mm/swap_state.c

   1 /*
   2  *  linux/mm/swap_state.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  *
   7  *  Rewritten to use page cache, (C) 1998 Stephen Tweedie
   8  */
   9 #include <linux/mm.h>
  10 #include <linux/gfp.h>
  11 #include <linux/kernel_stat.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/init.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/pagevec.h>
  19 #include <linux/migrate.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/swap_slots.h>
  22 #include <linux/huge_mm.h>
  23
  24 #include <asm/pgtable.h>
  25
  26 /*
  27  * swapper_space is a fiction, retained to simplify the path through
  28  * vmscan's shrink_page_list.
  29  */
  30 static const struct address_space_operations swap_aops = {
  31         .writepage      = swap_writepage,
  32         .set_page_dirty = swap_set_page_dirty,
  33 #ifdef CONFIG_MIGRATION
  34         .migratepage    = migrate_page,
  35 #endif
  36 };
  37
  38 struct address_space *swapper_spaces[MAX_SWAPFILES];
  39 static unsigned int nr_swapper_spaces[MAX_SWAPFILES];
  40 bool swap_vma_readahead = true;
  41
  42 #define SWAP_RA_MAX_ORDER_DEFAULT       3
  43
  44 static int swap_ra_max_order = SWAP_RA_MAX_ORDER_DEFAULT;
  45
  46 #define SWAP_RA_WIN_SHIFT       (PAGE_SHIFT / 2)
  47 #define SWAP_RA_HITS_MASK       ((1UL << SWAP_RA_WIN_SHIFT) - 1)
  48 #define SWAP_RA_HITS_MAX        SWAP_RA_HITS_MASK
  49 #define SWAP_RA_WIN_MASK        (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
  50
  51 #define SWAP_RA_HITS(v)         ((v) & SWAP_RA_HITS_MASK)
  52 #define SWAP_RA_WIN(v)          (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
  53 #define SWAP_RA_ADDR(v)         ((v) & PAGE_MASK)
  54
  55 #define SWAP_RA_VAL(addr, win, hits)                            \
  56         (((addr) & PAGE_MASK) |                                 \
  57          (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |    \
  58          ((hits) & SWAP_RA_HITS_MASK))
  59
  60 /* Initial readahead hits is 4 to start up with a small window */
  61 #define GET_SWAP_RA_VAL(vma)                                    \
  62         (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
  63
  64 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  65 #define ADD_CACHE_INFO(x, nr)   do { swap_cache_info.x += (nr); } while (0)
  66
  67 static struct {
  68         unsigned long add_total;
  69         unsigned long del_total;
  70         unsigned long find_success;
  71         unsigned long find_total;
  72 } swap_cache_info;
  73
  74 unsigned long total_swapcache_pages(void)
  75 {
  76         unsigned int i, j, nr;
  77         unsigned long ret = 0;
  78         struct address_space *spaces;
  79
  80         rcu_read_lock();
  81         for (i = 0; i < MAX_SWAPFILES; i++) {
  82                 /*
  83                  * The corresponding entries in nr_swapper_spaces and
  84                  * swapper_spaces will be reused only after at least
  85                  * one grace period.  So it is impossible for them
  86                  * belongs to different usage.
  87                  */
  88                 nr = nr_swapper_spaces[i];
  89                 spaces = rcu_dereference(swapper_spaces[i]);
  90                 if (!nr || !spaces)
  91                         continue;
  92                 for (j = 0; j < nr; j++)
  93                         ret += spaces[j].nrpages;
  94         }
  95         rcu_read_unlock();
  96         return ret;
  97 }
  98
  99 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
 100
 101 void show_swap_cache_info(void)
 102 {
 103         printk("%lu pages in swap cache\n", total_swapcache_pages());
 104         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
 105                 swap_cache_info.add_total, swap_cache_info.del_total,
 106                 swap_cache_info.find_success, swap_cache_info.find_total);
 107         printk("Free swap  = %ldkB\n",
 108                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
 109         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
 110 }
 111
 112 /*
 113  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
 114  * but sets SwapCache flag and private instead of mapping and index.
 115  */
 116 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
 117 {
 118         int error, i, nr = hpage_nr_pages(page);
 119         struct address_space *address_space;
 120         pgoff_t idx = swp_offset(entry);
 121
 122         VM_BUG_ON_PAGE(!PageLocked(page), page);
 123         VM_BUG_ON_PAGE(PageSwapCache(page), page);
 124         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 125
 126         page_ref_add(page, nr);
 127         SetPageSwapCache(page);
 128
 129         address_space = swap_address_space(entry);
 130         spin_lock_irq(&address_space->tree_lock);
 131         for (i = 0; i < nr; i++) {
 132                 set_page_private(page + i, entry.val + i);
 133                 error = radix_tree_insert(&address_space->page_tree,
 134                                           idx + i, page + i);
 135                 if (unlikely(error))
 136                         break;
 137         }
 138         if (likely(!error)) {
 139                 address_space->nrpages += nr;
 140                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
 141                 ADD_CACHE_INFO(add_total, nr);
 142         } else {
 143                 /*
 144                  * Only the context which have set SWAP_HAS_CACHE flag
 145                  * would call add_to_swap_cache().
 146                  * So add_to_swap_cache() doesn't returns -EEXIST.
 147                  */
 148                 VM_BUG_ON(error == -EEXIST);
 149                 set_page_private(page + i, 0UL);
 150                 while (i--) {
 151                         radix_tree_delete(&address_space->page_tree, idx + i);
 152                         set_page_private(page + i, 0UL);
 153                 }
 154                 ClearPageSwapCache(page);
 155                 page_ref_sub(page, nr);
 156         }
 157         spin_unlock_irq(&address_space->tree_lock);
 158
 159         return error;
 160 }
 161
 162
 163 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 164 {
 165         int error;
 166
 167         error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
 168         if (!error) {
 169                 error = __add_to_swap_cache(page, entry);
 170                 radix_tree_preload_end();
 171         }
 172         return error;
 173 }
 174
 175 /*
 176  * This must be called only on pages that have
 177  * been verified to be in the swap cache.
 178  */
 179 void __delete_from_swap_cache(struct page *page)
 180 {
 181         struct address_space *address_space;
 182         int i, nr = hpage_nr_pages(page);
 183         swp_entry_t entry;
 184         pgoff_t idx;
 185
 186         VM_BUG_ON_PAGE(!PageLocked(page), page);
 187         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 188         VM_BUG_ON_PAGE(PageWriteback(page), page);
 189
 190         entry.val = page_private(page);
 191         address_space = swap_address_space(entry);
 192         idx = swp_offset(entry);
 193         for (i = 0; i < nr; i++) {
 194                 radix_tree_delete(&address_space->page_tree, idx + i);
 195                 set_page_private(page + i, 0);
 196         }
 197         ClearPageSwapCache(page);
 198         address_space->nrpages -= nr;
 199         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
 200         ADD_CACHE_INFO(del_total, nr);
 201 }
 202
 203 /**
 204  * add_to_swap - allocate swap space for a page
 205  * @page: page we want to move to swap
 206  *
 207  * Allocate swap space for the page and add the page to the
 208  * swap cache.  Caller needs to hold the page lock.
 209  */
 210 int add_to_swap(struct page *page)
 211 {
 212         swp_entry_t entry;
 213         int err;
 214
 215         VM_BUG_ON_PAGE(!PageLocked(page), page);
 216         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 217
 218         entry = get_swap_page(page);
 219         if (!entry.val)
 220                 return 0;
 221
 222         if (mem_cgroup_try_charge_swap(page, entry))
 223                 goto fail;
 224
 225         /*
 226          * Radix-tree node allocations from PF_MEMALLOC contexts could
 227          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 228          * stops emergency reserves from being allocated.
 229          *
 230          * TODO: this could cause a theoretical memory reclaim
 231          * deadlock in the swap out path.
 232          */
 233         /*
 234          * Add it to the swap cache.
 235          */
 236         err = add_to_swap_cache(page, entry,
 237                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 238         /* -ENOMEM radix-tree allocation failure */
 239         if (err)
 240                 /*
 241                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 242                  * clear SWAP_HAS_CACHE flag.
 243                  */
 244                 goto fail;
 245
 246         return 1;
 247
 248 fail:
 249         put_swap_page(page, entry);
 250         return 0;
 251 }
 252
 253 /*
 254  * This must be called only on pages that have
 255  * been verified to be in the swap cache and locked.
 256  * It will never put the page into the free list,
 257  * the caller has a reference on the page.
 258  */
 259 void delete_from_swap_cache(struct page *page)
 260 {
 261         swp_entry_t entry;
 262         struct address_space *address_space;
 263
 264         entry.val = page_private(page);
 265
 266         address_space = swap_address_space(entry);
 267         spin_lock_irq(&address_space->tree_lock);
 268         __delete_from_swap_cache(page);
 269         spin_unlock_irq(&address_space->tree_lock);
 270
 271         put_swap_page(page, entry);
 272         page_ref_sub(page, hpage_nr_pages(page));
 273 }
 274
 275 /*
 276  * If we are the only user, then try to free up the swap cache.
 277  *
 278  * Its ok to check for PageSwapCache without the page lock
 279  * here because we are going to recheck again inside
 280  * try_to_free_swap() _with_ the lock.
 281  *                                      - Marcelo
 282  */
 283 static inline void free_swap_cache(struct page *page)
 284 {
 285         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 286                 try_to_free_swap(page);
 287                 unlock_page(page);
 288         }
 289 }
 290
 291 /*
 292  * Perform a free_page(), also freeing any swap cache associated with
 293  * this page if it is the last user of the page.
 294  */
 295 void free_page_and_swap_cache(struct page *page)
 296 {
 297         free_swap_cache(page);
 298         if (!is_huge_zero_page(page))
 299                 put_page(page);
 300 }
 301
 302 /*
 303  * Passed an array of pages, drop them all from swapcache and then release
 304  * them.  They are removed from the LRU and freed if this is their last use.
 305  */
 306 void free_pages_and_swap_cache(struct page **pages, int nr)
 307 {
 308         struct page **pagep = pages;
 309         int i;
 310
 311         lru_add_drain();
 312         for (i = 0; i < nr; i++)
 313                 free_swap_cache(pagep[i]);
 314         release_pages(pagep, nr, false);
 315 }
 316
 317 /*
 318  * Lookup a swap entry in the swap cache. A found page will be returned
 319  * unlocked and with its refcount incremented - we rely on the kernel
 320  * lock getting page table operations atomic even if we drop the page
 321  * lock before returning.
 322  */
 323 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
 324                                unsigned long addr)
 325 {
 326         struct page *page;
 327         unsigned long ra_info;
 328         int win, hits, readahead;
 329
 330         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 331
 332         INC_CACHE_INFO(find_total);
 333         if (page) {
 334                 INC_CACHE_INFO(find_success);
 335                 if (unlikely(PageTransCompound(page)))
 336                         return page;
 337                 readahead = TestClearPageReadahead(page);
 338                 if (vma) {
 339                         ra_info = GET_SWAP_RA_VAL(vma);
 340                         win = SWAP_RA_WIN(ra_info);
 341                         hits = SWAP_RA_HITS(ra_info);
 342                         if (readahead)
 343                                 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
 344                         atomic_long_set(&vma->swap_readahead_info,
 345                                         SWAP_RA_VAL(addr, win, hits));
 346                 }
 347                 if (readahead) {
 348                         count_vm_event(SWAP_RA_HIT);
 349                         if (!vma)
 350                                 atomic_inc(&swapin_readahead_hits);
 351                 }
 352         }
 353         return page;
 354 }
 355
 356 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 357                         struct vm_area_struct *vma, unsigned long addr,
 358                         bool *new_page_allocated)
 359 {
 360         struct page *found_page, *new_page = NULL;
 361         struct address_space *swapper_space = swap_address_space(entry);
 362         int err;
 363         *new_page_allocated = false;
 364
 365         do {
 366                 /*
 367                  * First check the swap cache.  Since this is normally
 368                  * called after lookup_swap_cache() failed, re-calling
 369                  * that would confuse statistics.
 370                  */
 371                 found_page = find_get_page(swapper_space, swp_offset(entry));
 372                 if (found_page)
 373                         break;
 374
 375                 /*
 376                  * Just skip read ahead for unused swap slot.
 377                  * During swap_off when swap_slot_cache is disabled,
 378                  * we have to handle the race between putting
 379                  * swap entry in swap cache and marking swap slot
 380                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 381                  * else swap_off will be aborted if we return NULL.
 382                  */
 383                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 384                         break;
 385
 386                 /*
 387                  * Get a new page to read into from swap.
 388                  */
 389                 if (!new_page) {
 390                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 391                         if (!new_page)
 392                                 break;          /* Out of memory */
 393                 }
 394
 395                 /*
 396                  * call radix_tree_preload() while we can wait.
 397                  */
 398                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 399                 if (err)
 400                         break;
 401
 402                 /*
 403                  * Swap entry may have been freed since our caller observed it.
 404                  */
 405                 err = swapcache_prepare(entry);
 406                 if (err == -EEXIST) {
 407                         radix_tree_preload_end();
 408                         /*
 409                          * We might race against get_swap_page() and stumble
 410                          * across a SWAP_HAS_CACHE swap_map entry whose page
 411                          * has not been brought into the swapcache yet.
 412                          */
 413                         cond_resched();
 414                         continue;
 415                 }
 416                 if (err) {              /* swp entry is obsolete ? */
 417                         radix_tree_preload_end();
 418                         break;
 419                 }
 420
 421                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 422                 __SetPageLocked(new_page);
 423                 __SetPageSwapBacked(new_page);
 424                 err = __add_to_swap_cache(new_page, entry);
 425                 if (likely(!err)) {
 426                         radix_tree_preload_end();
 427                         /*
 428                          * Initiate read into locked page and return.
 429                          */
 430                         lru_cache_add_anon(new_page);
 431                         *new_page_allocated = true;
 432                         return new_page;
 433                 }
 434                 radix_tree_preload_end();
 435                 __ClearPageLocked(new_page);
 436                 /*
 437                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 438                  * clear SWAP_HAS_CACHE flag.
 439                  */
 440                 put_swap_page(new_page, entry);
 441         } while (err != -ENOMEM);
 442
 443         if (new_page)
 444                 put_page(new_page);
 445         return found_page;
 446 }
 447
 448 /*
 449  * Locate a page of swap in physical memory, reserving swap cache space
 450  * and reading the disk if it is not already cached.
 451  * A failure return means that either the page allocation failed or that
 452  * the swap entry is no longer in use.
 453  */
 454 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 455                 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
 456 {
 457         bool page_was_allocated;
 458         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 459                         vma, addr, &page_was_allocated);
 460
 461         if (page_was_allocated)
 462                 swap_readpage(retpage, do_poll);
 463
 464         return retpage;
 465 }
 466
 467 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
 468                                       unsigned long offset,
 469                                       int hits,
 470                                       int max_pages,
 471                                       int prev_win)
 472 {
 473         unsigned int pages, last_ra;
 474
 475         /*
 476          * This heuristic has been found to work well on both sequential and
 477          * random loads, swapping to hard disk or to SSD: please don't ask
 478          * what the "+ 2" means, it just happens to work well, that's all.
 479          */
 480         pages = hits + 2;
 481         if (pages == 2) {
 482                 /*
 483                  * We can have no readahead hits to judge by: but must not get
 484                  * stuck here forever, so check for an adjacent offset instead
 485                  * (and don't even bother to check whether swap type is same).
 486                  */
 487                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 488                         pages = 1;
 489         } else {
 490                 unsigned int roundup = 4;
 491                 while (roundup < pages)
 492                         roundup <<= 1;
 493                 pages = roundup;
 494         }
 495
 496         if (pages > max_pages)
 497                 pages = max_pages;
 498
 499         /* Don't shrink readahead too fast */
 500         last_ra = prev_win / 2;
 501         if (pages < last_ra)
 502                 pages = last_ra;
 503
 504         return pages;
 505 }
 506
 507 static unsigned long swapin_nr_pages(unsigned long offset)
 508 {
 509         static unsigned long prev_offset;
 510         unsigned int hits, pages, max_pages;
 511         static atomic_t last_readahead_pages;
 512
 513         max_pages = 1 << READ_ONCE(page_cluster);
 514         if (max_pages <= 1)
 515                 return 1;
 516
 517         hits = atomic_xchg(&swapin_readahead_hits, 0);
 518         pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
 519                                   atomic_read(&last_readahead_pages));
 520         if (!hits)
 521                 prev_offset = offset;
 522         atomic_set(&last_readahead_pages, pages);
 523
 524         return pages;
 525 }
 526
 527 /**
 528  * swapin_readahead - swap in pages in hope we need them soon
 529  * @entry: swap entry of this memory
 530  * @gfp_mask: memory allocation flags
 531  * @vma: user vma this address belongs to
 532  * @addr: target address for mempolicy
 533  *
 534  * Returns the struct page for entry and addr, after queueing swapin.
 535  *
 536  * Primitive swap readahead code. We simply read an aligned block of
 537  * (1 << page_cluster) entries in the swap area. This method is chosen
 538  * because it doesn't cost us any seek time.  We also make sure to queue
 539  * the 'original' request together with the readahead ones...
 540  *
 541  * This has been extended to use the NUMA policies from the mm triggering
 542  * the readahead.
 543  *
 544  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 545  */
 546 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 547                         struct vm_area_struct *vma, unsigned long addr)
 548 {
 549         struct page *page;
 550         unsigned long entry_offset = swp_offset(entry);
 551         unsigned long offset = entry_offset;
 552         unsigned long start_offset, end_offset;
 553         unsigned long mask;
 554         struct blk_plug plug;
 555         bool do_poll = true, page_allocated;
 556
 557         mask = swapin_nr_pages(offset) - 1;
 558         if (!mask)
 559                 goto skip;
 560
 561         do_poll = false;
 562         /* Read a page_cluster sized and aligned cluster around offset. */
 563         start_offset = offset & ~mask;
 564         end_offset = offset | mask;
 565         if (!start_offset)      /* First page is swap header. */
 566                 start_offset++;
 567
 568         blk_start_plug(&plug);
 569         for (offset = start_offset; offset <= end_offset ; offset++) {
 570                 /* Ok, do the async read-ahead now */
 571                 page = __read_swap_cache_async(
 572                         swp_entry(swp_type(entry), offset),
 573                         gfp_mask, vma, addr, &page_allocated);
 574                 if (!page)
 575                         continue;
 576                 if (page_allocated) {
 577                         swap_readpage(page, false);
 578                         if (offset != entry_offset &&
 579                             likely(!PageTransCompound(page))) {
 580                                 SetPageReadahead(page);
 581                                 count_vm_event(SWAP_RA);
 582                         }
 583                 }
 584                 put_page(page);
 585         }
 586         blk_finish_plug(&plug);
 587
 588         lru_add_drain();        /* Push any new pages onto the LRU now */
 589 skip:
 590         return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
 591 }
 592
 593 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 594 {
 595         struct address_space *spaces, *space;
 596         unsigned int i, nr;
 597
 598         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 599         spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
 600         if (!spaces)
 601                 return -ENOMEM;
 602         for (i = 0; i < nr; i++) {
 603                 space = spaces + i;
 604                 INIT_RADIX_TREE(&space->page_tree, GFP_ATOMIC|__GFP_NOWARN);
 605                 atomic_set(&space->i_mmap_writable, 0);
 606                 space->a_ops = &swap_aops;
 607                 /* swap cache doesn't use writeback related tags */
 608                 mapping_set_no_writeback_tags(space);
 609                 spin_lock_init(&space->tree_lock);
 610         }
 611         nr_swapper_spaces[type] = nr;
 612         rcu_assign_pointer(swapper_spaces[type], spaces);
 613
 614         return 0;
 615 }
 616
 617 void exit_swap_address_space(unsigned int type)
 618 {
 619         struct address_space *spaces;
 620
 621         spaces = swapper_spaces[type];
 622         nr_swapper_spaces[type] = 0;
 623         rcu_assign_pointer(swapper_spaces[type], NULL);
 624         synchronize_rcu();
 625         kvfree(spaces);
 626 }
 627
 628 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
 629                                      unsigned long faddr,
 630                                      unsigned long lpfn,
 631                                      unsigned long rpfn,
 632                                      unsigned long *start,
 633                                      unsigned long *end)
 634 {
 635         *start = max3(lpfn, PFN_DOWN(vma->vm_start),
 636                       PFN_DOWN(faddr & PMD_MASK));
 637         *end = min3(rpfn, PFN_DOWN(vma->vm_end),
 638                     PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
 639 }
 640
 641 struct page *swap_readahead_detect(struct vm_fault *vmf,
 642                                    struct vma_swap_readahead *swap_ra)
 643 {
 644         struct vm_area_struct *vma = vmf->vma;
 645         unsigned long swap_ra_info;
 646         struct page *page;
 647         swp_entry_t entry;
 648         unsigned long faddr, pfn, fpfn;
 649         unsigned long start, end;
 650         pte_t *pte;
 651         unsigned int max_win, hits, prev_win, win, left;
 652 #ifndef CONFIG_64BIT
 653         pte_t *tpte;
 654 #endif
 655
 656         faddr = vmf->address;
 657         entry = pte_to_swp_entry(vmf->orig_pte);
 658         if ((unlikely(non_swap_entry(entry))))
 659                 return NULL;
 660         page = lookup_swap_cache(entry, vma, faddr);
 661         if (page)
 662                 return page;
 663
 664         max_win = 1 << READ_ONCE(swap_ra_max_order);
 665         if (max_win == 1) {
 666                 swap_ra->win = 1;
 667                 return NULL;
 668         }
 669
 670         fpfn = PFN_DOWN(faddr);
 671         swap_ra_info = GET_SWAP_RA_VAL(vma);
 672         pfn = PFN_DOWN(SWAP_RA_ADDR(swap_ra_info));
 673         prev_win = SWAP_RA_WIN(swap_ra_info);
 674         hits = SWAP_RA_HITS(swap_ra_info);
 675         swap_ra->win = win = __swapin_nr_pages(pfn, fpfn, hits,
 676                                                max_win, prev_win);
 677         atomic_long_set(&vma->swap_readahead_info,
 678                         SWAP_RA_VAL(faddr, win, 0));
 679
 680         if (win == 1)
 681                 return NULL;
 682
 683         /* Copy the PTEs because the page table may be unmapped */
 684         if (fpfn == pfn + 1)
 685                 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
 686         else if (pfn == fpfn + 1)
 687                 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
 688                                   &start, &end);
 689         else {
 690                 left = (win - 1) / 2;
 691                 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
 692                                   &start, &end);
 693         }
 694         swap_ra->nr_pte = end - start;
 695         swap_ra->offset = fpfn - start;
 696         pte = vmf->pte - swap_ra->offset;
 697 #ifdef CONFIG_64BIT
 698         swap_ra->ptes = pte;
 699 #else
 700         tpte = swap_ra->ptes;
 701         for (pfn = start; pfn != end; pfn++)
 702                 *tpte++ = *pte++;
 703 #endif
 704
 705         return NULL;
 706 }
 707
 708 struct page *do_swap_page_readahead(swp_entry_t fentry, gfp_t gfp_mask,
 709                                     struct vm_fault *vmf,
 710                                     struct vma_swap_readahead *swap_ra)
 711 {
 712         struct blk_plug plug;
 713         struct vm_area_struct *vma = vmf->vma;
 714         struct page *page;
 715         pte_t *pte, pentry;
 716         swp_entry_t entry;
 717         unsigned int i;
 718         bool page_allocated;
 719
 720         if (swap_ra->win == 1)
 721                 goto skip;
 722
 723         blk_start_plug(&plug);
 724         for (i = 0, pte = swap_ra->ptes; i < swap_ra->nr_pte;
 725              i++, pte++) {
 726                 pentry = *pte;
 727                 if (pte_none(pentry))
 728                         continue;
 729                 if (pte_present(pentry))
 730                         continue;
 731                 entry = pte_to_swp_entry(pentry);
 732                 if (unlikely(non_swap_entry(entry)))
 733                         continue;
 734                 page = __read_swap_cache_async(entry, gfp_mask, vma,
 735                                                vmf->address, &page_allocated);
 736                 if (!page)
 737                         continue;
 738                 if (page_allocated) {
 739                         swap_readpage(page, false);
 740                         if (i != swap_ra->offset &&
 741                             likely(!PageTransCompound(page))) {
 742                                 SetPageReadahead(page);
 743                                 count_vm_event(SWAP_RA);
 744                         }
 745                 }
 746                 put_page(page);
 747         }
 748         blk_finish_plug(&plug);
 749         lru_add_drain();
 750 skip:
 751         return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
 752                                      swap_ra->win == 1);
 753 }
 754
 755 #ifdef CONFIG_SYSFS
 756 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
 757                                      struct kobj_attribute *attr, char *buf)
 758 {
 759         return sprintf(buf, "%s\n", swap_vma_readahead ? "true" : "false");
 760 }
 761 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
 762                                       struct kobj_attribute *attr,
 763                                       const char *buf, size_t count)
 764 {
 765         if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
 766                 swap_vma_readahead = true;
 767         else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
 768                 swap_vma_readahead = false;
 769         else
 770                 return -EINVAL;
 771
 772         return count;
 773 }
 774 static struct kobj_attribute vma_ra_enabled_attr =
 775         __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
 776                vma_ra_enabled_store);
 777
 778 static ssize_t vma_ra_max_order_show(struct kobject *kobj,
 779                                      struct kobj_attribute *attr, char *buf)
 780 {
 781         return sprintf(buf, "%d\n", swap_ra_max_order);
 782 }
 783 static ssize_t vma_ra_max_order_store(struct kobject *kobj,
 784                                       struct kobj_attribute *attr,
 785                                       const char *buf, size_t count)
 786 {
 787         int err, v;
 788
 789         err = kstrtoint(buf, 10, &v);
 790         if (err || v > SWAP_RA_ORDER_CEILING || v <= 0)
 791                 return -EINVAL;
 792
 793         swap_ra_max_order = v;
 794
 795         return count;
 796 }
 797 static struct kobj_attribute vma_ra_max_order_attr =
 798         __ATTR(vma_ra_max_order, 0644, vma_ra_max_order_show,
 799                vma_ra_max_order_store);
 800
 801 static struct attribute *swap_attrs[] = {
 802         &vma_ra_enabled_attr.attr,
 803         &vma_ra_max_order_attr.attr,
 804         NULL,
 805 };
 806
 807 static struct attribute_group swap_attr_group = {
 808         .attrs = swap_attrs,
 809 };
 810
 811 static int __init swap_init_sysfs(void)
 812 {
 813         int err;
 814         struct kobject *swap_kobj;
 815
 816         swap_kobj = kobject_create_and_add("swap", mm_kobj);
 817         if (!swap_kobj) {
 818                 pr_err("failed to create swap kobject\n");
 819                 return -ENOMEM;
 820         }
 821         err = sysfs_create_group(swap_kobj, &swap_attr_group);
 822         if (err) {
 823                 pr_err("failed to register swap group\n");
 824                 goto delete_obj;
 825         }
 826         return 0;
 827
 828 delete_obj:
 829         kobject_put(swap_kobj);
 830         return err;
 831 }
 832 subsys_initcall(swap_init_sysfs);
 833 #endif