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/gfp.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/init.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/buffer_head.h>
  18 #include <linux/backing-dev.h>
  19 #include <linux/pagevec.h>
  20 #include <linux/migrate.h>
  21 #include <linux/page_cgroup.h>
  22
  23 #include <asm/pgtable.h>
  24
  25 /*
  26  * swapper_space is a fiction, retained to simplify the path through
  27  * vmscan's shrink_page_list.
  28  */
  29 static const struct address_space_operations swap_aops = {
  30         .writepage      = swap_writepage,
  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         .name           = "swap",
  37         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  38 };
  39
  40 struct address_space swapper_space = {
  41         .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  42         .tree_lock      = __SPIN_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("%lu pages in swap cache\n", total_swapcache_pages);
  60         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  61                 swap_cache_info.add_total, swap_cache_info.del_total,
  62                 swap_cache_info.find_success, swap_cache_info.find_total);
  63         printk("Free swap  = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10));
  64         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  65 }
  66
  67 /*
  68  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  69  * but sets SwapCache flag and private instead of mapping and index.
  70  */
  71 static int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  72 {
  73         int error;
  74
  75         VM_BUG_ON(!PageLocked(page));
  76         VM_BUG_ON(PageSwapCache(page));
  77         VM_BUG_ON(!PageSwapBacked(page));
  78
  79         page_cache_get(page);
  80         SetPageSwapCache(page);
  81         set_page_private(page, entry.val);
  82
  83         spin_lock_irq(&swapper_space.tree_lock);
  84         error = radix_tree_insert(&swapper_space.page_tree, entry.val, page);
  85         if (likely(!error)) {
  86                 total_swapcache_pages++;
  87                 __inc_zone_page_state(page, NR_FILE_PAGES);
  88                 INC_CACHE_INFO(add_total);
  89         }
  90         spin_unlock_irq(&swapper_space.tree_lock);
  91
  92         if (unlikely(error)) {
  93                 /*
  94                  * Only the context which have set SWAP_HAS_CACHE flag
  95                  * would call add_to_swap_cache().
  96                  * So add_to_swap_cache() doesn't returns -EEXIST.
  97                  */
  98                 VM_BUG_ON(error == -EEXIST);
  99                 set_page_private(page, 0UL);
 100                 ClearPageSwapCache(page);
 101                 page_cache_release(page);
 102         }
 103
 104         return error;
 105 }
 106
 107
 108 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 109 {
 110         int error;
 111
 112         error = radix_tree_preload(gfp_mask);
 113         if (!error) {
 114                 error = __add_to_swap_cache(page, entry);
 115                 radix_tree_preload_end();
 116         }
 117         return error;
 118 }
 119
 120 /*
 121  * This must be called only on pages that have
 122  * been verified to be in the swap cache.
 123  */
 124 void __delete_from_swap_cache(struct page *page)
 125 {
 126         VM_BUG_ON(!PageLocked(page));
 127         VM_BUG_ON(!PageSwapCache(page));
 128         VM_BUG_ON(PageWriteback(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)
 146 {
 147         swp_entry_t entry;
 148         int err;
 149
 150         VM_BUG_ON(!PageLocked(page));
 151         VM_BUG_ON(!PageUptodate(page));
 152
 153         entry = get_swap_page();
 154         if (!entry.val)
 155                 return 0;
 156
 157         if (unlikely(PageTransHuge(page)))
 158                 if (unlikely(split_huge_page(page))) {
 159                         swapcache_free(entry, NULL);
 160                         return 0;
 161                 }
 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_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 176
 177         if (!err) {     /* Success */
 178                 SetPageDirty(page);
 179                 return 1;
 180         } else {        /* -ENOMEM radix-tree allocation failure */
 181                 /*
 182                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 183                  * clear SWAP_HAS_CACHE flag.
 184                  */
 185                 swapcache_free(entry, NULL);
 186                 return 0;
 187         }
 188 }
 189
 190 /*
 191  * This must be called only on pages that have
 192  * been verified to be in the swap cache and locked.
 193  * It will never put the page into the free list,
 194  * the caller has a reference on the page.
 195  */
 196 void delete_from_swap_cache(struct page *page)
 197 {
 198         swp_entry_t entry;
 199
 200         entry.val = page_private(page);
 201
 202         spin_lock_irq(&swapper_space.tree_lock);
 203         __delete_from_swap_cache(page);
 204         spin_unlock_irq(&swapper_space.tree_lock);
 205
 206         swapcache_free(entry, page);
 207         page_cache_release(page);
 208 }
 209
 210 /*
 211  * If we are the only user, then try to free up the swap cache.
 212  *
 213  * Its ok to check for PageSwapCache without the page lock
 214  * here because we are going to recheck again inside
 215  * try_to_free_swap() _with_ the lock.
 216  *                                      - Marcelo
 217  */
 218 static inline void free_swap_cache(struct page *page)
 219 {
 220         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 221                 try_to_free_swap(page);
 222                 unlock_page(page);
 223         }
 224 }
 225
 226 /*
 227  * Perform a free_page(), also freeing any swap cache associated with
 228  * this page if it is the last user of the page.
 229  */
 230 void free_page_and_swap_cache(struct page *page)
 231 {
 232         free_swap_cache(page);
 233         page_cache_release(page);
 234 }
 235
 236 /*
 237  * Passed an array of pages, drop them all from swapcache and then release
 238  * them.  They are removed from the LRU and freed if this is their last use.
 239  */
 240 void free_pages_and_swap_cache(struct page **pages, int nr)
 241 {
 242         struct page **pagep = pages;
 243
 244         lru_add_drain();
 245         while (nr) {
 246                 int todo = min(nr, PAGEVEC_SIZE);
 247                 int i;
 248
 249                 for (i = 0; i < todo; i++)
 250                         free_swap_cache(pagep[i]);
 251                 release_pages(pagep, todo, 0);
 252                 pagep += todo;
 253                 nr -= todo;
 254         }
 255 }
 256
 257 /*
 258  * Lookup a swap entry in the swap cache. A found page will be returned
 259  * unlocked and with its refcount incremented - we rely on the kernel
 260  * lock getting page table operations atomic even if we drop the page
 261  * lock before returning.
 262  */
 263 struct page * lookup_swap_cache(swp_entry_t entry)
 264 {
 265         struct page *page;
 266
 267         page = find_get_page(&swapper_space, entry.val);
 268
 269         if (page)
 270                 INC_CACHE_INFO(find_success);
 271
 272         INC_CACHE_INFO(find_total);
 273         return page;
 274 }
 275
 276 /*
 277  * Locate a page of swap in physical memory, reserving swap cache space
 278  * and reading the disk if it is not already cached.
 279  * A failure return means that either the page allocation failed or that
 280  * the swap entry is no longer in use.
 281  */
 282 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 283                         struct vm_area_struct *vma, unsigned long addr)
 284 {
 285         struct page *found_page, *new_page = NULL;
 286         int err;
 287
 288         do {
 289                 /*
 290                  * First check the swap cache.  Since this is normally
 291                  * called after lookup_swap_cache() failed, re-calling
 292                  * that would confuse statistics.
 293                  */
 294                 found_page = find_get_page(&swapper_space, entry.val);
 295                 if (found_page)
 296                         break;
 297
 298                 /*
 299                  * Get a new page to read into from swap.
 300                  */
 301                 if (!new_page) {
 302                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 303                         if (!new_page)
 304                                 break;          /* Out of memory */
 305                 }
 306
 307                 /*
 308                  * call radix_tree_preload() while we can wait.
 309                  */
 310                 err = radix_tree_preload(gfp_mask & GFP_KERNEL);
 311                 if (err)
 312                         break;
 313
 314                 /*
 315                  * Swap entry may have been freed since our caller observed it.
 316                  */
 317                 err = swapcache_prepare(entry);
 318                 if (err == -EEXIST) {   /* seems racy */
 319                         radix_tree_preload_end();
 320                         continue;
 321                 }
 322                 if (err) {              /* swp entry is obsolete ? */
 323                         radix_tree_preload_end();
 324                         break;
 325                 }
 326
 327                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 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                 /*
 344                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 345                  * clear SWAP_HAS_CACHE flag.
 346                  */
 347                 swapcache_free(entry, NULL);
 348         } while (err != -ENOMEM);
 349
 350         if (new_page)
 351                 page_cache_release(new_page);
 352         return found_page;
 353 }
 354
 355 /**
 356  * swapin_readahead - swap in pages in hope we need them soon
 357  * @entry: swap entry of this memory
 358  * @gfp_mask: memory allocation flags
 359  * @vma: user vma this address belongs to
 360  * @addr: target address for mempolicy
 361  *
 362  * Returns the struct page for entry and addr, after queueing swapin.
 363  *
 364  * Primitive swap readahead code. We simply read an aligned block of
 365  * (1 << page_cluster) entries in the swap area. This method is chosen
 366  * because it doesn't cost us any seek time.  We also make sure to queue
 367  * the 'original' request together with the readahead ones...
 368  *
 369  * This has been extended to use the NUMA policies from the mm triggering
 370  * the readahead.
 371  *
 372  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 373  */
 374 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 375                         struct vm_area_struct *vma, unsigned long addr)
 376 {
 377         int nr_pages;
 378         struct page *page;
 379         unsigned long offset;
 380         unsigned long end_offset;
 381
 382         /*
 383          * Get starting offset for readaround, and number of pages to read.
 384          * Adjust starting address by readbehind (for NUMA interleave case)?
 385          * No, it's very unlikely that swap layout would follow vma layout,
 386          * more likely that neighbouring swap pages came from the same node:
 387          * so use the same "addr" to choose the same node for each swap read.
 388          */
 389         nr_pages = valid_swaphandles(entry, &offset);
 390         for (end_offset = offset + nr_pages; offset < end_offset; offset++) {
 391                 /* Ok, do the async read-ahead now */
 392                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 393                                                 gfp_mask, vma, addr);
 394                 if (!page)
 395                         break;
 396                 page_cache_release(page);
 397         }
 398         lru_add_drain();        /* Push any new pages onto the LRU now */
 399         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 400 }