readahead.c revision caca7cb748571a5b39943a9b3e7081feef055e5e
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 */ 24void 25file_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} 30EXPORT_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 */ 41static 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 */ 58static 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 */ 80int 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 106EXPORT_SYMBOL(read_cache_pages); 107 108static 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; 131out: 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 */ 146static 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)); 197out: 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 */ 205int 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 nr_to_read = max_sane_readahead(nr_to_read); 214 while (nr_to_read) { 215 int err; 216 217 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE; 218 219 if (this_chunk > nr_to_read) 220 this_chunk = nr_to_read; 221 err = __do_page_cache_readahead(mapping, filp, 222 offset, this_chunk, 0); 223 if (err < 0) { 224 ret = err; 225 break; 226 } 227 ret += err; 228 offset += this_chunk; 229 nr_to_read -= this_chunk; 230 } 231 return ret; 232} 233 234/* 235 * This version skips the IO if the queue is read-congested, and will tell the 236 * block layer to abandon the readahead if request allocation would block. 237 * 238 * force_page_cache_readahead() will ignore queue congestion and will block on 239 * request queues. 240 */ 241int do_page_cache_readahead(struct address_space *mapping, struct file *filp, 242 pgoff_t offset, unsigned long nr_to_read) 243{ 244 if (bdi_read_congested(mapping->backing_dev_info)) 245 return -1; 246 247 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0); 248} 249 250/* 251 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a 252 * sensible upper limit. 253 */ 254unsigned long max_sane_readahead(unsigned long nr) 255{ 256 return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE) 257 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2); 258} 259 260/* 261 * Submit IO for the read-ahead request in file_ra_state. 262 */ 263static unsigned long ra_submit(struct file_ra_state *ra, 264 struct address_space *mapping, struct file *filp) 265{ 266 int actual; 267 268 actual = __do_page_cache_readahead(mapping, filp, 269 ra->start, ra->size, ra->async_size); 270 271 return actual; 272} 273 274/* 275 * Set the initial window size, round to next power of 2 and square 276 * for small size, x 4 for medium, and x 2 for large 277 * for 128k (32 page) max ra 278 * 1-8 page = 32k initial, > 8 page = 128k initial 279 */ 280static unsigned long get_init_ra_size(unsigned long size, unsigned long max) 281{ 282 unsigned long newsize = roundup_pow_of_two(size); 283 284 if (newsize <= max / 32) 285 newsize = newsize * 4; 286 else if (newsize <= max / 4) 287 newsize = newsize * 2; 288 else 289 newsize = max; 290 291 return newsize; 292} 293 294/* 295 * Get the previous window size, ramp it up, and 296 * return it as the new window size. 297 */ 298static unsigned long get_next_ra_size(struct file_ra_state *ra, 299 unsigned long max) 300{ 301 unsigned long cur = ra->size; 302 unsigned long newsize; 303 304 if (cur < max / 16) 305 newsize = 4 * cur; 306 else 307 newsize = 2 * cur; 308 309 return min(newsize, max); 310} 311 312/* 313 * On-demand readahead design. 314 * 315 * The fields in struct file_ra_state represent the most-recently-executed 316 * readahead attempt: 317 * 318 * |<----- async_size ---------| 319 * |------------------- size -------------------->| 320 * |==================#===========================| 321 * ^start ^page marked with PG_readahead 322 * 323 * To overlap application thinking time and disk I/O time, we do 324 * `readahead pipelining': Do not wait until the application consumed all 325 * readahead pages and stalled on the missing page at readahead_index; 326 * Instead, submit an asynchronous readahead I/O as soon as there are 327 * only async_size pages left in the readahead window. Normally async_size 328 * will be equal to size, for maximum pipelining. 329 * 330 * In interleaved sequential reads, concurrent streams on the same fd can 331 * be invalidating each other's readahead state. So we flag the new readahead 332 * page at (start+size-async_size) with PG_readahead, and use it as readahead 333 * indicator. The flag won't be set on already cached pages, to avoid the 334 * readahead-for-nothing fuss, saving pointless page cache lookups. 335 * 336 * prev_pos tracks the last visited byte in the _previous_ read request. 337 * It should be maintained by the caller, and will be used for detecting 338 * small random reads. Note that the readahead algorithm checks loosely 339 * for sequential patterns. Hence interleaved reads might be served as 340 * sequential ones. 341 * 342 * There is a special-case: if the first page which the application tries to 343 * read happens to be the first page of the file, it is assumed that a linear 344 * read is about to happen and the window is immediately set to the initial size 345 * based on I/O request size and the max_readahead. 346 * 347 * The code ramps up the readahead size aggressively at first, but slow down as 348 * it approaches max_readhead. 349 */ 350 351/* 352 * A minimal readahead algorithm for trivial sequential/random reads. 353 */ 354static unsigned long 355ondemand_readahead(struct address_space *mapping, 356 struct file_ra_state *ra, struct file *filp, 357 bool hit_readahead_marker, pgoff_t offset, 358 unsigned long req_size) 359{ 360 unsigned long max = max_sane_readahead(ra->ra_pages); 361 pgoff_t prev_offset; 362 int sequential; 363 364 /* 365 * It's the expected callback offset, assume sequential access. 366 * Ramp up sizes, and push forward the readahead window. 367 */ 368 if (offset && (offset == (ra->start + ra->size - ra->async_size) || 369 offset == (ra->start + ra->size))) { 370 ra->start += ra->size; 371 ra->size = get_next_ra_size(ra, max); 372 ra->async_size = ra->size; 373 goto readit; 374 } 375 376 prev_offset = ra->prev_pos >> PAGE_CACHE_SHIFT; 377 sequential = offset - prev_offset <= 1UL || req_size > max; 378 379 /* 380 * Standalone, small read. 381 * Read as is, and do not pollute the readahead state. 382 */ 383 if (!hit_readahead_marker && !sequential) { 384 return __do_page_cache_readahead(mapping, filp, 385 offset, req_size, 0); 386 } 387 388 /* 389 * Hit a marked page without valid readahead state. 390 * E.g. interleaved reads. 391 * Query the pagecache for async_size, which normally equals to 392 * readahead size. Ramp it up and use it as the new readahead size. 393 */ 394 if (hit_readahead_marker) { 395 pgoff_t start; 396 397 rcu_read_lock(); 398 start = radix_tree_next_hole(&mapping->page_tree, offset+1,max); 399 rcu_read_unlock(); 400 401 if (!start || start - offset > max) 402 return 0; 403 404 ra->start = start; 405 ra->size = start - offset; /* old async_size */ 406 ra->size = get_next_ra_size(ra, max); 407 ra->async_size = ra->size; 408 goto readit; 409 } 410 411 /* 412 * It may be one of 413 * - first read on start of file 414 * - sequential cache miss 415 * - oversize random read 416 * Start readahead for it. 417 */ 418 ra->start = offset; 419 ra->size = get_init_ra_size(req_size, max); 420 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; 421 422readit: 423 return ra_submit(ra, mapping, filp); 424} 425 426/** 427 * page_cache_sync_readahead - generic file readahead 428 * @mapping: address_space which holds the pagecache and I/O vectors 429 * @ra: file_ra_state which holds the readahead state 430 * @filp: passed on to ->readpage() and ->readpages() 431 * @offset: start offset into @mapping, in pagecache page-sized units 432 * @req_size: hint: total size of the read which the caller is performing in 433 * pagecache pages 434 * 435 * page_cache_sync_readahead() should be called when a cache miss happened: 436 * it will submit the read. The readahead logic may decide to piggyback more 437 * pages onto the read request if access patterns suggest it will improve 438 * performance. 439 */ 440void page_cache_sync_readahead(struct address_space *mapping, 441 struct file_ra_state *ra, struct file *filp, 442 pgoff_t offset, unsigned long req_size) 443{ 444 /* no read-ahead */ 445 if (!ra->ra_pages) 446 return; 447 448 /* do read-ahead */ 449 ondemand_readahead(mapping, ra, filp, false, offset, req_size); 450} 451EXPORT_SYMBOL_GPL(page_cache_sync_readahead); 452 453/** 454 * page_cache_async_readahead - file readahead for marked pages 455 * @mapping: address_space which holds the pagecache and I/O vectors 456 * @ra: file_ra_state which holds the readahead state 457 * @filp: passed on to ->readpage() and ->readpages() 458 * @page: the page at @offset which has the PG_readahead flag set 459 * @offset: start offset into @mapping, in pagecache page-sized units 460 * @req_size: hint: total size of the read which the caller is performing in 461 * pagecache pages 462 * 463 * page_cache_async_ondemand() should be called when a page is used which 464 * has the PG_readahead flag; this is a marker to suggest that the application 465 * has used up enough of the readahead window that we should start pulling in 466 * more pages. 467 */ 468void 469page_cache_async_readahead(struct address_space *mapping, 470 struct file_ra_state *ra, struct file *filp, 471 struct page *page, pgoff_t offset, 472 unsigned long req_size) 473{ 474 /* no read-ahead */ 475 if (!ra->ra_pages) 476 return; 477 478 /* 479 * Same bit is used for PG_readahead and PG_reclaim. 480 */ 481 if (PageWriteback(page)) 482 return; 483 484 ClearPageReadahead(page); 485 486 /* 487 * Defer asynchronous read-ahead on IO congestion. 488 */ 489 if (bdi_read_congested(mapping->backing_dev_info)) 490 return; 491 492 /* do read-ahead */ 493 ondemand_readahead(mapping, ra, filp, true, offset, req_size); 494} 495EXPORT_SYMBOL_GPL(page_cache_async_readahead); 496