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
  33 #include "internal.h"
  34
  35 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  36
  37 /*
  38  * Isolate one page from the LRU lists. If successful put it onto
  39  * the indicated list with elevated page count.
  40  *
  41  * Result:
  42  *  -EBUSY: page not on LRU list
  43  *  0: page removed from LRU list and added to the specified list.
  44  */
  45 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  46 {
  47         int ret = -EBUSY;
  48
  49         if (PageLRU(page)) {
  50                 struct zone *zone = page_zone(page);
  51
  52                 spin_lock_irq(&zone->lru_lock);
  53                 if (PageLRU(page) && get_page_unless_zero(page)) {
  54                         ret = 0;
  55                         ClearPageLRU(page);
  56                         if (PageActive(page))
  57                                 del_page_from_active_list(zone, page);
  58                         else
  59                                 del_page_from_inactive_list(zone, page);
  60                         list_add_tail(&page->lru, pagelist);
  61                 }
  62                 spin_unlock_irq(&zone->lru_lock);
  63         }
  64         return ret;
  65 }
  66
  67 /*
  68  * migrate_prep() needs to be called before we start compiling a list of pages
  69  * to be migrated using isolate_lru_page().
  70  */
  71 int migrate_prep(void)
  72 {
  73         /*
  74          * Clear the LRU lists so pages can be isolated.
  75          * Note that pages may be moved off the LRU after we have
  76          * drained them. Those pages will fail to migrate like other
  77          * pages that may be busy.
  78          */
  79         lru_add_drain_all();
  80
  81         return 0;
  82 }
  83
  84 static inline void move_to_lru(struct page *page)
  85 {
  86         if (PageActive(page)) {
  87                 /*
  88                  * lru_cache_add_active checks that
  89                  * the PG_active bit is off.
  90                  */
  91                 ClearPageActive(page);
  92                 lru_cache_add_active(page);
  93         } else {
  94                 lru_cache_add(page);
  95         }
  96         put_page(page);
  97 }
  98
  99 /*
 100  * Add isolated pages on the list back to the LRU.
 101  *
 102  * returns the number of pages put back.
 103  */
 104 int putback_lru_pages(struct list_head *l)
 105 {
 106         struct page *page;
 107         struct page *page2;
 108         int count = 0;
 109
 110         list_for_each_entry_safe(page, page2, l, lru) {
 111                 list_del(&page->lru);
 112                 move_to_lru(page);
 113                 count++;
 114         }
 115         return count;
 116 }
 117
 118 static inline int is_swap_pte(pte_t pte)
 119 {
 120         return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
 121 }
 122
 123 /*
 124  * Restore a potential migration pte to a working pte entry
 125  */
 126 static void remove_migration_pte(struct vm_area_struct *vma,
 127                 struct page *old, struct page *new)
 128 {
 129         struct mm_struct *mm = vma->vm_mm;
 130         swp_entry_t entry;
 131         pgd_t *pgd;
 132         pud_t *pud;
 133         pmd_t *pmd;
 134         pte_t *ptep, pte;
 135         spinlock_t *ptl;
 136         unsigned long addr = page_address_in_vma(new, vma);
 137
 138         if (addr == -EFAULT)
 139                 return;
 140
 141         pgd = pgd_offset(mm, addr);
 142         if (!pgd_present(*pgd))
 143                 return;
 144
 145         pud = pud_offset(pgd, addr);
 146         if (!pud_present(*pud))
 147                 return;
 148
 149         pmd = pmd_offset(pud, addr);
 150         if (!pmd_present(*pmd))
 151                 return;
 152
 153         ptep = pte_offset_map(pmd, addr);
 154
 155         if (!is_swap_pte(*ptep)) {
 156                 pte_unmap(ptep);
 157                 return;
 158         }
 159
 160         ptl = pte_lockptr(mm, pmd);
 161         spin_lock(ptl);
 162         pte = *ptep;
 163         if (!is_swap_pte(pte))
 164                 goto out;
 165
 166         entry = pte_to_swp_entry(pte);
 167
 168         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 169                 goto out;
 170
 171         get_page(new);
 172         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 173         if (is_write_migration_entry(entry))
 174                 pte = pte_mkwrite(pte);
 175         flush_cache_page(vma, addr, pte_pfn(pte));
 176         set_pte_at(mm, addr, ptep, pte);
 177
 178         if (PageAnon(new))
 179                 page_add_anon_rmap(new, vma, addr);
 180         else
 181                 page_add_file_rmap(new);
 182
 183         /* No need to invalidate - it was non-present before */
 184         update_mmu_cache(vma, addr, pte);
 185
 186 out:
 187         pte_unmap_unlock(ptep, ptl);
 188 }
 189
 190 /*
 191  * Note that remove_file_migration_ptes will only work on regular mappings,
 192  * Nonlinear mappings do not use migration entries.
 193  */
 194 static void remove_file_migration_ptes(struct page *old, struct page *new)
 195 {
 196         struct vm_area_struct *vma;
 197         struct address_space *mapping = page_mapping(new);
 198         struct prio_tree_iter iter;
 199         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 200
 201         if (!mapping)
 202                 return;
 203
 204         spin_lock(&mapping->i_mmap_lock);
 205
 206         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 207                 remove_migration_pte(vma, old, new);
 208
 209         spin_unlock(&mapping->i_mmap_lock);
 210 }
 211
 212 /*
 213  * Must hold mmap_sem lock on at least one of the vmas containing
 214  * the page so that the anon_vma cannot vanish.
 215  */
 216 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 217 {
 218         struct anon_vma *anon_vma;
 219         struct vm_area_struct *vma;
 220         unsigned long mapping;
 221
 222         mapping = (unsigned long)new->mapping;
 223
 224         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 225                 return;
 226
 227         /*
 228          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 229          */
 230         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 231         spin_lock(&anon_vma->lock);
 232
 233         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 234                 remove_migration_pte(vma, old, new);
 235
 236         spin_unlock(&anon_vma->lock);
 237 }
 238
 239 /*
 240  * Get rid of all migration entries and replace them by
 241  * references to the indicated page.
 242  */
 243 static void remove_migration_ptes(struct page *old, struct page *new)
 244 {
 245         if (PageAnon(new))
 246                 remove_anon_migration_ptes(old, new);
 247         else
 248                 remove_file_migration_ptes(old, new);
 249 }
 250
 251 /*
 252  * Something used the pte of a page under migration. We need to
 253  * get to the page and wait until migration is finished.
 254  * When we return from this function the fault will be retried.
 255  *
 256  * This function is called from do_swap_page().
 257  */
 258 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 259                                 unsigned long address)
 260 {
 261         pte_t *ptep, pte;
 262         spinlock_t *ptl;
 263         swp_entry_t entry;
 264         struct page *page;
 265
 266         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 267         pte = *ptep;
 268         if (!is_swap_pte(pte))
 269                 goto out;
 270
 271         entry = pte_to_swp_entry(pte);
 272         if (!is_migration_entry(entry))
 273                 goto out;
 274
 275         page = migration_entry_to_page(entry);
 276
 277         get_page(page);
 278         pte_unmap_unlock(ptep, ptl);
 279         wait_on_page_locked(page);
 280         put_page(page);
 281         return;
 282 out:
 283         pte_unmap_unlock(ptep, ptl);
 284 }
 285
 286 /*
 287  * Replace the page in the mapping.
 288  *
 289  * The number of remaining references must be:
 290  * 1 for anonymous pages without a mapping
 291  * 2 for pages with a mapping
 292  * 3 for pages with a mapping and PagePrivate set.
 293  */
 294 static int migrate_page_move_mapping(struct address_space *mapping,
 295                 struct page *newpage, struct page *page)
 296 {
 297         void **pslot;
 298
 299         if (!mapping) {
 300                 /* Anonymous page without mapping */
 301                 if (page_count(page) != 1)
 302                         return -EAGAIN;
 303                 return 0;
 304         }
 305
 306         write_lock_irq(&mapping->tree_lock);
 307
 308         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 309                                         page_index(page));
 310
 311         if (page_count(page) != 2 + !!PagePrivate(page) ||
 312                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 313                 write_unlock_irq(&mapping->tree_lock);
 314                 return -EAGAIN;
 315         }
 316
 317         /*
 318          * Now we know that no one else is looking at the page.
 319          */
 320         get_page(newpage);      /* add cache reference */
 321 #ifdef CONFIG_SWAP
 322         if (PageSwapCache(page)) {
 323                 SetPageSwapCache(newpage);
 324                 set_page_private(newpage, page_private(page));
 325         }
 326 #endif
 327
 328         radix_tree_replace_slot(pslot, newpage);
 329
 330         /*
 331          * Drop cache reference from old page.
 332          * We know this isn't the last reference.
 333          */
 334         __put_page(page);
 335
 336         /*
 337          * If moved to a different zone then also account
 338          * the page for that zone. Other VM counters will be
 339          * taken care of when we establish references to the
 340          * new page and drop references to the old page.
 341          *
 342          * Note that anonymous pages are accounted for
 343          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 344          * are mapped to swap space.
 345          */
 346         __dec_zone_page_state(page, NR_FILE_PAGES);
 347         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 348
 349         write_unlock_irq(&mapping->tree_lock);
 350
 351         return 0;
 352 }
 353
 354 /*
 355  * Copy the page to its new location
 356  */
 357 static void migrate_page_copy(struct page *newpage, struct page *page)
 358 {
 359         copy_highpage(newpage, page);
 360
 361         if (PageError(page))
 362                 SetPageError(newpage);
 363         if (PageReferenced(page))
 364                 SetPageReferenced(newpage);
 365         if (PageUptodate(page))
 366                 SetPageUptodate(newpage);
 367         if (PageActive(page))
 368                 SetPageActive(newpage);
 369         if (PageChecked(page))
 370                 SetPageChecked(newpage);
 371         if (PageMappedToDisk(page))
 372                 SetPageMappedToDisk(newpage);
 373
 374         if (PageDirty(page)) {
 375                 clear_page_dirty_for_io(page);
 376                 set_page_dirty(newpage);
 377         }
 378
 379 #ifdef CONFIG_SWAP
 380         ClearPageSwapCache(page);
 381 #endif
 382         ClearPageActive(page);
 383         ClearPagePrivate(page);
 384         set_page_private(page, 0);
 385         page->mapping = NULL;
 386
 387         /*
 388          * If any waiters have accumulated on the new page then
 389          * wake them up.
 390          */
 391         if (PageWriteback(newpage))
 392                 end_page_writeback(newpage);
 393 }
 394
 395 /************************************************************
 396  *                    Migration functions
 397  ***********************************************************/
 398
 399 /* Always fail migration. Used for mappings that are not movable */
 400 int fail_migrate_page(struct address_space *mapping,
 401                         struct page *newpage, struct page *page)
 402 {
 403         return -EIO;
 404 }
 405 EXPORT_SYMBOL(fail_migrate_page);
 406
 407 /*
 408  * Common logic to directly migrate a single page suitable for
 409  * pages that do not use PagePrivate.
 410  *
 411  * Pages are locked upon entry and exit.
 412  */
 413 int migrate_page(struct address_space *mapping,
 414                 struct page *newpage, struct page *page)
 415 {
 416         int rc;
 417
 418         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 419
 420         rc = migrate_page_move_mapping(mapping, newpage, page);
 421
 422         if (rc)
 423                 return rc;
 424
 425         migrate_page_copy(newpage, page);
 426         return 0;
 427 }
 428 EXPORT_SYMBOL(migrate_page);
 429
 430 #ifdef CONFIG_BLOCK
 431 /*
 432  * Migration function for pages with buffers. This function can only be used
 433  * if the underlying filesystem guarantees that no other references to "page"
 434  * exist.
 435  */
 436 int buffer_migrate_page(struct address_space *mapping,
 437                 struct page *newpage, struct page *page)
 438 {
 439         struct buffer_head *bh, *head;
 440         int rc;
 441
 442         if (!page_has_buffers(page))
 443                 return migrate_page(mapping, newpage, page);
 444
 445         head = page_buffers(page);
 446
 447         rc = migrate_page_move_mapping(mapping, newpage, page);
 448
 449         if (rc)
 450                 return rc;
 451
 452         bh = head;
 453         do {
 454                 get_bh(bh);
 455                 lock_buffer(bh);
 456                 bh = bh->b_this_page;
 457
 458         } while (bh != head);
 459
 460         ClearPagePrivate(page);
 461         set_page_private(newpage, page_private(page));
 462         set_page_private(page, 0);
 463         put_page(page);
 464         get_page(newpage);
 465
 466         bh = head;
 467         do {
 468                 set_bh_page(bh, newpage, bh_offset(bh));
 469                 bh = bh->b_this_page;
 470
 471         } while (bh != head);
 472
 473         SetPagePrivate(newpage);
 474
 475         migrate_page_copy(newpage, page);
 476
 477         bh = head;
 478         do {
 479                 unlock_buffer(bh);
 480                 put_bh(bh);
 481                 bh = bh->b_this_page;
 482
 483         } while (bh != head);
 484
 485         return 0;
 486 }
 487 EXPORT_SYMBOL(buffer_migrate_page);
 488 #endif
 489
 490 /*
 491  * Writeback a page to clean the dirty state
 492  */
 493 static int writeout(struct address_space *mapping, struct page *page)
 494 {
 495         struct writeback_control wbc = {
 496                 .sync_mode = WB_SYNC_NONE,
 497                 .nr_to_write = 1,
 498                 .range_start = 0,
 499                 .range_end = LLONG_MAX,
 500                 .nonblocking = 1,
 501                 .for_reclaim = 1
 502         };
 503         int rc;
 504
 505         if (!mapping->a_ops->writepage)
 506                 /* No write method for the address space */
 507                 return -EINVAL;
 508
 509         if (!clear_page_dirty_for_io(page))
 510                 /* Someone else already triggered a write */
 511                 return -EAGAIN;
 512
 513         /*
 514          * A dirty page may imply that the underlying filesystem has
 515          * the page on some queue. So the page must be clean for
 516          * migration. Writeout may mean we loose the lock and the
 517          * page state is no longer what we checked for earlier.
 518          * At this point we know that the migration attempt cannot
 519          * be successful.
 520          */
 521         remove_migration_ptes(page, page);
 522
 523         rc = mapping->a_ops->writepage(page, &wbc);
 524         if (rc < 0)
 525                 /* I/O Error writing */
 526                 return -EIO;
 527
 528         if (rc != AOP_WRITEPAGE_ACTIVATE)
 529                 /* unlocked. Relock */
 530                 lock_page(page);
 531
 532         return -EAGAIN;
 533 }
 534
 535 /*
 536  * Default handling if a filesystem does not provide a migration function.
 537  */
 538 static int fallback_migrate_page(struct address_space *mapping,
 539         struct page *newpage, struct page *page)
 540 {
 541         if (PageDirty(page))
 542                 return writeout(mapping, page);
 543
 544         /*
 545          * Buffers may be managed in a filesystem specific way.
 546          * We must have no buffers or drop them.
 547          */
 548         if (PagePrivate(page) &&
 549             !try_to_release_page(page, GFP_KERNEL))
 550                 return -EAGAIN;
 551
 552         return migrate_page(mapping, newpage, page);
 553 }
 554
 555 /*
 556  * Move a page to a newly allocated page
 557  * The page is locked and all ptes have been successfully removed.
 558  *
 559  * The new page will have replaced the old page if this function
 560  * is successful.
 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 (TestSetPageLocked(newpage))
 573                 BUG();
 574
 575         /* Prepare mapping for the new page.*/
 576         newpage->index = page->index;
 577         newpage->mapping = page->mapping;
 578
 579         mapping = page_mapping(page);
 580         if (!mapping)
 581                 rc = migrate_page(mapping, newpage, page);
 582         else if (mapping->a_ops->migratepage)
 583                 /*
 584                  * Most pages have a mapping and most filesystems
 585                  * should provide a migration function. Anonymous
 586                  * pages are part of swap space which also has its
 587                  * own migration function. This is the most common
 588                  * path for page migration.
 589                  */
 590                 rc = mapping->a_ops->migratepage(mapping,
 591                                                 newpage, page);
 592         else
 593                 rc = fallback_migrate_page(mapping, newpage, page);
 594
 595         if (!rc)
 596                 remove_migration_ptes(page, newpage);
 597         else
 598                 newpage->mapping = NULL;
 599
 600         unlock_page(newpage);
 601
 602         return rc;
 603 }
 604
 605 /*
 606  * Obtain the lock on page, remove all ptes and migrate the page
 607  * to the newly allocated page in newpage.
 608  */
 609 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 610                         struct page *page, int force)
 611 {
 612         int rc = 0;
 613         int *result = NULL;
 614         struct page *newpage = get_new_page(page, private, &result);
 615         int rcu_locked = 0;
 616
 617         if (!newpage)
 618                 return -ENOMEM;
 619
 620         if (page_count(page) == 1)
 621                 /* page was freed from under us. So we are done. */
 622                 goto move_newpage;
 623
 624         rc = -EAGAIN;
 625         if (TestSetPageLocked(page)) {
 626                 if (!force)
 627                         goto move_newpage;
 628                 lock_page(page);
 629         }
 630
 631         if (PageWriteback(page)) {
 632                 if (!force)
 633                         goto unlock;
 634                 wait_on_page_writeback(page);
 635         }
 636         /*
 637          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 638          * we cannot notice that anon_vma is freed while we migrates a page.
 639          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 640          * of migration. File cache pages are no problem because of page_lock()
 641          * File Caches may use write_page() or lock_page() in migration, then,
 642          * just care Anon page here.
 643          */
 644         if (PageAnon(page)) {
 645                 rcu_read_lock();
 646                 rcu_locked = 1;
 647         }
 648         /*
 649          * This is a corner case handling.
 650          * When a new swap-cache is read into, it is linked to LRU
 651          * and treated as swapcache but has no rmap yet.
 652          * Calling try_to_unmap() against a page->mapping==NULL page is
 653          * BUG. So handle it here.
 654          */
 655         if (!page->mapping)
 656                 goto rcu_unlock;
 657         /* Establish migration ptes or remove ptes */
 658         try_to_unmap(page, 1);
 659
 660         if (!page_mapped(page))
 661                 rc = move_to_new_page(newpage, page);
 662
 663         if (rc)
 664                 remove_migration_ptes(page, page);
 665 rcu_unlock:
 666         if (rcu_locked)
 667                 rcu_read_unlock();
 668
 669 unlock:
 670
 671         unlock_page(page);
 672
 673         if (rc != -EAGAIN) {
 674                 /*
 675                  * A page that has been migrated has all references
 676                  * removed and will be freed. A page that has not been
 677                  * migrated will have kepts its references and be
 678                  * restored.
 679                  */
 680                 list_del(&page->lru);
 681                 move_to_lru(page);
 682         }
 683
 684 move_newpage:
 685         /*
 686          * Move the new page to the LRU. If migration was not successful
 687          * then this will free the page.
 688          */
 689         move_to_lru(newpage);
 690         if (result) {
 691                 if (rc)
 692                         *result = rc;
 693                 else
 694                         *result = page_to_nid(newpage);
 695         }
 696         return rc;
 697 }
 698
 699 /*
 700  * migrate_pages
 701  *
 702  * The function takes one list of pages to migrate and a function
 703  * that determines from the page to be migrated and the private data
 704  * the target of the move and allocates the page.
 705  *
 706  * The function returns after 10 attempts or if no pages
 707  * are movable anymore because to has become empty
 708  * or no retryable pages exist anymore. All pages will be
 709  * returned to the LRU or freed.
 710  *
 711  * Return: Number of pages not migrated or error code.
 712  */
 713 int migrate_pages(struct list_head *from,
 714                 new_page_t get_new_page, unsigned long private)
 715 {
 716         int retry = 1;
 717         int nr_failed = 0;
 718         int pass = 0;
 719         struct page *page;
 720         struct page *page2;
 721         int swapwrite = current->flags & PF_SWAPWRITE;
 722         int rc;
 723
 724         if (!swapwrite)
 725                 current->flags |= PF_SWAPWRITE;
 726
 727         for(pass = 0; pass < 10 && retry; pass++) {
 728                 retry = 0;
 729
 730                 list_for_each_entry_safe(page, page2, from, lru) {
 731                         cond_resched();
 732
 733                         rc = unmap_and_move(get_new_page, private,
 734                                                 page, pass > 2);
 735
 736                         switch(rc) {
 737                         case -ENOMEM:
 738                                 goto out;
 739                         case -EAGAIN:
 740                                 retry++;
 741                                 break;
 742                         case 0:
 743                                 break;
 744                         default:
 745                                 /* Permanent failure */
 746                                 nr_failed++;
 747                                 break;
 748                         }
 749                 }
 750         }
 751         rc = 0;
 752 out:
 753         if (!swapwrite)
 754                 current->flags &= ~PF_SWAPWRITE;
 755
 756         putback_lru_pages(from);
 757
 758         if (rc)
 759                 return rc;
 760
 761         return nr_failed + retry;
 762 }
 763
 764 #ifdef CONFIG_NUMA
 765 /*
 766  * Move a list of individual pages
 767  */
 768 struct page_to_node {
 769         unsigned long addr;
 770         struct page *page;
 771         int node;
 772         int status;
 773 };
 774
 775 static struct page *new_page_node(struct page *p, unsigned long private,
 776                 int **result)
 777 {
 778         struct page_to_node *pm = (struct page_to_node *)private;
 779
 780         while (pm->node != MAX_NUMNODES && pm->page != p)
 781                 pm++;
 782
 783         if (pm->node == MAX_NUMNODES)
 784                 return NULL;
 785
 786         *result = &pm->status;
 787
 788         return alloc_pages_node(pm->node,
 789                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 790 }
 791
 792 /*
 793  * Move a set of pages as indicated in the pm array. The addr
 794  * field must be set to the virtual address of the page to be moved
 795  * and the node number must contain a valid target node.
 796  */
 797 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 798                                 int migrate_all)
 799 {
 800         int err;
 801         struct page_to_node *pp;
 802         LIST_HEAD(pagelist);
 803
 804         down_read(&mm->mmap_sem);
 805
 806         /*
 807          * Build a list of pages to migrate
 808          */
 809         migrate_prep();
 810         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 811                 struct vm_area_struct *vma;
 812                 struct page *page;
 813
 814                 /*
 815                  * A valid page pointer that will not match any of the
 816                  * pages that will be moved.
 817                  */
 818                 pp->page = ZERO_PAGE(0);
 819
 820                 err = -EFAULT;
 821                 vma = find_vma(mm, pp->addr);
 822                 if (!vma || !vma_migratable(vma))
 823                         goto set_status;
 824
 825                 page = follow_page(vma, pp->addr, FOLL_GET);
 826                 err = -ENOENT;
 827                 if (!page)
 828                         goto set_status;
 829
 830                 if (PageReserved(page))         /* Check for zero page */
 831                         goto put_and_set;
 832
 833                 pp->page = page;
 834                 err = page_to_nid(page);
 835
 836                 if (err == pp->node)
 837                         /*
 838                          * Node already in the right place
 839                          */
 840                         goto put_and_set;
 841
 842                 err = -EACCES;
 843                 if (page_mapcount(page) > 1 &&
 844                                 !migrate_all)
 845                         goto put_and_set;
 846
 847                 err = isolate_lru_page(page, &pagelist);
 848 put_and_set:
 849                 /*
 850                  * Either remove the duplicate refcount from
 851                  * isolate_lru_page() or drop the page ref if it was
 852                  * not isolated.
 853                  */
 854                 put_page(page);
 855 set_status:
 856                 pp->status = err;
 857         }
 858
 859         if (!list_empty(&pagelist))
 860                 err = migrate_pages(&pagelist, new_page_node,
 861                                 (unsigned long)pm);
 862         else
 863                 err = -ENOENT;
 864
 865         up_read(&mm->mmap_sem);
 866         return err;
 867 }
 868
 869 /*
 870  * Determine the nodes of a list of pages. The addr in the pm array
 871  * must have been set to the virtual address of which we want to determine
 872  * the node number.
 873  */
 874 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 875 {
 876         down_read(&mm->mmap_sem);
 877
 878         for ( ; pm->node != MAX_NUMNODES; pm++) {
 879                 struct vm_area_struct *vma;
 880                 struct page *page;
 881                 int err;
 882
 883                 err = -EFAULT;
 884                 vma = find_vma(mm, pm->addr);
 885                 if (!vma)
 886                         goto set_status;
 887
 888                 page = follow_page(vma, pm->addr, 0);
 889                 err = -ENOENT;
 890                 /* Use PageReserved to check for zero page */
 891                 if (!page || PageReserved(page))
 892                         goto set_status;
 893
 894                 err = page_to_nid(page);
 895 set_status:
 896                 pm->status = err;
 897         }
 898
 899         up_read(&mm->mmap_sem);
 900         return 0;
 901 }
 902
 903 /*
 904  * Move a list of pages in the address space of the currently executing
 905  * process.
 906  */
 907 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 908                         const void __user * __user *pages,
 909                         const int __user *nodes,
 910                         int __user *status, int flags)
 911 {
 912         int err = 0;
 913         int i;
 914         struct task_struct *task;
 915         nodemask_t task_nodes;
 916         struct mm_struct *mm;
 917         struct page_to_node *pm = NULL;
 918
 919         /* Check flags */
 920         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 921                 return -EINVAL;
 922
 923         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 924                 return -EPERM;
 925
 926         /* Find the mm_struct */
 927         read_lock(&tasklist_lock);
 928         task = pid ? find_task_by_vpid(pid) : current;
 929         if (!task) {
 930                 read_unlock(&tasklist_lock);
 931                 return -ESRCH;
 932         }
 933         mm = get_task_mm(task);
 934         read_unlock(&tasklist_lock);
 935
 936         if (!mm)
 937                 return -EINVAL;
 938
 939         /*
 940          * Check if this process has the right to modify the specified
 941          * process. The right exists if the process has administrative
 942          * capabilities, superuser privileges or the same
 943          * userid as the target process.
 944          */
 945         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 946             (current->uid != task->suid) && (current->uid != task->uid) &&
 947             !capable(CAP_SYS_NICE)) {
 948                 err = -EPERM;
 949                 goto out2;
 950         }
 951
 952         err = security_task_movememory(task);
 953         if (err)
 954                 goto out2;
 955
 956
 957         task_nodes = cpuset_mems_allowed(task);
 958
 959         /* Limit nr_pages so that the multiplication may not overflow */
 960         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 961                 err = -E2BIG;
 962                 goto out2;
 963         }
 964
 965         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
 966         if (!pm) {
 967                 err = -ENOMEM;
 968                 goto out2;
 969         }
 970
 971         /*
 972          * Get parameters from user space and initialize the pm
 973          * array. Return various errors if the user did something wrong.
 974          */
 975         for (i = 0; i < nr_pages; i++) {
 976                 const void __user *p;
 977
 978                 err = -EFAULT;
 979                 if (get_user(p, pages + i))
 980                         goto out;
 981
 982                 pm[i].addr = (unsigned long)p;
 983                 if (nodes) {
 984                         int node;
 985
 986                         if (get_user(node, nodes + i))
 987                                 goto out;
 988
 989                         err = -ENODEV;
 990                         if (!node_state(node, N_HIGH_MEMORY))
 991                                 goto out;
 992
 993                         err = -EACCES;
 994                         if (!node_isset(node, task_nodes))
 995                                 goto out;
 996
 997                         pm[i].node = node;
 998                 } else
 999                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1000         }
1001         /* End marker */
1002         pm[nr_pages].node = MAX_NUMNODES;
1003
1004         if (nodes)
1005                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1006         else
1007                 err = do_pages_stat(mm, pm);
1008
1009         if (err >= 0)
1010                 /* Return status information */
1011                 for (i = 0; i < nr_pages; i++)
1012                         if (put_user(pm[i].status, status + i))
1013                                 err = -EFAULT;
1014
1015 out:
1016         vfree(pm);
1017 out2:
1018         mmput(mm);
1019         return err;
1020 }
1021 #endif
1022
1023 /*
1024  * Call migration functions in the vma_ops that may prepare
1025  * memory in a vm for migration. migration functions may perform
1026  * the migration for vmas that do not have an underlying page struct.
1027  */
1028 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1029         const nodemask_t *from, unsigned long flags)
1030 {
1031         struct vm_area_struct *vma;
1032         int err = 0;
1033
1034         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1035                 if (vma->vm_ops && vma->vm_ops->migrate) {
1036                         err = vma->vm_ops->migrate(vma, to, from, flags);
1037                         if (err)
1038                                 break;
1039                 }
1040         }
1041         return err;
1042 }