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