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
  23 #include <asm/pgtable.h>
  24
  25 /*
  26  * swapper_space is a fiction, retained to simplify the path through
  27  * vmscan's shrink_page_list.
  28  */
  29 static const struct address_space_operations swap_aops = {
  30         .writepage      = swap_writepage,
  31         .set_page_dirty = swap_set_page_dirty,
  32 #ifdef CONFIG_MIGRATION
  33         .migratepage    = migrate_page,
  34 #endif
  35 };
  36
  37 struct address_space *swapper_spaces[MAX_SWAPFILES];
  38 static unsigned int nr_swapper_spaces[MAX_SWAPFILES];
  39
  40 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  41
  42 static struct {
  43         unsigned long add_total;
  44         unsigned long del_total;
  45         unsigned long find_success;
  46         unsigned long find_total;
  47 } swap_cache_info;
  48
  49 unsigned long total_swapcache_pages(void)
  50 {
  51         unsigned int i, j, nr;
  52         unsigned long ret = 0;
  53         struct address_space *spaces;
  54
  55         rcu_read_lock();
  56         for (i = 0; i < MAX_SWAPFILES; i++) {
  57                 /*
  58                  * The corresponding entries in nr_swapper_spaces and
  59                  * swapper_spaces will be reused only after at least
  60                  * one grace period.  So it is impossible for them
  61                  * belongs to different usage.
  62                  */
  63                 nr = nr_swapper_spaces[i];
  64                 spaces = rcu_dereference(swapper_spaces[i]);
  65                 if (!nr || !spaces)
  66                         continue;
  67                 for (j = 0; j < nr; j++)
  68                         ret += spaces[j].nrpages;
  69         }
  70         rcu_read_unlock();
  71         return ret;
  72 }
  73
  74 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  75
  76 void show_swap_cache_info(void)
  77 {
  78         printk("%lu pages in swap cache\n", total_swapcache_pages());
  79         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  80                 swap_cache_info.add_total, swap_cache_info.del_total,
  81                 swap_cache_info.find_success, swap_cache_info.find_total);
  82         printk("Free swap  = %ldkB\n",
  83                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
  84         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  85 }
  86
  87 /*
  88  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  89  * but sets SwapCache flag and private instead of mapping and index.
  90  */
  91 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  92 {
  93         int error;
  94         struct address_space *address_space;
  95
  96         VM_BUG_ON_PAGE(!PageLocked(page), page);
  97         VM_BUG_ON_PAGE(PageSwapCache(page), page);
  98         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
  99
 100         get_page(page);
 101         SetPageSwapCache(page);
 102         set_page_private(page, entry.val);
 103
 104         address_space = swap_address_space(entry);
 105         spin_lock_irq(&address_space->tree_lock);
 106         error = radix_tree_insert(&address_space->page_tree,
 107                                   swp_offset(entry), page);
 108         if (likely(!error)) {
 109                 address_space->nrpages++;
 110                 __inc_node_page_state(page, NR_FILE_PAGES);
 111                 INC_CACHE_INFO(add_total);
 112         }
 113         spin_unlock_irq(&address_space->tree_lock);
 114
 115         if (unlikely(error)) {
 116                 /*
 117                  * Only the context which have set SWAP_HAS_CACHE flag
 118                  * would call add_to_swap_cache().
 119                  * So add_to_swap_cache() doesn't returns -EEXIST.
 120                  */
 121                 VM_BUG_ON(error == -EEXIST);
 122                 set_page_private(page, 0UL);
 123                 ClearPageSwapCache(page);
 124                 put_page(page);
 125         }
 126
 127         return error;
 128 }
 129
 130
 131 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 132 {
 133         int error;
 134
 135         error = radix_tree_maybe_preload(gfp_mask);
 136         if (!error) {
 137                 error = __add_to_swap_cache(page, entry);
 138                 radix_tree_preload_end();
 139         }
 140         return error;
 141 }
 142
 143 /*
 144  * This must be called only on pages that have
 145  * been verified to be in the swap cache.
 146  */
 147 void __delete_from_swap_cache(struct page *page)
 148 {
 149         swp_entry_t entry;
 150         struct address_space *address_space;
 151
 152         VM_BUG_ON_PAGE(!PageLocked(page), page);
 153         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 154         VM_BUG_ON_PAGE(PageWriteback(page), page);
 155
 156         entry.val = page_private(page);
 157         address_space = swap_address_space(entry);
 158         radix_tree_delete(&address_space->page_tree, swp_offset(entry));
 159         set_page_private(page, 0);
 160         ClearPageSwapCache(page);
 161         address_space->nrpages--;
 162         __dec_node_page_state(page, NR_FILE_PAGES);
 163         INC_CACHE_INFO(del_total);
 164 }
 165
 166 /**
 167  * add_to_swap - allocate swap space for a page
 168  * @page: page we want to move to swap
 169  *
 170  * Allocate swap space for the page and add the page to the
 171  * swap cache.  Caller needs to hold the page lock.
 172  */
 173 int add_to_swap(struct page *page, struct list_head *list)
 174 {
 175         swp_entry_t entry;
 176         int err;
 177
 178         VM_BUG_ON_PAGE(!PageLocked(page), page);
 179         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 180
 181         entry = get_swap_page();
 182         if (!entry.val)
 183                 return 0;
 184
 185         if (mem_cgroup_try_charge_swap(page, entry)) {
 186                 swapcache_free(entry);
 187                 return 0;
 188         }
 189
 190         if (unlikely(PageTransHuge(page)))
 191                 if (unlikely(split_huge_page_to_list(page, list))) {
 192                         swapcache_free(entry);
 193                         return 0;
 194                 }
 195
 196         /*
 197          * Radix-tree node allocations from PF_MEMALLOC contexts could
 198          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 199          * stops emergency reserves from being allocated.
 200          *
 201          * TODO: this could cause a theoretical memory reclaim
 202          * deadlock in the swap out path.
 203          */
 204         /*
 205          * Add it to the swap cache.
 206          */
 207         err = add_to_swap_cache(page, entry,
 208                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 209
 210         if (!err) {
 211                 return 1;
 212         } else {        /* -ENOMEM radix-tree allocation failure */
 213                 /*
 214                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 215                  * clear SWAP_HAS_CACHE flag.
 216                  */
 217                 swapcache_free(entry);
 218                 return 0;
 219         }
 220 }
 221
 222 /*
 223  * This must be called only on pages that have
 224  * been verified to be in the swap cache and locked.
 225  * It will never put the page into the free list,
 226  * the caller has a reference on the page.
 227  */
 228 void delete_from_swap_cache(struct page *page)
 229 {
 230         swp_entry_t entry;
 231         struct address_space *address_space;
 232
 233         entry.val = page_private(page);
 234
 235         address_space = swap_address_space(entry);
 236         spin_lock_irq(&address_space->tree_lock);
 237         __delete_from_swap_cache(page);
 238         spin_unlock_irq(&address_space->tree_lock);
 239
 240         swapcache_free(entry);
 241         put_page(page);
 242 }
 243
 244 /*
 245  * If we are the only user, then try to free up the swap cache.
 246  *
 247  * Its ok to check for PageSwapCache without the page lock
 248  * here because we are going to recheck again inside
 249  * try_to_free_swap() _with_ the lock.
 250  *                                      - Marcelo
 251  */
 252 static inline void free_swap_cache(struct page *page)
 253 {
 254         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 255                 try_to_free_swap(page);
 256                 unlock_page(page);
 257         }
 258 }
 259
 260 /*
 261  * Perform a free_page(), also freeing any swap cache associated with
 262  * this page if it is the last user of the page.
 263  */
 264 void free_page_and_swap_cache(struct page *page)
 265 {
 266         free_swap_cache(page);
 267         if (!is_huge_zero_page(page))
 268                 put_page(page);
 269 }
 270
 271 /*
 272  * Passed an array of pages, drop them all from swapcache and then release
 273  * them.  They are removed from the LRU and freed if this is their last use.
 274  */
 275 void free_pages_and_swap_cache(struct page **pages, int nr)
 276 {
 277         struct page **pagep = pages;
 278         int i;
 279
 280         lru_add_drain();
 281         for (i = 0; i < nr; i++)
 282                 free_swap_cache(pagep[i]);
 283         release_pages(pagep, nr, false);
 284 }
 285
 286 /*
 287  * Lookup a swap entry in the swap cache. A found page will be returned
 288  * unlocked and with its refcount incremented - we rely on the kernel
 289  * lock getting page table operations atomic even if we drop the page
 290  * lock before returning.
 291  */
 292 struct page * lookup_swap_cache(swp_entry_t entry)
 293 {
 294         struct page *page;
 295
 296         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 297
 298         if (page) {
 299                 INC_CACHE_INFO(find_success);
 300                 if (TestClearPageReadahead(page))
 301                         atomic_inc(&swapin_readahead_hits);
 302         }
 303
 304         INC_CACHE_INFO(find_total);
 305         return page;
 306 }
 307
 308 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 309                         struct vm_area_struct *vma, unsigned long addr,
 310                         bool *new_page_allocated)
 311 {
 312         struct page *found_page, *new_page = NULL;
 313         struct address_space *swapper_space = swap_address_space(entry);
 314         int err;
 315         *new_page_allocated = false;
 316
 317         do {
 318                 /*
 319                  * First check the swap cache.  Since this is normally
 320                  * called after lookup_swap_cache() failed, re-calling
 321                  * that would confuse statistics.
 322                  */
 323                 found_page = find_get_page(swapper_space, swp_offset(entry));
 324                 if (found_page)
 325                         break;
 326
 327                 /*
 328                  * Just skip read ahead for unused swap slot.
 329                  * During swap_off when swap_slot_cache is disabled,
 330                  * we have to handle the race between putting
 331                  * swap entry in swap cache and marking swap slot
 332                  * as SWAP_HAS_CACHE.  That's done in later part of code or
 333                  * else swap_off will be aborted if we return NULL.
 334                  */
 335                 if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
 336                         break;
 337
 338                 /*
 339                  * Get a new page to read into from swap.
 340                  */
 341                 if (!new_page) {
 342                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 343                         if (!new_page)
 344                                 break;          /* Out of memory */
 345                 }
 346
 347                 /*
 348                  * call radix_tree_preload() while we can wait.
 349                  */
 350                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 351                 if (err)
 352                         break;
 353
 354                 /*
 355                  * Swap entry may have been freed since our caller observed it.
 356                  */
 357                 err = swapcache_prepare(entry);
 358                 if (err == -EEXIST) {
 359                         radix_tree_preload_end();
 360                         /*
 361                          * We might race against get_swap_page() and stumble
 362                          * across a SWAP_HAS_CACHE swap_map entry whose page
 363                          * has not been brought into the swapcache yet, while
 364                          * the other end is scheduled away waiting on discard
 365                          * I/O completion at scan_swap_map().
 366                          *
 367                          * In order to avoid turning this transitory state
 368                          * into a permanent loop around this -EEXIST case
 369                          * if !CONFIG_PREEMPT and the I/O completion happens
 370                          * to be waiting on the CPU waitqueue where we are now
 371                          * busy looping, we just conditionally invoke the
 372                          * scheduler here, if there are some more important
 373                          * tasks to run.
 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                 swapcache_free(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)
 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);
 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
 502         mask = swapin_nr_pages(offset) - 1;
 503         if (!mask)
 504                 goto skip;
 505
 506         /* Read a page_cluster sized and aligned cluster around offset. */
 507         start_offset = offset & ~mask;
 508         end_offset = offset | mask;
 509         if (!start_offset)      /* First page is swap header. */
 510                 start_offset++;
 511
 512         blk_start_plug(&plug);
 513         for (offset = start_offset; offset <= end_offset ; offset++) {
 514                 /* Ok, do the async read-ahead now */
 515                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 516                                                 gfp_mask, vma, addr);
 517                 if (!page)
 518                         continue;
 519                 if (offset != entry_offset)
 520                         SetPageReadahead(page);
 521                 put_page(page);
 522         }
 523         blk_finish_plug(&plug);
 524
 525         lru_add_drain();        /* Push any new pages onto the LRU now */
 526 skip:
 527         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 528 }
 529
 530 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
 531 {
 532         struct address_space *spaces, *space;
 533         unsigned int i, nr;
 534
 535         nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
 536         spaces = vzalloc(sizeof(struct address_space) * nr);
 537         if (!spaces)
 538                 return -ENOMEM;
 539         for (i = 0; i < nr; i++) {
 540                 space = spaces + i;
 541                 INIT_RADIX_TREE(&space->page_tree, GFP_ATOMIC|__GFP_NOWARN);
 542                 atomic_set(&space->i_mmap_writable, 0);
 543                 space->a_ops = &swap_aops;
 544                 /* swap cache doesn't use writeback related tags */
 545                 mapping_set_no_writeback_tags(space);
 546                 spin_lock_init(&space->tree_lock);
 547         }
 548         nr_swapper_spaces[type] = nr;
 549         rcu_assign_pointer(swapper_spaces[type], spaces);
 550
 551         return 0;
 552 }
 553
 554 void exit_swap_address_space(unsigned int type)
 555 {
 556         struct address_space *spaces;
 557
 558         spaces = swapper_spaces[type];
 559         nr_swapper_spaces[type] = 0;
 560         rcu_assign_pointer(swapper_spaces[type], NULL);
 561         synchronize_rcu();
 562         kvfree(spaces);
 563 }