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 static bool enable_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         xa_lock_irq(&address_space->i_pages);
 128         for (i = 0; i < nr; i++) {
 129                 set_page_private(page + i, entry.val + i);
 130                 error = radix_tree_insert(&address_space->i_pages,
 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->i_pages, idx + i);
 149                         set_page_private(page + i, 0UL);
 150                 }
 151                 ClearPageSwapCache(page);
 152                 page_ref_sub(page, nr);
 153         }
 154         xa_unlock_irq(&address_space->i_pages);
 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->i_pages, 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         /*
 220          * Radix-tree node allocations from PF_MEMALLOC contexts could
 221          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 222          * stops emergency reserves from being allocated.
 223          *
 224          * TODO: this could cause a theoretical memory reclaim
 225          * deadlock in the swap out path.
 226          */
 227         /*
 228          * Add it to the swap cache.
 229          */
 230         err = add_to_swap_cache(page, entry,
 231                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 232         /* -ENOMEM radix-tree allocation failure */
 233         if (err)
 234                 /*
 235                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 236                  * clear SWAP_HAS_CACHE flag.
 237                  */
 238                 goto fail;
 239         /*
 240          * Normally the page will be dirtied in unmap because its pte should be
 241          * dirty. A special case is MADV_FREE page. The page'e pte could have
 242          * dirty bit cleared but the page's SwapBacked bit is still set because
 243          * clearing the dirty bit and SwapBacked bit has no lock protected. For
 244          * such page, unmap will not set dirty bit for it, so page reclaim will
 245          * not write the page out. This can cause data corruption when the page
 246          * is swap in later. Always setting the dirty bit for the page solves
 247          * the problem.
 248          */
 249         set_page_dirty(page);
 250
 251         return 1;
 252
 253 fail:
 254         put_swap_page(page, entry);
 255         return 0;
 256 }
 257
 258 /*
 259  * This must be called only on pages that have
 260  * been verified to be in the swap cache and locked.
 261  * It will never put the page into the free list,
 262  * the caller has a reference on the page.
 263  */
 264 void delete_from_swap_cache(struct page *page)
 265 {
 266         swp_entry_t entry;
 267         struct address_space *address_space;
 268
 269         entry.val = page_private(page);
 270
 271         address_space = swap_address_space(entry);
 272         xa_lock_irq(&address_space->i_pages);
 273         __delete_from_swap_cache(page);
 274         xa_unlock_irq(&address_space->i_pages);
 275
 276         put_swap_page(page, entry);
 277         page_ref_sub(page, hpage_nr_pages(page));
 278 }
 279
 280 /*
 281  * If we are the only user, then try to free up the swap cache.
 282  *
 283  * Its ok to check for PageSwapCache without the page lock
 284  * here because we are going to recheck again inside
 285  * try_to_free_swap() _with_ the lock.
 286  *                                      - Marcelo
 287  */
 288 static inline void free_swap_cache(struct page *page)
 289 {
 290         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 291                 try_to_free_swap(page);
 292                 unlock_page(page);
 293         }
 294 }
 295
 296 /*
 297  * Perform a free_page(), also freeing any swap cache associated with
 298  * this page if it is the last user of the page.
 299  */
 300 void free_page_and_swap_cache(struct page *page)
 301 {
 302         free_swap_cache(page);
 303         if (!is_huge_zero_page(page))
 304                 put_page(page);
 305 }
 306
 307 /*
 308  * Passed an array of pages, drop them all from swapcache and then release
 309  * them.  They are removed from the LRU and freed if this is their last use.
 310  */
 311 void free_pages_and_swap_cache(struct page **pages, int nr)
 312 {
 313         struct page **pagep = pages;
 314         int i;
 315
 316         lru_add_drain();
 317         for (i = 0; i < nr; i++)
 318                 free_swap_cache(pagep[i]);
 319         release_pages(pagep, nr);
 320 }
 321
 322 static inline bool swap_use_vma_readahead(void)
 323 {
 324         return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
 325 }
 326
 327 /*
 328  * Lookup a swap entry in the swap cache. A found page will be returned
 329  * unlocked and with its refcount incremented - we rely on the kernel
 330  * lock getting page table operations atomic even if we drop the page
 331  * lock before returning.
 332  */
 333 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
 334                                unsigned long addr)
 335 {
 336         struct page *page;
 337
 338         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 339
 340         INC_CACHE_INFO(find_total);
 341         if (page) {
 342                 bool vma_ra = swap_use_vma_readahead();
 343                 bool readahead;
 344
 345                 INC_CACHE_INFO(find_success);
 346                 /*
 347                  * At the moment, we don't support PG_readahead for anon THP
 348                  * so let's bail out rather than confusing the readahead stat.
 349                  */
 350                 if (unlikely(PageTransCompound(page)))
 351                         return page;
 352
 353                 readahead = TestClearPageReadahead(page);
 354                 if (vma && vma_ra) {
 355                         unsigned long ra_val;
 356                         int win, hits;
 357
 358                         ra_val = GET_SWAP_RA_VAL(vma);
 359                         win = SWAP_RA_WIN(ra_val);
 360                         hits = SWAP_RA_HITS(ra_val);
 361                         if (readahead)
 362                                 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
 363                         atomic_long_set(&vma->swap_readahead_info,
 364                                         SWAP_RA_VAL(addr, win, hits));
 365                 }
 366
 367                 if (readahead) {
 368                         count_vm_event(SWAP_RA_HIT);
 369                         if (!vma || !vma_ra)
 370                                 atomic_inc(&swapin_readahead_hits);
 371                 }
 372         }
 373
 374         return page;
 375 }
 376
 377 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 378                         struct vm_area_struct *vma, unsigned long addr,
 379                         bool *new_page_allocated)
 380 {
 381         struct page *found_page, *new_page = NULL;
 382         struct address_space *swapper_space = swap_address_space(entry);
 383         int err;
 384         *new_page_allocated = false;
 385
 386         do {
 387                 /*
 388                  * First check the swap cache.  Since this is normally
 389                  * called after lookup_swap_cache() failed, re-calling
 390                  * that would confuse statistics.
 391                  */
 392                 found_page = find_get_page(swapper_space, swp_offset(entry));
 393                 if (found_page)
 394                         break;
 395
 396                 /*
 397                  * Just skip read ahead for unused swap slot.
 398                  * During swap_off when swap_slot_cache is disabled,
 399                  * we have to handle the race between putting
 400                  * swap entry in swap cache and marking swap slot
 401                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 402                  * else swap_off will be aborted if we return NULL.
 403                  */
 404                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 405                         break;
 406
 407                 /*
 408                  * Get a new page to read into from swap.
 409                  */
 410                 if (!new_page) {
 411                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 412                         if (!new_page)
 413                                 break;          /* Out of memory */
 414                 }
 415
 416                 /*
 417                  * call radix_tree_preload() while we can wait.
 418                  */
 419                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 420                 if (err)
 421                         break;
 422
 423                 /*
 424                  * Swap entry may have been freed since our caller observed it.
 425                  */
 426                 err = swapcache_prepare(entry);
 427                 if (err == -EEXIST) {
 428                         radix_tree_preload_end();
 429                         /*
 430                          * We might race against get_swap_page() and stumble
 431                          * across a SWAP_HAS_CACHE swap_map entry whose page
 432                          * has not been brought into the swapcache yet.
 433                          */
 434                         cond_resched();
 435                         continue;
 436                 }
 437                 if (err) {              /* swp entry is obsolete ? */
 438                         radix_tree_preload_end();
 439                         break;
 440                 }
 441
 442                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 443                 __SetPageLocked(new_page);
 444                 __SetPageSwapBacked(new_page);
 445                 err = __add_to_swap_cache(new_page, entry);
 446                 if (likely(!err)) {
 447                         radix_tree_preload_end();
 448                         /*
 449                          * Initiate read into locked page and return.
 450                          */
 451                         lru_cache_add_anon(new_page);
 452                         *new_page_allocated = true;
 453                         return new_page;
 454                 }
 455                 radix_tree_preload_end();
 456                 __ClearPageLocked(new_page);
 457                 /*
 458                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 459                  * clear SWAP_HAS_CACHE flag.
 460                  */
 461                 put_swap_page(new_page, entry);
 462         } while (err != -ENOMEM);
 463
 464         if (new_page)
 465                 put_page(new_page);
 466         return found_page;
 467 }
 468
 469 /*
 470  * Locate a page of swap in physical memory, reserving swap cache space
 471  * and reading the disk if it is not already cached.
 472  * A failure return means that either the page allocation failed or that
 473  * the swap entry is no longer in use.
 474  */
 475 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 476                 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
 477 {
 478         bool page_was_allocated;
 479         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 480                         vma, addr, &page_was_allocated);
 481
 482         if (page_was_allocated)
 483                 swap_readpage(retpage, do_poll);
 484
 485         return retpage;
 486 }
 487
 488 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
 489                                       unsigned long offset,
 490                                       int hits,
 491                                       int max_pages,
 492                                       int prev_win)
 493 {
 494         unsigned int pages, last_ra;
 495
 496         /*
 497          * This heuristic has been found to work well on both sequential and
 498          * random loads, swapping to hard disk or to SSD: please don't ask
 499          * what the "+ 2" means, it just happens to work well, that's all.
 500          */
 501         pages = hits + 2;
 502         if (pages == 2) {
 503                 /*
 504                  * We can have no readahead hits to judge by: but must not get
 505                  * stuck here forever, so check for an adjacent offset instead
 506                  * (and don't even bother to check whether swap type is same).
 507                  */
 508                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 509                         pages = 1;
 510         } else {
 511                 unsigned int roundup = 4;
 512                 while (roundup < pages)
 513                         roundup <<= 1;
 514                 pages = roundup;
 515         }
 516
 517         if (pages > max_pages)
 518                 pages = max_pages;
 519
 520         /* Don't shrink readahead too fast */
 521         last_ra = prev_win / 2;
 522         if (pages < last_ra)
 523                 pages = last_ra;
 524
 525         return pages;
 526 }
 527
 528 static unsigned long swapin_nr_pages(unsigned long offset)
 529 {
 530         static unsigned long prev_offset;
 531         unsigned int hits, pages, max_pages;
 532         static atomic_t last_readahead_pages;
 533
 534         max_pages = 1 << READ_ONCE(page_cluster);
 535         if (max_pages <= 1)
 536                 return 1;
 537
 538         hits = atomic_xchg(&swapin_readahead_hits, 0);
 539         pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
 540                                   atomic_read(&last_readahead_pages));
 541         if (!hits)
 542                 prev_offset = offset;
 543         atomic_set(&last_readahead_pages, pages);
 544
 545         return pages;
 546 }
 547
 548 /**
 549  * swap_cluster_readahead - swap in pages in hope we need them soon
 550  * @entry: swap entry of this memory
 551  * @gfp_mask: memory allocation flags
 552  * @vmf: fault information
 553  *
 554  * Returns the struct page for entry and addr, after queueing swapin.
 555  *
 556  * Primitive swap readahead code. We simply read an aligned block of
 557  * (1 << page_cluster) entries in the swap area. This method is chosen
 558  * because it doesn't cost us any seek time.  We also make sure to queue
 559  * the 'original' request together with the readahead ones...
 560  *
 561  * This has been extended to use the NUMA policies from the mm triggering
 562  * the readahead.
 563  *
 564  * Caller must hold down_read on the vma->vm_mm if vmf->vma is not NULL.
 565  */
 566 struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
 567                                 struct vm_fault *vmf)
 568 {
 569         struct page *page;
 570         unsigned long entry_offset = swp_offset(entry);
 571         unsigned long offset = entry_offset;
 572         unsigned long start_offset, end_offset;
 573         unsigned long mask;
 574         struct swap_info_struct *si = swp_swap_info(entry);
 575         struct blk_plug plug;
 576         bool do_poll = true, page_allocated;
 577         struct vm_area_struct *vma = vmf->vma;
 578         unsigned long addr = vmf->address;
 579
 580         mask = swapin_nr_pages(offset) - 1;
 581         if (!mask)
 582                 goto skip;
 583
 584         do_poll = false;
 585         /* Read a page_cluster sized and aligned cluster around offset. */
 586         start_offset = offset & ~mask;
 587         end_offset = offset | mask;
 588         if (!start_offset)      /* First page is swap header. */
 589                 start_offset++;
 590         if (end_offset >= si->max)
 591                 end_offset = si->max - 1;
 592
 593         blk_start_plug(&plug);
 594         for (offset = start_offset; offset <= end_offset ; offset++) {
 595                 /* Ok, do the async read-ahead now */
 596                 page = __read_swap_cache_async(
 597                         swp_entry(swp_type(entry), offset),
 598                         gfp_mask, vma, addr, &page_allocated);
 599                 if (!page)
 600                         continue;
 601                 if (page_allocated) {
 602                         swap_readpage(page, false);
 603                         if (offset != entry_offset) {
 604                                 SetPageReadahead(page);
 605                                 count_vm_event(SWAP_RA);
 606                         }
 607                 }
 608                 put_page(page);
 609         }
 610         blk_finish_plug(&plug);
 611
 612         lru_add_drain();        /* Push any new pages onto the LRU now */
 613 skip:
 614         return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
 615 }
 616
 617 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 618 {
 619         struct address_space *spaces, *space;
 620         unsigned int i, nr;
 621
 622         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 623         spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
 624         if (!spaces)
 625                 return -ENOMEM;
 626         for (i = 0; i < nr; i++) {
 627                 space = spaces + i;
 628                 INIT_RADIX_TREE(&space->i_pages, GFP_ATOMIC|__GFP_NOWARN);
 629                 atomic_set(&space->i_mmap_writable, 0);
 630                 space->a_ops = &swap_aops;
 631                 /* swap cache doesn't use writeback related tags */
 632                 mapping_set_no_writeback_tags(space);
 633         }
 634         nr_swapper_spaces[type] = nr;
 635         rcu_assign_pointer(swapper_spaces[type], spaces);
 636
 637         return 0;
 638 }
 639
 640 void exit_swap_address_space(unsigned int type)
 641 {
 642         struct address_space *spaces;
 643
 644         spaces = swapper_spaces[type];
 645         nr_swapper_spaces[type] = 0;
 646         rcu_assign_pointer(swapper_spaces[type], NULL);
 647         synchronize_rcu();
 648         kvfree(spaces);
 649 }
 650
 651 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
 652                                      unsigned long faddr,
 653                                      unsigned long lpfn,
 654                                      unsigned long rpfn,
 655                                      unsigned long *start,
 656                                      unsigned long *end)
 657 {
 658         *start = max3(lpfn, PFN_DOWN(vma->vm_start),
 659                       PFN_DOWN(faddr & PMD_MASK));
 660         *end = min3(rpfn, PFN_DOWN(vma->vm_end),
 661                     PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
 662 }
 663
 664 static void swap_ra_info(struct vm_fault *vmf,
 665                         struct vma_swap_readahead *ra_info)
 666 {
 667         struct vm_area_struct *vma = vmf->vma;
 668         unsigned long ra_val;
 669         swp_entry_t entry;
 670         unsigned long faddr, pfn, fpfn;
 671         unsigned long start, end;
 672         pte_t *pte, *orig_pte;
 673         unsigned int max_win, hits, prev_win, win, left;
 674 #ifndef CONFIG_64BIT
 675         pte_t *tpte;
 676 #endif
 677
 678         max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
 679                              SWAP_RA_ORDER_CEILING);
 680         if (max_win == 1) {
 681                 ra_info->win = 1;
 682                 return;
 683         }
 684
 685         faddr = vmf->address;
 686         orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
 687         entry = pte_to_swp_entry(*pte);
 688         if ((unlikely(non_swap_entry(entry)))) {
 689                 pte_unmap(orig_pte);
 690                 return;
 691         }
 692
 693         fpfn = PFN_DOWN(faddr);
 694         ra_val = GET_SWAP_RA_VAL(vma);
 695         pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
 696         prev_win = SWAP_RA_WIN(ra_val);
 697         hits = SWAP_RA_HITS(ra_val);
 698         ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
 699                                                max_win, prev_win);
 700         atomic_long_set(&vma->swap_readahead_info,
 701                         SWAP_RA_VAL(faddr, win, 0));
 702
 703         if (win == 1) {
 704                 pte_unmap(orig_pte);
 705                 return;
 706         }
 707
 708         /* Copy the PTEs because the page table may be unmapped */
 709         if (fpfn == pfn + 1)
 710                 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
 711         else if (pfn == fpfn + 1)
 712                 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
 713                                   &start, &end);
 714         else {
 715                 left = (win - 1) / 2;
 716                 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
 717                                   &start, &end);
 718         }
 719         ra_info->nr_pte = end - start;
 720         ra_info->offset = fpfn - start;
 721         pte -= ra_info->offset;
 722 #ifdef CONFIG_64BIT
 723         ra_info->ptes = pte;
 724 #else
 725         tpte = ra_info->ptes;
 726         for (pfn = start; pfn != end; pfn++)
 727                 *tpte++ = *pte++;
 728 #endif
 729         pte_unmap(orig_pte);
 730 }
 731
 732 static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
 733                                        struct vm_fault *vmf)
 734 {
 735         struct blk_plug plug;
 736         struct vm_area_struct *vma = vmf->vma;
 737         struct page *page;
 738         pte_t *pte, pentry;
 739         swp_entry_t entry;
 740         unsigned int i;
 741         bool page_allocated;
 742         struct vma_swap_readahead ra_info = {0,};
 743
 744         swap_ra_info(vmf, &ra_info);
 745         if (ra_info.win == 1)
 746                 goto skip;
 747
 748         blk_start_plug(&plug);
 749         for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
 750              i++, pte++) {
 751                 pentry = *pte;
 752                 if (pte_none(pentry))
 753                         continue;
 754                 if (pte_present(pentry))
 755                         continue;
 756                 entry = pte_to_swp_entry(pentry);
 757                 if (unlikely(non_swap_entry(entry)))
 758                         continue;
 759                 page = __read_swap_cache_async(entry, gfp_mask, vma,
 760                                                vmf->address, &page_allocated);
 761                 if (!page)
 762                         continue;
 763                 if (page_allocated) {
 764                         swap_readpage(page, false);
 765                         if (i != ra_info.offset) {
 766                                 SetPageReadahead(page);
 767                                 count_vm_event(SWAP_RA);
 768                         }
 769                 }
 770                 put_page(page);
 771         }
 772         blk_finish_plug(&plug);
 773         lru_add_drain();
 774 skip:
 775         return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
 776                                      ra_info.win == 1);
 777 }
 778
 779 /**
 780  * swapin_readahead - swap in pages in hope we need them soon
 781  * @entry: swap entry of this memory
 782  * @gfp_mask: memory allocation flags
 783  * @vmf: fault information
 784  *
 785  * Returns the struct page for entry and addr, after queueing swapin.
 786  *
 787  * It's a main entry function for swap readahead. By the configuration,
 788  * it will read ahead blocks by cluster-based(ie, physical disk based)
 789  * or vma-based(ie, virtual address based on faulty address) readahead.
 790  */
 791 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 792                                 struct vm_fault *vmf)
 793 {
 794         return swap_use_vma_readahead() ?
 795                         swap_vma_readahead(entry, gfp_mask, vmf) :
 796                         swap_cluster_readahead(entry, gfp_mask, vmf);
 797 }
 798
 799 #ifdef CONFIG_SYSFS
 800 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
 801                                      struct kobj_attribute *attr, char *buf)
 802 {
 803         return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
 804 }
 805 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
 806                                       struct kobj_attribute *attr,
 807                                       const char *buf, size_t count)
 808 {
 809         if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
 810                 enable_vma_readahead = true;
 811         else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
 812                 enable_vma_readahead = false;
 813         else
 814                 return -EINVAL;
 815
 816         return count;
 817 }
 818 static struct kobj_attribute vma_ra_enabled_attr =
 819         __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
 820                vma_ra_enabled_store);
 821
 822 static struct attribute *swap_attrs[] = {
 823         &vma_ra_enabled_attr.attr,
 824         NULL,
 825 };
 826
 827 static struct attribute_group swap_attr_group = {
 828         .attrs = swap_attrs,
 829 };
 830
 831 static int __init swap_init_sysfs(void)
 832 {
 833         int err;
 834         struct kobject *swap_kobj;
 835
 836         swap_kobj = kobject_create_and_add("swap", mm_kobj);
 837         if (!swap_kobj) {
 838                 pr_err("failed to create swap kobject\n");
 839                 return -ENOMEM;
 840         }
 841         err = sysfs_create_group(swap_kobj, &swap_attr_group);
 842         if (err) {
 843                 pr_err("failed to register swap group\n");
 844                 goto delete_obj;
 845         }
 846         return 0;
 847
 848 delete_obj:
 849         kobject_put(swap_kobj);
 850         return err;
 851 }
 852 subsys_initcall(swap_init_sysfs);
 853 #endif