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