linux/mm/page_alloc.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/page_alloc.c
   3 *
   4 *  Manages the free list, the system allocates free pages here.
   5 *  Note that kmalloc() lives in slab.c
   6 *
   7 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   8 *  Swap reorganised 29.12.95, Stephen Tweedie
   9 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
  10 *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
  11 *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
  12 *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  13 *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  14 *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
  15 */
  16
  17#include <linux/stddef.h>
  18#include <linux/mm.h>
  19#include <linux/swap.h>
  20#include <linux/interrupt.h>
  21#include <linux/pagemap.h>
  22#include <linux/jiffies.h>
  23#include <linux/bootmem.h>
  24#include <linux/memblock.h>
  25#include <linux/compiler.h>
  26#include <linux/kernel.h>
  27#include <linux/kmemcheck.h>
  28#include <linux/module.h>
  29#include <linux/suspend.h>
  30#include <linux/pagevec.h>
  31#include <linux/blkdev.h>
  32#include <linux/slab.h>
  33#include <linux/ratelimit.h>
  34#include <linux/oom.h>
  35#include <linux/notifier.h>
  36#include <linux/topology.h>
  37#include <linux/sysctl.h>
  38#include <linux/cpu.h>
  39#include <linux/cpuset.h>
  40#include <linux/memory_hotplug.h>
  41#include <linux/nodemask.h>
  42#include <linux/vmalloc.h>
  43#include <linux/vmstat.h>
  44#include <linux/mempolicy.h>
  45#include <linux/stop_machine.h>
  46#include <linux/sort.h>
  47#include <linux/pfn.h>
  48#include <linux/backing-dev.h>
  49#include <linux/fault-inject.h>
  50#include <linux/page-isolation.h>
  51#include <linux/page_cgroup.h>
  52#include <linux/debugobjects.h>
  53#include <linux/kmemleak.h>
  54#include <linux/compaction.h>
  55#include <trace/events/kmem.h>
  56#include <linux/ftrace_event.h>
  57#include <linux/memcontrol.h>
  58#include <linux/prefetch.h>
  59#include <linux/migrate.h>
  60#include <linux/page-debug-flags.h>
  61#include <linux/sched/rt.h>
  62
  63#include <asm/tlbflush.h>
  64#include <asm/div64.h>
  65#include "internal.h"
  66
  67#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
  68DEFINE_PER_CPU(int, numa_node);
  69EXPORT_PER_CPU_SYMBOL(numa_node);
  70#endif
  71
  72#ifdef CONFIG_HAVE_MEMORYLESS_NODES
  73/*
  74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
  75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
  76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
  77 * defined in <linux/topology.h>.
  78 */
  79DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
  80EXPORT_PER_CPU_SYMBOL(_numa_mem_);
  81#endif
  82
  83/*
  84 * Array of node states.
  85 */
  86nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
  87        [N_POSSIBLE] = NODE_MASK_ALL,
  88        [N_ONLINE] = { { [0] = 1UL } },
  89#ifndef CONFIG_NUMA
  90        [N_NORMAL_MEMORY] = { { [0] = 1UL } },
  91#ifdef CONFIG_HIGHMEM
  92        [N_HIGH_MEMORY] = { { [0] = 1UL } },
  93#endif
  94#ifdef CONFIG_MOVABLE_NODE
  95        [N_MEMORY] = { { [0] = 1UL } },
  96#endif
  97        [N_CPU] = { { [0] = 1UL } },
  98#endif  /* NUMA */
  99};
 100EXPORT_SYMBOL(node_states);
 101
 102unsigned long totalram_pages __read_mostly;
 103unsigned long totalreserve_pages __read_mostly;
 104/*
 105 * When calculating the number of globally allowed dirty pages, there
 106 * is a certain number of per-zone reserves that should not be
 107 * considered dirtyable memory.  This is the sum of those reserves
 108 * over all existing zones that contribute dirtyable memory.
 109 */
 110unsigned long dirty_balance_reserve __read_mostly;
 111
 112int percpu_pagelist_fraction;
 113gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
 114
 115#ifdef CONFIG_PM_SLEEP
 116/*
 117 * The following functions are used by the suspend/hibernate code to temporarily
 118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
 119 * while devices are suspended.  To avoid races with the suspend/hibernate code,
 120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
 121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
 122 * guaranteed not to run in parallel with that modification).
 123 */
 124
 125static gfp_t saved_gfp_mask;
 126
 127void pm_restore_gfp_mask(void)
 128{
 129        WARN_ON(!mutex_is_locked(&pm_mutex));
 130        if (saved_gfp_mask) {
 131                gfp_allowed_mask = saved_gfp_mask;
 132                saved_gfp_mask = 0;
 133        }
 134}
 135
 136void pm_restrict_gfp_mask(void)
 137{
 138        WARN_ON(!mutex_is_locked(&pm_mutex));
 139        WARN_ON(saved_gfp_mask);
 140        saved_gfp_mask = gfp_allowed_mask;
 141        gfp_allowed_mask &= ~GFP_IOFS;
 142}
 143
 144bool pm_suspended_storage(void)
 145{
 146        if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
 147                return false;
 148        return true;
 149}
 150#endif /* CONFIG_PM_SLEEP */
 151
 152#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 153int pageblock_order __read_mostly;
 154#endif
 155
 156static void __free_pages_ok(struct page *page, unsigned int order);
 157
 158/*
 159 * results with 256, 32 in the lowmem_reserve sysctl:
 160 *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 161 *      1G machine -> (16M dma, 784M normal, 224M high)
 162 *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 163 *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 164 *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
 165 *
 166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 167 * don't need any ZONE_NORMAL reservation
 168 */
 169int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
 170#ifdef CONFIG_ZONE_DMA
 171         256,
 172#endif
 173#ifdef CONFIG_ZONE_DMA32
 174         256,
 175#endif
 176#ifdef CONFIG_HIGHMEM
 177         32,
 178#endif
 179         32,
 180};
 181
 182EXPORT_SYMBOL(totalram_pages);
 183
 184static char * const zone_names[MAX_NR_ZONES] = {
 185#ifdef CONFIG_ZONE_DMA
 186         "DMA",
 187#endif
 188#ifdef CONFIG_ZONE_DMA32
 189         "DMA32",
 190#endif
 191         "Normal",
 192#ifdef CONFIG_HIGHMEM
 193         "HighMem",
 194#endif
 195         "Movable",
 196};
 197
 198int min_free_kbytes = 1024;
 199
 200static unsigned long __meminitdata nr_kernel_pages;
 201static unsigned long __meminitdata nr_all_pages;
 202static unsigned long __meminitdata dma_reserve;
 203
 204#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
 205static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
 206static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
 207static unsigned long __initdata required_kernelcore;
 208static unsigned long __initdata required_movablecore;
 209static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
 210
 211/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 212int movable_zone;
 213EXPORT_SYMBOL(movable_zone);
 214#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
 215
 216#if MAX_NUMNODES > 1
 217int nr_node_ids __read_mostly = MAX_NUMNODES;
 218int nr_online_nodes __read_mostly = 1;
 219EXPORT_SYMBOL(nr_node_ids);
 220EXPORT_SYMBOL(nr_online_nodes);
 221#endif
 222
 223int page_group_by_mobility_disabled __read_mostly;
 224
 225void set_pageblock_migratetype(struct page *page, int migratetype)
 226{
 227
 228        if (unlikely(page_group_by_mobility_disabled))
 229                migratetype = MIGRATE_UNMOVABLE;
 230
 231        set_pageblock_flags_group(page, (unsigned long)migratetype,
 232                                        PB_migrate, PB_migrate_end);
 233}
 234
 235bool oom_killer_disabled __read_mostly;
 236
 237#ifdef CONFIG_DEBUG_VM
 238static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 239{
 240        int ret = 0;
 241        unsigned seq;
 242        unsigned long pfn = page_to_pfn(page);
 243        unsigned long sp, start_pfn;
 244
 245        do {
 246                seq = zone_span_seqbegin(zone);
 247                start_pfn = zone->zone_start_pfn;
 248                sp = zone->spanned_pages;
 249                if (!zone_spans_pfn(zone, pfn))
 250                        ret = 1;
 251        } while (zone_span_seqretry(zone, seq));
 252
 253        if (ret)
 254                pr_err("page %lu outside zone [ %lu - %lu ]\n",
 255                        pfn, start_pfn, start_pfn + sp);
 256
 257        return ret;
 258}
 259
 260static int page_is_consistent(struct zone *zone, struct page *page)
 261{
 262        if (!pfn_valid_within(page_to_pfn(page)))
 263                return 0;
 264        if (zone != page_zone(page))
 265                return 0;
 266
 267        return 1;
 268}
 269/*
 270 * Temporary debugging check for pages not lying within a given zone.
 271 */
 272static int bad_range(struct zone *zone, struct page *page)
 273{
 274        if (page_outside_zone_boundaries(zone, page))
 275                return 1;
 276        if (!page_is_consistent(zone, page))
 277                return 1;
 278
 279        return 0;
 280}
 281#else
 282static inline int bad_range(struct zone *zone, struct page *page)
 283{
 284        return 0;
 285}
 286#endif
 287
 288static void bad_page(struct page *page)
 289{
 290        static unsigned long resume;
 291        static unsigned long nr_shown;
 292        static unsigned long nr_unshown;
 293
 294        /* Don't complain about poisoned pages */
 295        if (PageHWPoison(page)) {
 296                page_mapcount_reset(page); /* remove PageBuddy */
 297                return;
 298        }
 299
 300        /*
 301         * Allow a burst of 60 reports, then keep quiet for that minute;
 302         * or allow a steady drip of one report per second.
 303         */
 304        if (nr_shown == 60) {
 305                if (time_before(jiffies, resume)) {
 306                        nr_unshown++;
 307                        goto out;
 308                }
 309                if (nr_unshown) {
 310                        printk(KERN_ALERT
 311                              "BUG: Bad page state: %lu messages suppressed\n",
 312                                nr_unshown);
 313                        nr_unshown = 0;
 314                }
 315                nr_shown = 0;
 316        }
 317        if (nr_shown++ == 0)
 318                resume = jiffies + 60 * HZ;
 319
 320        printk(KERN_ALERT "BUG: Bad page state in process %s  pfn:%05lx\n",
 321                current->comm, page_to_pfn(page));
 322        dump_page(page);
 323
 324        print_modules();
 325        dump_stack();
 326out:
 327        /* Leave bad fields for debug, except PageBuddy could make trouble */
 328        page_mapcount_reset(page); /* remove PageBuddy */
 329        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 330}
 331
 332/*
 333 * Higher-order pages are called "compound pages".  They are structured thusly:
 334 *
 335 * The first PAGE_SIZE page is called the "head page".
 336 *
 337 * The remaining PAGE_SIZE pages are called "tail pages".
 338 *
 339 * All pages have PG_compound set.  All tail pages have their ->first_page
 340 * pointing at the head page.
 341 *
 342 * The first tail page's ->lru.next holds the address of the compound page's
 343 * put_page() function.  Its ->lru.prev holds the order of allocation.
 344 * This usage means that zero-order pages may not be compound.
 345 */
 346
 347static void free_compound_page(struct page *page)
 348{
 349        __free_pages_ok(page, compound_order(page));
 350}
 351
 352void prep_compound_page(struct page *page, unsigned long order)
 353{
 354        int i;
 355        int nr_pages = 1 << order;
 356
 357        set_compound_page_dtor(page, free_compound_page);
 358        set_compound_order(page, order);
 359        __SetPageHead(page);
 360        for (i = 1; i < nr_pages; i++) {
 361                struct page *p = page + i;
 362                __SetPageTail(p);
 363                set_page_count(p, 0);
 364                p->first_page = page;
 365        }
 366}
 367
 368/* update __split_huge_page_refcount if you change this function */
 369static int destroy_compound_page(struct page *page, unsigned long order)
 370{
 371        int i;
 372        int nr_pages = 1 << order;
 373        int bad = 0;
 374
 375        if (unlikely(compound_order(page) != order)) {
 376                bad_page(page);
 377                bad++;
 378        }
 379
 380        __ClearPageHead(page);
 381
 382        for (i = 1; i < nr_pages; i++) {
 383                struct page *p = page + i;
 384
 385                if (unlikely(!PageTail(p) || (p->first_page != page))) {
 386                        bad_page(page);
 387                        bad++;
 388                }
 389                __ClearPageTail(p);
 390        }
 391
 392        return bad;
 393}
 394
 395static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
 396{
 397        int i;
 398
 399        /*
 400         * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
 401         * and __GFP_HIGHMEM from hard or soft interrupt context.
 402         */
 403        VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
 404        for (i = 0; i < (1 << order); i++)
 405                clear_highpage(page + i);
 406}
 407
 408#ifdef CONFIG_DEBUG_PAGEALLOC
 409unsigned int _debug_guardpage_minorder;
 410
 411static int __init debug_guardpage_minorder_setup(char *buf)
 412{
 413        unsigned long res;
 414
 415        if (kstrtoul(buf, 10, &res) < 0 ||  res > MAX_ORDER / 2) {
 416                printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
 417                return 0;
 418        }
 419        _debug_guardpage_minorder = res;
 420        printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
 421        return 0;
 422}
 423__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
 424
 425static inline void set_page_guard_flag(struct page *page)
 426{
 427        __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
 428}
 429
 430static inline void clear_page_guard_flag(struct page *page)
 431{
 432        __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
 433}
 434#else
 435static inline void set_page_guard_flag(struct page *page) { }
 436static inline void clear_page_guard_flag(struct page *page) { }
 437#endif
 438
 439static inline void set_page_order(struct page *page, int order)
 440{
 441        set_page_private(page, order);
 442        __SetPageBuddy(page);
 443}
 444
 445static inline void rmv_page_order(struct page *page)
 446{
 447        __ClearPageBuddy(page);
 448        set_page_private(page, 0);
 449}
 450
 451/*
 452 * Locate the struct page for both the matching buddy in our
 453 * pair (buddy1) and the combined O(n+1) page they form (page).
 454 *
 455 * 1) Any buddy B1 will have an order O twin B2 which satisfies
 456 * the following equation:
 457 *     B2 = B1 ^ (1 << O)
 458 * For example, if the starting buddy (buddy2) is #8 its order
 459 * 1 buddy is #10:
 460 *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 461 *
 462 * 2) Any buddy B will have an order O+1 parent P which
 463 * satisfies the following equation:
 464 *     P = B & ~(1 << O)
 465 *
 466 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
 467 */
 468static inline unsigned long
 469__find_buddy_index(unsigned long page_idx, unsigned int order)
 470{
 471        return page_idx ^ (1 << order);
 472}
 473
 474/*
 475 * This function checks whether a page is free && is the buddy
 476 * we can do coalesce a page and its buddy if
 477 * (a) the buddy is not in a hole &&
 478 * (b) the buddy is in the buddy system &&
 479 * (c) a page and its buddy have the same order &&
 480 * (d) a page and its buddy are in the same zone.
 481 *
 482 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
 483 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
 484 *
 485 * For recording page's order, we use page_private(page).
 486 */
 487static inline int page_is_buddy(struct page *page, struct page *buddy,
 488                                                                int order)
 489{
 490        if (!pfn_valid_within(page_to_pfn(buddy)))
 491                return 0;
 492
 493        if (page_zone_id(page) != page_zone_id(buddy))
 494                return 0;
 495
 496        if (page_is_guard(buddy) && page_order(buddy) == order) {
 497                VM_BUG_ON(page_count(buddy) != 0);
 498                return 1;
 499        }
 500
 501        if (PageBuddy(buddy) && page_order(buddy) == order) {
 502                VM_BUG_ON(page_count(buddy) != 0);
 503                return 1;
 504        }
 505        return 0;
 506}
 507
 508/*
 509 * Freeing function for a buddy system allocator.
 510 *
 511 * The concept of a buddy system is to maintain direct-mapped table
 512 * (containing bit values) for memory blocks of various "orders".
 513 * The bottom level table contains the map for the smallest allocatable
 514 * units of memory (here, pages), and each level above it describes
 515 * pairs of units from the levels below, hence, "buddies".
 516 * At a high level, all that happens here is marking the table entry
 517 * at the bottom level available, and propagating the changes upward
 518 * as necessary, plus some accounting needed to play nicely with other
 519 * parts of the VM system.
 520 * At each level, we keep a list of pages, which are heads of continuous
 521 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
 522 * order is recorded in page_private(page) field.
 523 * So when we are allocating or freeing one, we can derive the state of the
 524 * other.  That is, if we allocate a small block, and both were
 525 * free, the remainder of the region must be split into blocks.
 526 * If a block is freed, and its buddy is also free, then this
 527 * triggers coalescing into a block of larger size.
 528 *
 529 * -- nyc
 530 */
 531
 532static inline void __free_one_page(struct page *page,
 533                struct zone *zone, unsigned int order,
 534                int migratetype)
 535{
 536        unsigned long page_idx;
 537        unsigned long combined_idx;
 538        unsigned long uninitialized_var(buddy_idx);
 539        struct page *buddy;
 540
 541        VM_BUG_ON(!zone_is_initialized(zone));
 542
 543        if (unlikely(PageCompound(page)))
 544                if (unlikely(destroy_compound_page(page, order)))
 545                        return;
 546
 547        VM_BUG_ON(migratetype == -1);
 548
 549        page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 550
 551        VM_BUG_ON(page_idx & ((1 << order) - 1));
 552        VM_BUG_ON(bad_range(zone, page));
 553
 554        while (order < MAX_ORDER-1) {
 555                buddy_idx = __find_buddy_index(page_idx, order);
 556                buddy = page + (buddy_idx - page_idx);
 557                if (!page_is_buddy(page, buddy, order))
 558                        break;
 559                /*
 560                 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
 561                 * merge with it and move up one order.
 562                 */
 563                if (page_is_guard(buddy)) {
 564                        clear_page_guard_flag(buddy);
 565                        set_page_private(page, 0);
 566                        __mod_zone_freepage_state(zone, 1 << order,
 567                                                  migratetype);
 568                } else {
 569                        list_del(&buddy->lru);
 570                        zone->free_area[order].nr_free--;
 571                        rmv_page_order(buddy);
 572                }
 573                combined_idx = buddy_idx & page_idx;
 574                page = page + (combined_idx - page_idx);
 575                page_idx = combined_idx;
 576                order++;
 577        }
 578        set_page_order(page, order);
 579
 580        /*
 581         * If this is not the largest possible page, check if the buddy
 582         * of the next-highest order is free. If it is, it's possible
 583         * that pages are being freed that will coalesce soon. In case,
 584         * that is happening, add the free page to the tail of the list
 585         * so it's less likely to be used soon and more likely to be merged
 586         * as a higher order page
 587         */
 588        if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
 589                struct page *higher_page, *higher_buddy;
 590                combined_idx = buddy_idx & page_idx;
 591                higher_page = page + (combined_idx - page_idx);
 592                buddy_idx = __find_buddy_index(combined_idx, order + 1);
 593                higher_buddy = higher_page + (buddy_idx - combined_idx);
 594                if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
 595                        list_add_tail(&page->lru,
 596                                &zone->free_area[order].free_list[migratetype]);
 597                        goto out;
 598                }
 599        }
 600
 601        list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
 602out:
 603        zone->free_area[order].nr_free++;
 604}
 605
 606static inline int free_pages_check(struct page *page)
 607{
 608        if (unlikely(page_mapcount(page) |
 609                (page->mapping != NULL)  |
 610                (atomic_read(&page->_count) != 0) |
 611                (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
 612                (mem_cgroup_bad_page_check(page)))) {
 613                bad_page(page);
 614                return 1;
 615        }
 616        page_nid_reset_last(page);
 617        if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
 618                page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
 619        return 0;
 620}
 621
 622/*
 623 * Frees a number of pages from the PCP lists
 624 * Assumes all pages on list are in same zone, and of same order.
 625 * count is the number of pages to free.
 626 *
 627 * If the zone was previously in an "all pages pinned" state then look to
 628 * see if this freeing clears that state.
 629 *
 630 * And clear the zone's pages_scanned counter, to hold off the "all pages are
 631 * pinned" detection logic.
 632 */
 633static void free_pcppages_bulk(struct zone *zone, int count,
 634                                        struct per_cpu_pages *pcp)
 635{
 636        int migratetype = 0;
 637        int batch_free = 0;
 638        int to_free = count;
 639
 640        spin_lock(&zone->lock);
 641        zone->all_unreclaimable = 0;
 642        zone->pages_scanned = 0;
 643
 644        while (to_free) {
 645                struct page *page;
 646                struct list_head *list;
 647
 648                /*
 649                 * Remove pages from lists in a round-robin fashion. A
 650                 * batch_free count is maintained that is incremented when an
 651                 * empty list is encountered.  This is so more pages are freed
 652                 * off fuller lists instead of spinning excessively around empty
 653                 * lists
 654                 */
 655                do {
 656                        batch_free++;
 657                        if (++migratetype == MIGRATE_PCPTYPES)
 658                                migratetype = 0;
 659                        list = &pcp->lists[migratetype];
 660                } while (list_empty(list));
 661
 662                /* This is the only non-empty list. Free them all. */
 663                if (batch_free == MIGRATE_PCPTYPES)
 664                        batch_free = to_free;
 665
 666                do {
 667                        int mt; /* migratetype of the to-be-freed page */
 668
 669                        page = list_entry(list->prev, struct page, lru);
 670                        /* must delete as __free_one_page list manipulates */
 671                        list_del(&page->lru);
 672                        mt = get_freepage_migratetype(page);
 673                        /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
 674                        __free_one_page(page, zone, 0, mt);
 675                        trace_mm_page_pcpu_drain(page, 0, mt);
 676                        if (likely(!is_migrate_isolate_page(page))) {
 677                                __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
 678                                if (is_migrate_cma(mt))
 679                                        __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
 680                        }
 681                } while (--to_free && --batch_free && !list_empty(list));
 682        }
 683        spin_unlock(&zone->lock);
 684}
 685
 686static void free_one_page(struct zone *zone, struct page *page, int order,
 687                                int migratetype)
 688{
 689        spin_lock(&zone->lock);
 690        zone->all_unreclaimable = 0;
 691        zone->pages_scanned = 0;
 692
 693        __free_one_page(page, zone, order, migratetype);
 694        if (unlikely(!is_migrate_isolate(migratetype)))
 695                __mod_zone_freepage_state(zone, 1 << order, migratetype);
 696        spin_unlock(&zone->lock);
 697}
 698
 699static bool free_pages_prepare(struct page *page, unsigned int order)
 700{
 701        int i;
 702        int bad = 0;
 703
 704        trace_mm_page_free(page, order);
 705        kmemcheck_free_shadow(page, order);
 706
 707        if (PageAnon(page))
 708                page->mapping = NULL;
 709        for (i = 0; i < (1 << order); i++)
 710                bad += free_pages_check(page + i);
 711        if (bad)
 712                return false;
 713
 714        if (!PageHighMem(page)) {
 715                debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
 716                debug_check_no_obj_freed(page_address(page),
 717                                           PAGE_SIZE << order);
 718        }
 719        arch_free_page(page, order);
 720        kernel_map_pages(page, 1 << order, 0);
 721
 722        return true;
 723}
 724
 725static void __free_pages_ok(struct page *page, unsigned int order)
 726{
 727        unsigned long flags;
 728        int migratetype;
 729
 730        if (!free_pages_prepare(page, order))
 731                return;
 732
 733        local_irq_save(flags);
 734        __count_vm_events(PGFREE, 1 << order);
 735        migratetype = get_pageblock_migratetype(page);
 736        set_freepage_migratetype(page, migratetype);
 737        free_one_page(page_zone(page), page, order, migratetype);
 738        local_irq_restore(flags);
 739}
 740
 741/*
 742 * Read access to zone->managed_pages is safe because it's unsigned long,
 743 * but we still need to serialize writers. Currently all callers of
 744 * __free_pages_bootmem() except put_page_bootmem() should only be used
 745 * at boot time. So for shorter boot time, we shift the burden to
 746 * put_page_bootmem() to serialize writers.
 747 */
 748void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 749{
 750        unsigned int nr_pages = 1 << order;
 751        unsigned int loop;
 752
 753        prefetchw(page);
 754        for (loop = 0; loop < nr_pages; loop++) {
 755                struct page *p = &page[loop];
 756
 757                if (loop + 1 < nr_pages)
 758                        prefetchw(p + 1);
 759                __ClearPageReserved(p);
 760                set_page_count(p, 0);
 761        }
 762
 763        page_zone(page)->managed_pages += 1 << order;
 764        set_page_refcounted(page);
 765        __free_pages(page, order);
 766}
 767
 768#ifdef CONFIG_CMA
 769/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
 770void __init init_cma_reserved_pageblock(struct page *page)
 771{
 772        unsigned i = pageblock_nr_pages;
 773        struct page *p = page;
 774
 775        do {
 776                __ClearPageReserved(p);
 777                set_page_count(p, 0);
 778        } while (++p, --i);
 779
 780        set_page_refcounted(page);
 781        set_pageblock_migratetype(page, MIGRATE_CMA);
 782        __free_pages(page, pageblock_order);
 783        totalram_pages += pageblock_nr_pages;
 784#ifdef CONFIG_HIGHMEM
 785        if (PageHighMem(page))
 786                totalhigh_pages += pageblock_nr_pages;
 787#endif
 788}
 789#endif
 790
 791/*
 792 * The order of subdivision here is critical for the IO subsystem.
 793 * Please do not alter this order without good reasons and regression
 794 * testing. Specifically, as large blocks of memory are subdivided,
 795 * the order in which smaller blocks are delivered depends on the order
 796 * they're subdivided in this function. This is the primary factor
 797 * influencing the order in which pages are delivered to the IO
 798 * subsystem according to empirical testing, and this is also justified
 799 * by considering the behavior of a buddy system containing a single
 800 * large block of memory acted on by a series of small allocations.
 801 * This behavior is a critical factor in sglist merging's success.
 802 *
 803 * -- nyc
 804 */
 805static inline void expand(struct zone *zone, struct page *page,
 806        int low, int high, struct free_area *area,
 807        int migratetype)
 808{
 809        unsigned long size = 1 << high;
 810
 811        while (high > low) {
 812                area--;
 813                high--;
 814                size >>= 1;
 815                VM_BUG_ON(bad_range(zone, &page[size]));
 816
 817#ifdef CONFIG_DEBUG_PAGEALLOC
 818                if (high < debug_guardpage_minorder()) {
 819                        /*
 820                         * Mark as guard pages (or page), that will allow to
 821                         * merge back to allocator when buddy will be freed.
 822                         * Corresponding page table entries will not be touched,
 823                         * pages will stay not present in virtual address space
 824                         */
 825                        INIT_LIST_HEAD(&page[size].lru);
 826                        set_page_guard_flag(&page[size]);
 827                        set_page_private(&page[size], high);
 828                        /* Guard pages are not available for any usage */
 829                        __mod_zone_freepage_state(zone, -(1 << high),
 830                                                  migratetype);
 831                        continue;
 832                }
 833#endif
 834                list_add(&page[size].lru, &area->free_list[migratetype]);
 835                area->nr_free++;
 836                set_page_order(&page[size], high);
 837        }
 838}
 839
 840/*
 841 * This page is about to be returned from the page allocator
 842 */
 843static inline int check_new_page(struct page *page)
 844{
 845        if (unlikely(page_mapcount(page) |
 846                (page->mapping != NULL)  |
 847                (atomic_read(&page->_count) != 0)  |
 848                (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
 849                (mem_cgroup_bad_page_check(page)))) {
 850                bad_page(page);
 851                return 1;
 852        }
 853        return 0;
 854}
 855
 856static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 857{
 858        int i;
 859
 860        for (i = 0; i < (1 << order); i++) {
 861                struct page *p = page + i;
 862                if (unlikely(check_new_page(p)))
 863                        return 1;
 864        }
 865
 866        set_page_private(page, 0);
 867        set_page_refcounted(page);
 868
 869        arch_alloc_page(page, order);
 870        kernel_map_pages(page, 1 << order, 1);
 871
 872        if (gfp_flags & __GFP_ZERO)
 873                prep_zero_page(page, order, gfp_flags);
 874
 875        if (order && (gfp_flags & __GFP_COMP))
 876                prep_compound_page(page, order);
 877
 878        return 0;
 879}
 880
 881/*
 882 * Go through the free lists for the given migratetype and remove
 883 * the smallest available page from the freelists
 884 */
 885static inline
 886struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
 887                                                int migratetype)
 888{
 889        unsigned int current_order;
 890        struct free_area * area;
 891        struct page *page;
 892
 893        /* Find a page of the appropriate size in the preferred list */
 894        for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 895                area = &(zone->free_area[current_order]);
 896                if (list_empty(&area->free_list[migratetype]))
 897                        continue;
 898
 899                page = list_entry(area->free_list[migratetype].next,
 900                                                        struct page, lru);
 901                list_del(&page->lru);
 902                rmv_page_order(page);
 903                area->nr_free--;
 904                expand(zone, page, order, current_order, area, migratetype);
 905                return page;
 906        }
 907
 908        return NULL;
 909}
 910
 911
 912/*
 913 * This array describes the order lists are fallen back to when
 914 * the free lists for the desirable migrate type are depleted
 915 */
 916static int fallbacks[MIGRATE_TYPES][4] = {
 917        [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,     MIGRATE_RESERVE },
 918        [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,     MIGRATE_RESERVE },
 919#ifdef CONFIG_CMA
 920        [MIGRATE_MOVABLE]     = { MIGRATE_CMA,         MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
 921        [MIGRATE_CMA]         = { MIGRATE_RESERVE }, /* Never used */
 922#else
 923        [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE,   MIGRATE_RESERVE },
 924#endif
 925        [MIGRATE_RESERVE]     = { MIGRATE_RESERVE }, /* Never used */
 926#ifdef CONFIG_MEMORY_ISOLATION
 927        [MIGRATE_ISOLATE]     = { MIGRATE_RESERVE }, /* Never used */
 928#endif
 929};
 930
 931/*
 932 * Move the free pages in a range to the free lists of the requested type.
 933 * Note that start_page and end_pages are not aligned on a pageblock
 934 * boundary. If alignment is required, use move_freepages_block()
 935 */
 936int move_freepages(struct zone *zone,
 937                          struct page *start_page, struct page *end_page,
 938                          int migratetype)
 939{
 940        struct page *page;
 941        unsigned long order;
 942        int pages_moved = 0;
 943
 944#ifndef CONFIG_HOLES_IN_ZONE
 945        /*
 946         * page_zone is not safe to call in this context when
 947         * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
 948         * anyway as we check zone boundaries in move_freepages_block().
 949         * Remove at a later date when no bug reports exist related to
 950         * grouping pages by mobility
 951         */
 952        BUG_ON(page_zone(start_page) != page_zone(end_page));
 953#endif
 954
 955        for (page = start_page; page <= end_page;) {
 956                /* Make sure we are not inadvertently changing nodes */
 957                VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
 958
 959                if (!pfn_valid_within(page_to_pfn(page))) {
 960                        page++;
 961                        continue;
 962                }
 963
 964                if (!PageBuddy(page)) {
 965                        page++;
 966                        continue;
 967                }
 968
 969                order = page_order(page);
 970                list_move(&page->lru,
 971                          &zone->free_area[order].free_list[migratetype]);
 972                set_freepage_migratetype(page, migratetype);
 973                page += 1 << order;
 974                pages_moved += 1 << order;
 975        }
 976
 977        return pages_moved;
 978}
 979
 980int move_freepages_block(struct zone *zone, struct page *page,
 981                                int migratetype)
 982{
 983        unsigned long start_pfn, end_pfn;
 984        struct page *start_page, *end_page;
 985
 986        start_pfn = page_to_pfn(page);
 987        start_pfn = start_pfn & ~(pageblock_nr_pages-1);
 988        start_page = pfn_to_page(start_pfn);
 989        end_page = start_page + pageblock_nr_pages - 1;
 990        end_pfn = start_pfn + pageblock_nr_pages - 1;
 991
 992        /* Do not cross zone boundaries */
 993        if (!zone_spans_pfn(zone, start_pfn))
 994                start_page = page;
 995        if (!zone_spans_pfn(zone, end_pfn))
 996                return 0;
 997
 998        return move_freepages(zone, start_page, end_page, migratetype);
 999}
1000
1001static void change_pageblock_range(struct page *pageblock_page,
1002                                        int start_order, int migratetype)
1003{
1004        int nr_pageblocks = 1 << (start_order - pageblock_order);
1005
1006        while (nr_pageblocks--) {
1007                set_pageblock_migratetype(pageblock_page, migratetype);
1008                pageblock_page += pageblock_nr_pages;
1009        }
1010}
1011
1012/* Remove an element from the buddy allocator from the fallback list */
1013static inline struct page *
1014__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1015{
1016        struct free_area * area;
1017        int current_order;
1018        struct page *page;
1019        int migratetype, i;
1020
1021        /* Find the largest possible block of pages in the other list */
1022        for (current_order = MAX_ORDER-1; current_order >= order;
1023                                                --current_order) {
1024                for (i = 0;; i++) {
1025                        migratetype = fallbacks[start_migratetype][i];
1026
1027                        /* MIGRATE_RESERVE handled later if necessary */
1028                        if (migratetype == MIGRATE_RESERVE)
1029                                break;
1030
1031                        area = &(zone->free_area[current_order]);
1032                        if (list_empty(&area->free_list[migratetype]))
1033                                continue;
1034
1035                        page = list_entry(area->free_list[migratetype].next,
1036                                        struct page, lru);
1037                        area->nr_free--;
1038
1039                        /*
1040                         * If breaking a large block of pages, move all free
1041                         * pages to the preferred allocation list. If falling
1042                         * back for a reclaimable kernel allocation, be more
1043                         * aggressive about taking ownership of free pages
1044                         *
1045                         * On the other hand, never change migration
1046                         * type of MIGRATE_CMA pageblocks nor move CMA
1047                         * pages on different free lists. We don't
1048                         * want unmovable pages to be allocated from
1049                         * MIGRATE_CMA areas.
1050                         */
1051                        if (!is_migrate_cma(migratetype) &&
1052                            (unlikely(current_order >= pageblock_order / 2) ||
1053                             start_migratetype == MIGRATE_RECLAIMABLE ||
1054                             page_group_by_mobility_disabled)) {
1055                                int pages;
1056                                pages = move_freepages_block(zone, page,
1057                                                                start_migratetype);
1058
1059                                /* Claim the whole block if over half of it is free */
1060                                if (pages >= (1 << (pageblock_order-1)) ||
1061                                                page_group_by_mobility_disabled)
1062                                        set_pageblock_migratetype(page,
1063                                                                start_migratetype);
1064
1065                                migratetype = start_migratetype;
1066                        }
1067
1068                        /* Remove the page from the freelists */
1069                        list_del(&page->lru);
1070                        rmv_page_order(page);
1071
1072                        /* Take ownership for orders >= pageblock_order */
1073                        if (current_order >= pageblock_order &&
1074                            !is_migrate_cma(migratetype))
1075                                change_pageblock_range(page, current_order,
1076                                                        start_migratetype);
1077
1078                        expand(zone, page, order, current_order, area,
1079                               is_migrate_cma(migratetype)
1080                             ? migratetype : start_migratetype);
1081
1082                        trace_mm_page_alloc_extfrag(page, order, current_order,
1083                                start_migratetype, migratetype);
1084
1085                        return page;
1086                }
1087        }
1088
1089        return NULL;
1090}
1091
1092/*
1093 * Do the hard work of removing an element from the buddy allocator.
1094 * Call me with the zone->lock already held.
1095 */
1096static struct page *__rmqueue(struct zone *zone, unsigned int order,
1097                                                int migratetype)
1098{
1099        struct page *page;
1100
1101retry_reserve:
1102        page = __rmqueue_smallest(zone, order, migratetype);
1103
1104        if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1105                page = __rmqueue_fallback(zone, order, migratetype);
1106
1107                /*
1108                 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1109                 * is used because __rmqueue_smallest is an inline function
1110                 * and we want just one call site
1111                 */
1112                if (!page) {
1113                        migratetype = MIGRATE_RESERVE;
1114                        goto retry_reserve;
1115                }
1116        }
1117
1118        trace_mm_page_alloc_zone_locked(page, order, migratetype);
1119        return page;
1120}
1121
1122/*
1123 * Obtain a specified number of elements from the buddy allocator, all under
1124 * a single hold of the lock, for efficiency.  Add them to the supplied list.
1125 * Returns the number of new pages which were placed at *list.
1126 */
1127static int rmqueue_bulk(struct zone *zone, unsigned int order,
1128                        unsigned long count, struct list_head *list,
1129                        int migratetype, int cold)
1130{
1131        int mt = migratetype, i;
1132
1133        spin_lock(&zone->lock);
1134        for (i = 0; i < count; ++i) {
1135                struct page *page = __rmqueue(zone, order, migratetype);
1136                if (unlikely(page == NULL))
1137                        break;
1138
1139                /*
1140                 * Split buddy pages returned by expand() are received here
1141                 * in physical page order. The page is added to the callers and
1142                 * list and the list head then moves forward. From the callers
1143                 * perspective, the linked list is ordered by page number in
1144                 * some conditions. This is useful for IO devices that can
1145                 * merge IO requests if the physical pages are ordered
1146                 * properly.
1147                 */
1148                if (likely(cold == 0))
1149                        list_add(&page->lru, list);
1150                else
1151                        list_add_tail(&page->lru, list);
1152                if (IS_ENABLED(CONFIG_CMA)) {
1153                        mt = get_pageblock_migratetype(page);
1154                        if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1155                                mt = migratetype;
1156                }
1157                set_freepage_migratetype(page, mt);
1158                list = &page->lru;
1159                if (is_migrate_cma(mt))
1160                        __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1161                                              -(1 << order));
1162        }
1163        __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1164        spin_unlock(&zone->lock);
1165        return i;
1166}
1167
1168#ifdef CONFIG_NUMA
1169/*
1170 * Called from the vmstat counter updater to drain pagesets of this
1171 * currently executing processor on remote nodes after they have
1172 * expired.
1173 *
1174 * Note that this function must be called with the thread pinned to
1175 * a single processor.
1176 */
1177void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1178{
1179        unsigned long flags;
1180        int to_drain;
1181
1182        local_irq_save(flags);
1183        if (pcp->count >= pcp->batch)
1184                to_drain = pcp->batch;
1185        else
1186                to_drain = pcp->count;
1187        if (to_drain > 0) {
1188                free_pcppages_bulk(zone, to_drain, pcp);
1189                pcp->count -= to_drain;
1190        }
1191        local_irq_restore(flags);
1192}
1193#endif
1194
1195/*
1196 * Drain pages of the indicated processor.
1197 *
1198 * The processor must either be the current processor and the
1199 * thread pinned to the current processor or a processor that
1200 * is not online.
1201 */
1202static void drain_pages(unsigned int cpu)
1203{
1204        unsigned long flags;
1205        struct zone *zone;
1206
1207        for_each_populated_zone(zone) {
1208                struct per_cpu_pageset *pset;
1209                struct per_cpu_pages *pcp;
1210
1211                local_irq_save(flags);
1212                pset = per_cpu_ptr(zone->pageset, cpu);
1213
1214                pcp = &pset->pcp;
1215                if (pcp->count) {
1216                        free_pcppages_bulk(zone, pcp->count, pcp);
1217                        pcp->count = 0;
1218                }
1219                local_irq_restore(flags);
1220        }
1221}
1222
1223/*
1224 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1225 */
1226void drain_local_pages(void *arg)
1227{
1228        drain_pages(smp_processor_id());
1229}
1230
1231/*
1232 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1233 *
1234 * Note that this code is protected against sending an IPI to an offline
1235 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1236 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1237 * nothing keeps CPUs from showing up after we populated the cpumask and
1238 * before the call to on_each_cpu_mask().
1239 */
1240void drain_all_pages(void)
1241{
1242        int cpu;
1243        struct per_cpu_pageset *pcp;
1244        struct zone *zone;
1245
1246        /*
1247         * Allocate in the BSS so we wont require allocation in
1248         * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1249         */
1250        static cpumask_t cpus_with_pcps;
1251
1252        /*
1253         * We don't care about racing with CPU hotplug event
1254         * as offline notification will cause the notified
1255         * cpu to drain that CPU pcps and on_each_cpu_mask
1256         * disables preemption as part of its processing
1257         */
1258        for_each_online_cpu(cpu) {
1259                bool has_pcps = false;
1260                for_each_populated_zone(zone) {
1261                        pcp = per_cpu_ptr(zone->pageset, cpu);
1262                        if (pcp->pcp.count) {
1263                                has_pcps = true;
1264                                break;
1265                        }
1266                }
1267                if (has_pcps)
1268                        cpumask_set_cpu(cpu, &cpus_with_pcps);
1269                else
1270                        cpumask_clear_cpu(cpu, &cpus_with_pcps);
1271        }
1272        on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1273}
1274
1275#ifdef CONFIG_HIBERNATION
1276
1277void mark_free_pages(struct zone *zone)
1278{
1279        unsigned long pfn, max_zone_pfn;
1280        unsigned long flags;
1281        int order, t;
1282        struct list_head *curr;
1283
1284        if (!zone->spanned_pages)
1285                return;
1286
1287        spin_lock_irqsave(&zone->lock, flags);
1288
1289        max_zone_pfn = zone_end_pfn(zone);
1290        for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1291                if (pfn_valid(pfn)) {
1292                        struct page *page = pfn_to_page(pfn);
1293
1294                        if (!swsusp_page_is_forbidden(page))
1295                                swsusp_unset_page_free(page);
1296                }
1297
1298        for_each_migratetype_order(order, t) {
1299                list_for_each(curr, &zone->free_area[order].free_list[t]) {
1300                        unsigned long i;
1301
1302                        pfn = page_to_pfn(list_entry(curr, struct page, lru));
1303                        for (i = 0; i < (1UL << order); i++)
1304                                swsusp_set_page_free(pfn_to_page(pfn + i));
1305                }
1306        }
1307        spin_unlock_irqrestore(&zone->lock, flags);
1308}
1309#endif /* CONFIG_PM */
1310
1311/*
1312 * Free a 0-order page
1313 * cold == 1 ? free a cold page : free a hot page
1314 */
1315void free_hot_cold_page(struct page *page, int cold)
1316{
1317        struct zone *zone = page_zone(page);
1318        struct per_cpu_pages *pcp;
1319        unsigned long flags;
1320        int migratetype;
1321
1322        if (!free_pages_prepare(page, 0))
1323                return;
1324
1325        migratetype = get_pageblock_migratetype(page);
1326        set_freepage_migratetype(page, migratetype);
1327        local_irq_save(flags);
1328        __count_vm_event(PGFREE);
1329
1330        /*
1331         * We only track unmovable, reclaimable and movable on pcp lists.
1332         * Free ISOLATE pages back to the allocator because they are being
1333         * offlined but treat RESERVE as movable pages so we can get those
1334         * areas back if necessary. Otherwise, we may have to free
1335         * excessively into the page allocator
1336         */
1337        if (migratetype >= MIGRATE_PCPTYPES) {
1338                if (unlikely(is_migrate_isolate(migratetype))) {
1339                        free_one_page(zone, page, 0, migratetype);
1340                        goto out;
1341                }
1342                migratetype = MIGRATE_MOVABLE;
1343        }
1344
1345        pcp = &this_cpu_ptr(zone->pageset)->pcp;
1346        if (cold)
1347                list_add_tail(&page->lru, &pcp->lists[migratetype]);
1348        else
1349                list_add(&page->lru, &pcp->lists[migratetype]);
1350        pcp->count++;
1351        if (pcp->count >= pcp->high) {
1352                free_pcppages_bulk(zone, pcp->batch, pcp);
1353                pcp->count -= pcp->batch;
1354        }
1355
1356out:
1357        local_irq_restore(flags);
1358}
1359
1360/*
1361 * Free a list of 0-order pages
1362 */
1363void free_hot_cold_page_list(struct list_head *list, int cold)
1364{
1365        struct page *page, *next;
1366
1367        list_for_each_entry_safe(page, next, list, lru) {
1368                trace_mm_page_free_batched(page, cold);
1369                free_hot_cold_page(page, cold);
1370        }
1371}
1372
1373/*
1374 * split_page takes a non-compound higher-order page, and splits it into
1375 * n (1<<order) sub-pages: page[0..n]
1376 * Each sub-page must be freed individually.
1377 *
1378 * Note: this is probably too low level an operation for use in drivers.
1379 * Please consult with lkml before using this in your driver.
1380 */
1381void split_page(struct page *page, unsigned int order)
1382{
1383        int i;
1384
1385        VM_BUG_ON(PageCompound(page));
1386        VM_BUG_ON(!page_count(page));
1387
1388#ifdef CONFIG_KMEMCHECK
1389        /*
1390         * Split shadow pages too, because free(page[0]) would
1391         * otherwise free the whole shadow.
1392         */
1393        if (kmemcheck_page_is_tracked(page))
1394                split_page(virt_to_page(page[0].shadow), order);
1395#endif
1396
1397        for (i = 1; i < (1 << order); i++)
1398                set_page_refcounted(page + i);
1399}
1400
1401static int __isolate_free_page(struct page *page, unsigned int order)
1402{
1403        unsigned long watermark;
1404        struct zone *zone;
1405        int mt;
1406
1407        BUG_ON(!PageBuddy(page));
1408
1409        zone = page_zone(page);
1410        mt = get_pageblock_migratetype(page);
1411
1412        if (!is_migrate_isolate(mt)) {
1413                /* Obey watermarks as if the page was being allocated */
1414                watermark = low_wmark_pages(zone) + (1 << order);
1415                if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1416                        return 0;
1417
1418                __mod_zone_freepage_state(zone, -(1UL << order), mt);
1419        }
1420
1421        /* Remove page from free list */
1422        list_del(&page->lru);
1423        zone->free_area[order].nr_free--;
1424        rmv_page_order(page);
1425
1426        /* Set the pageblock if the isolated page is at least a pageblock */
1427        if (order >= pageblock_order - 1) {
1428                struct page *endpage = page + (1 << order) - 1;
1429                for (; page < endpage; page += pageblock_nr_pages) {
1430                        int mt = get_pageblock_migratetype(page);
1431                        if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1432                                set_pageblock_migratetype(page,
1433                                                          MIGRATE_MOVABLE);
1434                }
1435        }
1436
1437        return 1UL << order;
1438}
1439
1440/*
1441 * Similar to split_page except the page is already free. As this is only
1442 * being used for migration, the migratetype of the block also changes.
1443 * As this is called with interrupts disabled, the caller is responsible
1444 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1445 * are enabled.
1446 *
1447 * Note: this is probably too low level an operation for use in drivers.
1448 * Please consult with lkml before using this in your driver.
1449 */
1450int split_free_page(struct page *page)
1451{
1452        unsigned int order;
1453        int nr_pages;
1454
1455        order = page_order(page);
1456
1457        nr_pages = __isolate_free_page(page, order);
1458        if (!nr_pages)
1459                return 0;
1460
1461        /* Split into individual pages */
1462        set_page_refcounted(page);
1463        split_page(page, order);
1464        return nr_pages;
1465}
1466
1467/*
1468 * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
1469 * we cheat by calling it from here, in the order > 0 path.  Saves a branch
1470 * or two.
1471 */
1472static inline
1473struct page *buffered_rmqueue(struct zone *preferred_zone,
1474                        struct zone *zone, int order, gfp_t gfp_flags,
1475                        int migratetype)
1476{
1477        unsigned long flags;
1478        struct page *page;
1479        int cold = !!(gfp_flags & __GFP_COLD);
1480
1481again:
1482        if (likely(order == 0)) {
1483                struct per_cpu_pages *pcp;
1484                struct list_head *list;
1485
1486                local_irq_save(flags);
1487                pcp = &this_cpu_ptr(zone->pageset)->pcp;
1488                list = &pcp->lists[migratetype];
1489                if (list_empty(list)) {
1490                        pcp->count += rmqueue_bulk(zone, 0,
1491                                        pcp->batch, list,
1492                                        migratetype, cold);
1493                        if (unlikely(list_empty(list)))
1494                                goto failed;
1495                }
1496
1497                if (cold)
1498                        page = list_entry(list->prev, struct page, lru);
1499                else
1500                        page = list_entry(list->next, struct page, lru);
1501
1502                list_del(&page->lru);
1503                pcp->count--;
1504        } else {
1505                if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1506                        /*
1507                         * __GFP_NOFAIL is not to be used in new code.
1508                         *
1509                         * All __GFP_NOFAIL callers should be fixed so that they
1510                         * properly detect and handle allocation failures.
1511                         *
1512                         * We most definitely don't want callers attempting to
1513                         * allocate greater than order-1 page units with
1514                         * __GFP_NOFAIL.
1515                         */
1516                        WARN_ON_ONCE(order > 1);
1517                }
1518                spin_lock_irqsave(&zone->lock, flags);
1519                page = __rmqueue(zone, order, migratetype);
1520                spin_unlock(&zone->lock);
1521                if (!page)
1522                        goto failed;
1523                __mod_zone_freepage_state(zone, -(1 << order),
1524                                          get_pageblock_migratetype(page));
1525        }
1526
1527        __count_zone_vm_events(PGALLOC, zone, 1 << order);
1528        zone_statistics(preferred_zone, zone, gfp_flags);
1529        local_irq_restore(flags);
1530
1531        VM_BUG_ON(bad_range(zone, page));
1532        if (prep_new_page(page, order, gfp_flags))
1533                goto again;
1534        return page;
1535
1536failed:
1537        local_irq_restore(flags);
1538        return NULL;
1539}
1540
1541#ifdef CONFIG_FAIL_PAGE_ALLOC
1542
1543static struct {
1544        struct fault_attr attr;
1545
1546        u32 ignore_gfp_highmem;
1547        u32 ignore_gfp_wait;
1548        u32 min_order;
1549} fail_page_alloc = {
1550        .attr = FAULT_ATTR_INITIALIZER,
1551        .ignore_gfp_wait = 1,
1552        .ignore_gfp_highmem = 1,
1553        .min_order = 1,
1554};
1555
1556static int __init setup_fail_page_alloc(char *str)
1557{
1558        return setup_fault_attr(&fail_page_alloc.attr, str);
1559}
1560__setup("fail_page_alloc=", setup_fail_page_alloc);
1561
1562static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1563{
1564        if (order < fail_page_alloc.min_order)
1565                return false;
1566        if (gfp_mask & __GFP_NOFAIL)
1567                return false;
1568        if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1569                return false;
1570        if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1571                return false;
1572
1573        return should_fail(&fail_page_alloc.attr, 1 << order);
1574}
1575
1576#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1577
1578static int __init fail_page_alloc_debugfs(void)
1579{
1580        umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1581        struct dentry *dir;
1582
1583        dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1584                                        &fail_page_alloc.attr);
1585        if (IS_ERR(dir))
1586                return PTR_ERR(dir);
1587
1588        if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1589                                &fail_page_alloc.ignore_gfp_wait))
1590                goto fail;
1591        if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1592                                &fail_page_alloc.ignore_gfp_highmem))
1593                goto fail;
1594        if (!debugfs_create_u32("min-order", mode, dir,
1595                                &fail_page_alloc.min_order))
1596                goto fail;
1597
1598        return 0;
1599fail:
1600        debugfs_remove_recursive(dir);
1601
1602        return -ENOMEM;
1603}
1604
1605late_initcall(fail_page_alloc_debugfs);
1606
1607#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1608
1609#else /* CONFIG_FAIL_PAGE_ALLOC */
1610
1611static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1612{
1613        return false;
1614}
1615
1616#endif /* CONFIG_FAIL_PAGE_ALLOC */
1617
1618/*
1619 * Return true if free pages are above 'mark'. This takes into account the order
1620 * of the allocation.
1621 */
1622static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1623                      int classzone_idx, int alloc_flags, long free_pages)
1624{
1625        /* free_pages my go negative - that's OK */
1626        long min = mark;
1627        long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1628        int o;
1629
1630        free_pages -= (1 << order) - 1;
1631        if (alloc_flags & ALLOC_HIGH)
1632                min -= min / 2;
1633        if (alloc_flags & ALLOC_HARDER)
1634                min -= min / 4;
1635#ifdef CONFIG_CMA
1636        /* If allocation can't use CMA areas don't use free CMA pages */
1637        if (!(alloc_flags & ALLOC_CMA))
1638                free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1639#endif
1640        if (free_pages <= min + lowmem_reserve)
1641                return false;
1642        for (o = 0; o < order; o++) {
1643                /* At the next order, this order's pages become unavailable */
1644                free_pages -= z->free_area[o].nr_free << o;
1645
1646                /* Require fewer higher order pages to be free */
1647                min >>= 1;
1648
1649                if (free_pages <= min)
1650                        return false;
1651        }
1652        return true;
1653}
1654
1655bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1656                      int classzone_idx, int alloc_flags)
1657{
1658        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1659                                        zone_page_state(z, NR_FREE_PAGES));
1660}
1661
1662bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1663                      int classzone_idx, int alloc_flags)
1664{
1665        long free_pages = zone_page_state(z, NR_FREE_PAGES);
1666
1667        if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1668                free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1669
1670        return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1671                                                                free_pages);
1672}
1673
1674#ifdef CONFIG_NUMA
1675/*
1676 * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1677 * skip over zones that are not allowed by the cpuset, or that have
1678 * been recently (in last second) found to be nearly full.  See further
1679 * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1680 * that have to skip over a lot of full or unallowed zones.
1681 *
1682 * If the zonelist cache is present in the passed in zonelist, then
1683 * returns a pointer to the allowed node mask (either the current
1684 * tasks mems_allowed, or node_states[N_MEMORY].)
1685 *
1686 * If the zonelist cache is not available for this zonelist, does
1687 * nothing and returns NULL.
1688 *
1689 * If the fullzones BITMAP in the zonelist cache is stale (more than
1690 * a second since last zap'd) then we zap it out (clear its bits.)
1691 *
1692 * We hold off even calling zlc_setup, until after we've checked the
1693 * first zone in the zonelist, on the theory that most allocations will
1694 * be satisfied from that first zone, so best to examine that zone as
1695 * quickly as we can.
1696 */
1697static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1698{
1699        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1700        nodemask_t *allowednodes;       /* zonelist_cache approximation */
1701
1702        zlc = zonelist->zlcache_ptr;
1703        if (!zlc)
1704                return NULL;
1705
1706        if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1707                bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1708                zlc->last_full_zap = jiffies;
1709        }
1710
1711        allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1712                                        &cpuset_current_mems_allowed :
1713                                        &node_states[N_MEMORY];
1714        return allowednodes;
1715}
1716
1717/*
1718 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1719 * if it is worth looking at further for free memory:
1720 *  1) Check that the zone isn't thought to be full (doesn't have its
1721 *     bit set in the zonelist_cache fullzones BITMAP).
1722 *  2) Check that the zones node (obtained from the zonelist_cache
1723 *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1724 * Return true (non-zero) if zone is worth looking at further, or
1725 * else return false (zero) if it is not.
1726 *
1727 * This check -ignores- the distinction between various watermarks,
1728 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1729 * found to be full for any variation of these watermarks, it will
1730 * be considered full for up to one second by all requests, unless
1731 * we are so low on memory on all allowed nodes that we are forced
1732 * into the second scan of the zonelist.
1733 *
1734 * In the second scan we ignore this zonelist cache and exactly
1735 * apply the watermarks to all zones, even it is slower to do so.
1736 * We are low on memory in the second scan, and should leave no stone
1737 * unturned looking for a free page.
1738 */
1739static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1740                                                nodemask_t *allowednodes)
1741{
1742        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1743        int i;                          /* index of *z in zonelist zones */
1744        int n;                          /* node that zone *z is on */
1745
1746        zlc = zonelist->zlcache_ptr;
1747        if (!zlc)
1748                return 1;
1749
1750        i = z - zonelist->_zonerefs;
1751        n = zlc->z_to_n[i];
1752
1753        /* This zone is worth trying if it is allowed but not full */
1754        return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1755}
1756
1757/*
1758 * Given 'z' scanning a zonelist, set the corresponding bit in
1759 * zlc->fullzones, so that subsequent attempts to allocate a page
1760 * from that zone don't waste time re-examining it.
1761 */
1762static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1763{
1764        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1765        int i;                          /* index of *z in zonelist zones */
1766
1767        zlc = zonelist->zlcache_ptr;
1768        if (!zlc)
1769                return;
1770
1771        i = z - zonelist->_zonerefs;
1772
1773        set_bit(i, zlc->fullzones);
1774}
1775
1776/*
1777 * clear all zones full, called after direct reclaim makes progress so that
1778 * a zone that was recently full is not skipped over for up to a second
1779 */
1780static void zlc_clear_zones_full(struct zonelist *zonelist)
1781{
1782        struct zonelist_cache *zlc;     /* cached zonelist speedup info */
1783
1784        zlc = zonelist->zlcache_ptr;
1785        if (!zlc)
1786                return;
1787
1788        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1789}
1790
1791static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1792{
1793        return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1794}
1795
1796static void __paginginit init_zone_allows_reclaim(int nid)
1797{
1798        int i;
1799
1800        for_each_online_node(i)
1801                if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1802                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
1803                else
1804                        zone_reclaim_mode = 1;
1805}
1806
1807#else   /* CONFIG_NUMA */
1808
1809static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1810{
1811        return NULL;
1812}
1813
1814static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1815                                nodemask_t *allowednodes)
1816{
1817        return 1;
1818}
1819
1820static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1821{
1822}
1823
1824static void zlc_clear_zones_full(struct zonelist *zonelist)
1825{
1826}
1827
1828static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1829{
1830        return true;
1831}
1832
1833static inline void init_zone_allows_reclaim(int nid)
1834{
1835}
1836#endif  /* CONFIG_NUMA */
1837
1838/*
1839 * get_page_from_freelist goes through the zonelist trying to allocate
1840 * a page.
1841 */
1842static struct page *
1843get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1844                struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1845                struct zone *preferred_zone, int migratetype)
1846{
1847        struct zoneref *z;
1848        struct page *page = NULL;
1849        int classzone_idx;
1850        struct zone *zone;
1851        nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1852        int zlc_active = 0;             /* set if using zonelist_cache */
1853        int did_zlc_setup = 0;          /* just call zlc_setup() one time */
1854
1855        classzone_idx = zone_idx(preferred_zone);
1856zonelist_scan:
1857        /*
1858         * Scan zonelist, looking for a zone with enough free.
1859         * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1860         */
1861        for_each_zone_zonelist_nodemask(zone, z, zonelist,
1862                                                high_zoneidx, nodemask) {
1863                if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1864                        !zlc_zone_worth_trying(zonelist, z, allowednodes))
1865                                continue;
1866                if ((alloc_flags & ALLOC_CPUSET) &&
1867                        !cpuset_zone_allowed_softwall(zone, gfp_mask))
1868                                continue;
1869                /*
1870                 * When allocating a page cache page for writing, we
1871                 * want to get it from a zone that is within its dirty
1872                 * limit, such that no single zone holds more than its
1873                 * proportional share of globally allowed dirty pages.
1874                 * The dirty limits take into account the zone's
1875                 * lowmem reserves and high watermark so that kswapd
1876                 * should be able to balance it without having to
1877                 * write pages from its LRU list.
1878                 *
1879                 * This may look like it could increase pressure on
1880                 * lower zones by failing allocations in higher zones
1881                 * before they are full.  But the pages that do spill
1882                 * over are limited as the lower zones are protected
1883                 * by this very same mechanism.  It should not become
1884                 * a practical burden to them.
1885                 *
1886                 * XXX: For now, allow allocations to potentially
1887                 * exceed the per-zone dirty limit in the slowpath
1888                 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1889                 * which is important when on a NUMA setup the allowed
1890                 * zones are together not big enough to reach the
1891                 * global limit.  The proper fix for these situations
1892                 * will require awareness of zones in the
1893                 * dirty-throttling and the flusher threads.
1894                 */
1895                if ((alloc_flags & ALLOC_WMARK_LOW) &&
1896                    (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1897                        goto this_zone_full;
1898
1899                BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1900                if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1901                        unsigned long mark;
1902                        int ret;
1903
1904                        mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1905                        if (zone_watermark_ok(zone, order, mark,
1906                                    classzone_idx, alloc_flags))
1907                                goto try_this_zone;
1908
1909                        if (IS_ENABLED(CONFIG_NUMA) &&
1910                                        !did_zlc_setup && nr_online_nodes > 1) {
1911                                /*
1912                                 * we do zlc_setup if there are multiple nodes
1913                                 * and before considering the first zone allowed
1914                                 * by the cpuset.
1915                                 */
1916                                allowednodes = zlc_setup(zonelist, alloc_flags);
1917                                zlc_active = 1;
1918                                did_zlc_setup = 1;
1919                        }
1920
1921                        if (zone_reclaim_mode == 0 ||
1922                            !zone_allows_reclaim(preferred_zone, zone))
1923                                goto this_zone_full;
1924
1925                        /*
1926                         * As we may have just activated ZLC, check if the first
1927                         * eligible zone has failed zone_reclaim recently.
1928                         */
1929                        if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1930                                !zlc_zone_worth_trying(zonelist, z, allowednodes))
1931                                continue;
1932
1933                        ret = zone_reclaim(zone, gfp_mask, order);
1934                        switch (ret) {
1935                        case ZONE_RECLAIM_NOSCAN:
1936                                /* did not scan */
1937                                continue;
1938                        case ZONE_RECLAIM_FULL:
1939                                /* scanned but unreclaimable */
1940                                continue;
1941                        default:
1942                                /* did we reclaim enough */
1943                                if (!zone_watermark_ok(zone, order, mark,
1944                                                classzone_idx, alloc_flags))
1945                                        goto this_zone_full;
1946                        }
1947                }
1948
1949try_this_zone:
1950                page = buffered_rmqueue(preferred_zone, zone, order,
1951                                                gfp_mask, migratetype);
1952                if (page)
1953                        break;
1954this_zone_full:
1955                if (IS_ENABLED(CONFIG_NUMA))
1956                        zlc_mark_zone_full(zonelist, z);
1957        }
1958
1959        if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1960                /* Disable zlc cache for second zonelist scan */
1961                zlc_active = 0;
1962                goto zonelist_scan;
1963        }
1964
1965        if (page)
1966                /*
1967                 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1968                 * necessary to allocate the page. The expectation is
1969                 * that the caller is taking steps that will free more
1970                 * memory. The caller should avoid the page being used
1971                 * for !PFMEMALLOC purposes.
1972                 */
1973                page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1974
1975        return page;
1976}
1977
1978/*
1979 * Large machines with many possible nodes should not always dump per-node
1980 * meminfo in irq context.
1981 */
1982static inline bool should_suppress_show_mem(void)
1983{
1984        bool ret = false;
1985
1986#if NODES_SHIFT > 8
1987        ret = in_interrupt();
1988#endif
1989        return ret;
1990}
1991
1992static DEFINE_RATELIMIT_STATE(nopage_rs,
1993                DEFAULT_RATELIMIT_INTERVAL,
1994                DEFAULT_RATELIMIT_BURST);
1995
1996void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1997{
1998        unsigned int filter = SHOW_MEM_FILTER_NODES;
1999
2000        if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2001            debug_guardpage_minorder() > 0)
2002                return;
2003
2004        /*
2005         * This documents exceptions given to allocations in certain
2006         * contexts that are allowed to allocate outside current's set
2007         * of allowed nodes.
2008         */
2009        if (!(gfp_mask & __GFP_NOMEMALLOC))
2010                if (test_thread_flag(TIF_MEMDIE) ||
2011                    (current->flags & (PF_MEMALLOC | PF_EXITING)))
2012                        filter &= ~SHOW_MEM_FILTER_NODES;
2013        if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2014                filter &= ~SHOW_MEM_FILTER_NODES;
2015
2016        if (fmt) {
2017                struct va_format vaf;
2018                va_list args;
2019
2020                va_start(args, fmt);
2021
2022                vaf.fmt = fmt;
2023                vaf.va = &args;
2024
2025                pr_warn("%pV", &vaf);
2026
2027                va_end(args);
2028        }
2029
2030        pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2031                current->comm, order, gfp_mask);
2032
2033        dump_stack();
2034        if (!should_suppress_show_mem())
2035                show_mem(filter);
2036}
2037
2038static inline int
2039should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2040                                unsigned long did_some_progress,
2041                                unsigned long pages_reclaimed)
2042{
2043        /* Do not loop if specifically requested */
2044        if (gfp_mask & __GFP_NORETRY)
2045                return 0;
2046
2047        /* Always retry if specifically requested */
2048        if (gfp_mask & __GFP_NOFAIL)
2049                return 1;
2050
2051        /*
2052         * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2053         * making forward progress without invoking OOM. Suspend also disables
2054         * storage devices so kswapd will not help. Bail if we are suspending.
2055         */
2056        if (!did_some_progress && pm_suspended_storage())
2057                return 0;
2058
2059        /*
2060         * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2061         * means __GFP_NOFAIL, but that may not be true in other
2062         * implementations.
2063         */
2064        if (order <= PAGE_ALLOC_COSTLY_ORDER)
2065                return 1;
2066
2067        /*
2068         * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2069         * specified, then we retry until we no longer reclaim any pages
2070         * (above), or we've reclaimed an order of pages at least as
2071         * large as the allocation's order. In both cases, if the
2072         * allocation still fails, we stop retrying.
2073         */
2074        if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2075                return 1;
2076
2077        return 0;
2078}
2079
2080static inline struct page *
2081__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2082        struct zonelist *zonelist, enum zone_type high_zoneidx,
2083        nodemask_t *nodemask, struct zone *preferred_zone,
2084        int migratetype)
2085{
2086        struct page *page;
2087
2088        /* Acquire the OOM killer lock for the zones in zonelist */
2089        if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2090                schedule_timeout_uninterruptible(1);
2091                return NULL;
2092        }
2093
2094        /*
2095         * Go through the zonelist yet one more time, keep very high watermark
2096         * here, this is only to catch a parallel oom killing, we must fail if
2097         * we're still under heavy pressure.
2098         */
2099        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2100                order, zonelist, high_zoneidx,
2101                ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2102                preferred_zone, migratetype);
2103        if (page)
2104                goto out;
2105
2106        if (!(gfp_mask & __GFP_NOFAIL)) {
2107                /* The OOM killer will not help higher order allocs */
2108                if (order > PAGE_ALLOC_COSTLY_ORDER)
2109                        goto out;
2110                /* The OOM killer does not needlessly kill tasks for lowmem */
2111                if (high_zoneidx < ZONE_NORMAL)
2112                        goto out;
2113                /*
2114                 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2115                 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2116                 * The caller should handle page allocation failure by itself if
2117                 * it specifies __GFP_THISNODE.
2118                 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2119                 */
2120                if (gfp_mask & __GFP_THISNODE)
2121                        goto out;
2122        }
2123        /* Exhausted what can be done so it's blamo time */
2124        out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2125
2126out:
2127        clear_zonelist_oom(zonelist, gfp_mask);
2128        return page;
2129}
2130
2131#ifdef CONFIG_COMPACTION
2132/* Try memory compaction for high-order allocations before reclaim */
2133static struct page *
2134__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2135        struct zonelist *zonelist, enum zone_type high_zoneidx,
2136        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2137        int migratetype, bool sync_migration,
2138        bool *contended_compaction, bool *deferred_compaction,
2139        unsigned long *did_some_progress)
2140{
2141        if (!order)
2142                return NULL;
2143
2144        if (compaction_deferred(preferred_zone, order)) {
2145                *deferred_compaction = true;
2146                return NULL;
2147        }
2148
2149        current->flags |= PF_MEMALLOC;
2150        *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2151                                                nodemask, sync_migration,
2152                                                contended_compaction);
2153        current->flags &= ~PF_MEMALLOC;
2154
2155        if (*did_some_progress != COMPACT_SKIPPED) {
2156                struct page *page;
2157
2158                /* Page migration frees to the PCP lists but we want merging */
2159                drain_pages(get_cpu());
2160                put_cpu();
2161
2162                page = get_page_from_freelist(gfp_mask, nodemask,
2163                                order, zonelist, high_zoneidx,
2164                                alloc_flags & ~ALLOC_NO_WATERMARKS,
2165                                preferred_zone, migratetype);
2166                if (page) {
2167                        preferred_zone->compact_blockskip_flush = false;
2168                        preferred_zone->compact_considered = 0;
2169                        preferred_zone->compact_defer_shift = 0;
2170                        if (order >= preferred_zone->compact_order_failed)
2171                                preferred_zone->compact_order_failed = order + 1;
2172                        count_vm_event(COMPACTSUCCESS);
2173                        return page;
2174                }
2175
2176                /*
2177                 * It's bad if compaction run occurs and fails.
2178                 * The most likely reason is that pages exist,
2179                 * but not enough to satisfy watermarks.
2180                 */
2181                count_vm_event(COMPACTFAIL);
2182
2183                /*
2184                 * As async compaction considers a subset of pageblocks, only
2185                 * defer if the failure was a sync compaction failure.
2186                 */
2187                if (sync_migration)
2188                        defer_compaction(preferred_zone, order);
2189
2190                cond_resched();
2191        }
2192
2193        return NULL;
2194}
2195#else
2196static inline struct page *
2197__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2198        struct zonelist *zonelist, enum zone_type high_zoneidx,
2199        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2200        int migratetype, bool sync_migration,
2201        bool *contended_compaction, bool *deferred_compaction,
2202        unsigned long *did_some_progress)
2203{
2204        return NULL;
2205}
2206#endif /* CONFIG_COMPACTION */
2207
2208/* Perform direct synchronous page reclaim */
2209static int
2210__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2211                  nodemask_t *nodemask)
2212{
2213        struct reclaim_state reclaim_state;
2214        int progress;
2215
2216        cond_resched();
2217
2218        /* We now go into synchronous reclaim */
2219        cpuset_memory_pressure_bump();
2220        current->flags |= PF_MEMALLOC;
2221        lockdep_set_current_reclaim_state(gfp_mask);
2222        reclaim_state.reclaimed_slab = 0;
2223        current->reclaim_state = &reclaim_state;
2224
2225        progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2226
2227        current->reclaim_state = NULL;
2228        lockdep_clear_current_reclaim_state();
2229        current->flags &= ~PF_MEMALLOC;
2230
2231        cond_resched();
2232
2233        return progress;
2234}
2235
2236/* The really slow allocator path where we enter direct reclaim */
2237static inline struct page *
2238__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2239        struct zonelist *zonelist, enum zone_type high_zoneidx,
2240        nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2241        int migratetype, unsigned long *did_some_progress)
2242{
2243        struct page *page = NULL;
2244        bool drained = false;
2245
2246        *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2247                                               nodemask);
2248        if (unlikely(!(*did_some_progress)))
2249                return NULL;
2250
2251        /* After successful reclaim, reconsider all zones for allocation */
2252        if (IS_ENABLED(CONFIG_NUMA))
2253                zlc_clear_zones_full(zonelist);
2254
2255retry:
2256        page = get_page_from_freelist(gfp_mask, nodemask, order,
2257                                        zonelist, high_zoneidx,
2258                                        alloc_flags & ~ALLOC_NO_WATERMARKS,
2259                                        preferred_zone, migratetype);
2260
2261        /*
2262         * If an allocation failed after direct reclaim, it could be because
2263         * pages are pinned on the per-cpu lists. Drain them and try again
2264         */
2265        if (!page && !drained) {
2266                drain_all_pages();
2267                drained = true;
2268                goto retry;
2269        }
2270
2271        return page;
2272}
2273
2274/*
2275 * This is called in the allocator slow-path if the allocation request is of
2276 * sufficient urgency to ignore watermarks and take other desperate measures
2277 */
2278static inline struct page *
2279__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2280        struct zonelist *zonelist, enum zone_type high_zoneidx,
2281        nodemask_t *nodemask, struct zone *preferred_zone,
2282        int migratetype)
2283{
2284        struct page *page;
2285
2286        do {
2287                page = get_page_from_freelist(gfp_mask, nodemask, order,
2288                        zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2289                        preferred_zone, migratetype);
2290
2291                if (!page && gfp_mask & __GFP_NOFAIL)
2292                        wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2293        } while (!page && (gfp_mask & __GFP_NOFAIL));
2294
2295        return page;
2296}
2297
2298static inline
2299void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2300                                                enum zone_type high_zoneidx,
2301                                                enum zone_type classzone_idx)
2302{
2303        struct zoneref *z;
2304        struct zone *zone;
2305
2306        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2307                wakeup_kswapd(zone, order, classzone_idx);
2308}
2309
2310static inline int
2311gfp_to_alloc_flags(gfp_t gfp_mask)
2312{
2313        int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2314        const gfp_t wait = gfp_mask & __GFP_WAIT;
2315
2316        /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2317        BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2318
2319        /*
2320         * The caller may dip into page reserves a bit more if the caller
2321         * cannot run direct reclaim, or if the caller has realtime scheduling
2322         * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
2323         * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2324         */
2325        alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2326
2327        if (!wait) {
2328                /*
2329                 * Not worth trying to allocate harder for
2330                 * __GFP_NOMEMALLOC even if it can't schedule.
2331                 */
2332                if  (!(gfp_mask & __GFP_NOMEMALLOC))
2333                        alloc_flags |= ALLOC_HARDER;
2334                /*
2335                 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2336                 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2337                 */
2338                alloc_flags &= ~ALLOC_CPUSET;
2339        } else if (unlikely(rt_task(current)) && !in_interrupt())
2340                alloc_flags |= ALLOC_HARDER;
2341
2342        if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2343                if (gfp_mask & __GFP_MEMALLOC)
2344                        alloc_flags |= ALLOC_NO_WATERMARKS;
2345                else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2346                        alloc_flags |= ALLOC_NO_WATERMARKS;
2347                else if (!in_interrupt() &&
2348                                ((current->flags & PF_MEMALLOC) ||
2349                                 unlikely(test_thread_flag(TIF_MEMDIE))))
2350                        alloc_flags |= ALLOC_NO_WATERMARKS;
2351        }
2352#ifdef CONFIG_CMA
2353        if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2354                alloc_flags |= ALLOC_CMA;
2355#endif
2356        return alloc_flags;
2357}
2358
2359bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2360{
2361        return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2362}
2363
2364static inline struct page *
2365__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2366        struct zonelist *zonelist, enum zone_type high_zoneidx,
2367        nodemask_t *nodemask, struct zone *preferred_zone,
2368        int migratetype)
2369{
2370        const gfp_t wait = gfp_mask & __GFP_WAIT;
2371        struct page *page = NULL;
2372        int alloc_flags;
2373        unsigned long pages_reclaimed = 0;
2374        unsigned long did_some_progress;
2375        bool sync_migration = false;
2376        bool deferred_compaction = false;
2377        bool contended_compaction = false;
2378
2379        /*
2380         * In the slowpath, we sanity check order to avoid ever trying to
2381         * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2382         * be using allocators in order of preference for an area that is
2383         * too large.
2384         */
2385        if (order >= MAX_ORDER) {
2386                WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2387                return NULL;
2388        }
2389
2390        /*
2391         * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2392         * __GFP_NOWARN set) should not cause reclaim since the subsystem
2393         * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2394         * using a larger set of nodes after it has established that the
2395         * allowed per node queues are empty and that nodes are
2396         * over allocated.
2397         */
2398        if (IS_ENABLED(CONFIG_NUMA) &&
2399                        (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2400                goto nopage;
2401
2402restart:
2403        if (!(gfp_mask & __GFP_NO_KSWAPD))
2404                wake_all_kswapd(order, zonelist, high_zoneidx,
2405                                                zone_idx(preferred_zone));
2406
2407        /*
2408         * OK, we're below the kswapd watermark and have kicked background
2409         * reclaim. Now things get more complex, so set up alloc_flags according
2410         * to how we want to proceed.
2411         */
2412        alloc_flags = gfp_to_alloc_flags(gfp_mask);
2413
2414        /*
2415         * Find the true preferred zone if the allocation is unconstrained by
2416         * cpusets.
2417         */
2418        if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2419                first_zones_zonelist(zonelist, high_zoneidx, NULL,
2420                                        &preferred_zone);
2421
2422rebalance:
2423        /* This is the last chance, in general, before the goto nopage. */
2424        page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2425                        high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2426                        preferred_zone, migratetype);
2427        if (page)
2428                goto got_pg;
2429
2430        /* Allocate without watermarks if the context allows */
2431        if (alloc_flags & ALLOC_NO_WATERMARKS) {
2432                /*
2433                 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2434                 * the allocation is high priority and these type of
2435                 * allocations are system rather than user orientated
2436                 */
2437                zonelist = node_zonelist(numa_node_id(), gfp_mask);
2438
2439                page = __alloc_pages_high_priority(gfp_mask, order,
2440                                zonelist, high_zoneidx, nodemask,
2441                                preferred_zone, migratetype);
2442                if (page) {
2443                        goto got_pg;
2444                }
2445        }
2446
2447        /* Atomic allocations - we can't balance anything */
2448        if (!wait)
2449                goto nopage;
2450
2451        /* Avoid recursion of direct reclaim */
2452        if (current->flags & PF_MEMALLOC)
2453                goto nopage;
2454
2455        /* Avoid allocations with no watermarks from looping endlessly */
2456        if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2457                goto nopage;
2458
2459        /*
2460         * Try direct compaction. The first pass is asynchronous. Subsequent
2461         * attempts after direct reclaim are synchronous
2462         */
2463        page = __alloc_pages_direct_compact(gfp_mask, order,
2464                                        zonelist, high_zoneidx,
2465                                        nodemask,
2466                                        alloc_flags, preferred_zone,
2467                                        migratetype, sync_migration,
2468                                        &contended_compaction,
2469                                        &deferred_compaction,
2470                                        &did_some_progress);
2471        if (page)
2472                goto got_pg;
2473        sync_migration = true;
2474
2475        /*
2476         * If compaction is deferred for high-order allocations, it is because
2477         * sync compaction recently failed. In this is the case and the caller
2478         * requested a movable allocation that does not heavily disrupt the
2479         * system then fail the allocation instead of entering direct reclaim.
2480         */
2481        if ((deferred_compaction || contended_compaction) &&
2482                                                (gfp_mask & __GFP_NO_KSWAPD))
2483                goto nopage;
2484
2485        /* Try direct reclaim and then allocating */
2486        page = __alloc_pages_direct_reclaim(gfp_mask, order,
2487                                        zonelist, high_zoneidx,
2488                                        nodemask,
2489                                        alloc_flags, preferred_zone,
2490                                        migratetype, &did_some_progress);
2491        if (page)
2492                goto got_pg;
2493
2494        /*
2495         * If we failed to make any progress reclaiming, then we are
2496         * running out of options and have to consider going OOM
2497         */
2498        if (!did_some_progress) {
2499                if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2500                        if (oom_killer_disabled)
2501                                goto nopage;
2502                        /* Coredumps can quickly deplete all memory reserves */
2503                        if ((current->flags & PF_DUMPCORE) &&
2504                            !(gfp_mask & __GFP_NOFAIL))
2505                                goto nopage;
2506                        page = __alloc_pages_may_oom(gfp_mask, order,
2507                                        zonelist, high_zoneidx,
2508                                        nodemask, preferred_zone,
2509                                        migratetype);
2510                        if (page)
2511                                goto got_pg;
2512
2513                        if (!(gfp_mask & __GFP_NOFAIL)) {
2514                                /*
2515                                 * The oom killer is not called for high-order
2516                                 * allocations that may fail, so if no progress
2517                                 * is being made, there are no other options and
2518                                 * retrying is unlikely to help.
2519                                 */
2520                                if (order > PAGE_ALLOC_COSTLY_ORDER)
2521                                        goto nopage;
2522                                /*
2523                                 * The oom killer is not called for lowmem
2524                                 * allocations to prevent needlessly killing
2525                                 * innocent tasks.
2526                                 */
2527                                if (high_zoneidx < ZONE_NORMAL)
2528                                        goto nopage;
2529                        }
2530
2531                        goto restart;
2532                }
2533        }
2534
2535        /* Check if we should retry the allocation */
2536        pages_reclaimed += did_some_progress;
2537        if (should_alloc_retry(gfp_mask, order, did_some_progress,
2538                                                pages_reclaimed)) {
2539                /* Wait for some write requests to complete then retry */
2540                wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2541                goto rebalance;
2542        } else {
2543                /*
2544                 * High-order allocations do not necessarily loop after
2545                 * direct reclaim and reclaim/compaction depends on compaction
2546                 * being called after reclaim so call directly if necessary
2547                 */
2548                page = __alloc_pages_direct_compact(gfp_mask, order,
2549                                        zonelist, high_zoneidx,
2550                                        nodemask,
2551                                        alloc_flags, preferred_zone,
2552                                        migratetype, sync_migration,
2553                                        &contended_compaction,
2554                                        &deferred_compaction,
2555                                        &did_some_progress);
2556                if (page)
2557                        goto got_pg;
2558        }
2559
2560nopage:
2561        warn_alloc_failed(gfp_mask, order, NULL);
2562        return page;
2563got_pg:
2564        if (kmemcheck_enabled)
2565                kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2566
2567        return page;
2568}
2569
2570/*
2571 * This is the 'heart' of the zoned buddy allocator.
2572 */
2573struct page *
2574__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2575                        struct zonelist *zonelist, nodemask_t *nodemask)
2576{
2577        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2578        struct zone *preferred_zone;
2579        struct page *page = NULL;
2580        int migratetype = allocflags_to_migratetype(gfp_mask);
2581        unsigned int cpuset_mems_cookie;
2582        int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2583        struct mem_cgroup *memcg = NULL;
2584
2585        gfp_mask &= gfp_allowed_mask;
2586
2587        lockdep_trace_alloc(gfp_mask);
2588
2589        might_sleep_if(gfp_mask & __GFP_WAIT);
2590
2591        if (should_fail_alloc_page(gfp_mask, order))
2592                return NULL;
2593
2594        /*
2595         * Check the zones suitable for the gfp_mask contain at least one
2596         * valid zone. It's possible to have an empty zonelist as a result
2597         * of GFP_THISNODE and a memoryless node
2598         */
2599        if (unlikely(!zonelist->_zonerefs->zone))
2600                return NULL;
2601
2602        /*
2603         * Will only have any effect when __GFP_KMEMCG is set.  This is
2604         * verified in the (always inline) callee
2605         */
2606        if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2607                return NULL;
2608
2609retry_cpuset:
2610        cpuset_mems_cookie = get_mems_allowed();
2611
2612        /* The preferred zone is used for statistics later */
2613        first_zones_zonelist(zonelist, high_zoneidx,
2614                                nodemask ? : &cpuset_current_mems_allowed,
2615                                &preferred_zone);
2616        if (!preferred_zone)
2617                goto out;
2618
2619#ifdef CONFIG_CMA
2620        if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2621                alloc_flags |= ALLOC_CMA;
2622#endif
2623        /* First allocation attempt */
2624        page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2625                        zonelist, high_zoneidx, alloc_flags,
2626                        preferred_zone, migratetype);
2627        if (unlikely(!page)) {
2628                /*
2629                 * Runtime PM, block IO and its error handling path
2630                 * can deadlock because I/O on the device might not
2631                 * complete.
2632                 */
2633                gfp_mask = memalloc_noio_flags(gfp_mask);
2634                page = __alloc_pages_slowpath(gfp_mask, order,
2635                                zonelist, high_zoneidx, nodemask,
2636                                preferred_zone, migratetype);
2637        }
2638
2639        trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2640
2641out:
2642        /*
2643         * When updating a task's mems_allowed, it is possible to race with
2644         * parallel threads in such a way that an allocation can fail while
2645         * the mask is being updated. If a page allocation is about to fail,
2646         * check if the cpuset changed during allocation and if so, retry.
2647         */
2648        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2649                goto retry_cpuset;
2650
2651        memcg_kmem_commit_charge(page, memcg, order);
2652
2653        return page;
2654}
2655EXPORT_SYMBOL(__alloc_pages_nodemask);
2656
2657/*
2658 * Common helper functions.
2659 */
2660unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2661{
2662        struct page *page;
2663
2664        /*
2665         * __get_free_pages() returns a 32-bit address, which cannot represent
2666         * a highmem page
2667         */
2668        VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2669
2670        page = alloc_pages(gfp_mask, order);
2671        if (!page)
2672                return 0;
2673        return (unsigned long) page_address(page);
2674}
2675EXPORT_SYMBOL(__get_free_pages);
2676
2677unsigned long get_zeroed_page(gfp_t gfp_mask)
2678{
2679        return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2680}
2681EXPORT_SYMBOL(get_zeroed_page);
2682
2683void __free_pages(struct page *page, unsigned int order)
2684{
2685        if (put_page_testzero(page)) {
2686                if (order == 0)
2687                        free_hot_cold_page(page, 0);
2688                else
2689                        __free_pages_ok(page, order);
2690        }
2691}
2692
2693EXPORT_SYMBOL(__free_pages);
2694
2695void free_pages(unsigned long addr, unsigned int order)
2696{
2697        if (addr != 0) {
2698                VM_BUG_ON(!virt_addr_valid((void *)addr));
2699                __free_pages(virt_to_page((void *)addr), order);
2700        }
2701}
2702
2703EXPORT_SYMBOL(free_pages);
2704
2705/*
2706 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2707 * pages allocated with __GFP_KMEMCG.
2708 *
2709 * Those pages are accounted to a particular memcg, embedded in the
2710 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2711 * for that information only to find out that it is NULL for users who have no
2712 * interest in that whatsoever, we provide these functions.
2713 *
2714 * The caller knows better which flags it relies on.
2715 */
2716void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2717{
2718        memcg_kmem_uncharge_pages(page, order);
2719        __free_pages(page, order);
2720}
2721
2722void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2723{
2724        if (addr != 0) {
2725                VM_BUG_ON(!virt_addr_valid((void *)addr));
2726                __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2727        }
2728}
2729
2730static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2731{
2732        if (addr) {
2733                unsigned long alloc_end = addr + (PAGE_SIZE << order);
2734                unsigned long used = addr + PAGE_ALIGN(size);
2735
2736                split_page(virt_to_page((void *)addr), order);
2737                while (used < alloc_end) {
2738                        free_page(used);
2739                        used += PAGE_SIZE;
2740                }
2741        }
2742        return (void *)addr;
2743}
2744
2745/**
2746 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2747 * @size: the number of bytes to allocate
2748 * @gfp_mask: GFP flags for the allocation
2749 *
2750 * This function is similar to alloc_pages(), except that it allocates the
2751 * minimum number of pages to satisfy the request.  alloc_pages() can only
2752 * allocate memory in power-of-two pages.
2753 *
2754 * This function is also limited by MAX_ORDER.
2755 *
2756 * Memory allocated by this function must be released by free_pages_exact().
2757 */
2758void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2759{
2760        unsigned int order = get_order(size);
2761        unsigned long addr;
2762
2763        addr = __get_free_pages(gfp_mask, order);
2764        return make_alloc_exact(addr, order, size);
2765}
2766EXPORT_SYMBOL(alloc_pages_exact);
2767
2768/**
2769 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2770 *                         pages on a node.
2771 * @nid: the preferred node ID where memory should be allocated
2772 * @size: the number of bytes to allocate
2773 * @gfp_mask: GFP flags for the allocation
2774 *
2775 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2776 * back.
2777 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2778 * but is not exact.
2779 */
2780void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2781{
2782        unsigned order = get_order(size);
2783        struct page *p = alloc_pages_node(nid, gfp_mask, order);
2784        if (!p)
2785                return NULL;
2786        return make_alloc_exact((unsigned long)page_address(p), order, size);
2787}
2788EXPORT_SYMBOL(alloc_pages_exact_nid);
2789
2790/**
2791 * free_pages_exact - release memory allocated via alloc_pages_exact()
2792 * @virt: the value returned by alloc_pages_exact.
2793 * @size: size of allocation, same value as passed to alloc_pages_exact().
2794 *
2795 * Release the memory allocated by a previous call to alloc_pages_exact.
2796 */
2797void free_pages_exact(void *virt, size_t size)
2798{
2799        unsigned long addr = (unsigned long)virt;
2800        unsigned long end = addr + PAGE_ALIGN(size);
2801
2802        while (addr < end) {
2803                free_page(addr);
2804                addr += PAGE_SIZE;
2805        }
2806}
2807EXPORT_SYMBOL(free_pages_exact);
2808
2809/**
2810 * nr_free_zone_pages - count number of pages beyond high watermark
2811 * @offset: The zone index of the highest zone
2812 *
2813 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2814 * high watermark within all zones at or below a given zone index.  For each
2815 * zone, the number of pages is calculated as:
2816 *     present_pages - high_pages
2817 */
2818static unsigned long nr_free_zone_pages(int offset)
2819{
2820        struct zoneref *z;
2821        struct zone *zone;
2822
2823        /* Just pick one node, since fallback list is circular */
2824        unsigned long sum = 0;
2825
2826        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2827
2828        for_each_zone_zonelist(zone, z, zonelist, offset) {
2829                unsigned long size = zone->managed_pages;
2830                unsigned long high = high_wmark_pages(zone);
2831                if (size > high)
2832                        sum += size - high;
2833        }
2834
2835        return sum;
2836}
2837
2838/**
2839 * nr_free_buffer_pages - count number of pages beyond high watermark
2840 *
2841 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2842 * watermark within ZONE_DMA and ZONE_NORMAL.
2843 */
2844unsigned long nr_free_buffer_pages(void)
2845{
2846        return nr_free_zone_pages(gfp_zone(GFP_USER));
2847}
2848EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2849
2850/**
2851 * nr_free_pagecache_pages - count number of pages beyond high watermark
2852 *
2853 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2854 * high watermark within all zones.
2855 */
2856unsigned long nr_free_pagecache_pages(void)
2857{
2858        return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2859}
2860
2861static inline void show_node(struct zone *zone)
2862{
2863        if (IS_ENABLED(CONFIG_NUMA))
2864                printk("Node %d ", zone_to_nid(zone));
2865}
2866
2867void si_meminfo(struct sysinfo *val)
2868{
2869        val->totalram = totalram_pages;
2870        val->sharedram = 0;
2871        val->freeram = global_page_state(NR_FREE_PAGES);
2872        val->bufferram = nr_blockdev_pages();
2873        val->totalhigh = totalhigh_pages;
2874        val->freehigh = nr_free_highpages();
2875        val->mem_unit = PAGE_SIZE;
2876}
2877
2878EXPORT_SYMBOL(si_meminfo);
2879
2880#ifdef CONFIG_NUMA
2881void si_meminfo_node(struct sysinfo *val, int nid)
2882{
2883        pg_data_t *pgdat = NODE_DATA(nid);
2884
2885        val->totalram = pgdat->node_present_pages;
2886        val->freeram = node_page_state(nid, NR_FREE_PAGES);
2887#ifdef CONFIG_HIGHMEM
2888        val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2889        val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2890                        NR_FREE_PAGES);
2891#else
2892        val->totalhigh = 0;
2893        val->freehigh = 0;
2894#endif
2895        val->mem_unit = PAGE_SIZE;
2896}
2897#endif
2898
2899/*
2900 * Determine whether the node should be displayed or not, depending on whether
2901 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2902 */
2903bool skip_free_areas_node(unsigned int flags, int nid)
2904{
2905        bool ret = false;
2906        unsigned int cpuset_mems_cookie;
2907
2908        if (!(flags & SHOW_MEM_FILTER_NODES))
2909                goto out;
2910
2911        do {
2912                cpuset_mems_cookie = get_mems_allowed();
2913                ret = !node_isset(nid, cpuset_current_mems_allowed);
2914        } while (!put_mems_allowed(cpuset_mems_cookie));
2915out:
2916        return ret;
2917}
2918
2919#define K(x) ((x) << (PAGE_SHIFT-10))
2920
2921static void show_migration_types(unsigned char type)
2922{
2923        static const char types[MIGRATE_TYPES] = {
2924                [MIGRATE_UNMOVABLE]     = 'U',
2925                [MIGRATE_RECLAIMABLE]   = 'E',
2926                [MIGRATE_MOVABLE]       = 'M',
2927                [MIGRATE_RESERVE]       = 'R',
2928#ifdef CONFIG_CMA
2929                [MIGRATE_CMA]           = 'C',
2930#endif
2931#ifdef CONFIG_MEMORY_ISOLATION
2932                [MIGRATE_ISOLATE]       = 'I',
2933#endif
2934        };
2935        char tmp[MIGRATE_TYPES + 1];
2936        char *p = tmp;
2937        int i;
2938
2939        for (i = 0; i < MIGRATE_TYPES; i++) {
2940                if (type & (1 << i))
2941                        *p++ = types[i];
2942        }
2943
2944        *p = '\0';
2945        printk("(%s) ", tmp);
2946}
2947
2948/*
2949 * Show free area list (used inside shift_scroll-lock stuff)
2950 * We also calculate the percentage fragmentation. We do this by counting the
2951 * memory on each free list with the exception of the first item on the list.
2952 * Suppresses nodes that are not allowed by current's cpuset if
2953 * SHOW_MEM_FILTER_NODES is passed.
2954 */
2955void show_free_areas(unsigned int filter)
2956{
2957        int cpu;
2958        struct zone *zone;
2959
2960        for_each_populated_zone(zone) {
2961                if (skip_free_areas_node(filter, zone_to_nid(zone)))
2962                        continue;
2963                show_node(zone);
2964                printk("%s per-cpu:\n", zone->name);
2965
2966                for_each_online_cpu(cpu) {
2967                        struct per_cpu_pageset *pageset;
2968
2969                        pageset = per_cpu_ptr(zone->pageset, cpu);
2970
2971                        printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2972                               cpu, pageset->pcp.high,
2973                               pageset->pcp.batch, pageset->pcp.count);
2974                }
2975        }
2976
2977        printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2978                " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2979                " unevictable:%lu"
2980                " dirty:%lu writeback:%lu unstable:%lu\n"
2981                " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2982                " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2983                " free_cma:%lu\n",
2984                global_page_state(NR_ACTIVE_ANON),
2985                global_page_state(NR_INACTIVE_ANON),
2986                global_page_state(NR_ISOLATED_ANON),
2987                global_page_state(NR_ACTIVE_FILE),
2988                global_page_state(NR_INACTIVE_FILE),
2989                global_page_state(NR_ISOLATED_FILE),
2990                global_page_state(NR_UNEVICTABLE),
2991                global_page_state(NR_FILE_DIRTY),
2992                global_page_state(NR_WRITEBACK),
2993                global_page_state(NR_UNSTABLE_NFS),
2994                global_page_state(NR_FREE_PAGES),
2995                global_page_state(NR_SLAB_RECLAIMABLE),
2996                global_page_state(NR_SLAB_UNRECLAIMABLE),
2997                global_page_state(NR_FILE_MAPPED),
2998                global_page_state(NR_SHMEM),
2999                global_page_state(NR_PAGETABLE),
3000                global_page_state(NR_BOUNCE),
3001                global_page_state(NR_FREE_CMA_PAGES));
3002
3003        for_each_populated_zone(zone) {
3004                int i;
3005
3006                if (skip_free_areas_node(filter, zone_to_nid(zone)))
3007                        continue;
3008                show_node(zone);
3009                printk("%s"
3010                        " free:%lukB"
3011                        " min:%lukB"
3012                        " low:%lukB"
3013                        " high:%lukB"
3014                        " active_anon:%lukB"
3015                        " inactive_anon:%lukB"
3016                        " active_file:%lukB"
3017                        " inactive_file:%lukB"
3018                        " unevictable:%lukB"
3019                        " isolated(anon):%lukB"
3020                        " isolated(file):%lukB"
3021                        " present:%lukB"
3022                        " managed:%lukB"
3023                        " mlocked:%lukB"
3024                        " dirty:%lukB"
3025                        " writeback:%lukB"
3026                        " mapped:%lukB"
3027                        " shmem:%lukB"
3028                        " slab_reclaimable:%lukB"
3029                        " slab_unreclaimable:%lukB"
3030                        " kernel_stack:%lukB"
3031                        " pagetables:%lukB"
3032                        " unstable:%lukB"
3033                        " bounce:%lukB"
3034                        " free_cma:%lukB"
3035                        " writeback_tmp:%lukB"
3036                        " pages_scanned:%lu"
3037                        " all_unreclaimable? %s"
3038                        "\n",
3039                        zone->name,
3040                        K(zone_page_state(zone, NR_FREE_PAGES)),
3041                        K(min_wmark_pages(zone)),
3042                        K(low_wmark_pages(zone)),
3043                        K(high_wmark_pages(zone)),
3044                        K(zone_page_state(zone, NR_ACTIVE_ANON)),
3045                        K(zone_page_state(zone, NR_INACTIVE_ANON)),
3046                        K(zone_page_state(zone, NR_ACTIVE_FILE)),
3047                        K(zone_page_state(zone, NR_INACTIVE_FILE)),
3048                        K(zone_page_state(zone, NR_UNEVICTABLE)),
3049                        K(zone_page_state(zone, NR_ISOLATED_ANON)),
3050                        K(zone_page_state(zone, NR_ISOLATED_FILE)),
3051                        K(zone->present_pages),
3052                        K(zone->managed_pages),
3053                        K(zone_page_state(zone, NR_MLOCK)),
3054                        K(zone_page_state(zone, NR_FILE_DIRTY)),
3055                        K(zone_page_state(zone, NR_WRITEBACK)),
3056                        K(zone_page_state(zone, NR_FILE_MAPPED)),
3057                        K(zone_page_state(zone, NR_SHMEM)),
3058                        K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3059                        K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3060                        zone_page_state(zone, NR_KERNEL_STACK) *
3061                                THREAD_SIZE / 1024,
3062                        K(zone_page_state(zone, NR_PAGETABLE)),
3063                        K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3064                        K(zone_page_state(zone, NR_BOUNCE)),
3065                        K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3066                        K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3067                        zone->pages_scanned,
3068                        (zone->all_unreclaimable ? "yes" : "no")
3069                        );
3070                printk("lowmem_reserve[]:");
3071                for (i = 0; i < MAX_NR_ZONES; i++)
3072                        printk(" %lu", zone->lowmem_reserve[i]);
3073                printk("\n");
3074        }
3075
3076        for_each_populated_zone(zone) {
3077                unsigned long nr[MAX_ORDER], flags, order, total = 0;
3078                unsigned char types[MAX_ORDER];
3079
3080                if (skip_free_areas_node(filter, zone_to_nid(zone)))
3081                        continue;
3082                show_node(zone);
3083                printk("%s: ", zone->name);
3084
3085                spin_lock_irqsave(&zone->lock, flags);
3086                for (order = 0; order < MAX_ORDER; order++) {
3087                        struct free_area *area = &zone->free_area[order];
3088                        int type;
3089
3090                        nr[order] = area->nr_free;
3091                        total += nr[order] << order;
3092
3093                        types[order] = 0;
3094                        for (type = 0; type < MIGRATE_TYPES; type++) {
3095                                if (!list_empty(&area->free_list[type]))
3096                                        types[order] |= 1 << type;
3097                        }
3098                }
3099                spin_unlock_irqrestore(&zone->lock, flags);
3100                for (order = 0; order < MAX_ORDER; order++) {
3101                        printk("%lu*%lukB ", nr[order], K(1UL) << order);
3102                        if (nr[order])
3103                                show_migration_types(types[order]);
3104                }
3105                printk("= %lukB\n", K(total));
3106        }
3107
3108        printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3109
3110        show_swap_cache_info();
3111}
3112
3113static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3114{
3115        zoneref->zone = zone;
3116        zoneref->zone_idx = zone_idx(zone);
3117}
3118
3119/*
3120 * Builds allocation fallback zone lists.
3121 *
3122 * Add all populated zones of a node to the zonelist.
3123 */
3124static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3125                                int nr_zones, enum zone_type zone_type)
3126{
3127        struct zone *zone;
3128
3129        BUG_ON(zone_type >= MAX_NR_ZONES);
3130        zone_type++;
3131
3132        do {
3133                zone_type--;
3134                zone = pgdat->node_zones + zone_type;
3135                if (populated_zone(zone)) {
3136                        zoneref_set_zone(zone,
3137                                &zonelist->_zonerefs[nr_zones++]);
3138                        check_highest_zone(zone_type);
3139                }
3140
3141        } while (zone_type);
3142        return nr_zones;
3143}
3144
3145
3146/*
3147 *  zonelist_order:
3148 *  0 = automatic detection of better ordering.
3149 *  1 = order by ([node] distance, -zonetype)
3150 *  2 = order by (-zonetype, [node] distance)
3151 *
3152 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3153 *  the same zonelist. So only NUMA can configure this param.
3154 */
3155#define ZONELIST_ORDER_DEFAULT  0
3156#define ZONELIST_ORDER_NODE     1
3157#define ZONELIST_ORDER_ZONE     2
3158
3159/* zonelist order in the kernel.
3160 * set_zonelist_order() will set this to NODE or ZONE.
3161 */
3162static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3163static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3164
3165
3166#ifdef CONFIG_NUMA
3167/* The value user specified ....changed by config */
3168static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3169/* string for sysctl */
3170#define NUMA_ZONELIST_ORDER_LEN 16
3171char numa_zonelist_order[16] = "default";
3172
3173/*
3174 * interface for configure zonelist ordering.
3175 * command line option "numa_zonelist_order"
3176 *      = "[dD]efault   - default, automatic configuration.
3177 *      = "[nN]ode      - order by node locality, then by zone within node
3178 *      = "[zZ]one      - order by zone, then by locality within zone
3179 */
3180
3181static int __parse_numa_zonelist_order(char *s)
3182{
3183        if (*s == 'd' || *s == 'D') {
3184                user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3185        } else if (*s == 'n' || *s == 'N') {
3186                user_zonelist_order = ZONELIST_ORDER_NODE;
3187        } else if (*s == 'z' || *s == 'Z') {
3188                user_zonelist_order = ZONELIST_ORDER_ZONE;
3189        } else {
3190                printk(KERN_WARNING
3191                        "Ignoring invalid numa_zonelist_order value:  "
3192                        "%s\n", s);
3193                return -EINVAL;
3194        }
3195        return 0;
3196}
3197
3198static __init int setup_numa_zonelist_order(char *s)
3199{
3200        int ret;
3201
3202        if (!s)
3203                return 0;
3204
3205        ret = __parse_numa_zonelist_order(s);
3206        if (ret == 0)
3207                strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3208
3209        return ret;
3210}
3211early_param("numa_zonelist_order", setup_numa_zonelist_order);
3212
3213/*
3214 * sysctl handler for numa_zonelist_order
3215 */
3216int numa_zonelist_order_handler(ctl_table *table, int write,
3217                void __user *buffer, size_t *length,
3218                loff_t *ppos)
3219{
3220        char saved_string[NUMA_ZONELIST_ORDER_LEN];
3221        int ret;
3222        static DEFINE_MUTEX(zl_order_mutex);
3223
3224        mutex_lock(&zl_order_mutex);
3225        if (write)
3226                strcpy(saved_string, (char*)table->data);
3227        ret = proc_dostring(table, write, buffer, length, ppos);
3228        if (ret)
3229                goto out;
3230        if (write) {
3231                int oldval = user_zonelist_order;
3232                if (__parse_numa_zonelist_order((char*)table->data)) {
3233                        /*
3234                         * bogus value.  restore saved string
3235                         */
3236                        strncpy((char*)table->data, saved_string,
3237                                NUMA_ZONELIST_ORDER_LEN);
3238                        user_zonelist_order = oldval;
3239                } else if (oldval != user_zonelist_order) {
3240                        mutex_lock(&zonelists_mutex);
3241                        build_all_zonelists(NULL, NULL);
3242                        mutex_unlock(&zonelists_mutex);
3243                }
3244        }
3245out:
3246        mutex_unlock(&zl_order_mutex);
3247        return ret;
3248}
3249
3250
3251#define MAX_NODE_LOAD (nr_online_nodes)
3252static int node_load[MAX_NUMNODES];
3253
3254/**
3255 * find_next_best_node - find the next node that should appear in a given node's fallback list
3256 * @node: node whose fallback list we're appending
3257 * @used_node_mask: nodemask_t of already used nodes
3258 *
3259 * We use a number of factors to determine which is the next node that should
3260 * appear on a given node's fallback list.  The node should not have appeared
3261 * already in @node's fallback list, and it should be the next closest node
3262 * according to the distance array (which contains arbitrary distance values
3263 * from each node to each node in the system), and should also prefer nodes
3264 * with no CPUs, since presumably they'll have very little allocation pressure
3265 * on them otherwise.
3266 * It returns -1 if no node is found.
3267 */
3268static int find_next_best_node(int node, nodemask_t *used_node_mask)
3269{
3270        int n, val;
3271        int min_val = INT_MAX;
3272        int best_node = NUMA_NO_NODE;
3273        const struct cpumask *tmp = cpumask_of_node(0);
3274
3275        /* Use the local node if we haven't already */
3276        if (!node_isset(node, *used_node_mask)) {
3277                node_set(node, *used_node_mask);
3278                return node;
3279        }
3280
3281        for_each_node_state(n, N_MEMORY) {
3282
3283                /* Don't want a node to appear more than once */
3284                if (node_isset(n, *used_node_mask))
3285                        continue;
3286
3287                /* Use the distance array to find the distance */
3288                val = node_distance(node, n);
3289
3290                /* Penalize nodes under us ("prefer the next node") */
3291                val += (n < node);
3292
3293                /* Give preference to headless and unused nodes */
3294                tmp = cpumask_of_node(n);
3295                if (!cpumask_empty(tmp))
3296                        val += PENALTY_FOR_NODE_WITH_CPUS;
3297
3298                /* Slight preference for less loaded node */
3299                val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3300                val += node_load[n];
3301
3302                if (val < min_val) {
3303                        min_val = val;
3304                        best_node = n;
3305                }
3306        }
3307
3308        if (best_node >= 0)
3309                node_set(best_node, *used_node_mask);
3310
3311        return best_node;
3312}
3313
3314
3315/*
3316 * Build zonelists ordered by node and zones within node.
3317 * This results in maximum locality--normal zone overflows into local
3318 * DMA zone, if any--but risks exhausting DMA zone.
3319 */
3320static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3321{
3322        int j;
3323        struct zonelist *zonelist;
3324
3325        zonelist = &pgdat->node_zonelists[0];
3326        for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3327                ;
3328        j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3329                                                        MAX_NR_ZONES - 1);
3330        zonelist->_zonerefs[j].zone = NULL;
3331        zonelist->_zonerefs[j].zone_idx = 0;
3332}
3333
3334/*
3335 * Build gfp_thisnode zonelists
3336 */
3337static void build_thisnode_zonelists(pg_data_t *pgdat)
3338{
3339        int j;
3340        struct zonelist *zonelist;
3341
3342        zonelist = &pgdat->node_zonelists[1];
3343        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3344        zonelist->_zonerefs[j].zone = NULL;
3345        zonelist->_zonerefs[j].zone_idx = 0;
3346}
3347
3348/*
3349 * Build zonelists ordered by zone and nodes within zones.
3350 * This results in conserving DMA zone[s] until all Normal memory is
3351 * exhausted, but results in overflowing to remote node while memory
3352 * may still exist in local DMA zone.
3353 */
3354static int node_order[MAX_NUMNODES];
3355
3356static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3357{
3358        int pos, j, node;
3359        int zone_type;          /* needs to be signed */
3360        struct zone *z;
3361        struct zonelist *zonelist;
3362
3363        zonelist = &pgdat->node_zonelists[0];
3364        pos = 0;
3365        for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3366                for (j = 0; j < nr_nodes; j++) {
3367                        node = node_order[j];
3368                        z = &NODE_DATA(node)->node_zones[zone_type];
3369                        if (populated_zone(z)) {
3370                                zoneref_set_zone(z,
3371                                        &zonelist->_zonerefs[pos++]);
3372                                check_highest_zone(zone_type);
3373                        }
3374                }
3375        }
3376        zonelist->_zonerefs[pos].zone = NULL;
3377        zonelist->_zonerefs[pos].zone_idx = 0;
3378}
3379
3380static int default_zonelist_order(void)
3381{
3382        int nid, zone_type;
3383        unsigned long low_kmem_size,total_size;
3384        struct zone *z;
3385        int average_size;
3386        /*
3387         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3388         * If they are really small and used heavily, the system can fall
3389         * into OOM very easily.
3390         * This function detect ZONE_DMA/DMA32 size and configures zone order.
3391         */
3392        /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3393        low_kmem_size = 0;
3394        total_size = 0;
3395        for_each_online_node(nid) {
3396                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3397                        z = &NODE_DATA(nid)->node_zones[zone_type];
3398                        if (populated_zone(z)) {
3399                                if (zone_type < ZONE_NORMAL)
3400                                        low_kmem_size += z->present_pages;
3401                                total_size += z->present_pages;
3402                        } else if (zone_type == ZONE_NORMAL) {
3403                                /*
3404                                 * If any node has only lowmem, then node order
3405                                 * is preferred to allow kernel allocations
3406                                 * locally; otherwise, they can easily infringe
3407                                 * on other nodes when there is an abundance of
3408                                 * lowmem available to allocate from.
3409                                 */
3410                                return ZONELIST_ORDER_NODE;
3411                        }
3412                }
3413        }
3414        if (!low_kmem_size ||  /* there are no DMA area. */
3415            low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3416                return ZONELIST_ORDER_NODE;
3417        /*
3418         * look into each node's config.
3419         * If there is a node whose DMA/DMA32 memory is very big area on
3420         * local memory, NODE_ORDER may be suitable.
3421         */
3422        average_size = total_size /
3423                                (nodes_weight(node_states[N_MEMORY]) + 1);
3424        for_each_online_node(nid) {
3425                low_kmem_size = 0;
3426                total_size = 0;
3427                for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3428                        z = &NODE_DATA(nid)->node_zones[zone_type];
3429                        if (populated_zone(z)) {
3430                                if (zone_type < ZONE_NORMAL)
3431                                        low_kmem_size += z->present_pages;
3432                                total_size += z->present_pages;
3433                        }
3434                }
3435                if (low_kmem_size &&
3436                    total_size > average_size && /* ignore small node */
3437                    low_kmem_size > total_size * 70/100)
3438                        return ZONELIST_ORDER_NODE;
3439        }
3440        return ZONELIST_ORDER_ZONE;
3441}
3442
3443static void set_zonelist_order(void)
3444{
3445        if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3446                current_zonelist_order = default_zonelist_order();
3447        else
3448                current_zonelist_order = user_zonelist_order;
3449}
3450
3451static void build_zonelists(pg_data_t *pgdat)
3452{
3453        int j, node, load;
3454        enum zone_type i;
3455        nodemask_t used_mask;
3456        int local_node, prev_node;
3457        struct zonelist *zonelist;
3458        int order = current_zonelist_order;
3459
3460        /* initialize zonelists */
3461        for (i = 0; i < MAX_ZONELISTS; i++) {
3462                zonelist = pgdat->node_zonelists + i;
3463                zonelist->_zonerefs[0].zone = NULL;
3464                zonelist->_zonerefs[0].zone_idx = 0;
3465        }
3466
3467        /* NUMA-aware ordering of nodes */
3468        local_node = pgdat->node_id;
3469        load = nr_online_nodes;
3470        prev_node = local_node;
3471        nodes_clear(used_mask);
3472
3473        memset(node_order, 0, sizeof(node_order));
3474        j = 0;
3475
3476        while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3477                /*
3478                 * We don't want to pressure a particular node.
3479                 * So adding penalty to the first node in same
3480                 * distance group to make it round-robin.
3481                 */
3482                if (node_distance(local_node, node) !=
3483                    node_distance(local_node, prev_node))
3484                        node_load[node] = load;
3485
3486                prev_node = node;
3487                load--;
3488                if (order == ZONELIST_ORDER_NODE)
3489                        build_zonelists_in_node_order(pgdat, node);
3490                else
3491                        node_order[j++] = node; /* remember order */
3492        }
3493
3494        if (order == ZONELIST_ORDER_ZONE) {
3495                /* calculate node order -- i.e., DMA last! */
3496                build_zonelists_in_zone_order(pgdat, j);
3497        }
3498
3499        build_thisnode_zonelists(pgdat);
3500}
3501
3502/* Construct the zonelist performance cache - see further mmzone.h */
3503static void build_zonelist_cache(pg_data_t *pgdat)
3504{
3505        struct zonelist *zonelist;
3506        struct zonelist_cache *zlc;
3507        struct zoneref *z;
3508
3509        zonelist = &pgdat->node_zonelists[0];
3510        zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3511        bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3512        for (z = zonelist->_zonerefs; z->zone; z++)
3513                zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3514}
3515
3516#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3517/*
3518 * Return node id of node used for "local" allocations.
3519 * I.e., first node id of first zone in arg node's generic zonelist.
3520 * Used for initializing percpu 'numa_mem', which is used primarily
3521 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3522 */
3523int local_memory_node(int node)
3524{
3525        struct zone *zone;
3526
3527        (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3528                                   gfp_zone(GFP_KERNEL),
3529                                   NULL,
3530                                   &zone);
3531        return zone->node;
3532}
3533#endif
3534
3535#else   /* CONFIG_NUMA */
3536
3537static void set_zonelist_order(void)
3538{
3539        current_zonelist_order = ZONELIST_ORDER_ZONE;
3540}
3541
3542static void build_zonelists(pg_data_t *pgdat)
3543{
3544        int node, local_node;
3545        enum zone_type j;
3546        struct zonelist *zonelist;
3547
3548        local_node = pgdat->node_id;
3549
3550        zonelist = &pgdat->node_zonelists[0];
3551        j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3552
3553        /*
3554         * Now we build the zonelist so that it contains the zones
3555         * of all the other nodes.
3556         * We don't want to pressure a particular node, so when
3557         * building the zones for node N, we make sure that the
3558         * zones coming right after the local ones are those from
3559         * node N+1 (modulo N)
3560         */
3561        for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3562                if (!node_online(node))
3563                        continue;
3564                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3565                                                        MAX_NR_ZONES - 1);
3566        }
3567        for (node = 0; node < local_node; node++) {
3568                if (!node_online(node))
3569                        continue;
3570                j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3571                                                        MAX_NR_ZONES - 1);
3572        }
3573
3574        zonelist->_zonerefs[j].zone = NULL;
3575        zonelist->_zonerefs[j].zone_idx = 0;
3576}
3577
3578/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3579static void build_zonelist_cache(pg_data_t *pgdat)
3580{
3581        pgdat->node_zonelists[0].zlcache_ptr = NULL;
3582}
3583
3584#endif  /* CONFIG_NUMA */
3585
3586/*
3587 * Boot pageset table. One per cpu which is going to be used for all
3588 * zones and all nodes. The parameters will be set in such a way
3589 * that an item put on a list will immediately be handed over to
3590 * the buddy list. This is safe since pageset manipulation is done
3591 * with interrupts disabled.
3592 *
3593 * The boot_pagesets must be kept even after bootup is complete for
3594 * unused processors and/or zones. They do play a role for bootstrapping
3595 * hotplugged processors.
3596 *
3597 * zoneinfo_show() and maybe other functions do
3598 * not check if the processor is online before following the pageset pointer.
3599 * Other parts of the kernel may not check if the zone is available.
3600 */
3601static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3602static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3603static void setup_zone_pageset(struct zone *zone);
3604
3605/*
3606 * Global mutex to protect against size modification of zonelists
3607 * as well as to serialize pageset setup for the new populated zone.
3608 */
3609DEFINE_MUTEX(zonelists_mutex);
3610
3611/* return values int ....just for stop_machine() */
3612static int __build_all_zonelists(void *data)
3613{
3614        int nid;
3615        int cpu;
3616        pg_data_t *self = data;
3617
3618#ifdef CONFIG_NUMA
3619        memset(node_load, 0, sizeof(node_load));
3620#endif
3621
3622        if (self && !node_online(self->node_id)) {
3623                build_zonelists(self);
3624                build_zonelist_cache(self);
3625        }
3626
3627        for_each_online_node(nid) {
3628                pg_data_t *pgdat = NODE_DATA(nid);
3629
3630                build_zonelists(pgdat);
3631                build_zonelist_cache(pgdat);
3632        }
3633
3634        /*
3635         * Initialize the boot_pagesets that are going to be used
3636         * for bootstrapping processors. The real pagesets for
3637         * each zone will be allocated later when the per cpu
3638         * allocator is available.
3639         *
3640         * boot_pagesets are used also for bootstrapping offline
3641         * cpus if the system is already booted because the pagesets
3642         * are needed to initialize allocators on a specific cpu too.
3643         * F.e. the percpu allocator needs the page allocator which
3644         * needs the percpu allocator in order to allocate its pagesets
3645         * (a chicken-egg dilemma).
3646         */
3647        for_each_possible_cpu(cpu) {
3648                setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3649
3650#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3651                /*
3652                 * We now know the "local memory node" for each node--
3653                 * i.e., the node of the first zone in the generic zonelist.
3654                 * Set up numa_mem percpu variable for on-line cpus.  During
3655                 * boot, only the boot cpu should be on-line;  we'll init the
3656                 * secondary cpus' numa_mem as they come on-line.  During
3657                 * node/memory hotplug, we'll fixup all on-line cpus.
3658                 */
3659                if (cpu_online(cpu))
3660                        set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3661#endif
3662        }
3663
3664        return 0;
3665}
3666
3667/*
3668 * Called with zonelists_mutex held always
3669 * unless system_state == SYSTEM_BOOTING.
3670 */
3671void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3672{
3673        set_zonelist_order();
3674
3675        if (system_state == SYSTEM_BOOTING) {
3676                __build_all_zonelists(NULL);
3677                mminit_verify_zonelist();
3678                cpuset_init_current_mems_allowed();
3679        } else {
3680                /* we have to stop all cpus to guarantee there is no user
3681                   of zonelist */
3682#ifdef CONFIG_MEMORY_HOTPLUG
3683                if (zone)
3684                        setup_zone_pageset(zone);
3685#endif
3686                stop_machine(__build_all_zonelists, pgdat, NULL);
3687                /* cpuset refresh routine should be here */
3688        }
3689        vm_total_pages = nr_free_pagecache_pages();
3690        /*
3691         * Disable grouping by mobility if the number of pages in the
3692         * system is too low to allow the mechanism to work. It would be
3693         * more accurate, but expensive to check per-zone. This check is
3694         * made on memory-hotadd so a system can start with mobility
3695         * disabled and enable it later
3696         */
3697        if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3698                page_group_by_mobility_disabled = 1;
3699        else
3700                page_group_by_mobility_disabled = 0;
3701
3702        printk("Built %i zonelists in %s order, mobility grouping %s.  "
3703                "Total pages: %ld\n",
3704                        nr_online_nodes,
3705                        zonelist_order_name[current_zonelist_order],
3706                        page_group_by_mobility_disabled ? "off" : "on",
3707                        vm_total_pages);
3708#ifdef CONFIG_NUMA
3709        printk("Policy zone: %s\n", zone_names[policy_zone]);
3710#endif
3711}
3712
3713/*
3714 * Helper functions to size the waitqueue hash table.
3715 * Essentially these want to choose hash table sizes sufficiently
3716 * large so that collisions trying to wait on pages are rare.
3717 * But in fact, the number of active page waitqueues on typical
3718 * systems is ridiculously low, less than 200. So this is even
3719 * conservative, even though it seems large.
3720 *
3721 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3722 * waitqueues, i.e. the size of the waitq table given the number of pages.
3723 */
3724#define PAGES_PER_WAITQUEUE     256
3725
3726#ifndef CONFIG_MEMORY_HOTPLUG
3727static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3728{
3729        unsigned long size = 1;
3730
3731        pages /= PAGES_PER_WAITQUEUE;
3732
3733        while (size < pages)
3734                size <<= 1;
3735
3736        /*
3737         * Once we have dozens or even hundreds of threads sleeping
3738         * on IO we've got bigger problems than wait queue collision.
3739         * Limit the size of the wait table to a reasonable size.
3740         */
3741        size = min(size, 4096UL);
3742
3743        return max(size, 4UL);
3744}
3745#else
3746/*
3747 * A zone's size might be changed by hot-add, so it is not possible to determine
3748 * a suitable size for its wait_table.  So we use the maximum size now.
3749 *
3750 * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
3751 *
3752 *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
3753 *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3754 *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
3755 *
3756 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3757 * or more by the traditional way. (See above).  It equals:
3758 *
3759 *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
3760 *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
3761 *    powerpc (64K page size)             : =  (32G +16M)byte.
3762 */
3763static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3764{
3765        return 4096UL;
3766}
3767#endif
3768
3769/*
3770 * This is an integer logarithm so that shifts can be used later
3771 * to extract the more random high bits from the multiplicative
3772 * hash function before the remainder is taken.
3773 */
3774static inline unsigned long wait_table_bits(unsigned long size)
3775{
3776        return ffz(~size);
3777}
3778
3779#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3780
3781/*
3782 * Check if a pageblock contains reserved pages
3783 */
3784static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3785{
3786        unsigned long pfn;
3787
3788        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3789                if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3790                        return 1;
3791        }
3792        return 0;
3793}
3794
3795/*
3796 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3797 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3798 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3799 * higher will lead to a bigger reserve which will get freed as contiguous
3800 * blocks as reclaim kicks in
3801 */
3802static void setup_zone_migrate_reserve(struct zone *zone)
3803{
3804        unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3805        struct page *page;
3806        unsigned long block_migratetype;
3807        int reserve;
3808
3809        /*
3810         * Get the start pfn, end pfn and the number of blocks to reserve
3811         * We have to be careful to be aligned to pageblock_nr_pages to
3812         * make sure that we always check pfn_valid for the first page in
3813         * the block.
3814         */
3815        start_pfn = zone->zone_start_pfn;
3816        end_pfn = zone_end_pfn(zone);
3817        start_pfn = roundup(start_pfn, pageblock_nr_pages);
3818        reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3819                                                        pageblock_order;
3820
3821        /*
3822         * Reserve blocks are generally in place to help high-order atomic
3823         * allocations that are short-lived. A min_free_kbytes value that
3824         * would result in more than 2 reserve blocks for atomic allocations
3825         * is assumed to be in place to help anti-fragmentation for the
3826         * future allocation of hugepages at runtime.
3827         */
3828        reserve = min(2, reserve);
3829
3830        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3831                if (!pfn_valid(pfn))
3832                        continue;
3833                page = pfn_to_page(pfn);
3834
3835                /* Watch out for overlapping nodes */
3836                if (page_to_nid(page) != zone_to_nid(zone))
3837                        continue;
3838
3839                block_migratetype = get_pageblock_migratetype(page);
3840
3841                /* Only test what is necessary when the reserves are not met */
3842                if (reserve > 0) {
3843                        /*
3844                         * Blocks with reserved pages will never free, skip
3845                         * them.
3846                         */
3847                        block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3848                        if (pageblock_is_reserved(pfn, block_end_pfn))
3849                                continue;
3850
3851                        /* If this block is reserved, account for it */
3852                        if (block_migratetype == MIGRATE_RESERVE) {
3853                                reserve--;
3854                                continue;
3855                        }
3856
3857                        /* Suitable for reserving if this block is movable */
3858                        if (block_migratetype == MIGRATE_MOVABLE) {
3859                                set_pageblock_migratetype(page,
3860                                                        MIGRATE_RESERVE);
3861                                move_freepages_block(zone, page,
3862                                                        MIGRATE_RESERVE);
3863                                reserve--;
3864                                continue;
3865                        }
3866                }
3867
3868                /*
3869                 * If the reserve is met and this is a previous reserved block,
3870                 * take it back
3871                 */
3872                if (block_migratetype == MIGRATE_RESERVE) {
3873                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3874                        move_freepages_block(zone, page, MIGRATE_MOVABLE);
3875                }
3876        }
3877}
3878
3879/*
3880 * Initially all pages are reserved - free ones are freed
3881 * up by free_all_bootmem() once the early boot process is
3882 * done. Non-atomic initialization, single-pass.
3883 */
3884void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3885                unsigned long start_pfn, enum memmap_context context)
3886{
3887        struct page *page;
3888        unsigned long end_pfn = start_pfn + size;
3889        unsigned long pfn;
3890        struct zone *z;
3891
3892        if (highest_memmap_pfn < end_pfn - 1)
3893                highest_memmap_pfn = end_pfn - 1;
3894
3895        z = &NODE_DATA(nid)->node_zones[zone];
3896        for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3897                /*
3898                 * There can be holes in boot-time mem_map[]s
3899                 * handed to this function.  They do not
3900                 * exist on hotplugged memory.
3901                 */
3902                if (context == MEMMAP_EARLY) {
3903                        if (!early_pfn_valid(pfn))
3904                                continue;
3905                        if (!early_pfn_in_nid(pfn, nid))
3906                                continue;
3907                }
3908                page = pfn_to_page(pfn);
3909                set_page_links(page, zone, nid, pfn);
3910                mminit_verify_page_links(page, zone, nid, pfn);
3911                init_page_count(page);
3912                page_mapcount_reset(page);
3913                page_nid_reset_last(page);
3914                SetPageReserved(page);
3915                /*
3916                 * Mark the block movable so that blocks are reserved for
3917                 * movable at startup. This will force kernel allocations
3918                 * to reserve their blocks rather than leaking throughout
3919                 * the address space during boot when many long-lived
3920                 * kernel allocations are made. Later some blocks near
3921                 * the start are marked MIGRATE_RESERVE by
3922                 * setup_zone_migrate_reserve()
3923                 *
3924                 * bitmap is created for zone's valid pfn range. but memmap
3925                 * can be created for invalid pages (for alignment)
3926                 * check here not to call set_pageblock_migratetype() against
3927                 * pfn out of zone.
3928                 */
3929                if ((z->zone_start_pfn <= pfn)
3930                    && (pfn < zone_end_pfn(z))
3931                    && !(pfn & (pageblock_nr_pages - 1)))
3932                        set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3933
3934                INIT_LIST_HEAD(&page->lru);
3935#ifdef WANT_PAGE_VIRTUAL
3936                /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3937                if (!is_highmem_idx(zone))
3938                        set_page_address(page, __va(pfn << PAGE_SHIFT));
3939#endif
3940        }
3941}
3942
3943static void __meminit zone_init_free_lists(struct zone *zone)
3944{
3945        int order, t;
3946        for_each_migratetype_order(order, t) {
3947                INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3948                zone->free_area[order].nr_free = 0;
3949        }
3950}
3951
3952#ifndef __HAVE_ARCH_MEMMAP_INIT
3953#define memmap_init(size, nid, zone, start_pfn) \
3954        memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3955#endif
3956
3957static int __meminit zone_batchsize(struct zone *zone)
3958{
3959#ifdef CONFIG_MMU
3960        int batch;
3961
3962        /*
3963         * The per-cpu-pages pools are set to around 1000th of the
3964         * size of the zone.  But no more than 1/2 of a meg.
3965         *
3966         * OK, so we don't know how big the cache is.  So guess.
3967         */
3968        batch = zone->managed_pages / 1024;
3969        if (batch * PAGE_SIZE > 512 * 1024)
3970                batch = (512 * 1024) / PAGE_SIZE;
3971        batch /= 4;             /* We effectively *= 4 below */
3972        if (batch < 1)
3973                batch = 1;
3974
3975        /*
3976         * Clamp the batch to a 2^n - 1 value. Having a power
3977         * of 2 value was found to be more likely to have
3978         * suboptimal cache aliasing properties in some cases.
3979         *
3980         * For example if 2 tasks are alternately allocating
3981         * batches of pages, one task can end up with a lot
3982         * of pages of one half of the possible page colors
3983         * and the other with pages of the other colors.
3984         */
3985        batch = rounddown_pow_of_two(batch + batch/2) - 1;
3986
3987        return batch;
3988
3989#else
3990        /* The deferral and batching of frees should be suppressed under NOMMU
3991         * conditions.
3992         *
3993         * The problem is that NOMMU needs to be able to allocate large chunks
3994         * of contiguous memory as there's no hardware page translation to
3995         * assemble apparent contiguous memory from discontiguous pages.
3996         *
3997         * Queueing large contiguous runs of pages for batching, however,
3998         * causes the pages to actually be freed in smaller chunks.  As there
3999         * can be a significant delay between the individual batches being
4000         * recycled, this leads to the once large chunks of space being
4001         * fragmented and becoming unavailable for high-order allocations.
4002         */
4003        return 0;
4004#endif
4005}
4006
4007static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4008{
4009        struct per_cpu_pages *pcp;
4010        int migratetype;
4011
4012        memset(p, 0, sizeof(*p));
4013
4014        pcp = &p->pcp;
4015        pcp->count = 0;
4016        pcp->high = 6 * batch;
4017        pcp->batch = max(1UL, 1 * batch);
4018        for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4019                INIT_LIST_HEAD(&pcp->lists[migratetype]);
4020}
4021
4022/*
4023 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4024 * to the value high for the pageset p.
4025 */
4026
4027static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4028                                unsigned long high)
4029{
4030        struct per_cpu_pages *pcp;
4031
4032        pcp = &p->pcp;
4033        pcp->high = high;
4034        pcp->batch = max(1UL, high/4);
4035        if ((high/4) > (PAGE_SHIFT * 8))
4036                pcp->batch = PAGE_SHIFT * 8;
4037}
4038
4039static void __meminit setup_zone_pageset(struct zone *zone)
4040{
4041        int cpu;
4042
4043        zone->pageset = alloc_percpu(struct per_cpu_pageset);
4044
4045        for_each_possible_cpu(cpu) {
4046                struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4047
4048                setup_pageset(pcp, zone_batchsize(zone));
4049
4050                if (percpu_pagelist_fraction)
4051                        setup_pagelist_highmark(pcp,
4052                                (zone->managed_pages /
4053                                        percpu_pagelist_fraction));
4054        }
4055}
4056
4057/*
4058 * Allocate per cpu pagesets and initialize them.
4059 * Before this call only boot pagesets were available.
4060 */
4061void __init setup_per_cpu_pageset(void)
4062{
4063        struct zone *zone;
4064
4065        for_each_populated_zone(zone)
4066                setup_zone_pageset(zone);
4067}
4068
4069static noinline __init_refok
4070int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4071{
4072        int i;
4073        struct pglist_data *pgdat = zone->zone_pgdat;
4074        size_t alloc_size;
4075
4076        /*
4077         * The per-page waitqueue mechanism uses hashed waitqueues
4078         * per zone.
4079         */
4080        zone->wait_table_hash_nr_entries =
4081                 wait_table_hash_nr_entries(zone_size_pages);
4082        zone->wait_table_bits =
4083                wait_table_bits(zone->wait_table_hash_nr_entries);
4084        alloc_size = zone->wait_table_hash_nr_entries
4085                                        * sizeof(wait_queue_head_t);
4086
4087        if (!slab_is_available()) {
4088                zone->wait_table = (wait_queue_head_t *)
4089                        alloc_bootmem_node_nopanic(pgdat, alloc_size);
4090        } else {
4091                /*
4092                 * This case means that a zone whose size was 0 gets new memory
4093                 * via memory hot-add.
4094                 * But it may be the case that a new node was hot-added.  In
4095                 * this case vmalloc() will not be able to use this new node's
4096                 * memory - this wait_table must be initialized to use this new
4097                 * node itself as well.
4098                 * To use this new node's memory, further consideration will be
4099                 * necessary.
4100                 */
4101                zone->wait_table = vmalloc(alloc_size);
4102        }
4103        if (!zone->wait_table)
4104                return -ENOMEM;
4105
4106        for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4107                init_waitqueue_head(zone->wait_table + i);
4108
4109        return 0;
4110}
4111
4112static __meminit void zone_pcp_init(struct zone *zone)
4113{
4114        /*
4115         * per cpu subsystem is not up at this point. The following code
4116         * relies on the ability of the linker to provide the
4117         * offset of a (static) per cpu variable into the per cpu area.
4118         */
4119        zone->pageset = &boot_pageset;
4120
4121        if (zone->present_pages)
4122                printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
4123                        zone->name, zone->present_pages,
4124                                         zone_batchsize(zone));
4125}
4126
4127int __meminit init_currently_empty_zone(struct zone *zone,
4128                                        unsigned long zone_start_pfn,
4129                                        unsigned long size,
4130                                        enum memmap_context context)
4131{
4132        struct pglist_data *pgdat = zone->zone_pgdat;
4133        int ret;
4134        ret = zone_wait_table_init(zone, size);
4135        if (ret)
4136                return ret;
4137        pgdat->nr_zones = zone_idx(zone) + 1;
4138
4139        zone->zone_start_pfn = zone_start_pfn;
4140
4141        mminit_dprintk(MMINIT_TRACE, "memmap_init",
4142                        "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4143                        pgdat->node_id,
4144                        (unsigned long)zone_idx(zone),
4145                        zone_start_pfn, (zone_start_pfn + size));
4146
4147        zone_init_free_lists(zone);
4148
4149        return 0;
4150}
4151
4152#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4153#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4154/*
4155 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4156 * Architectures may implement their own version but if add_active_range()
4157 * was used and there are no special requirements, this is a convenient
4158 * alternative
4159 */
4160int __meminit __early_pfn_to_nid(unsigned long pfn)
4161{
4162        unsigned long start_pfn, end_pfn;
4163        int i, nid;
4164
4165        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4166                if (start_pfn <= pfn && pfn < end_pfn)
4167                        return nid;
4168        /* This is a memory hole */
4169        return -1;
4170}
4171#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4172
4173int __meminit early_pfn_to_nid(unsigned long pfn)
4174{
4175        int nid;
4176
4177        nid = __early_pfn_to_nid(pfn);
4178        if (nid >= 0)
4179                return nid;
4180        /* just returns 0 */
4181        return 0;
4182}
4183
4184#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4185bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4186{
4187        int nid;
4188
4189        nid = __early_pfn_to_nid(pfn);
4190        if (nid >= 0 && nid != node)
4191                return false;
4192        return true;
4193}
4194#endif
4195
4196/**
4197 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4198 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4199 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4200 *
4201 * If an architecture guarantees that all ranges registered with
4202 * add_active_ranges() contain no holes and may be freed, this
4203 * this function may be used instead of calling free_bootmem() manually.
4204 */
4205void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4206{
4207        unsigned long start_pfn, end_pfn;
4208        int i, this_nid;
4209
4210        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4211                start_pfn = min(start_pfn, max_low_pfn);
4212                end_pfn = min(end_pfn, max_low_pfn);
4213
4214                if (start_pfn < end_pfn)
4215                        free_bootmem_node(NODE_DATA(this_nid),
4216                                          PFN_PHYS(start_pfn),
4217                                          (end_pfn - start_pfn) << PAGE_SHIFT);
4218        }
4219}
4220
4221/**
4222 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4223 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4224 *
4225 * If an architecture guarantees that all ranges registered with
4226 * add_active_ranges() contain no holes and may be freed, this
4227 * function may be used instead of calling memory_present() manually.
4228 */
4229void __init sparse_memory_present_with_active_regions(int nid)
4230{
4231        unsigned long start_pfn, end_pfn;
4232        int i, this_nid;
4233
4234        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4235                memory_present(this_nid, start_pfn, end_pfn);
4236}
4237
4238/**
4239 * get_pfn_range_for_nid - Return the start and end page frames for a node
4240 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4241 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4242 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4243 *
4244 * It returns the start and end page frame of a node based on information
4245 * provided by an arch calling add_active_range(). If called for a node
4246 * with no available memory, a warning is printed and the start and end
4247 * PFNs will be 0.
4248 */
4249void __meminit get_pfn_range_for_nid(unsigned int nid,
4250                        unsigned long *start_pfn, unsigned long *end_pfn)
4251{
4252        unsigned long this_start_pfn, this_end_pfn;
4253        int i;
4254
4255        *start_pfn = -1UL;
4256        *end_pfn = 0;
4257
4258        for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4259                *start_pfn = min(*start_pfn, this_start_pfn);
4260                *end_pfn = max(*end_pfn, this_end_pfn);
4261        }
4262
4263        if (*start_pfn == -1UL)
4264                *start_pfn = 0;
4265}
4266
4267/*
4268 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4269 * assumption is made that zones within a node are ordered in monotonic
4270 * increasing memory addresses so that the "highest" populated zone is used
4271 */
4272static void __init find_usable_zone_for_movable(void)
4273{
4274        int zone_index;
4275        for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4276                if (zone_index == ZONE_MOVABLE)
4277                        continue;
4278
4279                if (arch_zone_highest_possible_pfn[zone_index] >
4280                                arch_zone_lowest_possible_pfn[zone_index])
4281                        break;
4282        }
4283
4284        VM_BUG_ON(zone_index == -1);
4285        movable_zone = zone_index;
4286}
4287
4288/*
4289 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4290 * because it is sized independent of architecture. Unlike the other zones,
4291 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4292 * in each node depending on the size of each node and how evenly kernelcore
4293 * is distributed. This helper function adjusts the zone ranges
4294 * provided by the architecture for a given node by using the end of the
4295 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4296 * zones within a node are in order of monotonic increases memory addresses
4297 */
4298static void __meminit adjust_zone_range_for_zone_movable(int nid,
4299                                        unsigned long zone_type,
4300                                        unsigned long node_start_pfn,
4301                                        unsigned long node_end_pfn,
4302                                        unsigned long *zone_start_pfn,
4303                                        unsigned long *zone_end_pfn)
4304{
4305        /* Only adjust if ZONE_MOVABLE is on this node */
4306        if (zone_movable_pfn[nid]) {
4307                /* Size ZONE_MOVABLE */
4308                if (zone_type == ZONE_MOVABLE) {
4309                        *zone_start_pfn = zone_movable_pfn[nid];
4310                        *zone_end_pfn = min(node_end_pfn,
4311                                arch_zone_highest_possible_pfn[movable_zone]);
4312
4313                /* Adjust for ZONE_MOVABLE starting within this range */
4314                } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4315                                *zone_end_pfn > zone_movable_pfn[nid]) {
4316                        *zone_end_pfn = zone_movable_pfn[nid];
4317
4318                /* Check if this whole range is within ZONE_MOVABLE */
4319                } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4320                        *zone_start_pfn = *zone_end_pfn;
4321        }
4322}
4323
4324/*
4325 * Return the number of pages a zone spans in a node, including holes
4326 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4327 */
4328static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4329                                        unsigned long zone_type,
4330                                        unsigned long *ignored)
4331{
4332        unsigned long node_start_pfn, node_end_pfn;
4333        unsigned long zone_start_pfn, zone_end_pfn;
4334
4335        /* Get the start and end of the node and zone */
4336        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4337        zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4338        zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4339        adjust_zone_range_for_zone_movable(nid, zone_type,
4340                                node_start_pfn, node_end_pfn,
4341                                &zone_start_pfn, &zone_end_pfn);
4342
4343        /* Check that this node has pages within the zone's required range */
4344        if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4345                return 0;
4346
4347        /* Move the zone boundaries inside the node if necessary */
4348        zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4349        zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4350
4351        /* Return the spanned pages */
4352        return zone_end_pfn - zone_start_pfn;
4353}
4354
4355/*
4356 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4357 * then all holes in the requested range will be accounted for.
4358 */
4359unsigned long __meminit __absent_pages_in_range(int nid,
4360                                unsigned long range_start_pfn,
4361                                unsigned long range_end_pfn)
4362{
4363        unsigned long nr_absent = range_end_pfn - range_start_pfn;
4364        unsigned long start_pfn, end_pfn;
4365        int i;
4366
4367        for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4368                start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4369                end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4370                nr_absent -= end_pfn - start_pfn;
4371        }
4372        return nr_absent;
4373}
4374
4375/**
4376 * absent_pages_in_range - Return number of page frames in holes within a range
4377 * @start_pfn: The start PFN to start searching for holes
4378 * @end_pfn: The end PFN to stop searching for holes
4379 *
4380 * It returns the number of pages frames in memory holes within a range.
4381 */
4382unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4383                                                        unsigned long end_pfn)
4384{
4385        return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4386}
4387
4388/* Return the number of page frames in holes in a zone on a node */
4389static unsigned long __meminit zone_absent_pages_in_node(int nid,
4390                                        unsigned long zone_type,
4391                                        unsigned long *ignored)
4392{
4393        unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4394        unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4395        unsigned long node_start_pfn, node_end_pfn;
4396        unsigned long zone_start_pfn, zone_end_pfn;
4397
4398        get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4399        zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4400        zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4401
4402        adjust_zone_range_for_zone_movable(nid, zone_type,
4403                        node_start_pfn, node_end_pfn,
4404                        &zone_start_pfn, &zone_end_pfn);
4405        return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4406}
4407
4408#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4409static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4410                                        unsigned long zone_type,
4411                                        unsigned long *zones_size)
4412{
4413        return zones_size[zone_type];
4414}
4415
4416static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4417                                                unsigned long zone_type,
4418                                                unsigned long *zholes_size)
4419{
4420        if (!zholes_size)
4421                return 0;
4422
4423        return zholes_size[zone_type];
4424}
4425
4426#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4427
4428static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4429                unsigned long *zones_size, unsigned long *zholes_size)
4430{
4431        unsigned long realtotalpages, totalpages = 0;
4432        enum zone_type i;
4433
4434        for (i = 0; i < MAX_NR_ZONES; i++)
4435                totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4436                                                                zones_size);
4437        pgdat->node_spanned_pages = totalpages;
4438
4439        realtotalpages = totalpages;
4440        for (i = 0; i < MAX_NR_ZONES; i++)
4441                realtotalpages -=
4442                        zone_absent_pages_in_node(pgdat->node_id, i,
4443                                                                zholes_size);
4444        pgdat->node_present_pages = realtotalpages;
4445        printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4446                                                        realtotalpages);
4447}
4448
4449#ifndef CONFIG_SPARSEMEM
4450/*
4451 * Calculate the size of the zone->blockflags rounded to an unsigned long
4452 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4453 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4454 * round what is now in bits to nearest long in bits, then return it in
4455 * bytes.
4456 */
4457static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4458{
4459        unsigned long usemapsize;
4460
4461        zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4462        usemapsize = roundup(zonesize, pageblock_nr_pages);
4463        usemapsize = usemapsize >> pageblock_order;
4464        usemapsize *= NR_PAGEBLOCK_BITS;
4465        usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4466
4467        return usemapsize / 8;
4468}
4469
4470static void __init setup_usemap(struct pglist_data *pgdat,
4471                                struct zone *zone,
4472                                unsigned long zone_start_pfn,
4473                                unsigned long zonesize)
4474{
4475        unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4476        zone->pageblock_flags = NULL;
4477        if (usemapsize)
4478                zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4479                                                                   usemapsize);
4480}
4481#else
4482static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4483                                unsigned long zone_start_pfn, unsigned long zonesize) {}
4484#endif /* CONFIG_SPARSEMEM */
4485
4486#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4487
4488/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4489void __init set_pageblock_order(void)
4490{
4491        unsigned int order;
4492
4493        /* Check that pageblock_nr_pages has not already been setup */
4494        if (pageblock_order)
4495                return;
4496
4497        if (HPAGE_SHIFT > PAGE_SHIFT)
4498                order = HUGETLB_PAGE_ORDER;
4499        else
4500                order = MAX_ORDER - 1;
4501
4502        /*
4503         * Assume the largest contiguous order of interest is a huge page.
4504         * This value may be variable depending on boot parameters on IA64 and
4505         * powerpc.
4506         */
4507        pageblock_order = order;
4508}
4509#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4510
4511/*
4512 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4513 * is unused as pageblock_order is set at compile-time. See
4514 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4515 * the kernel config
4516 */
4517void __init set_pageblock_order(void)
4518{
4519}
4520
4521#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4522
4523static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4524                                                   unsigned long present_pages)
4525{
4526        unsigned long pages = spanned_pages;
4527
4528        /*
4529         * Provide a more accurate estimation if there are holes within
4530         * the zone and SPARSEMEM is in use. If there are holes within the
4531         * zone, each populated memory region may cost us one or two extra
4532         * memmap pages due to alignment because memmap pages for each
4533         * populated regions may not naturally algined on page boundary.
4534         * So the (present_pages >> 4) heuristic is a tradeoff for that.
4535         */
4536        if (spanned_pages > present_pages + (present_pages >> 4) &&
4537            IS_ENABLED(CONFIG_SPARSEMEM))
4538                pages = present_pages;
4539
4540        return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4541}
4542
4543/*
4544 * Set up the zone data structures:
4545 *   - mark all pages reserved
4546 *   - mark all memory queues empty
4547 *   - clear the memory bitmaps
4548 *
4549 * NOTE: pgdat should get zeroed by caller.
4550 */
4551static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4552                unsigned long *zones_size, unsigned long *zholes_size)
4553{
4554        enum zone_type j;
4555        int nid = pgdat->node_id;
4556        unsigned long zone_start_pfn = pgdat->node_start_pfn;
4557        int ret;
4558
4559        pgdat_resize_init(pgdat);
4560#ifdef CONFIG_NUMA_BALANCING
4561        spin_lock_init(&pgdat->numabalancing_migrate_lock);
4562        pgdat->numabalancing_migrate_nr_pages = 0;
4563        pgdat->numabalancing_migrate_next_window = jiffies;
4564#endif
4565        init_waitqueue_head(&pgdat->kswapd_wait);
4566        init_waitqueue_head(&pgdat->pfmemalloc_wait);
4567        pgdat_page_cgroup_init(pgdat);
4568
4569        for (j = 0; j < MAX_NR_ZONES; j++) {
4570                struct zone *zone = pgdat->node_zones + j;
4571                unsigned long size, realsize, freesize, memmap_pages;
4572
4573                size = zone_spanned_pages_in_node(nid, j, zones_size);
4574                realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4575                                                                zholes_size);
4576
4577                /*
4578                 * Adjust freesize so that it accounts for how much memory
4579                 * is used by this zone for memmap. This affects the watermark
4580                 * and per-cpu initialisations
4581                 */
4582                memmap_pages = calc_memmap_size(size, realsize);
4583                if (freesize >= memmap_pages) {
4584                        freesize -= memmap_pages;
4585                        if (memmap_pages)
4586                                printk(KERN_DEBUG
4587                                       "  %s zone: %lu pages used for memmap\n",
4588                                       zone_names[j], memmap_pages);
4589                } else
4590                        printk(KERN_WARNING
4591                                "  %s zone: %lu pages exceeds freesize %lu\n",
4592                                zone_names[j], memmap_pages, freesize);
4593
4594                /* Account for reserved pages */
4595                if (j == 0 && freesize > dma_reserve) {
4596                        freesize -= dma_reserve;
4597                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
4598                                        zone_names[0], dma_reserve);
4599                }
4600
4601                if (!is_highmem_idx(j))
4602                        nr_kernel_pages += freesize;
4603                /* Charge for highmem memmap if there are enough kernel pages */
4604                else if (nr_kernel_pages > memmap_pages * 2)
4605                        nr_kernel_pages -= memmap_pages;
4606                nr_all_pages += freesize;
4607
4608                zone->spanned_pages = size;
4609                zone->present_pages = realsize;
4610                /*
4611                 * Set an approximate value for lowmem here, it will be adjusted
4612                 * when the bootmem allocator frees pages into the buddy system.
4613                 * And all highmem pages will be managed by the buddy system.
4614                 */
4615                zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4616#ifdef CONFIG_NUMA
4617                zone->node = nid;
4618                zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4619                                                / 100;
4620                zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4621#endif
4622                zone->name = zone_names[j];
4623                spin_lock_init(&zone->lock);
4624                spin_lock_init(&zone->lru_lock);
4625                zone_seqlock_init(zone);
4626                zone->zone_pgdat = pgdat;
4627
4628                zone_pcp_init(zone);
4629                lruvec_init(&zone->lruvec);
4630                if (!size)
4631                        continue;
4632
4633                set_pageblock_order();
4634                setup_usemap(pgdat, zone, zone_start_pfn, size);
4635                ret = init_currently_empty_zone(zone, zone_start_pfn,
4636                                                size, MEMMAP_EARLY);
4637                BUG_ON(ret);
4638                memmap_init(size, nid, j, zone_start_pfn);
4639                zone_start_pfn += size;
4640        }
4641}
4642
4643static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4644{
4645        /* Skip empty nodes */
4646        if (!pgdat->node_spanned_pages)
4647                return;
4648
4649#ifdef CONFIG_FLAT_NODE_MEM_MAP
4650        /* ia64 gets its own node_mem_map, before this, without bootmem */
4651        if (!pgdat->node_mem_map) {
4652                unsigned long size, start, end;
4653                struct page *map;
4654
4655                /*
4656                 * The zone's endpoints aren't required to be MAX_ORDER
4657                 * aligned but the node_mem_map endpoints must be in order
4658                 * for the buddy allocator to function correctly.
4659                 */
4660                start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4661                end = pgdat_end_pfn(pgdat);
4662                end = ALIGN(end, MAX_ORDER_NR_PAGES);
4663                size =  (end - start) * sizeof(struct page);
4664                map = alloc_remap(pgdat->node_id, size);
4665                if (!map)
4666                        map = alloc_bootmem_node_nopanic(pgdat, size);
4667                pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4668        }
4669#ifndef CONFIG_NEED_MULTIPLE_NODES
4670        /*
4671         * With no DISCONTIG, the global mem_map is just set as node 0's
4672         */
4673        if (pgdat == NODE_DATA(0)) {
4674                mem_map = NODE_DATA(0)->node_mem_map;
4675#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4676                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4677                        mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4678#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4679        }
4680#endif
4681#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4682}
4683
4684void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4685                unsigned long node_start_pfn, unsigned long *zholes_size)
4686{
4687        pg_data_t *pgdat = NODE_DATA(nid);
4688
4689        /* pg_data_t should be reset to zero when it's allocated */
4690        WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4691
4692        pgdat->node_id = nid;
4693        pgdat->node_start_pfn = node_start_pfn;
4694        init_zone_allows_reclaim(nid);
4695        calculate_node_totalpages(pgdat, zones_size, zholes_size);
4696
4697        alloc_node_mem_map(pgdat);
4698#ifdef CONFIG_FLAT_NODE_MEM_MAP
4699        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4700                nid, (unsigned long)pgdat,
4701                (unsigned long)pgdat->node_mem_map);
4702#endif
4703
4704        free_area_init_core(pgdat, zones_size, zholes_size);
4705}
4706
4707#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4708
4709#if MAX_NUMNODES > 1
4710/*
4711 * Figure out the number of possible node ids.
4712 */
4713static void __init setup_nr_node_ids(void)
4714{
4715        unsigned int node;
4716        unsigned int highest = 0;
4717
4718        for_each_node_mask(node, node_possible_map)
4719                highest = node;
4720        nr_node_ids = highest + 1;
4721}
4722#else
4723static inline void setup_nr_node_ids(void)
4724{
4725}
4726#endif
4727
4728/**
4729 * node_map_pfn_alignment - determine the maximum internode alignment
4730 *
4731 * This function should be called after node map is populated and sorted.
4732 * It calculates the maximum power of two alignment which can distinguish
4733 * all the nodes.
4734 *
4735 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4736 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
4737 * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
4738 * shifted, 1GiB is enough and this function will indicate so.
4739 *
4740 * This is used to test whether pfn -> nid mapping of the chosen memory
4741 * model has fine enough granularity to avoid incorrect mapping for the
4742 * populated node map.
4743 *
4744 * Returns the determined alignment in pfn's.  0 if there is no alignment
4745 * requirement (single node).
4746 */
4747unsigned long __init node_map_pfn_alignment(void)
4748{
4749        unsigned long accl_mask = 0, last_end = 0;
4750        unsigned long start, end, mask;
4751        int last_nid = -1;
4752        int i, nid;
4753
4754        for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4755                if (!start || last_nid < 0 || last_nid == nid) {
4756                        last_nid = nid;
4757                        last_end = end;
4758                        continue;
4759                }
4760
4761                /*
4762                 * Start with a mask granular enough to pin-point to the
4763                 * start pfn and tick off bits one-by-one until it becomes
4764                 * too coarse to separate the current node from the last.
4765                 */
4766                mask = ~((1 << __ffs(start)) - 1);
4767                while (mask && last_end <= (start & (mask << 1)))
4768                        mask <<= 1;
4769
4770                /* accumulate all internode masks */
4771                accl_mask |= mask;
4772        }
4773
4774        /* convert mask to number of pages */
4775        return ~accl_mask + 1;
4776}
4777
4778/* Find the lowest pfn for a node */
4779static unsigned long __init find_min_pfn_for_node(int nid)
4780{
4781        unsigned long min_pfn = ULONG_MAX;
4782        unsigned long start_pfn;
4783        int i;
4784
4785        for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4786                min_pfn = min(min_pfn, start_pfn);
4787
4788        if (min_pfn == ULONG_MAX) {
4789                printk(KERN_WARNING
4790                        "Could not find start_pfn for node %d\n", nid);
4791                return 0;
4792        }
4793
4794        return min_pfn;
4795}
4796
4797/**
4798 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4799 *
4800 * It returns the minimum PFN based on information provided via
4801 * add_active_range().
4802 */
4803unsigned long __init find_min_pfn_with_active_regions(void)
4804{
4805        return find_min_pfn_for_node(MAX_NUMNODES);
4806}
4807
4808/*
4809 * early_calculate_totalpages()
4810 * Sum pages in active regions for movable zone.
4811 * Populate N_MEMORY for calculating usable_nodes.
4812 */
4813static unsigned long __init early_calculate_totalpages(void)
4814{
4815        unsigned long totalpages = 0;
4816        unsigned long start_pfn, end_pfn;
4817        int i, nid;
4818
4819        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4820                unsigned long pages = end_pfn - start_pfn;
4821
4822                totalpages += pages;
4823                if (pages)
4824                        node_set_state(nid, N_MEMORY);
4825        }
4826        return totalpages;
4827}
4828
4829/*
4830 * Find the PFN the Movable zone begins in each node. Kernel memory
4831 * is spread evenly between nodes as long as the nodes have enough
4832 * memory. When they don't, some nodes will have more kernelcore than
4833 * others
4834 */
4835static void __init find_zone_movable_pfns_for_nodes(void)
4836{
4837        int i, nid;
4838        unsigned long usable_startpfn;
4839        unsigned long kernelcore_node, kernelcore_remaining;
4840        /* save the state before borrow the nodemask */
4841        nodemask_t saved_node_state = node_states[N_MEMORY];
4842        unsigned long totalpages = early_calculate_totalpages();
4843        int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4844
4845        /*
4846         * If movablecore was specified, calculate what size of
4847         * kernelcore that corresponds so that memory usable for
4848         * any allocation type is evenly spread. If both kernelcore
4849         * and movablecore are specified, then the value of kernelcore
4850         * will be used for required_kernelcore if it's greater than
4851         * what movablecore would have allowed.
4852         */
4853        if (required_movablecore) {
4854                unsigned long corepages;
4855
4856                /*
4857                 * Round-up so that ZONE_MOVABLE is at least as large as what
4858                 * was requested by the user
4859                 */
4860                required_movablecore =
4861                        roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4862                corepages = totalpages - required_movablecore;
4863
4864                required_kernelcore = max(required_kernelcore, corepages);
4865        }
4866
4867        /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4868        if (!required_kernelcore)
4869                goto out;
4870
4871        /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4872        find_usable_zone_for_movable();
4873        usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4874
4875restart:
4876        /* Spread kernelcore memory as evenly as possible throughout nodes */
4877        kernelcore_node = required_kernelcore / usable_nodes;
4878        for_each_node_state(nid, N_MEMORY) {
4879                unsigned long start_pfn, end_pfn;
4880
4881                /*
4882                 * Recalculate kernelcore_node if the division per node
4883                 * now exceeds what is necessary to satisfy the requested
4884                 * amount of memory for the kernel
4885                 */
4886                if (required_kernelcore < kernelcore_node)
4887                        kernelcore_node = required_kernelcore / usable_nodes;
4888
4889                /*
4890                 * As the map is walked, we track how much memory is usable
4891                 * by the kernel using kernelcore_remaining. When it is
4892                 * 0, the rest of the node is usable by ZONE_MOVABLE
4893                 */
4894                kernelcore_remaining = kernelcore_node;
4895
4896                /* Go through each range of PFNs within this node */
4897                for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4898                        unsigned long size_pages;
4899
4900                        start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4901                        if (start_pfn >= end_pfn)
4902                                continue;
4903
4904                        /* Account for what is only usable for kernelcore */
4905                        if (start_pfn < usable_startpfn) {
4906                                unsigned long kernel_pages;
4907                                kernel_pages = min(end_pfn, usable_startpfn)
4908                                                                - start_pfn;
4909
4910                                kernelcore_remaining -= min(kernel_pages,
4911                                                        kernelcore_remaining);
4912                                required_kernelcore -= min(kernel_pages,
4913                                                        required_kernelcore);
4914
4915                                /* Continue if range is now fully accounted */
4916                                if (end_pfn <= usable_startpfn) {
4917
4918                                        /*
4919                                         * Push zone_movable_pfn to the end so
4920                                         * that if we have to rebalance
4921                                         * kernelcore across nodes, we will
4922                                         * not double account here
4923                                         */
4924                                        zone_movable_pfn[nid] = end_pfn;
4925                                        continue;
4926                                }
4927                                start_pfn = usable_startpfn;
4928                        }
4929
4930                        /*
4931                         * The usable PFN range for ZONE_MOVABLE is from
4932                         * start_pfn->end_pfn. Calculate size_pages as the
4933                         * number of pages used as kernelcore
4934                         */
4935                        size_pages = end_pfn - start_pfn;
4936                        if (size_pages > kernelcore_remaining)
4937                                size_pages = kernelcore_remaining;
4938                        zone_movable_pfn[nid] = start_pfn + size_pages;
4939
4940                        /*
4941                         * Some kernelcore has been met, update counts and
4942                         * break if the kernelcore for this node has been
4943                         * satisified
4944                         */
4945                        required_kernelcore -= min(required_kernelcore,
4946                                                                size_pages);
4947                        kernelcore_remaining -= size_pages;
4948                        if (!kernelcore_remaining)
4949                                break;
4950                }
4951        }
4952
4953        /*
4954         * If there is still required_kernelcore, we do another pass with one
4955         * less node in the count. This will push zone_movable_pfn[nid] further
4956         * along on the nodes that still have memory until kernelcore is
4957         * satisified
4958         */
4959        usable_nodes--;
4960        if (usable_nodes && required_kernelcore > usable_nodes)
4961                goto restart;
4962
4963        /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4964        for (nid = 0; nid < MAX_NUMNODES; nid++)
4965                zone_movable_pfn[nid] =
4966                        roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4967
4968out:
4969        /* restore the node_state */
4970        node_states[N_MEMORY] = saved_node_state;
4971}
4972
4973/* Any regular or high memory on that node ? */
4974static void check_for_memory(pg_data_t *pgdat, int nid)
4975{
4976        enum zone_type zone_type;
4977
4978        if (N_MEMORY == N_NORMAL_MEMORY)
4979                return;
4980
4981        for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
4982                struct zone *zone = &pgdat->node_zones[zone_type];
4983                if (zone->present_pages) {
4984                        node_set_state(nid, N_HIGH_MEMORY);
4985                        if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
4986                            zone_type <= ZONE_NORMAL)
4987                                node_set_state(nid, N_NORMAL_MEMORY);
4988                        break;
4989                }
4990        }
4991}
4992
4993/**
4994 * free_area_init_nodes - Initialise all pg_data_t and zone data
4995 * @max_zone_pfn: an array of max PFNs for each zone
4996 *
4997 * This will call free_area_init_node() for each active node in the system.
4998 * Using the page ranges provided by add_active_range(), the size of each
4999 * zone in each node and their holes is calculated. If the maximum PFN
5000 * between two adjacent zones match, it is assumed that the zone is empty.
5001 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5002 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5003 * starts where the previous one ended. For example, ZONE_DMA32 starts
5004 * at arch_max_dma_pfn.
5005 */
5006void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5007{
5008        unsigned long start_pfn, end_pfn;
5009        int i, nid;
5010
5011        /* Record where the zone boundaries are */
5012        memset(arch_zone_lowest_possible_pfn, 0,
5013                                sizeof(arch_zone_lowest_possible_pfn));
5014        memset(arch_zone_highest_possible_pfn, 0,
5015                                sizeof(arch_zone_highest_possible_pfn));
5016        arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5017        arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5018        for (i = 1; i < MAX_NR_ZONES; i++) {
5019                if (i == ZONE_MOVABLE)
5020                        continue;
5021                arch_zone_lowest_possible_pfn[i] =
5022                        arch_zone_highest_possible_pfn[i-1];
5023                arch_zone_highest_possible_pfn[i] =
5024                        max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5025        }
5026        arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5027        arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5028
5029        /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5030        memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5031        find_zone_movable_pfns_for_nodes();
5032
5033        /* Print out the zone ranges */
5034        printk("Zone ranges:\n");
5035        for (i = 0; i < MAX_NR_ZONES; i++) {
5036                if (i == ZONE_MOVABLE)
5037                        continue;
5038                printk(KERN_CONT "  %-8s ", zone_names[i]);
5039                if (arch_zone_lowest_possible_pfn[i] ==
5040                                arch_zone_highest_possible_pfn[i])
5041                        printk(KERN_CONT "empty\n");
5042                else
5043                        printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5044                                arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5045                                (arch_zone_highest_possible_pfn[i]
5046                                        << PAGE_SHIFT) - 1);
5047        }
5048
5049        /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5050        printk("Movable zone start for each node\n");
5051        for (i = 0; i < MAX_NUMNODES; i++) {
5052                if (zone_movable_pfn[i])
5053                        printk("  Node %d: %#010lx\n", i,
5054                               zone_movable_pfn[i] << PAGE_SHIFT);
5055        }
5056
5057        /* Print out the early node map */
5058        printk("Early memory node ranges\n");
5059        for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5060                printk("  node %3d: [mem %#010lx-%#010lx]\n", nid,
5061                       start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5062
5063        /* Initialise every node */
5064        mminit_verify_pageflags_layout();
5065        setup_nr_node_ids();
5066        for_each_online_node(nid) {
5067                pg_data_t *pgdat = NODE_DATA(nid);
5068                free_area_init_node(nid, NULL,
5069                                find_min_pfn_for_node(nid), NULL);
5070
5071                /* Any memory on that node */
5072                if (pgdat->node_present_pages)
5073                        node_set_state(nid, N_MEMORY);
5074                check_for_memory(pgdat, nid);
5075        }
5076}
5077
5078static int __init cmdline_parse_core(char *p, unsigned long *core)
5079{
5080        unsigned long long coremem;
5081        if (!p)
5082                return -EINVAL;
5083
5084        coremem = memparse(p, &p);
5085        *core = coremem >> PAGE_SHIFT;
5086
5087        /* Paranoid check that UL is enough for the coremem value */
5088        WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5089
5090        return 0;
5091}
5092
5093/*
5094 * kernelcore=size sets the amount of memory for use for allocations that
5095 * cannot be reclaimed or migrated.
5096 */
5097static int __init cmdline_parse_kernelcore(char *p)
5098{
5099        return cmdline_parse_core(p, &required_kernelcore);
5100}
5101
5102/*
5103 * movablecore=size sets the amount of memory for use for allocations that
5104 * can be reclaimed or migrated.
5105 */
5106static int __init cmdline_parse_movablecore(char *p)
5107{
5108        return cmdline_parse_core(p, &required_movablecore);
5109}
5110
5111early_param("kernelcore", cmdline_parse_kernelcore);
5112early_param("movablecore", cmdline_parse_movablecore);
5113
5114#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5115
5116/**
5117 * set_dma_reserve - set the specified number of pages reserved in the first zone
5118 * @new_dma_reserve: The number of pages to mark reserved
5119 *
5120 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5121 * In the DMA zone, a significant percentage may be consumed by kernel image
5122 * and other unfreeable allocations which can skew the watermarks badly. This
5123 * function may optionally be used to account for unfreeable pages in the
5124 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5125 * smaller per-cpu batchsize.
5126 */
5127void __init set_dma_reserve(unsigned long new_dma_reserve)
5128{
5129        dma_reserve = new_dma_reserve;
5130}
5131
5132void __init free_area_init(unsigned long *zones_size)
5133{
5134        free_area_init_node(0, zones_size,
5135                        __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5136}
5137
5138static int page_alloc_cpu_notify(struct notifier_block *self,
5139                                 unsigned long action, void *hcpu)
5140{
5141        int cpu = (unsigned long)hcpu;
5142
5143        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5144                lru_add_drain_cpu(cpu);
5145                drain_pages(cpu);
5146
5147                /*
5148                 * Spill the event counters of the dead processor
5149                 * into the current processors event counters.
5150                 * This artificially elevates the count of the current
5151                 * processor.
5152                 */
5153                vm_events_fold_cpu(cpu);
5154
5155                /*
5156                 * Zero the differential counters of the dead processor
5157                 * so that the vm statistics are consistent.
5158                 *
5159                 * This is only okay since the processor is dead and cannot
5160                 * race with what we are doing.
5161                 */
5162                refresh_cpu_vm_stats(cpu);
5163        }
5164        return NOTIFY_OK;
5165}
5166
5167void __init page_alloc_init(void)
5168{
5169        hotcpu_notifier(page_alloc_cpu_notify, 0);
5170}
5171
5172/*
5173 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5174 *      or min_free_kbytes changes.
5175 */
5176static void calculate_totalreserve_pages(void)
5177{
5178        struct pglist_data *pgdat;
5179        unsigned long reserve_pages = 0;
5180        enum zone_type i, j;
5181
5182        for_each_online_pgdat(pgdat) {
5183                for (i = 0; i < MAX_NR_ZONES; i++) {
5184                        struct zone *zone = pgdat->node_zones + i;
5185                        unsigned long max = 0;
5186
5187                        /* Find valid and maximum lowmem_reserve in the zone */
5188                        for (j = i; j < MAX_NR_ZONES; j++) {
5189                                if (zone->lowmem_reserve[j] > max)
5190                                        max = zone->lowmem_reserve[j];
5191                        }
5192
5193                        /* we treat the high watermark as reserved pages. */
5194                        max += high_wmark_pages(zone);
5195
5196                        if (max > zone->managed_pages)
5197                                max = zone->managed_pages;
5198                        reserve_pages += max;
5199                        /*
5200                         * Lowmem reserves are not available to
5201                         * GFP_HIGHUSER page cache allocations and
5202                         * kswapd tries to balance zones to their high
5203                         * watermark.  As a result, neither should be
5204                         * regarded as dirtyable memory, to prevent a
5205                         * situation where reclaim has to clean pages
5206                         * in order to balance the zones.
5207                         */
5208                        zone->dirty_balance_reserve = max;
5209                }
5210        }
5211        dirty_balance_reserve = reserve_pages;
5212        totalreserve_pages = reserve_pages;
5213}
5214
5215/*
5216 * setup_per_zone_lowmem_reserve - called whenever
5217 *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
5218 *      has a correct pages reserved value, so an adequate number of
5219 *      pages are left in the zone after a successful __alloc_pages().
5220 */
5221static void setup_per_zone_lowmem_reserve(void)
5222{
5223        struct pglist_data *pgdat;
5224        enum zone_type j, idx;
5225
5226        for_each_online_pgdat(pgdat) {
5227                for (j = 0; j < MAX_NR_ZONES; j++) {
5228                        struct zone *zone = pgdat->node_zones + j;
5229                        unsigned long managed_pages = zone->managed_pages;
5230
5231                        zone->lowmem_reserve[j] = 0;
5232
5233                        idx = j;
5234                        while (idx) {
5235                                struct zone *lower_zone;
5236
5237                                idx--;
5238
5239                                if (sysctl_lowmem_reserve_ratio[idx] < 1)
5240                                        sysctl_lowmem_reserve_ratio[idx] = 1;
5241
5242                                lower_zone = pgdat->node_zones + idx;
5243                                lower_zone->lowmem_reserve[j] = managed_pages /
5244                                        sysctl_lowmem_reserve_ratio[idx];
5245                                managed_pages += lower_zone->managed_pages;
5246                        }
5247                }
5248        }
5249
5250        /* update totalreserve_pages */
5251        calculate_totalreserve_pages();
5252}
5253
5254static void __setup_per_zone_wmarks(void)
5255{
5256        unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5257        unsigned long lowmem_pages = 0;
5258        struct zone *zone;
5259        unsigned long flags;
5260
5261        /* Calculate total number of !ZONE_HIGHMEM pages */
5262        for_each_zone(zone) {
5263                if (!is_highmem(zone))
5264                        lowmem_pages += zone->managed_pages;
5265        }
5266
5267        for_each_zone(zone) {
5268                u64 tmp;
5269
5270                spin_lock_irqsave(&zone->lock, flags);
5271                tmp = (u64)pages_min * zone->managed_pages;
5272                do_div(tmp, lowmem_pages);
5273                if (is_highmem(zone)) {
5274                        /*
5275                         * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5276                         * need highmem pages, so cap pages_min to a small
5277                         * value here.
5278                         *
5279                         * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5280                         * deltas controls asynch page reclaim, and so should
5281                         * not be capped for highmem.
5282                         */
5283                        unsigned long min_pages;
5284
5285                        min_pages = zone->managed_pages / 1024;
5286                        min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5287                        zone->watermark[WMARK_MIN] = min_pages;
5288                } else {
5289                        /*
5290                         * If it's a lowmem zone, reserve a number of pages
5291                         * proportionate to the zone's size.
5292                         */
5293                        zone->watermark[WMARK_MIN] = tmp;
5294                }
5295
5296                zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + (tmp >> 2);
5297                zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5298
5299                setup_zone_migrate_reserve(zone);
5300                spin_unlock_irqrestore(&zone->lock, flags);
5301        }
5302
5303        /* update totalreserve_pages */
5304        calculate_totalreserve_pages();
5305}
5306
5307/**
5308 * setup_per_zone_wmarks - called when min_free_kbytes changes
5309 * or when memory is hot-{added|removed}
5310 *
5311 * Ensures that the watermark[min,low,high] values for each zone are set
5312 * correctly with respect to min_free_kbytes.
5313 */
5314void setup_per_zone_wmarks(void)
5315{
5316        mutex_lock(&zonelists_mutex);
5317        __setup_per_zone_wmarks();
5318        mutex_unlock(&zonelists_mutex);
5319}
5320
5321/*
5322 * The inactive anon list should be small enough that the VM never has to
5323 * do too much work, but large enough that each inactive page has a chance
5324 * to be referenced again before it is swapped out.
5325 *
5326 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5327 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5328 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5329 * the anonymous pages are kept on the inactive list.
5330 *
5331 * total     target    max
5332 * memory    ratio     inactive anon
5333 * -------------------------------------
5334 *   10MB       1         5MB
5335 *  100MB       1        50MB
5336 *    1GB       3       250MB
5337 *   10GB      10       0.9GB
5338 *  100GB      31         3GB
5339 *    1TB     101        10GB
5340 *   10TB     320        32GB
5341 */
5342static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5343{
5344        unsigned int gb, ratio;
5345
5346        /* Zone size in gigabytes */
5347        gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5348        if (gb)
5349                ratio = int_sqrt(10 * gb);
5350        else
5351                ratio = 1;
5352
5353        zone->inactive_ratio = ratio;
5354}
5355
5356static void __meminit setup_per_zone_inactive_ratio(void)
5357{
5358        struct zone *zone;
5359
5360        for_each_zone(zone)
5361                calculate_zone_inactive_ratio(zone);
5362}
5363
5364/*
5365 * Initialise min_free_kbytes.
5366 *
5367 * For small machines we want it small (128k min).  For large machines
5368 * we want it large (64MB max).  But it is not linear, because network
5369 * bandwidth does not increase linearly with machine size.  We use
5370 *
5371 *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5372 *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
5373 *
5374 * which yields
5375 *
5376 * 16MB:        512k
5377 * 32MB:        724k
5378 * 64MB:        1024k
5379 * 128MB:       1448k
5380 * 256MB:       2048k
5381 * 512MB:       2896k
5382 * 1024MB:      4096k
5383 * 2048MB:      5792k
5384 * 4096MB:      8192k
5385 * 8192MB:      11584k
5386 * 16384MB:     16384k
5387 */
5388int __meminit init_per_zone_wmark_min(void)
5389{
5390        unsigned long lowmem_kbytes;
5391
5392        lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5393
5394        min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5395        if (min_free_kbytes < 128)
5396                min_free_kbytes = 128;
5397        if (min_free_kbytes > 65536)
5398                min_free_kbytes = 65536;
5399        setup_per_zone_wmarks();
5400        refresh_zone_stat_thresholds();
5401        setup_per_zone_lowmem_reserve();
5402        setup_per_zone_inactive_ratio();
5403        return 0;
5404}
5405module_init(init_per_zone_wmark_min)
5406
5407/*
5408 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
5409 *      that we can call two helper functions whenever min_free_kbytes
5410 *      changes.
5411 */
5412int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
5413        void __user *buffer, size_t *length, loff_t *ppos)
5414{
5415        proc_dointvec(table, write, buffer, length, ppos);
5416        if (write)
5417                setup_per_zone_wmarks();
5418        return 0;
5419}
5420
5421#ifdef CONFIG_NUMA
5422int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5423        void __user *buffer, size_t *length, loff_t *ppos)
5424{
5425        struct zone *zone;
5426        int rc;
5427
5428        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5429        if (rc)
5430                return rc;
5431
5432        for_each_zone(zone)
5433                zone->min_unmapped_pages = (zone->managed_pages *
5434                                sysctl_min_unmapped_ratio) / 100;
5435        return 0;
5436}
5437
5438int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5439        void __user *buffer, size_t *length, loff_t *ppos)
5440{
5441        struct zone *zone;
5442        int rc;
5443
5444        rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5445        if (rc)
5446                return rc;
5447
5448        for_each_zone(zone)
5449                zone->min_slab_pages = (zone->managed_pages *
5450                                sysctl_min_slab_ratio) / 100;
5451        return 0;
5452}
5453#endif
5454
5455/*
5456 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5457 *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5458 *      whenever sysctl_lowmem_reserve_ratio changes.
5459 *
5460 * The reserve ratio obviously has absolutely no relation with the
5461 * minimum watermarks. The lowmem reserve ratio can only make sense
5462 * if in function of the boot time zone sizes.
5463 */
5464int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5465        void __user *buffer, size_t *length, loff_t *ppos)
5466{
5467        proc_dointvec_minmax(table, write, buffer, length, ppos);
5468        setup_per_zone_lowmem_reserve();
5469        return 0;
5470}
5471
5472/*
5473 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5474 * cpu.  It is the fraction of total pages in each zone that a hot per cpu pagelist
5475 * can have before it gets flushed back to buddy allocator.
5476 */
5477
5478int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5479        void __user *buffer, size_t *length, loff_t *ppos)
5480{
5481        struct zone *zone;
5482        unsigned int cpu;
5483        int ret;
5484
5485        ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5486        if (!write || (ret < 0))
5487                return ret;
5488        for_each_populated_zone(zone) {
5489                for_each_possible_cpu(cpu) {
5490                        unsigned long  high;
5491                        high = zone->managed_pages / percpu_pagelist_fraction;
5492                        setup_pagelist_highmark(
5493                                per_cpu_ptr(zone->pageset, cpu), high);
5494                }
5495        }
5496        return 0;
5497}
5498
5499int hashdist = HASHDIST_DEFAULT;
5500
5501#ifdef CONFIG_NUMA
5502static int __init set_hashdist(char *str)
5503{
5504        if (!str)
5505                return 0;
5506        hashdist = simple_strtoul(str, &str, 0);
5507        return 1;
5508}
5509__setup("hashdist=", set_hashdist);
5510#endif
5511
5512/*
5513 * allocate a large system hash table from bootmem
5514 * - it is assumed that the hash table must contain an exact power-of-2
5515 *   quantity of entries
5516 * - limit is the number of hash buckets, not the total allocation size
5517 */
5518void *__init alloc_large_system_hash(const char *tablename,
5519                                     unsigned long bucketsize,
5520                                     unsigned long numentries,
5521                                     int scale,
5522                                     int flags,
5523                                     unsigned int *_hash_shift,
5524                                     unsigned int *_hash_mask,
5525                                     unsigned long low_limit,
5526                                     unsigned long high_limit)
5527{
5528        unsigned long long max = high_limit;
5529        unsigned long log2qty, size;
5530        void *table = NULL;
5531
5532        /* allow the kernel cmdline to have a say */
5533        if (!numentries) {
5534                /* round applicable memory size up to nearest megabyte */
5535                numentries = nr_kernel_pages;
5536                numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5537                numentries >>= 20 - PAGE_SHIFT;
5538                numentries <<= 20 - PAGE_SHIFT;
5539
5540                /* limit to 1 bucket per 2^scale bytes of low memory */
5541                if (scale > PAGE_SHIFT)
5542                        numentries >>= (scale - PAGE_SHIFT);
5543                else
5544                        numentries <<= (PAGE_SHIFT - scale);
5545
5546                /* Make sure we've got at least a 0-order allocation.. */
5547                if (unlikely(flags & HASH_SMALL)) {
5548                        /* Makes no sense without HASH_EARLY */
5549                        WARN_ON(!(flags & HASH_EARLY));
5550                        if (!(numentries >> *_hash_shift)) {
5551                                numentries = 1UL << *_hash_shift;
5552                                BUG_ON(!numentries);
5553                        }
5554                } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5555                        numentries = PAGE_SIZE / bucketsize;
5556        }
5557        numentries = roundup_pow_of_two(numentries);
5558
5559        /* limit allocation size to 1/16 total memory by default */
5560        if (max == 0) {
5561                max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5562                do_div(max, bucketsize);
5563        }
5564        max = min(max, 0x80000000ULL);
5565
5566        if (numentries < low_limit)
5567                numentries = low_limit;
5568        if (numentries > max)
5569                numentries = max;
5570
5571        log2qty = ilog2(numentries);
5572
5573        do {
5574                size = bucketsize << log2qty;
5575                if (flags & HASH_EARLY)
5576                        table = alloc_bootmem_nopanic(size);
5577                else if (hashdist)
5578                        table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5579                else {
5580                        /*
5581                         * If bucketsize is not a power-of-two, we may free
5582                         * some pages at the end of hash table which
5583                         * alloc_pages_exact() automatically does
5584                         */
5585                        if (get_order(size) < MAX_ORDER) {
5586                                table = alloc_pages_exact(size, GFP_ATOMIC);
5587                                kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5588                        }
5589                }
5590        } while (!table && size > PAGE_SIZE && --log2qty);
5591
5592        if (!table)
5593                panic("Failed to allocate %s hash table\n", tablename);
5594
5595        printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5596               tablename,
5597               (1UL << log2qty),
5598               ilog2(size) - PAGE_SHIFT,
5599               size);
5600
5601        if (_hash_shift)
5602                *_hash_shift = log2qty;
5603        if (_hash_mask)
5604                *_hash_mask = (1 << log2qty) - 1;
5605
5606        return table;
5607}
5608
5609/* Return a pointer to the bitmap storing bits affecting a block of pages */
5610static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5611                                                        unsigned long pfn)
5612{
5613#ifdef CONFIG_SPARSEMEM
5614        return __pfn_to_section(pfn)->pageblock_flags;
5615#else
5616        return zone->pageblock_flags;
5617#endif /* CONFIG_SPARSEMEM */
5618}
5619
5620static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5621{
5622#ifdef CONFIG_SPARSEMEM
5623        pfn &= (PAGES_PER_SECTION-1);
5624        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5625#else
5626        pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5627        return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5628#endif /* CONFIG_SPARSEMEM */
5629}
5630
5631/**
5632 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5633 * @page: The page within the block of interest
5634 * @start_bitidx: The first bit of interest to retrieve
5635 * @end_bitidx: The last bit of interest
5636 * returns pageblock_bits flags
5637 */
5638unsigned long get_pageblock_flags_group(struct page *page,
5639                                        int start_bitidx, int end_bitidx)
5640{
5641        struct zone *zone;
5642        unsigned long *bitmap;
5643        unsigned long pfn, bitidx;
5644        unsigned long flags = 0;
5645        unsigned long value = 1;
5646
5647        zone = page_zone(page);
5648        pfn = page_to_pfn(page);
5649        bitmap = get_pageblock_bitmap(zone, pfn);
5650        bitidx = pfn_to_bitidx(zone, pfn);
5651
5652        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5653                if (test_bit(bitidx + start_bitidx, bitmap))
5654                        flags |= value;
5655
5656        return flags;
5657}
5658
5659/**
5660 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5661 * @page: The page within the block of interest
5662 * @start_bitidx: The first bit of interest
5663 * @end_bitidx: The last bit of interest
5664 * @flags: The flags to set
5665 */
5666void set_pageblock_flags_group(struct page *page, unsigned long flags,
5667                                        int start_bitidx, int end_bitidx)
5668{
5669        struct zone *zone;
5670        unsigned long *bitmap;
5671        unsigned long pfn, bitidx;
5672        unsigned long value = 1;
5673
5674        zone = page_zone(page);
5675        pfn = page_to_pfn(page);
5676        bitmap = get_pageblock_bitmap(zone, pfn);
5677        bitidx = pfn_to_bitidx(zone, pfn);
5678        VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5679
5680        for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5681                if (flags & value)
5682                        __set_bit(bitidx + start_bitidx, bitmap);
5683                else
5684                        __clear_bit(bitidx + start_bitidx, bitmap);
5685}
5686
5687/*
5688 * This function checks whether pageblock includes unmovable pages or not.
5689 * If @count is not zero, it is okay to include less @count unmovable pages
5690 *
5691 * PageLRU check wihtout isolation or lru_lock could race so that
5692 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5693 * expect this function should be exact.
5694 */
5695bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5696                         bool skip_hwpoisoned_pages)
5697{
5698        unsigned long pfn, iter, found;
5699        int mt;
5700
5701        /*
5702         * For avoiding noise data, lru_add_drain_all() should be called
5703         * If ZONE_MOVABLE, the zone never contains unmovable pages
5704         */
5705        if (zone_idx(zone) == ZONE_MOVABLE)
5706                return false;
5707        mt = get_pageblock_migratetype(page);
5708        if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5709                return false;
5710
5711        pfn = page_to_pfn(page);
5712        for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5713                unsigned long check = pfn + iter;
5714
5715                if (!pfn_valid_within(check))