(req_size, bdi->io_pages); | |
569 | ||
045a2529 WF |
570 | /* |
571 | * start of file | |
572 | */ | |
08eb9658 | 573 | if (!index) |
045a2529 | 574 | goto initial_readahead; |
122a21d1 FW |
575 | |
576 | /* | |
08eb9658 | 577 | * It's the expected callback index, assume sequential access. |
122a21d1 FW |
578 | * Ramp up sizes, and push forward the readahead window. |
579 | */ | |
b9ff43dd MWO |
580 | expected = round_up(ra->start + ra->size - ra->async_size, |
581 | 1UL << order); | |
582 | if (index == expected || index == (ra->start + ra->size)) { | |
f9acc8c7 | 583 | ra->start += ra->size; |
9491ae4a | 584 | ra->size = get_next_ra_size(ra, max_pages); |
f9acc8c7 FW |
585 | ra->async_size = ra->size; |
586 | goto readit; | |
122a21d1 FW |
587 | } |
588 | ||
6b10c6c9 | 589 | /* |
793917d9 | 590 | * Hit a marked folio without valid readahead state. |
6b10c6c9 FW |
591 | * E.g. interleaved reads. |
592 | * Query the pagecache for async_size, which normally equals to | |
593 | * readahead size. Ramp it up and use it as the new readahead size. | |
594 | */ | |
793917d9 | 595 | if (folio) { |
6b10c6c9 FW |
596 | pgoff_t start; |
597 | ||
30002ed2 | 598 | rcu_read_lock(); |
6e4af69a DH |
599 | start = page_cache_next_miss(ractl->mapping, index + 1, |
600 | max_pages); | |
30002ed2 | 601 | rcu_read_unlock(); |
6b10c6c9 | 602 | |
08eb9658 | 603 | if (!start || start - index > max_pages) |
9a42823a | 604 | return; |
6b10c6c9 FW |
605 | |
606 | ra->start = start; | |
08eb9658 | 607 | ra->size = start - index; /* old async_size */ |
160334a0 | 608 | ra->size += req_size; |
9491ae4a | 609 | ra->size = get_next_ra_size(ra, max_pages); |
6b10c6c9 FW |
610 | ra->async_size = ra->size; |
611 | goto readit; | |
612 | } | |
613 | ||
122a21d1 | 614 | /* |
045a2529 | 615 | * oversize read |
122a21d1 | 616 | */ |
9491ae4a | 617 | if (req_size > max_pages) |
045a2529 WF |
618 | goto initial_readahead; |
619 | ||
620 | /* | |
621 | * sequential cache miss | |
08eb9658 MWO |
622 | * trivial case: (index - prev_index) == 1 |
623 | * unaligned reads: (index - prev_index) == 0 | |
045a2529 | 624 | */ |
08eb9658 MWO |
625 | prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT; |
626 | if (index - prev_index <= 1UL) | |
045a2529 WF |
627 | goto initial_readahead; |
628 | ||
10be0b37 WF |
629 | /* |
630 | * Query the page cache and look for the traces(cached history pages) | |
631 | * that a sequential stream would leave behind. | |
632 | */ | |
6e4af69a DH |
633 | if (try_context_readahead(ractl->mapping, ra, index, req_size, |
634 | max_pages)) | |
10be0b37 WF |
635 | goto readit; |
636 | ||
045a2529 WF |
637 | /* |
638 | * standalone, small random read | |
639 | * Read as is, and do not pollute the readahead state. | |
640 | */ | |
6e4af69a | 641 | do_page_cache_ra(ractl, req_size, 0); |
9a42823a | 642 | return; |
045a2529 WF |
643 | |
644 | initial_readahead: | |
08eb9658 | 645 | ra->start = index; |
9491ae4a | 646 | ra->size = get_init_ra_size(req_size, max_pages); |
f9acc8c7 | 647 | ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; |
122a21d1 | 648 | |
f9acc8c7 | 649 | readit: |
51daa88e WF |
650 | /* |
651 | * Will this read hit the readahead marker made by itself? | |
652 | * If so, trigger the readahead marker hit now, and merge | |
653 | * the resulted next readahead window into the current one. | |
dc30b96a | 654 | * Take care of maximum IO pages as above. |
51daa88e | 655 | */ |
08eb9658 | 656 | if (index == ra->start && ra->size == ra->async_size) { |
dc30b96a MS |
657 | add_pages = get_next_ra_size(ra, max_pages); |
658 | if (ra->size + add_pages <= max_pages) { | |
659 | ra->async_size = add_pages; | |
660 | ra->size += add_pages; | |
661 | } else { | |
662 | ra->size = max_pages; | |
663 | ra->async_size = max_pages >> 1; | |
664 | } | |
51daa88e WF |
665 | } |
666 | ||
6e4af69a | 667 | ractl->_index = ra->start; |
b9ff43dd | 668 | page_cache_ra_order(ractl, ra, order); |
122a21d1 FW |
669 | } |
670 | ||
fefa7c47 | 671 | void page_cache_sync_ra(struct readahead_control *ractl, |
fcd9ae4f | 672 | unsigned long req_count) |
122a21d1 | 673 | { |
324bcf54 | 674 | bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM); |
cf914a7d | 675 | |
324bcf54 | 676 | /* |
1e470280 | 677 | * Even if readahead is disabled, issue this request as readahead |
324bcf54 | 678 | * as we'll need it to satisfy the requested range. The forced |
1e470280 | 679 | * readahead will do the right thing and limit the read to just the |
324bcf54 JA |
680 | * requested range, which we'll set to 1 page for this case. |
681 | */ | |
fcd9ae4f | 682 | if (!ractl->ra->ra_pages || blk_cgroup_congested()) { |
324bcf54 JA |
683 | if (!ractl->file) |
684 | return; | |
685 | req_count = 1; | |
686 | do_forced_ra = true; | |
687 | } | |
ca47e8c7 | 688 | |
0141450f | 689 | /* be dumb */ |
324bcf54 | 690 | if (do_forced_ra) { |
fcd9ae4f | 691 | force_page_cache_ra(ractl, req_count); |
0141450f WF |
692 | return; |
693 | } | |
694 | ||
793917d9 | 695 | ondemand_readahead(ractl, NULL, req_count); |
cf914a7d | 696 | } |
fefa7c47 | 697 | EXPORT_SYMBOL_GPL(page_cache_sync_ra); |
cf914a7d | 698 | |
fefa7c47 | 699 | void page_cache_async_ra(struct readahead_control *ractl, |
7836d999 | 700 | struct folio *folio, unsigned long req_count) |
cf914a7d | 701 | { |
1e470280 | 702 | /* no readahead */ |
fcd9ae4f | 703 | if (!ractl->ra->ra_pages) |
cf914a7d RR |
704 | return; |
705 | ||
706 | /* | |
707 | * Same bit is used for PG_readahead and PG_reclaim. | |
708 | */ | |
7836d999 | 709 | if (folio_test_writeback(folio)) |
cf914a7d RR |
710 | return; |
711 | ||
7836d999 | 712 | folio_clear_readahead(folio); |
cf914a7d | 713 | |
ca47e8c7 JB |
714 | if (blk_cgroup_congested()) |
715 | return; | |
716 | ||
793917d9 | 717 | ondemand_readahead(ractl, folio, req_count); |
122a21d1 | 718 | } |
fefa7c47 | 719 | EXPORT_SYMBOL_GPL(page_cache_async_ra); |
782182e5 | 720 | |
c7b95d51 | 721 | ssize_t ksys_readahead(int fd, loff_t offset, size_t count) |
782182e5 CW |
722 | { |
723 | ssize_t ret; | |
2903ff01 | 724 | struct fd f; |
782182e5 CW |
725 | |
726 | ret = -EBADF; | |
2903ff01 | 727 | f = fdget(fd); |
3d8f7615 AG |
728 | if (!f.file || !(f.file->f_mode & FMODE_READ)) |
729 | goto out; | |
730 | ||
731 | /* | |
732 | * The readahead() syscall is intended to run only on files | |
733 | * that can execute readahead. If readahead is not possible | |
734 | * on this file, then we must return -EINVAL. | |
735 | */ | |
736 | ret = -EINVAL; | |
737 | if (!f.file->f_mapping || !f.file->f_mapping->a_ops || | |
738 | !S_ISREG(file_inode(f.file)->i_mode)) | |
739 | goto out; | |
740 | ||
741 | ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED); | |
742 | out: | |
743 | fdput(f); | |
782182e5 CW |
744 | return ret; |
745 | } | |
c7b95d51 DB |
746 | |
747 | SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count) | |
748 | { | |
749 | return ksys_readahead(fd, offset, count); | |
750 | } | |
3ca23644 | 751 | |
59c10c52 GR |
752 | #if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD) |
753 | COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count) | |
754 | { | |
755 | return ksys_readahead(fd, compat_arg_u64_glue(offset), count); | |
756 | } | |
757 | #endif | |
758 | ||
3ca23644 DH |
759 | /** |
760 | * readahead_expand - Expand a readahead request | |
761 | * @ractl: The request to be expanded | |
762 | * @new_start: The revised start | |
763 | * @new_len: The revised size of the request | |
764 | * | |
765 | * Attempt to expand a readahead request outwards from the current size to the | |
766 | * specified size by inserting locked pages before and after the current window | |
767 | * to increase the size to the new window. This may involve the insertion of | |
768 | * THPs, in which case the window may get expanded even beyond what was | |
769 | * requested. | |
770 | * | |
771 | * The algorithm will stop if it encounters a conflicting page already in the | |
772 | * pagecache and leave a smaller expansion than requested. | |
773 | * | |
774 | * The caller must check for this by examining the revised @ractl object for a | |
775 | * different expansion than was requested. | |
776 | */ | |
777 | void readahead_expand(struct readahead_control *ractl, | |
778 | loff_t new_start, size_t new_len) | |
779 | { | |
780 | struct address_space *mapping = ractl->mapping; | |
781 | struct file_ra_state *ra = ractl->ra; | |
782 | pgoff_t new_index, new_nr_pages; | |
783 | gfp_t gfp_mask = readahead_gfp_mask(mapping); | |
784 | ||
785 | new_index = new_start / PAGE_SIZE; | |
786 | ||
787 | /* Expand the leading edge downwards */ | |
788 | while (ractl->_index > new_index) { | |
789 | unsigned long index = ractl->_index - 1; | |
790 | struct page *page = xa_load(&mapping->i_pages, index); | |
791 | ||
792 | if (page && !xa_is_value(page)) | |
793 | return; /* Page apparently present */ | |
794 | ||
795 | page = __page_cache_alloc(gfp_mask); | |
796 | if (!page) | |
797 | return; | |
798 | if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) { | |
799 | put_page(page); | |
800 | return; | |
801 | } | |
802 | ||
803 | ractl->_nr_pages++; | |
804 | ractl->_index = page->index; | |
805 | } | |
806 | ||
807 | new_len += new_start - readahead_pos(ractl); | |
808 | new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE); | |
809 | ||
810 | /* Expand the trailing edge upwards */ | |
811 | while (ractl->_nr_pages < new_nr_pages) { | |
812 | unsigned long index = ractl->_index + ractl->_nr_pages; | |
813 | struct page *page = xa_load(&mapping->i_pages, index); | |
814 | ||
815 | if (page && !xa_is_value(page)) | |
816 | return; /* Page apparently present */ | |
817 | ||
818 | page = __page_cache_alloc(gfp_mask); | |
819 | if (!page) | |
820 | return; | |
821 | if (add_to_page_cache_lru(page, mapping, index, gfp_mask) < 0) { | |
822 | put_page(page); | |
823 | return; | |
824 | } | |
825 | ractl->_nr_pages++; | |
826 | if (ra) { | |
827 | ra->size++; | |
828 | ra->async_size++; | |
829 | } | |
830 | } | |
831 | } | |
832 | EXPORT_SYMBOL(readahead_expand); |