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