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/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
  33 #include "internal.h"
  34
  35 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  36
  37 /*
  38  * Isolate one page from the LRU lists. If successful put it onto
  39  * the indicated list with elevated page count.
  40  *
  41  * Result:
  42  *  -EBUSY: page not on LRU list
  43  *  0: page removed from LRU list and added to the specified list.
  44  */
  45 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  46 {
  47         int ret = -EBUSY;
  48
  49         if (PageLRU(page)) {
  50                 struct zone *zone = page_zone(page);
  51
  52                 spin_lock_irq(&zone->lru_lock);
  53                 if (PageLRU(page) && get_page_unless_zero(page)) {
  54                         ret = 0;
  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 /*
 119  * Restore a potential migration pte to a working pte entry
 120  */
 121 static void remove_migration_pte(struct vm_area_struct *vma,
 122                 struct page *old, struct page *new)
 123 {
 124         struct mm_struct *mm = vma->vm_mm;
 125         swp_entry_t entry;
 126         pgd_t *pgd;
 127         pud_t *pud;
 128         pmd_t *pmd;
 129         pte_t *ptep, pte;
 130         spinlock_t *ptl;
 131         unsigned long addr = page_address_in_vma(new, vma);
 132
 133         if (addr == -EFAULT)
 134                 return;
 135
 136         pgd = pgd_offset(mm, addr);
 137         if (!pgd_present(*pgd))
 138                 return;
 139
 140         pud = pud_offset(pgd, addr);
 141         if (!pud_present(*pud))
 142                 return;
 143
 144         pmd = pmd_offset(pud, addr);
 145         if (!pmd_present(*pmd))
 146                 return;
 147
 148         ptep = pte_offset_map(pmd, addr);
 149
 150         if (!is_swap_pte(*ptep)) {
 151                 pte_unmap(ptep);
 152                 return;
 153         }
 154
 155         ptl = pte_lockptr(mm, pmd);
 156         spin_lock(ptl);
 157         pte = *ptep;
 158         if (!is_swap_pte(pte))
 159                 goto out;
 160
 161         entry = pte_to_swp_entry(pte);
 162
 163         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 164                 goto out;
 165
 166         get_page(new);
 167         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 168         if (is_write_migration_entry(entry))
 169                 pte = pte_mkwrite(pte);
 170         flush_cache_page(vma, addr, pte_pfn(pte));
 171         set_pte_at(mm, addr, ptep, pte);
 172
 173         if (PageAnon(new))
 174                 page_add_anon_rmap(new, vma, addr);
 175         else
 176                 page_add_file_rmap(new);
 177
 178         /* No need to invalidate - it was non-present before */
 179         update_mmu_cache(vma, addr, pte);
 180
 181 out:
 182         pte_unmap_unlock(ptep, ptl);
 183 }
 184
 185 /*
 186  * Note that remove_file_migration_ptes will only work on regular mappings,
 187  * Nonlinear mappings do not use migration entries.
 188  */
 189 static void remove_file_migration_ptes(struct page *old, struct page *new)
 190 {
 191         struct vm_area_struct *vma;
 192         struct address_space *mapping = page_mapping(new);
 193         struct prio_tree_iter iter;
 194         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 195
 196         if (!mapping)
 197                 return;
 198
 199         spin_lock(&mapping->i_mmap_lock);
 200
 201         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 202                 remove_migration_pte(vma, old, new);
 203
 204         spin_unlock(&mapping->i_mmap_lock);
 205 }
 206
 207 /*
 208  * Must hold mmap_sem lock on at least one of the vmas containing
 209  * the page so that the anon_vma cannot vanish.
 210  */
 211 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 212 {
 213         struct anon_vma *anon_vma;
 214         struct vm_area_struct *vma;
 215         unsigned long mapping;
 216
 217         mapping = (unsigned long)new->mapping;
 218
 219         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 220                 return;
 221
 222         /*
 223          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 224          */
 225         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 226         spin_lock(&anon_vma->lock);
 227
 228         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 229                 remove_migration_pte(vma, old, new);
 230
 231         spin_unlock(&anon_vma->lock);
 232 }
 233
 234 /*
 235  * Get rid of all migration entries and replace them by
 236  * references to the indicated page.
 237  */
 238 static void remove_migration_ptes(struct page *old, struct page *new)
 239 {
 240         if (PageAnon(new))
 241                 remove_anon_migration_ptes(old, new);
 242         else
 243                 remove_file_migration_ptes(old, new);
 244 }
 245
 246 /*
 247  * Something used the pte of a page under migration. We need to
 248  * get to the page and wait until migration is finished.
 249  * When we return from this function the fault will be retried.
 250  *
 251  * This function is called from do_swap_page().
 252  */
 253 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 254                                 unsigned long address)
 255 {
 256         pte_t *ptep, pte;
 257         spinlock_t *ptl;
 258         swp_entry_t entry;
 259         struct page *page;
 260
 261         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 262         pte = *ptep;
 263         if (!is_swap_pte(pte))
 264                 goto out;
 265
 266         entry = pte_to_swp_entry(pte);
 267         if (!is_migration_entry(entry))
 268                 goto out;
 269
 270         page = migration_entry_to_page(entry);
 271
 272         get_page(page);
 273         pte_unmap_unlock(ptep, ptl);
 274         wait_on_page_locked(page);
 275         put_page(page);
 276         return;
 277 out:
 278         pte_unmap_unlock(ptep, ptl);
 279 }
 280
 281 /*
 282  * Replace the page in the mapping.
 283  *
 284  * The number of remaining references must be:
 285  * 1 for anonymous pages without a mapping
 286  * 2 for pages with a mapping
 287  * 3 for pages with a mapping and PagePrivate set.
 288  */
 289 static int migrate_page_move_mapping(struct address_space *mapping,
 290                 struct page *newpage, struct page *page)
 291 {
 292         void **pslot;
 293
 294         if (!mapping) {
 295                 /* Anonymous page without mapping */
 296                 if (page_count(page) != 1)
 297                         return -EAGAIN;
 298                 return 0;
 299         }
 300
 301         write_lock_irq(&mapping->tree_lock);
 302
 303         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 304                                         page_index(page));
 305
 306         if (page_count(page) != 2 + !!PagePrivate(page) ||
 307                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 308                 write_unlock_irq(&mapping->tree_lock);
 309                 return -EAGAIN;
 310         }
 311
 312         /*
 313          * Now we know that no one else is looking at the page.
 314          */
 315         get_page(newpage);      /* add cache reference */
 316 #ifdef CONFIG_SWAP
 317         if (PageSwapCache(page)) {
 318                 SetPageSwapCache(newpage);
 319                 set_page_private(newpage, page_private(page));
 320         }
 321 #endif
 322
 323         radix_tree_replace_slot(pslot, newpage);
 324
 325         /*
 326          * Drop cache reference from old page.
 327          * We know this isn't the last reference.
 328          */
 329         __put_page(page);
 330
 331         /*
 332          * If moved to a different zone then also account
 333          * the page for that zone. Other VM counters will be
 334          * taken care of when we establish references to the
 335          * new page and drop references to the old page.
 336          *
 337          * Note that anonymous pages are accounted for
 338          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 339          * are mapped to swap space.
 340          */
 341         __dec_zone_page_state(page, NR_FILE_PAGES);
 342         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 343
 344         write_unlock_irq(&mapping->tree_lock);
 345
 346         return 0;
 347 }
 348
 349 /*
 350  * Copy the page to its new location
 351  */
 352 static void migrate_page_copy(struct page *newpage, struct page *page)
 353 {
 354         copy_highpage(newpage, page);
 355
 356         if (PageError(page))
 357                 SetPageError(newpage);
 358         if (PageReferenced(page))
 359                 SetPageReferenced(newpage);
 360         if (PageUptodate(page))
 361                 SetPageUptodate(newpage);
 362         if (PageActive(page))
 363                 SetPageActive(newpage);
 364         if (PageChecked(page))
 365                 SetPageChecked(newpage);
 366         if (PageMappedToDisk(page))
 367                 SetPageMappedToDisk(newpage);
 368
 369         if (PageDirty(page)) {
 370                 clear_page_dirty_for_io(page);
 371                 set_page_dirty(newpage);
 372         }
 373
 374 #ifdef CONFIG_SWAP
 375         ClearPageSwapCache(page);
 376 #endif
 377         ClearPageActive(page);
 378         ClearPagePrivate(page);
 379         set_page_private(page, 0);
 380         page->mapping = NULL;
 381
 382         /*
 383          * If any waiters have accumulated on the new page then
 384          * wake them up.
 385          */
 386         if (PageWriteback(newpage))
 387                 end_page_writeback(newpage);
 388 }
 389
 390 /************************************************************
 391  *                    Migration functions
 392  ***********************************************************/
 393
 394 /* Always fail migration. Used for mappings that are not movable */
 395 int fail_migrate_page(struct address_space *mapping,
 396                         struct page *newpage, struct page *page)
 397 {
 398         return -EIO;
 399 }
 400 EXPORT_SYMBOL(fail_migrate_page);
 401
 402 /*
 403  * Common logic to directly migrate a single page suitable for
 404  * pages that do not use PagePrivate.
 405  *
 406  * Pages are locked upon entry and exit.
 407  */
 408 int migrate_page(struct address_space *mapping,
 409                 struct page *newpage, struct page *page)
 410 {
 411         int rc;
 412
 413         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 414
 415         rc = migrate_page_move_mapping(mapping, newpage, page);
 416
 417         if (rc)
 418                 return rc;
 419
 420         migrate_page_copy(newpage, page);
 421         return 0;
 422 }
 423 EXPORT_SYMBOL(migrate_page);
 424
 425 #ifdef CONFIG_BLOCK
 426 /*
 427  * Migration function for pages with buffers. This function can only be used
 428  * if the underlying filesystem guarantees that no other references to "page"
 429  * exist.
 430  */
 431 int buffer_migrate_page(struct address_space *mapping,
 432                 struct page *newpage, struct page *page)
 433 {
 434         struct buffer_head *bh, *head;
 435         int rc;
 436
 437         if (!page_has_buffers(page))
 438                 return migrate_page(mapping, newpage, page);
 439
 440         head = page_buffers(page);
 441
 442         rc = migrate_page_move_mapping(mapping, newpage, page);
 443
 444         if (rc)
 445                 return rc;
 446
 447         bh = head;
 448         do {
 449                 get_bh(bh);
 450                 lock_buffer(bh);
 451                 bh = bh->b_this_page;
 452
 453         } while (bh != head);
 454
 455         ClearPagePrivate(page);
 456         set_page_private(newpage, page_private(page));
 457         set_page_private(page, 0);
 458         put_page(page);
 459         get_page(newpage);
 460
 461         bh = head;
 462         do {
 463                 set_bh_page(bh, newpage, bh_offset(bh));
 464                 bh = bh->b_this_page;
 465
 466         } while (bh != head);
 467
 468         SetPagePrivate(newpage);
 469
 470         migrate_page_copy(newpage, page);
 471
 472         bh = head;
 473         do {
 474                 unlock_buffer(bh);
 475                 put_bh(bh);
 476                 bh = bh->b_this_page;
 477
 478         } while (bh != head);
 479
 480         return 0;
 481 }
 482 EXPORT_SYMBOL(buffer_migrate_page);
 483 #endif
 484
 485 /*
 486  * Writeback a page to clean the dirty state
 487  */
 488 static int writeout(struct address_space *mapping, struct page *page)
 489 {
 490         struct writeback_control wbc = {
 491                 .sync_mode = WB_SYNC_NONE,
 492                 .nr_to_write = 1,
 493                 .range_start = 0,
 494                 .range_end = LLONG_MAX,
 495                 .nonblocking = 1,
 496                 .for_reclaim = 1
 497         };
 498         int rc;
 499
 500         if (!mapping->a_ops->writepage)
 501                 /* No write method for the address space */
 502                 return -EINVAL;
 503
 504         if (!clear_page_dirty_for_io(page))
 505                 /* Someone else already triggered a write */
 506                 return -EAGAIN;
 507
 508         /*
 509          * A dirty page may imply that the underlying filesystem has
 510          * the page on some queue. So the page must be clean for
 511          * migration. Writeout may mean we loose the lock and the
 512          * page state is no longer what we checked for earlier.
 513          * At this point we know that the migration attempt cannot
 514          * be successful.
 515          */
 516         remove_migration_ptes(page, page);
 517
 518         rc = mapping->a_ops->writepage(page, &wbc);
 519         if (rc < 0)
 520                 /* I/O Error writing */
 521                 return -EIO;
 522
 523         if (rc != AOP_WRITEPAGE_ACTIVATE)
 524                 /* unlocked. Relock */
 525                 lock_page(page);
 526
 527         return -EAGAIN;
 528 }
 529
 530 /*
 531  * Default handling if a filesystem does not provide a migration function.
 532  */
 533 static int fallback_migrate_page(struct address_space *mapping,
 534         struct page *newpage, struct page *page)
 535 {
 536         if (PageDirty(page))
 537                 return writeout(mapping, page);
 538
 539         /*
 540          * Buffers may be managed in a filesystem specific way.
 541          * We must have no buffers or drop them.
 542          */
 543         if (PagePrivate(page) &&
 544             !try_to_release_page(page, GFP_KERNEL))
 545                 return -EAGAIN;
 546
 547         return migrate_page(mapping, newpage, page);
 548 }
 549
 550 /*
 551  * Move a page to a newly allocated page
 552  * The page is locked and all ptes have been successfully removed.
 553  *
 554  * The new page will have replaced the old page if this function
 555  * is successful.
 556  */
 557 static int move_to_new_page(struct page *newpage, struct page *page)
 558 {
 559         struct address_space *mapping;
 560         int rc;
 561
 562         /*
 563          * Block others from accessing the page when we get around to
 564          * establishing additional references. We are the only one
 565          * holding a reference to the new page at this point.
 566          */
 567         if (TestSetPageLocked(newpage))
 568                 BUG();
 569
 570         /* Prepare mapping for the new page.*/
 571         newpage->index = page->index;
 572         newpage->mapping = page->mapping;
 573
 574         mapping = page_mapping(page);
 575         if (!mapping)
 576                 rc = migrate_page(mapping, newpage, page);
 577         else if (mapping->a_ops->migratepage)
 578                 /*
 579                  * Most pages have a mapping and most filesystems
 580                  * should provide a migration function. Anonymous
 581                  * pages are part of swap space which also has its
 582                  * own migration function. This is the most common
 583                  * path for page migration.
 584                  */
 585                 rc = mapping->a_ops->migratepage(mapping,
 586                                                 newpage, page);
 587         else
 588                 rc = fallback_migrate_page(mapping, newpage, page);
 589
 590         if (!rc)
 591                 remove_migration_ptes(page, newpage);
 592         else
 593                 newpage->mapping = NULL;
 594
 595         unlock_page(newpage);
 596
 597         return rc;
 598 }
 599
 600 /*
 601  * Obtain the lock on page, remove all ptes and migrate the page
 602  * to the newly allocated page in newpage.
 603  */
 604 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 605                         struct page *page, int force)
 606 {
 607         int rc = 0;
 608         int *result = NULL;
 609         struct page *newpage = get_new_page(page, private, &result);
 610         int rcu_locked = 0;
 611
 612         if (!newpage)
 613                 return -ENOMEM;
 614
 615         if (page_count(page) == 1)
 616                 /* page was freed from under us. So we are done. */
 617                 goto move_newpage;
 618
 619         rc = -EAGAIN;
 620         if (TestSetPageLocked(page)) {
 621                 if (!force)
 622                         goto move_newpage;
 623                 lock_page(page);
 624         }
 625
 626         if (PageWriteback(page)) {
 627                 if (!force)
 628                         goto unlock;
 629                 wait_on_page_writeback(page);
 630         }
 631         /*
 632          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 633          * we cannot notice that anon_vma is freed while we migrates a page.
 634          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 635          * of migration. File cache pages are no problem because of page_lock()
 636          * File Caches may use write_page() or lock_page() in migration, then,
 637          * just care Anon page here.
 638          */
 639         if (PageAnon(page)) {
 640                 rcu_read_lock();
 641                 rcu_locked = 1;
 642         }
 643
 644         /*
 645          * Corner case handling:
 646          * 1. When a new swap-cache page is read into, it is added to the LRU
 647          * and treated as swapcache but it has no rmap yet.
 648          * Calling try_to_unmap() against a page->mapping==NULL page will
 649          * trigger a BUG.  So handle it here.
 650          * 2. An orphaned page (see truncate_complete_page) might have
 651          * fs-private metadata. The page can be picked up due to memory
 652          * offlining.  Everywhere else except page reclaim, the page is
 653          * invisible to the vm, so the page can not be migrated.  So try to
 654          * free the metadata, so the page can be freed.
 655          */
 656         if (!page->mapping) {
 657                 if (!PageAnon(page) && PagePrivate(page)) {
 658                         /*
 659                          * Go direct to try_to_free_buffers() here because
 660                          * a) that's what try_to_release_page() would do anyway
 661                          * b) we may be under rcu_read_lock() here, so we can't
 662                          *    use GFP_KERNEL which is what try_to_release_page()
 663                          *    needs to be effective.
 664                          */
 665                         try_to_free_buffers(page);
 666                 }
 667                 goto rcu_unlock;
 668         }
 669
 670         /* Establish migration ptes or remove ptes */
 671         try_to_unmap(page, 1);
 672
 673         if (!page_mapped(page))
 674                 rc = move_to_new_page(newpage, page);
 675
 676         if (rc)
 677                 remove_migration_ptes(page, page);
 678 rcu_unlock:
 679         if (rcu_locked)
 680                 rcu_read_unlock();
 681
 682 unlock:
 683
 684         unlock_page(page);
 685
 686         if (rc != -EAGAIN) {
 687                 /*
 688                  * A page that has been migrated has all references
 689                  * removed and will be freed. A page that has not been
 690                  * migrated will have kepts its references and be
 691                  * restored.
 692                  */
 693                 list_del(&page->lru);
 694                 move_to_lru(page);
 695         }
 696
 697 move_newpage:
 698         /*
 699          * Move the new page to the LRU. If migration was not successful
 700          * then this will free the page.
 701          */
 702         move_to_lru(newpage);
 703         if (result) {
 704                 if (rc)
 705                         *result = rc;
 706                 else
 707                         *result = page_to_nid(newpage);
 708         }
 709         return rc;
 710 }
 711
 712 /*
 713  * migrate_pages
 714  *
 715  * The function takes one list of pages to migrate and a function
 716  * that determines from the page to be migrated and the private data
 717  * the target of the move and allocates the page.
 718  *
 719  * The function returns after 10 attempts or if no pages
 720  * are movable anymore because to has become empty
 721  * or no retryable pages exist anymore. All pages will be
 722  * returned to the LRU or freed.
 723  *
 724  * Return: Number of pages not migrated or error code.
 725  */
 726 int migrate_pages(struct list_head *from,
 727                 new_page_t get_new_page, unsigned long private)
 728 {
 729         int retry = 1;
 730         int nr_failed = 0;
 731         int pass = 0;
 732         struct page *page;
 733         struct page *page2;
 734         int swapwrite = current->flags & PF_SWAPWRITE;
 735         int rc;
 736
 737         if (!swapwrite)
 738                 current->flags |= PF_SWAPWRITE;
 739
 740         for(pass = 0; pass < 10 && retry; pass++) {
 741                 retry = 0;
 742
 743                 list_for_each_entry_safe(page, page2, from, lru) {
 744                         cond_resched();
 745
 746                         rc = unmap_and_move(get_new_page, private,
 747                                                 page, pass > 2);
 748
 749                         switch(rc) {
 750                         case -ENOMEM:
 751                                 goto out;
 752                         case -EAGAIN:
 753                                 retry++;
 754                                 break;
 755                         case 0:
 756                                 break;
 757                         default:
 758                                 /* Permanent failure */
 759                                 nr_failed++;
 760                                 break;
 761                         }
 762                 }
 763         }
 764         rc = 0;
 765 out:
 766         if (!swapwrite)
 767                 current->flags &= ~PF_SWAPWRITE;
 768
 769         putback_lru_pages(from);
 770
 771         if (rc)
 772                 return rc;
 773
 774         return nr_failed + retry;
 775 }
 776
 777 #ifdef CONFIG_NUMA
 778 /*
 779  * Move a list of individual pages
 780  */
 781 struct page_to_node {
 782         unsigned long addr;
 783         struct page *page;
 784         int node;
 785         int status;
 786 };
 787
 788 static struct page *new_page_node(struct page *p, unsigned long private,
 789                 int **result)
 790 {
 791         struct page_to_node *pm = (struct page_to_node *)private;
 792
 793         while (pm->node != MAX_NUMNODES && pm->page != p)
 794                 pm++;
 795
 796         if (pm->node == MAX_NUMNODES)
 797                 return NULL;
 798
 799         *result = &pm->status;
 800
 801         return alloc_pages_node(pm->node,
 802                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 803 }
 804
 805 /*
 806  * Move a set of pages as indicated in the pm array. The addr
 807  * field must be set to the virtual address of the page to be moved
 808  * and the node number must contain a valid target node.
 809  */
 810 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 811                                 int migrate_all)
 812 {
 813         int err;
 814         struct page_to_node *pp;
 815         LIST_HEAD(pagelist);
 816
 817         down_read(&mm->mmap_sem);
 818
 819         /*
 820          * Build a list of pages to migrate
 821          */
 822         migrate_prep();
 823         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 824                 struct vm_area_struct *vma;
 825                 struct page *page;
 826
 827                 /*
 828                  * A valid page pointer that will not match any of the
 829                  * pages that will be moved.
 830                  */
 831                 pp->page = ZERO_PAGE(0);
 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                 err = -ENOENT;
 840                 if (!page)
 841                         goto set_status;
 842
 843                 if (PageReserved(page))         /* Check for zero page */
 844                         goto put_and_set;
 845
 846                 pp->page = page;
 847                 err = page_to_nid(page);
 848
 849                 if (err == pp->node)
 850                         /*
 851                          * Node already in the right place
 852                          */
 853                         goto put_and_set;
 854
 855                 err = -EACCES;
 856                 if (page_mapcount(page) > 1 &&
 857                                 !migrate_all)
 858                         goto put_and_set;
 859
 860                 err = isolate_lru_page(page, &pagelist);
 861 put_and_set:
 862                 /*
 863                  * Either remove the duplicate refcount from
 864                  * isolate_lru_page() or drop the page ref if it was
 865                  * not isolated.
 866                  */
 867                 put_page(page);
 868 set_status:
 869                 pp->status = err;
 870         }
 871
 872         if (!list_empty(&pagelist))
 873                 err = migrate_pages(&pagelist, new_page_node,
 874                                 (unsigned long)pm);
 875         else
 876                 err = -ENOENT;
 877
 878         up_read(&mm->mmap_sem);
 879         return err;
 880 }
 881
 882 /*
 883  * Determine the nodes of a list of pages. The addr in the pm array
 884  * must have been set to the virtual address of which we want to determine
 885  * the node number.
 886  */
 887 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 888 {
 889         down_read(&mm->mmap_sem);
 890
 891         for ( ; pm->node != MAX_NUMNODES; pm++) {
 892                 struct vm_area_struct *vma;
 893                 struct page *page;
 894                 int err;
 895
 896                 err = -EFAULT;
 897                 vma = find_vma(mm, pm->addr);
 898                 if (!vma)
 899                         goto set_status;
 900
 901                 page = follow_page(vma, pm->addr, 0);
 902                 err = -ENOENT;
 903                 /* Use PageReserved to check for zero page */
 904                 if (!page || PageReserved(page))
 905                         goto set_status;
 906
 907                 err = page_to_nid(page);
 908 set_status:
 909                 pm->status = err;
 910         }
 911
 912         up_read(&mm->mmap_sem);
 913         return 0;
 914 }
 915
 916 /*
 917  * Move a list of pages in the address space of the currently executing
 918  * process.
 919  */
 920 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 921                         const void __user * __user *pages,
 922                         const int __user *nodes,
 923                         int __user *status, int flags)
 924 {
 925         int err = 0;
 926         int i;
 927         struct task_struct *task;
 928         nodemask_t task_nodes;
 929         struct mm_struct *mm;
 930         struct page_to_node *pm = NULL;
 931
 932         /* Check flags */
 933         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 934                 return -EINVAL;
 935
 936         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 937                 return -EPERM;
 938
 939         /* Find the mm_struct */
 940         read_lock(&tasklist_lock);
 941         task = pid ? find_task_by_vpid(pid) : current;
 942         if (!task) {
 943                 read_unlock(&tasklist_lock);
 944                 return -ESRCH;
 945         }
 946         mm = get_task_mm(task);
 947         read_unlock(&tasklist_lock);
 948
 949         if (!mm)
 950                 return -EINVAL;
 951
 952         /*
 953          * Check if this process has the right to modify the specified
 954          * process. The right exists if the process has administrative
 955          * capabilities, superuser privileges or the same
 956          * userid as the target process.
 957          */
 958         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 959             (current->uid != task->suid) && (current->uid != task->uid) &&
 960             !capable(CAP_SYS_NICE)) {
 961                 err = -EPERM;
 962                 goto out2;
 963         }
 964
 965         err = security_task_movememory(task);
 966         if (err)
 967                 goto out2;
 968
 969
 970         task_nodes = cpuset_mems_allowed(task);
 971
 972         /* Limit nr_pages so that the multiplication may not overflow */
 973         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 974                 err = -E2BIG;
 975                 goto out2;
 976         }
 977
 978         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
 979         if (!pm) {
 980                 err = -ENOMEM;
 981                 goto out2;
 982         }
 983
 984         /*
 985          * Get parameters from user space and initialize the pm
 986          * array. Return various errors if the user did something wrong.
 987          */
 988         for (i = 0; i < nr_pages; i++) {
 989                 const void __user *p;
 990
 991                 err = -EFAULT;
 992                 if (get_user(p, pages + i))
 993                         goto out;
 994
 995                 pm[i].addr = (unsigned long)p;
 996                 if (nodes) {
 997                         int node;
 998
 999                         if (get_user(node, nodes + i))
1000                                 goto out;
1001
1002                         err = -ENODEV;
1003                         if (!node_state(node, N_HIGH_MEMORY))
1004                                 goto out;
1005
1006                         err = -EACCES;
1007                         if (!node_isset(node, task_nodes))
1008                                 goto out;
1009
1010                         pm[i].node = node;
1011                 } else
1012                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1013         }
1014         /* End marker */
1015         pm[nr_pages].node = MAX_NUMNODES;
1016
1017         if (nodes)
1018                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1019         else
1020                 err = do_pages_stat(mm, pm);
1021
1022         if (err >= 0)
1023                 /* Return status information */
1024                 for (i = 0; i < nr_pages; i++)
1025                         if (put_user(pm[i].status, status + i))
1026                                 err = -EFAULT;
1027
1028 out:
1029         vfree(pm);
1030 out2:
1031         mmput(mm);
1032         return err;
1033 }
1034 #endif
1035
1036 /*
1037  * Call migration functions in the vma_ops that may prepare
1038  * memory in a vm for migration. migration functions may perform
1039  * the migration for vmas that do not have an underlying page struct.
1040  */
1041 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1042         const nodemask_t *from, unsigned long flags)
1043 {
1044         struct vm_area_struct *vma;
1045         int err = 0;
1046
1047         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1048                 if (vma->vm_ops && vma->vm_ops->migrate) {
1049                         err = vma->vm_ops->migrate(vma, to, from, flags);
1050                         if (err)
1051                                 break;
1052                 }
1053         }
1054         return err;
1055 }