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