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                 putback_lru_page(page);
  71                 count++;
  72         }
  73         return count;
  74 }
  75
  76 /*
  77  * Restore a potential migration pte to a working pte entry
  78  */
  79 static void remove_migration_pte(struct vm_area_struct *vma,
  80                 struct page *old, struct page *new)
  81 {
  82         struct mm_struct *mm = vma->vm_mm;
  83         swp_entry_t entry;
  84         pgd_t *pgd;
  85         pud_t *pud;
  86         pmd_t *pmd;
  87         pte_t *ptep, pte;
  88         spinlock_t *ptl;
  89         unsigned long addr = page_address_in_vma(new, vma);
  90
  91         if (addr == -EFAULT)
  92                 return;
  93
  94         pgd = pgd_offset(mm, addr);
  95         if (!pgd_present(*pgd))
  96                 return;
  97
  98         pud = pud_offset(pgd, addr);
  99         if (!pud_present(*pud))
 100                 return;
 101
 102         pmd = pmd_offset(pud, addr);
 103         if (!pmd_present(*pmd))
 104                 return;
 105
 106         ptep = pte_offset_map(pmd, addr);
 107
 108         if (!is_swap_pte(*ptep)) {
 109                 pte_unmap(ptep);
 110                 return;
 111         }
 112
 113         ptl = pte_lockptr(mm, pmd);
 114         spin_lock(ptl);
 115         pte = *ptep;
 116         if (!is_swap_pte(pte))
 117                 goto out;
 118
 119         entry = pte_to_swp_entry(pte);
 120
 121         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 122                 goto out;
 123
 124         /*
 125          * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
 126          * Failure is not an option here: we're now expected to remove every
 127          * migration pte, and will cause crashes otherwise.  Normally this
 128          * is not an issue: mem_cgroup_prepare_migration bumped up the old
 129          * page_cgroup count for safety, that's now attached to the new page,
 130          * so this charge should just be another incrementation of the count,
 131          * to keep in balance with rmap.c's mem_cgroup_uncharging.  But if
 132          * there's been a force_empty, those reference counts may no longer
 133          * be reliable, and this charge can actually fail: oh well, we don't
 134          * make the situation any worse by proceeding as if it had succeeded.
 135          */
 136         mem_cgroup_charge(new, mm, GFP_ATOMIC);
 137
 138         get_page(new);
 139         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 140         if (is_write_migration_entry(entry))
 141                 pte = pte_mkwrite(pte);
 142         flush_cache_page(vma, addr, pte_pfn(pte));
 143         set_pte_at(mm, addr, ptep, pte);
 144
 145         if (PageAnon(new))
 146                 page_add_anon_rmap(new, vma, addr);
 147         else
 148                 page_add_file_rmap(new);
 149
 150         /* No need to invalidate - it was non-present before */
 151         update_mmu_cache(vma, addr, pte);
 152
 153 out:
 154         pte_unmap_unlock(ptep, ptl);
 155 }
 156
 157 /*
 158  * Note that remove_file_migration_ptes will only work on regular mappings,
 159  * Nonlinear mappings do not use migration entries.
 160  */
 161 static void remove_file_migration_ptes(struct page *old, struct page *new)
 162 {
 163         struct vm_area_struct *vma;
 164         struct address_space *mapping = page_mapping(new);
 165         struct prio_tree_iter iter;
 166         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 167
 168         if (!mapping)
 169                 return;
 170
 171         spin_lock(&mapping->i_mmap_lock);
 172
 173         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 174                 remove_migration_pte(vma, old, new);
 175
 176         spin_unlock(&mapping->i_mmap_lock);
 177 }
 178
 179 /*
 180  * Must hold mmap_sem lock on at least one of the vmas containing
 181  * the page so that the anon_vma cannot vanish.
 182  */
 183 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 184 {
 185         struct anon_vma *anon_vma;
 186         struct vm_area_struct *vma;
 187         unsigned long mapping;
 188
 189         mapping = (unsigned long)new->mapping;
 190
 191         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 192                 return;
 193
 194         /*
 195          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 196          */
 197         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 198         spin_lock(&anon_vma->lock);
 199
 200         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 201                 remove_migration_pte(vma, old, new);
 202
 203         spin_unlock(&anon_vma->lock);
 204 }
 205
 206 /*
 207  * Get rid of all migration entries and replace them by
 208  * references to the indicated page.
 209  */
 210 static void remove_migration_ptes(struct page *old, struct page *new)
 211 {
 212         if (PageAnon(new))
 213                 remove_anon_migration_ptes(old, new);
 214         else
 215                 remove_file_migration_ptes(old, new);
 216 }
 217
 218 /*
 219  * Something used the pte of a page under migration. We need to
 220  * get to the page and wait until migration is finished.
 221  * When we return from this function the fault will be retried.
 222  *
 223  * This function is called from do_swap_page().
 224  */
 225 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 226                                 unsigned long address)
 227 {
 228         pte_t *ptep, pte;
 229         spinlock_t *ptl;
 230         swp_entry_t entry;
 231         struct page *page;
 232
 233         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 234         pte = *ptep;
 235         if (!is_swap_pte(pte))
 236                 goto out;
 237
 238         entry = pte_to_swp_entry(pte);
 239         if (!is_migration_entry(entry))
 240                 goto out;
 241
 242         page = migration_entry_to_page(entry);
 243
 244         /*
 245          * Once radix-tree replacement of page migration started, page_count
 246          * *must* be zero. And, we don't want to call wait_on_page_locked()
 247          * against a page without get_page().
 248          * So, we use get_page_unless_zero(), here. Even failed, page fault
 249          * will occur again.
 250          */
 251         if (!get_page_unless_zero(page))
 252                 goto out;
 253         pte_unmap_unlock(ptep, ptl);
 254         wait_on_page_locked(page);
 255         put_page(page);
 256         return;
 257 out:
 258         pte_unmap_unlock(ptep, ptl);
 259 }
 260
 261 /*
 262  * Replace the page in the mapping.
 263  *
 264  * The number of remaining references must be:
 265  * 1 for anonymous pages without a mapping
 266  * 2 for pages with a mapping
 267  * 3 for pages with a mapping and PagePrivate set.
 268  */
 269 static int migrate_page_move_mapping(struct address_space *mapping,
 270                 struct page *newpage, struct page *page)
 271 {
 272         int expected_count;
 273         void **pslot;
 274
 275         if (!mapping) {
 276                 /* Anonymous page without mapping */
 277                 if (page_count(page) != 1)
 278                         return -EAGAIN;
 279                 return 0;
 280         }
 281
 282         spin_lock_irq(&mapping->tree_lock);
 283
 284         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 285                                         page_index(page));
 286
 287         expected_count = 2 + !!PagePrivate(page);
 288         if (page_count(page) != expected_count ||
 289                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 290                 spin_unlock_irq(&mapping->tree_lock);
 291                 return -EAGAIN;
 292         }
 293
 294         if (!page_freeze_refs(page, expected_count)) {
 295                 spin_unlock_irq(&mapping->tree_lock);
 296                 return -EAGAIN;
 297         }
 298
 299         /*
 300          * Now we know that no one else is looking at the page.
 301          */
 302         get_page(newpage);      /* add cache reference */
 303 #ifdef CONFIG_SWAP
 304         if (PageSwapCache(page)) {
 305                 SetPageSwapCache(newpage);
 306                 set_page_private(newpage, page_private(page));
 307         }
 308 #endif
 309
 310         radix_tree_replace_slot(pslot, newpage);
 311
 312         page_unfreeze_refs(page, expected_count);
 313         /*
 314          * Drop cache reference from old page.
 315          * We know this isn't the last reference.
 316          */
 317         __put_page(page);
 318
 319         /*
 320          * If moved to a different zone then also account
 321          * the page for that zone. Other VM counters will be
 322          * taken care of when we establish references to the
 323          * new page and drop references to the old page.
 324          *
 325          * Note that anonymous pages are accounted for
 326          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 327          * are mapped to swap space.
 328          */
 329         __dec_zone_page_state(page, NR_FILE_PAGES);
 330         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 331
 332         spin_unlock_irq(&mapping->tree_lock);
 333         if (!PageSwapCache(newpage))
 334                 mem_cgroup_uncharge_cache_page(page);
 335
 336         return 0;
 337 }
 338
 339 /*
 340  * Copy the page to its new location
 341  */
 342 static void migrate_page_copy(struct page *newpage, struct page *page)
 343 {
 344         copy_highpage(newpage, page);
 345
 346         if (PageError(page))
 347                 SetPageError(newpage);
 348         if (PageReferenced(page))
 349                 SetPageReferenced(newpage);
 350         if (PageUptodate(page))
 351                 SetPageUptodate(newpage);
 352         if (TestClearPageActive(page)) {
 353                 VM_BUG_ON(PageUnevictable(page));
 354                 SetPageActive(newpage);
 355         } else
 356                 unevictable_migrate_page(newpage, page);
 357         if (PageChecked(page))
 358                 SetPageChecked(newpage);
 359         if (PageMappedToDisk(page))
 360                 SetPageMappedToDisk(newpage);
 361
 362         if (PageDirty(page)) {
 363                 clear_page_dirty_for_io(page);
 364                 /*
 365                  * Want to mark the page and the radix tree as dirty, and
 366                  * redo the accounting that clear_page_dirty_for_io undid,
 367                  * but we can't use set_page_dirty because that function
 368                  * is actually a signal that all of the page has become dirty.
 369                  * Wheras only part of our page may be dirty.
 370                  */
 371                 __set_page_dirty_nobuffers(newpage);
 372         }
 373
 374         mlock_migrate_page(newpage, page);
 375
 376 #ifdef CONFIG_SWAP
 377         ClearPageSwapCache(page);
 378 #endif
 379         ClearPagePrivate(page);
 380         set_page_private(page, 0);
 381         page->mapping = NULL;
 382
 383         /*
 384          * If any waiters have accumulated on the new page then
 385          * wake them up.
 386          */
 387         if (PageWriteback(newpage))
 388                 end_page_writeback(newpage);
 389 }
 390
 391 /************************************************************
 392  *                    Migration functions
 393  ***********************************************************/
 394
 395 /* Always fail migration. Used for mappings that are not movable */
 396 int fail_migrate_page(struct address_space *mapping,
 397                         struct page *newpage, struct page *page)
 398 {
 399         return -EIO;
 400 }
 401 EXPORT_SYMBOL(fail_migrate_page);
 402
 403 /*
 404  * Common logic to directly migrate a single page suitable for
 405  * pages that do not use PagePrivate.
 406  *
 407  * Pages are locked upon entry and exit.
 408  */
 409 int migrate_page(struct address_space *mapping,
 410                 struct page *newpage, struct page *page)
 411 {
 412         int rc;
 413
 414         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 415
 416         rc = migrate_page_move_mapping(mapping, newpage, page);
 417
 418         if (rc)
 419                 return rc;
 420
 421         migrate_page_copy(newpage, page);
 422         return 0;
 423 }
 424 EXPORT_SYMBOL(migrate_page);
 425
 426 #ifdef CONFIG_BLOCK
 427 /*
 428  * Migration function for pages with buffers. This function can only be used
 429  * if the underlying filesystem guarantees that no other references to "page"
 430  * exist.
 431  */
 432 int buffer_migrate_page(struct address_space *mapping,
 433                 struct page *newpage, struct page *page)
 434 {
 435         struct buffer_head *bh, *head;
 436         int rc;
 437
 438         if (!page_has_buffers(page))
 439                 return migrate_page(mapping, newpage, page);
 440
 441         head = page_buffers(page);
 442
 443         rc = migrate_page_move_mapping(mapping, newpage, page);
 444
 445         if (rc)
 446                 return rc;
 447
 448         bh = head;
 449         do {
 450                 get_bh(bh);
 451                 lock_buffer(bh);
 452                 bh = bh->b_this_page;
 453
 454         } while (bh != head);
 455
 456         ClearPagePrivate(page);
 457         set_page_private(newpage, page_private(page));
 458         set_page_private(page, 0);
 459         put_page(page);
 460         get_page(newpage);
 461
 462         bh = head;
 463         do {
 464                 set_bh_page(bh, newpage, bh_offset(bh));
 465                 bh = bh->b_this_page;
 466
 467         } while (bh != head);
 468
 469         SetPagePrivate(newpage);
 470
 471         migrate_page_copy(newpage, page);
 472
 473         bh = head;
 474         do {
 475                 unlock_buffer(bh);
 476                 put_bh(bh);
 477                 bh = bh->b_this_page;
 478
 479         } while (bh != head);
 480
 481         return 0;
 482 }
 483 EXPORT_SYMBOL(buffer_migrate_page);
 484 #endif
 485
 486 /*
 487  * Writeback a page to clean the dirty state
 488  */
 489 static int writeout(struct address_space *mapping, struct page *page)
 490 {
 491         struct writeback_control wbc = {
 492                 .sync_mode = WB_SYNC_NONE,
 493                 .nr_to_write = 1,
 494                 .range_start = 0,
 495                 .range_end = LLONG_MAX,
 496                 .nonblocking = 1,
 497                 .for_reclaim = 1
 498         };
 499         int rc;
 500
 501         if (!mapping->a_ops->writepage)
 502                 /* No write method for the address space */
 503                 return -EINVAL;
 504
 505         if (!clear_page_dirty_for_io(page))
 506                 /* Someone else already triggered a write */
 507                 return -EAGAIN;
 508
 509         /*
 510          * A dirty page may imply that the underlying filesystem has
 511          * the page on some queue. So the page must be clean for
 512          * migration. Writeout may mean we loose the lock and the
 513          * page state is no longer what we checked for earlier.
 514          * At this point we know that the migration attempt cannot
 515          * be successful.
 516          */
 517         remove_migration_ptes(page, page);
 518
 519         rc = mapping->a_ops->writepage(page, &wbc);
 520         if (rc < 0)
 521                 /* I/O Error writing */
 522                 return -EIO;
 523
 524         if (rc != AOP_WRITEPAGE_ACTIVATE)
 525                 /* unlocked. Relock */
 526                 lock_page(page);
 527
 528         return -EAGAIN;
 529 }
 530
 531 /*
 532  * Default handling if a filesystem does not provide a migration function.
 533  */
 534 static int fallback_migrate_page(struct address_space *mapping,
 535         struct page *newpage, struct page *page)
 536 {
 537         if (PageDirty(page))
 538                 return writeout(mapping, page);
 539
 540         /*
 541          * Buffers may be managed in a filesystem specific way.
 542          * We must have no buffers or drop them.
 543          */
 544         if (PagePrivate(page) &&
 545             !try_to_release_page(page, GFP_KERNEL))
 546                 return -EAGAIN;
 547
 548         return migrate_page(mapping, newpage, page);
 549 }
 550
 551 /*
 552  * Move a page to a newly allocated page
 553  * The page is locked and all ptes have been successfully removed.
 554  *
 555  * The new page will have replaced the old page if this function
 556  * is successful.
 557  *
 558  * Return value:
 559  *   < 0 - error code
 560  *  == 0 - success
 561  */
 562 static int move_to_new_page(struct page *newpage, struct page *page)
 563 {
 564         struct address_space *mapping;
 565         int rc;
 566
 567         /*
 568          * Block others from accessing the page when we get around to
 569          * establishing additional references. We are the only one
 570          * holding a reference to the new page at this point.
 571          */
 572         if (!trylock_page(newpage))
 573                 BUG();
 574
 575         /* Prepare mapping for the new page.*/
 576         newpage->index = page->index;
 577         newpage->mapping = page->mapping;
 578         if (PageSwapBacked(page))
 579                 SetPageSwapBacked(newpage);
 580
 581         mapping = page_mapping(page);
 582         if (!mapping)
 583                 rc = migrate_page(mapping, newpage, page);
 584         else if (mapping->a_ops->migratepage)
 585                 /*
 586                  * Most pages have a mapping and most filesystems
 587                  * should provide a migration function. Anonymous
 588                  * pages are part of swap space which also has its
 589                  * own migration function. This is the most common
 590                  * path for page migration.
 591                  */
 592                 rc = mapping->a_ops->migratepage(mapping,
 593                                                 newpage, page);
 594         else
 595                 rc = fallback_migrate_page(mapping, newpage, page);
 596
 597         if (!rc) {
 598                 remove_migration_ptes(page, newpage);
 599         } else
 600                 newpage->mapping = NULL;
 601
 602         unlock_page(newpage);
 603
 604         return rc;
 605 }
 606
 607 /*
 608  * Obtain the lock on page, remove all ptes and migrate the page
 609  * to the newly allocated page in newpage.
 610  */
 611 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 612                         struct page *page, int force)
 613 {
 614         int rc = 0;
 615         int *result = NULL;
 616         struct page *newpage = get_new_page(page, private, &result);
 617         int rcu_locked = 0;
 618         int charge = 0;
 619
 620         if (!newpage)
 621                 return -ENOMEM;
 622
 623         if (page_count(page) == 1) {
 624                 /* page was freed from under us. So we are done. */
 625                 goto move_newpage;
 626         }
 627
 628         charge = mem_cgroup_prepare_migration(page, newpage);
 629         if (charge == -ENOMEM) {
 630                 rc = -ENOMEM;
 631                 goto move_newpage;
 632         }
 633         /* prepare cgroup just returns 0 or -ENOMEM */
 634         BUG_ON(charge);
 635
 636         rc = -EAGAIN;
 637         if (!trylock_page(page)) {
 638                 if (!force)
 639                         goto move_newpage;
 640                 lock_page(page);
 641         }
 642
 643         if (PageWriteback(page)) {
 644                 if (!force)
 645                         goto unlock;
 646                 wait_on_page_writeback(page);
 647         }
 648         /*
 649          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 650          * we cannot notice that anon_vma is freed while we migrates a page.
 651          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 652          * of migration. File cache pages are no problem because of page_lock()
 653          * File Caches may use write_page() or lock_page() in migration, then,
 654          * just care Anon page here.
 655          */
 656         if (PageAnon(page)) {
 657                 rcu_read_lock();
 658                 rcu_locked = 1;
 659         }
 660
 661         /*
 662          * Corner case handling:
 663          * 1. When a new swap-cache page is read into, it is added to the LRU
 664          * and treated as swapcache but it has no rmap yet.
 665          * Calling try_to_unmap() against a page->mapping==NULL page will
 666          * trigger a BUG.  So handle it here.
 667          * 2. An orphaned page (see truncate_complete_page) might have
 668          * fs-private metadata. The page can be picked up due to memory
 669          * offlining.  Everywhere else except page reclaim, the page is
 670          * invisible to the vm, so the page can not be migrated.  So try to
 671          * free the metadata, so the page can be freed.
 672          */
 673         if (!page->mapping) {
 674                 if (!PageAnon(page) && PagePrivate(page)) {
 675                         /*
 676                          * Go direct to try_to_free_buffers() here because
 677                          * a) that's what try_to_release_page() would do anyway
 678                          * b) we may be under rcu_read_lock() here, so we can't
 679                          *    use GFP_KERNEL which is what try_to_release_page()
 680                          *    needs to be effective.
 681                          */
 682                         try_to_free_buffers(page);
 683                 }
 684                 goto rcu_unlock;
 685         }
 686
 687         /* Establish migration ptes or remove ptes */
 688         try_to_unmap(page, 1);
 689
 690         if (!page_mapped(page))
 691                 rc = move_to_new_page(newpage, page);
 692
 693         if (rc)
 694                 remove_migration_ptes(page, page);
 695 rcu_unlock:
 696         if (rcu_locked)
 697                 rcu_read_unlock();
 698
 699 unlock:
 700         unlock_page(page);
 701
 702         if (rc != -EAGAIN) {
 703                 /*
 704                  * A page that has been migrated has all references
 705                  * removed and will be freed. A page that has not been
 706                  * migrated will have kepts its references and be
 707                  * restored.
 708                  */
 709                 list_del(&page->lru);
 710                 putback_lru_page(page);
 711         }
 712
 713 move_newpage:
 714         if (!charge)
 715                 mem_cgroup_end_migration(newpage);
 716
 717         /*
 718          * Move the new page to the LRU. If migration was not successful
 719          * then this will free the page.
 720          */
 721         putback_lru_page(newpage);
 722
 723         if (result) {
 724                 if (rc)
 725                         *result = rc;
 726                 else
 727                         *result = page_to_nid(newpage);
 728         }
 729         return rc;
 730 }
 731
 732 /*
 733  * migrate_pages
 734  *
 735  * The function takes one list of pages to migrate and a function
 736  * that determines from the page to be migrated and the private data
 737  * the target of the move and allocates the page.
 738  *
 739  * The function returns after 10 attempts or if no pages
 740  * are movable anymore because to has become empty
 741  * or no retryable pages exist anymore. All pages will be
 742  * returned to the LRU or freed.
 743  *
 744  * Return: Number of pages not migrated or error code.
 745  */
 746 int migrate_pages(struct list_head *from,
 747                 new_page_t get_new_page, unsigned long private)
 748 {
 749         int retry = 1;
 750         int nr_failed = 0;
 751         int pass = 0;
 752         struct page *page;
 753         struct page *page2;
 754         int swapwrite = current->flags & PF_SWAPWRITE;
 755         int rc;
 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_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  */
 830 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 831                                 int migrate_all)
 832 {
 833         int err;
 834         struct page_to_node *pp;
 835         LIST_HEAD(pagelist);
 836
 837         down_read(&mm->mmap_sem);
 838
 839         /*
 840          * Build a list of pages to migrate
 841          */
 842         migrate_prep();
 843         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 844                 struct vm_area_struct *vma;
 845                 struct page *page;
 846
 847                 /*
 848                  * A valid page pointer that will not match any of the
 849                  * pages that will be moved.
 850                  */
 851                 pp->page = ZERO_PAGE(0);
 852
 853                 err = -EFAULT;
 854                 vma = find_vma(mm, pp->addr);
 855                 if (!vma || !vma_migratable(vma))
 856                         goto set_status;
 857
 858                 page = follow_page(vma, pp->addr, FOLL_GET);
 859
 860                 err = PTR_ERR(page);
 861                 if (IS_ERR(page))
 862                         goto set_status;
 863
 864                 err = -ENOENT;
 865                 if (!page)
 866                         goto set_status;
 867
 868                 if (PageReserved(page))         /* Check for zero page */
 869                         goto put_and_set;
 870
 871                 pp->page = page;
 872                 err = page_to_nid(page);
 873
 874                 if (err == pp->node)
 875                         /*
 876                          * Node already in the right place
 877                          */
 878                         goto put_and_set;
 879
 880                 err = -EACCES;
 881                 if (page_mapcount(page) > 1 &&
 882                                 !migrate_all)
 883                         goto put_and_set;
 884
 885                 err = isolate_lru_page(page);
 886                 if (!err)
 887                         list_add_tail(&page->lru, &pagelist);
 888 put_and_set:
 889                 /*
 890                  * Either remove the duplicate refcount from
 891                  * isolate_lru_page() or drop the page ref if it was
 892                  * not isolated.
 893                  */
 894                 put_page(page);
 895 set_status:
 896                 pp->status = err;
 897         }
 898
 899         if (!list_empty(&pagelist))
 900                 err = migrate_pages(&pagelist, new_page_node,
 901                                 (unsigned long)pm);
 902         else
 903                 err = -ENOENT;
 904
 905         up_read(&mm->mmap_sem);
 906         return err;
 907 }
 908
 909 /*
 910  * Determine the nodes of a list of pages. The addr in the pm array
 911  * must have been set to the virtual address of which we want to determine
 912  * the node number.
 913  */
 914 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 915 {
 916         down_read(&mm->mmap_sem);
 917
 918         for ( ; pm->node != MAX_NUMNODES; pm++) {
 919                 struct vm_area_struct *vma;
 920                 struct page *page;
 921                 int err;
 922
 923                 err = -EFAULT;
 924                 vma = find_vma(mm, pm->addr);
 925                 if (!vma)
 926                         goto set_status;
 927
 928                 page = follow_page(vma, pm->addr, 0);
 929
 930                 err = PTR_ERR(page);
 931                 if (IS_ERR(page))
 932                         goto set_status;
 933
 934                 err = -ENOENT;
 935                 /* Use PageReserved to check for zero page */
 936                 if (!page || PageReserved(page))
 937                         goto set_status;
 938
 939                 err = page_to_nid(page);
 940 set_status:
 941                 pm->status = err;
 942         }
 943
 944         up_read(&mm->mmap_sem);
 945         return 0;
 946 }
 947
 948 /*
 949  * Move a list of pages in the address space of the currently executing
 950  * process.
 951  */
 952 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 953                         const void __user * __user *pages,
 954                         const int __user *nodes,
 955                         int __user *status, int flags)
 956 {
 957         int err = 0;
 958         int i;
 959         struct task_struct *task;
 960         nodemask_t task_nodes;
 961         struct mm_struct *mm;
 962         struct page_to_node *pm = NULL;
 963
 964         /* Check flags */
 965         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 966                 return -EINVAL;
 967
 968         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 969                 return -EPERM;
 970
 971         /* Find the mm_struct */
 972         read_lock(&tasklist_lock);
 973         task = pid ? find_task_by_vpid(pid) : current;
 974         if (!task) {
 975                 read_unlock(&tasklist_lock);
 976                 return -ESRCH;
 977         }
 978         mm = get_task_mm(task);
 979         read_unlock(&tasklist_lock);
 980
 981         if (!mm)
 982                 return -EINVAL;
 983
 984         /*
 985          * Check if this process has the right to modify the specified
 986          * process. The right exists if the process has administrative
 987          * capabilities, superuser privileges or the same
 988          * userid as the target process.
 989          */
 990         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 991             (current->uid != task->suid) && (current->uid != task->uid) &&
 992             !capable(CAP_SYS_NICE)) {
 993                 err = -EPERM;
 994                 goto out2;
 995         }
 996
 997         err = security_task_movememory(task);
 998         if (err)
 999                 goto out2;
1000
1001
1002         task_nodes = cpuset_mems_allowed(task);
1003
1004         /* Limit nr_pages so that the multiplication may not overflow */
1005         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1006                 err = -E2BIG;
1007                 goto out2;
1008         }
1009
1010         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1011         if (!pm) {
1012                 err = -ENOMEM;
1013                 goto out2;
1014         }
1015
1016         /*
1017          * Get parameters from user space and initialize the pm
1018          * array. Return various errors if the user did something wrong.
1019          */
1020         for (i = 0; i < nr_pages; i++) {
1021                 const void __user *p;
1022
1023                 err = -EFAULT;
1024                 if (get_user(p, pages + i))
1025                         goto out;
1026
1027                 pm[i].addr = (unsigned long)p;
1028                 if (nodes) {
1029                         int node;
1030
1031                         if (get_user(node, nodes + i))
1032                                 goto out;
1033
1034                         err = -ENODEV;
1035                         if (!node_state(node, N_HIGH_MEMORY))
1036                                 goto out;
1037
1038                         err = -EACCES;
1039                         if (!node_isset(node, task_nodes))
1040                                 goto out;
1041
1042                         pm[i].node = node;
1043                 } else
1044                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1045         }
1046         /* End marker */
1047         pm[nr_pages].node = MAX_NUMNODES;
1048
1049         if (nodes)
1050                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1051         else
1052                 err = do_pages_stat(mm, pm);
1053
1054         if (err >= 0)
1055                 /* Return status information */
1056                 for (i = 0; i < nr_pages; i++)
1057                         if (put_user(pm[i].status, status + i))
1058                                 err = -EFAULT;
1059
1060 out:
1061         vfree(pm);
1062 out2:
1063         mmput(mm);
1064         return err;
1065 }
1066
1067 /*
1068  * Call migration functions in the vma_ops that may prepare
1069  * memory in a vm for migration. migration functions may perform
1070  * the migration for vmas that do not have an underlying page struct.
1071  */
1072 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1073         const nodemask_t *from, unsigned long flags)
1074 {
1075         struct vm_area_struct *vma;
1076         int err = 0;
1077
1078         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1079                 if (vma->vm_ops && vma->vm_ops->migrate) {
1080                         err = vma->vm_ops->migrate(vma, to, from, flags);
1081                         if (err)
1082                                 break;
1083                 }
1084         }
1085         return err;
1086 }
1087 #endif