mm/migrate.c

   1 /*
   2  * Memory Migration functionality - linux/mm/migration.c
   3  *
   4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5  *
   6  * Page migration was first developed in the context of the memory hotplug
   7  * project. The main authors of the migration code are:
   8  *
   9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10  * Hirokazu Takahashi <taka@valinux.co.jp>
  11  * Dave Hansen <haveblue@us.ibm.com>
  12  * Christoph Lameter <clameter@sgi.com>
  13  */
  14
  15 #include <linux/migrate.h>
  16 #include <linux/module.h>
  17 #include <linux/swap.h>
  18 #include <linux/swapops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/nsproxy.h>
  23 #include <linux/pagevec.h>
  24 #include <linux/rmap.h>
  25 #include <linux/topology.h>
  26 #include <linux/cpu.h>
  27 #include <linux/cpuset.h>
  28 #include <linux/writeback.h>
  29 #include <linux/mempolicy.h>
  30 #include <linux/vmalloc.h>
  31 #include <linux/security.h>
  32 #include <linux/memcontrol.h>
  33
  34 #include "internal.h"
  35
  36 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  37
  38 /*
  39  * Isolate one page from the LRU lists. If successful put it onto
  40  * the indicated list with elevated page count.
  41  *
  42  * Result:
  43  *  -EBUSY: page not on LRU list
  44  *  0: page removed from LRU list and added to the specified list.
  45  */
  46 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  47 {
  48         int ret = -EBUSY;
  49
  50         if (PageLRU(page)) {
  51                 struct zone *zone = page_zone(page);
  52
  53                 spin_lock_irq(&zone->lru_lock);
  54                 if (PageLRU(page) && get_page_unless_zero(page)) {
  55                         ret = 0;
  56                         ClearPageLRU(page);
  57                         if (PageActive(page))
  58                                 del_page_from_active_list(zone, page);
  59                         else
  60                                 del_page_from_inactive_list(zone, page);
  61                         list_add_tail(&page->lru, pagelist);
  62                 }
  63                 spin_unlock_irq(&zone->lru_lock);
  64         }
  65         return ret;
  66 }
  67
  68 /*
  69  * migrate_prep() needs to be called before we start compiling a list of pages
  70  * to be migrated using isolate_lru_page().
  71  */
  72 int migrate_prep(void)
  73 {
  74         /*
  75          * Clear the LRU lists so pages can be isolated.
  76          * Note that pages may be moved off the LRU after we have
  77          * drained them. Those pages will fail to migrate like other
  78          * pages that may be busy.
  79          */
  80         lru_add_drain_all();
  81
  82         return 0;
  83 }
  84
  85 static inline void move_to_lru(struct page *page)
  86 {
  87         if (PageActive(page)) {
  88                 /*
  89                  * lru_cache_add_active checks that
  90                  * the PG_active bit is off.
  91                  */
  92                 ClearPageActive(page);
  93                 lru_cache_add_active(page);
  94         } else {
  95                 lru_cache_add(page);
  96         }
  97         put_page(page);
  98 }
  99
 100 /*
 101  * Add isolated pages on the list back to the LRU.
 102  *
 103  * returns the number of pages put back.
 104  */
 105 int putback_lru_pages(struct list_head *l)
 106 {
 107         struct page *page;
 108         struct page *page2;
 109         int count = 0;
 110
 111         list_for_each_entry_safe(page, page2, l, lru) {
 112                 list_del(&page->lru);
 113                 move_to_lru(page);
 114                 count++;
 115         }
 116         return count;
 117 }
 118
 119 /*
 120  * Restore a potential migration pte to a working pte entry
 121  */
 122 static void remove_migration_pte(struct vm_area_struct *vma,
 123                 struct page *old, struct page *new)
 124 {
 125         struct mm_struct *mm = vma->vm_mm;
 126         swp_entry_t entry;
 127         pgd_t *pgd;
 128         pud_t *pud;
 129         pmd_t *pmd;
 130         pte_t *ptep, pte;
 131         spinlock_t *ptl;
 132         unsigned long addr = page_address_in_vma(new, vma);
 133
 134         if (addr == -EFAULT)
 135                 return;
 136
 137         pgd = pgd_offset(mm, addr);
 138         if (!pgd_present(*pgd))
 139                 return;
 140
 141         pud = pud_offset(pgd, addr);
 142         if (!pud_present(*pud))
 143                 return;
 144
 145         pmd = pmd_offset(pud, addr);
 146         if (!pmd_present(*pmd))
 147                 return;
 148
 149         ptep = pte_offset_map(pmd, addr);
 150
 151         if (!is_swap_pte(*ptep)) {
 152                 pte_unmap(ptep);
 153                 return;
 154         }
 155
 156         ptl = pte_lockptr(mm, pmd);
 157         spin_lock(ptl);
 158         pte = *ptep;
 159         if (!is_swap_pte(pte))
 160                 goto out;
 161
 162         entry = pte_to_swp_entry(pte);
 163
 164         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 165                 goto out;
 166
 167         /*
 168          * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
 169          * Failure is not an option here: we're now expected to remove every
 170          * migration pte, and will cause crashes otherwise.  Normally this
 171          * is not an issue: mem_cgroup_prepare_migration bumped up the old
 172          * page_cgroup count for safety, that's now attached to the new page,
 173          * so this charge should just be another incrementation of the count,
 174          * to keep in balance with rmap.c's mem_cgroup_uncharging.  But if
 175          * there's been a force_empty, those reference counts may no longer
 176          * be reliable, and this charge can actually fail: oh well, we don't
 177          * make the situation any worse by proceeding as if it had succeeded.
 178          */
 179         mem_cgroup_charge(new, mm, GFP_ATOMIC);
 180
 181         get_page(new);
 182         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 183         if (is_write_migration_entry(entry))
 184                 pte = pte_mkwrite(pte);
 185         flush_cache_page(vma, addr, pte_pfn(pte));
 186         set_pte_at(mm, addr, ptep, pte);
 187
 188         if (PageAnon(new))
 189                 page_add_anon_rmap(new, vma, addr);
 190         else
 191                 page_add_file_rmap(new);
 192
 193         /* No need to invalidate - it was non-present before */
 194         update_mmu_cache(vma, addr, pte);
 195
 196 out:
 197         pte_unmap_unlock(ptep, ptl);
 198 }
 199
 200 /*
 201  * Note that remove_file_migration_ptes will only work on regular mappings,
 202  * Nonlinear mappings do not use migration entries.
 203  */
 204 static void remove_file_migration_ptes(struct page *old, struct page *new)
 205 {
 206         struct vm_area_struct *vma;
 207         struct address_space *mapping = page_mapping(new);
 208         struct prio_tree_iter iter;
 209         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 210
 211         if (!mapping)
 212                 return;
 213
 214         spin_lock(&mapping->i_mmap_lock);
 215
 216         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 217                 remove_migration_pte(vma, old, new);
 218
 219         spin_unlock(&mapping->i_mmap_lock);
 220 }
 221
 222 /*
 223  * Must hold mmap_sem lock on at least one of the vmas containing
 224  * the page so that the anon_vma cannot vanish.
 225  */
 226 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 227 {
 228         struct anon_vma *anon_vma;
 229         struct vm_area_struct *vma;
 230         unsigned long mapping;
 231
 232         mapping = (unsigned long)new->mapping;
 233
 234         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 235                 return;
 236
 237         /*
 238          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 239          */
 240         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 241         spin_lock(&anon_vma->lock);
 242
 243         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 244                 remove_migration_pte(vma, old, new);
 245
 246         spin_unlock(&anon_vma->lock);
 247 }
 248
 249 /*
 250  * Get rid of all migration entries and replace them by
 251  * references to the indicated page.
 252  */
 253 static void remove_migration_ptes(struct page *old, struct page *new)
 254 {
 255         if (PageAnon(new))
 256                 remove_anon_migration_ptes(old, new);
 257         else
 258                 remove_file_migration_ptes(old, new);
 259 }
 260
 261 /*
 262  * Something used the pte of a page under migration. We need to
 263  * get to the page and wait until migration is finished.
 264  * When we return from this function the fault will be retried.
 265  *
 266  * This function is called from do_swap_page().
 267  */
 268 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 269                                 unsigned long address)
 270 {
 271         pte_t *ptep, pte;
 272         spinlock_t *ptl;
 273         swp_entry_t entry;
 274         struct page *page;
 275
 276         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 277         pte = *ptep;
 278         if (!is_swap_pte(pte))
 279                 goto out;
 280
 281         entry = pte_to_swp_entry(pte);
 282         if (!is_migration_entry(entry))
 283                 goto out;
 284
 285         page = migration_entry_to_page(entry);
 286
 287         get_page(page);
 288         pte_unmap_unlock(ptep, ptl);
 289         wait_on_page_locked(page);
 290         put_page(page);
 291         return;
 292 out:
 293         pte_unmap_unlock(ptep, ptl);
 294 }
 295
 296 /*
 297  * Replace the page in the mapping.
 298  *
 299  * The number of remaining references must be:
 300  * 1 for anonymous pages without a mapping
 301  * 2 for pages with a mapping
 302  * 3 for pages with a mapping and PagePrivate set.
 303  */
 304 static int migrate_page_move_mapping(struct address_space *mapping,
 305                 struct page *newpage, struct page *page)
 306 {
 307         void **pslot;
 308
 309         if (!mapping) {
 310                 /* Anonymous page without mapping */
 311                 if (page_count(page) != 1)
 312                         return -EAGAIN;
 313                 return 0;
 314         }
 315
 316         write_lock_irq(&mapping->tree_lock);
 317
 318         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 319                                         page_index(page));
 320
 321         if (page_count(page) != 2 + !!PagePrivate(page) ||
 322                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 323                 write_unlock_irq(&mapping->tree_lock);
 324                 return -EAGAIN;
 325         }
 326
 327         /*
 328          * Now we know that no one else is looking at the page.
 329          */
 330         get_page(newpage);      /* add cache reference */
 331 #ifdef CONFIG_SWAP
 332         if (PageSwapCache(page)) {
 333                 SetPageSwapCache(newpage);
 334                 set_page_private(newpage, page_private(page));
 335         }
 336 #endif
 337
 338         radix_tree_replace_slot(pslot, newpage);
 339
 340         /*
 341          * Drop cache reference from old page.
 342          * We know this isn't the last reference.
 343          */
 344         __put_page(page);
 345
 346         /*
 347          * If moved to a different zone then also account
 348          * the page for that zone. Other VM counters will be
 349          * taken care of when we establish references to the
 350          * new page and drop references to the old page.
 351          *
 352          * Note that anonymous pages are accounted for
 353          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 354          * are mapped to swap space.
 355          */
 356         __dec_zone_page_state(page, NR_FILE_PAGES);
 357         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 358
 359         write_unlock_irq(&mapping->tree_lock);
 360
 361         return 0;
 362 }
 363
 364 /*
 365  * Copy the page to its new location
 366  */
 367 static void migrate_page_copy(struct page *newpage, struct page *page)
 368 {
 369         copy_highpage(newpage, page);
 370
 371         if (PageError(page))
 372                 SetPageError(newpage);
 373         if (PageReferenced(page))
 374                 SetPageReferenced(newpage);
 375         if (PageUptodate(page))
 376                 SetPageUptodate(newpage);
 377         if (PageActive(page))
 378                 SetPageActive(newpage);
 379         if (PageChecked(page))
 380                 SetPageChecked(newpage);
 381         if (PageMappedToDisk(page))
 382                 SetPageMappedToDisk(newpage);
 383
 384         if (PageDirty(page)) {
 385                 clear_page_dirty_for_io(page);
 386                 set_page_dirty(newpage);
 387         }
 388
 389 #ifdef CONFIG_SWAP
 390         ClearPageSwapCache(page);
 391 #endif
 392         ClearPageActive(page);
 393         ClearPagePrivate(page);
 394         set_page_private(page, 0);
 395         page->mapping = NULL;
 396
 397         /*
 398          * If any waiters have accumulated on the new page then
 399          * wake them up.
 400          */
 401         if (PageWriteback(newpage))
 402                 end_page_writeback(newpage);
 403 }
 404
 405 /************************************************************
 406  *                    Migration functions
 407  ***********************************************************/
 408
 409 /* Always fail migration. Used for mappings that are not movable */
 410 int fail_migrate_page(struct address_space *mapping,
 411                         struct page *newpage, struct page *page)
 412 {
 413         return -EIO;
 414 }
 415 EXPORT_SYMBOL(fail_migrate_page);
 416
 417 /*
 418  * Common logic to directly migrate a single page suitable for
 419  * pages that do not use PagePrivate.
 420  *
 421  * Pages are locked upon entry and exit.
 422  */
 423 int migrate_page(struct address_space *mapping,
 424                 struct page *newpage, struct page *page)
 425 {
 426         int rc;
 427
 428         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 429
 430         rc = migrate_page_move_mapping(mapping, newpage, page);
 431
 432         if (rc)
 433                 return rc;
 434
 435         migrate_page_copy(newpage, page);
 436         return 0;
 437 }
 438 EXPORT_SYMBOL(migrate_page);
 439
 440 #ifdef CONFIG_BLOCK
 441 /*
 442  * Migration function for pages with buffers. This function can only be used
 443  * if the underlying filesystem guarantees that no other references to "page"
 444  * exist.
 445  */
 446 int buffer_migrate_page(struct address_space *mapping,
 447                 struct page *newpage, struct page *page)
 448 {
 449         struct buffer_head *bh, *head;
 450         int rc;
 451
 452         if (!page_has_buffers(page))
 453                 return migrate_page(mapping, newpage, page);
 454
 455         head = page_buffers(page);
 456
 457         rc = migrate_page_move_mapping(mapping, newpage, page);
 458
 459         if (rc)
 460                 return rc;
 461
 462         bh = head;
 463         do {
 464                 get_bh(bh);
 465                 lock_buffer(bh);
 466                 bh = bh->b_this_page;
 467
 468         } while (bh != head);
 469
 470         ClearPagePrivate(page);
 471         set_page_private(newpage, page_private(page));
 472         set_page_private(page, 0);
 473         put_page(page);
 474         get_page(newpage);
 475
 476         bh = head;
 477         do {
 478                 set_bh_page(bh, newpage, bh_offset(bh));
 479                 bh = bh->b_this_page;
 480
 481         } while (bh != head);
 482
 483         SetPagePrivate(newpage);
 484
 485         migrate_page_copy(newpage, page);
 486
 487         bh = head;
 488         do {
 489                 unlock_buffer(bh);
 490                 put_bh(bh);
 491                 bh = bh->b_this_page;
 492
 493         } while (bh != head);
 494
 495         return 0;
 496 }
 497 EXPORT_SYMBOL(buffer_migrate_page);
 498 #endif
 499
 500 /*
 501  * Writeback a page to clean the dirty state
 502  */
 503 static int writeout(struct address_space *mapping, struct page *page)
 504 {
 505         struct writeback_control wbc = {
 506                 .sync_mode = WB_SYNC_NONE,
 507                 .nr_to_write = 1,
 508                 .range_start = 0,
 509                 .range_end = LLONG_MAX,
 510                 .nonblocking = 1,
 511                 .for_reclaim = 1
 512         };
 513         int rc;
 514
 515         if (!mapping->a_ops->writepage)
 516                 /* No write method for the address space */
 517                 return -EINVAL;
 518
 519         if (!clear_page_dirty_for_io(page))
 520                 /* Someone else already triggered a write */
 521                 return -EAGAIN;
 522
 523         /*
 524          * A dirty page may imply that the underlying filesystem has
 525          * the page on some queue. So the page must be clean for
 526          * migration. Writeout may mean we loose the lock and the
 527          * page state is no longer what we checked for earlier.
 528          * At this point we know that the migration attempt cannot
 529          * be successful.
 530          */
 531         remove_migration_ptes(page, page);
 532
 533         rc = mapping->a_ops->writepage(page, &wbc);
 534         if (rc < 0)
 535                 /* I/O Error writing */
 536                 return -EIO;
 537
 538         if (rc != AOP_WRITEPAGE_ACTIVATE)
 539                 /* unlocked. Relock */
 540                 lock_page(page);
 541
 542         return -EAGAIN;
 543 }
 544
 545 /*
 546  * Default handling if a filesystem does not provide a migration function.
 547  */
 548 static int fallback_migrate_page(struct address_space *mapping,
 549         struct page *newpage, struct page *page)
 550 {
 551         if (PageDirty(page))
 552                 return writeout(mapping, page);
 553
 554         /*
 555          * Buffers may be managed in a filesystem specific way.
 556          * We must have no buffers or drop them.
 557          */
 558         if (PagePrivate(page) &&
 559             !try_to_release_page(page, GFP_KERNEL))
 560                 return -EAGAIN;
 561
 562         return migrate_page(mapping, newpage, page);
 563 }
 564
 565 /*
 566  * Move a page to a newly allocated page
 567  * The page is locked and all ptes have been successfully removed.
 568  *
 569  * The new page will have replaced the old page if this function
 570  * is successful.
 571  */
 572 static int move_to_new_page(struct page *newpage, struct page *page)
 573 {
 574         struct address_space *mapping;
 575         int rc;
 576
 577         /*
 578          * Block others from accessing the page when we get around to
 579          * establishing additional references. We are the only one
 580          * holding a reference to the new page at this point.
 581          */
 582         if (TestSetPageLocked(newpage))
 583                 BUG();
 584
 585         /* Prepare mapping for the new page.*/
 586         newpage->index = page->index;
 587         newpage->mapping = page->mapping;
 588
 589         mapping = page_mapping(page);
 590         if (!mapping)
 591                 rc = migrate_page(mapping, newpage, page);
 592         else if (mapping->a_ops->migratepage)
 593                 /*
 594                  * Most pages have a mapping and most filesystems
 595                  * should provide a migration function. Anonymous
 596                  * pages are part of swap space which also has its
 597                  * own migration function. This is the most common
 598                  * path for page migration.
 599                  */
 600                 rc = mapping->a_ops->migratepage(mapping,
 601                                                 newpage, page);
 602         else
 603                 rc = fallback_migrate_page(mapping, newpage, page);
 604
 605         if (!rc) {
 606                 mem_cgroup_page_migration(page, newpage);
 607                 remove_migration_ptes(page, newpage);
 608         } else
 609                 newpage->mapping = NULL;
 610
 611         unlock_page(newpage);
 612
 613         return rc;
 614 }
 615
 616 /*
 617  * Obtain the lock on page, remove all ptes and migrate the page
 618  * to the newly allocated page in newpage.
 619  */
 620 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 621                         struct page *page, int force)
 622 {
 623         int rc = 0;
 624         int *result = NULL;
 625         struct page *newpage = get_new_page(page, private, &result);
 626         int rcu_locked = 0;
 627         int charge = 0;
 628
 629         if (!newpage)
 630                 return -ENOMEM;
 631
 632         if (page_count(page) == 1)
 633                 /* page was freed from under us. So we are done. */
 634                 goto move_newpage;
 635
 636         rc = -EAGAIN;
 637         if (TestSetPageLocked(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         charge = mem_cgroup_prepare_migration(page);
 688         /* Establish migration ptes or remove ptes */
 689         try_to_unmap(page, 1);
 690
 691         if (!page_mapped(page))
 692                 rc = move_to_new_page(newpage, page);
 693
 694         if (rc) {
 695                 remove_migration_ptes(page, page);
 696                 if (charge)
 697                         mem_cgroup_end_migration(page);
 698         } else if (charge)
 699                 mem_cgroup_end_migration(newpage);
 700 rcu_unlock:
 701         if (rcu_locked)
 702                 rcu_read_unlock();
 703
 704 unlock:
 705
 706         unlock_page(page);
 707
 708         if (rc != -EAGAIN) {
 709                 /*
 710                  * A page that has been migrated has all references
 711                  * removed and will be freed. A page that has not been
 712                  * migrated will have kepts its references and be
 713                  * restored.
 714                  */
 715                 list_del(&page->lru);
 716                 move_to_lru(page);
 717         }
 718
 719 move_newpage:
 720         /*
 721          * Move the new page to the LRU. If migration was not successful
 722          * then this will free the page.
 723          */
 724         move_to_lru(newpage);
 725         if (result) {
 726                 if (rc)
 727                         *result = rc;
 728                 else
 729                         *result = page_to_nid(newpage);
 730         }
 731         return rc;
 732 }
 733
 734 /*
 735  * migrate_pages
 736  *
 737  * The function takes one list of pages to migrate and a function
 738  * that determines from the page to be migrated and the private data
 739  * the target of the move and allocates the page.
 740  *
 741  * The function returns after 10 attempts or if no pages
 742  * are movable anymore because to has become empty
 743  * or no retryable pages exist anymore. All pages will be
 744  * returned to the LRU or freed.
 745  *
 746  * Return: Number of pages not migrated or error code.
 747  */
 748 int migrate_pages(struct list_head *from,
 749                 new_page_t get_new_page, unsigned long private)
 750 {
 751         int retry = 1;
 752         int nr_failed = 0;
 753         int pass = 0;
 754         struct page *page;
 755         struct page *page2;
 756         int swapwrite = current->flags & PF_SWAPWRITE;
 757         int rc;
 758
 759         if (!swapwrite)
 760                 current->flags |= PF_SWAPWRITE;
 761
 762         for(pass = 0; pass < 10 && retry; pass++) {
 763                 retry = 0;
 764
 765                 list_for_each_entry_safe(page, page2, from, lru) {
 766                         cond_resched();
 767
 768                         rc = unmap_and_move(get_new_page, private,
 769                                                 page, pass > 2);
 770
 771                         switch(rc) {
 772                         case -ENOMEM:
 773                                 goto out;
 774                         case -EAGAIN:
 775                                 retry++;
 776                                 break;
 777                         case 0:
 778                                 break;
 779                         default:
 780                                 /* Permanent failure */
 781                                 nr_failed++;
 782                                 break;
 783                         }
 784                 }
 785         }
 786         rc = 0;
 787 out:
 788         if (!swapwrite)
 789                 current->flags &= ~PF_SWAPWRITE;
 790
 791         putback_lru_pages(from);
 792
 793         if (rc)
 794                 return rc;
 795
 796         return nr_failed + retry;
 797 }
 798
 799 #ifdef CONFIG_NUMA
 800 /*
 801  * Move a list of individual pages
 802  */
 803 struct page_to_node {
 804         unsigned long addr;
 805         struct page *page;
 806         int node;
 807         int status;
 808 };
 809
 810 static struct page *new_page_node(struct page *p, unsigned long private,
 811                 int **result)
 812 {
 813         struct page_to_node *pm = (struct page_to_node *)private;
 814
 815         while (pm->node != MAX_NUMNODES && pm->page != p)
 816                 pm++;
 817
 818         if (pm->node == MAX_NUMNODES)
 819                 return NULL;
 820
 821         *result = &pm->status;
 822
 823         return alloc_pages_node(pm->node,
 824                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 825 }
 826
 827 /*
 828  * Move a set of pages as indicated in the pm array. The addr
 829  * field must be set to the virtual address of the page to be moved
 830  * and the node number must contain a valid target node.
 831  */
 832 static int do_move_pages(struct mm_struct *mm, 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         migrate_prep();
 845         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 846                 struct vm_area_struct *vma;
 847                 struct page *page;
 848
 849                 /*
 850                  * A valid page pointer that will not match any of the
 851                  * pages that will be moved.
 852                  */
 853                 pp->page = ZERO_PAGE(0);
 854
 855                 err = -EFAULT;
 856                 vma = find_vma(mm, pp->addr);
 857                 if (!vma || !vma_migratable(vma))
 858                         goto set_status;
 859
 860                 page = follow_page(vma, pp->addr, FOLL_GET);
 861                 err = -ENOENT;
 862                 if (!page)
 863                         goto set_status;
 864
 865                 if (PageReserved(page))         /* Check for zero page */
 866                         goto put_and_set;
 867
 868                 pp->page = page;
 869                 err = page_to_nid(page);
 870
 871                 if (err == pp->node)
 872                         /*
 873                          * Node already in the right place
 874                          */
 875                         goto put_and_set;
 876
 877                 err = -EACCES;
 878                 if (page_mapcount(page) > 1 &&
 879                                 !migrate_all)
 880                         goto put_and_set;
 881
 882                 err = isolate_lru_page(page, &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         if (!list_empty(&pagelist))
 895                 err = migrate_pages(&pagelist, new_page_node,
 896                                 (unsigned long)pm);
 897         else
 898                 err = -ENOENT;
 899
 900         up_read(&mm->mmap_sem);
 901         return err;
 902 }
 903
 904 /*
 905  * Determine the nodes of a list of pages. The addr in the pm array
 906  * must have been set to the virtual address of which we want to determine
 907  * the node number.
 908  */
 909 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 910 {
 911         down_read(&mm->mmap_sem);
 912
 913         for ( ; pm->node != MAX_NUMNODES; pm++) {
 914                 struct vm_area_struct *vma;
 915                 struct page *page;
 916                 int err;
 917
 918                 err = -EFAULT;
 919                 vma = find_vma(mm, pm->addr);
 920                 if (!vma)
 921                         goto set_status;
 922
 923                 page = follow_page(vma, pm->addr, 0);
 924                 err = -ENOENT;
 925                 /* Use PageReserved to check for zero page */
 926                 if (!page || PageReserved(page))
 927                         goto set_status;
 928
 929                 err = page_to_nid(page);
 930 set_status:
 931                 pm->status = err;
 932         }
 933
 934         up_read(&mm->mmap_sem);
 935         return 0;
 936 }
 937
 938 /*
 939  * Move a list of pages in the address space of the currently executing
 940  * process.
 941  */
 942 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 943                         const void __user * __user *pages,
 944                         const int __user *nodes,
 945                         int __user *status, int flags)
 946 {
 947         int err = 0;
 948         int i;
 949         struct task_struct *task;
 950         nodemask_t task_nodes;
 951         struct mm_struct *mm;
 952         struct page_to_node *pm = NULL;
 953
 954         /* Check flags */
 955         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 956                 return -EINVAL;
 957
 958         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 959                 return -EPERM;
 960
 961         /* Find the mm_struct */
 962         read_lock(&tasklist_lock);
 963         task = pid ? find_task_by_vpid(pid) : current;
 964         if (!task) {
 965                 read_unlock(&tasklist_lock);
 966                 return -ESRCH;
 967         }
 968         mm = get_task_mm(task);
 969         read_unlock(&tasklist_lock);
 970
 971         if (!mm)
 972                 return -EINVAL;
 973
 974         /*
 975          * Check if this process has the right to modify the specified
 976          * process. The right exists if the process has administrative
 977          * capabilities, superuser privileges or the same
 978          * userid as the target process.
 979          */
 980         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 981             (current->uid != task->suid) && (current->uid != task->uid) &&
 982             !capable(CAP_SYS_NICE)) {
 983                 err = -EPERM;
 984                 goto out2;
 985         }
 986
 987         err = security_task_movememory(task);
 988         if (err)
 989                 goto out2;
 990
 991
 992         task_nodes = cpuset_mems_allowed(task);
 993
 994         /* Limit nr_pages so that the multiplication may not overflow */
 995         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 996                 err = -E2BIG;
 997                 goto out2;
 998         }
 999
1000         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1001         if (!pm) {
1002                 err = -ENOMEM;
1003                 goto out2;
1004         }
1005
1006         /*
1007          * Get parameters from user space and initialize the pm
1008          * array. Return various errors if the user did something wrong.
1009          */
1010         for (i = 0; i < nr_pages; i++) {
1011                 const void __user *p;
1012
1013                 err = -EFAULT;
1014                 if (get_user(p, pages + i))
1015                         goto out;
1016
1017                 pm[i].addr = (unsigned long)p;
1018                 if (nodes) {
1019                         int node;
1020
1021                         if (get_user(node, nodes + i))
1022                                 goto out;
1023
1024                         err = -ENODEV;
1025                         if (!node_state(node, N_HIGH_MEMORY))
1026                                 goto out;
1027
1028                         err = -EACCES;
1029                         if (!node_isset(node, task_nodes))
1030                                 goto out;
1031
1032                         pm[i].node = node;
1033                 } else
1034                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1035         }
1036         /* End marker */
1037         pm[nr_pages].node = MAX_NUMNODES;
1038
1039         if (nodes)
1040                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1041         else
1042                 err = do_pages_stat(mm, pm);
1043
1044         if (err >= 0)
1045                 /* Return status information */
1046                 for (i = 0; i < nr_pages; i++)
1047                         if (put_user(pm[i].status, status + i))
1048                                 err = -EFAULT;
1049
1050 out:
1051         vfree(pm);
1052 out2:
1053         mmput(mm);
1054         return err;
1055 }
1056 #endif
1057
1058 /*
1059  * Call migration functions in the vma_ops that may prepare
1060  * memory in a vm for migration. migration functions may perform
1061  * the migration for vmas that do not have an underlying page struct.
1062  */
1063 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1064         const nodemask_t *from, unsigned long flags)
1065 {
1066         struct vm_area_struct *vma;
1067         int err = 0;
1068
1069         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1070                 if (vma->vm_ops && vma->vm_ops->migrate) {
1071                         err = vma->vm_ops->migrate(vma, to, from, flags);
1072                         if (err)
1073                                 break;
1074                 }
1075         }
1076         return err;
1077 }