mm/readahead.c

   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 static int
 144 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
 145                         pgoff_t offset, unsigned long nr_to_read,
 146                         unsigned long lookahead_size)
 147 {
 148         struct inode *inode = mapping->host;
 149         struct page *page;
 150         unsigned long end_index;        /* The last page we want to read */
 151         LIST_HEAD(page_pool);
 152         int page_idx;
 153         int ret = 0;
 154         loff_t isize = i_size_read(inode);
 155
 156         if (isize == 0)
 157                 goto out;
 158
 159         end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
 160
 161         /*
 162          * Preallocate as many pages as we will need.
 163          */
 164         for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
 165                 pgoff_t page_offset = offset + page_idx;
 166
 167                 if (page_offset > end_index)
 168                         break;
 169
 170                 rcu_read_lock();
 171                 page = radix_tree_lookup(&mapping->page_tree, page_offset);
 172                 rcu_read_unlock();
 173                 if (page)
 174                         continue;
 175
 176                 page = page_cache_alloc_cold(mapping);
 177                 if (!page)
 178                         break;
 179                 page->index = page_offset;
 180                 list_add(&page->lru, &page_pool);
 181                 if (page_idx == nr_to_read - lookahead_size)
 182                         SetPageReadahead(page);
 183                 ret++;
 184         }
 185
 186         /*
 187          * Now start the IO.  We ignore I/O errors - if the page is not
 188          * uptodate then the caller will launch readpage again, and
 189          * will then handle the error.
 190          */
 191         if (ret)
 192                 read_pages(mapping, filp, &page_pool, ret);
 193         BUG_ON(!list_empty(&page_pool));
 194 out:
 195         return ret;
 196 }
 197
 198 /*
 199  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
 200  * memory at once.
 201  */
 202 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 203                 pgoff_t offset, unsigned long nr_to_read)
 204 {
 205         int ret = 0;
 206
 207         if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 208                 return -EINVAL;
 209
 210         nr_to_read = max_sane_readahead(nr_to_read);
 211         while (nr_to_read) {
 212                 int err;
 213
 214                 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
 215
 216                 if (this_chunk > nr_to_read)
 217                         this_chunk = nr_to_read;
 218                 err = __do_page_cache_readahead(mapping, filp,
 219                                                 offset, this_chunk, 0);
 220                 if (err < 0) {
 221                         ret = err;
 222                         break;
 223                 }
 224                 ret += err;
 225                 offset += this_chunk;
 226                 nr_to_read -= this_chunk;
 227         }
 228         return ret;
 229 }
 230
 231 /*
 232  * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
 233  * sensible upper limit.
 234  */
 235 unsigned long max_sane_readahead(unsigned long nr)
 236 {
 237         return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
 238                 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
 239 }
 240
 241 /*
 242  * Submit IO for the read-ahead request in file_ra_state.
 243  */
 244 unsigned long ra_submit(struct file_ra_state *ra,
 245                        struct address_space *mapping, struct file *filp)
 246 {
 247         int actual;
 248
 249         actual = __do_page_cache_readahead(mapping, filp,
 250                                         ra->start, ra->size, ra->async_size);
 251
 252         return actual;
 253 }
 254
 255 /*
 256  * Set the initial window size, round to next power of 2 and square
 257  * for small size, x 4 for medium, and x 2 for large
 258  * for 128k (32 page) max ra
 259  * 1-8 page = 32k initial, > 8 page = 128k initial
 260  */
 261 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 262 {
 263         unsigned long newsize = roundup_pow_of_two(size);
 264
 265         if (newsize <= max / 32)
 266                 newsize = newsize * 4;
 267         else if (newsize <= max / 4)
 268                 newsize = newsize * 2;
 269         else
 270                 newsize = max;
 271
 272         return newsize;
 273 }
 274
 275 /*
 276  *  Get the previous window size, ramp it up, and
 277  *  return it as the new window size.
 278  */
 279 static unsigned long get_next_ra_size(struct file_ra_state *ra,
 280                                                 unsigned long max)
 281 {
 282         unsigned long cur = ra->size;
 283         unsigned long newsize;
 284
 285         if (cur < max / 16)
 286                 newsize = 4 * cur;
 287         else
 288                 newsize = 2 * cur;
 289
 290         return min(newsize, max);
 291 }
 292
 293 /*
 294  * On-demand readahead design.
 295  *
 296  * The fields in struct file_ra_state represent the most-recently-executed
 297  * readahead attempt:
 298  *
 299  *                        |<----- async_size ---------|
 300  *     |------------------- size -------------------->|
 301  *     |==================#===========================|
 302  *     ^start             ^page marked with PG_readahead
 303  *
 304  * To overlap application thinking time and disk I/O time, we do
 305  * `readahead pipelining': Do not wait until the application consumed all
 306  * readahead pages and stalled on the missing page at readahead_index;
 307  * Instead, submit an asynchronous readahead I/O as soon as there are
 308  * only async_size pages left in the readahead window. Normally async_size
 309  * will be equal to size, for maximum pipelining.
 310  *
 311  * In interleaved sequential reads, concurrent streams on the same fd can
 312  * be invalidating each other's readahead state. So we flag the new readahead
 313  * page at (start+size-async_size) with PG_readahead, and use it as readahead
 314  * indicator. The flag won't be set on already cached pages, to avoid the
 315  * readahead-for-nothing fuss, saving pointless page cache lookups.
 316  *
 317  * prev_pos tracks the last visited byte in the _previous_ read request.
 318  * It should be maintained by the caller, and will be used for detecting
 319  * small random reads. Note that the readahead algorithm checks loosely
 320  * for sequential patterns. Hence interleaved reads might be served as
 321  * sequential ones.
 322  *
 323  * There is a special-case: if the first page which the application tries to
 324  * read happens to be the first page of the file, it is assumed that a linear
 325  * read is about to happen and the window is immediately set to the initial size
 326  * based on I/O request size and the max_readahead.
 327  *
 328  * The code ramps up the readahead size aggressively at first, but slow down as
 329  * it approaches max_readhead.
 330  */
 331
 332 /*
 333  * Count contiguously cached pages from @offset-1 to @offset-@max,
 334  * this count is a conservative estimation of
 335  *      - length of the sequential read sequence, or
 336  *      - thrashing threshold in memory tight systems
 337  */
 338 static pgoff_t count_history_pages(struct address_space *mapping,
 339                                    struct file_ra_state *ra,
 340                                    pgoff_t offset, unsigned long max)
 341 {
 342         pgoff_t head;
 343
 344         rcu_read_lock();
 345         head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
 346         rcu_read_unlock();
 347
 348         return offset - 1 - head;
 349 }
 350
 351 /*
 352  * page cache context based read-ahead
 353  */
 354 static int try_context_readahead(struct address_space *mapping,
 355                                  struct file_ra_state *ra,
 356                                  pgoff_t offset,
 357                                  unsigned long req_size,
 358                                  unsigned long max)
 359 {
 360         pgoff_t size;
 361
 362         size = count_history_pages(mapping, ra, offset, max);
 363
 364         /*
 365          * no history pages:
 366          * it could be a random read
 367          */
 368         if (!size)
 369                 return 0;
 370
 371         /*
 372          * starts from beginning of file:
 373          * it is a strong indication of long-run stream (or whole-file-read)
 374          */
 375         if (size >= offset)
 376                 size *= 2;
 377
 378         ra->start = offset;
 379         ra->size = get_init_ra_size(size + req_size, max);
 380         ra->async_size = ra->size;
 381
 382         return 1;
 383 }
 384
 385 /*
 386  * A minimal readahead algorithm for trivial sequential/random reads.
 387  */
 388 static unsigned long
 389 ondemand_readahead(struct address_space *mapping,
 390                    struct file_ra_state *ra, struct file *filp,
 391                    bool hit_readahead_marker, pgoff_t offset,
 392                    unsigned long req_size)
 393 {
 394         unsigned long max = max_sane_readahead(ra->ra_pages);
 395
 396         /*
 397          * start of file
 398          */
 399         if (!offset)
 400                 goto initial_readahead;
 401
 402         /*
 403          * It's the expected callback offset, assume sequential access.
 404          * Ramp up sizes, and push forward the readahead window.
 405          */
 406         if ((offset == (ra->start + ra->size - ra->async_size) ||
 407              offset == (ra->start + ra->size))) {
 408                 ra->start += ra->size;
 409                 ra->size = get_next_ra_size(ra, max);
 410                 ra->async_size = ra->size;
 411                 goto readit;
 412         }
 413
 414         /*
 415          * Hit a marked page without valid readahead state.
 416          * E.g. interleaved reads.
 417          * Query the pagecache for async_size, which normally equals to
 418          * readahead size. Ramp it up and use it as the new readahead size.
 419          */
 420         if (hit_readahead_marker) {
 421                 pgoff_t start;
 422
 423                 rcu_read_lock();
 424                 start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
 425                 rcu_read_unlock();
 426
 427                 if (!start || start - offset > max)
 428                         return 0;
 429
 430                 ra->start = start;
 431                 ra->size = start - offset;      /* old async_size */
 432                 ra->size += req_size;
 433                 ra->size = get_next_ra_size(ra, max);
 434                 ra->async_size = ra->size;
 435                 goto readit;
 436         }
 437
 438         /*
 439          * oversize read
 440          */
 441         if (req_size > max)
 442                 goto initial_readahead;
 443
 444         /*
 445          * sequential cache miss
 446          */
 447         if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
 448                 goto initial_readahead;
 449
 450         /*
 451          * Query the page cache and look for the traces(cached history pages)
 452          * that a sequential stream would leave behind.
 453          */
 454         if (try_context_readahead(mapping, ra, offset, req_size, max))
 455                 goto readit;
 456
 457         /*
 458          * standalone, small random read
 459          * Read as is, and do not pollute the readahead state.
 460          */
 461         return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
 462
 463 initial_readahead:
 464         ra->start = offset;
 465         ra->size = get_init_ra_size(req_size, max);
 466         ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 467
 468 readit:
 469         /*
 470          * Will this read hit the readahead marker made by itself?
 471          * If so, trigger the readahead marker hit now, and merge
 472          * the resulted next readahead window into the current one.
 473          */
 474         if (offset == ra->start && ra->size == ra->async_size) {
 475                 ra->async_size = get_next_ra_size(ra, max);
 476                 ra->size += ra->async_size;
 477         }
 478
 479         return ra_submit(ra, mapping, filp);
 480 }
 481
 482 /**
 483  * page_cache_sync_readahead - generic file readahead
 484  * @mapping: address_space which holds the pagecache and I/O vectors
 485  * @ra: file_ra_state which holds the readahead state
 486  * @filp: passed on to ->readpage() and ->readpages()
 487  * @offset: start offset into @mapping, in pagecache page-sized units
 488  * @req_size: hint: total size of the read which the caller is performing in
 489  *            pagecache pages
 490  *
 491  * page_cache_sync_readahead() should be called when a cache miss happened:
 492  * it will submit the read.  The readahead logic may decide to piggyback more
 493  * pages onto the read request if access patterns suggest it will improve
 494  * performance.
 495  */
 496 void page_cache_sync_readahead(struct address_space *mapping,
 497                                struct file_ra_state *ra, struct file *filp,
 498                                pgoff_t offset, unsigned long req_size)
 499 {
 500         /* no read-ahead */
 501         if (!ra->ra_pages)
 502                 return;
 503
 504         /* be dumb */
 505         if (filp && (filp->f_mode & FMODE_RANDOM)) {
 506                 force_page_cache_readahead(mapping, filp, offset, req_size);
 507                 return;
 508         }
 509
 510         /* do read-ahead */
 511         ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 512 }
 513 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 514
 515 /**
 516  * page_cache_async_readahead - file readahead for marked pages
 517  * @mapping: address_space which holds the pagecache and I/O vectors
 518  * @ra: file_ra_state which holds the readahead state
 519  * @filp: passed on to ->readpage() and ->readpages()
 520  * @page: the page at @offset which has the PG_readahead flag set
 521  * @offset: start offset into @mapping, in pagecache page-sized units
 522  * @req_size: hint: total size of the read which the caller is performing in
 523  *            pagecache pages
 524  *
 525  * page_cache_async_ondemand() should be called when a page is used which
 526  * has the PG_readahead flag; this is a marker to suggest that the application
 527  * has used up enough of the readahead window that we should start pulling in
 528  * more pages.
 529  */
 530 void
 531 page_cache_async_readahead(struct address_space *mapping,
 532                            struct file_ra_state *ra, struct file *filp,
 533                            struct page *page, pgoff_t offset,
 534                            unsigned long req_size)
 535 {
 536         /* no read-ahead */
 537         if (!ra->ra_pages)
 538                 return;
 539
 540         /*
 541          * Same bit is used for PG_readahead and PG_reclaim.
 542          */
 543         if (PageWriteback(page))
 544                 return;
 545
 546         ClearPageReadahead(page);
 547
 548         /*
 549          * Defer asynchronous read-ahead on IO congestion.
 550          */
 551         if (bdi_read_congested(mapping->backing_dev_info))
 552                 return;
 553
 554         /* do read-ahead */
 555         ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 556
 557 #ifdef CONFIG_BLOCK
 558         /*
 559          * Normally the current page is !uptodate and lock_page() will be
 560          * immediately called to implicitly unplug the device. However this
 561          * is not always true for RAID conifgurations, where data arrives
 562          * not strictly in their submission order. In this case we need to
 563          * explicitly kick off the IO.
 564          */
 565         if (PageUptodate(page))
 566                 blk_run_backing_dev(mapping->backing_dev_info, NULL);
 567 #endif
 568 }
 569 EXPORT_SYMBOL_GPL(page_cache_async_readahead);