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