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