mm/swap_state.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/mm/swap_state.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  *  Swap reorganised 29.12.95, Stephen Tweedie
   7  *
   8  *  Rewritten to use page cache, (C) 1998 Stephen Tweedie
   9  */
  10 #include <linux/mm.h>
  11 #include <linux/gfp.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/init.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/backing-dev.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/pagevec.h>
  20 #include <linux/migrate.h>
  21 #include <linux/vmalloc.h>
  22 #include <linux/swap_slots.h>
  23 #include <linux/huge_mm.h>
  24
  25 #include <asm/pgtable.h>
  26
  27 /*
  28  * swapper_space is a fiction, retained to simplify the path through
  29  * vmscan's shrink_page_list.
  30  */
  31 static const struct address_space_operations swap_aops = {
  32         .writepage      = swap_writepage,
  33         .set_page_dirty = swap_set_page_dirty,
  34 #ifdef CONFIG_MIGRATION
  35         .migratepage    = migrate_page,
  36 #endif
  37 };
  38
  39 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
  40 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
  41 bool swap_vma_readahead __read_mostly = true;
  42
  43 #define SWAP_RA_WIN_SHIFT       (PAGE_SHIFT / 2)
  44 #define SWAP_RA_HITS_MASK       ((1UL << SWAP_RA_WIN_SHIFT) - 1)
  45 #define SWAP_RA_HITS_MAX        SWAP_RA_HITS_MASK
  46 #define SWAP_RA_WIN_MASK        (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
  47
  48 #define SWAP_RA_HITS(v)         ((v) & SWAP_RA_HITS_MASK)
  49 #define SWAP_RA_WIN(v)          (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
  50 #define SWAP_RA_ADDR(v)         ((v) & PAGE_MASK)
  51
  52 #define SWAP_RA_VAL(addr, win, hits)                            \
  53         (((addr) & PAGE_MASK) |                                 \
  54          (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |    \
  55          ((hits) & SWAP_RA_HITS_MASK))
  56
  57 /* Initial readahead hits is 4 to start up with a small window */
  58 #define GET_SWAP_RA_VAL(vma)                                    \
  59         (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
  60
  61 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  62 #define ADD_CACHE_INFO(x, nr)   do { swap_cache_info.x += (nr); } while (0)
  63
  64 static struct {
  65         unsigned long add_total;
  66         unsigned long del_total;
  67         unsigned long find_success;
  68         unsigned long find_total;
  69 } swap_cache_info;
  70
  71 unsigned long total_swapcache_pages(void)
  72 {
  73         unsigned int i, j, nr;
  74         unsigned long ret = 0;
  75         struct address_space *spaces;
  76
  77         rcu_read_lock();
  78         for (i = 0; i < MAX_SWAPFILES; i++) {
  79                 /*
  80                  * The corresponding entries in nr_swapper_spaces and
  81                  * swapper_spaces will be reused only after at least
  82                  * one grace period.  So it is impossible for them
  83                  * belongs to different usage.
  84                  */
  85                 nr = nr_swapper_spaces[i];
  86                 spaces = rcu_dereference(swapper_spaces[i]);
  87                 if (!nr || !spaces)
  88                         continue;
  89                 for (j = 0; j < nr; j++)
  90                         ret += spaces[j].nrpages;
  91         }
  92         rcu_read_unlock();
  93         return ret;
  94 }
  95
  96 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  97
  98 void show_swap_cache_info(void)
  99 {
 100         printk("%lu pages in swap cache\n", total_swapcache_pages());
 101         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
 102                 swap_cache_info.add_total, swap_cache_info.del_total,
 103                 swap_cache_info.find_success, swap_cache_info.find_total);
 104         printk("Free swap  = %ldkB\n",
 105                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
 106         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
 107 }
 108
 109 /*
 110  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
 111  * but sets SwapCache flag and private instead of mapping and index.
 112  */
 113 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
 114 {
 115         int error, i, nr = hpage_nr_pages(page);
 116         struct address_space *address_space;
 117         pgoff_t idx = swp_offset(entry);
 118
 119         VM_BUG_ON_PAGE(!PageLocked(page), page);
 120         VM_BUG_ON_PAGE(PageSwapCache(page), page);
 121         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 122
 123         page_ref_add(page, nr);
 124         SetPageSwapCache(page);
 125
 126         address_space = swap_address_space(entry);
 127         spin_lock_irq(&address_space->tree_lock);
 128         for (i = 0; i < nr; i++) {
 129                 set_page_private(page + i, entry.val + i);
 130                 error = radix_tree_insert(&address_space->page_tree,
 131                                           idx + i, page + i);
 132                 if (unlikely(error))
 133                         break;
 134         }
 135         if (likely(!error)) {
 136                 address_space->nrpages += nr;
 137                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
 138                 ADD_CACHE_INFO(add_total, nr);
 139         } else {
 140                 /*
 141                  * Only the context which have set SWAP_HAS_CACHE flag
 142                  * would call add_to_swap_cache().
 143                  * So add_to_swap_cache() doesn't returns -EEXIST.
 144                  */
 145                 VM_BUG_ON(error == -EEXIST);
 146                 set_page_private(page + i, 0UL);
 147                 while (i--) {
 148                         radix_tree_delete(&address_space->page_tree, idx + i);
 149                         set_page_private(page + i, 0UL);
 150                 }
 151                 ClearPageSwapCache(page);
 152                 page_ref_sub(page, nr);
 153         }
 154         spin_unlock_irq(&address_space->tree_lock);
 155
 156         return error;
 157 }
 158
 159
 160 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 161 {
 162         int error;
 163
 164         error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
 165         if (!error) {
 166                 error = __add_to_swap_cache(page, entry);
 167                 radix_tree_preload_end();
 168         }
 169         return error;
 170 }
 171
 172 /*
 173  * This must be called only on pages that have
 174  * been verified to be in the swap cache.
 175  */
 176 void __delete_from_swap_cache(struct page *page)
 177 {
 178         struct address_space *address_space;
 179         int i, nr = hpage_nr_pages(page);
 180         swp_entry_t entry;
 181         pgoff_t idx;
 182
 183         VM_BUG_ON_PAGE(!PageLocked(page), page);
 184         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 185         VM_BUG_ON_PAGE(PageWriteback(page), page);
 186
 187         entry.val = page_private(page);
 188         address_space = swap_address_space(entry);
 189         idx = swp_offset(entry);
 190         for (i = 0; i < nr; i++) {
 191                 radix_tree_delete(&address_space->page_tree, idx + i);
 192                 set_page_private(page + i, 0);
 193         }
 194         ClearPageSwapCache(page);
 195         address_space->nrpages -= nr;
 196         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
 197         ADD_CACHE_INFO(del_total, nr);
 198 }
 199
 200 /**
 201  * add_to_swap - allocate swap space for a page
 202  * @page: page we want to move to swap
 203  *
 204  * Allocate swap space for the page and add the page to the
 205  * swap cache.  Caller needs to hold the page lock.
 206  */
 207 int add_to_swap(struct page *page)
 208 {
 209         swp_entry_t entry;
 210         int err;
 211
 212         VM_BUG_ON_PAGE(!PageLocked(page), page);
 213         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 214
 215         entry = get_swap_page(page);
 216         if (!entry.val)
 217                 return 0;
 218
 219         if (mem_cgroup_try_charge_swap(page, entry))
 220                 goto fail;
 221
 222         /*
 223          * Radix-tree node allocations from PF_MEMALLOC contexts could
 224          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 225          * stops emergency reserves from being allocated.
 226          *
 227          * TODO: this could cause a theoretical memory reclaim
 228          * deadlock in the swap out path.
 229          */
 230         /*
 231          * Add it to the swap cache.
 232          */
 233         err = add_to_swap_cache(page, entry,
 234                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 235         /* -ENOMEM radix-tree allocation failure */
 236         if (err)
 237                 /*
 238                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 239                  * clear SWAP_HAS_CACHE flag.
 240                  */
 241                 goto fail;
 242         /*
 243          * Normally the page will be dirtied in unmap because its pte should be
 244          * dirty. A special case is MADV_FREE page. The page'e pte could have
 245          * dirty bit cleared but the page's SwapBacked bit is still set because
 246          * clearing the dirty bit and SwapBacked bit has no lock protected. For
 247          * such page, unmap will not set dirty bit for it, so page reclaim will
 248          * not write the page out. This can cause data corruption when the page
 249          * is swap in later. Always setting the dirty bit for the page solves
 250          * the problem.
 251          */
 252         set_page_dirty(page);
 253
 254         return 1;
 255
 256 fail:
 257         put_swap_page(page, entry);
 258         return 0;
 259 }
 260
 261 /*
 262  * This must be called only on pages that have
 263  * been verified to be in the swap cache and locked.
 264  * It will never put the page into the free list,
 265  * the caller has a reference on the page.
 266  */
 267 void delete_from_swap_cache(struct page *page)
 268 {
 269         swp_entry_t entry;
 270         struct address_space *address_space;
 271
 272         entry.val = page_private(page);
 273
 274         address_space = swap_address_space(entry);
 275         spin_lock_irq(&address_space->tree_lock);
 276         __delete_from_swap_cache(page);
 277         spin_unlock_irq(&address_space->tree_lock);
 278
 279         put_swap_page(page, entry);
 280         page_ref_sub(page, hpage_nr_pages(page));
 281 }
 282
 283 /*
 284  * If we are the only user, then try to free up the swap cache.
 285  *
 286  * Its ok to check for PageSwapCache without the page lock
 287  * here because we are going to recheck again inside
 288  * try_to_free_swap() _with_ the lock.
 289  *                                      - Marcelo
 290  */
 291 static inline void free_swap_cache(struct page *page)
 292 {
 293         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 294                 try_to_free_swap(page);
 295                 unlock_page(page);
 296         }
 297 }
 298
 299 /*
 300  * Perform a free_page(), also freeing any swap cache associated with
 301  * this page if it is the last user of the page.
 302  */
 303 void free_page_and_swap_cache(struct page *page)
 304 {
 305         free_swap_cache(page);
 306         if (!is_huge_zero_page(page))
 307                 put_page(page);
 308 }
 309
 310 /*
 311  * Passed an array of pages, drop them all from swapcache and then release
 312  * them.  They are removed from the LRU and freed if this is their last use.
 313  */
 314 void free_pages_and_swap_cache(struct page **pages, int nr)
 315 {
 316         struct page **pagep = pages;
 317         int i;
 318
 319         lru_add_drain();
 320         for (i = 0; i < nr; i++)
 321                 free_swap_cache(pagep[i]);
 322         release_pages(pagep, nr);
 323 }
 324
 325 /*
 326  * Lookup a swap entry in the swap cache. A found page will be returned
 327  * unlocked and with its refcount incremented - we rely on the kernel
 328  * lock getting page table operations atomic even if we drop the page
 329  * lock before returning.
 330  */
 331 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
 332                                unsigned long addr)
 333 {
 334         struct page *page;
 335         unsigned long ra_info;
 336         int win, hits, readahead;
 337
 338         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 339
 340         INC_CACHE_INFO(find_total);
 341         if (page) {
 342                 INC_CACHE_INFO(find_success);
 343                 if (unlikely(PageTransCompound(page)))
 344                         return page;
 345                 readahead = TestClearPageReadahead(page);
 346                 if (vma) {
 347                         ra_info = GET_SWAP_RA_VAL(vma);
 348                         win = SWAP_RA_WIN(ra_info);
 349                         hits = SWAP_RA_HITS(ra_info);
 350                         if (readahead)
 351                                 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
 352                         atomic_long_set(&vma->swap_readahead_info,
 353                                         SWAP_RA_VAL(addr, win, hits));
 354                 }
 355                 if (readahead) {
 356                         count_vm_event(SWAP_RA_HIT);
 357                         if (!vma)
 358                                 atomic_inc(&swapin_readahead_hits);
 359                 }
 360         }
 361         return page;
 362 }
 363
 364 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 365                         struct vm_area_struct *vma, unsigned long addr,
 366                         bool *new_page_allocated)
 367 {
 368         struct page *found_page, *new_page = NULL;
 369         struct address_space *swapper_space = swap_address_space(entry);
 370         int err;
 371         *new_page_allocated = false;
 372
 373         do {
 374                 /*
 375                  * First check the swap cache.  Since this is normally
 376                  * called after lookup_swap_cache() failed, re-calling
 377                  * that would confuse statistics.
 378                  */
 379                 found_page = find_get_page(swapper_space, swp_offset(entry));
 380                 if (found_page)
 381                         break;
 382
 383                 /*
 384                  * Just skip read ahead for unused swap slot.
 385                  * During swap_off when swap_slot_cache is disabled,
 386                  * we have to handle the race between putting
 387                  * swap entry in swap cache and marking swap slot
 388                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 389                  * else swap_off will be aborted if we return NULL.
 390                  */
 391                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 392                         break;
 393
 394                 /*
 395                  * Get a new page to read into from swap.
 396                  */
 397                 if (!new_page) {
 398                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 399                         if (!new_page)
 400                                 break;          /* Out of memory */
 401                 }
 402
 403                 /*
 404                  * call radix_tree_preload() while we can wait.
 405                  */
 406                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 407                 if (err)
 408                         break;
 409
 410                 /*
 411                  * Swap entry may have been freed since our caller observed it.
 412                  */
 413                 err = swapcache_prepare(entry);
 414                 if (err == -EEXIST) {
 415                         radix_tree_preload_end();
 416                         /*
 417                          * We might race against get_swap_page() and stumble
 418                          * across a SWAP_HAS_CACHE swap_map entry whose page
 419                          * has not been brought into the swapcache yet.
 420                          */
 421                         cond_resched();
 422                         continue;
 423                 }
 424                 if (err) {              /* swp entry is obsolete ? */
 425                         radix_tree_preload_end();
 426                         break;
 427                 }
 428
 429                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 430                 __SetPageLocked(new_page);
 431                 __SetPageSwapBacked(new_page);
 432                 err = __add_to_swap_cache(new_page, entry);
 433                 if (likely(!err)) {
 434                         radix_tree_preload_end();
 435                         /*
 436                          * Initiate read into locked page and return.
 437                          */
 438                         lru_cache_add_anon(new_page);
 439                         *new_page_allocated = true;
 440                         return new_page;
 441                 }
 442                 radix_tree_preload_end();
 443                 __ClearPageLocked(new_page);
 444                 /*
 445                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 446                  * clear SWAP_HAS_CACHE flag.
 447                  */
 448                 put_swap_page(new_page, entry);
 449         } while (err != -ENOMEM);
 450
 451         if (new_page)
 452                 put_page(new_page);
 453         return found_page;
 454 }
 455
 456 /*
 457  * Locate a page of swap in physical memory, reserving swap cache space
 458  * and reading the disk if it is not already cached.
 459  * A failure return means that either the page allocation failed or that
 460  * the swap entry is no longer in use.
 461  */
 462 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 463                 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
 464 {
 465         bool page_was_allocated;
 466         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 467                         vma, addr, &page_was_allocated);
 468
 469         if (page_was_allocated)
 470                 swap_readpage(retpage, do_poll);
 471
 472         return retpage;
 473 }
 474
 475 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
 476                                       unsigned long offset,
 477                                       int hits,
 478                                       int max_pages,
 479                                       int prev_win)
 480 {
 481         unsigned int pages, last_ra;
 482
 483         /*
 484          * This heuristic has been found to work well on both sequential and
 485          * random loads, swapping to hard disk or to SSD: please don't ask
 486          * what the "+ 2" means, it just happens to work well, that's all.
 487          */
 488         pages = hits + 2;
 489         if (pages == 2) {
 490                 /*
 491                  * We can have no readahead hits to judge by: but must not get
 492                  * stuck here forever, so check for an adjacent offset instead
 493                  * (and don't even bother to check whether swap type is same).
 494                  */
 495                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 496                         pages = 1;
 497         } else {
 498                 unsigned int roundup = 4;
 499                 while (roundup < pages)
 500                         roundup <<= 1;
 501                 pages = roundup;
 502         }
 503
 504         if (pages > max_pages)
 505                 pages = max_pages;
 506
 507         /* Don't shrink readahead too fast */
 508         last_ra = prev_win / 2;
 509         if (pages < last_ra)
 510                 pages = last_ra;
 511
 512         return pages;
 513 }
 514
 515 static unsigned long swapin_nr_pages(unsigned long offset)
 516 {
 517         static unsigned long prev_offset;
 518         unsigned int hits, pages, max_pages;
 519         static atomic_t last_readahead_pages;
 520
 521         max_pages = 1 << READ_ONCE(page_cluster);
 522         if (max_pages <= 1)
 523                 return 1;
 524
 525         hits = atomic_xchg(&swapin_readahead_hits, 0);
 526         pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
 527                                   atomic_read(&last_readahead_pages));
 528         if (!hits)
 529                 prev_offset = offset;
 530         atomic_set(&last_readahead_pages, pages);
 531
 532         return pages;
 533 }
 534
 535 /**
 536  * swapin_readahead - swap in pages in hope we need them soon
 537  * @entry: swap entry of this memory
 538  * @gfp_mask: memory allocation flags
 539  * @vma: user vma this address belongs to
 540  * @addr: target address for mempolicy
 541  *
 542  * Returns the struct page for entry and addr, after queueing swapin.
 543  *
 544  * Primitive swap readahead code. We simply read an aligned block of
 545  * (1 << page_cluster) entries in the swap area. This method is chosen
 546  * because it doesn't cost us any seek time.  We also make sure to queue
 547  * the 'original' request together with the readahead ones...
 548  *
 549  * This has been extended to use the NUMA policies from the mm triggering
 550  * the readahead.
 551  *
 552  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 553  */
 554 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 555                         struct vm_area_struct *vma, unsigned long addr)
 556 {
 557         struct page *page;
 558         unsigned long entry_offset = swp_offset(entry);
 559         unsigned long offset = entry_offset;
 560         unsigned long start_offset, end_offset;
 561         unsigned long mask;
 562         struct swap_info_struct *si = swp_swap_info(entry);
 563         struct blk_plug plug;
 564         bool do_poll = true, page_allocated;
 565
 566         mask = swapin_nr_pages(offset) - 1;
 567         if (!mask)
 568                 goto skip;
 569
 570         do_poll = false;
 571         /* Read a page_cluster sized and aligned cluster around offset. */
 572         start_offset = offset & ~mask;
 573         end_offset = offset | mask;
 574         if (!start_offset)      /* First page is swap header. */
 575                 start_offset++;
 576         if (end_offset >= si->max)
 577                 end_offset = si->max - 1;
 578
 579         blk_start_plug(&plug);
 580         for (offset = start_offset; offset <= end_offset ; offset++) {
 581                 /* Ok, do the async read-ahead now */
 582                 page = __read_swap_cache_async(
 583                         swp_entry(swp_type(entry), offset),
 584                         gfp_mask, vma, addr, &page_allocated);
 585                 if (!page)
 586                         continue;
 587                 if (page_allocated) {
 588                         swap_readpage(page, false);
 589                         if (offset != entry_offset &&
 590                             likely(!PageTransCompound(page))) {
 591                                 SetPageReadahead(page);
 592                                 count_vm_event(SWAP_RA);
 593                         }
 594                 }
 595                 put_page(page);
 596         }
 597         blk_finish_plug(&plug);
 598
 599         lru_add_drain();        /* Push any new pages onto the LRU now */
 600 skip:
 601         return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
 602 }
 603
 604 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 605 {
 606         struct address_space *spaces, *space;
 607         unsigned int i, nr;
 608
 609         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 610         spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
 611         if (!spaces)
 612                 return -ENOMEM;
 613         for (i = 0; i < nr; i++) {
 614                 space = spaces + i;
 615                 INIT_RADIX_TREE(&space->page_tree, GFP_ATOMIC|__GFP_NOWARN);
 616                 atomic_set(&space->i_mmap_writable, 0);
 617                 space->a_ops = &swap_aops;
 618                 /* swap cache doesn't use writeback related tags */
 619                 mapping_set_no_writeback_tags(space);
 620                 spin_lock_init(&space->tree_lock);
 621         }
 622         nr_swapper_spaces[type] = nr;
 623         rcu_assign_pointer(swapper_spaces[type], spaces);
 624
 625         return 0;
 626 }
 627
 628 void exit_swap_address_space(unsigned int type)
 629 {
 630         struct address_space *spaces;
 631
 632         spaces = swapper_spaces[type];
 633         nr_swapper_spaces[type] = 0;
 634         rcu_assign_pointer(swapper_spaces[type], NULL);
 635         synchronize_rcu();
 636         kvfree(spaces);
 637 }
 638
 639 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
 640                                      unsigned long faddr,
 641                                      unsigned long lpfn,
 642                                      unsigned long rpfn,
 643                                      unsigned long *start,
 644                                      unsigned long *end)
 645 {
 646         *start = max3(lpfn, PFN_DOWN(vma->vm_start),
 647                       PFN_DOWN(faddr & PMD_MASK));
 648         *end = min3(rpfn, PFN_DOWN(vma->vm_end),
 649                     PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
 650 }
 651
 652 struct page *swap_readahead_detect(struct vm_fault *vmf,
 653                                    struct vma_swap_readahead *swap_ra)
 654 {
 655         struct vm_area_struct *vma = vmf->vma;
 656         unsigned long swap_ra_info;
 657         struct page *page;
 658         swp_entry_t entry;
 659         unsigned long faddr, pfn, fpfn;
 660         unsigned long start, end;
 661         pte_t *pte;
 662         unsigned int max_win, hits, prev_win, win, left;
 663 #ifndef CONFIG_64BIT
 664         pte_t *tpte;
 665 #endif
 666
 667         max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
 668                              SWAP_RA_ORDER_CEILING);
 669         if (max_win == 1) {
 670                 swap_ra->win = 1;
 671                 return NULL;
 672         }
 673
 674         faddr = vmf->address;
 675         entry = pte_to_swp_entry(vmf->orig_pte);
 676         if ((unlikely(non_swap_entry(entry))))
 677                 return NULL;
 678         page = lookup_swap_cache(entry, vma, faddr);
 679         if (page)
 680                 return page;
 681
 682         fpfn = PFN_DOWN(faddr);
 683         swap_ra_info = GET_SWAP_RA_VAL(vma);
 684         pfn = PFN_DOWN(SWAP_RA_ADDR(swap_ra_info));
 685         prev_win = SWAP_RA_WIN(swap_ra_info);
 686         hits = SWAP_RA_HITS(swap_ra_info);
 687         swap_ra->win = win = __swapin_nr_pages(pfn, fpfn, hits,
 688                                                max_win, prev_win);
 689         atomic_long_set(&vma->swap_readahead_info,
 690                         SWAP_RA_VAL(faddr, win, 0));
 691
 692         if (win == 1)
 693                 return NULL;
 694
 695         /* Copy the PTEs because the page table may be unmapped */
 696         if (fpfn == pfn + 1)
 697                 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
 698         else if (pfn == fpfn + 1)
 699                 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
 700                                   &start, &end);
 701         else {
 702                 left = (win - 1) / 2;
 703                 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
 704                                   &start, &end);
 705         }
 706         swap_ra->nr_pte = end - start;
 707         swap_ra->offset = fpfn - start;
 708         pte = vmf->pte - swap_ra->offset;
 709 #ifdef CONFIG_64BIT
 710         swap_ra->ptes = pte;
 711 #else
 712         tpte = swap_ra->ptes;
 713         for (pfn = start; pfn != end; pfn++)
 714                 *tpte++ = *pte++;
 715 #endif
 716
 717         return NULL;
 718 }
 719
 720 struct page *do_swap_page_readahead(swp_entry_t fentry, gfp_t gfp_mask,
 721                                     struct vm_fault *vmf,
 722                                     struct vma_swap_readahead *swap_ra)
 723 {
 724         struct blk_plug plug;
 725         struct vm_area_struct *vma = vmf->vma;
 726         struct page *page;
 727         pte_t *pte, pentry;
 728         swp_entry_t entry;
 729         unsigned int i;
 730         bool page_allocated;
 731
 732         if (swap_ra->win == 1)
 733                 goto skip;
 734
 735         blk_start_plug(&plug);
 736         for (i = 0, pte = swap_ra->ptes; i < swap_ra->nr_pte;
 737              i++, pte++) {
 738                 pentry = *pte;
 739                 if (pte_none(pentry))
 740                         continue;
 741                 if (pte_present(pentry))
 742                         continue;
 743                 entry = pte_to_swp_entry(pentry);
 744                 if (unlikely(non_swap_entry(entry)))
 745                         continue;
 746                 page = __read_swap_cache_async(entry, gfp_mask, vma,
 747                                                vmf->address, &page_allocated);
 748                 if (!page)
 749                         continue;
 750                 if (page_allocated) {
 751                         swap_readpage(page, false);
 752                         if (i != swap_ra->offset &&
 753                             likely(!PageTransCompound(page))) {
 754                                 SetPageReadahead(page);
 755                                 count_vm_event(SWAP_RA);
 756                         }
 757                 }
 758                 put_page(page);
 759         }
 760         blk_finish_plug(&plug);
 761         lru_add_drain();
 762 skip:
 763         return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
 764                                      swap_ra->win == 1);
 765 }
 766
 767 #ifdef CONFIG_SYSFS
 768 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
 769                                      struct kobj_attribute *attr, char *buf)
 770 {
 771         return sprintf(buf, "%s\n", swap_vma_readahead ? "true" : "false");
 772 }
 773 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
 774                                       struct kobj_attribute *attr,
 775                                       const char *buf, size_t count)
 776 {
 777         if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
 778                 swap_vma_readahead = true;
 779         else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
 780                 swap_vma_readahead = false;
 781         else
 782                 return -EINVAL;
 783
 784         return count;
 785 }
 786 static struct kobj_attribute vma_ra_enabled_attr =
 787         __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
 788                vma_ra_enabled_store);
 789
 790 static struct attribute *swap_attrs[] = {
 791         &vma_ra_enabled_attr.attr,
 792         NULL,
 793 };
 794
 795 static struct attribute_group swap_attr_group = {
 796         .attrs = swap_attrs,
 797 };
 798
 799 static int __init swap_init_sysfs(void)
 800 {
 801         int err;
 802         struct kobject *swap_kobj;
 803
 804         swap_kobj = kobject_create_and_add("swap", mm_kobj);
 805         if (!swap_kobj) {
 806                 pr_err("failed to create swap kobject\n");
 807                 return -ENOMEM;
 808         }
 809         err = sysfs_create_group(swap_kobj, &swap_attr_group);
 810         if (err) {
 811                 pr_err("failed to register swap group\n");
 812                 goto delete_obj;
 813         }
 814         return 0;
 815
 816 delete_obj:
 817         kobject_put(swap_kobj);
 818         return err;
 819 }
 820 subsys_initcall(swap_init_sysfs);
 821 #endif