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