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
  10 #include <linux/mm.h>
  11 #include <linux/kernel_stat.h>
  12 #include <linux/swap.h>
  13 #include <linux/init.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/backing-dev.h>
  16
  17 #include <asm/pgtable.h>
  18
  19 /*
  20  * swapper_space is a fiction, retained to simplify the path through
  21  * vmscan's shrink_list.  Only those fields initialized below are used.
  22  */
  23 static struct address_space_operations swap_aops = {
  24         .writepage      = swap_writepage,
  25 };
  26
  27 static struct backing_dev_info swap_backing_dev_info = {
  28         .state          = 0,    /* uncongested */
  29 };
  30
  31 struct address_space swapper_space = {
  32         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC),
  33         .tree_lock      = SPIN_LOCK_UNLOCKED,
  34         .nrpages        = 0,    /* total_swapcache_pages */
  35         .a_ops          = &swap_aops,
  36         .backing_dev_info = &swap_backing_dev_info,
  37 };
  38
  39 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  40
  41 static struct {
  42         unsigned long add_total;
  43         unsigned long del_total;
  44         unsigned long find_success;
  45         unsigned long find_total;
  46         unsigned long noent_race;
  47         unsigned long exist_race;
  48 } swap_cache_info;
  49
  50 void show_swap_cache_info(void)
  51 {
  52         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
  53                 swap_cache_info.add_total, swap_cache_info.del_total,
  54                 swap_cache_info.find_success, swap_cache_info.find_total,
  55                 swap_cache_info.noent_race, swap_cache_info.exist_race);
  56 }
  57
  58 /*
  59  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
  60  * but sets SwapCache flag and private instead of mapping and index.
  61  */
  62 static int __add_to_swap_cache(struct page *page,
  63                 swp_entry_t entry, int gfp_mask)
  64 {
  65         int error;
  66
  67         BUG_ON(PageSwapCache(page));
  68         BUG_ON(PagePrivate(page));
  69         error = radix_tree_preload(gfp_mask);
  70         if (!error) {
  71                 page_cache_get(page);
  72                 spin_lock(&swapper_space.tree_lock);
  73                 error = radix_tree_insert(&swapper_space.page_tree,
  74                                                 entry.val, page);
  75                 if (!error) {
  76                         SetPageLocked(page);
  77                         SetPageSwapCache(page);
  78                         page->private = entry.val;
  79                         total_swapcache_pages++;
  80                         pagecache_acct(1);
  81                 } else
  82                         page_cache_release(page);
  83                 spin_unlock(&swapper_space.tree_lock);
  84                 radix_tree_preload_end();
  85         }
  86         return error;
  87 }
  88
  89 static int add_to_swap_cache(struct page *page, swp_entry_t entry)
  90 {
  91         int error;
  92
  93         if (!swap_duplicate(entry)) {
  94                 INC_CACHE_INFO(noent_race);
  95                 return -ENOENT;
  96         }
  97         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
  98         /*
  99          * Anon pages are already on the LRU, we don't run lru_cache_add here.
 100          */
 101         if (error) {
 102                 swap_free(entry);
 103                 if (error == -EEXIST)
 104                         INC_CACHE_INFO(exist_race);
 105                 return error;
 106         }
 107         INC_CACHE_INFO(add_total);
 108         return 0;
 109 }
 110
 111 /*
 112  * This must be called only on pages that have
 113  * been verified to be in the swap cache.
 114  */
 115 void __delete_from_swap_cache(struct page *page)
 116 {
 117         BUG_ON(!PageLocked(page));
 118         BUG_ON(!PageSwapCache(page));
 119         BUG_ON(PageWriteback(page));
 120
 121         radix_tree_delete(&swapper_space.page_tree, page->private);
 122         page->private = 0;
 123         ClearPageSwapCache(page);
 124         total_swapcache_pages--;
 125         pagecache_acct(-1);
 126         INC_CACHE_INFO(del_total);
 127 }
 128
 129 /**
 130  * add_to_swap - allocate swap space for a page
 131  * @page: page we want to move to swap
 132  *
 133  * Allocate swap space for the page and add the page to the
 134  * swap cache.  Caller needs to hold the page lock.
 135  */
 136 int add_to_swap(struct page * page)
 137 {
 138         swp_entry_t entry;
 139         int pf_flags;
 140         int err;
 141
 142         if (!PageLocked(page))
 143                 BUG();
 144
 145         for (;;) {
 146                 entry = get_swap_page();
 147                 if (!entry.val)
 148                         return 0;
 149
 150                 /* Radix-tree node allocations are performing
 151                  * GFP_ATOMIC allocations under PF_MEMALLOC.
 152                  * They can completely exhaust the page allocator.
 153                  *
 154                  * So PF_MEMALLOC is dropped here.  This causes the slab
 155                  * allocations to fail earlier, so radix-tree nodes will
 156                  * then be allocated from the mempool reserves.
 157                  *
 158                  * We're still using __GFP_HIGH for radix-tree node
 159                  * allocations, so some of the emergency pools are available,
 160                  * just not all of them.
 161                  */
 162
 163                 pf_flags = current->flags;
 164                 current->flags &= ~PF_MEMALLOC;
 165
 166                 /*
 167                  * Add it to the swap cache and mark it dirty
 168                  */
 169                 err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 170
 171                 if (pf_flags & PF_MEMALLOC)
 172                         current->flags |= PF_MEMALLOC;
 173
 174                 switch (err) {
 175                 case 0:                         /* Success */
 176                         SetPageUptodate(page);
 177                         SetPageDirty(page);
 178                         INC_CACHE_INFO(add_total);
 179                         return 1;
 180                 case -EEXIST:
 181                         /* Raced with "speculative" read_swap_cache_async */
 182                         INC_CACHE_INFO(exist_race);
 183                         swap_free(entry);
 184                         continue;
 185                 default:
 186                         /* -ENOMEM radix-tree allocation failure */
 187                         swap_free(entry);
 188                         return 0;
 189                 }
 190         }
 191 }
 192
 193 /*
 194  * This must be called only on pages that have
 195  * been verified to be in the swap cache and locked.
 196  * It will never put the page into the free list,
 197  * the caller has a reference on the page.
 198  */
 199 void delete_from_swap_cache(struct page *page)
 200 {
 201         swp_entry_t entry;
 202
 203         BUG_ON(!PageSwapCache(page));
 204         BUG_ON(!PageLocked(page));
 205         BUG_ON(PageWriteback(page));
 206         BUG_ON(PagePrivate(page));
 207
 208         entry.val = page->private;
 209
 210         spin_lock(&swapper_space.tree_lock);
 211         __delete_from_swap_cache(page);
 212         spin_unlock(&swapper_space.tree_lock);
 213
 214         swap_free(entry);
 215         page_cache_release(page);
 216 }
 217
 218 /*
 219  * Strange swizzling function only for use by shmem_writepage
 220  */
 221 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 222 {
 223         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 224         if (!err) {
 225                 remove_from_page_cache(page);
 226                 page_cache_release(page);       /* pagecache ref */
 227                 if (!swap_duplicate(entry))
 228                         BUG();
 229                 SetPageDirty(page);
 230                 INC_CACHE_INFO(add_total);
 231         } else if (err == -EEXIST)
 232                 INC_CACHE_INFO(exist_race);
 233         return err;
 234 }
 235
 236 /*
 237  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 238  */
 239 int move_from_swap_cache(struct page *page, unsigned long index,
 240                 struct address_space *mapping)
 241 {
 242         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 243         if (!err) {
 244                 delete_from_swap_cache(page);
 245                 /* shift page from clean_pages to dirty_pages list */
 246                 ClearPageDirty(page);
 247                 set_page_dirty(page);
 248         }
 249         return err;
 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  * exclusive_swap_page() _with_ the lock.
 258  *                                      - Marcelo
 259  */
 260 static inline void free_swap_cache(struct page *page)
 261 {
 262         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 263                 remove_exclusive_swap_page(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. Can not do a lock_page,
 271  * as we are holding the page_table_lock spinlock.
 272  */
 273 void free_page_and_swap_cache(struct page *page)
 274 {
 275         free_swap_cache(page);
 276         page_cache_release(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         int chunk = 16;
 286         struct page **pagep = pages;
 287
 288         lru_add_drain();
 289         while (nr) {
 290                 int todo = min(chunk, nr);
 291                 int i;
 292
 293                 for (i = 0; i < todo; i++)
 294                         free_swap_cache(pagep[i]);
 295                 release_pages(pagep, todo, 0);
 296                 pagep += todo;
 297                 nr -= todo;
 298         }
 299 }
 300
 301 /*
 302  * Lookup a swap entry in the swap cache. A found page will be returned
 303  * unlocked and with its refcount incremented - we rely on the kernel
 304  * lock getting page table operations atomic even if we drop the page
 305  * lock before returning.
 306  */
 307 struct page * lookup_swap_cache(swp_entry_t entry)
 308 {
 309         struct page *page;
 310
 311         spin_lock(&swapper_space.tree_lock);
 312         page = radix_tree_lookup(&swapper_space.page_tree, entry.val);
 313         if (page) {
 314                 page_cache_get(page);
 315                 INC_CACHE_INFO(find_success);
 316         }
 317         spin_unlock(&swapper_space.tree_lock);
 318         INC_CACHE_INFO(find_total);
 319         return page;
 320 }
 321
 322 /*
 323  * Locate a page of swap in physical memory, reserving swap cache space
 324  * and reading the disk if it is not already cached.
 325  * A failure return means that either the page allocation failed or that
 326  * the swap entry is no longer in use.
 327  */
 328 struct page * read_swap_cache_async(swp_entry_t entry)
 329 {
 330         struct page *found_page, *new_page = NULL;
 331         int err;
 332
 333         do {
 334                 /*
 335                  * First check the swap cache.  Since this is normally
 336                  * called after lookup_swap_cache() failed, re-calling
 337                  * that would confuse statistics.
 338                  */
 339                 spin_lock(&swapper_space.tree_lock);
 340                 found_page = radix_tree_lookup(&swapper_space.page_tree,
 341                                                 entry.val);
 342                 if (found_page)
 343                         page_cache_get(found_page);
 344                 spin_unlock(&swapper_space.tree_lock);
 345                 if (found_page)
 346                         break;
 347
 348                 /*
 349                  * Get a new page to read into from swap.
 350                  */
 351                 if (!new_page) {
 352                         new_page = alloc_page(GFP_HIGHUSER);
 353                         if (!new_page)
 354                                 break;          /* Out of memory */
 355                 }
 356
 357                 /*
 358                  * Associate the page with swap entry in the swap cache.
 359                  * May fail (-ENOENT) if swap entry has been freed since
 360                  * our caller observed it.  May fail (-EEXIST) if there
 361                  * is already a page associated with this entry in the
 362                  * swap cache: added by a racing read_swap_cache_async,
 363                  * or by try_to_swap_out (or shmem_writepage) re-using
 364                  * the just freed swap entry for an existing page.
 365                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 366                  */
 367                 err = add_to_swap_cache(new_page, entry);
 368                 if (!err) {
 369                         /*
 370                          * Initiate read into locked page and return.
 371                          */
 372                         lru_cache_add_active(new_page);
 373                         swap_readpage(NULL, new_page);
 374                         return new_page;
 375                 }
 376         } while (err != -ENOENT && err != -ENOMEM);
 377
 378         if (new_page)
 379                 page_cache_release(new_page);
 380         return found_page;
 381 }