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 #include <linux/page_cgroup.h>
  21
  22 #include <asm/pgtable.h>
  23
  24 /*
  25  * swapper_space is a fiction, retained to simplify the path through
  26  * vmscan's shrink_page_list, to make sync_page look nicer, and to allow
  27  * future use of radix_tree tags in the swap cache.
  28  */
  29 static const struct address_space_operations swap_aops = {
  30         .writepage      = swap_writepage,
  31         .sync_page      = block_sync_page,
  32         .set_page_dirty = __set_page_dirty_nobuffers,
  33         .migratepage    = migrate_page,
  34 };
  35
  36 static struct backing_dev_info swap_backing_dev_info = {
  37         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  38         .unplug_io_fn   = swap_unplug_io_fn,
  39 };
  40
  41 struct address_space swapper_space = {
  42         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  43         .tree_lock      = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock),
  44         .a_ops          = &swap_aops,
  45         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
  46         .backing_dev_info = &swap_backing_dev_info,
  47 };
  48
  49 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  50
  51 static struct {
  52         unsigned long add_total;
  53         unsigned long del_total;
  54         unsigned long find_success;
  55         unsigned long find_total;
  56 } swap_cache_info;
  57
  58 void show_swap_cache_info(void)
  59 {
  60         printk("%lu pages in swap cache\n", total_swapcache_pages);
  61         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  62                 swap_cache_info.add_total, swap_cache_info.del_total,
  63                 swap_cache_info.find_success, swap_cache_info.find_total);
  64         printk("Free swap  = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  65         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  66 }
  67
  68 /*
  69  * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  70  * but sets SwapCache flag and private instead of mapping and index.
  71  */
  72 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
  73 {
  74         int error;
  75
  76         VM_BUG_ON(!PageLocked(page));
  77         VM_BUG_ON(PageSwapCache(page));
  78         VM_BUG_ON(!PageSwapBacked(page));
  79
  80         error = radix_tree_preload(gfp_mask);
  81         if (!error) {
  82                 page_cache_get(page);
  83                 SetPageSwapCache(page);
  84                 set_page_private(page, entry.val);
  85
  86                 spin_lock_irq(&swapper_space.tree_lock);
  87                 error = radix_tree_insert(&swapper_space.page_tree,
  88                                                 entry.val, page);
  89                 if (likely(!error)) {
  90                         total_swapcache_pages++;
  91                         __inc_zone_page_state(page, NR_FILE_PAGES);
  92                         INC_CACHE_INFO(add_total);
  93                 }
  94                 spin_unlock_irq(&swapper_space.tree_lock);
  95                 radix_tree_preload_end();
  96
  97                 if (unlikely(error)) {
  98                         set_page_private(page, 0UL);
  99                         ClearPageSwapCache(page);
 100                         page_cache_release(page);
 101                 }
 102         }
 103         return error;
 104 }
 105
 106 /*
 107  * This must be called only on pages that have
 108  * been verified to be in the swap cache.
 109  */
 110 void __delete_from_swap_cache(struct page *page)
 111 {
 112         VM_BUG_ON(!PageLocked(page));
 113         VM_BUG_ON(!PageSwapCache(page));
 114         VM_BUG_ON(PageWriteback(page));
 115
 116         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 117         set_page_private(page, 0);
 118         ClearPageSwapCache(page);
 119         total_swapcache_pages--;
 120         __dec_zone_page_state(page, NR_FILE_PAGES);
 121         INC_CACHE_INFO(del_total);
 122 }
 123
 124 /**
 125  * add_to_swap - allocate swap space for a page
 126  * @page: page we want to move to swap
 127  *
 128  * Allocate swap space for the page and add the page to the
 129  * swap cache.  Caller needs to hold the page lock.
 130  */
 131 int add_to_swap(struct page *page)
 132 {
 133         swp_entry_t entry;
 134         int err;
 135
 136         VM_BUG_ON(!PageLocked(page));
 137         VM_BUG_ON(!PageUptodate(page));
 138
 139         for (;;) {
 140                 entry = get_swap_page();
 141                 if (!entry.val)
 142                         return 0;
 143
 144                 /*
 145                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 146                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 147                  * stops emergency reserves from being allocated.
 148                  *
 149                  * TODO: this could cause a theoretical memory reclaim
 150                  * deadlock in the swap out path.
 151                  */
 152                 /*
 153                  * Add it to the swap cache and mark it dirty
 154                  */
 155                 err = add_to_swap_cache(page, entry,
 156                                 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 157
 158                 switch (err) {
 159                 case 0:                         /* Success */
 160                         SetPageDirty(page);
 161                         return 1;
 162                 case -EEXIST:
 163                         /* Raced with "speculative" read_swap_cache_async */
 164                         swapcache_free(entry, NULL);
 165                         continue;
 166                 default:
 167                         /* -ENOMEM radix-tree allocation failure */
 168                         swapcache_free(entry, NULL);
 169                         return 0;
 170                 }
 171         }
 172 }
 173
 174 /*
 175  * This must be called only on pages that have
 176  * been verified to be in the swap cache and locked.
 177  * It will never put the page into the free list,
 178  * the caller has a reference on the page.
 179  */
 180 void delete_from_swap_cache(struct page *page)
 181 {
 182         swp_entry_t entry;
 183
 184         entry.val = page_private(page);
 185
 186         spin_lock_irq(&swapper_space.tree_lock);
 187         __delete_from_swap_cache(page);
 188         spin_unlock_irq(&swapper_space.tree_lock);
 189
 190         swapcache_free(entry, page);
 191         page_cache_release(page);
 192 }
 193
 194 /*
 195  * If we are the only user, then try to free up the swap cache.
 196  *
 197  * Its ok to check for PageSwapCache without the page lock
 198  * here because we are going to recheck again inside
 199  * try_to_free_swap() _with_ the lock.
 200  *                                      - Marcelo
 201  */
 202 static inline void free_swap_cache(struct page *page)
 203 {
 204         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 205                 try_to_free_swap(page);
 206                 unlock_page(page);
 207         }
 208 }
 209
 210 /*
 211  * Perform a free_page(), also freeing any swap cache associated with
 212  * this page if it is the last user of the page.
 213  */
 214 void free_page_and_swap_cache(struct page *page)
 215 {
 216         free_swap_cache(page);
 217         page_cache_release(page);
 218 }
 219
 220 /*
 221  * Passed an array of pages, drop them all from swapcache and then release
 222  * them.  They are removed from the LRU and freed if this is their last use.
 223  */
 224 void free_pages_and_swap_cache(struct page **pages, int nr)
 225 {
 226         struct page **pagep = pages;
 227
 228         lru_add_drain();
 229         while (nr) {
 230                 int todo = min(nr, PAGEVEC_SIZE);
 231                 int i;
 232
 233                 for (i = 0; i < todo; i++)
 234                         free_swap_cache(pagep[i]);
 235                 release_pages(pagep, todo, 0);
 236                 pagep += todo;
 237                 nr -= todo;
 238         }
 239 }
 240
 241 /*
 242  * Lookup a swap entry in the swap cache. A found page will be returned
 243  * unlocked and with its refcount incremented - we rely on the kernel
 244  * lock getting page table operations atomic even if we drop the page
 245  * lock before returning.
 246  */
 247 struct page * lookup_swap_cache(swp_entry_t entry)
 248 {
 249         struct page *page;
 250
 251         page = find_get_page(&swapper_space, entry.val);
 252
 253         if (page)
 254                 INC_CACHE_INFO(find_success);
 255
 256         INC_CACHE_INFO(find_total);
 257         return page;
 258 }
 259
 260 /*
 261  * Locate a page of swap in physical memory, reserving swap cache space
 262  * and reading the disk if it is not already cached.
 263  * A failure return means that either the page allocation failed or that
 264  * the swap entry is no longer in use.
 265  */
 266 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 267                         struct vm_area_struct *vma, unsigned long addr)
 268 {
 269         struct page *found_page, *new_page = NULL;
 270         int err;
 271
 272         do {
 273                 /*
 274                  * First check the swap cache.  Since this is normally
 275                  * called after lookup_swap_cache() failed, re-calling
 276                  * that would confuse statistics.
 277                  */
 278                 found_page = find_get_page(&swapper_space, entry.val);
 279                 if (found_page)
 280                         break;
 281
 282                 /*
 283                  * Get a new page to read into from swap.
 284                  */
 285                 if (!new_page) {
 286                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 287                         if (!new_page)
 288                                 break;          /* Out of memory */
 289                 }
 290
 291                 /*
 292                  * Swap entry may have been freed since our caller observed it.
 293                  */
 294                 err = swapcache_prepare(entry);
 295                 if (err == -EEXIST) /* seems racy */
 296                         continue;
 297                 if (err)           /* swp entry is obsolete ? */
 298                         break;
 299
 300                 /*
 301                  * Associate the page with swap entry in the swap cache.
 302                  * May fail (-EEXIST) if there is already a page associated
 303                  * with this entry in the swap cache: added by a racing
 304                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 305                  * re-using the just freed swap entry for an existing page.
 306                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 307                  */
 308                 __set_page_locked(new_page);
 309                 SetPageSwapBacked(new_page);
 310                 err = add_to_swap_cache(new_page, entry, gfp_mask & GFP_KERNEL);
 311                 if (likely(!err)) {
 312                         /*
 313                          * Initiate read into locked page and return.
 314                          */
 315                         lru_cache_add_anon(new_page);
 316                         swap_readpage(new_page);
 317                         return new_page;
 318                 }
 319                 ClearPageSwapBacked(new_page);
 320                 __clear_page_locked(new_page);
 321                 swapcache_free(entry, NULL);
 322         } while (err != -ENOMEM);
 323
 324         if (new_page)
 325                 page_cache_release(new_page);
 326         return found_page;
 327 }
 328
 329 /**
 330  * swapin_readahead - swap in pages in hope we need them soon
 331  * @entry: swap entry of this memory
 332  * @gfp_mask: memory allocation flags
 333  * @vma: user vma this address belongs to
 334  * @addr: target address for mempolicy
 335  *
 336  * Returns the struct page for entry and addr, after queueing swapin.
 337  *
 338  * Primitive swap readahead code. We simply read an aligned block of
 339  * (1 << page_cluster) entries in the swap area. This method is chosen
 340  * because it doesn't cost us any seek time.  We also make sure to queue
 341  * the 'original' request together with the readahead ones...
 342  *
 343  * This has been extended to use the NUMA policies from the mm triggering
 344  * the readahead.
 345  *
 346  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 347  */
 348 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 349                         struct vm_area_struct *vma, unsigned long addr)
 350 {
 351         int nr_pages;
 352         struct page *page;
 353         unsigned long offset;
 354         unsigned long end_offset;
 355
 356         /*
 357          * Get starting offset for readaround, and number of pages to read.
 358          * Adjust starting address by readbehind (for NUMA interleave case)?
 359          * No, it's very unlikely that swap layout would follow vma layout,
 360          * more likely that neighbouring swap pages came from the same node:
 361          * so use the same "addr" to choose the same node for each swap read.
 362          */
 363         nr_pages = valid_swaphandles(entry, &offset);
 364         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 365                 /* Ok, do the async read-ahead now */
 366                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 367                                                 gfp_mask, vma, addr);
 368                 if (!page)
 369                         break;
 370                 page_cache_release(page);
 371         }
 372         lru_add_drain();        /* Push any new pages onto the LRU now */
 373         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 374 }