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/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, to make sync_page look nicer, and to allow
  23  * future use of radix_tree tags in the swap cache.
  24  */
  25 static struct address_space_operations swap_aops = {
  26         .writepage      = swap_writepage,
  27         .sync_page      = block_sync_page,
  28         .set_page_dirty = __set_page_dirty_nobuffers,
  29 };
  30
  31 static struct backing_dev_info swap_backing_dev_info = {
  32         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
  33         .unplug_io_fn   = swap_unplug_io_fn,
  34 };
  35
  36 struct address_space swapper_space = {
  37         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  38         .tree_lock      = RW_LOCK_UNLOCKED,
  39         .a_ops          = &swap_aops,
  40         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
  41         .backing_dev_info = &swap_backing_dev_info,
  42 };
  43 EXPORT_SYMBOL(swapper_space);
  44
  45 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  46
  47 static struct {
  48         unsigned long add_total;
  49         unsigned long del_total;
  50         unsigned long find_success;
  51         unsigned long find_total;
  52         unsigned long noent_race;
  53         unsigned long exist_race;
  54 } swap_cache_info;
  55
  56 void show_swap_cache_info(void)
  57 {
  58         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %lu+%lu\n",
  59                 swap_cache_info.add_total, swap_cache_info.del_total,
  60                 swap_cache_info.find_success, swap_cache_info.find_total,
  61                 swap_cache_info.noent_race, swap_cache_info.exist_race);
  62         printk("Free swap  = %lukB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  63         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  64 }
  65
  66 /*
  67  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
  68  * but sets SwapCache flag and private instead of mapping and index.
  69  */
  70 static int __add_to_swap_cache(struct page *page,
  71                 swp_entry_t entry, int gfp_mask)
  72 {
  73         int error;
  74
  75         BUG_ON(PageSwapCache(page));
  76         BUG_ON(PagePrivate(page));
  77         error = radix_tree_preload(gfp_mask);
  78         if (!error) {
  79                 write_lock_irq(&swapper_space.tree_lock);
  80                 error = radix_tree_insert(&swapper_space.page_tree,
  81                                                 entry.val, page);
  82                 if (!error) {
  83                         page_cache_get(page);
  84                         SetPageLocked(page);
  85                         SetPageSwapCache(page);
  86                         page->private = entry.val;
  87                         total_swapcache_pages++;
  88                         pagecache_acct(1);
  89                 }
  90                 write_unlock_irq(&swapper_space.tree_lock);
  91                 radix_tree_preload_end();
  92         }
  93         return error;
  94 }
  95
  96 static int add_to_swap_cache(struct page *page, swp_entry_t entry)
  97 {
  98         int error;
  99
 100         if (!swap_duplicate(entry)) {
 101                 INC_CACHE_INFO(noent_race);
 102                 return -ENOENT;
 103         }
 104         error = __add_to_swap_cache(page, entry, GFP_KERNEL);
 105         /*
 106          * Anon pages are already on the LRU, we don't run lru_cache_add here.
 107          */
 108         if (error) {
 109                 swap_free(entry);
 110                 if (error == -EEXIST)
 111                         INC_CACHE_INFO(exist_race);
 112                 return error;
 113         }
 114         INC_CACHE_INFO(add_total);
 115         return 0;
 116 }
 117
 118 /*
 119  * This must be called only on pages that have
 120  * been verified to be in the swap cache.
 121  */
 122 void __delete_from_swap_cache(struct page *page)
 123 {
 124         BUG_ON(!PageLocked(page));
 125         BUG_ON(!PageSwapCache(page));
 126         BUG_ON(PageWriteback(page));
 127
 128         radix_tree_delete(&swapper_space.page_tree, page->private);
 129         page->private = 0;
 130         ClearPageSwapCache(page);
 131         total_swapcache_pages--;
 132         pagecache_acct(-1);
 133         INC_CACHE_INFO(del_total);
 134 }
 135
 136 /**
 137  * add_to_swap - allocate swap space for a page
 138  * @page: page we want to move to swap
 139  *
 140  * Allocate swap space for the page and add the page to the
 141  * swap cache.  Caller needs to hold the page lock.
 142  */
 143 int add_to_swap(struct page * page)
 144 {
 145         swp_entry_t entry;
 146         int pf_flags;
 147         int err;
 148
 149         if (!PageLocked(page))
 150                 BUG();
 151
 152         for (;;) {
 153                 entry = get_swap_page();
 154                 if (!entry.val)
 155                         return 0;
 156
 157                 /* Radix-tree node allocations are performing
 158                  * GFP_ATOMIC allocations under PF_MEMALLOC.
 159                  * They can completely exhaust the page allocator.
 160                  *
 161                  * So PF_MEMALLOC is dropped here.  This causes the slab
 162                  * allocations to fail earlier, so radix-tree nodes will
 163                  * then be allocated from the mempool reserves.
 164                  *
 165                  * We're still using __GFP_HIGH for radix-tree node
 166                  * allocations, so some of the emergency pools are available,
 167                  * just not all of them.
 168                  */
 169
 170                 pf_flags = current->flags;
 171                 current->flags &= ~PF_MEMALLOC;
 172
 173                 /*
 174                  * Add it to the swap cache and mark it dirty
 175                  */
 176                 err = __add_to_swap_cache(page, entry, GFP_ATOMIC|__GFP_NOWARN);
 177
 178                 if (pf_flags & PF_MEMALLOC)
 179                         current->flags |= PF_MEMALLOC;
 180
 181                 switch (err) {
 182                 case 0:                         /* Success */
 183                         SetPageUptodate(page);
 184                         SetPageDirty(page);
 185                         INC_CACHE_INFO(add_total);
 186                         return 1;
 187                 case -EEXIST:
 188                         /* Raced with "speculative" read_swap_cache_async */
 189                         INC_CACHE_INFO(exist_race);
 190                         swap_free(entry);
 191                         continue;
 192                 default:
 193                         /* -ENOMEM radix-tree allocation failure */
 194                         swap_free(entry);
 195                         return 0;
 196                 }
 197         }
 198 }
 199
 200 /*
 201  * This must be called only on pages that have
 202  * been verified to be in the swap cache and locked.
 203  * It will never put the page into the free list,
 204  * the caller has a reference on the page.
 205  */
 206 void delete_from_swap_cache(struct page *page)
 207 {
 208         swp_entry_t entry;
 209
 210         BUG_ON(!PageSwapCache(page));
 211         BUG_ON(!PageLocked(page));
 212         BUG_ON(PageWriteback(page));
 213         BUG_ON(PagePrivate(page));
 214
 215         entry.val = page->private;
 216
 217         write_lock_irq(&swapper_space.tree_lock);
 218         __delete_from_swap_cache(page);
 219         write_unlock_irq(&swapper_space.tree_lock);
 220
 221         swap_free(entry);
 222         page_cache_release(page);
 223 }
 224
 225 /*
 226  * Strange swizzling function only for use by shmem_writepage
 227  */
 228 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 229 {
 230         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 231         if (!err) {
 232                 remove_from_page_cache(page);
 233                 page_cache_release(page);       /* pagecache ref */
 234                 if (!swap_duplicate(entry))
 235                         BUG();
 236                 SetPageDirty(page);
 237                 INC_CACHE_INFO(add_total);
 238         } else if (err == -EEXIST)
 239                 INC_CACHE_INFO(exist_race);
 240         return err;
 241 }
 242
 243 /*
 244  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 245  */
 246 int move_from_swap_cache(struct page *page, unsigned long index,
 247                 struct address_space *mapping)
 248 {
 249         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 250         if (!err) {
 251                 delete_from_swap_cache(page);
 252                 /* shift page from clean_pages to dirty_pages list */
 253                 ClearPageDirty(page);
 254                 set_page_dirty(page);
 255         }
 256         return err;
 257 }
 258
 259 /*
 260  * If we are the only user, then try to free up the swap cache.
 261  *
 262  * Its ok to check for PageSwapCache without the page lock
 263  * here because we are going to recheck again inside
 264  * exclusive_swap_page() _with_ the lock.
 265  *                                      - Marcelo
 266  */
 267 static inline void free_swap_cache(struct page *page)
 268 {
 269         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 270                 remove_exclusive_swap_page(page);
 271                 unlock_page(page);
 272         }
 273 }
 274
 275 /*
 276  * Perform a free_page(), also freeing any swap cache associated with
 277  * this page if it is the last user of the page. Can not do a lock_page,
 278  * as we are holding the page_table_lock spinlock.
 279  */
 280 void free_page_and_swap_cache(struct page *page)
 281 {
 282         free_swap_cache(page);
 283         page_cache_release(page);
 284 }
 285
 286 /*
 287  * Passed an array of pages, drop them all from swapcache and then release
 288  * them.  They are removed from the LRU and freed if this is their last use.
 289  */
 290 void free_pages_and_swap_cache(struct page **pages, int nr)
 291 {
 292         int chunk = 16;
 293         struct page **pagep = pages;
 294
 295         lru_add_drain();
 296         while (nr) {
 297                 int todo = min(chunk, nr);
 298                 int i;
 299
 300                 for (i = 0; i < todo; i++)
 301                         free_swap_cache(pagep[i]);
 302                 release_pages(pagep, todo, 0);
 303                 pagep += todo;
 304                 nr -= todo;
 305         }
 306 }
 307
 308 /*
 309  * Lookup a swap entry in the swap cache. A found page will be returned
 310  * unlocked and with its refcount incremented - we rely on the kernel
 311  * lock getting page table operations atomic even if we drop the page
 312  * lock before returning.
 313  */
 314 struct page * lookup_swap_cache(swp_entry_t entry)
 315 {
 316         struct page *page;
 317
 318         page = find_get_page(&swapper_space, entry.val);
 319
 320         if (page)
 321                 INC_CACHE_INFO(find_success);
 322
 323         INC_CACHE_INFO(find_total);
 324         return page;
 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                         struct vm_area_struct *vma, unsigned long addr)
 335 {
 336         struct page *found_page, *new_page = NULL;
 337         int err;
 338
 339         do {
 340                 /*
 341                  * First check the swap cache.  Since this is normally
 342                  * called after lookup_swap_cache() failed, re-calling
 343                  * that would confuse statistics.
 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_vma(GFP_HIGHUSER, vma, addr);
 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 }