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