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 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
  71 {
  72         int error;
  73
  74         BUG_ON(!PageLocked(page));
  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                         SetPageSwapCache(page);
  85                         set_page_private(page, entry.val);
  86                         total_swapcache_pages++;
  87                         __inc_zone_page_state(page, NR_FILE_PAGES);
  88                         INC_CACHE_INFO(add_total);
  89                 }
  90                 write_unlock_irq(&swapper_space.tree_lock);
  91                 radix_tree_preload_end();
  92         }
  93         return error;
  94 }
  95
  96 /*
  97  * This must be called only on pages that have
  98  * been verified to be in the swap cache.
  99  */
 100 void __delete_from_swap_cache(struct page *page)
 101 {
 102         BUG_ON(!PageLocked(page));
 103         BUG_ON(!PageSwapCache(page));
 104         BUG_ON(PageWriteback(page));
 105         BUG_ON(PagePrivate(page));
 106
 107         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 108         set_page_private(page, 0);
 109         ClearPageSwapCache(page);
 110         total_swapcache_pages--;
 111         __dec_zone_page_state(page, NR_FILE_PAGES);
 112         INC_CACHE_INFO(del_total);
 113 }
 114
 115 /**
 116  * add_to_swap - allocate swap space for a page
 117  * @page: page we want to move to swap
 118  *
 119  * Allocate swap space for the page and add the page to the
 120  * swap cache.  Caller needs to hold the page lock.
 121  */
 122 int add_to_swap(struct page * page, gfp_t gfp_mask)
 123 {
 124         swp_entry_t entry;
 125         int err;
 126
 127         BUG_ON(!PageLocked(page));
 128         BUG_ON(!PageUptodate(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                         SetPageDirty(page);
 152                         return 1;
 153                 case -EEXIST:
 154                         /* Raced with "speculative" read_swap_cache_async */
 155                         swap_free(entry);
 156                         continue;
 157                 default:
 158                         /* -ENOMEM radix-tree allocation failure */
 159                         swap_free(entry);
 160                         return 0;
 161                 }
 162         }
 163 }
 164
 165 /*
 166  * This must be called only on pages that have
 167  * been verified to be in the swap cache and locked.
 168  * It will never put the page into the free list,
 169  * the caller has a reference on the page.
 170  */
 171 void delete_from_swap_cache(struct page *page)
 172 {
 173         swp_entry_t entry;
 174
 175         entry.val = page_private(page);
 176
 177         write_lock_irq(&swapper_space.tree_lock);
 178         __delete_from_swap_cache(page);
 179         write_unlock_irq(&swapper_space.tree_lock);
 180
 181         swap_free(entry);
 182         page_cache_release(page);
 183 }
 184
 185 /*
 186  * If we are the only user, then try to free up the swap cache.
 187  *
 188  * Its ok to check for PageSwapCache without the page lock
 189  * here because we are going to recheck again inside
 190  * exclusive_swap_page() _with_ the lock.
 191  *                                      - Marcelo
 192  */
 193 static inline void free_swap_cache(struct page *page)
 194 {
 195         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 196                 remove_exclusive_swap_page(page);
 197                 unlock_page(page);
 198         }
 199 }
 200
 201 /*
 202  * Perform a free_page(), also freeing any swap cache associated with
 203  * this page if it is the last user of the page.
 204  */
 205 void free_page_and_swap_cache(struct page *page)
 206 {
 207         free_swap_cache(page);
 208         page_cache_release(page);
 209 }
 210
 211 /*
 212  * Passed an array of pages, drop them all from swapcache and then release
 213  * them.  They are removed from the LRU and freed if this is their last use.
 214  */
 215 void free_pages_and_swap_cache(struct page **pages, int nr)
 216 {
 217         struct page **pagep = pages;
 218
 219         lru_add_drain();
 220         while (nr) {
 221                 int todo = min(nr, PAGEVEC_SIZE);
 222                 int i;
 223
 224                 for (i = 0; i < todo; i++)
 225                         free_swap_cache(pagep[i]);
 226                 release_pages(pagep, todo, 0);
 227                 pagep += todo;
 228                 nr -= todo;
 229         }
 230 }
 231
 232 /*
 233  * Lookup a swap entry in the swap cache. A found page will be returned
 234  * unlocked and with its refcount incremented - we rely on the kernel
 235  * lock getting page table operations atomic even if we drop the page
 236  * lock before returning.
 237  */
 238 struct page * lookup_swap_cache(swp_entry_t entry)
 239 {
 240         struct page *page;
 241
 242         page = find_get_page(&swapper_space, entry.val);
 243
 244         if (page)
 245                 INC_CACHE_INFO(find_success);
 246
 247         INC_CACHE_INFO(find_total);
 248         return page;
 249 }
 250
 251 /*
 252  * Locate a page of swap in physical memory, reserving swap cache space
 253  * and reading the disk if it is not already cached.
 254  * A failure return means that either the page allocation failed or that
 255  * the swap entry is no longer in use.
 256  */
 257 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 258                         struct vm_area_struct *vma, unsigned long addr)
 259 {
 260         struct page *found_page, *new_page = NULL;
 261         int err;
 262
 263         do {
 264                 /*
 265                  * First check the swap cache.  Since this is normally
 266                  * called after lookup_swap_cache() failed, re-calling
 267                  * that would confuse statistics.
 268                  */
 269                 found_page = find_get_page(&swapper_space, entry.val);
 270                 if (found_page)
 271                         break;
 272
 273                 /*
 274                  * Get a new page to read into from swap.
 275                  */
 276                 if (!new_page) {
 277                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 278                         if (!new_page)
 279                                 break;          /* Out of memory */
 280                 }
 281
 282                 /*
 283                  * Swap entry may have been freed since our caller observed it.
 284                  */
 285                 if (!swap_duplicate(entry))
 286                         break;
 287
 288                 /*
 289                  * Associate the page with swap entry in the swap cache.
 290                  * May fail (-EEXIST) if there is already a page associated
 291                  * with this entry in the swap cache: added by a racing
 292                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 293                  * re-using the just freed swap entry for an existing page.
 294                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 295                  */
 296                 SetPageLocked(new_page);
 297                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
 298                 if (!err) {
 299                         /*
 300                          * Initiate read into locked page and return.
 301                          */
 302                         lru_cache_add_active(new_page);
 303                         swap_readpage(NULL, new_page);
 304                         return new_page;
 305                 }
 306                 ClearPageLocked(new_page);
 307                 swap_free(entry);
 308         } while (err != -ENOMEM);
 309
 310         if (new_page)
 311                 page_cache_release(new_page);
 312         return found_page;
 313 }
 314
 315 /**
 316  * swapin_readahead - swap in pages in hope we need them soon
 317  * @entry: swap entry of this memory
 318  * @vma: user vma this address belongs to
 319  * @addr: target address for mempolicy
 320  *
 321  * Returns the struct page for entry and addr, after queueing swapin.
 322  *
 323  * Primitive swap readahead code. We simply read an aligned block of
 324  * (1 << page_cluster) entries in the swap area. This method is chosen
 325  * because it doesn't cost us any seek time.  We also make sure to queue
 326  * the 'original' request together with the readahead ones...
 327  *
 328  * This has been extended to use the NUMA policies from the mm triggering
 329  * the readahead.
 330  *
 331  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 332  */
 333 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 334                         struct vm_area_struct *vma, unsigned long addr)
 335 {
 336         int nr_pages;
 337         struct page *page;
 338         unsigned long offset;
 339         unsigned long end_offset;
 340
 341         /*
 342          * Get starting offset for readaround, and number of pages to read.
 343          * Adjust starting address by readbehind (for NUMA interleave case)?
 344          * No, it's very unlikely that swap layout would follow vma layout,
 345          * more likely that neighbouring swap pages came from the same node:
 346          * so use the same "addr" to choose the same node for each swap read.
 347          */
 348         nr_pages = valid_swaphandles(entry, &offset);
 349         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 350                 /* Ok, do the async read-ahead now */
 351                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 352                                                 gfp_mask, vma, addr);
 353                 if (!page)
 354                         break;
 355                 page_cache_release(page);
 356         }
 357         lru_add_drain();        /* Push any new pages onto the LRU now */
 358         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 359 }