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 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  */
 150 int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
 151                         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         int ret = 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                         continue;
 182
 183                 page = __page_cache_alloc(gfp_mask);
 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, gfp_mask);
 200         BUG_ON(!list_empty(&page_pool));
 201 out:
 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  */
 209 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 210                                pgoff_t offset, unsigned long nr_to_read)
 211 {
 212         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 213         struct file_ra_state *ra = &filp->f_ra;
 214         unsigned long max_pages;
 215
 216         if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
 217                 return -EINVAL;
 218
 219         /*
 220          * If the request exceeds the readahead window, allow the read to
 221          * be up to the optimal hardware IO size
 222          */
 223         max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
 224         nr_to_read = min(nr_to_read, max_pages);
 225         while (nr_to_read) {
 226                 int err;
 227
 228                 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
 229
 230                 if (this_chunk > nr_to_read)
 231                         this_chunk = nr_to_read;
 232                 err = __do_page_cache_readahead(mapping, filp,
 233                                                 offset, this_chunk, 0);
 234                 if (err < 0)
 235                         return err;
 236
 237                 offset += this_chunk;
 238                 nr_to_read -= this_chunk;
 239         }
 240         return 0;
 241 }
 242
 243 /*
 244  * Set the initial window size, round to next power of 2 and square
 245  * for small size, x 4 for medium, and x 2 for large
 246  * for 128k (32 page) max ra
 247  * 1-8 page = 32k initial, > 8 page = 128k initial
 248  */
 249 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 250 {
 251         unsigned long newsize = roundup_pow_of_two(size);
 252
 253         if (newsize <= max / 32)
 254                 newsize = newsize * 4;
 255         else if (newsize <= max / 4)
 256                 newsize = newsize * 2;
 257         else
 258                 newsize = max;
 259
 260         return newsize;
 261 }
 262
 263 /*
 264  *  Get the previous window size, ramp it up, and
 265  *  return it as the new window size.
 266  */
 267 static unsigned long get_next_ra_size(struct file_ra_state *ra,
 268                                                 unsigned long max)
 269 {
 270         unsigned long cur = ra->size;
 271         unsigned long newsize;
 272
 273         if (cur < max / 16)
 274                 newsize = 4 * cur;
 275         else
 276                 newsize = 2 * cur;
 277
 278         return min(newsize, max);
 279 }
 280
 281 /*
 282  * On-demand readahead design.
 283  *
 284  * The fields in struct file_ra_state represent the most-recently-executed
 285  * readahead attempt:
 286  *
 287  *                        |<----- async_size ---------|
 288  *     |------------------- size -------------------->|
 289  *     |==================#===========================|
 290  *     ^start             ^page marked with PG_readahead
 291  *
 292  * To overlap application thinking time and disk I/O time, we do
 293  * `readahead pipelining': Do not wait until the application consumed all
 294  * readahead pages and stalled on the missing page at readahead_index;
 295  * Instead, submit an asynchronous readahead I/O as soon as there are
 296  * only async_size pages left in the readahead window. Normally async_size
 297  * will be equal to size, for maximum pipelining.
 298  *
 299  * In interleaved sequential reads, concurrent streams on the same fd can
 300  * be invalidating each other's readahead state. So we flag the new readahead
 301  * page at (start+size-async_size) with PG_readahead, and use it as readahead
 302  * indicator. The flag won't be set on already cached pages, to avoid the
 303  * readahead-for-nothing fuss, saving pointless page cache lookups.
 304  *
 305  * prev_pos tracks the last visited byte in the _previous_ read request.
 306  * It should be maintained by the caller, and will be used for detecting
 307  * small random reads. Note that the readahead algorithm checks loosely
 308  * for sequential patterns. Hence interleaved reads might be served as
 309  * sequential ones.
 310  *
 311  * There is a special-case: if the first page which the application tries to
 312  * read happens to be the first page of the file, it is assumed that a linear
 313  * read is about to happen and the window is immediately set to the initial size
 314  * based on I/O request size and the max_readahead.
 315  *
 316  * The code ramps up the readahead size aggressively at first, but slow down as
 317  * it approaches max_readhead.
 318  */
 319
 320 /*
 321  * Count contiguously cached pages from @offset-1 to @offset-@max,
 322  * this count is a conservative estimation of
 323  *      - length of the sequential read sequence, or
 324  *      - thrashing threshold in memory tight systems
 325  */
 326 static pgoff_t count_history_pages(struct address_space *mapping,
 327                                    pgoff_t offset, unsigned long max)
 328 {
 329         pgoff_t head;
 330
 331         rcu_read_lock();
 332         head = page_cache_prev_hole(mapping, offset - 1, max);
 333         rcu_read_unlock();
 334
 335         return offset - 1 - head;
 336 }
 337
 338 /*
 339  * page cache context based read-ahead
 340  */
 341 static int try_context_readahead(struct address_space *mapping,
 342                                  struct file_ra_state *ra,
 343                                  pgoff_t offset,
 344                                  unsigned long req_size,
 345                                  unsigned long max)
 346 {
 347         pgoff_t size;
 348
 349         size = count_history_pages(mapping, offset, max);
 350
 351         /*
 352          * not enough history pages:
 353          * it could be a random read
 354          */
 355         if (size <= req_size)
 356                 return 0;
 357
 358         /*
 359          * starts from beginning of file:
 360          * it is a strong indication of long-run stream (or whole-file-read)
 361          */
 362         if (size >= offset)
 363                 size *= 2;
 364
 365         ra->start = offset;
 366         ra->size = min(size + req_size, max);
 367         ra->async_size = 1;
 368
 369         return 1;
 370 }
 371
 372 /*
 373  * A minimal readahead algorithm for trivial sequential/random reads.
 374  */
 375 static unsigned long
 376 ondemand_readahead(struct address_space *mapping,
 377                    struct file_ra_state *ra, struct file *filp,
 378                    bool hit_readahead_marker, pgoff_t offset,
 379                    unsigned long req_size)
 380 {
 381         struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
 382         unsigned long max_pages = ra->ra_pages;
 383         pgoff_t prev_offset;
 384
 385         /*
 386          * If the request exceeds the readahead window, allow the read to
 387          * be up to the optimal hardware IO size
 388          */
 389         if (req_size > max_pages && bdi->io_pages > max_pages)
 390                 max_pages = min(req_size, bdi->io_pages);
 391
 392         /*
 393          * start of file
 394          */
 395         if (!offset)
 396                 goto initial_readahead;
 397
 398         /*
 399          * It's the expected callback offset, assume sequential access.
 400          * Ramp up sizes, and push forward the readahead window.
 401          */
 402         if ((offset == (ra->start + ra->size - ra->async_size) ||
 403              offset == (ra->start + ra->size))) {
 404                 ra->start += ra->size;
 405                 ra->size = get_next_ra_size(ra, max_pages);
 406                 ra->async_size = ra->size;
 407                 goto readit;
 408         }
 409
 410         /*
 411          * Hit a marked page without valid readahead state.
 412          * E.g. interleaved reads.
 413          * Query the pagecache for async_size, which normally equals to
 414          * readahead size. Ramp it up and use it as the new readahead size.
 415          */
 416         if (hit_readahead_marker) {
 417                 pgoff_t start;
 418
 419                 rcu_read_lock();
 420                 start = page_cache_next_hole(mapping, offset + 1, max_pages);
 421                 rcu_read_unlock();
 422
 423                 if (!start || start - offset > max_pages)
 424                         return 0;
 425
 426                 ra->start = start;
 427                 ra->size = start - offset;      /* old async_size */
 428                 ra->size += req_size;
 429                 ra->size = get_next_ra_size(ra, max_pages);
 430                 ra->async_size = ra->size;
 431                 goto readit;
 432         }
 433
 434         /*
 435          * oversize read
 436          */
 437         if (req_size > max_pages)
 438                 goto initial_readahead;
 439
 440         /*
 441          * sequential cache miss
 442          * trivial case: (offset - prev_offset) == 1
 443          * unaligned reads: (offset - prev_offset) == 0
 444          */
 445         prev_offset = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
 446         if (offset - prev_offset <= 1UL)
 447                 goto initial_readahead;
 448
 449         /*
 450          * Query the page cache and look for the traces(cached history pages)
 451          * that a sequential stream would leave behind.
 452          */
 453         if (try_context_readahead(mapping, ra, offset, req_size, max_pages))
 454                 goto readit;
 455
 456         /*
 457          * standalone, small random read
 458          * Read as is, and do not pollute the readahead state.
 459          */
 460         return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
 461
 462 initial_readahead:
 463         ra->start = offset;
 464         ra->size = get_init_ra_size(req_size, max_pages);
 465         ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
 466
 467 readit:
 468         /*
 469          * Will this read hit the readahead marker made by itself?
 470          * If so, trigger the readahead marker hit now, and merge
 471          * the resulted next readahead window into the current one.
 472          */
 473         if (offset == ra->start && ra->size == ra->async_size) {
 474                 ra->async_size = get_next_ra_size(ra, max_pages);
 475                 ra->size += ra->async_size;
 476         }
 477
 478         return ra_submit(ra, mapping, filp);
 479 }
 480
 481 /**
 482  * page_cache_sync_readahead - generic file readahead
 483  * @mapping: address_space which holds the pagecache and I/O vectors
 484  * @ra: file_ra_state which holds the readahead state
 485  * @filp: passed on to ->readpage() and ->readpages()
 486  * @offset: start offset into @mapping, in pagecache page-sized units
 487  * @req_size: hint: total size of the read which the caller is performing in
 488  *            pagecache pages
 489  *
 490  * page_cache_sync_readahead() should be called when a cache miss happened:
 491  * it will submit the read.  The readahead logic may decide to piggyback more
 492  * pages onto the read request if access patterns suggest it will improve
 493  * performance.
 494  */
 495 void page_cache_sync_readahead(struct address_space *mapping,
 496                                struct file_ra_state *ra, struct file *filp,
 497                                pgoff_t offset, unsigned long req_size)
 498 {
 499         /* no read-ahead */
 500         if (!ra->ra_pages)
 501                 return;
 502
 503         /* be dumb */
 504         if (filp && (filp->f_mode & FMODE_RANDOM)) {
 505                 force_page_cache_readahead(mapping, filp, offset, req_size);
 506                 return;
 507         }
 508
 509         /* do read-ahead */
 510         ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 511 }
 512 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
 513
 514 /**
 515  * page_cache_async_readahead - file readahead for marked pages
 516  * @mapping: address_space which holds the pagecache and I/O vectors
 517  * @ra: file_ra_state which holds the readahead state
 518  * @filp: passed on to ->readpage() and ->readpages()
 519  * @page: the page at @offset which has the PG_readahead flag set
 520  * @offset: start offset into @mapping, in pagecache page-sized units
 521  * @req_size: hint: total size of the read which the caller is performing in
 522  *            pagecache pages
 523  *
 524  * page_cache_async_readahead() should be called when a page is used which
 525  * has the PG_readahead flag; this is a marker to suggest that the application
 526  * has used up enough of the readahead window that we should start pulling in
 527  * more pages.
 528  */
 529 void
 530 page_cache_async_readahead(struct address_space *mapping,
 531                            struct file_ra_state *ra, struct file *filp,
 532                            struct page *page, pgoff_t offset,
 533                            unsigned long req_size)
 534 {
 535         /* no read-ahead */
 536         if (!ra->ra_pages)
 537                 return;
 538
 539         /*
 540          * Same bit is used for PG_readahead and PG_reclaim.
 541          */
 542         if (PageWriteback(page))
 543                 return;
 544
 545         ClearPageReadahead(page);
 546
 547         /*
 548          * Defer asynchronous read-ahead on IO congestion.
 549          */
 550         if (inode_read_congested(mapping->host))
 551                 return;
 552
 553         /* do read-ahead */
 554         ondemand_readahead(mapping, ra, filp, true, offset, req_size);
 555 }
 556 EXPORT_SYMBOL_GPL(page_cache_async_readahead);
 557
 558 static ssize_t
 559 do_readahead(struct address_space *mapping, struct file *filp,
 560              pgoff_t index, unsigned long nr)
 561 {
 562         if (!mapping || !mapping->a_ops)
 563                 return -EINVAL;
 564
 565         /*
 566          * Readahead doesn't make sense for DAX inodes, but we don't want it
 567          * to report a failure either.  Instead, we just return success and
 568          * don't do any work.
 569          */
 570         if (dax_mapping(mapping))
 571                 return 0;
 572
 573         return force_page_cache_readahead(mapping, filp, index, nr);
 574 }
 575
 576 ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
 577 {
 578         ssize_t ret;
 579         struct fd f;
 580
 581         ret = -EBADF;
 582         f = fdget(fd);
 583         if (f.file) {
 584                 if (f.file->f_mode & FMODE_READ) {
 585                         struct address_space *mapping = f.file->f_mapping;
 586                         pgoff_t start = offset >> PAGE_SHIFT;
 587                         pgoff_t end = (offset + count - 1) >> PAGE_SHIFT;
 588                         unsigned long len = end - start + 1;
 589                         ret = do_readahead(mapping, f.file, start, len);
 590                 }
 591                 fdput(f);
 592         }
 593         return ret;
 594 }
 595
 596 SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
 597 {
 598         return ksys_readahead(fd, offset, count);
 599 }