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  * @gfp_mask: memory allocation flags
 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         BUG_ON(!PageUptodate(page));
 130
 131         for (;;) {
 132                 entry = get_swap_page();
 133                 if (!entry.val)
 134                         return 0;
 135
 136                 /*
 137                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 138                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 139                  * stops emergency reserves from being allocated.
 140                  *
 141                  * TODO: this could cause a theoretical memory reclaim
 142                  * deadlock in the swap out path.
 143                  */
 144                 /*
 145                  * Add it to the swap cache and mark it dirty
 146                  */
 147                 err = add_to_swap_cache(page, entry,
 148                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
 149
 150                 switch (err) {
 151                 case 0:                         /* Success */
 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  * If we are the only user, then try to free up the swap cache.
 188  *
 189  * Its ok to check for PageSwapCache without the page lock
 190  * here because we are going to recheck again inside
 191  * exclusive_swap_page() _with_ the lock.
 192  *                                      - Marcelo
 193  */
 194 static inline void free_swap_cache(struct page *page)
 195 {
 196         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 197                 remove_exclusive_swap_page(page);
 198                 unlock_page(page);
 199         }
 200 }
 201
 202 /*
 203  * Perform a free_page(), also freeing any swap cache associated with
 204  * this page if it is the last user of the page.
 205  */
 206 void free_page_and_swap_cache(struct page *page)
 207 {
 208         free_swap_cache(page);
 209         page_cache_release(page);
 210 }
 211
 212 /*
 213  * Passed an array of pages, drop them all from swapcache and then release
 214  * them.  They are removed from the LRU and freed if this is their last use.
 215  */
 216 void free_pages_and_swap_cache(struct page **pages, int nr)
 217 {
 218         struct page **pagep = pages;
 219
 220         lru_add_drain();
 221         while (nr) {
 222                 int todo = min(nr, PAGEVEC_SIZE);
 223                 int i;
 224
 225                 for (i = 0; i < todo; i++)
 226                         free_swap_cache(pagep[i]);
 227                 release_pages(pagep, todo, 0);
 228                 pagep += todo;
 229                 nr -= todo;
 230         }
 231 }
 232
 233 /*
 234  * Lookup a swap entry in the swap cache. A found page will be returned
 235  * unlocked and with its refcount incremented - we rely on the kernel
 236  * lock getting page table operations atomic even if we drop the page
 237  * lock before returning.
 238  */
 239 struct page * lookup_swap_cache(swp_entry_t entry)
 240 {
 241         struct page *page;
 242
 243         page = find_get_page(&swapper_space, entry.val);
 244
 245         if (page)
 246                 INC_CACHE_INFO(find_success);
 247
 248         INC_CACHE_INFO(find_total);
 249         return page;
 250 }
 251
 252 /*
 253  * Locate a page of swap in physical memory, reserving swap cache space
 254  * and reading the disk if it is not already cached.
 255  * A failure return means that either the page allocation failed or that
 256  * the swap entry is no longer in use.
 257  */
 258 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 259                         struct vm_area_struct *vma, unsigned long addr)
 260 {
 261         struct page *found_page, *new_page = NULL;
 262         int err;
 263
 264         do {
 265                 /*
 266                  * First check the swap cache.  Since this is normally
 267                  * called after lookup_swap_cache() failed, re-calling
 268                  * that would confuse statistics.
 269                  */
 270                 found_page = find_get_page(&swapper_space, entry.val);
 271                 if (found_page)
 272                         break;
 273
 274                 /*
 275                  * Get a new page to read into from swap.
 276                  */
 277                 if (!new_page) {
 278                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 279                         if (!new_page)
 280                                 break;          /* Out of memory */
 281                 }
 282
 283                 /*
 284                  * Swap entry may have been freed since our caller observed it.
 285                  */
 286                 if (!swap_duplicate(entry))
 287                         break;
 288
 289                 /*
 290                  * Associate the page with swap entry in the swap cache.
 291                  * May fail (-EEXIST) if there is already a page associated
 292                  * with this entry in the swap cache: added by a racing
 293                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 294                  * re-using the just freed swap entry for an existing page.
 295                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 296                  */
 297                 SetPageLocked(new_page);
 298                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
 299                 if (!err) {
 300                         /*
 301                          * Initiate read into locked page and return.
 302                          */
 303                         lru_cache_add_active(new_page);
 304                         swap_readpage(NULL, new_page);
 305                         return new_page;
 306                 }
 307                 ClearPageLocked(new_page);
 308                 swap_free(entry);
 309         } while (err != -ENOMEM);
 310
 311         if (new_page)
 312                 page_cache_release(new_page);
 313         return found_page;
 314 }
 315
 316 /**
 317  * swapin_readahead - swap in pages in hope we need them soon
 318  * @entry: swap entry of this memory
 319  * @gfp_mask: memory allocation flags
 320  * @vma: user vma this address belongs to
 321  * @addr: target address for mempolicy
 322  *
 323  * Returns the struct page for entry and addr, after queueing swapin.
 324  *
 325  * Primitive swap readahead code. We simply read an aligned block of
 326  * (1 << page_cluster) entries in the swap area. This method is chosen
 327  * because it doesn't cost us any seek time.  We also make sure to queue
 328  * the 'original' request together with the readahead ones...
 329  *
 330  * This has been extended to use the NUMA policies from the mm triggering
 331  * the readahead.
 332  *
 333  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 334  */
 335 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 336                         struct vm_area_struct *vma, unsigned long addr)
 337 {
 338         int nr_pages;
 339         struct page *page;
 340         unsigned long offset;
 341         unsigned long end_offset;
 342
 343         /*
 344          * Get starting offset for readaround, and number of pages to read.
 345          * Adjust starting address by readbehind (for NUMA interleave case)?
 346          * No, it's very unlikely that swap layout would follow vma layout,
 347          * more likely that neighbouring swap pages came from the same node:
 348          * so use the same "addr" to choose the same node for each swap read.
 349          */
 350         nr_pages = valid_swaphandles(entry, &offset);
 351         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 352                 /* Ok, do the async read-ahead now */
 353                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 354                                                 gfp_mask, vma, addr);
 355                 if (!page)
 356                         break;
 357                 page_cache_release(page);
 358         }
 359         lru_add_drain();        /* Push any new pages onto the LRU now */
 360         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 361 }