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 #include "internal.h"
  27
  28 /*
  29  * swapper_space is a fiction, retained to simplify the path through
  30  * vmscan's shrink_page_list.
  31  */
  32 static const struct address_space_operations swap_aops = {
  33         .writepage      = swap_writepage,
  34         .set_page_dirty = swap_set_page_dirty,
  35 #ifdef CONFIG_MIGRATION
  36         .migratepage    = migrate_page,
  37 #endif
  38 };
  39
  40 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
  41 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
  42 static bool enable_vma_readahead __read_mostly = true;
  43
  44 #define SWAP_RA_WIN_SHIFT       (PAGE_SHIFT / 2)
  45 #define SWAP_RA_HITS_MASK       ((1UL << SWAP_RA_WIN_SHIFT) - 1)
  46 #define SWAP_RA_HITS_MAX        SWAP_RA_HITS_MASK
  47 #define SWAP_RA_WIN_MASK        (~PAGE_MASK & ~SWAP_RA_HITS_MASK)
  48
  49 #define SWAP_RA_HITS(v)         ((v) & SWAP_RA_HITS_MASK)
  50 #define SWAP_RA_WIN(v)          (((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
  51 #define SWAP_RA_ADDR(v)         ((v) & PAGE_MASK)
  52
  53 #define SWAP_RA_VAL(addr, win, hits)                            \
  54         (((addr) & PAGE_MASK) |                                 \
  55          (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |    \
  56          ((hits) & SWAP_RA_HITS_MASK))
  57
  58 /* Initial readahead hits is 4 to start up with a small window */
  59 #define GET_SWAP_RA_VAL(vma)                                    \
  60         (atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
  61
  62 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  63 #define ADD_CACHE_INFO(x, nr)   do { swap_cache_info.x += (nr); } while (0)
  64
  65 static struct {
  66         unsigned long add_total;
  67         unsigned long del_total;
  68         unsigned long find_success;
  69         unsigned long find_total;
  70 } swap_cache_info;
  71
  72 unsigned long total_swapcache_pages(void)
  73 {
  74         unsigned int i, j, nr;
  75         unsigned long ret = 0;
  76         struct address_space *spaces;
  77
  78         rcu_read_lock();
  79         for (i = 0; i < MAX_SWAPFILES; i++) {
  80                 /*
  81                  * The corresponding entries in nr_swapper_spaces and
  82                  * swapper_spaces will be reused only after at least
  83                  * one grace period.  So it is impossible for them
  84                  * belongs to different usage.
  85                  */
  86                 nr = nr_swapper_spaces[i];
  87                 spaces = rcu_dereference(swapper_spaces[i]);
  88                 if (!nr || !spaces)
  89                         continue;
  90                 for (j = 0; j < nr; j++)
  91                         ret += spaces[j].nrpages;
  92         }
  93         rcu_read_unlock();
  94         return ret;
  95 }
  96
  97 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  98
  99 void show_swap_cache_info(void)
 100 {
 101         printk("%lu pages in swap cache\n", total_swapcache_pages());
 102         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
 103                 swap_cache_info.add_total, swap_cache_info.del_total,
 104                 swap_cache_info.find_success, swap_cache_info.find_total);
 105         printk("Free swap  = %ldkB\n",
 106                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
 107         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
 108 }
 109
 110 /*
 111  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
 112  * but sets SwapCache flag and private instead of mapping and index.
 113  */
 114 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
 115 {
 116         int error, i, nr = hpage_nr_pages(page);
 117         struct address_space *address_space;
 118         pgoff_t idx = swp_offset(entry);
 119
 120         VM_BUG_ON_PAGE(!PageLocked(page), page);
 121         VM_BUG_ON_PAGE(PageSwapCache(page), page);
 122         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 123
 124         page_ref_add(page, nr);
 125         SetPageSwapCache(page);
 126
 127         address_space = swap_address_space(entry);
 128         xa_lock_irq(&address_space->i_pages);
 129         for (i = 0; i < nr; i++) {
 130                 set_page_private(page + i, entry.val + i);
 131                 error = radix_tree_insert(&address_space->i_pages,
 132                                           idx + i, page + i);
 133                 if (unlikely(error))
 134                         break;
 135         }
 136         if (likely(!error)) {
 137                 address_space->nrpages += nr;
 138                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
 139                 ADD_CACHE_INFO(add_total, nr);
 140         } else {
 141                 /*
 142                  * Only the context which have set SWAP_HAS_CACHE flag
 143                  * would call add_to_swap_cache().
 144                  * So add_to_swap_cache() doesn't returns -EEXIST.
 145                  */
 146                 VM_BUG_ON(error == -EEXIST);
 147                 set_page_private(page + i, 0UL);
 148                 while (i--) {
 149                         radix_tree_delete(&address_space->i_pages, idx + i);
 150                         set_page_private(page + i, 0UL);
 151                 }
 152                 ClearPageSwapCache(page);
 153                 page_ref_sub(page, nr);
 154         }
 155         xa_unlock_irq(&address_space->i_pages);
 156
 157         return error;
 158 }
 159
 160
 161 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 162 {
 163         int error;
 164
 165         error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
 166         if (!error) {
 167                 error = __add_to_swap_cache(page, entry);
 168                 radix_tree_preload_end();
 169         }
 170         return error;
 171 }
 172
 173 /*
 174  * This must be called only on pages that have
 175  * been verified to be in the swap cache.
 176  */
 177 void __delete_from_swap_cache(struct page *page)
 178 {
 179         struct address_space *address_space;
 180         int i, nr = hpage_nr_pages(page);
 181         swp_entry_t entry;
 182         pgoff_t idx;
 183
 184         VM_BUG_ON_PAGE(!PageLocked(page), page);
 185         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 186         VM_BUG_ON_PAGE(PageWriteback(page), page);
 187
 188         entry.val = page_private(page);
 189         address_space = swap_address_space(entry);
 190         idx = swp_offset(entry);
 191         for (i = 0; i < nr; i++) {
 192                 radix_tree_delete(&address_space->i_pages, idx + i);
 193                 set_page_private(page + i, 0);
 194         }
 195         ClearPageSwapCache(page);
 196         address_space->nrpages -= nr;
 197         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
 198         ADD_CACHE_INFO(del_total, nr);
 199 }
 200
 201 /**
 202  * add_to_swap - allocate swap space for a page
 203  * @page: page we want to move to swap
 204  *
 205  * Allocate swap space for the page and add the page to the
 206  * swap cache.  Caller needs to hold the page lock.
 207  */
 208 int add_to_swap(struct page *page)
 209 {
 210         swp_entry_t entry;
 211         int err;
 212
 213         VM_BUG_ON_PAGE(!PageLocked(page), page);
 214         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 215
 216         entry = get_swap_page(page);
 217         if (!entry.val)
 218                 return 0;
 219
 220         /*
 221          * Radix-tree node allocations from PF_MEMALLOC contexts could
 222          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 223          * stops emergency reserves from being allocated.
 224          *
 225          * TODO: this could cause a theoretical memory reclaim
 226          * deadlock in the swap out path.
 227          */
 228         /*
 229          * Add it to the swap cache.
 230          */
 231         err = add_to_swap_cache(page, entry,
 232                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 233         /* -ENOMEM radix-tree allocation failure */
 234         if (err)
 235                 /*
 236                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 237                  * clear SWAP_HAS_CACHE flag.
 238                  */
 239                 goto fail;
 240         /*
 241          * Normally the page will be dirtied in unmap because its pte should be
 242          * dirty. A special case is MADV_FREE page. The page'e pte could have
 243          * dirty bit cleared but the page's SwapBacked bit is still set because
 244          * clearing the dirty bit and SwapBacked bit has no lock protected. For
 245          * such page, unmap will not set dirty bit for it, so page reclaim will
 246          * not write the page out. This can cause data corruption when the page
 247          * is swap in later. Always setting the dirty bit for the page solves
 248          * the problem.
 249          */
 250         set_page_dirty(page);
 251
 252         return 1;
 253
 254 fail:
 255         put_swap_page(page, entry);
 256         return 0;
 257 }
 258
 259 /*
 260  * This must be called only on pages that have
 261  * been verified to be in the swap cache and locked.
 262  * It will never put the page into the free list,
 263  * the caller has a reference on the page.
 264  */
 265 void delete_from_swap_cache(struct page *page)
 266 {
 267         swp_entry_t entry;
 268         struct address_space *address_space;
 269
 270         entry.val = page_private(page);
 271
 272         address_space = swap_address_space(entry);
 273         xa_lock_irq(&address_space->i_pages);
 274         __delete_from_swap_cache(page);
 275         xa_unlock_irq(&address_space->i_pages);
 276
 277         put_swap_page(page, entry);
 278         page_ref_sub(page, hpage_nr_pages(page));
 279 }
 280
 281 /*
 282  * If we are the only user, then try to free up the swap cache.
 283  *
 284  * Its ok to check for PageSwapCache without the page lock
 285  * here because we are going to recheck again inside
 286  * try_to_free_swap() _with_ the lock.
 287  *                                      - Marcelo
 288  */
 289 static inline void free_swap_cache(struct page *page)
 290 {
 291         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 292                 try_to_free_swap(page);
 293                 unlock_page(page);
 294         }
 295 }
 296
 297 /*
 298  * Perform a free_page(), also freeing any swap cache associated with
 299  * this page if it is the last user of the page.
 300  */
 301 void free_page_and_swap_cache(struct page *page)
 302 {
 303         free_swap_cache(page);
 304         if (!is_huge_zero_page(page))
 305                 put_page(page);
 306 }
 307
 308 /*
 309  * Passed an array of pages, drop them all from swapcache and then release
 310  * them.  They are removed from the LRU and freed if this is their last use.
 311  */
 312 void free_pages_and_swap_cache(struct page **pages, int nr)
 313 {
 314         struct page **pagep = pages;
 315         int i;
 316
 317         lru_add_drain();
 318         for (i = 0; i < nr; i++)
 319                 free_swap_cache(pagep[i]);
 320         release_pages(pagep, nr);
 321 }
 322
 323 static inline bool swap_use_vma_readahead(void)
 324 {
 325         return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
 326 }
 327
 328 /*
 329  * Lookup a swap entry in the swap cache. A found page will be returned
 330  * unlocked and with its refcount incremented - we rely on the kernel
 331  * lock getting page table operations atomic even if we drop the page
 332  * lock before returning.
 333  */
 334 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
 335                                unsigned long addr)
 336 {
 337         struct page *page;
 338
 339         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 340
 341         INC_CACHE_INFO(find_total);
 342         if (page) {
 343                 bool vma_ra = swap_use_vma_readahead();
 344                 bool readahead;
 345
 346                 INC_CACHE_INFO(find_success);
 347                 /*
 348                  * At the moment, we don't support PG_readahead for anon THP
 349                  * so let's bail out rather than confusing the readahead stat.
 350                  */
 351                 if (unlikely(PageTransCompound(page)))
 352                         return page;
 353
 354                 readahead = TestClearPageReadahead(page);
 355                 if (vma && vma_ra) {
 356                         unsigned long ra_val;
 357                         int win, hits;
 358
 359                         ra_val = GET_SWAP_RA_VAL(vma);
 360                         win = SWAP_RA_WIN(ra_val);
 361                         hits = SWAP_RA_HITS(ra_val);
 362                         if (readahead)
 363                                 hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
 364                         atomic_long_set(&vma->swap_readahead_info,
 365                                         SWAP_RA_VAL(addr, win, hits));
 366                 }
 367
 368                 if (readahead) {
 369                         count_vm_event(SWAP_RA_HIT);
 370                         if (!vma || !vma_ra)
 371                                 atomic_inc(&swapin_readahead_hits);
 372                 }
 373         }
 374
 375         return page;
 376 }
 377
 378 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 379                         struct vm_area_struct *vma, unsigned long addr,
 380                         bool *new_page_allocated)
 381 {
 382         struct page *found_page, *new_page = NULL;
 383         struct address_space *swapper_space = swap_address_space(entry);
 384         int err;
 385         *new_page_allocated = false;
 386
 387         do {
 388                 /*
 389                  * First check the swap cache.  Since this is normally
 390                  * called after lookup_swap_cache() failed, re-calling
 391                  * that would confuse statistics.
 392                  */
 393                 found_page = find_get_page(swapper_space, swp_offset(entry));
 394                 if (found_page)
 395                         break;
 396
 397                 /*
 398                  * Just skip read ahead for unused swap slot.
 399                  * During swap_off when swap_slot_cache is disabled,
 400                  * we have to handle the race between putting
 401                  * swap entry in swap cache and marking swap slot
 402                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 403                  * else swap_off will be aborted if we return NULL.
 404                  */
 405                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 406                         break;
 407
 408                 /*
 409                  * Get a new page to read into from swap.
 410                  */
 411                 if (!new_page) {
 412                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 413                         if (!new_page)
 414                                 break;          /* Out of memory */
 415                 }
 416
 417                 /*
 418                  * call radix_tree_preload() while we can wait.
 419                  */
 420                 err = radix_tree_maybe_preload(gfp_mask & GFP_RECLAIM_MASK);
 421                 if (err)
 422                         break;
 423
 424                 /*
 425                  * Swap entry may have been freed since our caller observed it.
 426                  */
 427                 err = swapcache_prepare(entry);
 428                 if (err == -EEXIST) {
 429                         radix_tree_preload_end();
 430                         /*
 431                          * We might race against get_swap_page() and stumble
 432                          * across a SWAP_HAS_CACHE swap_map entry whose page
 433                          * has not been brought into the swapcache yet.
 434                          */
 435                         cond_resched();
 436                         continue;
 437                 }
 438                 if (err) {              /* swp entry is obsolete ? */
 439                         radix_tree_preload_end();
 440                         break;
 441                 }
 442
 443                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 444                 __SetPageLocked(new_page);
 445                 __SetPageSwapBacked(new_page);
 446                 err = __add_to_swap_cache(new_page, entry);
 447                 if (likely(!err)) {
 448                         radix_tree_preload_end();
 449                         /*
 450                          * Initiate read into locked page and return.
 451                          */
 452                         lru_cache_add_anon(new_page);
 453                         *new_page_allocated = true;
 454                         return new_page;
 455                 }
 456                 radix_tree_preload_end();
 457                 __ClearPageLocked(new_page);
 458                 /*
 459                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 460                  * clear SWAP_HAS_CACHE flag.
 461                  */
 462                 put_swap_page(new_page, entry);
 463         } while (err != -ENOMEM);
 464
 465         if (new_page)
 466                 put_page(new_page);
 467         return found_page;
 468 }
 469
 470 /*
 471  * Locate a page of swap in physical memory, reserving swap cache space
 472  * and reading the disk if it is not already cached.
 473  * A failure return means that either the page allocation failed or that
 474  * the swap entry is no longer in use.
 475  */
 476 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 477                 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
 478 {
 479         bool page_was_allocated;
 480         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 481                         vma, addr, &page_was_allocated);
 482
 483         if (page_was_allocated)
 484                 swap_readpage(retpage, do_poll);
 485
 486         return retpage;
 487 }
 488
 489 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
 490                                       unsigned long offset,
 491                                       int hits,
 492                                       int max_pages,
 493                                       int prev_win)
 494 {
 495         unsigned int pages, last_ra;
 496
 497         /*
 498          * This heuristic has been found to work well on both sequential and
 499          * random loads, swapping to hard disk or to SSD: please don't ask
 500          * what the "+ 2" means, it just happens to work well, that's all.
 501          */
 502         pages = hits + 2;
 503         if (pages == 2) {
 504                 /*
 505                  * We can have no readahead hits to judge by: but must not get
 506                  * stuck here forever, so check for an adjacent offset instead
 507                  * (and don't even bother to check whether swap type is same).
 508                  */
 509                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 510                         pages = 1;
 511         } else {
 512                 unsigned int roundup = 4;
 513                 while (roundup < pages)
 514                         roundup <<= 1;
 515                 pages = roundup;
 516         }
 517
 518         if (pages > max_pages)
 519                 pages = max_pages;
 520
 521         /* Don't shrink readahead too fast */
 522         last_ra = prev_win / 2;
 523         if (pages < last_ra)
 524                 pages = last_ra;
 525
 526         return pages;
 527 }
 528
 529 static unsigned long swapin_nr_pages(unsigned long offset)
 530 {
 531         static unsigned long prev_offset;
 532         unsigned int hits, pages, max_pages;
 533         static atomic_t last_readahead_pages;
 534
 535         max_pages = 1 << READ_ONCE(page_cluster);
 536         if (max_pages <= 1)
 537                 return 1;
 538
 539         hits = atomic_xchg(&swapin_readahead_hits, 0);
 540         pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
 541                                   max_pages,
 542                                   atomic_read(&last_readahead_pages));
 543         if (!hits)
 544                 WRITE_ONCE(prev_offset, offset);
 545         atomic_set(&last_readahead_pages, pages);
 546
 547         return pages;
 548 }
 549
 550 /**
 551  * swap_cluster_readahead - swap in pages in hope we need them soon
 552  * @entry: swap entry of this memory
 553  * @gfp_mask: memory allocation flags
 554  * @vmf: fault information
 555  *
 556  * Returns the struct page for entry and addr, after queueing swapin.
 557  *
 558  * Primitive swap readahead code. We simply read an aligned block of
 559  * (1 << page_cluster) entries in the swap area. This method is chosen
 560  * because it doesn't cost us any seek time.  We also make sure to queue
 561  * the 'original' request together with the readahead ones...
 562  *
 563  * This has been extended to use the NUMA policies from the mm triggering
 564  * the readahead.
 565  *
 566  * Caller must hold down_read on the vma->vm_mm if vmf->vma is not NULL.
 567  */
 568 struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
 569                                 struct vm_fault *vmf)
 570 {
 571         struct page *page;
 572         unsigned long entry_offset = swp_offset(entry);
 573         unsigned long offset = entry_offset;
 574         unsigned long start_offset, end_offset;
 575         unsigned long mask;
 576         struct swap_info_struct *si = swp_swap_info(entry);
 577         struct blk_plug plug;
 578         bool do_poll = true, page_allocated;
 579         struct vm_area_struct *vma = vmf->vma;
 580         unsigned long addr = vmf->address;
 581
 582         mask = swapin_nr_pages(offset) - 1;
 583         if (!mask)
 584                 goto skip;
 585
 586         do_poll = false;
 587         /* Read a page_cluster sized and aligned cluster around offset. */
 588         start_offset = offset & ~mask;
 589         end_offset = offset | mask;
 590         if (!start_offset)      /* First page is swap header. */
 591                 start_offset++;
 592         if (end_offset >= si->max)
 593                 end_offset = si->max - 1;
 594
 595         blk_start_plug(&plug);
 596         for (offset = start_offset; offset <= end_offset ; offset++) {
 597                 /* Ok, do the async read-ahead now */
 598                 page = __read_swap_cache_async(
 599                         swp_entry(swp_type(entry), offset),
 600                         gfp_mask, vma, addr, &page_allocated);
 601                 if (!page)
 602                         continue;
 603                 if (page_allocated) {
 604                         swap_readpage(page, false);
 605                         if (offset != entry_offset) {
 606                                 SetPageReadahead(page);
 607                                 count_vm_event(SWAP_RA);
 608                         }
 609                 }
 610                 put_page(page);
 611         }
 612         blk_finish_plug(&plug);
 613
 614         lru_add_drain();        /* Push any new pages onto the LRU now */
 615 skip:
 616         return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
 617 }
 618
 619 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 620 {
 621         struct address_space *spaces, *space;
 622         unsigned int i, nr;
 623
 624         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 625         spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
 626         if (!spaces)
 627                 return -ENOMEM;
 628         for (i = 0; i < nr; i++) {
 629                 space = spaces + i;
 630                 INIT_RADIX_TREE(&space->i_pages, GFP_ATOMIC|__GFP_NOWARN);
 631                 atomic_set(&space->i_mmap_writable, 0);
 632                 space->a_ops = &swap_aops;
 633                 /* swap cache doesn't use writeback related tags */
 634                 mapping_set_no_writeback_tags(space);
 635         }
 636         nr_swapper_spaces[type] = nr;
 637         rcu_assign_pointer(swapper_spaces[type], spaces);
 638
 639         return 0;
 640 }
 641
 642 void exit_swap_address_space(unsigned int type)
 643 {
 644         struct address_space *spaces;
 645
 646         spaces = swapper_spaces[type];
 647         nr_swapper_spaces[type] = 0;
 648         rcu_assign_pointer(swapper_spaces[type], NULL);
 649         synchronize_rcu();
 650         kvfree(spaces);
 651 }
 652
 653 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
 654                                      unsigned long faddr,
 655                                      unsigned long lpfn,
 656                                      unsigned long rpfn,
 657                                      unsigned long *start,
 658                                      unsigned long *end)
 659 {
 660         *start = max3(lpfn, PFN_DOWN(vma->vm_start),
 661                       PFN_DOWN(faddr & PMD_MASK));
 662         *end = min3(rpfn, PFN_DOWN(vma->vm_end),
 663                     PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
 664 }
 665
 666 static void swap_ra_info(struct vm_fault *vmf,
 667                         struct vma_swap_readahead *ra_info)
 668 {
 669         struct vm_area_struct *vma = vmf->vma;
 670         unsigned long ra_val;
 671         swp_entry_t entry;
 672         unsigned long faddr, pfn, fpfn;
 673         unsigned long start, end;
 674         pte_t *pte, *orig_pte;
 675         unsigned int max_win, hits, prev_win, win, left;
 676 #ifndef CONFIG_64BIT
 677         pte_t *tpte;
 678 #endif
 679
 680         max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
 681                              SWAP_RA_ORDER_CEILING);
 682         if (max_win == 1) {
 683                 ra_info->win = 1;
 684                 return;
 685         }
 686
 687         faddr = vmf->address;
 688         orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
 689         entry = pte_to_swp_entry(*pte);
 690         if ((unlikely(non_swap_entry(entry)))) {
 691                 pte_unmap(orig_pte);
 692                 return;
 693         }
 694
 695         fpfn = PFN_DOWN(faddr);
 696         ra_val = GET_SWAP_RA_VAL(vma);
 697         pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
 698         prev_win = SWAP_RA_WIN(ra_val);
 699         hits = SWAP_RA_HITS(ra_val);
 700         ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
 701                                                max_win, prev_win);
 702         atomic_long_set(&vma->swap_readahead_info,
 703                         SWAP_RA_VAL(faddr, win, 0));
 704
 705         if (win == 1) {
 706                 pte_unmap(orig_pte);
 707                 return;
 708         }
 709
 710         /* Copy the PTEs because the page table may be unmapped */
 711         if (fpfn == pfn + 1)
 712                 swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
 713         else if (pfn == fpfn + 1)
 714                 swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
 715                                   &start, &end);
 716         else {
 717                 left = (win - 1) / 2;
 718                 swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
 719                                   &start, &end);
 720         }
 721         ra_info->nr_pte = end - start;
 722         ra_info->offset = fpfn - start;
 723         pte -= ra_info->offset;
 724 #ifdef CONFIG_64BIT
 725         ra_info->ptes = pte;
 726 #else
 727         tpte = ra_info->ptes;
 728         for (pfn = start; pfn != end; pfn++)
 729                 *tpte++ = *pte++;
 730 #endif
 731         pte_unmap(orig_pte);
 732 }
 733
 734 static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
 735                                        struct vm_fault *vmf)
 736 {
 737         struct blk_plug plug;
 738         struct vm_area_struct *vma = vmf->vma;
 739         struct page *page;
 740         pte_t *pte, pentry;
 741         swp_entry_t entry;
 742         unsigned int i;
 743         bool page_allocated;
 744         struct vma_swap_readahead ra_info = {0,};
 745
 746         swap_ra_info(vmf, &ra_info);
 747         if (ra_info.win == 1)
 748                 goto skip;
 749
 750         blk_start_plug(&plug);
 751         for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
 752              i++, pte++) {
 753                 pentry = *pte;
 754                 if (pte_none(pentry))
 755                         continue;
 756                 if (pte_present(pentry))
 757                         continue;
 758                 entry = pte_to_swp_entry(pentry);
 759                 if (unlikely(non_swap_entry(entry)))
 760                         continue;
 761                 page = __read_swap_cache_async(entry, gfp_mask, vma,
 762                                                vmf->address, &page_allocated);
 763                 if (!page)
 764                         continue;
 765                 if (page_allocated) {
 766                         swap_readpage(page, false);
 767                         if (i != ra_info.offset) {
 768                                 SetPageReadahead(page);
 769                                 count_vm_event(SWAP_RA);
 770                         }
 771                 }
 772                 put_page(page);
 773         }
 774         blk_finish_plug(&plug);
 775         lru_add_drain();
 776 skip:
 777         return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
 778                                      ra_info.win == 1);
 779 }
 780
 781 /**
 782  * swapin_readahead - swap in pages in hope we need them soon
 783  * @entry: swap entry of this memory
 784  * @gfp_mask: memory allocation flags
 785  * @vmf: fault information
 786  *
 787  * Returns the struct page for entry and addr, after queueing swapin.
 788  *
 789  * It's a main entry function for swap readahead. By the configuration,
 790  * it will read ahead blocks by cluster-based(ie, physical disk based)
 791  * or vma-based(ie, virtual address based on faulty address) readahead.
 792  */
 793 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 794                                 struct vm_fault *vmf)
 795 {
 796         return swap_use_vma_readahead() ?
 797                         swap_vma_readahead(entry, gfp_mask, vmf) :
 798                         swap_cluster_readahead(entry, gfp_mask, vmf);
 799 }
 800
 801 #ifdef CONFIG_SYSFS
 802 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
 803                                      struct kobj_attribute *attr, char *buf)
 804 {
 805         return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
 806 }
 807 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
 808                                       struct kobj_attribute *attr,
 809                                       const char *buf, size_t count)
 810 {
 811         if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
 812                 enable_vma_readahead = true;
 813         else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
 814                 enable_vma_readahead = false;
 815         else
 816                 return -EINVAL;
 817
 818         return count;
 819 }
 820 static struct kobj_attribute vma_ra_enabled_attr =
 821         __ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
 822                vma_ra_enabled_store);
 823
 824 static struct attribute *swap_attrs[] = {
 825         &vma_ra_enabled_attr.attr,
 826         NULL,
 827 };
 828
 829 static struct attribute_group swap_attr_group = {
 830         .attrs = swap_attrs,
 831 };
 832
 833 static int __init swap_init_sysfs(void)
 834 {
 835         int err;
 836         struct kobject *swap_kobj;
 837
 838         swap_kobj = kobject_create_and_add("swap", mm_kobj);
 839         if (!swap_kobj) {
 840                 pr_err("failed to create swap kobject\n");
 841                 return -ENOMEM;
 842         }
 843         err = sysfs_create_group(swap_kobj, &swap_attr_group);
 844         if (err) {
 845                 pr_err("failed to register swap group\n");
 846                 goto delete_obj;
 847         }
 848         return 0;
 849
 850 delete_obj:
 851         kobject_put(swap_kobj);
 852         return err;
 853 }
 854 subsys_initcall(swap_init_sysfs);
 855 #endif