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 void show_swap_cache_info(void)
  67 {
  68         printk("%lu pages in swap cache\n", total_swapcache_pages());
  69         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  70                 swap_cache_info.add_total, swap_cache_info.del_total,
  71                 swap_cache_info.find_success, swap_cache_info.find_total);
  72         printk("Free swap  = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  73         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  74 }
  75
  76 /*
  77  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  78  * but sets SwapCache flag and private instead of mapping and index.
  79  */
  80 static int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  81 {
  82         int error;
  83         struct address_space *address_space;
  84
  85         VM_BUG_ON(!PageLocked(page));
  86         VM_BUG_ON(PageSwapCache(page));
  87         VM_BUG_ON(!PageSwapBacked(page));
  88
  89         page_cache_get(page);
  90         SetPageSwapCache(page);
  91         set_page_private(page, entry.val);
  92
  93         address_space = swap_address_space(entry);
  94         spin_lock_irq(&address_space->tree_lock);
  95         error = radix_tree_insert(&address_space->page_tree,
  96                                         entry.val, page);
  97         if (likely(!error)) {
  98                 address_space->nrpages++;
  99                 __inc_zone_page_state(page, NR_FILE_PAGES);
 100                 INC_CACHE_INFO(add_total);
 101         }
 102         spin_unlock_irq(&address_space->tree_lock);
 103
 104         if (unlikely(error)) {
 105                 /*
 106                  * Only the context which have set SWAP_HAS_CACHE flag
 107                  * would call add_to_swap_cache().
 108                  * So add_to_swap_cache() doesn't returns -EEXIST.
 109                  */
 110                 VM_BUG_ON(error == -EEXIST);
 111                 set_page_private(page, 0UL);
 112                 ClearPageSwapCache(page);
 113                 page_cache_release(page);
 114         }
 115
 116         return error;
 117 }
 118
 119
 120 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 121 {
 122         int error;
 123
 124         error = radix_tree_preload(gfp_mask);
 125         if (!error) {
 126                 error = __add_to_swap_cache(page, entry);
 127                 radix_tree_preload_end();
 128         }
 129         return error;
 130 }
 131
 132 /*
 133  * This must be called only on pages that have
 134  * been verified to be in the swap cache.
 135  */
 136 void __delete_from_swap_cache(struct page *page)
 137 {
 138         swp_entry_t entry;
 139         struct address_space *address_space;
 140
 141         VM_BUG_ON(!PageLocked(page));
 142         VM_BUG_ON(!PageSwapCache(page));
 143         VM_BUG_ON(PageWriteback(page));
 144
 145         entry.val = page_private(page);
 146         address_space = swap_address_space(entry);
 147         radix_tree_delete(&address_space->page_tree, page_private(page));
 148         set_page_private(page, 0);
 149         ClearPageSwapCache(page);
 150         address_space->nrpages--;
 151         __dec_zone_page_state(page, NR_FILE_PAGES);
 152         INC_CACHE_INFO(del_total);
 153 }
 154
 155 /**
 156  * add_to_swap - allocate swap space for a page
 157  * @page: page we want to move to swap
 158  *
 159  * Allocate swap space for the page and add the page to the
 160  * swap cache.  Caller needs to hold the page lock.
 161  */
 162 int add_to_swap(struct page *page)
 163 {
 164         swp_entry_t entry;
 165         int err;
 166
 167         VM_BUG_ON(!PageLocked(page));
 168         VM_BUG_ON(!PageUptodate(page));
 169
 170         entry = get_swap_page();
 171         if (!entry.val)
 172                 return 0;
 173
 174         if (unlikely(PageTransHuge(page)))
 175                 if (unlikely(split_huge_page(page))) {
 176                         swapcache_free(entry, NULL);
 177                         return 0;
 178                 }
 179
 180         /*
 181          * Radix-tree node allocations from PF_MEMALLOC contexts could
 182          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 183          * stops emergency reserves from being allocated.
 184          *
 185          * TODO: this could cause a theoretical memory reclaim
 186          * deadlock in the swap out path.
 187          */
 188         /*
 189          * Add it to the swap cache and mark it dirty
 190          */
 191         err = add_to_swap_cache(page, entry,
 192                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 193
 194         if (!err) {     /* Success */
 195                 SetPageDirty(page);
 196                 return 1;
 197         } else {        /* -ENOMEM radix-tree allocation failure */
 198                 /*
 199                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 200                  * clear SWAP_HAS_CACHE flag.
 201                  */
 202                 swapcache_free(entry, NULL);
 203                 return 0;
 204         }
 205 }
 206
 207 /*
 208  * This must be called only on pages that have
 209  * been verified to be in the swap cache and locked.
 210  * It will never put the page into the free list,
 211  * the caller has a reference on the page.
 212  */
 213 void delete_from_swap_cache(struct page *page)
 214 {
 215         swp_entry_t entry;
 216         struct address_space *address_space;
 217
 218         entry.val = page_private(page);
 219
 220         address_space = swap_address_space(entry);
 221         spin_lock_irq(&address_space->tree_lock);
 222         __delete_from_swap_cache(page);
 223         spin_unlock_irq(&address_space->tree_lock);
 224
 225         swapcache_free(entry, page);
 226         page_cache_release(page);
 227 }
 228
 229 /*
 230  * If we are the only user, then try to free up the swap cache.
 231  *
 232  * Its ok to check for PageSwapCache without the page lock
 233  * here because we are going to recheck again inside
 234  * try_to_free_swap() _with_ the lock.
 235  *                                      - Marcelo
 236  */
 237 static inline void free_swap_cache(struct page *page)
 238 {
 239         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 240                 try_to_free_swap(page);
 241                 unlock_page(page);
 242         }
 243 }
 244
 245 /*
 246  * Perform a free_page(), also freeing any swap cache associated with
 247  * this page if it is the last user of the page.
 248  */
 249 void free_page_and_swap_cache(struct page *page)
 250 {
 251         free_swap_cache(page);
 252         page_cache_release(page);
 253 }
 254
 255 /*
 256  * Passed an array of pages, drop them all from swapcache and then release
 257  * them.  They are removed from the LRU and freed if this is their last use.
 258  */
 259 void free_pages_and_swap_cache(struct page **pages, int nr)
 260 {
 261         struct page **pagep = pages;
 262
 263         lru_add_drain();
 264         while (nr) {
 265                 int todo = min(nr, PAGEVEC_SIZE);
 266                 int i;
 267
 268                 for (i = 0; i < todo; i++)
 269                         free_swap_cache(pagep[i]);
 270                 release_pages(pagep, todo, 0);
 271                 pagep += todo;
 272                 nr -= todo;
 273         }
 274 }
 275
 276 /*
 277  * Lookup a swap entry in the swap cache. A found page will be returned
 278  * unlocked and with its refcount incremented - we rely on the kernel
 279  * lock getting page table operations atomic even if we drop the page
 280  * lock before returning.
 281  */
 282 struct page * lookup_swap_cache(swp_entry_t entry)
 283 {
 284         struct page *page;
 285
 286         page = find_get_page(swap_address_space(entry), entry.val);
 287
 288         if (page)
 289                 INC_CACHE_INFO(find_success);
 290
 291         INC_CACHE_INFO(find_total);
 292         return page;
 293 }
 294
 295 /*
 296  * Locate a page of swap in physical memory, reserving swap cache space
 297  * and reading the disk if it is not already cached.
 298  * A failure return means that either the page allocation failed or that
 299  * the swap entry is no longer in use.
 300  */
 301 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 302                         struct vm_area_struct *vma, unsigned long addr)
 303 {
 304         struct page *found_page, *new_page = NULL;
 305         int err;
 306
 307         do {
 308                 /*
 309                  * First check the swap cache.  Since this is normally
 310                  * called after lookup_swap_cache() failed, re-calling
 311                  * that would confuse statistics.
 312                  */
 313                 found_page = find_get_page(swap_address_space(entry),
 314                                         entry.val);
 315                 if (found_page)
 316                         break;
 317
 318                 /*
 319                  * Get a new page to read into from swap.
 320                  */
 321                 if (!new_page) {
 322                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 323                         if (!new_page)
 324                                 break;          /* Out of memory */
 325                 }
 326
 327                 /*
 328                  * call radix_tree_preload() while we can wait.
 329                  */
 330                 err = radix_tree_preload(gfp_mask & GFP_KERNEL);
 331                 if (err)
 332                         break;
 333
 334                 /*
 335                  * Swap entry may have been freed since our caller observed it.
 336                  */
 337                 err = swapcache_prepare(entry);
 338                 if (err == -EEXIST) {   /* seems racy */
 339                         radix_tree_preload_end();
 340                         continue;
 341                 }
 342                 if (err) {              /* swp entry is obsolete ? */
 343                         radix_tree_preload_end();
 344                         break;
 345                 }
 346
 347                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 348                 __set_page_locked(new_page);
 349                 SetPageSwapBacked(new_page);
 350                 err = __add_to_swap_cache(new_page, entry);
 351                 if (likely(!err)) {
 352                         radix_tree_preload_end();
 353                         /*
 354                          * Initiate read into locked page and return.
 355                          */
 356                         lru_cache_add_anon(new_page);
 357                         swap_readpage(new_page);
 358                         return new_page;
 359                 }
 360                 radix_tree_preload_end();
 361                 ClearPageSwapBacked(new_page);
 362                 __clear_page_locked(new_page);
 363                 /*
 364                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 365                  * clear SWAP_HAS_CACHE flag.
 366                  */
 367                 swapcache_free(entry, NULL);
 368         } while (err != -ENOMEM);
 369
 370         if (new_page)
 371                 page_cache_release(new_page);
 372         return found_page;
 373 }
 374
 375 /**
 376  * swapin_readahead - swap in pages in hope we need them soon
 377  * @entry: swap entry of this memory
 378  * @gfp_mask: memory allocation flags
 379  * @vma: user vma this address belongs to
 380  * @addr: target address for mempolicy
 381  *
 382  * Returns the struct page for entry and addr, after queueing swapin.
 383  *
 384  * Primitive swap readahead code. We simply read an aligned block of
 385  * (1 << page_cluster) entries in the swap area. This method is chosen
 386  * because it doesn't cost us any seek time.  We also make sure to queue
 387  * the 'original' request together with the readahead ones...
 388  *
 389  * This has been extended to use the NUMA policies from the mm triggering
 390  * the readahead.
 391  *
 392  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 393  */
 394 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 395                         struct vm_area_struct *vma, unsigned long addr)
 396 {
 397         struct page *page;
 398         unsigned long offset = swp_offset(entry);
 399         unsigned long start_offset, end_offset;
 400         unsigned long mask = (1UL << page_cluster) - 1;
 401         struct blk_plug plug;
 402
 403         /* Read a page_cluster sized and aligned cluster around offset. */
 404         start_offset = offset & ~mask;
 405         end_offset = offset | mask;
 406         if (!start_offset)      /* First page is swap header. */
 407                 start_offset++;
 408
 409         blk_start_plug(&plug);
 410         for (offset = start_offset; offset <= end_offset ; offset++) {
 411                 /* Ok, do the async read-ahead now */
 412                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 413                                                 gfp_mask, vma, addr);
 414                 if (!page)
 415                         continue;
 416                 page_cache_release(page);
 417         }
 418         blk_finish_plug(&plug);
 419
 420         lru_add_drain();        /* Push any new pages onto the LRU now */
 421         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 422 }