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
  23 #include "internal.h"
  24
  25 /*
  26  * Initialise a struct file's readahead state.  Assumes that the caller has
  27  * memset *ra to zero.
  28  */
  29 void
  30 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
  31 {
  32         ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
  33         ra->prev_pos = -1;
  34 }
  35 EXPORT_SYMBOL_GPL(file_ra_state_init);
  36
  37 /*
  38  * see if a page needs releasing upon read_cache_pages() failure
  39  * - the caller of read_cache_pages() may have set PG_private or PG_fscache
  40  *   before calling, such as the NFS fs marking pages that are cached locally
  41  *   on disk, thus we need to give the fs a chance to clean up in the event of
  42  *   an error
  43  */
  44 static void read_cache_pages_invalidate_page(struct address_space *mapping,
  45                                              struct page *page)
  46 {
  47         if (page_has_private(page)) {
  48                 if (!trylock_page(page))
  49                         BUG();
  50                 page->mapping = mapping;
  51                 do_invalidatepage(page, 0, PAGE_SIZE);
  52                 page->mapping = NULL;
  53                 unlock_page(page);
  54         }
  55         put_page(page);
  56 }
  57
  58 /*
  59  * release a list of pages, invalidating them first if need be
  60  */
  61 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
  62                                               struct list_head *pages)
  63 {
  64         struct page *victim;
  65
  66         while (!list_empty(pages)) {
  67                 victim = lru_to_page(pages);
  68                 list_del(&victim->lru);
  69                 read_cache_pages_invalidate_page(mapping, victim);
  70         }
  71 }
  72
  73 /**
  74  * read_cache_pages - populate an address space with some pages & start reads against them
  75  * @mapping: the address_space
  76  * @pages: The address of a list_head which contains the target pages.  These
  77  *   pages have their ->index populated and are otherwise uninitialised.
  78  * @filler: callback routine for filling a single page.
  79  * @data: private data for the callback routine.
  80  *
  81  * Hides the details of the LRU cache etc from the filesystems.
  82  */
  83 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
  84                         int (*filler)(void *, struct page *), void *data)
  85 {
  86         struct page *page;
  87         int ret = 0;
  88
  89         while (!list_empty(pages)) {
  90                 page = lru_to_page(pages);
  91                 list_del(&page->lru);
  92                 if (add_to_page_cache_lru(page, mapping, page->index,
  93                                 readahead_gfp_mask(mapping))) {
  94                         read_cache_pages_invalidate_page(mapping, page);
  95                         continue;
  96                 }
  97                 put_page(page);
  98
  99                 ret = filler(data, page);
 100                 if (unlikely(ret)) {
 101                         read_cache_pages_invalidate_pages(mapping, pages);
 102                         break;
 103                 }
 104                 task_io_account_read(PAGE_SIZE);
 105         }
 106         return ret;
 107 }
 108
 109 EXPORT_SYMBOL(read_cache_pages);
 110
 111 static int read_pages(struct address_space *mapping, struct file *filp,
 112                 struct list_head *pages, unsigned int nr_pages, gfp_t gfp)
 113 {
 114         struct blk_plug plug;
 115         unsigned page_idx;
 116         int ret;
 117
 118         blk_start_plug(&plug);
 119
 120         if (mapping->a_ops->readpages) {
 121                 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
 122                 /* Clean up the remaining pages */
 123                 put_pages_list(pages);
 124                 goto out;
 125         }
 126
 127         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 128                 struct page *page = lru_to_page(pages);
 129                 list_del(&page->lru);
 130                 if (!add_to_page_cache_lru(page, mapping, page->index, gfp))
 131                         mapping->a_ops->readpage(filp, page);
 132                 put_page(page);
 133         }
 134         ret = 0;
 135
 136 out:
 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
 144  * the pages first, then submits them 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  */
 150 unsigned int __do_page_cache_readahead(struct address_space *mapping,
 151                 struct file *filp, pgoff_t offset, unsigned long nr_to_read,
 152                 unsigned long lookahead_size)
 153 {
 154         struct inode *inode = mapping->host;
 155         struct page *page;
 156         unsigned long end_index;        /* The last page we want to read */
 157         LIST_HEAD(page_pool);
 158         int page_idx;
 159         unsigned int nr_pages = 0;
 160         loff_t isize = i_size_read(inode);
 161         gfp_t gfp_mask = readahead_gfp_mask(mapping);
 162
 163         if (isize == 0)
 164                 goto out;
 165
 166         end_index = ((isize - 1) >> PAGE_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->i_pages, page_offset);
 179                 rcu_read_unlock();
 180                 if (page && !radix_tree_exceptional_entry(page)) {
 181                         /*
 182                          * Page already present?  Kick off the current batch of
 183                          * contiguous pages before continuing with the next
 184                          * batch.
 185                          */
 186                         if (nr_pages)
 187                                 read_pages(mapping, filp, &page_pool, nr_pages,
 188                                                 gfp_mask);
 189                         nr_pages = 0;
 190                         continue;
 191                 }
 192
 193                 page = __page_cache_alloc(gfp_mask);
 194                 if (!page)
 195                         break;
 196                 page->index = page_offset;
 197                 list_add(&page->lru, &page_pool);
 198                 if (page_idx == nr_to_read - lookahead_size)
 199                         SetPageReadahead(page);
 200                 nr_pages++;
 201         }
 202
 203         /*
 204          * Now start the IO.  We ignore I/O errors - if the page is not
 205          * uptodate then the caller will launch readpage again, and
 206          * will then handle the error.
 207          */
 208         if (nr_pages)
 209                 read_pages(mapping, filp, &page_pool, nr_pages, gfp_mask);
 210         BUG_ON(!list_empty(&page_pool));
 211 out:
 212         return nr_pages;
 213 }
 214
 215 /*
 216  * Chunk the readahead into 2 megabyte units, so that we don't pin too much
 217  * memory at once.
 218  */
 219 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 220                                pgoff_t offset, unsigned long nr_to_read)
 221 {
 222         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 223         struct file_ra_state *ra = &filp->f_ra;
 224         unsigned long max_pages;
 225
 226         if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 227                 return -EINVAL;
 228
 229         /*
 230          * If the request exceeds the readahead window, allow the read to
 231          * be up to the optimal hardware IO size
 232          */
 233         max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
 234         nr_to_read = min(nr_to_read, max_pages);
 235         while (nr_to_read) {
 236                 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
 237
 238                 if (this_chunk > nr_to_read)
 239                         this_chunk = nr_to_read;
 240                 __do_page_cache_readahead(mapping, filp, offset, this_chunk, 0);
 241
 242                 offset += this_chunk;
 243                 nr_to_read -= this_chunk;
 244         }
 245         return 0;
 246 }
 247
 248 /*
 249  * Set the initial window size, round to next power of 2 and square
 250  * for small size, x 4 for medium, and x 2 for large
 251  * for 128k (32 page) max ra
 252  * 1-8 page = 32k initial, > 8 page = 128k initial
 253  */
 254 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 255 {
 256         unsigned long newsize = roundup_pow_of_two(size);
 257
 258         if (newsize <= max / 32)
 259                 newsize = newsize * 4;
 260         else if (newsize <= max / 4)
 261                 newsize = newsize * 2;
 262         else
 263                 newsize = max;
 264
 265         return newsize;
 266 }
 267
 268 /*
 269  *  Get the previous window size, ramp it up, and
 270  *  return it as the new window size.
 271  */
 272 static unsigned long get_next_ra_size(struct file_ra_state *ra,
 273                                                 unsigned long max)
 274 {
 275         unsigned long cur = ra->size;
 276         unsigned long newsize;
 277
 278         if (cur < max / 16)
 279                 newsize = 4 * cur;
 280         else
 281                 newsize = 2 * cur;
 282
 283         return min(newsize, max);
 284 }
 285
 286 /*
 287  * On-demand readahead design.
 288  *
 289  * The fields in struct file_ra_state represent the most-recently-executed
 290  * readahead attempt:
 291  *
 292  *                        |<----- async_size ---------|
 293  *     |------------------- size -------------------->|
 294  *     |==================#===========================|
 295  *     ^start             ^page marked with PG_readahead
 296  *
 297  * To overlap application thinking time and disk I/O time, we do
 298  * `readahead pipelining': Do not wait until the application consumed all
 299  * readahead pages and stalled on the missing page at readahead_index;
 300  * Instead, submit an asynchronous readahead I/O as soon as there are
 301  * only async_size pages left in the readahead window. Normally async_size
 302  * will be equal to size, for maximum pipelining.
 303  *
 304  * In interleaved sequential reads, concurrent streams on the same fd can
 305  * be invalidating each other's readahead state. So we flag the new readahead
 306  * page at (start+size-async_size) with PG_readahead, and use it as readahead
 307  * indicator. The flag won't be set on already cached pages, to avoid the
 308  * readahead-for-nothing fuss, saving pointless page cache lookups.
 309  *
 310  * prev_pos tracks the last visited byte in the _previous_ read request.
 311  * It should be maintained by the caller, and will be used for detecting
 312  * small random reads. Note that the readahead algorithm checks loosely
 313  * for sequential patterns. Hence interleaved reads might be served as
 314  * sequential ones.
 315  *
 316  * There is a special-case: if the first page which the application tries to
 317  * read happens to be the first page of the file, it is assumed that a linear
 318  * read is about to happen and the window is immediately set to the initial size
 319  * based on I/O request size and the max_readahead.
 320  *
 321  * The code ramps up the readahead size aggressively at first, but slow down as
 322  * it approaches max_readhead.
 323  */
 324
 325 /*
 326  * Count contiguously cached pages from @offset-1 to @offset-@max,
 327  * this count is a conservative estimation of
 328  *      - length of the sequential read sequence, or
 329  *      - thrashing threshold in memory tight systems
 330  */
 331 static pgoff_t count_history_pages(struct address_space *mapping,
 332                                    pgoff_t offset, unsigned long max)
 333 {
 334         pgoff_t head;
 335
 336         rcu_read_lock();
 337         head = page_cache_prev_hole(mapping, offset - 1, max);
 338         rcu_read_unlock();
 339
 340         return offset - 1 - head;
 341 }
 342
 343 /*
 344  * page cache context based read-ahead
 345  */
 346 static int try_context_readahead(struct address_space *mapping,
 347                                  struct file_ra_state *ra,
 348                                  pgoff_t offset,
 349                                  unsigned long req_size,
 350                                  unsigned long max)
 351 {
 352         pgoff_t size;
 353
 354         size = count_history_pages(mapping, offset, max);
 355
 356         /*
 357          * not enough history pages:
 358          * it could be a random read
 359          */
 360         if (size <= req_size)
 361                 return 0;
 362
 363         /*
 364          * starts from beginning of file:
 365          * it is a strong indication of long-run stream (or whole-file-read)
 366          */
 367         if (size >= offset)
 368                 size *= 2;
 369
 370         ra->start = offset;
 371         ra->size = min(size + req_size, max);
 372         ra->async_size = 1;
 373
 374         return 1;
 375 }
 376
 377 /*
 378  * A minimal readahead algorithm for trivial sequential/random reads.
 379  */
 380 static unsigned long
 381 ondemand_readahead(struct address_space *mapping,
 382                    struct file_ra_state *ra, struct file *filp,
 383                    bool hit_readahead_marker, pgoff_t offset,
 384                    unsigned long req_size)
 385 {
 386         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 387         unsigned long max_pages = ra->ra_pages;
 388         pgoff_t prev_offset;
 389
 390         /*
 391          * If the request exceeds the readahead window, allow the read to
 392          * be up to the optimal hardware IO size
 393          */
 394         if (req_size > max_pages && bdi->io_pages > max_pages)
 395                 max_pages = min(req_size, bdi->io_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_pages);
 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 = page_cache_next_hole(mapping, offset + 1, max_pages);
 426                 rcu_read_unlock();
 427
 428                 if (!start || start - offset > max_pages)
 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_pages);
 435                 ra->async_size = ra->size;
 436                 goto readit;
 437         }
 438
 439         /*
 440          * oversize read
 441          */
 442         if (req_size > max_pages)
 443                 goto initial_readahead;
 444
 445         /*
 446          * sequential cache miss
 447          * trivial case: (offset - prev_offset) == 1
 448          * unaligned reads: (offset - prev_offset) == 0
 449          */
 450         prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
 451         if (offset - prev_offset <= 1UL)
 452                 goto initial_readahead;
 453
 454         /*
 455          * Query the page cache and look for the traces(cached history pages)
 456          * that a sequential stream would leave behind.
 457          */
 458         if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
 459                 goto readit;
 460
 461         /*
 462          * standalone, small random read
 463          * Read as is, and do not pollute the readahead state.
 464          */
 465         return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
 466
 467 initial_readahead:
 468         ra->start = offset;
 469         ra->size = get_init_ra_size(req_size, max_pages);
 470         ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 471
 472 readit:
 473         /*
 474          * Will this read hit the readahead marker made by itself?
 475          * If so, trigger the readahead marker hit now, and merge
 476          * the resulted next readahead window into the current one.
 477          */
 478         if (offset == ra->start && ra->size == ra->async_size) {
 479                 ra->async_size = get_next_ra_size(ra, max_pages);
 480                 ra->size += ra->async_size;
 481         }
 482
 483         return ra_submit(ra, mapping, filp);
 484 }
 485
 486 /**
 487  * page_cache_sync_readahead - generic file readahead
 488  * @mapping: address_space which holds the pagecache and I/O vectors
 489  * @ra: file_ra_state which holds the readahead state
 490  * @filp: passed on to ->readpage() and ->readpages()
 491  * @offset: start offset into @mapping, in pagecache page-sized units
 492  * @req_size: hint: total size of the read which the caller is performing in
 493  *            pagecache pages
 494  *
 495  * page_cache_sync_readahead() should be called when a cache miss happened:
 496  * it will submit the read.  The readahead logic may decide to piggyback more
 497  * pages onto the read request if access patterns suggest it will improve
 498  * performance.
 499  */
 500 void page_cache_sync_readahead(struct address_space *mapping,
 501                                struct file_ra_state *ra, struct file *filp,
 502                                pgoff_t offset, unsigned long req_size)
 503 {
 504         /* no read-ahead */
 505         if (!ra->ra_pages)
 506                 return;
 507
 508         /* be dumb */
 509         if (filp && (filp->f_mode & FMODE_RANDOM)) {
 510                 force_page_cache_readahead(mapping, filp, offset, req_size);
 511                 return;
 512         }
 513
 514         /* do read-ahead */
 515         ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 516 }
 517 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 518
 519 /**
 520  * page_cache_async_readahead - file readahead for marked pages
 521  * @mapping: address_space which holds the pagecache and I/O vectors
 522  * @ra: file_ra_state which holds the readahead state
 523  * @filp: passed on to ->readpage() and ->readpages()
 524  * @page: the page at @offset which has the PG_readahead flag set
 525  * @offset: start offset into @mapping, in pagecache page-sized units
 526  * @req_size: hint: total size of the read which the caller is performing in
 527  *            pagecache pages
 528  *
 529  * page_cache_async_readahead() should be called when a page is used which
 530  * has the PG_readahead flag; this is a marker to suggest that the application
 531  * has used up enough of the readahead window that we should start pulling in
 532  * more pages.
 533  */
 534 void
 535 page_cache_async_readahead(struct address_space *mapping,
 536                            struct file_ra_state *ra, struct file *filp,
 537                            struct page *page, pgoff_t offset,
 538                            unsigned long req_size)
 539 {
 540         /* no read-ahead */
 541         if (!ra->ra_pages)
 542                 return;
 543
 544         /*
 545          * Same bit is used for PG_readahead and PG_reclaim.
 546          */
 547         if (PageWriteback(page))
 548                 return;
 549
 550         ClearPageReadahead(page);
 551
 552         /*
 553          * Defer asynchronous read-ahead on IO congestion.
 554          */
 555         if (inode_read_congested(mapping->host))
 556                 return;
 557
 558         /* do read-ahead */
 559         ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 560 }
 561 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
 562
 563 static ssize_t
 564 do_readahead(struct address_space *mapping, struct file *filp,
 565              pgoff_t index, unsigned long nr)
 566 {
 567         if (!mapping || !mapping->a_ops)
 568                 return -EINVAL;
 569
 570         /*
 571          * Readahead doesn't make sense for DAX inodes, but we don't want it
 572          * to report a failure either.  Instead, we just return success and
 573          * don't do any work.
 574          */
 575         if (dax_mapping(mapping))
 576                 return 0;
 577
 578         return force_page_cache_readahead(mapping, filp, index, nr);
 579 }
 580
 581 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
 582 {
 583         ssize_t ret;
 584         struct fd f;
 585
 586         ret = -EBADF;
 587         f = fdget(fd);
 588         if (f.file) {
 589                 if (f.file->f_mode & FMODE_READ) {
 590                         struct address_space *mapping = f.file->f_mapping;
 591                         pgoff_t start = offset >> PAGE_SHIFT;
 592                         pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
 593                         unsigned long len = end - start + 1;
 594                         ret = do_readahead(mapping, f.file, start, len);
 595                 }
 596                 fdput(f);
 597         }
 598         return ret;
 599 }
 600
 601 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
 602 {
 603         return ksys_readahead(fd, offset, count);
 604 }