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/module.h>
  10 #include <linux/mm.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/buffer_head.h>
  17 #include <linux/backing-dev.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, to make sync_page look nicer, and to allow
  27  * future use of radix_tree tags in the swap cache.
  28  */
  29 static const struct address_space_operations swap_aops = {
  30         .writepage      = swap_writepage,
  31         .sync_page      = block_sync_page,
  32         .set_page_dirty = __set_page_dirty_nobuffers,
  33         .migratepage    = migrate_page,
  34 };
  35
  36 static struct backing_dev_info swap_backing_dev_info = {
  37         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  38         .unplug_io_fn   = swap_unplug_io_fn,
  39 };
  40
  41 struct address_space swapper_space = {
  42         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  43         .tree_lock      = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock),
  44         .a_ops          = &swap_aops,
  45         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
  46         .backing_dev_info = &swap_backing_dev_info,
  47 };
  48
  49 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  50
  51 static struct {
  52         unsigned long add_total;
  53         unsigned long del_total;
  54         unsigned long find_success;
  55         unsigned long find_total;
  56 } swap_cache_info;
  57
  58 void show_swap_cache_info(void)
  59 {
  60         printk("%lu pages in swap cache\n", total_swapcache_pages);
  61         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  62                 swap_cache_info.add_total, swap_cache_info.del_total,
  63                 swap_cache_info.find_success, swap_cache_info.find_total);
  64         printk("Free swap  = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  65         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  66 }
  67
  68 /*
  69  * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  70  * but sets SwapCache flag and private instead of mapping and index.
  71  */
  72 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
  73 {
  74         int error;
  75
  76         VM_BUG_ON(!PageLocked(page));
  77         VM_BUG_ON(PageSwapCache(page));
  78         VM_BUG_ON(!PageSwapBacked(page));
  79
  80         error = radix_tree_preload(gfp_mask);
  81         if (!error) {
  82                 page_cache_get(page);
  83                 SetPageSwapCache(page);
  84                 set_page_private(page, entry.val);
  85
  86                 spin_lock_irq(&swapper_space.tree_lock);
  87                 error = radix_tree_insert(&swapper_space.page_tree,
  88                                                 entry.val, page);
  89                 if (likely(!error)) {
  90                         total_swapcache_pages++;
  91                         __inc_zone_page_state(page, NR_FILE_PAGES);
  92                         INC_CACHE_INFO(add_total);
  93                 }
  94                 spin_unlock_irq(&swapper_space.tree_lock);
  95                 radix_tree_preload_end();
  96
  97                 if (unlikely(error)) {
  98                         set_page_private(page, 0UL);
  99                         ClearPageSwapCache(page);
 100                         page_cache_release(page);
 101                 }
 102         }
 103         return error;
 104 }
 105
 106 /*
 107  * This must be called only on pages that have
 108  * been verified to be in the swap cache.
 109  */
 110 void __delete_from_swap_cache(struct page *page)
 111 {
 112         VM_BUG_ON(!PageLocked(page));
 113         VM_BUG_ON(!PageSwapCache(page));
 114         VM_BUG_ON(PageWriteback(page));
 115
 116         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 117         set_page_private(page, 0);
 118         ClearPageSwapCache(page);
 119         total_swapcache_pages--;
 120         __dec_zone_page_state(page, NR_FILE_PAGES);
 121         INC_CACHE_INFO(del_total);
 122 }
 123
 124 /**
 125  * add_to_swap - allocate swap space for a page
 126  * @page: page we want to move to swap
 127  * @gfp_mask: memory allocation flags
 128  *
 129  * Allocate swap space for the page and add the page to the
 130  * swap cache.  Caller needs to hold the page lock.
 131  */
 132 int add_to_swap(struct page *page)
 133 {
 134         swp_entry_t entry;
 135         int err;
 136
 137         VM_BUG_ON(!PageLocked(page));
 138         VM_BUG_ON(!PageUptodate(page));
 139
 140         for (;;) {
 141                 entry = get_swap_page();
 142                 if (!entry.val)
 143                         return 0;
 144
 145                 /*
 146                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 147                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 148                  * stops emergency reserves from being allocated.
 149                  *
 150                  * TODO: this could cause a theoretical memory reclaim
 151                  * deadlock in the swap out path.
 152                  */
 153                 /*
 154                  * Add it to the swap cache and mark it dirty
 155                  */
 156                 err = add_to_swap_cache(page, entry,
 157                                 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 158
 159                 switch (err) {
 160                 case 0:                         /* Success */
 161                         SetPageDirty(page);
 162                         return 1;
 163                 case -EEXIST:
 164                         /* Raced with "speculative" read_swap_cache_async */
 165                         swapcache_free(entry, NULL);
 166                         continue;
 167                 default:
 168                         /* -ENOMEM radix-tree allocation failure */
 169                         swapcache_free(entry, NULL);
 170                         return 0;
 171                 }
 172         }
 173 }
 174
 175 /*
 176  * This must be called only on pages that have
 177  * been verified to be in the swap cache and locked.
 178  * It will never put the page into the free list,
 179  * the caller has a reference on the page.
 180  */
 181 void delete_from_swap_cache(struct page *page)
 182 {
 183         swp_entry_t entry;
 184
 185         entry.val = page_private(page);
 186
 187         spin_lock_irq(&swapper_space.tree_lock);
 188         __delete_from_swap_cache(page);
 189         spin_unlock_irq(&swapper_space.tree_lock);
 190
 191         swapcache_free(entry, page);
 192         page_cache_release(page);
 193 }
 194
 195 /*
 196  * If we are the only user, then try to free up the swap cache.
 197  *
 198  * Its ok to check for PageSwapCache without the page lock
 199  * here because we are going to recheck again inside
 200  * try_to_free_swap() _with_ the lock.
 201  *                                      - Marcelo
 202  */
 203 static inline void free_swap_cache(struct page *page)
 204 {
 205         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 206                 try_to_free_swap(page);
 207                 unlock_page(page);
 208         }
 209 }
 210
 211 /*
 212  * Perform a free_page(), also freeing any swap cache associated with
 213  * this page if it is the last user of the page.
 214  */
 215 void free_page_and_swap_cache(struct page *page)
 216 {
 217         free_swap_cache(page);
 218         page_cache_release(page);
 219 }
 220
 221 /*
 222  * Passed an array of pages, drop them all from swapcache and then release
 223  * them.  They are removed from the LRU and freed if this is their last use.
 224  */
 225 void free_pages_and_swap_cache(struct page **pages, int nr)
 226 {
 227         struct page **pagep = pages;
 228
 229         lru_add_drain();
 230         while (nr) {
 231                 int todo = min(nr, PAGEVEC_SIZE);
 232                 int i;
 233
 234                 for (i = 0; i < todo; i++)
 235                         free_swap_cache(pagep[i]);
 236                 release_pages(pagep, todo, 0);
 237                 pagep += todo;
 238                 nr -= todo;
 239         }
 240 }
 241
 242 /*
 243  * Lookup a swap entry in the swap cache. A found page will be returned
 244  * unlocked and with its refcount incremented - we rely on the kernel
 245  * lock getting page table operations atomic even if we drop the page
 246  * lock before returning.
 247  */
 248 struct page * lookup_swap_cache(swp_entry_t entry)
 249 {
 250         struct page *page;
 251
 252         page = find_get_page(&swapper_space, entry.val);
 253
 254         if (page)
 255                 INC_CACHE_INFO(find_success);
 256
 257         INC_CACHE_INFO(find_total);
 258         return page;
 259 }
 260
 261 /*
 262  * Locate a page of swap in physical memory, reserving swap cache space
 263  * and reading the disk if it is not already cached.
 264  * A failure return means that either the page allocation failed or that
 265  * the swap entry is no longer in use.
 266  */
 267 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 268                         struct vm_area_struct *vma, unsigned long addr)
 269 {
 270         struct page *found_page, *new_page = NULL;
 271         int err;
 272
 273         do {
 274                 /*
 275                  * First check the swap cache.  Since this is normally
 276                  * called after lookup_swap_cache() failed, re-calling
 277                  * that would confuse statistics.
 278                  */
 279                 found_page = find_get_page(&swapper_space, entry.val);
 280                 if (found_page)
 281                         break;
 282
 283                 /*
 284                  * Get a new page to read into from swap.
 285                  */
 286                 if (!new_page) {
 287                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 288                         if (!new_page)
 289                                 break;          /* Out of memory */
 290                 }
 291
 292                 /*
 293                  * Swap entry may have been freed since our caller observed it.
 294                  */
 295                 err = swapcache_prepare(entry);
 296                 if (err == -EEXIST) /* seems racy */
 297                         continue;
 298                 if (err)           /* swp entry is obsolete ? */
 299                         break;
 300
 301                 /*
 302                  * Associate the page with swap entry in the swap cache.
 303                  * May fail (-EEXIST) if there is already a page associated
 304                  * with this entry in the swap cache: added by a racing
 305                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 306                  * re-using the just freed swap entry for an existing page.
 307                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 308                  */
 309                 __set_page_locked(new_page);
 310                 SetPageSwapBacked(new_page);
 311                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
 312                 if (likely(!err)) {
 313                         /*
 314                          * Initiate read into locked page and return.
 315                          */
 316                         lru_cache_add_anon(new_page);
 317                         swap_readpage(NULL, new_page);
 318                         return new_page;
 319                 }
 320                 ClearPageSwapBacked(new_page);
 321                 __clear_page_locked(new_page);
 322                 swapcache_free(entry, NULL);
 323         } while (err != -ENOMEM);
 324
 325         if (new_page)
 326                 page_cache_release(new_page);
 327         return found_page;
 328 }
 329
 330 /**
 331  * swapin_readahead - swap in pages in hope we need them soon
 332  * @entry: swap entry of this memory
 333  * @gfp_mask: memory allocation flags
 334  * @vma: user vma this address belongs to
 335  * @addr: target address for mempolicy
 336  *
 337  * Returns the struct page for entry and addr, after queueing swapin.
 338  *
 339  * Primitive swap readahead code. We simply read an aligned block of
 340  * (1 << page_cluster) entries in the swap area. This method is chosen
 341  * because it doesn't cost us any seek time.  We also make sure to queue
 342  * the 'original' request together with the readahead ones...
 343  *
 344  * This has been extended to use the NUMA policies from the mm triggering
 345  * the readahead.
 346  *
 347  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 348  */
 349 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 350                         struct vm_area_struct *vma, unsigned long addr)
 351 {
 352         int nr_pages;
 353         struct page *page;
 354         unsigned long offset;
 355         unsigned long end_offset;
 356
 357         /*
 358          * Get starting offset for readaround, and number of pages to read.
 359          * Adjust starting address by readbehind (for NUMA interleave case)?
 360          * No, it's very unlikely that swap layout would follow vma layout,
 361          * more likely that neighbouring swap pages came from the same node:
 362          * so use the same "addr" to choose the same node for each swap read.
 363          */
 364         nr_pages = valid_swaphandles(entry, &offset);
 365         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 366                 /* Ok, do the async read-ahead now */
 367                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 368                                                 gfp_mask, vma, addr);
 369                 if (!page)
 370                         break;
 371                 page_cache_release(page);
 372         }
 373         lru_add_drain();        /* Push any new pages onto the LRU now */
 374         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 375 }