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/gfp.h>
  13 #include <linux/mm.h>
  14 #include <linux/module.h>
  15 #include <linux/blkdev.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/task_io_accounting_ops.h>
  18 #include <linux/pagevec.h>
  19 #include <linux/pagemap.h>
  20
  21 /*
  22  * Initialise a struct file's readahead state.  Assumes that the caller has
  23  * memset *ra to zero.
  24  */
  25 void
  26 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
  27 {
  28         ra->ra_pages = mapping->backing_dev_info->ra_pages;
  29         ra->prev_pos = -1;
  30 }
  31 EXPORT_SYMBOL_GPL(file_ra_state_init);
  32
  33 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
  34
  35 /*
  36  * see if a page needs releasing upon read_cache_pages() failure
  37  * - the caller of read_cache_pages() may have set PG_private or PG_fscache
  38  *   before calling, such as the NFS fs marking pages that are cached locally
  39  *   on disk, thus we need to give the fs a chance to clean up in the event of
  40  *   an error
  41  */
  42 static void read_cache_pages_invalidate_page(struct address_space *mapping,
  43                                              struct page *page)
  44 {
  45         if (page_has_private(page)) {
  46                 if (!trylock_page(page))
  47                         BUG();
  48                 page->mapping = mapping;
  49                 do_invalidatepage(page, 0);
  50                 page->mapping = NULL;
  51                 unlock_page(page);
  52         }
  53         page_cache_release(page);
  54 }
  55
  56 /*
  57  * release a list of pages, invalidating them first if need be
  58  */
  59 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
  60                                               struct list_head *pages)
  61 {
  62         struct page *victim;
  63
  64         while (!list_empty(pages)) {
  65                 victim = list_to_page(pages);
  66                 list_del(&victim->lru);
  67                 read_cache_pages_invalidate_page(mapping, victim);
  68         }
  69 }
  70
  71 /**
  72  * read_cache_pages - populate an address space with some pages & start reads against them
  73  * @mapping: the address_space
  74  * @pages: The address of a list_head which contains the target pages.  These
  75  *   pages have their ->index populated and are otherwise uninitialised.
  76  * @filler: callback routine for filling a single page.
  77  * @data: private data for the callback routine.
  78  *
  79  * Hides the details of the LRU cache etc from the filesystems.
  80  */
  81 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
  82                         int (*filler)(void *, struct page *), void *data)
  83 {
  84         struct page *page;
  85         int ret = 0;
  86
  87         while (!list_empty(pages)) {
  88                 page = list_to_page(pages);
  89                 list_del(&page->lru);
  90                 if (add_to_page_cache_lru(page, mapping,
  91                                         page->index, GFP_KERNEL)) {
  92                         read_cache_pages_invalidate_page(mapping, page);
  93                         continue;
  94                 }
  95                 page_cache_release(page);
  96
  97                 ret = filler(data, page);
  98                 if (unlikely(ret)) {
  99                         read_cache_pages_invalidate_pages(mapping, pages);
 100                         break;
 101                 }
 102                 task_io_account_read(PAGE_CACHE_SIZE);
 103         }
 104         return ret;
 105 }
 106
 107 EXPORT_SYMBOL(read_cache_pages);
 108
 109 static int read_pages(struct address_space *mapping, struct file *filp,
 110                 struct list_head *pages, unsigned nr_pages)
 111 {
 112         unsigned page_idx;
 113         int ret;
 114
 115         if (mapping->a_ops->readpages) {
 116                 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
 117                 /* Clean up the remaining pages */
 118                 put_pages_list(pages);
 119                 goto out;
 120         }
 121
 122         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 123                 struct page *page = list_to_page(pages);
 124                 list_del(&page->lru);
 125                 if (!add_to_page_cache_lru(page, mapping,
 126                                         page->index, GFP_KERNEL)) {
 127                         mapping->a_ops->readpage(filp, page);
 128                 }
 129                 page_cache_release(page);
 130         }
 131         ret = 0;
 132 out:
 133         return ret;
 134 }
 135
 136 /*
 137  * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
 138  * the pages first, then submits them all for I/O. This avoids the very bad
 139  * behaviour which would occur if page allocations are causing VM writeback.
 140  * We really don't want to intermingle reads and writes like that.
 141  *
 142  * Returns the number of pages requested, or the maximum amount of I/O allowed.
 143  */
 144 static int
 145 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
 146                         pgoff_t offset, unsigned long nr_to_read,
 147                         unsigned long lookahead_size)
 148 {
 149         struct inode *inode = mapping->host;
 150         struct page *page;
 151         unsigned long end_index;        /* The last page we want to read */
 152         LIST_HEAD(page_pool);
 153         int page_idx;
 154         int ret = 0;
 155         loff_t isize = i_size_read(inode);
 156
 157         if (isize == 0)
 158                 goto out;
 159
 160         end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
 161
 162         /*
 163          * Preallocate as many pages as we will need.
 164          */
 165         for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
 166                 pgoff_t page_offset = offset + page_idx;
 167
 168                 if (page_offset > end_index)
 169                         break;
 170
 171                 rcu_read_lock();
 172                 page = radix_tree_lookup(&mapping->page_tree, page_offset);
 173                 rcu_read_unlock();
 174                 if (page)
 175                         continue;
 176
 177                 page = page_cache_alloc_cold(mapping);
 178                 if (!page)
 179                         break;
 180                 page->index = page_offset;
 181                 list_add(&page->lru, &page_pool);
 182                 if (page_idx == nr_to_read - lookahead_size)
 183                         SetPageReadahead(page);
 184                 ret++;
 185         }
 186
 187         /*
 188          * Now start the IO.  We ignore I/O errors - if the page is not
 189          * uptodate then the caller will launch readpage again, and
 190          * will then handle the error.
 191          */
 192         if (ret)
 193                 read_pages(mapping, filp, &page_pool, ret);
 194         BUG_ON(!list_empty(&page_pool));
 195 out:
 196         return ret;
 197 }
 198
 199 /*
 200  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
 201  * memory at once.
 202  */
 203 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 204                 pgoff_t offset, unsigned long nr_to_read)
 205 {
 206         int ret = 0;
 207
 208         if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 209                 return -EINVAL;
 210
 211         nr_to_read = max_sane_readahead(nr_to_read);
 212         while (nr_to_read) {
 213                 int err;
 214
 215                 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
 216
 217                 if (this_chunk > nr_to_read)
 218                         this_chunk = nr_to_read;
 219                 err = __do_page_cache_readahead(mapping, filp,
 220                                                 offset, this_chunk, 0);
 221                 if (err < 0) {
 222                         ret = err;
 223                         break;
 224                 }
 225                 ret += err;
 226                 offset += this_chunk;
 227                 nr_to_read -= this_chunk;
 228         }
 229         return ret;
 230 }
 231
 232 /*
 233  * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
 234  * sensible upper limit.
 235  */
 236 unsigned long max_sane_readahead(unsigned long nr)
 237 {
 238         return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
 239                 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
 240 }
 241
 242 /*
 243  * Submit IO for the read-ahead request in file_ra_state.
 244  */
 245 unsigned long ra_submit(struct file_ra_state *ra,
 246                        struct address_space *mapping, struct file *filp)
 247 {
 248         int actual;
 249
 250         actual = __do_page_cache_readahead(mapping, filp,
 251                                         ra->start, ra->size, ra->async_size);
 252
 253         return actual;
 254 }
 255
 256 /*
 257  * Set the initial window size, round to next power of 2 and square
 258  * for small size, x 4 for medium, and x 2 for large
 259  * for 128k (32 page) max ra
 260  * 1-8 page = 32k initial, > 8 page = 128k initial
 261  */
 262 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 263 {
 264         unsigned long newsize = roundup_pow_of_two(size);
 265
 266         if (newsize <= max / 32)
 267                 newsize = newsize * 4;
 268         else if (newsize <= max / 4)
 269                 newsize = newsize * 2;
 270         else
 271                 newsize = max;
 272
 273         return newsize;
 274 }
 275
 276 /*
 277  *  Get the previous window size, ramp it up, and
 278  *  return it as the new window size.
 279  */
 280 static unsigned long get_next_ra_size(struct file_ra_state *ra,
 281                                                 unsigned long max)
 282 {
 283         unsigned long cur = ra->size;
 284         unsigned long newsize;
 285
 286         if (cur < max / 16)
 287                 newsize = 4 * cur;
 288         else
 289                 newsize = 2 * cur;
 290
 291         return min(newsize, max);
 292 }
 293
 294 /*
 295  * On-demand readahead design.
 296  *
 297  * The fields in struct file_ra_state represent the most-recently-executed
 298  * readahead attempt:
 299  *
 300  *                        |<----- async_size ---------|
 301  *     |------------------- size -------------------->|
 302  *     |==================#===========================|
 303  *     ^start             ^page marked with PG_readahead
 304  *
 305  * To overlap application thinking time and disk I/O time, we do
 306  * `readahead pipelining': Do not wait until the application consumed all
 307  * readahead pages and stalled on the missing page at readahead_index;
 308  * Instead, submit an asynchronous readahead I/O as soon as there are
 309  * only async_size pages left in the readahead window. Normally async_size
 310  * will be equal to size, for maximum pipelining.
 311  *
 312  * In interleaved sequential reads, concurrent streams on the same fd can
 313  * be invalidating each other's readahead state. So we flag the new readahead
 314  * page at (start+size-async_size) with PG_readahead, and use it as readahead
 315  * indicator. The flag won't be set on already cached pages, to avoid the
 316  * readahead-for-nothing fuss, saving pointless page cache lookups.
 317  *
 318  * prev_pos tracks the last visited byte in the _previous_ read request.
 319  * It should be maintained by the caller, and will be used for detecting
 320  * small random reads. Note that the readahead algorithm checks loosely
 321  * for sequential patterns. Hence interleaved reads might be served as
 322  * sequential ones.
 323  *
 324  * There is a special-case: if the first page which the application tries to
 325  * read happens to be the first page of the file, it is assumed that a linear
 326  * read is about to happen and the window is immediately set to the initial size
 327  * based on I/O request size and the max_readahead.
 328  *
 329  * The code ramps up the readahead size aggressively at first, but slow down as
 330  * it approaches max_readhead.
 331  */
 332
 333 /*
 334  * Count contiguously cached pages from @offset-1 to @offset-@max,
 335  * this count is a conservative estimation of
 336  *      - length of the sequential read sequence, or
 337  *      - thrashing threshold in memory tight systems
 338  */
 339 static pgoff_t count_history_pages(struct address_space *mapping,
 340                                    struct file_ra_state *ra,
 341                                    pgoff_t offset, unsigned long max)
 342 {
 343         pgoff_t head;
 344
 345         rcu_read_lock();
 346         head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
 347         rcu_read_unlock();
 348
 349         return offset - 1 - head;
 350 }
 351
 352 /*
 353  * page cache context based read-ahead
 354  */
 355 static int try_context_readahead(struct address_space *mapping,
 356                                  struct file_ra_state *ra,
 357                                  pgoff_t offset,
 358                                  unsigned long req_size,
 359                                  unsigned long max)
 360 {
 361         pgoff_t size;
 362
 363         size = count_history_pages(mapping, ra, offset, max);
 364
 365         /*
 366          * no history pages:
 367          * it could be a random read
 368          */
 369         if (!size)
 370                 return 0;
 371
 372         /*
 373          * starts from beginning of file:
 374          * it is a strong indication of long-run stream (or whole-file-read)
 375          */
 376         if (size >= offset)
 377                 size *= 2;
 378
 379         ra->start = offset;
 380         ra->size = get_init_ra_size(size + req_size, max);
 381         ra->async_size = ra->size;
 382
 383         return 1;
 384 }
 385
 386 /*
 387  * A minimal readahead algorithm for trivial sequential/random reads.
 388  */
 389 static unsigned long
 390 ondemand_readahead(struct address_space *mapping,
 391                    struct file_ra_state *ra, struct file *filp,
 392                    bool hit_readahead_marker, pgoff_t offset,
 393                    unsigned long req_size)
 394 {
 395         unsigned long max = max_sane_readahead(ra->ra_pages);
 396
 397         /*
 398          * start of file
 399          */
 400         if (!offset)
 401                 goto initial_readahead;
 402
 403         /*
 404          * It's the expected callback offset, assume sequential access.
 405          * Ramp up sizes, and push forward the readahead window.
 406          */
 407         if ((offset == (ra->start + ra->size - ra->async_size) ||
 408              offset == (ra->start + ra->size))) {
 409                 ra->start += ra->size;
 410                 ra->size = get_next_ra_size(ra, max);
 411                 ra->async_size = ra->size;
 412                 goto readit;
 413         }
 414
 415         /*
 416          * Hit a marked page without valid readahead state.
 417          * E.g. interleaved reads.
 418          * Query the pagecache for async_size, which normally equals to
 419          * readahead size. Ramp it up and use it as the new readahead size.
 420          */
 421         if (hit_readahead_marker) {
 422                 pgoff_t start;
 423
 424                 rcu_read_lock();
 425                 start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
 426                 rcu_read_unlock();
 427
 428                 if (!start || start - offset > max)
 429                         return 0;
 430
 431                 ra->start = start;
 432                 ra->size = start - offset;      /* old async_size */
 433                 ra->size += req_size;
 434                 ra->size = get_next_ra_size(ra, max);
 435                 ra->async_size = ra->size;
 436                 goto readit;
 437         }
 438
 439         /*
 440          * oversize read
 441          */
 442         if (req_size > max)
 443                 goto initial_readahead;
 444
 445         /*
 446          * sequential cache miss
 447          */
 448         if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
 449                 goto initial_readahead;
 450
 451         /*
 452          * Query the page cache and look for the traces(cached history pages)
 453          * that a sequential stream would leave behind.
 454          */
 455         if (try_context_readahead(mapping, ra, offset, req_size, max))
 456                 goto readit;
 457
 458         /*
 459          * standalone, small random read
 460          * Read as is, and do not pollute the readahead state.
 461          */
 462         return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
 463
 464 initial_readahead:
 465         ra->start = offset;
 466         ra->size = get_init_ra_size(req_size, max);
 467         ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 468
 469 readit:
 470         /*
 471          * Will this read hit the readahead marker made by itself?
 472          * If so, trigger the readahead marker hit now, and merge
 473          * the resulted next readahead window into the current one.
 474          */
 475         if (offset == ra->start && ra->size == ra->async_size) {
 476                 ra->async_size = get_next_ra_size(ra, max);
 477                 ra->size += ra->async_size;
 478         }
 479
 480         return ra_submit(ra, mapping, filp);
 481 }
 482
 483 /**
 484  * page_cache_sync_readahead - generic file readahead
 485  * @mapping: address_space which holds the pagecache and I/O vectors
 486  * @ra: file_ra_state which holds the readahead state
 487  * @filp: passed on to ->readpage() and ->readpages()
 488  * @offset: start offset into @mapping, in pagecache page-sized units
 489  * @req_size: hint: total size of the read which the caller is performing in
 490  *            pagecache pages
 491  *
 492  * page_cache_sync_readahead() should be called when a cache miss happened:
 493  * it will submit the read.  The readahead logic may decide to piggyback more
 494  * pages onto the read request if access patterns suggest it will improve
 495  * performance.
 496  */
 497 void page_cache_sync_readahead(struct address_space *mapping,
 498                                struct file_ra_state *ra, struct file *filp,
 499                                pgoff_t offset, unsigned long req_size)
 500 {
 501         /* no read-ahead */
 502         if (!ra->ra_pages)
 503                 return;
 504
 505         /* be dumb */
 506         if (filp && (filp->f_mode & FMODE_RANDOM)) {
 507                 force_page_cache_readahead(mapping, filp, offset, req_size);
 508                 return;
 509         }
 510
 511         /* do read-ahead */
 512         ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 513 }
 514 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 515
 516 /**
 517  * page_cache_async_readahead - file readahead for marked pages
 518  * @mapping: address_space which holds the pagecache and I/O vectors
 519  * @ra: file_ra_state which holds the readahead state
 520  * @filp: passed on to ->readpage() and ->readpages()
 521  * @page: the page at @offset which has the PG_readahead flag set
 522  * @offset: start offset into @mapping, in pagecache page-sized units
 523  * @req_size: hint: total size of the read which the caller is performing in
 524  *            pagecache pages
 525  *
 526  * page_cache_async_ondemand() should be called when a page is used which
 527  * has the PG_readahead flag; this is a marker to suggest that the application
 528  * has used up enough of the readahead window that we should start pulling in
 529  * more pages.
 530  */
 531 void
 532 page_cache_async_readahead(struct address_space *mapping,
 533                            struct file_ra_state *ra, struct file *filp,
 534                            struct page *page, pgoff_t offset,
 535                            unsigned long req_size)
 536 {
 537         /* no read-ahead */
 538         if (!ra->ra_pages)
 539                 return;
 540
 541         /*
 542          * Same bit is used for PG_readahead and PG_reclaim.
 543          */
 544         if (PageWriteback(page))
 545                 return;
 546
 547         ClearPageReadahead(page);
 548
 549         /*
 550          * Defer asynchronous read-ahead on IO congestion.
 551          */
 552         if (bdi_read_congested(mapping->backing_dev_info))
 553                 return;
 554
 555         /* do read-ahead */
 556         ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 557
 558 #ifdef CONFIG_BLOCK
 559         /*
 560          * Normally the current page is !uptodate and lock_page() will be
 561          * immediately called to implicitly unplug the device. However this
 562          * is not always true for RAID conifgurations, where data arrives
 563          * not strictly in their submission order. In this case we need to
 564          * explicitly kick off the IO.
 565          */
 566         if (PageUptodate(page))
 567                 blk_run_backing_dev(mapping->backing_dev_info, NULL);
 568 #endif
 569 }
 570 EXPORT_SYMBOL_GPL(page_cache_async_readahead);