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/page_cgroup.h>
  21
  22 #include <asm/pgtable.h>
  23
  24 /*
  25  * swapper_space is a fiction, retained to simplify the path through
  26  * vmscan's shrink_page_list.
  27  */
  28 static const struct address_space_operations swap_aops = {
  29         .writepage      = swap_writepage,
  30         .set_page_dirty = swap_set_page_dirty,
  31         .migratepage    = migrate_page,
  32 };
  33
  34 static struct backing_dev_info swap_backing_dev_info = {
  35         .name           = "swap",
  36         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  37 };
  38
  39 struct address_space swapper_spaces[MAX_SWAPFILES] = {
  40         [0 ... MAX_SWAPFILES - 1] = {
  41                 .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  42                 .a_ops          = &swap_aops,
  43                 .backing_dev_info = &swap_backing_dev_info,
  44         }
  45 };
  46
  47 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  48
  49 static struct {
  50         unsigned long add_total;
  51         unsigned long del_total;
  52         unsigned long find_success;
  53         unsigned long find_total;
  54 } swap_cache_info;
  55
  56 unsigned long total_swapcache_pages(void)
  57 {
  58         int i;
  59         unsigned long ret = 0;
  60
  61         for (i = 0; i < MAX_SWAPFILES; i++)
  62                 ret += swapper_spaces[i].nrpages;
  63         return ret;
  64 }
  65
  66 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  67
  68 void show_swap_cache_info(void)
  69 {
  70         printk("%lu pages in swap cache\n", total_swapcache_pages());
  71         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  72                 swap_cache_info.add_total, swap_cache_info.del_total,
  73                 swap_cache_info.find_success, swap_cache_info.find_total);
  74         printk("Free swap  = %ldkB\n",
  75                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
  76         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  77 }
  78
  79 /*
  80  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  81  * but sets SwapCache flag and private instead of mapping and index.
  82  */
  83 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  84 {
  85         int error;
  86         struct address_space *address_space;
  87
  88         VM_BUG_ON_PAGE(!PageLocked(page), page);
  89         VM_BUG_ON_PAGE(PageSwapCache(page), page);
  90         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
  91
  92         page_cache_get(page);
  93         SetPageSwapCache(page);
  94         set_page_private(page, entry.val);
  95
  96         address_space = swap_address_space(entry);
  97         spin_lock_irq(&address_space->tree_lock);
  98         error = radix_tree_insert(&address_space->page_tree,
  99                                         entry.val, page);
 100         if (likely(!error)) {
 101                 address_space->nrpages++;
 102                 __inc_zone_page_state(page, NR_FILE_PAGES);
 103                 INC_CACHE_INFO(add_total);
 104         }
 105         spin_unlock_irq(&address_space->tree_lock);
 106
 107         if (unlikely(error)) {
 108                 /*
 109                  * Only the context which have set SWAP_HAS_CACHE flag
 110                  * would call add_to_swap_cache().
 111                  * So add_to_swap_cache() doesn't returns -EEXIST.
 112                  */
 113                 VM_BUG_ON(error == -EEXIST);
 114                 set_page_private(page, 0UL);
 115                 ClearPageSwapCache(page);
 116                 page_cache_release(page);
 117         }
 118
 119         return error;
 120 }
 121
 122
 123 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 124 {
 125         int error;
 126
 127         error = radix_tree_maybe_preload(gfp_mask);
 128         if (!error) {
 129                 error = __add_to_swap_cache(page, entry);
 130                 radix_tree_preload_end();
 131         }
 132         return error;
 133 }
 134
 135 /*
 136  * This must be called only on pages that have
 137  * been verified to be in the swap cache.
 138  */
 139 void __delete_from_swap_cache(struct page *page)
 140 {
 141         swp_entry_t entry;
 142         struct address_space *address_space;
 143
 144         VM_BUG_ON_PAGE(!PageLocked(page), page);
 145         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 146         VM_BUG_ON_PAGE(PageWriteback(page), page);
 147
 148         entry.val = page_private(page);
 149         address_space = swap_address_space(entry);
 150         radix_tree_delete(&address_space->page_tree, page_private(page));
 151         set_page_private(page, 0);
 152         ClearPageSwapCache(page);
 153         address_space->nrpages--;
 154         __dec_zone_page_state(page, NR_FILE_PAGES);
 155         INC_CACHE_INFO(del_total);
 156 }
 157
 158 /**
 159  * add_to_swap - allocate swap space for a page
 160  * @page: page we want to move to swap
 161  *
 162  * Allocate swap space for the page and add the page to the
 163  * swap cache.  Caller needs to hold the page lock.
 164  */
 165 int add_to_swap(struct page *page, struct list_head *list)
 166 {
 167         swp_entry_t entry;
 168         int err;
 169
 170         VM_BUG_ON_PAGE(!PageLocked(page), page);
 171         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 172
 173         entry = get_swap_page();
 174         if (!entry.val)
 175                 return 0;
 176
 177         if (unlikely(PageTransHuge(page)))
 178                 if (unlikely(split_huge_page_to_list(page, list))) {
 179                         swapcache_free(entry, NULL);
 180                         return 0;
 181                 }
 182
 183         /*
 184          * Radix-tree node allocations from PF_MEMALLOC contexts could
 185          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 186          * stops emergency reserves from being allocated.
 187          *
 188          * TODO: this could cause a theoretical memory reclaim
 189          * deadlock in the swap out path.
 190          */
 191         /*
 192          * Add it to the swap cache and mark it dirty
 193          */
 194         err = add_to_swap_cache(page, entry,
 195                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 196
 197         if (!err) {     /* Success */
 198                 SetPageDirty(page);
 199                 return 1;
 200         } else {        /* -ENOMEM radix-tree allocation failure */
 201                 /*
 202                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 203                  * clear SWAP_HAS_CACHE flag.
 204                  */
 205                 swapcache_free(entry, NULL);
 206                 return 0;
 207         }
 208 }
 209
 210 /*
 211  * This must be called only on pages that have
 212  * been verified to be in the swap cache and locked.
 213  * It will never put the page into the free list,
 214  * the caller has a reference on the page.
 215  */
 216 void delete_from_swap_cache(struct page *page)
 217 {
 218         swp_entry_t entry;
 219         struct address_space *address_space;
 220
 221         entry.val = page_private(page);
 222
 223         address_space = swap_address_space(entry);
 224         spin_lock_irq(&address_space->tree_lock);
 225         __delete_from_swap_cache(page);
 226         spin_unlock_irq(&address_space->tree_lock);
 227
 228         swapcache_free(entry, page);
 229         page_cache_release(page);
 230 }
 231
 232 /*
 233  * If we are the only user, then try to free up the swap cache.
 234  *
 235  * Its ok to check for PageSwapCache without the page lock
 236  * here because we are going to recheck again inside
 237  * try_to_free_swap() _with_ the lock.
 238  *                                      - Marcelo
 239  */
 240 static inline void free_swap_cache(struct page *page)
 241 {
 242         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 243                 try_to_free_swap(page);
 244                 unlock_page(page);
 245         }
 246 }
 247
 248 /*
 249  * Perform a free_page(), also freeing any swap cache associated with
 250  * this page if it is the last user of the page.
 251  */
 252 void free_page_and_swap_cache(struct page *page)
 253 {
 254         free_swap_cache(page);
 255         page_cache_release(page);
 256 }
 257
 258 /*
 259  * Passed an array of pages, drop them all from swapcache and then release
 260  * them.  They are removed from the LRU and freed if this is their last use.
 261  */
 262 void free_pages_and_swap_cache(struct page **pages, int nr)
 263 {
 264         struct page **pagep = pages;
 265
 266         lru_add_drain();
 267         while (nr) {
 268                 int todo = min(nr, PAGEVEC_SIZE);
 269                 int i;
 270
 271                 for (i = 0; i < todo; i++)
 272                         free_swap_cache(pagep[i]);
 273                 release_pages(pagep, todo, 0);
 274                 pagep += todo;
 275                 nr -= todo;
 276         }
 277 }
 278
 279 /*
 280  * Lookup a swap entry in the swap cache. A found page will be returned
 281  * unlocked and with its refcount incremented - we rely on the kernel
 282  * lock getting page table operations atomic even if we drop the page
 283  * lock before returning.
 284  */
 285 struct page * lookup_swap_cache(swp_entry_t entry)
 286 {
 287         struct page *page;
 288
 289         page = find_get_page(swap_address_space(entry), entry.val);
 290
 291         if (page) {
 292                 INC_CACHE_INFO(find_success);
 293                 if (TestClearPageReadahead(page))
 294                         atomic_inc(&swapin_readahead_hits);
 295         }
 296
 297         INC_CACHE_INFO(find_total);
 298         return page;
 299 }
 300
 301 /*
 302  * Locate a page of swap in physical memory, reserving swap cache space
 303  * and reading the disk if it is not already cached.
 304  * A failure return means that either the page allocation failed or that
 305  * the swap entry is no longer in use.
 306  */
 307 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 308                         struct vm_area_struct *vma, unsigned long addr)
 309 {
 310         struct page *found_page, *new_page = NULL;
 311         int err;
 312
 313         do {
 314                 /*
 315                  * First check the swap cache.  Since this is normally
 316                  * called after lookup_swap_cache() failed, re-calling
 317                  * that would confuse statistics.
 318                  */
 319                 found_page = find_get_page(swap_address_space(entry),
 320                                         entry.val);
 321                 if (found_page)
 322                         break;
 323
 324                 /*
 325                  * Get a new page to read into from swap.
 326                  */
 327                 if (!new_page) {
 328                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 329                         if (!new_page)
 330                                 break;          /* Out of memory */
 331                 }
 332
 333                 /*
 334                  * call radix_tree_preload() while we can wait.
 335                  */
 336                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 337                 if (err)
 338                         break;
 339
 340                 /*
 341                  * Swap entry may have been freed since our caller observed it.
 342                  */
 343                 err = swapcache_prepare(entry);
 344                 if (err == -EEXIST) {
 345                         radix_tree_preload_end();
 346                         /*
 347                          * We might race against get_swap_page() and stumble
 348                          * across a SWAP_HAS_CACHE swap_map entry whose page
 349                          * has not been brought into the swapcache yet, while
 350                          * the other end is scheduled away waiting on discard
 351                          * I/O completion at scan_swap_map().
 352                          *
 353                          * In order to avoid turning this transitory state
 354                          * into a permanent loop around this -EEXIST case
 355                          * if !CONFIG_PREEMPT and the I/O completion happens
 356                          * to be waiting on the CPU waitqueue where we are now
 357                          * busy looping, we just conditionally invoke the
 358                          * scheduler here, if there are some more important
 359                          * tasks to run.
 360                          */
 361                         cond_resched();
 362                         continue;
 363                 }
 364                 if (err) {              /* swp entry is obsolete ? */
 365                         radix_tree_preload_end();
 366                         break;
 367                 }
 368
 369                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 370                 __set_page_locked(new_page);
 371                 SetPageSwapBacked(new_page);
 372                 err = __add_to_swap_cache(new_page, entry);
 373                 if (likely(!err)) {
 374                         radix_tree_preload_end();
 375                         /*
 376                          * Initiate read into locked page and return.
 377                          */
 378                         lru_cache_add_anon(new_page);
 379                         swap_readpage(new_page);
 380                         return new_page;
 381                 }
 382                 radix_tree_preload_end();
 383                 ClearPageSwapBacked(new_page);
 384                 __clear_page_locked(new_page);
 385                 /*
 386                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 387                  * clear SWAP_HAS_CACHE flag.
 388                  */
 389                 swapcache_free(entry, NULL);
 390         } while (err != -ENOMEM);
 391
 392         if (new_page)
 393                 page_cache_release(new_page);
 394         return found_page;
 395 }
 396
 397 static unsigned long swapin_nr_pages(unsigned long offset)
 398 {
 399         static unsigned long prev_offset;
 400         unsigned int pages, max_pages, last_ra;
 401         static atomic_t last_readahead_pages;
 402
 403         max_pages = 1 << ACCESS_ONCE(page_cluster);
 404         if (max_pages <= 1)
 405                 return 1;
 406
 407         /*
 408          * This heuristic has been found to work well on both sequential and
 409          * random loads, swapping to hard disk or to SSD: please don't ask
 410          * what the "+ 2" means, it just happens to work well, that's all.
 411          */
 412         pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
 413         if (pages == 2) {
 414                 /*
 415                  * We can have no readahead hits to judge by: but must not get
 416                  * stuck here forever, so check for an adjacent offset instead
 417                  * (and don't even bother to check whether swap type is same).
 418                  */
 419                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 420                         pages = 1;
 421                 prev_offset = offset;
 422         } else {
 423                 unsigned int roundup = 4;
 424                 while (roundup < pages)
 425                         roundup <<= 1;
 426                 pages = roundup;
 427         }
 428
 429         if (pages > max_pages)
 430                 pages = max_pages;
 431
 432         /* Don't shrink readahead too fast */
 433         last_ra = atomic_read(&last_readahead_pages) / 2;
 434         if (pages < last_ra)
 435                 pages = last_ra;
 436         atomic_set(&last_readahead_pages, pages);
 437
 438         return pages;
 439 }
 440
 441 /**
 442  * swapin_readahead - swap in pages in hope we need them soon
 443  * @entry: swap entry of this memory
 444  * @gfp_mask: memory allocation flags
 445  * @vma: user vma this address belongs to
 446  * @addr: target address for mempolicy
 447  *
 448  * Returns the struct page for entry and addr, after queueing swapin.
 449  *
 450  * Primitive swap readahead code. We simply read an aligned block of
 451  * (1 << page_cluster) entries in the swap area. This method is chosen
 452  * because it doesn't cost us any seek time.  We also make sure to queue
 453  * the 'original' request together with the readahead ones...
 454  *
 455  * This has been extended to use the NUMA policies from the mm triggering
 456  * the readahead.
 457  *
 458  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 459  */
 460 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 461                         struct vm_area_struct *vma, unsigned long addr)
 462 {
 463         struct page *page;
 464         unsigned long entry_offset = swp_offset(entry);
 465         unsigned long offset = entry_offset;
 466         unsigned long start_offset, end_offset;
 467         unsigned long mask;
 468         struct blk_plug plug;
 469
 470         mask = swapin_nr_pages(offset) - 1;
 471         if (!mask)
 472                 goto skip;
 473
 474         /* Read a page_cluster sized and aligned cluster around offset. */
 475         start_offset = offset & ~mask;
 476         end_offset = offset | mask;
 477         if (!start_offset)      /* First page is swap header. */
 478                 start_offset++;
 479
 480         blk_start_plug(&plug);
 481         for (offset = start_offset; offset <= end_offset ; offset++) {
 482                 /* Ok, do the async read-ahead now */
 483                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 484                                                 gfp_mask, vma, addr);
 485                 if (!page)
 486                         continue;
 487                 if (offset != entry_offset)
 488                         SetPageReadahead(page);
 489                 page_cache_release(page);
 490         }
 491         blk_finish_plug(&plug);
 492
 493         lru_add_drain();        /* Push any new pages onto the LRU now */
 494 skip:
 495         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 496 }