mm/swap_state.c

   1 /*
   2  *  linux/mm/swap_state.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  *
   7  *  Rewritten to use page cache, (C) 1998 Stephen Tweedie
   8  */
   9 #include <linux/mm.h>
  10 #include <linux/gfp.h>
  11 #include <linux/kernel_stat.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/init.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/pagevec.h>
  19 #include <linux/migrate.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/swap_slots.h>
  22 #include <linux/huge_mm.h>
  23
  24 #include <asm/pgtable.h>
  25
  26 /*
  27  * swapper_space is a fiction, retained to simplify the path through
  28  * vmscan's shrink_page_list.
  29  */
  30 static const struct address_space_operations swap_aops = {
  31         .writepage      = swap_writepage,
  32         .set_page_dirty = swap_set_page_dirty,
  33 #ifdef CONFIG_MIGRATION
  34         .migratepage    = migrate_page,
  35 #endif
  36 };
  37
  38 struct address_space *swapper_spaces[MAX_SWAPFILES];
  39 static unsigned int nr_swapper_spaces[MAX_SWAPFILES];
  40
  41 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  42 #define ADD_CACHE_INFO(x, nr)   do { swap_cache_info.x += (nr); } while (0)
  43
  44 static struct {
  45         unsigned long add_total;
  46         unsigned long del_total;
  47         unsigned long find_success;
  48         unsigned long find_total;
  49 } swap_cache_info;
  50
  51 unsigned long total_swapcache_pages(void)
  52 {
  53         unsigned int i, j, nr;
  54         unsigned long ret = 0;
  55         struct address_space *spaces;
  56
  57         rcu_read_lock();
  58         for (i = 0; i < MAX_SWAPFILES; i++) {
  59                 /*
  60                  * The corresponding entries in nr_swapper_spaces and
  61                  * swapper_spaces will be reused only after at least
  62                  * one grace period.  So it is impossible for them
  63                  * belongs to different usage.
  64                  */
  65                 nr = nr_swapper_spaces[i];
  66                 spaces = rcu_dereference(swapper_spaces[i]);
  67                 if (!nr || !spaces)
  68                         continue;
  69                 for (j = 0; j < nr; j++)
  70                         ret += spaces[j].nrpages;
  71         }
  72         rcu_read_unlock();
  73         return ret;
  74 }
  75
  76 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  77
  78 void show_swap_cache_info(void)
  79 {
  80         printk("%lu pages in swap cache\n", total_swapcache_pages());
  81         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  82                 swap_cache_info.add_total, swap_cache_info.del_total,
  83                 swap_cache_info.find_success, swap_cache_info.find_total);
  84         printk("Free swap  = %ldkB\n",
  85                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
  86         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  87 }
  88
  89 /*
  90  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  91  * but sets SwapCache flag and private instead of mapping and index.
  92  */
  93 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  94 {
  95         int error, i, nr = hpage_nr_pages(page);
  96         struct address_space *address_space;
  97         pgoff_t idx = swp_offset(entry);
  98
  99         VM_BUG_ON_PAGE(!PageLocked(page), page);
 100         VM_BUG_ON_PAGE(PageSwapCache(page), page);
 101         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 102
 103         page_ref_add(page, nr);
 104         SetPageSwapCache(page);
 105
 106         address_space = swap_address_space(entry);
 107         spin_lock_irq(&address_space->tree_lock);
 108         for (i = 0; i < nr; i++) {
 109                 set_page_private(page + i, entry.val + i);
 110                 error = radix_tree_insert(&address_space->page_tree,
 111                                           idx + i, page + i);
 112                 if (unlikely(error))
 113                         break;
 114         }
 115         if (likely(!error)) {
 116                 address_space->nrpages += nr;
 117                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
 118                 ADD_CACHE_INFO(add_total, nr);
 119         } else {
 120                 /*
 121                  * Only the context which have set SWAP_HAS_CACHE flag
 122                  * would call add_to_swap_cache().
 123                  * So add_to_swap_cache() doesn't returns -EEXIST.
 124                  */
 125                 VM_BUG_ON(error == -EEXIST);
 126                 set_page_private(page + i, 0UL);
 127                 while (i--) {
 128                         radix_tree_delete(&address_space->page_tree, idx + i);
 129                         set_page_private(page + i, 0UL);
 130                 }
 131                 ClearPageSwapCache(page);
 132                 page_ref_sub(page, nr);
 133         }
 134         spin_unlock_irq(&address_space->tree_lock);
 135
 136         return error;
 137 }
 138
 139
 140 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 141 {
 142         int error;
 143
 144         error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
 145         if (!error) {
 146                 error = __add_to_swap_cache(page, entry);
 147                 radix_tree_preload_end();
 148         }
 149         return error;
 150 }
 151
 152 /*
 153  * This must be called only on pages that have
 154  * been verified to be in the swap cache.
 155  */
 156 void __delete_from_swap_cache(struct page *page)
 157 {
 158         struct address_space *address_space;
 159         int i, nr = hpage_nr_pages(page);
 160         swp_entry_t entry;
 161         pgoff_t idx;
 162
 163         VM_BUG_ON_PAGE(!PageLocked(page), page);
 164         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 165         VM_BUG_ON_PAGE(PageWriteback(page), page);
 166
 167         entry.val = page_private(page);
 168         address_space = swap_address_space(entry);
 169         idx = swp_offset(entry);
 170         for (i = 0; i < nr; i++) {
 171                 radix_tree_delete(&address_space->page_tree, idx + i);
 172                 set_page_private(page + i, 0);
 173         }
 174         ClearPageSwapCache(page);
 175         address_space->nrpages -= nr;
 176         __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
 177         ADD_CACHE_INFO(del_total, nr);
 178 }
 179
 180 /**
 181  * add_to_swap - allocate swap space for a page
 182  * @page: page we want to move to swap
 183  *
 184  * Allocate swap space for the page and add the page to the
 185  * swap cache.  Caller needs to hold the page lock.
 186  */
 187 int add_to_swap(struct page *page)
 188 {
 189         swp_entry_t entry;
 190         int err;
 191
 192         VM_BUG_ON_PAGE(!PageLocked(page), page);
 193         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 194
 195         entry = get_swap_page(page);
 196         if (!entry.val)
 197                 return 0;
 198
 199         if (mem_cgroup_try_charge_swap(page, entry))
 200                 goto fail;
 201
 202         /*
 203          * Radix-tree node allocations from PF_MEMALLOC contexts could
 204          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 205          * stops emergency reserves from being allocated.
 206          *
 207          * TODO: this could cause a theoretical memory reclaim
 208          * deadlock in the swap out path.
 209          */
 210         /*
 211          * Add it to the swap cache.
 212          */
 213         err = add_to_swap_cache(page, entry,
 214                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 215         /* -ENOMEM radix-tree allocation failure */
 216         if (err)
 217                 /*
 218                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 219                  * clear SWAP_HAS_CACHE flag.
 220                  */
 221                 goto fail;
 222
 223         return 1;
 224
 225 fail:
 226         put_swap_page(page, entry);
 227         return 0;
 228 }
 229
 230 /*
 231  * This must be called only on pages that have
 232  * been verified to be in the swap cache and locked.
 233  * It will never put the page into the free list,
 234  * the caller has a reference on the page.
 235  */
 236 void delete_from_swap_cache(struct page *page)
 237 {
 238         swp_entry_t entry;
 239         struct address_space *address_space;
 240
 241         entry.val = page_private(page);
 242
 243         address_space = swap_address_space(entry);
 244         spin_lock_irq(&address_space->tree_lock);
 245         __delete_from_swap_cache(page);
 246         spin_unlock_irq(&address_space->tree_lock);
 247
 248         put_swap_page(page, entry);
 249         page_ref_sub(page, hpage_nr_pages(page));
 250 }
 251
 252 /*
 253  * If we are the only user, then try to free up the swap cache.
 254  *
 255  * Its ok to check for PageSwapCache without the page lock
 256  * here because we are going to recheck again inside
 257  * try_to_free_swap() _with_ the lock.
 258  *                                      - Marcelo
 259  */
 260 static inline void free_swap_cache(struct page *page)
 261 {
 262         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 263                 try_to_free_swap(page);
 264                 unlock_page(page);
 265         }
 266 }
 267
 268 /*
 269  * Perform a free_page(), also freeing any swap cache associated with
 270  * this page if it is the last user of the page.
 271  */
 272 void free_page_and_swap_cache(struct page *page)
 273 {
 274         free_swap_cache(page);
 275         if (!is_huge_zero_page(page))
 276                 put_page(page);
 277 }
 278
 279 /*
 280  * Passed an array of pages, drop them all from swapcache and then release
 281  * them.  They are removed from the LRU and freed if this is their last use.
 282  */
 283 void free_pages_and_swap_cache(struct page **pages, int nr)
 284 {
 285         struct page **pagep = pages;
 286         int i;
 287
 288         lru_add_drain();
 289         for (i = 0; i < nr; i++)
 290                 free_swap_cache(pagep[i]);
 291         release_pages(pagep, nr, false);
 292 }
 293
 294 /*
 295  * Lookup a swap entry in the swap cache. A found page will be returned
 296  * unlocked and with its refcount incremented - we rely on the kernel
 297  * lock getting page table operations atomic even if we drop the page
 298  * lock before returning.
 299  */
 300 struct page * lookup_swap_cache(swp_entry_t entry)
 301 {
 302         struct page *page;
 303
 304         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 305
 306         if (page && likely(!PageTransCompound(page))) {
 307                 INC_CACHE_INFO(find_success);
 308                 if (TestClearPageReadahead(page)) {
 309                         atomic_inc(&swapin_readahead_hits);
 310                         count_vm_event(SWAP_RA_HIT);
 311                 }
 312         }
 313
 314         INC_CACHE_INFO(find_total);
 315         return page;
 316 }
 317
 318 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 319                         struct vm_area_struct *vma, unsigned long addr,
 320                         bool *new_page_allocated)
 321 {
 322         struct page *found_page, *new_page = NULL;
 323         struct address_space *swapper_space = swap_address_space(entry);
 324         int err;
 325         *new_page_allocated = false;
 326
 327         do {
 328                 /*
 329                  * First check the swap cache.  Since this is normally
 330                  * called after lookup_swap_cache() failed, re-calling
 331                  * that would confuse statistics.
 332                  */
 333                 found_page = find_get_page(swapper_space, swp_offset(entry));
 334                 if (found_page)
 335                         break;
 336
 337                 /*
 338                  * Just skip read ahead for unused swap slot.
 339                  * During swap_off when swap_slot_cache is disabled,
 340                  * we have to handle the race between putting
 341                  * swap entry in swap cache and marking swap slot
 342                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 343                  * else swap_off will be aborted if we return NULL.
 344                  */
 345                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 346                         break;
 347
 348                 /*
 349                  * Get a new page to read into from swap.
 350                  */
 351                 if (!new_page) {
 352                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 353                         if (!new_page)
 354                                 break;          /* Out of memory */
 355                 }
 356
 357                 /*
 358                  * call radix_tree_preload() while we can wait.
 359                  */
 360                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 361                 if (err)
 362                         break;
 363
 364                 /*
 365                  * Swap entry may have been freed since our caller observed it.
 366                  */
 367                 err = swapcache_prepare(entry);
 368                 if (err == -EEXIST) {
 369                         radix_tree_preload_end();
 370                         /*
 371                          * We might race against get_swap_page() and stumble
 372                          * across a SWAP_HAS_CACHE swap_map entry whose page
 373                          * has not been brought into the swapcache yet.
 374                          */
 375                         cond_resched();
 376                         continue;
 377                 }
 378                 if (err) {              /* swp entry is obsolete ? */
 379                         radix_tree_preload_end();
 380                         break;
 381                 }
 382
 383                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 384                 __SetPageLocked(new_page);
 385                 __SetPageSwapBacked(new_page);
 386                 err = __add_to_swap_cache(new_page, entry);
 387                 if (likely(!err)) {
 388                         radix_tree_preload_end();
 389                         /*
 390                          * Initiate read into locked page and return.
 391                          */
 392                         lru_cache_add_anon(new_page);
 393                         *new_page_allocated = true;
 394                         return new_page;
 395                 }
 396                 radix_tree_preload_end();
 397                 __ClearPageLocked(new_page);
 398                 /*
 399                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 400                  * clear SWAP_HAS_CACHE flag.
 401                  */
 402                 put_swap_page(new_page, entry);
 403         } while (err != -ENOMEM);
 404
 405         if (new_page)
 406                 put_page(new_page);
 407         return found_page;
 408 }
 409
 410 /*
 411  * Locate a page of swap in physical memory, reserving swap cache space
 412  * and reading the disk if it is not already cached.
 413  * A failure return means that either the page allocation failed or that
 414  * the swap entry is no longer in use.
 415  */
 416 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 417                 struct vm_area_struct *vma, unsigned long addr, bool do_poll)
 418 {
 419         bool page_was_allocated;
 420         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 421                         vma, addr, &page_was_allocated);
 422
 423         if (page_was_allocated)
 424                 swap_readpage(retpage, do_poll);
 425
 426         return retpage;
 427 }
 428
 429 static unsigned long swapin_nr_pages(unsigned long offset)
 430 {
 431         static unsigned long prev_offset;
 432         unsigned int pages, max_pages, last_ra;
 433         static atomic_t last_readahead_pages;
 434
 435         max_pages = 1 << READ_ONCE(page_cluster);
 436         if (max_pages <= 1)
 437                 return 1;
 438
 439         /*
 440          * This heuristic has been found to work well on both sequential and
 441          * random loads, swapping to hard disk or to SSD: please don't ask
 442          * what the "+ 2" means, it just happens to work well, that's all.
 443          */
 444         pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
 445         if (pages == 2) {
 446                 /*
 447                  * We can have no readahead hits to judge by: but must not get
 448                  * stuck here forever, so check for an adjacent offset instead
 449                  * (and don't even bother to check whether swap type is same).
 450                  */
 451                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 452                         pages = 1;
 453                 prev_offset = offset;
 454         } else {
 455                 unsigned int roundup = 4;
 456                 while (roundup < pages)
 457                         roundup <<= 1;
 458                 pages = roundup;
 459         }
 460
 461         if (pages > max_pages)
 462                 pages = max_pages;
 463
 464         /* Don't shrink readahead too fast */
 465         last_ra = atomic_read(&last_readahead_pages) / 2;
 466         if (pages < last_ra)
 467                 pages = last_ra;
 468         atomic_set(&last_readahead_pages, pages);
 469
 470         return pages;
 471 }
 472
 473 /**
 474  * swapin_readahead - swap in pages in hope we need them soon
 475  * @entry: swap entry of this memory
 476  * @gfp_mask: memory allocation flags
 477  * @vma: user vma this address belongs to
 478  * @addr: target address for mempolicy
 479  *
 480  * Returns the struct page for entry and addr, after queueing swapin.
 481  *
 482  * Primitive swap readahead code. We simply read an aligned block of
 483  * (1 << page_cluster) entries in the swap area. This method is chosen
 484  * because it doesn't cost us any seek time.  We also make sure to queue
 485  * the 'original' request together with the readahead ones...
 486  *
 487  * This has been extended to use the NUMA policies from the mm triggering
 488  * the readahead.
 489  *
 490  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 491  */
 492 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 493                         struct vm_area_struct *vma, unsigned long addr)
 494 {
 495         struct page *page;
 496         unsigned long entry_offset = swp_offset(entry);
 497         unsigned long offset = entry_offset;
 498         unsigned long start_offset, end_offset;
 499         unsigned long mask;
 500         struct blk_plug plug;
 501         bool do_poll = true;
 502
 503         mask = swapin_nr_pages(offset) - 1;
 504         if (!mask)
 505                 goto skip;
 506
 507         do_poll = false;
 508         /* Read a page_cluster sized and aligned cluster around offset. */
 509         start_offset = offset & ~mask;
 510         end_offset = offset | mask;
 511         if (!start_offset)      /* First page is swap header. */
 512                 start_offset++;
 513
 514         blk_start_plug(&plug);
 515         for (offset = start_offset; offset <= end_offset ; offset++) {
 516                 /* Ok, do the async read-ahead now */
 517                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 518                                                 gfp_mask, vma, addr, false);
 519                 if (!page)
 520                         continue;
 521                 if (offset != entry_offset &&
 522                     likely(!PageTransCompound(page))) {
 523                         SetPageReadahead(page);
 524                         count_vm_event(SWAP_RA);
 525                 }
 526                 put_page(page);
 527         }
 528         blk_finish_plug(&plug);
 529
 530         lru_add_drain();        /* Push any new pages onto the LRU now */
 531 skip:
 532         return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
 533 }
 534
 535 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 536 {
 537         struct address_space *spaces, *space;
 538         unsigned int i, nr;
 539
 540         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 541         spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
 542         if (!spaces)
 543                 return -ENOMEM;
 544         for (i = 0; i < nr; i++) {
 545                 space = spaces + i;
 546                 INIT_RADIX_TREE(&space->page_tree, GFP_ATOMIC|__GFP_NOWARN);
 547                 atomic_set(&space->i_mmap_writable, 0);
 548                 space->a_ops = &swap_aops;
 549                 /* swap cache doesn't use writeback related tags */
 550                 mapping_set_no_writeback_tags(space);
 551                 spin_lock_init(&space->tree_lock);
 552         }
 553         nr_swapper_spaces[type] = nr;
 554         rcu_assign_pointer(swapper_spaces[type], spaces);
 555
 556         return 0;
 557 }
 558
 559 void exit_swap_address_space(unsigned int type)
 560 {
 561         struct address_space *spaces;
 562
 563         spaces = swapper_spaces[type];
 564         nr_swapper_spaces[type] = 0;
 565         rcu_assign_pointer(swapper_spaces[type], NULL);
 566         synchronize_rcu();
 567         kvfree(spaces);
 568 }