linux/mm/migrate.c
<<
>>
Prefs
   1/*
   2 * Memory Migration functionality - linux/mm/migration.c
   3 *
   4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   5 *
   6 * Page migration was first developed in the context of the memory hotplug
   7 * project. The main authors of the migration code are:
   8 *
   9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  10 * Hirokazu Takahashi <taka@valinux.co.jp>
  11 * Dave Hansen <haveblue@us.ibm.com>
  12 * Christoph Lameter
  13 */
  14
  15#include <linux/migrate.h>
  16#include <linux/export.h>
  17#include <linux/swap.h>
  18#include <linux/swapops.h>
  19#include <linux/pagemap.h>
  20#include <linux/buffer_head.h>
  21#include <linux/mm_inline.h>
  22#include <linux/nsproxy.h>
  23#include <linux/pagevec.h>
  24#include <linux/ksm.h>
  25#include <linux/rmap.h>
  26#include <linux/topology.h>
  27#include <linux/cpu.h>
  28#include <linux/cpuset.h>
  29#include <linux/writeback.h>
  30#include <linux/mempolicy.h>
  31#include <linux/vmalloc.h>
  32#include <linux/security.h>
  33#include <linux/memcontrol.h>
  34#include <linux/syscalls.h>
  35#include <linux/hugetlb.h>
  36#include <linux/hugetlb_cgroup.h>
  37#include <linux/gfp.h>
  38#include <linux/balloon_compaction.h>
  39
  40#include <asm/tlbflush.h>
  41
  42#define CREATE_TRACE_POINTS
  43#include <trace/events/migrate.h>
  44
  45#include "internal.h"
  46
  47/*
  48 * migrate_prep() needs to be called before we start compiling a list of pages
  49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
  50 * undesirable, use migrate_prep_local()
  51 */
  52int migrate_prep(void)
  53{
  54        /*
  55         * Clear the LRU lists so pages can be isolated.
  56         * Note that pages may be moved off the LRU after we have
  57         * drained them. Those pages will fail to migrate like other
  58         * pages that may be busy.
  59         */
  60        lru_add_drain_all();
  61
  62        return 0;
  63}
  64
  65/* Do the necessary work of migrate_prep but not if it involves other CPUs */
  66int migrate_prep_local(void)
  67{
  68        lru_add_drain();
  69
  70        return 0;
  71}
  72
  73/*
  74 * Add isolated pages on the list back to the LRU under page lock
  75 * to avoid leaking evictable pages back onto unevictable list.
  76 */
  77void putback_lru_pages(struct list_head *l)
  78{
  79        struct page *page;
  80        struct page *page2;
  81
  82        list_for_each_entry_safe(page, page2, l, lru) {
  83                list_del(&page->lru);
  84                dec_zone_page_state(page, NR_ISOLATED_ANON +
  85                                page_is_file_cache(page));
  86                        putback_lru_page(page);
  87        }
  88}
  89
  90/*
  91 * Put previously isolated pages back onto the appropriate lists
  92 * from where they were once taken off for compaction/migration.
  93 *
  94 * This function shall be used instead of putback_lru_pages(),
  95 * whenever the isolated pageset has been built by isolate_migratepages_range()
  96 */
  97void putback_movable_pages(struct list_head *l)
  98{
  99        struct page *page;
 100        struct page *page2;
 101
 102        list_for_each_entry_safe(page, page2, l, lru) {
 103                list_del(&page->lru);
 104                dec_zone_page_state(page, NR_ISOLATED_ANON +
 105                                page_is_file_cache(page));
 106                if (unlikely(balloon_page_movable(page)))
 107                        balloon_page_putback(page);
 108                else
 109                        putback_lru_page(page);
 110        }
 111}
 112
 113/*
 114 * Restore a potential migration pte to a working pte entry
 115 */
 116static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
 117                                 unsigned long addr, void *old)
 118{
 119        struct mm_struct *mm = vma->vm_mm;
 120        swp_entry_t entry;
 121        pmd_t *pmd;
 122        pte_t *ptep, pte;
 123        spinlock_t *ptl;
 124
 125        if (unlikely(PageHuge(new))) {
 126                ptep = huge_pte_offset(mm, addr);
 127                if (!ptep)
 128                        goto out;
 129                ptl = &mm->page_table_lock;
 130        } else {
 131                pmd = mm_find_pmd(mm, addr);
 132                if (!pmd)
 133                        goto out;
 134                if (pmd_trans_huge(*pmd))
 135                        goto out;
 136
 137                ptep = pte_offset_map(pmd, addr);
 138
 139                /*
 140                 * Peek to check is_swap_pte() before taking ptlock?  No, we
 141                 * can race mremap's move_ptes(), which skips anon_vma lock.
 142                 */
 143
 144                ptl = pte_lockptr(mm, pmd);
 145        }
 146
 147        spin_lock(ptl);
 148        pte = *ptep;
 149        if (!is_swap_pte(pte))
 150                goto unlock;
 151
 152        entry = pte_to_swp_entry(pte);
 153
 154        if (!is_migration_entry(entry) ||
 155            migration_entry_to_page(entry) != old)
 156                goto unlock;
 157
 158        get_page(new);
 159        pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 160        if (is_write_migration_entry(entry))
 161                pte = pte_mkwrite(pte);
 162#ifdef CONFIG_HUGETLB_PAGE
 163        if (PageHuge(new)) {
 164                pte = pte_mkhuge(pte);
 165                pte = arch_make_huge_pte(pte, vma, new, 0);
 166        }
 167#endif
 168        flush_cache_page(vma, addr, pte_pfn(pte));
 169        set_pte_at(mm, addr, ptep, pte);
 170
 171        if (PageHuge(new)) {
 172                if (PageAnon(new))
 173                        hugepage_add_anon_rmap(new, vma, addr);
 174                else
 175                        page_dup_rmap(new);
 176        } else if (PageAnon(new))
 177                page_add_anon_rmap(new, vma, addr);
 178        else
 179                page_add_file_rmap(new);
 180
 181        /* No need to invalidate - it was non-present before */
 182        update_mmu_cache(vma, addr, ptep);
 183unlock:
 184        pte_unmap_unlock(ptep, ptl);
 185out:
 186        return SWAP_AGAIN;
 187}
 188
 189/*
 190 * Get rid of all migration entries and replace them by
 191 * references to the indicated page.
 192 */
 193static void remove_migration_ptes(struct page *old, struct page *new)
 194{
 195        rmap_walk(new, remove_migration_pte, old);
 196}
 197
 198/*
 199 * Something used the pte of a page under migration. We need to
 200 * get to the page and wait until migration is finished.
 201 * When we return from this function the fault will be retried.
 202 */
 203void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 204                                unsigned long address)
 205{
 206        pte_t *ptep, pte;
 207        spinlock_t *ptl;
 208        swp_entry_t entry;
 209        struct page *page;
 210
 211        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 212        pte = *ptep;
 213        if (!is_swap_pte(pte))
 214                goto out;
 215
 216        entry = pte_to_swp_entry(pte);
 217        if (!is_migration_entry(entry))
 218                goto out;
 219
 220        page = migration_entry_to_page(entry);
 221
 222        /*
 223         * Once radix-tree replacement of page migration started, page_count
 224         * *must* be zero. And, we don't want to call wait_on_page_locked()
 225         * against a page without get_page().
 226         * So, we use get_page_unless_zero(), here. Even failed, page fault
 227         * will occur again.
 228         */
 229        if (!get_page_unless_zero(page))
 230                goto out;
 231        pte_unmap_unlock(ptep, ptl);
 232        wait_on_page_locked(page);
 233        put_page(page);
 234        return;
 235out:
 236        pte_unmap_unlock(ptep, ptl);
 237}
 238
 239#ifdef CONFIG_BLOCK
 240/* Returns true if all buffers are successfully locked */
 241static bool buffer_migrate_lock_buffers(struct buffer_head *head,
 242                                                        enum migrate_mode mode)
 243{
 244        struct buffer_head *bh = head;
 245
 246        /* Simple case, sync compaction */
 247        if (mode != MIGRATE_ASYNC) {
 248                do {
 249                        get_bh(bh);
 250                        lock_buffer(bh);
 251                        bh = bh->b_this_page;
 252
 253                } while (bh != head);
 254
 255                return true;
 256        }
 257
 258        /* async case, we cannot block on lock_buffer so use trylock_buffer */
 259        do {
 260                get_bh(bh);
 261                if (!trylock_buffer(bh)) {
 262                        /*
 263                         * We failed to lock the buffer and cannot stall in
 264                         * async migration. Release the taken locks
 265                         */
 266                        struct buffer_head *failed_bh = bh;
 267                        put_bh(failed_bh);
 268                        bh = head;
 269                        while (bh != failed_bh) {
 270                                unlock_buffer(bh);
 271                                put_bh(bh);
 272                                bh = bh->b_this_page;
 273                        }
 274                        return false;
 275                }
 276
 277                bh = bh->b_this_page;
 278        } while (bh != head);
 279        return true;
 280}
 281#else
 282static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
 283                                                        enum migrate_mode mode)
 284{
 285        return true;
 286}
 287#endif /* CONFIG_BLOCK */
 288
 289/*
 290 * Replace the page in the mapping.
 291 *
 292 * The number of remaining references must be:
 293 * 1 for anonymous pages without a mapping
 294 * 2 for pages with a mapping
 295 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 296 */
 297static int migrate_page_move_mapping(struct address_space *mapping,
 298                struct page *newpage, struct page *page,
 299                struct buffer_head *head, enum migrate_mode mode)
 300{
 301        int expected_count = 0;
 302        void **pslot;
 303
 304        if (!mapping) {
 305                /* Anonymous page without mapping */
 306                if (page_count(page) != 1)
 307                        return -EAGAIN;
 308                return MIGRATEPAGE_SUCCESS;
 309        }
 310
 311        spin_lock_irq(&mapping->tree_lock);
 312
 313        pslot = radix_tree_lookup_slot(&mapping->page_tree,
 314                                        page_index(page));
 315
 316        expected_count = 2 + page_has_private(page);
 317        if (page_count(page) != expected_count ||
 318                radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
 319                spin_unlock_irq(&mapping->tree_lock);
 320                return -EAGAIN;
 321        }
 322
 323        if (!page_freeze_refs(page, expected_count)) {
 324                spin_unlock_irq(&mapping->tree_lock);
 325                return -EAGAIN;
 326        }
 327
 328        /*
 329         * In the async migration case of moving a page with buffers, lock the
 330         * buffers using trylock before the mapping is moved. If the mapping
 331         * was moved, we later failed to lock the buffers and could not move
 332         * the mapping back due to an elevated page count, we would have to
 333         * block waiting on other references to be dropped.
 334         */
 335        if (mode == MIGRATE_ASYNC && head &&
 336                        !buffer_migrate_lock_buffers(head, mode)) {
 337                page_unfreeze_refs(page, expected_count);
 338                spin_unlock_irq(&mapping->tree_lock);
 339                return -EAGAIN;
 340        }
 341
 342        /*
 343         * Now we know that no one else is looking at the page.
 344         */
 345        get_page(newpage);      /* add cache reference */
 346        if (PageSwapCache(page)) {
 347                SetPageSwapCache(newpage);
 348                set_page_private(newpage, page_private(page));
 349        }
 350
 351        radix_tree_replace_slot(pslot, newpage);
 352
 353        /*
 354         * Drop cache reference from old page by unfreezing
 355         * to one less reference.
 356         * We know this isn't the last reference.
 357         */
 358        page_unfreeze_refs(page, expected_count - 1);
 359
 360        /*
 361         * If moved to a different zone then also account
 362         * the page for that zone. Other VM counters will be
 363         * taken care of when we establish references to the
 364         * new page and drop references to the old page.
 365         *
 366         * Note that anonymous pages are accounted for
 367         * via NR_FILE_PAGES and NR_ANON_PAGES if they
 368         * are mapped to swap space.
 369         */
 370        __dec_zone_page_state(page, NR_FILE_PAGES);
 371        __inc_zone_page_state(newpage, NR_FILE_PAGES);
 372        if (!PageSwapCache(page) && PageSwapBacked(page)) {
 373                __dec_zone_page_state(page, NR_SHMEM);
 374                __inc_zone_page_state(newpage, NR_SHMEM);
 375        }
 376        spin_unlock_irq(&mapping->tree_lock);
 377
 378        return MIGRATEPAGE_SUCCESS;
 379}
 380
 381/*
 382 * The expected number of remaining references is the same as that
 383 * of migrate_page_move_mapping().
 384 */
 385int migrate_huge_page_move_mapping(struct address_space *mapping,
 386                                   struct page *newpage, struct page *page)
 387{
 388        int expected_count;
 389        void **pslot;
 390
 391        if (!mapping) {
 392                if (page_count(page) != 1)
 393                        return -EAGAIN;
 394                return MIGRATEPAGE_SUCCESS;
 395        }
 396
 397        spin_lock_irq(&mapping->tree_lock);
 398
 399        pslot = radix_tree_lookup_slot(&mapping->page_tree,
 400                                        page_index(page));
 401
 402        expected_count = 2 + page_has_private(page);
 403        if (page_count(page) != expected_count ||
 404                radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
 405                spin_unlock_irq(&mapping->tree_lock);
 406                return -EAGAIN;
 407        }
 408
 409        if (!page_freeze_refs(page, expected_count)) {
 410                spin_unlock_irq(&mapping->tree_lock);
 411                return -EAGAIN;
 412        }
 413
 414        get_page(newpage);
 415
 416        radix_tree_replace_slot(pslot, newpage);
 417
 418        page_unfreeze_refs(page, expected_count - 1);
 419
 420        spin_unlock_irq(&mapping->tree_lock);
 421        return MIGRATEPAGE_SUCCESS;
 422}
 423
 424/*
 425 * Copy the page to its new location
 426 */
 427void migrate_page_copy(struct page *newpage, struct page *page)
 428{
 429        if (PageHuge(page) || PageTransHuge(page))
 430                copy_huge_page(newpage, page);
 431        else
 432                copy_highpage(newpage, page);
 433
 434        if (PageError(page))
 435                SetPageError(newpage);
 436        if (PageReferenced(page))
 437                SetPageReferenced(newpage);
 438        if (PageUptodate(page))
 439                SetPageUptodate(newpage);
 440        if (TestClearPageActive(page)) {
 441                VM_BUG_ON(PageUnevictable(page));
 442                SetPageActive(newpage);
 443        } else if (TestClearPageUnevictable(page))
 444                SetPageUnevictable(newpage);
 445        if (PageChecked(page))
 446                SetPageChecked(newpage);
 447        if (PageMappedToDisk(page))
 448                SetPageMappedToDisk(newpage);
 449
 450        if (PageDirty(page)) {
 451                clear_page_dirty_for_io(page);
 452                /*
 453                 * Want to mark the page and the radix tree as dirty, and
 454                 * redo the accounting that clear_page_dirty_for_io undid,
 455                 * but we can't use set_page_dirty because that function
 456                 * is actually a signal that all of the page has become dirty.
 457                 * Whereas only part of our page may be dirty.
 458                 */
 459                if (PageSwapBacked(page))
 460                        SetPageDirty(newpage);
 461                else
 462                        __set_page_dirty_nobuffers(newpage);
 463        }
 464
 465        mlock_migrate_page(newpage, page);
 466        ksm_migrate_page(newpage, page);
 467
 468        ClearPageSwapCache(page);
 469        ClearPagePrivate(page);
 470        set_page_private(page, 0);
 471
 472        /*
 473         * If any waiters have accumulated on the new page then
 474         * wake them up.
 475         */
 476        if (PageWriteback(newpage))
 477                end_page_writeback(newpage);
 478}
 479
 480/************************************************************
 481 *                    Migration functions
 482 ***********************************************************/
 483
 484/* Always fail migration. Used for mappings that are not movable */
 485int fail_migrate_page(struct address_space *mapping,
 486                        struct page *newpage, struct page *page)
 487{
 488        return -EIO;
 489}
 490EXPORT_SYMBOL(fail_migrate_page);
 491
 492/*
 493 * Common logic to directly migrate a single page suitable for
 494 * pages that do not use PagePrivate/PagePrivate2.
 495 *
 496 * Pages are locked upon entry and exit.
 497 */
 498int migrate_page(struct address_space *mapping,
 499                struct page *newpage, struct page *page,
 500                enum migrate_mode mode)
 501{
 502        int rc;
 503
 504        BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 505
 506        rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
 507
 508        if (rc != MIGRATEPAGE_SUCCESS)
 509                return rc;
 510
 511        migrate_page_copy(newpage, page);
 512        return MIGRATEPAGE_SUCCESS;
 513}
 514EXPORT_SYMBOL(migrate_page);
 515
 516#ifdef CONFIG_BLOCK
 517/*
 518 * Migration function for pages with buffers. This function can only be used
 519 * if the underlying filesystem guarantees that no other references to "page"
 520 * exist.
 521 */
 522int buffer_migrate_page(struct address_space *mapping,
 523                struct page *newpage, struct page *page, enum migrate_mode mode)
 524{
 525        struct buffer_head *bh, *head;
 526        int rc;
 527
 528        if (!page_has_buffers(page))
 529                return migrate_page(mapping, newpage, page, mode);
 530
 531        head = page_buffers(page);
 532
 533        rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
 534
 535        if (rc != MIGRATEPAGE_SUCCESS)
 536                return rc;
 537
 538        /*
 539         * In the async case, migrate_page_move_mapping locked the buffers
 540         * with an IRQ-safe spinlock held. In the sync case, the buffers
 541         * need to be locked now
 542         */
 543        if (mode != MIGRATE_ASYNC)
 544                BUG_ON(!buffer_migrate_lock_buffers(head, mode));
 545
 546        ClearPagePrivate(page);
 547        set_page_private(newpage, page_private(page));
 548        set_page_private(page, 0);
 549        put_page(page);
 550        get_page(newpage);
 551
 552        bh = head;
 553        do {
 554                set_bh_page(bh, newpage, bh_offset(bh));
 555                bh = bh->b_this_page;
 556
 557        } while (bh != head);
 558
 559        SetPagePrivate(newpage);
 560
 561        migrate_page_copy(newpage, page);
 562
 563        bh = head;
 564        do {
 565                unlock_buffer(bh);
 566                put_bh(bh);
 567                bh = bh->b_this_page;
 568
 569        } while (bh != head);
 570
 571        return MIGRATEPAGE_SUCCESS;
 572}
 573EXPORT_SYMBOL(buffer_migrate_page);
 574#endif
 575
 576/*
 577 * Writeback a page to clean the dirty state
 578 */
 579static int writeout(struct address_space *mapping, struct page *page)
 580{
 581        struct writeback_control wbc = {
 582                .sync_mode = WB_SYNC_NONE,
 583                .nr_to_write = 1,
 584                .range_start = 0,
 585                .range_end = LLONG_MAX,
 586                .for_reclaim = 1
 587        };
 588        int rc;
 589
 590        if (!mapping->a_ops->writepage)
 591                /* No write method for the address space */
 592                return -EINVAL;
 593
 594        if (!clear_page_dirty_for_io(page))
 595                /* Someone else already triggered a write */
 596                return -EAGAIN;
 597
 598        /*
 599         * A dirty page may imply that the underlying filesystem has
 600         * the page on some queue. So the page must be clean for
 601         * migration. Writeout may mean we loose the lock and the
 602         * page state is no longer what we checked for earlier.
 603         * At this point we know that the migration attempt cannot
 604         * be successful.
 605         */
 606        remove_migration_ptes(page, page);
 607
 608        rc = mapping->a_ops->writepage(page, &wbc);
 609
 610        if (rc != AOP_WRITEPAGE_ACTIVATE)
 611                /* unlocked. Relock */
 612                lock_page(page);
 613
 614        return (rc < 0) ? -EIO : -EAGAIN;
 615}
 616
 617/*
 618 * Default handling if a filesystem does not provide a migration function.
 619 */
 620static int fallback_migrate_page(struct address_space *mapping,
 621        struct page *newpage, struct page *page, enum migrate_mode mode)
 622{
 623        if (PageDirty(page)) {
 624                /* Only writeback pages in full synchronous migration */
 625                if (mode != MIGRATE_SYNC)
 626                        return -EBUSY;
 627                return writeout(mapping, page);
 628        }
 629
 630        /*
 631         * Buffers may be managed in a filesystem specific way.
 632         * We must have no buffers or drop them.
 633         */
 634        if (page_has_private(page) &&
 635            !try_to_release_page(page, GFP_KERNEL))
 636                return -EAGAIN;
 637
 638        return migrate_page(mapping, newpage, page, mode);
 639}
 640
 641/*
 642 * Move a page to a newly allocated page
 643 * The page is locked and all ptes have been successfully removed.
 644 *
 645 * The new page will have replaced the old page if this function
 646 * is successful.
 647 *
 648 * Return value:
 649 *   < 0 - error code
 650 *  MIGRATEPAGE_SUCCESS - success
 651 */
 652static int move_to_new_page(struct page *newpage, struct page *page,
 653                                int remap_swapcache, enum migrate_mode mode)
 654{
 655        struct address_space *mapping;
 656        int rc;
 657
 658        /*
 659         * Block others from accessing the page when we get around to
 660         * establishing additional references. We are the only one
 661         * holding a reference to the new page at this point.
 662         */
 663        if (!trylock_page(newpage))
 664                BUG();
 665
 666        /* Prepare mapping for the new page.*/
 667        newpage->index = page->index;
 668        newpage->mapping = page->mapping;
 669        if (PageSwapBacked(page))
 670                SetPageSwapBacked(newpage);
 671
 672        mapping = page_mapping(page);
 673        if (!mapping)
 674                rc = migrate_page(mapping, newpage, page, mode);
 675        else if (mapping->a_ops->migratepage)
 676                /*
 677                 * Most pages have a mapping and most filesystems provide a
 678                 * migratepage callback. Anonymous pages are part of swap
 679                 * space which also has its own migratepage callback. This
 680                 * is the most common path for page migration.
 681                 */
 682                rc = mapping->a_ops->migratepage(mapping,
 683                                                newpage, page, mode);
 684        else
 685                rc = fallback_migrate_page(mapping, newpage, page, mode);
 686
 687        if (rc != MIGRATEPAGE_SUCCESS) {
 688                newpage->mapping = NULL;
 689        } else {
 690                if (remap_swapcache)
 691                        remove_migration_ptes(page, newpage);
 692                page->mapping = NULL;
 693        }
 694
 695        unlock_page(newpage);
 696
 697        return rc;
 698}
 699
 700static int __unmap_and_move(struct page *page, struct page *newpage,
 701                        int force, bool offlining, enum migrate_mode mode)
 702{
 703        int rc = -EAGAIN;
 704        int remap_swapcache = 1;
 705        struct mem_cgroup *mem;
 706        struct anon_vma *anon_vma = NULL;
 707
 708        if (!trylock_page(page)) {
 709                if (!force || mode == MIGRATE_ASYNC)
 710                        goto out;
 711
 712                /*
 713                 * It's not safe for direct compaction to call lock_page.
 714                 * For example, during page readahead pages are added locked
 715                 * to the LRU. Later, when the IO completes the pages are
 716                 * marked uptodate and unlocked. However, the queueing
 717                 * could be merging multiple pages for one bio (e.g.
 718                 * mpage_readpages). If an allocation happens for the
 719                 * second or third page, the process can end up locking
 720                 * the same page twice and deadlocking. Rather than
 721                 * trying to be clever about what pages can be locked,
 722                 * avoid the use of lock_page for direct compaction
 723                 * altogether.
 724                 */
 725                if (current->flags & PF_MEMALLOC)
 726                        goto out;
 727
 728                lock_page(page);
 729        }
 730
 731        /*
 732         * Only memory hotplug's offline_pages() caller has locked out KSM,
 733         * and can safely migrate a KSM page.  The other cases have skipped
 734         * PageKsm along with PageReserved - but it is only now when we have
 735         * the page lock that we can be certain it will not go KSM beneath us
 736         * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
 737         * its pagecount raised, but only here do we take the page lock which
 738         * serializes that).
 739         */
 740        if (PageKsm(page) && !offlining) {
 741                rc = -EBUSY;
 742                goto unlock;
 743        }
 744
 745        /* charge against new page */
 746        mem_cgroup_prepare_migration(page, newpage, &mem);
 747
 748        if (PageWriteback(page)) {
 749                /*
 750                 * Only in the case of a full syncronous migration is it
 751                 * necessary to wait for PageWriteback. In the async case,
 752                 * the retry loop is too short and in the sync-light case,
 753                 * the overhead of stalling is too much
 754                 */
 755                if (mode != MIGRATE_SYNC) {
 756                        rc = -EBUSY;
 757                        goto uncharge;
 758                }
 759                if (!force)
 760                        goto uncharge;
 761                wait_on_page_writeback(page);
 762        }
 763        /*
 764         * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 765         * we cannot notice that anon_vma is freed while we migrates a page.
 766         * This get_anon_vma() delays freeing anon_vma pointer until the end
 767         * of migration. File cache pages are no problem because of page_lock()
 768         * File Caches may use write_page() or lock_page() in migration, then,
 769         * just care Anon page here.
 770         */
 771        if (PageAnon(page)) {
 772                /*
 773                 * Only page_lock_anon_vma_read() understands the subtleties of
 774                 * getting a hold on an anon_vma from outside one of its mms.
 775                 */
 776                anon_vma = page_get_anon_vma(page);
 777                if (anon_vma) {
 778                        /*
 779                         * Anon page
 780                         */
 781                } else if (PageSwapCache(page)) {
 782                        /*
 783                         * We cannot be sure that the anon_vma of an unmapped
 784                         * swapcache page is safe to use because we don't
 785                         * know in advance if the VMA that this page belonged
 786                         * to still exists. If the VMA and others sharing the
 787                         * data have been freed, then the anon_vma could
 788                         * already be invalid.
 789                         *
 790                         * To avoid this possibility, swapcache pages get
 791                         * migrated but are not remapped when migration
 792                         * completes
 793                         */
 794                        remap_swapcache = 0;
 795                } else {
 796                        goto uncharge;
 797                }
 798        }
 799
 800        if (unlikely(balloon_page_movable(page))) {
 801                /*
 802                 * A ballooned page does not need any special attention from
 803                 * physical to virtual reverse mapping procedures.
 804                 * Skip any attempt to unmap PTEs or to remap swap cache,
 805                 * in order to avoid burning cycles at rmap level, and perform
 806                 * the page migration right away (proteced by page lock).
 807                 */
 808                rc = balloon_page_migrate(newpage, page, mode);
 809                goto uncharge;
 810        }
 811
 812        /*
 813         * Corner case handling:
 814         * 1. When a new swap-cache page is read into, it is added to the LRU
 815         * and treated as swapcache but it has no rmap yet.
 816         * Calling try_to_unmap() against a page->mapping==NULL page will
 817         * trigger a BUG.  So handle it here.
 818         * 2. An orphaned page (see truncate_complete_page) might have
 819         * fs-private metadata. The page can be picked up due to memory
 820         * offlining.  Everywhere else except page reclaim, the page is
 821         * invisible to the vm, so the page can not be migrated.  So try to
 822         * free the metadata, so the page can be freed.
 823         */
 824        if (!page->mapping) {
 825                VM_BUG_ON(PageAnon(page));
 826                if (page_has_private(page)) {
 827                        try_to_free_buffers(page);
 828                        goto uncharge;
 829                }
 830                goto skip_unmap;
 831        }
 832
 833        /* Establish migration ptes or remove ptes */
 834        try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
 835
 836skip_unmap:
 837        if (!page_mapped(page))
 838                rc = move_to_new_page(newpage, page, remap_swapcache, mode);
 839
 840        if (rc && remap_swapcache)
 841                remove_migration_ptes(page, page);
 842
 843        /* Drop an anon_vma reference if we took one */
 844        if (anon_vma)
 845                put_anon_vma(anon_vma);
 846
 847uncharge:
 848        mem_cgroup_end_migration(mem, page, newpage,
 849                                 (rc == MIGRATEPAGE_SUCCESS ||
 850                                  rc == MIGRATEPAGE_BALLOON_SUCCESS));
 851unlock:
 852        unlock_page(page);
 853out:
 854        return rc;
 855}
 856
 857/*
 858 * Obtain the lock on page, remove all ptes and migrate the page
 859 * to the newly allocated page in newpage.
 860 */
 861static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 862                        struct page *page, int force, bool offlining,
 863                        enum migrate_mode mode)
 864{
 865        int rc = 0;
 866        int *result = NULL;
 867        struct page *newpage = get_new_page(page, private, &result);
 868
 869        if (!newpage)
 870                return -ENOMEM;
 871
 872        if (page_count(page) == 1) {
 873                /* page was freed from under us. So we are done. */
 874                goto out;
 875        }
 876
 877        if (unlikely(PageTransHuge(page)))
 878                if (unlikely(split_huge_page(page)))
 879                        goto out;
 880
 881        rc = __unmap_and_move(page, newpage, force, offlining, mode);
 882
 883        if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
 884                /*
 885                 * A ballooned page has been migrated already.
 886                 * Now, it's the time to wrap-up counters,
 887                 * handle the page back to Buddy and return.
 888                 */
 889                dec_zone_page_state(page, NR_ISOLATED_ANON +
 890                                    page_is_file_cache(page));
 891                balloon_page_free(page);
 892                return MIGRATEPAGE_SUCCESS;
 893        }
 894out:
 895        if (rc != -EAGAIN) {
 896                /*
 897                 * A page that has been migrated has all references
 898                 * removed and will be freed. A page that has not been
 899                 * migrated will have kepts its references and be
 900                 * restored.
 901                 */
 902                list_del(&page->lru);
 903                dec_zone_page_state(page, NR_ISOLATED_ANON +
 904                                page_is_file_cache(page));
 905                putback_lru_page(page);
 906        }
 907        /*
 908         * Move the new page to the LRU. If migration was not successful
 909         * then this will free the page.
 910         */
 911        putback_lru_page(newpage);
 912        if (result) {
 913                if (rc)
 914                        *result = rc;
 915                else
 916                        *result = page_to_nid(newpage);
 917        }
 918        return rc;
 919}
 920
 921/*
 922 * Counterpart of unmap_and_move_page() for hugepage migration.
 923 *
 924 * This function doesn't wait the completion of hugepage I/O
 925 * because there is no race between I/O and migration for hugepage.
 926 * Note that currently hugepage I/O occurs only in direct I/O
 927 * where no lock is held and PG_writeback is irrelevant,
 928 * and writeback status of all subpages are counted in the reference
 929 * count of the head page (i.e. if all subpages of a 2MB hugepage are
 930 * under direct I/O, the reference of the head page is 512 and a bit more.)
 931 * This means that when we try to migrate hugepage whose subpages are
 932 * doing direct I/O, some references remain after try_to_unmap() and
 933 * hugepage migration fails without data corruption.
 934 *
 935 * There is also no race when direct I/O is issued on the page under migration,
 936 * because then pte is replaced with migration swap entry and direct I/O code
 937 * will wait in the page fault for migration to complete.
 938 */
 939static int unmap_and_move_huge_page(new_page_t get_new_page,
 940                                unsigned long private, struct page *hpage,
 941                                int force, bool offlining,
 942                                enum migrate_mode mode)
 943{
 944        int rc = 0;
 945        int *result = NULL;
 946        struct page *new_hpage = get_new_page(hpage, private, &result);
 947        struct anon_vma *anon_vma = NULL;
 948
 949        if (!new_hpage)
 950                return -ENOMEM;
 951
 952        rc = -EAGAIN;
 953
 954        if (!trylock_page(hpage)) {
 955                if (!force || mode != MIGRATE_SYNC)
 956                        goto out;
 957                lock_page(hpage);
 958        }
 959
 960        if (PageAnon(hpage))
 961                anon_vma = page_get_anon_vma(hpage);
 962
 963        try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
 964
 965        if (!page_mapped(hpage))
 966                rc = move_to_new_page(new_hpage, hpage, 1, mode);
 967
 968        if (rc)
 969                remove_migration_ptes(hpage, hpage);
 970
 971        if (anon_vma)
 972                put_anon_vma(anon_vma);
 973
 974        if (!rc)
 975                hugetlb_cgroup_migrate(hpage, new_hpage);
 976
 977        unlock_page(hpage);
 978out:
 979        put_page(new_hpage);
 980        if (result) {
 981                if (rc)
 982                        *result = rc;
 983                else
 984                        *result = page_to_nid(new_hpage);
 985        }
 986        return rc;
 987}
 988
 989/*
 990 * migrate_pages
 991 *
 992 * The function takes one list of pages to migrate and a function
 993 * that determines from the page to be migrated and the private data
 994 * the target of the move and allocates the page.
 995 *
 996 * The function returns after 10 attempts or if no pages
 997 * are movable anymore because to has become empty
 998 * or no retryable pages exist anymore.
 999 * Caller should call putback_lru_pages to return pages to the LRU
1000 * or free list only if ret != 0.
1001 *
1002 * Return: Number of pages not migrated or error code.
1003 */
1004int migrate_pages(struct list_head *from,
1005                new_page_t get_new_page, unsigned long private, bool offlining,
1006                enum migrate_mode mode, int reason)
1007{
1008        int retry = 1;
1009        int nr_failed = 0;
1010        int nr_succeeded = 0;
1011        int pass = 0;
1012        struct page *page;
1013        struct page *page2;
1014        int swapwrite = current->flags & PF_SWAPWRITE;
1015        int rc;
1016
1017        if (!swapwrite)
1018                current->flags |= PF_SWAPWRITE;
1019
1020        for(pass = 0; pass < 10 && retry; pass++) {
1021                retry = 0;
1022
1023                list_for_each_entry_safe(page, page2, from, lru) {
1024                        cond_resched();
1025
1026                        rc = unmap_and_move(get_new_page, private,
1027                                                page, pass > 2, offlining,
1028                                                mode);
1029
1030                        switch(rc) {
1031                        case -ENOMEM:
1032                                goto out;
1033                        case -EAGAIN:
1034                                retry++;
1035                                break;
1036                        case MIGRATEPAGE_SUCCESS:
1037                                nr_succeeded++;
1038                                break;
1039                        default:
1040                                /* Permanent failure */
1041                                nr_failed++;
1042                                break;
1043                        }
1044                }
1045        }
1046        rc = nr_failed + retry;
1047out:
1048        if (nr_succeeded)
1049                count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1050        if (nr_failed)
1051                count_vm_events(PGMIGRATE_FAIL, nr_failed);
1052        trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1053
1054        if (!swapwrite)
1055                current->flags &= ~PF_SWAPWRITE;
1056
1057        return rc;
1058}
1059
1060int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1061                      unsigned long private, bool offlining,
1062                      enum migrate_mode mode)
1063{
1064        int pass, rc;
1065
1066        for (pass = 0; pass < 10; pass++) {
1067                rc = unmap_and_move_huge_page(get_new_page,
1068                                              private, hpage, pass > 2, offlining,
1069                                              mode);
1070                switch (rc) {
1071                case -ENOMEM:
1072                        goto out;
1073                case -EAGAIN:
1074                        /* try again */
1075                        cond_resched();
1076                        break;
1077                case MIGRATEPAGE_SUCCESS:
1078                        goto out;
1079                default:
1080                        rc = -EIO;
1081                        goto out;
1082                }
1083        }
1084out:
1085        return rc;
1086}
1087
1088#ifdef CONFIG_NUMA
1089/*
1090 * Move a list of individual pages
1091 */
1092struct page_to_node {
1093        unsigned long addr;
1094        struct page *page;
1095        int node;
1096        int status;
1097};
1098
1099static struct page *new_page_node(struct page *p, unsigned long private,
1100                int **result)
1101{
1102        struct page_to_node *pm = (struct page_to_node *)private;
1103
1104        while (pm->node != MAX_NUMNODES && pm->page != p)
1105                pm++;
1106
1107        if (pm->node == MAX_NUMNODES)
1108                return NULL;
1109
1110        *result = &pm->status;
1111
1112        return alloc_pages_exact_node(pm->node,
1113                                GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1114}
1115
1116/*
1117 * Move a set of pages as indicated in the pm array. The addr
1118 * field must be set to the virtual address of the page to be moved
1119 * and the node number must contain a valid target node.
1120 * The pm array ends with node = MAX_NUMNODES.
1121 */
1122static int do_move_page_to_node_array(struct mm_struct *mm,
1123                                      struct page_to_node *pm,
1124                                      int migrate_all)
1125{
1126        int err;
1127        struct page_to_node *pp;
1128        LIST_HEAD(pagelist);
1129
1130        down_read(&mm->mmap_sem);
1131
1132        /*
1133         * Build a list of pages to migrate
1134         */
1135        for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1136                struct vm_area_struct *vma;
1137                struct page *page;
1138
1139                err = -EFAULT;
1140                vma = find_vma(mm, pp->addr);
1141                if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1142                        goto set_status;
1143
1144                page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1145
1146                err = PTR_ERR(page);
1147                if (IS_ERR(page))
1148                        goto set_status;
1149
1150                err = -ENOENT;
1151                if (!page)
1152                        goto set_status;
1153
1154                /* Use PageReserved to check for zero page */
1155                if (PageReserved(page) || PageKsm(page))
1156                        goto put_and_set;
1157
1158                pp->page = page;
1159                err = page_to_nid(page);
1160
1161                if (err == pp->node)
1162                        /*
1163                         * Node already in the right place
1164                         */
1165                        goto put_and_set;
1166
1167                err = -EACCES;
1168                if (page_mapcount(page) > 1 &&
1169                                !migrate_all)
1170                        goto put_and_set;
1171
1172                err = isolate_lru_page(page);
1173                if (!err) {
1174                        list_add_tail(&page->lru, &pagelist);
1175                        inc_zone_page_state(page, NR_ISOLATED_ANON +
1176                                            page_is_file_cache(page));
1177                }
1178put_and_set:
1179                /*
1180                 * Either remove the duplicate refcount from
1181                 * isolate_lru_page() or drop the page ref if it was
1182                 * not isolated.
1183                 */
1184                put_page(page);
1185set_status:
1186                pp->status = err;
1187        }
1188
1189        err = 0;
1190        if (!list_empty(&pagelist)) {
1191                err = migrate_pages(&pagelist, new_page_node,
1192                                (unsigned long)pm, 0, MIGRATE_SYNC,
1193                                MR_SYSCALL);
1194                if (err)
1195                        putback_lru_pages(&pagelist);
1196        }
1197
1198        up_read(&mm->mmap_sem);
1199        return err;
1200}
1201
1202/*
1203 * Migrate an array of page address onto an array of nodes and fill
1204 * the corresponding array of status.
1205 */
1206static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1207                         unsigned long nr_pages,
1208                         const void __user * __user *pages,
1209                         const int __user *nodes,
1210                         int __user *status, int flags)
1211{
1212        struct page_to_node *pm;
1213        unsigned long chunk_nr_pages;
1214        unsigned long chunk_start;
1215        int err;
1216
1217        err = -ENOMEM;
1218        pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1219        if (!pm)
1220                goto out;
1221
1222        migrate_prep();
1223
1224        /*
1225         * Store a chunk of page_to_node array in a page,
1226         * but keep the last one as a marker
1227         */
1228        chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1229
1230        for (chunk_start = 0;
1231             chunk_start < nr_pages;
1232             chunk_start += chunk_nr_pages) {
1233                int j;
1234
1235                if (chunk_start + chunk_nr_pages > nr_pages)
1236                        chunk_nr_pages = nr_pages - chunk_start;
1237
1238                /* fill the chunk pm with addrs and nodes from user-space */
1239                for (j = 0; j < chunk_nr_pages; j++) {
1240                        const void __user *p;
1241                        int node;
1242
1243                        err = -EFAULT;
1244                        if (get_user(p, pages + j + chunk_start))
1245                                goto out_pm;
1246                        pm[j].addr = (unsigned long) p;
1247
1248                        if (get_user(node, nodes + j + chunk_start))
1249                                goto out_pm;
1250
1251                        err = -ENODEV;
1252                        if (node < 0 || node >= MAX_NUMNODES)
1253                                goto out_pm;
1254
1255                        if (!node_state(node, N_MEMORY))
1256                                goto out_pm;
1257
1258                        err = -EACCES;
1259                        if (!node_isset(node, task_nodes))
1260                                goto out_pm;
1261
1262                        pm[j].node = node;
1263                }
1264
1265                /* End marker for this chunk */
1266                pm[chunk_nr_pages].node = MAX_NUMNODES;
1267
1268                /* Migrate this chunk */
1269                err = do_move_page_to_node_array(mm, pm,
1270                                                 flags & MPOL_MF_MOVE_ALL);
1271                if (err < 0)
1272                        goto out_pm;
1273
1274                /* Return status information */
1275                for (j = 0; j < chunk_nr_pages; j++)
1276                        if (put_user(pm[j].status, status + j + chunk_start)) {
1277                                err = -EFAULT;
1278                                goto out_pm;
1279                        }
1280        }
1281        err = 0;
1282
1283out_pm:
1284        free_page((unsigned long)pm);
1285out:
1286        return err;
1287}
1288
1289/*
1290 * Determine the nodes of an array of pages and store it in an array of status.
1291 */
1292static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1293                                const void __user **pages, int *status)
1294{
1295        unsigned long i;
1296
1297        down_read(&mm->mmap_sem);
1298
1299        for (i = 0; i < nr_pages; i++) {
1300                unsigned long addr = (unsigned long)(*pages);
1301                struct vm_area_struct *vma;
1302                struct page *page;
1303                int err = -EFAULT;
1304
1305                vma = find_vma(mm, addr);
1306                if (!vma || addr < vma->vm_start)
1307                        goto set_status;
1308
1309                page = follow_page(vma, addr, 0);
1310
1311                err = PTR_ERR(page);
1312                if (IS_ERR(page))
1313                        goto set_status;
1314
1315                err = -ENOENT;
1316                /* Use PageReserved to check for zero page */
1317                if (!page || PageReserved(page) || PageKsm(page))
1318                        goto set_status;
1319
1320                err = page_to_nid(page);
1321set_status:
1322                *status = err;
1323
1324                pages++;
1325                status++;
1326        }
1327
1328        up_read(&mm->mmap_sem);
1329}
1330
1331/*
1332 * Determine the nodes of a user array of pages and store it in
1333 * a user array of status.
1334 */
1335static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1336                         const void __user * __user *pages,
1337                         int __user *status)
1338{
1339#define DO_PAGES_STAT_CHUNK_NR 16
1340        const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1341        int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1342
1343        while (nr_pages) {
1344                unsigned long chunk_nr;
1345
1346                chunk_nr = nr_pages;
1347                if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1348                        chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1349
1350                if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1351                        break;
1352
1353                do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1354
1355                if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1356                        break;
1357
1358                pages += chunk_nr;
1359                status += chunk_nr;
1360                nr_pages -= chunk_nr;
1361        }
1362        return nr_pages ? -EFAULT : 0;
1363}
1364
1365/*
1366 * Move a list of pages in the address space of the currently executing
1367 * process.
1368 */
1369SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1370                const void __user * __user *, pages,
1371                const int __user *, nodes,
1372                int __user *, status, int, flags)
1373{
1374        const struct cred *cred = current_cred(), *tcred;
1375        struct task_struct *task;
1376        struct mm_struct *mm;
1377        int err;
1378        nodemask_t task_nodes;
1379
1380        /* Check flags */
1381        if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1382                return -EINVAL;
1383
1384        if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1385                return -EPERM;
1386
1387        /* Find the mm_struct */
1388        rcu_read_lock();
1389        task = pid ? find_task_by_vpid(pid) : current;
1390        if (!task) {
1391                rcu_read_unlock();
1392                return -ESRCH;
1393        }
1394        get_task_struct(task);
1395
1396        /*
1397         * Check if this process has the right to modify the specified
1398         * process. The right exists if the process has administrative
1399         * capabilities, superuser privileges or the same
1400         * userid as the target process.
1401         */
1402        tcred = __task_cred(task);
1403        if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1404            !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1405            !capable(CAP_SYS_NICE)) {
1406                rcu_read_unlock();
1407                err = -EPERM;
1408                goto out;
1409        }
1410        rcu_read_unlock();
1411
1412        err = security_task_movememory(task);
1413        if (err)
1414                goto out;
1415
1416        task_nodes = cpuset_mems_allowed(task);
1417        mm = get_task_mm(task);
1418        put_task_struct(task);
1419
1420        if (!mm)
1421                return -EINVAL;
1422
1423        if (nodes)
1424                err = do_pages_move(mm, task_nodes, nr_pages, pages,
1425                                    nodes, status, flags);
1426        else
1427                err = do_pages_stat(mm, nr_pages, pages, status);
1428
1429        mmput(mm);
1430        return err;
1431
1432out:
1433        put_task_struct(task);
1434        return err;
1435}
1436
1437/*
1438 * Call migration functions in the vma_ops that may prepare
1439 * memory in a vm for migration. migration functions may perform
1440 * the migration for vmas that do not have an underlying page struct.
1441 */
1442int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1443        const nodemask_t *from, unsigned long flags)
1444{
1445        struct vm_area_struct *vma;
1446        int err = 0;
1447
1448        for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1449                if (vma->vm_ops && vma->vm_ops->migrate) {
1450                        err = vma->vm_ops->migrate(vma, to, from, flags);
1451                        if (err)
1452                                break;
1453                }
1454        }
1455        return err;
1456}
1457
1458#ifdef CONFIG_NUMA_BALANCING
1459/*
1460 * Returns true if this is a safe migration target node for misplaced NUMA
1461 * pages. Currently it only checks the watermarks which crude
1462 */
1463static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1464                                   int nr_migrate_pages)
1465{
1466        int z;
1467        for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1468                struct zone *zone = pgdat->node_zones + z;
1469
1470                if (!populated_zone(zone))
1471                        continue;
1472
1473                if (zone->all_unreclaimable)
1474                        continue;
1475
1476                /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1477                if (!zone_watermark_ok(zone, 0,
1478                                       high_wmark_pages(zone) +
1479                                       nr_migrate_pages,
1480                                       0, 0))
1481                        continue;
1482                return true;
1483        }
1484        return false;
1485}
1486
1487static struct page *alloc_misplaced_dst_page(struct page *page,
1488                                           unsigned long data,
1489                                           int **result)
1490{
1491        int nid = (int) data;
1492        struct page *newpage;
1493
1494        newpage = alloc_pages_exact_node(nid,
1495                                         (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1496                                          __GFP_NOMEMALLOC | __GFP_NORETRY |
1497                                          __GFP_NOWARN) &
1498                                         ~GFP_IOFS, 0);
1499        if (newpage)
1500                page_xchg_last_nid(newpage, page_last_nid(page));
1501
1502        return newpage;
1503}
1504
1505/*
1506 * page migration rate limiting control.
1507 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1508 * window of time. Default here says do not migrate more than 1280M per second.
1509 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1510 * as it is faults that reset the window, pte updates will happen unconditionally
1511 * if there has not been a fault since @pteupdate_interval_millisecs after the
1512 * throttle window closed.
1513 */
1514static unsigned int migrate_interval_millisecs __read_mostly = 100;
1515static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1516static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1517
1518/* Returns true if NUMA migration is currently rate limited */
1519bool migrate_ratelimited(int node)
1520{
1521        pg_data_t *pgdat = NODE_DATA(node);
1522
1523        if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1524                                msecs_to_jiffies(pteupdate_interval_millisecs)))
1525                return false;
1526
1527        if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1528                return false;
1529
1530        return true;
1531}
1532
1533/* Returns true if the node is migrate rate-limited after the update */
1534bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
1535{
1536        bool rate_limited = false;
1537
1538        /*
1539         * Rate-limit the amount of data that is being migrated to a node.
1540         * Optimal placement is no good if the memory bus is saturated and
1541         * all the time is being spent migrating!
1542         */
1543        spin_lock(&pgdat->numabalancing_migrate_lock);
1544        if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1545                pgdat->numabalancing_migrate_nr_pages = 0;
1546                pgdat->numabalancing_migrate_next_window = jiffies +
1547                        msecs_to_jiffies(migrate_interval_millisecs);
1548        }
1549        if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1550                rate_limited = true;
1551        else
1552                pgdat->numabalancing_migrate_nr_pages += nr_pages;
1553        spin_unlock(&pgdat->numabalancing_migrate_lock);
1554        
1555        return rate_limited;
1556}
1557
1558int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1559{
1560        int ret = 0;
1561
1562        /* Avoid migrating to a node that is nearly full */
1563        if (migrate_balanced_pgdat(pgdat, 1)) {
1564                int page_lru;
1565
1566                if (isolate_lru_page(page)) {
1567                        put_page(page);
1568                        return 0;
1569                }
1570
1571                /* Page is isolated */
1572                ret = 1;
1573                page_lru = page_is_file_cache(page);
1574                if (!PageTransHuge(page))
1575                        inc_zone_page_state(page, NR_ISOLATED_ANON + page_lru);
1576                else
1577                        mod_zone_page_state(page_zone(page),
1578                                        NR_ISOLATED_ANON + page_lru,
1579                                        HPAGE_PMD_NR);
1580        }
1581
1582        /*
1583         * Page is either isolated or there is not enough space on the target
1584         * node. If isolated, then it has taken a reference count and the
1585         * callers reference can be safely dropped without the page
1586         * disappearing underneath us during migration. Otherwise the page is
1587         * not to be migrated but the callers reference should still be
1588         * dropped so it does not leak.
1589         */
1590        put_page(page);
1591
1592        return ret;
1593}
1594
1595/*
1596 * Attempt to migrate a misplaced page to the specified destination
1597 * node. Caller is expected to have an elevated reference count on
1598 * the page that will be dropped by this function before returning.
1599 */
1600int migrate_misplaced_page(struct page *page, int node)
1601{
1602        pg_data_t *pgdat = NODE_DATA(node);
1603        int isolated = 0;
1604        int nr_remaining;
1605        LIST_HEAD(migratepages);
1606
1607        /*
1608         * Don't migrate pages that are mapped in multiple processes.
1609         * TODO: Handle false sharing detection instead of this hammer
1610         */
1611        if (page_mapcount(page) != 1) {
1612                put_page(page);
1613                goto out;
1614        }
1615
1616        /*
1617         * Rate-limit the amount of data that is being migrated to a node.
1618         * Optimal placement is no good if the memory bus is saturated and
1619         * all the time is being spent migrating!
1620         */
1621        if (numamigrate_update_ratelimit(pgdat, 1)) {
1622                put_page(page);
1623                goto out;
1624        }
1625
1626        isolated = numamigrate_isolate_page(pgdat, page);
1627        if (!isolated)
1628                goto out;
1629
1630        list_add(&page->lru, &migratepages);
1631        nr_remaining = migrate_pages(&migratepages,
1632                        alloc_misplaced_dst_page,
1633                        node, false, MIGRATE_ASYNC,
1634                        MR_NUMA_MISPLACED);
1635        if (nr_remaining) {
1636                putback_lru_pages(&migratepages);
1637                isolated = 0;
1638        } else
1639                count_vm_numa_event(NUMA_PAGE_MIGRATE);
1640        BUG_ON(!list_empty(&migratepages));
1641out:
1642        return isolated;
1643}
1644#endif /* CONFIG_NUMA_BALANCING */
1645
1646#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1647int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1648                                struct vm_area_struct *vma,
1649                                pmd_t *pmd, pmd_t entry,
1650                                unsigned long address,
1651                                struct page *page, int node)
1652{
1653        unsigned long haddr = address & HPAGE_PMD_MASK;
1654        pg_data_t *pgdat = NODE_DATA(node);
1655        int isolated = 0;
1656        struct page *new_page = NULL;
1657        struct mem_cgroup *memcg = NULL;
1658        int page_lru = page_is_file_cache(page);
1659
1660        /*
1661         * Don't migrate pages that are mapped in multiple processes.
1662         * TODO: Handle false sharing detection instead of this hammer
1663         */
1664        if (page_mapcount(page) != 1)
1665                goto out_dropref;
1666
1667        /*
1668         * Rate-limit the amount of data that is being migrated to a node.
1669         * Optimal placement is no good if the memory bus is saturated and
1670         * all the time is being spent migrating!
1671         */
1672        if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1673                goto out_dropref;
1674
1675        new_page = alloc_pages_node(node,
1676                (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1677        if (!new_page) {
1678                count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1679                goto out_dropref;
1680        }
1681        page_xchg_last_nid(new_page, page_last_nid(page));
1682
1683        isolated = numamigrate_isolate_page(pgdat, page);
1684
1685        /*
1686         * Failing to isolate or a GUP pin prevents migration. The expected
1687         * page count is 2. 1 for anonymous pages without a mapping and 1
1688         * for the callers pin. If the page was isolated, the page will
1689         * need to be put back on the LRU.
1690         */
1691        if (!isolated || page_count(page) != 2) {
1692                count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1693                put_page(new_page);
1694                if (isolated) {
1695                        putback_lru_page(page);
1696                        isolated = 0;
1697                        goto out;
1698                }
1699                goto out_keep_locked;
1700        }
1701
1702        /* Prepare a page as a migration target */
1703        __set_page_locked(new_page);
1704        SetPageSwapBacked(new_page);
1705
1706        /* anon mapping, we can simply copy page->mapping to the new page: */
1707        new_page->mapping = page->mapping;
1708        new_page->index = page->index;
1709        migrate_page_copy(new_page, page);
1710        WARN_ON(PageLRU(new_page));
1711
1712        /* Recheck the target PMD */
1713        spin_lock(&mm->page_table_lock);
1714        if (unlikely(!pmd_same(*pmd, entry))) {
1715                spin_unlock(&mm->page_table_lock);
1716
1717                /* Reverse changes made by migrate_page_copy() */
1718                if (TestClearPageActive(new_page))
1719                        SetPageActive(page);
1720                if (TestClearPageUnevictable(new_page))
1721                        SetPageUnevictable(page);
1722                mlock_migrate_page(page, new_page);
1723
1724                unlock_page(new_page);
1725                put_page(new_page);             /* Free it */
1726
1727                unlock_page(page);
1728                putback_lru_page(page);
1729
1730                count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1731                goto out;
1732        }
1733
1734        /*
1735         * Traditional migration needs to prepare the memcg charge
1736         * transaction early to prevent the old page from being
1737         * uncharged when installing migration entries.  Here we can
1738         * save the potential rollback and start the charge transfer
1739         * only when migration is already known to end successfully.
1740         */
1741        mem_cgroup_prepare_migration(page, new_page, &memcg);
1742
1743        entry = mk_pmd(new_page, vma->vm_page_prot);
1744        entry = pmd_mknonnuma(entry);
1745        entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1746        entry = pmd_mkhuge(entry);
1747
1748        page_add_new_anon_rmap(new_page, vma, haddr);
1749
1750        set_pmd_at(mm, haddr, pmd, entry);
1751        update_mmu_cache_pmd(vma, address, &entry);
1752        page_remove_rmap(page);
1753        /*
1754         * Finish the charge transaction under the page table lock to
1755         * prevent split_huge_page() from dividing up the charge
1756         * before it's fully transferred to the new page.
1757         */
1758        mem_cgroup_end_migration(memcg, page, new_page, true);
1759        spin_unlock(&mm->page_table_lock);
1760
1761        unlock_page(new_page);
1762        unlock_page(page);
1763        put_page(page);                 /* Drop the rmap reference */
1764        put_page(page);                 /* Drop the LRU isolation reference */
1765
1766        count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1767        count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1768
1769out:
1770        mod_zone_page_state(page_zone(page),
1771                        NR_ISOLATED_ANON + page_lru,
1772                        -HPAGE_PMD_NR);
1773        return isolated;
1774
1775out_dropref:
1776        put_page(page);
1777out_keep_locked:
1778        return 0;
1779}
1780#endif /* CONFIG_NUMA_BALANCING */
1781
1782#endif /* CONFIG_NUMA */
1783
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.