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