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 int nr = thp_nr_pages(page); 410 411 if (!mapping) { 412 /* Anonymous page without mapping */ 413 if (page_count(page) != expected_count) 414 return -EAGAIN; 415 416 /* No turning back from here */ 417 newpage->index = page->index; 418 newpage->mapping = page->mapping; 419 if (PageSwapBacked(page)) 420 __SetPageSwapBacked(newpage); 421 422 return MIGRATEPAGE_SUCCESS; 423 } 424 425 oldzone = page_zone(page); 426 newzone = page_zone(newpage); 427 428 xas_lock_irq(&xas); 429 if (page_count(page) != expected_count || xas_load(&xas) != page) { 430 xas_unlock_irq(&xas); 431 return -EAGAIN; 432 } 433 434 if (!page_ref_freeze(page, expected_count)) { 435 xas_unlock_irq(&xas); 436 return -EAGAIN; 437 } 438 439 /* 440 * Now we know that no one else is looking at the page: 441 * no turning back from here. 442 */ 443 newpage->index = page->index; 444 newpage->mapping = page->mapping; 445 page_ref_add(newpage, nr); /* add cache reference */ 446 if (PageSwapBacked(page)) { 447 __SetPageSwapBacked(newpage); 448 if (PageSwapCache(page)) { 449 SetPageSwapCache(newpage); 450 set_page_private(newpage, page_private(page)); 451 } 452 } else { 453 VM_BUG_ON_PAGE(PageSwapCache(page), page); 454 } 455 456 /* Move dirty while page refs frozen and newpage not yet exposed */ 457 dirty = PageDirty(page); 458 if (dirty) { 459 ClearPageDirty(page); 460 SetPageDirty(newpage); 461 } 462 463 xas_store(&xas, newpage); 464 if (PageTransHuge(page)) { 465 int i; 466 467 for (i = 1; i < nr; i++) { 468 xas_next(&xas); 469 xas_store(&xas, newpage); 470 } 471 } 472 473 /* 474 * Drop cache reference from old page by unfreezing 475 * to one less reference. 476 * We know this isn't the last reference. 477 */ 478 page_ref_unfreeze(page, expected_count - nr); 479 480 xas_unlock(&xas); 481 /* Leave irq disabled to prevent preemption while updating stats */ 482 483 /* 484 * If moved to a different zone then also account 485 * the page for that zone. Other VM counters will be 486 * taken care of when we establish references to the 487 * new page and drop references to the old page. 488 * 489 * Note that anonymous pages are accounted for 490 * via NR_FILE_PAGES and NR_ANON_MAPPED if they 491 * are mapped to swap space. 492 */ 493 if (newzone != oldzone) { 494 struct lruvec *old_lruvec, *new_lruvec; 495 struct mem_cgroup *memcg; 496 497 memcg = page_memcg(page); 498 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat); 499 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat); 500 501 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr); 502 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr); 503 if (PageSwapBacked(page) && !PageSwapCache(page)) { 504 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr); 505 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr); 506 } 507 if (dirty && mapping_can_writeback(mapping)) { 508 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr); 509 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr); 510 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr); 511 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr); 512 } 513 } 514 local_irq_enable(); 515 516 return MIGRATEPAGE_SUCCESS; 517 } 518 EXPORT_SYMBOL(migrate_page_move_mapping); 519 520 /* 521 * The expected number of remaining references is the same as that 522 * of migrate_page_move_mapping(). 523 */ 524 int migrate_huge_page_move_mapping(struct address_space *mapping, 525 struct page *newpage, struct page *page) 526 { 527 XA_STATE(xas, &mapping->i_pages, page_index(page)); 528 int expected_count; 529 530 xas_lock_irq(&xas); 531 expected_count = 2 + page_has_private(page); 532 if (page_count(page) != expected_count || xas_load(&xas) != page) { 533 xas_unlock_irq(&xas); 534 return -EAGAIN; 535 } 536 537 if (!page_ref_freeze(page, expected_count)) { 538 xas_unlock_irq(&xas); 539 return -EAGAIN; 540 } 541 542 newpage->index = page->index; 543 newpage->mapping = page->mapping; 544 545 get_page(newpage); 546 547 xas_store(&xas, newpage); 548 549 page_ref_unfreeze(page, expected_count - 1); 550 551 xas_unlock_irq(&xas); 552 553 return MIGRATEPAGE_SUCCESS; 554 } 555 556 /* 557 * Gigantic pages are so large that we do not guarantee that page++ pointer 558 * arithmetic will work across the entire page. We need something more 559 * specialized. 560 */ 561 static void __copy_gigantic_page(struct page *dst, struct page *src, 562 int nr_pages) 563 { 564 int i; 565 struct page *dst_base = dst; 566 struct page *src_base = src; 567 568 for (i = 0; i < nr_pages; ) { 569 cond_resched(); 570 copy_highpage(dst, src); 571 572 i++; 573 dst = mem_map_next(dst, dst_base, i); 574 src = mem_map_next(src, src_base, i); 575 } 576 } 577 578 static void copy_huge_page(struct page *dst, struct page *src) 579 { 580 int i; 581 int nr_pages; 582 583 if (PageHuge(src)) { 584 /* hugetlbfs page */ 585 struct hstate *h = page_hstate(src); 586 nr_pages = pages_per_huge_page(h); 587 588 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) { 589 __copy_gigantic_page(dst, src, nr_pages); 590 return; 591 } 592 } else { 593 /* thp page */ 594 BUG_ON(!PageTransHuge(src)); 595 nr_pages = thp_nr_pages(src); 596 } 597 598 for (i = 0; i < nr_pages; i++) { 599 cond_resched(); 600 copy_highpage(dst + i, src + i); 601 } 602 } 603 604 /* 605 * Copy the page to its new location 606 */ 607 void migrate_page_states(struct page *newpage, struct page *page) 608 { 609 int cpupid; 610 611 if (PageError(page)) 612 SetPageError(newpage); 613 if (PageReferenced(page)) 614 SetPageReferenced(newpage); 615 if (PageUptodate(page)) 616 SetPageUptodate(newpage); 617 if (TestClearPageActive(page)) { 618 VM_BUG_ON_PAGE(PageUnevictable(page), page); 619 SetPageActive(newpage); 620 } else if (TestClearPageUnevictable(page)) 621 SetPageUnevictable(newpage); 622 if (PageWorkingset(page)) 623 SetPageWorkingset(newpage); 624 if (PageChecked(page)) 625 SetPageChecked(newpage); 626 if (PageMappedToDisk(page)) 627 SetPageMappedToDisk(newpage); 628 629 /* Move dirty on pages not done by migrate_page_move_mapping() */ 630 if (PageDirty(page)) 631 SetPageDirty(newpage); 632 633 if (page_is_young(page)) 634 set_page_young(newpage); 635 if (page_is_idle(page)) 636 set_page_idle(newpage); 637 638 /* 639 * Copy NUMA information to the new page, to prevent over-eager 640 * future migrations of this same page. 641 */ 642 cpupid = page_cpupid_xchg_last(page, -1); 643 page_cpupid_xchg_last(newpage, cpupid); 644 645 ksm_migrate_page(newpage, page); 646 /* 647 * Please do not reorder this without considering how mm/ksm.c's 648 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). 649 */ 650 if (PageSwapCache(page)) 651 ClearPageSwapCache(page); 652 ClearPagePrivate(page); 653 set_page_private(page, 0); 654 655 /* 656 * If any waiters have accumulated on the new page then 657 * wake them up. 658 */ 659 if (PageWriteback(newpage)) 660 end_page_writeback(newpage); 661 662 /* 663 * PG_readahead shares the same bit with PG_reclaim. The above 664 * end_page_writeback() may clear PG_readahead mistakenly, so set the 665 * bit after that. 666 */ 667 if (PageReadahead(page)) 668 SetPageReadahead(newpage); 669 670 copy_page_owner(page, newpage); 671 672 if (!PageHuge(page)) 673 mem_cgroup_migrate(page, newpage); 674 } 675 EXPORT_SYMBOL(migrate_page_states); 676 677 void migrate_page_copy(struct page *newpage, struct page *page) 678 { 679 if (PageHuge(page) || PageTransHuge(page)) 680 copy_huge_page(newpage, page); 681 else 682 copy_highpage(newpage, page); 683 684 migrate_page_states(newpage, page); 685 } 686 EXPORT_SYMBOL(migrate_page_copy); 687 688 /************************************************************ 689 * Migration functions 690 ***********************************************************/ 691 692 /* 693 * Common logic to directly migrate a single LRU page suitable for 694 * pages that do not use PagePrivate/PagePrivate2. 695 * 696 * Pages are locked upon entry and exit. 697 */ 698 int migrate_page(struct address_space *mapping, 699 struct page *newpage, struct page *page, 700 enum migrate_mode mode) 701 { 702 int rc; 703 704 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 705 706 rc = migrate_page_move_mapping(mapping, newpage, page, 0); 707 708 if (rc != MIGRATEPAGE_SUCCESS) 709 return rc; 710 711 if (mode != MIGRATE_SYNC_NO_COPY) 712 migrate_page_copy(newpage, page); 713 else 714 migrate_page_states(newpage, page); 715 return MIGRATEPAGE_SUCCESS; 716 } 717 EXPORT_SYMBOL(migrate_page); 718 719 #ifdef CONFIG_BLOCK 720 /* Returns true if all buffers are successfully locked */ 721 static bool buffer_migrate_lock_buffers(struct buffer_head *head, 722 enum migrate_mode mode) 723 { 724 struct buffer_head *bh = head; 725 726 /* Simple case, sync compaction */ 727 if (mode != MIGRATE_ASYNC) { 728 do { 729 lock_buffer(bh); 730 bh = bh->b_this_page; 731 732 } while (bh != head); 733 734 return true; 735 } 736 737 /* async case, we cannot block on lock_buffer so use trylock_buffer */ 738 do { 739 if (!trylock_buffer(bh)) { 740 /* 741 * We failed to lock the buffer and cannot stall in 742 * async migration. Release the taken locks 743 */ 744 struct buffer_head *failed_bh = bh; 745 bh = head; 746 while (bh != failed_bh) { 747 unlock_buffer(bh); 748 bh = bh->b_this_page; 749 } 750 return false; 751 } 752 753 bh = bh->b_this_page; 754 } while (bh != head); 755 return true; 756 } 757 758 static int __buffer_migrate_page(struct address_space *mapping, 759 struct page *newpage, struct page *page, enum migrate_mode mode, 760 bool check_refs) 761 { 762 struct buffer_head *bh, *head; 763 int rc; 764 int expected_count; 765 766 if (!page_has_buffers(page)) 767 return migrate_page(mapping, newpage, page, mode); 768 769 /* Check whether page does not have extra refs before we do more work */ 770 expected_count = expected_page_refs(mapping, page); 771 if (page_count(page) != expected_count) 772 return -EAGAIN; 773 774 head = page_buffers(page); 775 if (!buffer_migrate_lock_buffers(head, mode)) 776 return -EAGAIN; 777 778 if (check_refs) { 779 bool busy; 780 bool invalidated = false; 781 782 recheck_buffers: 783 busy = false; 784 spin_lock(&mapping->private_lock); 785 bh = head; 786 do { 787 if (atomic_read(&bh->b_count)) { 788 busy = true; 789 break; 790 } 791 bh = bh->b_this_page; 792 } while (bh != head); 793 if (busy) { 794 if (invalidated) { 795 rc = -EAGAIN; 796 goto unlock_buffers; 797 } 798 spin_unlock(&mapping->private_lock); 799 invalidate_bh_lrus(); 800 invalidated = true; 801 goto recheck_buffers; 802 } 803 } 804 805 rc = migrate_page_move_mapping(mapping, newpage, page, 0); 806 if (rc != MIGRATEPAGE_SUCCESS) 807 goto unlock_buffers; 808 809 attach_page_private(newpage, detach_page_private(page)); 810 811 bh = head; 812 do { 813 set_bh_page(bh, newpage, bh_offset(bh)); 814 bh = bh->b_this_page; 815 816 } while (bh != head); 817 818 if (mode != MIGRATE_SYNC_NO_COPY) 819 migrate_page_copy(newpage, page); 820 else 821 migrate_page_states(newpage, page); 822 823 rc = MIGRATEPAGE_SUCCESS; 824 unlock_buffers: 825 if (check_refs) 826 spin_unlock(&mapping->private_lock); 827 bh = head; 828 do { 829 unlock_buffer(bh); 830 bh = bh->b_this_page; 831 832 } while (bh != head); 833 834 return rc; 835 } 836 837 /* 838 * Migration function for pages with buffers. This function can only be used 839 * if the underlying filesystem guarantees that no other references to "page" 840 * exist. For example attached buffer heads are accessed only under page lock. 841 */ 842 int buffer_migrate_page(struct address_space *mapping, 843 struct page *newpage, struct page *page, enum migrate_mode mode) 844 { 845 return __buffer_migrate_page(mapping, newpage, page, mode, false); 846 } 847 EXPORT_SYMBOL(buffer_migrate_page); 848 849 /* 850 * Same as above except that this variant is more careful and checks that there 851 * are also no buffer head references. This function is the right one for 852 * mappings where buffer heads are directly looked up and referenced (such as 853 * block device mappings). 854 */ 855 int buffer_migrate_page_norefs(struct address_space *mapping, 856 struct page *newpage, struct page *page, enum migrate_mode mode) 857 { 858 return __buffer_migrate_page(mapping, newpage, page, mode, true); 859 } 860 #endif 861 862 /* 863 * Writeback a page to clean the dirty state 864 */ 865 static int writeout(struct address_space *mapping, struct page *page) 866 { 867 struct writeback_control wbc = { 868 .sync_mode = WB_SYNC_NONE, 869 .nr_to_write = 1, 870 .range_start = 0, 871 .range_end = LLONG_MAX, 872 .for_reclaim = 1 873 }; 874 int rc; 875 876 if (!mapping->a_ops->writepage) 877 /* No write method for the address space */ 878 return -EINVAL; 879 880 if (!clear_page_dirty_for_io(page)) 881 /* Someone else already triggered a write */ 882 return -EAGAIN; 883 884 /* 885 * A dirty page may imply that the underlying filesystem has 886 * the page on some queue. So the page must be clean for 887 * migration. Writeout may mean we loose the lock and the 888 * page state is no longer what we checked for earlier. 889 * At this point we know that the migration attempt cannot 890 * be successful. 891 */ 892 remove_migration_ptes(page, page, false); 893 894 rc = mapping->a_ops->writepage(page, &wbc); 895 896 if (rc != AOP_WRITEPAGE_ACTIVATE) 897 /* unlocked. Relock */ 898 lock_page(page); 899 900 return (rc < 0) ? -EIO : -EAGAIN; 901 } 902 903 /* 904 * Default handling if a filesystem does not provide a migration function. 905 */ 906 static int fallback_migrate_page(struct address_space *mapping, 907 struct page *newpage, struct page *page, enum migrate_mode mode) 908 { 909 if (PageDirty(page)) { 910 /* Only writeback pages in full synchronous migration */ 911 switch (mode) { 912 case MIGRATE_SYNC: 913 case MIGRATE_SYNC_NO_COPY: 914 break; 915 default: 916 return -EBUSY; 917 } 918 return writeout(mapping, page); 919 } 920 921 /* 922 * Buffers may be managed in a filesystem specific way. 923 * We must have no buffers or drop them. 924 */ 925 if (page_has_private(page) && 926 !try_to_release_page(page, GFP_KERNEL)) 927 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY; 928 929 return migrate_page(mapping, newpage, page, mode); 930 } 931 932 /* 933 * Move a page to a newly allocated page 934 * The page is locked and all ptes have been successfully removed. 935 * 936 * The new page will have replaced the old page if this function 937 * is successful. 938 * 939 * Return value: 940 * < 0 - error code 941 * MIGRATEPAGE_SUCCESS - success 942 */ 943 static int move_to_new_page(struct page *newpage, struct page *page, 944 enum migrate_mode mode) 945 { 946 struct address_space *mapping; 947 int rc = -EAGAIN; 948 bool is_lru = !__PageMovable(page); 949 950 VM_BUG_ON_PAGE(!PageLocked(page), page); 951 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); 952 953 mapping = page_mapping(page); 954 955 if (likely(is_lru)) { 956 if (!mapping) 957 rc = migrate_page(mapping, newpage, page, mode); 958 else if (mapping->a_ops->migratepage) 959 /* 960 * Most pages have a mapping and most filesystems 961 * provide a migratepage callback. Anonymous pages 962 * are part of swap space which also has its own 963 * migratepage callback. This is the most common path 964 * for page migration. 965 */ 966 rc = mapping->a_ops->migratepage(mapping, newpage, 967 page, mode); 968 else 969 rc = fallback_migrate_page(mapping, newpage, 970 page, mode); 971 } else { 972 /* 973 * In case of non-lru page, it could be released after 974 * isolation step. In that case, we shouldn't try migration. 975 */ 976 VM_BUG_ON_PAGE(!PageIsolated(page), page); 977 if (!PageMovable(page)) { 978 rc = MIGRATEPAGE_SUCCESS; 979 __ClearPageIsolated(page); 980 goto out; 981 } 982 983 rc = mapping->a_ops->migratepage(mapping, newpage, 984 page, mode); 985 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS && 986 !PageIsolated(page)); 987 } 988 989 /* 990 * When successful, old pagecache page->mapping must be cleared before 991 * page is freed; but stats require that PageAnon be left as PageAnon. 992 */ 993 if (rc == MIGRATEPAGE_SUCCESS) { 994 if (__PageMovable(page)) { 995 VM_BUG_ON_PAGE(!PageIsolated(page), page); 996 997 /* 998 * We clear PG_movable under page_lock so any compactor 999 * cannot try to migrate this page. 1000 */ 1001 __ClearPageIsolated(page); 1002 } 1003 1004 /* 1005 * Anonymous and movable page->mapping will be cleared by 1006 * free_pages_prepare so don't reset it here for keeping 1007 * the type to work PageAnon, for example. 1008 */ 1009 if (!PageMappingFlags(page)) 1010 page->mapping = NULL; 1011 1012 if (likely(!is_zone_device_page(newpage))) 1013 flush_dcache_page(newpage); 1014 1015 } 1016 out: 1017 return rc; 1018 } 1019 1020 static int __unmap_and_move(struct page *page, struct page *newpage, 1021 int force, enum migrate_mode mode) 1022 { 1023 int rc = -EAGAIN; 1024 int page_was_mapped = 0; 1025 struct anon_vma *anon_vma = NULL; 1026 bool is_lru = !__PageMovable(page); 1027 1028 if (!trylock_page(page)) { 1029 if (!force || mode == MIGRATE_ASYNC) 1030 goto out; 1031 1032 /* 1033 * It's not safe for direct compaction to call lock_page. 1034 * For example, during page readahead pages are added locked 1035 * to the LRU. Later, when the IO completes the pages are 1036 * marked uptodate and unlocked. However, the queueing 1037 * could be merging multiple pages for one bio (e.g. 1038 * mpage_readahead). If an allocation happens for the 1039 * second or third page, the process can end up locking 1040 * the same page twice and deadlocking. Rather than 1041 * trying to be clever about what pages can be locked, 1042 * avoid the use of lock_page for direct compaction 1043 * altogether. 1044 */ 1045 if (current->flags & PF_MEMALLOC) 1046 goto out; 1047 1048 lock_page(page); 1049 } 1050 1051 if (PageWriteback(page)) { 1052 /* 1053 * Only in the case of a full synchronous migration is it 1054 * necessary to wait for PageWriteback. In the async case, 1055 * the retry loop is too short and in the sync-light case, 1056 * the overhead of stalling is too much 1057 */ 1058 switch (mode) { 1059 case MIGRATE_SYNC: 1060 case MIGRATE_SYNC_NO_COPY: 1061 break; 1062 default: 1063 rc = -EBUSY; 1064 goto out_unlock; 1065 } 1066 if (!force) 1067 goto out_unlock; 1068 wait_on_page_writeback(page); 1069 } 1070 1071 /* 1072 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, 1073 * we cannot notice that anon_vma is freed while we migrates a page. 1074 * This get_anon_vma() delays freeing anon_vma pointer until the end 1075 * of migration. File cache pages are no problem because of page_lock() 1076 * File Caches may use write_page() or lock_page() in migration, then, 1077 * just care Anon page here. 1078 * 1079 * Only page_get_anon_vma() understands the subtleties of 1080 * getting a hold on an anon_vma from outside one of its mms. 1081 * But if we cannot get anon_vma, then we won't need it anyway, 1082 * because that implies that the anon page is no longer mapped 1083 * (and cannot be remapped so long as we hold the page lock). 1084 */ 1085 if (PageAnon(page) && !PageKsm(page)) 1086 anon_vma = page_get_anon_vma(page); 1087 1088 /* 1089 * Block others from accessing the new page when we get around to 1090 * establishing additional references. We are usually the only one 1091 * holding a reference to newpage at this point. We used to have a BUG 1092 * here if trylock_page(newpage) fails, but would like to allow for 1093 * cases where there might be a race with the previous use of newpage. 1094 * This is much like races on refcount of oldpage: just don't BUG(). 1095 */ 1096 if (unlikely(!trylock_page(newpage))) 1097 goto out_unlock; 1098 1099 if (unlikely(!is_lru)) { 1100 rc = move_to_new_page(newpage, page, mode); 1101 goto out_unlock_both; 1102 } 1103 1104 /* 1105 * Corner case handling: 1106 * 1. When a new swap-cache page is read into, it is added to the LRU 1107 * and treated as swapcache but it has no rmap yet. 1108 * Calling try_to_unmap() against a page->mapping==NULL page will 1109 * trigger a BUG. So handle it here. 1110 * 2. An orphaned page (see truncate_complete_page) might have 1111 * fs-private metadata. The page can be picked up due to memory 1112 * offlining. Everywhere else except page reclaim, the page is 1113 * invisible to the vm, so the page can not be migrated. So try to 1114 * free the metadata, so the page can be freed. 1115 */ 1116 if (!page->mapping) { 1117 VM_BUG_ON_PAGE(PageAnon(page), page); 1118 if (page_has_private(page)) { 1119 try_to_free_buffers(page); 1120 goto out_unlock_both; 1121 } 1122 } else if (page_mapped(page)) { 1123 /* Establish migration ptes */ 1124 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma, 1125 page); 1126 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK); 1127 page_was_mapped = 1; 1128 } 1129 1130 if (!page_mapped(page)) 1131 rc = move_to_new_page(newpage, page, mode); 1132 1133 if (page_was_mapped) 1134 remove_migration_ptes(page, 1135 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false); 1136 1137 out_unlock_both: 1138 unlock_page(newpage); 1139 out_unlock: 1140 /* Drop an anon_vma reference if we took one */ 1141 if (anon_vma) 1142 put_anon_vma(anon_vma); 1143 unlock_page(page); 1144 out: 1145 /* 1146 * If migration is successful, decrease refcount of the newpage 1147 * which will not free the page because new page owner increased 1148 * refcounter. As well, if it is LRU page, add the page to LRU 1149 * list in here. Use the old state of the isolated source page to 1150 * determine if we migrated a LRU page. newpage was already unlocked 1151 * and possibly modified by its owner - don't rely on the page 1152 * state. 1153 */ 1154 if (rc == MIGRATEPAGE_SUCCESS) { 1155 if (unlikely(!is_lru)) 1156 put_page(newpage); 1157 else 1158 putback_lru_page(newpage); 1159 } 1160 1161 return rc; 1162 } 1163 1164 /* 1165 * Obtain the lock on page, remove all ptes and migrate the page 1166 * to the newly allocated page in newpage. 1167 */ 1168 static int unmap_and_move(new_page_t get_new_page, 1169 free_page_t put_new_page, 1170 unsigned long private, struct page *page, 1171 int force, enum migrate_mode mode, 1172 enum migrate_reason reason) 1173 { 1174 int rc = MIGRATEPAGE_SUCCESS; 1175 struct page *newpage = NULL; 1176 1177 if (!thp_migration_supported() && PageTransHuge(page)) 1178 return -ENOMEM; 1179 1180 if (page_count(page) == 1) { 1181 /* page was freed from under us. So we are done. */ 1182 ClearPageActive(page); 1183 ClearPageUnevictable(page); 1184 if (unlikely(__PageMovable(page))) { 1185 lock_page(page); 1186 if (!PageMovable(page)) 1187 __ClearPageIsolated(page); 1188 unlock_page(page); 1189 } 1190 goto out; 1191 } 1192 1193 newpage = get_new_page(page, private); 1194 if (!newpage) 1195 return -ENOMEM; 1196 1197 rc = __unmap_and_move(page, newpage, force, mode); 1198 if (rc == MIGRATEPAGE_SUCCESS) 1199 set_page_owner_migrate_reason(newpage, reason); 1200 1201 out: 1202 if (rc != -EAGAIN) { 1203 /* 1204 * A page that has been migrated has all references 1205 * removed and will be freed. A page that has not been 1206 * migrated will have kept its references and be restored. 1207 */ 1208 list_del(&page->lru); 1209 1210 /* 1211 * Compaction can migrate also non-LRU pages which are 1212 * not accounted to NR_ISOLATED_*. They can be recognized 1213 * as __PageMovable 1214 */ 1215 if (likely(!__PageMovable(page))) 1216 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + 1217 page_is_file_lru(page), -thp_nr_pages(page)); 1218 } 1219 1220 /* 1221 * If migration is successful, releases reference grabbed during 1222 * isolation. Otherwise, restore the page to right list unless 1223 * we want to retry. 1224 */ 1225 if (rc == MIGRATEPAGE_SUCCESS) { 1226 if (reason != MR_MEMORY_FAILURE) 1227 /* 1228 * We release the page in page_handle_poison. 1229 */ 1230 put_page(page); 1231 } else { 1232 if (rc != -EAGAIN) { 1233 if (likely(!__PageMovable(page))) { 1234 putback_lru_page(page); 1235 goto put_new; 1236 } 1237 1238 lock_page(page); 1239 if (PageMovable(page)) 1240 putback_movable_page(page); 1241 else 1242 __ClearPageIsolated(page); 1243 unlock_page(page); 1244 put_page(page); 1245 } 1246 put_new: 1247 if (put_new_page) 1248 put_new_page(newpage, private); 1249 else 1250 put_page(newpage); 1251 } 1252 1253 return rc; 1254 } 1255 1256 /* 1257 * Counterpart of unmap_and_move_page() for hugepage migration. 1258 * 1259 * This function doesn't wait the completion of hugepage I/O 1260 * because there is no race between I/O and migration for hugepage. 1261 * Note that currently hugepage I/O occurs only in direct I/O 1262 * where no lock is held and PG_writeback is irrelevant, 1263 * and writeback status of all subpages are counted in the reference 1264 * count of the head page (i.e. if all subpages of a 2MB hugepage are 1265 * under direct I/O, the reference of the head page is 512 and a bit more.) 1266 * This means that when we try to migrate hugepage whose subpages are 1267 * doing direct I/O, some references remain after try_to_unmap() and 1268 * hugepage migration fails without data corruption. 1269 * 1270 * There is also no race when direct I/O is issued on the page under migration, 1271 * because then pte is replaced with migration swap entry and direct I/O code 1272 * will wait in the page fault for migration to complete. 1273 */ 1274 static int unmap_and_move_huge_page(new_page_t get_new_page, 1275 free_page_t put_new_page, unsigned long private, 1276 struct page *hpage, int force, 1277 enum migrate_mode mode, int reason) 1278 { 1279 int rc = -EAGAIN; 1280 int page_was_mapped = 0; 1281 struct page *new_hpage; 1282 struct anon_vma *anon_vma = NULL; 1283 struct address_space *mapping = NULL; 1284 1285 /* 1286 * Migratability of hugepages depends on architectures and their size. 1287 * This check is necessary because some callers of hugepage migration 1288 * like soft offline and memory hotremove don't walk through page 1289 * tables or check whether the hugepage is pmd-based or not before 1290 * kicking migration. 1291 */ 1292 if (!hugepage_migration_supported(page_hstate(hpage))) { 1293 putback_active_hugepage(hpage); 1294 return -ENOSYS; 1295 } 1296 1297 new_hpage = get_new_page(hpage, private); 1298 if (!new_hpage) 1299 return -ENOMEM; 1300 1301 if (!trylock_page(hpage)) { 1302 if (!force) 1303 goto out; 1304 switch (mode) { 1305 case MIGRATE_SYNC: 1306 case MIGRATE_SYNC_NO_COPY: 1307 break; 1308 default: 1309 goto out; 1310 } 1311 lock_page(hpage); 1312 } 1313 1314 /* 1315 * Check for pages which are in the process of being freed. Without 1316 * page_mapping() set, hugetlbfs specific move page routine will not 1317 * be called and we could leak usage counts for subpools. 1318 */ 1319 if (page_private(hpage) && !page_mapping(hpage)) { 1320 rc = -EBUSY; 1321 goto out_unlock; 1322 } 1323 1324 if (PageAnon(hpage)) 1325 anon_vma = page_get_anon_vma(hpage); 1326 1327 if (unlikely(!trylock_page(new_hpage))) 1328 goto put_anon; 1329 1330 if (page_mapped(hpage)) { 1331 bool mapping_locked = false; 1332 enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK; 1333 1334 if (!PageAnon(hpage)) { 1335 /* 1336 * In shared mappings, try_to_unmap could potentially 1337 * call huge_pmd_unshare. Because of this, take 1338 * semaphore in write mode here and set TTU_RMAP_LOCKED 1339 * to let lower levels know we have taken the lock. 1340 */ 1341 mapping = hugetlb_page_mapping_lock_write(hpage); 1342 if (unlikely(!mapping)) 1343 goto unlock_put_anon; 1344 1345 mapping_locked = true; 1346 ttu |= TTU_RMAP_LOCKED; 1347 } 1348 1349 try_to_unmap(hpage, ttu); 1350 page_was_mapped = 1; 1351 1352 if (mapping_locked) 1353 i_mmap_unlock_write(mapping); 1354 } 1355 1356 if (!page_mapped(hpage)) 1357 rc = move_to_new_page(new_hpage, hpage, mode); 1358 1359 if (page_was_mapped) 1360 remove_migration_ptes(hpage, 1361 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false); 1362 1363 unlock_put_anon: 1364 unlock_page(new_hpage); 1365 1366 put_anon: 1367 if (anon_vma) 1368 put_anon_vma(anon_vma); 1369 1370 if (rc == MIGRATEPAGE_SUCCESS) { 1371 move_hugetlb_state(hpage, new_hpage, reason); 1372 put_new_page = NULL; 1373 } 1374 1375 out_unlock: 1376 unlock_page(hpage); 1377 out: 1378 if (rc != -EAGAIN) 1379 putback_active_hugepage(hpage); 1380 1381 /* 1382 * If migration was not successful and there's a freeing callback, use 1383 * it. Otherwise, put_page() will drop the reference grabbed during 1384 * isolation. 1385 */ 1386 if (put_new_page) 1387 put_new_page(new_hpage, private); 1388 else 1389 putback_active_hugepage(new_hpage); 1390 1391 return rc; 1392 } 1393 1394 /* 1395 * migrate_pages - migrate the pages specified in a list, to the free pages 1396 * supplied as the target for the page migration 1397 * 1398 * @from: The list of pages to be migrated. 1399 * @get_new_page: The function used to allocate free pages to be used 1400 * as the target of the page migration. 1401 * @put_new_page: The function used to free target pages if migration 1402 * fails, or NULL if no special handling is necessary. 1403 * @private: Private data to be passed on to get_new_page() 1404 * @mode: The migration mode that specifies the constraints for 1405 * page migration, if any. 1406 * @reason: The reason for page migration. 1407 * 1408 * The function returns after 10 attempts or if no pages are movable any more 1409 * because the list has become empty or no retryable pages exist any more. 1410 * The caller should call putback_movable_pages() to return pages to the LRU 1411 * or free list only if ret != 0. 1412 * 1413 * Returns the number of pages that were not migrated, or an error code. 1414 */ 1415 int migrate_pages(struct list_head *from, new_page_t get_new_page, 1416 free_page_t put_new_page, unsigned long private, 1417 enum migrate_mode mode, int reason) 1418 { 1419 int retry = 1; 1420 int thp_retry = 1; 1421 int nr_failed = 0; 1422 int nr_succeeded = 0; 1423 int nr_thp_succeeded = 0; 1424 int nr_thp_failed = 0; 1425 int nr_thp_split = 0; 1426 int pass = 0; 1427 bool is_thp = false; 1428 struct page *page; 1429 struct page *page2; 1430 int swapwrite = current->flags & PF_SWAPWRITE; 1431 int rc, nr_subpages; 1432 1433 if (!swapwrite) 1434 current->flags |= PF_SWAPWRITE; 1435 1436 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) { 1437 retry = 0; 1438 thp_retry = 0; 1439 1440 list_for_each_entry_safe(page, page2, from, lru) { 1441 retry: 1442 /* 1443 * THP statistics is based on the source huge page. 1444 * Capture required information that might get lost 1445 * during migration. 1446 */ 1447 is_thp = PageTransHuge(page) && !PageHuge(page); 1448 nr_subpages = thp_nr_pages(page); 1449 cond_resched(); 1450 1451 if (PageHuge(page)) 1452 rc = unmap_and_move_huge_page(get_new_page, 1453 put_new_page, private, page, 1454 pass > 2, mode, reason); 1455 else 1456 rc = unmap_and_move(get_new_page, put_new_page, 1457 private, page, pass > 2, mode, 1458 reason); 1459 1460 switch(rc) { 1461 case -ENOMEM: 1462 /* 1463 * THP migration might be unsupported or the 1464 * allocation could've failed so we should 1465 * retry on the same page with the THP split 1466 * to base pages. 1467 * 1468 * Head page is retried immediately and tail 1469 * pages are added to the tail of the list so 1470 * we encounter them after the rest of the list 1471 * is processed. 1472 */ 1473 if (is_thp) { 1474 lock_page(page); 1475 rc = split_huge_page_to_list(page, from); 1476 unlock_page(page); 1477 if (!rc) { 1478 list_safe_reset_next(page, page2, lru); 1479 nr_thp_split++; 1480 goto retry; 1481 } 1482 1483 nr_thp_failed++; 1484 nr_failed += nr_subpages; 1485 goto out; 1486 } 1487 nr_failed++; 1488 goto out; 1489 case -EAGAIN: 1490 if (is_thp) { 1491 thp_retry++; 1492 break; 1493 } 1494 retry++; 1495 break; 1496 case MIGRATEPAGE_SUCCESS: 1497 if (is_thp) { 1498 nr_thp_succeeded++; 1499 nr_succeeded += nr_subpages; 1500 break; 1501 } 1502 nr_succeeded++; 1503 break; 1504 default: 1505 /* 1506 * Permanent failure (-EBUSY, -ENOSYS, etc.): 1507 * unlike -EAGAIN case, the failed page is 1508 * removed from migration page list and not 1509 * retried in the next outer loop. 1510 */ 1511 if (is_thp) { 1512 nr_thp_failed++; 1513 nr_failed += nr_subpages; 1514 break; 1515 } 1516 nr_failed++; 1517 break; 1518 } 1519 } 1520 } 1521 nr_failed += retry + thp_retry; 1522 nr_thp_failed += thp_retry; 1523 rc = nr_failed; 1524 out: 1525 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); 1526 count_vm_events(PGMIGRATE_FAIL, nr_failed); 1527 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded); 1528 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed); 1529 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split); 1530 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded, 1531 nr_thp_failed, nr_thp_split, mode, reason); 1532 1533 if (!swapwrite) 1534 current->flags &= ~PF_SWAPWRITE; 1535 1536 return rc; 1537 } 1538 1539 struct page *alloc_migration_target(struct page *page, unsigned long private) 1540 { 1541 struct migration_target_control *mtc; 1542 gfp_t gfp_mask; 1543 unsigned int order = 0; 1544 struct page *new_page = NULL; 1545 int nid; 1546 int zidx; 1547 1548 mtc = (struct migration_target_control *)private; 1549 gfp_mask = mtc->gfp_mask; 1550 nid = mtc->nid; 1551 if (nid == NUMA_NO_NODE) 1552 nid = page_to_nid(page); 1553 1554 if (PageHuge(page)) { 1555 struct hstate *h = page_hstate(compound_head(page)); 1556 1557 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask); 1558 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask); 1559 } 1560 1561 if (PageTransHuge(page)) { 1562 /* 1563 * clear __GFP_RECLAIM to make the migration callback 1564 * consistent with regular THP allocations. 1565 */ 1566 gfp_mask &= ~__GFP_RECLAIM; 1567 gfp_mask |= GFP_TRANSHUGE; 1568 order = HPAGE_PMD_ORDER; 1569 } 1570 zidx = zone_idx(page_zone(page)); 1571 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE) 1572 gfp_mask |= __GFP_HIGHMEM; 1573 1574 new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask); 1575 1576 if (new_page && PageTransHuge(new_page)) 1577 prep_transhuge_page(new_page); 1578 1579 return new_page; 1580 } 1581 1582 #ifdef CONFIG_NUMA 1583 1584 static int store_status(int __user *status, int start, int value, int nr) 1585 { 1586 while (nr-- > 0) { 1587 if (put_user(value, status + start)) 1588 return -EFAULT; 1589 start++; 1590 } 1591 1592 return 0; 1593 } 1594 1595 static int do_move_pages_to_node(struct mm_struct *mm, 1596 struct list_head *pagelist, int node) 1597 { 1598 int err; 1599 struct migration_target_control mtc = { 1600 .nid = node, 1601 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 1602 }; 1603 1604 err = migrate_pages(pagelist, alloc_migration_target, NULL, 1605 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL); 1606 if (err) 1607 putback_movable_pages(pagelist); 1608 return err; 1609 } 1610 1611 /* 1612 * Resolves the given address to a struct page, isolates it from the LRU and 1613 * puts it to the given pagelist. 1614 * Returns: 1615 * errno - if the page cannot be found/isolated 1616 * 0 - when it doesn't have to be migrated because it is already on the 1617 * target node 1618 * 1 - when it has been queued 1619 */ 1620 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr, 1621 int node, struct list_head *pagelist, bool migrate_all) 1622 { 1623 struct vm_area_struct *vma; 1624 struct page *page; 1625 unsigned int follflags; 1626 int err; 1627 1628 mmap_read_lock(mm); 1629 err = -EFAULT; 1630 vma = find_vma(mm, addr); 1631 if (!vma || addr < vma->vm_start || !vma_migratable(vma)) 1632 goto out; 1633 1634 /* FOLL_DUMP to ignore special (like zero) pages */ 1635 follflags = FOLL_GET | FOLL_DUMP; 1636 page = follow_page(vma, addr, follflags); 1637 1638 err = PTR_ERR(page); 1639 if (IS_ERR(page)) 1640 goto out; 1641 1642 err = -ENOENT; 1643 if (!page) 1644 goto out; 1645 1646 err = 0; 1647 if (page_to_nid(page) == node) 1648 goto out_putpage; 1649 1650 err = -EACCES; 1651 if (page_mapcount(page) > 1 && !migrate_all) 1652 goto out_putpage; 1653 1654 if (PageHuge(page)) { 1655 if (PageHead(page)) { 1656 isolate_huge_page(page, pagelist); 1657 err = 1; 1658 } 1659 } else { 1660 struct page *head; 1661 1662 head = compound_head(page); 1663 err = isolate_lru_page(head); 1664 if (err) 1665 goto out_putpage; 1666 1667 err = 1; 1668 list_add_tail(&head->lru, pagelist); 1669 mod_node_page_state(page_pgdat(head), 1670 NR_ISOLATED_ANON + page_is_file_lru(head), 1671 thp_nr_pages(head)); 1672 } 1673 out_putpage: 1674 /* 1675 * Either remove the duplicate refcount from 1676 * isolate_lru_page() or drop the page ref if it was 1677 * not isolated. 1678 */ 1679 put_page(page); 1680 out: 1681 mmap_read_unlock(mm); 1682 return err; 1683 } 1684 1685 static int move_pages_and_store_status(struct mm_struct *mm, int node, 1686 struct list_head *pagelist, int __user *status, 1687 int start, int i, unsigned long nr_pages) 1688 { 1689 int err; 1690 1691 if (list_empty(pagelist)) 1692 return 0; 1693 1694 err = do_move_pages_to_node(mm, pagelist, node); 1695 if (err) { 1696 /* 1697 * Positive err means the number of failed 1698 * pages to migrate. Since we are going to 1699 * abort and return the number of non-migrated 1700 * pages, so need to incude the rest of the 1701 * nr_pages that have not been attempted as 1702 * well. 1703 */ 1704 if (err > 0) 1705 err += nr_pages - i - 1; 1706 return err; 1707 } 1708 return store_status(status, start, node, i - start); 1709 } 1710 1711 /* 1712 * Migrate an array of page address onto an array of nodes and fill 1713 * the corresponding array of status. 1714 */ 1715 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, 1716 unsigned long nr_pages, 1717 const void __user * __user *pages, 1718 const int __user *nodes, 1719 int __user *status, int flags) 1720 { 1721 int current_node = NUMA_NO_NODE; 1722 LIST_HEAD(pagelist); 1723 int start, i; 1724 int err = 0, err1; 1725 1726 migrate_prep(); 1727 1728 for (i = start = 0; i < nr_pages; i++) { 1729 const void __user *p; 1730 unsigned long addr; 1731 int node; 1732 1733 err = -EFAULT; 1734 if (get_user(p, pages + i)) 1735 goto out_flush; 1736 if (get_user(node, nodes + i)) 1737 goto out_flush; 1738 addr = (unsigned long)untagged_addr(p); 1739 1740 err = -ENODEV; 1741 if (node < 0 || node >= MAX_NUMNODES) 1742 goto out_flush; 1743 if (!node_state(node, N_MEMORY)) 1744 goto out_flush; 1745 1746 err = -EACCES; 1747 if (!node_isset(node, task_nodes)) 1748 goto out_flush; 1749 1750 if (current_node == NUMA_NO_NODE) { 1751 current_node = node; 1752 start = i; 1753 } else if (node != current_node) { 1754 err = move_pages_and_store_status(mm, current_node, 1755 &pagelist, status, start, i, nr_pages); 1756 if (err) 1757 goto out; 1758 start = i; 1759 current_node = node; 1760 } 1761 1762 /* 1763 * Errors in the page lookup or isolation are not fatal and we simply 1764 * report them via status 1765 */ 1766 err = add_page_for_migration(mm, addr, current_node, 1767 &pagelist, flags & MPOL_MF_MOVE_ALL); 1768 1769 if (err > 0) { 1770 /* The page is successfully queued for migration */ 1771 continue; 1772 } 1773 1774 /* 1775 * If the page is already on the target node (!err), store the 1776 * node, otherwise, store the err. 1777 */ 1778 err = store_status(status, i, err ? : current_node, 1); 1779 if (err) 1780 goto out_flush; 1781 1782 err = move_pages_and_store_status(mm, current_node, &pagelist, 1783 status, start, i, nr_pages); 1784 if (err) 1785 goto out; 1786 current_node = NUMA_NO_NODE; 1787 } 1788 out_flush: 1789 /* Make sure we do not overwrite the existing error */ 1790 err1 = move_pages_and_store_status(mm, current_node, &pagelist, 1791 status, start, i, nr_pages); 1792 if (err >= 0) 1793 err = err1; 1794 out: 1795 return err; 1796 } 1797 1798 /* 1799 * Determine the nodes of an array of pages and store it in an array of status. 1800 */ 1801 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, 1802 const void __user **pages, int *status) 1803 { 1804 unsigned long i; 1805 1806 mmap_read_lock(mm); 1807 1808 for (i = 0; i < nr_pages; i++) { 1809 unsigned long addr = (unsigned long)(*pages); 1810 struct vm_area_struct *vma; 1811 struct page *page; 1812 int err = -EFAULT; 1813 1814 vma = find_vma(mm, addr); 1815 if (!vma || addr < vma->vm_start) 1816 goto set_status; 1817 1818 /* FOLL_DUMP to ignore special (like zero) pages */ 1819 page = follow_page(vma, addr, FOLL_DUMP); 1820 1821 err = PTR_ERR(page); 1822 if (IS_ERR(page)) 1823 goto set_status; 1824 1825 err = page ? page_to_nid(page) : -ENOENT; 1826 set_status: 1827 *status = err; 1828 1829 pages++; 1830 status++; 1831 } 1832 1833 mmap_read_unlock(mm); 1834 } 1835 1836 /* 1837 * Determine the nodes of a user array of pages and store it in 1838 * a user array of status. 1839 */ 1840 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, 1841 const void __user * __user *pages, 1842 int __user *status) 1843 { 1844 #define DO_PAGES_STAT_CHUNK_NR 16 1845 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; 1846 int chunk_status[DO_PAGES_STAT_CHUNK_NR]; 1847 1848 while (nr_pages) { 1849 unsigned long chunk_nr; 1850 1851 chunk_nr = nr_pages; 1852 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) 1853 chunk_nr = DO_PAGES_STAT_CHUNK_NR; 1854 1855 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) 1856 break; 1857 1858 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); 1859 1860 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) 1861 break; 1862 1863 pages += chunk_nr; 1864 status += chunk_nr; 1865 nr_pages -= chunk_nr; 1866 } 1867 return nr_pages ? -EFAULT : 0; 1868 } 1869 1870 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes) 1871 { 1872 struct task_struct *task; 1873 struct mm_struct *mm; 1874 1875 /* 1876 * There is no need to check if current process has the right to modify 1877 * the specified process when they are same. 1878 */ 1879 if (!pid) { 1880 mmget(current->mm); 1881 *mem_nodes = cpuset_mems_allowed(current); 1882 return current->mm; 1883 } 1884 1885 /* Find the mm_struct */ 1886 rcu_read_lock(); 1887 task = find_task_by_vpid(pid); 1888 if (!task) { 1889 rcu_read_unlock(); 1890 return ERR_PTR(-ESRCH); 1891 } 1892 get_task_struct(task); 1893 1894 /* 1895 * Check if this process has the right to modify the specified 1896 * process. Use the regular "ptrace_may_access()" checks. 1897 */ 1898 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 1899 rcu_read_unlock(); 1900 mm = ERR_PTR(-EPERM); 1901 goto out; 1902 } 1903 rcu_read_unlock(); 1904 1905 mm = ERR_PTR(security_task_movememory(task)); 1906 if (IS_ERR(mm)) 1907 goto out; 1908 *mem_nodes = cpuset_mems_allowed(task); 1909 mm = get_task_mm(task); 1910 out: 1911 put_task_struct(task); 1912 if (!mm) 1913 mm = ERR_PTR(-EINVAL); 1914 return mm; 1915 } 1916 1917 /* 1918 * Move a list of pages in the address space of the currently executing 1919 * process. 1920 */ 1921 static int kernel_move_pages(pid_t pid, unsigned long nr_pages, 1922 const void __user * __user *pages, 1923 const int __user *nodes, 1924 int __user *status, int flags) 1925 { 1926 struct mm_struct *mm; 1927 int err; 1928 nodemask_t task_nodes; 1929 1930 /* Check flags */ 1931 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) 1932 return -EINVAL; 1933 1934 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1935 return -EPERM; 1936 1937 mm = find_mm_struct(pid, &task_nodes); 1938 if (IS_ERR(mm)) 1939 return PTR_ERR(mm); 1940 1941 if (nodes) 1942 err = do_pages_move(mm, task_nodes, nr_pages, pages, 1943 nodes, status, flags); 1944 else 1945 err = do_pages_stat(mm, nr_pages, pages, status); 1946 1947 mmput(mm); 1948 return err; 1949 } 1950 1951 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, 1952 const void __user * __user *, pages, 1953 const int __user *, nodes, 1954 int __user *, status, int, flags) 1955 { 1956 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 1957 } 1958 1959 #ifdef CONFIG_COMPAT 1960 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages, 1961 compat_uptr_t __user *, pages32, 1962 const int __user *, nodes, 1963 int __user *, status, 1964 int, flags) 1965 { 1966 const void __user * __user *pages; 1967 int i; 1968 1969 pages = compat_alloc_user_space(nr_pages * sizeof(void *)); 1970 for (i = 0; i < nr_pages; i++) { 1971 compat_uptr_t p; 1972 1973 if (get_user(p, pages32 + i) || 1974 put_user(compat_ptr(p), pages + i)) 1975 return -EFAULT; 1976 } 1977 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags); 1978 } 1979 #endif /* CONFIG_COMPAT */ 1980 1981 #ifdef CONFIG_NUMA_BALANCING 1982 /* 1983 * Returns true if this is a safe migration target node for misplaced NUMA 1984 * pages. Currently it only checks the watermarks which crude 1985 */ 1986 static bool migrate_balanced_pgdat(struct pglist_data *pgdat, 1987 unsigned long nr_migrate_pages) 1988 { 1989 int z; 1990 1991 for (z = pgdat->nr_zones - 1; z >= 0; z--) { 1992 struct zone *zone = pgdat->node_zones + z; 1993 1994 if (!populated_zone(zone)) 1995 continue; 1996 1997 /* Avoid waking kswapd by allocating pages_to_migrate pages. */ 1998 if (!zone_watermark_ok(zone, 0, 1999 high_wmark_pages(zone) + 2000 nr_migrate_pages, 2001 ZONE_MOVABLE, 0)) 2002 continue; 2003 return true; 2004 } 2005 return false; 2006 } 2007 2008 static struct page *alloc_misplaced_dst_page(struct page *page, 2009 unsigned long data) 2010 { 2011 int nid = (int) data; 2012 struct page *newpage; 2013 2014 newpage = __alloc_pages_node(nid, 2015 (GFP_HIGHUSER_MOVABLE | 2016 __GFP_THISNODE | __GFP_NOMEMALLOC | 2017 __GFP_NORETRY | __GFP_NOWARN) & 2018 ~__GFP_RECLAIM, 0); 2019 2020 return newpage; 2021 } 2022 2023 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) 2024 { 2025 int page_lru; 2026 2027 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page); 2028 2029 /* Avoid migrating to a node that is nearly full */ 2030 if (!migrate_balanced_pgdat(pgdat, compound_nr(page))) 2031 return 0; 2032 2033 if (isolate_lru_page(page)) 2034 return 0; 2035 2036 /* 2037 * migrate_misplaced_transhuge_page() skips page migration's usual 2038 * check on page_count(), so we must do it here, now that the page 2039 * has been isolated: a GUP pin, or any other pin, prevents migration. 2040 * The expected page count is 3: 1 for page's mapcount and 1 for the 2041 * caller's pin and 1 for the reference taken by isolate_lru_page(). 2042 */ 2043 if (PageTransHuge(page) && page_count(page) != 3) { 2044 putback_lru_page(page); 2045 return 0; 2046 } 2047 2048 page_lru = page_is_file_lru(page); 2049 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru, 2050 thp_nr_pages(page)); 2051 2052 /* 2053 * Isolating the page has taken another reference, so the 2054 * caller's reference can be safely dropped without the page 2055 * disappearing underneath us during migration. 2056 */ 2057 put_page(page); 2058 return 1; 2059 } 2060 2061 bool pmd_trans_migrating(pmd_t pmd) 2062 { 2063 struct page *page = pmd_page(pmd); 2064 return PageLocked(page); 2065 } 2066 2067 /* 2068 * Attempt to migrate a misplaced page to the specified destination 2069 * node. Caller is expected to have an elevated reference count on 2070 * the page that will be dropped by this function before returning. 2071 */ 2072 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, 2073 int node) 2074 { 2075 pg_data_t *pgdat = NODE_DATA(node); 2076 int isolated; 2077 int nr_remaining; 2078 LIST_HEAD(migratepages); 2079 2080 /* 2081 * Don't migrate file pages that are mapped in multiple processes 2082 * with execute permissions as they are probably shared libraries. 2083 */ 2084 if (page_mapcount(page) != 1 && page_is_file_lru(page) && 2085 (vma->vm_flags & VM_EXEC)) 2086 goto out; 2087 2088 /* 2089 * Also do not migrate dirty pages as not all filesystems can move 2090 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles. 2091 */ 2092 if (page_is_file_lru(page) && PageDirty(page)) 2093 goto out; 2094 2095 isolated = numamigrate_isolate_page(pgdat, page); 2096 if (!isolated) 2097 goto out; 2098 2099 list_add(&page->lru, &migratepages); 2100 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, 2101 NULL, node, MIGRATE_ASYNC, 2102 MR_NUMA_MISPLACED); 2103 if (nr_remaining) { 2104 if (!list_empty(&migratepages)) { 2105 list_del(&page->lru); 2106 dec_node_page_state(page, NR_ISOLATED_ANON + 2107 page_is_file_lru(page)); 2108 putback_lru_page(page); 2109 } 2110 isolated = 0; 2111 } else 2112 count_vm_numa_event(NUMA_PAGE_MIGRATE); 2113 BUG_ON(!list_empty(&migratepages)); 2114 return isolated; 2115 2116 out: 2117 put_page(page); 2118 return 0; 2119 } 2120 #endif /* CONFIG_NUMA_BALANCING */ 2121 2122 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) 2123 /* 2124 * Migrates a THP to a given target node. page must be locked and is unlocked 2125 * before returning. 2126 */ 2127 int migrate_misplaced_transhuge_page(struct mm_struct *mm, 2128 struct vm_area_struct *vma, 2129 pmd_t *pmd, pmd_t entry, 2130 unsigned long address, 2131 struct page *page, int node) 2132 { 2133 spinlock_t *ptl; 2134 pg_data_t *pgdat = NODE_DATA(node); 2135 int isolated = 0; 2136 struct page *new_page = NULL; 2137 int page_lru = page_is_file_lru(page); 2138 unsigned long start = address & HPAGE_PMD_MASK; 2139 2140 new_page = alloc_pages_node(node, 2141 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE), 2142 HPAGE_PMD_ORDER); 2143 if (!new_page) 2144 goto out_fail; 2145 prep_transhuge_page(new_page); 2146 2147 isolated = numamigrate_isolate_page(pgdat, page); 2148 if (!isolated) { 2149 put_page(new_page); 2150 goto out_fail; 2151 } 2152 2153 /* Prepare a page as a migration target */ 2154 __SetPageLocked(new_page); 2155 if (PageSwapBacked(page)) 2156 __SetPageSwapBacked(new_page); 2157 2158 /* anon mapping, we can simply copy page->mapping to the new page: */ 2159 new_page->mapping = page->mapping; 2160 new_page->index = page->index; 2161 /* flush the cache before copying using the kernel virtual address */ 2162 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE); 2163 migrate_page_copy(new_page, page); 2164 WARN_ON(PageLRU(new_page)); 2165 2166 /* Recheck the target PMD */ 2167 ptl = pmd_lock(mm, pmd); 2168 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) { 2169 spin_unlock(ptl); 2170 2171 /* Reverse changes made by migrate_page_copy() */ 2172 if (TestClearPageActive(new_page)) 2173 SetPageActive(page); 2174 if (TestClearPageUnevictable(new_page)) 2175 SetPageUnevictable(page); 2176 2177 unlock_page(new_page); 2178 put_page(new_page); /* Free it */ 2179 2180 /* Retake the callers reference and putback on LRU */ 2181 get_page(page); 2182 putback_lru_page(page); 2183 mod_node_page_state(page_pgdat(page), 2184 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); 2185 2186 goto out_unlock; 2187 } 2188 2189 entry = mk_huge_pmd(new_page, vma->vm_page_prot); 2190 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); 2191 2192 /* 2193 * Overwrite the old entry under pagetable lock and establish 2194 * the new PTE. Any parallel GUP will either observe the old 2195 * page blocking on the page lock, block on the page table 2196 * lock or observe the new page. The SetPageUptodate on the 2197 * new page and page_add_new_anon_rmap guarantee the copy is 2198 * visible before the pagetable update. 2199 */ 2200 page_add_anon_rmap(new_page, vma, start, true); 2201 /* 2202 * At this point the pmd is numa/protnone (i.e. non present) and the TLB 2203 * has already been flushed globally. So no TLB can be currently 2204 * caching this non present pmd mapping. There's no need to clear the 2205 * pmd before doing set_pmd_at(), nor to flush the TLB after 2206 * set_pmd_at(). Clearing the pmd here would introduce a race 2207 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the 2208 * mmap_lock for reading. If the pmd is set to NULL at any given time, 2209 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this 2210 * pmd. 2211 */ 2212 set_pmd_at(mm, start, pmd, entry); 2213 update_mmu_cache_pmd(vma, address, &entry); 2214 2215 page_ref_unfreeze(page, 2); 2216 mlock_migrate_page(new_page, page); 2217 page_remove_rmap(page, true); 2218 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED); 2219 2220 spin_unlock(ptl); 2221 2222 /* Take an "isolate" reference and put new page on the LRU. */ 2223 get_page(new_page); 2224 putback_lru_page(new_page); 2225 2226 unlock_page(new_page); 2227 unlock_page(page); 2228 put_page(page); /* Drop the rmap reference */ 2229 put_page(page); /* Drop the LRU isolation reference */ 2230 2231 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); 2232 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); 2233 2234 mod_node_page_state(page_pgdat(page), 2235 NR_ISOLATED_ANON + page_lru, 2236 -HPAGE_PMD_NR); 2237 return isolated; 2238 2239 out_fail: 2240 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); 2241 ptl = pmd_lock(mm, pmd); 2242 if (pmd_same(*pmd, entry)) { 2243 entry = pmd_modify(entry, vma->vm_page_prot); 2244 set_pmd_at(mm, start, pmd, entry); 2245 update_mmu_cache_pmd(vma, address, &entry); 2246 } 2247 spin_unlock(ptl); 2248 2249 out_unlock: 2250 unlock_page(page); 2251 put_page(page); 2252 return 0; 2253 } 2254 #endif /* CONFIG_NUMA_BALANCING */ 2255 2256 #endif /* CONFIG_NUMA */ 2257 2258 #ifdef CONFIG_DEVICE_PRIVATE 2259 static int migrate_vma_collect_hole(unsigned long start, 2260 unsigned long end, 2261 __always_unused int depth, 2262 struct mm_walk *walk) 2263 { 2264 struct migrate_vma *migrate = walk->private; 2265 unsigned long addr; 2266 2267 /* Only allow populating anonymous memory. */ 2268 if (!vma_is_anonymous(walk->vma)) { 2269 for (addr = start; addr < end; addr += PAGE_SIZE) { 2270 migrate->src[migrate->npages] = 0; 2271 migrate->dst[migrate->npages] = 0; 2272 migrate->npages++; 2273 } 2274 return 0; 2275 } 2276 2277 for (addr = start; addr < end; addr += PAGE_SIZE) { 2278 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE; 2279 migrate->dst[migrate->npages] = 0; 2280 migrate->npages++; 2281 migrate->cpages++; 2282 } 2283 2284 return 0; 2285 } 2286 2287 static int migrate_vma_collect_skip(unsigned long start, 2288 unsigned long end, 2289 struct mm_walk *walk) 2290 { 2291 struct migrate_vma *migrate = walk->private; 2292 unsigned long addr; 2293 2294 for (addr = start; addr < end; addr += PAGE_SIZE) { 2295 migrate->dst[migrate->npages] = 0; 2296 migrate->src[migrate->npages++] = 0; 2297 } 2298 2299 return 0; 2300 } 2301 2302 static int migrate_vma_collect_pmd(pmd_t *pmdp, 2303 unsigned long start, 2304 unsigned long end, 2305 struct mm_walk *walk) 2306 { 2307 struct migrate_vma *migrate = walk->private; 2308 struct vm_area_struct *vma = walk->vma; 2309 struct mm_struct *mm = vma->vm_mm; 2310 unsigned long addr = start, unmapped = 0; 2311 spinlock_t *ptl; 2312 pte_t *ptep; 2313 2314 again: 2315 if (pmd_none(*pmdp)) 2316 return migrate_vma_collect_hole(start, end, -1, walk); 2317 2318 if (pmd_trans_huge(*pmdp)) { 2319 struct page *page; 2320 2321 ptl = pmd_lock(mm, pmdp); 2322 if (unlikely(!pmd_trans_huge(*pmdp))) { 2323 spin_unlock(ptl); 2324 goto again; 2325 } 2326 2327 page = pmd_page(*pmdp); 2328 if (is_huge_zero_page(page)) { 2329 spin_unlock(ptl); 2330 split_huge_pmd(vma, pmdp, addr); 2331 if (pmd_trans_unstable(pmdp)) 2332 return migrate_vma_collect_skip(start, end, 2333 walk); 2334 } else { 2335 int ret; 2336 2337 get_page(page); 2338 spin_unlock(ptl); 2339 if (unlikely(!trylock_page(page))) 2340 return migrate_vma_collect_skip(start, end, 2341 walk); 2342 ret = split_huge_page(page); 2343 unlock_page(page); 2344 put_page(page); 2345 if (ret) 2346 return migrate_vma_collect_skip(start, end, 2347 walk); 2348 if (pmd_none(*pmdp)) 2349 return migrate_vma_collect_hole(start, end, -1, 2350 walk); 2351 } 2352 } 2353 2354 if (unlikely(pmd_bad(*pmdp))) 2355 return migrate_vma_collect_skip(start, end, walk); 2356 2357 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); 2358 arch_enter_lazy_mmu_mode(); 2359 2360 for (; addr < end; addr += PAGE_SIZE, ptep++) { 2361 unsigned long mpfn = 0, pfn; 2362 struct page *page; 2363 swp_entry_t entry; 2364 pte_t pte; 2365 2366 pte = *ptep; 2367 2368 if (pte_none(pte)) { 2369 if (vma_is_anonymous(vma)) { 2370 mpfn = MIGRATE_PFN_MIGRATE; 2371 migrate->cpages++; 2372 } 2373 goto next; 2374 } 2375 2376 if (!pte_present(pte)) { 2377 /* 2378 * Only care about unaddressable device page special 2379 * page table entry. Other special swap entries are not 2380 * migratable, and we ignore regular swapped page. 2381 */ 2382 entry = pte_to_swp_entry(pte); 2383 if (!is_device_private_entry(entry)) 2384 goto next; 2385 2386 page = device_private_entry_to_page(entry); 2387 if (!(migrate->flags & 2388 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) || 2389 page->pgmap->owner != migrate->pgmap_owner) 2390 goto next; 2391 2392 mpfn = migrate_pfn(page_to_pfn(page)) | 2393 MIGRATE_PFN_MIGRATE; 2394 if (is_write_device_private_entry(entry)) 2395 mpfn |= MIGRATE_PFN_WRITE; 2396 } else { 2397 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) 2398 goto next; 2399 pfn = pte_pfn(pte); 2400 if (is_zero_pfn(pfn)) { 2401 mpfn = MIGRATE_PFN_MIGRATE; 2402 migrate->cpages++; 2403 goto next; 2404 } 2405 page = vm_normal_page(migrate->vma, addr, pte); 2406 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE; 2407 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0; 2408 } 2409 2410 /* FIXME support THP */ 2411 if (!page || !page->mapping || PageTransCompound(page)) { 2412 mpfn = 0; 2413 goto next; 2414 } 2415 2416 /* 2417 * By getting a reference on the page we pin it and that blocks 2418 * any kind of migration. Side effect is that it "freezes" the 2419 * pte. 2420 * 2421 * We drop this reference after isolating the page from the lru 2422 * for non device page (device page are not on the lru and thus 2423 * can't be dropped from it). 2424 */ 2425 get_page(page); 2426 migrate->cpages++; 2427 2428 /* 2429 * Optimize for the common case where page is only mapped once 2430 * in one process. If we can lock the page, then we can safely 2431 * set up a special migration page table entry now. 2432 */ 2433 if (trylock_page(page)) { 2434 pte_t swp_pte; 2435 2436 mpfn |= MIGRATE_PFN_LOCKED; 2437 ptep_get_and_clear(mm, addr, ptep); 2438 2439 /* Setup special migration page table entry */ 2440 entry = make_migration_entry(page, mpfn & 2441 MIGRATE_PFN_WRITE); 2442 swp_pte = swp_entry_to_pte(entry); 2443 if (pte_present(pte)) { 2444 if (pte_soft_dirty(pte)) 2445 swp_pte = pte_swp_mksoft_dirty(swp_pte); 2446 if (pte_uffd_wp(pte)) 2447 swp_pte = pte_swp_mkuffd_wp(swp_pte); 2448 } else { 2449 if (pte_swp_soft_dirty(pte)) 2450 swp_pte = pte_swp_mksoft_dirty(swp_pte); 2451 if (pte_swp_uffd_wp(pte)) 2452 swp_pte = pte_swp_mkuffd_wp(swp_pte); 2453 } 2454 set_pte_at(mm, addr, ptep, swp_pte); 2455 2456 /* 2457 * This is like regular unmap: we remove the rmap and 2458 * drop page refcount. Page won't be freed, as we took 2459 * a reference just above. 2460 */ 2461 page_remove_rmap(page, false); 2462 put_page(page); 2463 2464 if (pte_present(pte)) 2465 unmapped++; 2466 } 2467 2468 next: 2469 migrate->dst[migrate->npages] = 0; 2470 migrate->src[migrate->npages++] = mpfn; 2471 } 2472 arch_leave_lazy_mmu_mode(); 2473 pte_unmap_unlock(ptep - 1, ptl); 2474 2475 /* Only flush the TLB if we actually modified any entries */ 2476 if (unmapped) 2477 flush_tlb_range(walk->vma, start, end); 2478 2479 return 0; 2480 } 2481 2482 static const struct mm_walk_ops migrate_vma_walk_ops = { 2483 .pmd_entry = migrate_vma_collect_pmd, 2484 .pte_hole = migrate_vma_collect_hole, 2485 }; 2486 2487 /* 2488 * migrate_vma_collect() - collect pages over a range of virtual addresses 2489 * @migrate: migrate struct containing all migration information 2490 * 2491 * This will walk the CPU page table. For each virtual address backed by a 2492 * valid page, it updates the src array and takes a reference on the page, in 2493 * order to pin the page until we lock it and unmap it. 2494 */ 2495 static void migrate_vma_collect(struct migrate_vma *migrate) 2496 { 2497 struct mmu_notifier_range range; 2498 2499 /* 2500 * Note that the pgmap_owner is passed to the mmu notifier callback so 2501 * that the registered device driver can skip invalidating device 2502 * private page mappings that won't be migrated. 2503 */ 2504 mmu_notifier_range_init_migrate(&range, 0, migrate->vma, 2505 migrate->vma->vm_mm, migrate->start, migrate->end, 2506 migrate->pgmap_owner); 2507 mmu_notifier_invalidate_range_start(&range); 2508 2509 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end, 2510 &migrate_vma_walk_ops, migrate); 2511 2512 mmu_notifier_invalidate_range_end(&range); 2513 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT); 2514 } 2515 2516 /* 2517 * migrate_vma_check_page() - check if page is pinned or not 2518 * @page: struct page to check 2519 * 2520 * Pinned pages cannot be migrated. This is the same test as in 2521 * migrate_page_move_mapping(), except that here we allow migration of a 2522 * ZONE_DEVICE page. 2523 */ 2524 static bool migrate_vma_check_page(struct page *page) 2525 { 2526 /* 2527 * One extra ref because caller holds an extra reference, either from 2528 * isolate_lru_page() for a regular page, or migrate_vma_collect() for 2529 * a device page. 2530 */ 2531 int extra = 1; 2532 2533 /* 2534 * FIXME support THP (transparent huge page), it is bit more complex to 2535 * check them than regular pages, because they can be mapped with a pmd 2536 * or with a pte (split pte mapping). 2537 */ 2538 if (PageCompound(page)) 2539 return false; 2540 2541 /* Page from ZONE_DEVICE have one extra reference */ 2542 if (is_zone_device_page(page)) { 2543 /* 2544 * Private page can never be pin as they have no valid pte and 2545 * GUP will fail for those. Yet if there is a pending migration 2546 * a thread might try to wait on the pte migration entry and 2547 * will bump the page reference count. Sadly there is no way to 2548 * differentiate a regular pin from migration wait. Hence to 2549 * avoid 2 racing thread trying to migrate back to CPU to enter 2550 * infinite loop (one stoping migration because the other is 2551 * waiting on pte migration entry). We always return true here. 2552 * 2553 * FIXME proper solution is to rework migration_entry_wait() so 2554 * it does not need to take a reference on page. 2555 */ 2556 return is_device_private_page(page); 2557 } 2558 2559 /* For file back page */ 2560 if (page_mapping(page)) 2561 extra += 1 + page_has_private(page); 2562 2563 if ((page_count(page) - extra) > page_mapcount(page)) 2564 return false; 2565 2566 return true; 2567 } 2568 2569 /* 2570 * migrate_vma_prepare() - lock pages and isolate them from the lru 2571 * @migrate: migrate struct containing all migration information 2572 * 2573 * This locks pages that have been collected by migrate_vma_collect(). Once each 2574 * page is locked it is isolated from the lru (for non-device pages). Finally, 2575 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be 2576 * migrated by concurrent kernel threads. 2577 */ 2578 static void migrate_vma_prepare(struct migrate_vma *migrate) 2579 { 2580 const unsigned long npages = migrate->npages; 2581 const unsigned long start = migrate->start; 2582 unsigned long addr, i, restore = 0; 2583 bool allow_drain = true; 2584 2585 lru_add_drain(); 2586 2587 for (i = 0; (i < npages) && migrate->cpages; i++) { 2588 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2589 bool remap = true; 2590 2591 if (!page) 2592 continue; 2593 2594 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) { 2595 /* 2596 * Because we are migrating several pages there can be 2597 * a deadlock between 2 concurrent migration where each 2598 * are waiting on each other page lock. 2599 * 2600 * Make migrate_vma() a best effort thing and backoff 2601 * for any page we can not lock right away. 2602 */ 2603 if (!trylock_page(page)) { 2604 migrate->src[i] = 0; 2605 migrate->cpages--; 2606 put_page(page); 2607 continue; 2608 } 2609 remap = false; 2610 migrate->src[i] |= MIGRATE_PFN_LOCKED; 2611 } 2612 2613 /* ZONE_DEVICE pages are not on LRU */ 2614 if (!is_zone_device_page(page)) { 2615 if (!PageLRU(page) && allow_drain) { 2616 /* Drain CPU's pagevec */ 2617 lru_add_drain_all(); 2618 allow_drain = false; 2619 } 2620 2621 if (isolate_lru_page(page)) { 2622 if (remap) { 2623 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2624 migrate->cpages--; 2625 restore++; 2626 } else { 2627 migrate->src[i] = 0; 2628 unlock_page(page); 2629 migrate->cpages--; 2630 put_page(page); 2631 } 2632 continue; 2633 } 2634 2635 /* Drop the reference we took in collect */ 2636 put_page(page); 2637 } 2638 2639 if (!migrate_vma_check_page(page)) { 2640 if (remap) { 2641 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2642 migrate->cpages--; 2643 restore++; 2644 2645 if (!is_zone_device_page(page)) { 2646 get_page(page); 2647 putback_lru_page(page); 2648 } 2649 } else { 2650 migrate->src[i] = 0; 2651 unlock_page(page); 2652 migrate->cpages--; 2653 2654 if (!is_zone_device_page(page)) 2655 putback_lru_page(page); 2656 else 2657 put_page(page); 2658 } 2659 } 2660 } 2661 2662 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) { 2663 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2664 2665 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2666 continue; 2667 2668 remove_migration_pte(page, migrate->vma, addr, page); 2669 2670 migrate->src[i] = 0; 2671 unlock_page(page); 2672 put_page(page); 2673 restore--; 2674 } 2675 } 2676 2677 /* 2678 * migrate_vma_unmap() - replace page mapping with special migration pte entry 2679 * @migrate: migrate struct containing all migration information 2680 * 2681 * Replace page mapping (CPU page table pte) with a special migration pte entry 2682 * and check again if it has been pinned. Pinned pages are restored because we 2683 * cannot migrate them. 2684 * 2685 * This is the last step before we call the device driver callback to allocate 2686 * destination memory and copy contents of original page over to new page. 2687 */ 2688 static void migrate_vma_unmap(struct migrate_vma *migrate) 2689 { 2690 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK; 2691 const unsigned long npages = migrate->npages; 2692 const unsigned long start = migrate->start; 2693 unsigned long addr, i, restore = 0; 2694 2695 for (i = 0; i < npages; i++) { 2696 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2697 2698 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2699 continue; 2700 2701 if (page_mapped(page)) { 2702 try_to_unmap(page, flags); 2703 if (page_mapped(page)) 2704 goto restore; 2705 } 2706 2707 if (migrate_vma_check_page(page)) 2708 continue; 2709 2710 restore: 2711 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2712 migrate->cpages--; 2713 restore++; 2714 } 2715 2716 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) { 2717 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2718 2719 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE)) 2720 continue; 2721 2722 remove_migration_ptes(page, page, false); 2723 2724 migrate->src[i] = 0; 2725 unlock_page(page); 2726 restore--; 2727 2728 if (is_zone_device_page(page)) 2729 put_page(page); 2730 else 2731 putback_lru_page(page); 2732 } 2733 } 2734 2735 /** 2736 * migrate_vma_setup() - prepare to migrate a range of memory 2737 * @args: contains the vma, start, and pfns arrays for the migration 2738 * 2739 * Returns: negative errno on failures, 0 when 0 or more pages were migrated 2740 * without an error. 2741 * 2742 * Prepare to migrate a range of memory virtual address range by collecting all 2743 * the pages backing each virtual address in the range, saving them inside the 2744 * src array. Then lock those pages and unmap them. Once the pages are locked 2745 * and unmapped, check whether each page is pinned or not. Pages that aren't 2746 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the 2747 * corresponding src array entry. Then restores any pages that are pinned, by 2748 * remapping and unlocking those pages. 2749 * 2750 * The caller should then allocate destination memory and copy source memory to 2751 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE 2752 * flag set). Once these are allocated and copied, the caller must update each 2753 * corresponding entry in the dst array with the pfn value of the destination 2754 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set 2755 * (destination pages must have their struct pages locked, via lock_page()). 2756 * 2757 * Note that the caller does not have to migrate all the pages that are marked 2758 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from 2759 * device memory to system memory. If the caller cannot migrate a device page 2760 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe 2761 * consequences for the userspace process, so it must be avoided if at all 2762 * possible. 2763 * 2764 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we 2765 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus 2766 * allowing the caller to allocate device memory for those unback virtual 2767 * address. For this the caller simply has to allocate device memory and 2768 * properly set the destination entry like for regular migration. Note that 2769 * this can still fails and thus inside the device driver must check if the 2770 * migration was successful for those entries after calling migrate_vma_pages() 2771 * just like for regular migration. 2772 * 2773 * After that, the callers must call migrate_vma_pages() to go over each entry 2774 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag 2775 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set, 2776 * then migrate_vma_pages() to migrate struct page information from the source 2777 * struct page to the destination struct page. If it fails to migrate the 2778 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the 2779 * src array. 2780 * 2781 * At this point all successfully migrated pages have an entry in the src 2782 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst 2783 * array entry with MIGRATE_PFN_VALID flag set. 2784 * 2785 * Once migrate_vma_pages() returns the caller may inspect which pages were 2786 * successfully migrated, and which were not. Successfully migrated pages will 2787 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry. 2788 * 2789 * It is safe to update device page table after migrate_vma_pages() because 2790 * both destination and source page are still locked, and the mmap_lock is held 2791 * in read mode (hence no one can unmap the range being migrated). 2792 * 2793 * Once the caller is done cleaning up things and updating its page table (if it 2794 * chose to do so, this is not an obligation) it finally calls 2795 * migrate_vma_finalize() to update the CPU page table to point to new pages 2796 * for successfully migrated pages or otherwise restore the CPU page table to 2797 * point to the original source pages. 2798 */ 2799 int migrate_vma_setup(struct migrate_vma *args) 2800 { 2801 long nr_pages = (args->end - args->start) >> PAGE_SHIFT; 2802 2803 args->start &= PAGE_MASK; 2804 args->end &= PAGE_MASK; 2805 if (!args->vma || is_vm_hugetlb_page(args->vma) || 2806 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma)) 2807 return -EINVAL; 2808 if (nr_pages <= 0) 2809 return -EINVAL; 2810 if (args->start < args->vma->vm_start || 2811 args->start >= args->vma->vm_end) 2812 return -EINVAL; 2813 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end) 2814 return -EINVAL; 2815 if (!args->src || !args->dst) 2816 return -EINVAL; 2817 2818 memset(args->src, 0, sizeof(*args->src) * nr_pages); 2819 args->cpages = 0; 2820 args->npages = 0; 2821 2822 migrate_vma_collect(args); 2823 2824 if (args->cpages) 2825 migrate_vma_prepare(args); 2826 if (args->cpages) 2827 migrate_vma_unmap(args); 2828 2829 /* 2830 * At this point pages are locked and unmapped, and thus they have 2831 * stable content and can safely be copied to destination memory that 2832 * is allocated by the drivers. 2833 */ 2834 return 0; 2835 2836 } 2837 EXPORT_SYMBOL(migrate_vma_setup); 2838 2839 /* 2840 * This code closely matches the code in: 2841 * __handle_mm_fault() 2842 * handle_pte_fault() 2843 * do_anonymous_page() 2844 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE 2845 * private page. 2846 */ 2847 static void migrate_vma_insert_page(struct migrate_vma *migrate, 2848 unsigned long addr, 2849 struct page *page, 2850 unsigned long *src, 2851 unsigned long *dst) 2852 { 2853 struct vm_area_struct *vma = migrate->vma; 2854 struct mm_struct *mm = vma->vm_mm; 2855 bool flush = false; 2856 spinlock_t *ptl; 2857 pte_t entry; 2858 pgd_t *pgdp; 2859 p4d_t *p4dp; 2860 pud_t *pudp; 2861 pmd_t *pmdp; 2862 pte_t *ptep; 2863 2864 /* Only allow populating anonymous memory */ 2865 if (!vma_is_anonymous(vma)) 2866 goto abort; 2867 2868 pgdp = pgd_offset(mm, addr); 2869 p4dp = p4d_alloc(mm, pgdp, addr); 2870 if (!p4dp) 2871 goto abort; 2872 pudp = pud_alloc(mm, p4dp, addr); 2873 if (!pudp) 2874 goto abort; 2875 pmdp = pmd_alloc(mm, pudp, addr); 2876 if (!pmdp) 2877 goto abort; 2878 2879 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp)) 2880 goto abort; 2881 2882 /* 2883 * Use pte_alloc() instead of pte_alloc_map(). We can't run 2884 * pte_offset_map() on pmds where a huge pmd might be created 2885 * from a different thread. 2886 * 2887 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when 2888 * parallel threads are excluded by other means. 2889 * 2890 * Here we only have mmap_read_lock(mm). 2891 */ 2892 if (pte_alloc(mm, pmdp)) 2893 goto abort; 2894 2895 /* See the comment in pte_alloc_one_map() */ 2896 if (unlikely(pmd_trans_unstable(pmdp))) 2897 goto abort; 2898 2899 if (unlikely(anon_vma_prepare(vma))) 2900 goto abort; 2901 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) 2902 goto abort; 2903 2904 /* 2905 * The memory barrier inside __SetPageUptodate makes sure that 2906 * preceding stores to the page contents become visible before 2907 * the set_pte_at() write. 2908 */ 2909 __SetPageUptodate(page); 2910 2911 if (is_zone_device_page(page)) { 2912 if (is_device_private_page(page)) { 2913 swp_entry_t swp_entry; 2914 2915 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE); 2916 entry = swp_entry_to_pte(swp_entry); 2917 } 2918 } else { 2919 entry = mk_pte(page, vma->vm_page_prot); 2920 if (vma->vm_flags & VM_WRITE) 2921 entry = pte_mkwrite(pte_mkdirty(entry)); 2922 } 2923 2924 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl); 2925 2926 if (check_stable_address_space(mm)) 2927 goto unlock_abort; 2928 2929 if (pte_present(*ptep)) { 2930 unsigned long pfn = pte_pfn(*ptep); 2931 2932 if (!is_zero_pfn(pfn)) 2933 goto unlock_abort; 2934 flush = true; 2935 } else if (!pte_none(*ptep)) 2936 goto unlock_abort; 2937 2938 /* 2939 * Check for userfaultfd but do not deliver the fault. Instead, 2940 * just back off. 2941 */ 2942 if (userfaultfd_missing(vma)) 2943 goto unlock_abort; 2944 2945 inc_mm_counter(mm, MM_ANONPAGES); 2946 page_add_new_anon_rmap(page, vma, addr, false); 2947 if (!is_zone_device_page(page)) 2948 lru_cache_add_inactive_or_unevictable(page, vma); 2949 get_page(page); 2950 2951 if (flush) { 2952 flush_cache_page(vma, addr, pte_pfn(*ptep)); 2953 ptep_clear_flush_notify(vma, addr, ptep); 2954 set_pte_at_notify(mm, addr, ptep, entry); 2955 update_mmu_cache(vma, addr, ptep); 2956 } else { 2957 /* No need to invalidate - it was non-present before */ 2958 set_pte_at(mm, addr, ptep, entry); 2959 update_mmu_cache(vma, addr, ptep); 2960 } 2961 2962 pte_unmap_unlock(ptep, ptl); 2963 *src = MIGRATE_PFN_MIGRATE; 2964 return; 2965 2966 unlock_abort: 2967 pte_unmap_unlock(ptep, ptl); 2968 abort: 2969 *src &= ~MIGRATE_PFN_MIGRATE; 2970 } 2971 2972 /** 2973 * migrate_vma_pages() - migrate meta-data from src page to dst page 2974 * @migrate: migrate struct containing all migration information 2975 * 2976 * This migrates struct page meta-data from source struct page to destination 2977 * struct page. This effectively finishes the migration from source page to the 2978 * destination page. 2979 */ 2980 void migrate_vma_pages(struct migrate_vma *migrate) 2981 { 2982 const unsigned long npages = migrate->npages; 2983 const unsigned long start = migrate->start; 2984 struct mmu_notifier_range range; 2985 unsigned long addr, i; 2986 bool notified = false; 2987 2988 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) { 2989 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); 2990 struct page *page = migrate_pfn_to_page(migrate->src[i]); 2991 struct address_space *mapping; 2992 int r; 2993 2994 if (!newpage) { 2995 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 2996 continue; 2997 } 2998 2999 if (!page) { 3000 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 3001 continue; 3002 if (!notified) { 3003 notified = true; 3004 3005 mmu_notifier_range_init(&range, 3006 MMU_NOTIFY_CLEAR, 0, 3007 NULL, 3008 migrate->vma->vm_mm, 3009 addr, migrate->end); 3010 mmu_notifier_invalidate_range_start(&range); 3011 } 3012 migrate_vma_insert_page(migrate, addr, newpage, 3013 &migrate->src[i], 3014 &migrate->dst[i]); 3015 continue; 3016 } 3017 3018 mapping = page_mapping(page); 3019 3020 if (is_zone_device_page(newpage)) { 3021 if (is_device_private_page(newpage)) { 3022 /* 3023 * For now only support private anonymous when 3024 * migrating to un-addressable device memory. 3025 */ 3026 if (mapping) { 3027 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3028 continue; 3029 } 3030 } else { 3031 /* 3032 * Other types of ZONE_DEVICE page are not 3033 * supported. 3034 */ 3035 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3036 continue; 3037 } 3038 } 3039 3040 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY); 3041 if (r != MIGRATEPAGE_SUCCESS) 3042 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE; 3043 } 3044 3045 /* 3046 * No need to double call mmu_notifier->invalidate_range() callback as 3047 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page() 3048 * did already call it. 3049 */ 3050 if (notified) 3051 mmu_notifier_invalidate_range_only_end(&range); 3052 } 3053 EXPORT_SYMBOL(migrate_vma_pages); 3054 3055 /** 3056 * migrate_vma_finalize() - restore CPU page table entry 3057 * @migrate: migrate struct containing all migration information 3058 * 3059 * This replaces the special migration pte entry with either a mapping to the 3060 * new page if migration was successful for that page, or to the original page 3061 * otherwise. 3062 * 3063 * This also unlocks the pages and puts them back on the lru, or drops the extra 3064 * refcount, for device pages. 3065 */ 3066 void migrate_vma_finalize(struct migrate_vma *migrate) 3067 { 3068 const unsigned long npages = migrate->npages; 3069 unsigned long i; 3070 3071 for (i = 0; i < npages; i++) { 3072 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]); 3073 struct page *page = migrate_pfn_to_page(migrate->src[i]); 3074 3075 if (!page) { 3076 if (newpage) { 3077 unlock_page(newpage); 3078 put_page(newpage); 3079 } 3080 continue; 3081 } 3082 3083 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) { 3084 if (newpage) { 3085 unlock_page(newpage); 3086 put_page(newpage); 3087 } 3088 newpage = page; 3089 } 3090 3091 remove_migration_ptes(page, newpage, false); 3092 unlock_page(page); 3093 3094 if (is_zone_device_page(page)) 3095 put_page(page); 3096 else 3097 putback_lru_page(page); 3098 3099 if (newpage != page) { 3100 unlock_page(newpage); 3101 if (is_zone_device_page(newpage)) 3102 put_page(newpage); 3103 else 3104 putback_lru_page(newpage); 3105 } 3106 } 3107 } 3108 EXPORT_SYMBOL(migrate_vma_finalize); 3109 #endif /* CONFIG_DEVICE_PRIVATE */ 3110