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 <clameter@sgi.com>
  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/pagevec.h>
  23 #include <linux/rmap.h>
  24 #include <linux/topology.h>
  25 #include <linux/cpu.h>
  26 #include <linux/cpuset.h>
  27 #include <linux/writeback.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/security.h>
  31
  32 #include "internal.h"
  33
  34 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  35
  36 /*
  37  * Isolate one page from the LRU lists. If successful put it onto
  38  * the indicated list with elevated page count.
  39  *
  40  * Result:
  41  *  -EBUSY: page not on LRU list
  42  *  0: page removed from LRU list and added to the specified list.
  43  */
  44 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  45 {
  46         int ret = -EBUSY;
  47
  48         if (PageLRU(page)) {
  49                 struct zone *zone = page_zone(page);
  50
  51                 spin_lock_irq(&zone->lru_lock);
  52                 if (PageLRU(page)) {
  53                         ret = 0;
  54                         get_page(page);
  55                         ClearPageLRU(page);
  56                         if (PageActive(page))
  57                                 del_page_from_active_list(zone, page);
  58                         else
  59                                 del_page_from_inactive_list(zone, page);
  60                         list_add_tail(&page->lru, pagelist);
  61                 }
  62                 spin_unlock_irq(&zone->lru_lock);
  63         }
  64         return ret;
  65 }
  66
  67 /*
  68  * migrate_prep() needs to be called before we start compiling a list of pages
  69  * to be migrated using isolate_lru_page().
  70  */
  71 int migrate_prep(void)
  72 {
  73         /*
  74          * Clear the LRU lists so pages can be isolated.
  75          * Note that pages may be moved off the LRU after we have
  76          * drained them. Those pages will fail to migrate like other
  77          * pages that may be busy.
  78          */
  79         lru_add_drain_all();
  80
  81         return 0;
  82 }
  83
  84 static inline void move_to_lru(struct page *page)
  85 {
  86         if (PageActive(page)) {
  87                 /*
  88                  * lru_cache_add_active checks that
  89                  * the PG_active bit is off.
  90                  */
  91                 ClearPageActive(page);
  92                 lru_cache_add_active(page);
  93         } else {
  94                 lru_cache_add(page);
  95         }
  96         put_page(page);
  97 }
  98
  99 /*
 100  * Add isolated pages on the list back to the LRU.
 101  *
 102  * returns the number of pages put back.
 103  */
 104 int putback_lru_pages(struct list_head *l)
 105 {
 106         struct page *page;
 107         struct page *page2;
 108         int count = 0;
 109
 110         list_for_each_entry_safe(page, page2, l, lru) {
 111                 list_del(&page->lru);
 112                 move_to_lru(page);
 113                 count++;
 114         }
 115         return count;
 116 }
 117
 118 static inline int is_swap_pte(pte_t pte)
 119 {
 120         return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
 121 }
 122
 123 /*
 124  * Restore a potential migration pte to a working pte entry
 125  */
 126 static void remove_migration_pte(struct vm_area_struct *vma,
 127                 struct page *old, struct page *new)
 128 {
 129         struct mm_struct *mm = vma->vm_mm;
 130         swp_entry_t entry;
 131         pgd_t *pgd;
 132         pud_t *pud;
 133         pmd_t *pmd;
 134         pte_t *ptep, pte;
 135         spinlock_t *ptl;
 136         unsigned long addr = page_address_in_vma(new, vma);
 137
 138         if (addr == -EFAULT)
 139                 return;
 140
 141         pgd = pgd_offset(mm, addr);
 142         if (!pgd_present(*pgd))
 143                 return;
 144
 145         pud = pud_offset(pgd, addr);
 146         if (!pud_present(*pud))
 147                 return;
 148
 149         pmd = pmd_offset(pud, addr);
 150         if (!pmd_present(*pmd))
 151                 return;
 152
 153         ptep = pte_offset_map(pmd, addr);
 154
 155         if (!is_swap_pte(*ptep)) {
 156                 pte_unmap(ptep);
 157                 return;
 158         }
 159
 160         ptl = pte_lockptr(mm, pmd);
 161         spin_lock(ptl);
 162         pte = *ptep;
 163         if (!is_swap_pte(pte))
 164                 goto out;
 165
 166         entry = pte_to_swp_entry(pte);
 167
 168         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 169                 goto out;
 170
 171         get_page(new);
 172         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 173         if (is_write_migration_entry(entry))
 174                 pte = pte_mkwrite(pte);
 175         set_pte_at(mm, addr, ptep, pte);
 176
 177         if (PageAnon(new))
 178                 page_add_anon_rmap(new, vma, addr);
 179         else
 180                 page_add_file_rmap(new);
 181
 182         /* No need to invalidate - it was non-present before */
 183         update_mmu_cache(vma, addr, pte);
 184         lazy_mmu_prot_update(pte);
 185
 186 out:
 187         pte_unmap_unlock(ptep, ptl);
 188 }
 189
 190 /*
 191  * Note that remove_file_migration_ptes will only work on regular mappings,
 192  * Nonlinear mappings do not use migration entries.
 193  */
 194 static void remove_file_migration_ptes(struct page *old, struct page *new)
 195 {
 196         struct vm_area_struct *vma;
 197         struct address_space *mapping = page_mapping(new);
 198         struct prio_tree_iter iter;
 199         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 200
 201         if (!mapping)
 202                 return;
 203
 204         spin_lock(&mapping->i_mmap_lock);
 205
 206         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 207                 remove_migration_pte(vma, old, new);
 208
 209         spin_unlock(&mapping->i_mmap_lock);
 210 }
 211
 212 /*
 213  * Must hold mmap_sem lock on at least one of the vmas containing
 214  * the page so that the anon_vma cannot vanish.
 215  */
 216 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 217 {
 218         struct anon_vma *anon_vma;
 219         struct vm_area_struct *vma;
 220         unsigned long mapping;
 221
 222         mapping = (unsigned long)new->mapping;
 223
 224         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 225                 return;
 226
 227         /*
 228          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 229          */
 230         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 231         spin_lock(&anon_vma->lock);
 232
 233         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 234                 remove_migration_pte(vma, old, new);
 235
 236         spin_unlock(&anon_vma->lock);
 237 }
 238
 239 /*
 240  * Get rid of all migration entries and replace them by
 241  * references to the indicated page.
 242  */
 243 static void remove_migration_ptes(struct page *old, struct page *new)
 244 {
 245         if (PageAnon(new))
 246                 remove_anon_migration_ptes(old, new);
 247         else
 248                 remove_file_migration_ptes(old, new);
 249 }
 250
 251 /*
 252  * Something used the pte of a page under migration. We need to
 253  * get to the page and wait until migration is finished.
 254  * When we return from this function the fault will be retried.
 255  *
 256  * This function is called from do_swap_page().
 257  */
 258 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 259                                 unsigned long address)
 260 {
 261         pte_t *ptep, pte;
 262         spinlock_t *ptl;
 263         swp_entry_t entry;
 264         struct page *page;
 265
 266         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 267         pte = *ptep;
 268         if (!is_swap_pte(pte))
 269                 goto out;
 270
 271         entry = pte_to_swp_entry(pte);
 272         if (!is_migration_entry(entry))
 273                 goto out;
 274
 275         page = migration_entry_to_page(entry);
 276
 277         get_page(page);
 278         pte_unmap_unlock(ptep, ptl);
 279         wait_on_page_locked(page);
 280         put_page(page);
 281         return;
 282 out:
 283         pte_unmap_unlock(ptep, ptl);
 284 }
 285
 286 /*
 287  * Replace the page in the mapping.
 288  *
 289  * The number of remaining references must be:
 290  * 1 for anonymous pages without a mapping
 291  * 2 for pages with a mapping
 292  * 3 for pages with a mapping and PagePrivate set.
 293  */
 294 static int migrate_page_move_mapping(struct address_space *mapping,
 295                 struct page *newpage, struct page *page)
 296 {
 297         struct page **radix_pointer;
 298
 299         if (!mapping) {
 300                 /* Anonymous page */
 301                 if (page_count(page) != 1)
 302                         return -EAGAIN;
 303                 return 0;
 304         }
 305
 306         write_lock_irq(&mapping->tree_lock);
 307
 308         radix_pointer = (struct page **)radix_tree_lookup_slot(
 309                                                 &mapping->page_tree,
 310                                                 page_index(page));
 311
 312         if (page_count(page) != 2 + !!PagePrivate(page) ||
 313                         *radix_pointer != page) {
 314                 write_unlock_irq(&mapping->tree_lock);
 315                 return -EAGAIN;
 316         }
 317
 318         /*
 319          * Now we know that no one else is looking at the page.
 320          */
 321         get_page(newpage);
 322 #ifdef CONFIG_SWAP
 323         if (PageSwapCache(page)) {
 324                 SetPageSwapCache(newpage);
 325                 set_page_private(newpage, page_private(page));
 326         }
 327 #endif
 328
 329         *radix_pointer = newpage;
 330         __put_page(page);
 331         write_unlock_irq(&mapping->tree_lock);
 332
 333         return 0;
 334 }
 335
 336 /*
 337  * Copy the page to its new location
 338  */
 339 static void migrate_page_copy(struct page *newpage, struct page *page)
 340 {
 341         copy_highpage(newpage, page);
 342
 343         if (PageError(page))
 344                 SetPageError(newpage);
 345         if (PageReferenced(page))
 346                 SetPageReferenced(newpage);
 347         if (PageUptodate(page))
 348                 SetPageUptodate(newpage);
 349         if (PageActive(page))
 350                 SetPageActive(newpage);
 351         if (PageChecked(page))
 352                 SetPageChecked(newpage);
 353         if (PageMappedToDisk(page))
 354                 SetPageMappedToDisk(newpage);
 355
 356         if (PageDirty(page)) {
 357                 clear_page_dirty_for_io(page);
 358                 set_page_dirty(newpage);
 359         }
 360
 361 #ifdef CONFIG_SWAP
 362         ClearPageSwapCache(page);
 363 #endif
 364         ClearPageActive(page);
 365         ClearPagePrivate(page);
 366         set_page_private(page, 0);
 367         page->mapping = NULL;
 368
 369         /*
 370          * If any waiters have accumulated on the new page then
 371          * wake them up.
 372          */
 373         if (PageWriteback(newpage))
 374                 end_page_writeback(newpage);
 375 }
 376
 377 /************************************************************
 378  *                    Migration functions
 379  ***********************************************************/
 380
 381 /* Always fail migration. Used for mappings that are not movable */
 382 int fail_migrate_page(struct address_space *mapping,
 383                         struct page *newpage, struct page *page)
 384 {
 385         return -EIO;
 386 }
 387 EXPORT_SYMBOL(fail_migrate_page);
 388
 389 /*
 390  * Common logic to directly migrate a single page suitable for
 391  * pages that do not use PagePrivate.
 392  *
 393  * Pages are locked upon entry and exit.
 394  */
 395 int migrate_page(struct address_space *mapping,
 396                 struct page *newpage, struct page *page)
 397 {
 398         int rc;
 399
 400         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 401
 402         rc = migrate_page_move_mapping(mapping, newpage, page);
 403
 404         if (rc)
 405                 return rc;
 406
 407         migrate_page_copy(newpage, page);
 408         return 0;
 409 }
 410 EXPORT_SYMBOL(migrate_page);
 411
 412 /*
 413  * Migration function for pages with buffers. This function can only be used
 414  * if the underlying filesystem guarantees that no other references to "page"
 415  * exist.
 416  */
 417 int buffer_migrate_page(struct address_space *mapping,
 418                 struct page *newpage, struct page *page)
 419 {
 420         struct buffer_head *bh, *head;
 421         int rc;
 422
 423         if (!page_has_buffers(page))
 424                 return migrate_page(mapping, newpage, page);
 425
 426         head = page_buffers(page);
 427
 428         rc = migrate_page_move_mapping(mapping, newpage, page);
 429
 430         if (rc)
 431                 return rc;
 432
 433         bh = head;
 434         do {
 435                 get_bh(bh);
 436                 lock_buffer(bh);
 437                 bh = bh->b_this_page;
 438
 439         } while (bh != head);
 440
 441         ClearPagePrivate(page);
 442         set_page_private(newpage, page_private(page));
 443         set_page_private(page, 0);
 444         put_page(page);
 445         get_page(newpage);
 446
 447         bh = head;
 448         do {
 449                 set_bh_page(bh, newpage, bh_offset(bh));
 450                 bh = bh->b_this_page;
 451
 452         } while (bh != head);
 453
 454         SetPagePrivate(newpage);
 455
 456         migrate_page_copy(newpage, page);
 457
 458         bh = head;
 459         do {
 460                 unlock_buffer(bh);
 461                 put_bh(bh);
 462                 bh = bh->b_this_page;
 463
 464         } while (bh != head);
 465
 466         return 0;
 467 }
 468 EXPORT_SYMBOL(buffer_migrate_page);
 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         if (rc < 0)
 505                 /* I/O Error writing */
 506                 return -EIO;
 507
 508         if (rc != AOP_WRITEPAGE_ACTIVATE)
 509                 /* unlocked. Relock */
 510                 lock_page(page);
 511
 512         return -EAGAIN;
 513 }
 514
 515 /*
 516  * Default handling if a filesystem does not provide a migration function.
 517  */
 518 static int fallback_migrate_page(struct address_space *mapping,
 519         struct page *newpage, struct page *page)
 520 {
 521         if (PageDirty(page))
 522                 return writeout(mapping, page);
 523
 524         /*
 525          * Buffers may be managed in a filesystem specific way.
 526          * We must have no buffers or drop them.
 527          */
 528         if (page_has_buffers(page) &&
 529             !try_to_release_page(page, GFP_KERNEL))
 530                 return -EAGAIN;
 531
 532         return migrate_page(mapping, newpage, page);
 533 }
 534
 535 /*
 536  * Move a page to a newly allocated page
 537  * The page is locked and all ptes have been successfully removed.
 538  *
 539  * The new page will have replaced the old page if this function
 540  * is successful.
 541  */
 542 static int move_to_new_page(struct page *newpage, struct page *page)
 543 {
 544         struct address_space *mapping;
 545         int rc;
 546
 547         /*
 548          * Block others from accessing the page when we get around to
 549          * establishing additional references. We are the only one
 550          * holding a reference to the new page at this point.
 551          */
 552         if (TestSetPageLocked(newpage))
 553                 BUG();
 554
 555         /* Prepare mapping for the new page.*/
 556         newpage->index = page->index;
 557         newpage->mapping = page->mapping;
 558
 559         mapping = page_mapping(page);
 560         if (!mapping)
 561                 rc = migrate_page(mapping, newpage, page);
 562         else if (mapping->a_ops->migratepage)
 563                 /*
 564                  * Most pages have a mapping and most filesystems
 565                  * should provide a migration function. Anonymous
 566                  * pages are part of swap space which also has its
 567                  * own migration function. This is the most common
 568                  * path for page migration.
 569                  */
 570                 rc = mapping->a_ops->migratepage(mapping,
 571                                                 newpage, page);
 572         else
 573                 rc = fallback_migrate_page(mapping, newpage, page);
 574
 575         if (!rc)
 576                 remove_migration_ptes(page, newpage);
 577         else
 578                 newpage->mapping = NULL;
 579
 580         unlock_page(newpage);
 581
 582         return rc;
 583 }
 584
 585 /*
 586  * Obtain the lock on page, remove all ptes and migrate the page
 587  * to the newly allocated page in newpage.
 588  */
 589 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 590                         struct page *page, int force)
 591 {
 592         int rc = 0;
 593         int *result = NULL;
 594         struct page *newpage = get_new_page(page, private, &result);
 595
 596         if (!newpage)
 597                 return -ENOMEM;
 598
 599         if (page_count(page) == 1)
 600                 /* page was freed from under us. So we are done. */
 601                 goto move_newpage;
 602
 603         rc = -EAGAIN;
 604         if (TestSetPageLocked(page)) {
 605                 if (!force)
 606                         goto move_newpage;
 607                 lock_page(page);
 608         }
 609
 610         if (PageWriteback(page)) {
 611                 if (!force)
 612                         goto unlock;
 613                 wait_on_page_writeback(page);
 614         }
 615
 616         /*
 617          * Establish migration ptes or remove ptes
 618          */
 619         try_to_unmap(page, 1);
 620         if (!page_mapped(page))
 621                 rc = move_to_new_page(newpage, page);
 622
 623         if (rc)
 624                 remove_migration_ptes(page, page);
 625
 626 unlock:
 627         unlock_page(page);
 628
 629         if (rc != -EAGAIN) {
 630                 /*
 631                  * A page that has been migrated has all references
 632                  * removed and will be freed. A page that has not been
 633                  * migrated will have kepts its references and be
 634                  * restored.
 635                  */
 636                 list_del(&page->lru);
 637                 move_to_lru(page);
 638         }
 639
 640 move_newpage:
 641         /*
 642          * Move the new page to the LRU. If migration was not successful
 643          * then this will free the page.
 644          */
 645         move_to_lru(newpage);
 646         if (result) {
 647                 if (rc)
 648                         *result = rc;
 649                 else
 650                         *result = page_to_nid(newpage);
 651         }
 652         return rc;
 653 }
 654
 655 /*
 656  * migrate_pages
 657  *
 658  * The function takes one list of pages to migrate and a function
 659  * that determines from the page to be migrated and the private data
 660  * the target of the move and allocates the page.
 661  *
 662  * The function returns after 10 attempts or if no pages
 663  * are movable anymore because to has become empty
 664  * or no retryable pages exist anymore. All pages will be
 665  * retruned to the LRU or freed.
 666  *
 667  * Return: Number of pages not migrated or error code.
 668  */
 669 int migrate_pages(struct list_head *from,
 670                 new_page_t get_new_page, unsigned long private)
 671 {
 672         int retry = 1;
 673         int nr_failed = 0;
 674         int pass = 0;
 675         struct page *page;
 676         struct page *page2;
 677         int swapwrite = current->flags & PF_SWAPWRITE;
 678         int rc;
 679
 680         if (!swapwrite)
 681                 current->flags |= PF_SWAPWRITE;
 682
 683         for(pass = 0; pass < 10 && retry; pass++) {
 684                 retry = 0;
 685
 686                 list_for_each_entry_safe(page, page2, from, lru) {
 687                         cond_resched();
 688
 689                         rc = unmap_and_move(get_new_page, private,
 690                                                 page, pass > 2);
 691
 692                         switch(rc) {
 693                         case -ENOMEM:
 694                                 goto out;
 695                         case -EAGAIN:
 696                                 retry++;
 697                                 break;
 698                         case 0:
 699                                 break;
 700                         default:
 701                                 /* Permanent failure */
 702                                 nr_failed++;
 703                                 break;
 704                         }
 705                 }
 706         }
 707         rc = 0;
 708 out:
 709         if (!swapwrite)
 710                 current->flags &= ~PF_SWAPWRITE;
 711
 712         putback_lru_pages(from);
 713
 714         if (rc)
 715                 return rc;
 716
 717         return nr_failed + retry;
 718 }
 719
 720 #ifdef CONFIG_NUMA
 721 /*
 722  * Move a list of individual pages
 723  */
 724 struct page_to_node {
 725         unsigned long addr;
 726         struct page *page;
 727         int node;
 728         int status;
 729 };
 730
 731 static struct page *new_page_node(struct page *p, unsigned long private,
 732                 int **result)
 733 {
 734         struct page_to_node *pm = (struct page_to_node *)private;
 735
 736         while (pm->node != MAX_NUMNODES && pm->page != p)
 737                 pm++;
 738
 739         if (pm->node == MAX_NUMNODES)
 740                 return NULL;
 741
 742         *result = &pm->status;
 743
 744         return alloc_pages_node(pm->node, GFP_HIGHUSER | GFP_THISNODE, 0);
 745 }
 746
 747 /*
 748  * Move a set of pages as indicated in the pm array. The addr
 749  * field must be set to the virtual address of the page to be moved
 750  * and the node number must contain a valid target node.
 751  */
 752 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 753                                 int migrate_all)
 754 {
 755         int err;
 756         struct page_to_node *pp;
 757         LIST_HEAD(pagelist);
 758
 759         down_read(&mm->mmap_sem);
 760
 761         /*
 762          * Build a list of pages to migrate
 763          */
 764         migrate_prep();
 765         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 766                 struct vm_area_struct *vma;
 767                 struct page *page;
 768
 769                 /*
 770                  * A valid page pointer that will not match any of the
 771                  * pages that will be moved.
 772                  */
 773                 pp->page = ZERO_PAGE(0);
 774
 775                 err = -EFAULT;
 776                 vma = find_vma(mm, pp->addr);
 777                 if (!vma)
 778                         goto set_status;
 779
 780                 page = follow_page(vma, pp->addr, FOLL_GET);
 781                 err = -ENOENT;
 782                 if (!page)
 783                         goto set_status;
 784
 785                 if (PageReserved(page))         /* Check for zero page */
 786                         goto put_and_set;
 787
 788                 pp->page = page;
 789                 err = page_to_nid(page);
 790
 791                 if (err == pp->node)
 792                         /*
 793                          * Node already in the right place
 794                          */
 795                         goto put_and_set;
 796
 797                 err = -EACCES;
 798                 if (page_mapcount(page) > 1 &&
 799                                 !migrate_all)
 800                         goto put_and_set;
 801
 802                 err = isolate_lru_page(page, &pagelist);
 803 put_and_set:
 804                 /*
 805                  * Either remove the duplicate refcount from
 806                  * isolate_lru_page() or drop the page ref if it was
 807                  * not isolated.
 808                  */
 809                 put_page(page);
 810 set_status:
 811                 pp->status = err;
 812         }
 813
 814         if (!list_empty(&pagelist))
 815                 err = migrate_pages(&pagelist, new_page_node,
 816                                 (unsigned long)pm);
 817         else
 818                 err = -ENOENT;
 819
 820         up_read(&mm->mmap_sem);
 821         return err;
 822 }
 823
 824 /*
 825  * Determine the nodes of a list of pages. The addr in the pm array
 826  * must have been set to the virtual address of which we want to determine
 827  * the node number.
 828  */
 829 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 830 {
 831         down_read(&mm->mmap_sem);
 832
 833         for ( ; pm->node != MAX_NUMNODES; pm++) {
 834                 struct vm_area_struct *vma;
 835                 struct page *page;
 836                 int err;
 837
 838                 err = -EFAULT;
 839                 vma = find_vma(mm, pm->addr);
 840                 if (!vma)
 841                         goto set_status;
 842
 843                 page = follow_page(vma, pm->addr, 0);
 844                 err = -ENOENT;
 845                 /* Use PageReserved to check for zero page */
 846                 if (!page || PageReserved(page))
 847                         goto set_status;
 848
 849                 err = page_to_nid(page);
 850 set_status:
 851                 pm->status = err;
 852         }
 853
 854         up_read(&mm->mmap_sem);
 855         return 0;
 856 }
 857
 858 /*
 859  * Move a list of pages in the address space of the currently executing
 860  * process.
 861  */
 862 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 863                         const void __user * __user *pages,
 864                         const int __user *nodes,
 865                         int __user *status, int flags)
 866 {
 867         int err = 0;
 868         int i;
 869         struct task_struct *task;
 870         nodemask_t task_nodes;
 871         struct mm_struct *mm;
 872         struct page_to_node *pm = NULL;
 873
 874         /* Check flags */
 875         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 876                 return -EINVAL;
 877
 878         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 879                 return -EPERM;
 880
 881         /* Find the mm_struct */
 882         read_lock(&tasklist_lock);
 883         task = pid ? find_task_by_pid(pid) : current;
 884         if (!task) {
 885                 read_unlock(&tasklist_lock);
 886                 return -ESRCH;
 887         }
 888         mm = get_task_mm(task);
 889         read_unlock(&tasklist_lock);
 890
 891         if (!mm)
 892                 return -EINVAL;
 893
 894         /*
 895          * Check if this process has the right to modify the specified
 896          * process. The right exists if the process has administrative
 897          * capabilities, superuser privileges or the same
 898          * userid as the target process.
 899          */
 900         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 901             (current->uid != task->suid) && (current->uid != task->uid) &&
 902             !capable(CAP_SYS_NICE)) {
 903                 err = -EPERM;
 904                 goto out2;
 905         }
 906
 907         err = security_task_movememory(task);
 908         if (err)
 909                 goto out2;
 910
 911
 912         task_nodes = cpuset_mems_allowed(task);
 913
 914         /* Limit nr_pages so that the multiplication may not overflow */
 915         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 916                 err = -E2BIG;
 917                 goto out2;
 918         }
 919
 920         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
 921         if (!pm) {
 922                 err = -ENOMEM;
 923                 goto out2;
 924         }
 925
 926         /*
 927          * Get parameters from user space and initialize the pm
 928          * array. Return various errors if the user did something wrong.
 929          */
 930         for (i = 0; i < nr_pages; i++) {
 931                 const void *p;
 932
 933                 err = -EFAULT;
 934                 if (get_user(p, pages + i))
 935                         goto out;
 936
 937                 pm[i].addr = (unsigned long)p;
 938                 if (nodes) {
 939                         int node;
 940
 941                         if (get_user(node, nodes + i))
 942                                 goto out;
 943
 944                         err = -ENODEV;
 945                         if (!node_online(node))
 946                                 goto out;
 947
 948                         err = -EACCES;
 949                         if (!node_isset(node, task_nodes))
 950                                 goto out;
 951
 952                         pm[i].node = node;
 953                 }
 954         }
 955         /* End marker */
 956         pm[nr_pages].node = MAX_NUMNODES;
 957
 958         if (nodes)
 959                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
 960         else
 961                 err = do_pages_stat(mm, pm);
 962
 963         if (err >= 0)
 964                 /* Return status information */
 965                 for (i = 0; i < nr_pages; i++)
 966                         if (put_user(pm[i].status, status + i))
 967                                 err = -EFAULT;
 968
 969 out:
 970         vfree(pm);
 971 out2:
 972         mmput(mm);
 973         return err;
 974 }
 975 #endif
 976
 977 /*
 978  * Call migration functions in the vma_ops that may prepare
 979  * memory in a vm for migration. migration functions may perform
 980  * the migration for vmas that do not have an underlying page struct.
 981  */
 982 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
 983         const nodemask_t *from, unsigned long flags)
 984 {
 985         struct vm_area_struct *vma;
 986         int err = 0;
 987
 988         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
 989                 if (vma->vm_ops && vma->vm_ops->migrate) {
 990                         err = vma->vm_ops->migrate(vma, to, from, flags);
 991                         if (err)
 992                                 break;
 993                 }
 994         }
 995         return err;
 996 }