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
  21 #include <asm/pgtable.h>
  22 #include "internal.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 struct address_space swapper_spaces[MAX_SWAPFILES] = {
  37         [0 ... MAX_SWAPFILES - 1] = {
  38                 .page_tree      = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN),
  39                 .i_mmap_writable = ATOMIC_INIT(0),
  40                 .a_ops          = &swap_aops,
  41                 /* swap cache doesn't use writeback related tags */
  42                 .flags          = 1 << AS_NO_WRITEBACK_TAGS,
  43         }
  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 unsigned long total_swapcache_pages(void)
  56 {
  57         int i;
  58         unsigned long ret = 0;
  59
  60         for (i = 0; i < MAX_SWAPFILES; i++)
  61                 ret += swapper_spaces[i].nrpages;
  62         return ret;
  63 }
  64
  65 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
  66
  67 void show_swap_cache_info(void)
  68 {
  69         printk("%lu pages in swap cache\n", total_swapcache_pages());
  70         printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
  71                 swap_cache_info.add_total, swap_cache_info.del_total,
  72                 swap_cache_info.find_success, swap_cache_info.find_total);
  73         printk("Free swap  = %ldkB\n",
  74                 get_nr_swap_pages() << (PAGE_SHIFT - 10));
  75         printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
  76 }
  77
  78 /*
  79  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
  80  * but sets SwapCache flag and private instead of mapping and index.
  81  */
  82 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
  83 {
  84         int error;
  85         struct address_space *address_space;
  86
  87         VM_BUG_ON_PAGE(!PageLocked(page), page);
  88         VM_BUG_ON_PAGE(PageSwapCache(page), page);
  89         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
  90
  91         get_page(page);
  92         SetPageSwapCache(page);
  93         set_page_private(page, entry.val);
  94
  95         address_space = swap_address_space(entry);
  96         spin_lock_irq(&address_space->tree_lock);
  97         error = radix_tree_insert(&address_space->page_tree,
  98                                   swp_offset(entry), page);
  99         if (likely(!error)) {
 100                 address_space->nrpages++;
 101                 __inc_node_page_state(page, NR_FILE_PAGES);
 102                 INC_CACHE_INFO(add_total);
 103         }
 104         spin_unlock_irq(&address_space->tree_lock);
 105
 106         if (unlikely(error)) {
 107                 /*
 108                  * Only the context which have set SWAP_HAS_CACHE flag
 109                  * would call add_to_swap_cache().
 110                  * So add_to_swap_cache() doesn't returns -EEXIST.
 111                  */
 112                 VM_BUG_ON(error == -EEXIST);
 113                 set_page_private(page, 0UL);
 114                 ClearPageSwapCache(page);
 115                 put_page(page);
 116         }
 117
 118         return error;
 119 }
 120
 121
 122 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
 123 {
 124         int error;
 125
 126         error = radix_tree_maybe_preload(gfp_mask);
 127         if (!error) {
 128                 error = __add_to_swap_cache(page, entry);
 129                 radix_tree_preload_end();
 130         }
 131         return error;
 132 }
 133
 134 /*
 135  * This must be called only on pages that have
 136  * been verified to be in the swap cache.
 137  */
 138 void __delete_from_swap_cache(struct page *page)
 139 {
 140         swp_entry_t entry;
 141         struct address_space *address_space;
 142
 143         VM_BUG_ON_PAGE(!PageLocked(page), page);
 144         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
 145         VM_BUG_ON_PAGE(PageWriteback(page), page);
 146
 147         entry.val = page_private(page);
 148         address_space = swap_address_space(entry);
 149         radix_tree_delete(&address_space->page_tree, swp_offset(entry));
 150         set_page_private(page, 0);
 151         ClearPageSwapCache(page);
 152         address_space->nrpages--;
 153         __dec_node_page_state(page, NR_FILE_PAGES);
 154         INC_CACHE_INFO(del_total);
 155 }
 156
 157 /**
 158  * add_to_swap - allocate swap space for a page
 159  * @page: page we want to move to swap
 160  *
 161  * Allocate swap space for the page and add the page to the
 162  * swap cache.  Caller needs to hold the page lock.
 163  */
 164 int add_to_swap(struct page *page, struct list_head *list)
 165 {
 166         swp_entry_t entry;
 167         int err;
 168
 169         VM_BUG_ON_PAGE(!PageLocked(page), page);
 170         VM_BUG_ON_PAGE(!PageUptodate(page), page);
 171
 172         entry = get_swap_page();
 173         if (!entry.val)
 174                 return 0;
 175
 176         if (mem_cgroup_try_charge_swap(page, entry)) {
 177                 swapcache_free(entry);
 178                 return 0;
 179         }
 180
 181         if (unlikely(PageTransHuge(page)))
 182                 if (unlikely(split_huge_page_to_list(page, list))) {
 183                         swapcache_free(entry);
 184                         return 0;
 185                 }
 186
 187         /*
 188          * Radix-tree node allocations from PF_MEMALLOC contexts could
 189          * completely exhaust the page allocator. __GFP_NOMEMALLOC
 190          * stops emergency reserves from being allocated.
 191          *
 192          * TODO: this could cause a theoretical memory reclaim
 193          * deadlock in the swap out path.
 194          */
 195         /*
 196          * Add it to the swap cache.
 197          */
 198         err = add_to_swap_cache(page, entry,
 199                         __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
 200
 201         if (!err) {
 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         put_page(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         if (!is_huge_zero_page(page))
 259                 put_page(page);
 260 }
 261
 262 /*
 263  * Passed an array of pages, drop them all from swapcache and then release
 264  * them.  They are removed from the LRU and freed if this is their last use.
 265  */
 266 void free_pages_and_swap_cache(struct page **pages, int nr)
 267 {
 268         struct page **pagep = pages;
 269         int i;
 270
 271         lru_add_drain();
 272         for (i = 0; i < nr; i++)
 273                 free_swap_cache(pagep[i]);
 274         release_pages(pagep, nr, false);
 275 }
 276
 277 /*
 278  * Lookup a swap entry in the swap cache. A found page will be returned
 279  * unlocked and with its refcount incremented - we rely on the kernel
 280  * lock getting page table operations atomic even if we drop the page
 281  * lock before returning.
 282  */
 283 struct page * lookup_swap_cache(swp_entry_t entry)
 284 {
 285         struct page *page;
 286
 287         page = find_get_page(swap_address_space(entry), swp_offset(entry));
 288
 289         if (page) {
 290                 INC_CACHE_INFO(find_success);
 291                 if (TestClearPageReadahead(page))
 292                         atomic_inc(&swapin_readahead_hits);
 293         }
 294
 295         INC_CACHE_INFO(find_total);
 296         return page;
 297 }
 298
 299 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 300                         struct vm_area_struct *vma, unsigned long addr,
 301                         bool *new_page_allocated)
 302 {
 303         struct page *found_page, *new_page = NULL;
 304         struct address_space *swapper_space = swap_address_space(entry);
 305         int err;
 306         *new_page_allocated = false;
 307
 308         do {
 309                 /*
 310                  * First check the swap cache.  Since this is normally
 311                  * called after lookup_swap_cache() failed, re-calling
 312                  * that would confuse statistics.
 313                  */
 314                 found_page = find_get_page(swapper_space, swp_offset(entry));
 315                 if (found_page)
 316                         break;
 317
 318                 /*
 319                  * Get a new page to read into from swap.
 320                  */
 321                 if (!new_page) {
 322                         new_page = alloc_page_vma(gfp_mask, vma, addr);
 323                         if (!new_page)
 324                                 break;          /* Out of memory */
 325                 }
 326
 327                 /*
 328                  * call radix_tree_preload() while we can wait.
 329                  */
 330                 err = radix_tree_maybe_preload(gfp_mask & GFP_RECLAIM_MASK);
 331                 if (err)
 332                         break;
 333
 334                 /*
 335                  * Swap entry may have been freed since our caller observed it.
 336                  */
 337                 err = swapcache_prepare(entry);
 338                 if (err == -EEXIST) {
 339                         radix_tree_preload_end();
 340                         /*
 341                          * We might race against get_swap_page() and stumble
 342                          * across a SWAP_HAS_CACHE swap_map entry whose page
 343                          * has not been brought into the swapcache yet, while
 344                          * the other end is scheduled away waiting on discard
 345                          * I/O completion at scan_swap_map().
 346                          *
 347                          * In order to avoid turning this transitory state
 348                          * into a permanent loop around this -EEXIST case
 349                          * if !CONFIG_PREEMPT and the I/O completion happens
 350                          * to be waiting on the CPU waitqueue where we are now
 351                          * busy looping, we just conditionally invoke the
 352                          * scheduler here, if there are some more important
 353                          * tasks to run.
 354                          */
 355                         cond_resched();
 356                         continue;
 357                 }
 358                 if (err) {              /* swp entry is obsolete ? */
 359                         radix_tree_preload_end();
 360                         break;
 361                 }
 362
 363                 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
 364                 __SetPageLocked(new_page);
 365                 __SetPageSwapBacked(new_page);
 366                 err = __add_to_swap_cache(new_page, entry);
 367                 if (likely(!err)) {
 368                         radix_tree_preload_end();
 369                         /*
 370                          * Initiate read into locked page and return.
 371                          */
 372                         lru_cache_add_anon(new_page);
 373                         *new_page_allocated = true;
 374                         return new_page;
 375                 }
 376                 radix_tree_preload_end();
 377                 __ClearPageLocked(new_page);
 378                 /*
 379                  * add_to_swap_cache() doesn't return -EEXIST, so we can safely
 380                  * clear SWAP_HAS_CACHE flag.
 381                  */
 382                 swapcache_free(entry);
 383         } while (err != -ENOMEM);
 384
 385         if (new_page)
 386                 put_page(new_page);
 387         return found_page;
 388 }
 389
 390 /*
 391  * Locate a page of swap in physical memory, reserving swap cache space
 392  * and reading the disk if it is not already cached.
 393  * A failure return means that either the page allocation failed or that
 394  * the swap entry is no longer in use.
 395  */
 396 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
 397                         struct vm_area_struct *vma, unsigned long addr)
 398 {
 399         bool page_was_allocated;
 400         struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
 401                         vma, addr, &page_was_allocated);
 402
 403         if (page_was_allocated)
 404                 swap_readpage(retpage);
 405
 406         return retpage;
 407 }
 408
 409 static unsigned long swapin_nr_pages(unsigned long offset)
 410 {
 411         static unsigned long prev_offset;
 412         unsigned int pages, max_pages, last_ra;
 413         static atomic_t last_readahead_pages;
 414
 415         max_pages = 1 << READ_ONCE(page_cluster);
 416         if (max_pages <= 1)
 417                 return 1;
 418
 419         /*
 420          * This heuristic has been found to work well on both sequential and
 421          * random loads, swapping to hard disk or to SSD: please don't ask
 422          * what the "+ 2" means, it just happens to work well, that's all.
 423          */
 424         pages = atomic_xchg(&swapin_readahead_hits, 0) + 2;
 425         if (pages == 2) {
 426                 /*
 427                  * We can have no readahead hits to judge by: but must not get
 428                  * stuck here forever, so check for an adjacent offset instead
 429                  * (and don't even bother to check whether swap type is same).
 430                  */
 431                 if (offset != prev_offset + 1 && offset != prev_offset - 1)
 432                         pages = 1;
 433                 prev_offset = offset;
 434         } else {
 435                 unsigned int roundup = 4;
 436                 while (roundup < pages)
 437                         roundup <<= 1;
 438                 pages = roundup;
 439         }
 440
 441         if (pages > max_pages)
 442                 pages = max_pages;
 443
 444         /* Don't shrink readahead too fast */
 445         last_ra = atomic_read(&last_readahead_pages) / 2;
 446         if (pages < last_ra)
 447                 pages = last_ra;
 448         atomic_set(&last_readahead_pages, pages);
 449
 450         return pages;
 451 }
 452
 453 /**
 454  * swapin_readahead - swap in pages in hope we need them soon
 455  * @entry: swap entry of this memory
 456  * @gfp_mask: memory allocation flags
 457  * @vma: user vma this address belongs to
 458  * @addr: target address for mempolicy
 459  *
 460  * Returns the struct page for entry and addr, after queueing swapin.
 461  *
 462  * Primitive swap readahead code. We simply read an aligned block of
 463  * (1 << page_cluster) entries in the swap area. This method is chosen
 464  * because it doesn't cost us any seek time.  We also make sure to queue
 465  * the 'original' request together with the readahead ones...
 466  *
 467  * This has been extended to use the NUMA policies from the mm triggering
 468  * the readahead.
 469  *
 470  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
 471  */
 472 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
 473                         struct vm_area_struct *vma, unsigned long addr)
 474 {
 475         struct page *page;
 476         unsigned long entry_offset = swp_offset(entry);
 477         unsigned long offset = entry_offset;
 478         unsigned long start_offset, end_offset;
 479         unsigned long mask;
 480         struct blk_plug plug;
 481
 482         mask = swapin_nr_pages(offset) - 1;
 483         if (!mask)
 484                 goto skip;
 485
 486         /* Read a page_cluster sized and aligned cluster around offset. */
 487         start_offset = offset & ~mask;
 488         end_offset = offset | mask;
 489         if (!start_offset)      /* First page is swap header. */
 490                 start_offset++;
 491
 492         blk_start_plug(&plug);
 493         for (offset = start_offset; offset <= end_offset ; offset++) {
 494                 /* Ok, do the async read-ahead now */
 495                 page = read_swap_cache_async(swp_entry(swp_type(entry), offset),
 496                                                 gfp_mask, vma, addr);
 497                 if (!page)
 498                         continue;
 499                 if (offset != entry_offset)
 500                         SetPageReadahead(page);
 501                 put_page(page);
 502         }
 503         blk_finish_plug(&plug);
 504
 505         lru_add_drain();        /* Push any new pages onto the LRU now */
 506 skip:
 507         return read_swap_cache_async(entry, gfp_mask, vma, addr);
 508 }