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                 /*
 387                  * Want to mark the page and the radix tree as dirty, and
 388                  * redo the accounting that clear_page_dirty_for_io undid,
 389                  * but we can't use set_page_dirty because that function
 390                  * is actually a signal that all of the page has become dirty.
 391                  * Wheras only part of our page may be dirty.
 392                  */
 393                 __set_page_dirty_nobuffers(newpage);
 394         }
 395
 396 #ifdef CONFIG_SWAP
 397         ClearPageSwapCache(page);
 398 #endif
 399         ClearPageActive(page);
 400         ClearPagePrivate(page);
 401         set_page_private(page, 0);
 402         page->mapping = NULL;
 403
 404         /*
 405          * If any waiters have accumulated on the new page then
 406          * wake them up.
 407          */
 408         if (PageWriteback(newpage))
 409                 end_page_writeback(newpage);
 410 }
 411
 412 /************************************************************
 413  *                    Migration functions
 414  ***********************************************************/
 415
 416 /* Always fail migration. Used for mappings that are not movable */
 417 int fail_migrate_page(struct address_space *mapping,
 418                         struct page *newpage, struct page *page)
 419 {
 420         return -EIO;
 421 }
 422 EXPORT_SYMBOL(fail_migrate_page);
 423
 424 /*
 425  * Common logic to directly migrate a single page suitable for
 426  * pages that do not use PagePrivate.
 427  *
 428  * Pages are locked upon entry and exit.
 429  */
 430 int migrate_page(struct address_space *mapping,
 431                 struct page *newpage, struct page *page)
 432 {
 433         int rc;
 434
 435         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 436
 437         rc = migrate_page_move_mapping(mapping, newpage, page);
 438
 439         if (rc)
 440                 return rc;
 441
 442         migrate_page_copy(newpage, page);
 443         return 0;
 444 }
 445 EXPORT_SYMBOL(migrate_page);
 446
 447 #ifdef CONFIG_BLOCK
 448 /*
 449  * Migration function for pages with buffers. This function can only be used
 450  * if the underlying filesystem guarantees that no other references to "page"
 451  * exist.
 452  */
 453 int buffer_migrate_page(struct address_space *mapping,
 454                 struct page *newpage, struct page *page)
 455 {
 456         struct buffer_head *bh, *head;
 457         int rc;
 458
 459         if (!page_has_buffers(page))
 460                 return migrate_page(mapping, newpage, page);
 461
 462         head = page_buffers(page);
 463
 464         rc = migrate_page_move_mapping(mapping, newpage, page);
 465
 466         if (rc)
 467                 return rc;
 468
 469         bh = head;
 470         do {
 471                 get_bh(bh);
 472                 lock_buffer(bh);
 473                 bh = bh->b_this_page;
 474
 475         } while (bh != head);
 476
 477         ClearPagePrivate(page);
 478         set_page_private(newpage, page_private(page));
 479         set_page_private(page, 0);
 480         put_page(page);
 481         get_page(newpage);
 482
 483         bh = head;
 484         do {
 485                 set_bh_page(bh, newpage, bh_offset(bh));
 486                 bh = bh->b_this_page;
 487
 488         } while (bh != head);
 489
 490         SetPagePrivate(newpage);
 491
 492         migrate_page_copy(newpage, page);
 493
 494         bh = head;
 495         do {
 496                 unlock_buffer(bh);
 497                 put_bh(bh);
 498                 bh = bh->b_this_page;
 499
 500         } while (bh != head);
 501
 502         return 0;
 503 }
 504 EXPORT_SYMBOL(buffer_migrate_page);
 505 #endif
 506
 507 /*
 508  * Writeback a page to clean the dirty state
 509  */
 510 static int writeout(struct address_space *mapping, struct page *page)
 511 {
 512         struct writeback_control wbc = {
 513                 .sync_mode = WB_SYNC_NONE,
 514                 .nr_to_write = 1,
 515                 .range_start = 0,
 516                 .range_end = LLONG_MAX,
 517                 .nonblocking = 1,
 518                 .for_reclaim = 1
 519         };
 520         int rc;
 521
 522         if (!mapping->a_ops->writepage)
 523                 /* No write method for the address space */
 524                 return -EINVAL;
 525
 526         if (!clear_page_dirty_for_io(page))
 527                 /* Someone else already triggered a write */
 528                 return -EAGAIN;
 529
 530         /*
 531          * A dirty page may imply that the underlying filesystem has
 532          * the page on some queue. So the page must be clean for
 533          * migration. Writeout may mean we loose the lock and the
 534          * page state is no longer what we checked for earlier.
 535          * At this point we know that the migration attempt cannot
 536          * be successful.
 537          */
 538         remove_migration_ptes(page, page);
 539
 540         rc = mapping->a_ops->writepage(page, &wbc);
 541         if (rc < 0)
 542                 /* I/O Error writing */
 543                 return -EIO;
 544
 545         if (rc != AOP_WRITEPAGE_ACTIVATE)
 546                 /* unlocked. Relock */
 547                 lock_page(page);
 548
 549         return -EAGAIN;
 550 }
 551
 552 /*
 553  * Default handling if a filesystem does not provide a migration function.
 554  */
 555 static int fallback_migrate_page(struct address_space *mapping,
 556         struct page *newpage, struct page *page)
 557 {
 558         if (PageDirty(page))
 559                 return writeout(mapping, page);
 560
 561         /*
 562          * Buffers may be managed in a filesystem specific way.
 563          * We must have no buffers or drop them.
 564          */
 565         if (PagePrivate(page) &&
 566             !try_to_release_page(page, GFP_KERNEL))
 567                 return -EAGAIN;
 568
 569         return migrate_page(mapping, newpage, page);
 570 }
 571
 572 /*
 573  * Move a page to a newly allocated page
 574  * The page is locked and all ptes have been successfully removed.
 575  *
 576  * The new page will have replaced the old page if this function
 577  * is successful.
 578  */
 579 static int move_to_new_page(struct page *newpage, struct page *page)
 580 {
 581         struct address_space *mapping;
 582         int rc;
 583
 584         /*
 585          * Block others from accessing the page when we get around to
 586          * establishing additional references. We are the only one
 587          * holding a reference to the new page at this point.
 588          */
 589         if (TestSetPageLocked(newpage))
 590                 BUG();
 591
 592         /* Prepare mapping for the new page.*/
 593         newpage->index = page->index;
 594         newpage->mapping = page->mapping;
 595
 596         mapping = page_mapping(page);
 597         if (!mapping)
 598                 rc = migrate_page(mapping, newpage, page);
 599         else if (mapping->a_ops->migratepage)
 600                 /*
 601                  * Most pages have a mapping and most filesystems
 602                  * should provide a migration function. Anonymous
 603                  * pages are part of swap space which also has its
 604                  * own migration function. This is the most common
 605                  * path for page migration.
 606                  */
 607                 rc = mapping->a_ops->migratepage(mapping,
 608                                                 newpage, page);
 609         else
 610                 rc = fallback_migrate_page(mapping, newpage, page);
 611
 612         if (!rc) {
 613                 mem_cgroup_page_migration(page, newpage);
 614                 remove_migration_ptes(page, newpage);
 615         } else
 616                 newpage->mapping = NULL;
 617
 618         unlock_page(newpage);
 619
 620         return rc;
 621 }
 622
 623 /*
 624  * Obtain the lock on page, remove all ptes and migrate the page
 625  * to the newly allocated page in newpage.
 626  */
 627 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 628                         struct page *page, int force)
 629 {
 630         int rc = 0;
 631         int *result = NULL;
 632         struct page *newpage = get_new_page(page, private, &result);
 633         int rcu_locked = 0;
 634         int charge = 0;
 635
 636         if (!newpage)
 637                 return -ENOMEM;
 638
 639         if (page_count(page) == 1)
 640                 /* page was freed from under us. So we are done. */
 641                 goto move_newpage;
 642
 643         rc = -EAGAIN;
 644         if (TestSetPageLocked(page)) {
 645                 if (!force)
 646                         goto move_newpage;
 647                 lock_page(page);
 648         }
 649
 650         if (PageWriteback(page)) {
 651                 if (!force)
 652                         goto unlock;
 653                 wait_on_page_writeback(page);
 654         }
 655         /*
 656          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 657          * we cannot notice that anon_vma is freed while we migrates a page.
 658          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 659          * of migration. File cache pages are no problem because of page_lock()
 660          * File Caches may use write_page() or lock_page() in migration, then,
 661          * just care Anon page here.
 662          */
 663         if (PageAnon(page)) {
 664                 rcu_read_lock();
 665                 rcu_locked = 1;
 666         }
 667
 668         /*
 669          * Corner case handling:
 670          * 1. When a new swap-cache page is read into, it is added to the LRU
 671          * and treated as swapcache but it has no rmap yet.
 672          * Calling try_to_unmap() against a page->mapping==NULL page will
 673          * trigger a BUG.  So handle it here.
 674          * 2. An orphaned page (see truncate_complete_page) might have
 675          * fs-private metadata. The page can be picked up due to memory
 676          * offlining.  Everywhere else except page reclaim, the page is
 677          * invisible to the vm, so the page can not be migrated.  So try to
 678          * free the metadata, so the page can be freed.
 679          */
 680         if (!page->mapping) {
 681                 if (!PageAnon(page) && PagePrivate(page)) {
 682                         /*
 683                          * Go direct to try_to_free_buffers() here because
 684                          * a) that's what try_to_release_page() would do anyway
 685                          * b) we may be under rcu_read_lock() here, so we can't
 686                          *    use GFP_KERNEL which is what try_to_release_page()
 687                          *    needs to be effective.
 688                          */
 689                         try_to_free_buffers(page);
 690                 }
 691                 goto rcu_unlock;
 692         }
 693
 694         charge = mem_cgroup_prepare_migration(page);
 695         /* Establish migration ptes or remove ptes */
 696         try_to_unmap(page, 1);
 697
 698         if (!page_mapped(page))
 699                 rc = move_to_new_page(newpage, page);
 700
 701         if (rc) {
 702                 remove_migration_ptes(page, page);
 703                 if (charge)
 704                         mem_cgroup_end_migration(page);
 705         } else if (charge)
 706                 mem_cgroup_end_migration(newpage);
 707 rcu_unlock:
 708         if (rcu_locked)
 709                 rcu_read_unlock();
 710
 711 unlock:
 712
 713         unlock_page(page);
 714
 715         if (rc != -EAGAIN) {
 716                 /*
 717                  * A page that has been migrated has all references
 718                  * removed and will be freed. A page that has not been
 719                  * migrated will have kepts its references and be
 720                  * restored.
 721                  */
 722                 list_del(&page->lru);
 723                 move_to_lru(page);
 724         }
 725
 726 move_newpage:
 727         /*
 728          * Move the new page to the LRU. If migration was not successful
 729          * then this will free the page.
 730          */
 731         move_to_lru(newpage);
 732         if (result) {
 733                 if (rc)
 734                         *result = rc;
 735                 else
 736                         *result = page_to_nid(newpage);
 737         }
 738         return rc;
 739 }
 740
 741 /*
 742  * migrate_pages
 743  *
 744  * The function takes one list of pages to migrate and a function
 745  * that determines from the page to be migrated and the private data
 746  * the target of the move and allocates the page.
 747  *
 748  * The function returns after 10 attempts or if no pages
 749  * are movable anymore because to has become empty
 750  * or no retryable pages exist anymore. All pages will be
 751  * returned to the LRU or freed.
 752  *
 753  * Return: Number of pages not migrated or error code.
 754  */
 755 int migrate_pages(struct list_head *from,
 756                 new_page_t get_new_page, unsigned long private)
 757 {
 758         int retry = 1;
 759         int nr_failed = 0;
 760         int pass = 0;
 761         struct page *page;
 762         struct page *page2;
 763         int swapwrite = current->flags & PF_SWAPWRITE;
 764         int rc;
 765
 766         if (!swapwrite)
 767                 current->flags |= PF_SWAPWRITE;
 768
 769         for(pass = 0; pass < 10 && retry; pass++) {
 770                 retry = 0;
 771
 772                 list_for_each_entry_safe(page, page2, from, lru) {
 773                         cond_resched();
 774
 775                         rc = unmap_and_move(get_new_page, private,
 776                                                 page, pass > 2);
 777
 778                         switch(rc) {
 779                         case -ENOMEM:
 780                                 goto out;
 781                         case -EAGAIN:
 782                                 retry++;
 783                                 break;
 784                         case 0:
 785                                 break;
 786                         default:
 787                                 /* Permanent failure */
 788                                 nr_failed++;
 789                                 break;
 790                         }
 791                 }
 792         }
 793         rc = 0;
 794 out:
 795         if (!swapwrite)
 796                 current->flags &= ~PF_SWAPWRITE;
 797
 798         putback_lru_pages(from);
 799
 800         if (rc)
 801                 return rc;
 802
 803         return nr_failed + retry;
 804 }
 805
 806 #ifdef CONFIG_NUMA
 807 /*
 808  * Move a list of individual pages
 809  */
 810 struct page_to_node {
 811         unsigned long addr;
 812         struct page *page;
 813         int node;
 814         int status;
 815 };
 816
 817 static struct page *new_page_node(struct page *p, unsigned long private,
 818                 int **result)
 819 {
 820         struct page_to_node *pm = (struct page_to_node *)private;
 821
 822         while (pm->node != MAX_NUMNODES && pm->page != p)
 823                 pm++;
 824
 825         if (pm->node == MAX_NUMNODES)
 826                 return NULL;
 827
 828         *result = &pm->status;
 829
 830         return alloc_pages_node(pm->node,
 831                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 832 }
 833
 834 /*
 835  * Move a set of pages as indicated in the pm array. The addr
 836  * field must be set to the virtual address of the page to be moved
 837  * and the node number must contain a valid target node.
 838  */
 839 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 840                                 int migrate_all)
 841 {
 842         int err;
 843         struct page_to_node *pp;
 844         LIST_HEAD(pagelist);
 845
 846         down_read(&mm->mmap_sem);
 847
 848         /*
 849          * Build a list of pages to migrate
 850          */
 851         migrate_prep();
 852         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 853                 struct vm_area_struct *vma;
 854                 struct page *page;
 855
 856                 /*
 857                  * A valid page pointer that will not match any of the
 858                  * pages that will be moved.
 859                  */
 860                 pp->page = ZERO_PAGE(0);
 861
 862                 err = -EFAULT;
 863                 vma = find_vma(mm, pp->addr);
 864                 if (!vma || !vma_migratable(vma))
 865                         goto set_status;
 866
 867                 page = follow_page(vma, pp->addr, FOLL_GET);
 868                 err = -ENOENT;
 869                 if (!page)
 870                         goto set_status;
 871
 872                 if (PageReserved(page))         /* Check for zero page */
 873                         goto put_and_set;
 874
 875                 pp->page = page;
 876                 err = page_to_nid(page);
 877
 878                 if (err == pp->node)
 879                         /*
 880                          * Node already in the right place
 881                          */
 882                         goto put_and_set;
 883
 884                 err = -EACCES;
 885                 if (page_mapcount(page) > 1 &&
 886                                 !migrate_all)
 887                         goto put_and_set;
 888
 889                 err = isolate_lru_page(page, &pagelist);
 890 put_and_set:
 891                 /*
 892                  * Either remove the duplicate refcount from
 893                  * isolate_lru_page() or drop the page ref if it was
 894                  * not isolated.
 895                  */
 896                 put_page(page);
 897 set_status:
 898                 pp->status = err;
 899         }
 900
 901         if (!list_empty(&pagelist))
 902                 err = migrate_pages(&pagelist, new_page_node,
 903                                 (unsigned long)pm);
 904         else
 905                 err = -ENOENT;
 906
 907         up_read(&mm->mmap_sem);
 908         return err;
 909 }
 910
 911 /*
 912  * Determine the nodes of a list of pages. The addr in the pm array
 913  * must have been set to the virtual address of which we want to determine
 914  * the node number.
 915  */
 916 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 917 {
 918         down_read(&mm->mmap_sem);
 919
 920         for ( ; pm->node != MAX_NUMNODES; pm++) {
 921                 struct vm_area_struct *vma;
 922                 struct page *page;
 923                 int err;
 924
 925                 err = -EFAULT;
 926                 vma = find_vma(mm, pm->addr);
 927                 if (!vma)
 928                         goto set_status;
 929
 930                 page = follow_page(vma, pm->addr, 0);
 931                 err = -ENOENT;
 932                 /* Use PageReserved to check for zero page */
 933                 if (!page || PageReserved(page))
 934                         goto set_status;
 935
 936                 err = page_to_nid(page);
 937 set_status:
 938                 pm->status = err;
 939         }
 940
 941         up_read(&mm->mmap_sem);
 942         return 0;
 943 }
 944
 945 /*
 946  * Move a list of pages in the address space of the currently executing
 947  * process.
 948  */
 949 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 950                         const void __user * __user *pages,
 951                         const int __user *nodes,
 952                         int __user *status, int flags)
 953 {
 954         int err = 0;
 955         int i;
 956         struct task_struct *task;
 957         nodemask_t task_nodes;
 958         struct mm_struct *mm;
 959         struct page_to_node *pm = NULL;
 960
 961         /* Check flags */
 962         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 963                 return -EINVAL;
 964
 965         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 966                 return -EPERM;
 967
 968         /* Find the mm_struct */
 969         read_lock(&tasklist_lock);
 970         task = pid ? find_task_by_vpid(pid) : current;
 971         if (!task) {
 972                 read_unlock(&tasklist_lock);
 973                 return -ESRCH;
 974         }
 975         mm = get_task_mm(task);
 976         read_unlock(&tasklist_lock);
 977
 978         if (!mm)
 979                 return -EINVAL;
 980
 981         /*
 982          * Check if this process has the right to modify the specified
 983          * process. The right exists if the process has administrative
 984          * capabilities, superuser privileges or the same
 985          * userid as the target process.
 986          */
 987         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 988             (current->uid != task->suid) && (current->uid != task->uid) &&
 989             !capable(CAP_SYS_NICE)) {
 990                 err = -EPERM;
 991                 goto out2;
 992         }
 993
 994         err = security_task_movememory(task);
 995         if (err)
 996                 goto out2;
 997
 998
 999         task_nodes = cpuset_mems_allowed(task);
1000
1001         /* Limit nr_pages so that the multiplication may not overflow */
1002         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1003                 err = -E2BIG;
1004                 goto out2;
1005         }
1006
1007         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1008         if (!pm) {
1009                 err = -ENOMEM;
1010                 goto out2;
1011         }
1012
1013         /*
1014          * Get parameters from user space and initialize the pm
1015          * array. Return various errors if the user did something wrong.
1016          */
1017         for (i = 0; i < nr_pages; i++) {
1018                 const void __user *p;
1019
1020                 err = -EFAULT;
1021                 if (get_user(p, pages + i))
1022                         goto out;
1023
1024                 pm[i].addr = (unsigned long)p;
1025                 if (nodes) {
1026                         int node;
1027
1028                         if (get_user(node, nodes + i))
1029                                 goto out;
1030
1031                         err = -ENODEV;
1032                         if (!node_state(node, N_HIGH_MEMORY))
1033                                 goto out;
1034
1035                         err = -EACCES;
1036                         if (!node_isset(node, task_nodes))
1037                                 goto out;
1038
1039                         pm[i].node = node;
1040                 } else
1041                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1042         }
1043         /* End marker */
1044         pm[nr_pages].node = MAX_NUMNODES;
1045
1046         if (nodes)
1047                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1048         else
1049                 err = do_pages_stat(mm, pm);
1050
1051         if (err >= 0)
1052                 /* Return status information */
1053                 for (i = 0; i < nr_pages; i++)
1054                         if (put_user(pm[i].status, status + i))
1055                                 err = -EFAULT;
1056
1057 out:
1058         vfree(pm);
1059 out2:
1060         mmput(mm);
1061         return err;
1062 }
1063 #endif
1064
1065 /*
1066  * Call migration functions in the vma_ops that may prepare
1067  * memory in a vm for migration. migration functions may perform
1068  * the migration for vmas that do not have an underlying page struct.
1069  */
1070 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1071         const nodemask_t *from, unsigned long flags)
1072 {
1073         struct vm_area_struct *vma;
1074         int err = 0;
1075
1076         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1077                 if (vma->vm_ops && vma->vm_ops->migrate) {
1078                         err = vma->vm_ops->migrate(vma, to, from, flags);
1079                         if (err)
1080                                 break;
1081                 }
1082         }
1083         return err;
1084 }