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
  21 #include <asm/pgtable.h>
  22
  23 /*
  24  * swapper_space is a fiction, retained to simplify the path through
  25  * vmscan's shrink_page_list, to make sync_page look nicer, and to allow
  26  * future use of radix_tree tags in the swap cache.
  27  */
  28 static const struct address_space_operations swap_aops = {
  29         .writepage      = swap_writepage,
  30         .sync_page      = block_sync_page,
  31         .set_page_dirty = __set_page_dirty_nobuffers,
  32         .migratepage    = migrate_page,
  33 };
  34
  35 static struct backing_dev_info swap_backing_dev_info = {
  36         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
  37         .unplug_io_fn   = swap_unplug_io_fn,
  38 };
  39
  40 struct address_space swapper_space = {
  41         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  42         .tree_lock      = __RW_LOCK_UNLOCKED(swapper_space.tree_lock),
  43         .a_ops          = &swap_aops,
  44         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
  45         .backing_dev_info = &swap_backing_dev_info,
  46 };
  47
  48 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  49
  50 static struct {
  51         unsigned long add_total;
  52         unsigned long del_total;
  53         unsigned long find_success;
  54         unsigned long find_total;
  55 } swap_cache_info;
  56
  57 void show_swap_cache_info(void)
  58 {
  59         printk("Swap cache: add %lu, delete %lu, find %lu/%lu\n",
  60                 swap_cache_info.add_total, swap_cache_info.del_total,
  61                 swap_cache_info.find_success, swap_cache_info.find_total);
  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, swp_entry_t entry,
  71                                gfp_t gfp_mask)
  72 {
  73         int error;
  74
  75         BUG_ON(!PageLocked(page));
  76         BUG_ON(PageSwapCache(page));
  77         BUG_ON(PagePrivate(page));
  78         error = radix_tree_preload(gfp_mask);
  79         if (!error) {
  80                 write_lock_irq(&swapper_space.tree_lock);
  81                 error = radix_tree_insert(&swapper_space.page_tree,
  82                                                 entry.val, page);
  83                 if (!error) {
  84                         page_cache_get(page);
  85                         SetPageSwapCache(page);
  86                         set_page_private(page, entry.val);
  87                         total_swapcache_pages++;
  88                         __inc_zone_page_state(page, NR_FILE_PAGES);
  89                         INC_CACHE_INFO(add_total);
  90                 }
  91                 write_unlock_irq(&swapper_space.tree_lock);
  92                 radix_tree_preload_end();
  93         }
  94         return error;
  95 }
  96
  97 /*
  98  * This must be called only on pages that have
  99  * been verified to be in the swap cache.
 100  */
 101 void __delete_from_swap_cache(struct page *page)
 102 {
 103         BUG_ON(!PageLocked(page));
 104         BUG_ON(!PageSwapCache(page));
 105         BUG_ON(PageWriteback(page));
 106         BUG_ON(PagePrivate(page));
 107
 108         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 109         set_page_private(page, 0);
 110         ClearPageSwapCache(page);
 111         total_swapcache_pages--;
 112         __dec_zone_page_state(page, NR_FILE_PAGES);
 113         INC_CACHE_INFO(del_total);
 114 }
 115
 116 /**
 117  * add_to_swap - allocate swap space for a page
 118  * @page: page we want to move to swap
 119  *
 120  * Allocate swap space for the page and add the page to the
 121  * swap cache.  Caller needs to hold the page lock.
 122  */
 123 int add_to_swap(struct page * page, gfp_t gfp_mask)
 124 {
 125         swp_entry_t entry;
 126         int err;
 127
 128         BUG_ON(!PageLocked(page));
 129
 130         for (;;) {
 131                 entry = get_swap_page();
 132                 if (!entry.val)
 133                         return 0;
 134
 135                 /*
 136                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 137                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 138                  * stops emergency reserves from being allocated.
 139                  *
 140                  * TODO: this could cause a theoretical memory reclaim
 141                  * deadlock in the swap out path.
 142                  */
 143                 /*
 144                  * Add it to the swap cache and mark it dirty
 145                  */
 146                 err = add_to_swap_cache(page, entry,
 147                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
 148
 149                 switch (err) {
 150                 case 0:                         /* Success */
 151                         SetPageUptodate(page);
 152                         SetPageDirty(page);
 153                         return 1;
 154                 case -EEXIST:
 155                         /* Raced with "speculative" read_swap_cache_async */
 156                         swap_free(entry);
 157                         continue;
 158                 default:
 159                         /* -ENOMEM radix-tree allocation failure */
 160                         swap_free(entry);
 161                         return 0;
 162                 }
 163         }
 164 }
 165
 166 /*
 167  * This must be called only on pages that have
 168  * been verified to be in the swap cache and locked.
 169  * It will never put the page into the free list,
 170  * the caller has a reference on the page.
 171  */
 172 void delete_from_swap_cache(struct page *page)
 173 {
 174         swp_entry_t entry;
 175
 176         entry.val = page_private(page);
 177
 178         write_lock_irq(&swapper_space.tree_lock);
 179         __delete_from_swap_cache(page);
 180         write_unlock_irq(&swapper_space.tree_lock);
 181
 182         swap_free(entry);
 183         page_cache_release(page);
 184 }
 185
 186 /*
 187  * Strange swizzling function only for use by shmem_writepage
 188  */
 189 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 190 {
 191         int err = add_to_swap_cache(page, entry, GFP_ATOMIC);
 192         if (!err) {
 193                 remove_from_page_cache(page);
 194                 page_cache_release(page);       /* pagecache ref */
 195                 if (!swap_duplicate(entry))
 196                         BUG();
 197                 SetPageDirty(page);
 198         }
 199         return err;
 200 }
 201
 202 /*
 203  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 204  */
 205 int move_from_swap_cache(struct page *page, unsigned long index,
 206                 struct address_space *mapping)
 207 {
 208         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 209         if (!err) {
 210                 delete_from_swap_cache(page);
 211                 /* shift page from clean_pages to dirty_pages list */
 212                 ClearPageDirty(page);
 213                 set_page_dirty(page);
 214         }
 215         return err;
 216 }
 217
 218 /*
 219  * If we are the only user, then try to free up the swap cache.
 220  *
 221  * Its ok to check for PageSwapCache without the page lock
 222  * here because we are going to recheck again inside
 223  * exclusive_swap_page() _with_ the lock.
 224  *                                      - Marcelo
 225  */
 226 static inline void free_swap_cache(struct page *page)
 227 {
 228         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 229                 remove_exclusive_swap_page(page);
 230                 unlock_page(page);
 231         }
 232 }
 233
 234 /*
 235  * Perform a free_page(), also freeing any swap cache associated with
 236  * this page if it is the last user of the page.
 237  */
 238 void free_page_and_swap_cache(struct page *page)
 239 {
 240         free_swap_cache(page);
 241         page_cache_release(page);
 242 }
 243
 244 /*
 245  * Passed an array of pages, drop them all from swapcache and then release
 246  * them.  They are removed from the LRU and freed if this is their last use.
 247  */
 248 void free_pages_and_swap_cache(struct page **pages, int nr)
 249 {
 250         struct page **pagep = pages;
 251
 252         lru_add_drain();
 253         while (nr) {
 254                 int todo = min(nr, PAGEVEC_SIZE);
 255                 int i;
 256
 257                 for (i = 0; i < todo; i++)
 258                         free_swap_cache(pagep[i]);
 259                 release_pages(pagep, todo, 0);
 260                 pagep += todo;
 261                 nr -= todo;
 262         }
 263 }
 264
 265 /*
 266  * Lookup a swap entry in the swap cache. A found page will be returned
 267  * unlocked and with its refcount incremented - we rely on the kernel
 268  * lock getting page table operations atomic even if we drop the page
 269  * lock before returning.
 270  */
 271 struct page * lookup_swap_cache(swp_entry_t entry)
 272 {
 273         struct page *page;
 274
 275         page = find_get_page(&swapper_space, entry.val);
 276
 277         if (page)
 278                 INC_CACHE_INFO(find_success);
 279
 280         INC_CACHE_INFO(find_total);
 281         return page;
 282 }
 283
 284 /*
 285  * Locate a page of swap in physical memory, reserving swap cache space
 286  * and reading the disk if it is not already cached.
 287  * A failure return means that either the page allocation failed or that
 288  * the swap entry is no longer in use.
 289  */
 290 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 291                         struct vm_area_struct *vma, unsigned long addr)
 292 {
 293         struct page *found_page, *new_page = NULL;
 294         int err;
 295
 296         do {
 297                 /*
 298                  * First check the swap cache.  Since this is normally
 299                  * called after lookup_swap_cache() failed, re-calling
 300                  * that would confuse statistics.
 301                  */
 302                 found_page = find_get_page(&swapper_space, entry.val);
 303                 if (found_page)
 304                         break;
 305
 306                 /*
 307                  * Get a new page to read into from swap.
 308                  */
 309                 if (!new_page) {
 310                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 311                         if (!new_page)
 312                                 break;          /* Out of memory */
 313                 }
 314
 315                 /*
 316                  * Swap entry may have been freed since our caller observed it.
 317                  */
 318                 if (!swap_duplicate(entry))
 319                         break;
 320
 321                 /*
 322                  * Associate the page with swap entry in the swap cache.
 323                  * May fail (-EEXIST) if there is already a page associated
 324                  * with this entry in the swap cache: added by a racing
 325                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 326                  * re-using the just freed swap entry for an existing page.
 327                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 328                  */
 329                 SetPageLocked(new_page);
 330                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
 331                 if (!err) {
 332                         /*
 333                          * Initiate read into locked page and return.
 334                          */
 335                         lru_cache_add_active(new_page);
 336                         swap_readpage(NULL, new_page);
 337                         return new_page;
 338                 }
 339                 ClearPageLocked(new_page);
 340                 swap_free(entry);
 341         } while (err != -ENOMEM);
 342
 343         if (new_page)
 344                 page_cache_release(new_page);
 345         return found_page;
 346 }
 347
 348 /**
 349  * swapin_readahead - swap in pages in hope we need them soon
 350  * @entry: swap entry of this memory
 351  * @vma: user vma this address belongs to
 352  * @addr: target address for mempolicy
 353  *
 354  * Returns the struct page for entry and addr, after queueing swapin.
 355  *
 356  * Primitive swap readahead code. We simply read an aligned block of
 357  * (1 << page_cluster) entries in the swap area. This method is chosen
 358  * because it doesn't cost us any seek time.  We also make sure to queue
 359  * the 'original' request together with the readahead ones...
 360  *
 361  * This has been extended to use the NUMA policies from the mm triggering
 362  * the readahead.
 363  *
 364  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 365  */
 366 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 367                         struct vm_area_struct *vma, unsigned long addr)
 368 {
 369         int nr_pages;
 370         struct page *page;
 371         unsigned long offset;
 372         unsigned long end_offset;
 373
 374         /*
 375          * Get starting offset for readaround, and number of pages to read.
 376          * Adjust starting address by readbehind (for NUMA interleave case)?
 377          * No, it's very unlikely that swap layout would follow vma layout,
 378          * more likely that neighbouring swap pages came from the same node:
 379          * so use the same "addr" to choose the same node for each swap read.
 380          */
 381         nr_pages = valid_swaphandles(entry, &offset);
 382         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 383                 /* Ok, do the async read-ahead now */
 384                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 385                                                 gfp_mask, vma, addr);
 386                 if (!page)
 387                         break;
 388                 page_cache_release(page);
 389         }
 390         lru_add_drain();        /* Push any new pages onto the LRU now */
 391         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 392 }