mm/migrate.c

   1 /*
   2  * Memory Migration functionality - linux/mm/migrate.c
   3  *
   4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5  *
   6  * Page migration was first developed in the context of the memory hotplug
   7  * project. The main authors of the migration code are:
   8  *
   9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10  * Hirokazu Takahashi <taka@valinux.co.jp>
  11  * Dave Hansen <haveblue@us.ibm.com>
  12  * Christoph Lameter
  13  */
  14
  15 #include <linux/migrate.h>
  16 #include <linux/export.h>
  17 #include <linux/swap.h>
  18 #include <linux/swapops.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/buffer_head.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/nsproxy.h>
  23 #include <linux/pagevec.h>
  24 #include <linux/ksm.h>
  25 #include <linux/rmap.h>
  26 #include <linux/topology.h>
  27 #include <linux/cpu.h>
  28 #include <linux/cpuset.h>
  29 #include <linux/writeback.h>
  30 #include <linux/mempolicy.h>
  31 #include <linux/vmalloc.h>
  32 #include <linux/security.h>
  33 #include <linux/backing-dev.h>
  34 #include <linux/compaction.h>
  35 #include <linux/syscalls.h>
  36 #include <linux/hugetlb.h>
  37 #include <linux/hugetlb_cgroup.h>
  38 #include <linux/gfp.h>
  39 #include <linux/balloon_compaction.h>
  40 #include <linux/mmu_notifier.h>
  41 #include <linux/page_idle.h>
  42 #include <linux/page_owner.h>
  43
  44 #include <asm/tlbflush.h>
  45
  46 #define CREATE_TRACE_POINTS
  47 #include <trace/events/migrate.h>
  48
  49 #include "internal.h"
  50
  51 /*
  52  * migrate_prep() needs to be called before we start compiling a list of pages
  53  * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
  54  * undesirable, use migrate_prep_local()
  55  */
  56 int migrate_prep(void)
  57 {
  58         /*
  59          * Clear the LRU lists so pages can be isolated.
  60          * Note that pages may be moved off the LRU after we have
  61          * drained them. Those pages will fail to migrate like other
  62          * pages that may be busy.
  63          */
  64         lru_add_drain_all();
  65
  66         return 0;
  67 }
  68
  69 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
  70 int migrate_prep_local(void)
  71 {
  72         lru_add_drain();
  73
  74         return 0;
  75 }
  76
  77 int isolate_movable_page(struct page *page, isolate_mode_t mode)
  78 {
  79         struct address_space *mapping;
  80
  81         /*
  82          * Avoid burning cycles with pages that are yet under __free_pages(),
  83          * or just got freed under us.
  84          *
  85          * In case we 'win' a race for a movable page being freed under us and
  86          * raise its refcount preventing __free_pages() from doing its job
  87          * the put_page() at the end of this block will take care of
  88          * release this page, thus avoiding a nasty leakage.
  89          */
  90         if (unlikely(!get_page_unless_zero(page)))
  91                 goto out;
  92
  93         /*
  94          * Check PageMovable before holding a PG_lock because page's owner
  95          * assumes anybody doesn't touch PG_lock of newly allocated page
  96          * so unconditionally grapping the lock ruins page's owner side.
  97          */
  98         if (unlikely(!__PageMovable(page)))
  99                 goto out_putpage;
 100         /*
 101          * As movable pages are not isolated from LRU lists, concurrent
 102          * compaction threads can race against page migration functions
 103          * as well as race against the releasing a page.
 104          *
 105          * In order to avoid having an already isolated movable page
 106          * being (wrongly) re-isolated while it is under migration,
 107          * or to avoid attempting to isolate pages being released,
 108          * lets be sure we have the page lock
 109          * before proceeding with the movable page isolation steps.
 110          */
 111         if (unlikely(!trylock_page(page)))
 112                 goto out_putpage;
 113
 114         if (!PageMovable(page) || PageIsolated(page))
 115                 goto out_no_isolated;
 116
 117         mapping = page_mapping(page);
 118         VM_BUG_ON_PAGE(!mapping, page);
 119
 120         if (!mapping->a_ops->isolate_page(page, mode))
 121                 goto out_no_isolated;
 122
 123         /* Driver shouldn't use PG_isolated bit of page->flags */
 124         WARN_ON_ONCE(PageIsolated(page));
 125         __SetPageIsolated(page);
 126         unlock_page(page);
 127
 128         return 0;
 129
 130 out_no_isolated:
 131         unlock_page(page);
 132 out_putpage:
 133         put_page(page);
 134 out:
 135         return -EBUSY;
 136 }
 137
 138 /* It should be called on page which is PG_movable */
 139 void putback_movable_page(struct page *page)
 140 {
 141         struct address_space *mapping;
 142
 143         VM_BUG_ON_PAGE(!PageLocked(page), page);
 144         VM_BUG_ON_PAGE(!PageMovable(page), page);
 145         VM_BUG_ON_PAGE(!PageIsolated(page), page);
 146
 147         mapping = page_mapping(page);
 148         mapping->a_ops->putback_page(page);
 149         __ClearPageIsolated(page);
 150 }
 151
 152 /*
 153  * Put previously isolated pages back onto the appropriate lists
 154  * from where they were once taken off for compaction/migration.
 155  *
 156  * This function shall be used whenever the isolated pageset has been
 157  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
 158  * and isolate_huge_page().
 159  */
 160 void putback_movable_pages(struct list_head *l)
 161 {
 162         struct page *page;
 163         struct page *page2;
 164
 165         list_for_each_entry_safe(page, page2, l, lru) {
 166                 if (unlikely(PageHuge(page))) {
 167                         putback_active_hugepage(page);
 168                         continue;
 169                 }
 170                 list_del(&page->lru);
 171                 /*
 172                  * We isolated non-lru movable page so here we can use
 173                  * __PageMovable because LRU page's mapping cannot have
 174                  * PAGE_MAPPING_MOVABLE.
 175                  */
 176                 if (unlikely(__PageMovable(page))) {
 177                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
 178                         lock_page(page);
 179                         if (PageMovable(page))
 180                                 putback_movable_page(page);
 181                         else
 182                                 __ClearPageIsolated(page);
 183                         unlock_page(page);
 184                         put_page(page);
 185                 } else {
 186                         putback_lru_page(page);
 187                         dec_node_page_state(page, NR_ISOLATED_ANON +
 188                                         page_is_file_cache(page));
 189                 }
 190         }
 191 }
 192
 193 /*
 194  * Restore a potential migration pte to a working pte entry
 195  */
 196 static int remove_migration_pte(struct page *page, struct vm_area_struct *vma,
 197                                  unsigned long addr, void *old)
 198 {
 199         struct page_vma_mapped_walk pvmw = {
 200                 .page = old,
 201                 .vma = vma,
 202                 .address = addr,
 203                 .flags = PVMW_SYNC | PVMW_MIGRATION,
 204         };
 205         struct page *new;
 206         pte_t pte;
 207         swp_entry_t entry;
 208
 209         VM_BUG_ON_PAGE(PageTail(page), page);
 210         while (page_vma_mapped_walk(&pvmw)) {
 211                 new = page - pvmw.page->index +
 212                         linear_page_index(vma, pvmw.address);
 213
 214                 get_page(new);
 215                 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
 216                 if (pte_swp_soft_dirty(*pvmw.pte))
 217                         pte = pte_mksoft_dirty(pte);
 218
 219                 /*
 220                  * Recheck VMA as permissions can change since migration started
 221                  */
 222                 entry = pte_to_swp_entry(*pvmw.pte);
 223                 if (is_write_migration_entry(entry))
 224                         pte = maybe_mkwrite(pte, vma);
 225
 226 #ifdef CONFIG_HUGETLB_PAGE
 227                 if (PageHuge(new)) {
 228                         pte = pte_mkhuge(pte);
 229                         pte = arch_make_huge_pte(pte, vma, new, 0);
 230                 }
 231 #endif
 232                 flush_dcache_page(new);
 233                 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
 234
 235                 if (PageHuge(new)) {
 236                         if (PageAnon(new))
 237                                 hugepage_add_anon_rmap(new, vma, pvmw.address);
 238                         else
 239                                 page_dup_rmap(new, true);
 240                 } else if (PageAnon(new))
 241                         page_add_anon_rmap(new, vma, pvmw.address, false);
 242                 else
 243                         page_add_file_rmap(new, false);
 244
 245                 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
 246                         mlock_vma_page(new);
 247
 248                 /* No need to invalidate - it was non-present before */
 249                 update_mmu_cache(vma, pvmw.address, pvmw.pte);
 250         }
 251
 252         return SWAP_AGAIN;
 253 }
 254
 255 /*
 256  * Get rid of all migration entries and replace them by
 257  * references to the indicated page.
 258  */
 259 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
 260 {
 261         struct rmap_walk_control rwc = {
 262                 .rmap_one = remove_migration_pte,
 263                 .arg = old,
 264         };
 265
 266         if (locked)
 267                 rmap_walk_locked(new, &rwc);
 268         else
 269                 rmap_walk(new, &rwc);
 270 }
 271
 272 /*
 273  * Something used the pte of a page under migration. We need to
 274  * get to the page and wait until migration is finished.
 275  * When we return from this function the fault will be retried.
 276  */
 277 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
 278                                 spinlock_t *ptl)
 279 {
 280         pte_t pte;
 281         swp_entry_t entry;
 282         struct page *page;
 283
 284         spin_lock(ptl);
 285         pte = *ptep;
 286         if (!is_swap_pte(pte))
 287                 goto out;
 288
 289         entry = pte_to_swp_entry(pte);
 290         if (!is_migration_entry(entry))
 291                 goto out;
 292
 293         page = migration_entry_to_page(entry);
 294
 295         /*
 296          * Once radix-tree replacement of page migration started, page_count
 297          * *must* be zero. And, we don't want to call wait_on_page_locked()
 298          * against a page without get_page().
 299          * So, we use get_page_unless_zero(), here. Even failed, page fault
 300          * will occur again.
 301          */
 302         if (!get_page_unless_zero(page))
 303                 goto out;
 304         pte_unmap_unlock(ptep, ptl);
 305         wait_on_page_locked(page);
 306         put_page(page);
 307         return;
 308 out:
 309         pte_unmap_unlock(ptep, ptl);
 310 }
 311
 312 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 313                                 unsigned long address)
 314 {
 315         spinlock_t *ptl = pte_lockptr(mm, pmd);
 316         pte_t *ptep = pte_offset_map(pmd, address);
 317         __migration_entry_wait(mm, ptep, ptl);
 318 }
 319
 320 void migration_entry_wait_huge(struct vm_area_struct *vma,
 321                 struct mm_struct *mm, pte_t *pte)
 322 {
 323         spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
 324         __migration_entry_wait(mm, pte, ptl);
 325 }
 326
 327 #ifdef CONFIG_BLOCK
 328 /* Returns true if all buffers are successfully locked */
 329 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
 330                                                         enum migrate_mode mode)
 331 {
 332         struct buffer_head *bh = head;
 333
 334         /* Simple case, sync compaction */
 335         if (mode != MIGRATE_ASYNC) {
 336                 do {
 337                         get_bh(bh);
 338                         lock_buffer(bh);
 339                         bh = bh->b_this_page;
 340
 341                 } while (bh != head);
 342
 343                 return true;
 344         }
 345
 346         /* async case, we cannot block on lock_buffer so use trylock_buffer */
 347         do {
 348                 get_bh(bh);
 349                 if (!trylock_buffer(bh)) {
 350                         /*
 351                          * We failed to lock the buffer and cannot stall in
 352                          * async migration. Release the taken locks
 353                          */
 354                         struct buffer_head *failed_bh = bh;
 355                         put_bh(failed_bh);
 356                         bh = head;
 357                         while (bh != failed_bh) {
 358                                 unlock_buffer(bh);
 359                                 put_bh(bh);
 360                                 bh = bh->b_this_page;
 361                         }
 362                         return false;
 363                 }
 364
 365                 bh = bh->b_this_page;
 366         } while (bh != head);
 367         return true;
 368 }
 369 #else
 370 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
 371                                                         enum migrate_mode mode)
 372 {
 373         return true;
 374 }
 375 #endif /* CONFIG_BLOCK */
 376
 377 /*
 378  * Replace the page in the mapping.
 379  *
 380  * The number of remaining references must be:
 381  * 1 for anonymous pages without a mapping
 382  * 2 for pages with a mapping
 383  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 384  */
 385 int migrate_page_move_mapping(struct address_space *mapping,
 386                 struct page *newpage, struct page *page,
 387                 struct buffer_head *head, enum migrate_mode mode,
 388                 int extra_count)
 389 {
 390         struct zone *oldzone, *newzone;
 391         int dirty;
 392         int expected_count = 1 + extra_count;
 393         void **pslot;
 394
 395         if (!mapping) {
 396                 /* Anonymous page without mapping */
 397                 if (page_count(page) != expected_count)
 398                         return -EAGAIN;
 399
 400                 /* No turning back from here */
 401                 newpage->index = page->index;
 402                 newpage->mapping = page->mapping;
 403                 if (PageSwapBacked(page))
 404                         __SetPageSwapBacked(newpage);
 405
 406                 return MIGRATEPAGE_SUCCESS;
 407         }
 408
 409         oldzone = page_zone(page);
 410         newzone = page_zone(newpage);
 411
 412         spin_lock_irq(&mapping->tree_lock);
 413
 414         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 415                                         page_index(page));
 416
 417         expected_count += 1 + page_has_private(page);
 418         if (page_count(page) != expected_count ||
 419                 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
 420                 spin_unlock_irq(&mapping->tree_lock);
 421                 return -EAGAIN;
 422         }
 423
 424         if (!page_ref_freeze(page, expected_count)) {
 425                 spin_unlock_irq(&mapping->tree_lock);
 426                 return -EAGAIN;
 427         }
 428
 429         /*
 430          * In the async migration case of moving a page with buffers, lock the
 431          * buffers using trylock before the mapping is moved. If the mapping
 432          * was moved, we later failed to lock the buffers and could not move
 433          * the mapping back due to an elevated page count, we would have to
 434          * block waiting on other references to be dropped.
 435          */
 436         if (mode == MIGRATE_ASYNC && head &&
 437                         !buffer_migrate_lock_buffers(head, mode)) {
 438                 page_ref_unfreeze(page, expected_count);
 439                 spin_unlock_irq(&mapping->tree_lock);
 440                 return -EAGAIN;
 441         }
 442
 443         /*
 444          * Now we know that no one else is looking at the page:
 445          * no turning back from here.
 446          */
 447         newpage->index = page->index;
 448         newpage->mapping = page->mapping;
 449         get_page(newpage);      /* add cache reference */
 450         if (PageSwapBacked(page)) {
 451                 __SetPageSwapBacked(newpage);
 452                 if (PageSwapCache(page)) {
 453                         SetPageSwapCache(newpage);
 454                         set_page_private(newpage, page_private(page));
 455                 }
 456         } else {
 457                 VM_BUG_ON_PAGE(PageSwapCache(page), page);
 458         }
 459
 460         /* Move dirty while page refs frozen and newpage not yet exposed */
 461         dirty = PageDirty(page);
 462         if (dirty) {
 463                 ClearPageDirty(page);
 464                 SetPageDirty(newpage);
 465         }
 466
 467         radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
 468
 469         /*
 470          * Drop cache reference from old page by unfreezing
 471          * to one less reference.
 472          * We know this isn't the last reference.
 473          */
 474         page_ref_unfreeze(page, expected_count - 1);
 475
 476         spin_unlock(&mapping->tree_lock);
 477         /* Leave irq disabled to prevent preemption while updating stats */
 478
 479         /*
 480          * If moved to a different zone then also account
 481          * the page for that zone. Other VM counters will be
 482          * taken care of when we establish references to the
 483          * new page and drop references to the old page.
 484          *
 485          * Note that anonymous pages are accounted for
 486          * via NR_FILE_PAGES and NR_ANON_MAPPED if they
 487          * are mapped to swap space.
 488          */
 489         if (newzone != oldzone) {
 490                 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
 491                 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
 492                 if (PageSwapBacked(page) && !PageSwapCache(page)) {
 493                         __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
 494                         __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
 495                 }
 496                 if (dirty && mapping_cap_account_dirty(mapping)) {
 497                         __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
 498                         __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
 499                         __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
 500                         __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
 501                 }
 502         }
 503         local_irq_enable();
 504
 505         return MIGRATEPAGE_SUCCESS;
 506 }
 507 EXPORT_SYMBOL(migrate_page_move_mapping);
 508
 509 /*
 510  * The expected number of remaining references is the same as that
 511  * of migrate_page_move_mapping().
 512  */
 513 int migrate_huge_page_move_mapping(struct address_space *mapping,
 514                                    struct page *newpage, struct page *page)
 515 {
 516         int expected_count;
 517         void **pslot;
 518
 519         spin_lock_irq(&mapping->tree_lock);
 520
 521         pslot = radix_tree_lookup_slot(&mapping->page_tree,
 522                                         page_index(page));
 523
 524         expected_count = 2 + page_has_private(page);
 525         if (page_count(page) != expected_count ||
 526                 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
 527                 spin_unlock_irq(&mapping->tree_lock);
 528                 return -EAGAIN;
 529         }
 530
 531         if (!page_ref_freeze(page, expected_count)) {
 532                 spin_unlock_irq(&mapping->tree_lock);
 533                 return -EAGAIN;
 534         }
 535
 536         newpage->index = page->index;
 537         newpage->mapping = page->mapping;
 538
 539         get_page(newpage);
 540
 541         radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
 542
 543         page_ref_unfreeze(page, expected_count - 1);
 544
 545         spin_unlock_irq(&mapping->tree_lock);
 546
 547         return MIGRATEPAGE_SUCCESS;
 548 }
 549
 550 /*
 551  * Gigantic pages are so large that we do not guarantee that page++ pointer
 552  * arithmetic will work across the entire page.  We need something more
 553  * specialized.
 554  */
 555 static void __copy_gigantic_page(struct page *dst, struct page *src,
 556                                 int nr_pages)
 557 {
 558         int i;
 559         struct page *dst_base = dst;
 560         struct page *src_base = src;
 561
 562         for (i = 0; i < nr_pages; ) {
 563                 cond_resched();
 564                 copy_highpage(dst, src);
 565
 566                 i++;
 567                 dst = mem_map_next(dst, dst_base, i);
 568                 src = mem_map_next(src, src_base, i);
 569         }
 570 }
 571
 572 static void copy_huge_page(struct page *dst, struct page *src)
 573 {
 574         int i;
 575         int nr_pages;
 576
 577         if (PageHuge(src)) {
 578                 /* hugetlbfs page */
 579                 struct hstate *h = page_hstate(src);
 580                 nr_pages = pages_per_huge_page(h);
 581
 582                 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
 583                         __copy_gigantic_page(dst, src, nr_pages);
 584                         return;
 585                 }
 586         } else {
 587                 /* thp page */
 588                 BUG_ON(!PageTransHuge(src));
 589                 nr_pages = hpage_nr_pages(src);
 590         }
 591
 592         for (i = 0; i < nr_pages; i++) {
 593                 cond_resched();
 594                 copy_highpage(dst + i, src + i);
 595         }
 596 }
 597
 598 /*
 599  * Copy the page to its new location
 600  */
 601 void migrate_page_copy(struct page *newpage, struct page *page)
 602 {
 603         int cpupid;
 604
 605         if (PageHuge(page) || PageTransHuge(page))
 606                 copy_huge_page(newpage, page);
 607         else
 608                 copy_highpage(newpage, page);
 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 (PageChecked(page))
 622                 SetPageChecked(newpage);
 623         if (PageMappedToDisk(page))
 624                 SetPageMappedToDisk(newpage);
 625
 626         /* Move dirty on pages not done by migrate_page_move_mapping() */
 627         if (PageDirty(page))
 628                 SetPageDirty(newpage);
 629
 630         if (page_is_young(page))
 631                 set_page_young(newpage);
 632         if (page_is_idle(page))
 633                 set_page_idle(newpage);
 634
 635         /*
 636          * Copy NUMA information to the new page, to prevent over-eager
 637          * future migrations of this same page.
 638          */
 639         cpupid = page_cpupid_xchg_last(page, -1);
 640         page_cpupid_xchg_last(newpage, cpupid);
 641
 642         ksm_migrate_page(newpage, page);
 643         /*
 644          * Please do not reorder this without considering how mm/ksm.c's
 645          * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
 646          */
 647         if (PageSwapCache(page))
 648                 ClearPageSwapCache(page);
 649         ClearPagePrivate(page);
 650         set_page_private(page, 0);
 651
 652         /*
 653          * If any waiters have accumulated on the new page then
 654          * wake them up.
 655          */
 656         if (PageWriteback(newpage))
 657                 end_page_writeback(newpage);
 658
 659         copy_page_owner(page, newpage);
 660
 661         mem_cgroup_migrate(page, newpage);
 662 }
 663 EXPORT_SYMBOL(migrate_page_copy);
 664
 665 /************************************************************
 666  *                    Migration functions
 667  ***********************************************************/
 668
 669 /*
 670  * Common logic to directly migrate a single LRU page suitable for
 671  * pages that do not use PagePrivate/PagePrivate2.
 672  *
 673  * Pages are locked upon entry and exit.
 674  */
 675 int migrate_page(struct address_space *mapping,
 676                 struct page *newpage, struct page *page,
 677                 enum migrate_mode mode)
 678 {
 679         int rc;
 680
 681         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 682
 683         rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 684
 685         if (rc != MIGRATEPAGE_SUCCESS)
 686                 return rc;
 687
 688         migrate_page_copy(newpage, page);
 689         return MIGRATEPAGE_SUCCESS;
 690 }
 691 EXPORT_SYMBOL(migrate_page);
 692
 693 #ifdef CONFIG_BLOCK
 694 /*
 695  * Migration function for pages with buffers. This function can only be used
 696  * if the underlying filesystem guarantees that no other references to "page"
 697  * exist.
 698  */
 699 int buffer_migrate_page(struct address_space *mapping,
 700                 struct page *newpage, struct page *page, enum migrate_mode mode)
 701 {
 702         struct buffer_head *bh, *head;
 703         int rc;
 704
 705         if (!page_has_buffers(page))
 706                 return migrate_page(mapping, newpage, page, mode);
 707
 708         head = page_buffers(page);
 709
 710         rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
 711
 712         if (rc != MIGRATEPAGE_SUCCESS)
 713                 return rc;
 714
 715         /*
 716          * In the async case, migrate_page_move_mapping locked the buffers
 717          * with an IRQ-safe spinlock held. In the sync case, the buffers
 718          * need to be locked now
 719          */
 720         if (mode != MIGRATE_ASYNC)
 721                 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
 722
 723         ClearPagePrivate(page);
 724         set_page_private(newpage, page_private(page));
 725         set_page_private(page, 0);
 726         put_page(page);
 727         get_page(newpage);
 728
 729         bh = head;
 730         do {
 731                 set_bh_page(bh, newpage, bh_offset(bh));
 732                 bh = bh->b_this_page;
 733
 734         } while (bh != head);
 735
 736         SetPagePrivate(newpage);
 737
 738         migrate_page_copy(newpage, page);
 739
 740         bh = head;
 741         do {
 742                 unlock_buffer(bh);
 743                 put_bh(bh);
 744                 bh = bh->b_this_page;
 745
 746         } while (bh != head);
 747
 748         return MIGRATEPAGE_SUCCESS;
 749 }
 750 EXPORT_SYMBOL(buffer_migrate_page);
 751 #endif
 752
 753 /*
 754  * Writeback a page to clean the dirty state
 755  */
 756 static int writeout(struct address_space *mapping, struct page *page)
 757 {
 758         struct writeback_control wbc = {
 759                 .sync_mode = WB_SYNC_NONE,
 760                 .nr_to_write = 1,
 761                 .range_start = 0,
 762                 .range_end = LLONG_MAX,
 763                 .for_reclaim = 1
 764         };
 765         int rc;
 766
 767         if (!mapping->a_ops->writepage)
 768                 /* No write method for the address space */
 769                 return -EINVAL;
 770
 771         if (!clear_page_dirty_for_io(page))
 772                 /* Someone else already triggered a write */
 773                 return -EAGAIN;
 774
 775         /*
 776          * A dirty page may imply that the underlying filesystem has
 777          * the page on some queue. So the page must be clean for
 778          * migration. Writeout may mean we loose the lock and the
 779          * page state is no longer what we checked for earlier.
 780          * At this point we know that the migration attempt cannot
 781          * be successful.
 782          */
 783         remove_migration_ptes(page, page, false);
 784
 785         rc = mapping->a_ops->writepage(page, &wbc);
 786
 787         if (rc != AOP_WRITEPAGE_ACTIVATE)
 788                 /* unlocked. Relock */
 789                 lock_page(page);
 790
 791         return (rc < 0) ? -EIO : -EAGAIN;
 792 }
 793
 794 /*
 795  * Default handling if a filesystem does not provide a migration function.
 796  */
 797 static int fallback_migrate_page(struct address_space *mapping,
 798         struct page *newpage, struct page *page, enum migrate_mode mode)
 799 {
 800         if (PageDirty(page)) {
 801                 /* Only writeback pages in full synchronous migration */
 802                 if (mode != MIGRATE_SYNC)
 803                         return -EBUSY;
 804                 return writeout(mapping, page);
 805         }
 806
 807         /*
 808          * Buffers may be managed in a filesystem specific way.
 809          * We must have no buffers or drop them.
 810          */
 811         if (page_has_private(page) &&
 812             !try_to_release_page(page, GFP_KERNEL))
 813                 return -EAGAIN;
 814
 815         return migrate_page(mapping, newpage, page, mode);
 816 }
 817
 818 /*
 819  * Move a page to a newly allocated page
 820  * The page is locked and all ptes have been successfully removed.
 821  *
 822  * The new page will have replaced the old page if this function
 823  * is successful.
 824  *
 825  * Return value:
 826  *   < 0 - error code
 827  *  MIGRATEPAGE_SUCCESS - success
 828  */
 829 static int move_to_new_page(struct page *newpage, struct page *page,
 830                                 enum migrate_mode mode)
 831 {
 832         struct address_space *mapping;
 833         int rc = -EAGAIN;
 834         bool is_lru = !__PageMovable(page);
 835
 836         VM_BUG_ON_PAGE(!PageLocked(page), page);
 837         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
 838
 839         mapping = page_mapping(page);
 840
 841         if (likely(is_lru)) {
 842                 if (!mapping)
 843                         rc = migrate_page(mapping, newpage, page, mode);
 844                 else if (mapping->a_ops->migratepage)
 845                         /*
 846                          * Most pages have a mapping and most filesystems
 847                          * provide a migratepage callback. Anonymous pages
 848                          * are part of swap space which also has its own
 849                          * migratepage callback. This is the most common path
 850                          * for page migration.
 851                          */
 852                         rc = mapping->a_ops->migratepage(mapping, newpage,
 853                                                         page, mode);
 854                 else
 855                         rc = fallback_migrate_page(mapping, newpage,
 856                                                         page, mode);
 857         } else {
 858                 /*
 859                  * In case of non-lru page, it could be released after
 860                  * isolation step. In that case, we shouldn't try migration.
 861                  */
 862                 VM_BUG_ON_PAGE(!PageIsolated(page), page);
 863                 if (!PageMovable(page)) {
 864                         rc = MIGRATEPAGE_SUCCESS;
 865                         __ClearPageIsolated(page);
 866                         goto out;
 867                 }
 868
 869                 rc = mapping->a_ops->migratepage(mapping, newpage,
 870                                                 page, mode);
 871                 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
 872                         !PageIsolated(page));
 873         }
 874
 875         /*
 876          * When successful, old pagecache page->mapping must be cleared before
 877          * page is freed; but stats require that PageAnon be left as PageAnon.
 878          */
 879         if (rc == MIGRATEPAGE_SUCCESS) {
 880                 if (__PageMovable(page)) {
 881                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
 882
 883                         /*
 884                          * We clear PG_movable under page_lock so any compactor
 885                          * cannot try to migrate this page.
 886                          */
 887                         __ClearPageIsolated(page);
 888                 }
 889
 890                 /*
 891                  * Anonymous and movable page->mapping will be cleard by
 892                  * free_pages_prepare so don't reset it here for keeping
 893                  * the type to work PageAnon, for example.
 894                  */
 895                 if (!PageMappingFlags(page))
 896                         page->mapping = NULL;
 897         }
 898 out:
 899         return rc;
 900 }
 901
 902 static int __unmap_and_move(struct page *page, struct page *newpage,
 903                                 int force, enum migrate_mode mode)
 904 {
 905         int rc = -EAGAIN;
 906         int page_was_mapped = 0;
 907         struct anon_vma *anon_vma = NULL;
 908         bool is_lru = !__PageMovable(page);
 909
 910         if (!trylock_page(page)) {
 911                 if (!force || mode == MIGRATE_ASYNC)
 912                         goto out;
 913
 914                 /*
 915                  * It's not safe for direct compaction to call lock_page.
 916                  * For example, during page readahead pages are added locked
 917                  * to the LRU. Later, when the IO completes the pages are
 918                  * marked uptodate and unlocked. However, the queueing
 919                  * could be merging multiple pages for one bio (e.g.
 920                  * mpage_readpages). If an allocation happens for the
 921                  * second or third page, the process can end up locking
 922                  * the same page twice and deadlocking. Rather than
 923                  * trying to be clever about what pages can be locked,
 924                  * avoid the use of lock_page for direct compaction
 925                  * altogether.
 926                  */
 927                 if (current->flags & PF_MEMALLOC)
 928                         goto out;
 929
 930                 lock_page(page);
 931         }
 932
 933         if (PageWriteback(page)) {
 934                 /*
 935                  * Only in the case of a full synchronous migration is it
 936                  * necessary to wait for PageWriteback. In the async case,
 937                  * the retry loop is too short and in the sync-light case,
 938                  * the overhead of stalling is too much
 939                  */
 940                 if (mode != MIGRATE_SYNC) {
 941                         rc = -EBUSY;
 942                         goto out_unlock;
 943                 }
 944                 if (!force)
 945                         goto out_unlock;
 946                 wait_on_page_writeback(page);
 947         }
 948
 949         /*
 950          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 951          * we cannot notice that anon_vma is freed while we migrates a page.
 952          * This get_anon_vma() delays freeing anon_vma pointer until the end
 953          * of migration. File cache pages are no problem because of page_lock()
 954          * File Caches may use write_page() or lock_page() in migration, then,
 955          * just care Anon page here.
 956          *
 957          * Only page_get_anon_vma() understands the subtleties of
 958          * getting a hold on an anon_vma from outside one of its mms.
 959          * But if we cannot get anon_vma, then we won't need it anyway,
 960          * because that implies that the anon page is no longer mapped
 961          * (and cannot be remapped so long as we hold the page lock).
 962          */
 963         if (PageAnon(page) && !PageKsm(page))
 964                 anon_vma = page_get_anon_vma(page);
 965
 966         /*
 967          * Block others from accessing the new page when we get around to
 968          * establishing additional references. We are usually the only one
 969          * holding a reference to newpage at this point. We used to have a BUG
 970          * here if trylock_page(newpage) fails, but would like to allow for
 971          * cases where there might be a race with the previous use of newpage.
 972          * This is much like races on refcount of oldpage: just don't BUG().
 973          */
 974         if (unlikely(!trylock_page(newpage)))
 975                 goto out_unlock;
 976
 977         if (unlikely(!is_lru)) {
 978                 rc = move_to_new_page(newpage, page, mode);
 979                 goto out_unlock_both;
 980         }
 981
 982         /*
 983          * Corner case handling:
 984          * 1. When a new swap-cache page is read into, it is added to the LRU
 985          * and treated as swapcache but it has no rmap yet.
 986          * Calling try_to_unmap() against a page->mapping==NULL page will
 987          * trigger a BUG.  So handle it here.
 988          * 2. An orphaned page (see truncate_complete_page) might have
 989          * fs-private metadata. The page can be picked up due to memory
 990          * offlining.  Everywhere else except page reclaim, the page is
 991          * invisible to the vm, so the page can not be migrated.  So try to
 992          * free the metadata, so the page can be freed.
 993          */
 994         if (!page->mapping) {
 995                 VM_BUG_ON_PAGE(PageAnon(page), page);
 996                 if (page_has_private(page)) {
 997                         try_to_free_buffers(page);
 998                         goto out_unlock_both;
 999                 }
1000         } else if (page_mapped(page)) {
1001                 /* Establish migration ptes */
1002                 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1003                                 page);
1004                 try_to_unmap(page,
1005                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1006                 page_was_mapped = 1;
1007         }
1008
1009         if (!page_mapped(page))
1010                 rc = move_to_new_page(newpage, page, mode);
1011
1012         if (page_was_mapped)
1013                 remove_migration_ptes(page,
1014                         rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1015
1016 out_unlock_both:
1017         unlock_page(newpage);
1018 out_unlock:
1019         /* Drop an anon_vma reference if we took one */
1020         if (anon_vma)
1021                 put_anon_vma(anon_vma);
1022         unlock_page(page);
1023 out:
1024         /*
1025          * If migration is successful, decrease refcount of the newpage
1026          * which will not free the page because new page owner increased
1027          * refcounter. As well, if it is LRU page, add the page to LRU
1028          * list in here.
1029          */
1030         if (rc == MIGRATEPAGE_SUCCESS) {
1031                 if (unlikely(__PageMovable(newpage)))
1032                         put_page(newpage);
1033                 else
1034                         putback_lru_page(newpage);
1035         }
1036
1037         return rc;
1038 }
1039
1040 /*
1041  * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move().  Work
1042  * around it.
1043  */
1044 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1045 #define ICE_noinline noinline
1046 #else
1047 #define ICE_noinline
1048 #endif
1049
1050 /*
1051  * Obtain the lock on page, remove all ptes and migrate the page
1052  * to the newly allocated page in newpage.
1053  */
1054 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1055                                    free_page_t put_new_page,
1056                                    unsigned long private, struct page *page,
1057                                    int force, enum migrate_mode mode,
1058                                    enum migrate_reason reason)
1059 {
1060         int rc = MIGRATEPAGE_SUCCESS;
1061         int *result = NULL;
1062         struct page *newpage;
1063
1064         newpage = get_new_page(page, private, &result);
1065         if (!newpage)
1066                 return -ENOMEM;
1067
1068         if (page_count(page) == 1) {
1069                 /* page was freed from under us. So we are done. */
1070                 ClearPageActive(page);
1071                 ClearPageUnevictable(page);
1072                 if (unlikely(__PageMovable(page))) {
1073                         lock_page(page);
1074                         if (!PageMovable(page))
1075                                 __ClearPageIsolated(page);
1076                         unlock_page(page);
1077                 }
1078                 if (put_new_page)
1079                         put_new_page(newpage, private);
1080                 else
1081                         put_page(newpage);
1082                 goto out;
1083         }
1084
1085         if (unlikely(PageTransHuge(page))) {
1086                 lock_page(page);
1087                 rc = split_huge_page(page);
1088                 unlock_page(page);
1089                 if (rc)
1090                         goto out;
1091         }
1092
1093         rc = __unmap_and_move(page, newpage, force, mode);
1094         if (rc == MIGRATEPAGE_SUCCESS)
1095                 set_page_owner_migrate_reason(newpage, reason);
1096
1097 out:
1098         if (rc != -EAGAIN) {
1099                 /*
1100                  * A page that has been migrated has all references
1101                  * removed and will be freed. A page that has not been
1102                  * migrated will have kepts its references and be
1103                  * restored.
1104                  */
1105                 list_del(&page->lru);
1106
1107                 /*
1108                  * Compaction can migrate also non-LRU pages which are
1109                  * not accounted to NR_ISOLATED_*. They can be recognized
1110                  * as __PageMovable
1111                  */
1112                 if (likely(!__PageMovable(page)))
1113                         dec_node_page_state(page, NR_ISOLATED_ANON +
1114                                         page_is_file_cache(page));
1115         }
1116
1117         /*
1118          * If migration is successful, releases reference grabbed during
1119          * isolation. Otherwise, restore the page to right list unless
1120          * we want to retry.
1121          */
1122         if (rc == MIGRATEPAGE_SUCCESS) {
1123                 put_page(page);
1124                 if (reason == MR_MEMORY_FAILURE) {
1125                         /*
1126                          * Set PG_HWPoison on just freed page
1127                          * intentionally. Although it's rather weird,
1128                          * it's how HWPoison flag works at the moment.
1129                          */
1130                         if (!test_set_page_hwpoison(page))
1131                                 num_poisoned_pages_inc();
1132                 }
1133         } else {
1134                 if (rc != -EAGAIN) {
1135                         if (likely(!__PageMovable(page))) {
1136                                 putback_lru_page(page);
1137                                 goto put_new;
1138                         }
1139
1140                         lock_page(page);
1141                         if (PageMovable(page))
1142                                 putback_movable_page(page);
1143                         else
1144                                 __ClearPageIsolated(page);
1145                         unlock_page(page);
1146                         put_page(page);
1147                 }
1148 put_new:
1149                 if (put_new_page)
1150                         put_new_page(newpage, private);
1151                 else
1152                         put_page(newpage);
1153         }
1154
1155         if (result) {
1156                 if (rc)
1157                         *result = rc;
1158                 else
1159                         *result = page_to_nid(newpage);
1160         }
1161         return rc;
1162 }
1163
1164 /*
1165  * Counterpart of unmap_and_move_page() for hugepage migration.
1166  *
1167  * This function doesn't wait the completion of hugepage I/O
1168  * because there is no race between I/O and migration for hugepage.
1169  * Note that currently hugepage I/O occurs only in direct I/O
1170  * where no lock is held and PG_writeback is irrelevant,
1171  * and writeback status of all subpages are counted in the reference
1172  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1173  * under direct I/O, the reference of the head page is 512 and a bit more.)
1174  * This means that when we try to migrate hugepage whose subpages are
1175  * doing direct I/O, some references remain after try_to_unmap() and
1176  * hugepage migration fails without data corruption.
1177  *
1178  * There is also no race when direct I/O is issued on the page under migration,
1179  * because then pte is replaced with migration swap entry and direct I/O code
1180  * will wait in the page fault for migration to complete.
1181  */
1182 static int unmap_and_move_huge_page(new_page_t get_new_page,
1183                                 free_page_t put_new_page, unsigned long private,
1184                                 struct page *hpage, int force,
1185                                 enum migrate_mode mode, int reason)
1186 {
1187         int rc = -EAGAIN;
1188         int *result = NULL;
1189         int page_was_mapped = 0;
1190         struct page *new_hpage;
1191         struct anon_vma *anon_vma = NULL;
1192
1193         /*
1194          * Movability of hugepages depends on architectures and hugepage size.
1195          * This check is necessary because some callers of hugepage migration
1196          * like soft offline and memory hotremove don't walk through page
1197          * tables or check whether the hugepage is pmd-based or not before
1198          * kicking migration.
1199          */
1200         if (!hugepage_migration_supported(page_hstate(hpage))) {
1201                 putback_active_hugepage(hpage);
1202                 return -ENOSYS;
1203         }
1204
1205         new_hpage = get_new_page(hpage, private, &result);
1206         if (!new_hpage)
1207                 return -ENOMEM;
1208
1209         if (!trylock_page(hpage)) {
1210                 if (!force || mode != MIGRATE_SYNC)
1211                         goto out;
1212                 lock_page(hpage);
1213         }
1214
1215         if (PageAnon(hpage))
1216                 anon_vma = page_get_anon_vma(hpage);
1217
1218         if (unlikely(!trylock_page(new_hpage)))
1219                 goto put_anon;
1220
1221         if (page_mapped(hpage)) {
1222                 try_to_unmap(hpage,
1223                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1224                 page_was_mapped = 1;
1225         }
1226
1227         if (!page_mapped(hpage))
1228                 rc = move_to_new_page(new_hpage, hpage, mode);
1229
1230         if (page_was_mapped)
1231                 remove_migration_ptes(hpage,
1232                         rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1233
1234         unlock_page(new_hpage);
1235
1236 put_anon:
1237         if (anon_vma)
1238                 put_anon_vma(anon_vma);
1239
1240         if (rc == MIGRATEPAGE_SUCCESS) {
1241                 hugetlb_cgroup_migrate(hpage, new_hpage);
1242                 put_new_page = NULL;
1243                 set_page_owner_migrate_reason(new_hpage, reason);
1244         }
1245
1246         unlock_page(hpage);
1247 out:
1248         if (rc != -EAGAIN)
1249                 putback_active_hugepage(hpage);
1250
1251         /*
1252          * If migration was not successful and there's a freeing callback, use
1253          * it.  Otherwise, put_page() will drop the reference grabbed during
1254          * isolation.
1255          */
1256         if (put_new_page)
1257                 put_new_page(new_hpage, private);
1258         else
1259                 putback_active_hugepage(new_hpage);
1260
1261         if (result) {
1262                 if (rc)
1263                         *result = rc;
1264                 else
1265                         *result = page_to_nid(new_hpage);
1266         }
1267         return rc;
1268 }
1269
1270 /*
1271  * migrate_pages - migrate the pages specified in a list, to the free pages
1272  *                 supplied as the target for the page migration
1273  *
1274  * @from:               The list of pages to be migrated.
1275  * @get_new_page:       The function used to allocate free pages to be used
1276  *                      as the target of the page migration.
1277  * @put_new_page:       The function used to free target pages if migration
1278  *                      fails, or NULL if no special handling is necessary.
1279  * @private:            Private data to be passed on to get_new_page()
1280  * @mode:               The migration mode that specifies the constraints for
1281  *                      page migration, if any.
1282  * @reason:             The reason for page migration.
1283  *
1284  * The function returns after 10 attempts or if no pages are movable any more
1285  * because the list has become empty or no retryable pages exist any more.
1286  * The caller should call putback_movable_pages() to return pages to the LRU
1287  * or free list only if ret != 0.
1288  *
1289  * Returns the number of pages that were not migrated, or an error code.
1290  */
1291 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1292                 free_page_t put_new_page, unsigned long private,
1293                 enum migrate_mode mode, int reason)
1294 {
1295         int retry = 1;
1296         int nr_failed = 0;
1297         int nr_succeeded = 0;
1298         int pass = 0;
1299         struct page *page;
1300         struct page *page2;
1301         int swapwrite = current->flags & PF_SWAPWRITE;
1302         int rc;
1303
1304         if (!swapwrite)
1305                 current->flags |= PF_SWAPWRITE;
1306
1307         for(pass = 0; pass < 10 && retry; pass++) {
1308                 retry = 0;
1309
1310                 list_for_each_entry_safe(page, page2, from, lru) {
1311                         cond_resched();
1312
1313                         if (PageHuge(page))
1314                                 rc = unmap_and_move_huge_page(get_new_page,
1315                                                 put_new_page, private, page,
1316                                                 pass > 2, mode, reason);
1317                         else
1318                                 rc = unmap_and_move(get_new_page, put_new_page,
1319                                                 private, page, pass > 2, mode,
1320                                                 reason);
1321
1322                         switch(rc) {
1323                         case -ENOMEM:
1324                                 nr_failed++;
1325                                 goto out;
1326                         case -EAGAIN:
1327                                 retry++;
1328                                 break;
1329                         case MIGRATEPAGE_SUCCESS:
1330                                 nr_succeeded++;
1331                                 break;
1332                         default:
1333                                 /*
1334                                  * Permanent failure (-EBUSY, -ENOSYS, etc.):
1335                                  * unlike -EAGAIN case, the failed page is
1336                                  * removed from migration page list and not
1337                                  * retried in the next outer loop.
1338                                  */
1339                                 nr_failed++;
1340                                 break;
1341                         }
1342                 }
1343         }
1344         nr_failed += retry;
1345         rc = nr_failed;
1346 out:
1347         if (nr_succeeded)
1348                 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1349         if (nr_failed)
1350                 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1351         trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1352
1353         if (!swapwrite)
1354                 current->flags &= ~PF_SWAPWRITE;
1355
1356         return rc;
1357 }
1358
1359 #ifdef CONFIG_NUMA
1360 /*
1361  * Move a list of individual pages
1362  */
1363 struct page_to_node {
1364         unsigned long addr;
1365         struct page *page;
1366         int node;
1367         int status;
1368 };
1369
1370 static struct page *new_page_node(struct page *p, unsigned long private,
1371                 int **result)
1372 {
1373         struct page_to_node *pm = (struct page_to_node *)private;
1374
1375         while (pm->node != MAX_NUMNODES && pm->page != p)
1376                 pm++;
1377
1378         if (pm->node == MAX_NUMNODES)
1379                 return NULL;
1380
1381         *result = &pm->status;
1382
1383         if (PageHuge(p))
1384                 return alloc_huge_page_node(page_hstate(compound_head(p)),
1385                                         pm->node);
1386         else
1387                 return __alloc_pages_node(pm->node,
1388                                 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1389 }
1390
1391 /*
1392  * Move a set of pages as indicated in the pm array. The addr
1393  * field must be set to the virtual address of the page to be moved
1394  * and the node number must contain a valid target node.
1395  * The pm array ends with node = MAX_NUMNODES.
1396  */
1397 static int do_move_page_to_node_array(struct mm_struct *mm,
1398                                       struct page_to_node *pm,
1399                                       int migrate_all)
1400 {
1401         int err;
1402         struct page_to_node *pp;
1403         LIST_HEAD(pagelist);
1404
1405         down_read(&mm->mmap_sem);
1406
1407         /*
1408          * Build a list of pages to migrate
1409          */
1410         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1411                 struct vm_area_struct *vma;
1412                 struct page *page;
1413
1414                 err = -EFAULT;
1415                 vma = find_vma(mm, pp->addr);
1416                 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1417                         goto set_status;
1418
1419                 /* FOLL_DUMP to ignore special (like zero) pages */
1420                 page = follow_page(vma, pp->addr,
1421                                 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1422
1423                 err = PTR_ERR(page);
1424                 if (IS_ERR(page))
1425                         goto set_status;
1426
1427                 err = -ENOENT;
1428                 if (!page)
1429                         goto set_status;
1430
1431                 pp->page = page;
1432                 err = page_to_nid(page);
1433
1434                 if (err == pp->node)
1435                         /*
1436                          * Node already in the right place
1437                          */
1438                         goto put_and_set;
1439
1440                 err = -EACCES;
1441                 if (page_mapcount(page) > 1 &&
1442                                 !migrate_all)
1443                         goto put_and_set;
1444
1445                 if (PageHuge(page)) {
1446                         if (PageHead(page))
1447                                 isolate_huge_page(page, &pagelist);
1448                         goto put_and_set;
1449                 }
1450
1451                 err = isolate_lru_page(page);
1452                 if (!err) {
1453                         list_add_tail(&page->lru, &pagelist);
1454                         inc_node_page_state(page, NR_ISOLATED_ANON +
1455                                             page_is_file_cache(page));
1456                 }
1457 put_and_set:
1458                 /*
1459                  * Either remove the duplicate refcount from
1460                  * isolate_lru_page() or drop the page ref if it was
1461                  * not isolated.
1462                  */
1463                 put_page(page);
1464 set_status:
1465                 pp->status = err;
1466         }
1467
1468         err = 0;
1469         if (!list_empty(&pagelist)) {
1470                 err = migrate_pages(&pagelist, new_page_node, NULL,
1471                                 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1472                 if (err)
1473                         putback_movable_pages(&pagelist);
1474         }
1475
1476         up_read(&mm->mmap_sem);
1477         return err;
1478 }
1479
1480 /*
1481  * Migrate an array of page address onto an array of nodes and fill
1482  * the corresponding array of status.
1483  */
1484 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1485                          unsigned long nr_pages,
1486                          const void __user * __user *pages,
1487                          const int __user *nodes,
1488                          int __user *status, int flags)
1489 {
1490         struct page_to_node *pm;
1491         unsigned long chunk_nr_pages;
1492         unsigned long chunk_start;
1493         int err;
1494
1495         err = -ENOMEM;
1496         pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1497         if (!pm)
1498                 goto out;
1499
1500         migrate_prep();
1501
1502         /*
1503          * Store a chunk of page_to_node array in a page,
1504          * but keep the last one as a marker
1505          */
1506         chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1507
1508         for (chunk_start = 0;
1509              chunk_start < nr_pages;
1510              chunk_start += chunk_nr_pages) {
1511                 int j;
1512
1513                 if (chunk_start + chunk_nr_pages > nr_pages)
1514                         chunk_nr_pages = nr_pages - chunk_start;
1515
1516                 /* fill the chunk pm with addrs and nodes from user-space */
1517                 for (j = 0; j < chunk_nr_pages; j++) {
1518                         const void __user *p;
1519                         int node;
1520
1521                         err = -EFAULT;
1522                         if (get_user(p, pages + j + chunk_start))
1523                                 goto out_pm;
1524                         pm[j].addr = (unsigned long) p;
1525
1526                         if (get_user(node, nodes + j + chunk_start))
1527                                 goto out_pm;
1528
1529                         err = -ENODEV;
1530                         if (node < 0 || node >= MAX_NUMNODES)
1531                                 goto out_pm;
1532
1533                         if (!node_state(node, N_MEMORY))
1534                                 goto out_pm;
1535
1536                         err = -EACCES;
1537                         if (!node_isset(node, task_nodes))
1538                                 goto out_pm;
1539
1540                         pm[j].node = node;
1541                 }
1542
1543                 /* End marker for this chunk */
1544                 pm[chunk_nr_pages].node = MAX_NUMNODES;
1545
1546                 /* Migrate this chunk */
1547                 err = do_move_page_to_node_array(mm, pm,
1548                                                  flags & MPOL_MF_MOVE_ALL);
1549                 if (err < 0)
1550                         goto out_pm;
1551
1552                 /* Return status information */
1553                 for (j = 0; j < chunk_nr_pages; j++)
1554                         if (put_user(pm[j].status, status + j + chunk_start)) {
1555                                 err = -EFAULT;
1556                                 goto out_pm;
1557                         }
1558         }
1559         err = 0;
1560
1561 out_pm:
1562         free_page((unsigned long)pm);
1563 out:
1564         return err;
1565 }
1566
1567 /*
1568  * Determine the nodes of an array of pages and store it in an array of status.
1569  */
1570 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1571                                 const void __user **pages, int *status)
1572 {
1573         unsigned long i;
1574
1575         down_read(&mm->mmap_sem);
1576
1577         for (i = 0; i < nr_pages; i++) {
1578                 unsigned long addr = (unsigned long)(*pages);
1579                 struct vm_area_struct *vma;
1580                 struct page *page;
1581                 int err = -EFAULT;
1582
1583                 vma = find_vma(mm, addr);
1584                 if (!vma || addr < vma->vm_start)
1585                         goto set_status;
1586
1587                 /* FOLL_DUMP to ignore special (like zero) pages */
1588                 page = follow_page(vma, addr, FOLL_DUMP);
1589
1590                 err = PTR_ERR(page);
1591                 if (IS_ERR(page))
1592                         goto set_status;
1593
1594                 err = page ? page_to_nid(page) : -ENOENT;
1595 set_status:
1596                 *status = err;
1597
1598                 pages++;
1599                 status++;
1600         }
1601
1602         up_read(&mm->mmap_sem);
1603 }
1604
1605 /*
1606  * Determine the nodes of a user array of pages and store it in
1607  * a user array of status.
1608  */
1609 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1610                          const void __user * __user *pages,
1611                          int __user *status)
1612 {
1613 #define DO_PAGES_STAT_CHUNK_NR 16
1614         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1615         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1616
1617         while (nr_pages) {
1618                 unsigned long chunk_nr;
1619
1620                 chunk_nr = nr_pages;
1621                 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1622                         chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1623
1624                 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1625                         break;
1626
1627                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1628
1629                 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1630                         break;
1631
1632                 pages += chunk_nr;
1633                 status += chunk_nr;
1634                 nr_pages -= chunk_nr;
1635         }
1636         return nr_pages ? -EFAULT : 0;
1637 }
1638
1639 /*
1640  * Move a list of pages in the address space of the currently executing
1641  * process.
1642  */
1643 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1644                 const void __user * __user *, pages,
1645                 const int __user *, nodes,
1646                 int __user *, status, int, flags)
1647 {
1648         const struct cred *cred = current_cred(), *tcred;
1649         struct task_struct *task;
1650         struct mm_struct *mm;
1651         int err;
1652         nodemask_t task_nodes;
1653
1654         /* Check flags */
1655         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1656                 return -EINVAL;
1657
1658         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1659                 return -EPERM;
1660
1661         /* Find the mm_struct */
1662         rcu_read_lock();
1663         task = pid ? find_task_by_vpid(pid) : current;
1664         if (!task) {
1665                 rcu_read_unlock();
1666                 return -ESRCH;
1667         }
1668         get_task_struct(task);
1669
1670         /*
1671          * Check if this process has the right to modify the specified
1672          * process. The right exists if the process has administrative
1673          * capabilities, superuser privileges or the same
1674          * userid as the target process.
1675          */
1676         tcred = __task_cred(task);
1677         if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1678             !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1679             !capable(CAP_SYS_NICE)) {
1680                 rcu_read_unlock();
1681                 err = -EPERM;
1682                 goto out;
1683         }
1684         rcu_read_unlock();
1685
1686         err = security_task_movememory(task);
1687         if (err)
1688                 goto out;
1689
1690         task_nodes = cpuset_mems_allowed(task);
1691         mm = get_task_mm(task);
1692         put_task_struct(task);
1693
1694         if (!mm)
1695                 return -EINVAL;
1696
1697         if (nodes)
1698                 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1699                                     nodes, status, flags);
1700         else
1701                 err = do_pages_stat(mm, nr_pages, pages, status);
1702
1703         mmput(mm);
1704         return err;
1705
1706 out:
1707         put_task_struct(task);
1708         return err;
1709 }
1710
1711 #ifdef CONFIG_NUMA_BALANCING
1712 /*
1713  * Returns true if this is a safe migration target node for misplaced NUMA
1714  * pages. Currently it only checks the watermarks which crude
1715  */
1716 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1717                                    unsigned long nr_migrate_pages)
1718 {
1719         int z;
1720
1721         if (!pgdat_reclaimable(pgdat))
1722                 return false;
1723
1724         for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1725                 struct zone *zone = pgdat->node_zones + z;
1726
1727                 if (!populated_zone(zone))
1728                         continue;
1729
1730                 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1731                 if (!zone_watermark_ok(zone, 0,
1732                                        high_wmark_pages(zone) +
1733                                        nr_migrate_pages,
1734                                        0, 0))
1735                         continue;
1736                 return true;
1737         }
1738         return false;
1739 }
1740
1741 static struct page *alloc_misplaced_dst_page(struct page *page,
1742                                            unsigned long data,
1743                                            int **result)
1744 {
1745         int nid = (int) data;
1746         struct page *newpage;
1747
1748         newpage = __alloc_pages_node(nid,
1749                                          (GFP_HIGHUSER_MOVABLE |
1750                                           __GFP_THISNODE | __GFP_NOMEMALLOC |
1751                                           __GFP_NORETRY | __GFP_NOWARN) &
1752                                          ~__GFP_RECLAIM, 0);
1753
1754         return newpage;
1755 }
1756
1757 /*
1758  * page migration rate limiting control.
1759  * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1760  * window of time. Default here says do not migrate more than 1280M per second.
1761  */
1762 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1763 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1764
1765 /* Returns true if the node is migrate rate-limited after the update */
1766 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1767                                         unsigned long nr_pages)
1768 {
1769         /*
1770          * Rate-limit the amount of data that is being migrated to a node.
1771          * Optimal placement is no good if the memory bus is saturated and
1772          * all the time is being spent migrating!
1773          */
1774         if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1775                 spin_lock(&pgdat->numabalancing_migrate_lock);
1776                 pgdat->numabalancing_migrate_nr_pages = 0;
1777                 pgdat->numabalancing_migrate_next_window = jiffies +
1778                         msecs_to_jiffies(migrate_interval_millisecs);
1779                 spin_unlock(&pgdat->numabalancing_migrate_lock);
1780         }
1781         if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1782                 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1783                                                                 nr_pages);
1784                 return true;
1785         }
1786
1787         /*
1788          * This is an unlocked non-atomic update so errors are possible.
1789          * The consequences are failing to migrate when we potentiall should
1790          * have which is not severe enough to warrant locking. If it is ever
1791          * a problem, it can be converted to a per-cpu counter.
1792          */
1793         pgdat->numabalancing_migrate_nr_pages += nr_pages;
1794         return false;
1795 }
1796
1797 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1798 {
1799         int page_lru;
1800
1801         VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1802
1803         /* Avoid migrating to a node that is nearly full */
1804         if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1805                 return 0;
1806
1807         if (isolate_lru_page(page))
1808                 return 0;
1809
1810         /*
1811          * migrate_misplaced_transhuge_page() skips page migration's usual
1812          * check on page_count(), so we must do it here, now that the page
1813          * has been isolated: a GUP pin, or any other pin, prevents migration.
1814          * The expected page count is 3: 1 for page's mapcount and 1 for the
1815          * caller's pin and 1 for the reference taken by isolate_lru_page().
1816          */
1817         if (PageTransHuge(page) && page_count(page) != 3) {
1818                 putback_lru_page(page);
1819                 return 0;
1820         }
1821
1822         page_lru = page_is_file_cache(page);
1823         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1824                                 hpage_nr_pages(page));
1825
1826         /*
1827          * Isolating the page has taken another reference, so the
1828          * caller's reference can be safely dropped without the page
1829          * disappearing underneath us during migration.
1830          */
1831         put_page(page);
1832         return 1;
1833 }
1834
1835 bool pmd_trans_migrating(pmd_t pmd)
1836 {
1837         struct page *page = pmd_page(pmd);
1838         return PageLocked(page);
1839 }
1840
1841 /*
1842  * Attempt to migrate a misplaced page to the specified destination
1843  * node. Caller is expected to have an elevated reference count on
1844  * the page that will be dropped by this function before returning.
1845  */
1846 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1847                            int node)
1848 {
1849         pg_data_t *pgdat = NODE_DATA(node);
1850         int isolated;
1851         int nr_remaining;
1852         LIST_HEAD(migratepages);
1853
1854         /*
1855          * Don't migrate file pages that are mapped in multiple processes
1856          * with execute permissions as they are probably shared libraries.
1857          */
1858         if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1859             (vma->vm_flags & VM_EXEC))
1860                 goto out;
1861
1862         /*
1863          * Rate-limit the amount of data that is being migrated to a node.
1864          * Optimal placement is no good if the memory bus is saturated and
1865          * all the time is being spent migrating!
1866          */
1867         if (numamigrate_update_ratelimit(pgdat, 1))
1868                 goto out;
1869
1870         isolated = numamigrate_isolate_page(pgdat, page);
1871         if (!isolated)
1872                 goto out;
1873
1874         list_add(&page->lru, &migratepages);
1875         nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1876                                      NULL, node, MIGRATE_ASYNC,
1877                                      MR_NUMA_MISPLACED);
1878         if (nr_remaining) {
1879                 if (!list_empty(&migratepages)) {
1880                         list_del(&page->lru);
1881                         dec_node_page_state(page, NR_ISOLATED_ANON +
1882                                         page_is_file_cache(page));
1883                         putback_lru_page(page);
1884                 }
1885                 isolated = 0;
1886         } else
1887                 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1888         BUG_ON(!list_empty(&migratepages));
1889         return isolated;
1890
1891 out:
1892         put_page(page);
1893         return 0;
1894 }
1895 #endif /* CONFIG_NUMA_BALANCING */
1896
1897 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1898 /*
1899  * Migrates a THP to a given target node. page must be locked and is unlocked
1900  * before returning.
1901  */
1902 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1903                                 struct vm_area_struct *vma,
1904                                 pmd_t *pmd, pmd_t entry,
1905                                 unsigned long address,
1906                                 struct page *page, int node)
1907 {
1908         spinlock_t *ptl;
1909         pg_data_t *pgdat = NODE_DATA(node);
1910         int isolated = 0;
1911         struct page *new_page = NULL;
1912         int page_lru = page_is_file_cache(page);
1913         unsigned long mmun_start = address & HPAGE_PMD_MASK;
1914         unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1915         pmd_t orig_entry;
1916
1917         /*
1918          * Rate-limit the amount of data that is being migrated to a node.
1919          * Optimal placement is no good if the memory bus is saturated and
1920          * all the time is being spent migrating!
1921          */
1922         if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1923                 goto out_dropref;
1924
1925         new_page = alloc_pages_node(node,
1926                 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1927                 HPAGE_PMD_ORDER);
1928         if (!new_page)
1929                 goto out_fail;
1930         prep_transhuge_page(new_page);
1931
1932         isolated = numamigrate_isolate_page(pgdat, page);
1933         if (!isolated) {
1934                 put_page(new_page);
1935                 goto out_fail;
1936         }
1937         /*
1938          * We are not sure a pending tlb flush here is for a huge page
1939          * mapping or not. Hence use the tlb range variant
1940          */
1941         if (mm_tlb_flush_pending(mm))
1942                 flush_tlb_range(vma, mmun_start, mmun_end);
1943
1944         /* Prepare a page as a migration target */
1945         __SetPageLocked(new_page);
1946         __SetPageSwapBacked(new_page);
1947
1948         /* anon mapping, we can simply copy page->mapping to the new page: */
1949         new_page->mapping = page->mapping;
1950         new_page->index = page->index;
1951         migrate_page_copy(new_page, page);
1952         WARN_ON(PageLRU(new_page));
1953
1954         /* Recheck the target PMD */
1955         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1956         ptl = pmd_lock(mm, pmd);
1957         if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1958 fail_putback:
1959                 spin_unlock(ptl);
1960                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1961
1962                 /* Reverse changes made by migrate_page_copy() */
1963                 if (TestClearPageActive(new_page))
1964                         SetPageActive(page);
1965                 if (TestClearPageUnevictable(new_page))
1966                         SetPageUnevictable(page);
1967
1968                 unlock_page(new_page);
1969                 put_page(new_page);             /* Free it */
1970
1971                 /* Retake the callers reference and putback on LRU */
1972                 get_page(page);
1973                 putback_lru_page(page);
1974                 mod_node_page_state(page_pgdat(page),
1975                          NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1976
1977                 goto out_unlock;
1978         }
1979
1980         orig_entry = *pmd;
1981         entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1982         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1983
1984         /*
1985          * Clear the old entry under pagetable lock and establish the new PTE.
1986          * Any parallel GUP will either observe the old page blocking on the
1987          * page lock, block on the page table lock or observe the new page.
1988          * The SetPageUptodate on the new page and page_add_new_anon_rmap
1989          * guarantee the copy is visible before the pagetable update.
1990          */
1991         flush_cache_range(vma, mmun_start, mmun_end);
1992         page_add_anon_rmap(new_page, vma, mmun_start, true);
1993         pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1994         set_pmd_at(mm, mmun_start, pmd, entry);
1995         update_mmu_cache_pmd(vma, address, &entry);
1996
1997         if (page_count(page) != 2) {
1998                 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1999                 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
2000                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2001                 update_mmu_cache_pmd(vma, address, &entry);
2002                 page_remove_rmap(new_page, true);
2003                 goto fail_putback;
2004         }
2005
2006         mlock_migrate_page(new_page, page);
2007         page_remove_rmap(page, true);
2008         set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2009
2010         spin_unlock(ptl);
2011         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2012
2013         /* Take an "isolate" reference and put new page on the LRU. */
2014         get_page(new_page);
2015         putback_lru_page(new_page);
2016
2017         unlock_page(new_page);
2018         unlock_page(page);
2019         put_page(page);                 /* Drop the rmap reference */
2020         put_page(page);                 /* Drop the LRU isolation reference */
2021
2022         count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2023         count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2024
2025         mod_node_page_state(page_pgdat(page),
2026                         NR_ISOLATED_ANON + page_lru,
2027                         -HPAGE_PMD_NR);
2028         return isolated;
2029
2030 out_fail:
2031         count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2032 out_dropref:
2033         ptl = pmd_lock(mm, pmd);
2034         if (pmd_same(*pmd, entry)) {
2035                 entry = pmd_modify(entry, vma->vm_page_prot);
2036                 set_pmd_at(mm, mmun_start, pmd, entry);
2037                 update_mmu_cache_pmd(vma, address, &entry);
2038         }
2039         spin_unlock(ptl);
2040
2041 out_unlock:
2042         unlock_page(page);
2043         put_page(page);
2044         return 0;
2045 }
2046 #endif /* CONFIG_NUMA_BALANCING */
2047
2048 #endif /* CONFIG_NUMA */