1 /* 2 * Memory Migration functionality - linux/mm/migrate.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/export.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/ksm.h> 25 #include <linux/rmap.h> 26 #include <linux/topology.h> 27 #include <linux/cpu.h> 28 #include <linux/cpuset.h> 29 #include <linux/writeback.h> 30 #include <linux/mempolicy.h> 31 #include <linux/vmalloc.h> 32 #include <linux/security.h> 33 #include <linux/backing-dev.h> 34 #include <linux/compaction.h> 35 #include <linux/syscalls.h> 36 #include <linux/hugetlb.h> 37 #include <linux/hugetlb_cgroup.h> 38 #include <linux/gfp.h> 39 #include <linux/balloon_compaction.h> 40 #include <linux/mmu_notifier.h> 41 #include <linux/page_idle.h> 42 #include <linux/page_owner.h> 43 44 #include <asm/tlbflush.h> 45 46 #define CREATE_TRACE_POINTS 47 #include <trace/events/migrate.h> 48 49 #include "internal.h" 50 51 /* 52 * migrate_prep() needs to be called before we start compiling a list of pages 53 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is 54 * undesirable, use migrate_prep_local() 55 */ 56 int migrate_prep(void) 57 { 58 /* 59 * Clear the LRU lists so pages can be isolated. 60 * Note that pages may be moved off the LRU after we have 61 * drained them. Those pages will fail to migrate like other 62 * pages that may be busy. 63 */ 64 lru_add_drain_all(); 65 66 return 0; 67 } 68 69 /* Do the necessary work of migrate_prep but not if it involves other CPUs */ 70 int migrate_prep_local(void) 71 { 72 lru_add_drain(); 73 74 return 0; 75 } 76 77 bool isolate_movable_page(struct page *page, isolate_mode_t mode) 78 { 79 struct address_space *mapping; 80 81 /* 82 * Avoid burning cycles with pages that are yet under __free_pages(), 83 * or just got freed under us. 84 * 85 * In case we 'win' a race for a movable page being freed under us and 86 * raise its refcount preventing __free_pages() from doing its job 87 * the put_page() at the end of this block will take care of 88 * release this page, thus avoiding a nasty leakage. 89 */ 90 if (unlikely(!get_page_unless_zero(page))) 91 goto out; 92 93 /* 94 * Check PageMovable before holding a PG_lock because page's owner 95 * assumes anybody doesn't touch PG_lock of newly allocated page 96 * so unconditionally grapping the lock ruins page's owner side. 97 */ 98 if (unlikely(!__PageMovable(page))) 99 goto out_putpage; 100 /* 101 * As movable pages are not isolated from LRU lists, concurrent 102 * compaction threads can race against page migration functions 103 * as well as race against the releasing a page. 104 * 105 * In order to avoid having an already isolated movable page 106 * being (wrongly) re-isolated while it is under migration, 107 * or to avoid attempting to isolate pages being released, 108 * lets be sure we have the page lock 109 * before proceeding with the movable page isolation steps. 110 */ 111 if (unlikely(!trylock_page(page))) 112 goto out_putpage; 113 114 if (!PageMovable(page) || PageIsolated(page)) 115 goto out_no_isolated; 116 117 mapping = page_mapping(page); 118 VM_BUG_ON_PAGE(!mapping, page); 119 120 if (!mapping->a_ops->isolate_page(page, mode)) 121 goto out_no_isolated; 122 123 /* Driver shouldn't use PG_isolated bit of page->flags */ 124 WARN_ON_ONCE(PageIsolated(page)); 125 __SetPageIsolated(page); 126 unlock_page(page); 127 128 return true; 129 130 out_no_isolated: 131 unlock_page(page); 132 out_putpage: 133 put_page(page); 134 out: 135 return false; 136 } 137 138 /* It should be called on page which is PG_movable */ 139 void putback_movable_page(struct page *page) 140 { 141 struct address_space *mapping; 142 143 VM_BUG_ON_PAGE(!PageLocked(page), page); 144 VM_BUG_ON_PAGE(!PageMovable(page), page); 145 VM_BUG_ON_PAGE(!PageIsolated(page), page); 146 147 mapping = page_mapping(page); 148 mapping->a_ops->putback_page(page); 149 __ClearPageIsolated(page); 150 } 151 152 /* 153 * Put previously isolated pages back onto the appropriate lists 154 * from where they were once taken off for compaction/migration. 155 * 156 * This function shall be used whenever the isolated pageset has been 157 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() 158 * and isolate_huge_page(). 159 */ 160 void putback_movable_pages(struct list_head *l) 161 { 162 struct page *page; 163 struct page *page2; 164 165 list_for_each_entry_safe(page, page2, l, lru) { 166 if (unlikely(PageHuge(page))) { 167 putback_active_hugepage(page); 168 continue; 169 } 170 list_del(&page->lru); 171 /* 172 * We isolated non-lru movable page so here we can use 173 * __PageMovable because LRU page's mapping cannot have 174 * PAGE_MAPPING_MOVABLE. 175 */ 176 if (unlikely(__PageMovable(page))) { 177 VM_BUG_ON_PAGE(!PageIsolated(page), page); 178 lock_page(page); 179 if (PageMovable(page)) 180 putback_movable_page(page); 181 else 182 __ClearPageIsolated(page); 183 unlock_page(page); 184 put_page(page); 185 } else { 186 putback_lru_page(page); 187 dec_node_page_state(page, NR_ISOLATED_ANON + 188 page_is_file_cache(page)); 189 } 190 } 191 } 192 193 /* 194 * Restore a potential migration pte to a working pte entry 195 */ 196 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, 197 unsigned long addr, void *old) 198 { 199 struct mm_struct *mm = vma->vm_mm; 200 swp_entry_t entry; 201 pmd_t *pmd; 202 pte_t *ptep, pte; 203 spinlock_t *ptl; 204 205 if (unlikely(PageHuge(new))) { 206 ptep = huge_pte_offset(mm, addr); 207 if (!ptep) 208 goto out; 209 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep); 210 } else { 211 pmd = mm_find_pmd(mm, addr); 212 if (!pmd) 213 goto out; 214 215 ptep = pte_offset_map(pmd, addr); 216 217 /* 218 * Peek to check is_swap_pte() before taking ptlock? No, we 219 * can race mremap's move_ptes(), which skips anon_vma lock. 220 */ 221 222 ptl = pte_lockptr(mm, pmd); 223 } 224 225 spin_lock(ptl); 226 pte = *ptep; 227 if (!is_swap_pte(pte)) 228 goto unlock; 229 230 entry = pte_to_swp_entry(pte); 231 232 if (!is_migration_entry(entry) || 233 migration_entry_to_page(entry) != old) 234 goto unlock; 235 236 get_page(new); 237 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); 238 if (pte_swp_soft_dirty(*ptep)) 239 pte = pte_mksoft_dirty(pte); 240 241 /* Recheck VMA as permissions can change since migration started */ 242 if (is_write_migration_entry(entry)) 243 pte = maybe_mkwrite(pte, vma); 244 245 #ifdef CONFIG_HUGETLB_PAGE 246 if (PageHuge(new)) { 247 pte = pte_mkhuge(pte); 248 pte = arch_make_huge_pte(pte, vma, new, 0); 249 } 250 #endif 251 flush_dcache_page(new); 252 set_pte_at(mm, addr, ptep, pte); 253 254 if (PageHuge(new)) { 255 if (PageAnon(new)) 256 hugepage_add_anon_rmap(new, vma, addr); 257 else 258 page_dup_rmap(new, true); 259 } else if (PageAnon(new)) 260 page_add_anon_rmap(new, vma, addr, false); 261 else 262 page_add_file_rmap(new, false); 263 264 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) 265 mlock_vma_page(new); 266 267 /* No need to invalidate - it was non-present before */ 268 update_mmu_cache(vma, addr, ptep); 269 unlock: 270 pte_unmap_unlock(ptep, ptl); 271 out: 272 return SWAP_AGAIN; 273 } 274 275 /* 276 * Get rid of all migration entries and replace them by 277 * references to the indicated page. 278 */ 279 void remove_migration_ptes(struct page *old, struct page *new, bool locked) 280 { 281 struct rmap_walk_control rwc = { 282 .rmap_one = remove_migration_pte, 283 .arg = old, 284 }; 285 286 if (locked) 287 rmap_walk_locked(new, &rwc); 288 else 289 rmap_walk(new, &rwc); 290 } 291 292 /* 293 * Something used the pte of a page under migration. We need to 294 * get to the page and wait until migration is finished. 295 * When we return from this function the fault will be retried. 296 */ 297 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, 298 spinlock_t *ptl) 299 { 300 pte_t pte; 301 swp_entry_t entry; 302 struct page *page; 303 304 spin_lock(ptl); 305 pte = *ptep; 306 if (!is_swap_pte(pte)) 307 goto out; 308 309 entry = pte_to_swp_entry(pte); 310 if (!is_migration_entry(entry)) 311 goto out; 312 313 page = migration_entry_to_page(entry); 314 315 /* 316 * Once radix-tree replacement of page migration started, page_count 317 * *must* be zero. And, we don't want to call wait_on_page_locked() 318 * against a page without get_page(). 319 * So, we use get_page_unless_zero(), here. Even failed, page fault 320 * will occur again. 321 */ 322 if (!get_page_unless_zero(page)) 323 goto out; 324 pte_unmap_unlock(ptep, ptl); 325 wait_on_page_locked(page); 326 put_page(page); 327 return; 328 out: 329 pte_unmap_unlock(ptep, ptl); 330 } 331 332 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 333 unsigned long address) 334 { 335 spinlock_t *ptl = pte_lockptr(mm, pmd); 336 pte_t *ptep = pte_offset_map(pmd, address); 337 __migration_entry_wait(mm, ptep, ptl); 338 } 339 340 void migration_entry_wait_huge(struct vm_area_struct *vma, 341 struct mm_struct *mm, pte_t *pte) 342 { 343 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); 344 __migration_entry_wait(mm, pte, ptl); 345 } 346 347 #ifdef CONFIG_BLOCK 348 /* Returns true if all buffers are successfully locked */ 349 static bool buffer_migrate_lock_buffers(struct buffer_head *head, 350 enum migrate_mode mode) 351 { 352 struct buffer_head *bh = head; 353 354 /* Simple case, sync compaction */ 355 if (mode != MIGRATE_ASYNC) { 356 do { 357 get_bh(bh); 358 lock_buffer(bh); 359 bh = bh->b_this_page; 360 361 } while (bh != head); 362 363 return true; 364 } 365 366 /* async case, we cannot block on lock_buffer so use trylock_buffer */ 367 do { 368 get_bh(bh); 369 if (!trylock_buffer(bh)) { 370 /* 371 * We failed to lock the buffer and cannot stall in 372 * async migration. Release the taken locks 373 */ 374 struct buffer_head *failed_bh = bh; 375 put_bh(failed_bh); 376 bh = head; 377 while (bh != failed_bh) { 378 unlock_buffer(bh); 379 put_bh(bh); 380 bh = bh->b_this_page; 381 } 382 return false; 383 } 384 385 bh = bh->b_this_page; 386 } while (bh != head); 387 return true; 388 } 389 #else 390 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, 391 enum migrate_mode mode) 392 { 393 return true; 394 } 395 #endif /* CONFIG_BLOCK */ 396 397 /* 398 * Replace the page in the mapping. 399 * 400 * The number of remaining references must be: 401 * 1 for anonymous pages without a mapping 402 * 2 for pages with a mapping 403 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. 404 */ 405 int migrate_page_move_mapping(struct address_space *mapping, 406 struct page *newpage, struct page *page, 407 struct buffer_head *head, enum migrate_mode mode, 408 int extra_count) 409 { 410 struct zone *oldzone, *newzone; 411 int dirty; 412 int expected_count = 1 + extra_count; 413 void **pslot; 414 415 if (!mapping) { 416 /* Anonymous page without mapping */ 417 if (page_count(page) != expected_count) 418 return -EAGAIN; 419 420 /* No turning back from here */ 421 newpage->index = page->index; 422 newpage->mapping = page->mapping; 423 if (PageSwapBacked(page)) 424 __SetPageSwapBacked(newpage); 425 426 return MIGRATEPAGE_SUCCESS; 427 } 428 429 oldzone = page_zone(page); 430 newzone = page_zone(newpage); 431 432 spin_lock_irq(&mapping->tree_lock); 433 434 pslot = radix_tree_lookup_slot(&mapping->page_tree, 435 page_index(page)); 436 437 expected_count += 1 + page_has_private(page); 438 if (page_count(page) != expected_count || 439 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { 440 spin_unlock_irq(&mapping->tree_lock); 441 return -EAGAIN; 442 } 443 444 if (!page_ref_freeze(page, expected_count)) { 445 spin_unlock_irq(&mapping->tree_lock); 446 return -EAGAIN; 447 } 448 449 /* 450 * In the async migration case of moving a page with buffers, lock the 451 * buffers using trylock before the mapping is moved. If the mapping 452 * was moved, we later failed to lock the buffers and could not move 453 * the mapping back due to an elevated page count, we would have to 454 * block waiting on other references to be dropped. 455 */ 456 if (mode == MIGRATE_ASYNC && head && 457 !buffer_migrate_lock_buffers(head, mode)) { 458 page_ref_unfreeze(page, expected_count); 459 spin_unlock_irq(&mapping->tree_lock); 460 return -EAGAIN; 461 } 462 463 /* 464 * Now we know that no one else is looking at the page: 465 * no turning back from here. 466 */ 467 newpage->index = page->index; 468 newpage->mapping = page->mapping; 469 get_page(newpage); /* add cache reference */ 470 if (PageSwapBacked(page)) { 471 __SetPageSwapBacked(newpage); 472 if (PageSwapCache(page)) { 473 SetPageSwapCache(newpage); 474 set_page_private(newpage, page_private(page)); 475 } 476 } else { 477 VM_BUG_ON_PAGE(PageSwapCache(page), page); 478 } 479 480 /* Move dirty while page refs frozen and newpage not yet exposed */ 481 dirty = PageDirty(page); 482 if (dirty) { 483 ClearPageDirty(page); 484 SetPageDirty(newpage); 485 } 486 487 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); 488 489 /* 490 * Drop cache reference from old page by unfreezing 491 * to one less reference. 492 * We know this isn't the last reference. 493 */ 494 page_ref_unfreeze(page, expected_count - 1); 495 496 spin_unlock(&mapping->tree_lock); 497 /* Leave irq disabled to prevent preemption while updating stats */ 498 499 /* 500 * If moved to a different zone then also account 501 * the page for that zone. Other VM counters will be 502 * taken care of when we establish references to the 503 * new page and drop references to the old page. 504 * 505 * Note that anonymous pages are accounted for 506 * via NR_FILE_PAGES and NR_ANON_MAPPED if they 507 * are mapped to swap space. 508 */ 509 if (newzone != oldzone) { 510 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES); 511 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES); 512 if (PageSwapBacked(page) && !PageSwapCache(page)) { 513 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM); 514 __inc_node_state(newzone->zone_pgdat, NR_SHMEM); 515 } 516 if (dirty && mapping_cap_account_dirty(mapping)) { 517 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); 518 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); 519 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); 520 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); 521 } 522 } 523 local_irq_enable(); 524 525 return MIGRATEPAGE_SUCCESS; 526 } 527 EXPORT_SYMBOL(migrate_page_move_mapping); 528 529 /* 530 * The expected number of remaining references is the same as that 531 * of migrate_page_move_mapping(). 532 */ 533 int migrate_huge_page_move_mapping(struct address_space *mapping, 534 struct page *newpage, struct page *page) 535 { 536 int expected_count; 537 void **pslot; 538 539 spin_lock_irq(&mapping->tree_lock); 540 541 pslot = radix_tree_lookup_slot(&mapping->page_tree, 542 page_index(page)); 543 544 expected_count = 2 + page_has_private(page); 545 if (page_count(page) != expected_count || 546 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { 547 spin_unlock_irq(&mapping->tree_lock); 548 return -EAGAIN; 549 } 550 551 if (!page_ref_freeze(page, expected_count)) { 552 spin_unlock_irq(&mapping->tree_lock); 553 return -EAGAIN; 554 } 555 556 newpage->index = page->index; 557 newpage->mapping = page->mapping; 558 559 get_page(newpage); 560 561 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage); 562 563 page_ref_unfreeze(page, expected_count - 1); 564 565 spin_unlock_irq(&mapping->tree_lock); 566 567 return MIGRATEPAGE_SUCCESS; 568 } 569 570 /* 571 * Gigantic pages are so large that we do not guarantee that page++ pointer 572 * arithmetic will work across the entire page. We need something more 573 * specialized. 574 */ 575 static void __copy_gigantic_page(struct page *dst, struct page *src, 576 int nr_pages) 577 { 578 int i; 579 struct page *dst_base = dst; 580 struct page *src_base = src; 581 582 for (i = 0; i < nr_pages; ) { 583 cond_resched(); 584 copy_highpage(dst, src); 585 586 i++; 587 dst = mem_map_next(dst, dst_base, i); 588 src = mem_map_next(src, src_base, i); 589 } 590 } 591 592 static void copy_huge_page(struct page *dst, struct page *src) 593 { 594 int i; 595 int nr_pages; 596 597 if (PageHuge(src)) { 598 /* hugetlbfs page */ 599 struct hstate *h = page_hstate(src); 600 nr_pages = pages_per_huge_page(h); 601 602 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { 603 __copy_gigantic_page(dst, src, nr_pages); 604 return; 605 } 606 } else { 607 /* thp page */ 608 BUG_ON(!PageTransHuge(src)); 609 nr_pages = hpage_nr_pages(src); 610 } 611 612 for (i = 0; i < nr_pages; i++) { 613 cond_resched(); 614 copy_highpage(dst + i, src + i); 615 } 616 } 617 618 /* 619 * Copy the page to its new location 620 */ 621 void migrate_page_copy(struct page *newpage, struct page *page) 622 { 623 int cpupid; 624 625 if (PageHuge(page) || PageTransHuge(page)) 626 copy_huge_page(newpage, page); 627 else 628 copy_highpage(newpage, page); 629 630 if (PageError(page)) 631 SetPageError(newpage); 632 if (PageReferenced(page)) 633 SetPageReferenced(newpage); 634 if (PageUptodate(page)) 635 SetPageUptodate(newpage); 636 if (TestClearPageActive(page)) { 637 VM_BUG_ON_PAGE(PageUnevictable(page), page); 638 SetPageActive(newpage); 639 } else if (TestClearPageUnevictable(page)) 640 SetPageUnevictable(newpage); 641 if (PageChecked(page)) 642 SetPageChecked(newpage); 643 if (PageMappedToDisk(page)) 644 SetPageMappedToDisk(newpage); 645 646 /* Move dirty on pages not done by migrate_page_move_mapping() */ 647 if (PageDirty(page)) 648 SetPageDirty(newpage); 649 650 if (page_is_young(page)) 651 set_page_young(newpage); 652 if (page_is_idle(page)) 653 set_page_idle(newpage); 654 655 /* 656 * Copy NUMA information to the new page, to prevent over-eager 657 * future migrations of this same page. 658 */ 659 cpupid = page_cpupid_xchg_last(page, -1); 660 page_cpupid_xchg_last(newpage, cpupid); 661 662 ksm_migrate_page(newpage, page); 663 /* 664 * Please do not reorder this without considering how mm/ksm.c's 665 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 666 */ 667 if (PageSwapCache(page)) 668 ClearPageSwapCache(page); 669 ClearPagePrivate(page); 670 set_page_private(page, 0); 671 672 /* 673 * If any waiters have accumulated on the new page then 674 * wake them up. 675 */ 676 if (PageWriteback(newpage)) 677 end_page_writeback(newpage); 678 679 copy_page_owner(page, newpage); 680 681 mem_cgroup_migrate(page, newpage); 682 } 683 EXPORT_SYMBOL(migrate_page_copy); 684 685 /************************************************************ 686 * Migration functions 687 ***********************************************************/ 688 689 /* 690 * Common logic to directly migrate a single LRU page suitable for 691 * pages that do not use PagePrivate/PagePrivate2. 692 * 693 * Pages are locked upon entry and exit. 694 */ 695 int migrate_page(struct address_space *mapping, 696 struct page *newpage, struct page *page, 697 enum migrate_mode mode) 698 { 699 int rc; 700 701 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 702 703 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0); 704 705 if (rc != MIGRATEPAGE_SUCCESS) 706 return rc; 707 708 migrate_page_copy(newpage, page); 709 return MIGRATEPAGE_SUCCESS; 710 } 711 EXPORT_SYMBOL(migrate_page); 712 713 #ifdef CONFIG_BLOCK 714 /* 715 * Migration function for pages with buffers. This function can only be used 716 * if the underlying filesystem guarantees that no other references to "page" 717 * exist. 718 */ 719 int buffer_migrate_page(struct address_space *mapping, 720 struct page *newpage, struct page *page, enum migrate_mode mode) 721 { 722 struct buffer_head *bh, *head; 723 int rc; 724 725 if (!page_has_buffers(page)) 726 return migrate_page(mapping, newpage, page, mode); 727 728 head = page_buffers(page); 729 730 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0); 731 732 if (rc != MIGRATEPAGE_SUCCESS) 733 return rc; 734 735 /* 736 * In the async case, migrate_page_move_mapping locked the buffers 737 * with an IRQ-safe spinlock held. In the sync case, the buffers 738 * need to be locked now 739 */ 740 if (mode != MIGRATE_ASYNC) 741 BUG_ON(!buffer_migrate_lock_buffers(head, mode)); 742 743 ClearPagePrivate(page); 744 set_page_private(newpage, page_private(page)); 745 set_page_private(page, 0); 746 put_page(page); 747 get_page(newpage); 748 749 bh = head; 750 do { 751 set_bh_page(bh, newpage, bh_offset(bh)); 752 bh = bh->b_this_page; 753 754 } while (bh != head); 755 756 SetPagePrivate(newpage); 757 758 migrate_page_copy(newpage, page); 759 760 bh = head; 761 do { 762 unlock_buffer(bh); 763 put_bh(bh); 764 bh = bh->b_this_page; 765 766 } while (bh != head); 767 768 return MIGRATEPAGE_SUCCESS; 769 } 770 EXPORT_SYMBOL(buffer_migrate_page); 771 #endif 772 773 /* 774 * Writeback a page to clean the dirty state 775 */ 776 static int writeout(struct address_space *mapping, struct page *page) 777 { 778 struct writeback_control wbc = { 779 .sync_mode = WB_SYNC_NONE, 780 .nr_to_write = 1, 781 .range_start = 0, 782 .range_end = LLONG_MAX, 783 .for_reclaim = 1 784 }; 785 int rc; 786 787 if (!mapping->a_ops->writepage) 788 /* No write method for the address space */ 789 return -EINVAL; 790 791 if (!clear_page_dirty_for_io(page)) 792 /* Someone else already triggered a write */ 793 return -EAGAIN; 794 795 /* 796 * A dirty page may imply that the underlying filesystem has 797 * the page on some queue. So the page must be clean for 798 * migration. Writeout may mean we loose the lock and the 799 * page state is no longer what we checked for earlier. 800 * At this point we know that the migration attempt cannot 801 * be successful. 802 */ 803 remove_migration_ptes(page, page, false); 804 805 rc = mapping->a_ops->writepage(page, &wbc); 806 807 if (rc != AOP_WRITEPAGE_ACTIVATE) 808 /* unlocked. Relock */ 809 lock_page(page); 810 811 return (rc < 0) ? -EIO : -EAGAIN; 812 } 813 814 /* 815 * Default handling if a filesystem does not provide a migration function. 816 */ 817 static int fallback_migrate_page(struct address_space *mapping, 818 struct page *newpage, struct page *page, enum migrate_mode mode) 819 { 820 if (PageDirty(page)) { 821 /* Only writeback pages in full synchronous migration */ 822 if (mode != MIGRATE_SYNC) 823 return -EBUSY; 824 return writeout(mapping, page); 825 } 826 827 /* 828 * Buffers may be managed in a filesystem specific way. 829 * We must have no buffers or drop them. 830 */ 831 if (page_has_private(page) && 832 !try_to_release_page(page, GFP_KERNEL)) 833 return -EAGAIN; 834 835 return migrate_page(mapping, newpage, page, mode); 836 } 837 838 /* 839 * Move a page to a newly allocated page 840 * The page is locked and all ptes have been successfully removed. 841 * 842 * The new page will have replaced the old page if this function 843 * is successful. 844 * 845 * Return value: 846 * < 0 - error code 847 * MIGRATEPAGE_SUCCESS - success 848 */ 849 static int move_to_new_page(struct page *newpage, struct page *page, 850 enum migrate_mode mode) 851 { 852 struct address_space *mapping; 853 int rc = -EAGAIN; 854 bool is_lru = !__PageMovable(page); 855 856 VM_BUG_ON_PAGE(!PageLocked(page), page); 857 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); 858 859 mapping = page_mapping(page); 860 861 if (likely(is_lru)) { 862 if (!mapping) 863 rc = migrate_page(mapping, newpage, page, mode); 864 else if (mapping->a_ops->migratepage) 865 /* 866 * Most pages have a mapping and most filesystems 867 * provide a migratepage callback. Anonymous pages 868 * are part of swap space which also has its own 869 * migratepage callback. This is the most common path 870 * for page migration. 871 */ 872 rc = mapping->a_ops->migratepage(mapping, newpage, 873 page, mode); 874 else 875 rc = fallback_migrate_page(mapping, newpage, 876 page, mode); 877 } else { 878 /* 879 * In case of non-lru page, it could be released after 880 * isolation step. In that case, we shouldn't try migration. 881 */ 882 VM_BUG_ON_PAGE(!PageIsolated(page), page); 883 if (!PageMovable(page)) { 884 rc = MIGRATEPAGE_SUCCESS; 885 __ClearPageIsolated(page); 886 goto out; 887 } 888 889 rc = mapping->a_ops->migratepage(mapping, newpage, 890 page, mode); 891 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 892 !PageIsolated(page)); 893 } 894 895 /* 896 * When successful, old pagecache page->mapping must be cleared before 897 * page is freed; but stats require that PageAnon be left as PageAnon. 898 */ 899 if (rc == MIGRATEPAGE_SUCCESS) { 900 if (__PageMovable(page)) { 901 VM_BUG_ON_PAGE(!PageIsolated(page), page); 902 903 /* 904 * We clear PG_movable under page_lock so any compactor 905 * cannot try to migrate this page. 906 */ 907 __ClearPageIsolated(page); 908 } 909 910 /* 911 * Anonymous and movable page->mapping will be cleard by 912 * free_pages_prepare so don't reset it here for keeping 913 * the type to work PageAnon, for example. 914 */ 915 if (!PageMappingFlags(page)) 916 page->mapping = NULL; 917 } 918 out: 919 return rc; 920 } 921 922 static int __unmap_and_move(struct page *page, struct page *newpage, 923 int force, enum migrate_mode mode) 924 { 925 int rc = -EAGAIN; 926 int page_was_mapped = 0; 927 struct anon_vma *anon_vma = NULL; 928 bool is_lru = !__PageMovable(page); 929 930 if (!trylock_page(page)) { 931 if (!force || mode == MIGRATE_ASYNC) 932 goto out; 933 934 /* 935 * It's not safe for direct compaction to call lock_page. 936 * For example, during page readahead pages are added locked 937 * to the LRU. Later, when the IO completes the pages are 938 * marked uptodate and unlocked. However, the queueing 939 * could be merging multiple pages for one bio (e.g. 940 * mpage_readpages). If an allocation happens for the 941 * second or third page, the process can end up locking 942 * the same page twice and deadlocking. Rather than 943 * trying to be clever about what pages can be locked, 944 * avoid the use of lock_page for direct compaction 945 * altogether. 946 */ 947 if (current->flags & PF_MEMALLOC) 948 goto out; 949 950 lock_page(page); 951 } 952 953 if (PageWriteback(page)) { 954 /* 955 * Only in the case of a full synchronous migration is it 956 * necessary to wait for PageWriteback. In the async case, 957 * the retry loop is too short and in the sync-light case, 958 * the overhead of stalling is too much 959 */ 960 if (mode != MIGRATE_SYNC) { 961 rc = -EBUSY; 962 goto out_unlock; 963 } 964 if (!force) 965 goto out_unlock; 966 wait_on_page_writeback(page); 967 } 968 969 /* 970 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, 971 * we cannot notice that anon_vma is freed while we migrates a page. 972 * This get_anon_vma() delays freeing anon_vma pointer until the end 973 * of migration. File cache pages are no problem because of page_lock() 974 * File Caches may use write_page() or lock_page() in migration, then, 975 * just care Anon page here. 976 * 977 * Only page_get_anon_vma() understands the subtleties of 978 * getting a hold on an anon_vma from outside one of its mms. 979 * But if we cannot get anon_vma, then we won't need it anyway, 980 * because that implies that the anon page is no longer mapped 981 * (and cannot be remapped so long as we hold the page lock). 982 */ 983 if (PageAnon(page) && !PageKsm(page)) 984 anon_vma = page_get_anon_vma(page); 985 986 /* 987 * Block others from accessing the new page when we get around to 988 * establishing additional references. We are usually the only one 989 * holding a reference to newpage at this point. We used to have a BUG 990 * here if trylock_page(newpage) fails, but would like to allow for 991 * cases where there might be a race with the previous use of newpage. 992 * This is much like races on refcount of oldpage: just don't BUG(). 993 */ 994 if (unlikely(!trylock_page(newpage))) 995 goto out_unlock; 996 997 if (unlikely(!is_lru)) { 998 rc = move_to_new_page(newpage, page, mode); 999 goto out_unlock_both; 1000 } 1001 1002 /* 1003 * Corner case handling: 1004 * 1. When a new swap-cache page is read into, it is added to the LRU 1005 * and treated as swapcache but it has no rmap yet. 1006 * Calling try_to_unmap() against a page->mapping==NULL page will 1007 * trigger a BUG. So handle it here. 1008 * 2. An orphaned page (see truncate_complete_page) might have 1009 * fs-private metadata. The page can be picked up due to memory 1010 * offlining. Everywhere else except page reclaim, the page is 1011 * invisible to the vm, so the page can not be migrated. So try to 1012 * free the metadata, so the page can be freed. 1013 */ 1014 if (!page->mapping) { 1015 VM_BUG_ON_PAGE(PageAnon(page), page); 1016 if (page_has_private(page)) { 1017 try_to_free_buffers(page); 1018 goto out_unlock_both; 1019 } 1020 } else if (page_mapped(page)) { 1021 /* Establish migration ptes */ 1022 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, 1023 page); 1024 try_to_unmap(page, 1025 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); 1026 page_was_mapped = 1; 1027 } 1028 1029 if (!page_mapped(page)) 1030 rc = move_to_new_page(newpage, page, mode); 1031 1032 if (page_was_mapped) 1033 remove_migration_ptes(page, 1034 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); 1035 1036 out_unlock_both: 1037 unlock_page(newpage); 1038 out_unlock: 1039 /* Drop an anon_vma reference if we took one */ 1040 if (anon_vma) 1041 put_anon_vma(anon_vma); 1042 unlock_page(page); 1043 out: 1044 /* 1045 * If migration is successful, decrease refcount of the newpage 1046 * which will not free the page because new page owner increased 1047 * refcounter. As well, if it is LRU page, add the page to LRU 1048 * list in here. 1049 */ 1050 if (rc == MIGRATEPAGE_SUCCESS) { 1051 if (unlikely(__PageMovable(newpage))) 1052 put_page(newpage); 1053 else 1054 putback_lru_page(newpage); 1055 } 1056 1057 return rc; 1058 } 1059 1060 /* 1061 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work 1062 * around it. 1063 */ 1064 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM) 1065 #define ICE_noinline noinline 1066 #else 1067 #define ICE_noinline 1068 #endif 1069 1070 /* 1071 * Obtain the lock on page, remove all ptes and migrate the page 1072 * to the newly allocated page in newpage. 1073 */ 1074 static ICE_noinline int unmap_and_move(new_page_t get_new_page, 1075 free_page_t put_new_page, 1076 unsigned long private, struct page *page, 1077 int force, enum migrate_mode mode, 1078 enum migrate_reason reason) 1079 { 1080 int rc = MIGRATEPAGE_SUCCESS; 1081 int *result = NULL; 1082 struct page *newpage; 1083 1084 newpage = get_new_page(page, private, &result); 1085 if (!newpage) 1086 return -ENOMEM; 1087 1088 if (page_count(page) == 1) { 1089 /* page was freed from under us. So we are done. */ 1090 ClearPageActive(page); 1091 ClearPageUnevictable(page); 1092 if (unlikely(__PageMovable(page))) { 1093 lock_page(page); 1094 if (!PageMovable(page)) 1095 __ClearPageIsolated(page); 1096 unlock_page(page); 1097 } 1098 if (put_new_page) 1099 put_new_page(newpage, private); 1100 else 1101 put_page(newpage); 1102 goto out; 1103 } 1104 1105 if (unlikely(PageTransHuge(page))) { 1106 lock_page(page); 1107 rc = split_huge_page(page); 1108 unlock_page(page); 1109 if (rc) 1110 goto out; 1111 } 1112 1113 rc = __unmap_and_move(page, newpage, force, mode); 1114 if (rc == MIGRATEPAGE_SUCCESS) 1115 set_page_owner_migrate_reason(newpage, reason); 1116 1117 out: 1118 if (rc != -EAGAIN) { 1119 /* 1120 * A page that has been migrated has all references 1121 * removed and will be freed. A page that has not been 1122 * migrated will have kepts its references and be 1123 * restored. 1124 */ 1125 list_del(&page->lru); 1126 1127 /* 1128 * Compaction can migrate also non-LRU pages which are 1129 * not accounted to NR_ISOLATED_*. They can be recognized 1130 * as __PageMovable 1131 */ 1132 if (likely(!__PageMovable(page))) 1133 dec_node_page_state(page, NR_ISOLATED_ANON + 1134 page_is_file_cache(page)); 1135 } 1136 1137 /* 1138 * If migration is successful, releases reference grabbed during 1139 * isolation. Otherwise, restore the page to right list unless 1140 * we want to retry. 1141 */ 1142 if (rc == MIGRATEPAGE_SUCCESS) { 1143 put_page(page); 1144 if (reason == MR_MEMORY_FAILURE) { 1145 /* 1146 * Set PG_HWPoison on just freed page 1147 * intentionally. Although it's rather weird, 1148 * it's how HWPoison flag works at the moment. 1149 */ 1150 if (!test_set_page_hwpoison(page)) 1151 num_poisoned_pages_inc(); 1152 } 1153 } else { 1154 if (rc != -EAGAIN) { 1155 if (likely(!__PageMovable(page))) { 1156 putback_lru_page(page); 1157 goto put_new; 1158 } 1159 1160 lock_page(page); 1161 if (PageMovable(page)) 1162 putback_movable_page(page); 1163 else 1164 __ClearPageIsolated(page); 1165 unlock_page(page); 1166 put_page(page); 1167 } 1168 put_new: 1169 if (put_new_page) 1170 put_new_page(newpage, private); 1171 else 1172 put_page(newpage); 1173 } 1174 1175 if (result) { 1176 if (rc) 1177 *result = rc; 1178 else 1179 *result = page_to_nid(newpage); 1180 } 1181 return rc; 1182 } 1183 1184 /* 1185 * Counterpart of unmap_and_move_page() for hugepage migration. 1186 * 1187 * This function doesn't wait the completion of hugepage I/O 1188 * because there is no race between I/O and migration for hugepage. 1189 * Note that currently hugepage I/O occurs only in direct I/O 1190 * where no lock is held and PG_writeback is irrelevant, 1191 * and writeback status of all subpages are counted in the reference 1192 * count of the head page (i.e. if all subpages of a 2MB hugepage are 1193 * under direct I/O, the reference of the head page is 512 and a bit more.) 1194 * This means that when we try to migrate hugepage whose subpages are 1195 * doing direct I/O, some references remain after try_to_unmap() and 1196 * hugepage migration fails without data corruption. 1197 * 1198 * There is also no race when direct I/O is issued on the page under migration, 1199 * because then pte is replaced with migration swap entry and direct I/O code 1200 * will wait in the page fault for migration to complete. 1201 */ 1202 static int unmap_and_move_huge_page(new_page_t get_new_page, 1203 free_page_t put_new_page, unsigned long private, 1204 struct page *hpage, int force, 1205 enum migrate_mode mode, int reason) 1206 { 1207 int rc = -EAGAIN; 1208 int *result = NULL; 1209 int page_was_mapped = 0; 1210 struct page *new_hpage; 1211 struct anon_vma *anon_vma = NULL; 1212 1213 /* 1214 * Movability of hugepages depends on architectures and hugepage size. 1215 * This check is necessary because some callers of hugepage migration 1216 * like soft offline and memory hotremove don't walk through page 1217 * tables or check whether the hugepage is pmd-based or not before 1218 * kicking migration. 1219 */ 1220 if (!hugepage_migration_supported(page_hstate(hpage))) { 1221 putback_active_hugepage(hpage); 1222 return -ENOSYS; 1223 } 1224 1225 new_hpage = get_new_page(hpage, private, &result); 1226 if (!new_hpage) 1227 return -ENOMEM; 1228 1229 if (!trylock_page(hpage)) { 1230 if (!force || mode != MIGRATE_SYNC) 1231 goto out; 1232 lock_page(hpage); 1233 } 1234 1235 if (PageAnon(hpage)) 1236 anon_vma = page_get_anon_vma(hpage); 1237 1238 if (unlikely(!trylock_page(new_hpage))) 1239 goto put_anon; 1240 1241 if (page_mapped(hpage)) { 1242 try_to_unmap(hpage, 1243 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); 1244 page_was_mapped = 1; 1245 } 1246 1247 if (!page_mapped(hpage)) 1248 rc = move_to_new_page(new_hpage, hpage, mode); 1249 1250 if (page_was_mapped) 1251 remove_migration_ptes(hpage, 1252 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); 1253 1254 unlock_page(new_hpage); 1255 1256 put_anon: 1257 if (anon_vma) 1258 put_anon_vma(anon_vma); 1259 1260 if (rc == MIGRATEPAGE_SUCCESS) { 1261 hugetlb_cgroup_migrate(hpage, new_hpage); 1262 put_new_page = NULL; 1263 set_page_owner_migrate_reason(new_hpage, reason); 1264 } 1265 1266 unlock_page(hpage); 1267 out: 1268 if (rc != -EAGAIN) 1269 putback_active_hugepage(hpage); 1270 1271 /* 1272 * If migration was not successful and there's a freeing callback, use 1273 * it. Otherwise, put_page() will drop the reference grabbed during 1274 * isolation. 1275 */ 1276 if (put_new_page) 1277 put_new_page(new_hpage, private); 1278 else 1279 putback_active_hugepage(new_hpage); 1280 1281 if (result) { 1282 if (rc) 1283 *result = rc; 1284 else 1285 *result = page_to_nid(new_hpage); 1286 } 1287 return rc; 1288 } 1289 1290 /* 1291 * migrate_pages - migrate the pages specified in a list, to the free pages 1292 * supplied as the target for the page migration 1293 * 1294 * @from: The list of pages to be migrated. 1295 * @get_new_page: The function used to allocate free pages to be used 1296 * as the target of the page migration. 1297 * @put_new_page: The function used to free target pages if migration 1298 * fails, or NULL if no special handling is necessary. 1299 * @private: Private data to be passed on to get_new_page() 1300 * @mode: The migration mode that specifies the constraints for 1301 * page migration, if any. 1302 * @reason: The reason for page migration. 1303 * 1304 * The function returns after 10 attempts or if no pages are movable any more 1305 * because the list has become empty or no retryable pages exist any more. 1306 * The caller should call putback_movable_pages() to return pages to the LRU 1307 * or free list only if ret != 0. 1308 * 1309 * Returns the number of pages that were not migrated, or an error code. 1310 */ 1311 int migrate_pages(struct list_head *from, new_page_t get_new_page, 1312 free_page_t put_new_page, unsigned long private, 1313 enum migrate_mode mode, int reason) 1314 { 1315 int retry = 1; 1316 int nr_failed = 0; 1317 int nr_succeeded = 0; 1318 int pass = 0; 1319 struct page *page; 1320 struct page *page2; 1321 int swapwrite = current->flags & PF_SWAPWRITE; 1322 int rc; 1323 1324 if (!swapwrite) 1325 current->flags |= PF_SWAPWRITE; 1326 1327 for(pass = 0; pass < 10 && retry; pass++) { 1328 retry = 0; 1329 1330 list_for_each_entry_safe(page, page2, from, lru) { 1331 cond_resched(); 1332 1333 if (PageHuge(page)) 1334 rc = unmap_and_move_huge_page(get_new_page, 1335 put_new_page, private, page, 1336 pass > 2, mode, reason); 1337 else 1338 rc = unmap_and_move(get_new_page, put_new_page, 1339 private, page, pass > 2, mode, 1340 reason); 1341 1342 switch(rc) { 1343 case -ENOMEM: 1344 nr_failed++; 1345 goto out; 1346 case -EAGAIN: 1347 retry++; 1348 break; 1349 case MIGRATEPAGE_SUCCESS: 1350 nr_succeeded++; 1351 break; 1352 default: 1353 /* 1354 * Permanent failure (-EBUSY, -ENOSYS, etc.): 1355 * unlike -EAGAIN case, the failed page is 1356 * removed from migration page list and not 1357 * retried in the next outer loop. 1358 */ 1359 nr_failed++; 1360 break; 1361 } 1362 } 1363 } 1364 nr_failed += retry; 1365 rc = nr_failed; 1366 out: 1367 if (nr_succeeded) 1368 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); 1369 if (nr_failed) 1370 count_vm_events(PGMIGRATE_FAIL, nr_failed); 1371 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); 1372 1373 if (!swapwrite) 1374 current->flags &= ~PF_SWAPWRITE; 1375 1376 return rc; 1377 } 1378 1379 #ifdef CONFIG_NUMA 1380 /* 1381 * Move a list of individual pages 1382 */ 1383 struct page_to_node { 1384 unsigned long addr; 1385 struct page *page; 1386 int node; 1387 int status; 1388 }; 1389 1390 static struct page *new_page_node(struct page *p, unsigned long private, 1391 int **result) 1392 { 1393 struct page_to_node *pm = (struct page_to_node *)private; 1394 1395 while (pm->node != MAX_NUMNODES && pm->page != p) 1396 pm++; 1397 1398 if (pm->node == MAX_NUMNODES) 1399 return NULL; 1400 1401 *result = &pm->status; 1402 1403 if (PageHuge(p)) 1404 return alloc_huge_page_node(page_hstate(compound_head(p)), 1405 pm->node); 1406 else 1407 return __alloc_pages_node(pm->node, 1408 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0); 1409 } 1410 1411 /* 1412 * Move a set of pages as indicated in the pm array. The addr 1413 * field must be set to the virtual address of the page to be moved 1414 * and the node number must contain a valid target node. 1415 * The pm array ends with node = MAX_NUMNODES. 1416 */ 1417 static int do_move_page_to_node_array(struct mm_struct *mm, 1418 struct page_to_node *pm, 1419 int migrate_all) 1420 { 1421 int err; 1422 struct page_to_node *pp; 1423 LIST_HEAD(pagelist); 1424 1425 down_read(&mm->mmap_sem); 1426 1427 /* 1428 * Build a list of pages to migrate 1429 */ 1430 for (pp = pm; pp->node != MAX_NUMNODES; pp++) { 1431 struct vm_area_struct *vma; 1432 struct page *page; 1433 1434 err = -EFAULT; 1435 vma = find_vma(mm, pp->addr); 1436 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) 1437 goto set_status; 1438 1439 /* FOLL_DUMP to ignore special (like zero) pages */ 1440 page = follow_page(vma, pp->addr, 1441 FOLL_GET | FOLL_SPLIT | FOLL_DUMP); 1442 1443 err = PTR_ERR(page); 1444 if (IS_ERR(page)) 1445 goto set_status; 1446 1447 err = -ENOENT; 1448 if (!page) 1449 goto set_status; 1450 1451 pp->page = page; 1452 err = page_to_nid(page); 1453 1454 if (err == pp->node) 1455 /* 1456 * Node already in the right place 1457 */ 1458 goto put_and_set; 1459 1460 err = -EACCES; 1461 if (page_mapcount(page) > 1 && 1462 !migrate_all) 1463 goto put_and_set; 1464 1465 if (PageHuge(page)) { 1466 if (PageHead(page)) 1467 isolate_huge_page(page, &pagelist); 1468 goto put_and_set; 1469 } 1470 1471 err = isolate_lru_page(page); 1472 if (!err) { 1473 list_add_tail(&page->lru, &pagelist); 1474 inc_node_page_state(page, NR_ISOLATED_ANON + 1475 page_is_file_cache(page)); 1476 } 1477 put_and_set: 1478 /* 1479 * Either remove the duplicate refcount from 1480 * isolate_lru_page() or drop the page ref if it was 1481 * not isolated. 1482 */ 1483 put_page(page); 1484 set_status: 1485 pp->status = err; 1486 } 1487 1488 err = 0; 1489 if (!list_empty(&pagelist)) { 1490 err = migrate_pages(&pagelist, new_page_node, NULL, 1491 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); 1492 if (err) 1493 putback_movable_pages(&pagelist); 1494 } 1495 1496 up_read(&mm->mmap_sem); 1497 return err; 1498 } 1499 1500 /* 1501 * Migrate an array of page address onto an array of nodes and fill 1502 * the corresponding array of status. 1503 */ 1504 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 1505 unsigned long nr_pages, 1506 const void __user * __user *pages, 1507 const int __user *nodes, 1508 int __user *status, int flags) 1509 { 1510 struct page_to_node *pm; 1511 unsigned long chunk_nr_pages; 1512 unsigned long chunk_start; 1513 int err; 1514 1515 err = -ENOMEM; 1516 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); 1517 if (!pm) 1518 goto out; 1519 1520 migrate_prep(); 1521 1522 /* 1523 * Store a chunk of page_to_node array in a page, 1524 * but keep the last one as a marker 1525 */ 1526 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; 1527 1528 for (chunk_start = 0; 1529 chunk_start < nr_pages; 1530 chunk_start += chunk_nr_pages) { 1531 int j; 1532 1533 if (chunk_start + chunk_nr_pages > nr_pages) 1534 chunk_nr_pages = nr_pages - chunk_start; 1535 1536 /* fill the chunk pm with addrs and nodes from user-space */ 1537 for (j = 0; j < chunk_nr_pages; j++) { 1538 const void __user *p; 1539 int node; 1540 1541 err = -EFAULT; 1542 if (get_user(p, pages + j + chunk_start)) 1543 goto out_pm; 1544 pm[j].addr = (unsigned long) p; 1545 1546 if (get_user(node, nodes + j + chunk_start)) 1547 goto out_pm; 1548 1549 err = -ENODEV; 1550 if (node < 0 || node >= MAX_NUMNODES) 1551 goto out_pm; 1552 1553 if (!node_state(node, N_MEMORY)) 1554 goto out_pm; 1555 1556 err = -EACCES; 1557 if (!node_isset(node, task_nodes)) 1558 goto out_pm; 1559 1560 pm[j].node = node; 1561 } 1562 1563 /* End marker for this chunk */ 1564 pm[chunk_nr_pages].node = MAX_NUMNODES; 1565 1566 /* Migrate this chunk */ 1567 err = do_move_page_to_node_array(mm, pm, 1568 flags & MPOL_MF_MOVE_ALL); 1569 if (err < 0) 1570 goto out_pm; 1571 1572 /* Return status information */ 1573 for (j = 0; j < chunk_nr_pages; j++) 1574 if (put_user(pm[j].status, status + j + chunk_start)) { 1575 err = -EFAULT; 1576 goto out_pm; 1577 } 1578 } 1579 err = 0; 1580 1581 out_pm: 1582 free_page((unsigned long)pm); 1583 out: 1584 return err; 1585 } 1586 1587 /* 1588 * Determine the nodes of an array of pages and store it in an array of status. 1589 */ 1590 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 1591 const void __user **pages, int *status) 1592 { 1593 unsigned long i; 1594 1595 down_read(&mm->mmap_sem); 1596 1597 for (i = 0; i < nr_pages; i++) { 1598 unsigned long addr = (unsigned long)(*pages); 1599 struct vm_area_struct *vma; 1600 struct page *page; 1601 int err = -EFAULT; 1602 1603 vma = find_vma(mm, addr); 1604 if (!vma || addr < vma->vm_start) 1605 goto set_status; 1606 1607 /* FOLL_DUMP to ignore special (like zero) pages */ 1608 page = follow_page(vma, addr, FOLL_DUMP); 1609 1610 err = PTR_ERR(page); 1611 if (IS_ERR(page)) 1612 goto set_status; 1613 1614 err = page ? page_to_nid(page) : -ENOENT; 1615 set_status: 1616 *status = err; 1617 1618 pages++; 1619 status++; 1620 } 1621 1622 up_read(&mm->mmap_sem); 1623 } 1624 1625 /* 1626 * Determine the nodes of a user array of pages and store it in 1627 * a user array of status. 1628 */ 1629 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 1630 const void __user * __user *pages, 1631 int __user *status) 1632 { 1633 #define DO_PAGES_STAT_CHUNK_NR 16 1634 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 1635 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 1636 1637 while (nr_pages) { 1638 unsigned long chunk_nr; 1639 1640 chunk_nr = nr_pages; 1641 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) 1642 chunk_nr = DO_PAGES_STAT_CHUNK_NR; 1643 1644 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) 1645 break; 1646 1647 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 1648 1649 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 1650 break; 1651 1652 pages += chunk_nr; 1653 status += chunk_nr; 1654 nr_pages -= chunk_nr; 1655 } 1656 return nr_pages ? -EFAULT : 0; 1657 } 1658 1659 /* 1660 * Move a list of pages in the address space of the currently executing 1661 * process. 1662 */ 1663 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 1664 const void __user * __user *, pages, 1665 const int __user *, nodes, 1666 int __user *, status, int, flags) 1667 { 1668 const struct cred *cred = current_cred(), *tcred; 1669 struct task_struct *task; 1670 struct mm_struct *mm; 1671 int err; 1672 nodemask_t task_nodes; 1673 1674 /* Check flags */ 1675 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 1676 return -EINVAL; 1677 1678 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1679 return -EPERM; 1680 1681 /* Find the mm_struct */ 1682 rcu_read_lock(); 1683 task = pid ? find_task_by_vpid(pid) : current; 1684 if (!task) { 1685 rcu_read_unlock(); 1686 return -ESRCH; 1687 } 1688 get_task_struct(task); 1689 1690 /* 1691 * Check if this process has the right to modify the specified 1692 * process. The right exists if the process has administrative 1693 * capabilities, superuser privileges or the same 1694 * userid as the target process. 1695 */ 1696 tcred = __task_cred(task); 1697 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && 1698 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && 1699 !capable(CAP_SYS_NICE)) { 1700 rcu_read_unlock(); 1701 err = -EPERM; 1702 goto out; 1703 } 1704 rcu_read_unlock(); 1705 1706 err = security_task_movememory(task); 1707 if (err) 1708 goto out; 1709 1710 task_nodes = cpuset_mems_allowed(task); 1711 mm = get_task_mm(task); 1712 put_task_struct(task); 1713 1714 if (!mm) 1715 return -EINVAL; 1716 1717 if (nodes) 1718 err = do_pages_move(mm, task_nodes, nr_pages, pages, 1719 nodes, status, flags); 1720 else 1721 err = do_pages_stat(mm, nr_pages, pages, status); 1722 1723 mmput(mm); 1724 return err; 1725 1726 out: 1727 put_task_struct(task); 1728 return err; 1729 } 1730 1731 #ifdef CONFIG_NUMA_BALANCING 1732 /* 1733 * Returns true if this is a safe migration target node for misplaced NUMA 1734 * pages. Currently it only checks the watermarks which crude 1735 */ 1736 static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 1737 unsigned long nr_migrate_pages) 1738 { 1739 int z; 1740 1741 if (!pgdat_reclaimable(pgdat)) 1742 return false; 1743 1744 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 1745 struct zone *zone = pgdat->node_zones + z; 1746 1747 if (!populated_zone(zone)) 1748 continue; 1749 1750 /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 1751 if (!zone_watermark_ok(zone, 0, 1752 high_wmark_pages(zone) + 1753 nr_migrate_pages, 1754 0, 0)) 1755 continue; 1756 return true; 1757 } 1758 return false; 1759 } 1760 1761 static struct page *alloc_misplaced_dst_page(struct page *page, 1762 unsigned long data, 1763 int **result) 1764 { 1765 int nid = (int) data; 1766 struct page *newpage; 1767 1768 newpage = __alloc_pages_node(nid, 1769 (GFP_HIGHUSER_MOVABLE | 1770 __GFP_THISNODE | __GFP_NOMEMALLOC | 1771 __GFP_NORETRY | __GFP_NOWARN) & 1772 ~__GFP_RECLAIM, 0); 1773 1774 return newpage; 1775 } 1776 1777 /* 1778 * page migration rate limiting control. 1779 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs 1780 * window of time. Default here says do not migrate more than 1280M per second. 1781 */ 1782 static unsigned int migrate_interval_millisecs __read_mostly = 100; 1783 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); 1784 1785 /* Returns true if the node is migrate rate-limited after the update */ 1786 static bool numamigrate_update_ratelimit(pg_data_t *pgdat, 1787 unsigned long nr_pages) 1788 { 1789 /* 1790 * Rate-limit the amount of data that is being migrated to a node. 1791 * Optimal placement is no good if the memory bus is saturated and 1792 * all the time is being spent migrating! 1793 */ 1794 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { 1795 spin_lock(&pgdat->numabalancing_migrate_lock); 1796 pgdat->numabalancing_migrate_nr_pages = 0; 1797 pgdat->numabalancing_migrate_next_window = jiffies + 1798 msecs_to_jiffies(migrate_interval_millisecs); 1799 spin_unlock(&pgdat->numabalancing_migrate_lock); 1800 } 1801 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) { 1802 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id, 1803 nr_pages); 1804 return true; 1805 } 1806 1807 /* 1808 * This is an unlocked non-atomic update so errors are possible. 1809 * The consequences are failing to migrate when we potentiall should 1810 * have which is not severe enough to warrant locking. If it is ever 1811 * a problem, it can be converted to a per-cpu counter. 1812 */ 1813 pgdat->numabalancing_migrate_nr_pages += nr_pages; 1814 return false; 1815 } 1816 1817 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 1818 { 1819 int page_lru; 1820 1821 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); 1822 1823 /* Avoid migrating to a node that is nearly full */ 1824 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) 1825 return 0; 1826 1827 if (isolate_lru_page(page)) 1828 return 0; 1829 1830 /* 1831 * migrate_misplaced_transhuge_page() skips page migration's usual 1832 * check on page_count(), so we must do it here, now that the page 1833 * has been isolated: a GUP pin, or any other pin, prevents migration. 1834 * The expected page count is 3: 1 for page's mapcount and 1 for the 1835 * caller's pin and 1 for the reference taken by isolate_lru_page(). 1836 */ 1837 if (PageTransHuge(page) && page_count(page) != 3) { 1838 putback_lru_page(page); 1839 return 0; 1840 } 1841 1842 page_lru = page_is_file_cache(page); 1843 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, 1844 hpage_nr_pages(page)); 1845 1846 /* 1847 * Isolating the page has taken another reference, so the 1848 * caller's reference can be safely dropped without the page 1849 * disappearing underneath us during migration. 1850 */ 1851 put_page(page); 1852 return 1; 1853 } 1854 1855 bool pmd_trans_migrating(pmd_t pmd) 1856 { 1857 struct page *page = pmd_page(pmd); 1858 return PageLocked(page); 1859 } 1860 1861 /* 1862 * Attempt to migrate a misplaced page to the specified destination 1863 * node. Caller is expected to have an elevated reference count on 1864 * the page that will be dropped by this function before returning. 1865 */ 1866 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 1867 int node) 1868 { 1869 pg_data_t *pgdat = NODE_DATA(node); 1870 int isolated; 1871 int nr_remaining; 1872 LIST_HEAD(migratepages); 1873 1874 /* 1875 * Don't migrate file pages that are mapped in multiple processes 1876 * with execute permissions as they are probably shared libraries. 1877 */ 1878 if (page_mapcount(page) != 1 && page_is_file_cache(page) && 1879 (vma->vm_flags & VM_EXEC)) 1880 goto out; 1881 1882 /* 1883 * Rate-limit the amount of data that is being migrated to a node. 1884 * Optimal placement is no good if the memory bus is saturated and 1885 * all the time is being spent migrating! 1886 */ 1887 if (numamigrate_update_ratelimit(pgdat, 1)) 1888 goto out; 1889 1890 isolated = numamigrate_isolate_page(pgdat, page); 1891 if (!isolated) 1892 goto out; 1893 1894 list_add(&page->lru, &migratepages); 1895 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, 1896 NULL, node, MIGRATE_ASYNC, 1897 MR_NUMA_MISPLACED); 1898 if (nr_remaining) { 1899 if (!list_empty(&migratepages)) { 1900 list_del(&page->lru); 1901 dec_node_page_state(page, NR_ISOLATED_ANON + 1902 page_is_file_cache(page)); 1903 putback_lru_page(page); 1904 } 1905 isolated = 0; 1906 } else 1907 count_vm_numa_event(NUMA_PAGE_MIGRATE); 1908 BUG_ON(!list_empty(&migratepages)); 1909 return isolated; 1910 1911 out: 1912 put_page(page); 1913 return 0; 1914 } 1915 #endif /* CONFIG_NUMA_BALANCING */ 1916 1917 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) 1918 /* 1919 * Migrates a THP to a given target node. page must be locked and is unlocked 1920 * before returning. 1921 */ 1922 int migrate_misplaced_transhuge_page(struct mm_struct *mm, 1923 struct vm_area_struct *vma, 1924 pmd_t *pmd, pmd_t entry, 1925 unsigned long address, 1926 struct page *page, int node) 1927 { 1928 spinlock_t *ptl; 1929 pg_data_t *pgdat = NODE_DATA(node); 1930 int isolated = 0; 1931 struct page *new_page = NULL; 1932 int page_lru = page_is_file_cache(page); 1933 unsigned long mmun_start = address & HPAGE_PMD_MASK; 1934 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE; 1935 pmd_t orig_entry; 1936 1937 /* 1938 * Rate-limit the amount of data that is being migrated to a node. 1939 * Optimal placement is no good if the memory bus is saturated and 1940 * all the time is being spent migrating! 1941 */ 1942 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) 1943 goto out_dropref; 1944 1945 new_page = alloc_pages_node(node, 1946 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), 1947 HPAGE_PMD_ORDER); 1948 if (!new_page) 1949 goto out_fail; 1950 prep_transhuge_page(new_page); 1951 1952 isolated = numamigrate_isolate_page(pgdat, page); 1953 if (!isolated) { 1954 put_page(new_page); 1955 goto out_fail; 1956 } 1957 /* 1958 * We are not sure a pending tlb flush here is for a huge page 1959 * mapping or not. Hence use the tlb range variant 1960 */ 1961 if (mm_tlb_flush_pending(mm)) 1962 flush_tlb_range(vma, mmun_start, mmun_end); 1963 1964 /* Prepare a page as a migration target */ 1965 __SetPageLocked(new_page); 1966 __SetPageSwapBacked(new_page); 1967 1968 /* anon mapping, we can simply copy page->mapping to the new page: */ 1969 new_page->mapping = page->mapping; 1970 new_page->index = page->index; 1971 migrate_page_copy(new_page, page); 1972 WARN_ON(PageLRU(new_page)); 1973 1974 /* Recheck the target PMD */ 1975 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 1976 ptl = pmd_lock(mm, pmd); 1977 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) { 1978 fail_putback: 1979 spin_unlock(ptl); 1980 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 1981 1982 /* Reverse changes made by migrate_page_copy() */ 1983 if (TestClearPageActive(new_page)) 1984 SetPageActive(page); 1985 if (TestClearPageUnevictable(new_page)) 1986 SetPageUnevictable(page); 1987 1988 unlock_page(new_page); 1989 put_page(new_page); /* Free it */ 1990 1991 /* Retake the callers reference and putback on LRU */ 1992 get_page(page); 1993 putback_lru_page(page); 1994 mod_node_page_state(page_pgdat(page), 1995 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); 1996 1997 goto out_unlock; 1998 } 1999 2000 orig_entry = *pmd; 2001 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 2002 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 2003 2004 /* 2005 * Clear the old entry under pagetable lock and establish the new PTE. 2006 * Any parallel GUP will either observe the old page blocking on the 2007 * page lock, block on the page table lock or observe the new page. 2008 * The SetPageUptodate on the new page and page_add_new_anon_rmap 2009 * guarantee the copy is visible before the pagetable update. 2010 */ 2011 flush_cache_range(vma, mmun_start, mmun_end); 2012 page_add_anon_rmap(new_page, vma, mmun_start, true); 2013 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd); 2014 set_pmd_at(mm, mmun_start, pmd, entry); 2015 update_mmu_cache_pmd(vma, address, &entry); 2016 2017 if (page_count(page) != 2) { 2018 set_pmd_at(mm, mmun_start, pmd, orig_entry); 2019 flush_pmd_tlb_range(vma, mmun_start, mmun_end); 2020 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 2021 update_mmu_cache_pmd(vma, address, &entry); 2022 page_remove_rmap(new_page, true); 2023 goto fail_putback; 2024 } 2025 2026 mlock_migrate_page(new_page, page); 2027 page_remove_rmap(page, true); 2028 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); 2029 2030 spin_unlock(ptl); 2031 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2032 2033 /* Take an "isolate" reference and put new page on the LRU. */ 2034 get_page(new_page); 2035 putback_lru_page(new_page); 2036 2037 unlock_page(new_page); 2038 unlock_page(page); 2039 put_page(page); /* Drop the rmap reference */ 2040 put_page(page); /* Drop the LRU isolation reference */ 2041 2042 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); 2043 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); 2044 2045 mod_node_page_state(page_pgdat(page), 2046 NR_ISOLATED_ANON + page_lru, 2047 -HPAGE_PMD_NR); 2048 return isolated; 2049 2050 out_fail: 2051 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); 2052 out_dropref: 2053 ptl = pmd_lock(mm, pmd); 2054 if (pmd_same(*pmd, entry)) { 2055 entry = pmd_modify(entry, vma->vm_page_prot); 2056 set_pmd_at(mm, mmun_start, pmd, entry); 2057 update_mmu_cache_pmd(vma, address, &entry); 2058 } 2059 spin_unlock(ptl); 2060 2061 out_unlock: 2062 unlock_page(page); 2063 put_page(page); 2064 return 0; 2065 } 2066 #endif /* CONFIG_NUMA_BALANCING */ 2067 2068 #endif /* CONFIG_NUMA */ 2069