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/buffer_head.h>
  16 #include <linux/backing-dev.h>
  17
  18 #include <asm/pgtable.h>
  19
  20 /*
  21  * swapper_space is a fiction, retained to simplify the path through
  22  * vmscan's shrink_list.  Only those fields initialized below are used.
  23  */
  24 static struct address_space_operations swap_aops = {
  25         .writepage      = swap_writepage,
  26         .sync_page      = block_sync_page,
  27         .set_page_dirty = __set_page_dirty_nobuffers,
  28 };
  29
  30 static struct backing_dev_info swap_backing_dev_info = {
  31         .memory_backed  = 1,    /* Does not contribute to dirty memory */
  32         .unplug_io_fn   = swap_unplug_io_fn,
  33 };
  34
  35 struct address_space swapper_space = {
  36         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC),
  37         .tree_lock      = SPIN_LOCK_UNLOCKED,
  38         .a_ops          = &swap_aops,
  39         .backing_dev_info = &swap_backing_dev_info,
  40 };
  41
  42 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  43
  44 static struct {
  45         unsigned long add_total;
  46         unsigned long del_total;
  47         unsigned long find_success;
  48         unsigned long find_total;
  49         unsigned long noent_race;
  50         unsigned long exist_race;
  51 } swap_cache_info;
  52
  53 void show_swap_cache_info(void)
  54 {
  55         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
  56                 swap_cache_info.add_total, swap_cache_info.del_total,
  57                 swap_cache_info.find_success, swap_cache_info.find_total,
  58                 swap_cache_info.noent_race, swap_cache_info.exist_race);
  59 }
  60
  61 /*
  62  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
  63  * but sets SwapCache flag and private instead of mapping and index.
  64  */
  65 static int __add_to_swap_cache(struct page *page,
  66                 swp_entry_t entry, int gfp_mask)
  67 {
  68         int error;
  69
  70         BUG_ON(PageSwapCache(page));
  71         BUG_ON(PagePrivate(page));
  72         error = radix_tree_preload(gfp_mask);
  73         if (!error) {
  74                 spin_lock_irq(&swapper_space.tree_lock);
  75                 error = radix_tree_insert(&swapper_space.page_tree,
  76                                                 entry.val, page);
  77                 if (!error) {
  78                         page_cache_get(page);
  79                         SetPageLocked(page);
  80                         SetPageSwapCache(page);
  81                         page->private = entry.val;
  82                         total_swapcache_pages++;
  83                         pagecache_acct(1);
  84                 }
  85                 spin_unlock_irq(&swapper_space.tree_lock);
  86                 radix_tree_preload_end();
  87         }
  88         return error;
  89 }
  90
  91 static int add_to_swap_cache(struct page *page, swp_entry_t entry)
  92 {
  93         int error;
  94
  95         if (!swap_duplicate(entry)) {
  96                 INC_CACHE_INFO(noent_race);
  97                 return -ENOENT;
  98         }
  99         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
 100         /*
 101          * Anon pages are already on the LRU, we don't run lru_cache_add here.
 102          */
 103         if (error) {
 104                 swap_free(entry);
 105                 if (error == -EEXIST)
 106                         INC_CACHE_INFO(exist_race);
 107                 return error;
 108         }
 109         INC_CACHE_INFO(add_total);
 110         return 0;
 111 }
 112
 113 /*
 114  * This must be called only on pages that have
 115  * been verified to be in the swap cache.
 116  */
 117 void __delete_from_swap_cache(struct page *page)
 118 {
 119         BUG_ON(!PageLocked(page));
 120         BUG_ON(!PageSwapCache(page));
 121         BUG_ON(PageWriteback(page));
 122
 123         radix_tree_delete(&swapper_space.page_tree, page->private);
 124         page->private = 0;
 125         ClearPageSwapCache(page);
 126         total_swapcache_pages--;
 127         pagecache_acct(-1);
 128         INC_CACHE_INFO(del_total);
 129 }
 130
 131 /**
 132  * add_to_swap - allocate swap space for a page
 133  * @page: page we want to move to swap
 134  *
 135  * Allocate swap space for the page and add the page to the
 136  * swap cache.  Caller needs to hold the page lock.
 137  */
 138 int add_to_swap(struct page * page)
 139 {
 140         swp_entry_t entry;
 141         int pf_flags;
 142         int err;
 143
 144         if (!PageLocked(page))
 145                 BUG();
 146
 147         for (;;) {
 148                 entry = get_swap_page();
 149                 if (!entry.val)
 150                         return 0;
 151
 152                 /* Radix-tree node allocations are performing
 153                  * GFP_ATOMIC allocations under PF_MEMALLOC.
 154                  * They can completely exhaust the page allocator.
 155                  *
 156                  * So PF_MEMALLOC is dropped here.  This causes the slab
 157                  * allocations to fail earlier, so radix-tree nodes will
 158                  * then be allocated from the mempool reserves.
 159                  *
 160                  * We're still using __GFP_HIGH for radix-tree node
 161                  * allocations, so some of the emergency pools are available,
 162                  * just not all of them.
 163                  */
 164
 165                 pf_flags = current->flags;
 166                 current->flags &= ~PF_MEMALLOC;
 167
 168                 /*
 169                  * Add it to the swap cache and mark it dirty
 170                  */
 171                 err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 172
 173                 if (pf_flags & PF_MEMALLOC)
 174                         current->flags |= PF_MEMALLOC;
 175
 176                 switch (err) {
 177                 case 0:                         /* Success */
 178                         SetPageUptodate(page);
 179                         SetPageDirty(page);
 180                         INC_CACHE_INFO(add_total);
 181                         return 1;
 182                 case -EEXIST:
 183                         /* Raced with "speculative" read_swap_cache_async */
 184                         INC_CACHE_INFO(exist_race);
 185                         swap_free(entry);
 186                         continue;
 187                 default:
 188                         /* -ENOMEM radix-tree allocation failure */
 189                         swap_free(entry);
 190                         return 0;
 191                 }
 192         }
 193 }
 194
 195 /*
 196  * This must be called only on pages that have
 197  * been verified to be in the swap cache and locked.
 198  * It will never put the page into the free list,
 199  * the caller has a reference on the page.
 200  */
 201 void delete_from_swap_cache(struct page *page)
 202 {
 203         swp_entry_t entry;
 204
 205         BUG_ON(!PageSwapCache(page));
 206         BUG_ON(!PageLocked(page));
 207         BUG_ON(PageWriteback(page));
 208         BUG_ON(PagePrivate(page));
 209
 210         entry.val = page->private;
 211
 212         spin_lock_irq(&swapper_space.tree_lock);
 213         __delete_from_swap_cache(page);
 214         spin_unlock_irq(&swapper_space.tree_lock);
 215
 216         swap_free(entry);
 217         page_cache_release(page);
 218 }
 219
 220 /*
 221  * Strange swizzling function only for use by shmem_writepage
 222  */
 223 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 224 {
 225         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 226         if (!err) {
 227                 remove_from_page_cache(page);
 228                 page_cache_release(page);       /* pagecache ref */
 229                 if (!swap_duplicate(entry))
 230                         BUG();
 231                 SetPageDirty(page);
 232                 INC_CACHE_INFO(add_total);
 233         } else if (err == -EEXIST)
 234                 INC_CACHE_INFO(exist_race);
 235         return err;
 236 }
 237
 238 /*
 239  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 240  */
 241 int move_from_swap_cache(struct page *page, unsigned long index,
 242                 struct address_space *mapping)
 243 {
 244         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 245         if (!err) {
 246                 delete_from_swap_cache(page);
 247                 /* shift page from clean_pages to dirty_pages list */
 248                 ClearPageDirty(page);
 249                 set_page_dirty(page);
 250         }
 251         return err;
 252 }
 253
 254 /*
 255  * If we are the only user, then try to free up the swap cache.
 256  *
 257  * Its ok to check for PageSwapCache without the page lock
 258  * here because we are going to recheck again inside
 259  * exclusive_swap_page() _with_ the lock.
 260  *                                      - Marcelo
 261  */
 262 static inline void free_swap_cache(struct page *page)
 263 {
 264         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 265                 remove_exclusive_swap_page(page);
 266                 unlock_page(page);
 267         }
 268 }
 269
 270 /*
 271  * Perform a free_page(), also freeing any swap cache associated with
 272  * this page if it is the last user of the page. Can not do a lock_page,
 273  * as we are holding the page_table_lock spinlock.
 274  */
 275 void free_page_and_swap_cache(struct page *page)
 276 {
 277         free_swap_cache(page);
 278         page_cache_release(page);
 279 }
 280
 281 /*
 282  * Passed an array of pages, drop them all from swapcache and then release
 283  * them.  They are removed from the LRU and freed if this is their last use.
 284  */
 285 void free_pages_and_swap_cache(struct page **pages, int nr)
 286 {
 287         int chunk = 16;
 288         struct page **pagep = pages;
 289
 290         lru_add_drain();
 291         while (nr) {
 292                 int todo = min(chunk, nr);
 293                 int i;
 294
 295                 for (i = 0; i < todo; i++)
 296                         free_swap_cache(pagep[i]);
 297                 release_pages(pagep, todo, 0);
 298                 pagep += todo;
 299                 nr -= todo;
 300         }
 301 }
 302
 303 /*
 304  * Lookup a swap entry in the swap cache. A found page will be returned
 305  * unlocked and with its refcount incremented - we rely on the kernel
 306  * lock getting page table operations atomic even if we drop the page
 307  * lock before returning.
 308  */
 309 struct page * lookup_swap_cache(swp_entry_t entry)
 310 {
 311         struct page *page;
 312
 313         spin_lock_irq(&swapper_space.tree_lock);
 314         page = radix_tree_lookup(&swapper_space.page_tree, entry.val);
 315         if (page) {
 316                 page_cache_get(page);
 317                 INC_CACHE_INFO(find_success);
 318         }
 319         spin_unlock_irq(&swapper_space.tree_lock);
 320         INC_CACHE_INFO(find_total);
 321         return page;
 322 }
 323
 324 /*
 325  * Locate a page of swap in physical memory, reserving swap cache space
 326  * and reading the disk if it is not already cached.
 327  * A failure return means that either the page allocation failed or that
 328  * the swap entry is no longer in use.
 329  */
 330 struct page *read_swap_cache_async(swp_entry_t entry,
 331                         struct vm_area_struct *vma, unsigned long addr)
 332 {
 333         struct page *found_page, *new_page = NULL;
 334         int err;
 335
 336         do {
 337                 /*
 338                  * First check the swap cache.  Since this is normally
 339                  * called after lookup_swap_cache() failed, re-calling
 340                  * that would confuse statistics.
 341                  */
 342                 spin_lock_irq(&swapper_space.tree_lock);
 343                 found_page = radix_tree_lookup(&swapper_space.page_tree,
 344                                                 entry.val);
 345                 if (found_page)
 346                         page_cache_get(found_page);
 347                 spin_unlock_irq(&swapper_space.tree_lock);
 348                 if (found_page)
 349                         break;
 350
 351                 /*
 352                  * Get a new page to read into from swap.
 353                  */
 354                 if (!new_page) {
 355                         new_page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 356                         if (!new_page)
 357                                 break;          /* Out of memory */
 358                 }
 359
 360                 /*
 361                  * Associate the page with swap entry in the swap cache.
 362                  * May fail (-ENOENT) if swap entry has been freed since
 363                  * our caller observed it.  May fail (-EEXIST) if there
 364                  * is already a page associated with this entry in the
 365                  * swap cache: added by a racing read_swap_cache_async,
 366                  * or by try_to_swap_out (or shmem_writepage) re-using
 367                  * the just freed swap entry for an existing page.
 368                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 369                  */
 370                 err = add_to_swap_cache(new_page, entry);
 371                 if (!err) {
 372                         /*
 373                          * Initiate read into locked page and return.
 374                          */
 375                         lru_cache_add_active(new_page);
 376                         swap_readpage(NULL, new_page);
 377                         return new_page;
 378                 }
 379         } while (err != -ENOENT && err != -ENOMEM);
 380
 381         if (new_page)
 382                 page_cache_release(new_page);
 383         return found_page;
 384 }