linux/mm/readahead.c
<<
>>
Prefs
   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 */
  25void
  26file_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}
  31EXPORT_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 */
  42static 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 */
  59static 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 */
  81int 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
 107EXPORT_SYMBOL(read_cache_pages);
 108
 109static int read_pages(struct address_space *mapping, struct file *filp,
 110                struct list_head *pages, unsigned nr_pages)
 111{
 112        struct blk_plug plug;
 113        unsigned page_idx;
 114        int ret;
 115
 116        blk_start_plug(&plug);
 117
 118        if (mapping->a_ops->readpages) {
 119                ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
 120                /* Clean up the remaining pages */
 121                put_pages_list(pages);
 122                goto out;
 123        }
 124
 125        for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 126                struct page *page = list_to_page(pages);
 127                list_del(&page->lru);
 128                if (!add_to_page_cache_lru(page, mapping,
 129                                        page->index, GFP_KERNEL)) {
 130                        mapping->a_ops->readpage(filp, page);
 131                }
 132                page_cache_release(page);
 133        }
 134        ret = 0;
 135
 136out:
 137        blk_finish_plug(&plug);
 138
 139        return ret;
 140}
 141
 142/*
 143 * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
 144 * the pages first, then submits them all for I/O. This avoids the very bad
 145 * behaviour which would occur if page allocations are causing VM writeback.
 146 * We really don't want to intermingle reads and writes like that.
 147 *
 148 * Returns the number of pages requested, or the maximum amount of I/O allowed.
 149 */
 150static int
 151__do_page_cache_readahead(struct address_space *mapping, struct file *filp,
 152                        pgoff_t offset, unsigned long nr_to_read,
 153                        unsigned long lookahead_size)
 154{
 155        struct inode *inode = mapping->host;
 156        struct page *page;
 157        unsigned long end_index;        /* The last page we want to read */
 158        LIST_HEAD(page_pool);
 159        int page_idx;
 160        int ret = 0;
 161        loff_t isize = i_size_read(inode);
 162
 163        if (isize == 0)
 164                goto out;
 165
 166        end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
 167
 168        /*
 169         * Preallocate as many pages as we will need.
 170         */
 171        for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
 172                pgoff_t page_offset = offset + page_idx;
 173
 174                if (page_offset > end_index)
 175                        break;
 176
 177                rcu_read_lock();
 178                page = radix_tree_lookup(&mapping->page_tree, page_offset);
 179                rcu_read_unlock();
 180                if (page)
 181                        continue;
 182
 183                page = page_cache_alloc_readahead(mapping);
 184                if (!page)
 185                        break;
 186                page->index = page_offset;
 187                list_add(&page->lru, &page_pool);
 188                if (page_idx == nr_to_read - lookahead_size)
 189                        SetPageReadahead(page);
 190                ret++;
 191        }
 192
 193        /*
 194         * Now start the IO.  We ignore I/O errors - if the page is not
 195         * uptodate then the caller will launch readpage again, and
 196         * will then handle the error.
 197         */
 198        if (ret)
 199                read_pages(mapping, filp, &page_pool, ret);
 200        BUG_ON(!list_empty(&page_pool));
 201out:
 202        return ret;
 203}
 204
 205/*
 206 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
 207 * memory at once.
 208 */
 209int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 210                pgoff_t offset, unsigned long nr_to_read)
 211{
 212        int ret = 0;
 213
 214        if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 215                return -EINVAL;
 216
 217        nr_to_read = max_sane_readahead(nr_to_read);
 218        while (nr_to_read) {
 219                int err;
 220
 221                unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
 222
 223                if (this_chunk > nr_to_read)
 224                        this_chunk = nr_to_read;
 225                err = __do_page_cache_readahead(mapping, filp,
 226                                                offset, this_chunk, 0);
 227                if (err < 0) {
 228                        ret = err;
 229                        break;
 230                }
 231                ret += err;
 232                offset += this_chunk;
 233                nr_to_read -= this_chunk;
 234        }
 235        return ret;
 236}
 237
 238/*
 239 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
 240 * sensible upper limit.
 241 */
 242unsigned long max_sane_readahead(unsigned long nr)
 243{
 244        return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
 245                + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
 246}
 247
 248/*
 249 * Submit IO for the read-ahead request in file_ra_state.
 250 */
 251unsigned long ra_submit(struct file_ra_state *ra,
 252                       struct address_space *mapping, struct file *filp)
 253{
 254        int actual;
 255
 256        actual = __do_page_cache_readahead(mapping, filp,
 257                                        ra->start, ra->size, ra->async_size);
 258
 259        return actual;
 260}
 261
 262/*
 263 * Set the initial window size, round to next power of 2 and square
 264 * for small size, x 4 for medium, and x 2 for large
 265 * for 128k (32 page) max ra
 266 * 1-8 page = 32k initial, > 8 page = 128k initial
 267 */
 268static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 269{
 270        unsigned long newsize = roundup_pow_of_two(size);
 271
 272        if (newsize <= max / 32)
 273                newsize = newsize * 4;
 274        else if (newsize <= max / 4)
 275                newsize = newsize * 2;
 276        else
 277                newsize = max;
 278
 279        return newsize;
 280}
 281
 282/*
 283 *  Get the previous window size, ramp it up, and
 284 *  return it as the new window size.
 285 */
 286static unsigned long get_next_ra_size(struct file_ra_state *ra,
 287                                                unsigned long max)
 288{
 289        unsigned long cur = ra->size;
 290        unsigned long newsize;
 291
 292        if (cur < max / 16)
 293                newsize = 4 * cur;
 294        else
 295                newsize = 2 * cur;
 296
 297        return min(newsize, max);
 298}
 299
 300/*
 301 * On-demand readahead design.
 302 *
 303 * The fields in struct file_ra_state represent the most-recently-executed
 304 * readahead attempt:
 305 *
 306 *                        |<----- async_size ---------|
 307 *     |------------------- size -------------------->|
 308 *     |==================#===========================|
 309 *     ^start             ^page marked with PG_readahead
 310 *
 311 * To overlap application thinking time and disk I/O time, we do
 312 * `readahead pipelining': Do not wait until the application consumed all
 313 * readahead pages and stalled on the missing page at readahead_index;
 314 * Instead, submit an asynchronous readahead I/O as soon as there are
 315 * only async_size pages left in the readahead window. Normally async_size
 316 * will be equal to size, for maximum pipelining.
 317 *
 318 * In interleaved sequential reads, concurrent streams on the same fd can
 319 * be invalidating each other's readahead state. So we flag the new readahead
 320 * page at (start+size-async_size) with PG_readahead, and use it as readahead
 321 * indicator. The flag won't be set on already cached pages, to avoid the
 322 * readahead-for-nothing fuss, saving pointless page cache lookups.
 323 *
 324 * prev_pos tracks the last visited byte in the _previous_ read request.
 325 * It should be maintained by the caller, and will be used for detecting
 326 * small random reads. Note that the readahead algorithm checks loosely
 327 * for sequential patterns. Hence interleaved reads might be served as
 328 * sequential ones.
 329 *
 330 * There is a special-case: if the first page which the application tries to
 331 * read happens to be the first page of the file, it is assumed that a linear
 332 * read is about to happen and the window is immediately set to the initial size
 333 * based on I/O request size and the max_readahead.
 334 *
 335 * The code ramps up the readahead size aggressively at first, but slow down as
 336 * it approaches max_readhead.
 337 */
 338
 339/*
 340 * Count contiguously cached pages from @offset-1 to @offset-@max,
 341 * this count is a conservative estimation of
 342 *      - length of the sequential read sequence, or
 343 *      - thrashing threshold in memory tight systems
 344 */
 345static pgoff_t count_history_pages(struct address_space *mapping,
 346                                   struct file_ra_state *ra,
 347                                   pgoff_t offset, unsigned long max)
 348{
 349        pgoff_t head;
 350
 351        rcu_read_lock();
 352        head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
 353        rcu_read_unlock();
 354
 355        return offset - 1 - head;
 356}
 357
 358/*
 359 * page cache context based read-ahead
 360 */
 361static int try_context_readahead(struct address_space *mapping,
 362                                 struct file_ra_state *ra,
 363                                 pgoff_t offset,
 364                                 unsigned long req_size,
 365                                 unsigned long max)
 366{
 367        pgoff_t size;
 368
 369        size = count_history_pages(mapping, ra, offset, max);
 370
 371        /*
 372         * no history pages:
 373         * it could be a random read
 374         */
 375        if (!size)
 376                return 0;
 377
 378        /*
 379         * starts from beginning of file:
 380         * it is a strong indication of long-run stream (or whole-file-read)
 381         */
 382        if (size >= offset)
 383                size *= 2;
 384
 385        ra->start = offset;
 386        ra->size = get_init_ra_size(size + req_size, max);
 387        ra->async_size = ra->size;
 388
 389        return 1;
 390}
 391
 392/*
 393 * A minimal readahead algorithm for trivial sequential/random reads.
 394 */
 395static unsigned long
 396ondemand_readahead(struct address_space *mapping,
 397                   struct file_ra_state *ra, struct file *filp,
 398                   bool hit_readahead_marker, pgoff_t offset,
 399                   unsigned long req_size)
 400{
 401        unsigned long max = max_sane_readahead(ra->ra_pages);
 402
 403        /*
 404         * start of file
 405         */
 406        if (!offset)
 407                goto initial_readahead;
 408
 409        /*
 410         * It's the expected callback offset, assume sequential access.
 411         * Ramp up sizes, and push forward the readahead window.
 412         */
 413        if ((offset == (ra->start + ra->size - ra->async_size) ||
 414             offset == (ra->start + ra->size))) {
 415                ra->start += ra->size;
 416                ra->size = get_next_ra_size(ra, max);
 417                ra->async_size = ra->size;
 418                goto readit;
 419        }
 420
 421        /*
 422         * Hit a marked page without valid readahead state.
 423         * E.g. interleaved reads.
 424         * Query the pagecache for async_size, which normally equals to
 425         * readahead size. Ramp it up and use it as the new readahead size.
 426         */
 427        if (hit_readahead_marker) {
 428                pgoff_t start;
 429
 430                rcu_read_lock();
 431                start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
 432                rcu_read_unlock();
 433
 434                if (!start || start - offset > max)
 435                        return 0;
 436
 437                ra->start = start;
 438                ra->size = start - offset;      /* old async_size */
 439                ra->size += req_size;
 440                ra->size = get_next_ra_size(ra, max);
 441                ra->async_size = ra->size;
 442                goto readit;
 443        }
 444
 445        /*
 446         * oversize read
 447         */
 448        if (req_size > max)
 449                goto initial_readahead;
 450
 451        /*
 452         * sequential cache miss
 453         */
 454        if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
 455                goto initial_readahead;
 456
 457        /*
 458         * Query the page cache and look for the traces(cached history pages)
 459         * that a sequential stream would leave behind.
 460         */
 461        if (try_context_readahead(mapping, ra, offset, req_size, max))
 462                goto readit;
 463
 464        /*
 465         * standalone, small random read
 466         * Read as is, and do not pollute the readahead state.
 467         */
 468        return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
 469
 470initial_readahead:
 471        ra->start = offset;
 472        ra->size = get_init_ra_size(req_size, max);
 473        ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 474
 475readit:
 476        /*
 477         * Will this read hit the readahead marker made by itself?
 478         * If so, trigger the readahead marker hit now, and merge
 479         * the resulted next readahead window into the current one.
 480         */
 481        if (offset == ra->start && ra->size == ra->async_size) {
 482                ra->async_size = get_next_ra_size(ra, max);
 483                ra->size += ra->async_size;
 484        }
 485
 486        return ra_submit(ra, mapping, filp);
 487}
 488
 489/**
 490 * page_cache_sync_readahead - generic file readahead
 491 * @mapping: address_space which holds the pagecache and I/O vectors
 492 * @ra: file_ra_state which holds the readahead state
 493 * @filp: passed on to ->readpage() and ->readpages()
 494 * @offset: start offset into @mapping, in pagecache page-sized units
 495 * @req_size: hint: total size of the read which the caller is performing in
 496 *            pagecache pages
 497 *
 498 * page_cache_sync_readahead() should be called when a cache miss happened:
 499 * it will submit the read.  The readahead logic may decide to piggyback more
 500 * pages onto the read request if access patterns suggest it will improve
 501 * performance.
 502 */
 503void page_cache_sync_readahead(struct address_space *mapping,
 504                               struct file_ra_state *ra, struct file *filp,
 505                               pgoff_t offset, unsigned long req_size)
 506{
 507        /* no read-ahead */
 508        if (!ra->ra_pages)
 509                return;
 510
 511        /* be dumb */
 512        if (filp && (filp->f_mode & FMODE_RANDOM)) {
 513                force_page_cache_readahead(mapping, filp, offset, req_size);
 514                return;
 515        }
 516
 517        /* do read-ahead */
 518        ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 519}
 520EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 521
 522/**
 523 * page_cache_async_readahead - file readahead for marked pages
 524 * @mapping: address_space which holds the pagecache and I/O vectors
 525 * @ra: file_ra_state which holds the readahead state
 526 * @filp: passed on to ->readpage() and ->readpages()
 527 * @page: the page at @offset which has the PG_readahead flag set
 528 * @offset: start offset into @mapping, in pagecache page-sized units
 529 * @req_size: hint: total size of the read which the caller is performing in
 530 *            pagecache pages
 531 *
 532 * page_cache_async_readahead() should be called when a page is used which
 533 * has the PG_readahead flag; this is a marker to suggest that the application
 534 * has used up enough of the readahead window that we should start pulling in
 535 * more pages.
 536 */
 537void
 538page_cache_async_readahead(struct address_space *mapping,
 539                           struct file_ra_state *ra, struct file *filp,
 540                           struct page *page, pgoff_t offset,
 541                           unsigned long req_size)
 542{
 543        /* no read-ahead */
 544        if (!ra->ra_pages)
 545                return;
 546
 547        /*
 548         * Same bit is used for PG_readahead and PG_reclaim.
 549         */
 550        if (PageWriteback(page))
 551                return;
 552
 553        ClearPageReadahead(page);
 554
 555        /*
 556         * Defer asynchronous read-ahead on IO congestion.
 557         */
 558        if (bdi_read_congested(mapping->backing_dev_info))
 559                return;
 560
 561        /* do read-ahead */
 562        ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 563}
 564EXPORT_SYMBOL_GPL(page_cache_async_readahead);
 565
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.