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         unsigned long noent_race;
  56         unsigned long exist_race;
  57 } swap_cache_info;
  58
  59 void show_swap_cache_info(void)
  60 {
  61         printk("Swap cache: add %lu, delete %lu, find %lu/%lu, race %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                 swap_cache_info.noent_race, swap_cache_info.exist_race);
  65         printk("Free swap  = %lukB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  66         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  67 }
  68
  69 /*
  70  * __add_to_swap_cache resembles add_to_page_cache on swapper_space,
  71  * but sets SwapCache flag and private instead of mapping and index.
  72  */
  73 static int __add_to_swap_cache(struct page *page, swp_entry_t entry,
  74                                gfp_t gfp_mask)
  75 {
  76         int error;
  77
  78         BUG_ON(!PageLocked(page));
  79         BUG_ON(PageSwapCache(page));
  80         BUG_ON(PagePrivate(page));
  81         error = radix_tree_preload(gfp_mask);
  82         if (!error) {
  83                 write_lock_irq(&swapper_space.tree_lock);
  84                 error = radix_tree_insert(&swapper_space.page_tree,
  85                                                 entry.val, page);
  86                 if (!error) {
  87                         page_cache_get(page);
  88                         SetPageSwapCache(page);
  89                         set_page_private(page, entry.val);
  90                         total_swapcache_pages++;
  91                         __inc_zone_page_state(page, NR_FILE_PAGES);
  92                 }
  93                 write_unlock_irq(&swapper_space.tree_lock);
  94                 radix_tree_preload_end();
  95         }
  96         return error;
  97 }
  98
  99 static int add_to_swap_cache(struct page *page, swp_entry_t entry,
 100                                 gfp_t gfp_mask)
 101 {
 102         int error;
 103
 104         BUG_ON(PageLocked(page));
 105         if (!swap_duplicate(entry)) {
 106                 INC_CACHE_INFO(noent_race);
 107                 return -ENOENT;
 108         }
 109         SetPageLocked(page);
 110         error = __add_to_swap_cache(page, entry, gfp_mask & GFP_KERNEL);
 111         /*
 112          * Anon pages are already on the LRU, we don't run lru_cache_add here.
 113          */
 114         if (error) {
 115                 ClearPageLocked(page);
 116                 swap_free(entry);
 117                 if (error == -EEXIST)
 118                         INC_CACHE_INFO(exist_race);
 119                 return error;
 120         }
 121         INC_CACHE_INFO(add_total);
 122         return 0;
 123 }
 124
 125 /*
 126  * This must be called only on pages that have
 127  * been verified to be in the swap cache.
 128  */
 129 void __delete_from_swap_cache(struct page *page)
 130 {
 131         BUG_ON(!PageLocked(page));
 132         BUG_ON(!PageSwapCache(page));
 133         BUG_ON(PageWriteback(page));
 134         BUG_ON(PagePrivate(page));
 135
 136         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 137         set_page_private(page, 0);
 138         ClearPageSwapCache(page);
 139         total_swapcache_pages--;
 140         __dec_zone_page_state(page, NR_FILE_PAGES);
 141         INC_CACHE_INFO(del_total);
 142 }
 143
 144 /**
 145  * add_to_swap - allocate swap space for a page
 146  * @page: page we want to move to swap
 147  *
 148  * Allocate swap space for the page and add the page to the
 149  * swap cache.  Caller needs to hold the page lock.
 150  */
 151 int add_to_swap(struct page * page, gfp_t gfp_mask)
 152 {
 153         swp_entry_t entry;
 154         int err;
 155
 156         BUG_ON(!PageLocked(page));
 157
 158         for (;;) {
 159                 entry = get_swap_page();
 160                 if (!entry.val)
 161                         return 0;
 162
 163                 /*
 164                  * Radix-tree node allocations from PF_MEMALLOC contexts could
 165                  * completely exhaust the page allocator. __GFP_NOMEMALLOC
 166                  * stops emergency reserves from being allocated.
 167                  *
 168                  * TODO: this could cause a theoretical memory reclaim
 169                  * deadlock in the swap out path.
 170                  */
 171                 /*
 172                  * Add it to the swap cache and mark it dirty
 173                  */
 174                 err = __add_to_swap_cache(page, entry,
 175                                 gfp_mask|__GFP_NOMEMALLOC|__GFP_NOWARN);
 176
 177                 switch (err) {
 178                 case 0:                         /* Success */
 179                         SetPageUptodate(page);
 180                         SetPageDirty(page);
 181                         INC_CACHE_INFO(add_total);
 182                         return 1;
 183                 case -EEXIST:
 184                         /* Raced with "speculative" read_swap_cache_async */
 185                         INC_CACHE_INFO(exist_race);
 186                         swap_free(entry);
 187                         continue;
 188                 default:
 189                         /* -ENOMEM radix-tree allocation failure */
 190                         swap_free(entry);
 191                         return 0;
 192                 }
 193         }
 194 }
 195
 196 /*
 197  * This must be called only on pages that have
 198  * been verified to be in the swap cache and locked.
 199  * It will never put the page into the free list,
 200  * the caller has a reference on the page.
 201  */
 202 void delete_from_swap_cache(struct page *page)
 203 {
 204         swp_entry_t entry;
 205
 206         entry.val = page_private(page);
 207
 208         write_lock_irq(&swapper_space.tree_lock);
 209         __delete_from_swap_cache(page);
 210         write_unlock_irq(&swapper_space.tree_lock);
 211
 212         swap_free(entry);
 213         page_cache_release(page);
 214 }
 215
 216 /*
 217  * Strange swizzling function only for use by shmem_writepage
 218  */
 219 int move_to_swap_cache(struct page *page, swp_entry_t entry)
 220 {
 221         int err = __add_to_swap_cache(page, entry, GFP_ATOMIC);
 222         if (!err) {
 223                 remove_from_page_cache(page);
 224                 page_cache_release(page);       /* pagecache ref */
 225                 if (!swap_duplicate(entry))
 226                         BUG();
 227                 SetPageDirty(page);
 228                 INC_CACHE_INFO(add_total);
 229         } else if (err == -EEXIST)
 230                 INC_CACHE_INFO(exist_race);
 231         return err;
 232 }
 233
 234 /*
 235  * Strange swizzling function for shmem_getpage (and shmem_unuse)
 236  */
 237 int move_from_swap_cache(struct page *page, unsigned long index,
 238                 struct address_space *mapping)
 239 {
 240         int err = add_to_page_cache(page, mapping, index, GFP_ATOMIC);
 241         if (!err) {
 242                 delete_from_swap_cache(page);
 243                 /* shift page from clean_pages to dirty_pages list */
 244                 ClearPageDirty(page);
 245                 set_page_dirty(page);
 246         }
 247         return err;
 248 }
 249
 250 /*
 251  * If we are the only user, then try to free up the swap cache.
 252  *
 253  * Its ok to check for PageSwapCache without the page lock
 254  * here because we are going to recheck again inside
 255  * exclusive_swap_page() _with_ the lock.
 256  *                                      - Marcelo
 257  */
 258 static inline void free_swap_cache(struct page *page)
 259 {
 260         if (PageSwapCache(page) && !TestSetPageLocked(page)) {
 261                 remove_exclusive_swap_page(page);
 262                 unlock_page(page);
 263         }
 264 }
 265
 266 /*
 267  * Perform a free_page(), also freeing any swap cache associated with
 268  * this page if it is the last user of the page.
 269  */
 270 void free_page_and_swap_cache(struct page *page)
 271 {
 272         free_swap_cache(page);
 273         page_cache_release(page);
 274 }
 275
 276 /*
 277  * Passed an array of pages, drop them all from swapcache and then release
 278  * them.  They are removed from the LRU and freed if this is their last use.
 279  */
 280 void free_pages_and_swap_cache(struct page **pages, int nr)
 281 {
 282         struct page **pagep = pages;
 283
 284         lru_add_drain();
 285         while (nr) {
 286                 int todo = min(nr, PAGEVEC_SIZE);
 287                 int i;
 288
 289                 for (i = 0; i < todo; i++)
 290                         free_swap_cache(pagep[i]);
 291                 release_pages(pagep, todo, 0);
 292                 pagep += todo;
 293                 nr -= todo;
 294         }
 295 }
 296
 297 /*
 298  * Lookup a swap entry in the swap cache. A found page will be returned
 299  * unlocked and with its refcount incremented - we rely on the kernel
 300  * lock getting page table operations atomic even if we drop the page
 301  * lock before returning.
 302  */
 303 struct page * lookup_swap_cache(swp_entry_t entry)
 304 {
 305         struct page *page;
 306
 307         page = find_get_page(&swapper_space, entry.val);
 308
 309         if (page)
 310                 INC_CACHE_INFO(find_success);
 311
 312         INC_CACHE_INFO(find_total);
 313         return page;
 314 }
 315
 316 /*
 317  * Locate a page of swap in physical memory, reserving swap cache space
 318  * and reading the disk if it is not already cached.
 319  * A failure return means that either the page allocation failed or that
 320  * the swap entry is no longer in use.
 321  */
 322 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 323                         struct vm_area_struct *vma, unsigned long addr)
 324 {
 325         struct page *found_page, *new_page = NULL;
 326         int err;
 327
 328         do {
 329                 /*
 330                  * First check the swap cache.  Since this is normally
 331                  * called after lookup_swap_cache() failed, re-calling
 332                  * that would confuse statistics.
 333                  */
 334                 found_page = find_get_page(&swapper_space, entry.val);
 335                 if (found_page)
 336                         break;
 337
 338                 /*
 339                  * Get a new page to read into from swap.
 340                  */
 341                 if (!new_page) {
 342                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 343                         if (!new_page)
 344                                 break;          /* Out of memory */
 345                 }
 346
 347                 /*
 348                  * Associate the page with swap entry in the swap cache.
 349                  * May fail (-ENOENT) if swap entry has been freed since
 350                  * our caller observed it.  May fail (-EEXIST) if there
 351                  * is already a page associated with this entry in the
 352                  * swap cache: added by a racing read_swap_cache_async,
 353                  * or by try_to_swap_out (or shmem_writepage) re-using
 354                  * the just freed swap entry for an existing page.
 355                  * May fail (-ENOMEM) if radix-tree node allocation failed.
 356                  */
 357                 err = add_to_swap_cache(new_page, entry, gfp_mask);
 358                 if (!err) {
 359                         /*
 360                          * Initiate read into locked page and return.
 361                          */
 362                         lru_cache_add_active(new_page);
 363                         swap_readpage(NULL, new_page);
 364                         return new_page;
 365                 }
 366         } while (err != -ENOENT && err != -ENOMEM);
 367
 368         if (new_page)
 369                 page_cache_release(new_page);
 370         return found_page;
 371 }
 372
 373 /**
 374  * swapin_readahead - swap in pages in hope we need them soon
 375  * @entry: swap entry of this memory
 376  * @vma: user vma this address belongs to
 377  * @addr: target address for mempolicy
 378  *
 379  * Returns the struct page for entry and addr, after queueing swapin.
 380  *
 381  * Primitive swap readahead code. We simply read an aligned block of
 382  * (1 << page_cluster) entries in the swap area. This method is chosen
 383  * because it doesn't cost us any seek time.  We also make sure to queue
 384  * the 'original' request together with the readahead ones...
 385  *
 386  * This has been extended to use the NUMA policies from the mm triggering
 387  * the readahead.
 388  *
 389  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 390  */
 391 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 392                         struct vm_area_struct *vma, unsigned long addr)
 393 {
 394         int nr_pages;
 395         struct page *page;
 396         unsigned long offset;
 397         unsigned long end_offset;
 398
 399         /*
 400          * Get starting offset for readaround, and number of pages to read.
 401          * Adjust starting address by readbehind (for NUMA interleave case)?
 402          * No, it's very unlikely that swap layout would follow vma layout,
 403          * more likely that neighbouring swap pages came from the same node:
 404          * so use the same "addr" to choose the same node for each swap read.
 405          */
 406         nr_pages = valid_swaphandles(entry, &offset);
 407         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 408                 /* Ok, do the async read-ahead now */
 409                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 410                                                 gfp_mask, vma, addr);
 411                 if (!page)
 412                         break;
 413                 page_cache_release(page);
 414         }
 415         lru_add_drain();        /* Push any new pages onto the LRU now */
 416         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 417 }