migrate.c revision bda8550deed96687f29992d711a88ea21cff4d26
1/* 2 * Memory Migration functionality - linux/mm/migration.c 3 * 4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter 5 * 6 * Page migration was first developed in the context of the memory hotplug 7 * project. The main authors of the migration code are: 8 * 9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> 10 * Hirokazu Takahashi <taka@valinux.co.jp> 11 * Dave Hansen <haveblue@us.ibm.com> 12 * Christoph Lameter 13 */ 14 15#include <linux/migrate.h> 16#include <linux/module.h> 17#include <linux/swap.h> 18#include <linux/swapops.h> 19#include <linux/pagemap.h> 20#include <linux/buffer_head.h> 21#include <linux/mm_inline.h> 22#include <linux/nsproxy.h> 23#include <linux/pagevec.h> 24#include <linux/rmap.h> 25#include <linux/topology.h> 26#include <linux/cpu.h> 27#include <linux/cpuset.h> 28#include <linux/writeback.h> 29#include <linux/mempolicy.h> 30#include <linux/vmalloc.h> 31#include <linux/security.h> 32#include <linux/memcontrol.h> 33#include <linux/syscalls.h> 34 35#include "internal.h" 36 37#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 38 39/* 40 * migrate_prep() needs to be called before we start compiling a list of pages 41 * to be migrated using isolate_lru_page(). 42 */ 43int migrate_prep(void) 44{ 45 /* 46 * Clear the LRU lists so pages can be isolated. 47 * Note that pages may be moved off the LRU after we have 48 * drained them. Those pages will fail to migrate like other 49 * pages that may be busy. 50 */ 51 lru_add_drain_all(); 52 53 return 0; 54} 55 56/* 57 * Add isolated pages on the list back to the LRU under page lock 58 * to avoid leaking evictable pages back onto unevictable list. 59 * 60 * returns the number of pages put back. 61 */ 62int putback_lru_pages(struct list_head *l) 63{ 64 struct page *page; 65 struct page *page2; 66 int count = 0; 67 68 list_for_each_entry_safe(page, page2, l, lru) { 69 list_del(&page->lru); 70 putback_lru_page(page); 71 count++; 72 } 73 return count; 74} 75 76/* 77 * Restore a potential migration pte to a working pte entry 78 */ 79static void remove_migration_pte(struct vm_area_struct *vma, 80 struct page *old, struct page *new) 81{ 82 struct mm_struct *mm = vma->vm_mm; 83 swp_entry_t entry; 84 pgd_t *pgd; 85 pud_t *pud; 86 pmd_t *pmd; 87 pte_t *ptep, pte; 88 spinlock_t *ptl; 89 unsigned long addr = page_address_in_vma(new, vma); 90 91 if (addr == -EFAULT) 92 return; 93 94 pgd = pgd_offset(mm, addr); 95 if (!pgd_present(*pgd)) 96 return; 97 98 pud = pud_offset(pgd, addr); 99 if (!pud_present(*pud)) 100 return; 101 102 pmd = pmd_offset(pud, addr); 103 if (!pmd_present(*pmd)) 104 return; 105 106 ptep = pte_offset_map(pmd, addr); 107 108 if (!is_swap_pte(*ptep)) { 109 pte_unmap(ptep); 110 return; 111 } 112 113 ptl = pte_lockptr(mm, pmd); 114 spin_lock(ptl); 115 pte = *ptep; 116 if (!is_swap_pte(pte)) 117 goto out; 118 119 entry = pte_to_swp_entry(pte); 120 121 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) 122 goto out; 123 124 /* 125 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge. 126 * Failure is not an option here: we're now expected to remove every 127 * migration pte, and will cause crashes otherwise. Normally this 128 * is not an issue: mem_cgroup_prepare_migration bumped up the old 129 * page_cgroup count for safety, that's now attached to the new page, 130 * so this charge should just be another incrementation of the count, 131 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if 132 * there's been a force_empty, those reference counts may no longer 133 * be reliable, and this charge can actually fail: oh well, we don't 134 * make the situation any worse by proceeding as if it had succeeded. 135 */ 136 mem_cgroup_charge(new, mm, GFP_ATOMIC); 137 138 get_page(new); 139 pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); 140 if (is_write_migration_entry(entry)) 141 pte = pte_mkwrite(pte); 142 flush_cache_page(vma, addr, pte_pfn(pte)); 143 set_pte_at(mm, addr, ptep, pte); 144 145 if (PageAnon(new)) 146 page_add_anon_rmap(new, vma, addr); 147 else 148 page_add_file_rmap(new); 149 150 /* No need to invalidate - it was non-present before */ 151 update_mmu_cache(vma, addr, pte); 152 153out: 154 pte_unmap_unlock(ptep, ptl); 155} 156 157/* 158 * Note that remove_file_migration_ptes will only work on regular mappings, 159 * Nonlinear mappings do not use migration entries. 160 */ 161static void remove_file_migration_ptes(struct page *old, struct page *new) 162{ 163 struct vm_area_struct *vma; 164 struct address_space *mapping = page_mapping(new); 165 struct prio_tree_iter iter; 166 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 167 168 if (!mapping) 169 return; 170 171 spin_lock(&mapping->i_mmap_lock); 172 173 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) 174 remove_migration_pte(vma, old, new); 175 176 spin_unlock(&mapping->i_mmap_lock); 177} 178 179/* 180 * Must hold mmap_sem lock on at least one of the vmas containing 181 * the page so that the anon_vma cannot vanish. 182 */ 183static void remove_anon_migration_ptes(struct page *old, struct page *new) 184{ 185 struct anon_vma *anon_vma; 186 struct vm_area_struct *vma; 187 unsigned long mapping; 188 189 mapping = (unsigned long)new->mapping; 190 191 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) 192 return; 193 194 /* 195 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. 196 */ 197 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); 198 spin_lock(&anon_vma->lock); 199 200 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) 201 remove_migration_pte(vma, old, new); 202 203 spin_unlock(&anon_vma->lock); 204} 205 206/* 207 * Get rid of all migration entries and replace them by 208 * references to the indicated page. 209 */ 210static void remove_migration_ptes(struct page *old, struct page *new) 211{ 212 if (PageAnon(new)) 213 remove_anon_migration_ptes(old, new); 214 else 215 remove_file_migration_ptes(old, new); 216} 217 218/* 219 * Something used the pte of a page under migration. We need to 220 * get to the page and wait until migration is finished. 221 * When we return from this function the fault will be retried. 222 * 223 * This function is called from do_swap_page(). 224 */ 225void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 226 unsigned long address) 227{ 228 pte_t *ptep, pte; 229 spinlock_t *ptl; 230 swp_entry_t entry; 231 struct page *page; 232 233 ptep = pte_offset_map_lock(mm, pmd, address, &ptl); 234 pte = *ptep; 235 if (!is_swap_pte(pte)) 236 goto out; 237 238 entry = pte_to_swp_entry(pte); 239 if (!is_migration_entry(entry)) 240 goto out; 241 242 page = migration_entry_to_page(entry); 243 244 /* 245 * Once radix-tree replacement of page migration started, page_count 246 * *must* be zero. And, we don't want to call wait_on_page_locked() 247 * against a page without get_page(). 248 * So, we use get_page_unless_zero(), here. Even failed, page fault 249 * will occur again. 250 */ 251 if (!get_page_unless_zero(page)) 252 goto out; 253 pte_unmap_unlock(ptep, ptl); 254 wait_on_page_locked(page); 255 put_page(page); 256 return; 257out: 258 pte_unmap_unlock(ptep, ptl); 259} 260 261/* 262 * Replace the page in the mapping. 263 * 264 * The number of remaining references must be: 265 * 1 for anonymous pages without a mapping 266 * 2 for pages with a mapping 267 * 3 for pages with a mapping and PagePrivate set. 268 */ 269static int migrate_page_move_mapping(struct address_space *mapping, 270 struct page *newpage, struct page *page) 271{ 272 int expected_count; 273 void **pslot; 274 275 if (!mapping) { 276 /* Anonymous page without mapping */ 277 if (page_count(page) != 1) 278 return -EAGAIN; 279 return 0; 280 } 281 282 spin_lock_irq(&mapping->tree_lock); 283 284 pslot = radix_tree_lookup_slot(&mapping->page_tree, 285 page_index(page)); 286 287 expected_count = 2 + !!PagePrivate(page); 288 if (page_count(page) != expected_count || 289 (struct page *)radix_tree_deref_slot(pslot) != page) { 290 spin_unlock_irq(&mapping->tree_lock); 291 return -EAGAIN; 292 } 293 294 if (!page_freeze_refs(page, expected_count)) { 295 spin_unlock_irq(&mapping->tree_lock); 296 return -EAGAIN; 297 } 298 299 /* 300 * Now we know that no one else is looking at the page. 301 */ 302 get_page(newpage); /* add cache reference */ 303#ifdef CONFIG_SWAP 304 if (PageSwapCache(page)) { 305 SetPageSwapCache(newpage); 306 set_page_private(newpage, page_private(page)); 307 } 308#endif 309 310 radix_tree_replace_slot(pslot, newpage); 311 312 page_unfreeze_refs(page, expected_count); 313 /* 314 * Drop cache reference from old page. 315 * We know this isn't the last reference. 316 */ 317 __put_page(page); 318 319 /* 320 * If moved to a different zone then also account 321 * the page for that zone. Other VM counters will be 322 * taken care of when we establish references to the 323 * new page and drop references to the old page. 324 * 325 * Note that anonymous pages are accounted for 326 * via NR_FILE_PAGES and NR_ANON_PAGES if they 327 * are mapped to swap space. 328 */ 329 __dec_zone_page_state(page, NR_FILE_PAGES); 330 __inc_zone_page_state(newpage, NR_FILE_PAGES); 331 332 spin_unlock_irq(&mapping->tree_lock); 333 334 return 0; 335} 336 337/* 338 * Copy the page to its new location 339 */ 340static void migrate_page_copy(struct page *newpage, struct page *page) 341{ 342 int anon; 343 344 copy_highpage(newpage, page); 345 346 if (PageError(page)) 347 SetPageError(newpage); 348 if (PageReferenced(page)) 349 SetPageReferenced(newpage); 350 if (PageUptodate(page)) 351 SetPageUptodate(newpage); 352 if (TestClearPageActive(page)) { 353 VM_BUG_ON(PageUnevictable(page)); 354 SetPageActive(newpage); 355 } else 356 unevictable_migrate_page(newpage, page); 357 if (PageChecked(page)) 358 SetPageChecked(newpage); 359 if (PageMappedToDisk(page)) 360 SetPageMappedToDisk(newpage); 361 362 if (PageDirty(page)) { 363 clear_page_dirty_for_io(page); 364 /* 365 * Want to mark the page and the radix tree as dirty, and 366 * redo the accounting that clear_page_dirty_for_io undid, 367 * but we can't use set_page_dirty because that function 368 * is actually a signal that all of the page has become dirty. 369 * Wheras only part of our page may be dirty. 370 */ 371 __set_page_dirty_nobuffers(newpage); 372 } 373 374 mlock_migrate_page(newpage, page); 375 376#ifdef CONFIG_SWAP 377 ClearPageSwapCache(page); 378#endif 379 ClearPagePrivate(page); 380 set_page_private(page, 0); 381 /* page->mapping contains a flag for PageAnon() */ 382 anon = PageAnon(page); 383 page->mapping = NULL; 384 385 if (!anon) /* This page was removed from radix-tree. */ 386 mem_cgroup_uncharge_cache_page(page); 387 388 /* 389 * If any waiters have accumulated on the new page then 390 * wake them up. 391 */ 392 if (PageWriteback(newpage)) 393 end_page_writeback(newpage); 394} 395 396/************************************************************ 397 * Migration functions 398 ***********************************************************/ 399 400/* Always fail migration. Used for mappings that are not movable */ 401int fail_migrate_page(struct address_space *mapping, 402 struct page *newpage, struct page *page) 403{ 404 return -EIO; 405} 406EXPORT_SYMBOL(fail_migrate_page); 407 408/* 409 * Common logic to directly migrate a single page suitable for 410 * pages that do not use PagePrivate. 411 * 412 * Pages are locked upon entry and exit. 413 */ 414int migrate_page(struct address_space *mapping, 415 struct page *newpage, struct page *page) 416{ 417 int rc; 418 419 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 420 421 rc = migrate_page_move_mapping(mapping, newpage, page); 422 423 if (rc) 424 return rc; 425 426 migrate_page_copy(newpage, page); 427 return 0; 428} 429EXPORT_SYMBOL(migrate_page); 430 431#ifdef CONFIG_BLOCK 432/* 433 * Migration function for pages with buffers. This function can only be used 434 * if the underlying filesystem guarantees that no other references to "page" 435 * exist. 436 */ 437int buffer_migrate_page(struct address_space *mapping, 438 struct page *newpage, struct page *page) 439{ 440 struct buffer_head *bh, *head; 441 int rc; 442 443 if (!page_has_buffers(page)) 444 return migrate_page(mapping, newpage, page); 445 446 head = page_buffers(page); 447 448 rc = migrate_page_move_mapping(mapping, newpage, page); 449 450 if (rc) 451 return rc; 452 453 bh = head; 454 do { 455 get_bh(bh); 456 lock_buffer(bh); 457 bh = bh->b_this_page; 458 459 } while (bh != head); 460 461 ClearPagePrivate(page); 462 set_page_private(newpage, page_private(page)); 463 set_page_private(page, 0); 464 put_page(page); 465 get_page(newpage); 466 467 bh = head; 468 do { 469 set_bh_page(bh, newpage, bh_offset(bh)); 470 bh = bh->b_this_page; 471 472 } while (bh != head); 473 474 SetPagePrivate(newpage); 475 476 migrate_page_copy(newpage, page); 477 478 bh = head; 479 do { 480 unlock_buffer(bh); 481 put_bh(bh); 482 bh = bh->b_this_page; 483 484 } while (bh != head); 485 486 return 0; 487} 488EXPORT_SYMBOL(buffer_migrate_page); 489#endif 490 491/* 492 * Writeback a page to clean the dirty state 493 */ 494static int writeout(struct address_space *mapping, struct page *page) 495{ 496 struct writeback_control wbc = { 497 .sync_mode = WB_SYNC_NONE, 498 .nr_to_write = 1, 499 .range_start = 0, 500 .range_end = LLONG_MAX, 501 .nonblocking = 1, 502 .for_reclaim = 1 503 }; 504 int rc; 505 506 if (!mapping->a_ops->writepage) 507 /* No write method for the address space */ 508 return -EINVAL; 509 510 if (!clear_page_dirty_for_io(page)) 511 /* Someone else already triggered a write */ 512 return -EAGAIN; 513 514 /* 515 * A dirty page may imply that the underlying filesystem has 516 * the page on some queue. So the page must be clean for 517 * migration. Writeout may mean we loose the lock and the 518 * page state is no longer what we checked for earlier. 519 * At this point we know that the migration attempt cannot 520 * be successful. 521 */ 522 remove_migration_ptes(page, page); 523 524 rc = mapping->a_ops->writepage(page, &wbc); 525 526 if (rc != AOP_WRITEPAGE_ACTIVATE) 527 /* unlocked. Relock */ 528 lock_page(page); 529 530 return (rc < 0) ? -EIO : -EAGAIN; 531} 532 533/* 534 * Default handling if a filesystem does not provide a migration function. 535 */ 536static int fallback_migrate_page(struct address_space *mapping, 537 struct page *newpage, struct page *page) 538{ 539 if (PageDirty(page)) 540 return writeout(mapping, page); 541 542 /* 543 * Buffers may be managed in a filesystem specific way. 544 * We must have no buffers or drop them. 545 */ 546 if (PagePrivate(page) && 547 !try_to_release_page(page, GFP_KERNEL)) 548 return -EAGAIN; 549 550 return migrate_page(mapping, newpage, page); 551} 552 553/* 554 * Move a page to a newly allocated page 555 * The page is locked and all ptes have been successfully removed. 556 * 557 * The new page will have replaced the old page if this function 558 * is successful. 559 * 560 * Return value: 561 * < 0 - error code 562 * == 0 - success 563 */ 564static int move_to_new_page(struct page *newpage, struct page *page) 565{ 566 struct address_space *mapping; 567 int rc; 568 569 /* 570 * Block others from accessing the page when we get around to 571 * establishing additional references. We are the only one 572 * holding a reference to the new page at this point. 573 */ 574 if (!trylock_page(newpage)) 575 BUG(); 576 577 /* Prepare mapping for the new page.*/ 578 newpage->index = page->index; 579 newpage->mapping = page->mapping; 580 if (PageSwapBacked(page)) 581 SetPageSwapBacked(newpage); 582 583 mapping = page_mapping(page); 584 if (!mapping) 585 rc = migrate_page(mapping, newpage, page); 586 else if (mapping->a_ops->migratepage) 587 /* 588 * Most pages have a mapping and most filesystems 589 * should provide a migration function. Anonymous 590 * pages are part of swap space which also has its 591 * own migration function. This is the most common 592 * path for page migration. 593 */ 594 rc = mapping->a_ops->migratepage(mapping, 595 newpage, page); 596 else 597 rc = fallback_migrate_page(mapping, newpage, page); 598 599 if (!rc) { 600 remove_migration_ptes(page, newpage); 601 } else 602 newpage->mapping = NULL; 603 604 unlock_page(newpage); 605 606 return rc; 607} 608 609/* 610 * Obtain the lock on page, remove all ptes and migrate the page 611 * to the newly allocated page in newpage. 612 */ 613static int unmap_and_move(new_page_t get_new_page, unsigned long private, 614 struct page *page, int force) 615{ 616 int rc = 0; 617 int *result = NULL; 618 struct page *newpage = get_new_page(page, private, &result); 619 int rcu_locked = 0; 620 int charge = 0; 621 622 if (!newpage) 623 return -ENOMEM; 624 625 if (page_count(page) == 1) { 626 /* page was freed from under us. So we are done. */ 627 goto move_newpage; 628 } 629 630 charge = mem_cgroup_prepare_migration(page, newpage); 631 if (charge == -ENOMEM) { 632 rc = -ENOMEM; 633 goto move_newpage; 634 } 635 /* prepare cgroup just returns 0 or -ENOMEM */ 636 BUG_ON(charge); 637 638 rc = -EAGAIN; 639 if (!trylock_page(page)) { 640 if (!force) 641 goto move_newpage; 642 lock_page(page); 643 } 644 645 if (PageWriteback(page)) { 646 if (!force) 647 goto unlock; 648 wait_on_page_writeback(page); 649 } 650 /* 651 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, 652 * we cannot notice that anon_vma is freed while we migrates a page. 653 * This rcu_read_lock() delays freeing anon_vma pointer until the end 654 * of migration. File cache pages are no problem because of page_lock() 655 * File Caches may use write_page() or lock_page() in migration, then, 656 * just care Anon page here. 657 */ 658 if (PageAnon(page)) { 659 rcu_read_lock(); 660 rcu_locked = 1; 661 } 662 663 /* 664 * Corner case handling: 665 * 1. When a new swap-cache page is read into, it is added to the LRU 666 * and treated as swapcache but it has no rmap yet. 667 * Calling try_to_unmap() against a page->mapping==NULL page will 668 * trigger a BUG. So handle it here. 669 * 2. An orphaned page (see truncate_complete_page) might have 670 * fs-private metadata. The page can be picked up due to memory 671 * offlining. Everywhere else except page reclaim, the page is 672 * invisible to the vm, so the page can not be migrated. So try to 673 * free the metadata, so the page can be freed. 674 */ 675 if (!page->mapping) { 676 if (!PageAnon(page) && PagePrivate(page)) { 677 /* 678 * Go direct to try_to_free_buffers() here because 679 * a) that's what try_to_release_page() would do anyway 680 * b) we may be under rcu_read_lock() here, so we can't 681 * use GFP_KERNEL which is what try_to_release_page() 682 * needs to be effective. 683 */ 684 try_to_free_buffers(page); 685 } 686 goto rcu_unlock; 687 } 688 689 /* Establish migration ptes or remove ptes */ 690 try_to_unmap(page, 1); 691 692 if (!page_mapped(page)) 693 rc = move_to_new_page(newpage, page); 694 695 if (rc) 696 remove_migration_ptes(page, page); 697rcu_unlock: 698 if (rcu_locked) 699 rcu_read_unlock(); 700 701unlock: 702 unlock_page(page); 703 704 if (rc != -EAGAIN) { 705 /* 706 * A page that has been migrated has all references 707 * removed and will be freed. A page that has not been 708 * migrated will have kepts its references and be 709 * restored. 710 */ 711 list_del(&page->lru); 712 putback_lru_page(page); 713 } 714 715move_newpage: 716 if (!charge) 717 mem_cgroup_end_migration(newpage); 718 719 /* 720 * Move the new page to the LRU. If migration was not successful 721 * then this will free the page. 722 */ 723 putback_lru_page(newpage); 724 725 if (result) { 726 if (rc) 727 *result = rc; 728 else 729 *result = page_to_nid(newpage); 730 } 731 return rc; 732} 733 734/* 735 * migrate_pages 736 * 737 * The function takes one list of pages to migrate and a function 738 * that determines from the page to be migrated and the private data 739 * the target of the move and allocates the page. 740 * 741 * The function returns after 10 attempts or if no pages 742 * are movable anymore because to has become empty 743 * or no retryable pages exist anymore. All pages will be 744 * returned to the LRU or freed. 745 * 746 * Return: Number of pages not migrated or error code. 747 */ 748int migrate_pages(struct list_head *from, 749 new_page_t get_new_page, unsigned long private) 750{ 751 int retry = 1; 752 int nr_failed = 0; 753 int pass = 0; 754 struct page *page; 755 struct page *page2; 756 int swapwrite = current->flags & PF_SWAPWRITE; 757 int rc; 758 759 if (!swapwrite) 760 current->flags |= PF_SWAPWRITE; 761 762 for(pass = 0; pass < 10 && retry; pass++) { 763 retry = 0; 764 765 list_for_each_entry_safe(page, page2, from, lru) { 766 cond_resched(); 767 768 rc = unmap_and_move(get_new_page, private, 769 page, pass > 2); 770 771 switch(rc) { 772 case -ENOMEM: 773 goto out; 774 case -EAGAIN: 775 retry++; 776 break; 777 case 0: 778 break; 779 default: 780 /* Permanent failure */ 781 nr_failed++; 782 break; 783 } 784 } 785 } 786 rc = 0; 787out: 788 if (!swapwrite) 789 current->flags &= ~PF_SWAPWRITE; 790 791 putback_lru_pages(from); 792 793 if (rc) 794 return rc; 795 796 return nr_failed + retry; 797} 798 799#ifdef CONFIG_NUMA 800/* 801 * Move a list of individual pages 802 */ 803struct page_to_node { 804 unsigned long addr; 805 struct page *page; 806 int node; 807 int status; 808}; 809 810static struct page *new_page_node(struct page *p, unsigned long private, 811 int **result) 812{ 813 struct page_to_node *pm = (struct page_to_node *)private; 814 815 while (pm->node != MAX_NUMNODES && pm->page != p) 816 pm++; 817 818 if (pm->node == MAX_NUMNODES) 819 return NULL; 820 821 *result = &pm->status; 822 823 return alloc_pages_node(pm->node, 824 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); 825} 826 827/* 828 * Move a set of pages as indicated in the pm array. The addr 829 * field must be set to the virtual address of the page to be moved 830 * and the node number must contain a valid target node. 831 * The pm array ends with node = MAX_NUMNODES. 832 */ 833static int do_move_page_to_node_array(struct mm_struct *mm, 834 struct page_to_node *pm, 835 int migrate_all) 836{ 837 int err; 838 struct page_to_node *pp; 839 LIST_HEAD(pagelist); 840 841 migrate_prep(); 842 down_read(&mm->mmap_sem); 843 844 /* 845 * Build a list of pages to migrate 846 */ 847 for (pp = pm; pp->node != MAX_NUMNODES; pp++) { 848 struct vm_area_struct *vma; 849 struct page *page; 850 851 /* 852 * A valid page pointer that will not match any of the 853 * pages that will be moved. 854 */ 855 pp->page = ZERO_PAGE(0); 856 857 err = -EFAULT; 858 vma = find_vma(mm, pp->addr); 859 if (!vma || !vma_migratable(vma)) 860 goto set_status; 861 862 page = follow_page(vma, pp->addr, FOLL_GET); 863 864 err = PTR_ERR(page); 865 if (IS_ERR(page)) 866 goto set_status; 867 868 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); 890 if (!err) 891 list_add_tail(&page->lru, &pagelist); 892put_and_set: 893 /* 894 * Either remove the duplicate refcount from 895 * isolate_lru_page() or drop the page ref if it was 896 * not isolated. 897 */ 898 put_page(page); 899set_status: 900 pp->status = err; 901 } 902 903 err = 0; 904 if (!list_empty(&pagelist)) 905 err = migrate_pages(&pagelist, new_page_node, 906 (unsigned long)pm); 907 908 up_read(&mm->mmap_sem); 909 return err; 910} 911 912/* 913 * Migrate an array of page address onto an array of nodes and fill 914 * the corresponding array of status. 915 */ 916static int do_pages_move(struct mm_struct *mm, struct task_struct *task, 917 unsigned long nr_pages, 918 const void __user * __user *pages, 919 const int __user *nodes, 920 int __user *status, int flags) 921{ 922 struct page_to_node *pm = NULL; 923 nodemask_t task_nodes; 924 int err = 0; 925 int i; 926 927 task_nodes = cpuset_mems_allowed(task); 928 929 /* Limit nr_pages so that the multiplication may not overflow */ 930 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { 931 err = -E2BIG; 932 goto out; 933 } 934 935 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); 936 if (!pm) { 937 err = -ENOMEM; 938 goto out; 939 } 940 941 /* 942 * Get parameters from user space and initialize the pm 943 * array. Return various errors if the user did something wrong. 944 */ 945 for (i = 0; i < nr_pages; i++) { 946 const void __user *p; 947 948 err = -EFAULT; 949 if (get_user(p, pages + i)) 950 goto out_pm; 951 952 pm[i].addr = (unsigned long)p; 953 if (nodes) { 954 int node; 955 956 if (get_user(node, nodes + i)) 957 goto out_pm; 958 959 err = -ENODEV; 960 if (!node_state(node, N_HIGH_MEMORY)) 961 goto out_pm; 962 963 err = -EACCES; 964 if (!node_isset(node, task_nodes)) 965 goto out_pm; 966 967 pm[i].node = node; 968 } else 969 pm[i].node = 0; /* anything to not match MAX_NUMNODES */ 970 } 971 /* End marker */ 972 pm[nr_pages].node = MAX_NUMNODES; 973 974 err = do_move_page_to_node_array(mm, pm, flags & MPOL_MF_MOVE_ALL); 975 if (err >= 0) 976 /* Return status information */ 977 for (i = 0; i < nr_pages; i++) 978 if (put_user(pm[i].status, status + i)) 979 err = -EFAULT; 980 981out_pm: 982 vfree(pm); 983out: 984 return err; 985} 986 987/* 988 * Determine the nodes of an array of pages and store it in an array of status. 989 */ 990static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 991 const void __user * __user *pages, 992 int __user *status) 993{ 994 unsigned long i; 995 int err; 996 997 down_read(&mm->mmap_sem); 998 999 for (i = 0; i < nr_pages; i++) { 1000 const void __user *p; 1001 unsigned long addr; 1002 struct vm_area_struct *vma; 1003 struct page *page; 1004 1005 err = -EFAULT; 1006 if (get_user(p, pages+i)) 1007 goto out; 1008 addr = (unsigned long) p; 1009 1010 vma = find_vma(mm, addr); 1011 if (!vma) 1012 goto set_status; 1013 1014 page = follow_page(vma, addr, 0); 1015 1016 err = PTR_ERR(page); 1017 if (IS_ERR(page)) 1018 goto set_status; 1019 1020 err = -ENOENT; 1021 /* Use PageReserved to check for zero page */ 1022 if (!page || PageReserved(page)) 1023 goto set_status; 1024 1025 err = page_to_nid(page); 1026set_status: 1027 put_user(err, status+i); 1028 } 1029 err = 0; 1030 1031out: 1032 up_read(&mm->mmap_sem); 1033 return err; 1034} 1035 1036/* 1037 * Move a list of pages in the address space of the currently executing 1038 * process. 1039 */ 1040asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, 1041 const void __user * __user *pages, 1042 const int __user *nodes, 1043 int __user *status, int flags) 1044{ 1045 struct task_struct *task; 1046 struct mm_struct *mm; 1047 int err; 1048 1049 /* Check flags */ 1050 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 1051 return -EINVAL; 1052 1053 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1054 return -EPERM; 1055 1056 /* Find the mm_struct */ 1057 read_lock(&tasklist_lock); 1058 task = pid ? find_task_by_vpid(pid) : current; 1059 if (!task) { 1060 read_unlock(&tasklist_lock); 1061 return -ESRCH; 1062 } 1063 mm = get_task_mm(task); 1064 read_unlock(&tasklist_lock); 1065 1066 if (!mm) 1067 return -EINVAL; 1068 1069 /* 1070 * Check if this process has the right to modify the specified 1071 * process. The right exists if the process has administrative 1072 * capabilities, superuser privileges or the same 1073 * userid as the target process. 1074 */ 1075 if ((current->euid != task->suid) && (current->euid != task->uid) && 1076 (current->uid != task->suid) && (current->uid != task->uid) && 1077 !capable(CAP_SYS_NICE)) { 1078 err = -EPERM; 1079 goto out; 1080 } 1081 1082 err = security_task_movememory(task); 1083 if (err) 1084 goto out; 1085 1086 if (nodes) { 1087 err = do_pages_move(mm, task, nr_pages, pages, nodes, status, 1088 flags); 1089 } else { 1090 err = do_pages_stat(mm, nr_pages, pages, status); 1091 } 1092 1093out: 1094 mmput(mm); 1095 return err; 1096} 1097 1098/* 1099 * Call migration functions in the vma_ops that may prepare 1100 * memory in a vm for migration. migration functions may perform 1101 * the migration for vmas that do not have an underlying page struct. 1102 */ 1103int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, 1104 const nodemask_t *from, unsigned long flags) 1105{ 1106 struct vm_area_struct *vma; 1107 int err = 0; 1108 1109 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { 1110 if (vma->vm_ops && vma->vm_ops->migrate) { 1111 err = vma->vm_ops->migrate(vma, to, from, flags); 1112 if (err) 1113 break; 1114 } 1115 } 1116 return err; 1117} 1118#endif 1119