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/oom.h>
  34#include <linux/notifier.h>
  35#include <linux/topology.h>
  36#include <linux/sysctl.h>
  37#include <linux/cpu.h>
  38#include <linux/cpuset.h>
  39#include <linux/memory_hotplug.h>
  40#include <linux/nodemask.h>
  41#include <linux/vmalloc.h>
  42#include <linux/mempolicy.h>
  43#include <linux/stop_machine.h>
  44#include <linux/sort.h>
  45#include <linux/pfn.h>
  46#include <linux/backing-dev.h>
  47#include <linux/fault-inject.h>
  48#include <linux/page-isolation.h>
  49#include <linux/page_cgroup.h>
  50#include <linux/debugobjects.h>
  51#include <linux/kmemleak.h>
  52#include <linux/memory.h>
  53#include <linux/compaction.h>
  54#include <trace/events/kmem.h>
  55#include <linux/ftrace_event.h>
  56
  57#include <asm/tlbflush.h>
  58#include <asm/div64.h>
  59#include "internal.h"
  60
  61#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
  62DEFINE_PER_CPU(int, numa_node);
  63EXPORT_PER_CPU_SYMBOL(numa_node);
  64#endif
  65
  66#ifdef CONFIG_HAVE_MEMORYLESS_NODES
  67/*
  68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
  69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
  70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
  71 * defined in <linux/topology.h>.
  72 */
  73DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
  74EXPORT_PER_CPU_SYMBOL(_numa_mem_);
  75#endif
  76
  77/*
  78 * Array of node states.
  79 */
  80nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
  81        [N_POSSIBLE] = NODE_MASK_ALL,
  82        [N_ONLINE] = { { [0] = 1UL } },
  83#ifndef CONFIG_NUMA
  84        [N_NORMAL_MEMORY] = { { [0] = 1UL } },
  85#ifdef CONFIG_HIGHMEM
  86        [N_HIGH_MEMORY] = { { [0] = 1UL } },
  87#endif
  88        [N_CPU] = { { [0] = 1UL } },
  89#endif  /* NUMA */
  90};
  91EXPORT_SYMBOL(node_states);
  92
  93unsigned long totalram_pages __read_mostly;
  94unsigned long totalreserve_pages __read_mostly;
  95int percpu_pagelist_fraction;
  96gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
  97
  98#ifdef CONFIG_PM_SLEEP
  99/*
 100 * The following functions are used by the suspend/hibernate code to temporarily
 101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
 102 * while devices are suspended.  To avoid races with the suspend/hibernate code,
 103 * they should always be called with pm_mutex held (gfp_allowed_mask also should
 104 * only be modified with pm_mutex held, unless the suspend/hibernate code is
 105 * guaranteed not to run in parallel with that modification).
 106 */
 107
 108static gfp_t saved_gfp_mask;
 109
 110void pm_restore_gfp_mask(void)
 111{
 112        WARN_ON(!mutex_is_locked(&pm_mutex));
 113        if (saved_gfp_mask) {
 114                gfp_allowed_mask = saved_gfp_mask;
 115                saved_gfp_mask = 0;
 116        }
 117}
 118
 119void pm_restrict_gfp_mask(void)
 120{
 121        WARN_ON(!mutex_is_locked(&pm_mutex));
 122        WARN_ON(saved_gfp_mask);
 123        saved_gfp_mask = gfp_allowed_mask;
 124        gfp_allowed_mask &= ~GFP_IOFS;
 125}
 126#endif /* CONFIG_PM_SLEEP */
 127
 128#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 129int pageblock_order __read_mostly;
 130#endif
 131
 132static void __free_pages_ok(struct page *page, unsigned int order);
 133
 134/*
 135 * results with 256, 32 in the lowmem_reserve sysctl:
 136 *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 137 *      1G machine -> (16M dma, 784M normal, 224M high)
 138 *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 139 *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 140 *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
 141 *
 142 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 143 * don't need any ZONE_NORMAL reservation
 144 */
 145int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
 146#ifdef CONFIG_ZONE_DMA
 147         256,
 148#endif
 149#ifdef CONFIG_ZONE_DMA32
 150         256,
 151#endif
 152#ifdef CONFIG_HIGHMEM
 153         32,
 154#endif
 155         32,
 156};
 157
 158EXPORT_SYMBOL(totalram_pages);
 159
 160static char * const zone_names[MAX_NR_ZONES] = {
 161#ifdef CONFIG_ZONE_DMA
 162         "DMA",
 163#endif
 164#ifdef CONFIG_ZONE_DMA32
 165         "DMA32",
 166#endif
 167         "Normal",
 168#ifdef CONFIG_HIGHMEM
 169         "HighMem",
 170#endif
 171         "Movable",
 172};
 173
 174int min_free_kbytes = 1024;
 175
 176static unsigned long __meminitdata nr_kernel_pages;
 177static unsigned long __meminitdata nr_all_pages;
 178static unsigned long __meminitdata dma_reserve;
 179
 180#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
 181  /*
 182   * MAX_ACTIVE_REGIONS determines the maximum number of distinct
 183   * ranges of memory (RAM) that may be registered with add_active_range().
 184   * Ranges passed to add_active_range() will be merged if possible
 185   * so the number of times add_active_range() can be called is
 186   * related to the number of nodes and the number of holes
 187   */
 188  #ifdef CONFIG_MAX_ACTIVE_REGIONS
 189    /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
 190    #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
 191  #else
 192    #if MAX_NUMNODES >= 32
 193      /* If there can be many nodes, allow up to 50 holes per node */
 194      #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
 195    #else
 196      /* By default, allow up to 256 distinct regions */
 197      #define MAX_ACTIVE_REGIONS 256
 198    #endif
 199  #endif
 200
 201  static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
 202  static int __meminitdata nr_nodemap_entries;
 203  static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
 204  static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
 205  static unsigned long __initdata required_kernelcore;
 206  static unsigned long __initdata required_movablecore;
 207  static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
 208
 209  /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 210  int movable_zone;
 211  EXPORT_SYMBOL(movable_zone);
 212#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
 213
 214#if MAX_NUMNODES > 1
 215int nr_node_ids __read_mostly = MAX_NUMNODES;
 216int nr_online_nodes __read_mostly = 1;
 217EXPORT_SYMBOL(nr_node_ids);
 218EXPORT_SYMBOL(nr_online_nodes);
 219#endif
 220
 221int page_group_by_mobility_disabled __read_mostly;
 222
 223static void set_pageblock_migratetype(struct page *page, int migratetype)
 224{
 225
 226        if (unlikely(page_group_by_mobility_disabled))
 227                migratetype = MIGRATE_UNMOVABLE;
 228
 229        set_pageblock_flags_group(page, (unsigned long)migratetype,
 230                                        PB_migrate, PB_migrate_end);
 231}
 232
 233bool oom_killer_disabled __read_mostly;
 234
 235#ifdef CONFIG_DEBUG_VM
 236static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 237{
 238        int ret = 0;
 239        unsigned seq;
 240        unsigned long pfn = page_to_pfn(page);
 241
 242        do {
 243                seq = zone_span_seqbegin(zone);
 244                if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
 245                        ret = 1;
 246                else if (pfn < zone->zone_start_pfn)
 247                        ret = 1;
 248        } while (zone_span_seqretry(zone, seq));
 249
 250        return ret;
 251}
 252
 253static int page_is_consistent(struct zone *zone, struct page *page)
 254{
 255        if (!pfn_valid_within(page_to_pfn(page)))
 256                return 0;
 257        if (zone != page_zone(page))
 258                return 0;
 259
 260        return 1;
 261}
 262/*
 263 * Temporary debugging check for pages not lying within a given zone.
 264 */
 265static int bad_range(struct zone *zone, struct page *page)
 266{
 267        if (page_outside_zone_boundaries(zone, page))
 268                return 1;
 269        if (!page_is_consistent(zone, page))
 270                return 1;
 271
 272        return 0;
 273}
 274#else
 275static inline int bad_range(struct zone *zone, struct page *page)
 276{
 277        return 0;
 278}
 279#endif
 280
 281static void bad_page(struct page *page)
 282{
 283        static unsigned long resume;
 284        static unsigned long nr_shown;
 285        static unsigned long nr_unshown;
 286
 287        /* Don't complain about poisoned pages */
 288        if (PageHWPoison(page)) {
 289                reset_page_mapcount(page); /* remove PageBuddy */
 290                return;
 291        }
 292
 293        /*
 294         * Allow a burst of 60 reports, then keep quiet for that minute;
 295         * or allow a steady drip of one report per second.
 296         */
 297        if (nr_shown == 60) {
 298                if (time_before(jiffies, resume)) {
 299                        nr_unshown++;
 300                        goto out;
 301                }
 302                if (nr_unshown) {
 303                        printk(KERN_ALERT
 304                              "BUG: Bad page state: %lu messages suppressed\n",
 305                                nr_unshown);
 306                        nr_unshown = 0;
 307                }
 308                nr_shown = 0;
 309        }
 310        if (nr_shown++ == 0)
 311                resume = jiffies + 60 * HZ;
 312
 313        printk(KERN_ALERT "BUG: Bad page state in process %s  pfn:%05lx\n",
 314                current->comm, page_to_pfn(page));
 315        dump_page(page);
 316
 317        dump_stack();
 318out:
 319        /* Leave bad fields for debug, except PageBuddy could make trouble */
 320        reset_page_mapcount(page); /* remove PageBuddy */
 321        add_taint(TAINT_BAD_PAGE);
 322}
 323
 324/*
 325 * Higher-order pages are called "compound pages".  They are structured thusly:
 326 *
 327 * The first PAGE_SIZE page is called the "head page".
 328 *
 329 * The remaining PAGE_SIZE pages are called "tail pages".
 330 *
 331 * All pages have PG_compound set.  All pages have their ->private pointing at
 332 * the head page (even the head page has this).
 333 *
 334 * The first tail page's ->lru.next holds the address of the compound page's
 335 * put_page() function.  Its ->lru.prev holds the order of allocation.
 336 * This usage means that zero-order pages may not be compound.
 337 */
 338
 339static void free_compound_page(struct page *page)
 340{
 341        __free_pages_ok(page, compound_order(page));
 342}
 343
 344void prep_compound_page(struct page *page, unsigned long order)
 345{
 346        int i;
 347        int nr_pages = 1 << order;
 348
 349        set_compound_page_dtor(page, free_compound_page);
 350        set_compound_order(page, order);
 351        __SetPageHead(page);
 352        for (i = 1; i < nr_pages; i++) {
 353                struct page *p = page + i;
 354
 355                __SetPageTail(p);
 356                p->first_page = page;
 357        }
 358}
 359
 360/* update __split_huge_page_refcount if you change this function */
 361static int destroy_compound_page(struct page *page, unsigned long order)
 362{
 363        int i;
 364        int nr_pages = 1 << order;
 365        int bad = 0;
 366
 367        if (unlikely(compound_order(page) != order) ||
 368            unlikely(!PageHead(page))) {
 369                bad_page(page);
 370                bad++;
 371        }
 372
 373        __ClearPageHead(page);
 374
 375        for (i = 1; i < nr_pages; i++) {
 376                struct page *p = page + i;
 377
 378                if (unlikely(!PageTail(p) || (p->first_page != page))) {
 379                        bad_page(page);
 380                        bad++;
 381                }
 382                __ClearPageTail(p);
 383        }
 384
 385        return bad;
 386}
 387
 388static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
 389{
 390        int i;
 391
 392        /*
 393         * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
 394         * and __GFP_HIGHMEM from hard or soft interrupt context.
 395         */
 396        VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
 397        for (i = 0; i < (1 << order); i++)
 398                clear_highpage(page + i);
 399}
 400
 401static inline void set_page_order(struct page *page, int order)
 402{
 403        set_page_private(page, order);
 404        __SetPageBuddy(page);
 405}
 406
 407static inline void rmv_page_order(struct page *page)
 408{
 409        __ClearPageBuddy(page);
 410        set_page_private(page, 0);
 411}
 412
 413/*
 414 * Locate the struct page for both the matching buddy in our
 415 * pair (buddy1) and the combined O(n+1) page they form (page).
 416 *
 417 * 1) Any buddy B1 will have an order O twin B2 which satisfies
 418 * the following equation:
 419 *     B2 = B1 ^ (1 << O)
 420 * For example, if the starting buddy (buddy2) is #8 its order
 421 * 1 buddy is #10:
 422 *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 423 *
 424 * 2) Any buddy B will have an order O+1 parent P which
 425 * satisfies the following equation:
 426 *     P = B & ~(1 << O)
 427 *
 428 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
 429 */
 430static inline unsigned long
 431__find_buddy_index(unsigned long page_idx, unsigned int order)
 432{
 433        return page_idx ^ (1 << order);
 434}
 435
 436/*
 437 * This function checks whether a page is free && is the buddy
 438 * we can do coalesce a page and its buddy if
 439 * (a) the buddy is not in a hole &&
 440 * (b) the buddy is in the buddy system &&
 441 * (c) a page and its buddy have the same order &&
 442 * (d) a page and its buddy are in the same zone.
 443 *
 444 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
 445 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
 446 *
 447 * For recording page's order, we use page_private(page).
 448 */
 449static inline int page_is_buddy(struct page *page, struct page *buddy,
 450                                                                int order)
 451{
 452        if (!pfn_valid_within(page_to_pfn(buddy)))
 453                return 0;
 454
 455        if (page_zone_id(page) != page_zone_id(buddy))
 456                return 0;
 457
 458        if (PageBuddy(buddy) && page_order(buddy) == order) {
 459                VM_BUG_ON(page_count(buddy) != 0);
 460                return 1;
 461        }
 462        return 0;
 463}
 464
 465/*
 466 * Freeing function for a buddy system allocator.
 467 *
 468 * The concept of a buddy system is to maintain direct-mapped table
 469 * (containing bit values) for memory blocks of various "orders".
 470 * The bottom level table contains the map for the smallest allocatable
 471 * units of memory (here, pages), and each level above it describes
 472 * pairs of units from the levels below, hence, "buddies".
 473 * At a high level, all that happens here is marking the table entry
 474 * at the bottom level available, and propagating the changes upward
 475 * as necessary, plus some accounting needed to play nicely with other
 476 * parts of the VM system.
 477 * At each level, we keep a list of pages, which are heads of continuous
 478 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
 479 * order is recorded in page_private(page) field.
 480 * So when we are allocating or freeing one, we can derive the state of the
 481 * other.  That is, if we allocate a small block, and both were   
 482 * free, the remainder of the region must be split into blocks.   
 483 * If a block is freed, and its buddy is also free, then this
 484 * triggers coalescing into a block of larger size.            
 485 *
 486 * -- wli
 487 */
 488
 489static inline void __free_one_page(struct page *page,
 490                struct zone *zone, unsigned int order,
 491                int migratetype)
 492{
 493        unsigned long page_idx;
 494        unsigned long combined_idx;
 495        unsigned long uninitialized_var(buddy_idx);
 496        struct page *buddy;
 497
 498        if (unlikely(PageCompound(page)))
 499                if (unlikely(destroy_compound_page(page, order)))
 500                        return;
 501
 502        VM_BUG_ON(migratetype == -1);
 503
 504        page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 505
 506        VM_BUG_ON(page_idx & ((1 << order) - 1));
 507        VM_BUG_ON(bad_range(zone, page));
 508
 509        while (order < MAX_ORDER-1) {
 510                buddy_idx = __find_buddy_index(page_idx, order);
 511                buddy = page + (buddy_idx - page_idx);
 512                if (!page_is_buddy(page, buddy, order))
 513                        break;
 514
 515                /* Our buddy is free, merge with it and move up one order. */
 516                list_del(&buddy->lru);
 517                zone->free_area[order].nr_free--;
 518                rmv_page_order(buddy);
 519                combined_idx = buddy_idx & page_idx;
 520                page = page + (combined_idx - page_idx);
 521                page_idx = combined_idx;
 522                order++;
 523        }
 524        set_page_order(page, order);
 525
 526        /*
 527         * If this is not the largest possible page, check if the buddy
 528         * of the next-highest order is free. If it is, it's possible
 529         * that pages are being freed that will coalesce soon. In case,
 530         * that is happening, add the free page to the tail of the list
 531         * so it's less likely to be used soon and more likely to be merged
 532         * as a higher order page
 533         */
 534        if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
 535                struct page *higher_page, *higher_buddy;
 536                combined_idx = buddy_idx & page_idx;
 537                higher_page = page + (combined_idx - page_idx);
 538                buddy_idx = __find_buddy_index(combined_idx, order + 1);
 539                higher_buddy = page + (buddy_idx - combined_idx);
 540                if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
 541                        list_add_tail(&page->lru,
 542                                &zone->free_area[order].free_list[migratetype]);
 543                        goto out;
 544                }
 545        }
 546
 547        list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
 548out:
 549        zone->free_area[order].nr_free++;
 550}
 551
 552/*
 553 * free_page_mlock() -- clean up attempts to free and mlocked() page.
 554 * Page should not be on lru, so no need to fix that up.
 555 * free_pages_check() will verify...
 556 */
 557static inline void free_page_mlock(struct page *page)
 558{
 559        __dec_zone_page_state(page, NR_MLOCK);
 560        __count_vm_event(UNEVICTABLE_MLOCKFREED);
 561}
 562
 563static inline int free_pages_check(struct page *page)
 564{
 565        if (unlikely(page_mapcount(page) |
 566                (page->mapping != NULL)  |
 567                (atomic_read(&page->_count) != 0) |
 568                (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
 569                bad_page(page);
 570                return 1;
 571        }
 572        if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
 573                page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
 574        return 0;
 575}
 576
 577/*
 578 * Frees a number of pages from the PCP lists
 579 * Assumes all pages on list are in same zone, and of same order.
 580 * count is the number of pages to free.
 581 *
 582 * If the zone was previously in an "all pages pinned" state then look to
 583 * see if this freeing clears that state.
 584 *
 585 * And clear the zone's pages_scanned counter, to hold off the "all pages are
 586 * pinned" detection logic.
 587 */
 588static void free_pcppages_bulk(struct zone *zone, int count,
 589                                        struct per_cpu_pages *pcp)
 590{
 591        int migratetype = 0;
 592        int batch_free = 0;
 593        int to_free = count;
 594
 595        spin_lock(&zone->lock);
 596        zone->all_unreclaimable = 0;
 597        zone->pages_scanned = 0;
 598
 599        while (to_free) {
 600                struct page *page;
 601                struct list_head *list;
 602
 603                /*
 604                 * Remove pages from lists in a round-robin fashion. A
 605                 * batch_free count is maintained that is incremented when an
 606                 * empty list is encountered.  This is so more pages are freed
 607                 * off fuller lists instead of spinning excessively around empty
 608                 * lists
 609                 */
 610                do {
 611                        batch_free++;
 612                        if (++migratetype == MIGRATE_PCPTYPES)
 613                                migratetype = 0;
 614                        list = &pcp->lists[migratetype];
 615                } while (list_empty(list));
 616
 617                do {
 618                        page = list_entry(list->prev, struct page, lru);
 619                        /* must delete as __free_one_page list manipulates */
 620                        list_del(&page->lru);
 621                        /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
 622                        __free_one_page(page, zone, 0, page_private(page));
 623                        trace_mm_page_pcpu_drain(page, 0, page_private(page));
 624                } while (--to_free && --batch_free && !list_empty(list));
 625        }
 626        __mod_zone_page_state(zone, NR_FREE_PAGES, count);
 627        spin_unlock(&zone->lock);
 628}
 629
 630static void free_one_page(struct zone *zone, struct page *page, int order,
 631                                int migratetype)
 632{
 633        spin_lock(&zone->lock);
 634        zone->all_unreclaimable = 0;
 635        zone->pages_scanned = 0;
 636
 637        __free_one_page(page, zone, order, migratetype);
 638        __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 639        spin_unlock(&zone->lock);
 640}
 641
 642static bool free_pages_prepare(struct page *page, unsigned int order)
 643{
 644        int i;
 645        int bad = 0;
 646
 647        trace_mm_page_free_direct(page, order);
 648        kmemcheck_free_shadow(page, order);
 649
 650        if (PageAnon(page))
 651                page->mapping = NULL;
 652        for (i = 0; i < (1 << order); i++)
 653                bad += free_pages_check(page + i);
 654        if (bad)
 655                return false;
 656
 657        if (!PageHighMem(page)) {
 658                debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
 659                debug_check_no_obj_freed(page_address(page),
 660                                           PAGE_SIZE << order);
 661        }
 662        arch_free_page(page, order);
 663        kernel_map_pages(page, 1 << order, 0);
 664
 665        return true;
 666}
 667
 668static void __free_pages_ok(struct page *page, unsigned int order)
 669{
 670        unsigned long flags;
 671        int wasMlocked = __TestClearPageMlocked(page);
 672
 673        if (!free_pages_prepare(page, order))
 674                return;
 675
 676        local_irq_save(flags);
 677        if (unlikely(wasMlocked))
 678                free_page_mlock(page);
 679        __count_vm_events(PGFREE, 1 << order);
 680        free_one_page(page_zone(page), page, order,
 681                                        get_pageblock_migratetype(page));
 682        local_irq_restore(flags);
 683}
 684
 685/*
 686 * permit the bootmem allocator to evade page validation on high-order frees
 687 */
 688void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 689{
 690        if (order == 0) {
 691                __ClearPageReserved(page);
 692                set_page_count(page, 0);
 693                set_page_refcounted(page);
 694                __free_page(page);
 695        } else {
 696                int loop;
 697
 698                prefetchw(page);
 699                for (loop = 0; loop < BITS_PER_LONG; loop++) {
 700                        struct page *p = &page[loop];
 701
 702                        if (loop + 1 < BITS_PER_LONG)
 703                                prefetchw(p + 1);
 704                        __ClearPageReserved(p);
 705                        set_page_count(p, 0);
 706                }
 707
 708                set_page_refcounted(page);
 709                __free_pages(page, order);
 710        }
 711}
 712
 713
 714/*
 715 * The order of subdivision here is critical for the IO subsystem.
 716 * Please do not alter this order without good reasons and regression
 717 * testing. Specifically, as large blocks of memory are subdivided,
 718 * the order in which smaller blocks are delivered depends on the order
 719 * they're subdivided in this function. This is the primary factor
 720 * influencing the order in which pages are delivered to the IO
 721 * subsystem according to empirical testing, and this is also justified
 722 * by considering the behavior of a buddy system containing a single
 723 * large block of memory acted on by a series of small allocations.
 724 * This behavior is a critical factor in sglist merging's success.
 725 *
 726 * -- wli
 727 */
 728static inline void expand(struct zone *zone, struct page *page,
 729        int low, int high, struct free_area *area,
 730        int migratetype)
 731{
 732        unsigned long size = 1 << high;
 733
 734        while (high > low) {
 735                area--;
 736                high--;
 737                size >>= 1;
 738                VM_BUG_ON(bad_range(zone, &page[size]));
 739                list_add(&page[size].lru, &area->free_list[migratetype]);
 740                area->nr_free++;
 741                set_page_order(&page[size], high);
 742        }
 743}
 744
 745/*
 746 * This page is about to be returned from the page allocator
 747 */
 748static inline int check_new_page(struct page *page)
 749{
 750        if (unlikely(page_mapcount(page) |
 751                (page->mapping != NULL)  |
 752                (atomic_read(&page->_count) != 0)  |
 753                (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
 754                bad_page(page);
 755                return 1;
 756        }
 757        return 0;
 758}
 759
 760static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 761{
 762        int i;
 763
 764        for (i = 0; i < (1 << order); i++) {
 765                struct page *p = page + i;
 766                if (unlikely(check_new_page(p)))
 767                        return 1;
 768        }
 769
 770        set_page_private(page, 0);
 771        set_page_refcounted(page);
 772
 773        arch_alloc_page(page, order);
 774        kernel_map_pages(page, 1 << order, 1);
 775
 776        if (gfp_flags & __GFP_ZERO)
 777                prep_zero_page(page, order, gfp_flags);
 778
 779        if (order && (gfp_flags & __GFP_COMP))
 780                prep_compound_page(page, order);
 781
 782        return 0;
 783}
 784
 785/*
 786 * Go through the free lists for the given migratetype and remove
 787 * the smallest available page from the freelists
 788 */
 789static inline
 790struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
 791                                                int migratetype)
 792{
 793        unsigned int current_order;
 794        struct free_area * area;
 795        struct page *page;
 796
 797        /* Find a page of the appropriate size in the preferred list */
 798        for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 799                area = &(zone->free_area[current_order]);
 800                if (list_empty(&area->free_list[migratetype]))
 801                        continue;
 802
 803                page = list_entry(area->free_list[migratetype].next,
 804                                                        struct page, lru);
 805                list_del(&page->lru);
 806                rmv_page_order(page);
 807                area->nr_free--;
 808                expand(zone, page, order, current_order, area, migratetype);
 809                return page;
 810        }
 811
 812        return NULL;
 813}
 814
 815
 816/*
 817 * This array describes the order lists are fallen back to when
 818 * the free lists for the desirable migrate type are depleted
 819 */
 820static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
 821        [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 822        [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 823        [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
 824        [MIGRATE_RESERVE]     = { MIGRATE_RESERVE,     MIGRATE_RESERVE,   MIGRATE_RESERVE }, /* Never used */
 825};
 826
 827/*
 828 * Move the free pages in a range to the free lists of the requested type.
 829 * Note that start_page and end_pages are not aligned on a pageblock
 830 * boundary. If alignment is required, use move_freepages_block()
 831 */
 832static int move_freepages(struct zone *zone,
 833                          struct page *start_page, struct page *end_page,
 834                          int migratetype)
 835{
 836        struct page *page;
 837        unsigned long order;
 838        int pages_moved = 0;
 839
 840#ifndef CONFIG_HOLES_IN_ZONE
 841        /*
 842         * page_zone is not safe to call in this context when
 843         * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
 844         * anyway as we check zone boundaries in move_freepages_block().
 845         * Remove at a later date when no bug reports exist related to
 846         * grouping pages by mobility
 847         */
 848        BUG_ON(page_zone(start_page) != page_zone(end_page));
 849#endif
 850
 851        for (page = start_page; page <= end_page;) {
 852                /* Make sure we are not inadvertently changing nodes */
 853                VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
 854
 855                if (!pfn_valid_within(page_to_pfn(page))) {
 856                        page++;
 857                        continue;
 858                }
 859
 860                if (!PageBuddy(page)) {
 861                        page++;
 862                        continue;
 863                }
 864
 865                order = page_order(page);
 866                list_del(&page->lru);
 867                list_add(&page->lru,
 868                        &zone->free_area[order].free_list[migratetype]);
 869                page += 1 << order;
 870                pages_moved += 1 << order;
 871        }
 872
 873        return pages_moved;
 874}
 875
 876static int move_freepages_block(struct zone *zone, struct page *page,
 877                                int migratetype)
 878{
 879        unsigned long start_pfn, end_pfn;
 880        struct page *start_page, *end_page;
 881
 882        start_pfn = page_to_pfn(page);
 883        start_pfn = start_pfn & ~(pageblock_nr_pages-1);
 884        start_page = pfn_to_page(start_pfn);
 885        end_page = start_page + pageblock_nr_pages - 1;
 886        end_pfn = start_pfn + pageblock_nr_pages - 1;
 887
 888        /* Do not cross zone boundaries */
 889        if (start_pfn < zone->zone_start_pfn)
 890                start_page = page;
 891        if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
 892                return 0;
 893
 894        return move_freepages(zone, start_page, end_page, migratetype);
 895}
 896
 897static void change_pageblock_range(struct page *pageblock_page,
 898                                        int start_order, int migratetype)
 899{
 900        int nr_pageblocks = 1 << (start_order - pageblock_order);
 901
 902        while (nr_pageblocks--) {
 903                set_pageblock_migratetype(pageblock_page, migratetype);
 904                pageblock_page += pageblock_nr_pages;
 905        }
 906}
 907
 908/* Remove an element from the buddy allocator from the fallback list */
 909static inline struct page *
 910__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
 911{
 912        struct free_area * area;
 913        int current_order;
 914        struct page *page;
 915        int migratetype, i;
 916
 917        /* Find the largest possible block of pages in the other list */
 918        for (current_order = MAX_ORDER-1; current_order >= order;
 919                                                --current_order) {
 920                for (i = 0; i < MIGRATE_TYPES - 1; i++) {
 921                        migratetype = fallbacks[start_migratetype][i];
 922
 923                        /* MIGRATE_RESERVE handled later if necessary */
 924                        if (migratetype == MIGRATE_RESERVE)
 925                                continue;
 926
 927                        area = &(zone->free_area[current_order]);
 928                        if (list_empty(&area->free_list[migratetype]))
 929                                continue;
 930
 931                        page = list_entry(area->free_list[migratetype].next,
 932                                        struct page, lru);
 933                        area->nr_free--;
 934
 935                        /*
 936                         * If breaking a large block of pages, move all free
 937                         * pages to the preferred allocation list. If falling
 938                         * back for a reclaimable kernel allocation, be more
 939                         * agressive about taking ownership of free pages
 940                         */
 941                        if (unlikely(current_order >= (pageblock_order >> 1)) ||
 942                                        start_migratetype == MIGRATE_RECLAIMABLE ||
 943                                        page_group_by_mobility_disabled) {
 944                                unsigned long pages;
 945                                pages = move_freepages_block(zone, page,
 946                                                                start_migratetype);
 947
 948                                /* Claim the whole block if over half of it is free */
 949                                if (pages >= (1 << (pageblock_order-1)) ||
 950                                                page_group_by_mobility_disabled)
 951                                        set_pageblock_migratetype(page,
 952                                                                start_migratetype);
 953
 954                                migratetype = start_migratetype;
 955                        }
 956
 957                        /* Remove the page from the freelists */
 958                        list_del(&page->lru);
 959                        rmv_page_order(page);
 960
 961                        /* Take ownership for orders >= pageblock_order */
 962                        if (current_order >= pageblock_order)
 963                                change_pageblock_range(page, current_order,
 964                                                        start_migratetype);
 965
 966                        expand(zone, page, order, current_order, area, migratetype);
 967
 968                        trace_mm_page_alloc_extfrag(page, order, current_order,
 969                                start_migratetype, migratetype);
 970
 971                        return page;
 972                }
 973        }
 974
 975        return NULL;
 976}
 977
 978/*
 979 * Do the hard work of removing an element from the buddy allocator.
 980 * Call me with the zone->lock already held.
 981 */
 982static struct page *__rmqueue(struct zone *zone, unsigned int order,
 983                                                int migratetype)
 984{
 985        struct page *page;
 986
 987retry_reserve:
 988        page = __rmqueue_smallest(zone, order, migratetype);
 989
 990        if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
 991                page = __rmqueue_fallback(zone, order, migratetype);
 992
 993                /*
 994                 * Use MIGRATE_RESERVE rather than fail an allocation. goto
 995                 * is used because __rmqueue_smallest is an inline function
 996                 * and we want just one call site
 997                 */
 998                if (!page) {
 999                        migratetype = MIGRATE_RESERVE;
1000                        goto retry_reserve;
1001                }
1002        }
1003
1004        trace_mm_page_alloc_zone_locked(page, order, migratetype);
1005        return page;
1006}
1007
1008/* 
1009 * Obtain a specified number of elements from the buddy allocator, all under
1010 * a single hold of the lock, for efficiency.  Add them to the supplied list.
1011 * Returns the number of new pages which were placed at *list.
1012 */
1013static int rmqueue_bulk(struct zone *zone, unsigned int order, 
1014                        unsigned long count, struct list_head *list,
1015                        int migratetype, int cold)
1016{
1017        int i;
1018        
1019        spin_lock(&zone->lock);
1020        for (i = 0; i < count; ++i) {
1021                struct page *page = __rmqueue(zone, order, migratetype);
1022                if (unlikely(page == NULL))
1023                        break;
1024
1025                /*
1026                 * Split buddy pages returned by expand() are received here
1027                 * in physical page order. The page is added to the callers and
1028                 * list and the list head then moves forward. From the callers
1029                 * perspective, the linked list is ordered by page number in
1030                 * some conditions. This is useful for IO devices that can
1031                 * merge IO requests if the physical pages are ordered
1032                 * properly.
1033                 */
1034                if (likely(cold == 0))
1035                        list_add(&page->lru, list);
1036                else
1037                        list_add_tail(&page->lru, list);
1038                set_page_private(page, migratetype);
1039                list = &page->lru;
1040        }
1041        __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1042        spin_unlock(&zone->lock);
1043        return i;
1044}
1045
1046#ifdef CONFIG_NUMA
1047/*
1048 * Called from the vmstat counter updater to drain pagesets of this
1049 * currently executing processor on remote nodes after they have
1050 * expired.
1051 *
1052 * Note that this function must be called with the thread pinned to
1053 * a single processor.
1054 */
1055void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1056{
1057        unsigned long flags;
1058        int to_drain;
1059
1060        local_irq_save(flags);
1061        if (pcp->count >= pcp->batch)
1062                to_drain = pcp->batch;
1063        else
1064                to_drain = pcp->count;
1065        free_pcppages_bulk(zone, to_drain, pcp);
1066        pcp->count -= to_drain;
1067        local_irq_restore(flags);
1068}
1069#endif
1070
1071/*
1072 * Drain pages of the indicated processor.
1073 *
1074 * The processor must either be the current processor and the
1075 * thread pinned to the current processor or a processor that
1076 * is not online.
1077 */
1078static void drain_pages(unsigned int cpu)
1079{
1080        unsigned long flags;
1081        struct zone *zone;
1082
1083        for_each_populated_zone(zone) {
1084                struct per_cpu_pageset *pset;
1085                struct per_cpu_pages *pcp;
1086
1087                local_irq_save(flags);
1088                pset = per_cpu_ptr(zone->pageset, cpu);
1089
1090                pcp = &pset->pcp;
1091                if (pcp->count) {
1092                        free_pcppages_bulk(zone, pcp->count, pcp);
1093                        pcp->count = 0;
1094                }
1095                local_irq_restore(flags);
1096        }
1097}
1098
1099/*
1100 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1101 */
1102void drain_local_pages(void *arg)
1103{
1104        drain_pages(smp_processor_id());
1105}
1106
1107/*
1108 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1109 */
1110void drain_all_pages(void)
1111{
1112        on_each_cpu(drain_local_pages, NULL, 1);
1113}
1114
1115#ifdef CONFIG_HIBERNATION
1116
1117void mark_free_pages(struct zone *zone)
1118{
1119        unsigned long pfn, max_zone_pfn;
1120        unsigned long flags;
1121        int order, t;
1122        struct list_head *curr;
1123
1124        if (!zone->spanned_pages)
1125                return;
1126
1127        spin_lock_irqsave(&zone->lock, flags);
1128
1129        max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1130        for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1131                if (pfn_valid(pfn)) {
1132                        struct page *page = pfn_to_page(pfn);
1133
1134                        if (!swsusp_page_is_forbidden(page))
1135                                swsusp_unset_page_free(page);
1136                }
1137
1138        for_each_migratetype_order(order, t) {
1139                list_for_each(curr, &zone->free_area[order].free_list[t]) {
1140                        unsigned long i;
1141
1142                        pfn = page_to_pfn(list_entry(curr, struct page, lru));
1143                        for (i = 0; i < (1UL << order); i++)
1144                                swsusp_set_page_free(pfn_to_page(pfn + i));
1145                }
1146        }
1147        spin_unlock_irqrestore(&zone->lock, flags);
1148}
1149#endif /* CONFIG_PM */
1150
1151/*
1152 * Free a 0-order page
1153 * cold == 1 ? free a cold page : free a hot page
1154 */
1155void free_hot_cold_page(struct page *page, int cold)
1156{
1157        struct zone *zone = page_zone(page);
1158        struct per_cpu_pages *pcp;
1159        unsigned long flags;
1160        int migratetype;
1161        int wasMlocked = __TestClearPageMlocked(page);
1162
1163        if (!free_pages_prepare(page, 0))
1164                return;
1165
1166        migratetype = get_pageblock_migratetype(page);
1167        set_page_private(page, migratetype);
1168        local_irq_save(flags);
1169        if (unlikely(wasMlocked))
1170                free_page_mlock(page);
1171        __count_vm_event(PGFREE);
1172
1173        /*
1174         * We only track unmovable, reclaimable and movable on pcp lists.
1175         * Free ISOLATE pages back to the allocator because they are being
1176         * offlined but treat RESERVE as movable pages so we can get those
1177         * areas back if necessary. Otherwise, we may have to free
1178         * excessively into the page allocator
1179         */
1180        if (migratetype >= MIGRATE_PCPTYPES) {
1181                if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1182                        free_one_page(zone, page, 0, migratetype);
1183                        goto out;
1184                }
1185                migratetype = MIGRATE_MOVABLE;
1186        }
1187
1188        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1189        if (cold)
1190                list_add_tail(&page->lru, &pcp->lists[migratetype]);
1191        else
1192                list_add(&page->lru, &pcp->lists[migratetype]);
1193        pcp->count++;
1194        if (pcp->count >= pcp->high) {
1195                free_pcppages_bulk(zone, pcp->batch, pcp);
1196                pcp->count -= pcp->batch;
1197        }
1198
1199out:
1200        local_irq_restore(flags);
1201}
1202
1203/*
1204 * split_page takes a non-compound higher-order page, and splits it into
1205 * n (1<<order) sub-pages: page[0..n]
1206 * Each sub-page must be freed individually.
1207 *
1208 * Note: this is probably too low level an operation for use in drivers.
1209 * Please consult with lkml before using this in your driver.
1210 */
1211void split_page(struct page *page, unsigned int order)
1212{
1213        int i;
1214
1215        VM_BUG_ON(PageCompound(page));
1216        VM_BUG_ON(!page_count(page));
1217
1218#ifdef CONFIG_KMEMCHECK
1219        /*
1220         * Split shadow pages too, because free(page[0]) would
1221         * otherwise free the whole shadow.
1222         */
1223        if (kmemcheck_page_is_tracked(page))
1224                split_page(virt_to_page(page[0].shadow), order);
1225#endif
1226
1227        for (i = 1; i < (1 << order); i++)
1228                set_page_refcounted(page + i);
1229}
1230
1231/*
1232 * Similar to split_page except the page is already free. As this is only
1233 * being used for migration, the migratetype of the block also changes.
1234 * As this is called with interrupts disabled, the caller is responsible
1235 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1236 * are enabled.
1237 *
1238 * Note: this is probably too low level an operation for use in drivers.
1239 * Please consult with lkml before using this in your driver.
1240 */
1241int split_free_page(struct page *page)
1242{
1243        unsigned int order;
1244        unsigned long watermark;
1245        struct zone *zone;
1246
1247        BUG_ON(!PageBuddy(page));
1248
1249        zone = page_zone(page);
1250        order = page_order(page);
1251
1252        /* Obey watermarks as if the page was being allocated */
1253        watermark = low_wmark_pages(zone) + (1 << order);
1254        if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1255                return 0;
1256
1257        /* Remove page from free list */
1258        list_del(&page->lru);
1259        zone->free_area[order].nr_free--;
1260        rmv_page_order(page);
1261        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1262
1263        /* Split into individual pages */
1264        set_page_refcounted(page);
1265        split_page(page, order);
1266
1267        if (order >= pageblock_order - 1) {
1268                struct page *endpage = page + (1 << order) - 1;
1269                for (; page < endpage; page += pageblock_nr_pages)
1270                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1271        }
1272
1273        return 1 << order;
1274}
1275
1276/*
1277 * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
1278 * we cheat by calling it from here, in the order > 0 path.  Saves a branch
1279 * or two.
1280 */
1281static inline
1282struct page *buffered_rmqueue(struct zone *preferred_zone,
1283                        struct zone *zone, int order, gfp_t gfp_flags,
1284                        int migratetype)
1285{
1286        unsigned long flags;
1287        struct page *page;
1288        int cold = !!(gfp_flags & __GFP_COLD);
1289
1290again:
1291        if (likely(order == 0)) {
1292                struct per_cpu_pages *pcp;
1293                struct list_head *list;
1294
1295                local_irq_save(flags);
1296                pcp = &this_cpu_ptr(zone->pageset)->pcp;
1297                list = &pcp->lists[migratetype];
1298                if (list_empty(list)) {
1299                        pcp->count += rmqueue_bulk(zone, 0,
1300                                        pcp->batch, list,
1301                                        migratetype, cold);
1302                        if (unlikely(list_empty(list)))
1303                                goto failed;
1304                }
1305
1306                if (cold)
1307                        page = list_entry(list->prev, struct page, lru);
1308                else
1309                        page = list_entry(list->next, struct page, lru);
1310
1311                list_del(&page->lru);
1312                pcp->count--;
1313        } else {
1314                if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1315                        /*
1316                         * __GFP_NOFAIL is not to be used in new code.
1317                         *
1318                         * All __GFP_NOFAIL callers should be fixed so that they
1319                         * properly detect and handle allocation failures.
1320                         *
1321                         * We most definitely don't want callers attempting to
1322                         * allocate greater than order-1 page units with
1323                         * __GFP_NOFAIL.
1324                         */
1325                        WARN_ON_ONCE(order > 1);
1326                }
1327                spin_lock_irqsave(&zone->lock, flags);
1328                page = __rmqueue(zone, order, migratetype);
1329                spin_unlock(&zone->lock);
1330                if (!page)
1331                        goto failed;
1332                __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1333        }
1334
1335        __count_zone_vm_events(PGALLOC, zone, 1 << order);
1336        zone_statistics(preferred_zone, zone);
1337        local_irq_restore(flags);
1338
1339        VM_BUG_ON(bad_range(zone, page));
1340        if (prep_new_page(page, order, gfp_flags))
1341                goto again;
1342        return page;
1343
1344failed:
1345        local_irq_restore(flags);
1346        return NULL;
1347}
1348
1349/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1350#define ALLOC_WMARK_MIN         WMARK_MIN
1351#define ALLOC_WMARK_LOW         WMARK_LOW
1352#define ALLOC_WMARK_HIGH        WMARK_HIGH
1353#define ALLOC_NO_WATERMARKS     0x04 /* don't check watermarks at all */
1354
1355/* Mask to get the watermark bits */
1356#define ALLOC_WMARK_MASK        (ALLOC_NO_WATERMARKS-1)
1357
1358#define ALLOC_HARDER            0x10 /* try to alloc harder */
1359#define ALLOC_HIGH              0x20 /* __GFP_HIGH set */
1360#define ALLOC_CPUSET            0x40 /* check for correct cpuset */
1361
1362#ifdef CONFIG_FAIL_PAGE_ALLOC
1363
1364static struct fail_page_alloc_attr {
1365        struct fault_attr attr;
1366
1367        u32 ignore_gfp_highmem;
1368        u32 ignore_gfp_wait;
1369        u32 min_order;
1370
1371#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1372
1373        struct dentry *ignore_gfp_highmem_file;
1374        struct dentry *ignore_gfp_wait_file;
1375        struct dentry *min_order_file;
1376
1377#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1378
1379} fail_page_alloc = {
1380        .attr = FAULT_ATTR_INITIALIZER,
1381        .ignore_gfp_wait = 1,
1382        .ignore_gfp_highmem = 1,
1383        .min_order = 1,
1384};
1385
1386static int __init setup_fail_page_alloc(char *str)
1387{
1388        return setup_fault_attr(&fail_page_alloc.attr, str);
1389}
1390__setup("fail_page_alloc=", setup_fail_page_alloc);
1391
1392static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393{
1394        if (order < fail_page_alloc.min_order)
1395                return 0;
1396        if (gfp_mask & __GFP_NOFAIL)
1397                return 0;
1398        if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399                return 0;
1400        if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1401                return 0;
1402
1403        return should_fail(&fail_page_alloc.attr, 1 << order);
1404}
1405
1406#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407
1408static int __init fail_page_alloc_debugfs(void)
1409{
1410        mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1411        struct dentry *dir;
1412        int err;
1413
1414        err = init_fault_attr_dentries(&fail_page_alloc.attr,
1415                                       "fail_page_alloc");
1416        if (err)
1417                return err;
1418        dir = fail_page_alloc.attr.dentries.dir;
1419
1420        fail_page_alloc.ignore_gfp_wait_file =
1421                debugfs_create_bool("ignore-gfp-wait", mode, dir,
1422                                      &fail_page_alloc.ignore_gfp_wait);
1423
1424        fail_page_alloc.ignore_gfp_highmem_file =
1425                debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1426                                      &fail_page_alloc.ignore_gfp_highmem);
1427        fail_page_alloc.min_order_file =
1428                debugfs_create_u32("min-order", mode, dir,
1429                                   &fail_page_alloc.min_order);
1430
1431        if (!fail_page_alloc.ignore_gfp_wait_file ||
1432            !fail_page_alloc.ignore_gfp_highmem_file ||
1433            !fail_page_alloc.min_order_file) {
1434                err = -ENOMEM;
1435                debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1436                debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1437                debugfs_remove(fail_page_alloc.min_order_file);
1438                cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1439        }
1440
1441        return err;
1442}
1443
1444late_initcall(fail_page_alloc_debugfs);
1445
1446#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1447
1448#else /* CONFIG_FAIL_PAGE_ALLOC */
1449
1450static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1451{
1452        return 0;
1453}
1454
1455#endif /* CONFIG_FAIL_PAGE_ALLOC */
1456
1457/*
1458 * Return true if free pages are above 'mark'. This takes into account the order
1459 * of the allocation.
1460 */
1461static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1462                      int classzone_idx, int alloc_flags, long free_pages)
1463{
1464        /* free_pages my go negative - that's OK */
1465        long min = mark;
1466        int o;
1467
1468        free_pages -= (1 << order) + 1;
1469        if (alloc_flags & ALLOC_HIGH)
1470                min -= min / 2;
1471        if (alloc_flags & ALLOC_HARDER)
1472                min -= min / 4;
1473
1474        if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1475                return false;
1476        for (o = 0; o < order; o++) {
1477                /* At the next order, this order's pages become unavailable */
1478                free_pages -= z->free_area[o].nr_free << o;
1479
1480                /* Require fewer higher order pages to be free */
1481                min >>= 1;
1482
1483                if (free_pages <= min)
1484                        return false;
1485        }
1486        return true;
1487}
1488
1489bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1490                      int classzone_idx, int alloc_flags)
1491{
1492        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1493                                        zone_page_state(z, NR_FREE_PAGES));
1494}
1495
1496bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1497                      int classzone_idx, int alloc_flags)
1498{
1499        long free_pages = zone_page_state(z, NR_FREE_PAGES);
1500
1501        if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1502                free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1503
1504        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1505                                                                free_pages);
1506}
1507
1508#ifdef CONFIG_NUMA
1509/*
1510 * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1511 * skip over zones that are not allowed by the cpuset, or that have
1512 * been recently (in last second) found to be nearly full.  See further
1513 * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1514 * that have to skip over a lot of full or unallowed zones.
1515 *
1516 * If the zonelist cache is present in the passed in zonelist, then
1517 * returns a pointer to the allowed node mask (either the current
1518 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1519 *
1520 * If the zonelist cache is not available for this zonelist, does
1521 * nothing and returns NULL.
1522 *
1523 * If the fullzones BITMAP in the zonelist cache is stale (more than
1524 * a second since last zap'd) then we zap it out (clear its bits.)
1525 *
1526 * We hold off even calling zlc_setup, until after we've checked the
1527 * first zone in the zonelist, on the theory that most allocations will
1528 * be satisfied from that first zone, so best to examine that zone as
1529 * quickly as we can.
1530 */
1531static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1532{
1533        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1534        nodemask_t *allowednodes;       /* zonelist_cache approximation */
1535
1536        zlc = zonelist->zlcache_ptr;
1537        if (!zlc)
1538                return NULL;
1539
1540        if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1541                bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1542                zlc->last_full_zap = jiffies;
1543        }
1544
1545        allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1546                                        &cpuset_current_mems_allowed :
1547                                        &node_states[N_HIGH_MEMORY];
1548        return allowednodes;
1549}
1550
1551/*
1552 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1553 * if it is worth looking at further for free memory:
1554 *  1) Check that the zone isn't thought to be full (doesn't have its
1555 *     bit set in the zonelist_cache fullzones BITMAP).
1556 *  2) Check that the zones node (obtained from the zonelist_cache
1557 *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1558 * Return true (non-zero) if zone is worth looking at further, or
1559 * else return false (zero) if it is not.
1560 *
1561 * This check -ignores- the distinction between various watermarks,
1562 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1563 * found to be full for any variation of these watermarks, it will
1564 * be considered full for up to one second by all requests, unless
1565 * we are so low on memory on all allowed nodes that we are forced
1566 * into the second scan of the zonelist.
1567 *
1568 * In the second scan we ignore this zonelist cache and exactly
1569 * apply the watermarks to all zones, even it is slower to do so.
1570 * We are low on memory in the second scan, and should leave no stone
1571 * unturned looking for a free page.
1572 */
1573static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1574                                                nodemask_t *allowednodes)
1575{
1576        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1577        int i;                          /* index of *z in zonelist zones */
1578        int n;                          /* node that zone *z is on */
1579
1580        zlc = zonelist->zlcache_ptr;
1581        if (!zlc)
1582                return 1;
1583
1584        i = z - zonelist->_zonerefs;
1585        n = zlc->z_to_n[i];
1586
1587        /* This zone is worth trying if it is allowed but not full */
1588        return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1589}
1590
1591/*
1592 * Given 'z' scanning a zonelist, set the corresponding bit in
1593 * zlc->fullzones, so that subsequent attempts to allocate a page
1594 * from that zone don't waste time re-examining it.
1595 */
1596static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1597{
1598        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1599        int i;                          /* index of *z in zonelist zones */
1600
1601        zlc = zonelist->zlcache_ptr;
1602        if (!zlc)
1603                return;
1604
1605        i = z - zonelist->_zonerefs;
1606
1607        set_bit(i, zlc->fullzones);
1608}
1609
1610#else   /* CONFIG_NUMA */
1611
1612static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1613{
1614        return NULL;
1615}
1616
1617static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1618                                nodemask_t *allowednodes)
1619{
1620        return 1;
1621}
1622
1623static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1624{
1625}
1626#endif  /* CONFIG_NUMA */
1627
1628/*
1629 * get_page_from_freelist goes through the zonelist trying to allocate
1630 * a page.
1631 */
1632static struct page *
1633get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1634                struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1635                struct zone *preferred_zone, int migratetype)
1636{
1637        struct zoneref *z;
1638        struct page *page = NULL;
1639        int classzone_idx;
1640        struct zone *zone;
1641        nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1642        int zlc_active = 0;             /* set if using zonelist_cache */
1643        int did_zlc_setup = 0;          /* just call zlc_setup() one time */
1644
1645        classzone_idx = zone_idx(preferred_zone);
1646zonelist_scan:
1647        /*
1648         * Scan zonelist, looking for a zone with enough free.
1649         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1650         */
1651        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1652                                                high_zoneidx, nodemask) {
1653                if (NUMA_BUILD && zlc_active &&
1654                        !zlc_zone_worth_trying(zonelist, z, allowednodes))
1655                                continue;
1656                if ((alloc_flags & ALLOC_CPUSET) &&
1657                        !cpuset_zone_allowed_softwall(zone, gfp_mask))
1658                                goto try_next_zone;
1659
1660                BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1661                if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1662                        unsigned long mark;
1663                        int ret;
1664
1665                        mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1666                        if (zone_watermark_ok(zone, order, mark,
1667                                    classzone_idx, alloc_flags))
1668                                goto try_this_zone;
1669
1670                        if (zone_reclaim_mode == 0)
1671                                goto this_zone_full;
1672
1673                        ret = zone_reclaim(zone, gfp_mask, order);
1674                        switch (ret) {
1675                        case ZONE_RECLAIM_NOSCAN:
1676                                /* did not scan */
1677                                goto try_next_zone;
1678                        case ZONE_RECLAIM_FULL:
1679                                /* scanned but unreclaimable */
1680                                goto this_zone_full;
1681                        default:
1682                                /* did we reclaim enough */
1683                                if (!zone_watermark_ok(zone, order, mark,
1684                                                classzone_idx, alloc_flags))
1685                                        goto this_zone_full;
1686                        }
1687                }
1688
1689try_this_zone:
1690                page = buffered_rmqueue(preferred_zone, zone, order,
1691                                                gfp_mask, migratetype);
1692                if (page)
1693                        break;
1694this_zone_full:
1695                if (NUMA_BUILD)
1696                        zlc_mark_zone_full(zonelist, z);
1697try_next_zone:
1698                if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1699                        /*
1700                         * we do zlc_setup after the first zone is tried but only
1701                         * if there are multiple nodes make it worthwhile
1702                         */
1703                        allowednodes = zlc_setup(zonelist, alloc_flags);
1704                        zlc_active = 1;
1705                        did_zlc_setup = 1;
1706                }
1707        }
1708
1709        if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1710                /* Disable zlc cache for second zonelist scan */
1711                zlc_active = 0;
1712                goto zonelist_scan;
1713        }
1714        return page;
1715}
1716
1717static inline int
1718should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1719                                unsigned long pages_reclaimed)
1720{
1721        /* Do not loop if specifically requested */
1722        if (gfp_mask & __GFP_NORETRY)
1723                return 0;
1724
1725        /*
1726         * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1727         * means __GFP_NOFAIL, but that may not be true in other
1728         * implementations.
1729         */
1730        if (order <= PAGE_ALLOC_COSTLY_ORDER)
1731                return 1;
1732
1733        /*
1734         * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1735         * specified, then we retry until we no longer reclaim any pages
1736         * (above), or we've reclaimed an order of pages at least as
1737         * large as the allocation's order. In both cases, if the
1738         * allocation still fails, we stop retrying.
1739         */
1740        if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1741                return 1;
1742
1743        /*
1744         * Don't let big-order allocations loop unless the caller
1745         * explicitly requests that.
1746         */
1747        if (gfp_mask & __GFP_NOFAIL)
1748                return 1;
1749
1750        return 0;
1751}
1752
1753static inline struct page *
1754__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1755        struct zonelist *zonelist, enum zone_type high_zoneidx,
1756        nodemask_t *nodemask, struct zone *preferred_zone,
1757        int migratetype)
1758{
1759        struct page *page;
1760
1761        /* Acquire the OOM killer lock for the zones in zonelist */
1762        if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1763                schedule_timeout_uninterruptible(1);
1764                return NULL;
1765        }
1766
1767        /*
1768         * Go through the zonelist yet one more time, keep very high watermark
1769         * here, this is only to catch a parallel oom killing, we must fail if
1770         * we're still under heavy pressure.
1771         */
1772        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1773                order, zonelist, high_zoneidx,
1774                ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1775                preferred_zone, migratetype);
1776        if (page)
1777                goto out;
1778
1779        if (!(gfp_mask & __GFP_NOFAIL)) {
1780                /* The OOM killer will not help higher order allocs */
1781                if (order > PAGE_ALLOC_COSTLY_ORDER)
1782                        goto out;
1783                /* The OOM killer does not needlessly kill tasks for lowmem */
1784                if (high_zoneidx < ZONE_NORMAL)
1785                        goto out;
1786                /*
1787                 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1788                 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1789                 * The caller should handle page allocation failure by itself if
1790                 * it specifies __GFP_THISNODE.
1791                 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1792                 */
1793                if (gfp_mask & __GFP_THISNODE)
1794                        goto out;
1795        }
1796        /* Exhausted what can be done so it's blamo time */
1797        out_of_memory(zonelist, gfp_mask, order, nodemask);
1798
1799out:
1800        clear_zonelist_oom(zonelist, gfp_mask);
1801        return page;
1802}
1803
1804#ifdef CONFIG_COMPACTION
1805/* Try memory compaction for high-order allocations before reclaim */
1806static struct page *
1807__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1808        struct zonelist *zonelist, enum zone_type high_zoneidx,
1809        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1810        int migratetype, unsigned long *did_some_progress,
1811        bool sync_migration)
1812{
1813        struct page *page;
1814
1815        if (!order || compaction_deferred(preferred_zone))
1816                return NULL;
1817
1818        current->flags |= PF_MEMALLOC;
1819        *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1820                                                nodemask, sync_migration);
1821        current->flags &= ~PF_MEMALLOC;
1822        if (*did_some_progress != COMPACT_SKIPPED) {
1823
1824                /* Page migration frees to the PCP lists but we want merging */
1825                drain_pages(get_cpu());
1826                put_cpu();
1827
1828                page = get_page_from_freelist(gfp_mask, nodemask,
1829                                order, zonelist, high_zoneidx,
1830                                alloc_flags, preferred_zone,
1831                                migratetype);
1832                if (page) {
1833                        preferred_zone->compact_considered = 0;
1834                        preferred_zone->compact_defer_shift = 0;
1835                        count_vm_event(COMPACTSUCCESS);
1836                        return page;
1837                }
1838
1839                /*
1840                 * It's bad if compaction run occurs and fails.
1841                 * The most likely reason is that pages exist,
1842                 * but not enough to satisfy watermarks.
1843                 */
1844                count_vm_event(COMPACTFAIL);
1845                defer_compaction(preferred_zone);
1846
1847                cond_resched();
1848        }
1849
1850        return NULL;
1851}
1852#else
1853static inline struct page *
1854__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1855        struct zonelist *zonelist, enum zone_type high_zoneidx,
1856        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1857        int migratetype, unsigned long *did_some_progress,
1858        bool sync_migration)
1859{
1860        return NULL;
1861}
1862#endif /* CONFIG_COMPACTION */
1863
1864/* The really slow allocator path where we enter direct reclaim */
1865static inline struct page *
1866__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1867        struct zonelist *zonelist, enum zone_type high_zoneidx,
1868        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1869        int migratetype, unsigned long *did_some_progress)
1870{
1871        struct page *page = NULL;
1872        struct reclaim_state reclaim_state;
1873        bool drained = false;
1874
1875        cond_resched();
1876
1877        /* We now go into synchronous reclaim */
1878        cpuset_memory_pressure_bump();
1879        current->flags |= PF_MEMALLOC;
1880        lockdep_set_current_reclaim_state(gfp_mask);
1881        reclaim_state.reclaimed_slab = 0;
1882        current->reclaim_state = &reclaim_state;
1883
1884        *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1885
1886        current->reclaim_state = NULL;
1887        lockdep_clear_current_reclaim_state();
1888        current->flags &= ~PF_MEMALLOC;
1889
1890        cond_resched();
1891
1892        if (unlikely(!(*did_some_progress)))
1893                return NULL;
1894
1895retry:
1896        page = get_page_from_freelist(gfp_mask, nodemask, order,
1897                                        zonelist, high_zoneidx,
1898                                        alloc_flags, preferred_zone,
1899                                        migratetype);
1900
1901        /*
1902         * If an allocation failed after direct reclaim, it could be because
1903         * pages are pinned on the per-cpu lists. Drain them and try again
1904         */
1905        if (!page && !drained) {
1906                drain_all_pages();
1907                drained = true;
1908                goto retry;
1909        }
1910
1911        return page;
1912}
1913
1914/*
1915 * This is called in the allocator slow-path if the allocation request is of
1916 * sufficient urgency to ignore watermarks and take other desperate measures
1917 */
1918static inline struct page *
1919__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1920        struct zonelist *zonelist, enum zone_type high_zoneidx,
1921        nodemask_t *nodemask, struct zone *preferred_zone,
1922        int migratetype)
1923{
1924        struct page *page;
1925
1926        do {
1927                page = get_page_from_freelist(gfp_mask, nodemask, order,
1928                        zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1929                        preferred_zone, migratetype);
1930
1931                if (!page && gfp_mask & __GFP_NOFAIL)
1932                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
1933        } while (!page && (gfp_mask & __GFP_NOFAIL));
1934
1935        return page;
1936}
1937
1938static inline
1939void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1940                                                enum zone_type high_zoneidx,
1941                                                enum zone_type classzone_idx)
1942{
1943        struct zoneref *z;
1944        struct zone *zone;
1945
1946        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1947                wakeup_kswapd(zone, order, classzone_idx);
1948}
1949
1950static inline int
1951gfp_to_alloc_flags(gfp_t gfp_mask)
1952{
1953        int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1954        const gfp_t wait = gfp_mask & __GFP_WAIT;
1955
1956        /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1957        BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
1958
1959        /*
1960         * The caller may dip into page reserves a bit more if the caller
1961         * cannot run direct reclaim, or if the caller has realtime scheduling
1962         * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
1963         * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1964         */
1965        alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
1966
1967        if (!wait) {
1968                /*
1969                 * Not worth trying to allocate harder for
1970                 * __GFP_NOMEMALLOC even if it can't schedule.
1971                 */
1972                if  (!(gfp_mask & __GFP_NOMEMALLOC))
1973                        alloc_flags |= ALLOC_HARDER;
1974                /*
1975                 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1976                 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1977                 */
1978                alloc_flags &= ~ALLOC_CPUSET;
1979        } else if (unlikely(rt_task(current)) && !in_interrupt())
1980                alloc_flags |= ALLOC_HARDER;
1981
1982        if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1983                if (!in_interrupt() &&
1984                    ((current->flags & PF_MEMALLOC) ||
1985                     unlikely(test_thread_flag(TIF_MEMDIE))))
1986                        alloc_flags |= ALLOC_NO_WATERMARKS;
1987        }
1988
1989        return alloc_flags;
1990}
1991
1992static inline struct page *
1993__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1994        struct zonelist *zonelist, enum zone_type high_zoneidx,
1995        nodemask_t *nodemask, struct zone *preferred_zone,
1996        int migratetype)
1997{
1998        const gfp_t wait = gfp_mask & __GFP_WAIT;
1999        struct page *page = NULL;
2000        int alloc_flags;
2001        unsigned long pages_reclaimed = 0;
2002        unsigned long did_some_progress;
2003        bool sync_migration = false;
2004
2005        /*
2006         * In the slowpath, we sanity check order to avoid ever trying to
2007         * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2008         * be using allocators in order of preference for an area that is
2009         * too large.
2010         */
2011        if (order >= MAX_ORDER) {
2012                WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2013                return NULL;
2014        }
2015
2016        /*
2017         * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2018         * __GFP_NOWARN set) should not cause reclaim since the subsystem
2019         * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2020         * using a larger set of nodes after it has established that the
2021         * allowed per node queues are empty and that nodes are
2022         * over allocated.
2023         */
2024        if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2025                goto nopage;
2026
2027restart:
2028        if (!(gfp_mask & __GFP_NO_KSWAPD))
2029                wake_all_kswapd(order, zonelist, high_zoneidx,
2030                                                zone_idx(preferred_zone));
2031
2032        /*
2033         * OK, we're below the kswapd watermark and have kicked background
2034         * reclaim. Now things get more complex, so set up alloc_flags according
2035         * to how we want to proceed.
2036         */
2037        alloc_flags = gfp_to_alloc_flags(gfp_mask);
2038
2039        /*
2040         * Find the true preferred zone if the allocation is unconstrained by
2041         * cpusets.
2042         */
2043        if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2044                first_zones_zonelist(zonelist, high_zoneidx, NULL,
2045                                        &preferred_zone);
2046
2047rebalance:
2048        /* This is the last chance, in general, before the goto nopage. */
2049        page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2050                        high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2051                        preferred_zone, migratetype);
2052        if (page)
2053                goto got_pg;
2054
2055        /* Allocate without watermarks if the context allows */
2056        if (alloc_flags & ALLOC_NO_WATERMARKS) {
2057                page = __alloc_pages_high_priority(gfp_mask, order,
2058                                zonelist, high_zoneidx, nodemask,
2059                                preferred_zone, migratetype);
2060                if (page)
2061                        goto got_pg;
2062        }
2063
2064        /* Atomic allocations - we can't balance anything */
2065        if (!wait)
2066                goto nopage;
2067
2068        /* Avoid recursion of direct reclaim */
2069        if (current->flags & PF_MEMALLOC)
2070                goto nopage;
2071
2072        /* Avoid allocations with no watermarks from looping endlessly */
2073        if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2074                goto nopage;
2075
2076        /*
2077         * Try direct compaction. The first pass is asynchronous. Subsequent
2078         * attempts after direct reclaim are synchronous
2079         */
2080        page = __alloc_pages_direct_compact(gfp_mask, order,
2081                                        zonelist, high_zoneidx,
2082                                        nodemask,
2083                                        alloc_flags, preferred_zone,
2084                                        migratetype, &did_some_progress,
2085                                        sync_migration);
2086        if (page)
2087                goto got_pg;
2088        sync_migration = true;
2089
2090        /* Try direct reclaim and then allocating */
2091        page = __alloc_pages_direct_reclaim(gfp_mask, order,
2092                                        zonelist, high_zoneidx,
2093                                        nodemask,
2094                                        alloc_flags, preferred_zone,
2095                                        migratetype, &did_some_progress);
2096        if (page)
2097                goto got_pg;
2098
2099        /*
2100         * If we failed to make any progress reclaiming, then we are
2101         * running out of options and have to consider going OOM
2102         */
2103        if (!did_some_progress) {
2104                if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2105                        if (oom_killer_disabled)
2106                                goto nopage;
2107                        page = __alloc_pages_may_oom(gfp_mask, order,
2108                                        zonelist, high_zoneidx,
2109                                        nodemask, preferred_zone,
2110                                        migratetype);
2111                        if (page)
2112                                goto got_pg;
2113
2114                        if (!(gfp_mask & __GFP_NOFAIL)) {
2115                                /*
2116                                 * The oom killer is not called for high-order
2117                                 * allocations that may fail, so if no progress
2118                                 * is being made, there are no other options and
2119                                 * retrying is unlikely to help.
2120                                 */
2121                                if (order > PAGE_ALLOC_COSTLY_ORDER)
2122                                        goto nopage;
2123                                /*
2124                                 * The oom killer is not called for lowmem
2125                                 * allocations to prevent needlessly killing
2126                                 * innocent tasks.
2127                                 */
2128                                if (high_zoneidx < ZONE_NORMAL)
2129                                        goto nopage;
2130                        }
2131
2132                        goto restart;
2133                }
2134        }
2135
2136        /* Check if we should retry the allocation */
2137        pages_reclaimed += did_some_progress;
2138        if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2139                /* Wait for some write requests to complete then retry */
2140                wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2141                goto rebalance;
2142        } else {
2143                /*
2144                 * High-order allocations do not necessarily loop after
2145                 * direct reclaim and reclaim/compaction depends on compaction
2146                 * being called after reclaim so call directly if necessary
2147                 */
2148                page = __alloc_pages_direct_compact(gfp_mask, order,
2149                                        zonelist, high_zoneidx,
2150                                        nodemask,
2151                                        alloc_flags, preferred_zone,
2152                                        migratetype, &did_some_progress,
2153                                        sync_migration);
2154                if (page)
2155                        goto got_pg;
2156        }
2157
2158nopage:
2159        if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2160                printk(KERN_WARNING "%s: page allocation failure."
2161                        " order:%d, mode:0x%x\n",
2162                        current->comm, order, gfp_mask);
2163                dump_stack();
2164                show_mem();
2165        }
2166        return page;
2167got_pg:
2168        if (kmemcheck_enabled)
2169                kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2170        return page;
2171
2172}
2173
2174/*
2175 * This is the 'heart' of the zoned buddy allocator.
2176 */
2177struct page *
2178__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2179                        struct zonelist *zonelist, nodemask_t *nodemask)
2180{
2181        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2182        struct zone *preferred_zone;
2183        struct page *page;
2184        int migratetype = allocflags_to_migratetype(gfp_mask);
2185
2186        gfp_mask &= gfp_allowed_mask;
2187
2188        lockdep_trace_alloc(gfp_mask);
2189
2190        might_sleep_if(gfp_mask & __GFP_WAIT);
2191
2192        if (should_fail_alloc_page(gfp_mask, order))
2193                return NULL;
2194
2195        /*
2196         * Check the zones suitable for the gfp_mask contain at least one
2197         * valid zone. It's possible to have an empty zonelist as a result
2198         * of GFP_THISNODE and a memoryless node
2199         */
2200        if (unlikely(!zonelist->_zonerefs->zone))
2201                return NULL;
2202
2203        get_mems_allowed();
2204        /* The preferred zone is used for statistics later */
2205        first_zones_zonelist(zonelist, high_zoneidx,
2206                                nodemask ? : &cpuset_current_mems_allowed,
2207                                &preferred_zone);
2208        if (!preferred_zone) {
2209                put_mems_allowed();
2210                return NULL;
2211        }
2212
2213        /* First allocation attempt */
2214        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2215                        zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2216                        preferred_zone, migratetype);
2217        if (unlikely(!page))
2218                page = __alloc_pages_slowpath(gfp_mask, order,
2219                                zonelist, high_zoneidx, nodemask,
2220                                preferred_zone, migratetype);
2221        put_mems_allowed();
2222
2223        trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2224        return page;
2225}
2226EXPORT_SYMBOL(__alloc_pages_nodemask);
2227
2228/*
2229 * Common helper functions.
2230 */
2231unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2232{
2233        struct page *page;
2234
2235        /*
2236         * __get_free_pages() returns a 32-bit address, which cannot represent
2237         * a highmem page
2238         */
2239        VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2240
2241        page = alloc_pages(gfp_mask, order);
2242        if (!page)
2243                return 0;
2244        return (unsigned long) page_address(page);
2245}
2246EXPORT_SYMBOL(__get_free_pages);
2247
2248unsigned long get_zeroed_page(gfp_t gfp_mask)
2249{
2250        return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2251}
2252EXPORT_SYMBOL(get_zeroed_page);
2253
2254void __pagevec_free(struct pagevec *pvec)
2255{
2256        int i = pagevec_count(pvec);
2257
2258        while (--i >= 0) {
2259                trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2260                free_hot_cold_page(pvec->pages[i], pvec->cold);
2261        }
2262}
2263
2264void __free_pages(struct page *page, unsigned int order)
2265{
2266        if (put_page_testzero(page)) {
2267                if (order == 0)
2268                        free_hot_cold_page(page, 0);
2269                else
2270                        __free_pages_ok(page, order);
2271        }
2272}
2273
2274EXPORT_SYMBOL(__free_pages);
2275
2276void free_pages(unsigned long addr, unsigned int order)
2277{
2278        if (addr != 0) {
2279                VM_BUG_ON(!virt_addr_valid((void *)addr));
2280                __free_pages(virt_to_page((void *)addr), order);
2281        }
2282}
2283
2284EXPORT_SYMBOL(free_pages);
2285
2286/**
2287 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2288 * @size: the number of bytes to allocate
2289 * @gfp_mask: GFP flags for the allocation
2290 *
2291 * This function is similar to alloc_pages(), except that it allocates the
2292 * minimum number of pages to satisfy the request.  alloc_pages() can only
2293 * allocate memory in power-of-two pages.
2294 *
2295 * This function is also limited by MAX_ORDER.
2296 *
2297 * Memory allocated by this function must be released by free_pages_exact().
2298 */
2299void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2300{
2301        unsigned int order = get_order(size);
2302        unsigned long addr;
2303
2304        addr = __get_free_pages(gfp_mask, order);
2305        if (addr) {
2306                unsigned long alloc_end = addr + (PAGE_SIZE << order);
2307                unsigned long used = addr + PAGE_ALIGN(size);
2308
2309                split_page(virt_to_page((void *)addr), order);
2310                while (used < alloc_end) {
2311                        free_page(used);
2312                        used += PAGE_SIZE;
2313                }
2314        }
2315
2316        return (void *)addr;
2317}
2318EXPORT_SYMBOL(alloc_pages_exact);
2319
2320/**
2321 * free_pages_exact - release memory allocated via alloc_pages_exact()
2322 * @virt: the value returned by alloc_pages_exact.
2323 * @size: size of allocation, same value as passed to alloc_pages_exact().
2324 *
2325 * Release the memory allocated by a previous call to alloc_pages_exact.
2326 */
2327void free_pages_exact(void *virt, size_t size)
2328{
2329        unsigned long addr = (unsigned long)virt;
2330        unsigned long end = addr + PAGE_ALIGN(size);
2331
2332        while (addr < end) {
2333                free_page(addr);
2334                addr += PAGE_SIZE;
2335        }
2336}
2337EXPORT_SYMBOL(free_pages_exact);
2338
2339static unsigned int nr_free_zone_pages(int offset)
2340{
2341        struct zoneref *z;
2342        struct zone *zone;
2343
2344        /* Just pick one node, since fallback list is circular */
2345        unsigned int sum = 0;
2346
2347        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2348
2349        for_each_zone_zonelist(zone, z, zonelist, offset) {
2350                unsigned long size = zone->present_pages;
2351                unsigned long high = high_wmark_pages(zone);
2352                if (size > high)
2353                        sum += size - high;
2354        }
2355
2356        return sum;
2357}
2358
2359/*
2360 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2361 */
2362unsigned int nr_free_buffer_pages(void)
2363{
2364        return nr_free_zone_pages(gfp_zone(GFP_USER));
2365}
2366EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2367
2368/*
2369 * Amount of free RAM allocatable within all zones
2370 */
2371unsigned int nr_free_pagecache_pages(void)
2372{
2373        return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2374}
2375
2376static inline void show_node(struct zone *zone)
2377{
2378        if (NUMA_BUILD)
2379                printk("Node %d ", zone_to_nid(zone));
2380}
2381
2382void si_meminfo(struct sysinfo *val)
2383{
2384        val->totalram = totalram_pages;
2385        val->sharedram = 0;
2386        val->freeram = global_page_state(NR_FREE_PAGES);
2387        val->bufferram = nr_blockdev_pages();
2388        val->totalhigh = totalhigh_pages;
2389        val->freehigh = nr_free_highpages();
2390        val->mem_unit = PAGE_SIZE;
2391}
2392
2393EXPORT_SYMBOL(si_meminfo);
2394
2395#ifdef CONFIG_NUMA
2396void si_meminfo_node(struct sysinfo *val, int nid)
2397{
2398        pg_data_t *pgdat = NODE_DATA(nid);
2399
2400        val->totalram = pgdat->node_present_pages;
2401        val->freeram = node_page_state(nid, NR_FREE_PAGES);
2402#ifdef CONFIG_HIGHMEM
2403        val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2404        val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2405                        NR_FREE_PAGES);
2406#else
2407        val->totalhigh = 0;
2408        val->freehigh = 0;
2409#endif
2410        val->mem_unit = PAGE_SIZE;
2411}
2412#endif
2413
2414#define K(x) ((x) << (PAGE_SHIFT-10))
2415
2416/*
2417 * Show free area list (used inside shift_scroll-lock stuff)
2418 * We also calculate the percentage fragmentation. We do this by counting the
2419 * memory on each free list with the exception of the first item on the list.
2420 */
2421void show_free_areas(void)
2422{
2423        int cpu;
2424        struct zone *zone;
2425
2426        for_each_populated_zone(zone) {
2427                show_node(zone);
2428                printk("%s per-cpu:\n", zone->name);
2429
2430                for_each_online_cpu(cpu) {
2431                        struct per_cpu_pageset *pageset;
2432
2433                        pageset = per_cpu_ptr(zone->pageset, cpu);
2434
2435                        printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2436                               cpu, pageset->pcp.high,
2437                               pageset->pcp.batch, pageset->pcp.count);
2438                }
2439        }
2440
2441        printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2442                " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2443                " unevictable:%lu"
2444                " dirty:%lu writeback:%lu unstable:%lu\n"
2445                " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2446                " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2447                global_page_state(NR_ACTIVE_ANON),
2448                global_page_state(NR_INACTIVE_ANON),
2449                global_page_state(NR_ISOLATED_ANON),
2450                global_page_state(NR_ACTIVE_FILE),
2451                global_page_state(NR_INACTIVE_FILE),
2452                global_page_state(NR_ISOLATED_FILE),
2453                global_page_state(NR_UNEVICTABLE),
2454                global_page_state(NR_FILE_DIRTY),
2455                global_page_state(NR_WRITEBACK),
2456                global_page_state(NR_UNSTABLE_NFS),
2457                global_page_state(NR_FREE_PAGES),
2458                global_page_state(NR_SLAB_RECLAIMABLE),
2459                global_page_state(NR_SLAB_UNRECLAIMABLE),
2460                global_page_state(NR_FILE_MAPPED),
2461                global_page_state(NR_SHMEM),
2462                global_page_state(NR_PAGETABLE),
2463                global_page_state(NR_BOUNCE));
2464
2465        for_each_populated_zone(zone) {
2466                int i;
2467
2468                show_node(zone);
2469                printk("%s"
2470                        " free:%lukB"
2471                        " min:%lukB"
2472                        " low:%lukB"
2473                        " high:%lukB"
2474                        " active_anon:%lukB"
2475                        " inactive_anon:%lukB"
2476                        " active_file:%lukB"
2477                        " inactive_file:%lukB"
2478                        " unevictable:%lukB"
2479                        " isolated(anon):%lukB"
2480                        " isolated(file):%lukB"
2481                        " present:%lukB"
2482                        " mlocked:%lukB"
2483                        " dirty:%lukB"
2484                        " writeback:%lukB"
2485                        " mapped:%lukB"
2486                        " shmem:%lukB"
2487                        " slab_reclaimable:%lukB"
2488                        " slab_unreclaimable:%lukB"
2489                        " kernel_stack:%lukB"
2490                        " pagetables:%lukB"
2491                        " unstable:%lukB"
2492                        " bounce:%lukB"
2493                        " writeback_tmp:%lukB"
2494                        " pages_scanned:%lu"
2495                        " all_unreclaimable? %s"
2496                        "\n",
2497                        zone->name,
2498                        K(zone_page_state(zone, NR_FREE_PAGES)),
2499                        K(min_wmark_pages(zone)),
2500                        K(low_wmark_pages(zone)),
2501                        K(high_wmark_pages(zone)),
2502                        K(zone_page_state(zone, NR_ACTIVE_ANON)),
2503                        K(zone_page_state(zone, NR_INACTIVE_ANON)),
2504                        K(zone_page_state(zone, NR_ACTIVE_FILE)),
2505                        K(zone_page_state(zone, NR_INACTIVE_FILE)),
2506                        K(zone_page_state(zone, NR_UNEVICTABLE)),
2507                        K(zone_page_state(zone, NR_ISOLATED_ANON)),
2508                        K(zone_page_state(zone, NR_ISOLATED_FILE)),
2509                        K(zone->present_pages),
2510                        K(zone_page_state(zone, NR_MLOCK)),
2511                        K(zone_page_state(zone, NR_FILE_DIRTY)),
2512                        K(zone_page_state(zone, NR_WRITEBACK)),
2513                        K(zone_page_state(zone, NR_FILE_MAPPED)),
2514                        K(zone_page_state(zone, NR_SHMEM)),
2515                        K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2516                        K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2517                        zone_page_state(zone, NR_KERNEL_STACK) *
2518                                THREAD_SIZE / 1024,
2519                        K(zone_page_state(zone, NR_PAGETABLE)),
2520                        K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2521                        K(zone_page_state(zone, NR_BOUNCE)),
2522                        K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2523                        zone->pages_scanned,
2524                        (zone->all_unreclaimable ? "yes" : "no")
2525                        );
2526                printk("lowmem_reserve[]:");
2527                for (i = 0; i < MAX_NR_ZONES; i++)
2528                        printk(" %lu", zone->lowmem_reserve[i]);
2529                printk("\n");
2530        }
2531
2532        for_each_populated_zone(zone) {
2533                unsigned long nr[MAX_ORDER], flags, order, total = 0;
2534
2535                show_node(zone);
2536                printk("%s: ", zone->name);
2537
2538                spin_lock_irqsave(&zone->lock, flags);
2539                for (order = 0; order < MAX_ORDER; order++) {
2540                        nr[order] = zone->free_area[order].nr_free;
2541                        total += nr[order] << order;
2542                }
2543                spin_unlock_irqrestore(&zone->lock, flags);
2544                for (order = 0; order < MAX_ORDER; order++)
2545                        printk("%lu*%lukB ", nr[order], K(1UL) << order);
2546                printk("= %lukB\n", K(total));
2547        }
2548
2549        printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2550
2551        show_swap_cache_info();
2552}
2553
2554static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2555{
2556        zoneref->zone = zone;
2557        zoneref->zone_idx = zone_idx(zone);
2558}
2559
2560/*
2561 * Builds allocation fallback zone lists.
2562 *
2563 * Add all populated zones of a node to the zonelist.
2564 */
2565static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2566                                int nr_zones, enum zone_type zone_type)
2567{
2568        struct zone *zone;
2569
2570        BUG_ON(zone_type >= MAX_NR_ZONES);
2571        zone_type++;
2572
2573        do {
2574                zone_type--;
2575                zone = pgdat->node_zones + zone_type;
2576                if (populated_zone(zone)) {
2577                        zoneref_set_zone(zone,
2578                                &zonelist->_zonerefs[nr_zones++]);
2579                        check_highest_zone(zone_type);
2580                }
2581
2582        } while (zone_type);
2583        return nr_zones;
2584}
2585
2586
2587/*
2588 *  zonelist_order:
2589 *  0 = automatic detection of better ordering.
2590 *  1 = order by ([node] distance, -zonetype)
2591 *  2 = order by (-zonetype, [node] distance)
2592 *
2593 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2594 *  the same zonelist. So only NUMA can configure this param.
2595 */
2596#define ZONELIST_ORDER_DEFAULT  0
2597#define ZONELIST_ORDER_NODE     1
2598#define ZONELIST_ORDER_ZONE     2
2599
2600/* zonelist order in the kernel.
2601 * set_zonelist_order() will set this to NODE or ZONE.
2602 */
2603static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2604static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2605
2606
2607#ifdef CONFIG_NUMA
2608/* The value user specified ....changed by config */
2609static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2610/* string for sysctl */
2611#define NUMA_ZONELIST_ORDER_LEN 16
2612char numa_zonelist_order[16] = "default";
2613
2614/*
2615 * interface for configure zonelist ordering.
2616 * command line option "numa_zonelist_order"
2617 *      = "[dD]efault   - default, automatic configuration.
2618 *      = "[nN]ode      - order by node locality, then by zone within node
2619 *      = "[zZ]one      - order by zone, then by locality within zone
2620 */
2621
2622static int __parse_numa_zonelist_order(char *s)
2623{
2624        if (*s == 'd' || *s == 'D') {
2625                user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2626        } else if (*s == 'n' || *s == 'N') {
2627                user_zonelist_order = ZONELIST_ORDER_NODE;
2628        } else if (*s == 'z' || *s == 'Z') {
2629                user_zonelist_order = ZONELIST_ORDER_ZONE;
2630        } else {
2631                printk(KERN_WARNING
2632                        "Ignoring invalid numa_zonelist_order value:  "
2633                        "%s\n", s);
2634                return -EINVAL;
2635        }
2636        return 0;
2637}
2638
2639static __init int setup_numa_zonelist_order(char *s)
2640{
2641        int ret;
2642
2643        if (!s)
2644                return 0;
2645
2646        ret = __parse_numa_zonelist_order(s);
2647        if (ret == 0)
2648                strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2649
2650        return ret;
2651}
2652early_param("numa_zonelist_order", setup_numa_zonelist_order);
2653
2654/*
2655 * sysctl handler for numa_zonelist_order
2656 */
2657int numa_zonelist_order_handler(ctl_table *table, int write,
2658                void __user *buffer, size_t *length,
2659                loff_t *ppos)
2660{
2661        char saved_string[NUMA_ZONELIST_ORDER_LEN];
2662        int ret;
2663        static DEFINE_MUTEX(zl_order_mutex);
2664
2665        mutex_lock(&zl_order_mutex);
2666        if (write)
2667                strcpy(saved_string, (char*)table->data);
2668        ret = proc_dostring(table, write, buffer, length, ppos);
2669        if (ret)
2670                goto out;
2671        if (write) {
2672                int oldval = user_zonelist_order;
2673                if (__parse_numa_zonelist_order((char*)table->data)) {
2674                        /*
2675                         * bogus value.  restore saved string
2676                         */
2677                        strncpy((char*)table->data, saved_string,
2678                                NUMA_ZONELIST_ORDER_LEN);
2679                        user_zonelist_order = oldval;
2680                } else if (oldval != user_zonelist_order) {
2681                        mutex_lock(&zonelists_mutex);
2682                        build_all_zonelists(NULL);
2683                        mutex_unlock(&zonelists_mutex);
2684                }
2685        }
2686out:
2687        mutex_unlock(&zl_order_mutex);
2688        return ret;
2689}
2690
2691
2692#define MAX_NODE_LOAD (nr_online_nodes)
2693static int node_load[MAX_NUMNODES];
2694
2695/**
2696 * find_next_best_node - find the next node that should appear in a given node's fallback list
2697 * @node: node whose fallback list we're appending
2698 * @used_node_mask: nodemask_t of already used nodes
2699 *
2700 * We use a number of factors to determine which is the next node that should
2701 * appear on a given node's fallback list.  The node should not have appeared
2702 * already in @node's fallback list, and it should be the next closest node
2703 * according to the distance array (which contains arbitrary distance values
2704 * from each node to each node in the system), and should also prefer nodes
2705 * with no CPUs, since presumably they'll have very little allocation pressure
2706 * on them otherwise.
2707 * It returns -1 if no node is found.
2708 */
2709static int find_next_best_node(int node, nodemask_t *used_node_mask)
2710{
2711        int n, val;
2712        int min_val = INT_MAX;
2713        int best_node = -1;
2714        const struct cpumask *tmp = cpumask_of_node(0);
2715
2716        /* Use the local node if we haven't already */
2717        if (!node_isset(node, *used_node_mask)) {
2718                node_set(node, *used_node_mask);
2719                return node;
2720        }
2721
2722        for_each_node_state(n, N_HIGH_MEMORY) {
2723
2724                /* Don't want a node to appear more than once */
2725                if (node_isset(n, *used_node_mask))
2726                        continue;
2727
2728                /* Use the distance array to find the distance */
2729                val = node_distance(node, n);
2730
2731                /* Penalize nodes under us ("prefer the next node") */
2732                val += (n < node);
2733
2734                /* Give preference to headless and unused nodes */
2735                tmp = cpumask_of_node(n);
2736                if (!cpumask_empty(tmp))
2737                        val += PENALTY_FOR_NODE_WITH_CPUS;
2738
2739                /* Slight preference for less loaded node */
2740                val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2741                val += node_load[n];
2742
2743                if (val < min_val) {
2744                        min_val = val;
2745                        best_node = n;
2746                }
2747        }
2748
2749        if (best_node >= 0)
2750                node_set(best_node, *used_node_mask);
2751
2752        return best_node;
2753}
2754
2755
2756/*
2757 * Build zonelists ordered by node and zones within node.
2758 * This results in maximum locality--normal zone overflows into local
2759 * DMA zone, if any--but risks exhausting DMA zone.
2760 */
2761static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2762{
2763        int j;
2764        struct zonelist *zonelist;
2765
2766        zonelist = &pgdat->node_zonelists[0];
2767        for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2768                ;
2769        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2770                                                        MAX_NR_ZONES - 1);
2771        zonelist->_zonerefs[j].zone = NULL;
2772        zonelist->_zonerefs[j].zone_idx = 0;
2773}
2774
2775/*
2776 * Build gfp_thisnode zonelists
2777 */
2778static void build_thisnode_zonelists(pg_data_t *pgdat)
2779{
2780        int j;
2781        struct zonelist *zonelist;
2782
2783        zonelist = &pgdat->node_zonelists[1];
2784        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2785        zonelist->_zonerefs[j].zone = NULL;
2786        zonelist->_zonerefs[j].zone_idx = 0;
2787}
2788
2789/*
2790 * Build zonelists ordered by zone and nodes within zones.
2791 * This results in conserving DMA zone[s] until all Normal memory is
2792 * exhausted, but results in overflowing to remote node while memory
2793 * may still exist in local DMA zone.
2794 */
2795static int node_order[MAX_NUMNODES];
2796
2797static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2798{
2799        int pos, j, node;
2800        int zone_type;          /* needs to be signed */
2801        struct zone *z;
2802        struct zonelist *zonelist;
2803
2804        zonelist = &pgdat->node_zonelists[0];
2805        pos = 0;
2806        for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2807                for (j = 0; j < nr_nodes; j++) {
2808                        node = node_order[j];
2809                        z = &NODE_DATA(node)->node_zones[zone_type];
2810                        if (populated_zone(z)) {
2811                                zoneref_set_zone(z,
2812                                        &zonelist->_zonerefs[pos++]);
2813                                check_highest_zone(zone_type);
2814                        }
2815                }
2816        }
2817        zonelist->_zonerefs[pos].zone = NULL;
2818        zonelist->_zonerefs[pos].zone_idx = 0;
2819}
2820
2821static int default_zonelist_order(void)
2822{
2823        int nid, zone_type;
2824        unsigned long low_kmem_size,total_size;
2825        struct zone *z;
2826        int average_size;
2827        /*
2828         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2829         * If they are really small and used heavily, the system can fall
2830         * into OOM very easily.
2831         * This function detect ZONE_DMA/DMA32 size and configures zone order.
2832         */
2833        /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2834        low_kmem_size = 0;
2835        total_size = 0;
2836        for_each_online_node(nid) {
2837                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2838                        z = &NODE_DATA(nid)->node_zones[zone_type];
2839                        if (populated_zone(z)) {
2840                                if (zone_type < ZONE_NORMAL)
2841                                        low_kmem_size += z->present_pages;
2842                                total_size += z->present_pages;
2843                        } else if (zone_type == ZONE_NORMAL) {
2844                                /*
2845                                 * If any node has only lowmem, then node order
2846                                 * is preferred to allow kernel allocations
2847                                 * locally; otherwise, they can easily infringe
2848                                 * on other nodes when there is an abundance of
2849                                 * lowmem available to allocate from.
2850                                 */
2851                                return ZONELIST_ORDER_NODE;
2852                        }
2853                }
2854        }
2855        if (!low_kmem_size ||  /* there are no DMA area. */
2856            low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2857                return ZONELIST_ORDER_NODE;
2858        /*
2859         * look into each node's config.
2860         * If there is a node whose DMA/DMA32 memory is very big area on
2861         * local memory, NODE_ORDER may be suitable.
2862         */
2863        average_size = total_size /
2864                                (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2865        for_each_online_node(nid) {
2866                low_kmem_size = 0;
2867                total_size = 0;
2868                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2869                        z = &NODE_DATA(nid)->node_zones[zone_type];
2870                        if (populated_zone(z)) {
2871                                if (zone_type < ZONE_NORMAL)
2872                                        low_kmem_size += z->present_pages;
2873                                total_size += z->present_pages;
2874                        }
2875                }
2876                if (low_kmem_size &&
2877                    total_size > average_size && /* ignore small node */
2878                    low_kmem_size > total_size * 70/100)
2879                        return ZONELIST_ORDER_NODE;
2880        }
2881        return ZONELIST_ORDER_ZONE;
2882}
2883
2884static void set_zonelist_order(void)
2885{
2886        if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2887                current_zonelist_order = default_zonelist_order();
2888        else
2889                current_zonelist_order = user_zonelist_order;
2890}
2891
2892static void build_zonelists(pg_data_t *pgdat)
2893{
2894        int j, node, load;
2895        enum zone_type i;
2896        nodemask_t used_mask;
2897        int local_node, prev_node;
2898        struct zonelist *zonelist;
2899        int order = current_zonelist_order;
2900
2901        /* initialize zonelists */
2902        for (i = 0; i < MAX_ZONELISTS; i++) {
2903                zonelist = pgdat->node_zonelists + i;
2904                zonelist->_zonerefs[0].zone = NULL;
2905                zonelist->_zonerefs[0].zone_idx = 0;
2906        }
2907
2908        /* NUMA-aware ordering of nodes */
2909        local_node = pgdat->node_id;
2910        load = nr_online_nodes;
2911        prev_node = local_node;
2912        nodes_clear(used_mask);
2913
2914        memset(node_order, 0, sizeof(node_order));
2915        j = 0;
2916
2917        while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2918                int distance = node_distance(local_node, node);
2919
2920                /*
2921                 * If another node is sufficiently far away then it is better
2922                 * to reclaim pages in a zone before going off node.
2923                 */
2924                if (distance > RECLAIM_DISTANCE)
2925                        zone_reclaim_mode = 1;
2926
2927                /*
2928                 * We don't want to pressure a particular node.
2929                 * So adding penalty to the first node in same
2930                 * distance group to make it round-robin.
2931                 */
2932                if (distance != node_distance(local_node, prev_node))
2933                        node_load[node] = load;
2934
2935                prev_node = node;
2936                load--;
2937                if (order == ZONELIST_ORDER_NODE)
2938                        build_zonelists_in_node_order(pgdat, node);
2939                else
2940                        node_order[j++] = node; /* remember order */
2941        }
2942
2943        if (order == ZONELIST_ORDER_ZONE) {
2944                /* calculate node order -- i.e., DMA last! */
2945                build_zonelists_in_zone_order(pgdat, j);
2946        }
2947
2948        build_thisnode_zonelists(pgdat);
2949}
2950
2951/* Construct the zonelist performance cache - see further mmzone.h */
2952static void build_zonelist_cache(pg_data_t *pgdat)
2953{
2954        struct zonelist *zonelist;
2955        struct zonelist_cache *zlc;
2956        struct zoneref *z;
2957
2958        zonelist = &pgdat->node_zonelists[0];
2959        zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2960        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2961        for (z = zonelist->_zonerefs; z->zone; z++)
2962                zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2963}
2964
2965#ifdef CONFIG_HAVE_MEMORYLESS_NODES
2966/*
2967 * Return node id of node used for "local" allocations.
2968 * I.e., first node id of first zone in arg node's generic zonelist.
2969 * Used for initializing percpu 'numa_mem', which is used primarily
2970 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2971 */
2972int local_memory_node(int node)
2973{
2974        struct zone *zone;
2975
2976        (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
2977                                   gfp_zone(GFP_KERNEL),
2978                                   NULL,
2979                                   &zone);
2980        return zone->node;
2981}
2982#endif
2983
2984#else   /* CONFIG_NUMA */
2985
2986static void set_zonelist_order(void)
2987{
2988        current_zonelist_order = ZONELIST_ORDER_ZONE;
2989}
2990
2991static void build_zonelists(pg_data_t *pgdat)
2992{
2993        int node, local_node;
2994        enum zone_type j;
2995        struct zonelist *zonelist;
2996
2997        local_node = pgdat->node_id;
2998
2999        zonelist = &pgdat->node_zonelists[0];
3000        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3001
3002        /*
3003         * Now we build the zonelist so that it contains the zones
3004         * of all the other nodes.
3005         * We don't want to pressure a particular node, so when
3006         * building the zones for node N, we make sure that the
3007         * zones coming right after the local ones are those from
3008         * node N+1 (modulo N)
3009         */
3010        for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3011                if (!node_online(node))
3012                        continue;
3013                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3014                                                        MAX_NR_ZONES - 1);
3015        }
3016        for (node = 0; node < local_node; node++) {
3017                if (!node_online(node))
3018                        continue;
3019                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3020                                                        MAX_NR_ZONES - 1);
3021        }
3022
3023        zonelist->_zonerefs[j].zone = NULL;
3024        zonelist->_zonerefs[j].zone_idx = 0;
3025}
3026
3027/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3028static void build_zonelist_cache(pg_data_t *pgdat)
3029{
3030        pgdat->node_zonelists[0].zlcache_ptr = NULL;
3031}
3032
3033#endif  /* CONFIG_NUMA */
3034
3035/*
3036 * Boot pageset table. One per cpu which is going to be used for all
3037 * zones and all nodes. The parameters will be set in such a way
3038 * that an item put on a list will immediately be handed over to
3039 * the buddy list. This is safe since pageset manipulation is done
3040 * with interrupts disabled.
3041 *
3042 * The boot_pagesets must be kept even after bootup is complete for
3043 * unused processors and/or zones. They do play a role for bootstrapping
3044 * hotplugged processors.
3045 *
3046 * zoneinfo_show() and maybe other functions do
3047 * not check if the processor is online before following the pageset pointer.
3048 * Other parts of the kernel may not check if the zone is available.
3049 */
3050static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3051static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3052static void setup_zone_pageset(struct zone *zone);
3053
3054/*
3055 * Global mutex to protect against size modification of zonelists
3056 * as well as to serialize pageset setup for the new populated zone.
3057 */
3058DEFINE_MUTEX(zonelists_mutex);
3059
3060/* return values int ....just for stop_machine() */
3061static __init_refok int __build_all_zonelists(void *data)
3062{
3063        int nid;
3064        int cpu;
3065
3066#ifdef CONFIG_NUMA
3067        memset(node_load, 0, sizeof(node_load));
3068#endif
3069        for_each_online_node(nid) {
3070                pg_data_t *pgdat = NODE_DATA(nid);
3071
3072                build_zonelists(pgdat);
3073                build_zonelist_cache(pgdat);
3074        }
3075
3076        /*
3077         * Initialize the boot_pagesets that are going to be used
3078         * for bootstrapping processors. The real pagesets for
3079         * each zone will be allocated later when the per cpu
3080         * allocator is available.
3081         *
3082         * boot_pagesets are used also for bootstrapping offline
3083         * cpus if the system is already booted because the pagesets
3084         * are needed to initialize allocators on a specific cpu too.
3085         * F.e. the percpu allocator needs the page allocator which
3086         * needs the percpu allocator in order to allocate its pagesets
3087         * (a chicken-egg dilemma).
3088         */
3089        for_each_possible_cpu(cpu) {
3090                setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3091
3092#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3093                /*
3094                 * We now know the "local memory node" for each node--
3095                 * i.e., the node of the first zone in the generic zonelist.
3096                 * Set up numa_mem percpu variable for on-line cpus.  During
3097                 * boot, only the boot cpu should be on-line;  we'll init the
3098                 * secondary cpus' numa_mem as they come on-line.  During
3099                 * node/memory hotplug, we'll fixup all on-line cpus.
3100                 */
3101                if (cpu_online(cpu))
3102                        set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3103#endif
3104        }
3105
3106        return 0;
3107}
3108
3109/*
3110 * Called with zonelists_mutex held always
3111 * unless system_state == SYSTEM_BOOTING.
3112 */
3113void build_all_zonelists(void *data)
3114{
3115        set_zonelist_order();
3116
3117        if (system_state == SYSTEM_BOOTING) {
3118                __build_all_zonelists(NULL);
3119                mminit_verify_zonelist();
3120                cpuset_init_current_mems_allowed();
3121        } else {
3122                /* we have to stop all cpus to guarantee there is no user
3123                   of zonelist */
3124#ifdef CONFIG_MEMORY_HOTPLUG
3125                if (data)
3126                        setup_zone_pageset((struct zone *)data);
3127#endif
3128                stop_machine(__build_all_zonelists, NULL, NULL);
3129                /* cpuset refresh routine should be here */
3130        }
3131        vm_total_pages = nr_free_pagecache_pages();
3132        /*
3133         * Disable grouping by mobility if the number of pages in the
3134         * system is too low to allow the mechanism to work. It would be
3135         * more accurate, but expensive to check per-zone. This check is
3136         * made on memory-hotadd so a system can start with mobility
3137         * disabled and enable it later
3138         */
3139        if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3140                page_group_by_mobility_disabled = 1;
3141        else
3142                page_group_by_mobility_disabled = 0;
3143
3144        printk("Built %i zonelists in %s order, mobility grouping %s.  "
3145                "Total pages: %ld\n",
3146                        nr_online_nodes,
3147                        zonelist_order_name[current_zonelist_order],
3148                        page_group_by_mobility_disabled ? "off" : "on",
3149                        vm_total_pages);
3150#ifdef CONFIG_NUMA
3151        printk("Policy zone: %s\n", zone_names[policy_zone]);
3152#endif
3153}
3154
3155/*
3156 * Helper functions to size the waitqueue hash table.
3157 * Essentially these want to choose hash table sizes sufficiently
3158 * large so that collisions trying to wait on pages are rare.
3159 * But in fact, the number of active page waitqueues on typical
3160 * systems is ridiculously low, less than 200. So this is even
3161 * conservative, even though it seems large.
3162 *
3163 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3164 * waitqueues, i.e. the size of the waitq table given the number of pages.
3165 */
3166#define PAGES_PER_WAITQUEUE     256
3167
3168#ifndef CONFIG_MEMORY_HOTPLUG
3169static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3170{
3171        unsigned long size = 1;
3172
3173        pages /= PAGES_PER_WAITQUEUE;
3174
3175        while (size < pages)
3176                size <<= 1;
3177
3178        /*
3179         * Once we have dozens or even hundreds of threads sleeping
3180         * on IO we've got bigger problems than wait queue collision.
3181         * Limit the size of the wait table to a reasonable size.
3182         */
3183        size = min(size, 4096UL);
3184
3185        return max(size, 4UL);
3186}
3187#else
3188/*
3189 * A zone's size might be changed by hot-add, so it is not possible to determine
3190 * a suitable size for its wait_table.  So we use the maximum size now.
3191 *
3192 * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
3193 *
3194 *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
3195 *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3196 *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
3197 *
3198 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3199 * or more by the traditional way. (See above).  It equals:
3200 *
3201 *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
3202 *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
3203 *    powerpc (64K page size)             : =  (32G +16M)byte.
3204 */
3205static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3206{
3207        return 4096UL;
3208}
3209#endif
3210
3211/*
3212 * This is an integer logarithm so that shifts can be used later
3213 * to extract the more random high bits from the multiplicative
3214 * hash function before the remainder is taken.
3215 */
3216static inline unsigned long wait_table_bits(unsigned long size)
3217{
3218        return ffz(~size);
3219}
3220
3221#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3222
3223/*
3224 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3225 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3226 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3227 * higher will lead to a bigger reserve which will get freed as contiguous
3228 * blocks as reclaim kicks in
3229 */
3230static void setup_zone_migrate_reserve(struct zone *zone)
3231{
3232        unsigned long start_pfn, pfn, end_pfn;
3233        struct page *page;
3234        unsigned long block_migratetype;
3235        int reserve;
3236
3237        /* Get the start pfn, end pfn and the number of blocks to reserve */
3238        start_pfn = zone->zone_start_pfn;
3239        end_pfn = start_pfn + zone->spanned_pages;
3240        reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3241                                                        pageblock_order;
3242
3243        /*
3244         * Reserve blocks are generally in place to help high-order atomic
3245         * allocations that are short-lived. A min_free_kbytes value that
3246         * would result in more than 2 reserve blocks for atomic allocations
3247         * is assumed to be in place to help anti-fragmentation for the
3248         * future allocation of hugepages at runtime.
3249         */
3250        reserve = min(2, reserve);
3251
3252        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3253                if (!pfn_valid(pfn))
3254                        continue;
3255                page = pfn_to_page(pfn);
3256
3257                /* Watch out for overlapping nodes */
3258                if (page_to_nid(page) != zone_to_nid(zone))
3259                        continue;
3260
3261                /* Blocks with reserved pages will never free, skip them. */
3262                if (PageReserved(page))
3263                        continue;
3264
3265                block_migratetype = get_pageblock_migratetype(page);
3266
3267                /* If this block is reserved, account for it */
3268                if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3269                        reserve--;
3270                        continue;
3271                }
3272
3273                /* Suitable for reserving if this block is movable */
3274                if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3275                        set_pageblock_migratetype(page, MIGRATE_RESERVE);
3276                        move_freepages_block(zone, page, MIGRATE_RESERVE);
3277                        reserve--;
3278                        continue;
3279                }
3280
3281                /*
3282                 * If the reserve is met and this is a previous reserved block,
3283                 * take it back
3284                 */
3285                if (block_migratetype == MIGRATE_RESERVE) {
3286                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3287                        move_freepages_block(zone, page, MIGRATE_MOVABLE);
3288                }
3289        }
3290}
3291
3292/*
3293 * Initially all pages are reserved - free ones are freed
3294 * up by free_all_bootmem() once the early boot process is
3295 * done. Non-atomic initialization, single-pass.
3296 */
3297void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3298                unsigned long start_pfn, enum memmap_context context)
3299{
3300        struct page *page;
3301        unsigned long end_pfn = start_pfn + size;
3302        unsigned long pfn;
3303        struct zone *z;
3304
3305        if (highest_memmap_pfn < end_pfn - 1)
3306                highest_memmap_pfn = end_pfn - 1;
3307
3308        z = &NODE_DATA(nid)->node_zones[zone];
3309        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3310                /*
3311                 * There can be holes in boot-time mem_map[]s
3312                 * handed to this function.  They do not
3313                 * exist on hotplugged memory.
3314                 */
3315                if (context == MEMMAP_EARLY) {
3316                        if (!early_pfn_valid(pfn))
3317                                continue;
3318                        if (!early_pfn_in_nid(pfn, nid))
3319                                continue;
3320                }
3321                page = pfn_to_page(pfn);
3322                set_page_links(page, zone, nid, pfn);
3323                mminit_verify_page_links(page, zone, nid, pfn);
3324                init_page_count(page);
3325                reset_page_mapcount(page);
3326                SetPageReserved(page);
3327                /*
3328                 * Mark the block movable so that blocks are reserved for
3329                 * movable at startup. This will force kernel allocations
3330                 * to reserve their blocks rather than leaking throughout
3331                 * the address space during boot when many long-lived
3332                 * kernel allocations are made. Later some blocks near
3333                 * the start are marked MIGRATE_RESERVE by
3334                 * setup_zone_migrate_reserve()
3335                 *
3336                 * bitmap is created for zone's valid pfn range. but memmap
3337                 * can be created for invalid pages (for alignment)
3338                 * check here not to call set_pageblock_migratetype() against
3339                 * pfn out of zone.
3340                 */
3341                if ((z->zone_start_pfn <= pfn)
3342                    && (pfn < z->zone_start_pfn + z->spanned_pages)
3343                    && !(pfn & (pageblock_nr_pages - 1)))
3344                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3345
3346                INIT_LIST_HEAD(&page->lru);
3347#ifdef WANT_PAGE_VIRTUAL
3348                /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3349                if (!is_highmem_idx(zone))
3350                        set_page_address(page, __va(pfn << PAGE_SHIFT));
3351#endif
3352        }
3353}
3354
3355static void __meminit zone_init_free_lists(struct zone *zone)
3356{
3357        int order, t;
3358        for_each_migratetype_order(order, t) {
3359                INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3360                zone->free_area[order].nr_free = 0;
3361        }
3362}
3363
3364#ifndef __HAVE_ARCH_MEMMAP_INIT
3365#define memmap_init(size, nid, zone, start_pfn) \
3366        memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3367#endif
3368
3369static int zone_batchsize(struct zone *zone)
3370{
3371#ifdef CONFIG_MMU
3372        int batch;
3373
3374        /*
3375         * The per-cpu-pages pools are set to around 1000th of the
3376         * size of the zone.  But no more than 1/2 of a meg.
3377         *
3378         * OK, so we don't know how big the cache is.  So guess.
3379         */
3380        batch = zone->present_pages / 1024;
3381        if (batch * PAGE_SIZE > 512 * 1024)
3382                batch = (512 * 1024) / PAGE_SIZE;
3383        batch /= 4;             /* We effectively *= 4 below */
3384        if (batch < 1)
3385                batch = 1;
3386
3387        /*
3388         * Clamp the batch to a 2^n - 1 value. Having a power
3389         * of 2 value was found to be more likely to have
3390         * suboptimal cache aliasing properties in some cases.
3391         *
3392         * For example if 2 tasks are alternately allocating
3393         * batches of pages, one task can end up with a lot
3394         * of pages of one half of the possible page colors
3395         * and the other with pages of the other colors.
3396         */
3397        batch = rounddown_pow_of_two(batch + batch/2) - 1;
3398
3399        return batch;
3400
3401#else
3402        /* The deferral and batching of frees should be suppressed under NOMMU
3403         * conditions.
3404         *
3405         * The problem is that NOMMU needs to be able to allocate large chunks
3406         * of contiguous memory as there's no hardware page translation to
3407         * assemble apparent contiguous memory from discontiguous pages.
3408         *
3409         * Queueing large contiguous runs of pages for batching, however,
3410         * causes the pages to actually be freed in smaller chunks.  As there
3411         * can be a significant delay between the individual batches being
3412         * recycled, this leads to the once large chunks of space being
3413         * fragmented and becoming unavailable for high-order allocations.
3414         */
3415        return 0;
3416#endif
3417}
3418
3419static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3420{
3421        struct per_cpu_pages *pcp;
3422        int migratetype;
3423
3424        memset(p, 0, sizeof(*p));
3425
3426        pcp = &p->pcp;
3427        pcp->count = 0;
3428        pcp->high = 6 * batch;
3429        pcp->batch = max(1UL, 1 * batch);
3430        for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3431                INIT_LIST_HEAD(&pcp->lists[migratetype]);
3432}
3433
3434/*
3435 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3436 * to the value high for the pageset p.
3437 */
3438
3439static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3440                                unsigned long high)
3441{
3442        struct per_cpu_pages *pcp;
3443
3444        pcp = &p->pcp;
3445        pcp->high = high;
3446        pcp->batch = max(1UL, high/4);
3447        if ((high/4) > (PAGE_SHIFT * 8))
3448                pcp->batch = PAGE_SHIFT * 8;
3449}
3450
3451static __meminit void setup_zone_pageset(struct zone *zone)
3452{
3453        int cpu;
3454
3455        zone->pageset = alloc_percpu(struct per_cpu_pageset);
3456
3457        for_each_possible_cpu(cpu) {
3458                struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3459
3460                setup_pageset(pcp, zone_batchsize(zone));
3461
3462                if (percpu_pagelist_fraction)
3463                        setup_pagelist_highmark(pcp,
3464                                (zone->present_pages /
3465                                        percpu_pagelist_fraction));
3466        }
3467}
3468
3469/*
3470 * Allocate per cpu pagesets and initialize them.
3471 * Before this call only boot pagesets were available.
3472 */
3473void __init setup_per_cpu_pageset(void)
3474{
3475        struct zone *zone;
3476
3477        for_each_populated_zone(zone)
3478                setup_zone_pageset(zone);
3479}
3480
3481static noinline __init_refok
3482int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3483{
3484        int i;
3485        struct pglist_data *pgdat = zone->zone_pgdat;
3486        size_t alloc_size;
3487
3488        /*
3489         * The per-page waitqueue mechanism uses hashed waitqueues
3490         * per zone.
3491         */
3492        zone->wait_table_hash_nr_entries =
3493                 wait_table_hash_nr_entries(zone_size_pages);
3494        zone->wait_table_bits =
3495                wait_table_bits(zone->wait_table_hash_nr_entries);
3496        alloc_size = zone->wait_table_hash_nr_entries
3497                                        * sizeof(wait_queue_head_t);
3498
3499        if (!slab_is_available()) {
3500                zone->wait_table = (wait_queue_head_t *)
3501                        alloc_bootmem_node_nopanic(pgdat, alloc_size);
3502        } else {
3503                /*
3504                 * This case means that a zone whose size was 0 gets new memory
3505                 * via memory hot-add.
3506                 * But it may be the case that a new node was hot-added.  In
3507                 * this case vmalloc() will not be able to use this new node's
3508                 * memory - this wait_table must be initialized to use this new
3509                 * node itself as well.
3510                 * To use this new node's memory, further consideration will be
3511                 * necessary.
3512                 */
3513                zone->wait_table = vmalloc(alloc_size);
3514        }
3515        if (!zone->wait_table)
3516                return -ENOMEM;
3517
3518        for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3519                init_waitqueue_head(zone->wait_table + i);
3520
3521        return 0;
3522}
3523
3524static int __zone_pcp_update(void *data)
3525{
3526        struct zone *zone = data;
3527        int cpu;
3528        unsigned long batch = zone_batchsize(zone), flags;
3529
3530        for_each_possible_cpu(cpu) {
3531                struct per_cpu_pageset *pset;
3532                struct per_cpu_pages *pcp;
3533
3534                pset = per_cpu_ptr(zone->pageset, cpu);
3535                pcp = &pset->pcp;
3536
3537                local_irq_save(flags);
3538                free_pcppages_bulk(zone, pcp->count, pcp);
3539                setup_pageset(pset, batch);
3540                local_irq_restore(flags);
3541        }
3542        return 0;
3543}
3544
3545void zone_pcp_update(struct zone *zone)
3546{
3547        stop_machine(__zone_pcp_update, zone, NULL);
3548}
3549
3550static __meminit void zone_pcp_init(struct zone *zone)
3551{
3552        /*
3553         * per cpu subsystem is not up at this point. The following code
3554         * relies on the ability of the linker to provide the
3555         * offset of a (static) per cpu variable into the per cpu area.
3556         */
3557        zone->pageset = &boot_pageset;
3558
3559        if (zone->present_pages)
3560                printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
3561                        zone->name, zone->present_pages,
3562                                         zone_batchsize(zone));
3563}
3564
3565__meminit int init_currently_empty_zone(struct zone *zone,
3566                                        unsigned long zone_start_pfn,
3567                                        unsigned long size,
3568                                        enum memmap_context context)
3569{
3570        struct pglist_data *pgdat = zone->zone_pgdat;
3571        int ret;
3572        ret = zone_wait_table_init(zone, size);
3573        if (ret)
3574                return ret;
3575        pgdat->nr_zones = zone_idx(zone) + 1;
3576
3577        zone->zone_start_pfn = zone_start_pfn;
3578
3579        mminit_dprintk(MMINIT_TRACE, "memmap_init",
3580                        "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3581                        pgdat->node_id,
3582                        (unsigned long)zone_idx(zone),
3583                        zone_start_pfn, (zone_start_pfn + size));
3584
3585        zone_init_free_lists(zone);
3586
3587        return 0;
3588}
3589
3590#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3591/*
3592 * Basic iterator support. Return the first range of PFNs for a node
3593 * Note: nid == MAX_NUMNODES returns first region regardless of node
3594 */
3595static int __meminit first_active_region_index_in_nid(int nid)
3596{
3597        int i;
3598
3599        for (i = 0; i < nr_nodemap_entries; i++)
3600                if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3601                        return i;
3602
3603        return -1;
3604}
3605
3606/*
3607 * Basic iterator support. Return the next active range of PFNs for a node
3608 * Note: nid == MAX_NUMNODES returns next region regardless of node
3609 */
3610static int __meminit next_active_region_index_in_nid(int index, int nid)
3611{
3612        for (index = index + 1; index < nr_nodemap_entries; index++)
3613                if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3614                        return index;
3615
3616        return -1;
3617}
3618
3619#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3620/*
3621 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3622 * Architectures may implement their own version but if add_active_range()
3623 * was used and there are no special requirements, this is a convenient
3624 * alternative
3625 */
3626int __meminit __early_pfn_to_nid(unsigned long pfn)
3627{
3628        int i;
3629
3630        for (i = 0; i < nr_nodemap_entries; i++) {
3631                unsigned long start_pfn = early_node_map[i].start_pfn;
3632                unsigned long end_pfn = early_node_map[i].end_pfn;
3633
3634                if (start_pfn <= pfn && pfn < end_pfn)
3635                        return early_node_map[i].nid;
3636        }
3637        /* This is a memory hole */
3638        return -1;
3639}
3640#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3641
3642int __meminit early_pfn_to_nid(unsigned long pfn)
3643{
3644        int nid;
3645
3646        nid = __early_pfn_to_nid(pfn);
3647        if (nid >= 0)
3648                return nid;
3649        /* just returns 0 */
3650        return 0;
3651}
3652
3653#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3654bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3655{
3656        int nid;
3657
3658        nid = __early_pfn_to_nid(pfn);
3659        if (nid >= 0 && nid != node)
3660                return false;
3661        return true;
3662}
3663#endif
3664
3665/* Basic iterator support to walk early_node_map[] */
3666#define for_each_active_range_index_in_nid(i, nid) \
3667        for (i = first_active_region_index_in_nid(nid); i != -1; \
3668                                i = next_active_region_index_in_nid(i, nid))
3669
3670/**
3671 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3672 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3673 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3674 *
3675 * If an architecture guarantees that all ranges registered with
3676 * add_active_ranges() contain no holes and may be freed, this
3677 * this function may be used instead of calling free_bootmem() manually.
3678 */
3679void __init free_bootmem_with_active_regions(int nid,
3680                                                unsigned long max_low_pfn)
3681{
3682        int i;
3683
3684        for_each_active_range_index_in_nid(i, nid) {
3685                unsigned long size_pages = 0;
3686                unsigned long end_pfn = early_node_map[i].end_pfn;
3687
3688                if (early_node_map[i].start_pfn >= max_low_pfn)
3689                        continue;
3690
3691                if (end_pfn > max_low_pfn)
3692                        end_pfn = max_low_pfn;
3693
3694                size_pages = end_pfn - early_node_map[i].start_pfn;
3695                free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3696                                PFN_PHYS(early_node_map[i].start_pfn),
3697                                size_pages << PAGE_SHIFT);
3698        }
3699}
3700
3701#ifdef CONFIG_HAVE_MEMBLOCK
3702u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3703                                        u64 goal, u64 limit)
3704{
3705        int i;
3706
3707        /* Need to go over early_node_map to find out good range for node */
3708        for_each_active_range_index_in_nid(i, nid) {
3709                u64 addr;
3710                u64 ei_start, ei_last;
3711                u64 final_start, final_end;
3712
3713                ei_last = early_node_map[i].end_pfn;
3714                ei_last <<= PAGE_SHIFT;
3715                ei_start = early_node_map[i].start_pfn;
3716                ei_start <<= PAGE_SHIFT;
3717
3718                final_start = max(ei_start, goal);
3719                final_end = min(ei_last, limit);
3720
3721                if (final_start >= final_end)
3722                        continue;
3723
3724                addr = memblock_find_in_range(final_start, final_end, size, align);
3725
3726                if (addr == MEMBLOCK_ERROR)
3727                        continue;
3728
3729                return addr;
3730        }
3731
3732        return MEMBLOCK_ERROR;
3733}
3734#endif
3735
3736int __init add_from_early_node_map(struct range *range, int az,
3737                                   int nr_range, int nid)
3738{
3739        int i;
3740        u64 start, end;
3741
3742        /* need to go over early_node_map to find out good range for node */
3743        for_each_active_range_index_in_nid(i, nid) {
3744                start = early_node_map[i].start_pfn;
3745                end = early_node_map[i].end_pfn;
3746                nr_range = add_range(range, az, nr_range, start, end);
3747        }
3748        return nr_range;
3749}
3750
3751#ifdef CONFIG_NO_BOOTMEM
3752void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3753                                        u64 goal, u64 limit)
3754{
3755        void *ptr;
3756        u64 addr;
3757
3758        if (limit > memblock.current_limit)
3759                limit = memblock.current_limit;
3760
3761        addr = find_memory_core_early(nid, size, align, goal, limit);
3762
3763        if (addr == MEMBLOCK_ERROR)
3764                return NULL;
3765
3766        ptr = phys_to_virt(addr);
3767        memset(ptr, 0, size);
3768        memblock_x86_reserve_range(addr, addr + size, "BOOTMEM");
3769        /*
3770         * The min_count is set to 0 so that bootmem allocated blocks
3771         * are never reported as leaks.
3772         */
3773        kmemleak_alloc(ptr, size, 0, 0);
3774        return ptr;
3775}
3776#endif
3777
3778
3779void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3780{
3781        int i;
3782        int ret;
3783
3784        for_each_active_range_index_in_nid(i, nid) {
3785                ret = work_fn(early_node_map[i].start_pfn,
3786                              early_node_map[i].end_pfn, data);
3787                if (ret)
3788                        break;
3789        }
3790}
3791/**
3792 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3793 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3794 *
3795 * If an architecture guarantees that all ranges registered with
3796 * add_active_ranges() contain no holes and may be freed, this
3797 * function may be used instead of calling memory_present() manually.
3798 */
3799void __init sparse_memory_present_with_active_regions(int nid)
3800{
3801        int i;
3802
3803        for_each_active_range_index_in_nid(i, nid)
3804                memory_present(early_node_map[i].nid,
3805                                early_node_map[i].start_pfn,
3806                                early_node_map[i].end_pfn);
3807}
3808
3809/**
3810 * get_pfn_range_for_nid - Return the start and end page frames for a node
3811 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3812 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3813 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3814 *
3815 * It returns the start and end page frame of a node based on information
3816 * provided by an arch calling add_active_range(). If called for a node
3817 * with no available memory, a warning is printed and the start and end
3818 * PFNs will be 0.
3819 */
3820void __meminit get_pfn_range_for_nid(unsigned int nid,
3821                        unsigned long *start_pfn, unsigned long *end_pfn)
3822{
3823        int i;
3824        *start_pfn = -1UL;
3825        *end_pfn = 0;
3826
3827        for_each_active_range_index_in_nid(i, nid) {
3828                *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3829                *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3830        }
3831
3832        if (*start_pfn == -1UL)
3833                *start_pfn = 0;
3834}
3835
3836/*
3837 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3838 * assumption is made that zones within a node are ordered in monotonic
3839 * increasing memory addresses so that the "highest" populated zone is used
3840 */
3841static void __init find_usable_zone_for_movable(void)
3842{
3843        int zone_index;
3844        for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3845                if (zone_index == ZONE_MOVABLE)
3846                        continue;
3847
3848                if (arch_zone_highest_possible_pfn[zone_index] >
3849                                arch_zone_lowest_possible_pfn[zone_index])
3850                        break;
3851        }
3852
3853        VM_BUG_ON(zone_index == -1);
3854        movable_zone = zone_index;
3855}
3856
3857/*
3858 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3859 * because it is sized independant of architecture. Unlike the other zones,
3860 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3861 * in each node depending on the size of each node and how evenly kernelcore
3862 * is distributed. This helper function adjusts the zone ranges
3863 * provided by the architecture for a given node by using the end of the
3864 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3865 * zones within a node are in order of monotonic increases memory addresses
3866 */
3867static void __meminit adjust_zone_range_for_zone_movable(int nid,
3868                                        unsigned long zone_type,
3869                                        unsigned long node_start_pfn,
3870                                        unsigned long node_end_pfn,
3871                                        unsigned long *zone_start_pfn,
3872                                        unsigned long *zone_end_pfn)
3873{
3874        /* Only adjust if ZONE_MOVABLE is on this node */
3875        if (zone_movable_pfn[nid]) {
3876                /* Size ZONE_MOVABLE */
3877                if (zone_type == ZONE_MOVABLE) {
3878                        *zone_start_pfn = zone_movable_pfn[nid];
3879                        *zone_end_pfn = min(node_end_pfn,
3880                                arch_zone_highest_possible_pfn[movable_zone]);
3881
3882                /* Adjust for ZONE_MOVABLE starting within this range */
3883                } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3884                                *zone_end_pfn > zone_movable_pfn[nid]) {
3885                        *zone_end_pfn = zone_movable_pfn[nid];
3886
3887                /* Check if this whole range is within ZONE_MOVABLE */
3888                } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3889                        *zone_start_pfn = *zone_end_pfn;
3890        }
3891}
3892
3893/*
3894 * Return the number of pages a zone spans in a node, including holes
3895 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3896 */
3897static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3898                                        unsigned long zone_type,
3899                                        unsigned long *ignored)
3900{
3901        unsigned long node_start_pfn, node_end_pfn;
3902        unsigned long zone_start_pfn, zone_end_pfn;
3903
3904        /* Get the start and end of the node and zone */
3905        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3906        zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3907        zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3908        adjust_zone_range_for_zone_movable(nid, zone_type,
3909                                node_start_pfn, node_end_pfn,
3910                                &zone_start_pfn, &zone_end_pfn);
3911
3912        /* Check that this node has pages within the zone's required range */
3913        if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3914                return 0;
3915
3916        /* Move the zone boundaries inside the node if necessary */
3917        zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3918        zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3919
3920        /* Return the spanned pages */
3921        return zone_end_pfn - zone_start_pfn;
3922}
3923
3924/*
3925 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3926 * then all holes in the requested range will be accounted for.
3927 */
3928unsigned long __meminit __absent_pages_in_range(int nid,
3929                                unsigned long range_start_pfn,
3930                                unsigned long range_end_pfn)
3931{
3932        int i = 0;
3933        unsigned long prev_end_pfn = 0, hole_pages = 0;
3934        unsigned long start_pfn;
3935
3936        /* Find the end_pfn of the first active range of pfns in the node */
3937        i = first_active_region_index_in_nid(nid);
3938        if (i == -1)
3939                return 0;
3940
3941        prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3942
3943        /* Account for ranges before physical memory on this node */
3944        if (early_node_map[i].start_pfn > range_start_pfn)
3945                hole_pages = prev_end_pfn - range_start_pfn;
3946
3947        /* Find all holes for the zone within the node */
3948        for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3949
3950                /* No need to continue if prev_end_pfn is outside the zone */
3951                if (prev_end_pfn >= range_end_pfn)
3952                        break;
3953
3954                /* Make sure the end of the zone is not within the hole */
3955                start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3956                prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3957
3958                /* Update the hole size cound and move on */
3959                if (start_pfn > range_start_pfn) {
3960                        BUG_ON(prev_end_pfn > start_pfn);
3961                        hole_pages += start_pfn - prev_end_pfn;
3962                }
3963                prev_end_pfn = early_node_map[i].end_pfn;
3964        }
3965
3966        /* Account for ranges past physical memory on this node */
3967        if (range_end_pfn > prev_end_pfn)
3968                hole_pages += range_end_pfn -
3969                                max(range_start_pfn, prev_end_pfn);
3970
3971        return hole_pages;
3972}
3973
3974/**
3975 * absent_pages_in_range - Return number of page frames in holes within a range
3976 * @start_pfn: The start PFN to start searching for holes
3977 * @end_pfn: The end PFN to stop searching for holes
3978 *
3979 * It returns the number of pages frames in memory holes within a range.
3980 */
3981unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3982                                                        unsigned long end_pfn)
3983{
3984        return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3985}
3986
3987/* Return the number of page frames in holes in a zone on a node */
3988static unsigned long __meminit zone_absent_pages_in_node(int nid,
3989                                        unsigned long zone_type,
3990                                        unsigned long *ignored)
3991{
3992        unsigned long node_start_pfn, node_end_pfn;
3993        unsigned long zone_start_pfn, zone_end_pfn;
3994
3995        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3996        zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3997                                                        node_start_pfn);
3998        zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3999                                                        node_end_pfn);
4000
4001        adjust_zone_range_for_zone_movable(nid, zone_type,
4002                        node_start_pfn, node_end_pfn,
4003                        &zone_start_pfn, &zone_end_pfn);
4004        return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4005}
4006
4007#else
4008static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4009                                        unsigned long zone_type,
4010                                        unsigned long *zones_size)
4011{
4012        return zones_size[zone_type];
4013}
4014
4015static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4016                                                unsigned long zone_type,
4017                                                unsigned long *zholes_size)
4018{
4019        if (!zholes_size)
4020                return 0;
4021
4022        return zholes_size[zone_type];
4023}
4024
4025#endif
4026
4027static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4028                unsigned long *zones_size, unsigned long *zholes_size)
4029{
4030        unsigned long realtotalpages, totalpages = 0;
4031        enum zone_type i;
4032
4033        for (i = 0; i < MAX_NR_ZONES; i++)
4034                totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4035                                                                zones_size);
4036        pgdat->node_spanned_pages = totalpages;
4037
4038        realtotalpages = totalpages;
4039        for (i = 0; i < MAX_NR_ZONES; i++)
4040                realtotalpages -=
4041                        zone_absent_pages_in_node(pgdat->node_id, i,
4042                                                                zholes_size);
4043        pgdat->node_present_pages = realtotalpages;
4044        printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4045                                                        realtotalpages);
4046}
4047
4048#ifndef CONFIG_SPARSEMEM
4049/*
4050 * Calculate the size of the zone->blockflags rounded to an unsigned long
4051 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4052 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4053 * round what is now in bits to nearest long in bits, then return it in
4054 * bytes.
4055 */
4056static unsigned long __init usemap_size(unsigned long zonesize)
4057{
4058        unsigned long usemapsize;
4059
4060        usemapsize = roundup(zonesize, pageblock_nr_pages);
4061        usemapsize = usemapsize >> pageblock_order;
4062        usemapsize *= NR_PAGEBLOCK_BITS;
4063        usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4064
4065        return usemapsize / 8;
4066}
4067
4068static void __init setup_usemap(struct pglist_data *pgdat,
4069                                struct zone *zone, unsigned long zonesize)
4070{
4071        unsigned long usemapsize = usemap_size(zonesize);
4072        zone->pageblock_flags = NULL;
4073        if (usemapsize)
4074                zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4075                                                                   usemapsize);
4076}
4077#else
4078static inline void setup_usemap(struct pglist_data *pgdat,
4079                                struct zone *zone, unsigned long zonesize) {}
4080#endif /* CONFIG_SPARSEMEM */
4081
4082#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4083
4084/* Return a sensible default order for the pageblock size. */
4085static inline int pageblock_default_order(void)
4086{
4087        if (HPAGE_SHIFT > PAGE_SHIFT)
4088                return HUGETLB_PAGE_ORDER;
4089
4090        return MAX_ORDER-1;
4091}
4092
4093/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4094static inline void __init set_pageblock_order(unsigned int order)
4095{
4096        /* Check that pageblock_nr_pages has not already been setup */
4097        if (pageblock_order)
4098                return;
4099
4100        /*
4101         * Assume the largest contiguous order of interest is a huge page.
4102         * This value may be variable depending on boot parameters on IA64
4103         */
4104        pageblock_order = order;
4105}
4106#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4107
4108/*
4109 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4110 * and pageblock_default_order() are unused as pageblock_order is set
4111 * at compile-time. See include/linux/pageblock-flags.h for the values of
4112 * pageblock_order based on the kernel config
4113 */
4114static inline int pageblock_default_order(unsigned int order)
4115{
4116        return MAX_ORDER-1;
4117}
4118#define set_pageblock_order(x)  do {} while (0)
4119
4120#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4121
4122/*
4123 * Set up the zone data structures:
4124 *   - mark all pages reserved
4125 *   - mark all memory queues empty
4126 *   - clear the memory bitmaps
4127 */
4128static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4129                unsigned long *zones_size, unsigned long *zholes_size)
4130{
4131        enum zone_type j;
4132        int nid = pgdat->node_id;
4133        unsigned long zone_start_pfn = pgdat->node_start_pfn;
4134        int ret;
4135
4136        pgdat_resize_init(pgdat);
4137        pgdat->nr_zones = 0;
4138        init_waitqueue_head(&pgdat->kswapd_wait);
4139        pgdat->kswapd_max_order = 0;
4140        pgdat_page_cgroup_init(pgdat);
4141        
4142        for (j = 0; j < MAX_NR_ZONES; j++) {
4143                struct zone *zone = pgdat->node_zones + j;
4144                unsigned long size, realsize, memmap_pages;
4145                enum lru_list l;
4146
4147                size = zone_spanned_pages_in_node(nid, j, zones_size);
4148                realsize = size - zone_absent_pages_in_node(nid, j,
4149                                                                zholes_size);
4150
4151                /*
4152                 * Adjust realsize so that it accounts for how much memory
4153                 * is used by this zone for memmap. This affects the watermark
4154                 * and per-cpu initialisations
4155                 */
4156                memmap_pages =
4157                        PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4158                if (realsize >= memmap_pages) {
4159                        realsize -= memmap_pages;
4160                        if (memmap_pages)
4161                                printk(KERN_DEBUG
4162                                       "  %s zone: %lu pages used for memmap\n",
4163                                       zone_names[j], memmap_pages);
4164                } else
4165                        printk(KERN_WARNING
4166                                "  %s zone: %lu pages exceeds realsize %lu\n",
4167                                zone_names[j], memmap_pages, realsize);
4168
4169                /* Account for reserved pages */
4170                if (j == 0 && realsize > dma_reserve) {
4171                        realsize -= dma_reserve;
4172                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
4173                                        zone_names[0], dma_reserve);
4174                }
4175
4176                if (!is_highmem_idx(j))
4177                        nr_kernel_pages += realsize;
4178                nr_all_pages += realsize;
4179
4180                zone->spanned_pages = size;
4181                zone->present_pages = realsize;
4182#ifdef CONFIG_NUMA
4183                zone->node = nid;
4184                zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4185                                                / 100;
4186                zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4187#endif
4188                zone->name = zone_names[j];
4189                spin_lock_init(&zone->lock);
4190                spin_lock_init(&zone->lru_lock);
4191                zone_seqlock_init(zone);
4192                zone->zone_pgdat = pgdat;
4193
4194                zone_pcp_init(zone);
4195                for_each_lru(l) {
4196                        INIT_LIST_HEAD(&zone->lru[l].list);
4197                        zone->reclaim_stat.nr_saved_scan[l] = 0;
4198                }
4199                zone->reclaim_stat.recent_rotated[0] = 0;
4200                zone->reclaim_stat.recent_rotated[1] = 0;
4201                zone->reclaim_stat.recent_scanned[0] = 0;
4202                zone->reclaim_stat.recent_scanned[1] = 0;
4203                zap_zone_vm_stats(zone);
4204                zone->flags = 0;
4205                if (!size)
4206                        continue;
4207
4208                set_pageblock_order(pageblock_default_order());
4209                setup_usemap(pgdat, zone, size);
4210                ret = init_currently_empty_zone(zone, zone_start_pfn,
4211                                                size, MEMMAP_EARLY);
4212                BUG_ON(ret);
4213                memmap_init(size, nid, j, zone_start_pfn);
4214                zone_start_pfn += size;
4215        }
4216}
4217
4218static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4219{
4220        /* Skip empty nodes */
4221        if (!pgdat->node_spanned_pages)
4222                return;
4223
4224#ifdef CONFIG_FLAT_NODE_MEM_MAP
4225        /* ia64 gets its own node_mem_map, before this, without bootmem */
4226        if (!pgdat->node_mem_map) {
4227                unsigned long size, start, end;
4228                struct page *map;
4229
4230                /*
4231                 * The zone's endpoints aren't required to be MAX_ORDER
4232                 * aligned but the node_mem_map endpoints must be in order
4233                 * for the buddy allocator to function correctly.
4234                 */
4235                start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4236                end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4237                end = ALIGN(end, MAX_ORDER_NR_PAGES);
4238                size =  (end - start) * sizeof(struct page);
4239                map = alloc_remap(pgdat->node_id, size);
4240                if (!map)
4241                        map = alloc_bootmem_node_nopanic(pgdat, size);
4242                pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4243        }
4244#ifndef CONFIG_NEED_MULTIPLE_NODES
4245        /*
4246         * With no DISCONTIG, the global mem_map is just set as node 0's
4247         */
4248        if (pgdat == NODE_DATA(0)) {
4249                mem_map = NODE_DATA(0)->node_mem_map;
4250#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4251                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4252                        mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4253#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4254        }
4255#endif
4256#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4257}
4258
4259void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4260                unsigned long node_start_pfn, unsigned long *zholes_size)
4261{
4262        pg_data_t *pgdat = NODE_DATA(nid);
4263
4264        pgdat->node_id = nid;
4265        pgdat->node_start_pfn = node_start_pfn;
4266        calculate_node_totalpages(pgdat, zones_size, zholes_size);
4267
4268        alloc_node_mem_map(pgdat);
4269#ifdef CONFIG_FLAT_NODE_MEM_MAP
4270        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4271                nid, (unsigned long)pgdat,
4272                (unsigned long)pgdat->node_mem_map);
4273#endif
4274
4275        free_area_init_core(pgdat, zones_size, zholes_size);
4276}
4277
4278#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4279
4280#if MAX_NUMNODES > 1
4281/*
4282 * Figure out the number of possible node ids.
4283 */
4284static void __init setup_nr_node_ids(void)
4285{
4286        unsigned int node;
4287        unsigned int highest = 0;
4288
4289        for_each_node_mask(node, node_possible_map)
4290                highest = node;
4291        nr_node_ids = highest + 1;
4292}
4293#else
4294static inline void setup_nr_node_ids(void)
4295{
4296}
4297#endif
4298
4299/**
4300 * add_active_range - Register a range of PFNs backed by physical memory
4301 * @nid: The node ID the range resides on
4302 * @start_pfn: The start PFN of the available physical memory
4303 * @end_pfn: The end PFN of the available physical memory
4304 *
4305 * These ranges are stored in an early_node_map[] and later used by
4306 * free_area_init_nodes() to calculate zone sizes and holes. If the
4307 * range spans a memory hole, it is up to the architecture to ensure
4308 * the memory is not freed by the bootmem allocator. If possible
4309 * the range being registered will be merged with existing ranges.
4310 */
4311void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4312                                                unsigned long end_pfn)
4313{
4314        int i;
4315
4316        mminit_dprintk(MMINIT_TRACE, "memory_register",
4317                        "Entering add_active_range(%d, %#lx, %#lx) "
4318                        "%d entries of %d used\n",
4319                        nid, start_pfn, end_pfn,
4320                        nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4321
4322        mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4323
4324        /* Merge with existing active regions if possible */
4325        for (i = 0; i < nr_nodemap_entries; i++) {
4326                if (early_node_map[i].nid != nid)
4327                        continue;
4328
4329                /* Skip if an existing region covers this new one */
4330                if (start_pfn >= early_node_map[i].start_pfn &&
4331                                end_pfn <= early_node_map[i].end_pfn)
4332                        return;
4333
4334                /* Merge forward if suitable */
4335                if (start_pfn <= early_node_map[i].end_pfn &&
4336                                end_pfn > early_node_map[i].end_pfn) {
4337                        early_node_map[i].end_pfn = end_pfn;
4338                        return;
4339                }
4340
4341                /* Merge backward if suitable */
4342                if (start_pfn < early_node_map[i].start_pfn &&
4343                                end_pfn >= early_node_map[i].start_pfn) {
4344                        early_node_map[i].start_pfn = start_pfn;
4345                        return;
4346                }
4347        }
4348
4349        /* Check that early_node_map is large enough */
4350        if (i >= MAX_ACTIVE_REGIONS) {
4351                printk(KERN_CRIT "More than %d memory regions, truncating\n",
4352                                                        MAX_ACTIVE_REGIONS);
4353                return;
4354        }
4355
4356        early_node_map[i].nid = nid;
4357        early_node_map[i].start_pfn = start_pfn;
4358        early_node_map[i].end_pfn = end_pfn;
4359        nr_nodemap_entries = i + 1;
4360}
4361
4362/**
4363 * remove_active_range - Shrink an existing registered range of PFNs
4364 * @nid: The node id the range is on that should be shrunk
4365 * @start_pfn: The new PFN of the range
4366 * @end_pfn: The new PFN of the range
4367 *
4368 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4369 * The map is kept near the end physical page range that has already been
4370 * registered. This function allows an arch to shrink an existing registered
4371 * range.
4372 */
4373void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4374                                unsigned long end_pfn)
4375{
4376        int i, j;
4377        int removed = 0;
4378
4379        printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4380                          nid, start_pfn, end_pfn);
4381
4382        /* Find the old active region end and shrink */
4383        for_each_active_range_index_in_nid(i, nid) {
4384                if (early_node_map[i].start_pfn >= start_pfn &&
4385                    early_node_map[i].end_pfn <= end_pfn) {
4386                        /* clear it */
4387                        early_node_map[i].start_pfn = 0;
4388                        early_node_map[i].end_pfn = 0;
4389                        removed = 1;
4390                        continue;
4391                }
4392                if (early_node_map[i].start_pfn < start_pfn &&
4393                    early_node_map[i].end_pfn > start_pfn) {
4394                        unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4395                        early_node_map[i].end_pfn = start_pfn;
4396                        if (temp_end_pfn > end_pfn)
4397                                add_active_range(nid, end_pfn, temp_end_pfn);
4398                        continue;
4399                }
4400                if (early_node_map[i].start_pfn >= start_pfn &&
4401                    early_node_map[i].end_pfn > end_pfn &&
4402                    early_node_map[i].start_pfn < end_pfn) {
4403                        early_node_map[i].start_pfn = end_pfn;
4404                        continue;
4405                }
4406        }
4407
4408        if (!removed)
4409                return;
4410
4411        /* remove the blank ones */
4412        for (i = nr_nodemap_entries - 1; i > 0; i--) {
4413                if (early_node_map[i].nid != nid)
4414                        continue;
4415                if (early_node_map[i].end_pfn)
4416                        continue;
4417                /* we found it, get rid of it */
4418                for (j = i; j < nr_nodemap_entries - 1; j++)
4419                        memcpy(&early_node_map[j], &early_node_map[j+1],
4420                                sizeof(early_node_map[j]));
4421                j = nr_nodemap_entries - 1;
4422                memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4423                nr_nodemap_entries--;
4424        }
4425}
4426
4427/**
4428 * remove_all_active_ranges - Remove all currently registered regions
4429 *
4430 * During discovery, it may be found that a table like SRAT is invalid
4431 * and an alternative discovery method must be used. This function removes
4432 * all currently registered regions.
4433 */
4434void __init remove_all_active_ranges(void)
4435{
4436        memset(early_node_map, 0, sizeof(early_node_map));
4437        nr_nodemap_entries = 0;
4438}
4439
4440/* Compare two active node_active_regions */
4441static int __init cmp_node_active_region(const void *a, const void *b)
4442{
4443        struct node_active_region *arange = (struct node_active_region *)a;
4444        struct node_active_region *brange = (struct node_active_region *)b;
4445
4446        /* Done this way to avoid overflows */
4447        if (arange->start_pfn > brange->start_pfn)
4448                return 1;
4449        if (arange->start_pfn < brange->start_pfn)
4450                return -1;
4451
4452        return 0;
4453}
4454
4455/* sort the node_map by start_pfn */
4456void __init sort_node_map(void)
4457{
4458        sort(early_node_map, (size_t)nr_nodemap_entries,
4459                        sizeof(struct node_active_region),
4460                        cmp_node_active_region, NULL);
4461}
4462
4463/* Find the lowest pfn for a node */
4464static unsigned long __init find_min_pfn_for_node(int nid)
4465{
4466        int i;
4467        unsigned long min_pfn = ULONG_MAX;
4468
4469        /* Assuming a sorted map, the first range found has the starting pfn */
4470        for_each_active_range_index_in_nid(i, nid)
4471                min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4472
4473        if (min_pfn == ULONG_MAX) {
4474                printk(KERN_WARNING
4475                        "Could not find start_pfn for node %d\n", nid);
4476                return 0;
4477        }
4478
4479        return min_pfn;
4480}
4481
4482/**
4483 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4484 *
4485 * It returns the minimum PFN based on information provided via
4486 * add_active_range().
4487 */
4488unsigned long __init find_min_pfn_with_active_regions(void)
4489{
4490        return find_min_pfn_for_node(MAX_NUMNODES);
4491}
4492
4493/*
4494 * early_calculate_totalpages()
4495 * Sum pages in active regions for movable zone.
4496 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4497 */
4498static unsigned long __init early_calculate_totalpages(void)
4499{
4500        int i;
4501        unsigned long totalpages = 0;
4502
4503        for (i = 0; i < nr_nodemap_entries; i++) {
4504                unsigned long pages = early_node_map[i].end_pfn -
4505                                                early_node_map[i].start_pfn;
4506                totalpages += pages;
4507                if (pages)
4508                        node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4509        }
4510        return totalpages;
4511}
4512
4513/*
4514 * Find the PFN the Movable zone begins in each node. Kernel memory
4515 * is spread evenly between nodes as long as the nodes have enough
4516 * memory. When they don't, some nodes will have more kernelcore than
4517 * others
4518 */
4519static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4520{
4521        int i, nid;
4522        unsigned long usable_startpfn;
4523        unsigned long kernelcore_node, kernelcore_remaining;
4524        /* save the state before borrow the nodemask */
4525        nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4526        unsigned long totalpages = early_calculate_totalpages();
4527        int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4528
4529        /*
4530         * If movablecore was specified, calculate what size of
4531         * kernelcore that corresponds so that memory usable for
4532         * any allocation type is evenly spread. If both kernelcore
4533         * and movablecore are specified, then the value of kernelcore
4534         * will be used for required_kernelcore if it's greater than
4535         * what movablecore would have allowed.
4536         */
4537        if (required_movablecore) {
4538                unsigned long corepages;
4539
4540                /*
4541                 * Round-up so that ZONE_MOVABLE is at least as large as what
4542                 * was requested by the user
4543                 */
4544                required_movablecore =
4545                        roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4546                corepages = totalpages - required_movablecore;
4547
4548                required_kernelcore = max(required_kernelcore, corepages);
4549        }
4550
4551        /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4552        if (!required_kernelcore)
4553                goto out;
4554
4555        /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4556        find_usable_zone_for_movable();
4557        usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4558
4559restart:
4560        /* Spread kernelcore memory as evenly as possible throughout nodes */
4561        kernelcore_node = required_kernelcore / usable_nodes;
4562        for_each_node_state(nid, N_HIGH_MEMORY) {
4563                /*
4564                 * Recalculate kernelcore_node if the division per node
4565                 * now exceeds what is necessary to satisfy the requested
4566                 * amount of memory for the kernel
4567                 */
4568                if (required_kernelcore < kernelcore_node)
4569                        kernelcore_node = required_kernelcore / usable_nodes;
4570
4571                /*
4572                 * As the map is walked, we track how much memory is usable
4573                 * by the kernel using kernelcore_remaining. When it is
4574                 * 0, the rest of the node is usable by ZONE_MOVABLE
4575                 */
4576                kernelcore_remaining = kernelcore_node;
4577
4578                /* Go through each range of PFNs within this node */
4579                for_each_active_range_index_in_nid(i, nid) {
4580                        unsigned long start_pfn, end_pfn;
4581                        unsigned long size_pages;
4582
4583                        start_pfn = max(early_node_map[i].start_pfn,
4584                                                zone_movable_pfn[nid]);
4585                        end_pfn = early_node_map[i].end_pfn;
4586                        if (start_pfn >= end_pfn)
4587                                continue;
4588
4589                        /* Account for what is only usable for kernelcore */
4590                        if (start_pfn < usable_startpfn) {
4591                                unsigned long kernel_pages;
4592                                kernel_pages = min(end_pfn, usable_startpfn)
4593                                                                - start_pfn;
4594
4595                                kernelcore_remaining -= min(kernel_pages,
4596                                                        kernelcore_remaining);
4597                                required_kernelcore -= min(kernel_pages,
4598                                                        required_kernelcore);
4599
4600                                /* Continue if range is now fully accounted */
4601                                if (end_pfn <= usable_startpfn) {
4602
4603                                        /*
4604                                         * Push zone_movable_pfn to the end so
4605                                         * that if we have to rebalance
4606                                         * kernelcore across nodes, we will
4607                                         * not double account here
4608                                         */
4609                                        zone_movable_pfn[nid] = end_pfn;
4610                                        continue;
4611                                }
4612                                start_pfn = usable_startpfn;
4613                        }
4614
4615                        /*
4616                         * The usable PFN range for ZONE_MOVABLE is from
4617                         * start_pfn->end_pfn. Calculate size_pages as the
4618                         * number of pages used as kernelcore
4619                         */
4620                        size_pages = end_pfn - start_pfn;
4621                        if (size_pages > kernelcore_remaining)
4622                                size_pages = kernelcore_remaining;
4623                        zone_movable_pfn[nid] = start_pfn + size_pages;
4624
4625                        /*
4626                         * Some kernelcore has been met, update counts and
4627                         * break if the kernelcore for this node has been
4628                         * satisified
4629                         */
4630                        required_kernelcore -= min(required_kernelcore,
4631                                                                size_pages);
4632                        kernelcore_remaining -= size_pages;
4633                        if (!kernelcore_remaining)
4634                                break;
4635                }
4636        }
4637
4638        /*
4639         * If there is still required_kernelcore, we do another pass with one
4640         * less node in the count. This will push zone_movable_pfn[nid] further
4641         * along on the nodes that still have memory until kernelcore is
4642         * satisified
4643         */
4644        usable_nodes--;
4645        if (usable_nodes && required_kernelcore > usable_nodes)
4646                goto restart;
4647
4648        /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4649        for (nid = 0; nid < MAX_NUMNODES; nid++)
4650                zone_movable_pfn[nid] =
4651                        roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4652
4653out:
4654        /* restore the node_state */
4655        node_states[N_HIGH_MEMORY] = saved_node_state;
4656}
4657
4658/* Any regular memory on that node ? */
4659static void check_for_regular_memory(pg_data_t *pgdat)
4660{
4661#ifdef CONFIG_HIGHMEM
4662        enum zone_type zone_type;
4663
4664        for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4665                struct zone *zone = &pgdat->node_zones[zone_type];
4666                if (zone->present_pages)
4667                        node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4668        }
4669#endif
4670}
4671
4672/**
4673 * free_area_init_nodes - Initialise all pg_data_t and zone data
4674 * @max_zone_pfn: an array of max PFNs for each zone
4675 *
4676 * This will call free_area_init_node() for each active node in the system.
4677 * Using the page ranges provided by add_active_range(), the size of each
4678 * zone in each node and their holes is calculated. If the maximum PFN
4679 * between two adjacent zones match, it is assumed that the zone is empty.
4680 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4681 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4682 * starts where the previous one ended. For example, ZONE_DMA32 starts
4683 * at arch_max_dma_pfn.
4684 */
4685void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4686{
4687        unsigned long nid;
4688        int i;
4689
4690        /* Sort early_node_map as initialisation assumes it is sorted */
4691        sort_node_map();
4692
4693        /* Record where the zone boundaries are */
4694        memset(arch_zone_lowest_possible_pfn, 0,
4695                                sizeof(arch_zone_lowest_possible_pfn));
4696        memset(arch_zone_highest_possible_pfn, 0,
4697                                sizeof(arch_zone_highest_possible_pfn));
4698        arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4699        arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4700        for (i = 1; i < MAX_NR_ZONES; i++) {
4701                if (i == ZONE_MOVABLE)
4702                        continue;
4703                arch_zone_lowest_possible_pfn[i] =
4704                        arch_zone_highest_possible_pfn[i-1];
4705                arch_zone_highest_possible_pfn[i] =
4706                        max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4707        }
4708        arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4709        arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4710
4711        /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4712        memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4713        find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4714
4715        /* Print out the zone ranges */
4716        printk("Zone PFN ranges:\n");
4717        for (i = 0; i < MAX_NR_ZONES; i++) {
4718                if (i == ZONE_MOVABLE)
4719                        continue;
4720                printk("  %-8s ", zone_names[i]);
4721                if (arch_zone_lowest_possible_pfn[i] ==
4722                                arch_zone_highest_possible_pfn[i])
4723                        printk("empty\n");
4724                else
4725                        printk("%0#10lx -> %0#10lx\n",
4726                                arch_zone_lowest_possible_pfn[i],
4727                                arch_zone_highest_possible_pfn[i]);
4728        }
4729
4730        /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4731        printk("Movable zone start PFN for each node\n");
4732        for (i = 0; i < MAX_NUMNODES; i++) {
4733                if (zone_movable_pfn[i])
4734                        printk("  Node %d: %lu\n", i, zone_movable_pfn[i]);
4735        }
4736
4737        /* Print out the early_node_map[] */
4738        printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4739        for (i = 0; i < nr_nodemap_entries; i++)
4740                printk("  %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4741                                                early_node_map[i].start_pfn,
4742                                                early_node_map[i].end_pfn);
4743
4744        /* Initialise every node */
4745        mminit_verify_pageflags_layout();
4746        setup_nr_node_ids();
4747        for_each_online_node(nid) {
4748                pg_data_t *pgdat = NODE_DATA(nid);
4749                free_area_init_node(nid, NULL,
4750                                find_min_pfn_for_node(nid), NULL);
4751
4752                /* Any memory on that node */
4753                if (pgdat->node_present_pages)
4754                        node_set_state(nid, N_HIGH_MEMORY);
4755                check_for_regular_memory(pgdat);
4756        }
4757}
4758
4759static int __init cmdline_parse_core(char *p, unsigned long *core)
4760{
4761        unsigned long long coremem;
4762        if (!p)
4763                return -EINVAL;
4764
4765        coremem = memparse(p, &p);
4766        *core = coremem >> PAGE_SHIFT;
4767
4768        /* Paranoid check that UL is enough for the coremem value */
4769        WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4770
4771        return 0;
4772}
4773
4774/*
4775 * kernelcore=size sets the amount of memory for use for allocations that
4776 * cannot be reclaimed or migrated.
4777 */
4778static int __init cmdline_parse_kernelcore(char *p)
4779{
4780        return cmdline_parse_core(p, &required_kernelcore);
4781}
4782
4783/*
4784 * movablecore=size sets the amount of memory for use for allocations that
4785 * can be reclaimed or migrated.
4786 */
4787static int __init cmdline_parse_movablecore(char *p)
4788{
4789        return cmdline_parse_core(p, &required_movablecore);
4790}
4791
4792early_param("kernelcore", cmdline_parse_kernelcore);
4793early_param("movablecore", cmdline_parse_movablecore);
4794
4795#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4796
4797/**
4798 * set_dma_reserve - set the specified number of pages reserved in the first zone
4799 * @new_dma_reserve: The number of pages to mark reserved
4800 *
4801 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4802 * In the DMA zone, a significant percentage may be consumed by kernel image
4803 * and other unfreeable allocations which can skew the watermarks badly. This
4804 * function may optionally be used to account for unfreeable pages in the
4805 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4806 * smaller per-cpu batchsize.
4807 */
4808void __init set_dma_reserve(unsigned long new_dma_reserve)
4809{
4810        dma_reserve = new_dma_reserve;
4811}
4812
4813#ifndef CONFIG_NEED_MULTIPLE_NODES
4814struct pglist_data __refdata contig_page_data = {
4815#ifndef CONFIG_NO_BOOTMEM
4816 .bdata = &bootmem_node_data[0]
4817#endif
4818 };
4819EXPORT_SYMBOL(contig_page_data);
4820#endif
4821
4822void __init free_area_init(unsigned long *zones_size)
4823{
4824        free_area_init_node(0, zones_size,
4825                        __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4826}
4827
4828static int page_alloc_cpu_notify(struct notifier_block *self,
4829                                 unsigned long action, void *hcpu)
4830{
4831        int cpu = (unsigned long)hcpu;
4832
4833        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4834                drain_pages(cpu);
4835
4836                /*
4837                 * Spill the event counters of the dead processor
4838                 * into the current processors event counters.
4839                 * This artificially elevates the count of the current
4840                 * processor.
4841                 */
4842                vm_events_fold_cpu(cpu);
4843
4844                /*
4845                 * Zero the differential counters of the dead processor
4846                 * so that the vm statistics are consistent.
4847                 *
4848                 * This is only okay since the processor is dead and cannot
4849                 * race with what we are doing.
4850                 */
4851                refresh_cpu_vm_stats(cpu);
4852        }
4853        return NOTIFY_OK;
4854}
4855
4856void __init page_alloc_init(void)
4857{
4858        hotcpu_notifier(page_alloc_cpu_notify, 0);
4859}
4860
4861/*
4862 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4863 *      or min_free_kbytes changes.
4864 */
4865static void calculate_totalreserve_pages(void)
4866{
4867        struct pglist_data *pgdat;
4868        unsigned long reserve_pages = 0;
4869        enum zone_type i, j;
4870
4871        for_each_online_pgdat(pgdat) {
4872                for (i = 0; i < MAX_NR_ZONES; i++) {
4873                        struct zone *zone = pgdat->node_zones + i;
4874                        unsigned long max = 0;
4875
4876                        /* Find valid and maximum lowmem_reserve in the zone */
4877                        for (j = i; j < MAX_NR_ZONES; j++) {
4878                                if (zone->lowmem_reserve[j] > max)
4879                                        max = zone->lowmem_reserve[j];
4880                        }
4881
4882                        /* we treat the high watermark as reserved pages. */
4883                        max += high_wmark_pages(zone);
4884
4885                        if (max > zone->present_pages)
4886                                max = zone->present_pages;
4887                        reserve_pages += max;
4888                }
4889        }
4890        totalreserve_pages = reserve_pages;
4891}
4892
4893/*
4894 * setup_per_zone_lowmem_reserve - called whenever
4895 *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
4896 *      has a correct pages reserved value, so an adequate number of
4897 *      pages are left in the zone after a successful __alloc_pages().
4898 */
4899static void setup_per_zone_lowmem_reserve(void)
4900{
4901        struct pglist_data *pgdat;
4902        enum zone_type j, idx;
4903
4904        for_each_online_pgdat(pgdat) {
4905                for (j = 0; j < MAX_NR_ZONES; j++) {
4906                        struct zone *zone = pgdat->node_zones + j;
4907                        unsigned long present_pages = zone->present_pages;
4908
4909                        zone->lowmem_reserve[j] = 0;
4910
4911                        idx = j;
4912                        while (idx) {
4913                                struct zone *lower_zone;
4914
4915                                idx--;
4916
4917                                if (sysctl_lowmem_reserve_ratio[idx] < 1)
4918                                        sysctl_lowmem_reserve_ratio[