linux/mm/page_alloc.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/page_alloc.c
   3 *
   4 *  Manages the free list, the system allocates free pages here.
   5 *  Note that kmalloc() lives in slab.c
   6 *
   7 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   8 *  Swap reorganised 29.12.95, Stephen Tweedie
   9 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
  10 *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
  11 *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
  12 *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  13 *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  14 *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
  15 */
  16
  17#include <linux/stddef.h>
  18#include <linux/mm.h>
  19#include <linux/swap.h>
  20#include <linux/interrupt.h>
  21#include <linux/pagemap.h>
  22#include <linux/jiffies.h>
  23#include <linux/bootmem.h>
  24#include <linux/memblock.h>
  25#include <linux/compiler.h>
  26#include <linux/kernel.h>
  27#include <linux/kmemcheck.h>
  28#include <linux/module.h>
  29#include <linux/suspend.h>
  30#include <linux/pagevec.h>
  31#include <linux/blkdev.h>
  32#include <linux/slab.h>
  33#include <linux/ratelimit.h>
  34#include <linux/oom.h>
  35#include <linux/notifier.h>
  36#include <linux/topology.h>
  37#include <linux/sysctl.h>
  38#include <linux/cpu.h>
  39#include <linux/cpuset.h>
  40#include <linux/memory_hotplug.h>
  41#include <linux/nodemask.h>
  42#include <linux/vmalloc.h>
  43#include <linux/vmstat.h>
  44#include <linux/mempolicy.h>
  45#include <linux/stop_machine.h>
  46#include <linux/sort.h>
  47#include <linux/pfn.h>
  48#include <linux/backing-dev.h>
  49#include <linux/fault-inject.h>
  50#include <linux/page-isolation.h>
  51#include <linux/page_cgroup.h>
  52#include <linux/debugobjects.h>
  53#include <linux/kmemleak.h>
  54#include <linux/compaction.h>
  55#include <trace/events/kmem.h>
  56#include <linux/ftrace_event.h>
  57#include <linux/memcontrol.h>
  58#include <linux/prefetch.h>
  59#include <linux/migrate.h>
  60#include <linux/page-debug-flags.h>
  61
  62#include <asm/tlbflush.h>
  63#include <asm/div64.h>
  64#include "internal.h"
  65
  66#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
  67DEFINE_PER_CPU(int, numa_node);
  68EXPORT_PER_CPU_SYMBOL(numa_node);
  69#endif
  70
  71#ifdef CONFIG_HAVE_MEMORYLESS_NODES
  72/*
  73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
  74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
  75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
  76 * defined in <linux/topology.h>.
  77 */
  78DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
  79EXPORT_PER_CPU_SYMBOL(_numa_mem_);
  80#endif
  81
  82/*
  83 * Array of node states.
  84 */
  85nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
  86        [N_POSSIBLE] = NODE_MASK_ALL,
  87        [N_ONLINE] = { { [0] = 1UL } },
  88#ifndef CONFIG_NUMA
  89        [N_NORMAL_MEMORY] = { { [0] = 1UL } },
  90#ifdef CONFIG_HIGHMEM
  91        [N_HIGH_MEMORY] = { { [0] = 1UL } },
  92#endif
  93        [N_CPU] = { { [0] = 1UL } },
  94#endif  /* NUMA */
  95};
  96EXPORT_SYMBOL(node_states);
  97
  98unsigned long totalram_pages __read_mostly;
  99unsigned long totalreserve_pages __read_mostly;
 100/*
 101 * When calculating the number of globally allowed dirty pages, there
 102 * is a certain number of per-zone reserves that should not be
 103 * considered dirtyable memory.  This is the sum of those reserves
 104 * over all existing zones that contribute dirtyable memory.
 105 */
 106unsigned long dirty_balance_reserve __read_mostly;
 107
 108int percpu_pagelist_fraction;
 109gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
 110
 111#ifdef CONFIG_PM_SLEEP
 112/*
 113 * The following functions are used by the suspend/hibernate code to temporarily
 114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
 115 * while devices are suspended.  To avoid races with the suspend/hibernate code,
 116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
 117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
 118 * guaranteed not to run in parallel with that modification).
 119 */
 120
 121static gfp_t saved_gfp_mask;
 122
 123void pm_restore_gfp_mask(void)
 124{
 125        WARN_ON(!mutex_is_locked(&pm_mutex));
 126        if (saved_gfp_mask) {
 127                gfp_allowed_mask = saved_gfp_mask;
 128                saved_gfp_mask = 0;
 129        }
 130}
 131
 132void pm_restrict_gfp_mask(void)
 133{
 134        WARN_ON(!mutex_is_locked(&pm_mutex));
 135        WARN_ON(saved_gfp_mask);
 136        saved_gfp_mask = gfp_allowed_mask;
 137        gfp_allowed_mask &= ~GFP_IOFS;
 138}
 139
 140bool pm_suspended_storage(void)
 141{
 142        if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
 143                return false;
 144        return true;
 145}
 146#endif /* CONFIG_PM_SLEEP */
 147
 148#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 149int pageblock_order __read_mostly;
 150#endif
 151
 152static void __free_pages_ok(struct page *page, unsigned int order);
 153
 154/*
 155 * results with 256, 32 in the lowmem_reserve sysctl:
 156 *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 157 *      1G machine -> (16M dma, 784M normal, 224M high)
 158 *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 159 *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 160 *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
 161 *
 162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 163 * don't need any ZONE_NORMAL reservation
 164 */
 165int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
 166#ifdef CONFIG_ZONE_DMA
 167         256,
 168#endif
 169#ifdef CONFIG_ZONE_DMA32
 170         256,
 171#endif
 172#ifdef CONFIG_HIGHMEM
 173         32,
 174#endif
 175         32,
 176};
 177
 178EXPORT_SYMBOL(totalram_pages);
 179
 180static char * const zone_names[MAX_NR_ZONES] = {
 181#ifdef CONFIG_ZONE_DMA
 182         "DMA",
 183#endif
 184#ifdef CONFIG_ZONE_DMA32
 185         "DMA32",
 186#endif
 187         "Normal",
 188#ifdef CONFIG_HIGHMEM
 189         "HighMem",
 190#endif
 191         "Movable",
 192};
 193
 194int min_free_kbytes = 1024;
 195
 196static unsigned long __meminitdata nr_kernel_pages;
 197static unsigned long __meminitdata nr_all_pages;
 198static unsigned long __meminitdata dma_reserve;
 199
 200#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
 201static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
 202static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
 203static unsigned long __initdata required_kernelcore;
 204static unsigned long __initdata required_movablecore;
 205static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
 206
 207/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 208int movable_zone;
 209EXPORT_SYMBOL(movable_zone);
 210#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
 211
 212#if MAX_NUMNODES > 1
 213int nr_node_ids __read_mostly = MAX_NUMNODES;
 214int nr_online_nodes __read_mostly = 1;
 215EXPORT_SYMBOL(nr_node_ids);
 216EXPORT_SYMBOL(nr_online_nodes);
 217#endif
 218
 219int page_group_by_mobility_disabled __read_mostly;
 220
 221/*
 222 * NOTE:
 223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
 224 * Instead, use {un}set_pageblock_isolate.
 225 */
 226void set_pageblock_migratetype(struct page *page, int migratetype)
 227{
 228
 229        if (unlikely(page_group_by_mobility_disabled))
 230                migratetype = MIGRATE_UNMOVABLE;
 231
 232        set_pageblock_flags_group(page, (unsigned long)migratetype,
 233                                        PB_migrate, PB_migrate_end);
 234}
 235
 236bool oom_killer_disabled __read_mostly;
 237
 238#ifdef CONFIG_DEBUG_VM
 239static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 240{
 241        int ret = 0;
 242        unsigned seq;
 243        unsigned long pfn = page_to_pfn(page);
 244
 245        do {
 246                seq = zone_span_seqbegin(zone);
 247                if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
 248                        ret = 1;
 249                else if (pfn < zone->zone_start_pfn)
 250                        ret = 1;
 251        } while (zone_span_seqretry(zone, seq));
 252
 253        return ret;
 254}
 255
 256static int page_is_consistent(struct zone *zone, struct page *page)
 257{
 258        if (!pfn_valid_within(page_to_pfn(page)))
 259                return 0;
 260        if (zone != page_zone(page))
 261                return 0;
 262
 263        return 1;
 264}
 265/*
 266 * Temporary debugging check for pages not lying within a given zone.
 267 */
 268static int bad_range(struct zone *zone, struct page *page)
 269{
 270        if (page_outside_zone_boundaries(zone, page))
 271                return 1;
 272        if (!page_is_consistent(zone, page))
 273                return 1;
 274
 275        return 0;
 276}
 277#else
 278static inline int bad_range(struct zone *zone, struct page *page)
 279{
 280        return 0;
 281}
 282#endif
 283
 284static void bad_page(struct page *page)
 285{
 286        static unsigned long resume;
 287        static unsigned long nr_shown;
 288        static unsigned long nr_unshown;
 289
 290        /* Don't complain about poisoned pages */
 291        if (PageHWPoison(page)) {
 292                reset_page_mapcount(page); /* remove PageBuddy */
 293                return;
 294        }
 295
 296        /*
 297         * Allow a burst of 60 reports, then keep quiet for that minute;
 298         * or allow a steady drip of one report per second.
 299         */
 300        if (nr_shown == 60) {
 301                if (time_before(jiffies, resume)) {
 302                        nr_unshown++;
 303                        goto out;
 304                }
 305                if (nr_unshown) {
 306                        printk(KERN_ALERT
 307                              "BUG: Bad page state: %lu messages suppressed\n",
 308                                nr_unshown);
 309                        nr_unshown = 0;
 310                }
 311                nr_shown = 0;
 312        }
 313        if (nr_shown++ == 0)
 314                resume = jiffies + 60 * HZ;
 315
 316        printk(KERN_ALERT "BUG: Bad page state in process %s  pfn:%05lx\n",
 317                current->comm, page_to_pfn(page));
 318        dump_page(page);
 319
 320        print_modules();
 321        dump_stack();
 322out:
 323        /* Leave bad fields for debug, except PageBuddy could make trouble */
 324        reset_page_mapcount(page); /* remove PageBuddy */
 325        add_taint(TAINT_BAD_PAGE);
 326}
 327
 328/*
 329 * Higher-order pages are called "compound pages".  They are structured thusly:
 330 *
 331 * The first PAGE_SIZE page is called the "head page".
 332 *
 333 * The remaining PAGE_SIZE pages are called "tail pages".
 334 *
 335 * All pages have PG_compound set.  All tail pages have their ->first_page
 336 * pointing at the head page.
 337 *
 338 * The first tail page's ->lru.next holds the address of the compound page's
 339 * put_page() function.  Its ->lru.prev holds the order of allocation.
 340 * This usage means that zero-order pages may not be compound.
 341 */
 342
 343static void free_compound_page(struct page *page)
 344{
 345        __free_pages_ok(page, compound_order(page));
 346}
 347
 348void prep_compound_page(struct page *page, unsigned long order)
 349{
 350        int i;
 351        int nr_pages = 1 << order;
 352
 353        set_compound_page_dtor(page, free_compound_page);
 354        set_compound_order(page, order);
 355        __SetPageHead(page);
 356        for (i = 1; i < nr_pages; i++) {
 357                struct page *p = page + i;
 358                __SetPageTail(p);
 359                set_page_count(p, 0);
 360                p->first_page = page;
 361        }
 362}
 363
 364/* update __split_huge_page_refcount if you change this function */
 365static int destroy_compound_page(struct page *page, unsigned long order)
 366{
 367        int i;
 368        int nr_pages = 1 << order;
 369        int bad = 0;
 370
 371        if (unlikely(compound_order(page) != order) ||
 372            unlikely(!PageHead(page))) {
 373                bad_page(page);
 374                bad++;
 375        }
 376
 377        __ClearPageHead(page);
 378
 379        for (i = 1; i < nr_pages; i++) {
 380                struct page *p = page + i;
 381
 382                if (unlikely(!PageTail(p) || (p->first_page != page))) {
 383                        bad_page(page);
 384                        bad++;
 385                }
 386                __ClearPageTail(p);
 387        }
 388
 389        return bad;
 390}
 391
 392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
 393{
 394        int i;
 395
 396        /*
 397         * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
 398         * and __GFP_HIGHMEM from hard or soft interrupt context.
 399         */
 400        VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
 401        for (i = 0; i < (1 << order); i++)
 402                clear_highpage(page + i);
 403}
 404
 405#ifdef CONFIG_DEBUG_PAGEALLOC
 406unsigned int _debug_guardpage_minorder;
 407
 408static int __init debug_guardpage_minorder_setup(char *buf)
 409{
 410        unsigned long res;
 411
 412        if (kstrtoul(buf, 10, &res) < 0 ||  res > MAX_ORDER / 2) {
 413                printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
 414                return 0;
 415        }
 416        _debug_guardpage_minorder = res;
 417        printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
 418        return 0;
 419}
 420__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
 421
 422static inline void set_page_guard_flag(struct page *page)
 423{
 424        __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
 425}
 426
 427static inline void clear_page_guard_flag(struct page *page)
 428{
 429        __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
 430}
 431#else
 432static inline void set_page_guard_flag(struct page *page) { }
 433static inline void clear_page_guard_flag(struct page *page) { }
 434#endif
 435
 436static inline void set_page_order(struct page *page, int order)
 437{
 438        set_page_private(page, order);
 439        __SetPageBuddy(page);
 440}
 441
 442static inline void rmv_page_order(struct page *page)
 443{
 444        __ClearPageBuddy(page);
 445        set_page_private(page, 0);
 446}
 447
 448/*
 449 * Locate the struct page for both the matching buddy in our
 450 * pair (buddy1) and the combined O(n+1) page they form (page).
 451 *
 452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
 453 * the following equation:
 454 *     B2 = B1 ^ (1 << O)
 455 * For example, if the starting buddy (buddy2) is #8 its order
 456 * 1 buddy is #10:
 457 *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 458 *
 459 * 2) Any buddy B will have an order O+1 parent P which
 460 * satisfies the following equation:
 461 *     P = B & ~(1 << O)
 462 *
 463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
 464 */
 465static inline unsigned long
 466__find_buddy_index(unsigned long page_idx, unsigned int order)
 467{
 468        return page_idx ^ (1 << order);
 469}
 470
 471/*
 472 * This function checks whether a page is free && is the buddy
 473 * we can do coalesce a page and its buddy if
 474 * (a) the buddy is not in a hole &&
 475 * (b) the buddy is in the buddy system &&
 476 * (c) a page and its buddy have the same order &&
 477 * (d) a page and its buddy are in the same zone.
 478 *
 479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
 480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
 481 *
 482 * For recording page's order, we use page_private(page).
 483 */
 484static inline int page_is_buddy(struct page *page, struct page *buddy,
 485                                                                int order)
 486{
 487        if (!pfn_valid_within(page_to_pfn(buddy)))
 488                return 0;
 489
 490        if (page_zone_id(page) != page_zone_id(buddy))
 491                return 0;
 492
 493        if (page_is_guard(buddy) && page_order(buddy) == order) {
 494                VM_BUG_ON(page_count(buddy) != 0);
 495                return 1;
 496        }
 497
 498        if (PageBuddy(buddy) && page_order(buddy) == order) {
 499                VM_BUG_ON(page_count(buddy) != 0);
 500                return 1;
 501        }
 502        return 0;
 503}
 504
 505/*
 506 * Freeing function for a buddy system allocator.
 507 *
 508 * The concept of a buddy system is to maintain direct-mapped table
 509 * (containing bit values) for memory blocks of various "orders".
 510 * The bottom level table contains the map for the smallest allocatable
 511 * units of memory (here, pages), and each level above it describes
 512 * pairs of units from the levels below, hence, "buddies".
 513 * At a high level, all that happens here is marking the table entry
 514 * at the bottom level available, and propagating the changes upward
 515 * as necessary, plus some accounting needed to play nicely with other
 516 * parts of the VM system.
 517 * At each level, we keep a list of pages, which are heads of continuous
 518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
 519 * order is recorded in page_private(page) field.
 520 * So when we are allocating or freeing one, we can derive the state of the
 521 * other.  That is, if we allocate a small block, and both were
 522 * free, the remainder of the region must be split into blocks.
 523 * If a block is freed, and its buddy is also free, then this
 524 * triggers coalescing into a block of larger size.
 525 *
 526 * -- wli
 527 */
 528
 529static inline void __free_one_page(struct page *page,
 530                struct zone *zone, unsigned int order,
 531                int migratetype)
 532{
 533        unsigned long page_idx;
 534        unsigned long combined_idx;
 535        unsigned long uninitialized_var(buddy_idx);
 536        struct page *buddy;
 537
 538        if (unlikely(PageCompound(page)))
 539                if (unlikely(destroy_compound_page(page, order)))
 540                        return;
 541
 542        VM_BUG_ON(migratetype == -1);
 543
 544        page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 545
 546        VM_BUG_ON(page_idx & ((1 << order) - 1));
 547        VM_BUG_ON(bad_range(zone, page));
 548
 549        while (order < MAX_ORDER-1) {
 550                buddy_idx = __find_buddy_index(page_idx, order);
 551                buddy = page + (buddy_idx - page_idx);
 552                if (!page_is_buddy(page, buddy, order))
 553                        break;
 554                /*
 555                 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
 556                 * merge with it and move up one order.
 557                 */
 558                if (page_is_guard(buddy)) {
 559                        clear_page_guard_flag(buddy);
 560                        set_page_private(page, 0);
 561                        __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 562                } else {
 563                        list_del(&buddy->lru);
 564                        zone->free_area[order].nr_free--;
 565                        rmv_page_order(buddy);
 566                }
 567                combined_idx = buddy_idx & page_idx;
 568                page = page + (combined_idx - page_idx);
 569                page_idx = combined_idx;
 570                order++;
 571        }
 572        set_page_order(page, order);
 573
 574        /*
 575         * If this is not the largest possible page, check if the buddy
 576         * of the next-highest order is free. If it is, it's possible
 577         * that pages are being freed that will coalesce soon. In case,
 578         * that is happening, add the free page to the tail of the list
 579         * so it's less likely to be used soon and more likely to be merged
 580         * as a higher order page
 581         */
 582        if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
 583                struct page *higher_page, *higher_buddy;
 584                combined_idx = buddy_idx & page_idx;
 585                higher_page = page + (combined_idx - page_idx);
 586                buddy_idx = __find_buddy_index(combined_idx, order + 1);
 587                higher_buddy = higher_page + (buddy_idx - combined_idx);
 588                if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
 589                        list_add_tail(&page->lru,
 590                                &zone->free_area[order].free_list[migratetype]);
 591                        goto out;
 592                }
 593        }
 594
 595        list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
 596out:
 597        zone->free_area[order].nr_free++;
 598}
 599
 600/*
 601 * free_page_mlock() -- clean up attempts to free and mlocked() page.
 602 * Page should not be on lru, so no need to fix that up.
 603 * free_pages_check() will verify...
 604 */
 605static inline void free_page_mlock(struct page *page)
 606{
 607        __dec_zone_page_state(page, NR_MLOCK);
 608        __count_vm_event(UNEVICTABLE_MLOCKFREED);
 609}
 610
 611static inline int free_pages_check(struct page *page)
 612{
 613        if (unlikely(page_mapcount(page) |
 614                (page->mapping != NULL)  |
 615                (atomic_read(&page->_count) != 0) |
 616                (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
 617                (mem_cgroup_bad_page_check(page)))) {
 618                bad_page(page);
 619                return 1;
 620        }
 621        if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
 622                page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
 623        return 0;
 624}
 625
 626/*
 627 * Frees a number of pages from the PCP lists
 628 * Assumes all pages on list are in same zone, and of same order.
 629 * count is the number of pages to free.
 630 *
 631 * If the zone was previously in an "all pages pinned" state then look to
 632 * see if this freeing clears that state.
 633 *
 634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
 635 * pinned" detection logic.
 636 */
 637static void free_pcppages_bulk(struct zone *zone, int count,
 638                                        struct per_cpu_pages *pcp)
 639{
 640        int migratetype = 0;
 641        int batch_free = 0;
 642        int to_free = count;
 643
 644        spin_lock(&zone->lock);
 645        zone->all_unreclaimable = 0;
 646        zone->pages_scanned = 0;
 647
 648        while (to_free) {
 649                struct page *page;
 650                struct list_head *list;
 651
 652                /*
 653                 * Remove pages from lists in a round-robin fashion. A
 654                 * batch_free count is maintained that is incremented when an
 655                 * empty list is encountered.  This is so more pages are freed
 656                 * off fuller lists instead of spinning excessively around empty
 657                 * lists
 658                 */
 659                do {
 660                        batch_free++;
 661                        if (++migratetype == MIGRATE_PCPTYPES)
 662                                migratetype = 0;
 663                        list = &pcp->lists[migratetype];
 664                } while (list_empty(list));
 665
 666                /* This is the only non-empty list. Free them all. */
 667                if (batch_free == MIGRATE_PCPTYPES)
 668                        batch_free = to_free;
 669
 670                do {
 671                        page = list_entry(list->prev, struct page, lru);
 672                        /* must delete as __free_one_page list manipulates */
 673                        list_del(&page->lru);
 674                        /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
 675                        __free_one_page(page, zone, 0, page_private(page));
 676                        trace_mm_page_pcpu_drain(page, 0, page_private(page));
 677                } while (--to_free && --batch_free && !list_empty(list));
 678        }
 679        __mod_zone_page_state(zone, NR_FREE_PAGES, count);
 680        spin_unlock(&zone->lock);
 681}
 682
 683static void free_one_page(struct zone *zone, struct page *page, int order,
 684                                int migratetype)
 685{
 686        spin_lock(&zone->lock);
 687        zone->all_unreclaimable = 0;
 688        zone->pages_scanned = 0;
 689
 690        __free_one_page(page, zone, order, migratetype);
 691        __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 692        spin_unlock(&zone->lock);
 693}
 694
 695static bool free_pages_prepare(struct page *page, unsigned int order)
 696{
 697        int i;
 698        int bad = 0;
 699
 700        trace_mm_page_free(page, order);
 701        kmemcheck_free_shadow(page, order);
 702
 703        if (PageAnon(page))
 704                page->mapping = NULL;
 705        for (i = 0; i < (1 << order); i++)
 706                bad += free_pages_check(page + i);
 707        if (bad)
 708                return false;
 709
 710        if (!PageHighMem(page)) {
 711                debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
 712                debug_check_no_obj_freed(page_address(page),
 713                                           PAGE_SIZE << order);
 714        }
 715        arch_free_page(page, order);
 716        kernel_map_pages(page, 1 << order, 0);
 717
 718        return true;
 719}
 720
 721static void __free_pages_ok(struct page *page, unsigned int order)
 722{
 723        unsigned long flags;
 724        int wasMlocked = __TestClearPageMlocked(page);
 725
 726        if (!free_pages_prepare(page, order))
 727                return;
 728
 729        local_irq_save(flags);
 730        if (unlikely(wasMlocked))
 731                free_page_mlock(page);
 732        __count_vm_events(PGFREE, 1 << order);
 733        free_one_page(page_zone(page), page, order,
 734                                        get_pageblock_migratetype(page));
 735        local_irq_restore(flags);
 736}
 737
 738void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 739{
 740        unsigned int nr_pages = 1 << order;
 741        unsigned int loop;
 742
 743        prefetchw(page);
 744        for (loop = 0; loop < nr_pages; loop++) {
 745                struct page *p = &page[loop];
 746
 747                if (loop + 1 < nr_pages)
 748                        prefetchw(p + 1);
 749                __ClearPageReserved(p);
 750                set_page_count(p, 0);
 751        }
 752
 753        set_page_refcounted(page);
 754        __free_pages(page, order);
 755}
 756
 757#ifdef CONFIG_CMA
 758/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
 759void __init init_cma_reserved_pageblock(struct page *page)
 760{
 761        unsigned i = pageblock_nr_pages;
 762        struct page *p = page;
 763
 764        do {
 765                __ClearPageReserved(p);
 766                set_page_count(p, 0);
 767        } while (++p, --i);
 768
 769        set_page_refcounted(page);
 770        set_pageblock_migratetype(page, MIGRATE_CMA);
 771        __free_pages(page, pageblock_order);
 772        totalram_pages += pageblock_nr_pages;
 773}
 774#endif
 775
 776/*
 777 * The order of subdivision here is critical for the IO subsystem.
 778 * Please do not alter this order without good reasons and regression
 779 * testing. Specifically, as large blocks of memory are subdivided,
 780 * the order in which smaller blocks are delivered depends on the order
 781 * they're subdivided in this function. This is the primary factor
 782 * influencing the order in which pages are delivered to the IO
 783 * subsystem according to empirical testing, and this is also justified
 784 * by considering the behavior of a buddy system containing a single
 785 * large block of memory acted on by a series of small allocations.
 786 * This behavior is a critical factor in sglist merging's success.
 787 *
 788 * -- wli
 789 */
 790static inline void expand(struct zone *zone, struct page *page,
 791        int low, int high, struct free_area *area,
 792        int migratetype)
 793{
 794        unsigned long size = 1 << high;
 795
 796        while (high > low) {
 797                area--;
 798                high--;
 799                size >>= 1;
 800                VM_BUG_ON(bad_range(zone, &page[size]));
 801
 802#ifdef CONFIG_DEBUG_PAGEALLOC
 803                if (high < debug_guardpage_minorder()) {
 804                        /*
 805                         * Mark as guard pages (or page), that will allow to
 806                         * merge back to allocator when buddy will be freed.
 807                         * Corresponding page table entries will not be touched,
 808                         * pages will stay not present in virtual address space
 809                         */
 810                        INIT_LIST_HEAD(&page[size].lru);
 811                        set_page_guard_flag(&page[size]);
 812                        set_page_private(&page[size], high);
 813                        /* Guard pages are not available for any usage */
 814                        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
 815                        continue;
 816                }
 817#endif
 818                list_add(&page[size].lru, &area->free_list[migratetype]);
 819                area->nr_free++;
 820                set_page_order(&page[size], high);
 821        }
 822}
 823
 824/*
 825 * This page is about to be returned from the page allocator
 826 */
 827static inline int check_new_page(struct page *page)
 828{
 829        if (unlikely(page_mapcount(page) |
 830                (page->mapping != NULL)  |
 831                (atomic_read(&page->_count) != 0)  |
 832                (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
 833                (mem_cgroup_bad_page_check(page)))) {
 834                bad_page(page);
 835                return 1;
 836        }
 837        return 0;
 838}
 839
 840static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 841{
 842        int i;
 843
 844        for (i = 0; i < (1 << order); i++) {
 845                struct page *p = page + i;
 846                if (unlikely(check_new_page(p)))
 847                        return 1;
 848        }
 849
 850        set_page_private(page, 0);
 851        set_page_refcounted(page);
 852
 853        arch_alloc_page(page, order);
 854        kernel_map_pages(page, 1 << order, 1);
 855
 856        if (gfp_flags & __GFP_ZERO)
 857                prep_zero_page(page, order, gfp_flags);
 858
 859        if (order && (gfp_flags & __GFP_COMP))
 860                prep_compound_page(page, order);
 861
 862        return 0;
 863}
 864
 865/*
 866 * Go through the free lists for the given migratetype and remove
 867 * the smallest available page from the freelists
 868 */
 869static inline
 870struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
 871                                                int migratetype)
 872{
 873        unsigned int current_order;
 874        struct free_area * area;
 875        struct page *page;
 876
 877        /* Find a page of the appropriate size in the preferred list */
 878        for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 879                area = &(zone->free_area[current_order]);
 880                if (list_empty(&area->free_list[migratetype]))
 881                        continue;
 882
 883                page = list_entry(area->free_list[migratetype].next,
 884                                                        struct page, lru);
 885                list_del(&page->lru);
 886                rmv_page_order(page);
 887                area->nr_free--;
 888                expand(zone, page, order, current_order, area, migratetype);
 889                return page;
 890        }
 891
 892        return NULL;
 893}
 894
 895
 896/*
 897 * This array describes the order lists are fallen back to when
 898 * the free lists for the desirable migrate type are depleted
 899 */
 900static int fallbacks[MIGRATE_TYPES][4] = {
 901        [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,     MIGRATE_RESERVE },
 902        [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,     MIGRATE_RESERVE },
 903#ifdef CONFIG_CMA
 904        [MIGRATE_MOVABLE]     = { MIGRATE_CMA,         MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
 905        [MIGRATE_CMA]         = { MIGRATE_RESERVE }, /* Never used */
 906#else
 907        [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE,   MIGRATE_RESERVE },
 908#endif
 909        [MIGRATE_RESERVE]     = { MIGRATE_RESERVE }, /* Never used */
 910        [MIGRATE_ISOLATE]     = { MIGRATE_RESERVE }, /* Never used */
 911};
 912
 913/*
 914 * Move the free pages in a range to the free lists of the requested type.
 915 * Note that start_page and end_pages are not aligned on a pageblock
 916 * boundary. If alignment is required, use move_freepages_block()
 917 */
 918static int move_freepages(struct zone *zone,
 919                          struct page *start_page, struct page *end_page,
 920                          int migratetype)
 921{
 922        struct page *page;
 923        unsigned long order;
 924        int pages_moved = 0;
 925
 926#ifndef CONFIG_HOLES_IN_ZONE
 927        /*
 928         * page_zone is not safe to call in this context when
 929         * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
 930         * anyway as we check zone boundaries in move_freepages_block().
 931         * Remove at a later date when no bug reports exist related to
 932         * grouping pages by mobility
 933         */
 934        BUG_ON(page_zone(start_page) != page_zone(end_page));
 935#endif
 936
 937        for (page = start_page; page <= end_page;) {
 938                /* Make sure we are not inadvertently changing nodes */
 939                VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
 940
 941                if (!pfn_valid_within(page_to_pfn(page))) {
 942                        page++;
 943                        continue;
 944                }
 945
 946                if (!PageBuddy(page)) {
 947                        page++;
 948                        continue;
 949                }
 950
 951                order = page_order(page);
 952                list_move(&page->lru,
 953                          &zone->free_area[order].free_list[migratetype]);
 954                page += 1 << order;
 955                pages_moved += 1 << order;
 956        }
 957
 958        return pages_moved;
 959}
 960
 961int move_freepages_block(struct zone *zone, struct page *page,
 962                                int migratetype)
 963{
 964        unsigned long start_pfn, end_pfn;
 965        struct page *start_page, *end_page;
 966
 967        start_pfn = page_to_pfn(page);
 968        start_pfn = start_pfn & ~(pageblock_nr_pages-1);
 969        start_page = pfn_to_page(start_pfn);
 970        end_page = start_page + pageblock_nr_pages - 1;
 971        end_pfn = start_pfn + pageblock_nr_pages - 1;
 972
 973        /* Do not cross zone boundaries */
 974        if (start_pfn < zone->zone_start_pfn)
 975                start_page = page;
 976        if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
 977                return 0;
 978
 979        return move_freepages(zone, start_page, end_page, migratetype);
 980}
 981
 982static void change_pageblock_range(struct page *pageblock_page,
 983                                        int start_order, int migratetype)
 984{
 985        int nr_pageblocks = 1 << (start_order - pageblock_order);
 986
 987        while (nr_pageblocks--) {
 988                set_pageblock_migratetype(pageblock_page, migratetype);
 989                pageblock_page += pageblock_nr_pages;
 990        }
 991}
 992
 993/* Remove an element from the buddy allocator from the fallback list */
 994static inline struct page *
 995__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
 996{
 997        struct free_area * area;
 998        int current_order;
 999        struct page *page;
1000        int migratetype, i;
1001
1002        /* Find the largest possible block of pages in the other list */
1003        for (current_order = MAX_ORDER-1; current_order >= order;
1004                                                --current_order) {
1005                for (i = 0;; i++) {
1006                        migratetype = fallbacks[start_migratetype][i];
1007
1008                        /* MIGRATE_RESERVE handled later if necessary */
1009                        if (migratetype == MIGRATE_RESERVE)
1010                                break;
1011
1012                        area = &(zone->free_area[current_order]);
1013                        if (list_empty(&area->free_list[migratetype]))
1014                                continue;
1015
1016                        page = list_entry(area->free_list[migratetype].next,
1017                                        struct page, lru);
1018                        area->nr_free--;
1019
1020                        /*
1021                         * If breaking a large block of pages, move all free
1022                         * pages to the preferred allocation list. If falling
1023                         * back for a reclaimable kernel allocation, be more
1024                         * aggressive about taking ownership of free pages
1025                         *
1026                         * On the other hand, never change migration
1027                         * type of MIGRATE_CMA pageblocks nor move CMA
1028                         * pages on different free lists. We don't
1029                         * want unmovable pages to be allocated from
1030                         * MIGRATE_CMA areas.
1031                         */
1032                        if (!is_migrate_cma(migratetype) &&
1033                            (unlikely(current_order >= pageblock_order / 2) ||
1034                             start_migratetype == MIGRATE_RECLAIMABLE ||
1035                             page_group_by_mobility_disabled)) {
1036                                int pages;
1037                                pages = move_freepages_block(zone, page,
1038                                                                start_migratetype);
1039
1040                                /* Claim the whole block if over half of it is free */
1041                                if (pages >= (1 << (pageblock_order-1)) ||
1042                                                page_group_by_mobility_disabled)
1043                                        set_pageblock_migratetype(page,
1044                                                                start_migratetype);
1045
1046                                migratetype = start_migratetype;
1047                        }
1048
1049                        /* Remove the page from the freelists */
1050                        list_del(&page->lru);
1051                        rmv_page_order(page);
1052
1053                        /* Take ownership for orders >= pageblock_order */
1054                        if (current_order >= pageblock_order &&
1055                            !is_migrate_cma(migratetype))
1056                                change_pageblock_range(page, current_order,
1057                                                        start_migratetype);
1058
1059                        expand(zone, page, order, current_order, area,
1060                               is_migrate_cma(migratetype)
1061                             ? migratetype : start_migratetype);
1062
1063                        trace_mm_page_alloc_extfrag(page, order, current_order,
1064                                start_migratetype, migratetype);
1065
1066                        return page;
1067                }
1068        }
1069
1070        return NULL;
1071}
1072
1073/*
1074 * Do the hard work of removing an element from the buddy allocator.
1075 * Call me with the zone->lock already held.
1076 */
1077static struct page *__rmqueue(struct zone *zone, unsigned int order,
1078                                                int migratetype)
1079{
1080        struct page *page;
1081
1082retry_reserve:
1083        page = __rmqueue_smallest(zone, order, migratetype);
1084
1085        if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1086                page = __rmqueue_fallback(zone, order, migratetype);
1087
1088                /*
1089                 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1090                 * is used because __rmqueue_smallest is an inline function
1091                 * and we want just one call site
1092                 */
1093                if (!page) {
1094                        migratetype = MIGRATE_RESERVE;
1095                        goto retry_reserve;
1096                }
1097        }
1098
1099        trace_mm_page_alloc_zone_locked(page, order, migratetype);
1100        return page;
1101}
1102
1103/*
1104 * Obtain a specified number of elements from the buddy allocator, all under
1105 * a single hold of the lock, for efficiency.  Add them to the supplied list.
1106 * Returns the number of new pages which were placed at *list.
1107 */
1108static int rmqueue_bulk(struct zone *zone, unsigned int order,
1109                        unsigned long count, struct list_head *list,
1110                        int migratetype, int cold)
1111{
1112        int mt = migratetype, i;
1113
1114        spin_lock(&zone->lock);
1115        for (i = 0; i < count; ++i) {
1116                struct page *page = __rmqueue(zone, order, migratetype);
1117                if (unlikely(page == NULL))
1118                        break;
1119
1120                /*
1121                 * Split buddy pages returned by expand() are received here
1122                 * in physical page order. The page is added to the callers and
1123                 * list and the list head then moves forward. From the callers
1124                 * perspective, the linked list is ordered by page number in
1125                 * some conditions. This is useful for IO devices that can
1126                 * merge IO requests if the physical pages are ordered
1127                 * properly.
1128                 */
1129                if (likely(cold == 0))
1130                        list_add(&page->lru, list);
1131                else
1132                        list_add_tail(&page->lru, list);
1133                if (IS_ENABLED(CONFIG_CMA)) {
1134                        mt = get_pageblock_migratetype(page);
1135                        if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1136                                mt = migratetype;
1137                }
1138                set_page_private(page, mt);
1139                list = &page->lru;
1140        }
1141        __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1142        spin_unlock(&zone->lock);
1143        return i;
1144}
1145
1146#ifdef CONFIG_NUMA
1147/*
1148 * Called from the vmstat counter updater to drain pagesets of this
1149 * currently executing processor on remote nodes after they have
1150 * expired.
1151 *
1152 * Note that this function must be called with the thread pinned to
1153 * a single processor.
1154 */
1155void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1156{
1157        unsigned long flags;
1158        int to_drain;
1159
1160        local_irq_save(flags);
1161        if (pcp->count >= pcp->batch)
1162                to_drain = pcp->batch;
1163        else
1164                to_drain = pcp->count;
1165        if (to_drain > 0) {
1166                free_pcppages_bulk(zone, to_drain, pcp);
1167                pcp->count -= to_drain;
1168        }
1169        local_irq_restore(flags);
1170}
1171#endif
1172
1173/*
1174 * Drain pages of the indicated processor.
1175 *
1176 * The processor must either be the current processor and the
1177 * thread pinned to the current processor or a processor that
1178 * is not online.
1179 */
1180static void drain_pages(unsigned int cpu)
1181{
1182        unsigned long flags;
1183        struct zone *zone;
1184
1185        for_each_populated_zone(zone) {
1186                struct per_cpu_pageset *pset;
1187                struct per_cpu_pages *pcp;
1188
1189                local_irq_save(flags);
1190                pset = per_cpu_ptr(zone->pageset, cpu);
1191
1192                pcp = &pset->pcp;
1193                if (pcp->count) {
1194                        free_pcppages_bulk(zone, pcp->count, pcp);
1195                        pcp->count = 0;
1196                }
1197                local_irq_restore(flags);
1198        }
1199}
1200
1201/*
1202 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1203 */
1204void drain_local_pages(void *arg)
1205{
1206        drain_pages(smp_processor_id());
1207}
1208
1209/*
1210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1211 *
1212 * Note that this code is protected against sending an IPI to an offline
1213 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1214 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1215 * nothing keeps CPUs from showing up after we populated the cpumask and
1216 * before the call to on_each_cpu_mask().
1217 */
1218void drain_all_pages(void)
1219{
1220        int cpu;
1221        struct per_cpu_pageset *pcp;
1222        struct zone *zone;
1223
1224        /*
1225         * Allocate in the BSS so we wont require allocation in
1226         * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1227         */
1228        static cpumask_t cpus_with_pcps;
1229
1230        /*
1231         * We don't care about racing with CPU hotplug event
1232         * as offline notification will cause the notified
1233         * cpu to drain that CPU pcps and on_each_cpu_mask
1234         * disables preemption as part of its processing
1235         */
1236        for_each_online_cpu(cpu) {
1237                bool has_pcps = false;
1238                for_each_populated_zone(zone) {
1239                        pcp = per_cpu_ptr(zone->pageset, cpu);
1240                        if (pcp->pcp.count) {
1241                                has_pcps = true;
1242                                break;
1243                        }
1244                }
1245                if (has_pcps)
1246                        cpumask_set_cpu(cpu, &cpus_with_pcps);
1247                else
1248                        cpumask_clear_cpu(cpu, &cpus_with_pcps);
1249        }
1250        on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1251}
1252
1253#ifdef CONFIG_HIBERNATION
1254
1255void mark_free_pages(struct zone *zone)
1256{
1257        unsigned long pfn, max_zone_pfn;
1258        unsigned long flags;
1259        int order, t;
1260        struct list_head *curr;
1261
1262        if (!zone->spanned_pages)
1263                return;
1264
1265        spin_lock_irqsave(&zone->lock, flags);
1266
1267        max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1268        for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1269                if (pfn_valid(pfn)) {
1270                        struct page *page = pfn_to_page(pfn);
1271
1272                        if (!swsusp_page_is_forbidden(page))
1273                                swsusp_unset_page_free(page);
1274                }
1275
1276        for_each_migratetype_order(order, t) {
1277                list_for_each(curr, &zone->free_area[order].free_list[t]) {
1278                        unsigned long i;
1279
1280                        pfn = page_to_pfn(list_entry(curr, struct page, lru));
1281                        for (i = 0; i < (1UL << order); i++)
1282                                swsusp_set_page_free(pfn_to_page(pfn + i));
1283                }
1284        }
1285        spin_unlock_irqrestore(&zone->lock, flags);
1286}
1287#endif /* CONFIG_PM */
1288
1289/*
1290 * Free a 0-order page
1291 * cold == 1 ? free a cold page : free a hot page
1292 */
1293void free_hot_cold_page(struct page *page, int cold)
1294{
1295        struct zone *zone = page_zone(page);
1296        struct per_cpu_pages *pcp;
1297        unsigned long flags;
1298        int migratetype;
1299        int wasMlocked = __TestClearPageMlocked(page);
1300
1301        if (!free_pages_prepare(page, 0))
1302                return;
1303
1304        migratetype = get_pageblock_migratetype(page);
1305        set_page_private(page, migratetype);
1306        local_irq_save(flags);
1307        if (unlikely(wasMlocked))
1308                free_page_mlock(page);
1309        __count_vm_event(PGFREE);
1310
1311        /*
1312         * We only track unmovable, reclaimable and movable on pcp lists.
1313         * Free ISOLATE pages back to the allocator because they are being
1314         * offlined but treat RESERVE as movable pages so we can get those
1315         * areas back if necessary. Otherwise, we may have to free
1316         * excessively into the page allocator
1317         */
1318        if (migratetype >= MIGRATE_PCPTYPES) {
1319                if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1320                        free_one_page(zone, page, 0, migratetype);
1321                        goto out;
1322                }
1323                migratetype = MIGRATE_MOVABLE;
1324        }
1325
1326        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1327        if (cold)
1328                list_add_tail(&page->lru, &pcp->lists[migratetype]);
1329        else
1330                list_add(&page->lru, &pcp->lists[migratetype]);
1331        pcp->count++;
1332        if (pcp->count >= pcp->high) {
1333                free_pcppages_bulk(zone, pcp->batch, pcp);
1334                pcp->count -= pcp->batch;
1335        }
1336
1337out:
1338        local_irq_restore(flags);
1339}
1340
1341/*
1342 * Free a list of 0-order pages
1343 */
1344void free_hot_cold_page_list(struct list_head *list, int cold)
1345{
1346        struct page *page, *next;
1347
1348        list_for_each_entry_safe(page, next, list, lru) {
1349                trace_mm_page_free_batched(page, cold);
1350                free_hot_cold_page(page, cold);
1351        }
1352}
1353
1354/*
1355 * split_page takes a non-compound higher-order page, and splits it into
1356 * n (1<<order) sub-pages: page[0..n]
1357 * Each sub-page must be freed individually.
1358 *
1359 * Note: this is probably too low level an operation for use in drivers.
1360 * Please consult with lkml before using this in your driver.
1361 */
1362void split_page(struct page *page, unsigned int order)
1363{
1364        int i;
1365
1366        VM_BUG_ON(PageCompound(page));
1367        VM_BUG_ON(!page_count(page));
1368
1369#ifdef CONFIG_KMEMCHECK
1370        /*
1371         * Split shadow pages too, because free(page[0]) would
1372         * otherwise free the whole shadow.
1373         */
1374        if (kmemcheck_page_is_tracked(page))
1375                split_page(virt_to_page(page[0].shadow), order);
1376#endif
1377
1378        for (i = 1; i < (1 << order); i++)
1379                set_page_refcounted(page + i);
1380}
1381
1382/*
1383 * Similar to split_page except the page is already free. As this is only
1384 * being used for migration, the migratetype of the block also changes.
1385 * As this is called with interrupts disabled, the caller is responsible
1386 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1387 * are enabled.
1388 *
1389 * Note: this is probably too low level an operation for use in drivers.
1390 * Please consult with lkml before using this in your driver.
1391 */
1392int split_free_page(struct page *page)
1393{
1394        unsigned int order;
1395        unsigned long watermark;
1396        struct zone *zone;
1397
1398        BUG_ON(!PageBuddy(page));
1399
1400        zone = page_zone(page);
1401        order = page_order(page);
1402
1403        /* Obey watermarks as if the page was being allocated */
1404        watermark = low_wmark_pages(zone) + (1 << order);
1405        if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1406                return 0;
1407
1408        /* Remove page from free list */
1409        list_del(&page->lru);
1410        zone->free_area[order].nr_free--;
1411        rmv_page_order(page);
1412        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1413
1414        /* Split into individual pages */
1415        set_page_refcounted(page);
1416        split_page(page, order);
1417
1418        if (order >= pageblock_order - 1) {
1419                struct page *endpage = page + (1 << order) - 1;
1420                for (; page < endpage; page += pageblock_nr_pages) {
1421                        int mt = get_pageblock_migratetype(page);
1422                        if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1423                                set_pageblock_migratetype(page,
1424                                                          MIGRATE_MOVABLE);
1425                }
1426        }
1427
1428        return 1 << order;
1429}
1430
1431/*
1432 * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
1433 * we cheat by calling it from here, in the order > 0 path.  Saves a branch
1434 * or two.
1435 */
1436static inline
1437struct page *buffered_rmqueue(struct zone *preferred_zone,
1438                        struct zone *zone, int order, gfp_t gfp_flags,
1439                        int migratetype)
1440{
1441        unsigned long flags;
1442        struct page *page;
1443        int cold = !!(gfp_flags & __GFP_COLD);
1444
1445again:
1446        if (likely(order == 0)) {
1447                struct per_cpu_pages *pcp;
1448                struct list_head *list;
1449
1450                local_irq_save(flags);
1451                pcp = &this_cpu_ptr(zone->pageset)->pcp;
1452                list = &pcp->lists[migratetype];
1453                if (list_empty(list)) {
1454                        pcp->count += rmqueue_bulk(zone, 0,
1455                                        pcp->batch, list,
1456                                        migratetype, cold);
1457                        if (unlikely(list_empty(list)))
1458                                goto failed;
1459                }
1460
1461                if (cold)
1462                        page = list_entry(list->prev, struct page, lru);
1463                else
1464                        page = list_entry(list->next, struct page, lru);
1465
1466                list_del(&page->lru);
1467                pcp->count--;
1468        } else {
1469                if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1470                        /*
1471                         * __GFP_NOFAIL is not to be used in new code.
1472                         *
1473                         * All __GFP_NOFAIL callers should be fixed so that they
1474                         * properly detect and handle allocation failures.
1475                         *
1476                         * We most definitely don't want callers attempting to
1477                         * allocate greater than order-1 page units with
1478                         * __GFP_NOFAIL.
1479                         */
1480                        WARN_ON_ONCE(order > 1);
1481                }
1482                spin_lock_irqsave(&zone->lock, flags);
1483                page = __rmqueue(zone, order, migratetype);
1484                spin_unlock(&zone->lock);
1485                if (!page)
1486                        goto failed;
1487                __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1488        }
1489
1490        __count_zone_vm_events(PGALLOC, zone, 1 << order);
1491        zone_statistics(preferred_zone, zone, gfp_flags);
1492        local_irq_restore(flags);
1493
1494        VM_BUG_ON(bad_range(zone, page));
1495        if (prep_new_page(page, order, gfp_flags))
1496                goto again;
1497        return page;
1498
1499failed:
1500        local_irq_restore(flags);
1501        return NULL;
1502}
1503
1504/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1505#define ALLOC_WMARK_MIN         WMARK_MIN
1506#define ALLOC_WMARK_LOW         WMARK_LOW
1507#define ALLOC_WMARK_HIGH        WMARK_HIGH
1508#define ALLOC_NO_WATERMARKS     0x04 /* don't check watermarks at all */
1509
1510/* Mask to get the watermark bits */
1511#define ALLOC_WMARK_MASK        (ALLOC_NO_WATERMARKS-1)
1512
1513#define ALLOC_HARDER            0x10 /* try to alloc harder */
1514#define ALLOC_HIGH              0x20 /* __GFP_HIGH set */
1515#define ALLOC_CPUSET            0x40 /* check for correct cpuset */
1516
1517#ifdef CONFIG_FAIL_PAGE_ALLOC
1518
1519static struct {
1520        struct fault_attr attr;
1521
1522        u32 ignore_gfp_highmem;
1523        u32 ignore_gfp_wait;
1524        u32 min_order;
1525} fail_page_alloc = {
1526        .attr = FAULT_ATTR_INITIALIZER,
1527        .ignore_gfp_wait = 1,
1528        .ignore_gfp_highmem = 1,
1529        .min_order = 1,
1530};
1531
1532static int __init setup_fail_page_alloc(char *str)
1533{
1534        return setup_fault_attr(&fail_page_alloc.attr, str);
1535}
1536__setup("fail_page_alloc=", setup_fail_page_alloc);
1537
1538static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1539{
1540        if (order < fail_page_alloc.min_order)
1541                return false;
1542        if (gfp_mask & __GFP_NOFAIL)
1543                return false;
1544        if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1545                return false;
1546        if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1547                return false;
1548
1549        return should_fail(&fail_page_alloc.attr, 1 << order);
1550}
1551
1552#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1553
1554static int __init fail_page_alloc_debugfs(void)
1555{
1556        umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1557        struct dentry *dir;
1558
1559        dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1560                                        &fail_page_alloc.attr);
1561        if (IS_ERR(dir))
1562                return PTR_ERR(dir);
1563
1564        if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1565                                &fail_page_alloc.ignore_gfp_wait))
1566                goto fail;
1567        if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1568                                &fail_page_alloc.ignore_gfp_highmem))
1569                goto fail;
1570        if (!debugfs_create_u32("min-order", mode, dir,
1571                                &fail_page_alloc.min_order))
1572                goto fail;
1573
1574        return 0;
1575fail:
1576        debugfs_remove_recursive(dir);
1577
1578        return -ENOMEM;
1579}
1580
1581late_initcall(fail_page_alloc_debugfs);
1582
1583#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1584
1585#else /* CONFIG_FAIL_PAGE_ALLOC */
1586
1587static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1588{
1589        return false;
1590}
1591
1592#endif /* CONFIG_FAIL_PAGE_ALLOC */
1593
1594/*
1595 * Return true if free pages are above 'mark'. This takes into account the order
1596 * of the allocation.
1597 */
1598static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1599                      int classzone_idx, int alloc_flags, long free_pages)
1600{
1601        /* free_pages my go negative - that's OK */
1602        long min = mark;
1603        long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1604        int o;
1605
1606        free_pages -= (1 << order) - 1;
1607        if (alloc_flags & ALLOC_HIGH)
1608                min -= min / 2;
1609        if (alloc_flags & ALLOC_HARDER)
1610                min -= min / 4;
1611
1612        if (free_pages <= min + lowmem_reserve)
1613                return false;
1614        for (o = 0; o < order; o++) {
1615                /* At the next order, this order's pages become unavailable */
1616                free_pages -= z->free_area[o].nr_free << o;
1617
1618                /* Require fewer higher order pages to be free */
1619                min >>= 1;
1620
1621                if (free_pages <= min)
1622                        return false;
1623        }
1624        return true;
1625}
1626
1627#ifdef CONFIG_MEMORY_ISOLATION
1628static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1629{
1630        if (unlikely(zone->nr_pageblock_isolate))
1631                return zone->nr_pageblock_isolate * pageblock_nr_pages;
1632        return 0;
1633}
1634#else
1635static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1636{
1637        return 0;
1638}
1639#endif
1640
1641bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1642                      int classzone_idx, int alloc_flags)
1643{
1644        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1645                                        zone_page_state(z, NR_FREE_PAGES));
1646}
1647
1648bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1649                      int classzone_idx, int alloc_flags)
1650{
1651        long free_pages = zone_page_state(z, NR_FREE_PAGES);
1652
1653        if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1654                free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1655
1656        /*
1657         * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1658         * it.  nr_zone_isolate_freepages is never accurate so kswapd might not
1659         * sleep although it could do so.  But this is more desirable for memory
1660         * hotplug than sleeping which can cause a livelock in the direct
1661         * reclaim path.
1662         */
1663        free_pages -= nr_zone_isolate_freepages(z);
1664        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1665                                                                free_pages);
1666}
1667
1668#ifdef CONFIG_NUMA
1669/*
1670 * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1671 * skip over zones that are not allowed by the cpuset, or that have
1672 * been recently (in last second) found to be nearly full.  See further
1673 * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1674 * that have to skip over a lot of full or unallowed zones.
1675 *
1676 * If the zonelist cache is present in the passed in zonelist, then
1677 * returns a pointer to the allowed node mask (either the current
1678 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1679 *
1680 * If the zonelist cache is not available for this zonelist, does
1681 * nothing and returns NULL.
1682 *
1683 * If the fullzones BITMAP in the zonelist cache is stale (more than
1684 * a second since last zap'd) then we zap it out (clear its bits.)
1685 *
1686 * We hold off even calling zlc_setup, until after we've checked the
1687 * first zone in the zonelist, on the theory that most allocations will
1688 * be satisfied from that first zone, so best to examine that zone as
1689 * quickly as we can.
1690 */
1691static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1692{
1693        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1694        nodemask_t *allowednodes;       /* zonelist_cache approximation */
1695
1696        zlc = zonelist->zlcache_ptr;
1697        if (!zlc)
1698                return NULL;
1699
1700        if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1701                bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1702                zlc->last_full_zap = jiffies;
1703        }
1704
1705        allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1706                                        &cpuset_current_mems_allowed :
1707                                        &node_states[N_HIGH_MEMORY];
1708        return allowednodes;
1709}
1710
1711/*
1712 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1713 * if it is worth looking at further for free memory:
1714 *  1) Check that the zone isn't thought to be full (doesn't have its
1715 *     bit set in the zonelist_cache fullzones BITMAP).
1716 *  2) Check that the zones node (obtained from the zonelist_cache
1717 *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1718 * Return true (non-zero) if zone is worth looking at further, or
1719 * else return false (zero) if it is not.
1720 *
1721 * This check -ignores- the distinction between various watermarks,
1722 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1723 * found to be full for any variation of these watermarks, it will
1724 * be considered full for up to one second by all requests, unless
1725 * we are so low on memory on all allowed nodes that we are forced
1726 * into the second scan of the zonelist.
1727 *
1728 * In the second scan we ignore this zonelist cache and exactly
1729 * apply the watermarks to all zones, even it is slower to do so.
1730 * We are low on memory in the second scan, and should leave no stone
1731 * unturned looking for a free page.
1732 */
1733static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1734                                                nodemask_t *allowednodes)
1735{
1736        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1737        int i;                          /* index of *z in zonelist zones */
1738        int n;                          /* node that zone *z is on */
1739
1740        zlc = zonelist->zlcache_ptr;
1741        if (!zlc)
1742                return 1;
1743
1744        i = z - zonelist->_zonerefs;
1745        n = zlc->z_to_n[i];
1746
1747        /* This zone is worth trying if it is allowed but not full */
1748        return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1749}
1750
1751/*
1752 * Given 'z' scanning a zonelist, set the corresponding bit in
1753 * zlc->fullzones, so that subsequent attempts to allocate a page
1754 * from that zone don't waste time re-examining it.
1755 */
1756static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1757{
1758        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1759        int i;                          /* index of *z in zonelist zones */
1760
1761        zlc = zonelist->zlcache_ptr;
1762        if (!zlc)
1763                return;
1764
1765        i = z - zonelist->_zonerefs;
1766
1767        set_bit(i, zlc->fullzones);
1768}
1769
1770/*
1771 * clear all zones full, called after direct reclaim makes progress so that
1772 * a zone that was recently full is not skipped over for up to a second
1773 */
1774static void zlc_clear_zones_full(struct zonelist *zonelist)
1775{
1776        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1777
1778        zlc = zonelist->zlcache_ptr;
1779        if (!zlc)
1780                return;
1781
1782        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1783}
1784
1785#else   /* CONFIG_NUMA */
1786
1787static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1788{
1789        return NULL;
1790}
1791
1792static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1793                                nodemask_t *allowednodes)
1794{
1795        return 1;
1796}
1797
1798static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1799{
1800}
1801
1802static void zlc_clear_zones_full(struct zonelist *zonelist)
1803{
1804}
1805#endif  /* CONFIG_NUMA */
1806
1807/*
1808 * get_page_from_freelist goes through the zonelist trying to allocate
1809 * a page.
1810 */
1811static struct page *
1812get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1813                struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1814                struct zone *preferred_zone, int migratetype)
1815{
1816        struct zoneref *z;
1817        struct page *page = NULL;
1818        int classzone_idx;
1819        struct zone *zone;
1820        nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1821        int zlc_active = 0;             /* set if using zonelist_cache */
1822        int did_zlc_setup = 0;          /* just call zlc_setup() one time */
1823
1824        classzone_idx = zone_idx(preferred_zone);
1825zonelist_scan:
1826        /*
1827         * Scan zonelist, looking for a zone with enough free.
1828         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1829         */
1830        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1831                                                high_zoneidx, nodemask) {
1832                if (NUMA_BUILD && zlc_active &&
1833                        !zlc_zone_worth_trying(zonelist, z, allowednodes))
1834                                continue;
1835                if ((alloc_flags & ALLOC_CPUSET) &&
1836                        !cpuset_zone_allowed_softwall(zone, gfp_mask))
1837                                continue;
1838                /*
1839                 * When allocating a page cache page for writing, we
1840                 * want to get it from a zone that is within its dirty
1841                 * limit, such that no single zone holds more than its
1842                 * proportional share of globally allowed dirty pages.
1843                 * The dirty limits take into account the zone's
1844                 * lowmem reserves and high watermark so that kswapd
1845                 * should be able to balance it without having to
1846                 * write pages from its LRU list.
1847                 *
1848                 * This may look like it could increase pressure on
1849                 * lower zones by failing allocations in higher zones
1850                 * before they are full.  But the pages that do spill
1851                 * over are limited as the lower zones are protected
1852                 * by this very same mechanism.  It should not become
1853                 * a practical burden to them.
1854                 *
1855                 * XXX: For now, allow allocations to potentially
1856                 * exceed the per-zone dirty limit in the slowpath
1857                 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1858                 * which is important when on a NUMA setup the allowed
1859                 * zones are together not big enough to reach the
1860                 * global limit.  The proper fix for these situations
1861                 * will require awareness of zones in the
1862                 * dirty-throttling and the flusher threads.
1863                 */
1864                if ((alloc_flags & ALLOC_WMARK_LOW) &&
1865                    (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1866                        goto this_zone_full;
1867
1868                BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1869                if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1870                        unsigned long mark;
1871                        int ret;
1872
1873                        mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1874                        if (zone_watermark_ok(zone, order, mark,
1875                                    classzone_idx, alloc_flags))
1876                                goto try_this_zone;
1877
1878                        if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1879                                /*
1880                                 * we do zlc_setup if there are multiple nodes
1881                                 * and before considering the first zone allowed
1882                                 * by the cpuset.
1883                                 */
1884                                allowednodes = zlc_setup(zonelist, alloc_flags);
1885                                zlc_active = 1;
1886                                did_zlc_setup = 1;
1887                        }
1888
1889                        if (zone_reclaim_mode == 0)
1890                                goto this_zone_full;
1891
1892                        /*
1893                         * As we may have just activated ZLC, check if the first
1894                         * eligible zone has failed zone_reclaim recently.
1895                         */
1896                        if (NUMA_BUILD && zlc_active &&
1897                                !zlc_zone_worth_trying(zonelist, z, allowednodes))
1898                                continue;
1899
1900                        ret = zone_reclaim(zone, gfp_mask, order);
1901                        switch (ret) {
1902                        case ZONE_RECLAIM_NOSCAN:
1903                                /* did not scan */
1904                                continue;
1905                        case ZONE_RECLAIM_FULL:
1906                                /* scanned but unreclaimable */
1907                                continue;
1908                        default:
1909                                /* did we reclaim enough */
1910                                if (!zone_watermark_ok(zone, order, mark,
1911                                                classzone_idx, alloc_flags))
1912                                        goto this_zone_full;
1913                        }
1914                }
1915
1916try_this_zone:
1917                page = buffered_rmqueue(preferred_zone, zone, order,
1918                                                gfp_mask, migratetype);
1919                if (page)
1920                        break;
1921this_zone_full:
1922                if (NUMA_BUILD)
1923                        zlc_mark_zone_full(zonelist, z);
1924        }
1925
1926        if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1927                /* Disable zlc cache for second zonelist scan */
1928                zlc_active = 0;
1929                goto zonelist_scan;
1930        }
1931
1932        if (page)
1933                /*
1934                 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1935                 * necessary to allocate the page. The expectation is
1936                 * that the caller is taking steps that will free more
1937                 * memory. The caller should avoid the page being used
1938                 * for !PFMEMALLOC purposes.
1939                 */
1940                page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1941
1942        return page;
1943}
1944
1945/*
1946 * Large machines with many possible nodes should not always dump per-node
1947 * meminfo in irq context.
1948 */
1949static inline bool should_suppress_show_mem(void)
1950{
1951        bool ret = false;
1952
1953#if NODES_SHIFT > 8
1954        ret = in_interrupt();
1955#endif
1956        return ret;
1957}
1958
1959static DEFINE_RATELIMIT_STATE(nopage_rs,
1960                DEFAULT_RATELIMIT_INTERVAL,
1961                DEFAULT_RATELIMIT_BURST);
1962
1963void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1964{
1965        unsigned int filter = SHOW_MEM_FILTER_NODES;
1966
1967        if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1968            debug_guardpage_minorder() > 0)
1969                return;
1970
1971        /*
1972         * This documents exceptions given to allocations in certain
1973         * contexts that are allowed to allocate outside current's set
1974         * of allowed nodes.
1975         */
1976        if (!(gfp_mask & __GFP_NOMEMALLOC))
1977                if (test_thread_flag(TIF_MEMDIE) ||
1978                    (current->flags & (PF_MEMALLOC | PF_EXITING)))
1979                        filter &= ~SHOW_MEM_FILTER_NODES;
1980        if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1981                filter &= ~SHOW_MEM_FILTER_NODES;
1982
1983        if (fmt) {
1984                struct va_format vaf;
1985                va_list args;
1986
1987                va_start(args, fmt);
1988
1989                vaf.fmt = fmt;
1990                vaf.va = &args;
1991
1992                pr_warn("%pV", &vaf);
1993
1994                va_end(args);
1995        }
1996
1997        pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1998                current->comm, order, gfp_mask);
1999
2000        dump_stack();
2001        if (!should_suppress_show_mem())
2002                show_mem(filter);
2003}
2004
2005static inline int
2006should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2007                                unsigned long did_some_progress,
2008                                unsigned long pages_reclaimed)
2009{
2010        /* Do not loop if specifically requested */
2011        if (gfp_mask & __GFP_NORETRY)
2012                return 0;
2013
2014        /* Always retry if specifically requested */
2015        if (gfp_mask & __GFP_NOFAIL)
2016                return 1;
2017
2018        /*
2019         * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2020         * making forward progress without invoking OOM. Suspend also disables
2021         * storage devices so kswapd will not help. Bail if we are suspending.
2022         */
2023        if (!did_some_progress && pm_suspended_storage())
2024                return 0;
2025
2026        /*
2027         * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2028         * means __GFP_NOFAIL, but that may not be true in other
2029         * implementations.
2030         */
2031        if (order <= PAGE_ALLOC_COSTLY_ORDER)
2032                return 1;
2033
2034        /*
2035         * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2036         * specified, then we retry until we no longer reclaim any pages
2037         * (above), or we've reclaimed an order of pages at least as
2038         * large as the allocation's order. In both cases, if the
2039         * allocation still fails, we stop retrying.
2040         */
2041        if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2042                return 1;
2043
2044        return 0;
2045}
2046
2047static inline struct page *
2048__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2049        struct zonelist *zonelist, enum zone_type high_zoneidx,
2050        nodemask_t *nodemask, struct zone *preferred_zone,
2051        int migratetype)
2052{
2053        struct page *page;
2054
2055        /* Acquire the OOM killer lock for the zones in zonelist */
2056        if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2057                schedule_timeout_uninterruptible(1);
2058                return NULL;
2059        }
2060
2061        /*
2062         * Go through the zonelist yet one more time, keep very high watermark
2063         * here, this is only to catch a parallel oom killing, we must fail if
2064         * we're still under heavy pressure.
2065         */
2066        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2067                order, zonelist, high_zoneidx,
2068                ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2069                preferred_zone, migratetype);
2070        if (page)
2071                goto out;
2072
2073        if (!(gfp_mask & __GFP_NOFAIL)) {
2074                /* The OOM killer will not help higher order allocs */
2075                if (order > PAGE_ALLOC_COSTLY_ORDER)
2076                        goto out;
2077                /* The OOM killer does not needlessly kill tasks for lowmem */
2078                if (high_zoneidx < ZONE_NORMAL)
2079                        goto out;
2080                /*
2081                 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2082                 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2083                 * The caller should handle page allocation failure by itself if
2084                 * it specifies __GFP_THISNODE.
2085                 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2086                 */
2087                if (gfp_mask & __GFP_THISNODE)
2088                        goto out;
2089        }
2090        /* Exhausted what can be done so it's blamo time */
2091        out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2092
2093out:
2094        clear_zonelist_oom(zonelist, gfp_mask);
2095        return page;
2096}
2097
2098#ifdef CONFIG_COMPACTION
2099/* Try memory compaction for high-order allocations before reclaim */
2100static struct page *
2101__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2102        struct zonelist *zonelist, enum zone_type high_zoneidx,
2103        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2104        int migratetype, bool sync_migration,
2105        bool *contended_compaction, bool *deferred_compaction,
2106        unsigned long *did_some_progress)
2107{
2108        struct page *page;
2109
2110        if (!order)
2111                return NULL;
2112
2113        if (compaction_deferred(preferred_zone, order)) {
2114                *deferred_compaction = true;
2115                return NULL;
2116        }
2117
2118        current->flags |= PF_MEMALLOC;
2119        *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2120                                                nodemask, sync_migration,
2121                                                contended_compaction);
2122        current->flags &= ~PF_MEMALLOC;
2123        if (*did_some_progress != COMPACT_SKIPPED) {
2124
2125                /* Page migration frees to the PCP lists but we want merging */
2126                drain_pages(get_cpu());
2127                put_cpu();
2128
2129                page = get_page_from_freelist(gfp_mask, nodemask,
2130                                order, zonelist, high_zoneidx,
2131                                alloc_flags & ~ALLOC_NO_WATERMARKS,
2132                                preferred_zone, migratetype);
2133                if (page) {
2134                        preferred_zone->compact_considered = 0;
2135                        preferred_zone->compact_defer_shift = 0;
2136                        if (order >= preferred_zone->compact_order_failed)
2137                                preferred_zone->compact_order_failed = order + 1;
2138                        count_vm_event(COMPACTSUCCESS);
2139                        return page;
2140                }
2141
2142                /*
2143                 * It's bad if compaction run occurs and fails.
2144                 * The most likely reason is that pages exist,
2145                 * but not enough to satisfy watermarks.
2146                 */
2147                count_vm_event(COMPACTFAIL);
2148
2149                /*
2150                 * As async compaction considers a subset of pageblocks, only
2151                 * defer if the failure was a sync compaction failure.
2152                 */
2153                if (sync_migration)
2154                        defer_compaction(preferred_zone, order);
2155
2156                cond_resched();
2157        }
2158
2159        return NULL;
2160}
2161#else
2162static inline struct page *
2163__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2164        struct zonelist *zonelist, enum zone_type high_zoneidx,
2165        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2166        int migratetype, bool sync_migration,
2167        bool *contended_compaction, bool *deferred_compaction,
2168        unsigned long *did_some_progress)
2169{
2170        return NULL;
2171}
2172#endif /* CONFIG_COMPACTION */
2173
2174/* Perform direct synchronous page reclaim */
2175static int
2176__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2177                  nodemask_t *nodemask)
2178{
2179        struct reclaim_state reclaim_state;
2180        int progress;
2181
2182        cond_resched();
2183
2184        /* We now go into synchronous reclaim */
2185        cpuset_memory_pressure_bump();
2186        current->flags |= PF_MEMALLOC;
2187        lockdep_set_current_reclaim_state(gfp_mask);
2188        reclaim_state.reclaimed_slab = 0;
2189        current->reclaim_state = &reclaim_state;
2190
2191        progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2192
2193        current->reclaim_state = NULL;
2194        lockdep_clear_current_reclaim_state();
2195        current->flags &= ~PF_MEMALLOC;
2196
2197        cond_resched();
2198
2199        return progress;
2200}
2201
2202/* The really slow allocator path where we enter direct reclaim */
2203static inline struct page *
2204__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2205        struct zonelist *zonelist, enum zone_type high_zoneidx,
2206        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2207        int migratetype, unsigned long *did_some_progress)
2208{
2209        struct page *page = NULL;
2210        bool drained = false;
2211
2212        *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2213                                               nodemask);
2214        if (unlikely(!(*did_some_progress)))
2215                return NULL;
2216
2217        /* After successful reclaim, reconsider all zones for allocation */
2218        if (NUMA_BUILD)
2219                zlc_clear_zones_full(zonelist);
2220
2221retry:
2222        page = get_page_from_freelist(gfp_mask, nodemask, order,
2223                                        zonelist, high_zoneidx,
2224                                        alloc_flags & ~ALLOC_NO_WATERMARKS,
2225                                        preferred_zone, migratetype);
2226
2227        /*
2228         * If an allocation failed after direct reclaim, it could be because
2229         * pages are pinned on the per-cpu lists. Drain them and try again
2230         */
2231        if (!page && !drained) {
2232                drain_all_pages();
2233                drained = true;
2234                goto retry;
2235        }
2236
2237        return page;
2238}
2239
2240/*
2241 * This is called in the allocator slow-path if the allocation request is of
2242 * sufficient urgency to ignore watermarks and take other desperate measures
2243 */
2244static inline struct page *
2245__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2246        struct zonelist *zonelist, enum zone_type high_zoneidx,
2247        nodemask_t *nodemask, struct zone *preferred_zone,
2248        int migratetype)
2249{
2250        struct page *page;
2251
2252        do {
2253                page = get_page_from_freelist(gfp_mask, nodemask, order,
2254                        zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2255                        preferred_zone, migratetype);
2256
2257                if (!page && gfp_mask & __GFP_NOFAIL)
2258                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2259        } while (!page && (gfp_mask & __GFP_NOFAIL));
2260
2261        return page;
2262}
2263
2264static inline
2265void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2266                                                enum zone_type high_zoneidx,
2267                                                enum zone_type classzone_idx)
2268{
2269        struct zoneref *z;
2270        struct zone *zone;
2271
2272        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2273                wakeup_kswapd(zone, order, classzone_idx);
2274}
2275
2276static inline int
2277gfp_to_alloc_flags(gfp_t gfp_mask)
2278{
2279        int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2280        const gfp_t wait = gfp_mask & __GFP_WAIT;
2281
2282        /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2283        BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2284
2285        /*
2286         * The caller may dip into page reserves a bit more if the caller
2287         * cannot run direct reclaim, or if the caller has realtime scheduling
2288         * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
2289         * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2290         */
2291        alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2292
2293        if (!wait) {
2294                /*
2295                 * Not worth trying to allocate harder for
2296                 * __GFP_NOMEMALLOC even if it can't schedule.
2297                 */
2298                if  (!(gfp_mask & __GFP_NOMEMALLOC))
2299                        alloc_flags |= ALLOC_HARDER;
2300                /*
2301                 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2302                 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2303                 */
2304                alloc_flags &= ~ALLOC_CPUSET;
2305        } else if (unlikely(rt_task(current)) && !in_interrupt())
2306                alloc_flags |= ALLOC_HARDER;
2307
2308        if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2309                if (gfp_mask & __GFP_MEMALLOC)
2310                        alloc_flags |= ALLOC_NO_WATERMARKS;
2311                else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2312                        alloc_flags |= ALLOC_NO_WATERMARKS;
2313                else if (!in_interrupt() &&
2314                                ((current->flags & PF_MEMALLOC) ||
2315                                 unlikely(test_thread_flag(TIF_MEMDIE))))
2316                        alloc_flags |= ALLOC_NO_WATERMARKS;
2317        }
2318
2319        return alloc_flags;
2320}
2321
2322bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2323{
2324        return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2325}
2326
2327static inline struct page *
2328__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2329        struct zonelist *zonelist, enum zone_type high_zoneidx,
2330        nodemask_t *nodemask, struct zone *preferred_zone,
2331        int migratetype)
2332{
2333        const gfp_t wait = gfp_mask & __GFP_WAIT;
2334        struct page *page = NULL;
2335        int alloc_flags;
2336        unsigned long pages_reclaimed = 0;
2337        unsigned long did_some_progress;
2338        bool sync_migration = false;
2339        bool deferred_compaction = false;
2340        bool contended_compaction = false;
2341
2342        /*
2343         * In the slowpath, we sanity check order to avoid ever trying to
2344         * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2345         * be using allocators in order of preference for an area that is
2346         * too large.
2347         */
2348        if (order >= MAX_ORDER) {
2349                WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2350                return NULL;
2351        }
2352
2353        /*
2354         * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2355         * __GFP_NOWARN set) should not cause reclaim since the subsystem
2356         * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2357         * using a larger set of nodes after it has established that the
2358         * allowed per node queues are empty and that nodes are
2359         * over allocated.
2360         */
2361        if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2362                goto nopage;
2363
2364restart:
2365        if (!(gfp_mask & __GFP_NO_KSWAPD))
2366                wake_all_kswapd(order, zonelist, high_zoneidx,
2367                                                zone_idx(preferred_zone));
2368
2369        /*
2370         * OK, we're below the kswapd watermark and have kicked background
2371         * reclaim. Now things get more complex, so set up alloc_flags according
2372         * to how we want to proceed.
2373         */
2374        alloc_flags = gfp_to_alloc_flags(gfp_mask);
2375
2376        /*
2377         * Find the true preferred zone if the allocation is unconstrained by
2378         * cpusets.
2379         */
2380        if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2381                first_zones_zonelist(zonelist, high_zoneidx, NULL,
2382                                        &preferred_zone);
2383
2384rebalance:
2385        /* This is the last chance, in general, before the goto nopage. */
2386        page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2387                        high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2388                        preferred_zone, migratetype);
2389        if (page)
2390                goto got_pg;
2391
2392        /* Allocate without watermarks if the context allows */
2393        if (alloc_flags & ALLOC_NO_WATERMARKS) {
2394                /*
2395                 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2396                 * the allocation is high priority and these type of
2397                 * allocations are system rather than user orientated
2398                 */
2399                zonelist = node_zonelist(numa_node_id(), gfp_mask);
2400
2401                page = __alloc_pages_high_priority(gfp_mask, order,
2402                                zonelist, high_zoneidx, nodemask,
2403                                preferred_zone, migratetype);
2404                if (page) {
2405                        goto got_pg;
2406                }
2407        }
2408
2409        /* Atomic allocations - we can't balance anything */
2410        if (!wait)
2411                goto nopage;
2412
2413        /* Avoid recursion of direct reclaim */
2414        if (current->flags & PF_MEMALLOC)
2415                goto nopage;
2416
2417        /* Avoid allocations with no watermarks from looping endlessly */
2418        if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2419                goto nopage;
2420
2421        /*
2422         * Try direct compaction. The first pass is asynchronous. Subsequent
2423         * attempts after direct reclaim are synchronous
2424         */
2425        page = __alloc_pages_direct_compact(gfp_mask, order,
2426                                        zonelist, high_zoneidx,
2427                                        nodemask,
2428                                        alloc_flags, preferred_zone,
2429                                        migratetype, sync_migration,
2430                                        &contended_compaction,
2431                                        &deferred_compaction,
2432                                        &did_some_progress);
2433        if (page)
2434                goto got_pg;
2435        sync_migration = true;
2436
2437        /*
2438         * If compaction is deferred for high-order allocations, it is because
2439         * sync compaction recently failed. In this is the case and the caller
2440         * requested a movable allocation that does not heavily disrupt the
2441         * system then fail the allocation instead of entering direct reclaim.
2442         */
2443        if ((deferred_compaction || contended_compaction) &&
2444                                                (gfp_mask & __GFP_NO_KSWAPD))
2445                goto nopage;
2446
2447        /* Try direct reclaim and then allocating */
2448        page = __alloc_pages_direct_reclaim(gfp_mask, order,
2449                                        zonelist, high_zoneidx,
2450                                        nodemask,
2451                                        alloc_flags, preferred_zone,
2452                                        migratetype, &did_some_progress);
2453        if (page)
2454                goto got_pg;
2455
2456        /*
2457         * If we failed to make any progress reclaiming, then we are
2458         * running out of options and have to consider going OOM
2459         */
2460        if (!did_some_progress) {
2461                if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2462                        if (oom_killer_disabled)
2463                                goto nopage;
2464                        /* Coredumps can quickly deplete all memory reserves */
2465                        if ((current->flags & PF_DUMPCORE) &&
2466                            !(gfp_mask & __GFP_NOFAIL))
2467                                goto nopage;
2468                        page = __alloc_pages_may_oom(gfp_mask, order,
2469                                        zonelist, high_zoneidx,
2470                                        nodemask, preferred_zone,
2471                                        migratetype);
2472                        if (page)
2473                                goto got_pg;
2474
2475                        if (!(gfp_mask & __GFP_NOFAIL)) {
2476                                /*
2477                                 * The oom killer is not called for high-order
2478                                 * allocations that may fail, so if no progress
2479                                 * is being made, there are no other options and
2480                                 * retrying is unlikely to help.
2481                                 */
2482                                if (order > PAGE_ALLOC_COSTLY_ORDER)
2483                                        goto nopage;
2484                                /*
2485                                 * The oom killer is not called for lowmem
2486                                 * allocations to prevent needlessly killing
2487                                 * innocent tasks.
2488                                 */
2489                                if (high_zoneidx < ZONE_NORMAL)
2490                                        goto nopage;
2491                        }
2492
2493                        goto restart;
2494                }
2495        }
2496
2497        /* Check if we should retry the allocation */
2498        pages_reclaimed += did_some_progress;
2499        if (should_alloc_retry(gfp_mask, order, did_some_progress,
2500                                                pages_reclaimed)) {
2501                /* Wait for some write requests to complete then retry */
2502                wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2503                goto rebalance;
2504        } else {
2505                /*
2506                 * High-order allocations do not necessarily loop after
2507                 * direct reclaim and reclaim/compaction depends on compaction
2508                 * being called after reclaim so call directly if necessary
2509                 */
2510                page = __alloc_pages_direct_compact(gfp_mask, order,
2511                                        zonelist, high_zoneidx,
2512                                        nodemask,
2513                                        alloc_flags, preferred_zone,
2514                                        migratetype, sync_migration,
2515                                        &contended_compaction,
2516                                        &deferred_compaction,
2517                                        &did_some_progress);
2518                if (page)
2519                        goto got_pg;
2520        }
2521
2522nopage:
2523        warn_alloc_failed(gfp_mask, order, NULL);
2524        return page;
2525got_pg:
2526        if (kmemcheck_enabled)
2527                kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2528
2529        return page;
2530}
2531
2532/*
2533 * This is the 'heart' of the zoned buddy allocator.
2534 */
2535struct page *
2536__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2537                        struct zonelist *zonelist, nodemask_t *nodemask)
2538{
2539        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2540        struct zone *preferred_zone;
2541        struct page *page = NULL;
2542        int migratetype = allocflags_to_migratetype(gfp_mask);
2543        unsigned int cpuset_mems_cookie;
2544
2545        gfp_mask &= gfp_allowed_mask;
2546
2547        lockdep_trace_alloc(gfp_mask);
2548
2549        might_sleep_if(gfp_mask & __GFP_WAIT);
2550
2551        if (should_fail_alloc_page(gfp_mask, order))
2552                return NULL;
2553
2554        /*
2555         * Check the zones suitable for the gfp_mask contain at least one
2556         * valid zone. It's possible to have an empty zonelist as a result
2557         * of GFP_THISNODE and a memoryless node
2558         */
2559        if (unlikely(!zonelist->_zonerefs->zone))
2560                return NULL;
2561
2562retry_cpuset:
2563        cpuset_mems_cookie = get_mems_allowed();
2564
2565        /* The preferred zone is used for statistics later */
2566        first_zones_zonelist(zonelist, high_zoneidx,
2567                                nodemask ? : &cpuset_current_mems_allowed,
2568                                &preferred_zone);
2569        if (!preferred_zone)
2570                goto out;
2571
2572        /* First allocation attempt */
2573        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2574                        zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2575                        preferred_zone, migratetype);
2576        if (unlikely(!page))
2577                page = __alloc_pages_slowpath(gfp_mask, order,
2578                                zonelist, high_zoneidx, nodemask,
2579                                preferred_zone, migratetype);
2580
2581        trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2582
2583out:
2584        /*
2585         * When updating a task's mems_allowed, it is possible to race with
2586         * parallel threads in such a way that an allocation can fail while
2587         * the mask is being updated. If a page allocation is about to fail,
2588         * check if the cpuset changed during allocation and if so, retry.
2589         */
2590        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2591                goto retry_cpuset;
2592
2593        return page;
2594}
2595EXPORT_SYMBOL(__alloc_pages_nodemask);
2596
2597/*
2598 * Common helper functions.
2599 */
2600unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2601{
2602        struct page *page;
2603
2604        /*
2605         * __get_free_pages() returns a 32-bit address, which cannot represent
2606         * a highmem page
2607         */
2608        VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2609
2610        page = alloc_pages(gfp_mask, order);
2611        if (!page)
2612                return 0;
2613        return (unsigned long) page_address(page);
2614}
2615EXPORT_SYMBOL(__get_free_pages);
2616
2617unsigned long get_zeroed_page(gfp_t gfp_mask)
2618{
2619        return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2620}
2621EXPORT_SYMBOL(get_zeroed_page);
2622
2623void __free_pages(struct page *page, unsigned int order)
2624{
2625        if (put_page_testzero(page)) {
2626                if (order == 0)
2627                        free_hot_cold_page(page, 0);
2628                else
2629                        __free_pages_ok(page, order);
2630        }
2631}
2632
2633EXPORT_SYMBOL(__free_pages);
2634
2635void free_pages(unsigned long addr, unsigned int order)
2636{
2637        if (addr != 0) {
2638                VM_BUG_ON(!virt_addr_valid((void *)addr));
2639                __free_pages(virt_to_page((void *)addr), order);
2640        }
2641}
2642
2643EXPORT_SYMBOL(free_pages);
2644
2645static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2646{
2647        if (addr) {
2648                unsigned long alloc_end = addr + (PAGE_SIZE << order);
2649                unsigned long used = addr + PAGE_ALIGN(size);
2650
2651                split_page(virt_to_page((void *)addr), order);
2652                while (used < alloc_end) {
2653                        free_page(used);
2654                        used += PAGE_SIZE;
2655                }
2656        }
2657        return (void *)addr;
2658}
2659
2660/**
2661 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2662 * @size: the number of bytes to allocate
2663 * @gfp_mask: GFP flags for the allocation
2664 *
2665 * This function is similar to alloc_pages(), except that it allocates the
2666 * minimum number of pages to satisfy the request.  alloc_pages() can only
2667 * allocate memory in power-of-two pages.
2668 *
2669 * This function is also limited by MAX_ORDER.
2670 *
2671 * Memory allocated by this function must be released by free_pages_exact().
2672 */
2673void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2674{
2675        unsigned int order = get_order(size);
2676        unsigned long addr;
2677
2678        addr = __get_free_pages(gfp_mask, order);
2679        return make_alloc_exact(addr, order, size);
2680}
2681EXPORT_SYMBOL(alloc_pages_exact);
2682
2683/**
2684 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2685 *                         pages on a node.
2686 * @nid: the preferred node ID where memory should be allocated
2687 * @size: the number of bytes to allocate
2688 * @gfp_mask: GFP flags for the allocation
2689 *
2690 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2691 * back.
2692 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2693 * but is not exact.
2694 */
2695void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2696{
2697        unsigned order = get_order(size);
2698        struct page *p = alloc_pages_node(nid, gfp_mask, order);
2699        if (!p)
2700                return NULL;
2701        return make_alloc_exact((unsigned long)page_address(p), order, size);
2702}
2703EXPORT_SYMBOL(alloc_pages_exact_nid);
2704
2705/**
2706 * free_pages_exact - release memory allocated via alloc_pages_exact()
2707 * @virt: the value returned by alloc_pages_exact.
2708 * @size: size of allocation, same value as passed to alloc_pages_exact().
2709 *
2710 * Release the memory allocated by a previous call to alloc_pages_exact.
2711 */
2712void free_pages_exact(void *virt, size_t size)
2713{
2714        unsigned long addr = (unsigned long)virt;
2715        unsigned long end = addr + PAGE_ALIGN(size);
2716
2717        while (addr < end) {
2718                free_page(addr);
2719                addr += PAGE_SIZE;
2720        }
2721}
2722EXPORT_SYMBOL(free_pages_exact);
2723
2724static unsigned int nr_free_zone_pages(int offset)
2725{
2726        struct zoneref *z;
2727        struct zone *zone;
2728
2729        /* Just pick one node, since fallback list is circular */
2730        unsigned int sum = 0;
2731
2732        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2733
2734        for_each_zone_zonelist(zone, z, zonelist, offset) {
2735                unsigned long size = zone->present_pages;
2736                unsigned long high = high_wmark_pages(zone);
2737                if (size > high)
2738                        sum += size - high;
2739        }
2740
2741        return sum;
2742}
2743
2744/*
2745 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2746 */
2747unsigned int nr_free_buffer_pages(void)
2748{
2749        return nr_free_zone_pages(gfp_zone(GFP_USER));
2750}
2751EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2752
2753/*
2754 * Amount of free RAM allocatable within all zones
2755 */
2756unsigned int nr_free_pagecache_pages(void)
2757{
2758        return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2759}
2760
2761static inline void show_node(struct zone *zone)
2762{
2763        if (NUMA_BUILD)
2764                printk("Node %d ", zone_to_nid(zone));
2765}
2766
2767void si_meminfo(struct sysinfo *val)
2768{
2769        val->totalram = totalram_pages;
2770        val->sharedram = 0;
2771        val->freeram = global_page_state(NR_FREE_PAGES);
2772        val->bufferram = nr_blockdev_pages();
2773        val->totalhigh = totalhigh_pages;
2774        val->freehigh = nr_free_highpages();
2775        val->mem_unit = PAGE_SIZE;
2776}
2777
2778EXPORT_SYMBOL(si_meminfo);
2779
2780#ifdef CONFIG_NUMA
2781void si_meminfo_node(struct sysinfo *val, int nid)
2782{
2783        pg_data_t *pgdat = NODE_DATA(nid);
2784
2785        val->totalram = pgdat->node_present_pages;
2786        val->freeram = node_page_state(nid, NR_FREE_PAGES);
2787#ifdef CONFIG_HIGHMEM
2788        val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2789        val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2790                        NR_FREE_PAGES);
2791#else
2792        val->totalhigh = 0;
2793        val->freehigh = 0;
2794#endif
2795        val->mem_unit = PAGE_SIZE;
2796}
2797#endif
2798
2799/*
2800 * Determine whether the node should be displayed or not, depending on whether
2801 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2802 */
2803bool skip_free_areas_node(unsigned int flags, int nid)
2804{
2805        bool ret = false;
2806        unsigned int cpuset_mems_cookie;
2807
2808        if (!(flags & SHOW_MEM_FILTER_NODES))
2809                goto out;
2810
2811        do {
2812                cpuset_mems_cookie = get_mems_allowed();
2813                ret = !node_isset(nid, cpuset_current_mems_allowed);
2814        } while (!put_mems_allowed(cpuset_mems_cookie));
2815out:
2816        return ret;
2817}
2818
2819#define K(x) ((x) << (PAGE_SHIFT-10))
2820
2821/*
2822 * Show free area list (used inside shift_scroll-lock stuff)
2823 * We also calculate the percentage fragmentation. We do this by counting the
2824 * memory on each free list with the exception of the first item on the list.
2825 * Suppresses nodes that are not allowed by current's cpuset if
2826 * SHOW_MEM_FILTER_NODES is passed.
2827 */
2828void show_free_areas(unsigned int filter)
2829{
2830        int cpu;
2831        struct zone *zone;
2832
2833        for_each_populated_zone(zone) {
2834                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2835                        continue;
2836                show_node(zone);
2837                printk("%s per-cpu:\n", zone->name);
2838
2839                for_each_online_cpu(cpu) {
2840                        struct per_cpu_pageset *pageset;
2841
2842                        pageset = per_cpu_ptr(zone->pageset, cpu);
2843
2844                        printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2845                               cpu, pageset->pcp.high,
2846                               pageset->pcp.batch, pageset->pcp.count);
2847                }
2848        }
2849
2850        printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2851                " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2852                " unevictable:%lu"
2853                " dirty:%lu writeback:%lu unstable:%lu\n"
2854                " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2855                " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2856                global_page_state(NR_ACTIVE_ANON),
2857                global_page_state(NR_INACTIVE_ANON),
2858                global_page_state(NR_ISOLATED_ANON),
2859                global_page_state(NR_ACTIVE_FILE),
2860                global_page_state(NR_INACTIVE_FILE),
2861                global_page_state(NR_ISOLATED_FILE),
2862                global_page_state(NR_UNEVICTABLE),
2863                global_page_state(NR_FILE_DIRTY),
2864                global_page_state(NR_WRITEBACK),
2865                global_page_state(NR_UNSTABLE_NFS),
2866                global_page_state(NR_FREE_PAGES),
2867                global_page_state(NR_SLAB_RECLAIMABLE),
2868                global_page_state(NR_SLAB_UNRECLAIMABLE),
2869                global_page_state(NR_FILE_MAPPED),
2870                global_page_state(NR_SHMEM),
2871                global_page_state(NR_PAGETABLE),
2872                global_page_state(NR_BOUNCE));
2873
2874        for_each_populated_zone(zone) {
2875                int i;
2876
2877                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2878                        continue;
2879                show_node(zone);
2880                printk("%s"
2881                        " free:%lukB"
2882                        " min:%lukB"
2883                        " low:%lukB"
2884                        " high:%lukB"
2885                        " active_anon:%lukB"
2886                        " inactive_anon:%lukB"
2887                        " active_file:%lukB"
2888                        " inactive_file:%lukB"
2889                        " unevictable:%lukB"
2890                        " isolated(anon):%lukB"
2891                        " isolated(file):%lukB"
2892                        " present:%lukB"
2893                        " mlocked:%lukB"
2894                        " dirty:%lukB"
2895                        " writeback:%lukB"
2896                        " mapped:%lukB"
2897                        " shmem:%lukB"
2898                        " slab_reclaimable:%lukB"
2899                        " slab_unreclaimable:%lukB"
2900                        " kernel_stack:%lukB"
2901                        " pagetables:%lukB"
2902                        " unstable:%lukB"
2903                        " bounce:%lukB"
2904                        " writeback_tmp:%lukB"
2905                        " pages_scanned:%lu"
2906                        " all_unreclaimable? %s"
2907                        "\n",
2908                        zone->name,
2909                        K(zone_page_state(zone, NR_FREE_PAGES)),
2910                        K(min_wmark_pages(zone)),
2911                        K(low_wmark_pages(zone)),
2912                        K(high_wmark_pages(zone)),
2913                        K(zone_page_state(zone, NR_ACTIVE_ANON)),
2914                        K(zone_page_state(zone, NR_INACTIVE_ANON)),
2915                        K(zone_page_state(zone, NR_ACTIVE_FILE)),
2916                        K(zone_page_state(zone, NR_INACTIVE_FILE)),
2917                        K(zone_page_state(zone, NR_UNEVICTABLE)),
2918                        K(zone_page_state(zone, NR_ISOLATED_ANON)),
2919                        K(zone_page_state(zone, NR_ISOLATED_FILE)),
2920                        K(zone->present_pages),
2921                        K(zone_page_state(zone, NR_MLOCK)),
2922                        K(zone_page_state(zone, NR_FILE_DIRTY)),
2923                        K(zone_page_state(zone, NR_WRITEBACK)),
2924                        K(zone_page_state(zone, NR_FILE_MAPPED)),
2925                        K(zone_page_state(zone, NR_SHMEM)),
2926                        K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2927                        K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2928                        zone_page_state(zone, NR_KERNEL_STACK) *
2929                                THREAD_SIZE / 1024,
2930                        K(zone_page_state(zone, NR_PAGETABLE)),
2931                        K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2932                        K(zone_page_state(zone, NR_BOUNCE)),
2933                        K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2934                        zone->pages_scanned,
2935                        (zone->all_unreclaimable ? "yes" : "no")
2936                        );
2937                printk("lowmem_reserve[]:");
2938                for (i = 0; i < MAX_NR_ZONES; i++)
2939                        printk(" %lu", zone->lowmem_reserve[i]);
2940                printk("\n");
2941        }
2942
2943        for_each_populated_zone(zone) {
2944                unsigned long nr[MAX_ORDER], flags, order, total = 0;
2945
2946                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2947                        continue;
2948                show_node(zone);
2949                printk("%s: ", zone->name);
2950
2951                spin_lock_irqsave(&zone->lock, flags);
2952                for (order = 0; order < MAX_ORDER; order++) {
2953                        nr[order] = zone->free_area[order].nr_free;
2954                        total += nr[order] << order;
2955                }
2956                spin_unlock_irqrestore(&zone->lock, flags);
2957                for (order = 0; order < MAX_ORDER; order++)
2958                        printk("%lu*%lukB ", nr[order], K(1UL) << order);
2959                printk("= %lukB\n", K(total));
2960        }
2961
2962        printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2963
2964        show_swap_cache_info();
2965}
2966
2967static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2968{
2969        zoneref->zone = zone;
2970        zoneref->zone_idx = zone_idx(zone);
2971}
2972
2973/*
2974 * Builds allocation fallback zone lists.
2975 *
2976 * Add all populated zones of a node to the zonelist.
2977 */
2978static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2979                                int nr_zones, enum zone_type zone_type)
2980{
2981        struct zone *zone;
2982
2983        BUG_ON(zone_type >= MAX_NR_ZONES);
2984        zone_type++;
2985
2986        do {
2987                zone_type--;
2988                zone = pgdat->node_zones + zone_type;
2989                if (populated_zone(zone)) {
2990                        zoneref_set_zone(zone,
2991                                &zonelist->_zonerefs[nr_zones++]);
2992                        check_highest_zone(zone_type);
2993                }
2994
2995        } while (zone_type);
2996        return nr_zones;
2997}
2998
2999
3000/*
3001 *  zonelist_order:
3002 *  0 = automatic detection of better ordering.
3003 *  1 = order by ([node] distance, -zonetype)
3004 *  2 = order by (-zonetype, [node] distance)
3005 *
3006 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3007 *  the same zonelist. So only NUMA can configure this param.
3008 */
3009#define ZONELIST_ORDER_DEFAULT  0
3010#define ZONELIST_ORDER_NODE     1
3011#define ZONELIST_ORDER_ZONE     2
3012
3013/* zonelist order in the kernel.
3014 * set_zonelist_order() will set this to NODE or ZONE.
3015 */
3016static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3017static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3018
3019
3020#ifdef CONFIG_NUMA
3021/* The value user specified ....changed by config */
3022static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3023/* string for sysctl */
3024#define NUMA_ZONELIST_ORDER_LEN 16
3025char numa_zonelist_order[16] = "default";
3026
3027/*
3028 * interface for configure zonelist ordering.
3029 * command line option "numa_zonelist_order"
3030 *      = "[dD]efault   - default, automatic configuration.
3031 *      = "[nN]ode      - order by node locality, then by zone within node
3032 *      = "[zZ]one      - order by zone, then by locality within zone
3033 */
3034
3035static int __parse_numa_zonelist_order(char *s)
3036{
3037        if (*s == 'd' || *s == 'D') {
3038                user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3039        } else if (*s == 'n' || *s == 'N') {
3040                user_zonelist_order = ZONELIST_ORDER_NODE;
3041        } else if (*s == 'z' || *s == 'Z') {
3042                user_zonelist_order = ZONELIST_ORDER_ZONE;
3043        } else {
3044                printk(KERN_WARNING
3045                        "Ignoring invalid numa_zonelist_order value:  "
3046                        "%s\n", s);
3047                return -EINVAL;
3048        }
3049        return 0;
3050}
3051
3052static __init int setup_numa_zonelist_order(char *s)
3053{
3054        int ret;
3055
3056        if (!s)
3057                return 0;
3058
3059        ret = __parse_numa_zonelist_order(s);
3060        if (ret == 0)
3061                strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3062
3063        return ret;
3064}
3065early_param("numa_zonelist_order", setup_numa_zonelist_order);
3066
3067/*
3068 * sysctl handler for numa_zonelist_order
3069 */
3070int numa_zonelist_order_handler(ctl_table *table, int write,
3071                void __user *buffer, size_t *length,
3072                loff_t *ppos)
3073{
3074        char saved_string[NUMA_ZONELIST_ORDER_LEN];
3075        int ret;
3076        static DEFINE_MUTEX(zl_order_mutex);
3077
3078        mutex_lock(&zl_order_mutex);
3079        if (write)
3080                strcpy(saved_string, (char*)table->data);
3081        ret = proc_dostring(table, write, buffer, length, ppos);
3082        if (ret)
3083                goto out;
3084        if (write) {
3085                int oldval = user_zonelist_order;
3086                if (__parse_numa_zonelist_order((char*)table->data)) {
3087                        /*
3088                         * bogus value.  restore saved string
3089                         */
3090                        strncpy((char*)table->data, saved_string,
3091                                NUMA_ZONELIST_ORDER_LEN);
3092                        user_zonelist_order = oldval;
3093                } else if (oldval != user_zonelist_order) {
3094                        mutex_lock(&zonelists_mutex);
3095                        build_all_zonelists(NULL, NULL);
3096                        mutex_unlock(&zonelists_mutex);
3097                }
3098        }
3099out:
3100        mutex_unlock(&zl_order_mutex);
3101        return ret;
3102}
3103
3104
3105#define MAX_NODE_LOAD (nr_online_nodes)
3106static int node_load[MAX_NUMNODES];
3107
3108/**
3109 * find_next_best_node - find the next node that should appear in a given node's fallback list
3110 * @node: node whose fallback list we're appending
3111 * @used_node_mask: nodemask_t of already used nodes
3112 *
3113 * We use a number of factors to determine which is the next node that should
3114 * appear on a given node's fallback list.  The node should not have appeared
3115 * already in @node's fallback list, and it should be the next closest node
3116 * according to the distance array (which contains arbitrary distance values
3117 * from each node to each node in the system), and should also prefer nodes
3118 * with no CPUs, since presumably they'll have very little allocation pressure
3119 * on them otherwise.
3120 * It returns -1 if no node is found.
3121 */
3122static int find_next_best_node(int node, nodemask_t *used_node_mask)
3123{
3124        int n, val;
3125        int min_val = INT_MAX;
3126        int best_node = -1;
3127        const struct cpumask *tmp = cpumask_of_node(0);
3128
3129        /* Use the local node if we haven't already */
3130        if (!node_isset(node, *used_node_mask)) {
3131                node_set(node, *used_node_mask);
3132                return node;
3133        }
3134
3135        for_each_node_state(n, N_HIGH_MEMORY) {
3136
3137                /* Don't want a node to appear more than once */
3138                if (node_isset(n, *used_node_mask))
3139                        continue;
3140
3141                /* Use the distance array to find the distance */
3142                val = node_distance(node, n);
3143
3144                /* Penalize nodes under us ("prefer the next node") */
3145                val += (n < node);
3146
3147                /* Give preference to headless and unused nodes */
3148                tmp = cpumask_of_node(n);
3149                if (!cpumask_empty(tmp))
3150                        val += PENALTY_FOR_NODE_WITH_CPUS;
3151
3152                /* Slight preference for less loaded node */
3153                val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3154                val += node_load[n];
3155
3156                if (val < min_val) {
3157                        min_val = val;
3158                        best_node = n;
3159                }
3160        }
3161
3162        if (best_node >= 0)
3163                node_set(best_node, *used_node_mask);
3164
3165        return best_node;
3166}
3167
3168
3169/*
3170 * Build zonelists ordered by node and zones within node.
3171 * This results in maximum locality--normal zone overflows into local
3172 * DMA zone, if any--but risks exhausting DMA zone.
3173 */
3174static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3175{
3176        int j;
3177        struct zonelist *zonelist;
3178
3179        zonelist = &pgdat->node_zonelists[0];
3180        for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3181                ;
3182        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3183                                                        MAX_NR_ZONES - 1);
3184        zonelist->_zonerefs[j].zone = NULL;
3185        zonelist->_zonerefs[j].zone_idx = 0;
3186}
3187
3188/*
3189 * Build gfp_thisnode zonelists
3190 */
3191static void build_thisnode_zonelists(pg_data_t *pgdat)
3192{
3193        int j;
3194        struct zonelist *zonelist;
3195
3196        zonelist = &pgdat->node_zonelists[1];
3197        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3198        zonelist->_zonerefs[j].zone = NULL;
3199        zonelist->_zonerefs[j].zone_idx = 0;
3200}
3201
3202/*
3203 * Build zonelists ordered by zone and nodes within zones.
3204 * This results in conserving DMA zone[s] until all Normal memory is
3205 * exhausted, but results in overflowing to remote node while memory
3206 * may still exist in local DMA zone.
3207 */
3208static int node_order[MAX_NUMNODES];
3209
3210static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3211{
3212        int pos, j, node;
3213        int zone_type;          /* needs to be signed */
3214        struct zone *z;
3215        struct zonelist *zonelist;
3216
3217        zonelist = &pgdat->node_zonelists[0];
3218        pos = 0;
3219        for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3220                for (j = 0; j < nr_nodes; j++) {
3221                        node = node_order[j];
3222                        z = &NODE_DATA(node)->node_zones[zone_type];
3223                        if (populated_zone(z)) {
3224                                zoneref_set_zone(z,
3225                                        &zonelist->_zonerefs[pos++]);
3226                                check_highest_zone(zone_type);
3227                        }
3228                }
3229        }
3230        zonelist->_zonerefs[pos].zone = NULL;
3231        zonelist->_zonerefs[pos].zone_idx = 0;
3232}
3233
3234static int default_zonelist_order(void)
3235{
3236        int nid, zone_type;
3237        unsigned long low_kmem_size,total_size;
3238        struct zone *z;
3239        int average_size;
3240        /*
3241         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3242         * If they are really small and used heavily, the system can fall
3243         * into OOM very easily.
3244         * This function detect ZONE_DMA/DMA32 size and configures zone order.
3245         */
3246        /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3247        low_kmem_size = 0;
3248        total_size = 0;
3249        for_each_online_node(nid) {
3250                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3251                        z = &NODE_DATA(nid)->node_zones[zone_type];
3252                        if (populated_zone(z)) {
3253                                if (zone_type < ZONE_NORMAL)
3254                                        low_kmem_size += z->present_pages;
3255                                total_size += z->present_pages;
3256                        } else if (zone_type == ZONE_NORMAL) {
3257                                /*
3258                                 * If any node has only lowmem, then node order
3259                                 * is preferred to allow kernel allocations
3260                                 * locally; otherwise, they can easily infringe
3261                                 * on other nodes when there is an abundance of
3262                                 * lowmem available to allocate from.
3263                                 */
3264                                return ZONELIST_ORDER_NODE;
3265                        }
3266                }
3267        }
3268        if (!low_kmem_size ||  /* there are no DMA area. */
3269            low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3270                return ZONELIST_ORDER_NODE;
3271        /*
3272         * look into each node's config.
3273         * If there is a node whose DMA/DMA32 memory is very big area on
3274         * local memory, NODE_ORDER may be suitable.
3275         */
3276        average_size = total_size /
3277                                (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3278        for_each_online_node(nid) {
3279                low_kmem_size = 0;
3280                total_size = 0;
3281                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3282                        z = &NODE_DATA(nid)->node_zones[zone_type];
3283                        if (populated_zone(z)) {
3284                                if (zone_type < ZONE_NORMAL)
3285                                        low_kmem_size += z->present_pages;
3286                                total_size += z->present_pages;
3287                        }
3288                }
3289                if (low_kmem_size &&
3290                    total_size > average_size && /* ignore small node */
3291                    low_kmem_size > total_size * 70/100)
3292                        return ZONELIST_ORDER_NODE;
3293        }
3294        return ZONELIST_ORDER_ZONE;
3295}
3296
3297static void set_zonelist_order(void)
3298{
3299        if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3300                current_zonelist_order = default_zonelist_order();
3301        else
3302                current_zonelist_order = user_zonelist_order;
3303}
3304
3305static void build_zonelists(pg_data_t *pgdat)
3306{
3307        int j, node, load;
3308        enum zone_type i;
3309        nodemask_t used_mask;
3310        int local_node, prev_node;
3311        struct zonelist *zonelist;
3312        int order = current_zonelist_order;
3313
3314        /* initialize zonelists */
3315        for (i = 0; i < MAX_ZONELISTS; i++) {
3316                zonelist = pgdat->node_zonelists + i;
3317                zonelist->_zonerefs[0].zone = NULL;
3318                zonelist->_zonerefs[0].zone_idx = 0;
3319        }
3320
3321        /* NUMA-aware ordering of nodes */
3322        local_node = pgdat->node_id;
3323        load = nr_online_nodes;
3324        prev_node = local_node;
3325        nodes_clear(used_mask);
3326
3327        memset(node_order, 0, sizeof(node_order));
3328        j = 0;
3329
3330        while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3331                int distance = node_distance(local_node, node);
3332
3333                /*
3334                 * If another node is sufficiently far away then it is better
3335                 * to reclaim pages in a zone before going off node.
3336                 */
3337                if (distance > RECLAIM_DISTANCE)
3338                        zone_reclaim_mode = 1;
3339
3340                /*
3341                 * We don't want to pressure a particular node.
3342                 * So adding penalty to the first node in same
3343                 * distance group to make it round-robin.
3344                 */
3345                if (distance != node_distance(local_node, prev_node))
3346                        node_load[node] = load;
3347
3348                prev_node = node;
3349                load--;
3350                if (order == ZONELIST_ORDER_NODE)
3351                        build_zonelists_in_node_order(pgdat, node);
3352                else
3353                        node_order[j++] = node; /* remember order */
3354        }
3355
3356        if (order == ZONELIST_ORDER_ZONE) {
3357                /* calculate node order -- i.e., DMA last! */
3358                build_zonelists_in_zone_order(pgdat, j);
3359        }
3360
3361        build_thisnode_zonelists(pgdat);
3362}
3363
3364/* Construct the zonelist performance cache - see further mmzone.h */
3365static void build_zonelist_cache(pg_data_t *pgdat)
3366{
3367        struct zonelist *zonelist;
3368        struct zonelist_cache *zlc;
3369        struct zoneref *z;
3370
3371        zonelist = &pgdat->node_zonelists[0];
3372        zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3373        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3374        for (z = zonelist->_zonerefs; z->zone; z++)
3375                zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3376}
3377
3378#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3379/*
3380 * Return node id of node used for "local" allocations.
3381 * I.e., first node id of first zone in arg node's generic zonelist.
3382 * Used for initializing percpu 'numa_mem', which is used primarily
3383 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3384 */
3385int local_memory_node(int node)
3386{
3387        struct zone *zone;
3388
3389        (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3390                                   gfp_zone(GFP_KERNEL),
3391                                   NULL,
3392                                   &zone);
3393        return zone->node;
3394}
3395#endif
3396
3397#else   /* CONFIG_NUMA */
3398
3399static void set_zonelist_order(void)
3400{
3401        current_zonelist_order = ZONELIST_ORDER_ZONE;
3402}
3403
3404static void build_zonelists(pg_data_t *pgdat)
3405{
3406        int node, local_node;
3407        enum zone_type j;
3408        struct zonelist *zonelist;
3409
3410        local_node = pgdat->node_id;
3411
3412        zonelist = &pgdat->node_zonelists[0];
3413        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3414
3415        /*
3416         * Now we build the zonelist so that it contains the zones
3417         * of all the other nodes.
3418         * We don't want to pressure a particular node, so when
3419         * building the zones for node N, we make sure that the
3420         * zones coming right after the local ones are those from
3421         * node N+1 (modulo N)
3422         */
3423        for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3424                if (!node_online(node))
3425                        continue;
3426                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3427                                                        MAX_NR_ZONES - 1);
3428        }
3429        for (node = 0; node < local_node; node++) {
3430                if (!node_online(node))
3431                        continue;
3432                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3433                                                        MAX_NR_ZONES - 1);
3434        }
3435
3436        zonelist->_zonerefs[j].zone = NULL;
3437        zonelist->_zonerefs[j].zone_idx = 0;
3438}
3439
3440/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3441static void build_zonelist_cache(pg_data_t *pgdat)
3442{
3443        pgdat->node_zonelists[0].zlcache_ptr = NULL;
3444}
3445
3446#endif  /* CONFIG_NUMA */
3447
3448/*
3449 * Boot pageset table. One per cpu which is going to be used for all
3450 * zones and all nodes. The parameters will be set in such a way
3451 * that an item put on a list will immediately be handed over to
3452 * the buddy list. This is safe since pageset manipulation is done
3453 * with interrupts disabled.
3454 *
3455 * The boot_pagesets must be kept even after bootup is complete for
3456 * unused processors and/or zones. They do play a role for bootstrapping
3457 * hotplugged processors.
3458 *
3459 * zoneinfo_show() and maybe other functions do
3460 * not check if the processor is online before following the pageset pointer.
3461 * Other parts of the kernel may not check if the zone is available.
3462 */
3463static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3464static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3465static void setup_zone_pageset(struct zone *zone);
3466
3467/*
3468 * Global mutex to protect against size modification of zonelists
3469 * as well as to serialize pageset setup for the new populated zone.
3470 */
3471DEFINE_MUTEX(zonelists_mutex);
3472
3473/* return values int ....just for stop_machine() */
3474static int __build_all_zonelists(void *data)
3475{
3476        int nid;
3477        int cpu;
3478        pg_data_t *self = data;
3479
3480#ifdef CONFIG_NUMA
3481        memset(node_load, 0, sizeof(node_load));
3482#endif
3483
3484        if (self && !node_online(self->node_id)) {
3485                build_zonelists(self);
3486                build_zonelist_cache(self);
3487        }
3488
3489        for_each_online_node(nid) {
3490                pg_data_t *pgdat = NODE_DATA(nid);
3491
3492                build_zonelists(pgdat);
3493                build_zonelist_cache(pgdat);
3494        }
3495
3496        /*
3497         * Initialize the boot_pagesets that are going to be used
3498         * for bootstrapping processors. The real pagesets for
3499         * each zone will be allocated later when the per cpu
3500         * allocator is available.
3501         *
3502         * boot_pagesets are used also for bootstrapping offline
3503         * cpus if the system is already booted because the pagesets
3504         * are needed to initialize allocators on a specific cpu too.
3505         * F.e. the percpu allocator needs the page allocator which
3506         * needs the percpu allocator in order to allocate its pagesets
3507         * (a chicken-egg dilemma).
3508         */
3509        for_each_possible_cpu(cpu) {
3510                setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3511
3512#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3513                /*
3514                 * We now know the "local memory node" for each node--
3515                 * i.e., the node of the first zone in the generic zonelist.
3516                 * Set up numa_mem percpu variable for on-line cpus.  During
3517                 * boot, only the boot cpu should be on-line;  we'll init the
3518                 * secondary cpus' numa_mem as they come on-line.  During
3519                 * node/memory hotplug, we'll fixup all on-line cpus.
3520                 */
3521                if (cpu_online(cpu))
3522                        set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3523#endif
3524        }
3525
3526        return 0;
3527}
3528
3529/*
3530 * Called with zonelists_mutex held always
3531 * unless system_state == SYSTEM_BOOTING.
3532 */
3533void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3534{
3535        set_zonelist_order();
3536
3537        if (system_state == SYSTEM_BOOTING) {
3538                __build_all_zonelists(NULL);
3539                mminit_verify_zonelist();
3540                cpuset_init_current_mems_allowed();
3541        } else {
3542                /* we have to stop all cpus to guarantee there is no user
3543                   of zonelist */
3544#ifdef CONFIG_MEMORY_HOTPLUG
3545                if (zone)
3546                        setup_zone_pageset(zone);
3547#endif
3548                stop_machine(__build_all_zonelists, pgdat, NULL);
3549                /* cpuset refresh routine should be here */
3550        }
3551        vm_total_pages = nr_free_pagecache_pages();
3552        /*
3553         * Disable grouping by mobility if the number of pages in the
3554         * system is too low to allow the mechanism to work. It would be
3555         * more accurate, but expensive to check per-zone. This check is
3556         * made on memory-hotadd so a system can start with mobility
3557         * disabled and enable it later
3558         */
3559        if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3560                page_group_by_mobility_disabled = 1;
3561        else
3562                page_group_by_mobility_disabled = 0;
3563
3564        printk("Built %i zonelists in %s order, mobility grouping %s.  "
3565                "Total pages: %ld\n",
3566                        nr_online_nodes,
3567                        zonelist_order_name[current_zonelist_order],
3568                        page_group_by_mobility_disabled ? "off" : "on",
3569                        vm_total_pages);
3570#ifdef CONFIG_NUMA
3571        printk("Policy zone: %s\n", zone_names[policy_zone]);
3572#endif
3573}
3574
3575/*
3576 * Helper functions to size the waitqueue hash table.
3577 * Essentially these want to choose hash table sizes sufficiently
3578 * large so that collisions trying to wait on pages are rare.
3579 * But in fact, the number of active page waitqueues on typical
3580 * systems is ridiculously low, less than 200. So this is even
3581 * conservative, even though it seems large.
3582 *
3583 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3584 * waitqueues, i.e. the size of the waitq table given the number of pages.
3585 */
3586#define PAGES_PER_WAITQUEUE     256
3587
3588#ifndef CONFIG_MEMORY_HOTPLUG
3589static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3590{
3591        unsigned long size = 1;
3592
3593        pages /= PAGES_PER_WAITQUEUE;
3594
3595        while (size < pages)
3596                size <<= 1;
3597
3598        /*
3599         * Once we have dozens or even hundreds of threads sleeping
3600         * on IO we've got bigger problems than wait queue collision.
3601         * Limit the size of the wait table to a reasonable size.
3602         */
3603        size = min(size, 4096UL);
3604
3605        return max(size, 4UL);
3606}
3607#else
3608/*
3609 * A zone's size might be changed by hot-add, so it is not possible to determine
3610 * a suitable size for its wait_table.  So we use the maximum size now.
3611 *
3612 * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
3613 *
3614 *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
3615 *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3616 *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
3617 *
3618 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3619 * or more by the traditional way. (See above).  It equals:
3620 *
3621 *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
3622 *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
3623 *    powerpc (64K page size)             : =  (32G +16M)byte.
3624 */
3625static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3626{
3627        return 4096UL;
3628}
3629#endif
3630
3631/*
3632 * This is an integer logarithm so that shifts can be used later
3633 * to extract the more random high bits from the multiplicative
3634 * hash function before the remainder is taken.
3635 */
3636static inline unsigned long wait_table_bits(unsigned long size)
3637{
3638        return ffz(~size);
3639}
3640
3641#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3642
3643/*
3644 * Check if a pageblock contains reserved pages
3645 */
3646static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3647{
3648        unsigned long pfn;
3649
3650        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3651                if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3652                        return 1;
3653        }
3654        return 0;
3655}
3656
3657/*
3658 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3659 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3660 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3661 * higher will lead to a bigger reserve which will get freed as contiguous
3662 * blocks as reclaim kicks in
3663 */
3664static void setup_zone_migrate_reserve(struct zone *zone)
3665{
3666        unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3667        struct page *page;
3668        unsigned long block_migratetype;
3669        int reserve;
3670
3671        /*
3672         * Get the start pfn, end pfn and the number of blocks to reserve
3673         * We have to be careful to be aligned to pageblock_nr_pages to
3674         * make sure that we always check pfn_valid for the first page in
3675         * the block.
3676         */
3677        start_pfn = zone->zone_start_pfn;
3678        end_pfn = start_pfn + zone->spanned_pages;
3679        start_pfn = roundup(start_pfn, pageblock_nr_pages);
3680        reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3681                                                        pageblock_order;
3682
3683        /*
3684         * Reserve blocks are generally in place to help high-order atomic
3685         * allocations that are short-lived. A min_free_kbytes value that
3686         * would result in more than 2 reserve blocks for atomic allocations
3687         * is assumed to be in place to help anti-fragmentation for the
3688         * future allocation of hugepages at runtime.
3689         */
3690        reserve = min(2, reserve);
3691
3692        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3693                if (!pfn_valid(pfn))
3694                        continue;
3695                page = pfn_to_page(pfn);
3696
3697                /* Watch out for overlapping nodes */
3698                if (page_to_nid(page) != zone_to_nid(zone))
3699                        continue;
3700
3701                block_migratetype = get_pageblock_migratetype(page);
3702
3703                /* Only test what is necessary when the reserves are not met */
3704                if (reserve > 0) {
3705                        /*
3706                         * Blocks with reserved pages will never free, skip
3707                         * them.
3708                         */
3709                        block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3710                        if (pageblock_is_reserved(pfn, block_end_pfn))
3711                                continue;
3712
3713                        /* If this block is reserved, account for it */
3714                        if (block_migratetype == MIGRATE_RESERVE) {
3715                                reserve--;
3716                                continue;
3717                        }
3718
3719                        /* Suitable for reserving if this block is movable */
3720                        if (block_migratetype == MIGRATE_MOVABLE) {
3721                                set_pageblock_migratetype(page,
3722                                                        MIGRATE_RESERVE);
3723                                move_freepages_block(zone, page,
3724                                                        MIGRATE_RESERVE);
3725                                reserve--;
3726                                continue;
3727                        }
3728                }
3729
3730                /*
3731                 * If the reserve is met and this is a previous reserved block,
3732                 * take it back
3733                 */
3734                if (block_migratetype == MIGRATE_RESERVE) {
3735                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3736                        move_freepages_block(zone, page, MIGRATE_MOVABLE);
3737                }
3738        }
3739}
3740
3741/*
3742 * Initially all pages are reserved - free ones are freed
3743 * up by free_all_bootmem() once the early boot process is
3744 * done. Non-atomic initialization, single-pass.
3745 */
3746void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3747                unsigned long start_pfn, enum memmap_context context)
3748{
3749        struct page *page;
3750        unsigned long end_pfn = start_pfn + size;
3751        unsigned long pfn;
3752        struct zone *z;
3753
3754        if (highest_memmap_pfn < end_pfn - 1)
3755                highest_memmap_pfn = end_pfn - 1;
3756
3757        z = &NODE_DATA(nid)->node_zones[zone];
3758        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3759                /*
3760                 * There can be holes in boot-time mem_map[]s
3761                 * handed to this function.  They do not
3762                 * exist on hotplugged memory.
3763                 */
3764                if (context == MEMMAP_EARLY) {
3765                        if (!early_pfn_valid(pfn))
3766                                continue;
3767                        if (!early_pfn_in_nid(pfn, nid))
3768                                continue;
3769                }
3770                page = pfn_to_page(pfn);
3771                set_page_links(page, zone, nid, pfn);
3772                mminit_verify_page_links(page, zone, nid, pfn);
3773                init_page_count(page);
3774                reset_page_mapcount(page);
3775                SetPageReserved(page);
3776                /*
3777                 * Mark the block movable so that blocks are reserved for
3778                 * movable at startup. This will force kernel allocations
3779                 * to reserve their blocks rather than leaking throughout
3780                 * the address space during boot when many long-lived
3781                 * kernel allocations are made. Later some blocks near
3782                 * the start are marked MIGRATE_RESERVE by
3783                 * setup_zone_migrate_reserve()
3784                 *
3785                 * bitmap is created for zone's valid pfn range. but memmap
3786                 * can be created for invalid pages (for alignment)
3787                 * check here not to call set_pageblock_migratetype() against
3788                 * pfn out of zone.
3789                 */
3790                if ((z->zone_start_pfn <= pfn)
3791                    && (pfn < z->zone_start_pfn + z->spanned_pages)
3792                    && !(pfn & (pageblock_nr_pages - 1)))
3793                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3794
3795                INIT_LIST_HEAD(&page->lru);
3796#ifdef WANT_PAGE_VIRTUAL
3797                /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3798                if (!is_highmem_idx(zone))
3799                        set_page_address(page, __va(pfn << PAGE_SHIFT));
3800#endif
3801        }
3802}
3803
3804static void __meminit zone_init_free_lists(struct zone *zone)
3805{
3806        int order, t;
3807        for_each_migratetype_order(order, t) {
3808                INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3809                zone->free_area[order].nr_free = 0;
3810        }
3811}
3812
3813#ifndef __HAVE_ARCH_MEMMAP_INIT
3814#define memmap_init(size, nid, zone, start_pfn) \
3815        memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3816#endif
3817
3818static int __meminit zone_batchsize(struct zone *zone)
3819{
3820#ifdef CONFIG_MMU
3821        int batch;
3822
3823        /*
3824         * The per-cpu-pages pools are set to around 1000th of the
3825         * size of the zone.  But no more than 1/2 of a meg.
3826         *
3827         * OK, so we don't know how big the cache is.  So guess.
3828         */
3829        batch = zone->present_pages / 1024;
3830        if (batch * PAGE_SIZE > 512 * 1024)
3831                batch = (512 * 1024) / PAGE_SIZE;
3832        batch /= 4;             /* We effectively *= 4 below */
3833        if (batch < 1)
3834                batch = 1;
3835
3836        /*
3837         * Clamp the batch to a 2^n - 1 value. Having a power
3838         * of 2 value was found to be more likely to have
3839         * suboptimal cache aliasing properties in some cases.
3840         *
3841         * For example if 2 tasks are alternately allocating
3842         * batches of pages, one task can end up with a lot
3843         * of pages of one half of the possible page colors
3844         * and the other with pages of the other colors.
3845         */
3846        batch = rounddown_pow_of_two(batch + batch/2) - 1;
3847
3848        return batch;
3849
3850#else
3851        /* The deferral and batching of frees should be suppressed under NOMMU
3852         * conditions.
3853         *
3854         * The problem is that NOMMU needs to be able to allocate large chunks
3855         * of contiguous memory as there's no hardware page translation to
3856         * assemble apparent contiguous memory from discontiguous pages.
3857         *
3858         * Queueing large contiguous runs of pages for batching, however,
3859         * causes the pages to actually be freed in smaller chunks.  As there
3860         * can be a significant delay between the individual batches being
3861         * recycled, this leads to the once large chunks of space being
3862         * fragmented and becoming unavailable for high-order allocations.
3863         */
3864        return 0;
3865#endif
3866}
3867
3868static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3869{
3870        struct per_cpu_pages *pcp;
3871        int migratetype;
3872
3873        memset(p, 0, sizeof(*p));
3874
3875        pcp = &p->pcp;
3876        pcp->count = 0;
3877        pcp->high = 6 * batch;
3878        pcp->batch = max(1UL, 1 * batch);
3879        for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3880                INIT_LIST_HEAD(&pcp->lists[migratetype]);
3881}
3882
3883/*
3884 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3885 * to the value high for the pageset p.
3886 */
3887
3888static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3889                                unsigned long high)
3890{
3891        struct per_cpu_pages *pcp;
3892
3893        pcp = &p->pcp;
3894        pcp->high = high;
3895        pcp->batch = max(1UL, high/4);
3896        if ((high/4) > (PAGE_SHIFT * 8))
3897                pcp->batch = PAGE_SHIFT * 8;
3898}
3899
3900static void __meminit setup_zone_pageset(struct zone *zone)
3901{
3902        int cpu;
3903
3904        zone->pageset = alloc_percpu(struct per_cpu_pageset);
3905
3906        for_each_possible_cpu(cpu) {
3907                struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3908
3909                setup_pageset(pcp, zone_batchsize(zone));
3910
3911                if (percpu_pagelist_fraction)
3912                        setup_pagelist_highmark(pcp,
3913                                (zone->present_pages /
3914                                        percpu_pagelist_fraction));
3915        }
3916}
3917
3918/*
3919 * Allocate per cpu pagesets and initialize them.
3920 * Before this call only boot pagesets were available.
3921 */
3922void __init setup_per_cpu_pageset(void)
3923{
3924        struct zone *zone;
3925
3926        for_each_populated_zone(zone)
3927                setup_zone_pageset(zone);
3928}
3929
3930static noinline __init_refok
3931int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3932{
3933        int i;
3934        struct pglist_data *pgdat = zone->zone_pgdat;
3935        size_t alloc_size;
3936
3937        /*
3938         * The per-page waitqueue mechanism uses hashed waitqueues
3939         * per zone.
3940         */
3941        zone->wait_table_hash_nr_entries =
3942                 wait_table_hash_nr_entries(zone_size_pages);
3943        zone->wait_table_bits =
3944                wait_table_bits(zone->wait_table_hash_nr_entries);
3945        alloc_size = zone->wait_table_hash_nr_entries
3946                                        * sizeof(wait_queue_head_t);
3947
3948        if (!slab_is_available()) {
3949                zone->wait_table = (wait_queue_head_t *)
3950                        alloc_bootmem_node_nopanic(pgdat, alloc_size);
3951        } else {
3952                /*
3953                 * This case means that a zone whose size was 0 gets new memory
3954                 * via memory hot-add.
3955                 * But it may be the case that a new node was hot-added.  In
3956                 * this case vmalloc() will not be able to use this new node's
3957                 * memory - this wait_table must be initialized to use this new
3958                 * node itself as well.
3959                 * To use this new node's memory, further consideration will be
3960                 * necessary.
3961                 */
3962                zone->wait_table = vmalloc(alloc_size);
3963        }
3964        if (!zone->wait_table)
3965                return -ENOMEM;
3966
3967        for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3968                init_waitqueue_head(zone->wait_table + i);
3969
3970        return 0;
3971}
3972
3973static __meminit void zone_pcp_init(struct zone *zone)
3974{
3975        /*
3976         * per cpu subsystem is not up at this point. The following code
3977         * relies on the ability of the linker to provide the
3978         * offset of a (static) per cpu variable into the per cpu area.
3979         */
3980        zone->pageset = &boot_pageset;
3981
3982        if (zone->present_pages)
3983                printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
3984                        zone->name, zone->present_pages,
3985                                         zone_batchsize(zone));
3986}
3987
3988int __meminit init_currently_empty_zone(struct zone *zone,
3989                                        unsigned long zone_start_pfn,
3990                                        unsigned long size,
3991                                        enum memmap_context context)
3992{
3993        struct pglist_data *pgdat = zone->zone_pgdat;
3994        int ret;
3995        ret = zone_wait_table_init(zone, size);
3996        if (ret)
3997                return ret;
3998        pgdat->nr_zones = zone_idx(zone) + 1;
3999
4000        zone->zone_start_pfn = zone_start_pfn;
4001
4002        mminit_dprintk(MMINIT_TRACE, "memmap_init",
4003                        "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4004                        pgdat->node_id,
4005                        (unsigned long)zone_idx(zone),
4006                        zone_start_pfn, (zone_start_pfn + size));
4007
4008        zone_init_free_lists(zone);
4009
4010        return 0;
4011}
4012
4013#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4014#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4015/*
4016 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4017 * Architectures may implement their own version but if add_active_range()
4018 * was used and there are no special requirements, this is a convenient
4019 * alternative
4020 */
4021int __meminit __early_pfn_to_nid(unsigned long pfn)
4022{
4023        unsigned long start_pfn, end_pfn;
4024        int i, nid;
4025
4026        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4027                if (start_pfn <= pfn && pfn < end_pfn)
4028                        return nid;
4029        /* This is a memory hole */
4030        return -1;
4031}
4032#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4033
4034int __meminit early_pfn_to_nid(unsigned long pfn)
4035{
4036        int nid;
4037
4038        nid = __early_pfn_to_nid(pfn);
4039        if (nid >= 0)
4040                return nid;
4041        /* just returns 0 */
4042        return 0;
4043}
4044
4045#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4046bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4047{
4048        int nid;
4049
4050        nid = __early_pfn_to_nid(pfn);
4051        if (nid >= 0 && nid != node)
4052                return false;
4053        return true;
4054}
4055#endif
4056
4057/**
4058 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4059 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4060 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4061 *
4062 * If an architecture guarantees that all ranges registered with
4063 * add_active_ranges() contain no holes and may be freed, this
4064 * this function may be used instead of calling free_bootmem() manually.
4065 */
4066void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4067{
4068        unsigned long start_pfn, end_pfn;
4069        int i, this_nid;
4070
4071        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4072                start_pfn = min(start_pfn, max_low_pfn);
4073                end_pfn = min(end_pfn, max_low_pfn);
4074
4075                if (start_pfn < end_pfn)
4076                        free_bootmem_node(NODE_DATA(this_nid),
4077                                          PFN_PHYS(start_pfn),
4078                                          (end_pfn - start_pfn) << PAGE_SHIFT);
4079        }
4080}
4081
4082/**
4083 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4084 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4085 *
4086 * If an architecture guarantees that all ranges registered with
4087 * add_active_ranges() contain no holes and may be freed, this
4088 * function may be used instead of calling memory_present() manually.
4089 */
4090void __init sparse_memory_present_with_active_regions(int nid)
4091{
4092        unsigned long start_pfn, end_pfn;
4093        int i, this_nid;
4094
4095        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4096                memory_present(this_nid, start_pfn, end_pfn);
4097}
4098
4099/**
4100 * get_pfn_range_for_nid - Return the start and end page frames for a node
4101 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4102 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4103 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4104 *
4105 * It returns the start and end page frame of a node based on information
4106 * provided by an arch calling add_active_range(). If called for a node
4107 * with no available memory, a warning is printed and the start and end
4108 * PFNs will be 0.
4109 */
4110void __meminit get_pfn_range_for_nid(unsigned int nid,
4111                        unsigned long *start_pfn, unsigned long *end_pfn)
4112{
4113        unsigned long this_start_pfn, this_end_pfn;
4114        int i;
4115
4116        *start_pfn = -1UL;
4117        *end_pfn = 0;
4118
4119        for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4120                *start_pfn = min(*start_pfn, this_start_pfn);
4121                *end_pfn = max(*end_pfn, this_end_pfn);
4122        }
4123
4124        if (*start_pfn == -1UL)
4125                *start_pfn = 0;
4126}
4127
4128/*
4129 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4130 * assumption is made that zones within a node are ordered in monotonic
4131 * increasing memory addresses so that the "highest" populated zone is used
4132 */
4133static void __init find_usable_zone_for_movable(void)
4134{
4135        int zone_index;
4136        for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4137                if (zone_index == ZONE_MOVABLE)
4138                        continue;
4139
4140                if (arch_zone_highest_possible_pfn[zone_index] >
4141                                arch_zone_lowest_possible_pfn[zone_index])
4142                        break;
4143        }
4144
4145        VM_BUG_ON(zone_index == -1);
4146        movable_zone = zone_index;
4147}
4148
4149/*
4150 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4151 * because it is sized independent of architecture. Unlike the other zones,
4152 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4153 * in each node depending on the size of each node and how evenly kernelcore
4154 * is distributed. This helper function adjusts the zone ranges
4155 * provided by the architecture for a given node by using the end of the
4156 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4157 * zones within a node are in order of monotonic increases memory addresses
4158 */
4159static void __meminit adjust_zone_range_for_zone_movable(int nid,
4160                                        unsigned long zone_type,
4161                                        unsigned long node_start_pfn,
4162                                        unsigned long node_end_pfn,
4163                                        unsigned long *zone_start_pfn,
4164                                        unsigned long *zone_end_pfn)
4165{
4166        /* Only adjust if ZONE_MOVABLE is on this node */
4167        if (zone_movable_pfn[nid]) {
4168                /* Size ZONE_MOVABLE */
4169                if (zone_type == ZONE_MOVABLE) {
4170                        *zone_start_pfn = zone_movable_pfn[nid];
4171                        *zone_end_pfn = min(node_end_pfn,
4172                                arch_zone_highest_possible_pfn[movable_zone]);
4173
4174                /* Adjust for ZONE_MOVABLE starting within this range */
4175                } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4176                                *zone_end_pfn > zone_movable_pfn[nid]) {
4177                        *zone_end_pfn = zone_movable_pfn[nid];
4178
4179                /* Check if this whole range is within ZONE_MOVABLE */
4180                } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4181                        *zone_start_pfn = *zone_end_pfn;
4182        }
4183}
4184
4185/*
4186 * Return the number of pages a zone spans in a node, including holes
4187 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4188 */
4189static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4190                                        unsigned long zone_type,
4191                                        unsigned long *ignored)
4192{
4193        unsigned long node_start_pfn, node_end_pfn;
4194        unsigned long zone_start_pfn, zone_end_pfn;
4195
4196        /* Get the start and end of the node and zone */
4197        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4198        zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4199        zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4200        adjust_zone_range_for_zone_movable(nid, zone_type,
4201                                node_start_pfn, node_end_pfn,
4202                                &zone_start_pfn, &zone_end_pfn);
4203
4204        /* Check that this node has pages within the zone's required range */
4205        if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4206                return 0;
4207
4208        /* Move the zone boundaries inside the node if necessary */
4209        zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4210        zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4211
4212        /* Return the spanned pages */
4213        return zone_end_pfn - zone_start_pfn;
4214}
4215
4216/*
4217 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4218 * then all holes in the requested range will be accounted for.
4219 */
4220unsigned long __meminit __absent_pages_in_range(int nid,
4221                                unsigned long range_start_pfn,
4222                                unsigned long range_end_pfn)
4223{
4224        unsigned long nr_absent = range_end_pfn - range_start_pfn;
4225        unsigned long start_pfn, end_pfn;
4226        int i;
4227
4228        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4229                start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4230                end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4231                nr_absent -= end_pfn - start_pfn;
4232        }
4233        return nr_absent;
4234}
4235
4236/**
4237 * absent_pages_in_range - Return number of page frames in holes within a range
4238 * @start_pfn: The start PFN to start searching for holes
4239 * @end_pfn: The end PFN to stop searching for holes
4240 *
4241 * It returns the number of pages frames in memory holes within a range.
4242 */
4243unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4244                                                        unsigned long end_pfn)
4245{
4246        return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4247}
4248
4249/* Return the number of page frames in holes in a zone on a node */
4250static unsigned long __meminit zone_absent_pages_in_node(int nid,
4251                                        unsigned long zone_type,
4252                                        unsigned long *ignored)
4253{
4254        unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4255        unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4256        unsigned long node_start_pfn, node_end_pfn;
4257        unsigned long zone_start_pfn, zone_end_pfn;
4258
4259        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4260        zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4261        zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4262
4263        adjust_zone_range_for_zone_movable(nid, zone_type,
4264                        node_start_pfn, node_end_pfn,
4265                        &zone_start_pfn, &zone_end_pfn);
4266        return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4267}
4268
4269#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4270static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4271                                        unsigned long zone_type,
4272                                        unsigned long *zones_size)
4273{
4274        return zones_size[zone_type];
4275}
4276
4277static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4278                                                unsigned long zone_type,
4279                                                unsigned long *zholes_size)
4280{
4281        if (!zholes_size)
4282                return 0;
4283
4284        return zholes_size[zone_type];
4285}
4286
4287#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4288
4289static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4290                unsigned long *zones_size, unsigned long *zholes_size)
4291{
4292        unsigned long realtotalpages, totalpages = 0;
4293        enum zone_type i;
4294
4295        for (i = 0; i < MAX_NR_ZONES; i++)
4296                totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4297                                                                zones_size);
4298        pgdat->node_spanned_pages = totalpages;
4299
4300        realtotalpages = totalpages;
4301        for (i = 0; i < MAX_NR_ZONES; i++)
4302                realtotalpages -=
4303                        zone_absent_pages_in_node(pgdat->node_id, i,
4304                                                                zholes_size);
4305        pgdat->node_present_pages = realtotalpages;
4306        printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4307                                                        realtotalpages);
4308}
4309
4310#ifndef CONFIG_SPARSEMEM
4311/*
4312 * Calculate the size of the zone->blockflags rounded to an unsigned long
4313 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4314 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4315 * round what is now in bits to nearest long in bits, then return it in
4316 * bytes.
4317 */
4318static unsigned long __init usemap_size(unsigned long zonesize)
4319{
4320        unsigned long usemapsize;
4321
4322        usemapsize = roundup(zonesize, pageblock_nr_pages);
4323        usemapsize = usemapsize >> pageblock_order;
4324        usemapsize *= NR_PAGEBLOCK_BITS;
4325        usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4326
4327        return usemapsize / 8;
4328}
4329
4330static void __init setup_usemap(struct pglist_data *pgdat,
4331                                struct zone *zone, unsigned long zonesize)
4332{
4333        unsigned long usemapsize = usemap_size(zonesize);
4334        zone->pageblock_flags = NULL;
4335        if (usemapsize)
4336                zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4337                                                                   usemapsize);
4338}
4339#else
4340static inline void setup_usemap(struct pglist_data *pgdat,
4341                                struct zone *zone, unsigned long zonesize) {}
4342#endif /* CONFIG_SPARSEMEM */
4343
4344#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4345
4346/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4347void __init set_pageblock_order(void)
4348{
4349        unsigned int order;
4350
4351        /* Check that pageblock_nr_pages has not already been setup */
4352        if (pageblock_order)
4353                return;
4354
4355        if (HPAGE_SHIFT > PAGE_SHIFT)
4356                order = HUGETLB_PAGE_ORDER;
4357        else
4358                order = MAX_ORDER - 1;
4359
4360        /*
4361         * Assume the largest contiguous order of interest is a huge page.
4362         * This value may be variable depending on boot parameters on IA64 and
4363         * powerpc.
4364         */
4365        pageblock_order = order;
4366}
4367#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4368
4369/*
4370 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4371 * is unused as pageblock_order is set at compile-time. See
4372 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4373 * the kernel config
4374 */
4375void __init set_pageblock_order(void)
4376{
4377}
4378
4379#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4380
4381/*
4382 * Set up the zone data structures:
4383 *   - mark all pages reserved
4384 *   - mark all memory queues empty
4385 *   - clear the memory bitmaps
4386 *
4387 * NOTE: pgdat should get zeroed by caller.
4388 */
4389static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4390                unsigned long *zones_size, unsigned long *zholes_size)
4391{
4392        enum zone_type j;
4393        int nid = pgdat->node_id;
4394        unsigned long zone_start_pfn = pgdat->node_start_pfn;
4395        int ret;
4396
4397        pgdat_resize_init(pgdat);
4398        init_waitqueue_head(&pgdat->kswapd_wait);
4399        init_waitqueue_head(&pgdat->pfmemalloc_wait);
4400        pgdat_page_cgroup_init(pgdat);
4401
4402        for (j = 0; j < MAX_NR_ZONES; j++) {
4403                struct zone *zone = pgdat->node_zones + j;
4404                unsigned long size, realsize, memmap_pages;
4405
4406                size = zone_spanned_pages_in_node(nid, j, zones_size);
4407                realsize = size - zone_absent_pages_in_node(nid, j,
4408                                                                zholes_size);
4409
4410                /*
4411                 * Adjust realsize so that it accounts for how much memory
4412                 * is used by this zone for memmap. This affects the watermark
4413                 * and per-cpu initialisations
4414                 */
4415                memmap_pages =
4416                        PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4417                if (realsize >= memmap_pages) {
4418                        realsize -= memmap_pages;
4419                        if (memmap_pages)
4420                                printk(KERN_DEBUG
4421                                       "  %s zone: %lu pages used for memmap\n",
4422                                       zone_names[j], memmap_pages);
4423                } else
4424                        printk(KERN_WARNING
4425                                "  %s zone: %lu pages exceeds realsize %lu\n",
4426                                zone_names[j], memmap_pages, realsize);
4427
4428                /* Account for reserved pages */
4429                if (j == 0 && realsize > dma_reserve) {
4430                        realsize -= dma_reserve;
4431                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
4432                                        zone_names[0], dma_reserve);
4433                }
4434
4435                if (!is_highmem_idx(j))
4436                        nr_kernel_pages += realsize;
4437                nr_all_pages += realsize;
4438
4439                zone->spanned_pages = size;
4440                zone->present_pages = realsize;
4441#if defined CONFIG_COMPACTION || defined CONFIG_CMA
4442                zone->compact_cached_free_pfn = zone->zone_start_pfn +
4443                                                zone->spanned_pages;
4444                zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4445#endif
4446#ifdef CONFIG_NUMA
4447                zone->node = nid;
4448                zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4449                                                / 100;
4450                zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4451#endif
4452                zone->name = zone_names[j];
4453                spin_lock_init(&zone->lock);
4454                spin_lock_init(&zone->lru_lock);
4455                zone_seqlock_init(zone);
4456                zone->zone_pgdat = pgdat;
4457
4458                zone_pcp_init(zone);
4459                lruvec_init(&zone->lruvec, zone);
4460                if (!size)
4461                        continue;
4462
4463                set_pageblock_order();
4464                setup_usemap(pgdat, zone, size);
4465                ret = init_currently_empty_zone(zone, zone_start_pfn,
4466                                                size, MEMMAP_EARLY);
4467                BUG_ON(ret);
4468                memmap_init(size, nid, j, zone_start_pfn);
4469                zone_start_pfn += size;
4470        }
4471}
4472
4473static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4474{
4475        /* Skip empty nodes */
4476        if (!pgdat->node_spanned_pages)
4477                return;
4478
4479#ifdef CONFIG_FLAT_NODE_MEM_MAP
4480        /* ia64 gets its own node_mem_map, before this, without bootmem */
4481        if (!pgdat->node_mem_map) {
4482                unsigned long size, start, end;
4483                struct page *map;
4484
4485                /*
4486                 * The zone's endpoints aren't required to be MAX_ORDER
4487                 * aligned but the node_mem_map endpoints must be in order
4488                 * for the buddy allocator to function correctly.
4489                 */
4490                start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4491                end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4492                end = ALIGN(end, MAX_ORDER_NR_PAGES);
4493                size =  (end - start) * sizeof(struct page);
4494                map = alloc_remap(pgdat->node_id, size);
4495                if (!map)
4496                        map = alloc_bootmem_node_nopanic(pgdat, size);
4497                pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4498        }
4499#ifndef CONFIG_NEED_MULTIPLE_NODES
4500        /*
4501         * With no DISCONTIG, the global mem_map is just set as node 0's
4502         */
4503        if (pgdat == NODE_DATA(0)) {
4504                mem_map = NODE_DATA(0)->node_mem_map;
4505#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4506                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4507                        mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4508#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4509        }
4510#endif
4511#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4512}
4513
4514void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4515                unsigned long node_start_pfn, unsigned long *zholes_size)
4516{
4517        pg_data_t *pgdat = NODE_DATA(nid);
4518
4519        /* pg_data_t should be reset to zero when it's allocated */
4520        WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4521
4522        pgdat->node_id = nid;
4523        pgdat->node_start_pfn = node_start_pfn;
4524        calculate_node_totalpages(pgdat, zones_size, zholes_size);
4525
4526        alloc_node_mem_map(pgdat);
4527#ifdef CONFIG_FLAT_NODE_MEM_MAP
4528        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4529                nid, (unsigned long)pgdat,
4530                (unsigned long)pgdat->node_mem_map);
4531#endif
4532
4533        free_area_init_core(pgdat, zones_size, zholes_size);
4534}
4535
4536#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4537
4538#if MAX_NUMNODES > 1
4539/*
4540 * Figure out the number of possible node ids.
4541 */
4542static void __init setup_nr_node_ids(void)
4543{
4544        unsigned int node;
4545        unsigned int highest = 0;
4546
4547        for_each_node_mask(node, node_possible_map)
4548                highest = node;
4549        nr_node_ids = highest + 1;
4550}
4551#else
4552static inline void setup_nr_node_ids(void)
4553{
4554}
4555#endif
4556
4557/**
4558 * node_map_pfn_alignment - determine the maximum internode alignment
4559 *
4560 * This function should be called after node map is populated and sorted.
4561 * It calculates the maximum power of two alignment which can distinguish
4562 * all the nodes.
4563 *
4564 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4565 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
4566 * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
4567 * shifted, 1GiB is enough and this function will indicate so.
4568 *
4569 * This is used to test whether pfn -> nid mapping of the chosen memory
4570 * model has fine enough granularity to avoid incorrect mapping for the
4571 * populated node map.
4572 *
4573 * Returns the determined alignment in pfn's.  0 if there is no alignment
4574 * requirement (single node).
4575 */
4576unsigned long __init node_map_pfn_alignment(void)
4577{
4578        unsigned long accl_mask = 0, last_end = 0;
4579        unsigned long start, end, mask;
4580        int last_nid = -1;
4581        int i, nid;
4582
4583        for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4584                if (!start || last_nid < 0 || last_nid == nid) {
4585                        last_nid = nid;
4586                        last_end = end;
4587                        continue;
4588                }
4589
4590                /*
4591                 * Start with a mask granular enough to pin-point to the
4592                 * start pfn and tick off bits one-by-one until it becomes
4593                 * too coarse to separate the current node from the last.
4594                 */
4595                mask = ~((1 << __ffs(start)) - 1);
4596                while (mask && last_end <= (start & (mask << 1)))
4597                        mask <<= 1;
4598
4599                /* accumulate all internode masks */
4600                accl_mask |= mask;
4601        }
4602
4603        /* convert mask to number of pages */
4604        return ~accl_mask + 1;
4605}
4606
4607/* Find the lowest pfn for a node */
4608static unsigned long __init find_min_pfn_for_node(int nid)
4609{
4610        unsigned long min_pfn = ULONG_MAX;
4611        unsigned long start_pfn;
4612        int i;
4613
4614        for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4615                min_pfn = min(min_pfn, start_pfn);
4616
4617        if (min_pfn == ULONG_MAX) {
4618                printk(KERN_WARNING
4619                        "Could not find start_pfn for node %d\n", nid);
4620                return 0;
4621        }
4622
4623        return min_pfn;
4624}
4625
4626/**
4627 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4628 *
4629 * It returns the minimum PFN based on information provided via
4630 * add_active_range().
4631 */
4632unsigned long __init find_min_pfn_with_active_regions(void)
4633{
4634        return find_min_pfn_for_node(MAX_NUMNODES);
4635}
4636
4637/*
4638 * early_calculate_totalpages()
4639 * Sum pages in active regions for movable zone.
4640 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4641 */
4642static unsigned long __init early_calculate_totalpages(void)
4643{
4644        unsigned long totalpages = 0;
4645        unsigned long start_pfn, end_pfn;
4646        int i, nid;
4647
4648        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4649                unsigned long pages = end_pfn - start_pfn;
4650
4651                totalpages += pages;
4652                if (pages)
4653                        node_set_state(nid, N_HIGH_MEMORY);
4654        }
4655        return totalpages;
4656}
4657
4658/*
4659 * Find the PFN the Movable zone begins in each node. Kernel memory
4660 * is spread evenly between nodes as long as the nodes have enough
4661 * memory. When they don't, some nodes will have more kernelcore than
4662 * others
4663 */
4664static void __init find_zone_movable_pfns_for_nodes(void)
4665{
4666        int i, nid;
4667        unsigned long usable_startpfn;
4668        unsigned long kernelcore_node, kernelcore_remaining;
4669        /* save the state before borrow the nodemask */
4670        nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4671        unsigned long totalpages = early_calculate_totalpages();
4672        int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4673
4674        /*
4675         * If movablecore was specified, calculate what size of
4676         * kernelcore that corresponds so that memory usable for
4677         * any allocation type is evenly spread. If both kernelcore
4678         * and movablecore are specified, then the value of kernelcore
4679         * will be used for required_kernelcore if it's greater than
4680         * what movablecore would have allowed.
4681         */
4682        if (required_movablecore) {
4683                unsigned long corepages;
4684
4685                /*
4686                 * Round-up so that ZONE_MOVABLE is at least as large as what
4687                 * was requested by the user
4688                 */
4689                required_movablecore =
4690                        roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4691                corepages = totalpages - required_movablecore;
4692
4693                required_kernelcore = max(required_kernelcore, corepages);
4694        }
4695
4696        /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4697        if (!required_kernelcore)
4698                goto out;
4699
4700        /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4701        find_usable_zone_for_movable();
4702        usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4703
4704restart:
4705        /* Spread kernelcore memory as evenly as possible throughout nodes */
4706        kernelcore_node = required_kernelcore / usable_nodes;
4707        for_each_node_state(nid, N_HIGH_MEMORY) {
4708                unsigned long start_pfn, end_pfn;
4709
4710                /*
4711                 * Recalculate kernelcore_node if the division per node
4712                 * now exceeds what is necessary to satisfy the requested
4713                 * amount of memory for the kernel
4714                 */
4715                if (required_kernelcore < kernelcore_node)
4716                        kernelcore_node = required_kernelcore / usable_nodes;
4717
4718                /*
4719                 * As the map is walked, we track how much memory is usable
4720                 * by the kernel using kernelcore_remaining. When it is
4721                 * 0, the rest of the node is usable by ZONE_MOVABLE
4722                 */
4723                kernelcore_remaining = kernelcore_node;
4724
4725                /* Go through each range of PFNs within this node */
4726                for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4727                        unsigned long size_pages;
4728
4729                        start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4730                        if (start_pfn >= end_pfn)
4731                                continue;
4732
4733                        /* Account for what is only usable for kernelcore */
4734                        if (start_pfn < usable_startpfn) {
4735                                unsigned long kernel_pages;
4736                                kernel_pages = min(end_pfn, usable_startpfn)
4737                                                                - start_pfn;
4738
4739                                kernelcore_remaining -= min(kernel_pages,
4740                                                        kernelcore_remaining);
4741                                required_kernelcore -= min(kernel_pages,
4742                                                        required_kernelcore);
4743
4744                                /* Continue if range is now fully accounted */
4745                                if (end_pfn <= usable_startpfn) {
4746
4747                                        /*
4748                                         * Push zone_movable_pfn to the end so
4749                                         * that if we have to rebalance
4750                                         * kernelcore across nodes, we will
4751                                         * not double account here
4752                                         */
4753                                        zone_movable_pfn[nid] = end_pfn;
4754                                        continue;
4755                                }
4756                                start_pfn = usable_startpfn;
4757                        }
4758
4759                        /*
4760                         * The usable PFN range for ZONE_MOVABLE is from
4761                         * start_pfn->end_pfn. Calculate size_pages as the
4762                         * number of pages used as kernelcore
4763                         */
4764                        size_pages = end_pfn - start_pfn;
4765                        if (size_pages > kernelcore_remaining)
4766                                size_pages = kernelcore_remaining;
4767                        zone_movable_pfn[nid] = start_pfn + size_pages;
4768
4769                        /*
4770                         * Some kernelcore has been met, update counts and
4771                         * break if the kernelcore for this node has been
4772                         * satisified
4773                         */
4774                        required_kernelcore -= min(required_kernelcore,
4775                                                                size_pages);
4776                        kernelcore_remaining -= size_pages;
4777                        if (!kernelcore_remaining)
4778                                break;
4779                }
4780        }
4781
4782        /*
4783         * If there is still required_kernelcore, we do another pass with one
4784         * less node in the count. This will push zone_movable_pfn[nid] further
4785         * along on the nodes that still have memory until kernelcore is
4786         * satisified
4787         */
4788        usable_nodes--;
4789        if (usable_nodes && required_kernelcore > usable_nodes)
4790                goto restart;
4791
4792        /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4793        for (nid = 0; nid < MAX_NUMNODES; nid++)
4794                zone_movable_pfn[nid] =
4795                        roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4796
4797out:
4798        /* restore the node_state */
4799        node_states[N_HIGH_MEMORY] = saved_node_state;
4800}
4801
4802/* Any regular memory on that node ? */
4803static void __init check_for_regular_memory(pg_data_t *pgdat)
4804{
4805#ifdef CONFIG_HIGHMEM
4806        enum zone_type zone_type;
4807
4808        for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4809                struct zone *zone = &pgdat->node_zones[zone_type];
4810                if (zone->present_pages) {
4811                        node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4812                        break;
4813                }
4814        }
4815#endif
4816}
4817
4818/**
4819 * free_area_init_nodes - Initialise all pg_data_t and zone data
4820 * @max_zone_pfn: an array of max PFNs for each zone
4821 *
4822 * This will call free_area_init_node() for each active node in the system.
4823 * Using the page ranges provided by add_active_range(), the size of each
4824 * zone in each node and their holes is calculated. If the maximum PFN
4825 * between two adjacent zones match, it is assumed that the zone is empty.
4826 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4827 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4828 * starts where the previous one ended. For example, ZONE_DMA32 starts
4829 * at arch_max_dma_pfn.
4830 */
4831void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4832{
4833        unsigned long start_pfn, end_pfn;
4834        int i, nid;
4835
4836        /* Record where the zone boundaries are */
4837        memset(arch_zone_lowest_possible_pfn, 0,
4838                                sizeof(arch_zone_lowest_possible_pfn));
4839        memset(arch_zone_highest_possible_pfn, 0,
4840                                sizeof(arch_zone_highest_possible_pfn));
4841        arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4842        arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4843        for (i = 1; i < MAX_NR_ZONES; i++) {
4844                if (i == ZONE_MOVABLE)
4845                        continue;
4846                arch_zone_lowest_possible_pfn[i] =
4847                        arch_zone_highest_possible_pfn[i-1];
4848                arch_zone_highest_possible_pfn[i] =
4849                        max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4850        }
4851        arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4852        arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4853
4854        /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4855        memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4856        find_zone_movable_pfns_for_nodes();
4857
4858        /* Print out the zone ranges */
4859        printk("Zone ranges:\n");
4860        for (i = 0; i < MAX_NR_ZONES; i++) {
4861                if (i == ZONE_MOVABLE)
4862                        continue;
4863                printk(KERN_CONT "  %-8s ", zone_names[i]);
4864                if (arch_zone_lowest_possible_pfn[i] ==
4865                                arch_zone_highest_possible_pfn[i])
4866                        printk(KERN_CONT "empty\n");
4867                else
4868                        printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4869                                arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4870                                (arch_zone_highest_possible_pfn[i]
4871                                        << PAGE_SHIFT) - 1);
4872        }
4873
4874        /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4875        printk("Movable zone start for each node\n");
4876        for (i = 0; i < MAX_NUMNODES; i++) {
4877                if (zone_movable_pfn[i])
4878                        printk("  Node %d: %#010lx\n", i,
4879                               zone_movable_pfn[i] << PAGE_SHIFT);
4880        }
4881
4882        /* Print out the early_node_map[] */
4883        printk("Early memory node ranges\n");
4884        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4885                printk("  node %3d: [mem %#010lx-%#010lx]\n", nid,
4886                       start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4887
4888        /* Initialise every node */
4889        mminit_verify_pageflags_layout();
4890        setup_nr_node_ids();
4891        for_each_online_node(nid) {
4892                pg_data_t *pgdat = NODE_DATA(nid);
4893                free_area_init_node(nid, NULL,
4894                                find_min_pfn_for_node(nid), NULL);
4895
4896                /* Any memory on that node */
4897                if (pgdat->node_present_pages)
4898                        node_set_state(nid, N_HIGH_MEMORY);
4899                check_for_regular_memory(pgdat);
4900        }
4901}
4902
4903static int __init cmdline_parse_core(char *p, unsigned long *core)
4904{
4905        unsigned long long coremem;
4906        if (!p)
4907                return -EINVAL;
4908
4909        coremem = memparse(p, &p);
4910        *core = coremem >> PAGE_SHIFT;
4911
4912        /* Paranoid check that UL is enough for the coremem value */
4913        WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4914
4915        return 0;
4916}
4917
4918/*
4919 * kernelcore=size sets the amount of memory for use for allocations that
4920 * cannot be reclaimed or migrated.
4921 */
4922static int __init cmdline_parse_kernelcore(char *p)
4923{
4924        return cmdline_parse_core(p, &required_kernelcore);
4925}
4926
4927/*
4928 * movablecore=size sets the amount of memory for use for allocations that
4929 * can be reclaimed or migrated.
4930 */
4931static int __init cmdline_parse_movablecore(char *p)
4932{
4933        return cmdline_parse_core(p, &required_movablecore);
4934}
4935
4936early_param("kernelcore", cmdline_parse_kernelcore);
4937early_param("movablecore", cmdline_parse_movablecore);
4938
4939#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4940
4941/**
4942 * set_dma_reserve - set the specified number of pages reserved in the first zone
4943 * @new_dma_reserve: The number of pages to mark reserved
4944 *
4945 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4946 * In the DMA zone, a significant percentage may be consumed by kernel image
4947 * and other unfreeable allocations which can skew the watermarks badly. This
4948 * function may optionally be used to account for unfreeable pages in the
4949 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4950 * smaller per-cpu batchsize.
4951 */
4952void __init set_dma_reserve(unsigned long new_dma_reserve)
4953{
4954        dma_reserve = new_dma_reserve;
4955}
4956
4957void __init free_area_init(unsigned long *zones_size)
4958{
4959        free_area_init_node(0, zones_size,
4960                        __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4961}
4962
4963static int page_alloc_cpu_notify(struct notifier_block *self,
4964                                 unsigned long action, void *hcpu)
4965{
4966        int cpu = (unsigned long)hcpu;
4967
4968        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4969                lru_add_drain_cpu(cpu);
4970                drain_pages(cpu);
4971
4972                /*
4973                 * Spill the event counters of the dead processor
4974                 * into the current processors event counters.
4975                 * This artificially elevates the count of the current
4976                 * processor.
4977                 */
4978                vm_events_fold_cpu(cpu);
4979
4980                /*
4981                 * Zero the differential counters of the dead processor
4982                 * so that the vm statistics are consistent.
4983                 *
4984                 * This is only okay since the processor is dead and cannot
4985                 * race with what we are doing.
4986                 */
4987                refresh_cpu_vm_stats(cpu);
4988        }
4989        return NOTIFY_OK;
4990}
4991
4992void __init page_alloc_init(void)
4993{
4994        hotcpu_notifier(page_alloc_cpu_notify, 0);
4995}
4996
4997/*
4998 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4999 *      or min_free_kbytes changes.
5000 */
5001static void calculate_totalreserve_pages(void)
5002{
5003        struct pglist_data *pgdat;
5004        unsigned long reserve_pages = 0;
5005        enum zone_type i, j;
5006
5007        for_each_online_pgdat(pgdat) {
5008                for (i = 0; i < MAX_NR_ZONES; i++) {
5009                        struct zone *zone = pgdat->node_zones + i;
5010                        unsigned long max = 0;
5011
5012                        /* Find valid and maximum lowmem_reserve in the zone */
5013                        for (j = i; j < MAX_NR_ZONES; j++) {
5014                                if (zone->lowmem_reserve[j] > max)
5015                                        max = zone->lowmem_reserve[j];
5016                        }
5017
5018                        /* we treat the high watermark as reserved pages. */
5019                        max += high_wmark_pages(zone);
5020
5021                        if (max > zone->present_pages)
5022                                max = zone->present_pages;
5023                        reserve_pages += max;
5024                        /*
5025                         * Lowmem reserves are not available to
5026                         * GFP_HIGHUSER page cache allocations and
5027                         * kswapd tries to balance zones to their high
5028                         * watermark.  As a result, neither should be
5029                         * regarded as dirtyable memory, to prevent a
5030                         * situation where reclaim has to clean pages
5031                         * in order to balance the zones.
5032                         */
5033                        zone->dirty_balance_reserve = max;
5034                }
5035        }
5036        dirty_balance_reserve = reserve_pages;
5037        totalreserve_pages = reserve_pages;
5038}
5039
5040/*
5041 * setup_per_zone_lowmem_reserve - called whenever
5042 *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
5043 *      has a correct pages reserved value, so an adequate number of
5044 *      pages are left in the zone after a successful __alloc_pages().
5045 */
5046static void setup_per_zone_lowmem_reserve(void)
5047{
5048        struct pglist_data *pgdat;
5049        enum zone_type j, idx;
5050
5051        for_each_online_pgdat(pgdat) {
5052                for (j = 0; j < MAX_NR_ZONES; j++) {
5053                        struct zone *zone = pgdat->node_zones + j;
5054                        unsigned long present_pages = zone->present_pages;
5055
5056                        zone->lowmem_reserve[j] = 0;
5057
5058                        idx = j;
5059                        while (idx) {
5060                                struct zone *lower_zone;
5061
5062                                idx--;
5063
5064                                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5065                                        sysctl_lowmem_reserve_ratio[idx] = 1;
5066
5067                                lower_zone = pgdat->node_zones + idx;
5068                                lower_zone->lowmem_reserve[j] = present_pages /
5069                                        sysctl_lowmem_reserve_ratio[idx];
5070                                present_pages += lower_zone->present_pages;
5071                        }
5072                }
5073        }
5074
5075        /* update totalreserve_pages */
5076        calculate_totalreserve_pages();
5077}
5078
5079static void __setup_per_zone_wmarks(void)
5080{
5081        unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5082        unsigned long lowmem_pages = 0;
5083        struct zone *zone;
5084        unsigned long flags;
5085
5086        /* Calculate total number of !ZONE_HIGHMEM pages */
5087        for_each_zone(zone) {
5088                if (!is_highmem(zone))
5089                        lowmem_pages += zone->present_pages;
5090        }
5091
5092        for_each_zone(zone) {
5093                u64 tmp;
5094
5095                spin_lock_irqsave(&zone->lock, flags);
5096                tmp = (u64)pages_min * zone->present_pages;
5097                do_div(tmp, lowmem_pages);
5098                if (is_highmem(zone)) {
5099                        /*
5100                         * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5101                         * need highmem pages, so cap pages_min to a small
5102                         * value here.
5103                         *
5104                         * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5105                         * deltas controls asynch page reclaim, and so should
5106                         * not be capped for highmem.
5107                         */
5108                        int min_pages;
5109
5110                        min_pages = zone->present_pages / 1024;
5111                        if (min_pages < SWAP_CLUSTER_MAX)
5112                                min_pages = SWAP_CLUSTER_MAX;
5113                        if (min_pages > 128)
5114                                min_pages = 128;
5115                        zone->watermark[WMARK_MIN] = min_pages;
5116                } else {
5117                        /*
5118                         * If it's a lowmem zone, reserve a number of pages
5119                         * proportionate to the zone's size.
5120                         */
5121                        zone->watermark[WMARK_MIN] = tmp;
5122                }
5123
5124                zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + (tmp >> 2);
5125                zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5126
5127                zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5128                zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5129                zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5130
5131                setup_zone_migrate_reserve(zone);
5132                spin_unlock_irqrestore(&zone->lock, flags);
5133        }
5134
5135        /* update totalreserve_pages */
5136        calculate_totalreserve_pages();
5137}
5138
5139/**
5140 * setup_per_zone_wmarks - called when min_free_kbytes changes
5141 * or when memory is hot-{added|removed}
5142 *
5143 * Ensures that the watermark[min,low,high] values for each zone are set
5144 * correctly with respect to min_free_kbytes.
5145 */
5146void setup_per_zone_wmarks(void)
5147{
5148        mutex_lock(&zonelists_mutex);
5149        __setup_per_zone_wmarks();
5150        mutex_unlock(&zonelists_mutex);
5151}
5152
5153/*
5154 * The inactive anon list should be small enough that the VM never has to
5155 * do too much work, but large enough that each inactive page has a chance
5156 * to be referenced again before it is swapped out.
5157 *
5158 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5159 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5160 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5161 * the anonymous pages are kept on the inactive list.
5162 *
5163 * total     target    max
5164 * memory    ratio     inactive anon
5165 * -------------------------------------
5166 *   10MB       1         5MB
5167 *  100MB       1        50MB
5168 *    1GB       3       250MB
5169 *   10GB      10       0.9GB
5170 *  100GB      31         3GB
5171 *    1TB     101        10GB
5172 *   10TB     320        32GB
5173 */
5174static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5175{
5176        unsigned int gb, ratio;
5177
5178        /* Zone size in gigabytes */
5179        gb = zone->present_pages >> (30 - PAGE_SHIFT);
5180        if (gb)
5181                ratio = int_sqrt(10 * gb);
5182        else
5183                ratio = 1;
5184
5185        zone->inactive_ratio = ratio;
5186}
5187
5188static void __meminit setup_per_zone_inactive_ratio(void)
5189{
5190        struct zone *zone;
5191
5192        for_each_zone(zone)
5193                calculate_zone_inactive_ratio(zone);
5194}
5195
5196/*
5197 * Initialise min_free_kbytes.
5198 *
5199 * For small machines we want it small (128k min).  For large machines
5200 * we want it large (64MB max).  But it is not linear, because network
5201 * bandwidth does not increase linearly with machine size.  We use
5202 *
5203 *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5204 *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
5205 *
5206 * which yields
5207 *
5208 * 16MB:        512k
5209 * 32MB:        724k
5210 * 64MB:        1024k
5211 * 128MB:       1448k
5212 * 256MB:       2048k
5213 * 512MB:       2896k
5214 * 1024MB:      4096k
5215 * 2048MB:      5792k
5216 * 4096MB:      8192k
5217 * 8192MB:      11584k
5218 * 16384MB:     16384k
5219 */
5220int __meminit init_per_zone_wmark_min(void)
5221{
5222        unsigned long lowmem_kbytes;
5223
5224        lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5225
5226        min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5227        if (min_free_kbytes < 128)
5228                min_free_kbytes = 128;
5229        if (min_free_kbytes > 65536)
5230                min_free_kbytes = 65536;
5231        setup_per_zone_wmarks();
5232        refresh_zone_stat_thresholds();
5233        setup_per_zone_lowmem_reserve();
5234        setup_per_zone_inactive_ratio();
5235        return 0;
5236}
5237module_init(init_per_zone_wmark_min)
5238
5239/*
5240 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
5241 *      that we can call two helper functions whenever min_free_kbytes
5242 *      changes.
5243 */
5244int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
5245        void __user *buffer, size_t *length, loff_t *ppos)
5246{
5247        proc_dointvec(table, write, buffer, length, ppos);
5248        if (write)
5249                setup_per_zone_wmarks();
5250        return 0;
5251}
5252
5253#ifdef CONFIG_NUMA
5254int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5255        void __user *buffer, size_t *length, loff_t *ppos)
5256{
5257        struct zone *zone;
5258        int rc;
5259
5260        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5261        if (rc)
5262                return rc;
5263
5264        for_each_zone(zone)
5265                zone->min_unmapped_pages = (zone->present_pages *
5266                                sysctl_min_unmapped_ratio) / 100;
5267        return 0;
5268}
5269
5270int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5271        void __user *buffer, size_t *length, loff_t *ppos)
5272{
5273        struct zone *zone;
5274        int rc;
5275
5276        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5277        if (rc)
5278                return rc;
5279
5280        for_each_zone(zone)
5281                zone->min_slab_pages = (zone->present_pages *
5282                                sysctl_min_slab_ratio) / 100;
5283        return 0;
5284}
5285#endif
5286
5287/*
5288 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5289 *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5290 *      whenever sysctl_lowmem_reserve_ratio changes.
5291 *
5292 * The reserve ratio obviously has absolutely no relation with the
5293 * minimum watermarks. The lowmem reserve ratio can only make sense
5294 * if in function of the boot time zone sizes.
5295 */
5296int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5297        void __user *buffer, size_t *length, loff_t *ppos)
5298{
5299        proc_dointvec_minmax(table, write, buffer, length, ppos);
5300        setup_per_zone_lowmem_reserve();
5301        return 0;
5302}
5303
5304/*
5305 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5306 * cpu.  It is the fraction of total pages in each zone that a hot per cpu pagelist
5307 * can have before it gets flushed back to buddy allocator.
5308 */
5309
5310int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5311        void __user *buffer, size_t *length, loff_t *ppos)
5312{
5313        struct zone *zone;
5314        unsigned int cpu;
5315        int ret;
5316
5317        ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5318        if (!write || (ret < 0))
5319                return ret;
5320        for_each_populated_zone(zone) {
5321                for_each_possible_cpu(cpu) {
5322                        unsigned long  high;
5323                        high = zone->present_pages / percpu_pagelist_fraction;
5324                        setup_pagelist_highmark(
5325                                per_cpu_ptr(zone->pageset, cpu), high);
5326                }
5327        }
5328        return 0;
5329}
5330
5331int hashdist = HASHDIST_DEFAULT;
5332
5333#ifdef CONFIG_NUMA
5334static int __init set_hashdist(char *str)
5335{
5336        if (!str)
5337                return 0;
5338        hashdist = simple_strtoul(str, &str, 0);
5339        return 1;
5340}
5341__setup("hashdist=", set_hashdist);
5342#endif
5343
5344/*
5345 * allocate a large system hash table from bootmem
5346 * - it is assumed that the hash table must contain an exact power-of-2
5347 *   quantity of entries
5348 * - limit is the number of hash buckets, not the total allocation size
5349 */
5350void *__init alloc_large_system_hash(const char *tablename,
5351                                     unsigned long bucketsize,
5352                                     unsigned long numentries,
5353                                     int scale,
5354                                     int flags,
5355                                     unsigned int *_hash_shift,
5356                                     unsigned int *_hash_mask,
5357                                     unsigned long low_limit,
5358                                     unsigned long high_limit)
5359{
5360        unsigned long long max = high_limit;
5361        unsigned long log2qty, size;
5362        void *table = NULL;
5363
5364        /* allow the kernel cmdline to have a say */
5365        if (!numentries) {
5366                /* round applicable memory size up to nearest megabyte */
5367                numentries = nr_kernel_pages;
5368                numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5369                numentries >>= 20 - PAGE_SHIFT;
5370                numentries <<= 20 - PAGE_SHIFT;
5371
5372                /* limit to 1 bucket per 2^scale bytes of low memory */
5373                if (scale > PAGE_SHIFT)
5374                        numentries >>= (scale - PAGE_SHIFT);
5375                else
5376                        numentries <<= (PAGE_SHIFT - scale);
5377
5378                /* Make sure we've got at least a 0-order allocation.. */
5379                if (unlikely(flags & HASH_SMALL)) {
5380                        /* Makes no sense without HASH_EARLY */
5381                        WARN_ON(!(flags & HASH_EARLY));
5382                        if (!(numentries >> *_hash_shift)) {
5383                                numentries = 1UL << *_hash_shift;
5384                                BUG_ON(!numentries);
5385                        }
5386                } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5387                        numentries = PAGE_SIZE / bucketsize;
5388        }
5389        numentries = roundup_pow_of_two(numentries);
5390
5391        /* limit allocation size to 1/16 total memory by default */
5392        if (max == 0) {
5393                max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5394                do_div(max, bucketsize);
5395        }
5396        max = min(max, 0x80000000ULL);
5397
5398        if (numentries < low_limit)
5399                numentries = low_limit;
5400        if (numentries > max)
5401                numentries = max;
5402
5403        log2qty = ilog2(numentries);
5404
5405        do {
5406                size = bucketsize << log2qty;
5407                if (flags & HASH_EARLY)
5408                        table = alloc_bootmem_nopanic(size);
5409                else if (hashdist)
5410                        table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5411                else {
5412                        /*
5413                         * If bucketsize is not a power-of-two, we may free
5414                         * some pages at the end of hash table which
5415                         * alloc_pages_exact() automatically does
5416                         */
5417                        if (get_order(size) < MAX_ORDER) {
5418                                table = alloc_pages_exact(size, GFP_ATOMIC);
5419                                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5420                        }
5421                }
5422        } while (!table && size > PAGE_SIZE && --log2qty);
5423
5424        if (!table)
5425                panic("Failed to allocate %s hash table\n", tablename);
5426
5427        printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5428               tablename,
5429               (1UL << log2qty),
5430               ilog2(size) - PAGE_SHIFT,
5431               size);
5432
5433        if (_hash_shift)
5434                *_hash_shift = log2qty;
5435        if (_hash_mask)
5436                *_hash_mask = (1 << log2qty) - 1;
5437
5438        return table;
5439}
5440
5441/* Return a pointer to the bitmap storing bits affecting a block of pages */
5442static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5443                                                        unsigned long pfn)
5444{
5445#ifdef CONFIG_SPARSEMEM
5446        return __pfn_to_section(pfn)->pageblock_flags;
5447#else
5448        return zone->pageblock_flags;
5449#endif /* CONFIG_SPARSEMEM */
5450}
5451
5452static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5453{
5454#ifdef CONFIG_SPARSEMEM
5455        pfn &= (PAGES_PER_SECTION-1);
5456        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5457#else
5458        pfn = pfn - zone->zone_start_pfn;
5459        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5460#endif /* CONFIG_SPARSEMEM */
5461}
5462
5463/**
5464 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5465 * @page: The page within the block of interest
5466 * @start_bitidx: The first bit of interest to retrieve
5467 * @end_bitidx: The last bit of interest
5468 * returns pageblock_bits flags
5469 */
5470unsigned long get_pageblock_flags_group(struct page *page,
5471                                        int start_bitidx, int end_bitidx)
5472{
5473        struct zone *zone;
5474        unsigned long *bitmap;
5475        unsigned long pfn, bitidx;
5476        unsigned long flags = 0;
5477        unsigned long value = 1;
5478
5479        zone = page_zone(page);
5480        pfn = page_to_pfn(page);
5481        bitmap = get_pageblock_bitmap(zone, pfn);
5482        bitidx = pfn_to_bitidx(zone, pfn);
5483
5484        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5485                if (test_bit(bitidx + start_bitidx, bitmap))
5486                        flags |= value;
5487
5488        return flags;
5489}
5490
5491/**
5492 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5493 * @page: The page within the block of interest
5494 * @start_bitidx: The first bit of interest
5495 * @end_bitidx: The last bit of interest
5496 * @flags: The flags to set
5497 */
5498void set_pageblock_flags_group(struct page *page, unsigned long flags,
5499                                        int start_bitidx, int end_bitidx)
5500{
5501        struct zone *zone;
5502        unsigned long *bitmap;
5503        unsigned long pfn, bitidx;
5504        unsigned long value = 1;
5505
5506        zone = page_zone(page);
5507        pfn = page_to_pfn(page);
5508        bitmap = get_pageblock_bitmap(zone, pfn);
5509        bitidx = pfn_to_bitidx(zone, pfn);
5510        VM_BUG_ON(pfn < zone->zone_start_pfn);
5511        VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5512
5513        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5514                if (flags & value)
5515                        __set_bit(bitidx + start_bitidx, bitmap);
5516                else
5517                        __clear_bit(bitidx + start_bitidx, bitmap);
5518}
5519
5520/*
5521 * This function checks whether pageblock includes unmovable pages or not.
5522 * If @count is not zero, it is okay to include less @count unmovable pages
5523 *
5524 * PageLRU check wihtout isolation or lru_lock could race so that
5525 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5526 * expect this function should be exact.
5527 */
5528bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5529{
5530        unsigned long pfn, iter, found;
5531        int mt;
5532
5533        /*
5534         * For avoiding noise data, lru_add_drain_all() should be called
5535         * If ZONE_MOVABLE, the zone never contains unmovable pages
5536         */
5537        if (zone_idx(zone) == ZONE_MOVABLE)
5538                return false;
5539        mt = get_pageblock_migratetype(page);
5540        if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5541                return false;
5542
5543        pfn = page_to_pfn(page);
5544        for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5545                unsigned long check = pfn + iter;
5546
5547                if (!pfn_valid_within(check))
5548                        continue;
5549
5550                page = pfn_to_page(check);
5551                /*
5552                 * We can't use page_count without pin a page
5553                 * because another CPU can free compound page.
5554                 * This check already skips compound tails of THP
5555                 * because their page->_count is zero at all time.
5556                 */
5557                if (!atomic_read(&page->_count)) {
5558                        if (PageBuddy(page))
5559                                iter += (1 << page_order(page)) - 1;
5560                        continue;
5561                }
5562
5563                if (!PageLRU(page))
5564                        found++;
5565                /*
5566                 * If there are RECLAIMABLE pages, we need to check it.
5567                 * But now, memory offline itself doesn't call shrink_slab()
5568                 * and it still to be fixed.
5569                 */
5570                /*
5571                 * If the page is not RAM, page_count()should be 0.
5572                 * we don't need more check. This is an _used_ not-movable page.
5573                 *
5574                 * The problematic thing here is PG_reserved pages. PG_reserved
5575                 * is set to both of a memory hole page and a _used_ kernel
5576                 * page at boot.
5577                 */
5578                if (found > count)
5579                        return true;
5580        }
5581        return false;
5582}
5583
5584bool is_pageblock_removable_nolock(struct page *page)
5585{
5586        struct zone *zone;
5587        unsigned long pfn;
5588
5589        /*
5590         * We have to be careful here because we are iterating over memory
5591         * sections which are not zone aware so we might end up outside of
5592         * the zone but still within the section.
5593         * We have to take care about the node as well. If the node is offline
5594         * its NODE_DATA will be NULL - see page_zone.
5595         */
5596        if (!node_online(page_to_nid(page)))
5597                return false;
5598
5599        zone = page_zone(page);
5600        pfn = page_to_pfn(page);
5601        if (zone->zone_start_pfn > pfn ||
5602                        zone->zone_start_pfn + zone->spanned_pages <= pfn)
5603                return false;
5604
5605        return !has_unmovable_pages(zone, page, 0);
5606}
5607
5608#ifdef CONFIG_CMA
5609
5610static unsigned long pfn_max_align_down(unsigned long pfn)
5611{
5612        return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5613                             pageblock_nr_pages) - 1);
5614}
5615
5616static unsigned long pfn_max_align_up(unsigned long pfn)
5617{
5618        return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5619                                pageblock_nr_pages));
5620}
5621
5622static struct page *
5623__alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5624                             int **resultp)
5625{
5626        gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5627
5628        if (PageHighMem(page))
5629                gfp_mask |= __GFP_HIGHMEM;
5630
5631        return alloc_page(gfp_mask);
5632}
5633
5634/* [start, end) must belong to a single zone. */
5635static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5636{
5637        /* This function is based on compact_zone() from compaction.c. */
5638
5639        unsigned long pfn = start;
5640        unsigned int tries = 0;
5641        int ret = 0;
5642
5643        struct compact_control cc = {
5644                .nr_migratepages = 0,
5645                .order = -1,
5646                .zone = page_zone(pfn_to_page(start)),
5647                .sync = true,
5648        };
5649        INIT_LIST_HEAD(&cc.migratepages);
5650
5651        migrate_prep_local();
5652
5653        while (pfn < end || !list_empty(&cc.migratepages)) {
5654                if (fatal_signal_pending(current)) {
5655                        ret = -EINTR;
5656                        break;
5657                }
5658
5659                if (list_empty(&cc.migratepages)) {
5660                        cc.nr_migratepages = 0;
5661                        pfn = isolate_migratepages_range(cc.zone, &cc,
5662                                                         pfn, end);
5663                        if (!pfn) {
5664                                ret = -EINTR;
5665                                break;
5666                        }
5667                        tries = 0;
5668                } else if (++tries == 5) {
5669                        ret = ret < 0 ? ret : -EBUSY;
5670                        break;
5671                }
5672
5673                ret = migrate_pages(&cc.migratepages,
5674                                    __alloc_contig_migrate_alloc,
5675                                    0, false, MIGRATE_SYNC);
5676        }
5677
5678        putback_lru_pages(&cc.migratepages);
5679        return ret > 0 ? 0 : ret;
5680}
5681
5682/*
5683 * Update zone's cma pages counter used for watermark level calculation.
5684 */
5685static inline void __update_cma_watermarks(struct zone *zone, int count)
5686{
5687        unsigned long flags;
5688        spin_lock_irqsave(&zone->lock, flags);
5689        zone->min_cma_pages += count;
5690        spin_unlock_irqrestore(&zone->lock, flags);
5691        setup_per_zone_wmarks();
5692}
5693
5694/*
5695 * Trigger memory pressure bump to reclaim some pages in order to be able to
5696 * allocate 'count' pages in single page units. Does similar work as
5697 *__alloc_pages_slowpath() function.
5698 */
5699static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5700{
5701        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5702        struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5703        int did_some_progress = 0;
5704        int order = 1;
5705
5706        /*
5707         * Increase level of watermarks to force kswapd do his job
5708         * to stabilise at new watermark level.
5709         */
5710        __update_cma_watermarks(zone, count);
5711
5712        /* Obey watermarks as if the page was being allocated */
5713        while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5714                wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5715
5716                did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5717                                                      NULL);
5718                if (!did_some_progress) {
5719                        /* Exhausted what can be done so it's blamo time */
5720                        out_of_memory(zonelist, gfp_mask, order, NULL, false);
5721                }
5722        }
5723
5724        /* Restore original watermark levels. */
5725        __update_cma_watermarks(zone, -count);
5726
5727        return count;
5728}
5729
5730/**
5731 * alloc_contig_range() -- tries to allocate given range of pages
5732 * @start:      start PFN to allocate
5733 * @end:        one-past-the-last PFN to allocate
5734 * @migratetype:        migratetype of the underlaying pageblocks (either
5735 *                      #MIGRATE_MOVABLE or #MIGRATE_CMA).  All pageblocks
5736 *                      in range must have the same migratetype and it must
5737 *                      be either of the two.
5738 *
5739 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5740 * aligned, however it's the caller's responsibility to guarantee that
5741 * we are the only thread that changes migrate type of pageblocks the
5742 * pages fall in.
5743 *
5744 * The PFN range must belong to a single zone.
5745 *
5746 * Returns zero on success or negative error code.  On success all
5747 * pages which PFN is in [start, end) are allocated for the caller and
5748 * need to be freed with free_contig_range().
5749 */
5750int alloc_contig_range(unsigned long start, unsigned long end,
5751                       unsigned migratetype)
5752{
5753        struct zone *zone = page_zone(pfn_to_page(start));
5754        unsigned long