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 #ifdef CONFIG_BLOCK
 413 /*
 414  * Migration function for pages with buffers. This function can only be used
 415  * if the underlying filesystem guarantees that no other references to "page"
 416  * exist.
 417  */
 418 int buffer_migrate_page(struct address_space *mapping,
 419                 struct page *newpage, struct page *page)
 420 {
 421         struct buffer_head *bh, *head;
 422         int rc;
 423
 424         if (!page_has_buffers(page))
 425                 return migrate_page(mapping, newpage, page);
 426
 427         head = page_buffers(page);
 428
 429         rc = migrate_page_move_mapping(mapping, newpage, page);
 430
 431         if (rc)
 432                 return rc;
 433
 434         bh = head;
 435         do {
 436                 get_bh(bh);
 437                 lock_buffer(bh);
 438                 bh = bh->b_this_page;
 439
 440         } while (bh != head);
 441
 442         ClearPagePrivate(page);
 443         set_page_private(newpage, page_private(page));
 444         set_page_private(page, 0);
 445         put_page(page);
 446         get_page(newpage);
 447
 448         bh = head;
 449         do {
 450                 set_bh_page(bh, newpage, bh_offset(bh));
 451                 bh = bh->b_this_page;
 452
 453         } while (bh != head);
 454
 455         SetPagePrivate(newpage);
 456
 457         migrate_page_copy(newpage, page);
 458
 459         bh = head;
 460         do {
 461                 unlock_buffer(bh);
 462                 put_bh(bh);
 463                 bh = bh->b_this_page;
 464
 465         } while (bh != head);
 466
 467         return 0;
 468 }
 469 EXPORT_SYMBOL(buffer_migrate_page);
 470 #endif
 471
 472 /*
 473  * Writeback a page to clean the dirty state
 474  */
 475 static int writeout(struct address_space *mapping, struct page *page)
 476 {
 477         struct writeback_control wbc = {
 478                 .sync_mode = WB_SYNC_NONE,
 479                 .nr_to_write = 1,
 480                 .range_start = 0,
 481                 .range_end = LLONG_MAX,
 482                 .nonblocking = 1,
 483                 .for_reclaim = 1
 484         };
 485         int rc;
 486
 487         if (!mapping->a_ops->writepage)
 488                 /* No write method for the address space */
 489                 return -EINVAL;
 490
 491         if (!clear_page_dirty_for_io(page))
 492                 /* Someone else already triggered a write */
 493                 return -EAGAIN;
 494
 495         /*
 496          * A dirty page may imply that the underlying filesystem has
 497          * the page on some queue. So the page must be clean for
 498          * migration. Writeout may mean we loose the lock and the
 499          * page state is no longer what we checked for earlier.
 500          * At this point we know that the migration attempt cannot
 501          * be successful.
 502          */
 503         remove_migration_ptes(page, page);
 504
 505         rc = mapping->a_ops->writepage(page, &wbc);
 506         if (rc < 0)
 507                 /* I/O Error writing */
 508                 return -EIO;
 509
 510         if (rc != AOP_WRITEPAGE_ACTIVATE)
 511                 /* unlocked. Relock */
 512                 lock_page(page);
 513
 514         return -EAGAIN;
 515 }
 516
 517 /*
 518  * Default handling if a filesystem does not provide a migration function.
 519  */
 520 static int fallback_migrate_page(struct address_space *mapping,
 521         struct page *newpage, struct page *page)
 522 {
 523         if (PageDirty(page))
 524                 return writeout(mapping, page);
 525
 526         /*
 527          * Buffers may be managed in a filesystem specific way.
 528          * We must have no buffers or drop them.
 529          */
 530         if (PagePrivate(page) &&
 531             !try_to_release_page(page, GFP_KERNEL))
 532                 return -EAGAIN;
 533
 534         return migrate_page(mapping, newpage, page);
 535 }
 536
 537 /*
 538  * Move a page to a newly allocated page
 539  * The page is locked and all ptes have been successfully removed.
 540  *
 541  * The new page will have replaced the old page if this function
 542  * is successful.
 543  */
 544 static int move_to_new_page(struct page *newpage, struct page *page)
 545 {
 546         struct address_space *mapping;
 547         int rc;
 548
 549         /*
 550          * Block others from accessing the page when we get around to
 551          * establishing additional references. We are the only one
 552          * holding a reference to the new page at this point.
 553          */
 554         if (TestSetPageLocked(newpage))
 555                 BUG();
 556
 557         /* Prepare mapping for the new page.*/
 558         newpage->index = page->index;
 559         newpage->mapping = page->mapping;
 560
 561         mapping = page_mapping(page);
 562         if (!mapping)
 563                 rc = migrate_page(mapping, newpage, page);
 564         else if (mapping->a_ops->migratepage)
 565                 /*
 566                  * Most pages have a mapping and most filesystems
 567                  * should provide a migration function. Anonymous
 568                  * pages are part of swap space which also has its
 569                  * own migration function. This is the most common
 570                  * path for page migration.
 571                  */
 572                 rc = mapping->a_ops->migratepage(mapping,
 573                                                 newpage, page);
 574         else
 575                 rc = fallback_migrate_page(mapping, newpage, page);
 576
 577         if (!rc)
 578                 remove_migration_ptes(page, newpage);
 579         else
 580                 newpage->mapping = NULL;
 581
 582         unlock_page(newpage);
 583
 584         return rc;
 585 }
 586
 587 /*
 588  * Obtain the lock on page, remove all ptes and migrate the page
 589  * to the newly allocated page in newpage.
 590  */
 591 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 592                         struct page *page, int force)
 593 {
 594         int rc = 0;
 595         int *result = NULL;
 596         struct page *newpage = get_new_page(page, private, &result);
 597
 598         if (!newpage)
 599                 return -ENOMEM;
 600
 601         if (page_count(page) == 1)
 602                 /* page was freed from under us. So we are done. */
 603                 goto move_newpage;
 604
 605         rc = -EAGAIN;
 606         if (TestSetPageLocked(page)) {
 607                 if (!force)
 608                         goto move_newpage;
 609                 lock_page(page);
 610         }
 611
 612         if (PageWriteback(page)) {
 613                 if (!force)
 614                         goto unlock;
 615                 wait_on_page_writeback(page);
 616         }
 617
 618         /*
 619          * Establish migration ptes or remove ptes
 620          */
 621         try_to_unmap(page, 1);
 622         if (!page_mapped(page))
 623                 rc = move_to_new_page(newpage, page);
 624
 625         if (rc)
 626                 remove_migration_ptes(page, page);
 627
 628 unlock:
 629         unlock_page(page);
 630
 631         if (rc != -EAGAIN) {
 632                 /*
 633                  * A page that has been migrated has all references
 634                  * removed and will be freed. A page that has not been
 635                  * migrated will have kepts its references and be
 636                  * restored.
 637                  */
 638                 list_del(&page->lru);
 639                 move_to_lru(page);
 640         }
 641
 642 move_newpage:
 643         /*
 644          * Move the new page to the LRU. If migration was not successful
 645          * then this will free the page.
 646          */
 647         move_to_lru(newpage);
 648         if (result) {
 649                 if (rc)
 650                         *result = rc;
 651                 else
 652                         *result = page_to_nid(newpage);
 653         }
 654         return rc;
 655 }
 656
 657 /*
 658  * migrate_pages
 659  *
 660  * The function takes one list of pages to migrate and a function
 661  * that determines from the page to be migrated and the private data
 662  * the target of the move and allocates the page.
 663  *
 664  * The function returns after 10 attempts or if no pages
 665  * are movable anymore because to has become empty
 666  * or no retryable pages exist anymore. All pages will be
 667  * retruned to the LRU or freed.
 668  *
 669  * Return: Number of pages not migrated or error code.
 670  */
 671 int migrate_pages(struct list_head *from,
 672                 new_page_t get_new_page, unsigned long private)
 673 {
 674         int retry = 1;
 675         int nr_failed = 0;
 676         int pass = 0;
 677         struct page *page;
 678         struct page *page2;
 679         int swapwrite = current->flags & PF_SWAPWRITE;
 680         int rc;
 681
 682         if (!swapwrite)
 683                 current->flags |= PF_SWAPWRITE;
 684
 685         for(pass = 0; pass < 10 && retry; pass++) {
 686                 retry = 0;
 687
 688                 list_for_each_entry_safe(page, page2, from, lru) {
 689                         cond_resched();
 690
 691                         rc = unmap_and_move(get_new_page, private,
 692                                                 page, pass > 2);
 693
 694                         switch(rc) {
 695                         case -ENOMEM:
 696                                 goto out;
 697                         case -EAGAIN:
 698                                 retry++;
 699                                 break;
 700                         case 0:
 701                                 break;
 702                         default:
 703                                 /* Permanent failure */
 704                                 nr_failed++;
 705                                 break;
 706                         }
 707                 }
 708         }
 709         rc = 0;
 710 out:
 711         if (!swapwrite)
 712                 current->flags &= ~PF_SWAPWRITE;
 713
 714         putback_lru_pages(from);
 715
 716         if (rc)
 717                 return rc;
 718
 719         return nr_failed + retry;
 720 }
 721
 722 #ifdef CONFIG_NUMA
 723 /*
 724  * Move a list of individual pages
 725  */
 726 struct page_to_node {
 727         unsigned long addr;
 728         struct page *page;
 729         int node;
 730         int status;
 731 };
 732
 733 static struct page *new_page_node(struct page *p, unsigned long private,
 734                 int **result)
 735 {
 736         struct page_to_node *pm = (struct page_to_node *)private;
 737
 738         while (pm->node != MAX_NUMNODES && pm->page != p)
 739                 pm++;
 740
 741         if (pm->node == MAX_NUMNODES)
 742                 return NULL;
 743
 744         *result = &pm->status;
 745
 746         return alloc_pages_node(pm->node, GFP_HIGHUSER | GFP_THISNODE, 0);
 747 }
 748
 749 /*
 750  * Move a set of pages as indicated in the pm array. The addr
 751  * field must be set to the virtual address of the page to be moved
 752  * and the node number must contain a valid target node.
 753  */
 754 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 755                                 int migrate_all)
 756 {
 757         int err;
 758         struct page_to_node *pp;
 759         LIST_HEAD(pagelist);
 760
 761         down_read(&mm->mmap_sem);
 762
 763         /*
 764          * Build a list of pages to migrate
 765          */
 766         migrate_prep();
 767         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 768                 struct vm_area_struct *vma;
 769                 struct page *page;
 770
 771                 /*
 772                  * A valid page pointer that will not match any of the
 773                  * pages that will be moved.
 774                  */
 775                 pp->page = ZERO_PAGE(0);
 776
 777                 err = -EFAULT;
 778                 vma = find_vma(mm, pp->addr);
 779                 if (!vma)
 780                         goto set_status;
 781
 782                 page = follow_page(vma, pp->addr, FOLL_GET);
 783                 err = -ENOENT;
 784                 if (!page)
 785                         goto set_status;
 786
 787                 if (PageReserved(page))         /* Check for zero page */
 788                         goto put_and_set;
 789
 790                 pp->page = page;
 791                 err = page_to_nid(page);
 792
 793                 if (err == pp->node)
 794                         /*
 795                          * Node already in the right place
 796                          */
 797                         goto put_and_set;
 798
 799                 err = -EACCES;
 800                 if (page_mapcount(page) > 1 &&
 801                                 !migrate_all)
 802                         goto put_and_set;
 803
 804                 err = isolate_lru_page(page, &pagelist);
 805 put_and_set:
 806                 /*
 807                  * Either remove the duplicate refcount from
 808                  * isolate_lru_page() or drop the page ref if it was
 809                  * not isolated.
 810                  */
 811                 put_page(page);
 812 set_status:
 813                 pp->status = err;
 814         }
 815
 816         if (!list_empty(&pagelist))
 817                 err = migrate_pages(&pagelist, new_page_node,
 818                                 (unsigned long)pm);
 819         else
 820                 err = -ENOENT;
 821
 822         up_read(&mm->mmap_sem);
 823         return err;
 824 }
 825
 826 /*
 827  * Determine the nodes of a list of pages. The addr in the pm array
 828  * must have been set to the virtual address of which we want to determine
 829  * the node number.
 830  */
 831 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 832 {
 833         down_read(&mm->mmap_sem);
 834
 835         for ( ; pm->node != MAX_NUMNODES; pm++) {
 836                 struct vm_area_struct *vma;
 837                 struct page *page;
 838                 int err;
 839
 840                 err = -EFAULT;
 841                 vma = find_vma(mm, pm->addr);
 842                 if (!vma)
 843                         goto set_status;
 844
 845                 page = follow_page(vma, pm->addr, 0);
 846                 err = -ENOENT;
 847                 /* Use PageReserved to check for zero page */
 848                 if (!page || PageReserved(page))
 849                         goto set_status;
 850
 851                 err = page_to_nid(page);
 852 set_status:
 853                 pm->status = err;
 854         }
 855
 856         up_read(&mm->mmap_sem);
 857         return 0;
 858 }
 859
 860 /*
 861  * Move a list of pages in the address space of the currently executing
 862  * process.
 863  */
 864 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 865                         const void __user * __user *pages,
 866                         const int __user *nodes,
 867                         int __user *status, int flags)
 868 {
 869         int err = 0;
 870         int i;
 871         struct task_struct *task;
 872         nodemask_t task_nodes;
 873         struct mm_struct *mm;
 874         struct page_to_node *pm = NULL;
 875
 876         /* Check flags */
 877         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 878                 return -EINVAL;
 879
 880         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 881                 return -EPERM;
 882
 883         /* Find the mm_struct */
 884         read_lock(&tasklist_lock);
 885         task = pid ? find_task_by_pid(pid) : current;
 886         if (!task) {
 887                 read_unlock(&tasklist_lock);
 888                 return -ESRCH;
 889         }
 890         mm = get_task_mm(task);
 891         read_unlock(&tasklist_lock);
 892
 893         if (!mm)
 894                 return -EINVAL;
 895
 896         /*
 897          * Check if this process has the right to modify the specified
 898          * process. The right exists if the process has administrative
 899          * capabilities, superuser privileges or the same
 900          * userid as the target process.
 901          */
 902         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 903             (current->uid != task->suid) && (current->uid != task->uid) &&
 904             !capable(CAP_SYS_NICE)) {
 905                 err = -EPERM;
 906                 goto out2;
 907         }
 908
 909         err = security_task_movememory(task);
 910         if (err)
 911                 goto out2;
 912
 913
 914         task_nodes = cpuset_mems_allowed(task);
 915
 916         /* Limit nr_pages so that the multiplication may not overflow */
 917         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 918                 err = -E2BIG;
 919                 goto out2;
 920         }
 921
 922         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
 923         if (!pm) {
 924                 err = -ENOMEM;
 925                 goto out2;
 926         }
 927
 928         /*
 929          * Get parameters from user space and initialize the pm
 930          * array. Return various errors if the user did something wrong.
 931          */
 932         for (i = 0; i < nr_pages; i++) {
 933                 const void *p;
 934
 935                 err = -EFAULT;
 936                 if (get_user(p, pages + i))
 937                         goto out;
 938
 939                 pm[i].addr = (unsigned long)p;
 940                 if (nodes) {
 941                         int node;
 942
 943                         if (get_user(node, nodes + i))
 944                                 goto out;
 945
 946                         err = -ENODEV;
 947                         if (!node_online(node))
 948                                 goto out;
 949
 950                         err = -EACCES;
 951                         if (!node_isset(node, task_nodes))
 952                                 goto out;
 953
 954                         pm[i].node = node;
 955                 }
 956         }
 957         /* End marker */
 958         pm[nr_pages].node = MAX_NUMNODES;
 959
 960         if (nodes)
 961                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
 962         else
 963                 err = do_pages_stat(mm, pm);
 964
 965         if (err >= 0)
 966                 /* Return status information */
 967                 for (i = 0; i < nr_pages; i++)
 968                         if (put_user(pm[i].status, status + i))
 969                                 err = -EFAULT;
 970
 971 out:
 972         vfree(pm);
 973 out2:
 974         mmput(mm);
 975         return err;
 976 }
 977 #endif
 978
 979 /*
 980  * Call migration functions in the vma_ops that may prepare
 981  * memory in a vm for migration. migration functions may perform
 982  * the migration for vmas that do not have an underlying page struct.
 983  */
 984 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
 985         const nodemask_t *from, unsigned long flags)
 986 {
 987         struct vm_area_struct *vma;
 988         int err = 0;
 989
 990         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
 991                 if (vma->vm_ops && vma->vm_ops->migrate) {
 992                         err = vma->vm_ops->migrate(vma, to, from, flags);
 993                         if (err)
 994                                 break;
 995                 }
 996         }
 997         return err;
 998 }