migrate.c revision 7a81b88cb53e335ff7d019e6398c95792c817d93
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_migrate_fixup(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 if (PageSwapCache(page)) { 304 SetPageSwapCache(newpage); 305 set_page_private(newpage, page_private(page)); 306 } 307 308 radix_tree_replace_slot(pslot, newpage); 309 310 page_unfreeze_refs(page, expected_count); 311 /* 312 * Drop cache reference from old page. 313 * We know this isn't the last reference. 314 */ 315 __put_page(page); 316 317 /* 318 * If moved to a different zone then also account 319 * the page for that zone. Other VM counters will be 320 * taken care of when we establish references to the 321 * new page and drop references to the old page. 322 * 323 * Note that anonymous pages are accounted for 324 * via NR_FILE_PAGES and NR_ANON_PAGES if they 325 * are mapped to swap space. 326 */ 327 __dec_zone_page_state(page, NR_FILE_PAGES); 328 __inc_zone_page_state(newpage, NR_FILE_PAGES); 329 330 spin_unlock_irq(&mapping->tree_lock); 331 332 return 0; 333} 334 335/* 336 * Copy the page to its new location 337 */ 338static void migrate_page_copy(struct page *newpage, struct page *page) 339{ 340 int anon; 341 342 copy_highpage(newpage, page); 343 344 if (PageError(page)) 345 SetPageError(newpage); 346 if (PageReferenced(page)) 347 SetPageReferenced(newpage); 348 if (PageUptodate(page)) 349 SetPageUptodate(newpage); 350 if (TestClearPageActive(page)) { 351 VM_BUG_ON(PageUnevictable(page)); 352 SetPageActive(newpage); 353 } else 354 unevictable_migrate_page(newpage, page); 355 if (PageChecked(page)) 356 SetPageChecked(newpage); 357 if (PageMappedToDisk(page)) 358 SetPageMappedToDisk(newpage); 359 360 if (PageDirty(page)) { 361 clear_page_dirty_for_io(page); 362 /* 363 * Want to mark the page and the radix tree as dirty, and 364 * redo the accounting that clear_page_dirty_for_io undid, 365 * but we can't use set_page_dirty because that function 366 * is actually a signal that all of the page has become dirty. 367 * Wheras only part of our page may be dirty. 368 */ 369 __set_page_dirty_nobuffers(newpage); 370 } 371 372 mlock_migrate_page(newpage, page); 373 374 ClearPageSwapCache(page); 375 ClearPagePrivate(page); 376 set_page_private(page, 0); 377 /* page->mapping contains a flag for PageAnon() */ 378 anon = PageAnon(page); 379 page->mapping = NULL; 380 381 if (!anon) /* This page was removed from radix-tree. */ 382 mem_cgroup_uncharge_cache_page(page); 383 384 /* 385 * If any waiters have accumulated on the new page then 386 * wake them up. 387 */ 388 if (PageWriteback(newpage)) 389 end_page_writeback(newpage); 390} 391 392/************************************************************ 393 * Migration functions 394 ***********************************************************/ 395 396/* Always fail migration. Used for mappings that are not movable */ 397int fail_migrate_page(struct address_space *mapping, 398 struct page *newpage, struct page *page) 399{ 400 return -EIO; 401} 402EXPORT_SYMBOL(fail_migrate_page); 403 404/* 405 * Common logic to directly migrate a single page suitable for 406 * pages that do not use PagePrivate. 407 * 408 * Pages are locked upon entry and exit. 409 */ 410int migrate_page(struct address_space *mapping, 411 struct page *newpage, struct page *page) 412{ 413 int rc; 414 415 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 416 417 rc = migrate_page_move_mapping(mapping, newpage, page); 418 419 if (rc) 420 return rc; 421 422 migrate_page_copy(newpage, page); 423 return 0; 424} 425EXPORT_SYMBOL(migrate_page); 426 427#ifdef CONFIG_BLOCK 428/* 429 * Migration function for pages with buffers. This function can only be used 430 * if the underlying filesystem guarantees that no other references to "page" 431 * exist. 432 */ 433int buffer_migrate_page(struct address_space *mapping, 434 struct page *newpage, struct page *page) 435{ 436 struct buffer_head *bh, *head; 437 int rc; 438 439 if (!page_has_buffers(page)) 440 return migrate_page(mapping, newpage, page); 441 442 head = page_buffers(page); 443 444 rc = migrate_page_move_mapping(mapping, newpage, page); 445 446 if (rc) 447 return rc; 448 449 bh = head; 450 do { 451 get_bh(bh); 452 lock_buffer(bh); 453 bh = bh->b_this_page; 454 455 } while (bh != head); 456 457 ClearPagePrivate(page); 458 set_page_private(newpage, page_private(page)); 459 set_page_private(page, 0); 460 put_page(page); 461 get_page(newpage); 462 463 bh = head; 464 do { 465 set_bh_page(bh, newpage, bh_offset(bh)); 466 bh = bh->b_this_page; 467 468 } while (bh != head); 469 470 SetPagePrivate(newpage); 471 472 migrate_page_copy(newpage, page); 473 474 bh = head; 475 do { 476 unlock_buffer(bh); 477 put_bh(bh); 478 bh = bh->b_this_page; 479 480 } while (bh != head); 481 482 return 0; 483} 484EXPORT_SYMBOL(buffer_migrate_page); 485#endif 486 487/* 488 * Writeback a page to clean the dirty state 489 */ 490static int writeout(struct address_space *mapping, struct page *page) 491{ 492 struct writeback_control wbc = { 493 .sync_mode = WB_SYNC_NONE, 494 .nr_to_write = 1, 495 .range_start = 0, 496 .range_end = LLONG_MAX, 497 .nonblocking = 1, 498 .for_reclaim = 1 499 }; 500 int rc; 501 502 if (!mapping->a_ops->writepage) 503 /* No write method for the address space */ 504 return -EINVAL; 505 506 if (!clear_page_dirty_for_io(page)) 507 /* Someone else already triggered a write */ 508 return -EAGAIN; 509 510 /* 511 * A dirty page may imply that the underlying filesystem has 512 * the page on some queue. So the page must be clean for 513 * migration. Writeout may mean we loose the lock and the 514 * page state is no longer what we checked for earlier. 515 * At this point we know that the migration attempt cannot 516 * be successful. 517 */ 518 remove_migration_ptes(page, page); 519 520 rc = mapping->a_ops->writepage(page, &wbc); 521 522 if (rc != AOP_WRITEPAGE_ACTIVATE) 523 /* unlocked. Relock */ 524 lock_page(page); 525 526 return (rc < 0) ? -EIO : -EAGAIN; 527} 528 529/* 530 * Default handling if a filesystem does not provide a migration function. 531 */ 532static int fallback_migrate_page(struct address_space *mapping, 533 struct page *newpage, struct page *page) 534{ 535 if (PageDirty(page)) 536 return writeout(mapping, page); 537 538 /* 539 * Buffers may be managed in a filesystem specific way. 540 * We must have no buffers or drop them. 541 */ 542 if (PagePrivate(page) && 543 !try_to_release_page(page, GFP_KERNEL)) 544 return -EAGAIN; 545 546 return migrate_page(mapping, newpage, page); 547} 548 549/* 550 * Move a page to a newly allocated page 551 * The page is locked and all ptes have been successfully removed. 552 * 553 * The new page will have replaced the old page if this function 554 * is successful. 555 * 556 * Return value: 557 * < 0 - error code 558 * == 0 - success 559 */ 560static int move_to_new_page(struct page *newpage, struct page *page) 561{ 562 struct address_space *mapping; 563 int rc; 564 565 /* 566 * Block others from accessing the page when we get around to 567 * establishing additional references. We are the only one 568 * holding a reference to the new page at this point. 569 */ 570 if (!trylock_page(newpage)) 571 BUG(); 572 573 /* Prepare mapping for the new page.*/ 574 newpage->index = page->index; 575 newpage->mapping = page->mapping; 576 if (PageSwapBacked(page)) 577 SetPageSwapBacked(newpage); 578 579 mapping = page_mapping(page); 580 if (!mapping) 581 rc = migrate_page(mapping, newpage, page); 582 else if (mapping->a_ops->migratepage) 583 /* 584 * Most pages have a mapping and most filesystems 585 * should provide a migration function. Anonymous 586 * pages are part of swap space which also has its 587 * own migration function. This is the most common 588 * path for page migration. 589 */ 590 rc = mapping->a_ops->migratepage(mapping, 591 newpage, page); 592 else 593 rc = fallback_migrate_page(mapping, newpage, page); 594 595 if (!rc) { 596 remove_migration_ptes(page, newpage); 597 } else 598 newpage->mapping = NULL; 599 600 unlock_page(newpage); 601 602 return rc; 603} 604 605/* 606 * Obtain the lock on page, remove all ptes and migrate the page 607 * to the newly allocated page in newpage. 608 */ 609static int unmap_and_move(new_page_t get_new_page, unsigned long private, 610 struct page *page, int force) 611{ 612 int rc = 0; 613 int *result = NULL; 614 struct page *newpage = get_new_page(page, private, &result); 615 int rcu_locked = 0; 616 int charge = 0; 617 618 if (!newpage) 619 return -ENOMEM; 620 621 if (page_count(page) == 1) { 622 /* page was freed from under us. So we are done. */ 623 goto move_newpage; 624 } 625 626 charge = mem_cgroup_prepare_migration(page, newpage); 627 if (charge == -ENOMEM) { 628 rc = -ENOMEM; 629 goto move_newpage; 630 } 631 /* prepare cgroup just returns 0 or -ENOMEM */ 632 BUG_ON(charge); 633 634 rc = -EAGAIN; 635 if (!trylock_page(page)) { 636 if (!force) 637 goto move_newpage; 638 lock_page(page); 639 } 640 641 if (PageWriteback(page)) { 642 if (!force) 643 goto unlock; 644 wait_on_page_writeback(page); 645 } 646 /* 647 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, 648 * we cannot notice that anon_vma is freed while we migrates a page. 649 * This rcu_read_lock() delays freeing anon_vma pointer until the end 650 * of migration. File cache pages are no problem because of page_lock() 651 * File Caches may use write_page() or lock_page() in migration, then, 652 * just care Anon page here. 653 */ 654 if (PageAnon(page)) { 655 rcu_read_lock(); 656 rcu_locked = 1; 657 } 658 659 /* 660 * Corner case handling: 661 * 1. When a new swap-cache page is read into, it is added to the LRU 662 * and treated as swapcache but it has no rmap yet. 663 * Calling try_to_unmap() against a page->mapping==NULL page will 664 * trigger a BUG. So handle it here. 665 * 2. An orphaned page (see truncate_complete_page) might have 666 * fs-private metadata. The page can be picked up due to memory 667 * offlining. Everywhere else except page reclaim, the page is 668 * invisible to the vm, so the page can not be migrated. So try to 669 * free the metadata, so the page can be freed. 670 */ 671 if (!page->mapping) { 672 if (!PageAnon(page) && PagePrivate(page)) { 673 /* 674 * Go direct to try_to_free_buffers() here because 675 * a) that's what try_to_release_page() would do anyway 676 * b) we may be under rcu_read_lock() here, so we can't 677 * use GFP_KERNEL which is what try_to_release_page() 678 * needs to be effective. 679 */ 680 try_to_free_buffers(page); 681 } 682 goto rcu_unlock; 683 } 684 685 /* Establish migration ptes or remove ptes */ 686 try_to_unmap(page, 1); 687 688 if (!page_mapped(page)) 689 rc = move_to_new_page(newpage, page); 690 691 if (rc) 692 remove_migration_ptes(page, page); 693rcu_unlock: 694 if (rcu_locked) 695 rcu_read_unlock(); 696 697unlock: 698 unlock_page(page); 699 700 if (rc != -EAGAIN) { 701 /* 702 * A page that has been migrated has all references 703 * removed and will be freed. A page that has not been 704 * migrated will have kepts its references and be 705 * restored. 706 */ 707 list_del(&page->lru); 708 putback_lru_page(page); 709 } 710 711move_newpage: 712 if (!charge) 713 mem_cgroup_end_migration(newpage); 714 715 /* 716 * Move the new page to the LRU. If migration was not successful 717 * then this will free the page. 718 */ 719 putback_lru_page(newpage); 720 721 if (result) { 722 if (rc) 723 *result = rc; 724 else 725 *result = page_to_nid(newpage); 726 } 727 return rc; 728} 729 730/* 731 * migrate_pages 732 * 733 * The function takes one list of pages to migrate and a function 734 * that determines from the page to be migrated and the private data 735 * the target of the move and allocates the page. 736 * 737 * The function returns after 10 attempts or if no pages 738 * are movable anymore because to has become empty 739 * or no retryable pages exist anymore. All pages will be 740 * returned to the LRU or freed. 741 * 742 * Return: Number of pages not migrated or error code. 743 */ 744int migrate_pages(struct list_head *from, 745 new_page_t get_new_page, unsigned long private) 746{ 747 int retry = 1; 748 int nr_failed = 0; 749 int pass = 0; 750 struct page *page; 751 struct page *page2; 752 int swapwrite = current->flags & PF_SWAPWRITE; 753 int rc; 754 755 if (!swapwrite) 756 current->flags |= PF_SWAPWRITE; 757 758 for(pass = 0; pass < 10 && retry; pass++) { 759 retry = 0; 760 761 list_for_each_entry_safe(page, page2, from, lru) { 762 cond_resched(); 763 764 rc = unmap_and_move(get_new_page, private, 765 page, pass > 2); 766 767 switch(rc) { 768 case -ENOMEM: 769 goto out; 770 case -EAGAIN: 771 retry++; 772 break; 773 case 0: 774 break; 775 default: 776 /* Permanent failure */ 777 nr_failed++; 778 break; 779 } 780 } 781 } 782 rc = 0; 783out: 784 if (!swapwrite) 785 current->flags &= ~PF_SWAPWRITE; 786 787 putback_lru_pages(from); 788 789 if (rc) 790 return rc; 791 792 return nr_failed + retry; 793} 794 795#ifdef CONFIG_NUMA 796/* 797 * Move a list of individual pages 798 */ 799struct page_to_node { 800 unsigned long addr; 801 struct page *page; 802 int node; 803 int status; 804}; 805 806static struct page *new_page_node(struct page *p, unsigned long private, 807 int **result) 808{ 809 struct page_to_node *pm = (struct page_to_node *)private; 810 811 while (pm->node != MAX_NUMNODES && pm->page != p) 812 pm++; 813 814 if (pm->node == MAX_NUMNODES) 815 return NULL; 816 817 *result = &pm->status; 818 819 return alloc_pages_node(pm->node, 820 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); 821} 822 823/* 824 * Move a set of pages as indicated in the pm array. The addr 825 * field must be set to the virtual address of the page to be moved 826 * and the node number must contain a valid target node. 827 * The pm array ends with node = MAX_NUMNODES. 828 */ 829static int do_move_page_to_node_array(struct mm_struct *mm, 830 struct page_to_node *pm, 831 int migrate_all) 832{ 833 int err; 834 struct page_to_node *pp; 835 LIST_HEAD(pagelist); 836 837 migrate_prep(); 838 down_read(&mm->mmap_sem); 839 840 /* 841 * Build a list of pages to migrate 842 */ 843 for (pp = pm; pp->node != MAX_NUMNODES; pp++) { 844 struct vm_area_struct *vma; 845 struct page *page; 846 847 err = -EFAULT; 848 vma = find_vma(mm, pp->addr); 849 if (!vma || !vma_migratable(vma)) 850 goto set_status; 851 852 page = follow_page(vma, pp->addr, FOLL_GET); 853 854 err = PTR_ERR(page); 855 if (IS_ERR(page)) 856 goto set_status; 857 858 err = -ENOENT; 859 if (!page) 860 goto set_status; 861 862 if (PageReserved(page)) /* Check for zero page */ 863 goto put_and_set; 864 865 pp->page = page; 866 err = page_to_nid(page); 867 868 if (err == pp->node) 869 /* 870 * Node already in the right place 871 */ 872 goto put_and_set; 873 874 err = -EACCES; 875 if (page_mapcount(page) > 1 && 876 !migrate_all) 877 goto put_and_set; 878 879 err = isolate_lru_page(page); 880 if (!err) 881 list_add_tail(&page->lru, &pagelist); 882put_and_set: 883 /* 884 * Either remove the duplicate refcount from 885 * isolate_lru_page() or drop the page ref if it was 886 * not isolated. 887 */ 888 put_page(page); 889set_status: 890 pp->status = err; 891 } 892 893 err = 0; 894 if (!list_empty(&pagelist)) 895 err = migrate_pages(&pagelist, new_page_node, 896 (unsigned long)pm); 897 898 up_read(&mm->mmap_sem); 899 return err; 900} 901 902/* 903 * Migrate an array of page address onto an array of nodes and fill 904 * the corresponding array of status. 905 */ 906static int do_pages_move(struct mm_struct *mm, struct task_struct *task, 907 unsigned long nr_pages, 908 const void __user * __user *pages, 909 const int __user *nodes, 910 int __user *status, int flags) 911{ 912 struct page_to_node *pm; 913 nodemask_t task_nodes; 914 unsigned long chunk_nr_pages; 915 unsigned long chunk_start; 916 int err; 917 918 task_nodes = cpuset_mems_allowed(task); 919 920 err = -ENOMEM; 921 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); 922 if (!pm) 923 goto out; 924 /* 925 * Store a chunk of page_to_node array in a page, 926 * but keep the last one as a marker 927 */ 928 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; 929 930 for (chunk_start = 0; 931 chunk_start < nr_pages; 932 chunk_start += chunk_nr_pages) { 933 int j; 934 935 if (chunk_start + chunk_nr_pages > nr_pages) 936 chunk_nr_pages = nr_pages - chunk_start; 937 938 /* fill the chunk pm with addrs and nodes from user-space */ 939 for (j = 0; j < chunk_nr_pages; j++) { 940 const void __user *p; 941 int node; 942 943 err = -EFAULT; 944 if (get_user(p, pages + j + chunk_start)) 945 goto out_pm; 946 pm[j].addr = (unsigned long) p; 947 948 if (get_user(node, nodes + j + chunk_start)) 949 goto out_pm; 950 951 err = -ENODEV; 952 if (!node_state(node, N_HIGH_MEMORY)) 953 goto out_pm; 954 955 err = -EACCES; 956 if (!node_isset(node, task_nodes)) 957 goto out_pm; 958 959 pm[j].node = node; 960 } 961 962 /* End marker for this chunk */ 963 pm[chunk_nr_pages].node = MAX_NUMNODES; 964 965 /* Migrate this chunk */ 966 err = do_move_page_to_node_array(mm, pm, 967 flags & MPOL_MF_MOVE_ALL); 968 if (err < 0) 969 goto out_pm; 970 971 /* Return status information */ 972 for (j = 0; j < chunk_nr_pages; j++) 973 if (put_user(pm[j].status, status + j + chunk_start)) { 974 err = -EFAULT; 975 goto out_pm; 976 } 977 } 978 err = 0; 979 980out_pm: 981 free_page((unsigned long)pm); 982out: 983 return err; 984} 985 986/* 987 * Determine the nodes of an array of pages and store it in an array of status. 988 */ 989static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 990 const void __user **pages, int *status) 991{ 992 unsigned long i; 993 994 down_read(&mm->mmap_sem); 995 996 for (i = 0; i < nr_pages; i++) { 997 unsigned long addr = (unsigned long)(*pages); 998 struct vm_area_struct *vma; 999 struct page *page; 1000 int err = -EFAULT; 1001 1002 vma = find_vma(mm, addr); 1003 if (!vma) 1004 goto set_status; 1005 1006 page = follow_page(vma, addr, 0); 1007 1008 err = PTR_ERR(page); 1009 if (IS_ERR(page)) 1010 goto set_status; 1011 1012 err = -ENOENT; 1013 /* Use PageReserved to check for zero page */ 1014 if (!page || PageReserved(page)) 1015 goto set_status; 1016 1017 err = page_to_nid(page); 1018set_status: 1019 *status = err; 1020 1021 pages++; 1022 status++; 1023 } 1024 1025 up_read(&mm->mmap_sem); 1026} 1027 1028/* 1029 * Determine the nodes of a user array of pages and store it in 1030 * a user array of status. 1031 */ 1032static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 1033 const void __user * __user *pages, 1034 int __user *status) 1035{ 1036#define DO_PAGES_STAT_CHUNK_NR 16 1037 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 1038 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 1039 unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR; 1040 int err; 1041 1042 for (i = 0; i < nr_pages; i += chunk_nr) { 1043 if (chunk_nr + i > nr_pages) 1044 chunk_nr = nr_pages - i; 1045 1046 err = copy_from_user(chunk_pages, &pages[i], 1047 chunk_nr * sizeof(*chunk_pages)); 1048 if (err) { 1049 err = -EFAULT; 1050 goto out; 1051 } 1052 1053 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 1054 1055 err = copy_to_user(&status[i], chunk_status, 1056 chunk_nr * sizeof(*chunk_status)); 1057 if (err) { 1058 err = -EFAULT; 1059 goto out; 1060 } 1061 } 1062 err = 0; 1063 1064out: 1065 return err; 1066} 1067 1068/* 1069 * Move a list of pages in the address space of the currently executing 1070 * process. 1071 */ 1072asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, 1073 const void __user * __user *pages, 1074 const int __user *nodes, 1075 int __user *status, int flags) 1076{ 1077 const struct cred *cred = current_cred(), *tcred; 1078 struct task_struct *task; 1079 struct mm_struct *mm; 1080 int err; 1081 1082 /* Check flags */ 1083 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 1084 return -EINVAL; 1085 1086 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1087 return -EPERM; 1088 1089 /* Find the mm_struct */ 1090 read_lock(&tasklist_lock); 1091 task = pid ? find_task_by_vpid(pid) : current; 1092 if (!task) { 1093 read_unlock(&tasklist_lock); 1094 return -ESRCH; 1095 } 1096 mm = get_task_mm(task); 1097 read_unlock(&tasklist_lock); 1098 1099 if (!mm) 1100 return -EINVAL; 1101 1102 /* 1103 * Check if this process has the right to modify the specified 1104 * process. The right exists if the process has administrative 1105 * capabilities, superuser privileges or the same 1106 * userid as the target process. 1107 */ 1108 rcu_read_lock(); 1109 tcred = __task_cred(task); 1110 if (cred->euid != tcred->suid && cred->euid != tcred->uid && 1111 cred->uid != tcred->suid && cred->uid != tcred->uid && 1112 !capable(CAP_SYS_NICE)) { 1113 rcu_read_unlock(); 1114 err = -EPERM; 1115 goto out; 1116 } 1117 rcu_read_unlock(); 1118 1119 err = security_task_movememory(task); 1120 if (err) 1121 goto out; 1122 1123 if (nodes) { 1124 err = do_pages_move(mm, task, nr_pages, pages, nodes, status, 1125 flags); 1126 } else { 1127 err = do_pages_stat(mm, nr_pages, pages, status); 1128 } 1129 1130out: 1131 mmput(mm); 1132 return err; 1133} 1134 1135/* 1136 * Call migration functions in the vma_ops that may prepare 1137 * memory in a vm for migration. migration functions may perform 1138 * the migration for vmas that do not have an underlying page struct. 1139 */ 1140int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, 1141 const nodemask_t *from, unsigned long flags) 1142{ 1143 struct vm_area_struct *vma; 1144 int err = 0; 1145 1146 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { 1147 if (vma->vm_ops && vma->vm_ops->migrate) { 1148 err = vma->vm_ops->migrate(vma, to, from, flags); 1149 if (err) 1150 break; 1151 } 1152 } 1153 return err; 1154} 1155#endif 1156