mm/migrate.c

   1 /*
   2  * Memory Migration functionality - linux/mm/migration.c
   3  *
   4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5  *
   6  * Page migration was first developed in the context of the memory hotplug
   7  * project. The main authors of the migration code are:
   8  *
   9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10  * Hirokazu Takahashi <taka@valinux.co.jp>
  11  * Dave Hansen <haveblue@us.ibm.com>
  12  * Christoph Lameter
  13  */
  14
  15 #include <linux/migrate.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/swapops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/nsproxy.h>
  23 #include <linux/pagevec.h>
  24 #include <linux/rmap.h>
  25 #include <linux/topology.h>
  26 #include <linux/cpu.h>
  27 #include <linux/cpuset.h>
  28 #include <linux/writeback.h>
  29 #include <linux/mempolicy.h>
  30 #include <linux/vmalloc.h>
  31 #include <linux/security.h>
  32 #include <linux/memcontrol.h>
  33 #include <linux/syscalls.h>
  34
  35 #include "internal.h"
  36
  37 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  38
  39 /*
  40  * migrate_prep() needs to be called before we start compiling a list of pages
  41  * to be migrated using isolate_lru_page().
  42  */
  43 int migrate_prep(void)
  44 {
  45         /*
  46          * Clear the LRU lists so pages can be isolated.
  47          * Note that pages may be moved off the LRU after we have
  48          * drained them. Those pages will fail to migrate like other
  49          * pages that may be busy.
  50          */
  51         lru_add_drain_all();
  52
  53         return 0;
  54 }
  55
  56 /*
  57  * Add isolated pages on the list back to the LRU under page lock
  58  * to avoid leaking evictable pages back onto unevictable list.
  59  *
  60  * returns the number of pages put back.
  61  */
  62 int putback_lru_pages(struct list_head *l)
  63 {
  64         struct page *page;
  65         struct page *page2;
  66         int count = 0;
  67
  68         list_for_each_entry_safe(page, page2, l, lru) {
  69                 list_del(&page->lru);
  70                 putback_lru_page(page);
  71                 count++;
  72         }
  73         return count;
  74 }
  75
  76 /*
  77  * Restore a potential migration pte to a working pte entry
  78  */
  79 static void remove_migration_pte(struct vm_area_struct *vma,
  80                 struct page *old, struct page *new)
  81 {
  82         struct mm_struct *mm = vma->vm_mm;
  83         swp_entry_t entry;
  84         pgd_t *pgd;
  85         pud_t *pud;
  86         pmd_t *pmd;
  87         pte_t *ptep, pte;
  88         spinlock_t *ptl;
  89         unsigned long addr = page_address_in_vma(new, vma);
  90
  91         if (addr == -EFAULT)
  92                 return;
  93
  94         pgd = pgd_offset(mm, addr);
  95         if (!pgd_present(*pgd))
  96                 return;
  97
  98         pud = pud_offset(pgd, addr);
  99         if (!pud_present(*pud))
 100                 return;
 101
 102         pmd = pmd_offset(pud, addr);
 103         if (!pmd_present(*pmd))
 104                 return;
 105
 106         ptep = pte_offset_map(pmd, addr);
 107
 108         if (!is_swap_pte(*ptep)) {
 109                 pte_unmap(ptep);
 110                 return;
 111         }
 112
 113         ptl = pte_lockptr(mm, pmd);
 114         spin_lock(ptl);
 115         pte = *ptep;
 116         if (!is_swap_pte(pte))
 117                 goto out;
 118
 119         entry = pte_to_swp_entry(pte);
 120
 121         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 122                 goto out;
 123
 124         get_page(new);
 125         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 126         if (is_write_migration_entry(entry))
 127                 pte = pte_mkwrite(pte);
 128         flush_cache_page(vma, addr, pte_pfn(pte));
 129         set_pte_at(mm, addr, ptep, pte);
 130
 131         if (PageAnon(new))
 132                 page_add_anon_rmap(new, vma, addr);
 133         else
 134                 page_add_file_rmap(new);
 135
 136         /* No need to invalidate - it was non-present before */
 137         update_mmu_cache(vma, addr, pte);
 138
 139 out:
 140         pte_unmap_unlock(ptep, ptl);
 141 }
 142
 143 /*
 144  * Note that remove_file_migration_ptes will only work on regular mappings,
 145  * Nonlinear mappings do not use migration entries.
 146  */
 147 static void remove_file_migration_ptes(struct page *old, struct page *new)
 148 {
 149         struct vm_area_struct *vma;
 150         struct address_space *mapping = page_mapping(new);
 151         struct prio_tree_iter iter;
 152         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 153
 154         if (!mapping)
 155                 return;
 156
 157         spin_lock(&mapping->i_mmap_lock);
 158
 159         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 160                 remove_migration_pte(vma, old, new);
 161
 162         spin_unlock(&mapping->i_mmap_lock);
 163 }
 164
 165 /*
 166  * Must hold mmap_sem lock on at least one of the vmas containing
 167  * the page so that the anon_vma cannot vanish.
 168  */
 169 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 170 {
 171         struct anon_vma *anon_vma;
 172         struct vm_area_struct *vma;
 173         unsigned long mapping;
 174
 175         mapping = (unsigned long)new->mapping;
 176
 177         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 178                 return;
 179
 180         /*
 181          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 182          */
 183         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 184         spin_lock(&anon_vma->lock);
 185
 186         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 187                 remove_migration_pte(vma, old, new);
 188
 189         spin_unlock(&anon_vma->lock);
 190 }
 191
 192 /*
 193  * Get rid of all migration entries and replace them by
 194  * references to the indicated page.
 195  */
 196 static void remove_migration_ptes(struct page *old, struct page *new)
 197 {
 198         if (PageAnon(new))
 199                 remove_anon_migration_ptes(old, new);
 200         else
 201                 remove_file_migration_ptes(old, new);
 202 }
 203
 204 /*
 205  * Something used the pte of a page under migration. We need to
 206  * get to the page and wait until migration is finished.
 207  * When we return from this function the fault will be retried.
 208  *
 209  * This function is called from do_swap_page().
 210  */
 211 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 212                                 unsigned long address)
 213 {
 214         pte_t *ptep, pte;
 215         spinlock_t *ptl;
 216         swp_entry_t entry;
 217         struct page *page;
 218
 219         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 220         pte = *ptep;
 221         if (!is_swap_pte(pte))
 222                 goto out;
 223
 224         entry = pte_to_swp_entry(pte);
 225         if (!is_migration_entry(entry))
 226                 goto out;
 227
 228         page = migration_entry_to_page(entry);
 229
 230         /*
 231          * Once radix-tree replacement of page migration started, page_count
 232          * *must* be zero. And, we don't want to call wait_on_page_locked()
 233          * against a page without get_page().
 234          * So, we use get_page_unless_zero(), here. Even failed, page fault
 235          * will occur again.
 236          */
 237         if (!get_page_unless_zero(page))
 238                 goto out;
 239         pte_unmap_unlock(ptep, ptl);
 240         wait_on_page_locked(page);
 241         put_page(page);
 242         return;
 243 out:
 244         pte_unmap_unlock(ptep, ptl);
 245 }
 246
 247 /*
 248  * Replace the page in the mapping.
 249  *
 250  * The number of remaining references must be:
 251  * 1 for anonymous pages without a mapping
 252  * 2 for pages with a mapping
 253  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 254  */
 255 static int migrate_page_move_mapping(struct address_space *mapping,
 256                 struct page *newpage, struct page *page)
 257 {
 258         int expected_count;
 259         void **pslot;
 260
 261         if (!mapping) {
 262                 /* Anonymous page without mapping */
 263                 if (page_count(page) != 1)
 264                         return -EAGAIN;
 265                 return 0;
 266         }
 267
 268         spin_lock_irq(&mapping->tree_lock);
 269
 270         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 271                                         page_index(page));
 272
 273         expected_count = 2 + !!page_has_private(page);
 274         if (page_count(page) != expected_count ||
 275                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 276                 spin_unlock_irq(&mapping->tree_lock);
 277                 return -EAGAIN;
 278         }
 279
 280         if (!page_freeze_refs(page, expected_count)) {
 281                 spin_unlock_irq(&mapping->tree_lock);
 282                 return -EAGAIN;
 283         }
 284
 285         /*
 286          * Now we know that no one else is looking at the page.
 287          */
 288         get_page(newpage);      /* add cache reference */
 289         if (PageSwapCache(page)) {
 290                 SetPageSwapCache(newpage);
 291                 set_page_private(newpage, page_private(page));
 292         }
 293
 294         radix_tree_replace_slot(pslot, newpage);
 295
 296         page_unfreeze_refs(page, expected_count);
 297         /*
 298          * Drop cache reference from old page.
 299          * We know this isn't the last reference.
 300          */
 301         __put_page(page);
 302
 303         /*
 304          * If moved to a different zone then also account
 305          * the page for that zone. Other VM counters will be
 306          * taken care of when we establish references to the
 307          * new page and drop references to the old page.
 308          *
 309          * Note that anonymous pages are accounted for
 310          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 311          * are mapped to swap space.
 312          */
 313         __dec_zone_page_state(page, NR_FILE_PAGES);
 314         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 315
 316         spin_unlock_irq(&mapping->tree_lock);
 317
 318         return 0;
 319 }
 320
 321 /*
 322  * Copy the page to its new location
 323  */
 324 static void migrate_page_copy(struct page *newpage, struct page *page)
 325 {
 326         int anon;
 327
 328         copy_highpage(newpage, page);
 329
 330         if (PageError(page))
 331                 SetPageError(newpage);
 332         if (PageReferenced(page))
 333                 SetPageReferenced(newpage);
 334         if (PageUptodate(page))
 335                 SetPageUptodate(newpage);
 336         if (TestClearPageActive(page)) {
 337                 VM_BUG_ON(PageUnevictable(page));
 338                 SetPageActive(newpage);
 339         } else
 340                 unevictable_migrate_page(newpage, page);
 341         if (PageChecked(page))
 342                 SetPageChecked(newpage);
 343         if (PageMappedToDisk(page))
 344                 SetPageMappedToDisk(newpage);
 345
 346         if (PageDirty(page)) {
 347                 clear_page_dirty_for_io(page);
 348                 /*
 349                  * Want to mark the page and the radix tree as dirty, and
 350                  * redo the accounting that clear_page_dirty_for_io undid,
 351                  * but we can't use set_page_dirty because that function
 352                  * is actually a signal that all of the page has become dirty.
 353                  * Wheras only part of our page may be dirty.
 354                  */
 355                 __set_page_dirty_nobuffers(newpage);
 356         }
 357
 358         mlock_migrate_page(newpage, page);
 359
 360         ClearPageSwapCache(page);
 361         ClearPagePrivate(page);
 362         set_page_private(page, 0);
 363         /* page->mapping contains a flag for PageAnon() */
 364         anon = PageAnon(page);
 365         page->mapping = NULL;
 366
 367         /*
 368          * If any waiters have accumulated on the new page then
 369          * wake them up.
 370          */
 371         if (PageWriteback(newpage))
 372                 end_page_writeback(newpage);
 373 }
 374
 375 /************************************************************
 376  *                    Migration functions
 377  ***********************************************************/
 378
 379 /* Always fail migration. Used for mappings that are not movable */
 380 int fail_migrate_page(struct address_space *mapping,
 381                         struct page *newpage, struct page *page)
 382 {
 383         return -EIO;
 384 }
 385 EXPORT_SYMBOL(fail_migrate_page);
 386
 387 /*
 388  * Common logic to directly migrate a single page suitable for
 389  * pages that do not use PagePrivate/PagePrivate2.
 390  *
 391  * Pages are locked upon entry and exit.
 392  */
 393 int migrate_page(struct address_space *mapping,
 394                 struct page *newpage, struct page *page)
 395 {
 396         int rc;
 397
 398         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 399
 400         rc = migrate_page_move_mapping(mapping, newpage, page);
 401
 402         if (rc)
 403                 return rc;
 404
 405         migrate_page_copy(newpage, page);
 406         return 0;
 407 }
 408 EXPORT_SYMBOL(migrate_page);
 409
 410 #ifdef CONFIG_BLOCK
 411 /*
 412  * Migration function for pages with buffers. This function can only be used
 413  * if the underlying filesystem guarantees that no other references to "page"
 414  * exist.
 415  */
 416 int buffer_migrate_page(struct address_space *mapping,
 417                 struct page *newpage, struct page *page)
 418 {
 419         struct buffer_head *bh, *head;
 420         int rc;
 421
 422         if (!page_has_buffers(page))
 423                 return migrate_page(mapping, newpage, page);
 424
 425         head = page_buffers(page);
 426
 427         rc = migrate_page_move_mapping(mapping, newpage, page);
 428
 429         if (rc)
 430                 return rc;
 431
 432         bh = head;
 433         do {
 434                 get_bh(bh);
 435                 lock_buffer(bh);
 436                 bh = bh->b_this_page;
 437
 438         } while (bh != head);
 439
 440         ClearPagePrivate(page);
 441         set_page_private(newpage, page_private(page));
 442         set_page_private(page, 0);
 443         put_page(page);
 444         get_page(newpage);
 445
 446         bh = head;
 447         do {
 448                 set_bh_page(bh, newpage, bh_offset(bh));
 449                 bh = bh->b_this_page;
 450
 451         } while (bh != head);
 452
 453         SetPagePrivate(newpage);
 454
 455         migrate_page_copy(newpage, page);
 456
 457         bh = head;
 458         do {
 459                 unlock_buffer(bh);
 460                 put_bh(bh);
 461                 bh = bh->b_this_page;
 462
 463         } while (bh != head);
 464
 465         return 0;
 466 }
 467 EXPORT_SYMBOL(buffer_migrate_page);
 468 #endif
 469
 470 /*
 471  * Writeback a page to clean the dirty state
 472  */
 473 static int writeout(struct address_space *mapping, struct page *page)
 474 {
 475         struct writeback_control wbc = {
 476                 .sync_mode = WB_SYNC_NONE,
 477                 .nr_to_write = 1,
 478                 .range_start = 0,
 479                 .range_end = LLONG_MAX,
 480                 .nonblocking = 1,
 481                 .for_reclaim = 1
 482         };
 483         int rc;
 484
 485         if (!mapping->a_ops->writepage)
 486                 /* No write method for the address space */
 487                 return -EINVAL;
 488
 489         if (!clear_page_dirty_for_io(page))
 490                 /* Someone else already triggered a write */
 491                 return -EAGAIN;
 492
 493         /*
 494          * A dirty page may imply that the underlying filesystem has
 495          * the page on some queue. So the page must be clean for
 496          * migration. Writeout may mean we loose the lock and the
 497          * page state is no longer what we checked for earlier.
 498          * At this point we know that the migration attempt cannot
 499          * be successful.
 500          */
 501         remove_migration_ptes(page, page);
 502
 503         rc = mapping->a_ops->writepage(page, &wbc);
 504
 505         if (rc != AOP_WRITEPAGE_ACTIVATE)
 506                 /* unlocked. Relock */
 507                 lock_page(page);
 508
 509         return (rc < 0) ? -EIO : -EAGAIN;
 510 }
 511
 512 /*
 513  * Default handling if a filesystem does not provide a migration function.
 514  */
 515 static int fallback_migrate_page(struct address_space *mapping,
 516         struct page *newpage, struct page *page)
 517 {
 518         if (PageDirty(page))
 519                 return writeout(mapping, page);
 520
 521         /*
 522          * Buffers may be managed in a filesystem specific way.
 523          * We must have no buffers or drop them.
 524          */
 525         if (page_has_private(page) &&
 526             !try_to_release_page(page, GFP_KERNEL))
 527                 return -EAGAIN;
 528
 529         return migrate_page(mapping, newpage, page);
 530 }
 531
 532 /*
 533  * Move a page to a newly allocated page
 534  * The page is locked and all ptes have been successfully removed.
 535  *
 536  * The new page will have replaced the old page if this function
 537  * is successful.
 538  *
 539  * Return value:
 540  *   < 0 - error code
 541  *  == 0 - success
 542  */
 543 static int move_to_new_page(struct page *newpage, struct page *page)
 544 {
 545         struct address_space *mapping;
 546         int rc;
 547
 548         /*
 549          * Block others from accessing the page when we get around to
 550          * establishing additional references. We are the only one
 551          * holding a reference to the new page at this point.
 552          */
 553         if (!trylock_page(newpage))
 554                 BUG();
 555
 556         /* Prepare mapping for the new page.*/
 557         newpage->index = page->index;
 558         newpage->mapping = page->mapping;
 559         if (PageSwapBacked(page))
 560                 SetPageSwapBacked(newpage);
 561
 562         mapping = page_mapping(page);
 563         if (!mapping)
 564                 rc = migrate_page(mapping, newpage, page);
 565         else if (mapping->a_ops->migratepage)
 566                 /*
 567                  * Most pages have a mapping and most filesystems
 568                  * should provide a migration function. Anonymous
 569                  * pages are part of swap space which also has its
 570                  * own migration function. This is the most common
 571                  * path for page migration.
 572                  */
 573                 rc = mapping->a_ops->migratepage(mapping,
 574                                                 newpage, page);
 575         else
 576                 rc = fallback_migrate_page(mapping, newpage, page);
 577
 578         if (!rc) {
 579                 remove_migration_ptes(page, newpage);
 580         } else
 581                 newpage->mapping = NULL;
 582
 583         unlock_page(newpage);
 584
 585         return rc;
 586 }
 587
 588 /*
 589  * Obtain the lock on page, remove all ptes and migrate the page
 590  * to the newly allocated page in newpage.
 591  */
 592 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 593                         struct page *page, int force)
 594 {
 595         int rc = 0;
 596         int *result = NULL;
 597         struct page *newpage = get_new_page(page, private, &result);
 598         int rcu_locked = 0;
 599         int charge = 0;
 600         struct mem_cgroup *mem;
 601
 602         if (!newpage)
 603                 return -ENOMEM;
 604
 605         if (page_count(page) == 1) {
 606                 /* page was freed from under us. So we are done. */
 607                 goto move_newpage;
 608         }
 609
 610         /* prepare cgroup just returns 0 or -ENOMEM */
 611         rc = -EAGAIN;
 612
 613         if (!trylock_page(page)) {
 614                 if (!force)
 615                         goto move_newpage;
 616                 lock_page(page);
 617         }
 618
 619         /* charge against new page */
 620         charge = mem_cgroup_prepare_migration(page, &mem);
 621         if (charge == -ENOMEM) {
 622                 rc = -ENOMEM;
 623                 goto unlock;
 624         }
 625         BUG_ON(charge);
 626
 627         if (PageWriteback(page)) {
 628                 if (!force)
 629                         goto uncharge;
 630                 wait_on_page_writeback(page);
 631         }
 632         /*
 633          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 634          * we cannot notice that anon_vma is freed while we migrates a page.
 635          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 636          * of migration. File cache pages are no problem because of page_lock()
 637          * File Caches may use write_page() or lock_page() in migration, then,
 638          * just care Anon page here.
 639          */
 640         if (PageAnon(page)) {
 641                 rcu_read_lock();
 642                 rcu_locked = 1;
 643         }
 644
 645         /*
 646          * Corner case handling:
 647          * 1. When a new swap-cache page is read into, it is added to the LRU
 648          * and treated as swapcache but it has no rmap yet.
 649          * Calling try_to_unmap() against a page->mapping==NULL page will
 650          * trigger a BUG.  So handle it here.
 651          * 2. An orphaned page (see truncate_complete_page) might have
 652          * fs-private metadata. The page can be picked up due to memory
 653          * offlining.  Everywhere else except page reclaim, the page is
 654          * invisible to the vm, so the page can not be migrated.  So try to
 655          * free the metadata, so the page can be freed.
 656          */
 657         if (!page->mapping) {
 658                 if (!PageAnon(page) && page_has_private(page)) {
 659                         /*
 660                          * Go direct to try_to_free_buffers() here because
 661                          * a) that's what try_to_release_page() would do anyway
 662                          * b) we may be under rcu_read_lock() here, so we can't
 663                          *    use GFP_KERNEL which is what try_to_release_page()
 664                          *    needs to be effective.
 665                          */
 666                         try_to_free_buffers(page);
 667                 }
 668                 goto rcu_unlock;
 669         }
 670
 671         /* Establish migration ptes or remove ptes */
 672         try_to_unmap(page, 1);
 673
 674         if (!page_mapped(page))
 675                 rc = move_to_new_page(newpage, page);
 676
 677         if (rc)
 678                 remove_migration_ptes(page, page);
 679 rcu_unlock:
 680         if (rcu_locked)
 681                 rcu_read_unlock();
 682 uncharge:
 683         if (!charge)
 684                 mem_cgroup_end_migration(mem, page, newpage);
 685 unlock:
 686         unlock_page(page);
 687
 688         if (rc != -EAGAIN) {
 689                 /*
 690                  * A page that has been migrated has all references
 691                  * removed and will be freed. A page that has not been
 692                  * migrated will have kepts its references and be
 693                  * restored.
 694                  */
 695                 list_del(&page->lru);
 696                 putback_lru_page(page);
 697         }
 698
 699 move_newpage:
 700
 701         /*
 702          * Move the new page to the LRU. If migration was not successful
 703          * then this will free the page.
 704          */
 705         putback_lru_page(newpage);
 706
 707         if (result) {
 708                 if (rc)
 709                         *result = rc;
 710                 else
 711                         *result = page_to_nid(newpage);
 712         }
 713         return rc;
 714 }
 715
 716 /*
 717  * migrate_pages
 718  *
 719  * The function takes one list of pages to migrate and a function
 720  * that determines from the page to be migrated and the private data
 721  * the target of the move and allocates the page.
 722  *
 723  * The function returns after 10 attempts or if no pages
 724  * are movable anymore because to has become empty
 725  * or no retryable pages exist anymore. All pages will be
 726  * returned to the LRU or freed.
 727  *
 728  * Return: Number of pages not migrated or error code.
 729  */
 730 int migrate_pages(struct list_head *from,
 731                 new_page_t get_new_page, unsigned long private)
 732 {
 733         int retry = 1;
 734         int nr_failed = 0;
 735         int pass = 0;
 736         struct page *page;
 737         struct page *page2;
 738         int swapwrite = current->flags & PF_SWAPWRITE;
 739         int rc;
 740
 741         if (!swapwrite)
 742                 current->flags |= PF_SWAPWRITE;
 743
 744         for(pass = 0; pass < 10 && retry; pass++) {
 745                 retry = 0;
 746
 747                 list_for_each_entry_safe(page, page2, from, lru) {
 748                         cond_resched();
 749
 750                         rc = unmap_and_move(get_new_page, private,
 751                                                 page, pass > 2);
 752
 753                         switch(rc) {
 754                         case -ENOMEM:
 755                                 goto out;
 756                         case -EAGAIN:
 757                                 retry++;
 758                                 break;
 759                         case 0:
 760                                 break;
 761                         default:
 762                                 /* Permanent failure */
 763                                 nr_failed++;
 764                                 break;
 765                         }
 766                 }
 767         }
 768         rc = 0;
 769 out:
 770         if (!swapwrite)
 771                 current->flags &= ~PF_SWAPWRITE;
 772
 773         putback_lru_pages(from);
 774
 775         if (rc)
 776                 return rc;
 777
 778         return nr_failed + retry;
 779 }
 780
 781 #ifdef CONFIG_NUMA
 782 /*
 783  * Move a list of individual pages
 784  */
 785 struct page_to_node {
 786         unsigned long addr;
 787         struct page *page;
 788         int node;
 789         int status;
 790 };
 791
 792 static struct page *new_page_node(struct page *p, unsigned long private,
 793                 int **result)
 794 {
 795         struct page_to_node *pm = (struct page_to_node *)private;
 796
 797         while (pm->node != MAX_NUMNODES && pm->page != p)
 798                 pm++;
 799
 800         if (pm->node == MAX_NUMNODES)
 801                 return NULL;
 802
 803         *result = &pm->status;
 804
 805         return alloc_pages_node(pm->node,
 806                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 807 }
 808
 809 /*
 810  * Move a set of pages as indicated in the pm array. The addr
 811  * field must be set to the virtual address of the page to be moved
 812  * and the node number must contain a valid target node.
 813  * The pm array ends with node = MAX_NUMNODES.
 814  */
 815 static int do_move_page_to_node_array(struct mm_struct *mm,
 816                                       struct page_to_node *pm,
 817                                       int migrate_all)
 818 {
 819         int err;
 820         struct page_to_node *pp;
 821         LIST_HEAD(pagelist);
 822
 823         migrate_prep();
 824         down_read(&mm->mmap_sem);
 825
 826         /*
 827          * Build a list of pages to migrate
 828          */
 829         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 830                 struct vm_area_struct *vma;
 831                 struct page *page;
 832
 833                 err = -EFAULT;
 834                 vma = find_vma(mm, pp->addr);
 835                 if (!vma || !vma_migratable(vma))
 836                         goto set_status;
 837
 838                 page = follow_page(vma, pp->addr, FOLL_GET);
 839
 840                 err = PTR_ERR(page);
 841                 if (IS_ERR(page))
 842                         goto set_status;
 843
 844                 err = -ENOENT;
 845                 if (!page)
 846                         goto set_status;
 847
 848                 if (PageReserved(page))         /* Check for zero page */
 849                         goto put_and_set;
 850
 851                 pp->page = page;
 852                 err = page_to_nid(page);
 853
 854                 if (err == pp->node)
 855                         /*
 856                          * Node already in the right place
 857                          */
 858                         goto put_and_set;
 859
 860                 err = -EACCES;
 861                 if (page_mapcount(page) > 1 &&
 862                                 !migrate_all)
 863                         goto put_and_set;
 864
 865                 err = isolate_lru_page(page);
 866                 if (!err)
 867                         list_add_tail(&page->lru, &pagelist);
 868 put_and_set:
 869                 /*
 870                  * Either remove the duplicate refcount from
 871                  * isolate_lru_page() or drop the page ref if it was
 872                  * not isolated.
 873                  */
 874                 put_page(page);
 875 set_status:
 876                 pp->status = err;
 877         }
 878
 879         err = 0;
 880         if (!list_empty(&pagelist))
 881                 err = migrate_pages(&pagelist, new_page_node,
 882                                 (unsigned long)pm);
 883
 884         up_read(&mm->mmap_sem);
 885         return err;
 886 }
 887
 888 /*
 889  * Migrate an array of page address onto an array of nodes and fill
 890  * the corresponding array of status.
 891  */
 892 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
 893                          unsigned long nr_pages,
 894                          const void __user * __user *pages,
 895                          const int __user *nodes,
 896                          int __user *status, int flags)
 897 {
 898         struct page_to_node *pm;
 899         nodemask_t task_nodes;
 900         unsigned long chunk_nr_pages;
 901         unsigned long chunk_start;
 902         int err;
 903
 904         task_nodes = cpuset_mems_allowed(task);
 905
 906         err = -ENOMEM;
 907         pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
 908         if (!pm)
 909                 goto out;
 910         /*
 911          * Store a chunk of page_to_node array in a page,
 912          * but keep the last one as a marker
 913          */
 914         chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
 915
 916         for (chunk_start = 0;
 917              chunk_start < nr_pages;
 918              chunk_start += chunk_nr_pages) {
 919                 int j;
 920
 921                 if (chunk_start + chunk_nr_pages > nr_pages)
 922                         chunk_nr_pages = nr_pages - chunk_start;
 923
 924                 /* fill the chunk pm with addrs and nodes from user-space */
 925                 for (j = 0; j < chunk_nr_pages; j++) {
 926                         const void __user *p;
 927                         int node;
 928
 929                         err = -EFAULT;
 930                         if (get_user(p, pages + j + chunk_start))
 931                                 goto out_pm;
 932                         pm[j].addr = (unsigned long) p;
 933
 934                         if (get_user(node, nodes + j + chunk_start))
 935                                 goto out_pm;
 936
 937                         err = -ENODEV;
 938                         if (!node_state(node, N_HIGH_MEMORY))
 939                                 goto out_pm;
 940
 941                         err = -EACCES;
 942                         if (!node_isset(node, task_nodes))
 943                                 goto out_pm;
 944
 945                         pm[j].node = node;
 946                 }
 947
 948                 /* End marker for this chunk */
 949                 pm[chunk_nr_pages].node = MAX_NUMNODES;
 950
 951                 /* Migrate this chunk */
 952                 err = do_move_page_to_node_array(mm, pm,
 953                                                  flags & MPOL_MF_MOVE_ALL);
 954                 if (err < 0)
 955                         goto out_pm;
 956
 957                 /* Return status information */
 958                 for (j = 0; j < chunk_nr_pages; j++)
 959                         if (put_user(pm[j].status, status + j + chunk_start)) {
 960                                 err = -EFAULT;
 961                                 goto out_pm;
 962                         }
 963         }
 964         err = 0;
 965
 966 out_pm:
 967         free_page((unsigned long)pm);
 968 out:
 969         return err;
 970 }
 971
 972 /*
 973  * Determine the nodes of an array of pages and store it in an array of status.
 974  */
 975 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
 976                                 const void __user **pages, int *status)
 977 {
 978         unsigned long i;
 979
 980         down_read(&mm->mmap_sem);
 981
 982         for (i = 0; i < nr_pages; i++) {
 983                 unsigned long addr = (unsigned long)(*pages);
 984                 struct vm_area_struct *vma;
 985                 struct page *page;
 986                 int err = -EFAULT;
 987
 988                 vma = find_vma(mm, addr);
 989                 if (!vma)
 990                         goto set_status;
 991
 992                 page = follow_page(vma, addr, 0);
 993
 994                 err = PTR_ERR(page);
 995                 if (IS_ERR(page))
 996                         goto set_status;
 997
 998                 err = -ENOENT;
 999                 /* Use PageReserved to check for zero page */
1000                 if (!page || PageReserved(page))
1001                         goto set_status;
1002
1003                 err = page_to_nid(page);
1004 set_status:
1005                 *status = err;
1006
1007                 pages++;
1008                 status++;
1009         }
1010
1011         up_read(&mm->mmap_sem);
1012 }
1013
1014 /*
1015  * Determine the nodes of a user array of pages and store it in
1016  * a user array of status.
1017  */
1018 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1019                          const void __user * __user *pages,
1020                          int __user *status)
1021 {
1022 #define DO_PAGES_STAT_CHUNK_NR 16
1023         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1024         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1025         unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1026         int err;
1027
1028         for (i = 0; i < nr_pages; i += chunk_nr) {
1029                 if (chunk_nr + i > nr_pages)
1030                         chunk_nr = nr_pages - i;
1031
1032                 err = copy_from_user(chunk_pages, &pages[i],
1033                                      chunk_nr * sizeof(*chunk_pages));
1034                 if (err) {
1035                         err = -EFAULT;
1036                         goto out;
1037                 }
1038
1039                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1040
1041                 err = copy_to_user(&status[i], chunk_status,
1042                                    chunk_nr * sizeof(*chunk_status));
1043                 if (err) {
1044                         err = -EFAULT;
1045                         goto out;
1046                 }
1047         }
1048         err = 0;
1049
1050 out:
1051         return err;
1052 }
1053
1054 /*
1055  * Move a list of pages in the address space of the currently executing
1056  * process.
1057  */
1058 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1059                 const void __user * __user *, pages,
1060                 const int __user *, nodes,
1061                 int __user *, status, int, flags)
1062 {
1063         const struct cred *cred = current_cred(), *tcred;
1064         struct task_struct *task;
1065         struct mm_struct *mm;
1066         int err;
1067
1068         /* Check flags */
1069         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1070                 return -EINVAL;
1071
1072         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1073                 return -EPERM;
1074
1075         /* Find the mm_struct */
1076         read_lock(&tasklist_lock);
1077         task = pid ? find_task_by_vpid(pid) : current;
1078         if (!task) {
1079                 read_unlock(&tasklist_lock);
1080                 return -ESRCH;
1081         }
1082         mm = get_task_mm(task);
1083         read_unlock(&tasklist_lock);
1084
1085         if (!mm)
1086                 return -EINVAL;
1087
1088         /*
1089          * Check if this process has the right to modify the specified
1090          * process. The right exists if the process has administrative
1091          * capabilities, superuser privileges or the same
1092          * userid as the target process.
1093          */
1094         rcu_read_lock();
1095         tcred = __task_cred(task);
1096         if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1097             cred->uid  != tcred->suid && cred->uid  != tcred->uid &&
1098             !capable(CAP_SYS_NICE)) {
1099                 rcu_read_unlock();
1100                 err = -EPERM;
1101                 goto out;
1102         }
1103         rcu_read_unlock();
1104
1105         err = security_task_movememory(task);
1106         if (err)
1107                 goto out;
1108
1109         if (nodes) {
1110                 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1111                                     flags);
1112         } else {
1113                 err = do_pages_stat(mm, nr_pages, pages, status);
1114         }
1115
1116 out:
1117         mmput(mm);
1118         return err;
1119 }
1120
1121 /*
1122  * Call migration functions in the vma_ops that may prepare
1123  * memory in a vm for migration. migration functions may perform
1124  * the migration for vmas that do not have an underlying page struct.
1125  */
1126 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1127         const nodemask_t *from, unsigned long flags)
1128 {
1129         struct vm_area_struct *vma;
1130         int err = 0;
1131
1132         for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1133                 if (vma->vm_ops && vma->vm_ops->migrate) {
1134                         err = vma->vm_ops->migrate(vma, to, from, flags);
1135                         if (err)
1136                                 break;
1137                 }
1138         }
1139         return err;
1140 }
1141 #endif