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