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/dax.h>
  12 #include <linux/gfp.h>
  13 #include <linux/export.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 #include <linux/syscalls.h>
  20 #include <linux/file.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/blk-cgroup.h>
  23 #include <linux/fadvise.h>
  24
  25 #include "internal.h"
  26
  27 /*
  28  * Initialise a struct file's readahead state.  Assumes that the caller has
  29  * memset *ra to zero.
  30  */
  31 void
  32 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
  33 {
  34         ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
  35         ra->prev_pos = -1;
  36 }
  37 EXPORT_SYMBOL_GPL(file_ra_state_init);
  38
  39 /*
  40  * see if a page needs releasing upon read_cache_pages() failure
  41  * - the caller of read_cache_pages() may have set PG_private or PG_fscache
  42  *   before calling, such as the NFS fs marking pages that are cached locally
  43  *   on disk, thus we need to give the fs a chance to clean up in the event of
  44  *   an error
  45  */
  46 static void read_cache_pages_invalidate_page(struct address_space *mapping,
  47                                              struct page *page)
  48 {
  49         if (page_has_private(page)) {
  50                 if (!trylock_page(page))
  51                         BUG();
  52                 page->mapping = mapping;
  53                 do_invalidatepage(page, 0, PAGE_SIZE);
  54                 page->mapping = NULL;
  55                 unlock_page(page);
  56         }
  57         put_page(page);
  58 }
  59
  60 /*
  61  * release a list of pages, invalidating them first if need be
  62  */
  63 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
  64                                               struct list_head *pages)
  65 {
  66         struct page *victim;
  67
  68         while (!list_empty(pages)) {
  69                 victim = lru_to_page(pages);
  70                 list_del(&victim->lru);
  71                 read_cache_pages_invalidate_page(mapping, victim);
  72         }
  73 }
  74
  75 /**
  76  * read_cache_pages - populate an address space with some pages & start reads against them
  77  * @mapping: the address_space
  78  * @pages: The address of a list_head which contains the target pages.  These
  79  *   pages have their ->index populated and are otherwise uninitialised.
  80  * @filler: callback routine for filling a single page.
  81  * @data: private data for the callback routine.
  82  *
  83  * Hides the details of the LRU cache etc from the filesystems.
  84  */
  85 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
  86                         int (*filler)(void *, struct page *), void *data)
  87 {
  88         struct page *page;
  89         int ret = 0;
  90
  91         while (!list_empty(pages)) {
  92                 page = lru_to_page(pages);
  93                 list_del(&page->lru);
  94                 if (add_to_page_cache_lru(page, mapping, page->index,
  95                                 readahead_gfp_mask(mapping))) {
  96                         read_cache_pages_invalidate_page(mapping, page);
  97                         continue;
  98                 }
  99                 put_page(page);
 100
 101                 ret = filler(data, page);
 102                 if (unlikely(ret)) {
 103                         read_cache_pages_invalidate_pages(mapping, pages);
 104                         break;
 105                 }
 106                 task_io_account_read(PAGE_SIZE);
 107         }
 108         return ret;
 109 }
 110
 111 EXPORT_SYMBOL(read_cache_pages);
 112
 113 static int read_pages(struct address_space *mapping, struct file *filp,
 114                 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
 115 {
 116         struct blk_plug plug;
 117         unsigned page_idx;
 118         int ret;
 119
 120         blk_start_plug(&plug);
 121
 122         if (mapping->a_ops->readpages) {
 123                 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
 124                 /* Clean up the remaining pages */
 125                 put_pages_list(pages);
 126                 goto out;
 127         }
 128
 129         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 130                 struct page *page = lru_to_page(pages);
 131                 list_del(&page->lru);
 132                 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
 133                         mapping->a_ops->readpage(filp, page);
 134                 put_page(page);
 135         }
 136         ret = 0;
 137
 138 out:
 139         blk_finish_plug(&plug);
 140
 141         return ret;
 142 }
 143
 144 /*
 145  * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates
 146  * the pages first, then submits them for I/O. This avoids the very bad
 147  * behaviour which would occur if page allocations are causing VM writeback.
 148  * We really don't want to intermingle reads and writes like that.
 149  *
 150  * Returns the number of pages requested, or the maximum amount of I/O allowed.
 151  */
 152 unsigned int __do_page_cache_readahead(struct address_space *mapping,
 153                 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
 154                 unsigned long lookahead_size)
 155 {
 156         struct inode *inode = mapping->host;
 157         struct page *page;
 158         unsigned long end_index;        /* The last page we want to read */
 159         LIST_HEAD(page_pool);
 160         int page_idx;
 161         unsigned int nr_pages = 0;
 162         loff_t isize = i_size_read(inode);
 163         gfp_t gfp_mask = readahead_gfp_mask(mapping);
 164
 165         if (isize == 0)
 166                 goto out;
 167
 168         end_index = ((isize - 1) >> PAGE_SHIFT);
 169
 170         /*
 171          * Preallocate as many pages as we will need.
 172          */
 173         for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
 174                 pgoff_t page_offset = offset + page_idx;
 175
 176                 if (page_offset > end_index)
 177                         break;
 178
 179                 rcu_read_lock();
 180                 page = radix_tree_lookup(&mapping->i_pages, page_offset);
 181                 rcu_read_unlock();
 182                 if (page && !radix_tree_exceptional_entry(page)) {
 183                         /*
 184                          * Page already present?  Kick off the current batch of
 185                          * contiguous pages before continuing with the next
 186                          * batch.
 187                          */
 188                         if (nr_pages)
 189                                 read_pages(mapping, filp, &page_pool, nr_pages,
 190                                                 gfp_mask);
 191                         nr_pages = 0;
 192                         continue;
 193                 }
 194
 195                 page = __page_cache_alloc(gfp_mask);
 196                 if (!page)
 197                         break;
 198                 page->index = page_offset;
 199                 list_add(&page->lru, &page_pool);
 200                 if (page_idx == nr_to_read - lookahead_size)
 201                         SetPageReadahead(page);
 202                 nr_pages++;
 203         }
 204
 205         /*
 206          * Now start the IO.  We ignore I/O errors - if the page is not
 207          * uptodate then the caller will launch readpage again, and
 208          * will then handle the error.
 209          */
 210         if (nr_pages)
 211                 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
 212         BUG_ON(!list_empty(&page_pool));
 213 out:
 214         return nr_pages;
 215 }
 216
 217 /*
 218  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
 219  * memory at once.
 220  */
 221 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 222                                pgoff_t offset, unsigned long nr_to_read)
 223 {
 224         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 225         struct file_ra_state *ra = &filp->f_ra;
 226         unsigned long max_pages;
 227
 228         if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 229                 return -EINVAL;
 230
 231         /*
 232          * If the request exceeds the readahead window, allow the read to
 233          * be up to the optimal hardware IO size
 234          */
 235         max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
 236         nr_to_read = min(nr_to_read, max_pages);
 237         while (nr_to_read) {
 238                 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
 239
 240                 if (this_chunk > nr_to_read)
 241                         this_chunk = nr_to_read;
 242                 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
 243
 244                 offset += this_chunk;
 245                 nr_to_read -= this_chunk;
 246         }
 247         return 0;
 248 }
 249
 250 /*
 251  * Set the initial window size, round to next power of 2 and square
 252  * for small size, x 4 for medium, and x 2 for large
 253  * for 128k (32 page) max ra
 254  * 1-8 page = 32k initial, > 8 page = 128k initial
 255  */
 256 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 257 {
 258         unsigned long newsize = roundup_pow_of_two(size);
 259
 260         if (newsize <= max / 32)
 261                 newsize = newsize * 4;
 262         else if (newsize <= max / 4)
 263                 newsize = newsize * 2;
 264         else
 265                 newsize = max;
 266
 267         return newsize;
 268 }
 269
 270 /*
 271  *  Get the previous window size, ramp it up, and
 272  *  return it as the new window size.
 273  */
 274 static unsigned long get_next_ra_size(struct file_ra_state *ra,
 275                                                 unsigned long max)
 276 {
 277         unsigned long cur = ra->size;
 278         unsigned long newsize;
 279
 280         if (cur < max / 16)
 281                 newsize = 4 * cur;
 282         else
 283                 newsize = 2 * cur;
 284
 285         return min(newsize, max);
 286 }
 287
 288 /*
 289  * On-demand readahead design.
 290  *
 291  * The fields in struct file_ra_state represent the most-recently-executed
 292  * readahead attempt:
 293  *
 294  *                        |<----- async_size ---------|
 295  *     |------------------- size -------------------->|
 296  *     |==================#===========================|
 297  *     ^start             ^page marked with PG_readahead
 298  *
 299  * To overlap application thinking time and disk I/O time, we do
 300  * `readahead pipelining': Do not wait until the application consumed all
 301  * readahead pages and stalled on the missing page at readahead_index;
 302  * Instead, submit an asynchronous readahead I/O as soon as there are
 303  * only async_size pages left in the readahead window. Normally async_size
 304  * will be equal to size, for maximum pipelining.
 305  *
 306  * In interleaved sequential reads, concurrent streams on the same fd can
 307  * be invalidating each other's readahead state. So we flag the new readahead
 308  * page at (start+size-async_size) with PG_readahead, and use it as readahead
 309  * indicator. The flag won't be set on already cached pages, to avoid the
 310  * readahead-for-nothing fuss, saving pointless page cache lookups.
 311  *
 312  * prev_pos tracks the last visited byte in the _previous_ read request.
 313  * It should be maintained by the caller, and will be used for detecting
 314  * small random reads. Note that the readahead algorithm checks loosely
 315  * for sequential patterns. Hence interleaved reads might be served as
 316  * sequential ones.
 317  *
 318  * There is a special-case: if the first page which the application tries to
 319  * read happens to be the first page of the file, it is assumed that a linear
 320  * read is about to happen and the window is immediately set to the initial size
 321  * based on I/O request size and the max_readahead.
 322  *
 323  * The code ramps up the readahead size aggressively at first, but slow down as
 324  * it approaches max_readhead.
 325  */
 326
 327 /*
 328  * Count contiguously cached pages from @offset-1 to @offset-@max,
 329  * this count is a conservative estimation of
 330  *      - length of the sequential read sequence, or
 331  *      - thrashing threshold in memory tight systems
 332  */
 333 static pgoff_t count_history_pages(struct address_space *mapping,
 334                                    pgoff_t offset, unsigned long max)
 335 {
 336         pgoff_t head;
 337
 338         rcu_read_lock();
 339         head = page_cache_prev_hole(mapping, offset - 1, max);
 340         rcu_read_unlock();
 341
 342         return offset - 1 - head;
 343 }
 344
 345 /*
 346  * page cache context based read-ahead
 347  */
 348 static int try_context_readahead(struct address_space *mapping,
 349                                  struct file_ra_state *ra,
 350                                  pgoff_t offset,
 351                                  unsigned long req_size,
 352                                  unsigned long max)
 353 {
 354         pgoff_t size;
 355
 356         size = count_history_pages(mapping, offset, max);
 357
 358         /*
 359          * not enough history pages:
 360          * it could be a random read
 361          */
 362         if (size <= req_size)
 363                 return 0;
 364
 365         /*
 366          * starts from beginning of file:
 367          * it is a strong indication of long-run stream (or whole-file-read)
 368          */
 369         if (size >= offset)
 370                 size *= 2;
 371
 372         ra->start = offset;
 373         ra->size = min(size + req_size, max);
 374         ra->async_size = 1;
 375
 376         return 1;
 377 }
 378
 379 /*
 380  * A minimal readahead algorithm for trivial sequential/random reads.
 381  */
 382 static unsigned long
 383 ondemand_readahead(struct address_space *mapping,
 384                    struct file_ra_state *ra, struct file *filp,
 385                    bool hit_readahead_marker, pgoff_t offset,
 386                    unsigned long req_size)
 387 {
 388         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 389         unsigned long max_pages = ra->ra_pages;
 390         unsigned long add_pages;
 391         pgoff_t prev_offset;
 392
 393         /*
 394          * If the request exceeds the readahead window, allow the read to
 395          * be up to the optimal hardware IO size
 396          */
 397         if (req_size > max_pages && bdi->io_pages > max_pages)
 398                 max_pages = min(req_size, bdi->io_pages);
 399
 400         /*
 401          * start of file
 402          */
 403         if (!offset)
 404                 goto initial_readahead;
 405
 406         /*
 407          * It's the expected callback offset, assume sequential access.
 408          * Ramp up sizes, and push forward the readahead window.
 409          */
 410         if ((offset == (ra->start + ra->size - ra->async_size) ||
 411              offset == (ra->start + ra->size))) {
 412                 ra->start += ra->size;
 413                 ra->size = get_next_ra_size(ra, max_pages);
 414                 ra->async_size = ra->size;
 415                 goto readit;
 416         }
 417
 418         /*
 419          * Hit a marked page without valid readahead state.
 420          * E.g. interleaved reads.
 421          * Query the pagecache for async_size, which normally equals to
 422          * readahead size. Ramp it up and use it as the new readahead size.
 423          */
 424         if (hit_readahead_marker) {
 425                 pgoff_t start;
 426
 427                 rcu_read_lock();
 428                 start = page_cache_next_hole(mapping, offset + 1, max_pages);
 429                 rcu_read_unlock();
 430
 431                 if (!start || start - offset > max_pages)
 432                         return 0;
 433
 434                 ra->start = start;
 435                 ra->size = start - offset;      /* old async_size */
 436                 ra->size += req_size;
 437                 ra->size = get_next_ra_size(ra, max_pages);
 438                 ra->async_size = ra->size;
 439                 goto readit;
 440         }
 441
 442         /*
 443          * oversize read
 444          */
 445         if (req_size > max_pages)
 446                 goto initial_readahead;
 447
 448         /*
 449          * sequential cache miss
 450          * trivial case: (offset - prev_offset) == 1
 451          * unaligned reads: (offset - prev_offset) == 0
 452          */
 453         prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
 454         if (offset - prev_offset <= 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_pages))
 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
 470 initial_readahead:
 471         ra->start = offset;
 472         ra->size = get_init_ra_size(req_size, max_pages);
 473         ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 474
 475 readit:
 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          * Take care of maximum IO pages as above.
 481          */
 482         if (offset == ra->start && ra->size == ra->async_size) {
 483                 add_pages = get_next_ra_size(ra, max_pages);
 484                 if (ra->size + add_pages <= max_pages) {
 485                         ra->async_size = add_pages;
 486                         ra->size += add_pages;
 487                 } else {
 488                         ra->size = max_pages;
 489                         ra->async_size = max_pages >> 1;
 490                 }
 491         }
 492
 493         return ra_submit(ra, mapping, filp);
 494 }
 495
 496 /**
 497  * page_cache_sync_readahead - generic file readahead
 498  * @mapping: address_space which holds the pagecache and I/O vectors
 499  * @ra: file_ra_state which holds the readahead state
 500  * @filp: passed on to ->readpage() and ->readpages()
 501  * @offset: start offset into @mapping, in pagecache page-sized units
 502  * @req_size: hint: total size of the read which the caller is performing in
 503  *            pagecache pages
 504  *
 505  * page_cache_sync_readahead() should be called when a cache miss happened:
 506  * it will submit the read.  The readahead logic may decide to piggyback more
 507  * pages onto the read request if access patterns suggest it will improve
 508  * performance.
 509  */
 510 void page_cache_sync_readahead(struct address_space *mapping,
 511                                struct file_ra_state *ra, struct file *filp,
 512                                pgoff_t offset, unsigned long req_size)
 513 {
 514         /* no read-ahead */
 515         if (!ra->ra_pages)
 516                 return;
 517
 518         if (blk_cgroup_congested())
 519                 return;
 520
 521         /* be dumb */
 522         if (filp && (filp->f_mode & FMODE_RANDOM)) {
 523                 force_page_cache_readahead(mapping, filp, offset, req_size);
 524                 return;
 525         }
 526
 527         /* do read-ahead */
 528         ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 529 }
 530 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 531
 532 /**
 533  * page_cache_async_readahead - file readahead for marked pages
 534  * @mapping: address_space which holds the pagecache and I/O vectors
 535  * @ra: file_ra_state which holds the readahead state
 536  * @filp: passed on to ->readpage() and ->readpages()
 537  * @page: the page at @offset which has the PG_readahead flag set
 538  * @offset: start offset into @mapping, in pagecache page-sized units
 539  * @req_size: hint: total size of the read which the caller is performing in
 540  *            pagecache pages
 541  *
 542  * page_cache_async_readahead() should be called when a page is used which
 543  * has the PG_readahead flag; this is a marker to suggest that the application
 544  * has used up enough of the readahead window that we should start pulling in
 545  * more pages.
 546  */
 547 void
 548 page_cache_async_readahead(struct address_space *mapping,
 549                            struct file_ra_state *ra, struct file *filp,
 550                            struct page *page, pgoff_t offset,
 551                            unsigned long req_size)
 552 {
 553         /* no read-ahead */
 554         if (!ra->ra_pages)
 555                 return;
 556
 557         /*
 558          * Same bit is used for PG_readahead and PG_reclaim.
 559          */
 560         if (PageWriteback(page))
 561                 return;
 562
 563         ClearPageReadahead(page);
 564
 565         /*
 566          * Defer asynchronous read-ahead on IO congestion.
 567          */
 568         if (inode_read_congested(mapping->host))
 569                 return;
 570
 571         if (blk_cgroup_congested())
 572                 return;
 573
 574         /* do read-ahead */
 575         ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 576 }
 577 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
 578
 579 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
 580 {
 581         ssize_t ret;
 582         struct fd f;
 583
 584         ret = -EBADF;
 585         f = fdget(fd);
 586         if (!f.file || !(f.file->f_mode & FMODE_READ))
 587                 goto out;
 588
 589         /*
 590          * The readahead() syscall is intended to run only on files
 591          * that can execute readahead. If readahead is not possible
 592          * on this file, then we must return -EINVAL.
 593          */
 594         ret = -EINVAL;
 595         if (!f.file->f_mapping || !f.file->f_mapping->a_ops ||
 596             !S_ISREG(file_inode(f.file)->i_mode))
 597                 goto out;
 598
 599         ret = vfs_fadvise(f.file, offset, count, POSIX_FADV_WILLNEED);
 600 out:
 601         fdput(f);
 602         return ret;
 603 }
 604
 605 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
 606 {
 607         return ksys_readahead(fd, offset, count);
 608 }