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