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/blkdev.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.
  27  */
  28 static const struct address_space_operations swap_aops = {
  29         .writepage      = swap_writepage,
  30         .set_page_dirty = swap_set_page_dirty,
  31 #ifdef CONFIG_MIGRATION
  32         .migratepage    = migrate_page,
  33 #endif
  34 };
  35
  36 static struct backing_dev_info swap_backing_dev_info = {
  37         .name           = "swap",
  38         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
  39 };
  40
  41 struct address_space swapper_spaces[MAX_SWAPFILES] = {
  42         [0 ... MAX_SWAPFILES - 1] = {
  43                 .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  44                 .i_mmap_writable = ATOMIC_INIT(0),
  45                 .a_ops          = &swap_aops,
  46                 .backing_dev_info = &swap_backing_dev_info,
  47         }
  48 };
  49
  50 #define INC_CACHE_INFO(x)       do { swap_cache_info.x++; } while (0)
  51
  52 static struct {
  53         unsigned long add_total;
  54         unsigned long del_total;
  55         unsigned long find_success;
  56         unsigned long find_total;
  57 } swap_cache_info;
  58
  59 unsigned long total_swapcache_pages(void)
  60 {
  61         int i;
  62         unsigned long ret = 0;
  63
  64         for (i = 0; i < MAX_SWAPFILES; i++)
  65                 ret += swapper_spaces[i].nrpages;
  66         return ret;
  67 }
  68
  69 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  70
  71 void show_swap_cache_info(void)
  72 {
  73         printk("%lu pages in swap cache\n", total_swapcache_pages());
  74         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  75                 swap_cache_info.add_total, swap_cache_info.del_total,
  76                 swap_cache_info.find_success, swap_cache_info.find_total);
  77         printk("Free swap  = %ldkB\n",
  78                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
  79         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  80 }
  81
  82 /*
  83  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  84  * but sets SwapCache flag and private instead of mapping and index.
  85  */
  86 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  87 {
  88         int error;
  89         struct address_space *address_space;
  90
  91         VM_BUG_ON_PAGE(!PageLocked(page), page);
  92         VM_BUG_ON_PAGE(PageSwapCache(page), page);
  93         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
  94
  95         page_cache_get(page);
  96         SetPageSwapCache(page);
  97         set_page_private(page, entry.val);
  98
  99         address_space = swap_address_space(entry);
 100         spin_lock_irq(&address_space->tree_lock);
 101         error = radix_tree_insert(&address_space->page_tree,
 102                                         entry.val, page);
 103         if (likely(!error)) {
 104                 address_space->nrpages++;
 105                 __inc_zone_page_state(page, NR_FILE_PAGES);
 106                 INC_CACHE_INFO(add_total);
 107         }
 108         spin_unlock_irq(&address_space->tree_lock);
 109
 110         if (unlikely(error)) {
 111                 /*
 112                  * Only the context which have set SWAP_HAS_CACHE flag
 113                  * would call add_to_swap_cache().
 114                  * So add_to_swap_cache() doesn't returns -EEXIST.
 115                  */
 116                 VM_BUG_ON(error == -EEXIST);
 117                 set_page_private(page, 0UL);
 118                 ClearPageSwapCache(page);
 119                 page_cache_release(page);
 120         }
 121
 122         return error;
 123 }
 124
 125
 126 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 127 {
 128         int error;
 129
 130         error = radix_tree_maybe_preload(gfp_mask);
 131         if (!error) {
 132                 error = __add_to_swap_cache(page, entry);
 133                 radix_tree_preload_end();
 134         }
 135         return error;
 136 }
 137
 138 /*
 139  * This must be called only on pages that have
 140  * been verified to be in the swap cache.
 141  */
 142 void __delete_from_swap_cache(struct page *page)
 143 {
 144         swp_entry_t entry;
 145         struct address_space *address_space;
 146
 147         VM_BUG_ON_PAGE(!PageLocked(page), page);
 148         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 149         VM_BUG_ON_PAGE(PageWriteback(page), page);
 150
 151         entry.val = page_private(page);
 152         address_space = swap_address_space(entry);
 153         radix_tree_delete(&address_space->page_tree, page_private(page));
 154         set_page_private(page, 0);
 155         ClearPageSwapCache(page);
 156         address_space->nrpages--;
 157         __dec_zone_page_state(page, NR_FILE_PAGES);
 158         INC_CACHE_INFO(del_total);
 159 }
 160
 161 /**
 162  * add_to_swap - allocate swap space for a page
 163  * @page: page we want to move to swap
 164  *
 165  * Allocate swap space for the page and add the page to the
 166  * swap cache.  Caller needs to hold the page lock.
 167  */
 168 int add_to_swap(struct page *page, struct list_head *list)
 169 {
 170         swp_entry_t entry;
 171         int err;
 172
 173         VM_BUG_ON_PAGE(!PageLocked(page), page);
 174         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 175
 176         entry = get_swap_page();
 177         if (!entry.val)
 178                 return 0;
 179
 180         if (unlikely(PageTransHuge(page)))
 181                 if (unlikely(split_huge_page_to_list(page, list))) {
 182                         swapcache_free(entry);
 183                         return 0;
 184                 }
 185
 186         /*
 187          * Radix-tree node allocations from PF_MEMALLOC contexts could
 188          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 189          * stops emergency reserves from being allocated.
 190          *
 191          * TODO: this could cause a theoretical memory reclaim
 192          * deadlock in the swap out path.
 193          */
 194         /*
 195          * Add it to the swap cache and mark it dirty
 196          */
 197         err = add_to_swap_cache(page, entry,
 198                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 199
 200         if (!err) {     /* Success */
 201                 SetPageDirty(page);
 202                 return 1;
 203         } else {        /* -ENOMEM radix-tree allocation failure */
 204                 /*
 205                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 206                  * clear SWAP_HAS_CACHE flag.
 207                  */
 208                 swapcache_free(entry);
 209                 return 0;
 210         }
 211 }
 212
 213 /*
 214  * This must be called only on pages that have
 215  * been verified to be in the swap cache and locked.
 216  * It will never put the page into the free list,
 217  * the caller has a reference on the page.
 218  */
 219 void delete_from_swap_cache(struct page *page)
 220 {
 221         swp_entry_t entry;
 222         struct address_space *address_space;
 223
 224         entry.val = page_private(page);
 225
 226         address_space = swap_address_space(entry);
 227         spin_lock_irq(&address_space->tree_lock);
 228         __delete_from_swap_cache(page);
 229         spin_unlock_irq(&address_space->tree_lock);
 230
 231         swapcache_free(entry);
 232         page_cache_release(page);
 233 }
 234
 235 /*
 236  * If we are the only user, then try to free up the swap cache.
 237  *
 238  * Its ok to check for PageSwapCache without the page lock
 239  * here because we are going to recheck again inside
 240  * try_to_free_swap() _with_ the lock.
 241  *                                      - Marcelo
 242  */
 243 static inline void free_swap_cache(struct page *page)
 244 {
 245         if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
 246                 try_to_free_swap(page);
 247                 unlock_page(page);
 248         }
 249 }
 250
 251 /*
 252  * Perform a free_page(), also freeing any swap cache associated with
 253  * this page if it is the last user of the page.
 254  */
 255 void free_page_and_swap_cache(struct page *page)
 256 {
 257         free_swap_cache(page);
 258         page_cache_release(page);
 259 }
 260
 261 /*
 262  * Passed an array of pages, drop them all from swapcache and then release
 263  * them.  They are removed from the LRU and freed if this is their last use.
 264  */
 265 void free_pages_and_swap_cache(struct page **pages, int nr)
 266 {
 267         struct page **pagep = pages;
 268         int i;
 269
 270         lru_add_drain();
 271         for (i = 0; i < nr; i++)
 272                 free_swap_cache(pagep[i]);
 273         release_pages(pagep, nr, false);
 274 }
 275
 276 /*
 277  * Lookup a swap entry in the swap cache. A found page will be returned
 278  * unlocked and with its refcount incremented - we rely on the kernel
 279  * lock getting page table operations atomic even if we drop the page
 280  * lock before returning.
 281  */
 282 struct page * lookup_swap_cache(swp_entry_t entry)
 283 {
 284         struct page *page;
 285
 286         page = find_get_page(swap_address_space(entry), entry.val);
 287
 288         if (page) {
 289                 INC_CACHE_INFO(find_success);
 290                 if (TestClearPageReadahead(page))
 291                         atomic_inc(&swapin_readahead_hits);
 292         }
 293
 294         INC_CACHE_INFO(find_total);
 295         return page;
 296 }
 297
 298 /*
 299  * Locate a page of swap in physical memory, reserving swap cache space
 300  * and reading the disk if it is not already cached.
 301  * A failure return means that either the page allocation failed or that
 302  * the swap entry is no longer in use.
 303  */
 304 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 305                         struct vm_area_struct *vma, unsigned long addr)
 306 {
 307         struct page *found_page, *new_page = NULL;
 308         int err;
 309
 310         do {
 311                 /*
 312                  * First check the swap cache.  Since this is normally
 313                  * called after lookup_swap_cache() failed, re-calling
 314                  * that would confuse statistics.
 315                  */
 316                 found_page = find_get_page(swap_address_space(entry),
 317                                         entry.val);
 318                 if (found_page)
 319                         break;
 320
 321                 /*
 322                  * Get a new page to read into from swap.
 323                  */
 324                 if (!new_page) {
 325                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 326                         if (!new_page)
 327                                 break;          /* Out of memory */
 328                 }
 329
 330                 /*
 331                  * call radix_tree_preload() while we can wait.
 332                  */
 333                 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
 334                 if (err)
 335                         break;
 336
 337                 /*
 338                  * Swap entry may have been freed since our caller observed it.
 339                  */
 340                 err = swapcache_prepare(entry);
 341                 if (err == -EEXIST) {
 342                         radix_tree_preload_end();
 343                         /*
 344                          * We might race against get_swap_page() and stumble
 345                          * across a SWAP_HAS_CACHE swap_map entry whose page
 346                          * has not been brought into the swapcache yet, while
 347                          * the other end is scheduled away waiting on discard
 348                          * I/O completion at scan_swap_map().
 349                          *
 350                          * In order to avoid turning this transitory state
 351                          * into a permanent loop around this -EEXIST case
 352                          * if !CONFIG_PREEMPT and the I/O completion happens
 353                          * to be waiting on the CPU waitqueue where we are now
 354                          * busy looping, we just conditionally invoke the
 355                          * scheduler here, if there are some more important
 356                          * tasks to run.
 357                          */
 358                         cond_resched();
 359                         continue;
 360                 }
 361                 if (err) {              /* swp entry is obsolete ? */
 362                         radix_tree_preload_end();
 363                         break;
 364                 }
 365
 366                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 367                 __set_page_locked(new_page);
 368                 SetPageSwapBacked(new_page);
 369                 err = __add_to_swap_cache(new_page, entry);
 370                 if (likely(!err)) {
 371                         radix_tree_preload_end();
 372                         /*
 373                          * Initiate read into locked page and return.
 374                          */
 375                         lru_cache_add_anon(new_page);
 376                         swap_readpage(new_page);
 377                         return new_page;
 378                 }
 379                 radix_tree_preload_end();
 380                 ClearPageSwapBacked(new_page);
 381                 __clear_page_locked(new_page);
 382                 /*
 383                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 384                  * clear SWAP_HAS_CACHE flag.
 385                  */
 386                 swapcache_free(entry);
 387         } while (err != -ENOMEM);
 388
 389         if (new_page)
 390                 page_cache_release(new_page);
 391         return found_page;
 392 }
 393
 394 static unsigned long swapin_nr_pages(unsigned long offset)
 395 {
 396         static unsigned long prev_offset;
 397         unsigned int pages, max_pages, last_ra;
 398         static atomic_t last_readahead_pages;
 399
 400         max_pages = 1 << ACCESS_ONCE(page_cluster);
 401         if (max_pages <= 1)
 402                 return 1;
 403
 404         /*
 405          * This heuristic has been found to work well on both sequential and
 406          * random loads, swapping to hard disk or to SSD: please don't ask
 407          * what the "+ 2" means, it just happens to work well, that's all.
 408          */
 409         pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
 410         if (pages == 2) {
 411                 /*
 412                  * We can have no readahead hits to judge by: but must not get
 413                  * stuck here forever, so check for an adjacent offset instead
 414                  * (and don't even bother to check whether swap type is same).
 415                  */
 416                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 417                         pages = 1;
 418                 prev_offset = offset;
 419         } else {
 420                 unsigned int roundup = 4;
 421                 while (roundup < pages)
 422                         roundup <<= 1;
 423                 pages = roundup;
 424         }
 425
 426         if (pages > max_pages)
 427                 pages = max_pages;
 428
 429         /* Don't shrink readahead too fast */
 430         last_ra = atomic_read(&last_readahead_pages) / 2;
 431         if (pages < last_ra)
 432                 pages = last_ra;
 433         atomic_set(&last_readahead_pages, pages);
 434
 435         return pages;
 436 }
 437
 438 /**
 439  * swapin_readahead - swap in pages in hope we need them soon
 440  * @entry: swap entry of this memory
 441  * @gfp_mask: memory allocation flags
 442  * @vma: user vma this address belongs to
 443  * @addr: target address for mempolicy
 444  *
 445  * Returns the struct page for entry and addr, after queueing swapin.
 446  *
 447  * Primitive swap readahead code. We simply read an aligned block of
 448  * (1 << page_cluster) entries in the swap area. This method is chosen
 449  * because it doesn't cost us any seek time.  We also make sure to queue
 450  * the 'original' request together with the readahead ones...
 451  *
 452  * This has been extended to use the NUMA policies from the mm triggering
 453  * the readahead.
 454  *
 455  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 456  */
 457 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 458                         struct vm_area_struct *vma, unsigned long addr)
 459 {
 460         struct page *page;
 461         unsigned long entry_offset = swp_offset(entry);
 462         unsigned long offset = entry_offset;
 463         unsigned long start_offset, end_offset;
 464         unsigned long mask;
 465         struct blk_plug plug;
 466
 467         mask = swapin_nr_pages(offset) - 1;
 468         if (!mask)
 469                 goto skip;
 470
 471         /* Read a page_cluster sized and aligned cluster around offset. */
 472         start_offset = offset & ~mask;
 473         end_offset = offset | mask;
 474         if (!start_offset)      /* First page is swap header. */
 475                 start_offset++;
 476
 477         blk_start_plug(&plug);
 478         for (offset = start_offset; offset <= end_offset ; offset++) {
 479                 /* Ok, do the async read-ahead now */
 480                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 481                                                 gfp_mask, vma, addr);
 482                 if (!page)
 483                         continue;
 484                 if (offset != entry_offset)
 485                         SetPageReadahead(page);
 486                 page_cache_release(page);
 487         }
 488         blk_finish_plug(&plug);
 489
 490         lru_add_drain();        /* Push any new pages onto the LRU now */
 491 skip:
 492         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 493 }