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_dcache_page(new);
 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         * Please do not reorder this without considering how mm/ksm.c's
 469         * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
 470         */
 471        ClearPageSwapCache(page);
 472        ClearPagePrivate(page);
 473        set_page_private(page, 0);
 474
 475        /*
 476         * If any waiters have accumulated on the new page then
 477         * wake them up.
 478         */
 479        if (PageWriteback(newpage))
 480                end_page_writeback(newpage);
 481}
 482
 483/************************************************************
 484 *                    Migration functions
 485 ***********************************************************/
 486
 487/* Always fail migration. Used for mappings that are not movable */
 488int fail_migrate_page(struct address_space *mapping,
 489                        struct page *newpage, struct page *page)
 490{
 491        return -EIO;
 492}
 493EXPORT_SYMBOL(fail_migrate_page);
 494
 495/*
 496 * Common logic to directly migrate a single page suitable for
 497 * pages that do not use PagePrivate/PagePrivate2.
 498 *
 499 * Pages are locked upon entry and exit.
 500 */
 501int migrate_page(struct address_space *mapping,
 502                struct page *newpage, struct page *page,
 503                enum migrate_mode mode)
 504{
 505        int rc;
 506
 507        BUG_ON(PageWriteback(page));    /* Writeback must be complete */
 508
 509        rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
 510
 511        if (rc != MIGRATEPAGE_SUCCESS)
 512                return rc;
 513
 514        migrate_page_copy(newpage, page);
 515        return MIGRATEPAGE_SUCCESS;
 516}
 517EXPORT_SYMBOL(migrate_page);
 518
 519#ifdef CONFIG_BLOCK
 520/*
 521 * Migration function for pages with buffers. This function can only be used
 522 * if the underlying filesystem guarantees that no other references to "page"
 523 * exist.
 524 */
 525int buffer_migrate_page(struct address_space *mapping,
 526                struct page *newpage, struct page *page, enum migrate_mode mode)
 527{
 528        struct buffer_head *bh, *head;
 529        int rc;
 530
 531        if (!page_has_buffers(page))
 532                return migrate_page(mapping, newpage, page, mode);
 533
 534        head = page_buffers(page);
 535
 536        rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
 537
 538        if (rc != MIGRATEPAGE_SUCCESS)
 539                return rc;
 540
 541        /*
 542         * In the async case, migrate_page_move_mapping locked the buffers
 543         * with an IRQ-safe spinlock held. In the sync case, the buffers
 544         * need to be locked now
 545         */
 546        if (mode != MIGRATE_ASYNC)
 547                BUG_ON(!buffer_migrate_lock_buffers(head, mode));
 548
 549        ClearPagePrivate(page);
 550        set_page_private(newpage, page_private(page));
 551        set_page_private(page, 0);
 552        put_page(page);
 553        get_page(newpage);
 554
 555        bh = head;
 556        do {
 557                set_bh_page(bh, newpage, bh_offset(bh));
 558                bh = bh->b_this_page;
 559
 560        } while (bh != head);
 561
 562        SetPagePrivate(newpage);
 563
 564        migrate_page_copy(newpage, page);
 565
 566        bh = head;
 567        do {
 568                unlock_buffer(bh);
 569                put_bh(bh);
 570                bh = bh->b_this_page;
 571
 572        } while (bh != head);
 573
 574        return MIGRATEPAGE_SUCCESS;
 575}
 576EXPORT_SYMBOL(buffer_migrate_page);
 577#endif
 578
 579/*
 580 * Writeback a page to clean the dirty state
 581 */
 582static int writeout(struct address_space *mapping, struct page *page)
 583{
 584        struct writeback_control wbc = {
 585                .sync_mode = WB_SYNC_NONE,
 586                .nr_to_write = 1,
 587                .range_start = 0,
 588                .range_end = LLONG_MAX,
 589                .for_reclaim = 1
 590        };
 591        int rc;
 592
 593        if (!mapping->a_ops->writepage)
 594                /* No write method for the address space */
 595                return -EINVAL;
 596
 597        if (!clear_page_dirty_for_io(page))
 598                /* Someone else already triggered a write */
 599                return -EAGAIN;
 600
 601        /*
 602         * A dirty page may imply that the underlying filesystem has
 603         * the page on some queue. So the page must be clean for
 604         * migration. Writeout may mean we loose the lock and the
 605         * page state is no longer what we checked for earlier.
 606         * At this point we know that the migration attempt cannot
 607         * be successful.
 608         */
 609        remove_migration_ptes(page, page);
 610
 611        rc = mapping->a_ops->writepage(page, &wbc);
 612
 613        if (rc != AOP_WRITEPAGE_ACTIVATE)
 614                /* unlocked. Relock */
 615                lock_page(page);
 616
 617        return (rc < 0) ? -EIO : -EAGAIN;
 618}
 619
 620/*
 621 * Default handling if a filesystem does not provide a migration function.
 622 */
 623static int fallback_migrate_page(struct address_space *mapping,
 624        struct page *newpage, struct page *page, enum migrate_mode mode)
 625{
 626        if (PageDirty(page)) {
 627                /* Only writeback pages in full synchronous migration */
 628                if (mode != MIGRATE_SYNC)
 629                        return -EBUSY;
 630                return writeout(mapping, page);
 631        }
 632
 633        /*
 634         * Buffers may be managed in a filesystem specific way.
 635         * We must have no buffers or drop them.
 636         */
 637        if (page_has_private(page) &&
 638            !try_to_release_page(page, GFP_KERNEL))
 639                return -EAGAIN;
 640
 641        return migrate_page(mapping, newpage, page, mode);
 642}
 643
 644/*
 645 * Move a page to a newly allocated page
 646 * The page is locked and all ptes have been successfully removed.
 647 *
 648 * The new page will have replaced the old page if this function
 649 * is successful.
 650 *
 651 * Return value:
 652 *   < 0 - error code
 653 *  MIGRATEPAGE_SUCCESS - success
 654 */
 655static int move_to_new_page(struct page *newpage, struct page *page,
 656                                int remap_swapcache, enum migrate_mode mode)
 657{
 658        struct address_space *mapping;
 659        int rc;
 660
 661        /*
 662         * Block others from accessing the page when we get around to
 663         * establishing additional references. We are the only one
 664         * holding a reference to the new page at this point.
 665         */
 666        if (!trylock_page(newpage))
 667                BUG();
 668
 669        /* Prepare mapping for the new page.*/
 670        newpage->index = page->index;
 671        newpage->mapping = page->mapping;
 672        if (PageSwapBacked(page))
 673                SetPageSwapBacked(newpage);
 674
 675        mapping = page_mapping(page);
 676        if (!mapping)
 677                rc = migrate_page(mapping, newpage, page, mode);
 678        else if (mapping->a_ops->migratepage)
 679                /*
 680                 * Most pages have a mapping and most filesystems provide a
 681                 * migratepage callback. Anonymous pages are part of swap
 682                 * space which also has its own migratepage callback. This
 683                 * is the most common path for page migration.
 684                 */
 685                rc = mapping->a_ops->migratepage(mapping,
 686                                                newpage, page, mode);
 687        else
 688                rc = fallback_migrate_page(mapping, newpage, page, mode);
 689
 690        if (rc != MIGRATEPAGE_SUCCESS) {
 691                newpage->mapping = NULL;
 692        } else {
 693                if (remap_swapcache)
 694                        remove_migration_ptes(page, newpage);
 695                page->mapping = NULL;
 696        }
 697
 698        unlock_page(newpage);
 699
 700        return rc;
 701}
 702
 703static int __unmap_and_move(struct page *page, struct page *newpage,
 704                                int force, enum migrate_mode mode)
 705{
 706        int rc = -EAGAIN;
 707        int remap_swapcache = 1;
 708        struct mem_cgroup *mem;
 709        struct anon_vma *anon_vma = NULL;
 710
 711        if (!trylock_page(page)) {
 712                if (!force || mode == MIGRATE_ASYNC)
 713                        goto out;
 714
 715                /*
 716                 * It's not safe for direct compaction to call lock_page.
 717                 * For example, during page readahead pages are added locked
 718                 * to the LRU. Later, when the IO completes the pages are
 719                 * marked uptodate and unlocked. However, the queueing
 720                 * could be merging multiple pages for one bio (e.g.
 721                 * mpage_readpages). If an allocation happens for the
 722                 * second or third page, the process can end up locking
 723                 * the same page twice and deadlocking. Rather than
 724                 * trying to be clever about what pages can be locked,
 725                 * avoid the use of lock_page for direct compaction
 726                 * altogether.
 727                 */
 728                if (current->flags & PF_MEMALLOC)
 729                        goto out;
 730
 731                lock_page(page);
 732        }
 733
 734        /* charge against new page */
 735        mem_cgroup_prepare_migration(page, newpage, &mem);
 736
 737        if (PageWriteback(page)) {
 738                /*
 739                 * Only in the case of a full syncronous migration is it
 740                 * necessary to wait for PageWriteback. In the async case,
 741                 * the retry loop is too short and in the sync-light case,
 742                 * the overhead of stalling is too much
 743                 */
 744                if (mode != MIGRATE_SYNC) {
 745                        rc = -EBUSY;
 746                        goto uncharge;
 747                }
 748                if (!force)
 749                        goto uncharge;
 750                wait_on_page_writeback(page);
 751        }
 752        /*
 753         * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
 754         * we cannot notice that anon_vma is freed while we migrates a page.
 755         * This get_anon_vma() delays freeing anon_vma pointer until the end
 756         * of migration. File cache pages are no problem because of page_lock()
 757         * File Caches may use write_page() or lock_page() in migration, then,
 758         * just care Anon page here.
 759         */
 760        if (PageAnon(page) && !PageKsm(page)) {
 761                /*
 762                 * Only page_lock_anon_vma_read() understands the subtleties of
 763                 * getting a hold on an anon_vma from outside one of its mms.
 764                 */
 765                anon_vma = page_get_anon_vma(page);
 766                if (anon_vma) {
 767                        /*
 768                         * Anon page
 769                         */
 770                } else if (PageSwapCache(page)) {
 771                        /*
 772                         * We cannot be sure that the anon_vma of an unmapped
 773                         * swapcache page is safe to use because we don't
 774                         * know in advance if the VMA that this page belonged
 775                         * to still exists. If the VMA and others sharing the
 776                         * data have been freed, then the anon_vma could
 777                         * already be invalid.
 778                         *
 779                         * To avoid this possibility, swapcache pages get
 780                         * migrated but are not remapped when migration
 781                         * completes
 782                         */
 783                        remap_swapcache = 0;
 784                } else {
 785                        goto uncharge;
 786                }
 787        }
 788
 789        if (unlikely(balloon_page_movable(page))) {
 790                /*
 791                 * A ballooned page does not need any special attention from
 792                 * physical to virtual reverse mapping procedures.
 793                 * Skip any attempt to unmap PTEs or to remap swap cache,
 794                 * in order to avoid burning cycles at rmap level, and perform
 795                 * the page migration right away (proteced by page lock).
 796                 */
 797                rc = balloon_page_migrate(newpage, page, mode);
 798                goto uncharge;
 799        }
 800
 801        /*
 802         * Corner case handling:
 803         * 1. When a new swap-cache page is read into, it is added to the LRU
 804         * and treated as swapcache but it has no rmap yet.
 805         * Calling try_to_unmap() against a page->mapping==NULL page will
 806         * trigger a BUG.  So handle it here.
 807         * 2. An orphaned page (see truncate_complete_page) might have
 808         * fs-private metadata. The page can be picked up due to memory
 809         * offlining.  Everywhere else except page reclaim, the page is
 810         * invisible to the vm, so the page can not be migrated.  So try to
 811         * free the metadata, so the page can be freed.
 812         */
 813        if (!page->mapping) {
 814                VM_BUG_ON(PageAnon(page));
 815                if (page_has_private(page)) {
 816                        try_to_free_buffers(page);
 817                        goto uncharge;
 818                }
 819                goto skip_unmap;
 820        }
 821
 822        /* Establish migration ptes or remove ptes */
 823        try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
 824
 825skip_unmap:
 826        if (!page_mapped(page))
 827                rc = move_to_new_page(newpage, page, remap_swapcache, mode);
 828
 829        if (rc && remap_swapcache)
 830                remove_migration_ptes(page, page);
 831
 832        /* Drop an anon_vma reference if we took one */
 833        if (anon_vma)
 834                put_anon_vma(anon_vma);
 835
 836uncharge:
 837        mem_cgroup_end_migration(mem, page, newpage,
 838                                 (rc == MIGRATEPAGE_SUCCESS ||
 839                                  rc == MIGRATEPAGE_BALLOON_SUCCESS));
 840        unlock_page(page);
 841out:
 842        return rc;
 843}
 844
 845/*
 846 * Obtain the lock on page, remove all ptes and migrate the page
 847 * to the newly allocated page in newpage.
 848 */
 849static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 850                        struct page *page, int force, enum migrate_mode mode)
 851{
 852        int rc = 0;
 853        int *result = NULL;
 854        struct page *newpage = get_new_page(page, private, &result);
 855
 856        if (!newpage)
 857                return -ENOMEM;
 858
 859        if (page_count(page) == 1) {
 860                /* page was freed from under us. So we are done. */
 861                goto out;
 862        }
 863
 864        if (unlikely(PageTransHuge(page)))
 865                if (unlikely(split_huge_page(page)))
 866                        goto out;
 867
 868        rc = __unmap_and_move(page, newpage, force, mode);
 869
 870        if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
 871                /*
 872                 * A ballooned page has been migrated already.
 873                 * Now, it's the time to wrap-up counters,
 874                 * handle the page back to Buddy and return.
 875                 */
 876                dec_zone_page_state(page, NR_ISOLATED_ANON +
 877                                    page_is_file_cache(page));
 878                balloon_page_free(page);
 879                return MIGRATEPAGE_SUCCESS;
 880        }
 881out:
 882        if (rc != -EAGAIN) {
 883                /*
 884                 * A page that has been migrated has all references
 885                 * removed and will be freed. A page that has not been
 886                 * migrated will have kepts its references and be
 887                 * restored.
 888                 */
 889                list_del(&page->lru);
 890                dec_zone_page_state(page, NR_ISOLATED_ANON +
 891                                page_is_file_cache(page));
 892                putback_lru_page(page);
 893        }
 894        /*
 895         * Move the new page to the LRU. If migration was not successful
 896         * then this will free the page.
 897         */
 898        putback_lru_page(newpage);
 899        if (result) {
 900                if (rc)
 901                        *result = rc;
 902                else
 903                        *result = page_to_nid(newpage);
 904        }
 905        return rc;
 906}
 907
 908/*
 909 * Counterpart of unmap_and_move_page() for hugepage migration.
 910 *
 911 * This function doesn't wait the completion of hugepage I/O
 912 * because there is no race between I/O and migration for hugepage.
 913 * Note that currently hugepage I/O occurs only in direct I/O
 914 * where no lock is held and PG_writeback is irrelevant,
 915 * and writeback status of all subpages are counted in the reference
 916 * count of the head page (i.e. if all subpages of a 2MB hugepage are
 917 * under direct I/O, the reference of the head page is 512 and a bit more.)
 918 * This means that when we try to migrate hugepage whose subpages are
 919 * doing direct I/O, some references remain after try_to_unmap() and
 920 * hugepage migration fails without data corruption.
 921 *
 922 * There is also no race when direct I/O is issued on the page under migration,
 923 * because then pte is replaced with migration swap entry and direct I/O code
 924 * will wait in the page fault for migration to complete.
 925 */
 926static int unmap_and_move_huge_page(new_page_t get_new_page,
 927                                unsigned long private, struct page *hpage,
 928                                int force, enum migrate_mode mode)
 929{
 930        int rc = 0;
 931        int *result = NULL;
 932        struct page *new_hpage = get_new_page(hpage, private, &result);
 933        struct anon_vma *anon_vma = NULL;
 934
 935        if (!new_hpage)
 936                return -ENOMEM;
 937
 938        rc = -EAGAIN;
 939
 940        if (!trylock_page(hpage)) {
 941                if (!force || mode != MIGRATE_SYNC)
 942                        goto out;
 943                lock_page(hpage);
 944        }
 945
 946        if (PageAnon(hpage))
 947                anon_vma = page_get_anon_vma(hpage);
 948
 949        try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
 950
 951        if (!page_mapped(hpage))
 952                rc = move_to_new_page(new_hpage, hpage, 1, mode);
 953
 954        if (rc)
 955                remove_migration_ptes(hpage, hpage);
 956
 957        if (anon_vma)
 958                put_anon_vma(anon_vma);
 959
 960        if (!rc)
 961                hugetlb_cgroup_migrate(hpage, new_hpage);
 962
 963        unlock_page(hpage);
 964out:
 965        put_page(new_hpage);
 966        if (result) {
 967                if (rc)
 968                        *result = rc;
 969                else
 970                        *result = page_to_nid(new_hpage);
 971        }
 972        return rc;
 973}
 974
 975/*
 976 * migrate_pages
 977 *
 978 * The function takes one list of pages to migrate and a function
 979 * that determines from the page to be migrated and the private data
 980 * the target of the move and allocates the page.
 981 *
 982 * The function returns after 10 attempts or if no pages
 983 * are movable anymore because to has become empty
 984 * or no retryable pages exist anymore.
 985 * Caller should call putback_lru_pages to return pages to the LRU
 986 * or free list only if ret != 0.
 987 *
 988 * Return: Number of pages not migrated or error code.
 989 */
 990int migrate_pages(struct list_head *from, new_page_t get_new_page,
 991                unsigned long private, enum migrate_mode mode, int reason)
 992{
 993        int retry = 1;
 994        int nr_failed = 0;
 995        int nr_succeeded = 0;
 996        int pass = 0;
 997        struct page *page;
 998        struct page *page2;
 999        int swapwrite = current->flags & PF_SWAPWRITE;
1000        int rc;
1001
1002        if (!swapwrite)
1003                current->flags |= PF_SWAPWRITE;
1004
1005        for(pass = 0; pass < 10 && retry; pass++) {
1006                retry = 0;
1007
1008                list_for_each_entry_safe(page, page2, from, lru) {
1009                        cond_resched();
1010
1011                        rc = unmap_and_move(get_new_page, private,
1012                                                page, pass > 2, mode);
1013
1014                        switch(rc) {
1015                        case -ENOMEM:
1016                                goto out;
1017                        case -EAGAIN:
1018                                retry++;
1019                                break;
1020                        case MIGRATEPAGE_SUCCESS:
1021                                nr_succeeded++;
1022                                break;
1023                        default:
1024                                /* Permanent failure */
1025                                nr_failed++;
1026                                break;
1027                        }
1028                }
1029        }
1030        rc = nr_failed + retry;
1031out:
1032        if (nr_succeeded)
1033                count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1034        if (nr_failed)
1035                count_vm_events(PGMIGRATE_FAIL, nr_failed);
1036        trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1037
1038        if (!swapwrite)
1039                current->flags &= ~PF_SWAPWRITE;
1040
1041        return rc;
1042}
1043
1044int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1045                      unsigned long private, enum migrate_mode mode)
1046{
1047        int pass, rc;
1048
1049        for (pass = 0; pass < 10; pass++) {
1050                rc = unmap_and_move_huge_page(get_new_page, private,
1051                                                hpage, pass > 2, mode);
1052                switch (rc) {
1053                case -ENOMEM:
1054                        goto out;
1055                case -EAGAIN:
1056                        /* try again */
1057                        cond_resched();
1058                        break;
1059                case MIGRATEPAGE_SUCCESS:
1060                        goto out;
1061                default:
1062                        rc = -EIO;
1063                        goto out;
1064                }
1065        }
1066out:
1067        return rc;
1068}
1069
1070#ifdef CONFIG_NUMA
1071/*
1072 * Move a list of individual pages
1073 */
1074struct page_to_node {
1075        unsigned long addr;
1076        struct page *page;
1077        int node;
1078        int status;
1079};
1080
1081static struct page *new_page_node(struct page *p, unsigned long private,
1082                int **result)
1083{
1084        struct page_to_node *pm = (struct page_to_node *)private;
1085
1086        while (pm->node != MAX_NUMNODES && pm->page != p)
1087                pm++;
1088
1089        if (pm->node == MAX_NUMNODES)
1090                return NULL;
1091
1092        *result = &pm->status;
1093
1094        return alloc_pages_exact_node(pm->node,
1095                                GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1096}
1097
1098/*
1099 * Move a set of pages as indicated in the pm array. The addr
1100 * field must be set to the virtual address of the page to be moved
1101 * and the node number must contain a valid target node.
1102 * The pm array ends with node = MAX_NUMNODES.
1103 */
1104static int do_move_page_to_node_array(struct mm_struct *mm,
1105                                      struct page_to_node *pm,
1106                                      int migrate_all)
1107{
1108        int err;
1109        struct page_to_node *pp;
1110        LIST_HEAD(pagelist);
1111
1112        down_read(&mm->mmap_sem);
1113
1114        /*
1115         * Build a list of pages to migrate
1116         */
1117        for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1118                struct vm_area_struct *vma;
1119                struct page *page;
1120
1121                err = -EFAULT;
1122                vma = find_vma(mm, pp->addr);
1123                if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1124                        goto set_status;
1125
1126                page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1127
1128                err = PTR_ERR(page);
1129                if (IS_ERR(page))
1130                        goto set_status;
1131
1132                err = -ENOENT;
1133                if (!page)
1134                        goto set_status;
1135
1136                /* Use PageReserved to check for zero page */
1137                if (PageReserved(page))
1138                        goto put_and_set;
1139
1140                pp->page = page;
1141                err = page_to_nid(page);
1142
1143                if (err == pp->node)
1144                        /*
1145                         * Node already in the right place
1146                         */
1147                        goto put_and_set;
1148
1149                err = -EACCES;
1150                if (page_mapcount(page) > 1 &&
1151                                !migrate_all)
1152                        goto put_and_set;
1153
1154                err = isolate_lru_page(page);
1155                if (!err) {
1156                        list_add_tail(&page->lru, &pagelist);
1157                        inc_zone_page_state(page, NR_ISOLATED_ANON +
1158                                            page_is_file_cache(page));
1159                }
1160put_and_set:
1161                /*
1162                 * Either remove the duplicate refcount from
1163                 * isolate_lru_page() or drop the page ref if it was
1164                 * not isolated.
1165                 */
1166                put_page(page);
1167set_status:
1168                pp->status = err;
1169        }
1170
1171        err = 0;
1172        if (!list_empty(&pagelist)) {
1173                err = migrate_pages(&pagelist, new_page_node,
1174                                (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1175                if (err)
1176                        putback_lru_pages(&pagelist);
1177        }
1178
1179        up_read(&mm->mmap_sem);
1180        return err;
1181}
1182
1183/*
1184 * Migrate an array of page address onto an array of nodes and fill
1185 * the corresponding array of status.
1186 */
1187static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1188                         unsigned long nr_pages,
1189                         const void __user * __user *pages,
1190                         const int __user *nodes,
1191                         int __user *status, int flags)
1192{
1193        struct page_to_node *pm;
1194        unsigned long chunk_nr_pages;
1195        unsigned long chunk_start;
1196        int err;
1197
1198        err = -ENOMEM;
1199        pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1200        if (!pm)
1201                goto out;
1202
1203        migrate_prep();
1204
1205        /*
1206         * Store a chunk of page_to_node array in a page,
1207         * but keep the last one as a marker
1208         */
1209        chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1210
1211        for (chunk_start = 0;
1212             chunk_start < nr_pages;
1213             chunk_start += chunk_nr_pages) {
1214                int j;
1215
1216                if (chunk_start + chunk_nr_pages > nr_pages)
1217                        chunk_nr_pages = nr_pages - chunk_start;
1218
1219                /* fill the chunk pm with addrs and nodes from user-space */
1220                for (j = 0; j < chunk_nr_pages; j++) {
1221                        const void __user *p;
1222                        int node;
1223
1224                        err = -EFAULT;
1225                        if (get_user(p, pages + j + chunk_start))
1226                                goto out_pm;
1227                        pm[j].addr = (unsigned long) p;
1228
1229                        if (get_user(node, nodes + j + chunk_start))
1230                                goto out_pm;
1231
1232                        err = -ENODEV;
1233                        if (node < 0 || node >= MAX_NUMNODES)
1234                                goto out_pm;
1235
1236                        if (!node_state(node, N_MEMORY))
1237                                goto out_pm;
1238
1239                        err = -EACCES;
1240                        if (!node_isset(node, task_nodes))
1241                                goto out_pm;
1242
1243                        pm[j].node = node;
1244                }
1245
1246                /* End marker for this chunk */
1247                pm[chunk_nr_pages].node = MAX_NUMNODES;
1248
1249                /* Migrate this chunk */
1250                err = do_move_page_to_node_array(mm, pm,
1251                                                 flags & MPOL_MF_MOVE_ALL);
1252                if (err < 0)
1253                        goto out_pm;
1254
1255                /* Return status information */
1256                for (j = 0; j < chunk_nr_pages; j++)
1257                        if (put_user(pm[j].status, status + j + chunk_start)) {
1258                                err = -EFAULT;
1259                                goto out_pm;
1260                        }
1261        }
1262        err = 0;
1263
1264out_pm:
1265        free_page((unsigned long)pm);
1266out:
1267        return err;
1268}
1269
1270/*
1271 * Determine the nodes of an array of pages and store it in an array of status.
1272 */
1273static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1274                                const void __user **pages, int *status)
1275{
1276        unsigned long i;
1277
1278        down_read(&mm->mmap_sem);
1279
1280        for (i = 0; i < nr_pages; i++) {
1281                unsigned long addr = (unsigned long)(*pages);
1282                struct vm_area_struct *vma;
1283                struct page *page;
1284                int err = -EFAULT;
1285
1286                vma = find_vma(mm, addr);
1287                if (!vma || addr < vma->vm_start)
1288                        goto set_status;
1289
1290                page = follow_page(vma, addr, 0);
1291
1292                err = PTR_ERR(page);
1293                if (IS_ERR(page))
1294                        goto set_status;
1295
1296                err = -ENOENT;
1297                /* Use PageReserved to check for zero page */
1298                if (!page || PageReserved(page))
1299                        goto set_status;
1300
1301                err = page_to_nid(page);
1302set_status:
1303                *status = err;
1304
1305                pages++;
1306                status++;
1307        }
1308
1309        up_read(&mm->mmap_sem);
1310}
1311
1312/*
1313 * Determine the nodes of a user array of pages and store it in
1314 * a user array of status.
1315 */
1316static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1317                         const void __user * __user *pages,
1318                         int __user *status)
1319{
1320#define DO_PAGES_STAT_CHUNK_NR 16
1321        const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1322        int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1323
1324        while (nr_pages) {
1325                unsigned long chunk_nr;
1326
1327                chunk_nr = nr_pages;
1328                if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1329                        chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1330
1331                if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1332                        break;
1333
1334                do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1335
1336                if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1337                        break;
1338
1339                pages += chunk_nr;
1340                status += chunk_nr;
1341                nr_pages -= chunk_nr;
1342        }
1343        return nr_pages ? -EFAULT : 0;
1344}
1345
1346/*
1347 * Move a list of pages in the address space of the currently executing
1348 * process.
1349 */
1350SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1351                const void __user * __user *, pages,
1352                const int __user *, nodes,
1353                int __user *, status, int, flags)
1354{
1355        const struct cred *cred = current_cred(), *tcred;
1356        struct task_struct *task;
1357        struct mm_struct *mm;
1358        int err;
1359        nodemask_t task_nodes;
1360
1361        /* Check flags */
1362        if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1363                return -EINVAL;
1364
1365        if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1366                return -EPERM;
1367
1368        /* Find the mm_struct */
1369        rcu_read_lock();
1370        task = pid ? find_task_by_vpid(pid) : current;
1371        if (!task) {
1372                rcu_read_unlock();
1373                return -ESRCH;
1374        }
1375        get_task_struct(task);
1376
1377        /*
1378         * Check if this process has the right to modify the specified
1379         * process. The right exists if the process has administrative
1380         * capabilities, superuser privileges or the same
1381         * userid as the target process.
1382         */
1383        tcred = __task_cred(task);
1384        if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1385            !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1386            !capable(CAP_SYS_NICE)) {
1387                rcu_read_unlock();
1388                err = -EPERM;
1389                goto out;
1390        }
1391        rcu_read_unlock();
1392
1393        err = security_task_movememory(task);
1394        if (err)
1395                goto out;
1396
1397        task_nodes = cpuset_mems_allowed(task);
1398        mm = get_task_mm(task);
1399        put_task_struct(task);
1400
1401        if (!mm)
1402                return -EINVAL;
1403
1404        if (nodes)
1405                err = do_pages_move(mm, task_nodes, nr_pages, pages,
1406                                    nodes, status, flags);
1407        else
1408                err = do_pages_stat(mm, nr_pages, pages, status);
1409
1410        mmput(mm);
1411        return err;
1412
1413out:
1414        put_task_struct(task);
1415        return err;
1416}
1417
1418/*
1419 * Call migration functions in the vma_ops that may prepare
1420 * memory in a vm for migration. migration functions may perform
1421 * the migration for vmas that do not have an underlying page struct.
1422 */
1423int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1424        const nodemask_t *from, unsigned long flags)
1425{
1426        struct vm_area_struct *vma;
1427        int err = 0;
1428
1429        for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1430                if (vma->vm_ops && vma->vm_ops->migrate) {
1431                        err = vma->vm_ops->migrate(vma, to, from, flags);
1432                        if (err)
1433                                break;
1434                }
1435        }
1436        return err;
1437}
1438
1439#ifdef CONFIG_NUMA_BALANCING
1440/*
1441 * Returns true if this is a safe migration target node for misplaced NUMA
1442 * pages. Currently it only checks the watermarks which crude
1443 */
1444static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1445                                   unsigned long nr_migrate_pages)
1446{
1447        int z;
1448        for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1449                struct zone *zone = pgdat->node_zones + z;
1450
1451                if (!populated_zone(zone))
1452                        continue;
1453
1454                if (zone->all_unreclaimable)
1455                        continue;
1456
1457                /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1458                if (!zone_watermark_ok(zone, 0,
1459                                       high_wmark_pages(zone) +
1460                                       nr_migrate_pages,
1461                                       0, 0))
1462                        continue;
1463                return true;
1464        }
1465        return false;
1466}
1467
1468static struct page *alloc_misplaced_dst_page(struct page *page,
1469                                           unsigned long data,
1470                                           int **result)
1471{
1472        int nid = (int) data;
1473        struct page *newpage;
1474
1475        newpage = alloc_pages_exact_node(nid,
1476                                         (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1477                                          __GFP_NOMEMALLOC | __GFP_NORETRY |
1478                                          __GFP_NOWARN) &
1479                                         ~GFP_IOFS, 0);
1480        if (newpage)
1481                page_nid_xchg_last(newpage, page_nid_last(page));
1482
1483        return newpage;
1484}
1485
1486/*
1487 * page migration rate limiting control.
1488 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1489 * window of time. Default here says do not migrate more than 1280M per second.
1490 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1491 * as it is faults that reset the window, pte updates will happen unconditionally
1492 * if there has not been a fault since @pteupdate_interval_millisecs after the
1493 * throttle window closed.
1494 */
1495static unsigned int migrate_interval_millisecs __read_mostly = 100;
1496static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1497static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1498
1499/* Returns true if NUMA migration is currently rate limited */
1500bool migrate_ratelimited(int node)
1501{
1502        pg_data_t *pgdat = NODE_DATA(node);
1503
1504        if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1505                                msecs_to_jiffies(pteupdate_interval_millisecs)))
1506                return false;
1507
1508        if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1509                return false;
1510
1511        return true;
1512}
1513
1514/* Returns true if the node is migrate rate-limited after the update */
1515bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
1516{
1517        bool rate_limited = false;
1518
1519        /*
1520         * Rate-limit the amount of data that is being migrated to a node.
1521         * Optimal placement is no good if the memory bus is saturated and
1522         * all the time is being spent migrating!
1523         */
1524        spin_lock(&pgdat->numabalancing_migrate_lock);
1525        if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1526                pgdat->numabalancing_migrate_nr_pages = 0;
1527                pgdat->numabalancing_migrate_next_window = jiffies +
1528                        msecs_to_jiffies(migrate_interval_millisecs);
1529        }
1530        if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1531                rate_limited = true;
1532        else
1533                pgdat->numabalancing_migrate_nr_pages += nr_pages;
1534        spin_unlock(&pgdat->numabalancing_migrate_lock);
1535        
1536        return rate_limited;
1537}
1538
1539int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1540{
1541        int page_lru;
1542
1543        VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1544
1545        /* Avoid migrating to a node that is nearly full */
1546        if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1547                return 0;
1548
1549        if (isolate_lru_page(page))
1550                return 0;
1551
1552        /*
1553         * migrate_misplaced_transhuge_page() skips page migration's usual
1554         * check on page_count(), so we must do it here, now that the page
1555         * has been isolated: a GUP pin, or any other pin, prevents migration.
1556         * The expected page count is 3: 1 for page's mapcount and 1 for the
1557         * caller's pin and 1 for the reference taken by isolate_lru_page().
1558         */
1559        if (PageTransHuge(page) && page_count(page) != 3) {
1560                putback_lru_page(page);
1561                return 0;
1562        }
1563
1564        page_lru = page_is_file_cache(page);
1565        mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1566                                hpage_nr_pages(page));
1567
1568        /*
1569         * Isolating the page has taken another reference, so the
1570         * caller's reference can be safely dropped without the page
1571         * disappearing underneath us during migration.
1572         */
1573        put_page(page);
1574        return 1;
1575}
1576
1577/*
1578 * Attempt to migrate a misplaced page to the specified destination
1579 * node. Caller is expected to have an elevated reference count on
1580 * the page that will be dropped by this function before returning.
1581 */
1582int migrate_misplaced_page(struct page *page, int node)
1583{
1584        pg_data_t *pgdat = NODE_DATA(node);
1585        int isolated;
1586        int nr_remaining;
1587        LIST_HEAD(migratepages);
1588
1589        /*
1590         * Don't migrate pages that are mapped in multiple processes.
1591         * TODO: Handle false sharing detection instead of this hammer
1592         */
1593        if (page_mapcount(page) != 1)
1594                goto out;
1595
1596        /*
1597         * Rate-limit the amount of data that is being migrated to a node.
1598         * Optimal placement is no good if the memory bus is saturated and
1599         * all the time is being spent migrating!
1600         */
1601        if (numamigrate_update_ratelimit(pgdat, 1))
1602                goto out;
1603
1604        isolated = numamigrate_isolate_page(pgdat, page);
1605        if (!isolated)
1606                goto out;
1607
1608        list_add(&page->lru, &migratepages);
1609        nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1610                                     node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1611        if (nr_remaining) {
1612                putback_lru_pages(&migratepages);
1613                isolated = 0;
1614        } else
1615                count_vm_numa_event(NUMA_PAGE_MIGRATE);
1616        BUG_ON(!list_empty(&migratepages));
1617        return isolated;
1618
1619out:
1620        put_page(page);
1621        return 0;
1622}
1623#endif /* CONFIG_NUMA_BALANCING */
1624
1625#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1626/*
1627 * Migrates a THP to a given target node. page must be locked and is unlocked
1628 * before returning.
1629 */
1630int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1631                                struct vm_area_struct *vma,
1632                                pmd_t *pmd, pmd_t entry,
1633                                unsigned long address,
1634                                struct page *page, int node)
1635{
1636        unsigned long haddr = address & HPAGE_PMD_MASK;
1637        pg_data_t *pgdat = NODE_DATA(node);
1638        int isolated = 0;
1639        struct page *new_page = NULL;
1640        struct mem_cgroup *memcg = NULL;
1641        int page_lru = page_is_file_cache(page);
1642
1643        /*
1644         * Don't migrate pages that are mapped in multiple processes.
1645         * TODO: Handle false sharing detection instead of this hammer
1646         */
1647        if (page_mapcount(page) != 1)
1648                goto out_dropref;
1649
1650        /*
1651         * Rate-limit the amount of data that is being migrated to a node.
1652         * Optimal placement is no good if the memory bus is saturated and
1653         * all the time is being spent migrating!
1654         */
1655        if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1656                goto out_dropref;
1657
1658        new_page = alloc_pages_node(node,
1659                (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1660        if (!new_page)
1661                goto out_fail;
1662
1663        page_nid_xchg_last(new_page, page_nid_last(page));
1664
1665        isolated = numamigrate_isolate_page(pgdat, page);
1666        if (!isolated) {
1667                put_page(new_page);
1668                goto out_fail;
1669        }
1670
1671        /* Prepare a page as a migration target */
1672        __set_page_locked(new_page);
1673        SetPageSwapBacked(new_page);
1674
1675        /* anon mapping, we can simply copy page->mapping to the new page: */
1676        new_page->mapping = page->mapping;
1677        new_page->index = page->index;
1678        migrate_page_copy(new_page, page);
1679        WARN_ON(PageLRU(new_page));
1680
1681        /* Recheck the target PMD */
1682        spin_lock(&mm->page_table_lock);
1683        if (unlikely(!pmd_same(*pmd, entry))) {
1684                spin_unlock(&mm->page_table_lock);
1685
1686                /* Reverse changes made by migrate_page_copy() */
1687                if (TestClearPageActive(new_page))
1688                        SetPageActive(page);
1689                if (TestClearPageUnevictable(new_page))
1690                        SetPageUnevictable(page);
1691                mlock_migrate_page(page, new_page);
1692
1693                unlock_page(new_page);
1694                put_page(new_page);             /* Free it */
1695
1696                unlock_page(page);
1697                putback_lru_page(page);
1698
1699                count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1700                isolated = 0;
1701                goto out;
1702        }
1703
1704        /*
1705         * Traditional migration needs to prepare the memcg charge
1706         * transaction early to prevent the old page from being
1707         * uncharged when installing migration entries.  Here we can
1708         * save the potential rollback and start the charge transfer
1709         * only when migration is already known to end successfully.
1710         */
1711        mem_cgroup_prepare_migration(page, new_page, &memcg);
1712
1713        entry = mk_pmd(new_page, vma->vm_page_prot);
1714        entry = pmd_mknonnuma(entry);
1715        entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1716        entry = pmd_mkhuge(entry);
1717
1718        page_add_new_anon_rmap(new_page, vma, haddr);
1719
1720        set_pmd_at(mm, haddr, pmd, entry);
1721        update_mmu_cache_pmd(vma, address, &entry);
1722        page_remove_rmap(page);
1723        /*
1724         * Finish the charge transaction under the page table lock to
1725         * prevent split_huge_page() from dividing up the charge
1726         * before it's fully transferred to the new page.
1727         */
1728        mem_cgroup_end_migration(memcg, page, new_page, true);
1729        spin_unlock(&mm->page_table_lock);
1730
1731        unlock_page(new_page);
1732        unlock_page(page);
1733        put_page(page);                 /* Drop the rmap reference */
1734        put_page(page);                 /* Drop the LRU isolation reference */
1735
1736        count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1737        count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1738
1739out:
1740        mod_zone_page_state(page_zone(page),
1741                        NR_ISOLATED_ANON + page_lru,
1742                        -HPAGE_PMD_NR);
1743        return isolated;
1744
1745out_fail:
1746        count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1747out_dropref:
1748        unlock_page(page);
1749        put_page(page);
1750        return 0;
1751}
1752#endif /* CONFIG_NUMA_BALANCING */
1753
1754#endif /* CONFIG_NUMA */
1755
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.