linux/mm/page_alloc.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/page_alloc.c
   3 *
   4 *  Manages the free list, the system allocates free pages here.
   5 *  Note that kmalloc() lives in slab.c
   6 *
   7 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   8 *  Swap reorganised 29.12.95, Stephen Tweedie
   9 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
  10 *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
  11 *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
  12 *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  13 *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  14 *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
  15 */
  16
  17#include <linux/stddef.h>
  18#include <linux/mm.h>
  19#include <linux/swap.h>
  20#include <linux/interrupt.h>
  21#include <linux/pagemap.h>
  22#include <linux/jiffies.h>
  23#include <linux/bootmem.h>
  24#include <linux/memblock.h>
  25#include <linux/compiler.h>
  26#include <linux/kernel.h>
  27#include <linux/kmemcheck.h>
  28#include <linux/module.h>
  29#include <linux/suspend.h>
  30#include <linux/pagevec.h>
  31#include <linux/blkdev.h>
  32#include <linux/slab.h>
  33#include <linux/ratelimit.h>
  34#include <linux/oom.h>
  35#include <linux/notifier.h>
  36#include <linux/topology.h>
  37#include <linux/sysctl.h>
  38#include <linux/cpu.h>
  39#include <linux/cpuset.h>
  40#include <linux/memory_hotplug.h>
  41#include <linux/nodemask.h>
  42#include <linux/vmalloc.h>
  43#include <linux/vmstat.h>
  44#include <linux/mempolicy.h>
  45#include <linux/stop_machine.h>
  46#include <linux/sort.h>
  47#include <linux/pfn.h>
  48#include <linux/backing-dev.h>
  49#include <linux/fault-inject.h>
  50#include <linux/page-isolation.h>
  51#include <linux/page_cgroup.h>
  52#include <linux/debugobjects.h>
  53#include <linux/kmemleak.h>
  54#include <linux/memory.h>
  55#include <linux/compaction.h>
  56#include <trace/events/kmem.h>
  57#include <linux/ftrace_event.h>
  58#include <linux/memcontrol.h>
  59#include <linux/prefetch.h>
  60
  61#include <asm/tlbflush.h>
  62#include <asm/div64.h>
  63#include "internal.h"
  64
  65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
  66DEFINE_PER_CPU(int, numa_node);
  67EXPORT_PER_CPU_SYMBOL(numa_node);
  68#endif
  69
  70#ifdef CONFIG_HAVE_MEMORYLESS_NODES
  71/*
  72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
  73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
  74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
  75 * defined in <linux/topology.h>.
  76 */
  77DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
  78EXPORT_PER_CPU_SYMBOL(_numa_mem_);
  79#endif
  80
  81/*
  82 * Array of node states.
  83 */
  84nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
  85        [N_POSSIBLE] = NODE_MASK_ALL,
  86        [N_ONLINE] = { { [0] = 1UL } },
  87#ifndef CONFIG_NUMA
  88        [N_NORMAL_MEMORY] = { { [0] = 1UL } },
  89#ifdef CONFIG_HIGHMEM
  90        [N_HIGH_MEMORY] = { { [0] = 1UL } },
  91#endif
  92        [N_CPU] = { { [0] = 1UL } },
  93#endif  /* NUMA */
  94};
  95EXPORT_SYMBOL(node_states);
  96
  97unsigned long totalram_pages __read_mostly;
  98unsigned long totalreserve_pages __read_mostly;
  99int percpu_pagelist_fraction;
 100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
 101
 102#ifdef CONFIG_PM_SLEEP
 103/*
 104 * The following functions are used by the suspend/hibernate code to temporarily
 105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
 106 * while devices are suspended.  To avoid races with the suspend/hibernate code,
 107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
 108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
 109 * guaranteed not to run in parallel with that modification).
 110 */
 111
 112static gfp_t saved_gfp_mask;
 113
 114void pm_restore_gfp_mask(void)
 115{
 116        WARN_ON(!mutex_is_locked(&pm_mutex));
 117        if (saved_gfp_mask) {
 118                gfp_allowed_mask = saved_gfp_mask;
 119                saved_gfp_mask = 0;
 120        }
 121}
 122
 123void pm_restrict_gfp_mask(void)
 124{
 125        WARN_ON(!mutex_is_locked(&pm_mutex));
 126        WARN_ON(saved_gfp_mask);
 127        saved_gfp_mask = gfp_allowed_mask;
 128        gfp_allowed_mask &= ~GFP_IOFS;
 129}
 130#endif /* CONFIG_PM_SLEEP */
 131
 132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 133int pageblock_order __read_mostly;
 134#endif
 135
 136static void __free_pages_ok(struct page *page, unsigned int order);
 137
 138/*
 139 * results with 256, 32 in the lowmem_reserve sysctl:
 140 *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 141 *      1G machine -> (16M dma, 784M normal, 224M high)
 142 *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 143 *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 144 *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
 145 *
 146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 147 * don't need any ZONE_NORMAL reservation
 148 */
 149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
 150#ifdef CONFIG_ZONE_DMA
 151         256,
 152#endif
 153#ifdef CONFIG_ZONE_DMA32
 154         256,
 155#endif
 156#ifdef CONFIG_HIGHMEM
 157         32,
 158#endif
 159         32,
 160};
 161
 162EXPORT_SYMBOL(totalram_pages);
 163
 164static char * const zone_names[MAX_NR_ZONES] = {
 165#ifdef CONFIG_ZONE_DMA
 166         "DMA",
 167#endif
 168#ifdef CONFIG_ZONE_DMA32
 169         "DMA32",
 170#endif
 171         "Normal",
 172#ifdef CONFIG_HIGHMEM
 173         "HighMem",
 174#endif
 175         "Movable",
 176};
 177
 178int min_free_kbytes = 1024;
 179
 180static unsigned long __meminitdata nr_kernel_pages;
 181static unsigned long __meminitdata nr_all_pages;
 182static unsigned long __meminitdata dma_reserve;
 183
 184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
 185  /*
 186   * MAX_ACTIVE_REGIONS determines the maximum number of distinct
 187   * ranges of memory (RAM) that may be registered with add_active_range().
 188   * Ranges passed to add_active_range() will be merged if possible
 189   * so the number of times add_active_range() can be called is
 190   * related to the number of nodes and the number of holes
 191   */
 192  #ifdef CONFIG_MAX_ACTIVE_REGIONS
 193    /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
 194    #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
 195  #else
 196    #if MAX_NUMNODES >= 32
 197      /* If there can be many nodes, allow up to 50 holes per node */
 198      #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
 199    #else
 200      /* By default, allow up to 256 distinct regions */
 201      #define MAX_ACTIVE_REGIONS 256
 202    #endif
 203  #endif
 204
 205  static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
 206  static int __meminitdata nr_nodemap_entries;
 207  static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
 208  static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
 209  static unsigned long __initdata required_kernelcore;
 210  static unsigned long __initdata required_movablecore;
 211  static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
 212
 213  /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 214  int movable_zone;
 215  EXPORT_SYMBOL(movable_zone);
 216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
 217
 218#if MAX_NUMNODES > 1
 219int nr_node_ids __read_mostly = MAX_NUMNODES;
 220int nr_online_nodes __read_mostly = 1;
 221EXPORT_SYMBOL(nr_node_ids);
 222EXPORT_SYMBOL(nr_online_nodes);
 223#endif
 224
 225int page_group_by_mobility_disabled __read_mostly;
 226
 227static void set_pageblock_migratetype(struct page *page, int migratetype)
 228{
 229
 230        if (unlikely(page_group_by_mobility_disabled))
 231                migratetype = MIGRATE_UNMOVABLE;
 232
 233        set_pageblock_flags_group(page, (unsigned long)migratetype,
 234                                        PB_migrate, PB_migrate_end);
 235}
 236
 237bool oom_killer_disabled __read_mostly;
 238
 239#ifdef CONFIG_DEBUG_VM
 240static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 241{
 242        int ret = 0;
 243        unsigned seq;
 244        unsigned long pfn = page_to_pfn(page);
 245
 246        do {
 247                seq = zone_span_seqbegin(zone);
 248                if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
 249                        ret = 1;
 250                else if (pfn < zone->zone_start_pfn)
 251                        ret = 1;
 252        } while (zone_span_seqretry(zone, seq));
 253
 254        return ret;
 255}
 256
 257static int page_is_consistent(struct zone *zone, struct page *page)
 258{
 259        if (!pfn_valid_within(page_to_pfn(page)))
 260                return 0;
 261        if (zone != page_zone(page))
 262                return 0;
 263
 264        return 1;
 265}
 266/*
 267 * Temporary debugging check for pages not lying within a given zone.
 268 */
 269static int bad_range(struct zone *zone, struct page *page)
 270{
 271        if (page_outside_zone_boundaries(zone, page))
 272                return 1;
 273        if (!page_is_consistent(zone, page))
 274                return 1;
 275
 276        return 0;
 277}
 278#else
 279static inline int bad_range(struct zone *zone, struct page *page)
 280{
 281        return 0;
 282}
 283#endif
 284
 285static void bad_page(struct page *page)
 286{
 287        static unsigned long resume;
 288        static unsigned long nr_shown;
 289        static unsigned long nr_unshown;
 290
 291        /* Don't complain about poisoned pages */
 292        if (PageHWPoison(page)) {
 293                reset_page_mapcount(page); /* remove PageBuddy */
 294                return;
 295        }
 296
 297        /*
 298         * Allow a burst of 60 reports, then keep quiet for that minute;
 299         * or allow a steady drip of one report per second.
 300         */
 301        if (nr_shown == 60) {
 302                if (time_before(jiffies, resume)) {
 303                        nr_unshown++;
 304                        goto out;
 305                }
 306                if (nr_unshown) {
 307                        printk(KERN_ALERT
 308                              "BUG: Bad page state: %lu messages suppressed\n",
 309                                nr_unshown);
 310                        nr_unshown = 0;
 311                }
 312                nr_shown = 0;
 313        }
 314        if (nr_shown++ == 0)
 315                resume = jiffies + 60 * HZ;
 316
 317        printk(KERN_ALERT "BUG: Bad page state in process %s  pfn:%05lx\n",
 318                current->comm, page_to_pfn(page));
 319        dump_page(page);
 320
 321        dump_stack();
 322out:
 323        /* Leave bad fields for debug, except PageBuddy could make trouble */
 324        reset_page_mapcount(page); /* remove PageBuddy */
 325        add_taint(TAINT_BAD_PAGE);
 326}
 327
 328/*
 329 * Higher-order pages are called "compound pages".  They are structured thusly:
 330 *
 331 * The first PAGE_SIZE page is called the "head page".
 332 *
 333 * The remaining PAGE_SIZE pages are called "tail pages".
 334 *
 335 * All pages have PG_compound set.  All pages have their ->private pointing at
 336 * the head page (even the head page has this).
 337 *
 338 * The first tail page's ->lru.next holds the address of the compound page's
 339 * put_page() function.  Its ->lru.prev holds the order of allocation.
 340 * This usage means that zero-order pages may not be compound.
 341 */
 342
 343static void free_compound_page(struct page *page)
 344{
 345        __free_pages_ok(page, compound_order(page));
 346}
 347
 348void prep_compound_page(struct page *page, unsigned long order)
 349{
 350        int i;
 351        int nr_pages = 1 << order;
 352
 353        set_compound_page_dtor(page, free_compound_page);
 354        set_compound_order(page, order);
 355        __SetPageHead(page);
 356        for (i = 1; i < nr_pages; i++) {
 357                struct page *p = page + i;
 358
 359                __SetPageTail(p);
 360                p->first_page = page;
 361        }
 362}
 363
 364/* update __split_huge_page_refcount if you change this function */
 365static int destroy_compound_page(struct page *page, unsigned long order)
 366{
 367        int i;
 368        int nr_pages = 1 << order;
 369        int bad = 0;
 370
 371        if (unlikely(compound_order(page) != order) ||
 372            unlikely(!PageHead(page))) {
 373                bad_page(page);
 374                bad++;
 375        }
 376
 377        __ClearPageHead(page);
 378
 379        for (i = 1; i < nr_pages; i++) {
 380                struct page *p = page + i;
 381
 382                if (unlikely(!PageTail(p) || (p->first_page != page))) {
 383                        bad_page(page);
 384                        bad++;
 385                }
 386                __ClearPageTail(p);
 387        }
 388
 389        return bad;
 390}
 391
 392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
 393{
 394        int i;
 395
 396        /*
 397         * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
 398         * and __GFP_HIGHMEM from hard or soft interrupt context.
 399         */
 400        VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
 401        for (i = 0; i < (1 << order); i++)
 402                clear_highpage(page + i);
 403}
 404
 405static inline void set_page_order(struct page *page, int order)
 406{
 407        set_page_private(page, order);
 408        __SetPageBuddy(page);
 409}
 410
 411static inline void rmv_page_order(struct page *page)
 412{
 413        __ClearPageBuddy(page);
 414        set_page_private(page, 0);
 415}
 416
 417/*
 418 * Locate the struct page for both the matching buddy in our
 419 * pair (buddy1) and the combined O(n+1) page they form (page).
 420 *
 421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
 422 * the following equation:
 423 *     B2 = B1 ^ (1 << O)
 424 * For example, if the starting buddy (buddy2) is #8 its order
 425 * 1 buddy is #10:
 426 *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 427 *
 428 * 2) Any buddy B will have an order O+1 parent P which
 429 * satisfies the following equation:
 430 *     P = B & ~(1 << O)
 431 *
 432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
 433 */
 434static inline unsigned long
 435__find_buddy_index(unsigned long page_idx, unsigned int order)
 436{
 437        return page_idx ^ (1 << order);
 438}
 439
 440/*
 441 * This function checks whether a page is free && is the buddy
 442 * we can do coalesce a page and its buddy if
 443 * (a) the buddy is not in a hole &&
 444 * (b) the buddy is in the buddy system &&
 445 * (c) a page and its buddy have the same order &&
 446 * (d) a page and its buddy are in the same zone.
 447 *
 448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
 449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
 450 *
 451 * For recording page's order, we use page_private(page).
 452 */
 453static inline int page_is_buddy(struct page *page, struct page *buddy,
 454                                                                int order)
 455{
 456        if (!pfn_valid_within(page_to_pfn(buddy)))
 457                return 0;
 458
 459        if (page_zone_id(page) != page_zone_id(buddy))
 460                return 0;
 461
 462        if (PageBuddy(buddy) && page_order(buddy) == order) {
 463                VM_BUG_ON(page_count(buddy) != 0);
 464                return 1;
 465        }
 466        return 0;
 467}
 468
 469/*
 470 * Freeing function for a buddy system allocator.
 471 *
 472 * The concept of a buddy system is to maintain direct-mapped table
 473 * (containing bit values) for memory blocks of various "orders".
 474 * The bottom level table contains the map for the smallest allocatable
 475 * units of memory (here, pages), and each level above it describes
 476 * pairs of units from the levels below, hence, "buddies".
 477 * At a high level, all that happens here is marking the table entry
 478 * at the bottom level available, and propagating the changes upward
 479 * as necessary, plus some accounting needed to play nicely with other
 480 * parts of the VM system.
 481 * At each level, we keep a list of pages, which are heads of continuous
 482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
 483 * order is recorded in page_private(page) field.
 484 * So when we are allocating or freeing one, we can derive the state of the
 485 * other.  That is, if we allocate a small block, and both were   
 486 * free, the remainder of the region must be split into blocks.   
 487 * If a block is freed, and its buddy is also free, then this
 488 * triggers coalescing into a block of larger size.            
 489 *
 490 * -- wli
 491 */
 492
 493static inline void __free_one_page(struct page *page,
 494                struct zone *zone, unsigned int order,
 495                int migratetype)
 496{
 497        unsigned long page_idx;
 498        unsigned long combined_idx;
 499        unsigned long uninitialized_var(buddy_idx);
 500        struct page *buddy;
 501
 502        if (unlikely(PageCompound(page)))
 503                if (unlikely(destroy_compound_page(page, order)))
 504                        return;
 505
 506        VM_BUG_ON(migratetype == -1);
 507
 508        page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 509
 510        VM_BUG_ON(page_idx & ((1 << order) - 1));
 511        VM_BUG_ON(bad_range(zone, page));
 512
 513        while (order < MAX_ORDER-1) {
 514                buddy_idx = __find_buddy_index(page_idx, order);
 515                buddy = page + (buddy_idx - page_idx);
 516                if (!page_is_buddy(page, buddy, order))
 517                        break;
 518
 519                /* Our buddy is free, merge with it and move up one order. */
 520                list_del(&buddy->lru);
 521                zone->free_area[order].nr_free--;
 522                rmv_page_order(buddy);
 523                combined_idx = buddy_idx & page_idx;
 524                page = page + (combined_idx - page_idx);
 525                page_idx = combined_idx;
 526                order++;
 527        }
 528        set_page_order(page, order);
 529
 530        /*
 531         * If this is not the largest possible page, check if the buddy
 532         * of the next-highest order is free. If it is, it's possible
 533         * that pages are being freed that will coalesce soon. In case,
 534         * that is happening, add the free page to the tail of the list
 535         * so it's less likely to be used soon and more likely to be merged
 536         * as a higher order page
 537         */
 538        if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
 539                struct page *higher_page, *higher_buddy;
 540                combined_idx = buddy_idx & page_idx;
 541                higher_page = page + (combined_idx - page_idx);
 542                buddy_idx = __find_buddy_index(combined_idx, order + 1);
 543                higher_buddy = page + (buddy_idx - combined_idx);
 544                if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
 545                        list_add_tail(&page->lru,
 546                                &zone->free_area[order].free_list[migratetype]);
 547                        goto out;
 548                }
 549        }
 550
 551        list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
 552out:
 553        zone->free_area[order].nr_free++;
 554}
 555
 556/*
 557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
 558 * Page should not be on lru, so no need to fix that up.
 559 * free_pages_check() will verify...
 560 */
 561static inline void free_page_mlock(struct page *page)
 562{
 563        __dec_zone_page_state(page, NR_MLOCK);
 564        __count_vm_event(UNEVICTABLE_MLOCKFREED);
 565}
 566
 567static inline int free_pages_check(struct page *page)
 568{
 569        if (unlikely(page_mapcount(page) |
 570                (page->mapping != NULL)  |
 571                (atomic_read(&page->_count) != 0) |
 572                (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
 573                (mem_cgroup_bad_page_check(page)))) {
 574                bad_page(page);
 575                return 1;
 576        }
 577        if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
 578                page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
 579        return 0;
 580}
 581
 582/*
 583 * Frees a number of pages from the PCP lists
 584 * Assumes all pages on list are in same zone, and of same order.
 585 * count is the number of pages to free.
 586 *
 587 * If the zone was previously in an "all pages pinned" state then look to
 588 * see if this freeing clears that state.
 589 *
 590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
 591 * pinned" detection logic.
 592 */
 593static void free_pcppages_bulk(struct zone *zone, int count,
 594                                        struct per_cpu_pages *pcp)
 595{
 596        int migratetype = 0;
 597        int batch_free = 0;
 598        int to_free = count;
 599
 600        spin_lock(&zone->lock);
 601        zone->all_unreclaimable = 0;
 602        zone->pages_scanned = 0;
 603
 604        while (to_free) {
 605                struct page *page;
 606                struct list_head *list;
 607
 608                /*
 609                 * Remove pages from lists in a round-robin fashion. A
 610                 * batch_free count is maintained that is incremented when an
 611                 * empty list is encountered.  This is so more pages are freed
 612                 * off fuller lists instead of spinning excessively around empty
 613                 * lists
 614                 */
 615                do {
 616                        batch_free++;
 617                        if (++migratetype == MIGRATE_PCPTYPES)
 618                                migratetype = 0;
 619                        list = &pcp->lists[migratetype];
 620                } while (list_empty(list));
 621
 622                /* This is the only non-empty list. Free them all. */
 623                if (batch_free == MIGRATE_PCPTYPES)
 624                        batch_free = to_free;
 625
 626                do {
 627                        page = list_entry(list->prev, struct page, lru);
 628                        /* must delete as __free_one_page list manipulates */
 629                        list_del(&page->lru);
 630                        /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
 631                        __free_one_page(page, zone, 0, page_private(page));
 632                        trace_mm_page_pcpu_drain(page, 0, page_private(page));
 633                } while (--to_free && --batch_free && !list_empty(list));
 634        }
 635        __mod_zone_page_state(zone, NR_FREE_PAGES, count);
 636        spin_unlock(&zone->lock);
 637}
 638
 639static void free_one_page(struct zone *zone, struct page *page, int order,
 640                                int migratetype)
 641{
 642        spin_lock(&zone->lock);
 643        zone->all_unreclaimable = 0;
 644        zone->pages_scanned = 0;
 645
 646        __free_one_page(page, zone, order, migratetype);
 647        __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 648        spin_unlock(&zone->lock);
 649}
 650
 651static bool free_pages_prepare(struct page *page, unsigned int order)
 652{
 653        int i;
 654        int bad = 0;
 655
 656        trace_mm_page_free_direct(page, order);
 657        kmemcheck_free_shadow(page, order);
 658
 659        if (PageAnon(page))
 660                page->mapping = NULL;
 661        for (i = 0; i < (1 << order); i++)
 662                bad += free_pages_check(page + i);
 663        if (bad)
 664                return false;
 665
 666        if (!PageHighMem(page)) {
 667                debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
 668                debug_check_no_obj_freed(page_address(page),
 669                                           PAGE_SIZE << order);
 670        }
 671        arch_free_page(page, order);
 672        kernel_map_pages(page, 1 << order, 0);
 673
 674        return true;
 675}
 676
 677static void __free_pages_ok(struct page *page, unsigned int order)
 678{
 679        unsigned long flags;
 680        int wasMlocked = __TestClearPageMlocked(page);
 681
 682        if (!free_pages_prepare(page, order))
 683                return;
 684
 685        local_irq_save(flags);
 686        if (unlikely(wasMlocked))
 687                free_page_mlock(page);
 688        __count_vm_events(PGFREE, 1 << order);
 689        free_one_page(page_zone(page), page, order,
 690                                        get_pageblock_migratetype(page));
 691        local_irq_restore(flags);
 692}
 693
 694/*
 695 * permit the bootmem allocator to evade page validation on high-order frees
 696 */
 697void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 698{
 699        if (order == 0) {
 700                __ClearPageReserved(page);
 701                set_page_count(page, 0);
 702                set_page_refcounted(page);
 703                __free_page(page);
 704        } else {
 705                int loop;
 706
 707                prefetchw(page);
 708                for (loop = 0; loop < BITS_PER_LONG; loop++) {
 709                        struct page *p = &page[loop];
 710
 711                        if (loop + 1 < BITS_PER_LONG)
 712                                prefetchw(p + 1);
 713                        __ClearPageReserved(p);
 714                        set_page_count(p, 0);
 715                }
 716
 717                set_page_refcounted(page);
 718                __free_pages(page, order);
 719        }
 720}
 721
 722
 723/*
 724 * The order of subdivision here is critical for the IO subsystem.
 725 * Please do not alter this order without good reasons and regression
 726 * testing. Specifically, as large blocks of memory are subdivided,
 727 * the order in which smaller blocks are delivered depends on the order
 728 * they're subdivided in this function. This is the primary factor
 729 * influencing the order in which pages are delivered to the IO
 730 * subsystem according to empirical testing, and this is also justified
 731 * by considering the behavior of a buddy system containing a single
 732 * large block of memory acted on by a series of small allocations.
 733 * This behavior is a critical factor in sglist merging's success.
 734 *
 735 * -- wli
 736 */
 737static inline void expand(struct zone *zone, struct page *page,
 738        int low, int high, struct free_area *area,
 739        int migratetype)
 740{
 741        unsigned long size = 1 << high;
 742
 743        while (high > low) {
 744                area--;
 745                high--;
 746                size >>= 1;
 747                VM_BUG_ON(bad_range(zone, &page[size]));
 748                list_add(&page[size].lru, &area->free_list[migratetype]);
 749                area->nr_free++;
 750                set_page_order(&page[size], high);
 751        }
 752}
 753
 754/*
 755 * This page is about to be returned from the page allocator
 756 */
 757static inline int check_new_page(struct page *page)
 758{
 759        if (unlikely(page_mapcount(page) |
 760                (page->mapping != NULL)  |
 761                (atomic_read(&page->_count) != 0)  |
 762                (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
 763                (mem_cgroup_bad_page_check(page)))) {
 764                bad_page(page);
 765                return 1;
 766        }
 767        return 0;
 768}
 769
 770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 771{
 772        int i;
 773
 774        for (i = 0; i < (1 << order); i++) {
 775                struct page *p = page + i;
 776                if (unlikely(check_new_page(p)))
 777                        return 1;
 778        }
 779
 780        set_page_private(page, 0);
 781        set_page_refcounted(page);
 782
 783        arch_alloc_page(page, order);
 784        kernel_map_pages(page, 1 << order, 1);
 785
 786        if (gfp_flags & __GFP_ZERO)
 787                prep_zero_page(page, order, gfp_flags);
 788
 789        if (order && (gfp_flags & __GFP_COMP))
 790                prep_compound_page(page, order);
 791
 792        return 0;
 793}
 794
 795/*
 796 * Go through the free lists for the given migratetype and remove
 797 * the smallest available page from the freelists
 798 */
 799static inline
 800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
 801                                                int migratetype)
 802{
 803        unsigned int current_order;
 804        struct free_area * area;
 805        struct page *page;
 806
 807        /* Find a page of the appropriate size in the preferred list */
 808        for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 809                area = &(zone->free_area[current_order]);
 810                if (list_empty(&area->free_list[migratetype]))
 811                        continue;
 812
 813                page = list_entry(area->free_list[migratetype].next,
 814                                                        struct page, lru);
 815                list_del(&page->lru);
 816                rmv_page_order(page);
 817                area->nr_free--;
 818                expand(zone, page, order, current_order, area, migratetype);
 819                return page;
 820        }
 821
 822        return NULL;
 823}
 824
 825
 826/*
 827 * This array describes the order lists are fallen back to when
 828 * the free lists for the desirable migrate type are depleted
 829 */
 830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
 831        [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 832        [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 833        [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
 834        [MIGRATE_RESERVE]     = { MIGRATE_RESERVE,     MIGRATE_RESERVE,   MIGRATE_RESERVE }, /* Never used */
 835};
 836
 837/*
 838 * Move the free pages in a range to the free lists of the requested type.
 839 * Note that start_page and end_pages are not aligned on a pageblock
 840 * boundary. If alignment is required, use move_freepages_block()
 841 */
 842static int move_freepages(struct zone *zone,
 843                          struct page *start_page, struct page *end_page,
 844                          int migratetype)
 845{
 846        struct page *page;
 847        unsigned long order;
 848        int pages_moved = 0;
 849
 850#ifndef CONFIG_HOLES_IN_ZONE
 851        /*
 852         * page_zone is not safe to call in this context when
 853         * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
 854         * anyway as we check zone boundaries in move_freepages_block().
 855         * Remove at a later date when no bug reports exist related to
 856         * grouping pages by mobility
 857         */
 858        BUG_ON(page_zone(start_page) != page_zone(end_page));
 859#endif
 860
 861        for (page = start_page; page <= end_page;) {
 862                /* Make sure we are not inadvertently changing nodes */
 863                VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
 864
 865                if (!pfn_valid_within(page_to_pfn(page))) {
 866                        page++;
 867                        continue;
 868                }
 869
 870                if (!PageBuddy(page)) {
 871                        page++;
 872                        continue;
 873                }
 874
 875                order = page_order(page);
 876                list_move(&page->lru,
 877                          &zone->free_area[order].free_list[migratetype]);
 878                page += 1 << order;
 879                pages_moved += 1 << order;
 880        }
 881
 882        return pages_moved;
 883}
 884
 885static int move_freepages_block(struct zone *zone, struct page *page,
 886                                int migratetype)
 887{
 888        unsigned long start_pfn, end_pfn;
 889        struct page *start_page, *end_page;
 890
 891        start_pfn = page_to_pfn(page);
 892        start_pfn = start_pfn & ~(pageblock_nr_pages-1);
 893        start_page = pfn_to_page(start_pfn);
 894        end_page = start_page + pageblock_nr_pages - 1;
 895        end_pfn = start_pfn + pageblock_nr_pages - 1;
 896
 897        /* Do not cross zone boundaries */
 898        if (start_pfn < zone->zone_start_pfn)
 899                start_page = page;
 900        if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
 901                return 0;
 902
 903        return move_freepages(zone, start_page, end_page, migratetype);
 904}
 905
 906static void change_pageblock_range(struct page *pageblock_page,
 907                                        int start_order, int migratetype)
 908{
 909        int nr_pageblocks = 1 << (start_order - pageblock_order);
 910
 911        while (nr_pageblocks--) {
 912                set_pageblock_migratetype(pageblock_page, migratetype);
 913                pageblock_page += pageblock_nr_pages;
 914        }
 915}
 916
 917/* Remove an element from the buddy allocator from the fallback list */
 918static inline struct page *
 919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
 920{
 921        struct free_area * area;
 922        int current_order;
 923        struct page *page;
 924        int migratetype, i;
 925
 926        /* Find the largest possible block of pages in the other list */
 927        for (current_order = MAX_ORDER-1; current_order >= order;
 928                                                --current_order) {
 929                for (i = 0; i < MIGRATE_TYPES - 1; i++) {
 930                        migratetype = fallbacks[start_migratetype][i];
 931
 932                        /* MIGRATE_RESERVE handled later if necessary */
 933                        if (migratetype == MIGRATE_RESERVE)
 934                                continue;
 935
 936                        area = &(zone->free_area[current_order]);
 937                        if (list_empty(&area->free_list[migratetype]))
 938                                continue;
 939
 940                        page = list_entry(area->free_list[migratetype].next,
 941                                        struct page, lru);
 942                        area->nr_free--;
 943
 944                        /*
 945                         * If breaking a large block of pages, move all free
 946                         * pages to the preferred allocation list. If falling
 947                         * back for a reclaimable kernel allocation, be more
 948                         * aggressive about taking ownership of free pages
 949                         */
 950                        if (unlikely(current_order >= (pageblock_order >> 1)) ||
 951                                        start_migratetype == MIGRATE_RECLAIMABLE ||
 952                                        page_group_by_mobility_disabled) {
 953                                unsigned long pages;
 954                                pages = move_freepages_block(zone, page,
 955                                                                start_migratetype);
 956
 957                                /* Claim the whole block if over half of it is free */
 958                                if (pages >= (1 << (pageblock_order-1)) ||
 959                                                page_group_by_mobility_disabled)
 960                                        set_pageblock_migratetype(page,
 961                                                                start_migratetype);
 962
 963                                migratetype = start_migratetype;
 964                        }
 965
 966                        /* Remove the page from the freelists */
 967                        list_del(&page->lru);
 968                        rmv_page_order(page);
 969
 970                        /* Take ownership for orders >= pageblock_order */
 971                        if (current_order >= pageblock_order)
 972                                change_pageblock_range(page, current_order,
 973                                                        start_migratetype);
 974
 975                        expand(zone, page, order, current_order, area, migratetype);
 976
 977                        trace_mm_page_alloc_extfrag(page, order, current_order,
 978                                start_migratetype, migratetype);
 979
 980                        return page;
 981                }
 982        }
 983
 984        return NULL;
 985}
 986
 987/*
 988 * Do the hard work of removing an element from the buddy allocator.
 989 * Call me with the zone->lock already held.
 990 */
 991static struct page *__rmqueue(struct zone *zone, unsigned int order,
 992                                                int migratetype)
 993{
 994        struct page *page;
 995
 996retry_reserve:
 997        page = __rmqueue_smallest(zone, order, migratetype);
 998
 999        if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000                page = __rmqueue_fallback(zone, order, migratetype);
1001
1002                /*
1003                 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004                 * is used because __rmqueue_smallest is an inline function
1005                 * and we want just one call site
1006                 */
1007                if (!page) {
1008                        migratetype = MIGRATE_RESERVE;
1009                        goto retry_reserve;
1010                }
1011        }
1012
1013        trace_mm_page_alloc_zone_locked(page, order, migratetype);
1014        return page;
1015}
1016
1017/* 
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency.  Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1021 */
1022static int rmqueue_bulk(struct zone *zone, unsigned int order, 
1023                        unsigned long count, struct list_head *list,
1024                        int migratetype, int cold)
1025{
1026        int i;
1027        
1028        spin_lock(&zone->lock);
1029        for (i = 0; i < count; ++i) {
1030                struct page *page = __rmqueue(zone, order, migratetype);
1031                if (unlikely(page == NULL))
1032                        break;
1033
1034                /*
1035                 * Split buddy pages returned by expand() are received here
1036                 * in physical page order. The page is added to the callers and
1037                 * list and the list head then moves forward. From the callers
1038                 * perspective, the linked list is ordered by page number in
1039                 * some conditions. This is useful for IO devices that can
1040                 * merge IO requests if the physical pages are ordered
1041                 * properly.
1042                 */
1043                if (likely(cold == 0))
1044                        list_add(&page->lru, list);
1045                else
1046                        list_add_tail(&page->lru, list);
1047                set_page_private(page, migratetype);
1048                list = &page->lru;
1049        }
1050        __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051        spin_unlock(&zone->lock);
1052        return i;
1053}
1054
1055#ifdef CONFIG_NUMA
1056/*
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1059 * expired.
1060 *
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1063 */
1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1065{
1066        unsigned long flags;
1067        int to_drain;
1068
1069        local_irq_save(flags);
1070        if (pcp->count >= pcp->batch)
1071                to_drain = pcp->batch;
1072        else
1073                to_drain = pcp->count;
1074        free_pcppages_bulk(zone, to_drain, pcp);
1075        pcp->count -= to_drain;
1076        local_irq_restore(flags);
1077}
1078#endif
1079
1080/*
1081 * Drain pages of the indicated processor.
1082 *
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1085 * is not online.
1086 */
1087static void drain_pages(unsigned int cpu)
1088{
1089        unsigned long flags;
1090        struct zone *zone;
1091
1092        for_each_populated_zone(zone) {
1093                struct per_cpu_pageset *pset;
1094                struct per_cpu_pages *pcp;
1095
1096                local_irq_save(flags);
1097                pset = per_cpu_ptr(zone->pageset, cpu);
1098
1099                pcp = &pset->pcp;
1100                if (pcp->count) {
1101                        free_pcppages_bulk(zone, pcp->count, pcp);
1102                        pcp->count = 0;
1103                }
1104                local_irq_restore(flags);
1105        }
1106}
1107
1108/*
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1110 */
1111void drain_local_pages(void *arg)
1112{
1113        drain_pages(smp_processor_id());
1114}
1115
1116/*
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1118 */
1119void drain_all_pages(void)
1120{
1121        on_each_cpu(drain_local_pages, NULL, 1);
1122}
1123
1124#ifdef CONFIG_HIBERNATION
1125
1126void mark_free_pages(struct zone *zone)
1127{
1128        unsigned long pfn, max_zone_pfn;
1129        unsigned long flags;
1130        int order, t;
1131        struct list_head *curr;
1132
1133        if (!zone->spanned_pages)
1134                return;
1135
1136        spin_lock_irqsave(&zone->lock, flags);
1137
1138        max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139        for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140                if (pfn_valid(pfn)) {
1141                        struct page *page = pfn_to_page(pfn);
1142
1143                        if (!swsusp_page_is_forbidden(page))
1144                                swsusp_unset_page_free(page);
1145                }
1146
1147        for_each_migratetype_order(order, t) {
1148                list_for_each(curr, &zone->free_area[order].free_list[t]) {
1149                        unsigned long i;
1150
1151                        pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152                        for (i = 0; i < (1UL << order); i++)
1153                                swsusp_set_page_free(pfn_to_page(pfn + i));
1154                }
1155        }
1156        spin_unlock_irqrestore(&zone->lock, flags);
1157}
1158#endif /* CONFIG_PM */
1159
1160/*
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1163 */
1164void free_hot_cold_page(struct page *page, int cold)
1165{
1166        struct zone *zone = page_zone(page);
1167        struct per_cpu_pages *pcp;
1168        unsigned long flags;
1169        int migratetype;
1170        int wasMlocked = __TestClearPageMlocked(page);
1171
1172        if (!free_pages_prepare(page, 0))
1173                return;
1174
1175        migratetype = get_pageblock_migratetype(page);
1176        set_page_private(page, migratetype);
1177        local_irq_save(flags);
1178        if (unlikely(wasMlocked))
1179                free_page_mlock(page);
1180        __count_vm_event(PGFREE);
1181
1182        /*
1183         * We only track unmovable, reclaimable and movable on pcp lists.
1184         * Free ISOLATE pages back to the allocator because they are being
1185         * offlined but treat RESERVE as movable pages so we can get those
1186         * areas back if necessary. Otherwise, we may have to free
1187         * excessively into the page allocator
1188         */
1189        if (migratetype >= MIGRATE_PCPTYPES) {
1190                if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191                        free_one_page(zone, page, 0, migratetype);
1192                        goto out;
1193                }
1194                migratetype = MIGRATE_MOVABLE;
1195        }
1196
1197        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1198        if (cold)
1199                list_add_tail(&page->lru, &pcp->lists[migratetype]);
1200        else
1201                list_add(&page->lru, &pcp->lists[migratetype]);
1202        pcp->count++;
1203        if (pcp->count >= pcp->high) {
1204                free_pcppages_bulk(zone, pcp->batch, pcp);
1205                pcp->count -= pcp->batch;
1206        }
1207
1208out:
1209        local_irq_restore(flags);
1210}
1211
1212/*
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1216 *
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1219 */
1220void split_page(struct page *page, unsigned int order)
1221{
1222        int i;
1223
1224        VM_BUG_ON(PageCompound(page));
1225        VM_BUG_ON(!page_count(page));
1226
1227#ifdef CONFIG_KMEMCHECK
1228        /*
1229         * Split shadow pages too, because free(page[0]) would
1230         * otherwise free the whole shadow.
1231         */
1232        if (kmemcheck_page_is_tracked(page))
1233                split_page(virt_to_page(page[0].shadow), order);
1234#endif
1235
1236        for (i = 1; i < (1 << order); i++)
1237                set_page_refcounted(page + i);
1238}
1239
1240/*
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * are enabled.
1246 *
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1249 */
1250int split_free_page(struct page *page)
1251{
1252        unsigned int order;
1253        unsigned long watermark;
1254        struct zone *zone;
1255
1256        BUG_ON(!PageBuddy(page));
1257
1258        zone = page_zone(page);
1259        order = page_order(page);
1260
1261        /* Obey watermarks as if the page was being allocated */
1262        watermark = low_wmark_pages(zone) + (1 << order);
1263        if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1264                return 0;
1265
1266        /* Remove page from free list */
1267        list_del(&page->lru);
1268        zone->free_area[order].nr_free--;
1269        rmv_page_order(page);
1270        __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271
1272        /* Split into individual pages */
1273        set_page_refcounted(page);
1274        split_page(page, order);
1275
1276        if (order >= pageblock_order - 1) {
1277                struct page *endpage = page + (1 << order) - 1;
1278                for (; page < endpage; page += pageblock_nr_pages)
1279                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1280        }
1281
1282        return 1 << order;
1283}
1284
1285/*
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
1287 * we cheat by calling it from here, in the order > 0 path.  Saves a branch
1288 * or two.
1289 */
1290static inline
1291struct page *buffered_rmqueue(struct zone *preferred_zone,
1292                        struct zone *zone, int order, gfp_t gfp_flags,
1293                        int migratetype)
1294{
1295        unsigned long flags;
1296        struct page *page;
1297        int cold = !!(gfp_flags & __GFP_COLD);
1298
1299again:
1300        if (likely(order == 0)) {
1301                struct per_cpu_pages *pcp;
1302                struct list_head *list;
1303
1304                local_irq_save(flags);
1305                pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306                list = &pcp->lists[migratetype];
1307                if (list_empty(list)) {
1308                        pcp->count += rmqueue_bulk(zone, 0,
1309                                        pcp->batch, list,
1310                                        migratetype, cold);
1311                        if (unlikely(list_empty(list)))
1312                                goto failed;
1313                }
1314
1315                if (cold)
1316                        page = list_entry(list->prev, struct page, lru);
1317                else
1318                        page = list_entry(list->next, struct page, lru);
1319
1320                list_del(&page->lru);
1321                pcp->count--;
1322        } else {
1323                if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324                        /*
1325                         * __GFP_NOFAIL is not to be used in new code.
1326                         *
1327                         * All __GFP_NOFAIL callers should be fixed so that they
1328                         * properly detect and handle allocation failures.
1329                         *
1330                         * We most definitely don't want callers attempting to
1331                         * allocate greater than order-1 page units with
1332                         * __GFP_NOFAIL.
1333                         */
1334                        WARN_ON_ONCE(order > 1);
1335                }
1336                spin_lock_irqsave(&zone->lock, flags);
1337                page = __rmqueue(zone, order, migratetype);
1338                spin_unlock(&zone->lock);
1339                if (!page)
1340                        goto failed;
1341                __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1342        }
1343
1344        __count_zone_vm_events(PGALLOC, zone, 1 << order);
1345        zone_statistics(preferred_zone, zone, gfp_flags);
1346        local_irq_restore(flags);
1347
1348        VM_BUG_ON(bad_range(zone, page));
1349        if (prep_new_page(page, order, gfp_flags))
1350                goto again;
1351        return page;
1352
1353failed:
1354        local_irq_restore(flags);
1355        return NULL;
1356}
1357
1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359#define ALLOC_WMARK_MIN         WMARK_MIN
1360#define ALLOC_WMARK_LOW         WMARK_LOW
1361#define ALLOC_WMARK_HIGH        WMARK_HIGH
1362#define ALLOC_NO_WATERMARKS     0x04 /* don't check watermarks at all */
1363
1364/* Mask to get the watermark bits */
1365#define ALLOC_WMARK_MASK        (ALLOC_NO_WATERMARKS-1)
1366
1367#define ALLOC_HARDER            0x10 /* try to alloc harder */
1368#define ALLOC_HIGH              0x20 /* __GFP_HIGH set */
1369#define ALLOC_CPUSET            0x40 /* check for correct cpuset */
1370
1371#ifdef CONFIG_FAIL_PAGE_ALLOC
1372
1373static struct {
1374        struct fault_attr attr;
1375
1376        u32 ignore_gfp_highmem;
1377        u32 ignore_gfp_wait;
1378        u32 min_order;
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
1413        dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1414                                        &fail_page_alloc.attr);
1415        if (IS_ERR(dir))
1416                return PTR_ERR(dir);
1417
1418        if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1419                                &fail_page_alloc.ignore_gfp_wait))
1420                goto fail;
1421        if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1422                                &fail_page_alloc.ignore_gfp_highmem))
1423                goto fail;
1424        if (!debugfs_create_u32("min-order", mode, dir,
1425                                &fail_page_alloc.min_order))
1426                goto fail;
1427
1428        return 0;
1429fail:
1430        debugfs_remove_recursive(dir);
1431
1432        return -ENOMEM;
1433}
1434
1435late_initcall(fail_page_alloc_debugfs);
1436
1437#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1438
1439#else /* CONFIG_FAIL_PAGE_ALLOC */
1440
1441static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1442{
1443        return 0;
1444}
1445
1446#endif /* CONFIG_FAIL_PAGE_ALLOC */
1447
1448/*
1449 * Return true if free pages are above 'mark'. This takes into account the order
1450 * of the allocation.
1451 */
1452static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1453                      int classzone_idx, int alloc_flags, long free_pages)
1454{
1455        /* free_pages my go negative - that's OK */
1456        long min = mark;
1457        int o;
1458
1459        free_pages -= (1 << order) + 1;
1460        if (alloc_flags & ALLOC_HIGH)
1461                min -= min / 2;
1462        if (alloc_flags & ALLOC_HARDER)
1463                min -= min / 4;
1464
1465        if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1466                return false;
1467        for (o = 0; o < order; o++) {
1468                /* At the next order, this order's pages become unavailable */
1469                free_pages -= z->free_area[o].nr_free << o;
1470
1471                /* Require fewer higher order pages to be free */
1472                min >>= 1;
1473
1474                if (free_pages <= min)
1475                        return false;
1476        }
1477        return true;
1478}
1479
1480bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1481                      int classzone_idx, int alloc_flags)
1482{
1483        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1484                                        zone_page_state(z, NR_FREE_PAGES));
1485}
1486
1487bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1488                      int classzone_idx, int alloc_flags)
1489{
1490        long free_pages = zone_page_state(z, NR_FREE_PAGES);
1491
1492        if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1493                free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1494
1495        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1496                                                                free_pages);
1497}
1498
1499#ifdef CONFIG_NUMA
1500/*
1501 * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1502 * skip over zones that are not allowed by the cpuset, or that have
1503 * been recently (in last second) found to be nearly full.  See further
1504 * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1505 * that have to skip over a lot of full or unallowed zones.
1506 *
1507 * If the zonelist cache is present in the passed in zonelist, then
1508 * returns a pointer to the allowed node mask (either the current
1509 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1510 *
1511 * If the zonelist cache is not available for this zonelist, does
1512 * nothing and returns NULL.
1513 *
1514 * If the fullzones BITMAP in the zonelist cache is stale (more than
1515 * a second since last zap'd) then we zap it out (clear its bits.)
1516 *
1517 * We hold off even calling zlc_setup, until after we've checked the
1518 * first zone in the zonelist, on the theory that most allocations will
1519 * be satisfied from that first zone, so best to examine that zone as
1520 * quickly as we can.
1521 */
1522static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1523{
1524        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1525        nodemask_t *allowednodes;       /* zonelist_cache approximation */
1526
1527        zlc = zonelist->zlcache_ptr;
1528        if (!zlc)
1529                return NULL;
1530
1531        if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1532                bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1533                zlc->last_full_zap = jiffies;
1534        }
1535
1536        allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1537                                        &cpuset_current_mems_allowed :
1538                                        &node_states[N_HIGH_MEMORY];
1539        return allowednodes;
1540}
1541
1542/*
1543 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1544 * if it is worth looking at further for free memory:
1545 *  1) Check that the zone isn't thought to be full (doesn't have its
1546 *     bit set in the zonelist_cache fullzones BITMAP).
1547 *  2) Check that the zones node (obtained from the zonelist_cache
1548 *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1549 * Return true (non-zero) if zone is worth looking at further, or
1550 * else return false (zero) if it is not.
1551 *
1552 * This check -ignores- the distinction between various watermarks,
1553 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1554 * found to be full for any variation of these watermarks, it will
1555 * be considered full for up to one second by all requests, unless
1556 * we are so low on memory on all allowed nodes that we are forced
1557 * into the second scan of the zonelist.
1558 *
1559 * In the second scan we ignore this zonelist cache and exactly
1560 * apply the watermarks to all zones, even it is slower to do so.
1561 * We are low on memory in the second scan, and should leave no stone
1562 * unturned looking for a free page.
1563 */
1564static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1565                                                nodemask_t *allowednodes)
1566{
1567        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1568        int i;                          /* index of *z in zonelist zones */
1569        int n;                          /* node that zone *z is on */
1570
1571        zlc = zonelist->zlcache_ptr;
1572        if (!zlc)
1573                return 1;
1574
1575        i = z - zonelist->_zonerefs;
1576        n = zlc->z_to_n[i];
1577
1578        /* This zone is worth trying if it is allowed but not full */
1579        return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1580}
1581
1582/*
1583 * Given 'z' scanning a zonelist, set the corresponding bit in
1584 * zlc->fullzones, so that subsequent attempts to allocate a page
1585 * from that zone don't waste time re-examining it.
1586 */
1587static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1588{
1589        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1590        int i;                          /* index of *z in zonelist zones */
1591
1592        zlc = zonelist->zlcache_ptr;
1593        if (!zlc)
1594                return;
1595
1596        i = z - zonelist->_zonerefs;
1597
1598        set_bit(i, zlc->fullzones);
1599}
1600
1601/*
1602 * clear all zones full, called after direct reclaim makes progress so that
1603 * a zone that was recently full is not skipped over for up to a second
1604 */
1605static void zlc_clear_zones_full(struct zonelist *zonelist)
1606{
1607        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1608
1609        zlc = zonelist->zlcache_ptr;
1610        if (!zlc)
1611                return;
1612
1613        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1614}
1615
1616#else   /* CONFIG_NUMA */
1617
1618static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1619{
1620        return NULL;
1621}
1622
1623static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1624                                nodemask_t *allowednodes)
1625{
1626        return 1;
1627}
1628
1629static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1630{
1631}
1632
1633static void zlc_clear_zones_full(struct zonelist *zonelist)
1634{
1635}
1636#endif  /* CONFIG_NUMA */
1637
1638/*
1639 * get_page_from_freelist goes through the zonelist trying to allocate
1640 * a page.
1641 */
1642static struct page *
1643get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1644                struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1645                struct zone *preferred_zone, int migratetype)
1646{
1647        struct zoneref *z;
1648        struct page *page = NULL;
1649        int classzone_idx;
1650        struct zone *zone;
1651        nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1652        int zlc_active = 0;             /* set if using zonelist_cache */
1653        int did_zlc_setup = 0;          /* just call zlc_setup() one time */
1654
1655        classzone_idx = zone_idx(preferred_zone);
1656zonelist_scan:
1657        /*
1658         * Scan zonelist, looking for a zone with enough free.
1659         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1660         */
1661        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1662                                                high_zoneidx, nodemask) {
1663                if (NUMA_BUILD && zlc_active &&
1664                        !zlc_zone_worth_trying(zonelist, z, allowednodes))
1665                                continue;
1666                if ((alloc_flags & ALLOC_CPUSET) &&
1667                        !cpuset_zone_allowed_softwall(zone, gfp_mask))
1668                                continue;
1669
1670                BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1671                if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1672                        unsigned long mark;
1673                        int ret;
1674
1675                        mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1676                        if (zone_watermark_ok(zone, order, mark,
1677                                    classzone_idx, alloc_flags))
1678                                goto try_this_zone;
1679
1680                        if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681                                /*
1682                                 * we do zlc_setup if there are multiple nodes
1683                                 * and before considering the first zone allowed
1684                                 * by the cpuset.
1685                                 */
1686                                allowednodes = zlc_setup(zonelist, alloc_flags);
1687                                zlc_active = 1;
1688                                did_zlc_setup = 1;
1689                        }
1690
1691                        if (zone_reclaim_mode == 0)
1692                                goto this_zone_full;
1693
1694                        /*
1695                         * As we may have just activated ZLC, check if the first
1696                         * eligible zone has failed zone_reclaim recently.
1697                         */
1698                        if (NUMA_BUILD && zlc_active &&
1699                                !zlc_zone_worth_trying(zonelist, z, allowednodes))
1700                                continue;
1701
1702                        ret = zone_reclaim(zone, gfp_mask, order);
1703                        switch (ret) {
1704                        case ZONE_RECLAIM_NOSCAN:
1705                                /* did not scan */
1706                                continue;
1707                        case ZONE_RECLAIM_FULL:
1708                                /* scanned but unreclaimable */
1709                                continue;
1710                        default:
1711                                /* did we reclaim enough */
1712                                if (!zone_watermark_ok(zone, order, mark,
1713                                                classzone_idx, alloc_flags))
1714                                        goto this_zone_full;
1715                        }
1716                }
1717
1718try_this_zone:
1719                page = buffered_rmqueue(preferred_zone, zone, order,
1720                                                gfp_mask, migratetype);
1721                if (page)
1722                        break;
1723this_zone_full:
1724                if (NUMA_BUILD)
1725                        zlc_mark_zone_full(zonelist, z);
1726        }
1727
1728        if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1729                /* Disable zlc cache for second zonelist scan */
1730                zlc_active = 0;
1731                goto zonelist_scan;
1732        }
1733        return page;
1734}
1735
1736/*
1737 * Large machines with many possible nodes should not always dump per-node
1738 * meminfo in irq context.
1739 */
1740static inline bool should_suppress_show_mem(void)
1741{
1742        bool ret = false;
1743
1744#if NODES_SHIFT > 8
1745        ret = in_interrupt();
1746#endif
1747        return ret;
1748}
1749
1750static DEFINE_RATELIMIT_STATE(nopage_rs,
1751                DEFAULT_RATELIMIT_INTERVAL,
1752                DEFAULT_RATELIMIT_BURST);
1753
1754void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1755{
1756        va_list args;
1757        unsigned int filter = SHOW_MEM_FILTER_NODES;
1758
1759        if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1760                return;
1761
1762        /*
1763         * This documents exceptions given to allocations in certain
1764         * contexts that are allowed to allocate outside current's set
1765         * of allowed nodes.
1766         */
1767        if (!(gfp_mask & __GFP_NOMEMALLOC))
1768                if (test_thread_flag(TIF_MEMDIE) ||
1769                    (current->flags & (PF_MEMALLOC | PF_EXITING)))
1770                        filter &= ~SHOW_MEM_FILTER_NODES;
1771        if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1772                filter &= ~SHOW_MEM_FILTER_NODES;
1773
1774        if (fmt) {
1775                printk(KERN_WARNING);
1776                va_start(args, fmt);
1777                vprintk(fmt, args);
1778                va_end(args);
1779        }
1780
1781        pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1782                   current->comm, order, gfp_mask);
1783
1784        dump_stack();
1785        if (!should_suppress_show_mem())
1786                show_mem(filter);
1787}
1788
1789static inline int
1790should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1791                                unsigned long pages_reclaimed)
1792{
1793        /* Do not loop if specifically requested */
1794        if (gfp_mask & __GFP_NORETRY)
1795                return 0;
1796
1797        /*
1798         * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1799         * means __GFP_NOFAIL, but that may not be true in other
1800         * implementations.
1801         */
1802        if (order <= PAGE_ALLOC_COSTLY_ORDER)
1803                return 1;
1804
1805        /*
1806         * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1807         * specified, then we retry until we no longer reclaim any pages
1808         * (above), or we've reclaimed an order of pages at least as
1809         * large as the allocation's order. In both cases, if the
1810         * allocation still fails, we stop retrying.
1811         */
1812        if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1813                return 1;
1814
1815        /*
1816         * Don't let big-order allocations loop unless the caller
1817         * explicitly requests that.
1818         */
1819        if (gfp_mask & __GFP_NOFAIL)
1820                return 1;
1821
1822        return 0;
1823}
1824
1825static inline struct page *
1826__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1827        struct zonelist *zonelist, enum zone_type high_zoneidx,
1828        nodemask_t *nodemask, struct zone *preferred_zone,
1829        int migratetype)
1830{
1831        struct page *page;
1832
1833        /* Acquire the OOM killer lock for the zones in zonelist */
1834        if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1835                schedule_timeout_uninterruptible(1);
1836                return NULL;
1837        }
1838
1839        /*
1840         * Go through the zonelist yet one more time, keep very high watermark
1841         * here, this is only to catch a parallel oom killing, we must fail if
1842         * we're still under heavy pressure.
1843         */
1844        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1845                order, zonelist, high_zoneidx,
1846                ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1847                preferred_zone, migratetype);
1848        if (page)
1849                goto out;
1850
1851        if (!(gfp_mask & __GFP_NOFAIL)) {
1852                /* The OOM killer will not help higher order allocs */
1853                if (order > PAGE_ALLOC_COSTLY_ORDER)
1854                        goto out;
1855                /* The OOM killer does not needlessly kill tasks for lowmem */
1856                if (high_zoneidx < ZONE_NORMAL)
1857                        goto out;
1858                /*
1859                 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1860                 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1861                 * The caller should handle page allocation failure by itself if
1862                 * it specifies __GFP_THISNODE.
1863                 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1864                 */
1865                if (gfp_mask & __GFP_THISNODE)
1866                        goto out;
1867        }
1868        /* Exhausted what can be done so it's blamo time */
1869        out_of_memory(zonelist, gfp_mask, order, nodemask);
1870
1871out:
1872        clear_zonelist_oom(zonelist, gfp_mask);
1873        return page;
1874}
1875
1876#ifdef CONFIG_COMPACTION
1877/* Try memory compaction for high-order allocations before reclaim */
1878static struct page *
1879__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1880        struct zonelist *zonelist, enum zone_type high_zoneidx,
1881        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1882        int migratetype, unsigned long *did_some_progress,
1883        bool sync_migration)
1884{
1885        struct page *page;
1886
1887        if (!order || compaction_deferred(preferred_zone))
1888                return NULL;
1889
1890        current->flags |= PF_MEMALLOC;
1891        *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1892                                                nodemask, sync_migration);
1893        current->flags &= ~PF_MEMALLOC;
1894        if (*did_some_progress != COMPACT_SKIPPED) {
1895
1896                /* Page migration frees to the PCP lists but we want merging */
1897                drain_pages(get_cpu());
1898                put_cpu();
1899
1900                page = get_page_from_freelist(gfp_mask, nodemask,
1901                                order, zonelist, high_zoneidx,
1902                                alloc_flags, preferred_zone,
1903                                migratetype);
1904                if (page) {
1905                        preferred_zone->compact_considered = 0;
1906                        preferred_zone->compact_defer_shift = 0;
1907                        count_vm_event(COMPACTSUCCESS);
1908                        return page;
1909                }
1910
1911                /*
1912                 * It's bad if compaction run occurs and fails.
1913                 * The most likely reason is that pages exist,
1914                 * but not enough to satisfy watermarks.
1915                 */
1916                count_vm_event(COMPACTFAIL);
1917                defer_compaction(preferred_zone);
1918
1919                cond_resched();
1920        }
1921
1922        return NULL;
1923}
1924#else
1925static inline struct page *
1926__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1927        struct zonelist *zonelist, enum zone_type high_zoneidx,
1928        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1929        int migratetype, unsigned long *did_some_progress,
1930        bool sync_migration)
1931{
1932        return NULL;
1933}
1934#endif /* CONFIG_COMPACTION */
1935
1936/* The really slow allocator path where we enter direct reclaim */
1937static inline struct page *
1938__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1939        struct zonelist *zonelist, enum zone_type high_zoneidx,
1940        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1941        int migratetype, unsigned long *did_some_progress)
1942{
1943        struct page *page = NULL;
1944        struct reclaim_state reclaim_state;
1945        bool drained = false;
1946
1947        cond_resched();
1948
1949        /* We now go into synchronous reclaim */
1950        cpuset_memory_pressure_bump();
1951        current->flags |= PF_MEMALLOC;
1952        lockdep_set_current_reclaim_state(gfp_mask);
1953        reclaim_state.reclaimed_slab = 0;
1954        current->reclaim_state = &reclaim_state;
1955
1956        *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1957
1958        current->reclaim_state = NULL;
1959        lockdep_clear_current_reclaim_state();
1960        current->flags &= ~PF_MEMALLOC;
1961
1962        cond_resched();
1963
1964        if (unlikely(!(*did_some_progress)))
1965                return NULL;
1966
1967        /* After successful reclaim, reconsider all zones for allocation */
1968        if (NUMA_BUILD)
1969                zlc_clear_zones_full(zonelist);
1970
1971retry:
1972        page = get_page_from_freelist(gfp_mask, nodemask, order,
1973                                        zonelist, high_zoneidx,
1974                                        alloc_flags, preferred_zone,
1975                                        migratetype);
1976
1977        /*
1978         * If an allocation failed after direct reclaim, it could be because
1979         * pages are pinned on the per-cpu lists. Drain them and try again
1980         */
1981        if (!page && !drained) {
1982                drain_all_pages();
1983                drained = true;
1984                goto retry;
1985        }
1986
1987        return page;
1988}
1989
1990/*
1991 * This is called in the allocator slow-path if the allocation request is of
1992 * sufficient urgency to ignore watermarks and take other desperate measures
1993 */
1994static inline struct page *
1995__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1996        struct zonelist *zonelist, enum zone_type high_zoneidx,
1997        nodemask_t *nodemask, struct zone *preferred_zone,
1998        int migratetype)
1999{
2000        struct page *page;
2001
2002        do {
2003                page = get_page_from_freelist(gfp_mask, nodemask, order,
2004                        zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2005                        preferred_zone, migratetype);
2006
2007                if (!page && gfp_mask & __GFP_NOFAIL)
2008                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2009        } while (!page && (gfp_mask & __GFP_NOFAIL));
2010
2011        return page;
2012}
2013
2014static inline
2015void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2016                                                enum zone_type high_zoneidx,
2017                                                enum zone_type classzone_idx)
2018{
2019        struct zoneref *z;
2020        struct zone *zone;
2021
2022        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2023                wakeup_kswapd(zone, order, classzone_idx);
2024}
2025
2026static inline int
2027gfp_to_alloc_flags(gfp_t gfp_mask)
2028{
2029        int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2030        const gfp_t wait = gfp_mask & __GFP_WAIT;
2031
2032        /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2033        BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2034
2035        /*
2036         * The caller may dip into page reserves a bit more if the caller
2037         * cannot run direct reclaim, or if the caller has realtime scheduling
2038         * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
2039         * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2040         */
2041        alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2042
2043        if (!wait) {
2044                /*
2045                 * Not worth trying to allocate harder for
2046                 * __GFP_NOMEMALLOC even if it can't schedule.
2047                 */
2048                if  (!(gfp_mask & __GFP_NOMEMALLOC))
2049                        alloc_flags |= ALLOC_HARDER;
2050                /*
2051                 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2052                 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2053                 */
2054                alloc_flags &= ~ALLOC_CPUSET;
2055        } else if (unlikely(rt_task(current)) && !in_interrupt())
2056                alloc_flags |= ALLOC_HARDER;
2057
2058        if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2059                if (!in_interrupt() &&
2060                    ((current->flags & PF_MEMALLOC) ||
2061                     unlikely(test_thread_flag(TIF_MEMDIE))))
2062                        alloc_flags |= ALLOC_NO_WATERMARKS;
2063        }
2064
2065        return alloc_flags;
2066}
2067
2068static inline struct page *
2069__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2070        struct zonelist *zonelist, enum zone_type high_zoneidx,
2071        nodemask_t *nodemask, struct zone *preferred_zone,
2072        int migratetype)
2073{
2074        const gfp_t wait = gfp_mask & __GFP_WAIT;
2075        struct page *page = NULL;
2076        int alloc_flags;
2077        unsigned long pages_reclaimed = 0;
2078        unsigned long did_some_progress;
2079        bool sync_migration = false;
2080
2081        /*
2082         * In the slowpath, we sanity check order to avoid ever trying to
2083         * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2084         * be using allocators in order of preference for an area that is
2085         * too large.
2086         */
2087        if (order >= MAX_ORDER) {
2088                WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2089                return NULL;
2090        }
2091
2092        /*
2093         * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2094         * __GFP_NOWARN set) should not cause reclaim since the subsystem
2095         * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2096         * using a larger set of nodes after it has established that the
2097         * allowed per node queues are empty and that nodes are
2098         * over allocated.
2099         */
2100        if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2101                goto nopage;
2102
2103restart:
2104        if (!(gfp_mask & __GFP_NO_KSWAPD))
2105                wake_all_kswapd(order, zonelist, high_zoneidx,
2106                                                zone_idx(preferred_zone));
2107
2108        /*
2109         * OK, we're below the kswapd watermark and have kicked background
2110         * reclaim. Now things get more complex, so set up alloc_flags according
2111         * to how we want to proceed.
2112         */
2113        alloc_flags = gfp_to_alloc_flags(gfp_mask);
2114
2115        /*
2116         * Find the true preferred zone if the allocation is unconstrained by
2117         * cpusets.
2118         */
2119        if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2120                first_zones_zonelist(zonelist, high_zoneidx, NULL,
2121                                        &preferred_zone);
2122
2123rebalance:
2124        /* This is the last chance, in general, before the goto nopage. */
2125        page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2126                        high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2127                        preferred_zone, migratetype);
2128        if (page)
2129                goto got_pg;
2130
2131        /* Allocate without watermarks if the context allows */
2132        if (alloc_flags & ALLOC_NO_WATERMARKS) {
2133                page = __alloc_pages_high_priority(gfp_mask, order,
2134                                zonelist, high_zoneidx, nodemask,
2135                                preferred_zone, migratetype);
2136                if (page)
2137                        goto got_pg;
2138        }
2139
2140        /* Atomic allocations - we can't balance anything */
2141        if (!wait)
2142                goto nopage;
2143
2144        /* Avoid recursion of direct reclaim */
2145        if (current->flags & PF_MEMALLOC)
2146                goto nopage;
2147
2148        /* Avoid allocations with no watermarks from looping endlessly */
2149        if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2150                goto nopage;
2151
2152        /*
2153         * Try direct compaction. The first pass is asynchronous. Subsequent
2154         * attempts after direct reclaim are synchronous
2155         */
2156        page = __alloc_pages_direct_compact(gfp_mask, order,
2157                                        zonelist, high_zoneidx,
2158                                        nodemask,
2159                                        alloc_flags, preferred_zone,
2160                                        migratetype, &did_some_progress,
2161                                        sync_migration);
2162        if (page)
2163                goto got_pg;
2164        sync_migration = true;
2165
2166        /* Try direct reclaim and then allocating */
2167        page = __alloc_pages_direct_reclaim(gfp_mask, order,
2168                                        zonelist, high_zoneidx,
2169                                        nodemask,
2170                                        alloc_flags, preferred_zone,
2171                                        migratetype, &did_some_progress);
2172        if (page)
2173                goto got_pg;
2174
2175        /*
2176         * If we failed to make any progress reclaiming, then we are
2177         * running out of options and have to consider going OOM
2178         */
2179        if (!did_some_progress) {
2180                if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2181                        if (oom_killer_disabled)
2182                                goto nopage;
2183                        page = __alloc_pages_may_oom(gfp_mask, order,
2184                                        zonelist, high_zoneidx,
2185                                        nodemask, preferred_zone,
2186                                        migratetype);
2187                        if (page)
2188                                goto got_pg;
2189
2190                        if (!(gfp_mask & __GFP_NOFAIL)) {
2191                                /*
2192                                 * The oom killer is not called for high-order
2193                                 * allocations that may fail, so if no progress
2194                                 * is being made, there are no other options and
2195                                 * retrying is unlikely to help.
2196                                 */
2197                                if (order > PAGE_ALLOC_COSTLY_ORDER)
2198                                        goto nopage;
2199                                /*
2200                                 * The oom killer is not called for lowmem
2201                                 * allocations to prevent needlessly killing
2202                                 * innocent tasks.
2203                                 */
2204                                if (high_zoneidx < ZONE_NORMAL)
2205                                        goto nopage;
2206                        }
2207
2208                        goto restart;
2209                }
2210        }
2211
2212        /* Check if we should retry the allocation */
2213        pages_reclaimed += did_some_progress;
2214        if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2215                /* Wait for some write requests to complete then retry */
2216                wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2217                goto rebalance;
2218        } else {
2219                /*
2220                 * High-order allocations do not necessarily loop after
2221                 * direct reclaim and reclaim/compaction depends on compaction
2222                 * being called after reclaim so call directly if necessary
2223                 */
2224                page = __alloc_pages_direct_compact(gfp_mask, order,
2225                                        zonelist, high_zoneidx,
2226                                        nodemask,
2227                                        alloc_flags, preferred_zone,
2228                                        migratetype, &did_some_progress,
2229                                        sync_migration);
2230                if (page)
2231                        goto got_pg;
2232        }
2233
2234nopage:
2235        warn_alloc_failed(gfp_mask, order, NULL);
2236        return page;
2237got_pg:
2238        if (kmemcheck_enabled)
2239                kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2240        return page;
2241
2242}
2243
2244/*
2245 * This is the 'heart' of the zoned buddy allocator.
2246 */
2247struct page *
2248__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2249                        struct zonelist *zonelist, nodemask_t *nodemask)
2250{
2251        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2252        struct zone *preferred_zone;
2253        struct page *page;
2254        int migratetype = allocflags_to_migratetype(gfp_mask);
2255
2256        gfp_mask &= gfp_allowed_mask;
2257
2258        lockdep_trace_alloc(gfp_mask);
2259
2260        might_sleep_if(gfp_mask & __GFP_WAIT);
2261
2262        if (should_fail_alloc_page(gfp_mask, order))
2263                return NULL;
2264
2265        /*
2266         * Check the zones suitable for the gfp_mask contain at least one
2267         * valid zone. It's possible to have an empty zonelist as a result
2268         * of GFP_THISNODE and a memoryless node
2269         */
2270        if (unlikely(!zonelist->_zonerefs->zone))
2271                return NULL;
2272
2273        get_mems_allowed();
2274        /* The preferred zone is used for statistics later */
2275        first_zones_zonelist(zonelist, high_zoneidx,
2276                                nodemask ? : &cpuset_current_mems_allowed,
2277                                &preferred_zone);
2278        if (!preferred_zone) {
2279                put_mems_allowed();
2280                return NULL;
2281        }
2282
2283        /* First allocation attempt */
2284        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2285                        zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2286                        preferred_zone, migratetype);
2287        if (unlikely(!page))
2288                page = __alloc_pages_slowpath(gfp_mask, order,
2289                                zonelist, high_zoneidx, nodemask,
2290                                preferred_zone, migratetype);
2291        put_mems_allowed();
2292
2293        trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2294        return page;
2295}
2296EXPORT_SYMBOL(__alloc_pages_nodemask);
2297
2298/*
2299 * Common helper functions.
2300 */
2301unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2302{
2303        struct page *page;
2304
2305        /*
2306         * __get_free_pages() returns a 32-bit address, which cannot represent
2307         * a highmem page
2308         */
2309        VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2310
2311        page = alloc_pages(gfp_mask, order);
2312        if (!page)
2313                return 0;
2314        return (unsigned long) page_address(page);
2315}
2316EXPORT_SYMBOL(__get_free_pages);
2317
2318unsigned long get_zeroed_page(gfp_t gfp_mask)
2319{
2320        return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2321}
2322EXPORT_SYMBOL(get_zeroed_page);
2323
2324void __pagevec_free(struct pagevec *pvec)
2325{
2326        int i = pagevec_count(pvec);
2327
2328        while (--i >= 0) {
2329                trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2330                free_hot_cold_page(pvec->pages[i], pvec->cold);
2331        }
2332}
2333
2334void __free_pages(struct page *page, unsigned int order)
2335{
2336        if (put_page_testzero(page)) {
2337                if (order == 0)
2338                        free_hot_cold_page(page, 0);
2339                else
2340                        __free_pages_ok(page, order);
2341        }
2342}
2343
2344EXPORT_SYMBOL(__free_pages);
2345
2346void free_pages(unsigned long addr, unsigned int order)
2347{
2348        if (addr != 0) {
2349                VM_BUG_ON(!virt_addr_valid((void *)addr));
2350                __free_pages(virt_to_page((void *)addr), order);
2351        }
2352}
2353
2354EXPORT_SYMBOL(free_pages);
2355
2356static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2357{
2358        if (addr) {
2359                unsigned long alloc_end = addr + (PAGE_SIZE << order);
2360                unsigned long used = addr + PAGE_ALIGN(size);
2361
2362                split_page(virt_to_page((void *)addr), order);
2363                while (used < alloc_end) {
2364                        free_page(used);
2365                        used += PAGE_SIZE;
2366                }
2367        }
2368        return (void *)addr;
2369}
2370
2371/**
2372 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2373 * @size: the number of bytes to allocate
2374 * @gfp_mask: GFP flags for the allocation
2375 *
2376 * This function is similar to alloc_pages(), except that it allocates the
2377 * minimum number of pages to satisfy the request.  alloc_pages() can only
2378 * allocate memory in power-of-two pages.
2379 *
2380 * This function is also limited by MAX_ORDER.
2381 *
2382 * Memory allocated by this function must be released by free_pages_exact().
2383 */
2384void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2385{
2386        unsigned int order = get_order(size);
2387        unsigned long addr;
2388
2389        addr = __get_free_pages(gfp_mask, order);
2390        return make_alloc_exact(addr, order, size);
2391}
2392EXPORT_SYMBOL(alloc_pages_exact);
2393
2394/**
2395 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2396 *                         pages on a node.
2397 * @nid: the preferred node ID where memory should be allocated
2398 * @size: the number of bytes to allocate
2399 * @gfp_mask: GFP flags for the allocation
2400 *
2401 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2402 * back.
2403 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2404 * but is not exact.
2405 */
2406void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2407{
2408        unsigned order = get_order(size);
2409        struct page *p = alloc_pages_node(nid, gfp_mask, order);
2410        if (!p)
2411                return NULL;
2412        return make_alloc_exact((unsigned long)page_address(p), order, size);
2413}
2414EXPORT_SYMBOL(alloc_pages_exact_nid);
2415
2416/**
2417 * free_pages_exact - release memory allocated via alloc_pages_exact()
2418 * @virt: the value returned by alloc_pages_exact.
2419 * @size: size of allocation, same value as passed to alloc_pages_exact().
2420 *
2421 * Release the memory allocated by a previous call to alloc_pages_exact.
2422 */
2423void free_pages_exact(void *virt, size_t size)
2424{
2425        unsigned long addr = (unsigned long)virt;
2426        unsigned long end = addr + PAGE_ALIGN(size);
2427
2428        while (addr < end) {
2429                free_page(addr);
2430                addr += PAGE_SIZE;
2431        }
2432}
2433EXPORT_SYMBOL(free_pages_exact);
2434
2435static unsigned int nr_free_zone_pages(int offset)
2436{
2437        struct zoneref *z;
2438        struct zone *zone;
2439
2440        /* Just pick one node, since fallback list is circular */
2441        unsigned int sum = 0;
2442
2443        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2444
2445        for_each_zone_zonelist(zone, z, zonelist, offset) {
2446                unsigned long size = zone->present_pages;
2447                unsigned long high = high_wmark_pages(zone);
2448                if (size > high)
2449                        sum += size - high;
2450        }
2451
2452        return sum;
2453}
2454
2455/*
2456 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2457 */
2458unsigned int nr_free_buffer_pages(void)
2459{
2460        return nr_free_zone_pages(gfp_zone(GFP_USER));
2461}
2462EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2463
2464/*
2465 * Amount of free RAM allocatable within all zones
2466 */
2467unsigned int nr_free_pagecache_pages(void)
2468{
2469        return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2470}
2471
2472static inline void show_node(struct zone *zone)
2473{
2474        if (NUMA_BUILD)
2475                printk("Node %d ", zone_to_nid(zone));
2476}
2477
2478void si_meminfo(struct sysinfo *val)
2479{
2480        val->totalram = totalram_pages;
2481        val->sharedram = 0;
2482        val->freeram = global_page_state(NR_FREE_PAGES);
2483        val->bufferram = nr_blockdev_pages();
2484        val->totalhigh = totalhigh_pages;
2485        val->freehigh = nr_free_highpages();
2486        val->mem_unit = PAGE_SIZE;
2487}
2488
2489EXPORT_SYMBOL(si_meminfo);
2490
2491#ifdef CONFIG_NUMA
2492void si_meminfo_node(struct sysinfo *val, int nid)
2493{
2494        pg_data_t *pgdat = NODE_DATA(nid);
2495
2496        val->totalram = pgdat->node_present_pages;
2497        val->freeram = node_page_state(nid, NR_FREE_PAGES);
2498#ifdef CONFIG_HIGHMEM
2499        val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2500        val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2501                        NR_FREE_PAGES);
2502#else
2503        val->totalhigh = 0;
2504        val->freehigh = 0;
2505#endif
2506        val->mem_unit = PAGE_SIZE;
2507}
2508#endif
2509
2510/*
2511 * Determine whether the node should be displayed or not, depending on whether
2512 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2513 */
2514bool skip_free_areas_node(unsigned int flags, int nid)
2515{
2516        bool ret = false;
2517
2518        if (!(flags & SHOW_MEM_FILTER_NODES))
2519                goto out;
2520
2521        get_mems_allowed();
2522        ret = !node_isset(nid, cpuset_current_mems_allowed);
2523        put_mems_allowed();
2524out:
2525        return ret;
2526}
2527
2528#define K(x) ((x) << (PAGE_SHIFT-10))
2529
2530/*
2531 * Show free area list (used inside shift_scroll-lock stuff)
2532 * We also calculate the percentage fragmentation. We do this by counting the
2533 * memory on each free list with the exception of the first item on the list.
2534 * Suppresses nodes that are not allowed by current's cpuset if
2535 * SHOW_MEM_FILTER_NODES is passed.
2536 */
2537void show_free_areas(unsigned int filter)
2538{
2539        int cpu;
2540        struct zone *zone;
2541
2542        for_each_populated_zone(zone) {
2543                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2544                        continue;
2545                show_node(zone);
2546                printk("%s per-cpu:\n", zone->name);
2547
2548                for_each_online_cpu(cpu) {
2549                        struct per_cpu_pageset *pageset;
2550
2551                        pageset = per_cpu_ptr(zone->pageset, cpu);
2552
2553                        printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2554                               cpu, pageset->pcp.high,
2555                               pageset->pcp.batch, pageset->pcp.count);
2556                }
2557        }
2558
2559        printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2560                " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2561                " unevictable:%lu"
2562                " dirty:%lu writeback:%lu unstable:%lu\n"
2563                " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2564                " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2565                global_page_state(NR_ACTIVE_ANON),
2566                global_page_state(NR_INACTIVE_ANON),
2567                global_page_state(NR_ISOLATED_ANON),
2568                global_page_state(NR_ACTIVE_FILE),
2569                global_page_state(NR_INACTIVE_FILE),
2570                global_page_state(NR_ISOLATED_FILE),
2571                global_page_state(NR_UNEVICTABLE),
2572                global_page_state(NR_FILE_DIRTY),
2573                global_page_state(NR_WRITEBACK),
2574                global_page_state(NR_UNSTABLE_NFS),
2575                global_page_state(NR_FREE_PAGES),
2576                global_page_state(NR_SLAB_RECLAIMABLE),
2577                global_page_state(NR_SLAB_UNRECLAIMABLE),
2578                global_page_state(NR_FILE_MAPPED),
2579                global_page_state(NR_SHMEM),
2580                global_page_state(NR_PAGETABLE),
2581                global_page_state(NR_BOUNCE));
2582
2583        for_each_populated_zone(zone) {
2584                int i;
2585
2586                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2587                        continue;
2588                show_node(zone);
2589                printk("%s"
2590                        " free:%lukB"
2591                        " min:%lukB"
2592                        " low:%lukB"
2593                        " high:%lukB"
2594                        " active_anon:%lukB"
2595                        " inactive_anon:%lukB"
2596                        " active_file:%lukB"
2597                        " inactive_file:%lukB"
2598                        " unevictable:%lukB"
2599                        " isolated(anon):%lukB"
2600                        " isolated(file):%lukB"
2601                        " present:%lukB"
2602                        " mlocked:%lukB"
2603                        " dirty:%lukB"
2604                        " writeback:%lukB"
2605                        " mapped:%lukB"
2606                        " shmem:%lukB"
2607                        " slab_reclaimable:%lukB"
2608                        " slab_unreclaimable:%lukB"
2609                        " kernel_stack:%lukB"
2610                        " pagetables:%lukB"
2611                        " unstable:%lukB"
2612                        " bounce:%lukB"
2613                        " writeback_tmp:%lukB"
2614                        " pages_scanned:%lu"
2615                        " all_unreclaimable? %s"
2616                        "\n",
2617                        zone->name,
2618                        K(zone_page_state(zone, NR_FREE_PAGES)),
2619                        K(min_wmark_pages(zone)),
2620                        K(low_wmark_pages(zone)),
2621                        K(high_wmark_pages(zone)),
2622                        K(zone_page_state(zone, NR_ACTIVE_ANON)),
2623                        K(zone_page_state(zone, NR_INACTIVE_ANON)),
2624                        K(zone_page_state(zone, NR_ACTIVE_FILE)),
2625                        K(zone_page_state(zone, NR_INACTIVE_FILE)),
2626                        K(zone_page_state(zone, NR_UNEVICTABLE)),
2627                        K(zone_page_state(zone, NR_ISOLATED_ANON)),
2628                        K(zone_page_state(zone, NR_ISOLATED_FILE)),
2629                        K(zone->present_pages),
2630                        K(zone_page_state(zone, NR_MLOCK)),
2631                        K(zone_page_state(zone, NR_FILE_DIRTY)),
2632                        K(zone_page_state(zone, NR_WRITEBACK)),
2633                        K(zone_page_state(zone, NR_FILE_MAPPED)),
2634                        K(zone_page_state(zone, NR_SHMEM)),
2635                        K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2636                        K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2637                        zone_page_state(zone, NR_KERNEL_STACK) *
2638                                THREAD_SIZE / 1024,
2639                        K(zone_page_state(zone, NR_PAGETABLE)),
2640                        K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2641                        K(zone_page_state(zone, NR_BOUNCE)),
2642                        K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2643                        zone->pages_scanned,
2644                        (zone->all_unreclaimable ? "yes" : "no")
2645                        );
2646                printk("lowmem_reserve[]:");
2647                for (i = 0; i < MAX_NR_ZONES; i++)
2648                        printk(" %lu", zone->lowmem_reserve[i]);
2649                printk("\n");
2650        }
2651
2652        for_each_populated_zone(zone) {
2653                unsigned long nr[MAX_ORDER], flags, order, total = 0;
2654
2655                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2656                        continue;
2657                show_node(zone);
2658                printk("%s: ", zone->name);
2659
2660                spin_lock_irqsave(&zone->lock, flags);
2661                for (order = 0; order < MAX_ORDER; order++) {
2662                        nr[order] = zone->free_area[order].nr_free;
2663                        total += nr[order] << order;
2664                }
2665                spin_unlock_irqrestore(&zone->lock, flags);
2666                for (order = 0; order < MAX_ORDER; order++)
2667                        printk("%lu*%lukB ", nr[order], K(1UL) << order);
2668                printk("= %lukB\n", K(total));
2669        }
2670
2671        printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2672
2673        show_swap_cache_info();
2674}
2675
2676static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2677{
2678        zoneref->zone = zone;
2679        zoneref->zone_idx = zone_idx(zone);
2680}
2681
2682/*
2683 * Builds allocation fallback zone lists.
2684 *
2685 * Add all populated zones of a node to the zonelist.
2686 */
2687static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2688                                int nr_zones, enum zone_type zone_type)
2689{
2690        struct zone *zone;
2691
2692        BUG_ON(zone_type >= MAX_NR_ZONES);
2693        zone_type++;
2694
2695        do {
2696                zone_type--;
2697                zone = pgdat->node_zones + zone_type;
2698                if (populated_zone(zone)) {
2699                        zoneref_set_zone(zone,
2700                                &zonelist->_zonerefs[nr_zones++]);
2701                        check_highest_zone(zone_type);
2702                }
2703
2704        } while (zone_type);
2705        return nr_zones;
2706}
2707
2708
2709/*
2710 *  zonelist_order:
2711 *  0 = automatic detection of better ordering.
2712 *  1 = order by ([node] distance, -zonetype)
2713 *  2 = order by (-zonetype, [node] distance)
2714 *
2715 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2716 *  the same zonelist. So only NUMA can configure this param.
2717 */
2718#define ZONELIST_ORDER_DEFAULT  0
2719#define ZONELIST_ORDER_NODE     1
2720#define ZONELIST_ORDER_ZONE     2
2721
2722/* zonelist order in the kernel.
2723 * set_zonelist_order() will set this to NODE or ZONE.
2724 */
2725static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2726static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2727
2728
2729#ifdef CONFIG_NUMA
2730/* The value user specified ....changed by config */
2731static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2732/* string for sysctl */
2733#define NUMA_ZONELIST_ORDER_LEN 16
2734char numa_zonelist_order[16] = "default";
2735
2736/*
2737 * interface for configure zonelist ordering.
2738 * command line option "numa_zonelist_order"
2739 *      = "[dD]efault   - default, automatic configuration.
2740 *      = "[nN]ode      - order by node locality, then by zone within node
2741 *      = "[zZ]one      - order by zone, then by locality within zone
2742 */
2743
2744static int __parse_numa_zonelist_order(char *s)
2745{
2746        if (*s == 'd' || *s == 'D') {
2747                user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2748        } else if (*s == 'n' || *s == 'N') {
2749                user_zonelist_order = ZONELIST_ORDER_NODE;
2750        } else if (*s == 'z' || *s == 'Z') {
2751                user_zonelist_order = ZONELIST_ORDER_ZONE;
2752        } else {
2753                printk(KERN_WARNING
2754                        "Ignoring invalid numa_zonelist_order value:  "
2755                        "%s\n", s);
2756                return -EINVAL;
2757        }
2758        return 0;
2759}
2760
2761static __init int setup_numa_zonelist_order(char *s)
2762{
2763        int ret;
2764
2765        if (!s)
2766                return 0;
2767
2768        ret = __parse_numa_zonelist_order(s);
2769        if (ret == 0)
2770                strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2771
2772        return ret;
2773}
2774early_param("numa_zonelist_order", setup_numa_zonelist_order);
2775
2776/*
2777 * sysctl handler for numa_zonelist_order
2778 */
2779int numa_zonelist_order_handler(ctl_table *table, int write,
2780                void __user *buffer, size_t *length,
2781                loff_t *ppos)
2782{
2783        char saved_string[NUMA_ZONELIST_ORDER_LEN];
2784        int ret;
2785        static DEFINE_MUTEX(zl_order_mutex);
2786
2787        mutex_lock(&zl_order_mutex);
2788        if (write)
2789                strcpy(saved_string, (char*)table->data);
2790        ret = proc_dostring(table, write, buffer, length, ppos);
2791        if (ret)
2792                goto out;
2793        if (write) {
2794                int oldval = user_zonelist_order;
2795                if (__parse_numa_zonelist_order((char*)table->data)) {
2796                        /*
2797                         * bogus value.  restore saved string
2798                         */
2799                        strncpy((char*)table->data, saved_string,
2800                                NUMA_ZONELIST_ORDER_LEN);
2801                        user_zonelist_order = oldval;
2802                } else if (oldval != user_zonelist_order) {
2803                        mutex_lock(&zonelists_mutex);
2804                        build_all_zonelists(NULL);
2805                        mutex_unlock(&zonelists_mutex);
2806                }
2807        }
2808out:
2809        mutex_unlock(&zl_order_mutex);
2810        return ret;
2811}
2812
2813
2814#define MAX_NODE_LOAD (nr_online_nodes)
2815static int node_load[MAX_NUMNODES];
2816
2817/**
2818 * find_next_best_node - find the next node that should appear in a given node's fallback list
2819 * @node: node whose fallback list we're appending
2820 * @used_node_mask: nodemask_t of already used nodes
2821 *
2822 * We use a number of factors to determine which is the next node that should
2823 * appear on a given node's fallback list.  The node should not have appeared
2824 * already in @node's fallback list, and it should be the next closest node
2825 * according to the distance array (which contains arbitrary distance values
2826 * from each node to each node in the system), and should also prefer nodes
2827 * with no CPUs, since presumably they'll have very little allocation pressure
2828 * on them otherwise.
2829 * It returns -1 if no node is found.
2830 */
2831static int find_next_best_node(int node, nodemask_t *used_node_mask)
2832{
2833        int n, val;
2834        int min_val = INT_MAX;
2835        int best_node = -1;
2836        const struct cpumask *tmp = cpumask_of_node(0);
2837
2838        /* Use the local node if we haven't already */
2839        if (!node_isset(node, *used_node_mask)) {
2840                node_set(node, *used_node_mask);
2841                return node;
2842        }
2843
2844        for_each_node_state(n, N_HIGH_MEMORY) {
2845
2846                /* Don't want a node to appear more than once */
2847                if (node_isset(n, *used_node_mask))
2848                        continue;
2849
2850                /* Use the distance array to find the distance */
2851                val = node_distance(node, n);
2852
2853                /* Penalize nodes under us ("prefer the next node") */
2854                val += (n < node);
2855
2856                /* Give preference to headless and unused nodes */
2857                tmp = cpumask_of_node(n);
2858                if (!cpumask_empty(tmp))
2859                        val += PENALTY_FOR_NODE_WITH_CPUS;
2860
2861                /* Slight preference for less loaded node */
2862                val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2863                val += node_load[n];
2864
2865                if (val < min_val) {
2866                        min_val = val;
2867                        best_node = n;
2868                }
2869        }
2870
2871        if (best_node >= 0)
2872                node_set(best_node, *used_node_mask);
2873
2874        return best_node;
2875}
2876
2877
2878/*
2879 * Build zonelists ordered by node and zones within node.
2880 * This results in maximum locality--normal zone overflows into local
2881 * DMA zone, if any--but risks exhausting DMA zone.
2882 */
2883static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2884{
2885        int j;
2886        struct zonelist *zonelist;
2887
2888        zonelist = &pgdat->node_zonelists[0];
2889        for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2890                ;
2891        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2892                                                        MAX_NR_ZONES - 1);
2893        zonelist->_zonerefs[j].zone = NULL;
2894        zonelist->_zonerefs[j].zone_idx = 0;
2895}
2896
2897/*
2898 * Build gfp_thisnode zonelists
2899 */
2900static void build_thisnode_zonelists(pg_data_t *pgdat)
2901{
2902        int j;
2903        struct zonelist *zonelist;
2904
2905        zonelist = &pgdat->node_zonelists[1];
2906        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2907        zonelist->_zonerefs[j].zone = NULL;
2908        zonelist->_zonerefs[j].zone_idx = 0;
2909}
2910
2911/*
2912 * Build zonelists ordered by zone and nodes within zones.
2913 * This results in conserving DMA zone[s] until all Normal memory is
2914 * exhausted, but results in overflowing to remote node while memory
2915 * may still exist in local DMA zone.
2916 */
2917static int node_order[MAX_NUMNODES];
2918
2919static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2920{
2921        int pos, j, node;
2922        int zone_type;          /* needs to be signed */
2923        struct zone *z;
2924        struct zonelist *zonelist;
2925
2926        zonelist = &pgdat->node_zonelists[0];
2927        pos = 0;
2928        for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2929                for (j = 0; j < nr_nodes; j++) {
2930                        node = node_order[j];
2931                        z = &NODE_DATA(node)->node_zones[zone_type];
2932                        if (populated_zone(z)) {
2933                                zoneref_set_zone(z,
2934                                        &zonelist->_zonerefs[pos++]);
2935                                check_highest_zone(zone_type);
2936                        }
2937                }
2938        }
2939        zonelist->_zonerefs[pos].zone = NULL;
2940        zonelist->_zonerefs[pos].zone_idx = 0;
2941}
2942
2943static int default_zonelist_order(void)
2944{
2945        int nid, zone_type;
2946        unsigned long low_kmem_size,total_size;
2947        struct zone *z;
2948        int average_size;
2949        /*
2950         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2951         * If they are really small and used heavily, the system can fall
2952         * into OOM very easily.
2953         * This function detect ZONE_DMA/DMA32 size and configures zone order.
2954         */
2955        /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2956        low_kmem_size = 0;
2957        total_size = 0;
2958        for_each_online_node(nid) {
2959                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2960                        z = &NODE_DATA(nid)->node_zones[zone_type];
2961                        if (populated_zone(z)) {
2962                                if (zone_type < ZONE_NORMAL)
2963                                        low_kmem_size += z->present_pages;
2964                                total_size += z->present_pages;
2965                        } else if (zone_type == ZONE_NORMAL) {
2966                                /*
2967                                 * If any node has only lowmem, then node order
2968                                 * is preferred to allow kernel allocations
2969                                 * locally; otherwise, they can easily infringe
2970                                 * on other nodes when there is an abundance of
2971                                 * lowmem available to allocate from.
2972                                 */
2973                                return ZONELIST_ORDER_NODE;
2974                        }
2975                }
2976        }
2977        if (!low_kmem_size ||  /* there are no DMA area. */
2978            low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2979                return ZONELIST_ORDER_NODE;
2980        /*
2981         * look into each node's config.
2982         * If there is a node whose DMA/DMA32 memory is very big area on
2983         * local memory, NODE_ORDER may be suitable.
2984         */
2985        average_size = total_size /
2986                                (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2987        for_each_online_node(nid) {
2988                low_kmem_size = 0;
2989                total_size = 0;
2990                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2991                        z = &NODE_DATA(nid)->node_zones[zone_type];
2992                        if (populated_zone(z)) {
2993                                if (zone_type < ZONE_NORMAL)
2994                                        low_kmem_size += z->present_pages;
2995                                total_size += z->present_pages;
2996                        }
2997                }
2998                if (low_kmem_size &&
2999                    total_size > average_size && /* ignore small node */
3000                    low_kmem_size > total_size * 70/100)
3001                        return ZONELIST_ORDER_NODE;
3002        }
3003        return ZONELIST_ORDER_ZONE;
3004}
3005
3006static void set_zonelist_order(void)
3007{
3008        if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3009                current_zonelist_order = default_zonelist_order();
3010        else
3011                current_zonelist_order = user_zonelist_order;
3012}
3013
3014static void build_zonelists(pg_data_t *pgdat)
3015{
3016        int j, node, load;
3017        enum zone_type i;
3018        nodemask_t used_mask;
3019        int local_node, prev_node;
3020        struct zonelist *zonelist;
3021        int order = current_zonelist_order;
3022
3023        /* initialize zonelists */
3024        for (i = 0; i < MAX_ZONELISTS; i++) {
3025                zonelist = pgdat->node_zonelists + i;
3026                zonelist->_zonerefs[0].zone = NULL;
3027                zonelist->_zonerefs[0].zone_idx = 0;
3028        }
3029
3030        /* NUMA-aware ordering of nodes */
3031        local_node = pgdat->node_id;
3032        load = nr_online_nodes;
3033        prev_node = local_node;
3034        nodes_clear(used_mask);
3035
3036        memset(node_order, 0, sizeof(node_order));
3037        j = 0;
3038
3039        while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3040                int distance = node_distance(local_node, node);
3041
3042                /*
3043                 * If another node is sufficiently far away then it is better
3044                 * to reclaim pages in a zone before going off node.
3045                 */
3046                if (distance > RECLAIM_DISTANCE)
3047                        zone_reclaim_mode = 1;
3048
3049                /*
3050                 * We don't want to pressure a particular node.
3051                 * So adding penalty to the first node in same
3052                 * distance group to make it round-robin.
3053                 */
3054                if (distance != node_distance(local_node, prev_node))
3055                        node_load[node] = load;
3056
3057                prev_node = node;
3058                load--;
3059                if (order == ZONELIST_ORDER_NODE)
3060                        build_zonelists_in_node_order(pgdat, node);
3061                else
3062                        node_order[j++] = node; /* remember order */
3063        }
3064
3065        if (order == ZONELIST_ORDER_ZONE) {
3066                /* calculate node order -- i.e., DMA last! */
3067                build_zonelists_in_zone_order(pgdat, j);
3068        }
3069
3070        build_thisnode_zonelists(pgdat);
3071}
3072
3073/* Construct the zonelist performance cache - see further mmzone.h */
3074static void build_zonelist_cache(pg_data_t *pgdat)
3075{
3076        struct zonelist *zonelist;
3077        struct zonelist_cache *zlc;
3078        struct zoneref *z;
3079
3080        zonelist = &pgdat->node_zonelists[0];
3081        zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3082        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3083        for (z = zonelist->_zonerefs; z->zone; z++)
3084                zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3085}
3086
3087#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3088/*
3089 * Return node id of node used for "local" allocations.
3090 * I.e., first node id of first zone in arg node's generic zonelist.
3091 * Used for initializing percpu 'numa_mem', which is used primarily
3092 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3093 */
3094int local_memory_node(int node)
3095{
3096        struct zone *zone;
3097
3098        (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3099                                   gfp_zone(GFP_KERNEL),
3100                                   NULL,
3101                                   &zone);
3102        return zone->node;
3103}
3104#endif
3105
3106#else   /* CONFIG_NUMA */
3107
3108static void set_zonelist_order(void)
3109{
3110        current_zonelist_order = ZONELIST_ORDER_ZONE;
3111}
3112
3113static void build_zonelists(pg_data_t *pgdat)
3114{
3115        int node, local_node;
3116        enum zone_type j;
3117        struct zonelist *zonelist;
3118
3119        local_node = pgdat->node_id;
3120
3121        zonelist = &pgdat->node_zonelists[0];
3122        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3123
3124        /*
3125         * Now we build the zonelist so that it contains the zones
3126         * of all the other nodes.
3127         * We don't want to pressure a particular node, so when
3128         * building the zones for node N, we make sure that the
3129         * zones coming right after the local ones are those from
3130         * node N+1 (modulo N)
3131         */
3132        for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3133                if (!node_online(node))
3134                        continue;
3135                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3136                                                        MAX_NR_ZONES - 1);
3137        }
3138        for (node = 0; node < local_node; node++) {
3139                if (!node_online(node))
3140                        continue;
3141                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3142                                                        MAX_NR_ZONES - 1);
3143        }
3144
3145        zonelist->_zonerefs[j].zone = NULL;
3146        zonelist->_zonerefs[j].zone_idx = 0;
3147}
3148
3149/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3150static void build_zonelist_cache(pg_data_t *pgdat)
3151{
3152        pgdat->node_zonelists[0].zlcache_ptr = NULL;
3153}
3154
3155#endif  /* CONFIG_NUMA */
3156
3157/*
3158 * Boot pageset table. One per cpu which is going to be used for all
3159 * zones and all nodes. The parameters will be set in such a way
3160 * that an item put on a list will immediately be handed over to
3161 * the buddy list. This is safe since pageset manipulation is done
3162 * with interrupts disabled.
3163 *
3164 * The boot_pagesets must be kept even after bootup is complete for
3165 * unused processors and/or zones. They do play a role for bootstrapping
3166 * hotplugged processors.
3167 *
3168 * zoneinfo_show() and maybe other functions do
3169 * not check if the processor is online before following the pageset pointer.
3170 * Other parts of the kernel may not check if the zone is available.
3171 */
3172static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3173static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3174static void setup_zone_pageset(struct zone *zone);
3175
3176/*
3177 * Global mutex to protect against size modification of zonelists
3178 * as well as to serialize pageset setup for the new populated zone.
3179 */
3180DEFINE_MUTEX(zonelists_mutex);
3181
3182/* return values int ....just for stop_machine() */
3183static __init_refok int __build_all_zonelists(void *data)
3184{
3185        int nid;
3186        int cpu;
3187
3188#ifdef CONFIG_NUMA
3189        memset(node_load, 0, sizeof(node_load));
3190#endif
3191        for_each_online_node(nid) {
3192                pg_data_t *pgdat = NODE_DATA(nid);
3193
3194                build_zonelists(pgdat);
3195                build_zonelist_cache(pgdat);
3196        }
3197
3198        /*
3199         * Initialize the boot_pagesets that are going to be used
3200         * for bootstrapping processors. The real pagesets for
3201         * each zone will be allocated later when the per cpu
3202         * allocator is available.
3203         *
3204         * boot_pagesets are used also for bootstrapping offline
3205         * cpus if the system is already booted because the pagesets
3206         * are needed to initialize allocators on a specific cpu too.
3207         * F.e. the percpu allocator needs the page allocator which
3208         * needs the percpu allocator in order to allocate its pagesets
3209         * (a chicken-egg dilemma).
3210         */
3211        for_each_possible_cpu(cpu) {
3212                setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3213
3214#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3215                /*
3216                 * We now know the "local memory node" for each node--
3217                 * i.e., the node of the first zone in the generic zonelist.
3218                 * Set up numa_mem percpu variable for on-line cpus.  During
3219                 * boot, only the boot cpu should be on-line;  we'll init the
3220                 * secondary cpus' numa_mem as they come on-line.  During
3221                 * node/memory hotplug, we'll fixup all on-line cpus.
3222                 */
3223                if (cpu_online(cpu))
3224                        set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3225#endif
3226        }
3227
3228        return 0;
3229}
3230
3231/*
3232 * Called with zonelists_mutex held always
3233 * unless system_state == SYSTEM_BOOTING.
3234 */
3235void __ref build_all_zonelists(void *data)
3236{
3237        set_zonelist_order();
3238
3239        if (system_state == SYSTEM_BOOTING) {
3240                __build_all_zonelists(NULL);
3241                mminit_verify_zonelist();
3242                cpuset_init_current_mems_allowed();
3243        } else {
3244                /* we have to stop all cpus to guarantee there is no user
3245                   of zonelist */
3246#ifdef CONFIG_MEMORY_HOTPLUG
3247                if (data)
3248                        setup_zone_pageset((struct zone *)data);
3249#endif
3250                stop_machine(__build_all_zonelists, NULL, NULL);
3251                /* cpuset refresh routine should be here */
3252        }
3253        vm_total_pages = nr_free_pagecache_pages();
3254        /*
3255         * Disable grouping by mobility if the number of pages in the
3256         * system is too low to allow the mechanism to work. It would be
3257         * more accurate, but expensive to check per-zone. This check is
3258         * made on memory-hotadd so a system can start with mobility
3259         * disabled and enable it later
3260         */
3261        if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3262                page_group_by_mobility_disabled = 1;
3263        else
3264                page_group_by_mobility_disabled = 0;
3265
3266        printk("Built %i zonelists in %s order, mobility grouping %s.  "
3267                "Total pages: %ld\n",
3268                        nr_online_nodes,
3269                        zonelist_order_name[current_zonelist_order],
3270                        page_group_by_mobility_disabled ? "off" : "on",
3271                        vm_total_pages);
3272#ifdef CONFIG_NUMA
3273        printk("Policy zone: %s\n", zone_names[policy_zone]);
3274#endif
3275}
3276
3277/*
3278 * Helper functions to size the waitqueue hash table.
3279 * Essentially these want to choose hash table sizes sufficiently
3280 * large so that collisions trying to wait on pages are rare.
3281 * But in fact, the number of active page waitqueues on typical
3282 * systems is ridiculously low, less than 200. So this is even
3283 * conservative, even though it seems large.
3284 *
3285 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3286 * waitqueues, i.e. the size of the waitq table given the number of pages.
3287 */
3288#define PAGES_PER_WAITQUEUE     256
3289
3290#ifndef CONFIG_MEMORY_HOTPLUG
3291static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3292{
3293        unsigned long size = 1;
3294
3295        pages /= PAGES_PER_WAITQUEUE;
3296
3297        while (size < pages)
3298                size <<= 1;
3299
3300        /*
3301         * Once we have dozens or even hundreds of threads sleeping
3302         * on IO we've got bigger problems than wait queue collision.
3303         * Limit the size of the wait table to a reasonable size.
3304         */
3305        size = min(size, 4096UL);
3306
3307        return max(size, 4UL);
3308}
3309#else
3310/*
3311 * A zone's size might be changed by hot-add, so it is not possible to determine
3312 * a suitable size for its wait_table.  So we use the maximum size now.
3313 *
3314 * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
3315 *
3316 *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
3317 *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3318 *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
3319 *
3320 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3321 * or more by the traditional way. (See above).  It equals:
3322 *
3323 *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
3324 *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
3325 *    powerpc (64K page size)             : =  (32G +16M)byte.
3326 */
3327static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3328{
3329        return 4096UL;
3330}
3331#endif
3332
3333/*
3334 * This is an integer logarithm so that shifts can be used later
3335 * to extract the more random high bits from the multiplicative
3336 * hash function before the remainder is taken.
3337 */
3338static inline unsigned long wait_table_bits(unsigned long size)
3339{
3340        return ffz(~size);
3341}
3342
3343#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3344
3345/*
3346 * Check if a pageblock contains reserved pages
3347 */
3348static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3349{
3350        unsigned long pfn;
3351
3352        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3353                if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3354                        return 1;
3355        }
3356        return 0;
3357}
3358
3359/*
3360 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3361 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3362 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3363 * higher will lead to a bigger reserve which will get freed as contiguous
3364 * blocks as reclaim kicks in
3365 */
3366static void setup_zone_migrate_reserve(struct zone *zone)
3367{
3368        unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3369        struct page *page;
3370        unsigned long block_migratetype;
3371        int reserve;
3372
3373        /* Get the start pfn, end pfn and the number of blocks to reserve */
3374        start_pfn = zone->zone_start_pfn;
3375        end_pfn = start_pfn + zone->spanned_pages;
3376        reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3377                                                        pageblock_order;
3378
3379        /*
3380         * Reserve blocks are generally in place to help high-order atomic
3381         * allocations that are short-lived. A min_free_kbytes value that
3382         * would result in more than 2 reserve blocks for atomic allocations
3383         * is assumed to be in place to help anti-fragmentation for the
3384         * future allocation of hugepages at runtime.
3385         */
3386        reserve = min(2, reserve);
3387
3388        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3389                if (!pfn_valid(pfn))
3390                        continue;
3391                page = pfn_to_page(pfn);
3392
3393                /* Watch out for overlapping nodes */
3394                if (page_to_nid(page) != zone_to_nid(zone))
3395                        continue;
3396
3397                /* Blocks with reserved pages will never free, skip them. */
3398                block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3399                if (pageblock_is_reserved(pfn, block_end_pfn))
3400                        continue;
3401
3402                block_migratetype = get_pageblock_migratetype(page);
3403
3404                /* If this block is reserved, account for it */
3405                if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3406                        reserve--;
3407                        continue;
3408                }
3409
3410                /* Suitable for reserving if this block is movable */
3411                if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3412                        set_pageblock_migratetype(page, MIGRATE_RESERVE);
3413                        move_freepages_block(zone, page, MIGRATE_RESERVE);
3414                        reserve--;
3415                        continue;
3416                }
3417
3418                /*
3419                 * If the reserve is met and this is a previous reserved block,
3420                 * take it back
3421                 */
3422                if (block_migratetype == MIGRATE_RESERVE) {
3423                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3424                        move_freepages_block(zone, page, MIGRATE_MOVABLE);
3425                }
3426        }
3427}
3428
3429/*
3430 * Initially all pages are reserved - free ones are freed
3431 * up by free_all_bootmem() once the early boot process is
3432 * done. Non-atomic initialization, single-pass.
3433 */
3434void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3435                unsigned long start_pfn, enum memmap_context context)
3436{
3437        struct page *page;
3438        unsigned long end_pfn = start_pfn + size;
3439        unsigned long pfn;
3440        struct zone *z;
3441
3442        if (highest_memmap_pfn < end_pfn - 1)
3443                highest_memmap_pfn = end_pfn - 1;
3444
3445        z = &NODE_DATA(nid)->node_zones[zone];
3446        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3447                /*
3448                 * There can be holes in boot-time mem_map[]s
3449                 * handed to this function.  They do not
3450                 * exist on hotplugged memory.
3451                 */
3452                if (context == MEMMAP_EARLY) {
3453                        if (!early_pfn_valid(pfn))
3454                                continue;
3455                        if (!early_pfn_in_nid(pfn, nid))
3456                                continue;
3457                }
3458                page = pfn_to_page(pfn);
3459                set_page_links(page, zone, nid, pfn);
3460                mminit_verify_page_links(page, zone, nid, pfn);
3461                init_page_count(page);
3462                reset_page_mapcount(page);
3463                SetPageReserved(page);
3464                /*
3465                 * Mark the block movable so that blocks are reserved for
3466                 * movable at startup. This will force kernel allocations
3467                 * to reserve their blocks rather than leaking throughout
3468                 * the address space during boot when many long-lived
3469                 * kernel allocations are made. Later some blocks near
3470                 * the start are marked MIGRATE_RESERVE by
3471                 * setup_zone_migrate_reserve()
3472                 *
3473                 * bitmap is created for zone's valid pfn range. but memmap
3474                 * can be created for invalid pages (for alignment)
3475                 * check here not to call set_pageblock_migratetype() against
3476                 * pfn out of zone.
3477                 */
3478                if ((z->zone_start_pfn <= pfn)
3479                    && (pfn < z->zone_start_pfn + z->spanned_pages)
3480                    && !(pfn & (pageblock_nr_pages - 1)))
3481                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3482
3483                INIT_LIST_HEAD(&page->lru);
3484#ifdef WANT_PAGE_VIRTUAL
3485                /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3486                if (!is_highmem_idx(zone))
3487                        set_page_address(page, __va(pfn << PAGE_SHIFT));
3488#endif
3489        }
3490}
3491
3492static void __meminit zone_init_free_lists(struct zone *zone)
3493{
3494        int order, t;
3495        for_each_migratetype_order(order, t) {
3496                INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3497                zone->free_area[order].nr_free = 0;
3498        }
3499}
3500
3501#ifndef __HAVE_ARCH_MEMMAP_INIT
3502#define memmap_init(size, nid, zone, start_pfn) \
3503        memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3504#endif
3505
3506static int zone_batchsize(struct zone *zone)
3507{
3508#ifdef CONFIG_MMU
3509        int batch;
3510
3511        /*
3512         * The per-cpu-pages pools are set to around 1000th of the
3513         * size of the zone.  But no more than 1/2 of a meg.
3514         *
3515         * OK, so we don't know how big the cache is.  So guess.
3516         */
3517        batch = zone->present_pages / 1024;
3518        if (batch * PAGE_SIZE > 512 * 1024)
3519                batch = (512 * 1024) / PAGE_SIZE;
3520        batch /= 4;             /* We effectively *= 4 below */
3521        if (batch < 1)
3522                batch = 1;
3523
3524        /*
3525         * Clamp the batch to a 2^n - 1 value. Having a power
3526         * of 2 value was found to be more likely to have
3527         * suboptimal cache aliasing properties in some cases.
3528         *
3529         * For example if 2 tasks are alternately allocating
3530         * batches of pages, one task can end up with a lot
3531         * of pages of one half of the possible page colors
3532         * and the other with pages of the other colors.
3533         */
3534        batch = rounddown_pow_of_two(batch + batch/2) - 1;
3535
3536        return batch;
3537
3538#else
3539        /* The deferral and batching of frees should be suppressed under NOMMU
3540         * conditions.
3541         *
3542         * The problem is that NOMMU needs to be able to allocate large chunks
3543         * of contiguous memory as there's no hardware page translation to
3544         * assemble apparent contiguous memory from discontiguous pages.
3545         *
3546         * Queueing large contiguous runs of pages for batching, however,
3547         * causes the pages to actually be freed in smaller chunks.  As there
3548         * can be a significant delay between the individual batches being
3549         * recycled, this leads to the once large chunks of space being
3550         * fragmented and becoming unavailable for high-order allocations.
3551         */
3552        return 0;
3553#endif
3554}
3555
3556static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3557{
3558        struct per_cpu_pages *pcp;
3559        int migratetype;
3560
3561        memset(p, 0, sizeof(*p));
3562
3563        pcp = &p->pcp;
3564        pcp->count = 0;
3565        pcp->high = 6 * batch;
3566        pcp->batch = max(1UL, 1 * batch);
3567        for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3568                INIT_LIST_HEAD(&pcp->lists[migratetype]);
3569}
3570
3571/*
3572 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3573 * to the value high for the pageset p.
3574 */
3575
3576static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3577                                unsigned long high)
3578{
3579        struct per_cpu_pages *pcp;
3580
3581        pcp = &p->pcp;
3582        pcp->high = high;
3583        pcp->batch = max(1UL, high/4);
3584        if ((high/4) > (PAGE_SHIFT * 8))
3585                pcp->batch = PAGE_SHIFT * 8;
3586}
3587
3588static void setup_zone_pageset(struct zone *zone)
3589{
3590        int cpu;
3591
3592        zone->pageset = alloc_percpu(struct per_cpu_pageset);
3593
3594        for_each_possible_cpu(cpu) {
3595                struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3596
3597                setup_pageset(pcp, zone_batchsize(zone));
3598
3599                if (percpu_pagelist_fraction)
3600                        setup_pagelist_highmark(pcp,
3601                                (zone->present_pages /
3602                                        percpu_pagelist_fraction));
3603        }
3604}
3605
3606/*
3607 * Allocate per cpu pagesets and initialize them.
3608 * Before this call only boot pagesets were available.
3609 */
3610void __init setup_per_cpu_pageset(void)
3611{
3612        struct zone *zone;
3613
3614        for_each_populated_zone(zone)
3615                setup_zone_pageset(zone);
3616}
3617
3618static noinline __init_refok
3619int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3620{
3621        int i;
3622        struct pglist_data *pgdat = zone->zone_pgdat;
3623        size_t alloc_size;
3624
3625        /*
3626         * The per-page waitqueue mechanism uses hashed waitqueues
3627         * per zone.
3628         */
3629        zone->wait_table_hash_nr_entries =
3630                 wait_table_hash_nr_entries(zone_size_pages);
3631        zone->wait_table_bits =
3632                wait_table_bits(zone->wait_table_hash_nr_entries);
3633        alloc_size = zone->wait_table_hash_nr_entries
3634                                        * sizeof(wait_queue_head_t);
3635
3636        if (!slab_is_available()) {
3637                zone->wait_table = (wait_queue_head_t *)
3638                        alloc_bootmem_node_nopanic(pgdat, alloc_size);
3639        } else {
3640                /*
3641                 * This case means that a zone whose size was 0 gets new memory
3642                 * via memory hot-add.
3643                 * But it may be the case that a new node was hot-added.  In
3644                 * this case vmalloc() will not be able to use this new node's
3645                 * memory - this wait_table must be initialized to use this new
3646                 * node itself as well.
3647                 * To use this new node's memory, further consideration will be
3648                 * necessary.
3649                 */
3650                zone->wait_table = vmalloc(alloc_size);
3651        }
3652        if (!zone->wait_table)
3653                return -ENOMEM;
3654
3655        for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3656                init_waitqueue_head(zone->wait_table + i);
3657
3658        return 0;
3659}
3660
3661static int __zone_pcp_update(void *data)
3662{
3663        struct zone *zone = data;
3664        int cpu;
3665        unsigned long batch = zone_batchsize(zone), flags;
3666
3667        for_each_possible_cpu(cpu) {
3668                struct per_cpu_pageset *pset;
3669                struct per_cpu_pages *pcp;
3670
3671                pset = per_cpu_ptr(zone->pageset, cpu);
3672                pcp = &pset->pcp;
3673
3674                local_irq_save(flags);
3675                free_pcppages_bulk(zone, pcp->count, pcp);
3676                setup_pageset(pset, batch);
3677                local_irq_restore(flags);
3678        }
3679        return 0;
3680}
3681
3682void zone_pcp_update(struct zone *zone)
3683{
3684        stop_machine(__zone_pcp_update, zone, NULL);
3685}
3686
3687static __meminit void zone_pcp_init(struct zone *zone)
3688{
3689        /*
3690         * per cpu subsystem is not up at this point. The following code
3691         * relies on the ability of the linker to provide the
3692         * offset of a (static) per cpu variable into the per cpu area.
3693         */
3694        zone->pageset = &boot_pageset;
3695
3696        if (zone->present_pages)
3697                printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
3698                        zone->name, zone->present_pages,
3699                                         zone_batchsize(zone));
3700}
3701
3702__meminit int init_currently_empty_zone(struct zone *zone,
3703                                        unsigned long zone_start_pfn,
3704                                        unsigned long size,
3705                                        enum memmap_context context)
3706{
3707        struct pglist_data *pgdat = zone->zone_pgdat;
3708        int ret;
3709        ret = zone_wait_table_init(zone, size);
3710        if (ret)
3711                return ret;
3712        pgdat->nr_zones = zone_idx(zone) + 1;
3713
3714        zone->zone_start_pfn = zone_start_pfn;
3715
3716        mminit_dprintk(MMINIT_TRACE, "memmap_init",
3717                        "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3718                        pgdat->node_id,
3719                        (unsigned long)zone_idx(zone),
3720                        zone_start_pfn, (zone_start_pfn + size));
3721
3722        zone_init_free_lists(zone);
3723
3724        return 0;
3725}
3726
3727#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3728/*
3729 * Basic iterator support. Return the first range of PFNs for a node
3730 * Note: nid == MAX_NUMNODES returns first region regardless of node
3731 */
3732static int __meminit first_active_region_index_in_nid(int nid)
3733{
3734        int i;
3735
3736        for (i = 0; i < nr_nodemap_entries; i++)
3737                if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3738                        return i;
3739
3740        return -1;
3741}
3742
3743/*
3744 * Basic iterator support. Return the next active range of PFNs for a node
3745 * Note: nid == MAX_NUMNODES returns next region regardless of node
3746 */
3747static int __meminit next_active_region_index_in_nid(int index, int nid)
3748{
3749        for (index = index + 1; index < nr_nodemap_entries; index++)
3750                if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3751                        return index;
3752
3753        return -1;
3754}
3755
3756#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3757/*
3758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3759 * Architectures may implement their own version but if add_active_range()
3760 * was used and there are no special requirements, this is a convenient
3761 * alternative
3762 */
3763int __meminit __early_pfn_to_nid(unsigned long pfn)
3764{
3765        int i;
3766
3767        for (i = 0; i < nr_nodemap_entries; i++) {
3768                unsigned long start_pfn = early_node_map[i].start_pfn;
3769                unsigned long end_pfn = early_node_map[i].end_pfn;
3770
3771                if (start_pfn <= pfn && pfn < end_pfn)
3772                        return early_node_map[i].nid;
3773        }
3774        /* This is a memory hole */
3775        return -1;
3776}
3777#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3778
3779int __meminit early_pfn_to_nid(unsigned long pfn)
3780{
3781        int nid;
3782
3783        nid = __early_pfn_to_nid(pfn);
3784        if (nid >= 0)
3785                return nid;
3786        /* just returns 0 */
3787        return 0;
3788}
3789
3790#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3791bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3792{
3793        int nid;
3794
3795        nid = __early_pfn_to_nid(pfn);
3796        if (nid >= 0 && nid != node)
3797                return false;
3798        return true;
3799}
3800#endif
3801
3802/* Basic iterator support to walk early_node_map[] */
3803#define for_each_active_range_index_in_nid(i, nid) \
3804        for (i = first_active_region_index_in_nid(nid); i != -1; \
3805                                i = next_active_region_index_in_nid(i, nid))
3806
3807/**
3808 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3809 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3810 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3811 *
3812 * If an architecture guarantees that all ranges registered with
3813 * add_active_ranges() contain no holes and may be freed, this
3814 * this function may be used instead of calling free_bootmem() manually.
3815 */
3816void __init free_bootmem_with_active_regions(int nid,
3817                                                unsigned long max_low_pfn)
3818{
3819        int i;
3820
3821        for_each_active_range_index_in_nid(i, nid) {
3822                unsigned long size_pages = 0;
3823                unsigned long end_pfn = early_node_map[i].end_pfn;
3824
3825                if (early_node_map[i].start_pfn >= max_low_pfn)
3826                        continue;
3827
3828                if (end_pfn > max_low_pfn)
3829                        end_pfn = max_low_pfn;
3830
3831                size_pages = end_pfn - early_node_map[i].start_pfn;
3832                free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3833                                PFN_PHYS(early_node_map[i].start_pfn),
3834                                size_pages << PAGE_SHIFT);
3835        }
3836}
3837
3838#ifdef CONFIG_HAVE_MEMBLOCK
3839/*
3840 * Basic iterator support. Return the last range of PFNs for a node
3841 * Note: nid == MAX_NUMNODES returns last region regardless of node
3842 */
3843static int __meminit last_active_region_index_in_nid(int nid)
3844{
3845        int i;
3846
3847        for (i = nr_nodemap_entries - 1; i >= 0; i--)
3848                if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3849                        return i;
3850
3851        return -1;
3852}
3853
3854/*
3855 * Basic iterator support. Return the previous active range of PFNs for a node
3856 * Note: nid == MAX_NUMNODES returns next region regardless of node
3857 */
3858static int __meminit previous_active_region_index_in_nid(int index, int nid)
3859{
3860        for (index = index - 1; index >= 0; index--)
3861                if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3862                        return index;
3863
3864        return -1;
3865}
3866
3867#define for_each_active_range_index_in_nid_reverse(i, nid) \
3868        for (i = last_active_region_index_in_nid(nid); i != -1; \
3869                                i = previous_active_region_index_in_nid(i, nid))
3870
3871u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3872                                        u64 goal, u64 limit)
3873{
3874        int i;
3875
3876        /* Need to go over early_node_map to find out good range for node */
3877        for_each_active_range_index_in_nid_reverse(i, nid) {
3878                u64 addr;
3879                u64 ei_start, ei_last;
3880                u64 final_start, final_end;
3881
3882                ei_last = early_node_map[i].end_pfn;
3883                ei_last <<= PAGE_SHIFT;
3884                ei_start = early_node_map[i].start_pfn;
3885                ei_start <<= PAGE_SHIFT;
3886
3887                final_start = max(ei_start, goal);
3888                final_end = min(ei_last, limit);
3889
3890                if (final_start >= final_end)
3891                        continue;
3892
3893                addr = memblock_find_in_range(final_start, final_end, size, align);
3894
3895                if (addr == MEMBLOCK_ERROR)
3896                        continue;
3897
3898                return addr;
3899        }
3900
3901        return MEMBLOCK_ERROR;
3902}
3903#endif
3904
3905int __init add_from_early_node_map(struct range *range, int az,
3906                                   int nr_range, int nid)
3907{
3908        int i;
3909        u64 start, end;
3910
3911        /* need to go over early_node_map to find out good range for node */
3912        for_each_active_range_index_in_nid(i, nid) {
3913                start = early_node_map[i].start_pfn;
3914                end = early_node_map[i].end_pfn;
3915                nr_range = add_range(range, az, nr_range, start, end);
3916        }
3917        return nr_range;
3918}
3919
3920void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3921{
3922        int i;
3923        int ret;
3924
3925        for_each_active_range_index_in_nid(i, nid) {
3926                ret = work_fn(early_node_map[i].start_pfn,
3927                              early_node_map[i].end_pfn, data);
3928                if (ret)
3929                        break;
3930        }
3931}
3932/**
3933 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3934 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3935 *
3936 * If an architecture guarantees that all ranges registered with
3937 * add_active_ranges() contain no holes and may be freed, this
3938 * function may be used instead of calling memory_present() manually.
3939 */
3940void __init sparse_memory_present_with_active_regions(int nid)
3941{
3942        int i;
3943
3944        for_each_active_range_index_in_nid(i, nid)
3945                memory_present(early_node_map[i].nid,
3946                                early_node_map[i].start_pfn,
3947                                early_node_map[i].end_pfn);
3948}
3949
3950/**
3951 * get_pfn_range_for_nid - Return the start and end page frames for a node
3952 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3953 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3954 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3955 *
3956 * It returns the start and end page frame of a node based on information
3957 * provided by an arch calling add_active_range(). If called for a node
3958 * with no available memory, a warning is printed and the start and end
3959 * PFNs will be 0.
3960 */
3961void __meminit get_pfn_range_for_nid(unsigned int nid,
3962                        unsigned long *start_pfn, unsigned long *end_pfn)
3963{
3964        int i;
3965        *start_pfn = -1UL;
3966        *end_pfn = 0;
3967
3968        for_each_active_range_index_in_nid(i, nid) {
3969                *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3970                *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3971        }
3972
3973        if (*start_pfn == -1UL)
3974                *start_pfn = 0;
3975}
3976
3977/*
3978 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3979 * assumption is made that zones within a node are ordered in monotonic
3980 * increasing memory addresses so that the "highest" populated zone is used
3981 */
3982static void __init find_usable_zone_for_movable(void)
3983{
3984        int zone_index;
3985        for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3986                if (zone_index == ZONE_MOVABLE)
3987                        continue;
3988
3989                if (arch_zone_highest_possible_pfn[zone_index] >
3990                                arch_zone_lowest_possible_pfn[zone_index])
3991                        break;
3992        }
3993
3994        VM_BUG_ON(zone_index == -1);
3995        movable_zone = zone_index;
3996}
3997
3998/*
3999 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4000 * because it is sized independent of architecture. Unlike the other zones,
4001 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4002 * in each node depending on the size of each node and how evenly kernelcore
4003 * is distributed. This helper function adjusts the zone ranges
4004 * provided by the architecture for a given node by using the end of the
4005 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4006 * zones within a node are in order of monotonic increases memory addresses
4007 */
4008static void __meminit adjust_zone_range_for_zone_movable(int nid,
4009                                        unsigned long zone_type,
4010                                        unsigned long node_start_pfn,
4011                                        unsigned long node_end_pfn,
4012                                        unsigned long *zone_start_pfn,
4013                                        unsigned long *zone_end_pfn)
4014{
4015        /* Only adjust if ZONE_MOVABLE is on this node */
4016        if (zone_movable_pfn[nid]) {
4017                /* Size ZONE_MOVABLE */
4018                if (zone_type == ZONE_MOVABLE) {
4019                        *zone_start_pfn = zone_movable_pfn[nid];
4020                        *zone_end_pfn = min(node_end_pfn,
4021                                arch_zone_highest_possible_pfn[movable_zone]);
4022
4023                /* Adjust for ZONE_MOVABLE starting within this range */
4024                } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4025                                *zone_end_pfn > zone_movable_pfn[nid]) {
4026                        *zone_end_pfn = zone_movable_pfn[nid];
4027
4028                /* Check if this whole range is within ZONE_MOVABLE */
4029                } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4030                        *zone_start_pfn = *zone_end_pfn;
4031        }
4032}
4033
4034/*
4035 * Return the number of pages a zone spans in a node, including holes
4036 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4037 */
4038static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4039                                        unsigned long zone_type,
4040                                        unsigned long *ignored)
4041{
4042        unsigned long node_start_pfn, node_end_pfn;
4043        unsigned long zone_start_pfn, zone_end_pfn;
4044
4045        /* Get the start and end of the node and zone */
4046        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4047        zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4048        zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4049        adjust_zone_range_for_zone_movable(nid, zone_type,
4050                                node_start_pfn, node_end_pfn,
4051                                &zone_start_pfn, &zone_end_pfn);
4052
4053        /* Check that this node has pages within the zone's required range */
4054        if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4055                return 0;
4056
4057        /* Move the zone boundaries inside the node if necessary */
4058        zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4059        zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4060
4061        /* Return the spanned pages */
4062        return zone_end_pfn - zone_start_pfn;
4063}
4064
4065/*
4066 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4067 * then all holes in the requested range will be accounted for.
4068 */
4069unsigned long __meminit __absent_pages_in_range(int nid,
4070                                unsigned long range_start_pfn,
4071                                unsigned long range_end_pfn)
4072{
4073        int i = 0;
4074        unsigned long prev_end_pfn = 0, hole_pages = 0;
4075        unsigned long start_pfn;
4076
4077        /* Find the end_pfn of the first active range of pfns in the node */
4078        i = first_active_region_index_in_nid(nid);
4079        if (i == -1)
4080                return 0;
4081
4082        prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4083
4084        /* Account for ranges before physical memory on this node */
4085        if (early_node_map[i].start_pfn > range_start_pfn)
4086                hole_pages = prev_end_pfn - range_start_pfn;
4087
4088        /* Find all holes for the zone within the node */
4089        for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4090
4091                /* No need to continue if prev_end_pfn is outside the zone */
4092                if (prev_end_pfn >= range_end_pfn)
4093                        break;
4094
4095                /* Make sure the end of the zone is not within the hole */
4096                start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4097                prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4098
4099                /* Update the hole size cound and move on */
4100                if (start_pfn > range_start_pfn) {
4101                        BUG_ON(prev_end_pfn > start_pfn);
4102                        hole_pages += start_pfn - prev_end_pfn;
4103                }
4104                prev_end_pfn = early_node_map[i].end_pfn;
4105        }
4106
4107        /* Account for ranges past physical memory on this node */
4108        if (range_end_pfn > prev_end_pfn)
4109                hole_pages += range_end_pfn -
4110                                max(range_start_pfn, prev_end_pfn);
4111
4112        return hole_pages;
4113}
4114
4115/**
4116 * absent_pages_in_range - Return number of page frames in holes within a range
4117 * @start_pfn: The start PFN to start searching for holes
4118 * @end_pfn: The end PFN to stop searching for holes
4119 *
4120 * It returns the number of pages frames in memory holes within a range.
4121 */
4122unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4123                                                        unsigned long end_pfn)
4124{
4125        return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4126}
4127
4128/* Return the number of page frames in holes in a zone on a node */
4129static unsigned long __meminit zone_absent_pages_in_node(int nid,
4130                                        unsigned long zone_type,
4131                                        unsigned long *ignored)
4132{
4133        unsigned long node_start_pfn, node_end_pfn;
4134        unsigned long zone_start_pfn, zone_end_pfn;
4135
4136        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4137        zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4138                                                        node_start_pfn);
4139        zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4140                                                        node_end_pfn);
4141
4142        adjust_zone_range_for_zone_movable(nid, zone_type,
4143                        node_start_pfn, node_end_pfn,
4144                        &zone_start_pfn, &zone_end_pfn);
4145        return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4146}
4147
4148#else
4149static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4150                                        unsigned long zone_type,
4151                                        unsigned long *zones_size)
4152{
4153        return zones_size[zone_type];
4154}
4155
4156static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4157                                                unsigned long zone_type,
4158                                                unsigned long *zholes_size)
4159{
4160        if (!zholes_size)
4161                return 0;
4162
4163        return zholes_size[zone_type];
4164}
4165
4166#endif
4167
4168static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4169                unsigned long *zones_size, unsigned long *zholes_size)
4170{
4171        unsigned long realtotalpages, totalpages = 0;
4172        enum zone_type i;
4173
4174        for (i = 0; i < MAX_NR_ZONES; i++)
4175                totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4176                                                                zones_size);
4177        pgdat->node_spanned_pages = totalpages;
4178
4179        realtotalpages = totalpages;
4180        for (i = 0; i < MAX_NR_ZONES; i++)
4181                realtotalpages -=
4182                        zone_absent_pages_in_node(pgdat->node_id, i,
4183                                                                zholes_size);
4184        pgdat->node_present_pages = realtotalpages;
4185        printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4186                                                        realtotalpages);
4187}
4188
4189#ifndef CONFIG_SPARSEMEM
4190/*
4191 * Calculate the size of the zone->blockflags rounded to an unsigned long
4192 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4193 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4194 * round what is now in bits to nearest long in bits, then return it in
4195 * bytes.
4196 */
4197static unsigned long __init usemap_size(unsigned long zonesize)
4198{
4199        unsigned long usemapsize;
4200
4201        usemapsize = roundup(zonesize, pageblock_nr_pages);
4202        usemapsize = usemapsize >> pageblock_order;
4203        usemapsize *= NR_PAGEBLOCK_BITS;
4204        usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4205
4206        return usemapsize / 8;
4207}
4208
4209static void __init setup_usemap(struct pglist_data *pgdat,
4210                                struct zone *zone, unsigned long zonesize)
4211{
4212        unsigned long usemapsize = usemap_size(zonesize);
4213        zone->pageblock_flags = NULL;
4214        if (usemapsize)
4215                zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4216                                                                   usemapsize);
4217}
4218#else
4219static inline void setup_usemap(struct pglist_data *pgdat,
4220                                struct zone *zone, unsigned long zonesize) {}
4221#endif /* CONFIG_SPARSEMEM */
4222
4223#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4224
4225/* Return a sensible default order for the pageblock size. */
4226static inline int pageblock_default_order(void)
4227{
4228        if (HPAGE_SHIFT > PAGE_SHIFT)
4229                return HUGETLB_PAGE_ORDER;
4230
4231        return MAX_ORDER-1;
4232}
4233
4234/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4235static inline void __init set_pageblock_order(unsigned int order)
4236{
4237        /* Check that pageblock_nr_pages has not already been setup */
4238        if (pageblock_order)
4239                return;
4240
4241        /*
4242         * Assume the largest contiguous order of interest is a huge page.
4243         * This value may be variable depending on boot parameters on IA64
4244         */
4245        pageblock_order = order;
4246}
4247#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4248
4249/*
4250 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4251 * and pageblock_default_order() are unused as pageblock_order is set
4252 * at compile-time. See include/linux/pageblock-flags.h for the values of
4253 * pageblock_order based on the kernel config
4254 */
4255static inline int pageblock_default_order(unsigned int order)
4256{
4257        return MAX_ORDER-1;
4258}
4259#define set_pageblock_order(x)  do {} while (0)
4260
4261#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4262
4263/*
4264 * Set up the zone data structures:
4265 *   - mark all pages reserved
4266 *   - mark all memory queues empty
4267 *   - clear the memory bitmaps
4268 */
4269static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4270                unsigned long *zones_size, unsigned long *zholes_size)
4271{
4272        enum zone_type j;
4273        int nid = pgdat->node_id;
4274        unsigned long zone_start_pfn = pgdat->node_start_pfn;
4275        int ret;
4276
4277        pgdat_resize_init(pgdat);
4278        pgdat->nr_zones = 0;
4279        init_waitqueue_head(&pgdat->kswapd_wait);
4280        pgdat->kswapd_max_order = 0;
4281        pgdat_page_cgroup_init(pgdat);
4282        
4283        for (j = 0; j < MAX_NR_ZONES; j++) {
4284                struct zone *zone = pgdat->node_zones + j;
4285                unsigned long size, realsize, memmap_pages;
4286                enum lru_list l;
4287
4288                size = zone_spanned_pages_in_node(nid, j, zones_size);
4289                realsize = size - zone_absent_pages_in_node(nid, j,
4290                                                                zholes_size);
4291
4292                /*
4293                 * Adjust realsize so that it accounts for how much memory
4294                 * is used by this zone for memmap. This affects the watermark
4295                 * and per-cpu initialisations
4296                 */
4297                memmap_pages =
4298                        PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4299                if (realsize >= memmap_pages) {
4300                        realsize -= memmap_pages;
4301                        if (memmap_pages)
4302                                printk(KERN_DEBUG
4303                                       "  %s zone: %lu pages used for memmap\n",
4304                                       zone_names[j], memmap_pages);
4305                } else
4306                        printk(KERN_WARNING
4307                                "  %s zone: %lu pages exceeds realsize %lu\n",
4308                                zone_names[j], memmap_pages, realsize);
4309
4310                /* Account for reserved pages */
4311                if (j == 0 && realsize > dma_reserve) {
4312                        realsize -= dma_reserve;
4313                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
4314                                        zone_names[0], dma_reserve);
4315                }
4316
4317                if (!is_highmem_idx(j))
4318                        nr_kernel_pages += realsize;
4319                nr_all_pages += realsize;
4320
4321                zone->spanned_pages = size;
4322                zone->present_pages = realsize;
4323#ifdef CONFIG_NUMA
4324                zone->node = nid;
4325                zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4326                                                / 100;
4327                zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4328#endif
4329                zone->name = zone_names[j];
4330                spin_lock_init(&zone->lock);
4331                spin_lock_init(&zone->lru_lock);
4332                zone_seqlock_init(zone);
4333                zone->zone_pgdat = pgdat;
4334
4335                zone_pcp_init(zone);
4336                for_each_lru(l)
4337                        INIT_LIST_HEAD(&zone->lru[l].list);
4338                zone->reclaim_stat.recent_rotated[0] = 0;
4339                zone->reclaim_stat.recent_rotated[1] = 0;
4340                zone->reclaim_stat.recent_scanned[0] = 0;
4341                zone->reclaim_stat.recent_scanned[1] = 0;
4342                zap_zone_vm_stats(zone);
4343                zone->flags = 0;
4344                if (!size)
4345                        continue;
4346
4347                set_pageblock_order(pageblock_default_order());
4348                setup_usemap(pgdat, zone, size);
4349                ret = init_currently_empty_zone(zone, zone_start_pfn,
4350                                                size, MEMMAP_EARLY);
4351                BUG_ON(ret);
4352                memmap_init(size, nid, j, zone_start_pfn);
4353                zone_start_pfn += size;
4354        }
4355}
4356
4357static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4358{
4359        /* Skip empty nodes */
4360        if (!pgdat->node_spanned_pages)
4361                return;
4362
4363#ifdef CONFIG_FLAT_NODE_MEM_MAP
4364        /* ia64 gets its own node_mem_map, before this, without bootmem */
4365        if (!pgdat->node_mem_map) {
4366                unsigned long size, start, end;
4367                struct page *map;
4368
4369                /*
4370                 * The zone's endpoints aren't required to be MAX_ORDER
4371                 * aligned but the node_mem_map endpoints must be in order
4372                 * for the buddy allocator to function correctly.
4373                 */
4374                start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4375                end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4376                end = ALIGN(end, MAX_ORDER_NR_PAGES);
4377                size =  (end - start) * sizeof(struct page);
4378                map = alloc_remap(pgdat->node_id, size);
4379                if (!map)
4380                        map = alloc_bootmem_node_nopanic(pgdat, size);
4381                pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4382        }
4383#ifndef CONFIG_NEED_MULTIPLE_NODES
4384        /*
4385         * With no DISCONTIG, the global mem_map is just set as node 0's
4386         */
4387        if (pgdat == NODE_DATA(0)) {
4388                mem_map = NODE_DATA(0)->node_mem_map;
4389#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4390                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4391                        mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4392#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4393        }
4394#endif
4395#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4396}
4397
4398void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4399                unsigned long node_start_pfn, unsigned long *zholes_size)
4400{
4401        pg_data_t *pgdat = NODE_DATA(nid);
4402
4403        pgdat->node_id = nid;
4404        pgdat->node_start_pfn = node_start_pfn;
4405        calculate_node_totalpages(pgdat, zones_size, zholes_size);
4406
4407        alloc_node_mem_map(pgdat);
4408#ifdef CONFIG_FLAT_NODE_MEM_MAP
4409        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4410                nid, (unsigned long)pgdat,
4411                (unsigned long)pgdat->node_mem_map);
4412#endif
4413
4414        free_area_init_core(pgdat, zones_size, zholes_size);
4415}
4416
4417#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4418
4419#if MAX_NUMNODES > 1
4420/*
4421 * Figure out the number of possible node ids.
4422 */
4423static void __init setup_nr_node_ids(void)
4424{
4425        unsigned int node;
4426        unsigned int highest = 0;
4427
4428        for_each_node_mask(node, node_possible_map)
4429                highest = node;
4430        nr_node_ids = highest + 1;
4431}
4432#else
4433static inline void setup_nr_node_ids(void)
4434{
4435}
4436#endif
4437
4438/**
4439 * add_active_range - Register a range of PFNs backed by physical memory
4440 * @nid: The node ID the range resides on
4441 * @start_pfn: The start PFN of the available physical memory
4442 * @end_pfn: The end PFN of the available physical memory
4443 *
4444 * These ranges are stored in an early_node_map[] and later used by
4445 * free_area_init_nodes() to calculate zone sizes and holes. If the
4446 * range spans a memory hole, it is up to the architecture to ensure
4447 * the memory is not freed by the bootmem allocator. If possible
4448 * the range being registered will be merged with existing ranges.
4449 */
4450void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4451                                                unsigned long end_pfn)
4452{
4453        int i;
4454
4455        mminit_dprintk(MMINIT_TRACE, "memory_register",
4456                        "Entering add_active_range(%d, %#lx, %#lx) "
4457                        "%d entries of %d used\n",
4458                        nid, start_pfn, end_pfn,
4459                        nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4460
4461        mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4462
4463        /* Merge with existing active regions if possible */
4464        for (i = 0; i < nr_nodemap_entries; i++) {
4465                if (early_node_map[i].nid != nid)
4466                        continue;
4467
4468                /* Skip if an existing region covers this new one */
4469                if (start_pfn >= early_node_map[i].start_pfn &&
4470                                end_pfn <= early_node_map[i].end_pfn)
4471                        return;
4472
4473                /* Merge forward if suitable */
4474                if (start_pfn <= early_node_map[i].end_pfn &&
4475                                end_pfn > early_node_map[i].end_pfn) {
4476                        early_node_map[i].end_pfn = end_pfn;
4477                        return;
4478                }
4479
4480                /* Merge backward if suitable */
4481                if (start_pfn < early_node_map[i].start_pfn &&
4482                                end_pfn >= early_node_map[i].start_pfn) {
4483                        early_node_map[i].start_pfn = start_pfn;
4484                        return;
4485                }
4486        }
4487
4488        /* Check that early_node_map is large enough */
4489        if (i >= MAX_ACTIVE_REGIONS) {
4490                printk(KERN_CRIT "More than %d memory regions, truncating\n",
4491                                                        MAX_ACTIVE_REGIONS);
4492                return;
4493        }
4494
4495        early_node_map[i].nid = nid;
4496        early_node_map[i].start_pfn = start_pfn;
4497        early_node_map[i].end_pfn = end_pfn;
4498        nr_nodemap_entries = i + 1;
4499}
4500
4501/**
4502 * remove_active_range - Shrink an existing registered range of PFNs
4503 * @nid: The node id the range is on that should be shrunk
4504 * @start_pfn: The new PFN of the range
4505 * @end_pfn: The new PFN of the range
4506 *
4507 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4508 * The map is kept near the end physical page range that has already been
4509 * registered. This function allows an arch to shrink an existing registered
4510 * range.
4511 */
4512void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4513                                unsigned long end_pfn)
4514{
4515        int i, j;
4516        int removed = 0;
4517
4518        printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4519                          nid, start_pfn, end_pfn);
4520
4521        /* Find the old active region end and shrink */
4522        for_each_active_range_index_in_nid(i, nid) {
4523                if (early_node_map[i].start_pfn >= start_pfn &&
4524                    early_node_map[i].end_pfn <= end_pfn) {
4525                        /* clear it */
4526                        early_node_map[i].start_pfn = 0;
4527                        early_node_map[i].end_pfn = 0;
4528                        removed = 1;
4529                        continue;
4530                }
4531                if (early_node_map[i].start_pfn < start_pfn &&
4532                    early_node_map[i].end_pfn > start_pfn) {
4533                        unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4534                        early_node_map[i].end_pfn = start_pfn;
4535                        if (temp_end_pfn > end_pfn)
4536                                add_active_range(nid, end_pfn, temp_end_pfn);
4537                        continue;
4538                }
4539                if (early_node_map[i].start_pfn >= start_pfn &&
4540                    early_node_map[i].end_pfn > end_pfn &&
4541                    early_node_map[i].start_pfn < end_pfn) {
4542                        early_node_map[i].start_pfn = end_pfn;
4543                        continue;
4544                }
4545        }
4546
4547        if (!removed)
4548                return;
4549
4550        /* remove the blank ones */
4551        for (i = nr_nodemap_entries - 1; i > 0; i--) {
4552                if (early_node_map[i].nid != nid)
4553                        continue;
4554                if (early_node_map[i].end_pfn)
4555                        continue;
4556                /* we found it, get rid of it */
4557                for (j = i; j < nr_nodemap_entries - 1; j++)
4558                        memcpy(&early_node_map[j], &early_node_map[j+1],
4559                                sizeof(early_node_map[j]));
4560                j = nr_nodemap_entries - 1;
4561                memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4562                nr_nodemap_entries--;
4563        }
4564}
4565
4566/**
4567 * remove_all_active_ranges - Remove all currently registered regions
4568 *
4569 * During discovery, it may be found that a table like SRAT is invalid
4570 * and an alternative discovery method must be used. This function removes
4571 * all currently registered regions.
4572 */
4573void __init remove_all_active_ranges(void)
4574{
4575        memset(early_node_map, 0, sizeof(early_node_map));
4576        nr_nodemap_entries = 0;
4577}
4578
4579/* Compare two active node_active_regions */
4580static int __init cmp_node_active_region(const void *a, const void *b)
4581{
4582        struct node_active_region *arange = (struct node_active_region *)a;
4583        struct node_active_region *brange = (struct node_active_region *)b;
4584
4585        /* Done this way to avoid overflows */
4586        if (arange->start_pfn > brange->start_pfn)
4587                return 1;
4588        if (arange->start_pfn < brange->start_pfn)
4589                return -1;
4590
4591        return 0;
4592}
4593
4594/* sort the node_map by start_pfn */
4595void __init sort_node_map(void)
4596{
4597        sort(early_node_map, (size_t)nr_nodemap_entries,
4598                        sizeof(struct node_active_region),
4599                        cmp_node_active_region, NULL);
4600}
4601
4602/**
4603 * node_map_pfn_alignment - determine the maximum internode alignment
4604 *
4605 * This function should be called after node map is populated and sorted.
4606 * It calculates the maximum power of two alignment which can distinguish
4607 * all the nodes.
4608 *
4609 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4610 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
4611 * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
4612 * shifted, 1GiB is enough and this function will indicate so.
4613 *
4614 * This is used to test whether pfn -> nid mapping of the chosen memory
4615 * model has fine enough granularity to avoid incorrect mapping for the
4616 * populated node map.
4617 *
4618 * Returns the determined alignment in pfn's.  0 if there is no alignment
4619 * requirement (single node).
4620 */
4621unsigned long __init node_map_pfn_alignment(void)
4622{
4623        unsigned long accl_mask = 0, last_end = 0;
4624        int last_nid = -1;
4625        int i;
4626
4627        for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4628                int nid = early_node_map[i].nid;
4629                unsigned long start = early_node_map[i].start_pfn;
4630                unsigned long end = early_node_map[i].end_pfn;
4631                unsigned long mask;
4632
4633                if (!start || last_nid < 0 || last_nid == nid) {
4634                        last_nid = nid;
4635                        last_end = end;
4636                        continue;
4637                }
4638
4639                /*
4640                 * Start with a mask granular enough to pin-point to the
4641                 * start pfn and tick off bits one-by-one until it becomes
4642                 * too coarse to separate the current node from the last.
4643                 */
4644                mask = ~((1 << __ffs(start)) - 1);
4645                while (mask && last_end <= (start & (mask << 1)))
4646                        mask <<= 1;
4647
4648                /* accumulate all internode masks */
4649                accl_mask |= mask;
4650        }
4651
4652        /* convert mask to number of pages */
4653        return ~accl_mask + 1;
4654}
4655
4656/* Find the lowest pfn for a node */
4657static unsigned long __init find_min_pfn_for_node(int nid)
4658{
4659        int i;
4660        unsigned long min_pfn = ULONG_MAX;
4661
4662        /* Assuming a sorted map, the first range found has the starting pfn */
4663        for_each_active_range_index_in_nid(i, nid)
4664                min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4665
4666        if (min_pfn == ULONG_MAX) {
4667                printk(KERN_WARNING
4668                        "Could not find start_pfn for node %d\n", nid);
4669                return 0;
4670        }
4671
4672        return min_pfn;
4673}
4674
4675/**
4676 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4677 *
4678 * It returns the minimum PFN based on information provided via
4679 * add_active_range().
4680 */
4681unsigned long __init find_min_pfn_with_active_regions(void)
4682{
4683        return find_min_pfn_for_node(MAX_NUMNODES);
4684}
4685
4686/*
4687 * early_calculate_totalpages()
4688 * Sum pages in active regions for movable zone.
4689 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4690 */
4691static unsigned long __init early_calculate_totalpages(void)
4692{
4693        int i;
4694        unsigned long totalpages = 0;
4695
4696        for (i = 0; i < nr_nodemap_entries; i++) {
4697                unsigned long pages = early_node_map[i].end_pfn -
4698                                                early_node_map[i].start_pfn;
4699                totalpages += pages;
4700                if (pages)
4701                        node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4702        }
4703        return totalpages;
4704}
4705
4706/*
4707 * Find the PFN the Movable zone begins in each node. Kernel memory
4708 * is spread evenly between nodes as long as the nodes have enough
4709 * memory. When they don't, some nodes will have more kernelcore than
4710 * others
4711 */
4712static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4713{
4714        int i, nid;
4715        unsigned long usable_startpfn;
4716        unsigned long kernelcore_node, kernelcore_remaining;
4717        /* save the state before borrow the nodemask */
4718        nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4719        unsigned long totalpages = early_calculate_totalpages();
4720        int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4721
4722        /*
4723         * If movablecore was specified, calculate what size of
4724         * kernelcore that corresponds so that memory usable for
4725         * any allocation type is evenly spread. If both kernelcore
4726         * and movablecore are specified, then the value of kernelcore
4727         * will be used for required_kernelcore if it's greater than
4728         * what movablecore would have allowed.
4729         */
4730        if (required_movablecore) {
4731                unsigned long corepages;
4732
4733                /*
4734                 * Round-up so that ZONE_MOVABLE is at least as large as what
4735                 * was requested by the user
4736                 */
4737                required_movablecore =
4738                        roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4739                corepages = totalpages - required_movablecore;
4740
4741                required_kernelcore = max(required_kernelcore, corepages);
4742        }
4743
4744        /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4745        if (!required_kernelcore)
4746                goto out;
4747
4748        /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4749        find_usable_zone_for_movable();
4750        usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4751
4752restart:
4753        /* Spread kernelcore memory as evenly as possible throughout nodes */
4754        kernelcore_node = required_kernelcore / usable_nodes;
4755        for_each_node_state(nid, N_HIGH_MEMORY) {
4756                /*
4757                 * Recalculate kernelcore_node if the division per node
4758                 * now exceeds what is necessary to satisfy the requested
4759                 * amount of memory for the kernel
4760                 */
4761                if (required_kernelcore < kernelcore_node)
4762                        kernelcore_node = required_kernelcore / usable_nodes;
4763
4764                /*
4765                 * As the map is walked, we track how much memory is usable
4766                 * by the kernel using kernelcore_remaining. When it is
4767                 * 0, the rest of the node is usable by ZONE_MOVABLE
4768                 */
4769                kernelcore_remaining = kernelcore_node;
4770
4771                /* Go through each range of PFNs within this node */
4772                for_each_active_range_index_in_nid(i, nid) {
4773                        unsigned long start_pfn, end_pfn;
4774                        unsigned long size_pages;
4775
4776                        start_pfn = max(early_node_map[i].start_pfn,
4777                                                zone_movable_pfn[nid]);
4778                        end_pfn = early_node_map[i].end_pfn;
4779                        if (start_pfn >= end_pfn)
4780                                continue;
4781
4782                        /* Account for what is only usable for kernelcore */
4783                        if (start_pfn < usable_startpfn) {
4784                                unsigned long kernel_pages;
4785                                kernel_pages = min(end_pfn, usable_startpfn)
4786                                                                - start_pfn;
4787
4788                                kernelcore_remaining -= min(kernel_pages,
4789                                                        kernelcore_remaining);
4790                                required_kernelcore -= min(kernel_pages,
4791                                                        required_kernelcore);
4792
4793                                /* Continue if range is now fully accounted */
4794                                if (end_pfn <= usable_startpfn) {
4795
4796                                        /*
4797                                         * Push zone_movable_pfn to the end so
4798                                         * that if we have to rebalance
4799                                         * kernelcore across nodes, we will
4800                                         * not double account here
4801                                         */
4802                                        zone_movable_pfn[nid] = end_pfn;
4803                                        continue;
4804                                }
4805                                start_pfn = usable_startpfn;
4806                        }
4807
4808                        /*
4809                         * The usable PFN range for ZONE_MOVABLE is from
4810                         * start_pfn->end_pfn. Calculate size_pages as the
4811                         * number of pages used as kernelcore
4812                         */
4813                        size_pages = end_pfn - start_pfn;
4814                        if (size_pages > kernelcore_remaining)
4815                                size_pages = kernelcore_remaining;
4816                        zone_movable_pfn[nid] = start_pfn + size_pages;
4817
4818                        /*
4819                         * Some kernelcore has been met, update counts and
4820                         * break if the kernelcore for this node has been
4821                         * satisified
4822                         */
4823                        required_kernelcore -= min(required_kernelcore,
4824                                                                size_pages);
4825                        kernelcore_remaining -= size_pages;
4826                        if (!kernelcore_remaining)
4827                                break;
4828                }
4829        }
4830
4831        /*
4832         * If there is still required_kernelcore, we do another pass with one
4833         * less node in the count. This will push zone_movable_pfn[nid] further
4834         * along on the nodes that still have memory until kernelcore is
4835         * satisified
4836         */
4837        usable_nodes--;
4838        if (usable_nodes && required_kernelcore > usable_nodes)
4839                goto restart;
4840
4841        /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4842        for (nid = 0; nid < MAX_NUMNODES; nid++)
4843                zone_movable_pfn[nid] =
4844                        roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4845
4846out:
4847        /* restore the node_state */
4848        node_states[N_HIGH_MEMORY] = saved_node_state;
4849}
4850
4851/* Any regular memory on that node ? */
4852static void check_for_regular_memory(pg_data_t *pgdat)
4853{
4854#ifdef CONFIG_HIGHMEM
4855        enum zone_type zone_type;
4856
4857        for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4858                struct zone *zone = &pgdat->node_zones[zone_type];
4859                if (zone->present_pages)
4860                        node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4861        }
4862#endif
4863}
4864
4865/**
4866 * free_area_init_nodes - Initialise all pg_data_t and zone data
4867 * @max_zone_pfn: an array of max PFNs for each zone
4868 *
4869 * This will call free_area_init_node() for each active node in the system.
4870 * Using the page ranges provided by add_active_range(), the size of each
4871 * zone in each node and their holes is calculated. If the maximum PFN
4872 * between two adjacent zones match, it is assumed that the zone is empty.
4873 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4874 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4875 * starts where the previous one ended. For example, ZONE_DMA32 starts
4876 * at arch_max_dma_pfn.
4877 */
4878void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4879{
4880        unsigned long nid;
4881        int i;
4882
4883        /* Sort early_node_map as initialisation assumes it is sorted */
4884        sort_node_map();
4885
4886        /* Record where the zone boundaries are */
4887        memset(arch_zone_lowest_possible_pfn, 0,
4888                                sizeof(arch_zone_lowest_possible_pfn));
4889        memset(arch_zone_highest_possible_pfn, 0,
4890                                sizeof(arch_zone_highest_possible_pfn));
4891        arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4892        arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4893        for (i = 1; i < MAX_NR_ZONES; i++) {
4894                if (i == ZONE_MOVABLE)
4895                        continue;
4896                arch_zone_lowest_possible_pfn[i] =
4897                        arch_zone_highest_possible_pfn[i-1];
4898                arch_zone_highest_possible_pfn[i] =
4899                        max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900        }
4901        arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4902        arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903
4904        /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4905        memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4906        find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4907
4908        /* Print out the zone ranges */
4909        printk("Zone PFN ranges:\n");
4910        for (i = 0; i < MAX_NR_ZONES; i++) {
4911                if (i == ZONE_MOVABLE)
4912                        continue;
4913                printk("  %-8s ", zone_names[i]);
4914                if (arch_zone_lowest_possible_pfn[i] ==
4915                                arch_zone_highest_possible_pfn[i])
4916                        printk("empty\n");
4917                else
4918                        printk("%0#10lx -> %0#10lx\n",
4919                                arch_zone_lowest_possible_pfn[i],
4920                                arch_zone_highest_possible_pfn[i]);
4921        }
4922
4923        /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4924        printk("Movable zone start PFN for each node\n");
4925        for (i = 0; i < MAX_NUMNODES; i++) {
4926                if (zone_movable_pfn[i])
4927                        printk("  Node %d: %lu\n", i, zone_movable_pfn[i]);
4928        }
4929
4930        /* Print out the early_node_map[] */
4931        printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4932        for (i = 0; i < nr_nodemap_entries; i++)
4933                printk("  %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4934                                                early_node_map[i].start_pfn,
4935                                                early_node_map[i].end_pfn);
4936
4937        /* Initialise every node */
4938        mminit_verify_pageflags_layout();
4939        setup_nr_node_ids();
4940        for_each_online_node(nid) {
4941                pg_data_t *pgdat = NODE_DATA(nid);
4942                free_area_init_node(nid, NULL,
4943                                find_min_pfn_for_node(nid), NULL);
4944
4945                /* Any memory on that node */
4946                if (pgdat->node_present_pages)
4947                        node_set_state(nid, N_HIGH_MEMORY);
4948                check_for_regular_memory(pgdat);
4949        }
4950}
4951
4952static int __init cmdline_parse_core(char *p, unsigned long *core)
4953{
4954        unsigned long long coremem;
4955        if (!p)
4956                return -EINVAL;
4957
4958        coremem = memparse(p, &p);
4959        *core = coremem >> PAGE_SHIFT;
4960
4961        /* Paranoid check that UL is enough for the coremem value */
4962        WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4963
4964        return 0;
4965}
4966
4967/*
4968 * kernelcore=size sets the amount of memory for use for allocations that
4969 * cannot be reclaimed or migrated.
4970 */
4971static int __init cmdline_parse_kernelcore(char *p)
4972{
4973        return cmdline_parse_core(p, &required_kernelcore);
4974}
4975
4976/*
4977 * movablecore=size sets the amount of memory for use for allocations that
4978 * can be reclaimed or migrated.
4979 */
4980static int __init cmdline_parse_movablecore(char *p)
4981{
4982        return cmdline_parse_core(p, &required_movablecore);
4983}
4984
4985early_param("kernelcore", cmdline_parse_kernelcore);
4986early_param("movablecore", cmdline_parse_movablecore);
4987
4988#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4989
4990/**
4991 * set_dma_reserve - set the specified number of pages reserved in the first zone
4992 * @new_dma_reserve: The number of pages to mark reserved
4993 *
4994 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4995 * In the DMA zone, a significant percentage may be consumed by kernel image
4996 * and other unfreeable allocations which can skew the watermarks badly. This
4997 * function may optionally be used to account for unfreeable pages in the
4998 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4999 * smaller per-cpu batchsize.
5000 */
5001void __init set_dma_reserve(unsigned long new_dma_reserve)
5002{
5003        dma_reserve = new_dma_reserve;
5004}
5005
5006void __init free_area_init(unsigned long *zones_size)
5007{
5008        free_area_init_node(0, zones_size,
5009                        __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5010}
5011
5012static int page_alloc_cpu_notify(struct notifier_block *self,
5013                                 unsigned long action, void *hcpu)
5014{
5015        int cpu = (unsigned long)hcpu;
5016
5017        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5018                drain_pages(cpu);
5019
5020                /*
5021                 * Spill the event counters of the dead processor
5022                 * into the current processors event counters.
5023                 * This artificially elevates the count of the current
5024                 * processor.
5025                 */
5026                vm_events_fold_cpu(cpu);
5027
5028                /*
5029                 * Zero the differential counters of the dead processor
5030                 * so that the vm statistics are consistent.
5031                 *
5032                 * This is only okay since the processor is dead and cannot
5033                 * race with what we are doing.
5034                 */
5035                refresh_cpu_vm_stats(cpu);
5036        }
5037        return NOTIFY_OK;
5038}
5039
5040void __init page_alloc_init(void)
5041{
5042        hotcpu_notifier(page_alloc_cpu_notify, 0);
5043}
5044
5045/*
5046 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5047 *      or min_free_kbytes changes.
5048 */
5049static void calculate_totalreserve_pages(void)
5050{
5051        struct pglist_data *pgdat;
5052        unsigned long reserve_pages = 0;
5053        enum zone_type i, j;
5054
5055        for_each_online_pgdat(pgdat) {
5056                for (i = 0; i < MAX_NR_ZONES; i++) {
5057                        struct zone *zone = pgdat->node_zones + i;
5058                        unsigned long max = 0;
5059
5060                        /* Find valid and maximum lowmem_reserve in the zone */
5061                        for (j = i; j < MAX_NR_ZONES; j++) {
5062                                if (zone->lowmem_reserve[j] > max)
5063                                        max = zone->lowmem_reserve[j];
5064                        }
5065
5066                        /* we treat the high watermark as reserved pages. */
5067                        max += high_wmark_pages(zone);
5068
5069                        if (max > zone->present_pages)
5070                                max = zone->present_pages;
5071                        reserve_pages += max;
5072                }
5073        }
5074        totalreserve_pages = reserve_pages;
5075}
5076
5077/*
5078 * setup_per_zone_lowmem_reserve - called whenever
5079 *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
5080 *      has a correct pages reserved value, so an adequate number of
5081 *      pages are left in the zone after a successful __alloc_pages().
5082 */
5083static void setup_per_zone_lowmem_reserve(void)
5084{
5085        struct pglist_data *pgdat;
5086        enum zone_type j, idx;
5087
5088        for_each_online_pgdat(pgdat) {
5089                for (j = 0; j < MAX_NR_ZONES; j++) {
5090                        struct zone *zone = pgdat->node_zones + j;
5091                        unsigned long present_pages = zone->present_pages;
5092
5093                        zone->lowmem_reserve[j] = 0;
5094
5095                        idx = j;
5096                        while (idx) {
5097                                struct zone *lower_zone;
5098
5099                                idx--;
5100
5101                                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5102                                        sysctl_lowmem_reserve_ratio[idx] = 1;
5103
5104                                lower_zone = pgdat->node_zones + idx;
5105                                lower_zone->lowmem_reserve[j] = present_pages /
5106                                        sysctl_lowmem_reserve_ratio[idx];
5107                                present_pages += lower_zone->present_pages;
5108                        }
5109                }
5110        }
5111
5112        /* update totalreserve_pages */
5113        calculate_totalreserve_pages();
5114}
5115
5116/**
5117 * setup_per_zone_wmarks - called when min_free_kbytes changes
5118 * or when memory is hot-{added|removed}
5119 *
5120 * Ensures that the watermark[min,low,high] values for each zone are set
5121 * correctly with respect to min_free_kbytes.
5122 */
5123void setup_per_zone_wmarks(void)
5124{
5125        unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5126        unsigned long lowmem_pages = 0;
5127        struct zone *zone;
5128        unsigned long flags;
5129
5130        /* Calculate total number of !ZONE_HIGHMEM pages */
5131        for_each_zone(zone) {
5132                if (!is_highmem(zone))
5133                        lowmem_pages += zone->present_pages;
5134        }
5135
5136        for_each_zone(zone) {
5137                u64 tmp;
5138
5139                spin_lock_irqsave(&zone->lock, flags);
5140                tmp = (u64)pages_min * zone->present_pages;
5141                do_div(tmp, lowmem_pages);
5142                if (is_highmem(zone)) {
5143                        /*
5144                         * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5145                         * need highmem pages, so cap pages_min to a small
5146                         * value here.
5147                         *
5148                         * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5149                         * deltas controls asynch page reclaim, and so should
5150                         * not be capped for highmem.
5151                         */
5152                        int min_pages;
5153
5154                        min_pages = zone->present_pages / 1024;
5155                        if (min_pages < SWAP_CLUSTER_MAX)
5156                                min_pages = SWAP_CLUSTER_MAX;
5157                        if (min_pages > 128)
5158                                min_pages = 128;
5159                        zone->watermark[WMARK_MIN] = min_pages;
5160                } else {
5161                        /*
5162                         * If it's a lowmem zone, reserve a number of pages
5163                         * proportionate to the zone's size.
5164                         */
5165                        zone->watermark[WMARK_MIN] = tmp;
5166                }
5167
5168                zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + (tmp >> 2);
5169                zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5170                setup_zone_migrate_reserve(zone);
5171                spin_unlock_irqrestore(&zone->lock, flags);
5172        }
5173
5174        /* update totalreserve_pages */
5175        calculate_totalreserve_pages();
5176}
5177
5178/*
5179 * The inactive anon list should be small enough that the VM never has to
5180 * do too much work, but large enough that each inactive page has a chance
5181 * to be referenced again before it is swapped out.
5182 *
5183 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5184 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5185 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5186 * the anonymous pages are kept on the inactive list.
5187 *
5188 * total     target    max
5189 * memory    ratio     inactive anon
5190 * -------------------------------------
5191 *   10MB       1         5MB
5192 *  100MB       1        50MB
5193 *    1GB       3       250MB
5194 *   10GB      10       0.9GB
5195 *  100GB      31         3GB
5196 *    1TB     101        10GB
5197 *   10TB     320        32GB
5198 */
5199static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5200{
5201        unsigned int gb, ratio;
5202
5203        /* Zone size in gigabytes */
5204        gb = zone->present_pages >> (30 - PAGE_SHIFT);
5205        if (gb)
5206                ratio = int_sqrt(10 * gb);
5207        else
5208                ratio = 1;
5209
5210        zone->inactive_ratio = ratio;
5211}
5212
5213static void __meminit setup_per_zone_inactive_ratio(void)
5214{
5215        struct zone *zone;
5216
5217        for_each_zone(zone)
5218                calculate_zone_inactive_ratio(zone);
5219}
5220
5221/*
5222 * Initialise min_free_kbytes.
5223 *
5224 * For small machines we want it small (128k min).  For large machines
5225 * we want it large (64MB max).  But it is not linear, because network
5226 * bandwidth does not increase linearly with machine size.  We use
5227 *
5228 *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5229 *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
5230 *
5231 * which yields
5232 *
5233 * 16MB:        512k
5234 * 32MB:        724k
5235 * 64MB:        1024k
5236 * 128MB:       1448k
5237 * 256MB:       2048k
5238 * 512MB:       2896k
5239 * 1024MB:      4096k
5240 * 2048MB:      5792k
5241 * 4096MB:      8192k
5242 * 8192MB:      11584k
5243 * 16384MB:     16384k
5244 */
5245int __meminit init_per_zone_wmark_min(void)
5246{
5247        unsigned long lowmem_kbytes;
5248
5249        lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5250
5251        min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5252        if (min_free_kbytes < 128)
5253                min_free_kbytes = 128;
5254        if (min_free_kbytes > 65536)
5255                min_free_kbytes = 65536;
5256        setup_per_zone_wmarks();
5257        refresh_zone_stat_thresholds();
5258        setup_per_zone_lowmem_reserve();
5259        setup_per_zone_inactive_ratio();
5260        return 0;
5261}
5262module_init(init_per_zone_wmark_min)
5263
5264/*
5265 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
5266 *      that we can call two helper functions whenever min_free_kbytes
5267 *      changes.
5268 */
5269int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
5270        void __user *buffer, size_t *length, loff_t *ppos)
5271{
5272        proc_dointvec(table, write, buffer, length, ppos);
5273        if (write)
5274                setup_per_zone_wmarks();
5275        return 0;
5276}
5277
5278#ifdef CONFIG_NUMA
5279int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5280        void __user *buffer, size_t *length, loff_t *ppos)
5281{
5282        struct zone *zone;
5283        int rc;
5284
5285        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5286        if (rc)
5287                return rc;
5288
5289        for_each_zone(zone)
5290                zone->min_unmapped_pages = (zone->present_pages *
5291                                sysctl_min_unmapped_ratio) / 100;
5292        return 0;
5293}
5294
5295int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5296        void __user *buffer, size_t *length, loff_t *ppos)
5297{
5298        struct zone *zone;
5299        int rc;
5300
5301        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5302        if (rc)
5303                return rc;
5304
5305        for_each_zone(zone)
5306                zone->min_slab_pages = (zone->present_pages *
5307                                sysctl_min_slab_ratio) / 100;
5308        return 0;
5309}
5310#endif
5311
5312/*
5313 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5314 *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5315 *      whenever sysctl_lowmem_reserve_ratio changes.
5316 *
5317 * The reserve ratio obviously has absolutely no relation with the
5318 * minimum watermarks. The lowmem reserve ratio can only make sense
5319 * if in function of the boot time zone sizes.
5320 */
5321int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5322        void __user *buffer, size_t *length, loff_t *ppos)
5323{
5324        proc_dointvec_minmax(table, write, buffer, length, ppos);
5325        setup_per_zone_lowmem_reserve();
5326        return 0;
5327}
5328
5329/*
5330 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5331 * cpu.  It is the fraction of total pages in each zone that a hot per cpu pagelist
5332 * can have before it gets flushed back to buddy allocator.
5333 */
5334
5335int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5336        void __user *buffer, size_t *length, loff_t *ppos)
5337{
5338        struct zone *zone;
5339        unsigned int cpu;
5340        int ret;
5341
5342        ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5343        if (!write || (ret == -EINVAL))
5344                return ret;
5345        for_each_populated_zone(zone) {
5346                for_each_possible_cpu(cpu) {
5347                        unsigned long  high;
5348                        high = zone->present_pages / percpu_pagelist_fraction;
5349                        setup_pagelist_highmark(
5350                                per_cpu_ptr(zone->pageset, cpu), high);
5351                }
5352        }
5353        return 0;
5354}
5355
5356int hashdist = HASHDIST_DEFAULT;
5357
5358#ifdef CONFIG_NUMA
5359static int __init set_hashdist(char *str)
5360{
5361        if (!str)
5362                return 0;
5363        hashdist = simple_strtoul(str, &str, 0);
5364        return 1;
5365}
5366__setup("hashdist=", set_hashdist);
5367#endif
5368
5369/*
5370 * allocate a large system hash table from bootmem
5371 * - it is assumed that the hash table must contain an exact power-of-2
5372 *   quantity of entries
5373 * - limit is the number of hash buckets, not the total allocation size
5374 */
5375void *__init alloc_large_system_hash(const char *tablename,
5376                                     unsigned long bucketsize,
5377                                     unsigned long numentries,
5378                                     int scale,
5379                                     int flags,
5380                                     unsigned int *_hash_shift,
5381                                     unsigned int *_hash_mask,
5382                                     unsigned long limit)
5383{
5384        unsigned long long max = limit;
5385        unsigned long log2qty, size;
5386        void *table = NULL;
5387
5388        /* allow the kernel cmdline to have a say */
5389        if (!numentries) {
5390                /* round applicable memory size up to nearest megabyte */
5391                numentries = nr_kernel_pages;
5392                numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5393                numentries >>= 20 - PAGE_SHIFT;
5394                numentries <<= 20 - PAGE_SHIFT;
5395
5396                /* limit to 1 bucket per 2^scale bytes of low memory */
5397                if (scale > PAGE_SHIFT)
5398                        numentries >>= (scale - PAGE_SHIFT);
5399                else
5400                        numentries <<= (PAGE_SHIFT - scale);
5401
5402                /* Make sure we've got at least a 0-order allocation.. */
5403                if (unlikely(flags & HASH_SMALL)) {
5404                        /* Makes no sense without HASH_EARLY */
5405                        WARN_ON(!(flags & HASH_EARLY));
5406                        if (!(numentries >> *_hash_shift)) {
5407                                numentries = 1UL << *_hash_shift;
5408                                BUG_ON(!numentries);
5409                        }
5410                } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5411                        numentries = PAGE_SIZE / bucketsize;
5412        }
5413        numentries = roundup_pow_of_two(numentries);
5414
5415        /* limit allocation size to 1/16 total memory by default */
5416        if (max == 0) {
5417                max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5418                do_div(max, bucketsize);
5419        }
5420
5421        if (numentries > max)
5422                numentries = max;
5423
5424        log2qty = ilog2(numentries);
5425
5426        do {
5427                size = bucketsize << log2qty;
5428                if (flags & HASH_EARLY)
5429                        table = alloc_bootmem_nopanic(size);
5430                else if (hashdist)
5431                        table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5432                else {
5433                        /*
5434                         * If bucketsize is not a power-of-two, we may free
5435                         * some pages at the end of hash table which
5436                         * alloc_pages_exact() automatically does
5437                         */
5438                        if (get_order(size) < MAX_ORDER) {
5439                                table = alloc_pages_exact(size, GFP_ATOMIC);
5440                                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5441                        }
5442                }
5443        } while (!table && size > PAGE_SIZE && --log2qty);
5444
5445        if (!table)
5446                panic("Failed to allocate %s hash table\n", tablename);
5447
5448        printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5449               tablename,
5450               (1UL << log2qty),
5451               ilog2(size) - PAGE_SHIFT,
5452               size);
5453
5454        if (_hash_shift)
5455                *_hash_shift = log2qty;
5456        if (_hash_mask)
5457                *_hash_mask = (1 << log2qty) - 1;
5458
5459        return table;
5460}
5461
5462/* Return a pointer to the bitmap storing bits affecting a block of pages */
5463static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5464                                                        unsigned long pfn)
5465{
5466#ifdef CONFIG_SPARSEMEM
5467        return __pfn_to_section(pfn)->pageblock_flags;
5468#else
5469        return zone->pageblock_flags;
5470#endif /* CONFIG_SPARSEMEM */
5471}
5472
5473static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5474{
5475#ifdef CONFIG_SPARSEMEM
5476        pfn &= (PAGES_PER_SECTION-1);
5477        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5478#else
5479        pfn = pfn - zone->zone_start_pfn;
5480        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5481#endif /* CONFIG_SPARSEMEM */
5482}
5483
5484/**
5485 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5486 * @page: The page within the block of interest
5487 * @start_bitidx: The first bit of interest to retrieve
5488 * @end_bitidx: The last bit of interest
5489 * returns pageblock_bits flags
5490 */
5491unsigned long get_pageblock_flags_group(struct page *page,
5492                                        int start_bitidx, int end_bitidx)
5493{
5494        struct zone *zone;
5495        unsigned long *bitmap;
5496        unsigned long pfn, bitidx;
5497        unsigned long flags = 0;
5498        unsigned long value = 1;
5499
5500        zone = page_zone(page);
5501        pfn = page_to_pfn(page);
5502        bitmap = get_pageblock_bitmap(zone, pfn);
5503        bitidx = pfn_to_bitidx(zone, pfn);
5504
5505        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5506                if (test_bit(bitidx + start_bitidx, bitmap))
5507                        flags |= value;
5508
5509        return flags;
5510}
5511
5512/**
5513 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5514 * @page: The page within the block of interest
5515 * @start_bitidx: The first bit of interest
5516 * @end_bitidx: The last bit of interest
5517 * @flags: The flags to set
5518 */
5519void set_pageblock_flags_group(struct page *page, unsigned long flags,
5520                                        int start_bitidx, int end_bitidx)
5521{
5522        struct zone *zone;
5523        unsigned long *bitmap;
5524        unsigned long pfn, bitidx;
5525        unsigned long value = 1;
5526
5527        zone = page_zone(page);
5528        pfn = page_to_pfn(page);
5529        bitmap = get_pageblock_bitmap(zone, pfn);
5530        bitidx = pfn_to_bitidx(zone, pfn);
5531        VM_BUG_ON(pfn < zone->zone_start_pfn);
5532        VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5533
5534        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5535                if (flags & value)
5536                        __set_bit(bitidx + start_bitidx, bitmap);
5537                else
5538                        __clear_bit(bitidx + start_bitidx, bitmap);
5539}
5540
5541/*
5542 * This is designed as sub function...plz see page_isolation.c also.
5543 * set/clear page block's type to be ISOLATE.
5544 * page allocater never alloc memory from ISOLATE block.
5545 */
5546
5547static int
5548__count_immobile_pages(struct zone *zone, struct page *page, int count)
5549{
5550        unsigned long pfn, iter, found;
5551        /*
5552         * For avoiding noise data, lru_add_drain_all() should be called
5553         * If ZONE_MOVABLE, the zone never contains immobile pages
5554         */
5555        if (zone_idx(zone) == ZONE_MOVABLE)
5556                return true;
5557
5558        if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5559                return true;
5560
5561        pfn = page_to_pfn(page);
5562        for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5563                unsigned long check = pfn + iter;
5564
5565                if (!pfn_valid_within(check))
5566                        continue;
5567
5568                page = pfn_to_page(check);
5569                if (!page_count(page)) {
5570                        if (PageBuddy(page))
5571                                iter += (1 << page_order(page)) - 1;
5572                        continue;
5573                }
5574                if (!PageLRU(page))
5575                        found++;
5576                /*
5577                 * If there are RECLAIMABLE pages, we need to check it.
5578                 * But now, memory offline itself doesn't call shrink_slab()
5579                 * and it still to be fixed.
5580                 */
5581                /*
5582                 * If the page is not RAM, page_count()should be 0.
5583                 * we don't need more check. This is an _used_ not-movable page.
5584                 *
5585                 * The problematic thing here is PG_reserved pages. PG_reserved
5586                 * is set to both of a memory hole page and a _used_ kernel
5587                 * page at boot.
5588                 */
5589                if (found > count)
5590                        return false;
5591        }
5592        return true;
5593}
5594
5595bool is_pageblock_removable_nolock(struct page *page)
5596{
5597        struct zone *zone = page_zone(page);
5598        return __count_immobile_pages(zone, page, 0);
5599}
5600
5601int set_migratetype_isolate(struct page *page)
5602{
5603        struct zone *zone;
5604        unsigned long flags, pfn;
5605        struct memory_isolate_notify arg;
5606        int notifier_ret;
5607        int ret = -EBUSY;
5608
5609        zone = page_zone(page);
5610
5611        spin_lock_irqsave(&zone->lock, flags);
5612
5613        pfn = page_to_pfn(page);
5614        arg.start_pfn = pfn;
5615        arg.nr_pages = pageblock_nr_pages;
5616        arg.pages_found = 0;
5617
5618        /*
5619         * It may be possible to isolate a pageblock even if the
5620         * migratetype is not MIGRATE_MOVABLE. The memory isolation
5621         * notifier chain is used by balloon drivers to return the
5622         * number of pages in a range that are held by the balloon
5623         * driver to shrink memory. If all the pages are accounted for
5624         * by balloons, are free, or on the LRU, isolation can continue.
5625         * Later, for example, when memory hotplug notifier runs, these
5626         * pages reported as "can be isolated" should be isolated(freed)
5627         * by the balloon driver through the memory notifier chain.
5628         */
5629        notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5630        notifier_ret = notifier_to_errno(notifier_ret);
5631        if (notifier_ret)
5632                goto out;
5633        /*
5634         * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5635         * We just check MOVABLE pages.
5636         */
5637        if (__count_immobile_pages(zone, page, arg.pages_found))
5638                ret = 0;
5639
5640        /*
5641         * immobile means "not-on-lru" paes. If immobile is larger than
5642         * removable-by-driver pages reported by notifier, we'll fail.
5643         */
5644
5645out:
5646        if (!ret) {
5647                set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5648                move_freepages_block(zone, page, MIGRATE_ISOLATE);
5649        }
5650
5651        spin_unlock_irqrestore(&zone->lock, flags);
5652        if (!ret)
5653                drain_all_pages();
5654        return ret;
5655}
5656
5657void unset_migratetype_isolate(struct page *page)
5658{
5659        struct zone *zone;
5660        unsigned long flags;
5661        zone = page_zone(page);
5662        spin_lock_irqsave(&zone->lock, flags);
5663        if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5664                goto out;
5665        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5666        move_freepages_block(zone, page, MIGRATE_MOVABLE);
5667out:
5668        spin_unlock_irqrestore(&zone->lock, flags);
5669}
5670
5671#ifdef CONFIG_MEMORY_HOTREMOVE
5672/*
5673 * All pages in the range must be isolated before calling this.
5674 */
5675void
5676__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5677{
5678        struct page *page;
5679        struct zone *zone;
5680        int order, i;
5681        unsigned long pfn;
5682        unsigned long flags;
5683        /* find the first valid pfn */
5684        for (pfn = start_pfn; pfn < end_pfn; pfn++)
5685                if (pfn_valid(pfn))
5686                        break;
5687        if (pfn == end_pfn)
5688                return;
5689        zone = page_zone(pfn_to_page(pfn));
5690        spin_lock_irqsave(&zone->lock, flags);
5691        pfn = start_pfn;
5692        while (pfn < end_pfn) {
5693                if (!pfn_valid(pfn)) {
5694                        pfn++;
5695                        continue;
5696                }
5697