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 static int add_to_swap_cache(struct page *page, swp_entry_t entry,
  98                                 gfp_t gfp_mask)
  99 {
 100         int error;
 101
 102         BUG_ON(PageLocked(page));
 103         if (!swap_duplicate(entry))
 104                 return -ENOENT;
 105
 106         SetPageLocked(page);
 107         error = __add_to_swap_cache(page, entry, gfp_mask & GFP_KERNEL);
 108         /*
 109          * Anon pages are already on the LRU, we don't run lru_cache_add here.
 110          */
 111         if (error) {
 112                 ClearPageLocked(page);
 113                 swap_free(entry);
 114                 return error;
 115         }
 116         return 0;
 117 }
 118
 119 /*
 120  * This must be called only on pages that have
 121  * been verified to be in the swap cache.
 122  */
 123 void __delete_from_swap_cache(struct page *page)
 124 {
 125         BUG_ON(!PageLocked(page));
 126         BUG_ON(!PageSwapCache(page));
 127         BUG_ON(PageWriteback(page));
 128         BUG_ON(PagePrivate(page));
 129
 130         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 131         set_page_private(page, 0);
 132         ClearPageSwapCache(page);
 133         total_swapcache_pages--;
 134         __dec_zone_page_state(page, NR_FILE_PAGES);
 135         INC_CACHE_INFO(del_total);
 136 }
 137
 138 /**
 139  * add_to_swap - allocate swap space for a page
 140  * @page: page we want to move to swap
 141  *
 142  * Allocate swap space for the page and add the page to the
 143  * swap cache.  Caller needs to hold the page lock.
 144  */
 145 int add_to_swap(struct page * page, gfp_t gfp_mask)
 146 {
 147         swp_entry_t entry;
 148         int err;
 149
 150         BUG_ON(!PageLocked(page));
 151
 152         for (;;) {
 153                 entry = get_swap_page();
 154                 if (!entry.val)
 155                         return 0;
 156
 157                 /*
 158                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 159                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 160                  * stops emergency reserves from being allocated.
 161                  *
 162                  * TODO: this could cause a theoretical memory reclaim
 163                  * deadlock in the swap out path.
 164                  */
 165                 /*
 166                  * Add it to the swap cache and mark it dirty
 167                  */
 168                 err = __add_to_swap_cache(page, entry,
 169                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
 170
 171                 switch (err) {
 172                 case 0:                         /* Success */
 173                         SetPageUptodate(page);
 174                         SetPageDirty(page);
 175                         return 1;
 176                 case -EEXIST:
 177                         /* Raced with "speculative" read_swap_cache_async */
 178                         swap_free(entry);
 179                         continue;
 180                 default:
 181                         /* -ENOMEM radix-tree allocation failure */
 182                         swap_free(entry);
 183                         return 0;
 184                 }
 185         }
 186 }
 187
 188 /*
 189  * This must be called only on pages that have
 190  * been verified to be in the swap cache and locked.
 191  * It will never put the page into the free list,
 192  * the caller has a reference on the page.
 193  */
 194 void delete_from_swap_cache(struct page *page)
 195 {
 196         swp_entry_t entry;
 197
 198         entry.val = page_private(page);
 199
 200         write_lock_irq(&swapper_space.tree_lock);
 201         __delete_from_swap_cache(page);
 202         write_unlock_irq(&swapper_space.tree_lock);
 203
 204         swap_free(entry);
 205         page_cache_release(page);
 206 }
 207
 208 /*
 209  * Strange swizzling function only for use by shmem_writepage
 210  */
 211 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 212 {
 213         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 214         if (!err) {
 215                 remove_from_page_cache(page);
 216                 page_cache_release(page);       /* pagecache ref */
 217                 if (!swap_duplicate(entry))
 218                         BUG();
 219                 SetPageDirty(page);
 220         }
 221         return err;
 222 }
 223
 224 /*
 225  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 226  */
 227 int move_from_swap_cache(struct page *page, unsigned long index,
 228                 struct address_space *mapping)
 229 {
 230         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 231         if (!err) {
 232                 delete_from_swap_cache(page);
 233                 /* shift page from clean_pages to dirty_pages list */
 234                 ClearPageDirty(page);
 235                 set_page_dirty(page);
 236         }
 237         return err;
 238 }
 239
 240 /*
 241  * If we are the only user, then try to free up the swap cache.
 242  *
 243  * Its ok to check for PageSwapCache without the page lock
 244  * here because we are going to recheck again inside
 245  * exclusive_swap_page() _with_ the lock.
 246  *                                      - Marcelo
 247  */
 248 static inline void free_swap_cache(struct page *page)
 249 {
 250         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 251                 remove_exclusive_swap_page(page);
 252                 unlock_page(page);
 253         }
 254 }
 255
 256 /*
 257  * Perform a free_page(), also freeing any swap cache associated with
 258  * this page if it is the last user of the page.
 259  */
 260 void free_page_and_swap_cache(struct page *page)
 261 {
 262         free_swap_cache(page);
 263         page_cache_release(page);
 264 }
 265
 266 /*
 267  * Passed an array of pages, drop them all from swapcache and then release
 268  * them.  They are removed from the LRU and freed if this is their last use.
 269  */
 270 void free_pages_and_swap_cache(struct page **pages, int nr)
 271 {
 272         struct page **pagep = pages;
 273
 274         lru_add_drain();
 275         while (nr) {
 276                 int todo = min(nr, PAGEVEC_SIZE);
 277                 int i;
 278
 279                 for (i = 0; i < todo; i++)
 280                         free_swap_cache(pagep[i]);
 281                 release_pages(pagep, todo, 0);
 282                 pagep += todo;
 283                 nr -= todo;
 284         }
 285 }
 286
 287 /*
 288  * Lookup a swap entry in the swap cache. A found page will be returned
 289  * unlocked and with its refcount incremented - we rely on the kernel
 290  * lock getting page table operations atomic even if we drop the page
 291  * lock before returning.
 292  */
 293 struct page * lookup_swap_cache(swp_entry_t entry)
 294 {
 295         struct page *page;
 296
 297         page = find_get_page(&swapper_space, entry.val);
 298
 299         if (page)
 300                 INC_CACHE_INFO(find_success);
 301
 302         INC_CACHE_INFO(find_total);
 303         return page;
 304 }
 305
 306 /*
 307  * Locate a page of swap in physical memory, reserving swap cache space
 308  * and reading the disk if it is not already cached.
 309  * A failure return means that either the page allocation failed or that
 310  * the swap entry is no longer in use.
 311  */
 312 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 313                         struct vm_area_struct *vma, unsigned long addr)
 314 {
 315         struct page *found_page, *new_page = NULL;
 316         int err;
 317
 318         do {
 319                 /*
 320                  * First check the swap cache.  Since this is normally
 321                  * called after lookup_swap_cache() failed, re-calling
 322                  * that would confuse statistics.
 323                  */
 324                 found_page = find_get_page(&swapper_space, entry.val);
 325                 if (found_page)
 326                         break;
 327
 328                 /*
 329                  * Get a new page to read into from swap.
 330                  */
 331                 if (!new_page) {
 332                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 333                         if (!new_page)
 334                                 break;          /* Out of memory */
 335                 }
 336
 337                 /*
 338                  * Associate the page with swap entry in the swap cache.
 339                  * May fail (-ENOENT) if swap entry has been freed since
 340                  * our caller observed it.  May fail (-EEXIST) if there
 341                  * is already a page associated with this entry in the
 342                  * swap cache: added by a racing read_swap_cache_async,
 343                  * or by try_to_swap_out (or shmem_writepage) re-using
 344                  * the just freed swap entry for an existing page.
 345                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 346                  */
 347                 err = add_to_swap_cache(new_page, entry, gfp_mask);
 348                 if (!err) {
 349                         /*
 350                          * Initiate read into locked page and return.
 351                          */
 352                         lru_cache_add_active(new_page);
 353                         swap_readpage(NULL, new_page);
 354                         return new_page;
 355                 }
 356         } while (err != -ENOENT && err != -ENOMEM);
 357
 358         if (new_page)
 359                 page_cache_release(new_page);
 360         return found_page;
 361 }
 362
 363 /**
 364  * swapin_readahead - swap in pages in hope we need them soon
 365  * @entry: swap entry of this memory
 366  * @vma: user vma this address belongs to
 367  * @addr: target address for mempolicy
 368  *
 369  * Returns the struct page for entry and addr, after queueing swapin.
 370  *
 371  * Primitive swap readahead code. We simply read an aligned block of
 372  * (1 << page_cluster) entries in the swap area. This method is chosen
 373  * because it doesn't cost us any seek time.  We also make sure to queue
 374  * the 'original' request together with the readahead ones...
 375  *
 376  * This has been extended to use the NUMA policies from the mm triggering
 377  * the readahead.
 378  *
 379  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 380  */
 381 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 382                         struct vm_area_struct *vma, unsigned long addr)
 383 {
 384         int nr_pages;
 385         struct page *page;
 386         unsigned long offset;
 387         unsigned long end_offset;
 388
 389         /*
 390          * Get starting offset for readaround, and number of pages to read.
 391          * Adjust starting address by readbehind (for NUMA interleave case)?
 392          * No, it's very unlikely that swap layout would follow vma layout,
 393          * more likely that neighbouring swap pages came from the same node:
 394          * so use the same "addr" to choose the same node for each swap read.
 395          */
 396         nr_pages = valid_swaphandles(entry, &offset);
 397         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 398                 /* Ok, do the async read-ahead now */
 399                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 400                                                 gfp_mask, vma, addr);
 401                 if (!page)
 402                         break;
 403                 page_cache_release(page);
 404         }
 405         lru_add_drain();        /* Push any new pages onto the LRU now */
 406         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 407 }