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/pagevec.h>
  23 #include <linux/rmap.h>
  24 #include <linux/topology.h>
  25 #include <linux/cpu.h>
  26 #include <linux/cpuset.h>
  27 #include <linux/writeback.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/security.h>
  31
  32 #include "internal.h"
  33
  34 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  35
  36 /*
  37  * Isolate one page from the LRU lists. If successful put it onto
  38  * the indicated list with elevated page count.
  39  *
  40  * Result:
  41  *  -EBUSY: page not on LRU list
  42  *  0: page removed from LRU list and added to the specified list.
  43  */
  44 int isolate_lru_page(struct page *page, struct list_head *pagelist)
  45 {
  46         int ret = -EBUSY;
  47
  48         if (PageLRU(page)) {
  49                 struct zone *zone = page_zone(page);
  50
  51                 spin_lock_irq(&zone->lru_lock);
  52                 if (PageLRU(page) && get_page_unless_zero(page)) {
  53                         ret = 0;
  54                         ClearPageLRU(page);
  55                         if (PageActive(page))
  56                                 del_page_from_active_list(zone, page);
  57                         else
  58                                 del_page_from_inactive_list(zone, page);
  59                         list_add_tail(&page->lru, pagelist);
  60                 }
  61                 spin_unlock_irq(&zone->lru_lock);
  62         }
  63         return ret;
  64 }
  65
  66 /*
  67  * migrate_prep() needs to be called before we start compiling a list of pages
  68  * to be migrated using isolate_lru_page().
  69  */
  70 int migrate_prep(void)
  71 {
  72         /*
  73          * Clear the LRU lists so pages can be isolated.
  74          * Note that pages may be moved off the LRU after we have
  75          * drained them. Those pages will fail to migrate like other
  76          * pages that may be busy.
  77          */
  78         lru_add_drain_all();
  79
  80         return 0;
  81 }
  82
  83 static inline void move_to_lru(struct page *page)
  84 {
  85         if (PageActive(page)) {
  86                 /*
  87                  * lru_cache_add_active checks that
  88                  * the PG_active bit is off.
  89                  */
  90                 ClearPageActive(page);
  91                 lru_cache_add_active(page);
  92         } else {
  93                 lru_cache_add(page);
  94         }
  95         put_page(page);
  96 }
  97
  98 /*
  99  * Add isolated pages on the list back to the LRU.
 100  *
 101  * returns the number of pages put back.
 102  */
 103 int putback_lru_pages(struct list_head *l)
 104 {
 105         struct page *page;
 106         struct page *page2;
 107         int count = 0;
 108
 109         list_for_each_entry_safe(page, page2, l, lru) {
 110                 list_del(&page->lru);
 111                 move_to_lru(page);
 112                 count++;
 113         }
 114         return count;
 115 }
 116
 117 static inline int is_swap_pte(pte_t pte)
 118 {
 119         return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
 120 }
 121
 122 /*
 123  * Restore a potential migration pte to a working pte entry
 124  */
 125 static void remove_migration_pte(struct vm_area_struct *vma,
 126                 struct page *old, struct page *new)
 127 {
 128         struct mm_struct *mm = vma->vm_mm;
 129         swp_entry_t entry;
 130         pgd_t *pgd;
 131         pud_t *pud;
 132         pmd_t *pmd;
 133         pte_t *ptep, pte;
 134         spinlock_t *ptl;
 135         unsigned long addr = page_address_in_vma(new, vma);
 136
 137         if (addr == -EFAULT)
 138                 return;
 139
 140         pgd = pgd_offset(mm, addr);
 141         if (!pgd_present(*pgd))
 142                 return;
 143
 144         pud = pud_offset(pgd, addr);
 145         if (!pud_present(*pud))
 146                 return;
 147
 148         pmd = pmd_offset(pud, addr);
 149         if (!pmd_present(*pmd))
 150                 return;
 151
 152         ptep = pte_offset_map(pmd, addr);
 153
 154         if (!is_swap_pte(*ptep)) {
 155                 pte_unmap(ptep);
 156                 return;
 157         }
 158
 159         ptl = pte_lockptr(mm, pmd);
 160         spin_lock(ptl);
 161         pte = *ptep;
 162         if (!is_swap_pte(pte))
 163                 goto out;
 164
 165         entry = pte_to_swp_entry(pte);
 166
 167         if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 168                 goto out;
 169
 170         get_page(new);
 171         pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 172         if (is_write_migration_entry(entry))
 173                 pte = pte_mkwrite(pte);
 174         set_pte_at(mm, addr, ptep, pte);
 175
 176         if (PageAnon(new))
 177                 page_add_anon_rmap(new, vma, addr);
 178         else
 179                 page_add_file_rmap(new);
 180
 181         /* No need to invalidate - it was non-present before */
 182         update_mmu_cache(vma, addr, pte);
 183         lazy_mmu_prot_update(pte);
 184
 185 out:
 186         pte_unmap_unlock(ptep, ptl);
 187 }
 188
 189 /*
 190  * Note that remove_file_migration_ptes will only work on regular mappings,
 191  * Nonlinear mappings do not use migration entries.
 192  */
 193 static void remove_file_migration_ptes(struct page *old, struct page *new)
 194 {
 195         struct vm_area_struct *vma;
 196         struct address_space *mapping = page_mapping(new);
 197         struct prio_tree_iter iter;
 198         pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 199
 200         if (!mapping)
 201                 return;
 202
 203         spin_lock(&mapping->i_mmap_lock);
 204
 205         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
 206                 remove_migration_pte(vma, old, new);
 207
 208         spin_unlock(&mapping->i_mmap_lock);
 209 }
 210
 211 /*
 212  * Must hold mmap_sem lock on at least one of the vmas containing
 213  * the page so that the anon_vma cannot vanish.
 214  */
 215 static void remove_anon_migration_ptes(struct page *old, struct page *new)
 216 {
 217         struct anon_vma *anon_vma;
 218         struct vm_area_struct *vma;
 219         unsigned long mapping;
 220
 221         mapping = (unsigned long)new->mapping;
 222
 223         if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
 224                 return;
 225
 226         /*
 227          * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
 228          */
 229         anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
 230         spin_lock(&anon_vma->lock);
 231
 232         list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
 233                 remove_migration_pte(vma, old, new);
 234
 235         spin_unlock(&anon_vma->lock);
 236 }
 237
 238 /*
 239  * Get rid of all migration entries and replace them by
 240  * references to the indicated page.
 241  */
 242 static void remove_migration_ptes(struct page *old, struct page *new)
 243 {
 244         if (PageAnon(new))
 245                 remove_anon_migration_ptes(old, new);
 246         else
 247                 remove_file_migration_ptes(old, new);
 248 }
 249
 250 /*
 251  * Something used the pte of a page under migration. We need to
 252  * get to the page and wait until migration is finished.
 253  * When we return from this function the fault will be retried.
 254  *
 255  * This function is called from do_swap_page().
 256  */
 257 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 258                                 unsigned long address)
 259 {
 260         pte_t *ptep, pte;
 261         spinlock_t *ptl;
 262         swp_entry_t entry;
 263         struct page *page;
 264
 265         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 266         pte = *ptep;
 267         if (!is_swap_pte(pte))
 268                 goto out;
 269
 270         entry = pte_to_swp_entry(pte);
 271         if (!is_migration_entry(entry))
 272                 goto out;
 273
 274         page = migration_entry_to_page(entry);
 275
 276         get_page(page);
 277         pte_unmap_unlock(ptep, ptl);
 278         wait_on_page_locked(page);
 279         put_page(page);
 280         return;
 281 out:
 282         pte_unmap_unlock(ptep, ptl);
 283 }
 284
 285 /*
 286  * Replace the page in the mapping.
 287  *
 288  * The number of remaining references must be:
 289  * 1 for anonymous pages without a mapping
 290  * 2 for pages with a mapping
 291  * 3 for pages with a mapping and PagePrivate set.
 292  */
 293 static int migrate_page_move_mapping(struct address_space *mapping,
 294                 struct page *newpage, struct page *page)
 295 {
 296         void **pslot;
 297
 298         if (!mapping) {
 299                 /* Anonymous page without mapping */
 300                 if (page_count(page) != 1)
 301                         return -EAGAIN;
 302                 return 0;
 303         }
 304
 305         write_lock_irq(&mapping->tree_lock);
 306
 307         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 308                                         page_index(page));
 309
 310         if (page_count(page) != 2 + !!PagePrivate(page) ||
 311                         (struct page *)radix_tree_deref_slot(pslot) != page) {
 312                 write_unlock_irq(&mapping->tree_lock);
 313                 return -EAGAIN;
 314         }
 315
 316         /*
 317          * Now we know that no one else is looking at the page.
 318          */
 319         get_page(newpage);      /* add cache reference */
 320 #ifdef CONFIG_SWAP
 321         if (PageSwapCache(page)) {
 322                 SetPageSwapCache(newpage);
 323                 set_page_private(newpage, page_private(page));
 324         }
 325 #endif
 326
 327         radix_tree_replace_slot(pslot, newpage);
 328
 329         /*
 330          * Drop cache reference from old page.
 331          * We know this isn't the last reference.
 332          */
 333         __put_page(page);
 334
 335         /*
 336          * If moved to a different zone then also account
 337          * the page for that zone. Other VM counters will be
 338          * taken care of when we establish references to the
 339          * new page and drop references to the old page.
 340          *
 341          * Note that anonymous pages are accounted for
 342          * via NR_FILE_PAGES and NR_ANON_PAGES if they
 343          * are mapped to swap space.
 344          */
 345         __dec_zone_page_state(page, NR_FILE_PAGES);
 346         __inc_zone_page_state(newpage, NR_FILE_PAGES);
 347
 348         write_unlock_irq(&mapping->tree_lock);
 349
 350         return 0;
 351 }
 352
 353 /*
 354  * Copy the page to its new location
 355  */
 356 static void migrate_page_copy(struct page *newpage, struct page *page)
 357 {
 358         copy_highpage(newpage, page);
 359
 360         if (PageError(page))
 361                 SetPageError(newpage);
 362         if (PageReferenced(page))
 363                 SetPageReferenced(newpage);
 364         if (PageUptodate(page))
 365                 SetPageUptodate(newpage);
 366         if (PageActive(page))
 367                 SetPageActive(newpage);
 368         if (PageChecked(page))
 369                 SetPageChecked(newpage);
 370         if (PageMappedToDisk(page))
 371                 SetPageMappedToDisk(newpage);
 372
 373         if (PageDirty(page)) {
 374                 clear_page_dirty_for_io(page);
 375                 set_page_dirty(newpage);
 376         }
 377
 378 #ifdef CONFIG_SWAP
 379         ClearPageSwapCache(page);
 380 #endif
 381         ClearPageActive(page);
 382         ClearPagePrivate(page);
 383         set_page_private(page, 0);
 384         page->mapping = NULL;
 385
 386         /*
 387          * If any waiters have accumulated on the new page then
 388          * wake them up.
 389          */
 390         if (PageWriteback(newpage))
 391                 end_page_writeback(newpage);
 392 }
 393
 394 /************************************************************
 395  *                    Migration functions
 396  ***********************************************************/
 397
 398 /* Always fail migration. Used for mappings that are not movable */
 399 int fail_migrate_page(struct address_space *mapping,
 400                         struct page *newpage, struct page *page)
 401 {
 402         return -EIO;
 403 }
 404 EXPORT_SYMBOL(fail_migrate_page);
 405
 406 /*
 407  * Common logic to directly migrate a single page suitable for
 408  * pages that do not use PagePrivate.
 409  *
 410  * Pages are locked upon entry and exit.
 411  */
 412 int migrate_page(struct address_space *mapping,
 413                 struct page *newpage, struct page *page)
 414 {
 415         int rc;
 416
 417         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 418
 419         rc = migrate_page_move_mapping(mapping, newpage, page);
 420
 421         if (rc)
 422                 return rc;
 423
 424         migrate_page_copy(newpage, page);
 425         return 0;
 426 }
 427 EXPORT_SYMBOL(migrate_page);
 428
 429 #ifdef CONFIG_BLOCK
 430 /*
 431  * Migration function for pages with buffers. This function can only be used
 432  * if the underlying filesystem guarantees that no other references to "page"
 433  * exist.
 434  */
 435 int buffer_migrate_page(struct address_space *mapping,
 436                 struct page *newpage, struct page *page)
 437 {
 438         struct buffer_head *bh, *head;
 439         int rc;
 440
 441         if (!page_has_buffers(page))
 442                 return migrate_page(mapping, newpage, page);
 443
 444         head = page_buffers(page);
 445
 446         rc = migrate_page_move_mapping(mapping, newpage, page);
 447
 448         if (rc)
 449                 return rc;
 450
 451         bh = head;
 452         do {
 453                 get_bh(bh);
 454                 lock_buffer(bh);
 455                 bh = bh->b_this_page;
 456
 457         } while (bh != head);
 458
 459         ClearPagePrivate(page);
 460         set_page_private(newpage, page_private(page));
 461         set_page_private(page, 0);
 462         put_page(page);
 463         get_page(newpage);
 464
 465         bh = head;
 466         do {
 467                 set_bh_page(bh, newpage, bh_offset(bh));
 468                 bh = bh->b_this_page;
 469
 470         } while (bh != head);
 471
 472         SetPagePrivate(newpage);
 473
 474         migrate_page_copy(newpage, page);
 475
 476         bh = head;
 477         do {
 478                 unlock_buffer(bh);
 479                 put_bh(bh);
 480                 bh = bh->b_this_page;
 481
 482         } while (bh != head);
 483
 484         return 0;
 485 }
 486 EXPORT_SYMBOL(buffer_migrate_page);
 487 #endif
 488
 489 /*
 490  * Writeback a page to clean the dirty state
 491  */
 492 static int writeout(struct address_space *mapping, struct page *page)
 493 {
 494         struct writeback_control wbc = {
 495                 .sync_mode = WB_SYNC_NONE,
 496                 .nr_to_write = 1,
 497                 .range_start = 0,
 498                 .range_end = LLONG_MAX,
 499                 .nonblocking = 1,
 500                 .for_reclaim = 1
 501         };
 502         int rc;
 503
 504         if (!mapping->a_ops->writepage)
 505                 /* No write method for the address space */
 506                 return -EINVAL;
 507
 508         if (!clear_page_dirty_for_io(page))
 509                 /* Someone else already triggered a write */
 510                 return -EAGAIN;
 511
 512         /*
 513          * A dirty page may imply that the underlying filesystem has
 514          * the page on some queue. So the page must be clean for
 515          * migration. Writeout may mean we loose the lock and the
 516          * page state is no longer what we checked for earlier.
 517          * At this point we know that the migration attempt cannot
 518          * be successful.
 519          */
 520         remove_migration_ptes(page, page);
 521
 522         rc = mapping->a_ops->writepage(page, &wbc);
 523         if (rc < 0)
 524                 /* I/O Error writing */
 525                 return -EIO;
 526
 527         if (rc != AOP_WRITEPAGE_ACTIVATE)
 528                 /* unlocked. Relock */
 529                 lock_page(page);
 530
 531         return -EAGAIN;
 532 }
 533
 534 /*
 535  * Default handling if a filesystem does not provide a migration function.
 536  */
 537 static int fallback_migrate_page(struct address_space *mapping,
 538         struct page *newpage, struct page *page)
 539 {
 540         if (PageDirty(page))
 541                 return writeout(mapping, page);
 542
 543         /*
 544          * Buffers may be managed in a filesystem specific way.
 545          * We must have no buffers or drop them.
 546          */
 547         if (PagePrivate(page) &&
 548             !try_to_release_page(page, GFP_KERNEL))
 549                 return -EAGAIN;
 550
 551         return migrate_page(mapping, newpage, page);
 552 }
 553
 554 /*
 555  * Move a page to a newly allocated page
 556  * The page is locked and all ptes have been successfully removed.
 557  *
 558  * The new page will have replaced the old page if this function
 559  * is successful.
 560  */
 561 static int move_to_new_page(struct page *newpage, struct page *page)
 562 {
 563         struct address_space *mapping;
 564         int rc;
 565
 566         /*
 567          * Block others from accessing the page when we get around to
 568          * establishing additional references. We are the only one
 569          * holding a reference to the new page at this point.
 570          */
 571         if (TestSetPageLocked(newpage))
 572                 BUG();
 573
 574         /* Prepare mapping for the new page.*/
 575         newpage->index = page->index;
 576         newpage->mapping = page->mapping;
 577
 578         mapping = page_mapping(page);
 579         if (!mapping)
 580                 rc = migrate_page(mapping, newpage, page);
 581         else if (mapping->a_ops->migratepage)
 582                 /*
 583                  * Most pages have a mapping and most filesystems
 584                  * should provide a migration function. Anonymous
 585                  * pages are part of swap space which also has its
 586                  * own migration function. This is the most common
 587                  * path for page migration.
 588                  */
 589                 rc = mapping->a_ops->migratepage(mapping,
 590                                                 newpage, page);
 591         else
 592                 rc = fallback_migrate_page(mapping, newpage, page);
 593
 594         if (!rc)
 595                 remove_migration_ptes(page, newpage);
 596         else
 597                 newpage->mapping = NULL;
 598
 599         unlock_page(newpage);
 600
 601         return rc;
 602 }
 603
 604 /*
 605  * Obtain the lock on page, remove all ptes and migrate the page
 606  * to the newly allocated page in newpage.
 607  */
 608 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 609                         struct page *page, int force)
 610 {
 611         int rc = 0;
 612         int *result = NULL;
 613         struct page *newpage = get_new_page(page, private, &result);
 614         int rcu_locked = 0;
 615
 616         if (!newpage)
 617                 return -ENOMEM;
 618
 619         if (page_count(page) == 1)
 620                 /* page was freed from under us. So we are done. */
 621                 goto move_newpage;
 622
 623         rc = -EAGAIN;
 624         if (TestSetPageLocked(page)) {
 625                 if (!force)
 626                         goto move_newpage;
 627                 lock_page(page);
 628         }
 629
 630         if (PageWriteback(page)) {
 631                 if (!force)
 632                         goto unlock;
 633                 wait_on_page_writeback(page);
 634         }
 635         /*
 636          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 637          * we cannot notice that anon_vma is freed while we migrates a page.
 638          * This rcu_read_lock() delays freeing anon_vma pointer until the end
 639          * of migration. File cache pages are no problem because of page_lock()
 640          * File Caches may use write_page() or lock_page() in migration, then,
 641          * just care Anon page here.
 642          */
 643         if (PageAnon(page)) {
 644                 rcu_read_lock();
 645                 rcu_locked = 1;
 646         }
 647         /*
 648          * This is a corner case handling.
 649          * When a new swap-cache is read into, it is linked to LRU
 650          * and treated as swapcache but has no rmap yet.
 651          * Calling try_to_unmap() against a page->mapping==NULL page is
 652          * BUG. So handle it here.
 653          */
 654         if (!page->mapping)
 655                 goto rcu_unlock;
 656         /* Establish migration ptes or remove ptes */
 657         try_to_unmap(page, 1);
 658
 659         if (!page_mapped(page))
 660                 rc = move_to_new_page(newpage, page);
 661
 662         if (rc)
 663                 remove_migration_ptes(page, page);
 664 rcu_unlock:
 665         if (rcu_locked)
 666                 rcu_read_unlock();
 667
 668 unlock:
 669
 670         unlock_page(page);
 671
 672         if (rc != -EAGAIN) {
 673                 /*
 674                  * A page that has been migrated has all references
 675                  * removed and will be freed. A page that has not been
 676                  * migrated will have kepts its references and be
 677                  * restored.
 678                  */
 679                 list_del(&page->lru);
 680                 move_to_lru(page);
 681         }
 682
 683 move_newpage:
 684         /*
 685          * Move the new page to the LRU. If migration was not successful
 686          * then this will free the page.
 687          */
 688         move_to_lru(newpage);
 689         if (result) {
 690                 if (rc)
 691                         *result = rc;
 692                 else
 693                         *result = page_to_nid(newpage);
 694         }
 695         return rc;
 696 }
 697
 698 /*
 699  * migrate_pages
 700  *
 701  * The function takes one list of pages to migrate and a function
 702  * that determines from the page to be migrated and the private data
 703  * the target of the move and allocates the page.
 704  *
 705  * The function returns after 10 attempts or if no pages
 706  * are movable anymore because to has become empty
 707  * or no retryable pages exist anymore. All pages will be
 708  * retruned to the LRU or freed.
 709  *
 710  * Return: Number of pages not migrated or error code.
 711  */
 712 int migrate_pages(struct list_head *from,
 713                 new_page_t get_new_page, unsigned long private)
 714 {
 715         int retry = 1;
 716         int nr_failed = 0;
 717         int pass = 0;
 718         struct page *page;
 719         struct page *page2;
 720         int swapwrite = current->flags & PF_SWAPWRITE;
 721         int rc;
 722
 723         if (!swapwrite)
 724                 current->flags |= PF_SWAPWRITE;
 725
 726         for(pass = 0; pass < 10 && retry; pass++) {
 727                 retry = 0;
 728
 729                 list_for_each_entry_safe(page, page2, from, lru) {
 730                         cond_resched();
 731
 732                         rc = unmap_and_move(get_new_page, private,
 733                                                 page, pass > 2);
 734
 735                         switch(rc) {
 736                         case -ENOMEM:
 737                                 goto out;
 738                         case -EAGAIN:
 739                                 retry++;
 740                                 break;
 741                         case 0:
 742                                 break;
 743                         default:
 744                                 /* Permanent failure */
 745                                 nr_failed++;
 746                                 break;
 747                         }
 748                 }
 749         }
 750         rc = 0;
 751 out:
 752         if (!swapwrite)
 753                 current->flags &= ~PF_SWAPWRITE;
 754
 755         putback_lru_pages(from);
 756
 757         if (rc)
 758                 return rc;
 759
 760         return nr_failed + retry;
 761 }
 762
 763 #ifdef CONFIG_NUMA
 764 /*
 765  * Move a list of individual pages
 766  */
 767 struct page_to_node {
 768         unsigned long addr;
 769         struct page *page;
 770         int node;
 771         int status;
 772 };
 773
 774 static struct page *new_page_node(struct page *p, unsigned long private,
 775                 int **result)
 776 {
 777         struct page_to_node *pm = (struct page_to_node *)private;
 778
 779         while (pm->node != MAX_NUMNODES && pm->page != p)
 780                 pm++;
 781
 782         if (pm->node == MAX_NUMNODES)
 783                 return NULL;
 784
 785         *result = &pm->status;
 786
 787         return alloc_pages_node(pm->node,
 788                                 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
 789 }
 790
 791 /*
 792  * Move a set of pages as indicated in the pm array. The addr
 793  * field must be set to the virtual address of the page to be moved
 794  * and the node number must contain a valid target node.
 795  */
 796 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
 797                                 int migrate_all)
 798 {
 799         int err;
 800         struct page_to_node *pp;
 801         LIST_HEAD(pagelist);
 802
 803         down_read(&mm->mmap_sem);
 804
 805         /*
 806          * Build a list of pages to migrate
 807          */
 808         migrate_prep();
 809         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
 810                 struct vm_area_struct *vma;
 811                 struct page *page;
 812
 813                 /*
 814                  * A valid page pointer that will not match any of the
 815                  * pages that will be moved.
 816                  */
 817                 pp->page = ZERO_PAGE(0);
 818
 819                 err = -EFAULT;
 820                 vma = find_vma(mm, pp->addr);
 821                 if (!vma || !vma_migratable(vma))
 822                         goto set_status;
 823
 824                 page = follow_page(vma, pp->addr, FOLL_GET);
 825                 err = -ENOENT;
 826                 if (!page)
 827                         goto set_status;
 828
 829                 if (PageReserved(page))         /* Check for zero page */
 830                         goto put_and_set;
 831
 832                 pp->page = page;
 833                 err = page_to_nid(page);
 834
 835                 if (err == pp->node)
 836                         /*
 837                          * Node already in the right place
 838                          */
 839                         goto put_and_set;
 840
 841                 err = -EACCES;
 842                 if (page_mapcount(page) > 1 &&
 843                                 !migrate_all)
 844                         goto put_and_set;
 845
 846                 err = isolate_lru_page(page, &pagelist);
 847 put_and_set:
 848                 /*
 849                  * Either remove the duplicate refcount from
 850                  * isolate_lru_page() or drop the page ref if it was
 851                  * not isolated.
 852                  */
 853                 put_page(page);
 854 set_status:
 855                 pp->status = err;
 856         }
 857
 858         if (!list_empty(&pagelist))
 859                 err = migrate_pages(&pagelist, new_page_node,
 860                                 (unsigned long)pm);
 861         else
 862                 err = -ENOENT;
 863
 864         up_read(&mm->mmap_sem);
 865         return err;
 866 }
 867
 868 /*
 869  * Determine the nodes of a list of pages. The addr in the pm array
 870  * must have been set to the virtual address of which we want to determine
 871  * the node number.
 872  */
 873 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
 874 {
 875         down_read(&mm->mmap_sem);
 876
 877         for ( ; pm->node != MAX_NUMNODES; pm++) {
 878                 struct vm_area_struct *vma;
 879                 struct page *page;
 880                 int err;
 881
 882                 err = -EFAULT;
 883                 vma = find_vma(mm, pm->addr);
 884                 if (!vma)
 885                         goto set_status;
 886
 887                 page = follow_page(vma, pm->addr, 0);
 888                 err = -ENOENT;
 889                 /* Use PageReserved to check for zero page */
 890                 if (!page || PageReserved(page))
 891                         goto set_status;
 892
 893                 err = page_to_nid(page);
 894 set_status:
 895                 pm->status = err;
 896         }
 897
 898         up_read(&mm->mmap_sem);
 899         return 0;
 900 }
 901
 902 /*
 903  * Move a list of pages in the address space of the currently executing
 904  * process.
 905  */
 906 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
 907                         const void __user * __user *pages,
 908                         const int __user *nodes,
 909                         int __user *status, int flags)
 910 {
 911         int err = 0;
 912         int i;
 913         struct task_struct *task;
 914         nodemask_t task_nodes;
 915         struct mm_struct *mm;
 916         struct page_to_node *pm = NULL;
 917
 918         /* Check flags */
 919         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
 920                 return -EINVAL;
 921
 922         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
 923                 return -EPERM;
 924
 925         /* Find the mm_struct */
 926         read_lock(&tasklist_lock);
 927         task = pid ? find_task_by_pid(pid) : current;
 928         if (!task) {
 929                 read_unlock(&tasklist_lock);
 930                 return -ESRCH;
 931         }
 932         mm = get_task_mm(task);
 933         read_unlock(&tasklist_lock);
 934
 935         if (!mm)
 936                 return -EINVAL;
 937
 938         /*
 939          * Check if this process has the right to modify the specified
 940          * process. The right exists if the process has administrative
 941          * capabilities, superuser privileges or the same
 942          * userid as the target process.
 943          */
 944         if ((current->euid != task->suid) && (current->euid != task->uid) &&
 945             (current->uid != task->suid) && (current->uid != task->uid) &&
 946             !capable(CAP_SYS_NICE)) {
 947                 err = -EPERM;
 948                 goto out2;
 949         }
 950
 951         err = security_task_movememory(task);
 952         if (err)
 953                 goto out2;
 954
 955
 956         task_nodes = cpuset_mems_allowed(task);
 957
 958         /* Limit nr_pages so that the multiplication may not overflow */
 959         if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
 960                 err = -E2BIG;
 961                 goto out2;
 962         }
 963
 964         pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
 965         if (!pm) {
 966                 err = -ENOMEM;
 967                 goto out2;
 968         }
 969
 970         /*
 971          * Get parameters from user space and initialize the pm
 972          * array. Return various errors if the user did something wrong.
 973          */
 974         for (i = 0; i < nr_pages; i++) {
 975                 const void *p;
 976
 977                 err = -EFAULT;
 978                 if (get_user(p, pages + i))
 979                         goto out;
 980
 981                 pm[i].addr = (unsigned long)p;
 982                 if (nodes) {
 983                         int node;
 984
 985                         if (get_user(node, nodes + i))
 986                                 goto out;
 987
 988                         err = -ENODEV;
 989                         if (!node_online(node))
 990                                 goto out;
 991
 992                         err = -EACCES;
 993                         if (!node_isset(node, task_nodes))
 994                                 goto out;
 995
 996                         pm[i].node = node;
 997                 } else
 998                         pm[i].node = 0; /* anything to not match MAX_NUMNODES */
 999         }
1000         /* End marker */
1001         pm[nr_pages].node = MAX_NUMNODES;
1002
1003         if (nodes)
1004                 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1005         else
1006                 err = do_pages_stat(mm, pm);
1007
1008         if (err >= 0)
1009                 /* Return status information */
1010                 for (i = 0; i < nr_pages; i++)
1011                         if (put_user(pm[i].status, status + i))
1012                                 err = -EFAULT;
1013
1014 out:
1015         vfree(pm);
1016 out2:
1017         mmput(mm);
1018         return err;
1019 }
1020 #endif
1021
1022 /*
1023  * Call migration functions in the vma_ops that may prepare
1024  * memory in a vm for migration. migration functions may perform
1025  * the migration for vmas that do not have an underlying page struct.
1026  */
1027 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1028         const nodemask_t *from, unsigned long flags)
1029 {
1030         struct vm_area_struct *vma;
1031         int err = 0;
1032
1033         for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1034                 if (vma->vm_ops && vma->vm_ops->migrate) {
1035                         err = vma->vm_ops->migrate(vma, to, from, flags);
1036                         if (err)
1037                                 break;
1038                 }
1039         }
1040         return err;
1041 }