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 static int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  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         page_cache_get(page);
  81         SetPageSwapCache(page);
  82         set_page_private(page, entry.val);
  83
  84         spin_lock_irq(&swapper_space.tree_lock);
  85         error = radix_tree_insert(&swapper_space.page_tree, entry.val, page);
  86         if (likely(!error)) {
  87                 total_swapcache_pages++;
  88                 __inc_zone_page_state(page, NR_FILE_PAGES);
  89                 INC_CACHE_INFO(add_total);
  90         }
  91         spin_unlock_irq(&swapper_space.tree_lock);
  92
  93         if (unlikely(error)) {
  94                 set_page_private(page, 0UL);
  95                 ClearPageSwapCache(page);
  96                 page_cache_release(page);
  97         }
  98
  99         return error;
 100 }
 101
 102
 103 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 104 {
 105         int error;
 106
 107         error = radix_tree_preload(gfp_mask);
 108         if (!error) {
 109                 error = __add_to_swap_cache(page, entry);
 110                 radix_tree_preload_end();
 111         }
 112         return error;
 113 }
 114
 115 /*
 116  * This must be called only on pages that have
 117  * been verified to be in the swap cache.
 118  */
 119 void __delete_from_swap_cache(struct page *page)
 120 {
 121         VM_BUG_ON(!PageLocked(page));
 122         VM_BUG_ON(!PageSwapCache(page));
 123         VM_BUG_ON(PageWriteback(page));
 124
 125         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 126         set_page_private(page, 0);
 127         ClearPageSwapCache(page);
 128         total_swapcache_pages--;
 129         __dec_zone_page_state(page, NR_FILE_PAGES);
 130         INC_CACHE_INFO(del_total);
 131 }
 132
 133 /**
 134  * add_to_swap - allocate swap space for a page
 135  * @page: page we want to move to swap
 136  *
 137  * Allocate swap space for the page and add the page to the
 138  * swap cache.  Caller needs to hold the page lock.
 139  */
 140 int add_to_swap(struct page *page)
 141 {
 142         swp_entry_t entry;
 143         int err;
 144
 145         VM_BUG_ON(!PageLocked(page));
 146         VM_BUG_ON(!PageUptodate(page));
 147
 148         for (;;) {
 149                 entry = get_swap_page();
 150                 if (!entry.val)
 151                         return 0;
 152
 153                 /*
 154                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 155                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 156                  * stops emergency reserves from being allocated.
 157                  *
 158                  * TODO: this could cause a theoretical memory reclaim
 159                  * deadlock in the swap out path.
 160                  */
 161                 /*
 162                  * Add it to the swap cache and mark it dirty
 163                  */
 164                 err = add_to_swap_cache(page, entry,
 165                                 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 166
 167                 switch (err) {
 168                 case 0:                         /* Success */
 169                         SetPageDirty(page);
 170                         return 1;
 171                 case -EEXIST:
 172                         /* Raced with "speculative" read_swap_cache_async */
 173                         swapcache_free(entry, NULL);
 174                         continue;
 175                 default:
 176                         /* -ENOMEM radix-tree allocation failure */
 177                         swapcache_free(entry, NULL);
 178                         return 0;
 179                 }
 180         }
 181 }
 182
 183 /*
 184  * This must be called only on pages that have
 185  * been verified to be in the swap cache and locked.
 186  * It will never put the page into the free list,
 187  * the caller has a reference on the page.
 188  */
 189 void delete_from_swap_cache(struct page *page)
 190 {
 191         swp_entry_t entry;
 192
 193         entry.val = page_private(page);
 194
 195         spin_lock_irq(&swapper_space.tree_lock);
 196         __delete_from_swap_cache(page);
 197         spin_unlock_irq(&swapper_space.tree_lock);
 198
 199         swapcache_free(entry, page);
 200         page_cache_release(page);
 201 }
 202
 203 /*
 204  * If we are the only user, then try to free up the swap cache.
 205  *
 206  * Its ok to check for PageSwapCache without the page lock
 207  * here because we are going to recheck again inside
 208  * try_to_free_swap() _with_ the lock.
 209  *                                      - Marcelo
 210  */
 211 static inline void free_swap_cache(struct page *page)
 212 {
 213         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 214                 try_to_free_swap(page);
 215                 unlock_page(page);
 216         }
 217 }
 218
 219 /*
 220  * Perform a free_page(), also freeing any swap cache associated with
 221  * this page if it is the last user of the page.
 222  */
 223 void free_page_and_swap_cache(struct page *page)
 224 {
 225         free_swap_cache(page);
 226         page_cache_release(page);
 227 }
 228
 229 /*
 230  * Passed an array of pages, drop them all from swapcache and then release
 231  * them.  They are removed from the LRU and freed if this is their last use.
 232  */
 233 void free_pages_and_swap_cache(struct page **pages, int nr)
 234 {
 235         struct page **pagep = pages;
 236
 237         lru_add_drain();
 238         while (nr) {
 239                 int todo = min(nr, PAGEVEC_SIZE);
 240                 int i;
 241
 242                 for (i = 0; i < todo; i++)
 243                         free_swap_cache(pagep[i]);
 244                 release_pages(pagep, todo, 0);
 245                 pagep += todo;
 246                 nr -= todo;
 247         }
 248 }
 249
 250 /*
 251  * Lookup a swap entry in the swap cache. A found page will be returned
 252  * unlocked and with its refcount incremented - we rely on the kernel
 253  * lock getting page table operations atomic even if we drop the page
 254  * lock before returning.
 255  */
 256 struct page * lookup_swap_cache(swp_entry_t entry)
 257 {
 258         struct page *page;
 259
 260         page = find_get_page(&swapper_space, entry.val);
 261
 262         if (page)
 263                 INC_CACHE_INFO(find_success);
 264
 265         INC_CACHE_INFO(find_total);
 266         return page;
 267 }
 268
 269 /*
 270  * Locate a page of swap in physical memory, reserving swap cache space
 271  * and reading the disk if it is not already cached.
 272  * A failure return means that either the page allocation failed or that
 273  * the swap entry is no longer in use.
 274  */
 275 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 276                         struct vm_area_struct *vma, unsigned long addr)
 277 {
 278         struct page *found_page, *new_page = NULL;
 279         int err;
 280
 281         do {
 282                 /*
 283                  * First check the swap cache.  Since this is normally
 284                  * called after lookup_swap_cache() failed, re-calling
 285                  * that would confuse statistics.
 286                  */
 287                 found_page = find_get_page(&swapper_space, entry.val);
 288                 if (found_page)
 289                         break;
 290
 291                 /*
 292                  * Get a new page to read into from swap.
 293                  */
 294                 if (!new_page) {
 295                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 296                         if (!new_page)
 297                                 break;          /* Out of memory */
 298                 }
 299
 300                 /*
 301                  * call radix_tree_preload() while we can wait.
 302                  */
 303                 err = radix_tree_preload(gfp_mask & GFP_KERNEL);
 304                 if (err)
 305                         break;
 306
 307                 /*
 308                  * Swap entry may have been freed since our caller observed it.
 309                  */
 310                 err = swapcache_prepare(entry);
 311                 if (err == -EEXIST) {   /* seems racy */
 312                         radix_tree_preload_end();
 313                         continue;
 314                 }
 315                 if (err) {              /* swp entry is obsolete ? */
 316                         radix_tree_preload_end();
 317                         break;
 318                 }
 319
 320                 /*
 321                  * Associate the page with swap entry in the swap cache.
 322                  * May fail (-EEXIST) if there is already a page associated
 323                  * with this entry in the swap cache: added by a racing
 324                  * read_swap_cache_async, or add_to_swap or shmem_writepage
 325                  * re-using the just freed swap entry for an existing page.
 326                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 327                  */
 328                 __set_page_locked(new_page);
 329                 SetPageSwapBacked(new_page);
 330                 err = __add_to_swap_cache(new_page, entry);
 331                 if (likely(!err)) {
 332                         radix_tree_preload_end();
 333                         /*
 334                          * Initiate read into locked page and return.
 335                          */
 336                         lru_cache_add_anon(new_page);
 337                         swap_readpage(new_page);
 338                         return new_page;
 339                 }
 340                 radix_tree_preload_end();
 341                 ClearPageSwapBacked(new_page);
 342                 __clear_page_locked(new_page);
 343                 swapcache_free(entry, NULL);
 344         } while (err != -ENOMEM);
 345
 346         if (new_page)
 347                 page_cache_release(new_page);
 348         return found_page;
 349 }
 350
 351 /**
 352  * swapin_readahead - swap in pages in hope we need them soon
 353  * @entry: swap entry of this memory
 354  * @gfp_mask: memory allocation flags
 355  * @vma: user vma this address belongs to
 356  * @addr: target address for mempolicy
 357  *
 358  * Returns the struct page for entry and addr, after queueing swapin.
 359  *
 360  * Primitive swap readahead code. We simply read an aligned block of
 361  * (1 << page_cluster) entries in the swap area. This method is chosen
 362  * because it doesn't cost us any seek time.  We also make sure to queue
 363  * the 'original' request together with the readahead ones...
 364  *
 365  * This has been extended to use the NUMA policies from the mm triggering
 366  * the readahead.
 367  *
 368  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 369  */
 370 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 371                         struct vm_area_struct *vma, unsigned long addr)
 372 {
 373         int nr_pages;
 374         struct page *page;
 375         unsigned long offset;
 376         unsigned long end_offset;
 377
 378         /*
 379          * Get starting offset for readaround, and number of pages to read.
 380          * Adjust starting address by readbehind (for NUMA interleave case)?
 381          * No, it's very unlikely that swap layout would follow vma layout,
 382          * more likely that neighbouring swap pages came from the same node:
 383          * so use the same "addr" to choose the same node for each swap read.
 384          */
 385         nr_pages = valid_swaphandles(entry, &offset);
 386         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 387                 /* Ok, do the async read-ahead now */
 388                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 389                                                 gfp_mask, vma, addr);
 390                 if (!page)
 391                         break;
 392                 page_cache_release(page);
 393         }
 394         lru_add_drain();        /* Push any new pages onto the LRU now */
 395         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 396 }