mm/migrate.c

   1 /*
   2  * Memory Migration functionality - linux/mm/migration.c
   3  *
   4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5  *
   6  * Page migration was first developed in the context of the memory hotplug
   7  * project. The main authors of the migration code are:
   8  *
   9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10  * Hirokazu Takahashi <taka@valinux.co.jp>
  11  * Dave Hansen <haveblue@us.ibm.com>
  12  * Christoph Lameter
  13  */
  14
  15 #include <linux/migrate.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/swapops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/nsproxy.h>
  23 #include <linux/pagevec.h>
  24 #include <linux/rmap.h>
  25 #include <linux/topology.h>
  26 #include <linux/cpu.h>
  27 #include <linux/cpuset.h>
  28 #include <linux/writeback.h>
  29 #include <linux/mempolicy.h>
  30 #include <linux/vmalloc.h>
  31 #include <linux/security.h>
  32 #include <linux/memcontrol.h>
  33 #include <linux/syscalls.h>
  34
  35 #include "internal.h"
  36
  37 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  38
  39 /*
  40  * migrate_prep() needs to be called before we start compiling a list of pages
  41  * to be migrated using isolate_lru_page().
  42  */
  43 int migrate_prep(void)
  44 {
  45         /*
  46          * Clear the LRU lists so pages can be isolated.
  47          * Note that pages may be moved off the LRU after we have
  48          * drained them. Those pages will fail to migrate like other
  49          * pages that may be busy.
  50          */
  51         lru_add_drain_all();
  52
  53         return 0;
  54 }
  55
  56 /*
  57  * Add isolated pages on the list back to the LRU under page lock
  58  * to avoid leaking evictable pages back onto unevictable list.
  59  *
  60  * returns the number of pages put back.
  61  */
  62 int putback_lru_pages(struct list_head *l)
  63 {
  64         struct page *page;
  65         struct page *page2;
  66         int count = 0;
  67
  68         list_for_each_entry_safe(page, page2, l, lru) {
  69                 list_del(&page->lru);
  70                 dec_zone_page_state(page, NR_ISOLATED_ANON +
  71                                 page_is_file_cache(page));
  72                 putback_lru_page(page);
  73                 count++;
  74         }
  75         return count;
  76 }
  77
  78 /*
  79  * Restore a potential migration pte to a working pte entry
  80  */
  81 static void remove_migration_pte(struct vm_area_struct *vma,
  82                 struct page *old, struct page *new)
  83 {
  84         struct mm_struct *mm = vma->vm_mm;
  85         swp_entry_t entry;
  86         pgd_t *pgd;
  87         pud_t *pud;
  88         pmd_t *pmd;
  89         pte_t *ptep, pte;
  90         spinlock_t *ptl;
  91         unsigned long addr = page_address_in_vma(new, vma);
  92
  93         if (addr == -EFAULT)
  94                 return;
  95
  96         pgd = pgd_offset(mm, addr);
  97         if (!pgd_present(*pgd))
  98                 return;
  99
 100         pud = pud_offset(pgd, addr);
 101         if (!pud_present(*pud))
 102                 return;
 103
 104         pmd = pmd_offset(pud, addr);
 105         if (!pmd_present(*pmd))
 106                 return;
 107
 108         ptep = pte_offset_map(pmd, addr);
 109
 110         if (!is_swap_pte(*ptep)) {
 111                 pte_unmap(ptep);
 112                 return;
 113         }
 114
 115         ptl = pte_lockptr(mm, pmd);
 116         spin_lock(ptl);
 117         pte = *ptep;
 118         if (!is_swap_pte(pte))
 119                 goto out;
 120
 121         entry = pte_to_swp_entry(pte);
 122
 123         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 124                 goto out;
 125
 126         get_page(new);
 127         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 128         if (is_write_migration_entry(entry))
 129                 pte = pte_mkwrite(pte);
 130         flush_cache_page(vma, addr, pte_pfn(pte));
 131         set_pte_at(mm, addr, ptep, pte);
 132
 133         if (PageAnon(new))
 134                 page_add_anon_rmap(new, vma, addr);
 135         else
 136                 page_add_file_rmap(new);
 137
 138         /* No need to invalidate - it was non-present before */
 139         update_mmu_cache(vma, addr, pte);
 140
 141 out:
 142         pte_unmap_unlock(ptep, ptl);
 143 }
 144
 145 /*
 146  * Note that remove_file_migration_ptes will only work on regular mappings,
 147  * Nonlinear mappings do not use migration entries.
 148  */
 149 static void remove_file_migration_ptes(struct page *old, struct page *new)
 150 {
 151         struct vm_area_struct *vma;
 152         struct address_space *mapping = new->mapping;
 153         struct prio_tree_iter iter;
 154         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 155
 156         if (!mapping)
 157                 return;
 158
 159         spin_lock(&mapping->i_mmap_lock);
 160
 161         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 162                 remove_migration_pte(vma, old, new);
 163
 164         spin_unlock(&mapping->i_mmap_lock);
 165 }
 166
 167 /*
 168  * Must hold mmap_sem lock on at least one of the vmas containing
 169  * the page so that the anon_vma cannot vanish.
 170  */
 171 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 172 {
 173         struct anon_vma *anon_vma;
 174         struct vm_area_struct *vma;
 175         unsigned long mapping;
 176
 177         mapping = (unsigned long)new->mapping;
 178
 179         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 180                 return;
 181
 182         /*
 183          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 184          */
 185         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 186         spin_lock(&anon_vma->lock);
 187
 188         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 189                 remove_migration_pte(vma, old, new);
 190
 191         spin_unlock(&anon_vma->lock);
 192 }
 193
 194 /*
 195  * Get rid of all migration entries and replace them by
 196  * references to the indicated page.
 197  */
 198 static void remove_migration_ptes(struct page *old, struct page *new)
 199 {
 200         if (PageAnon(new))
 201                 remove_anon_migration_ptes(old, new);
 202         else
 203                 remove_file_migration_ptes(old, new);
 204 }
 205
 206 /*
 207  * Something used the pte of a page under migration. We need to
 208  * get to the page and wait until migration is finished.
 209  * When we return from this function the fault will be retried.
 210  *
 211  * This function is called from do_swap_page().
 212  */
 213 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 214                                 unsigned long address)
 215 {
 216         pte_t *ptep, pte;
 217         spinlock_t *ptl;
 218         swp_entry_t entry;
 219         struct page *page;
 220
 221         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 222         pte = *ptep;
 223         if (!is_swap_pte(pte))
 224                 goto out;
 225
 226         entry = pte_to_swp_entry(pte);
 227         if (!is_migration_entry(entry))
 228                 goto out;
 229
 230         page = migration_entry_to_page(entry);
 231
 232         /*
 233          * Once radix-tree replacement of page migration started, page_count
 234          * *must* be zero. And, we don't want to call wait_on_page_locked()
 235          * against a page without get_page().
 236          * So, we use get_page_unless_zero(), here. Even failed, page fault
 237          * will occur again.
 238          */
 239         if (!get_page_unless_zero(page))
 240                 goto out;
 241         pte_unmap_unlock(ptep, ptl);
 242         wait_on_page_locked(page);
 243         put_page(page);
 244         return;
 245 out:
 246         pte_unmap_unlock(ptep, ptl);
 247 }
 248
 249 /*
 250  * Replace the page in the mapping.
 251  *
 252  * The number of remaining references must be:
 253  * 1 for anonymous pages without a mapping
 254  * 2 for pages with a mapping
 255  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 256  */
 257 static int migrate_page_move_mapping(struct address_space *mapping,
 258                 struct page *newpage, struct page *page)
 259 {
 260         int expected_count;
 261         void **pslot;
 262
 263         if (!mapping) {
 264                 /* Anonymous page without mapping */
 265                 if (page_count(page) != 1)
 266                         return -EAGAIN;
 267                 return 0;
 268         }
 269
 270         spin_lock_irq(&mapping->tree_lock);
 271
 272         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 273                                         page_index(page));
 274
 275         expected_count = 2 + page_has_private(page);
 276         if (page_count(page) != expected_count ||
 277                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 278                 spin_unlock_irq(&mapping->tree_lock);
 279                 return -EAGAIN;
 280         }
 281
 282         if (!page_freeze_refs(page, expected_count)) {
 283                 spin_unlock_irq(&mapping->tree_lock);
 284                 return -EAGAIN;
 285         }
 286
 287         /*
 288          * Now we know that no one else is looking at the page.
 289          */
 290         get_page(newpage);      /* add cache reference */
 291         if (PageSwapCache(page)) {
 292                 SetPageSwapCache(newpage);
 293                 set_page_private(newpage, page_private(page));
 294         }
 295
 296         radix_tree_replace_slot(pslot, newpage);
 297
 298         page_unfreeze_refs(page, expected_count);
 299         /*
 300          * Drop cache reference from old page.
 301          * We know this isn't the last reference.
 302          */
 303         __put_page(page);
 304
 305         /*
 306          * If moved to a different zone then also account
 307          * the page for that zone. Other VM counters will be
 308          * taken care of when we establish references to the
 309          * new page and drop references to the old page.
 310          *
 311          * Note that anonymous pages are accounted for
 312          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 313          * are mapped to swap space.
 314          */
 315         __dec_zone_page_state(page, NR_FILE_PAGES);
 316         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 317         if (!PageSwapCache(page) && PageSwapBacked(page)) {
 318                 __dec_zone_page_state(page, NR_SHMEM);
 319                 __inc_zone_page_state(newpage, NR_SHMEM);
 320         }
 321         spin_unlock_irq(&mapping->tree_lock);
 322
 323         return 0;
 324 }
 325
 326 /*
 327  * Copy the page to its new location
 328  */
 329 static void migrate_page_copy(struct page *newpage, struct page *page)
 330 {
 331         int anon;
 332
 333         copy_highpage(newpage, page);
 334
 335         if (PageError(page))
 336                 SetPageError(newpage);
 337         if (PageReferenced(page))
 338                 SetPageReferenced(newpage);
 339         if (PageUptodate(page))
 340                 SetPageUptodate(newpage);
 341         if (TestClearPageActive(page)) {
 342                 VM_BUG_ON(PageUnevictable(page));
 343                 SetPageActive(newpage);
 344         } else
 345                 unevictable_migrate_page(newpage, page);
 346         if (PageChecked(page))
 347                 SetPageChecked(newpage);
 348         if (PageMappedToDisk(page))
 349                 SetPageMappedToDisk(newpage);
 350
 351         if (PageDirty(page)) {
 352                 clear_page_dirty_for_io(page);
 353                 /*
 354                  * Want to mark the page and the radix tree as dirty, and
 355                  * redo the accounting that clear_page_dirty_for_io undid,
 356                  * but we can't use set_page_dirty because that function
 357                  * is actually a signal that all of the page has become dirty.
 358                  * Wheras only part of our page may be dirty.
 359                  */
 360                 __set_page_dirty_nobuffers(newpage);
 361         }
 362
 363         mlock_migrate_page(newpage, page);
 364
 365         ClearPageSwapCache(page);
 366         ClearPagePrivate(page);
 367         set_page_private(page, 0);
 368         /* page->mapping contains a flag for PageAnon() */
 369         anon = PageAnon(page);
 370         page->mapping = NULL;
 371
 372         /*
 373          * If any waiters have accumulated on the new page then
 374          * wake them up.
 375          */
 376         if (PageWriteback(newpage))
 377                 end_page_writeback(newpage);
 378 }
 379
 380 /************************************************************
 381  *                    Migration functions
 382  ***********************************************************/
 383
 384 /* Always fail migration. Used for mappings that are not movable */
 385 int fail_migrate_page(struct address_space *mapping,
 386                         struct page *newpage, struct page *page)
 387 {
 388         return -EIO;
 389 }
 390 EXPORT_SYMBOL(fail_migrate_page);
 391
 392 /*
 393  * Common logic to directly migrate a single page suitable for
 394  * pages that do not use PagePrivate/PagePrivate2.
 395  *
 396  * Pages are locked upon entry and exit.
 397  */
 398 int migrate_page(struct address_space *mapping,
 399                 struct page *newpage, struct page *page)
 400 {
 401         int rc;
 402
 403         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 404
 405         rc = migrate_page_move_mapping(mapping, newpage, page);
 406
 407         if (rc)
 408                 return rc;
 409
 410         migrate_page_copy(newpage, page);
 411         return 0;
 412 }
 413 EXPORT_SYMBOL(migrate_page);
 414
 415 #ifdef CONFIG_BLOCK
 416 /*
 417  * Migration function for pages with buffers. This function can only be used
 418  * if the underlying filesystem guarantees that no other references to "page"
 419  * exist.
 420  */
 421 int buffer_migrate_page(struct address_space *mapping,
 422                 struct page *newpage, struct page *page)
 423 {
 424         struct buffer_head *bh, *head;
 425         int rc;
 426
 427         if (!page_has_buffers(page))
 428                 return migrate_page(mapping, newpage, page);
 429
 430         head = page_buffers(page);
 431
 432         rc = migrate_page_move_mapping(mapping, newpage, page);
 433
 434         if (rc)
 435                 return rc;
 436
 437         bh = head;
 438         do {
 439                 get_bh(bh);
 440                 lock_buffer(bh);
 441                 bh = bh->b_this_page;
 442
 443         } while (bh != head);
 444
 445         ClearPagePrivate(page);
 446         set_page_private(newpage, page_private(page));
 447         set_page_private(page, 0);
 448         put_page(page);
 449         get_page(newpage);
 450
 451         bh = head;
 452         do {
 453                 set_bh_page(bh, newpage, bh_offset(bh));
 454                 bh = bh->b_this_page;
 455
 456         } while (bh != head);
 457
 458         SetPagePrivate(newpage);
 459
 460         migrate_page_copy(newpage, page);
 461
 462         bh = head;
 463         do {
 464                 unlock_buffer(bh);
 465                 put_bh(bh);
 466                 bh = bh->b_this_page;
 467
 468         } while (bh != head);
 469
 470         return 0;
 471 }
 472 EXPORT_SYMBOL(buffer_migrate_page);
 473 #endif
 474
 475 /*
 476  * Writeback a page to clean the dirty state
 477  */
 478 static int writeout(struct address_space *mapping, struct page *page)
 479 {
 480         struct writeback_control wbc = {
 481                 .sync_mode = WB_SYNC_NONE,
 482                 .nr_to_write = 1,
 483                 .range_start = 0,
 484                 .range_end = LLONG_MAX,
 485                 .nonblocking = 1,
 486                 .for_reclaim = 1
 487         };
 488         int rc;
 489
 490         if (!mapping->a_ops->writepage)
 491                 /* No write method for the address space */
 492                 return -EINVAL;
 493
 494         if (!clear_page_dirty_for_io(page))
 495                 /* Someone else already triggered a write */
 496                 return -EAGAIN;
 497
 498         /*
 499          * A dirty page may imply that the underlying filesystem has
 500          * the page on some queue. So the page must be clean for
 501          * migration. Writeout may mean we loose the lock and the
 502          * page state is no longer what we checked for earlier.
 503          * At this point we know that the migration attempt cannot
 504          * be successful.
 505          */
 506         remove_migration_ptes(page, page);
 507
 508         rc = mapping->a_ops->writepage(page, &wbc);
 509
 510         if (rc != AOP_WRITEPAGE_ACTIVATE)
 511                 /* unlocked. Relock */
 512                 lock_page(page);
 513
 514         return (rc < 0) ? -EIO : -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 (page_has_private(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  * Return value:
 545  *   < 0 - error code
 546  *  == 0 - success
 547  */
 548 static int move_to_new_page(struct page *newpage, struct page *page)
 549 {
 550         struct address_space *mapping;
 551         int rc;
 552
 553         /*
 554          * Block others from accessing the page when we get around to
 555          * establishing additional references. We are the only one
 556          * holding a reference to the new page at this point.
 557          */
 558         if (!trylock_page(newpage))
 559                 BUG();
 560
 561         /* Prepare mapping for the new page.*/
 562         newpage->index = page->index;
 563         newpage->mapping = page->mapping;
 564         if (PageSwapBacked(page))
 565                 SetPageSwapBacked(newpage);
 566
 567         mapping = page_mapping(page);
 568         if (!mapping)
 569                 rc = migrate_page(mapping, newpage, page);
 570         else if (mapping->a_ops->migratepage)
 571                 /*
 572                  * Most pages have a mapping and most filesystems
 573                  * should provide a migration function. Anonymous
 574                  * pages are part of swap space which also has its
 575                  * own migration function. This is the most common
 576                  * path for page migration.
 577                  */
 578                 rc = mapping->a_ops->migratepage(mapping,
 579                                                 newpage, page);
 580         else
 581                 rc = fallback_migrate_page(mapping, newpage, page);
 582
 583         if (!rc) {
 584                 remove_migration_ptes(page, newpage);
 585         } else
 586                 newpage->mapping = NULL;
 587
 588         unlock_page(newpage);
 589
 590         return rc;
 591 }
 592
 593 /*
 594  * Obtain the lock on page, remove all ptes and migrate the page
 595  * to the newly allocated page in newpage.
 596  */
 597 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 598                         struct page *page, int force)
 599 {
 600         int rc = 0;
 601         int *result = NULL;
 602         struct page *newpage = get_new_page(page, private, &result);
 603         int rcu_locked = 0;
 604         int charge = 0;
 605         struct mem_cgroup *mem = NULL;
 606
 607         if (!newpage)
 608                 return -ENOMEM;
 609
 610         if (page_count(page) == 1) {
 611                 /* page was freed from under us. So we are done. */
 612                 goto move_newpage;
 613         }
 614
 615         /* prepare cgroup just returns 0 or -ENOMEM */
 616         rc = -EAGAIN;
 617
 618         if (!trylock_page(page)) {
 619                 if (!force)
 620                         goto move_newpage;
 621                 lock_page(page);
 622         }
 623
 624         /* charge against new page */
 625         charge = mem_cgroup_prepare_migration(page, &mem);
 626         if (charge == -ENOMEM) {
 627                 rc = -ENOMEM;
 628                 goto unlock;
 629         }
 630         BUG_ON(charge);
 631
 632         if (PageWriteback(page)) {
 633                 if (!force)
 634                         goto uncharge;
 635                 wait_on_page_writeback(page);
 636         }
 637         /*
 638          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 639          * we cannot notice that anon_vma is freed while we migrates a page.
 640          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 641          * of migration. File cache pages are no problem because of page_lock()
 642          * File Caches may use write_page() or lock_page() in migration, then,
 643          * just care Anon page here.
 644          */
 645         if (PageAnon(page)) {
 646                 rcu_read_lock();
 647                 rcu_locked = 1;
 648         }
 649
 650         /*
 651          * Corner case handling:
 652          * 1. When a new swap-cache page is read into, it is added to the LRU
 653          * and treated as swapcache but it has no rmap yet.
 654          * Calling try_to_unmap() against a page->mapping==NULL page will
 655          * trigger a BUG.  So handle it here.
 656          * 2. An orphaned page (see truncate_complete_page) might have
 657          * fs-private metadata. The page can be picked up due to memory
 658          * offlining.  Everywhere else except page reclaim, the page is
 659          * invisible to the vm, so the page can not be migrated.  So try to
 660          * free the metadata, so the page can be freed.
 661          */
 662         if (!page->mapping) {
 663                 if (!PageAnon(page) && page_has_private(page)) {
 664                         /*
 665                          * Go direct to try_to_free_buffers() here because
 666                          * a) that's what try_to_release_page() would do anyway
 667                          * b) we may be under rcu_read_lock() here, so we can't
 668                          *    use GFP_KERNEL which is what try_to_release_page()
 669                          *    needs to be effective.
 670                          */
 671                         try_to_free_buffers(page);
 672                         goto rcu_unlock;
 673                 }
 674                 goto skip_unmap;
 675         }
 676
 677         /* Establish migration ptes or remove ptes */
 678         try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
 679
 680 skip_unmap:
 681         if (!page_mapped(page))
 682                 rc = move_to_new_page(newpage, page);
 683
 684         if (rc)
 685                 remove_migration_ptes(page, page);
 686 rcu_unlock:
 687         if (rcu_locked)
 688                 rcu_read_unlock();
 689 uncharge:
 690         if (!charge)
 691                 mem_cgroup_end_migration(mem, page, newpage);
 692 unlock:
 693         unlock_page(page);
 694
 695         if (rc != -EAGAIN) {
 696                 /*
 697                  * A page that has been migrated has all references
 698                  * removed and will be freed. A page that has not been
 699                  * migrated will have kepts its references and be
 700                  * restored.
 701                  */
 702                 list_del(&page->lru);
 703                 dec_zone_page_state(page, NR_ISOLATED_ANON +
 704                                 page_is_file_cache(page));
 705                 putback_lru_page(page);
 706         }
 707
 708 move_newpage:
 709
 710         /*
 711          * Move the new page to the LRU. If migration was not successful
 712          * then this will free the page.
 713          */
 714         putback_lru_page(newpage);
 715
 716         if (result) {
 717                 if (rc)
 718                         *result = rc;
 719                 else
 720                         *result = page_to_nid(newpage);
 721         }
 722         return rc;
 723 }
 724
 725 /*
 726  * migrate_pages
 727  *
 728  * The function takes one list of pages to migrate and a function
 729  * that determines from the page to be migrated and the private data
 730  * the target of the move and allocates the page.
 731  *
 732  * The function returns after 10 attempts or if no pages
 733  * are movable anymore because to has become empty
 734  * or no retryable pages exist anymore. All pages will be
 735  * returned to the LRU or freed.
 736  *
 737  * Return: Number of pages not migrated or error code.
 738  */
 739 int migrate_pages(struct list_head *from,
 740                 new_page_t get_new_page, unsigned long private)
 741 {
 742         int retry = 1;
 743         int nr_failed = 0;
 744         int pass = 0;
 745         struct page *page;
 746         struct page *page2;
 747         int swapwrite = current->flags & PF_SWAPWRITE;
 748         int rc;
 749         unsigned long flags;
 750
 751         local_irq_save(flags);
 752         list_for_each_entry(page, from, lru)
 753                 __inc_zone_page_state(page, NR_ISOLATED_ANON +
 754                                 page_is_file_cache(page));
 755         local_irq_restore(flags);
 756
 757         if (!swapwrite)
 758                 current->flags |= PF_SWAPWRITE;
 759
 760         for(pass = 0; pass < 10 && retry; pass++) {
 761                 retry = 0;
 762
 763                 list_for_each_entry_safe(page, page2, from, lru) {
 764                         cond_resched();
 765
 766                         rc = unmap_and_move(get_new_page, private,
 767                                                 page, pass > 2);
 768
 769                         switch(rc) {
 770                         case -ENOMEM:
 771                                 goto out;
 772                         case -EAGAIN:
 773                                 retry++;
 774                                 break;
 775                         case 0:
 776                                 break;
 777                         default:
 778                                 /* Permanent failure */
 779                                 nr_failed++;
 780                                 break;
 781                         }
 782                 }
 783         }
 784         rc = 0;
 785 out:
 786         if (!swapwrite)
 787                 current->flags &= ~PF_SWAPWRITE;
 788
 789         putback_lru_pages(from);
 790
 791         if (rc)
 792                 return rc;
 793
 794         return nr_failed + retry;
 795 }
 796
 797 #ifdef CONFIG_NUMA
 798 /*
 799  * Move a list of individual pages
 800  */
 801 struct page_to_node {
 802         unsigned long addr;
 803         struct page *page;
 804         int node;
 805         int status;
 806 };
 807
 808 static struct page *new_page_node(struct page *p, unsigned long private,
 809                 int **result)
 810 {
 811         struct page_to_node *pm = (struct page_to_node *)private;
 812
 813         while (pm->node != MAX_NUMNODES && pm->page != p)
 814                 pm++;
 815
 816         if (pm->node == MAX_NUMNODES)
 817                 return NULL;
 818
 819         *result = &pm->status;
 820
 821         return alloc_pages_exact_node(pm->node,
 822                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 823 }
 824
 825 /*
 826  * Move a set of pages as indicated in the pm array. The addr
 827  * field must be set to the virtual address of the page to be moved
 828  * and the node number must contain a valid target node.
 829  * The pm array ends with node = MAX_NUMNODES.
 830  */
 831 static int do_move_page_to_node_array(struct mm_struct *mm,
 832                                       struct page_to_node *pm,
 833                                       int migrate_all)
 834 {
 835         int err;
 836         struct page_to_node *pp;
 837         LIST_HEAD(pagelist);
 838
 839         down_read(&mm->mmap_sem);
 840
 841         /*
 842          * Build a list of pages to migrate
 843          */
 844         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 845                 struct vm_area_struct *vma;
 846                 struct page *page;
 847
 848                 err = -EFAULT;
 849                 vma = find_vma(mm, pp->addr);
 850                 if (!vma || !vma_migratable(vma))
 851                         goto set_status;
 852
 853                 page = follow_page(vma, pp->addr, FOLL_GET);
 854
 855                 err = PTR_ERR(page);
 856                 if (IS_ERR(page))
 857                         goto set_status;
 858
 859                 err = -ENOENT;
 860                 if (!page)
 861                         goto set_status;
 862
 863                 if (PageReserved(page))         /* Check for zero page */
 864                         goto put_and_set;
 865
 866                 pp->page = page;
 867                 err = page_to_nid(page);
 868
 869                 if (err == pp->node)
 870                         /*
 871                          * Node already in the right place
 872                          */
 873                         goto put_and_set;
 874
 875                 err = -EACCES;
 876                 if (page_mapcount(page) > 1 &&
 877                                 !migrate_all)
 878                         goto put_and_set;
 879
 880                 err = isolate_lru_page(page);
 881                 if (!err)
 882                         list_add_tail(&page->lru, &pagelist);
 883 put_and_set:
 884                 /*
 885                  * Either remove the duplicate refcount from
 886                  * isolate_lru_page() or drop the page ref if it was
 887                  * not isolated.
 888                  */
 889                 put_page(page);
 890 set_status:
 891                 pp->status = err;
 892         }
 893
 894         err = 0;
 895         if (!list_empty(&pagelist))
 896                 err = migrate_pages(&pagelist, new_page_node,
 897                                 (unsigned long)pm);
 898
 899         up_read(&mm->mmap_sem);
 900         return err;
 901 }
 902
 903 /*
 904  * Migrate an array of page address onto an array of nodes and fill
 905  * the corresponding array of status.
 906  */
 907 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
 908                          unsigned long nr_pages,
 909                          const void __user * __user *pages,
 910                          const int __user *nodes,
 911                          int __user *status, int flags)
 912 {
 913         struct page_to_node *pm;
 914         nodemask_t task_nodes;
 915         unsigned long chunk_nr_pages;
 916         unsigned long chunk_start;
 917         int err;
 918
 919         task_nodes = cpuset_mems_allowed(task);
 920
 921         err = -ENOMEM;
 922         pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
 923         if (!pm)
 924                 goto out;
 925
 926         migrate_prep();
 927
 928         /*
 929          * Store a chunk of page_to_node array in a page,
 930          * but keep the last one as a marker
 931          */
 932         chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
 933
 934         for (chunk_start = 0;
 935              chunk_start < nr_pages;
 936              chunk_start += chunk_nr_pages) {
 937                 int j;
 938
 939                 if (chunk_start + chunk_nr_pages > nr_pages)
 940                         chunk_nr_pages = nr_pages - chunk_start;
 941
 942                 /* fill the chunk pm with addrs and nodes from user-space */
 943                 for (j = 0; j < chunk_nr_pages; j++) {
 944                         const void __user *p;
 945                         int node;
 946
 947                         err = -EFAULT;
 948                         if (get_user(p, pages + j + chunk_start))
 949                                 goto out_pm;
 950                         pm[j].addr = (unsigned long) p;
 951
 952                         if (get_user(node, nodes + j + chunk_start))
 953                                 goto out_pm;
 954
 955                         err = -ENODEV;
 956                         if (node < 0 || node >= MAX_NUMNODES)
 957                                 goto out_pm;
 958
 959                         if (!node_state(node, N_HIGH_MEMORY))
 960                                 goto out_pm;
 961
 962                         err = -EACCES;
 963                         if (!node_isset(node, task_nodes))
 964                                 goto out_pm;
 965
 966                         pm[j].node = node;
 967                 }
 968
 969                 /* End marker for this chunk */
 970                 pm[chunk_nr_pages].node = MAX_NUMNODES;
 971
 972                 /* Migrate this chunk */
 973                 err = do_move_page_to_node_array(mm, pm,
 974                                                  flags & MPOL_MF_MOVE_ALL);
 975                 if (err < 0)
 976                         goto out_pm;
 977
 978                 /* Return status information */
 979                 for (j = 0; j < chunk_nr_pages; j++)
 980                         if (put_user(pm[j].status, status + j + chunk_start)) {
 981                                 err = -EFAULT;
 982                                 goto out_pm;
 983                         }
 984         }
 985         err = 0;
 986
 987 out_pm:
 988         free_page((unsigned long)pm);
 989 out:
 990         return err;
 991 }
 992
 993 /*
 994  * Determine the nodes of an array of pages and store it in an array of status.
 995  */
 996 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
 997                                 const void __user **pages, int *status)
 998 {
 999         unsigned long i;
1000
1001         down_read(&mm->mmap_sem);
1002
1003         for (i = 0; i < nr_pages; i++) {
1004                 unsigned long addr = (unsigned long)(*pages);
1005                 struct vm_area_struct *vma;
1006                 struct page *page;
1007                 int err = -EFAULT;
1008
1009                 vma = find_vma(mm, addr);
1010                 if (!vma)
1011                         goto set_status;
1012
1013                 page = follow_page(vma, addr, 0);
1014
1015                 err = PTR_ERR(page);
1016                 if (IS_ERR(page))
1017                         goto set_status;
1018
1019                 err = -ENOENT;
1020                 /* Use PageReserved to check for zero page */
1021                 if (!page || PageReserved(page))
1022                         goto set_status;
1023
1024                 err = page_to_nid(page);
1025 set_status:
1026                 *status = err;
1027
1028                 pages++;
1029                 status++;
1030         }
1031
1032         up_read(&mm->mmap_sem);
1033 }
1034
1035 /*
1036  * Determine the nodes of a user array of pages and store it in
1037  * a user array of status.
1038  */
1039 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1040                          const void __user * __user *pages,
1041                          int __user *status)
1042 {
1043 #define DO_PAGES_STAT_CHUNK_NR 16
1044         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1045         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1046         unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1047         int err;
1048
1049         for (i = 0; i < nr_pages; i += chunk_nr) {
1050                 if (chunk_nr + i > nr_pages)
1051                         chunk_nr = nr_pages - i;
1052
1053                 err = copy_from_user(chunk_pages, &pages[i],
1054                                      chunk_nr * sizeof(*chunk_pages));
1055                 if (err) {
1056                         err = -EFAULT;
1057                         goto out;
1058                 }
1059
1060                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1061
1062                 err = copy_to_user(&status[i], chunk_status,
1063                                    chunk_nr * sizeof(*chunk_status));
1064                 if (err) {
1065                         err = -EFAULT;
1066                         goto out;
1067                 }
1068         }
1069         err = 0;
1070
1071 out:
1072         return err;
1073 }
1074
1075 /*
1076  * Move a list of pages in the address space of the currently executing
1077  * process.
1078  */
1079 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1080                 const void __user * __user *, pages,
1081                 const int __user *, nodes,
1082                 int __user *, status, int, flags)
1083 {
1084         const struct cred *cred = current_cred(), *tcred;
1085         struct task_struct *task;
1086         struct mm_struct *mm;
1087         int err;
1088
1089         /* Check flags */
1090         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1091                 return -EINVAL;
1092
1093         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1094                 return -EPERM;
1095
1096         /* Find the mm_struct */
1097         read_lock(&tasklist_lock);
1098         task = pid ? find_task_by_vpid(pid) : current;
1099         if (!task) {
1100                 read_unlock(&tasklist_lock);
1101                 return -ESRCH;
1102         }
1103         mm = get_task_mm(task);
1104         read_unlock(&tasklist_lock);
1105
1106         if (!mm)
1107                 return -EINVAL;
1108
1109         /*
1110          * Check if this process has the right to modify the specified
1111          * process. The right exists if the process has administrative
1112          * capabilities, superuser privileges or the same
1113          * userid as the target process.
1114          */
1115         rcu_read_lock();
1116         tcred = __task_cred(task);
1117         if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1118             cred->uid  != tcred->suid && cred->uid  != tcred->uid &&
1119             !capable(CAP_SYS_NICE)) {
1120                 rcu_read_unlock();
1121                 err = -EPERM;
1122                 goto out;
1123         }
1124         rcu_read_unlock();
1125
1126         err = security_task_movememory(task);
1127         if (err)
1128                 goto out;
1129
1130         if (nodes) {
1131                 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1132                                     flags);
1133         } else {
1134                 err = do_pages_stat(mm, nr_pages, pages, status);
1135         }
1136
1137 out:
1138         mmput(mm);
1139         return err;
1140 }
1141
1142 /*
1143  * Call migration functions in the vma_ops that may prepare
1144  * memory in a vm for migration. migration functions may perform
1145  * the migration for vmas that do not have an underlying page struct.
1146  */
1147 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1148         const nodemask_t *from, unsigned long flags)
1149 {
1150         struct vm_area_struct *vma;
1151         int err = 0;
1152
1153         for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1154                 if (vma->vm_ops && vma->vm_ops->migrate) {
1155                         err = vma->vm_ops->migrate(vma, to, from, flags);
1156                         if (err)
1157                                 break;
1158                 }
1159         }
1160         return err;
1161 }
1162 #endif