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/mm.h>
  10 #include <linux/gfp.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/backing-dev.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/migrate.h>
  19 #include <linux/page_cgroup.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.
  26  */
  27 static const struct address_space_operations swap_aops = {
  28         .writepage      = swap_writepage,
  29         .set_page_dirty = __set_page_dirty_no_writeback,
  30         .migratepage    = migrate_page,
  31 };
  32
  33 static struct backing_dev_info swap_backing_dev_info = {
  34         .name           = "swap",
  35         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  36 };
  37
  38 struct address_space swapper_space = {
  39         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  40         .tree_lock      = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock),
  41         .a_ops          = &swap_aops,
  42         .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear),
  43         .backing_dev_info = &swap_backing_dev_info,
  44 };
  45
  46 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  47
  48 static struct {
  49         unsigned long add_total;
  50         unsigned long del_total;
  51         unsigned long find_success;
  52         unsigned long find_total;
  53 } swap_cache_info;
  54
  55 void show_swap_cache_info(void)
  56 {
  57         printk("%lu pages in swap cache\n", total_swapcache_pages);
  58         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  59                 swap_cache_info.add_total, swap_cache_info.del_total,
  60                 swap_cache_info.find_success, swap_cache_info.find_total);
  61         printk("Free swap  = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  62         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  63 }
  64
  65 /*
  66  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  67  * but sets SwapCache flag and private instead of mapping and index.
  68  */
  69 static int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  70 {
  71         int error;
  72
  73         VM_BUG_ON(!PageLocked(page));
  74         VM_BUG_ON(PageSwapCache(page));
  75         VM_BUG_ON(!PageSwapBacked(page));
  76
  77         page_cache_get(page);
  78         SetPageSwapCache(page);
  79         set_page_private(page, entry.val);
  80
  81         spin_lock_irq(&swapper_space.tree_lock);
  82         error = radix_tree_insert(&swapper_space.page_tree, entry.val, page);
  83         if (likely(!error)) {
  84                 total_swapcache_pages++;
  85                 __inc_zone_page_state(page, NR_FILE_PAGES);
  86                 INC_CACHE_INFO(add_total);
  87         }
  88         spin_unlock_irq(&swapper_space.tree_lock);
  89
  90         if (unlikely(error)) {
  91                 /*
  92                  * Only the context which have set SWAP_HAS_CACHE flag
  93                  * would call add_to_swap_cache().
  94                  * So add_to_swap_cache() doesn't returns -EEXIST.
  95                  */
  96                 VM_BUG_ON(error == -EEXIST);
  97                 set_page_private(page, 0UL);
  98                 ClearPageSwapCache(page);
  99                 page_cache_release(page);
 100         }
 101
 102         return error;
 103 }
 104
 105
 106 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 107 {
 108         int error;
 109
 110         error = radix_tree_preload(gfp_mask);
 111         if (!error) {
 112                 error = __add_to_swap_cache(page, entry);
 113                 radix_tree_preload_end();
 114         }
 115         return error;
 116 }
 117
 118 /*
 119  * This must be called only on pages that have
 120  * been verified to be in the swap cache.
 121  */
 122 void __delete_from_swap_cache(struct page *page)
 123 {
 124         VM_BUG_ON(!PageLocked(page));
 125         VM_BUG_ON(!PageSwapCache(page));
 126         VM_BUG_ON(PageWriteback(page));
 127
 128         radix_tree_delete(&swapper_space.page_tree, page_private(page));
 129         set_page_private(page, 0);
 130         ClearPageSwapCache(page);
 131         total_swapcache_pages--;
 132         __dec_zone_page_state(page, NR_FILE_PAGES);
 133         INC_CACHE_INFO(del_total);
 134 }
 135
 136 /**
 137  * add_to_swap - allocate swap space for a page
 138  * @page: page we want to move to swap
 139  *
 140  * Allocate swap space for the page and add the page to the
 141  * swap cache.  Caller needs to hold the page lock.
 142  */
 143 int add_to_swap(struct page *page)
 144 {
 145         swp_entry_t entry;
 146         int err;
 147
 148         VM_BUG_ON(!PageLocked(page));
 149         VM_BUG_ON(!PageUptodate(page));
 150
 151         entry = get_swap_page();
 152         if (!entry.val)
 153                 return 0;
 154
 155         if (unlikely(PageTransHuge(page)))
 156                 if (unlikely(split_huge_page(page))) {
 157                         swapcache_free(entry, NULL);
 158                         return 0;
 159                 }
 160
 161         /*
 162          * Radix-tree node allocations from PF_MEMALLOC contexts could
 163          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 164          * stops emergency reserves from being allocated.
 165          *
 166          * TODO: this could cause a theoretical memory reclaim
 167          * deadlock in the swap out path.
 168          */
 169         /*
 170          * Add it to the swap cache and mark it dirty
 171          */
 172         err = add_to_swap_cache(page, entry,
 173                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 174
 175         if (!err) {     /* Success */
 176                 SetPageDirty(page);
 177                 return 1;
 178         } else {        /* -ENOMEM radix-tree allocation failure */
 179                 /*
 180                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 181                  * clear SWAP_HAS_CACHE flag.
 182                  */
 183                 swapcache_free(entry, NULL);
 184                 return 0;
 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         spin_lock_irq(&swapper_space.tree_lock);
 201         __delete_from_swap_cache(page);
 202         spin_unlock_irq(&swapper_space.tree_lock);
 203
 204         swapcache_free(entry, page);
 205         page_cache_release(page);
 206 }
 207
 208 /*
 209  * If we are the only user, then try to free up the swap cache.
 210  *
 211  * Its ok to check for PageSwapCache without the page lock
 212  * here because we are going to recheck again inside
 213  * try_to_free_swap() _with_ the lock.
 214  *                                      - Marcelo
 215  */
 216 static inline void free_swap_cache(struct page *page)
 217 {
 218         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 219                 try_to_free_swap(page);
 220                 unlock_page(page);
 221         }
 222 }
 223
 224 /*
 225  * Perform a free_page(), also freeing any swap cache associated with
 226  * this page if it is the last user of the page.
 227  */
 228 void free_page_and_swap_cache(struct page *page)
 229 {
 230         free_swap_cache(page);
 231         page_cache_release(page);
 232 }
 233
 234 /*
 235  * Passed an array of pages, drop them all from swapcache and then release
 236  * them.  They are removed from the LRU and freed if this is their last use.
 237  */
 238 void free_pages_and_swap_cache(struct page **pages, int nr)
 239 {
 240         struct page **pagep = pages;
 241
 242         lru_add_drain();
 243         while (nr) {
 244                 int todo = min(nr, PAGEVEC_SIZE);
 245                 int i;
 246
 247                 for (i = 0; i < todo; i++)
 248                         free_swap_cache(pagep[i]);
 249                 release_pages(pagep, todo, 0);
 250                 pagep += todo;
 251                 nr -= todo;
 252         }
 253 }
 254
 255 /*
 256  * Lookup a swap entry in the swap cache. A found page will be returned
 257  * unlocked and with its refcount incremented - we rely on the kernel
 258  * lock getting page table operations atomic even if we drop the page
 259  * lock before returning.
 260  */
 261 struct page * lookup_swap_cache(swp_entry_t entry)
 262 {
 263         struct page *page;
 264
 265         page = find_get_page(&swapper_space, entry.val);
 266
 267         if (page)
 268                 INC_CACHE_INFO(find_success);
 269
 270         INC_CACHE_INFO(find_total);
 271         return page;
 272 }
 273
 274 /*
 275  * Locate a page of swap in physical memory, reserving swap cache space
 276  * and reading the disk if it is not already cached.
 277  * A failure return means that either the page allocation failed or that
 278  * the swap entry is no longer in use.
 279  */
 280 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 281                         struct vm_area_struct *vma, unsigned long addr)
 282 {
 283         struct page *found_page, *new_page = NULL;
 284         int err;
 285
 286         do {
 287                 /*
 288                  * First check the swap cache.  Since this is normally
 289                  * called after lookup_swap_cache() failed, re-calling
 290                  * that would confuse statistics.
 291                  */
 292                 found_page = find_get_page(&swapper_space, entry.val);
 293                 if (found_page)
 294                         break;
 295
 296                 /*
 297                  * Get a new page to read into from swap.
 298                  */
 299                 if (!new_page) {
 300                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 301                         if (!new_page)
 302                                 break;          /* Out of memory */
 303                 }
 304
 305                 /*
 306                  * call radix_tree_preload() while we can wait.
 307                  */
 308                 err = radix_tree_preload(gfp_mask & GFP_KERNEL);
 309                 if (err)
 310                         break;
 311
 312                 /*
 313                  * Swap entry may have been freed since our caller observed it.
 314                  */
 315                 err = swapcache_prepare(entry);
 316                 if (err == -EEXIST) {
 317                         radix_tree_preload_end();
 318                         /*
 319                          * We might race against get_swap_page() and stumble
 320                          * across a SWAP_HAS_CACHE swap_map entry whose page
 321                          * has not been brought into the swapcache yet, while
 322                          * the other end is scheduled away waiting on discard
 323                          * I/O completion at scan_swap_map().
 324                          *
 325                          * In order to avoid turning this transitory state
 326                          * into a permanent loop around this -EEXIST case
 327                          * if !CONFIG_PREEMPT and the I/O completion happens
 328                          * to be waiting on the CPU waitqueue where we are now
 329                          * busy looping, we just conditionally invoke the
 330                          * scheduler here, if there are some more important
 331                          * tasks to run.
 332                          */
 333                         cond_resched();
 334                         continue;
 335                 }
 336                 if (err) {              /* swp entry is obsolete ? */
 337                         radix_tree_preload_end();
 338                         break;
 339                 }
 340
 341                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 342                 __set_page_locked(new_page);
 343                 SetPageSwapBacked(new_page);
 344                 err = __add_to_swap_cache(new_page, entry);
 345                 if (likely(!err)) {
 346                         radix_tree_preload_end();
 347                         /*
 348                          * Initiate read into locked page and return.
 349                          */
 350                         lru_cache_add_anon(new_page);
 351                         swap_readpage(new_page);
 352                         return new_page;
 353                 }
 354                 radix_tree_preload_end();
 355                 ClearPageSwapBacked(new_page);
 356                 __clear_page_locked(new_page);
 357                 /*
 358                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 359                  * clear SWAP_HAS_CACHE flag.
 360                  */
 361                 swapcache_free(entry, NULL);
 362         } while (err != -ENOMEM);
 363
 364         if (new_page)
 365                 page_cache_release(new_page);
 366         return found_page;
 367 }
 368
 369 /**
 370  * swapin_readahead - swap in pages in hope we need them soon
 371  * @entry: swap entry of this memory
 372  * @gfp_mask: memory allocation flags
 373  * @vma: user vma this address belongs to
 374  * @addr: target address for mempolicy
 375  *
 376  * Returns the struct page for entry and addr, after queueing swapin.
 377  *
 378  * Primitive swap readahead code. We simply read an aligned block of
 379  * (1 << page_cluster) entries in the swap area. This method is chosen
 380  * because it doesn't cost us any seek time.  We also make sure to queue
 381  * the 'original' request together with the readahead ones...
 382  *
 383  * This has been extended to use the NUMA policies from the mm triggering
 384  * the readahead.
 385  *
 386  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 387  */
 388 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 389                         struct vm_area_struct *vma, unsigned long addr)
 390 {
 391         struct page *page;
 392         unsigned long offset = swp_offset(entry);
 393         unsigned long start_offset, end_offset;
 394         unsigned long mask = (1UL << page_cluster) - 1;
 395
 396         /* Read a page_cluster sized and aligned cluster around offset. */
 397         start_offset = offset & ~mask;
 398         end_offset = offset | mask;
 399         if (!start_offset)      /* First page is swap header. */
 400                 start_offset++;
 401
 402         for (offset = start_offset; offset <= end_offset ; offset++) {
 403                 /* Ok, do the async read-ahead now */
 404                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 405                                                 gfp_mask, vma, addr);
 406                 if (!page)
 407                         continue;
 408                 page_cache_release(page);
 409         }
 410         lru_add_drain();        /* Push any new pages onto the LRU now */
 411         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 412 }