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  * Isolate one page from the LRU lists. If successful put it onto
  41  * the indicated list with elevated page count.
  42  *
  43  * Result:
  44  *  -EBUSY: page not on LRU list
  45  *  0: page removed from LRU list and added to the specified list.
  46  */
  47 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  48 {
  49         int ret = -EBUSY;
  50
  51         if (PageLRU(page)) {
  52                 struct zone *zone = page_zone(page);
  53
  54                 spin_lock_irq(&zone->lru_lock);
  55                 if (PageLRU(page) && get_page_unless_zero(page)) {
  56                         ret = 0;
  57                         ClearPageLRU(page);
  58                         if (PageActive(page))
  59                                 del_page_from_active_list(zone, page);
  60                         else
  61                                 del_page_from_inactive_list(zone, page);
  62                         list_add_tail(&page->lru, pagelist);
  63                 }
  64                 spin_unlock_irq(&zone->lru_lock);
  65         }
  66         return ret;
  67 }
  68
  69 /*
  70  * migrate_prep() needs to be called before we start compiling a list of pages
  71  * to be migrated using isolate_lru_page().
  72  */
  73 int migrate_prep(void)
  74 {
  75         /*
  76          * Clear the LRU lists so pages can be isolated.
  77          * Note that pages may be moved off the LRU after we have
  78          * drained them. Those pages will fail to migrate like other
  79          * pages that may be busy.
  80          */
  81         lru_add_drain_all();
  82
  83         return 0;
  84 }
  85
  86 static inline void move_to_lru(struct page *page)
  87 {
  88         if (PageActive(page)) {
  89                 /*
  90                  * lru_cache_add_active checks that
  91                  * the PG_active bit is off.
  92                  */
  93                 ClearPageActive(page);
  94                 lru_cache_add_active(page);
  95         } else {
  96                 lru_cache_add(page);
  97         }
  98         put_page(page);
  99 }
 100
 101 /*
 102  * Add isolated pages on the list back to the LRU.
 103  *
 104  * returns the number of pages put back.
 105  */
 106 int putback_lru_pages(struct list_head *l)
 107 {
 108         struct page *page;
 109         struct page *page2;
 110         int count = 0;
 111
 112         list_for_each_entry_safe(page, page2, l, lru) {
 113                 list_del(&page->lru);
 114                 move_to_lru(page);
 115                 count++;
 116         }
 117         return count;
 118 }
 119
 120 /*
 121  * Restore a potential migration pte to a working pte entry
 122  */
 123 static void remove_migration_pte(struct vm_area_struct *vma,
 124                 struct page *old, struct page *new)
 125 {
 126         struct mm_struct *mm = vma->vm_mm;
 127         swp_entry_t entry;
 128         pgd_t *pgd;
 129         pud_t *pud;
 130         pmd_t *pmd;
 131         pte_t *ptep, pte;
 132         spinlock_t *ptl;
 133         unsigned long addr = page_address_in_vma(new, vma);
 134
 135         if (addr == -EFAULT)
 136                 return;
 137
 138         pgd = pgd_offset(mm, addr);
 139         if (!pgd_present(*pgd))
 140                 return;
 141
 142         pud = pud_offset(pgd, addr);
 143         if (!pud_present(*pud))
 144                 return;
 145
 146         pmd = pmd_offset(pud, addr);
 147         if (!pmd_present(*pmd))
 148                 return;
 149
 150         ptep = pte_offset_map(pmd, addr);
 151
 152         if (!is_swap_pte(*ptep)) {
 153                 pte_unmap(ptep);
 154                 return;
 155         }
 156
 157         ptl = pte_lockptr(mm, pmd);
 158         spin_lock(ptl);
 159         pte = *ptep;
 160         if (!is_swap_pte(pte))
 161                 goto out;
 162
 163         entry = pte_to_swp_entry(pte);
 164
 165         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 166                 goto out;
 167
 168         /*
 169          * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
 170          * Failure is not an option here: we're now expected to remove every
 171          * migration pte, and will cause crashes otherwise.  Normally this
 172          * is not an issue: mem_cgroup_prepare_migration bumped up the old
 173          * page_cgroup count for safety, that's now attached to the new page,
 174          * so this charge should just be another incrementation of the count,
 175          * to keep in balance with rmap.c's mem_cgroup_uncharging.  But if
 176          * there's been a force_empty, those reference counts may no longer
 177          * be reliable, and this charge can actually fail: oh well, we don't
 178          * make the situation any worse by proceeding as if it had succeeded.
 179          */
 180         mem_cgroup_charge(new, mm, GFP_ATOMIC);
 181
 182         get_page(new);
 183         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 184         if (is_write_migration_entry(entry))
 185                 pte = pte_mkwrite(pte);
 186         flush_cache_page(vma, addr, pte_pfn(pte));
 187         set_pte_at(mm, addr, ptep, pte);
 188
 189         if (PageAnon(new))
 190                 page_add_anon_rmap(new, vma, addr);
 191         else
 192                 page_add_file_rmap(new);
 193
 194         /* No need to invalidate - it was non-present before */
 195         update_mmu_cache(vma, addr, pte);
 196
 197 out:
 198         pte_unmap_unlock(ptep, ptl);
 199 }
 200
 201 /*
 202  * Note that remove_file_migration_ptes will only work on regular mappings,
 203  * Nonlinear mappings do not use migration entries.
 204  */
 205 static void remove_file_migration_ptes(struct page *old, struct page *new)
 206 {
 207         struct vm_area_struct *vma;
 208         struct address_space *mapping = page_mapping(new);
 209         struct prio_tree_iter iter;
 210         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 211
 212         if (!mapping)
 213                 return;
 214
 215         spin_lock(&mapping->i_mmap_lock);
 216
 217         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 218                 remove_migration_pte(vma, old, new);
 219
 220         spin_unlock(&mapping->i_mmap_lock);
 221 }
 222
 223 /*
 224  * Must hold mmap_sem lock on at least one of the vmas containing
 225  * the page so that the anon_vma cannot vanish.
 226  */
 227 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 228 {
 229         struct anon_vma *anon_vma;
 230         struct vm_area_struct *vma;
 231         unsigned long mapping;
 232
 233         mapping = (unsigned long)new->mapping;
 234
 235         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 236                 return;
 237
 238         /*
 239          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 240          */
 241         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 242         spin_lock(&anon_vma->lock);
 243
 244         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 245                 remove_migration_pte(vma, old, new);
 246
 247         spin_unlock(&anon_vma->lock);
 248 }
 249
 250 /*
 251  * Get rid of all migration entries and replace them by
 252  * references to the indicated page.
 253  */
 254 static void remove_migration_ptes(struct page *old, struct page *new)
 255 {
 256         if (PageAnon(new))
 257                 remove_anon_migration_ptes(old, new);
 258         else
 259                 remove_file_migration_ptes(old, new);
 260 }
 261
 262 /*
 263  * Something used the pte of a page under migration. We need to
 264  * get to the page and wait until migration is finished.
 265  * When we return from this function the fault will be retried.
 266  *
 267  * This function is called from do_swap_page().
 268  */
 269 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 270                                 unsigned long address)
 271 {
 272         pte_t *ptep, pte;
 273         spinlock_t *ptl;
 274         swp_entry_t entry;
 275         struct page *page;
 276
 277         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 278         pte = *ptep;
 279         if (!is_swap_pte(pte))
 280                 goto out;
 281
 282         entry = pte_to_swp_entry(pte);
 283         if (!is_migration_entry(entry))
 284                 goto out;
 285
 286         page = migration_entry_to_page(entry);
 287
 288         /*
 289          * Once radix-tree replacement of page migration started, page_count
 290          * *must* be zero. And, we don't want to call wait_on_page_locked()
 291          * against a page without get_page().
 292          * So, we use get_page_unless_zero(), here. Even failed, page fault
 293          * will occur again.
 294          */
 295         if (!get_page_unless_zero(page))
 296                 goto out;
 297         pte_unmap_unlock(ptep, ptl);
 298         wait_on_page_locked(page);
 299         put_page(page);
 300         return;
 301 out:
 302         pte_unmap_unlock(ptep, ptl);
 303 }
 304
 305 /*
 306  * Replace the page in the mapping.
 307  *
 308  * The number of remaining references must be:
 309  * 1 for anonymous pages without a mapping
 310  * 2 for pages with a mapping
 311  * 3 for pages with a mapping and PagePrivate set.
 312  */
 313 static int migrate_page_move_mapping(struct address_space *mapping,
 314                 struct page *newpage, struct page *page)
 315 {
 316         int expected_count;
 317         void **pslot;
 318
 319         if (!mapping) {
 320                 /* Anonymous page without mapping */
 321                 if (page_count(page) != 1)
 322                         return -EAGAIN;
 323                 return 0;
 324         }
 325
 326         spin_lock_irq(&mapping->tree_lock);
 327
 328         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 329                                         page_index(page));
 330
 331         expected_count = 2 + !!PagePrivate(page);
 332         if (page_count(page) != expected_count ||
 333                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 334                 spin_unlock_irq(&mapping->tree_lock);
 335                 return -EAGAIN;
 336         }
 337
 338         if (!page_freeze_refs(page, expected_count)) {
 339                 spin_unlock_irq(&mapping->tree_lock);
 340                 return -EAGAIN;
 341         }
 342
 343         /*
 344          * Now we know that no one else is looking at the page.
 345          */
 346         get_page(newpage);      /* add cache reference */
 347 #ifdef CONFIG_SWAP
 348         if (PageSwapCache(page)) {
 349                 SetPageSwapCache(newpage);
 350                 set_page_private(newpage, page_private(page));
 351         }
 352 #endif
 353
 354         radix_tree_replace_slot(pslot, newpage);
 355
 356         page_unfreeze_refs(page, expected_count);
 357         /*
 358          * Drop cache reference from old page.
 359          * We know this isn't the last reference.
 360          */
 361         __put_page(page);
 362
 363         /*
 364          * If moved to a different zone then also account
 365          * the page for that zone. Other VM counters will be
 366          * taken care of when we establish references to the
 367          * new page and drop references to the old page.
 368          *
 369          * Note that anonymous pages are accounted for
 370          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 371          * are mapped to swap space.
 372          */
 373         __dec_zone_page_state(page, NR_FILE_PAGES);
 374         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 375
 376         spin_unlock_irq(&mapping->tree_lock);
 377         if (!PageSwapCache(newpage))
 378                 mem_cgroup_uncharge_cache_page(page);
 379
 380         return 0;
 381 }
 382
 383 /*
 384  * Copy the page to its new location
 385  */
 386 static void migrate_page_copy(struct page *newpage, struct page *page)
 387 {
 388         copy_highpage(newpage, page);
 389
 390         if (PageError(page))
 391                 SetPageError(newpage);
 392         if (PageReferenced(page))
 393                 SetPageReferenced(newpage);
 394         if (PageUptodate(page))
 395                 SetPageUptodate(newpage);
 396         if (PageActive(page))
 397                 SetPageActive(newpage);
 398         if (PageChecked(page))
 399                 SetPageChecked(newpage);
 400         if (PageMappedToDisk(page))
 401                 SetPageMappedToDisk(newpage);
 402
 403         if (PageDirty(page)) {
 404                 clear_page_dirty_for_io(page);
 405                 /*
 406                  * Want to mark the page and the radix tree as dirty, and
 407                  * redo the accounting that clear_page_dirty_for_io undid,
 408                  * but we can't use set_page_dirty because that function
 409                  * is actually a signal that all of the page has become dirty.
 410                  * Wheras only part of our page may be dirty.
 411                  */
 412                 __set_page_dirty_nobuffers(newpage);
 413         }
 414
 415 #ifdef CONFIG_SWAP
 416         ClearPageSwapCache(page);
 417 #endif
 418         ClearPageActive(page);
 419         ClearPagePrivate(page);
 420         set_page_private(page, 0);
 421         page->mapping = NULL;
 422
 423         /*
 424          * If any waiters have accumulated on the new page then
 425          * wake them up.
 426          */
 427         if (PageWriteback(newpage))
 428                 end_page_writeback(newpage);
 429 }
 430
 431 /************************************************************
 432  *                    Migration functions
 433  ***********************************************************/
 434
 435 /* Always fail migration. Used for mappings that are not movable */
 436 int fail_migrate_page(struct address_space *mapping,
 437                         struct page *newpage, struct page *page)
 438 {
 439         return -EIO;
 440 }
 441 EXPORT_SYMBOL(fail_migrate_page);
 442
 443 /*
 444  * Common logic to directly migrate a single page suitable for
 445  * pages that do not use PagePrivate.
 446  *
 447  * Pages are locked upon entry and exit.
 448  */
 449 int migrate_page(struct address_space *mapping,
 450                 struct page *newpage, struct page *page)
 451 {
 452         int rc;
 453
 454         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 455
 456         rc = migrate_page_move_mapping(mapping, newpage, page);
 457
 458         if (rc)
 459                 return rc;
 460
 461         migrate_page_copy(newpage, page);
 462         return 0;
 463 }
 464 EXPORT_SYMBOL(migrate_page);
 465
 466 #ifdef CONFIG_BLOCK
 467 /*
 468  * Migration function for pages with buffers. This function can only be used
 469  * if the underlying filesystem guarantees that no other references to "page"
 470  * exist.
 471  */
 472 int buffer_migrate_page(struct address_space *mapping,
 473                 struct page *newpage, struct page *page)
 474 {
 475         struct buffer_head *bh, *head;
 476         int rc;
 477
 478         if (!page_has_buffers(page))
 479                 return migrate_page(mapping, newpage, page);
 480
 481         head = page_buffers(page);
 482
 483         rc = migrate_page_move_mapping(mapping, newpage, page);
 484
 485         if (rc)
 486                 return rc;
 487
 488         bh = head;
 489         do {
 490                 get_bh(bh);
 491                 lock_buffer(bh);
 492                 bh = bh->b_this_page;
 493
 494         } while (bh != head);
 495
 496         ClearPagePrivate(page);
 497         set_page_private(newpage, page_private(page));
 498         set_page_private(page, 0);
 499         put_page(page);
 500         get_page(newpage);
 501
 502         bh = head;
 503         do {
 504                 set_bh_page(bh, newpage, bh_offset(bh));
 505                 bh = bh->b_this_page;
 506
 507         } while (bh != head);
 508
 509         SetPagePrivate(newpage);
 510
 511         migrate_page_copy(newpage, page);
 512
 513         bh = head;
 514         do {
 515                 unlock_buffer(bh);
 516                 put_bh(bh);
 517                 bh = bh->b_this_page;
 518
 519         } while (bh != head);
 520
 521         return 0;
 522 }
 523 EXPORT_SYMBOL(buffer_migrate_page);
 524 #endif
 525
 526 /*
 527  * Writeback a page to clean the dirty state
 528  */
 529 static int writeout(struct address_space *mapping, struct page *page)
 530 {
 531         struct writeback_control wbc = {
 532                 .sync_mode = WB_SYNC_NONE,
 533                 .nr_to_write = 1,
 534                 .range_start = 0,
 535                 .range_end = LLONG_MAX,
 536                 .nonblocking = 1,
 537                 .for_reclaim = 1
 538         };
 539         int rc;
 540
 541         if (!mapping->a_ops->writepage)
 542                 /* No write method for the address space */
 543                 return -EINVAL;
 544
 545         if (!clear_page_dirty_for_io(page))
 546                 /* Someone else already triggered a write */
 547                 return -EAGAIN;
 548
 549         /*
 550          * A dirty page may imply that the underlying filesystem has
 551          * the page on some queue. So the page must be clean for
 552          * migration. Writeout may mean we loose the lock and the
 553          * page state is no longer what we checked for earlier.
 554          * At this point we know that the migration attempt cannot
 555          * be successful.
 556          */
 557         remove_migration_ptes(page, page);
 558
 559         rc = mapping->a_ops->writepage(page, &wbc);
 560         if (rc < 0)
 561                 /* I/O Error writing */
 562                 return -EIO;
 563
 564         if (rc != AOP_WRITEPAGE_ACTIVATE)
 565                 /* unlocked. Relock */
 566                 lock_page(page);
 567
 568         return -EAGAIN;
 569 }
 570
 571 /*
 572  * Default handling if a filesystem does not provide a migration function.
 573  */
 574 static int fallback_migrate_page(struct address_space *mapping,
 575         struct page *newpage, struct page *page)
 576 {
 577         if (PageDirty(page))
 578                 return writeout(mapping, page);
 579
 580         /*
 581          * Buffers may be managed in a filesystem specific way.
 582          * We must have no buffers or drop them.
 583          */
 584         if (PagePrivate(page) &&
 585             !try_to_release_page(page, GFP_KERNEL))
 586                 return -EAGAIN;
 587
 588         return migrate_page(mapping, newpage, page);
 589 }
 590
 591 /*
 592  * Move a page to a newly allocated page
 593  * The page is locked and all ptes have been successfully removed.
 594  *
 595  * The new page will have replaced the old page if this function
 596  * is successful.
 597  */
 598 static int move_to_new_page(struct page *newpage, struct page *page)
 599 {
 600         struct address_space *mapping;
 601         int rc;
 602
 603         /*
 604          * Block others from accessing the page when we get around to
 605          * establishing additional references. We are the only one
 606          * holding a reference to the new page at this point.
 607          */
 608         if (!trylock_page(newpage))
 609                 BUG();
 610
 611         /* Prepare mapping for the new page.*/
 612         newpage->index = page->index;
 613         newpage->mapping = page->mapping;
 614
 615         mapping = page_mapping(page);
 616         if (!mapping)
 617                 rc = migrate_page(mapping, newpage, page);
 618         else if (mapping->a_ops->migratepage)
 619                 /*
 620                  * Most pages have a mapping and most filesystems
 621                  * should provide a migration function. Anonymous
 622                  * pages are part of swap space which also has its
 623                  * own migration function. This is the most common
 624                  * path for page migration.
 625                  */
 626                 rc = mapping->a_ops->migratepage(mapping,
 627                                                 newpage, page);
 628         else
 629                 rc = fallback_migrate_page(mapping, newpage, page);
 630
 631         if (!rc) {
 632                 remove_migration_ptes(page, newpage);
 633         } else
 634                 newpage->mapping = NULL;
 635
 636         unlock_page(newpage);
 637
 638         return rc;
 639 }
 640
 641 /*
 642  * Obtain the lock on page, remove all ptes and migrate the page
 643  * to the newly allocated page in newpage.
 644  */
 645 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 646                         struct page *page, int force)
 647 {
 648         int rc = 0;
 649         int *result = NULL;
 650         struct page *newpage = get_new_page(page, private, &result);
 651         int rcu_locked = 0;
 652         int charge = 0;
 653
 654         if (!newpage)
 655                 return -ENOMEM;
 656
 657         if (page_count(page) == 1)
 658                 /* page was freed from under us. So we are done. */
 659                 goto move_newpage;
 660
 661         charge = mem_cgroup_prepare_migration(page, newpage);
 662         if (charge == -ENOMEM) {
 663                 rc = -ENOMEM;
 664                 goto move_newpage;
 665         }
 666         /* prepare cgroup just returns 0 or -ENOMEM */
 667         BUG_ON(charge);
 668
 669         rc = -EAGAIN;
 670         if (!trylock_page(page)) {
 671                 if (!force)
 672                         goto move_newpage;
 673                 lock_page(page);
 674         }
 675
 676         if (PageWriteback(page)) {
 677                 if (!force)
 678                         goto unlock;
 679                 wait_on_page_writeback(page);
 680         }
 681         /*
 682          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 683          * we cannot notice that anon_vma is freed while we migrates a page.
 684          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 685          * of migration. File cache pages are no problem because of page_lock()
 686          * File Caches may use write_page() or lock_page() in migration, then,
 687          * just care Anon page here.
 688          */
 689         if (PageAnon(page)) {
 690                 rcu_read_lock();
 691                 rcu_locked = 1;
 692         }
 693
 694         /*
 695          * Corner case handling:
 696          * 1. When a new swap-cache page is read into, it is added to the LRU
 697          * and treated as swapcache but it has no rmap yet.
 698          * Calling try_to_unmap() against a page->mapping==NULL page will
 699          * trigger a BUG.  So handle it here.
 700          * 2. An orphaned page (see truncate_complete_page) might have
 701          * fs-private metadata. The page can be picked up due to memory
 702          * offlining.  Everywhere else except page reclaim, the page is
 703          * invisible to the vm, so the page can not be migrated.  So try to
 704          * free the metadata, so the page can be freed.
 705          */
 706         if (!page->mapping) {
 707                 if (!PageAnon(page) && PagePrivate(page)) {
 708                         /*
 709                          * Go direct to try_to_free_buffers() here because
 710                          * a) that's what try_to_release_page() would do anyway
 711                          * b) we may be under rcu_read_lock() here, so we can't
 712                          *    use GFP_KERNEL which is what try_to_release_page()
 713                          *    needs to be effective.
 714                          */
 715                         try_to_free_buffers(page);
 716                 }
 717                 goto rcu_unlock;
 718         }
 719
 720         /* Establish migration ptes or remove ptes */
 721         try_to_unmap(page, 1);
 722
 723         if (!page_mapped(page))
 724                 rc = move_to_new_page(newpage, page);
 725
 726         if (rc)
 727                 remove_migration_ptes(page, page);
 728 rcu_unlock:
 729         if (rcu_locked)
 730                 rcu_read_unlock();
 731
 732 unlock:
 733
 734         unlock_page(page);
 735
 736         if (rc != -EAGAIN) {
 737                 /*
 738                  * A page that has been migrated has all references
 739                  * removed and will be freed. A page that has not been
 740                  * migrated will have kepts its references and be
 741                  * restored.
 742                  */
 743                 list_del(&page->lru);
 744                 move_to_lru(page);
 745         }
 746
 747 move_newpage:
 748         if (!charge)
 749                 mem_cgroup_end_migration(newpage);
 750         /*
 751          * Move the new page to the LRU. If migration was not successful
 752          * then this will free the page.
 753          */
 754         move_to_lru(newpage);
 755         if (result) {
 756                 if (rc)
 757                         *result = rc;
 758                 else
 759                         *result = page_to_nid(newpage);
 760         }
 761         return rc;
 762 }
 763
 764 /*
 765  * migrate_pages
 766  *
 767  * The function takes one list of pages to migrate and a function
 768  * that determines from the page to be migrated and the private data
 769  * the target of the move and allocates the page.
 770  *
 771  * The function returns after 10 attempts or if no pages
 772  * are movable anymore because to has become empty
 773  * or no retryable pages exist anymore. All pages will be
 774  * returned to the LRU or freed.
 775  *
 776  * Return: Number of pages not migrated or error code.
 777  */
 778 int migrate_pages(struct list_head *from,
 779                 new_page_t get_new_page, unsigned long private)
 780 {
 781         int retry = 1;
 782         int nr_failed = 0;
 783         int pass = 0;
 784         struct page *page;
 785         struct page *page2;
 786         int swapwrite = current->flags & PF_SWAPWRITE;
 787         int rc;
 788
 789         if (!swapwrite)
 790                 current->flags |= PF_SWAPWRITE;
 791
 792         for(pass = 0; pass < 10 && retry; pass++) {
 793                 retry = 0;
 794
 795                 list_for_each_entry_safe(page, page2, from, lru) {
 796                         cond_resched();
 797
 798                         rc = unmap_and_move(get_new_page, private,
 799                                                 page, pass > 2);
 800
 801                         switch(rc) {
 802                         case -ENOMEM:
 803                                 goto out;
 804                         case -EAGAIN:
 805                                 retry++;
 806                                 break;
 807                         case 0:
 808                                 break;
 809                         default:
 810                                 /* Permanent failure */
 811                                 nr_failed++;
 812                                 break;
 813                         }
 814                 }
 815         }
 816         rc = 0;
 817 out:
 818         if (!swapwrite)
 819                 current->flags &= ~PF_SWAPWRITE;
 820
 821         putback_lru_pages(from);
 822
 823         if (rc)
 824                 return rc;
 825
 826         return nr_failed + retry;
 827 }
 828
 829 #ifdef CONFIG_NUMA
 830 /*
 831  * Move a list of individual pages
 832  */
 833 struct page_to_node {
 834         unsigned long addr;
 835         struct page *page;
 836         int node;
 837         int status;
 838 };
 839
 840 static struct page *new_page_node(struct page *p, unsigned long private,
 841                 int **result)
 842 {
 843         struct page_to_node *pm = (struct page_to_node *)private;
 844
 845         while (pm->node != MAX_NUMNODES && pm->page != p)
 846                 pm++;
 847
 848         if (pm->node == MAX_NUMNODES)
 849                 return NULL;
 850
 851         *result = &pm->status;
 852
 853         return alloc_pages_node(pm->node,
 854                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 855 }
 856
 857 /*
 858  * Move a set of pages as indicated in the pm array. The addr
 859  * field must be set to the virtual address of the page to be moved
 860  * and the node number must contain a valid target node.
 861  */
 862 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 863                                 int migrate_all)
 864 {
 865         int err;
 866         struct page_to_node *pp;
 867         LIST_HEAD(pagelist);
 868
 869         down_read(&mm->mmap_sem);
 870
 871         /*
 872          * Build a list of pages to migrate
 873          */
 874         migrate_prep();
 875         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 876                 struct vm_area_struct *vma;
 877                 struct page *page;
 878
 879                 /*
 880                  * A valid page pointer that will not match any of the
 881                  * pages that will be moved.
 882                  */
 883                 pp->page = ZERO_PAGE(0);
 884
 885                 err = -EFAULT;
 886                 vma = find_vma(mm, pp->addr);
 887                 if (!vma || !vma_migratable(vma))
 888                         goto set_status;
 889
 890                 page = follow_page(vma, pp->addr, FOLL_GET);
 891
 892                 err = PTR_ERR(page);
 893                 if (IS_ERR(page))
 894                         goto set_status;
 895
 896                 err = -ENOENT;
 897                 if (!page)
 898                         goto set_status;
 899
 900                 if (PageReserved(page))         /* Check for zero page */
 901                         goto put_and_set;
 902
 903                 pp->page = page;
 904                 err = page_to_nid(page);
 905
 906                 if (err == pp->node)
 907                         /*
 908                          * Node already in the right place
 909                          */
 910                         goto put_and_set;
 911
 912                 err = -EACCES;
 913                 if (page_mapcount(page) > 1 &&
 914                                 !migrate_all)
 915                         goto put_and_set;
 916
 917                 err = isolate_lru_page(page, &pagelist);
 918 put_and_set:
 919                 /*
 920                  * Either remove the duplicate refcount from
 921                  * isolate_lru_page() or drop the page ref if it was
 922                  * not isolated.
 923                  */
 924                 put_page(page);
 925 set_status:
 926                 pp->status = err;
 927         }
 928
 929         if (!list_empty(&pagelist))
 930                 err = migrate_pages(&pagelist, new_page_node,
 931                                 (unsigned long)pm);
 932         else
 933                 err = -ENOENT;
 934
 935         up_read(&mm->mmap_sem);
 936         return err;
 937 }
 938
 939 /*
 940  * Determine the nodes of a list of pages. The addr in the pm array
 941  * must have been set to the virtual address of which we want to determine
 942  * the node number.
 943  */
 944 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 945 {
 946         down_read(&mm->mmap_sem);
 947
 948         for ( ; pm->node != MAX_NUMNODES; pm++) {
 949                 struct vm_area_struct *vma;
 950                 struct page *page;
 951                 int err;
 952
 953                 err = -EFAULT;
 954                 vma = find_vma(mm, pm->addr);
 955                 if (!vma)
 956                         goto set_status;
 957
 958                 page = follow_page(vma, pm->addr, 0);
 959
 960                 err = PTR_ERR(page);
 961                 if (IS_ERR(page))
 962                         goto set_status;
 963
 964                 err = -ENOENT;
 965                 /* Use PageReserved to check for zero page */
 966                 if (!page || PageReserved(page))
 967                         goto set_status;
 968
 969                 err = page_to_nid(page);
 970 set_status:
 971                 pm->status = err;
 972         }
 973
 974         up_read(&mm->mmap_sem);
 975         return 0;
 976 }
 977
 978 /*
 979  * Move a list of pages in the address space of the currently executing
 980  * process.
 981  */
 982 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 983                         const void __user * __user *pages,
 984                         const int __user *nodes,
 985                         int __user *status, int flags)
 986 {
 987         int err = 0;
 988         int i;
 989         struct task_struct *task;
 990         nodemask_t task_nodes;
 991         struct mm_struct *mm;
 992         struct page_to_node *pm = NULL;
 993
 994         /* Check flags */
 995         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 996                 return -EINVAL;
 997
 998         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 999                 return -EPERM;
1000
1001         /* Find the mm_struct */
1002         read_lock(&tasklist_lock);
1003         task = pid ? find_task_by_vpid(pid) : current;
1004         if (!task) {
1005                 read_unlock(&tasklist_lock);
1006                 return -ESRCH;
1007         }
1008         mm = get_task_mm(task);
1009         read_unlock(&tasklist_lock);
1010
1011         if (!mm)
1012                 return -EINVAL;
1013
1014         /*
1015          * Check if this process has the right to modify the specified
1016          * process. The right exists if the process has administrative
1017          * capabilities, superuser privileges or the same
1018          * userid as the target process.
1019          */
1020         if ((current->euid != task->suid) && (current->euid != task->uid) &&
1021             (current->uid != task->suid) && (current->uid != task->uid) &&
1022             !capable(CAP_SYS_NICE)) {
1023                 err = -EPERM;
1024                 goto out2;
1025         }
1026
1027         err = security_task_movememory(task);
1028         if (err)
1029                 goto out2;
1030
1031
1032         task_nodes = cpuset_mems_allowed(task);
1033
1034         /* Limit nr_pages so that the multiplication may not overflow */
1035         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1036                 err = -E2BIG;
1037                 goto out2;
1038         }
1039
1040         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1041         if (!pm) {
1042                 err = -ENOMEM;
1043                 goto out2;
1044         }
1045
1046         /*
1047          * Get parameters from user space and initialize the pm
1048          * array. Return various errors if the user did something wrong.
1049          */
1050         for (i = 0; i < nr_pages; i++) {
1051                 const void __user *p;
1052
1053                 err = -EFAULT;
1054                 if (get_user(p, pages + i))
1055                         goto out;
1056
1057                 pm[i].addr = (unsigned long)p;
1058                 if (nodes) {
1059                         int node;
1060
1061                         if (get_user(node, nodes + i))
1062                                 goto out;
1063
1064                         err = -ENODEV;
1065                         if (!node_state(node, N_HIGH_MEMORY))
1066                                 goto out;
1067
1068                         err = -EACCES;
1069                         if (!node_isset(node, task_nodes))
1070                                 goto out;
1071
1072                         pm[i].node = node;
1073                 } else
1074                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1075         }
1076         /* End marker */
1077         pm[nr_pages].node = MAX_NUMNODES;
1078
1079         if (nodes)
1080                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1081         else
1082                 err = do_pages_stat(mm, pm);
1083
1084         if (err >= 0)
1085                 /* Return status information */
1086                 for (i = 0; i < nr_pages; i++)
1087                         if (put_user(pm[i].status, status + i))
1088                                 err = -EFAULT;
1089
1090 out:
1091         vfree(pm);
1092 out2:
1093         mmput(mm);
1094         return err;
1095 }
1096
1097 /*
1098  * Call migration functions in the vma_ops that may prepare
1099  * memory in a vm for migration. migration functions may perform
1100  * the migration for vmas that do not have an underlying page struct.
1101  */
1102 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1103         const nodemask_t *from, unsigned long flags)
1104 {
1105         struct vm_area_struct *vma;
1106         int err = 0;
1107
1108         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1109                 if (vma->vm_ops && vma->vm_ops->migrate) {
1110                         err = vma->vm_ops->migrate(vma, to, from, flags);
1111                         if (err)
1112                                 break;
1113                 }
1114         }
1115         return err;
1116 }
1117 #endif