1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Memory Migration functionality - linux/mm/migrate.c 4 * 5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter 6 * 7 * Page migration was first developed in the context of the memory hotplug 8 * project. The main authors of the migration code are: 9 * 10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> 11 * Hirokazu Takahashi <taka@valinux.co.jp> 12 * Dave Hansen <haveblue@us.ibm.com> 13 * Christoph Lameter 14 */ 15 16 #include <linux/migrate.h> 17 #include <linux/export.h> 18 #include <linux/swap.h> 19 #include <linux/swapops.h> 20 #include <linux/pagemap.h> 21 #include <linux/buffer_head.h> 22 #include <linux/mm_inline.h> 23 #include <linux/nsproxy.h> 24 #include <linux/pagevec.h> 25 #include <linux/ksm.h> 26 #include <linux/rmap.h> 27 #include <linux/topology.h> 28 #include <linux/cpu.h> 29 #include <linux/cpuset.h> 30 #include <linux/writeback.h> 31 #include <linux/mempolicy.h> 32 #include <linux/vmalloc.h> 33 #include <linux/security.h> 34 #include <linux/backing-dev.h> 35 #include <linux/compaction.h> 36 #include <linux/syscalls.h> 37 #include <linux/compat.h> 38 #include <linux/hugetlb.h> 39 #include <linux/hugetlb_cgroup.h> 40 #include <linux/gfp.h> 41 #include <linux/pagewalk.h> 42 #include <linux/pfn_t.h> 43 #include <linux/memremap.h> 44 #include <linux/userfaultfd_k.h> 45 #include <linux/balloon_compaction.h> 46 #include <linux/mmu_notifier.h> 47 #include <linux/page_idle.h> 48 #include <linux/page_owner.h> 49 #include <linux/sched/mm.h> 50 #include <linux/ptrace.h> 51 #include <linux/oom.h> 52 53 #include <asm/tlbflush.h> 54 55 #define CREATE_TRACE_POINTS 56 #include <trace/events/migrate.h> 57 58 #include "internal.h" 59 60 /* 61 * migrate_prep() needs to be called before we start compiling a list of pages 62 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is 63 * undesirable, use migrate_prep_local() 64 */ 65 int migrate_prep(void) 66 { 67 /* 68 * Clear the LRU lists so pages can be isolated. 69 * Note that pages may be moved off the LRU after we have 70 * drained them. Those pages will fail to migrate like other 71 * pages that may be busy. 72 */ 73 lru_add_drain_all(); 74 75 return 0; 76 } 77 78 /* Do the necessary work of migrate_prep but not if it involves other CPUs */ 79 int migrate_prep_local(void) 80 { 81 lru_add_drain(); 82 83 return 0; 84 } 85 86 int isolate_movable_page(struct page *page, isolate_mode_t mode) 87 { 88 struct address_space *mapping; 89 90 /* 91 * Avoid burning cycles with pages that are yet under __free_pages(), 92 * or just got freed under us. 93 * 94 * In case we 'win' a race for a movable page being freed under us and 95 * raise its refcount preventing __free_pages() from doing its job 96 * the put_page() at the end of this block will take care of 97 * release this page, thus avoiding a nasty leakage. 98 */ 99 if (unlikely(!get_page_unless_zero(page))) 100 goto out; 101 102 /* 103 * Check PageMovable before holding a PG_lock because page's owner 104 * assumes anybody doesn't touch PG_lock of newly allocated page 105 * so unconditionally grabbing the lock ruins page's owner side. 106 */ 107 if (unlikely(!__PageMovable(page))) 108 goto out_putpage; 109 /* 110 * As movable pages are not isolated from LRU lists, concurrent 111 * compaction threads can race against page migration functions 112 * as well as race against the releasing a page. 113 * 114 * In order to avoid having an already isolated movable page 115 * being (wrongly) re-isolated while it is under migration, 116 * or to avoid attempting to isolate pages being released, 117 * lets be sure we have the page lock 118 * before proceeding with the movable page isolation steps. 119 */ 120 if (unlikely(!trylock_page(page))) 121 goto out_putpage; 122 123 if (!PageMovable(page) || PageIsolated(page)) 124 goto out_no_isolated; 125 126 mapping = page_mapping(page); 127 VM_BUG_ON_PAGE(!mapping, page); 128 129 if (!mapping->a_ops->isolate_page(page, mode)) 130 goto out_no_isolated; 131 132 /* Driver shouldn't use PG_isolated bit of page->flags */ 133 WARN_ON_ONCE(PageIsolated(page)); 134 __SetPageIsolated(page); 135 unlock_page(page); 136 137 return 0; 138 139 out_no_isolated: 140 unlock_page(page); 141 out_putpage: 142 put_page(page); 143 out: 144 return -EBUSY; 145 } 146 147 /* It should be called on page which is PG_movable */ 148 void putback_movable_page(struct page *page) 149 { 150 struct address_space *mapping; 151 152 VM_BUG_ON_PAGE(!PageLocked(page), page); 153 VM_BUG_ON_PAGE(!PageMovable(page), page); 154 VM_BUG_ON_PAGE(!PageIsolated(page), page); 155 156 mapping = page_mapping(page); 157 mapping->a_ops->putback_page(page); 158 __ClearPageIsolated(page); 159 } 160 161 /* 162 * Put previously isolated pages back onto the appropriate lists 163 * from where they were once taken off for compaction/migration. 164 * 165 * This function shall be used whenever the isolated pageset has been 166 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range() 167 * and isolate_huge_page(). 168 */ 169 void putback_movable_pages(struct list_head *l) 170 { 171 struct page *page; 172 struct page *page2; 173 174 list_for_each_entry_safe(page, page2, l, lru) { 175 if (unlikely(PageHuge(page))) { 176 putback_active_hugepage(page); 177 continue; 178 } 179 list_del(&page->lru); 180 /* 181 * We isolated non-lru movable page so here we can use 182 * __PageMovable because LRU page's mapping cannot have 183 * PAGE_MAPPING_MOVABLE. 184 */ 185 if (unlikely(__PageMovable(page))) { 186 VM_BUG_ON_PAGE(!PageIsolated(page), page); 187 lock_page(page); 188 if (PageMovable(page)) 189 putback_movable_page(page); 190 else 191 __ClearPageIsolated(page); 192 unlock_page(page); 193 put_page(page); 194 } else { 195 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 196 page_is_file_lru(page), -thp_nr_pages(page)); 197 putback_lru_page(page); 198 } 199 } 200 } 201 202 /* 203 * Restore a potential migration pte to a working pte entry 204 */ 205 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma, 206 unsigned long addr, void *old) 207 { 208 struct page_vma_mapped_walk pvmw = { 209 .page = old, 210 .vma = vma, 211 .address = addr, 212 .flags = PVMW_SYNC | PVMW_MIGRATION, 213 }; 214 struct page *new; 215 pte_t pte; 216 swp_entry_t entry; 217 218 VM_BUG_ON_PAGE(PageTail(page), page); 219 while (page_vma_mapped_walk(&pvmw)) { 220 if (PageKsm(page)) 221 new = page; 222 else 223 new = page - pvmw.page->index + 224 linear_page_index(vma, pvmw.address); 225 226 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 227 /* PMD-mapped THP migration entry */ 228 if (!pvmw.pte) { 229 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); 230 remove_migration_pmd(&pvmw, new); 231 continue; 232 } 233 #endif 234 235 get_page(new); 236 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot))); 237 if (pte_swp_soft_dirty(*pvmw.pte)) 238 pte = pte_mksoft_dirty(pte); 239 240 /* 241 * Recheck VMA as permissions can change since migration started 242 */ 243 entry = pte_to_swp_entry(*pvmw.pte); 244 if (is_write_migration_entry(entry)) 245 pte = maybe_mkwrite(pte, vma); 246 else if (pte_swp_uffd_wp(*pvmw.pte)) 247 pte = pte_mkuffd_wp(pte); 248 249 if (unlikely(is_device_private_page(new))) { 250 entry = make_device_private_entry(new, pte_write(pte)); 251 pte = swp_entry_to_pte(entry); 252 if (pte_swp_soft_dirty(*pvmw.pte)) 253 pte = pte_swp_mksoft_dirty(pte); 254 if (pte_swp_uffd_wp(*pvmw.pte)) 255 pte = pte_swp_mkuffd_wp(pte); 256 } 257 258 #ifdef CONFIG_HUGETLB_PAGE 259 if (PageHuge(new)) { 260 pte = pte_mkhuge(pte); 261 pte = arch_make_huge_pte(pte, vma, new, 0); 262 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 263 if (PageAnon(new)) 264 hugepage_add_anon_rmap(new, vma, pvmw.address); 265 else 266 page_dup_rmap(new, true); 267 } else 268 #endif 269 { 270 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte); 271 272 if (PageAnon(new)) 273 page_add_anon_rmap(new, vma, pvmw.address, false); 274 else 275 page_add_file_rmap(new, false); 276 } 277 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new)) 278 mlock_vma_page(new); 279 280 if (PageTransHuge(page) && PageMlocked(page)) 281 clear_page_mlock(page); 282 283 /* No need to invalidate - it was non-present before */ 284 update_mmu_cache(vma, pvmw.address, pvmw.pte); 285 } 286 287 return true; 288 } 289 290 /* 291 * Get rid of all migration entries and replace them by 292 * references to the indicated page. 293 */ 294 void remove_migration_ptes(struct page *old, struct page *new, bool locked) 295 { 296 struct rmap_walk_control rwc = { 297 .rmap_one = remove_migration_pte, 298 .arg = old, 299 }; 300 301 if (locked) 302 rmap_walk_locked(new, &rwc); 303 else 304 rmap_walk(new, &rwc); 305 } 306 307 /* 308 * Something used the pte of a page under migration. We need to 309 * get to the page and wait until migration is finished. 310 * When we return from this function the fault will be retried. 311 */ 312 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, 313 spinlock_t *ptl) 314 { 315 pte_t pte; 316 swp_entry_t entry; 317 struct page *page; 318 319 spin_lock(ptl); 320 pte = *ptep; 321 if (!is_swap_pte(pte)) 322 goto out; 323 324 entry = pte_to_swp_entry(pte); 325 if (!is_migration_entry(entry)) 326 goto out; 327 328 page = migration_entry_to_page(entry); 329 330 /* 331 * Once page cache replacement of page migration started, page_count 332 * is zero; but we must not call put_and_wait_on_page_locked() without 333 * a ref. Use get_page_unless_zero(), and just fault again if it fails. 334 */ 335 if (!get_page_unless_zero(page)) 336 goto out; 337 pte_unmap_unlock(ptep, ptl); 338 put_and_wait_on_page_locked(page); 339 return; 340 out: 341 pte_unmap_unlock(ptep, ptl); 342 } 343 344 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, 345 unsigned long address) 346 { 347 spinlock_t *ptl = pte_lockptr(mm, pmd); 348 pte_t *ptep = pte_offset_map(pmd, address); 349 __migration_entry_wait(mm, ptep, ptl); 350 } 351 352 void migration_entry_wait_huge(struct vm_area_struct *vma, 353 struct mm_struct *mm, pte_t *pte) 354 { 355 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte); 356 __migration_entry_wait(mm, pte, ptl); 357 } 358 359 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 360 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd) 361 { 362 spinlock_t *ptl; 363 struct page *page; 364 365 ptl = pmd_lock(mm, pmd); 366 if (!is_pmd_migration_entry(*pmd)) 367 goto unlock; 368 page = migration_entry_to_page(pmd_to_swp_entry(*pmd)); 369 if (!get_page_unless_zero(page)) 370 goto unlock; 371 spin_unlock(ptl); 372 put_and_wait_on_page_locked(page); 373 return; 374 unlock: 375 spin_unlock(ptl); 376 } 377 #endif 378 379 static int expected_page_refs(struct address_space *mapping, struct page *page) 380 { 381 int expected_count = 1; 382 383 /* 384 * Device private pages have an extra refcount as they are 385 * ZONE_DEVICE pages. 386 */ 387 expected_count += is_device_private_page(page); 388 if (mapping) 389 expected_count += thp_nr_pages(page) + page_has_private(page); 390 391 return expected_count; 392 } 393 394 /* 395 * Replace the page in the mapping. 396 * 397 * The number of remaining references must be: 398 * 1 for anonymous pages without a mapping 399 * 2 for pages with a mapping 400 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. 401 */ 402 int migrate_page_move_mapping(struct address_space *mapping, 403 struct page *newpage, struct page *page, int extra_count) 404 { 405 XA_STATE(xas, &mapping->i_pages, page_index(page)); 406 struct zone *oldzone, *newzone; 407 int dirty; 408 int expected_count = expected_page_refs(mapping, page) + extra_count; 409 410 if (!mapping) { 411 /* Anonymous page without mapping */ 412 if (page_count(page) != expected_count) 413 return -EAGAIN; 414 415 /* No turning back from here */ 416 newpage->index = page->index; 417 newpage->mapping = page->mapping; 418 if (PageSwapBacked(page)) 419 __SetPageSwapBacked(newpage); 420 421 return MIGRATEPAGE_SUCCESS; 422 } 423 424 oldzone = page_zone(page); 425 newzone = page_zone(newpage); 426 427 xas_lock_irq(&xas); 428 if (page_count(page) != expected_count || xas_load(&xas) != page) { 429 xas_unlock_irq(&xas); 430 return -EAGAIN; 431 } 432 433 if (!page_ref_freeze(page, expected_count)) { 434 xas_unlock_irq(&xas); 435 return -EAGAIN; 436 } 437 438 /* 439 * Now we know that no one else is looking at the page: 440 * no turning back from here. 441 */ 442 newpage->index = page->index; 443 newpage->mapping = page->mapping; 444 page_ref_add(newpage, thp_nr_pages(page)); /* add cache reference */ 445 if (PageSwapBacked(page)) { 446 __SetPageSwapBacked(newpage); 447 if (PageSwapCache(page)) { 448 SetPageSwapCache(newpage); 449 set_page_private(newpage, page_private(page)); 450 } 451 } else { 452 VM_BUG_ON_PAGE(PageSwapCache(page), page); 453 } 454 455 /* Move dirty while page refs frozen and newpage not yet exposed */ 456 dirty = PageDirty(page); 457 if (dirty) { 458 ClearPageDirty(page); 459 SetPageDirty(newpage); 460 } 461 462 xas_store(&xas, newpage); 463 if (PageTransHuge(page)) { 464 int i; 465 466 for (i = 1; i < HPAGE_PMD_NR; i++) { 467 xas_next(&xas); 468 xas_store(&xas, newpage); 469 } 470 } 471 472 /* 473 * Drop cache reference from old page by unfreezing 474 * to one less reference. 475 * We know this isn't the last reference. 476 */ 477 page_ref_unfreeze(page, expected_count - thp_nr_pages(page)); 478 479 xas_unlock(&xas); 480 /* Leave irq disabled to prevent preemption while updating stats */ 481 482 /* 483 * If moved to a different zone then also account 484 * the page for that zone. Other VM counters will be 485 * taken care of when we establish references to the 486 * new page and drop references to the old page. 487 * 488 * Note that anonymous pages are accounted for 489 * via NR_FILE_PAGES and NR_ANON_MAPPED if they 490 * are mapped to swap space. 491 */ 492 if (newzone != oldzone) { 493 struct lruvec *old_lruvec, *new_lruvec; 494 struct mem_cgroup *memcg; 495 496 memcg = page_memcg(page); 497 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 498 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 499 500 __dec_lruvec_state(old_lruvec, NR_FILE_PAGES); 501 __inc_lruvec_state(new_lruvec, NR_FILE_PAGES); 502 if (PageSwapBacked(page) && !PageSwapCache(page)) { 503 __dec_lruvec_state(old_lruvec, NR_SHMEM); 504 __inc_lruvec_state(new_lruvec, NR_SHMEM); 505 } 506 if (dirty && mapping_can_writeback(mapping)) { 507 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY); 508 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING); 509 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY); 510 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING); 511 } 512 } 513 local_irq_enable(); 514 515 return MIGRATEPAGE_SUCCESS; 516 } 517 EXPORT_SYMBOL(migrate_page_move_mapping); 518 519 /* 520 * The expected number of remaining references is the same as that 521 * of migrate_page_move_mapping(). 522 */ 523 int migrate_huge_page_move_mapping(struct address_space *mapping, 524 struct page *newpage, struct page *page) 525 { 526 XA_STATE(xas, &mapping->i_pages, page_index(page)); 527 int expected_count; 528 529 xas_lock_irq(&xas); 530 expected_count = 2 + page_has_private(page); 531 if (page_count(page) != expected_count || xas_load(&xas) != page) { 532 xas_unlock_irq(&xas); 533 return -EAGAIN; 534 } 535 536 if (!page_ref_freeze(page, expected_count)) { 537 xas_unlock_irq(&xas); 538 return -EAGAIN; 539 } 540 541 newpage->index = page->index; 542 newpage->mapping = page->mapping; 543 544 get_page(newpage); 545 546 xas_store(&xas, newpage); 547 548 page_ref_unfreeze(page, expected_count - 1); 549 550 xas_unlock_irq(&xas); 551 552 return MIGRATEPAGE_SUCCESS; 553 } 554 555 /* 556 * Gigantic pages are so large that we do not guarantee that page++ pointer 557 * arithmetic will work across the entire page. We need something more 558 * specialized. 559 */ 560 static void __copy_gigantic_page(struct page *dst, struct page *src, 561 int nr_pages) 562 { 563 int i; 564 struct page *dst_base = dst; 565 struct page *src_base = src; 566 567 for (i = 0; i < nr_pages; ) { 568 cond_resched(); 569 copy_highpage(dst, src); 570 571 i++; 572 dst = mem_map_next(dst, dst_base, i); 573 src = mem_map_next(src, src_base, i); 574 } 575 } 576 577 static void copy_huge_page(struct page *dst, struct page *src) 578 { 579 int i; 580 int nr_pages; 581 582 if (PageHuge(src)) { 583 /* hugetlbfs page */ 584 struct hstate *h = page_hstate(src); 585 nr_pages = pages_per_huge_page(h); 586 587 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { 588 __copy_gigantic_page(dst, src, nr_pages); 589 return; 590 } 591 } else { 592 /* thp page */ 593 BUG_ON(!PageTransHuge(src)); 594 nr_pages = thp_nr_pages(src); 595 } 596 597 for (i = 0; i < nr_pages; i++) { 598 cond_resched(); 599 copy_highpage(dst + i, src + i); 600 } 601 } 602 603 /* 604 * Copy the page to its new location 605 */ 606 void migrate_page_states(struct page *newpage, struct page *page) 607 { 608 int cpupid; 609 610 if (PageError(page)) 611 SetPageError(newpage); 612 if (PageReferenced(page)) 613 SetPageReferenced(newpage); 614 if (PageUptodate(page)) 615 SetPageUptodate(newpage); 616 if (TestClearPageActive(page)) { 617 VM_BUG_ON_PAGE(PageUnevictable(page), page); 618 SetPageActive(newpage); 619 } else if (TestClearPageUnevictable(page)) 620 SetPageUnevictable(newpage); 621 if (PageWorkingset(page)) 622 SetPageWorkingset(newpage); 623 if (PageChecked(page)) 624 SetPageChecked(newpage); 625 if (PageMappedToDisk(page)) 626 SetPageMappedToDisk(newpage); 627 628 /* Move dirty on pages not done by migrate_page_move_mapping() */ 629 if (PageDirty(page)) 630 SetPageDirty(newpage); 631 632 if (page_is_young(page)) 633 set_page_young(newpage); 634 if (page_is_idle(page)) 635 set_page_idle(newpage); 636 637 /* 638 * Copy NUMA information to the new page, to prevent over-eager 639 * future migrations of this same page. 640 */ 641 cpupid = page_cpupid_xchg_last(page, -1); 642 page_cpupid_xchg_last(newpage, cpupid); 643 644 ksm_migrate_page(newpage, page); 645 /* 646 * Please do not reorder this without considering how mm/ksm.c's 647 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 648 */ 649 if (PageSwapCache(page)) 650 ClearPageSwapCache(page); 651 ClearPagePrivate(page); 652 set_page_private(page, 0); 653 654 /* 655 * If any waiters have accumulated on the new page then 656 * wake them up. 657 */ 658 if (PageWriteback(newpage)) 659 end_page_writeback(newpage); 660 661 /* 662 * PG_readahead shares the same bit with PG_reclaim. The above 663 * end_page_writeback() may clear PG_readahead mistakenly, so set the 664 * bit after that. 665 */ 666 if (PageReadahead(page)) 667 SetPageReadahead(newpage); 668 669 copy_page_owner(page, newpage); 670 671 if (!PageHuge(page)) 672 mem_cgroup_migrate(page, newpage); 673 } 674 EXPORT_SYMBOL(migrate_page_states); 675 676 void migrate_page_copy(struct page *newpage, struct page *page) 677 { 678 if (PageHuge(page) || PageTransHuge(page)) 679 copy_huge_page(newpage, page); 680 else 681 copy_highpage(newpage, page); 682 683 migrate_page_states(newpage, page); 684 } 685 EXPORT_SYMBOL(migrate_page_copy); 686 687 /************************************************************ 688 * Migration functions 689 ***********************************************************/ 690 691 /* 692 * Common logic to directly migrate a single LRU page suitable for 693 * pages that do not use PagePrivate/PagePrivate2. 694 * 695 * Pages are locked upon entry and exit. 696 */ 697 int migrate_page(struct address_space *mapping, 698 struct page *newpage, struct page *page, 699 enum migrate_mode mode) 700 { 701 int rc; 702 703 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 704 705 rc = migrate_page_move_mapping(mapping, newpage, page, 0); 706 707 if (rc != MIGRATEPAGE_SUCCESS) 708 return rc; 709 710 if (mode != MIGRATE_SYNC_NO_COPY) 711 migrate_page_copy(newpage, page); 712 else 713 migrate_page_states(newpage, page); 714 return MIGRATEPAGE_SUCCESS; 715 } 716 EXPORT_SYMBOL(migrate_page); 717 718 #ifdef CONFIG_BLOCK 719 /* Returns true if all buffers are successfully locked */ 720 static bool buffer_migrate_lock_buffers(struct buffer_head *head, 721 enum migrate_mode mode) 722 { 723 struct buffer_head *bh = head; 724 725 /* Simple case, sync compaction */ 726 if (mode != MIGRATE_ASYNC) { 727 do { 728 lock_buffer(bh); 729 bh = bh->b_this_page; 730 731 } while (bh != head); 732 733 return true; 734 } 735 736 /* async case, we cannot block on lock_buffer so use trylock_buffer */ 737 do { 738 if (!trylock_buffer(bh)) { 739 /* 740 * We failed to lock the buffer and cannot stall in 741 * async migration. Release the taken locks 742 */ 743 struct buffer_head *failed_bh = bh; 744 bh = head; 745 while (bh != failed_bh) { 746 unlock_buffer(bh); 747 bh = bh->b_this_page; 748 } 749 return false; 750 } 751 752 bh = bh->b_this_page; 753 } while (bh != head); 754 return true; 755 } 756 757 static int __buffer_migrate_page(struct address_space *mapping, 758 struct page *newpage, struct page *page, enum migrate_mode mode, 759 bool check_refs) 760 { 761 struct buffer_head *bh, *head; 762 int rc; 763 int expected_count; 764 765 if (!page_has_buffers(page)) 766 return migrate_page(mapping, newpage, page, mode); 767 768 /* Check whether page does not have extra refs before we do more work */ 769 expected_count = expected_page_refs(mapping, page); 770 if (page_count(page) != expected_count) 771 return -EAGAIN; 772 773 head = page_buffers(page); 774 if (!buffer_migrate_lock_buffers(head, mode)) 775 return -EAGAIN; 776 777 if (check_refs) { 778 bool busy; 779 bool invalidated = false; 780 781 recheck_buffers: 782 busy = false; 783 spin_lock(&mapping->private_lock); 784 bh = head; 785 do { 786 if (atomic_read(&bh->b_count)) { 787 busy = true; 788 break; 789 } 790 bh = bh->b_this_page; 791 } while (bh != head); 792 if (busy) { 793 if (invalidated) { 794 rc = -EAGAIN; 795 goto unlock_buffers; 796 } 797 spin_unlock(&mapping->private_lock); 798 invalidate_bh_lrus(); 799 invalidated = true; 800 goto recheck_buffers; 801 } 802 } 803 804 rc = migrate_page_move_mapping(mapping, newpage, page, 0); 805 if (rc != MIGRATEPAGE_SUCCESS) 806 goto unlock_buffers; 807 808 attach_page_private(newpage, detach_page_private(page)); 809 810 bh = head; 811 do { 812 set_bh_page(bh, newpage, bh_offset(bh)); 813 bh = bh->b_this_page; 814 815 } while (bh != head); 816 817 if (mode != MIGRATE_SYNC_NO_COPY) 818 migrate_page_copy(newpage, page); 819 else 820 migrate_page_states(newpage, page); 821 822 rc = MIGRATEPAGE_SUCCESS; 823 unlock_buffers: 824 if (check_refs) 825 spin_unlock(&mapping->private_lock); 826 bh = head; 827 do { 828 unlock_buffer(bh); 829 bh = bh->b_this_page; 830 831 } while (bh != head); 832 833 return rc; 834 } 835 836 /* 837 * Migration function for pages with buffers. This function can only be used 838 * if the underlying filesystem guarantees that no other references to "page" 839 * exist. For example attached buffer heads are accessed only under page lock. 840 */ 841 int buffer_migrate_page(struct address_space *mapping, 842 struct page *newpage, struct page *page, enum migrate_mode mode) 843 { 844 return __buffer_migrate_page(mapping, newpage, page, mode, false); 845 } 846 EXPORT_SYMBOL(buffer_migrate_page); 847 848 /* 849 * Same as above except that this variant is more careful and checks that there 850 * are also no buffer head references. This function is the right one for 851 * mappings where buffer heads are directly looked up and referenced (such as 852 * block device mappings). 853 */ 854 int buffer_migrate_page_norefs(struct address_space *mapping, 855 struct page *newpage, struct page *page, enum migrate_mode mode) 856 { 857 return __buffer_migrate_page(mapping, newpage, page, mode, true); 858 } 859 #endif 860 861 /* 862 * Writeback a page to clean the dirty state 863 */ 864 static int writeout(struct address_space *mapping, struct page *page) 865 { 866 struct writeback_control wbc = { 867 .sync_mode = WB_SYNC_NONE, 868 .nr_to_write = 1, 869 .range_start = 0, 870 .range_end = LLONG_MAX, 871 .for_reclaim = 1 872 }; 873 int rc; 874 875 if (!mapping->a_ops->writepage) 876 /* No write method for the address space */ 877 return -EINVAL; 878 879 if (!clear_page_dirty_for_io(page)) 880 /* Someone else already triggered a write */ 881 return -EAGAIN; 882 883 /* 884 * A dirty page may imply that the underlying filesystem has 885 * the page on some queue. So the page must be clean for 886 * migration. Writeout may mean we loose the lock and the 887 * page state is no longer what we checked for earlier. 888 * At this point we know that the migration attempt cannot 889 * be successful. 890 */ 891 remove_migration_ptes(page, page, false); 892 893 rc = mapping->a_ops->writepage(page, &wbc); 894 895 if (rc != AOP_WRITEPAGE_ACTIVATE) 896 /* unlocked. Relock */ 897 lock_page(page); 898 899 return (rc < 0) ? -EIO : -EAGAIN; 900 } 901 902 /* 903 * Default handling if a filesystem does not provide a migration function. 904 */ 905 static int fallback_migrate_page(struct address_space *mapping, 906 struct page *newpage, struct page *page, enum migrate_mode mode) 907 { 908 if (PageDirty(page)) { 909 /* Only writeback pages in full synchronous migration */ 910 switch (mode) { 911 case MIGRATE_SYNC: 912 case MIGRATE_SYNC_NO_COPY: 913 break; 914 default: 915 return -EBUSY; 916 } 917 return writeout(mapping, page); 918 } 919 920 /* 921 * Buffers may be managed in a filesystem specific way. 922 * We must have no buffers or drop them. 923 */ 924 if (page_has_private(page) && 925 !try_to_release_page(page, GFP_KERNEL)) 926 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 927 928 return migrate_page(mapping, newpage, page, mode); 929 } 930 931 /* 932 * Move a page to a newly allocated page 933 * The page is locked and all ptes have been successfully removed. 934 * 935 * The new page will have replaced the old page if this function 936 * is successful. 937 * 938 * Return value: 939 * < 0 - error code 940 * MIGRATEPAGE_SUCCESS - success 941 */ 942 static int move_to_new_page(struct page *newpage, struct page *page, 943 enum migrate_mode mode) 944 { 945 struct address_space *mapping; 946 int rc = -EAGAIN; 947 bool is_lru = !__PageMovable(page); 948 949 VM_BUG_ON_PAGE(!PageLocked(page), page); 950 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); 951 952 mapping = page_mapping(page); 953 954 if (likely(is_lru)) { 955 if (!mapping) 956 rc = migrate_page(mapping, newpage, page, mode); 957 else if (mapping->a_ops->migratepage) 958 /* 959 * Most pages have a mapping and most filesystems 960 * provide a migratepage callback. Anonymous pages 961 * are part of swap space which also has its own 962 * migratepage callback. This is the most common path 963 * for page migration. 964 */ 965 rc = mapping->a_ops->migratepage(mapping, newpage, 966 page, mode); 967 else 968 rc = fallback_migrate_page(mapping, newpage, 969 page, mode); 970 } else { 971 /* 972 * In case of non-lru page, it could be released after 973 * isolation step. In that case, we shouldn't try migration. 974 */ 975 VM_BUG_ON_PAGE(!PageIsolated(page), page); 976 if (!PageMovable(page)) { 977 rc = MIGRATEPAGE_SUCCESS; 978 __ClearPageIsolated(page); 979 goto out; 980 } 981 982 rc = mapping->a_ops->migratepage(mapping, newpage, 983 page, mode); 984 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 985 !PageIsolated(page)); 986 } 987 988 /* 989 * When successful, old pagecache page->mapping must be cleared before 990 * page is freed; but stats require that PageAnon be left as PageAnon. 991 */ 992 if (rc == MIGRATEPAGE_SUCCESS) { 993 if (__PageMovable(page)) { 994 VM_BUG_ON_PAGE(!PageIsolated(page), page); 995 996 /* 997 * We clear PG_movable under page_lock so any compactor 998 * cannot try to migrate this page. 999 */ 1000 __ClearPageIsolated(page); 1001 } 1002 1003 /* 1004 * Anonymous and movable page->mapping will be cleared by 1005 * free_pages_prepare so don't reset it here for keeping 1006 * the type to work PageAnon, for example. 1007 */ 1008 if (!PageMappingFlags(page)) 1009 page->mapping = NULL; 1010 1011 if (likely(!is_zone_device_page(newpage))) 1012 flush_dcache_page(newpage); 1013 1014 } 1015 out: 1016 return rc; 1017 } 1018 1019 static int __unmap_and_move(struct page *page, struct page *newpage, 1020 int force, enum migrate_mode mode) 1021 { 1022 int rc = -EAGAIN; 1023 int page_was_mapped = 0; 1024 struct anon_vma *anon_vma = NULL; 1025 bool is_lru = !__PageMovable(page); 1026 1027 if (!trylock_page(page)) { 1028 if (!force || mode == MIGRATE_ASYNC) 1029 goto out; 1030 1031 /* 1032 * It's not safe for direct compaction to call lock_page. 1033 * For example, during page readahead pages are added locked 1034 * to the LRU. Later, when the IO completes the pages are 1035 * marked uptodate and unlocked. However, the queueing 1036 * could be merging multiple pages for one bio (e.g. 1037 * mpage_readahead). If an allocation happens for the 1038 * second or third page, the process can end up locking 1039 * the same page twice and deadlocking. Rather than 1040 * trying to be clever about what pages can be locked, 1041 * avoid the use of lock_page for direct compaction 1042 * altogether. 1043 */ 1044 if (current->flags & PF_MEMALLOC) 1045 goto out; 1046 1047 lock_page(page); 1048 } 1049 1050 if (PageWriteback(page)) { 1051 /* 1052 * Only in the case of a full synchronous migration is it 1053 * necessary to wait for PageWriteback. In the async case, 1054 * the retry loop is too short and in the sync-light case, 1055 * the overhead of stalling is too much 1056 */ 1057 switch (mode) { 1058 case MIGRATE_SYNC: 1059 case MIGRATE_SYNC_NO_COPY: 1060 break; 1061 default: 1062 rc = -EBUSY; 1063 goto out_unlock; 1064 } 1065 if (!force) 1066 goto out_unlock; 1067 wait_on_page_writeback(page); 1068 } 1069 1070 /* 1071 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, 1072 * we cannot notice that anon_vma is freed while we migrates a page. 1073 * This get_anon_vma() delays freeing anon_vma pointer until the end 1074 * of migration. File cache pages are no problem because of page_lock() 1075 * File Caches may use write_page() or lock_page() in migration, then, 1076 * just care Anon page here. 1077 * 1078 * Only page_get_anon_vma() understands the subtleties of 1079 * getting a hold on an anon_vma from outside one of its mms. 1080 * But if we cannot get anon_vma, then we won't need it anyway, 1081 * because that implies that the anon page is no longer mapped 1082 * (and cannot be remapped so long as we hold the page lock). 1083 */ 1084 if (PageAnon(page) && !PageKsm(page)) 1085 anon_vma = page_get_anon_vma(page); 1086 1087 /* 1088 * Block others from accessing the new page when we get around to 1089 * establishing additional references. We are usually the only one 1090 * holding a reference to newpage at this point. We used to have a BUG 1091 * here if trylock_page(newpage) fails, but would like to allow for 1092 * cases where there might be a race with the previous use of newpage. 1093 * This is much like races on refcount of oldpage: just don't BUG(). 1094 */ 1095 if (unlikely(!trylock_page(newpage))) 1096 goto out_unlock; 1097 1098 if (unlikely(!is_lru)) { 1099 rc = move_to_new_page(newpage, page, mode); 1100 goto out_unlock_both; 1101 } 1102 1103 /* 1104 * Corner case handling: 1105 * 1. When a new swap-cache page is read into, it is added to the LRU 1106 * and treated as swapcache but it has no rmap yet. 1107 * Calling try_to_unmap() against a page->mapping==NULL page will 1108 * trigger a BUG. So handle it here. 1109 * 2. An orphaned page (see truncate_complete_page) might have 1110 * fs-private metadata. The page can be picked up due to memory 1111 * offlining. Everywhere else except page reclaim, the page is 1112 * invisible to the vm, so the page can not be migrated. So try to 1113 * free the metadata, so the page can be freed. 1114 */ 1115 if (!page->mapping) { 1116 VM_BUG_ON_PAGE(PageAnon(page), page); 1117 if (page_has_private(page)) { 1118 try_to_free_buffers(page); 1119 goto out_unlock_both; 1120 } 1121 } else if (page_mapped(page)) { 1122 /* Establish migration ptes */ 1123 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, 1124 page); 1125 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK); 1126 page_was_mapped = 1; 1127 } 1128 1129 if (!page_mapped(page)) 1130 rc = move_to_new_page(newpage, page, mode); 1131 1132 if (page_was_mapped) 1133 remove_migration_ptes(page, 1134 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); 1135 1136 out_unlock_both: 1137 unlock_page(newpage); 1138 out_unlock: 1139 /* Drop an anon_vma reference if we took one */ 1140 if (anon_vma) 1141 put_anon_vma(anon_vma); 1142 unlock_page(page); 1143 out: 1144 /* 1145 * If migration is successful, decrease refcount of the newpage 1146 * which will not free the page because new page owner increased 1147 * refcounter. As well, if it is LRU page, add the page to LRU 1148 * list in here. Use the old state of the isolated source page to 1149 * determine if we migrated a LRU page. newpage was already unlocked 1150 * and possibly modified by its owner - don't rely on the page 1151 * state. 1152 */ 1153 if (rc == MIGRATEPAGE_SUCCESS) { 1154 if (unlikely(!is_lru)) 1155 put_page(newpage); 1156 else 1157 putback_lru_page(newpage); 1158 } 1159 1160 return rc; 1161 } 1162 1163 /* 1164 * Obtain the lock on page, remove all ptes and migrate the page 1165 * to the newly allocated page in newpage. 1166 */ 1167 static int unmap_and_move(new_page_t get_new_page, 1168 free_page_t put_new_page, 1169 unsigned long private, struct page *page, 1170 int force, enum migrate_mode mode, 1171 enum migrate_reason reason) 1172 { 1173 int rc = MIGRATEPAGE_SUCCESS; 1174 struct page *newpage = NULL; 1175 1176 if (!thp_migration_supported() && PageTransHuge(page)) 1177 return -ENOMEM; 1178 1179 if (page_count(page) == 1) { 1180 /* page was freed from under us. So we are done. */ 1181 ClearPageActive(page); 1182 ClearPageUnevictable(page); 1183 if (unlikely(__PageMovable(page))) { 1184 lock_page(page); 1185 if (!PageMovable(page)) 1186 __ClearPageIsolated(page); 1187 unlock_page(page); 1188 } 1189 goto out; 1190 } 1191 1192 newpage = get_new_page(page, private); 1193 if (!newpage) 1194 return -ENOMEM; 1195 1196 rc = __unmap_and_move(page, newpage, force, mode); 1197 if (rc == MIGRATEPAGE_SUCCESS) 1198 set_page_owner_migrate_reason(newpage, reason); 1199 1200 out: 1201 if (rc != -EAGAIN) { 1202 /* 1203 * A page that has been migrated has all references 1204 * removed and will be freed. A page that has not been 1205 * migrated will have kept its references and be restored. 1206 */ 1207 list_del(&page->lru); 1208 1209 /* 1210 * Compaction can migrate also non-LRU pages which are 1211 * not accounted to NR_ISOLATED_*. They can be recognized 1212 * as __PageMovable 1213 */ 1214 if (likely(!__PageMovable(page))) 1215 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 1216 page_is_file_lru(page), -thp_nr_pages(page)); 1217 } 1218 1219 /* 1220 * If migration is successful, releases reference grabbed during 1221 * isolation. Otherwise, restore the page to right list unless 1222 * we want to retry. 1223 */ 1224 if (rc == MIGRATEPAGE_SUCCESS) { 1225 if (reason != MR_MEMORY_FAILURE) 1226 /* 1227 * We release the page in page_handle_poison. 1228 */ 1229 put_page(page); 1230 } else { 1231 if (rc != -EAGAIN) { 1232 if (likely(!__PageMovable(page))) { 1233 putback_lru_page(page); 1234 goto put_new; 1235 } 1236 1237 lock_page(page); 1238 if (PageMovable(page)) 1239 putback_movable_page(page); 1240 else 1241 __ClearPageIsolated(page); 1242 unlock_page(page); 1243 put_page(page); 1244 } 1245 put_new: 1246 if (put_new_page) 1247 put_new_page(newpage, private); 1248 else 1249 put_page(newpage); 1250 } 1251 1252 return rc; 1253 } 1254 1255 /* 1256 * Counterpart of unmap_and_move_page() for hugepage migration. 1257 * 1258 * This function doesn't wait the completion of hugepage I/O 1259 * because there is no race between I/O and migration for hugepage. 1260 * Note that currently hugepage I/O occurs only in direct I/O 1261 * where no lock is held and PG_writeback is irrelevant, 1262 * and writeback status of all subpages are counted in the reference 1263 * count of the head page (i.e. if all subpages of a 2MB hugepage are 1264 * under direct I/O, the reference of the head page is 512 and a bit more.) 1265 * This means that when we try to migrate hugepage whose subpages are 1266 * doing direct I/O, some references remain after try_to_unmap() and 1267 * hugepage migration fails without data corruption. 1268 * 1269 * There is also no race when direct I/O is issued on the page under migration, 1270 * because then pte is replaced with migration swap entry and direct I/O code 1271 * will wait in the page fault for migration to complete. 1272 */ 1273 static int unmap_and_move_huge_page(new_page_t get_new_page, 1274 free_page_t put_new_page, unsigned long private, 1275 struct page *hpage, int force, 1276 enum migrate_mode mode, int reason) 1277 { 1278 int rc = -EAGAIN; 1279 int page_was_mapped = 0; 1280 struct page *new_hpage; 1281 struct anon_vma *anon_vma = NULL; 1282 struct address_space *mapping = NULL; 1283 1284 /* 1285 * Migratability of hugepages depends on architectures and their size. 1286 * This check is necessary because some callers of hugepage migration 1287 * like soft offline and memory hotremove don't walk through page 1288 * tables or check whether the hugepage is pmd-based or not before 1289 * kicking migration. 1290 */ 1291 if (!hugepage_migration_supported(page_hstate(hpage))) { 1292 putback_active_hugepage(hpage); 1293 return -ENOSYS; 1294 } 1295 1296 new_hpage = get_new_page(hpage, private); 1297 if (!new_hpage) 1298 return -ENOMEM; 1299 1300 if (!trylock_page(hpage)) { 1301 if (!force) 1302 goto out; 1303 switch (mode) { 1304 case MIGRATE_SYNC: 1305 case MIGRATE_SYNC_NO_COPY: 1306 break; 1307 default: 1308 goto out; 1309 } 1310 lock_page(hpage); 1311 } 1312 1313 /* 1314 * Check for pages which are in the process of being freed. Without 1315 * page_mapping() set, hugetlbfs specific move page routine will not 1316 * be called and we could leak usage counts for subpools. 1317 */ 1318 if (page_private(hpage) && !page_mapping(hpage)) { 1319 rc = -EBUSY; 1320 goto out_unlock; 1321 } 1322 1323 if (PageAnon(hpage)) 1324 anon_vma = page_get_anon_vma(hpage); 1325 1326 if (unlikely(!trylock_page(new_hpage))) 1327 goto put_anon; 1328 1329 if (page_mapped(hpage)) { 1330 bool mapping_locked = false; 1331 enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK; 1332 1333 if (!PageAnon(hpage)) { 1334 /* 1335 * In shared mappings, try_to_unmap could potentially 1336 * call huge_pmd_unshare. Because of this, take 1337 * semaphore in write mode here and set TTU_RMAP_LOCKED 1338 * to let lower levels know we have taken the lock. 1339 */ 1340 mapping = hugetlb_page_mapping_lock_write(hpage); 1341 if (unlikely(!mapping)) 1342 goto unlock_put_anon; 1343 1344 mapping_locked = true; 1345 ttu |= TTU_RMAP_LOCKED; 1346 } 1347 1348 try_to_unmap(hpage, ttu); 1349 page_was_mapped = 1; 1350 1351 if (mapping_locked) 1352 i_mmap_unlock_write(mapping); 1353 } 1354 1355 if (!page_mapped(hpage)) 1356 rc = move_to_new_page(new_hpage, hpage, mode); 1357 1358 if (page_was_mapped) 1359 remove_migration_ptes(hpage, 1360 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); 1361 1362 unlock_put_anon: 1363 unlock_page(new_hpage); 1364 1365 put_anon: 1366 if (anon_vma) 1367 put_anon_vma(anon_vma); 1368 1369 if (rc == MIGRATEPAGE_SUCCESS) { 1370 move_hugetlb_state(hpage, new_hpage, reason); 1371 put_new_page = NULL; 1372 } 1373 1374 out_unlock: 1375 unlock_page(hpage); 1376 out: 1377 if (rc != -EAGAIN) 1378 putback_active_hugepage(hpage); 1379 1380 /* 1381 * If migration was not successful and there's a freeing callback, use 1382 * it. Otherwise, put_page() will drop the reference grabbed during 1383 * isolation. 1384 */ 1385 if (put_new_page) 1386 put_new_page(new_hpage, private); 1387 else 1388 putback_active_hugepage(new_hpage); 1389 1390 return rc; 1391 } 1392 1393 /* 1394 * migrate_pages - migrate the pages specified in a list, to the free pages 1395 * supplied as the target for the page migration 1396 * 1397 * @from: The list of pages to be migrated. 1398 * @get_new_page: The function used to allocate free pages to be used 1399 * as the target of the page migration. 1400 * @put_new_page: The function used to free target pages if migration 1401 * fails, or NULL if no special handling is necessary. 1402 * @private: Private data to be passed on to get_new_page() 1403 * @mode: The migration mode that specifies the constraints for 1404 * page migration, if any. 1405 * @reason: The reason for page migration. 1406 * 1407 * The function returns after 10 attempts or if no pages are movable any more 1408 * because the list has become empty or no retryable pages exist any more. 1409 * The caller should call putback_movable_pages() to return pages to the LRU 1410 * or free list only if ret != 0. 1411 * 1412 * Returns the number of pages that were not migrated, or an error code. 1413 */ 1414 int migrate_pages(struct list_head *from, new_page_t get_new_page, 1415 free_page_t put_new_page, unsigned long private, 1416 enum migrate_mode mode, int reason) 1417 { 1418 int retry = 1; 1419 int thp_retry = 1; 1420 int nr_failed = 0; 1421 int nr_succeeded = 0; 1422 int nr_thp_succeeded = 0; 1423 int nr_thp_failed = 0; 1424 int nr_thp_split = 0; 1425 int pass = 0; 1426 bool is_thp = false; 1427 struct page *page; 1428 struct page *page2; 1429 int swapwrite = current->flags & PF_SWAPWRITE; 1430 int rc, nr_subpages; 1431 1432 if (!swapwrite) 1433 current->flags |= PF_SWAPWRITE; 1434 1435 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { 1436 retry = 0; 1437 thp_retry = 0; 1438 1439 list_for_each_entry_safe(page, page2, from, lru) { 1440 retry: 1441 /* 1442 * THP statistics is based on the source huge page. 1443 * Capture required information that might get lost 1444 * during migration. 1445 */ 1446 is_thp = PageTransHuge(page) && !PageHuge(page); 1447 nr_subpages = thp_nr_pages(page); 1448 cond_resched(); 1449 1450 if (PageHuge(page)) 1451 rc = unmap_and_move_huge_page(get_new_page, 1452 put_new_page, private, page, 1453 pass > 2, mode, reason); 1454 else 1455 rc = unmap_and_move(get_new_page, put_new_page, 1456 private, page, pass > 2, mode, 1457 reason); 1458 1459 switch(rc) { 1460 case -ENOMEM: 1461 /* 1462 * THP migration might be unsupported or the 1463 * allocation could've failed so we should 1464 * retry on the same page with the THP split 1465 * to base pages. 1466 * 1467 * Head page is retried immediately and tail 1468 * pages are added to the tail of the list so 1469 * we encounter them after the rest of the list 1470 * is processed. 1471 */ 1472 if (is_thp) { 1473 lock_page(page); 1474 rc = split_huge_page_to_list(page, from); 1475 unlock_page(page); 1476 if (!rc) { 1477 list_safe_reset_next(page, page2, lru); 1478 nr_thp_split++; 1479 goto retry; 1480 } 1481 1482 nr_thp_failed++; 1483 nr_failed += nr_subpages; 1484 goto out; 1485 } 1486 nr_failed++; 1487 goto out; 1488 case -EAGAIN: 1489 if (is_thp) { 1490 thp_retry++; 1491 break; 1492 } 1493 retry++; 1494 break; 1495 case MIGRATEPAGE_SUCCESS: 1496 if (is_thp) { 1497 nr_thp_succeeded++; 1498 nr_succeeded += nr_subpages; 1499 break; 1500 } 1501 nr_succeeded++; 1502 break; 1503 default: 1504 /* 1505 * Permanent failure (-EBUSY, -ENOSYS, etc.): 1506 * unlike -EAGAIN case, the failed page is 1507 * removed from migration page list and not 1508 * retried in the next outer loop. 1509 */ 1510 if (is_thp) { 1511 nr_thp_failed++; 1512 nr_failed += nr_subpages; 1513 break; 1514 } 1515 nr_failed++; 1516 break; 1517 } 1518 } 1519 } 1520 nr_failed += retry + thp_retry; 1521 nr_thp_failed += thp_retry; 1522 rc = nr_failed; 1523 out: 1524 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); 1525 count_vm_events(PGMIGRATE_FAIL, nr_failed); 1526 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); 1527 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); 1528 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); 1529 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded, 1530 nr_thp_failed, nr_thp_split, mode, reason); 1531 1532 if (!swapwrite) 1533 current->flags &= ~PF_SWAPWRITE; 1534 1535 return rc; 1536 } 1537 1538 struct page *alloc_migration_target(struct page *page, unsigned long private) 1539 { 1540 struct migration_target_control *mtc; 1541 gfp_t gfp_mask; 1542 unsigned int order = 0; 1543 struct page *new_page = NULL; 1544 int nid; 1545 int zidx; 1546 1547 mtc = (struct migration_target_control *)private; 1548 gfp_mask = mtc->gfp_mask; 1549 nid = mtc->nid; 1550 if (nid == NUMA_NO_NODE) 1551 nid = page_to_nid(page); 1552 1553 if (PageHuge(page)) { 1554 struct hstate *h = page_hstate(compound_head(page)); 1555 1556 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 1557 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); 1558 } 1559 1560 if (PageTransHuge(page)) { 1561 /* 1562 * clear __GFP_RECLAIM to make the migration callback 1563 * consistent with regular THP allocations. 1564 */ 1565 gfp_mask &= ~__GFP_RECLAIM; 1566 gfp_mask |= GFP_TRANSHUGE; 1567 order = HPAGE_PMD_ORDER; 1568 } 1569 zidx = zone_idx(page_zone(page)); 1570 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 1571 gfp_mask |= __GFP_HIGHMEM; 1572 1573 new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask); 1574 1575 if (new_page && PageTransHuge(new_page)) 1576 prep_transhuge_page(new_page); 1577 1578 return new_page; 1579 } 1580 1581 #ifdef CONFIG_NUMA 1582 1583 static int store_status(int __user *status, int start, int value, int nr) 1584 { 1585 while (nr-- > 0) { 1586 if (put_user(value, status + start)) 1587 return -EFAULT; 1588 start++; 1589 } 1590 1591 return 0; 1592 } 1593 1594 static int do_move_pages_to_node(struct mm_struct *mm, 1595 struct list_head *pagelist, int node) 1596 { 1597 int err; 1598 struct migration_target_control mtc = { 1599 .nid = node, 1600 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 1601 }; 1602 1603 err = migrate_pages(pagelist, alloc_migration_target, NULL, 1604 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL); 1605 if (err) 1606 putback_movable_pages(pagelist); 1607 return err; 1608 } 1609 1610 /* 1611 * Resolves the given address to a struct page, isolates it from the LRU and 1612 * puts it to the given pagelist. 1613 * Returns: 1614 * errno - if the page cannot be found/isolated 1615 * 0 - when it doesn't have to be migrated because it is already on the 1616 * target node 1617 * 1 - when it has been queued 1618 */ 1619 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, 1620 int node, struct list_head *pagelist, bool migrate_all) 1621 { 1622 struct vm_area_struct *vma; 1623 struct page *page; 1624 unsigned int follflags; 1625 int err; 1626 1627 mmap_read_lock(mm); 1628 err = -EFAULT; 1629 vma = find_vma(mm, addr); 1630 if (!vma || addr < vma->vm_start || !vma_migratable(vma)) 1631 goto out; 1632 1633 /* FOLL_DUMP to ignore special (like zero) pages */ 1634 follflags = FOLL_GET | FOLL_DUMP; 1635 page = follow_page(vma, addr, follflags); 1636 1637 err = PTR_ERR(page); 1638 if (IS_ERR(page)) 1639 goto out; 1640 1641 err = -ENOENT; 1642 if (!page) 1643 goto out; 1644 1645 err = 0; 1646 if (page_to_nid(page) == node) 1647 goto out_putpage; 1648 1649 err = -EACCES; 1650 if (page_mapcount(page) > 1 && !migrate_all) 1651 goto out_putpage; 1652 1653 if (PageHuge(page)) { 1654 if (PageHead(page)) { 1655 isolate_huge_page(page, pagelist); 1656 err = 1; 1657 } 1658 } else { 1659 struct page *head; 1660 1661 head = compound_head(page); 1662 err = isolate_lru_page(head); 1663 if (err) 1664 goto out_putpage; 1665 1666 err = 1; 1667 list_add_tail(&head->lru, pagelist); 1668 mod_node_page_state(page_pgdat(head), 1669 NR_ISOLATED_ANON + page_is_file_lru(head), 1670 thp_nr_pages(head)); 1671 } 1672 out_putpage: 1673 /* 1674 * Either remove the duplicate refcount from 1675 * isolate_lru_page() or drop the page ref if it was 1676 * not isolated. 1677 */ 1678 put_page(page); 1679 out: 1680 mmap_read_unlock(mm); 1681 return err; 1682 } 1683 1684 static int move_pages_and_store_status(struct mm_struct *mm, int node, 1685 struct list_head *pagelist, int __user *status, 1686 int start, int i, unsigned long nr_pages) 1687 { 1688 int err; 1689 1690 if (list_empty(pagelist)) 1691 return 0; 1692 1693 err = do_move_pages_to_node(mm, pagelist, node); 1694 if (err) { 1695 /* 1696 * Positive err means the number of failed 1697 * pages to migrate. Since we are going to 1698 * abort and return the number of non-migrated 1699 * pages, so need to incude the rest of the 1700 * nr_pages that have not been attempted as 1701 * well. 1702 */ 1703 if (err > 0) 1704 err += nr_pages - i - 1; 1705 return err; 1706 } 1707 return store_status(status, start, node, i - start); 1708 } 1709 1710 /* 1711 * Migrate an array of page address onto an array of nodes and fill 1712 * the corresponding array of status. 1713 */ 1714 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 1715 unsigned long nr_pages, 1716 const void __user * __user *pages, 1717 const int __user *nodes, 1718 int __user *status, int flags) 1719 { 1720 int current_node = NUMA_NO_NODE; 1721 LIST_HEAD(pagelist); 1722 int start, i; 1723 int err = 0, err1; 1724 1725 migrate_prep(); 1726 1727 for (i = start = 0; i < nr_pages; i++) { 1728 const void __user *p; 1729 unsigned long addr; 1730 int node; 1731 1732 err = -EFAULT; 1733 if (get_user(p, pages + i)) 1734 goto out_flush; 1735 if (get_user(node, nodes + i)) 1736 goto out_flush; 1737 addr = (unsigned long)untagged_addr(p); 1738 1739 err = -ENODEV; 1740 if (node < 0 || node >= MAX_NUMNODES) 1741 goto out_flush; 1742 if (!node_state(node, N_MEMORY)) 1743 goto out_flush; 1744 1745 err = -EACCES; 1746 if (!node_isset(node, task_nodes)) 1747 goto out_flush; 1748 1749 if (current_node == NUMA_NO_NODE) { 1750 current_node = node; 1751 start = i; 1752 } else if (node != current_node) { 1753 err = move_pages_and_store_status(mm, current_node, 1754 &pagelist, status, start, i, nr_pages); 1755 if (err) 1756 goto out; 1757 start = i; 1758 current_node = node; 1759 } 1760 1761 /* 1762 * Errors in the page lookup or isolation are not fatal and we simply 1763 * report them via status 1764 */ 1765 err = add_page_for_migration(mm, addr, current_node, 1766 &pagelist, flags & MPOL_MF_MOVE_ALL); 1767 1768 if (err > 0) { 1769 /* The page is successfully queued for migration */ 1770 continue; 1771 } 1772 1773 /* 1774 * If the page is already on the target node (!err), store the 1775 * node, otherwise, store the err. 1776 */ 1777 err = store_status(status, i, err ? : current_node, 1); 1778 if (err) 1779 goto out_flush; 1780 1781 err = move_pages_and_store_status(mm, current_node, &pagelist, 1782 status, start, i, nr_pages); 1783 if (err) 1784 goto out; 1785 current_node = NUMA_NO_NODE; 1786 } 1787 out_flush: 1788 /* Make sure we do not overwrite the existing error */ 1789 err1 = move_pages_and_store_status(mm, current_node, &pagelist, 1790 status, start, i, nr_pages); 1791 if (err >= 0) 1792 err = err1; 1793 out: 1794 return err; 1795 } 1796 1797 /* 1798 * Determine the nodes of an array of pages and store it in an array of status. 1799 */ 1800 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 1801 const void __user **pages, int *status) 1802 { 1803 unsigned long i; 1804 1805 mmap_read_lock(mm); 1806 1807 for (i = 0; i < nr_pages; i++) { 1808 unsigned long addr = (unsigned long)(*pages); 1809 struct vm_area_struct *vma; 1810 struct page *page; 1811 int err = -EFAULT; 1812 1813 vma = find_vma(mm, addr); 1814 if (!vma || addr < vma->vm_start) 1815 goto set_status; 1816 1817 /* FOLL_DUMP to ignore special (like zero) pages */ 1818 page = follow_page(vma, addr, FOLL_DUMP); 1819 1820 err = PTR_ERR(page); 1821 if (IS_ERR(page)) 1822 goto set_status; 1823 1824 err = page ? page_to_nid(page) : -ENOENT; 1825 set_status: 1826 *status = err; 1827 1828 pages++; 1829 status++; 1830 } 1831 1832 mmap_read_unlock(mm); 1833 } 1834 1835 /* 1836 * Determine the nodes of a user array of pages and store it in 1837 * a user array of status. 1838 */ 1839 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 1840 const void __user * __user *pages, 1841 int __user *status) 1842 { 1843 #define DO_PAGES_STAT_CHUNK_NR 16 1844 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 1845 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 1846 1847 while (nr_pages) { 1848 unsigned long chunk_nr; 1849 1850 chunk_nr = nr_pages; 1851 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) 1852 chunk_nr = DO_PAGES_STAT_CHUNK_NR; 1853 1854 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) 1855 break; 1856 1857 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 1858 1859 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 1860 break; 1861 1862 pages += chunk_nr; 1863 status += chunk_nr; 1864 nr_pages -= chunk_nr; 1865 } 1866 return nr_pages ? -EFAULT : 0; 1867 } 1868 1869 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 1870 { 1871 struct task_struct *task; 1872 struct mm_struct *mm; 1873 1874 /* 1875 * There is no need to check if current process has the right to modify 1876 * the specified process when they are same. 1877 */ 1878 if (!pid) { 1879 mmget(current->mm); 1880 *mem_nodes = cpuset_mems_allowed(current); 1881 return current->mm; 1882 } 1883 1884 /* Find the mm_struct */ 1885 rcu_read_lock(); 1886 task = find_task_by_vpid(pid); 1887 if (!task) { 1888 rcu_read_unlock(); 1889 return ERR_PTR(-ESRCH); 1890 } 1891 get_task_struct(task); 1892 1893 /* 1894 * Check if this process has the right to modify the specified 1895 * process. Use the regular "ptrace_may_access()" checks. 1896 */ 1897 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 1898 rcu_read_unlock(); 1899 mm = ERR_PTR(-EPERM); 1900 goto out; 1901 } 1902 rcu_read_unlock(); 1903 1904 mm = ERR_PTR(security_task_movememory(task)); 1905 if (IS_ERR(mm)) 1906 goto out; 1907 *mem_nodes = cpuset_mems_allowed(task); 1908 mm = get_task_mm(task); 1909 out: 1910 put_task_struct(task); 1911 if (!mm) 1912 mm = ERR_PTR(-EINVAL); 1913 return mm; 1914 } 1915 1916 /* 1917 * Move a list of pages in the address space of the currently executing 1918 * process. 1919 */ 1920 static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 1921 const void __user * __user *pages, 1922 const int __user *nodes, 1923 int __user *status, int flags) 1924 { 1925 struct mm_struct *mm; 1926 int err; 1927 nodemask_t task_nodes; 1928 1929 /* Check flags */ 1930 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 1931 return -EINVAL; 1932 1933 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1934 return -EPERM; 1935 1936 mm = find_mm_struct(pid, &task_nodes); 1937 if (IS_ERR(mm)) 1938 return PTR_ERR(mm); 1939 1940 if (nodes) 1941 err = do_pages_move(mm, task_nodes, nr_pages, pages, 1942 nodes, status, flags); 1943 else 1944 err = do_pages_stat(mm, nr_pages, pages, status); 1945 1946 mmput(mm); 1947 return err; 1948 } 1949 1950 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 1951 const void __user * __user *, pages, 1952 const int __user *, nodes, 1953 int __user *, status, int, flags) 1954 { 1955 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 1956 } 1957 1958 #ifdef CONFIG_COMPAT 1959 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, 1960 compat_uptr_t __user *, pages32, 1961 const int __user *, nodes, 1962 int __user *, status, 1963 int, flags) 1964 { 1965 const void __user * __user *pages; 1966 int i; 1967 1968 pages = compat_alloc_user_space(nr_pages * sizeof(void *)); 1969 for (i = 0; i < nr_pages; i++) { 1970 compat_uptr_t p; 1971 1972 if (get_user(p, pages32 + i) || 1973 put_user(compat_ptr(p), pages + i)) 1974 return -EFAULT; 1975 } 1976 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 1977 } 1978 #endif /* CONFIG_COMPAT */ 1979 1980 #ifdef CONFIG_NUMA_BALANCING 1981 /* 1982 * Returns true if this is a safe migration target node for misplaced NUMA 1983 * pages. Currently it only checks the watermarks which crude 1984 */ 1985 static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 1986 unsigned long nr_migrate_pages) 1987 { 1988 int z; 1989 1990 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 1991 struct zone *zone = pgdat->node_zones + z; 1992 1993 if (!populated_zone(zone)) 1994 continue; 1995 1996 /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 1997 if (!zone_watermark_ok(zone, 0, 1998 high_wmark_pages(zone) + 1999 nr_migrate_pages, 2000 ZONE_MOVABLE, 0)) 2001 continue; 2002 return true; 2003 } 2004 return false; 2005 } 2006 2007 static struct page *alloc_misplaced_dst_page(struct page *page, 2008 unsigned long data) 2009 { 2010 int nid = (int) data; 2011 struct page *newpage; 2012 2013 newpage = __alloc_pages_node(nid, 2014 (GFP_HIGHUSER_MOVABLE | 2015 __GFP_THISNODE | __GFP_NOMEMALLOC | 2016 __GFP_NORETRY | __GFP_NOWARN) & 2017 ~__GFP_RECLAIM, 0); 2018 2019 return newpage; 2020 } 2021 2022 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 2023 { 2024 int page_lru; 2025 2026 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); 2027 2028 /* Avoid migrating to a node that is nearly full */ 2029 if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) 2030 return 0; 2031 2032 if (isolate_lru_page(page)) 2033 return 0; 2034 2035 /* 2036 * migrate_misplaced_transhuge_page() skips page migration's usual 2037 * check on page_count(), so we must do it here, now that the page 2038 * has been isolated: a GUP pin, or any other pin, prevents migration. 2039 * The expected page count is 3: 1 for page's mapcount and 1 for the 2040 * caller's pin and 1 for the reference taken by isolate_lru_page(). 2041 */ 2042 if (PageTransHuge(page) && page_count(page) != 3) { 2043 putback_lru_page(page); 2044 return 0; 2045 } 2046 2047 page_lru = page_is_file_lru(page); 2048 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, 2049 thp_nr_pages(page)); 2050 2051 /* 2052 * Isolating the page has taken another reference, so the 2053 * caller's reference can be safely dropped without the page 2054 * disappearing underneath us during migration. 2055 */ 2056 put_page(page); 2057 return 1; 2058 } 2059 2060 bool pmd_trans_migrating(pmd_t pmd) 2061 { 2062 struct page *page = pmd_page(pmd); 2063 return PageLocked(page); 2064 } 2065 2066 /* 2067 * Attempt to migrate a misplaced page to the specified destination 2068 * node. Caller is expected to have an elevated reference count on 2069 * the page that will be dropped by this function before returning. 2070 */ 2071 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 2072 int node) 2073 { 2074 pg_data_t *pgdat = NODE_DATA(node); 2075 int isolated; 2076 int nr_remaining; 2077 LIST_HEAD(migratepages); 2078 2079 /* 2080 * Don't migrate file pages that are mapped in multiple processes 2081 * with execute permissions as they are probably shared libraries. 2082 */ 2083 if (page_mapcount(page) != 1 && page_is_file_lru(page) && 2084 (vma->vm_flags & VM_EXEC)) 2085 goto out; 2086 2087 /* 2088 * Also do not migrate dirty pages as not all filesystems can move 2089 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. 2090 */ 2091 if (page_is_file_lru(page) && PageDirty(page)) 2092 goto out; 2093 2094 isolated = numamigrate_isolate_page(pgdat, page); 2095 if (!isolated) 2096 goto out; 2097 2098 list_add(&page->lru, &migratepages); 2099 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, 2100 NULL, node, MIGRATE_ASYNC, 2101 MR_NUMA_MISPLACED); 2102 if (nr_remaining) { 2103 if (!list_empty(&migratepages)) { 2104 list_del(&page->lru); 2105 dec_node_page_state(page, NR_ISOLATED_ANON + 2106 page_is_file_lru(page)); 2107 putback_lru_page(page); 2108 } 2109 isolated = 0; 2110 } else 2111 count_vm_numa_event(NUMA_PAGE_MIGRATE); 2112 BUG_ON(!list_empty(&migratepages)); 2113 return isolated; 2114 2115 out: 2116 put_page(page); 2117 return 0; 2118 } 2119 #endif /* CONFIG_NUMA_BALANCING */ 2120 2121 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) 2122 /* 2123 * Migrates a THP to a given target node. page must be locked and is unlocked 2124 * before returning. 2125 */ 2126 int migrate_misplaced_transhuge_page(struct mm_struct *mm, 2127 struct vm_area_struct *vma, 2128 pmd_t *pmd, pmd_t entry, 2129 unsigned long address, 2130 struct page *page, int node) 2131 { 2132 spinlock_t *ptl; 2133 pg_data_t *pgdat = NODE_DATA(node); 2134 int isolated = 0; 2135 struct page *new_page = NULL; 2136 int page_lru = page_is_file_lru(page); 2137 unsigned long start = address & HPAGE_PMD_MASK; 2138 2139 new_page = alloc_pages_node(node, 2140 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), 2141 HPAGE_PMD_ORDER); 2142 if (!new_page) 2143 goto out_fail; 2144 prep_transhuge_page(new_page); 2145 2146 isolated = numamigrate_isolate_page(pgdat, page); 2147 if (!isolated) { 2148 put_page(new_page); 2149 goto out_fail; 2150 } 2151 2152 /* Prepare a page as a migration target */ 2153 __SetPageLocked(new_page); 2154 if (PageSwapBacked(page)) 2155 __SetPageSwapBacked(new_page); 2156 2157 /* anon mapping, we can simply copy page->mapping to the new page: */ 2158 new_page->mapping = page->mapping; 2159 new_page->index = page->index; 2160 /* flush the cache before copying using the kernel virtual address */ 2161 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); 2162 migrate_page_copy(new_page, page); 2163 WARN_ON(PageLRU(new_page)); 2164 2165 /* Recheck the target PMD */ 2166 ptl = pmd_lock(mm, pmd); 2167 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { 2168 spin_unlock(ptl); 2169 2170 /* Reverse changes made by migrate_page_copy() */ 2171 if (TestClearPageActive(new_page)) 2172 SetPageActive(page); 2173 if (TestClearPageUnevictable(new_page)) 2174 SetPageUnevictable(page); 2175 2176 unlock_page(new_page); 2177 put_page(new_page); /* Free it */ 2178 2179 /* Retake the callers reference and putback on LRU */ 2180 get_page(page); 2181 putback_lru_page(page); 2182 mod_node_page_state(page_pgdat(page), 2183 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); 2184 2185 goto out_unlock; 2186 } 2187 2188 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 2189 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 2190 2191 /* 2192 * Overwrite the old entry under pagetable lock and establish 2193 * the new PTE. Any parallel GUP will either observe the old 2194 * page blocking on the page lock, block on the page table 2195 * lock or observe the new page. The SetPageUptodate on the 2196 * new page and page_add_new_anon_rmap guarantee the copy is 2197 * visible before the pagetable update. 2198 */ 2199 page_add_anon_rmap(new_page, vma, start, true); 2200 /* 2201 * At this point the pmd is numa/protnone (i.e. non present) and the TLB 2202 * has already been flushed globally. So no TLB can be currently 2203 * caching this non present pmd mapping. There's no need to clear the 2204 * pmd before doing set_pmd_at(), nor to flush the TLB after 2205 * set_pmd_at(). Clearing the pmd here would introduce a race 2206 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the 2207 * mmap_lock for reading. If the pmd is set to NULL at any given time, 2208 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this 2209 * pmd. 2210 */ 2211 set_pmd_at(mm, start, pmd, entry); 2212 update_mmu_cache_pmd(vma, address, &entry); 2213 2214 page_ref_unfreeze(page, 2); 2215 mlock_migrate_page(new_page, page); 2216 page_remove_rmap(page, true); 2217 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); 2218 2219 spin_unlock(ptl); 2220 2221 /* Take an "isolate" reference and put new page on the LRU. */ 2222 get_page(new_page); 2223 putback_lru_page(new_page); 2224 2225 unlock_page(new_page); 2226 unlock_page(page); 2227 put_page(page); /* Drop the rmap reference */ 2228 put_page(page); /* Drop the LRU isolation reference */ 2229 2230 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); 2231 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); 2232 2233 mod_node_page_state(page_pgdat(page), 2234 NR_ISOLATED_ANON + page_lru, 2235 -HPAGE_PMD_NR); 2236 return isolated; 2237 2238 out_fail: 2239 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); 2240 ptl = pmd_lock(mm, pmd); 2241 if (pmd_same(*pmd, entry)) { 2242 entry = pmd_modify(entry, vma->vm_page_prot); 2243 set_pmd_at(mm, start, pmd, entry); 2244 update_mmu_cache_pmd(vma, address, &entry); 2245 } 2246 spin_unlock(ptl); 2247 2248 out_unlock: 2249 unlock_page(page); 2250 put_page(page); 2251 return 0; 2252 } 2253 #endif /* CONFIG_NUMA_BALANCING */ 2254 2255 #endif /* CONFIG_NUMA */ 2256 2257 #ifdef CONFIG_DEVICE_PRIVATE 2258 static int migrate_vma_collect_hole(unsigned long start, 2259 unsigned long end, 2260 __always_unused int depth, 2261 struct mm_walk *walk) 2262 { 2263 struct migrate_vma *migrate = walk->private; 2264 unsigned long addr; 2265 2266 /* Only allow populating anonymous memory. */ 2267 if (!vma_is_anonymous(walk->vma)) { 2268 for (addr = start; addr < end; addr += PAGE_SIZE) { 2269 migrate->src[migrate->npages] = 0; 2270 migrate->dst[migrate->npages] = 0; 2271 migrate->npages++; 2272 } 2273 return 0; 2274 } 2275 2276 for (addr = start; addr < end; addr += PAGE_SIZE) { 2277 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; 2278 migrate->dst[migrate->npages] = 0; 2279 migrate->npages++; 2280 migrate->cpages++; 2281 } 2282 2283 return 0; 2284 } 2285 2286 static int migrate_vma_collect_skip(unsigned long start, 2287 unsigned long end, 2288 struct mm_walk *walk) 2289 { 2290 struct migrate_vma *migrate = walk->private; 2291 unsigned long addr; 2292 2293 for (addr = start; addr < end; addr += PAGE_SIZE) { 2294 migrate->dst[migrate->npages] = 0; 2295 migrate->src[migrate->npages++] = 0; 2296 } 2297 2298 return 0; 2299 } 2300 2301 static int migrate_vma_collect_pmd(pmd_t *pmdp, 2302 unsigned long start, 2303 unsigned long end, 2304 struct mm_walk *walk) 2305 { 2306 struct migrate_vma *migrate = walk->private; 2307 struct vm_area_struct *vma = walk->vma; 2308 struct mm_struct *mm = vma->vm_mm; 2309 unsigned long addr = start, unmapped = 0; 2310 spinlock_t *ptl; 2311 pte_t *ptep; 2312 2313 again: 2314 if (pmd_none(*pmdp)) 2315 return migrate_vma_collect_hole(start, end, -1, walk); 2316 2317 if (pmd_trans_huge(*pmdp)) { 2318 struct page *page; 2319 2320 ptl = pmd_lock(mm, pmdp); 2321 if (unlikely(!pmd_trans_huge(*pmdp))) { 2322 spin_unlock(ptl); 2323 goto again; 2324 } 2325 2326 page = pmd_page(*pmdp); 2327 if (is_huge_zero_page(page)) { 2328 spin_unlock(ptl); 2329 split_huge_pmd(vma, pmdp, addr); 2330 if (pmd_trans_unstable(pmdp)) 2331 return migrate_vma_collect_skip(start, end, 2332 walk); 2333 } else { 2334 int ret; 2335 2336 get_page(page); 2337 spin_unlock(ptl); 2338 if (unlikely(!trylock_page(page))) 2339 return migrate_vma_collect_skip(start, end, 2340 walk); 2341 ret = split_huge_page(page); 2342 unlock_page(page); 2343 put_page(page); 2344 if (ret) 2345 return migrate_vma_collect_skip(start, end, 2346 walk); 2347 if (pmd_none(*pmdp)) 2348 return migrate_vma_collect_hole(start, end, -1, 2349 walk); 2350 } 2351 } 2352 2353 if (unlikely(pmd_bad(*pmdp))) 2354 return migrate_vma_collect_skip(start, end, walk); 2355 2356 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); 2357 arch_enter_lazy_mmu_mode(); 2358 2359 for (; addr < end; addr += PAGE_SIZE, ptep++) { 2360 unsigned long mpfn = 0, pfn; 2361 struct page *page; 2362 swp_entry_t entry; 2363 pte_t pte; 2364 2365 pte = *ptep; 2366 2367 if (pte_none(pte)) { 2368 if (vma_is_anonymous(vma)) { 2369 mpfn = MIGRATE_PFN_MIGRATE; 2370 migrate->cpages++; 2371 } 2372 goto next; 2373 } 2374 2375 if (!pte_present(pte)) { 2376 /* 2377 * Only care about unaddressable device page special 2378 * page table entry. Other special swap entries are not 2379 * migratable, and we ignore regular swapped page. 2380 */ 2381 entry = pte_to_swp_entry(pte); 2382 if (!is_device_private_entry(entry)) 2383 goto next; 2384 2385 page = device_private_entry_to_page(entry); 2386 if (!(migrate->flags & 2387 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || 2388 page->pgmap->owner != migrate->pgmap_owner) 2389 goto next; 2390 2391 mpfn = migrate_pfn(page_to_pfn(page)) | 2392 MIGRATE_PFN_MIGRATE; 2393 if (is_write_device_private_entry(entry)) 2394 mpfn |= MIGRATE_PFN_WRITE; 2395 } else { 2396 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) 2397 goto next; 2398 pfn = pte_pfn(pte); 2399 if (is_zero_pfn(pfn)) { 2400 mpfn = MIGRATE_PFN_MIGRATE; 2401 migrate->cpages++; 2402 goto next; 2403 } 2404 page = vm_normal_page(migrate->vma, addr, pte); 2405 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; 2406 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; 2407 } 2408 2409 /* FIXME support THP */ 2410 if (!page || !page->mapping || PageTransCompound(page)) { 2411 mpfn = 0; 2412 goto next; 2413 } 2414 2415 /* 2416 * By getting a reference on the page we pin it and that blocks 2417 * any kind of migration. Side effect is that it "freezes" the 2418 * pte. 2419 * 2420 * We drop this reference after isolating the page from the lru 2421 * for non device page (device page are not on the lru and thus 2422 * can't be dropped from it). 2423 */ 2424 get_page(page); 2425 migrate->cpages++; 2426 2427 /* 2428 * Optimize for the common case where page is only mapped once 2429 * in one process. If we can lock the page, then we can safely 2430 * set up a special migration page table entry now. 2431 */ 2432 if (trylock_page(page)) { 2433 pte_t swp_pte; 2434 2435 mpfn |= MIGRATE_PFN_LOCKED; 2436 ptep_get_and_clear(mm, addr, ptep); 2437 2438 /* Setup special migration page table entry */ 2439 entry = make_migration_entry(page, mpfn & 2440 MIGRATE_PFN_WRITE); 2441 swp_pte = swp_entry_to_pte(entry); 2442 if (pte_present(pte)) { 2443 if (pte_soft_dirty(pte)) 2444 swp_pte = pte_swp_mksoft_dirty(swp_pte); 2445 if (pte_uffd_wp(pte)) 2446 swp_pte = pte_swp_mkuffd_wp(swp_pte); 2447 } else { 2448 if (pte_swp_soft_dirty(pte)) 2449 swp_pte = pte_swp_mksoft_dirty(swp_pte); 2450 if (pte_swp_uffd_wp(pte)) 2451 swp_pte = pte_swp_mkuffd_wp(swp_pte); 2452 } 2453 set_pte_at(mm, addr, ptep, swp_pte); 2454 2455 /* 2456 * This is like regular unmap: we remove the rmap and 2457 * drop page refcount. Page won't be freed, as we took 2458 * a reference just above. 2459 */ 2460 page_remove_rmap(page, false); 2461 put_page(page); 2462 2463 if (pte_present(pte)) 2464 unmapped++; 2465 } 2466 2467 next: 2468 migrate->dst[migrate->npages] = 0; 2469 migrate->src[migrate->npages++] = mpfn; 2470 } 2471 arch_leave_lazy_mmu_mode(); 2472 pte_unmap_unlock(ptep - 1, ptl); 2473 2474 /* Only flush the TLB if we actually modified any entries */ 2475 if (unmapped) 2476 flush_tlb_range(walk->vma, start, end); 2477 2478 return 0; 2479 } 2480 2481 static const struct mm_walk_ops migrate_vma_walk_ops = { 2482 .pmd_entry = migrate_vma_collect_pmd, 2483 .pte_hole = migrate_vma_collect_hole, 2484 }; 2485 2486 /* 2487 * migrate_vma_collect() - collect pages over a range of virtual addresses 2488 * @migrate: migrate struct containing all migration information 2489 * 2490 * This will walk the CPU page table. For each virtual address backed by a 2491 * valid page, it updates the src array and takes a reference on the page, in 2492 * order to pin the page until we lock it and unmap it. 2493 */ 2494 static void migrate_vma_collect(struct migrate_vma *migrate) 2495 { 2496 struct mmu_notifier_range range; 2497 2498 /* 2499 * Note that the pgmap_owner is passed to the mmu notifier callback so 2500 * that the registered device driver can skip invalidating device 2501 * private page mappings that won't be migrated. 2502 */ 2503 mmu_notifier_range_init_migrate(&range, 0, migrate->vma, 2504 migrate->vma->vm_mm, migrate->start, migrate->end, 2505 migrate->pgmap_owner); 2506 mmu_notifier_invalidate_range_start(&range); 2507 2508 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, 2509 &migrate_vma_walk_ops, migrate); 2510 2511 mmu_notifier_invalidate_range_end(&range); 2512 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); 2513 } 2514 2515 /* 2516 * migrate_vma_check_page() - check if page is pinned or not 2517 * @page: struct page to check 2518 * 2519 * Pinned pages cannot be migrated. This is the same test as in 2520 * migrate_page_move_mapping(), except that here we allow migration of a 2521 * ZONE_DEVICE page. 2522 */ 2523 static bool migrate_vma_check_page(struct page *page) 2524 { 2525 /* 2526 * One extra ref because caller holds an extra reference, either from 2527 * isolate_lru_page() for a regular page, or migrate_vma_collect() for 2528 * a device page. 2529 */ 2530 int extra = 1; 2531 2532 /* 2533 * FIXME support THP (transparent huge page), it is bit more complex to 2534 * check them than regular pages, because they can be mapped with a pmd 2535 * or with a pte (split pte mapping). 2536 */ 2537 if (PageCompound(page)) 2538 return false; 2539 2540 /* Page from ZONE_DEVICE have one extra reference */ 2541 if (is_zone_device_page(page)) { 2542 /* 2543 * Private page can never be pin as they have no valid pte and 2544 * GUP will fail for those. Yet if there is a pending migration 2545 * a thread might try to wait on the pte migration entry and 2546 * will bump the page reference count. Sadly there is no way to 2547 * differentiate a regular pin from migration wait. Hence to 2548 * avoid 2 racing thread trying to migrate back to CPU to enter 2549 * infinite loop (one stoping migration because the other is 2550 * waiting on pte migration entry). We always return true here. 2551 * 2552 * FIXME proper solution is to rework migration_entry_wait() so 2553 * it does not need to take a reference on page. 2554 */ 2555 return is_device_private_page(page); 2556 } 2557 2558 /* For file back page */ 2559 if (page_mapping(page)) 2560 extra += 1 + page_has_private(page); 2561 2562 if ((page_count(page) - extra) > page_mapcount(page)) 2563 return false; 2564 2565 return true; 2566 } 2567 2568 /* 2569 * migrate_vma_prepare() - lock pages and isolate them from the lru 2570 * @migrate: migrate struct containing all migration information 2571 * 2572 * This locks pages that have been collected by migrate_vma_collect(). Once each 2573 * page is locked it is isolated from the lru (for non-device pages). Finally, 2574 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be 2575 * migrated by concurrent kernel threads. 2576 */ 2577 static void migrate_vma_prepare(struct migrate_vma *migrate) 2578 { 2579 const unsigned long npages = migrate->npages; 2580 const unsigned long start = migrate->start; 2581 unsigned long addr, i, restore = 0; 2582 bool allow_drain = true; 2583 2584 lru_add_drain(); 2585 2586 for (i = 0; (i < npages) && migrate->cpages; i++) { 2587 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2588 bool remap = true; 2589 2590 if (!page) 2591 continue; 2592 2593 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { 2594 /* 2595 * Because we are migrating several pages there can be 2596 * a deadlock between 2 concurrent migration where each 2597 * are waiting on each other page lock. 2598 * 2599 * Make migrate_vma() a best effort thing and backoff 2600 * for any page we can not lock right away. 2601 */ 2602 if (!trylock_page(page)) { 2603 migrate->src[i] = 0; 2604 migrate->cpages--; 2605 put_page(page); 2606 continue; 2607 } 2608 remap = false; 2609 migrate->src[i] |= MIGRATE_PFN_LOCKED; 2610 } 2611 2612 /* ZONE_DEVICE pages are not on LRU */ 2613 if (!is_zone_device_page(page)) { 2614 if (!PageLRU(page) && allow_drain) { 2615 /* Drain CPU's pagevec */ 2616 lru_add_drain_all(); 2617 allow_drain = false; 2618 } 2619 2620 if (isolate_lru_page(page)) { 2621 if (remap) { 2622 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2623 migrate->cpages--; 2624 restore++; 2625 } else { 2626 migrate->src[i] = 0; 2627 unlock_page(page); 2628 migrate->cpages--; 2629 put_page(page); 2630 } 2631 continue; 2632 } 2633 2634 /* Drop the reference we took in collect */ 2635 put_page(page); 2636 } 2637 2638 if (!migrate_vma_check_page(page)) { 2639 if (remap) { 2640 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2641 migrate->cpages--; 2642 restore++; 2643 2644 if (!is_zone_device_page(page)) { 2645 get_page(page); 2646 putback_lru_page(page); 2647 } 2648 } else { 2649 migrate->src[i] = 0; 2650 unlock_page(page); 2651 migrate->cpages--; 2652 2653 if (!is_zone_device_page(page)) 2654 putback_lru_page(page); 2655 else 2656 put_page(page); 2657 } 2658 } 2659 } 2660 2661 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { 2662 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2663 2664 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2665 continue; 2666 2667 remove_migration_pte(page, migrate->vma, addr, page); 2668 2669 migrate->src[i] = 0; 2670 unlock_page(page); 2671 put_page(page); 2672 restore--; 2673 } 2674 } 2675 2676 /* 2677 * migrate_vma_unmap() - replace page mapping with special migration pte entry 2678 * @migrate: migrate struct containing all migration information 2679 * 2680 * Replace page mapping (CPU page table pte) with a special migration pte entry 2681 * and check again if it has been pinned. Pinned pages are restored because we 2682 * cannot migrate them. 2683 * 2684 * This is the last step before we call the device driver callback to allocate 2685 * destination memory and copy contents of original page over to new page. 2686 */ 2687 static void migrate_vma_unmap(struct migrate_vma *migrate) 2688 { 2689 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK; 2690 const unsigned long npages = migrate->npages; 2691 const unsigned long start = migrate->start; 2692 unsigned long addr, i, restore = 0; 2693 2694 for (i = 0; i < npages; i++) { 2695 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2696 2697 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2698 continue; 2699 2700 if (page_mapped(page)) { 2701 try_to_unmap(page, flags); 2702 if (page_mapped(page)) 2703 goto restore; 2704 } 2705 2706 if (migrate_vma_check_page(page)) 2707 continue; 2708 2709 restore: 2710 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2711 migrate->cpages--; 2712 restore++; 2713 } 2714 2715 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { 2716 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2717 2718 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2719 continue; 2720 2721 remove_migration_ptes(page, page, false); 2722 2723 migrate->src[i] = 0; 2724 unlock_page(page); 2725 restore--; 2726 2727 if (is_zone_device_page(page)) 2728 put_page(page); 2729 else 2730 putback_lru_page(page); 2731 } 2732 } 2733 2734 /** 2735 * migrate_vma_setup() - prepare to migrate a range of memory 2736 * @args: contains the vma, start, and pfns arrays for the migration 2737 * 2738 * Returns: negative errno on failures, 0 when 0 or more pages were migrated 2739 * without an error. 2740 * 2741 * Prepare to migrate a range of memory virtual address range by collecting all 2742 * the pages backing each virtual address in the range, saving them inside the 2743 * src array. Then lock those pages and unmap them. Once the pages are locked 2744 * and unmapped, check whether each page is pinned or not. Pages that aren't 2745 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the 2746 * corresponding src array entry. Then restores any pages that are pinned, by 2747 * remapping and unlocking those pages. 2748 * 2749 * The caller should then allocate destination memory and copy source memory to 2750 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE 2751 * flag set). Once these are allocated and copied, the caller must update each 2752 * corresponding entry in the dst array with the pfn value of the destination 2753 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set 2754 * (destination pages must have their struct pages locked, via lock_page()). 2755 * 2756 * Note that the caller does not have to migrate all the pages that are marked 2757 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from 2758 * device memory to system memory. If the caller cannot migrate a device page 2759 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe 2760 * consequences for the userspace process, so it must be avoided if at all 2761 * possible. 2762 * 2763 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we 2764 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus 2765 * allowing the caller to allocate device memory for those unback virtual 2766 * address. For this the caller simply has to allocate device memory and 2767 * properly set the destination entry like for regular migration. Note that 2768 * this can still fails and thus inside the device driver must check if the 2769 * migration was successful for those entries after calling migrate_vma_pages() 2770 * just like for regular migration. 2771 * 2772 * After that, the callers must call migrate_vma_pages() to go over each entry 2773 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag 2774 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, 2775 * then migrate_vma_pages() to migrate struct page information from the source 2776 * struct page to the destination struct page. If it fails to migrate the 2777 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the 2778 * src array. 2779 * 2780 * At this point all successfully migrated pages have an entry in the src 2781 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst 2782 * array entry with MIGRATE_PFN_VALID flag set. 2783 * 2784 * Once migrate_vma_pages() returns the caller may inspect which pages were 2785 * successfully migrated, and which were not. Successfully migrated pages will 2786 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. 2787 * 2788 * It is safe to update device page table after migrate_vma_pages() because 2789 * both destination and source page are still locked, and the mmap_lock is held 2790 * in read mode (hence no one can unmap the range being migrated). 2791 * 2792 * Once the caller is done cleaning up things and updating its page table (if it 2793 * chose to do so, this is not an obligation) it finally calls 2794 * migrate_vma_finalize() to update the CPU page table to point to new pages 2795 * for successfully migrated pages or otherwise restore the CPU page table to 2796 * point to the original source pages. 2797 */ 2798 int migrate_vma_setup(struct migrate_vma *args) 2799 { 2800 long nr_pages = (args->end - args->start) >> PAGE_SHIFT; 2801 2802 args->start &= PAGE_MASK; 2803 args->end &= PAGE_MASK; 2804 if (!args->vma || is_vm_hugetlb_page(args->vma) || 2805 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) 2806 return -EINVAL; 2807 if (nr_pages <= 0) 2808 return -EINVAL; 2809 if (args->start < args->vma->vm_start || 2810 args->start >= args->vma->vm_end) 2811 return -EINVAL; 2812 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) 2813 return -EINVAL; 2814 if (!args->src || !args->dst) 2815 return -EINVAL; 2816 2817 memset(args->src, 0, sizeof(*args->src) * nr_pages); 2818 args->cpages = 0; 2819 args->npages = 0; 2820 2821 migrate_vma_collect(args); 2822 2823 if (args->cpages) 2824 migrate_vma_prepare(args); 2825 if (args->cpages) 2826 migrate_vma_unmap(args); 2827 2828 /* 2829 * At this point pages are locked and unmapped, and thus they have 2830 * stable content and can safely be copied to destination memory that 2831 * is allocated by the drivers. 2832 */ 2833 return 0; 2834 2835 } 2836 EXPORT_SYMBOL(migrate_vma_setup); 2837 2838 /* 2839 * This code closely matches the code in: 2840 * __handle_mm_fault() 2841 * handle_pte_fault() 2842 * do_anonymous_page() 2843 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE 2844 * private page. 2845 */ 2846 static void migrate_vma_insert_page(struct migrate_vma *migrate, 2847 unsigned long addr, 2848 struct page *page, 2849 unsigned long *src, 2850 unsigned long *dst) 2851 { 2852 struct vm_area_struct *vma = migrate->vma; 2853 struct mm_struct *mm = vma->vm_mm; 2854 bool flush = false; 2855 spinlock_t *ptl; 2856 pte_t entry; 2857 pgd_t *pgdp; 2858 p4d_t *p4dp; 2859 pud_t *pudp; 2860 pmd_t *pmdp; 2861 pte_t *ptep; 2862 2863 /* Only allow populating anonymous memory */ 2864 if (!vma_is_anonymous(vma)) 2865 goto abort; 2866 2867 pgdp = pgd_offset(mm, addr); 2868 p4dp = p4d_alloc(mm, pgdp, addr); 2869 if (!p4dp) 2870 goto abort; 2871 pudp = pud_alloc(mm, p4dp, addr); 2872 if (!pudp) 2873 goto abort; 2874 pmdp = pmd_alloc(mm, pudp, addr); 2875 if (!pmdp) 2876 goto abort; 2877 2878 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) 2879 goto abort; 2880 2881 /* 2882 * Use pte_alloc() instead of pte_alloc_map(). We can't run 2883 * pte_offset_map() on pmds where a huge pmd might be created 2884 * from a different thread. 2885 * 2886 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 2887 * parallel threads are excluded by other means. 2888 * 2889 * Here we only have mmap_read_lock(mm). 2890 */ 2891 if (pte_alloc(mm, pmdp)) 2892 goto abort; 2893 2894 /* See the comment in pte_alloc_one_map() */ 2895 if (unlikely(pmd_trans_unstable(pmdp))) 2896 goto abort; 2897 2898 if (unlikely(anon_vma_prepare(vma))) 2899 goto abort; 2900 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) 2901 goto abort; 2902 2903 /* 2904 * The memory barrier inside __SetPageUptodate makes sure that 2905 * preceding stores to the page contents become visible before 2906 * the set_pte_at() write. 2907 */ 2908 __SetPageUptodate(page); 2909 2910 if (is_zone_device_page(page)) { 2911 if (is_device_private_page(page)) { 2912 swp_entry_t swp_entry; 2913 2914 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); 2915 entry = swp_entry_to_pte(swp_entry); 2916 } 2917 } else { 2918 entry = mk_pte(page, vma->vm_page_prot); 2919 if (vma->vm_flags & VM_WRITE) 2920 entry = pte_mkwrite(pte_mkdirty(entry)); 2921 } 2922 2923 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); 2924 2925 if (check_stable_address_space(mm)) 2926 goto unlock_abort; 2927 2928 if (pte_present(*ptep)) { 2929 unsigned long pfn = pte_pfn(*ptep); 2930 2931 if (!is_zero_pfn(pfn)) 2932 goto unlock_abort; 2933 flush = true; 2934 } else if (!pte_none(*ptep)) 2935 goto unlock_abort; 2936 2937 /* 2938 * Check for userfaultfd but do not deliver the fault. Instead, 2939 * just back off. 2940 */ 2941 if (userfaultfd_missing(vma)) 2942 goto unlock_abort; 2943 2944 inc_mm_counter(mm, MM_ANONPAGES); 2945 page_add_new_anon_rmap(page, vma, addr, false); 2946 if (!is_zone_device_page(page)) 2947 lru_cache_add_inactive_or_unevictable(page, vma); 2948 get_page(page); 2949 2950 if (flush) { 2951 flush_cache_page(vma, addr, pte_pfn(*ptep)); 2952 ptep_clear_flush_notify(vma, addr, ptep); 2953 set_pte_at_notify(mm, addr, ptep, entry); 2954 update_mmu_cache(vma, addr, ptep); 2955 } else { 2956 /* No need to invalidate - it was non-present before */ 2957 set_pte_at(mm, addr, ptep, entry); 2958 update_mmu_cache(vma, addr, ptep); 2959 } 2960 2961 pte_unmap_unlock(ptep, ptl); 2962 *src = MIGRATE_PFN_MIGRATE; 2963 return; 2964 2965 unlock_abort: 2966 pte_unmap_unlock(ptep, ptl); 2967 abort: 2968 *src &= ~MIGRATE_PFN_MIGRATE; 2969 } 2970 2971 /** 2972 * migrate_vma_pages() - migrate meta-data from src page to dst page 2973 * @migrate: migrate struct containing all migration information 2974 * 2975 * This migrates struct page meta-data from source struct page to destination 2976 * struct page. This effectively finishes the migration from source page to the 2977 * destination page. 2978 */ 2979 void migrate_vma_pages(struct migrate_vma *migrate) 2980 { 2981 const unsigned long npages = migrate->npages; 2982 const unsigned long start = migrate->start; 2983 struct mmu_notifier_range range; 2984 unsigned long addr, i; 2985 bool notified = false; 2986 2987 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { 2988 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); 2989 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2990 struct address_space *mapping; 2991 int r; 2992 2993 if (!newpage) { 2994 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2995 continue; 2996 } 2997 2998 if (!page) { 2999 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 3000 continue; 3001 if (!notified) { 3002 notified = true; 3003 3004 mmu_notifier_range_init(&range, 3005 MMU_NOTIFY_CLEAR, 0, 3006 NULL, 3007 migrate->vma->vm_mm, 3008 addr, migrate->end); 3009 mmu_notifier_invalidate_range_start(&range); 3010 } 3011 migrate_vma_insert_page(migrate, addr, newpage, 3012 &migrate->src[i], 3013 &migrate->dst[i]); 3014 continue; 3015 } 3016 3017 mapping = page_mapping(page); 3018 3019 if (is_zone_device_page(newpage)) { 3020 if (is_device_private_page(newpage)) { 3021 /* 3022 * For now only support private anonymous when 3023 * migrating to un-addressable device memory. 3024 */ 3025 if (mapping) { 3026 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3027 continue; 3028 } 3029 } else { 3030 /* 3031 * Other types of ZONE_DEVICE page are not 3032 * supported. 3033 */ 3034 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3035 continue; 3036 } 3037 } 3038 3039 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); 3040 if (r != MIGRATEPAGE_SUCCESS) 3041 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3042 } 3043 3044 /* 3045 * No need to double call mmu_notifier->invalidate_range() callback as 3046 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() 3047 * did already call it. 3048 */ 3049 if (notified) 3050 mmu_notifier_invalidate_range_only_end(&range); 3051 } 3052 EXPORT_SYMBOL(migrate_vma_pages); 3053 3054 /** 3055 * migrate_vma_finalize() - restore CPU page table entry 3056 * @migrate: migrate struct containing all migration information 3057 * 3058 * This replaces the special migration pte entry with either a mapping to the 3059 * new page if migration was successful for that page, or to the original page 3060 * otherwise. 3061 * 3062 * This also unlocks the pages and puts them back on the lru, or drops the extra 3063 * refcount, for device pages. 3064 */ 3065 void migrate_vma_finalize(struct migrate_vma *migrate) 3066 { 3067 const unsigned long npages = migrate->npages; 3068 unsigned long i; 3069 3070 for (i = 0; i < npages; i++) { 3071 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); 3072 struct page *page = migrate_pfn_to_page(migrate->src[i]); 3073 3074 if (!page) { 3075 if (newpage) { 3076 unlock_page(newpage); 3077 put_page(newpage); 3078 } 3079 continue; 3080 } 3081 3082 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { 3083 if (newpage) { 3084 unlock_page(newpage); 3085 put_page(newpage); 3086 } 3087 newpage = page; 3088 } 3089 3090 remove_migration_ptes(page, newpage, false); 3091 unlock_page(page); 3092 3093 if (is_zone_device_page(page)) 3094 put_page(page); 3095 else 3096 putback_lru_page(page); 3097 3098 if (newpage != page) { 3099 unlock_page(newpage); 3100 if (is_zone_device_page(newpage)) 3101 put_page(newpage); 3102 else 3103 putback_lru_page(newpage); 3104 } 3105 } 3106 } 3107 EXPORT_SYMBOL(migrate_vma_finalize); 3108 #endif /* CONFIG_DEVICE_PRIVATE */ 3109